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 int szedata2_logtype_init;
 106 int szedata2_logtype_driver;
 107
 108 static struct rte_ether_addr eth_addr = {
 109         .addr_bytes = { 0x00, 0x11, 0x17, 0x00, 0x00, 0x00 }
 110 };
 111
 112 static uint16_t
 113 eth_szedata2_rx(void *queue,
 114                 struct rte_mbuf **bufs,
 115                 uint16_t nb_pkts)
 116 {
 117         unsigned int i;
 118         struct rte_mbuf *mbuf;
 119         struct szedata2_rx_queue *sze_q = queue;
 120         struct rte_pktmbuf_pool_private *mbp_priv;
 121         uint16_t num_rx = 0;
 122         uint16_t buf_size;
 123         uint16_t sg_size;
 124         uint16_t hw_size;
 125         uint16_t packet_size;
 126         uint64_t num_bytes = 0;
 127         struct szedata *sze = sze_q->sze;
 128         uint8_t *header_ptr = NULL; /* header of packet */
 129         uint8_t *packet_ptr1 = NULL;
 130         uint8_t *packet_ptr2 = NULL;
 131         uint16_t packet_len1 = 0;
 132         uint16_t packet_len2 = 0;
 133         uint16_t hw_data_align;
 134
 135         if (unlikely(sze_q->sze == NULL || nb_pkts == 0))
 136                 return 0;
 137
 138         /*
 139          * Reads the given number of packets from szedata2 channel given
 140          * by queue and copies the packet data into a newly allocated mbuf
 141          * to return.
 142          */
 143         for (i = 0; i < nb_pkts; i++) {
 144                 mbuf = rte_pktmbuf_alloc(sze_q->mb_pool);
 145
 146                 if (unlikely(mbuf == NULL)) {
 147                         sze_q->priv->dev->data->rx_mbuf_alloc_failed++;
 148                         break;
 149                 }
 150
 151                 /* get the next sze packet */
 152                 if (sze->ct_rx_lck != NULL && !sze->ct_rx_rem_bytes &&
 153                                 sze->ct_rx_lck->next == NULL) {
 154                         /* unlock old data */
 155                         szedata_rx_unlock_data(sze_q->sze, sze->ct_rx_lck_orig);
 156                         sze->ct_rx_lck_orig = NULL;
 157                         sze->ct_rx_lck = NULL;
 158                 }
 159
 160                 if (!sze->ct_rx_rem_bytes && sze->ct_rx_lck_orig == NULL) {
 161                         /* nothing to read, lock new data */
 162                         sze->ct_rx_lck = szedata_rx_lock_data(sze_q->sze, ~0U);
 163                         sze->ct_rx_lck_orig = sze->ct_rx_lck;
 164
 165                         if (sze->ct_rx_lck == NULL) {
 166                                 /* nothing to lock */
 167                                 rte_pktmbuf_free(mbuf);
 168                                 break;
 169                         }
 170
 171                         sze->ct_rx_cur_ptr = sze->ct_rx_lck->start;
 172                         sze->ct_rx_rem_bytes = sze->ct_rx_lck->len;
 173
 174                         if (!sze->ct_rx_rem_bytes) {
 175                                 rte_pktmbuf_free(mbuf);
 176                                 break;
 177                         }
 178                 }
 179
 180                 if (sze->ct_rx_rem_bytes < RTE_SZE2_PACKET_HEADER_SIZE) {
 181                         /*
 182                          * cut in header
 183                          * copy parts of header to merge buffer
 184                          */
 185                         if (sze->ct_rx_lck->next == NULL) {
 186                                 rte_pktmbuf_free(mbuf);
 187                                 break;
 188                         }
 189
 190                         /* copy first part of header */
 191                         rte_memcpy(sze->ct_rx_buffer, sze->ct_rx_cur_ptr,
 192                                         sze->ct_rx_rem_bytes);
 193
 194                         /* copy second part of header */
 195                         sze->ct_rx_lck = sze->ct_rx_lck->next;
 196                         sze->ct_rx_cur_ptr = sze->ct_rx_lck->start;
 197                         rte_memcpy(sze->ct_rx_buffer + sze->ct_rx_rem_bytes,
 198                                 sze->ct_rx_cur_ptr,
 199                                 RTE_SZE2_PACKET_HEADER_SIZE -
 200                                 sze->ct_rx_rem_bytes);
 201
 202                         sze->ct_rx_cur_ptr += RTE_SZE2_PACKET_HEADER_SIZE -
 203                                 sze->ct_rx_rem_bytes;
 204                         sze->ct_rx_rem_bytes = sze->ct_rx_lck->len -
 205                                 RTE_SZE2_PACKET_HEADER_SIZE +
 206                                 sze->ct_rx_rem_bytes;
 207
 208                         header_ptr = (uint8_t *)sze->ct_rx_buffer;
 209                 } else {
 210                         /* not cut */
 211                         header_ptr = (uint8_t *)sze->ct_rx_cur_ptr;
 212                         sze->ct_rx_cur_ptr += RTE_SZE2_PACKET_HEADER_SIZE;
 213                         sze->ct_rx_rem_bytes -= RTE_SZE2_PACKET_HEADER_SIZE;
 214                 }
 215
 216                 sg_size = le16toh(*((uint16_t *)header_ptr));
 217                 hw_size = le16toh(*(((uint16_t *)header_ptr) + 1));
 218                 packet_size = sg_size -
 219                         RTE_SZE2_ALIGN8(RTE_SZE2_PACKET_HEADER_SIZE + hw_size);
 220
 221
 222                 /* checks if packet all right */
 223                 if (!sg_size)
 224                         errx(5, "Zero segsize");
 225
 226                 /* check sg_size and hwsize */
 227                 if (hw_size > sg_size - RTE_SZE2_PACKET_HEADER_SIZE) {
 228                         errx(10, "Hwsize bigger than expected. Segsize: %d, "
 229                                 "hwsize: %d", sg_size, hw_size);
 230                 }
 231
 232                 hw_data_align =
 233                         RTE_SZE2_ALIGN8(RTE_SZE2_PACKET_HEADER_SIZE + hw_size) -
 234                         RTE_SZE2_PACKET_HEADER_SIZE;
 235
 236                 if (sze->ct_rx_rem_bytes >=
 237                                 (uint16_t)(sg_size -
 238                                 RTE_SZE2_PACKET_HEADER_SIZE)) {
 239                         /* no cut */
 240                         /* one packet ready - go to another */
 241                         packet_ptr1 = sze->ct_rx_cur_ptr + hw_data_align;
 242                         packet_len1 = packet_size;
 243                         packet_ptr2 = NULL;
 244                         packet_len2 = 0;
 245
 246                         sze->ct_rx_cur_ptr += RTE_SZE2_ALIGN8(sg_size) -
 247                                 RTE_SZE2_PACKET_HEADER_SIZE;
 248                         sze->ct_rx_rem_bytes -= RTE_SZE2_ALIGN8(sg_size) -
 249                                 RTE_SZE2_PACKET_HEADER_SIZE;
 250                 } else {
 251                         /* cut in data */
 252                         if (sze->ct_rx_lck->next == NULL) {
 253                                 errx(6, "Need \"next\" lock, "
 254                                         "but it is missing: %u",
 255                                         sze->ct_rx_rem_bytes);
 256                         }
 257
 258                         /* skip hw data */
 259                         if (sze->ct_rx_rem_bytes <= hw_data_align) {
 260                                 uint16_t rem_size = hw_data_align -
 261                                         sze->ct_rx_rem_bytes;
 262
 263                                 /* MOVE to next lock */
 264                                 sze->ct_rx_lck = sze->ct_rx_lck->next;
 265                                 sze->ct_rx_cur_ptr =
 266                                         (void *)(((uint8_t *)
 267                                         (sze->ct_rx_lck->start)) + rem_size);
 268
 269                                 packet_ptr1 = sze->ct_rx_cur_ptr;
 270                                 packet_len1 = packet_size;
 271                                 packet_ptr2 = NULL;
 272                                 packet_len2 = 0;
 273
 274                                 sze->ct_rx_cur_ptr +=
 275                                         RTE_SZE2_ALIGN8(packet_size);
 276                                 sze->ct_rx_rem_bytes = sze->ct_rx_lck->len -
 277                                         rem_size - RTE_SZE2_ALIGN8(packet_size);
 278                         } else {
 279                                 /* get pointer and length from first part */
 280                                 packet_ptr1 = sze->ct_rx_cur_ptr +
 281                                         hw_data_align;
 282                                 packet_len1 = sze->ct_rx_rem_bytes -
 283                                         hw_data_align;
 284
 285                                 /* MOVE to next lock */
 286                                 sze->ct_rx_lck = sze->ct_rx_lck->next;
 287                                 sze->ct_rx_cur_ptr = sze->ct_rx_lck->start;
 288
 289                                 /* get pointer and length from second part */
 290                                 packet_ptr2 = sze->ct_rx_cur_ptr;
 291                                 packet_len2 = packet_size - packet_len1;
 292
 293                                 sze->ct_rx_cur_ptr +=
 294                                         RTE_SZE2_ALIGN8(packet_size) -
 295                                         packet_len1;
 296                                 sze->ct_rx_rem_bytes = sze->ct_rx_lck->len -
 297                                         (RTE_SZE2_ALIGN8(packet_size) -
 298                                          packet_len1);
 299                         }
 300                 }
 301
 302                 if (unlikely(packet_ptr1 == NULL)) {
 303                         rte_pktmbuf_free(mbuf);
 304                         break;
 305                 }
 306
 307                 /* get the space available for data in the mbuf */
 308                 mbp_priv = rte_mempool_get_priv(sze_q->mb_pool);
 309                 buf_size = (uint16_t)(mbp_priv->mbuf_data_room_size -
 310                                 RTE_PKTMBUF_HEADROOM);
 311
 312                 if (packet_size <= buf_size) {
 313                         /* sze packet will fit in one mbuf, go ahead and copy */
 314                         rte_memcpy(rte_pktmbuf_mtod(mbuf, void *),
 315                                         packet_ptr1, packet_len1);
 316                         if (packet_ptr2 != NULL) {
 317                                 rte_memcpy((void *)(rte_pktmbuf_mtod(mbuf,
 318                                         uint8_t *) + packet_len1),
 319                                         packet_ptr2, packet_len2);
 320                         }
 321                         mbuf->data_len = (uint16_t)packet_size;
 322
 323                         mbuf->pkt_len = packet_size;
 324                         mbuf->port = sze_q->in_port;
 325                         bufs[num_rx] = mbuf;
 326                         num_rx++;
 327                         num_bytes += packet_size;
 328                 } else {
 329                         /*
 330                          * sze packet will not fit in one mbuf,
 331                          * scattered mode is not enabled, drop packet
 332                          */
 333                         PMD_DRV_LOG(ERR,
 334                                 "SZE segment %d bytes will not fit in one mbuf "
 335                                 "(%d bytes), scattered mode is not enabled, "
 336                                 "drop packet!!",
 337                                 packet_size, buf_size);
 338                         rte_pktmbuf_free(mbuf);
 339                 }
 340         }
 341
 342         sze_q->rx_pkts += num_rx;
 343         sze_q->rx_bytes += num_bytes;
 344         return num_rx;
 345 }
 346
 347 static uint16_t
 348 eth_szedata2_rx_scattered(void *queue,
 349                 struct rte_mbuf **bufs,
 350                 uint16_t nb_pkts)
 351 {
 352         unsigned int i;
 353         struct rte_mbuf *mbuf;
 354         struct szedata2_rx_queue *sze_q = queue;
 355         struct rte_pktmbuf_pool_private *mbp_priv;
 356         uint16_t num_rx = 0;
 357         uint16_t buf_size;
 358         uint16_t sg_size;
 359         uint16_t hw_size;
 360         uint16_t packet_size;
 361         uint64_t num_bytes = 0;
 362         struct szedata *sze = sze_q->sze;
 363         uint8_t *header_ptr = NULL; /* header of packet */
 364         uint8_t *packet_ptr1 = NULL;
 365         uint8_t *packet_ptr2 = NULL;
 366         uint16_t packet_len1 = 0;
 367         uint16_t packet_len2 = 0;
 368         uint16_t hw_data_align;
 369         uint64_t *mbuf_failed_ptr =
 370                 &sze_q->priv->dev->data->rx_mbuf_alloc_failed;
 371
 372         if (unlikely(sze_q->sze == NULL || nb_pkts == 0))
 373                 return 0;
 374
 375         /*
 376          * Reads the given number of packets from szedata2 channel given
 377          * by queue and copies the packet data into a newly allocated mbuf
 378          * to return.
