drivers/baseband/turbo_sw/bbdev_turbo_software.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2017 Intel Corporation
   3  */
   4
   5 #include <string.h>
   6
   7 #include <rte_common.h>
   8 #include <rte_bus_vdev.h>
   9 #include <rte_malloc.h>
  10 #include <rte_ring.h>
  11 #include <rte_kvargs.h>
  12 #include <rte_cycles.h>
  13
  14 #include <rte_bbdev.h>
  15 #include <rte_bbdev_pmd.h>
  16
  17 #ifdef RTE_BBDEV_SDK_AVX2
  18 #include <phy_turbo.h>
  19 #include <phy_crc.h>
  20 #include <phy_rate_match.h>
  21 #endif
  22
  23 #define DRIVER_NAME baseband_turbo_sw
  24
  25 /* Turbo SW PMD logging ID */
  26 static int bbdev_turbo_sw_logtype;
  27
  28 /* Helper macro for logging */
  29 #define rte_bbdev_log(level, fmt, ...) \
  30         rte_log(RTE_LOG_ ## level, bbdev_turbo_sw_logtype, fmt "\n", \
  31                 ##__VA_ARGS__)
  32
  33 #define rte_bbdev_log_debug(fmt, ...) \
  34         rte_bbdev_log(DEBUG, RTE_STR(__LINE__) ":%s() " fmt, __func__, \
  35                 ##__VA_ARGS__)
  36
  37 #define DEINT_INPUT_BUF_SIZE (((RTE_BBDEV_TURBO_MAX_CB_SIZE >> 3) + 1) * 48)
  38 #define DEINT_OUTPUT_BUF_SIZE (DEINT_INPUT_BUF_SIZE * 6)
  39 #define ADAPTER_OUTPUT_BUF_SIZE ((RTE_BBDEV_TURBO_MAX_CB_SIZE + 4) * 48)
  40
  41 /* private data structure */
  42 struct bbdev_private {
  43         unsigned int max_nb_queues;  /**< Max number of queues */
  44 };
  45
  46 /*  Initialisation params structure that can be used by Turbo SW driver */
  47 struct turbo_sw_params {
  48         int socket_id;  /*< Turbo SW device socket */
  49         uint16_t queues_num;  /*< Turbo SW device queues number */
  50 };
  51
  52 /* Accecptable params for Turbo SW devices */
  53 #define TURBO_SW_MAX_NB_QUEUES_ARG  "max_nb_queues"
  54 #define TURBO_SW_SOCKET_ID_ARG      "socket_id"
  55
  56 static const char * const turbo_sw_valid_params[] = {
  57         TURBO_SW_MAX_NB_QUEUES_ARG,
  58         TURBO_SW_SOCKET_ID_ARG
  59 };
  60
  61 /* queue */
  62 struct turbo_sw_queue {
  63         /* Ring for processed (encoded/decoded) operations which are ready to
  64          * be dequeued.
  65          */
  66         struct rte_ring *processed_pkts;
  67         /* Stores input for turbo encoder (used when CRC attachment is
  68          * performed
  69          */
  70         uint8_t *enc_in;
  71         /* Stores output from turbo encoder */
  72         uint8_t *enc_out;
  73         /* Alpha gamma buf for bblib_turbo_decoder() function */
  74         int8_t *ag;
  75         /* Temp buf for bblib_turbo_decoder() function */
  76         uint16_t *code_block;
  77         /* Input buf for bblib_rate_dematching_lte() function */
  78         uint8_t *deint_input;
  79         /* Output buf for bblib_rate_dematching_lte() function */
  80         uint8_t *deint_output;
  81         /* Output buf for bblib_turbodec_adapter_lte() function */
  82         uint8_t *adapter_output;
  83         /* Operation type of this queue */
  84         enum rte_bbdev_op_type type;
  85 } __rte_cache_aligned;
  86
  87 #ifdef RTE_BBDEV_SDK_AVX2
  88 static inline char *
  89 mbuf_append(struct rte_mbuf *m_head, struct rte_mbuf *m, uint16_t len)
  90 {
  91         if (unlikely(len > rte_pktmbuf_tailroom(m)))
  92                 return NULL;
  93
  94         char *tail = (char *)m->buf_addr + m->data_off + m->data_len;
  95         m->data_len = (uint16_t)(m->data_len + len);
  96         m_head->pkt_len  = (m_head->pkt_len + len);
  97         return tail;
  98 }
  99
 100 /* Calculate index based on Table 5.1.3-3 from TS34.212 */
 101 static inline int32_t
 102 compute_idx(uint16_t k)
 103 {
 104         int32_t result = 0;
 105
 106         if (k < RTE_BBDEV_TURBO_MIN_CB_SIZE || k > RTE_BBDEV_TURBO_MAX_CB_SIZE)
 107                 return -1;
 108
 109         if (k > 2048) {
 110                 if ((k - 2048) % 64 != 0)
 111                         result = -1;
 112
 113                 result = 124 + (k - 2048) / 64;
 114         } else if (k <= 512) {
 115                 if ((k - 40) % 8 != 0)
 116                         result = -1;
 117
 118                 result = (k - 40) / 8 + 1;
 119         } else if (k <= 1024) {
 120                 if ((k - 512) % 16 != 0)
 121                         result = -1;
 122
 123                 result = 60 + (k - 512) / 16;
 124         } else { /* 1024 < k <= 2048 */
 125                 if ((k - 1024) % 32 != 0)
 126                         result = -1;
 127
 128                 result = 92 + (k - 1024) / 32;
 129         }
 130
 131         return result;
 132 }
 133 #endif
 134
 135 /* Read flag value 0/1 from bitmap */
 136 static inline bool
 137 check_bit(uint32_t bitmap, uint32_t bitmask)
 138 {
 139         return bitmap & bitmask;
 140 }
 141
 142 /* Get device info */
 143 static void
 144 info_get(struct rte_bbdev *dev, struct rte_bbdev_driver_info *dev_info)
 145 {
 146         struct bbdev_private *internals = dev->data->dev_private;
 147
 148         static const struct rte_bbdev_op_cap bbdev_capabilities[] = {
 149 #ifdef RTE_BBDEV_SDK_AVX2
 150                 {
 151                         .type = RTE_BBDEV_OP_TURBO_DEC,
 152                         .cap.turbo_dec = {
 153                                 .capability_flags =
 154                                         RTE_BBDEV_TURBO_SUBBLOCK_DEINTERLEAVE |
 155                                         RTE_BBDEV_TURBO_POS_LLR_1_BIT_IN |
 156                                         RTE_BBDEV_TURBO_NEG_LLR_1_BIT_IN |
 157                                         RTE_BBDEV_TURBO_CRC_TYPE_24B |
 158                                         RTE_BBDEV_TURBO_DEC_TB_CRC_24B_KEEP |
 159                                         RTE_BBDEV_TURBO_EARLY_TERMINATION,
 160                                 .max_llr_modulus = 16,
 161                                 .num_buffers_src =
 162                                                 RTE_BBDEV_TURBO_MAX_CODE_BLOCKS,
 163                                 .num_buffers_hard_out =
 164                                                 RTE_BBDEV_TURBO_MAX_CODE_BLOCKS,
 165                                 .num_buffers_soft_out = 0,
 166                         }
 167                 },
 168                 {
 169                         .type   = RTE_BBDEV_OP_TURBO_ENC,
 170                         .cap.turbo_enc = {
 171                                 .capability_flags =
 172                                                 RTE_BBDEV_TURBO_CRC_24B_ATTACH |
 173                                                 RTE_BBDEV_TURBO_CRC_24A_ATTACH |
 174                                                 RTE_BBDEV_TURBO_RATE_MATCH |
 175                                                 RTE_BBDEV_TURBO_RV_INDEX_BYPASS,
 176                                 .num_buffers_src =
 177                                                 RTE_BBDEV_TURBO_MAX_CODE_BLOCKS,
 178                                 .num_buffers_dst =
 179                                                 RTE_BBDEV_TURBO_MAX_CODE_BLOCKS,
 180                         }
 181                 },
 182 #endif
 183                 RTE_BBDEV_END_OF_CAPABILITIES_LIST()
 184         };
 185
 186         static struct rte_bbdev_queue_conf default_queue_conf = {
