xref: /dpdk/drivers/net/mlx5/mlx5_rxtx_vec_neon.h (revision 43fd3624fdfe3a33904a9b64d94306dd3d4f2c13)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2017 6WIND S.A.
3  * Copyright 2017 Mellanox Technologies, Ltd
4  */
5 
6 #ifndef RTE_PMD_MLX5_RXTX_VEC_NEON_H_
7 #define RTE_PMD_MLX5_RXTX_VEC_NEON_H_
8 
9 #include <stdint.h>
10 #include <string.h>
11 #include <stdlib.h>
12 #include <arm_neon.h>
13 
14 #include <rte_bitops.h>
15 #include <rte_mbuf.h>
16 #include <rte_mempool.h>
17 #include <rte_prefetch.h>
18 
19 #include <mlx5_prm.h>
20 
21 #include "mlx5_defs.h"
22 #include "mlx5.h"
23 #include "mlx5_utils.h"
24 #include "mlx5_rxtx.h"
25 #include "mlx5_rxtx_vec.h"
26 #include "mlx5_autoconf.h"
27 
28 /**
29  * Store free buffers to RX SW ring.
30  *
31  * @param elts
32  *   Pointer to SW ring to be filled.
33  * @param pkts
34  *   Pointer to array of packets to be stored.
35  * @param pkts_n
36  *   Number of packets to be stored.
37  */
38 static inline void
39 rxq_copy_mbuf_v(struct rte_mbuf **elts, struct rte_mbuf **pkts, uint16_t n)
40 {
41 	unsigned int pos;
42 	uint16_t p = n & -2;
43 
44 	for (pos = 0; pos < p; pos += 2) {
45 		uint64x2_t mbp;
46 
47 		mbp = vld1q_u64((void *)&elts[pos]);
48 		vst1q_u64((void *)&pkts[pos], mbp);
49 	}
50 	if (n & 1)
51 		pkts[pos] = elts[pos];
52 }
53 
54 /**
55  * Decompress a compressed completion and fill in mbufs in RX SW ring with data
56  * extracted from the title completion descriptor.
57  *
58  * @param rxq
59  *   Pointer to RX queue structure.
60  * @param cq
61  *   Pointer to completion array having a compressed completion at first.
62  * @param elts
63  *   Pointer to SW ring to be filled. The first mbuf has to be pre-built from
64  *   the title completion descriptor to be copied to the rest of mbufs.
65  * @param keep
66  *   Keep unzipping if the next CQE is the miniCQE array.
67  *
68  * @return
69  *   Number of mini-CQEs successfully decompressed.
70  */
71 static inline uint16_t
72 rxq_cq_decompress_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq,
73 		    struct rte_mbuf **elts, bool keep)
74 {
75 	volatile struct mlx5_mini_cqe8 *mcq =
76 		(volatile struct mlx5_mini_cqe8 *)&(cq + !rxq->cqe_comp_layout)->pkt_info;
77 	/* Title packet is pre-built. */
78 	struct rte_mbuf *t_pkt = rxq->cqe_comp_layout ? &rxq->title_pkt : elts[0];
79 	unsigned int pos;
80 	unsigned int i;
81 	unsigned int inv = 0;
82 	/* Mask to shuffle from extracted mini CQE to mbuf. */
83 	const uint8x16_t mcqe_shuf_m1 = {
84 		-1, -1, -1, -1, /* skip packet_type */
85 		 7,  6, -1, -1, /* pkt_len, bswap16 */
86 		 7,  6,         /* data_len, bswap16 */
87 		-1, -1,         /* skip vlan_tci */
88 		 3,  2,  1,  0  /* hash.rss, bswap32 */
89 	};
90 	const uint8x16_t mcqe_shuf_m2 = {
91 		-1, -1, -1, -1, /* skip packet_type */
92 		15, 14, -1, -1, /* pkt_len, bswap16 */
93 		15, 14,         /* data_len, bswap16 */
94 		-1, -1,         /* skip vlan_tci */
95 		11, 10,  9,  8  /* hash.rss, bswap32 */
96 	};
97 	/* Restore the compressed count. Must be 16 bits. */
98 	uint16_t mcqe_n = (rxq->cqe_comp_layout) ?
99 		(MLX5_CQE_NUM_MINIS(cq->op_own) + 1U) : rte_be_to_cpu_32(cq->byte_cnt);
100 	uint16_t pkts_n = mcqe_n;
101 	const uint64x2_t rearm =
102 		vld1q_u64((void *)&t_pkt->rearm_data);
103 	const uint32x4_t rxdf_mask = {
104 		0xffffffff, /* packet_type */
105 		0,          /* skip pkt_len */
106 		0xffff0000, /* vlan_tci, skip data_len */
107 		0,          /* skip hash.rss */
108 	};
109 	const uint8x16_t rxdf =
110 		vandq_u8(vld1q_u8((void *)&t_pkt->rx_descriptor_fields1),
111 			 vreinterpretq_u8_u32(rxdf_mask));
112 	const uint16x8_t crc_adj = {
113 		0, 0,
114 		rxq->crc_present * RTE_ETHER_CRC_LEN, 0,
115 		rxq->crc_present * RTE_ETHER_CRC_LEN, 0,
116 		0, 0
117 	};
118 	uint32x4_t ol_flags = {0, 0, 0, 0};
119 	uint32x4_t ol_flags_mask = {0, 0, 0, 0};
120 #ifdef MLX5_PMD_SOFT_COUNTERS
121 	uint32_t rcvd_byte = 0;
122 #endif
123 	/* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */
124 	const uint8x8_t len_shuf_m = {
125 		 7,  6,         /* 1st mCQE */
126 		15, 14,         /* 2nd mCQE */
127 		23, 22,         /* 3rd mCQE */
128 		31, 30          /* 4th mCQE */
129 	};
130 
131 	/*
132 	 * A. load mCQEs into a 128bit register.
133 	 * B. store rearm data to mbuf.
134 	 * C. combine data from mCQEs with rx_descriptor_fields1.
135 	 * D. store rx_descriptor_fields1.
136 	 * E. store flow tag (rte_flow mark).
137 	 */
138 cycle:
139 	if (rxq->cqe_comp_layout)
140 		rte_prefetch0((volatile void *)(cq + mcqe_n));
141 	for (pos = 0; pos < mcqe_n; ) {
142 		uint8_t *p = RTE_CAST_PTR(uint8_t *, &mcq[pos % 8]);
143 		uint8_t *e0 = (void *)&elts[pos]->rearm_data;
144 		uint8_t *e1 = (void *)&elts[pos + 1]->rearm_data;
145 		uint8_t *e2 = (void *)&elts[pos + 2]->rearm_data;
146 		uint8_t *e3 = (void *)&elts[pos + 3]->rearm_data;
147 		uint16x4_t byte_cnt;
148 #ifdef MLX5_PMD_SOFT_COUNTERS
149 		uint16x4_t invalid_mask =
150 			vcreate_u16(mcqe_n - pos < MLX5_VPMD_DESCS_PER_LOOP ?
