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 #include <phy_turbo.h>
  18 #include <phy_crc.h>
  19 #include <phy_rate_match.h>
  20 #include <divide.h>
  21
  22 #define DRIVER_NAME turbo_sw
  23
  24 /* Turbo SW PMD logging ID */
  25 static int bbdev_turbo_sw_logtype;
  26
  27 /* Helper macro for logging */
  28 #define rte_bbdev_log(level, fmt, ...) \
  29         rte_log(RTE_LOG_ ## level, bbdev_turbo_sw_logtype, fmt "\n", \
  30                 ##__VA_ARGS__)
  31
  32 #define rte_bbdev_log_debug(fmt, ...) \
  33         rte_bbdev_log(DEBUG, RTE_STR(__LINE__) ":%s() " fmt, __func__, \
  34                 ##__VA_ARGS__)
  35
  36 #define DEINT_INPUT_BUF_SIZE (((RTE_BBDEV_MAX_CB_SIZE >> 3) + 1) * 48)
  37 #define DEINT_OUTPUT_BUF_SIZE (DEINT_INPUT_BUF_SIZE * 6)
  38 #define ADAPTER_OUTPUT_BUF_SIZE ((RTE_BBDEV_MAX_CB_SIZE + 4) * 48)
  39
  40 /* private data structure */
  41 struct bbdev_private {
  42         unsigned int max_nb_queues;  /**< Max number of queues */
  43 };
  44
  45 /*  Initialisation params structure that can be used by Turbo SW driver */
  46 struct turbo_sw_params {
  47         int socket_id;  /*< Turbo SW device socket */
  48         uint16_t queues_num;  /*< Turbo SW device queues number */
  49 };
  50
  51 /* Accecptable params for Turbo SW devices */
  52 #define TURBO_SW_MAX_NB_QUEUES_ARG  "max_nb_queues"
  53 #define TURBO_SW_SOCKET_ID_ARG      "socket_id"
  54
  55 static const char * const turbo_sw_valid_params[] = {
  56         TURBO_SW_MAX_NB_QUEUES_ARG,
  57         TURBO_SW_SOCKET_ID_ARG
  58 };
  59
  60 /* queue */
  61 struct turbo_sw_queue {
  62         /* Ring for processed (encoded/decoded) operations which are ready to
  63          * be dequeued.
  64          */
  65         struct rte_ring *processed_pkts;
  66         /* Stores input for turbo encoder (used when CRC attachment is
  67          * performed
  68          */
  69         uint8_t *enc_in;
  70         /* Stores output from turbo encoder */
  71         uint8_t *enc_out;
  72         /* Alpha gamma buf for bblib_turbo_decoder() function */
  73         int8_t *ag;
  74         /* Temp buf for bblib_turbo_decoder() function */
  75         uint16_t *code_block;
  76         /* Input buf for bblib_rate_dematching_lte() function */
  77         uint8_t *deint_input;
  78         /* Output buf for bblib_rate_dematching_lte() function */
  79         uint8_t *deint_output;
  80         /* Output buf for bblib_turbodec_adapter_lte() function */
  81         uint8_t *adapter_output;
  82         /* Operation type of this queue */
  83         enum rte_bbdev_op_type type;
  84 } __rte_cache_aligned;
  85
  86 /* Calculate index based on Table 5.1.3-3 from TS34.212 */
  87 static inline int32_t
  88 compute_idx(uint16_t k)
  89 {
  90         int32_t result = 0;
  91
  92         if (k < RTE_BBDEV_MIN_CB_SIZE || k > RTE_BBDEV_MAX_CB_SIZE)
  93                 return -1;
  94
  95         if (k > 2048) {
  96                 if ((k - 2048) % 64 != 0)
  97                         result = -1;
  98
  99                 result = 124 + (k - 2048) / 64;
 100         } else if (k <= 512) {
 101                 if ((k - 40) % 8 != 0)
 102                         result = -1;
 103
 104                 result = (k - 40) / 8 + 1;
 105         } else if (k <= 1024) {
 106                 if ((k - 512) % 16 != 0)
 107                         result = -1;
 108
 109                 result = 60 + (k - 512) / 16;
 110         } else { /* 1024 < k <= 2048 */
 111                 if ((k - 1024) % 32 != 0)
 112                         result = -1;
 113
 114                 result = 92 + (k - 1024) / 32;
 115         }
 116
 117         return result;
 118 }
 119
 120 /* Read flag value 0/1 from bitmap */
 121 static inline bool
 122 check_bit(uint32_t bitmap, uint32_t bitmask)
 123 {
 124         return bitmap & bitmask;
 125 }
 126
 127 /* Get device info */
 128 static void
 129 info_get(struct rte_bbdev *dev, struct rte_bbdev_driver_info *dev_info)
 130 {
 131         struct bbdev_private *internals = dev->data->dev_private;
 132
 133         static const struct rte_bbdev_op_cap bbdev_capabilities[] = {
 134                 {
 135                         .type = RTE_BBDEV_OP_TURBO_DEC,
 136                         .cap.turbo_dec = {
 137                                 .capability_flags =
 138                                         RTE_BBDEV_TURBO_SUBBLOCK_DEINTERLEAVE |
 139                                         RTE_BBDEV_TURBO_POS_LLR_1_BIT_IN |
 140                                         RTE_BBDEV_TURBO_NEG_LLR_1_BIT_IN |
 141                                         RTE_BBDEV_TURBO_CRC_TYPE_24B |
 142                                         RTE_BBDEV_TURBO_DEC_TB_CRC_24B_KEEP |
 143                                         RTE_BBDEV_TURBO_EARLY_TERMINATION,
 144                                 .max_llr_modulus = 16,
 145                                 .num_buffers_src = RTE_BBDEV_MAX_CODE_BLOCKS,
 146                                 .num_buffers_hard_out =
 147                                                 RTE_BBDEV_MAX_CODE_BLOCKS,
 148                                 .num_buffers_soft_out = 0,
 149                         }
 150                 },
 151                 {
 152                         .type   = RTE_BBDEV_OP_TURBO_ENC,
 153                         .cap.turbo_enc = {
 154                                 .capability_flags =
 155                                                 RTE_BBDEV_TURBO_CRC_24B_ATTACH |
 156                                                 RTE_BBDEV_TURBO_CRC_24A_ATTACH |
 157                                                 RTE_BBDEV_TURBO_RATE_MATCH |
 158                                                 RTE_BBDEV_TURBO_RV_INDEX_BYPASS,
 159                                 .num_buffers_src = RTE_BBDEV_MAX_CODE_BLOCKS,
 160                                 .num_buffers_dst = RTE_BBDEV_MAX_CODE_BLOCKS,
 161                         }
 162                 },
 163                 RTE_BBDEV_END_OF_CAPABILITIES_LIST()
 164         };
 165
 166         static struct rte_bbdev_queue_conf default_queue_conf = {
 167                 .queue_size = RTE_BBDEV_QUEUE_SIZE_LIMIT,
 168         };
 169
 170         static const enum rte_cpu_flag_t cpu_flag = RTE_CPUFLAG_SSE4_2;
 171
 172         default_queue_conf.socket = dev->data->socket_id;
 173
 174         dev_info->driver_name = RTE_STR(DRIVER_NAME);
 175         dev_info->max_num_queues = internals->max_nb_queues;
 176         dev_info->queue_size_lim = RTE_BBDEV_QUEUE_SIZE_LIMIT;
 177         dev_info->hardware_accelerated = false;
 178         dev_info->max_queue_priority = 0;
 179         dev_info->default_queue_conf = default_queue_conf;
 180         dev_info->capabilities = bbdev_capabilities;
 181         dev_info->cpu_flag_reqs = &cpu_flag;
 182         dev_info->min_alignment = 64;
 183
 184         rte_bbdev_log_debug("got device info from %u\n", dev->data->dev_id);
 185 }
 186
 187 /* Release queue */
 188 static int
 189 q_release(struct rte_bbdev *dev, uint16_t q_id)
 190 {
 191         struct turbo_sw_queue *q = dev->data->queues[q_id].queue_private;
 192
 193         if (q != NULL) {
 194                 rte_ring_free(q->processed_pkts);
 195                 rte_free(q->enc_out);
 196                 rte_free(q->enc_in);
 197                 rte_free(q->ag);
