drivers/net/szedata2/rte_eth_szedata2.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2015 - 2016 CESNET
   3  */
   4
   5 #include <stdint.h>
   6 #include <unistd.h>
   7 #include <stdbool.h>
   8 #include <err.h>
   9 #include <sys/types.h>
  10 #include <dirent.h>
  11 #include <sys/stat.h>
  12 #include <fcntl.h>
  13 #include <sys/mman.h>
  14
  15 #include <libsze2.h>
  16
  17 #include <rte_mbuf.h>
  18 #include <rte_ethdev_driver.h>
  19 #include <rte_ethdev_pci.h>
  20 #include <rte_malloc.h>
  21 #include <rte_memcpy.h>
  22 #include <rte_kvargs.h>
  23 #include <rte_dev.h>
  24
  25 #include "rte_eth_szedata2.h"
  26 #include "szedata2_logs.h"
  27
  28 #define RTE_ETH_SZEDATA2_MAX_RX_QUEUES 32
  29 #define RTE_ETH_SZEDATA2_MAX_TX_QUEUES 32
  30 #define RTE_ETH_SZEDATA2_TX_LOCK_SIZE (32 * 1024 * 1024)
  31
  32 /**
  33  * size of szedata2_packet header with alignment
  34  */
  35 #define RTE_SZE2_PACKET_HEADER_SIZE_ALIGNED 8
  36
  37 #define RTE_SZEDATA2_DRIVER_NAME net_szedata2
  38
  39 #define SZEDATA2_DEV_PATH_FMT "/dev/szedataII%u"
  40
  41 /**
  42  * Format string for suffix used to differentiate between Ethernet ports
  43  * on the same PCI device.
  44  */
  45 #define SZEDATA2_ETH_DEV_NAME_SUFFIX_FMT "-port%u"
  46
  47 /**
  48  * Maximum number of ports for one device.
  49  */
  50 #define SZEDATA2_MAX_PORTS 2
  51
  52 /**
  53  * Entry in list of PCI devices for this driver.
  54  */
  55 struct pci_dev_list_entry;
  56 struct pci_dev_list_entry {
  57         LIST_ENTRY(pci_dev_list_entry) next;
  58         struct rte_pci_device *pci_dev;
  59         unsigned int port_count;
  60 };
  61
  62 /* List of PCI devices with number of ports for this driver. */
  63 LIST_HEAD(pci_dev_list, pci_dev_list_entry) szedata2_pci_dev_list =
  64         LIST_HEAD_INITIALIZER(szedata2_pci_dev_list);
  65
  66 struct port_info {
  67         unsigned int rx_base_id;
  68         unsigned int tx_base_id;
  69         unsigned int rx_count;
  70         unsigned int tx_count;
  71         int numa_node;
  72 };
  73
  74 struct pmd_internals {
  75         struct rte_eth_dev *dev;
  76         uint16_t max_rx_queues;
  77         uint16_t max_tx_queues;
  78         unsigned int rxq_base_id;
  79         unsigned int txq_base_id;
  80         char *sze_dev_path;
  81 };
  82
  83 struct szedata2_rx_queue {
  84         struct pmd_internals *priv;
  85         struct szedata *sze;
  86         uint8_t rx_channel;
  87         uint16_t qid;
  88         uint16_t in_port;
  89         struct rte_mempool *mb_pool;
  90         volatile uint64_t rx_pkts;
  91         volatile uint64_t rx_bytes;
  92         volatile uint64_t err_pkts;
  93 };
  94
  95 struct szedata2_tx_queue {
  96         struct pmd_internals *priv;
  97         struct szedata *sze;
  98         uint8_t tx_channel;
  99         uint16_t qid;
 100         volatile uint64_t tx_pkts;
 101         volatile uint64_t tx_bytes;
 102         volatile uint64_t err_pkts;
 103 };
 104
 105 static struct rte_ether_addr eth_addr = {
 106         .addr_bytes = { 0x00, 0x11, 0x17, 0x00, 0x00, 0x00 }
 107 };
 108
 109 static uint16_t
 110 eth_szedata2_rx(void *queue,
 111                 struct rte_mbuf **bufs,
 112                 uint16_t nb_pkts)
 113 {
 114         unsigned int i;
 115         struct rte_mbuf *mbuf;
 116         struct szedata2_rx_queue *sze_q = queue;
 117         struct rte_pktmbuf_pool_private *mbp_priv;
 118         uint16_t num_rx = 0;
 119         uint16_t buf_size;
 120         uint16_t sg_size;
 121         uint16_t hw_size;
 122         uint16_t packet_size;
 123         uint64_t num_bytes = 0;
 124         struct szedata *sze = sze_q->sze;
 125         uint8_t *header_ptr = NULL; /* header of packet */
 126         uint8_t *packet_ptr1 = NULL;
 127         uint8_t *packet_ptr2 = NULL;
 128         uint16_t packet_len1 = 0;
 129         uint16_t packet_len2 = 0;
 130         uint16_t hw_data_align;
 131
 132         if (unlikely(sze_q->sze == NULL || nb_pkts == 0))
 133                 return 0;
 134
 135         /*
 136          * Reads the given number of packets from szedata2 channel given
 137          * by queue and copies the packet data into a newly allocated mbuf
 138          * to return.
 139          */
 140         for (i = 0; i < nb_pkts; i++) {
 141                 mbuf = rte_pktmbuf_alloc(sze_q->mb_pool);
 142
 143                 if (unlikely(mbuf == NULL)) {
 144                         sze_q->priv->dev->data->rx_mbuf_alloc_failed++;
 145                         break;
 146                 }
 147
 148                 /* get the next sze packet */
 149                 if (sze->ct_rx_lck != NULL && !sze->ct_rx_rem_bytes &&
 150                                 sze->ct_rx_lck->next == NULL) {
 151                         /* unlock old data */
 152                         szedata_rx_unlock_data(sze_q->sze, sze->ct_rx_lck_orig);
 153                         sze->ct_rx_lck_orig = NULL;
 154                         sze->ct_rx_lck = NULL;
 155                 }
 156
 157                 if (!sze->ct_rx_rem_bytes && sze->ct_rx_lck_orig == NULL) {
 158                         /* nothing to read, lock new data */
 159                         sze->ct_rx_lck = szedata_rx_lock_data(sze_q->sze, ~0U);
 160                         sze->ct_rx_lck_orig = sze->ct_rx_lck;
 161
 162                         if (sze->ct_rx_lck == NULL) {
 163                                 /* nothing to lock */
 164                                 rte_pktmbuf_free(mbuf);
 165                                 break;
 166                         }
 167
 168                         sze->ct_rx_cur_ptr = sze->ct_rx_lck->start;
 169                         sze->ct_rx_rem_bytes = sze->ct_rx_lck->len;
 170
 171                         if (!sze->ct_rx_rem_bytes) {
 172                                 rte_pktmbuf_free(mbuf);
 173                                 break;
 174                         }
 175                 }
 176
 177                 if (sze->ct_rx_rem_bytes < RTE_SZE2_PACKET_HEADER_SIZE) {
 178                         /*
 179                          * cut in header
 180                          * copy parts of header to merge buffer
 181                          */
 182                         if (sze->ct_rx_lck->next == NULL) {
 183                                 rte_pktmbuf_free(mbuf);
 184                                 break;
 185                         }
 186
 187                         /* copy first part of header */
 188                         rte_memcpy(sze->ct_rx_buffer, sze->ct_rx_cur_ptr,
 189                                         sze->ct_rx_rem_bytes);
 190
 191                         /* copy second part of header */
 192                         sze->ct_rx_lck = sze->ct_rx_lck->next;
 193                         sze->ct_rx_cur_ptr = sze->ct_rx_lck->start;
 194                         rte_memcpy(sze->ct_rx_buffer + sze->ct_rx_rem_bytes,
 195                                 sze->ct_rx_cur_ptr,
 196                                 RTE_SZE2_PACKET_HEADER_SIZE -
 197                                 sze->ct_rx_rem_bytes);
 198
 199                         sze->ct_rx_cur_ptr += RTE_SZE2_PACKET_HEADER_SIZE -
 200                                 sze->ct_rx_rem_bytes;
 201                         sze->ct_rx_rem_bytes = sze->ct_rx_lck->len -
 202                                 RTE_SZE2_PACKET_HEADER_SIZE +
 203                                 sze->ct_rx_rem_bytes;
 204
 205                         header_ptr = (uint8_t *)sze->ct_rx_buffer;
 206                 } else {
 207                         /* not cut */
 208                         header_ptr = (uint8_t *)sze->ct_rx_cur_ptr;
 209                         sze->ct_rx_cur_ptr += RTE_SZE2_PACKET_HEADER_SIZE;
 210                         sze->ct_rx_rem_bytes -= RTE_SZE2_PACKET_HEADER_SIZE;
 211                 }
 212
 213                 sg_size = le16toh(*((uint16_t *)header_ptr));
 214                 hw_size = le16toh(*(((uint16_t *)header_ptr) + 1));
 215                 packet_size = sg_size -
 216                         RTE_SZE2_ALIGN8(RTE_SZE2_PACKET_HEADER_SIZE + hw_size);
 217
 218
 219                 /* checks if packet all right */
 220                 if (!sg_size)
 221                         errx(5, "Zero segsize");
 222
 223                 /* check sg_size and hwsize */
 224                 if (hw_size > sg_size - RTE_SZE2_PACKET_HEADER_SIZE) {
 225                         errx(10, "Hwsize bigger than expected. Segsize: %d, "
 226                                 "hwsize: %d", sg_size, hw_size);
 227                 }
 228
 229                 hw_data_align =
 230                         RTE_SZE2_ALIGN8(RTE_SZE2_PACKET_HEADER_SIZE + hw_size) -
 231                         RTE_SZE2_PACKET_HEADER_SIZE;
