drivers/net/memif/rte_eth_memif.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2018-2019 Cisco Systems, Inc.  All rights reserved.
   3  */
   4
   5 #include <stdlib.h>
   6 #include <fcntl.h>
   7 #include <unistd.h>
   8 #include <sys/types.h>
   9 #include <sys/socket.h>
  10 #include <sys/un.h>
  11 #include <sys/ioctl.h>
  12 #include <sys/mman.h>
  13 #include <linux/if_ether.h>
  14 #include <errno.h>
  15 #include <sys/eventfd.h>
  16
  17 #include <rte_version.h>
  18 #include <rte_mbuf.h>
  19 #include <rte_ether.h>
  20 #include <rte_ethdev_driver.h>
  21 #include <rte_ethdev_vdev.h>
  22 #include <rte_malloc.h>
  23 #include <rte_kvargs.h>
  24 #include <rte_bus_vdev.h>
  25 #include <rte_string_fns.h>
  26
  27 #include "rte_eth_memif.h"
  28 #include "memif_socket.h"
  29
  30 #define ETH_MEMIF_ID_ARG                "id"
  31 #define ETH_MEMIF_ROLE_ARG              "role"
  32 #define ETH_MEMIF_PKT_BUFFER_SIZE_ARG   "bsize"
  33 #define ETH_MEMIF_RING_SIZE_ARG         "rsize"
  34 #define ETH_MEMIF_SOCKET_ARG            "socket"
  35 #define ETH_MEMIF_MAC_ARG               "mac"
  36 #define ETH_MEMIF_ZC_ARG                "zero-copy"
  37 #define ETH_MEMIF_SECRET_ARG            "secret"
  38
  39 static const char * const valid_arguments[] = {
  40         ETH_MEMIF_ID_ARG,
  41         ETH_MEMIF_ROLE_ARG,
  42         ETH_MEMIF_PKT_BUFFER_SIZE_ARG,
  43         ETH_MEMIF_RING_SIZE_ARG,
  44         ETH_MEMIF_SOCKET_ARG,
  45         ETH_MEMIF_MAC_ARG,
  46         ETH_MEMIF_ZC_ARG,
  47         ETH_MEMIF_SECRET_ARG,
  48         NULL
  49 };
  50
  51 const char *
  52 memif_version(void)
  53 {
  54         return ("memif-" RTE_STR(MEMIF_VERSION_MAJOR) "." RTE_STR(MEMIF_VERSION_MINOR));
  55 }
  56
  57 static void
  58 memif_dev_info(struct rte_eth_dev *dev __rte_unused, struct rte_eth_dev_info *dev_info)
  59 {
  60         dev_info->max_mac_addrs = 1;
  61         dev_info->max_rx_pktlen = (uint32_t)ETH_FRAME_LEN;
  62         dev_info->max_rx_queues = ETH_MEMIF_MAX_NUM_Q_PAIRS;
  63         dev_info->max_tx_queues = ETH_MEMIF_MAX_NUM_Q_PAIRS;
  64         dev_info->min_rx_bufsize = 0;
  65 }
  66
  67 static memif_ring_t *
  68 memif_get_ring(struct pmd_internals *pmd, memif_ring_type_t type, uint16_t ring_num)
  69 {
  70         /* rings only in region 0 */
  71         void *p = pmd->regions[0]->addr;
  72         int ring_size = sizeof(memif_ring_t) + sizeof(memif_desc_t) *
  73             (1 << pmd->run.log2_ring_size);
  74
  75         p = (uint8_t *)p + (ring_num + type * pmd->run.num_s2m_rings) * ring_size;
  76
  77         return (memif_ring_t *)p;
  78 }
  79
  80 static void *
  81 memif_get_buffer(struct pmd_internals *pmd, memif_desc_t *d)
  82 {
  83         return ((uint8_t *)pmd->regions[d->region]->addr + d->offset);
  84 }
  85
  86 static int
  87 memif_pktmbuf_chain(struct rte_mbuf *head, struct rte_mbuf *cur_tail,
  88                     struct rte_mbuf *tail)
  89 {
  90         /* Check for number-of-segments-overflow */
  91         if (unlikely(head->nb_segs + tail->nb_segs > RTE_MBUF_MAX_NB_SEGS))
  92                 return -EOVERFLOW;
  93
  94         /* Chain 'tail' onto the old tail */
  95         cur_tail->next = tail;
  96
  97         /* accumulate number of segments and total length. */
  98         head->nb_segs = (uint16_t)(head->nb_segs + tail->nb_segs);
  99
 100         tail->pkt_len = tail->data_len;
 101         head->pkt_len += tail->pkt_len;
 102
 103         return 0;
 104 }
 105
 106 static uint16_t
 107 eth_memif_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 108 {
 109         struct memif_queue *mq = queue;
 110         struct pmd_internals *pmd = mq->pmd;
 111         memif_ring_t *ring = mq->ring;
 112         uint16_t cur_slot, last_slot, n_slots, ring_size, mask, s0;
 113         uint16_t n_rx_pkts = 0;
 114         uint16_t mbuf_size = rte_pktmbuf_data_room_size(mq->mempool) -
 115                 RTE_PKTMBUF_HEADROOM;
 116         uint16_t src_len, src_off, dst_len, dst_off, cp_len;
 117         memif_ring_type_t type = mq->type;
 118         memif_desc_t *d0;
 119         struct rte_mbuf *mbuf, *mbuf_head, *mbuf_tail;
 120         uint64_t b;
 121         ssize_t size __rte_unused;
 122         uint16_t head;
 123         int ret;
 124
 125         if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
 126                 return 0;
 127         if (unlikely(ring == NULL))
 128                 return 0;
 129
 130         /* consume interrupt */
 131         if ((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0)
 132                 size = read(mq->intr_handle.fd, &b, sizeof(b));
 133
 134         ring_size = 1 << mq->log2_ring_size;
 135         mask = ring_size - 1;
 136
 137         cur_slot = (type == MEMIF_RING_S2M) ? mq->last_head : mq->last_tail;
 138         last_slot = (type == MEMIF_RING_S2M) ? ring->head : ring->tail;
 139         if (cur_slot == last_slot)
 140                 goto refill;
 141         n_slots = last_slot - cur_slot;
 142
 143         while (n_slots && n_rx_pkts < nb_pkts) {
 144                 mbuf_head = rte_pktmbuf_alloc(mq->mempool);
 145                 if (unlikely(mbuf_head == NULL))
 146                         goto no_free_bufs;
 147                 mbuf = mbuf_head;
 148                 mbuf->port = mq->in_port;
 149
 150 next_slot:
 151                 s0 = cur_slot & mask;
 152                 d0 = &ring->desc[s0];
 153
