xref: /dpdk/drivers/net/mlx5/mlx5_rxtx_vec_neon.h (revision 14ad4f01845331a0ae98c681efa3086eeed3343a)
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 <assert.h>
10 #include <stdint.h>
11 #include <string.h>
12 #include <stdlib.h>
13 #include <arm_neon.h>
14 
15 #include <rte_mbuf.h>
16 #include <rte_mempool.h>
17 #include <rte_prefetch.h>
18 
19 #include "mlx5.h"
20 #include "mlx5_utils.h"
21 #include "mlx5_rxtx.h"
22 #include "mlx5_rxtx_vec.h"
23 #include "mlx5_autoconf.h"
24 #include "mlx5_defs.h"
25 #include "mlx5_prm.h"
26 
27 #pragma GCC diagnostic ignored "-Wcast-qual"
28 
29 /**
30  * Fill in buffer descriptors in a multi-packet send descriptor.
31  *
32  * @param txq
33  *   Pointer to TX queue structure.
34  * @param dseg
35  *   Pointer to buffer descriptor to be written.
36  * @param pkts
37  *   Pointer to array of packets to be sent.
38  * @param n
39  *   Number of packets to be filled.
40  */
41 static inline void
42 txq_wr_dseg_v(struct mlx5_txq_data *txq, uint8_t *dseg,
43 	      struct rte_mbuf **pkts, unsigned int n)
44 {
45 	unsigned int pos;
46 	uintptr_t addr;
47 	const uint8x16_t dseg_shuf_m = {
48 		 3,  2,  1,  0, /* length, bswap32 */
49 		 4,  5,  6,  7, /* lkey */
50 		15, 14, 13, 12, /* addr, bswap64 */
51 		11, 10,  9,  8
52 	};
53 #ifdef MLX5_PMD_SOFT_COUNTERS
54 	uint32_t tx_byte = 0;
55 #endif
56 
57 	for (pos = 0; pos < n; ++pos, dseg += MLX5_WQE_DWORD_SIZE) {
58 		uint8x16_t desc;
59 		struct rte_mbuf *pkt = pkts[pos];
60 
61 		addr = rte_pktmbuf_mtod(pkt, uintptr_t);
62 		desc = vreinterpretq_u8_u32((uint32x4_t) {
63 				DATA_LEN(pkt),
64 				mlx5_tx_mb2mr(txq, pkt),
65 				addr,
66 				addr >> 32 });
67 		desc = vqtbl1q_u8(desc, dseg_shuf_m);
68 		vst1q_u8(dseg, desc);
69 #ifdef MLX5_PMD_SOFT_COUNTERS
70 		tx_byte += DATA_LEN(pkt);
71 #endif
72 	}
73 #ifdef MLX5_PMD_SOFT_COUNTERS
74 	txq->stats.obytes += tx_byte;
75 #endif
76 }
77 
78 /**
79  * Send multi-segmented packets until it encounters a single segment packet in
80  * the pkts list.
81  *
82  * @param txq
83  *   Pointer to TX queue structure.
84  * @param pkts
85  *   Pointer to array of packets to be sent.
86  * @param pkts_n
87  *   Number of packets to be sent.
88  *
89  * @return
90  *   Number of packets successfully transmitted (<= pkts_n).
91  */
92 static uint16_t
93 txq_scatter_v(struct mlx5_txq_data *txq, struct rte_mbuf **pkts,
94 	      uint16_t pkts_n)
95 {
96 	uint16_t elts_head = txq->elts_head;
97 	const uint16_t elts_n = 1 << txq->elts_n;
98 	const uint16_t elts_m = elts_n - 1;
99 	const uint16_t wq_n = 1 << txq->wqe_n;
100 	const uint16_t wq_mask = wq_n - 1;
101 	const unsigned int nb_dword_per_wqebb =
102 		MLX5_WQE_SIZE / MLX5_WQE_DWORD_SIZE;
103 	const unsigned int nb_dword_in_hdr =
104 		sizeof(struct mlx5_wqe) / MLX5_WQE_DWORD_SIZE;
105 	unsigned int n;
106 	volatile struct mlx5_wqe *wqe = NULL;
107 	bool metadata_ol =
108 		txq->offloads & DEV_TX_OFFLOAD_MATCH_METADATA ? true : false;
109 
110 	assert(elts_n > pkts_n);
111 	mlx5_tx_complete(txq);
112 	if (unlikely(!pkts_n))
113 		return 0;
114 	for (n = 0; n < pkts_n; ++n) {
115 		struct rte_mbuf *buf = pkts[n];
116 		unsigned int segs_n = buf->nb_segs;
117 		unsigned int ds = nb_dword_in_hdr;
118 		unsigned int len = PKT_LEN(buf);
119 		uint16_t wqe_ci = txq->wqe_ci;
120 		const uint8x16_t ctrl_shuf_m = {
121 			3,  2,  1,  0, /* bswap32 */
122 			7,  6,  5,  4, /* bswap32 */
123 			11, 10,  9,  8, /* bswap32 */
124 			12, 13, 14, 15
125 		};
126 		uint8_t cs_flags;
127 		uint16_t max_elts;
128 		uint16_t max_wqe;
129 		uint8x16_t *t_wqe;
130 		uint8_t *dseg;
131 		uint8x16_t ctrl;
132 		rte_be32_t metadata =
133 			metadata_ol && (buf->ol_flags & PKT_TX_METADATA) ?
134 			buf->tx_metadata : 0;
135 
136 		assert(segs_n);
137 		max_elts = elts_n - (elts_head - txq->elts_tail);
138 		max_wqe = wq_n - (txq->wqe_ci - txq->wqe_pi);
139 		/*
140 		 * A MPW session consumes 2 WQEs at most to
141 		 * include MLX5_MPW_DSEG_MAX pointers.
