1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright 2017 6WIND S.A.
3 * Copyright 2017 Mellanox Technologies, Ltd
6 #ifndef RTE_PMD_MLX5_RXTX_VEC_NEON_H_
7 #define RTE_PMD_MLX5_RXTX_VEC_NEON_H_
15 #include <rte_mempool.h>
16 #include <rte_prefetch.h>
20 #include "mlx5_defs.h"
22 #include "mlx5_utils.h"
23 #include "mlx5_rxtx.h"
24 #include "mlx5_rxtx_vec.h"
25 #include "mlx5_autoconf.h"
27 #pragma GCC diagnostic ignored "-Wcast-qual"
30 * Store free buffers to RX SW ring.
33 * Pointer to SW ring to be filled.
35 * Pointer to array of packets to be stored.
37 * Number of packets to be stored.
40 rxq_copy_mbuf_v(struct rte_mbuf **elts, struct rte_mbuf **pkts, uint16_t n)
45 for (pos = 0; pos < p; pos += 2) {
48 mbp = vld1q_u64((void *)&elts[pos]);
49 vst1q_u64((void *)&pkts[pos], mbp);
52 pkts[pos] = elts[pos];
56 * Decompress a compressed completion and fill in mbufs in RX SW ring with data
57 * extracted from the title completion descriptor.
60 * Pointer to RX queue structure.
62 * Pointer to completion array having a compressed completion at first.
64 * Pointer to SW ring to be filled. The first mbuf has to be pre-built from
65 * the title completion descriptor to be copied to the rest of mbufs.
68 * Number of mini-CQEs successfully decompressed.
70 static inline uint16_t
71 rxq_cq_decompress_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq,
72 struct rte_mbuf **elts)
74 volatile struct mlx5_mini_cqe8 *mcq = (void *)&(cq + 1)->pkt_info;
75 struct rte_mbuf *t_pkt = elts[0]; /* Title packet is pre-built. */
79 /* Mask to shuffle from extracted mini CQE to mbuf. */
80 const uint8x16_t mcqe_shuf_m1 = {
81 -1, -1, -1, -1, /* skip packet_type */
82 7, 6, -1, -1, /* pkt_len, bswap16 */
83 7, 6, /* data_len, bswap16 */
84 -1, -1, /* skip vlan_tci */
85 3, 2, 1, 0 /* hash.rss, bswap32 */
87 const uint8x16_t mcqe_shuf_m2 = {
88 -1, -1, -1, -1, /* skip packet_type */
89 15, 14, -1, -1, /* pkt_len, bswap16 */
90 15, 14, /* data_len, bswap16 */
91 -1, -1, /* skip vlan_tci */
92 11, 10, 9, 8 /* hash.rss, bswap32 */
94 /* Restore the compressed count. Must be 16 bits. */
95 const uint16_t mcqe_n = t_pkt->data_len +
96 (rxq->crc_present * RTE_ETHER_CRC_LEN);
97 const uint64x2_t rearm =
98 vld1q_u64((void *)&t_pkt->rearm_data);
99 const uint32x4_t rxdf_mask = {
100 0xffffffff, /* packet_type */
101 0, /* skip pkt_len */
102 0xffff0000, /* vlan_tci, skip data_len */
103 0, /* skip hash.rss */
105 const uint8x16_t rxdf =
106 vandq_u8(vld1q_u8((void *)&t_pkt->rx_descriptor_fields1),
107 vreinterpretq_u8_u32(rxdf_mask));
108 const uint16x8_t crc_adj = {
110 rxq->crc_present * RTE_ETHER_CRC_LEN, 0,
111 rxq->crc_present * RTE_ETHER_CRC_LEN, 0,
114 uint32x4_t ol_flags = {0, 0, 0, 0};
115 uint32x4_t ol_flags_mask = {0, 0, 0, 0};
116 #ifdef MLX5_PMD_SOFT_COUNTERS
117 uint32_t rcvd_byte = 0;
119 /* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */
120 const uint8x8_t len_shuf_m = {
122 15, 14, /* 2nd mCQE */
123 23, 22, /* 3rd mCQE */
124 31, 30 /* 4th mCQE */
128 * A. load mCQEs into a 128bit register.
129 * B. store rearm data to mbuf.
130 * C. combine data from mCQEs with rx_descriptor_fields1.
131 * D. store rx_descriptor_fields1.
132 * E. store flow tag (rte_flow mark).
134 for (pos = 0; pos < mcqe_n; ) {
135 uint8_t *p = (void *)&mcq[pos % 8];
136 uint8_t *e0 = (void *)&elts[pos]->rearm_data;
137 uint8_t *e1 = (void *)&elts[pos + 1]->rearm_data;
138 uint8_t *e2 = (void *)&elts[pos + 2]->rearm_data;
139 uint8_t *e3 = (void *)&elts[pos + 3]->rearm_data;
141 #ifdef MLX5_PMD_SOFT_COUNTERS
142 uint16x4_t invalid_mask =
143 vcreate_u16(mcqe_n - pos < MLX5_VPMD_DESCS_PER_LOOP ?
