xref: /dpdk/drivers/net/mlx5/mlx5_rxtx_vec_sse.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_SSE_H_
7 #define RTE_PMD_MLX5_RXTX_VEC_SSE_H_
8 
9 #include <assert.h>
10 #include <stdint.h>
11 #include <string.h>
12 #include <stdlib.h>
13 #include <smmintrin.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 #ifndef __INTEL_COMPILER
28 #pragma GCC diagnostic ignored "-Wcast-qual"
29 #endif
30 
31 /**
32  * Fill in buffer descriptors in a multi-packet send descriptor.
33  *
34  * @param txq
35  *   Pointer to TX queue structure.
36  * @param dseg
37  *   Pointer to buffer descriptor to be written.
38  * @param pkts
39  *   Pointer to array of packets to be sent.
40  * @param n
41  *   Number of packets to be filled.
42  */
43 static inline void
44 txq_wr_dseg_v(struct mlx5_txq_data *txq, __m128i *dseg,
45 	      struct rte_mbuf **pkts, unsigned int n)
46 {
47 	unsigned int pos;
48 	uintptr_t addr;
49 	const __m128i shuf_mask_dseg =
50 		_mm_set_epi8(8,  9, 10, 11, /* addr, bswap64 */
51 			    12, 13, 14, 15,
52 			     7,  6,  5,  4, /* lkey */
53 			     0,  1,  2,  3  /* length, bswap32 */);
54 #ifdef MLX5_PMD_SOFT_COUNTERS
55 	uint32_t tx_byte = 0;
56 #endif
57 
58 	for (pos = 0; pos < n; ++pos, ++dseg) {
59 		__m128i desc;
60 		struct rte_mbuf *pkt = pkts[pos];
61 
62 		addr = rte_pktmbuf_mtod(pkt, uintptr_t);
63 		desc = _mm_set_epi32(addr >> 32,
64 				     addr,
65 				     mlx5_tx_mb2mr(txq, pkt),
66 				     DATA_LEN(pkt));
67 		desc = _mm_shuffle_epi8(desc, shuf_mask_dseg);
68 		_mm_store_si128(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 __m128i shuf_mask_ctrl =
121 			_mm_set_epi8(15, 14, 13, 12,
122 				      8,  9, 10, 11, /* bswap32 */
123 				      4,  5,  6,  7, /* bswap32 */
124 				      0,  1,  2,  3  /* bswap32 */);
125 		uint8_t cs_flags;
126 		uint16_t max_elts;
127 		uint16_t max_wqe;
128 		__m128i *t_wqe, *dseg;
129 		__m128i ctrl;
130 		rte_be32_t metadata =
131 			metadata_ol && (buf->ol_flags & PKT_TX_METADATA) ?
132 			buf->tx_metadata : 0;
133 
134 		assert(segs_n);
135 		max_elts = elts_n - (elts_head - txq->elts_tail);
136 		max_wqe = wq_n - (txq->wqe_ci - txq->wqe_pi);
137 		/*
138 		 * A MPW session consumes 2 WQEs at most to
139 		 * include MLX5_MPW_DSEG_MAX pointers.
140 		 */
141 		if (segs_n == 1 ||
142 		    max_elts < segs_n || max_wqe < 2)
143 			break;
144 		if (segs_n > MLX5_MPW_DSEG_MAX) {
145 			txq->stats.oerrors++;
146 			break;
147 		}
148 		wqe = &((volatile struct mlx5_wqe64 *)
149 			 txq->wqes)[wqe_ci & wq_mask].hdr;
150 		cs_flags = txq_ol_cksum_to_cs(buf);
151 		/* Title WQEBB pointer. */
152 		t_wqe = (__m128i *)wqe;
153 		dseg = (__m128i *)(wqe + 1);
154 		do {
155 			if (!(ds++ % nb_dword_per_wqebb)) {
156 				dseg = (__m128i *)
157 					&((volatile struct mlx5_wqe64 *)
158 					   txq->wqes)[++wqe_ci & wq_mask];
159 			}
160 			txq_wr_dseg_v(txq, dseg++, &buf, 1);
161 			(*txq->elts)[elts_head++ & elts_m] = buf;
162 			buf = buf->next;
163 		} while (--segs_n);
164 		++wqe_ci;
165 		/* Fill CTRL in the header. */
166 		ctrl = _mm_set_epi32(0, 0, txq->qp_num_8s | ds,
167 				     MLX5_OPC_MOD_MPW << 24 |
168 				     txq->wqe_ci << 8 | MLX5_OPCODE_TSO);
169 		ctrl = _mm_shuffle_epi8(ctrl, shuf_mask_ctrl);
170 		_mm_store_si128(t_wqe, ctrl);
171 		/* Fill ESEG in the header. */
172 		_mm_store_si128(t_wqe + 1,
173 				_mm_set_epi32(0, metadata,
174 					      (rte_cpu_to_be_16(len) << 16) |
175 					      cs_flags, 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 __m128i shuf_mask_ctrl =
240 		_mm_set_epi8(15, 14, 13, 12,
241 			      8,  9, 10, 11, /* bswap32 */
242 			      4,  5,  6,  7, /* bswap32 */
243 			      0,  1,  2,  3  /* bswap32 */);
244 	__m128i *t_wqe, *dseg;
245 	__m128i ctrl;
246 
247 	/* Make sure all packets can fit into a single WQE. */
248 	assert(elts_n > pkts_n);
249 	mlx5_tx_complete(txq);
250 	max_elts = (elts_n - (elts_head - txq->elts_tail));
251 	max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
252 	pkts_n = RTE_MIN((unsigned int)RTE_MIN(pkts_n, max_wqe), max_elts);
253 	assert(pkts_n <= MLX5_DSEG_MAX - nb_dword_in_hdr);
254 	if (unlikely(!pkts_n))
255 		return 0;
256 	elts = &(*txq->elts)[elts_head & elts_m];
257 	/* Loop for available tailroom first. */
258 	n = RTE_MIN(elts_n - (elts_head & elts_m), pkts_n);
259 	for (pos = 0; pos < (n & -2); pos += 2)
260 		_mm_storeu_si128((__m128i *)&elts[pos],
261 				 _mm_loadu_si128((__m128i *)&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 = (__m128i *)wqe;
273 	dseg = (__m128i *)(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 = (__m128i *)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 = _mm_set_epi32(txq->elts_head, comp_req,
296 			     txq->qp_num_8s | (pkts_n + 2),
297 			     MLX5_OPC_MOD_ENHANCED_MPSW << 24 |
298 				txq->wqe_ci << 8 | MLX5_OPCODE_ENHANCED_MPSW);
299 	ctrl = _mm_shuffle_epi8(ctrl, shuf_mask_ctrl);
300 	_mm_store_si128(t_wqe, ctrl);
301 	/* Fill ESEG in the header. */
302 	_mm_store_si128(t_wqe + 1, _mm_set_epi32(0, metadata, cs_flags, 0));
303 #ifdef MLX5_PMD_SOFT_COUNTERS
304 	txq->stats.opackets += pkts_n;
305 #endif
306 	txq->wqe_ci += (nb_dword_in_hdr + pkts_n + (nb_dword_per_wqebb - 1)) /
307 		       nb_dword_per_wqebb;
308 	/* Ring QP doorbell. */
309 	mlx5_tx_dbrec_cond_wmb(txq, wqe, pkts_n < MLX5_VPMD_TX_MAX_BURST);
310 	return pkts_n;
311 }
312 
313 /**
314  * Store free buffers to RX SW ring.
