1 /* SPDX-License-Identifier: BSD-3-Clause 2 * Copyright(c) 2017 Intel Corporation 3 */ 4 5 #include <string.h> 6 7 #include <rte_common.h> 8 #include <rte_bus_vdev.h> 9 #include <rte_malloc.h> 10 #include <rte_ring.h> 11 #include <rte_kvargs.h> 12 #include <rte_cycles.h> 13 14 #include <rte_bbdev.h> 15 #include <rte_bbdev_pmd.h> 16 17 #include <phy_turbo.h> 18 #include <phy_crc.h> 19 #include <phy_rate_match.h> 20 #include <divide.h> 21 22 #define DRIVER_NAME baseband_turbo_sw 23 24 /* Turbo SW PMD logging ID */ 25 static int bbdev_turbo_sw_logtype; 26 27 /* Helper macro for logging */ 28 #define rte_bbdev_log(level, fmt, ...) \ 29 rte_log(RTE_LOG_ ## level, bbdev_turbo_sw_logtype, fmt "\n", \ 30 ##__VA_ARGS__) 31 32 #define rte_bbdev_log_debug(fmt, ...) \ 33 rte_bbdev_log(DEBUG, RTE_STR(__LINE__) ":%s() " fmt, __func__, \ 34 ##__VA_ARGS__) 35 36 #define DEINT_INPUT_BUF_SIZE (((RTE_BBDEV_MAX_CB_SIZE >> 3) + 1) * 48) 37 #define DEINT_OUTPUT_BUF_SIZE (DEINT_INPUT_BUF_SIZE * 6) 38 #define ADAPTER_OUTPUT_BUF_SIZE ((RTE_BBDEV_MAX_CB_SIZE + 4) * 48) 39 40 /* private data structure */ 41 struct bbdev_private { 42 unsigned int max_nb_queues; /**< Max number of queues */ 43 }; 44 45 /* Initialisation params structure that can be used by Turbo SW driver */ 46 struct turbo_sw_params { 47 int socket_id; /*< Turbo SW device socket */ 48 uint16_t queues_num; /*< Turbo SW device queues number */ 49 }; 50 51 /* Accecptable params for Turbo SW devices */ 52 #define TURBO_SW_MAX_NB_QUEUES_ARG "max_nb_queues" 53 #define TURBO_SW_SOCKET_ID_ARG "socket_id" 54 55 static const char * const turbo_sw_valid_params[] = { 56 TURBO_SW_MAX_NB_QUEUES_ARG, 57 TURBO_SW_SOCKET_ID_ARG 58 }; 59 60 /* queue */ 61 struct turbo_sw_queue { 62 /* Ring for processed (encoded/decoded) operations which are ready to 63 * be dequeued. 64 */ 65 struct rte_ring *processed_pkts; 66 /* Stores input for turbo encoder (used when CRC attachment is 67 * performed 68 */ 69 uint8_t *enc_in; 70 /* Stores output from turbo encoder */ 71 uint8_t *enc_out; 72 /* Alpha gamma buf for bblib_turbo_decoder() function */ 73 int8_t *ag; 74 /* Temp buf for bblib_turbo_decoder() function */ 75 uint16_t *code_block; 76 /* Input buf for bblib_rate_dematching_lte() function */ 77 uint8_t *deint_input; 78 /* Output buf for bblib_rate_dematching_lte() function */ 79 uint8_t *deint_output; 80 /* Output buf for bblib_turbodec_adapter_lte() function */ 81 uint8_t *adapter_output; 82 /* Operation type of this queue */ 83 enum rte_bbdev_op_type type; 84 } __rte_cache_aligned; 85 86 static inline char * 87 mbuf_append(struct rte_mbuf *m_head, struct rte_mbuf *m, uint16_t len) 88 { 89 if (unlikely(len > rte_pktmbuf_tailroom(m))) 90 return NULL; 91 92 char *tail = (char *)m->buf_addr + m->data_off + m->data_len; 93 m->data_len = (uint16_t)(m->data_len + len); 94 m_head->pkt_len = (m_head->pkt_len + len); 95 return tail; 96 } 97 98 /* Calculate index based on Table 5.1.3-3 from TS34.212 */ 99 static inline int32_t 100 compute_idx(uint16_t k) 101 { 102 int32_t result = 0; 103 104 if (k < RTE_BBDEV_MIN_CB_SIZE || k > RTE_BBDEV_MAX_CB_SIZE) 105 return -1; 106 107 if (k > 2048) { 108 if ((k - 2048) % 64 != 0) 109 result = -1; 110 111 result = 124 + (k - 2048) / 64; 112 } else if (k <= 512) { 113 if ((k - 40) % 8 != 0) 114 result = -1; 115 116 result = (k - 40) / 8 + 1; 117 } else if (k <= 1024) { 118 if ((k - 512) % 16 != 0) 119 result = -1; 120 121 result = 60 + (k - 512) / 16; 122 } else { /* 1024 < k <= 2048 */ 123 if ((k - 1024) % 32 != 0) 124 result = -1; 125 126 result = 92 + (k - 1024) / 32; 127 } 128 129 return result; 130 } 131 132 /* Read flag value 0/1 from bitmap */ 133 static inline bool 134 check_bit(uint32_t bitmap, uint32_t bitmask) 135 { 136 return bitmap & bitmask; 137 } 138 139 /* Get device info */ 140 static void 141 info_get(struct rte_bbdev *dev, struct rte_bbdev_driver_info *dev_info) 142 { 143 struct bbdev_private *internals = dev->data->dev_private; 144 145 static const struct rte_bbdev_op_cap bbdev_capabilities[] = { 146 { 147 .type = RTE_BBDEV_OP_TURBO_DEC, 148 .cap.turbo_dec = { 149 .capability_flags = 150 RTE_BBDEV_TURBO_SUBBLOCK_DEINTERLEAVE | 151 RTE_BBDEV_TURBO_POS_LLR_1_BIT_IN | 152 RTE_BBDEV_TURBO_NEG_LLR_1_BIT_IN | 153 RTE_BBDEV_TURBO_CRC_TYPE_24B | 154 RTE_BBDEV_TURBO_DEC_TB_CRC_24B_KEEP | 155 RTE_BBDEV_TURBO_EARLY_TERMINATION, 156 .max_llr_modulus = 16, 157 .num_buffers_src = RTE_BBDEV_MAX_CODE_BLOCKS, 158 .num_buffers_hard_out = 159 RTE_BBDEV_MAX_CODE_BLOCKS, 160 .num_buffers_soft_out = 0, 161 } 162 }, 163 { 164 .type = RTE_BBDEV_OP_TURBO_ENC, 165 .cap.turbo_enc = { 166 .capability_flags = 167 RTE_BBDEV_TURBO_CRC_24B_ATTACH | 168 RTE_BBDEV_TURBO_CRC_24A_ATTACH | 169 RTE_BBDEV_TURBO_RATE_MATCH | 170 RTE_BBDEV_TURBO_RV_INDEX_BYPASS, 171 .num_buffers_src = RTE_BBDEV_MAX_CODE_BLOCKS, 172 .num_buffers_dst = RTE_BBDEV_MAX_CODE_BLOCKS, 173 } 174 }, 175 RTE_BBDEV_END_OF_CAPABILITIES_LIST() 176 }; 177 178 static struct rte_bbdev_queue_conf default_queue_conf = { 179 .queue_size = RTE_BBDEV_QUEUE_SIZE_LIMIT, 180 }; 181 182 static const enum rte_cpu_flag_t cpu_flag = RTE_CPUFLAG_SSE4_2; 183 184 default_queue_conf.socket = dev->data->socket_id; 185 186 dev_info->driver_name = RTE_STR(DRIVER_NAME); 187 dev_info->max_num_queues = internals->max_nb_queues; 188 dev_info->queue_size_lim = RTE_BBDEV_QUEUE_SIZE_LIMIT; 189 dev_info->hardware_accelerated = false; 190 dev_info->max_dl_queue_priority = 0; 191 dev_info->max_ul_queue_priority = 0; 192 dev_info->default_queue_conf = default_queue_conf; 193 dev_info->capabilities = bbdev_capabilities; 194 dev_info->cpu_flag_reqs = &cpu_flag; 195 dev_info->min_alignment = 64; 196 197 rte_bbdev_log_debug("got device info from %u\n", dev->data->dev_id); 198 } 199 200 /* Release queue */ 201 static int 202 q_release(struct rte_bbdev *dev, uint16_t q_id) 203 { 204 struct turbo_sw_queue *q = dev->data->queues[q_id].queue_private; 205 206 if (q != NULL) { 207 rte_ring_free(q->processed_pkts); 208 rte_free(q->enc_out); 209 rte_free(q->enc_in); 210 rte_free(q->ag); 211 rte_free(q->code_block); 212 rte_free(q->deint_input); 213 rte_free(q->deint_output); 214 rte_free(q->adapter_output); 215 rte_free(q); 216 dev->data->queues[q_id].queue_private = NULL; 217 } 218 219 rte_bbdev_log_debug("released device queue %u:%u", 220 dev->data->dev_id, q_id); 221 return 0; 222 } 223 224 /* Setup a queue */ 225 static int 226 q_setup(struct rte_bbdev *dev, uint16_t q_id, 227 const struct rte_bbdev_queue_conf *queue_conf) 228 { 229 int ret; 230 struct turbo_sw_queue *q; 231 char name[RTE_RING_NAMESIZE]; 232 233 /* Allocate the queue data structure. */ 234 q = rte_zmalloc_socket(RTE_STR(DRIVER_NAME), sizeof(*q), 235 RTE_CACHE_LINE_SIZE, queue_conf->socket); 236 if (q == NULL) { 237 rte_bbdev_log(ERR, "Failed to allocate queue memory"); 238 return -ENOMEM; 239 } 240 241 /* Allocate memory for encoder output. */ 242 ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"_enc_o%u:%u", 243 dev->data->dev_id, q_id); 244 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) { 245 rte_bbdev_log(ERR, 246 "Creating queue name for device %u queue %u failed", 247 dev->data->dev_id, q_id); 248 return -ENAMETOOLONG; 249 } 250 q->enc_out = rte_zmalloc_socket(name, 251 ((RTE_BBDEV_MAX_TB_SIZE >> 3) + 3) * 252 sizeof(*q->enc_out) * 3, 253 RTE_CACHE_LINE_SIZE, queue_conf->socket); 254 if (q->enc_out == NULL) { 255 rte_bbdev_log(ERR, 256 "Failed to allocate queue memory for %s", name); 257 goto free_q; 258 } 259 260 /* Allocate memory for rate matching output. */ 261 ret = snprintf(name, RTE_RING_NAMESIZE, 262 RTE_STR(DRIVER_NAME)"_enc_i%u:%u", dev->data->dev_id, 263 q_id); 264 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) { 265 rte_bbdev_log(ERR, 266 "Creating queue name for device %u queue %u failed", 267 dev->data->dev_id, q_id); 268 return -ENAMETOOLONG; 269 } 270 q->enc_in = rte_zmalloc_socket(name, 271 (RTE_BBDEV_MAX_CB_SIZE >> 3) * sizeof(*q->enc_in), 272 RTE_CACHE_LINE_SIZE, queue_conf->socket); 273 if (q->enc_in == NULL) { 274 rte_bbdev_log(ERR, 275 "Failed to allocate queue memory for %s", name); 276 goto free_q; 277 } 278 279 /* Allocate memory for Aplha Gamma temp buffer. */ 280 ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"_ag%u:%u", 281 dev->data->dev_id, q_id); 282 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) { 283 rte_bbdev_log(ERR, 284 "Creating queue name for device %u queue %u failed", 285 dev->data->dev_id, q_id); 286 return -ENAMETOOLONG; 287 } 288 q->ag = rte_zmalloc_socket(name, 289 RTE_BBDEV_MAX_CB_SIZE * 10 * sizeof(*q->ag), 290 RTE_CACHE_LINE_SIZE, queue_conf->socket); 291 if (q->ag == NULL) { 292 rte_bbdev_log(ERR, 293 "Failed to allocate queue memory for %s", name); 294 goto free_q; 295 } 296 297 /* Allocate memory for code block temp buffer. */ 298 ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"_cb%u:%u", 299 dev->data->dev_id, q_id); 300 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) { 301 rte_bbdev_log(ERR, 302 "Creating queue name for device %u queue %u failed", 303 dev->data->dev_id, q_id); 304 return -ENAMETOOLONG; 305 } 306 q->code_block = rte_zmalloc_socket(name, 307 RTE_BBDEV_MAX_CB_SIZE * sizeof(*q->code_block), 308 RTE_CACHE_LINE_SIZE, queue_conf->socket); 309 if (q->code_block == NULL) { 310 rte_bbdev_log(ERR, 311 "Failed to allocate queue memory for %s", name); 312 goto free_q; 313 } 314 315 /* Allocate memory for Deinterleaver input. */ 316 ret = snprintf(name, RTE_RING_NAMESIZE, 317 RTE_STR(DRIVER_NAME)"_de_i%u:%u", 318 dev->data->dev_id, q_id); 319 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) { 320 rte_bbdev_log(ERR, 321 "Creating queue name for device %u queue %u failed", 322 dev->data->dev_id, q_id); 323 return -ENAMETOOLONG; 324 } 325 q->deint_input = rte_zmalloc_socket(name, 326 DEINT_INPUT_BUF_SIZE * sizeof(*q->deint_input), 327 RTE_CACHE_LINE_SIZE, queue_conf->socket); 328 if (q->deint_input == NULL) { 329 rte_bbdev_log(ERR, 330 "Failed to allocate queue memory for %s", name); 331 goto free_q; 332 } 333 334 /* Allocate memory for Deinterleaver output. */ 335 ret = snprintf(name, RTE_RING_NAMESIZE, 336 RTE_STR(DRIVER_NAME)"_de_o%u:%u", 337 dev->data->dev_id, q_id); 338 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) { 339 rte_bbdev_log(ERR, 340 "Creating queue name for device %u queue %u failed", 341 dev->data->dev_id, q_id); 342 return -ENAMETOOLONG; 343 } 344 q->deint_output = rte_zmalloc_socket(NULL, 345 DEINT_OUTPUT_BUF_SIZE * sizeof(*q->deint_output), 346 RTE_CACHE_LINE_SIZE, queue_conf->socket); 347 if (q->deint_output == NULL) { 348 rte_bbdev_log(ERR, 349 "Failed to allocate queue memory for %s", name); 350 goto free_q; 351 } 352 353 /* Allocate memory for Adapter output. */ 354 ret = snprintf(name, RTE_RING_NAMESIZE, 355 