 198                 rte_free(q->code_block);
 199                 rte_free(q->deint_input);
 200                 rte_free(q->deint_output);
 201                 rte_free(q->adapter_output);
 202                 rte_free(q);
 203                 dev->data->queues[q_id].queue_private = NULL;
 204         }
 205
 206         rte_bbdev_log_debug("released device queue %u:%u",
 207                         dev->data->dev_id, q_id);
 208         return 0;
 209 }
 210
 211 /* Setup a queue */
 212 static int
 213 q_setup(struct rte_bbdev *dev, uint16_t q_id,
 214                 const struct rte_bbdev_queue_conf *queue_conf)
 215 {
 216         int ret;
 217         struct turbo_sw_queue *q;
 218         char name[RTE_RING_NAMESIZE];
 219
 220         /* Allocate the queue data structure. */
 221         q = rte_zmalloc_socket(RTE_STR(DRIVER_NAME), sizeof(*q),
 222                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 223         if (q == NULL) {
 224                 rte_bbdev_log(ERR, "Failed to allocate queue memory");
 225                 return -ENOMEM;
 226         }
 227
 228         /* Allocate memory for encoder output. */
 229         ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"_enc_out%u:%u",
 230                         dev->data->dev_id, q_id);
 231         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 232                 rte_bbdev_log(ERR,
 233                                 "Creating queue name for device %u queue %u failed",
 234                                 dev->data->dev_id, q_id);
 235                 return -ENAMETOOLONG;
 236         }
 237         q->enc_out = rte_zmalloc_socket(name,
 238                         ((RTE_BBDEV_MAX_TB_SIZE >> 3) + 3) *
 239                         sizeof(*q->enc_out) * 3,
 240                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 241         if (q->enc_out == NULL) {
 242                 rte_bbdev_log(ERR,
 243                         "Failed to allocate queue memory for %s", name);
 244                 goto free_q;
 245         }
 246
 247         /* Allocate memory for rate matching output. */
 248         ret = snprintf(name, RTE_RING_NAMESIZE,
 249                         RTE_STR(DRIVER_NAME)"_enc_in%u:%u", dev->data->dev_id,
 250                         q_id);
 251         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 252                 rte_bbdev_log(ERR,
 253                                 "Creating queue name for device %u queue %u failed",
 254                                 dev->data->dev_id, q_id);
 255                 return -ENAMETOOLONG;
 256         }
 257         q->enc_in = rte_zmalloc_socket(name,
 258                         (RTE_BBDEV_MAX_CB_SIZE >> 3) * sizeof(*q->enc_in),
 259                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 260         if (q->enc_in == NULL) {
 261                 rte_bbdev_log(ERR,
 262                         "Failed to allocate queue memory for %s", name);
 263                 goto free_q;
 264         }
 265
 266         /* Allocate memory for Aplha Gamma temp buffer. */
 267         ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"_ag%u:%u",
 268                         dev->data->dev_id, q_id);
 269         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 270                 rte_bbdev_log(ERR,
 271                                 "Creating queue name for device %u queue %u failed",
 272                                 dev->data->dev_id, q_id);
 273                 return -ENAMETOOLONG;
 274         }
 275         q->ag = rte_zmalloc_socket(name,
 276                         RTE_BBDEV_MAX_CB_SIZE * 10 * sizeof(*q->ag),
 277                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 278         if (q->ag == NULL) {
 279                 rte_bbdev_log(ERR,
 280                         "Failed to allocate queue memory for %s", name);
 281                 goto free_q;
 282         }
 283
 284         /* Allocate memory for code block temp buffer. */
 285         ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"_cb%u:%u",
 286                         dev->data->dev_id, q_id);
 287         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 288                 rte_bbdev_log(ERR,
 289                                 "Creating queue name for device %u queue %u failed",
 290                                 dev->data->dev_id, q_id);
 291                 return -ENAMETOOLONG;
 292         }
 293         q->code_block = rte_zmalloc_socket(name,
 294                         RTE_BBDEV_MAX_CB_SIZE * sizeof(*q->code_block),
 295                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 296         if (q->code_block == NULL) {
 297                 rte_bbdev_log(ERR,
 298                         "Failed to allocate queue memory for %s", name);
 299                 goto free_q;
 300         }
 301
 302         /* Allocate memory for Deinterleaver input. */
 303         ret = snprintf(name, RTE_RING_NAMESIZE,
 304                         RTE_STR(DRIVER_NAME)"_deint_input%u:%u",
 305                         dev->data->dev_id, q_id);
 306         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 307                 rte_bbdev_log(ERR,
 308                                 "Creating queue name for device %u queue %u failed",
 309                                 dev->data->dev_id, q_id);
 310                 return -ENAMETOOLONG;
 311         }
 312         q->deint_input = rte_zmalloc_socket(name,
 313                         DEINT_INPUT_BUF_SIZE * sizeof(*q->deint_input),
 314                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 315         if (q->deint_input == NULL) {
 316                 rte_bbdev_log(ERR,
 317                         "Failed to allocate queue memory for %s", name);
 318                 goto free_q;
 319         }
 320
 321         /* Allocate memory for Deinterleaver output. */
 322         ret = snprintf(name, RTE_RING_NAMESIZE,
 323                         RTE_STR(DRIVER_NAME)"_deint_output%u:%u",
 324                         dev->data->dev_id, q_id);
 325         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 326                 rte_bbdev_log(ERR,
 327                                 "Creating queue name for device %u queue %u failed",
 328                                 dev->data->dev_id, q_id);
 329                 return -ENAMETOOLONG;
 330         }
 331         q->deint_output = rte_zmalloc_socket(NULL,
 332                         DEINT_OUTPUT_BUF_SIZE * sizeof(*q->deint_output),
 333                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 334         if (q->deint_output == NULL) {
 335                 rte_bbdev_log(ERR,
 336                         "Failed to allocate queue memory for %s", name);
 337                 goto free_q;
 338         }
 339
 340         /* Allocate memory for Adapter output. */
 341         ret = snprintf(name, RTE_RING_NAMESIZE,
 342                         RTE_STR(DRIVER_NAME)"_adapter_output%u:%u",
 343                         dev->data->dev_id, q_id);
 344         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 345                 rte_bbdev_log(ERR,
 346                                 "Creating queue name for device %u queue %u failed",
 347                                 dev->data->dev_id, q_id);
 348                 return -ENAMETOOLONG;
