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         dev_info->tx_offload_capa = DEV_TX_OFFLOAD_MULTI_SEGS;
 203
 204         return 0;
 205 }
 206
 207 static memif_ring_t *
 208 memif_get_ring(struct pmd_internals *pmd, struct pmd_process_private *proc_private,
 209                memif_ring_type_t type, uint16_t ring_num)
 210 {
 211         /* rings only in region 0 */
 212         void *p = proc_private->regions[0]->addr;
 213         int ring_size = sizeof(memif_ring_t) + sizeof(memif_desc_t) *
 214             (1 << pmd->run.log2_ring_size);
 215
 216         p = (uint8_t *)p + (ring_num + type * pmd->run.num_c2s_rings) * ring_size;
 217
 218         return (memif_ring_t *)p;
 219 }
 220
 221 static memif_region_offset_t
 222 memif_get_ring_offset(struct rte_eth_dev *dev, struct memif_queue *mq,
 223                       memif_ring_type_t type, uint16_t num)
 224 {
 225         struct pmd_internals *pmd = dev->data->dev_private;
 226         struct pmd_process_private *proc_private = dev->process_private;
 227
 228         return ((uint8_t *)memif_get_ring(pmd, proc_private, type, num) -
 229                 (uint8_t *)proc_private->regions[mq->region]->addr);
 230 }
 231
 232 static memif_ring_t *
 233 memif_get_ring_from_queue(struct pmd_process_private *proc_private,
 234                           struct memif_queue *mq)
 235 {
 236         struct memif_region *r;
 237
 238         r = proc_private->regions[mq->region];
 239         if (r == NULL)
 240                 return NULL;
 241
 242         return (memif_ring_t *)((uint8_t *)r->addr + mq->ring_offset);
 243 }
 244
 245 static void *
 246 memif_get_buffer(struct pmd_process_private *proc_private, memif_desc_t *d)
 247 {
 248         return ((uint8_t *)proc_private->regions[d->region]->addr + d->offset);
 249 }
 250
 251 /* Free mbufs received by server */
 252 static void
 253 memif_free_stored_mbufs(struct pmd_process_private *proc_private, struct memif_queue *mq)
 254 {
 255         uint16_t cur_tail;
 256         uint16_t mask = (1 << mq->log2_ring_size) - 1;
 257         memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
 258
 259         /* FIXME: improve performance */
 260         /* The ring->tail acts as a guard variable between Tx and Rx
 261          * threads, so using load-acquire pairs with store-release
 262          * in function eth_memif_rx for C2S queues.
 263          */
 264         cur_tail = __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
 265         while (mq->last_tail != cur_tail) {
 266                 RTE_MBUF_PREFETCH_TO_FREE(mq->buffers[(mq->last_tail + 1) & mask]);
 267                 /* Decrement refcnt and free mbuf. (current segment) */
 268                 rte_mbuf_refcnt_update(mq->buffers[mq->last_tail & mask], -1);
 269                 rte_pktmbuf_free_seg(mq->buffers[mq->last_tail & mask]);
 270                 mq->last_tail++;
 271         }
 272 }
 273
 274 static int
 275 memif_pktmbuf_chain(struct rte_mbuf *head, struct rte_mbuf *cur_tail,
 276                     struct rte_mbuf *tail)
 277 {
 278         /* Check for number-of-segments-overflow */
 279         if (unlikely(head->nb_segs + tail->nb_segs > RTE_MBUF_MAX_NB_SEGS))
 280                 return -EOVERFLOW;
 281
 282         /* Chain 'tail' onto the old tail */
 283         cur_tail->next = tail;
 284
 285         /* accumulate number of segments and total length. */
 286         head->nb_segs = (uint16_t)(head->nb_segs + tail->nb_segs);
 287
 288         tail->pkt_len = tail->data_len;
 289         head->pkt_len += tail->pkt_len;
 290
 291         return 0;
 292 }
 293
 294 static uint16_t
 295 eth_memif_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 296 {
 297         struct memif_queue *mq = queue;
 298         struct pmd_internals *pmd = rte_eth_devices[mq->in_port].data->dev_private;
 299         struct pmd_process_private *proc_private =
 300                 rte_eth_devices[mq->in_port].process_private;
 301         memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
 302         uint16_t cur_slot, last_slot, n_slots, ring_size, mask, s0;
 303         uint16_t n_rx_pkts = 0;
 304         uint16_t mbuf_size = rte_pktmbuf_data_room_size(mq->mempool) -
 305                 RTE_PKTMBUF_HEADROOM;
 306         uint16_t src_len, src_off, dst_len, dst_off, cp_len;
 307         memif_ring_type_t type = mq->type;
 308         memif_desc_t *d0;
 309         struct rte_mbuf *mbuf, *mbuf_head, *mbuf_tail;
 310         uint64_t b;
 311         ssize_t size __rte_unused;
 312         uint16_t head;
 313         int ret;
 314         struct rte_eth_link link;
 315
 316         if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
 317                 return 0;
 318         if (unlikely(ring == NULL)) {
 319                 /* Secondary process will attempt to request regions. */
 320                 ret = rte_eth_link_get(mq->in_port, &link);
 321                 if (ret < 0)
 322                         MIF_LOG(ERR, "Failed to get port %u link info: %s",
 323                                 mq->in_port, rte_strerror(-ret));
 324                 return 0;
 325         }
 326
 327         /* consume interrupt */
 328         if ((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0)
 329                 size = read(mq->intr_handle.fd, &b, sizeof(b));
 330
 331         ring_size = 1 << mq->log2_ring_size;
 332         mask = ring_size - 1;
 333
 334         if (type == MEMIF_RING_C2S) {
 335                 cur_slot = mq->last_head;
 336                 last_slot = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE);
 337         } else {
 338                 cur_slot = mq->last_tail;
 339                 last_slot = __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
 340         }
 341
 342         if (cur_slot == last_slot)
 343                 goto refill;
 344         n_slots = last_slot - cur_slot;
 345
 346         while (n_slots && n_rx_pkts < nb_pkts) {
 347                 mbuf_head = rte_pktmbuf_alloc(mq->mempool);
 348                 if (unlikely(mbuf_head == NULL))
 349                         goto no_free_bufs;
 350                 mbuf = mbuf_head;
 351                 mbuf->port = mq->in_port;
 352
 353 next_slot:
 354                 s0 = cur_slot & mask;
 355                 d0 = &ring->desc[s0];
 356
 357                 src_len = d0->length;
 358                 dst_off = 0;
 359                 src_off = 0;
 360
 361                 do {
 362                         dst_len = mbuf_size - dst_off;
 363                         if (dst_len == 0) {
 364                                 dst_off = 0;
 365                                 dst_len = mbuf_size;
 366
 367                                 /* store pointer to tail */
 368                                 mbuf_tail = mbuf;
 369                                 mbuf = rte_pktmbuf_alloc(mq->mempool);
 370                                 if (unlikely(mbuf == NULL))
 371                                         goto no_free_bufs;
 372                                 mbuf->port = mq->in_port;
 373                                 ret = memif_pktmbuf_chain(mbuf_head, mbuf_tail, mbuf);
 374                                 if (unlikely(ret < 0)) {
 375                                         MIF_LOG(ERR, "number-of-segments-overflow");
 376                                         rte_pktmbuf_free(mbuf);
 377                                         goto no_free_bufs;
 378                                 }
 379                         }
 380                         cp_len = RTE_MIN(dst_len, src_len);
 381
 382                         rte_pktmbuf_data_len(mbuf) += cp_len;
 383                         rte_pktmbuf_pkt_len(mbuf) = rte_pktmbuf_data_len(mbuf);
 384                         if (mbuf != mbuf_head)
 385                                 rte_pktmbuf_pkt_len(mbuf_head) += cp_len;
 386
 387                         rte_memcpy(rte_pktmbuf_mtod_offset(mbuf, void *,
 388                                                            dst_off),
 389                                 (uint8_t *)memif_get_buffer(proc_private, d0) +
 390                                 src_off, cp_len);
 391
 392                         src_off += cp_len;
 393                         dst_off += cp_len;
 394                         src_len -= cp_len;
 395                 } while (src_len);
 396
 397                 cur_slot++;
 398                 n_slots--;
 399
 400                 if (d0->flags & MEMIF_DESC_FLAG_NEXT)
 401                         goto next_slot;
 402
 403                 mq->n_bytes += rte_pktmbuf_pkt_len(mbuf_head);
 404                 *bufs++ = mbuf_head;
 405                 n_rx_pkts++;
 406         }
 407
 408 no_free_bufs:
 409         if (type == MEMIF_RING_C2S) {
 410                 __atomic_store_n(&ring->tail, cur_slot, __ATOMIC_RELEASE);
 411                 mq->last_head = cur_slot;
 412         } else {
 413                 mq->last_tail = cur_slot;
 414         }
 415
 416 refill:
 417         if (type == MEMIF_RING_S2C) {
 418                 /* ring->head is updated by the receiver and this function
 419                  * is called in the context of receiver thread. The loads in
 420                  * the receiver do not need to synchronize with its own stores.
