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
  13 #include <rte_bbdev.h>
  14 #include <rte_bbdev_pmd.h>
  15
  16 #include <phy_turbo.h>
  17 #include <phy_crc.h>
  18 #include <phy_rate_match.h>
  19 #include <divide.h>
  20
  21 #define DRIVER_NAME turbo_sw
  22
  23 /* Turbo SW PMD logging ID */
  24 static int bbdev_turbo_sw_logtype;
  25
  26 /* Helper macro for logging */
  27 #define rte_bbdev_log(level, fmt, ...) \
  28         rte_log(RTE_LOG_ ## level, bbdev_turbo_sw_logtype, fmt "\n", \
  29                 ##__VA_ARGS__)
  30
  31 #define rte_bbdev_log_debug(fmt, ...) \
  32         rte_bbdev_log(DEBUG, RTE_STR(__LINE__) ":%s() " fmt, __func__, \
  33                 ##__VA_ARGS__)
  34
  35 #define DEINT_INPUT_BUF_SIZE (((RTE_BBDEV_MAX_CB_SIZE >> 3) + 1) * 48)
  36 #define DEINT_OUTPUT_BUF_SIZE (DEINT_INPUT_BUF_SIZE * 6)
  37 #define ADAPTER_OUTPUT_BUF_SIZE ((RTE_BBDEV_MAX_CB_SIZE + 4) * 48)
  38
  39 /* private data structure */
  40 struct bbdev_private {
  41         unsigned int max_nb_queues;  /**< Max number of queues */
  42 };
  43
  44 /*  Initialisation params structure that can be used by Turbo SW driver */
  45 struct turbo_sw_params {
  46         int socket_id;  /*< Turbo SW device socket */
  47         uint16_t queues_num;  /*< Turbo SW device queues number */
  48 };
  49
  50 /* Accecptable params for Turbo SW devices */
  51 #define TURBO_SW_MAX_NB_QUEUES_ARG  "max_nb_queues"
  52 #define TURBO_SW_SOCKET_ID_ARG      "socket_id"
  53
  54 static const char * const turbo_sw_valid_params[] = {
  55         TURBO_SW_MAX_NB_QUEUES_ARG,
  56         TURBO_SW_SOCKET_ID_ARG
  57 };
  58
  59 /* queue */
  60 struct turbo_sw_queue {
  61         /* Ring for processed (encoded/decoded) operations which are ready to
  62          * be dequeued.
  63          */
  64         struct rte_ring *processed_pkts;
  65         /* Stores input for turbo encoder (used when CRC attachment is
  66          * performed
  67          */
  68         uint8_t *enc_in;
  69         /* Stores output from turbo encoder */
  70         uint8_t *enc_out;
  71         /* Alpha gamma buf for bblib_turbo_decoder() function */
  72         int8_t *ag;
  73         /* Temp buf for bblib_turbo_decoder() function */
  74         uint16_t *code_block;
  75         /* Input buf for bblib_rate_dematching_lte() function */
  76         uint8_t *deint_input;
  77         /* Output buf for bblib_rate_dematching_lte() function */
  78         uint8_t *deint_output;
  79         /* Output buf for bblib_turbodec_adapter_lte() function */
  80         uint8_t *adapter_output;
  81         /* Operation type of this queue */
  82         enum rte_bbdev_op_type type;
  83 } __rte_cache_aligned;
  84
  85 /* Calculate index based on Table 5.1.3-3 from TS34.212 */
  86 static inline int32_t
  87 compute_idx(uint16_t k)
  88 {
  89         int32_t result = 0;
  90
  91         if (k < RTE_BBDEV_MIN_CB_SIZE || k > RTE_BBDEV_MAX_CB_SIZE)
  92                 return -1;
  93
  94         if (k > 2048) {
  95                 if ((k - 2048) % 64 != 0)
  96                         result = -1;
  97
  98                 result = 124 + (k - 2048) / 64;
  99         } else if (k <= 512) {
 100                 if ((k - 40) % 8 != 0)
 101                         result = -1;
 102
 103                 result = (k - 40) / 8 + 1;
 104         } else if (k <= 1024) {
 105                 if ((k - 512) % 16 != 0)
 106                         result = -1;
 107
 108                 result = 60 + (k - 512) / 16;
 109         } else { /* 1024 < k <= 2048 */
 110                 if ((k - 1024) % 32 != 0)
 111                         result = -1;
 112
 113                 result = 92 + (k - 1024) / 32;
 114         }
 115
 116         return result;
 117 }
 118
 119 /* Read flag value 0/1 from bitmap */
 120 static inline bool
 121 check_bit(uint32_t bitmap, uint32_t bitmask)
 122 {
 123         return bitmap & bitmask;
 124 }
 125
 126 /* Get device info */
 127 static void
 128 info_get(struct rte_bbdev *dev, struct rte_bbdev_driver_info *dev_info)
 129 {
 130         struct bbdev_private *internals = dev->data->dev_private;
 131
 132         static const struct rte_bbdev_op_cap bbdev_capabilities[] = {
 133                 {
 134                         .type = RTE_BBDEV_OP_TURBO_DEC,
 135                         .cap.turbo_dec = {
 136                                 .capability_flags =
 137                                         RTE_BBDEV_TURBO_SUBBLOCK_DEINTERLEAVE |
 138                                         RTE_BBDEV_TURBO_POS_LLR_1_BIT_IN |
 139                                         RTE_BBDEV_TURBO_NEG_LLR_1_BIT_IN |
 140                                         RTE_BBDEV_TURBO_CRC_TYPE_24B |
 141                                         RTE_BBDEV_TURBO_EARLY_TERMINATION,
 142                                 .max_llr_modulus = 16,
 143                                 .num_buffers_src = RTE_BBDEV_MAX_CODE_BLOCKS,
 144                                 .num_buffers_hard_out =
 145                                                 RTE_BBDEV_MAX_CODE_BLOCKS,
 146                                 .num_buffers_soft_out = 0,
 147                         }
 148                 },
 149                 {
 150                         .type   = RTE_BBDEV_OP_TURBO_ENC,
 151                         .cap.turbo_enc = {
 152                                 .capability_flags =
 153                                                 RTE_BBDEV_TURBO_CRC_24B_ATTACH |
 154                                                 RTE_BBDEV_TURBO_CRC_24A_ATTACH |
 155                                                 RTE_BBDEV_TURBO_RATE_MATCH |
 156                                                 RTE_BBDEV_TURBO_RV_INDEX_BYPASS,
 157                                 .num_buffers_src = RTE_BBDEV_MAX_CODE_BLOCKS,
 158                                 .num_buffers_dst = RTE_BBDEV_MAX_CODE_BLOCKS,
 159                         }
 160                 },
 161                 RTE_BBDEV_END_OF_CAPABILITIES_LIST()
 162         };
 163
 164         static struct rte_bbdev_queue_conf default_queue_conf = {
 165                 .queue_size = RTE_BBDEV_QUEUE_SIZE_LIMIT,
 166         };
 167
 168         static const enum rte_cpu_flag_t cpu_flag = RTE_CPUFLAG_SSE4_2;
 169
 170         default_queue_conf.socket = dev->data->socket_id;
 171
 172         dev_info->driver_name = RTE_STR(DRIVER_NAME);
 173         dev_info->max_num_queues = internals->max_nb_queues;
 174         dev_info->queue_size_lim = RTE_BBDEV_QUEUE_SIZE_LIMIT;
 175         dev_info->hardware_accelerated = false;
 176         dev_info->max_queue_priority = 0;
 177         dev_info->default_queue_conf = default_queue_conf;
 178         dev_info->capabilities = bbdev_capabilities;
 179         dev_info->cpu_flag_reqs = &cpu_flag;
 180         dev_info->min_alignment = 64;
 181
 182         rte_bbdev_log_debug("got device info from %u\n", dev->data->dev_id);
