1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2017 Intel Corporation
6 #include <rte_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"
15 #include <x86intrin.h>
17 #ifndef __INTEL_COMPILER
18 #pragma GCC diagnostic ignored "-Wcast-qual"
22 i40e_rxq_rearm(struct i40e_rx_queue *rxq)
26 volatile union i40e_rx_desc *rxdp;
27 struct i40e_rx_entry *rxep = &rxq->sw_ring[rxq->rxrearm_start];
29 rxdp = rxq->rx_ring + rxq->rxrearm_start;
31 /* Pull 'n' more MBUFs into the software ring */
32 if (rte_mempool_get_bulk(rxq->mp,
34 RTE_I40E_RXQ_REARM_THRESH) < 0) {
35 if (rxq->rxrearm_nb + RTE_I40E_RXQ_REARM_THRESH >=
38 dma_addr0 = _mm_setzero_si128();
39 for (i = 0; i < RTE_I40E_DESCS_PER_LOOP; i++) {
40 rxep[i].mbuf = &rxq->fake_mbuf;
41 _mm_store_si128((__m128i *)&rxdp[i].read,
45 rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
46 RTE_I40E_RXQ_REARM_THRESH;
50 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
51 struct rte_mbuf *mb0, *mb1;
52 __m128i dma_addr0, dma_addr1;
53 __m128i hdr_room = _mm_set_epi64x(RTE_PKTMBUF_HEADROOM,
54 RTE_PKTMBUF_HEADROOM);
55 /* Initialize the mbufs in vector, process 2 mbufs in one loop */
56 for (i = 0; i < RTE_I40E_RXQ_REARM_THRESH; i += 2, rxep += 2) {
57 __m128i vaddr0, vaddr1;
62 /* load buf_addr(lo 64bit) and buf_physaddr(hi 64bit) */
63 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, buf_physaddr) !=
64 offsetof(struct rte_mbuf, buf_addr) + 8);
65 vaddr0 = _mm_loadu_si128((__m128i *)&mb0->buf_addr);
66 vaddr1 = _mm_loadu_si128((__m128i *)&mb1->buf_addr);
68 /* convert pa to dma_addr hdr/data */
69 dma_addr0 = _mm_unpackhi_epi64(vaddr0, vaddr0);
70 dma_addr1 = _mm_unpackhi_epi64(vaddr1, vaddr1);
72 /* add headroom to pa values */
73 dma_addr0 = _mm_add_epi64(dma_addr0, hdr_room);
74 dma_addr1 = _mm_add_epi64(dma_addr1, hdr_room);
76 /* flush desc with pa dma_addr */
77 _mm_store_si128((__m128i *)&rxdp++->read, dma_addr0);
78 _mm_store_si128((__m128i *)&rxdp++->read, dma_addr1);
81 struct rte_mbuf *mb0, *mb1, *mb2, *mb3;
82 __m256i dma_addr0_1, dma_addr2_3;
83 __m256i hdr_room = _mm256_set1_epi64x(RTE_PKTMBUF_HEADROOM);
84 /* Initialize the mbufs in vector, process 4 mbufs in one loop */
85 for (i = 0; i < RTE_I40E_RXQ_REARM_THRESH;
86 i += 4, rxep += 4, rxdp += 4) {
87 __m128i vaddr0, vaddr1, vaddr2, vaddr3;
88 __m256i vaddr0_1, vaddr2_3;
95 /* load buf_addr(lo 64bit) and buf_physaddr(hi 64bit) */
96 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, buf_physaddr) !=
97 offsetof(struct rte_mbuf, buf_addr) + 8);
98 vaddr0 = _mm_loadu_si128((__m128i *)&mb0->buf_addr);
99 vaddr1 = _mm_loadu_si128((__m128i *)&mb1->buf_addr);
100 vaddr2 = _mm_loadu_si128((__m128i *)&mb2->buf_addr);
101 vaddr3 = _mm_loadu_si128((__m128i *)&mb3->buf_addr);
104 * merge 0 & 1, by casting 0 to 256-bit and inserting 1
105 * into the high lanes. Similarly for 2 & 3
107 vaddr0_1 = _mm256_inserti128_si256(
108 _mm256_castsi128_si256(vaddr0), vaddr1, 1);
109 vaddr2_3 = _mm256_inserti128_si256(
110 _mm256_castsi128_si256(vaddr2), vaddr3, 1);
112 /* convert pa to dma_addr hdr/data */
113 dma_addr0_1 = _mm256_unpackhi_epi64(vaddr0_1, vaddr0_1);
114 dma_addr2_3 = _mm256_unpackhi_epi64(vaddr2_3, vaddr2_3);
116 /* add headroom to pa values */
117 dma_addr0_1 = _mm256_add_epi64(dma_addr0_1, hdr_room);
118 dma_addr2_3 = _mm256_add_epi64(dma_addr2_3, hdr_room);
120 /* flush desc with pa dma_addr */
121 _mm256_store_si256((__m256i *)&rxdp->read, dma_addr0_1);
122 _mm256_store_si256((__m256i *)&(rxdp + 2)->read, dma_addr2_3);
127 rxq->rxrearm_start += RTE_I40E_RXQ_REARM_THRESH;
128 if (rxq->rxrearm_start >= rxq->nb_rx_desc)
129 rxq->rxrearm_start = 0;
131 rxq->rxrearm_nb -= RTE_I40E_RXQ_REARM_THRESH;
133 rx_id = (uint16_t)((rxq->rxrearm_start == 0) ?
