xref: /dpdk/drivers/net/mlx5/mlx5_rxtx_vec_neon.h (revision 200bc52e5aa0d72e70464c9cd22b55cf536ed13c)
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, 0, 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 					 cs_flags << 16 | rte_cpu_to_be_16(len),
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(8);
186 		wqe->ctrl[3] = txq->elts_head;
187 		txq->elts_comp = 0;
188 	}
189 #ifdef MLX5_PMD_SOFT_COUNTERS
190 	txq->stats.opackets += n;
191 #endif
192 	mlx5_tx_dbrec(txq, wqe);
193 	return n;
194 }
195 
196 /**
197  * Send burst of packets with Enhanced MPW. If it encounters a multi-seg packet,
198  * it returns to make it processed by txq_scatter_v(). All the packets in
199  * the pkts list should be single segment packets having same offload flags.
200  * This must be checked by txq_count_contig_single_seg() and txq_calc_offload().
201  *
202  * @param txq
203  *   Pointer to TX queue structure.
204  * @param pkts
205  *   Pointer to array of packets to be sent.
206  * @param pkts_n
207  *   Number of packets to be sent (<= MLX5_VPMD_TX_MAX_BURST).
208  * @param cs_flags
209  *   Checksum offload flags to be written in the descriptor.
210  * @param metadata
211  *   Metadata value to be written in the descriptor.
212  *
213  * @return
214  *   Number of packets successfully transmitted (<= pkts_n).
215  */
216 static inline uint16_t
217 txq_burst_v(struct mlx5_txq_data *txq, struct rte_mbuf **pkts, uint16_t pkts_n,
218 	    uint8_t cs_flags, rte_be32_t metadata)
219 {
220 	struct rte_mbuf **elts;
221 	uint16_t elts_head = txq->elts_head;
222 	const uint16_t elts_n = 1 << txq->elts_n;
223 	const uint16_t elts_m = elts_n - 1;
224 	const unsigned int nb_dword_per_wqebb =
225 		MLX5_WQE_SIZE / MLX5_WQE_DWORD_SIZE;
226 	const unsigned int nb_dword_in_hdr =
227 		sizeof(struct mlx5_wqe) / MLX5_WQE_DWORD_SIZE;
228 	unsigned int n = 0;
229 	unsigned int pos;
230 	uint16_t max_elts;
231 	uint16_t max_wqe;
232 	uint32_t comp_req = 0;
233 	const uint16_t wq_n = 1 << txq->wqe_n;
234 	const uint16_t wq_mask = wq_n - 1;
235 	uint16_t wq_idx = txq->wqe_ci & wq_mask;
236 	volatile struct mlx5_wqe64 *wq =
237 		&((volatile struct mlx5_wqe64 *)txq->wqes)[wq_idx];
238 	volatile struct mlx5_wqe *wqe = (volatile struct mlx5_wqe *)wq;
239 	const uint8x16_t ctrl_shuf_m = {
240 		 3,  2,  1,  0, /* bswap32 */
241 		 7,  6,  5,  4, /* bswap32 */
242 		11, 10,  9,  8, /* bswap32 */
243 		12, 13, 14, 15
244 	};
245 	uint8x16_t *t_wqe;
246 	uint8_t *dseg;
247 	uint8x16_t ctrl;
248 
249 	/* Make sure all packets can fit into a single WQE. */
250 	assert(elts_n > pkts_n);
251 	mlx5_tx_complete(txq);
252 	max_elts = (elts_n - (elts_head - txq->elts_tail));
253 	max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
254 	pkts_n = RTE_MIN((unsigned int)RTE_MIN(pkts_n, max_wqe), max_elts);
255 	if (unlikely(!pkts_n))
256 		return 0;
257 	elts = &(*txq->elts)[elts_head & elts_m];
258 	/* Loop for available tailroom first. */
