xref: /dpdk/drivers/net/mlx5/mlx5_rxtx_vec_neon.h (revision a2854c4de12970dc351eaefe9fbd3c77d403caf7)
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 rxq
33  *   Pointer to RX queue structure.
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 mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t n)
41 {
42 	const uint16_t q_mask = (1 << rxq->elts_n) - 1;
43 	struct rte_mbuf **elts = &(*rxq->elts)[rxq->rq_pi & q_mask];
44 	unsigned int pos;
45 	uint16_t p = n & -2;
46 
47 	for (pos = 0; pos < p; pos += 2) {
48 		uint64x2_t mbp;
49 
50 		mbp = vld1q_u64((void *)&elts[pos]);
51 		vst1q_u64((void *)&pkts[pos], mbp);
52 	}
53 	if (n & 1)
54 		pkts[pos] = elts[pos];
55 }
56 
57 /**
58  * Decompress a compressed completion and fill in mbufs in RX SW ring with data
59  * extracted from the title completion descriptor.
60  *
61  * @param rxq
62  *   Pointer to RX queue structure.
63  * @param cq
64  *   Pointer to completion array having a compressed completion at first.
65  * @param elts
66  *   Pointer to SW ring to be filled. The first mbuf has to be pre-built from
67  *   the title completion descriptor to be copied to the rest of mbufs.
68  *
69  * @return
70  *   Number of mini-CQEs successfully decompressed.
71  */
72 static inline uint16_t
73 rxq_cq_decompress_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq,
74 		    struct rte_mbuf **elts)
75 {
76 	volatile struct mlx5_mini_cqe8 *mcq = (void *)&(cq + 1)->pkt_info;
77 	struct rte_mbuf *t_pkt = elts[0]; /* Title packet is pre-built. */
78 	unsigned int pos;
79 	unsigned int i;
80 	unsigned int inv = 0;
81 	/* Mask to shuffle from extracted mini CQE to mbuf. */
82 	const uint8x16_t mcqe_shuf_m1 = {
83 		-1, -1, -1, -1, /* skip packet_type */
84 		 7,  6, -1, -1, /* pkt_len, bswap16 */
85 		 7,  6,         /* data_len, bswap16 */
86 		-1, -1,         /* skip vlan_tci */
87 		 3,  2,  1,  0  /* hash.rss, bswap32 */
88 	};
89 	const uint8x16_t mcqe_shuf_m2 = {
90 		-1, -1, -1, -1, /* skip packet_type */
91 		15, 14, -1, -1, /* pkt_len, bswap16 */
92 		15, 14,         /* data_len, bswap16 */
93 		-1, -1,         /* skip vlan_tci */
94 		11, 10,  9,  8  /* hash.rss, bswap32 */
95 	};
96 	/* Restore the compressed count. Must be 16 bits. */
97 	const uint16_t mcqe_n = t_pkt->data_len +
98 				(rxq->crc_present * RTE_ETHER_CRC_LEN);
99 	const uint64x2_t rearm =
100 		vld1q_u64((void *)&t_pkt->rearm_data);
101 	const uint32x4_t rxdf_mask = {
102 		0xffffffff, /* packet_type */
103 		0,          /* skip pkt_len */
104 		0xffff0000, /* vlan_tci, skip data_len */
105 		0,          /* skip hash.rss */
106 	};
107 	const uint8x16_t rxdf =
108 		vandq_u8(vld1q_u8((void *)&t_pkt->rx_descriptor_fields1),
109 			 vreinterpretq_u8_u32(rxdf_mask));
110 	const uint16x8_t crc_adj = {
111 		0, 0,
112 		rxq->crc_present * RTE_ETHER_CRC_LEN, 0,
113 		rxq->crc_present * RTE_ETHER_CRC_LEN, 0,
114 		0, 0
115 	};
116 	const uint32_t flow_tag = t_pkt->hash.fdir.hi;
117 #ifdef MLX5_PMD_SOFT_COUNTERS
118 	uint32_t rcvd_byte = 0;
119 #endif
120 	/* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */
121 	const uint8x8_t len_shuf_m = {
122 		 7,  6,         /* 1st mCQE */
123 		15, 14,         /* 2nd mCQE */
124 		23, 22,         /* 3rd mCQE */
125 		31, 30          /* 4th mCQE */
126 	};
127 
128 	/*
129 	 * A. load mCQEs into a 128bit register.
130 	 * B. store rearm data to mbuf.
131 	 * C. combine data from mCQEs with rx_descriptor_fields1.
132 	 * D. store rx_descriptor_fields1.
133 	 * E. store flow tag (rte_flow mark).
