1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright(c) 2017 Intel Corporation 3 */ 4 5 #include <stdlib.h> 6 #include <string.h> 7 8 #include <rte_common.h> 9 #include <bus_vdev_driver.h> 10 #include <rte_malloc.h> 11 #include <rte_ring.h> 12 #include <rte_kvargs.h> 13 #include <rte_cycles.h> 14 #include <rte_errno.h> 15 16 #include <rte_bbdev.h> 17 #include <rte_bbdev_pmd.h> 18 19 #include <rte_hexdump.h> 20 #include <rte_log.h> 21 22 #ifdef RTE_BBDEV_SDK_AVX2 23 #include <ipp.h> 24 #include <ipps.h> 25 #include <phy_turbo.h> 26 #include <phy_crc.h> 27 #include <phy_rate_match.h> 28 #endif 29 #ifdef RTE_BBDEV_SDK_AVX512 30 #include <bit_reverse.h> 31 #include <phy_ldpc_encoder_5gnr.h> 32 #include <phy_ldpc_decoder_5gnr.h> 33 #include <phy_LDPC_ratematch_5gnr.h> 34 #include <phy_rate_dematching_5gnr.h> 35 #endif 36 37 #define DRIVER_NAME baseband_turbo_sw 38 39 RTE_LOG_REGISTER_DEFAULT(bbdev_turbo_sw_logtype, NOTICE); 40 41 /* Helper macro for logging */ 42 #define rte_bbdev_log(level, fmt, ...) \ 43 rte_log(RTE_LOG_ ## level, bbdev_turbo_sw_logtype, fmt "\n", \ 44 ##__VA_ARGS__) 45 46 #define rte_bbdev_log_debug(fmt, ...) \ 47 rte_bbdev_log(DEBUG, RTE_STR(__LINE__) ":%s() " fmt, __func__, \ 48 ##__VA_ARGS__) 49 50 #define DEINT_INPUT_BUF_SIZE (((RTE_BBDEV_TURBO_MAX_CB_SIZE >> 3) + 1) * 48) 51 #define DEINT_OUTPUT_BUF_SIZE (DEINT_INPUT_BUF_SIZE * 6) 52 #define ADAPTER_OUTPUT_BUF_SIZE ((RTE_BBDEV_TURBO_MAX_CB_SIZE + 4) * 48) 53 54 /* private data structure */ 55 struct bbdev_private { 56 unsigned int max_nb_queues; /**< Max number of queues */ 57 }; 58 59 /* Initialisation params structure that can be used by Turbo SW driver */ 60 struct turbo_sw_params { 61 int socket_id; /*< Turbo SW device socket */ 62 uint16_t queues_num; /*< Turbo SW device queues number */ 63 }; 64 65 /* Acceptable params for Turbo SW devices */ 66 #define TURBO_SW_MAX_NB_QUEUES_ARG "max_nb_queues" 67 #define TURBO_SW_SOCKET_ID_ARG "socket_id" 68 69 static const char * const turbo_sw_valid_params[] = { 70 TURBO_SW_MAX_NB_QUEUES_ARG, 71 TURBO_SW_SOCKET_ID_ARG 72 }; 73 74 /* queue */ 75 struct turbo_sw_queue { 76 /* Ring for processed (encoded/decoded) operations which are ready to 77 * be dequeued. 78 */ 79 struct rte_ring *processed_pkts; 80 /* Stores input for turbo encoder (used when CRC attachment is 81 * performed 82 */ 83 uint8_t *enc_in; 84 /* Stores output from turbo encoder */ 85 uint8_t *enc_out; 86 /* Alpha gamma buf for bblib_turbo_decoder() function */ 87 int8_t *ag; 88 /* Temp buf for bblib_turbo_decoder() function */ 89 uint16_t *code_block; 90 /* Input buf for bblib_rate_dematching_lte() function */ 91 uint8_t *deint_input; 92 /* Output buf for bblib_rate_dematching_lte() function */ 93 uint8_t *deint_output; 94 /* Output buf for bblib_turbodec_adapter_lte() function */ 95 uint8_t *adapter_output; 96 /* Operation type of this queue */ 97 enum rte_bbdev_op_type type; 98 } __rte_cache_aligned; 99 100 101 #ifdef RTE_BBDEV_SDK_AVX2 102 static inline char * 103 mbuf_append(struct rte_mbuf *m_head, struct rte_mbuf *m, uint16_t len) 104 { 105 if (unlikely(len > rte_pktmbuf_tailroom(m))) 106 return NULL; 107 108 char *tail = (char *)m->buf_addr + m->data_off + m->data_len; 109 m->data_len = (uint16_t)(m->data_len + len); 110 m_head->pkt_len = (m_head->pkt_len + len); 111 return tail; 112 } 113 114 /* Calculate index based on Table 5.1.3-3 from TS34.212 */ 115 static inline int32_t 116 compute_idx(uint16_t k) 117 { 118 int32_t result = 0; 119 120 if (k < RTE_BBDEV_TURBO_MIN_CB_SIZE || k > RTE_BBDEV_TURBO_MAX_CB_SIZE) 121 return -1; 122 123 if (k > 2048) { 124 if ((k - 2048) % 64 != 0) 125 result = -1; 126 127 result = 124 + (k - 2048) / 64; 128 } else if (k <= 512) { 129 if ((k - 40) % 8 != 0) 130 result = -1; 131 132 result = (k - 40) / 8 + 1; 133 } else if (k <= 1024) { 134 if ((k - 512) % 16 != 0) 135 result = -1; 136 137 result = 60 + (k - 512) / 16; 138 } else { /* 1024 < k <= 2048 */ 139 if ((k - 1024) % 32 != 0) 140 result = -1; 141 142 result = 92 + (k - 1024) / 32; 143 } 144 145 return result; 146 } 147 #endif 148 149 /* Read flag value 0/1 from bitmap */ 150 static inline bool 151 check_bit(uint32_t bitmap, uint32_t bitmask) 152 { 153 return bitmap & bitmask; 154 } 155 156 /* Get device info */ 157 static void 158 info_get(struct rte_bbdev *dev, struct rte_bbdev_driver_info *dev_info) 159 { 160 struct bbdev_private *internals = dev->data->dev_private; 161 162 static const struct rte_bbdev_op_cap bbdev_capabilities[] = { 163 #ifdef RTE_BBDEV_SDK_AVX2 164 { 165 .type = RTE_BBDEV_OP_TURBO_DEC, 166 .cap.turbo_dec = { 167 .capability_flags = 168 RTE_BBDEV_TURBO_SUBBLOCK_DEINTERLEAVE | 169 RTE_BBDEV_TURBO_POS_LLR_1_BIT_IN | 170 RTE_BBDEV_TURBO_NEG_LLR_1_BIT_IN | 171 RTE_BBDEV_TURBO_CRC_TYPE_24B | 172 RTE_BBDEV_TURBO_DEC_TB_CRC_24B_KEEP | 173 RTE_BBDEV_TURBO_EARLY_TERMINATION, 174 .max_llr_modulus = 16, 175 .num_buffers_src = 176 RTE_BBDEV_TURBO_MAX_CODE_BLOCKS, 177 .num_buffers_hard_out = 178 RTE_BBDEV_TURBO_MAX_CODE_BLOCKS, 179 .num_buffers_soft_out = 0, 180 } 181 }, 182 { 183 .type = RTE_BBDEV_OP_TURBO_ENC, 184 .cap.turbo_enc = { 185 .capability_flags = 186 RTE_BBDEV_TURBO_CRC_24B_ATTACH | 187 RTE_BBDEV_TURBO_CRC_24A_ATTACH | 188 RTE_BBDEV_TURBO_RATE_MATCH | 189 RTE_BBDEV_TURBO_RV_INDEX_BYPASS, 190 .num_buffers_src = 191 RTE_BBDEV_TURBO_MAX_CODE_BLOCKS, 192 .num_buffers_dst = 193 RTE_BBDEV_TURBO_MAX_CODE_BLOCKS, 194 } 195 }, 196 #endif 197 #ifdef RTE_BBDEV_SDK_AVX512 198 { 199 .type = RTE_BBDEV_OP_LDPC_ENC, 200 .cap.ldpc_enc = { 201 .capability_flags = 202 RTE_BBDEV_LDPC_RATE_MATCH | 203 RTE_BBDEV_LDPC_CRC_16_ATTACH | 204 RTE_BBDEV_LDPC_CRC_24A_ATTACH | 205 RTE_BBDEV_LDPC_CRC_24B_ATTACH, 206 .num_buffers_src = 207 RTE_BBDEV_LDPC_MAX_CODE_BLOCKS, 208 .num_buffers_dst = 209 RTE_BBDEV_LDPC_MAX_CODE_BLOCKS, 210 } 211 }, 212 { 213 .type = RTE_BBDEV_OP_LDPC_DEC, 214 .cap.ldpc_dec = { 215 .capability_flags = 216 RTE_BBDEV_LDPC_CRC_TYPE_16_CHECK | 217 RTE_BBDEV_LDPC_CRC_TYPE_24B_CHECK | 218 RTE_BBDEV_LDPC_CRC_TYPE_24A_CHECK | 219 RTE_BBDEV_LDPC_CRC_TYPE_24B_DROP | 220 RTE_BBDEV_LDPC_HQ_COMBINE_IN_ENABLE | 221 RTE_BBDEV_LDPC_HQ_COMBINE_OUT_ENABLE | 222 RTE_BBDEV_LDPC_ITERATION_STOP_ENABLE, 223 .llr_size = 8, 224 .llr_decimals = 4, 225 .num_buffers_src = 226 RTE_BBDEV_LDPC_MAX_CODE_BLOCKS, 227 .num_buffers_hard_out = 228 RTE_BBDEV_LDPC_MAX_CODE_BLOCKS, 229 .num_buffers_soft_out = 0, 230 } 231 }, 232 #endif 233 RTE_BBDEV_END_OF_CAPABILITIES_LIST() 234 }; 235 236 static struct rte_bbdev_queue_conf default_queue_conf = { 237 .queue_size = RTE_BBDEV_QUEUE_SIZE_LIMIT, 238 }; 239 #ifdef RTE_BBDEV_SDK_AVX2 240 static const enum rte_cpu_flag_t cpu_flag = RTE_CPUFLAG_SSE4_2; 241 dev_info->cpu_flag_reqs = &cpu_flag; 242 #else 243 dev_info->cpu_flag_reqs = NULL; 244 #endif 245 default_queue_conf.socket = dev->data->socket_id; 246 247 dev_info->driver_name = RTE_STR(DRIVER_NAME); 248 dev_info->max_num_queues = internals->max_nb_queues; 249 dev_info->queue_size_lim = RTE_BBDEV_QUEUE_SIZE_LIMIT; 250 dev_info->hardware_accelerated = false; 251 dev_info->max_dl_queue_priority = 0; 252 dev_info->max_ul_queue_priority = 0; 253 dev_info->default_queue_conf = default_queue_conf; 254 dev_info->capabilities = bbdev_capabilities; 255 dev_info->min_alignment = 64; 256 dev_info->harq_buffer_size = 0; 257 dev_info->data_endianness = RTE_LITTLE_ENDIAN; 258 dev_info->device_status = RTE_BBDEV_DEV_NOT_SUPPORTED; 259 260 rte_bbdev_log_debug("got device info from %u\n", dev->data->dev_id); 261 } 262 263 /* Release queue */ 264 static int 265 q_release(struct rte_bbdev *dev, uint16_t q_id) 266 { 267 struct turbo_sw_queue *q = dev->data->queues[q_id].queue_private; 268 269 if (q != NULL) { 270 rte_ring_free(q->processed_pkts); 271 rte_free(q->enc_out); 272 rte_free(q->enc_in); 273 rte_free(q->ag); 274 rte_free(q->code_block); 275 rte_free(q->deint_input); 276 rte_free(q->deint_output); 277 