 349         }
 350         q->adapter_output = rte_zmalloc_socket(NULL,
 351                         ADAPTER_OUTPUT_BUF_SIZE * sizeof(*q->adapter_output),
 352                         RTE_CACHE_LINE_SIZE, queue_conf->socket);
 353         if (q->adapter_output == NULL) {
 354                 rte_bbdev_log(ERR,
 355                         "Failed to allocate queue memory for %s", name);
 356                 goto free_q;
 357         }
 358
 359         /* Create ring for packets awaiting to be dequeued. */
 360         ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"%u:%u",
 361                         dev->data->dev_id, q_id);
 362         if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
 363                 rte_bbdev_log(ERR,
 364                                 "Creating queue name for device %u queue %u failed",
 365                                 dev->data->dev_id, q_id);
 366                 return -ENAMETOOLONG;
 367         }
 368         q->processed_pkts = rte_ring_create(name, queue_conf->queue_size,
 369                         queue_conf->socket, RING_F_SP_ENQ | RING_F_SC_DEQ);
 370         if (q->processed_pkts == NULL) {
 371                 rte_bbdev_log(ERR, "Failed to create ring for %s", name);
 372                 goto free_q;
 373         }
 374
 375         q->type = queue_conf->op_type;
 376
 377         dev->data->queues[q_id].queue_private = q;
 378         rte_bbdev_log_debug("setup device queue %s", name);
 379         return 0;
 380
 381 free_q:
 382         rte_ring_free(q->processed_pkts);
 383         rte_free(q->enc_out);
 384         rte_free(q->enc_in);
 385         rte_free(q->ag);
 386         rte_free(q->code_block);
 387         rte_free(q->deint_input);
 388         rte_free(q->deint_output);
 389         rte_free(q->adapter_output);
 390         rte_free(q);
 391         return -EFAULT;
 392 }
 393
 394 static const struct rte_bbdev_ops pmd_ops = {
 395         .info_get = info_get,
 396         .queue_setup = q_setup,
 397         .queue_release = q_release
 398 };
 399
 400 /* Checks if the encoder input buffer is correct.
 401  * Returns 0 if it's valid, -1 otherwise.
 402  */
 403 static inline int
 404 is_enc_input_valid(const uint16_t k, const int32_t k_idx,
 405                 const uint16_t in_length)
 406 {
 407         if (k_idx < 0) {
 408                 rte_bbdev_log(ERR, "K Index is invalid");
 409                 return -1;
 410         }
 411
 412         if (in_length - (k >> 3) < 0) {
 413                 rte_bbdev_log(ERR,
 414                                 "Mismatch between input length (%u bytes) and K (%u bits)",
 415                                 in_length, k);
 416                 return -1;
 417         }
 418
 419         if (k > RTE_BBDEV_MAX_CB_SIZE) {
 420                 rte_bbdev_log(ERR, "CB size (%u) is too big, max: %d",
 421                                 k, RTE_BBDEV_MAX_CB_SIZE);
 422                 return -1;
 423         }
 424
 425         return 0;
 426 }
 427
 428 /* Checks if the decoder input buffer is correct.
 429  * Returns 0 if it's valid, -1 otherwise.
 430  */
 431 static inline int
 432 is_dec_input_valid(int32_t k_idx, int16_t kw, int16_t in_length)
 433 {
 434         if (k_idx < 0) {
 435                 rte_bbdev_log(ERR, "K index is invalid");
 436                 return -1;
 437         }
 438
 439         if (in_length - kw < 0) {
 440                 rte_bbdev_log(ERR,
 441                                 "Mismatch between input length (%u) and kw (%u)",
 442                                 in_length, kw);
 443                 return -1;
 444         }
 445
 446         if (kw > RTE_BBDEV_MAX_KW) {
 447                 rte_bbdev_log(ERR, "Input length (%u) is too big, max: %d",
 448                                 kw, RTE_BBDEV_MAX_KW);
 449                 return -1;
 450         }
 451
 452         return 0;
 453 }
 454
 455 static inline void
 456 process_enc_cb(struct turbo_sw_queue *q, struct rte_bbdev_enc_op *op,
 457                 uint8_t r, uint8_t c, uint16_t k, uint16_t ncb,
 458                 uint32_t e, struct rte_mbuf *m_in, struct rte_mbuf *m_out,
 459                 uint16_t in_offset, uint16_t out_offset, uint16_t total_left,
 460                 struct rte_bbdev_stats *q_stats)
 461 {
 462         int ret;
 463         int16_t k_idx;
 464         uint16_t m;
 465         uint8_t *in, *out0, *out1, *out2, *tmp_out, *rm_out;
 466         uint64_t first_3_bytes = 0;
 467         struct rte_bbdev_op_turbo_enc *enc = &op->turbo_enc;
 468         struct bblib_crc_request crc_req;
 469         struct bblib_crc_response crc_resp;
 470         struct bblib_turbo_encoder_request turbo_req;
 471         struct bblib_turbo_encoder_response turbo_resp;
 472         struct bblib_rate_match_dl_request rm_req;
 473         struct bblib_rate_match_dl_response rm_resp;
 474 #ifdef RTE_BBDEV_OFFLOAD_COST
 475         uint64_t start_time;
 476 #else
 477         RTE_SET_USED(q_stats);
 478 #endif
 479
 480         k_idx = compute_idx(k);
 481         in = rte_pktmbuf_mtod_offset(m_in, uint8_t *, in_offset);
 482
 483         /* CRC24A (for TB) */
 484         if ((enc->op_flags & RTE_BBDEV_TURBO_CRC_24A_ATTACH) &&
 485                 (enc->code_block_mode == 1)) {
 486                 ret = is_enc_input_valid(k - 24, k_idx, total_left);
 487                 if (ret != 0) {
 488                         op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 489                         return;
 490                 }
 491                 crc_req.data = in;
 492                 crc_req.len = (k - 24) >> 3;
 493                 /* Check if there is a room for CRC bits. If not use
 494                  * the temporary buffer.
 495                  */
 496                 if (rte_pktmbuf_append(m_in, 3) == NULL) {
 497                         rte_memcpy(q->enc_in, in, (k - 24) >> 3);
 498                         in = q->enc_in;
 499                 } else {
 500                         /* Store 3 first bytes of next CB as they will be
 501                          * overwritten by CRC bytes. If it is the last CB then
 502                          * there is no point to store 3 next bytes and this
 503                          * if..else branch will be omitted.
 504                          */
 505                         first_3_bytes = *((uint64_t *)&in[(k - 32) >> 3]);
 506                 }
 507
 508                 crc_resp.data = in;
 509 #ifdef RTE_BBDEV_OFFLOAD_COST
 510                 start_time = rte_rdtsc_precise();
 511 #endif
 512                 bblib_lte_crc24a_gen(&crc_req, &crc_resp);
 513 #ifdef RTE_BBDEV_OFFLOAD_COST
 514                 q_stats->offload_time += rte_rdtsc_precise() - start_time;
 515 #endif
 516         } else if (enc->op_flags & RTE_BBDEV_TURBO_CRC_24B_ATTACH) {
 517                 /* CRC24B */
 518                 ret = is_enc_input_valid(k - 24, k_idx, total_left);
 519                 if (ret != 0) {
 520                         op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 521                         return;
 522                 }
 523                 crc_req.data = in;
 524                 crc_req.len = (k - 24) >> 3;
 525                 /* Check if there is a room for CRC bits. If this is the last
 526                  * CB in TB. If not use temporary buffer.
