1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2017 Intel Corporation
6 #include <ethdev_driver.h>
7 #include <rte_malloc.h>
9 #include "base/i40e_prototype.h"
10 #include "base/i40e_type.h"
11 #include "i40e_ethdev.h"
12 #include "i40e_rxtx.h"
13 #include "i40e_rxtx_vec_common.h"
17 #ifndef __INTEL_COMPILER
18 #pragma GCC diagnostic ignored "-Wcast-qual"
21 static __rte_always_inline void
22 i40e_rxq_rearm(struct i40e_rx_queue *rxq)
24 return i40e_rxq_rearm_common(rxq, false);
27 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
28 /* Handles 32B descriptor FDIR ID processing:
29 * rxdp: receive descriptor ring, required to load 2nd 16B half of each desc
30 * rx_pkts: required to store metadata back to mbufs
31 * pkt_idx: offset into the burst, increments in vector widths
32 * desc_idx: required to select the correct shift at compile time
35 desc_fdir_processing_32b(volatile union i40e_rx_desc *rxdp,
36 struct rte_mbuf **rx_pkts,
37 const uint32_t pkt_idx,
38 const uint32_t desc_idx)
40 /* 32B desc path: load rxdp.wb.qword2 for EXT_STATUS and FLEXBH_STAT */
41 __m128i *rxdp_desc_0 = (void *)(&rxdp[desc_idx + 0].wb.qword2);
42 __m128i *rxdp_desc_1 = (void *)(&rxdp[desc_idx + 1].wb.qword2);
43 const __m128i desc_qw2_0 = _mm_load_si128(rxdp_desc_0);
44 const __m128i desc_qw2_1 = _mm_load_si128(rxdp_desc_1);
46 /* Mask for FLEXBH_STAT, and the FDIR_ID value to compare against. The
47 * remaining data is set to all 1's to pass through data.
49 const __m256i flexbh_mask = _mm256_set_epi32(-1, -1, -1, 3 << 4,
51 const __m256i flexbh_id = _mm256_set_epi32(-1, -1, -1, 1 << 4,
54 /* Load descriptor, check for FLEXBH bits, generate a mask for both
55 * packets in the register.
57 __m256i desc_qw2_0_1 =
58 _mm256_inserti128_si256(_mm256_castsi128_si256(desc_qw2_0),
60 __m256i desc_tmp_msk = _mm256_and_si256(flexbh_mask, desc_qw2_0_1);
61 __m256i fdir_mask = _mm256_cmpeq_epi32(flexbh_id, desc_tmp_msk);
62 __m256i fdir_data = _mm256_alignr_epi8(desc_qw2_0_1, desc_qw2_0_1, 12);
63 __m256i desc_fdir_data = _mm256_and_si256(fdir_mask, fdir_data);
65 /* Write data out to the mbuf. There is no store to this area of the
66 * mbuf today, so we cannot combine it with another store.
68 const uint32_t idx_0 = pkt_idx + desc_idx;
69 const uint32_t idx_1 = pkt_idx + desc_idx + 1;
70 rx_pkts[idx_0]->hash.fdir.hi = _mm256_extract_epi32(desc_fdir_data, 0);
71 rx_pkts[idx_1]->hash.fdir.hi = _mm256_extract_epi32(desc_fdir_data, 4);
73 /* Create mbuf flags as required for mbuf_flags layout
74 * (That's high lane [1,3,5,7, 0,2,4,6] as u32 lanes).
76 * - Mask away bits not required from the fdir_mask
77 * - Leave the PKT_FDIR_ID bit (1 << 13)
78 * - Position that bit correctly based on packet number
79 * - OR in the resulting bit to mbuf_flags
81 RTE_BUILD_BUG_ON(RTE_MBUF_F_RX_FDIR_ID != (1 << 13));
82 __m256i mbuf_flag_mask = _mm256_set_epi32(0, 0, 0, 1 << 13,
84 __m256i desc_flag_bit = _mm256_and_si256(mbuf_flag_mask, fdir_mask);
86 /* For static-inline function, this will be stripped out
87 * as the desc_idx is a hard-coded constant.