259 	n = RTE_MIN(elts_n - (elts_head & elts_m), pkts_n);
260 	for (pos = 0; pos < (n & -2); pos += 2)
261 		vst1q_u64((void *)&elts[pos], vld1q_u64((void *)&pkts[pos]));
262 	if (n & 1)
263 		elts[pos] = pkts[pos];
264 	/* Check if it crosses the end of the queue. */
265 	if (unlikely(n < pkts_n)) {
266 		elts = &(*txq->elts)[0];
267 		for (pos = 0; pos < pkts_n - n; ++pos)
268 			elts[pos] = pkts[n + pos];
269 	}
270 	txq->elts_head += pkts_n;
271 	/* Save title WQEBB pointer. */
272 	t_wqe = (uint8x16_t *)wqe;
273 	dseg = (uint8_t *)(wqe + 1);
274 	/* Calculate the number of entries to the end. */
275 	n = RTE_MIN(
276 		(wq_n - wq_idx) * nb_dword_per_wqebb - nb_dword_in_hdr,
277 		pkts_n);
278 	/* Fill DSEGs. */
279 	txq_wr_dseg_v(txq, dseg, pkts, n);
280 	/* Check if it crosses the end of the queue. */
281 	if (n < pkts_n) {
282 		dseg = (uint8_t *)txq->wqes;
283 		txq_wr_dseg_v(txq, dseg, &pkts[n], pkts_n - n);
284 	}
285 	if (txq->elts_comp + pkts_n < MLX5_TX_COMP_THRESH) {
286 		txq->elts_comp += pkts_n;
287 	} else {
288 		/* A CQE slot must always be available. */
289 		assert((1u << txq->cqe_n) - (txq->cq_pi++ - txq->cq_ci));
290 		/* Request a completion. */
291 		txq->elts_comp = 0;
292 		comp_req = 8;
293 	}
294 	/* Fill CTRL in the header. */
295 	ctrl = vreinterpretq_u8_u32((uint32x4_t) {
296 			MLX5_OPC_MOD_ENHANCED_MPSW << 24 |
297 			txq->wqe_ci << 8 | MLX5_OPCODE_ENHANCED_MPSW,
298 			txq->qp_num_8s | (pkts_n + 2),
299 			comp_req,
300 			txq->elts_head });
301 	ctrl = vqtbl1q_u8(ctrl, ctrl_shuf_m);
302 	vst1q_u8((void *)t_wqe, ctrl);
303 	/* Fill ESEG in the header. */
304 	vst1q_u32((void *)(t_wqe + 1),
305 		 ((uint32x4_t) { 0, cs_flags, metadata, 0 }));
306 #ifdef MLX5_PMD_SOFT_COUNTERS
307 	txq->stats.opackets += pkts_n;
308 #endif
309 	txq->wqe_ci += (nb_dword_in_hdr + pkts_n + (nb_dword_per_wqebb - 1)) /
310 		       nb_dword_per_wqebb;
311 	/* Ring QP doorbell. */
312 	mlx5_tx_dbrec_cond_wmb(txq, wqe, pkts_n < MLX5_VPMD_TX_MAX_BURST);
313 	return pkts_n;
314 }
315 
316 /**
317  * Store free buffers to RX SW ring.
318  *
319  * @param rxq
320  *   Pointer to RX queue structure.
321  * @param pkts
322  *   Pointer to array of packets to be stored.
323  * @param pkts_n
324  *   Number of packets to be stored.
325  */
326 static inline void
327 rxq_copy_mbuf_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t n)
328 {
329 	const uint16_t q_mask = (1 << rxq->elts_n) - 1;
330 	struct rte_mbuf **elts = &(*rxq->elts)[rxq->rq_pi & q_mask];
331 	unsigned int pos;
332 	uint16_t p = n & -2;
333 
334 	for (pos = 0; pos < p; pos += 2) {
335 		uint64x2_t mbp;
336 
337 		mbp = vld1q_u64((void *)&elts[pos]);
338 		vst1q_u64((void *)&pkts[pos], mbp);
339 	}
340 	if (n & 1)
341 		pkts[pos] = elts[pos];
342 }
343 
344 /**
345  * Decompress a compressed completion and fill in mbufs in RX SW ring with data
346  * extracted from the title completion descriptor.