 183 }
 184
 185 /* Release queue */
 186 static int
 187 q_release(struct rte_bbdev *dev, uint16_t q_id)
 188 {
 189         struct turbo_sw_queue *q = dev->data->queues[q_id].queue_private;
 190
 191         if (q != NULL) {
 192                 rte_ring_free(q->processed_pkts);
 193                 rte_free(q->enc_out);
 194                 rte_free(q->enc_in);
 195                 rte_free(q->ag);
 196                 rte_free(q->code_block);
 197                 rte_free(q->deint_input);
 198                 rte_free(q->deint_output);
 199                 rte_free(q->adapter_output);
 200                 rte_free(q);
 201                 dev->data->queues[q_id].queue_private = NULL;
 202         }
 203
 204         rte_bbdev_log_debug("released device queue %u:%u",
 205                         dev->data->dev_id, q_id);
 206         return 0;
 207 }
 208
 209 /* Setup a queue */
 210 static int
 211 q_setup(struct rte_bbdev *dev, uint16_t q_id,
 212                 const struct rte_bbdev_queue_conf *queue_conf)
 213 {
 214         int ret;
 215         struct turbo_sw_queue *q;
 216         char name[RTE_RING_NAMESIZE];
 217
 218         /* Allocate the queue data structure. */
 219         q = rte_zmalloc_socket(RTE_STR(DRIVER_NAME), sizeof(*q),
 220                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 221         if (q == NULL) {
 222                 rte_bbdev_log(ERR, "Failed to allocate queue memory");
 223                 return -ENOMEM;
 224         }
 225
 226         /* Allocate memory for encoder output. */
 227         ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"_enc_out%u:%u",
 228                         dev->data->dev_id, q_id);
 229         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 230                 rte_bbdev_log(ERR,
 231                                 "Creating queue name for device %u queue %u failed",
 232                                 dev->data->dev_id, q_id);
 233                 return -ENAMETOOLONG;
 234         }
 235         q->enc_out = rte_zmalloc_socket(name,
 236                         ((RTE_BBDEV_MAX_TB_SIZE >> 3) + 3) *
 237                         sizeof(*q->enc_out) * 3,
 238                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 239         if (q->enc_out == NULL) {
 240                 rte_bbdev_log(ERR,
 241                         "Failed to allocate queue memory for %s", name);
 242                 goto free_q;
 243         }
 244
 245         /* Allocate memory for rate matching output. */
 246         ret = snprintf(name, RTE_RING_NAMESIZE,
 247                         RTE_STR(DRIVER_NAME)"_enc_in%u:%u", dev->data->dev_id,
 248                         q_id);
 249         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 250                 rte_bbdev_log(ERR,
 251                                 "Creating queue name for device %u queue %u failed",
 252                                 dev->data->dev_id, q_id);
 253                 return -ENAMETOOLONG;
 254         }
 255         q->enc_in = rte_zmalloc_socket(name,
 256                         (RTE_BBDEV_MAX_CB_SIZE >> 3) * sizeof(*q->enc_in),
 257                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 258         if (q->enc_in == NULL) {
 259                 rte_bbdev_log(ERR,
 260                         "Failed to allocate queue memory for %s", name);
 261                 goto free_q;
 262         }
 263
 264         /* Allocate memory for Aplha Gamma temp buffer. */
 265         ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"_ag%u:%u",
 266                         dev->data->dev_id, q_id);
 267         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 268                 rte_bbdev_log(ERR,
 269                                 "Creating queue name for device %u queue %u failed",
 270                                 dev->data->dev_id, q_id);
 271                 return -ENAMETOOLONG;
 272         }
 273         q->ag = rte_zmalloc_socket(name,
 274                         RTE_BBDEV_MAX_CB_SIZE * 10 * sizeof(*q->ag),
 275                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 276         if (q->ag == NULL) {
 277                 rte_bbdev_log(ERR,
 278                         "Failed to allocate queue memory for %s", name);
 279                 goto free_q;
 280         }
 281
 282         /* Allocate memory for code block temp buffer. */
 283         ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"_cb%u:%u",
 284                         dev->data->dev_id, q_id);
 285         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 286                 rte_bbdev_log(ERR,
 287                                 "Creating queue name for device %u queue %u failed",
 288                                 dev->data->dev_id, q_id);
 289                 return -ENAMETOOLONG;
 290         }
 291         q->code_block = rte_zmalloc_socket(name,
 292                         RTE_BBDEV_MAX_CB_SIZE * sizeof(*q->code_block),
 293                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 294         if (q->code_block == NULL) {
 295                 rte_bbdev_log(ERR,
 296                         "Failed to allocate queue memory for %s", name);
 297                 goto free_q;
 298         }
 299
 300         /* Allocate memory for Deinterleaver input. */
 301         ret = snprintf(name, RTE_RING_NAMESIZE,
 302                         RTE_STR(DRIVER_NAME)"_deint_input%u:%u",
 303                         dev->data->dev_id, q_id);
 304         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 305                 rte_bbdev_log(ERR,
 306                                 "Creating queue name for device %u queue %u failed",
 307                                 dev->data->dev_id, q_id);
 308                 return -ENAMETOOLONG;
 309         }
 310         q->deint_input = rte_zmalloc_socket(name,
 311                         DEINT_INPUT_BUF_SIZE * sizeof(*q->deint_input),
 312                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 313         if (q->deint_input == NULL) {
 314                 rte_bbdev_log(ERR,
 315                         "Failed to allocate queue memory for %s", name);
 316                 goto free_q;
 317         }
 318
 319         /* Allocate memory for Deinterleaver output. */
 320         ret = snprintf(name, RTE_RING_NAMESIZE,
 321                         RTE_STR(DRIVER_NAME)"_deint_output%u:%u",
 322                         dev->data->dev_id, q_id);
 323         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 324                 rte_bbdev_log(ERR,
 325                                 "Creating queue name for device %u queue %u failed",
 326                                 dev->data->dev_id, q_id);
 327                 return -ENAMETOOLONG;
 328         }
 329         q->deint_output = rte_zmalloc_socket(NULL,
 330                         DEINT_OUTPUT_BUF_SIZE * sizeof(*q->deint_output),