315  *
316  * @param rxq
317  *   Pointer to RX queue structure.
318  * @param pkts
319  *   Pointer to array of packets to be stored.
320  * @param pkts_n
321  *   Number of packets to be stored.
322  */
323 static inline void
324 rxq_copy_mbuf_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t n)
325 {
326 	const uint16_t q_mask = (1 << rxq->elts_n) - 1;
327 	struct rte_mbuf **elts = &(*rxq->elts)[rxq->rq_pi & q_mask];
328 	unsigned int pos;
329 	uint16_t p = n & -2;
330 
331 	for (pos = 0; pos < p; pos += 2) {
332 		__m128i mbp;
333 
334 		mbp = _mm_loadu_si128((__m128i *)&elts[pos]);
335 		_mm_storeu_si128((__m128i *)&pkts[pos], mbp);
336 	}
337 	if (n & 1)
338 		pkts[pos] = elts[pos];
339 }
340 
341 /**
342  * Decompress a compressed completion and fill in mbufs in RX SW ring with data
343  * extracted from the title completion descriptor.
344  *
345  * @param rxq
346  *   Pointer to RX queue structure.
347  * @param cq
348  *   Pointer to completion array having a compressed completion at first.
349  * @param elts
350  *   Pointer to SW ring to be filled. The first mbuf has to be pre-built from
351  *   the title completion descriptor to be copied to the rest of mbufs.
352  */
353 static inline void
354 rxq_cq_decompress_v(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cq,
355 		    struct rte_mbuf **elts)
356 {
357 	volatile struct mlx5_mini_cqe8 *mcq = (void *)(cq + 1);
358 	struct rte_mbuf *t_pkt = elts[0]; /* Title packet is pre-built. */
359 	unsigned int pos;
360 	unsigned int i;
361 	unsigned int inv = 0;
362 	/* Mask to shuffle from extracted mini CQE to mbuf. */
363 	const __m128i shuf_mask1 =
364 		_mm_set_epi8(0,  1,  2,  3, /* rss, bswap32 */
365 			    -1, -1,         /* skip vlan_tci */
366 			     6,  7,         /* data_len, bswap16 */
367 			    -1, -1,  6,  7, /* pkt_len, bswap16 */
368 			    -1, -1, -1, -1  /* skip packet_type */);
369 	const __m128i shuf_mask2 =
370 		_mm_set_epi8(8,  9, 10, 11, /* rss, bswap32 */
371 			    -1, -1,         /* skip vlan_tci */
372 			    14, 15,         /* data_len, bswap16 */
373 			    -1, -1, 14, 15, /* pkt_len, bswap16 */
374 			    -1, -1, -1, -1  /* skip packet_type */);
375 	/* Restore the compressed count. Must be 16 bits. */
376 	const uint16_t mcqe_n = t_pkt->data_len +
377 				(rxq->crc_present * RTE_ETHER_CRC_LEN);
378 	const __m128i rearm =
379 		_mm_loadu_si128((__m128i *)&t_pkt->rearm_data);
380 	const __m128i rxdf =
381 		_mm_loadu_si128((__m128i *)&t_pkt->rx_descriptor_fields1);
382 	const __m128i crc_adj =
383 		_mm_set_epi16(0, 0, 0,
384 			      rxq->crc_present * RTE_ETHER_CRC_LEN,
385 			      0,
386 			      rxq->crc_present * RTE_ETHER_CRC_LEN,
387 			      0, 0);
388 	const uint32_t flow_tag = t_pkt->hash.fdir.hi;
389 #ifdef MLX5_PMD_SOFT_COUNTERS
390 	const __m128i zero = _mm_setzero_si128();
391 	const __m128i ones = _mm_cmpeq_epi32(zero, zero);
392 	uint32_t rcvd_byte = 0;
393 	/* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */
394 	const __m128i len_shuf_mask =
395 		_mm_set_epi8(-1, -1, -1, -1,
396 			     -1, -1, -1, -1,
397 			     14, 15,  6,  7,
398 			     10, 11,  2,  3);
399 #endif
400 
401 	/*
402 	 * A. load mCQEs into a 128bit register.
403 	 * B. store rearm data to mbuf.
404 	 * C. combine data from mCQEs with rx_descriptor_fields1.
405 	 * D. store rx_descriptor_fields1.