 187                 .queue_size = RTE_BBDEV_QUEUE_SIZE_LIMIT,
 188         };
 189
 190 #ifdef RTE_BBDEV_SDK_AVX2
 191         static const enum rte_cpu_flag_t cpu_flag = RTE_CPUFLAG_SSE4_2;
 192         dev_info->cpu_flag_reqs = &cpu_flag;
 193 #else
 194         dev_info->cpu_flag_reqs = NULL;
 195 #endif
 196
 197         default_queue_conf.socket = dev->data->socket_id;
 198
 199         dev_info->driver_name = RTE_STR(DRIVER_NAME);
 200         dev_info->max_num_queues = internals->max_nb_queues;
 201         dev_info->queue_size_lim = RTE_BBDEV_QUEUE_SIZE_LIMIT;
 202         dev_info->hardware_accelerated = false;
 203         dev_info->max_dl_queue_priority = 0;
 204         dev_info->max_ul_queue_priority = 0;
 205         dev_info->default_queue_conf = default_queue_conf;
 206         dev_info->capabilities = bbdev_capabilities;
 207         dev_info->min_alignment = 64;
 208
 209         rte_bbdev_log_debug("got device info from %u\n", dev->data->dev_id);
 210 }
 211
 212 /* Release queue */
 213 static int
 214 q_release(struct rte_bbdev *dev, uint16_t q_id)
 215 {
 216         struct turbo_sw_queue *q = dev->data->queues[q_id].queue_private;
 217
 218         if (q != NULL) {
 219                 rte_ring_free(q->processed_pkts);
 220                 rte_free(q->enc_out);
 221                 rte_free(q->enc_in);
 222                 rte_free(q->ag);
 223                 rte_free(q->code_block);
 224                 rte_free(q->deint_input);
 225                 rte_free(q->deint_output);
 226                 rte_free(q->adapter_output);
 227                 rte_free(q);
 228                 dev->data->queues[q_id].queue_private = NULL;
 229         }
 230
 231         rte_bbdev_log_debug("released device queue %u:%u",
 232                         dev->data->dev_id, q_id);
 233         return 0;
 234 }
 235
 236 /* Setup a queue */
 237 static int
 238 q_setup(struct rte_bbdev *dev, uint16_t q_id,
 239                 const struct rte_bbdev_queue_conf *queue_conf)
 240 {
 241         int ret;
 242         struct turbo_sw_queue *q;
 243         char name[RTE_RING_NAMESIZE];
 244
 245         /* Allocate the queue data structure. */
 246         q = rte_zmalloc_socket(RTE_STR(DRIVER_NAME), sizeof(*q),
 247                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 248         if (q == NULL) {
 249                 rte_bbdev_log(ERR, "Failed to allocate queue memory");
 250                 return -ENOMEM;
 251         }
 252
 253         /* Allocate memory for encoder output. */
 254         ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"_enc_o%u:%u",
 255                         dev->data->dev_id, q_id);
 256         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 257                 rte_bbdev_log(ERR,
 258                                 "Creating queue name for device %u queue %u failed",
 259                                 dev->data->dev_id, q_id);
 260                 return -ENAMETOOLONG;
 261         }
 262         q->enc_out = rte_zmalloc_socket(name,
 263                         ((RTE_BBDEV_TURBO_MAX_TB_SIZE >> 3) + 3) *
 264                         sizeof(*q->enc_out) * 3,
 265                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 266         if (q->enc_out == NULL) {
 267                 rte_bbdev_log(ERR,
 268                         "Failed to allocate queue memory for %s", name);
 269                 goto free_q;
 270         }
 271
 272         /* Allocate memory for rate matching output. */
 273         ret = snprintf(name, RTE_RING_NAMESIZE,
 274                         RTE_STR(DRIVER_NAME)"_enc_i%u:%u", dev->data->dev_id,
 275                         q_id);
 276         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 277                 rte_bbdev_log(ERR,
 278                                 "Creating queue name for device %u queue %u failed",
 279                                 dev->data->dev_id, q_id);
 280                 return -ENAMETOOLONG;
 281         }
 282         q->enc_in = rte_zmalloc_socket(name,
 283                         (RTE_BBDEV_TURBO_MAX_CB_SIZE >> 3) * sizeof(*q->enc_in),
 284                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 285         if (q->enc_in == NULL) {
 286                 rte_bbdev_log(ERR,
 287                         "Failed to allocate queue memory for %s", name);
 288                 goto free_q;
 289         }
 290
 291         /* Allocate memory for Aplha Gamma temp buffer. */
 292         ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"_ag%u:%u",
 293                         dev->data->dev_id, q_id);
 294         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 295                 rte_bbdev_log(ERR,
 296                                 "Creating queue name for device %u queue %u failed",
 297                                 dev->data->dev_id, q_id);
 298                 return -ENAMETOOLONG;
 299         }
 300         q->ag = rte_zmalloc_socket(name,
 301                         RTE_BBDEV_TURBO_MAX_CB_SIZE * 10 * sizeof(*q->ag),
 302                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 303         if (q->ag == NULL) {
 304                 rte_bbdev_log(ERR,
 305                         "Failed to allocate queue memory for %s", name);
 306                 goto free_q;
 307         }
 308
 309         /* Allocate memory for code block temp buffer. */
 310         ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"_cb%u:%u",
 311                         dev->data->dev_id, q_id);
 312         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 313                 rte_bbdev_log(ERR,
 314                                 "Creating queue name for device %u queue %u failed",
 315                                 dev->data->dev_id, q_id);
 316                 return -ENAMETOOLONG;
 317         }
 318         q->code_block = rte_zmalloc_socket(name,
 319                         RTE_BBDEV_TURBO_MAX_CB_SIZE * sizeof(*q->code_block),
 320                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 321         if (q->code_block == NULL) {
 322                 rte_bbdev_log(ERR,
 323                         "Failed to allocate queue memory for %s", name);
 324                 goto free_q;
 325         }
 326
 327         /* Allocate memory for Deinterleaver input. */
 328         ret = snprintf(name, RTE_RING_NAMESIZE,
 329                         RTE_STR(DRIVER_NAME)"_de_i%u:%u",
 330                         dev->data->dev_id, q_id);
 331         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 332                 rte_bbdev_log(ERR,
 333                                 "Creating queue name for device %u queue %u failed",
 334                                 dev->data->dev_id, q_id);
 335                 return -ENAMETOOLONG;
 336         }
 337         q->deint_input = rte_zmalloc_socket(name,
 338                         DEINT_INPUT_BUF_SIZE * sizeof(*q->deint_input),
 339                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 340         if (q->deint_input == NULL) {
 341                 rte_bbdev_log(ERR,
 342                         "Failed to allocate queue memory for %s", name);
 343                 goto free_q;
 344         }
 345
 346         /* Allocate memory for Deinterleaver output. */