 421                  */
 422                 head = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 423                 n_slots = ring_size - head + mq->last_tail;
 424
 425                 while (n_slots--) {
 426                         s0 = head++ & mask;
 427                         d0 = &ring->desc[s0];
 428                         d0->length = pmd->run.pkt_buffer_size;
 429                 }
 430                 __atomic_store_n(&ring->head, head, __ATOMIC_RELEASE);
 431         }
 432
 433         mq->n_pkts += n_rx_pkts;
 434         return n_rx_pkts;
 435 }
 436
 437 static uint16_t
 438 eth_memif_rx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 439 {
 440         struct memif_queue *mq = queue;
 441         struct pmd_internals *pmd = rte_eth_devices[mq->in_port].data->dev_private;
 442         struct pmd_process_private *proc_private =
 443                 rte_eth_devices[mq->in_port].process_private;
 444         memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
 445         uint16_t cur_slot, last_slot, n_slots, ring_size, mask, s0, head;
 446         uint16_t n_rx_pkts = 0;
 447         memif_desc_t *d0;
 448         struct rte_mbuf *mbuf, *mbuf_tail;
 449         struct rte_mbuf *mbuf_head = NULL;
 450         int ret;
 451         struct rte_eth_link link;
 452
 453         if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
 454                 return 0;
 455         if (unlikely(ring == NULL)) {
 456                 /* Secondary process will attempt to request regions. */
 457                 rte_eth_link_get(mq->in_port, &link);
 458                 return 0;
 459         }
 460
 461         /* consume interrupt */
 462         if ((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0) {
 463                 uint64_t b;
 464                 ssize_t size __rte_unused;
 465                 size = read(mq->intr_handle.fd, &b, sizeof(b));
 466         }
 467
 468         ring_size = 1 << mq->log2_ring_size;
 469         mask = ring_size - 1;
 470
 471         cur_slot = mq->last_tail;
 472         /* The ring->tail acts as a guard variable between Tx and Rx
 473          * threads, so using load-acquire pairs with store-release
 474          * to synchronize it between threads.
 475          */
 476         last_slot = __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
 477         if (cur_slot == last_slot)
 478                 goto refill;
 479         n_slots = last_slot - cur_slot;
 480
 481         while (n_slots && n_rx_pkts < nb_pkts) {
 482                 s0 = cur_slot & mask;
 483
 484                 d0 = &ring->desc[s0];
 485                 mbuf_head = mq->buffers[s0];
 486                 mbuf = mbuf_head;
 487
 488 next_slot:
 489                 /* prefetch next descriptor */
 490                 if (n_rx_pkts + 1 < nb_pkts)
 491                         rte_prefetch0(&ring->desc[(cur_slot + 1) & mask]);
 492
 493                 mbuf->port = mq->in_port;
 494                 rte_pktmbuf_data_len(mbuf) = d0->length;
 495                 rte_pktmbuf_pkt_len(mbuf) = rte_pktmbuf_data_len(mbuf);
 496
 497                 mq->n_bytes += rte_pktmbuf_data_len(mbuf);
 498
 499                 cur_slot++;
 500                 n_slots--;
 501                 if (d0->flags & MEMIF_DESC_FLAG_NEXT) {
 502                         s0 = cur_slot & mask;
 503                         d0 = &ring->desc[s0];
 504                         mbuf_tail = mbuf;
 505                         mbuf = mq->buffers[s0];
 506                         ret = memif_pktmbuf_chain(mbuf_head, mbuf_tail, mbuf);
 507                         if (unlikely(ret < 0)) {
 508                                 MIF_LOG(ERR, "number-of-segments-overflow");
 509                                 goto refill;
 510                         }
 511                         goto next_slot;
 512                 }
 513
 514                 *bufs++ = mbuf_head;
 515                 n_rx_pkts++;
 516         }
 517
 518         mq->last_tail = cur_slot;
 519
 520 /* Supply server with new buffers */
 521 refill:
 522         /* ring->head is updated by the receiver and this function
 523          * is called in the context of receiver thread. The loads in
 524          * the receiver do not need to synchronize with its own stores.
 525          */
 526         head = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 527         n_slots = ring_size - head + mq->last_tail;
 528
 529         if (n_slots < 32)
 530                 goto no_free_mbufs;
 531
 532         ret = rte_pktmbuf_alloc_bulk(mq->mempool, &mq->buffers[head & mask], n_slots);
 533         if (unlikely(ret < 0))
 534                 goto no_free_mbufs;
 535
 536         while (n_slots--) {
 537                 s0 = head++ & mask;
 538                 if (n_slots > 0)
 539                         rte_prefetch0(mq->buffers[head & mask]);
 540                 d0 = &ring->desc[s0];
 541                 /* store buffer header */
 542                 mbuf = mq->buffers[s0];
 543                 /* populate descriptor */
 544                 d0->length = rte_pktmbuf_data_room_size(mq->mempool) -
 545                                 RTE_PKTMBUF_HEADROOM;
 546                 d0->region = 1;
 547                 d0->offset = rte_pktmbuf_mtod(mbuf, uint8_t *) -
 548                         (uint8_t *)proc_private->regions[d0->region]->addr;
 549         }
 550 no_free_mbufs:
 551         /* The ring->head acts as a guard variable between Tx and Rx
 552          * threads, so using store-release pairs with load-acquire
 553          * in function eth_memif_tx.
 554          */
 555         __atomic_store_n(&ring->head, head, __ATOMIC_RELEASE);
 556
 557         mq->n_pkts += n_rx_pkts;
 558
 559         return n_rx_pkts;
 560 }
 561
 562 static uint16_t
 563 eth_memif_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 564 {
 565         struct memif_queue *mq = queue;
 566         struct pmd_internals *pmd = rte_eth_devices[mq->in_port].data->dev_private;
 567         struct pmd_process_private *proc_private =
 568                 rte_eth_devices[mq->in_port].process_private;
 569         memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
 570         uint16_t slot, saved_slot, n_free, ring_size, mask, n_tx_pkts = 0;
 571         uint16_t src_len, src_off, dst_len, dst_off, cp_len, nb_segs;
 572         memif_ring_type_t type = mq->type;
 573         memif_desc_t *d0;
 574         struct rte_mbuf *mbuf;
 575         struct rte_mbuf *mbuf_head;
 576         uint64_t a;
 577         ssize_t size;
 578         struct rte_eth_link link;
 579
 580         if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
 581                 return 0;
 582         if (unlikely(ring == NULL)) {
 583                 int ret;
 584
 585                 /* Secondary process will attempt to request regions. */
 586                 ret = rte_eth_link_get(mq->in_port, &link);
 587                 if (ret < 0)
 588                         MIF_LOG(ERR, "Failed to get port %u link info: %s",
 589                                 mq->in_port, rte_strerror(-ret));
 590                 return 0;
 591         }
 592
 593         ring_size = 1 << mq->log2_ring_size;
 594         mask = ring_size - 1;
 595
 596         if (type == MEMIF_RING_C2S) {
 597                 /* For C2S queues ring->head is updated by the sender and
 598                  * this function is called in the context of sending thread.
 599                  * The loads in the sender do not need to synchronize with
 600                  * its own stores. Hence, the following load can be a
 601                  * relaxed load.
 602                  */
 603                 slot = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 604                 n_free = ring_size - slot +
 605                                 __atomic_load_n(&ring->tail, __ATOMIC_ACQUIRE);
 606         } else {
 607                 /* For S2C queues ring->tail is updated by the sender and
 608                  * this function is called in the context of sending thread.
 609                  * The loads in the sender do not need to synchronize with
 610                  * its own stores. Hence, the following load can be a
 611                  * relaxed load.
 612                  */
 613                 slot = __atomic_load_n(&ring->tail, __ATOMIC_RELAXED);
 614                 n_free = __atomic_load_n(&ring->head, __ATOMIC_ACQUIRE) - slot;
 615         }
 616
 617         while (n_tx_pkts < nb_pkts && n_free) {
 618                 mbuf_head = *bufs++;
 619                 nb_segs = mbuf_head->nb_segs;
 620                 mbuf = mbuf_head;
 621
 622                 saved_slot = slot;
 623                 d0 = &ring->desc[slot & mask];
 624                 dst_off = 0;
 625                 dst_len = (type == MEMIF_RING_C2S) ?