144 -1UL << ((mcqe_n - pos) *
145 sizeof(uint16_t) * 8) : 0);
148 for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i)
149 if (likely(pos + i < mcqe_n))
150 rte_prefetch0((void *)(cq + pos + i));
152 /* A.1 load mCQEs into a 128bit register. */
153 "ld1 {v16.16b - v17.16b}, [%[mcq]] \n\t"
154 /* B.1 store rearm data to mbuf. */
155 "st1 {%[rearm].2d}, [%[e0]] \n\t"
156 "add %[e0], %[e0], #16 \n\t"
157 "st1 {%[rearm].2d}, [%[e1]] \n\t"
158 "add %[e1], %[e1], #16 \n\t"
159 /* C.1 combine data from mCQEs with rx_descriptor_fields1. */
160 "tbl v18.16b, {v16.16b}, %[mcqe_shuf_m1].16b \n\t"
161 "tbl v19.16b, {v16.16b}, %[mcqe_shuf_m2].16b \n\t"
162 "sub v18.8h, v18.8h, %[crc_adj].8h \n\t"
163 "sub v19.8h, v19.8h, %[crc_adj].8h \n\t"
164 "orr v18.16b, v18.16b, %[rxdf].16b \n\t"
165 "orr v19.16b, v19.16b, %[rxdf].16b \n\t"
166 /* D.1 store rx_descriptor_fields1. */
167 "st1 {v18.2d}, [%[e0]] \n\t"
168 "st1 {v19.2d}, [%[e1]] \n\t"
169 /* B.1 store rearm data to mbuf. */
170 "st1 {%[rearm].2d}, [%[e2]] \n\t"
171 "add %[e2], %[e2], #16 \n\t"
172 "st1 {%[rearm].2d}, [%[e3]] \n\t"
173 "add %[e3], %[e3], #16 \n\t"
174 /* C.1 combine data from mCQEs with rx_descriptor_fields1. */
175 "tbl v18.16b, {v17.16b}, %[mcqe_shuf_m1].16b \n\t"
176 "tbl v19.16b, {v17.16b}, %[mcqe_shuf_m2].16b \n\t"
177 "sub v18.8h, v18.8h, %[crc_adj].8h \n\t"
178 "sub v19.8h, v19.8h, %[crc_adj].8h \n\t"
179 "orr v18.16b, v18.16b, %[rxdf].16b \n\t"
180 "orr v19.16b, v19.16b, %[rxdf].16b \n\t"
181 /* D.1 store rx_descriptor_fields1. */
182 "st1 {v18.2d}, [%[e2]] \n\t"
183 "st1 {v19.2d}, [%[e3]] \n\t"
184 #ifdef MLX5_PMD_SOFT_COUNTERS
185 "tbl %[byte_cnt].8b, {v16.16b - v17.16b}, %[len_shuf_m].8b \n\t"
187 :[byte_cnt]"=&w"(byte_cnt)
191 [e3]"r"(e3), [e2]"r"(e2), [e1]"r"(e1), [e0]"r"(e0),
192 [mcqe_shuf_m1]"w"(mcqe_shuf_m1),
193 [mcqe_shuf_m2]"w"(mcqe_shuf_m2),
194 [crc_adj]"w"(crc_adj),
195 [len_shuf_m]"w"(len_shuf_m)
196 :"memory", "v16", "v17", "v18", "v19");
197 #ifdef MLX5_PMD_SOFT_COUNTERS
198 byte_cnt = vbic_u16(byte_cnt, invalid_mask);
199 rcvd_byte += vget_lane_u64(vpaddl_u32(vpaddl_u16(byte_cnt)), 0);
202 if (rxq->mcqe_format !=
203 MLX5_CQE_RESP_FORMAT_FTAG_STRIDX) {
204 const uint32_t flow_tag = t_pkt->hash.fdir.hi;
206 /* E.1 store flow tag (rte_flow mark). */
207 elts[pos]->hash.fdir.hi = flow_tag;
208 elts[pos + 1]->hash.fdir.hi = flow_tag;
209 elts[pos + 2]->hash.fdir.hi = flow_tag;
210 elts[pos + 3]->hash.fdir.hi = flow_tag;
212 const uint32x4_t flow_mark_adj = {
214 const uint8x16_t flow_mark_shuf = {
219 /* Extract flow_tag field. */
220 const uint32x4_t ft_mask =
221 vdupq_n_u32(MLX5_FLOW_MARK_DEFAULT);
222 const uint32x4_t fdir_flags =
223 vdupq_n_u32(PKT_RX_FDIR);
224 const uint32x4_t fdir_all_flags =
225 vdupq_n_u32(PKT_RX_FDIR |
227 uint32x4_t fdir_id_flags =
228 vdupq_n_u32(PKT_RX_FDIR_ID);
229 uint32x4_t invalid_mask, ftag;
232 /* A.1 load mCQEs into a 128bit register. */
233 ("ld1 {v16.16b - v17.16b}, [%[mcq]]\n\t"
234 /* Extract flow_tag. */
235 "tbl %[ftag].16b, {v16.16b - v17.16b}, %[flow_mark_shuf].16b\n\t"
238 [flow_mark_shuf]"w"(flow_mark_shuf)
239 : "memory", "v16", "v17");
240 invalid_mask = vceqzq_u32(ftag);
241 ol_flags_mask = vorrq_u32(ol_flags_mask,
243 /* Set PKT_RX_FDIR if flow tag is non-zero. */
244 ol_flags = vorrq_u32(ol_flags,
245 vbicq_u32(fdir_flags, invalid_mask));
246 /* Mask out invalid entries. */
247 fdir_id_flags = vbicq_u32(fdir_id_flags,
249 /* Check if flow tag MLX5_FLOW_MARK_DEFAULT. */
250 ol_flags = vorrq_u32(ol_flags,