142 		 */
143 		if (segs_n == 1 ||
144 		    max_elts < segs_n || max_wqe < 2)
145 			break;
146 		wqe = &((volatile struct mlx5_wqe64 *)
147 			 txq->wqes)[wqe_ci & wq_mask].hdr;
148 		cs_flags = txq_ol_cksum_to_cs(buf);
149 		/* Title WQEBB pointer. */
150 		t_wqe = (uint8x16_t *)wqe;
151 		dseg = (uint8_t *)(wqe + 1);
152 		do {
153 			if (!(ds++ % nb_dword_per_wqebb)) {
154 				dseg = (uint8_t *)
155 					&((volatile struct mlx5_wqe64 *)
156 					   txq->wqes)[++wqe_ci & wq_mask];
157 			}
158 			txq_wr_dseg_v(txq, dseg, &buf, 1);
159 			dseg += MLX5_WQE_DWORD_SIZE;
160 			(*txq->elts)[elts_head++ & elts_m] = buf;
161 			buf = buf->next;
162 		} while (--segs_n);
163 		++wqe_ci;
164 		/* Fill CTRL in the header. */
165 		ctrl = vreinterpretq_u8_u32((uint32x4_t) {
166 				MLX5_OPC_MOD_MPW << 24 |
167 				txq->wqe_ci << 8 | MLX5_OPCODE_TSO,
168 				txq->qp_num_8s | ds, 4, 0});
169 		ctrl = vqtbl1q_u8(ctrl, ctrl_shuf_m);
170 		vst1q_u8((void *)t_wqe, ctrl);
171 		/* Fill ESEG in the header. */
172 		vst1q_u32((void *)(t_wqe + 1),
173 			  ((uint32x4_t){ 0,
174 					 rte_cpu_to_be_16(len) << 16 | cs_flags,
175 					 metadata, 0 }));
176 		txq->wqe_ci = wqe_ci;
177 	}
178 	if (!n)
179 		return 0;
180 	txq->elts_comp += (uint16_t)(elts_head - txq->elts_head);
181 	txq->elts_head = elts_head;
182 	if (txq->elts_comp >= MLX5_TX_COMP_THRESH) {
183 		/* A CQE slot must always be available. */
184 		assert((1u << txq->cqe_n) - (txq->cq_pi++ - txq->cq_ci));
185 		wqe->ctrl[2] = rte_cpu_to_be_32(MLX5_COMP_ALWAYS <<
186 						MLX5_COMP_MODE_OFFSET);
187 		wqe->ctrl[3] = txq->elts_head;
188 		txq->elts_comp = 0;
189 	}
190 #ifdef MLX5_PMD_SOFT_COUNTERS
191 	txq->stats.opackets += n;
192 #endif
193 	mlx5_tx_dbrec(txq, wqe);
194 	return n;
195 }
196 
197 /**
198  * Send burst of packets with Enhanced MPW. If it encounters a multi-seg packet,
199  * it returns to make it processed by txq_scatter_v(). All the packets in
200  * the pkts list should be single segment packets having same offload flags.
201  * This must be checked by txq_count_contig_single_seg() and txq_calc_offload().
202  *
203  * @param txq
204  *   Pointer to TX queue structure.
205  * @param pkts
206  *   Pointer to array of packets to be sent.
207  * @param pkts_n
208  *   Number of packets to be sent (<= MLX5_VPMD_TX_MAX_BURST).
209  * @param cs_flags
210  *   Checksum offload flags to be written in the descriptor.
211  * @param metadata
212  *   Metadata value to be written in the descriptor.
213  *
214  * @return
215  *   Number of packets successfully transmitted (<= pkts_n).
216  */
217 static inline uint16_t
218 txq_burst_v(struct mlx5_txq_data *txq, struct rte_mbuf **pkts, uint16_t pkts_n,
219 	    uint8_t cs_flags, rte_be32_t metadata)
220 {
221 	struct rte_mbuf **elts;
222 	uint16_t elts_head = txq->elts_head;
223 	const uint16_t elts_n = 1 << txq->elts_n;
224 	const uint16_t elts_m = elts_n - 1;
225 	const unsigned int nb_dword_per_wqebb =
226 		MLX5_WQE_SIZE / MLX5_WQE_DWORD_SIZE;
227 	const unsigned int nb_dword_in_hdr =
228 		sizeof(struct mlx5_wqe) / MLX5_WQE_DWORD_SIZE;
229 	unsigned int n = 0;
230 	unsigned int pos;
231 	uint16_t max_elts;
232 	uint16_t max_wqe;
233 	uint32_t comp_req;
234 	const uint16_t wq_n = 1 << txq->wqe_n;
235 	const uint16_t wq_mask = wq_n - 1;
236 	uint16_t wq_idx = txq->wqe_ci & wq_mask;
237 	volatile struct mlx5_wqe64 *wq =
238 		&((volatile struct mlx5_wqe64 *)txq->wqes)[wq_idx];
239 	volatile struct mlx5_wqe *wqe = (volatile struct mlx5_wqe *)wq;
240 	const uint8x16_t ctrl_shuf_m = {
241 		 3,  2,  1,  0, /* bswap32 */
242 		 7,  6,  5,  4, /* bswap32 */
243 		11, 10,  9,  8, /* bswap32 */
244 		12, 13, 14, 15
245 	};
246 	uint8x16_t *t_wqe;
247 	uint8_t *dseg;
248 	uint8x16_t ctrl;
249 
250 	/* Make sure all packets can fit into a single WQE. */
251 	assert(elts_n > pkts_n);
252 	mlx5_tx_complete(txq);
253 	max_elts = (elts_n - (elts_head - txq->elts_tail));
254 	max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
255 	pkts_n = RTE_MIN((unsigned int)RTE_MIN(pkts_n, max_wqe), max_elts);
256 	if (unlikely(!pkts_n))
257 		return 0;
258 	elts = &(*txq->elts)[elts_head & elts_m];
259 	/* Loop for available tailroom first. */
260 	n = RTE_MIN(elts_n - (elts_head & elts_m), pkts_n);
261 	for (pos = 0; pos < (n & -2); pos += 2)
262 		vst1q_u64((void *)&elts[pos], vld1q_u64((void *)&pkts[pos]));
263 	if (n & 1)
264 		elts[pos] = pkts[pos];
265 	/* Check if it crosses the end of the queue. */
266 	if (unlikely(n < pkts_n)) {
267 		elts = &(*txq->elts)[0];
268 		for (pos = 0; pos < pkts_n - n; ++pos)
269 			elts[pos] = pkts[n + pos];
270 	}
271 	txq->elts_head += pkts_n;
272 	/* Save title WQEBB pointer. */
273 	t_wqe = (uint8x16_t *)wqe;
274 	dseg = (uint8_t *)(wqe + 1);
275 	/* Calculate the number of entries to the end. */
276 	n = RTE_MIN(
277 		(wq_n - wq_idx) * nb_dword_per_wqebb - nb_dword_in_hdr,
278 		pkts_n);
279 	/* Fill DSEGs. */
280 	txq_wr_dseg_v(txq, dseg, pkts, n);
281 	/* Check if it crosses the end of the queue. */
282 	if (n < pkts_n) {
283 		dseg = (uint8_t *)txq->wqes;
284 		txq_wr_dseg_v(txq, dseg, &pkts[n], pkts_n - n);
285 	}
286 	if (txq->elts_comp + pkts_n < MLX5_TX_COMP_THRESH) {
287 		txq->elts_comp += pkts_n;
288 		comp_req = MLX5_COMP_ONLY_FIRST_ERR << MLX5_COMP_MODE_OFFSET;
289 	} else {
290 		/* A CQE slot must always be available. */
291 		assert((1u << txq->cqe_n) - (txq->cq_pi++ - txq->cq_ci));
292 		/* Request a completion. */
293 		txq->elts_comp = 0;
294 		comp_req = MLX5_COMP_ALWAYS << MLX5_COMP_MODE_OFFSET;
295 	}
296 	/* Fill CTRL in the header. */
297 	ctrl = vreinterpretq_u8_u32((uint32x4_t) {
298 			MLX5_OPC_MOD_ENHANCED_MPSW << 24 |
299 			txq->wqe_ci << 8 | MLX5_OPCODE_ENHANCED_MPSW,
300 			txq->qp_num_8s | (pkts_n + 2),
301 			comp_req,
302 			txq->elts_head });
303 	ctrl = vqtbl1q_u8(ctrl, ctrl_shuf_m);
304 	vst1q_u8((void *)t_wqe, ctrl);
305 	/* Fill ESEG in the header. */
306 	vst1q_u32((void *)(t_wqe + 1),
307 		 ((uint32x4_t) { 0, cs_flags, metadata, 0 }));
308 #ifdef MLX5_PMD_SOFT_COUNTERS
309 	txq->stats.opackets += pkts_n;
310 #endif
311 	txq->wqe_ci += (nb_dword_in_hdr + pkts_n + (nb_dword_per_wqebb - 1)) /
312 		       nb_dword_per_wqebb;
313 	/* Ring QP doorbell. */
314 	mlx5_tx_dbrec_cond_wmb(txq, wqe, pkts_n < MLX5_VPMD_TX_MAX_BURST);
315 	return pkts_n;
316 }
317 
318 /**
319  * Store free buffers to RX SW ring.