347  *
348  * @param rxq
349  *   Pointer to RX queue structure.
350  * @param cq
351  *   Pointer to completion array having a compressed completion at first.
352  * @param elts
353  *   Pointer to SW ring to be filled. The first mbuf has to be pre-built from
354  *   the title completion descriptor to be copied to the rest of mbufs.
355  */
356 static inline void
357 rxq_cq_decompress_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq,
358 		    struct rte_mbuf **elts)
359 {
360 	volatile struct mlx5_mini_cqe8 *mcq = (void *)&(cq + 1)->pkt_info;
361 	struct rte_mbuf *t_pkt = elts[0]; /* Title packet is pre-built. */
362 	unsigned int pos;
363 	unsigned int i;
364 	unsigned int inv = 0;
365 	/* Mask to shuffle from extracted mini CQE to mbuf. */
366 	const uint8x16_t mcqe_shuf_m1 = {
367 		-1, -1, -1, -1, /* skip packet_type */
368 		 7,  6, -1, -1, /* pkt_len, bswap16 */
369 		 7,  6,         /* data_len, bswap16 */
370 		-1, -1,         /* skip vlan_tci */
371 		 3,  2,  1,  0  /* hash.rss, bswap32 */
372 	};
373 	const uint8x16_t mcqe_shuf_m2 = {
374 		-1, -1, -1, -1, /* skip packet_type */
375 		15, 14, -1, -1, /* pkt_len, bswap16 */
376 		15, 14,         /* data_len, bswap16 */
377 		-1, -1,         /* skip vlan_tci */
378 		11, 10,  9,  8  /* hash.rss, bswap32 */
379 	};
380 	/* Restore the compressed count. Must be 16 bits. */
381 	const uint16_t mcqe_n = t_pkt->data_len +
382 				(rxq->crc_present * RTE_ETHER_CRC_LEN);
383 	const uint64x2_t rearm =
384 		vld1q_u64((void *)&t_pkt->rearm_data);
385 	const uint32x4_t rxdf_mask = {
386 		0xffffffff, /* packet_type */
387 		0,          /* skip pkt_len */
388 		0xffff0000, /* vlan_tci, skip data_len */
389 		0,          /* skip hash.rss */
390 	};
391 	const uint8x16_t rxdf =
392 		vandq_u8(vld1q_u8((void *)&t_pkt->rx_descriptor_fields1),
393 			 vreinterpretq_u8_u32(rxdf_mask));
394 	const uint16x8_t crc_adj = {
395 		0, 0,
396 		rxq->crc_present * RTE_ETHER_CRC_LEN, 0,
397 		rxq->crc_present * RTE_ETHER_CRC_LEN, 0,
398 		0, 0
399 	};
400 	const uint32_t flow_tag = t_pkt->hash.fdir.hi;
401 #ifdef MLX5_PMD_SOFT_COUNTERS
402 	uint32_t rcvd_byte = 0;
403 #endif
404 	/* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */
405 	const uint8x8_t len_shuf_m = {
406 		 7,  6,         /* 1st mCQE */
407 		15, 14,         /* 2nd mCQE */
408 		23, 22,         /* 3rd mCQE */
409 		31, 30          /* 4th mCQE */
410 	};
411 
412 	/*
413 	 * A. load mCQEs into a 128bit register.
414 	 * B. store rearm data to mbuf.
415 	 * C. combine data from mCQEs with rx_descriptor_fields1.
416 	 * D. store rx_descriptor_fields1.
417 	 * E. store flow tag (rte_flow mark).
418 	 */
419 	for (pos = 0; pos < mcqe_n; ) {
420 		uint8_t *p = (void *)&mcq[pos % 8];
421 		uint8_t *e0 = (void *)&elts[pos]->rearm_data;
422 		uint8_t *e1 = (void *)&elts[pos + 1]->rearm_data;
423 		uint8_t *e2 = (void *)&elts[pos + 2]->rearm_data;
424 		uint8_t *e3 = (void *)&elts[pos + 3]->rearm_data;
425 		uint16x4_t byte_cnt;
426 #ifdef MLX5_PMD_SOFT_COUNTERS
427 		uint16x4_t invalid_mask =
428 			vcreate_u16(mcqe_n - pos < MLX5_VPMD_DESCS_PER_LOOP ?