 331                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 332         if (q->deint_output == NULL) {
 333                 rte_bbdev_log(ERR,
 334                         "Failed to allocate queue memory for %s", name);
 335                 goto free_q;
 336         }
 337
 338         /* Allocate memory for Adapter output. */
 339         ret = snprintf(name, RTE_RING_NAMESIZE,
 340                         RTE_STR(DRIVER_NAME)"_adapter_output%u:%u",
 341                         dev->data->dev_id, q_id);
 342         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 343                 rte_bbdev_log(ERR,
 344                                 "Creating queue name for device %u queue %u failed",
 345                                 dev->data->dev_id, q_id);
 346                 return -ENAMETOOLONG;
 347         }
 348         q->adapter_output = rte_zmalloc_socket(NULL,
 349                         ADAPTER_OUTPUT_BUF_SIZE * sizeof(*q->adapter_output),
 350                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 351         if (q->adapter_output == NULL) {
 352                 rte_bbdev_log(ERR,
 353                         "Failed to allocate queue memory for %s", name);
 354                 goto free_q;
 355         }
 356
 357         /* Create ring for packets awaiting to be dequeued. */
 358         ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"%u:%u",
 359                         dev->data->dev_id, q_id);
 360         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 361                 rte_bbdev_log(ERR,
 362                                 "Creating queue name for device %u queue %u failed",
 363                                 dev->data->dev_id, q_id);
 364                 return -ENAMETOOLONG;
 365         }
 366         q->processed_pkts = rte_ring_create(name, queue_conf->queue_size,
 367                         queue_conf->socket, RING_F_SP_ENQ | RING_F_SC_DEQ);
 368         if (q->processed_pkts == NULL) {
 369                 rte_bbdev_log(ERR, "Failed to create ring for %s", name);
 370                 goto free_q;
 371         }
 372
 373         q->type = queue_conf->op_type;
 374
 375         dev->data->queues[q_id].queue_private = q;
 376         rte_bbdev_log_debug("setup device queue %s", name);
 377         return 0;
 378
 379 free_q:
 380         rte_ring_free(q->processed_pkts);
 381         rte_free(q->enc_out);
 382         rte_free(q->enc_in);
 383         rte_free(q->ag);
 384         rte_free(q->code_block);
 385         rte_free(q->deint_input);
 386         rte_free(q->deint_output);
 387         rte_free(q->adapter_output);
 388         rte_free(q);
 389         return -EFAULT;
 390 }
 391
 392 static const struct rte_bbdev_ops pmd_ops = {
 393         .info_get = info_get,
 394         .queue_setup = q_setup,
 395         .queue_release = q_release
 396 };
 397
 398 /* Checks if the encoder input buffer is correct.
 399  * Returns 0 if it's valid, -1 otherwise.
 400  */
 401 static inline int
 402 is_enc_input_valid(const uint16_t k, const int32_t k_idx,
 403                 const uint16_t in_length)
 404 {
 405         if (k_idx < 0) {
 406                 rte_bbdev_log(ERR, "K Index is invalid");
 407                 return -1;
 408         }
 409
 410         if (in_length - (k >> 3) < 0) {
 411                 rte_bbdev_log(ERR,
 412                                 "Mismatch between input length (%u bytes) and K (%u bits)",
 413                                 in_length, k);
 414                 return -1;
 415         }
 416
 417         if (k > RTE_BBDEV_MAX_CB_SIZE) {
 418                 rte_bbdev_log(ERR, "CB size (%u) is too big, max: %d",
 419                                 k, RTE_BBDEV_MAX_CB_SIZE);
 420                 return -1;
 421         }
 422
 423         return 0;
 424 }
 425
 426 /* Checks if the decoder input buffer is correct.
 427  * Returns 0 if it's valid, -1 otherwise.
 428  */
 429 static inline int
 430 is_dec_input_valid(int32_t k_idx, int16_t kw, int16_t in_length)
 431 {
 432         if (k_idx < 0) {
 433                 rte_bbdev_log(ERR, "K index is invalid");
 434                 return -1;
 435         }
 436
 437         if (in_length - kw < 0) {
 438                 rte_bbdev_log(ERR,
 439                                 "Mismatch between input length (%u) and kw (%u)",
 440                                 in_length, kw);
 441                 return -1;
 442         }
 443
 444         if (kw > RTE_BBDEV_MAX_KW) {
 445                 rte_bbdev_log(ERR, "Input length (%u) is too big, max: %d",
 446                                 kw, RTE_BBDEV_MAX_KW);
 447                 return -1;
 448         }
 449
 450         return 0;
 451 }
 452
 453 static inline void
 454 process_enc_cb(struct turbo_sw_queue *q, struct rte_bbdev_enc_op *op,
 455                 uint8_t r, uint8_t c, uint16_t k, uint16_t ncb,
 456                 uint32_t e, struct rte_mbuf *m_in, struct rte_mbuf *m_out,
 457                 uint16_t in_offset, uint16_t out_offset, uint16_t total_left)
 458 {
 459         int ret;
 460         int16_t k_idx;
 461         uint16_t m;
 462         uint8_t *in, *out0, *out1, *out2, *tmp_out, *rm_out;
 463         uint64_t first_3_bytes = 0;
 464         struct rte_bbdev_op_turbo_enc *enc = &op->turbo_enc;
 465         struct bblib_crc_request crc_req;
 466         struct bblib_crc_response crc_resp;
 467         struct bblib_turbo_encoder_request turbo_req;
 468         struct bblib_turbo_encoder_response turbo_resp;
 469         struct bblib_rate_match_dl_request rm_req;
 470         struct bblib_rate_match_dl_response rm_resp;
 471
 472         k_idx = compute_idx(k);
 473         in = rte_pktmbuf_mtod_offset(m_in, uint8_t *, in_offset);
 474
 475         /* CRC24A (for TB) */
 476         if ((enc->op_flags & RTE_BBDEV_TURBO_CRC_24A_ATTACH) &&
 477                 (enc->code_block_mode == 1)) {
 478                 ret = is_enc_input_valid(k - 24, k_idx, total_left);
 479                 if (ret != 0) {
 480                         op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 481                         return;
 482                 }
 483                 crc_req.data = in;
 484                 crc_req.len = (k - 24) >> 3;
 485                 /* Check if there is a room for CRC bits. If not use
 486                  * the temporary buffer.
 487                  */
 488                 if (rte_pktmbuf_append(m_in, 3) == NULL) {
 489                         rte_memcpy(q->enc_in, in, (k - 24) >> 3);
 490                         in = q->enc_in;
 491                 } else {
 492                         /* Store 3 first bytes of next CB as they will be
 493                          * overwritten by CRC bytes. If it is the last CB then
 494                          * there is no point to store 3 next bytes and this
 495                          * if..else branch will be omitted.