rte_free(q->adapter_output); 278 rte_free(q); 279 dev->data->queues[q_id].queue_private = NULL; 280 } 281 282 rte_bbdev_log_debug("released device queue %u:%u", 283 dev->data->dev_id, q_id); 284 return 0; 285 } 286 287 /* Setup a queue */ 288 static int 289 q_setup(struct rte_bbdev *dev, uint16_t q_id, 290 const struct rte_bbdev_queue_conf *queue_conf) 291 { 292 int ret; 293 struct turbo_sw_queue *q; 294 char name[RTE_RING_NAMESIZE]; 295 296 /* Allocate the queue data structure. */ 297 q = rte_zmalloc_socket(RTE_STR(DRIVER_NAME), sizeof(*q), 298 RTE_CACHE_LINE_SIZE, queue_conf->socket); 299 if (q == NULL) { 300 rte_bbdev_log(ERR, "Failed to allocate queue memory"); 301 return -ENOMEM; 302 } 303 304 /* Allocate memory for encoder output. */ 305 ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"_enc_o%u:%u", 306 dev->data->dev_id, q_id); 307 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) { 308 rte_bbdev_log(ERR, 309 "Creating queue name for device %u queue %u failed", 310 dev->data->dev_id, q_id); 311 ret = -ENAMETOOLONG; 312 goto free_q; 313 } 314 q->enc_out = rte_zmalloc_socket(name, 315 ((RTE_BBDEV_TURBO_MAX_TB_SIZE >> 3) + 3) * 316 sizeof(*q->enc_out) * 3, 317 RTE_CACHE_LINE_SIZE, queue_conf->socket); 318 if (q->enc_out == NULL) { 319 rte_bbdev_log(ERR, 320 "Failed to allocate queue memory for %s", name); 321 ret = -ENOMEM; 322 goto free_q; 323 } 324 325 /* Allocate memory for rate matching output. */ 326 ret = snprintf(name, RTE_RING_NAMESIZE, 327 RTE_STR(DRIVER_NAME)"_enc_i%u:%u", dev->data->dev_id, 328 q_id); 329 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) { 330 rte_bbdev_log(ERR, 331 "Creating queue name for device %u queue %u failed", 332 dev->data->dev_id, q_id); 333 ret = -ENAMETOOLONG; 334 goto free_q; 335 } 336 q->enc_in = rte_zmalloc_socket(name, 337 (RTE_BBDEV_LDPC_MAX_CB_SIZE >> 3) * sizeof(*q->enc_in), 338 RTE_CACHE_LINE_SIZE, queue_conf->socket); 339 if (q->enc_in == NULL) { 340 rte_bbdev_log(ERR, 341 "Failed to allocate queue memory for %s", name); 342 ret = -ENOMEM; 343 goto free_q; 344 } 345 346 /* Allocate memory for Alpha Gamma temp buffer. */ 347 ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"_ag%u:%u", 348 dev->data->dev_id, q_id); 349 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) { 350 rte_bbdev_log(ERR, 351 "Creating queue name for device %u queue %u failed", 352 dev->data->dev_id, q_id); 353 ret = -ENAMETOOLONG; 354 goto free_q; 355 } 356 q->ag = rte_zmalloc_socket(name, 357 RTE_BBDEV_TURBO_MAX_CB_SIZE * 10 * sizeof(*q->ag), 358 RTE_CACHE_LINE_SIZE, queue_conf->socket); 359 if (q->ag == NULL) { 360 rte_bbdev_log(ERR, 361 "Failed to allocate queue memory for %s", name); 362 ret = -ENOMEM; 363 goto free_q; 364 } 365 366 /* Allocate memory for code block temp buffer. */ 367 ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"_cb%u:%u", 368 dev->data->dev_id, q_id); 369 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) { 370 rte_bbdev_log(ERR, 371 "Creating queue name for device %u queue %u failed", 372 dev->data->dev_id, q_id); 373 ret = -ENAMETOOLONG; 374 goto free_q; 375 } 376 q->code_block = rte_zmalloc_socket(name, 377 RTE_BBDEV_TURBO_MAX_CB_SIZE * sizeof(*q->code_block), 378 RTE_CACHE_LINE_SIZE, queue_conf->socket); 379 if (q->code_block == NULL) { 380 rte_bbdev_log(ERR, 381 "Failed to allocate queue memory for %s", name); 382 ret = -ENOMEM; 383 goto free_q; 384 } 385 386 /* Allocate memory for Deinterleaver input. */ 387 ret = snprintf(name, RTE_RING_NAMESIZE, 388 RTE_STR(DRIVER_NAME)"_de_i%u:%u", 389 dev->data->dev_id, q_id); 390 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) { 391 rte_bbdev_log(ERR, 392 "Creating queue name for device %u queue %u failed", 393 dev->data->dev_id, q_id); 394 ret = -ENAMETOOLONG; 395 goto free_q; 396 } 397 q->deint_input = rte_zmalloc_socket(name, 398 DEINT_INPUT_BUF_SIZE * sizeof(*q->deint_input), 399 RTE_CACHE_LINE_SIZE, queue_conf->socket); 400 if (q->deint_input == NULL) { 401 rte_bbdev_log(ERR, 402 "Failed to allocate queue memory for %s", name); 403 ret = -ENOMEM; 404 goto free_q; 405 } 406 407 /* Allocate memory for Deinterleaver output. */ 408 ret = snprintf(name, RTE_RING_NAMESIZE, 409 RTE_STR(DRIVER_NAME)"_de_o%u:%u", 410 dev->data->dev_id, q_id); 411 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) { 412 rte_bbdev_log(ERR, 413 "Creating queue name for device %u queue %u failed", 414 dev->data->dev_id, q_id); 415 ret = -ENAMETOOLONG; 416 goto free_q; 417 } 418 q->deint_output = rte_zmalloc_socket(NULL, 419 DEINT_OUTPUT_BUF_SIZE * sizeof(*q->deint_output), 420 RTE_CACHE_LINE_SIZE, queue_conf->socket); 421 if (q->deint_output == NULL) { 422 rte_bbdev_log(ERR, 423 "Failed to allocate queue memory for %s", name); 424 ret = -ENOMEM; 425 goto free_q; 426 } 427 428 /* Allocate memory for Adapter output. */ 429 ret = snprintf(name, RTE_RING_NAMESIZE, 430 RTE_STR(DRIVER_NAME)"_ada_o%u:%u", 431 dev->data->dev_id, q_id); 432 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) { 433 rte_bbdev_log(ERR, 434 "Creating queue name for device %u queue %u failed", 435 dev->data->dev_id, q_id); 436 ret = -ENAMETOOLONG; 437 goto free_q; 438 } 439 q->adapter_output = rte_zmalloc_socket(NULL, 440 ADAPTER_OUTPUT_BUF_SIZE * sizeof(*q->adapter_output), 441 RTE_CACHE_LINE_SIZE, queue_conf->socket); 442 if (q->adapter_output == NULL) { 443 rte_bbdev_log(ERR, 444 "Failed to allocate queue memory for %s", name); 445 ret = -ENOMEM; 446 goto free_q; 447 } 448 449 /* Create ring for packets awaiting to be dequeued. */ 450 ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"%u:%u", 451 dev->data->dev_id, q_id); 452 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) { 453 rte_bbdev_log(ERR, 454 "Creating queue name for device %u queue %u failed", 455 dev->data->dev_id, q_id); 456 ret = -ENAMETOOLONG; 457 goto free_q; 458 } 459 q->processed_pkts = rte_ring_create(name, queue_conf->queue_size, 460 queue_conf->socket, RING_F_SP_ENQ | RING_F_SC_DEQ); 461 if (q->processed_pkts == NULL) { 462 rte_bbdev_log(ERR, "Failed to create ring for %s", name); 463 ret = -rte_errno; 464 goto free_q; 465 } 466 467 q->type = queue_conf->op_type; 468 469 dev->data->queues[q_id].queue_private = q; 470 rte_bbdev_log_debug("setup device queue %s", name); 471 return 0; 472 473 free_q: 474 rte_ring_free(q->processed_pkts); 475 rte_free(q->enc_out); 476 rte_free(q->enc_in); 477 rte_free(q->ag); 478 rte_free(q->code_block); 479 rte_free(q->deint_input); 480 rte_free(q->deint_output); 481 rte_free(q->adapter_output); 482 rte_free(q); 483 return ret; 484 } 485 486 static const struct rte_bbdev_ops pmd_ops = { 487 .info_get = info_get, 488 .queue_setup = q_setup, 489 .queue_release = q_release 490 }; 491 492 #ifdef RTE_BBDEV_SDK_AVX2 493 #ifdef RTE_LIBRTE_BBDEV_DEBUG 494 /* Checks if the encoder input buffer is correct. 495 * Returns 0 if it's valid, -1 otherwise. 496 */ 497 static inline int 498 is_enc_input_valid(const uint16_t k, const int32_t k_idx, 499 const uint16_t in_length) 500 { 501 if (k_idx < 0) { 502 rte_bbdev_log(ERR, "K Index is invalid"); 503 return -1; 504 } 505 506 if (in_length - (k >> 3) < 0) { 507 rte_bbdev_log(ERR, 508 "Mismatch between input length (%u bytes) and K (%u bits)", 509 in_length, k); 510 return -1; 511 } 512 513 if (k > RTE_BBDEV_TURBO_MAX_CB_SIZE) { 514 rte_bbdev_log(ERR, "CB size (%u) is too big, max: %d", 515 k, RTE_BBDEV_TURBO_MAX_CB_SIZE); 516 return -1; 517 } 518 519 return 0; 520 } 521 522 /* Checks if the decoder input buffer is correct. 523 * Returns 0 if it's valid, -1 otherwise. 