134 (rxq->nb_rx_desc - 1) : (rxq->rxrearm_start - 1));
136 /* Update the tail pointer on the NIC */
137 I40E_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
140 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
141 /* Handles 32B descriptor FDIR ID processing:
142 * rxdp: receive descriptor ring, required to load 2nd 16B half of each desc
143 * rx_pkts: required to store metadata back to mbufs
144 * pkt_idx: offset into the burst, increments in vector widths
145 * desc_idx: required to select the correct shift at compile time
147 static inline __m256i
148 desc_fdir_processing_32b(volatile union i40e_rx_desc *rxdp,
149 struct rte_mbuf **rx_pkts,
150 const uint32_t pkt_idx,
151 const uint32_t desc_idx)
153 /* 32B desc path: load rxdp.wb.qword2 for EXT_STATUS and FLEXBH_STAT */
154 __m128i *rxdp_desc_0 = (void *)(&rxdp[desc_idx + 0].wb.qword2);
155 __m128i *rxdp_desc_1 = (void *)(&rxdp[desc_idx + 1].wb.qword2);
156 const __m128i desc_qw2_0 = _mm_load_si128(rxdp_desc_0);
157 const __m128i desc_qw2_1 = _mm_load_si128(rxdp_desc_1);
159 /* Mask for FLEXBH_STAT, and the FDIR_ID value to compare against. The
160 * remaining data is set to all 1's to pass through data.
162 const __m256i flexbh_mask = _mm256_set_epi32(-1, -1, -1, 3 << 4,
164 const __m256i flexbh_id = _mm256_set_epi32(-1, -1, -1, 1 << 4,
167 /* Load descriptor, check for FLEXBH bits, generate a mask for both
168 * packets in the register.
170 __m256i desc_qw2_0_1 =
171 _mm256_inserti128_si256(_mm256_castsi128_si256(desc_qw2_0),
173 __m256i desc_tmp_msk = _mm256_and_si256(flexbh_mask, desc_qw2_0_1);
174 __m256i fdir_mask = _mm256_cmpeq_epi32(flexbh_id, desc_tmp_msk);
175 __m256i fdir_data = _mm256_alignr_epi8(desc_qw2_0_1, desc_qw2_0_1, 12);
176 __m256i desc_fdir_data = _mm256_and_si256(fdir_mask, fdir_data);
178 /* Write data out to the mbuf. There is no store to this area of the
179 * mbuf today, so we cannot combine it with another store.
181 const uint32_t idx_0 = pkt_idx + desc_idx;
182 const uint32_t idx_1 = pkt_idx + desc_idx + 1;
183 rx_pkts[idx_0]->hash.fdir.hi = _mm256_extract_epi32(desc_fdir_data, 0);
184 rx_pkts[idx_1]->hash.fdir.hi = _mm256_extract_epi32(desc_fdir_data, 4);
186 /* Create mbuf flags as required for mbuf_flags layout
187 * (That's high lane [1,3,5,7, 0,2,4,6] as u32 lanes).