RTE_STR(DRIVER_NAME)"_ada_o%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->adapter_output = rte_zmalloc_socket(NULL, 364 ADAPTER_OUTPUT_BUF_SIZE * sizeof(*q->adapter_output), 365 RTE_CACHE_LINE_SIZE, queue_conf->socket); 366 if (q->adapter_output == NULL) { 367 rte_bbdev_log(ERR, 368 "Failed to allocate queue memory for %s", name); 369 goto free_q; 370 } 371 372 /* Create ring for packets awaiting to be dequeued. */ 373 ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"%u:%u", 374 dev->data->dev_id, q_id); 375 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) { 376 rte_bbdev_log(ERR, 377 "Creating queue name for device %u queue %u failed", 378 dev->data->dev_id, q_id); 379 return -ENAMETOOLONG; 380 } 381 q->processed_pkts = rte_ring_create(name, queue_conf->queue_size, 382 queue_conf->socket, RING_F_SP_ENQ | RING_F_SC_DEQ); 383 if (q->processed_pkts == NULL) { 384 rte_bbdev_log(ERR, "Failed to create ring for %s", name); 385 goto free_q; 386 } 387 388 q->type = queue_conf->op_type; 389 390 dev->data->queues[q_id].queue_private = q; 391 rte_bbdev_log_debug("setup device queue %s", name); 392 return 0; 393 394 free_q: 395 rte_ring_free(q->processed_pkts); 396 rte_free(q->enc_out); 397 rte_free(q->enc_in); 398 rte_free(q->ag); 399 rte_free(q->code_block); 400 rte_free(q->deint_input); 401 rte_free(q->deint_output); 402 rte_free(q->adapter_output); 403 rte_free(q); 404 return -EFAULT; 405 } 406 407 static const struct rte_bbdev_ops pmd_ops = { 408 .info_get = info_get, 409 .queue_setup = q_setup, 410 .queue_release = q_release 411 }; 412 413 /* Checks if the encoder input buffer is correct. 414 * Returns 0 if it's valid, -1 otherwise. 415 */ 416 static inline int 417 is_enc_input_valid(const uint16_t k, const int32_t k_idx, 418 const uint16_t in_length) 419 { 420 if (k_idx < 0) { 421 rte_bbdev_log(ERR, "K Index is invalid"); 422 return -1; 423 } 424 425 if (in_length - (k >> 3) < 0) { 426 rte_bbdev_log(ERR, 427 "Mismatch between input length (%u bytes) and K (%u bits)", 428 in_length, k); 429 return -1; 430 } 431 432 if (k > RTE_BBDEV_MAX_CB_SIZE) { 433 rte_bbdev_log(ERR, "CB size (%u) is too big, max: %d", 434 k, RTE_BBDEV_MAX_CB_SIZE); 435 return -1; 436 } 437 438 return 0; 439 } 440 441 /* Checks if the decoder input buffer is correct. 442 * Returns 0 if it's valid, -1 otherwise. 443 */ 444 static inline int 445 is_dec_input_valid(int32_t k_idx, int16_t kw, int16_t in_length) 446 { 447 if (k_idx < 0) { 448 rte_bbdev_log(ERR, "K index is invalid"); 449 return -1; 450 } 451 452 if (in_length < kw) { 453 rte_bbdev_log(ERR, 454 "Mismatch between input length (%u) and kw (%u)", 455 in_length, kw); 456 return -1; 457 } 458 459 if (kw > RTE_BBDEV_MAX_KW) { 460 rte_bbdev_log(ERR, "Input length (%u) is too big, max: %d", 461 kw, RTE_BBDEV_MAX_KW); 462 return -1; 463 } 464 465 return 0; 466 } 467 468 static inline void 469 process_enc_cb(struct turbo_sw_queue *q, struct rte_bbdev_enc_op *op, 470 uint8_t r, uint8_t c, uint16_t k, uint16_t ncb, 471 uint32_t e, struct rte_mbuf *m_in, struct rte_mbuf *m_out_head, 472 struct rte_mbuf *m_out, uint16_t in_offset, uint16_t out_offset, 473 uint16_t in_length, struct rte_bbdev_stats *q_stats) 474 { 475 int ret; 476 int16_t k_idx; 477 uint16_t m; 478 uint8_t *in, *out0, *out1, *out2, *tmp_out, *rm_out; 479 uint64_t first_3_bytes = 0; 480 struct rte_bbdev_op_turbo_enc *enc = &op->turbo_enc; 481 struct bblib_crc_request crc_req; 482 struct bblib_crc_response crc_resp; 483 struct bblib_turbo_encoder_request turbo_req; 484 struct bblib_turbo_encoder_response turbo_resp; 485 struct bblib_rate_match_dl_request rm_req; 486 struct bblib_rate_match_dl_response rm_resp; 487 #ifdef RTE_BBDEV_OFFLOAD_COST 488 uint64_t start_time; 489 #else 490 RTE_SET_USED(q_stats); 491 #endif 492 493 k_idx = compute_idx(k); 494 in = rte_pktmbuf_mtod_offset(m_in, uint8_t *, in_offset); 495 496 /* CRC24A (for TB) */ 497 if ((enc->op_flags & RTE_BBDEV_TURBO_CRC_24A_ATTACH) && 498 (enc->code_block_mode == 1)) { 499 ret = is_enc_input_valid(k - 24, k_idx, in_length); 500 if (ret != 0) { 501 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 502 return; 503 } 504 crc_req.data = in; 505 crc_req.len = k - 24; 506 /* Check if there is a room for CRC bits if not use 507 * the temporary buffer. 508 */ 509 if (mbuf_append(m_in, m_in, 3) == NULL) { 510 rte_memcpy(q->enc_in, in, (k - 24) >> 3); 511 in = q->enc_in; 512 } else { 513 /* Store 3 first bytes of next CB as they will be 514 * overwritten by CRC bytes. If it is the last CB then 515 * there is no point to store 3 next bytes and this 516 * if..else branch will be omitted. 517 */ 518 first_3_bytes = *((uint64_t *)&in[(k - 32) >> 3]); 519 } 520 521 crc_resp.data = in; 522 #ifdef RTE_BBDEV_OFFLOAD_COST 523 start_time = rte_rdtsc_precise(); 524 #endif 525 /* CRC24A generation */ 526 bblib_lte_crc24a_gen(&crc_req, &crc_resp); 527 #ifdef RTE_BBDEV_OFFLOAD_COST 528 q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time; 529 #endif 530 } else if (enc->op_flags & RTE_BBDEV_TURBO_CRC_24B_ATTACH) { 531 /* CRC24B */ 532 ret = is_enc_input_valid(k - 24, k_idx, in_length); 533 if (ret != 0) { 534 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 535 return; 536 } 537 crc_req.data = in; 538 crc_req.len = k - 24; 539 /* Check if there is a room for CRC bits if this is the last 540 * CB in TB. If not use temporary buffer. 