251 vbicq_u32(fdir_id_flags,
252 vceqq_u32(ftag, ft_mask)));
253 ftag = vaddq_u32(ftag, flow_mark_adj);
254 elts[pos]->hash.fdir.hi =
255 vgetq_lane_u32(ftag, 3);
256 elts[pos + 1]->hash.fdir.hi =
257 vgetq_lane_u32(ftag, 2);
258 elts[pos + 2]->hash.fdir.hi =
259 vgetq_lane_u32(ftag, 1);
260 elts[pos + 3]->hash.fdir.hi =
261 vgetq_lane_u32(ftag, 0);
264 if (unlikely(rxq->mcqe_format !=
265 MLX5_CQE_RESP_FORMAT_HASH)) {
266 if (rxq->mcqe_format ==
267 MLX5_CQE_RESP_FORMAT_L34H_STRIDX) {
268 const uint8_t pkt_info =
269 (cq->pkt_info & 0x3) << 6;
270 const uint8_t pkt_hdr0 =
271 mcq[pos % 8].hdr_type;
272 const uint8_t pkt_hdr1 =
273 mcq[pos % 8 + 1].hdr_type;
274 const uint8_t pkt_hdr2 =
275 mcq[pos % 8 + 2].hdr_type;
276 const uint8_t pkt_hdr3 =
277 mcq[pos % 8 + 3].hdr_type;
278 const uint32x4_t vlan_mask =
279 vdupq_n_u32(PKT_RX_VLAN |
280 PKT_RX_VLAN_STRIPPED);
281 const uint32x4_t cv_mask =
282 vdupq_n_u32(MLX5_CQE_VLAN_STRIPPED);
283 const uint32x4_t pkt_cv = {
284 pkt_hdr0 & 0x1, pkt_hdr1 & 0x1,
285 pkt_hdr2 & 0x1, pkt_hdr3 & 0x1};
287 ol_flags_mask = vorrq_u32(ol_flags_mask,
289 ol_flags = vorrq_u32(ol_flags,
291 vceqq_u32(pkt_cv, cv_mask)));
292 elts[pos]->packet_type =
293 mlx5_ptype_table[(pkt_hdr0 >> 2) |
295 elts[pos + 1]->packet_type =
296 mlx5_ptype_table[(pkt_hdr1 >> 2) |
298 elts[pos + 2]->packet_type =
299 mlx5_ptype_table[(pkt_hdr2 >> 2) |
301 elts[pos + 3]->packet_type =
302 mlx5_ptype_table[(pkt_hdr3 >> 2) |
305 elts[pos]->packet_type |=
306 !!(((pkt_hdr0 >> 2) |
307 pkt_info) & (1 << 6));
308 elts[pos + 1]->packet_type |=
309 !!(((pkt_hdr1 >> 2) |
310 pkt_info) & (1 << 6));
311 elts[pos + 2]->packet_type |=
312 !!(((pkt_hdr2 >> 2) |
313 pkt_info) & (1 << 6));
314 elts[pos + 3]->packet_type |=
315 !!(((pkt_hdr3 >> 2) |
316 pkt_info) & (1 << 6));
319 const uint32x4_t hash_flags =
320 vdupq_n_u32(PKT_RX_RSS_HASH);
321 const uint32x4_t rearm_flags =
322 vdupq_n_u32((uint32_t)t_pkt->ol_flags);
324 ol_flags_mask = vorrq_u32(ol_flags_mask, hash_flags);
325 ol_flags = vorrq_u32(ol_flags,
326 vbicq_u32(rearm_flags, ol_flags_mask));
327 elts[pos]->ol_flags = vgetq_lane_u32(ol_flags, 3);
328 elts[pos + 1]->ol_flags = vgetq_lane_u32(ol_flags, 2);
329 elts[pos + 2]->ol_flags = vgetq_lane_u32(ol_flags, 1);
330 elts[pos + 3]->ol_flags = vgetq_lane_u32(ol_flags, 0);
331 elts[pos]->hash.rss = 0;
332 elts[pos + 1]->hash.rss = 0;
333 elts[pos + 2]->hash.rss = 0;
334 elts[pos + 3]->hash.rss = 0;
336 if (rxq->dynf_meta) {
337 int32_t offs = rxq->flow_meta_offset;
338 const uint32_t meta =
339 *RTE_MBUF_DYNFIELD(t_pkt, offs, uint32_t *);
341 /* Check if title packet has valid metadata. */
343 MLX5_ASSERT(t_pkt->ol_flags &
344 rxq->flow_meta_mask);
345 *RTE_MBUF_DYNFIELD(elts[pos], offs,
347 *RTE_MBUF_DYNFIELD(elts[pos + 1], offs,
349 *RTE_MBUF_DYNFIELD(elts[pos + 2], offs,
351 *RTE_MBUF_DYNFIELD(elts[pos + 3], offs,
355 pos += MLX5_VPMD_DESCS_PER_LOOP;
356 /* Move to next CQE and invalidate consumed CQEs. */
357 if (!(pos & 0x7) && pos < mcqe_n) {
358 if (pos + 8 < mcqe_n)
359 rte_prefetch0((void *)(cq + pos + 8));
360 mcq = (void *)&(cq + pos)->pkt_info;
361 for (i = 0; i < 8; ++i)
362 cq[inv++].op_own = MLX5_CQE_INVALIDATE;
365 /* Invalidate the rest of CQEs. */
366 for (; inv < mcqe_n; ++inv)
367 cq[inv].op_own = MLX5_CQE_INVALIDATE;
368 #ifdef MLX5_PMD_SOFT_COUNTERS
369 rxq->stats.ipackets += mcqe_n;
370 rxq->stats.ibytes += rcvd_byte;
376 * Calculate packet type and offload flag for mbuf and store it.