 232
 233                 if (sze->ct_rx_rem_bytes >=
 234                                 (uint16_t)(sg_size -
 235                                 RTE_SZE2_PACKET_HEADER_SIZE)) {
 236                         /* no cut */
 237                         /* one packet ready - go to another */
 238                         packet_ptr1 = sze->ct_rx_cur_ptr + hw_data_align;
 239                         packet_len1 = packet_size;
 240                         packet_ptr2 = NULL;
 241                         packet_len2 = 0;
 242
 243                         sze->ct_rx_cur_ptr += RTE_SZE2_ALIGN8(sg_size) -
 244                                 RTE_SZE2_PACKET_HEADER_SIZE;
 245                         sze->ct_rx_rem_bytes -= RTE_SZE2_ALIGN8(sg_size) -
 246                                 RTE_SZE2_PACKET_HEADER_SIZE;
 247                 } else {
 248                         /* cut in data */
 249                         if (sze->ct_rx_lck->next == NULL) {
 250                                 errx(6, "Need \"next\" lock, "
 251                                         "but it is missing: %u",
 252                                         sze->ct_rx_rem_bytes);
 253                         }
 254
 255                         /* skip hw data */
 256                         if (sze->ct_rx_rem_bytes <= hw_data_align) {
 257                                 uint16_t rem_size = hw_data_align -
 258                                         sze->ct_rx_rem_bytes;
 259
 260                                 /* MOVE to next lock */
 261                                 sze->ct_rx_lck = sze->ct_rx_lck->next;
 262                                 sze->ct_rx_cur_ptr =
 263                                         (void *)(((uint8_t *)
 264                                         (sze->ct_rx_lck->start)) + rem_size);
 265
 266                                 packet_ptr1 = sze->ct_rx_cur_ptr;
 267                                 packet_len1 = packet_size;
 268                                 packet_ptr2 = NULL;
 269                                 packet_len2 = 0;
 270
 271                                 sze->ct_rx_cur_ptr +=
 272                                         RTE_SZE2_ALIGN8(packet_size);
 273                                 sze->ct_rx_rem_bytes = sze->ct_rx_lck->len -
 274                                         rem_size - RTE_SZE2_ALIGN8(packet_size);
 275                         } else {
 276                                 /* get pointer and length from first part */
 277                                 packet_ptr1 = sze->ct_rx_cur_ptr +
 278                                         hw_data_align;
 279                                 packet_len1 = sze->ct_rx_rem_bytes -
 280                                         hw_data_align;
 281
 282                                 /* MOVE to next lock */
 283                                 sze->ct_rx_lck = sze->ct_rx_lck->next;
 284                                 sze->ct_rx_cur_ptr = sze->ct_rx_lck->start;
 285
 286                                 /* get pointer and length from second part */
 287                                 packet_ptr2 = sze->ct_rx_cur_ptr;
 288                                 packet_len2 = packet_size - packet_len1;
 289
 290                                 sze->ct_rx_cur_ptr +=
 291                                         RTE_SZE2_ALIGN8(packet_size) -
 292                                         packet_len1;
 293                                 sze->ct_rx_rem_bytes = sze->ct_rx_lck->len -
 294                                         (RTE_SZE2_ALIGN8(packet_size) -
 295                                          packet_len1);
 296                         }
 297                 }
 298
 299                 if (unlikely(packet_ptr1 == NULL)) {
 300                         rte_pktmbuf_free(mbuf);
 301                         break;
 302                 }
 303
 304                 /* get the space available for data in the mbuf */
 305                 mbp_priv = rte_mempool_get_priv(sze_q->mb_pool);
 306                 buf_size = (uint16_t)(mbp_priv->mbuf_data_room_size -
 307                                 RTE_PKTMBUF_HEADROOM);
 308
 309                 if (packet_size <= buf_size) {
 310                         /* sze packet will fit in one mbuf, go ahead and copy */
 311                         rte_memcpy(rte_pktmbuf_mtod(mbuf, void *),
 312                                         packet_ptr1, packet_len1);
 313                         if (packet_ptr2 != NULL) {
 314                                 rte_memcpy((void *)(rte_pktmbuf_mtod(mbuf,
 315                                         uint8_t *) + packet_len1),
 316                                         packet_ptr2, packet_len2);
 317                         }
 318                         mbuf->data_len = (uint16_t)packet_size;
 319
 320                         mbuf->pkt_len = packet_size;
 321                         mbuf->port = sze_q->in_port;
 322                         bufs[num_rx] = mbuf;
 323                         num_rx++;
 324                         num_bytes += packet_size;
 325                 } else {
 326                         /*
 327                          * sze packet will not fit in one mbuf,
 328                          * scattered mode is not enabled, drop packet
 329                          */
 330                         PMD_DRV_LOG(ERR,
 331                                 "SZE segment %d bytes will not fit in one mbuf "
 332                                 "(%d bytes), scattered mode is not enabled, "
 333                                 "drop packet!!",
 334                                 packet_size, buf_size);
 335                         rte_pktmbuf_free(mbuf);
 336                 }
 337         }
 338
 339         sze_q->rx_pkts += num_rx;
 340         sze_q->rx_bytes += num_bytes;
 341         return num_rx;
 342 }
 343
 344 static uint16_t
 345 eth_szedata2_rx_scattered(void *queue,
 346                 struct rte_mbuf **bufs,
 347                 uint16_t nb_pkts)
 348 {
 349         unsigned int i;
 350         struct rte_mbuf *mbuf;
 351         struct szedata2_rx_queue *sze_q = queue;
 352         struct rte_pktmbuf_pool_private *mbp_priv;
 353         uint16_t num_rx = 0;
 354         uint16_t buf_size;
 355         uint16_t sg_size;
 356         uint16_t hw_size;
 357         uint16_t packet_size;
 358         uint64_t num_bytes = 0;
 359         struct szedata *sze = sze_q->sze;
 360         uint8_t *header_ptr = NULL; /* header of packet */
 361         uint8_t *packet_ptr1 = NULL;
 362         uint8_t *packet_ptr2 = NULL;
 363         uint16_t packet_len1 = 0;
 364         uint16_t packet_len2 = 0;
 365         uint16_t hw_data_align;
 366         uint64_t *mbuf_failed_ptr =
 367                 &sze_q->priv->dev->data->rx_mbuf_alloc_failed;
 368
 369         if (unlikely(sze_q->sze == NULL || nb_pkts == 0))
 370                 return 0;
 371
 372         /*
 373          * Reads the given number of packets from szedata2 channel given
 374          * by queue and copies the packet data into a newly allocated mbuf
 375          * to return.
 376          */
 377         for (i = 0; i < nb_pkts; i++) {
 378                 const struct szedata_lock *ct_rx_lck_backup;
 379                 unsigned int ct_rx_rem_bytes_backup;
 380                 unsigned char *ct_rx_cur_ptr_backup;
 381
 382                 /* get the next sze packet */
 383                 if (sze->ct_rx_lck != NULL && !sze->ct_rx_rem_bytes &&
 384                                 sze->ct_rx_lck->next == NULL) {
 385                         /* unlock old data */
 386                         szedata_rx_unlock_data(sze_q->sze, sze->ct_rx_lck_orig);
 387                         sze->ct_rx_lck_orig = NULL;
 388                         sze->ct_rx_lck = NULL;
 389                 }
 390
 391                 /*
 392                  * Store items from sze structure which can be changed
 393                  * before mbuf allocating. Use these items in case of mbuf
 394                  * allocating failure.
 395                  */
 396                 ct_rx_lck_backup = sze->ct_rx_lck;
 397                 ct_rx_rem_bytes_backup = sze->ct_rx_rem_bytes;
 398                 ct_rx_cur_ptr_backup = sze->ct_rx_cur_ptr;
 399
 400                 if (!sze->ct_rx_rem_bytes && sze->ct_rx_lck_orig == NULL) {
 401                         /* nothing to read, lock new data */
 402                         sze->ct_rx_lck = szedata_rx_lock_data(sze_q->sze, ~0U);
 403                         sze->ct_rx_lck_orig = sze->ct_rx_lck;
 404
 405                         /*
 406                          * Backup items from sze structure must be updated
 407                          * after locking to contain pointers to new locks.