 347         ret = snprintf(name, RTE_RING_NAMESIZE,
 348                         RTE_STR(DRIVER_NAME)"_de_o%u:%u",
 349                         dev->data->dev_id, q_id);
 350         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 351                 rte_bbdev_log(ERR,
 352                                 "Creating queue name for device %u queue %u failed",
 353                                 dev->data->dev_id, q_id);
 354                 return -ENAMETOOLONG;
 355         }
 356         q->deint_output = rte_zmalloc_socket(NULL,
 357                         DEINT_OUTPUT_BUF_SIZE * sizeof(*q->deint_output),
 358                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 359         if (q->deint_output == NULL) {
 360                 rte_bbdev_log(ERR,
 361                         "Failed to allocate queue memory for %s", name);
 362                 goto free_q;
 363         }
 364
 365         /* Allocate memory for Adapter output. */
 366         ret = snprintf(name, RTE_RING_NAMESIZE,
 367                         RTE_STR(DRIVER_NAME)"_ada_o%u:%u",
 368                         dev->data->dev_id, q_id);
 369         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 370                 rte_bbdev_log(ERR,
 371                                 "Creating queue name for device %u queue %u failed",
 372                                 dev->data->dev_id, q_id);
 373                 return -ENAMETOOLONG;
 374         }
 375         q->adapter_output = rte_zmalloc_socket(NULL,
 376                         ADAPTER_OUTPUT_BUF_SIZE * sizeof(*q->adapter_output),
 377                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 378         if (q->adapter_output == NULL) {
 379                 rte_bbdev_log(ERR,
 380                         "Failed to allocate queue memory for %s", name);
 381                 goto free_q;
 382         }
 383
 384         /* Create ring for packets awaiting to be dequeued. */
 385         ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"%u:%u",
 386                         dev->data->dev_id, q_id);
 387         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 388                 rte_bbdev_log(ERR,
 389                                 "Creating queue name for device %u queue %u failed",
 390                                 dev->data->dev_id, q_id);
 391                 return -ENAMETOOLONG;
 392         }
 393         q->processed_pkts = rte_ring_create(name, queue_conf->queue_size,
 394                         queue_conf->socket, RING_F_SP_ENQ | RING_F_SC_DEQ);
 395         if (q->processed_pkts == NULL) {
 396                 rte_bbdev_log(ERR, "Failed to create ring for %s", name);
 397                 goto free_q;
 398         }
 399
 400         q->type = queue_conf->op_type;
 401
 402         dev->data->queues[q_id].queue_private = q;
 403         rte_bbdev_log_debug("setup device queue %s", name);
 404         return 0;
 405
 406 free_q:
 407         rte_ring_free(q->processed_pkts);
 408         rte_free(q->enc_out);
 409         rte_free(q->enc_in);
 410         rte_free(q->ag);
 411         rte_free(q->code_block);
 412         rte_free(q->deint_input);
 413         rte_free(q->deint_output);
 414         rte_free(q->adapter_output);
 415         rte_free(q);
 416         return -EFAULT;
 417 }
 418
 419 static const struct rte_bbdev_ops pmd_ops = {
 420         .info_get = info_get,
 421         .queue_setup = q_setup,
 422         .queue_release = q_release
 423 };
 424
 425 #ifdef RTE_BBDEV_SDK_AVX2
 426 /* Checks if the encoder input buffer is correct.
 427  * Returns 0 if it's valid, -1 otherwise.
 428  */
 429 static inline int
 430 is_enc_input_valid(const uint16_t k, const int32_t k_idx,
 431                 const uint16_t in_length)
 432 {
 433         if (k_idx < 0) {
 434                 rte_bbdev_log(ERR, "K Index is invalid");
 435                 return -1;
 436         }
 437
 438         if (in_length - (k >> 3) < 0) {
 439                 rte_bbdev_log(ERR,
 440                                 "Mismatch between input length (%u bytes) and K (%u bits)",
 441                                 in_length, k);
 442                 return -1;
 443         }
 444
 445         if (k > RTE_BBDEV_TURBO_MAX_CB_SIZE) {
 446                 rte_bbdev_log(ERR, "CB size (%u) is too big, max: %d",
 447                                 k, RTE_BBDEV_TURBO_MAX_CB_SIZE);
 448                 return -1;
 449         }
 450
 451         return 0;
 452 }
 453
 454 /* Checks if the decoder input buffer is correct.
 455  * Returns 0 if it's valid, -1 otherwise.
 456  */
 457 static inline int
 458 is_dec_input_valid(int32_t k_idx, int16_t kw, int16_t in_length)
 459 {
 460         if (k_idx < 0) {
 461                 rte_bbdev_log(ERR, "K index is invalid");
 462                 return -1;
 463         }
 464
 465         if (in_length < kw) {
 466                 rte_bbdev_log(ERR,
 467                                 "Mismatch between input length (%u) and kw (%u)",
 468                                 in_length, kw);
 469                 return -1;
 470         }
 471
 472         if (kw > RTE_BBDEV_TURBO_MAX_KW) {
 473                 rte_bbdev_log(ERR, "Input length (%u) is too big, max: %d",
 474                                 kw, RTE_BBDEV_TURBO_MAX_KW);
 475                 return -1;
 476         }
 477
 478         return 0;
 479 }
 480 #endif
 481
 482 static inline void
 483 process_enc_cb(struct turbo_sw_queue *q, struct rte_bbdev_enc_op *op,
 484                 uint8_t r, uint8_t c, uint16_t k, uint16_t ncb,
 485                 uint32_t e, struct rte_mbuf *m_in, struct rte_mbuf *m_out_head,
 486                 struct rte_mbuf *m_out, uint16_t in_offset, uint16_t out_offset,
 487                 uint16_t in_length, struct rte_bbdev_stats *q_stats)
 488 {
 489 #ifdef RTE_BBDEV_SDK_AVX2
 490         int ret;
 491         int16_t k_idx;
 492         uint16_t m;
 493         uint8_t *in, *out0, *out1, *out2, *tmp_out, *rm_out;
 494         uint64_t first_3_bytes = 0;
 495         struct rte_bbdev_op_turbo_enc *enc = &op->turbo_enc;
 496         struct bblib_crc_request crc_req;
 497         struct bblib_crc_response crc_resp;
 498         struct bblib_turbo_encoder_request turbo_req;
 499         struct bblib_turbo_encoder_response turbo_resp;
 500         struct bblib_rate_match_dl_request rm_req;
 501         struct bblib_rate_match_dl_response rm_resp;
 502 #ifdef RTE_BBDEV_OFFLOAD_COST
 503         uint64_t start_time;
 504 #else
 505         RTE_SET_USED(q_stats);
 506 #endif
 507
 508         k_idx = compute_idx(k);
 509         in = rte_pktmbuf_mtod_offset(m_in, uint8_t *, in_offset);
 510
 511         /* CRC24A (for TB) */
 512         if ((enc->op_flags & RTE_BBDEV_TURBO_CRC_24A_ATTACH) &&
 513                 (enc->code_block_mode == 1)) {
 514                 ret = is_enc_input_valid(k - 24, k_idx, in_length);
 515                 if (ret != 0) {
 516                         op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 517                         return;
 518                 }
 519                 crc_req.data = in;
 520                 crc_req.len = k - 24;
 521                 /* Check if there is a room for CRC bits if not use
 522                  * the temporary buffer.