134 	 */
135 	for (pos = 0; pos < mcqe_n; ) {
136 		uint8_t *p = (void *)&mcq[pos % 8];
137 		uint8_t *e0 = (void *)&elts[pos]->rearm_data;
138 		uint8_t *e1 = (void *)&elts[pos + 1]->rearm_data;
139 		uint8_t *e2 = (void *)&elts[pos + 2]->rearm_data;
140 		uint8_t *e3 = (void *)&elts[pos + 3]->rearm_data;
141 		uint16x4_t byte_cnt;
142 #ifdef MLX5_PMD_SOFT_COUNTERS
143 		uint16x4_t invalid_mask =
144 			vcreate_u16(mcqe_n - pos < MLX5_VPMD_DESCS_PER_LOOP ?
145 				    -1UL << ((mcqe_n - pos) *
146 					     sizeof(uint16_t) * 8) : 0);
147 #endif
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));
151 		__asm__ volatile (
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"
186 #endif
187 		:[byte_cnt]"=&w"(byte_cnt)
188 		:[mcq]"r"(p),
189 		 [rxdf]"w"(rxdf),
190 		 [rearm]"w"(rearm),
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);
200 #endif
201 		if (rxq->mark) {
202 			/* E.1 store flow tag (rte_flow mark). */
203 			elts[pos]->hash.fdir.hi = flow_tag;
204 			elts[pos + 1]->hash.fdir.hi = flow_tag;
205 			elts[pos + 2]->hash.fdir.hi = flow_tag;
206 			elts[pos + 3]->hash.fdir.hi = flow_tag;
207 		}
208 		if (rxq->dynf_meta) {
209 			int32_t offs = rxq->flow_meta_offset;
210 			const uint32_t meta =
211 				*RTE_MBUF_DYNFIELD(t_pkt, offs, uint32_t *);
212 
213 			/* Check if title packet has valid metadata. */
214 			if (meta) {
215 				MLX5_ASSERT(t_pkt->ol_flags &
216 					    rxq->flow_meta_mask);
217 				*RTE_MBUF_DYNFIELD(elts[pos], offs,
218 							uint32_t *) = meta;
219 				*RTE_MBUF_DYNFIELD(elts[pos + 1], offs,
220 							uint32_t *) = meta;
221 				*RTE_MBUF_DYNFIELD(elts[pos + 2], offs,
222 							uint32_t *) = meta;
223 				*RTE_MBUF_DYNFIELD(elts[pos + 3], offs,
224 							uint32_t *) = meta;
225 			}
226 		}
227 		pos += MLX5_VPMD_DESCS_PER_LOOP;
228 		/* Move to next CQE and invalidate consumed CQEs. */
229 		if (!(pos & 0x7) && pos < mcqe_n) {
230 			mcq = (void *)&(cq + pos)->pkt_info;
231 			for (i = 0; i < 8; ++i)
232 				cq[inv++].op_own = MLX5_CQE_INVALIDATE;
233 		}
234 	}
235 	/* Invalidate the rest of CQEs. */
236 	for (; inv < mcqe_n; ++inv)
237 		cq[inv].op_own = MLX5_CQE_INVALIDATE;
238 #ifdef MLX5_PMD_SOFT_COUNTERS
239 	rxq->stats.ipackets += mcqe_n;
240 	rxq->stats.ibytes += rcvd_byte;
241 #endif
242 	rxq->cq_ci += mcqe_n;
243 	return mcqe_n;
244 }
245 
246 /**
247  * Calculate packet type and offload flag for mbuf and store it.
248  *
249  * @param rxq
250  *   Pointer to RX queue structure.
251  * @param ptype_info
252  *   Array of four 4bytes packet type info extracted from the original
253  *   completion descriptor.
254  * @param flow_tag
255  *   Array of four 4bytes flow ID extracted from the original completion
256  *   descriptor.
257  * @param op_err
258  *   Opcode vector having responder error status. Each field is 4B.
259  * @param pkts
260  *   Pointer to array of packets to be filled.