406 	 * E. store flow tag (rte_flow mark).
407 	 */
408 	for (pos = 0; pos < mcqe_n; ) {
409 		__m128i mcqe1, mcqe2;
410 		__m128i rxdf1, rxdf2;
411 #ifdef MLX5_PMD_SOFT_COUNTERS
412 		__m128i byte_cnt, invalid_mask;
413 #endif
414 
415 		if (!(pos & 0x7) && pos + 8 < mcqe_n)
416 			rte_prefetch0((void *)(cq + pos + 8));
417 		/* A.1 load mCQEs into a 128bit register. */
418 		mcqe1 = _mm_loadu_si128((__m128i *)&mcq[pos % 8]);
419 		mcqe2 = _mm_loadu_si128((__m128i *)&mcq[pos % 8 + 2]);
420 		/* B.1 store rearm data to mbuf. */
421 		_mm_storeu_si128((__m128i *)&elts[pos]->rearm_data, rearm);
422 		_mm_storeu_si128((__m128i *)&elts[pos + 1]->rearm_data, rearm);
423 		/* C.1 combine data from mCQEs with rx_descriptor_fields1. */
424 		rxdf1 = _mm_shuffle_epi8(mcqe1, shuf_mask1);
425 		rxdf2 = _mm_shuffle_epi8(mcqe1, shuf_mask2);
426 		rxdf1 = _mm_sub_epi16(rxdf1, crc_adj);
427 		rxdf2 = _mm_sub_epi16(rxdf2, crc_adj);
428 		rxdf1 = _mm_blend_epi16(rxdf1, rxdf, 0x23);
429 		rxdf2 = _mm_blend_epi16(rxdf2, rxdf, 0x23);
430 		/* D.1 store rx_descriptor_fields1. */
431 		_mm_storeu_si128((__m128i *)
432 				  &elts[pos]->rx_descriptor_fields1,
433 				 rxdf1);
434 		_mm_storeu_si128((__m128i *)
435 				  &elts[pos + 1]->rx_descriptor_fields1,
436 				 rxdf2);
437 		/* B.1 store rearm data to mbuf. */
438 		_mm_storeu_si128((__m128i *)&elts[pos + 2]->rearm_data, rearm);
439 		_mm_storeu_si128((__m128i *)&elts[pos + 3]->rearm_data, rearm);
440 		/* C.1 combine data from mCQEs with rx_descriptor_fields1. */
441 		rxdf1 = _mm_shuffle_epi8(mcqe2, shuf_mask1);
442 		rxdf2 = _mm_shuffle_epi8(mcqe2, shuf_mask2);
443 		rxdf1 = _mm_sub_epi16(rxdf1, crc_adj);
444 		rxdf2 = _mm_sub_epi16(rxdf2, crc_adj);
445 		rxdf1 = _mm_blend_epi16(rxdf1, rxdf, 0x23);
446 		rxdf2 = _mm_blend_epi16(rxdf2, rxdf, 0x23);
447 		/* D.1 store rx_descriptor_fields1. */
448 		_mm_storeu_si128((__m128i *)
449 				  &elts[pos + 2]->rx_descriptor_fields1,
450 				 rxdf1);
451 		_mm_storeu_si128((__m128i *)
452 				  &elts[pos + 3]->rx_descriptor_fields1,
453 				 rxdf2);
454 #ifdef MLX5_PMD_SOFT_COUNTERS
455 		invalid_mask = _mm_set_epi64x(0,
456 					      (mcqe_n - pos) *
457 					      sizeof(uint16_t) * 8);
458 		invalid_mask = _mm_sll_epi64(ones, invalid_mask);
459 		mcqe1 = _mm_srli_si128(mcqe1, 4);
460 		byte_cnt = _mm_blend_epi16(mcqe1, mcqe2, 0xcc);
461 		byte_cnt = _mm_shuffle_epi8(byte_cnt, len_shuf_mask);
462 		byte_cnt = _mm_andnot_si128(invalid_mask, byte_cnt);
463 		byte_cnt = _mm_hadd_epi16(byte_cnt, zero);
464 		rcvd_byte += _mm_cvtsi128_si64(_mm_hadd_epi16(byte_cnt, zero));
465 #endif
466 		if (rxq->mark) {
467 			/* E.1 store flow tag (rte_flow mark). */
468 			elts[pos]->hash.fdir.hi = flow_tag;
469 			elts[pos + 1]->hash.fdir.hi = flow_tag;
470 			elts[pos + 2]->hash.fdir.hi = flow_tag;
471 			elts[pos + 3]->hash.fdir.hi = flow_tag;
472 		}
473 		pos += MLX5_VPMD_DESCS_PER_LOOP;
474 		/* Move to next CQE and invalidate consumed CQEs. */
475 		if (!(pos & 0x7) && pos < mcqe_n) {
476 			mcq = (void *)(cq + pos);
477 			for (i = 0; i < 8; ++i)
478 				cq[inv++].op_own = MLX5_CQE_INVALIDATE;
479 		}
480 	}
481 	/* Invalidate the rest of CQEs. */
482 	for (; inv < mcqe_n; ++inv)
483 		cq[inv].op_own = MLX5_CQE_INVALIDATE;
484 #ifdef MLX5_PMD_SOFT_COUNTERS
485 	rxq->stats.ipackets += mcqe_n;
486 	rxq->stats.ibytes += rcvd_byte;
487 #endif
488 	rxq->cq_ci += mcqe_n;
489 }
490 
491 /**
492  * Calculate packet type and offload flag for mbuf and store it.
493  *
494  * @param rxq
495  *   Pointer to RX queue structure.
496  * @param cqes[4]
497  *   Array of four 16bytes completions extracted from the original completion
498  *   descriptor.
499  * @param op_err
500  *   Opcode vector having responder error status. Each field is 4B.
501  * @param pkts
502  *   Pointer to array of packets to be filled.
