common/sfc_efx/base: implement Tx control path for Riverhead
[dpdk.git] / drivers / net / iavf / iavf_rxtx_vec_avx2.c
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2019 Intel Corporation
3  */
4
5 #include "iavf_rxtx_vec_common.h"
6
7 #include <x86intrin.h>
8
9 #ifndef __INTEL_COMPILER
10 #pragma GCC diagnostic ignored "-Wcast-qual"
11 #endif
12
13 static inline void
14 iavf_rxq_rearm(struct iavf_rx_queue *rxq)
15 {
16         int i;
17         uint16_t rx_id;
18         volatile union iavf_rx_desc *rxdp;
19         struct rte_mbuf **rxp = &rxq->sw_ring[rxq->rxrearm_start];
20
21         rxdp = rxq->rx_ring + rxq->rxrearm_start;
22
23         /* Pull 'n' more MBUFs into the software ring */
24         if (rte_mempool_get_bulk(rxq->mp,
25                                  (void *)rxp,
26                                  IAVF_RXQ_REARM_THRESH) < 0) {
27                 if (rxq->rxrearm_nb + IAVF_RXQ_REARM_THRESH >=
28                     rxq->nb_rx_desc) {
29                         __m128i dma_addr0;
30
31                         dma_addr0 = _mm_setzero_si128();
32                         for (i = 0; i < IAVF_VPMD_DESCS_PER_LOOP; i++) {
33                                 rxp[i] = &rxq->fake_mbuf;
34                                 _mm_store_si128((__m128i *)&rxdp[i].read,
35                                                 dma_addr0);
36                         }
37                 }
38                 rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
39                         IAVF_RXQ_REARM_THRESH;
40                 return;
41         }
42
43 #ifndef RTE_LIBRTE_IAVF_16BYTE_RX_DESC
44         struct rte_mbuf *mb0, *mb1;
45         __m128i dma_addr0, dma_addr1;
46         __m128i hdr_room = _mm_set_epi64x(RTE_PKTMBUF_HEADROOM,
47                         RTE_PKTMBUF_HEADROOM);
48         /* Initialize the mbufs in vector, process 2 mbufs in one loop */
49         for (i = 0; i < IAVF_RXQ_REARM_THRESH; i += 2, rxp += 2) {
50                 __m128i vaddr0, vaddr1;
51
52                 mb0 = rxp[0];
53                 mb1 = rxp[1];
54
55                 /* load buf_addr(lo 64bit) and buf_iova(hi 64bit) */
56                 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, buf_iova) !=
57                                 offsetof(struct rte_mbuf, buf_addr) + 8);
58                 vaddr0 = _mm_loadu_si128((__m128i *)&mb0->buf_addr);
59                 vaddr1 = _mm_loadu_si128((__m128i *)&mb1->buf_addr);
60
61                 /* convert pa to dma_addr hdr/data */
62                 dma_addr0 = _mm_unpackhi_epi64(vaddr0, vaddr0);
63                 dma_addr1 = _mm_unpackhi_epi64(vaddr1, vaddr1);
64
65                 /* add headroom to pa values */
66                 dma_addr0 = _mm_add_epi64(dma_addr0, hdr_room);
67                 dma_addr1 = _mm_add_epi64(dma_addr1, hdr_room);
68
69                 /* flush desc with pa dma_addr */
70                 _mm_store_si128((__m128i *)&rxdp++->read, dma_addr0);
71                 _mm_store_si128((__m128i *)&rxdp++->read, dma_addr1);
72         }
73 #else
74         struct rte_mbuf *mb0, *mb1, *mb2, *mb3;
75         __m256i dma_addr0_1, dma_addr2_3;
76         __m256i hdr_room = _mm256_set1_epi64x(RTE_PKTMBUF_HEADROOM);
77         /* Initialize the mbufs in vector, process 4 mbufs in one loop */
78         for (i = 0; i < IAVF_RXQ_REARM_THRESH;
79                         i += 4, rxp += 4, rxdp += 4) {
80                 __m128i vaddr0, vaddr1, vaddr2, vaddr3;
81                 __m256i vaddr0_1, vaddr2_3;
82
83                 mb0 = rxp[0];
84                 mb1 = rxp[1];
85                 mb2 = rxp[2];
86                 mb3 = rxp[3];
87
88                 /* load buf_addr(lo 64bit) and buf_iova(hi 64bit) */
89                 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, buf_iova) !=
90                                 offsetof(struct rte_mbuf, buf_addr) + 8);
91                 vaddr0 = _mm_loadu_si128((__m128i *)&mb0->buf_addr);
92                 vaddr1 = _mm_loadu_si128((__m128i *)&mb1->buf_addr);
93                 vaddr2 = _mm_loadu_si128((__m128i *)&mb2->buf_addr);
94                 vaddr3 = _mm_loadu_si128((__m128i *)&mb3->buf_addr);
95
96                 /**
97                  * merge 0 & 1, by casting 0 to 256-bit and inserting 1
98                  * into the high lanes. Similarly for 2 & 3
99                  */
100                 vaddr0_1 =
101                         _mm256_inserti128_si256(_mm256_castsi128_si256(vaddr0),
102                                                 vaddr1, 1);
103                 vaddr2_3 =
104                         _mm256_inserti128_si256(_mm256_castsi128_si256(vaddr2),
105                                                 vaddr3, 1);
106
107                 /* convert pa to dma_addr hdr/data */
108                 dma_addr0_1 = _mm256_unpackhi_epi64(vaddr0_1, vaddr0_1);
109                 dma_addr2_3 = _mm256_unpackhi_epi64(vaddr2_3, vaddr2_3);
110
111                 /* add headroom to pa values */
112                 dma_addr0_1 = _mm256_add_epi64(dma_addr0_1, hdr_room);
113                 dma_addr2_3 = _mm256_add_epi64(dma_addr2_3, hdr_room);
114
115                 /* flush desc with pa dma_addr */
116                 _mm256_store_si256((__m256i *)&rxdp->read, dma_addr0_1);
117                 _mm256_store_si256((__m256i *)&(rxdp + 2)->read, dma_addr2_3);
118         }
119
120 #endif
121
122         rxq->rxrearm_start += IAVF_RXQ_REARM_THRESH;
123         if (rxq->rxrearm_start >= rxq->nb_rx_desc)
124                 rxq->rxrearm_start = 0;
125
126         rxq->rxrearm_nb -= IAVF_RXQ_REARM_THRESH;
127
128         rx_id = (uint16_t)((rxq->rxrearm_start == 0) ?
129                              (rxq->nb_rx_desc - 1) : (rxq->rxrearm_start - 1));
130
131         /* Update the tail pointer on the NIC */
132         IAVF_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
133 }
134
135 #define PKTLEN_SHIFT     10
136
137 static inline uint16_t
138 _iavf_recv_raw_pkts_vec_avx2(struct iavf_rx_queue *rxq,
139                              struct rte_mbuf **rx_pkts,
140                              uint16_t nb_pkts, uint8_t *split_packet)
141 {
142 #define IAVF_DESCS_PER_LOOP_AVX 8
143
144         /* const uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl; */
145         const uint32_t *type_table = rxq->vsi->adapter->ptype_tbl;
146
147         const __m256i mbuf_init = _mm256_set_epi64x(0, 0,
148                         0, rxq->mbuf_initializer);
149         /* struct iavf_rx_entry *sw_ring = &rxq->sw_ring[rxq->rx_tail]; */
150         struct rte_mbuf **sw_ring = &rxq->sw_ring[rxq->rx_tail];
151         volatile union iavf_rx_desc *rxdp = rxq->rx_ring + rxq->rx_tail;
152         const int avx_aligned = ((rxq->rx_tail & 1) == 0);
153
154         rte_prefetch0(rxdp);
155
156         /* nb_pkts has to be floor-aligned to IAVF_DESCS_PER_LOOP_AVX */
157         nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, IAVF_DESCS_PER_LOOP_AVX);
158
159         /* See if we need to rearm the RX queue - gives the prefetch a bit
160          * of time to act
161          */
162         if (rxq->rxrearm_nb > IAVF_RXQ_REARM_THRESH)
163                 iavf_rxq_rearm(rxq);
164
165         /* Before we start moving massive data around, check to see if
166          * there is actually a packet available
167          */
168         if (!(rxdp->wb.qword1.status_error_len &
169                         rte_cpu_to_le_32(1 << IAVF_RX_DESC_STATUS_DD_SHIFT)))
170                 return 0;
171
172         /* constants used in processing loop */
173         const __m256i crc_adjust =
174                 _mm256_set_epi16
175                         (/* first descriptor */
176                          0, 0, 0,       /* ignore non-length fields */
177                          -rxq->crc_len, /* sub crc on data_len */
178                          0,             /* ignore high-16bits of pkt_len */
179                          -rxq->crc_len, /* sub crc on pkt_len */
180                          0, 0,          /* ignore pkt_type field */
181                          /* second descriptor */
182                          0, 0, 0,       /* ignore non-length fields */
183                          -rxq->crc_len, /* sub crc on data_len */
184                          0,             /* ignore high-16bits of pkt_len */
185                          -rxq->crc_len, /* sub crc on pkt_len */
186                          0, 0           /* ignore pkt_type field */
187                         );
188
189         /* 8 packets DD mask, LSB in each 32-bit value */
190         const __m256i dd_check = _mm256_set1_epi32(1);
191
192         /* 8 packets EOP mask, second-LSB in each 32-bit value */
193         const __m256i eop_check = _mm256_slli_epi32(dd_check,
194                         IAVF_RX_DESC_STATUS_EOF_SHIFT);
195
196         /* mask to shuffle from desc. to mbuf (2 descriptors)*/
197         const __m256i shuf_msk =
198                 _mm256_set_epi8
199                         (/* first descriptor */
200                          7, 6, 5, 4,  /* octet 4~7, 32bits rss */
201                          3, 2,        /* octet 2~3, low 16 bits vlan_macip */
202                          15, 14,      /* octet 15~14, 16 bits data_len */
203                          0xFF, 0xFF,  /* skip high 16 bits pkt_len, zero out */
204                          15, 14,      /* octet 15~14, low 16 bits pkt_len */
205                          0xFF, 0xFF,  /* pkt_type set as unknown */
206                          0xFF, 0xFF,  /*pkt_type set as unknown */
207                          /* second descriptor */
208                          7, 6, 5, 4,  /* octet 4~7, 32bits rss */
209                          3, 2,        /* octet 2~3, low 16 bits vlan_macip */
210                          15, 14,      /* octet 15~14, 16 bits data_len */
211                          0xFF, 0xFF,  /* skip high 16 bits pkt_len, zero out */
212                          15, 14,      /* octet 15~14, low 16 bits pkt_len */
213                          0xFF, 0xFF,  /* pkt_type set as unknown */
214                          0xFF, 0xFF   /*pkt_type set as unknown */
215                         );
216         /**
217          * compile-time check the above crc and shuffle layout is correct.