524 */ 525 static inline int 526 is_dec_input_valid(int32_t k_idx, int16_t kw, int16_t in_length) 527 { 528 if (k_idx < 0) { 529 rte_bbdev_log(ERR, "K index is invalid"); 530 return -1; 531 } 532 533 if (in_length < kw) { 534 rte_bbdev_log(ERR, 535 "Mismatch between input length (%u) and kw (%u)", 536 in_length, kw); 537 return -1; 538 } 539 540 if (kw > RTE_BBDEV_TURBO_MAX_KW) { 541 rte_bbdev_log(ERR, "Input length (%u) is too big, max: %d", 542 kw, RTE_BBDEV_TURBO_MAX_KW); 543 return -1; 544 } 545 546 return 0; 547 } 548 #endif 549 #endif 550 551 static inline void 552 process_enc_cb(struct turbo_sw_queue *q, struct rte_bbdev_enc_op *op, 553 uint8_t r, uint8_t c, uint16_t k, uint16_t ncb, 554 uint32_t e, struct rte_mbuf *m_in, struct rte_mbuf *m_out_head, 555 struct rte_mbuf *m_out, uint16_t in_offset, uint16_t out_offset, 556 uint16_t in_length, struct rte_bbdev_stats *q_stats) 557 { 558 #ifdef RTE_BBDEV_SDK_AVX2 559 #ifdef RTE_LIBRTE_BBDEV_DEBUG 560 int ret; 561 #else 562 RTE_SET_USED(in_length); 563 #endif 564 int16_t k_idx; 565 uint16_t m; 566 uint8_t *in, *out0, *out1, *out2, *tmp_out, *rm_out; 567 uint64_t first_3_bytes = 0; 568 struct rte_bbdev_op_turbo_enc *enc = &op->turbo_enc; 569 struct bblib_crc_request crc_req; 570 struct bblib_crc_response crc_resp; 571 struct bblib_turbo_encoder_request turbo_req; 572 struct bblib_turbo_encoder_response turbo_resp; 573 struct bblib_rate_match_dl_request rm_req; 574 struct bblib_rate_match_dl_response rm_resp; 575 #ifdef RTE_BBDEV_OFFLOAD_COST 576 uint64_t start_time; 577 #else 578 RTE_SET_USED(q_stats); 579 #endif 580 581 k_idx = compute_idx(k); 582 in = rte_pktmbuf_mtod_offset(m_in, uint8_t *, in_offset); 583 584 /* CRC24A (for TB) */ 585 if ((enc->op_flags & RTE_BBDEV_TURBO_CRC_24A_ATTACH) && 586 (enc->code_block_mode == RTE_BBDEV_CODE_BLOCK)) { 587 #ifdef RTE_LIBRTE_BBDEV_DEBUG 588 ret = is_enc_input_valid(k - 24, k_idx, in_length); 589 if (ret != 0) { 590 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 591 return; 592 } 593 #endif 594 595 crc_req.data = in; 596 crc_req.len = k - 24; 597 /* Check if there is a room for CRC bits if not use 598 * the temporary buffer. 599 */ 600 if (mbuf_append(m_in, m_in, 3) == NULL) { 601 rte_memcpy(q->enc_in, in, (k - 24) >> 3); 602 in = q->enc_in; 603 } else { 604 /* Store 3 first bytes of next CB as they will be 605 * overwritten by CRC bytes. If it is the last CB then 606 * there is no point to store 3 next bytes and this 607 * if..else branch will be omitted. 608 */ 609 first_3_bytes = *((uint64_t *)&in[(k - 32) >> 3]); 610 } 611 612 crc_resp.data = in; 613 #ifdef RTE_BBDEV_OFFLOAD_COST 614 start_time = rte_rdtsc_precise(); 615 #endif 616 /* CRC24A generation */ 617 bblib_lte_crc24a_gen(&crc_req, &crc_resp); 618 #ifdef RTE_BBDEV_OFFLOAD_COST 619 q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time; 620 #endif 621 } else if (enc->op_flags & RTE_BBDEV_TURBO_CRC_24B_ATTACH) { 622 /* CRC24B */ 623 #ifdef RTE_LIBRTE_BBDEV_DEBUG 624 ret = is_enc_input_valid(k - 24, k_idx, in_length); 625 if (ret != 0) { 626 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 627 return; 628 } 629 #endif 630 631 crc_req.data = in; 632 crc_req.len = k - 24; 633 /* Check if there is a room for CRC bits if this is the last 634 * CB in TB. If not use temporary buffer. 635 */ 636 if ((c - r == 1) && (mbuf_append(m_in, m_in, 3) == NULL)) { 637 rte_memcpy(q->enc_in, in, (k - 24) >> 3); 638 in = q->enc_in; 639 } else if (c - r > 1) { 640 /* Store 3 first bytes of next CB as they will be 641 * overwritten by CRC bytes. If it is the last CB then 642 * there is no point to store 3 next bytes and this 643 * if..else branch will be omitted. 644 */ 645 first_3_bytes = *((uint64_t *)&in[(k - 32) >> 3]); 646 } 647 648 crc_resp.data = in; 649 #ifdef RTE_BBDEV_OFFLOAD_COST 650 start_time = rte_rdtsc_precise(); 651 #endif 652 /* CRC24B generation */ 653 bblib_lte_crc24b_gen(&crc_req, &crc_resp); 654 #ifdef RTE_BBDEV_OFFLOAD_COST 655 q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time; 656 #endif 657 } 658 #ifdef RTE_LIBRTE_BBDEV_DEBUG 659 else { 660 ret = is_enc_input_valid(k, k_idx, in_length); 661 if (ret != 0) { 662 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 663 return; 664 } 665 } 666 #endif 667 668 /* Turbo encoder */ 669 670 /* Each bit layer output from turbo encoder is (k+4) bits long, i.e. 671 * input length + 4 tail bits. That's (k/8) + 1 bytes after rounding up. 672 * So dst_data's length should be 3*(k/8) + 3 bytes. 673 * In Rate-matching bypass case outputs pointers passed to encoder 674 * (out0, out1 and out2) can directly point to addresses of output from 675 * turbo_enc entity. 676 */ 677 if (enc->op_flags & RTE_BBDEV_TURBO_RATE_MATCH) { 678 out0 = q->enc_out; 679 out1 = RTE_PTR_ADD(out0, (k >> 3) + 1); 680 out2 = RTE_PTR_ADD(out1, (k >> 3) + 1); 681 } else { 682 out0 = (uint8_t *)mbuf_append(m_out_head, m_out, 683 (k >> 3) * 3 + 2); 684 if (out0 == NULL) { 685 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 686 rte_bbdev_log(ERR, 687 "Too little space in output mbuf"); 688 return; 689 } 690 enc->output.length += (k >> 3) * 3 + 2; 691 /* rte_bbdev_op_data.offset can be different than the 692 * offset of the appended bytes 693 */ 694 out0 = rte_pktmbuf_mtod_offset(m_out, uint8_t *, out_offset); 695 out1 = rte_pktmbuf_mtod_offset(m_out, uint8_t *, 696 out_offset + (k >> 3) + 1); 697 out2 = rte_pktmbuf_mtod_offset(m_out, uint8_t *, 698 out_offset + 2 * ((k >> 3) + 1)); 699 } 700 701 turbo_req.case_id = k_idx; 702 turbo_req.input_win = in; 703 turbo_req.length = k >> 3; 704 turbo_resp.output_win_0 = out0; 705 turbo_resp.output_win_1 = out1; 706 turbo_resp.output_win_2 = out2; 707 708 #ifdef RTE_BBDEV_OFFLOAD_COST 709 start_time = rte_rdtsc_precise(); 710 #endif 711 /* Turbo encoding */ 712 if (bblib_turbo_encoder(&turbo_req, &turbo_resp) != 0) { 713 op->status |= 1 << RTE_BBDEV_DRV_ERROR; 714 rte_bbdev_log(ERR, "Turbo Encoder failed"); 715 return; 716 } 717 #ifdef RTE_BBDEV_OFFLOAD_COST 718 q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time; 719 #endif 720 721 /* Restore 3 first bytes of next CB if they were overwritten by CRC*/ 722 if (first_3_bytes != 0) 723 *((uint64_t *)&in[(k - 32) >> 3]) = first_3_bytes; 724 725 /* Rate-matching */ 726 if (enc->op_flags & RTE_BBDEV_TURBO_RATE_MATCH) { 727 uint8_t mask_id; 728 /* Integer round up division by 8 */ 729 uint16_t out_len = (e + 7) >> 3; 730 /* The mask array is indexed using E%8. E is an even number so 731 * there are only 4 possible values. 732 */ 733 const uint8_t mask_out[] = {0xFF, 0xC0, 0xF0, 0xFC}; 734 735 /* get output data starting address */ 736 rm_out = (uint8_t *)mbuf_append(m_out_head, m_out, out_len); 737 if (rm_out == NULL) { 738 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 739 rte_bbdev_log(ERR, 740 "Too little space in output mbuf"); 741 return; 742 } 743 /* rte_bbdev_op_data.offset can be different than the offset 744 * of the appended bytes 745 */ 746 rm_out = rte_pktmbuf_mtod_offset(m_out, uint8_t *, out_offset); 747 748 /* index of current code block */ 749 rm_req.r = r; 750 /* total number of code block */ 751 rm_req.C = c; 752 /* For DL - 1, UL - 0 */ 753 rm_req.direction = 1; 754 /* According to 3ggp 36.212 Spec 5.1.4.1.2 section Nsoft, KMIMO 755 * and MDL_HARQ are used for Ncb calculation. As Ncb is already 756 * known we can adjust those parameters 757 */ 758 rm_req.Nsoft = ncb * rm_req.C; 759 rm_req.KMIMO = 1; 760 rm_req.MDL_HARQ = 1; 761 /* According to 3ggp 36.212 Spec 5.1.4.1.2 section Nl, Qm and G 762 * are used for E calculation. As E is already known we can 763 * adjust those parameters 764 */ 765 rm_req.NL = e; 766 rm_req.Qm = 1; 767 rm_req.G = rm_req.NL * rm_req.Qm * rm_req.C; 768 769 rm_req.rvidx = enc->rv_index; 770 rm_req.Kidx = k_idx - 1; 771 rm_req.nLen = k + 4; 772 rm_req.tin0 = out0; 773 rm_req.tin1 = out1; 774 rm_req.tin2 = out2; 775 rm_resp.output = rm_out; 776 rm_resp.OutputLen = out_len; 777 if (enc->op_flags & RTE_BBDEV_TURBO_RV_INDEX_BYPASS) 778 rm_req.bypass_rvidx = 1; 779 else 780 rm_req.bypass_rvidx = 0; 781 782 #ifdef RTE_BBDEV_OFFLOAD_COST 783 start_time = rte_rdtsc_precise(); 784 #endif 785 /* Rate-Matching */ 786 if (bblib_rate_match_dl(&rm_req, &rm_resp) != 0) { 787 op->status |= 1 << RTE_BBDEV_DRV_ERROR; 788 rte_bbdev_log(ERR, "Rate matching failed"); 789 return; 790 } 791 #ifdef RTE_BBDEV_OFFLOAD_COST 792 q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time; 793 #endif 794 795 /* SW fills an entire last byte even if E%8 != 0. Clear the 796 * superfluous data bits for consistency with HW device. 