 626                         pmd->run.pkt_buffer_size : d0->length;
 627
 628 next_in_chain:
 629                 src_off = 0;
 630                 src_len = rte_pktmbuf_data_len(mbuf);
 631
 632                 while (src_len) {
 633                         if (dst_len == 0) {
 634                                 if (n_free) {
 635                                         slot++;
 636                                         n_free--;
 637                                         d0->flags |= MEMIF_DESC_FLAG_NEXT;
 638                                         d0 = &ring->desc[slot & mask];
 639                                         dst_off = 0;
 640                                         dst_len = (type == MEMIF_RING_C2S) ?
 641                                             pmd->run.pkt_buffer_size : d0->length;
 642                                         d0->flags = 0;
 643                                 } else {
 644                                         slot = saved_slot;
 645                                         goto no_free_slots;
 646                                 }
 647                         }
 648                         cp_len = RTE_MIN(dst_len, src_len);
 649
 650                         rte_memcpy((uint8_t *)memif_get_buffer(proc_private,
 651                                                                d0) + dst_off,
 652                                 rte_pktmbuf_mtod_offset(mbuf, void *, src_off),
 653                                 cp_len);
 654
 655                         mq->n_bytes += cp_len;
 656                         src_off += cp_len;
 657                         dst_off += cp_len;
 658                         src_len -= cp_len;
 659                         dst_len -= cp_len;
 660
 661                         d0->length = dst_off;
 662                 }
 663
 664                 if (--nb_segs > 0) {
 665                         mbuf = mbuf->next;
 666                         goto next_in_chain;
 667                 }
 668
 669                 n_tx_pkts++;
 670                 slot++;
 671                 n_free--;
 672                 rte_pktmbuf_free(mbuf_head);
 673         }
 674
 675 no_free_slots:
 676         if (type == MEMIF_RING_C2S)
 677                 __atomic_store_n(&ring->head, slot, __ATOMIC_RELEASE);
 678         else
 679                 __atomic_store_n(&ring->tail, slot, __ATOMIC_RELEASE);
 680
 681         if ((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0) {
 682                 a = 1;
 683                 size = write(mq->intr_handle.fd, &a, sizeof(a));
 684                 if (unlikely(size < 0)) {
 685                         MIF_LOG(WARNING,
 686                                 "Failed to send interrupt. %s", strerror(errno));
 687                 }
 688         }
 689
 690         mq->n_pkts += n_tx_pkts;
 691         return n_tx_pkts;
 692 }
 693
 694
 695 static int
 696 memif_tx_one_zc(struct pmd_process_private *proc_private, struct memif_queue *mq,
 697                 memif_ring_t *ring, struct rte_mbuf *mbuf, const uint16_t mask,
 698                 uint16_t slot, uint16_t n_free)
 699 {
 700         memif_desc_t *d0;
 701         uint16_t nb_segs = mbuf->nb_segs;
 702         int used_slots = 1;
 703
 704 next_in_chain:
 705         /* store pointer to mbuf to free it later */
 706         mq->buffers[slot & mask] = mbuf;
 707         /* Increment refcnt to make sure the buffer is not freed before server
 708          * receives it. (current segment)
 709          */
 710         rte_mbuf_refcnt_update(mbuf, 1);
 711         /* populate descriptor */
 712         d0 = &ring->desc[slot & mask];
 713         d0->length = rte_pktmbuf_data_len(mbuf);
 714         mq->n_bytes += rte_pktmbuf_data_len(mbuf);
 715         /* FIXME: get region index */
 716         d0->region = 1;
 717         d0->offset = rte_pktmbuf_mtod(mbuf, uint8_t *) -
 718                 (uint8_t *)proc_private->regions[d0->region]->addr;
 719         d0->flags = 0;
 720
 721         /* check if buffer is chained */
 722         if (--nb_segs > 0) {
 723                 if (n_free < 2)
 724                         return 0;
 725                 /* mark buffer as chained */
 726                 d0->flags |= MEMIF_DESC_FLAG_NEXT;
 727                 /* advance mbuf */
 728                 mbuf = mbuf->next;
 729                 /* update counters */
 730                 used_slots++;
 731                 slot++;
 732                 n_free--;
 733                 goto next_in_chain;
 734         }
 735         return used_slots;
 736 }
 737
 738 static uint16_t
 739 eth_memif_tx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 740 {
 741         struct memif_queue *mq = queue;
 742         struct pmd_internals *pmd = rte_eth_devices[mq->in_port].data->dev_private;
 743         struct pmd_process_private *proc_private =
 744                 rte_eth_devices[mq->in_port].process_private;
 745         memif_ring_t *ring = memif_get_ring_from_queue(proc_private, mq);
 746         uint16_t slot, n_free, ring_size, mask, n_tx_pkts = 0;
 747         struct rte_eth_link link;
 748
 749         if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
 750                 return 0;
 751         if (unlikely(ring == NULL)) {
 752                 /* Secondary process will attempt to request regions. */
 753                 rte_eth_link_get(mq->in_port, &link);
 754                 return 0;
 755         }
 756
 757         ring_size = 1 << mq->log2_ring_size;
 758         mask = ring_size - 1;
 759
 760         /* free mbufs received by server */
 761         memif_free_stored_mbufs(proc_private, mq);
 762
 763         /* ring type always MEMIF_RING_C2S */
 764         /* For C2S queues ring->head is updated by the sender and
 765          * this function is called in the context of sending thread.
 766          * The loads in the sender do not need to synchronize with
 767          * its own stores. Hence, the following load can be a
 768          * relaxed load.
 769          */
 770         slot = __atomic_load_n(&ring->head, __ATOMIC_RELAXED);
 771         n_free = ring_size - slot + mq->last_tail;
 772
 773         int used_slots;
 774
 775         while (n_free && (n_tx_pkts < nb_pkts)) {
 776                 while ((n_free > 4) && ((nb_pkts - n_tx_pkts) > 4)) {
 777                         if ((nb_pkts - n_tx_pkts) > 8) {
 778                                 rte_prefetch0(*bufs + 4);
 779                                 rte_prefetch0(*bufs + 5);
 780                                 rte_prefetch0(*bufs + 6);
 781                                 rte_prefetch0(*bufs + 7);
 782                         }
 783                         used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 784                                 mask, slot, n_free);
 785                         if (unlikely(used_slots < 1))
 786                                 goto no_free_slots;
 787                         n_tx_pkts++;
 788                         slot += used_slots;
 789                         n_free -= used_slots;
 790
 791                         used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 792                                 mask, slot, n_free);
 793                         if (unlikely(used_slots < 1))
 794                                 goto no_free_slots;
 795                         n_tx_pkts++;
 796                         slot += used_slots;
 797                         n_free -= used_slots;
 798
 799                         used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 800                                 mask, slot, n_free);
 801                         if (unlikely(used_slots < 1))
 802                                 goto no_free_slots;
 803                         n_tx_pkts++;
 804                         slot += used_slots;
 805                         n_free -= used_slots;
 806
 807                         used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 808                                 mask, slot, n_free);
 809                         if (unlikely(used_slots < 1))
 810                                 goto no_free_slots;
 811                         n_tx_pkts++;
 812                         slot += used_slots;
 813                         n_free -= used_slots;
 814                 }
 815                 used_slots = memif_tx_one_zc(proc_private, mq, ring, *bufs++,
 816                         mask, slot, n_free);
 817                 if (unlikely(used_slots < 1))
 818                         goto no_free_slots;
 819                 n_tx_pkts++;
 820                 slot += used_slots;
 821                 n_free -= used_slots;
 822         }
 823
 824 no_free_slots:
 825         /* ring type always MEMIF_RING_C2S */
 826         /* The ring->head acts as a guard variable between Tx and Rx
 827          * threads, so using store-release pairs with load-acquire
 828          * in function eth_memif_rx for C2S rings.