151 				    -1UL << ((mcqe_n - pos) *
152 					     sizeof(uint16_t) * 8) : 0);
153 #endif
154 
155 		if (!rxq->cqe_comp_layout)
156 			for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i)
157 				if (likely(pos + i < mcqe_n))
158 					rte_prefetch0((volatile void *)(cq + pos + i));
159 		__asm__ volatile (
160 		/* A.1 load mCQEs into a 128bit register. */
161 		"ld1 {v16.16b - v17.16b}, [%[mcq]] \n\t"
162 		/* B.1 store rearm data to mbuf. */
163 		"st1 {%[rearm].2d}, [%[e0]] \n\t"
164 		"add %[e0], %[e0], #16 \n\t"
165 		"st1 {%[rearm].2d}, [%[e1]] \n\t"
166 		"add %[e1], %[e1], #16 \n\t"
167 		/* C.1 combine data from mCQEs with rx_descriptor_fields1. */
168 		"tbl v18.16b, {v16.16b}, %[mcqe_shuf_m1].16b \n\t"
169 		"tbl v19.16b, {v16.16b}, %[mcqe_shuf_m2].16b \n\t"
170 		"sub v18.8h, v18.8h, %[crc_adj].8h \n\t"
171 		"sub v19.8h, v19.8h, %[crc_adj].8h \n\t"
172 		"orr v18.16b, v18.16b, %[rxdf].16b \n\t"
173 		"orr v19.16b, v19.16b, %[rxdf].16b \n\t"
174 		/* D.1 store rx_descriptor_fields1. */
175 		"st1 {v18.2d}, [%[e0]] \n\t"
176 		"st1 {v19.2d}, [%[e1]] \n\t"
177 		/* B.1 store rearm data to mbuf. */
178 		"st1 {%[rearm].2d}, [%[e2]] \n\t"
179 		"add %[e2], %[e2], #16 \n\t"
180 		"st1 {%[rearm].2d}, [%[e3]] \n\t"
181 		"add %[e3], %[e3], #16 \n\t"
182 		/* C.1 combine data from mCQEs with rx_descriptor_fields1. */
183 		"tbl v18.16b, {v17.16b}, %[mcqe_shuf_m1].16b \n\t"
184 		"tbl v19.16b, {v17.16b}, %[mcqe_shuf_m2].16b \n\t"
185 		"sub v18.8h, v18.8h, %[crc_adj].8h \n\t"
186 		"sub v19.8h, v19.8h, %[crc_adj].8h \n\t"
187 		"orr v18.16b, v18.16b, %[rxdf].16b \n\t"
188 		"orr v19.16b, v19.16b, %[rxdf].16b \n\t"
189 		/* D.1 store rx_descriptor_fields1. */
190 		"st1 {v18.2d}, [%[e2]] \n\t"
191 		"st1 {v19.2d}, [%[e3]] \n\t"
192 #ifdef MLX5_PMD_SOFT_COUNTERS
193 		"tbl %[byte_cnt].8b, {v16.16b - v17.16b}, %[len_shuf_m].8b \n\t"
194 #endif
195 		:[byte_cnt]"=&w"(byte_cnt)
196 		:[mcq]"r"(p),
197 		 [rxdf]"w"(rxdf),
198 		 [rearm]"w"(rearm),
199 		 [e3]"r"(e3), [e2]"r"(e2), [e1]"r"(e1), [e0]"r"(e0),
200 		 [mcqe_shuf_m1]"w"(mcqe_shuf_m1),
201 		 [mcqe_shuf_m2]"w"(mcqe_shuf_m2),
202 		 [crc_adj]"w"(crc_adj),
203 		 [len_shuf_m]"w"(len_shuf_m)
204 		:"memory", "v16", "v17", "v18", "v19");
205 #ifdef MLX5_PMD_SOFT_COUNTERS
206 		byte_cnt = vbic_u16(byte_cnt, invalid_mask);
207 		rcvd_byte += vget_lane_u64(vpaddl_u32(vpaddl_u16(byte_cnt)), 0);
208 #endif
209 		if (rxq->mark) {
210 			if (rxq->mcqe_format !=
211 			    MLX5_CQE_RESP_FORMAT_FTAG_STRIDX) {
212 				const uint32_t flow_tag = t_pkt->hash.fdir.hi;
213 
214 				/* E.1 store flow tag (rte_flow mark). */
215 				elts[pos]->hash.fdir.hi = flow_tag;
216 				elts[pos + 1]->hash.fdir.hi = flow_tag;
217 				elts[pos + 2]->hash.fdir.hi = flow_tag;
218 				elts[pos + 3]->hash.fdir.hi = flow_tag;
219 			}  else {
220 				const uint32x4_t flow_mark_adj = {
221 					-1, -1, -1, -1 };
222 				const uint8x16_t flow_mark_shuf = {
223 					28, 24, 25, -1,
224 					20, 16, 17, -1,
225 					12,  8,  9, -1,
226 					 4,  0,  1, -1};
227 				/* Extract flow_tag field. */
228 				const uint32x4_t ft_mask =
229 					vdupq_n_u32(MLX5_FLOW_MARK_DEFAULT);
230 				const uint32x4_t fdir_flags =
231 					vdupq_n_u32(RTE_MBUF_F_RX_FDIR);
232 				const uint32x4_t fdir_all_flags =
233 					vdupq_n_u32(RTE_MBUF_F_RX_FDIR |
234 						    rxq->mark_flag);
235 				uint32x4_t fdir_id_flags =
236 					vdupq_n_u32(rxq->mark_flag);
237 				uint32x4_t invalid_mask, ftag;
238 
239 				__asm__ volatile
240 				/* A.1 load mCQEs into a 128bit register. */
241 				("ld1 {v16.16b - v17.16b}, [%[mcq]]\n\t"
242 				/* Extract flow_tag. */
243 				 "tbl %[ftag].16b, {v16.16b - v17.16b}, %[flow_mark_shuf].16b\n\t"
244 				: [ftag]"=&w"(ftag)
245 				: [mcq]"r"(p),
246 				  [flow_mark_shuf]"w"(flow_mark_shuf)
247 				: "memory", "v16", "v17");
248 				invalid_mask = vceqzq_u32(ftag);
249 				ol_flags_mask = vorrq_u32(ol_flags_mask,
250 							  fdir_all_flags);
251 				/* Set RTE_MBUF_F_RX_FDIR if flow tag is non-zero. */
252 				ol_flags = vorrq_u32(ol_flags,
253 					vbicq_u32(fdir_flags, invalid_mask));
254 				/* Mask out invalid entries. */
255 				fdir_id_flags = vbicq_u32(fdir_id_flags,
256 							  invalid_mask);
257 				/* Check if flow tag MLX5_FLOW_MARK_DEFAULT. */
258 				ol_flags = vorrq_u32(ol_flags,
259 					vbicq_u32(fdir_id_flags,
260 						  vceqq_u32(ftag, ft_mask)));
261 				ftag = vaddq_u32(ftag, flow_mark_adj);
262 				elts[pos]->hash.fdir.hi =
263 					vgetq_lane_u32(ftag, 3);
264 				elts[pos + 1]->hash.fdir.hi =
265 					vgetq_lane_u32(ftag, 2);
266 				elts[pos + 2]->hash.fdir.hi =
267 					vgetq_lane_u32(ftag, 1);
