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 void
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
1023         for (i = 0; i < nb_tx; i++)
1024                 eth_tx_queue_stop(dev, i);
1025
1026         for (i = 0; i < nb_rx; i++)
1027                 eth_rx_queue_stop(dev, i);
1028 }
1029
1030 static int
1031 eth_dev_configure(struct rte_eth_dev *dev)
1032 {
1033         struct rte_eth_dev_data *data = dev->data;
1034         if (data->dev_conf.rxmode.offloads & DEV_RX_OFFLOAD_SCATTER) {
1035                 dev->rx_pkt_burst = eth_szedata2_rx_scattered;
1036                 data->scattered_rx = 1;
1037         } else {
1038                 dev->rx_pkt_burst = eth_szedata2_rx;
1039                 data->scattered_rx = 0;
1040         }
1041         return 0;
1042 }
1043
1044 static int
1045 eth_dev_info(struct rte_eth_dev *dev,
1046                 struct rte_eth_dev_info *dev_info)
1047 {
1048         struct pmd_internals *internals = dev->data->dev_private;
1049
1050         dev_info->if_index = 0;
1051         dev_info->max_mac_addrs = 1;
1052         dev_info->max_rx_pktlen = (uint32_t)-1;
1053         dev_info->max_rx_queues = internals->max_rx_queues;
1054         dev_info->max_tx_queues = internals->max_tx_queues;
1055         dev_info->min_rx_bufsize = 0;
1056         dev_info->rx_offload_capa = DEV_RX_OFFLOAD_SCATTER;
1057         dev_info->tx_offload_capa = 0;
1058         dev_info->rx_queue_offload_capa = 0;
1059         dev_info->tx_queue_offload_capa = 0;
1060         dev_info->speed_capa = ETH_LINK_SPEED_100G;
1061
1062         return 0;
1063 }
1064
1065 static int
1066 eth_stats_get(struct rte_eth_dev *dev,
1067                 struct rte_eth_stats *stats)
1068 {
1069         uint16_t i;
1070         uint16_t nb_rx = dev->data->nb_rx_queues;
1071         uint16_t nb_tx = dev->data->nb_tx_queues;
1072         uint64_t rx_total = 0;
1073         uint64_t tx_total = 0;
1074         uint64_t tx_err_total = 0;
1075         uint64_t rx_total_bytes = 0;
1076         uint64_t tx_total_bytes = 0;
1077
1078         for (i = 0; i < nb_rx; i++) {
1079                 struct szedata2_rx_queue *rxq = dev->data->rx_queues[i];
1080
1081                 if (i < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
1082                         stats->q_ipackets[i] = rxq->rx_pkts;
1083                         stats->q_ibytes[i] = rxq->rx_bytes;
1084                 }
1085                 rx_total += rxq->rx_pkts;
1086                 rx_total_bytes += rxq->rx_bytes;
1087         }
1088
1089         for (i = 0; i < nb_tx; i++) {
1090                 struct szedata2_tx_queue *txq = dev->data->tx_queues[i];
1091
1092                 if (i < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
1093                         stats->q_opackets[i] = txq->tx_pkts;
1094                         stats->q_obytes[i] = txq->tx_bytes;
1095                 }
1096                 tx_total += txq->tx_pkts;
1097                 tx_total_bytes += txq->tx_bytes;
1098                 tx_err_total += txq->err_pkts;
1099         }
1100
1101         stats->ipackets = rx_total;
1102         stats->opackets = tx_total;
1103         stats->ibytes = rx_total_bytes;
1104         stats->obytes = tx_total_bytes;
1105         stats->oerrors = tx_err_total;
1106         stats->rx_nombuf = dev->data->rx_mbuf_alloc_failed;
1107
1108         return 0;
1109 }
1110
1111 static int
1112 eth_stats_reset(struct rte_eth_dev *dev)
1113 {
1114         uint16_t i;
1115         uint16_t nb_rx = dev->data->nb_rx_queues;
1116         uint16_t nb_tx = dev->data->nb_tx_queues;
1117
1118         for (i = 0; i < nb_rx; i++) {
1119                 struct szedata2_rx_queue *rxq = dev->data->rx_queues[i];
1120                 rxq->rx_pkts = 0;
1121                 rxq->rx_bytes = 0;
1122                 rxq->err_pkts = 0;
1123         }
1124         for (i = 0; i < nb_tx; i++) {
1125                 struct szedata2_tx_queue *txq = dev->data->tx_queues[i];
1126                 txq->tx_pkts = 0;
1127                 txq->tx_bytes = 0;
1128                 txq->err_pkts = 0;
1129         }
1130
1131         return 0;
1132 }
1133
1134 static void
1135 eth_rx_queue_release(void *q)
1136 {
1137         struct szedata2_rx_queue *rxq = (struct szedata2_rx_queue *)q;
1138
1139         if (rxq != NULL) {
1140                 if (rxq->sze != NULL)
1141                         szedata_close(rxq->sze);
1142                 rte_free(rxq);
1143         }
1144 }
1145
1146 static void
1147 eth_tx_queue_release(void *q)
1148 {
1149         struct szedata2_tx_queue *txq = (struct szedata2_tx_queue *)q;
1150
1151         if (txq != NULL) {
1152                 if (txq->sze != NULL)
1153                         szedata_close(txq->sze);
1154                 rte_free(txq);
1155         }