 408                          */
 409                         ct_rx_lck_backup = sze->ct_rx_lck;
 410                         ct_rx_rem_bytes_backup = sze->ct_rx_rem_bytes;
 411                         ct_rx_cur_ptr_backup = sze->ct_rx_cur_ptr;
 412
 413                         if (sze->ct_rx_lck == NULL)
 414                                 /* nothing to lock */
 415                                 break;
 416
 417                         sze->ct_rx_cur_ptr = sze->ct_rx_lck->start;
 418                         sze->ct_rx_rem_bytes = sze->ct_rx_lck->len;
 419
 420                         if (!sze->ct_rx_rem_bytes)
 421                                 break;
 422                 }
 423
 424                 if (sze->ct_rx_rem_bytes < RTE_SZE2_PACKET_HEADER_SIZE) {
 425                         /*
 426                          * cut in header - copy parts of header to merge buffer
 427                          */
 428                         if (sze->ct_rx_lck->next == NULL)
 429                                 break;
 430
 431                         /* copy first part of header */
 432                         rte_memcpy(sze->ct_rx_buffer, sze->ct_rx_cur_ptr,
 433                                         sze->ct_rx_rem_bytes);
 434
 435                         /* copy second part of header */
 436                         sze->ct_rx_lck = sze->ct_rx_lck->next;
 437                         sze->ct_rx_cur_ptr = sze->ct_rx_lck->start;
 438                         rte_memcpy(sze->ct_rx_buffer + sze->ct_rx_rem_bytes,
 439                                 sze->ct_rx_cur_ptr,
 440                                 RTE_SZE2_PACKET_HEADER_SIZE -
 441                                 sze->ct_rx_rem_bytes);
 442
 443                         sze->ct_rx_cur_ptr += RTE_SZE2_PACKET_HEADER_SIZE -
 444                                 sze->ct_rx_rem_bytes;
 445                         sze->ct_rx_rem_bytes = sze->ct_rx_lck->len -
 446                                 RTE_SZE2_PACKET_HEADER_SIZE +
 447                                 sze->ct_rx_rem_bytes;
 448
 449                         header_ptr = (uint8_t *)sze->ct_rx_buffer;
 450                 } else {
 451                         /* not cut */
 452                         header_ptr = (uint8_t *)sze->ct_rx_cur_ptr;
 453                         sze->ct_rx_cur_ptr += RTE_SZE2_PACKET_HEADER_SIZE;
 454                         sze->ct_rx_rem_bytes -= RTE_SZE2_PACKET_HEADER_SIZE;
 455                 }
 456
 457                 sg_size = le16toh(*((uint16_t *)header_ptr));
 458                 hw_size = le16toh(*(((uint16_t *)header_ptr) + 1));
 459                 packet_size = sg_size -
 460                         RTE_SZE2_ALIGN8(RTE_SZE2_PACKET_HEADER_SIZE + hw_size);
 461
 462
 463                 /* checks if packet all right */
 464                 if (!sg_size)
 465                         errx(5, "Zero segsize");
 466
 467                 /* check sg_size and hwsize */
 468                 if (hw_size > sg_size - RTE_SZE2_PACKET_HEADER_SIZE) {
 469                         errx(10, "Hwsize bigger than expected. Segsize: %d, "
 470                                         "hwsize: %d", sg_size, hw_size);
 471                 }
 472
 473                 hw_data_align =
 474                         RTE_SZE2_ALIGN8((RTE_SZE2_PACKET_HEADER_SIZE +
 475                         hw_size)) - RTE_SZE2_PACKET_HEADER_SIZE;
 476
 477                 if (sze->ct_rx_rem_bytes >=
 478                                 (uint16_t)(sg_size -
 479                                 RTE_SZE2_PACKET_HEADER_SIZE)) {
 480                         /* no cut */
 481                         /* one packet ready - go to another */
 482                         packet_ptr1 = sze->ct_rx_cur_ptr + hw_data_align;
 483                         packet_len1 = packet_size;
 484                         packet_ptr2 = NULL;
 485                         packet_len2 = 0;
 486
 487                         sze->ct_rx_cur_ptr += RTE_SZE2_ALIGN8(sg_size) -
 488                                 RTE_SZE2_PACKET_HEADER_SIZE;
 489                         sze->ct_rx_rem_bytes -= RTE_SZE2_ALIGN8(sg_size) -
 490                                 RTE_SZE2_PACKET_HEADER_SIZE;
 491                 } else {
 492                         /* cut in data */
 493                         if (sze->ct_rx_lck->next == NULL) {
 494                                 errx(6, "Need \"next\" lock, but it is "
 495                                         "missing: %u", sze->ct_rx_rem_bytes);
 496                         }
 497
 498                         /* skip hw data */
 499                         if (sze->ct_rx_rem_bytes <= hw_data_align) {
 500                                 uint16_t rem_size = hw_data_align -
 501                                         sze->ct_rx_rem_bytes;
 502
 503                                 /* MOVE to next lock */
 504                                 sze->ct_rx_lck = sze->ct_rx_lck->next;
 505                                 sze->ct_rx_cur_ptr =
 506                                         (void *)(((uint8_t *)
 507                                         (sze->ct_rx_lck->start)) + rem_size);
 508
 509                                 packet_ptr1 = sze->ct_rx_cur_ptr;
 510                                 packet_len1 = packet_size;
 511                                 packet_ptr2 = NULL;
 512                                 packet_len2 = 0;
 513
 514                                 sze->ct_rx_cur_ptr +=
 515                                         RTE_SZE2_ALIGN8(packet_size);
 516                                 sze->ct_rx_rem_bytes = sze->ct_rx_lck->len -
 517                                         rem_size - RTE_SZE2_ALIGN8(packet_size);
 518                         } else {
 519                                 /* get pointer and length from first part */
 520                                 packet_ptr1 = sze->ct_rx_cur_ptr +
 521                                         hw_data_align;
 522                                 packet_len1 = sze->ct_rx_rem_bytes -
 523                                         hw_data_align;
 524
 525                                 /* MOVE to next lock */
 526                                 sze->ct_rx_lck = sze->ct_rx_lck->next;
 527                                 sze->ct_rx_cur_ptr = sze->ct_rx_lck->start;
 528
 529                                 /* get pointer and length from second part */
 530                                 packet_ptr2 = sze->ct_rx_cur_ptr;
 531                                 packet_len2 = packet_size - packet_len1;
 532
 533                                 sze->ct_rx_cur_ptr +=
 534                                         RTE_SZE2_ALIGN8(packet_size) -
 535                                         packet_len1;
 536                                 sze->ct_rx_rem_bytes = sze->ct_rx_lck->len -
 537                                         (RTE_SZE2_ALIGN8(packet_size) -
 538                                          packet_len1);
 539                         }
 540                 }
 541
 542                 if (unlikely(packet_ptr1 == NULL))
 543                         break;
 544
 545                 mbuf = rte_pktmbuf_alloc(sze_q->mb_pool);
 546
 547                 if (unlikely(mbuf == NULL)) {
 548                         /*
 549                          * Restore items from sze structure to state after
 550                          * unlocking (eventually locking).
 551                          */
 552                         sze->ct_rx_lck = ct_rx_lck_backup;
 553                         sze->ct_rx_rem_bytes = ct_rx_rem_bytes_backup;
 554                         sze->ct_rx_cur_ptr = ct_rx_cur_ptr_backup;
 555                         sze_q->priv->dev->data->rx_mbuf_alloc_failed++;
 556                         break;
 557                 }
 558
 559                 /* get the space available for data in the mbuf */
 560                 mbp_priv = rte_mempool_get_priv(sze_q->mb_pool);
 561                 buf_size = (uint16_t)(mbp_priv->mbuf_data_room_size -
 562                                 RTE_PKTMBUF_HEADROOM);
 563
 564                 if (packet_size <= buf_size) {
 565                         /* sze packet will fit in one mbuf, go ahead and copy */
 566                         rte_memcpy(rte_pktmbuf_mtod(mbuf, void *),
 567                                         packet_ptr1, packet_len1);
 568                         if (packet_ptr2 != NULL) {
 569                                 rte_memcpy((void *)
 570                                         (rte_pktmbuf_mtod(mbuf, uint8_t *) +
 571                                         packet_len1), packet_ptr2, packet_len2);
 572                         }
 573                         mbuf->data_len = (uint16_t)packet_size;
 574                 } else {
 575                         /*
 576                          * sze packet will not fit in one mbuf,
 577                          * scatter packet into more mbufs
 578                          */
 579                         struct rte_mbuf *m = mbuf;
 580                         uint16_t len = rte_pktmbuf_tailroom(mbuf);
 581
 582                         /* copy first part of packet */
 583                         /* fill first mbuf */
 584                         rte_memcpy(rte_pktmbuf_append(mbuf, len), packet_ptr1,
 585                                 len);
 586                         packet_len1 -= len;
 587                         packet_ptr1 = ((uint8_t *)packet_ptr1) + len;
 588
 589                         while (packet_len1 > 0) {
 590                                 /* fill new mbufs */
 591                                 m->next = rte_pktmbuf_alloc(sze_q->mb_pool);
 592
 593                                 if (unlikely(m->next == NULL)) {
 594                                         rte_pktmbuf_free(mbuf);
 595                                         /*
 596                                          * Restore items from sze structure
 597                                          * to state after unlocking (eventually
 598                                          * locking).
 599                                          */
 600                                         sze->ct_rx_lck = ct_rx_lck_backup;
 601                                         sze->ct_rx_rem_bytes =
 602                                                 ct_rx_rem_bytes_backup;
 603                                         sze->ct_rx_cur_ptr =
 604                                                 ct_rx_cur_ptr_backup;
 605                                         (*mbuf_failed_ptr)++;
 606                                         goto finish;
 607                                 }
 608
 609                                 m = m->next;
 610
 611                                 len = RTE_MIN(rte_pktmbuf_tailroom(m),
 612                                         packet_len1);
 613                                 rte_memcpy(rte_pktmbuf_append(mbuf, len),
 614                                         packet_ptr1, len);
 615
 616                                 (mbuf->nb_segs)++;
 617                                 packet_len1 -= len;
 618                                 packet_ptr1 = ((uint8_t *)packet_ptr1) + len;
 619                         }
 620
 621                         if (packet_ptr2 != NULL) {
 622                                 /* copy second part of packet, if exists */
 623                                 /* fill the rest of currently last mbuf */
 624                                 len = rte_pktmbuf_tailroom(m);
 625                                 rte_memcpy(rte_pktmbuf_append(mbuf, len),
 626                                         packet_ptr2, len);
 627                                 packet_len2 -= len;
 628                                 packet_ptr2 = ((uint8_t *)packet_ptr2) + len;
 629
 630                                 while (packet_len2 > 0) {
 631                                         /* fill new mbufs */
 632                                         m->next = rte_pktmbuf_alloc(
 633                                                         sze_q->mb_pool);
 634
 635                                         if (unlikely(m->next == NULL)) {
 636                                                 rte_pktmbuf_free(mbuf);
 637                                                 /*
 638                                                  * Restore items from sze
 639                                                  * structure to state after
 640                                                  * unlocking (eventually
 641                                                  * locking).