268 				elts[pos + 3]->hash.fdir.hi =
269 					vgetq_lane_u32(ftag, 0);
270 				}
271 		}
272 		if (unlikely(rxq->mcqe_format !=
273 			     MLX5_CQE_RESP_FORMAT_HASH)) {
274 			if (rxq->mcqe_format ==
275 			    MLX5_CQE_RESP_FORMAT_L34H_STRIDX) {
276 				const uint8_t pkt_info =
277 					(cq->pkt_info & 0x3) << 6;
278 				const uint8_t pkt_hdr0 =
279 					mcq[pos % 8].hdr_type;
280 				const uint8_t pkt_hdr1 =
281 					mcq[pos % 8 + 1].hdr_type;
282 				const uint8_t pkt_hdr2 =
283 					mcq[pos % 8 + 2].hdr_type;
284 				const uint8_t pkt_hdr3 =
285 					mcq[pos % 8 + 3].hdr_type;
286 				const uint32x4_t vlan_mask =
287 					vdupq_n_u32(RTE_MBUF_F_RX_VLAN |
288 						    RTE_MBUF_F_RX_VLAN_STRIPPED);
289 				const uint32x4_t cv_mask =
290 					vdupq_n_u32(MLX5_CQE_VLAN_STRIPPED);
291 				const uint32x4_t pkt_cv = {
292 					pkt_hdr0 & 0x1, pkt_hdr1 & 0x1,
293 					pkt_hdr2 & 0x1, pkt_hdr3 & 0x1};
294 
295 				ol_flags_mask = vorrq_u32(ol_flags_mask,
296 							  vlan_mask);
297 				ol_flags = vorrq_u32(ol_flags,
298 						vandq_u32(vlan_mask,
299 						vceqq_u32(pkt_cv, cv_mask)));
300 				elts[pos]->packet_type =
301 					mlx5_ptype_table[(pkt_hdr0 >> 2) |
302 							 pkt_info];
303 				elts[pos + 1]->packet_type =
304 					mlx5_ptype_table[(pkt_hdr1 >> 2) |
305 							 pkt_info];
306 				elts[pos + 2]->packet_type =
307 					mlx5_ptype_table[(pkt_hdr2 >> 2) |
308 							 pkt_info];
309 				elts[pos + 3]->packet_type =
310 					mlx5_ptype_table[(pkt_hdr3 >> 2) |
311 							 pkt_info];
312 				if (rxq->tunnel) {
313 					elts[pos]->packet_type |=
314 						!!(((pkt_hdr0 >> 2) |
315 						pkt_info) & (1 << 6));
316 					elts[pos + 1]->packet_type |=
317 						!!(((pkt_hdr1 >> 2) |
318 						pkt_info) & (1 << 6));
319 					elts[pos + 2]->packet_type |=
320 						!!(((pkt_hdr2 >> 2) |
321 						pkt_info) & (1 << 6));
322 					elts[pos + 3]->packet_type |=
323 						!!(((pkt_hdr3 >> 2) |
324 						pkt_info) & (1 << 6));
325 				}
326 			}
327 			const uint32x4_t hash_flags =
328 				vdupq_n_u32(RTE_MBUF_F_RX_RSS_HASH);
329 			const uint32x4_t rearm_flags =
330 				vdupq_n_u32((uint32_t)t_pkt->ol_flags);
331 
332 			ol_flags_mask = vorrq_u32(ol_flags_mask, hash_flags);
333 			ol_flags = vorrq_u32(ol_flags,
334 					vbicq_u32(rearm_flags, ol_flags_mask));
335 			elts[pos]->ol_flags = vgetq_lane_u32(ol_flags, 3);
336 			elts[pos + 1]->ol_flags = vgetq_lane_u32(ol_flags, 2);
337 			elts[pos + 2]->ol_flags = vgetq_lane_u32(ol_flags, 1);
338 			elts[pos + 3]->ol_flags = vgetq_lane_u32(ol_flags, 0);
339 			elts[pos]->hash.rss = 0;
340 			elts[pos + 1]->hash.rss = 0;
341 			elts[pos + 2]->hash.rss = 0;
342 			elts[pos + 3]->hash.rss = 0;
343 		}
344 		if (rxq->dynf_meta) {
345 			int32_t offs = rxq->flow_meta_offset;
346 			const uint32_t meta =
347 				*RTE_MBUF_DYNFIELD(t_pkt, offs, uint32_t *);
348 
349 			/* Check if title packet has valid metadata. */
350 			if (meta) {
351 				MLX5_ASSERT(t_pkt->ol_flags &
352 					    rxq->flow_meta_mask);
353 				*RTE_MBUF_DYNFIELD(elts[pos], offs,
354 							uint32_t *) = meta;
355 				*RTE_MBUF_DYNFIELD(elts[pos + 1], offs,
356 							uint32_t *) = meta;
357 				*RTE_MBUF_DYNFIELD(elts[pos + 2], offs,
358 							uint32_t *) = meta;
359 				*RTE_MBUF_DYNFIELD(elts[pos + 3], offs,
360 							uint32_t *) = meta;
361 			}
362 		}
363 		pos += MLX5_VPMD_DESCS_PER_LOOP;
364 		/* Move to next CQE and invalidate consumed CQEs. */
365 		if (!rxq->cqe_comp_layout) {
366 			if (!(pos & 0x7) && pos < mcqe_n) {
367 				if (pos + 8 < mcqe_n)
368 					rte_prefetch0((volatile void *)(cq + pos + 8));
369 				mcq = (volatile struct mlx5_mini_cqe8 *)&(cq + pos)->pkt_info;
370 				for (i = 0; i < 8; ++i)
371 					cq[inv++].op_own = MLX5_CQE_INVALIDATE;
372 			}
373 		}
374 	}
375 	if (rxq->cqe_comp_layout && keep) {
376 		int ret;
377 		/* Keep unzipping if the next CQE is the miniCQE array. */
378 		cq = &cq[mcqe_n];
379 		ret = check_cqe_iteration(cq, rxq->cqe_n, rxq->cq_ci + pkts_n);
380 		if (ret == MLX5_CQE_STATUS_SW_OWN &&
381 		    MLX5_CQE_FORMAT(cq->op_own) == MLX5_COMPRESSED) {
382 			pos = 0;
383 			elts = &elts[mcqe_n];
384 			mcq = (volatile struct mlx5_mini_cqe8 *)cq;
385 			mcqe_n = MLX5_CQE_NUM_MINIS(cq->op_own) + 1;
386 			pkts_n += mcqe_n;
387 			goto cycle;
388 		}
389 	} else {
390 		/* Invalidate the rest of CQEs. */
391 		for (; inv < pkts_n; ++inv)
392 			cq[inv].op_own = MLX5_CQE_INVALIDATE;
393 	}
394 #ifdef MLX5_PMD_SOFT_COUNTERS
395 	rxq->stats.ipackets += pkts_n;
396 	rxq->stats.ibytes += rcvd_byte;
397 #endif
398 	return pkts_n;
399 }
400 
401 /**
402  * Calculate packet type and offload flag for mbuf and store it.
