xref: /dpdk/drivers/net/mlx5/mlx5_rxtx_vec_neon.h (revision 9f3b3a96dec2f4c01cc92a132d763b8887d29e6a)
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_mbuf.h>
15 #include <rte_mempool.h>
16 #include <rte_prefetch.h>
17 
18 #include <mlx5_prm.h>
19 
20 #include "mlx5_defs.h"
21 #include "mlx5.h"
22 #include "mlx5_utils.h"
23 #include "mlx5_rxtx.h"
24 #include "mlx5_rxtx_vec.h"
25 #include "mlx5_autoconf.h"
26 
27 #pragma GCC diagnostic ignored "-Wcast-qual"
28 
29 /**
30  * Store free buffers to RX SW ring.
31  *
32  * @param elts
33  *   Pointer to SW ring to be filled.
34  * @param pkts
35  *   Pointer to array of packets to be stored.
36  * @param pkts_n
37  *   Number of packets to be stored.
38  */
39 static inline void
40 rxq_copy_mbuf_v(struct rte_mbuf **elts, struct rte_mbuf **pkts, uint16_t n)
41 {
42 	unsigned int pos;
43 	uint16_t p = n & -2;
44 
45 	for (pos = 0; pos < p; pos += 2) {
46 		uint64x2_t mbp;
47 
48 		mbp = vld1q_u64((void *)&elts[pos]);
49 		vst1q_u64((void *)&pkts[pos], mbp);
50 	}
51 	if (n & 1)
52 		pkts[pos] = elts[pos];
53 }
54 
55 /**
56  * Decompress a compressed completion and fill in mbufs in RX SW ring with data
57  * extracted from the title completion descriptor.
58  *
59  * @param rxq
60  *   Pointer to RX queue structure.
61  * @param cq
62  *   Pointer to completion array having a compressed completion at first.
63  * @param elts
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.
66  *
67  * @return
68  *   Number of mini-CQEs successfully decompressed.
69  */
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)
73 {
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. */
76 	unsigned int pos;
77 	unsigned int i;
78 	unsigned int inv = 0;
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 */
86 	};
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 */
93 	};
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 */
104 	};
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 = {
109 		0, 0,
110 		rxq->crc_present * RTE_ETHER_CRC_LEN, 0,
111 		rxq->crc_present * RTE_ETHER_CRC_LEN, 0,
112 		0, 0
113 	};
114 	const uint32_t flow_tag = t_pkt->hash.fdir.hi;
115 #ifdef MLX5_PMD_SOFT_COUNTERS
116 	uint32_t rcvd_byte = 0;
117 #endif
118 	/* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */
119 	const uint8x8_t len_shuf_m = {
120 		 7,  6,         /* 1st mCQE */
121 		15, 14,         /* 2nd mCQE */
122 		23, 22,         /* 3rd mCQE */
123 		31, 30          /* 4th mCQE */
124 	};
125 
126 	/*
127 	 * A. load mCQEs into a 128bit register.
128 	 * B. store rearm data to mbuf.
129 	 * C. combine data from mCQEs with rx_descriptor_fields1.
130 	 * D. store rx_descriptor_fields1.
131 	 * E. store flow tag (rte_flow mark).
132 	 */
133 	for (pos = 0; pos < mcqe_n; ) {
134 		uint8_t *p = (void *)&mcq[pos % 8];
135 		uint8_t *e0 = (void *)&elts[pos]->rearm_data;
136 		uint8_t *e1 = (void *)&elts[pos + 1]->rearm_data;
137 		uint8_t *e2 = (void *)&elts[pos + 2]->rearm_data;
138 		uint8_t *e3 = (void *)&elts[pos + 3]->rearm_data;
139 		uint16x4_t byte_cnt;
140 #ifdef MLX5_PMD_SOFT_COUNTERS
141 		uint16x4_t invalid_mask =
142 			vcreate_u16(mcqe_n - pos < MLX5_VPMD_DESCS_PER_LOOP ?