 642                                                  */
 643                                                 sze->ct_rx_lck =
 644                                                         ct_rx_lck_backup;
 645                                                 sze->ct_rx_rem_bytes =
 646                                                         ct_rx_rem_bytes_backup;
 647                                                 sze->ct_rx_cur_ptr =
 648                                                         ct_rx_cur_ptr_backup;
 649                                                 (*mbuf_failed_ptr)++;
 650                                                 goto finish;
 651                                         }
 652
 653                                         m = m->next;
 654
 655                                         len = RTE_MIN(rte_pktmbuf_tailroom(m),
 656                                                 packet_len2);
 657                                         rte_memcpy(
 658                                                 rte_pktmbuf_append(mbuf, len),
 659                                                 packet_ptr2, len);
 660
 661                                         (mbuf->nb_segs)++;
 662                                         packet_len2 -= len;
 663                                         packet_ptr2 = ((uint8_t *)packet_ptr2) +
 664                                                 len;
 665                                 }
 666                         }
 667                 }
 668                 mbuf->pkt_len = packet_size;
 669                 mbuf->port = sze_q->in_port;
 670                 bufs[num_rx] = mbuf;
 671                 num_rx++;
 672                 num_bytes += packet_size;
 673         }
 674
 675 finish:
 676         sze_q->rx_pkts += num_rx;
 677         sze_q->rx_bytes += num_bytes;
 678         return num_rx;
 679 }
 680
 681 static uint16_t
 682 eth_szedata2_tx(void *queue,
 683                 struct rte_mbuf **bufs,
 684                 uint16_t nb_pkts)
 685 {
 686         struct rte_mbuf *mbuf;
 687         struct szedata2_tx_queue *sze_q = queue;
 688         uint16_t num_tx = 0;
 689         uint64_t num_bytes = 0;
 690
 691         const struct szedata_lock *lck;
 692         uint32_t lock_size;
 693         uint32_t lock_size2;
 694         void *dst;
 695         uint32_t pkt_len;
 696         uint32_t hwpkt_len;
 697         uint32_t unlock_size;
 698         uint32_t rem_len;
 699         uint16_t mbuf_segs;
 700         uint16_t pkt_left = nb_pkts;
 701
 702         if (sze_q->sze == NULL || nb_pkts == 0)
 703                 return 0;
 704
 705         while (pkt_left > 0) {
 706                 unlock_size = 0;
 707                 lck = szedata_tx_lock_data(sze_q->sze,
 708                         RTE_ETH_SZEDATA2_TX_LOCK_SIZE,
 709                         sze_q->tx_channel);
 710                 if (lck == NULL)
 711                         continue;
 712
 713                 dst = lck->start;
 714                 lock_size = lck->len;
 715                 lock_size2 = lck->next ? lck->next->len : 0;
 716
 717 next_packet:
 718                 mbuf = bufs[nb_pkts - pkt_left];
 719
 720                 pkt_len = mbuf->pkt_len;
 721                 mbuf_segs = mbuf->nb_segs;
 722
 723                 hwpkt_len = RTE_SZE2_PACKET_HEADER_SIZE_ALIGNED +
 724                         RTE_SZE2_ALIGN8(pkt_len);
 725
 726                 if (lock_size + lock_size2 < hwpkt_len) {
 727                         szedata_tx_unlock_data(sze_q->sze, lck, unlock_size);
 728                         continue;
 729                 }
 730
 731                 num_bytes += pkt_len;
 732
 733                 if (lock_size > hwpkt_len) {
 734                         void *tmp_dst;
 735
 736                         rem_len = 0;
 737
 738                         /* write packet length at first 2 bytes in 8B header */
 739                         *((uint16_t *)dst) = htole16(
 740                                         RTE_SZE2_PACKET_HEADER_SIZE_ALIGNED +
 741                                         pkt_len);
 742                         *(((uint16_t *)dst) + 1) = htole16(0);
 743
 744                         /* copy packet from mbuf */
 745                         tmp_dst = ((uint8_t *)(dst)) +
 746                                 RTE_SZE2_PACKET_HEADER_SIZE_ALIGNED;
 747                         if (mbuf_segs == 1) {
 748                                 /*
 749                                  * non-scattered packet,
 750                                  * transmit from one mbuf
 751                                  */
 752                                 rte_memcpy(tmp_dst,
 753                                         rte_pktmbuf_mtod(mbuf, const void *),
 754                                         pkt_len);
 755                         } else {
 756                                 /* scattered packet, transmit from more mbufs */
 757                                 struct rte_mbuf *m = mbuf;
 758                                 while (m) {
 759                                         rte_memcpy(tmp_dst,
 760                                                 rte_pktmbuf_mtod(m,
 761                                                 const void *),
 762                                                 m->data_len);
 763                                         tmp_dst = ((uint8_t *)(tmp_dst)) +
 764                                                 m->data_len;
 765                                         m = m->next;
 766                                 }
 767                         }
 768
 769
 770                         dst = ((uint8_t *)dst) + hwpkt_len;
 771                         unlock_size += hwpkt_len;
 772                         lock_size -= hwpkt_len;
 773
 774                         rte_pktmbuf_free(mbuf);
 775                         num_tx++;
 776                         pkt_left--;
 777                         if (pkt_left == 0) {
 778                                 szedata_tx_unlock_data(sze_q->sze, lck,
 779                                         unlock_size);
 780                                 break;
 781                         }
 782                         goto next_packet;
 783                 } else if (lock_size + lock_size2 >= hwpkt_len) {
 784                         void *tmp_dst;
 785                         uint16_t write_len;
 786
 787                         /* write packet length at first 2 bytes in 8B header */
 788                         *((uint16_t *)dst) =
 789                                 htole16(RTE_SZE2_PACKET_HEADER_SIZE_ALIGNED +
 790                                         pkt_len);
 791                         *(((uint16_t *)dst) + 1) = htole16(0);
 792
 793                         /*
 794                          * If the raw packet (pkt_len) is smaller than lock_size
 795                          * get the correct length for memcpy
 796                          */
 797                         write_len =
 798                                 pkt_len < lock_size -
 799                                 RTE_SZE2_PACKET_HEADER_SIZE_ALIGNED ?
 800                                 pkt_len :
 801                                 lock_size - RTE_SZE2_PACKET_HEADER_SIZE_ALIGNED;
 802
 803                         rem_len = hwpkt_len - lock_size;
 804
 805                         tmp_dst = ((uint8_t *)(dst)) +
 806                                 RTE_SZE2_PACKET_HEADER_SIZE_ALIGNED;
 807                         if (mbuf_segs == 1) {
 808                                 /*
 809                                  * non-scattered packet,
 810                                  * transmit from one mbuf
 811                                  */
 812                                 /* copy part of packet to first area */
 813                                 rte_memcpy(tmp_dst,
 814                                         rte_pktmbuf_mtod(mbuf, const void *),
 815                                         write_len);
 816
 817                                 if (lck->next)
 818                                         dst = lck->next->start;
 819
 820                                 /* copy part of packet to second area */
 821                                 rte_memcpy(dst,
 822                                         (const void *)(rte_pktmbuf_mtod(mbuf,
 823                                                         const uint8_t *) +
 824                                         write_len), pkt_len - write_len);
 825                         } else {
 826                                 /* scattered packet, transmit from more mbufs */
 827                                 struct rte_mbuf *m = mbuf;
 828                                 uint16_t written = 0;
 829                                 uint16_t to_write = 0;
 830                                 bool new_mbuf = true;
 831                                 uint16_t write_off = 0;
 832
 833                                 /* copy part of packet to first area */
 834                                 while (m && written < write_len) {
 835                                         to_write = RTE_MIN(m->data_len,
 836                                                         write_len - written);
 837                                         rte_memcpy(tmp_dst,
 838                                                 rte_pktmbuf_mtod(m,
 839                                                         const void *),
 840                                                 to_write);
 841
 842                                         tmp_dst = ((uint8_t *)(tmp_dst)) +
 843                                                 to_write;
 844                                         if (m->data_len <= write_len -
 845                                                         written) {
 846                                                 m = m->next;
 847                                                 new_mbuf = true;
 848                                         } else {
 849                                                 new_mbuf = false;
 850                                         }
 851                                         written += to_write;
 852                                 }
 853
 854                                 if (lck->next)
 855                                         dst = lck->next->start;
 856
 857                                 tmp_dst = dst;
 858                                 written = 0;
 859                                 write_off = new_mbuf ? 0 : to_write;
 860
 861                                 /* copy part of packet to second area */
 862                                 while (m && written < pkt_len - write_len) {
 863                                         rte_memcpy(tmp_dst, (const void *)
 864                                                 (rte_pktmbuf_mtod(m,
 865                                                 uint8_t *) + write_off),
 866                                                 m->data_len - write_off);
 867
 868                                         tmp_dst = ((uint8_t *)(tmp_dst)) +