189 * - Mask away bits not required from the fdir_mask
190 * - Leave the PKT_FDIR_ID bit (1 << 13)
191 * - Position that bit correctly based on packet number
192 * - OR in the resulting bit to mbuf_flags
194 RTE_BUILD_BUG_ON(PKT_RX_FDIR_ID != (1 << 13));
195 __m256i mbuf_flag_mask = _mm256_set_epi32(0, 0, 0, 1 << 13,
197 __m256i desc_flag_bit = _mm256_and_si256(mbuf_flag_mask, fdir_mask);
199 /* For static-inline function, this will be stripped out
200 * as the desc_idx is a hard-coded constant.
204 return _mm256_alignr_epi8(desc_flag_bit, desc_flag_bit, 4);
206 return _mm256_alignr_epi8(desc_flag_bit, desc_flag_bit, 8);
208 return _mm256_alignr_epi8(desc_flag_bit, desc_flag_bit, 12);
210 return desc_flag_bit;
215 /* NOT REACHED, see above switch returns */
216 return _mm256_setzero_si256();
218 #endif /* RTE_LIBRTE_I40E_16BYTE_RX_DESC */
220 #define PKTLEN_SHIFT 10
222 /* Force inline as some compilers will not inline by default. */
223 static __rte_always_inline uint16_t
224 _recv_raw_pkts_vec_avx2(struct i40e_rx_queue *rxq, struct rte_mbuf **rx_pkts,
225 uint16_t nb_pkts, uint8_t *split_packet)
227 #define RTE_I40E_DESCS_PER_LOOP_AVX 8
229 const uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
230 const __m256i mbuf_init = _mm256_set_epi64x(0, 0,
231 0, rxq->mbuf_initializer);
232 struct i40e_rx_entry *sw_ring = &rxq->sw_ring[rxq->rx_tail];
233 volatile union i40e_rx_desc *rxdp = rxq->rx_ring + rxq->rx_tail;
234 const int avx_aligned = ((rxq->rx_tail & 1) == 0);
237 /* nb_pkts has to be floor-aligned to RTE_I40E_DESCS_PER_LOOP_AVX */
238 nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_I40E_DESCS_PER_LOOP_AVX);
240 /* See if we need to rearm the RX queue - gives the prefetch a bit
243 if (rxq->rxrearm_nb > RTE_I40E_RXQ_REARM_THRESH)
246 /* Before we start moving massive data around, check to see if
247 * there is actually a packet available
249 if (!(rxdp->wb.qword1.status_error_len &
250 rte_cpu_to_le_32(1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
253 /* constants used in processing loop */
254 const __m256i crc_adjust = _mm256_set_epi16(
255 /* first descriptor */
256 0, 0, 0, /* ignore non-length fields */
257 -rxq->crc_len, /* sub crc on data_len */
258 0, /* ignore high-16bits of pkt_len */
259 -rxq->crc_len, /* sub crc on pkt_len */
260 0, 0, /* ignore pkt_type field */
261 /* second descriptor */
262 0, 0, 0, /* ignore non-length fields */
263 -rxq->crc_len, /* sub crc on data_len */
264 0, /* ignore high-16bits of pkt_len */
265 -rxq->crc_len, /* sub crc on pkt_len */
266 0, 0 /* ignore pkt_type field */
269 /* 8 packets DD mask, LSB in each 32-bit value */
270 const __m256i dd_check = _mm256_set1_epi32(1);
272 /* 8 packets EOP mask, second-LSB in each 32-bit value */
273 const __m256i eop_check = _mm256_slli_epi32(dd_check,
274 I40E_RX_DESC_STATUS_EOF_SHIFT);
276 /* mask to shuffle from desc. to mbuf (2 descriptors)*/
277 const __m256i shuf_msk = _mm256_set_epi8(
278 /* first descriptor */
279 7, 6, 5, 4, /* octet 4~7, 32bits rss */
280 3, 2, /* octet 2~3, low 16 bits vlan_macip */
281 15, 14, /* octet 15~14, 16 bits data_len */
282 0xFF, 0xFF, /* skip high 16 bits pkt_len, zero out */
283 15, 14, /* octet 15~14, low 16 bits pkt_len */
284 0xFF, 0xFF, /* pkt_type set as unknown */
285 0xFF, 0xFF, /*pkt_type set as unknown */
286 /* second descriptor */
287 7, 6, 5, 4, /* octet 4~7, 32bits rss */
288 3, 2, /* octet 2~3, low 16 bits vlan_macip */
289 15, 14, /* octet 15~14, 16 bits data_len */
290 0xFF, 0xFF, /* skip high 16 bits pkt_len, zero out */
291 15, 14, /* octet 15~14, low 16 bits pkt_len */
292 0xFF, 0xFF, /* pkt_type set as unknown */
293 0xFF, 0xFF /*pkt_type set as unknown */
296 * compile-time check the above crc and shuffle layout is correct.