 496                          */
 497                         first_3_bytes = *((uint64_t *)&in[(k - 32) >> 3]);
 498                 }
 499
 500                 crc_resp.data = in;
 501                 bblib_lte_crc24a_gen(&crc_req, &crc_resp);
 502         } else if (enc->op_flags & RTE_BBDEV_TURBO_CRC_24B_ATTACH) {
 503                 /* CRC24B */
 504                 ret = is_enc_input_valid(k - 24, k_idx, total_left);
 505                 if (ret != 0) {
 506                         op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 507                         return;
 508                 }
 509                 crc_req.data = in;
 510                 crc_req.len = (k - 24) >> 3;
 511                 /* Check if there is a room for CRC bits. If this is the last
 512                  * CB in TB. If not use temporary buffer.
 513                  */
 514                 if ((c - r == 1) && (rte_pktmbuf_append(m_in, 3) == NULL)) {
 515                         rte_memcpy(q->enc_in, in, (k - 24) >> 3);
 516                         in = q->enc_in;
 517                 } else if (c - r > 1) {
 518                         /* Store 3 first bytes of next CB as they will be
 519                          * overwritten by CRC bytes. If it is the last CB then
 520                          * there is no point to store 3 next bytes and this
 521                          * if..else branch will be omitted.
 522                          */
 523                         first_3_bytes = *((uint64_t *)&in[(k - 32) >> 3]);
 524                 }
 525
 526                 crc_resp.data = in;
 527                 bblib_lte_crc24b_gen(&crc_req, &crc_resp);
 528         } else {
 529                 ret = is_enc_input_valid(k, k_idx, total_left);
 530                 if (ret != 0) {
 531                         op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 532                         return;
 533                 }
 534         }
 535
 536         /* Turbo encoder */
 537
 538         /* Each bit layer output from turbo encoder is (k+4) bits long, i.e.
 539          * input length + 4 tail bits. That's (k/8) + 1 bytes after rounding up.
 540          * So dst_data's length should be 3*(k/8) + 3 bytes.
 541          * In Rate-matching bypass case outputs pointers passed to encoder
 542          * (out0, out1 and out2) can directly point to addresses of output from
 543          * turbo_enc entity.
 544          */
 545         if (enc->op_flags & RTE_BBDEV_TURBO_RATE_MATCH) {
 546                 out0 = q->enc_out;
 547                 out1 = RTE_PTR_ADD(out0, (k >> 3) + 1);
 548                 out2 = RTE_PTR_ADD(out1, (k >> 3) + 1);
 549         } else {
 550                 out0 = (uint8_t *)rte_pktmbuf_append(m_out, (k >> 3) * 3 + 2);
 551                 if (out0 == NULL) {
 552                         op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 553                         rte_bbdev_log(ERR,
 554                                         "Too little space in output mbuf");
 555                         return;
 556                 }
 557                 enc->output.length += (k >> 3) * 3 + 2;
 558                 /* rte_bbdev_op_data.offset can be different than the
 559                  * offset of the appended bytes
 560                  */
 561                 out0 = rte_pktmbuf_mtod_offset(m_out, uint8_t *, out_offset);
 562                 out1 = rte_pktmbuf_mtod_offset(m_out, uint8_t *,
 563                                 out_offset + (k >> 3) + 1);
 564                 out2 = rte_pktmbuf_mtod_offset(m_out, uint8_t *,
 565                                 out_offset + 2 * ((k >> 3) + 1));
 566         }
 567
 568         turbo_req.case_id = k_idx;
 569         turbo_req.input_win = in;
 570         turbo_req.length = k >> 3;
 571         turbo_resp.output_win_0 = out0;
 572         turbo_resp.output_win_1 = out1;
 573         turbo_resp.output_win_2 = out2;
 574         if (bblib_turbo_encoder(&turbo_req, &turbo_resp) != 0) {
 575                 op->status |= 1 << RTE_BBDEV_DRV_ERROR;
 576                 rte_bbdev_log(ERR, "Turbo Encoder failed");
 577                 return;
 578         }
 579
 580         /* Restore 3 first bytes of next CB if they were overwritten by CRC*/
 581         if (first_3_bytes != 0)
 582                 *((uint64_t *)&in[(k - 32) >> 3]) = first_3_bytes;
 583
 584         /* Rate-matching */
 585         if (enc->op_flags & RTE_BBDEV_TURBO_RATE_MATCH) {
 586                 uint8_t mask_id;
 587                 /* Integer round up division by 8 */
 588                 uint16_t out_len = (e + 7) >> 3;
 589                 /* The mask array is indexed using E%8. E is an even number so
 590                  * there are only 4 possible values.
 591                  */
 592                 const uint8_t mask_out[] = {0xFF, 0xC0, 0xF0, 0xFC};
 593
 594                 /* get output data starting address */
 595                 rm_out = (uint8_t *)rte_pktmbuf_append(m_out, out_len);
 596                 if (rm_out == NULL) {
 597                         op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 598                         rte_bbdev_log(ERR,
 599                                         "Too little space in output mbuf");
 600                         return;
 601                 }
 602                 /* rte_bbdev_op_data.offset can be different than the offset
 603                  * of the appended bytes
 604                  */
 605                 rm_out = rte_pktmbuf_mtod_offset(m_out, uint8_t *, out_offset);
 606
 607                 /* index of current code block */
 608                 rm_req.r = r;
 609                 /* total number of code block */
 610                 rm_req.C = c;
 611                 /* For DL - 1, UL - 0 */
 612                 rm_req.direction = 1;
 613                 /* According to 3ggp 36.212 Spec 5.1.4.1.2 section Nsoft, KMIMO
 614                  * and MDL_HARQ are used for Ncb calculation. As Ncb is already
 615                  * known we can adjust those parameters
 616                  */
 617                 rm_req.Nsoft = ncb * rm_req.C;
 618                 rm_req.KMIMO = 1;
 619                 rm_req.MDL_HARQ = 1;
 620                 /* According to 3ggp 36.212 Spec 5.1.4.1.2 section Nl, Qm and G
 621                  * are used for E calculation. As E is already known we can
 622                  * adjust those parameters
 623                  */
 624                 rm_req.NL = e;
 625                 rm_req.Qm = 1;
 626                 rm_req.G = rm_req.NL * rm_req.Qm * rm_req.C;
 627
 628                 rm_req.rvidx = enc->rv_index;
 629                 rm_req.Kidx = k_idx - 1;
 630                 rm_req.nLen = k + 4;
 631                 rm_req.tin0 = out0;
 632                 rm_req.tin1 = out1;
 633                 rm_req.tin2 = out2;
 634                 rm_resp.output = rm_out;
 635                 rm_resp.OutputLen = out_len;
 636                 if (enc->op_flags & RTE_BBDEV_TURBO_RV_INDEX_BYPASS)
 637                         rm_req.bypass_rvidx = 1;
 638                 else
 639                         rm_req.bypass_rvidx = 0;
 640
 641                 if (bblib_rate_match_dl(&rm_req, &rm_resp) != 0) {
 642                         op->status |= 1 << RTE_BBDEV_DRV_ERROR;
 643                         rte_bbdev_log(ERR, "Rate matching failed");
 644                         return;
 645                 }
 646
 647                 /* SW fills an entire last byte even if E%8 != 0. Clear the
 648                  * superfluous data bits for consistency with HW device.
