xref: /dpdk/drivers/net/mlx5/mlx5_rxtx_vec_neon.h (revision 89f0711f9ddfb5822da9d34f384b92f72a61c4dc)
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright 2017 6WIND S.A.
5  *   Copyright 2017 Mellanox.
6  *
7  *   Redistribution and use in source and binary forms, with or without
8  *   modification, are permitted provided that the following conditions
9  *   are met:
10  *
11  *     * Redistributions of source code must retain the above copyright
12  *       notice, this list of conditions and the following disclaimer.
13  *     * Redistributions in binary form must reproduce the above copyright
14  *       notice, this list of conditions and the following disclaimer in
15  *       the documentation and/or other materials provided with the
16  *       distribution.
17  *     * Neither the name of 6WIND S.A. nor the names of its
18  *       contributors may be used to endorse or promote products derived
19  *       from this software without specific prior written permission.
20  *
21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  */
33 
34 #ifndef RTE_PMD_MLX5_RXTX_VEC_NEON_H_
35 #define RTE_PMD_MLX5_RXTX_VEC_NEON_H_
36 
37 #include <assert.h>
38 #include <stdint.h>
39 #include <string.h>
40 #include <stdlib.h>
41 #include <arm_neon.h>
42 
43 #include <rte_mbuf.h>
44 #include <rte_mempool.h>
45 #include <rte_prefetch.h>
46 
47 #include "mlx5.h"
48 #include "mlx5_utils.h"
49 #include "mlx5_rxtx.h"
50 #include "mlx5_rxtx_vec.h"
51 #include "mlx5_autoconf.h"
52 #include "mlx5_defs.h"
53 #include "mlx5_prm.h"
54 
55 #pragma GCC diagnostic ignored "-Wcast-qual"
56 
57 /**
58  * Fill in buffer descriptors in a multi-packet send descriptor.
59  *
60  * @param txq
61  *   Pointer to TX queue structure.
62  * @param dseg
63  *   Pointer to buffer descriptor to be written.
64  * @param pkts
65  *   Pointer to array of packets to be sent.
66  * @param n
67  *   Number of packets to be filled.
68  */
69 static inline void
70 txq_wr_dseg_v(struct mlx5_txq_data *txq, uint8_t *dseg,
71 	      struct rte_mbuf **pkts, unsigned int n)
72 {
73 	unsigned int pos;
74 	uintptr_t addr;
75 	const uint8x16_t dseg_shuf_m = {
76 		 3,  2,  1,  0, /* length, bswap32 */
77 		 4,  5,  6,  7, /* lkey */
78 		15, 14, 13, 12, /* addr, bswap64 */
79 		11, 10,  9,  8
80 	};
81 #ifdef MLX5_PMD_SOFT_COUNTERS
82 	uint32_t tx_byte = 0;
83 #endif
84 
85 	for (pos = 0; pos < n; ++pos, dseg += MLX5_WQE_DWORD_SIZE) {
86 		uint8x16_t desc;
87 		struct rte_mbuf *pkt = pkts[pos];
88 
89 		addr = rte_pktmbuf_mtod(pkt, uintptr_t);
90 		desc = vreinterpretq_u8_u32((uint32x4_t) {
91 				DATA_LEN(pkt),
92 				mlx5_tx_mb2mr(txq, pkt),
93 				addr,
94 				addr >> 32 });
95 		desc = vqtbl1q_u8(desc, dseg_shuf_m);
96 		vst1q_u8(dseg, desc);
97 #ifdef MLX5_PMD_SOFT_COUNTERS
98 		tx_byte += DATA_LEN(pkt);
99 #endif
100 	}
101 #ifdef MLX5_PMD_SOFT_COUNTERS
102 	txq->stats.obytes += tx_byte;
103 #endif
104 }
105 
106 /**
107  * Send multi-segmented packets until it encounters a single segment packet in
108  * the pkts list.
109  *
110  * @param txq
111  *   Pointer to TX queue structure.
112  * @param pkts
113  *   Pointer to array of packets to be sent.
114  * @param pkts_n
115  *   Number of packets to be sent.
116  *
117  * @return
118  *   Number of packets successfully transmitted (<= pkts_n).
119  */
120 static uint16_t
121 txq_scatter_v(struct mlx5_txq_data *txq, struct rte_mbuf **pkts,
122 	      uint16_t pkts_n)
123 {
124 	uint16_t elts_head = txq->elts_head;
125 	const uint16_t elts_n = 1 << txq->elts_n;
126 	const uint16_t elts_m = elts_n - 1;
127 	const uint16_t wq_n = 1 << txq->wqe_n;
128 	const uint16_t wq_mask = wq_n - 1;
129 	const unsigned int nb_dword_per_wqebb =
130 		MLX5_WQE_SIZE / MLX5_WQE_DWORD_SIZE;
131 	const unsigned int nb_dword_in_hdr =
132 		sizeof(struct mlx5_wqe) / MLX5_WQE_DWORD_SIZE;
133 	unsigned int n;
134 	volatile struct mlx5_wqe *wqe = NULL;
135 
136 	assert(elts_n > pkts_n);
137 	mlx5_tx_complete(txq);
138 	/* A CQE slot must always be available. */
139 	assert((1u << txq->cqe_n) - (txq->cq_pi - txq->cq_ci));
140 	if (unlikely(!pkts_n))
141 		return 0;
142 	for (n = 0; n < pkts_n; ++n) {
143 		struct rte_mbuf *buf = pkts[n];
144 		unsigned int segs_n = buf->nb_segs;
145 		unsigned int ds = nb_dword_in_hdr;
146 		unsigned int len = PKT_LEN(buf);
147 		uint16_t wqe_ci = txq->wqe_ci;
148 		const uint8x16_t ctrl_shuf_m = {
149 			3,  2,  1,  0, /* bswap32 */
150 			7,  6,  5,  4, /* bswap32 */
151 			11, 10,  9,  8, /* bswap32 */
152 			12, 13, 14, 15
153 		};
154 		uint8_t cs_flags;
155 		uint16_t max_elts;
156 		uint16_t max_wqe;
157 		uint8x16_t *t_wqe;
158 		uint8_t *dseg;
159 		uint8x16_t ctrl;
160 
161 		assert(segs_n);
162 		max_elts = elts_n - (elts_head - txq->elts_tail);
163 		max_wqe = wq_n - (txq->wqe_ci - txq->wqe_pi);
164 		/*
165 		 * A MPW session consumes 2 WQEs at most to
166 		 * include MLX5_MPW_DSEG_MAX pointers.