320  *
321  * @param rxq
322  *   Pointer to RX queue structure.
323  * @param pkts
324  *   Pointer to array of packets to be stored.
325  * @param pkts_n
326  *   Number of packets to be stored.
327  */
328 static inline void
329 rxq_copy_mbuf_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t n)
330 {
331 	const uint16_t q_mask = (1 << rxq->elts_n) - 1;
332 	struct rte_mbuf **elts = &(*rxq->elts)[rxq->rq_pi & q_mask];
333 	unsigned int pos;
334 	uint16_t p = n & -2;
335 
336 	for (pos = 0; pos < p; pos += 2) {
337 		uint64x2_t mbp;
338 
339 		mbp = vld1q_u64((void *)&elts[pos]);
340 		vst1q_u64((void *)&pkts[pos], mbp);
341 	}
342 	if (n & 1)
343 		pkts[pos] = elts[pos];
344 }
345 
346 /**
347  * Decompress a compressed completion and fill in mbufs in RX SW ring with data
348  * extracted from the title completion descriptor.
349  *
350  * @param rxq
351  *   Pointer to RX queue structure.
352  * @param cq
353  *   Pointer to completion array having a compressed completion at first.
354  * @param elts
355  *   Pointer to SW ring to be filled. The first mbuf has to be pre-built from
356  *   the title completion descriptor to be copied to the rest of mbufs.
357  *
358  * @return
359  *   Number of mini-CQEs successfully decompressed.
360  */
361 static inline uint16_t
362 rxq_cq_decompress_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq,
363 		    struct rte_mbuf **elts)
364 {
365 	volatile struct mlx5_mini_cqe8 *mcq = (void *)&(cq + 1)->pkt_info;
366 	struct rte_mbuf *t_pkt = elts[0]; /* Title packet is pre-built. */
367 	unsigned int pos;
368 	unsigned int i;
369 	unsigned int inv = 0;
370 	/* Mask to shuffle from extracted mini CQE to mbuf. */
371 	const uint8x16_t mcqe_shuf_m1 = {
372 		-1, -1, -1, -1, /* skip packet_type */
373 		 7,  6, -1, -1, /* pkt_len, bswap16 */
374 		 7,  6,         /* data_len, bswap16 */
375 		-1, -1,         /* skip vlan_tci */
376 		 3,  2,  1,  0  /* hash.rss, bswap32 */
377 	};
378 	const uint8x16_t mcqe_shuf_m2 = {
379 		-1, -1, -1, -1, /* skip packet_type */
380 		15, 14, -1, -1, /* pkt_len, bswap16 */
381 		15, 14,         /* data_len, bswap16 */
382 		-1, -1,         /* skip vlan_tci */
383 		11, 10,  9,  8  /* hash.rss, bswap32 */
384 	};
385 	/* Restore the compressed count. Must be 16 bits. */
386 	const uint16_t mcqe_n = t_pkt->data_len +
387 				(rxq->crc_present * RTE_ETHER_CRC_LEN);
388 	const uint64x2_t rearm =
389 		vld1q_u64((void *)&t_pkt->rearm_data);
390 	const uint32x4_t rxdf_mask = {
391 		0xffffffff, /* packet_type */
392 		0,          /* skip pkt_len */
393 		0xffff0000, /* vlan_tci, skip data_len */
394 		0,          /* skip hash.rss */
395 	};
396 	const uint8x16_t rxdf =
397 		vandq_u8(vld1q_u8((void *)&t_pkt->rx_descriptor_fields1),
398 			 vreinterpretq_u8_u32(rxdf_mask));
399 	const uint16x8_t crc_adj = {
400 		0, 0,
401 		rxq->crc_present * RTE_ETHER_CRC_LEN, 0,
402 		rxq->crc_present * RTE_ETHER_CRC_LEN, 0,
403 		0, 0
404 	};
405 	const uint32_t flow_tag = t_pkt->hash.fdir.hi;
406 #ifdef MLX5_PMD_SOFT_COUNTERS
407 	uint32_t rcvd_byte = 0;
408 #endif
409 	/* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */
410 	const uint8x8_t len_shuf_m = {
411 		 7,  6,         /* 1st mCQE */
412 		15, 14,         /* 2nd mCQE */
413 		23, 22,         /* 3rd mCQE */
414 		31, 30          /* 4th mCQE */
415 	};
416 
417 	/*
418 	 * A. load mCQEs into a 128bit register.
419 	 * B. store rearm data to mbuf.
420 	 * C. combine data from mCQEs with rx_descriptor_fields1.
421 	 * D. store rx_descriptor_fields1.
422 	 * E. store flow tag (rte_flow mark).
423 	 */
424 	for (pos = 0; pos < mcqe_n; ) {
425 		uint8_t *p = (void *)&mcq[pos % 8];
426 		uint8_t *e0 = (void *)&elts[pos]->rearm_data;
427 		uint8_t *e1 = (void *)&elts[pos + 1]->rearm_data;
428 		uint8_t *e2 = (void *)&elts[pos + 2]->rearm_data;
429 		uint8_t *e3 = (void *)&elts[pos + 3]->rearm_data;
430 		uint16x4_t byte_cnt;
431 #ifdef MLX5_PMD_SOFT_COUNTERS
432 		uint16x4_t invalid_mask =
433 			vcreate_u16(mcqe_n - pos < MLX5_VPMD_DESCS_PER_LOOP ?