1156 }
1157
1158 static int
1159 eth_dev_close(struct rte_eth_dev *dev)
1160 {
1161         struct pmd_internals *internals = dev->data->dev_private;
1162         uint16_t i;
1163         uint16_t nb_rx = dev->data->nb_rx_queues;
1164         uint16_t nb_tx = dev->data->nb_tx_queues;
1165
1166         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1167                 return 0;
1168
1169         eth_dev_stop(dev);
1170
1171         free(internals->sze_dev_path);
1172
1173         for (i = 0; i < nb_rx; i++) {
1174                 eth_rx_queue_release(dev->data->rx_queues[i]);
1175                 dev->data->rx_queues[i] = NULL;
1176         }
1177         dev->data->nb_rx_queues = 0;
1178         for (i = 0; i < nb_tx; i++) {
1179                 eth_tx_queue_release(dev->data->tx_queues[i]);
1180                 dev->data->tx_queues[i] = NULL;
1181         }
1182         dev->data->nb_tx_queues = 0;
1183
1184         return 0;
1185 }
1186
1187 static int
1188 eth_link_update(struct rte_eth_dev *dev,
1189                 int wait_to_complete __rte_unused)
1190 {
1191         struct rte_eth_link link;
1192
1193         memset(&link, 0, sizeof(link));
1194
1195         link.link_speed = ETH_SPEED_NUM_100G;
1196         link.link_duplex = ETH_LINK_FULL_DUPLEX;
1197         link.link_status = ETH_LINK_UP;
1198         link.link_autoneg = ETH_LINK_FIXED;
1199
1200         rte_eth_linkstatus_set(dev, &link);
1201         return 0;
1202 }
1203
1204 static int
1205 eth_dev_set_link_up(struct rte_eth_dev *dev __rte_unused)
1206 {
1207         PMD_DRV_LOG(WARNING, "Setting link up is not supported.");
1208         return 0;
1209 }
1210
1211 static int
1212 eth_dev_set_link_down(struct rte_eth_dev *dev __rte_unused)
1213 {
1214         PMD_DRV_LOG(WARNING, "Setting link down is not supported.");
1215         return 0;
1216 }
1217
1218 static int
1219 eth_rx_queue_setup(struct rte_eth_dev *dev,
1220                 uint16_t rx_queue_id,
1221                 uint16_t nb_rx_desc __rte_unused,
1222                 unsigned int socket_id,
1223                 const struct rte_eth_rxconf *rx_conf __rte_unused,
1224                 struct rte_mempool *mb_pool)
1225 {
1226         struct szedata2_rx_queue *rxq;
1227         int ret;
1228         struct pmd_internals *internals = dev->data->dev_private;
1229         uint8_t rx_channel = internals->rxq_base_id + rx_queue_id;
1230         uint32_t rx = 1 << rx_channel;
1231         uint32_t tx = 0;
1232
1233         PMD_INIT_FUNC_TRACE();
1234
1235         if (dev->data->rx_queues[rx_queue_id] != NULL) {
1236                 eth_rx_queue_release(dev->data->rx_queues[rx_queue_id]);
1237                 dev->data->rx_queues[rx_queue_id] = NULL;
1238         }
1239
1240         rxq = rte_zmalloc_socket("szedata2 rx queue",
1241                         sizeof(struct szedata2_rx_queue),
1242                         RTE_CACHE_LINE_SIZE, socket_id);
1243         if (rxq == NULL) {
1244                 PMD_INIT_LOG(ERR, "rte_zmalloc_socket() failed for rx queue id "
1245                                 "%" PRIu16 "!", rx_queue_id);
1246                 return -ENOMEM;
1247         }
1248
1249         rxq->priv = internals;
1250         rxq->sze = szedata_open(internals->sze_dev_path);
1251         if (rxq->sze == NULL) {
1252                 PMD_INIT_LOG(ERR, "szedata_open() failed for rx queue id "
1253                                 "%" PRIu16 "!", rx_queue_id);
1254                 eth_rx_queue_release(rxq);
1255                 return -EINVAL;
1256         }
1257         ret = szedata_subscribe3(rxq->sze, &rx, &tx);
1258         if (ret != 0 || rx == 0) {
1259                 PMD_INIT_LOG(ERR, "szedata_subscribe3() failed for rx queue id "
1260                                 "%" PRIu16 "!", rx_queue_id);
1261                 eth_rx_queue_release(rxq);
1262                 return -EINVAL;
1263         }
1264         rxq->rx_channel = rx_channel;
1265         rxq->qid = rx_queue_id;
1266         rxq->in_port = dev->data->port_id;
1267         rxq->mb_pool = mb_pool;
1268         rxq->rx_pkts = 0;
1269         rxq->rx_bytes = 0;
1270         rxq->err_pkts = 0;
1271
1272         dev->data->rx_queues[rx_queue_id] = rxq;
1273
1274         PMD_INIT_LOG(DEBUG, "Configured rx queue id %" PRIu16 " on socket "
1275                         "%u (channel id %u).", rxq->qid, socket_id,
1276                         rxq->rx_channel);
1277
1278         return 0;
1279 }
1280
1281 static int
1282 eth_tx_queue_setup(struct rte_eth_dev *dev,
1283                 uint16_t tx_queue_id,
1284                 uint16_t nb_tx_desc __rte_unused,
1285                 unsigned int socket_id,
1286                 const struct rte_eth_txconf *tx_conf __rte_unused)
1287 {
1288         struct szedata2_tx_queue *txq;