297 * NOTE: the first field (lowest address) is given last in set_epi
300 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
301 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
302 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
303 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
304 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, vlan_tci) !=
305 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 10);
306 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, hash) !=
307 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 12);
309 /* Status/Error flag masks */
311 * mask everything except RSS, flow director and VLAN flags
312 * bit2 is for VLAN tag, bit11 for flow director indication
313 * bit13:12 for RSS indication. Bits 3-5 of error
314 * field (bits 22-24) are for IP/L4 checksum errors
316 const __m256i flags_mask = _mm256_set1_epi32(
317 (1 << 2) | (1 << 11) | (3 << 12) | (7 << 22));
319 * data to be shuffled by result of flag mask. If VLAN bit is set,
320 * (bit 2), then position 4 in this array will be used in the
323 const __m256i vlan_flags_shuf = _mm256_set_epi32(
324 0, 0, PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED, 0,
325 0, 0, PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED, 0);
327 * data to be shuffled by result of flag mask, shifted down 11.
328 * If RSS/FDIR bits are set, shuffle moves appropriate flags in
331 const __m256i rss_flags_shuf = _mm256_set_epi8(
332 0, 0, 0, 0, 0, 0, 0, 0,
333 PKT_RX_RSS_HASH | PKT_RX_FDIR, PKT_RX_RSS_HASH, 0, 0,
334 0, 0, PKT_RX_FDIR, 0, /* end up 128-bits */
335 0, 0, 0, 0, 0, 0, 0, 0,
336 PKT_RX_RSS_HASH | PKT_RX_FDIR, PKT_RX_RSS_HASH, 0, 0,
337 0, 0, PKT_RX_FDIR, 0);
340 * data to be shuffled by the result of the flags mask shifted by 22
341 * bits. This gives use the l3_l4 flags.
343 const __m256i l3_l4_flags_shuf = _mm256_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
344 /* shift right 1 bit to make sure it not exceed 255 */
345 (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
346 (PKT_RX_IP_CKSUM_GOOD | PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD) >> 1,
347 (PKT_RX_EIP_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
348 (PKT_RX_IP_CKSUM_GOOD | PKT_RX_EIP_CKSUM_BAD) >> 1,
349 (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
350 (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD) >> 1,
351 PKT_RX_IP_CKSUM_BAD >> 1,
352 (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1,
353 /* second 128-bits */
354 0, 0, 0, 0, 0, 0, 0, 0,
355 (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
356 (PKT_RX_IP_CKSUM_GOOD | PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD) >> 1,
357 (PKT_RX_EIP_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
358 (PKT_RX_IP_CKSUM_GOOD | PKT_RX_EIP_CKSUM_BAD) >> 1,
359 (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
360 (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD) >> 1,
361 PKT_RX_IP_CKSUM_BAD >> 1,
362 (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1);
364 const __m256i cksum_mask = _mm256_set1_epi32(
365 PKT_RX_IP_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD |
366 PKT_RX_L4_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD |
367 PKT_RX_EIP_CKSUM_BAD);
369 RTE_SET_USED(avx_aligned); /* for 32B descriptors we don't use this */
371 uint16_t i, received;
372 for (i = 0, received = 0; i < nb_pkts;
373 i += RTE_I40E_DESCS_PER_LOOP_AVX,
374 rxdp += RTE_I40E_DESCS_PER_LOOP_AVX) {
375 /* step 1, copy over 8 mbuf pointers to rx_pkts array */
376 _mm256_storeu_si256((void *)&rx_pkts[i],
377 _mm256_loadu_si256((void *)&sw_ring[i]));
378 #ifdef RTE_ARCH_X86_64
379 _mm256_storeu_si256((void *)&rx_pkts[i + 4],
380 _mm256_loadu_si256((void *)&sw_ring[i + 4]));
383 __m256i raw_desc0_1, raw_desc2_3, raw_desc4_5, raw_desc6_7;
384 #ifdef RTE_LIBRTE_I40E_16BYTE_RX_DESC
385 /* for AVX we need alignment otherwise loads are not atomic */
387 /* load in descriptors, 2 at a time, in reverse order */
388 raw_desc6_7 = _mm256_load_si256((void *)(rxdp + 6));
389 rte_compiler_barrier();
390 raw_desc4_5 = _mm256_load_si256((void *)(rxdp + 4));
391 rte_compiler_barrier();
392 raw_desc2_3 = _mm256_load_si256((void *)(rxdp + 2));
393 rte_compiler_barrier();
394 raw_desc0_1 = _mm256_load_si256((void *)(rxdp + 0));
398 const __m128i raw_desc7 = _mm_load_si128((void *)(rxdp + 7));
399 rte_compiler_barrier();
400 const __m128i raw_desc6 = _mm_load_si128((void *)(rxdp + 6));
401 rte_compiler_barrier();
402 const __m128i raw_desc5 = _mm_load_si128((void *)(rxdp + 5));
403 rte_compiler_barrier();
404 const __m128i raw_desc4 = _mm_load_si128((void *)(rxdp + 4));
405 rte_compiler_barrier();
406 const __m128i raw_desc3 = _mm_load_si128((void *)(rxdp + 3));
407 rte_compiler_barrier();
408 const __m128i raw_desc2 = _mm_load_si128((void *)(rxdp + 2));
409 rte_compiler_barrier();
410 const __m128i raw_desc1 = _mm_load_si128((void *)(rxdp + 1));
411 rte_compiler_barrier();
412 const __m128i raw_desc0 = _mm_load_si128((void *)(rxdp + 0));
414 raw_desc6_7 = _mm256_inserti128_si256(
415 _mm256_castsi128_si256(raw_desc6), raw_desc7, 1);
416 raw_desc4_5 = _mm256_inserti128_si256(
417 _mm256_castsi128_si256(raw_desc4), raw_desc5, 1);
418 raw_desc2_3 = _mm256_inserti128_si256(
419 _mm256_castsi128_si256(raw_desc2), raw_desc3, 1);
420 raw_desc0_1 = _mm256_inserti128_si256(
421 _mm256_castsi128_si256(raw_desc0), raw_desc1, 1);
426 for (j = 0; j < RTE_I40E_DESCS_PER_LOOP_AVX; j++)
427 rte_mbuf_prefetch_part2(rx_pkts[i + j]);
431 * convert descriptors 4-7 into mbufs, adjusting length and
432 * re-arranging fields. Then write into the mbuf
434 const __m256i len6_7 = _mm256_slli_epi32(raw_desc6_7, PKTLEN_SHIFT);
435 const __m256i len4_5 = _mm256_slli_epi32(raw_desc4_5, PKTLEN_SHIFT);
436 const __m256i desc6_7 = _mm256_blend_epi16(raw_desc6_7, len6_7, 0x80);
437 const __m256i desc4_5 = _mm256_blend_epi16(raw_desc4_5, len4_5, 0x80);
438 __m256i mb6_7 = _mm256_shuffle_epi8(desc6_7, shuf_msk);
439 __m256i mb4_5 = _mm256_shuffle_epi8(desc4_5, shuf_msk);
440 mb6_7 = _mm256_add_epi16(mb6_7, crc_adjust);
441 mb4_5 = _mm256_add_epi16(mb4_5, crc_adjust);
443 * to get packet types, shift 64-bit values down 30 bits
444 * and so ptype is in lower 8-bits in each
446 const __m256i ptypes6_7 = _mm256_srli_epi64(desc6_7, 30);
447 const __m256i ptypes4_5 = _mm256_srli_epi64(desc4_5, 30);
448 const uint8_t ptype7 = _mm256_extract_epi8(ptypes6_7, 24);