 523                  */
 524                 if (mbuf_append(m_in, m_in, 3) == NULL) {
 525                         rte_memcpy(q->enc_in, in, (k - 24) >> 3);
 526                         in = q->enc_in;
 527                 } else {
 528                         /* Store 3 first bytes of next CB as they will be
 529                          * overwritten by CRC bytes. If it is the last CB then
 530                          * there is no point to store 3 next bytes and this
 531                          * if..else branch will be omitted.
 532                          */
 533                         first_3_bytes = *((uint64_t *)&in[(k - 32) >> 3]);
 534                 }
 535
 536                 crc_resp.data = in;
 537 #ifdef RTE_BBDEV_OFFLOAD_COST
 538                 start_time = rte_rdtsc_precise();
 539 #endif
 540                 /* CRC24A generation */
 541                 bblib_lte_crc24a_gen(&crc_req, &crc_resp);
 542 #ifdef RTE_BBDEV_OFFLOAD_COST
 543                 q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time;
 544 #endif
 545         } else if (enc->op_flags & RTE_BBDEV_TURBO_CRC_24B_ATTACH) {
 546                 /* CRC24B */
 547                 ret = is_enc_input_valid(k - 24, k_idx, in_length);
 548                 if (ret != 0) {
 549                         op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 550                         return;
 551                 }
 552                 crc_req.data = in;
 553                 crc_req.len = k - 24;
 554                 /* Check if there is a room for CRC bits if this is the last
 555                  * CB in TB. If not use temporary buffer.
 556                  */
 557                 if ((c - r == 1) && (mbuf_append(m_in, m_in, 3) == NULL)) {
 558                         rte_memcpy(q->enc_in, in, (k - 24) >> 3);
 559                         in = q->enc_in;
 560                 } else if (c - r > 1) {
 561                         /* Store 3 first bytes of next CB as they will be
 562                          * overwritten by CRC bytes. If it is the last CB then
 563                          * there is no point to store 3 next bytes and this
 564                          * if..else branch will be omitted.
 565                          */
 566                         first_3_bytes = *((uint64_t *)&in[(k - 32) >> 3]);
 567                 }
 568
 569                 crc_resp.data = in;
 570 #ifdef RTE_BBDEV_OFFLOAD_COST
 571                 start_time = rte_rdtsc_precise();
 572 #endif
 573                 /* CRC24B generation */
 574                 bblib_lte_crc24b_gen(&crc_req, &crc_resp);
 575 #ifdef RTE_BBDEV_OFFLOAD_COST
 576                 q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time;
 577 #endif
 578         } else {
 579                 ret = is_enc_input_valid(k, k_idx, in_length);
 580                 if (ret != 0) {
 581                         op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 582                         return;
 583                 }
 584         }
 585
 586         /* Turbo encoder */
 587
 588         /* Each bit layer output from turbo encoder is (k+4) bits long, i.e.
 589          * input length + 4 tail bits. That's (k/8) + 1 bytes after rounding up.
 590          * So dst_data's length should be 3*(k/8) + 3 bytes.
 591          * In Rate-matching bypass case outputs pointers passed to encoder
 592          * (out0, out1 and out2) can directly point to addresses of output from
 593          * turbo_enc entity.
 594          */
 595         if (enc->op_flags & RTE_BBDEV_TURBO_RATE_MATCH) {
 596                 out0 = q->enc_out;
 597                 out1 = RTE_PTR_ADD(out0, (k >> 3) + 1);
 598                 out2 = RTE_PTR_ADD(out1, (k >> 3) + 1);
 599         } else {
 600                 out0 = (uint8_t *)mbuf_append(m_out_head, m_out,
 601                                 (k >> 3) * 3 + 2);
 602                 if (out0 == NULL) {
 603                         op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 604                         rte_bbdev_log(ERR,
 605                                         "Too little space in output mbuf");
 606                         return;
 607                 }
 608                 enc->output.length += (k >> 3) * 3 + 2;
 609                 /* rte_bbdev_op_data.offset can be different than the
 610                  * offset of the appended bytes
 611                  */
 612                 out0 = rte_pktmbuf_mtod_offset(m_out, uint8_t *, out_offset);
 613                 out1 = rte_pktmbuf_mtod_offset(m_out, uint8_t *,
 614                                 out_offset + (k >> 3) + 1);
 615                 out2 = rte_pktmbuf_mtod_offset(m_out, uint8_t *,
 616                                 out_offset + 2 * ((k >> 3) + 1));
 617         }
 618
 619         turbo_req.case_id = k_idx;
 620         turbo_req.input_win = in;
 621         turbo_req.length = k >> 3;
 622         turbo_resp.output_win_0 = out0;
 623         turbo_resp.output_win_1 = out1;
 624         turbo_resp.output_win_2 = out2;
 625
 626 #ifdef RTE_BBDEV_OFFLOAD_COST
 627         start_time = rte_rdtsc_precise();
 628 #endif
 629         /* Turbo encoding */
 630         if (bblib_turbo_encoder(&turbo_req, &turbo_resp) != 0) {
 631                 op->status |= 1 << RTE_BBDEV_DRV_ERROR;
 632                 rte_bbdev_log(ERR, "Turbo Encoder failed");
 633                 return;
 634         }
 635 #ifdef RTE_BBDEV_OFFLOAD_COST
 636         q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time;
 637 #endif
 638
 639         /* Restore 3 first bytes of next CB if they were overwritten by CRC*/
 640         if (first_3_bytes != 0)
 641                 *((uint64_t *)&in[(k - 32) >> 3]) = first_3_bytes;
 642
 643         /* Rate-matching */
 644         if (enc->op_flags & RTE_BBDEV_TURBO_RATE_MATCH) {
 645                 uint8_t mask_id;
 646                 /* Integer round up division by 8 */
 647                 uint16_t out_len = (e + 7) >> 3;
 648                 /* The mask array is indexed using E%8. E is an even number so
 649                  * there are only 4 possible values.