1289         int ret;
1290         struct pmd_internals *internals = dev->data->dev_private;
1291         uint8_t tx_channel = internals->txq_base_id + tx_queue_id;
1292         uint32_t rx = 0;
1293         uint32_t tx = 1 << tx_channel;
1294
1295         PMD_INIT_FUNC_TRACE();
1296
1297         if (dev->data->tx_queues[tx_queue_id] != NULL) {
1298                 eth_tx_queue_release(dev->data->tx_queues[tx_queue_id]);
1299                 dev->data->tx_queues[tx_queue_id] = NULL;
1300         }
1301
1302         txq = rte_zmalloc_socket("szedata2 tx queue",
1303                         sizeof(struct szedata2_tx_queue),
1304                         RTE_CACHE_LINE_SIZE, socket_id);
1305         if (txq == NULL) {
1306                 PMD_INIT_LOG(ERR, "rte_zmalloc_socket() failed for tx queue id "
1307                                 "%" PRIu16 "!", tx_queue_id);
1308                 return -ENOMEM;
1309         }
1310
1311         txq->priv = internals;
1312         txq->sze = szedata_open(internals->sze_dev_path);
1313         if (txq->sze == NULL) {
1314                 PMD_INIT_LOG(ERR, "szedata_open() failed for tx queue id "
1315                                 "%" PRIu16 "!", tx_queue_id);
1316                 eth_tx_queue_release(txq);
1317                 return -EINVAL;
1318         }
1319         ret = szedata_subscribe3(txq->sze, &rx, &tx);
1320         if (ret != 0 || tx == 0) {
1321                 PMD_INIT_LOG(ERR, "szedata_subscribe3() failed for tx queue id "
1322                                 "%" PRIu16 "!", tx_queue_id);
1323                 eth_tx_queue_release(txq);
1324                 return -EINVAL;
1325         }
1326         txq->tx_channel = tx_channel;
1327         txq->qid = tx_queue_id;
1328         txq->tx_pkts = 0;
1329         txq->tx_bytes = 0;
1330         txq->err_pkts = 0;
1331
1332         dev->data->tx_queues[tx_queue_id] = txq;
1333
1334         PMD_INIT_LOG(DEBUG, "Configured tx queue id %" PRIu16 " on socket "
1335                         "%u (channel id %u).", txq->qid, socket_id,
1336                         txq->tx_channel);
1337
1338         return 0;
1339 }
1340
1341 static int
1342 eth_mac_addr_set(struct rte_eth_dev *dev __rte_unused,
1343                 struct rte_ether_addr *mac_addr __rte_unused)
1344 {
1345         return 0;
1346 }
1347
1348 static int
1349 eth_promiscuous_enable(struct rte_eth_dev *dev __rte_unused)
1350 {
1351         PMD_DRV_LOG(WARNING, "Enabling promiscuous mode is not supported. "
1352                         "The card is always in promiscuous mode.");
1353         return 0;
1354 }
1355
1356 static int
1357 eth_promiscuous_disable(struct rte_eth_dev *dev __rte_unused)
1358 {
1359         PMD_DRV_LOG(WARNING, "Disabling promiscuous mode is not supported. "
1360                         "The card is always in promiscuous mode.");
1361         return -ENOTSUP;
1362 }
1363
1364 static int
1365 eth_allmulticast_enable(struct rte_eth_dev *dev __rte_unused)
1366 {
1367         PMD_DRV_LOG(WARNING, "Enabling allmulticast mode is not supported.");
1368         return -ENOTSUP;
1369 }
1370
1371 static int
1372 eth_allmulticast_disable(struct rte_eth_dev *dev __rte_unused)
1373 {
1374         PMD_DRV_LOG(WARNING, "Disabling allmulticast mode is not supported.");
1375         return -ENOTSUP;
1376 }
1377
1378 static const struct eth_dev_ops ops = {
1379         .dev_start          = eth_dev_start,
1380         .dev_stop           = eth_dev_stop,
1381         .dev_set_link_up    = eth_dev_set_link_up,
1382         .dev_set_link_down  = eth_dev_set_link_down,
1383         .dev_close          = eth_dev_close,
1384         .dev_configure      = eth_dev_configure,
1385         .dev_infos_get      = eth_dev_info,
1386         .promiscuous_enable   = eth_promiscuous_enable,
1387         .promiscuous_disable  = eth_promiscuous_disable,
1388         .allmulticast_enable  = eth_allmulticast_enable,
1389         .allmulticast_disable = eth_allmulticast_disable,
1390         .rx_queue_start     = eth_rx_queue_start,
1391         .rx_queue_stop      = eth_rx_queue_stop,
1392         .tx_queue_start     = eth_tx_queue_start,
1393         .tx_queue_stop      = eth_tx_queue_stop,
1394         .rx_queue_setup     = eth_rx_queue_setup,
1395         .tx_queue_setup     = eth_tx_queue_setup,
1396         .rx_queue_release   = eth_rx_queue_release,
1397         .tx_queue_release   = eth_tx_queue_release,
1398         .link_update        = eth_link_update,
1399         .stats_get          = eth_stats_get,
1400         .stats_reset        = eth_stats_reset,
1401         .mac_addr_set       = eth_mac_addr_set,
1402 };
1403
1404 /*
1405  * This function goes through sysfs and looks for an index of szedata2
1406  * device file (/dev/szedataIIX, where X is the index).