218          * NOTE: the first field (lowest address) is given last in set_epi
219          * calls above.
220          */
221         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
222                         offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
223         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
224                         offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
225         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, vlan_tci) !=
226                         offsetof(struct rte_mbuf, rx_descriptor_fields1) + 10);
227         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, hash) !=
228                         offsetof(struct rte_mbuf, rx_descriptor_fields1) + 12);
229
230         /* Status/Error flag masks */
231         /**
232          * mask everything except RSS, flow director and VLAN flags
233          * bit2 is for VLAN tag, bit11 for flow director indication
234          * bit13:12 for RSS indication. Bits 3-5 of error
235          * field (bits 22-24) are for IP/L4 checksum errors
236          */
237         const __m256i flags_mask =
238                  _mm256_set1_epi32((1 << 2) | (1 << 11) |
239                                    (3 << 12) | (7 << 22));
240         /**
241          * data to be shuffled by result of flag mask. If VLAN bit is set,
242          * (bit 2), then position 4 in this array will be used in the
243          * destination
244          */
245         const __m256i vlan_flags_shuf =
246                 _mm256_set_epi32(0, 0, PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED, 0,
247                                  0, 0, PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED, 0);
248         /**
249          * data to be shuffled by result of flag mask, shifted down 11.
250          * If RSS/FDIR bits are set, shuffle moves appropriate flags in
251          * place.
252          */
253         const __m256i rss_flags_shuf =
254                 _mm256_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
255                                 PKT_RX_RSS_HASH | PKT_RX_FDIR, PKT_RX_RSS_HASH,
256                                 0, 0, 0, 0, PKT_RX_FDIR, 0,/* end up 128-bits */
257                                 0, 0, 0, 0, 0, 0, 0, 0,
258                                 PKT_RX_RSS_HASH | PKT_RX_FDIR, PKT_RX_RSS_HASH,
259                                 0, 0, 0, 0, PKT_RX_FDIR, 0);
260
261         /**
262          * data to be shuffled by the result of the flags mask shifted by 22
263          * bits.  This gives use the l3_l4 flags.
264          */
265         const __m256i l3_l4_flags_shuf = _mm256_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
266                         /* shift right 1 bit to make sure it not exceed 255 */
267                         (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
268                          PKT_RX_IP_CKSUM_BAD) >> 1,
269                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_EIP_CKSUM_BAD |
270                          PKT_RX_L4_CKSUM_BAD) >> 1,
271                         (PKT_RX_EIP_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
272                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_EIP_CKSUM_BAD) >> 1,
273                         (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
274                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD) >> 1,
275                         PKT_RX_IP_CKSUM_BAD >> 1,
276                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1,
277                         /* second 128-bits */
278                         0, 0, 0, 0, 0, 0, 0, 0,
279                         (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
280                          PKT_RX_IP_CKSUM_BAD) >> 1,
281                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_EIP_CKSUM_BAD |
282                          PKT_RX_L4_CKSUM_BAD) >> 1,
283                         (PKT_RX_EIP_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
284                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_EIP_CKSUM_BAD) >> 1,
285                         (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
286                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD) >> 1,
287                         PKT_RX_IP_CKSUM_BAD >> 1,
288                         (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1);
289
290         const __m256i cksum_mask =
291                  _mm256_set1_epi32(PKT_RX_IP_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD |
292                                    PKT_RX_L4_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD |
293                                    PKT_RX_EIP_CKSUM_BAD);
294
295         RTE_SET_USED(avx_aligned); /* for 32B descriptors we don't use this */
296
297         uint16_t i, received;
298
299         for (i = 0, received = 0; i < nb_pkts;
300              i += IAVF_DESCS_PER_LOOP_AVX,
301              rxdp += IAVF_DESCS_PER_LOOP_AVX) {
302                 /* step 1, copy over 8 mbuf pointers to rx_pkts array */
303                 _mm256_storeu_si256((void *)&rx_pkts[i],
304                                     _mm256_loadu_si256((void *)&sw_ring[i]));
305 #ifdef RTE_ARCH_X86_64
306                 _mm256_storeu_si256
307                         ((void *)&rx_pkts[i + 4],
308                          _mm256_loadu_si256((void *)&sw_ring[i + 4]));
309 #endif
310
311                 __m256i raw_desc0_1, raw_desc2_3, raw_desc4_5, raw_desc6_7;
312 #ifdef RTE_LIBRTE_IAVF_16BYTE_RX_DESC
313                 /* for AVX we need alignment otherwise loads are not atomic */
314                 if (avx_aligned) {
315                         /* load in descriptors, 2 at a time, in reverse order */
316                         raw_desc6_7 = _mm256_load_si256((void *)(rxdp + 6));
317                         rte_compiler_barrier();
318                         raw_desc4_5 = _mm256_load_si256((void *)(rxdp + 4));
319                         rte_compiler_barrier();
320                         raw_desc2_3 = _mm256_load_si256((void *)(rxdp + 2));
321                         rte_compiler_barrier();
322                         raw_desc0_1 = _mm256_load_si256((void *)(rxdp + 0));
323                 } else
324 #endif
325                 {
326                         const __m128i raw_desc7 =
327                                 _mm_load_si128((void *)(rxdp + 7));
328                         rte_compiler_barrier();
329                         const __m128i raw_desc6 =
330                                 _mm_load_si128((void *)(rxdp + 6));
331                         rte_compiler_barrier();
332                         const __m128i raw_desc5 =
333                                 _mm_load_si128((void *)(rxdp + 5));
334                         rte_compiler_barrier();
335                         const __m128i raw_desc4 =
336                                 _mm_load_si128((void *)(rxdp + 4));
337                         rte_compiler_barrier();
338                         const __m128i raw_desc3 =
339                                 _mm_load_si128((void *)(rxdp + 3));
340                         rte_compiler_barrier();
341                         const __m128i raw_desc2 =
342                                 _mm_load_si128((void *)(rxdp + 2));
343                         rte_compiler_barrier();
344                         const __m128i raw_desc1 =
345                                 _mm_load_si128((void *)(rxdp + 1));
346                         rte_compiler_barrier();
347                         const __m128i raw_desc0 =
348                                 _mm_load_si128((void *)(rxdp + 0));
349
350                         raw_desc6_7 =
351                                 _mm256_inserti128_si256
352                                         (_mm256_castsi128_si256(raw_desc6),
353                                          raw_desc7, 1);
354                         raw_desc4_5 =
355                                 _mm256_inserti128_si256
356                                         (_mm256_castsi128_si256(raw_desc4),
357                                          raw_desc5, 1);
358                         raw_desc2_3 =
359                                 _mm256_inserti128_si256
360                                         (_mm256_castsi128_si256(raw_desc2),
361                                          raw_desc3, 1);
362                         raw_desc0_1 =
363                                 _mm256_inserti128_si256
364                                         (_mm256_castsi128_si256(raw_desc0),
365                                          raw_desc1, 1);
366                 }
367
368                 if (split_packet) {
369                         int j;
370
371                         for (j = 0; j < IAVF_DESCS_PER_LOOP_AVX; j++)
372                                 rte_mbuf_prefetch_part2(rx_pkts[i + j]);
373                 }
374
375                 /**
376                  * convert descriptors 4-7 into mbufs, adjusting length and
377                  * re-arranging fields. Then write into the mbuf
378                  */
379                 const __m256i len6_7 = _mm256_slli_epi32(raw_desc6_7,
380                                                          PKTLEN_SHIFT);
381                 const __m256i len4_5 = _mm256_slli_epi32(raw_desc4_5,
382                                                          PKTLEN_SHIFT);
383                 const __m256i desc6_7 = _mm256_blend_epi16(raw_desc6_7,
384                                                            len6_7, 0x80);
385                 const __m256i desc4_5 = _mm256_blend_epi16(raw_desc4_5,
386                                                            len4_5, 0x80);
387                 __m256i mb6_7 = _mm256_shuffle_epi8(desc6_7, shuf_msk);
388                 __m256i mb4_5 = _mm256_shuffle_epi8(desc4_5, shuf_msk);
389
390                 mb6_7 = _mm256_add_epi16(mb6_7, crc_adjust);
391                 mb4_5 = _mm256_add_epi16(mb4_5, crc_adjust);
392                 /**
393                  * to get packet types, shift 64-bit values down 30 bits
394                  * and so ptype is in lower 8-bits in each
395                  */
396                 const __m256i ptypes6_7 = _mm256_srli_epi64(desc6_7, 30);
397                 const __m256i ptypes4_5 = _mm256_srli_epi64(desc4_5, 30);
398                 const uint8_t ptype7 = _mm256_extract_epi8(ptypes6_7, 24);
399                 const uint8_t ptype6 = _mm256_extract_epi8(ptypes6_7, 8);