403  *
404  * @param rxq
405  *   Pointer to RX queue structure.
406  * @param ptype_info
407  *   Array of four 4bytes packet type info extracted from the original
408  *   completion descriptor.
409  * @param flow_tag
410  *   Array of four 4bytes flow ID extracted from the original completion
411  *   descriptor.
412  * @param op_err
413  *   Opcode vector having responder error status. Each field is 4B.
414  * @param pkts
415  *   Pointer to array of packets to be filled.
416  */
417 static inline void
418 rxq_cq_to_ptype_oflags_v(struct mlx5_rxq_data *rxq,
419 			 uint32x4_t ptype_info, uint32x4_t flow_tag,
420 			 uint16x4_t op_err, struct rte_mbuf **pkts)
421 {
422 	uint16x4_t ptype;
423 	uint32x4_t pinfo, cv_flags;
424 	uint32x4_t ol_flags =
425 		vdupq_n_u32(rxq->rss_hash * RTE_MBUF_F_RX_RSS_HASH |
426 			    rxq->hw_timestamp * rxq->timestamp_rx_flag);
427 	const uint32x4_t ptype_ol_mask = { 0x106, 0x106, 0x106, 0x106 };
428 	const uint8x16_t cv_flag_sel = {
429 		0,
430 		(uint8_t)(RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED),
431 		(uint8_t)(RTE_MBUF_F_RX_IP_CKSUM_GOOD >> 1),
432 		0,
433 		(uint8_t)(RTE_MBUF_F_RX_L4_CKSUM_GOOD >> 1),
434 		0,
435 		(uint8_t)((RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_GOOD) >> 1),
436 		0, 0, 0, 0, 0, 0, 0, 0, 0
437 	};
438 	const uint32x4_t cv_mask =
439 		vdupq_n_u32(RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_GOOD |
440 			    RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED);
441 	const uint64x2_t mbuf_init = vld1q_u64
442 				((const uint64_t *)&rxq->mbuf_initializer);
443 	uint64x2_t rearm0, rearm1, rearm2, rearm3;
444 	uint8_t pt_idx0, pt_idx1, pt_idx2, pt_idx3;
445 
446 	if (rxq->mark) {
447 		const uint32x4_t ft_def = vdupq_n_u32(MLX5_FLOW_MARK_DEFAULT);
448 		const uint32x4_t fdir_flags = vdupq_n_u32(RTE_MBUF_F_RX_FDIR);
449 		uint32x4_t fdir_id_flags = vdupq_n_u32(rxq->mark_flag);
450 		uint32x4_t invalid_mask;
451 
452 		/* Check if flow tag is non-zero then set RTE_MBUF_F_RX_FDIR. */
453 		invalid_mask = vceqzq_u32(flow_tag);
454 		ol_flags = vorrq_u32(ol_flags,
455 				     vbicq_u32(fdir_flags, invalid_mask));
456 		/* Mask out invalid entries. */
457 		fdir_id_flags = vbicq_u32(fdir_id_flags, invalid_mask);
458 		/* Check if flow tag MLX5_FLOW_MARK_DEFAULT. */
459 		ol_flags = vorrq_u32(ol_flags,
460 				     vbicq_u32(fdir_id_flags,
461 					       vceqq_u32(flow_tag, ft_def)));
462 	}
463 	/*
464 	 * ptype_info has the following:
465 	 * bit[1]     = l3_ok
466 	 * bit[2]     = l4_ok
467 	 * bit[8]     = cv
468 	 * bit[11:10] = l3_hdr_type
469 	 * bit[14:12] = l4_hdr_type
470 	 * bit[15]    = ip_frag
471 	 * bit[16]    = tunneled
472 	 * bit[17]    = outer_l3_type
473 	 */
474 	ptype = vshrn_n_u32(ptype_info, 10);
475 	/* Errored packets will have RTE_PTYPE_ALL_MASK. */
476 	ptype = vorr_u16(ptype, op_err);
477 	pt_idx0 = vget_lane_u8(vreinterpret_u8_u16(ptype), 6);
478 	pt_idx1 = vget_lane_u8(vreinterpret_u8_u16(ptype), 4);
479 	pt_idx2 = vget_lane_u8(vreinterpret_u8_u16(ptype), 2);
480 	pt_idx3 = vget_lane_u8(vreinterpret_u8_u16(ptype), 0);
481 	pkts[0]->packet_type = mlx5_ptype_table[pt_idx0] |
482 			       !!(pt_idx0 & (1 << 6)) * rxq->tunnel;
483 	pkts[1]->packet_type = mlx5_ptype_table[pt_idx1] |
484 			       !!(pt_idx1 & (1 << 6)) * rxq->tunnel;
485 	pkts[2]->packet_type = mlx5_ptype_table[pt_idx2] |
486 			       !!(pt_idx2 & (1 << 6)) * rxq->tunnel;
487 	pkts[3]->packet_type = mlx5_ptype_table[pt_idx3] |
488 			       !!(pt_idx3 & (1 << 6)) * rxq->tunnel;
489 	/* Fill flags for checksum and VLAN. */
490 	pinfo = vandq_u32(ptype_info, ptype_ol_mask);
491 	pinfo = vreinterpretq_u32_u8(
492 		vqtbl1q_u8(cv_flag_sel, vreinterpretq_u8_u32(pinfo)));
493 	/* Locate checksum flags at byte[2:1] and merge with VLAN flags. */
494 	cv_flags = vshlq_n_u32(pinfo, 9);
495 	cv_flags = vorrq_u32(pinfo, cv_flags);
496 	/* Move back flags to start from byte[0]. */
497 	cv_flags = vshrq_n_u32(cv_flags, 8);
498 	/* Mask out garbage bits. */
499 	cv_flags = vandq_u32(cv_flags, cv_mask);
500 	/* Merge to ol_flags. */
501 	ol_flags = vorrq_u32(ol_flags, cv_flags);
502 	/* Merge mbuf_init and ol_flags, and store. */
503 	rearm0 = vreinterpretq_u64_u32(vsetq_lane_u32
504 					(vgetq_lane_u32(ol_flags, 3),
505 					 vreinterpretq_u32_u64(mbuf_init), 2));
506 	rearm1 = vreinterpretq_u64_u32(vsetq_lane_u32
507 					(vgetq_lane_u32(ol_flags, 2),
508 					 vreinterpretq_u32_u64(mbuf_init), 2));
509 	rearm2 = vreinterpretq_u64_u32(vsetq_lane_u32
510 					(vgetq_lane_u32(ol_flags, 1),
511 					 vreinterpretq_u32_u64(mbuf_init), 2));
512 	rearm3 = vreinterpretq_u64_u32(vsetq_lane_u32
513 					(vgetq_lane_u32(ol_flags, 0),
514 					 vreinterpretq_u32_u64(mbuf_init), 2));
515 
516 	vst1q_u64((void *)&pkts[0]->rearm_data, rearm0);
517 	vst1q_u64((void *)&pkts[1]->rearm_data, rearm1);
518 	vst1q_u64((void *)&pkts[2]->rearm_data, rearm2);
519 	vst1q_u64((void *)&pkts[3]->rearm_data, rearm3);
520 }
521 
522 /**
523  * Process a non-compressed completion and fill in mbufs in RX SW ring
524  * with data extracted from the title completion descriptor.