91 return _mm256_alignr_epi8(desc_flag_bit, desc_flag_bit, 4);
93 return _mm256_alignr_epi8(desc_flag_bit, desc_flag_bit, 8);
95 return _mm256_alignr_epi8(desc_flag_bit, desc_flag_bit, 12);
102 /* NOT REACHED, see above switch returns */
103 return _mm256_setzero_si256();
105 #endif /* RTE_LIBRTE_I40E_16BYTE_RX_DESC */
107 #define PKTLEN_SHIFT 10
109 /* Force inline as some compilers will not inline by default. */
110 static __rte_always_inline uint16_t
111 _recv_raw_pkts_vec_avx2(struct i40e_rx_queue *rxq, struct rte_mbuf **rx_pkts,
112 uint16_t nb_pkts, uint8_t *split_packet)
114 #define RTE_I40E_DESCS_PER_LOOP_AVX 8
116 const uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
117 const __m256i mbuf_init = _mm256_set_epi64x(0, 0,
118 0, rxq->mbuf_initializer);
119 struct i40e_rx_entry *sw_ring = &rxq->sw_ring[rxq->rx_tail];
120 volatile union i40e_rx_desc *rxdp = rxq->rx_ring + rxq->rx_tail;
121 const int avx_aligned = ((rxq->rx_tail & 1) == 0);
124 /* nb_pkts has to be floor-aligned to RTE_I40E_DESCS_PER_LOOP_AVX */
125 nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_I40E_DESCS_PER_LOOP_AVX);
127 /* See if we need to rearm the RX queue - gives the prefetch a bit
130 if (rxq->rxrearm_nb > RTE_I40E_RXQ_REARM_THRESH)
133 /* Before we start moving massive data around, check to see if
134 * there is actually a packet available
136 if (!(rxdp->wb.qword1.status_error_len &
137 rte_cpu_to_le_32(1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
140 /* constants used in processing loop */
141 const __m256i crc_adjust = _mm256_set_epi16(
142 /* first descriptor */
143 0, 0, 0, /* ignore non-length fields */
144 -rxq->crc_len, /* sub crc on data_len */
145 0, /* ignore high-16bits of pkt_len */
146 -rxq->crc_len, /* sub crc on pkt_len */
147 0, 0, /* ignore pkt_type field */
148 /* second descriptor */
149 0, 0, 0, /* ignore non-length fields */
150 -rxq->crc_len, /* sub crc on data_len */
151 0, /* ignore high-16bits of pkt_len */
152 -rxq->crc_len, /* sub crc on pkt_len */
153 0, 0 /* ignore pkt_type field */
156 /* 8 packets DD mask, LSB in each 32-bit value */
157 const __m256i dd_check = _mm256_set1_epi32(1);
159 /* 8 packets EOP mask, second-LSB in each 32-bit value */
160 const __m256i eop_check = _mm256_slli_epi32(dd_check,
161 I40E_RX_DESC_STATUS_EOF_SHIFT);
163 /* mask to shuffle from desc. to mbuf (2 descriptors)*/
164 const __m256i shuf_msk = _mm256_set_epi8(
165 /* first descriptor */
166 7, 6, 5, 4, /* octet 4~7, 32bits rss */
167 3, 2, /* octet 2~3, low 16 bits vlan_macip */
168 15, 14, /* octet 15~14, 16 bits data_len */
169 0xFF, 0xFF, /* skip high 16 bits pkt_len, zero out */
170 15, 14, /* octet 15~14, low 16 bits pkt_len */
171 0xFF, 0xFF, /* pkt_type set as unknown */
172 0xFF, 0xFF, /*pkt_type set as unknown */
173 /* second descriptor */
174 7, 6, 5, 4, /* octet 4~7, 32bits rss */
175 3, 2, /* octet 2~3, low 16 bits vlan_macip */
176 15, 14, /* octet 15~14, 16 bits data_len */
177 0xFF, 0xFF, /* skip high 16 bits pkt_len, zero out */
178 15, 14, /* octet 15~14, low 16 bits pkt_len */
179 0xFF, 0xFF, /* pkt_type set as unknown */
180 0xFF, 0xFF /*pkt_type set as unknown */
183 * compile-time check the above crc and shuffle layout is correct.
184 * NOTE: the first field (lowest address) is given last in set_epi
187 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
188 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
189 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
190 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
191 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, vlan_tci) !=
192 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 10);
193 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, hash) !=
194 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 12);
196 /* Status/Error flag masks */
198 * mask everything except RSS, flow director and VLAN flags
199 * bit2 is for VLAN tag, bit11 for flow director indication
200 * bit13:12 for RSS indication. Bits 3-5 of error
201 * field (bits 22-24) are for IP/L4 checksum errors
203 const __m256i flags_mask = _mm256_set1_epi32(
204 (1 << 2) | (1 << 11) | (3 << 12) | (7 << 22));
206 * data to be shuffled by result of flag mask. If VLAN bit is set,
207 * (bit 2), then position 4 in this array will be used in the
210 const __m256i vlan_flags_shuf = _mm256_set_epi32(
212 RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED, 0,
214 RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED, 0);
216 * data to be shuffled by result of flag mask, shifted down 11.
217 * If RSS/FDIR bits are set, shuffle moves appropriate flags in
220 const __m256i rss_flags_shuf = _mm256_set_epi8(
221 0, 0, 0, 0, 0, 0, 0, 0,
222 RTE_MBUF_F_RX_RSS_HASH | RTE_MBUF_F_RX_FDIR,
223 RTE_MBUF_F_RX_RSS_HASH, 0, 0,
224 0, 0, RTE_MBUF_F_RX_FDIR, 0, /* end up 128-bits */
225 0, 0, 0, 0, 0, 0, 0, 0,
226 RTE_MBUF_F_RX_RSS_HASH | RTE_MBUF_F_RX_FDIR,
227 RTE_MBUF_F_RX_RSS_HASH, 0, 0,
228 0, 0, RTE_MBUF_F_RX_FDIR, 0);
231 * data to be shuffled by the result of the flags mask shifted by 22
232 * bits. This gives use the l3_l4 flags.