 379          */
 380         for (i = 0; i < nb_pkts; i++) {
 381                 const struct szedata_lock *ct_rx_lck_backup;
 382                 unsigned int ct_rx_rem_bytes_backup;
 383                 unsigned char *ct_rx_cur_ptr_backup;
 384
 385                 /* get the next sze packet */
 386                 if (sze->ct_rx_lck != NULL && !sze->ct_rx_rem_bytes &&
 387                                 sze->ct_rx_lck->next == NULL) {
 388                         /* unlock old data */
 389                         szedata_rx_unlock_data(sze_q->sze, sze->ct_rx_lck_orig);
 390                         sze->ct_rx_lck_orig = NULL;
 391                         sze->ct_rx_lck = NULL;
 392                 }
 393
 394                 /*
 395                  * Store items from sze structure which can be changed
 396                  * before mbuf allocating. Use these items in case of mbuf
 397                  * allocating failure.
 398                  */
 399                 ct_rx_lck_backup = sze->ct_rx_lck;
 400                 ct_rx_rem_bytes_backup = sze->ct_rx_rem_bytes;
 401                 ct_rx_cur_ptr_backup = sze->ct_rx_cur_ptr;
 402
 403                 if (!sze->ct_rx_rem_bytes && sze->ct_rx_lck_orig == NULL) {
 404                         /* nothing to read, lock new data */
 405                         sze->ct_rx_lck = szedata_rx_lock_data(sze_q->sze, ~0U);
 406                         sze->ct_rx_lck_orig = sze->ct_rx_lck;
 407
 408                         /*
 409                          * Backup items from sze structure must be updated
 410                          * after locking to contain pointers to new locks.
 411                          */
 412                         ct_rx_lck_backup = sze->ct_rx_lck;
 413                         ct_rx_rem_bytes_backup = sze->ct_rx_rem_bytes;
 414                         ct_rx_cur_ptr_backup = sze->ct_rx_cur_ptr;
 415
 416                         if (sze->ct_rx_lck == NULL)
 417                                 /* nothing to lock */
 418                                 break;
 419
 420                         sze->ct_rx_cur_ptr = sze->ct_rx_lck->start;
 421                         sze->ct_rx_rem_bytes = sze->ct_rx_lck->len;
 422
 423                         if (!sze->ct_rx_rem_bytes)
 424                                 break;
 425                 }
 426
 427                 if (sze->ct_rx_rem_bytes < RTE_SZE2_PACKET_HEADER_SIZE) {
 428                         /*
 429                          * cut in header - copy parts of header to merge buffer
 430                          */
 431                         if (sze->ct_rx_lck->next == NULL)
 432                                 break;
 433
 434                         /* copy first part of header */
 435                         rte_memcpy(sze->ct_rx_buffer, sze->ct_rx_cur_ptr,
 436                                         sze->ct_rx_rem_bytes);
 437
 438                         /* copy second part of header */
 439                         sze->ct_rx_lck = sze->ct_rx_lck->next;
 440                         sze->ct_rx_cur_ptr = sze->ct_rx_lck->start;
 441                         rte_memcpy(sze->ct_rx_buffer + sze->ct_rx_rem_bytes,
 442                                 sze->ct_rx_cur_ptr,
 443                                 RTE_SZE2_PACKET_HEADER_SIZE -
 444                                 sze->ct_rx_rem_bytes);
 445
 446                         sze->ct_rx_cur_ptr += RTE_SZE2_PACKET_HEADER_SIZE -
 447                                 sze->ct_rx_rem_bytes;
 448                         sze->ct_rx_rem_bytes = sze->ct_rx_lck->len -
 449                                 RTE_SZE2_PACKET_HEADER_SIZE +
 450                                 sze->ct_rx_rem_bytes;
 451
 452                         header_ptr = (uint8_t *)sze->ct_rx_buffer;
 453                 } else {
 454                         /* not cut */
 455                         header_ptr = (uint8_t *)sze->ct_rx_cur_ptr;
 456                         sze->ct_rx_cur_ptr += RTE_SZE2_PACKET_HEADER_SIZE;
 457                         sze->ct_rx_rem_bytes -= RTE_SZE2_PACKET_HEADER_SIZE;
 458                 }
 459
 460                 sg_size = le16toh(*((uint16_t *)header_ptr));
 461                 hw_size = le16toh(*(((uint16_t *)header_ptr) + 1));
 462                 packet_size = sg_size -
 463                         RTE_SZE2_ALIGN8(RTE_SZE2_PACKET_HEADER_SIZE + hw_size);
 464
 465
 466                 /* checks if packet all right */
 467                 if (!sg_size)
 468                         errx(5, "Zero segsize");
 469
 470                 /* check sg_size and hwsize */
 471                 if (hw_size > sg_size - RTE_SZE2_PACKET_HEADER_SIZE) {
 472                         errx(10, "Hwsize bigger than expected. Segsize: %d, "
 473                                         "hwsize: %d", sg_size, hw_size);
 474                 }
 475
 476                 hw_data_align =
 477                         RTE_SZE2_ALIGN8((RTE_SZE2_PACKET_HEADER_SIZE +
 478                         hw_size)) - RTE_SZE2_PACKET_HEADER_SIZE;
 479
 480                 if (sze->ct_rx_rem_bytes >=
 481                                 (uint16_t)(sg_size -
 482                                 RTE_SZE2_PACKET_HEADER_SIZE)) {
 483                         /* no cut */
 484                         /* one packet ready - go to another */
 485                         packet_ptr1 = sze->ct_rx_cur_ptr + hw_data_align;
 486                         packet_len1 = packet_size;
 487                         packet_ptr2 = NULL;
 488                         packet_len2 = 0;
 489
 490                         sze->ct_rx_cur_ptr += RTE_SZE2_ALIGN8(sg_size) -
 491                                 RTE_SZE2_PACKET_HEADER_SIZE;
 492                         sze->ct_rx_rem_bytes -= RTE_SZE2_ALIGN8(sg_size) -
 493                                 RTE_SZE2_PACKET_HEADER_SIZE;
 494                 } else {
 495                         /* cut in data */
 496                         if (sze->ct_rx_lck->next == NULL) {
 497                                 errx(6, "Need \"next\" lock, but it is "
 498                                         "missing: %u", sze->ct_rx_rem_bytes);
 499                         }
 500
 501                         /* skip hw data */
 502                         if (sze->ct_rx_rem_bytes <= hw_data_align) {
 503                                 uint16_t rem_size = hw_data_align -
 504                                         sze->ct_rx_rem_bytes;
 505
 506                                 /* MOVE to next lock */
 507                                 sze->ct_rx_lck = sze->ct_rx_lck->next;
 508                                 sze->ct_rx_cur_ptr =
 509                                         (void *)(((uint8_t *)
 510                                         (sze->ct_rx_lck->start)) + rem_size);
 511
 512                                 packet_ptr1 = sze->ct_rx_cur_ptr;
 513                                 packet_len1 = packet_size;
 514                                 packet_ptr2 = NULL;
 515                                 packet_len2 = 0;
 516
 517                                 sze->ct_rx_cur_ptr +=
 518                                         RTE_SZE2_ALIGN8(packet_size);
 519                                 sze->ct_rx_rem_bytes = sze->ct_rx_lck->len -
 520                                         rem_size - RTE_SZE2_ALIGN8(packet_size);
 521                         } else {
 522                                 /* get pointer and length from first part */
 523                                 packet_ptr1 = sze->ct_rx_cur_ptr +
 524                                         hw_data_align;
 525                                 packet_len1 = sze->ct_rx_rem_bytes -
 526                                         hw_data_align;
 527
 528                                 /* MOVE to next lock */
 529                                 sze->ct_rx_lck = sze->ct_rx_lck->next;
 530                                 sze->ct_rx_cur_ptr = sze->ct_rx_lck->start;
 531
 532                                 /* get pointer and length from second part */
 533                                 packet_ptr2 = sze->ct_rx_cur_ptr;
 534                                 packet_len2 = packet_size - packet_len1;
 535
 536                                 sze->ct_rx_cur_ptr +=
 537                                         RTE_SZE2_ALIGN8(packet_size) -
 538                                         packet_len1;
 539                                 sze->ct_rx_rem_bytes = sze->ct_rx_lck->len -
 540                                         (RTE_SZE2_ALIGN8(packet_size) -
 541                                          packet_len1);
 542                         }
 543                 }
 544
 545                 if (unlikely(packet_ptr1 == NULL))
 546                         break;
 547
 548                 mbuf = rte_pktmbuf_alloc(sze_q->mb_pool);
 549
 550                 if (unlikely(mbuf == NULL)) {
 551                         /*
 552                          * Restore items from sze structure to state after
 553                          * unlocking (eventually locking).
