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>
12 #include <rte_cycles.h>
14 #include <rte_bbdev.h>
15 #include <rte_bbdev_pmd.h>
17 #include <phy_turbo.h>
19 #include <phy_rate_match.h>
22 #define DRIVER_NAME turbo_sw
24 /* Turbo SW PMD logging ID */
25 static int bbdev_turbo_sw_logtype;
27 /* Helper macro for logging */
28 #define rte_bbdev_log(level, fmt, ...) \
29 rte_log(RTE_LOG_ ## level, bbdev_turbo_sw_logtype, fmt "\n", \
32 #define rte_bbdev_log_debug(fmt, ...) \
33 rte_bbdev_log(DEBUG, RTE_STR(__LINE__) ":%s() " fmt, __func__, \
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)
40 /* private data structure */
41 struct bbdev_private {
42 unsigned int max_nb_queues; /**< Max number of queues */
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 */
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"
55 static const char * const turbo_sw_valid_params[] = {
56 TURBO_SW_MAX_NB_QUEUES_ARG,
57 TURBO_SW_SOCKET_ID_ARG
61 struct turbo_sw_queue {
62 /* Ring for processed (encoded/decoded) operations which are ready to
65 struct rte_ring *processed_pkts;
66 /* Stores input for turbo encoder (used when CRC attachment is
70 /* Stores output from turbo encoder */
72 /* Alpha gamma buf for bblib_turbo_decoder() function */
74 /* Temp buf for bblib_turbo_decoder() function */
76 /* Input buf for bblib_rate_dematching_lte() function */
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;
86 /* Calculate index based on Table 5.1.3-3 from TS34.212 */
88 compute_idx(uint16_t k)
92 if (k < RTE_BBDEV_MIN_CB_SIZE || k > RTE_BBDEV_MAX_CB_SIZE)
96 if ((k - 2048) % 64 != 0)
99 result = 124 + (k - 2048) / 64;
100 } else if (k <= 512) {
101 if ((k - 40) % 8 != 0)
104 result = (k - 40) / 8 + 1;
105 } else if (k <= 1024) {
106 if ((k - 512) % 16 != 0)
109 result = 60 + (k - 512) / 16;
110 } else { /* 1024 < k <= 2048 */
111 if ((k - 1024) % 32 != 0)
114 result = 92 + (k - 1024) / 32;
120 /* Read flag value 0/1 from bitmap */
122 check_bit(uint32_t bitmap, uint32_t bitmask)
124 return bitmap & bitmask;
127 /* Get device info */
129 info_get(struct rte_bbdev *dev, struct rte_bbdev_driver_info *dev_info)
131 struct bbdev_private *internals = dev->data->dev_private;
133 static const struct rte_bbdev_op_cap bbdev_capabilities[] = {
135 .type = RTE_BBDEV_OP_TURBO_DEC,
138 RTE_BBDEV_TURBO_SUBBLOCK_DEINTERLEAVE |
139 RTE_BBDEV_TURBO_POS_LLR_1_BIT_IN |
140 RTE_BBDEV_TURBO_NEG_LLR_1_BIT_IN |
141 RTE_BBDEV_TURBO_CRC_TYPE_24B |
142 RTE_BBDEV_TURBO_DEC_TB_CRC_24B_KEEP |
143 RTE_BBDEV_TURBO_EARLY_TERMINATION,
144 .max_llr_modulus = 16,
145 .num_buffers_src = RTE_BBDEV_MAX_CODE_BLOCKS,
146 .num_buffers_hard_out =
147 RTE_BBDEV_MAX_CODE_BLOCKS,
148 .num_buffers_soft_out = 0,
152 .type = RTE_BBDEV_OP_TURBO_ENC,
155 RTE_BBDEV_TURBO_CRC_24B_ATTACH |
156 RTE_BBDEV_TURBO_CRC_24A_ATTACH |
157 RTE_BBDEV_TURBO_RATE_MATCH |
158 RTE_BBDEV_TURBO_RV_INDEX_BYPASS,
159 .num_buffers_src = RTE_BBDEV_MAX_CODE_BLOCKS,
160 .num_buffers_dst = RTE_BBDEV_MAX_CODE_BLOCKS,
163 RTE_BBDEV_END_OF_CAPABILITIES_LIST()
166 static struct rte_bbdev_queue_conf default_queue_conf = {
167 .queue_size = RTE_BBDEV_QUEUE_SIZE_LIMIT,
170 static const enum rte_cpu_flag_t cpu_flag = RTE_CPUFLAG_SSE4_2;
172 default_queue_conf.socket = dev->data->socket_id;
174 dev_info->driver_name = RTE_STR(DRIVER_NAME);
175 dev_info->max_num_queues = internals->max_nb_queues;
176 dev_info->queue_size_lim = RTE_BBDEV_QUEUE_SIZE_LIMIT;
177 dev_info->hardware_accelerated = false;
178 dev_info->max_queue_priority = 0;
179 dev_info->default_queue_conf = default_queue_conf;
180 dev_info->capabilities = bbdev_capabilities;
181 dev_info->cpu_flag_reqs = &cpu_flag;
182 dev_info->min_alignment = 64;
184 rte_bbdev_log_debug("got device info from %u\n", dev->data->dev_id);
189 q_release(struct rte_bbdev *dev, uint16_t q_id)
191 struct turbo_sw_queue *q = dev->data->queues[q_id].queue_private;
194 rte_ring_free(q->processed_pkts);
195 rte_free(q->enc_out);
198 rte_free(q->code_block);
199 rte_free(q->deint_input);
200 rte_free(q->deint_output);
201 rte_free(q->adapter_output);
203 dev->data->queues[q_id].queue_private = NULL;
206 rte_bbdev_log_debug("released device queue %u:%u",
207 dev->data->dev_id, q_id);
213 q_setup(struct rte_bbdev *dev, uint16_t q_id,
214 const struct rte_bbdev_queue_conf *queue_conf)
217 struct turbo_sw_queue *q;
218 char name[RTE_RING_NAMESIZE];
220 /* Allocate the queue data structure. */
221 q = rte_zmalloc_socket(RTE_STR(DRIVER_NAME), sizeof(*q),
222 RTE_CACHE_LINE_SIZE, queue_conf->socket);
224 rte_bbdev_log(ERR, "Failed to allocate queue memory");
228 /* Allocate memory for encoder output. */
229 ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"_enc_out%u:%u",
230 dev->data->dev_id, q_id);
231 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