 154                 src_len = d0->length;
 155                 dst_off = 0;
 156                 src_off = 0;
 157
 158                 do {
 159                         dst_len = mbuf_size - dst_off;
 160                         if (dst_len == 0) {
 161                                 dst_off = 0;
 162                                 dst_len = mbuf_size;
 163
 164                                 /* store pointer to tail */
 165                                 mbuf_tail = mbuf;
 166                                 mbuf = rte_pktmbuf_alloc(mq->mempool);
 167                                 if (unlikely(mbuf == NULL))
 168                                         goto no_free_bufs;
 169                                 mbuf->port = mq->in_port;
 170                                 ret = memif_pktmbuf_chain(mbuf_head, mbuf_tail, mbuf);
 171                                 if (unlikely(ret < 0)) {
 172                                         MIF_LOG(ERR, "%s: number-of-segments-overflow",
 173                                                 rte_vdev_device_name(pmd->vdev));
 174                                         rte_pktmbuf_free(mbuf);
 175                                         goto no_free_bufs;
 176                                 }
 177                         }
 178                         cp_len = RTE_MIN(dst_len, src_len);
 179
 180                         rte_pktmbuf_data_len(mbuf) += cp_len;
 181                         rte_pktmbuf_pkt_len(mbuf) = rte_pktmbuf_data_len(mbuf);
 182                         if (mbuf != mbuf_head)
 183                                 rte_pktmbuf_pkt_len(mbuf_head) += cp_len;
 184
 185                         memcpy(rte_pktmbuf_mtod_offset(mbuf, void *, dst_off),
 186                                (uint8_t *)memif_get_buffer(pmd, d0) + src_off, cp_len);
 187
 188                         src_off += cp_len;
 189                         dst_off += cp_len;
 190                         src_len -= cp_len;
 191                 } while (src_len);
 192
 193                 cur_slot++;
 194                 n_slots--;
 195
 196                 if (d0->flags & MEMIF_DESC_FLAG_NEXT)
 197                         goto next_slot;
 198
 199                 mq->n_bytes += rte_pktmbuf_pkt_len(mbuf_head);
 200                 *bufs++ = mbuf_head;
 201                 n_rx_pkts++;
 202         }
 203
 204 no_free_bufs:
 205         if (type == MEMIF_RING_S2M) {
 206                 rte_mb();
 207                 ring->tail = cur_slot;
 208                 mq->last_head = cur_slot;
 209         } else {
 210                 mq->last_tail = cur_slot;
 211         }
 212
 213 refill:
 214         if (type == MEMIF_RING_M2S) {
 215                 head = ring->head;
 216                 n_slots = ring_size - head + mq->last_tail;
 217
 218                 while (n_slots--) {
 219                         s0 = head++ & mask;
 220                         d0 = &ring->desc[s0];
 221                         d0->length = pmd->run.pkt_buffer_size;
 222                 }
 223                 rte_mb();
 224                 ring->head = head;
 225         }
 226
 227         mq->n_pkts += n_rx_pkts;
 228         return n_rx_pkts;
 229 }
 230
 231 static uint16_t
 232 eth_memif_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 233 {
 234         struct memif_queue *mq = queue;
 235         struct pmd_internals *pmd = mq->pmd;
 236         memif_ring_t *ring = mq->ring;
 237         uint16_t slot, saved_slot, n_free, ring_size, mask, n_tx_pkts = 0;
 238         uint16_t src_len, src_off, dst_len, dst_off, cp_len;
 239         memif_ring_type_t type = mq->type;
 240         memif_desc_t *d0;
 241         struct rte_mbuf *mbuf;
 242         struct rte_mbuf *mbuf_head;
 243         uint64_t a;
 244         ssize_t size;
 245
 246         if (unlikely((pmd->flags & ETH_MEMIF_FLAG_CONNECTED) == 0))
 247                 return 0;
 248         if (unlikely(ring == NULL))
 249                 return 0;
 250
 251         ring_size = 1 << mq->log2_ring_size;
 252         mask = ring_size - 1;
 253
 254         n_free = ring->tail - mq->last_tail;
 255         mq->last_tail += n_free;
 256         slot = (type == MEMIF_RING_S2M) ? ring->head : ring->tail;
 257
 258         if (type == MEMIF_RING_S2M)
 259                 n_free = ring_size - ring->head + mq->last_tail;
 260         else
 261                 n_free = ring->head - ring->tail;
 262
 263         while (n_tx_pkts < nb_pkts && n_free) {
 264                 mbuf_head = *bufs++;
 265                 mbuf = mbuf_head;
 266
 267                 saved_slot = slot;
 268                 d0 = &ring->desc[slot & mask];
 269                 dst_off = 0;
 270                 dst_len = (type == MEMIF_RING_S2M) ?
 271                         pmd->run.pkt_buffer_size : d0->length;
 272
 273 next_in_chain:
 274                 src_off = 0;
 275                 src_len = rte_pktmbuf_data_len(mbuf);
 276
 277                 while (src_len) {
 278                         if (dst_len == 0) {
 279                                 if (n_free) {
 280                                         slot++;
 281                                         n_free--;
 282                                         d0->flags |= MEMIF_DESC_FLAG_NEXT;
 283                                         d0 = &ring->desc[slot & mask];
 284                                         dst_off = 0;
 285                                         dst_len = (type == MEMIF_RING_S2M) ?