234 const __m256i l3_l4_flags_shuf = _mm256_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
235 /* shift right 1 bit to make sure it not exceed 255 */
236 (RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
237 RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
238 (RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
239 RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
240 (RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_L4_CKSUM_GOOD |
241 RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
242 (RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_L4_CKSUM_GOOD |
243 RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
244 (RTE_MBUF_F_RX_L4_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
245 (RTE_MBUF_F_RX_L4_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
246 (RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
247 (RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
248 /* second 128-bits */
249 0, 0, 0, 0, 0, 0, 0, 0,
250 (RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
251 RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
252 (RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
253 RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
254 (RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_L4_CKSUM_GOOD |
255 RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
256 (RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_L4_CKSUM_GOOD |
257 RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
258 (RTE_MBUF_F_RX_L4_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
259 (RTE_MBUF_F_RX_L4_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
260 (RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
261 (RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1);
263 const __m256i cksum_mask = _mm256_set1_epi32(RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_BAD |
264 RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
265 RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD);
267 RTE_SET_USED(avx_aligned); /* for 32B descriptors we don't use this */
269 uint16_t i, received;
270 for (i = 0, received = 0; i < nb_pkts;
271 i += RTE_I40E_DESCS_PER_LOOP_AVX,
272 rxdp += RTE_I40E_DESCS_PER_LOOP_AVX) {
273 /* step 1, copy over 8 mbuf pointers to rx_pkts array */
274 _mm256_storeu_si256((void *)&rx_pkts[i],
275 _mm256_loadu_si256((void *)&sw_ring[i]));
276 #ifdef RTE_ARCH_X86_64
277 _mm256_storeu_si256((void *)&rx_pkts[i + 4],
278 _mm256_loadu_si256((void *)&sw_ring[i + 4]));
281 __m256i raw_desc0_1, raw_desc2_3, raw_desc4_5, raw_desc6_7;
282 #ifdef RTE_LIBRTE_I40E_16BYTE_RX_DESC
283 /* for AVX we need alignment otherwise loads are not atomic */
285 /* load in descriptors, 2 at a time, in reverse order */
286 raw_desc6_7 = _mm256_load_si256((void *)(rxdp + 6));
287 rte_compiler_barrier();
288 raw_desc4_5 = _mm256_load_si256((void *)(rxdp + 4));
289 rte_compiler_barrier();
290 raw_desc2_3 = _mm256_load_si256((void *)(rxdp + 2));
291 rte_compiler_barrier();
292 raw_desc0_1 = _mm256_load_si256((void *)(rxdp + 0));
296 const __m128i raw_desc7 = _mm_load_si128((void *)(rxdp + 7));
297 rte_compiler_barrier();
298 const __m128i raw_desc6 = _mm_load_si128((void *)(rxdp + 6));
299 rte_compiler_barrier();
300 const __m128i raw_desc5 = _mm_load_si128((void *)(rxdp + 5));
301 rte_compiler_barrier();
302 const __m128i raw_desc4 = _mm_load_si128((void *)(rxdp + 4));
303 rte_compiler_barrier();
304 const __m128i raw_desc3 = _mm_load_si128((void *)(rxdp + 3));
305 rte_compiler_barrier();
306 const __m128i raw_desc2 = _mm_load_si128((void *)(rxdp + 2));
307 rte_compiler_barrier();
308 const __m128i raw_desc1 = _mm_load_si128((void *)(rxdp + 1));
309 rte_compiler_barrier();
310 const __m128i raw_desc0 = _mm_load_si128((void *)(rxdp + 0));
312 raw_desc6_7 = _mm256_inserti128_si256(
313 _mm256_castsi128_si256(raw_desc6), raw_desc7, 1);
314 raw_desc4_5 = _mm256_inserti128_si256(
315 _mm256_castsi128_si256(raw_desc4), raw_desc5, 1);
316 raw_desc2_3 = _mm256_inserti128_si256(
317 _mm256_castsi128_si256(raw_desc2), raw_desc3, 1);
318 raw_desc0_1 = _mm256_inserti128_si256(
319 _mm256_castsi128_si256(raw_desc0), raw_desc1, 1);
324 for (j = 0; j < RTE_I40E_DESCS_PER_LOOP_AVX; j++)
