1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2017 Intel Corporation
7 #include <rte_common.h>
8 #include <rte_bus_vdev.h>
9 #include <rte_malloc.h>
11 #include <rte_kvargs.h>
13 #include <rte_bbdev.h>
14 #include <rte_bbdev_pmd.h>
16 #include <phy_turbo.h>
18 #include <phy_rate_match.h>
21 #define DRIVER_NAME turbo_sw
23 /* Turbo SW PMD logging ID */
24 static int bbdev_turbo_sw_logtype;
26 /* Helper macro for logging */
27 #define rte_bbdev_log(level, fmt, ...) \
28 rte_log(RTE_LOG_ ## level, bbdev_turbo_sw_logtype, fmt "\n", \
31 #define rte_bbdev_log_debug(fmt, ...) \
32 rte_bbdev_log(DEBUG, RTE_STR(__LINE__) ":%s() " fmt, __func__, \
35 #define DEINT_INPUT_BUF_SIZE (((RTE_BBDEV_MAX_CB_SIZE >> 3) + 1) * 48)
36 #define DEINT_OUTPUT_BUF_SIZE (DEINT_INPUT_BUF_SIZE * 6)
37 #define ADAPTER_OUTPUT_BUF_SIZE ((RTE_BBDEV_MAX_CB_SIZE + 4) * 48)
39 /* private data structure */
40 struct bbdev_private {
41 unsigned int max_nb_queues; /**< Max number of queues */
44 /* Initialisation params structure that can be used by Turbo SW driver */
45 struct turbo_sw_params {
46 int socket_id; /*< Turbo SW device socket */
47 uint16_t queues_num; /*< Turbo SW device queues number */
50 /* Accecptable params for Turbo SW devices */
51 #define TURBO_SW_MAX_NB_QUEUES_ARG "max_nb_queues"
52 #define TURBO_SW_SOCKET_ID_ARG "socket_id"
54 static const char * const turbo_sw_valid_params[] = {
55 TURBO_SW_MAX_NB_QUEUES_ARG,
56 TURBO_SW_SOCKET_ID_ARG
60 struct turbo_sw_queue {
61 /* Ring for processed (encoded/decoded) operations which are ready to
64 struct rte_ring *processed_pkts;
65 /* Stores input for turbo encoder (used when CRC attachment is
69 /* Stores output from turbo encoder */
71 /* Alpha gamma buf for bblib_turbo_decoder() function */
73 /* Temp buf for bblib_turbo_decoder() function */
75 /* Input buf for bblib_rate_dematching_lte() function */
77 /* Output buf for bblib_rate_dematching_lte() function */
78 uint8_t *deint_output;
79 /* Output buf for bblib_turbodec_adapter_lte() function */
80 uint8_t *adapter_output;
81 /* Operation type of this queue */
82 enum rte_bbdev_op_type type;
83 } __rte_cache_aligned;
85 /* Calculate index based on Table 5.1.3-3 from TS34.212 */
87 compute_idx(uint16_t k)
91 if (k < RTE_BBDEV_MIN_CB_SIZE || k > RTE_BBDEV_MAX_CB_SIZE)
95 if ((k - 2048) % 64 != 0)
98 result = 124 + (k - 2048) / 64;
99 } else if (k <= 512) {
100 if ((k - 40) % 8 != 0)
103 result = (k - 40) / 8 + 1;
104 } else if (k <= 1024) {
105 if ((k - 512) % 16 != 0)
108 result = 60 + (k - 512) / 16;
109 } else { /* 1024 < k <= 2048 */
110 if ((k - 1024) % 32 != 0)
113 result = 92 + (k - 1024) / 32;
119 /* Read flag value 0/1 from bitmap */
121 check_bit(uint32_t bitmap, uint32_t bitmask)
123 return bitmap & bitmask;
126 /* Get device info */
128 info_get(struct rte_bbdev *dev, struct rte_bbdev_driver_info *dev_info)
130 struct bbdev_private *internals = dev->data->dev_private;
132 static const struct rte_bbdev_op_cap bbdev_capabilities[] = {
134 .type = RTE_BBDEV_OP_TURBO_DEC,
137 RTE_BBDEV_TURBO_SUBBLOCK_DEINTERLEAVE |
138 RTE_BBDEV_TURBO_POS_LLR_1_BIT_IN |
139 RTE_BBDEV_TURBO_NEG_LLR_1_BIT_IN |
140 RTE_BBDEV_TURBO_CRC_TYPE_24B |
141 RTE_BBDEV_TURBO_EARLY_TERMINATION,
142 .max_llr_modulus = 16,
143 .num_buffers_src = RTE_BBDEV_MAX_CODE_BLOCKS,
144 .num_buffers_hard_out =
145 RTE_BBDEV_MAX_CODE_BLOCKS,
146 .num_buffers_soft_out = 0,
150 .type = RTE_BBDEV_OP_TURBO_ENC,
153 RTE_BBDEV_TURBO_CRC_24B_ATTACH |
154 RTE_BBDEV_TURBO_CRC_24A_ATTACH |
155 RTE_BBDEV_TURBO_RATE_MATCH |
156 RTE_BBDEV_TURBO_RV_INDEX_BYPASS,
157 .num_buffers_src = RTE_BBDEV_MAX_CODE_BLOCKS,
158 .num_buffers_dst = RTE_BBDEV_MAX_CODE_BLOCKS,
161 RTE_BBDEV_END_OF_CAPABILITIES_LIST()
164 static struct rte_bbdev_queue_conf default_queue_conf = {
165 .queue_size = RTE_BBDEV_QUEUE_SIZE_LIMIT,
168 static const enum rte_cpu_flag_t cpu_flag = RTE_CPUFLAG_SSE4_2;
170 default_queue_conf.socket = dev->data->socket_id;
172 dev_info->driver_name = RTE_STR(DRIVER_NAME);
173 dev_info->max_num_queues = internals->max_nb_queues;
174 dev_info->queue_size_lim = RTE_BBDEV_QUEUE_SIZE_LIMIT;
175 dev_info->hardware_accelerated = false;