261  */
262 static inline void
263 rxq_cq_to_ptype_oflags_v(struct mlx5_rxq_data *rxq,
264 			 uint32x4_t ptype_info, uint32x4_t flow_tag,
265 			 uint16x4_t op_err, struct rte_mbuf **pkts)
266 {
267 	uint16x4_t ptype;
268 	uint32x4_t pinfo, cv_flags;
269 	uint32x4_t ol_flags =
270 		vdupq_n_u32(rxq->rss_hash * PKT_RX_RSS_HASH |
271 			    rxq->hw_timestamp * PKT_RX_TIMESTAMP);
272 	const uint32x4_t ptype_ol_mask = { 0x106, 0x106, 0x106, 0x106 };
273 	const uint8x16_t cv_flag_sel = {
274 		0,
275 		(uint8_t)(PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED),
276 		(uint8_t)(PKT_RX_IP_CKSUM_GOOD >> 1),
277 		0,
278 		(uint8_t)(PKT_RX_L4_CKSUM_GOOD >> 1),
279 		0,
280 		(uint8_t)((PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1),
281 		0, 0, 0, 0, 0, 0, 0, 0, 0
282 	};
283 	const uint32x4_t cv_mask =
284 		vdupq_n_u32(PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
285 			    PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED);
286 	const uint64x2_t mbuf_init = vld1q_u64
287 				((const uint64_t *)&rxq->mbuf_initializer);
288 	uint64x2_t rearm0, rearm1, rearm2, rearm3;
289 	uint8_t pt_idx0, pt_idx1, pt_idx2, pt_idx3;
290 
291 	if (rxq->mark) {
292 		const uint32x4_t ft_def = vdupq_n_u32(MLX5_FLOW_MARK_DEFAULT);
293 		const uint32x4_t fdir_flags = vdupq_n_u32(PKT_RX_FDIR);
294 		uint32x4_t fdir_id_flags = vdupq_n_u32(PKT_RX_FDIR_ID);
295 		uint32x4_t invalid_mask;
296 
297 		/* Check if flow tag is non-zero then set PKT_RX_FDIR. */
298 		invalid_mask = vceqzq_u32(flow_tag);
299 		ol_flags = vorrq_u32(ol_flags,
300 				     vbicq_u32(fdir_flags, invalid_mask));
301 		/* Mask out invalid entries. */
302 		fdir_id_flags = vbicq_u32(fdir_id_flags, invalid_mask);
303 		/* Check if flow tag MLX5_FLOW_MARK_DEFAULT. */
304 		ol_flags = vorrq_u32(ol_flags,
305 				     vbicq_u32(fdir_id_flags,
306 					       vceqq_u32(flow_tag, ft_def)));
307 	}
308 	/*
309 	 * ptype_info has the following:
310 	 * bit[1]     = l3_ok
311 	 * bit[2]     = l4_ok
312 	 * bit[8]     = cv
313 	 * bit[11:10] = l3_hdr_type
314 	 * bit[14:12] = l4_hdr_type
315 	 * bit[15]    = ip_frag
316 	 * bit[16]    = tunneled
317 	 * bit[17]    = outer_l3_type
318 	 */
319 	ptype = vshrn_n_u32(ptype_info, 10);
320 	/* Errored packets will have RTE_PTYPE_ALL_MASK. */
321 	ptype = vorr_u16(ptype, op_err);
322 	pt_idx0 = vget_lane_u8(vreinterpret_u8_u16(ptype), 6);
323 	pt_idx1 = vget_lane_u8(vreinterpret_u8_u16(ptype), 4);
324 	pt_idx2 = vget_lane_u8(vreinterpret_u8_u16(ptype), 2);
325 	pt_idx3 = vget_lane_u8(vreinterpret_u8_u16(ptype), 0);
326 	pkts[0]->packet_type = mlx5_ptype_table[pt_idx0] |
327 			       !!(pt_idx0 & (1 << 6)) * rxq->tunnel;
328 	pkts[1]->packet_type = mlx5_ptype_table[pt_idx1] |
329 			       !!(pt_idx1 & (1 << 6)) * rxq->tunnel;
330 	pkts[2]->packet_type = mlx5_ptype_table[pt_idx2] |
331 			       !!(pt_idx2 & (1 << 6)) * rxq->tunnel;
332 	pkts[3]->packet_type = mlx5_ptype_table[pt_idx3] |
333 			       !!(pt_idx3 & (1 << 6)) * rxq->tunnel;
334 	/* Fill flags for checksum and VLAN. */
335 	pinfo = vandq_u32(ptype_info, ptype_ol_mask);
336 	pinfo = vreinterpretq_u32_u8(
337 		vqtbl1q_u8(cv_flag_sel, vreinterpretq_u8_u32(pinfo)));
338 	/* Locate checksum flags at byte[2:1] and merge with VLAN flags. */
339 	cv_flags = vshlq_n_u32(pinfo, 9);
340 	cv_flags = vorrq_u32(pinfo, cv_flags);
341 	/* Move back flags to start from byte[0]. */
342 	cv_flags = vshrq_n_u32(cv_flags, 8);
343 	/* Mask out garbage bits. */
344 	cv_flags = vandq_u32(cv_flags, cv_mask);
345 	/* Merge to ol_flags. */
346 	ol_flags = vorrq_u32(ol_flags, cv_flags);
347 	/* Merge mbuf_init and ol_flags, and store. */
348 	rearm0 = vreinterpretq_u64_u32(vsetq_lane_u32
349 					(vgetq_lane_u32(ol_flags, 3),
350 					 vreinterpretq_u32_u64(mbuf_init), 2));
351 	rearm1 = vreinterpretq_u64_u32(vsetq_lane_u32
352 					(vgetq_lane_u32(ol_flags, 2),
353 					 vreinterpretq_u32_u64(mbuf_init), 2));
354 	rearm2 = vreinterpretq_u64_u32(vsetq_lane_u32
355 					(vgetq_lane_u32(ol_flags, 1),
356 					 vreinterpretq_u32_u64(mbuf_init), 2));
357 	rearm3 = vreinterpretq_u64_u32(vsetq_lane_u32
358 					(vgetq_lane_u32(ol_flags, 0),
359 					 vreinterpretq_u32_u64(mbuf_init), 2));
360 
361 	vst1q_u64((void *)&pkts[0]->rearm_data, rearm0);
362 	vst1q_u64((void *)&pkts[1]->rearm_data, rearm1);
363 	vst1q_u64((void *)&pkts[2]->rearm_data, rearm2);
364 	vst1q_u64((void *)&pkts[3]->rearm_data, rearm3);
365 }
366 
367 /**
368  * Receive burst of packets. An errored completion also consumes a mbuf, but the
369  * packet_type is set to be RTE_PTYPE_ALL_MASK. Marked mbufs should be freed
370  * before returning to application.