400                 const uint8_t ptype5 = _mm256_extract_epi8(ptypes4_5, 24);
401                 const uint8_t ptype4 = _mm256_extract_epi8(ptypes4_5, 8);
402
403                 mb6_7 = _mm256_insert_epi32(mb6_7, type_table[ptype7], 4);
404                 mb6_7 = _mm256_insert_epi32(mb6_7, type_table[ptype6], 0);
405                 mb4_5 = _mm256_insert_epi32(mb4_5, type_table[ptype5], 4);
406                 mb4_5 = _mm256_insert_epi32(mb4_5, type_table[ptype4], 0);
407                 /* merge the status bits into one register */
408                 const __m256i status4_7 = _mm256_unpackhi_epi32(desc6_7,
409                                 desc4_5);
410
411                 /**
412                  * convert descriptors 0-3 into mbufs, adjusting length and
413                  * re-arranging fields. Then write into the mbuf
414                  */
415                 const __m256i len2_3 = _mm256_slli_epi32(raw_desc2_3,
416                                                          PKTLEN_SHIFT);
417                 const __m256i len0_1 = _mm256_slli_epi32(raw_desc0_1,
418                                                          PKTLEN_SHIFT);
419                 const __m256i desc2_3 = _mm256_blend_epi16(raw_desc2_3,
420                                                            len2_3, 0x80);
421                 const __m256i desc0_1 = _mm256_blend_epi16(raw_desc0_1,
422                                                            len0_1, 0x80);
423                 __m256i mb2_3 = _mm256_shuffle_epi8(desc2_3, shuf_msk);
424                 __m256i mb0_1 = _mm256_shuffle_epi8(desc0_1, shuf_msk);
425
426                 mb2_3 = _mm256_add_epi16(mb2_3, crc_adjust);
427                 mb0_1 = _mm256_add_epi16(mb0_1, crc_adjust);
428                 /* get the packet types */
429                 const __m256i ptypes2_3 = _mm256_srli_epi64(desc2_3, 30);
430                 const __m256i ptypes0_1 = _mm256_srli_epi64(desc0_1, 30);
431                 const uint8_t ptype3 = _mm256_extract_epi8(ptypes2_3, 24);
432                 const uint8_t ptype2 = _mm256_extract_epi8(ptypes2_3, 8);
433                 const uint8_t ptype1 = _mm256_extract_epi8(ptypes0_1, 24);
434                 const uint8_t ptype0 = _mm256_extract_epi8(ptypes0_1, 8);
435
436                 mb2_3 = _mm256_insert_epi32(mb2_3, type_table[ptype3], 4);
437                 mb2_3 = _mm256_insert_epi32(mb2_3, type_table[ptype2], 0);
438                 mb0_1 = _mm256_insert_epi32(mb0_1, type_table[ptype1], 4);
439                 mb0_1 = _mm256_insert_epi32(mb0_1, type_table[ptype0], 0);
440                 /* merge the status bits into one register */
441                 const __m256i status0_3 = _mm256_unpackhi_epi32(desc2_3,
442                                                                 desc0_1);
443
444                 /**
445                  * take the two sets of status bits and merge to one
446                  * After merge, the packets status flags are in the
447                  * order (hi->lo): [1, 3, 5, 7, 0, 2, 4, 6]
448                  */
449                 __m256i status0_7 = _mm256_unpacklo_epi64(status4_7,
450                                                           status0_3);
451
452                 /* now do flag manipulation */
453
454                 /* get only flag/error bits we want */
455                 const __m256i flag_bits =
456                         _mm256_and_si256(status0_7, flags_mask);
457                 /* set vlan and rss flags */
458                 const __m256i vlan_flags =
459                         _mm256_shuffle_epi8(vlan_flags_shuf, flag_bits);
460                 const __m256i rss_flags =
461                         _mm256_shuffle_epi8(rss_flags_shuf,
462                                             _mm256_srli_epi32(flag_bits, 11));
463                 /**
464                  * l3_l4_error flags, shuffle, then shift to correct adjustment
465                  * of flags in flags_shuf, and finally mask out extra bits
466                  */
467                 __m256i l3_l4_flags = _mm256_shuffle_epi8(l3_l4_flags_shuf,
468                                 _mm256_srli_epi32(flag_bits, 22));
469                 l3_l4_flags = _mm256_slli_epi32(l3_l4_flags, 1);
470                 l3_l4_flags = _mm256_and_si256(l3_l4_flags, cksum_mask);
471
472                 /* merge flags */
473                 const __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
474                                 _mm256_or_si256(rss_flags, vlan_flags));
475                 /**
476                  * At this point, we have the 8 sets of flags in the low 16-bits
477                  * of each 32-bit value in vlan0.
478                  * We want to extract these, and merge them with the mbuf init
479                  * data so we can do a single write to the mbuf to set the flags
480                  * and all the other initialization fields. Extracting the
481                  * appropriate flags means that we have to do a shift and blend
482                  * for each mbuf before we do the write. However, we can also
483                  * add in the previously computed rx_descriptor fields to
484                  * make a single 256-bit write per mbuf
485                  */
486                 /* check the structure matches expectations */
487                 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, ol_flags) !=
488                                  offsetof(struct rte_mbuf, rearm_data) + 8);
489                 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, rearm_data) !=
490                                  RTE_ALIGN(offsetof(struct rte_mbuf,
491                                                     rearm_data),
492                                            16));
493                 /* build up data and do writes */
494                 __m256i rearm0, rearm1, rearm2, rearm3, rearm4, rearm5,
495                         rearm6, rearm7;
496                 rearm6 = _mm256_blend_epi32(mbuf_init,
497                                             _mm256_slli_si256(mbuf_flags, 8),
498                                             0x04);
499                 rearm4 = _mm256_blend_epi32(mbuf_init,
500                                             _mm256_slli_si256(mbuf_flags, 4),
501                                             0x04);
502                 rearm2 = _mm256_blend_epi32(mbuf_init, mbuf_flags, 0x04);
503                 rearm0 = _mm256_blend_epi32(mbuf_init,
504                                             _mm256_srli_si256(mbuf_flags, 4),
505                                             0x04);
506                 /* permute to add in the rx_descriptor e.g. rss fields */
507                 rearm6 = _mm256_permute2f128_si256(rearm6, mb6_7, 0x20);
508                 rearm4 = _mm256_permute2f128_si256(rearm4, mb4_5, 0x20);
509                 rearm2 = _mm256_permute2f128_si256(rearm2, mb2_3, 0x20);
510                 rearm0 = _mm256_permute2f128_si256(rearm0, mb0_1, 0x20);
511                 /* write to mbuf */
512                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 6]->rearm_data,
513                                     rearm6);
514                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 4]->rearm_data,
515                                     rearm4);
516                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 2]->rearm_data,
517                                     rearm2);
518                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 0]->rearm_data,
519                                     rearm0);
520
521                 /* repeat for the odd mbufs */
522                 const __m256i odd_flags =
523                         _mm256_castsi128_si256
524                                 (_mm256_extracti128_si256(mbuf_flags, 1));
525                 rearm7 = _mm256_blend_epi32(mbuf_init,
526                                             _mm256_slli_si256(odd_flags, 8),
527                                             0x04);
528                 rearm5 = _mm256_blend_epi32(mbuf_init,
529                                             _mm256_slli_si256(odd_flags, 4),
530                                             0x04);
531                 rearm3 = _mm256_blend_epi32(mbuf_init, odd_flags, 0x04);
532                 rearm1 = _mm256_blend_epi32(mbuf_init,
533                                             _mm256_srli_si256(odd_flags, 4),
534                                             0x04);
535                 /* since odd mbufs are already in hi 128-bits use blend */
536                 rearm7 = _mm256_blend_epi32(rearm7, mb6_7, 0xF0);
537                 rearm5 = _mm256_blend_epi32(rearm5, mb4_5, 0xF0);
538                 rearm3 = _mm256_blend_epi32(rearm3, mb2_3, 0xF0);
539                 rearm1 = _mm256_blend_epi32(rearm1, mb0_1, 0xF0);
540                 /* again write to mbufs */
541                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 7]->rearm_data,
542                                     rearm7);
543                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 5]->rearm_data,
544                                     rearm5);
545                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 3]->rearm_data,
546                                     rearm3);
547                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 1]->rearm_data,
548                                     rearm1);
549
550                 /* extract and record EOP bit */
551                 if (split_packet) {
552                         const __m128i eop_mask =
553                                 _mm_set1_epi16(1 << IAVF_RX_DESC_STATUS_EOF_SHIFT);
554                         const __m256i eop_bits256 = _mm256_and_si256(status0_7,
555                                                                      eop_check);
556                         /* pack status bits into a single 128-bit register */
557                         const __m128i eop_bits =
558                                 _mm_packus_epi32
559                                         (_mm256_castsi256_si128(eop_bits256),
560                                          _mm256_extractf128_si256(eop_bits256,
561                                                                   1));
562                         /**
563                          * flip bits, and mask out the EOP bit, which is now
564                          * a split-packet bit i.e. !EOP, rather than EOP one.