503  */
504 static inline void
505 rxq_cq_to_ptype_oflags_v(struct mlx5_rxq_data *rxq, __m128i cqes[4],
506 			 __m128i op_err, struct rte_mbuf **pkts)
507 {
508 	__m128i pinfo0, pinfo1;
509 	__m128i pinfo, ptype;
510 	__m128i ol_flags = _mm_set1_epi32(rxq->rss_hash * PKT_RX_RSS_HASH |
511 					  rxq->hw_timestamp * PKT_RX_TIMESTAMP);
512 	__m128i cv_flags;
513 	const __m128i zero = _mm_setzero_si128();
514 	const __m128i ptype_mask =
515 		_mm_set_epi32(0xfd06, 0xfd06, 0xfd06, 0xfd06);
516 	const __m128i ptype_ol_mask =
517 		_mm_set_epi32(0x106, 0x106, 0x106, 0x106);
518 	const __m128i pinfo_mask =
519 		_mm_set_epi32(0x3, 0x3, 0x3, 0x3);
520 	const __m128i cv_flag_sel =
521 		_mm_set_epi8(0, 0, 0, 0, 0, 0, 0, 0, 0,
522 			     (uint8_t)((PKT_RX_IP_CKSUM_GOOD |
523 					PKT_RX_L4_CKSUM_GOOD) >> 1),
524 			     0,
525 			     (uint8_t)(PKT_RX_L4_CKSUM_GOOD >> 1),
526 			     0,
527 			     (uint8_t)(PKT_RX_IP_CKSUM_GOOD >> 1),
528 			     (uint8_t)(PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED),
529 			     0);
530 	const __m128i cv_mask =
531 		_mm_set_epi32(PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
532 			      PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
533 			      PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
534 			      PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
535 			      PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
536 			      PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
537 			      PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD |
538 			      PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED);
539 	const __m128i mbuf_init =
540 		_mm_loadl_epi64((__m128i *)&rxq->mbuf_initializer);
541 	__m128i rearm0, rearm1, rearm2, rearm3;
542 	uint8_t pt_idx0, pt_idx1, pt_idx2, pt_idx3;
543 
544 	/* Extract pkt_info field. */
545 	pinfo0 = _mm_unpacklo_epi32(cqes[0], cqes[1]);
546 	pinfo1 = _mm_unpacklo_epi32(cqes[2], cqes[3]);
547 	pinfo = _mm_unpacklo_epi64(pinfo0, pinfo1);
548 	/* Extract hdr_type_etc field. */
549 	pinfo0 = _mm_unpackhi_epi32(cqes[0], cqes[1]);
550 	pinfo1 = _mm_unpackhi_epi32(cqes[2], cqes[3]);
551 	ptype = _mm_unpacklo_epi64(pinfo0, pinfo1);
552 	if (rxq->mark) {
553 		const __m128i pinfo_ft_mask =
554 			_mm_set_epi32(0xffffff00, 0xffffff00,
555 				      0xffffff00, 0xffffff00);
556 		const __m128i fdir_flags = _mm_set1_epi32(PKT_RX_FDIR);
557 		__m128i fdir_id_flags = _mm_set1_epi32(PKT_RX_FDIR_ID);
558 		__m128i flow_tag, invalid_mask;
559 
560 		flow_tag = _mm_and_si128(pinfo, pinfo_ft_mask);
561 		/* Check if flow tag is non-zero then set PKT_RX_FDIR. */
562 		invalid_mask = _mm_cmpeq_epi32(flow_tag, zero);
563 		ol_flags = _mm_or_si128(ol_flags,
564 					_mm_andnot_si128(invalid_mask,
565 							 fdir_flags));
566 		/* Mask out invalid entries. */
567 		fdir_id_flags = _mm_andnot_si128(invalid_mask, fdir_id_flags);
568 		/* Check if flow tag MLX5_FLOW_MARK_DEFAULT. */
569 		ol_flags = _mm_or_si128(ol_flags,
570 					_mm_andnot_si128(
571 						_mm_cmpeq_epi32(flow_tag,
572 								pinfo_ft_mask),
573 						fdir_id_flags));
574 	}
575 	/*
576 	 * Merge the two fields to generate the following:
577 	 * bit[1]     = l3_ok
578 	 * bit[2]     = l4_ok
579 	 * bit[8]     = cv
580 	 * bit[11:10] = l3_hdr_type
581 	 * bit[14:12] = l4_hdr_type
582 	 * bit[15]    = ip_frag
583 	 * bit[16]    = tunneled
584 	 * bit[17]    = outer_l3_type
585 	 */
586 	ptype = _mm_and_si128(ptype, ptype_mask);
587 	pinfo = _mm_and_si128(pinfo, pinfo_mask);
588 	pinfo = _mm_slli_epi32(pinfo, 16);
589 	/* Make pinfo has merged fields for ol_flags calculation. */
590 	pinfo = _mm_or_si128(ptype, pinfo);
591 	ptype = _mm_srli_epi32(pinfo, 10);
592 	ptype = _mm_packs_epi32(ptype, zero);
593 	/* Errored packets will have RTE_PTYPE_ALL_MASK. */
594 	op_err = _mm_srli_epi16(op_err, 8);
595 	ptype = _mm_or_si128(ptype, op_err);
596 	pt_idx0 = _mm_extract_epi8(ptype, 0);
597 	pt_idx1 = _mm_extract_epi8(ptype, 2);
598 	pt_idx2 = _mm_extract_epi8(ptype, 4);
599 	pt_idx3 = _mm_extract_epi8(ptype, 6);
600 	pkts[0]->packet_type = mlx5_ptype_table[pt_idx0] |
601 			       !!(pt_idx0 & (1 << 6)) * rxq->tunnel;
602 	pkts[1]->packet_type = mlx5_ptype_table[pt_idx1] |
603 			       !!(pt_idx1 & (1 << 6)) * rxq->tunnel;
604 	pkts[2]->packet_type = mlx5_ptype_table[pt_idx2] |
605 			       !!(pt_idx2 & (1 << 6)) * rxq->tunnel;
606 	pkts[3]->packet_type = mlx5_ptype_table[pt_idx3] |
607 			       !!(pt_idx3 & (1 << 6)) * rxq->tunnel;
608 	/* Fill flags for checksum and VLAN. */