176 dev_info->max_queue_priority = 0;
177 dev_info->default_queue_conf = default_queue_conf;
178 dev_info->capabilities = bbdev_capabilities;
179 dev_info->cpu_flag_reqs = &cpu_flag;
180 dev_info->min_alignment = 64;
182 rte_bbdev_log_debug("got device info from %u\n", dev->data->dev_id);
187 q_release(struct rte_bbdev *dev, uint16_t q_id)
189 struct turbo_sw_queue *q = dev->data->queues[q_id].queue_private;
192 rte_ring_free(q->processed_pkts);
193 rte_free(q->enc_out);
196 rte_free(q->code_block);
197 rte_free(q->deint_input);
198 rte_free(q->deint_output);
199 rte_free(q->adapter_output);
201 dev->data->queues[q_id].queue_private = NULL;
204 rte_bbdev_log_debug("released device queue %u:%u",
205 dev->data->dev_id, q_id);
211 q_setup(struct rte_bbdev *dev, uint16_t q_id,
212 const struct rte_bbdev_queue_conf *queue_conf)
215 struct turbo_sw_queue *q;
216 char name[RTE_RING_NAMESIZE];
218 /* Allocate the queue data structure. */
219 q = rte_zmalloc_socket(RTE_STR(DRIVER_NAME), sizeof(*q),
220 RTE_CACHE_LINE_SIZE, queue_conf->socket);
222 rte_bbdev_log(ERR, "Failed to allocate queue memory");
226 /* Allocate memory for encoder output. */
227 ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"_enc_out%u:%u",
228 dev->data->dev_id, q_id);
229 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
231 "Creating queue name for device %u queue %u failed",
232 dev->data->dev_id, q_id);
233 return -ENAMETOOLONG;
235 q->enc_out = rte_zmalloc_socket(name,
236 ((RTE_BBDEV_MAX_TB_SIZE >> 3) + 3) *
237 sizeof(*q->enc_out) * 3,
238 RTE_CACHE_LINE_SIZE, queue_conf->socket);
239 if (q->enc_out == NULL) {
241 "Failed to allocate queue memory for %s", name);
245 /* Allocate memory for rate matching output. */
246 ret = snprintf(name, RTE_RING_NAMESIZE,
247 RTE_STR(DRIVER_NAME)"_enc_in%u:%u", dev->data->dev_id,
249 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
251 "Creating queue name for device %u queue %u failed",
252 dev->data->dev_id, q_id);
253 return -ENAMETOOLONG;
255 q->enc_in = rte_zmalloc_socket(name,
256 (RTE_BBDEV_MAX_CB_SIZE >> 3) * sizeof(*q->enc_in),
257 RTE_CACHE_LINE_SIZE, queue_conf->socket);
258 if (q->enc_in == NULL) {
260 "Failed to allocate queue memory for %s", name);
264 /* Allocate memory for Aplha Gamma temp buffer. */
265 ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"_ag%u:%u",
266 dev->data->dev_id, q_id);
267 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
269 "Creating queue name for device %u queue %u failed",
270 dev->data->dev_id, q_id);
271 return -ENAMETOOLONG;
273 q->ag = rte_zmalloc_socket(name,
274 RTE_BBDEV_MAX_CB_SIZE * 10 * sizeof(*q->ag),
275 RTE_CACHE_LINE_SIZE, queue_conf->socket);
278 "Failed to allocate queue memory for %s", name);
282 /* Allocate memory for code block temp buffer. */
283 ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"_cb%u:%u",
284 dev->data->dev_id, q_id);
285 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
287 "Creating queue name for device %u queue %u failed",
288 dev->data->dev_id, q_id);
289 return -ENAMETOOLONG;
291 q->code_block = rte_zmalloc_socket(name,
292 RTE_BBDEV_MAX_CB_SIZE * sizeof(*q->code_block),
293 RTE_CACHE_LINE_SIZE, queue_conf->socket);
294 if (q->code_block == NULL) {
296 "Failed to allocate queue memory for %s", name);
300 /* Allocate memory for Deinterleaver input. */
301 ret = snprintf(name, RTE_RING_NAMESIZE,
302 RTE_STR(DRIVER_NAME)"_deint_input%u:%u",
303 dev->data->dev_id, q_id);
304 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
306 "Creating queue name for device %u queue %u failed",
307 dev->data->dev_id, q_id);
308 return -ENAMETOOLONG;
310 q->deint_input = rte_zmalloc_socket(name,
311 DEINT_INPUT_BUF_SIZE * sizeof(*q->deint_input),
312 RTE_CACHE_LINE_SIZE, queue_conf->socket);
313 if (q->deint_input == NULL) {
315 "Failed to allocate queue memory for %s", name);
319 /* Allocate memory for Deinterleaver output. */
320 ret = snprintf(name, RTE_RING_NAMESIZE,
321 RTE_STR(DRIVER_NAME)"_deint_output%u:%u",
322 dev->data->dev_id, q_id);
323 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
325 "Creating queue name for device %u queue %u failed",
326 dev->data->dev_id, q_id);