565                          */
566                         __m128i split_bits = _mm_andnot_si128(eop_bits,
567                                         eop_mask);
568                         /**
569                          * eop bits are out of order, so we need to shuffle them
570                          * back into order again. In doing so, only use low 8
571                          * bits, which acts like another pack instruction
572                          * The original order is (hi->lo): 1,3,5,7,0,2,4,6
573                          * [Since we use epi8, the 16-bit positions are
574                          * multiplied by 2 in the eop_shuffle value.]
575                          */
576                         __m128i eop_shuffle =
577                                 _mm_set_epi8(/* zero hi 64b */
578                                              0xFF, 0xFF, 0xFF, 0xFF,
579                                              0xFF, 0xFF, 0xFF, 0xFF,
580                                              /* move values to lo 64b */
581                                              8, 0, 10, 2,
582                                              12, 4, 14, 6);
583                         split_bits = _mm_shuffle_epi8(split_bits, eop_shuffle);
584                         *(uint64_t *)split_packet =
585                                 _mm_cvtsi128_si64(split_bits);
586                         split_packet += IAVF_DESCS_PER_LOOP_AVX;
587                 }
588
589                 /* perform dd_check */
590                 status0_7 = _mm256_and_si256(status0_7, dd_check);
591                 status0_7 = _mm256_packs_epi32(status0_7,
592                                                _mm256_setzero_si256());
593
594                 uint64_t burst = __builtin_popcountll
595                                         (_mm_cvtsi128_si64
596                                                 (_mm256_extracti128_si256
597                                                         (status0_7, 1)));
598                 burst += __builtin_popcountll
599                                 (_mm_cvtsi128_si64
600                                         (_mm256_castsi256_si128(status0_7)));
601                 received += burst;
602                 if (burst != IAVF_DESCS_PER_LOOP_AVX)
603                         break;
604         }
605
606         /* update tail pointers */
607         rxq->rx_tail += received;
608         rxq->rx_tail &= (rxq->nb_rx_desc - 1);
609         if ((rxq->rx_tail & 1) == 1 && received > 1) { /* keep avx2 aligned */
610                 rxq->rx_tail--;
611                 received--;
612         }
613         rxq->rxrearm_nb += received;
614         return received;
615 }
616
617 static inline __m256i
618 flex_rxd_to_fdir_flags_vec_avx2(const __m256i fdir_id0_7)
619 {
620 #define FDID_MIS_MAGIC 0xFFFFFFFF
621         RTE_BUILD_BUG_ON(PKT_RX_FDIR != (1 << 2));
622         RTE_BUILD_BUG_ON(PKT_RX_FDIR_ID != (1 << 13));
623         const __m256i pkt_fdir_bit = _mm256_set1_epi32(PKT_RX_FDIR |
624                         PKT_RX_FDIR_ID);
625         /* desc->flow_id field == 0xFFFFFFFF means fdir mismatch */
626         const __m256i fdir_mis_mask = _mm256_set1_epi32(FDID_MIS_MAGIC);
627         __m256i fdir_mask = _mm256_cmpeq_epi32(fdir_id0_7,
628                         fdir_mis_mask);
629         /* this XOR op results to bit-reverse the fdir_mask */
630         fdir_mask = _mm256_xor_si256(fdir_mask, fdir_mis_mask);
631         const __m256i fdir_flags = _mm256_and_si256(fdir_mask, pkt_fdir_bit);
632
633         return fdir_flags;
634 }
635
636 static inline uint16_t
637 _iavf_recv_raw_pkts_vec_avx2_flex_rxd(struct iavf_rx_queue *rxq,
638                                       struct rte_mbuf **rx_pkts,
639                                       uint16_t nb_pkts, uint8_t *split_packet)
640 {
641 #define IAVF_DESCS_PER_LOOP_AVX 8
642
643         const uint32_t *type_table = rxq->vsi->adapter->ptype_tbl;
644
645         const __m256i mbuf_init = _mm256_set_epi64x(0, 0,
646                         0, rxq->mbuf_initializer);
647         struct rte_mbuf **sw_ring = &rxq->sw_ring[rxq->rx_tail];
648         volatile union iavf_rx_flex_desc *rxdp =
649                 (union iavf_rx_flex_desc *)rxq->rx_ring + rxq->rx_tail;
650
651         rte_prefetch0(rxdp);
652
653         /* nb_pkts has to be floor-aligned to IAVF_DESCS_PER_LOOP_AVX */
654         nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, IAVF_DESCS_PER_LOOP_AVX);
655
656         /* See if we need to rearm the RX queue - gives the prefetch a bit
657          * of time to act
658          */
659         if (rxq->rxrearm_nb > IAVF_RXQ_REARM_THRESH)
660                 iavf_rxq_rearm(rxq);
661
662         /* Before we start moving massive data around, check to see if
663          * there is actually a packet available
664          */
665         if (!(rxdp->wb.status_error0 &
666                         rte_cpu_to_le_32(1 << IAVF_RX_FLEX_DESC_STATUS0_DD_S)))
667                 return 0;
668
669         /* constants used in processing loop */
670         const __m256i crc_adjust =
671                 _mm256_set_epi16
672                         (/* first descriptor */
673                          0, 0, 0,       /* ignore non-length fields */
674                          -rxq->crc_len, /* sub crc on data_len */
675                          0,             /* ignore high-16bits of pkt_len */
676                          -rxq->crc_len, /* sub crc on pkt_len */
677                          0, 0,          /* ignore pkt_type field */
678                          /* second descriptor */
679                          0, 0, 0,       /* ignore non-length fields */
680                          -rxq->crc_len, /* sub crc on data_len */
681                          0,             /* ignore high-16bits of pkt_len */
682                          -rxq->crc_len, /* sub crc on pkt_len */
683                          0, 0           /* ignore pkt_type field */
684                         );
685
686         /* 8 packets DD mask, LSB in each 32-bit value */
687         const __m256i dd_check = _mm256_set1_epi32(1);
688
689         /* 8 packets EOP mask, second-LSB in each 32-bit value */
690         const __m256i eop_check = _mm256_slli_epi32(dd_check,
691                         IAVF_RX_FLEX_DESC_STATUS0_EOF_S);
692
693         /* mask to shuffle from desc. to mbuf (2 descriptors)*/
694         const __m256i shuf_msk =
695                 _mm256_set_epi8
696                         (/* first descriptor */
697                          0xFF, 0xFF,
698                          0xFF, 0xFF,    /* rss hash parsed separately */
699                          11, 10,        /* octet 10~11, 16 bits vlan_macip */
700                          5, 4,          /* octet 4~5, 16 bits data_len */
701                          0xFF, 0xFF,    /* skip hi 16 bits pkt_len, zero out */
702                          5, 4,          /* octet 4~5, 16 bits pkt_len */
703                          0xFF, 0xFF,    /* pkt_type set as unknown */
704                          0xFF, 0xFF,    /*pkt_type set as unknown */
705                          /* second descriptor */
706                          0xFF, 0xFF,
707                          0xFF, 0xFF,    /* rss hash parsed separately */
708                          11, 10,        /* octet 10~11, 16 bits vlan_macip */
709                          5, 4,          /* octet 4~5, 16 bits data_len */
710                          0xFF, 0xFF,    /* skip hi 16 bits pkt_len, zero out */
711                          5, 4,          /* octet 4~5, 16 bits pkt_len */
712                          0xFF, 0xFF,    /* pkt_type set as unknown */
713                          0xFF, 0xFF     /*pkt_type set as unknown */
714                         );
715         /**
716          * compile-time check the above crc and shuffle layout is correct.
717          * NOTE: the first field (lowest address) is given last in set_epi
718          * calls above.
719          */
720         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
721                         offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
722         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
723                         offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
724         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, vlan_tci) !=
725                         offsetof(struct rte_mbuf, rx_descriptor_fields1) + 10);
726         RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, hash) !=
727                         offsetof(struct rte_mbuf, rx_descriptor_fields1) + 12);
728
729         /* Status/Error flag masks */
730         /**
731          * mask everything except Checksum Reports, RSS indication
732          * and VLAN indication.
733          * bit6:4 for IP/L4 checksum errors.
734          * bit12 is for RSS indication.
735          * bit13 is for VLAN indication.
736          */
737         const __m256i flags_mask =
738                  _mm256_set1_epi32((7 << 4) | (1 << 12) | (1 << 13));
739         /**
740          * data to be shuffled by the result of the flags mask shifted by 4
741          * bits.  This gives use the l3_l4 flags.