 527                  */
 528                 if ((c - r == 1) && (rte_pktmbuf_append(m_in, 3) == NULL)) {
 529                         rte_memcpy(q->enc_in, in, (k - 24) >> 3);
 530                         in = q->enc_in;
 531                 } else if (c - r > 1) {
 532                         /* Store 3 first bytes of next CB as they will be
 533                          * overwritten by CRC bytes. If it is the last CB then
 534                          * there is no point to store 3 next bytes and this
 535                          * if..else branch will be omitted.
 536                          */
 537                         first_3_bytes = *((uint64_t *)&in[(k - 32) >> 3]);
 538                 }
 539
 540                 crc_resp.data = in;
 541 #ifdef RTE_BBDEV_OFFLOAD_COST
 542                 start_time = rte_rdtsc_precise();
 543 #endif
 544                 bblib_lte_crc24b_gen(&crc_req, &crc_resp);
 545 #ifdef RTE_BBDEV_OFFLOAD_COST
 546                 q_stats->offload_time += rte_rdtsc_precise() - start_time;
 547 #endif
 548         } else {
 549                 ret = is_enc_input_valid(k, k_idx, total_left);
 550                 if (ret != 0) {
 551                         op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 552                         return;
 553                 }
 554         }
 555
 556         /* Turbo encoder */
 557
 558         /* Each bit layer output from turbo encoder is (k+4) bits long, i.e.
 559          * input length + 4 tail bits. That's (k/8) + 1 bytes after rounding up.
 560          * So dst_data's length should be 3*(k/8) + 3 bytes.
 561          * In Rate-matching bypass case outputs pointers passed to encoder
 562          * (out0, out1 and out2) can directly point to addresses of output from
 563          * turbo_enc entity.
 564          */
 565         if (enc->op_flags & RTE_BBDEV_TURBO_RATE_MATCH) {
 566                 out0 = q->enc_out;
 567                 out1 = RTE_PTR_ADD(out0, (k >> 3) + 1);
 568                 out2 = RTE_PTR_ADD(out1, (k >> 3) + 1);
 569         } else {
 570                 out0 = (uint8_t *)rte_pktmbuf_append(m_out, (k >> 3) * 3 + 2);
 571                 if (out0 == NULL) {
 572                         op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 573                         rte_bbdev_log(ERR,
 574                                         "Too little space in output mbuf");
 575                         return;
 576                 }
 577                 enc->output.length += (k >> 3) * 3 + 2;
 578                 /* rte_bbdev_op_data.offset can be different than the
 579                  * offset of the appended bytes
 580                  */
 581                 out0 = rte_pktmbuf_mtod_offset(m_out, uint8_t *, out_offset);
 582                 out1 = rte_pktmbuf_mtod_offset(m_out, uint8_t *,
 583                                 out_offset + (k >> 3) + 1);
 584                 out2 = rte_pktmbuf_mtod_offset(m_out, uint8_t *,
 585                                 out_offset + 2 * ((k >> 3) + 1));
 586         }
 587
 588         turbo_req.case_id = k_idx;
 589         turbo_req.input_win = in;
 590         turbo_req.length = k >> 3;
 591         turbo_resp.output_win_0 = out0;
 592         turbo_resp.output_win_1 = out1;
 593         turbo_resp.output_win_2 = out2;
 594
 595 #ifdef RTE_BBDEV_OFFLOAD_COST
 596         start_time = rte_rdtsc_precise();
 597 #endif
 598
 599         if (bblib_turbo_encoder(&turbo_req, &turbo_resp) != 0) {
 600                 op->status |= 1 << RTE_BBDEV_DRV_ERROR;
 601                 rte_bbdev_log(ERR, "Turbo Encoder failed");
 602                 return;
 603         }
 604
 605 #ifdef RTE_BBDEV_OFFLOAD_COST
 606         q_stats->offload_time += rte_rdtsc_precise() - start_time;
 607 #endif
 608
 609         /* Restore 3 first bytes of next CB if they were overwritten by CRC*/
 610         if (first_3_bytes != 0)
 611                 *((uint64_t *)&in[(k - 32) >> 3]) = first_3_bytes;
 612
 613         /* Rate-matching */
 614         if (enc->op_flags & RTE_BBDEV_TURBO_RATE_MATCH) {
 615                 uint8_t mask_id;
 616                 /* Integer round up division by 8 */
 617                 uint16_t out_len = (e + 7) >> 3;
 618                 /* The mask array is indexed using E%8. E is an even number so
 619                  * there are only 4 possible values.
 620                  */
 621                 const uint8_t mask_out[] = {0xFF, 0xC0, 0xF0, 0xFC};
 622
 623                 /* get output data starting address */
 624                 rm_out = (uint8_t *)rte_pktmbuf_append(m_out, out_len);
 625                 if (rm_out == NULL) {
 626                         op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 627                         rte_bbdev_log(ERR,
 628                                         "Too little space in output mbuf");
 629                         return;
 630                 }
 631                 /* rte_bbdev_op_data.offset can be different than the offset
 632                  * of the appended bytes
 633                  */
 634                 rm_out = rte_pktmbuf_mtod_offset(m_out, uint8_t *, out_offset);
 635
 636                 /* index of current code block */
 637                 rm_req.r = r;
 638                 /* total number of code block */
 639                 rm_req.C = c;
 640                 /* For DL - 1, UL - 0 */
 641                 rm_req.direction = 1;
 642                 /* According to 3ggp 36.212 Spec 5.1.4.1.2 section Nsoft, KMIMO
 643                  * and MDL_HARQ are used for Ncb calculation. As Ncb is already
 644                  * known we can adjust those parameters
 645                  */
 646                 rm_req.Nsoft = ncb * rm_req.C;
 647                 rm_req.KMIMO = 1;
 648                 rm_req.MDL_HARQ = 1;
 649                 /* According to 3ggp 36.212 Spec 5.1.4.1.2 section Nl, Qm and G
 650                  * are used for E calculation. As E is already known we can
 651                  * adjust those parameters
 652                  */
 653                 rm_req.NL = e;
 654                 rm_req.Qm = 1;
 655                 rm_req.G = rm_req.NL * rm_req.Qm * rm_req.C;
 656
 657                 rm_req.rvidx = enc->rv_index;
 658                 rm_req.Kidx = k_idx - 1;
 659                 rm_req.nLen = k + 4;
 660                 rm_req.tin0 = out0;
 661                 rm_req.tin1 = out1;
 662                 rm_req.tin2 = out2;
 663                 rm_resp.output = rm_out;
 664                 rm_resp.OutputLen = out_len;
 665                 if (enc->op_flags & RTE_BBDEV_TURBO_RV_INDEX_BYPASS)
 666                         rm_req.bypass_rvidx = 1;
 667                 else
 668                         rm_req.bypass_rvidx = 0;
 669
 670 #ifdef RTE_BBDEV_OFFLOAD_COST
 671                 start_time = rte_rdtsc_precise();
 672 #endif
 673
 674                 if (bblib_rate_match_dl(&rm_req, &rm_resp) != 0) {
 675                         op->status |= 1 << RTE_BBDEV_DRV_ERROR;
 676                         rte_bbdev_log(ERR, "Rate matching failed");
 677                         return;
 678                 }
 679
 680                 /* SW fills an entire last byte even if E%8 != 0. Clear the
 681                  * superfluous data bits for consistency with HW device.