525  *
526  * @param rxq
527  *   Pointer to RX queue structure.
528  * @param cq
529  *   Pointer to completion array having a non-compressed completion at first.
530  * @param elts
531  *   Pointer to SW ring to be filled. The first mbuf has to be pre-built from
532  *   the title completion descriptor to be copied to the rest of mbufs.
533  * @param[out] pkts
534  *   Array to store received packets.
535  * @param pkts_n
536  *   Maximum number of packets in array.
537  * @param[out] err
538  *   Pointer to a flag. Set non-zero value if pkts array has at least one error
539  *   packet to handle.
540  * @param[out] comp
541  *   Pointer to a index. Set it to the first compressed completion if any.
542  *
543  * @return
544  *   Number of CQEs successfully processed.
545  */
546 static inline uint16_t
547 rxq_cq_process_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq,
548 		 struct rte_mbuf **elts, struct rte_mbuf **pkts,
549 		 uint16_t pkts_n, uint64_t *err, uint64_t *comp)
550 {
551 	const uint16_t q_n = 1 << rxq->cqe_n;
552 	const uint16_t q_mask = q_n - 1;
553 	unsigned int pos, adj;
554 	uint64_t n = 0;
555 	uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
556 	uint16_t nocmp_n = 0;
557 	const uint16x4_t validity = vdup_n_u16((rxq->cq_ci >> rxq->cqe_n) << 8);
558 	const uint16x4_t ownership = vdup_n_u16(!(rxq->cq_ci & (q_mask + 1)));
559 	const uint16x4_t vic_check = vcreate_u16(0xff00ff00ff00ff00);
560 	const uint16x4_t owner_check = vcreate_u16(0x0001000100010001);
561 	const uint16x4_t opcode_check = vcreate_u16(0x00f000f000f000f0);
562 	const uint16x4_t format_check = vcreate_u16(0x000c000c000c000c);
563 	const uint16x4_t resp_err_check = vcreate_u16(0x00e000e000e000e0);
564 #ifdef MLX5_PMD_SOFT_COUNTERS
565 	uint32_t rcvd_byte = 0;
566 #endif
567 	/* Mask to generate 16B length vector. */
568 	const uint8x8_t len_shuf_m = {
569 		52, 53,         /* 4th CQE */
570 		36, 37,         /* 3rd CQE */
571 		20, 21,         /* 2nd CQE */
572 		 4,  5          /* 1st CQE */
573 	};
574 	/* Mask to extract 16B data from a 64B CQE. */
575 	const uint8x16_t cqe_shuf_m = {
576 		28, 29,         /* hdr_type_etc */
577 		 0,             /* pkt_info */
578 		62,             /* validity_iteration_count */
579 		47, 46,         /* byte_cnt, bswap16 */
580 		31, 30,         /* vlan_info, bswap16 */
581 		15, 14, 13, 12, /* rx_hash_res, bswap32 */
582 		57, 58, 59,     /* flow_tag */
583 		63              /* op_own */
584 	};
585 	/* Mask to generate 16B data for mbuf. */
586 	const uint8x16_t mb_shuf_m = {
587 		 4,  5, -1, -1, /* pkt_len */
588 		 4,  5,         /* data_len */
589 		 6,  7,         /* vlan_tci */
590 		 8,  9, 10, 11, /* hash.rss */
591 		12, 13, 14, -1  /* hash.fdir.hi */
592 	};
593 	/* Mask to generate 16B owner vector. */
594 	const uint8x8_t owner_shuf_m = {
595 		63, 51,         /* 4th CQE */
596 		47, 35,         /* 3rd CQE */
597 		31, 19,         /* 2nd CQE */
598 		15,  3          /* 1st CQE */
599 	};
600 	/* Mask to generate a vector having packet_type/ol_flags. */
601 	const uint8x16_t ptype_shuf_m = {
602 		48, 49, 50, -1, /* 4th CQE */
603 		32, 33, 34, -1, /* 3rd CQE */
604 		16, 17, 18, -1, /* 2nd CQE */
605 		 0,  1,  2, -1  /* 1st CQE */
606 	};
607 	/* Mask to generate a vector having flow tags. */
608 	const uint8x16_t ftag_shuf_m = {
609 		60, 61, 62, -1, /* 4th CQE */
610 		44, 45, 46, -1, /* 3rd CQE */
611 		28, 29, 30, -1, /* 2nd CQE */
612 		12, 13, 14, -1  /* 1st CQE */
613 	};
614 	const uint16x8_t crc_adj = {
615 		0, 0, rxq->crc_present * RTE_ETHER_CRC_LEN, 0, 0, 0, 0, 0
616 	};
617 	const uint32x4_t flow_mark_adj = { 0, 0, 0, rxq->mark * (-1) };
618 
619 	/*
620 	 * Note that vectors have reverse order - {v3, v2, v1, v0}, because
621 	 * there's no instruction to count trailing zeros. rte_clz64() is
622 	 * used instead.