449 const uint8_t ptype6 = _mm256_extract_epi8(ptypes6_7, 8);
450 const uint8_t ptype5 = _mm256_extract_epi8(ptypes4_5, 24);
451 const uint8_t ptype4 = _mm256_extract_epi8(ptypes4_5, 8);
452 mb6_7 = _mm256_insert_epi32(mb6_7, ptype_tbl[ptype7], 4);
453 mb6_7 = _mm256_insert_epi32(mb6_7, ptype_tbl[ptype6], 0);
454 mb4_5 = _mm256_insert_epi32(mb4_5, ptype_tbl[ptype5], 4);
455 mb4_5 = _mm256_insert_epi32(mb4_5, ptype_tbl[ptype4], 0);
456 /* merge the status bits into one register */
457 const __m256i status4_7 = _mm256_unpackhi_epi32(desc6_7,
461 * convert descriptors 0-3 into mbufs, adjusting length and
462 * re-arranging fields. Then write into the mbuf
464 const __m256i len2_3 = _mm256_slli_epi32(raw_desc2_3, PKTLEN_SHIFT);
465 const __m256i len0_1 = _mm256_slli_epi32(raw_desc0_1, PKTLEN_SHIFT);
466 const __m256i desc2_3 = _mm256_blend_epi16(raw_desc2_3, len2_3, 0x80);
467 const __m256i desc0_1 = _mm256_blend_epi16(raw_desc0_1, len0_1, 0x80);
468 __m256i mb2_3 = _mm256_shuffle_epi8(desc2_3, shuf_msk);
469 __m256i mb0_1 = _mm256_shuffle_epi8(desc0_1, shuf_msk);
470 mb2_3 = _mm256_add_epi16(mb2_3, crc_adjust);
471 mb0_1 = _mm256_add_epi16(mb0_1, crc_adjust);
472 /* get the packet types */
473 const __m256i ptypes2_3 = _mm256_srli_epi64(desc2_3, 30);
474 const __m256i ptypes0_1 = _mm256_srli_epi64(desc0_1, 30);
475 const uint8_t ptype3 = _mm256_extract_epi8(ptypes2_3, 24);
476 const uint8_t ptype2 = _mm256_extract_epi8(ptypes2_3, 8);
477 const uint8_t ptype1 = _mm256_extract_epi8(ptypes0_1, 24);
478 const uint8_t ptype0 = _mm256_extract_epi8(ptypes0_1, 8);
479 mb2_3 = _mm256_insert_epi32(mb2_3, ptype_tbl[ptype3], 4);
480 mb2_3 = _mm256_insert_epi32(mb2_3, ptype_tbl[ptype2], 0);
481 mb0_1 = _mm256_insert_epi32(mb0_1, ptype_tbl[ptype1], 4);
482 mb0_1 = _mm256_insert_epi32(mb0_1, ptype_tbl[ptype0], 0);
483 /* merge the status bits into one register */
484 const __m256i status0_3 = _mm256_unpackhi_epi32(desc2_3,
488 * take the two sets of status bits and merge to one
489 * After merge, the packets status flags are in the
490 * order (hi->lo): [1, 3, 5, 7, 0, 2, 4, 6]
492 __m256i status0_7 = _mm256_unpacklo_epi64(status4_7,
495 /* now do flag manipulation */
497 /* get only flag/error bits we want */
498 const __m256i flag_bits = _mm256_and_si256(
499 status0_7, flags_mask);
500 /* set vlan and rss flags */
501 const __m256i vlan_flags = _mm256_shuffle_epi8(
502 vlan_flags_shuf, flag_bits);
503 const __m256i rss_fdir_bits = _mm256_srli_epi32(flag_bits, 11);
504 const __m256i rss_flags = _mm256_shuffle_epi8(rss_flags_shuf,
508 * l3_l4_error flags, shuffle, then shift to correct adjustment
509 * of flags in flags_shuf, and finally mask out extra bits
511 __m256i l3_l4_flags = _mm256_shuffle_epi8(l3_l4_flags_shuf,
512 _mm256_srli_epi32(flag_bits, 22));
513 l3_l4_flags = _mm256_slli_epi32(l3_l4_flags, 1);
514 l3_l4_flags = _mm256_and_si256(l3_l4_flags, cksum_mask);
517 __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags,
518 _mm256_or_si256(rss_flags, vlan_flags));
520 /* If the rxq has FDIR enabled, read and process the FDIR info
521 * from the descriptor. This can cause more loads/stores, so is
522 * not always performed. Branch over the code when not enabled.