143 				    -1UL << ((mcqe_n - pos) *
144 					     sizeof(uint16_t) * 8) : 0);
145 #endif
146 
147 		for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i)
148 			if (likely(pos + i < mcqe_n))
149 				rte_prefetch0((void *)(cq + pos + i));
150 		__asm__ volatile (
151 		/* A.1 load mCQEs into a 128bit register. */
152 		"ld1 {v16.16b - v17.16b}, [%[mcq]] \n\t"
153 		/* B.1 store rearm data to mbuf. */
154 		"st1 {%[rearm].2d}, [%[e0]] \n\t"
155 		"add %[e0], %[e0], #16 \n\t"
156 		"st1 {%[rearm].2d}, [%[e1]] \n\t"
157 		"add %[e1], %[e1], #16 \n\t"
158 		/* C.1 combine data from mCQEs with rx_descriptor_fields1. */
159 		"tbl v18.16b, {v16.16b}, %[mcqe_shuf_m1].16b \n\t"
160 		"tbl v19.16b, {v16.16b}, %[mcqe_shuf_m2].16b \n\t"
161 		"sub v18.8h, v18.8h, %[crc_adj].8h \n\t"
162 		"sub v19.8h, v19.8h, %[crc_adj].8h \n\t"
163 		"orr v18.16b, v18.16b, %[rxdf].16b \n\t"
164 		"orr v19.16b, v19.16b, %[rxdf].16b \n\t"
165 		/* D.1 store rx_descriptor_fields1. */
166 		"st1 {v18.2d}, [%[e0]] \n\t"
167 		"st1 {v19.2d}, [%[e1]] \n\t"
168 		/* B.1 store rearm data to mbuf. */
169 		"st1 {%[rearm].2d}, [%[e2]] \n\t"
170 		"add %[e2], %[e2], #16 \n\t"
171 		"st1 {%[rearm].2d}, [%[e3]] \n\t"
172 		"add %[e3], %[e3], #16 \n\t"
173 		/* C.1 combine data from mCQEs with rx_descriptor_fields1. */
174 		"tbl v18.16b, {v17.16b}, %[mcqe_shuf_m1].16b \n\t"
175 		"tbl v19.16b, {v17.16b}, %[mcqe_shuf_m2].16b \n\t"
176 		"sub v18.8h, v18.8h, %[crc_adj].8h \n\t"
177 		"sub v19.8h, v19.8h, %[crc_adj].8h \n\t"
178 		"orr v18.16b, v18.16b, %[rxdf].16b \n\t"
179 		"orr v19.16b, v19.16b, %[rxdf].16b \n\t"
180 		/* D.1 store rx_descriptor_fields1. */
181 		"st1 {v18.2d}, [%[e2]] \n\t"
182 		"st1 {v19.2d}, [%[e3]] \n\t"
183 #ifdef MLX5_PMD_SOFT_COUNTERS
184 		"tbl %[byte_cnt].8b, {v16.16b - v17.16b}, %[len_shuf_m].8b \n\t"
185 #endif
186 		:[byte_cnt]"=&w"(byte_cnt)
187 		:[mcq]"r"(p),
188 		 [rxdf]"w"(rxdf),
189 		 [rearm]"w"(rearm),
190 		 [e3]"r"(e3), [e2]"r"(e2), [e1]"r"(e1), [e0]"r"(e0),
191 		 [mcqe_shuf_m1]"w"(mcqe_shuf_m1),
192 		 [mcqe_shuf_m2]"w"(mcqe_shuf_m2),
193 		 [crc_adj]"w"(crc_adj),
194 		 [len_shuf_m]"w"(len_shuf_m)
195 		:"memory", "v16", "v17", "v18", "v19");
196 #ifdef MLX5_PMD_SOFT_COUNTERS
197 		byte_cnt = vbic_u16(byte_cnt, invalid_mask);
198 		rcvd_byte += vget_lane_u64(vpaddl_u32(vpaddl_u16(byte_cnt)), 0);
199 #endif
200 		if (rxq->mark) {
201 			/* E.1 store flow tag (rte_flow mark). */
202 			elts[pos]->hash.fdir.hi = flow_tag;
203 			elts[pos + 1]->hash.fdir.hi = flow_tag;
204 			elts[pos + 2]->hash.fdir.hi = flow_tag;
205 			elts[pos + 3]->hash.fdir.hi = flow_tag;
206 		}
207 		if (rxq->dynf_meta) {
208 			int32_t offs = rxq->flow_meta_offset;
209 			const uint32_t meta =
210 				*RTE_MBUF_DYNFIELD(t_pkt, offs, uint32_t *);
211 
212 			/* Check if title packet has valid metadata. */
213 			if (meta) {
214 				MLX5_ASSERT(t_pkt->ol_flags &
215 					    rxq->flow_meta_mask);
216 				*RTE_MBUF_DYNFIELD(elts[pos], offs,
217 							uint32_t *) = meta;
218 				*RTE_MBUF_DYNFIELD(elts[pos + 1], offs,
219 							uint32_t *) = meta;
220 				*RTE_MBUF_DYNFIELD(elts[pos + 2], offs,
221 							uint32_t *) = meta;
222 				*RTE_MBUF_DYNFIELD(elts[pos + 3], offs,
223 							uint32_t *) = meta;
224 			}
225 		}
226 		pos += MLX5_VPMD_DESCS_PER_LOOP;
227 		/* Move to next CQE and invalidate consumed CQEs. */
228 		if (!(pos & 0x7) && pos < mcqe_n) {
229 			if (pos + 8 < mcqe_n)
230 				rte_prefetch0((void *)(cq + pos + 8));
231 			mcq = (void *)&(cq + pos)->pkt_info;
232 			for (i = 0; i < 8; ++i)
233 				cq[inv++].op_own = MLX5_CQE_INVALIDATE;
234 		}
235 	}
236 	/* Invalidate the rest of CQEs. */
237 	for (; inv < mcqe_n; ++inv)
238 		cq[inv].op_own = MLX5_CQE_INVALIDATE;
239 #ifdef MLX5_PMD_SOFT_COUNTERS
240 	rxq->stats.ipackets += mcqe_n;
241 	rxq->stats.ibytes += rcvd_byte;
242 #endif
243 	rxq->cq_ci += mcqe_n;
244 	return mcqe_n;
245 }
246 
247 /**
248  * Calculate packet type and offload flag for mbuf and store it.