327 return -ENAMETOOLONG;
329 q->deint_output = rte_zmalloc_socket(NULL,
330 DEINT_OUTPUT_BUF_SIZE * sizeof(*q->deint_output),
331 RTE_CACHE_LINE_SIZE, queue_conf->socket);
332 if (q->deint_output == NULL) {
334 "Failed to allocate queue memory for %s", name);
338 /* Allocate memory for Adapter output. */
339 ret = snprintf(name, RTE_RING_NAMESIZE,
340 RTE_STR(DRIVER_NAME)"_adapter_output%u:%u",
341 dev->data->dev_id, q_id);
342 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
344 "Creating queue name for device %u queue %u failed",
345 dev->data->dev_id, q_id);
346 return -ENAMETOOLONG;
348 q->adapter_output = rte_zmalloc_socket(NULL,
349 ADAPTER_OUTPUT_BUF_SIZE * sizeof(*q->adapter_output),
350 RTE_CACHE_LINE_SIZE, queue_conf->socket);
351 if (q->adapter_output == NULL) {
353 "Failed to allocate queue memory for %s", name);
357 /* Create ring for packets awaiting to be dequeued. */
358 ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"%u:%u",
359 dev->data->dev_id, q_id);
360 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
362 "Creating queue name for device %u queue %u failed",
363 dev->data->dev_id, q_id);
364 return -ENAMETOOLONG;
366 q->processed_pkts = rte_ring_create(name, queue_conf->queue_size,
367 queue_conf->socket, RING_F_SP_ENQ | RING_F_SC_DEQ);
368 if (q->processed_pkts == NULL) {
369 rte_bbdev_log(ERR, "Failed to create ring for %s", name);
373 q->type = queue_conf->op_type;
375 dev->data->queues[q_id].queue_private = q;
376 rte_bbdev_log_debug("setup device queue %s", name);
380 rte_ring_free(q->processed_pkts);
381 rte_free(q->enc_out);
384 rte_free(q->code_block);
385 rte_free(q->deint_input);
386 rte_free(q->deint_output);
387 rte_free(q->adapter_output);
392 static const struct rte_bbdev_ops pmd_ops = {
393 .info_get = info_get,
394 .queue_setup = q_setup,
395 .queue_release = q_release
398 /* Checks if the encoder input buffer is correct.
399 * Returns 0 if it's valid, -1 otherwise.
402 is_enc_input_valid(const uint16_t k, const int32_t k_idx,
403 const uint16_t in_length)
406 rte_bbdev_log(ERR, "K Index is invalid");
410 if (in_length - (k >> 3) < 0) {
412 "Mismatch between input length (%u bytes) and K (%u bits)",
417 if (k > RTE_BBDEV_MAX_CB_SIZE) {
418 rte_bbdev_log(ERR, "CB size (%u) is too big, max: %d",
419 k, RTE_BBDEV_MAX_CB_SIZE);
426 /* Checks if the decoder input buffer is correct.
427 * Returns 0 if it's valid, -1 otherwise.
430 is_dec_input_valid(int32_t k_idx, int16_t kw, int16_t in_length)
433 rte_bbdev_log(ERR, "K index is invalid");
437 if (in_length - kw < 0) {
439 "Mismatch between input length (%u) and kw (%u)",
444 if (kw > RTE_BBDEV_MAX_KW) {
445 rte_bbdev_log(ERR, "Input length (%u) is too big, max: %d",
446 kw, RTE_BBDEV_MAX_KW);
454 process_enc_cb(struct turbo_sw_queue *q, struct rte_bbdev_enc_op *op,
455 uint8_t r, uint8_t c, uint16_t k, uint16_t ncb,
456 uint32_t e, struct rte_mbuf *m_in, struct rte_mbuf *m_out,
457 uint16_t in_offset, uint16_t out_offset, uint16_t total_left)
462 uint8_t *in, *out0, *out1, *out2, *tmp_out, *rm_out;
463 uint64_t first_3_bytes = 0;
464 struct rte_bbdev_op_turbo_enc *enc = &op->turbo_enc;
465 struct bblib_crc_request crc_req;
466 struct bblib_crc_response crc_resp;
467 struct bblib_turbo_encoder_request turbo_req;
468 struct bblib_turbo_encoder_response turbo_resp;
469 struct bblib_rate_match_dl_request rm_req;
470 struct bblib_rate_match_dl_response rm_resp;
472 k_idx = compute_idx(k);
473 in = rte_pktmbuf_mtod_offset(m_in, uint8_t *, in_offset);
475 /* CRC24A (for TB) */
476 if ((enc->op_flags & RTE_BBDEV_TURBO_CRC_24A_ATTACH) &&
477 (enc->code_block_mode == 1)) {
478 ret = is_enc_input_valid(k - 24, k_idx, total_left);
480 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
484 crc_req.len = (k - 24) >> 3;
485 /* Check if there is a room for CRC bits. If not use
486 * the temporary buffer.
488 if (rte_pktmbuf_append(m_in, 3) == NULL) {
489 rte_memcpy(q->enc_in, in, (k - 24) >> 3);
492 /* Store 3 first bytes of next CB as they will be
493 * overwritten by CRC bytes. If it is the last CB then
494 * there is no point to store 3 next bytes and this
495 * if..else branch will be omitted.