 682                  */
 683                 mask_id = (e & 7) >> 1;
 684                 rm_out[out_len - 1] &= mask_out[mask_id];
 685
 686 #ifdef RTE_BBDEV_OFFLOAD_COST
 687                 q_stats->offload_time += rte_rdtsc_precise() - start_time;
 688 #endif
 689
 690                 enc->output.length += rm_resp.OutputLen;
 691         } else {
 692                 /* Rate matching is bypassed */
 693
 694                 /* Completing last byte of out0 (where 4 tail bits are stored)
 695                  * by moving first 4 bits from out1
 696                  */
 697                 tmp_out = (uint8_t *) --out1;
 698                 *tmp_out = *tmp_out | ((*(tmp_out + 1) & 0xF0) >> 4);
 699                 tmp_out++;
 700                 /* Shifting out1 data by 4 bits to the left */
 701                 for (m = 0; m < k >> 3; ++m) {
 702                         uint8_t *first = tmp_out;
 703                         uint8_t second = *(tmp_out + 1);
 704                         *first = (*first << 4) | ((second & 0xF0) >> 4);
 705                         tmp_out++;
 706                 }
 707                 /* Shifting out2 data by 8 bits to the left */
 708                 for (m = 0; m < (k >> 3) + 1; ++m) {
 709                         *tmp_out = *(tmp_out + 1);
 710                         tmp_out++;
 711                 }
 712                 *tmp_out = 0;
 713         }
 714 }
 715
 716 static inline void
 717 enqueue_enc_one_op(struct turbo_sw_queue *q, struct rte_bbdev_enc_op *op,
 718                 struct rte_bbdev_stats *queue_stats)
 719 {
 720         uint8_t c, r, crc24_bits = 0;
 721         uint16_t k, ncb;
 722         uint32_t e;
 723         struct rte_bbdev_op_turbo_enc *enc = &op->turbo_enc;
 724         uint16_t in_offset = enc->input.offset;
 725         uint16_t out_offset = enc->output.offset;
 726         struct rte_mbuf *m_in = enc->input.data;
 727         struct rte_mbuf *m_out = enc->output.data;
 728         uint16_t total_left = enc->input.length;
 729
 730         /* Clear op status */
 731         op->status = 0;
 732
 733         if (total_left > RTE_BBDEV_MAX_TB_SIZE >> 3) {
 734                 rte_bbdev_log(ERR, "TB size (%u) is too big, max: %d",
 735                                 total_left, RTE_BBDEV_MAX_TB_SIZE);
 736                 op->status = 1 << RTE_BBDEV_DATA_ERROR;
 737                 return;
 738         }
 739
 740         if (m_in == NULL || m_out == NULL) {
 741                 rte_bbdev_log(ERR, "Invalid mbuf pointer");
 742                 op->status = 1 << RTE_BBDEV_DATA_ERROR;
 743                 return;
 744         }
 745
 746         if ((enc->op_flags & RTE_BBDEV_TURBO_CRC_24B_ATTACH) ||
 747                 (enc->op_flags & RTE_BBDEV_TURBO_CRC_24A_ATTACH))
 748                 crc24_bits = 24;
 749
 750         if (enc->code_block_mode == 0) { /* For Transport Block mode */
 751                 c = enc->tb_params.c;
 752                 r = enc->tb_params.r;
 753         } else {/* For Code Block mode */
 754                 c = 1;
 755                 r = 0;
 756         }
 757
 758         while (total_left > 0 && r < c) {
 759                 if (enc->code_block_mode == 0) {
 760                         k = (r < enc->tb_params.c_neg) ?
 761                                 enc->tb_params.k_neg : enc->tb_params.k_pos;
 762                         ncb = (r < enc->tb_params.c_neg) ?
 763                                 enc->tb_params.ncb_neg : enc->tb_params.ncb_pos;
 764                         e = (r < enc->tb_params.cab) ?
 765                                 enc->tb_params.ea : enc->tb_params.eb;
 766                 } else {
 767                         k = enc->cb_params.k;
 768                         ncb = enc->cb_params.ncb;
 769                         e = enc->cb_params.e;
 770                 }
 771
 772                 process_enc_cb(q, op, r, c, k, ncb, e, m_in,
 773                                 m_out, in_offset, out_offset, total_left,
 774                                 queue_stats);
 775                 /* Update total_left */
 776                 total_left -= (k - crc24_bits) >> 3;
 777                 /* Update offsets for next CBs (if exist) */
 778                 in_offset += (k - crc24_bits) >> 3;
 779                 if (enc->op_flags & RTE_BBDEV_TURBO_RATE_MATCH)
 780                         out_offset += e >> 3;
 781                 else
 782                         out_offset += (k >> 3) * 3 + 2;
 783                 r++;
 784         }
 785
 786         /* check if all input data was processed */
 787         if (total_left != 0) {
 788                 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 789                 rte_bbdev_log(ERR,
 790                                 "Mismatch between mbuf length and included CBs sizes");
 791         }
 792 }
 793
 794 static inline uint16_t
 795 enqueue_enc_all_ops(struct turbo_sw_queue *q, struct rte_bbdev_enc_op **ops,
 796                 uint16_t nb_ops, struct rte_bbdev_stats *queue_stats)
 797 {
 798         uint16_t i;
 799 #ifdef RTE_BBDEV_OFFLOAD_COST
 800         queue_stats->offload_time = 0;
 801 #endif
 802
 803         for (i = 0; i < nb_ops; ++i)
 804                 enqueue_enc_one_op(q, ops[i], queue_stats);
 805
 806         return rte_ring_enqueue_burst(q->processed_pkts, (void **)ops, nb_ops,
 807                         NULL);
 808 }
 809
 810 /* Remove the padding bytes from a cyclic buffer.
 811  * The input buffer is a data stream wk as described in 3GPP TS 36.212 section
 812  * 5.1.4.1.2 starting from w0 and with length Ncb bytes.
 813  * The output buffer is a data stream wk with pruned padding bytes. It's length
 814  * is 3*D bytes and the order of non-padding bytes is preserved.