609 	pinfo = _mm_and_si128(pinfo, ptype_ol_mask);
610 	pinfo = _mm_shuffle_epi8(cv_flag_sel, pinfo);
611 	/* Locate checksum flags at byte[2:1] and merge with VLAN flags. */
612 	cv_flags = _mm_slli_epi32(pinfo, 9);
613 	cv_flags = _mm_or_si128(pinfo, cv_flags);
614 	/* Move back flags to start from byte[0]. */
615 	cv_flags = _mm_srli_epi32(cv_flags, 8);
616 	/* Mask out garbage bits. */
617 	cv_flags = _mm_and_si128(cv_flags, cv_mask);
618 	/* Merge to ol_flags. */
619 	ol_flags = _mm_or_si128(ol_flags, cv_flags);
620 	/* Merge mbuf_init and ol_flags. */
621 	rearm0 = _mm_blend_epi16(mbuf_init, _mm_slli_si128(ol_flags, 8), 0x30);
622 	rearm1 = _mm_blend_epi16(mbuf_init, _mm_slli_si128(ol_flags, 4), 0x30);
623 	rearm2 = _mm_blend_epi16(mbuf_init, ol_flags, 0x30);
624 	rearm3 = _mm_blend_epi16(mbuf_init, _mm_srli_si128(ol_flags, 4), 0x30);
625 	/* Write 8B rearm_data and 8B ol_flags. */
626 	_mm_store_si128((__m128i *)&pkts[0]->rearm_data, rearm0);
627 	_mm_store_si128((__m128i *)&pkts[1]->rearm_data, rearm1);
628 	_mm_store_si128((__m128i *)&pkts[2]->rearm_data, rearm2);
629 	_mm_store_si128((__m128i *)&pkts[3]->rearm_data, rearm3);
630 }
631 
632 /**
633  * Receive burst of packets. An errored completion also consumes a mbuf, but the
634  * packet_type is set to be RTE_PTYPE_ALL_MASK. Marked mbufs should be freed
635  * before returning to application.
636  *
637  * @param rxq
638  *   Pointer to RX queue structure.
639  * @param[out] pkts
640  *   Array to store received packets.
641  * @param pkts_n
642  *   Maximum number of packets in array.
643  * @param[out] err
644  *   Pointer to a flag. Set non-zero value if pkts array has at least one error
645  *   packet to handle.
646  *
647  * @return
648  *   Number of packets received including errors (<= pkts_n).
649  */
650 static inline uint16_t
651 rxq_burst_v(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts, uint16_t pkts_n,
652 	    uint64_t *err)
653 {
654 	const uint16_t q_n = 1 << rxq->cqe_n;
655 	const uint16_t q_mask = q_n - 1;
656 	volatile struct mlx5_cqe *cq;
657 	struct rte_mbuf **elts;
658 	unsigned int pos;
659 	uint64_t n;
660 	uint16_t repl_n;
661 	uint64_t comp_idx = MLX5_VPMD_DESCS_PER_LOOP;
662 	uint16_t nocmp_n = 0;
663 	uint16_t rcvd_pkt = 0;
664 	unsigned int cq_idx = rxq->cq_ci & q_mask;
665 	unsigned int elts_idx;
666 	unsigned int ownership = !!(rxq->cq_ci & (q_mask + 1));
667 	const __m128i owner_check =
668 		_mm_set_epi64x(0x0100000001000000LL, 0x0100000001000000LL);
669 	const __m128i opcode_check =
670 		_mm_set_epi64x(0xf0000000f0000000LL, 0xf0000000f0000000LL);
671 	const __m128i format_check =
672 		_mm_set_epi64x(0x0c0000000c000000LL, 0x0c0000000c000000LL);
673 	const __m128i resp_err_check =
674 		_mm_set_epi64x(0xe0000000e0000000LL, 0xe0000000e0000000LL);
675 #ifdef MLX5_PMD_SOFT_COUNTERS
676 	uint32_t rcvd_byte = 0;
677 	/* Mask to shuffle byte_cnt to add up stats. Do bswap16 for all. */
678 	const __m128i len_shuf_mask =
679 		_mm_set_epi8(-1, -1, -1, -1,
680 			     -1, -1, -1, -1,
681 			     12, 13,  8,  9,
682 			      4,  5,  0,  1);
683 #endif
684 	/* Mask to shuffle from extracted CQE to mbuf. */
685 	const __m128i shuf_mask =
686 		_mm_set_epi8(-1,  3,  2,  1, /* fdir.hi */
687 			     12, 13, 14, 15, /* rss, bswap32 */
688 			     10, 11,         /* vlan_tci, bswap16 */
689 			      4,  5,         /* data_len, bswap16 */
690 			     -1, -1,         /* zero out 2nd half of pkt_len */
691 			      4,  5          /* pkt_len, bswap16 */);
692 	/* Mask to blend from the last Qword to the first DQword. */
693 	const __m128i blend_mask =
694 		_mm_set_epi8(-1, -1, -1, -1,
695 			     -1, -1, -1, -1,
696 			      0,  0,  0,  0,
697 			      0,  0,  0, -1);
698 	const __m128i zero = _mm_setzero_si128();
699 	const __m128i ones = _mm_cmpeq_epi32(zero, zero);
700 	const __m128i crc_adj =
701 		_mm_set_epi16(0, 0, 0, 0, 0,
702 			      rxq->crc_present * RTE_ETHER_CRC_LEN,
703 			      0,
704 			      rxq->crc_present * RTE_ETHER_CRC_LEN);
705 	const __m128i flow_mark_adj = _mm_set_epi32(rxq->mark * (-1), 0, 0, 0);
706 
707 	assert(rxq->sges_n == 0);
708 	assert(rxq->cqe_n == rxq->elts_n);
709 	cq = &(*rxq->cqes)[cq_idx];
710 	rte_prefetch0(cq);
711 	rte_prefetch0(cq + 1);
712 	rte_prefetch0(cq + 2);
713 	rte_prefetch0(cq + 3);
714 	pkts_n = RTE_MIN(pkts_n, MLX5_VPMD_RX_MAX_BURST);
715 	/*
716 	 * Order of indexes:
717 	 *   rq_ci >= cq_ci >= rq_pi
718 	 * Definition of indexes:
719 	 *   rq_ci - cq_ci := # of buffers owned by HW (posted).