541 */ 542 if ((c - r == 1) && (mbuf_append(m_in, m_in, 3) == NULL)) { 543 rte_memcpy(q->enc_in, in, (k - 24) >> 3); 544 in = q->enc_in; 545 } else if (c - r > 1) { 546 /* Store 3 first bytes of next CB as they will be 547 * overwritten by CRC bytes. If it is the last CB then 548 * there is no point to store 3 next bytes and this 549 * if..else branch will be omitted. 550 */ 551 first_3_bytes = *((uint64_t *)&in[(k - 32) >> 3]); 552 } 553 554 crc_resp.data = in; 555 #ifdef RTE_BBDEV_OFFLOAD_COST 556 start_time = rte_rdtsc_precise(); 557 #endif 558 /* CRC24B generation */ 559 bblib_lte_crc24b_gen(&crc_req, &crc_resp); 560 #ifdef RTE_BBDEV_OFFLOAD_COST 561 q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time; 562 #endif 563 } else { 564 ret = is_enc_input_valid(k, k_idx, in_length); 565 if (ret != 0) { 566 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 567 return; 568 } 569 } 570 571 /* Turbo encoder */ 572 573 /* Each bit layer output from turbo encoder is (k+4) bits long, i.e. 574 * input length + 4 tail bits. That's (k/8) + 1 bytes after rounding up. 575 * So dst_data's length should be 3*(k/8) + 3 bytes. 576 * In Rate-matching bypass case outputs pointers passed to encoder 577 * (out0, out1 and out2) can directly point to addresses of output from 578 * turbo_enc entity. 579 */ 580 if (enc->op_flags & RTE_BBDEV_TURBO_RATE_MATCH) { 581 out0 = q->enc_out; 582 out1 = RTE_PTR_ADD(out0, (k >> 3) + 1); 583 out2 = RTE_PTR_ADD(out1, (k >> 3) + 1); 584 } else { 585 out0 = (uint8_t *)mbuf_append(m_out_head, m_out, 586 (k >> 3) * 3 + 2); 587 if (out0 == NULL) { 588 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 589 rte_bbdev_log(ERR, 590 "Too little space in output mbuf"); 591 return; 592 } 593 enc->output.length += (k >> 3) * 3 + 2; 594 /* rte_bbdev_op_data.offset can be different than the 595 * offset of the appended bytes 596 */ 597 out0 = rte_pktmbuf_mtod_offset(m_out, uint8_t *, out_offset); 598 out1 = rte_pktmbuf_mtod_offset(m_out, uint8_t *, 599 out_offset + (k >> 3) + 1); 600 out2 = rte_pktmbuf_mtod_offset(m_out, uint8_t *, 601 out_offset + 2 * ((k >> 3) + 1)); 602 } 603 604 turbo_req.case_id = k_idx; 605 turbo_req.input_win = in; 606 turbo_req.length = k >> 3; 607 turbo_resp.output_win_0 = out0; 608 turbo_resp.output_win_1 = out1; 609 turbo_resp.output_win_2 = out2; 610 611 #ifdef RTE_BBDEV_OFFLOAD_COST 612 start_time = rte_rdtsc_precise(); 613 #endif 614 /* Turbo encoding */ 615 if (bblib_turbo_encoder(&turbo_req, &turbo_resp) != 0) { 616 op->status |= 1 << RTE_BBDEV_DRV_ERROR; 617 rte_bbdev_log(ERR, "Turbo Encoder failed"); 618 return; 619 } 620 #ifdef RTE_BBDEV_OFFLOAD_COST 621 q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time; 622 #endif 623 624 /* Restore 3 first bytes of next CB if they were overwritten by CRC*/ 625 if (first_3_bytes != 0) 626 *((uint64_t *)&in[(k - 32) >> 3]) = first_3_bytes; 627 628 /* Rate-matching */ 629 if (enc->op_flags & RTE_BBDEV_TURBO_RATE_MATCH) { 630 uint8_t mask_id; 631 /* Integer round up division by 8 */ 632 uint16_t out_len = (e + 7) >> 3; 633 /* The mask array is indexed using E%8. E is an even number so 634 * there are only 4 possible values. 635 */ 636 const uint8_t mask_out[] = {0xFF, 0xC0, 0xF0, 0xFC}; 637 638 /* get output data starting address */ 639 rm_out = (uint8_t *)mbuf_append(m_out_head, m_out, out_len); 640 if (rm_out == NULL) { 641 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 642 rte_bbdev_log(ERR, 643 "Too little space in output mbuf"); 644 return; 645 } 646 /* rte_bbdev_op_data.offset can be different than the offset 647 * of the appended bytes 648 */ 649 rm_out = rte_pktmbuf_mtod_offset(m_out, uint8_t *, out_offset); 650 651 /* index of current code block */ 652 rm_req.r = r; 653 /* total number of code block */ 654 rm_req.C = c; 655 /* For DL - 1, UL - 0 */ 656 rm_req.direction = 1; 657 /* According to 3ggp 36.212 Spec 5.1.4.1.2 section Nsoft, KMIMO 658 * and MDL_HARQ are used for Ncb calculation. As Ncb is already 659 * known we can adjust those parameters 660 */ 661 rm_req.Nsoft = ncb * rm_req.C; 662 rm_req.KMIMO = 1; 663 rm_req.MDL_HARQ = 1; 664 /* According to 3ggp 36.212 Spec 5.1.4.1.2 section Nl, Qm and G 665 * are used for E calculation. As E is already known we can 666 * adjust those parameters 667 */ 668 rm_req.NL = e; 669 rm_req.Qm = 1; 670 rm_req.G = rm_req.NL * rm_req.Qm * rm_req.C; 671 672 rm_req.rvidx = enc->rv_index; 673 rm_req.Kidx = k_idx - 1; 674 rm_req.nLen = k + 4; 675 rm_req.tin0 = out0; 676 rm_req.tin1 = out1; 677 rm_req.tin2 = out2; 678 rm_resp.output = rm_out; 679 rm_resp.OutputLen = out_len; 680 if (enc->op_flags & RTE_BBDEV_TURBO_RV_INDEX_BYPASS) 681 rm_req.bypass_rvidx = 1; 682 else 683 rm_req.bypass_rvidx = 0; 684 685 #ifdef RTE_BBDEV_OFFLOAD_COST 686 start_time = rte_rdtsc_precise(); 687 #endif 688 /* Rate-Matching */ 689 if (bblib_rate_match_dl(&rm_req, &rm_resp) != 0) { 690 op->status |= 1 << RTE_BBDEV_DRV_ERROR; 691 rte_bbdev_log(ERR, "Rate matching failed"); 692 return; 693 } 694 #ifdef RTE_BBDEV_OFFLOAD_COST 695 q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time; 696 #endif 697 698 /* SW fills an entire last byte even if E%8 != 0. Clear the 699 * superfluous data bits for consistency with HW device. 