 286                                             pmd->run.pkt_buffer_size : d0->length;
 287                                         d0->flags = 0;
 288                                 } else {
 289                                         slot = saved_slot;
 290                                         goto no_free_slots;
 291                                 }
 292                         }
 293                         cp_len = RTE_MIN(dst_len, src_len);
 294
 295                         memcpy((uint8_t *)memif_get_buffer(pmd, d0) + dst_off,
 296                                rte_pktmbuf_mtod_offset(mbuf, void *, src_off),
 297                                cp_len);
 298
 299                         mq->n_bytes += cp_len;
 300                         src_off += cp_len;
 301                         dst_off += cp_len;
 302                         src_len -= cp_len;
 303                         dst_len -= cp_len;
 304
 305                         d0->length = dst_off;
 306                 }
 307
 308                 if (rte_pktmbuf_is_contiguous(mbuf) == 0) {
 309                         mbuf = mbuf->next;
 310                         goto next_in_chain;
 311                 }
 312
 313                 n_tx_pkts++;
 314                 slot++;
 315                 n_free--;
 316                 rte_pktmbuf_free(mbuf_head);
 317         }
 318
 319 no_free_slots:
 320         rte_mb();
 321         if (type == MEMIF_RING_S2M)
 322                 ring->head = slot;
 323         else
 324                 ring->tail = slot;
 325
 326         if ((ring->flags & MEMIF_RING_FLAG_MASK_INT) == 0) {
 327                 a = 1;
 328                 size = write(mq->intr_handle.fd, &a, sizeof(a));
 329                 if (unlikely(size < 0)) {
 330                         MIF_LOG(WARNING,
 331                                 "%s: Failed to send interrupt. %s",
 332                                 rte_vdev_device_name(pmd->vdev), strerror(errno));
 333                 }
 334         }
 335
 336         mq->n_err += nb_pkts - n_tx_pkts;
 337         mq->n_pkts += n_tx_pkts;
 338         return n_tx_pkts;
 339 }
 340
 341 void
 342 memif_free_regions(struct pmd_internals *pmd)
 343 {
 344         int i;
 345         struct memif_region *r;
 346
 347         /* regions are allocated contiguously, so it's
 348          * enough to loop until 'pmd->regions_num'
 349          */
 350         for (i = 0; i < pmd->regions_num; i++) {
 351                 r = pmd->regions[i];
 352                 if (r != NULL) {
 353                         if (r->addr != NULL) {
 354                                 munmap(r->addr, r->region_size);
 355                                 if (r->fd > 0) {
 356                                         close(r->fd);
 357                                         r->fd = -1;
 358                                 }
 359                         }
 360                         rte_free(r);
 361                         pmd->regions[i] = NULL;
 362                 }
 363         }
 364         pmd->regions_num = 0;
 365 }
 366
 367 static int
 368 memif_region_init_shm(struct pmd_internals *pmd, uint8_t has_buffers)
 369 {
 370         char shm_name[ETH_MEMIF_SHM_NAME_SIZE];
 371         int ret = 0;
 372         struct memif_region *r;
 373
 374         if (pmd->regions_num >= ETH_MEMIF_MAX_REGION_NUM) {
 375                 MIF_LOG(ERR, "%s: Too many regions.", rte_vdev_device_name(pmd->vdev));
 376                 return -1;
 377         }
 378
 379         r = rte_zmalloc("region", sizeof(struct memif_region), 0);
 380         if (r == NULL) {
 381                 MIF_LOG(ERR, "%s: Failed to alloc memif region.",
 382                         rte_vdev_device_name(pmd->vdev));
 383                 return -ENOMEM;
 384         }
 385
 386         /* calculate buffer offset */
 387         r->pkt_buffer_offset = (pmd->run.num_s2m_rings + pmd->run.num_m2s_rings) *
 388             (sizeof(memif_ring_t) + sizeof(memif_desc_t) *
 389             (1 << pmd->run.log2_ring_size));
 390
 391         r->region_size = r->pkt_buffer_offset;
 392         /* if region has buffers, add buffers size to region_size */
 393         if (has_buffers == 1)
 394                 r->region_size += (uint32_t)(pmd->run.pkt_buffer_size *
 395                         (1 << pmd->run.log2_ring_size) *
 396                         (pmd->run.num_s2m_rings +
 397                          pmd->run.num_m2s_rings));
 398
 399         memset(shm_name, 0, sizeof(char) * ETH_MEMIF_SHM_NAME_SIZE);
 400         snprintf(shm_name, ETH_MEMIF_SHM_NAME_SIZE, "memif_region_%d",
 401                  pmd->regions_num);
 402
 403         r->fd = memfd_create(shm_name, MFD_ALLOW_SEALING);
 404         if (r->fd < 0) {
 405                 MIF_LOG(ERR, "%s: Failed to create shm file: %s.",
 406                         rte_vdev_device_name(pmd->vdev),
 407                         strerror(errno));
 408                 ret = -1;
 409                 goto error;
 410         }
 411
 412         ret = fcntl(r->fd, F_ADD_SEALS, F_SEAL_SHRINK);
 413         if (ret < 0) {
 414                 MIF_LOG(ERR, "%s: Failed to add seals to shm file: %s.",
 415                         rte_vdev_device_name(pmd->vdev),
 416                         strerror(errno));
 417                 goto error;
 418         }
 419
 420         ret = ftruncate(r->fd, r->region_size);
 421         if (ret < 0) {
 422                 MIF_LOG(ERR, "%s: Failed to truncate shm file: %s.",
 423                         rte_vdev_device_name(pmd->vdev),
 424                         strerror(errno));
 425                 goto error;
 426         }
 427
 428         r->addr = mmap(NULL, r->region_size, PROT_READ |
 429                        PROT_WRITE, MAP_SHARED, r->fd, 0);
 430         if (r->addr == MAP_FAILED) {
 431                 MIF_LOG(ERR, "%s: Failed to mmap shm region: %s.",
 432                         rte_vdev_device_name(pmd->vdev),
 433                         strerror(ret));
 434                 ret = -1;
 435                 goto error;
 436         }
 437
 438         pmd->regions[pmd->regions_num] = r;
 439         pmd->regions_num++;
 440
 441         return ret;
 442
 443 error:
 444         if (r->fd > 0)
 445                 close(r->fd);
 446         r->fd = -1;
 447
 448         return ret;
 449 }
 450
 451 static int
 452 memif_regions_init(struct pmd_internals *pmd)
 453 {
 454         int ret;
 455
 456         /* create one buffer region */