 649                  */
 650                 mask_id = (e & 7) >> 1;
 651                 rm_out[out_len - 1] &= mask_out[mask_id];
 652
 653                 enc->output.length += rm_resp.OutputLen;
 654         } else {
 655                 /* Rate matching is bypassed */
 656
 657                 /* Completing last byte of out0 (where 4 tail bits are stored)
 658                  * by moving first 4 bits from out1
 659                  */
 660                 tmp_out = (uint8_t *) --out1;
 661                 *tmp_out = *tmp_out | ((*(tmp_out + 1) & 0xF0) >> 4);
 662                 tmp_out++;
 663                 /* Shifting out1 data by 4 bits to the left */
 664                 for (m = 0; m < k >> 3; ++m) {
 665                         uint8_t *first = tmp_out;
 666                         uint8_t second = *(tmp_out + 1);
 667                         *first = (*first << 4) | ((second & 0xF0) >> 4);
 668                         tmp_out++;
 669                 }
 670                 /* Shifting out2 data by 8 bits to the left */
 671                 for (m = 0; m < (k >> 3) + 1; ++m) {
 672                         *tmp_out = *(tmp_out + 1);
 673                         tmp_out++;
 674                 }
 675                 *tmp_out = 0;
 676         }
 677 }
 678
 679 static inline void
 680 enqueue_enc_one_op(struct turbo_sw_queue *q, struct rte_bbdev_enc_op *op)
 681 {
 682         uint8_t c, r, crc24_bits = 0;
 683         uint16_t k, ncb;
 684         uint32_t e;
 685         struct rte_bbdev_op_turbo_enc *enc = &op->turbo_enc;
 686         uint16_t in_offset = enc->input.offset;
 687         uint16_t out_offset = enc->output.offset;
 688         struct rte_mbuf *m_in = enc->input.data;
 689         struct rte_mbuf *m_out = enc->output.data;
 690         uint16_t total_left = enc->input.length;
 691
 692         /* Clear op status */
 693         op->status = 0;
 694
 695         if (total_left > RTE_BBDEV_MAX_TB_SIZE >> 3) {
 696                 rte_bbdev_log(ERR, "TB size (%u) is too big, max: %d",
 697                                 total_left, RTE_BBDEV_MAX_TB_SIZE);
 698                 op->status = 1 << RTE_BBDEV_DATA_ERROR;
 699                 return;
 700         }
 701
 702         if (m_in == NULL || m_out == NULL) {
 703                 rte_bbdev_log(ERR, "Invalid mbuf pointer");
 704                 op->status = 1 << RTE_BBDEV_DATA_ERROR;
 705                 return;
 706         }
 707
 708         if ((enc->op_flags & RTE_BBDEV_TURBO_CRC_24B_ATTACH) ||
 709                 (enc->op_flags & RTE_BBDEV_TURBO_CRC_24A_ATTACH))
 710                 crc24_bits = 24;
 711
 712         if (enc->code_block_mode == 0) { /* For Transport Block mode */
 713                 c = enc->tb_params.c;
 714                 r = enc->tb_params.r;
 715         } else {/* For Code Block mode */
 716                 c = 1;
 717                 r = 0;
 718         }
 719
 720         while (total_left > 0 && r < c) {
 721                 if (enc->code_block_mode == 0) {
 722                         k = (r < enc->tb_params.c_neg) ?
 723                                 enc->tb_params.k_neg : enc->tb_params.k_pos;
 724                         ncb = (r < enc->tb_params.c_neg) ?
 725                                 enc->tb_params.ncb_neg : enc->tb_params.ncb_pos;
 726                         e = (r < enc->tb_params.cab) ?
 727                                 enc->tb_params.ea : enc->tb_params.eb;
 728                 } else {
 729                         k = enc->cb_params.k;
 730                         ncb = enc->cb_params.ncb;
 731                         e = enc->cb_params.e;
 732                 }
 733
 734                 process_enc_cb(q, op, r, c, k, ncb, e, m_in,
 735                                 m_out, in_offset, out_offset, total_left);
 736                 /* Update total_left */
 737                 total_left -= (k - crc24_bits) >> 3;
 738                 /* Update offsets for next CBs (if exist) */
 739                 in_offset += (k - crc24_bits) >> 3;
 740                 if (enc->op_flags & RTE_BBDEV_TURBO_RATE_MATCH)
 741                         out_offset += e >> 3;
 742                 else
 743                         out_offset += (k >> 3) * 3 + 2;
 744                 r++;
 745         }
 746
 747         /* check if all input data was processed */
 748         if (total_left != 0) {
 749                 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 750                 rte_bbdev_log(ERR,
 751                                 "Mismatch between mbuf length and included CBs sizes");
 752         }
 753 }
 754
 755 static inline uint16_t
 756 enqueue_enc_all_ops(struct turbo_sw_queue *q, struct rte_bbdev_enc_op **ops,
 757                 uint16_t nb_ops)
 758 {
 759         uint16_t i;
 760
 761         for (i = 0; i < nb_ops; ++i)
 762                 enqueue_enc_one_op(q, ops[i]);
 763
 764         return rte_ring_enqueue_burst(q->processed_pkts, (void **)ops, nb_ops,
 765                         NULL);
 766 }
 767
 768 /* Remove the padding bytes from a cyclic buffer.
 769  * The input buffer is a data stream wk as described in 3GPP TS 36.212 section
 770  * 5.1.4.1.2 starting from w0 and with length Ncb bytes.
 771  * The output buffer is a data stream wk with pruned padding bytes. It's length
 772  * is 3*D bytes and the order of non-padding bytes is preserved.