742          */
743         const __m256i l3_l4_flags_shuf = _mm256_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
744                         /* shift right 1 bit to make sure it not exceed 255 */
745                         (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
746                          PKT_RX_IP_CKSUM_BAD) >> 1,
747                         (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
748                          PKT_RX_IP_CKSUM_GOOD) >> 1,
749                         (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
750                          PKT_RX_IP_CKSUM_BAD) >> 1,
751                         (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
752                          PKT_RX_IP_CKSUM_GOOD) >> 1,
753                         (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
754                         (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1,
755                         (PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1,
756                         (PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_GOOD) >> 1,
757                         /* second 128-bits */
758                         0, 0, 0, 0, 0, 0, 0, 0,
759                         (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
760                          PKT_RX_IP_CKSUM_BAD) >> 1,
761                         (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
762                          PKT_RX_IP_CKSUM_GOOD) >> 1,
763                         (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
764                          PKT_RX_IP_CKSUM_BAD) >> 1,
765                         (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
766                          PKT_RX_IP_CKSUM_GOOD) >> 1,
767                         (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
768                         (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1,
769                         (PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1,
770                         (PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_GOOD) >> 1);
771         const __m256i cksum_mask =
772                  _mm256_set1_epi32(PKT_RX_IP_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD |
773                                    PKT_RX_L4_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD |
774                                    PKT_RX_EIP_CKSUM_BAD);
775         /**
776          * data to be shuffled by result of flag mask, shifted down 12.
777          * If RSS(bit12)/VLAN(bit13) are set,
778          * shuffle moves appropriate flags in place.
779          */
780         const __m256i rss_vlan_flags_shuf = _mm256_set_epi8(0, 0, 0, 0,
781                         0, 0, 0, 0,
782                         0, 0, 0, 0,
783                         PKT_RX_RSS_HASH | PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
784                         PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
785                         PKT_RX_RSS_HASH, 0,
786                         /* end up 128-bits */
787                         0, 0, 0, 0,
788                         0, 0, 0, 0,
789                         0, 0, 0, 0,
790                         PKT_RX_RSS_HASH | PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
791                         PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
792                         PKT_RX_RSS_HASH, 0);
793
794         uint16_t i, received;
795
796         for (i = 0, received = 0; i < nb_pkts;
797              i += IAVF_DESCS_PER_LOOP_AVX,
798              rxdp += IAVF_DESCS_PER_LOOP_AVX) {
799                 /* step 1, copy over 8 mbuf pointers to rx_pkts array */
800                 _mm256_storeu_si256((void *)&rx_pkts[i],
801                                     _mm256_loadu_si256((void *)&sw_ring[i]));
802 #ifdef RTE_ARCH_X86_64
803                 _mm256_storeu_si256
804                         ((void *)&rx_pkts[i + 4],
805                          _mm256_loadu_si256((void *)&sw_ring[i + 4]));
806 #endif
807
808                 __m256i raw_desc0_1, raw_desc2_3, raw_desc4_5, raw_desc6_7;
809
810                 const __m128i raw_desc7 =
811                         _mm_load_si128((void *)(rxdp + 7));
812                 rte_compiler_barrier();
813                 const __m128i raw_desc6 =
814                         _mm_load_si128((void *)(rxdp + 6));
815                 rte_compiler_barrier();
816                 const __m128i raw_desc5 =
817                         _mm_load_si128((void *)(rxdp + 5));
818                 rte_compiler_barrier();
819                 const __m128i raw_desc4 =
820                         _mm_load_si128((void *)(rxdp + 4));
821                 rte_compiler_barrier();
822                 const __m128i raw_desc3 =
823                         _mm_load_si128((void *)(rxdp + 3));
824                 rte_compiler_barrier();
825                 const __m128i raw_desc2 =
826                         _mm_load_si128((void *)(rxdp + 2));
827                 rte_compiler_barrier();
828                 const __m128i raw_desc1 =
829                         _mm_load_si128((void *)(rxdp + 1));
830                 rte_compiler_barrier();
831                 const __m128i raw_desc0 =
832                         _mm_load_si128((void *)(rxdp + 0));
833
834                 raw_desc6_7 =
835                         _mm256_inserti128_si256
836                                 (_mm256_castsi128_si256(raw_desc6),
837                                  raw_desc7, 1);
838                 raw_desc4_5 =
839                         _mm256_inserti128_si256
840                                 (_mm256_castsi128_si256(raw_desc4),
841                                  raw_desc5, 1);
842                 raw_desc2_3 =
843                         _mm256_inserti128_si256
844                                 (_mm256_castsi128_si256(raw_desc2),
845                                  raw_desc3, 1);
846                 raw_desc0_1 =
847                         _mm256_inserti128_si256
848                                 (_mm256_castsi128_si256(raw_desc0),
849                                  raw_desc1, 1);
850
851                 if (split_packet) {
852                         int j;
853
854                         for (j = 0; j < IAVF_DESCS_PER_LOOP_AVX; j++)
855                                 rte_mbuf_prefetch_part2(rx_pkts[i + j]);
856                 }
857
858                 /**
859                  * convert descriptors 4-7 into mbufs, re-arrange fields.
860                  * Then write into the mbuf.
861                  */
862                 __m256i mb6_7 = _mm256_shuffle_epi8(raw_desc6_7, shuf_msk);
863                 __m256i mb4_5 = _mm256_shuffle_epi8(raw_desc4_5, shuf_msk);
864
865                 mb6_7 = _mm256_add_epi16(mb6_7, crc_adjust);
866                 mb4_5 = _mm256_add_epi16(mb4_5, crc_adjust);
867                 /**
868                  * to get packet types, ptype is located in bit16-25
869                  * of each 128bits
870                  */
871                 const __m256i ptype_mask =
872                         _mm256_set1_epi16(IAVF_RX_FLEX_DESC_PTYPE_M);
873                 const __m256i ptypes6_7 =
874                         _mm256_and_si256(raw_desc6_7, ptype_mask);
875                 const __m256i ptypes4_5 =
876                         _mm256_and_si256(raw_desc4_5, ptype_mask);
877                 const uint16_t ptype7 = _mm256_extract_epi16(ptypes6_7, 9);
878                 const uint16_t ptype6 = _mm256_extract_epi16(ptypes6_7, 1);
879                 const uint16_t ptype5 = _mm256_extract_epi16(ptypes4_5, 9);
880                 const uint16_t ptype4 = _mm256_extract_epi16(ptypes4_5, 1);
881
882                 mb6_7 = _mm256_insert_epi32(mb6_7, type_table[ptype7], 4);
883                 mb6_7 = _mm256_insert_epi32(mb6_7, type_table[ptype6], 0);
884                 mb4_5 = _mm256_insert_epi32(mb4_5, type_table[ptype5], 4);
885                 mb4_5 = _mm256_insert_epi32(mb4_5, type_table[ptype4], 0);
886                 /* merge the status bits into one register */
887                 const __m256i status4_7 = _mm256_unpackhi_epi32(raw_desc6_7,
888                                 raw_desc4_5);
889
890                 /**
891                  * convert descriptors 0-3 into mbufs, re-arrange fields.
892                  * Then write into the mbuf.