797 */ 798 mask_id = (e & 7) >> 1; 799 rm_out[out_len - 1] &= mask_out[mask_id]; 800 enc->output.length += rm_resp.OutputLen; 801 } else { 802 /* Rate matching is bypassed */ 803 804 /* Completing last byte of out0 (where 4 tail bits are stored) 805 * by moving first 4 bits from out1 806 */ 807 tmp_out = (uint8_t *) --out1; 808 *tmp_out = *tmp_out | ((*(tmp_out + 1) & 0xF0) >> 4); 809 tmp_out++; 810 /* Shifting out1 data by 4 bits to the left */ 811 for (m = 0; m < k >> 3; ++m) { 812 uint8_t *first = tmp_out; 813 uint8_t second = *(tmp_out + 1); 814 *first = (*first << 4) | ((second & 0xF0) >> 4); 815 tmp_out++; 816 } 817 /* Shifting out2 data by 8 bits to the left */ 818 for (m = 0; m < (k >> 3) + 1; ++m) { 819 *tmp_out = *(tmp_out + 1); 820 tmp_out++; 821 } 822 *tmp_out = 0; 823 } 824 #else 825 RTE_SET_USED(q); 826 RTE_SET_USED(op); 827 RTE_SET_USED(r); 828 RTE_SET_USED(c); 829 RTE_SET_USED(k); 830 RTE_SET_USED(ncb); 831 RTE_SET_USED(e); 832 RTE_SET_USED(m_in); 833 RTE_SET_USED(m_out_head); 834 RTE_SET_USED(m_out); 835 RTE_SET_USED(in_offset); 836 RTE_SET_USED(out_offset); 837 RTE_SET_USED(in_length); 838 RTE_SET_USED(q_stats); 839 #endif 840 } 841 842 843 static inline void 844 process_ldpc_enc_cb(struct turbo_sw_queue *q, struct rte_bbdev_enc_op *op, 845 uint32_t e, struct rte_mbuf *m_in, struct rte_mbuf *m_out_head, 846 struct rte_mbuf *m_out, uint16_t in_offset, uint16_t out_offset, 847 uint16_t seg_total_left, struct rte_bbdev_stats *q_stats) 848 { 849 #ifdef RTE_BBDEV_SDK_AVX512 850 RTE_SET_USED(seg_total_left); 851 uint8_t *in, *rm_out; 852 struct rte_bbdev_op_ldpc_enc *enc = &op->ldpc_enc; 853 struct bblib_ldpc_encoder_5gnr_request ldpc_req; 854 struct bblib_ldpc_encoder_5gnr_response ldpc_resp; 855 struct bblib_LDPC_ratematch_5gnr_request rm_req; 856 struct bblib_LDPC_ratematch_5gnr_response rm_resp; 857 struct bblib_crc_request crc_req; 858 struct bblib_crc_response crc_resp; 859 uint16_t msgLen, puntBits, parity_offset, out_len; 860 uint16_t K = (enc->basegraph == 1 ? 22 : 10) * enc->z_c; 861 uint16_t in_length_in_bits = K - enc->n_filler; 862 uint16_t in_length_in_bytes = (in_length_in_bits + 7) >> 3; 863 864 #ifdef RTE_BBDEV_OFFLOAD_COST 865 uint64_t start_time = rte_rdtsc_precise(); 866 #else 867 RTE_SET_USED(q_stats); 868 #endif 869 870 in = rte_pktmbuf_mtod_offset(m_in, uint8_t *, in_offset); 871 872 /* Masking the Filler bits explicitly */ 873 memset(q->enc_in + (in_length_in_bytes - 3), 0, 874 ((K + 7) >> 3) - (in_length_in_bytes - 3)); 875 /* CRC Generation */ 876 if (enc->op_flags & RTE_BBDEV_LDPC_CRC_24A_ATTACH) { 877 rte_memcpy(q->enc_in, in, in_length_in_bytes - 3); 878 crc_req.data = in; 879 crc_req.len = in_length_in_bits - 24; 880 crc_resp.data = q->enc_in; 881 bblib_lte_crc24a_gen(&crc_req, &crc_resp); 882 } else if (enc->op_flags & RTE_BBDEV_LDPC_CRC_24B_ATTACH) { 883 rte_memcpy(q->enc_in, in, in_length_in_bytes - 3); 884 crc_req.data = in; 885 crc_req.len = in_length_in_bits - 24; 886 crc_resp.data = q->enc_in; 887 bblib_lte_crc24b_gen(&crc_req, &crc_resp); 888 } else if (enc->op_flags & RTE_BBDEV_LDPC_CRC_16_ATTACH) { 889 rte_memcpy(q->enc_in, in, in_length_in_bytes - 2); 890 crc_req.data = in; 891 crc_req.len = in_length_in_bits - 16; 892 crc_resp.data = q->enc_in; 893 bblib_lte_crc16_gen(&crc_req, &crc_resp); 894 } else 895 rte_memcpy(q->enc_in, in, in_length_in_bytes); 896 897 /* LDPC Encoding */ 898 ldpc_req.Zc = enc->z_c; 899 ldpc_req.baseGraph = enc->basegraph; 900 /* Number of rows set to maximum */ 901 ldpc_req.nRows = ldpc_req.baseGraph == 1 ? 46 : 42; 902 ldpc_req.numberCodeblocks = 1; 903 ldpc_req.input[0] = (int8_t *) q->enc_in; 904 ldpc_resp.output[0] = (int8_t *) q->enc_out; 905 906 bblib_bit_reverse(ldpc_req.input[0], in_length_in_bytes << 3); 907 908 if (bblib_ldpc_encoder_5gnr(&ldpc_req, &ldpc_resp) != 0) { 909 op->status |= 1 << RTE_BBDEV_DRV_ERROR; 910 rte_bbdev_log(ERR, "LDPC Encoder failed"); 911 return; 912 } 913 914 /* 915 * Systematic + Parity : Recreating stream with filler bits, ideally 916 * the bit select could handle this in the RM SDK 917 */ 918 msgLen = (ldpc_req.baseGraph == 1 ? 22 : 10) * ldpc_req.Zc; 919 puntBits = 2 * ldpc_req.Zc; 920 parity_offset = msgLen - puntBits; 921 ippsCopyBE_1u(((uint8_t *) ldpc_req.input[0]) + (puntBits / 8), 922 puntBits%8, q->adapter_output, 0, parity_offset); 923 ippsCopyBE_1u(q->enc_out, 0, q->adapter_output + (parity_offset / 8), 924 parity_offset % 8, ldpc_req.nRows * ldpc_req.Zc); 925 926 out_len = (e + 7) >> 3; 927 /* get output data starting address */ 928 rm_out = (uint8_t *)mbuf_append(m_out_head, m_out, out_len); 929 if (rm_out == NULL) { 930 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 931 rte_bbdev_log(ERR, 932 "Too little space in output mbuf"); 933 return; 934 } 935 /* 936 * rte_bbdev_op_data.offset can be different than the offset 937 * of the appended bytes 938 */ 939 rm_out = rte_pktmbuf_mtod_offset(m_out, uint8_t *, out_offset); 940 941 /* Rate-Matching */ 942 rm_req.E = e; 943 rm_req.Ncb = enc->n_cb; 944 rm_req.Qm = enc->q_m; 945 rm_req.Zc = enc->z_c; 946 rm_req.baseGraph = enc->basegraph; 947 rm_req.input = q->adapter_output; 948 rm_req.nLen = enc->n_filler; 949 rm_req.nullIndex = parity_offset - enc->n_filler; 950 rm_req.rvidx = enc->rv_index; 951 rm_resp.output = q->deint_output; 952 953 if (bblib_LDPC_ratematch_5gnr(&rm_req, &rm_resp) != 0) { 954 op->status |= 1 << RTE_BBDEV_DRV_ERROR; 955 rte_bbdev_log(ERR, "Rate matching failed"); 956 return; 957 } 958 959 /* RM SDK may provide non zero bits on last byte */ 960 if ((e % 8) != 0) 961 q->deint_output[out_len-1] &= (1 << (e % 8)) - 1; 962 963 bblib_bit_reverse((int8_t *) q->deint_output, out_len << 3); 964 965 rte_memcpy(rm_out, q->deint_output, out_len); 966 enc->output.length += out_len; 967 968 #ifdef RTE_BBDEV_OFFLOAD_COST 969 q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time; 970 #endif 971 #else 972 RTE_SET_USED(q); 973 RTE_SET_USED(op); 974 RTE_SET_USED(e); 975 RTE_SET_USED(m_in); 976 RTE_SET_USED(m_out_head); 977 RTE_SET_USED(m_out); 978 RTE_SET_USED(in_offset); 979 RTE_SET_USED(out_offset); 980 RTE_SET_USED(seg_total_left); 981 RTE_SET_USED(q_stats); 982 #endif 983 } 984 985 static inline void 986 enqueue_enc_one_op(struct turbo_sw_queue *q, struct rte_bbdev_enc_op *op, 987 struct rte_bbdev_stats *queue_stats) 988 { 989 uint8_t c, r, crc24_bits = 0; 990 uint16_t k, ncb; 991 uint32_t e; 992 struct rte_bbdev_op_turbo_enc *enc = &op->turbo_enc; 993 uint16_t in_offset = enc->input.offset; 994 uint16_t out_offset = enc->output.offset; 995 struct rte_mbuf *m_in = enc->input.data; 996 struct rte_mbuf *m_out = enc->output.data; 997 struct rte_mbuf *m_out_head = enc->output.data; 998 uint32_t in_length, mbuf_total_left = enc->input.length; 999 uint16_t seg_total_left; 1000 1001 /* Clear op status */ 1002 op->status = 0; 1003 1004 if (mbuf_total_left > RTE_BBDEV_TURBO_MAX_TB_SIZE >> 3) { 1005 rte_bbdev_log(ERR, "TB size (%u) is too big, max: %d", 1006 mbuf_total_left, RTE_BBDEV_TURBO_MAX_TB_SIZE); 1007 op->status = 1 << RTE_BBDEV_DATA_ERROR; 1008 return; 1009 } 1010 1011 if (m_in == NULL || m_out == NULL) { 1012 rte_bbdev_log(ERR, "Invalid mbuf pointer"); 1013 op->status = 1 << RTE_BBDEV_DATA_ERROR; 1014 return; 1015 } 1016 1017 if ((enc->op_flags & RTE_BBDEV_TURBO_CRC_24B_ATTACH) || 1018 (enc->op_flags & RTE_BBDEV_TURBO_CRC_24A_ATTACH)) 1019 crc24_bits = 24; 1020 1021 if (enc->code_block_mode == RTE_BBDEV_TRANSPORT_BLOCK) { 1022 c = enc->tb_params.c; 1023 r = enc->tb_params.r; 1024 } else {/* For Code Block mode */ 1025 c = 1; 1026 r = 0; 1027 } 1028 1029 while (mbuf_total_left > 0 && r < c) { 1030 1031 seg_total_left = rte_pktmbuf_data_len(m_in) - in_offset; 1032 1033 if (enc->code_block_mode == RTE_BBDEV_TRANSPORT_BLOCK) { 1034 k = (r < enc->tb_params.c_neg) ? 