1407  *
1408  * @return
1409  *           0 on success
1410  *          -1 on error
1411  */
1412 static int
1413 get_szedata2_index(const struct rte_pci_addr *pcislot_addr, uint32_t *index)
1414 {
1415         DIR *dir;
1416         struct dirent *entry;
1417         int ret;
1418         uint32_t tmp_index;
1419         FILE *fd;
1420         char pcislot_path[PATH_MAX];
1421         uint32_t domain;
1422         uint8_t bus;
1423         uint8_t devid;
1424         uint8_t function;
1425
1426         dir = opendir("/sys/class/combo");
1427         if (dir == NULL)
1428                 return -1;
1429
1430         /*
1431          * Iterate through all combosixX directories.
1432          * When the value in /sys/class/combo/combosixX/device/pcislot
1433          * file is the location of the ethernet device dev, "X" is the
1434          * index of the device.
1435          */
1436         while ((entry = readdir(dir)) != NULL) {
1437                 ret = sscanf(entry->d_name, "combosix%u", &tmp_index);
1438                 if (ret != 1)
1439                         continue;
1440
1441                 snprintf(pcislot_path, PATH_MAX,
1442                         "/sys/class/combo/combosix%u/device/pcislot",
1443                         tmp_index);
1444
1445                 fd = fopen(pcislot_path, "r");
1446                 if (fd == NULL)
1447                         continue;
1448
1449                 ret = fscanf(fd, "%8" SCNx32 ":%2" SCNx8 ":%2" SCNx8 ".%" SCNx8,
1450                                 &domain, &bus, &devid, &function);
1451                 fclose(fd);
1452                 if (ret != 4)
1453                         continue;
1454
1455                 if (pcislot_addr->domain == domain &&
1456                                 pcislot_addr->bus == bus &&
1457                                 pcislot_addr->devid == devid &&
1458                                 pcislot_addr->function == function) {
1459                         *index = tmp_index;
1460                         closedir(dir);
1461                         return 0;
1462                 }
1463         }
1464
1465         closedir(dir);
1466         return -1;
1467 }
1468
1469 /**
1470  * @brief Initializes rte_eth_dev device.
1471  * @param dev Device to initialize.
1472  * @param pi Structure with info about DMA queues.
1473  * @return 0 on success, negative error code on error.