 815  */
 816 static inline void
 817 remove_nulls_from_circular_buf(const uint8_t *in, uint8_t *out, uint16_t k,
 818                 uint16_t ncb)
 819 {
 820         uint32_t in_idx, out_idx, c_idx;
 821         const uint32_t d = k + 4;
 822         const uint32_t kw = (ncb / 3);
 823         const uint32_t nd = kw - d;
 824         const uint32_t r_subblock = kw / RTE_BBDEV_C_SUBBLOCK;
 825         /* Inter-column permutation pattern */
 826         const uint32_t P[RTE_BBDEV_C_SUBBLOCK] = {0, 16, 8, 24, 4, 20, 12, 28,
 827                         2, 18, 10, 26, 6, 22, 14, 30, 1, 17, 9, 25, 5, 21, 13,
 828                         29, 3, 19, 11, 27, 7, 23, 15, 31};
 829         in_idx = 0;
 830         out_idx = 0;
 831
 832         /* The padding bytes are at the first Nd positions in the first row. */
 833         for (c_idx = 0; in_idx < kw; in_idx += r_subblock, ++c_idx) {
 834                 if (P[c_idx] < nd) {
 835                         rte_memcpy(&out[out_idx], &in[in_idx + 1],
 836                                         r_subblock - 1);
 837                         out_idx += r_subblock - 1;
 838                 } else {
 839                         rte_memcpy(&out[out_idx], &in[in_idx], r_subblock);
 840                         out_idx += r_subblock;
 841                 }
 842         }
 843
 844         /* First and second parity bits sub-blocks are interlaced. */
 845         for (c_idx = 0; in_idx < ncb - 2 * r_subblock;
 846                         in_idx += 2 * r_subblock, ++c_idx) {
 847                 uint32_t second_block_c_idx = P[c_idx];
 848                 uint32_t third_block_c_idx = P[c_idx] + 1;
 849
 850                 if (second_block_c_idx < nd && third_block_c_idx < nd) {
 851                         rte_memcpy(&out[out_idx], &in[in_idx + 2],
 852                                         2 * r_subblock - 2);
 853                         out_idx += 2 * r_subblock - 2;
 854                 } else if (second_block_c_idx >= nd &&
 855                                 third_block_c_idx >= nd) {
 856                         rte_memcpy(&out[out_idx], &in[in_idx], 2 * r_subblock);
 857                         out_idx += 2 * r_subblock;
 858                 } else if (second_block_c_idx < nd) {
 859                         out[out_idx++] = in[in_idx];
 860                         rte_memcpy(&out[out_idx], &in[in_idx + 2],
 861                                         2 * r_subblock - 2);
 862                         out_idx += 2 * r_subblock - 2;
 863                 } else {
 864                         rte_memcpy(&out[out_idx], &in[in_idx + 1],
 865                                         2 * r_subblock - 1);
 866                         out_idx += 2 * r_subblock - 1;
 867                 }
 868         }
 869
 870         /* Last interlaced row is different - its last byte is the only padding
 871          * byte. We can have from 4 up to 28 padding bytes (Nd) per sub-block.
 872          * After interlacing the 1st and 2nd parity sub-blocks we can have 0, 1
 873          * or 2 padding bytes each time we make a step of 2 * R_SUBBLOCK bytes
 874          * (moving to another column). 2nd parity sub-block uses the same
 875          * inter-column permutation pattern as the systematic and 1st parity
 876          * sub-blocks but it adds '1' to the resulting index and calculates the
 877          * modulus of the result and Kw. Last column is mapped to itself (id 31)
 878          * so the first byte taken from the 2nd parity sub-block will be the
 879          * 32nd (31+1) byte, then 64th etc. (step is C_SUBBLOCK == 32) and the
 880          * last byte will be the first byte from the sub-block:
 881          * (32 + 32 * (R_SUBBLOCK-1)) % Kw == Kw % Kw == 0. Nd can't  be smaller
 882          * than 4 so we know that bytes with ids 0, 1, 2 and 3 must be the
 883          * padding bytes. The bytes from the 1st parity sub-block are the bytes
 884          * from the 31st column - Nd can't be greater than 28 so we are sure
 885          * that there are no padding bytes in 31st column.
 886          */
 887         rte_memcpy(&out[out_idx], &in[in_idx], 2 * r_subblock - 1);
 888 }
 889
 890 static inline void
 891 move_padding_bytes(const uint8_t *in, uint8_t *out, uint16_t k,
 892                 uint16_t ncb)
 893 {
 894         uint16_t d = k + 4;
 895         uint16_t kpi = ncb / 3;
 896         uint16_t nd = kpi - d;
 897
 898         rte_memcpy(&out[nd], in, d);
 899         rte_memcpy(&out[nd + kpi + 64], &in[kpi], d);
 900         rte_memcpy(&out[(nd - 1) + 2 * (kpi + 64)], &in[2 * kpi], d);
 901 }
 902
 903 static inline void
 904 process_dec_cb(struct turbo_sw_queue *q, struct rte_bbdev_dec_op *op,
 905                 uint8_t c, uint16_t k, uint16_t kw, struct rte_mbuf *m_in,
 906                 struct rte_mbuf *m_out, uint16_t in_offset, uint16_t out_offset,
 907                 bool check_crc_24b, uint16_t crc24_overlap, uint16_t total_left)
 908 {
 909         int ret;
 910         int32_t k_idx;
 911         int32_t iter_cnt;
 912         uint8_t *in, *out, *adapter_input;
 913         int32_t ncb, ncb_without_null;
 914         struct bblib_turbo_adapter_ul_response adapter_resp;
 915         struct bblib_turbo_adapter_ul_request adapter_req;
 916         struct bblib_turbo_decoder_request turbo_req;
 917         struct bblib_turbo_decoder_response turbo_resp;
 918         struct rte_bbdev_op_turbo_dec *dec = &op->turbo_dec;
 919
 920         k_idx = compute_idx(k);
 921
 922         ret = is_dec_input_valid(k_idx, kw, total_left);
 923         if (ret != 0) {
 924                 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 925                 return;
 926         }
 927
 928         in = rte_pktmbuf_mtod_offset(m_in, uint8_t *, in_offset);
 929         ncb = kw;
 930         ncb_without_null = (k + 4) * 3;
 931
 932         if (check_bit(dec->op_flags, RTE_BBDEV_TURBO_SUBBLOCK_DEINTERLEAVE)) {
 933                 struct bblib_deinterleave_ul_request deint_req;
 934                 struct bblib_deinterleave_ul_response deint_resp;
 935
 936                 /* SW decoder accepts only a circular buffer without NULL bytes
 937                  * so the input needs to be converted.