379 * Pointer to RX queue structure.
381 * Array of four 4bytes packet type info extracted from the original
382 * completion descriptor.
384 * Array of four 4bytes flow ID extracted from the original completion
387 * Opcode vector having responder error status. Each field is 4B.
389 * Pointer to array of packets to be filled.
392 rxq_cq_to_ptype_oflags_v(struct mlx5_rxq_data *rxq,
393 uint32x4_t ptype_info, uint32x4_t flow_tag,
394 uint16x4_t op_err, struct rte_mbuf **pkts)
397 uint32x4_t pinfo, cv_flags;
398 uint32x4_t ol_flags =
399 vdupq_n_u32(rxq->rss_hash * PKT_RX_RSS_HASH |
400 rxq->hw_timestamp * rxq->timestamp_rx_flag);
401 const uint32x4_t ptype_ol_mask = { 0x106, 0x106, 0x106, 0x106 };
402 const uint8x16_t cv_flag_sel = {
404 (uint8_t)(PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED),
405 (uint8_t)(PKT_RX_IP_CKSUM_GOOD >> 1),
407 (uint8_t)(PKT_RX_L4_CKSUM_GOOD >> 1),
409 (uint8_t)((PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1),
410 0, 0, 0, 0, 0, 0, 0, 0, 0
412 const uint32x4_t cv_mask =
413 vdupq_n_u32(PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
414 PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED);
415 const uint64x2_t mbuf_init = vld1q_u64
416 ((const uint64_t *)&rxq->mbuf_initializer);
417 uint64x2_t rearm0, rearm1, rearm2, rearm3;
418 uint8_t pt_idx0, pt_idx1, pt_idx2, pt_idx3;
421 const uint32x4_t ft_def = vdupq_n_u32(MLX5_FLOW_MARK_DEFAULT);
422 const uint32x4_t fdir_flags = vdupq_n_u32(PKT_RX_FDIR);
423 uint32x4_t fdir_id_flags = vdupq_n_u32(PKT_RX_FDIR_ID);
424 uint32x4_t invalid_mask;
426 /* Check if flow tag is non-zero then set PKT_RX_FDIR. */
427 invalid_mask = vceqzq_u32(flow_tag);
428 ol_flags = vorrq_u32(ol_flags,
429 vbicq_u32(fdir_flags, invalid_mask));
430 /* Mask out invalid entries. */
431 fdir_id_flags = vbicq_u32(fdir_id_flags, invalid_mask);
432 /* Check if flow tag MLX5_FLOW_MARK_DEFAULT. */
433 ol_flags = vorrq_u32(ol_flags,
434 vbicq_u32(fdir_id_flags,
435 vceqq_u32(flow_tag, ft_def)));
438 * ptype_info has the following:
442 * bit[11:10] = l3_hdr_type
443 * bit[14:12] = l4_hdr_type
446 * bit[17] = outer_l3_type
448 ptype = vshrn_n_u32(ptype_info, 10);
449 /* Errored packets will have RTE_PTYPE_ALL_MASK. */
450 ptype = vorr_u16(ptype, op_err);
451 pt_idx0 = vget_lane_u8(vreinterpret_u8_u16(ptype), 6);
452 pt_idx1 = vget_lane_u8(vreinterpret_u8_u16(ptype), 4);
453 pt_idx2 = vget_lane_u8(vreinterpret_u8_u16(ptype), 2);
454 pt_idx3 = vget_lane_u8(vreinterpret_u8_u16(ptype), 0);
455 pkts[0]->packet_type = mlx5_ptype_table[pt_idx0] |
456 !!(pt_idx0 & (1 << 6)) * rxq->tunnel;
457 pkts[1]->packet_type = mlx5_ptype_table[pt_idx1] |
458 !!(pt_idx1 & (1 << 6)) * rxq->tunnel;
459 pkts[2]->packet_type = mlx5_ptype_table[pt_idx2] |
460 !!(pt_idx2 & (1 << 6)) * rxq->tunnel;
461 pkts[3]->packet_type = mlx5_ptype_table[pt_idx3] |
462 !!(pt_idx3 & (1 << 6)) * rxq->tunnel;
463 /* Fill flags for checksum and VLAN. */
464 pinfo = vandq_u32(ptype_info, ptype_ol_mask);
465 pinfo = vreinterpretq_u32_u8(
466 vqtbl1q_u8(cv_flag_sel, vreinterpretq_u8_u32(pinfo)));
467 /* Locate checksum flags at byte[2:1] and merge with VLAN flags. */
468 cv_flags = vshlq_n_u32(pinfo, 9);
469 cv_flags = vorrq_u32(pinfo, cv_flags);
470 /* Move back flags to start from byte[0]. */
471 cv_flags = vshrq_n_u32(cv_flags, 8);
472 /* Mask out garbage bits. */
473 cv_flags = vandq_u32(cv_flags, cv_mask);
474 /* Merge to ol_flags. */
475 ol_flags = vorrq_u32(ol_flags, cv_flags);
476 /* Merge mbuf_init and ol_flags, and store. */
477 rearm0 = vreinterpretq_u64_u32(vsetq_lane_u32
478 (vgetq_lane_u32(ol_flags, 3),
479 vreinterpretq_u32_u64(mbuf_init), 2));
480 rearm1 = vreinterpretq_u64_u32(vsetq_lane_u32
481 (vgetq_lane_u32(ol_flags, 2),
482 vreinterpretq_u32_u64(mbuf_init), 2));
483 rearm2 = vreinterpretq_u64_u32(vsetq_lane_u32
484 (vgetq_lane_u32(ol_flags, 1),
485 vreinterpretq_u32_u64(mbuf_init), 2));
486 rearm3 = vreinterpretq_u64_u32(vsetq_lane_u32
487 (vgetq_lane_u32(ol_flags, 0),
488 vreinterpretq_u32_u64(mbuf_init), 2));
490 vst1q_u64((void *)&pkts[0]->rearm_data, rearm0);
491 vst1q_u64((void *)&pkts[1]->rearm_data, rearm1);
492 vst1q_u64((void *)&pkts[2]->rearm_data, rearm2);
493 vst1q_u64((void *)&pkts[3]->rearm_data, rearm3);
497 * Process a non-compressed completion and fill in mbufs in RX SW ring
498 * with data extracted from the title completion descriptor.