 869                                                 (m->data_len - write_off);
 870                                         written += m->data_len - write_off;
 871                                         m = m->next;
 872                                         write_off = 0;
 873                                 }
 874                         }
 875
 876                         dst = ((uint8_t *)dst) + rem_len;
 877                         unlock_size += hwpkt_len;
 878                         lock_size = lock_size2 - rem_len;
 879                         lock_size2 = 0;
 880
 881                         rte_pktmbuf_free(mbuf);
 882                         num_tx++;
 883                 }
 884
 885                 szedata_tx_unlock_data(sze_q->sze, lck, unlock_size);
 886                 pkt_left--;
 887         }
 888
 889         sze_q->tx_pkts += num_tx;
 890         sze_q->err_pkts += nb_pkts - num_tx;
 891         sze_q->tx_bytes += num_bytes;
 892         return num_tx;
 893 }
 894
 895 static int
 896 eth_rx_queue_start(struct rte_eth_dev *dev, uint16_t rxq_id)
 897 {
 898         struct szedata2_rx_queue *rxq = dev->data->rx_queues[rxq_id];
 899         int ret;
 900         struct pmd_internals *internals = (struct pmd_internals *)
 901                 dev->data->dev_private;
 902
 903         if (rxq->sze == NULL) {
 904                 uint32_t rx = 1 << rxq->rx_channel;
 905                 uint32_t tx = 0;
 906                 rxq->sze = szedata_open(internals->sze_dev_path);
 907                 if (rxq->sze == NULL)
 908                         return -EINVAL;
 909                 ret = szedata_subscribe3(rxq->sze, &rx, &tx);
 910                 if (ret != 0 || rx == 0)
 911                         goto err;
 912         }
 913
 914         ret = szedata_start(rxq->sze);
 915         if (ret != 0)
 916                 goto err;
 917         dev->data->rx_queue_state[rxq_id] = RTE_ETH_QUEUE_STATE_STARTED;
 918         return 0;
 919
 920 err:
 921         szedata_close(rxq->sze);
 922         rxq->sze = NULL;
 923         return -EINVAL;
 924 }
 925
 926 static int
 927 eth_rx_queue_stop(struct rte_eth_dev *dev, uint16_t rxq_id)
 928 {
 929         struct szedata2_rx_queue *rxq = dev->data->rx_queues[rxq_id];
 930
 931         if (rxq->sze != NULL) {
 932                 szedata_close(rxq->sze);
 933                 rxq->sze = NULL;
 934         }
 935
 936         dev->data->rx_queue_state[rxq_id] = RTE_ETH_QUEUE_STATE_STOPPED;
 937         return 0;
 938 }
 939
 940 static int
 941 eth_tx_queue_start(struct rte_eth_dev *dev, uint16_t txq_id)
 942 {
 943         struct szedata2_tx_queue *txq = dev->data->tx_queues[txq_id];
 944         int ret;
 945         struct pmd_internals *internals = (struct pmd_internals *)
 946                 dev->data->dev_private;
 947
 948         if (txq->sze == NULL) {
 949                 uint32_t rx = 0;
 950                 uint32_t tx = 1 << txq->tx_channel;
 951                 txq->sze = szedata_open(internals->sze_dev_path);
 952                 if (txq->sze == NULL)
 953                         return -EINVAL;
 954                 ret = szedata_subscribe3(txq->sze, &rx, &tx);
 955                 if (ret != 0 || tx == 0)
 956                         goto err;
 957         }
 958
 959         ret = szedata_start(txq->sze);
 960         if (ret != 0)
 961                 goto err;
 962         dev->data->tx_queue_state[txq_id] = RTE_ETH_QUEUE_STATE_STARTED;
 963         return 0;
 964
 965 err:
 966         szedata_close(txq->sze);
 967         txq->sze = NULL;
 968         return -EINVAL;
 969 }
 970
 971 static int
 972 eth_tx_queue_stop(struct rte_eth_dev *dev, uint16_t txq_id)
 973 {
 974         struct szedata2_tx_queue *txq = dev->data->tx_queues[txq_id];
 975
 976         if (txq->sze != NULL) {
 977                 szedata_close(txq->sze);
 978                 txq->sze = NULL;
 979         }
 980
 981         dev->data->tx_queue_state[txq_id] = RTE_ETH_QUEUE_STATE_STOPPED;
 982         return 0;
 983 }
 984
 985 static int
 986 eth_dev_start(struct rte_eth_dev *dev)
 987 {
 988         int ret;
 989         uint16_t i;
 990         uint16_t nb_rx = dev->data->nb_rx_queues;
 991         uint16_t nb_tx = dev->data->nb_tx_queues;
 992
 993         for (i = 0; i < nb_rx; i++) {
 994                 ret = eth_rx_queue_start(dev, i);
 995                 if (ret != 0)
 996                         goto err_rx;
 997         }
 998
 999         for (i = 0; i < nb_tx; i++) {
1000                 ret = eth_tx_queue_start(dev, i);
1001                 if (ret != 0)
1002                         goto err_tx;
1003         }
1004
1005         return 0;
1006
1007 err_tx:
1008         for (i = 0; i < nb_tx; i++)
1009                 eth_tx_queue_stop(dev, i);
1010 err_rx:
1011         for (i = 0; i < nb_rx; i++)
1012                 eth_rx_queue_stop(dev, i);
1013         return ret;
1014 }
1015
1016 static int
1017 eth_dev_stop(struct rte_eth_dev *dev)
1018 {
1019         uint16_t i;
1020         uint16_t nb_rx = dev->data->nb_rx_queues;
1021         uint16_t nb_tx = dev->data->nb_tx_queues;
1022         int ret;
1023
1024         dev->data->dev_started = 0;
1025
1026         for (i = 0; i < nb_tx; i++) {
1027                 ret = eth_tx_queue_stop(dev, i);
1028                 if (ret != 0)
1029                         return ret;
1030         }
1031
1032         for (i = 0; i < nb_rx; i++) {
1033                 ret = eth_rx_queue_stop(dev, i);
1034                 if (ret != 0)
1035                         return ret;
1036         }
1037
1038         return 0;
1039 }
1040
1041 static int
1042 eth_dev_configure(struct rte_eth_dev *dev)
1043 {
1044         struct rte_eth_dev_data *data = dev->data;
1045         if (data->dev_conf.rxmode.offloads & DEV_RX_OFFLOAD_SCATTER) {
1046                 dev->rx_pkt_burst = eth_szedata2_rx_scattered;
1047                 data->scattered_rx = 1;
1048         } else {
1049                 dev->rx_pkt_burst = eth_szedata2_rx;
1050                 data->scattered_rx = 0;
1051         }
1052         return 0;
1053 }
1054
1055 static int
1056 eth_dev_info(struct rte_eth_dev *dev,
1057                 struct rte_eth_dev_info *dev_info)
1058 {
1059         struct pmd_internals *internals = dev->data->dev_private;
1060
1061         dev_info->if_index = 0;
1062         dev_info->max_mac_addrs = 1;
1063         dev_info->max_rx_pktlen = (uint32_t)-1;
1064         dev_info->max_rx_queues = internals->max_rx_queues;
1065         dev_info->max_tx_queues = internals->max_tx_queues;
1066         dev_info->min_rx_bufsize = 0;
1067         dev_info->rx_offload_capa = DEV_RX_OFFLOAD_SCATTER;
1068         dev_info->tx_offload_capa = 0;
1069         dev_info->rx_queue_offload_capa = 0;
1070         dev_info->tx_queue_offload_capa = 0;
1071         dev_info->speed_capa = ETH_LINK_SPEED_100G;
1072
1073         return 0;
1074 }
1075
1076 static int
1077 eth_stats_get(struct rte_eth_dev *dev,
1078                 struct rte_eth_stats *stats)
1079 {
1080         uint16_t i;
1081         uint16_t nb_rx = dev->data->nb_rx_queues;
1082         uint16_t nb_tx = dev->data->nb_tx_queues;
1083         uint64_t rx_total = 0;
1084         uint64_t tx_total = 0;
1085         uint64_t tx_err_total = 0;
1086         uint64_t rx_total_bytes = 0;
1087         uint64_t tx_total_bytes = 0;
1088
1089         for (i = 0; i < nb_rx; i++) {
1090                 struct szedata2_rx_queue *rxq = dev->data->rx_queues[i];
1091
1092                 if (i < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
1093                         stats->q_ipackets[i] = rxq->rx_pkts;
1094                         stats->q_ibytes[i] = rxq->rx_bytes;
1095                 }
1096                 rx_total += rxq->rx_pkts;
1097                 rx_total_bytes += rxq->rx_bytes;
1098         }
1099
1100         for (i = 0; i < nb_tx; i++) {
1101                 struct szedata2_tx_queue *txq = dev->data->tx_queues[i];
1102
1103                 if (i < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
1104                         stats->q_opackets[i] = txq->tx_pkts;
1105                         stats->q_obytes[i] = txq->tx_bytes;
1106                 }
1107                 tx_total += txq->tx_pkts;
1108                 tx_total_bytes += txq->tx_bytes;
1109                 tx_err_total += txq->err_pkts;
1110         }
1111
1112         stats->ipackets = rx_total;
1113         stats->opackets = tx_total;
1114         stats->ibytes = rx_total_bytes;
1115         stats->obytes = tx_total_bytes;
1116         stats->oerrors = tx_err_total;
1117         stats->rx_nombuf = dev->data->rx_mbuf_alloc_failed;
1118
1119         return 0;
1120 }
1121
1122 static int
1123 eth_stats_reset(struct rte_eth_dev *dev)
1124 {
1125         uint16_t i;
1126         uint16_t nb_rx = dev->data->nb_rx_queues;
1127         uint16_t nb_tx = dev->data->nb_tx_queues;
1128
1129         for (i = 0; i < nb_rx; i++) {
1130                 struct szedata2_rx_queue *rxq = dev->data->rx_queues[i];
1131                 rxq->rx_pkts = 0;
1132                 rxq->rx_bytes = 0;
1133                 rxq->err_pkts = 0;
1134         }
1135         for (i = 0; i < nb_tx; i++) {
1136                 struct szedata2_tx_queue *txq = dev->data->tx_queues[i];
1137                 txq->tx_pkts = 0;
1138                 txq->tx_bytes = 0;
1139                 txq->err_pkts = 0;
1140         }
1141
1142         return 0;
1143 }
1144
1145 static void
1146 eth_rx_queue_release(void *q)
1147 {
1148         struct szedata2_rx_queue *rxq = (struct szedata2_rx_queue *)q;
1149
1150         if (rxq != NULL) {
1151                 if (rxq->sze != NULL)
1152                         szedata_close(rxq->sze);
1153                 rte_free(rxq);
1154         }
1155 }
1156
1157 static void
1158 eth_tx_queue_release(void *q)
1159 {
1160         struct szedata2_tx_queue *txq = (struct szedata2_tx_queue *)q;
1161
1162         if (txq != NULL) {
1163                 if (txq->sze != NULL)
1164                         szedata_close(txq->sze);
1165                 rte_free(txq);
1166         }
1167 }
1168
1169 static int
1170 eth_dev_close(struct rte_eth_dev *dev)
1171 {
1172         struct pmd_internals *internals = dev->data->dev_private;
1173         uint16_t i;
1174         uint16_t nb_rx = dev->data->nb_rx_queues;
1175         uint16_t nb_tx = dev->data->nb_tx_queues;
1176         int ret;
1177
1178         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1179                 return 0;
1180
1181         ret = eth_dev_stop(dev);
1182
1183         free(internals->sze_dev_path);
1184
1185         for (i = 0; i < nb_rx; i++) {
1186                 eth_rx_queue_release(dev->data->rx_queues[i]);
1187                 dev->data->rx_queues[i] = NULL;
1188         }
1189         dev->data->nb_rx_queues = 0;
1190         for (i = 0; i < nb_tx; i++) {
1191                 eth_tx_queue_release(dev->data->tx_queues[i]);
1192                 dev->data->tx_queues[i] = NULL;
1193         }
1194         dev->data->nb_tx_queues = 0;