 554                          */
 555                         sze->ct_rx_lck = ct_rx_lck_backup;
 556                         sze->ct_rx_rem_bytes = ct_rx_rem_bytes_backup;
 557                         sze->ct_rx_cur_ptr = ct_rx_cur_ptr_backup;
 558                         sze_q->priv->dev->data->rx_mbuf_alloc_failed++;
 559                         break;
 560                 }
 561
 562                 /* get the space available for data in the mbuf */
 563                 mbp_priv = rte_mempool_get_priv(sze_q->mb_pool);
 564                 buf_size = (uint16_t)(mbp_priv->mbuf_data_room_size -
 565                                 RTE_PKTMBUF_HEADROOM);
 566
 567                 if (packet_size <= buf_size) {
 568                         /* sze packet will fit in one mbuf, go ahead and copy */
 569                         rte_memcpy(rte_pktmbuf_mtod(mbuf, void *),
 570                                         packet_ptr1, packet_len1);
 571                         if (packet_ptr2 != NULL) {
 572                                 rte_memcpy((void *)
 573                                         (rte_pktmbuf_mtod(mbuf, uint8_t *) +
 574                                         packet_len1), packet_ptr2, packet_len2);
 575                         }
 576                         mbuf->data_len = (uint16_t)packet_size;
 577                 } else {
 578                         /*
 579                          * sze packet will not fit in one mbuf,
 580                          * scatter packet into more mbufs
 581                          */
 582                         struct rte_mbuf *m = mbuf;
 583                         uint16_t len = rte_pktmbuf_tailroom(mbuf);
 584
 585                         /* copy first part of packet */
 586                         /* fill first mbuf */
 587                         rte_memcpy(rte_pktmbuf_append(mbuf, len), packet_ptr1,
 588                                 len);
 589                         packet_len1 -= len;
 590                         packet_ptr1 = ((uint8_t *)packet_ptr1) + len;
 591
 592                         while (packet_len1 > 0) {
 593                                 /* fill new mbufs */
 594                                 m->next = rte_pktmbuf_alloc(sze_q->mb_pool);
 595
 596                                 if (unlikely(m->next == NULL)) {
 597                                         rte_pktmbuf_free(mbuf);
 598                                         /*
 599                                          * Restore items from sze structure
 600                                          * to state after unlocking (eventually
 601                                          * locking).
 602                                          */
 603                                         sze->ct_rx_lck = ct_rx_lck_backup;
 604                                         sze->ct_rx_rem_bytes =
 605                                                 ct_rx_rem_bytes_backup;
 606                                         sze->ct_rx_cur_ptr =
 607                                                 ct_rx_cur_ptr_backup;
 608                                         (*mbuf_failed_ptr)++;
 609                                         goto finish;
 610                                 }
 611
 612                                 m = m->next;
 613
 614                                 len = RTE_MIN(rte_pktmbuf_tailroom(m),
 615                                         packet_len1);
 616                                 rte_memcpy(rte_pktmbuf_append(mbuf, len),
 617                                         packet_ptr1, len);
 618
 619                                 (mbuf->nb_segs)++;
 620                                 packet_len1 -= len;
 621                                 packet_ptr1 = ((uint8_t *)packet_ptr1) + len;
 622                         }
 623
 624                         if (packet_ptr2 != NULL) {
 625                                 /* copy second part of packet, if exists */
 626                                 /* fill the rest of currently last mbuf */
 627                                 len = rte_pktmbuf_tailroom(m);
 628                                 rte_memcpy(rte_pktmbuf_append(mbuf, len),
 629                                         packet_ptr2, len);
 630                                 packet_len2 -= len;
 631                                 packet_ptr2 = ((uint8_t *)packet_ptr2) + len;
 632
 633                                 while (packet_len2 > 0) {
 634                                         /* fill new mbufs */
 635                                         m->next = rte_pktmbuf_alloc(
 636                                                         sze_q->mb_pool);
 637
 638                                         if (unlikely(m->next == NULL)) {
 639                                                 rte_pktmbuf_free(mbuf);
 640                                                 /*
 641                                                  * Restore items from sze
 642                                                  * structure to state after
 643                                                  * unlocking (eventually
 644                                                  * locking).
 645                                                  */
 646                                                 sze->ct_rx_lck =
 647                                                         ct_rx_lck_backup;
 648                                                 sze->ct_rx_rem_bytes =
 649                                                         ct_rx_rem_bytes_backup;
 650                                                 sze->ct_rx_cur_ptr =
 651                                                         ct_rx_cur_ptr_backup;
 652                                                 (*mbuf_failed_ptr)++;
 653                                                 goto finish;
 654                                         }
 655
 656                                         m = m->next;
 657
 658                                         len = RTE_MIN(rte_pktmbuf_tailroom(m),
 659                                                 packet_len2);
 660                                         rte_memcpy(
 661                                                 rte_pktmbuf_append(mbuf, len),
 662                                                 packet_ptr2, len);
 663
 664                                         (mbuf->nb_segs)++;
 665                                         packet_len2 -= len;
 666                                         packet_ptr2 = ((uint8_t *)packet_ptr2) +
 667                                                 len;
 668                                 }
 669                         }
 670                 }
 671                 mbuf->pkt_len = packet_size;
 672                 mbuf->port = sze_q->in_port;
 673                 bufs[num_rx] = mbuf;
 674                 num_rx++;
 675                 num_bytes += packet_size;
 676         }
 677
 678 finish:
 679         sze_q->rx_pkts += num_rx;
 680         sze_q->rx_bytes += num_bytes;
 681         return num_rx;
 682 }
 683
 684 static uint16_t
 685 eth_szedata2_tx(void *queue,
 686                 struct rte_mbuf **bufs,
 687                 uint16_t nb_pkts)
 688 {
 689         struct rte_mbuf *mbuf;
 690         struct szedata2_tx_queue *sze_q = queue;
 691         uint16_t num_tx = 0;
 692         uint64_t num_bytes = 0;
 693
 694         const struct szedata_lock *lck;
 695         uint32_t lock_size;
 696         uint32_t lock_size2;
 697         void *dst;
 698         uint32_t pkt_len;
 699         uint32_t hwpkt_len;
 700         uint32_t unlock_size;
 701         uint32_t rem_len;
 702         uint16_t mbuf_segs;
 703         uint16_t pkt_left = nb_pkts;
 704
 705         if (sze_q->sze == NULL || nb_pkts == 0)
 706                 return 0;
 707
 708         while (pkt_left > 0) {
 709                 unlock_size = 0;
 710                 lck = szedata_tx_lock_data(sze_q->sze,
 711                         RTE_ETH_SZEDATA2_TX_LOCK_SIZE,
 712                         sze_q->tx_channel);
 713                 if (lck == NULL)
 714                         continue;
 715
 716                 dst = lck->start;
 717                 lock_size = lck->len;
 718                 lock_size2 = lck->next ? lck->next->len : 0;
 719
 720 next_packet:
 721                 mbuf = bufs[nb_pkts - pkt_left];
 722
 723                 pkt_len = mbuf->pkt_len;
 724                 mbuf_segs = mbuf->nb_segs;
 725
 726                 hwpkt_len = RTE_SZE2_PACKET_HEADER_SIZE_ALIGNED +
 727                         RTE_SZE2_ALIGN8(pkt_len);
 728
 729                 if (lock_size + lock_size2 < hwpkt_len) {
 730                         szedata_tx_unlock_data(sze_q->sze, lck, unlock_size);
 731                         continue;
 732                 }
 733
 734                 num_bytes += pkt_len;
 735
 736                 if (lock_size > hwpkt_len) {
 737                         void *tmp_dst;
 738
 739                         rem_len = 0;
 740
 741                         /* write packet length at first 2 bytes in 8B header */
 742                         *((uint16_t *)dst) = htole16(
 743                                         RTE_SZE2_PACKET_HEADER_SIZE_ALIGNED +
 744                                         pkt_len);
 745                         *(((uint16_t *)dst) + 1) = htole16(0);
 746
 747                         /* copy packet from mbuf */
 748                         tmp_dst = ((uint8_t *)(dst)) +
 749                                 RTE_SZE2_PACKET_HEADER_SIZE_ALIGNED;
 750                         if (mbuf_segs == 1) {
 751                                 /*
 752                                  * non-scattered packet,
 753                                  * transmit from one mbuf
 754                                  */
 755                                 rte_memcpy(tmp_dst,
 756                                         rte_pktmbuf_mtod(mbuf, const void *),
 757                                         pkt_len);
 758                         } else {
 759                                 /* scattered packet, transmit from more mbufs */
 760                                 struct rte_mbuf *m = mbuf;
 761                                 while (m) {
 762                                         rte_memcpy(tmp_dst,
 763                                                 rte_pktmbuf_mtod(m,
 764                                                 const void *),
 765                                                 m->data_len);
 766                                         tmp_dst = ((uint8_t *)(tmp_dst)) +
 767                                                 m->data_len;
 768                                         m = m->next;
 769                                 }
 770                         }
 771
 772
 773                         dst = ((uint8_t *)dst) + hwpkt_len;
 774                         unlock_size += hwpkt_len;
 775                         lock_size -= hwpkt_len;
 776
 777                         rte_pktmbuf_free(mbuf);
 778                         num_tx++;
 779                         pkt_left--;
 780                         if (pkt_left == 0) {
 781                                 szedata_tx_unlock_data(sze_q->sze, lck,
 782                                         unlock_size);
 783                                 break;
 784                         }
 785                         goto next_packet;
 786                 } else if (lock_size + lock_size2 >= hwpkt_len) {
 787                         void *tmp_dst;
 788                         uint16_t write_len;
 789
 790                         /* write packet length at first 2 bytes in 8B header */
 791                         *((uint16_t *)dst) =
 792                                 htole16(RTE_SZE2_PACKET_HEADER_SIZE_ALIGNED +
 793                                         pkt_len);
 794                         *(((uint16_t *)dst) + 1) = htole16(0);
 795
 796                         /*
 797                          * If the raw packet (pkt_len) is smaller than lock_size