233 "Creating queue name for device %u queue %u failed",
234 dev->data->dev_id, q_id);
235 return -ENAMETOOLONG;
237 q->enc_out = rte_zmalloc_socket(name,
238 ((RTE_BBDEV_MAX_TB_SIZE >> 3) + 3) *
239 sizeof(*q->enc_out) * 3,
240 RTE_CACHE_LINE_SIZE, queue_conf->socket);
241 if (q->enc_out == NULL) {
243 "Failed to allocate queue memory for %s", name);
247 /* Allocate memory for rate matching output. */
248 ret = snprintf(name, RTE_RING_NAMESIZE,
249 RTE_STR(DRIVER_NAME)"_enc_in%u:%u", dev->data->dev_id,
251 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
253 "Creating queue name for device %u queue %u failed",
254 dev->data->dev_id, q_id);
255 return -ENAMETOOLONG;
257 q->enc_in = rte_zmalloc_socket(name,
258 (RTE_BBDEV_MAX_CB_SIZE >> 3) * sizeof(*q->enc_in),
259 RTE_CACHE_LINE_SIZE, queue_conf->socket);
260 if (q->enc_in == NULL) {
262 "Failed to allocate queue memory for %s", name);
266 /* Allocate memory for Aplha Gamma temp buffer. */
267 ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"_ag%u:%u",
268 dev->data->dev_id, q_id);
269 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
271 "Creating queue name for device %u queue %u failed",
272 dev->data->dev_id, q_id);
273 return -ENAMETOOLONG;
275 q->ag = rte_zmalloc_socket(name,
276 RTE_BBDEV_MAX_CB_SIZE * 10 * sizeof(*q->ag),
277 RTE_CACHE_LINE_SIZE, queue_conf->socket);
280 "Failed to allocate queue memory for %s", name);
284 /* Allocate memory for code block temp buffer. */
285 ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"_cb%u:%u",
286 dev->data->dev_id, q_id);
287 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
289 "Creating queue name for device %u queue %u failed",
290 dev->data->dev_id, q_id);
291 return -ENAMETOOLONG;
293 q->code_block = rte_zmalloc_socket(name,
294 RTE_BBDEV_MAX_CB_SIZE * sizeof(*q->code_block),
295 RTE_CACHE_LINE_SIZE, queue_conf->socket);
296 if (q->code_block == NULL) {
298 "Failed to allocate queue memory for %s", name);
302 /* Allocate memory for Deinterleaver input. */
303 ret = snprintf(name, RTE_RING_NAMESIZE,
304 RTE_STR(DRIVER_NAME)"_deint_input%u:%u",
305 dev->data->dev_id, q_id);
306 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
308 "Creating queue name for device %u queue %u failed",
309 dev->data->dev_id, q_id);
310 return -ENAMETOOLONG;
312 q->deint_input = rte_zmalloc_socket(name,
313 DEINT_INPUT_BUF_SIZE * sizeof(*q->deint_input),
314 RTE_CACHE_LINE_SIZE, queue_conf->socket);
315 if (q->deint_input == NULL) {
317 "Failed to allocate queue memory for %s", name);
321 /* Allocate memory for Deinterleaver output. */
322 ret = snprintf(name, RTE_RING_NAMESIZE,
323 RTE_STR(DRIVER_NAME)"_deint_output%u:%u",
324 dev->data->dev_id, q_id);
325 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
327 "Creating queue name for device %u queue %u failed",
328 dev->data->dev_id, q_id);
329 return -ENAMETOOLONG;
331 q->deint_output = rte_zmalloc_socket(NULL,
332 DEINT_OUTPUT_BUF_SIZE * sizeof(*q->deint_output),
333 RTE_CACHE_LINE_SIZE, queue_conf->socket);
334 if (q->deint_output == NULL) {
336 "Failed to allocate queue memory for %s", name);
340 /* Allocate memory for Adapter output. */
341 ret = snprintf(name, RTE_RING_NAMESIZE,
342 RTE_STR(DRIVER_NAME)"_adapter_output%u:%u",
343 dev->data->dev_id, q_id);
344 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
346 "Creating queue name for device %u queue %u failed",
347 dev->data->dev_id, q_id);
348 return -ENAMETOOLONG;
350 q->adapter_output = rte_zmalloc_socket(NULL,
351 ADAPTER_OUTPUT_BUF_SIZE * sizeof(*q->adapter_output),
352 RTE_CACHE_LINE_SIZE, queue_conf->socket);
353 if (q->adapter_output == NULL) {
355 "Failed to allocate queue memory for %s", name);
359 /* Create ring for packets awaiting to be dequeued. */
360 ret = snprintf(name, RTE_RING_NAMESIZE, RTE_STR(DRIVER_NAME)"%u:%u",
361 dev->data->dev_id, q_id);
362 if ((ret < 0) || (ret >= (int)RTE_RING_NAMESIZE)) {
364 "Creating queue name for device %u queue %u failed",
365 dev->data->dev_id, q_id);
366 return -ENAMETOOLONG;
368 q->processed_pkts = rte_ring_create(name, queue_conf->queue_size,
369 queue_conf->socket, RING_F_SP_ENQ | RING_F_SC_DEQ);
370 if (q->processed_pkts == NULL) {
371 rte_bbdev_log(ERR, "Failed to create ring for %s", name);
375 q->type = queue_conf->op_type;
377 dev->data->queues[q_id].queue_private = q;
378 rte_bbdev_log_debug("setup device queue %s", name);
382 rte_ring_free(q->processed_pkts);
383 rte_free(q->enc_out);
386 rte_free(q->code_block);
387 rte_free(q->deint_input);
388 rte_free(q->deint_output);
389 rte_free(q->adapter_output);
394 static const struct rte_bbdev_ops pmd_ops = {
395 .info_get = info_get,
396 .queue_setup = q_setup,
397 .queue_release = q_release
400 /* Checks if the encoder input buffer is correct.