720 	 *   cq_ci - rq_pi := # of buffers not returned to app (decompressed).
721 	 *   N - (rq_ci - rq_pi) := # of buffers consumed (to be replenished).
722 	 */
723 	repl_n = q_n - (rxq->rq_ci - rxq->rq_pi);
724 	if (repl_n >= rxq->rq_repl_thresh)
725 		mlx5_rx_replenish_bulk_mbuf(rxq, repl_n);
726 	/* See if there're unreturned mbufs from compressed CQE. */
727 	rcvd_pkt = rxq->cq_ci - rxq->rq_pi;
728 	if (rcvd_pkt > 0) {
729 		rcvd_pkt = RTE_MIN(rcvd_pkt, pkts_n);
730 		rxq_copy_mbuf_v(rxq, pkts, rcvd_pkt);
731 		rxq->rq_pi += rcvd_pkt;
732 		pkts += rcvd_pkt;
733 	}
734 	elts_idx = rxq->rq_pi & q_mask;
735 	elts = &(*rxq->elts)[elts_idx];
736 	/* Not to overflow pkts array. */
737 	pkts_n = RTE_ALIGN_FLOOR(pkts_n - rcvd_pkt, MLX5_VPMD_DESCS_PER_LOOP);
738 	/* Not to cross queue end. */
739 	pkts_n = RTE_MIN(pkts_n, q_n - elts_idx);
740 	if (!pkts_n)
741 		return rcvd_pkt;
742 	/* At this point, there shouldn't be any remained packets. */
743 	assert(rxq->rq_pi == rxq->cq_ci);
744 	/*
745 	 * A. load first Qword (8bytes) in one loop.
746 	 * B. copy 4 mbuf pointers from elts ring to returing pkts.
747 	 * C. load remained CQE data and extract necessary fields.
748 	 *    Final 16bytes cqes[] extracted from original 64bytes CQE has the
749 	 *    following structure:
750 	 *        struct {
751 	 *          uint8_t  pkt_info;
752 	 *          uint8_t  flow_tag[3];
753 	 *          uint16_t byte_cnt;
754 	 *          uint8_t  rsvd4;
755 	 *          uint8_t  op_own;
756 	 *          uint16_t hdr_type_etc;
757 	 *          uint16_t vlan_info;
758 	 *          uint32_t rx_has_res;
759 	 *        } c;
760 	 * D. fill in mbuf.
761 	 * E. get valid CQEs.
762 	 * F. find compressed CQE.
763 	 */
764 	for (pos = 0;
765 	     pos < pkts_n;
766 	     pos += MLX5_VPMD_DESCS_PER_LOOP) {
767 		__m128i cqes[MLX5_VPMD_DESCS_PER_LOOP];
768 		__m128i cqe_tmp1, cqe_tmp2;
769 		__m128i pkt_mb0, pkt_mb1, pkt_mb2, pkt_mb3;
770 		__m128i op_own, op_own_tmp1, op_own_tmp2;
771 		__m128i opcode, owner_mask, invalid_mask;
772 		__m128i comp_mask;
773 		__m128i mask;
774 #ifdef MLX5_PMD_SOFT_COUNTERS
775 		__m128i byte_cnt;
776 #endif
777 		__m128i mbp1, mbp2;
778 		__m128i p = _mm_set_epi16(0, 0, 0, 0, 3, 2, 1, 0);
779 		unsigned int p1, p2, p3;
780 
781 		/* Prefetch next 4 CQEs. */
782 		if (pkts_n - pos >= 2 * MLX5_VPMD_DESCS_PER_LOOP) {
783 			rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP]);
784 			rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 1]);
785 			rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 2]);
786 			rte_prefetch0(&cq[pos + MLX5_VPMD_DESCS_PER_LOOP + 3]);
787 		}
788 		/* A.0 do not cross the end of CQ. */
789 		mask = _mm_set_epi64x(0, (pkts_n - pos) * sizeof(uint16_t) * 8);
790 		mask = _mm_sll_epi64(ones, mask);
791 		p = _mm_andnot_si128(mask, p);
792 		/* A.1 load cqes. */
793 		p3 = _mm_extract_epi16(p, 3);
794 		cqes[3] = _mm_loadl_epi64((__m128i *)
795 					   &cq[pos + p3].sop_drop_qpn);
796 		rte_compiler_barrier();
797 		p2 = _mm_extract_epi16(p, 2);
798 		cqes[2] = _mm_loadl_epi64((__m128i *)
799 					   &cq[pos + p2].sop_drop_qpn);
800 		rte_compiler_barrier();
801 		/* B.1 load mbuf pointers. */
802 		mbp1 = _mm_loadu_si128((__m128i *)&elts[pos]);
803 		mbp2 = _mm_loadu_si128((__m128i *)&elts[pos + 2]);
804 		/* A.1 load a block having op_own. */
805 		p1 = _mm_extract_epi16(p, 1);
806 		cqes[1] = _mm_loadl_epi64((__m128i *)
807 					   &cq[pos + p1].sop_drop_qpn);
808 		rte_compiler_barrier();
809 		cqes[0] = _mm_loadl_epi64((__m128i *)
810 					   &cq[pos].sop_drop_qpn);
811 		/* B.2 copy mbuf pointers. */
812 		_mm_storeu_si128((__m128i *)&pkts[pos], mbp1);
813 		_mm_storeu_si128((__m128i *)&pkts[pos + 2], mbp2);
814 		rte_cio_rmb();
815 		/* C.1 load remained CQE data and extract necessary fields. */
816 		cqe_tmp2 = _mm_load_si128((__m128i *)&cq[pos + p3]);
817 		cqe_tmp1 = _mm_load_si128((__m128i *)&cq[pos + p2]);
818 		cqes[3] = _mm_blendv_epi8(cqes[3], cqe_tmp2, blend_mask);
819 		cqes[2] = _mm_blendv_epi8(cqes[2], cqe_tmp1, blend_mask);