371  *
372  * @param rxq
373  *   Pointer to RX queue structure.
374  * @param[out] pkts
375  *   Array to store received packets.
376  * @param pkts_n
377  *   Maximum number of packets in array.
378  * @param[out] err
379  *   Pointer to a flag. Set non-zero value if pkts array has at least one error
380  *   packet to handle.
381  * @param[out] no_cq
382  *   Pointer to a boolean. Set true if no new CQE seen.
383  *
384  * @return
385  *   Number of packets received including errors (<= pkts_n).
386  */
387 static inline uint16_t
388 rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
389 	    uint64_t *err, bool *no_cq)
390 {
391 	const uint16_t q_n = 1 << rxq->cqe_n;
392 	const uint16_t q_mask = q_n - 1;
393 	volatile struct mlx5_cqe *cq;
394 	struct rte_mbuf **elts;
395 	unsigned int pos;
396 	uint64_t n;
397 	uint16_t repl_n;
398 	uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
399 	uint16_t nocmp_n = 0;
400 	uint16_t rcvd_pkt = 0;
401 	unsigned int cq_idx = rxq->cq_ci & q_mask;
402 	unsigned int elts_idx;
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 	MLX5_ASSERT(rxq->sges_n == 0);
464 	MLX5_ASSERT(rxq->cqe_n == rxq->elts_n);
465 	cq = &(*rxq->cqes)[cq_idx];
466 	rte_prefetch_non_temporal(cq);
467 	rte_prefetch_non_temporal(cq + 1);
468 	rte_prefetch_non_temporal(cq + 2);
469 	rte_prefetch_non_temporal(cq + 3);
470 	pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
471 	repl_n = q_n - (rxq->rq_ci - rxq->rq_pi);
472 	if (repl_n >= rxq->rq_repl_thresh)
473 		mlx5_rx_replenish_bulk_mbuf(rxq, repl_n);
474 	/* See if there're unreturned mbufs from compressed CQE. */
475 	rcvd_pkt = rxq->decompressed;
476 	if (rcvd_pkt > 0) {
477 		rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
478 		rxq_copy_mbuf_v(rxq, pkts, rcvd_pkt);
479 		rxq->rq_pi += rcvd_pkt;
480 		pkts += rcvd_pkt;
481 		rxq->decompressed -= rcvd_pkt;
482 	}
483 	elts_idx = rxq->rq_pi & q_mask;
484 	elts = &(*rxq->elts)[elts_idx];
485 	/* Not to overflow pkts array. */
486 	pkts_n = RTE_ALIGN_FLOOR(pkts_n - rcvd_pkt, MLX5_VPMD_DESCS_PER_LOOP);
487 	/* Not to cross queue end. */
488 	pkts_n = RTE_MIN(pkts_n, q_n - elts_idx);
489 	pkts_n = RTE_MIN(pkts_n, q_n - cq_idx);
490 	if (!pkts_n) {
491 		*no_cq = !rcvd_pkt;
492 		return rcvd_pkt;
493 	}
494 	/* At this point, there shouldn't be any remained packets. */
495 	MLX5_ASSERT(rxq->decompressed == 0);
496 	/*
497 	 * Note that vectors have reverse order - {v3, v2, v1, v0}, because
498 	 * there's no instruction to count trailing zeros. __builtin_clzl() is
499 	 * used instead.
500 	 *
501 	 * A. copy 4 mbuf pointers from elts ring to returing pkts.