623 	 *
624 	 * A. copy 4 mbuf pointers from elts ring to returning pkts.
625 	 * B. load 64B CQE and extract necessary fields
626 	 *    Final 16bytes cqes[] extracted from original 64bytes CQE has the
627 	 *    following structure:
628 	 *        struct {
629 	 *          uint16_t hdr_type_etc;
630 	 *          uint8_t  pkt_info;
631 	 *          uint8_t  validity_iteration_count;
632 	 *          uint16_t byte_cnt;
633 	 *          uint16_t vlan_info;
634 	 *          uint32_t rx_has_res;
635 	 *          uint8_t  flow_tag[3];
636 	 *          uint8_t  op_own;
637 	 *        } c;
638 	 * C. fill in mbuf.
639 	 * D. get valid CQEs.
640 	 * E. find compressed CQE.
641 	 */
642 	for (pos = 0;
643 	     pos < pkts_n;
644 	     pos += MLX5_VPMD_DESCS_PER_LOOP) {
645 		uint16x4_t op_own;
646 		uint16x4_t opcode, owner_mask, invalid_mask;
647 		uint16x4_t comp_mask, mini_mask;
648 		uint16x4_t mask;
649 		uint16x4_t byte_cnt;
650 		uint32x4_t ptype_info, flow_tag;
651 		register uint64x2_t c0, c1, c2, c3;
652 		uint8_t *p0, *p1, *p2, *p3;
653 		uint8_t *e0 = (void *)&elts[pos]->pkt_len;
654 		uint8_t *e1 = (void *)&elts[pos + 1]->pkt_len;
655 		uint8_t *e2 = (void *)&elts[pos + 2]->pkt_len;
656 		uint8_t *e3 = (void *)&elts[pos + 3]->pkt_len;
657 		void *elts_p = (void *)&elts[pos];
658 		void *pkts_p = (void *)&pkts[pos];
659 
660 		/* A.0 do not cross the end of CQ. */
661 		mask = vcreate_u16(pkts_n - pos < MLX5_VPMD_DESCS_PER_LOOP ?
662 				   -1UL >> ((pkts_n - pos) *
663 					    sizeof(uint16_t) * 8) : 0);
664 		p0 = RTE_PTR_UNQUAL(&cq[pos].pkt_info);
665 		p1 = p0 + (pkts_n - pos > 1) * sizeof(struct mlx5_cqe);
666 		p2 = p1 + (pkts_n - pos > 2) * sizeof(struct mlx5_cqe);
667 		p3 = p2 + (pkts_n - pos > 3) * sizeof(struct mlx5_cqe);
668 		/* B.0 (CQE 3) load a block having op_own. */
669 		c3 = vld1q_u64((uint64_t *)(p3 + 48));
670 		/* B.0 (CQE 2) load a block having op_own. */
671 		c2 = vld1q_u64((uint64_t *)(p2 + 48));
672 		/* B.0 (CQE 1) load a block having op_own. */
673 		c1 = vld1q_u64((uint64_t *)(p1 + 48));
674 		/* B.0 (CQE 0) load a block having op_own. */
675 		c0 = vld1q_u64((uint64_t *)(p0 + 48));
676 		/* Synchronize for loading the rest of blocks. */
677 		rte_io_rmb();
678 		/* B.0 (CQE 3) reload lower half of the block. */
679 		c3 = vld1q_lane_u64((uint64_t *)(p3 + 48), c3, 0);
680 		/* B.0 (CQE 2) reload lower half of the block. */
681 		c2 = vld1q_lane_u64((uint64_t *)(p2 + 48), c2, 0);
682 		/* B.0 (CQE 1) reload lower half of the block. */
683 		c1 = vld1q_lane_u64((uint64_t *)(p1 + 48), c1, 0);
684 		/* B.0 (CQE 0) reload lower half of the block. */
685 		c0 = vld1q_lane_u64((uint64_t *)(p0 + 48), c0, 0);
686 		/* Prefetch next 4 CQEs. */
687 		if (pkts_n - pos >= 2 * MLX5_VPMD_DESCS_PER_LOOP) {
688 			unsigned int next = pos + MLX5_VPMD_DESCS_PER_LOOP;
689 			rte_prefetch_non_temporal(&cq[next]);
690 			rte_prefetch_non_temporal(&cq[next + 1]);
691 			rte_prefetch_non_temporal(&cq[next + 2]);
692 			rte_prefetch_non_temporal(&cq[next + 3]);
693 		}
694 		__asm__ volatile (
695 		/* B.1 (CQE 3) load the rest of blocks. */
696 		"ld1 {v16.16b - v18.16b}, [%[p3]] \n\t"
697 		/* B.2 (CQE 3) move the block having op_own. */
698 		"mov v19.16b, %[c3].16b \n\t"
699 		/* B.3 (CQE 3) extract 16B fields. */
700 		"tbl v23.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
701 		/* B.1 (CQE 2) load the rest of blocks. */
702 		"ld1 {v16.16b - v18.16b}, [%[p2]] \n\t"
703 		/* B.4 (CQE 3) adjust CRC length. */
704 		"sub v23.8h, v23.8h, %[crc_adj].8h \n\t"
705 		/* C.1 (CQE 3) generate final structure for mbuf. */
706 		"tbl v15.16b, {v23.16b}, %[mb_shuf_m].16b \n\t"
707 		/* B.2 (CQE 2) move the block having op_own. */
708 		"mov v19.16b, %[c2].16b \n\t"
709 		/* B.3 (CQE 2) extract 16B fields. */
710 		"tbl v22.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
711 		/* B.1 (CQE 1) load the rest of blocks. */
712 		"ld1 {v16.16b - v18.16b}, [%[p1]] \n\t"
713 		/* B.4 (CQE 2) adjust CRC length. */
714 		"sub v22.8h, v22.8h, %[crc_adj].8h \n\t"
715 		/* C.1 (CQE 2) generate final structure for mbuf. */
716 		"tbl v14.16b, {v22.16b}, %[mb_shuf_m].16b \n\t"
717 		/* B.2 (CQE 1) move the block having op_own. */
718 		"mov v19.16b, %[c1].16b \n\t"
719 		/* B.3 (CQE 1) extract 16B fields. */
720 		"tbl v21.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
721 		/* B.1 (CQE 0) load the rest of blocks. */
722 		"ld1 {v16.16b - v18.16b}, [%[p0]] \n\t"
723 		/* B.4 (CQE 1) adjust CRC length. */
724 		"sub v21.8h, v21.8h, %[crc_adj].8h \n\t"