 798                          * get the correct length for memcpy
 799                          */
 800                         write_len =
 801                                 pkt_len < lock_size -
 802                                 RTE_SZE2_PACKET_HEADER_SIZE_ALIGNED ?
 803                                 pkt_len :
 804                                 lock_size - RTE_SZE2_PACKET_HEADER_SIZE_ALIGNED;
 805
 806                         rem_len = hwpkt_len - lock_size;
 807
 808                         tmp_dst = ((uint8_t *)(dst)) +
 809                                 RTE_SZE2_PACKET_HEADER_SIZE_ALIGNED;
 810                         if (mbuf_segs == 1) {
 811                                 /*
 812                                  * non-scattered packet,
 813                                  * transmit from one mbuf
 814                                  */
 815                                 /* copy part of packet to first area */
 816                                 rte_memcpy(tmp_dst,
 817                                         rte_pktmbuf_mtod(mbuf, const void *),
 818                                         write_len);
 819
 820                                 if (lck->next)
 821                                         dst = lck->next->start;
 822
 823                                 /* copy part of packet to second area */
 824                                 rte_memcpy(dst,
 825                                         (const void *)(rte_pktmbuf_mtod(mbuf,
 826                                                         const uint8_t *) +
 827                                         write_len), pkt_len - write_len);
 828                         } else {
 829                                 /* scattered packet, transmit from more mbufs */
 830                                 struct rte_mbuf *m = mbuf;
 831                                 uint16_t written = 0;
 832                                 uint16_t to_write = 0;
 833                                 bool new_mbuf = true;
 834                                 uint16_t write_off = 0;
 835
 836                                 /* copy part of packet to first area */
 837                                 while (m && written < write_len) {
 838                                         to_write = RTE_MIN(m->data_len,
 839                                                         write_len - written);
 840                                         rte_memcpy(tmp_dst,
 841                                                 rte_pktmbuf_mtod(m,
 842                                                         const void *),
 843                                                 to_write);
 844
 845                                         tmp_dst = ((uint8_t *)(tmp_dst)) +
 846                                                 to_write;
 847                                         if (m->data_len <= write_len -
 848                                                         written) {
 849                                                 m = m->next;
 850                                                 new_mbuf = true;
 851                                         } else {
 852                                                 new_mbuf = false;
 853                                         }
 854                                         written += to_write;
 855                                 }
 856
 857                                 if (lck->next)
 858                                         dst = lck->next->start;
 859
 860                                 tmp_dst = dst;
 861                                 written = 0;
 862                                 write_off = new_mbuf ? 0 : to_write;
 863
 864                                 /* copy part of packet to second area */
 865                                 while (m && written < pkt_len - write_len) {
 866                                         rte_memcpy(tmp_dst, (const void *)
 867                                                 (rte_pktmbuf_mtod(m,
 868                                                 uint8_t *) + write_off),
 869                                                 m->data_len - write_off);
 870
 871                                         tmp_dst = ((uint8_t *)(tmp_dst)) +
 872                                                 (m->data_len - write_off);
 873                                         written += m->data_len - write_off;
 874                                         m = m->next;
 875                                         write_off = 0;
 876                                 }
 877                         }
 878
 879                         dst = ((uint8_t *)dst) + rem_len;
 880                         unlock_size += hwpkt_len;
 881                         lock_size = lock_size2 - rem_len;
 882                         lock_size2 = 0;
 883
 884                         rte_pktmbuf_free(mbuf);
 885                         num_tx++;
 886                 }
 887
 888                 szedata_tx_unlock_data(sze_q->sze, lck, unlock_size);
 889                 pkt_left--;
 890         }
 891
 892         sze_q->tx_pkts += num_tx;
 893         sze_q->err_pkts += nb_pkts - num_tx;
 894         sze_q->tx_bytes += num_bytes;
 895         return num_tx;
 896 }
 897
 898 static int
 899 eth_rx_queue_start(struct rte_eth_dev *dev, uint16_t rxq_id)
 900 {
 901         struct szedata2_rx_queue *rxq = dev->data->rx_queues[rxq_id];
 902         int ret;
 903         struct pmd_internals *internals = (struct pmd_internals *)
 904                 dev->data->dev_private;
 905
 906         if (rxq->sze == NULL) {
 907                 uint32_t rx = 1 << rxq->rx_channel;
 908                 uint32_t tx = 0;
 909                 rxq->sze = szedata_open(internals->sze_dev_path);
 910                 if (rxq->sze == NULL)
 911                         return -EINVAL;
 912                 ret = szedata_subscribe3(rxq->sze, &rx, &tx);
 913                 if (ret != 0 || rx == 0)
 914                         goto err;
 915         }
 916
 917         ret = szedata_start(rxq->sze);
 918         if (ret != 0)
 919                 goto err;
 920         dev->data->rx_queue_state[rxq_id] = RTE_ETH_QUEUE_STATE_STARTED;
 921         return 0;
 922
 923 err:
 924         szedata_close(rxq->sze);
 925         rxq->sze = NULL;
 926         return -EINVAL;
 927 }
 928
 929 static int
 930 eth_rx_queue_stop(struct rte_eth_dev *dev, uint16_t rxq_id)
 931 {
 932         struct szedata2_rx_queue *rxq = dev->data->rx_queues[rxq_id];
 933
 934         if (rxq->sze != NULL) {
 935                 szedata_close(rxq->sze);
 936                 rxq->sze = NULL;
 937         }
 938
 939         dev->data->rx_queue_state[rxq_id] = RTE_ETH_QUEUE_STATE_STOPPED;
 940         return 0;
 941 }
 942
 943 static int
 944 eth_tx_queue_start(struct rte_eth_dev *dev, uint16_t txq_id)
 945 {
 946         struct szedata2_tx_queue *txq = dev->data->tx_queues[txq_id];
 947         int ret;
 948         struct pmd_internals *internals = (struct pmd_internals *)
 949                 dev->data->dev_private;
 950
 951         if (txq->sze == NULL) {
 952                 uint32_t rx = 0;
 953                 uint32_t tx = 1 << txq->tx_channel;
 954                 txq->sze = szedata_open(internals->sze_dev_path);
 955                 if (txq->sze == NULL)
 956                         return -EINVAL;
 957                 ret = szedata_subscribe3(txq->sze, &rx, &tx);
 958                 if (ret != 0 || tx == 0)
 959                         goto err;
 960         }
 961
 962         ret = szedata_start(txq->sze);
 963         if (ret != 0)
 964                 goto err;
 965         dev->data->tx_queue_state[txq_id] = RTE_ETH_QUEUE_STATE_STARTED;
 966         return 0;
 967
 968 err:
 969         szedata_close(txq->sze);
 970         txq->sze = NULL;
 971         return -EINVAL;
 972 }
 973
 974 static int
 975 eth_tx_queue_stop(struct rte_eth_dev *dev, uint16_t txq_id)
 976 {
 977         struct szedata2_tx_queue *txq = dev->data->tx_queues[txq_id];
 978
 979         if (txq->sze != NULL) {
 980                 szedata_close(txq->sze);
 981                 txq->sze = NULL;
 982         }
 983
 984         dev->data->tx_queue_state[txq_id] = RTE_ETH_QUEUE_STATE_STOPPED;
 985         return 0;
 986 }
 987
 988 static int
 989 eth_dev_start(struct rte_eth_dev *dev)
 990 {
 991         int ret;
 992         uint16_t i;
 993         uint16_t nb_rx = dev->data->nb_rx_queues;
 994         uint16_t nb_tx = dev->data->nb_tx_queues;
 995
 996         for (i = 0; i < nb_rx; i++) {
 997                 ret = eth_rx_queue_start(dev, i);