401 * Returns 0 if it's valid, -1 otherwise.
404 is_enc_input_valid(const uint16_t k, const int32_t k_idx,
405 const uint16_t in_length)
408 rte_bbdev_log(ERR, "K Index is invalid");
412 if (in_length - (k >> 3) < 0) {
414 "Mismatch between input length (%u bytes) and K (%u bits)",
419 if (k > RTE_BBDEV_MAX_CB_SIZE) {
420 rte_bbdev_log(ERR, "CB size (%u) is too big, max: %d",
421 k, RTE_BBDEV_MAX_CB_SIZE);
428 /* Checks if the decoder input buffer is correct.
429 * Returns 0 if it's valid, -1 otherwise.
432 is_dec_input_valid(int32_t k_idx, int16_t kw, int16_t in_length)
435 rte_bbdev_log(ERR, "K index is invalid");
439 if (in_length - kw < 0) {
441 "Mismatch between input length (%u) and kw (%u)",
446 if (kw > RTE_BBDEV_MAX_KW) {
447 rte_bbdev_log(ERR, "Input length (%u) is too big, max: %d",
448 kw, RTE_BBDEV_MAX_KW);
456 process_enc_cb(struct turbo_sw_queue *q, struct rte_bbdev_enc_op *op,
457 uint8_t r, uint8_t c, uint16_t k, uint16_t ncb,
458 uint32_t e, struct rte_mbuf *m_in, struct rte_mbuf *m_out,
459 uint16_t in_offset, uint16_t out_offset, uint16_t total_left,
460 struct rte_bbdev_stats *q_stats)
465 uint8_t *in, *out0, *out1, *out2, *tmp_out, *rm_out;
466 uint64_t first_3_bytes = 0;
467 struct rte_bbdev_op_turbo_enc *enc = &op->turbo_enc;
468 struct bblib_crc_request crc_req;
469 struct bblib_crc_response crc_resp;
470 struct bblib_turbo_encoder_request turbo_req;
471 struct bblib_turbo_encoder_response turbo_resp;
472 struct bblib_rate_match_dl_request rm_req;
473 struct bblib_rate_match_dl_response rm_resp;
474 #ifdef RTE_BBDEV_OFFLOAD_COST
477 RTE_SET_USED(q_stats);
480 k_idx = compute_idx(k);
481 in = rte_pktmbuf_mtod_offset(m_in, uint8_t *, in_offset);
483 /* CRC24A (for TB) */
484 if ((enc->op_flags & RTE_BBDEV_TURBO_CRC_24A_ATTACH) &&
485 (enc->code_block_mode == 1)) {
486 ret = is_enc_input_valid(k - 24, k_idx, total_left);
488 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
492 crc_req.len = (k - 24) >> 3;
493 /* Check if there is a room for CRC bits. If not use
494 * the temporary buffer.
496 if (rte_pktmbuf_append(m_in, 3) == NULL) {
497 rte_memcpy(q->enc_in, in, (k - 24) >> 3);
500 /* Store 3 first bytes of next CB as they will be
501 * overwritten by CRC bytes. If it is the last CB then
502 * there is no point to store 3 next bytes and this
503 * if..else branch will be omitted.
505 first_3_bytes = *((uint64_t *)&in[(k - 32) >> 3]);
509 #ifdef RTE_BBDEV_OFFLOAD_COST
510 start_time = rte_rdtsc_precise();
512 bblib_lte_crc24a_gen(&crc_req, &crc_resp);
513 #ifdef RTE_BBDEV_OFFLOAD_COST
514 q_stats->offload_time += rte_rdtsc_precise() - start_time;
516 } else if (enc->op_flags & RTE_BBDEV_TURBO_CRC_24B_ATTACH) {
518 ret = is_enc_input_valid(k - 24, k_idx, total_left);
520 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
524 crc_req.len = (k - 24) >> 3;
525 /* Check if there is a room for CRC bits. If this is the last
526 * CB in TB. If not use temporary buffer.
528 if ((c - r == 1) && (rte_pktmbuf_append(m_in, 3) == NULL)) {
529 rte_memcpy(q->enc_in, in, (k - 24) >> 3);
531 } else if (c - r > 1) {
532 /* Store 3 first bytes of next CB as they will be
533 * overwritten by CRC bytes. If it is the last CB then
534 * there is no point to store 3 next bytes and this
535 * if..else branch will be omitted.