1035 enc->tb_params.k_neg : enc->tb_params.k_pos; 1036 ncb = (r < enc->tb_params.c_neg) ? 1037 enc->tb_params.ncb_neg : enc->tb_params.ncb_pos; 1038 e = (r < enc->tb_params.cab) ? 1039 enc->tb_params.ea : enc->tb_params.eb; 1040 } else { 1041 k = enc->cb_params.k; 1042 ncb = enc->cb_params.ncb; 1043 e = enc->cb_params.e; 1044 } 1045 1046 process_enc_cb(q, op, r, c, k, ncb, e, m_in, m_out_head, 1047 m_out, in_offset, out_offset, seg_total_left, 1048 queue_stats); 1049 /* Update total_left */ 1050 in_length = ((k - crc24_bits) >> 3); 1051 mbuf_total_left -= in_length; 1052 /* Update offsets for next CBs (if exist) */ 1053 in_offset += (k - crc24_bits) >> 3; 1054 if (enc->op_flags & RTE_BBDEV_TURBO_RATE_MATCH) 1055 out_offset += e >> 3; 1056 else 1057 out_offset += (k >> 3) * 3 + 2; 1058 1059 /* Update offsets */ 1060 if (seg_total_left == in_length) { 1061 /* Go to the next mbuf */ 1062 m_in = m_in->next; 1063 m_out = m_out->next; 1064 in_offset = 0; 1065 out_offset = 0; 1066 } 1067 r++; 1068 } 1069 1070 /* check if all input data was processed */ 1071 if (mbuf_total_left != 0) { 1072 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 1073 rte_bbdev_log(ERR, 1074 "Mismatch between mbuf length and included CBs sizes"); 1075 } 1076 } 1077 1078 1079 static inline void 1080 enqueue_ldpc_enc_one_op(struct turbo_sw_queue *q, struct rte_bbdev_enc_op *op, 1081 struct rte_bbdev_stats *queue_stats) 1082 { 1083 uint8_t c, r, crc24_bits = 0; 1084 uint32_t e; 1085 struct rte_bbdev_op_ldpc_enc *enc = &op->ldpc_enc; 1086 uint16_t in_offset = enc->input.offset; 1087 uint16_t out_offset = enc->output.offset; 1088 struct rte_mbuf *m_in = enc->input.data; 1089 struct rte_mbuf *m_out = enc->output.data; 1090 struct rte_mbuf *m_out_head = enc->output.data; 1091 uint32_t in_length, mbuf_total_left = enc->input.length; 1092 1093 uint16_t seg_total_left; 1094 1095 /* Clear op status */ 1096 op->status = 0; 1097 1098 if (mbuf_total_left > RTE_BBDEV_TURBO_MAX_TB_SIZE >> 3) { 1099 rte_bbdev_log(ERR, "TB size (%u) is too big, max: %d", 1100 mbuf_total_left, RTE_BBDEV_TURBO_MAX_TB_SIZE); 1101 op->status = 1 << RTE_BBDEV_DATA_ERROR; 1102 return; 1103 } 1104 1105 if (m_in == NULL || m_out == NULL) { 1106 rte_bbdev_log(ERR, "Invalid mbuf pointer"); 1107 op->status = 1 << RTE_BBDEV_DATA_ERROR; 1108 return; 1109 } 1110 1111 if ((enc->op_flags & RTE_BBDEV_TURBO_CRC_24B_ATTACH) || 1112 (enc->op_flags & RTE_BBDEV_TURBO_CRC_24A_ATTACH)) 1113 crc24_bits = 24; 1114 1115 if (enc->code_block_mode == RTE_BBDEV_TRANSPORT_BLOCK) { 1116 c = enc->tb_params.c; 1117 r = enc->tb_params.r; 1118 } else { /* For Code Block mode */ 1119 c = 1; 1120 r = 0; 1121 } 1122 1123 while (mbuf_total_left > 0 && r < c) { 1124 1125 seg_total_left = rte_pktmbuf_data_len(m_in) - in_offset; 1126 1127 if (enc->code_block_mode == RTE_BBDEV_TRANSPORT_BLOCK) { 1128 e = (r < enc->tb_params.cab) ? 1129 enc->tb_params.ea : enc->tb_params.eb; 1130 } else { 1131 e = enc->cb_params.e; 1132 } 1133 1134 process_ldpc_enc_cb(q, op, e, m_in, m_out_head, 1135 m_out, in_offset, out_offset, seg_total_left, 1136 queue_stats); 1137 /* Update total_left */ 1138 in_length = (enc->basegraph == 1 ? 22 : 10) * enc->z_c; 1139 in_length = ((in_length - crc24_bits - enc->n_filler) >> 3); 1140 mbuf_total_left -= in_length; 1141 /* Update offsets for next CBs (if exist) */ 1142 in_offset += in_length; 1143 out_offset += (e + 7) >> 3; 1144 1145 /* Update offsets */ 1146 if (seg_total_left == in_length) { 1147 /* Go to the next mbuf */ 1148 m_in = m_in->next; 1149 m_out = m_out->next; 1150 in_offset = 0; 1151 out_offset = 0; 1152 } 1153 r++; 1154 } 1155 1156 /* check if all input data was processed */ 1157 if (mbuf_total_left != 0) { 1158 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 1159 rte_bbdev_log(ERR, 1160 "Mismatch between mbuf length and included CBs sizes %d", 1161 mbuf_total_left); 1162 } 1163 } 1164 1165 static inline uint16_t 1166 enqueue_enc_all_ops(struct turbo_sw_queue *q, struct rte_bbdev_enc_op **ops, 1167 uint16_t nb_ops, struct rte_bbdev_stats *queue_stats) 1168 { 1169 uint16_t i; 1170 #ifdef RTE_BBDEV_OFFLOAD_COST 1171 queue_stats->acc_offload_cycles = 0; 1172 #endif 1173 1174 for (i = 0; i < nb_ops; ++i) 1175 enqueue_enc_one_op(q, ops[i], queue_stats); 1176 1177 return rte_ring_enqueue_burst(q->processed_pkts, (void **)ops, nb_ops, 1178 NULL); 1179 } 1180 1181 static inline uint16_t 1182 enqueue_ldpc_enc_all_ops(struct turbo_sw_queue *q, 1183 struct rte_bbdev_enc_op **ops, 1184 uint16_t nb_ops, struct rte_bbdev_stats *queue_stats) 1185 { 1186 uint16_t i; 1187 #ifdef RTE_BBDEV_OFFLOAD_COST 1188 queue_stats->acc_offload_cycles = 0; 1189 #endif 1190 1191 for (i = 0; i < nb_ops; ++i) 1192 enqueue_ldpc_enc_one_op(q, ops[i], queue_stats); 1193 1194 return rte_ring_enqueue_burst(q->processed_pkts, (void **)ops, nb_ops, 1195 NULL); 1196 } 1197 1198 #ifdef RTE_BBDEV_SDK_AVX2 1199 static inline void 1200 move_padding_bytes(const uint8_t *in, uint8_t *out, uint16_t k, 1201 uint16_t ncb) 1202 { 1203 uint16_t d = k + 4; 1204 uint16_t kpi = ncb / 3; 1205 uint16_t nd = kpi - d; 1206 1207 rte_memcpy(&out[nd], in, d); 1208 rte_memcpy(&out[nd + kpi + 64], &in[kpi], d); 1209 rte_memcpy(&out[(nd - 1) + 2 * (kpi + 64)], &in[2 * kpi], d); 1210 } 1211 #endif 1212 1213 static inline void 1214 process_dec_cb(struct turbo_sw_queue *q, struct rte_bbdev_dec_op *op, 1215 uint8_t c, uint16_t k, uint16_t kw, struct rte_mbuf *m_in, 1216 struct rte_mbuf *m_out_head, struct rte_mbuf *m_out, 1217 uint16_t in_offset, uint16_t out_offset, bool check_crc_24b, 1218 uint16_t crc24_overlap, uint16_t in_length, 1219 struct rte_bbdev_stats *q_stats) 1220 { 1221 #ifdef RTE_BBDEV_SDK_AVX2 1222 #ifdef RTE_LIBRTE_BBDEV_DEBUG 1223 int ret; 1224 #else 1225 RTE_SET_USED(in_length); 1226 #endif 1227 int32_t k_idx; 1228 int32_t iter_cnt; 1229 uint8_t *in, *out, *adapter_input; 1230 int32_t ncb, ncb_without_null; 1231 struct bblib_turbo_adapter_ul_response adapter_resp; 1232 struct bblib_turbo_adapter_ul_request adapter_req; 1233 struct bblib_turbo_decoder_request turbo_req; 1234 struct bblib_turbo_decoder_response turbo_resp; 1235 struct rte_bbdev_op_turbo_dec *dec = &op->turbo_dec; 1236 #ifdef RTE_BBDEV_OFFLOAD_COST 1237 uint64_t start_time; 1238 #else 1239 RTE_SET_USED(q_stats); 1240 #endif 1241 1242 k_idx = compute_idx(k); 1243 1244 #ifdef RTE_LIBRTE_BBDEV_DEBUG 1245 ret = is_dec_input_valid(k_idx, kw, in_length); 1246 if (ret != 0) { 1247 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 1248 return; 1249 } 1250 #endif 1251 1252 in = rte_pktmbuf_mtod_offset(m_in, uint8_t *, in_offset); 1253 ncb = kw; 1254 ncb_without_null = (k + 4) * 3; 1255 1256 if (check_bit(dec->op_flags, RTE_BBDEV_TURBO_SUBBLOCK_DEINTERLEAVE)) { 1257 struct bblib_deinterleave_ul_request deint_req; 1258 struct bblib_deinterleave_ul_response deint_resp; 1259 1260 deint_req.circ_buffer = BBLIB_FULL_CIRCULAR_BUFFER; 1261 deint_req.pharqbuffer = in; 1262 deint_req.ncb = ncb; 1263 deint_resp.pinteleavebuffer = q->deint_output; 1264 1265 #ifdef RTE_BBDEV_OFFLOAD_COST 1266 start_time = rte_rdtsc_precise(); 1267 #endif 1268 /* Sub-block De-Interleaving */ 1269 bblib_deinterleave_ul(&deint_req, &deint_resp); 1270 #ifdef RTE_BBDEV_OFFLOAD_COST 1271 q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time; 1272 #endif 1273 } else 1274 move_padding_bytes(in, q->deint_output, k, ncb); 1275 1276 adapter_input = q->deint_output; 1277 1278 if (dec->op_flags & RTE_BBDEV_TURBO_POS_LLR_1_BIT_IN) 1279 adapter_req.isinverted = 1; 1280 else if (dec->op_flags & RTE_BBDEV_TURBO_NEG_LLR_1_BIT_IN) 1281 adapter_req.isinverted = 0; 1282 else { 1283 op->status |= 1 << RTE_BBDEV_DRV_ERROR; 