249  *
250  * @param rxq
251  *   Pointer to RX queue structure.
252  * @param ptype_info
253  *   Array of four 4bytes packet type info extracted from the original
254  *   completion descriptor.
255  * @param flow_tag
256  *   Array of four 4bytes flow ID extracted from the original completion
257  *   descriptor.
258  * @param op_err
259  *   Opcode vector having responder error status. Each field is 4B.
260  * @param pkts
261  *   Pointer to array of packets to be filled.
262  */
263 static inline void
264 rxq_cq_to_ptype_oflags_v(struct mlx5_rxq_data *rxq,
265 			 uint32x4_t ptype_info, uint32x4_t flow_tag,
266 			 uint16x4_t op_err, struct rte_mbuf **pkts)
267 {
268 	uint16x4_t ptype;
269 	uint32x4_t pinfo, cv_flags;
270 	uint32x4_t ol_flags =
271 		vdupq_n_u32(rxq->rss_hash * PKT_RX_RSS_HASH |
272 			    rxq->hw_timestamp * rxq->timestamp_rx_flag);
273 	const uint32x4_t ptype_ol_mask = { 0x106, 0x106, 0x106, 0x106 };
274 	const uint8x16_t cv_flag_sel = {
275 		0,
276 		(uint8_t)(PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED),
277 		(uint8_t)(PKT_RX_IP_CKSUM_GOOD >> 1),
278 		0,
279 		(uint8_t)(PKT_RX_L4_CKSUM_GOOD >> 1),
280 		0,
281 		(uint8_t)((PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1),
282 		0, 0, 0, 0, 0, 0, 0, 0, 0
283 	};
284 	const uint32x4_t cv_mask =
285 		vdupq_n_u32(PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
286 			    PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED);
287 	const uint64x2_t mbuf_init = vld1q_u64
288 				((const uint64_t *)&rxq->mbuf_initializer);
289 	uint64x2_t rearm0, rearm1, rearm2, rearm3;
290 	uint8_t pt_idx0, pt_idx1, pt_idx2, pt_idx3;
291 
292 	if (rxq->mark) {
293 		const uint32x4_t ft_def = vdupq_n_u32(MLX5_FLOW_MARK_DEFAULT);
294 		const uint32x4_t fdir_flags = vdupq_n_u32(PKT_RX_FDIR);
295 		uint32x4_t fdir_id_flags = vdupq_n_u32(PKT_RX_FDIR_ID);
296 		uint32x4_t invalid_mask;
297 
298 		/* Check if flow tag is non-zero then set PKT_RX_FDIR. */
299 		invalid_mask = vceqzq_u32(flow_tag);
300 		ol_flags = vorrq_u32(ol_flags,
301 				     vbicq_u32(fdir_flags, invalid_mask));
302 		/* Mask out invalid entries. */
303 		fdir_id_flags = vbicq_u32(fdir_id_flags, invalid_mask);
304 		/* Check if flow tag MLX5_FLOW_MARK_DEFAULT. */
305 		ol_flags = vorrq_u32(ol_flags,
306 				     vbicq_u32(fdir_id_flags,
307 					       vceqq_u32(flow_tag, ft_def)));
308 	}
309 	/*
310 	 * ptype_info has the following:
311 	 * bit[1]     = l3_ok
312 	 * bit[2]     = l4_ok
313 	 * bit[8]     = cv
314 	 * bit[11:10] = l3_hdr_type
315 	 * bit[14:12] = l4_hdr_type
316 	 * bit[15]    = ip_frag
317 	 * bit[16]    = tunneled
318 	 * bit[17]    = outer_l3_type
319 	 */
320 	ptype = vshrn_n_u32(ptype_info, 10);
321 	/* Errored packets will have RTE_PTYPE_ALL_MASK. */
322 	ptype = vorr_u16(ptype, op_err);
323 	pt_idx0 = vget_lane_u8(vreinterpret_u8_u16(ptype), 6);
324 	pt_idx1 = vget_lane_u8(vreinterpret_u8_u16(ptype), 4);
325 	pt_idx2 = vget_lane_u8(vreinterpret_u8_u16(ptype), 2);
326 	pt_idx3 = vget_lane_u8(vreinterpret_u8_u16(ptype), 0);
327 	pkts[0]->packet_type = mlx5_ptype_table[pt_idx0] |
328 			       !!(pt_idx0 & (1 << 6)) * rxq->tunnel;
329 	pkts[1]->packet_type = mlx5_ptype_table[pt_idx1] |
330 			       !!(pt_idx1 & (1 << 6)) * rxq->tunnel;
331 	pkts[2]->packet_type = mlx5_ptype_table[pt_idx2] |
332 			       !!(pt_idx2 & (1 << 6)) * rxq->tunnel;