501 * Pointer to RX queue structure.
503 * Pointer to completion array having a non-compressed completion at first.
505 * Pointer to SW ring to be filled. The first mbuf has to be pre-built from
506 * the title completion descriptor to be copied to the rest of mbufs.
508 * Array to store received packets.
510 * Maximum number of packets in array.
512 * Pointer to a flag. Set non-zero value if pkts array has at least one error
515 * Pointer to a index. Set it to the first compressed completion if any.
518 * Number of CQEs successfully processed.
520 static inline uint16_t
521 rxq_cq_process_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq,
522 struct rte_mbuf **elts, struct rte_mbuf **pkts,
523 uint16_t pkts_n, uint64_t *err, uint64_t *comp)
525 const uint16_t q_n = 1 << rxq->cqe_n;
526 const uint16_t q_mask = q_n - 1;
529 uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
530 uint16_t nocmp_n = 0;
531 const uint16x4_t ownership = vdup_n_u16(!(rxq->cq_ci & (q_mask + 1)));
532 const uint16x4_t owner_check = vcreate_u16(0x0001000100010001);
533 const uint16x4_t opcode_check = vcreate_u16(0x00f000f000f000f0);
534 const uint16x4_t format_check = vcreate_u16(0x000c000c000c000c);
535 const uint16x4_t resp_err_check = vcreate_u16(0x00e000e000e000e0);
536 #ifdef MLX5_PMD_SOFT_COUNTERS
537 uint32_t rcvd_byte = 0;
539 /* Mask to generate 16B length vector. */
540 const uint8x8_t len_shuf_m = {
541 52, 53, /* 4th CQE */
542 36, 37, /* 3rd CQE */
543 20, 21, /* 2nd CQE */
546 /* Mask to extract 16B data from a 64B CQE. */
547 const uint8x16_t cqe_shuf_m = {
548 28, 29, /* hdr_type_etc */
551 47, 46, /* byte_cnt, bswap16 */
552 31, 30, /* vlan_info, bswap16 */
553 15, 14, 13, 12, /* rx_hash_res, bswap32 */
554 57, 58, 59, /* flow_tag */
557 /* Mask to generate 16B data for mbuf. */
558 const uint8x16_t mb_shuf_m = {
559 4, 5, -1, -1, /* pkt_len */
562 8, 9, 10, 11, /* hash.rss */
563 12, 13, 14, -1 /* hash.fdir.hi */
565 /* Mask to generate 16B owner vector. */
566 const uint8x8_t owner_shuf_m = {
567 63, -1, /* 4th CQE */
568 47, -1, /* 3rd CQE */
569 31, -1, /* 2nd CQE */
572 /* Mask to generate a vector having packet_type/ol_flags. */
573 const uint8x16_t ptype_shuf_m = {
574 48, 49, 50, -1, /* 4th CQE */
575 32, 33, 34, -1, /* 3rd CQE */
576 16, 17, 18, -1, /* 2nd CQE */
577 0, 1, 2, -1 /* 1st CQE */
579 /* Mask to generate a vector having flow tags. */
580 const uint8x16_t ftag_shuf_m = {
581 60, 61, 62, -1, /* 4th CQE */
582 44, 45, 46, -1, /* 3rd CQE */
583 28, 29, 30, -1, /* 2nd CQE */
584 12, 13, 14, -1 /* 1st CQE */
586 const uint16x8_t crc_adj = {
587 0, 0, rxq->crc_present * RTE_ETHER_CRC_LEN, 0, 0, 0, 0, 0
589 const uint32x4_t flow_mark_adj = { 0, 0, 0, rxq->mark * (-1) };
592 * Note that vectors have reverse order - {v3, v2, v1, v0}, because
593 * there's no instruction to count trailing zeros. __builtin_clzl() is
596 * A. copy 4 mbuf pointers from elts ring to returning pkts.