497 first_3_bytes = *((uint64_t *)&in[(k - 32) >> 3]);
501 bblib_lte_crc24a_gen(&crc_req, &crc_resp);
502 } else if (enc->op_flags & RTE_BBDEV_TURBO_CRC_24B_ATTACH) {
504 ret = is_enc_input_valid(k - 24, k_idx, total_left);
506 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
510 crc_req.len = (k - 24) >> 3;
511 /* Check if there is a room for CRC bits. If this is the last
512 * CB in TB. If not use temporary buffer.
514 if ((c - r == 1) && (rte_pktmbuf_append(m_in, 3) == NULL)) {
515 rte_memcpy(q->enc_in, in, (k - 24) >> 3);
517 } else if (c - r > 1) {
518 /* Store 3 first bytes of next CB as they will be
519 * overwritten by CRC bytes. If it is the last CB then
520 * there is no point to store 3 next bytes and this
521 * if..else branch will be omitted.
523 first_3_bytes = *((uint64_t *)&in[(k - 32) >> 3]);
527 bblib_lte_crc24b_gen(&crc_req, &crc_resp);
529 ret = is_enc_input_valid(k, k_idx, total_left);
531 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
538 /* Each bit layer output from turbo encoder is (k+4) bits long, i.e.
539 * input length + 4 tail bits. That's (k/8) + 1 bytes after rounding up.
540 * So dst_data's length should be 3*(k/8) + 3 bytes.
541 * In Rate-matching bypass case outputs pointers passed to encoder
542 * (out0, out1 and out2) can directly point to addresses of output from
545 if (enc->op_flags & RTE_BBDEV_TURBO_RATE_MATCH) {
547 out1 = RTE_PTR_ADD(out0, (k >> 3) + 1);
548 out2 = RTE_PTR_ADD(out1, (k >> 3) + 1);
550 out0 = (uint8_t *)rte_pktmbuf_append(m_out, (k >> 3) * 3 + 2);
552 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
554 "Too little space in output mbuf");
557 enc->output.length += (k >> 3) * 3 + 2;
558 /* rte_bbdev_op_data.offset can be different than the
559 * offset of the appended bytes
561 out0 = rte_pktmbuf_mtod_offset(m_out, uint8_t *, out_offset);
562 out1 = rte_pktmbuf_mtod_offset(m_out, uint8_t *,
563 out_offset + (k >> 3) + 1);
564 out2 = rte_pktmbuf_mtod_offset(m_out, uint8_t *,
565 out_offset + 2 * ((k >> 3) + 1));
568 turbo_req.case_id = k_idx;
569 turbo_req.input_win = in;
570 turbo_req.length = k >> 3;
571 turbo_resp.output_win_0 = out0;
572 turbo_resp.output_win_1 = out1;
573 turbo_resp.output_win_2 = out2;
574 if (bblib_turbo_encoder(&turbo_req, &turbo_resp) != 0) {
575 op->status |= 1 << RTE_BBDEV_DRV_ERROR;
576 rte_bbdev_log(ERR, "Turbo Encoder failed");
580 /* Restore 3 first bytes of next CB if they were overwritten by CRC*/
581 if (first_3_bytes != 0)
582 *((uint64_t *)&in[(k - 32) >> 3]) = first_3_bytes;
585 if (enc->op_flags & RTE_BBDEV_TURBO_RATE_MATCH) {
587 /* Integer round up division by 8 */
588 uint16_t out_len = (e + 7) >> 3;
589 /* The mask array is indexed using E%8. E is an even number so
590 * there are only 4 possible values.
592 const uint8_t mask_out[] = {0xFF, 0xC0, 0xF0, 0xFC};
594 /* get output data starting address */
595 rm_out = (uint8_t *)rte_pktmbuf_append(m_out, out_len);
596 if (rm_out == NULL) {
597 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
599 "Too little space in output mbuf");
602 /* rte_bbdev_op_data.offset can be different than the offset
603 * of the appended bytes
605 rm_out = rte_pktmbuf_mtod_offset(m_out, uint8_t *, out_offset);
607 /* index of current code block */
609 /* total number of code block */
611 /* For DL - 1, UL - 0 */
612 rm_req.direction = 1;
613 /* According to 3ggp 36.212 Spec 5.1.4.1.2 section Nsoft, KMIMO
614 * and MDL_HARQ are used for Ncb calculation. As Ncb is already
615 * known we can adjust those parameters
617 rm_req.Nsoft = ncb * rm_req.C;
620 /* According to 3ggp 36.212 Spec 5.1.4.1.2 section Nl, Qm and G
621 * are used for E calculation. As E is already known we can
622 * adjust those parameters
626 rm_req.G = rm_req.NL * rm_req.Qm * rm_req.C;
628 rm_req.rvidx = enc->rv_index;
629 rm_req.Kidx = k_idx - 1;
634 rm_resp.output = rm_out;
635 rm_resp.OutputLen = out_len;
636 if (enc->op_flags & RTE_BBDEV_TURBO_RV_INDEX_BYPASS)
637 rm_req.bypass_rvidx = 1;
639 rm_req.bypass_rvidx = 0;
641 if (bblib_rate_match_dl(&rm_req, &rm_resp) != 0) {
642 op->status |= 1 << RTE_BBDEV_DRV_ERROR;
643 rte_bbdev_log(ERR, "Rate matching failed");
647 /* SW fills an entire last byte even if E%8 != 0. Clear the
648 * superfluous data bits for consistency with HW device.