333 	pkts[3]->packet_type = mlx5_ptype_table[pt_idx3] |
334 			       !!(pt_idx3 & (1 << 6)) * rxq->tunnel;
335 	/* Fill flags for checksum and VLAN. */
336 	pinfo = vandq_u32(ptype_info, ptype_ol_mask);
337 	pinfo = vreinterpretq_u32_u8(
338 		vqtbl1q_u8(cv_flag_sel, vreinterpretq_u8_u32(pinfo)));
339 	/* Locate checksum flags at byte[2:1] and merge with VLAN flags. */
340 	cv_flags = vshlq_n_u32(pinfo, 9);
341 	cv_flags = vorrq_u32(pinfo, cv_flags);
342 	/* Move back flags to start from byte[0]. */
343 	cv_flags = vshrq_n_u32(cv_flags, 8);
344 	/* Mask out garbage bits. */
345 	cv_flags = vandq_u32(cv_flags, cv_mask);
346 	/* Merge to ol_flags. */
347 	ol_flags = vorrq_u32(ol_flags, cv_flags);
348 	/* Merge mbuf_init and ol_flags, and store. */
349 	rearm0 = vreinterpretq_u64_u32(vsetq_lane_u32
350 					(vgetq_lane_u32(ol_flags, 3),
351 					 vreinterpretq_u32_u64(mbuf_init), 2));
352 	rearm1 = vreinterpretq_u64_u32(vsetq_lane_u32
353 					(vgetq_lane_u32(ol_flags, 2),
354 					 vreinterpretq_u32_u64(mbuf_init), 2));
355 	rearm2 = vreinterpretq_u64_u32(vsetq_lane_u32
356 					(vgetq_lane_u32(ol_flags, 1),
357 					 vreinterpretq_u32_u64(mbuf_init), 2));
358 	rearm3 = vreinterpretq_u64_u32(vsetq_lane_u32
359 					(vgetq_lane_u32(ol_flags, 0),
360 					 vreinterpretq_u32_u64(mbuf_init), 2));
361 
362 	vst1q_u64((void *)&pkts[0]->rearm_data, rearm0);
363 	vst1q_u64((void *)&pkts[1]->rearm_data, rearm1);
364 	vst1q_u64((void *)&pkts[2]->rearm_data, rearm2);
365 	vst1q_u64((void *)&pkts[3]->rearm_data, rearm3);
366 }
367 
368 /**
369  * Process a non-compressed completion and fill in mbufs in RX SW ring
370  * with data extracted from the title completion descriptor.
371  *
372  * @param rxq
373  *   Pointer to RX queue structure.
374  * @param cq
375  *   Pointer to completion array having a non-compressed completion at first.
376  * @param elts
377  *   Pointer to SW ring to be filled. The first mbuf has to be pre-built from
378  *   the title completion descriptor to be copied to the rest of mbufs.
379  * @param[out] pkts
380  *   Array to store received packets.
381  * @param pkts_n
382  *   Maximum number of packets in array.
383  * @param[out] err
384  *   Pointer to a flag. Set non-zero value if pkts array has at least one error
385  *   packet to handle.
386  * @param[out] comp
387  *   Pointer to a index. Set it to the first compressed completion if any.
388  *
389  * @return
390  *   Number of CQEs successfully processed.
391  */
392 static inline uint16_t
393 rxq_cq_process_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq,
394 		 struct rte_mbuf **elts, struct rte_mbuf **pkts,
395 		 uint16_t pkts_n, uint64_t *err, uint64_t *comp)
396 {
397 	const uint16_t q_n = 1 << rxq->cqe_n;
398 	const uint16_t q_mask = q_n - 1;
399 	unsigned int pos;
400 	uint64_t n = 0;
401 	uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
402 	uint16_t nocmp_n = 0;
403 	const uint16x4_t ownership = vdup_n_u16(!(rxq->cq_ci & (q_mask + 1)));
404 	const uint16x4_t owner_check = vcreate_u16(0x0001000100010001);
405 	const uint16x4_t opcode_check = vcreate_u16(0x00f000f000f000f0);
406 	const uint16x4_t format_check = vcreate_u16(0x000c000c000c000c);
407 	const uint16x4_t resp_err_check = vcreate_u16(0x00e000e000e000e0);
408 #ifdef MLX5_PMD_SOFT_COUNTERS
409 	uint32_t rcvd_byte = 0;
410 #endif
411 	/* Mask to generate 16B length vector. */
412 	const uint8x8_t len_shuf_m = {