597 * B. load 64B CQE and extract necessary fields
598 * Final 16bytes cqes[] extracted from original 64bytes CQE has the
599 * following structure:
601 * uint16_t hdr_type_etc;
605 * uint16_t vlan_info;
606 * uint32_t rx_has_res;
607 * uint8_t flow_tag[3];
612 * E. find compressed CQE.
616 pos += MLX5_VPMD_DESCS_PER_LOOP) {
618 uint16x4_t opcode, owner_mask, invalid_mask;
619 uint16x4_t comp_mask;
622 uint32x4_t ptype_info, flow_tag;
623 register uint64x2_t c0, c1, c2, c3;
624 uint8_t *p0, *p1, *p2, *p3;
625 uint8_t *e0 = (void *)&elts[pos]->pkt_len;
626 uint8_t *e1 = (void *)&elts[pos + 1]->pkt_len;
627 uint8_t *e2 = (void *)&elts[pos + 2]->pkt_len;
628 uint8_t *e3 = (void *)&elts[pos + 3]->pkt_len;
629 void *elts_p = (void *)&elts[pos];
630 void *pkts_p = (void *)&pkts[pos];
632 /* A.0 do not cross the end of CQ. */
633 mask = vcreate_u16(pkts_n - pos < MLX5_VPMD_DESCS_PER_LOOP ?
634 -1UL >> ((pkts_n - pos) *
635 sizeof(uint16_t) * 8) : 0);
636 p0 = (void *)&cq[pos].pkt_info;
637 p1 = p0 + (pkts_n - pos > 1) * sizeof(struct mlx5_cqe);
638 p2 = p1 + (pkts_n - pos > 2) * sizeof(struct mlx5_cqe);
639 p3 = p2 + (pkts_n - pos > 3) * sizeof(struct mlx5_cqe);
640 /* B.0 (CQE 3) load a block having op_own. */
641 c3 = vld1q_u64((uint64_t *)(p3 + 48));
642 /* B.0 (CQE 2) load a block having op_own. */
643 c2 = vld1q_u64((uint64_t *)(p2 + 48));
644 /* B.0 (CQE 1) load a block having op_own. */
645 c1 = vld1q_u64((uint64_t *)(p1 + 48));
646 /* B.0 (CQE 0) load a block having op_own. */
647 c0 = vld1q_u64((uint64_t *)(p0 + 48));
648 /* Synchronize for loading the rest of blocks. */
650 /* Prefetch next 4 CQEs. */
651 if (pkts_n - pos >= 2 * MLX5_VPMD_DESCS_PER_LOOP) {
652 unsigned int next = pos + MLX5_VPMD_DESCS_PER_LOOP;
653 rte_prefetch_non_temporal(&cq[next]);
654 rte_prefetch_non_temporal(&cq[next + 1]);
655 rte_prefetch_non_temporal(&cq[next + 2]);
656 rte_prefetch_non_temporal(&cq[next + 3]);
659 /* B.1 (CQE 3) load the rest of blocks. */
660 "ld1 {v16.16b - v18.16b}, [%[p3]] \n\t"
661 /* B.2 (CQE 3) move the block having op_own. */
662 "mov v19.16b, %[c3].16b \n\t"
663 /* B.3 (CQE 3) extract 16B fields. */
664 "tbl v23.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
665 /* B.1 (CQE 2) load the rest of blocks. */
666 "ld1 {v16.16b - v18.16b}, [%[p2]] \n\t"
667 /* B.4 (CQE 3) adjust CRC length. */
668 "sub v23.8h, v23.8h, %[crc_adj].8h \n\t"
669 /* C.1 (CQE 3) generate final structure for mbuf. */
670 "tbl v15.16b, {v23.16b}, %[mb_shuf_m].16b \n\t"
671 /* B.2 (CQE 2) move the block having op_own. */
672 "mov v19.16b, %[c2].16b \n\t"
673 /* B.3 (CQE 2) extract 16B fields. */
674 "tbl v22.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
675 /* B.1 (CQE 1) load the rest of blocks. */
676 "ld1 {v16.16b - v18.16b}, [%[p1]] \n\t"
677 /* B.4 (CQE 2) adjust CRC length. */
678 "sub v22.8h, v22.8h, %[crc_adj].8h \n\t"
679 /* C.1 (CQE 2) generate final structure for mbuf. */
680 "tbl v14.16b, {v22.16b}, %[mb_shuf_m].16b \n\t"
681 /* B.2 (CQE 1) move the block having op_own. */
682 "mov v19.16b, %[c1].16b \n\t"
683 /* B.3 (CQE 1) extract 16B fields. */
684 "tbl v21.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
685 /* B.1 (CQE 0) load the rest of blocks. */
686 "ld1 {v16.16b - v18.16b}, [%[p0]] \n\t"
687 /* B.4 (CQE 1) adjust CRC length. */
688 "sub v21.8h, v21.8h, %[crc_adj].8h \n\t"
689 /* C.1 (CQE 1) generate final structure for mbuf. */
690 "tbl v13.16b, {v21.16b}, %[mb_shuf_m].16b \n\t"
691 /* B.2 (CQE 0) move the block having op_own. */
692 "mov v19.16b, %[c0].16b \n\t"