650 mask_id = (e & 7) >> 1;
651 rm_out[out_len - 1] &= mask_out[mask_id];
653 enc->output.length += rm_resp.OutputLen;
655 /* Rate matching is bypassed */
657 /* Completing last byte of out0 (where 4 tail bits are stored)
658 * by moving first 4 bits from out1
660 tmp_out = (uint8_t *) --out1;
661 *tmp_out = *tmp_out | ((*(tmp_out + 1) & 0xF0) >> 4);
663 /* Shifting out1 data by 4 bits to the left */
664 for (m = 0; m < k >> 3; ++m) {
665 uint8_t *first = tmp_out;
666 uint8_t second = *(tmp_out + 1);
667 *first = (*first << 4) | ((second & 0xF0) >> 4);
670 /* Shifting out2 data by 8 bits to the left */
671 for (m = 0; m < (k >> 3) + 1; ++m) {
672 *tmp_out = *(tmp_out + 1);
680 enqueue_enc_one_op(struct turbo_sw_queue *q, struct rte_bbdev_enc_op *op)
682 uint8_t c, r, crc24_bits = 0;
685 struct rte_bbdev_op_turbo_enc *enc = &op->turbo_enc;
686 uint16_t in_offset = enc->input.offset;
687 uint16_t out_offset = enc->output.offset;
688 struct rte_mbuf *m_in = enc->input.data;
689 struct rte_mbuf *m_out = enc->output.data;
690 uint16_t total_left = enc->input.length;
692 /* Clear op status */
695 if (total_left > RTE_BBDEV_MAX_TB_SIZE >> 3) {
696 rte_bbdev_log(ERR, "TB size (%u) is too big, max: %d",
697 total_left, RTE_BBDEV_MAX_TB_SIZE);
698 op->status = 1 << RTE_BBDEV_DATA_ERROR;
702 if (m_in == NULL || m_out == NULL) {
703 rte_bbdev_log(ERR, "Invalid mbuf pointer");
704 op->status = 1 << RTE_BBDEV_DATA_ERROR;
708 if ((enc->op_flags & RTE_BBDEV_TURBO_CRC_24B_ATTACH) ||
709 (enc->op_flags & RTE_BBDEV_TURBO_CRC_24A_ATTACH))
712 if (enc->code_block_mode == 0) { /* For Transport Block mode */
713 c = enc->tb_params.c;
714 r = enc->tb_params.r;
715 } else {/* For Code Block mode */
720 while (total_left > 0 && r < c) {
721 if (enc->code_block_mode == 0) {
722 k = (r < enc->tb_params.c_neg) ?
723 enc->tb_params.k_neg : enc->tb_params.k_pos;
724 ncb = (r < enc->tb_params.c_neg) ?
725 enc->tb_params.ncb_neg : enc->tb_params.ncb_pos;
726 e = (r < enc->tb_params.cab) ?
727 enc->tb_params.ea : enc->tb_params.eb;
729 k = enc->cb_params.k;
730 ncb = enc->cb_params.ncb;
731 e = enc->cb_params.e;
734 process_enc_cb(q, op, r, c, k, ncb, e, m_in,
735 m_out, in_offset, out_offset, total_left);
736 /* Update total_left */
737 total_left -= (k - crc24_bits) >> 3;
738 /* Update offsets for next CBs (if exist) */
739 in_offset += (k - crc24_bits) >> 3;
740 if (enc->op_flags & RTE_BBDEV_TURBO_RATE_MATCH)
741 out_offset += e >> 3;
743 out_offset += (k >> 3) * 3 + 2;
747 /* check if all input data was processed */
748 if (total_left != 0) {
749 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
751 "Mismatch between mbuf length and included CBs sizes");
755 static inline uint16_t
756 enqueue_enc_all_ops(struct turbo_sw_queue *q, struct rte_bbdev_enc_op **ops,
761 for (i = 0; i < nb_ops; ++i)
762 enqueue_enc_one_op(q, ops[i]);
764 return rte_ring_enqueue_burst(q->processed_pkts, (void **)ops, nb_ops,
768 /* Remove the padding bytes from a cyclic buffer.
769 * The input buffer is a data stream wk as described in 3GPP TS 36.212 section
770 * 5.1.4.1.2 starting from w0 and with length Ncb bytes.
771 * The output buffer is a data stream wk with pruned padding bytes. It's length
772 * is 3*D bytes and the order of non-padding bytes is preserved.