 998                 if (ret != 0)
 999                         goto err_rx;
1000         }
1001
1002         for (i = 0; i < nb_tx; i++) {
1003                 ret = eth_tx_queue_start(dev, i);
1004                 if (ret != 0)
1005                         goto err_tx;
1006         }
1007
1008         return 0;
1009
1010 err_tx:
1011         for (i = 0; i < nb_tx; i++)
1012                 eth_tx_queue_stop(dev, i);
1013 err_rx:
1014         for (i = 0; i < nb_rx; i++)
1015                 eth_rx_queue_stop(dev, i);
1016         return ret;
1017 }
1018
1019 static void
1020 eth_dev_stop(struct rte_eth_dev *dev)
1021 {
1022         uint16_t i;
1023         uint16_t nb_rx = dev->data->nb_rx_queues;
1024         uint16_t nb_tx = dev->data->nb_tx_queues;
1025
1026         for (i = 0; i < nb_tx; i++)
1027                 eth_tx_queue_stop(dev, i);
1028
1029         for (i = 0; i < nb_rx; i++)
1030                 eth_rx_queue_stop(dev, i);
1031 }
1032
1033 static int
1034 eth_dev_configure(struct rte_eth_dev *dev)
1035 {
1036         struct rte_eth_dev_data *data = dev->data;
1037         if (data->dev_conf.rxmode.offloads & DEV_RX_OFFLOAD_SCATTER) {
1038                 dev->rx_pkt_burst = eth_szedata2_rx_scattered;
1039                 data->scattered_rx = 1;
1040         } else {
1041                 dev->rx_pkt_burst = eth_szedata2_rx;
1042                 data->scattered_rx = 0;
1043         }
1044         return 0;
1045 }
1046
1047 static void
1048 eth_dev_info(struct rte_eth_dev *dev,
1049                 struct rte_eth_dev_info *dev_info)
1050 {
1051         struct pmd_internals *internals = dev->data->dev_private;
1052
1053         dev_info->if_index = 0;
1054         dev_info->max_mac_addrs = 1;
1055         dev_info->max_rx_pktlen = (uint32_t)-1;
1056         dev_info->max_rx_queues = internals->max_rx_queues;
1057         dev_info->max_tx_queues = internals->max_tx_queues;
1058         dev_info->min_rx_bufsize = 0;
1059         dev_info->rx_offload_capa = DEV_RX_OFFLOAD_SCATTER;
1060         dev_info->tx_offload_capa = 0;
1061         dev_info->rx_queue_offload_capa = 0;
1062         dev_info->tx_queue_offload_capa = 0;
1063         dev_info->speed_capa = ETH_LINK_SPEED_100G;
1064 }
1065
1066 static int
1067 eth_stats_get(struct rte_eth_dev *dev,
1068                 struct rte_eth_stats *stats)
1069 {
1070         uint16_t i;
1071         uint16_t nb_rx = dev->data->nb_rx_queues;
1072         uint16_t nb_tx = dev->data->nb_tx_queues;
1073         uint64_t rx_total = 0;
1074         uint64_t tx_total = 0;
1075         uint64_t tx_err_total = 0;
1076         uint64_t rx_total_bytes = 0;
1077         uint64_t tx_total_bytes = 0;
1078
1079         for (i = 0; i < nb_rx; i++) {
1080                 struct szedata2_rx_queue *rxq = dev->data->rx_queues[i];
1081
1082                 if (i < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
1083                         stats->q_ipackets[i] = rxq->rx_pkts;
1084                         stats->q_ibytes[i] = rxq->rx_bytes;
1085                 }
1086                 rx_total += rxq->rx_pkts;
1087                 rx_total_bytes += rxq->rx_bytes;
1088         }
1089
1090         for (i = 0; i < nb_tx; i++) {
1091                 struct szedata2_tx_queue *txq = dev->data->tx_queues[i];
1092
1093                 if (i < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
1094                         stats->q_opackets[i] = txq->tx_pkts;
1095                         stats->q_obytes[i] = txq->tx_bytes;
1096                 }
1097                 tx_total += txq->tx_pkts;
1098                 tx_total_bytes += txq->tx_bytes;
1099                 tx_err_total += txq->err_pkts;
1100         }
1101
1102         stats->ipackets = rx_total;
1103         stats->opackets = tx_total;
1104         stats->ibytes = rx_total_bytes;
1105         stats->obytes = tx_total_bytes;
1106         stats->oerrors = tx_err_total;
1107         stats->rx_nombuf = dev->data->rx_mbuf_alloc_failed;
1108
1109         return 0;
1110 }
1111
1112 static void
1113 eth_stats_reset(struct rte_eth_dev *dev)
1114 {
1115         uint16_t i;
1116         uint16_t nb_rx = dev->data->nb_rx_queues;
1117         uint16_t nb_tx = dev->data->nb_tx_queues;
1118
1119         for (i = 0; i < nb_rx; i++) {
1120                 struct szedata2_rx_queue *rxq = dev->data->rx_queues[i];
1121                 rxq->rx_pkts = 0;
1122                 rxq->rx_bytes = 0;
1123                 rxq->err_pkts = 0;
1124         }
1125         for (i = 0; i < nb_tx; i++) {
1126                 struct szedata2_tx_queue *txq = dev->data->tx_queues[i];
1127                 txq->tx_pkts = 0;
1128                 txq->tx_bytes = 0;
1129                 txq->err_pkts = 0;
1130         }
1131 }
1132
1133 static void
1134 eth_rx_queue_release(void *q)
1135 {
1136         struct szedata2_rx_queue *rxq = (struct szedata2_rx_queue *)q;
1137
1138         if (rxq != NULL) {
1139                 if (rxq->sze != NULL)
1140                         szedata_close(rxq->sze);
1141                 rte_free(rxq);
1142         }
1143 }
1144
1145 static void
1146 eth_tx_queue_release(void *q)
1147 {
1148         struct szedata2_tx_queue *txq = (struct szedata2_tx_queue *)q;
1149
1150         if (txq != NULL) {
1151                 if (txq->sze != NULL)
1152                         szedata_close(txq->sze);
1153                 rte_free(txq);
1154         }
1155 }
1156
1157 static void
1158 eth_dev_close(struct rte_eth_dev *dev)
1159 {
1160         struct pmd_internals *internals = dev->data->dev_private;
1161         uint16_t i;
1162         uint16_t nb_rx = dev->data->nb_rx_queues;
1163         uint16_t nb_tx = dev->data->nb_tx_queues;
1164
1165         eth_dev_stop(dev);
1166
1167         free(internals->sze_dev_path);
1168
1169         for (i = 0; i < nb_rx; i++) {
1170                 eth_rx_queue_release(dev->data->rx_queues[i]);
1171                 dev->data->rx_queues[i] = NULL;
1172         }
1173         dev->data->nb_rx_queues = 0;
1174         for (i = 0; i < nb_tx; i++) {
1175                 eth_tx_queue_release(dev->data->tx_queues[i]);
1176                 dev->data->tx_queues[i] = NULL;
1177         }
1178         dev->data->nb_tx_queues = 0;
1179
1180         rte_free(dev->data->mac_addrs);
1181         dev->data->mac_addrs = NULL;
1182 }
1183
1184 static int
1185 eth_link_update(struct rte_eth_dev *dev,
1186                 int wait_to_complete __rte_unused)
1187 {
1188         struct rte_eth_link link;
1189
1190         memset(&link, 0, sizeof(link));
1191
1192         link.link_speed = ETH_SPEED_NUM_100G;
1193         link.link_duplex = ETH_LINK_FULL_DUPLEX;
1194         link.link_status = ETH_LINK_UP;
1195         link.link_autoneg = ETH_LINK_FIXED;
1196
1197         rte_eth_linkstatus_set(dev, &link);
1198         return 0;
1199 }
1200
1201 static int
1202 eth_dev_set_link_up(struct rte_eth_dev *dev __rte_unused)
1203 {
1204         PMD_DRV_LOG(WARNING, "Setting link up is not supported.");
1205         return 0;
1206 }
1207
1208 static int
1209 eth_dev_set_link_down(struct rte_eth_dev *dev __rte_unused)
1210 {
1211         PMD_DRV_LOG(WARNING, "Setting link down is not supported.");
1212         return 0;
1213 }
1214
1215 static int
1216 eth_rx_queue_setup(struct rte_eth_dev *dev,
1217                 uint16_t rx_queue_id,
1218                 uint16_t nb_rx_desc __rte_unused,
1219                 unsigned int socket_id,
1220                 const struct rte_eth_rxconf *rx_conf __rte_unused,
1221                 struct rte_mempool *mb_pool)
1222 {
1223         struct szedata2_rx_queue *rxq;
1224         int ret;
1225         struct pmd_internals *internals = dev->data->dev_private;
1226         uint8_t rx_channel = internals->rxq_base_id + rx_queue_id;
1227         uint32_t rx = 1 << rx_channel;
1228         uint32_t tx = 0;
1229
1230         PMD_INIT_FUNC_TRACE();
1231
1232         if (dev->data->rx_queues[rx_queue_id] != NULL) {
1233                 eth_rx_queue_release(dev->data->rx_queues[rx_queue_id]);
1234                 dev->data->rx_queues[rx_queue_id] = NULL;
1235         }
1236
1237         rxq = rte_zmalloc_socket("szedata2 rx queue",
1238                         sizeof(struct szedata2_rx_queue),
1239                         RTE_CACHE_LINE_SIZE, socket_id);
1240         if (rxq == NULL) {
1241                 PMD_INIT_LOG(ERR, "rte_zmalloc_socket() failed for rx queue id "
1242                                 "%" PRIu16 "!", rx_queue_id);
1243                 return -ENOMEM;
1244         }
1245
1246         rxq->priv = internals;
1247         rxq->sze = szedata_open(internals->sze_dev_path);