1474  */
1475 static int
1476 rte_szedata2_eth_dev_init(struct rte_eth_dev *dev, struct port_info *pi)
1477 {
1478         int ret;
1479         uint32_t szedata2_index;
1480         char name[PATH_MAX];
1481         struct rte_eth_dev_data *data = dev->data;
1482         struct pmd_internals *internals = (struct pmd_internals *)
1483                 data->dev_private;
1484         struct rte_pci_device *pci_dev = RTE_ETH_DEV_TO_PCI(dev);
1485
1486         PMD_INIT_FUNC_TRACE();
1487
1488         PMD_INIT_LOG(INFO, "Initializing eth_dev %s (driver %s)", data->name,
1489                         RTE_STR(RTE_SZEDATA2_DRIVER_NAME));
1490
1491         /* Fill internal private structure. */
1492         internals->dev = dev;
1493         /* Get index of szedata2 device file and create path to device file */
1494         ret = get_szedata2_index(&pci_dev->addr, &szedata2_index);
1495         if (ret != 0) {
1496                 PMD_INIT_LOG(ERR, "Failed to get szedata2 device index!");
1497                 return -ENODEV;
1498         }
1499         snprintf(name, PATH_MAX, SZEDATA2_DEV_PATH_FMT, szedata2_index);
1500         internals->sze_dev_path = strdup(name);
1501         if (internals->sze_dev_path == NULL) {
1502                 PMD_INIT_LOG(ERR, "strdup() failed!");
1503                 return -ENOMEM;
1504         }
1505         PMD_INIT_LOG(INFO, "SZEDATA2 path: %s", internals->sze_dev_path);
1506         internals->max_rx_queues = pi->rx_count;
1507         internals->max_tx_queues = pi->tx_count;
1508         internals->rxq_base_id = pi->rx_base_id;
1509         internals->txq_base_id = pi->tx_base_id;
1510         PMD_INIT_LOG(INFO, "%u RX DMA channels from id %u",
1511                         internals->max_rx_queues, internals->rxq_base_id);
1512         PMD_INIT_LOG(INFO, "%u TX DMA channels from id %u",
1513                         internals->max_tx_queues, internals->txq_base_id);
1514
1515         /* Set rx, tx burst functions */
1516         if (data->scattered_rx == 1)
1517                 dev->rx_pkt_burst = eth_szedata2_rx_scattered;
1518         else
1519                 dev->rx_pkt_burst = eth_szedata2_rx;
1520         dev->tx_pkt_burst = eth_szedata2_tx;
1521
1522         /* Set function callbacks for Ethernet API */
1523         dev->dev_ops = &ops;
1524
1525         /* Get link state */
1526         eth_link_update(dev, 0);
1527
1528         /* Allocate space for one mac address */
1529         data->mac_addrs = rte_zmalloc(data->name, sizeof(struct rte_ether_addr),
1530                         RTE_CACHE_LINE_SIZE);
1531         if (data->mac_addrs == NULL) {
1532                 PMD_INIT_LOG(ERR, "Could not alloc space for MAC address!");
1533                 free(internals->sze_dev_path);
1534                 return -ENOMEM;
1535         }
1536
1537         rte_ether_addr_copy(&eth_addr, data->mac_addrs);
1538
1539         PMD_INIT_LOG(INFO, "%s device %s successfully initialized",
1540                         RTE_STR(RTE_SZEDATA2_DRIVER_NAME), data->name);
1541
1542         return 0;
1543 }
1544
1545 /**
1546  * @brief Unitializes rte_eth_dev device.
1547  * @param dev Device to uninitialize.
1548  * @return 0 on success, negative error code on error.
1549  */
1550 static int
1551 rte_szedata2_eth_dev_uninit(struct rte_eth_dev *dev)
1552 {
1553         PMD_INIT_FUNC_TRACE();
1554
1555         eth_dev_close(dev);
1556
1557         PMD_DRV_LOG(INFO, "%s device %s successfully uninitialized",
1558                         RTE_STR(RTE_SZEDATA2_DRIVER_NAME), dev->data->name);
1559
1560         return 0;
1561 }
1562
1563 static const struct rte_pci_id rte_szedata2_pci_id_table[] = {
1564         {
1565                 RTE_PCI_DEVICE(PCI_VENDOR_ID_NETCOPE,
1566                                 PCI_DEVICE_ID_NETCOPE_COMBO80G)
1567         },
1568         {
1569                 RTE_PCI_DEVICE(PCI_VENDOR_ID_NETCOPE,
1570                                 PCI_DEVICE_ID_NETCOPE_COMBO100G)
1571         },
1572         {
1573                 RTE_PCI_DEVICE(PCI_VENDOR_ID_NETCOPE,
1574                                 PCI_DEVICE_ID_NETCOPE_COMBO100G2)
1575         },
1576         {
1577                 RTE_PCI_DEVICE(PCI_VENDOR_ID_NETCOPE,
1578                                 PCI_DEVICE_ID_NETCOPE_NFB200G2QL)
1579         },
1580         {
1581                 RTE_PCI_DEVICE(PCI_VENDOR_ID_SILICOM,
1582                                 PCI_DEVICE_ID_FB2CGG3)
1583         },
1584         {
1585                 RTE_PCI_DEVICE(PCI_VENDOR_ID_SILICOM,
1586                                 PCI_DEVICE_ID_FB2CGG3D)
1587         },
1588         {
1589                 .vendor_id = 0,
1590         }
1591 };
1592
1593 /**
1594  * @brief Gets info about DMA queues for ports.
1595  * @param pci_dev PCI device structure.
1596  * @param port_count Pointer to variable set with number of ports.
1597  * @param pi Pointer to array of structures with info about DMA queues
1598  *           for ports.
1599  * @param max_ports Maximum number of ports.
1600  * @return 0 on success, negative error code on error.