434 				    -1UL << ((mcqe_n - pos) *
435 					     sizeof(uint16_t) * 8) : 0);
436 #endif
437 
438 		if (!(pos & 0x7) && pos + 8 < mcqe_n)
439 			rte_prefetch0((void *)(cq + pos + 8));
440 		__asm__ volatile (
441 		/* A.1 load mCQEs into a 128bit register. */
442 		"ld1 {v16.16b - v17.16b}, [%[mcq]] \n\t"
443 		/* B.1 store rearm data to mbuf. */
444 		"st1 {%[rearm].2d}, [%[e0]] \n\t"
445 		"add %[e0], %[e0], #16 \n\t"
446 		"st1 {%[rearm].2d}, [%[e1]] \n\t"
447 		"add %[e1], %[e1], #16 \n\t"
448 		/* C.1 combine data from mCQEs with rx_descriptor_fields1. */
449 		"tbl v18.16b, {v16.16b}, %[mcqe_shuf_m1].16b \n\t"
450 		"tbl v19.16b, {v16.16b}, %[mcqe_shuf_m2].16b \n\t"
451 		"sub v18.8h, v18.8h, %[crc_adj].8h \n\t"
452 		"sub v19.8h, v19.8h, %[crc_adj].8h \n\t"
453 		"orr v18.16b, v18.16b, %[rxdf].16b \n\t"
454 		"orr v19.16b, v19.16b, %[rxdf].16b \n\t"
455 		/* D.1 store rx_descriptor_fields1. */
456 		"st1 {v18.2d}, [%[e0]] \n\t"
457 		"st1 {v19.2d}, [%[e1]] \n\t"
458 		/* B.1 store rearm data to mbuf. */
459 		"st1 {%[rearm].2d}, [%[e2]] \n\t"
460 		"add %[e2], %[e2], #16 \n\t"
461 		"st1 {%[rearm].2d}, [%[e3]] \n\t"
462 		"add %[e3], %[e3], #16 \n\t"
463 		/* C.1 combine data from mCQEs with rx_descriptor_fields1. */
464 		"tbl v18.16b, {v17.16b}, %[mcqe_shuf_m1].16b \n\t"
465 		"tbl v19.16b, {v17.16b}, %[mcqe_shuf_m2].16b \n\t"
466 		"sub v18.8h, v18.8h, %[crc_adj].8h \n\t"
467 		"sub v19.8h, v19.8h, %[crc_adj].8h \n\t"
468 		"orr v18.16b, v18.16b, %[rxdf].16b \n\t"
469 		"orr v19.16b, v19.16b, %[rxdf].16b \n\t"
470 		/* D.1 store rx_descriptor_fields1. */
471 		"st1 {v18.2d}, [%[e2]] \n\t"
472 		"st1 {v19.2d}, [%[e3]] \n\t"
473 #ifdef MLX5_PMD_SOFT_COUNTERS
474 		"tbl %[byte_cnt].8b, {v16.16b - v17.16b}, %[len_shuf_m].8b \n\t"
475 #endif
476 		:[byte_cnt]"=&w"(byte_cnt)
477 		:[mcq]"r"(p),
478 		 [rxdf]"w"(rxdf),
479 		 [rearm]"w"(rearm),
480 		 [e3]"r"(e3), [e2]"r"(e2), [e1]"r"(e1), [e0]"r"(e0),
481 		 [mcqe_shuf_m1]"w"(mcqe_shuf_m1),
482 		 [mcqe_shuf_m2]"w"(mcqe_shuf_m2),
483 		 [crc_adj]"w"(crc_adj),
484 		 [len_shuf_m]"w"(len_shuf_m)
485 		:"memory", "v16", "v17", "v18", "v19");
486 #ifdef MLX5_PMD_SOFT_COUNTERS
487 		byte_cnt = vbic_u16(byte_cnt, invalid_mask);
488 		rcvd_byte += vget_lane_u64(vpaddl_u32(vpaddl_u16(byte_cnt)), 0);
489 #endif
490 		if (rxq->mark) {
491 			/* E.1 store flow tag (rte_flow mark). */
492 			elts[pos]->hash.fdir.hi = flow_tag;
493 			elts[pos + 1]->hash.fdir.hi = flow_tag;
494 			elts[pos + 2]->hash.fdir.hi = flow_tag;
495 			elts[pos + 3]->hash.fdir.hi = flow_tag;
496 		}
497 		pos += MLX5_VPMD_DESCS_PER_LOOP;
498 		/* Move to next CQE and invalidate consumed CQEs. */
499 		if (!(pos & 0x7) && pos < mcqe_n) {
500 			mcq = (void *)&(cq + pos)->pkt_info;
501 			for (i = 0; i < 8; ++i)
502 				cq[inv++].op_own = MLX5_CQE_INVALIDATE;
503 		}
504 	}
505 	/* Invalidate the rest of CQEs. */
506 	for (; inv < mcqe_n; ++inv)
507 		cq[inv].op_own = MLX5_CQE_INVALIDATE;
508 #ifdef MLX5_PMD_SOFT_COUNTERS
509 	rxq->stats.ipackets += mcqe_n;
510 	rxq->stats.ibytes += rcvd_byte;
511 #endif
512 	rxq->cq_ci += mcqe_n;
513 	return mcqe_n;
514 }
515 
516 /**
517  * Calculate packet type and offload flag for mbuf and store it.
518  *
519  * @param rxq
520  *   Pointer to RX queue structure.
521  * @param ptype_info
522  *   Array of four 4bytes packet type info extracted from the original
523  *   completion descriptor.
524  * @param flow_tag
525  *   Array of four 4bytes flow ID extracted from the original completion
526  *   descriptor.
527  * @param op_err
528  *   Opcode vector having responder error status. Each field is 4B.
529  * @param pkts
530  *   Pointer to array of packets to be filled.