1248         if (rxq->sze == NULL) {
1249                 PMD_INIT_LOG(ERR, "szedata_open() failed for rx queue id "
1250                                 "%" PRIu16 "!", rx_queue_id);
1251                 eth_rx_queue_release(rxq);
1252                 return -EINVAL;
1253         }
1254         ret = szedata_subscribe3(rxq->sze, &rx, &tx);
1255         if (ret != 0 || rx == 0) {
1256                 PMD_INIT_LOG(ERR, "szedata_subscribe3() failed for rx queue id "
1257                                 "%" PRIu16 "!", rx_queue_id);
1258                 eth_rx_queue_release(rxq);
1259                 return -EINVAL;
1260         }
1261         rxq->rx_channel = rx_channel;
1262         rxq->qid = rx_queue_id;
1263         rxq->in_port = dev->data->port_id;
1264         rxq->mb_pool = mb_pool;
1265         rxq->rx_pkts = 0;
1266         rxq->rx_bytes = 0;
1267         rxq->err_pkts = 0;
1268
1269         dev->data->rx_queues[rx_queue_id] = rxq;
1270
1271         PMD_INIT_LOG(DEBUG, "Configured rx queue id %" PRIu16 " on socket "
1272                         "%u (channel id %u).", rxq->qid, socket_id,
1273                         rxq->rx_channel);
1274
1275         return 0;
1276 }
1277
1278 static int
1279 eth_tx_queue_setup(struct rte_eth_dev *dev,
1280                 uint16_t tx_queue_id,
1281                 uint16_t nb_tx_desc __rte_unused,
1282                 unsigned int socket_id,
1283                 const struct rte_eth_txconf *tx_conf __rte_unused)
1284 {
1285         struct szedata2_tx_queue *txq;
1286         int ret;
1287         struct pmd_internals *internals = dev->data->dev_private;
1288         uint8_t tx_channel = internals->txq_base_id + tx_queue_id;
1289         uint32_t rx = 0;
1290         uint32_t tx = 1 << tx_channel;
1291
1292         PMD_INIT_FUNC_TRACE();
1293
1294         if (dev->data->tx_queues[tx_queue_id] != NULL) {
1295                 eth_tx_queue_release(dev->data->tx_queues[tx_queue_id]);
1296                 dev->data->tx_queues[tx_queue_id] = NULL;
1297         }
1298
1299         txq = rte_zmalloc_socket("szedata2 tx queue",
1300                         sizeof(struct szedata2_tx_queue),
1301                         RTE_CACHE_LINE_SIZE, socket_id);
1302         if (txq == NULL) {
1303                 PMD_INIT_LOG(ERR, "rte_zmalloc_socket() failed for tx queue id "
1304                                 "%" PRIu16 "!", tx_queue_id);
1305                 return -ENOMEM;
1306         }
1307
1308         txq->priv = internals;
1309         txq->sze = szedata_open(internals->sze_dev_path);
1310         if (txq->sze == NULL) {
1311                 PMD_INIT_LOG(ERR, "szedata_open() failed for tx queue id "
1312                                 "%" PRIu16 "!", tx_queue_id);
1313                 eth_tx_queue_release(txq);
1314                 return -EINVAL;
1315         }
1316         ret = szedata_subscribe3(txq->sze, &rx, &tx);
1317         if (ret != 0 || tx == 0) {
1318                 PMD_INIT_LOG(ERR, "szedata_subscribe3() failed for tx queue id "
1319                                 "%" PRIu16 "!", tx_queue_id);
1320                 eth_tx_queue_release(txq);
1321                 return -EINVAL;
1322         }
1323         txq->tx_channel = tx_channel;
1324         txq->qid = tx_queue_id;
1325         txq->tx_pkts = 0;
1326         txq->tx_bytes = 0;
1327         txq->err_pkts = 0;
1328
1329         dev->data->tx_queues[tx_queue_id] = txq;
1330
1331         PMD_INIT_LOG(DEBUG, "Configured tx queue id %" PRIu16 " on socket "
1332                         "%u (channel id %u).", txq->qid, socket_id,
1333                         txq->tx_channel);
1334
1335         return 0;
1336 }
1337
1338 static int
1339 eth_mac_addr_set(struct rte_eth_dev *dev __rte_unused,
1340                 struct rte_ether_addr *mac_addr __rte_unused)
1341 {
1342         return 0;
1343 }
1344
1345 static void
1346 eth_promiscuous_enable(struct rte_eth_dev *dev __rte_unused)
1347 {
1348         PMD_DRV_LOG(WARNING, "Enabling promiscuous mode is not supported. "
1349                         "The card is always in promiscuous mode.");
1350 }
1351
1352 static void
1353 eth_promiscuous_disable(struct rte_eth_dev *dev __rte_unused)
1354 {
1355         PMD_DRV_LOG(WARNING, "Disabling promiscuous mode is not supported. "
1356                         "The card is always in promiscuous mode.");
1357 }
1358
1359 static void
1360 eth_allmulticast_enable(struct rte_eth_dev *dev __rte_unused)
1361 {
1362         PMD_DRV_LOG(WARNING, "Enabling allmulticast mode is not supported.");
1363 }
1364
1365 static void
1366 eth_allmulticast_disable(struct rte_eth_dev *dev __rte_unused)
1367 {
1368         PMD_DRV_LOG(WARNING, "Disabling allmulticast mode is not supported.");
1369 }
1370
1371 static const struct eth_dev_ops ops = {
1372         .dev_start          = eth_dev_start,
1373         .dev_stop           = eth_dev_stop,
1374         .dev_set_link_up    = eth_dev_set_link_up,
1375         .dev_set_link_down  = eth_dev_set_link_down,
1376         .dev_close          = eth_dev_close,
1377         .dev_configure      = eth_dev_configure,
1378         .dev_infos_get      = eth_dev_info,
1379         .promiscuous_enable   = eth_promiscuous_enable,
1380         .promiscuous_disable  = eth_promiscuous_disable,
1381         .allmulticast_enable  = eth_allmulticast_enable,
1382         .allmulticast_disable = eth_allmulticast_disable,
1383         .rx_queue_start     = eth_rx_queue_start,
1384         .rx_queue_stop      = eth_rx_queue_stop,
1385         .tx_queue_start     = eth_tx_queue_start,
1386         .tx_queue_stop      = eth_tx_queue_stop,
1387         .rx_queue_setup     = eth_rx_queue_setup,
1388         .tx_queue_setup     = eth_tx_queue_setup,
1389         .rx_queue_release   = eth_rx_queue_release,
1390         .tx_queue_release   = eth_tx_queue_release,
1391         .link_update        = eth_link_update,
1392         .stats_get          = eth_stats_get,
1393         .stats_reset        = eth_stats_reset,
1394         .mac_addr_set       = eth_mac_addr_set,
1395 };
1396
1397 /*
1398  * This function goes through sysfs and looks for an index of szedata2
1399  * device file (/dev/szedataIIX, where X is the index).
1400  *
1401  * @return
1402  *           0 on success
1403  *          -1 on error
1404  */
1405 static int
1406 get_szedata2_index(const struct rte_pci_addr *pcislot_addr, uint32_t *index)
1407 {
1408         DIR *dir;
1409         struct dirent *entry;
1410         int ret;
1411         uint32_t tmp_index;
1412         FILE *fd;
1413         char pcislot_path[PATH_MAX];
1414         uint32_t domain;
1415         uint8_t bus;
1416         uint8_t devid;
1417         uint8_t function;
1418
1419         dir = opendir("/sys/class/combo");
1420         if (dir == NULL)
1421                 return -1;
1422
1423         /*
1424          * Iterate through all combosixX directories.
1425          * When the value in /sys/class/combo/combosixX/device/pcislot
1426          * file is the location of the ethernet device dev, "X" is the
1427          * index of the device.
1428          */
1429         while ((entry = readdir(dir)) != NULL) {
1430                 ret = sscanf(entry->d_name, "combosix%u", &tmp_index);
1431                 if (ret != 1)
1432                         continue;
1433
1434                 snprintf(pcislot_path, PATH_MAX,
1435                         "/sys/class/combo/combosix%u/device/pcislot",
1436                         tmp_index);
1437
1438                 fd = fopen(pcislot_path, "r");
1439                 if (fd == NULL)
1440                         continue;
1441
1442                 ret = fscanf(fd, "%8" SCNx32 ":%2" SCNx8 ":%2" SCNx8 ".%" SCNx8,
1443                                 &domain, &bus, &devid, &function);
1444                 fclose(fd);
1445                 if (ret != 4)
1446                         continue;
1447
1448                 if (pcislot_addr->domain == domain &&
1449                                 pcislot_addr->bus == bus &&
1450                                 pcislot_addr->devid == devid &&
1451                                 pcislot_addr->function == function) {
1452                         *index = tmp_index;
1453                         closedir(dir);
1454                         return 0;
1455                 }
1456         }
1457
1458         closedir(dir);
1459         return -1;
1460 }
1461
1462 /**
1463  * @brief Initializes rte_eth_dev device.
1464  * @param dev Device to initialize.
1465  * @param pi Structure with info about DMA queues.
1466  * @return 0 on success, negative error code on error.