167 		 */
168 		if (segs_n == 1 ||
169 		    max_elts < segs_n || max_wqe < 2)
170 			break;
171 		wqe = &((volatile struct mlx5_wqe64 *)
172 			 txq->wqes)[wqe_ci & wq_mask].hdr;
173 		cs_flags = txq_ol_cksum_to_cs(txq, buf);
174 		/* Title WQEBB pointer. */
175 		t_wqe = (uint8x16_t *)wqe;
176 		dseg = (uint8_t *)(wqe + 1);
177 		do {
178 			if (!(ds++ % nb_dword_per_wqebb)) {
179 				dseg = (uint8_t *)
180 					&((volatile struct mlx5_wqe64 *)
181 					   txq->wqes)[++wqe_ci & wq_mask];
182 			}
183 			txq_wr_dseg_v(txq, dseg, &buf, 1);
184 			dseg += MLX5_WQE_DWORD_SIZE;
185 			(*txq->elts)[elts_head++ & elts_m] = buf;
186 			buf = buf->next;
187 		} while (--segs_n);
188 		++wqe_ci;
189 		/* Fill CTRL in the header. */
190 		ctrl = vreinterpretq_u8_u32((uint32x4_t) {
191 				MLX5_OPC_MOD_MPW << 24 |
192 				txq->wqe_ci << 8 | MLX5_OPCODE_TSO,
193 				txq->qp_num_8s | ds, 0, 0});
194 		ctrl = vqtbl1q_u8(ctrl, ctrl_shuf_m);
195 		vst1q_u8((void *)t_wqe, ctrl);
196 		/* Fill ESEG in the header. */
197 		vst1q_u16((void *)(t_wqe + 1),
198 			  (uint16x8_t) { 0, 0, cs_flags, rte_cpu_to_be_16(len),
199 					 0, 0, 0, 0 });
200 		txq->wqe_ci = wqe_ci;
201 	}
202 	if (!n)
203 		return 0;
204 	txq->elts_comp += (uint16_t)(elts_head - txq->elts_head);
205 	txq->elts_head = elts_head;
206 	if (txq->elts_comp >= MLX5_TX_COMP_THRESH) {
207 		wqe->ctrl[2] = rte_cpu_to_be_32(8);
208 		wqe->ctrl[3] = txq->elts_head;
209 		txq->elts_comp = 0;
210 #ifndef NDEBUG
211 		++txq->cq_pi;
212 #endif
213 	}
214 #ifdef MLX5_PMD_SOFT_COUNTERS
215 	txq->stats.opackets += n;
216 #endif
217 	mlx5_tx_dbrec(txq, wqe);
218 	return n;
219 }
220 
221 /**
222  * Send burst of packets with Enhanced MPW. If it encounters a multi-seg packet,
223  * it returns to make it processed by txq_scatter_v(). All the packets in
224  * the pkts list should be single segment packets having same offload flags.
225  * This must be checked by txq_count_contig_single_seg() and txq_calc_offload().
226  *
227  * @param txq
228  *   Pointer to TX queue structure.
229  * @param pkts
230  *   Pointer to array of packets to be sent.
231  * @param pkts_n
232  *   Number of packets to be sent (<= MLX5_VPMD_TX_MAX_BURST).
233  * @param cs_flags
234  *   Checksum offload flags to be written in the descriptor.
235  *
236  * @return
237  *   Number of packets successfully transmitted (<= pkts_n).
238  */
239 static inline uint16_t
240 txq_burst_v(struct mlx5_txq_data *txq, struct rte_mbuf **pkts, uint16_t pkts_n,
241 	    uint8_t cs_flags)
242 {
243 	struct rte_mbuf **elts;
244 	uint16_t elts_head = txq->elts_head;
245 	const uint16_t elts_n = 1 << txq->elts_n;
246 	const uint16_t elts_m = elts_n - 1;
247 	const unsigned int nb_dword_per_wqebb =
248 		MLX5_WQE_SIZE / MLX5_WQE_DWORD_SIZE;
249 	const unsigned int nb_dword_in_hdr =
250 		sizeof(struct mlx5_wqe) / MLX5_WQE_DWORD_SIZE;
251 	unsigned int n = 0;
252 	unsigned int pos;
253 	uint16_t max_elts;
254 	uint16_t max_wqe;
255 	uint32_t comp_req = 0;
256 	const uint16_t wq_n = 1 << txq->wqe_n;
257 	const uint16_t wq_mask = wq_n - 1;
258 	uint16_t wq_idx = txq->wqe_ci & wq_mask;
259 	volatile struct mlx5_wqe64 *wq =
260 		&((volatile struct mlx5_wqe64 *)txq->wqes)[wq_idx];
261 	volatile struct mlx5_wqe *wqe = (volatile struct mlx5_wqe *)wq;
262 	const uint8x16_t ctrl_shuf_m = {
263 		 3,  2,  1,  0, /* bswap32 */
264 		 7,  6,  5,  4, /* bswap32 */
265 		11, 10,  9,  8, /* bswap32 */
266 		12, 13, 14, 15
267 	};
268 	uint8x16_t *t_wqe;
269 	uint8_t *dseg;
270 	uint8x16_t ctrl;
271 
272 	/* Make sure all packets can fit into a single WQE. */
273 	assert(elts_n > pkts_n);
274 	mlx5_tx_complete(txq);
275 	max_elts = (elts_n - (elts_head - txq->elts_tail));
276 	/* A CQE slot must always be available. */
277 	assert((1u << txq->cqe_n) - (txq->cq_pi - txq->cq_ci));
278 	max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
279 	pkts_n = RTE_MIN((unsigned int)RTE_MIN(pkts_n, max_wqe), max_elts);
280 	if (unlikely(!pkts_n))
281 		return 0;
282 	elts = &(*txq->elts)[elts_head & elts_m];
283 	/* Loop for available tailroom first. */
284 	n = RTE_MIN(elts_n - (elts_head & elts_m), pkts_n);
285 	for (pos = 0; pos < (n & -2); pos += 2)
286 		vst1q_u64((void *)&elts[pos], vld1q_u64((void *)&pkts[pos]));
287 	if (n & 1)
288 		elts[pos] = pkts[pos];
289 	/* Check if it crosses the end of the queue. */
290 	if (unlikely(n < pkts_n)) {
291 		elts = &(*txq->elts)[0];
292 		for (pos = 0; pos < pkts_n - n; ++pos)
293 			elts[pos] = pkts[n + pos];
294 	}
295 	txq->elts_head += pkts_n;
296 	/* Save title WQEBB pointer. */
297 	t_wqe = (uint8x16_t *)wqe;
298 	dseg = (uint8_t *)(wqe + 1);
299 	/* Calculate the number of entries to the end. */
300 	n = RTE_MIN(
301 		(wq_n - wq_idx) * nb_dword_per_wqebb - nb_dword_in_hdr,
302 		pkts_n);
303 	/* Fill DSEGs. */
304 	txq_wr_dseg_v(txq, dseg, pkts, n);
305 	/* Check if it crosses the end of the queue. */
306 	if (n < pkts_n) {
307 		dseg = (uint8_t *)txq->wqes;
308 		txq_wr_dseg_v(txq, dseg, &pkts[n], pkts_n - n);
309 	}
310 	if (txq->elts_comp + pkts_n < MLX5_TX_COMP_THRESH) {
311 		txq->elts_comp += pkts_n;
312 	} else {
313 		/* Request a completion. */
314 		txq->elts_comp = 0;
315 #ifndef NDEBUG
316 		++txq->cq_pi;
317 #endif
318 		comp_req = 8;
319 	}
320 	/* Fill CTRL in the header. */
321 	ctrl = vreinterpretq_u8_u32((uint32x4_t) {
322 			MLX5_OPC_MOD_ENHANCED_MPSW << 24 |
323 			txq->wqe_ci << 8 | MLX5_OPCODE_ENHANCED_MPSW,
324 			txq->qp_num_8s | (pkts_n + 2),
325 			comp_req,
326 			txq->elts_head });
327 	ctrl = vqtbl1q_u8(ctrl, ctrl_shuf_m);
328 	vst1q_u8((void *)t_wqe, ctrl);
329 	/* Fill ESEG in the header. */
330 	vst1q_u8((void *)(t_wqe + 1),
331 		 (uint8x16_t) { 0, 0, 0, 0,
332 				cs_flags, 0, 0, 0,
333 				0, 0, 0, 0,
334 				0, 0, 0, 0 });
335 #ifdef MLX5_PMD_SOFT_COUNTERS
336 	txq->stats.opackets += pkts_n;
337 #endif
338 	txq->wqe_ci += (nb_dword_in_hdr + pkts_n + (nb_dword_per_wqebb - 1)) /
339 		       nb_dword_per_wqebb;
340 	/* Ring QP doorbell. */
341 	mlx5_tx_dbrec_cond_wmb(txq, wqe, pkts_n < MLX5_VPMD_TX_MAX_BURST);
342 	return pkts_n;
343 }
344 
345 /**
346  * Store free buffers to RX SW ring.