413 		52, 53,         /* 4th CQE */
414 		36, 37,         /* 3rd CQE */
415 		20, 21,         /* 2nd CQE */
416 		 4,  5          /* 1st CQE */
417 	};
418 	/* Mask to extract 16B data from a 64B CQE. */
419 	const uint8x16_t cqe_shuf_m = {
420 		28, 29,         /* hdr_type_etc */
421 		 0,             /* pkt_info */
422 		-1,             /* null */
423 		47, 46,         /* byte_cnt, bswap16 */
424 		31, 30,         /* vlan_info, bswap16 */
425 		15, 14, 13, 12, /* rx_hash_res, bswap32 */
426 		57, 58, 59,     /* flow_tag */
427 		63              /* op_own */
428 	};
429 	/* Mask to generate 16B data for mbuf. */
430 	const uint8x16_t mb_shuf_m = {
431 		 4,  5, -1, -1, /* pkt_len */
432 		 4,  5,         /* data_len */
433 		 6,  7,         /* vlan_tci */
434 		 8,  9, 10, 11, /* hash.rss */
435 		12, 13, 14, -1  /* hash.fdir.hi */
436 	};
437 	/* Mask to generate 16B owner vector. */
438 	const uint8x8_t owner_shuf_m = {
439 		63, -1,         /* 4th CQE */
440 		47, -1,         /* 3rd CQE */
441 		31, -1,         /* 2nd CQE */
442 		15, -1          /* 1st CQE */
443 	};
444 	/* Mask to generate a vector having packet_type/ol_flags. */
445 	const uint8x16_t ptype_shuf_m = {
446 		48, 49, 50, -1, /* 4th CQE */
447 		32, 33, 34, -1, /* 3rd CQE */
448 		16, 17, 18, -1, /* 2nd CQE */
449 		 0,  1,  2, -1  /* 1st CQE */
450 	};
451 	/* Mask to generate a vector having flow tags. */
452 	const uint8x16_t ftag_shuf_m = {
453 		60, 61, 62, -1, /* 4th CQE */
454 		44, 45, 46, -1, /* 3rd CQE */
455 		28, 29, 30, -1, /* 2nd CQE */
456 		12, 13, 14, -1  /* 1st CQE */
457 	};
458 	const uint16x8_t crc_adj = {
459 		0, 0, rxq->crc_present * RTE_ETHER_CRC_LEN, 0, 0, 0, 0, 0
460 	};
461 	const uint32x4_t flow_mark_adj = { 0, 0, 0, rxq->mark * (-1) };
462 
463 	/*
464 	 * Note that vectors have reverse order - {v3, v2, v1, v0}, because
465 	 * there's no instruction to count trailing zeros. __builtin_clzl() is
466 	 * used instead.
467 	 *
468 	 * A. copy 4 mbuf pointers from elts ring to returing pkts.
469 	 * B. load 64B CQE and extract necessary fields
470 	 *    Final 16bytes cqes[] extracted from original 64bytes CQE has the
471 	 *    following structure:
472 	 *        struct {
473 	 *          uint16_t hdr_type_etc;
474 	 *          uint8_t  pkt_info;
475 	 *          uint8_t  rsvd;
476 	 *          uint16_t byte_cnt;
477 	 *          uint16_t vlan_info;
478 	 *          uint32_t rx_has_res;
479 	 *          uint8_t  flow_tag[3];
480 	 *          uint8_t  op_own;
481 	 *        } c;
482 	 * C. fill in mbuf.
483 	 * D. get valid CQEs.
484 	 * E. find compressed CQE.
485 	 */
486 	for (pos = 0;
487 	     pos < pkts_n;
488 	     pos += MLX5_VPMD_DESCS_PER_LOOP) {
489 		uint16x4_t op_own;
490 		uint16x4_t opcode, owner_mask, invalid_mask;
491 		uint16x4_t comp_mask;
492 		uint16x4_t mask;
493 		uint16x4_t byte_cnt;
494 		uint32x4_t ptype_info, flow_tag;
495 		register uint64x2_t c0, c1, c2, c3;
496 		uint8_t *p0, *p1, *p2, *p3;
497 		uint8_t *e0 = (void *)&elts[pos]->pkt_len;
498 		uint8_t *e1 = (void *)&elts[pos + 1]->pkt_len;
499 		uint8_t *e2 = (void *)&elts[pos + 2]->pkt_len;
500 		uint8_t *e3 = (void *)&elts[pos + 3]->pkt_len;
501 		void *elts_p = (void *)&elts[pos];
502 		void *pkts_p = (void *)&pkts[pos];
503 
504 		/* A.0 do not cross the end of CQ. */
505 		mask = vcreate_u16(pkts_n - pos < MLX5_VPMD_DESCS_PER_LOOP ?