693 /* A.1 load mbuf pointers. */
694 "ld1 {v24.2d - v25.2d}, [%[elts_p]] \n\t"
695 /* B.3 (CQE 0) extract 16B fields. */
696 "tbl v20.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
697 /* B.4 (CQE 0) adjust CRC length. */
698 "sub v20.8h, v20.8h, %[crc_adj].8h \n\t"
699 /* D.1 extract op_own byte. */
700 "tbl %[op_own].8b, {v20.16b - v23.16b}, %[owner_shuf_m].8b \n\t"
701 /* C.2 (CQE 3) adjust flow mark. */
702 "add v15.4s, v15.4s, %[flow_mark_adj].4s \n\t"
703 /* C.3 (CQE 3) fill in mbuf - rx_descriptor_fields1. */
704 "st1 {v15.2d}, [%[e3]] \n\t"
705 /* C.2 (CQE 2) adjust flow mark. */
706 "add v14.4s, v14.4s, %[flow_mark_adj].4s \n\t"
707 /* C.3 (CQE 2) fill in mbuf - rx_descriptor_fields1. */
708 "st1 {v14.2d}, [%[e2]] \n\t"
709 /* C.1 (CQE 0) generate final structure for mbuf. */
710 "tbl v12.16b, {v20.16b}, %[mb_shuf_m].16b \n\t"
711 /* C.2 (CQE 1) adjust flow mark. */
712 "add v13.4s, v13.4s, %[flow_mark_adj].4s \n\t"
713 /* C.3 (CQE 1) fill in mbuf - rx_descriptor_fields1. */
714 "st1 {v13.2d}, [%[e1]] \n\t"
715 #ifdef MLX5_PMD_SOFT_COUNTERS
716 /* Extract byte_cnt. */
717 "tbl %[byte_cnt].8b, {v20.16b - v23.16b}, %[len_shuf_m].8b \n\t"
719 /* Extract ptype_info. */
720 "tbl %[ptype_info].16b, {v20.16b - v23.16b}, %[ptype_shuf_m].16b \n\t"
721 /* Extract flow_tag. */
722 "tbl %[flow_tag].16b, {v20.16b - v23.16b}, %[ftag_shuf_m].16b \n\t"
723 /* A.2 copy mbuf pointers. */
724 "st1 {v24.2d - v25.2d}, [%[pkts_p]] \n\t"
725 /* C.2 (CQE 0) adjust flow mark. */
726 "add v12.4s, v12.4s, %[flow_mark_adj].4s \n\t"
727 /* C.3 (CQE 1) fill in mbuf - rx_descriptor_fields1. */
728 "st1 {v12.2d}, [%[e0]] \n\t"
729 :[op_own]"=&w"(op_own),
730 [byte_cnt]"=&w"(byte_cnt),
731 [ptype_info]"=&w"(ptype_info),
732 [flow_tag]"=&w"(flow_tag)
733 :[p3]"r"(p3), [p2]"r"(p2), [p1]"r"(p1), [p0]"r"(p0),
734 [e3]"r"(e3), [e2]"r"(e2), [e1]"r"(e1), [e0]"r"(e0),
735 [c3]"w"(c3), [c2]"w"(c2), [c1]"w"(c1), [c0]"w"(c0),
738 [cqe_shuf_m]"w"(cqe_shuf_m),
739 [mb_shuf_m]"w"(mb_shuf_m),
740 [owner_shuf_m]"w"(owner_shuf_m),
741 [len_shuf_m]"w"(len_shuf_m),
742 [ptype_shuf_m]"w"(ptype_shuf_m),
743 [ftag_shuf_m]"w"(ftag_shuf_m),
744 [crc_adj]"w"(crc_adj),
745 [flow_mark_adj]"w"(flow_mark_adj)
747 "v12", "v13", "v14", "v15",
748 "v16", "v17", "v18", "v19",
749 "v20", "v21", "v22", "v23",
751 /* D.2 flip owner bit to mark CQEs from last round. */
752 owner_mask = vand_u16(op_own, owner_check);
753 owner_mask = vceq_u16(owner_mask, ownership);
754 /* D.3 get mask for invalidated CQEs. */
755 opcode = vand_u16(op_own, opcode_check);
756 invalid_mask = vceq_u16(opcode_check, opcode);
757 /* E.1 find compressed CQE format. */
758 comp_mask = vand_u16(op_own, format_check);
759 comp_mask = vceq_u16(comp_mask, format_check);
760 /* D.4 mask out beyond boundary. */
761 invalid_mask = vorr_u16(invalid_mask, mask);
762 /* D.5 merge invalid_mask with invalid owner. */
763 invalid_mask = vorr_u16(invalid_mask, owner_mask);
764 /* E.2 mask out invalid entries. */
765 comp_mask = vbic_u16(comp_mask, invalid_mask);
766 /* E.3 get the first compressed CQE. */
767 comp_idx = __builtin_clzl(vget_lane_u64(vreinterpret_u64_u16(
769 (sizeof(uint16_t) * 8);
770 invalid_mask = vorr_u16(invalid_mask, comp_mask);
771 /* D.7 count non-compressed valid CQEs. */
772 n = __builtin_clzl(vget_lane_u64(vreinterpret_u64_u16(
773 invalid_mask), 0)) / (sizeof(uint16_t) * 8);
776 * D.2 mask out entries after the compressed CQE.
777 * get the final invalid mask.
779 mask = vcreate_u16(n < MLX5_VPMD_DESCS_PER_LOOP ?