537 first_3_bytes = *((uint64_t *)&in[(k - 32) >> 3]);
541 #ifdef RTE_BBDEV_OFFLOAD_COST
542 start_time = rte_rdtsc_precise();
544 bblib_lte_crc24b_gen(&crc_req, &crc_resp);
545 #ifdef RTE_BBDEV_OFFLOAD_COST
546 q_stats->offload_time += rte_rdtsc_precise() - start_time;
549 ret = is_enc_input_valid(k, k_idx, total_left);
551 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
558 /* Each bit layer output from turbo encoder is (k+4) bits long, i.e.
559 * input length + 4 tail bits. That's (k/8) + 1 bytes after rounding up.
560 * So dst_data's length should be 3*(k/8) + 3 bytes.
561 * In Rate-matching bypass case outputs pointers passed to encoder
562 * (out0, out1 and out2) can directly point to addresses of output from
565 if (enc->op_flags & RTE_BBDEV_TURBO_RATE_MATCH) {
567 out1 = RTE_PTR_ADD(out0, (k >> 3) + 1);
568 out2 = RTE_PTR_ADD(out1, (k >> 3) + 1);
570 out0 = (uint8_t *)rte_pktmbuf_append(m_out, (k >> 3) * 3 + 2);
572 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
574 "Too little space in output mbuf");
577 enc->output.length += (k >> 3) * 3 + 2;
578 /* rte_bbdev_op_data.offset can be different than the
579 * offset of the appended bytes
581 out0 = rte_pktmbuf_mtod_offset(m_out, uint8_t *, out_offset);
582 out1 = rte_pktmbuf_mtod_offset(m_out, uint8_t *,
583 out_offset + (k >> 3) + 1);
584 out2 = rte_pktmbuf_mtod_offset(m_out, uint8_t *,
585 out_offset + 2 * ((k >> 3) + 1));
588 turbo_req.case_id = k_idx;
589 turbo_req.input_win = in;
590 turbo_req.length = k >> 3;
591 turbo_resp.output_win_0 = out0;
592 turbo_resp.output_win_1 = out1;
593 turbo_resp.output_win_2 = out2;
595 #ifdef RTE_BBDEV_OFFLOAD_COST
596 start_time = rte_rdtsc_precise();
599 if (bblib_turbo_encoder(&turbo_req, &turbo_resp) != 0) {
600 op->status |= 1 << RTE_BBDEV_DRV_ERROR;
601 rte_bbdev_log(ERR, "Turbo Encoder failed");
605 #ifdef RTE_BBDEV_OFFLOAD_COST
606 q_stats->offload_time += rte_rdtsc_precise() - start_time;
609 /* Restore 3 first bytes of next CB if they were overwritten by CRC*/
610 if (first_3_bytes != 0)
611 *((uint64_t *)&in[(k - 32) >> 3]) = first_3_bytes;
614 if (enc->op_flags & RTE_BBDEV_TURBO_RATE_MATCH) {
616 /* Integer round up division by 8 */
617 uint16_t out_len = (e + 7) >> 3;
618 /* The mask array is indexed using E%8. E is an even number so
619 * there are only 4 possible values.
621 const uint8_t mask_out[] = {0xFF, 0xC0, 0xF0, 0xFC};
623 /* get output data starting address */
624 rm_out = (uint8_t *)rte_pktmbuf_append(m_out, out_len);
625 if (rm_out == NULL) {
626 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
628 "Too little space in output mbuf");
631 /* rte_bbdev_op_data.offset can be different than the offset
632 * of the appended bytes
634 rm_out = rte_pktmbuf_mtod_offset(m_out, uint8_t *, out_offset);
636 /* index of current code block */
638 /* total number of code block */
640 /* For DL - 1, UL - 0 */
641 rm_req.direction = 1;
642 /* According to 3ggp 36.212 Spec 5.1.4.1.2 section Nsoft, KMIMO
643 * and MDL_HARQ are used for Ncb calculation. As Ncb is already
644 * known we can adjust those parameters
646 rm_req.Nsoft = ncb * rm_req.C;
649 /* According to 3ggp 36.212 Spec 5.1.4.1.2 section Nl, Qm and G
650 * are used for E calculation. As E is already known we can
651 * adjust those parameters
655 rm_req.G = rm_req.NL * rm_req.Qm * rm_req.C;
657 rm_req.rvidx = enc->rv_index;
658 rm_req.Kidx = k_idx - 1;
663 rm_resp.output = rm_out;
664 rm_resp.OutputLen = out_len;
665 if (enc->op_flags & RTE_BBDEV_TURBO_RV_INDEX_BYPASS)
666 rm_req.bypass_rvidx = 1;
668 rm_req.bypass_rvidx = 0;
670 #ifdef RTE_BBDEV_OFFLOAD_COST
671 start_time = rte_rdtsc_precise();
674 if (bblib_rate_match_dl(&rm_req, &rm_resp) != 0) {
675 op->status |= 1 << RTE_BBDEV_DRV_ERROR;
676 rte_bbdev_log(ERR, "Rate matching failed");
680 /* SW fills an entire last byte even if E%8 != 0. Clear the
681 * superfluous data bits for consistency with HW device.