506 				   -1UL >> ((pkts_n - pos) *
507 					    sizeof(uint16_t) * 8) : 0);
508 		p0 = (void *)&cq[pos].pkt_info;
509 		p1 = p0 + (pkts_n - pos > 1) * sizeof(struct mlx5_cqe);
510 		p2 = p1 + (pkts_n - pos > 2) * sizeof(struct mlx5_cqe);
511 		p3 = p2 + (pkts_n - pos > 3) * sizeof(struct mlx5_cqe);
512 		/* B.0 (CQE 3) load a block having op_own. */
513 		c3 = vld1q_u64((uint64_t *)(p3 + 48));
514 		/* B.0 (CQE 2) load a block having op_own. */
515 		c2 = vld1q_u64((uint64_t *)(p2 + 48));
516 		/* B.0 (CQE 1) load a block having op_own. */
517 		c1 = vld1q_u64((uint64_t *)(p1 + 48));
518 		/* B.0 (CQE 0) load a block having op_own. */
519 		c0 = vld1q_u64((uint64_t *)(p0 + 48));
520 		/* Synchronize for loading the rest of blocks. */
521 		rte_io_rmb();
522 		/* Prefetch next 4 CQEs. */
523 		if (pkts_n - pos >= 2 * MLX5_VPMD_DESCS_PER_LOOP) {
524 			unsigned int next = pos + MLX5_VPMD_DESCS_PER_LOOP;
525 			rte_prefetch_non_temporal(&cq[next]);
526 			rte_prefetch_non_temporal(&cq[next + 1]);
527 			rte_prefetch_non_temporal(&cq[next + 2]);
528 			rte_prefetch_non_temporal(&cq[next + 3]);
529 		}
530 		__asm__ volatile (
531 		/* B.1 (CQE 3) load the rest of blocks. */
532 		"ld1 {v16.16b - v18.16b}, [%[p3]] \n\t"
533 		/* B.2 (CQE 3) move the block having op_own. */
534 		"mov v19.16b, %[c3].16b \n\t"
535 		/* B.3 (CQE 3) extract 16B fields. */
536 		"tbl v23.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
537 		/* B.1 (CQE 2) load the rest of blocks. */
538 		"ld1 {v16.16b - v18.16b}, [%[p2]] \n\t"
539 		/* B.4 (CQE 3) adjust CRC length. */
540 		"sub v23.8h, v23.8h, %[crc_adj].8h \n\t"
541 		/* C.1 (CQE 3) generate final structure for mbuf. */
542 		"tbl v15.16b, {v23.16b}, %[mb_shuf_m].16b \n\t"
543 		/* B.2 (CQE 2) move the block having op_own. */
544 		"mov v19.16b, %[c2].16b \n\t"
545 		/* B.3 (CQE 2) extract 16B fields. */
546 		"tbl v22.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
547 		/* B.1 (CQE 1) load the rest of blocks. */
548 		"ld1 {v16.16b - v18.16b}, [%[p1]] \n\t"
549 		/* B.4 (CQE 2) adjust CRC length. */
550 		"sub v22.8h, v22.8h, %[crc_adj].8h \n\t"
551 		/* C.1 (CQE 2) generate final structure for mbuf. */
552 		"tbl v14.16b, {v22.16b}, %[mb_shuf_m].16b \n\t"
553 		/* B.2 (CQE 1) move the block having op_own. */
554 		"mov v19.16b, %[c1].16b \n\t"
555 		/* B.3 (CQE 1) extract 16B fields. */
556 		"tbl v21.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
557 		/* B.1 (CQE 0) load the rest of blocks. */
558 		"ld1 {v16.16b - v18.16b}, [%[p0]] \n\t"
559 		/* B.4 (CQE 1) adjust CRC length. */
560 		"sub v21.8h, v21.8h, %[crc_adj].8h \n\t"
561 		/* C.1 (CQE 1) generate final structure for mbuf. */
562 		"tbl v13.16b, {v21.16b}, %[mb_shuf_m].16b \n\t"
563 		/* B.2 (CQE 0) move the block having op_own. */
564 		"mov v19.16b, %[c0].16b \n\t"
565 		/* A.1 load mbuf pointers. */
566 		"ld1 {v24.2d - v25.2d}, [%[elts_p]] \n\t"
567 		/* B.3 (CQE 0) extract 16B fields. */
568 		"tbl v20.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
569 		/* B.4 (CQE 0) adjust CRC length. */
570 		"sub v20.8h, v20.8h, %[crc_adj].8h \n\t"
571 		/* D.1 extract op_own byte. */
572 		"tbl %[op_own].8b, {v20.16b - v23.16b}, %[owner_shuf_m].8b \n\t"
573 		/* C.2 (CQE 3) adjust flow mark. */
574 		"add v15.4s, v15.4s, %[flow_mark_adj].4s \n\t"
575 		/* C.3 (CQE 3) fill in mbuf - rx_descriptor_fields1. */
576 		"st1 {v15.2d}, [%[e3]] \n\t"