347  *
348  * @param rxq
349  *   Pointer to RX queue structure.
350  * @param pkts
351  *   Pointer to array of packets to be stored.
352  * @param pkts_n
353  *   Number of packets to be stored.
354  */
355 static inline void
356 rxq_copy_mbuf_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t n)
357 {
358 	const uint16_t q_mask = (1 << rxq->elts_n) - 1;
359 	struct rte_mbuf **elts = &(*rxq->elts)[rxq->rq_pi & q_mask];
360 	unsigned int pos;
361 	uint16_t p = n & -2;
362 
363 	for (pos = 0; pos < p; pos += 2) {
364 		uint64x2_t mbp;
365 
366 		mbp = vld1q_u64((void *)&elts[pos]);
367 		vst1q_u64((void *)&pkts[pos], mbp);
368 	}
369 	if (n & 1)
370 		pkts[pos] = elts[pos];
371 }
372 
373 /**
374  * Decompress a compressed completion and fill in mbufs in RX SW ring with data
375  * extracted from the title completion descriptor.
376  *
377  * @param rxq
378  *   Pointer to RX queue structure.
379  * @param cq
380  *   Pointer to completion array having a compressed completion at first.
381  * @param elts
382  *   Pointer to SW ring to be filled. The first mbuf has to be pre-built from
383  *   the title completion descriptor to be copied to the rest of mbufs.
384  */
385 static inline void
386 rxq_cq_decompress_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq,
387 		    struct rte_mbuf **elts)
388 {
389 	volatile struct mlx5_mini_cqe8 *mcq = (void *)&(cq + 1)->pkt_info;
390 	struct rte_mbuf *t_pkt = elts[0]; /* Title packet is pre-built. */
391 	unsigned int pos;
392 	unsigned int i;
393 	unsigned int inv = 0;
394 	/* Mask to shuffle from extracted mini CQE to mbuf. */
395 	const uint8x16_t mcqe_shuf_m1 = {
396 		-1, -1, -1, -1, /* skip packet_type */
397 		 7,  6, -1, -1, /* pkt_len, bswap16 */
398 		 7,  6,         /* data_len, bswap16 */
399 		-1, -1,         /* skip vlan_tci */
400 		 3,  2,  1,  0  /* hash.rss, bswap32 */
401 	};
402 	const uint8x16_t mcqe_shuf_m2 = {
403 		-1, -1, -1, -1, /* skip packet_type */
404 		15, 14, -1, -1, /* pkt_len, bswap16 */
405 		15, 14,         /* data_len, bswap16 */
406 		-1, -1,         /* skip vlan_tci */
407 		11, 10,  9,  8  /* hash.rss, bswap32 */
408 	};
409 	/* Restore the compressed count. Must be 16 bits. */
410 	const uint16_t mcqe_n = t_pkt->data_len +
411 				(rxq->crc_present * ETHER_CRC_LEN);
412 	const uint64x2_t rearm =
413 		vld1q_u64((void *)&t_pkt->rearm_data);
414 	const uint32x4_t rxdf_mask = {
415 		0xffffffff, /* packet_type */
416 		0,          /* skip pkt_len */
417 		0xffff0000, /* vlan_tci, skip data_len */
418 		0,          /* skip hash.rss */
419 	};
420 	const uint8x16_t rxdf =
421 		vandq_u8(vld1q_u8((void *)&t_pkt->rx_descriptor_fields1),
422 			 vreinterpretq_u8_u32(rxdf_mask));
423 	const uint16x8_t crc_adj = {
424 		0, 0,
425 		rxq->crc_present * ETHER_CRC_LEN, 0,
426 		rxq->crc_present * ETHER_CRC_LEN, 0,
427 		0, 0
428 	};
429 	const uint32_t flow_tag = t_pkt->hash.fdir.hi;
430 #ifdef MLX5_PMD_SOFT_COUNTERS
431 	uint32_t rcvd_byte = 0;
432 #endif
433 	/* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */
434 	const uint8x8_t len_shuf_m = {
435 		 7,  6,         /* 1st mCQE */
436 		15, 14,         /* 2nd mCQE */
437 		23, 22,         /* 3rd mCQE */
438 		31, 30          /* 4th mCQE */
439 	};
440 
441 	/*
442 	 * A. load mCQEs into a 128bit register.
443 	 * B. store rearm data to mbuf.
444 	 * C. combine data from mCQEs with rx_descriptor_fields1.
445 	 * D. store rx_descriptor_fields1.
446 	 * E. store flow tag (rte_flow mark).
447 	 */
448 	for (pos = 0; pos < mcqe_n; ) {
449 		uint8_t *p = (void *)&mcq[pos % 8];
450 		uint8_t *e0 = (void *)&elts[pos]->rearm_data;
451 		uint8_t *e1 = (void *)&elts[pos + 1]->rearm_data;
452 		uint8_t *e2 = (void *)&elts[pos + 2]->rearm_data;
453 		uint8_t *e3 = (void *)&elts[pos + 3]->rearm_data;
454 		uint16x4_t byte_cnt;
455 #ifdef MLX5_PMD_SOFT_COUNTERS
456 		uint16x4_t invalid_mask =
457 			vcreate_u16(mcqe_n - pos < MLX5_VPMD_DESCS_PER_LOOP ?