820 		cqe_tmp2 = _mm_loadu_si128((__m128i *)&cq[pos + p3].rsvd1[3]);
821 		cqe_tmp1 = _mm_loadu_si128((__m128i *)&cq[pos + p2].rsvd1[3]);
822 		cqes[3] = _mm_blend_epi16(cqes[3], cqe_tmp2, 0x30);
823 		cqes[2] = _mm_blend_epi16(cqes[2], cqe_tmp1, 0x30);
824 		cqe_tmp2 = _mm_loadl_epi64((__m128i *)&cq[pos + p3].rsvd2[10]);
825 		cqe_tmp1 = _mm_loadl_epi64((__m128i *)&cq[pos + p2].rsvd2[10]);
826 		cqes[3] = _mm_blend_epi16(cqes[3], cqe_tmp2, 0x04);
827 		cqes[2] = _mm_blend_epi16(cqes[2], cqe_tmp1, 0x04);
828 		/* C.2 generate final structure for mbuf with swapping bytes. */
829 		pkt_mb3 = _mm_shuffle_epi8(cqes[3], shuf_mask);
830 		pkt_mb2 = _mm_shuffle_epi8(cqes[2], shuf_mask);
831 		/* C.3 adjust CRC length. */
832 		pkt_mb3 = _mm_sub_epi16(pkt_mb3, crc_adj);
833 		pkt_mb2 = _mm_sub_epi16(pkt_mb2, crc_adj);
834 		/* C.4 adjust flow mark. */
835 		pkt_mb3 = _mm_add_epi32(pkt_mb3, flow_mark_adj);
836 		pkt_mb2 = _mm_add_epi32(pkt_mb2, flow_mark_adj);
837 		/* D.1 fill in mbuf - rx_descriptor_fields1. */
838 		_mm_storeu_si128((void *)&pkts[pos + 3]->pkt_len, pkt_mb3);
839 		_mm_storeu_si128((void *)&pkts[pos + 2]->pkt_len, pkt_mb2);
840 		/* E.1 extract op_own field. */
841 		op_own_tmp2 = _mm_unpacklo_epi32(cqes[2], cqes[3]);
842 		/* C.1 load remained CQE data and extract necessary fields. */
843 		cqe_tmp2 = _mm_load_si128((__m128i *)&cq[pos + p1]);
844 		cqe_tmp1 = _mm_load_si128((__m128i *)&cq[pos]);
845 		cqes[1] = _mm_blendv_epi8(cqes[1], cqe_tmp2, blend_mask);
846 		cqes[0] = _mm_blendv_epi8(cqes[0], cqe_tmp1, blend_mask);
847 		cqe_tmp2 = _mm_loadu_si128((__m128i *)&cq[pos + p1].rsvd1[3]);
848 		cqe_tmp1 = _mm_loadu_si128((__m128i *)&cq[pos].rsvd1[3]);
849 		cqes[1] = _mm_blend_epi16(cqes[1], cqe_tmp2, 0x30);
850 		cqes[0] = _mm_blend_epi16(cqes[0], cqe_tmp1, 0x30);
851 		cqe_tmp2 = _mm_loadl_epi64((__m128i *)&cq[pos + p1].rsvd2[10]);
852 		cqe_tmp1 = _mm_loadl_epi64((__m128i *)&cq[pos].rsvd2[10]);
853 		cqes[1] = _mm_blend_epi16(cqes[1], cqe_tmp2, 0x04);
854 		cqes[0] = _mm_blend_epi16(cqes[0], cqe_tmp1, 0x04);
855 		/* C.2 generate final structure for mbuf with swapping bytes. */
856 		pkt_mb1 = _mm_shuffle_epi8(cqes[1], shuf_mask);
857 		pkt_mb0 = _mm_shuffle_epi8(cqes[0], shuf_mask);
858 		/* C.3 adjust CRC length. */
859 		pkt_mb1 = _mm_sub_epi16(pkt_mb1, crc_adj);
860 		pkt_mb0 = _mm_sub_epi16(pkt_mb0, crc_adj);
861 		/* C.4 adjust flow mark. */
862 		pkt_mb1 = _mm_add_epi32(pkt_mb1, flow_mark_adj);
863 		pkt_mb0 = _mm_add_epi32(pkt_mb0, flow_mark_adj);
864 		/* E.1 extract op_own byte. */
865 		op_own_tmp1 = _mm_unpacklo_epi32(cqes[0], cqes[1]);
866 		op_own = _mm_unpackhi_epi64(op_own_tmp1, op_own_tmp2);
867 		/* D.1 fill in mbuf - rx_descriptor_fields1. */
868 		_mm_storeu_si128((void *)&pkts[pos + 1]->pkt_len, pkt_mb1);
869 		_mm_storeu_si128((void *)&pkts[pos]->pkt_len, pkt_mb0);
870 		/* E.2 flip owner bit to mark CQEs from last round. */
871 		owner_mask = _mm_and_si128(op_own, owner_check);
872 		if (ownership)
873 			owner_mask = _mm_xor_si128(owner_mask, owner_check);
874 		owner_mask = _mm_cmpeq_epi32(owner_mask, owner_check);
875 		owner_mask = _mm_packs_epi32(owner_mask, zero);
876 		/* E.3 get mask for invalidated CQEs. */
877 		opcode = _mm_and_si128(op_own, opcode_check);
878 		invalid_mask = _mm_cmpeq_epi32(opcode_check, opcode);
879 		invalid_mask = _mm_packs_epi32(invalid_mask, zero);
880 		/* E.4 mask out beyond boundary. */
881 		invalid_mask = _mm_or_si128(invalid_mask, mask);
882 		/* E.5 merge invalid_mask with invalid owner. */
883 		invalid_mask = _mm_or_si128(invalid_mask, owner_mask);
884 		/* F.1 find compressed CQE format. */
885 		comp_mask = _mm_and_si128(op_own, format_check);
886 		comp_mask = _mm_cmpeq_epi32(comp_mask, format_check);
887 		comp_mask = _mm_packs_epi32(comp_mask, zero);
888 		/* F.2 mask out invalid entries. */
889 		comp_mask = _mm_andnot_si128(invalid_mask, comp_mask);
890 		comp_idx = _mm_cvtsi128_si64(comp_mask);
891 		/* F.3 get the first compressed CQE. */
892 		comp_idx = comp_idx ?