775 remove_nulls_from_circular_buf(const uint8_t *in, uint8_t *out, uint16_t k,
778 uint32_t in_idx, out_idx, c_idx;
779 const uint32_t d = k + 4;
780 const uint32_t kw = (ncb / 3);
781 const uint32_t nd = kw - d;
782 const uint32_t r_subblock = kw / RTE_BBDEV_C_SUBBLOCK;
783 /* Inter-column permutation pattern */
784 const uint32_t P[RTE_BBDEV_C_SUBBLOCK] = {0, 16, 8, 24, 4, 20, 12, 28,
785 2, 18, 10, 26, 6, 22, 14, 30, 1, 17, 9, 25, 5, 21, 13,
786 29, 3, 19, 11, 27, 7, 23, 15, 31};
790 /* The padding bytes are at the first Nd positions in the first row. */
791 for (c_idx = 0; in_idx < kw; in_idx += r_subblock, ++c_idx) {
793 rte_memcpy(&out[out_idx], &in[in_idx + 1],
795 out_idx += r_subblock - 1;
797 rte_memcpy(&out[out_idx], &in[in_idx], r_subblock);
798 out_idx += r_subblock;
802 /* First and second parity bits sub-blocks are interlaced. */
803 for (c_idx = 0; in_idx < ncb - 2 * r_subblock;
804 in_idx += 2 * r_subblock, ++c_idx) {
805 uint32_t second_block_c_idx = P[c_idx];
806 uint32_t third_block_c_idx = P[c_idx] + 1;
808 if (second_block_c_idx < nd && third_block_c_idx < nd) {
809 rte_memcpy(&out[out_idx], &in[in_idx + 2],
811 out_idx += 2 * r_subblock - 2;
812 } else if (second_block_c_idx >= nd &&
813 third_block_c_idx >= nd) {
814 rte_memcpy(&out[out_idx], &in[in_idx], 2 * r_subblock);
815 out_idx += 2 * r_subblock;
816 } else if (second_block_c_idx < nd) {
817 out[out_idx++] = in[in_idx];
818 rte_memcpy(&out[out_idx], &in[in_idx + 2],
820 out_idx += 2 * r_subblock - 2;
822 rte_memcpy(&out[out_idx], &in[in_idx + 1],
824 out_idx += 2 * r_subblock - 1;
828 /* Last interlaced row is different - its last byte is the only padding
829 * byte. We can have from 4 up to 28 padding bytes (Nd) per sub-block.
830 * After interlacing the 1st and 2nd parity sub-blocks we can have 0, 1
831 * or 2 padding bytes each time we make a step of 2 * R_SUBBLOCK bytes
832 * (moving to another column). 2nd parity sub-block uses the same
833 * inter-column permutation pattern as the systematic and 1st parity
834 * sub-blocks but it adds '1' to the resulting index and calculates the
835 * modulus of the result and Kw. Last column is mapped to itself (id 31)
836 * so the first byte taken from the 2nd parity sub-block will be the
837 * 32nd (31+1) byte, then 64th etc. (step is C_SUBBLOCK == 32) and the
838 * last byte will be the first byte from the sub-block:
839 * (32 + 32 * (R_SUBBLOCK-1)) % Kw == Kw % Kw == 0. Nd can't be smaller
840 * than 4 so we know that bytes with ids 0, 1, 2 and 3 must be the
841 * padding bytes. The bytes from the 1st parity sub-block are the bytes
842 * from the 31st column - Nd can't be greater than 28 so we are sure
843 * that there are no padding bytes in 31st column.
845 rte_memcpy(&out[out_idx], &in[in_idx], 2 * r_subblock - 1);
849 move_padding_bytes(const uint8_t *in, uint8_t *out, uint16_t k,
853 uint16_t kpi = ncb / 3;
854 uint16_t nd = kpi - d;
856 rte_memcpy(&out[nd], in, d);
857 rte_memcpy(&out[nd + kpi + 64], &in[kpi], d);
858 rte_memcpy(&out[(nd - 1) + 2 * (kpi + 64)], &in[2 * kpi], d);
862 process_dec_cb(struct turbo_sw_queue *q, struct rte_bbdev_dec_op *op,
863 uint8_t c, uint16_t k, uint16_t kw, struct rte_mbuf *m_in,
864 struct rte_mbuf *m_out, uint16_t in_offset, uint16_t out_offset,
865 bool check_crc_24b, uint16_t total_left)
870 uint8_t *in, *out, *adapter_input;
871 int32_t ncb, ncb_without_null;
872 struct bblib_turbo_adapter_ul_response adapter_resp;
873 struct bblib_turbo_adapter_ul_request adapter_req;
874 struct bblib_turbo_decoder_request turbo_req;
875 struct bblib_turbo_decoder_response turbo_resp;
876 struct rte_bbdev_op_turbo_dec *dec = &op->turbo_dec;
878 k_idx = compute_idx(k);
880 ret = is_dec_input_valid(k_idx, kw, total_left);
882 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
886 in = rte_pktmbuf_mtod_offset(m_in, uint8_t *, in_offset);
888 ncb_without_null = (k + 4) * 3;
890 if (check_bit(dec->op_flags, RTE_BBDEV_TURBO_SUBBLOCK_DEINTERLEAVE)) {
891 struct bblib_deinterleave_ul_request deint_req;
892 struct bblib_deinterleave_ul_response deint_resp;
894 /* SW decoder accepts only a circular buffer without NULL bytes
895 * so the input needs to be converted.