 829          */
 830         __atomic_store_n(&ring->head, slot, __ATOMIC_RELEASE);
 831
 832         /* Send interrupt, if enabled. */
 833         if ((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0) {
 834                 uint64_t a = 1;
 835                 ssize_t size = write(mq->intr_handle.fd, &a, sizeof(a));
 836                 if (unlikely(size < 0)) {
 837                         MIF_LOG(WARNING,
 838                                 "Failed to send interrupt. %s", strerror(errno));
 839                 }
 840         }
 841
 842         /* increment queue counters */
 843         mq->n_pkts += n_tx_pkts;
 844
 845         return n_tx_pkts;
 846 }
 847
 848 void
 849 memif_free_regions(struct rte_eth_dev *dev)
 850 {
 851         struct pmd_process_private *proc_private = dev->process_private;
 852         struct pmd_internals *pmd = dev->data->dev_private;
 853         int i;
 854         struct memif_region *r;
 855
 856         /* regions are allocated contiguously, so it's
 857          * enough to loop until 'proc_private->regions_num'
 858          */
 859         for (i = 0; i < proc_private->regions_num; i++) {
 860                 r = proc_private->regions[i];
 861                 if (r != NULL) {
 862                         /* This is memzone */
 863                         if (i > 0 && (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)) {
 864                                 r->addr = NULL;
 865                                 if (r->fd > 0)
 866                                         close(r->fd);
 867                         }
 868                         if (r->addr != NULL) {
 869                                 munmap(r->addr, r->region_size);
 870                                 if (r->fd > 0) {
 871                                         close(r->fd);
 872                                         r->fd = -1;
 873                                 }
 874                         }
 875                         rte_free(r);
 876                         proc_private->regions[i] = NULL;
 877                 }
 878         }
 879         proc_private->regions_num = 0;
 880 }
 881
 882 static int
 883 memif_region_init_zc(const struct rte_memseg_list *msl, const struct rte_memseg *ms,
 884                      void *arg)
 885 {
 886         struct pmd_process_private *proc_private = (struct pmd_process_private *)arg;
 887         struct memif_region *r;
 888
 889         if (proc_private->regions_num < 1) {
 890                 MIF_LOG(ERR, "Missing descriptor region");
 891                 return -1;
 892         }
 893
 894         r = proc_private->regions[proc_private->regions_num - 1];
 895
 896         if (r->addr != msl->base_va)
 897                 r = proc_private->regions[++proc_private->regions_num - 1];
 898
 899         if (r == NULL) {
 900                 r = rte_zmalloc("region", sizeof(struct memif_region), 0);
 901                 if (r == NULL) {
 902                         MIF_LOG(ERR, "Failed to alloc memif region.");
 903                         return -ENOMEM;
 904                 }
 905
 906                 r->addr = msl->base_va;
 907                 r->region_size = ms->len;
 908                 r->fd = rte_memseg_get_fd(ms);
 909                 if (r->fd < 0)
 910                         return -1;
 911                 r->pkt_buffer_offset = 0;
 912
 913                 proc_private->regions[proc_private->regions_num - 1] = r;
 914         } else {
 915                 r->region_size += ms->len;
 916         }
 917
 918         return 0;
 919 }
 920
 921 static int
 922 memif_region_init_shm(struct rte_eth_dev *dev, uint8_t has_buffers)
 923 {
 924         struct pmd_internals *pmd = dev->data->dev_private;
 925         struct pmd_process_private *proc_private = dev->process_private;
 926         char shm_name[ETH_MEMIF_SHM_NAME_SIZE];
 927         int ret = 0;
 928         struct memif_region *r;
 929
 930         if (proc_private->regions_num >= ETH_MEMIF_MAX_REGION_NUM) {
 931                 MIF_LOG(ERR, "Too many regions.");
 932                 return -1;
 933         }
 934
 935         r = rte_zmalloc("region", sizeof(struct memif_region), 0);
 936         if (r == NULL) {
 937                 MIF_LOG(ERR, "Failed to alloc memif region.");
 938                 return -ENOMEM;
 939         }
 940
 941         /* calculate buffer offset */
 942         r->pkt_buffer_offset = (pmd->run.num_c2s_rings + pmd->run.num_s2c_rings) *
 943             (sizeof(memif_ring_t) + sizeof(memif_desc_t) *
 944             (1 << pmd->run.log2_ring_size));
 945
 946         r->region_size = r->pkt_buffer_offset;
 947         /* if region has buffers, add buffers size to region_size */
 948         if (has_buffers == 1)
 949                 r->region_size += (uint32_t)(pmd->run.pkt_buffer_size *
 950                         (1 << pmd->run.log2_ring_size) *
 951                         (pmd->run.num_c2s_rings +
 952                          pmd->run.num_s2c_rings));
 953
 954         memset(shm_name, 0, sizeof(char) * ETH_MEMIF_SHM_NAME_SIZE);
 955         snprintf(shm_name, ETH_MEMIF_SHM_NAME_SIZE, "memif_region_%d",
 956                  proc_private->regions_num);
 957
 958         r->fd = memfd_create(shm_name, MFD_ALLOW_SEALING);
 959         if (r->fd < 0) {
 960                 MIF_LOG(ERR, "Failed to create shm file: %s.", strerror(errno));
 961                 ret = -1;
 962                 goto error;
 963         }
 964
 965         ret = fcntl(r->fd, F_ADD_SEALS, F_SEAL_SHRINK);
 966         if (ret < 0) {
 967                 MIF_LOG(ERR, "Failed to add seals to shm file: %s.", strerror(errno));
 968                 goto error;
 969         }
 970
 971         ret = ftruncate(r->fd, r->region_size);
 972         if (ret < 0) {
 973                 MIF_LOG(ERR, "Failed to truncate shm file: %s.", strerror(errno));
 974                 goto error;
 975         }
 976
 977         r->addr = mmap(NULL, r->region_size, PROT_READ |
 978                        PROT_WRITE, MAP_SHARED, r->fd, 0);
 979         if (r->addr == MAP_FAILED) {
 980                 MIF_LOG(ERR, "Failed to mmap shm region: %s.", strerror(ret));
 981                 ret = -1;
 982                 goto error;
 983         }
 984
 985         proc_private->regions[proc_private->regions_num] = r;
 986         proc_private->regions_num++;
 987
 988         return ret;
 989
 990 error:
 991         if (r->fd > 0)
 992                 close(r->fd);
 993         r->fd = -1;
 994
 995         return ret;
 996 }
 997
 998 static int
 999 memif_regions_init(struct rte_eth_dev *dev)
1000 {
1001         struct pmd_internals *pmd = dev->data->dev_private;
1002         int ret;
1003
1004         /*
1005          * Zero-copy exposes dpdk memory.
1006          * Each memseg list will be represented by memif region.
1007          * Zero-copy regions indexing: memseg list idx + 1,
1008          * as we already have region 0 reserved for descriptors.