 650                  */
 651                 const uint8_t mask_out[] = {0xFF, 0xC0, 0xF0, 0xFC};
 652
 653                 /* get output data starting address */
 654                 rm_out = (uint8_t *)mbuf_append(m_out_head, m_out, out_len);
 655                 if (rm_out == NULL) {
 656                         op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 657                         rte_bbdev_log(ERR,
 658                                         "Too little space in output mbuf");
 659                         return;
 660                 }
 661                 /* rte_bbdev_op_data.offset can be different than the offset
 662                  * of the appended bytes
 663                  */
 664                 rm_out = rte_pktmbuf_mtod_offset(m_out, uint8_t *, out_offset);
 665
 666                 /* index of current code block */
 667                 rm_req.r = r;
 668                 /* total number of code block */
 669                 rm_req.C = c;
 670                 /* For DL - 1, UL - 0 */
 671                 rm_req.direction = 1;
 672                 /* According to 3ggp 36.212 Spec 5.1.4.1.2 section Nsoft, KMIMO
 673                  * and MDL_HARQ are used for Ncb calculation. As Ncb is already
 674                  * known we can adjust those parameters
 675                  */
 676                 rm_req.Nsoft = ncb * rm_req.C;
 677                 rm_req.KMIMO = 1;
 678                 rm_req.MDL_HARQ = 1;
 679                 /* According to 3ggp 36.212 Spec 5.1.4.1.2 section Nl, Qm and G
 680                  * are used for E calculation. As E is already known we can
 681                  * adjust those parameters
 682                  */
 683                 rm_req.NL = e;
 684                 rm_req.Qm = 1;
 685                 rm_req.G = rm_req.NL * rm_req.Qm * rm_req.C;
 686
 687                 rm_req.rvidx = enc->rv_index;
 688                 rm_req.Kidx = k_idx - 1;
 689                 rm_req.nLen = k + 4;
 690                 rm_req.tin0 = out0;
 691                 rm_req.tin1 = out1;
 692                 rm_req.tin2 = out2;
 693                 rm_resp.output = rm_out;
 694                 rm_resp.OutputLen = out_len;
 695                 if (enc->op_flags & RTE_BBDEV_TURBO_RV_INDEX_BYPASS)
 696                         rm_req.bypass_rvidx = 1;
 697                 else
 698                         rm_req.bypass_rvidx = 0;
 699
 700 #ifdef RTE_BBDEV_OFFLOAD_COST
 701                 start_time = rte_rdtsc_precise();
 702 #endif
 703                 /* Rate-Matching */
 704                 if (bblib_rate_match_dl(&rm_req, &rm_resp) != 0) {
 705                         op->status |= 1 << RTE_BBDEV_DRV_ERROR;
 706                         rte_bbdev_log(ERR, "Rate matching failed");
 707                         return;
 708                 }
 709 #ifdef RTE_BBDEV_OFFLOAD_COST
 710                 q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time;
 711 #endif
 712
 713                 /* SW fills an entire last byte even if E%8 != 0. Clear the
 714                  * superfluous data bits for consistency with HW device.
 715                  */
 716                 mask_id = (e & 7) >> 1;
 717                 rm_out[out_len - 1] &= mask_out[mask_id];
 718                 enc->output.length += rm_resp.OutputLen;
 719         } else {
 720                 /* Rate matching is bypassed */
 721
 722                 /* Completing last byte of out0 (where 4 tail bits are stored)
 723                  * by moving first 4 bits from out1
 724                  */
 725                 tmp_out = (uint8_t *) --out1;
 726                 *tmp_out = *tmp_out | ((*(tmp_out + 1) & 0xF0) >> 4);
 727                 tmp_out++;
 728                 /* Shifting out1 data by 4 bits to the left */
 729                 for (m = 0; m < k >> 3; ++m) {
 730                         uint8_t *first = tmp_out;
 731                         uint8_t second = *(tmp_out + 1);
 732                         *first = (*first << 4) | ((second & 0xF0) >> 4);
 733                         tmp_out++;
 734                 }
 735                 /* Shifting out2 data by 8 bits to the left */
 736                 for (m = 0; m < (k >> 3) + 1; ++m) {
 737                         *tmp_out = *(tmp_out + 1);
 738                         tmp_out++;
 739                 }
 740                 *tmp_out = 0;
 741         }
 742 #else
 743         RTE_SET_USED(q);
 744         RTE_SET_USED(op);
 745         RTE_SET_USED(r);
 746         RTE_SET_USED(c);
 747         RTE_SET_USED(k);
 748         RTE_SET_USED(ncb);
 749         RTE_SET_USED(e);
 750         RTE_SET_USED(m_in);
 751         RTE_SET_USED(m_out_head);
 752         RTE_SET_USED(m_out);
 753         RTE_SET_USED(in_offset);
 754         RTE_SET_USED(out_offset);
 755         RTE_SET_USED(in_length);
 756         RTE_SET_USED(q_stats);
 757 #endif
 758 }
 759
 760 static inline void
 761 enqueue_enc_one_op(struct turbo_sw_queue *q, struct rte_bbdev_enc_op *op,
 762                 struct rte_bbdev_stats *queue_stats)
 763 {
 764         uint8_t c, r, crc24_bits = 0;
 765         uint16_t k, ncb;
 766         uint32_t e;
 767         struct rte_bbdev_op_turbo_enc *enc = &op->turbo_enc;
 768         uint16_t in_offset = enc->input.offset;
 769         uint16_t out_offset = enc->output.offset;
 770         struct rte_mbuf *m_in = enc->input.data;
 771         struct rte_mbuf *m_out = enc->output.data;
 772         struct rte_mbuf *m_out_head = enc->output.data;
 773         uint32_t in_length, mbuf_total_left = enc->input.length;
 774         uint16_t seg_total_left;
 775
 776         /* Clear op status */
 777         op->status = 0;
 778
 779         if (mbuf_total_left > RTE_BBDEV_TURBO_MAX_TB_SIZE >> 3) {
 780                 rte_bbdev_log(ERR, "TB size (%u) is too big, max: %d",
 781                                 mbuf_total_left, RTE_BBDEV_TURBO_MAX_TB_SIZE);
 782                 op->status = 1 << RTE_BBDEV_DATA_ERROR;
 783                 return;
 784         }
 785
 786         if (m_in == NULL || m_out == NULL) {
 787                 rte_bbdev_log(ERR, "Invalid mbuf pointer");
 788                 op->status = 1 << RTE_BBDEV_DATA_ERROR;
 789                 return;
 790         }
 791
 792         if ((enc->op_flags & RTE_BBDEV_TURBO_CRC_24B_ATTACH) ||
 793                 (enc->op_flags & RTE_BBDEV_TURBO_CRC_24A_ATTACH))
 794                 crc24_bits = 24;
 795
 796         if (enc->code_block_mode == 0) { /* For Transport Block mode */
 797                 c = enc->tb_params.c;
 798                 r = enc->tb_params.r;
 799         } else {/* For Code Block mode */
 800                 c = 1;
 801                 r = 0;
 802         }
 803
 804         while (mbuf_total_left > 0 && r < c) {
 805
 806                 seg_total_left = rte_pktmbuf_data_len(m_in) - in_offset;
 807
 808                 if (enc->code_block_mode == 0) {
 809                         k = (r < enc->tb_params.c_neg) ?
 810                                 enc->tb_params.k_neg : enc->tb_params.k_pos;
 811                         ncb = (r < enc->tb_params.c_neg) ?
 812                                 enc->tb_params.ncb_neg : enc->tb_params.ncb_pos;
 813                         e = (r < enc->tb_params.cab) ?