531  */
532 static inline void
533 rxq_cq_to_ptype_oflags_v(struct mlx5_rxq_data *rxq,
534 			 uint32x4_t ptype_info, uint32x4_t flow_tag,
535 			 uint16x4_t op_err, struct rte_mbuf **pkts)
536 {
537 	uint16x4_t ptype;
538 	uint32x4_t pinfo, cv_flags;
539 	uint32x4_t ol_flags =
540 		vdupq_n_u32(rxq->rss_hash * PKT_RX_RSS_HASH |
541 			    rxq->hw_timestamp * PKT_RX_TIMESTAMP);
542 	const uint32x4_t ptype_ol_mask = { 0x106, 0x106, 0x106, 0x106 };
543 	const uint8x16_t cv_flag_sel = {
544 		0,
545 		(uint8_t)(PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED),
546 		(uint8_t)(PKT_RX_IP_CKSUM_GOOD >> 1),
547 		0,
548 		(uint8_t)(PKT_RX_L4_CKSUM_GOOD >> 1),
549 		0,
550 		(uint8_t)((PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1),
551 		0, 0, 0, 0, 0, 0, 0, 0, 0
552 	};
553 	const uint32x4_t cv_mask =
554 		vdupq_n_u32(PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
555 			    PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED);
556 	const uint64x1_t mbuf_init = vld1_u64(&rxq->mbuf_initializer);
557 	const uint64x1_t r32_mask = vcreate_u64(0xffffffff);
558 	uint64x2_t rearm0, rearm1, rearm2, rearm3;
559 	uint8_t pt_idx0, pt_idx1, pt_idx2, pt_idx3;
560 
561 	if (rxq->mark) {
562 		const uint32x4_t ft_def = vdupq_n_u32(MLX5_FLOW_MARK_DEFAULT);
563 		const uint32x4_t fdir_flags = vdupq_n_u32(PKT_RX_FDIR);
564 		uint32x4_t fdir_id_flags = vdupq_n_u32(PKT_RX_FDIR_ID);
565 		uint32x4_t invalid_mask;
566 
567 		/* Check if flow tag is non-zero then set PKT_RX_FDIR. */
568 		invalid_mask = vceqzq_u32(flow_tag);
569 		ol_flags = vorrq_u32(ol_flags,
570 				     vbicq_u32(fdir_flags, invalid_mask));
571 		/* Mask out invalid entries. */
572 		fdir_id_flags = vbicq_u32(fdir_id_flags, invalid_mask);
573 		/* Check if flow tag MLX5_FLOW_MARK_DEFAULT. */
574 		ol_flags = vorrq_u32(ol_flags,
575 				     vbicq_u32(fdir_id_flags,
576 					       vceqq_u32(flow_tag, ft_def)));
577 	}
578 	/*
579 	 * ptype_info has the following:
580 	 * bit[1]     = l3_ok
581 	 * bit[2]     = l4_ok
582 	 * bit[8]     = cv
583 	 * bit[11:10] = l3_hdr_type
584 	 * bit[14:12] = l4_hdr_type
585 	 * bit[15]    = ip_frag
586 	 * bit[16]    = tunneled
587 	 * bit[17]    = outer_l3_type
588 	 */
589 	ptype = vshrn_n_u32(ptype_info, 10);
590 	/* Errored packets will have RTE_PTYPE_ALL_MASK. */
591 	ptype = vorr_u16(ptype, op_err);
592 	pt_idx0 = vget_lane_u8(vreinterpret_u8_u16(ptype), 6);
593 	pt_idx1 = vget_lane_u8(vreinterpret_u8_u16(ptype), 4);
594 	pt_idx2 = vget_lane_u8(vreinterpret_u8_u16(ptype), 2);
595 	pt_idx3 = vget_lane_u8(vreinterpret_u8_u16(ptype), 0);
596 	pkts[0]->packet_type = mlx5_ptype_table[pt_idx0] |
597 			       !!(pt_idx0 & (1 << 6)) * rxq->tunnel;
598 	pkts[1]->packet_type = mlx5_ptype_table[pt_idx1] |
599 			       !!(pt_idx1 & (1 << 6)) * rxq->tunnel;
600 	pkts[2]->packet_type = mlx5_ptype_table[pt_idx2] |
601 			       !!(pt_idx2 & (1 << 6)) * rxq->tunnel;
602 	pkts[3]->packet_type = mlx5_ptype_table[pt_idx3] |
603 			       !!(pt_idx3 & (1 << 6)) * rxq->tunnel;
604 	/* Fill flags for checksum and VLAN. */
605 	pinfo = vandq_u32(ptype_info, ptype_ol_mask);
606 	pinfo = vreinterpretq_u32_u8(
607 		vqtbl1q_u8(cv_flag_sel, vreinterpretq_u8_u32(pinfo)));
608 	/* Locate checksum flags at byte[2:1] and merge with VLAN flags. */
609 	cv_flags = vshlq_n_u32(pinfo, 9);
610 	cv_flags = vorrq_u32(pinfo, cv_flags);
611 	/* Move back flags to start from byte[0]. */
612 	cv_flags = vshrq_n_u32(cv_flags, 8);
613 	/* Mask out garbage bits. */
614 	cv_flags = vandq_u32(cv_flags, cv_mask);
615 	/* Merge to ol_flags. */
616 	ol_flags = vorrq_u32(ol_flags, cv_flags);
617 	/* Merge mbuf_init and ol_flags, and store. */
618 	rearm0 = vcombine_u64(mbuf_init,
619 			      vshr_n_u64(vget_high_u64(vreinterpretq_u64_u32(
620 						       ol_flags)), 32));
621 	rearm1 = vcombine_u64(mbuf_init,
622 			      vand_u64(vget_high_u64(vreinterpretq_u64_u32(
623 						     ol_flags)), r32_mask));
624 	rearm2 = vcombine_u64(mbuf_init,
625 			      vshr_n_u64(vget_low_u64(vreinterpretq_u64_u32(
626 						      ol_flags)), 32));
627 	rearm3 = vcombine_u64(mbuf_init,
628 			      vand_u64(vget_low_u64(vreinterpretq_u64_u32(
629 						    ol_flags)), r32_mask));
630 	vst1q_u64((void *)&pkts[0]->rearm_data, rearm0);
631 	vst1q_u64((void *)&pkts[1]->rearm_data, rearm1);
632 	vst1q_u64((void *)&pkts[2]->rearm_data, rearm2);
633 	vst1q_u64((void *)&pkts[3]->rearm_data, rearm3);
634 }
635 
636 /**
637  * Receive burst of packets. An errored completion also consumes a mbuf, but the
638  * packet_type is set to be RTE_PTYPE_ALL_MASK. Marked mbufs should be freed
639  * before returning to application.
640  *
641  * @param rxq
642  *   Pointer to RX queue structure.
643  * @param[out] pkts
644  *   Array to store received packets.
645  * @param pkts_n
646  *   Maximum number of packets in array.
647  * @param[out] err
648  *   Pointer to a flag. Set non-zero value if pkts array has at least one error
649  *   packet to handle.
650  *
651  * @return
652  *   Number of packets received including errors (<= pkts_n).