458 				    -1UL << ((mcqe_n - pos) *
459 					     sizeof(uint16_t) * 8) : 0);
460 #endif
461 
462 		if (!(pos & 0x7) && pos + 8 < mcqe_n)
463 			rte_prefetch0((void *)(cq + pos + 8));
464 		__asm__ volatile (
465 		/* A.1 load mCQEs into a 128bit register. */
466 		"ld1 {v16.16b - v17.16b}, [%[mcq]] \n\t"
467 		/* B.1 store rearm data to mbuf. */
468 		"st1 {%[rearm].2d}, [%[e0]] \n\t"
469 		"add %[e0], %[e0], #16 \n\t"
470 		"st1 {%[rearm].2d}, [%[e1]] \n\t"
471 		"add %[e1], %[e1], #16 \n\t"
472 		/* C.1 combine data from mCQEs with rx_descriptor_fields1. */
473 		"tbl v18.16b, {v16.16b}, %[mcqe_shuf_m1].16b \n\t"
474 		"tbl v19.16b, {v16.16b}, %[mcqe_shuf_m2].16b \n\t"
475 		"sub v18.8h, v18.8h, %[crc_adj].8h \n\t"
476 		"sub v19.8h, v19.8h, %[crc_adj].8h \n\t"
477 		"orr v18.16b, v18.16b, %[rxdf].16b \n\t"
478 		"orr v19.16b, v19.16b, %[rxdf].16b \n\t"
479 		/* D.1 store rx_descriptor_fields1. */
480 		"st1 {v18.2d}, [%[e0]] \n\t"
481 		"st1 {v19.2d}, [%[e1]] \n\t"
482 		/* B.1 store rearm data to mbuf. */
483 		"st1 {%[rearm].2d}, [%[e2]] \n\t"
484 		"add %[e2], %[e2], #16 \n\t"
485 		"st1 {%[rearm].2d}, [%[e3]] \n\t"
486 		"add %[e3], %[e3], #16 \n\t"
487 		/* C.1 combine data from mCQEs with rx_descriptor_fields1. */
488 		"tbl v18.16b, {v17.16b}, %[mcqe_shuf_m1].16b \n\t"
489 		"tbl v19.16b, {v17.16b}, %[mcqe_shuf_m2].16b \n\t"
490 		"sub v18.8h, v18.8h, %[crc_adj].8h \n\t"
491 		"sub v19.8h, v19.8h, %[crc_adj].8h \n\t"
492 		"orr v18.16b, v18.16b, %[rxdf].16b \n\t"
493 		"orr v19.16b, v19.16b, %[rxdf].16b \n\t"
494 		/* D.1 store rx_descriptor_fields1. */
495 		"st1 {v18.2d}, [%[e2]] \n\t"
496 		"st1 {v19.2d}, [%[e3]] \n\t"
497 #ifdef MLX5_PMD_SOFT_COUNTERS
498 		"tbl %[byte_cnt].8b, {v16.16b - v17.16b}, %[len_shuf_m].8b \n\t"
499 #endif
500 		:[byte_cnt]"=&w"(byte_cnt)
501 		:[mcq]"r"(p),
502 		 [rxdf]"w"(rxdf),
503 		 [rearm]"w"(rearm),
504 		 [e3]"r"(e3), [e2]"r"(e2), [e1]"r"(e1), [e0]"r"(e0),
505 		 [mcqe_shuf_m1]"w"(mcqe_shuf_m1),
506 		 [mcqe_shuf_m2]"w"(mcqe_shuf_m2),
507 		 [crc_adj]"w"(crc_adj),
508 		 [len_shuf_m]"w"(len_shuf_m)
509 		:"memory", "v16", "v17", "v18", "v19");
510 #ifdef MLX5_PMD_SOFT_COUNTERS
511 		byte_cnt = vbic_u16(byte_cnt, invalid_mask);
512 		rcvd_byte += vget_lane_u64(vpaddl_u32(vpaddl_u16(byte_cnt)), 0);
513 #endif
514 		if (rxq->mark) {
515 			/* E.1 store flow tag (rte_flow mark). */
516 			elts[pos]->hash.fdir.hi = flow_tag;
517 			elts[pos + 1]->hash.fdir.hi = flow_tag;
518 			elts[pos + 2]->hash.fdir.hi = flow_tag;
519 			elts[pos + 3]->hash.fdir.hi = flow_tag;
520 		}
521 		pos += MLX5_VPMD_DESCS_PER_LOOP;
522 		/* Move to next CQE and invalidate consumed CQEs. */
523 		if (!(pos & 0x7) && pos < mcqe_n) {
524 			mcq = (void *)&(cq + pos)->pkt_info;
525 			for (i = 0; i < 8; ++i)
526 				cq[inv++].op_own = MLX5_CQE_INVALIDATE;
527 		}
528 	}
529 	/* Invalidate the rest of CQEs. */
530 	for (; inv < mcqe_n; ++inv)
531 		cq[inv].op_own = MLX5_CQE_INVALIDATE;
532 #ifdef MLX5_PMD_SOFT_COUNTERS
533 	rxq->stats.ipackets += mcqe_n;
534 	rxq->stats.ibytes += rcvd_byte;
535 #endif
536 	rxq->cq_ci += mcqe_n;
537 }
538 
539 /**
540  * Calculate packet type and offload flag for mbuf and store it.
541  *
542  * @param rxq
543  *   Pointer to RX queue structure.
544  * @param ptype_info
545  *   Array of four 4bytes packet type info extracted from the original
546  *   completion descriptor.
547  * @param flow_tag
548  *   Array of four 4bytes flow ID extracted from the original completion
549  *   descriptor.
550  * @param op_err
551  *   Opcode vector having responder error status. Each field is 4B.
552  * @param pkts
553  *   Pointer to array of packets to be filled.