 938                  */
 939                 remove_nulls_from_circular_buf(in, q->deint_input, k, ncb);
 940
 941                 deint_req.pharqbuffer = q->deint_input;
 942                 deint_req.ncb = ncb_without_null;
 943                 deint_resp.pinteleavebuffer = q->deint_output;
 944                 bblib_deinterleave_ul(&deint_req, &deint_resp);
 945         } else
 946                 move_padding_bytes(in, q->deint_output, k, ncb);
 947
 948         adapter_input = q->deint_output;
 949
 950         if (dec->op_flags & RTE_BBDEV_TURBO_POS_LLR_1_BIT_IN)
 951                 adapter_req.isinverted = 1;
 952         else if (dec->op_flags & RTE_BBDEV_TURBO_NEG_LLR_1_BIT_IN)
 953                 adapter_req.isinverted = 0;
 954         else {
 955                 op->status |= 1 << RTE_BBDEV_DRV_ERROR;
 956                 rte_bbdev_log(ERR, "LLR format wasn't specified");
 957                 return;
 958         }
 959
 960         adapter_req.ncb = ncb_without_null;
 961         adapter_req.pinteleavebuffer = adapter_input;
 962         adapter_resp.pharqout = q->adapter_output;
 963         bblib_turbo_adapter_ul(&adapter_req, &adapter_resp);
 964
 965         out = (uint8_t *)rte_pktmbuf_append(m_out, ((k - crc24_overlap) >> 3));
 966         if (out == NULL) {
 967                 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 968                 rte_bbdev_log(ERR, "Too little space in output mbuf");
 969                 return;
 970         }
 971         /* rte_bbdev_op_data.offset can be different than the offset of the
 972          * appended bytes
 973          */
 974         out = rte_pktmbuf_mtod_offset(m_out, uint8_t *, out_offset);
 975         if (check_crc_24b)
 976                 turbo_req.c = c + 1;
 977         else
 978                 turbo_req.c = c;
 979         turbo_req.input = (int8_t *)q->adapter_output;
 980         turbo_req.k = k;
 981         turbo_req.k_idx = k_idx;
 982         turbo_req.max_iter_num = dec->iter_max;
 983         turbo_req.early_term_disable = !check_bit(dec->op_flags,
 984                         RTE_BBDEV_TURBO_EARLY_TERMINATION);
 985         turbo_resp.ag_buf = q->ag;
 986         turbo_resp.cb_buf = q->code_block;
 987         turbo_resp.output = out;
 988         iter_cnt = bblib_turbo_decoder(&turbo_req, &turbo_resp);
 989         dec->hard_output.length += (k >> 3);
 990
 991         if (iter_cnt > 0) {
 992                 /* Temporary solution for returned iter_count from SDK */
 993                 iter_cnt = (iter_cnt - 1) / 2;
 994                 dec->iter_count = RTE_MAX(iter_cnt, dec->iter_count);
 995         } else {
 996                 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
 997                 rte_bbdev_log(ERR, "Turbo Decoder failed");
 998                 return;
 999         }
1000 }
1001
1002 static inline void
1003 enqueue_dec_one_op(struct turbo_sw_queue *q, struct rte_bbdev_dec_op *op)
1004 {
1005         uint8_t c, r = 0;
1006         uint16_t kw, k = 0;
1007         uint16_t crc24_overlap = 0;
1008         struct rte_bbdev_op_turbo_dec *dec = &op->turbo_dec;
1009         struct rte_mbuf *m_in = dec->input.data;
1010         struct rte_mbuf *m_out = dec->hard_output.data;
1011         uint16_t in_offset = dec->input.offset;
1012         uint16_t total_left = dec->input.length;
1013         uint16_t out_offset = dec->hard_output.offset;
1014
1015         /* Clear op status */
1016         op->status = 0;
1017
1018         if (m_in == NULL || m_out == NULL) {
1019                 rte_bbdev_log(ERR, "Invalid mbuf pointer");
1020                 op->status = 1 << RTE_BBDEV_DATA_ERROR;
1021                 return;
1022         }
1023
1024         if (dec->code_block_mode == 0) { /* For Transport Block mode */
1025                 c = dec->tb_params.c;
1026         } else { /* For Code Block mode */
1027                 k = dec->cb_params.k;
1028                 c = 1;
1029         }
1030
1031         if ((c > 1) && !check_bit(dec->op_flags,
1032                 RTE_BBDEV_TURBO_DEC_TB_CRC_24B_KEEP))
1033                 crc24_overlap = 24;
1034
1035         while (total_left > 0) {
1036                 if (dec->code_block_mode == 0)
1037                         k = (r < dec->tb_params.c_neg) ?
1038                                 dec->tb_params.k_neg : dec->tb_params.k_pos;
1039
1040                 /* Calculates circular buffer size (Kw).
1041                  * According to 3gpp 36.212 section 5.1.4.2
1042                  *   Kw = 3 * Kpi,
1043                  * where:
1044                  *   Kpi = nCol * nRow
1045                  * where nCol is 32 and nRow can be calculated from:
1046                  *   D =< nCol * nRow
1047                  * where D is the size of each output from turbo encoder block
1048                  * (k + 4).
1049                  */
1050                 kw = RTE_ALIGN_CEIL(k + 4, RTE_BBDEV_C_SUBBLOCK) * 3;
1051
1052                 process_dec_cb(q, op, c, k, kw, m_in, m_out, in_offset,
1053                                 out_offset, check_bit(dec->op_flags,
1054                                 RTE_BBDEV_TURBO_CRC_TYPE_24B), crc24_overlap,
1055                                 total_left);
1056                 /* To keep CRC24 attached to end of Code block, use
1057                  * RTE_BBDEV_TURBO_DEC_TB_CRC_24B_KEEP flag as it
1058                  * removed by default once verified.
1059                  */
1060
1061                 /* Update total_left */
1062                 total_left -= kw;
1063                 /* Update offsets for next CBs (if exist) */
1064                 in_offset += kw;
1065                 out_offset += ((k - crc24_overlap) >> 3);
1066                 r++;
1067         }
1068         if (total_left != 0) {
1069                 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
1070                 rte_bbdev_log(ERR,
1071                                 "Mismatch between mbuf length and included Circular buffer sizes");
1072         }
1073 }
1074
1075 static inline uint16_t
1076 enqueue_dec_all_ops(struct turbo_sw_queue *q, struct rte_bbdev_dec_op **ops,
1077                 uint16_t nb_ops)
1078 {
1079         uint16_t i;
1080
1081         for (i = 0; i < nb_ops; ++i)
1082                 enqueue_dec_one_op(q, ops[i]);
1083
1084         return rte_ring_enqueue_burst(q->processed_pkts, (void **)ops, nb_ops,
1085                         NULL);
1086 }
1087
1088 /* Enqueue burst */
1089 static uint16_t
1090 enqueue_enc_ops(struct rte_bbdev_queue_data *q_data,
1091                 struct rte_bbdev_enc_op **ops, uint16_t nb_ops)
1092 {
1093         void *queue = q_data->queue_private;
1094         struct turbo_sw_queue *q = queue;
1095         uint16_t nb_enqueued = 0;
1096
1097         nb_enqueued = enqueue_enc_all_ops(q, ops, nb_ops, &q_data->queue_stats);
1098
1099         q_data->queue_stats.enqueue_err_count += nb_ops - nb_enqueued;
1100         q_data->queue_stats.enqueued_count += nb_enqueued;
1101
1102         return nb_enqueued;
1103 }
1104
1105 /* Enqueue burst */
1106 static uint16_t
1107 enqueue_dec_ops(struct rte_bbdev_queue_data *q_data,