653  */
654 static inline uint16_t
655 rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
656 	    uint64_t *err)
657 {
658 	const uint16_t q_n = 1 << rxq->cqe_n;
659 	const uint16_t q_mask = q_n - 1;
660 	volatile struct mlx5_cqe *cq;
661 	struct rte_mbuf **elts;
662 	unsigned int pos;
663 	uint64_t n;
664 	uint16_t repl_n;
665 	uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
666 	uint16_t nocmp_n = 0;
667 	uint16_t rcvd_pkt = 0;
668 	unsigned int cq_idx = rxq->cq_ci & q_mask;
669 	unsigned int elts_idx;
670 	const uint16x4_t ownership = vdup_n_u16(!(rxq->cq_ci & (q_mask + 1)));
671 	const uint16x4_t owner_check = vcreate_u16(0x0001000100010001);
672 	const uint16x4_t opcode_check = vcreate_u16(0x00f000f000f000f0);
673 	const uint16x4_t format_check = vcreate_u16(0x000c000c000c000c);
674 	const uint16x4_t resp_err_check = vcreate_u16(0x00e000e000e000e0);
675 #ifdef MLX5_PMD_SOFT_COUNTERS
676 	uint32_t rcvd_byte = 0;
677 #endif
678 	/* Mask to generate 16B length vector. */
679 	const uint8x8_t len_shuf_m = {
680 		52, 53,         /* 4th CQE */
681 		36, 37,         /* 3rd CQE */
682 		20, 21,         /* 2nd CQE */
683 		 4,  5          /* 1st CQE */
684 	};
685 	/* Mask to extract 16B data from a 64B CQE. */
686 	const uint8x16_t cqe_shuf_m = {
687 		28, 29,         /* hdr_type_etc */
688 		 0,             /* pkt_info */
689 		-1,             /* null */
690 		47, 46,         /* byte_cnt, bswap16 */
691 		31, 30,         /* vlan_info, bswap16 */
692 		15, 14, 13, 12, /* rx_hash_res, bswap32 */
693 		57, 58, 59,     /* flow_tag */
694 		63              /* op_own */
695 	};
696 	/* Mask to generate 16B data for mbuf. */
697 	const uint8x16_t mb_shuf_m = {
698 		 4,  5, -1, -1, /* pkt_len */
699 		 4,  5,         /* data_len */
700 		 6,  7,         /* vlan_tci */
701 		 8,  9, 10, 11, /* hash.rss */
702 		12, 13, 14, -1  /* hash.fdir.hi */
703 	};
704 	/* Mask to generate 16B owner vector. */
705 	const uint8x8_t owner_shuf_m = {
706 		63, -1,         /* 4th CQE */
707 		47, -1,         /* 3rd CQE */
708 		31, -1,         /* 2nd CQE */
709 		15, -1          /* 1st CQE */
710 	};
711 	/* Mask to generate a vector having packet_type/ol_flags. */
712 	const uint8x16_t ptype_shuf_m = {
713 		48, 49, 50, -1, /* 4th CQE */
714 		32, 33, 34, -1, /* 3rd CQE */
715 		16, 17, 18, -1, /* 2nd CQE */
716 		 0,  1,  2, -1  /* 1st CQE */
717 	};
718 	/* Mask to generate a vector having flow tags. */
719 	const uint8x16_t ftag_shuf_m = {
720 		60, 61, 62, -1, /* 4th CQE */
721 		44, 45, 46, -1, /* 3rd CQE */
722 		28, 29, 30, -1, /* 2nd CQE */
723 		12, 13, 14, -1  /* 1st CQE */
724 	};
725 	const uint16x8_t crc_adj = {
726 		0, 0, rxq->crc_present * RTE_ETHER_CRC_LEN, 0, 0, 0, 0, 0
727 	};
728 	const uint32x4_t flow_mark_adj = { 0, 0, 0, rxq->mark * (-1) };
729 
730 	assert(rxq->sges_n == 0);
731 	assert(rxq->cqe_n == rxq->elts_n);
732 	cq = &(*rxq->cqes)[cq_idx];
733 	rte_prefetch_non_temporal(cq);
734 	rte_prefetch_non_temporal(cq + 1);
735 	rte_prefetch_non_temporal(cq + 2);
736 	rte_prefetch_non_temporal(cq + 3);
737 	pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
738 	repl_n = q_n - (rxq->rq_ci - rxq->rq_pi);
739 	if (repl_n >= rxq->rq_repl_thresh)
740 		mlx5_rx_replenish_bulk_mbuf(rxq, repl_n);
741 	/* See if there're unreturned mbufs from compressed CQE. */
742 	rcvd_pkt = rxq->decompressed;
743 	if (rcvd_pkt > 0) {
744 		rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
745 		rxq_copy_mbuf_v(rxq, pkts, rcvd_pkt);
746 		rxq->rq_pi += rcvd_pkt;
747 		pkts += rcvd_pkt;
748 		rxq->decompressed -= rcvd_pkt;
749 	}
750 	elts_idx = rxq->rq_pi & q_mask;
751 	elts = &(*rxq->elts)[elts_idx];
752 	/* Not to overflow pkts array. */
753 	pkts_n = RTE_ALIGN_FLOOR(pkts_n - rcvd_pkt, MLX5_VPMD_DESCS_PER_LOOP);
754 	/* Not to cross queue end. */
755 	pkts_n = RTE_MIN(pkts_n, q_n - elts_idx);
756 	pkts_n = RTE_MIN(pkts_n, q_n - cq_idx);
757 	if (!pkts_n)
758 		return rcvd_pkt;
759 	/* At this point, there shouldn't be any remained packets. */
760 	assert(rxq->decompressed == 0);
761 	/*
762 	 * Note that vectors have reverse order - {v3, v2, v1, v0}, because
763 	 * there's no instruction to count trailing zeros. __builtin_clzl() is
764 	 * used instead.
765 	 *
766 	 * A. copy 4 mbuf pointers from elts ring to returing pkts.
767 	 * B. load 64B CQE and extract necessary fields
768 	 *    Final 16bytes cqes[] extracted from original 64bytes CQE has the
769 	 *    following structure:
770 	 *        struct {
771 	 *          uint16_t hdr_type_etc;
772 	 *          uint8_t  pkt_info;
773 	 *          uint8_t  rsvd;
774 	 *          uint16_t byte_cnt;
775 	 *          uint16_t vlan_info;
776 	 *          uint32_t rx_has_res;
777 	 *          uint8_t  flow_tag[3];
778 	 *          uint8_t  op_own;
779 	 *        } c;
780 	 * C. fill in mbuf.
781 	 * D. get valid CQEs.
782 	 * E. find compressed CQE.
783 	 */
784 	for (pos = 0;
785 	     pos < pkts_n;
786 	     pos += MLX5_VPMD_DESCS_PER_LOOP) {
787 		uint16x4_t op_own;
788 		uint16x4_t opcode, owner_mask, invalid_mask;
789 		uint16x4_t comp_mask;
790 		uint16x4_t mask;
791 		uint16x4_t byte_cnt;
792 		uint32x4_t ptype_info, flow_tag;
793 		register uint64x2_t c0, c1, c2, c3;
794 		uint8_t *p0, *p1, *p2, *p3;
795 		uint8_t *e0 = (void *)&elts[pos]->pkt_len;
796 		uint8_t *e1 = (void *)&elts[pos + 1]->pkt_len;
797 		uint8_t *e2 = (void *)&elts[pos + 2]->pkt_len;
798 		uint8_t *e3 = (void *)&elts[pos + 3]->pkt_len;
799 		void *elts_p = (void *)&elts[pos];
800 		void *pkts_p = (void *)&pkts[pos];
801 
802 		/* A.0 do not cross the end of CQ. */
803 		mask = vcreate_u16(pkts_n - pos < MLX5_VPMD_DESCS_PER_LOOP ?