 773  */
 774 static inline void
 775 remove_nulls_from_circular_buf(const uint8_t *in, uint8_t *out, uint16_t k,
 776                 uint16_t ncb)
 777 {
 778         uint32_t in_idx, out_idx, c_idx;
 779         const uint32_t d = k + 4;
 780         const uint32_t kw = (ncb / 3);
 781         const uint32_t nd = kw - d;
 782         const uint32_t r_subblock = kw / RTE_BBDEV_C_SUBBLOCK;
 783         /* Inter-column permutation pattern */
 784         const uint32_t P[RTE_BBDEV_C_SUBBLOCK] = {0, 16, 8, 24, 4, 20, 12, 28,
 785                         2, 18, 10, 26, 6, 22, 14, 30, 1, 17, 9, 25, 5, 21, 13,
 786                         29, 3, 19, 11, 27, 7, 23, 15, 31};
 787         in_idx = 0;
 788         out_idx = 0;
 789
 790         /* The padding bytes are at the first Nd positions in the first row. */
 791         for (c_idx = 0; in_idx < kw; in_idx += r_subblock, ++c_idx) {
 792                 if (P[c_idx] < nd) {
 793                         rte_memcpy(&out[out_idx], &in[in_idx + 1],
 794                                         r_subblock - 1);
 795                         out_idx += r_subblock - 1;
 796                 } else {
 797                         rte_memcpy(&out[out_idx], &in[in_idx], r_subblock);
 798                         out_idx += r_subblock;
 799                 }
 800         }
 801
 802         /* First and second parity bits sub-blocks are interlaced. */
 803         for (c_idx = 0; in_idx < ncb - 2 * r_subblock;
 804                         in_idx += 2 * r_subblock, ++c_idx) {
 805                 uint32_t second_block_c_idx = P[c_idx];
 806                 uint32_t third_block_c_idx = P[c_idx] + 1;
 807
 808                 if (second_block_c_idx < nd && third_block_c_idx < nd) {
 809                         rte_memcpy(&out[out_idx], &in[in_idx + 2],
 810                                         2 * r_subblock - 2);
 811                         out_idx += 2 * r_subblock - 2;
 812                 } else if (second_block_c_idx >= nd &&
 813                                 third_block_c_idx >= nd) {
 814                         rte_memcpy(&out[out_idx], &in[in_idx], 2 * r_subblock);
 815                         out_idx += 2 * r_subblock;
 816                 } else if (second_block_c_idx < nd) {
 817                         out[out_idx++] = in[in_idx];
 818                         rte_memcpy(&out[out_idx], &in[in_idx + 2],
 819                                         2 * r_subblock - 2);
 820                         out_idx += 2 * r_subblock - 2;
 821                 } else {
 822                         rte_memcpy(&out[out_idx], &in[in_idx + 1],
 823                                         2 * r_subblock - 1);
 824                         out_idx += 2 * r_subblock - 1;
 825                 }
 826         }
 827
 828         /* Last interlaced row is different - its last byte is the only padding
 829          * byte. We can have from 4 up to 28 padding bytes (Nd) per sub-block.
 830          * After interlacing the 1st and 2nd parity sub-blocks we can have 0, 1
 831          * or 2 padding bytes each time we make a step of 2 * R_SUBBLOCK bytes
 832          * (moving to another column). 2nd parity sub-block uses the same
 833          * inter-column permutation pattern as the systematic and 1st parity
 834          * sub-blocks but it adds '1' to the resulting index and calculates the
 835          * modulus of the result and Kw. Last column is mapped to itself (id 31)
 836          * so the first byte taken from the 2nd parity sub-block will be the
 837          * 32nd (31+1) byte, then 64th etc. (step is C_SUBBLOCK == 32) and the
 838          * last byte will be the first byte from the sub-block:
 839          * (32 + 32 * (R_SUBBLOCK-1)) % Kw == Kw % Kw == 0. Nd can't  be smaller
 840          * than 4 so we know that bytes with ids 0, 1, 2 and 3 must be the
 841          * padding bytes. The bytes from the 1st parity sub-block are the bytes
 842          * from the 31st column - Nd can't be greater than 28 so we are sure
 843          * that there are no padding bytes in 31st column.
 844          */
 845         rte_memcpy(&out[out_idx], &in[in_idx], 2 * r_subblock - 1);
 846 }
 847
 848 static inline void
 849 move_padding_bytes(const uint8_t *in, uint8_t *out, uint16_t k,
 850                 uint16_t ncb)
 851 {
 852         uint16_t d = k + 4;
 853         uint16_t kpi = ncb / 3;
 854         uint16_t nd = kpi - d;
 855
 856         rte_memcpy(&out[nd], in, d);
 857         rte_memcpy(&out[nd + kpi + 64], &in[kpi], d);
 858         rte_memcpy(&out[(nd - 1) + 2 * (kpi + 64)], &in[2 * kpi], d);
 859 }
 860
 861 static inline void
 862 process_dec_cb(struct turbo_sw_queue *q, struct rte_bbdev_dec_op *op,
 863                 uint8_t c, uint16_t k, uint16_t kw, struct rte_mbuf *m_in,
 864                 struct rte_mbuf *m_out, uint16_t in_offset, uint16_t out_offset,
 865                 bool check_crc_24b, uint16_t total_left)
 866 {
 867         int ret;
 868         int32_t k_idx;
 869         int32_t iter_cnt;
 870         uint8_t *in, *out, *adapter_input;
 871         int32_t ncb, ncb_without_null;
 872         struct bblib_turbo_adapter_ul_response adapter_resp;
 873         struct bblib_turbo_adapter_ul_request adapter_req;
 874         struct bblib_turbo_decoder_request turbo_req;
 875         struct bblib_turbo_decoder_response turbo_resp;
 876         struct rte_bbdev_op_turbo_dec *dec = &op->turbo_dec;
 877
 878         k_idx = compute_idx(k);
 879
 880         ret = is_dec_input_valid(k_idx, kw, total_left);
 881         if (ret != 0) {
 882                 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 883                 return;
 884         }
 885
 886         in = rte_pktmbuf_mtod_offset(m_in, uint8_t *, in_offset);
 887         ncb = kw;
 888         ncb_without_null = (k + 4) * 3;
 889
 890         if (check_bit(dec->op_flags, RTE_BBDEV_TURBO_SUBBLOCK_DEINTERLEAVE)) {
 891                 struct bblib_deinterleave_ul_request deint_req;
 892                 struct bblib_deinterleave_ul_response deint_resp;
 893
 894                 /* SW decoder accepts only a circular buffer without NULL bytes
 895                  * so the input needs to be converted.