 814                                 enc->tb_params.ea : enc->tb_params.eb;
 815                 } else {
 816                         k = enc->cb_params.k;
 817                         ncb = enc->cb_params.ncb;
 818                         e = enc->cb_params.e;
 819                 }
 820
 821                 process_enc_cb(q, op, r, c, k, ncb, e, m_in, m_out_head,
 822                                 m_out, in_offset, out_offset, seg_total_left,
 823                                 queue_stats);
 824                 /* Update total_left */
 825                 in_length = ((k - crc24_bits) >> 3);
 826                 mbuf_total_left -= in_length;
 827                 /* Update offsets for next CBs (if exist) */
 828                 in_offset += (k - crc24_bits) >> 3;
 829                 if (enc->op_flags & RTE_BBDEV_TURBO_RATE_MATCH)
 830                         out_offset += e >> 3;
 831                 else
 832                         out_offset += (k >> 3) * 3 + 2;
 833
 834                 /* Update offsets */
 835                 if (seg_total_left == in_length) {
 836                         /* Go to the next mbuf */
 837                         m_in = m_in->next;
 838                         m_out = m_out->next;
 839                         in_offset = 0;
 840                         out_offset = 0;
 841                 }
 842                 r++;
 843         }
 844
 845         /* check if all input data was processed */
 846         if (mbuf_total_left != 0) {
 847                 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 848                 rte_bbdev_log(ERR,
 849                                 "Mismatch between mbuf length and included CBs sizes");
 850         }
 851 }
 852
 853 static inline uint16_t
 854 enqueue_enc_all_ops(struct turbo_sw_queue *q, struct rte_bbdev_enc_op **ops,
 855                 uint16_t nb_ops, struct rte_bbdev_stats *queue_stats)
 856 {
 857         uint16_t i;
 858 #ifdef RTE_BBDEV_OFFLOAD_COST
 859         queue_stats->acc_offload_cycles = 0;
 860 #endif
 861
 862         for (i = 0; i < nb_ops; ++i)
 863                 enqueue_enc_one_op(q, ops[i], queue_stats);
 864
 865         return rte_ring_enqueue_burst(q->processed_pkts, (void **)ops, nb_ops,
 866                         NULL);
 867 }
 868
 869 #ifdef RTE_BBDEV_SDK_AVX2
 870 static inline void
 871 move_padding_bytes(const uint8_t *in, uint8_t *out, uint16_t k,
 872                 uint16_t ncb)
 873 {
 874         uint16_t d = k + 4;
 875         uint16_t kpi = ncb / 3;
 876         uint16_t nd = kpi - d;
 877
 878         rte_memcpy(&out[nd], in, d);
 879         rte_memcpy(&out[nd + kpi + 64], &in[kpi], d);
 880         rte_memcpy(&out[(nd - 1) + 2 * (kpi + 64)], &in[2 * kpi], d);
 881 }
 882 #endif
 883
 884 static inline void
 885 process_dec_cb(struct turbo_sw_queue *q, struct rte_bbdev_dec_op *op,
 886                 uint8_t c, uint16_t k, uint16_t kw, struct rte_mbuf *m_in,
 887                 struct rte_mbuf *m_out_head, struct rte_mbuf *m_out,
 888                 uint16_t in_offset, uint16_t out_offset, bool check_crc_24b,
 889                 uint16_t crc24_overlap, uint16_t in_length,
 890                 struct rte_bbdev_stats *q_stats)
 891 {
 892 #ifdef RTE_BBDEV_SDK_AVX2
 893         int ret;
 894         int32_t k_idx;
 895         int32_t iter_cnt;
 896         uint8_t *in, *out, *adapter_input;
 897         int32_t ncb, ncb_without_null;
 898         struct bblib_turbo_adapter_ul_response adapter_resp;
 899         struct bblib_turbo_adapter_ul_request adapter_req;
 900         struct bblib_turbo_decoder_request turbo_req;
 901         struct bblib_turbo_decoder_response turbo_resp;
 902         struct rte_bbdev_op_turbo_dec *dec = &op->turbo_dec;
 903 #ifdef RTE_BBDEV_OFFLOAD_COST
 904         uint64_t start_time;
 905 #else
 906         RTE_SET_USED(q_stats);
 907 #endif
 908
 909         k_idx = compute_idx(k);
 910
 911         ret = is_dec_input_valid(k_idx, kw, in_length);
 912         if (ret != 0) {
 913                 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 914                 return;
 915         }
 916
 917         in = rte_pktmbuf_mtod_offset(m_in, uint8_t *, in_offset);
 918         ncb = kw;
 919         ncb_without_null = (k + 4) * 3;
 920
 921         if (check_bit(dec->op_flags, RTE_BBDEV_TURBO_SUBBLOCK_DEINTERLEAVE)) {
 922                 struct bblib_deinterleave_ul_request deint_req;
 923                 struct bblib_deinterleave_ul_response deint_resp;
 924
 925                 deint_req.circ_buffer = BBLIB_FULL_CIRCULAR_BUFFER;
 926                 deint_req.pharqbuffer = in;
 927                 deint_req.ncb = ncb;
 928                 deint_resp.pinteleavebuffer = q->deint_output;
 929
 930 #ifdef RTE_BBDEV_OFFLOAD_COST
 931                 start_time = rte_rdtsc_precise();
 932 #endif
 933                 bblib_deinterleave_ul(&deint_req, &deint_resp);
 934 #ifdef RTE_BBDEV_OFFLOAD_COST
 935                 q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time;
 936 #endif
 937         } else
 938                 move_padding_bytes(in, q->deint_output, k, ncb);
 939
 940         adapter_input = q->deint_output;
 941
 942         if (dec->op_flags & RTE_BBDEV_TURBO_POS_LLR_1_BIT_IN)
 943                 adapter_req.isinverted = 1;
 944         else if (dec->op_flags & RTE_BBDEV_TURBO_NEG_LLR_1_BIT_IN)
 945                 adapter_req.isinverted = 0;
 946         else {
 947                 op->status |= 1 << RTE_BBDEV_DRV_ERROR;
 948                 rte_bbdev_log(ERR, "LLR format wasn't specified");
 949                 return;
 950         }
 951
 952         adapter_req.ncb = ncb_without_null;
 953         adapter_req.pinteleavebuffer = adapter_input;
 954         adapter_resp.pharqout = q->adapter_output;
 955
 956 #ifdef RTE_BBDEV_OFFLOAD_COST
 957         start_time = rte_rdtsc_precise();
 958 #endif
 959         /* Turbo decode adaptation */
 960         bblib_turbo_adapter_ul(&adapter_req, &adapter_resp);
 961 #ifdef RTE_BBDEV_OFFLOAD_COST
 962         q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time;
 963 #endif
 964
 965         out = (uint8_t *)mbuf_append(m_out_head, m_out,
 966                         ((k - crc24_overlap) >> 3));
 967         if (out == NULL) {
 968                 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 969                 rte_bbdev_log(ERR, "Too little space in output mbuf");
 970                 return;
 971         }
 972         /* rte_bbdev_op_data.offset can be different than the offset of the
 973          * appended bytes
 974          */
 975         out = rte_pktmbuf_mtod_offset(m_out, uint8_t *, out_offset);
 976         if (check_crc_24b)
 977                 turbo_req.c = c + 1;
 978         else
 979                 turbo_req.c = c;
 980         turbo_req.input = (int8_t *)q->adapter_output;
 981         turbo_req.k = k;
 982         turbo_req.k_idx = k_idx;
 983         turbo_req.max_iter_num = dec->iter_max;