1108                  struct rte_bbdev_dec_op **ops, uint16_t nb_ops)
1109 {
1110         void *queue = q_data->queue_private;
1111         struct turbo_sw_queue *q = queue;
1112         uint16_t nb_enqueued = 0;
1113
1114         nb_enqueued = enqueue_dec_all_ops(q, ops, nb_ops);
1115
1116         q_data->queue_stats.enqueue_err_count += nb_ops - nb_enqueued;
1117         q_data->queue_stats.enqueued_count += nb_enqueued;
1118
1119         return nb_enqueued;
1120 }
1121
1122 /* Dequeue decode burst */
1123 static uint16_t
1124 dequeue_dec_ops(struct rte_bbdev_queue_data *q_data,
1125                 struct rte_bbdev_dec_op **ops, uint16_t nb_ops)
1126 {
1127         struct turbo_sw_queue *q = q_data->queue_private;
1128         uint16_t nb_dequeued = rte_ring_dequeue_burst(q->processed_pkts,
1129                         (void **)ops, nb_ops, NULL);
1130         q_data->queue_stats.dequeued_count += nb_dequeued;
1131
1132         return nb_dequeued;
1133 }
1134
1135 /* Dequeue encode burst */
1136 static uint16_t
1137 dequeue_enc_ops(struct rte_bbdev_queue_data *q_data,
1138                 struct rte_bbdev_enc_op **ops, uint16_t nb_ops)
1139 {
1140         struct turbo_sw_queue *q = q_data->queue_private;
1141         uint16_t nb_dequeued = rte_ring_dequeue_burst(q->processed_pkts,
1142                         (void **)ops, nb_ops, NULL);
1143         q_data->queue_stats.dequeued_count += nb_dequeued;
1144
1145         return nb_dequeued;
1146 }
1147
1148 /* Parse 16bit integer from string argument */
1149 static inline int
1150 parse_u16_arg(const char *key, const char *value, void *extra_args)
1151 {
1152         uint16_t *u16 = extra_args;
1153         unsigned int long result;
1154
1155         if ((value == NULL) || (extra_args == NULL))
1156                 return -EINVAL;
1157         errno = 0;
1158         result = strtoul(value, NULL, 0);
1159         if ((result >= (1 << 16)) || (errno != 0)) {
1160                 rte_bbdev_log(ERR, "Invalid value %lu for %s", result, key);
1161                 return -ERANGE;
1162         }
1163         *u16 = (uint16_t)result;
1164         return 0;
1165 }
1166
1167 /* Parse parameters used to create device */
1168 static int
1169 parse_turbo_sw_params(struct turbo_sw_params *params, const char *input_args)
1170 {
1171         struct rte_kvargs *kvlist = NULL;
1172         int ret = 0;
1173
1174         if (params == NULL)
1175                 return -EINVAL;
1176         if (input_args) {
1177                 kvlist = rte_kvargs_parse(input_args, turbo_sw_valid_params);
1178                 if (kvlist == NULL)
1179                         return -EFAULT;
1180
1181                 ret = rte_kvargs_process(kvlist, turbo_sw_valid_params[0],
1182                                         &parse_u16_arg, &params->queues_num);
1183                 if (ret < 0)
1184                         goto exit;
1185
1186                 ret = rte_kvargs_process(kvlist, turbo_sw_valid_params[1],
1187                                         &parse_u16_arg, &params->socket_id);
1188                 if (ret < 0)
1189                         goto exit;
1190
1191                 if (params->socket_id >= RTE_MAX_NUMA_NODES) {
1192                         rte_bbdev_log(ERR, "Invalid socket, must be < %u",
1193                                         RTE_MAX_NUMA_NODES);
1194                         goto exit;
1195                 }
1196         }
1197
1198 exit:
1199         if (kvlist)
1200                 rte_kvargs_free(kvlist);
1201         return ret;
1202 }
1203
1204 /* Create device */
1205 static int
1206 turbo_sw_bbdev_create(struct rte_vdev_device *vdev,
1207                 struct turbo_sw_params *init_params)
1208 {
1209         struct rte_bbdev *bbdev;
1210         const char *name = rte_vdev_device_name(vdev);
1211
1212         bbdev = rte_bbdev_allocate(name);
1213         if (bbdev == NULL)
1214                 return -ENODEV;
1215
1216         bbdev->data->dev_private = rte_zmalloc_socket(name,
1217                         sizeof(struct bbdev_private), RTE_CACHE_LINE_SIZE,
1218                         init_params->socket_id);
1219         if (bbdev->data->dev_private == NULL) {
1220                 rte_bbdev_release(bbdev);
1221                 return -ENOMEM;
1222         }
1223
1224         bbdev->dev_ops = &pmd_ops;
1225         bbdev->device = &vdev->device;
1226         bbdev->data->socket_id = init_params->socket_id;
1227         bbdev->intr_handle = NULL;
1228
1229         /* register rx/tx burst functions for data path */
1230         bbdev->dequeue_enc_ops = dequeue_enc_ops;
1231         bbdev->dequeue_dec_ops = dequeue_dec_ops;
1232         bbdev->enqueue_enc_ops = enqueue_enc_ops;
1233         bbdev->enqueue_dec_ops = enqueue_dec_ops;
1234         ((struct bbdev_private *) bbdev->data->dev_private)->max_nb_queues =
1235                         init_params->queues_num;
1236
1237         return 0;
1238 }
1239
1240 /* Initialise device */
1241 static int
1242 turbo_sw_bbdev_probe(struct rte_vdev_device *vdev)
1243 {
1244         struct turbo_sw_params init_params = {
1245                 rte_socket_id(),
1246                 RTE_BBDEV_DEFAULT_MAX_NB_QUEUES
1247         };
1248         const char *name;
1249         const char *input_args;
1250
1251         if (vdev == NULL)
1252                 return -EINVAL;
1253
1254         name = rte_vdev_device_name(vdev);
1255         if (name == NULL)
1256                 return -EINVAL;
1257         input_args = rte_vdev_device_args(vdev);
1258         parse_turbo_sw_params(&init_params, input_args);
1259
1260         rte_bbdev_log_debug(
1261                         "Initialising %s on NUMA node %d with max queues: %d\n",
1262                         name, init_params.socket_id, init_params.queues_num);
1263
1264         return turbo_sw_bbdev_create(vdev, &init_params);
1265 }
1266
1267 /* Uninitialise device */
1268 static int
1269 turbo_sw_bbdev_remove(struct rte_vdev_device *vdev)
1270 {
1271         struct rte_bbdev *bbdev;
1272         const char *name;
1273
1274         if (vdev == NULL)
1275                 return -EINVAL;
1276
1277         name = rte_vdev_device_name(vdev);
1278         if (name == NULL)
1279                 return -EINVAL;
1280
1281         bbdev = rte_bbdev_get_named_dev(name);
1282         if (bbdev == NULL)
1283                 return -EINVAL;
1284
1285         rte_free(bbdev->data->dev_private);
1286
1287         return rte_bbdev_release(bbdev);
1288 }
1289
1290 static struct rte_vdev_driver bbdev_turbo_sw_pmd_drv = {
1291         .probe = turbo_sw_bbdev_probe,
1292         .remove = turbo_sw_bbdev_remove
1293 };
1294
1295 RTE_PMD_REGISTER_VDEV(DRIVER_NAME, bbdev_turbo_sw_pmd_drv);
1296 RTE_PMD_REGISTER_PARAM_STRING(DRIVER_NAME,
1297         TURBO_SW_MAX_NB_QUEUES_ARG"=<int> "
1298         TURBO_SW_SOCKET_ID_ARG"=<int>");
1299
1300 RTE_INIT(null_bbdev_init_log);
1301 static void
1302 null_bbdev_init_log(void)
1303 {
1304         bbdev_turbo_sw_logtype = rte_log_register("pmd.bb.turbo_sw");
1305         if (bbdev_turbo_sw_logtype >= 0)
1306                 rte_log_set_level(bbdev_turbo_sw_logtype, RTE_LOG_NOTICE);
1307 }