429 				    -1UL << ((mcqe_n - pos) *
430 					     sizeof(uint16_t) * 8) : 0);
431 #endif
432 
433 		if (!(pos & 0x7) && pos + 8 < mcqe_n)
434 			rte_prefetch0((void *)(cq + pos + 8));
435 		__asm__ volatile (
436 		/* A.1 load mCQEs into a 128bit register. */
437 		"ld1 {v16.16b - v17.16b}, [%[mcq]] \n\t"
438 		/* B.1 store rearm data to mbuf. */
439 		"st1 {%[rearm].2d}, [%[e0]] \n\t"
440 		"add %[e0], %[e0], #16 \n\t"
441 		"st1 {%[rearm].2d}, [%[e1]] \n\t"
442 		"add %[e1], %[e1], #16 \n\t"
443 		/* C.1 combine data from mCQEs with rx_descriptor_fields1. */
444 		"tbl v18.16b, {v16.16b}, %[mcqe_shuf_m1].16b \n\t"
445 		"tbl v19.16b, {v16.16b}, %[mcqe_shuf_m2].16b \n\t"
446 		"sub v18.8h, v18.8h, %[crc_adj].8h \n\t"
447 		"sub v19.8h, v19.8h, %[crc_adj].8h \n\t"
448 		"orr v18.16b, v18.16b, %[rxdf].16b \n\t"
449 		"orr v19.16b, v19.16b, %[rxdf].16b \n\t"
450 		/* D.1 store rx_descriptor_fields1. */
451 		"st1 {v18.2d}, [%[e0]] \n\t"
452 		"st1 {v19.2d}, [%[e1]] \n\t"
453 		/* B.1 store rearm data to mbuf. */
454 		"st1 {%[rearm].2d}, [%[e2]] \n\t"
455 		"add %[e2], %[e2], #16 \n\t"
456 		"st1 {%[rearm].2d}, [%[e3]] \n\t"
457 		"add %[e3], %[e3], #16 \n\t"
458 		/* C.1 combine data from mCQEs with rx_descriptor_fields1. */
459 		"tbl v18.16b, {v17.16b}, %[mcqe_shuf_m1].16b \n\t"
460 		"tbl v19.16b, {v17.16b}, %[mcqe_shuf_m2].16b \n\t"
461 		"sub v18.8h, v18.8h, %[crc_adj].8h \n\t"
462 		"sub v19.8h, v19.8h, %[crc_adj].8h \n\t"
463 		"orr v18.16b, v18.16b, %[rxdf].16b \n\t"
464 		"orr v19.16b, v19.16b, %[rxdf].16b \n\t"
465 		/* D.1 store rx_descriptor_fields1. */
466 		"st1 {v18.2d}, [%[e2]] \n\t"
467 		"st1 {v19.2d}, [%[e3]] \n\t"
468 #ifdef MLX5_PMD_SOFT_COUNTERS
469 		"tbl %[byte_cnt].8b, {v16.16b - v17.16b}, %[len_shuf_m].8b \n\t"
470 #endif
471 		:[byte_cnt]"=&w"(byte_cnt)
472 		:[mcq]"r"(p),
473 		 [rxdf]"w"(rxdf),
474 		 [rearm]"w"(rearm),
475 		 [e3]"r"(e3), [e2]"r"(e2), [e1]"r"(e1), [e0]"r"(e0),
476 		 [mcqe_shuf_m1]"w"(mcqe_shuf_m1),
477 		 [mcqe_shuf_m2]"w"(mcqe_shuf_m2),
478 		 [crc_adj]"w"(crc_adj),
479 		 [len_shuf_m]"w"(len_shuf_m)
480 		:"memory", "v16", "v17", "v18", "v19");
481 #ifdef MLX5_PMD_SOFT_COUNTERS
482 		byte_cnt = vbic_u16(byte_cnt, invalid_mask);
483 		rcvd_byte += vget_lane_u64(vpaddl_u32(vpaddl_u16(byte_cnt)), 0);
484 #endif
485 		if (rxq->mark) {
486 			/* E.1 store flow tag (rte_flow mark). */
487 			elts[pos]->hash.fdir.hi = flow_tag;
488 			elts[pos + 1]->hash.fdir.hi = flow_tag;
489 			elts[pos + 2]->hash.fdir.hi = flow_tag;
490 			elts[pos + 3]->hash.fdir.hi = flow_tag;
491 		}
492 		pos += MLX5_VPMD_DESCS_PER_LOOP;
493 		/* Move to next CQE and invalidate consumed CQEs. */
494 		if (!(pos & 0x7) && pos < mcqe_n) {
495 			mcq = (void *)&(cq + pos)->pkt_info;
496 			for (i = 0; i < 8; ++i)