1009          */
1010         if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1011                 /* create region idx 0 containing descriptors */
1012                 ret = memif_region_init_shm(dev, 0);
1013                 if (ret < 0)
1014                         return ret;
1015                 ret = rte_memseg_walk(memif_region_init_zc, (void *)dev->process_private);
1016                 if (ret < 0)
1017                         return ret;
1018         } else {
1019                 /* create one memory region contaning rings and buffers */
1020                 ret = memif_region_init_shm(dev, /* has buffers */ 1);
1021                 if (ret < 0)
1022                         return ret;
1023         }
1024
1025         return 0;
1026 }
1027
1028 static void
1029 memif_init_rings(struct rte_eth_dev *dev)
1030 {
1031         struct pmd_internals *pmd = dev->data->dev_private;
1032         struct pmd_process_private *proc_private = dev->process_private;
1033         memif_ring_t *ring;
1034         int i, j;
1035         uint16_t slot;
1036
1037         for (i = 0; i < pmd->run.num_c2s_rings; i++) {
1038                 ring = memif_get_ring(pmd, proc_private, MEMIF_RING_C2S, i);
1039                 __atomic_store_n(&ring->head, 0, __ATOMIC_RELAXED);
1040                 __atomic_store_n(&ring->tail, 0, __ATOMIC_RELAXED);
1041                 ring->cookie = MEMIF_COOKIE;
1042                 ring->flags = 0;
1043
1044                 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)
1045                         continue;
1046
1047                 for (j = 0; j < (1 << pmd->run.log2_ring_size); j++) {
1048                         slot = i * (1 << pmd->run.log2_ring_size) + j;
1049                         ring->desc[j].region = 0;
1050                         ring->desc[j].offset =
1051                                 proc_private->regions[0]->pkt_buffer_offset +
1052                                 (uint32_t)(slot * pmd->run.pkt_buffer_size);
1053                         ring->desc[j].length = pmd->run.pkt_buffer_size;
1054                 }
1055         }
1056
1057         for (i = 0; i < pmd->run.num_s2c_rings; i++) {
1058                 ring = memif_get_ring(pmd, proc_private, MEMIF_RING_S2C, i);
1059                 __atomic_store_n(&ring->head, 0, __ATOMIC_RELAXED);
1060                 __atomic_store_n(&ring->tail, 0, __ATOMIC_RELAXED);
1061                 ring->cookie = MEMIF_COOKIE;
1062                 ring->flags = 0;
1063
1064                 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)
1065                         continue;
1066
1067                 for (j = 0; j < (1 << pmd->run.log2_ring_size); j++) {
1068                         slot = (i + pmd->run.num_c2s_rings) *
1069                             (1 << pmd->run.log2_ring_size) + j;
1070                         ring->desc[j].region = 0;
1071                         ring->desc[j].offset =
1072                                 proc_private->regions[0]->pkt_buffer_offset +
1073                                 (uint32_t)(slot * pmd->run.pkt_buffer_size);
1074                         ring->desc[j].length = pmd->run.pkt_buffer_size;
1075                 }
1076         }
1077 }
1078
1079 /* called only by client */
1080 static int
1081 memif_init_queues(struct rte_eth_dev *dev)
1082 {
1083         struct pmd_internals *pmd = dev->data->dev_private;
1084         struct memif_queue *mq;
1085         int i;
1086
1087         for (i = 0; i < pmd->run.num_c2s_rings; i++) {
1088                 mq = dev->data->tx_queues[i];
1089                 mq->log2_ring_size = pmd->run.log2_ring_size;
1090                 /* queues located only in region 0 */
1091                 mq->region = 0;
1092                 mq->ring_offset = memif_get_ring_offset(dev, mq, MEMIF_RING_C2S, i);
1093                 mq->last_head = 0;
1094                 mq->last_tail = 0;
1095                 mq->intr_handle.fd = eventfd(0, EFD_NONBLOCK);
1096                 if (mq->intr_handle.fd < 0) {
1097                         MIF_LOG(WARNING,
1098                                 "Failed to create eventfd for tx queue %d: %s.", i,
1099                                 strerror(errno));
1100                 }
1101                 mq->buffers = NULL;
1102                 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1103                         mq->buffers = rte_zmalloc("bufs", sizeof(struct rte_mbuf *) *
1104                                                   (1 << mq->log2_ring_size), 0);
1105                         if (mq->buffers == NULL)
1106                                 return -ENOMEM;
1107                 }
1108         }
1109
1110         for (i = 0; i < pmd->run.num_s2c_rings; i++) {
1111                 mq = dev->data->rx_queues[i];
1112                 mq->log2_ring_size = pmd->run.log2_ring_size;
1113                 /* queues located only in region 0 */
1114                 mq->region = 0;
1115                 mq->ring_offset = memif_get_ring_offset(dev, mq, MEMIF_RING_S2C, i);
1116                 mq->last_head = 0;
1117                 mq->last_tail = 0;
1118                 mq->intr_handle.fd = eventfd(0, EFD_NONBLOCK);
1119                 if (mq->intr_handle.fd < 0) {
1120                         MIF_LOG(WARNING,
1121                                 "Failed to create eventfd for rx queue %d: %s.", i,
1122                                 strerror(errno));
1123                 }
1124                 mq->buffers = NULL;
1125                 if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1126                         mq->buffers = rte_zmalloc("bufs", sizeof(struct rte_mbuf *) *
1127                                                   (1 << mq->log2_ring_size), 0);
1128                         if (mq->buffers == NULL)
1129                                 return -ENOMEM;
1130                 }
1131         }
1132         return 0;
1133 }
1134
1135 int
1136 memif_init_regions_and_queues(struct rte_eth_dev *dev)
1137 {
1138         int ret;
1139
1140         ret = memif_regions_init(dev);
1141         if (ret < 0)
1142                 return ret;
1143
1144         memif_init_rings(dev);
1145
1146         ret = memif_init_queues(dev);
1147         if (ret < 0)
1148                 return ret;
1149
1150         return 0;
1151 }
1152
1153 int
1154 memif_connect(struct rte_eth_dev *dev)
1155 {
1156         struct pmd_internals *pmd = dev->data->dev_private;
1157         struct pmd_process_private *proc_private = dev->process_private;
1158         struct memif_region *mr;
1159         struct memif_queue *mq;
1160         memif_ring_t *ring;
1161         int i;
1162
1163         for (i = 0; i < proc_private->regions_num; i++) {
1164                 mr = proc_private->regions[i];
1165                 if (mr != NULL) {
1166                         if (mr->addr == NULL) {
1167                                 if (mr->fd < 0)
1168                                         return -1;
1169                                 mr->addr = mmap(NULL, mr->region_size,
1170                                                 PROT_READ | PROT_WRITE,
1171                                                 MAP_SHARED, mr->fd, 0);
1172                                 if (mr->addr == MAP_FAILED) {
1173                                         MIF_LOG(ERR, "mmap failed: %s\n",
1174                                                 strerror(errno));
1175                                         return -1;
1176                                 }
1177                         }
1178                         if (i > 0 && (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY)) {
1179                                 /* close memseg file */
1180                                 close(mr->fd);
1181                                 mr->fd = -1;
1182                         }
1183                 }
1184         }
1185
1186         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1187                 for (i = 0; i < pmd->run.num_c2s_rings; i++) {
1188                         mq = (pmd->role == MEMIF_ROLE_CLIENT) ?
1189                             dev->data->tx_queues[i] : dev->data->rx_queues[i];
1190                         ring = memif_get_ring_from_queue(proc_private, mq);
1191                         if (ring == NULL || ring->cookie != MEMIF_COOKIE) {
1192                                 MIF_LOG(ERR, "Wrong ring");
1193                                 return -1;
1194                         }
1195                         __atomic_store_n(&ring->head, 0, __ATOMIC_RELAXED);
1196                         __atomic_store_n(&ring->tail, 0, __ATOMIC_RELAXED);
1197                         mq->last_head = 0;
1198                         mq->last_tail = 0;
1199                         /* enable polling mode */
1200                         if (pmd->role == MEMIF_ROLE_SERVER)
1201                                 ring->flags = MEMIF_RING_FLAG_MASK_INT;
1202                 }
1203                 for (i = 0; i < pmd->run.num_s2c_rings; i++) {
1204                         mq = (pmd->role == MEMIF_ROLE_CLIENT) ?
1205                             dev->data->rx_queues[i] : dev->data->tx_queues[i];
1206                         ring = memif_get_ring_from_queue(proc_private, mq);
1207                         if (ring == NULL || ring->cookie != MEMIF_COOKIE) {
1208                                 MIF_LOG(ERR, "Wrong ring");
1209                                 return -1;
1210                         }
1211                         __atomic_store_n(&ring->head, 0, __ATOMIC_RELAXED);
1212                         __atomic_store_n(&ring->tail, 0, __ATOMIC_RELAXED);
1213                         mq->last_head = 0;
1214                         mq->last_tail = 0;
1215                         /* enable polling mode */
1216                         if (pmd->role == MEMIF_ROLE_CLIENT)
1217                                 ring->flags = MEMIF_RING_FLAG_MASK_INT;
1218                 }
1219
1220                 pmd->flags &= ~ETH_MEMIF_FLAG_CONNECTING;
1221                 pmd->flags |= ETH_MEMIF_FLAG_CONNECTED;
1222                 dev->data->dev_link.link_status = ETH_LINK_UP;
1223         }
1224         MIF_LOG(INFO, "Connected.");
1225         return 0;
1226 }
1227
1228 static int
1229 memif_dev_start(struct rte_eth_dev *dev)
1230 {
1231         struct pmd_internals *pmd = dev->data->dev_private;
1232         int ret = 0;
1233
1234         switch (pmd->role) {
1235         case MEMIF_ROLE_CLIENT:
1236                 ret = memif_connect_client(dev);
1237                 break;
1238         case MEMIF_ROLE_SERVER:
1239                 ret = memif_connect_server(dev);
1240                 break;
1241         default:
1242                 MIF_LOG(ERR, "Unknown role: %d.", pmd->role);
1243                 ret = -1;
1244                 break;
1245         }
1246
1247         return ret;
1248 }
1249
1250 static int
1251 memif_dev_close(struct rte_eth_dev *dev)
1252 {
1253         struct pmd_internals *pmd = dev->data->dev_private;
1254         int i;
1255
1256         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1257                 memif_msg_enq_disconnect(pmd->cc, "Device closed", 0);
1258                 memif_disconnect(dev);
1259
1260                 for (i = 0; i < dev->data->nb_rx_queues; i++)
1261                         (*dev->dev_ops->rx_queue_release)(dev->data->rx_queues[i]);
1262                 for (i = 0; i < dev->data->nb_tx_queues; i++)
1263                         (*dev->dev_ops->tx_queue_release)(dev->data->tx_queues[i]);
1264
1265                 memif_socket_remove_device(dev);
1266         } else {
1267                 memif_disconnect(dev);
1268         }
1269
1270         rte_free(dev->process_private);
1271
1272         return 0;
1273 }
1274
1275 static int
1276 memif_dev_configure(struct rte_eth_dev *dev)
1277 {
1278         struct pmd_internals *pmd = dev->data->dev_private;
1279
1280         /*
1281          * CLIENT - TXQ
1282          * SERVER - RXQ
1283          */
1284         pmd->cfg.num_c2s_rings = (pmd->role == MEMIF_ROLE_CLIENT) ?