 896                  */
 897                 remove_nulls_from_circular_buf(in, q->deint_input, k, ncb);
 898
 899                 deint_req.pharqbuffer = q->deint_input;
 900                 deint_req.ncb = ncb_without_null;
 901                 deint_resp.pinteleavebuffer = q->deint_output;
 902                 bblib_deinterleave_ul(&deint_req, &deint_resp);
 903         } else
 904                 move_padding_bytes(in, q->deint_output, k, ncb);
 905
 906         adapter_input = q->deint_output;
 907
 908         if (dec->op_flags & RTE_BBDEV_TURBO_POS_LLR_1_BIT_IN)
 909                 adapter_req.isinverted = 1;
 910         else if (dec->op_flags & RTE_BBDEV_TURBO_NEG_LLR_1_BIT_IN)
 911                 adapter_req.isinverted = 0;
 912         else {
 913                 op->status |= 1 << RTE_BBDEV_DRV_ERROR;
 914                 rte_bbdev_log(ERR, "LLR format wasn't specified");
 915                 return;
 916         }
 917
 918         adapter_req.ncb = ncb_without_null;
 919         adapter_req.pinteleavebuffer = adapter_input;
 920         adapter_resp.pharqout = q->adapter_output;
 921         bblib_turbo_adapter_ul(&adapter_req, &adapter_resp);
 922
 923         out = (uint8_t *)rte_pktmbuf_append(m_out, (k >> 3));
 924         if (out == NULL) {
 925                 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 926                 rte_bbdev_log(ERR, "Too little space in output mbuf");
 927                 return;
 928         }
 929         /* rte_bbdev_op_data.offset can be different than the offset of the
 930          * appended bytes
 931          */
 932         out = rte_pktmbuf_mtod_offset(m_out, uint8_t *, out_offset);
 933         if (check_crc_24b)
 934                 turbo_req.c = c + 1;
 935         else
 936                 turbo_req.c = c;
 937         turbo_req.input = (int8_t *)q->adapter_output;
 938         turbo_req.k = k;
 939         turbo_req.k_idx = k_idx;
 940         turbo_req.max_iter_num = dec->iter_max;
 941         turbo_resp.ag_buf = q->ag;
 942         turbo_resp.cb_buf = q->code_block;
 943         turbo_resp.output = out;
 944         iter_cnt = bblib_turbo_decoder(&turbo_req, &turbo_resp);
 945         dec->hard_output.length += (k >> 3);
 946
 947         if (iter_cnt > 0) {
 948                 /* Temporary solution for returned iter_count from SDK */
 949                 iter_cnt = (iter_cnt - 1) / 2;
 950                 dec->iter_count = RTE_MAX(iter_cnt, dec->iter_count);
 951         } else {
 952                 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 953                 rte_bbdev_log(ERR, "Turbo Decoder failed");
 954                 return;
 955         }
 956 }
 957
 958 static inline void
 959 enqueue_dec_one_op(struct turbo_sw_queue *q, struct rte_bbdev_dec_op *op)
 960 {
 961         uint8_t c, r = 0;
 962         uint16_t kw, k = 0;
 963         struct rte_bbdev_op_turbo_dec *dec = &op->turbo_dec;
 964         struct rte_mbuf *m_in = dec->input.data;
 965         struct rte_mbuf *m_out = dec->hard_output.data;
 966         uint16_t in_offset = dec->input.offset;
 967         uint16_t total_left = dec->input.length;
 968         uint16_t out_offset = dec->hard_output.offset;
 969
 970         /* Clear op status */
 971         op->status = 0;
 972
 973         if (m_in == NULL || m_out == NULL) {
 974                 rte_bbdev_log(ERR, "Invalid mbuf pointer");
 975                 op->status = 1 << RTE_BBDEV_DATA_ERROR;
 976                 return;
 977         }
 978
 979         if (dec->code_block_mode == 0) { /* For Transport Block mode */
 980                 c = dec->tb_params.c;
 981         } else { /* For Code Block mode */
 982                 k = dec->cb_params.k;
 983                 c = 1;
 984         }
 985
 986         while (total_left > 0) {
 987                 if (dec->code_block_mode == 0)
 988                         k = (r < dec->tb_params.c_neg) ?
 989                                 dec->tb_params.k_neg : dec->tb_params.k_pos;
 990
 991                 /* Calculates circular buffer size (Kw).
 992                  * According to 3gpp 36.212 section 5.1.4.2
 993                  *   Kw = 3 * Kpi,
 994                  * where:
 995                  *   Kpi = nCol * nRow
 996                  * where nCol is 32 and nRow can be calculated from:
 997                  *   D =< nCol * nRow
 998                  * where D is the size of each output from turbo encoder block
 999                  * (k + 4).
1000                  */
1001                 kw = RTE_ALIGN_CEIL(k + 4, RTE_BBDEV_C_SUBBLOCK) * 3;
1002
1003                 process_dec_cb(q, op, c, k, kw, m_in, m_out, in_offset,
1004                                 out_offset, check_bit(dec->op_flags,
1005                                 RTE_BBDEV_TURBO_CRC_TYPE_24B), total_left);
1006                 /* As a result of decoding we get Code Block with included
1007                  * decoded CRC24 at the end of Code Block. Type of CRC24 is
1008                  * specified by flag.
1009                  */
1010
1011                 /* Update total_left */
1012                 total_left -= kw;
1013                 /* Update offsets for next CBs (if exist) */
1014                 in_offset += kw;
1015                 out_offset += (k >> 3);
1016                 r++;
1017         }
1018         if (total_left != 0) {
1019                 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
1020                 rte_bbdev_log(ERR,
1021                                 "Mismatch between mbuf length and included Circular buffer sizes");
1022         }
1023 }
1024
1025 static inline uint16_t
1026 enqueue_dec_all_ops(struct turbo_sw_queue *q, struct rte_bbdev_dec_op **ops,
1027                 uint16_t nb_ops)
1028 {
1029         uint16_t i;
1030
1031         for (i = 0; i < nb_ops; ++i)
1032                 enqueue_dec_one_op(q, ops[i]);
1033
1034         return rte_ring_enqueue_burst(q->processed_pkts, (void **)ops, nb_ops,
1035                         NULL);
1036 }
1037
1038 /* Enqueue burst */
1039 static uint16_t
1040 enqueue_enc_ops(struct rte_bbdev_queue_data *q_data,
1041                 struct rte_bbdev_enc_op **ops, uint16_t nb_ops)
1042 {
1043         void *queue = q_data->queue_private;
1044         struct turbo_sw_queue *q = queue;
1045         uint16_t nb_enqueued = 0;
1046
1047         nb_enqueued = enqueue_enc_all_ops(q, ops, nb_ops);
1048
1049         q_data->queue_stats.enqueue_err_count += nb_ops - nb_enqueued;
1050         q_data->queue_stats.enqueued_count += nb_enqueued;
1051
1052         return nb_enqueued;
1053 }
1054
1055 /* Enqueue burst */
1056 static uint16_t
1057 enqueue_dec_ops(struct rte_bbdev_queue_data *q_data,
1058                  struct rte_bbdev_dec_op **ops, uint16_t nb_ops)
1059 {
1060         void *queue = q_data->queue_private;
1061         struct turbo_sw_queue *q = queue;
1062         uint16_t nb_enqueued = 0;
1063