497 				cq[inv++].op_own = MLX5_CQE_INVALIDATE;
498 		}
499 	}
500 	/* Invalidate the rest of CQEs. */
501 	for (; inv < mcqe_n; ++inv)
502 		cq[inv].op_own = MLX5_CQE_INVALIDATE;
503 #ifdef MLX5_PMD_SOFT_COUNTERS
504 	rxq->stats.ipackets += mcqe_n;
505 	rxq->stats.ibytes += rcvd_byte;
506 #endif
507 	rxq->cq_ci += mcqe_n;
508 }
509 
510 /**
511  * Calculate packet type and offload flag for mbuf and store it.
512  *
513  * @param rxq
514  *   Pointer to RX queue structure.
515  * @param ptype_info
516  *   Array of four 4bytes packet type info extracted from the original
517  *   completion descriptor.
518  * @param flow_tag
519  *   Array of four 4bytes flow ID extracted from the original completion
520  *   descriptor.
521  * @param op_err
522  *   Opcode vector having responder error status. Each field is 4B.
523  * @param pkts
524  *   Pointer to array of packets to be filled.
525  */
526 static inline void
527 rxq_cq_to_ptype_oflags_v(struct mlx5_rxq_data *rxq,
528 			 uint32x4_t ptype_info, uint32x4_t flow_tag,
529 			 uint16x4_t op_err, struct rte_mbuf **pkts)
530 {
531 	uint16x4_t ptype;
532 	uint32x4_t pinfo, cv_flags;
533 	uint32x4_t ol_flags =
534 		vdupq_n_u32(rxq->rss_hash * PKT_RX_RSS_HASH |
535 			    rxq->hw_timestamp * PKT_RX_TIMESTAMP);
536 	const uint32x4_t ptype_ol_mask = { 0x106, 0x106, 0x106, 0x106 };
537 	const uint8x16_t cv_flag_sel = {
538 		0,
539 		(uint8_t)(PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED),
540 		(uint8_t)(PKT_RX_IP_CKSUM_GOOD >> 1),
541 		0,
542 		(uint8_t)(PKT_RX_L4_CKSUM_GOOD >> 1),
543 		0,
544 		(uint8_t)((PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1),
545 		0, 0, 0, 0, 0, 0, 0, 0, 0
546 	};
547 	const uint32x4_t cv_mask =
548 		vdupq_n_u32(PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
549 			    PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED);
550 	const uint64x1_t mbuf_init = vld1_u64(&rxq->mbuf_initializer);
551 	const uint64x1_t r32_mask = vcreate_u64(0xffffffff);
552 	uint64x2_t rearm0, rearm1, rearm2, rearm3;
553 	uint8_t pt_idx0, pt_idx1, pt_idx2, pt_idx3;
554 
555 	if (rxq->mark) {
556 		const uint32x4_t ft_def = vdupq_n_u32(MLX5_FLOW_MARK_DEFAULT);
557 		const uint32x4_t fdir_flags = vdupq_n_u32(PKT_RX_FDIR);
558 		uint32x4_t fdir_id_flags = vdupq_n_u32(PKT_RX_FDIR_ID);
559 		uint32x4_t invalid_mask;
560 
561 		/* Check if flow tag is non-zero then set PKT_RX_FDIR. */
562 		invalid_mask = vceqzq_u32(flow_tag);
563 		ol_flags = vorrq_u32(ol_flags,
564 				     vbicq_u32(fdir_flags, invalid_mask));
565 		/* Mask out invalid entries. */
566 		fdir_id_flags = vbicq_u32(fdir_id_flags, invalid_mask);
567 		/* Check if flow tag MLX5_FLOW_MARK_DEFAULT. */
568 		ol_flags = vorrq_u32(ol_flags,
569 				     vbicq_u32(fdir_id_flags,