1285                                   dev->data->nb_tx_queues : dev->data->nb_rx_queues;
1286
1287         /*
1288          * CLIENT - RXQ
1289          * SERVER - TXQ
1290          */
1291         pmd->cfg.num_s2c_rings = (pmd->role == MEMIF_ROLE_CLIENT) ?
1292                                   dev->data->nb_rx_queues : dev->data->nb_tx_queues;
1293
1294         return 0;
1295 }
1296
1297 static int
1298 memif_tx_queue_setup(struct rte_eth_dev *dev,
1299                      uint16_t qid,
1300                      uint16_t nb_tx_desc __rte_unused,
1301                      unsigned int socket_id __rte_unused,
1302                      const struct rte_eth_txconf *tx_conf __rte_unused)
1303 {
1304         struct pmd_internals *pmd = dev->data->dev_private;
1305         struct memif_queue *mq;
1306
1307         mq = rte_zmalloc("tx-queue", sizeof(struct memif_queue), 0);
1308         if (mq == NULL) {
1309                 MIF_LOG(ERR, "Failed to allocate tx queue id: %u", qid);
1310                 return -ENOMEM;
1311         }
1312
1313         mq->type =
1314             (pmd->role == MEMIF_ROLE_CLIENT) ? MEMIF_RING_C2S : MEMIF_RING_S2C;
1315         mq->n_pkts = 0;
1316         mq->n_bytes = 0;
1317         mq->intr_handle.fd = -1;
1318         mq->intr_handle.type = RTE_INTR_HANDLE_EXT;
1319         mq->in_port = dev->data->port_id;
1320         dev->data->tx_queues[qid] = mq;
1321
1322         return 0;
1323 }
1324
1325 static int
1326 memif_rx_queue_setup(struct rte_eth_dev *dev,
1327                      uint16_t qid,
1328                      uint16_t nb_rx_desc __rte_unused,
1329                      unsigned int socket_id __rte_unused,
1330                      const struct rte_eth_rxconf *rx_conf __rte_unused,
1331                      struct rte_mempool *mb_pool)
1332 {
1333         struct pmd_internals *pmd = dev->data->dev_private;
1334         struct memif_queue *mq;
1335
1336         mq = rte_zmalloc("rx-queue", sizeof(struct memif_queue), 0);
1337         if (mq == NULL) {
1338                 MIF_LOG(ERR, "Failed to allocate rx queue id: %u", qid);
1339                 return -ENOMEM;
1340         }
1341
1342         mq->type = (pmd->role == MEMIF_ROLE_CLIENT) ? MEMIF_RING_S2C : MEMIF_RING_C2S;
1343         mq->n_pkts = 0;
1344         mq->n_bytes = 0;
1345         mq->intr_handle.fd = -1;
1346         mq->intr_handle.type = RTE_INTR_HANDLE_EXT;
1347         mq->mempool = mb_pool;
1348         mq->in_port = dev->data->port_id;
1349         dev->data->rx_queues[qid] = mq;
1350
1351         return 0;
1352 }
1353
1354 static void
1355 memif_queue_release(void *queue)
1356 {
1357         struct memif_queue *mq = (struct memif_queue *)queue;
1358
1359         if (!mq)
1360                 return;
1361
1362         rte_free(mq);
1363 }
1364
1365 static int
1366 memif_link_update(struct rte_eth_dev *dev,
1367                   int wait_to_complete __rte_unused)
1368 {
1369         struct pmd_process_private *proc_private;
1370
1371         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1372                 proc_private = dev->process_private;
1373                 if (dev->data->dev_link.link_status == ETH_LINK_UP &&
1374                                 proc_private->regions_num == 0) {
1375                         memif_mp_request_regions(dev);
1376                 } else if (dev->data->dev_link.link_status == ETH_LINK_DOWN &&
1377                                 proc_private->regions_num > 0) {
1378                         memif_free_regions(dev);
1379                 }
1380         }
1381         return 0;
1382 }
1383
1384 static int
1385 memif_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
1386 {
1387         struct pmd_internals *pmd = dev->data->dev_private;
1388         struct memif_queue *mq;
1389         int i;
1390         uint8_t tmp, nq;
1391
1392         stats->ipackets = 0;
1393         stats->ibytes = 0;
1394         stats->opackets = 0;
1395         stats->obytes = 0;
1396
1397         tmp = (pmd->role == MEMIF_ROLE_CLIENT) ? pmd->run.num_c2s_rings :
1398             pmd->run.num_s2c_rings;
1399         nq = (tmp < RTE_ETHDEV_QUEUE_STAT_CNTRS) ? tmp :
1400             RTE_ETHDEV_QUEUE_STAT_CNTRS;
1401
1402         /* RX stats */
1403         for (i = 0; i < nq; i++) {
1404                 mq = dev->data->rx_queues[i];
1405                 stats->q_ipackets[i] = mq->n_pkts;
1406                 stats->q_ibytes[i] = mq->n_bytes;
1407                 stats->ipackets += mq->n_pkts;
1408                 stats->ibytes += mq->n_bytes;
1409         }
1410
1411         tmp = (pmd->role == MEMIF_ROLE_CLIENT) ? pmd->run.num_s2c_rings :
1412             pmd->run.num_c2s_rings;
1413         nq = (tmp < RTE_ETHDEV_QUEUE_STAT_CNTRS) ? tmp :
1414             RTE_ETHDEV_QUEUE_STAT_CNTRS;
1415
1416         /* TX stats */
1417         for (i = 0; i < nq; i++) {
1418                 mq = dev->data->tx_queues[i];
1419                 stats->q_opackets[i] = mq->n_pkts;
1420                 stats->q_obytes[i] = mq->n_bytes;
1421                 stats->opackets += mq->n_pkts;
1422                 stats->obytes += mq->n_bytes;
1423         }
1424         return 0;
1425 }
1426
1427 static int
1428 memif_stats_reset(struct rte_eth_dev *dev)
1429 {
1430         struct pmd_internals *pmd = dev->data->dev_private;
1431         int i;
1432         struct memif_queue *mq;
1433
1434         for (i = 0; i < pmd->run.num_c2s_rings; i++) {
1435                 mq = (pmd->role == MEMIF_ROLE_CLIENT) ? dev->data->tx_queues[i] :
1436                     dev->data->rx_queues[i];
1437                 mq->n_pkts = 0;
1438                 mq->n_bytes = 0;
1439         }
1440         for (i = 0; i < pmd->run.num_s2c_rings; i++) {
1441                 mq = (pmd->role == MEMIF_ROLE_CLIENT) ? dev->data->rx_queues[i] :
1442                     dev->data->tx_queues[i];
1443                 mq->n_pkts = 0;
1444                 mq->n_bytes = 0;
1445         }
1446
1447         return 0;
1448 }
1449
1450 static int
1451 memif_rx_queue_intr_enable(struct rte_eth_dev *dev __rte_unused,
1452                            uint16_t qid __rte_unused)
1453 {
1454         MIF_LOG(WARNING, "Interrupt mode not supported.");
1455
1456         return -1;
1457 }
1458
1459 static int
1460 memif_rx_queue_intr_disable(struct rte_eth_dev *dev, uint16_t qid __rte_unused)
1461 {
1462         struct pmd_internals *pmd __rte_unused = dev->data->dev_private;
1463
1464         return 0;
1465 }
1466
1467 static const struct eth_dev_ops ops = {
1468         .dev_start = memif_dev_start,
1469         .dev_close = memif_dev_close,
1470         .dev_infos_get = memif_dev_info,
1471         .dev_configure = memif_dev_configure,
1472         .tx_queue_setup = memif_tx_queue_setup,
1473         .rx_queue_setup = memif_rx_queue_setup,
1474         .rx_queue_release = memif_queue_release,
1475         .tx_queue_release = memif_queue_release,
1476         .rx_queue_intr_enable = memif_rx_queue_intr_enable,
1477         .rx_queue_intr_disable = memif_rx_queue_intr_disable,
1478         .link_update = memif_link_update,
1479         .stats_get = memif_stats_get,
1480         .stats_reset = memif_stats_reset,
1481 };
1482
1483 static int
1484 memif_create(struct rte_vdev_device *vdev, enum memif_role_t role,
1485              memif_interface_id_t id, uint32_t flags,
1486              const char *socket_filename,
1487              memif_log2_ring_size_t log2_ring_size,
1488              uint16_t pkt_buffer_size, const char *secret,
1489              struct rte_ether_addr *ether_addr)
1490 {
1491         int ret = 0;
1492         struct rte_eth_dev *eth_dev;
1493         struct rte_eth_dev_data *data;
1494         struct pmd_internals *pmd;
1495         struct pmd_process_private *process_private;
1496         const unsigned int numa_node = vdev->device.numa_node;