 457         ret = memif_region_init_shm(pmd, /* has buffer */ 1);
 458         if (ret < 0)
 459                 return ret;
 460
 461         return 0;
 462 }
 463
 464 static void
 465 memif_init_rings(struct rte_eth_dev *dev)
 466 {
 467         struct pmd_internals *pmd = dev->data->dev_private;
 468         memif_ring_t *ring;
 469         int i, j;
 470         uint16_t slot;
 471
 472         for (i = 0; i < pmd->run.num_s2m_rings; i++) {
 473                 ring = memif_get_ring(pmd, MEMIF_RING_S2M, i);
 474                 ring->head = 0;
 475                 ring->tail = 0;
 476                 ring->cookie = MEMIF_COOKIE;
 477                 ring->flags = 0;
 478                 for (j = 0; j < (1 << pmd->run.log2_ring_size); j++) {
 479                         slot = i * (1 << pmd->run.log2_ring_size) + j;
 480                         ring->desc[j].region = 0;
 481                         ring->desc[j].offset = pmd->regions[0]->pkt_buffer_offset +
 482                                 (uint32_t)(slot * pmd->run.pkt_buffer_size);
 483                         ring->desc[j].length = pmd->run.pkt_buffer_size;
 484                 }
 485         }
 486
 487         for (i = 0; i < pmd->run.num_m2s_rings; i++) {
 488                 ring = memif_get_ring(pmd, MEMIF_RING_M2S, i);
 489                 ring->head = 0;
 490                 ring->tail = 0;
 491                 ring->cookie = MEMIF_COOKIE;
 492                 ring->flags = 0;
 493                 for (j = 0; j < (1 << pmd->run.log2_ring_size); j++) {
 494                         slot = (i + pmd->run.num_s2m_rings) *
 495                             (1 << pmd->run.log2_ring_size) + j;
 496                         ring->desc[j].region = 0;
 497                         ring->desc[j].offset = pmd->regions[0]->pkt_buffer_offset +
 498                                 (uint32_t)(slot * pmd->run.pkt_buffer_size);
 499                         ring->desc[j].length = pmd->run.pkt_buffer_size;
 500                 }
 501         }
 502 }
 503
 504 /* called only by slave */
 505 static void
 506 memif_init_queues(struct rte_eth_dev *dev)
 507 {
 508         struct pmd_internals *pmd = dev->data->dev_private;
 509         struct memif_queue *mq;
 510         int i;
 511
 512         for (i = 0; i < pmd->run.num_s2m_rings; i++) {
 513                 mq = dev->data->tx_queues[i];
 514                 mq->ring = memif_get_ring(pmd, MEMIF_RING_S2M, i);
 515                 mq->log2_ring_size = pmd->run.log2_ring_size;
 516                 /* queues located only in region 0 */
 517                 mq->region = 0;
 518                 mq->ring_offset = (uint8_t *)mq->ring - (uint8_t *)pmd->regions[0]->addr;
 519                 mq->last_head = 0;
 520                 mq->last_tail = 0;
 521                 mq->intr_handle.fd = eventfd(0, EFD_NONBLOCK);
 522                 if (mq->intr_handle.fd < 0) {
 523                         MIF_LOG(WARNING,
 524                                 "%s: Failed to create eventfd for tx queue %d: %s.",
 525                                 rte_vdev_device_name(pmd->vdev), i,
 526                                 strerror(errno));
 527                 }
 528         }
 529
 530         for (i = 0; i < pmd->run.num_m2s_rings; i++) {
 531                 mq = dev->data->rx_queues[i];
 532                 mq->ring = memif_get_ring(pmd, MEMIF_RING_M2S, i);
 533                 mq->log2_ring_size = pmd->run.log2_ring_size;
 534                 /* queues located only in region 0 */
 535                 mq->region = 0;
 536                 mq->ring_offset = (uint8_t *)mq->ring - (uint8_t *)pmd->regions[0]->addr;
 537                 mq->last_head = 0;
 538                 mq->last_tail = 0;
 539                 mq->intr_handle.fd = eventfd(0, EFD_NONBLOCK);
 540                 if (mq->intr_handle.fd < 0) {
 541                         MIF_LOG(WARNING,
 542                                 "%s: Failed to create eventfd for rx queue %d: %s.",
 543                                 rte_vdev_device_name(pmd->vdev), i,
 544                                 strerror(errno));
 545                 }
 546         }
 547 }
 548
 549 int
 550 memif_init_regions_and_queues(struct rte_eth_dev *dev)
 551 {
 552         int ret;
 553
 554         ret = memif_regions_init(dev->data->dev_private);
 555         if (ret < 0)
 556                 return ret;
 557
 558         memif_init_rings(dev);
 559
 560         memif_init_queues(dev);
 561
 562         return 0;
 563 }
 564
 565 int
 566 memif_connect(struct rte_eth_dev *dev)
 567 {
 568         struct pmd_internals *pmd = dev->data->dev_private;
 569         struct memif_region *mr;
 570         struct memif_queue *mq;
 571         int i;
 572
 573         for (i = 0; i < pmd->regions_num; i++) {
 574                 mr = pmd->regions[i];
 575                 if (mr != NULL) {
 576                         if (mr->addr == NULL) {
 577                                 if (mr->fd < 0)
 578                                         return -1;
 579                                 mr->addr = mmap(NULL, mr->region_size,
 580                                                 PROT_READ | PROT_WRITE,
 581                                                 MAP_SHARED, mr->fd, 0);
 582                                 if (mr->addr == NULL)
 583                                         return -1;
 584                         }
 585                 }
 586         }
 587
 588         for (i = 0; i < pmd->run.num_s2m_rings; i++) {
 589                 mq = (pmd->role == MEMIF_ROLE_SLAVE) ?
 590                     dev->data->tx_queues[i] : dev->data->rx_queues[i];
 591                 mq->ring = (memif_ring_t *)((uint8_t *)pmd->regions[mq->region]->addr +
 592                             mq->ring_offset);
 593                 if (mq->ring->cookie != MEMIF_COOKIE) {
 594                         MIF_LOG(ERR, "%s: Wrong cookie",
 595                                 rte_vdev_device_name(pmd->vdev));
 596                         return -1;
 597                 }
 598                 mq->ring->head = 0;
 599                 mq->ring->tail = 0;
 600                 mq->last_head = 0;
 601                 mq->last_tail = 0;
 602                 /* enable polling mode */
 603                 if (pmd->role == MEMIF_ROLE_MASTER)
 604                         mq->ring->flags = MEMIF_RING_FLAG_MASK_INT;
 605         }
 606         for (i = 0; i < pmd->run.num_m2s_rings; i++) {
 607                 mq = (pmd->role == MEMIF_ROLE_SLAVE) ?