804 				   -1UL >> ((pkts_n - pos) *
805 					    sizeof(uint16_t) * 8) : 0);
806 		p0 = (void *)&cq[pos].pkt_info;
807 		p1 = p0 + (pkts_n - pos > 1) * sizeof(struct mlx5_cqe);
808 		p2 = p1 + (pkts_n - pos > 2) * sizeof(struct mlx5_cqe);
809 		p3 = p2 + (pkts_n - pos > 3) * sizeof(struct mlx5_cqe);
810 		/* B.0 (CQE 3) load a block having op_own. */
811 		c3 = vld1q_u64((uint64_t *)(p3 + 48));
812 		/* B.0 (CQE 2) load a block having op_own. */
813 		c2 = vld1q_u64((uint64_t *)(p2 + 48));
814 		/* B.0 (CQE 1) load a block having op_own. */
815 		c1 = vld1q_u64((uint64_t *)(p1 + 48));
816 		/* B.0 (CQE 0) load a block having op_own. */
817 		c0 = vld1q_u64((uint64_t *)(p0 + 48));
818 		/* Synchronize for loading the rest of blocks. */
819 		rte_cio_rmb();
820 		/* Prefetch next 4 CQEs. */
821 		if (pkts_n - pos >= 2 * MLX5_VPMD_DESCS_PER_LOOP) {
822 			unsigned int next = pos + MLX5_VPMD_DESCS_PER_LOOP;
823 			rte_prefetch_non_temporal(&cq[next]);
824 			rte_prefetch_non_temporal(&cq[next + 1]);
825 			rte_prefetch_non_temporal(&cq[next + 2]);
826 			rte_prefetch_non_temporal(&cq[next + 3]);
827 		}
828 		__asm__ volatile (
829 		/* B.1 (CQE 3) load the rest of blocks. */
830 		"ld1 {v16.16b - v18.16b}, [%[p3]] \n\t"
831 		/* B.2 (CQE 3) move the block having op_own. */
832 		"mov v19.16b, %[c3].16b \n\t"
833 		/* B.3 (CQE 3) extract 16B fields. */
834 		"tbl v23.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
835 		/* B.1 (CQE 2) load the rest of blocks. */
836 		"ld1 {v16.16b - v18.16b}, [%[p2]] \n\t"
837 		/* B.4 (CQE 3) adjust CRC length. */
838 		"sub v23.8h, v23.8h, %[crc_adj].8h \n\t"
839 		/* C.1 (CQE 3) generate final structure for mbuf. */
840 		"tbl v15.16b, {v23.16b}, %[mb_shuf_m].16b \n\t"
841 		/* B.2 (CQE 2) move the block having op_own. */
842 		"mov v19.16b, %[c2].16b \n\t"
843 		/* B.3 (CQE 2) extract 16B fields. */
844 		"tbl v22.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
845 		/* B.1 (CQE 1) load the rest of blocks. */
846 		"ld1 {v16.16b - v18.16b}, [%[p1]] \n\t"
847 		/* B.4 (CQE 2) adjust CRC length. */
848 		"sub v22.8h, v22.8h, %[crc_adj].8h \n\t"
849 		/* C.1 (CQE 2) generate final structure for mbuf. */
850 		"tbl v14.16b, {v22.16b}, %[mb_shuf_m].16b \n\t"
851 		/* B.2 (CQE 1) move the block having op_own. */
852 		"mov v19.16b, %[c1].16b \n\t"
853 		/* B.3 (CQE 1) extract 16B fields. */
854 		"tbl v21.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
855 		/* B.1 (CQE 0) load the rest of blocks. */
856 		"ld1 {v16.16b - v18.16b}, [%[p0]] \n\t"
857 		/* B.4 (CQE 1) adjust CRC length. */
858 		"sub v21.8h, v21.8h, %[crc_adj].8h \n\t"
859 		/* C.1 (CQE 1) generate final structure for mbuf. */
860 		"tbl v13.16b, {v21.16b}, %[mb_shuf_m].16b \n\t"
861 		/* B.2 (CQE 0) move the block having op_own. */
862 		"mov v19.16b, %[c0].16b \n\t"
863 		/* A.1 load mbuf pointers. */
864 		"ld1 {v24.2d - v25.2d}, [%[elts_p]] \n\t"
865 		/* B.3 (CQE 0) extract 16B fields. */
866 		"tbl v20.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
867 		/* B.4 (CQE 0) adjust CRC length. */
868 		"sub v20.8h, v20.8h, %[crc_adj].8h \n\t"
869 		/* D.1 extract op_own byte. */
870 		"tbl %[op_own].8b, {v20.16b - v23.16b}, %[owner_shuf_m].8b \n\t"
871 		/* C.2 (CQE 3) adjust flow mark. */
872 		"add v15.4s, v15.4s, %[flow_mark_adj].4s \n\t"
873 		/* C.3 (CQE 3) fill in mbuf - rx_descriptor_fields1. */
874 		"st1 {v15.2d}, [%[e3]] \n\t"
875 		/* C.2 (CQE 2) adjust flow mark. */
876 		"add v14.4s, v14.4s, %[flow_mark_adj].4s \n\t"
877 		/* C.3 (CQE 2) fill in mbuf - rx_descriptor_fields1. */
878 		"st1 {v14.2d}, [%[e2]] \n\t"
879 		/* C.1 (CQE 0) generate final structure for mbuf. */
880 		"tbl v12.16b, {v20.16b}, %[mb_shuf_m].16b \n\t"
881 		/* C.2 (CQE 1) adjust flow mark. */
882 		"add v13.4s, v13.4s, %[flow_mark_adj].4s \n\t"
883 		/* C.3 (CQE 1) fill in mbuf - rx_descriptor_fields1. */
884 		"st1 {v13.2d}, [%[e1]] \n\t"
885 #ifdef MLX5_PMD_SOFT_COUNTERS
886 		/* Extract byte_cnt. */
887 		"tbl %[byte_cnt].8b, {v20.16b - v23.16b}, %[len_shuf_m].8b \n\t"
888 #endif
889 		/* Extract ptype_info. */
890 		"tbl %[ptype_info].16b, {v20.16b - v23.16b}, %[ptype_shuf_m].16b \n\t"
891 		/* Extract flow_tag. */
892 		"tbl %[flow_tag].16b, {v20.16b - v23.16b}, %[ftag_shuf_m].16b \n\t"
893 		/* A.2 copy mbuf pointers. */
894 		"st1 {v24.2d - v25.2d}, [%[pkts_p]] \n\t"
895 		/* C.2 (CQE 0) adjust flow mark. */
896 		"add v12.4s, v12.4s, %[flow_mark_adj].4s \n\t"
897 		/* C.3 (CQE 1) fill in mbuf - rx_descriptor_fields1. */
898 		"st1 {v12.2d}, [%[e0]] \n\t"
899 		:[op_own]"=&w"(op_own),
900 		 [byte_cnt]"=&w"(byte_cnt),
901 		 [ptype_info]"=&w"(ptype_info),
902 		 [flow_tag]"=&w"(flow_tag)
903 		:[p3]"r"(p3), [p2]"r"(p2), [p1]"r"(p1), [p0]"r"(p0),