897 remove_nulls_from_circular_buf(in, q->deint_input, k, ncb);
899 deint_req.pharqbuffer = q->deint_input;
900 deint_req.ncb = ncb_without_null;
901 deint_resp.pinteleavebuffer = q->deint_output;
902 bblib_deinterleave_ul(&deint_req, &deint_resp);
904 move_padding_bytes(in, q->deint_output, k, ncb);
906 adapter_input = q->deint_output;
908 if (dec->op_flags & RTE_BBDEV_TURBO_POS_LLR_1_BIT_IN)
909 adapter_req.isinverted = 1;
910 else if (dec->op_flags & RTE_BBDEV_TURBO_NEG_LLR_1_BIT_IN)
911 adapter_req.isinverted = 0;
913 op->status |= 1 << RTE_BBDEV_DRV_ERROR;
914 rte_bbdev_log(ERR, "LLR format wasn't specified");
918 adapter_req.ncb = ncb_without_null;
919 adapter_req.pinteleavebuffer = adapter_input;
920 adapter_resp.pharqout = q->adapter_output;
921 bblib_turbo_adapter_ul(&adapter_req, &adapter_resp);
923 out = (uint8_t *)rte_pktmbuf_append(m_out, (k >> 3));
925 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
926 rte_bbdev_log(ERR, "Too little space in output mbuf");
929 /* rte_bbdev_op_data.offset can be different than the offset of the
932 out = rte_pktmbuf_mtod_offset(m_out, uint8_t *, out_offset);
937 turbo_req.input = (int8_t *)q->adapter_output;
939 turbo_req.k_idx = k_idx;
940 turbo_req.max_iter_num = dec->iter_max;
941 turbo_resp.ag_buf = q->ag;
942 turbo_resp.cb_buf = q->code_block;
943 turbo_resp.output = out;
944 iter_cnt = bblib_turbo_decoder(&turbo_req, &turbo_resp);
945 dec->hard_output.length += (k >> 3);
948 /* Temporary solution for returned iter_count from SDK */
949 iter_cnt = (iter_cnt - 1) / 2;
950 dec->iter_count = RTE_MAX(iter_cnt, dec->iter_count);
952 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
953 rte_bbdev_log(ERR, "Turbo Decoder failed");
959 enqueue_dec_one_op(struct turbo_sw_queue *q, struct rte_bbdev_dec_op *op)
963 struct rte_bbdev_op_turbo_dec *dec = &op->turbo_dec;
964 struct rte_mbuf *m_in = dec->input.data;
965 struct rte_mbuf *m_out = dec->hard_output.data;
966 uint16_t in_offset = dec->input.offset;
967 uint16_t total_left = dec->input.length;
968 uint16_t out_offset = dec->hard_output.offset;
970 /* Clear op status */
973 if (m_in == NULL || m_out == NULL) {
974 rte_bbdev_log(ERR, "Invalid mbuf pointer");
975 op->status = 1 << RTE_BBDEV_DATA_ERROR;
979 if (dec->code_block_mode == 0) { /* For Transport Block mode */
980 c = dec->tb_params.c;
981 } else { /* For Code Block mode */
982 k = dec->cb_params.k;
986 while (total_left > 0) {
987 if (dec->code_block_mode == 0)
988 k = (r < dec->tb_params.c_neg) ?
989 dec->tb_params.k_neg : dec->tb_params.k_pos;
991 /* Calculates circular buffer size (Kw).
992 * According to 3gpp 36.212 section 5.1.4.2
996 * where nCol is 32 and nRow can be calculated from:
998 * where D is the size of each output from turbo encoder block
1001 kw = RTE_ALIGN_CEIL(k + 4, RTE_BBDEV_C_SUBBLOCK) * 3;
1003 process_dec_cb(q, op, c, k, kw, m_in, m_out, in_offset,
1004 out_offset, check_bit(dec->op_flags,
1005 RTE_BBDEV_TURBO_CRC_TYPE_24B), total_left);
1006 /* As a result of decoding we get Code Block with included
1007 * decoded CRC24 at the end of Code Block. Type of CRC24 is
1008 * specified by flag.
1011 /* Update total_left */
1013 /* Update offsets for next CBs (if exist) */
1015 out_offset += (k >> 3);
1018 if (total_left != 0) {
1019 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
1021 "Mismatch between mbuf length and included Circular buffer sizes");
1025 static inline uint16_t
1026 enqueue_dec_all_ops(struct turbo_sw_queue *q, struct rte_bbdev_dec_op **ops,
1031 for (i = 0; i < nb_ops; ++i)
1032 enqueue_dec_one_op(q, ops[i]);
1034 return rte_ring_enqueue_burst(q->processed_pkts, (void **)ops, nb_ops,
1040 enqueue_enc_ops(struct rte_bbdev_queue_data *q_data,
1041 struct rte_bbdev_enc_op **ops, uint16_t nb_ops)
1043 void *queue = q_data->queue_private;
1044 struct turbo_sw_queue *q = queue;
1045 uint16_t nb_enqueued = 0;
1047 nb_enqueued = enqueue_enc_all_ops(q, ops, nb_ops);
1049 q_data->queue_stats.enqueue_err_count += nb_ops - nb_enqueued;
1050 q_data->queue_stats.enqueued_count += nb_enqueued;
1057 enqueue_dec_ops(struct rte_bbdev_queue_data *q_data,
1058 struct rte_bbdev_dec_op **ops, uint16_t nb_ops)
1060 void *queue = q_data->queue_private;
1061 struct turbo_sw_queue *q = queue;