 608                     dev->data->rx_queues[i] : dev->data->tx_queues[i];
 609                 mq->ring = (memif_ring_t *)((uint8_t *)pmd->regions[mq->region]->addr +
 610                             mq->ring_offset);
 611                 if (mq->ring->cookie != MEMIF_COOKIE) {
 612                         MIF_LOG(ERR, "%s: Wrong cookie",
 613                                 rte_vdev_device_name(pmd->vdev));
 614                         return -1;
 615                 }
 616                 mq->ring->head = 0;
 617                 mq->ring->tail = 0;
 618                 mq->last_head = 0;
 619                 mq->last_tail = 0;
 620                 /* enable polling mode */
 621                 if (pmd->role == MEMIF_ROLE_SLAVE)
 622                         mq->ring->flags = MEMIF_RING_FLAG_MASK_INT;
 623         }
 624
 625         pmd->flags &= ~ETH_MEMIF_FLAG_CONNECTING;
 626         pmd->flags |= ETH_MEMIF_FLAG_CONNECTED;
 627         dev->data->dev_link.link_status = ETH_LINK_UP;
 628         MIF_LOG(INFO, "%s: Connected.", rte_vdev_device_name(pmd->vdev));
 629         return 0;
 630 }
 631
 632 static int
 633 memif_dev_start(struct rte_eth_dev *dev)
 634 {
 635         struct pmd_internals *pmd = dev->data->dev_private;
 636         int ret = 0;
 637
 638         switch (pmd->role) {
 639         case MEMIF_ROLE_SLAVE:
 640                 ret = memif_connect_slave(dev);
 641                 break;
 642         case MEMIF_ROLE_MASTER:
 643                 ret = memif_connect_master(dev);
 644                 break;
 645         default:
 646                 MIF_LOG(ERR, "%s: Unknown role: %d.",
 647                         rte_vdev_device_name(pmd->vdev), pmd->role);
 648                 ret = -1;
 649                 break;
 650         }
 651
 652         return ret;
 653 }
 654
 655 static void
 656 memif_dev_close(struct rte_eth_dev *dev)
 657 {
 658         struct pmd_internals *pmd = dev->data->dev_private;
 659         int i;
 660
 661         memif_msg_enq_disconnect(pmd->cc, "Device closed", 0);
 662         memif_disconnect(dev);
 663
 664         for (i = 0; i < dev->data->nb_rx_queues; i++)
 665                 (*dev->dev_ops->rx_queue_release)(dev->data->rx_queues[i]);
 666         for (i = 0; i < dev->data->nb_tx_queues; i++)
 667                 (*dev->dev_ops->tx_queue_release)(dev->data->tx_queues[i]);
 668
 669         memif_socket_remove_device(dev);
 670 }
 671
 672 static int
 673 memif_dev_configure(struct rte_eth_dev *dev)
 674 {
 675         struct pmd_internals *pmd = dev->data->dev_private;
 676
 677         /*
 678          * SLAVE - TXQ
 679          * MASTER - RXQ
 680          */
 681         pmd->cfg.num_s2m_rings = (pmd->role == MEMIF_ROLE_SLAVE) ?
 682                                   dev->data->nb_tx_queues : dev->data->nb_rx_queues;
 683
 684         /*
 685          * SLAVE - RXQ
 686          * MASTER - TXQ
 687          */
 688         pmd->cfg.num_m2s_rings = (pmd->role == MEMIF_ROLE_SLAVE) ?
 689                                   dev->data->nb_rx_queues : dev->data->nb_tx_queues;
 690
 691         return 0;
 692 }
 693
 694 static int
 695 memif_tx_queue_setup(struct rte_eth_dev *dev,
 696                      uint16_t qid,
 697                      uint16_t nb_tx_desc __rte_unused,
 698                      unsigned int socket_id __rte_unused,
 699                      const struct rte_eth_txconf *tx_conf __rte_unused)
 700 {
 701         struct pmd_internals *pmd = dev->data->dev_private;
 702         struct memif_queue *mq;
 703
 704         mq = rte_zmalloc("tx-queue", sizeof(struct memif_queue), 0);
 705         if (mq == NULL) {
 706                 MIF_LOG(ERR, "%s: Failed to allocate tx queue id: %u",
 707                         rte_vdev_device_name(pmd->vdev), qid);
 708                 return -ENOMEM;
 709         }
 710
 711         mq->type =
 712             (pmd->role == MEMIF_ROLE_SLAVE) ? MEMIF_RING_S2M : MEMIF_RING_M2S;
 713         mq->n_pkts = 0;
 714         mq->n_bytes = 0;
 715         mq->n_err = 0;
 716         mq->intr_handle.fd = -1;
 717         mq->intr_handle.type = RTE_INTR_HANDLE_EXT;
 718         mq->pmd = pmd;
 719         dev->data->tx_queues[qid] = mq;
 720
 721         return 0;
 722 }
 723
 724 static int
 725 memif_rx_queue_setup(struct rte_eth_dev *dev,
 726                      uint16_t qid,
 727                      uint16_t nb_rx_desc __rte_unused,
 728                      unsigned int socket_id __rte_unused,
 729                      const struct rte_eth_rxconf *rx_conf __rte_unused,
 730                      struct rte_mempool *mb_pool)
 731 {
 732         struct pmd_internals *pmd = dev->data->dev_private;
 733         struct memif_queue *mq;
 734
 735         mq = rte_zmalloc("rx-queue", sizeof(struct memif_queue), 0);
 736         if (mq == NULL) {
 737                 MIF_LOG(ERR, "%s: Failed to allocate rx queue id: %u",
 738                         rte_vdev_device_name(pmd->vdev), qid);
 739                 return -ENOMEM;
 740         }
 741
 742         mq->type = (pmd->role == MEMIF_ROLE_SLAVE) ? MEMIF_RING_M2S : MEMIF_RING_S2M;
 743         mq->n_pkts = 0;
 744         mq->n_bytes = 0;
 745         mq->n_err = 0;
 746         mq->intr_handle.fd = -1;
 747         mq->intr_handle.type = RTE_INTR_HANDLE_EXT;
 748         mq->mempool = mb_pool;
 749         mq->in_port = dev->data->port_id;
 750         mq->pmd = pmd;
 751         dev->data->rx_queues[qid] = mq;
 752
 753         return 0;
 754 }
 755
 756 static void
 757 memif_queue_release(void *queue)
 758 {