725 		/* C.1 (CQE 1) generate final structure for mbuf. */
726 		"tbl v13.16b, {v21.16b}, %[mb_shuf_m].16b \n\t"
727 		/* B.2 (CQE 0) move the block having op_own. */
728 		"mov v19.16b, %[c0].16b \n\t"
729 		/* A.1 load mbuf pointers. */
730 		"ld1 {v24.2d - v25.2d}, [%[elts_p]] \n\t"
731 		/* B.3 (CQE 0) extract 16B fields. */
732 		"tbl v20.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
733 		/* B.4 (CQE 0) adjust CRC length. */
734 		"sub v20.8h, v20.8h, %[crc_adj].8h \n\t"
735 		/* D.1 extract op_own byte. */
736 		"tbl %[op_own].8b, {v20.16b - v23.16b}, %[owner_shuf_m].8b \n\t"
737 		/* C.2 (CQE 3) adjust flow mark. */
738 		"add v15.4s, v15.4s, %[flow_mark_adj].4s \n\t"
739 		/* C.3 (CQE 3) fill in mbuf - rx_descriptor_fields1. */
740 		"st1 {v15.2d}, [%[e3]] \n\t"
741 		/* C.2 (CQE 2) adjust flow mark. */
742 		"add v14.4s, v14.4s, %[flow_mark_adj].4s \n\t"
743 		/* C.3 (CQE 2) fill in mbuf - rx_descriptor_fields1. */
744 		"st1 {v14.2d}, [%[e2]] \n\t"
745 		/* C.1 (CQE 0) generate final structure for mbuf. */
746 		"tbl v12.16b, {v20.16b}, %[mb_shuf_m].16b \n\t"
747 		/* C.2 (CQE 1) adjust flow mark. */
748 		"add v13.4s, v13.4s, %[flow_mark_adj].4s \n\t"
749 		/* C.3 (CQE 1) fill in mbuf - rx_descriptor_fields1. */
750 		"st1 {v13.2d}, [%[e1]] \n\t"
751 #ifdef MLX5_PMD_SOFT_COUNTERS
752 		/* Extract byte_cnt. */
753 		"tbl %[byte_cnt].8b, {v20.16b - v23.16b}, %[len_shuf_m].8b \n\t"
754 #endif
755 		/* Extract ptype_info. */
756 		"tbl %[ptype_info].16b, {v20.16b - v23.16b}, %[ptype_shuf_m].16b \n\t"
757 		/* Extract flow_tag. */
758 		"tbl %[flow_tag].16b, {v20.16b - v23.16b}, %[ftag_shuf_m].16b \n\t"
759 		/* A.2 copy mbuf pointers. */
760 		"st1 {v24.2d - v25.2d}, [%[pkts_p]] \n\t"
761 		/* C.2 (CQE 0) adjust flow mark. */
762 		"add v12.4s, v12.4s, %[flow_mark_adj].4s \n\t"
763 		/* C.3 (CQE 1) fill in mbuf - rx_descriptor_fields1. */
764 		"st1 {v12.2d}, [%[e0]] \n\t"
765 		:[op_own]"=&w"(op_own),
766 		 [byte_cnt]"=&w"(byte_cnt),
767 		 [ptype_info]"=&w"(ptype_info),
768 		 [flow_tag]"=&w"(flow_tag)
769 		:[p3]"r"(p3), [p2]"r"(p2), [p1]"r"(p1), [p0]"r"(p0),
770 		 [e3]"r"(e3), [e2]"r"(e2), [e1]"r"(e1), [e0]"r"(e0),
771 		 [c3]"w"(c3), [c2]"w"(c2), [c1]"w"(c1), [c0]"w"(c0),
772 		 [elts_p]"r"(elts_p),
773 		 [pkts_p]"r"(pkts_p),
774 		 [cqe_shuf_m]"w"(cqe_shuf_m),
775 		 [mb_shuf_m]"w"(mb_shuf_m),
776 		 [owner_shuf_m]"w"(owner_shuf_m),
777 		 [len_shuf_m]"w"(len_shuf_m),
778 		 [ptype_shuf_m]"w"(ptype_shuf_m),
779 		 [ftag_shuf_m]"w"(ftag_shuf_m),
780 		 [crc_adj]"w"(crc_adj),
781 		 [flow_mark_adj]"w"(flow_mark_adj)
782 		:"memory",
783 		 "v12", "v13", "v14", "v15",
784 		 "v16", "v17", "v18", "v19",
785 		 "v20", "v21", "v22", "v23",
786 		 "v24", "v25");
787 		/* D.2 mask out CQEs belonging to HW. */
788 		if (rxq->cqe_comp_layout) {
789 			owner_mask = vand_u16(op_own, vic_check);
790 			owner_mask = vceq_u16(owner_mask, validity);
791 			owner_mask = vmvn_u16(owner_mask);
792 		} else {
793 			owner_mask = vand_u16(op_own, owner_check);
794 			owner_mask = vceq_u16(owner_mask, ownership);
795 		}
796 		/* D.3 get mask for invalidated CQEs. */
797 		opcode = vand_u16(op_own, opcode_check);
798 		invalid_mask = vceq_u16(opcode_check, opcode);
799 		/* E.1 find compressed CQE format. */
800 		comp_mask = vand_u16(op_own, format_check);
801 		comp_mask = vceq_u16(comp_mask, format_check);
802 		/* D.4 mask out beyond boundary. */
803 		invalid_mask = vorr_u16(invalid_mask, mask);
804 		/* D.5 merge invalid_mask with invalid owner. */
805 		invalid_mask = vorr_u16(invalid_mask, owner_mask);
806 		/* E.2 mask out invalid entries. */
807 		comp_mask = vbic_u16(comp_mask, invalid_mask);
808 		/* E.3 get the first compressed CQE. */
809 		comp_idx = rte_clz64(vget_lane_u64(vreinterpret_u64_u16(comp_mask), 0)) /
810 			(sizeof(uint16_t) * 8);
811 		invalid_mask = vorr_u16(invalid_mask, comp_mask);
812 		/* D.7 count non-compressed valid CQEs. */
813 		n = rte_clz64(vget_lane_u64(vreinterpret_u64_u16(invalid_mask), 0)) /
814 			(sizeof(uint16_t) * 8);
815 		nocmp_n += n;
816 		/*
817 		 * D.2 mask out entries after the compressed CQE.
818 		 *     get the final invalid mask.
819 		 */
820 		mask = vcreate_u16(n < MLX5_VPMD_DESCS_PER_LOOP ?
821 				   -1UL >> (n * sizeof(uint16_t) * 8) : 0);
822 		invalid_mask = vorr_u16(invalid_mask, mask);
823 		/* D.3 check error in opcode. */
824 		adj = (!rxq->cqe_comp_layout &&
825 		       comp_idx != MLX5_VPMD_DESCS_PER_LOOP && comp_idx == n);
826 		mask = vcreate_u16(adj ?