1284 rte_bbdev_log(ERR, "LLR format wasn't specified"); 1285 return; 1286 } 1287 1288 adapter_req.ncb = ncb_without_null; 1289 adapter_req.pinteleavebuffer = adapter_input; 1290 adapter_resp.pharqout = q->adapter_output; 1291 1292 #ifdef RTE_BBDEV_OFFLOAD_COST 1293 start_time = rte_rdtsc_precise(); 1294 #endif 1295 /* Turbo decode adaptation */ 1296 bblib_turbo_adapter_ul(&adapter_req, &adapter_resp); 1297 #ifdef RTE_BBDEV_OFFLOAD_COST 1298 q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time; 1299 #endif 1300 1301 out = (uint8_t *)mbuf_append(m_out_head, m_out, 1302 ((k - crc24_overlap) >> 3)); 1303 if (out == NULL) { 1304 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 1305 rte_bbdev_log(ERR, "Too little space in output mbuf"); 1306 return; 1307 } 1308 /* rte_bbdev_op_data.offset can be different than the offset of the 1309 * appended bytes 1310 */ 1311 out = rte_pktmbuf_mtod_offset(m_out, uint8_t *, out_offset); 1312 if (check_crc_24b) 1313 turbo_req.c = c + 1; 1314 else 1315 turbo_req.c = c; 1316 turbo_req.input = (int8_t *)q->adapter_output; 1317 turbo_req.k = k; 1318 turbo_req.k_idx = k_idx; 1319 turbo_req.max_iter_num = dec->iter_max; 1320 turbo_req.early_term_disable = !check_bit(dec->op_flags, 1321 RTE_BBDEV_TURBO_EARLY_TERMINATION); 1322 turbo_resp.ag_buf = q->ag; 1323 turbo_resp.cb_buf = q->code_block; 1324 turbo_resp.output = out; 1325 1326 #ifdef RTE_BBDEV_OFFLOAD_COST 1327 start_time = rte_rdtsc_precise(); 1328 #endif 1329 /* Turbo decode */ 1330 iter_cnt = bblib_turbo_decoder(&turbo_req, &turbo_resp); 1331 #ifdef RTE_BBDEV_OFFLOAD_COST 1332 q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time; 1333 #endif 1334 dec->hard_output.length += (k >> 3); 1335 1336 if (iter_cnt > 0) { 1337 /* Temporary solution for returned iter_count from SDK */ 1338 iter_cnt = (iter_cnt - 1) >> 1; 1339 dec->iter_count = RTE_MAX(iter_cnt, dec->iter_count); 1340 } else { 1341 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 1342 rte_bbdev_log(ERR, "Turbo Decoder failed"); 1343 return; 1344 } 1345 #else 1346 RTE_SET_USED(q); 1347 RTE_SET_USED(op); 1348 RTE_SET_USED(c); 1349 RTE_SET_USED(k); 1350 RTE_SET_USED(kw); 1351 RTE_SET_USED(m_in); 1352 RTE_SET_USED(m_out_head); 1353 RTE_SET_USED(m_out); 1354 RTE_SET_USED(in_offset); 1355 RTE_SET_USED(out_offset); 1356 RTE_SET_USED(check_crc_24b); 1357 RTE_SET_USED(crc24_overlap); 1358 RTE_SET_USED(in_length); 1359 RTE_SET_USED(q_stats); 1360 #endif 1361 } 1362 1363 static inline void 1364 process_ldpc_dec_cb(struct turbo_sw_queue *q, struct rte_bbdev_dec_op *op, 1365 uint8_t c, uint16_t out_length, uint32_t e, 1366 struct rte_mbuf *m_in, 1367 struct rte_mbuf *m_out_head, struct rte_mbuf *m_out, 1368 struct rte_mbuf *m_harq_in, 1369 struct rte_mbuf *m_harq_out_head, struct rte_mbuf *m_harq_out, 1370 uint16_t in_offset, uint16_t out_offset, 1371 uint16_t harq_in_offset, uint16_t harq_out_offset, 1372 bool check_crc_24b, 1373 uint16_t crc24_overlap, uint16_t in_length, 1374 struct rte_bbdev_stats *q_stats) 1375 { 1376 #ifdef RTE_BBDEV_SDK_AVX512 1377 RTE_SET_USED(in_length); 1378 RTE_SET_USED(c); 1379 uint8_t *in, *out, *harq_in, *harq_out, *adapter_input; 1380 struct bblib_rate_dematching_5gnr_request derm_req; 1381 struct bblib_rate_dematching_5gnr_response derm_resp; 1382 struct bblib_ldpc_decoder_5gnr_request dec_req; 1383 struct bblib_ldpc_decoder_5gnr_response dec_resp; 1384 struct bblib_crc_request crc_req; 1385 struct bblib_crc_response crc_resp; 1386 struct rte_bbdev_op_ldpc_dec *dec = &op->ldpc_dec; 1387 uint16_t K, parity_offset, sys_cols, outLenWithCrc; 1388 int16_t deRmOutSize, numRows; 1389 1390 /* Compute some LDPC BG lengths */ 1391 outLenWithCrc = out_length + (crc24_overlap >> 3); 1392 sys_cols = (dec->basegraph == 1) ? 22 : 10; 1393 K = sys_cols * dec->z_c; 1394 parity_offset = K - 2 * dec->z_c; 1395 1396 #ifdef RTE_BBDEV_OFFLOAD_COST 1397 uint64_t start_time = rte_rdtsc_precise(); 1398 #else 1399 RTE_SET_USED(q_stats); 1400 #endif 1401 1402 in = rte_pktmbuf_mtod_offset(m_in, uint8_t *, in_offset); 1403 1404 if (check_bit(dec->op_flags, RTE_BBDEV_LDPC_HQ_COMBINE_IN_ENABLE)) { 1405 /** 1406 * Single contiguous block from the first LLR of the 1407 * circular buffer. 1408 */ 1409 harq_in = NULL; 1410 if (m_harq_in != NULL) 1411 harq_in = rte_pktmbuf_mtod_offset(m_harq_in, 1412 uint8_t *, harq_in_offset); 1413 if (harq_in == NULL) { 1414 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 1415 rte_bbdev_log(ERR, "No space in harq input mbuf"); 1416 return; 1417 } 1418 uint16_t harq_in_length = RTE_MIN( 1419 dec->harq_combined_input.length, 1420 (uint32_t) dec->n_cb); 1421 memset(q->ag + harq_in_length, 0, 1422 dec->n_cb - harq_in_length); 1423 rte_memcpy(q->ag, harq_in, harq_in_length); 1424 } 1425 1426 derm_req.p_in = (int8_t *) in; 1427 derm_req.p_harq = q->ag; /* This doesn't include the filler bits */ 1428 derm_req.base_graph = dec->basegraph; 1429 derm_req.zc = dec->z_c; 1430 derm_req.ncb = dec->n_cb; 1431 derm_req.e = e; 1432 derm_req.k0 = 0; /* Actual output from SDK */ 1433 derm_req.isretx = check_bit(dec->op_flags, 1434 RTE_BBDEV_LDPC_HQ_COMBINE_IN_ENABLE); 1435 derm_req.rvid = dec->rv_index; 1436 derm_req.modulation_order = dec->q_m; 1437 derm_req.start_null_index = parity_offset - dec->n_filler; 1438 derm_req.num_of_null = dec->n_filler; 1439 1440 bblib_rate_dematching_5gnr(&derm_req, &derm_resp); 1441 1442 /* Compute RM out size and number of rows */ 1443 deRmOutSize = RTE_MIN( 1444 derm_req.k0 + derm_req.e - 1445 ((derm_req.k0 < derm_req.start_null_index) ? 1446 0 : dec->n_filler), 1447 dec->n_cb - dec->n_filler); 1448 if (m_harq_in != NULL) 1449 deRmOutSize = RTE_MAX(deRmOutSize, 1450 RTE_MIN(dec->n_cb - dec->n_filler, 1451 m_harq_in->data_len)); 1452 numRows = ((deRmOutSize + dec->n_filler + dec->z_c - 1) / dec->z_c) 1453 - sys_cols + 2; 1454 numRows = RTE_MAX(4, numRows); 1455 1456 /* get output data starting address */ 1457 out = (uint8_t *)mbuf_append(m_out_head, m_out, out_length); 1458 if (out == NULL) { 1459 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 1460 rte_bbdev_log(ERR, 1461 "Too little space in LDPC decoder output mbuf"); 1462 return; 1463 } 1464 1465 /* rte_bbdev_op_data.offset can be different than the offset 1466 * of the appended bytes 1467 */ 1468 out = rte_pktmbuf_mtod_offset(m_out, uint8_t *, out_offset); 1469 adapter_input = q->enc_out; 1470 1471 dec_req.Zc = dec->z_c; 1472 dec_req.baseGraph = dec->basegraph; 1473 dec_req.nRows = numRows; 1474 dec_req.numChannelLlrs = deRmOutSize; 1475 dec_req.varNodes = derm_req.p_harq; 1476 dec_req.numFillerBits = dec->n_filler; 1477 dec_req.maxIterations = dec->iter_max; 1478 dec_req.enableEarlyTermination = check_bit(dec->op_flags, 1479 RTE_BBDEV_LDPC_ITERATION_STOP_ENABLE); 1480 dec_resp.varNodes = (int16_t *) q->adapter_output; 1481 dec_resp.compactedMessageBytes = q->enc_out; 1482 1483 bblib_ldpc_decoder_5gnr(&dec_req, &dec_resp); 1484 1485 dec->iter_count = RTE_MAX(dec_resp.iterationAtTermination, 1486 dec->iter_count); 1487 if (!dec_resp.parityPassedAtTermination) 1488 op->status |= 1 << RTE_BBDEV_SYNDROME_ERROR; 1489 1490 bblib_bit_reverse((int8_t *) q->enc_out, outLenWithCrc << 3); 1491 1492 if (check_bit(dec->op_flags, RTE_BBDEV_LDPC_CRC_TYPE_24A_CHECK) || 1493 check_bit(dec->op_flags, 1494 RTE_BBDEV_LDPC_CRC_TYPE_24B_CHECK)) { 1495 crc_req.data = adapter_input; 1496 crc_req.len = K - dec->n_filler - 24; 1497 crc_resp.check_passed = false; 1498 crc_resp.data = adapter_input; 1499 if (check_crc_24b) 1500 bblib_lte_crc24b_check(&crc_req, &crc_resp); 1501 else 1502 bblib_lte_crc24a_check(&crc_req, &crc_resp); 1503 if (!crc_resp.check_passed) 1504 op->status |= 1 << RTE_BBDEV_CRC_ERROR; 1505 } else if (check_bit(dec->op_flags, RTE_BBDEV_LDPC_CRC_TYPE_16_CHECK)) { 1506 crc_req.data = adapter_input; 1507 crc_req.len = K - dec->n_filler - 16; 1508 crc_resp.check_passed = false; 1509 crc_resp.data = adapter_input; 1510 bblib_lte_crc16_check(&crc_req, &crc_resp); 1511 if (!crc_resp.check_passed) 1512 op->status |= 1 << RTE_BBDEV_CRC_ERROR; 1513 } 1514 1515 #ifdef RTE_BBDEV_OFFLOAD_COST 1516 q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time; 1517 #endif 1518 if (check_bit(dec->op_flags, RTE_BBDEV_LDPC_HQ_COMBINE_OUT_ENABLE)) { 1519 harq_out = NULL; 1520 if (m_harq_out != NULL) { 1521 /* Initialize HARQ data length since we overwrite */ 1522 m_harq_out->data_len = 0; 1523 /* Check there is enough space 1524 * in the HARQ outbound buffer 1525 */ 1526 harq_out = (uint8_t *)mbuf_append(m_harq_out_head, 1527 m_harq_out, deRmOutSize); 1528 } 1529 if (harq_out == NULL) { 1530 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 1531 rte_bbdev_log(ERR, "No space in HARQ output mbuf"); 1532 return; 1533 } 1534 /* get output data starting address and overwrite the data */ 1535 harq_out = rte_pktmbuf_mtod_offset(m_harq_out, uint8_t *, 1536 harq_out_offset); 1537 rte_memcpy(harq_out, derm_req.p_harq, deRmOutSize); 1538 dec->harq_combined_output.length += deRmOutSize; 1539 } 1540 1541 rte_memcpy(out, adapter_input, out_length); 1542 dec->hard_output.length += out_length; 1543 #else 1544 RTE_SET_USED(q); 1545 RTE_SET_USED(op); 1546 RTE_SET_USED(c); 1547 RTE_SET_USED(out_length); 1548 RTE_SET_USED(e); 1549 RTE_SET_USED(m_in); 1550 RTE_SET_USED(m_out_head); 1551 RTE_SET_USED(m_out); 1552 RTE_SET_USED(m_harq_in); 1553 RTE_SET_USED(m_harq_out_head); 1554 RTE_SET_USED(m_harq_out); 1555 RTE_SET_USED(harq_in_offset); 1556 RTE_SET_USED(harq_out_offset); 1557 RTE_SET_USED(in_offset); 1558 RTE_SET_USED(out_offset); 1559 RTE_SET_USED(check_crc_24b); 1560 RTE_SET_USED(crc24_overlap); 1561 RTE_SET_USED(in_length); 1562 RTE_SET_USED(q_stats); 1563 #endif 1564 } 1565 1566 1567 static inline void 1568 enqueue_dec_one_op(struct turbo_sw_queue *q, struct rte_bbdev_dec_op *op, 1569 struct rte_bbdev_stats *queue_stats) 1570 { 1571 uint8_t c, r = 0; 1572 uint16_t kw, k = 0; 1573 uint16_t crc24_overlap = 0; 1574 struct rte_bbdev_op_turbo_dec *dec = &op->turbo_dec; 1575 struct rte_mbuf *m_in = dec->input.data; 1576 struct rte_mbuf *m_out = dec->hard_output.data; 1577 struct rte_mbuf *m_out_head = dec->hard_output.data; 1578 uint16_t in_offset = dec->input.offset; 1579 uint16_t out_offset = dec->hard_output.offset; 1580 uint32_t mbuf_total_left = dec->input.length; 1581 uint16_t seg_total_left; 1582 1583 /* Clear op status */ 1584 op->status = 0; 1585 1586 if (m_in == NULL || m_out == NULL) { 1587 rte_bbdev_log(ERR, "Invalid mbuf pointer"); 1588 op->status = 1 << RTE_BBDEV_DATA_ERROR; 1589 return; 1590 } 1591 1592 if (dec->code_block_mode == RTE_BBDEV_TRANSPORT_BLOCK) { 1593 c = dec->tb_params.c; 1594 } else { /* For Code Block mode */ 1595 k = dec->cb_params.k; 1596 c = 1; 1597 } 1598 1599 if ((c > 1) && !check_bit(dec->op_flags, 1600 RTE_BBDEV_TURBO_DEC_TB_CRC_24B_KEEP)) 1601 crc24_overlap = 24; 1602 1603 while (mbuf_total_left > 0) { 1604 if (dec->code_block_mode == RTE_BBDEV_TRANSPORT_BLOCK) 1605 k = (r < dec->tb_params.c_neg) ? 1606 dec->tb_params.k_neg : dec->tb_params.k_pos; 1607 1608 seg_total_left = rte_pktmbuf_data_len(m_in) - in_offset; 1609 1610 /* Calculates circular buffer size (Kw). 1611 * According to 3gpp 36.212 section 5.1.4.2 1612 * Kw = 3 * Kpi, 1613 * where: 1614 * Kpi = nCol * nRow 1615 * where nCol is 32 and nRow can be calculated from: 1616 * D =< nCol * nRow 1617 * where D is the size of each output from turbo encoder block 1618 * (k + 4). 1619 */ 1620 kw = RTE_ALIGN_CEIL(k + 4, RTE_BBDEV_TURBO_C_SUBBLOCK) * 3; 1621 1622 process_dec_cb(q, op, c, k, kw, m_in, m_out_head, m_out, 1623 in_offset, out_offset, check_bit(dec->op_flags, 1624 RTE_BBDEV_TURBO_CRC_TYPE_24B), crc24_overlap, 1625 seg_total_left, queue_stats); 1626 1627 /* To keep CRC24 attached to end of Code block, use 1628 * RTE_BBDEV_TURBO_DEC_TB_CRC_24B_KEEP flag as it 1629 * removed by default once verified. 1630 */ 1631 1632 mbuf_total_left -= kw; 1633 1634 /* Update offsets */ 1635 if (seg_total_left == kw) { 1636 /* Go to the next mbuf */ 1637 m_in = m_in->next; 1638 m_out = m_out->next; 1639 in_offset = 0; 1640 out_offset = 0; 1641 } else { 1642 /* Update offsets for next CBs (if exist) */ 1643 in_offset += kw; 1644 out_offset += ((k - crc24_overlap) >> 3); 1645 } 1646 r++; 1647 } 1648 } 1649 1650 static inline void 1651 enqueue_ldpc_dec_one_op(struct turbo_sw_queue *q, struct rte_bbdev_dec_op *op, 1652 struct rte_bbdev_stats *queue_stats) 1653 { 1654 uint8_t c, r = 0; 1655 uint32_t e; 1656 uint16_t out_length, crc24_overlap = 0; 1657 struct rte_bbdev_op_ldpc_dec *dec = &op->ldpc_dec; 1658 struct rte_mbuf *m_in = dec->input.data; 1659 struct rte_mbuf *m_harq_in = dec->harq_combined_input.data; 1660 struct rte_mbuf *m_harq_out = dec->harq_combined_output.data; 1661 struct rte_mbuf *m_harq_out_head = dec->harq_combined_output.data; 1662 struct rte_mbuf *m_out = dec->hard_output.data; 1663 struct rte_mbuf *m_out_head = dec->hard_output.data; 1664 uint16_t in_offset = dec->input.offset; 1665 uint16_t harq_in_offset = dec->harq_combined_input.offset; 1666 uint16_t harq_out_offset = dec->harq_combined_output.offset; 1667 uint16_t out_offset = dec->hard_output.offset; 1668 uint32_t mbuf_total_left = dec->input.length; 1669 uint16_t seg_total_left; 1670 1671 /* Clear op status */ 1672 op->status = 0; 1673 1674 if (m_in == NULL || m_out == NULL) { 1675 rte_bbdev_log(ERR, "Invalid mbuf pointer"); 1676 op->status = 1 << RTE_BBDEV_DATA_ERROR; 1677 return; 1678 } 1679 1680 if (dec->code_block_mode == RTE_BBDEV_TRANSPORT_BLOCK) { 1681 c = dec->tb_params.c; 1682 e = dec->tb_params.ea; 1683 } else { /* For Code Block mode */ 1684 c = 1; 1685 e = dec->cb_params.e; 1686 } 1687 1688 if (check_bit(dec->op_flags, RTE_BBDEV_LDPC_CRC_TYPE_24B_DROP)) 1689 crc24_overlap = 24; 1690 1691 out_length = (dec->basegraph == 1 ? 22 : 10) * dec->z_c; /* K */ 1692 out_length = ((out_length - crc24_overlap - dec->n_filler) >> 3); 1693 1694 while (mbuf_total_left > 0) { 1695 if (dec->code_block_mode == RTE_BBDEV_TRANSPORT_BLOCK) 1696 e = (r < dec->tb_params.cab) ? 1697 dec->tb_params.ea : dec->tb_params.eb; 1698 /* Special case handling when overusing mbuf */ 1699 if (e < RTE_BBDEV_LDPC_E_MAX_MBUF) 1700 seg_total_left = rte_pktmbuf_data_len(m_in) - in_offset; 1701 else 1702 seg_total_left = e; 1703 1704 process_ldpc_dec_cb(q, op, c, out_length, e, 1705 m_in, m_out_head, m_out, 1706 m_harq_in, m_harq_out_head, m_harq_out, 1707 in_offset, out_offset, harq_in_offset, 1708 harq_out_offset, 1709 check_bit(dec->op_flags, 1710 RTE_BBDEV_LDPC_CRC_TYPE_24B_CHECK), 1711 crc24_overlap, 1712 seg_total_left, queue_stats); 1713 1714 /* To keep CRC24 attached to end of Code block, use 1715 * RTE_BBDEV_LDPC_DEC_TB_CRC_24B_KEEP flag as it 1716 * removed by default once verified. 