570 					       vceqq_u32(flow_tag, ft_def)));
571 	}
572 	/*
573 	 * ptype_info has the following:
574 	 * bit[1]     = l3_ok
575 	 * bit[2]     = l4_ok
576 	 * bit[8]     = cv
577 	 * bit[11:10] = l3_hdr_type
578 	 * bit[14:12] = l4_hdr_type
579 	 * bit[15]    = ip_frag
580 	 * bit[16]    = tunneled
581 	 * bit[17]    = outer_l3_type
582 	 */
583 	ptype = vshrn_n_u32(ptype_info, 10);
584 	/* Errored packets will have RTE_PTYPE_ALL_MASK. */
585 	ptype = vorr_u16(ptype, op_err);
586 	pt_idx0 = vget_lane_u8(vreinterpret_u8_u16(ptype), 6);
587 	pt_idx1 = vget_lane_u8(vreinterpret_u8_u16(ptype), 4);
588 	pt_idx2 = vget_lane_u8(vreinterpret_u8_u16(ptype), 2);
589 	pt_idx3 = vget_lane_u8(vreinterpret_u8_u16(ptype), 0);
590 	pkts[0]->packet_type = mlx5_ptype_table[pt_idx0] |
591 			       !!(pt_idx0 & (1 << 6)) * rxq->tunnel;
592 	pkts[1]->packet_type = mlx5_ptype_table[pt_idx1] |
593 			       !!(pt_idx1 & (1 << 6)) * rxq->tunnel;
594 	pkts[2]->packet_type = mlx5_ptype_table[pt_idx2] |
595 			       !!(pt_idx2 & (1 << 6)) * rxq->tunnel;
596 	pkts[3]->packet_type = mlx5_ptype_table[pt_idx3] |
597 			       !!(pt_idx3 & (1 << 6)) * rxq->tunnel;
598 	/* Fill flags for checksum and VLAN. */
599 	pinfo = vandq_u32(ptype_info, ptype_ol_mask);
600 	pinfo = vreinterpretq_u32_u8(
601 		vqtbl1q_u8(cv_flag_sel, vreinterpretq_u8_u32(pinfo)));
602 	/* Locate checksum flags at byte[2:1] and merge with VLAN flags. */
603 	cv_flags = vshlq_n_u32(pinfo, 9);
604 	cv_flags = vorrq_u32(pinfo, cv_flags);
605 	/* Move back flags to start from byte[0]. */
606 	cv_flags = vshrq_n_u32(cv_flags, 8);
607 	/* Mask out garbage bits. */
608 	cv_flags = vandq_u32(cv_flags, cv_mask);
609 	/* Merge to ol_flags. */
610 	ol_flags = vorrq_u32(ol_flags, cv_flags);
611 	/* Merge mbuf_init and ol_flags, and store. */
612 	rearm0 = vcombine_u64(mbuf_init,
613 			      vshr_n_u64(vget_high_u64(vreinterpretq_u64_u32(
614 						       ol_flags)), 32));
615 	rearm1 = vcombine_u64(mbuf_init,
616 			      vand_u64(vget_high_u64(vreinterpretq_u64_u32(
617 						     ol_flags)), r32_mask));
618 	rearm2 = vcombine_u64(mbuf_init,
619 			      vshr_n_u64(vget_low_u64(vreinterpretq_u64_u32(
620 						      ol_flags)), 32));
621 	rearm3 = vcombine_u64(mbuf_init,
622 			      vand_u64(vget_low_u64(vreinterpretq_u64_u32(
623 						    ol_flags)), r32_mask));
624 	vst1q_u64((void *)&pkts[0]->rearm_data, rearm0);
625 	vst1q_u64((void *)&pkts[1]->rearm_data, rearm1);
626 	vst1q_u64((void *)&pkts[2]->rearm_data, rearm2);
627 	vst1q_u64((void *)&pkts[3]->rearm_data, rearm3);
628 }
629 
630 /**
631  * Receive burst of packets. An errored completion also consumes a mbuf, but the
632  * packet_type is set to be RTE_PTYPE_ALL_MASK. Marked mbufs should be freed
633  * before returning to application.