325 rte_mbuf_prefetch_part2(rx_pkts[i + j]);
329 * convert descriptors 4-7 into mbufs, adjusting length and
330 * re-arranging fields. Then write into the mbuf
332 const __m256i len6_7 = _mm256_slli_epi32(raw_desc6_7, PKTLEN_SHIFT);
333 const __m256i len4_5 = _mm256_slli_epi32(raw_desc4_5, PKTLEN_SHIFT);
334 const __m256i desc6_7 = _mm256_blend_epi16(raw_desc6_7, len6_7, 0x80);
335 const __m256i desc4_5 = _mm256_blend_epi16(raw_desc4_5, len4_5, 0x80);
336 __m256i mb6_7 = _mm256_shuffle_epi8(desc6_7, shuf_msk);
337 __m256i mb4_5 = _mm256_shuffle_epi8(desc4_5, shuf_msk);
338 mb6_7 = _mm256_add_epi16(mb6_7, crc_adjust);
339 mb4_5 = _mm256_add_epi16(mb4_5, crc_adjust);
341 * to get packet types, shift 64-bit values down 30 bits
342 * and so ptype is in lower 8-bits in each
344 const __m256i ptypes6_7 = _mm256_srli_epi64(desc6_7, 30);
345 const __m256i ptypes4_5 = _mm256_srli_epi64(desc4_5, 30);
346 const uint8_t ptype7 = _mm256_extract_epi8(ptypes6_7, 24);
347 const uint8_t ptype6 = _mm256_extract_epi8(ptypes6_7, 8);
348 const uint8_t ptype5 = _mm256_extract_epi8(ptypes4_5, 24);
349 const uint8_t ptype4 = _mm256_extract_epi8(ptypes4_5, 8);
350 mb6_7 = _mm256_insert_epi32(mb6_7, ptype_tbl[ptype7], 4);
351 mb6_7 = _mm256_insert_epi32(mb6_7, ptype_tbl[ptype6], 0);
352 mb4_5 = _mm256_insert_epi32(mb4_5, ptype_tbl[ptype5], 4);
353 mb4_5 = _mm256_insert_epi32(mb4_5, ptype_tbl[ptype4], 0);
354 /* merge the status bits into one register */
355 const __m256i status4_7 = _mm256_unpackhi_epi32(desc6_7,
359 * convert descriptors 0-3 into mbufs, adjusting length and
360 * re-arranging fields. Then write into the mbuf
362 const __m256i len2_3 = _mm256_slli_epi32(raw_desc2_3, PKTLEN_SHIFT);
363 const __m256i len0_1 = _mm256_slli_epi32(raw_desc0_1, PKTLEN_SHIFT);
364 const __m256i desc2_3 = _mm256_blend_epi16(raw_desc2_3, len2_3, 0x80);
365 const __m256i desc0_1 = _mm256_blend_epi16(raw_desc0_1, len0_1, 0x80);
366 __m256i mb2_3 = _mm256_shuffle_epi8(desc2_3, shuf_msk);
367 __m256i mb0_1 = _mm256_shuffle_epi8(desc0_1, shuf_msk);
368 mb2_3 = _mm256_add_epi16(mb2_3, crc_adjust);
369 mb0_1 = _mm256_add_epi16(mb0_1, crc_adjust);
370 /* get the packet types */
371 const __m256i ptypes2_3 = _mm256_srli_epi64(desc2_3, 30);
372 const __m256i ptypes0_1 = _mm256_srli_epi64(desc0_1, 30);
373 const uint8_t ptype3 = _mm256_extract_epi8(ptypes2_3, 24);
374 const uint8_t ptype2 = _mm256_extract_epi8(ptypes2_3, 8);
375 const uint8_t ptype1 = _mm256_extract_epi8(ptypes0_1, 24);
376 const uint8_t ptype0 = _mm256_extract_epi8(ptypes0_1, 8);
377 mb2_3 = _mm256_insert_epi32(mb2_3, ptype_tbl[ptype3], 4);
378 mb2_3 = _mm256_insert_epi32(mb2_3, ptype_tbl[ptype2], 0);
379 mb0_1 = _mm256_insert_epi32(mb0_1, ptype_tbl[ptype1], 4);
380 mb0_1 = _mm256_insert_epi32(mb0_1, ptype_tbl[ptype0], 0);
381 /* merge the status bits into one register */
382 const __m256i status0_3 = _mm256_unpackhi_epi32(desc2_3,
386 * take the two sets of status bits and merge to one
387 * After merge, the packets status flags are in the
388 * order (hi->lo): [1, 3, 5, 7, 0, 2, 4, 6]
390 __m256i status0_7 = _mm256_unpacklo_epi64(status4_7,
393 /* now do flag manipulation */
395 /* get only flag/error bits we want */
396 const __m256i flag_bits = _mm256_and_si256(
397 status0_7, flags_mask);
398 /* set vlan and rss flags */
399 const __m256i vlan_flags = _mm256_shuffle_epi8(
400 vlan_flags_shuf, flag_bits);
401 const __m256i rss_fdir_bits = _mm256_srli_epi32(flag_bits, 11);
402 const __m256i rss_flags = _mm256_shuffle_epi8(rss_flags_shuf,
406 * l3_l4_error flags, shuffle, then shift to correct adjustment
407 * of flags in flags_shuf, and finally mask out extra bits
409 __m256i l3_l4_flags = _mm256_shuffle_epi8(l3_l4_flags_shuf,
410 _mm256_srli_epi32(flag_bits, 22));
411 l3_l4_flags = _mm256_slli_epi32(l3_l4_flags, 1);
412 l3_l4_flags = _mm256_and_si256(l3_l4_flags, cksum_mask);
415 __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
416 _mm256_or_si256(rss_flags, vlan_flags));
418 /* If the rxq has FDIR enabled, read and process the FDIR info
419 * from the descriptor. This can cause more loads/stores, so is
420 * not always performed. Branch over the code when not enabled.