1717 */ 1718 1719 mbuf_total_left -= e; 1720 1721 /* Update offsets */ 1722 if (seg_total_left == e) { 1723 /* Go to the next mbuf */ 1724 m_in = m_in->next; 1725 m_out = m_out->next; 1726 if (m_harq_in != NULL) 1727 m_harq_in = m_harq_in->next; 1728 if (m_harq_out != NULL) 1729 m_harq_out = m_harq_out->next; 1730 in_offset = 0; 1731 out_offset = 0; 1732 harq_in_offset = 0; 1733 harq_out_offset = 0; 1734 } else { 1735 /* Update offsets for next CBs (if exist) */ 1736 in_offset += e; 1737 out_offset += out_length; 1738 } 1739 r++; 1740 } 1741 } 1742 1743 static inline uint16_t 1744 enqueue_dec_all_ops(struct turbo_sw_queue *q, struct rte_bbdev_dec_op **ops, 1745 uint16_t nb_ops, struct rte_bbdev_stats *queue_stats) 1746 { 1747 uint16_t i; 1748 #ifdef RTE_BBDEV_OFFLOAD_COST 1749 queue_stats->acc_offload_cycles = 0; 1750 #endif 1751 1752 for (i = 0; i < nb_ops; ++i) 1753 enqueue_dec_one_op(q, ops[i], queue_stats); 1754 1755 return rte_ring_enqueue_burst(q->processed_pkts, (void **)ops, nb_ops, 1756 NULL); 1757 } 1758 1759 static inline uint16_t 1760 enqueue_ldpc_dec_all_ops(struct turbo_sw_queue *q, 1761 struct rte_bbdev_dec_op **ops, 1762 uint16_t nb_ops, struct rte_bbdev_stats *queue_stats) 1763 { 1764 uint16_t i; 1765 #ifdef RTE_BBDEV_OFFLOAD_COST 1766 queue_stats->acc_offload_cycles = 0; 1767 #endif 1768 1769 for (i = 0; i < nb_ops; ++i) 1770 enqueue_ldpc_dec_one_op(q, ops[i], queue_stats); 1771 1772 return rte_ring_enqueue_burst(q->processed_pkts, (void **)ops, nb_ops, 1773 NULL); 1774 } 1775 1776 /* Enqueue burst */ 1777 static uint16_t 1778 enqueue_enc_ops(struct rte_bbdev_queue_data *q_data, 1779 struct rte_bbdev_enc_op **ops, uint16_t nb_ops) 1780 { 1781 void *queue = q_data->queue_private; 1782 struct turbo_sw_queue *q = queue; 1783 uint16_t nb_enqueued = 0; 1784 1785 nb_enqueued = enqueue_enc_all_ops(q, ops, nb_ops, &q_data->queue_stats); 1786 1787 q_data->queue_stats.enqueue_err_count += nb_ops - nb_enqueued; 1788 q_data->queue_stats.enqueued_count += nb_enqueued; 1789 1790 return nb_enqueued; 1791 } 1792 1793 /* Enqueue burst */ 1794 static uint16_t 1795 enqueue_ldpc_enc_ops(struct rte_bbdev_queue_data *q_data, 1796 struct rte_bbdev_enc_op **ops, uint16_t nb_ops) 1797 { 1798 void *queue = q_data->queue_private; 1799 struct turbo_sw_queue *q = queue; 1800 uint16_t nb_enqueued = 0; 1801 1802 nb_enqueued = enqueue_ldpc_enc_all_ops( 1803 q, ops, nb_ops, &q_data->queue_stats); 1804 1805 q_data->queue_stats.enqueue_err_count += nb_ops - nb_enqueued; 1806 q_data->queue_stats.enqueued_count += nb_enqueued; 1807 1808 return nb_enqueued; 1809 } 1810 1811 /* Enqueue burst */ 1812 static uint16_t 1813 enqueue_dec_ops(struct rte_bbdev_queue_data *q_data, 1814 struct rte_bbdev_dec_op **ops, uint16_t nb_ops) 1815 { 1816 void *queue = q_data->queue_private; 1817 struct turbo_sw_queue *q = queue; 1818 uint16_t nb_enqueued = 0; 1819 1820 nb_enqueued = enqueue_dec_all_ops(q, ops, nb_ops, &q_data->queue_stats); 1821 1822 q_data->queue_stats.enqueue_err_count += nb_ops - nb_enqueued; 1823 q_data->queue_stats.enqueued_count += nb_enqueued; 1824 1825 return nb_enqueued; 1826 } 1827 1828 /* Enqueue burst */ 1829 static uint16_t 1830 enqueue_ldpc_dec_ops(struct rte_bbdev_queue_data *q_data, 1831 struct rte_bbdev_dec_op **ops, uint16_t nb_ops) 1832 { 1833 void *queue = q_data->queue_private; 1834 struct turbo_sw_queue *q = queue; 1835 uint16_t nb_enqueued = 0; 1836 1837 nb_enqueued = enqueue_ldpc_dec_all_ops(q, ops, nb_ops, 1838 &q_data->queue_stats); 1839 1840 q_data->queue_stats.enqueue_err_count += nb_ops - nb_enqueued; 1841 q_data->queue_stats.enqueued_count += nb_enqueued; 1842 1843 return nb_enqueued; 1844 } 1845 1846 /* Dequeue decode burst */ 1847 static uint16_t 1848 dequeue_dec_ops(struct rte_bbdev_queue_data *q_data, 1849 struct rte_bbdev_dec_op **ops, uint16_t nb_ops) 1850 { 1851 struct turbo_sw_queue *q = q_data->queue_private; 1852 uint16_t nb_dequeued = rte_ring_dequeue_burst(q->processed_pkts, 1853 (void **)ops, nb_ops, NULL); 1854 q_data->queue_stats.dequeued_count += nb_dequeued; 1855 1856 return nb_dequeued; 1857 } 1858 1859 /* Dequeue encode burst */ 1860 static uint16_t 1861 dequeue_enc_ops(struct rte_bbdev_queue_data *q_data, 1862 struct rte_bbdev_enc_op **ops, uint16_t nb_ops) 1863 { 1864 struct turbo_sw_queue *q = q_data->queue_private; 1865 uint16_t nb_dequeued = rte_ring_dequeue_burst(q->processed_pkts, 1866 (void **)ops, nb_ops, NULL); 1867 q_data->queue_stats.dequeued_count += nb_dequeued; 1868 1869 return nb_dequeued; 1870 } 1871 1872 /* Parse 16bit integer from string argument */ 1873 static inline int 1874 parse_u16_arg(const char *key, const char *value, void *extra_args) 1875 { 1876 uint16_t *u16 = extra_args; 1877 unsigned int long result; 1878 1879 if ((value == NULL) || (extra_args == NULL)) 1880 return -EINVAL; 1881 errno = 0; 1882 result = strtoul(value, NULL, 0); 1883 if ((result >= (1 << 16)) || (errno != 0)) { 1884 rte_bbdev_log(ERR, "Invalid value %lu for %s", result, key); 1885 return -ERANGE; 1886 } 1887 *u16 = (uint16_t)result; 1888 return 0; 1889 } 1890 1891 /* Parse parameters used to create device */ 1892 static int 1893 parse_turbo_sw_params(struct turbo_sw_params *params, const char *input_args) 1894 { 1895 struct rte_kvargs *kvlist = NULL; 1896 int ret = 0; 1897 1898 if (params == NULL) 1899 return -EINVAL; 1900 if (input_args) { 1901 kvlist = rte_kvargs_parse(input_args, turbo_sw_valid_params); 1902 if (kvlist == NULL) 1903 return -EFAULT; 1904 1905 ret = rte_kvargs_process(kvlist, turbo_sw_valid_params[0], 1906 &parse_u16_arg, ¶ms->queues_num); 1907 if (ret < 0) 1908 goto exit; 1909 1910 ret = rte_kvargs_process(kvlist, turbo_sw_valid_params[1], 1911 &parse_u16_arg, ¶ms->socket_id); 1912 if (ret < 0) 1913 goto exit; 1914 1915 if (params->socket_id >= RTE_MAX_NUMA_NODES) { 1916 rte_bbdev_log(ERR, "Invalid socket, must be < %u", 1917 RTE_MAX_NUMA_NODES); 1918 goto exit; 1919 } 1920 } 1921 1922 exit: 1923 rte_kvargs_free(kvlist); 1924 return ret; 1925 } 1926 1927 /* Create device */ 1928 static int 1929 turbo_sw_bbdev_create(struct rte_vdev_device *vdev, 1930 struct turbo_sw_params *init_params) 1931 { 1932 struct rte_bbdev *bbdev; 1933 const char *name = rte_vdev_device_name(vdev); 1934 1935 bbdev = rte_bbdev_allocate(name); 1936 if (bbdev == NULL) 1937 return -ENODEV; 1938 1939 bbdev->data->dev_private = rte_zmalloc_socket(name, 1940 sizeof(struct bbdev_private), RTE_CACHE_LINE_SIZE, 1941 init_params->socket_id); 1942 if (bbdev->data->dev_private == NULL) { 1943 rte_bbdev_release(bbdev); 1944 return -ENOMEM; 1945 } 1946 1947 bbdev->dev_ops = &pmd_ops; 1948 bbdev->device = &vdev->device; 1949 bbdev->data->socket_id = init_params->socket_id; 1950 bbdev->intr_handle = NULL; 1951 1952 /* register rx/tx burst functions for data path */ 1953 bbdev->dequeue_enc_ops = dequeue_enc_ops; 1954 bbdev->dequeue_dec_ops = dequeue_dec_ops; 1955 bbdev->enqueue_enc_ops = enqueue_enc_ops; 1956 bbdev->enqueue_dec_ops = enqueue_dec_ops; 1957 bbdev->dequeue_ldpc_enc_ops = dequeue_enc_ops; 1958 bbdev->dequeue_ldpc_dec_ops = dequeue_dec_ops; 1959 bbdev->enqueue_ldpc_enc_ops = enqueue_ldpc_enc_ops; 1960 bbdev->enqueue_ldpc_dec_ops = enqueue_ldpc_dec_ops; 1961 ((struct bbdev_private *) bbdev->data->dev_private)->max_nb_queues = 1962 init_params->queues_num; 1963 1964 return 0; 1965 } 1966 1967 /* Initialise device */ 1968 static int 1969 turbo_sw_bbdev_probe(struct rte_vdev_device *vdev) 1970 { 1971 struct turbo_sw_params init_params = { 1972 rte_socket_id(), 1973 RTE_BBDEV_DEFAULT_MAX_NB_QUEUES 1974 }; 1975 const char *name; 1976 const char *input_args; 1977 1978 if (vdev == NULL) 1979 return -EINVAL; 1980 1981 name = rte_vdev_device_name(vdev); 1982 if (name == NULL) 1983 return -EINVAL; 1984 input_args = rte_vdev_device_args(vdev); 1985 parse_turbo_sw_params(&init_params, input_args); 1986 1987 rte_bbdev_log_debug( 1988 "Initialising %s on NUMA node %d with max queues: %d\n", 1989 name, init_params.socket_id, init_params.queues_num); 1990 1991 return turbo_sw_bbdev_create(vdev, &init_params); 1992 } 1993 1994 /* Uninitialise device */ 1995 static int 1996 turbo_sw_bbdev_remove(struct rte_vdev_device *vdev) 1997 { 1998 struct rte_bbdev *bbdev; 1999 const char *name; 2000 2001 if (vdev == NULL) 2002 return -EINVAL; 2003 2004 name = rte_vdev_device_name(vdev); 2005 if (name == NULL) 2006 return -EINVAL; 2007 2008 bbdev = rte_bbdev_get_named_dev(name); 2009 if (bbdev == NULL) 2010 return -EINVAL; 2011 2012 rte_free(bbdev->data->dev_private); 2013 2014 return rte_bbdev_release(bbdev); 2015 } 2016 2017 static struct rte_vdev_driver bbdev_turbo_sw_pmd_drv = { 2018 .probe = turbo_sw_bbdev_probe, 2019 .remove = turbo_sw_bbdev_remove 2020 }; 2021 2022 RTE_PMD_REGISTER_VDEV(DRIVER_NAME, bbdev_turbo_sw_pmd_drv); 2023 RTE_PMD_REGISTER_PARAM_STRING(DRIVER_NAME, 2024 TURBO_SW_MAX_NB_QUEUES_ARG"=<int> " 2025 TURBO_SW_SOCKET_ID_ARG"=<int>"); 2026 RTE_PMD_REGISTER_ALIAS(DRIVER_NAME, turbo_sw); 2027