893                  */
894                 __m256i mb2_3 = _mm256_shuffle_epi8(raw_desc2_3, shuf_msk);
895                 __m256i mb0_1 = _mm256_shuffle_epi8(raw_desc0_1, shuf_msk);
896
897                 mb2_3 = _mm256_add_epi16(mb2_3, crc_adjust);
898                 mb0_1 = _mm256_add_epi16(mb0_1, crc_adjust);
899                 /**
900                  * to get packet types, ptype is located in bit16-25
901                  * of each 128bits
902                  */
903                 const __m256i ptypes2_3 =
904                         _mm256_and_si256(raw_desc2_3, ptype_mask);
905                 const __m256i ptypes0_1 =
906                         _mm256_and_si256(raw_desc0_1, ptype_mask);
907                 const uint16_t ptype3 = _mm256_extract_epi16(ptypes2_3, 9);
908                 const uint16_t ptype2 = _mm256_extract_epi16(ptypes2_3, 1);
909                 const uint16_t ptype1 = _mm256_extract_epi16(ptypes0_1, 9);
910                 const uint16_t ptype0 = _mm256_extract_epi16(ptypes0_1, 1);
911
912                 mb2_3 = _mm256_insert_epi32(mb2_3, type_table[ptype3], 4);
913                 mb2_3 = _mm256_insert_epi32(mb2_3, type_table[ptype2], 0);
914                 mb0_1 = _mm256_insert_epi32(mb0_1, type_table[ptype1], 4);
915                 mb0_1 = _mm256_insert_epi32(mb0_1, type_table[ptype0], 0);
916                 /* merge the status bits into one register */
917                 const __m256i status0_3 = _mm256_unpackhi_epi32(raw_desc2_3,
918                                                                 raw_desc0_1);
919
920                 /**
921                  * take the two sets of status bits and merge to one
922                  * After merge, the packets status flags are in the
923                  * order (hi->lo): [1, 3, 5, 7, 0, 2, 4, 6]
924                  */
925                 __m256i status0_7 = _mm256_unpacklo_epi64(status4_7,
926                                                           status0_3);
927
928                 /* now do flag manipulation */
929
930                 /* get only flag/error bits we want */
931                 const __m256i flag_bits =
932                         _mm256_and_si256(status0_7, flags_mask);
933                 /**
934                  * l3_l4_error flags, shuffle, then shift to correct adjustment
935                  * of flags in flags_shuf, and finally mask out extra bits
936                  */
937                 __m256i l3_l4_flags = _mm256_shuffle_epi8(l3_l4_flags_shuf,
938                                 _mm256_srli_epi32(flag_bits, 4));
939                 l3_l4_flags = _mm256_slli_epi32(l3_l4_flags, 1);
940                 l3_l4_flags = _mm256_and_si256(l3_l4_flags, cksum_mask);
941                 /* set rss and vlan flags */
942                 const __m256i rss_vlan_flag_bits =
943                         _mm256_srli_epi32(flag_bits, 12);
944                 const __m256i rss_vlan_flags =
945                         _mm256_shuffle_epi8(rss_vlan_flags_shuf,
946                                             rss_vlan_flag_bits);
947
948                 /* merge flags */
949                 __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
950                                 rss_vlan_flags);
951
952                 if (rxq->fdir_enabled) {
953                         const __m256i fdir_id4_7 =
954                                 _mm256_unpackhi_epi32(raw_desc6_7, raw_desc4_5);
955
956                         const __m256i fdir_id0_3 =
957                                 _mm256_unpackhi_epi32(raw_desc2_3, raw_desc0_1);
958
959                         const __m256i fdir_id0_7 =
960                                 _mm256_unpackhi_epi64(fdir_id4_7, fdir_id0_3);
961
962                         const __m256i fdir_flags =
963                                 flex_rxd_to_fdir_flags_vec_avx2(fdir_id0_7);
964
965                         /* merge with fdir_flags */
966                         mbuf_flags = _mm256_or_si256(mbuf_flags, fdir_flags);
967
968                         /* write to mbuf: have to use scalar store here */
969                         rx_pkts[i + 0]->hash.fdir.hi =
970                                 _mm256_extract_epi32(fdir_id0_7, 3);
971
972                         rx_pkts[i + 1]->hash.fdir.hi =
973                                 _mm256_extract_epi32(fdir_id0_7, 7);
974
975                         rx_pkts[i + 2]->hash.fdir.hi =
976                                 _mm256_extract_epi32(fdir_id0_7, 2);
977
978                         rx_pkts[i + 3]->hash.fdir.hi =
979                                 _mm256_extract_epi32(fdir_id0_7, 6);
980
981                         rx_pkts[i + 4]->hash.fdir.hi =
982                                 _mm256_extract_epi32(fdir_id0_7, 1);
983
984                         rx_pkts[i + 5]->hash.fdir.hi =
985                                 _mm256_extract_epi32(fdir_id0_7, 5);
986
987                         rx_pkts[i + 6]->hash.fdir.hi =
988                                 _mm256_extract_epi32(fdir_id0_7, 0);
989
990                         rx_pkts[i + 7]->hash.fdir.hi =
991                                 _mm256_extract_epi32(fdir_id0_7, 4);
992                 } /* if() on fdir_enabled */
993
994 #ifndef RTE_LIBRTE_IAVF_16BYTE_RX_DESC
995                 /**
996                  * needs to load 2nd 16B of each desc for RSS hash parsing,
997                  * will cause performance drop to get into this context.
998                  */
999                 if (rxq->vsi->adapter->eth_dev->data->dev_conf.rxmode.offloads &
1000                                 DEV_RX_OFFLOAD_RSS_HASH) {
1001                         /* load bottom half of every 32B desc */
1002                         const __m128i raw_desc_bh7 =
1003                                 _mm_load_si128
1004                                         ((void *)(&rxdp[7].wb.status_error1));
1005                         rte_compiler_barrier();
1006                         const __m128i raw_desc_bh6 =
1007                                 _mm_load_si128
1008                                         ((void *)(&rxdp[6].wb.status_error1));
1009                         rte_compiler_barrier();
1010                         const __m128i raw_desc_bh5 =
1011                                 _mm_load_si128
1012                                         ((void *)(&rxdp[5].wb.status_error1));
1013                         rte_compiler_barrier();
1014                         const __m128i raw_desc_bh4 =
1015                                 _mm_load_si128
1016                                         ((void *)(&rxdp[4].wb.status_error1));
1017                         rte_compiler_barrier();
1018                         const __m128i raw_desc_bh3 =
1019                                 _mm_load_si128
1020                                         ((void *)(&rxdp[3].wb.status_error1));
1021                         rte_compiler_barrier();
1022                         const __m128i raw_desc_bh2 =
1023                                 _mm_load_si128
1024                                         ((void *)(&rxdp[2].wb.status_error1));
1025                         rte_compiler_barrier();
1026                         const __m128i raw_desc_bh1 =
1027                                 _mm_load_si128
1028                                         ((void *)(&rxdp[1].wb.status_error1));
1029                         rte_compiler_barrier();
1030                         const __m128i raw_desc_bh0 =
1031                                 _mm_load_si128
1032                                         ((void *)(&rxdp[0].wb.status_error1));
1033
1034                         __m256i raw_desc_bh6_7 =
1035                                 _mm256_inserti128_si256
1036                                         (_mm256_castsi128_si256(raw_desc_bh6),
1037                                         raw_desc_bh7, 1);
1038                         __m256i raw_desc_bh4_5 =
1039                                 _mm256_inserti128_si256
1040                                         (_mm256_castsi128_si256(raw_desc_bh4),
1041                                         raw_desc_bh5, 1);
1042                         __m256i raw_desc_bh2_3 =
1043                                 _mm256_inserti128_si256
1044                                         (_mm256_castsi128_si256(raw_desc_bh2),
1045                                         raw_desc_bh3, 1);
1046                         __m256i raw_desc_bh0_1 =
1047                                 _mm256_inserti128_si256
1048                                         (_mm256_castsi128_si256(raw_desc_bh0),
1049                                         raw_desc_bh1, 1);
1050
1051                         /**
1052                          * to shift the 32b RSS hash value to the
1053                          * highest 32b of each 128b before mask
1054                          */
1055                         __m256i rss_hash6_7 =
1056                                 _mm256_slli_epi64(raw_desc_bh6_7, 32);
1057                         __m256i rss_hash4_5 =
1058                                 _mm256_slli_epi64(raw_desc_bh4_5, 32);
1059                         __m256i rss_hash2_3 =
1060                                 _mm256_slli_epi64(raw_desc_bh2_3, 32);
1061                         __m256i rss_hash0_1 =
1062                                 _mm256_slli_epi64(raw_desc_bh0_1, 32);
1063
1064                         __m256i rss_hash_msk =
1065                                 _mm256_set_epi32(0xFFFFFFFF, 0, 0, 0,
1066                                                  0xFFFFFFFF, 0, 0, 0);
1067
1068                         rss_hash6_7 = _mm256_and_si256
1069                                         (rss_hash6_7, rss_hash_msk);
1070                         rss_hash4_5 = _mm256_and_si256
1071                                         (rss_hash4_5, rss_hash_msk);
1072                         rss_hash2_3 = _mm256_and_si256
1073                                         (rss_hash2_3, rss_hash_msk);
1074                         rss_hash0_1 = _mm256_and_si256
1075                                         (rss_hash0_1, rss_hash_msk);
1076
1077                         mb6_7 = _mm256_or_si256(mb6_7, rss_hash6_7);
1078                         mb4_5 = _mm256_or_si256(mb4_5, rss_hash4_5);
1079                         mb2_3 = _mm256_or_si256(mb2_3, rss_hash2_3);
1080                         mb0_1 = _mm256_or_si256(mb0_1, rss_hash0_1);
1081                 } /* if() on RSS hash parsing */
1082 #endif
1083
1084                 /**
1085                  * At this point, we have the 8 sets of flags in the low 16-bits
1086                  * of each 32-bit value in vlan0.