422 if (rxq->fdir_enabled) {
423 #ifdef RTE_LIBRTE_I40E_16BYTE_RX_DESC
424 /* 16B descriptor code path:
425 * RSS and FDIR ID use the same offset in the desc, so
426 * only one can be present at a time. The code below
427 * identifies an FDIR ID match, and zeros the RSS value
428 * in the mbuf on FDIR match to keep mbuf data clean.
430 #define FDIR_BLEND_MASK ((1 << 3) | (1 << 7))
433 * - Take flags, shift bits to null out
434 * - CMPEQ with known FDIR ID, to get 0xFFFF or 0 mask
435 * - Strip bits from mask, leaving 0 or 1 for FDIR ID
436 * - Merge with mbuf_flags
438 /* FLM = 1, FLTSTAT = 0b01, (FLM | FLTSTAT) == 3.
439 * Shift left by 28 to avoid having to mask.
441 const __m256i fdir = _mm256_slli_epi32(rss_fdir_bits, 28);
442 const __m256i fdir_id = _mm256_set1_epi32(3 << 28);
444 /* As above, the fdir_mask to packet mapping is this:
445 * order (hi->lo): [1, 3, 5, 7, 0, 2, 4, 6]
446 * Then OR FDIR flags to mbuf_flags on FDIR ID hit.
448 RTE_BUILD_BUG_ON(RTE_MBUF_F_RX_FDIR_ID != (1 << 13));
449 const __m256i pkt_fdir_bit = _mm256_set1_epi32(1 << 13);
450 const __m256i fdir_mask = _mm256_cmpeq_epi32(fdir, fdir_id);
451 __m256i fdir_bits = _mm256_and_si256(fdir_mask, pkt_fdir_bit);
452 mbuf_flags = _mm256_or_si256(mbuf_flags, fdir_bits);
454 /* Based on FDIR_MASK, clear the RSS or FDIR value.
455 * The FDIR ID value is masked to zero if not a hit,
456 * otherwise the mb0_1 register RSS field is zeroed.
458 const __m256i fdir_zero_mask = _mm256_setzero_si256();
459 __m256i tmp0_1 = _mm256_blend_epi32(fdir_zero_mask,
460 fdir_mask, FDIR_BLEND_MASK);
461 __m256i fdir_mb0_1 = _mm256_and_si256(mb0_1, fdir_mask);
462 mb0_1 = _mm256_andnot_si256(tmp0_1, mb0_1);
464 /* Write to mbuf: no stores to combine with, so just a
465 * scalar store to push data here.
467 rx_pkts[i + 0]->hash.fdir.hi = _mm256_extract_epi32(fdir_mb0_1, 3);
468 rx_pkts[i + 1]->hash.fdir.hi = _mm256_extract_epi32(fdir_mb0_1, 7);
470 /* Same as above, only shift the fdir_mask to align
471 * the packet FDIR mask with the FDIR_ID desc lane.
473 __m256i tmp2_3 = _mm256_alignr_epi8(fdir_mask, fdir_mask, 12);
474 __m256i fdir_mb2_3 = _mm256_and_si256(mb2_3, tmp2_3);
475 tmp2_3 = _mm256_blend_epi32(fdir_zero_mask, tmp2_3,
477 mb2_3 = _mm256_andnot_si256(tmp2_3, mb2_3);
478 rx_pkts[i + 2]->hash.fdir.hi = _mm256_extract_epi32(fdir_mb2_3, 3);
479 rx_pkts[i + 3]->hash.fdir.hi = _mm256_extract_epi32(fdir_mb2_3, 7);
481 __m256i tmp4_5 = _mm256_alignr_epi8(fdir_mask, fdir_mask, 8);
482 __m256i fdir_mb4_5 = _mm256_and_si256(mb4_5, tmp4_5);
483 tmp4_5 = _mm256_blend_epi32(fdir_zero_mask, tmp4_5,
485 mb4_5 = _mm256_andnot_si256(tmp4_5, mb4_5);
486 rx_pkts[i + 4]->hash.fdir.hi = _mm256_extract_epi32(fdir_mb4_5, 3);
487 rx_pkts[i + 5]->hash.fdir.hi = _mm256_extract_epi32(fdir_mb4_5, 7);
489 __m256i tmp6_7 = _mm256_alignr_epi8(fdir_mask, fdir_mask, 4);
490 __m256i fdir_mb6_7 = _mm256_and_si256(mb6_7, tmp6_7);
491 tmp6_7 = _mm256_blend_epi32(fdir_zero_mask, tmp6_7,
493 mb6_7 = _mm256_andnot_si256(tmp6_7, mb6_7);
494 rx_pkts[i + 6]->hash.fdir.hi = _mm256_extract_epi32(fdir_mb6_7, 3);
495 rx_pkts[i + 7]->hash.fdir.hi = _mm256_extract_epi32(fdir_mb6_7, 7);
497 /* End of 16B descriptor handling */
499 /* 32B descriptor FDIR ID mark handling. Returns bits
500 * to be OR-ed into the mbuf olflags.