904 		 [e3]"r"(e3), [e2]"r"(e2), [e1]"r"(e1), [e0]"r"(e0),
905 		 [c3]"w"(c3), [c2]"w"(c2), [c1]"w"(c1), [c0]"w"(c0),
906 		 [elts_p]"r"(elts_p),
907 		 [pkts_p]"r"(pkts_p),
908 		 [cqe_shuf_m]"w"(cqe_shuf_m),
909 		 [mb_shuf_m]"w"(mb_shuf_m),
910 		 [owner_shuf_m]"w"(owner_shuf_m),
911 		 [len_shuf_m]"w"(len_shuf_m),
912 		 [ptype_shuf_m]"w"(ptype_shuf_m),
913 		 [ftag_shuf_m]"w"(ftag_shuf_m),
914 		 [crc_adj]"w"(crc_adj),
915 		 [flow_mark_adj]"w"(flow_mark_adj)
916 		:"memory",
917 		 "v12", "v13", "v14", "v15",
918 		 "v16", "v17", "v18", "v19",
919 		 "v20", "v21", "v22", "v23",
920 		 "v24", "v25");
921 		/* D.2 flip owner bit to mark CQEs from last round. */
922 		owner_mask = vand_u16(op_own, owner_check);
923 		owner_mask = vceq_u16(owner_mask, ownership);
924 		/* D.3 get mask for invalidated CQEs. */
925 		opcode = vand_u16(op_own, opcode_check);
926 		invalid_mask = vceq_u16(opcode_check, opcode);
927 		/* E.1 find compressed CQE format. */
928 		comp_mask = vand_u16(op_own, format_check);
929 		comp_mask = vceq_u16(comp_mask, format_check);
930 		/* D.4 mask out beyond boundary. */
931 		invalid_mask = vorr_u16(invalid_mask, mask);
932 		/* D.5 merge invalid_mask with invalid owner. */
933 		invalid_mask = vorr_u16(invalid_mask, owner_mask);
934 		/* E.2 mask out invalid entries. */
935 		comp_mask = vbic_u16(comp_mask, invalid_mask);
936 		/* E.3 get the first compressed CQE. */
937 		comp_idx = __builtin_clzl(vget_lane_u64(vreinterpret_u64_u16(
938 					  comp_mask), 0)) /
939 					  (sizeof(uint16_t) * 8);
940 		/* D.6 mask out entries after the compressed CQE. */
941 		mask = vcreate_u16(comp_idx < MLX5_VPMD_DESCS_PER_LOOP ?
942 				   -1UL >> (comp_idx * sizeof(uint16_t) * 8) :
943 				   0);
944 		invalid_mask = vorr_u16(invalid_mask, mask);
945 		/* D.7 count non-compressed valid CQEs. */
946 		n = __builtin_clzl(vget_lane_u64(vreinterpret_u64_u16(
947 				   invalid_mask), 0)) / (sizeof(uint16_t) * 8);
948 		nocmp_n += n;
949 		/* D.2 get the final invalid mask. */
950 		mask = vcreate_u16(n < MLX5_VPMD_DESCS_PER_LOOP ?
951 				   -1UL >> (n * sizeof(uint16_t) * 8) : 0);
952 		invalid_mask = vorr_u16(invalid_mask, mask);
953 		/* D.3 check error in opcode. */
954 		opcode = vceq_u16(resp_err_check, opcode);
955 		opcode = vbic_u16(opcode, invalid_mask);
956 		/* D.4 mark if any error is set */
957 		*err |= vget_lane_u64(vreinterpret_u64_u16(opcode), 0);
958 		/* C.4 fill in mbuf - rearm_data and packet_type. */
959 		rxq_cq_to_ptype_oflags_v(rxq, ptype_info, flow_tag,
960 					 opcode, &elts[pos]);
961 		if (rxq->hw_timestamp) {
962 			elts[pos]->timestamp =
963 				rte_be_to_cpu_64(
964 					container_of(p0, struct mlx5_cqe,
965 						     pkt_info)->timestamp);
966 			elts[pos + 1]->timestamp =
967 				rte_be_to_cpu_64(
968 					container_of(p1, struct mlx5_cqe,
969 						     pkt_info)->timestamp);
970 			elts[pos + 2]->timestamp =
971 				rte_be_to_cpu_64(
972 					container_of(p2, struct mlx5_cqe,
973 						     pkt_info)->timestamp);
974 			elts[pos + 3]->timestamp =
975 				rte_be_to_cpu_64(
976 					container_of(p3, struct mlx5_cqe,
977 						     pkt_info)->timestamp);
978 		}
979 #ifdef MLX5_PMD_SOFT_COUNTERS
980 		/* Add up received bytes count. */
981 		byte_cnt = vbic_u16(byte_cnt, invalid_mask);
982 		rcvd_byte += vget_lane_u64(vpaddl_u32(vpaddl_u16(byte_cnt)), 0);
983 #endif
984 		/*
985 		 * Break the loop unless more valid CQE is expected, or if
986 		 * there's a compressed CQE.
987 		 */
988 		if (n != MLX5_VPMD_DESCS_PER_LOOP)
989 			break;
990 	}
991 	/* If no new CQE seen, return without updating cq_db. */
992 	if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP))
993 		return rcvd_pkt;
994 	/* Update the consumer indexes for non-compressed CQEs. */
995 	assert(nocmp_n <= pkts_n);
996 	rxq->cq_ci += nocmp_n;
997 	rxq->rq_pi += nocmp_n;
998 	rcvd_pkt += nocmp_n;
999 #ifdef MLX5_PMD_SOFT_COUNTERS
1000 	rxq->stats.ipackets += nocmp_n;
1001 	rxq->stats.ibytes += rcvd_byte;
1002 #endif
1003 	/* Decompress the last CQE if compressed. */
1004 	if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP && comp_idx == n) {
1005 		assert(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
1006 		rxq->decompressed = rxq_cq_decompress_v(rxq, &cq[nocmp_n],
1007 							&elts[nocmp_n]);
1008 		/* Return more packets if needed. */
1009 		if (nocmp_n < pkts_n) {
1010 			uint16_t n = rxq->decompressed;
1011 
1012 			n = RTE_MIN(n, pkts_n - nocmp_n);
1013 			rxq_copy_mbuf_v(rxq, &pkts[nocmp_n], n);
1014 			rxq->rq_pi += n;
1015 			rcvd_pkt += n;
1016 			rxq->decompressed -= n;
1017 		}
1018 	}
1019 	rte_compiler_barrier();
1020 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1021 	return rcvd_pkt;
1022 }
1023 
1024 #endif /* RTE_PMD_MLX5_RXTX_VEC_NEON_H_ */
1025