1064         nb_enqueued = enqueue_dec_all_ops(q, ops, nb_ops);
1065
1066         q_data->queue_stats.enqueue_err_count += nb_ops - nb_enqueued;
1067         q_data->queue_stats.enqueued_count += nb_enqueued;
1068
1069         return nb_enqueued;
1070 }
1071
1072 /* Dequeue decode burst */
1073 static uint16_t
1074 dequeue_dec_ops(struct rte_bbdev_queue_data *q_data,
1075                 struct rte_bbdev_dec_op **ops, uint16_t nb_ops)
1076 {
1077         struct turbo_sw_queue *q = q_data->queue_private;
1078         uint16_t nb_dequeued = rte_ring_dequeue_burst(q->processed_pkts,
1079                         (void **)ops, nb_ops, NULL);
1080         q_data->queue_stats.dequeued_count += nb_dequeued;
1081
1082         return nb_dequeued;
1083 }
1084
1085 /* Dequeue encode burst */
1086 static uint16_t
1087 dequeue_enc_ops(struct rte_bbdev_queue_data *q_data,
1088                 struct rte_bbdev_enc_op **ops, uint16_t nb_ops)
1089 {
1090         struct turbo_sw_queue *q = q_data->queue_private;
1091         uint16_t nb_dequeued = rte_ring_dequeue_burst(q->processed_pkts,
1092                         (void **)ops, nb_ops, NULL);
1093         q_data->queue_stats.dequeued_count += nb_dequeued;
1094
1095         return nb_dequeued;
1096 }
1097
1098 /* Parse 16bit integer from string argument */
1099 static inline int
1100 parse_u16_arg(const char *key, const char *value, void *extra_args)
1101 {
1102         uint16_t *u16 = extra_args;
1103         unsigned int long result;
1104
1105         if ((value == NULL) || (extra_args == NULL))
1106                 return -EINVAL;
1107         errno = 0;
1108         result = strtoul(value, NULL, 0);
1109         if ((result >= (1 << 16)) || (errno != 0)) {
1110                 rte_bbdev_log(ERR, "Invalid value %lu for %s", result, key);
1111                 return -ERANGE;
1112         }
1113         *u16 = (uint16_t)result;
1114         return 0;
1115 }
1116
1117 /* Parse parameters used to create device */
1118 static int
1119 parse_turbo_sw_params(struct turbo_sw_params *params, const char *input_args)
1120 {
1121         struct rte_kvargs *kvlist = NULL;
1122         int ret = 0;
1123
1124         if (params == NULL)
1125                 return -EINVAL;
1126         if (input_args) {
1127                 kvlist = rte_kvargs_parse(input_args, turbo_sw_valid_params);
1128                 if (kvlist == NULL)
1129                         return -EFAULT;
1130
1131                 ret = rte_kvargs_process(kvlist, turbo_sw_valid_params[0],
1132                                         &parse_u16_arg, &params->queues_num);
1133                 if (ret < 0)
1134                         goto exit;
1135
1136                 ret = rte_kvargs_process(kvlist, turbo_sw_valid_params[1],
1137                                         &parse_u16_arg, &params->socket_id);
1138                 if (ret < 0)
1139                         goto exit;
1140
1141                 if (params->socket_id >= RTE_MAX_NUMA_NODES) {
1142                         rte_bbdev_log(ERR, "Invalid socket, must be < %u",
1143                                         RTE_MAX_NUMA_NODES);
1144                         goto exit;
1145                 }
1146         }
1147
1148 exit:
1149         if (kvlist)
1150                 rte_kvargs_free(kvlist);
1151         return ret;
1152 }
1153
1154 /* Create device */
1155 static int
1156 turbo_sw_bbdev_create(struct rte_vdev_device *vdev,
1157                 struct turbo_sw_params *init_params)
1158 {
1159         struct rte_bbdev *bbdev;
1160         const char *name = rte_vdev_device_name(vdev);
1161
1162         bbdev = rte_bbdev_allocate(name);
1163         if (bbdev == NULL)
1164                 return -ENODEV;
1165
1166         bbdev->data->dev_private = rte_zmalloc_socket(name,
1167                         sizeof(struct bbdev_private), RTE_CACHE_LINE_SIZE,
1168                         init_params->socket_id);
1169         if (bbdev->data->dev_private == NULL) {
1170                 rte_bbdev_release(bbdev);
1171                 return -ENOMEM;
1172         }
1173
1174         bbdev->dev_ops = &pmd_ops;
1175         bbdev->device = &vdev->device;
1176         bbdev->data->socket_id = init_params->socket_id;
1177         bbdev->intr_handle = NULL;
1178
1179         /* register rx/tx burst functions for data path */
1180         bbdev->dequeue_enc_ops = dequeue_enc_ops;
1181         bbdev->dequeue_dec_ops = dequeue_dec_ops;
1182         bbdev->enqueue_enc_ops = enqueue_enc_ops;
1183         bbdev->enqueue_dec_ops = enqueue_dec_ops;
1184         ((struct bbdev_private *) bbdev->data->dev_private)->max_nb_queues =
1185                         init_params->queues_num;
1186
1187         return 0;
1188 }
1189
1190 /* Initialise device */
1191 static int
1192 turbo_sw_bbdev_probe(struct rte_vdev_device *vdev)
1193 {
1194         struct turbo_sw_params init_params = {
1195                 rte_socket_id(),
1196                 RTE_BBDEV_DEFAULT_MAX_NB_QUEUES
1197         };
1198         const char *name;
1199         const char *input_args;
1200
1201         if (vdev == NULL)
1202                 return -EINVAL;
1203
1204         name = rte_vdev_device_name(vdev);
1205         if (name == NULL)
1206                 return -EINVAL;
1207         input_args = rte_vdev_device_args(vdev);
1208         parse_turbo_sw_params(&init_params, input_args);
1209
1210         rte_bbdev_log_debug(
1211                         "Initialising %s on NUMA node %d with max queues: %d\n",
1212                         name, init_params.socket_id, init_params.queues_num);
1213
1214         return turbo_sw_bbdev_create(vdev, &init_params);
1215 }
1216
1217 /* Uninitialise device */
1218 static int
1219 turbo_sw_bbdev_remove(struct rte_vdev_device *vdev)
1220 {
1221         struct rte_bbdev *bbdev;
1222         const char *name;
1223
1224         if (vdev == NULL)
1225                 return -EINVAL;
1226
1227         name = rte_vdev_device_name(vdev);
1228         if (name == NULL)
1229                 return -EINVAL;
1230
1231         bbdev = rte_bbdev_get_named_dev(name);
1232         if (bbdev == NULL)
1233                 return -EINVAL;
1234
1235         rte_free(bbdev->data->dev_private);
1236
1237         return rte_bbdev_release(bbdev);
1238 }
1239
1240 static struct rte_vdev_driver bbdev_turbo_sw_pmd_drv = {
1241         .probe = turbo_sw_bbdev_probe,
1242         .remove = turbo_sw_bbdev_remove
1243 };
1244
1245 RTE_PMD_REGISTER_VDEV(DRIVER_NAME, bbdev_turbo_sw_pmd_drv);
1246 RTE_PMD_REGISTER_PARAM_STRING(DRIVER_NAME,
1247         TURBO_SW_MAX_NB_QUEUES_ARG"=<int> "
1248         TURBO_SW_SOCKET_ID_ARG"=<int>");
1249
1250 RTE_INIT(null_bbdev_init_log);
1251 static void
1252 null_bbdev_init_log(void)
1253 {
1254         bbdev_turbo_sw_logtype = rte_log_register("pmd.bb.turbo_sw");
1255         if (bbdev_turbo_sw_logtype >= 0)
1256                 rte_log_set_level(bbdev_turbo_sw_logtype, RTE_LOG_NOTICE);
1257 }