 984         turbo_req.early_term_disable = !check_bit(dec->op_flags,
 985                         RTE_BBDEV_TURBO_EARLY_TERMINATION);
 986         turbo_resp.ag_buf = q->ag;
 987         turbo_resp.cb_buf = q->code_block;
 988         turbo_resp.output = out;
 989
 990 #ifdef RTE_BBDEV_OFFLOAD_COST
 991         start_time = rte_rdtsc_precise();
 992 #endif
 993         /* Turbo decode */
 994         iter_cnt = bblib_turbo_decoder(&turbo_req, &turbo_resp);
 995 #ifdef RTE_BBDEV_OFFLOAD_COST
 996         q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time;
 997 #endif
 998         dec->hard_output.length += (k >> 3);
 999
1000         if (iter_cnt > 0) {
1001                 /* Temporary solution for returned iter_count from SDK */
1002                 iter_cnt = (iter_cnt - 1) >> 1;
1003                 dec->iter_count = RTE_MAX(iter_cnt, dec->iter_count);
1004         } else {
1005                 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
1006                 rte_bbdev_log(ERR, "Turbo Decoder failed");
1007                 return;
1008         }
1009 #else
1010         RTE_SET_USED(q);
1011         RTE_SET_USED(op);
1012         RTE_SET_USED(c);
1013         RTE_SET_USED(k);
1014         RTE_SET_USED(kw);
1015         RTE_SET_USED(m_in);
1016         RTE_SET_USED(m_out_head);
1017         RTE_SET_USED(m_out);
1018         RTE_SET_USED(in_offset);
1019         RTE_SET_USED(out_offset);
1020         RTE_SET_USED(check_crc_24b);
1021         RTE_SET_USED(crc24_overlap);
1022         RTE_SET_USED(in_length);
1023         RTE_SET_USED(q_stats);
1024 #endif
1025 }
1026
1027 static inline void
1028 enqueue_dec_one_op(struct turbo_sw_queue *q, struct rte_bbdev_dec_op *op,
1029                 struct rte_bbdev_stats *queue_stats)
1030 {
1031         uint8_t c, r = 0;
1032         uint16_t kw, k = 0;
1033         uint16_t crc24_overlap = 0;
1034         struct rte_bbdev_op_turbo_dec *dec = &op->turbo_dec;
1035         struct rte_mbuf *m_in = dec->input.data;
1036         struct rte_mbuf *m_out = dec->hard_output.data;
1037         struct rte_mbuf *m_out_head = dec->hard_output.data;
1038         uint16_t in_offset = dec->input.offset;
1039         uint16_t out_offset = dec->hard_output.offset;
1040         uint32_t mbuf_total_left = dec->input.length;
1041         uint16_t seg_total_left;
1042
1043         /* Clear op status */
1044         op->status = 0;
1045
1046         if (m_in == NULL || m_out == NULL) {
1047                 rte_bbdev_log(ERR, "Invalid mbuf pointer");
1048                 op->status = 1 << RTE_BBDEV_DATA_ERROR;
1049                 return;
1050         }
1051
1052         if (dec->code_block_mode == 0) { /* For Transport Block mode */
1053                 c = dec->tb_params.c;
1054         } else { /* For Code Block mode */
1055                 k = dec->cb_params.k;
1056                 c = 1;
1057         }
1058
1059         if ((c > 1) && !check_bit(dec->op_flags,
1060                 RTE_BBDEV_TURBO_DEC_TB_CRC_24B_KEEP))
1061                 crc24_overlap = 24;
1062
1063         while (mbuf_total_left > 0) {
1064                 if (dec->code_block_mode == 0)
1065                         k = (r < dec->tb_params.c_neg) ?
1066                                 dec->tb_params.k_neg : dec->tb_params.k_pos;
1067
1068                 seg_total_left = rte_pktmbuf_data_len(m_in) - in_offset;
1069
1070                 /* Calculates circular buffer size (Kw).
1071                  * According to 3gpp 36.212 section 5.1.4.2
1072                  *   Kw = 3 * Kpi,
1073                  * where:
1074                  *   Kpi = nCol * nRow
1075                  * where nCol is 32 and nRow can be calculated from:
1076                  *   D =< nCol * nRow
1077                  * where D is the size of each output from turbo encoder block
1078                  * (k + 4).
1079                  */
1080                 kw = RTE_ALIGN_CEIL(k + 4, RTE_BBDEV_TURBO_C_SUBBLOCK) * 3;
1081
1082                 process_dec_cb(q, op, c, k, kw, m_in, m_out_head, m_out,
1083                                 in_offset, out_offset, check_bit(dec->op_flags,
1084                                 RTE_BBDEV_TURBO_CRC_TYPE_24B), crc24_overlap,
1085                                 seg_total_left, queue_stats);
1086                 /* To keep CRC24 attached to end of Code block, use
1087                  * RTE_BBDEV_TURBO_DEC_TB_CRC_24B_KEEP flag as it
1088                  * removed by default once verified.
1089                  */
1090
1091                 mbuf_total_left -= kw;
1092
1093                 /* Update offsets */
1094                 if (seg_total_left == kw) {
1095                         /* Go to the next mbuf */
1096                         m_in = m_in->next;
1097                         m_out = m_out->next;
1098                         in_offset = 0;
1099                         out_offset = 0;
1100                 } else {
1101                         /* Update offsets for next CBs (if exist) */
1102                         in_offset += kw;
1103                         out_offset += ((k - crc24_overlap) >> 3);
1104                 }
1105                 r++;
1106         }
1107         if (mbuf_total_left != 0) {
1108                 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
1109                 rte_bbdev_log(ERR,
1110                                 "Mismatch between mbuf length and included Circular buffer sizes");
1111         }
1112 }
1113
1114 static inline uint16_t
1115 enqueue_dec_all_ops(struct turbo_sw_queue *q, struct rte_bbdev_dec_op **ops,
1116                 uint16_t nb_ops, struct rte_bbdev_stats *queue_stats)
1117 {
1118         uint16_t i;
1119 #ifdef RTE_BBDEV_OFFLOAD_COST
1120         queue_stats->acc_offload_cycles = 0;
1121 #endif
1122
1123         for (i = 0; i < nb_ops; ++i)
1124                 enqueue_dec_one_op(q, ops[i], queue_stats);
1125
1126         return rte_ring_enqueue_burst(q->processed_pkts, (void **)ops, nb_ops,
1127                         NULL);
1128 }
1129
1130 /* Enqueue burst */
1131 static uint16_t
1132 enqueue_enc_ops(struct rte_bbdev_queue_data *q_data,
1133                 struct rte_bbdev_enc_op **ops, uint16_t nb_ops)
1134 {
1135         void *queue = q_data->queue_private;
1136         struct turbo_sw_queue *q = queue;
1137         uint16_t nb_enqueued = 0;
1138
1139         nb_enqueued = enqueue_enc_all_ops(q, ops, nb_ops, &q_data->queue_stats);
1140
1141         q_data->queue_stats.enqueue_err_count += nb_ops - nb_enqueued;
1142         q_data->queue_stats.enqueued_count += nb_enqueued;
1143
1144         return nb_enqueued;
1145 }
1146
1147 /* Enqueue burst */
1148 static uint16_t
1149 enqueue_dec_ops(struct rte_bbdev_queue_data *q_data,
1150                  struct rte_bbdev_dec_op **ops, uint16_t nb_ops)
1151 {
1152         void *queue = q_data->queue_private;
1153         struct turbo_sw_queue *q = queue;
1154         uint16_t nb_enqueued = 0;
1155