1195
1196         return ret;
1197 }
1198
1199 static int
1200 eth_link_update(struct rte_eth_dev *dev,
1201                 int wait_to_complete __rte_unused)
1202 {
1203         struct rte_eth_link link;
1204
1205         memset(&link, 0, sizeof(link));
1206
1207         link.link_speed = ETH_SPEED_NUM_100G;
1208         link.link_duplex = ETH_LINK_FULL_DUPLEX;
1209         link.link_status = ETH_LINK_UP;
1210         link.link_autoneg = ETH_LINK_FIXED;
1211
1212         rte_eth_linkstatus_set(dev, &link);
1213         return 0;
1214 }
1215
1216 static int
1217 eth_dev_set_link_up(struct rte_eth_dev *dev __rte_unused)
1218 {
1219         PMD_DRV_LOG(WARNING, "Setting link up is not supported.");
1220         return 0;
1221 }
1222
1223 static int
1224 eth_dev_set_link_down(struct rte_eth_dev *dev __rte_unused)
1225 {
1226         PMD_DRV_LOG(WARNING, "Setting link down is not supported.");
1227         return 0;
1228 }
1229
1230 static int
1231 eth_rx_queue_setup(struct rte_eth_dev *dev,
1232                 uint16_t rx_queue_id,
1233                 uint16_t nb_rx_desc __rte_unused,
1234                 unsigned int socket_id,
1235                 const struct rte_eth_rxconf *rx_conf __rte_unused,
1236                 struct rte_mempool *mb_pool)
1237 {
1238         struct szedata2_rx_queue *rxq;
1239         int ret;
1240         struct pmd_internals *internals = dev->data->dev_private;
1241         uint8_t rx_channel = internals->rxq_base_id + rx_queue_id;
1242         uint32_t rx = 1 << rx_channel;
1243         uint32_t tx = 0;
1244
1245         PMD_INIT_FUNC_TRACE();
1246
1247         if (dev->data->rx_queues[rx_queue_id] != NULL) {
1248                 eth_rx_queue_release(dev->data->rx_queues[rx_queue_id]);
1249                 dev->data->rx_queues[rx_queue_id] = NULL;
1250         }
1251
1252         rxq = rte_zmalloc_socket("szedata2 rx queue",
1253                         sizeof(struct szedata2_rx_queue),
1254                         RTE_CACHE_LINE_SIZE, socket_id);
1255         if (rxq == NULL) {
1256                 PMD_INIT_LOG(ERR, "rte_zmalloc_socket() failed for rx queue id "
1257                                 "%" PRIu16 "!", rx_queue_id);
1258                 return -ENOMEM;
1259         }
1260
1261         rxq->priv = internals;
1262         rxq->sze = szedata_open(internals->sze_dev_path);
1263         if (rxq->sze == NULL) {
1264                 PMD_INIT_LOG(ERR, "szedata_open() failed for rx queue id "
1265                                 "%" PRIu16 "!", rx_queue_id);
1266                 eth_rx_queue_release(rxq);
1267                 return -EINVAL;
1268         }
1269         ret = szedata_subscribe3(rxq->sze, &rx, &tx);
1270         if (ret != 0 || rx == 0) {
1271                 PMD_INIT_LOG(ERR, "szedata_subscribe3() failed for rx queue id "
1272                                 "%" PRIu16 "!", rx_queue_id);
1273                 eth_rx_queue_release(rxq);
1274                 return -EINVAL;
1275         }
1276         rxq->rx_channel = rx_channel;
1277         rxq->qid = rx_queue_id;
1278         rxq->in_port = dev->data->port_id;
1279         rxq->mb_pool = mb_pool;
1280         rxq->rx_pkts = 0;
1281         rxq->rx_bytes = 0;
1282         rxq->err_pkts = 0;
1283
1284         dev->data->rx_queues[rx_queue_id] = rxq;
1285
1286         PMD_INIT_LOG(DEBUG, "Configured rx queue id %" PRIu16 " on socket "
1287                         "%u (channel id %u).", rxq->qid, socket_id,
1288                         rxq->rx_channel);
1289
1290         return 0;
1291 }
1292
1293 static int
1294 eth_tx_queue_setup(struct rte_eth_dev *dev,
1295                 uint16_t tx_queue_id,
1296                 uint16_t nb_tx_desc __rte_unused,
1297                 unsigned int socket_id,
1298                 const struct rte_eth_txconf *tx_conf __rte_unused)
1299 {
1300         struct szedata2_tx_queue *txq;
1301         int ret;
1302         struct pmd_internals *internals = dev->data->dev_private;
1303         uint8_t tx_channel = internals->txq_base_id + tx_queue_id;
1304         uint32_t rx = 0;
1305         uint32_t tx = 1 << tx_channel;
1306
1307         PMD_INIT_FUNC_TRACE();
1308
1309         if (dev->data->tx_queues[tx_queue_id] != NULL) {
1310                 eth_tx_queue_release(dev->data->tx_queues[tx_queue_id]);
1311                 dev->data->tx_queues[tx_queue_id] = NULL;
1312         }
1313
1314         txq = rte_zmalloc_socket("szedata2 tx queue",
1315                         sizeof(struct szedata2_tx_queue),
1316                         RTE_CACHE_LINE_SIZE, socket_id);
1317         if (txq == NULL) {
1318                 PMD_INIT_LOG(ERR, "rte_zmalloc_socket() failed for tx queue id "
1319                                 "%" PRIu16 "!", tx_queue_id);
1320                 return -ENOMEM;
1321         }
1322
1323         txq->priv = internals;
1324         txq->sze = szedata_open(internals->sze_dev_path);
1325         if (txq->sze == NULL) {
1326                 PMD_INIT_LOG(ERR, "szedata_open() failed for tx queue id "
1327                                 "%" PRIu16 "!", tx_queue_id);
1328                 eth_tx_queue_release(txq);
1329                 return -EINVAL;
1330         }
1331         ret = szedata_subscribe3(txq->sze, &rx, &tx);
1332         if (ret != 0 || tx == 0) {
1333                 PMD_INIT_LOG(ERR, "szedata_subscribe3() failed for tx queue id "
1334                                 "%" PRIu16 "!", tx_queue_id);
1335                 eth_tx_queue_release(txq);
1336                 return -EINVAL;
1337         }
1338         txq->tx_channel = tx_channel;
1339         txq->qid = tx_queue_id;
1340         txq->tx_pkts = 0;
1341         txq->tx_bytes = 0;
1342         txq->err_pkts = 0;
1343
1344         dev->data->tx_queues[tx_queue_id] = txq;
1345
1346         PMD_INIT_LOG(DEBUG, "Configured tx queue id %" PRIu16 " on socket "
1347                         "%u (channel id %u).", txq->qid, socket_id,
1348                         txq->tx_channel);
1349
1350         return 0;
1351 }
1352
1353 static int
1354 eth_mac_addr_set(struct rte_eth_dev *dev __rte_unused,
1355                 struct rte_ether_addr *mac_addr __rte_unused)
1356 {
1357         return 0;
1358 }
1359
1360 static int
1361 eth_promiscuous_enable(struct rte_eth_dev *dev __rte_unused)
1362 {
1363         PMD_DRV_LOG(WARNING, "Enabling promiscuous mode is not supported. "
1364                         "The card is always in promiscuous mode.");
1365         return 0;
1366 }
1367
1368 static int
1369 eth_promiscuous_disable(struct rte_eth_dev *dev __rte_unused)
1370 {
1371         PMD_DRV_LOG(WARNING, "Disabling promiscuous mode is not supported. "
1372                         "The card is always in promiscuous mode.");
1373         return -ENOTSUP;
1374 }
1375
1376 static int
1377 eth_allmulticast_enable(struct rte_eth_dev *dev __rte_unused)
1378 {
1379         PMD_DRV_LOG(WARNING, "Enabling allmulticast mode is not supported.");
1380         return -ENOTSUP;
1381 }
1382
1383 static int
1384 eth_allmulticast_disable(struct rte_eth_dev *dev __rte_unused)
1385 {
1386         PMD_DRV_LOG(WARNING, "Disabling allmulticast mode is not supported.");
1387         return -ENOTSUP;
1388 }
1389
1390 static const struct eth_dev_ops ops = {
1391         .dev_start          = eth_dev_start,
1392         .dev_stop           = eth_dev_stop,
1393         .dev_set_link_up    = eth_dev_set_link_up,
1394         .dev_set_link_down  = eth_dev_set_link_down,
1395         .dev_close          = eth_dev_close,
1396         .dev_configure      = eth_dev_configure,
1397         .dev_infos_get      = eth_dev_info,
1398         .promiscuous_enable   = eth_promiscuous_enable,
1399         .promiscuous_disable  = eth_promiscuous_disable,
1400         .allmulticast_enable  = eth_allmulticast_enable,
1401         .allmulticast_disable = eth_allmulticast_disable,
1402         .rx_queue_start     = eth_rx_queue_start,
1403         .rx_queue_stop      = eth_rx_queue_stop,
1404         .tx_queue_start     = eth_tx_queue_start,
1405         .tx_queue_stop      = eth_tx_queue_stop,
1406         .rx_queue_setup     = eth_rx_queue_setup,
1407         .tx_queue_setup     = eth_tx_queue_setup,
1408         .rx_queue_release   = eth_rx_queue_release,
1409         .tx_queue_release   = eth_tx_queue_release,
1410         .link_update        = eth_link_update,
1411         .stats_get          = eth_stats_get,
1412         .stats_reset        = eth_stats_reset,
1413         .mac_addr_set       = eth_mac_addr_set,
1414 };
1415
1416 /*
1417  * This function goes through sysfs and looks for an index of szedata2
1418  * device file (/dev/szedataIIX, where X is the index).
1419  *
1420  * @return
1421  *           0 on success
1422  *          -1 on error
1423  */
1424 static int
1425 get_szedata2_index(const struct rte_pci_addr *pcislot_addr, uint32_t *index)
1426 {
1427         DIR *dir;
1428         struct dirent *entry;
1429         int ret;
1430         uint32_t tmp_index;
1431         FILE *fd;
1432         char pcislot_path[PATH_MAX];
1433         uint32_t domain;
1434         uint8_t bus;
1435         uint8_t devid;
1436         uint8_t function;
1437
1438         dir = opendir("/sys/class/combo");
1439         if (dir == NULL)
1440                 return -1;
1441
1442         /*
1443          * Iterate through all combosixX directories.
1444          * When the value in /sys/class/combo/combosixX/device/pcislot
1445          * file is the location of the ethernet device dev, "X" is the
1446          * index of the device.
1447          */
1448         while ((entry = readdir(dir)) != NULL) {
1449                 ret = sscanf(entry->d_name, "combosix%u", &tmp_index);
1450                 if (ret != 1)
1451                         continue;
1452
1453                 snprintf(pcislot_path, PATH_MAX,
1454                         "/sys/class/combo/combosix%u/device/pcislot",
1455                         tmp_index);
1456
1457                 fd = fopen(pcislot_path, "r");
1458                 if (fd == NULL)
1459                         continue;
1460
1461                 ret = fscanf(fd, "%8" SCNx32 ":%2" SCNx8 ":%2" SCNx8 ".%" SCNx8,
1462                                 &domain, &bus, &devid, &function);
1463                 fclose(fd);
1464                 if (ret != 4)
1465                         continue;
1466
1467                 if (pcislot_addr->domain == domain &&
1468                                 pcislot_addr->bus == bus &&
1469                                 pcislot_addr->devid == devid &&
1470                                 pcislot_addr->function == function) {
1471                         *index = tmp_index;
1472                         closedir(dir);
1473                         return 0;
1474                 }
1475         }
1476
1477         closedir(dir);
1478         return -1;
1479 }
1480
1481 /**
1482  * @brief Initializes rte_eth_dev device.
1483  * @param dev Device to initialize.
1484  * @param pi Structure with info about DMA queues.
1485  * @return 0 on success, negative error code on error.