683 mask_id = (e & 7) >> 1;
684 rm_out[out_len - 1] &= mask_out[mask_id];
686 #ifdef RTE_BBDEV_OFFLOAD_COST
687 q_stats->offload_time += rte_rdtsc_precise() - start_time;
690 enc->output.length += rm_resp.OutputLen;
692 /* Rate matching is bypassed */
694 /* Completing last byte of out0 (where 4 tail bits are stored)
695 * by moving first 4 bits from out1
697 tmp_out = (uint8_t *) --out1;
698 *tmp_out = *tmp_out | ((*(tmp_out + 1) & 0xF0) >> 4);
700 /* Shifting out1 data by 4 bits to the left */
701 for (m = 0; m < k >> 3; ++m) {
702 uint8_t *first = tmp_out;
703 uint8_t second = *(tmp_out + 1);
704 *first = (*first << 4) | ((second & 0xF0) >> 4);
707 /* Shifting out2 data by 8 bits to the left */
708 for (m = 0; m < (k >> 3) + 1; ++m) {
709 *tmp_out = *(tmp_out + 1);
717 enqueue_enc_one_op(struct turbo_sw_queue *q, struct rte_bbdev_enc_op *op,
718 struct rte_bbdev_stats *queue_stats)
720 uint8_t c, r, crc24_bits = 0;
723 struct rte_bbdev_op_turbo_enc *enc = &op->turbo_enc;
724 uint16_t in_offset = enc->input.offset;
725 uint16_t out_offset = enc->output.offset;
726 struct rte_mbuf *m_in = enc->input.data;
727 struct rte_mbuf *m_out = enc->output.data;
728 uint16_t total_left = enc->input.length;
730 /* Clear op status */
733 if (total_left > RTE_BBDEV_MAX_TB_SIZE >> 3) {
734 rte_bbdev_log(ERR, "TB size (%u) is too big, max: %d",
735 total_left, RTE_BBDEV_MAX_TB_SIZE);
736 op->status = 1 << RTE_BBDEV_DATA_ERROR;
740 if (m_in == NULL || m_out == NULL) {
741 rte_bbdev_log(ERR, "Invalid mbuf pointer");
742 op->status = 1 << RTE_BBDEV_DATA_ERROR;
746 if ((enc->op_flags & RTE_BBDEV_TURBO_CRC_24B_ATTACH) ||
747 (enc->op_flags & RTE_BBDEV_TURBO_CRC_24A_ATTACH))
750 if (enc->code_block_mode == 0) { /* For Transport Block mode */
751 c = enc->tb_params.c;
752 r = enc->tb_params.r;
753 } else {/* For Code Block mode */
758 while (total_left > 0 && r < c) {
759 if (enc->code_block_mode == 0) {
760 k = (r < enc->tb_params.c_neg) ?
761 enc->tb_params.k_neg : enc->tb_params.k_pos;
762 ncb = (r < enc->tb_params.c_neg) ?
763 enc->tb_params.ncb_neg : enc->tb_params.ncb_pos;
764 e = (r < enc->tb_params.cab) ?
765 enc->tb_params.ea : enc->tb_params.eb;
767 k = enc->cb_params.k;
768 ncb = enc->cb_params.ncb;
769 e = enc->cb_params.e;
772 process_enc_cb(q, op, r, c, k, ncb, e, m_in,
773 m_out, in_offset, out_offset, total_left,
775 /* Update total_left */
776 total_left -= (k - crc24_bits) >> 3;
777 /* Update offsets for next CBs (if exist) */
778 in_offset += (k - crc24_bits) >> 3;
779 if (enc->op_flags & RTE_BBDEV_TURBO_RATE_MATCH)
780 out_offset += e >> 3;
782 out_offset += (k >> 3) * 3 + 2;
786 /* check if all input data was processed */
787 if (total_left != 0) {
788 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
790 "Mismatch between mbuf length and included CBs sizes");
794 static inline uint16_t
795 enqueue_enc_all_ops(struct turbo_sw_queue *q, struct rte_bbdev_enc_op **ops,
796 uint16_t nb_ops, struct rte_bbdev_stats *queue_stats)
799 #ifdef RTE_BBDEV_OFFLOAD_COST
800 queue_stats->offload_time = 0;
803 for (i = 0; i < nb_ops; ++i)
804 enqueue_enc_one_op(q, ops[i], queue_stats);
806 return rte_ring_enqueue_burst(q->processed_pkts, (void **)ops, nb_ops,
810 /* Remove the padding bytes from a cyclic buffer.
811 * The input buffer is a data stream wk as described in 3GPP TS 36.212 section
812 * 5.1.4.1.2 starting from w0 and with length Ncb bytes.
813 * The output buffer is a data stream wk with pruned padding bytes. It's length
814 * is 3*D bytes and the order of non-padding bytes is preserved.