1156         nb_enqueued = enqueue_dec_all_ops(q, ops, nb_ops, &q_data->queue_stats);
1157
1158         q_data->queue_stats.enqueue_err_count += nb_ops - nb_enqueued;
1159         q_data->queue_stats.enqueued_count += nb_enqueued;
1160
1161         return nb_enqueued;
1162 }
1163
1164 /* Dequeue decode burst */
1165 static uint16_t
1166 dequeue_dec_ops(struct rte_bbdev_queue_data *q_data,
1167                 struct rte_bbdev_dec_op **ops, uint16_t nb_ops)
1168 {
1169         struct turbo_sw_queue *q = q_data->queue_private;
1170         uint16_t nb_dequeued = rte_ring_dequeue_burst(q->processed_pkts,
1171                         (void **)ops, nb_ops, NULL);
1172         q_data->queue_stats.dequeued_count += nb_dequeued;
1173
1174         return nb_dequeued;
1175 }
1176
1177 /* Dequeue encode burst */
1178 static uint16_t
1179 dequeue_enc_ops(struct rte_bbdev_queue_data *q_data,
1180                 struct rte_bbdev_enc_op **ops, uint16_t nb_ops)
1181 {
1182         struct turbo_sw_queue *q = q_data->queue_private;
1183         uint16_t nb_dequeued = rte_ring_dequeue_burst(q->processed_pkts,
1184                         (void **)ops, nb_ops, NULL);
1185         q_data->queue_stats.dequeued_count += nb_dequeued;
1186
1187         return nb_dequeued;
1188 }
1189
1190 /* Parse 16bit integer from string argument */
1191 static inline int
1192 parse_u16_arg(const char *key, const char *value, void *extra_args)
1193 {
1194         uint16_t *u16 = extra_args;
1195         unsigned int long result;
1196
1197         if ((value == NULL) || (extra_args == NULL))
1198                 return -EINVAL;
1199         errno = 0;
1200         result = strtoul(value, NULL, 0);
1201         if ((result >= (1 << 16)) || (errno != 0)) {
1202                 rte_bbdev_log(ERR, "Invalid value %lu for %s", result, key);
1203                 return -ERANGE;
1204         }
1205         *u16 = (uint16_t)result;
1206         return 0;
1207 }
1208
1209 /* Parse parameters used to create device */
1210 static int
1211 parse_turbo_sw_params(struct turbo_sw_params *params, const char *input_args)
1212 {
1213         struct rte_kvargs *kvlist = NULL;
1214         int ret = 0;
1215
1216         if (params == NULL)
1217                 return -EINVAL;
1218         if (input_args) {
1219                 kvlist = rte_kvargs_parse(input_args, turbo_sw_valid_params);
1220                 if (kvlist == NULL)
1221                         return -EFAULT;
1222
1223                 ret = rte_kvargs_process(kvlist, turbo_sw_valid_params[0],
1224                                         &parse_u16_arg, &params->queues_num);
1225                 if (ret < 0)
1226                         goto exit;
1227
1228                 ret = rte_kvargs_process(kvlist, turbo_sw_valid_params[1],
1229                                         &parse_u16_arg, &params->socket_id);
1230                 if (ret < 0)
1231                         goto exit;
1232
1233                 if (params->socket_id >= RTE_MAX_NUMA_NODES) {
1234                         rte_bbdev_log(ERR, "Invalid socket, must be < %u",
1235                                         RTE_MAX_NUMA_NODES);
1236                         goto exit;
1237                 }
1238         }
1239
1240 exit:
1241         if (kvlist)
1242                 rte_kvargs_free(kvlist);
1243         return ret;
1244 }
1245
1246 /* Create device */
1247 static int
1248 turbo_sw_bbdev_create(struct rte_vdev_device *vdev,
1249                 struct turbo_sw_params *init_params)
1250 {
1251         struct rte_bbdev *bbdev;
1252         const char *name = rte_vdev_device_name(vdev);
1253
1254         bbdev = rte_bbdev_allocate(name);
1255         if (bbdev == NULL)
1256                 return -ENODEV;
1257
1258         bbdev->data->dev_private = rte_zmalloc_socket(name,
1259                         sizeof(struct bbdev_private), RTE_CACHE_LINE_SIZE,
1260                         init_params->socket_id);
1261         if (bbdev->data->dev_private == NULL) {
1262                 rte_bbdev_release(bbdev);
1263                 return -ENOMEM;
1264         }
1265
1266         bbdev->dev_ops = &pmd_ops;
1267         bbdev->device = &vdev->device;
1268         bbdev->data->socket_id = init_params->socket_id;
1269         bbdev->intr_handle = NULL;
1270
1271         /* register rx/tx burst functions for data path */
1272         bbdev->dequeue_enc_ops = dequeue_enc_ops;
1273         bbdev->dequeue_dec_ops = dequeue_dec_ops;
1274         bbdev->enqueue_enc_ops = enqueue_enc_ops;
1275         bbdev->enqueue_dec_ops = enqueue_dec_ops;
1276         ((struct bbdev_private *) bbdev->data->dev_private)->max_nb_queues =
1277                         init_params->queues_num;
1278
1279         return 0;
1280 }
1281
1282 /* Initialise device */
1283 static int
1284 turbo_sw_bbdev_probe(struct rte_vdev_device *vdev)
1285 {
1286         struct turbo_sw_params init_params = {
1287                 rte_socket_id(),
1288                 RTE_BBDEV_DEFAULT_MAX_NB_QUEUES
1289         };
1290         const char *name;
1291         const char *input_args;
1292
1293         if (vdev == NULL)
1294                 return -EINVAL;
1295
1296         name = rte_vdev_device_name(vdev);
1297         if (name == NULL)
1298                 return -EINVAL;
1299         input_args = rte_vdev_device_args(vdev);
1300         parse_turbo_sw_params(&init_params, input_args);
1301
1302         rte_bbdev_log_debug(
1303                         "Initialising %s on NUMA node %d with max queues: %d\n",
1304                         name, init_params.socket_id, init_params.queues_num);
1305
1306         return turbo_sw_bbdev_create(vdev, &init_params);
1307 }
1308
1309 /* Uninitialise device */
1310 static int
1311 turbo_sw_bbdev_remove(struct rte_vdev_device *vdev)
1312 {
1313         struct rte_bbdev *bbdev;
1314         const char *name;
1315
1316         if (vdev == NULL)
1317                 return -EINVAL;
1318
1319         name = rte_vdev_device_name(vdev);
1320         if (name == NULL)
1321                 return -EINVAL;
1322
1323         bbdev = rte_bbdev_get_named_dev(name);
1324         if (bbdev == NULL)
1325                 return -EINVAL;
1326
1327         rte_free(bbdev->data->dev_private);
1328
1329         return rte_bbdev_release(bbdev);
1330 }
1331
1332 static struct rte_vdev_driver bbdev_turbo_sw_pmd_drv = {
1333         .probe = turbo_sw_bbdev_probe,
1334         .remove = turbo_sw_bbdev_remove
1335 };
1336
1337 RTE_PMD_REGISTER_VDEV(DRIVER_NAME, bbdev_turbo_sw_pmd_drv);
1338 RTE_PMD_REGISTER_PARAM_STRING(DRIVER_NAME,
1339         TURBO_SW_MAX_NB_QUEUES_ARG"=<int> "
1340         TURBO_SW_SOCKET_ID_ARG"=<int>");
1341 RTE_PMD_REGISTER_ALIAS(DRIVER_NAME, turbo_sw);
1342
1343 RTE_INIT(turbo_sw_bbdev_init_log)
1344 {
1345         bbdev_turbo_sw_logtype = rte_log_register("pmd.bb.turbo_sw");
1346         if (bbdev_turbo_sw_logtype >= 0)
1347                 rte_log_set_level(bbdev_turbo_sw_logtype, RTE_LOG_NOTICE);
1348 }