554  */
555 static inline void
556 rxq_cq_to_ptype_oflags_v(struct mlx5_rxq_data *rxq,
557 			 uint32x4_t ptype_info, uint32x4_t flow_tag,
558 			 uint16x4_t op_err, struct rte_mbuf **pkts)
559 {
560 	uint16x4_t ptype;
561 	uint32x4_t pinfo, cv_flags;
562 	uint32x4_t ol_flags =
563 		vdupq_n_u32(rxq->rss_hash * PKT_RX_RSS_HASH |
564 			    rxq->hw_timestamp * PKT_RX_TIMESTAMP);
565 	const uint32x4_t ptype_ol_mask = { 0x106, 0x106, 0x106, 0x106 };
566 	const uint8x16_t cv_flag_sel = {
567 		0,
568 		(uint8_t)(PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED),
569 		(uint8_t)(PKT_RX_IP_CKSUM_GOOD >> 1),
570 		0,
571 		(uint8_t)(PKT_RX_L4_CKSUM_GOOD >> 1),
572 		0,
573 		(uint8_t)((PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1),
574 		0, 0, 0, 0, 0, 0, 0, 0, 0
575 	};
576 	const uint32x4_t cv_mask =
577 		vdupq_n_u32(PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
578 			    PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED);
579 	const uint64x1_t mbuf_init = vld1_u64(&rxq->mbuf_initializer);
580 	const uint64x1_t r32_mask = vcreate_u64(0xffffffff);
581 	uint64x2_t rearm0, rearm1, rearm2, rearm3;
582 
583 	if (rxq->mark) {
584 		const uint32x4_t ft_def = vdupq_n_u32(MLX5_FLOW_MARK_DEFAULT);
585 		const uint32x4_t fdir_flags = vdupq_n_u32(PKT_RX_FDIR);
586 		uint32x4_t fdir_id_flags = vdupq_n_u32(PKT_RX_FDIR_ID);
587 		uint32x4_t invalid_mask;
588 
589 		/* Check if flow tag is non-zero then set PKT_RX_FDIR. */
590 		invalid_mask = vceqzq_u32(flow_tag);
591 		ol_flags = vorrq_u32(ol_flags,
592 				     vbicq_u32(fdir_flags, invalid_mask));
593 		/* Mask out invalid entries. */
594 		fdir_id_flags = vbicq_u32(fdir_id_flags, invalid_mask);
595 		/* Check if flow tag MLX5_FLOW_MARK_DEFAULT. */
596 		ol_flags = vorrq_u32(ol_flags,
597 				     vbicq_u32(fdir_id_flags,
598 					       vceqq_u32(flow_tag, ft_def)));
599 	}
600 	/*
601 	 * ptype_info has the following:
602 	 * bit[1]     = l3_ok
603 	 * bit[2]     = l4_ok
604 	 * bit[8]     = cv
605 	 * bit[11:10] = l3_hdr_type
606 	 * bit[14:12] = l4_hdr_type
607 	 * bit[15]    = ip_frag
608 	 * bit[16]    = tunneled
609 	 * bit[17]    = outer_l3_type
610 	 */
611 	ptype = vshrn_n_u32(ptype_info, 10);
612 	/* Errored packets will have RTE_PTYPE_ALL_MASK. */
613 	ptype = vorr_u16(ptype, op_err);
614 	pkts[0]->packet_type =
615 		mlx5_ptype_table[vget_lane_u8(vreinterpret_u8_u16(ptype), 6)];
616 	pkts[1]->packet_type =
617 		mlx5_ptype_table[vget_lane_u8(vreinterpret_u8_u16(ptype), 4)];
618 	pkts[2]->packet_type =
619 		mlx5_ptype_table[vget_lane_u8(vreinterpret_u8_u16(ptype), 2)];
620 	pkts[3]->packet_type =
621 		mlx5_ptype_table[vget_lane_u8(vreinterpret_u8_u16(ptype), 0)];
622 	/* Fill flags for checksum and VLAN. */
623 	pinfo = vandq_u32(ptype_info, ptype_ol_mask);
624 	pinfo = vreinterpretq_u32_u8(
625 		vqtbl1q_u8(cv_flag_sel, vreinterpretq_u8_u32(pinfo)));
626 	/* Locate checksum flags at byte[2:1] and merge with VLAN flags. */
627 	cv_flags = vshlq_n_u32(pinfo, 9);
628 	cv_flags = vorrq_u32(pinfo, cv_flags);
629 	/* Move back flags to start from byte[0]. */
630 	cv_flags = vshrq_n_u32(cv_flags, 8);
631 	/* Mask out garbage bits. */
632 	cv_flags = vandq_u32(cv_flags, cv_mask);
633 	/* Merge to ol_flags. */
634 	ol_flags = vorrq_u32(ol_flags, cv_flags);
635 	/* Merge mbuf_init and ol_flags, and store. */
636 	rearm0 = vcombine_u64(mbuf_init,
637 			      vshr_n_u64(vget_high_u64(vreinterpretq_u64_u32(
638 						       ol_flags)), 32));
639 	rearm1 = vcombine_u64(mbuf_init,
640 			      vand_u64(vget_high_u64(vreinterpretq_u64_u32(
641 						     ol_flags)), r32_mask));
642 	rearm2 = vcombine_u64(mbuf_init,
643 			      vshr_n_u64(vget_low_u64(vreinterpretq_u64_u32(
644 						      ol_flags)), 32));
645 	rearm3 = vcombine_u64(mbuf_init,
646 			      vand_u64(vget_low_u64(vreinterpretq_u64_u32(
647 						    ol_flags)), r32_mask));
648 	vst1q_u64((void *)&pkts[0]->rearm_data, rearm0);
649 	vst1q_u64((void *)&pkts[1]->rearm_data, rearm1);
650 	vst1q_u64((void *)&pkts[2]->rearm_data, rearm2);
651 	vst1q_u64((void *)&pkts[3]->rearm_data, rearm3);
652 }
653 
654 /**
655  * Receive burst of packets. An errored completion also consumes a mbuf, but the
656  * packet_type is set to be RTE_PTYPE_ALL_MASK. Marked mbufs should be freed
657  * before returning to application.
658  *
659  * @param rxq
660  *   Pointer to RX queue structure.
661  * @param[out] pkts
662  *   Array to store received packets.
663  * @param pkts_n
664  *   Maximum number of packets in array.
665  * @param[out] err
666  *   Pointer to a flag. Set non-zero value if pkts array has at least one error
667  *   packet to handle.
668  *
669  * @return
670  *   Number of packets received including errors (<= pkts_n).