 759         struct memif_queue *mq = (struct memif_queue *)queue;
 760
 761         if (!mq)
 762                 return;
 763
 764         rte_free(mq);
 765 }
 766
 767 static int
 768 memif_link_update(struct rte_eth_dev *dev __rte_unused,
 769                   int wait_to_complete __rte_unused)
 770 {
 771         return 0;
 772 }
 773
 774 static int
 775 memif_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 776 {
 777         struct pmd_internals *pmd = dev->data->dev_private;
 778         struct memif_queue *mq;
 779         int i;
 780         uint8_t tmp, nq;
 781
 782         stats->ipackets = 0;
 783         stats->ibytes = 0;
 784         stats->opackets = 0;
 785         stats->obytes = 0;
 786         stats->oerrors = 0;
 787
 788         tmp = (pmd->role == MEMIF_ROLE_SLAVE) ? pmd->run.num_s2m_rings :
 789             pmd->run.num_m2s_rings;
 790         nq = (tmp < RTE_ETHDEV_QUEUE_STAT_CNTRS) ? tmp :
 791             RTE_ETHDEV_QUEUE_STAT_CNTRS;
 792
 793         /* RX stats */
 794         for (i = 0; i < nq; i++) {
 795                 mq = dev->data->rx_queues[i];
 796                 stats->q_ipackets[i] = mq->n_pkts;
 797                 stats->q_ibytes[i] = mq->n_bytes;
 798                 stats->ipackets += mq->n_pkts;
 799                 stats->ibytes += mq->n_bytes;
 800         }
 801
 802         tmp = (pmd->role == MEMIF_ROLE_SLAVE) ? pmd->run.num_m2s_rings :
 803             pmd->run.num_s2m_rings;
 804         nq = (tmp < RTE_ETHDEV_QUEUE_STAT_CNTRS) ? tmp :
 805             RTE_ETHDEV_QUEUE_STAT_CNTRS;
 806
 807         /* TX stats */
 808         for (i = 0; i < nq; i++) {
 809                 mq = dev->data->tx_queues[i];
 810                 stats->q_opackets[i] = mq->n_pkts;
 811                 stats->q_obytes[i] = mq->n_bytes;
 812                 stats->opackets += mq->n_pkts;
 813                 stats->obytes += mq->n_bytes;
 814                 stats->oerrors += mq->n_err;
 815         }
 816         return 0;
 817 }
 818
 819 static void
 820 memif_stats_reset(struct rte_eth_dev *dev)
 821 {
 822         struct pmd_internals *pmd = dev->data->dev_private;
 823         int i;
 824         struct memif_queue *mq;
 825
 826         for (i = 0; i < pmd->run.num_s2m_rings; i++) {
 827                 mq = (pmd->role == MEMIF_ROLE_SLAVE) ? dev->data->tx_queues[i] :
 828                     dev->data->rx_queues[i];
 829                 mq->n_pkts = 0;
 830                 mq->n_bytes = 0;
 831                 mq->n_err = 0;
 832         }
 833         for (i = 0; i < pmd->run.num_m2s_rings; i++) {
 834                 mq = (pmd->role == MEMIF_ROLE_SLAVE) ? dev->data->rx_queues[i] :
 835                     dev->data->tx_queues[i];
 836                 mq->n_pkts = 0;
 837                 mq->n_bytes = 0;
 838                 mq->n_err = 0;
 839         }
 840 }
 841
 842 static int
 843 memif_rx_queue_intr_enable(struct rte_eth_dev *dev, uint16_t qid __rte_unused)
 844 {
 845         struct pmd_internals *pmd = dev->data->dev_private;
 846
 847         MIF_LOG(WARNING, "%s: Interrupt mode not supported.",
 848                 rte_vdev_device_name(pmd->vdev));
 849
 850         return -1;
 851 }
 852
 853 static int
 854 memif_rx_queue_intr_disable(struct rte_eth_dev *dev, uint16_t qid __rte_unused)
 855 {
 856         struct pmd_internals *pmd __rte_unused = dev->data->dev_private;
 857
 858         return 0;
 859 }
 860
 861 static const struct eth_dev_ops ops = {
 862         .dev_start = memif_dev_start,
 863         .dev_close = memif_dev_close,
 864         .dev_infos_get = memif_dev_info,
 865         .dev_configure = memif_dev_configure,
 866         .tx_queue_setup = memif_tx_queue_setup,
 867         .rx_queue_setup = memif_rx_queue_setup,
 868         .rx_queue_release = memif_queue_release,
 869         .tx_queue_release = memif_queue_release,
 870         .rx_queue_intr_enable = memif_rx_queue_intr_enable,
 871         .rx_queue_intr_disable = memif_rx_queue_intr_disable,
 872         .link_update = memif_link_update,
 873         .stats_get = memif_stats_get,
 874         .stats_reset = memif_stats_reset,
 875 };
 876
 877 static int
 878 memif_create(struct rte_vdev_device *vdev, enum memif_role_t role,
 879              memif_interface_id_t id, uint32_t flags,
 880              const char *socket_filename,
 881              memif_log2_ring_size_t log2_ring_size,
 882              uint16_t pkt_buffer_size, const char *secret,
 883              struct rte_ether_addr *ether_addr)
 884 {
 885         int ret = 0;
 886         struct rte_eth_dev *eth_dev;
 887         struct rte_eth_dev_data *data;
 888         struct pmd_internals *pmd;
 889         const unsigned int numa_node = vdev->device.numa_node;
 890         const char *name = rte_vdev_device_name(vdev);
 891
 892         if (flags & ETH_MEMIF_FLAG_ZERO_COPY) {
 893                 MIF_LOG(ERR, "Zero-copy slave not supported.");
 894                 return -1;
 895         }
 896
 897         eth_dev = rte_eth_vdev_allocate(vdev, sizeof(*pmd));
 898         if (eth_dev == NULL) {
 899                 MIF_LOG(ERR, "%s: Unable to allocate device struct.", name);
 900                 return -1;
 901         }
 902
 903         pmd = eth_dev->data->dev_private;
 904         memset(pmd, 0, sizeof(*pmd));
 905
 906         pmd->vdev = vdev;
 907         pmd->id = id;
 908         pmd->flags = flags;
 909         pmd->flags |= ETH_MEMIF_FLAG_DISABLED;
 910         pmd->role = role;
 911
 912         ret = memif_socket_init(eth_dev, socket_filename);
 913         if (ret < 0)
 914                 return ret;
 915