700 */ 701 mask_id = (e & 7) >> 1; 702 rm_out[out_len - 1] &= mask_out[mask_id]; 703 enc->output.length += rm_resp.OutputLen; 704 } else { 705 /* Rate matching is bypassed */ 706 707 /* Completing last byte of out0 (where 4 tail bits are stored) 708 * by moving first 4 bits from out1 709 */ 710 tmp_out = (uint8_t *) --out1; 711 *tmp_out = *tmp_out | ((*(tmp_out + 1) & 0xF0) >> 4); 712 tmp_out++; 713 /* Shifting out1 data by 4 bits to the left */ 714 for (m = 0; m < k >> 3; ++m) { 715 uint8_t *first = tmp_out; 716 uint8_t second = *(tmp_out + 1); 717 *first = (*first << 4) | ((second & 0xF0) >> 4); 718 tmp_out++; 719 } 720 /* Shifting out2 data by 8 bits to the left */ 721 for (m = 0; m < (k >> 3) + 1; ++m) { 722 *tmp_out = *(tmp_out + 1); 723 tmp_out++; 724 } 725 *tmp_out = 0; 726 } 727 } 728 729 static inline void 730 enqueue_enc_one_op(struct turbo_sw_queue *q, struct rte_bbdev_enc_op *op, 731 struct rte_bbdev_stats *queue_stats) 732 { 733 uint8_t c, r, crc24_bits = 0; 734 uint16_t k, ncb; 735 uint32_t e; 736 struct rte_bbdev_op_turbo_enc *enc = &op->turbo_enc; 737 uint16_t in_offset = enc->input.offset; 738 uint16_t out_offset = enc->output.offset; 739 struct rte_mbuf *m_in = enc->input.data; 740 struct rte_mbuf *m_out = enc->output.data; 741 struct rte_mbuf *m_out_head = enc->output.data; 742 uint32_t in_length, mbuf_total_left = enc->input.length; 743 uint16_t seg_total_left; 744 745 /* Clear op status */ 746 op->status = 0; 747 748 if (mbuf_total_left > RTE_BBDEV_MAX_TB_SIZE >> 3) { 749 rte_bbdev_log(ERR, "TB size (%u) is too big, max: %d", 750 mbuf_total_left, RTE_BBDEV_MAX_TB_SIZE); 751 op->status = 1 << RTE_BBDEV_DATA_ERROR; 752 return; 753 } 754 755 if (m_in == NULL || m_out == NULL) { 756 rte_bbdev_log(ERR, "Invalid mbuf pointer"); 757 op->status = 1 << RTE_BBDEV_DATA_ERROR; 758 return; 759 } 760 761 if ((enc->op_flags & RTE_BBDEV_TURBO_CRC_24B_ATTACH) || 762 (enc->op_flags & RTE_BBDEV_TURBO_CRC_24A_ATTACH)) 763 crc24_bits = 24; 764 765 if (enc->code_block_mode == 0) { /* For Transport Block mode */ 766 c = enc->tb_params.c; 767 r = enc->tb_params.r; 768 } else {/* For Code Block mode */ 769 c = 1; 770 r = 0; 771 } 772 773 while (mbuf_total_left > 0 && r < c) { 774 775 seg_total_left = rte_pktmbuf_data_len(m_in) - in_offset; 776 777 if (enc->code_block_mode == 0) { 778 k = (r < enc->tb_params.c_neg) ? 779 enc->tb_params.k_neg : enc->tb_params.k_pos; 780 ncb = (r < enc->tb_params.c_neg) ? 781 enc->tb_params.ncb_neg : enc->tb_params.ncb_pos; 782 e = (r < enc->tb_params.cab) ? 783 enc->tb_params.ea : enc->tb_params.eb; 784 } else { 785 k = enc->cb_params.k; 786 ncb = enc->cb_params.ncb; 787 e = enc->cb_params.e; 788 } 789 790 process_enc_cb(q, op, r, c, k, ncb, e, m_in, m_out_head, 791 m_out, in_offset, out_offset, seg_total_left, 792 queue_stats); 793 /* Update total_left */ 794 in_length = ((k - crc24_bits) >> 3); 795 mbuf_total_left -= in_length; 796 /* Update offsets for next CBs (if exist) */ 797 in_offset += (k - crc24_bits) >> 3; 798 if (enc->op_flags & RTE_BBDEV_TURBO_RATE_MATCH) 799 out_offset += e >> 3; 800 else 801 out_offset += (k >> 3) * 3 + 2; 802 803 /* Update offsets */ 804 if (seg_total_left == in_length) { 805 /* Go to the next mbuf */ 806 m_in = m_in->next; 807 m_out = m_out->next; 808 in_offset = 0; 809 out_offset = 0; 810 } 811 r++; 812 } 813 814 /* check if all input data was processed */ 815 if (mbuf_total_left != 0) { 816 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 817 rte_bbdev_log(ERR, 818 "Mismatch between mbuf length and included CBs sizes"); 819 } 820 } 821 822 static inline uint16_t 823 enqueue_enc_all_ops(struct turbo_sw_queue *q, struct rte_bbdev_enc_op **ops, 824 uint16_t nb_ops, struct rte_bbdev_stats *queue_stats) 825 { 826 uint16_t i; 827 #ifdef RTE_BBDEV_OFFLOAD_COST 828 queue_stats->acc_offload_cycles = 0; 829 #endif 830 831 for (i = 0; i < nb_ops; ++i) 832 enqueue_enc_one_op(q, ops[i], queue_stats); 833 834 return rte_ring_enqueue_burst(q->processed_pkts, (void **)ops, nb_ops, 835 NULL); 836 } 837 838 /* Remove the padding bytes from a cyclic buffer. 839 * The input buffer is a data stream wk as described in 3GPP TS 36.212 section 840 * 5.1.4.1.2 starting from w0 and with length Ncb bytes. 841 * The output buffer is a data stream wk with pruned padding bytes. It's length 842 * is 3*D bytes and the order of non-padding bytes is preserved. 843 */ 844 static inline void 845 remove_nulls_from_circular_buf(const uint8_t *in, uint8_t *out, uint16_t k, 846 uint16_t ncb) 847 { 848 uint32_t in_idx, out_idx, c_idx; 849 const uint32_t d = k + 4; 850 const uint32_t kw = (ncb / 3); 851 const uint32_t nd = kw - d; 852 const uint32_t r_subblock = kw / RTE_BBDEV_C_SUBBLOCK; 853 /* Inter-column permutation pattern */ 854 const uint32_t P[RTE_BBDEV_C_SUBBLOCK] = {0, 16, 8, 24, 4, 20, 12, 28, 855 2, 18, 10, 26, 6, 22, 14, 30, 1, 17, 9, 25, 5, 21, 13, 856 29, 3, 19, 11, 27, 7, 23, 15, 31}; 857 in_idx = 0; 858 out_idx = 0; 859 860 /* The padding bytes are at the first Nd positions in the first row. */ 861 for (c_idx = 0; in_idx < kw; in_idx += r_subblock, ++c_idx) { 862 if (P[c_idx] < nd) { 863 rte_memcpy(&out[out_idx], &in[in_idx + 1], 864 r_subblock - 1); 865 out_idx += r_subblock - 1; 866 } else { 867 rte_memcpy(&out[out_idx], &in[in_idx], r_subblock); 868 out_idx += r_subblock; 869 } 870 } 871 872 /* First and second parity bits sub-blocks are interlaced. */ 873 for (c_idx = 0; in_idx < ncb - 2 * r_subblock; 874 in_idx += 2 * r_subblock, ++c_idx) { 875 uint32_t second_block_c_idx = P[c_idx]; 876 uint32_t