502 	 * B. load 64B CQE and extract necessary fields
503 	 *    Final 16bytes cqes[] extracted from original 64bytes CQE has the
504 	 *    following structure:
505 	 *        struct {
506 	 *          uint16_t hdr_type_etc;
507 	 *          uint8_t  pkt_info;
508 	 *          uint8_t  rsvd;
509 	 *          uint16_t byte_cnt;
510 	 *          uint16_t vlan_info;
511 	 *          uint32_t rx_has_res;
512 	 *          uint8_t  flow_tag[3];
513 	 *          uint8_t  op_own;
514 	 *        } c;
515 	 * C. fill in mbuf.
516 	 * D. get valid CQEs.
517 	 * E. find compressed CQE.
518 	 */
519 	for (pos = 0;
520 	     pos < pkts_n;
521 	     pos += MLX5_VPMD_DESCS_PER_LOOP) {
522 		uint16x4_t op_own;
523 		uint16x4_t opcode, owner_mask, invalid_mask;
524 		uint16x4_t comp_mask;
525 		uint16x4_t mask;
526 		uint16x4_t byte_cnt;
527 		uint32x4_t ptype_info, flow_tag;
528 		register uint64x2_t c0, c1, c2, c3;
529 		uint8_t *p0, *p1, *p2, *p3;
530 		uint8_t *e0 = (void *)&elts[pos]->pkt_len;
531 		uint8_t *e1 = (void *)&elts[pos + 1]->pkt_len;
532 		uint8_t *e2 = (void *)&elts[pos + 2]->pkt_len;
533 		uint8_t *e3 = (void *)&elts[pos + 3]->pkt_len;
534 		void *elts_p = (void *)&elts[pos];
535 		void *pkts_p = (void *)&pkts[pos];
536 
537 		/* A.0 do not cross the end of CQ. */
538 		mask = vcreate_u16(pkts_n - pos < MLX5_VPMD_DESCS_PER_LOOP ?
539 				   -1UL >> ((pkts_n - pos) *
540 					    sizeof(uint16_t) * 8) : 0);
541 		p0 = (void *)&cq[pos].pkt_info;
542 		p1 = p0 + (pkts_n - pos > 1) * sizeof(struct mlx5_cqe);
543 		p2 = p1 + (pkts_n - pos > 2) * sizeof(struct mlx5_cqe);
544 		p3 = p2 + (pkts_n - pos > 3) * sizeof(struct mlx5_cqe);
545 		/* B.0 (CQE 3) load a block having op_own. */
546 		c3 = vld1q_u64((uint64_t *)(p3 + 48));
547 		/* B.0 (CQE 2) load a block having op_own. */
548 		c2 = vld1q_u64((uint64_t *)(p2 + 48));
549 		/* B.0 (CQE 1) load a block having op_own. */
550 		c1 = vld1q_u64((uint64_t *)(p1 + 48));
551 		/* B.0 (CQE 0) load a block having op_own. */
552 		c0 = vld1q_u64((uint64_t *)(p0 + 48));
553 		/* Synchronize for loading the rest of blocks. */
554 		rte_cio_rmb();
555 		/* Prefetch next 4 CQEs. */
556 		if (pkts_n - pos >= 2 * MLX5_VPMD_DESCS_PER_LOOP) {
557 			unsigned int next = pos + MLX5_VPMD_DESCS_PER_LOOP;
558 			rte_prefetch_non_temporal(&cq[next]);
559 			rte_prefetch_non_temporal(&cq[next + 1]);
560 			rte_prefetch_non_temporal(&cq[next + 2]);
561 			rte_prefetch_non_temporal(&cq[next + 3]);
562 		}
563 		__asm__ volatile (
564 		/* B.1 (CQE 3) load the rest of blocks. */
565 		"ld1 {v16.16b - v18.16b}, [%[p3]] \n\t"
566 		/* B.2 (CQE 3) move the block having op_own. */
567 		"mov v19.16b, %[c3].16b \n\t"
568 		/* B.3 (CQE 3) extract 16B fields. */
569 		"tbl v23.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
570 		/* B.1 (CQE 2) load the rest of blocks. */
571 		"ld1 {v16.16b - v18.16b}, [%[p2]] \n\t"
572 		/* B.4 (CQE 3) adjust CRC length. */
573 		"sub v23.8h, v23.8h, %[crc_adj].8h \n\t"
574 		/* C.1 (CQE 3) generate final structure for mbuf. */
575 		"tbl v15.16b, {v23.16b}, %[mb_shuf_m].16b \n\t"
576 		/* B.2 (CQE 2) move the block having op_own. */
577 		"mov v19.16b, %[c2].16b \n\t"
578 		/* B.3 (CQE 2) extract 16B fields. */