893 				__builtin_ctzll(comp_idx) /
894 					(sizeof(uint16_t) * 8) :
895 				MLX5_VPMD_DESCS_PER_LOOP;
896 		/* E.6 mask out entries after the compressed CQE. */
897 		mask = _mm_set_epi64x(0, comp_idx * sizeof(uint16_t) * 8);
898 		mask = _mm_sll_epi64(ones, mask);
899 		invalid_mask = _mm_or_si128(invalid_mask, mask);
900 		/* E.7 count non-compressed valid CQEs. */
901 		n = _mm_cvtsi128_si64(invalid_mask);
902 		n = n ? __builtin_ctzll(n) / (sizeof(uint16_t) * 8) :
903 			MLX5_VPMD_DESCS_PER_LOOP;
904 		nocmp_n += n;
905 		/* D.2 get the final invalid mask. */
906 		mask = _mm_set_epi64x(0, n * sizeof(uint16_t) * 8);
907 		mask = _mm_sll_epi64(ones, mask);
908 		invalid_mask = _mm_or_si128(invalid_mask, mask);
909 		/* D.3 check error in opcode. */
910 		opcode = _mm_cmpeq_epi32(resp_err_check, opcode);
911 		opcode = _mm_packs_epi32(opcode, zero);
912 		opcode = _mm_andnot_si128(invalid_mask, opcode);
913 		/* D.4 mark if any error is set */
914 		*err |= _mm_cvtsi128_si64(opcode);
915 		/* D.5 fill in mbuf - rearm_data and packet_type. */
916 		rxq_cq_to_ptype_oflags_v(rxq, cqes, opcode, &pkts[pos]);
917 		if (rxq->hw_timestamp) {
918 			pkts[pos]->timestamp =
919 				rte_be_to_cpu_64(cq[pos].timestamp);
920 			pkts[pos + 1]->timestamp =
921 				rte_be_to_cpu_64(cq[pos + p1].timestamp);
922 			pkts[pos + 2]->timestamp =
923 				rte_be_to_cpu_64(cq[pos + p2].timestamp);
924 			pkts[pos + 3]->timestamp =
925 				rte_be_to_cpu_64(cq[pos + p3].timestamp);
926 		}
927 #ifdef MLX5_PMD_SOFT_COUNTERS
928 		/* Add up received bytes count. */
929 		byte_cnt = _mm_shuffle_epi8(op_own, len_shuf_mask);
930 		byte_cnt = _mm_andnot_si128(invalid_mask, byte_cnt);
931 		byte_cnt = _mm_hadd_epi16(byte_cnt, zero);
932 		rcvd_byte += _mm_cvtsi128_si64(_mm_hadd_epi16(byte_cnt, zero));
933 #endif
934 		/*
935 		 * Break the loop unless more valid CQE is expected, or if
936 		 * there's a compressed CQE.
937 		 */
938 		if (n != MLX5_VPMD_DESCS_PER_LOOP)
939 			break;
940 	}
941 	/* If no new CQE seen, return without updating cq_db. */
942 	if (unlikely(!nocmp_n && comp_idx == MLX5_VPMD_DESCS_PER_LOOP))
943 		return rcvd_pkt;
944 	/* Update the consumer indexes for non-compressed CQEs. */
945 	assert(nocmp_n <= pkts_n);
946 	rxq->cq_ci += nocmp_n;
947 	rxq->rq_pi += nocmp_n;
948 	rcvd_pkt += nocmp_n;
949 #ifdef MLX5_PMD_SOFT_COUNTERS
950 	rxq->stats.ipackets += nocmp_n;
951 	rxq->stats.ibytes += rcvd_byte;
952 #endif
953 	/* Decompress the last CQE if compressed. */
954 	if (comp_idx < MLX5_VPMD_DESCS_PER_LOOP && comp_idx == n) {
955 		assert(comp_idx == (nocmp_n % MLX5_VPMD_DESCS_PER_LOOP));
956 		rxq_cq_decompress_v(rxq, &cq[nocmp_n], &elts[nocmp_n]);
957 		/* Return more packets if needed. */
958 		if (nocmp_n < pkts_n) {
959 			uint16_t n = rxq->cq_ci - rxq->rq_pi;
960 
961 			n = RTE_MIN(n, pkts_n - nocmp_n);
962 			rxq_copy_mbuf_v(rxq, &pkts[nocmp_n], n);
963 			rxq->rq_pi += n;
964 			rcvd_pkt += n;
965 		}
966 	}
967 	rte_compiler_barrier();
968 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
969 	return rcvd_pkt;
970 }
971 
972 #endif /* RTE_PMD_MLX5_RXTX_VEC_SSE_H_ */
973