third_block_c_idx = P[c_idx] + 1; 877 878 if (second_block_c_idx < nd && third_block_c_idx < nd) { 879 rte_memcpy(&out[out_idx], &in[in_idx + 2], 880 2 * r_subblock - 2); 881 out_idx += 2 * r_subblock - 2; 882 } else if (second_block_c_idx >= nd && 883 third_block_c_idx >= nd) { 884 rte_memcpy(&out[out_idx], &in[in_idx], 2 * r_subblock); 885 out_idx += 2 * r_subblock; 886 } else if (second_block_c_idx < nd) { 887 out[out_idx++] = in[in_idx]; 888 rte_memcpy(&out[out_idx], &in[in_idx + 2], 889 2 * r_subblock - 2); 890 out_idx += 2 * r_subblock - 2; 891 } else { 892 rte_memcpy(&out[out_idx], &in[in_idx + 1], 893 2 * r_subblock - 1); 894 out_idx += 2 * r_subblock - 1; 895 } 896 } 897 898 /* Last interlaced row is different - its last byte is the only padding 899 * byte. We can have from 4 up to 28 padding bytes (Nd) per sub-block. 900 * After interlacing the 1st and 2nd parity sub-blocks we can have 0, 1 901 * or 2 padding bytes each time we make a step of 2 * R_SUBBLOCK bytes 902 * (moving to another column). 2nd parity sub-block uses the same 903 * inter-column permutation pattern as the systematic and 1st parity 904 * sub-blocks but it adds '1' to the resulting index and calculates the 905 * modulus of the result and Kw. Last column is mapped to itself (id 31) 906 * so the first byte taken from the 2nd parity sub-block will be the 907 * 32nd (31+1) byte, then 64th etc. (step is C_SUBBLOCK == 32) and the 908 * last byte will be the first byte from the sub-block: 909 * (32 + 32 * (R_SUBBLOCK-1)) % Kw == Kw % Kw == 0. Nd can't be smaller 910 * than 4 so we know that bytes with ids 0, 1, 2 and 3 must be the 911 * padding bytes. The bytes from the 1st parity sub-block are the bytes 912 * from the 31st column - Nd can't be greater than 28 so we are sure 913 * that there are no padding bytes in 31st column. 914 */ 915 rte_memcpy(&out[out_idx], &in[in_idx], 2 * r_subblock - 1); 916 } 917 918 static inline void 919 move_padding_bytes(const uint8_t *in, uint8_t *out, uint16_t k, 920 uint16_t ncb) 921 { 922 uint16_t d = k + 4; 923 uint16_t kpi = ncb / 3; 924 uint16_t nd = kpi - d; 925 926 rte_memcpy(&out[nd], in, d); 927 rte_memcpy(&out[nd + kpi + 64], &in[kpi], d); 928 rte_memcpy(&out[(nd - 1) + 2 * (kpi + 64)], &in[2 * kpi], d); 929 } 930 931 static inline void 932 process_dec_cb(struct turbo_sw_queue *q, struct rte_bbdev_dec_op *op, 933 uint8_t c, uint16_t k, uint16_t kw, struct rte_mbuf *m_in, 934 struct rte_mbuf *m_out_head, struct rte_mbuf *m_out, 935 uint16_t in_offset, uint16_t out_offset, bool check_crc_24b, 936 uint16_t crc24_overlap, uint16_t in_length, 937 struct rte_bbdev_stats *q_stats) 938 { 939 int ret; 940 int32_t k_idx; 941 int32_t iter_cnt; 942 uint8_t *in, *out, *adapter_input; 943 int32_t ncb, ncb_without_null; 944 struct bblib_turbo_adapter_ul_response adapter_resp; 945 struct bblib_turbo_adapter_ul_request adapter_req; 946 struct bblib_turbo_decoder_request turbo_req; 947 struct bblib_turbo_decoder_response turbo_resp; 948 struct rte_bbdev_op_turbo_dec *dec = &op->turbo_dec; 949 #ifdef RTE_BBDEV_OFFLOAD_COST 950 uint64_t start_time; 951 #else 952 RTE_SET_USED(q_stats); 953 #endif 954 955 k_idx = compute_idx(k); 956 957 ret = is_dec_input_valid(k_idx, kw, in_length); 958 if (ret != 0) { 959 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 960 return; 961 } 962 963 in = rte_pktmbuf_mtod_offset(m_in, uint8_t *, in_offset); 964 ncb = kw; 965 ncb_without_null = (k + 4) * 3; 966 967 if (check_bit(dec->op_flags, RTE_BBDEV_TURBO_SUBBLOCK_DEINTERLEAVE)) { 968 struct bblib_deinterleave_ul_request deint_req; 969 struct bblib_deinterleave_ul_response deint_resp; 970 971 /* SW decoder accepts only a circular buffer without NULL bytes 972 * so the input needs to be converted. 973 */ 974 remove_nulls_from_circular_buf(in, q->deint_input, k, ncb); 975 976 deint_req.pharqbuffer = q->deint_input; 977 deint_req.ncb = ncb_without_null; 978 deint_resp.pinteleavebuffer = q->deint_output; 979 980 #ifdef RTE_BBDEV_OFFLOAD_COST 981 start_time = rte_rdtsc_precise(); 982 #endif 983 bblib_deinterleave_ul(&deint_req, &deint_resp); 984 #ifdef RTE_BBDEV_OFFLOAD_COST 985 q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time; 986 #endif 987 } else 988 move_padding_bytes(in, q->deint_output, k, ncb); 989 990 adapter_input = q->deint_output; 991 992 if (dec->op_flags & RTE_BBDEV_TURBO_POS_LLR_1_BIT_IN) 993 adapter_req.isinverted = 1; 994 else if (dec->op_flags & RTE_BBDEV_TURBO_NEG_LLR_1_BIT_IN) 995 adapter_req.isinverted = 0; 996 else { 997 op->status |= 1 << RTE_BBDEV_DRV_ERROR; 998 rte_bbdev_log(ERR, "LLR format wasn't specified"); 999 return; 1000 } 1001 1002 adapter_req.ncb = ncb_without_null; 1003 adapter_req.pinteleavebuffer = adapter_input; 1004 adapter_resp.pharqout = q->adapter_output; 1005 1006 #ifdef RTE_BBDEV_OFFLOAD_COST 1007 start_time = rte_rdtsc_precise(); 1008 #endif 1009 /* Turbo decode adaptation */ 1010 bblib_turbo_adapter_ul(&adapter_req, &adapter_resp); 1011 #ifdef RTE_BBDEV_OFFLOAD_COST 1012 q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time; 1013 #endif 1014 1015 out = (uint8_t *)mbuf_append(m_out_head, m_out, 1016 ((k - crc24_overlap) >> 3)); 1017 if (out == NULL) { 1018 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 1019 rte_bbdev_log(ERR, "Too little space in output mbuf"); 1020 return; 1021 } 1022 /* rte_bbdev_op_data.offset can be different than the offset