671  */
672 static inline uint16_t
673 rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
674 	    uint64_t *err)
675 {
676 	const uint16_t q_n = 1 << rxq->cqe_n;
677 	const uint16_t q_mask = q_n - 1;
678 	volatile struct mlx5_cqe *cq;
679 	struct rte_mbuf **elts;
680 	unsigned int pos;
681 	uint64_t n;
682 	uint16_t repl_n;
683 	uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
684 	uint16_t nocmp_n = 0;
685 	uint16_t rcvd_pkt = 0;
686 	unsigned int cq_idx = rxq->cq_ci & q_mask;
687 	unsigned int elts_idx;
688 	const uint16x4_t ownership = vdup_n_u16(!(rxq->cq_ci & (q_mask + 1)));
689 	const uint16x4_t owner_check = vcreate_u16(0x0001000100010001);
690 	const uint16x4_t opcode_check = vcreate_u16(0x00f000f000f000f0);
691 	const uint16x4_t format_check = vcreate_u16(0x000c000c000c000c);
692 	const uint16x4_t resp_err_check = vcreate_u16(0x00e000e000e000e0);
693 #ifdef MLX5_PMD_SOFT_COUNTERS
694 	uint32_t rcvd_byte = 0;
695 #endif
696 	/* Mask to generate 16B length vector. */
697 	const uint8x8_t len_shuf_m = {
698 		52, 53,         /* 4th CQE */
699 		36, 37,         /* 3rd CQE */
700 		20, 21,         /* 2nd CQE */
701 		 4,  5          /* 1st CQE */
702 	};
703 	/* Mask to extract 16B data from a 64B CQE. */
704 	const uint8x16_t cqe_shuf_m = {
705 		28, 29,         /* hdr_type_etc */
706 		 0,             /* pkt_info */
707 		-1,             /* null */
708 		47, 46,         /* byte_cnt, bswap16 */
709 		31, 30,         /* vlan_info, bswap16 */
710 		15, 14, 13, 12, /* rx_hash_res, bswap32 */
711 		57, 58, 59,     /* flow_tag */
712 		63              /* op_own */
713 	};
714 	/* Mask to generate 16B data for mbuf. */
715 	const uint8x16_t mb_shuf_m = {
716 		 4,  5, -1, -1, /* pkt_len */
717 		 4,  5,         /* data_len */
718 		 6,  7,         /* vlan_tci */
719 		 8,  9, 10, 11, /* hash.rss */
720 		12, 13, 14, -1  /* hash.fdir.hi */
721 	};
722 	/* Mask to generate 16B owner vector. */
723 	const uint8x8_t owner_shuf_m = {
724 		63, -1,         /* 4th CQE */
725 		47, -1,         /* 3rd CQE */
726 		31, -1,         /* 2nd CQE */
727 		15, -1          /* 1st CQE */
728 	};
729 	/* Mask to generate a vector having packet_type/ol_flags. */
730 	const uint8x16_t ptype_shuf_m = {
731 		48, 49, 50, -1, /* 4th CQE */
732 		32, 33, 34, -1, /* 3rd CQE */
733 		16, 17, 18, -1, /* 2nd CQE */
734 		 0,  1,  2, -1  /* 1st CQE */
735 	};
736 	/* Mask to generate a vector having flow tags. */
737 	const uint8x16_t ftag_shuf_m = {
738 		60, 61, 62, -1, /* 4th CQE */
739 		44, 45, 46, -1, /* 3rd CQE */
740 		28, 29, 30, -1, /* 2nd CQE */
741 		12, 13, 14, -1  /* 1st CQE */
742 	};
743 	const uint16x8_t crc_adj = {
744 		0, 0, rxq->crc_present * ETHER_CRC_LEN, 0, 0, 0, 0, 0
745 	};
746 	const uint32x4_t flow_mark_adj = { 0, 0, 0, rxq->mark * (-1) };
747 
748 	assert(rxq->sges_n == 0);
749 	assert(rxq->cqe_n == rxq->elts_n);
750 	cq = &(*rxq->cqes)[cq_idx];
751 	rte_prefetch_non_temporal(cq);
752 	rte_prefetch_non_temporal(cq + 1);
753 	rte_prefetch_non_temporal(cq + 2);
754 	rte_prefetch_non_temporal(cq + 3);
755 	pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
756 	/*
757 	 * Order of indexes:
758 	 *   rq_ci >= cq_ci >= rq_pi
759 	 * Definition of indexes:
760 	 *   rq_ci - cq_ci := # of buffers owned by HW (posted).
761 	 *   cq_ci - rq_pi := # of buffers not returned to app (decompressed).
762 	 *   N - (rq_ci - rq_pi) := # of buffers consumed (to be replenished).
763 	 */
764 	repl_n = q_n - (rxq->rq_ci - rxq->rq_pi);
765 	if (repl_n >= MLX5_VPMD_RXQ_RPLNSH_THRESH)
766 		mlx5_rx_replenish_bulk_mbuf(rxq, repl_n);
767 	/* See if there're unreturned mbufs from compressed CQE. */
768 	rcvd_pkt = rxq->cq_ci - rxq->rq_pi;
769 	if (rcvd_pkt > 0) {
770 		rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
771 		rxq_copy_mbuf_v(rxq, pkts, rcvd_pkt);
772 		rxq->rq_pi += rcvd_pkt;
773 		pkts += rcvd_pkt;
774 	}
775 	elts_idx = rxq->rq_pi & q_mask;
776 	elts = &(*rxq->elts)[elts_idx];
777 	/* Not to overflow pkts array. */
778 	pkts_n = RTE_ALIGN_FLOOR(pkts_n - rcvd_pkt, MLX5_VPMD_DESCS_PER_LOOP);
779 	/* Not to cross queue end. */
780 	pkts_n = RTE_MIN(pkts_n, q_n - elts_idx);
781 	if (!pkts_n)
782 		return rcvd_pkt;
783 	/* At this point, there shouldn't be any remained packets. */
784 	assert(rxq->rq_pi == rxq->cq_ci);
785 	/*
786 	 * Note that vectors have reverse order - {v3, v2, v1, v0}, because
787 	 * there's no instruction to count trailing zeros. __builtin_clzl() is
788 	 * used instead.
789 	 *
790 	 * A. copy 4 mbuf pointers from elts ring to returing pkts.
791 	 * B. load 64B CQE and extract necessary fields
792 	 *    Final 16bytes cqes[] extracted from original 64bytes CQE has the
793 	 *    following structure:
794 	 *        struct {
795 	 *          uint16_t hdr_type_etc;
796 	 *          uint8_t  pkt_info;
797 	 *          uint8_t  rsvd;
798 	 *          uint16_t byte_cnt;
799 	 *          uint16_t vlan_info;
800 	 *          uint32_t rx_has_res;
801 	 *          uint8_t  flow_tag[3];
802 	 *          uint8_t  op_own;
803 	 *        } c;
804 	 * C. fill in mbuf.
805 	 * D. get valid CQEs.
806 	 * E. find compressed CQE.
807 	 */
808 	for (pos = 0;
809 	     pos < pkts_n;
810 	     pos += MLX5_VPMD_DESCS_PER_LOOP) {
811 		uint16x4_t op_own;
812 		uint16x4_t opcode, owner_mask, invalid_mask;
813 		uint16x4_t comp_mask;
814 		uint16x4_t mask;
815 		uint16x4_t byte_cnt;
816 		uint32x4_t ptype_info, flow_tag;
817 		register uint64x2_t c0, c1, c2, c3;
818 		uint8_t *p0, *p1, *p2, *p3;
819 		uint8_t *e0 = (void *)&elts[pos]->pkt_len;
820 		uint8_t *e1 = (void *)&elts[pos + 1]->pkt_len;
821 		uint8_t *e2 = (void *)&elts[pos + 2]->pkt_len;
822 		uint8_t *e3 = (void *)&elts[pos + 3]->pkt_len;
823 		void *elts_p = (void *)&elts[pos];
824 		void *pkts_p = (void *)&pkts[pos];
825 
826 		/* A.0 do not cross the end of CQ. */
827 		mask = vcreate_u16(pkts_n - pos < MLX5_VPMD_DESCS_PER_LOOP ?