502 __m256i fdir_add_flags;
503 fdir_add_flags = desc_fdir_processing_32b(rxdp, rx_pkts, i, 0);
504 mbuf_flags = _mm256_or_si256(mbuf_flags, fdir_add_flags);
506 fdir_add_flags = desc_fdir_processing_32b(rxdp, rx_pkts, i, 2);
507 mbuf_flags = _mm256_or_si256(mbuf_flags, fdir_add_flags);
509 fdir_add_flags = desc_fdir_processing_32b(rxdp, rx_pkts, i, 4);
510 mbuf_flags = _mm256_or_si256(mbuf_flags, fdir_add_flags);
512 fdir_add_flags = desc_fdir_processing_32b(rxdp, rx_pkts, i, 6);
513 mbuf_flags = _mm256_or_si256(mbuf_flags, fdir_add_flags);
514 /* End 32B desc handling */
515 #endif /* RTE_LIBRTE_I40E_16BYTE_RX_DESC */
517 } /* if() on FDIR enabled */
520 * At this point, we have the 8 sets of flags in the low 16-bits
521 * of each 32-bit value in vlan0.
522 * We want to extract these, and merge them with the mbuf init data
523 * so we can do a single write to the mbuf to set the flags
524 * and all the other initialization fields. Extracting the
525 * appropriate flags means that we have to do a shift and blend for
526 * each mbuf before we do the write. However, we can also
527 * add in the previously computed rx_descriptor fields to
528 * make a single 256-bit write per mbuf
530 /* check the structure matches expectations */
531 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, ol_flags) !=
532 offsetof(struct rte_mbuf, rearm_data) + 8);
533 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, rearm_data) !=
534 RTE_ALIGN(offsetof(struct rte_mbuf, rearm_data), 16));
535 /* build up data and do writes */
536 __m256i rearm0, rearm1, rearm2, rearm3, rearm4, rearm5,
538 rearm6 = _mm256_blend_epi32(mbuf_init, _mm256_slli_si256(mbuf_flags, 8), 0x04);
539 rearm4 = _mm256_blend_epi32(mbuf_init, _mm256_slli_si256(mbuf_flags, 4), 0x04);
540 rearm2 = _mm256_blend_epi32(mbuf_init, mbuf_flags, 0x04);
541 rearm0 = _mm256_blend_epi32(mbuf_init, _mm256_srli_si256(mbuf_flags, 4), 0x04);
542 /* permute to add in the rx_descriptor e.g. rss fields */
543 rearm6 = _mm256_permute2f128_si256(rearm6, mb6_7, 0x20);
544 rearm4 = _mm256_permute2f128_si256(rearm4, mb4_5, 0x20);
545 rearm2 = _mm256_permute2f128_si256(rearm2, mb2_3, 0x20);
546 rearm0 = _mm256_permute2f128_si256(rearm0, mb0_1, 0x20);
548 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 6]->rearm_data, rearm6);
549 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 4]->rearm_data, rearm4);
550 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 2]->rearm_data, rearm2);
551 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 0]->rearm_data, rearm0);
553 /* repeat for the odd mbufs */
554 const __m256i odd_flags = _mm256_castsi128_si256(
555 _mm256_extracti128_si256(mbuf_flags, 1));
556 rearm7 = _mm256_blend_epi32(mbuf_init, _mm256_slli_si256(odd_flags, 8), 0x04);
557 rearm5 = _mm256_blend_epi32(mbuf_init, _mm256_slli_si256(odd_flags, 4), 0x04);
558 rearm3 = _mm256_blend_epi32(mbuf_init, odd_flags, 0x04);
559 rearm1 = _mm256_blend_epi32(mbuf_init, _mm256_srli_si256(odd_flags, 4), 0x04);
560 /* since odd mbufs are already in hi 128-bits use blend */
561 rearm7 = _mm256_blend_epi32(rearm7, mb6_7, 0xF0);
562 rearm5 = _mm256_blend_epi32(rearm5, mb4_5, 0xF0);
563 rearm3 = _mm256_blend_epi32(rearm3, mb2_3, 0xF0);
564 rearm1 = _mm256_blend_epi32(rearm1, mb0_1, 0xF0);
565 /* again write to mbufs */
566 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 7]->rearm_data, rearm7);
567 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 5]->rearm_data, rearm5);
568 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 3]->rearm_data, rearm3);
569 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 1]->rearm_data, rearm1);
571 /* extract and record EOP bit */
573 const __m128i eop_mask = _mm_set1_epi16(
574 1 << I40E_RX_DESC_STATUS_EOF_SHIFT);
575 const __m256i eop_bits256 = _mm256_and_si256(status0_7,
577 /* pack status bits into a single 128-bit register */
578 const __m128i eop_bits = _mm_packus_epi32(
579 _mm256_castsi256_si128(eop_bits256),
580 _mm256_extractf128_si256(eop_bits256, 1));
582 * flip bits, and mask out the EOP bit, which is now
583 * a split-packet bit i.e. !EOP, rather than EOP one.