1486  */
1487 static int
1488 rte_szedata2_eth_dev_init(struct rte_eth_dev *dev, struct port_info *pi)
1489 {
1490         int ret;
1491         uint32_t szedata2_index;
1492         char name[PATH_MAX];
1493         struct rte_eth_dev_data *data = dev->data;
1494         struct pmd_internals *internals = (struct pmd_internals *)
1495                 data->dev_private;
1496         struct rte_pci_device *pci_dev = RTE_ETH_DEV_TO_PCI(dev);
1497
1498         PMD_INIT_FUNC_TRACE();
1499
1500         PMD_INIT_LOG(INFO, "Initializing eth_dev %s (driver %s)", data->name,
1501                         RTE_STR(RTE_SZEDATA2_DRIVER_NAME));
1502
1503         /* Fill internal private structure. */
1504         internals->dev = dev;
1505         /* Get index of szedata2 device file and create path to device file */
1506         ret = get_szedata2_index(&pci_dev->addr, &szedata2_index);
1507         if (ret != 0) {
1508                 PMD_INIT_LOG(ERR, "Failed to get szedata2 device index!");
1509                 return -ENODEV;
1510         }
1511         snprintf(name, PATH_MAX, SZEDATA2_DEV_PATH_FMT, szedata2_index);
1512         internals->sze_dev_path = strdup(name);
1513         if (internals->sze_dev_path == NULL) {
1514                 PMD_INIT_LOG(ERR, "strdup() failed!");
1515                 return -ENOMEM;
1516         }
1517         PMD_INIT_LOG(INFO, "SZEDATA2 path: %s", internals->sze_dev_path);
1518         internals->max_rx_queues = pi->rx_count;
1519         internals->max_tx_queues = pi->tx_count;
1520         internals->rxq_base_id = pi->rx_base_id;
1521         internals->txq_base_id = pi->tx_base_id;
1522         PMD_INIT_LOG(INFO, "%u RX DMA channels from id %u",
1523                         internals->max_rx_queues, internals->rxq_base_id);
1524         PMD_INIT_LOG(INFO, "%u TX DMA channels from id %u",
1525                         internals->max_tx_queues, internals->txq_base_id);
1526
1527         /* Set rx, tx burst functions */
1528         if (data->scattered_rx == 1)
1529                 dev->rx_pkt_burst = eth_szedata2_rx_scattered;
1530         else
1531                 dev->rx_pkt_burst = eth_szedata2_rx;
1532         dev->tx_pkt_burst = eth_szedata2_tx;
1533
1534         /* Set function callbacks for Ethernet API */
1535         dev->dev_ops = &ops;
1536
1537         /* Get link state */
1538         eth_link_update(dev, 0);
1539
1540         /* Allocate space for one mac address */
1541         data->mac_addrs = rte_zmalloc(data->name, sizeof(struct rte_ether_addr),
1542                         RTE_CACHE_LINE_SIZE);
1543         if (data->mac_addrs == NULL) {
1544                 PMD_INIT_LOG(ERR, "Could not alloc space for MAC address!");
1545                 free(internals->sze_dev_path);
1546                 return -ENOMEM;
1547         }
1548
1549         rte_ether_addr_copy(&eth_addr, data->mac_addrs);
1550
1551         PMD_INIT_LOG(INFO, "%s device %s successfully initialized",
1552                         RTE_STR(RTE_SZEDATA2_DRIVER_NAME), data->name);
1553
1554         return 0;
1555 }
1556
1557 /**
1558  * @brief Unitializes rte_eth_dev device.
1559  * @param dev Device to uninitialize.
1560  * @return 0 on success, negative error code on error.
1561  */
1562 static int
1563 rte_szedata2_eth_dev_uninit(struct rte_eth_dev *dev)
1564 {
1565         PMD_INIT_FUNC_TRACE();
1566
1567         eth_dev_close(dev);
1568
1569         PMD_DRV_LOG(INFO, "%s device %s successfully uninitialized",
1570                         RTE_STR(RTE_SZEDATA2_DRIVER_NAME), dev->data->name);
1571
1572         return 0;
1573 }
1574
1575 static const struct rte_pci_id rte_szedata2_pci_id_table[] = {
1576         {
1577                 RTE_PCI_DEVICE(PCI_VENDOR_ID_NETCOPE,
1578                                 PCI_DEVICE_ID_NETCOPE_COMBO80G)
1579         },
1580         {
1581                 RTE_PCI_DEVICE(PCI_VENDOR_ID_NETCOPE,
1582                                 PCI_DEVICE_ID_NETCOPE_COMBO100G)
1583         },
1584         {
1585                 RTE_PCI_DEVICE(PCI_VENDOR_ID_NETCOPE,
1586                                 PCI_DEVICE_ID_NETCOPE_COMBO100G2)
1587         },
1588         {
1589                 RTE_PCI_DEVICE(PCI_VENDOR_ID_NETCOPE,
1590                                 PCI_DEVICE_ID_NETCOPE_NFB200G2QL)
1591         },
1592         {
1593                 RTE_PCI_DEVICE(PCI_VENDOR_ID_SILICOM,
1594                                 PCI_DEVICE_ID_FB2CGG3)
1595         },
1596         {
1597                 RTE_PCI_DEVICE(PCI_VENDOR_ID_SILICOM,
1598                                 PCI_DEVICE_ID_FB2CGG3D)
1599         },
1600         {
1601                 .vendor_id = 0,
1602         }
1603 };
1604
1605 /**
1606  * @brief Gets info about DMA queues for ports.
1607  * @param pci_dev PCI device structure.
1608  * @param port_count Pointer to variable set with number of ports.
1609  * @param pi Pointer to array of structures with info about DMA queues
1610  *           for ports.
1611  * @param max_ports Maximum number of ports.
1612  * @return 0 on success, negative error code on error.
1613  */
1614 static int
1615 get_port_info(struct rte_pci_device *pci_dev, unsigned int *port_count,
1616                 struct port_info *pi, unsigned int max_ports)
1617 {
1618         struct szedata *szedata_temp;
1619         char sze_dev_path[PATH_MAX];
1620         uint32_t szedata2_index;
1621         int ret;
1622         uint16_t max_rx_queues;
1623         uint16_t max_tx_queues;
1624
1625         if (max_ports == 0)
1626                 return -EINVAL;
1627
1628         memset(pi, 0, max_ports * sizeof(struct port_info));
1629         *port_count = 0;
1630
1631         /* Get index of szedata2 device file and create path to device file */
1632         ret = get_szedata2_index(&pci_dev->addr, &szedata2_index);
1633         if (ret != 0) {
1634                 PMD_INIT_LOG(ERR, "Failed to get szedata2 device index!");
1635                 return -ENODEV;
1636         }
1637         snprintf(sze_dev_path, PATH_MAX, SZEDATA2_DEV_PATH_FMT, szedata2_index);
1638
1639         /*
1640          * Get number of available DMA RX and TX channels, which is maximum
1641          * number of queues that can be created.
1642          */
1643         szedata_temp = szedata_open(sze_dev_path);
1644         if (szedata_temp == NULL) {
1645                 PMD_INIT_LOG(ERR, "szedata_open(%s) failed", sze_dev_path);
1646                 return -EINVAL;
1647         }
1648         max_rx_queues = szedata_ifaces_available(szedata_temp, SZE2_DIR_RX);
1649         max_tx_queues = szedata_ifaces_available(szedata_temp, SZE2_DIR_TX);
1650         PMD_INIT_LOG(INFO, "Available DMA channels RX: %u TX: %u",
1651                         max_rx_queues, max_tx_queues);
1652         if (max_rx_queues > RTE_ETH_SZEDATA2_MAX_RX_QUEUES) {
1653                 PMD_INIT_LOG(ERR, "%u RX queues exceeds supported number %u",
1654                                 max_rx_queues, RTE_ETH_SZEDATA2_MAX_RX_QUEUES);
1655                 szedata_close(szedata_temp);
1656                 return -EINVAL;
1657         }
1658         if (max_tx_queues > RTE_ETH_SZEDATA2_MAX_TX_QUEUES) {
1659                 PMD_INIT_LOG(ERR, "%u TX queues exceeds supported number %u",
1660                                 max_tx_queues, RTE_ETH_SZEDATA2_MAX_TX_QUEUES);
1661                 szedata_close(szedata_temp);
1662                 return -EINVAL;
1663         }
1664
1665         if (pci_dev->id.device_id == PCI_DEVICE_ID_NETCOPE_NFB200G2QL) {
1666                 unsigned int i;
1667                 unsigned int rx_queues = max_rx_queues / max_ports;
1668                 unsigned int tx_queues = max_tx_queues / max_ports;
1669
1670                 /*
1671                  * Number of queues reported by szedata_ifaces_available()
1672                  * is the number of all queues from all DMA controllers which
1673                  * may reside at different numa locations.
1674                  * All queues from the same DMA controller have the same numa
1675                  * node.
1676                  * Numa node from the first queue of each DMA controller is
1677                  * retrieved.
1678                  * If the numa node differs from the numa node of the queues
1679                  * from the previous DMA controller the queues are assigned
1680                  * to the next port.
1681                  */
1682
1683                 for (i = 0; i < max_ports; i++) {
1684                         int numa_rx = szedata_get_area_numa_node(szedata_temp,
1685                                 SZE2_DIR_RX, rx_queues * i);
1686                         int numa_tx = szedata_get_area_numa_node(szedata_temp,
1687                                 SZE2_DIR_TX, tx_queues * i);
1688                         unsigned int port_rx_queues = numa_rx != -1 ?
1689                                 rx_queues : 0;
1690                         unsigned int port_tx_queues = numa_tx != -1 ?
1691                                 tx_queues : 0;
1692                         PMD_INIT_LOG(DEBUG, "%u rx queues from id %u, numa %d",
1693                                         rx_queues, rx_queues * i, numa_rx);
1694                         PMD_INIT_LOG(DEBUG, "%u tx queues from id %u, numa %d",
1695                                         tx_queues, tx_queues * i, numa_tx);
1696
1697                         if (port_rx_queues != 0 && port_tx_queues != 0 &&
1698                                         numa_rx != numa_tx) {
1699                                 PMD_INIT_LOG(ERR, "RX queue %u numa %d differs "
1700                                                 "from TX queue %u numa %d "
1701                                                 "unexpectedly",
1702                                                 rx_queues * i, numa_rx,
1703                                                 tx_queues * i, numa_tx);
1704                                 szedata_close(szedata_temp);
1705                                 return -EINVAL;
1706                         } else if (port_rx_queues == 0 && port_tx_queues == 0) {
1707                                 continue;
1708                         } else {
1709                                 unsigned int j;
1710                                 unsigned int current = *port_count;
1711                                 int port_numa = port_rx_queues != 0 ?