828 				   -1UL >> ((pkts_n - pos) *
829 					    sizeof(uint16_t) * 8) : 0);
830 		p0 = (void *)&cq[pos].pkt_info;
831 		p1 = p0 + (pkts_n - pos > 1) * sizeof(struct mlx5_cqe);
832 		p2 = p1 + (pkts_n - pos > 2) * sizeof(struct mlx5_cqe);
833 		p3 = p2 + (pkts_n - pos > 3) * sizeof(struct mlx5_cqe);
834 		/* B.0 (CQE 3) load a block having op_own. */
835 		c3 = vld1q_u64((uint64_t *)(p3 + 48));
836 		/* B.0 (CQE 2) load a block having op_own. */
837 		c2 = vld1q_u64((uint64_t *)(p2 + 48));
838 		/* B.0 (CQE 1) load a block having op_own. */
839 		c1 = vld1q_u64((uint64_t *)(p1 + 48));
840 		/* B.0 (CQE 0) load a block having op_own. */
841 		c0 = vld1q_u64((uint64_t *)(p0 + 48));
842 		/* Synchronize for loading the rest of blocks. */
843 		rte_cio_rmb();
844 		/* Prefetch next 4 CQEs. */
845 		if (pkts_n - pos >= 2 * MLX5_VPMD_DESCS_PER_LOOP) {
846 			unsigned int next = pos + MLX5_VPMD_DESCS_PER_LOOP;
847 			rte_prefetch_non_temporal(&cq[next]);
848 			rte_prefetch_non_temporal(&cq[next + 1]);
849 			rte_prefetch_non_temporal(&cq[next + 2]);
850 			rte_prefetch_non_temporal(&cq[next + 3]);
851 		}
852 		__asm__ volatile (
853 		/* B.1 (CQE 3) load the rest of blocks. */
854 		"ld1 {v16.16b - v18.16b}, [%[p3]] \n\t"
855 		/* B.2 (CQE 3) move the block having op_own. */
856 		"mov v19.16b, %[c3].16b \n\t"
857 		/* B.3 (CQE 3) extract 16B fields. */
858 		"tbl v23.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
859 		/* B.1 (CQE 2) load the rest of blocks. */
860 		"ld1 {v16.16b - v18.16b}, [%[p2]] \n\t"
861 		/* B.4 (CQE 3) adjust CRC length. */
862 		"sub v23.8h, v23.8h, %[crc_adj].8h \n\t"
863 		/* C.1 (CQE 3) generate final structure for mbuf. */
864 		"tbl v15.16b, {v23.16b}, %[mb_shuf_m].16b \n\t"
865 		/* B.2 (CQE 2) move the block having op_own. */
866 		"mov v19.16b, %[c2].16b \n\t"
867 		/* B.3 (CQE 2) extract 16B fields. */
868 		"tbl v22.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
869 		/* B.1 (CQE 1) load the rest of blocks. */
870 		"ld1 {v16.16b - v18.16b}, [%[p1]] \n\t"
871 		/* B.4 (CQE 2) adjust CRC length. */
872 		"sub v22.8h, v22.8h, %[crc_adj].8h \n\t"
873 		/* C.1 (CQE 2) generate final structure for mbuf. */
874 		"tbl v14.16b, {v22.16b}, %[mb_shuf_m].16b \n\t"
875 		/* B.2 (CQE 1) move the block having op_own. */
876 		"mov v19.16b, %[c1].16b \n\t"
877 		/* B.3 (CQE 1) extract 16B fields. */
878 		"tbl v21.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
879 		/* B.1 (CQE 0) load the rest of blocks. */
880 		"ld1 {v16.16b - v18.16b}, [%[p0]] \n\t"
881 		/* B.4 (CQE 1) adjust CRC length. */
882 		"sub v21.8h, v21.8h, %[crc_adj].8h \n\t"
883 		/* C.1 (CQE 1) generate final structure for mbuf. */
884 		"tbl v13.16b, {v21.16b}, %[mb_shuf_m].16b \n\t"
885 		/* B.2 (CQE 0) move the block having op_own. */
886 		"mov v19.16b, %[c0].16b \n\t"
887 		/* A.1 load mbuf pointers. */
888 		"ld1 {v24.2d - v25.2d}, [%[elts_p]] \n\t"
889 		/* B.3 (CQE 0) extract 16B fields. */
890 		"tbl v20.16b, {v16.16b - v19.16b}, %[cqe_shuf_m].16b \n\t"
891 		/* B.4 (CQE 0) adjust CRC length. */
892 		"sub v20.8h, v20.8h, %[crc_adj].8h \n\t"
893 		/* D.1 extract op_own byte. */
894 		"tbl %[op_own].8b, {v20.16b - v23.16b}, %[owner_shuf_m].8b \n\t"
895 		/* C.2 (CQE 3) adjust flow mark. */
896 		"add v15.4s, v15.4s, %[flow_mark_adj].4s \n\t"
897 		/* C.3 (CQE 3) fill in mbuf - rx_descriptor_fields1. */
898 		"st1 {v15.2d}, [%[e3]] \n\t"
899 		/* C.2 (CQE 2) adjust flow mark. */
900 		"add v14.4s, v14.4s, %[flow_mark_adj].4s \n\t"
901 		/* C.3 (CQE 2) fill in mbuf - rx_descriptor_fields1. */
902 		"st1 {v14.2d}, [%[e2]] \n\t"
903 		/* C.1 (CQE 0) generate final structure for mbuf. */
904 		"tbl v12.16b, {v20.16b}, %[mb_shuf_m].16b \n\t"
905 		/* C.2 (CQE 1) adjust flow mark. */
906 		"add v13.4s, v13.4s, %[flow_mark_adj].4s \n\t"
907 		/* C.3 (CQE 1) fill in mbuf - rx_descriptor_fields1. */
908 		"st1 {v13.2d}, [%[e1]] \n\t"
909 #ifdef MLX5_PMD_SOFT_COUNTERS
910 		/* Extract byte_cnt. */
911 		"tbl %[byte_cnt].8b, {v20.16b - v23.16b}, %[len_shuf_m].8b \n\t"
912 #endif
913 		/* Extract ptype_info. */
914 		"tbl %[ptype_info].16b, {v20.16b - v23.16b}, %[ptype_shuf_m].16b \n\t"
915 		/* Extract flow_tag. */
916 		"tbl %[flow_tag].16b, {v20.16b - v23.16b}, %[ftag_shuf_m].16b \n\t"
917 		/* A.2 copy mbuf pointers. */
918 		"st1 {v24.2d - v25.2d}, [%[pkts_p]] \n\t"
919 		/* C.2 (CQE 0) adjust flow mark. */
920 		"add v12.4s, v12.4s, %[flow_mark_adj].4s \n\t"
921 		/* C.3 (CQE 1) fill in mbuf - rx_descriptor_fields1. */
922 		"st1 {v12.2d}, [%[e0]] \n\t"
923 		:[op_own]"=&w"(op_own),
924 		 [byte_cnt]"=&w"(byte_cnt),
925 		 [ptype_info]"=&w"(ptype_info),
926 		 [flow_tag]"=&w"(flow_tag)
927 		:[p3]"r"(p3), [p2]"r"(p2), [p1]"r"(p1), [p0]"r"(p0),