577 		/* C.2 (CQE 2) adjust flow mark. */
578 		"add v14.4s, v14.4s, %[flow_mark_adj].4s \n\t"
579 		/* C.3 (CQE 2) fill in mbuf - rx_descriptor_fields1. */
580 		"st1 {v14.2d}, [%[e2]] \n\t"
581 		/* C.1 (CQE 0) generate final structure for mbuf. */
582 		"tbl v12.16b, {v20.16b}, %[mb_shuf_m].16b \n\t"
583 		/* C.2 (CQE 1) adjust flow mark. */
584 		"add v13.4s, v13.4s, %[flow_mark_adj].4s \n\t"
585 		/* C.3 (CQE 1) fill in mbuf - rx_descriptor_fields1. */
586 		"st1 {v13.2d}, [%[e1]] \n\t"
587 #ifdef MLX5_PMD_SOFT_COUNTERS
588 		/* Extract byte_cnt. */
589 		"tbl %[byte_cnt].8b, {v20.16b - v23.16b}, %[len_shuf_m].8b \n\t"
590 #endif
591 		/* Extract ptype_info. */
592 		"tbl %[ptype_info].16b, {v20.16b - v23.16b}, %[ptype_shuf_m].16b \n\t"
593 		/* Extract flow_tag. */
594 		"tbl %[flow_tag].16b, {v20.16b - v23.16b}, %[ftag_shuf_m].16b \n\t"
595 		/* A.2 copy mbuf pointers. */
596 		"st1 {v24.2d - v25.2d}, [%[pkts_p]] \n\t"
597 		/* C.2 (CQE 0) adjust flow mark. */
598 		"add v12.4s, v12.4s, %[flow_mark_adj].4s \n\t"
599 		/* C.3 (CQE 1) fill in mbuf - rx_descriptor_fields1. */
600 		"st1 {v12.2d}, [%[e0]] \n\t"
601 		:[op_own]"=&w"(op_own),
602 		 [byte_cnt]"=&w"(byte_cnt),
603 		 [ptype_info]"=&w"(ptype_info),
604 		 [flow_tag]"=&w"(flow_tag)
605 		:[p3]"r"(p3), [p2]"r"(p2), [p1]"r"(p1), [p0]"r"(p0),
606 		 [e3]"r"(e3), [e2]"r"(e2), [e1]"r"(e1), [e0]"r"(e0),
607 		 [c3]"w"(c3), [c2]"w"(c2), [c1]"w"(c1), [c0]"w"(c0),
608 		 [elts_p]"r"(elts_p),
609 		 [pkts_p]"r"(pkts_p),
610 		 [cqe_shuf_m]"w"(cqe_shuf_m),
611 		 [mb_shuf_m]"w"(mb_shuf_m),
612 		 [owner_shuf_m]"w"(owner_shuf_m),
613 		 [len_shuf_m]"w"(len_shuf_m),
614 		 [ptype_shuf_m]"w"(ptype_shuf_m),
615 		 [ftag_shuf_m]"w"(ftag_shuf_m),
616 		 [crc_adj]"w"(crc_adj),
617 		 [flow_mark_adj]"w"(flow_mark_adj)
618 		:"memory",
619 		 "v12", "v13", "v14", "v15",
620 		 "v16", "v17", "v18", "v19",
621 		 "v20", "v21", "v22", "v23",
622 		 "v24", "v25");
623 		/* D.2 flip owner bit to mark CQEs from last round. */
624 		owner_mask = vand_u16(op_own, owner_check);
625 		owner_mask = vceq_u16(owner_mask, ownership);
626 		/* D.3 get mask for invalidated CQEs. */
627 		opcode = vand_u16(op_own, opcode_check);
628 		invalid_mask = vceq_u16(opcode_check, opcode);
629 		/* E.1 find compressed CQE format. */
630 		comp_mask = vand_u16(op_own, format_check);
631 		comp_mask = vceq_u16(comp_mask, format_check);
632 		/* D.4 mask out beyond boundary. */
633 		invalid_mask = vorr_u16(invalid_mask, mask);
634 		/* D.5 merge invalid_mask with invalid owner. */
635 		invalid_mask = vorr_u16(invalid_mask, owner_mask);
636 		/* E.2 mask out invalid entries. */
637 		comp_mask = vbic_u16(comp_mask, invalid_mask);
638 		/* E.3 get the first compressed CQE. */
639 		comp_idx = __builtin_clzl(vget_lane_u64(vreinterpret_u64_u16(
640 					  comp_mask), 0)) /
641 					  (sizeof(uint16_t) * 8);
642 		/* D.6 mask out entries after the compressed CQE. */
643 		mask = vcreate_u16(comp_idx < MLX5_VPMD_DESCS_PER_LOOP ?
644 				   -1UL >> (comp_idx * sizeof(uint16_t) * 8) :
645 				   0);
646 		invalid_mask = vorr_u16(invalid_mask, mask);
647 		/* D.7 count non-compressed valid CQEs. */
648 		n = __builtin_clzl(vget_lane_u64(vreinterpret_u64_u16(
649 				   invalid_mask), 0)) / (sizeof(uint16_t) * 8);
650 		nocmp_n += n;
651 		/* D.2 get the final invalid mask. */
652 		mask = vcreate_u16(n < MLX5_VPMD_DESCS_PER_LOOP ?