1712                                         numa_rx : numa_tx;
1713
1714                                 for (j = 0; j < *port_count; j++) {
1715                                         if (pi[j].numa_node ==
1716                                                         port_numa) {
1717                                                 current = j;
1718                                                 break;
1719                                         }
1720                                 }
1721                                 if (pi[current].rx_count == 0 &&
1722                                                 pi[current].tx_count == 0) {
1723                                         pi[current].rx_base_id = rx_queues * i;
1724                                         pi[current].tx_base_id = tx_queues * i;
1725                                         (*port_count)++;
1726                                 } else if ((rx_queues * i !=
1727                                                 pi[current].rx_base_id +
1728                                                 pi[current].rx_count) ||
1729                                                 (tx_queues * i !=
1730                                                  pi[current].tx_base_id +
1731                                                  pi[current].tx_count)) {
1732                                         PMD_INIT_LOG(ERR, "Queue ids does not "
1733                                                         "fulfill constraints");
1734                                         szedata_close(szedata_temp);
1735                                         return -EINVAL;
1736                                 }
1737                                 pi[current].rx_count += port_rx_queues;
1738                                 pi[current].tx_count += port_tx_queues;
1739                                 pi[current].numa_node = port_numa;
1740                         }
1741                 }
1742         } else {
1743                 pi[0].rx_count = max_rx_queues;
1744                 pi[0].tx_count = max_tx_queues;
1745                 pi[0].numa_node = pci_dev->device.numa_node;
1746                 *port_count = 1;
1747         }
1748
1749         szedata_close(szedata_temp);
1750         return 0;
1751 }
1752
1753 /**
1754  * @brief Allocates rte_eth_dev device.
1755  * @param pci_dev Corresponding PCI device.
1756  * @param numa_node NUMA node on which device is allocated.
1757  * @param port_no Id of rte_eth_device created on PCI device pci_dev.
1758  * @return Pointer to allocated device or NULL on error.
1759  */
1760 static struct rte_eth_dev *
1761 szedata2_eth_dev_allocate(struct rte_pci_device *pci_dev, int numa_node,
1762                 unsigned int port_no)
1763 {
1764         struct rte_eth_dev *eth_dev;
1765         char name[RTE_ETH_NAME_MAX_LEN];
1766
1767         PMD_INIT_FUNC_TRACE();
1768
1769         snprintf(name, RTE_ETH_NAME_MAX_LEN, "%s"
1770                         SZEDATA2_ETH_DEV_NAME_SUFFIX_FMT,
1771                         pci_dev->device.name, port_no);
1772         PMD_INIT_LOG(DEBUG, "Allocating eth_dev %s", name);
1773
1774         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1775                 eth_dev = rte_eth_dev_allocate(name);
1776                 if (!eth_dev)
1777                         return NULL;
1778
1779                 eth_dev->data->dev_private = rte_zmalloc_socket(name,
1780                         sizeof(struct pmd_internals), RTE_CACHE_LINE_SIZE,
1781                         numa_node);
1782                 if (!eth_dev->data->dev_private) {
1783                         rte_eth_dev_release_port(eth_dev);
1784                         return NULL;
1785                 }
1786         } else {
1787                 eth_dev = rte_eth_dev_attach_secondary(name);
1788                 if (!eth_dev)
1789                         return NULL;
1790         }
1791
1792         eth_dev->device = &pci_dev->device;
1793         rte_eth_copy_pci_info(eth_dev, pci_dev);
1794         eth_dev->data->numa_node = numa_node;
1795         return eth_dev;
1796 }
1797
1798 /**
1799  * @brief Releases interval of rte_eth_dev devices from array.
1800  * @param eth_devs Array of pointers to rte_eth_dev devices.
1801  * @param from Index in array eth_devs to start with.
1802  * @param to Index in array right after the last element to release.
1803  *
1804  * Used for releasing at failed initialization.
1805  */
1806 static void
1807 szedata2_eth_dev_release_interval(struct rte_eth_dev **eth_devs,
1808                 unsigned int from, unsigned int to)
1809 {
1810         unsigned int i;
1811
1812         PMD_INIT_FUNC_TRACE();
1813
1814         for (i = from; i < to; i++) {
1815                 rte_szedata2_eth_dev_uninit(eth_devs[i]);
1816                 rte_eth_dev_release_port(eth_devs[i]);
1817         }
1818 }
1819
1820 /**
1821  * @brief Callback .probe for struct rte_pci_driver.
1822  */
1823 static int szedata2_eth_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1824         struct rte_pci_device *pci_dev)
1825 {
1826         struct port_info port_info[SZEDATA2_MAX_PORTS];
1827         unsigned int port_count;
1828         int ret;
1829         unsigned int i;
1830         struct pci_dev_list_entry *list_entry;
1831         struct rte_eth_dev *eth_devs[SZEDATA2_MAX_PORTS] = {NULL,};
1832
1833         PMD_INIT_FUNC_TRACE();
1834
1835         ret = get_port_info(pci_dev, &port_count, port_info,
1836                         SZEDATA2_MAX_PORTS);
1837         if (ret != 0)
1838                 return ret;
1839
1840         if (port_count == 0) {
1841                 PMD_INIT_LOG(ERR, "No available ports!");
1842                 return -ENODEV;
1843         }
1844
1845         list_entry = rte_zmalloc(NULL, sizeof(struct pci_dev_list_entry),
1846                         RTE_CACHE_LINE_SIZE);
1847         if (list_entry == NULL) {
1848                 PMD_INIT_LOG(ERR, "rte_zmalloc() failed!");
1849                 return -ENOMEM;
1850         }
1851
1852         for (i = 0; i < port_count; i++) {
1853                 eth_devs[i] = szedata2_eth_dev_allocate(pci_dev,
1854                                 port_info[i].numa_node, i);
1855                 if (eth_devs[i] == NULL) {
1856                         PMD_INIT_LOG(ERR, "Failed to alloc eth_dev for port %u",
1857                                         i);
1858                         szedata2_eth_dev_release_interval(eth_devs, 0, i);
1859                         rte_free(list_entry);
1860                         return -ENOMEM;
1861                 }
1862
1863                 ret = rte_szedata2_eth_dev_init(eth_devs[i], &port_info[i]);
1864                 if (ret != 0) {
1865                         PMD_INIT_LOG(ERR, "Failed to init eth_dev for port %u",
1866                                         i);
1867                         rte_eth_dev_release_port(eth_devs[i]);
1868                         szedata2_eth_dev_release_interval(eth_devs, 0, i);
1869                         rte_free(list_entry);
1870                         return ret;
1871                 }
1872
1873                 rte_eth_dev_probing_finish(eth_devs[i]);
1874         }
1875
1876         /*
1877          * Add pci_dev to list of PCI devices for this driver
1878          * which is used at remove callback to release all created eth_devs.
1879          */
1880         list_entry->pci_dev = pci_dev;
1881         list_entry->port_count = port_count;
1882         LIST_INSERT_HEAD(&szedata2_pci_dev_list, list_entry, next);
1883         return 0;
1884 }
1885
1886 /**
1887  * @brief Callback .remove for struct rte_pci_driver.
1888  */
1889 static int szedata2_eth_pci_remove(struct rte_pci_device *pci_dev)
1890 {
1891         unsigned int i;
1892         unsigned int port_count;
1893         char name[RTE_ETH_NAME_MAX_LEN];
1894         struct rte_eth_dev *eth_dev;
1895         int ret;
1896         int retval = 0;
1897         bool found = false;
1898         struct pci_dev_list_entry *list_entry = NULL;
1899
1900         PMD_INIT_FUNC_TRACE();
1901
1902         LIST_FOREACH(list_entry, &szedata2_pci_dev_list, next) {
1903                 if (list_entry->pci_dev == pci_dev) {
1904                         port_count = list_entry->port_count;
1905                         found = true;
1906                         break;
1907                 }
1908         }
1909         LIST_REMOVE(list_entry, next);
1910         rte_free(list_entry);
1911
1912         if (!found) {
1913                 PMD_DRV_LOG(ERR, "PCI device " PCI_PRI_FMT " not found",
1914                                 pci_dev->addr.domain, pci_dev->addr.bus,
1915                                 pci_dev->addr.devid, pci_dev->addr.function);
1916                 return -ENODEV;
1917         }
1918
1919         for (i = 0; i < port_count; i++) {
1920                 snprintf(name, RTE_ETH_NAME_MAX_LEN, "%s"
1921                                 SZEDATA2_ETH_DEV_NAME_SUFFIX_FMT,
1922                                 pci_dev->device.name, i);
1923                 PMD_DRV_LOG(DEBUG, "Removing eth_dev %s", name);
1924                 eth_dev = rte_eth_dev_allocated(name);
1925                 if (eth_dev == NULL)
1926                         continue; /* port already released */
1927
1928                 ret = rte_szedata2_eth_dev_uninit(eth_dev);
1929                 if (ret != 0) {
1930                         PMD_DRV_LOG(ERR, "eth_dev %s uninit failed", name);
1931                         retval = retval ? retval : ret;
1932                 }
1933
1934                 rte_eth_dev_release_port(eth_dev);
1935         }
1936
1937         return retval;
1938 }
1939
1940 static struct rte_pci_driver szedata2_eth_driver = {
1941         .id_table = rte_szedata2_pci_id_table,
1942         .probe = szedata2_eth_pci_probe,
1943         .remove = szedata2_eth_pci_remove,
1944 };
1945
1946 RTE_PMD_REGISTER_PCI(RTE_SZEDATA2_DRIVER_NAME, szedata2_eth_driver);
1947 RTE_PMD_REGISTER_PCI_TABLE(RTE_SZEDATA2_DRIVER_NAME, rte_szedata2_pci_id_table);
1948 RTE_PMD_REGISTER_KMOD_DEP(RTE_SZEDATA2_DRIVER_NAME,
1949         "* combo6core & combov3 & szedata2 & ( szedata2_cv3 | szedata2_cv3_fdt )");
1950 RTE_LOG_REGISTER(szedata2_logtype_init, pmd.net.szedata2.init, NOTICE);
1951 RTE_LOG_REGISTER(szedata2_logtype_driver, pmd.net.szedata2.driver, NOTICE);