817 remove_nulls_from_circular_buf(const uint8_t *in, uint8_t *out, uint16_t k,
820 uint32_t in_idx, out_idx, c_idx;
821 const uint32_t d = k + 4;
822 const uint32_t kw = (ncb / 3);
823 const uint32_t nd = kw - d;
824 const uint32_t r_subblock = kw / RTE_BBDEV_C_SUBBLOCK;
825 /* Inter-column permutation pattern */
826 const uint32_t P[RTE_BBDEV_C_SUBBLOCK] = {0, 16, 8, 24, 4, 20, 12, 28,
827 2, 18, 10, 26, 6, 22, 14, 30, 1, 17, 9, 25, 5, 21, 13,
828 29, 3, 19, 11, 27, 7, 23, 15, 31};
832 /* The padding bytes are at the first Nd positions in the first row. */
833 for (c_idx = 0; in_idx < kw; in_idx += r_subblock, ++c_idx) {
835 rte_memcpy(&out[out_idx], &in[in_idx + 1],
837 out_idx += r_subblock - 1;
839 rte_memcpy(&out[out_idx], &in[in_idx], r_subblock);
840 out_idx += r_subblock;
844 /* First and second parity bits sub-blocks are interlaced. */
845 for (c_idx = 0; in_idx < ncb - 2 * r_subblock;
846 in_idx += 2 * r_subblock, ++c_idx) {
847 uint32_t second_block_c_idx = P[c_idx];
848 uint32_t third_block_c_idx = P[c_idx] + 1;
850 if (second_block_c_idx < nd && third_block_c_idx < nd) {
851 rte_memcpy(&out[out_idx], &in[in_idx + 2],
853 out_idx += 2 * r_subblock - 2;
854 } else if (second_block_c_idx >= nd &&
855 third_block_c_idx >= nd) {
856 rte_memcpy(&out[out_idx], &in[in_idx], 2 * r_subblock);
857 out_idx += 2 * r_subblock;
858 } else if (second_block_c_idx < nd) {
859 out[out_idx++] = in[in_idx];
860 rte_memcpy(&out[out_idx], &in[in_idx + 2],
862 out_idx += 2 * r_subblock - 2;
864 rte_memcpy(&out[out_idx], &in[in_idx + 1],
866 out_idx += 2 * r_subblock - 1;
870 /* Last interlaced row is different - its last byte is the only padding
871 * byte. We can have from 4 up to 28 padding bytes (Nd) per sub-block.
872 * After interlacing the 1st and 2nd parity sub-blocks we can have 0, 1
873 * or 2 padding bytes each time we make a step of 2 * R_SUBBLOCK bytes
874 * (moving to another column). 2nd parity sub-block uses the same
875 * inter-column permutation pattern as the systematic and 1st parity
876 * sub-blocks but it adds '1' to the resulting index and calculates the
877 * modulus of the result and Kw. Last column is mapped to itself (id 31)
878 * so the first byte taken from the 2nd parity sub-block will be the
879 * 32nd (31+1) byte, then 64th etc. (step is C_SUBBLOCK == 32) and the
880 * last byte will be the first byte from the sub-block:
881 * (32 + 32 * (R_SUBBLOCK-1)) % Kw == Kw % Kw == 0. Nd can't be smaller
882 * than 4 so we know that bytes with ids 0, 1, 2 and 3 must be the
883 * padding bytes. The bytes from the 1st parity sub-block are the bytes
884 * from the 31st column - Nd can't be greater than 28 so we are sure
885 * that there are no padding bytes in 31st column.
887 rte_memcpy(&out[out_idx], &in[in_idx], 2 * r_subblock - 1);
891 move_padding_bytes(const uint8_t *in, uint8_t *out, uint16_t k,
895 uint16_t kpi = ncb / 3;
896 uint16_t nd = kpi - d;
898 rte_memcpy(&out[nd], in, d);
899 rte_memcpy(&out[nd + kpi + 64], &in[kpi], d);
900 rte_memcpy(&out[(nd - 1) + 2 * (kpi + 64)], &in[2 * kpi], d);
904 process_dec_cb(struct turbo_sw_queue *q, struct rte_bbdev_dec_op *op,
905 uint8_t c, uint16_t k, uint16_t kw, struct rte_mbuf *m_in,
906 struct rte_mbuf *m_out, uint16_t in_offset, uint16_t out_offset,
907 bool check_crc_24b, uint16_t crc24_overlap, uint16_t total_left)
912 uint8_t *in, *out, *adapter_input;
913 int32_t ncb, ncb_without_null;
914 struct bblib_turbo_adapter_ul_response adapter_resp;
915 struct bblib_turbo_adapter_ul_request adapter_req;
916 struct bblib_turbo_decoder_request turbo_req;
917 struct bblib_turbo_decoder_response turbo_resp;
918 struct rte_bbdev_op_turbo_dec *dec = &op->turbo_dec;
920 k_idx = compute_idx(k);
922 ret = is_dec_input_valid(k_idx, kw, total_left);
924 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
928 in = rte_pktmbuf_mtod_offset(m_in, uint8_t *, in_offset);
930 ncb_without_null = (k + 4) * 3;
932 if (check_bit(dec->op_flags, RTE_BBDEV_TURBO_SUBBLOCK_DEINTERLEAVE)) {
933 struct bblib_deinterleave_ul_request deint_req;
934 struct bblib_deinterleave_ul_response deint_resp;
936 /* SW decoder accepts only a circular buffer without NULL bytes
937 * so the input needs to be converted.