of the 1023 * appended bytes 1024 */ 1025 out = rte_pktmbuf_mtod_offset(m_out, uint8_t *, out_offset); 1026 if (check_crc_24b) 1027 turbo_req.c = c + 1; 1028 else 1029 turbo_req.c = c; 1030 turbo_req.input = (int8_t *)q->adapter_output; 1031 turbo_req.k = k; 1032 turbo_req.k_idx = k_idx; 1033 turbo_req.max_iter_num = dec->iter_max; 1034 turbo_req.early_term_disable = !check_bit(dec->op_flags, 1035 RTE_BBDEV_TURBO_EARLY_TERMINATION); 1036 turbo_resp.ag_buf = q->ag; 1037 turbo_resp.cb_buf = q->code_block; 1038 turbo_resp.output = out; 1039 1040 #ifdef RTE_BBDEV_OFFLOAD_COST 1041 start_time = rte_rdtsc_precise(); 1042 #endif 1043 /* Turbo decode */ 1044 iter_cnt = bblib_turbo_decoder(&turbo_req, &turbo_resp); 1045 #ifdef RTE_BBDEV_OFFLOAD_COST 1046 q_stats->acc_offload_cycles += rte_rdtsc_precise() - start_time; 1047 #endif 1048 dec->hard_output.length += (k >> 3); 1049 1050 if (iter_cnt > 0) { 1051 /* Temporary solution for returned iter_count from SDK */ 1052 iter_cnt = (iter_cnt - 1) >> 1; 1053 dec->iter_count = RTE_MAX(iter_cnt, dec->iter_count); 1054 } else { 1055 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 1056 rte_bbdev_log(ERR, "Turbo Decoder failed"); 1057 return; 1058 } 1059 } 1060 1061 static inline void 1062 enqueue_dec_one_op(struct turbo_sw_queue *q, struct rte_bbdev_dec_op *op, 1063 struct rte_bbdev_stats *queue_stats) 1064 { 1065 uint8_t c, r = 0; 1066 uint16_t kw, k = 0; 1067 uint16_t crc24_overlap = 0; 1068 struct rte_bbdev_op_turbo_dec *dec = &op->turbo_dec; 1069 struct rte_mbuf *m_in = dec->input.data; 1070 struct rte_mbuf *m_out = dec->hard_output.data; 1071 struct rte_mbuf *m_out_head = dec->hard_output.data; 1072 uint16_t in_offset = dec->input.offset; 1073 uint16_t out_offset = dec->hard_output.offset; 1074 uint32_t mbuf_total_left = dec->input.length; 1075 uint16_t seg_total_left; 1076 1077 /* Clear op status */ 1078 op->status = 0; 1079 1080 if (m_in == NULL || m_out == NULL) { 1081 rte_bbdev_log(ERR, "Invalid mbuf pointer"); 1082 op->status = 1 << RTE_BBDEV_DATA_ERROR; 1083 return; 1084 } 1085 1086 if (dec->code_block_mode == 0) { /* For Transport Block mode */ 1087 c = dec->tb_params.c; 1088 } else { /* For Code Block mode */ 1089 k = dec->cb_params.k; 1090 c = 1; 1091 } 1092 1093 if ((c > 1) && !check_bit(dec->op_flags, 1094 RTE_BBDEV_TURBO_DEC_TB_CRC_24B_KEEP)) 1095 crc24_overlap = 24; 1096 1097 while (mbuf_total_left > 0) { 1098 if (dec->code_block_mode == 0) 1099 k = (r < dec->tb_params.c_neg) ? 1100 dec->tb_params.k_neg : dec->tb_params.k_pos; 1101 1102 seg_total_left = rte_pktmbuf_data_len(m_in) - in_offset; 1103 1104 /* Calculates circular buffer size (Kw). 1105 * According to 3gpp 36.212 section 5.1.4.2 1106 * Kw = 3 * Kpi, 1107 * where: 1108 * Kpi = nCol * nRow 1109 * where nCol is 32 and nRow can be calculated from: 1110 * D =< nCol * nRow 1111 * where D is the size of each output from turbo encoder block 1112 * (k + 4). 1113 */ 1114 kw = RTE_ALIGN_CEIL(k + 4, RTE_BBDEV_C_SUBBLOCK) * 3; 1115 1116 process_dec_cb(q, op, c, k, kw, m_in, m_out_head, m_out, 1117 in_offset, out_offset, check_bit(dec->op_flags, 1118 RTE_BBDEV_TURBO_CRC_TYPE_24B), crc24_overlap, 1119 seg_total_left, queue_stats); 1120 /* To keep CRC24 attached to end of Code block, use 1121 * RTE_BBDEV_TURBO_DEC_TB_CRC_24B_KEEP flag as it 1122 * removed by default once verified. 1123 */ 1124 1125 mbuf_total_left -= kw; 1126 1127 /* Update offsets */ 1128 if (seg_total_left == kw) { 1129 /* Go to the next mbuf */ 1130 m_in = m_in->next; 1131 m_out = m_out->next; 1132 in_offset = 0; 1133 out_offset = 0; 1134 } else { 1135 /* Update offsets for next CBs (if exist) */ 1136 in_offset += kw; 1137 out_offset += ((k - crc24_overlap) >> 3); 1138 } 1139 r++; 1140 } 1141 if (mbuf_total_left != 0) { 1142 op->status |= 1 << RTE_BBDEV_DATA_ERROR; 1143 rte_bbdev_log(ERR, 1144 "Mismatch between mbuf length and included Circular buffer sizes"); 1145 } 1146 } 1147 1148 static inline uint16_t 1149 enqueue_dec_all_ops(struct turbo_sw_queue *q, struct rte_bbdev_dec_op **ops, 1150 uint16_t nb_ops, struct rte_bbdev_stats *queue_stats) 1151 { 1152 uint16_t i; 1153 #ifdef RTE_BBDEV_OFFLOAD_COST 1154 queue_stats->acc_offload_cycles = 0; 1155 #endif 1156 1157 for (i = 0; i < nb_ops; ++i) 1158 enqueue_dec_one_op(q, ops[i], queue_stats); 1159 1160 return rte_ring_enqueue_burst(q->processed_pkts, (void **)ops, nb_ops, 1161 NULL); 1162 } 1163 1164 /* Enqueue burst */ 1165 static uint16_t 1166 enqueue_enc_ops(struct rte_bbdev_queue_data *q_data, 1167 struct rte_bbdev_enc_op **ops, uint16_t nb_ops) 1168 { 1169 void *queue = q_data->queue_private; 1170 struct turbo_sw_queue *q = queue; 1171 uint16_t nb_enqueued = 0; 1172 1173 nb_enqueued = enqueue_enc_all_ops(q, ops, nb_ops, &q_data->queue_stats); 1174 1175 q_data->queue_stats.enqueue_err_count += nb_ops - nb_enqueued; 1176 q_data->queue_stats.enqueued_count += nb_enqueued; 1177 1178 return nb_enqueued; 1179 } 1180 1181 /* Enqueue burst */ 1182 static uint16_t 1183 enqueue_dec_ops(struct rte_bbdev_queue_data *q_data, 1184 struct rte_bbdev_dec_op **ops, uint16_t nb_ops) 1185 { 1186 void *queue = q_data->queue_private; 1187 struct turbo_sw_queue *q = queue; 1188 uint16_t nb_enqueued = 0; 1189 1190 nb_enqueued = enqueue_dec_all_ops(q, ops, nb_ops, &q_data->queue_stats); 1191 1192 q_data->queue_stats.enqueue_err_count += nb_ops - nb_enqueued; 1193 q_data->queue_stats.enqueued_count += nb_enqueued; 1194 1195 return nb_enqueued; 1196 } 1197 1198 /* Dequeue decode burst */ 1199 static uint16_t 1200 dequeue_dec_ops(struct rte_bbdev_queue_data *q_data, 1201 struct rte_bbdev_dec_op **ops, uint16_t nb_ops) 1202 { 1203 struct turbo_sw_queue *q = q_data->queue_private; 1204 uint16_t nb_dequeued = rte_ring_dequeue_burst(q->processed_pkts, 1205 (void **)ops, nb_ops, NULL); 1206 q_data->queue_stats.dequeued_count += nb_dequeued; 1207 1208 return nb_dequeued; 1209 } 1210 1211 /* Dequeue encode burst */ 1212 static uint16_t 1213 dequeue_enc_ops(struct rte_bbdev_queue_data *q_data, 1214 struct rte_bbdev_enc_op **ops, uint16_t nb_ops) 1215 { 1216 struct turbo_sw_queue *q = q_data->queue_private; 1217 uint16_t nb_dequeued = rte_ring_dequeue_burst(q->processed_pkts, 1218 (void **)ops, nb_ops, NULL); 1219 q_data->queue_stats.dequeued_count += nb_dequeued; 1220 1221 return nb_dequeued; 1222 } 1223 1224 /* Parse 16bit integer from string argument */ 1225 static inline int 1226 parse_u16_arg(const char *key, const char *value, void *extra_args) 1227 { 1228 uint16_t *u16 = extra_args; 1229 unsigned int long result; 1230 1231 if ((value == NULL) || (extra_args == NULL)) 1232 return -EINVAL; 1233 errno = 0; 1234 result = strtoul(value, NULL, 0); 1235 if ((result >= (1 << 16)) || (errno != 0)) { 1236 rte_bbdev_log(ERR, "Invalid value %lu for %s", result, key); 1237 return -ERANGE; 1238 } 1239 *u16 = (uint16_t)result; 1240 return 0; 1241 } 1242 1243 /* Parse parameters used to create device */ 1244 static int 1245 parse_turbo_sw_params(struct turbo_sw_params *params, const char *input_args) 1246 { 1247 struct rte_kvargs *kvlist = NULL; 1248 int ret = 0; 1249 1250 if (params == NULL) 1251 return -EINVAL; 1252 if (input_args) { 1253 kvlist = rte_kvargs_parse(input_args, turbo_sw_valid_params); 1254 if (kvlist == NULL) 1255 return -EFAULT; 1256 1257 ret = rte_kvargs_process(kvlist, turbo_sw_valid_params[0], 1258 &parse_u16_arg, ¶ms->queues_num); 1259 if (ret < 0) 1260 goto exit; 1261 1262 ret = rte_kvargs_process(kvlist, turbo_sw_valid_params[1], 1263 &parse_u16_arg, ¶ms->socket_id); 1264 if (ret < 0) 1265 goto exit; 1266 1267 if (params->socket_id >= RTE_MAX_NUMA_NODES) { 1268 rte_bbdev_log(ERR, "Invalid socket, must be < %u", 1269 RTE_MAX_NUMA_NODES); 1270 goto exit; 1271 } 1272 } 1273 1274 exit: 1275 if (kvlist) 1276 rte_kvargs_free(kvlist); 1277 return ret; 1278 } 1279 1280 /* Create device */ 1281 static int 1282 turbo_sw_bbdev_create(struct rte_vdev_device *vdev, 1283 struct turbo_sw_params *init_params) 1284 { 1285 struct rte_bbdev *bbdev; 1286 const char *name = rte_vdev_device_name(vdev); 1287 1288 bbdev = rte_bbdev_allocate(name); 1289 if (bbdev == NULL) 1290 return -ENODEV; 1291 1292 bbdev->data->dev_private = rte_zmalloc_socket(name, 1293 sizeof(struct bbdev_private), RTE_CACHE_LINE_SIZE, 1294 init_params->socket_id); 1295 if (bbdev->data->dev_private == NULL) { 1296 rte_bbdev_release(bbdev); 1297 return -ENOMEM; 1298 } 1299 1300 bbdev->dev_ops = &pmd_ops; 1301 bbdev->device = &vdev->device; 1302 bbdev->data->socket_id = init_params->socket_id; 1303 bbdev->intr_handle = NULL; 1304 1305 /* register rx/tx burst functions for data path */ 1306 bbdev->dequeue_enc_ops = dequeue_enc_ops; 1307 bbdev->dequeue_dec_ops = dequeue_dec_ops; 1308 bbdev->enqueue_enc_ops = enqueue_enc_ops; 1309 bbdev->enqueue_dec_ops = enqueue_dec_ops; 1310 ((struct bbdev_private *) bbdev->data->dev_private)->max_nb_queues = 1311 init_params->queues_num; 1312 1313 return 0; 1314 } 1315 1316 /* Initialise device */ 1317 static int 1318 turbo_sw_bbdev_probe(struct rte_vdev_device *vdev) 1319 { 1320 struct turbo_sw_params init_params = { 1321 rte_socket_id(), 1322 RTE_BBDEV_DEFAULT_MAX_NB_QUEUES 1323 }; 1324 const char *name; 1325 const char *input_args; 1326 1327 if (vdev == NULL) 1328 return -EINVAL; 1329 1330 name = rte_vdev_device_name(vdev); 1331 if (name == NULL) 1332 return -EINVAL; 1333 input_args = rte_vdev_device_args(vdev); 1334 parse_turbo_sw_params(&init_params, input_args); 1335 1336 rte_bbdev_log_debug( 1337 "Initialising %s on NUMA node %d with max queues: %d\n", 1338 name, init_params.socket_id, init_params.queues_num); 1339 1340 return turbo_sw_bbdev_create(vdev, &init_params); 1341 } 1342 1343 /* Uninitialise device */ 1344 static int 1345 turbo_sw_bbdev_remove(struct rte_vdev_device *vdev) 1346 { 1347 struct rte_bbdev *bbdev; 1348 const char *name; 1349 1350 if (vdev == NULL) 1351 return -EINVAL; 1352 1353 name = rte_vdev_device_name(vdev); 1354 if (name == NULL) 1355 return -EINVAL; 1356 1357 bbdev = rte_bbdev_get_named_dev(name); 1358 if (bbdev == NULL) 1359 return -EINVAL; 1360 1361 rte_free(bbdev->data->dev_private); 1362 1363 return rte_bbdev_release(bbdev); 1364 } 1365 1366 static struct rte_vdev_driver bbdev_turbo_sw_pmd_drv = { 1367 .probe = turbo_sw_bbdev_probe, 1368 .remove = turbo_sw_bbdev_remove 1369 }; 1370 1371 RTE_PMD_REGISTER_VDEV(DRIVER_NAME, bbdev_turbo_sw_pmd_drv); 1372 RTE_PMD_REGISTER_PARAM_STRING(DRIVER_NAME, 1373 TURBO_SW_MAX_NB_QUEUES_ARG"=<int> " 1374 TURBO_SW_SOCKET_ID_ARG"=<int>"); 1375 RTE_PMD_REGISTER_ALIAS(DRIVER_NAME, turbo_sw); 1376 1377 RTE_INIT(turbo_sw_bbdev_init_log) 1378 { 1379 bbdev_turbo_sw_logtype = rte_log_register("pmd.bb.turbo_sw"); 1380 if (bbdev_turbo_sw_logtype >= 0) 1381 rte_log_set_level(bbdev_turbo_sw_logtype, RTE_LOG_NOTICE); 1382 } 1383