 916         memset(pmd->secret, 0, sizeof(char) * ETH_MEMIF_SECRET_SIZE);
 917         if (secret != NULL)
 918                 strlcpy(pmd->secret, secret, sizeof(pmd->secret));
 919
 920         pmd->cfg.log2_ring_size = log2_ring_size;
 921         /* set in .dev_configure() */
 922         pmd->cfg.num_s2m_rings = 0;
 923         pmd->cfg.num_m2s_rings = 0;
 924
 925         pmd->cfg.pkt_buffer_size = pkt_buffer_size;
 926
 927         data = eth_dev->data;
 928         data->dev_private = pmd;
 929         data->numa_node = numa_node;
 930         data->mac_addrs = ether_addr;
 931
 932         eth_dev->dev_ops = &ops;
 933         eth_dev->device = &vdev->device;
 934         eth_dev->rx_pkt_burst = eth_memif_rx;
 935         eth_dev->tx_pkt_burst = eth_memif_tx;
 936
 937         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
 938
 939         rte_eth_dev_probing_finish(eth_dev);
 940
 941         return 0;
 942 }
 943
 944 static int
 945 memif_set_role(const char *key __rte_unused, const char *value,
 946                void *extra_args)
 947 {
 948         enum memif_role_t *role = (enum memif_role_t *)extra_args;
 949
 950         if (strstr(value, "master") != NULL) {
 951                 *role = MEMIF_ROLE_MASTER;
 952         } else if (strstr(value, "slave") != NULL) {
 953                 *role = MEMIF_ROLE_SLAVE;
 954         } else {
 955                 MIF_LOG(ERR, "Unknown role: %s.", value);
 956                 return -EINVAL;
 957         }
 958         return 0;
 959 }
 960
 961 static int
 962 memif_set_zc(const char *key __rte_unused, const char *value, void *extra_args)
 963 {
 964         uint32_t *flags = (uint32_t *)extra_args;
 965
 966         if (strstr(value, "yes") != NULL) {
 967                 *flags |= ETH_MEMIF_FLAG_ZERO_COPY;
 968         } else if (strstr(value, "no") != NULL) {
 969                 *flags &= ~ETH_MEMIF_FLAG_ZERO_COPY;
 970         } else {
 971                 MIF_LOG(ERR, "Failed to parse zero-copy param: %s.", value);
 972                 return -EINVAL;
 973         }
 974         return 0;
 975 }
 976
 977 static int
 978 memif_set_id(const char *key __rte_unused, const char *value, void *extra_args)
 979 {
 980         memif_interface_id_t *id = (memif_interface_id_t *)extra_args;
 981
 982         /* even if parsing fails, 0 is a valid id */
 983         *id = strtoul(value, NULL, 10);
 984         return 0;
 985 }
 986
 987 static int
 988 memif_set_bs(const char *key __rte_unused, const char *value, void *extra_args)
 989 {
 990         unsigned long tmp;
 991         uint16_t *pkt_buffer_size = (uint16_t *)extra_args;
 992
 993         tmp = strtoul(value, NULL, 10);
 994         if (tmp == 0 || tmp > 0xFFFF) {
 995                 MIF_LOG(ERR, "Invalid buffer size: %s.", value);
 996                 return -EINVAL;
 997         }
 998         *pkt_buffer_size = tmp;
 999         return 0;
1000 }
1001
1002 static int
1003 memif_set_rs(const char *key __rte_unused, const char *value, void *extra_args)
1004 {
1005         unsigned long tmp;
1006         memif_log2_ring_size_t *log2_ring_size =
1007             (memif_log2_ring_size_t *)extra_args;
1008
1009         tmp = strtoul(value, NULL, 10);
1010         if (tmp == 0 || tmp > ETH_MEMIF_MAX_LOG2_RING_SIZE) {
1011                 MIF_LOG(ERR, "Invalid ring size: %s (max %u).",
1012                         value, ETH_MEMIF_MAX_LOG2_RING_SIZE);
1013                 return -EINVAL;
1014         }
1015         *log2_ring_size = tmp;
1016         return 0;
1017 }
1018
1019 /* check if directory exists and if we have permission to read/write */
1020 static int
1021 memif_check_socket_filename(const char *filename)
1022 {
1023         char *dir = NULL, *tmp;
1024         uint32_t idx;
1025         int ret = 0;
1026
1027         tmp = strrchr(filename, '/');
1028         if (tmp != NULL) {
1029                 idx = tmp - filename;
1030                 dir = rte_zmalloc("memif_tmp", sizeof(char) * (idx + 1), 0);
1031                 if (dir == NULL) {
1032                         MIF_LOG(ERR, "Failed to allocate memory.");
1033                         return -1;
1034                 }
1035                 strlcpy(dir, filename, sizeof(char) * (idx + 1));
1036         }
1037
1038         if (dir == NULL || (faccessat(-1, dir, F_OK | R_OK |
1039                                         W_OK, AT_EACCESS) < 0)) {
1040                 MIF_LOG(ERR, "Invalid socket directory.");
1041                 ret = -EINVAL;
1042         }
1043
1044         if (dir != NULL)
1045                 rte_free(dir);
1046
1047         return ret;
1048 }
1049
1050 static int
1051 memif_set_socket_filename(const char *key __rte_unused, const char *value,
1052                           void *extra_args)
1053 {
1054         const char **socket_filename = (const char **)extra_args;
1055
1056         *socket_filename = value;
1057         return memif_check_socket_filename(*socket_filename);
1058 }
1059
1060 static int
1061 memif_set_mac(const char *key __rte_unused, const char *value, void *extra_args)
1062 {
1063         struct rte_ether_addr *ether_addr = (struct rte_ether_addr *)extra_args;
1064         int ret = 0;
1065
1066         ret = sscanf(value, "%hhx:%hhx:%hhx:%hhx:%hhx:%hhx",
1067                &ether_addr->addr_bytes[0], &ether_addr->addr_bytes[1],
1068                &ether_addr->addr_bytes[2], &ether_addr->addr_bytes[3],
1069                &ether_addr->addr_bytes[4], &ether_addr->addr_bytes[5]);
1070         if (ret != 6)
1071                 MIF_LOG(WARNING, "Failed to parse mac '%s'.", value);