1062 uint16_t nb_enqueued = 0;
1064 nb_enqueued = enqueue_dec_all_ops(q, ops, nb_ops);
1066 q_data->queue_stats.enqueue_err_count += nb_ops - nb_enqueued;
1067 q_data->queue_stats.enqueued_count += nb_enqueued;
1072 /* Dequeue decode burst */
1074 dequeue_dec_ops(struct rte_bbdev_queue_data *q_data,
1075 struct rte_bbdev_dec_op **ops, uint16_t nb_ops)
1077 struct turbo_sw_queue *q = q_data->queue_private;
1078 uint16_t nb_dequeued = rte_ring_dequeue_burst(q->processed_pkts,
1079 (void **)ops, nb_ops, NULL);
1080 q_data->queue_stats.dequeued_count += nb_dequeued;
1085 /* Dequeue encode burst */
1087 dequeue_enc_ops(struct rte_bbdev_queue_data *q_data,
1088 struct rte_bbdev_enc_op **ops, uint16_t nb_ops)
1090 struct turbo_sw_queue *q = q_data->queue_private;
1091 uint16_t nb_dequeued = rte_ring_dequeue_burst(q->processed_pkts,
1092 (void **)ops, nb_ops, NULL);
1093 q_data->queue_stats.dequeued_count += nb_dequeued;
1098 /* Parse 16bit integer from string argument */
1100 parse_u16_arg(const char *key, const char *value, void *extra_args)
1102 uint16_t *u16 = extra_args;
1103 unsigned int long result;
1105 if ((value == NULL) || (extra_args == NULL))
1108 result = strtoul(value, NULL, 0);
1109 if ((result >= (1 << 16)) || (errno != 0)) {
1110 rte_bbdev_log(ERR, "Invalid value %lu for %s", result, key);
1113 *u16 = (uint16_t)result;
1117 /* Parse parameters used to create device */
1119 parse_turbo_sw_params(struct turbo_sw_params *params, const char *input_args)
1121 struct rte_kvargs *kvlist = NULL;
1127 kvlist = rte_kvargs_parse(input_args, turbo_sw_valid_params);
1131 ret = rte_kvargs_process(kvlist, turbo_sw_valid_params[0],
1132 &parse_u16_arg, ¶ms->queues_num);
1136 ret = rte_kvargs_process(kvlist, turbo_sw_valid_params[1],
1137 &parse_u16_arg, ¶ms->socket_id);
1141 if (params->socket_id >= RTE_MAX_NUMA_NODES) {
1142 rte_bbdev_log(ERR, "Invalid socket, must be < %u",
1143 RTE_MAX_NUMA_NODES);
1150 rte_kvargs_free(kvlist);
1156 turbo_sw_bbdev_create(struct rte_vdev_device *vdev,
1157 struct turbo_sw_params *init_params)
1159 struct rte_bbdev *bbdev;
1160 const char *name = rte_vdev_device_name(vdev);
1162 bbdev = rte_bbdev_allocate(name);
1166 bbdev->data->dev_private = rte_zmalloc_socket(name,
1167 sizeof(struct bbdev_private), RTE_CACHE_LINE_SIZE,
1168 init_params->socket_id);
1169 if (bbdev->data->dev_private == NULL) {
1170 rte_bbdev_release(bbdev);
1174 bbdev->dev_ops = &pmd_ops;
1175 bbdev->device = &vdev->device;
1176 bbdev->data->socket_id = init_params->socket_id;
1177 bbdev->intr_handle = NULL;
1179 /* register rx/tx burst functions for data path */
1180 bbdev->dequeue_enc_ops = dequeue_enc_ops;
1181 bbdev->dequeue_dec_ops = dequeue_dec_ops;
1182 bbdev->enqueue_enc_ops = enqueue_enc_ops;
1183 bbdev->enqueue_dec_ops = enqueue_dec_ops;
1184 ((struct bbdev_private *) bbdev->data->dev_private)->max_nb_queues =
1185 init_params->queues_num;
1190 /* Initialise device */
1192 turbo_sw_bbdev_probe(struct rte_vdev_device *vdev)
1194 struct turbo_sw_params init_params = {
1196 RTE_BBDEV_DEFAULT_MAX_NB_QUEUES
1199 const char *input_args;
1204 name = rte_vdev_device_name(vdev);
1207 input_args = rte_vdev_device_args(vdev);
1208 parse_turbo_sw_params(&init_params, input_args);
1210 rte_bbdev_log_debug(
1211 "Initialising %s on NUMA node %d with max queues: %d\n",
1212 name, init_params.socket_id, init_params.queues_num);
1214 return turbo_sw_bbdev_create(vdev, &init_params);
1217 /* Uninitialise device */
1219 turbo_sw_bbdev_remove(struct rte_vdev_device *vdev)
1221 struct rte_bbdev *bbdev;
1227 name = rte_vdev_device_name(vdev);
1231 bbdev = rte_bbdev_get_named_dev(name);
1235 rte_free(bbdev->data->dev_private);
1237 return rte_bbdev_release(bbdev);
1240 static struct rte_vdev_driver bbdev_turbo_sw_pmd_drv = {
1241 .probe = turbo_sw_bbdev_probe,
1242 .remove = turbo_sw_bbdev_remove
1245 RTE_PMD_REGISTER_VDEV(DRIVER_NAME, bbdev_turbo_sw_pmd_drv);
1246 RTE_PMD_REGISTER_PARAM_STRING(DRIVER_NAME,
1247 TURBO_SW_MAX_NB_QUEUES_ARG"=<int> "
1248 TURBO_SW_SOCKET_ID_ARG"=<int>");
1250 RTE_INIT(null_bbdev_init_log);
1252 null_bbdev_init_log(void)
1254 bbdev_turbo_sw_logtype = rte_log_register("pmd.bb.turbo_sw");
1255 if (bbdev_turbo_sw_logtype >= 0)
1256 rte_log_set_level(bbdev_turbo_sw_logtype, RTE_LOG_NOTICE);