1497         const char *name = rte_vdev_device_name(vdev);
1498
1499         eth_dev = rte_eth_vdev_allocate(vdev, sizeof(*pmd));
1500         if (eth_dev == NULL) {
1501                 MIF_LOG(ERR, "%s: Unable to allocate device struct.", name);
1502                 return -1;
1503         }
1504
1505         process_private = (struct pmd_process_private *)
1506                 rte_zmalloc(name, sizeof(struct pmd_process_private),
1507                             RTE_CACHE_LINE_SIZE);
1508
1509         if (process_private == NULL) {
1510                 MIF_LOG(ERR, "Failed to alloc memory for process private");
1511                 return -1;
1512         }
1513         eth_dev->process_private = process_private;
1514
1515         pmd = eth_dev->data->dev_private;
1516         memset(pmd, 0, sizeof(*pmd));
1517
1518         pmd->id = id;
1519         pmd->flags = flags;
1520         pmd->flags |= ETH_MEMIF_FLAG_DISABLED;
1521         pmd->role = role;
1522         /* Zero-copy flag irelevant to server. */
1523         if (pmd->role == MEMIF_ROLE_SERVER)
1524                 pmd->flags &= ~ETH_MEMIF_FLAG_ZERO_COPY;
1525
1526         ret = memif_socket_init(eth_dev, socket_filename);
1527         if (ret < 0)
1528                 return ret;
1529
1530         memset(pmd->secret, 0, sizeof(char) * ETH_MEMIF_SECRET_SIZE);
1531         if (secret != NULL)
1532                 strlcpy(pmd->secret, secret, sizeof(pmd->secret));
1533
1534         pmd->cfg.log2_ring_size = log2_ring_size;
1535         /* set in .dev_configure() */
1536         pmd->cfg.num_c2s_rings = 0;
1537         pmd->cfg.num_s2c_rings = 0;
1538
1539         pmd->cfg.pkt_buffer_size = pkt_buffer_size;
1540         rte_spinlock_init(&pmd->cc_lock);
1541
1542         data = eth_dev->data;
1543         data->dev_private = pmd;
1544         data->numa_node = numa_node;
1545         data->dev_link = pmd_link;
1546         data->mac_addrs = ether_addr;
1547         data->promiscuous = 1;
1548         data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1549
1550         eth_dev->dev_ops = &ops;
1551         eth_dev->device = &vdev->device;
1552         if (pmd->flags & ETH_MEMIF_FLAG_ZERO_COPY) {
1553                 eth_dev->rx_pkt_burst = eth_memif_rx_zc;
1554                 eth_dev->tx_pkt_burst = eth_memif_tx_zc;
1555         } else {
1556                 eth_dev->rx_pkt_burst = eth_memif_rx;
1557                 eth_dev->tx_pkt_burst = eth_memif_tx;
1558         }
1559
1560         rte_eth_dev_probing_finish(eth_dev);
1561
1562         return 0;
1563 }
1564
1565 static int
1566 memif_set_role(const char *key __rte_unused, const char *value,
1567                void *extra_args)
1568 {
1569         enum memif_role_t *role = (enum memif_role_t *)extra_args;
1570
1571         if (strstr(value, "server") != NULL) {
1572                 *role = MEMIF_ROLE_SERVER;
1573         } else if (strstr(value, "client") != NULL) {
1574                 *role = MEMIF_ROLE_CLIENT;
1575         } else if (strstr(value, "master") != NULL) {
1576                 MIF_LOG(NOTICE, "Role argument \"master\" is deprecated, use \"server\"");
1577                 *role = MEMIF_ROLE_SERVER;
1578         } else if (strstr(value, "slave") != NULL) {
1579                 MIF_LOG(NOTICE, "Role argument \"slave\" is deprecated, use \"client\"");
1580                 *role = MEMIF_ROLE_CLIENT;
1581         } else {
1582                 MIF_LOG(ERR, "Unknown role: %s.", value);
1583                 return -EINVAL;
1584         }
1585         return 0;
1586 }
1587
1588 static int
1589 memif_set_zc(const char *key __rte_unused, const char *value, void *extra_args)
1590 {
1591         uint32_t *flags = (uint32_t *)extra_args;
1592
1593         if (strstr(value, "yes") != NULL) {
1594                 if (!rte_mcfg_get_single_file_segments()) {
1595                         MIF_LOG(ERR, "Zero-copy doesn't support multi-file segments.");
1596                         return -ENOTSUP;
1597                 }
1598                 *flags |= ETH_MEMIF_FLAG_ZERO_COPY;
1599         } else if (strstr(value, "no") != NULL) {
1600                 *flags &= ~ETH_MEMIF_FLAG_ZERO_COPY;
1601         } else {
1602                 MIF_LOG(ERR, "Failed to parse zero-copy param: %s.", value);
1603                 return -EINVAL;
1604         }
1605         return 0;
1606 }
1607
1608 static int
1609 memif_set_id(const char *key __rte_unused, const char *value, void *extra_args)
1610 {
1611         memif_interface_id_t *id = (memif_interface_id_t *)extra_args;
1612
1613         /* even if parsing fails, 0 is a valid id */
1614         *id = strtoul(value, NULL, 10);
1615         return 0;
1616 }
1617
1618 static int
1619 memif_set_bs(const char *key __rte_unused, const char *value, void *extra_args)
1620 {
1621         unsigned long tmp;
1622         uint16_t *pkt_buffer_size = (uint16_t *)extra_args;
1623
1624         tmp = strtoul(value, NULL, 10);
1625         if (tmp == 0 || tmp > 0xFFFF) {
1626                 MIF_LOG(ERR, "Invalid buffer size: %s.", value);
1627                 return -EINVAL;
1628         }
1629         *pkt_buffer_size = tmp;
1630         return 0;
1631 }
1632
1633 static int
1634 memif_set_rs(const char *key __rte_unused, const char *value, void *extra_args)
1635 {
1636         unsigned long tmp;
1637         memif_log2_ring_size_t *log2_ring_size =
1638             (memif_log2_ring_size_t *)extra_args;
1639
1640         tmp = strtoul(value, NULL, 10);
1641         if (tmp == 0 || tmp > ETH_MEMIF_MAX_LOG2_RING_SIZE) {
1642                 MIF_LOG(ERR, "Invalid ring size: %s (max %u).",
1643                         value, ETH_MEMIF_MAX_LOG2_RING_SIZE);
1644                 return -EINVAL;
1645         }
1646         *log2_ring_size = tmp;
1647         return 0;
1648 }
1649
1650 /* check if directory exists and if we have permission to read/write */
1651 static int
1652 memif_check_socket_filename(const char *filename)
1653 {
1654         char *dir = NULL, *tmp;
1655         uint32_t idx;
1656         int ret = 0;
1657
1658         if (strlen(filename) >= MEMIF_SOCKET_UN_SIZE) {
1659                 MIF_LOG(ERR, "Unix socket address too long (max 108).");
1660                 return -1;
1661         }
1662
1663         tmp = strrchr(filename, '/');
1664         if (tmp != NULL) {
1665                 idx = tmp - filename;
1666                 dir = rte_zmalloc("memif_tmp", sizeof(char) * (idx + 1), 0);
1667                 if (dir == NULL) {
1668                         MIF_LOG(ERR, "Failed to allocate memory.");
1669                         return -1;
1670                 }
1671                 strlcpy(dir, filename, sizeof(char) * (idx + 1));
1672         }
1673
1674         if (dir == NULL || (faccessat(-1, dir, F_OK | R_OK |
1675                                         W_OK, AT_EACCESS) < 0)) {
1676                 MIF_LOG(ERR, "Invalid socket directory.");
1677                 ret = -EINVAL;
1678         }
1679
1680         if (dir != NULL)
1681                 rte_free(dir);
1682
1683         return ret;
1684 }
1685
1686 static int
1687 memif_set_socket_filename(const char *key __rte_unused, const char *value,
1688                           void *extra_args)
1689 {
1690         const char **socket_filename = (const char **)extra_args;
1691
1692         *socket_filename = value;
1693         return 0;
1694 }
1695
1696 static int
1697 memif_set_is_socket_abstract(const char *key __rte_unused, const char *value, void *extra_args)
1698 {
1699         uint32_t *flags = (uint32_t *)extra_args;
1700
1701         if (strstr(value, "yes") != NULL) {
1702                 *flags |= ETH_MEMIF_FLAG_SOCKET_ABSTRACT;
1703         } else if (strstr(value, "no") != NULL) {
1704                 *flags &= ~ETH_MEMIF_FLAG_SOCKET_ABSTRACT;
1705         } else {
1706                 MIF_LOG(ERR, "Failed to parse socket-abstract param: %s.", value);