1601  */
1602 static int
1603 get_port_info(struct rte_pci_device *pci_dev, unsigned int *port_count,
1604                 struct port_info *pi, unsigned int max_ports)
1605 {
1606         struct szedata *szedata_temp;
1607         char sze_dev_path[PATH_MAX];
1608         uint32_t szedata2_index;
1609         int ret;
1610         uint16_t max_rx_queues;
1611         uint16_t max_tx_queues;
1612
1613         if (max_ports == 0)
1614                 return -EINVAL;
1615
1616         memset(pi, 0, max_ports * sizeof(struct port_info));
1617         *port_count = 0;
1618
1619         /* Get index of szedata2 device file and create path to device file */
1620         ret = get_szedata2_index(&pci_dev->addr, &szedata2_index);
1621         if (ret != 0) {
1622                 PMD_INIT_LOG(ERR, "Failed to get szedata2 device index!");
1623                 return -ENODEV;
1624         }
1625         snprintf(sze_dev_path, PATH_MAX, SZEDATA2_DEV_PATH_FMT, szedata2_index);
1626
1627         /*
1628          * Get number of available DMA RX and TX channels, which is maximum
1629          * number of queues that can be created.
1630          */
1631         szedata_temp = szedata_open(sze_dev_path);
1632         if (szedata_temp == NULL) {
1633                 PMD_INIT_LOG(ERR, "szedata_open(%s) failed", sze_dev_path);
1634                 return -EINVAL;
1635         }
1636         max_rx_queues = szedata_ifaces_available(szedata_temp, SZE2_DIR_RX);
1637         max_tx_queues = szedata_ifaces_available(szedata_temp, SZE2_DIR_TX);
1638         PMD_INIT_LOG(INFO, "Available DMA channels RX: %u TX: %u",
1639                         max_rx_queues, max_tx_queues);
1640         if (max_rx_queues > RTE_ETH_SZEDATA2_MAX_RX_QUEUES) {
1641                 PMD_INIT_LOG(ERR, "%u RX queues exceeds supported number %u",
1642                                 max_rx_queues, RTE_ETH_SZEDATA2_MAX_RX_QUEUES);
1643                 szedata_close(szedata_temp);
1644                 return -EINVAL;
1645         }
1646         if (max_tx_queues > RTE_ETH_SZEDATA2_MAX_TX_QUEUES) {
1647                 PMD_INIT_LOG(ERR, "%u TX queues exceeds supported number %u",
1648                                 max_tx_queues, RTE_ETH_SZEDATA2_MAX_TX_QUEUES);
1649                 szedata_close(szedata_temp);
1650                 return -EINVAL;
1651         }
1652
1653         if (pci_dev->id.device_id == PCI_DEVICE_ID_NETCOPE_NFB200G2QL) {
1654                 unsigned int i;
1655                 unsigned int rx_queues = max_rx_queues / max_ports;
1656                 unsigned int tx_queues = max_tx_queues / max_ports;
1657
1658                 /*
1659                  * Number of queues reported by szedata_ifaces_available()
1660                  * is the number of all queues from all DMA controllers which
1661                  * may reside at different numa locations.
1662                  * All queues from the same DMA controller have the same numa
1663                  * node.
1664                  * Numa node from the first queue of each DMA controller is
1665                  * retrieved.
1666                  * If the numa node differs from the numa node of the queues
1667                  * from the previous DMA controller the queues are assigned
1668                  * to the next port.
1669                  */
1670
1671                 for (i = 0; i < max_ports; i++) {
1672                         int numa_rx = szedata_get_area_numa_node(szedata_temp,
1673                                 SZE2_DIR_RX, rx_queues * i);
1674                         int numa_tx = szedata_get_area_numa_node(szedata_temp,
1675                                 SZE2_DIR_TX, tx_queues * i);
1676                         unsigned int port_rx_queues = numa_rx != -1 ?
1677                                 rx_queues : 0;
1678                         unsigned int port_tx_queues = numa_tx != -1 ?
1679                                 tx_queues : 0;
1680                         PMD_INIT_LOG(DEBUG, "%u rx queues from id %u, numa %d",
1681                                         rx_queues, rx_queues * i, numa_rx);
1682                         PMD_INIT_LOG(DEBUG, "%u tx queues from id %u, numa %d",
1683                                         tx_queues, tx_queues * i, numa_tx);
1684
1685                         if (port_rx_queues != 0 && port_tx_queues != 0 &&
1686                                         numa_rx != numa_tx) {
1687                                 PMD_INIT_LOG(ERR, "RX queue %u numa %d differs "
1688                                                 "from TX queue %u numa %d "
1689                                                 "unexpectedly",
1690                                                 rx_queues * i, numa_rx,
1691                                                 tx_queues * i, numa_tx);
1692                                 szedata_close(szedata_temp);
1693                                 return -EINVAL;
1694                         } else if (port_rx_queues == 0 && port_tx_queues == 0) {
1695                                 continue;
1696                         } else {
1697                                 unsigned int j;
1698                                 unsigned int current = *port_count;
1699                                 int port_numa = port_rx_queues != 0 ?