585 __m128i split_bits = _mm_andnot_si128(eop_bits,
588 * eop bits are out of order, so we need to shuffle them
589 * back into order again. In doing so, only use low 8
590 * bits, which acts like another pack instruction
591 * The original order is (hi->lo): 1,3,5,7,0,2,4,6
592 * [Since we use epi8, the 16-bit positions are
593 * multiplied by 2 in the eop_shuffle value.]
595 __m128i eop_shuffle = _mm_set_epi8(
596 0xFF, 0xFF, 0xFF, 0xFF, /* zero hi 64b */
597 0xFF, 0xFF, 0xFF, 0xFF,
598 8, 0, 10, 2, /* move values to lo 64b */
600 split_bits = _mm_shuffle_epi8(split_bits, eop_shuffle);
601 *(uint64_t *)split_packet = _mm_cvtsi128_si64(split_bits);
602 split_packet += RTE_I40E_DESCS_PER_LOOP_AVX;
605 /* perform dd_check */
606 status0_7 = _mm256_and_si256(status0_7, dd_check);
607 status0_7 = _mm256_packs_epi32(status0_7,
608 _mm256_setzero_si256());
610 uint64_t burst = __builtin_popcountll(_mm_cvtsi128_si64(
611 _mm256_extracti128_si256(status0_7, 1)));
612 burst += __builtin_popcountll(_mm_cvtsi128_si64(
613 _mm256_castsi256_si128(status0_7)));
615 if (burst != RTE_I40E_DESCS_PER_LOOP_AVX)
619 /* update tail pointers */
620 rxq->rx_tail += received;
621 rxq->rx_tail &= (rxq->nb_rx_desc - 1);
622 if ((rxq->rx_tail & 1) == 1 && received > 1) { /* keep avx2 aligned */
626 rxq->rxrearm_nb += received;
632 * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
635 i40e_recv_pkts_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts,
638 return _recv_raw_pkts_vec_avx2(rx_queue, rx_pkts, nb_pkts, NULL);
642 * vPMD receive routine that reassembles single burst of 32 scattered packets
644 * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
647 i40e_recv_scattered_burst_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts,
650 struct i40e_rx_queue *rxq = rx_queue;
651 uint8_t split_flags[RTE_I40E_VPMD_RX_BURST] = {0};
653 /* get some new buffers */
654 uint16_t nb_bufs = _recv_raw_pkts_vec_avx2(rxq, rx_pkts, nb_pkts,
659 /* happy day case, full burst + no packets to be joined */
660 const uint64_t *split_fl64 = (uint64_t *)split_flags;
662 if (rxq->pkt_first_seg == NULL &&
663 split_fl64[0] == 0 && split_fl64[1] == 0 &&
664 split_fl64[2] == 0 && split_fl64[3] == 0)
667 /* reassemble any packets that need reassembly*/
670 if (rxq->pkt_first_seg == NULL) {
671 /* find the first split flag, and only reassemble then*/
672 while (i < nb_bufs && !split_flags[i])
676 rxq->pkt_first_seg = rx_pkts[i];
678 return i + reassemble_packets(rxq, &rx_pkts[i], nb_bufs - i,
683 * vPMD receive routine that reassembles scattered packets.