579 		"tbl v22.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
580 		/* B.1 (CQE 1) load the rest of blocks. */
581 		"ld1 {v16.16b - v18.16b}, [%[p1]] \n\t"
582 		/* B.4 (CQE 2) adjust CRC length. */
583 		"sub v22.8h, v22.8h, %[crc_adj].8h \n\t"
584 		/* C.1 (CQE 2) generate final structure for mbuf. */
585 		"tbl v14.16b, {v22.16b}, %[mb_shuf_m].16b \n\t"
586 		/* B.2 (CQE 1) move the block having op_own. */
587 		"mov v19.16b, %[c1].16b \n\t"
588 		/* B.3 (CQE 1) extract 16B fields. */
589 		"tbl v21.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
590 		/* B.1 (CQE 0) load the rest of blocks. */
591 		"ld1 {v16.16b - v18.16b}, [%[p0]] \n\t"
592 		/* B.4 (CQE 1) adjust CRC length. */
593 		"sub v21.8h, v21.8h, %[crc_adj].8h \n\t"
594 		/* C.1 (CQE 1) generate final structure for mbuf. */
595 		"tbl v13.16b, {v21.16b}, %[mb_shuf_m].16b \n\t"
596 		/* B.2 (CQE 0) move the block having op_own. */
597 		"mov v19.16b, %[c0].16b \n\t"
598 		/* A.1 load mbuf pointers. */
599 		"ld1 {v24.2d - v25.2d}, [%[elts_p]] \n\t"
600 		/* B.3 (CQE 0) extract 16B fields. */
601 		"tbl v20.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
602 		/* B.4 (CQE 0) adjust CRC length. */
603 		"sub v20.8h, v20.8h, %[crc_adj].8h \n\t"
604 		/* D.1 extract op_own byte. */
605 		"tbl %[op_own].8b, {v20.16b - v23.16b}, %[owner_shuf_m].8b \n\t"
606 		/* C.2 (CQE 3) adjust flow mark. */
607 		"add v15.4s, v15.4s, %[flow_mark_adj].4s \n\t"
608 		/* C.3 (CQE 3) fill in mbuf - rx_descriptor_fields1. */
609 		"st1 {v15.2d}, [%[e3]] \n\t"
610 		/* C.2 (CQE 2) adjust flow mark. */
611 		"add v14.4s, v14.4s, %[flow_mark_adj].4s \n\t"
612 		/* C.3 (CQE 2) fill in mbuf - rx_descriptor_fields1. */
613 		"st1 {v14.2d}, [%[e2]] \n\t"
614 		/* C.1 (CQE 0) generate final structure for mbuf. */
615 		"tbl v12.16b, {v20.16b}, %[mb_shuf_m].16b \n\t"
616 		/* C.2 (CQE 1) adjust flow mark. */
617 		"add v13.4s, v13.4s, %[flow_mark_adj].4s \n\t"
618 		/* C.3 (CQE 1) fill in mbuf - rx_descriptor_fields1. */
619 		"st1 {v13.2d}, [%[e1]] \n\t"
620 #ifdef MLX5_PMD_SOFT_COUNTERS
621 		/* Extract byte_cnt. */
622 		"tbl %[byte_cnt].8b, {v20.16b - v23.16b}, %[len_shuf_m].8b \n\t"
623 #endif
624 		/* Extract ptype_info. */
625 		"tbl %[ptype_info].16b, {v20.16b - v23.16b}, %[ptype_shuf_m].16b \n\t"
626 		/* Extract flow_tag. */
627 		"tbl %[flow_tag].16b, {v20.16b - v23.16b}, %[ftag_shuf_m].16b \n\t"
628 		/* A.2 copy mbuf pointers. */
629 		"st1 {v24.2d - v25.2d}, [%[pkts_p]] \n\t"
630 		/* C.2 (CQE 0) adjust flow mark. */
631 		"add v12.4s, v12.4s, %[flow_mark_adj].4s \n\t"
632 		/* C.3 (CQE 1) fill in mbuf - rx_descriptor_fields1. */
633 		"st1 {v12.2d}, [%[e0]] \n\t"
634 		:[op_own]"=&w"(op_own),
635 		 [byte_cnt]"=&w"(byte_cnt),
636 		 [ptype_info]"=&w"(ptype_info),
637 		 [flow_tag]"=&w"(flow_tag)
638 		:[p3]"r"(p3), [p2]"r"(p2), [p1]"r"(p1), [p0]"r"(p0),
639 		 [e3]"r"(e3), [e2]"r"(e2), [e1]"r"(e1), [e0]"r"(e0),
640 		 [c3]"w"(c3), [c2]"w"(c2), [c1]"w"(c1), [c0]"w"(c0),
641 		 [elts_p]"r"(elts_p),
642 		 [pkts_p]"r"(pkts_p),
643 		 [cqe_shuf_m]"w"(cqe_shuf_m),
644 		 [mb_shuf_m]"w"(mb_shuf_m),
645 		 [owner_shuf_m]"w"(owner_shuf_m),
646 		 [len_shuf_m]"w"(len_shuf_m),
647 		 [ptype_shuf_m]"w"(ptype_shuf_m),
648 		 [ftag_shuf_m]"w"(ftag_shuf_m),