634  *
635  * @param rxq
636  *   Pointer to RX queue structure.
637  * @param[out] pkts
638  *   Array to store received packets.
639  * @param pkts_n
640  *   Maximum number of packets in array.
641  * @param[out] err
642  *   Pointer to a flag. Set non-zero value if pkts array has at least one error
643  *   packet to handle.
644  *
645  * @return
646  *   Number of packets received including errors (<= pkts_n).
647  */
648 static inline uint16_t
649 rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
650 	    uint64_t *err)
651 {
652 	const uint16_t q_n = 1 << rxq->cqe_n;
653 	const uint16_t q_mask = q_n - 1;
654 	volatile struct mlx5_cqe *cq;
655 	struct rte_mbuf **elts;
656 	unsigned int pos;
657 	uint64_t n;
658 	uint16_t repl_n;
659 	uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
660 	uint16_t nocmp_n = 0;
661 	uint16_t rcvd_pkt = 0;
662 	unsigned int cq_idx = rxq->cq_ci & q_mask;
663 	unsigned int elts_idx;
664 	const uint16x4_t ownership = vdup_n_u16(!(rxq->cq_ci & (q_mask + 1)));
665 	const uint16x4_t owner_check = vcreate_u16(0x0001000100010001);
666 	const uint16x4_t opcode_check = vcreate_u16(0x00f000f000f000f0);
667 	const uint16x4_t format_check = vcreate_u16(0x000c000c000c000c);
668 	const uint16x4_t resp_err_check = vcreate_u16(0x00e000e000e000e0);
669 #ifdef MLX5_PMD_SOFT_COUNTERS
670 	uint32_t rcvd_byte = 0;
671 #endif
672 	/* Mask to generate 16B length vector. */
673 	const uint8x8_t len_shuf_m = {
674 		52, 53,         /* 4th CQE */
675 		36, 37,         /* 3rd CQE */
676 		20, 21,         /* 2nd CQE */
677 		 4,  5          /* 1st CQE */
678 	};
679 	/* Mask to extract 16B data from a 64B CQE. */
680 	const uint8x16_t cqe_shuf_m = {
681 		28, 29,         /* hdr_type_etc */
682 		 0,             /* pkt_info */
683 		-1,             /* null */
684 		47, 46,         /* byte_cnt, bswap16 */
685 		31, 30,         /* vlan_info, bswap16 */
686 		15, 14, 13, 12, /* rx_hash_res, bswap32 */
687 		57, 58, 59,     /* flow_tag */
688 		63              /* op_own */
689 	};
690 	/* Mask to generate 16B data for mbuf. */
691 	const uint8x16_t mb_shuf_m = {
692 		 4,  5, -1, -1, /* pkt_len */
693 		 4,  5,         /* data_len */
694 		 6,  7,         /* vlan_tci */
695 		 8,  9, 10, 11, /* hash.rss */
696 		12, 13, 14, -1  /* hash.fdir.hi */
697 	};
698 	/* Mask to generate 16B owner vector. */
699 	const uint8x8_t owner_shuf_m = {
700 		63, -1,         /* 4th CQE */
701 		47, -1,         /* 3rd CQE */
702 		31, -1,         /* 2nd CQE */
703 		15, -1          /* 1st CQE */
704 	};
705 	/* Mask to generate a vector having packet_type/ol_flags. */
706 	const uint8x16_t ptype_shuf_m = {