939 remove_nulls_from_circular_buf(in, q->deint_input, k, ncb);
941 deint_req.pharqbuffer = q->deint_input;
942 deint_req.ncb = ncb_without_null;
943 deint_resp.pinteleavebuffer = q->deint_output;
944 bblib_deinterleave_ul(&deint_req, &deint_resp);
946 move_padding_bytes(in, q->deint_output, k, ncb);
948 adapter_input = q->deint_output;
950 if (dec->op_flags & RTE_BBDEV_TURBO_POS_LLR_1_BIT_IN)
951 adapter_req.isinverted = 1;
952 else if (dec->op_flags & RTE_BBDEV_TURBO_NEG_LLR_1_BIT_IN)
953 adapter_req.isinverted = 0;
955 op->status |= 1 << RTE_BBDEV_DRV_ERROR;
956 rte_bbdev_log(ERR, "LLR format wasn't specified");
960 adapter_req.ncb = ncb_without_null;
961 adapter_req.pinteleavebuffer = adapter_input;
962 adapter_resp.pharqout = q->adapter_output;
963 bblib_turbo_adapter_ul(&adapter_req, &adapter_resp);
965 out = (uint8_t *)rte_pktmbuf_append(m_out, ((k - crc24_overlap) >> 3));
967 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
968 rte_bbdev_log(ERR, "Too little space in output mbuf");
971 /* rte_bbdev_op_data.offset can be different than the offset of the
974 out = rte_pktmbuf_mtod_offset(m_out, uint8_t *, out_offset);
979 turbo_req.input = (int8_t *)q->adapter_output;
981 turbo_req.k_idx = k_idx;
982 turbo_req.max_iter_num = dec->iter_max;
983 turbo_req.early_term_disable = !check_bit(dec->op_flags,
984 RTE_BBDEV_TURBO_EARLY_TERMINATION);
985 turbo_resp.ag_buf = q->ag;
986 turbo_resp.cb_buf = q->code_block;
987 turbo_resp.output = out;
988 iter_cnt = bblib_turbo_decoder(&turbo_req, &turbo_resp);
989 dec->hard_output.length += (k >> 3);
992 /* Temporary solution for returned iter_count from SDK */
993 iter_cnt = (iter_cnt - 1) / 2;
994 dec->iter_count = RTE_MAX(iter_cnt, dec->iter_count);
996 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
997 rte_bbdev_log(ERR, "Turbo Decoder failed");
1003 enqueue_dec_one_op(struct turbo_sw_queue *q, struct rte_bbdev_dec_op *op)
1007 uint16_t crc24_overlap = 0;
1008 struct rte_bbdev_op_turbo_dec *dec = &op->turbo_dec;
1009 struct rte_mbuf *m_in = dec->input.data;
1010 struct rte_mbuf *m_out = dec->hard_output.data;
1011 uint16_t in_offset = dec->input.offset;
1012 uint16_t total_left = dec->input.length;
1013 uint16_t out_offset = dec->hard_output.offset;
1015 /* Clear op status */
1018 if (m_in == NULL || m_out == NULL) {
1019 rte_bbdev_log(ERR, "Invalid mbuf pointer");
1020 op->status = 1 << RTE_BBDEV_DATA_ERROR;
1024 if (dec->code_block_mode == 0) { /* For Transport Block mode */
1025 c = dec->tb_params.c;
1026 } else { /* For Code Block mode */
1027 k = dec->cb_params.k;
1031 if ((c > 1) && !check_bit(dec->op_flags,
1032 RTE_BBDEV_TURBO_DEC_TB_CRC_24B_KEEP))
1035 while (total_left > 0) {
1036 if (dec->code_block_mode == 0)
1037 k = (r < dec->tb_params.c_neg) ?
1038 dec->tb_params.k_neg : dec->tb_params.k_pos;
1040 /* Calculates circular buffer size (Kw).
1041 * According to 3gpp 36.212 section 5.1.4.2
1045 * where nCol is 32 and nRow can be calculated from:
1047 * where D is the size of each output from turbo encoder block
1050 kw = RTE_ALIGN_CEIL(k + 4, RTE_BBDEV_C_SUBBLOCK) * 3;
1052 process_dec_cb(q, op, c, k, kw, m_in, m_out, in_offset,
1053 out_offset, check_bit(dec->op_flags,
1054 RTE_BBDEV_TURBO_CRC_TYPE_24B), crc24_overlap,
1056 /* To keep CRC24 attached to end of Code block, use
1057 * RTE_BBDEV_TURBO_DEC_TB_CRC_24B_KEEP flag as it
1058 * removed by default once verified.
1061 /* Update total_left */
1063 /* Update offsets for next CBs (if exist) */
1065 out_offset += ((k - crc24_overlap) >> 3);
1068 if (total_left != 0) {
1069 op->status |= 1 << RTE_BBDEV_DATA_ERROR;
1071 "Mismatch between mbuf length and included Circular buffer sizes");
1075 static inline uint16_t
1076 enqueue_dec_all_ops(struct turbo_sw_queue *q, struct rte_bbdev_dec_op **ops,
1081 for (i = 0; i < nb_ops; ++i)
1082 enqueue_dec_one_op(q, ops[i]);
1084 return rte_ring_enqueue_burst(q->processed_pkts, (void **)ops, nb_ops,
1090 enqueue_enc_ops(struct rte_bbdev_queue_data *q_data,
1091 struct rte_bbdev_enc_op **ops, uint16_t nb_ops)
1093 void *queue = q_data->queue_private;
1094 struct turbo_sw_queue *q = queue;
1095 uint16_t nb_enqueued = 0;
1097 nb_enqueued = enqueue_enc_all_ops(q, ops, nb_ops, &q_data->queue_stats);
1099 q_data->queue_stats.enqueue_err_count += nb_ops - nb_enqueued;
1100 q_data->queue_stats.enqueued_count += nb_enqueued;
1107 enqueue_dec_ops(struct rte_bbdev_queue_data *q_data,