649 		 [crc_adj]"w"(crc_adj),
650 		 [flow_mark_adj]"w"(flow_mark_adj)
651 		:"memory",
652 		 "v12", "v13", "v14", "v15",
653 		 "v16", "v17", "v18", "v19",
654 		 "v20", "v21", "v22", "v23",
655 		 "v24", "v25");
656 		/* D.2 flip owner bit to mark CQEs from last round. */
657 		owner_mask = vand_u16(op_own, owner_check);
658 		owner_mask = vceq_u16(owner_mask, ownership);
659 		/* D.3 get mask for invalidated CQEs. */
660 		opcode = vand_u16(op_own, opcode_check);
661 		invalid_mask = vceq_u16(opcode_check, opcode);
662 		/* E.1 find compressed CQE format. */
663 		comp_mask = vand_u16(op_own, format_check);
664 		comp_mask = vceq_u16(comp_mask, format_check);
665 		/* D.4 mask out beyond boundary. */
666 		invalid_mask = vorr_u16(invalid_mask, mask);
667 		/* D.5 merge invalid_mask with invalid owner. */
668 		invalid_mask = vorr_u16(invalid_mask, owner_mask);
669 		/* E.2 mask out invalid entries. */
670 		comp_mask = vbic_u16(comp_mask, invalid_mask);
671 		/* E.3 get the first compressed CQE. */
672 		comp_idx = __builtin_clzl(vget_lane_u64(vreinterpret_u64_u16(
673 					  comp_mask), 0)) /
674 					  (sizeof(uint16_t) * 8);
675 		/* D.6 mask out entries after the compressed CQE. */
676 		mask = vcreate_u16(comp_idx < MLX5_VPMD_DESCS_PER_LOOP ?
677 				   -1UL >> (comp_idx * sizeof(uint16_t) * 8) :
678 				   0);
679 		invalid_mask = vorr_u16(invalid_mask, mask);
680 		/* D.7 count non-compressed valid CQEs. */
681 		n = __builtin_clzl(vget_lane_u64(vreinterpret_u64_u16(
682 				   invalid_mask), 0)) / (sizeof(uint16_t) * 8);
683 		nocmp_n += n;
684 		/* D.2 get the final invalid mask. */
685 		mask = vcreate_u16(n < MLX5_VPMD_DESCS_PER_LOOP ?
686 				   -1UL >> (n * sizeof(uint16_t) * 8) : 0);
687 		invalid_mask = vorr_u16(invalid_mask, mask);
688 		/* D.3 check error in opcode. */
689 		opcode = vceq_u16(resp_err_check, opcode);
690 		opcode = vbic_u16(opcode, invalid_mask);
691 		/* D.4 mark if any error is set */
692 		*err |= vget_lane_u64(vreinterpret_u64_u16(opcode), 0);
693 		/* C.4 fill in mbuf - rearm_data and packet_type. */
694 		rxq_cq_to_ptype_oflags_v(rxq, ptype_info, flow_tag,
695 					 opcode, &elts[pos]);
696 		if (rxq->hw_timestamp) {
697 			if (rxq->rt_timestamp) {
698 				struct mlx5_dev_ctx_shared *sh = rxq->sh;
699 				uint64_t ts;
700 
701 				ts = rte_be_to_cpu_64
702 					(container_of(p0, struct mlx5_cqe,
703 						      pkt_info)->timestamp);
704 				elts[pos]->timestamp =
705 					mlx5_txpp_convert_rx_ts(sh, ts);
706 				ts = rte_be_to_cpu_64
707 					(container_of(p1, struct mlx5_cqe,
708 						      pkt_info)->timestamp);
709 				elts[pos + 1]->timestamp =
710 					mlx5_txpp_convert_rx_ts(sh, ts);
711 				ts = rte_be_to_cpu_64
712 					(container_of(p2, struct mlx5_cqe,
713 						      pkt_info)->timestamp);
714 				elts[pos + 2]->timestamp =
715 					mlx5_txpp_convert_rx_ts(sh, ts);
716 				ts = rte_be_to_cpu_64
717 					(container_of(p3, struct mlx5_cqe,
718 						      pkt_info)->timestamp);
719 				elts[pos + 3]->timestamp =
720 					mlx5_txpp_convert_rx_ts(sh, ts);
721 			} else {
722 				elts[pos]->timestamp = rte_be_to_cpu_64
723 					(container_of(p0, struct mlx5_cqe,
724 						      pkt_info)->timestamp);
725 				elts[pos + 1]->timestamp = rte_be_to_cpu_64
