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