928 		 [e3]"r"(e3), [e2]"r"(e2), [e1]"r"(e1), [e0]"r"(e0),
929 		 [c3]"w"(c3), [c2]"w"(c2), [c1]"w"(c1), [c0]"w"(c0),
930 		 [elts_p]"r"(elts_p),
931 		 [pkts_p]"r"(pkts_p),
932 		 [cqe_shuf_m]"w"(cqe_shuf_m),
933 		 [mb_shuf_m]"w"(mb_shuf_m),
934 		 [owner_shuf_m]"w"(owner_shuf_m),
935 		 [len_shuf_m]"w"(len_shuf_m),
936 		 [ptype_shuf_m]"w"(ptype_shuf_m),
937 		 [ftag_shuf_m]"w"(ftag_shuf_m),
938 		 [crc_adj]"w"(crc_adj),
939 		 [flow_mark_adj]"w"(flow_mark_adj)
940 		:"memory",
941 		 "v12", "v13", "v14", "v15",
942 		 "v16", "v17", "v18", "v19",
943 		 "v20", "v21", "v22", "v23",
944 		 "v24", "v25");
945 		/* D.2 flip owner bit to mark CQEs from last round. */
946 		owner_mask = vand_u16(op_own, owner_check);
947 		owner_mask = vceq_u16(owner_mask, ownership);
948 		/* D.3 get mask for invalidated CQEs. */
949 		opcode = vand_u16(op_own, opcode_check);
950 		invalid_mask = vceq_u16(opcode_check, opcode);
951 		/* E.1 find compressed CQE format. */
952 		comp_mask = vand_u16(op_own, format_check);
953 		comp_mask = vceq_u16(comp_mask, format_check);
954 		/* D.4 mask out beyond boundary. */
955 		invalid_mask = vorr_u16(invalid_mask, mask);
956 		/* D.5 merge invalid_mask with invalid owner. */
957 		invalid_mask = vorr_u16(invalid_mask, owner_mask);
958 		/* E.2 mask out invalid entries. */
959 		comp_mask = vbic_u16(comp_mask, invalid_mask);
960 		/* E.3 get the first compressed CQE. */
961 		comp_idx = __builtin_clzl(vget_lane_u64(vreinterpret_u64_u16(
962 					  comp_mask), 0)) /
963 					  (sizeof(uint16_t) * 8);
964 		/* D.6 mask out entries after the compressed CQE. */
965 		mask = vcreate_u16(comp_idx < MLX5_VPMD_DESCS_PER_LOOP ?
966 				   -1UL >> (comp_idx * sizeof(uint16_t) * 8) :
967 				   0);
968 		invalid_mask = vorr_u16(invalid_mask, mask);
969 		/* D.7 count non-compressed valid CQEs. */
970 		n = __builtin_clzl(vget_lane_u64(vreinterpret_u64_u16(
971 				   invalid_mask), 0)) / (sizeof(uint16_t) * 8);
972 		nocmp_n += n;
973 		/* D.2 get the final invalid mask. */
974 		mask = vcreate_u16(n < MLX5_VPMD_DESCS_PER_LOOP ?
975 				   -1UL >> (n * sizeof(uint16_t) * 8) : 0);
976 		invalid_mask = vorr_u16(invalid_mask, mask);
977 		/* D.3 check error in opcode. */
978 		opcode = vceq_u16(resp_err_check, opcode);
979 		opcode = vbic_u16(opcode, invalid_mask);
980 		/* D.4 mark if any error is set */
981 		*err |= vget_lane_u64(vreinterpret_u64_u16(opcode), 0);
982 		/* C.4 fill in mbuf - rearm_data and packet_type. */
983 		rxq_cq_to_ptype_oflags_v(rxq, ptype_info, flow_tag,
984 					 opcode, &elts[pos]);
985 		if (rxq->hw_timestamp) {
986 			elts[pos]->timestamp =
987 				rte_be_to_cpu_64(
988 					container_of(p0, struct mlx5_cqe,
989 						     pkt_info)->timestamp);
990 			elts[pos + 1]->timestamp =
991 				rte_be_to_cpu_64(
992 					container_of(p1, struct mlx5_cqe,
993 						     pkt_info)->timestamp);
994 			elts[pos + 2]->timestamp =
995 				rte_be_to_cpu_64(
996 					container_of(p2, struct mlx5_cqe,
997 						     pkt_info)->timestamp);
998 			elts[pos + 3]->timestamp =
999 				rte_be_to_cpu_64(
1000 					container_of(p3, struct mlx5_cqe,
1001 						     pkt_info)->timestamp);
1002 		}
1003 #ifdef MLX5_PMD_SOFT_COUNTERS
1004 		/* Add up received bytes count. */
1005 		byte_cnt = vbic_u16(byte_cnt, invalid_mask);
1006 		rcvd_byte += vget_lane_u64(vpaddl_u32(vpaddl_u16(byte_cnt)), 0);
1007 #endif
1008 		/*
1009 		 * Break the loop unless more valid CQE is expected, or if
1010 		 * there's a compressed CQE.
1011 		 */
1012 		if (n != MLX5_VPMD_DESCS_PER_LOOP)
1013 			break;
1014 	}
1015 	/* If no new CQE seen, return without updating cq_db. */
1016 	if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP))
1017 		return rcvd_pkt;
1018 	/* Update the consumer indexes for non-compressed CQEs. */
1019 	assert(nocmp_n <= pkts_n);
1020 	rxq->cq_ci += nocmp_n;
1021 	rxq->rq_pi += nocmp_n;
1022 	rcvd_pkt += nocmp_n;
1023 #ifdef MLX5_PMD_SOFT_COUNTERS
1024 	rxq->stats.ipackets += nocmp_n;
1025 	rxq->stats.ibytes += rcvd_byte;
1026 #endif
1027 	/* Decompress the last CQE if compressed. */
1028 	if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP && comp_idx == n) {
1029 		assert(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
1030 		rxq_cq_decompress_v(rxq, &cq[nocmp_n], &elts[nocmp_n]);
1031 		/* Return more packets if needed. */
1032 		if (nocmp_n < pkts_n) {
1033 			uint16_t n = rxq->cq_ci - rxq->rq_pi;
1034 
1035 			n = RTE_MIN(n, pkts_n - nocmp_n);
1036 			rxq_copy_mbuf_v(rxq, &pkts[nocmp_n], n);
1037 			rxq->rq_pi += n;
1038 			rcvd_pkt += n;
1039 		}
1040 	}
1041 	rte_compiler_barrier();
1042 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1043 	return rcvd_pkt;
1044 }
1045 
1046 #endif /* RTE_PMD_MLX5_RXTX_VEC_NEON_H_ */
1047