1087                  * We want to extract these, and merge them with the mbuf init
1088                  * data so we can do a single write to the mbuf to set the flags
1089                  * and all the other initialization fields. Extracting the
1090                  * appropriate flags means that we have to do a shift and blend
1091                  * for each mbuf before we do the write. However, we can also
1092                  * add in the previously computed rx_descriptor fields to
1093                  * make a single 256-bit write per mbuf
1094                  */
1095                 /* check the structure matches expectations */
1096                 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, ol_flags) !=
1097                                  offsetof(struct rte_mbuf, rearm_data) + 8);
1098                 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, rearm_data) !=
1099                                  RTE_ALIGN(offsetof(struct rte_mbuf,
1100                                                     rearm_data),
1101                                            16));
1102                 /* build up data and do writes */
1103                 __m256i rearm0, rearm1, rearm2, rearm3, rearm4, rearm5,
1104                         rearm6, rearm7;
1105                 rearm6 = _mm256_blend_epi32(mbuf_init,
1106                                             _mm256_slli_si256(mbuf_flags, 8),
1107                                             0x04);
1108                 rearm4 = _mm256_blend_epi32(mbuf_init,
1109                                             _mm256_slli_si256(mbuf_flags, 4),
1110                                             0x04);
1111                 rearm2 = _mm256_blend_epi32(mbuf_init, mbuf_flags, 0x04);
1112                 rearm0 = _mm256_blend_epi32(mbuf_init,
1113                                             _mm256_srli_si256(mbuf_flags, 4),
1114                                             0x04);
1115                 /* permute to add in the rx_descriptor e.g. rss fields */
1116                 rearm6 = _mm256_permute2f128_si256(rearm6, mb6_7, 0x20);
1117                 rearm4 = _mm256_permute2f128_si256(rearm4, mb4_5, 0x20);
1118                 rearm2 = _mm256_permute2f128_si256(rearm2, mb2_3, 0x20);
1119                 rearm0 = _mm256_permute2f128_si256(rearm0, mb0_1, 0x20);
1120                 /* write to mbuf */
1121                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 6]->rearm_data,
1122                                     rearm6);
1123                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 4]->rearm_data,
1124                                     rearm4);
1125                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 2]->rearm_data,
1126                                     rearm2);
1127                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 0]->rearm_data,
1128                                     rearm0);
1129
1130                 /* repeat for the odd mbufs */
1131                 const __m256i odd_flags =
1132                         _mm256_castsi128_si256
1133                                 (_mm256_extracti128_si256(mbuf_flags, 1));
1134                 rearm7 = _mm256_blend_epi32(mbuf_init,
1135                                             _mm256_slli_si256(odd_flags, 8),
1136                                             0x04);
1137                 rearm5 = _mm256_blend_epi32(mbuf_init,
1138                                             _mm256_slli_si256(odd_flags, 4),
1139                                             0x04);
1140                 rearm3 = _mm256_blend_epi32(mbuf_init, odd_flags, 0x04);
1141                 rearm1 = _mm256_blend_epi32(mbuf_init,
1142                                             _mm256_srli_si256(odd_flags, 4),
1143                                             0x04);
1144                 /* since odd mbufs are already in hi 128-bits use blend */
1145                 rearm7 = _mm256_blend_epi32(rearm7, mb6_7, 0xF0);
1146                 rearm5 = _mm256_blend_epi32(rearm5, mb4_5, 0xF0);
1147                 rearm3 = _mm256_blend_epi32(rearm3, mb2_3, 0xF0);
1148                 rearm1 = _mm256_blend_epi32(rearm1, mb0_1, 0xF0);
1149                 /* again write to mbufs */
1150                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 7]->rearm_data,
1151                                     rearm7);
1152                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 5]->rearm_data,
1153                                     rearm5);
1154                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 3]->rearm_data,
1155                                     rearm3);
1156                 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 1]->rearm_data,
1157                                     rearm1);
1158
1159                 /* extract and record EOP bit */
1160                 if (split_packet) {
1161                         const __m128i eop_mask =
1162                                 _mm_set1_epi16(1 <<
1163                                                IAVF_RX_FLEX_DESC_STATUS0_EOF_S);
1164                         const __m256i eop_bits256 = _mm256_and_si256(status0_7,
1165                                                                      eop_check);
1166                         /* pack status bits into a single 128-bit register */
1167                         const __m128i eop_bits =
1168                                 _mm_packus_epi32
1169                                         (_mm256_castsi256_si128(eop_bits256),
1170                                          _mm256_extractf128_si256(eop_bits256,
1171                                                                   1));
1172                         /**
1173                          * flip bits, and mask out the EOP bit, which is now
1174                          * a split-packet bit i.e. !EOP, rather than EOP one.
1175                          */
1176                         __m128i split_bits = _mm_andnot_si128(eop_bits,
1177                                         eop_mask);
1178                         /**
1179                          * eop bits are out of order, so we need to shuffle them
1180                          * back into order again. In doing so, only use low 8
1181                          * bits, which acts like another pack instruction
1182                          * The original order is (hi->lo): 1,3,5,7,0,2,4,6
1183                          * [Since we use epi8, the 16-bit positions are
1184                          * multiplied by 2 in the eop_shuffle value.]
1185                          */
1186                         __m128i eop_shuffle =
1187                                 _mm_set_epi8(/* zero hi 64b */
1188                                              0xFF, 0xFF, 0xFF, 0xFF,
1189                                              0xFF, 0xFF, 0xFF, 0xFF,
1190                                              /* move values to lo 64b */
1191                                              8, 0, 10, 2,
1192                                              12, 4, 14, 6);
1193                         split_bits = _mm_shuffle_epi8(split_bits, eop_shuffle);
1194                         *(uint64_t *)split_packet =
1195                                 _mm_cvtsi128_si64(split_bits);
1196                         split_packet += IAVF_DESCS_PER_LOOP_AVX;
1197                 }
1198
1199                 /* perform dd_check */
1200                 status0_7 = _mm256_and_si256(status0_7, dd_check);
1201                 status0_7 = _mm256_packs_epi32(status0_7,
1202                                                _mm256_setzero_si256());
1203
1204                 uint64_t burst = __builtin_popcountll
1205                                         (_mm_cvtsi128_si64
1206                                                 (_mm256_extracti128_si256
1207                                                         (status0_7, 1)));
1208                 burst += __builtin_popcountll
1209                                 (_mm_cvtsi128_si64
1210                                         (_mm256_castsi256_si128(status0_7)));
1211                 received += burst;
1212                 if (burst != IAVF_DESCS_PER_LOOP_AVX)
1213                         break;
1214         }
1215
1216         /* update tail pointers */
1217         rxq->rx_tail += received;
1218         rxq->rx_tail &= (rxq->nb_rx_desc - 1);
1219         if ((rxq->rx_tail & 1) == 1 && received > 1) { /* keep avx2 aligned */
1220                 rxq->rx_tail--;
1221                 received--;
1222         }
1223         rxq->rxrearm_nb += received;
1224         return received;
1225 }
1226
1227 /**
1228  * Notice:
1229  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
1230  */
1231 uint16_t
1232 iavf_recv_pkts_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts,
1233                         uint16_t nb_pkts)
1234 {
1235         return _iavf_recv_raw_pkts_vec_avx2(rx_queue, rx_pkts, nb_pkts, NULL);
1236 }
1237
1238 /**
1239  * Notice:
1240  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
1241  */
1242 uint16_t
1243 iavf_recv_pkts_vec_avx2_flex_rxd(void *rx_queue, struct rte_mbuf **rx_pkts,
1244                                  uint16_t nb_pkts)
1245 {
1246         return _iavf_recv_raw_pkts_vec_avx2_flex_rxd(rx_queue, rx_pkts,
1247                                                      nb_pkts, NULL);
1248 }
1249
1250 /**
1251  * vPMD receive routine that reassembles single burst of 32 scattered packets
1252  * Notice:
1253  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
1254  */
1255 static uint16_t
1256 iavf_recv_scattered_burst_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts,
1257                                    uint16_t nb_pkts)
1258 {
1259         struct iavf_rx_queue *rxq = rx_queue;
1260         uint8_t split_flags[IAVF_VPMD_RX_MAX_BURST] = {0};
1261
1262         /* get some new buffers */
1263         uint16_t nb_bufs = _iavf_recv_raw_pkts_vec_avx2(rxq, rx_pkts, nb_pkts,
1264                                                        split_flags);
1265         if (nb_bufs == 0)
1266                 return 0;
1267
1268         /* happy day case, full burst + no packets to be joined */
1269         const uint64_t *split_fl64 = (uint64_t *)split_flags;
1270
1271         if (!rxq->pkt_first_seg &&
1272             split_fl64[0] == 0 && split_fl64[1] == 0 &&
1273             split_fl64[2] == 0 && split_fl64[3] == 0)
1274                 return nb_bufs;
1275
1276         /* reassemble any packets that need reassembly*/
1277         unsigned int i = 0;
1278
1279         if (!rxq->pkt_first_seg) {
1280                 /* find the first split flag, and only reassemble then*/
1281                 while (i < nb_bufs && !split_flags[i])
1282                         i++;
1283                 if (i == nb_bufs)
1284                         return nb_bufs;
1285                 rxq->pkt_first_seg = rx_pkts[i];
1286         }
1287         return i + reassemble_packets(rxq, &rx_pkts[i], nb_bufs - i,
1288                                              &split_flags[i]);
1289 }
1290
1291 /**
1292  * vPMD receive routine that reassembles scattered packets.
1293  * Main receive routine that can handle arbitrary burst sizes
1294  * Notice:
1295  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
1296  */
1297 uint16_t
1298 iavf_recv_scattered_pkts_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts,
1299                                   uint16_t nb_pkts)
1300 {
1301         uint16_t retval = 0;
1302
1303         while (nb_pkts > IAVF_VPMD_RX_MAX_BURST) {
1304                 uint16_t burst = iavf_recv_scattered_burst_vec_avx2(rx_queue,
1305                                 rx_pkts + retval, IAVF_VPMD_RX_MAX_BURST);
1306                 retval += burst;
1307                 nb_pkts -= burst;
1308                 if (burst < IAVF_VPMD_RX_MAX_BURST)
1309                         return retval;
1310         }
1311         return retval + iavf_recv_scattered_burst_vec_avx2(rx_queue,
1312                                 rx_pkts + retval, nb_pkts);
1313 }
1314
1315 /**
1316  * vPMD receive routine that reassembles single burst of
1317  * 32 scattered packets for flex RxD
1318  * Notice:
1319  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
1320  */
1321 static uint16_t
1322 iavf_recv_scattered_burst_vec_avx2_flex_rxd(void *rx_queue,
1323                                             struct rte_mbuf **rx_pkts,
1324                                             uint16_t nb_pkts)
1325 {
1326         struct iavf_rx_queue *rxq = rx_queue;
1327         uint8_t split_flags[IAVF_VPMD_RX_MAX_BURST] = {0};
1328
1329         /* get some new buffers */
1330         uint16_t nb_bufs = _iavf_recv_raw_pkts_vec_avx2_flex_rxd(rxq,
1331                                         rx_pkts, nb_pkts, split_flags);
1332         if (nb_bufs == 0)
1333                 return 0;
1334
1335         /* happy day case, full burst + no packets to be joined */
1336         const uint64_t *split_fl64 = (uint64_t *)split_flags;
1337
1338         if (!rxq->pkt_first_seg &&
1339             split_fl64[0] == 0 && split_fl64[1] == 0 &&
1340             split_fl64[2] == 0 && split_fl64[3] == 0)
1341                 return nb_bufs;
1342
1343         /* reassemble any packets that need reassembly*/
1344         unsigned int i = 0;
1345
1346         if (!rxq->pkt_first_seg) {
1347                 /* find the first split flag, and only reassemble then*/
1348                 while (i < nb_bufs && !split_flags[i])
1349                         i++;
1350                 if (i == nb_bufs)
1351                         return nb_bufs;
1352                 rxq->pkt_first_seg = rx_pkts[i];
1353         }
1354         return i + reassemble_packets(rxq, &rx_pkts[i], nb_bufs - i,
1355                                              &split_flags[i]);
1356 }
1357
1358 /**
1359  * vPMD receive routine that reassembles scattered packets for flex RxD.