1108 struct rte_bbdev_dec_op **ops, uint16_t nb_ops)
1110 void *queue = q_data->queue_private;
1111 struct turbo_sw_queue *q = queue;
1112 uint16_t nb_enqueued = 0;
1114 nb_enqueued = enqueue_dec_all_ops(q, ops, nb_ops);
1116 q_data->queue_stats.enqueue_err_count += nb_ops - nb_enqueued;
1117 q_data->queue_stats.enqueued_count += nb_enqueued;
1122 /* Dequeue decode burst */
1124 dequeue_dec_ops(struct rte_bbdev_queue_data *q_data,
1125 struct rte_bbdev_dec_op **ops, uint16_t nb_ops)
1127 struct turbo_sw_queue *q = q_data->queue_private;
1128 uint16_t nb_dequeued = rte_ring_dequeue_burst(q->processed_pkts,
1129 (void **)ops, nb_ops, NULL);
1130 q_data->queue_stats.dequeued_count += nb_dequeued;
1135 /* Dequeue encode burst */
1137 dequeue_enc_ops(struct rte_bbdev_queue_data *q_data,
1138 struct rte_bbdev_enc_op **ops, uint16_t nb_ops)
1140 struct turbo_sw_queue *q = q_data->queue_private;
1141 uint16_t nb_dequeued = rte_ring_dequeue_burst(q->processed_pkts,
1142 (void **)ops, nb_ops, NULL);
1143 q_data->queue_stats.dequeued_count += nb_dequeued;
1148 /* Parse 16bit integer from string argument */
1150 parse_u16_arg(const char *key, const char *value, void *extra_args)
1152 uint16_t *u16 = extra_args;
1153 unsigned int long result;
1155 if ((value == NULL) || (extra_args == NULL))
1158 result = strtoul(value, NULL, 0);
1159 if ((result >= (1 << 16)) || (errno != 0)) {
1160 rte_bbdev_log(ERR, "Invalid value %lu for %s", result, key);
1163 *u16 = (uint16_t)result;
1167 /* Parse parameters used to create device */
1169 parse_turbo_sw_params(struct turbo_sw_params *params, const char *input_args)
1171 struct rte_kvargs *kvlist = NULL;
1177 kvlist = rte_kvargs_parse(input_args, turbo_sw_valid_params);
1181 ret = rte_kvargs_process(kvlist, turbo_sw_valid_params[0],
1182 &parse_u16_arg, ¶ms->queues_num);
1186 ret = rte_kvargs_process(kvlist, turbo_sw_valid_params[1],
1187 &parse_u16_arg, ¶ms->socket_id);
1191 if (params->socket_id >= RTE_MAX_NUMA_NODES) {
1192 rte_bbdev_log(ERR, "Invalid socket, must be < %u",
1193 RTE_MAX_NUMA_NODES);
1200 rte_kvargs_free(kvlist);
1206 turbo_sw_bbdev_create(struct rte_vdev_device *vdev,
1207 struct turbo_sw_params *init_params)
1209 struct rte_bbdev *bbdev;
1210 const char *name = rte_vdev_device_name(vdev);
1212 bbdev = rte_bbdev_allocate(name);
1216 bbdev->data->dev_private = rte_zmalloc_socket(name,
1217 sizeof(struct bbdev_private), RTE_CACHE_LINE_SIZE,
1218 init_params->socket_id);
1219 if (bbdev->data->dev_private == NULL) {
1220 rte_bbdev_release(bbdev);
1224 bbdev->dev_ops = &pmd_ops;
1225 bbdev->device = &vdev->device;
1226 bbdev->data->socket_id = init_params->socket_id;
1227 bbdev->intr_handle = NULL;
1229 /* register rx/tx burst functions for data path */
1230 bbdev->dequeue_enc_ops = dequeue_enc_ops;
1231 bbdev->dequeue_dec_ops = dequeue_dec_ops;
1232 bbdev->enqueue_enc_ops = enqueue_enc_ops;
1233 bbdev->enqueue_dec_ops = enqueue_dec_ops;
1234 ((struct bbdev_private *) bbdev->data->dev_private)->max_nb_queues =
1235 init_params->queues_num;
1240 /* Initialise device */
1242 turbo_sw_bbdev_probe(struct rte_vdev_device *vdev)
1244 struct turbo_sw_params init_params = {
1246 RTE_BBDEV_DEFAULT_MAX_NB_QUEUES
1249 const char *input_args;
1254 name = rte_vdev_device_name(vdev);
1257 input_args = rte_vdev_device_args(vdev);
1258 parse_turbo_sw_params(&init_params, input_args);
1260 rte_bbdev_log_debug(
1261 "Initialising %s on NUMA node %d with max queues: %d\n",
1262 name, init_params.socket_id, init_params.queues_num);
1264 return turbo_sw_bbdev_create(vdev, &init_params);
1267 /* Uninitialise device */
1269 turbo_sw_bbdev_remove(struct rte_vdev_device *vdev)
1271 struct rte_bbdev *bbdev;
1277 name = rte_vdev_device_name(vdev);
1281 bbdev = rte_bbdev_get_named_dev(name);
1285 rte_free(bbdev->data->dev_private);
1287 return rte_bbdev_release(bbdev);
1290 static struct rte_vdev_driver bbdev_turbo_sw_pmd_drv = {
1291 .probe = turbo_sw_bbdev_probe,
1292 .remove = turbo_sw_bbdev_remove
1295 RTE_PMD_REGISTER_VDEV(DRIVER_NAME, bbdev_turbo_sw_pmd_drv);
1296 RTE_PMD_REGISTER_PARAM_STRING(DRIVER_NAME,
1297 TURBO_SW_MAX_NB_QUEUES_ARG"=<int> "
1298 TURBO_SW_SOCKET_ID_ARG"=<int>");
1300 RTE_INIT(null_bbdev_init_log);
1302 null_bbdev_init_log(void)
1304 bbdev_turbo_sw_logtype = rte_log_register("pmd.bb.turbo_sw");
1305 if (bbdev_turbo_sw_logtype >= 0)
1306 rte_log_set_level(bbdev_turbo_sw_logtype, RTE_LOG_NOTICE);