684 * Main receive routine that can handle arbitrary burst sizes
686 * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
689 i40e_recv_scattered_pkts_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts,
693 while (nb_pkts > RTE_I40E_VPMD_RX_BURST) {
694 uint16_t burst = i40e_recv_scattered_burst_vec_avx2(rx_queue,
695 rx_pkts + retval, RTE_I40E_VPMD_RX_BURST);
698 if (burst < RTE_I40E_VPMD_RX_BURST)
701 return retval + i40e_recv_scattered_burst_vec_avx2(rx_queue,
702 rx_pkts + retval, nb_pkts);
707 vtx1(volatile struct i40e_tx_desc *txdp,
708 struct rte_mbuf *pkt, uint64_t flags)
710 uint64_t high_qw = (I40E_TX_DESC_DTYPE_DATA |
711 ((uint64_t)flags << I40E_TXD_QW1_CMD_SHIFT) |
712 ((uint64_t)pkt->data_len << I40E_TXD_QW1_TX_BUF_SZ_SHIFT));
714 __m128i descriptor = _mm_set_epi64x(high_qw,
715 pkt->buf_iova + pkt->data_off);
716 _mm_store_si128((__m128i *)txdp, descriptor);
720 vtx(volatile struct i40e_tx_desc *txdp,
721 struct rte_mbuf **pkt, uint16_t nb_pkts, uint64_t flags)
723 const uint64_t hi_qw_tmpl = (I40E_TX_DESC_DTYPE_DATA |
724 ((uint64_t)flags << I40E_TXD_QW1_CMD_SHIFT));
726 /* if unaligned on 32-bit boundary, do one to align */
727 if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
728 vtx1(txdp, *pkt, flags);
729 nb_pkts--, txdp++, pkt++;
732 /* do two at a time while possible, in bursts */
733 for (; nb_pkts > 3; txdp += 4, pkt += 4, nb_pkts -= 4) {
734 uint64_t hi_qw3 = hi_qw_tmpl |
735 ((uint64_t)pkt[3]->data_len << I40E_TXD_QW1_TX_BUF_SZ_SHIFT);
736 uint64_t hi_qw2 = hi_qw_tmpl |
737 ((uint64_t)pkt[2]->data_len << I40E_TXD_QW1_TX_BUF_SZ_SHIFT);
738 uint64_t hi_qw1 = hi_qw_tmpl |
739 ((uint64_t)pkt[1]->data_len << I40E_TXD_QW1_TX_BUF_SZ_SHIFT);
740 uint64_t hi_qw0 = hi_qw_tmpl |
741 ((uint64_t)pkt[0]->data_len << I40E_TXD_QW1_TX_BUF_SZ_SHIFT);
743 __m256i desc2_3 = _mm256_set_epi64x(
744 hi_qw3, pkt[3]->buf_iova + pkt[3]->data_off,
745 hi_qw2, pkt[2]->buf_iova + pkt[2]->data_off);
746 __m256i desc0_1 = _mm256_set_epi64x(
747 hi_qw1, pkt[1]->buf_iova + pkt[1]->data_off,
748 hi_qw0, pkt[0]->buf_iova + pkt[0]->data_off);
749 _mm256_store_si256((void *)(txdp + 2), desc2_3);
750 _mm256_store_si256((void *)txdp, desc0_1);
753 /* do any last ones */
755 vtx1(txdp, *pkt, flags);
756 txdp++, pkt++, nb_pkts--;
760 static inline uint16_t
761 i40e_xmit_fixed_burst_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
764 struct i40e_tx_queue *txq = (struct i40e_tx_queue *)tx_queue;
765 volatile struct i40e_tx_desc *txdp;
766 struct i40e_tx_entry *txep;
767 uint16_t n, nb_commit, tx_id;
768 uint64_t flags = I40E_TD_CMD;
769 uint64_t rs = I40E_TX_DESC_CMD_RS | I40E_TD_CMD;
771 /* cross rx_thresh boundary is not allowed */
772 nb_pkts = RTE_MIN(nb_pkts, txq->tx_rs_thresh);
774 if (txq->nb_tx_free < txq->tx_free_thresh)
775 i40e_tx_free_bufs(txq);
777 nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
778 if (unlikely(nb_pkts == 0))
781 tx_id = txq->tx_tail;
782 txdp = &txq->tx_ring[tx_id];
783 txep = &txq->sw_ring[tx_id];
785 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
787 n = (uint16_t)(txq->nb_tx_desc - tx_id);
788 if (nb_commit >= n) {
789 tx_backlog_entry(txep, tx_pkts, n);
791 vtx(txdp, tx_pkts, n - 1, flags);
795 vtx1(txdp, *tx_pkts++, rs);
797 nb_commit = (uint16_t)(nb_commit - n);
800 txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
802 /* avoid reach the end of ring */
803 txdp = &txq->tx_ring[tx_id];
804 txep = &txq->sw_ring[tx_id];
807 tx_backlog_entry(txep, tx_pkts, nb_commit);
809 vtx(txdp, tx_pkts, nb_commit, flags);
811 tx_id = (uint16_t)(tx_id + nb_commit);
812 if (tx_id > txq->tx_next_rs) {
813 txq->tx_ring[txq->tx_next_rs].cmd_type_offset_bsz |=
814 rte_cpu_to_le_64(((uint64_t)I40E_TX_DESC_CMD_RS) <<
815 I40E_TXD_QW1_CMD_SHIFT);
817 (uint16_t)(txq->tx_next_rs + txq->tx_rs_thresh);
820 txq->tx_tail = tx_id;
822 I40E_PCI_REG_WC_WRITE(txq->qtx_tail, txq->tx_tail);
828 i40e_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
832 struct i40e_tx_queue *txq = (struct i40e_tx_queue *)tx_queue;
837 num = (uint16_t)RTE_MIN(nb_pkts, txq->tx_rs_thresh);
838 ret = i40e_xmit_fixed_burst_vec_avx2(tx_queue, &tx_pkts[nb_tx],