707 		48, 49, 50, -1, /* 4th CQE */
708 		32, 33, 34, -1, /* 3rd CQE */
709 		16, 17, 18, -1, /* 2nd CQE */
710 		 0,  1,  2, -1  /* 1st CQE */
711 	};
712 	/* Mask to generate a vector having flow tags. */
713 	const uint8x16_t ftag_shuf_m = {
714 		60, 61, 62, -1, /* 4th CQE */
715 		44, 45, 46, -1, /* 3rd CQE */
716 		28, 29, 30, -1, /* 2nd CQE */
717 		12, 13, 14, -1  /* 1st CQE */
718 	};
719 	const uint16x8_t crc_adj = {
720 		0, 0, rxq->crc_present * RTE_ETHER_CRC_LEN, 0, 0, 0, 0, 0
721 	};
722 	const uint32x4_t flow_mark_adj = { 0, 0, 0, rxq->mark * (-1) };
723 
724 	assert(rxq->sges_n == 0);
725 	assert(rxq->cqe_n == rxq->elts_n);
726 	cq = &(*rxq->cqes)[cq_idx];
727 	rte_prefetch_non_temporal(cq);
728 	rte_prefetch_non_temporal(cq + 1);
729 	rte_prefetch_non_temporal(cq + 2);
730 	rte_prefetch_non_temporal(cq + 3);
731 	pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
732 	/*
733 	 * Order of indexes:
734 	 *   rq_ci >= cq_ci >= rq_pi
735 	 * Definition of indexes:
736 	 *   rq_ci - cq_ci := # of buffers owned by HW (posted).
737 	 *   cq_ci - rq_pi := # of buffers not returned to app (decompressed).
738 	 *   N - (rq_ci - rq_pi) := # of buffers consumed (to be replenished).
739 	 */
740 	repl_n = q_n - (rxq->rq_ci - rxq->rq_pi);
741 	if (repl_n >= rxq->rq_repl_thresh)
742 		mlx5_rx_replenish_bulk_mbuf(rxq, repl_n);
743 	/* See if there're unreturned mbufs from compressed CQE. */
744 	rcvd_pkt = rxq->cq_ci - rxq->rq_pi;
745 	if (rcvd_pkt > 0) {
746 		rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
747 		rxq_copy_mbuf_v(rxq, pkts, rcvd_pkt);
748 		rxq->rq_pi += rcvd_pkt;
749 		pkts += rcvd_pkt;
750 	}
751 	elts_idx = rxq->rq_pi & q_mask;
752 	elts = &(*rxq->elts)[elts_idx];
753 	/* Not to overflow pkts array. */
754 	pkts_n = RTE_ALIGN_FLOOR(pkts_n - rcvd_pkt, MLX5_VPMD_DESCS_PER_LOOP);
755 	/* Not to cross queue end. */
756 	pkts_n = RTE_MIN(pkts_n, q_n - elts_idx);
757 	if (!pkts_n)
758 		return rcvd_pkt;
759 	/* At this point, there shouldn't be any remained packets. */
760 	assert(rxq->rq_pi == rxq->cq_ci);
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_cq_decompress_v(rxq, &cq[nocmp_n], &elts[nocmp_n]);
1007 		/* Return more packets if needed. */
1008 		if (nocmp_n < pkts_n) {
1009 			uint16_t n = rxq->cq_ci - rxq->rq_pi;
1010 
1011 			n = RTE_MIN(n, pkts_n - nocmp_n);
1012 			rxq_copy_mbuf_v(rxq, &pkts[nocmp_n], n);
1013 			rxq->rq_pi += n;
1014 			rcvd_pkt += n;
1015 		}
1016 	}
1017 	rte_compiler_barrier();
1018 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1019 	return rcvd_pkt;
1020 }
1021 
1022 #endif /* RTE_PMD_MLX5_RXTX_VEC_NEON_H_ */
1023