726 					(container_of(p1, struct mlx5_cqe,
727 						      pkt_info)->timestamp);
728 				elts[pos + 2]->timestamp = rte_be_to_cpu_64
729 					(container_of(p2, struct mlx5_cqe,
730 						      pkt_info)->timestamp);
731 				elts[pos + 3]->timestamp = rte_be_to_cpu_64
732 					(container_of(p3, struct mlx5_cqe,
733 						      pkt_info)->timestamp);
734 			}
735 		}
736 		if (!!rxq->flow_meta_mask) {
737 			/* This code is subject for futher optimization. */
738 			int32_t offs = rxq->flow_meta_offset;
739 
740 			*RTE_MBUF_DYNFIELD(pkts[pos], offs, uint32_t *) =
741 				container_of(p0, struct mlx5_cqe,
742 					     pkt_info)->flow_table_metadata;
743 			*RTE_MBUF_DYNFIELD(pkts[pos], offs, uint32_t *) =
744 				container_of(p1, struct mlx5_cqe,
745 					     pkt_info)->flow_table_metadata;
746 			*RTE_MBUF_DYNFIELD(pkts[pos], offs, uint32_t *) =
747 				container_of(p2, struct mlx5_cqe,
748 					     pkt_info)->flow_table_metadata;
749 			*RTE_MBUF_DYNFIELD(pkts[pos], offs, uint32_t *) =
750 				container_of(p3, struct mlx5_cqe,
751 					     pkt_info)->flow_table_metadata;
752 			if (*RTE_MBUF_DYNFIELD(pkts[pos], offs, uint32_t *))
753 				elts[pos]->ol_flags |= rxq->flow_meta_mask;
754 			if (*RTE_MBUF_DYNFIELD(pkts[pos + 1], offs, uint32_t *))
755 				elts[pos + 1]->ol_flags |= rxq->flow_meta_mask;
756 			if (*RTE_MBUF_DYNFIELD(pkts[pos + 2], offs, uint32_t *))
757 				elts[pos + 2]->ol_flags |= rxq->flow_meta_mask;
758 			if (*RTE_MBUF_DYNFIELD(pkts[pos + 3], offs, uint32_t *))
759 				elts[pos + 3]->ol_flags |= rxq->flow_meta_mask;
760 		}
761 #ifdef MLX5_PMD_SOFT_COUNTERS
762 		/* Add up received bytes count. */
763 		byte_cnt = vbic_u16(byte_cnt, invalid_mask);
764 		rcvd_byte += vget_lane_u64(vpaddl_u32(vpaddl_u16(byte_cnt)), 0);
765 #endif
766 		/*
767 		 * Break the loop unless more valid CQE is expected, or if
768 		 * there's a compressed CQE.
769 		 */
770 		if (n != MLX5_VPMD_DESCS_PER_LOOP)
771 			break;
772 	}
773 	/* If no new CQE seen, return without updating cq_db. */
774 	if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP)) {
775 		*no_cq = true;
776 		return rcvd_pkt;
777 	}
778 	/* Update the consumer indexes for non-compressed CQEs. */
779 	MLX5_ASSERT(nocmp_n <= pkts_n);
780 	rxq->cq_ci += nocmp_n;
781 	rxq->rq_pi += nocmp_n;
782 	rcvd_pkt += nocmp_n;
783 #ifdef MLX5_PMD_SOFT_COUNTERS
784 	rxq->stats.ipackets += nocmp_n;
785 	rxq->stats.ibytes += rcvd_byte;
786 #endif
787 	/* Decompress the last CQE if compressed. */
788 	if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP && comp_idx == n) {
789 		MLX5_ASSERT(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
790 		rxq->decompressed = rxq_cq_decompress_v(rxq, &cq[nocmp_n],
791 							&elts[nocmp_n]);
792 		/* Return more packets if needed. */
793 		if (nocmp_n < pkts_n) {
794 			uint16_t n = rxq->decompressed;
795 
796 			n = RTE_MIN(n, pkts_n - nocmp_n);
797 			rxq_copy_mbuf_v(rxq, &pkts[nocmp_n], n);
798 			rxq->rq_pi += n;
799 			rcvd_pkt += n;
800 			rxq->decompressed -= n;
801 		}
802 	}
803 	rte_cio_wmb();
804 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
805 	*no_cq = !rcvd_pkt;
806 	return rcvd_pkt;
807 }
808 
809 #endif /* RTE_PMD_MLX5_RXTX_VEC_NEON_H_ */
810