1700                                         numa_rx : numa_tx;
1701
1702                                 for (j = 0; j < *port_count; j++) {
1703                                         if (pi[j].numa_node ==
1704                                                         port_numa) {
1705                                                 current = j;
1706                                                 break;
1707                                         }
1708                                 }
1709                                 if (pi[current].rx_count == 0 &&
1710                                                 pi[current].tx_count == 0) {
1711                                         pi[current].rx_base_id = rx_queues * i;
1712                                         pi[current].tx_base_id = tx_queues * i;
1713                                         (*port_count)++;
1714                                 } else if ((rx_queues * i !=
1715                                                 pi[current].rx_base_id +
1716                                                 pi[current].rx_count) ||
1717                                                 (tx_queues * i !=
1718                                                  pi[current].tx_base_id +
1719                                                  pi[current].tx_count)) {
1720                                         PMD_INIT_LOG(ERR, "Queue ids does not "
1721                                                         "fulfill constraints");
1722                                         szedata_close(szedata_temp);
1723                                         return -EINVAL;
1724                                 }
1725                                 pi[current].rx_count += port_rx_queues;
1726                                 pi[current].tx_count += port_tx_queues;
1727                                 pi[current].numa_node = port_numa;
1728                         }
1729                 }
1730         } else {
1731                 pi[0].rx_count = max_rx_queues;
1732                 pi[0].tx_count = max_tx_queues;
1733                 pi[0].numa_node = pci_dev->device.numa_node;
1734                 *port_count = 1;
1735         }
1736
1737         szedata_close(szedata_temp);
1738         return 0;
1739 }
1740
1741 /**
1742  * @brief Allocates rte_eth_dev device.
1743  * @param pci_dev Corresponding PCI device.
1744  * @param numa_node NUMA node on which device is allocated.
1745  * @param port_no Id of rte_eth_device created on PCI device pci_dev.
1746  * @return Pointer to allocated device or NULL on error.
1747  */
1748 static struct rte_eth_dev *
1749 szedata2_eth_dev_allocate(struct rte_pci_device *pci_dev, int numa_node,
1750                 unsigned int port_no)
1751 {
1752         struct rte_eth_dev *eth_dev;
1753         char name[RTE_ETH_NAME_MAX_LEN];
1754
1755         PMD_INIT_FUNC_TRACE();
1756
1757         snprintf(name, RTE_ETH_NAME_MAX_LEN, "%s"
1758                         SZEDATA2_ETH_DEV_NAME_SUFFIX_FMT,
1759                         pci_dev->device.name, port_no);
1760         PMD_INIT_LOG(DEBUG, "Allocating eth_dev %s", name);
1761
1762         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1763                 eth_dev = rte_eth_dev_allocate(name);
1764                 if (!eth_dev)
1765                         return NULL;
1766
1767                 eth_dev->data->dev_private = rte_zmalloc_socket(name,
1768                         sizeof(struct pmd_internals), RTE_CACHE_LINE_SIZE,
1769                         numa_node);
1770                 if (!eth_dev->data->dev_private) {
1771                         rte_eth_dev_release_port(eth_dev);
1772                         return NULL;
1773                 }
1774         } else {
1775                 eth_dev = rte_eth_dev_attach_secondary(name);
1776                 if (!eth_dev)
1777                         return NULL;
1778         }
1779
1780         eth_dev->device = &pci_dev->device;
1781         rte_eth_copy_pci_info(eth_dev, pci_dev);
1782         eth_dev->data->numa_node = numa_node;
1783         return eth_dev;
1784 }
1785
1786 /**
1787  * @brief Releases interval of rte_eth_dev devices from array.
1788  * @param eth_devs Array of pointers to rte_eth_dev devices.
1789  * @param from Index in array eth_devs to start with.
1790  * @param to Index in array right after the last element to release.
1791  *
1792  * Used for releasing at failed initialization.
1793  */
1794 static void
1795 szedata2_eth_dev_release_interval(struct rte_eth_dev **eth_devs,
1796                 unsigned int from, unsigned int to)
1797 {
1798         unsigned int i;
1799
1800         PMD_INIT_FUNC_TRACE();
1801
1802         for (i = from; i < to; i++) {
1803                 rte_szedata2_eth_dev_uninit(eth_devs[i]);
1804                 rte_eth_dev_release_port(eth_devs[i]);
1805         }
1806 }
1807
1808 /**
1809  * @brief Callback .probe for struct rte_pci_driver.