1467  */
1468 static int
1469 rte_szedata2_eth_dev_init(struct rte_eth_dev *dev, struct port_info *pi)
1470 {
1471         int ret;
1472         uint32_t szedata2_index;
1473         char name[PATH_MAX];
1474         struct rte_eth_dev_data *data = dev->data;
1475         struct pmd_internals *internals = (struct pmd_internals *)
1476                 data->dev_private;
1477         struct rte_pci_device *pci_dev = RTE_ETH_DEV_TO_PCI(dev);
1478
1479         PMD_INIT_FUNC_TRACE();
1480
1481         PMD_INIT_LOG(INFO, "Initializing eth_dev %s (driver %s)", data->name,
1482                         RTE_STR(RTE_SZEDATA2_DRIVER_NAME));
1483
1484         /* Let rte_eth_dev_close() release the port resources */
1485         dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1486
1487         /* Fill internal private structure. */
1488         internals->dev = dev;
1489         /* Get index of szedata2 device file and create path to device file */
1490         ret = get_szedata2_index(&pci_dev->addr, &szedata2_index);
1491         if (ret != 0) {
1492                 PMD_INIT_LOG(ERR, "Failed to get szedata2 device index!");
1493                 return -ENODEV;
1494         }
1495         snprintf(name, PATH_MAX, SZEDATA2_DEV_PATH_FMT, szedata2_index);
1496         internals->sze_dev_path = strdup(name);
1497         if (internals->sze_dev_path == NULL) {
1498                 PMD_INIT_LOG(ERR, "strdup() failed!");
1499                 return -ENOMEM;
1500         }
1501         PMD_INIT_LOG(INFO, "SZEDATA2 path: %s", internals->sze_dev_path);
1502         internals->max_rx_queues = pi->rx_count;
1503         internals->max_tx_queues = pi->tx_count;
1504         internals->rxq_base_id = pi->rx_base_id;
1505         internals->txq_base_id = pi->tx_base_id;
1506         PMD_INIT_LOG(INFO, "%u RX DMA channels from id %u",
1507                         internals->max_rx_queues, internals->rxq_base_id);
1508         PMD_INIT_LOG(INFO, "%u TX DMA channels from id %u",
1509                         internals->max_tx_queues, internals->txq_base_id);
1510
1511         /* Set rx, tx burst functions */
1512         if (data->scattered_rx == 1)
1513                 dev->rx_pkt_burst = eth_szedata2_rx_scattered;
1514         else
1515                 dev->rx_pkt_burst = eth_szedata2_rx;
1516         dev->tx_pkt_burst = eth_szedata2_tx;
1517
1518         /* Set function callbacks for Ethernet API */
1519         dev->dev_ops = &ops;
1520
1521         /* Get link state */
1522         eth_link_update(dev, 0);
1523
1524         /* Allocate space for one mac address */
1525         data->mac_addrs = rte_zmalloc(data->name, sizeof(struct rte_ether_addr),
1526                         RTE_CACHE_LINE_SIZE);
1527         if (data->mac_addrs == NULL) {
1528                 PMD_INIT_LOG(ERR, "Could not alloc space for MAC address!");
1529                 free(internals->sze_dev_path);
1530                 return -ENOMEM;
1531         }
1532
1533         rte_ether_addr_copy(&eth_addr, data->mac_addrs);
1534
1535         PMD_INIT_LOG(INFO, "%s device %s successfully initialized",
1536                         RTE_STR(RTE_SZEDATA2_DRIVER_NAME), data->name);
1537
1538         return 0;
1539 }
1540
1541 /**
1542  * @brief Unitializes rte_eth_dev device.
1543  * @param dev Device to uninitialize.
1544  * @return 0 on success, negative error code on error.
1545  */
1546 static int
1547 rte_szedata2_eth_dev_uninit(struct rte_eth_dev *dev)
1548 {
1549         PMD_INIT_FUNC_TRACE();
1550
1551         eth_dev_close(dev);
1552
1553         PMD_DRV_LOG(INFO, "%s device %s successfully uninitialized",
1554                         RTE_STR(RTE_SZEDATA2_DRIVER_NAME), dev->data->name);
1555
1556         return 0;
1557 }
1558
1559 static const struct rte_pci_id rte_szedata2_pci_id_table[] = {
1560         {
1561                 RTE_PCI_DEVICE(PCI_VENDOR_ID_NETCOPE,
1562                                 PCI_DEVICE_ID_NETCOPE_COMBO80G)
1563         },
1564         {
1565                 RTE_PCI_DEVICE(PCI_VENDOR_ID_NETCOPE,
1566                                 PCI_DEVICE_ID_NETCOPE_COMBO100G)
1567         },
1568         {
1569                 RTE_PCI_DEVICE(PCI_VENDOR_ID_NETCOPE,
1570                                 PCI_DEVICE_ID_NETCOPE_COMBO100G2)
1571         },
1572         {
1573                 RTE_PCI_DEVICE(PCI_VENDOR_ID_NETCOPE,
1574                                 PCI_DEVICE_ID_NETCOPE_NFB200G2QL)
1575         },
1576         {
1577                 RTE_PCI_DEVICE(PCI_VENDOR_ID_SILICOM,
1578                                 PCI_DEVICE_ID_FB2CGG3)
1579         },
1580         {
1581                 RTE_PCI_DEVICE(PCI_VENDOR_ID_SILICOM,
1582                                 PCI_DEVICE_ID_FB2CGG3D)
1583         },
1584         {
1585                 .vendor_id = 0,
1586         }
1587 };
1588
1589 /**
1590  * @brief Gets info about DMA queues for ports.
1591  * @param pci_dev PCI device structure.
1592  * @param port_count Pointer to variable set with number of ports.
1593  * @param pi Pointer to array of structures with info about DMA queues
1594  *           for ports.
1595  * @param max_ports Maximum number of ports.
1596  * @return 0 on success, negative error code on error.
1597  */
1598 static int
1599 get_port_info(struct rte_pci_device *pci_dev, unsigned int *port_count,
1600                 struct port_info *pi, unsigned int max_ports)
1601 {
1602         struct szedata *szedata_temp;
1603         char sze_dev_path[PATH_MAX];
1604         uint32_t szedata2_index;
1605         int ret;
1606         uint16_t max_rx_queues;
1607         uint16_t max_tx_queues;
1608
1609         if (max_ports == 0)
1610                 return -EINVAL;
1611
1612         memset(pi, 0, max_ports * sizeof(struct port_info));
1613         *port_count = 0;
1614
1615         /* Get index of szedata2 device file and create path to device file */
1616         ret = get_szedata2_index(&pci_dev->addr, &szedata2_index);
1617         if (ret != 0) {
1618                 PMD_INIT_LOG(ERR, "Failed to get szedata2 device index!");
1619                 return -ENODEV;
1620         }
1621         snprintf(sze_dev_path, PATH_MAX, SZEDATA2_DEV_PATH_FMT, szedata2_index);
1622
1623         /*
1624          * Get number of available DMA RX and TX channels, which is maximum
1625          * number of queues that can be created.
1626          */
1627         szedata_temp = szedata_open(sze_dev_path);
1628         if (szedata_temp == NULL) {
1629                 PMD_INIT_LOG(ERR, "szedata_open(%s) failed", sze_dev_path);
1630                 return -EINVAL;
1631         }
1632         max_rx_queues = szedata_ifaces_available(szedata_temp, SZE2_DIR_RX);
1633         max_tx_queues = szedata_ifaces_available(szedata_temp, SZE2_DIR_TX);
1634         PMD_INIT_LOG(INFO, "Available DMA channels RX: %u TX: %u",
1635                         max_rx_queues, max_tx_queues);
1636         if (max_rx_queues > RTE_ETH_SZEDATA2_MAX_RX_QUEUES) {
1637                 PMD_INIT_LOG(ERR, "%u RX queues exceeds supported number %u",
1638                                 max_rx_queues, RTE_ETH_SZEDATA2_MAX_RX_QUEUES);
1639                 szedata_close(szedata_temp);
1640                 return -EINVAL;
1641         }
1642         if (max_tx_queues > RTE_ETH_SZEDATA2_MAX_TX_QUEUES) {
1643                 PMD_INIT_LOG(ERR, "%u TX queues exceeds supported number %u",
1644                                 max_tx_queues, RTE_ETH_SZEDATA2_MAX_TX_QUEUES);
1645                 szedata_close(szedata_temp);
1646                 return -EINVAL;
1647         }
1648
1649         if (pci_dev->id.device_id == PCI_DEVICE_ID_NETCOPE_NFB200G2QL) {
1650                 unsigned int i;
1651                 unsigned int rx_queues = max_rx_queues / max_ports;
1652                 unsigned int tx_queues = max_tx_queues / max_ports;
1653
1654                 /*
1655                  * Number of queues reported by szedata_ifaces_available()
1656                  * is the number of all queues from all DMA controllers which
1657                  * may reside at different numa locations.
1658                  * All queues from the same DMA controller have the same numa
1659                  * node.
1660                  * Numa node from the first queue of each DMA controller is
1661                  * retrieved.
1662                  * If the numa node differs from the numa node of the queues
1663                  * from the previous DMA controller the queues are assigned
1664                  * to the next port.
1665                  */
1666
1667                 for (i = 0; i < max_ports; i++) {
1668                         int numa_rx = szedata_get_area_numa_node(szedata_temp,
1669                                 SZE2_DIR_RX, rx_queues * i);
1670                         int numa_tx = szedata_get_area_numa_node(szedata_temp,
1671                                 SZE2_DIR_TX, tx_queues * i);
1672                         unsigned int port_rx_queues = numa_rx != -1 ?
1673                                 rx_queues : 0;
1674                         unsigned int port_tx_queues = numa_tx != -1 ?
1675                                 tx_queues : 0;
1676                         PMD_INIT_LOG(DEBUG, "%u rx queues from id %u, numa %d",
1677                                         rx_queues, rx_queues * i, numa_rx);
1678                         PMD_INIT_LOG(DEBUG, "%u tx queues from id %u, numa %d",
1679                                         tx_queues, tx_queues * i, numa_tx);
1680
1681                         if (port_rx_queues != 0 && port_tx_queues != 0 &&
1682                                         numa_rx != numa_tx) {
1683                                 PMD_INIT_LOG(ERR, "RX queue %u numa %d differs "
1684                                                 "from TX queue %u numa %d "
1685                                                 "unexpectedly",
1686                                                 rx_queues * i, numa_rx,
1687                                                 tx_queues * i, numa_tx);
1688                                 szedata_close(szedata_temp);
1689                                 return -EINVAL;
1690                         } else if (port_rx_queues == 0 && port_tx_queues == 0) {
1691                                 continue;
1692                         } else {
1693                                 unsigned int j;
1694                                 unsigned int current = *port_count;
1695                                 int port_numa = port_rx_queues != 0 ?