1072         return 0;
1073 }
1074
1075 static int
1076 memif_set_secret(const char *key __rte_unused, const char *value, void *extra_args)
1077 {
1078         const char **secret = (const char **)extra_args;
1079
1080         *secret = value;
1081         return 0;
1082 }
1083
1084 static int
1085 rte_pmd_memif_probe(struct rte_vdev_device *vdev)
1086 {
1087         RTE_BUILD_BUG_ON(sizeof(memif_msg_t) != 128);
1088         RTE_BUILD_BUG_ON(sizeof(memif_desc_t) != 16);
1089         int ret = 0;
1090         struct rte_kvargs *kvlist;
1091         const char *name = rte_vdev_device_name(vdev);
1092         enum memif_role_t role = MEMIF_ROLE_SLAVE;
1093         memif_interface_id_t id = 0;
1094         uint16_t pkt_buffer_size = ETH_MEMIF_DEFAULT_PKT_BUFFER_SIZE;
1095         memif_log2_ring_size_t log2_ring_size = ETH_MEMIF_DEFAULT_RING_SIZE;
1096         const char *socket_filename = ETH_MEMIF_DEFAULT_SOCKET_FILENAME;
1097         uint32_t flags = 0;
1098         const char *secret = NULL;
1099         struct rte_ether_addr *ether_addr = rte_zmalloc("",
1100                 sizeof(struct rte_ether_addr), 0);
1101
1102         rte_eth_random_addr(ether_addr->addr_bytes);
1103
1104         MIF_LOG(INFO, "Initialize MEMIF: %s.", name);
1105
1106         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1107                 MIF_LOG(ERR, "Multi-processing not supported for memif.");
1108                 /* TODO:
1109                  * Request connection information.
1110                  *
1111                  * Once memif in the primary process is connected,
1112                  * broadcast connection information.
1113                  */
1114                 return -1;
1115         }
1116
1117         kvlist = rte_kvargs_parse(rte_vdev_device_args(vdev), valid_arguments);
1118
1119         /* parse parameters */
1120         if (kvlist != NULL) {
1121                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_ROLE_ARG,
1122                                          &memif_set_role, &role);
1123                 if (ret < 0)
1124                         goto exit;
1125                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_ID_ARG,
1126                                          &memif_set_id, &id);
1127                 if (ret < 0)
1128                         goto exit;
1129                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_PKT_BUFFER_SIZE_ARG,
1130                                          &memif_set_bs, &pkt_buffer_size);
1131                 if (ret < 0)
1132                         goto exit;
1133                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_RING_SIZE_ARG,
1134                                          &memif_set_rs, &log2_ring_size);
1135                 if (ret < 0)
1136                         goto exit;
1137                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_SOCKET_ARG,
1138                                          &memif_set_socket_filename,
1139                                          (void *)(&socket_filename));
1140                 if (ret < 0)
1141                         goto exit;
1142                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_MAC_ARG,
1143                                          &memif_set_mac, ether_addr);
1144                 if (ret < 0)
1145                         goto exit;
1146                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_ZC_ARG,
1147                                          &memif_set_zc, &flags);
1148                 if (ret < 0)
1149                         goto exit;
1150                 ret = rte_kvargs_process(kvlist, ETH_MEMIF_SECRET_ARG,
1151                                          &memif_set_secret, (void *)(&secret));
1152                 if (ret < 0)
1153                         goto exit;
1154         }
1155
1156         /* create interface */
1157         ret = memif_create(vdev, role, id, flags, socket_filename,
1158                            log2_ring_size, pkt_buffer_size, secret, ether_addr);
1159
1160 exit:
1161         if (kvlist != NULL)
1162                 rte_kvargs_free(kvlist);
1163         return ret;
1164 }
1165
1166 static int
1167 rte_pmd_memif_remove(struct rte_vdev_device *vdev)
1168 {
1169         struct rte_eth_dev *eth_dev;
1170
1171         eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(vdev));
1172         if (eth_dev == NULL)
1173                 return 0;
1174
1175         rte_eth_dev_close(eth_dev->data->port_id);
1176
1177         return 0;
1178 }
1179
1180 static struct rte_vdev_driver pmd_memif_drv = {
1181         .probe = rte_pmd_memif_probe,
1182         .remove = rte_pmd_memif_remove,
1183 };
1184
1185 RTE_PMD_REGISTER_VDEV(net_memif, pmd_memif_drv);
1186
1187 RTE_PMD_REGISTER_PARAM_STRING(net_memif,
1188                               ETH_MEMIF_ID_ARG "=<int>"
1189                               ETH_MEMIF_ROLE_ARG "=master|slave"
1190                               ETH_MEMIF_PKT_BUFFER_SIZE_ARG "=<int>"
1191                               ETH_MEMIF_RING_SIZE_ARG "=<int>"
1192                               ETH_MEMIF_SOCKET_ARG "=<string>"
1193                               ETH_MEMIF_MAC_ARG "=xx:xx:xx:xx:xx:xx"
1194                               ETH_MEMIF_ZC_ARG "=yes|no"
1195                               ETH_MEMIF_SECRET_ARG "=<string>");
1196
1197 int memif_logtype;
1198
1199 RTE_INIT(memif_init_log)
1200 {
1201         memif_logtype = rte_log_register("pmd.net.memif");
1202         if (memif_logtype >= 0)
1203                 rte_log_set_level(memif_logtype, RTE_LOG_NOTICE);
1204 }