1810  */
1811 static int szedata2_eth_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1812         struct rte_pci_device *pci_dev)
1813 {
1814         struct port_info port_info[SZEDATA2_MAX_PORTS];
1815         unsigned int port_count;
1816         int ret;
1817         unsigned int i;
1818         struct pci_dev_list_entry *list_entry;
1819         struct rte_eth_dev *eth_devs[SZEDATA2_MAX_PORTS] = {NULL,};
1820
1821         PMD_INIT_FUNC_TRACE();
1822
1823         ret = get_port_info(pci_dev, &port_count, port_info,
1824                         SZEDATA2_MAX_PORTS);
1825         if (ret != 0)
1826                 return ret;
1827
1828         if (port_count == 0) {
1829                 PMD_INIT_LOG(ERR, "No available ports!");
1830                 return -ENODEV;
1831         }
1832
1833         list_entry = rte_zmalloc(NULL, sizeof(struct pci_dev_list_entry),
1834                         RTE_CACHE_LINE_SIZE);
1835         if (list_entry == NULL) {
1836                 PMD_INIT_LOG(ERR, "rte_zmalloc() failed!");
1837                 return -ENOMEM;
1838         }
1839
1840         for (i = 0; i < port_count; i++) {
1841                 eth_devs[i] = szedata2_eth_dev_allocate(pci_dev,
1842                                 port_info[i].numa_node, i);
1843                 if (eth_devs[i] == NULL) {
1844                         PMD_INIT_LOG(ERR, "Failed to alloc eth_dev for port %u",
1845                                         i);
1846                         szedata2_eth_dev_release_interval(eth_devs, 0, i);
1847                         rte_free(list_entry);
1848                         return -ENOMEM;
1849                 }
1850
1851                 ret = rte_szedata2_eth_dev_init(eth_devs[i], &port_info[i]);
1852                 if (ret != 0) {
1853                         PMD_INIT_LOG(ERR, "Failed to init eth_dev for port %u",
1854                                         i);
1855                         rte_eth_dev_release_port(eth_devs[i]);
1856                         szedata2_eth_dev_release_interval(eth_devs, 0, i);
1857                         rte_free(list_entry);
1858                         return ret;
1859                 }
1860
1861                 rte_eth_dev_probing_finish(eth_devs[i]);
1862         }
1863
1864         /*
1865          * Add pci_dev to list of PCI devices for this driver
1866          * which is used at remove callback to release all created eth_devs.
1867          */
1868         list_entry->pci_dev = pci_dev;
1869         list_entry->port_count = port_count;
1870         LIST_INSERT_HEAD(&szedata2_pci_dev_list, list_entry, next);
1871         return 0;
1872 }
1873
1874 /**
1875  * @brief Callback .remove for struct rte_pci_driver.
1876  */
1877 static int szedata2_eth_pci_remove(struct rte_pci_device *pci_dev)
1878 {
1879         unsigned int i;
1880         unsigned int port_count;
1881         char name[RTE_ETH_NAME_MAX_LEN];
1882         struct rte_eth_dev *eth_dev;
1883         int ret;
1884         int retval = 0;
1885         bool found = false;
1886         struct pci_dev_list_entry *list_entry = NULL;
1887
1888         PMD_INIT_FUNC_TRACE();
1889
1890         LIST_FOREACH(list_entry, &szedata2_pci_dev_list, next) {
1891                 if (list_entry->pci_dev == pci_dev) {
1892                         port_count = list_entry->port_count;
1893                         found = true;
1894                         break;
1895                 }
1896         }
1897         LIST_REMOVE(list_entry, next);
1898         rte_free(list_entry);
1899
1900         if (!found) {
1901                 PMD_DRV_LOG(ERR, "PCI device " PCI_PRI_FMT " not found",
1902                                 pci_dev->addr.domain, pci_dev->addr.bus,
1903                                 pci_dev->addr.devid, pci_dev->addr.function);
1904                 return -ENODEV;
1905         }
1906
1907         for (i = 0; i < port_count; i++) {
1908                 snprintf(name, RTE_ETH_NAME_MAX_LEN, "%s"
1909                                 SZEDATA2_ETH_DEV_NAME_SUFFIX_FMT,
1910                                 pci_dev->device.name, i);
1911                 PMD_DRV_LOG(DEBUG, "Removing eth_dev %s", name);
1912                 eth_dev = rte_eth_dev_allocated(name);
1913                 if (eth_dev == NULL)
1914                         continue; /* port already released */
1915
1916                 ret = rte_szedata2_eth_dev_uninit(eth_dev);
1917                 if (ret != 0) {
1918                         PMD_DRV_LOG(ERR, "eth_dev %s uninit failed", name);
1919                         retval = retval ? retval : ret;
1920                 }
1921
1922                 rte_eth_dev_release_port(eth_dev);
1923         }
1924
1925         return retval;
1926 }
1927
1928 static struct rte_pci_driver szedata2_eth_driver = {
1929         .id_table = rte_szedata2_pci_id_table,
1930         .probe = szedata2_eth_pci_probe,
1931         .remove = szedata2_eth_pci_remove,
1932 };
1933
1934 RTE_PMD_REGISTER_PCI(RTE_SZEDATA2_DRIVER_NAME, szedata2_eth_driver);
1935 RTE_PMD_REGISTER_PCI_TABLE(RTE_SZEDATA2_DRIVER_NAME, rte_szedata2_pci_id_table);
1936 RTE_PMD_REGISTER_KMOD_DEP(RTE_SZEDATA2_DRIVER_NAME,
1937         "* combo6core & combov3 & szedata2 & ( szedata2_cv3 | szedata2_cv3_fdt )");
1938 RTE_LOG_REGISTER(szedata2_logtype_init, pmd.net.szedata2.init, NOTICE);
1939 RTE_LOG_REGISTER(szedata2_logtype_driver, pmd.net.szedata2.driver, NOTICE);