1360  * Main receive routine that can handle arbitrary burst sizes
1361  * Notice:
1362  * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
1363  */
1364 uint16_t
1365 iavf_recv_scattered_pkts_vec_avx2_flex_rxd(void *rx_queue,
1366                                            struct rte_mbuf **rx_pkts,
1367                                            uint16_t nb_pkts)
1368 {
1369         uint16_t retval = 0;
1370
1371         while (nb_pkts > IAVF_VPMD_RX_MAX_BURST) {
1372                 uint16_t burst =
1373                         iavf_recv_scattered_burst_vec_avx2_flex_rxd
1374                         (rx_queue, rx_pkts + retval, IAVF_VPMD_RX_MAX_BURST);
1375                 retval += burst;
1376                 nb_pkts -= burst;
1377                 if (burst < IAVF_VPMD_RX_MAX_BURST)
1378                         return retval;
1379         }
1380         return retval + iavf_recv_scattered_burst_vec_avx2_flex_rxd(rx_queue,
1381                                 rx_pkts + retval, nb_pkts);
1382 }
1383
1384 static inline void
1385 iavf_vtx1(volatile struct iavf_tx_desc *txdp,
1386           struct rte_mbuf *pkt, uint64_t flags)
1387 {
1388         uint64_t high_qw =
1389                 (IAVF_TX_DESC_DTYPE_DATA |
1390                  ((uint64_t)flags  << IAVF_TXD_QW1_CMD_SHIFT) |
1391                  ((uint64_t)pkt->data_len << IAVF_TXD_QW1_TX_BUF_SZ_SHIFT));
1392
1393         __m128i descriptor = _mm_set_epi64x(high_qw,
1394                                 pkt->buf_iova + pkt->data_off);
1395         _mm_store_si128((__m128i *)txdp, descriptor);
1396 }
1397
1398 static inline void
1399 iavf_vtx(volatile struct iavf_tx_desc *txdp,
1400          struct rte_mbuf **pkt, uint16_t nb_pkts,  uint64_t flags)
1401 {
1402         const uint64_t hi_qw_tmpl = (IAVF_TX_DESC_DTYPE_DATA |
1403                         ((uint64_t)flags  << IAVF_TXD_QW1_CMD_SHIFT));
1404
1405         /* if unaligned on 32-bit boundary, do one to align */
1406         if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
1407                 iavf_vtx1(txdp, *pkt, flags);
1408                 nb_pkts--, txdp++, pkt++;
1409         }
1410
1411         /* do two at a time while possible, in bursts */
1412         for (; nb_pkts > 3; txdp += 4, pkt += 4, nb_pkts -= 4) {
1413                 uint64_t hi_qw3 =
1414                         hi_qw_tmpl |
1415                         ((uint64_t)pkt[3]->data_len <<
1416                          IAVF_TXD_QW1_TX_BUF_SZ_SHIFT);
1417                 uint64_t hi_qw2 =
1418                         hi_qw_tmpl |
1419                         ((uint64_t)pkt[2]->data_len <<
1420                          IAVF_TXD_QW1_TX_BUF_SZ_SHIFT);
1421                 uint64_t hi_qw1 =
1422                         hi_qw_tmpl |
1423                         ((uint64_t)pkt[1]->data_len <<
1424                          IAVF_TXD_QW1_TX_BUF_SZ_SHIFT);
1425                 uint64_t hi_qw0 =
1426                         hi_qw_tmpl |
1427                         ((uint64_t)pkt[0]->data_len <<
1428                          IAVF_TXD_QW1_TX_BUF_SZ_SHIFT);
1429
1430                 __m256i desc2_3 =
1431                         _mm256_set_epi64x
1432                                 (hi_qw3,
1433                                  pkt[3]->buf_iova + pkt[3]->data_off,
1434                                  hi_qw2,
1435                                  pkt[2]->buf_iova + pkt[2]->data_off);
1436                 __m256i desc0_1 =
1437                         _mm256_set_epi64x
1438                                 (hi_qw1,
1439                                  pkt[1]->buf_iova + pkt[1]->data_off,
1440                                  hi_qw0,
1441                                  pkt[0]->buf_iova + pkt[0]->data_off);
1442                 _mm256_store_si256((void *)(txdp + 2), desc2_3);
1443                 _mm256_store_si256((void *)txdp, desc0_1);
1444         }
1445
1446         /* do any last ones */
1447         while (nb_pkts) {
1448                 iavf_vtx1(txdp, *pkt, flags);
1449                 txdp++, pkt++, nb_pkts--;
1450         }
1451 }
1452
1453 static inline uint16_t
1454 iavf_xmit_fixed_burst_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
1455                                uint16_t nb_pkts)
1456 {
1457         struct iavf_tx_queue *txq = (struct iavf_tx_queue *)tx_queue;
1458         volatile struct iavf_tx_desc *txdp;
1459         struct iavf_tx_entry *txep;
1460         uint16_t n, nb_commit, tx_id;
1461         /* bit2 is reserved and must be set to 1 according to Spec */
1462         uint64_t flags = IAVF_TX_DESC_CMD_EOP | IAVF_TX_DESC_CMD_ICRC;
1463         uint64_t rs = IAVF_TX_DESC_CMD_RS | flags;
1464
1465         /* cross rx_thresh boundary is not allowed */
1466         nb_pkts = RTE_MIN(nb_pkts, txq->rs_thresh);
1467
1468         if (txq->nb_free < txq->free_thresh)
1469                 iavf_tx_free_bufs(txq);
1470
1471         nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_free, nb_pkts);
1472         if (unlikely(nb_pkts == 0))
1473                 return 0;
1474
1475         tx_id = txq->tx_tail;
1476         txdp = &txq->tx_ring[tx_id];
1477         txep = &txq->sw_ring[tx_id];
1478
1479         txq->nb_free = (uint16_t)(txq->nb_free - nb_pkts);
1480
1481         n = (uint16_t)(txq->nb_tx_desc - tx_id);
1482         if (nb_commit >= n) {
1483                 tx_backlog_entry(txep, tx_pkts, n);
1484
1485                 iavf_vtx(txdp, tx_pkts, n - 1, flags);
1486                 tx_pkts += (n - 1);
1487                 txdp += (n - 1);
1488
1489                 iavf_vtx1(txdp, *tx_pkts++, rs);
1490
1491                 nb_commit = (uint16_t)(nb_commit - n);
1492
1493                 tx_id = 0;
1494                 txq->next_rs = (uint16_t)(txq->rs_thresh - 1);
1495
1496                 /* avoid reach the end of ring */
1497                 txdp = &txq->tx_ring[tx_id];
1498                 txep = &txq->sw_ring[tx_id];
1499         }
1500
1501         tx_backlog_entry(txep, tx_pkts, nb_commit);
1502
1503         iavf_vtx(txdp, tx_pkts, nb_commit, flags);
1504
1505         tx_id = (uint16_t)(tx_id + nb_commit);
1506         if (tx_id > txq->next_rs) {
1507                 txq->tx_ring[txq->next_rs].cmd_type_offset_bsz |=
1508                         rte_cpu_to_le_64(((uint64_t)IAVF_TX_DESC_CMD_RS) <<
1509                                          IAVF_TXD_QW1_CMD_SHIFT);
1510                 txq->next_rs =
1511                         (uint16_t)(txq->next_rs + txq->rs_thresh);
1512         }
1513
1514         txq->tx_tail = tx_id;
1515
1516         IAVF_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
1517
1518         return nb_pkts;
1519 }
1520
1521 uint16_t
1522 iavf_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
1523                         uint16_t nb_pkts)
1524 {
1525         uint16_t nb_tx = 0;
1526         struct iavf_tx_queue *txq = (struct iavf_tx_queue *)tx_queue;
1527
1528         while (nb_pkts) {
1529                 uint16_t ret, num;
1530
1531                 num = (uint16_t)RTE_MIN(nb_pkts, txq->rs_thresh);
1532                 ret = iavf_xmit_fixed_burst_vec_avx2(tx_queue, &tx_pkts[nb_tx],
1533                                                      num);
1534                 nb_tx += ret;
1535                 nb_pkts -= ret;
1536                 if (ret < num)
1537                         break;
1538         }
1539
1540         return nb_tx;
1541 }