827 			   -1UL >> ((n + 1) * sizeof(uint16_t) * 8) : -1UL);
828 		mini_mask = vand_u16(invalid_mask, mask);
829 		opcode = vceq_u16(resp_err_check, opcode);
830 		opcode = vbic_u16(opcode, mini_mask);
831 		/* D.4 mark if any error is set */
832 		*err |= vget_lane_u64(vreinterpret_u64_u16(opcode), 0);
833 		/* C.4 fill in mbuf - rearm_data and packet_type. */
834 		rxq_cq_to_ptype_oflags_v(rxq, ptype_info, flow_tag,
835 					 opcode, &elts[pos]);
836 		if (unlikely(rxq->shared)) {
837 			pkts[pos]->port = container_of(p0, struct mlx5_cqe,
838 					      pkt_info)->user_index_low;
839 			pkts[pos + 1]->port = container_of(p1, struct mlx5_cqe,
840 					      pkt_info)->user_index_low;
841 			pkts[pos + 2]->port = container_of(p2, struct mlx5_cqe,
842 					      pkt_info)->user_index_low;
843 			pkts[pos + 3]->port = container_of(p3, struct mlx5_cqe,
844 					      pkt_info)->user_index_low;
845 		}
846 		if (unlikely(rxq->hw_timestamp)) {
847 			int offset = rxq->timestamp_offset;
848 			if (rxq->rt_timestamp) {
849 				struct mlx5_dev_ctx_shared *sh = rxq->sh;
850 				uint64_t ts;
851 
852 				ts = rte_be_to_cpu_64
853 					(container_of(p0, struct mlx5_cqe,
854 						      pkt_info)->timestamp);
855 				mlx5_timestamp_set(pkts[pos], offset,
856 					mlx5_txpp_convert_rx_ts(sh, ts));
857 				ts = rte_be_to_cpu_64
858 					(container_of(p1, struct mlx5_cqe,
859 						      pkt_info)->timestamp);
860 				mlx5_timestamp_set(pkts[pos + 1], offset,
861 					mlx5_txpp_convert_rx_ts(sh, ts));
862 				ts = rte_be_to_cpu_64
863 					(container_of(p2, struct mlx5_cqe,
864 						      pkt_info)->timestamp);
865 				mlx5_timestamp_set(pkts[pos + 2], offset,
866 					mlx5_txpp_convert_rx_ts(sh, ts));
867 				ts = rte_be_to_cpu_64
868 					(container_of(p3, struct mlx5_cqe,
869 						      pkt_info)->timestamp);
870 				mlx5_timestamp_set(pkts[pos + 3], offset,
871 					mlx5_txpp_convert_rx_ts(sh, ts));
872 			} else {
873 				mlx5_timestamp_set(pkts[pos], offset,
874 					rte_be_to_cpu_64(container_of(p0,
875 					struct mlx5_cqe, pkt_info)->timestamp));
876 				mlx5_timestamp_set(pkts[pos + 1], offset,
877 					rte_be_to_cpu_64(container_of(p1,
878 					struct mlx5_cqe, pkt_info)->timestamp));
879 				mlx5_timestamp_set(pkts[pos + 2], offset,
880 					rte_be_to_cpu_64(container_of(p2,
881 					struct mlx5_cqe, pkt_info)->timestamp));
882 				mlx5_timestamp_set(pkts[pos + 3], offset,
883 					rte_be_to_cpu_64(container_of(p3,
884 					struct mlx5_cqe, pkt_info)->timestamp));
885 			}
886 		}
887 		if (rxq->dynf_meta) {
888 			/* This code is subject for further optimization. */
889 			int32_t offs = rxq->flow_meta_offset;
890 			uint32_t mask = rxq->flow_meta_port_mask;
891 
892 			*RTE_MBUF_DYNFIELD(pkts[pos], offs, uint32_t *) =
893 				rte_be_to_cpu_32(container_of
894 				(p0, struct mlx5_cqe,
895 				pkt_info)->flow_table_metadata) & mask;
896 			*RTE_MBUF_DYNFIELD(pkts[pos + 1], offs, uint32_t *) =
897 				rte_be_to_cpu_32(container_of
898 				(p1, struct mlx5_cqe,
899 				pkt_info)->flow_table_metadata) & mask;
900 			*RTE_MBUF_DYNFIELD(pkts[pos + 2], offs, uint32_t *) =
901 				rte_be_to_cpu_32(container_of
902 				(p2, struct mlx5_cqe,
903 				pkt_info)->flow_table_metadata) & mask;
904 			*RTE_MBUF_DYNFIELD(pkts[pos + 3], offs, uint32_t *) =
905 				rte_be_to_cpu_32(container_of
906 				(p3, struct mlx5_cqe,
907 				pkt_info)->flow_table_metadata) & mask;
908 			if (*RTE_MBUF_DYNFIELD(pkts[pos], offs, uint32_t *))
909 				elts[pos]->ol_flags |= rxq->flow_meta_mask;
910 			if (*RTE_MBUF_DYNFIELD(pkts[pos + 1], offs, uint32_t *))
911 				elts[pos + 1]->ol_flags |= rxq->flow_meta_mask;
912 			if (*RTE_MBUF_DYNFIELD(pkts[pos + 2], offs, uint32_t *))
913 				elts[pos + 2]->ol_flags |= rxq->flow_meta_mask;
914 			if (*RTE_MBUF_DYNFIELD(pkts[pos + 3], offs, uint32_t *))
915 				elts[pos + 3]->ol_flags |= rxq->flow_meta_mask;
916 		}
917 #ifdef MLX5_PMD_SOFT_COUNTERS
918 		/* Add up received bytes count. */
919 		byte_cnt = vbic_u16(byte_cnt, invalid_mask);
920 		rcvd_byte += vget_lane_u64(vpaddl_u32(vpaddl_u16(byte_cnt)), 0);
921 #endif
922 		/*
923 		 * Break the loop unless more valid CQE is expected, or if
924 		 * there's a compressed CQE.
925 		 */
926 		if (n != MLX5_VPMD_DESCS_PER_LOOP)
927 			break;
928 	}
929 #ifdef MLX5_PMD_SOFT_COUNTERS
930 	rxq->stats.ipackets += nocmp_n;
931 	rxq->stats.ibytes += rcvd_byte;
932 #endif
933 	if (comp_idx == n)
934 		*comp = comp_idx;
935 	return nocmp_n;
936 }
937 
938 #endif /* RTE_PMD_MLX5_RXTX_VEC_NEON_H_ */
939