780 -1UL >> (n * sizeof(uint16_t) * 8) : 0);
781 invalid_mask = vorr_u16(invalid_mask, mask);
782 /* D.3 check error in opcode. */
783 opcode = vceq_u16(resp_err_check, opcode);
784 opcode = vbic_u16(opcode, invalid_mask);
785 /* D.4 mark if any error is set */
786 *err |= vget_lane_u64(vreinterpret_u64_u16(opcode), 0);
787 /* C.4 fill in mbuf - rearm_data and packet_type. */
788 rxq_cq_to_ptype_oflags_v(rxq, ptype_info, flow_tag,
790 if (rxq->hw_timestamp) {
791 int offset = rxq->timestamp_offset;
792 if (rxq->rt_timestamp) {
793 struct mlx5_dev_ctx_shared *sh = rxq->sh;
796 ts = rte_be_to_cpu_64
797 (container_of(p0, struct mlx5_cqe,
798 pkt_info)->timestamp);
799 mlx5_timestamp_set(elts[pos], offset,
800 mlx5_txpp_convert_rx_ts(sh, ts));
801 ts = rte_be_to_cpu_64
802 (container_of(p1, struct mlx5_cqe,
803 pkt_info)->timestamp);
804 mlx5_timestamp_set(elts[pos + 1], offset,
805 mlx5_txpp_convert_rx_ts(sh, ts));
806 ts = rte_be_to_cpu_64
807 (container_of(p2, struct mlx5_cqe,
808 pkt_info)->timestamp);
809 mlx5_timestamp_set(elts[pos + 2], offset,
810 mlx5_txpp_convert_rx_ts(sh, ts));
811 ts = rte_be_to_cpu_64
812 (container_of(p3, struct mlx5_cqe,
813 pkt_info)->timestamp);
814 mlx5_timestamp_set(elts[pos + 3], offset,
815 mlx5_txpp_convert_rx_ts(sh, ts));
817 mlx5_timestamp_set(elts[pos], offset,
818 rte_be_to_cpu_64(container_of(p0,
819 struct mlx5_cqe, pkt_info)->timestamp));
820 mlx5_timestamp_set(elts[pos + 1], offset,
821 rte_be_to_cpu_64(container_of(p1,
822 struct mlx5_cqe, pkt_info)->timestamp));
823 mlx5_timestamp_set(elts[pos + 2], offset,
824 rte_be_to_cpu_64(container_of(p2,
825 struct mlx5_cqe, pkt_info)->timestamp));
826 mlx5_timestamp_set(elts[pos + 3], offset,
827 rte_be_to_cpu_64(container_of(p3,
828 struct mlx5_cqe, pkt_info)->timestamp));
831 if (rxq->dynf_meta) {
832 /* This code is subject for futher optimization. */
833 int32_t offs = rxq->flow_meta_offset;
834 uint32_t mask = rxq->flow_meta_port_mask;
836 *RTE_MBUF_DYNFIELD(pkts[pos], offs, uint32_t *) =
837 rte_be_to_cpu_32(container_of
838 (p0, struct mlx5_cqe,
839 pkt_info)->flow_table_metadata) & mask;
840 *RTE_MBUF_DYNFIELD(pkts[pos + 1], offs, uint32_t *) =
841 rte_be_to_cpu_32(container_of
842 (p1, struct mlx5_cqe,
843 pkt_info)->flow_table_metadata) & mask;
844 *RTE_MBUF_DYNFIELD(pkts[pos + 2], offs, uint32_t *) =
845 rte_be_to_cpu_32(container_of
846 (p2, struct mlx5_cqe,
847 pkt_info)->flow_table_metadata) & mask;
848 *RTE_MBUF_DYNFIELD(pkts[pos + 3], offs, uint32_t *) =
849 rte_be_to_cpu_32(container_of
850 (p3, struct mlx5_cqe,
851 pkt_info)->flow_table_metadata) & mask;
852 if (*RTE_MBUF_DYNFIELD(pkts[pos], offs, uint32_t *))
853 elts[pos]->ol_flags |= rxq->flow_meta_mask;
854 if (*RTE_MBUF_DYNFIELD(pkts[pos + 1], offs, uint32_t *))
855 elts[pos + 1]->ol_flags |= rxq->flow_meta_mask;
856 if (*RTE_MBUF_DYNFIELD(pkts[pos + 2], offs, uint32_t *))
857 elts[pos + 2]->ol_flags |= rxq->flow_meta_mask;
858 if (*RTE_MBUF_DYNFIELD(pkts[pos + 3], offs, uint32_t *))
859 elts[pos + 3]->ol_flags |= rxq->flow_meta_mask;
861 #ifdef MLX5_PMD_SOFT_COUNTERS
862 /* Add up received bytes count. */
863 byte_cnt = vbic_u16(byte_cnt, invalid_mask);
864 rcvd_byte += vget_lane_u64(vpaddl_u32(vpaddl_u16(byte_cnt)), 0);
867 * Break the loop unless more valid CQE is expected, or if
868 * there's a compressed CQE.
870 if (n != MLX5_VPMD_DESCS_PER_LOOP)
873 #ifdef MLX5_PMD_SOFT_COUNTERS
874 rxq->stats.ipackets += nocmp_n;
875 rxq->stats.ibytes += rcvd_byte;
882 #endif /* RTE_PMD_MLX5_RXTX_VEC_NEON_H_ */