653 				   -1UL >> (n * sizeof(uint16_t) * 8) : 0);
654 		invalid_mask = vorr_u16(invalid_mask, mask);
655 		/* D.3 check error in opcode. */
656 		opcode = vceq_u16(resp_err_check, opcode);
657 		opcode = vbic_u16(opcode, invalid_mask);
658 		/* D.4 mark if any error is set */
659 		*err |= vget_lane_u64(vreinterpret_u64_u16(opcode), 0);
660 		/* C.4 fill in mbuf - rearm_data and packet_type. */
661 		rxq_cq_to_ptype_oflags_v(rxq, ptype_info, flow_tag,
662 					 opcode, &elts[pos]);
663 		if (rxq->hw_timestamp) {
664 			int offset = rxq->timestamp_offset;
665 			if (rxq->rt_timestamp) {
666 				struct mlx5_dev_ctx_shared *sh = rxq->sh;
667 				uint64_t ts;
668 
669 				ts = rte_be_to_cpu_64
670 					(container_of(p0, struct mlx5_cqe,
671 						      pkt_info)->timestamp);
672 				mlx5_timestamp_set(elts[pos], offset,
673 					mlx5_txpp_convert_rx_ts(sh, ts));
674 				ts = rte_be_to_cpu_64
675 					(container_of(p1, struct mlx5_cqe,
676 						      pkt_info)->timestamp);
677 				mlx5_timestamp_set(elts[pos + 1], offset,
678 					mlx5_txpp_convert_rx_ts(sh, ts));
679 				ts = rte_be_to_cpu_64
680 					(container_of(p2, struct mlx5_cqe,
681 						      pkt_info)->timestamp);
682 				mlx5_timestamp_set(elts[pos + 2], offset,
683 					mlx5_txpp_convert_rx_ts(sh, ts));
684 				ts = rte_be_to_cpu_64
685 					(container_of(p3, struct mlx5_cqe,
686 						      pkt_info)->timestamp);
687 				mlx5_timestamp_set(elts[pos + 3], offset,
688 					mlx5_txpp_convert_rx_ts(sh, ts));
689 			} else {
690 				mlx5_timestamp_set(elts[pos], offset,
691 					rte_be_to_cpu_64(container_of(p0,
692 					struct mlx5_cqe, pkt_info)->timestamp));
693 				mlx5_timestamp_set(elts[pos + 1], offset,
694 					rte_be_to_cpu_64(container_of(p1,
695 					struct mlx5_cqe, pkt_info)->timestamp));
696 				mlx5_timestamp_set(elts[pos + 2], offset,
697 					rte_be_to_cpu_64(container_of(p2,
698 					struct mlx5_cqe, pkt_info)->timestamp));
699 				mlx5_timestamp_set(elts[pos + 3], offset,
700 					rte_be_to_cpu_64(container_of(p3,
701 					struct mlx5_cqe, pkt_info)->timestamp));
702 			}
703 		}
704 		if (rxq->dynf_meta) {
705 			/* This code is subject for futher optimization. */
706 			int32_t offs = rxq->flow_meta_offset;
707 
708 			*RTE_MBUF_DYNFIELD(pkts[pos], offs, uint32_t *) =
709 				container_of(p0, struct mlx5_cqe,
710 					     pkt_info)->flow_table_metadata;
711 			*RTE_MBUF_DYNFIELD(pkts[pos + 1], offs, uint32_t *) =
712 				container_of(p1, struct mlx5_cqe,
713 					     pkt_info)->flow_table_metadata;
714 			*RTE_MBUF_DYNFIELD(pkts[pos + 2], offs, uint32_t *) =
715 				container_of(p2, struct mlx5_cqe,
716 					     pkt_info)->flow_table_metadata;
717 			*RTE_MBUF_DYNFIELD(pkts[pos + 3], offs, uint32_t *) =
718 				container_of(p3, struct mlx5_cqe,
719 					     pkt_info)->flow_table_metadata;
720 			if (*RTE_MBUF_DYNFIELD(pkts[pos], offs, uint32_t *))
721 				elts[pos]->ol_flags |= rxq->flow_meta_mask;
722 			if (*RTE_MBUF_DYNFIELD(pkts[pos + 1], offs, uint32_t *))
723 				elts[pos + 1]->ol_flags |= rxq->flow_meta_mask;
724 			if (*RTE_MBUF_DYNFIELD(pkts[pos + 2], offs, uint32_t *))
725 				elts[pos + 2]->ol_flags |= rxq->flow_meta_mask;
726 			if (*RTE_MBUF_DYNFIELD(pkts[pos + 3], offs, uint32_t *))
727 				elts[pos + 3]->ol_flags |= rxq->flow_meta_mask;
728 		}
729 #ifdef MLX5_PMD_SOFT_COUNTERS
730 		/* Add up received bytes count. */
731 		byte_cnt = vbic_u16(byte_cnt, invalid_mask);
732 		rcvd_byte += vget_lane_u64(vpaddl_u32(vpaddl_u16(byte_cnt)), 0);
733 #endif
734 		/*
735 		 * Break the loop unless more valid CQE is expected, or if
736 		 * there's a compressed CQE.
737 		 */
738 		if (n != MLX5_VPMD_DESCS_PER_LOOP)
739 			break;
740 	}
741 #ifdef MLX5_PMD_SOFT_COUNTERS
742 	rxq->stats.ipackets += nocmp_n;
743 	rxq->stats.ibytes += rcvd_byte;
744 #endif
745 	if (comp_idx == n)
746 		*comp = comp_idx;
747 	return nocmp_n;
748 }
749 
750 #endif /* RTE_PMD_MLX5_RXTX_VEC_NEON_H_ */
751