1696                                         numa_rx : numa_tx;
1697
1698                                 for (j = 0; j < *port_count; j++) {
1699                                         if (pi[j].numa_node ==
1700                                                         port_numa) {
1701                                                 current = j;
1702                                                 break;
1703                                         }
1704                                 }
1705                                 if (pi[current].rx_count == 0 &&
1706                                                 pi[current].tx_count == 0) {
1707                                         pi[current].rx_base_id = rx_queues * i;
1708                                         pi[current].tx_base_id = tx_queues * i;
1709                                         (*port_count)++;
1710                                 } else if ((rx_queues * i !=
1711                                                 pi[current].rx_base_id +
1712                                                 pi[current].rx_count) ||
1713                                                 (tx_queues * i !=
1714                                                  pi[current].tx_base_id +
1715                                                  pi[current].tx_count)) {
1716                                         PMD_INIT_LOG(ERR, "Queue ids does not "
1717                                                         "fulfill constraints");
1718                                         szedata_close(szedata_temp);
1719                                         return -EINVAL;
1720                                 }
1721                                 pi[current].rx_count += port_rx_queues;
1722                                 pi[current].tx_count += port_tx_queues;
1723                                 pi[current].numa_node = port_numa;
1724                         }
1725                 }
1726         } else {
1727                 pi[0].rx_count = max_rx_queues;
1728                 pi[0].tx_count = max_tx_queues;
1729                 pi[0].numa_node = pci_dev->device.numa_node;
1730                 *port_count = 1;
1731         }
1732
1733         szedata_close(szedata_temp);
1734         return 0;
1735 }
1736
1737 /**
1738  * @brief Allocates rte_eth_dev device.
1739  * @param pci_dev Corresponding PCI device.
1740  * @param numa_node NUMA node on which device is allocated.
1741  * @param port_no Id of rte_eth_device created on PCI device pci_dev.
1742  * @return Pointer to allocated device or NULL on error.
1743  */
1744 static struct rte_eth_dev *
1745 szedata2_eth_dev_allocate(struct rte_pci_device *pci_dev, int numa_node,
1746                 unsigned int port_no)
1747 {
1748         struct rte_eth_dev *eth_dev;
1749         char name[RTE_ETH_NAME_MAX_LEN];
1750
1751         PMD_INIT_FUNC_TRACE();
1752
1753         snprintf(name, RTE_ETH_NAME_MAX_LEN, "%s"
1754                         SZEDATA2_ETH_DEV_NAME_SUFFIX_FMT,
1755                         pci_dev->device.name, port_no);
1756         PMD_INIT_LOG(DEBUG, "Allocating eth_dev %s", name);
1757
1758         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
1759                 eth_dev = rte_eth_dev_allocate(name);
1760                 if (!eth_dev)
1761                         return NULL;
1762
1763                 eth_dev->data->dev_private = rte_zmalloc_socket(name,
1764                         sizeof(struct pmd_internals), RTE_CACHE_LINE_SIZE,
1765                         numa_node);
1766                 if (!eth_dev->data->dev_private) {
1767                         rte_eth_dev_release_port(eth_dev);
1768                         return NULL;
1769                 }
1770         } else {
1771                 eth_dev = rte_eth_dev_attach_secondary(name);
1772                 if (!eth_dev)
1773                         return NULL;
1774         }
1775
1776         eth_dev->device = &pci_dev->device;
1777         rte_eth_copy_pci_info(eth_dev, pci_dev);
1778         eth_dev->data->numa_node = numa_node;
1779         return eth_dev;
1780 }
1781
1782 /**
1783  * @brief Releases interval of rte_eth_dev devices from array.
1784  * @param eth_devs Array of pointers to rte_eth_dev devices.
1785  * @param from Index in array eth_devs to start with.
1786  * @param to Index in array right after the last element to release.
1787  *
1788  * Used for releasing at failed initialization.
1789  */
1790 static void
1791 szedata2_eth_dev_release_interval(struct rte_eth_dev **eth_devs,
1792                 unsigned int from, unsigned int to)
1793 {
1794         unsigned int i;
1795
1796         PMD_INIT_FUNC_TRACE();
1797
1798         for (i = from; i < to; i++) {
1799                 rte_szedata2_eth_dev_uninit(eth_devs[i]);
1800                 rte_eth_dev_pci_release(eth_devs[i]);
1801         }
1802 }
1803
1804 /**
1805  * @brief Callback .probe for struct rte_pci_driver.
1806  */
1807 static int szedata2_eth_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1808         struct rte_pci_device *pci_dev)
1809 {
1810         struct port_info port_info[SZEDATA2_MAX_PORTS];
1811         unsigned int port_count;
1812         int ret;
1813         unsigned int i;
1814         struct pci_dev_list_entry *list_entry;
1815         struct rte_eth_dev *eth_devs[SZEDATA2_MAX_PORTS] = {NULL,};
1816
1817         PMD_INIT_FUNC_TRACE();
1818
1819         ret = get_port_info(pci_dev, &port_count, port_info,
1820                         SZEDATA2_MAX_PORTS);
1821         if (ret != 0)
1822                 return ret;
1823
1824         if (port_count == 0) {
1825                 PMD_INIT_LOG(ERR, "No available ports!");
1826                 return -ENODEV;
1827         }
1828
1829         list_entry = rte_zmalloc(NULL, sizeof(struct pci_dev_list_entry),
1830                         RTE_CACHE_LINE_SIZE);
1831         if (list_entry == NULL) {
1832                 PMD_INIT_LOG(ERR, "rte_zmalloc() failed!");
1833                 return -ENOMEM;
1834         }
1835
1836         for (i = 0; i < port_count; i++) {
1837                 eth_devs[i] = szedata2_eth_dev_allocate(pci_dev,
1838                                 port_info[i].numa_node, i);
1839                 if (eth_devs[i] == NULL) {
1840                         PMD_INIT_LOG(ERR, "Failed to alloc eth_dev for port %u",
1841                                         i);
1842                         szedata2_eth_dev_release_interval(eth_devs, 0, i);
1843                         rte_free(list_entry);
1844                         return -ENOMEM;
1845                 }
1846
1847                 ret = rte_szedata2_eth_dev_init(eth_devs[i], &port_info[i]);
1848                 if (ret != 0) {
1849                         PMD_INIT_LOG(ERR, "Failed to init eth_dev for port %u",
1850                                         i);
1851                         rte_eth_dev_pci_release(eth_devs[i]);
1852                         szedata2_eth_dev_release_interval(eth_devs, 0, i);
1853                         rte_free(list_entry);
1854                         return ret;
1855                 }
1856
1857                 rte_eth_dev_probing_finish(eth_devs[i]);
1858         }
1859
1860         /*
1861          * Add pci_dev to list of PCI devices for this driver
1862          * which is used at remove callback to release all created eth_devs.
1863          */
1864         list_entry->pci_dev = pci_dev;
1865         list_entry->port_count = port_count;
1866         LIST_INSERT_HEAD(&szedata2_pci_dev_list, list_entry, next);
1867         return 0;
1868 }
1869
1870 /**
1871  * @brief Callback .remove for struct rte_pci_driver.
1872  */
1873 static int szedata2_eth_pci_remove(struct rte_pci_device *pci_dev)
1874 {
1875         unsigned int i;
1876         unsigned int port_count;
1877         char name[RTE_ETH_NAME_MAX_LEN];
1878         struct rte_eth_dev *eth_dev;
1879         int ret;
1880         int retval = 0;
1881         bool found = false;
1882         struct pci_dev_list_entry *list_entry = NULL;
1883
1884         PMD_INIT_FUNC_TRACE();
1885
1886         LIST_FOREACH(list_entry, &szedata2_pci_dev_list, next) {
1887                 if (list_entry->pci_dev == pci_dev) {
1888                         port_count = list_entry->port_count;
1889                         found = true;
1890                         break;
1891                 }
1892         }
1893         LIST_REMOVE(list_entry, next);
1894         rte_free(list_entry);
1895
1896         if (!found) {
1897                 PMD_DRV_LOG(ERR, "PCI device " PCI_PRI_FMT " not found",
1898                                 pci_dev->addr.domain, pci_dev->addr.bus,
1899                                 pci_dev->addr.devid, pci_dev->addr.function);
1900                 return -ENODEV;
1901         }
1902
1903         for (i = 0; i < port_count; i++) {
1904                 snprintf(name, RTE_ETH_NAME_MAX_LEN, "%s"
1905                                 SZEDATA2_ETH_DEV_NAME_SUFFIX_FMT,
1906                                 pci_dev->device.name, i);
1907                 PMD_DRV_LOG(DEBUG, "Removing eth_dev %s", name);
1908                 eth_dev = rte_eth_dev_allocated(name);
1909                 if (!eth_dev) {
1910                         PMD_DRV_LOG(ERR, "eth_dev %s not found", name);
1911                         retval = retval ? retval : -ENODEV;
1912                 }
1913
1914                 ret = rte_szedata2_eth_dev_uninit(eth_dev);
1915                 if (ret != 0) {
1916                         PMD_DRV_LOG(ERR, "eth_dev %s uninit failed", name);
1917                         retval = retval ? retval : ret;
1918                 }
1919
1920                 rte_eth_dev_pci_release(eth_dev);
1921         }
1922
1923         return retval;
1924 }
1925
1926 static struct rte_pci_driver szedata2_eth_driver = {
1927         .id_table = rte_szedata2_pci_id_table,
1928         .probe = szedata2_eth_pci_probe,
1929         .remove = szedata2_eth_pci_remove,
1930 };
1931
1932 RTE_PMD_REGISTER_PCI(RTE_SZEDATA2_DRIVER_NAME, szedata2_eth_driver);
1933 RTE_PMD_REGISTER_PCI_TABLE(RTE_SZEDATA2_DRIVER_NAME, rte_szedata2_pci_id_table);
1934 RTE_PMD_REGISTER_KMOD_DEP(RTE_SZEDATA2_DRIVER_NAME,
1935         "* combo6core & combov3 & szedata2 & ( szedata2_cv3 | szedata2_cv3_fdt )");
1936
1937 RTE_INIT(szedata2_init_log)
1938 {
1939         szedata2_logtype_init = rte_log_register("pmd.net.szedata2.init");
1940         if (szedata2_logtype_init >= 0)
1941                 rte_log_set_level(szedata2_logtype_init, RTE_LOG_NOTICE);
1942         szedata2_logtype_driver = rte_log_register("pmd.net.szedata2.driver");
1943         if (szedata2_logtype_driver >= 0)
1944                 rte_log_set_level(szedata2_logtype_driver, RTE_LOG_NOTICE);
1945 }