1707                 return -EINVAL;
1708         }
1709         return 0;
1710 }
1711
1712 static int
1713 memif_set_mac(const char *key __rte_unused, const char *value, void *extra_args)
1714 {
1715         struct rte_ether_addr *ether_addr = (struct rte_ether_addr *)extra_args;
1716
1717         if (rte_ether_unformat_addr(value, ether_addr) < 0)
1718                 MIF_LOG(WARNING, "Failed to parse mac '%s'.", value);
1719         return 0;
1720 }
1721
1722 static int
1723 memif_set_secret(const char *key __rte_unused, const char *value, void *extra_args)
1724 {
1725         const char **secret = (const char **)extra_args;
1726
1727         *secret = value;
1728         return 0;
1729 }
1730
1731 static int
1732 rte_pmd_memif_probe(struct rte_vdev_device *vdev)
1733 {
1734         RTE_BUILD_BUG_ON(sizeof(memif_msg_t) != 128);
1735         RTE_BUILD_BUG_ON(sizeof(memif_desc_t) != 16);
1736         int ret = 0;
1737         struct rte_kvargs *kvlist;
1738         const char *name = rte_vdev_device_name(vdev);
1739         enum memif_role_t role = MEMIF_ROLE_CLIENT;
1740         memif_interface_id_t id = 0;
1741         uint16_t pkt_buffer_size = ETH_MEMIF_DEFAULT_PKT_BUFFER_SIZE;
1742         memif_log2_ring_size_t log2_ring_size = ETH_MEMIF_DEFAULT_RING_SIZE;
1743         const char *socket_filename = ETH_MEMIF_DEFAULT_SOCKET_FILENAME;
1744         uint32_t flags = 0;
1745         const char *secret = NULL;
1746         struct rte_ether_addr *ether_addr = rte_zmalloc("",
1747                 sizeof(struct rte_ether_addr), 0);
1748         struct rte_eth_dev *eth_dev;
1749
1750         rte_eth_random_addr(ether_addr->addr_bytes);
1751
1752         MIF_LOG(INFO, "Initialize MEMIF: %s.", name);
1753
1754         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1755                 eth_dev = rte_eth_dev_attach_secondary(name);
1756                 if (!eth_dev) {
1757                         MIF_LOG(ERR, "Failed to probe %s", name);
1758                         return -1;
1759                 }
1760
1761                 eth_dev->dev_ops = &ops;
1762                 eth_dev->device = &vdev->device;
1763                 eth_dev->rx_pkt_burst = eth_memif_rx;
1764                 eth_dev->tx_pkt_burst = eth_memif_tx;
1765
1766                 if (!rte_eal_primary_proc_alive(NULL)) {
1767                         MIF_LOG(ERR, "Primary process is missing");
1768                         return -1;
1769                 }
1770
1771                 eth_dev->process_private = (struct pmd_process_private *)
1772                         rte_zmalloc(name,
1773                                 sizeof(struct pmd_process_private),
1774                                 RTE_CACHE_LINE_SIZE);
1775                 if (eth_dev->process_private == NULL) {
1776                         MIF_LOG(ERR,
1777                                 "Failed to alloc memory for process private");
1778                         return -1;
1779                 }
1780
1781                 rte_eth_dev_probing_finish(eth_dev);
1782
1783                 return 0;
1784         }
1785
1786         ret = rte_mp_action_register(MEMIF_MP_SEND_REGION, memif_mp_send_region);
1787         /*
1788          * Primary process can continue probing, but secondary process won't
1789          * be able to get memory regions information
1790          */
1791         if (ret < 0 && rte_errno != EEXIST)
1792                 MIF_LOG(WARNING, "Failed to register mp action callback: %s",
1793                         strerror(rte_errno));
1794
1795         /* use abstract address by default */
1796         flags |= ETH_MEMIF_FLAG_SOCKET_ABSTRACT;
1797
1798         kvlist = rte_kvargs_parse(rte_vdev_device_args(vdev), valid_arguments);
1799
1800         /* parse parameters */
1801         if (kvlist != NULL) {
1802                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_ROLE_ARG,
1803                                          &memif_set_role, &role);
1804                 if (ret < 0)
1805                         goto exit;
1806                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_ID_ARG,
1807                                          &memif_set_id, &id);
1808                 if (ret < 0)
1809                         goto exit;
1810                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_PKT_BUFFER_SIZE_ARG,
1811                                          &memif_set_bs, &pkt_buffer_size);
1812                 if (ret < 0)
1813                         goto exit;
1814                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_RING_SIZE_ARG,
1815                                          &memif_set_rs, &log2_ring_size);
1816                 if (ret < 0)
1817                         goto exit;
1818                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_SOCKET_ARG,
1819                                          &memif_set_socket_filename,
1820                                          (void *)(&socket_filename));
1821                 if (ret < 0)
1822                         goto exit;
1823                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_SOCKET_ABSTRACT_ARG,
1824                                          &memif_set_is_socket_abstract, &flags);
1825                 if (ret < 0)
1826                         goto exit;
1827                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_MAC_ARG,
1828                                          &memif_set_mac, ether_addr);
1829                 if (ret < 0)
1830                         goto exit;
1831                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_ZC_ARG,
1832                                          &memif_set_zc, &flags);
1833                 if (ret < 0)
1834                         goto exit;
1835                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_SECRET_ARG,
1836                                          &memif_set_secret, (void *)(&secret));
1837                 if (ret < 0)
1838                         goto exit;
1839         }
1840
1841         if (!(flags & ETH_MEMIF_FLAG_SOCKET_ABSTRACT)) {
1842                 ret = memif_check_socket_filename(socket_filename);
1843                 if (ret < 0)
1844                         goto exit;
1845         }
1846
1847         /* create interface */
1848         ret = memif_create(vdev, role, id, flags, socket_filename,
1849                            log2_ring_size, pkt_buffer_size, secret, ether_addr);
1850
1851 exit:
1852         if (kvlist != NULL)
1853                 rte_kvargs_free(kvlist);
1854         return ret;
1855 }
1856
1857 static int
1858 rte_pmd_memif_remove(struct rte_vdev_device *vdev)
1859 {
1860         struct rte_eth_dev *eth_dev;
1861
1862         eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(vdev));
1863         if (eth_dev == NULL)
1864                 return 0;
1865
1866         return rte_eth_dev_close(eth_dev->data->port_id);
1867 }
1868
1869 static struct rte_vdev_driver pmd_memif_drv = {
1870         .probe = rte_pmd_memif_probe,
1871         .remove = rte_pmd_memif_remove,
1872 };
1873
1874 RTE_PMD_REGISTER_VDEV(net_memif, pmd_memif_drv);
1875
1876 RTE_PMD_REGISTER_PARAM_STRING(net_memif,
1877                               ETH_MEMIF_ID_ARG "=<int>"
1878                               ETH_MEMIF_ROLE_ARG "=server|client"
1879                               ETH_MEMIF_PKT_BUFFER_SIZE_ARG "=<int>"
1880                               ETH_MEMIF_RING_SIZE_ARG "=<int>"
1881                               ETH_MEMIF_SOCKET_ARG "=<string>"
1882                                   ETH_MEMIF_SOCKET_ABSTRACT_ARG "=yes|no"
1883                               ETH_MEMIF_MAC_ARG "=xx:xx:xx:xx:xx:xx"
1884                               ETH_MEMIF_ZC_ARG "=yes|no"
1885                               ETH_MEMIF_SECRET_ARG "=<string>");
1886
1887 RTE_LOG_REGISTER_DEFAULT(memif_logtype, NOTICE);