524 if (rxq->fdir_enabled) {
525 #ifdef RTE_LIBRTE_I40E_16BYTE_RX_DESC
526 /* 16B descriptor code path:
527 * RSS and FDIR ID use the same offset in the desc, so
528 * only one can be present at a time. The code below
529 * identifies an FDIR ID match, and zeros the RSS value
530 * in the mbuf on FDIR match to keep mbuf data clean.
532 #define FDIR_BLEND_MASK ((1 << 3) | (1 << 7))
535 * - Take flags, shift bits to null out
536 * - CMPEQ with known FDIR ID, to get 0xFFFF or 0 mask
537 * - Strip bits from mask, leaving 0 or 1 for FDIR ID
538 * - Merge with mbuf_flags
540 /* FLM = 1, FLTSTAT = 0b01, (FLM | FLTSTAT) == 3.
541 * Shift left by 28 to avoid having to mask.
543 const __m256i fdir = _mm256_slli_epi32(rss_fdir_bits, 28);
544 const __m256i fdir_id = _mm256_set1_epi32(3 << 28);
546 /* As above, the fdir_mask to packet mapping is this:
547 * order (hi->lo): [1, 3, 5, 7, 0, 2, 4, 6]
548 * Then OR FDIR flags to mbuf_flags on FDIR ID hit.
550 RTE_BUILD_BUG_ON(PKT_RX_FDIR_ID != (1 << 13));
551 const __m256i pkt_fdir_bit = _mm256_set1_epi32(1 << 13);
552 const __m256i fdir_mask = _mm256_cmpeq_epi32(fdir, fdir_id);
553 __m256i fdir_bits = _mm256_and_si256(fdir_mask, pkt_fdir_bit);
554 mbuf_flags = _mm256_or_si256(mbuf_flags, fdir_bits);
556 /* Based on FDIR_MASK, clear the RSS or FDIR value.
557 * The FDIR ID value is masked to zero if not a hit,
558 * otherwise the mb0_1 register RSS field is zeroed.
560 const __m256i fdir_zero_mask = _mm256_setzero_si256();
561 __m256i tmp0_1 = _mm256_blend_epi32(fdir_zero_mask,
562 fdir_mask, FDIR_BLEND_MASK);
563 __m256i fdir_mb0_1 = _mm256_and_si256(mb0_1, fdir_mask);
564 mb0_1 = _mm256_andnot_si256(tmp0_1, mb0_1);
566 /* Write to mbuf: no stores to combine with, so just a
567 * scalar store to push data here.
569 rx_pkts[i + 0]->hash.fdir.hi = _mm256_extract_epi32(fdir_mb0_1, 3);
570 rx_pkts[i + 1]->hash.fdir.hi = _mm256_extract_epi32(fdir_mb0_1, 7);
572 /* Same as above, only shift the fdir_mask to align
573 * the packet FDIR mask with the FDIR_ID desc lane.
575 __m256i tmp2_3 = _mm256_alignr_epi8(fdir_mask, fdir_mask, 12);
576 __m256i fdir_mb2_3 = _mm256_and_si256(mb2_3, tmp2_3);
577 tmp2_3 = _mm256_blend_epi32(fdir_zero_mask, tmp2_3,
579 mb2_3 = _mm256_andnot_si256(tmp2_3, mb2_3);
580 rx_pkts[i + 2]->hash.fdir.hi = _mm256_extract_epi32(fdir_mb2_3, 3);
581 rx_pkts[i + 3]->hash.fdir.hi = _mm256_extract_epi32(fdir_mb2_3, 7);
583 __m256i tmp4_5 = _mm256_alignr_epi8(fdir_mask, fdir_mask, 8);
584 __m256i fdir_mb4_5 = _mm256_and_si256(mb4_5, tmp4_5);
585 tmp4_5 = _mm256_blend_epi32(fdir_zero_mask, tmp4_5,
587 mb4_5 = _mm256_andnot_si256(tmp4_5, mb4_5);
588 rx_pkts[i + 4]->hash.fdir.hi = _mm256_extract_epi32(fdir_mb4_5, 3);
589 rx_pkts[i + 5]->hash.fdir.hi = _mm256_extract_epi32(fdir_mb4_5, 7);
591 __m256i tmp6_7 = _mm256_alignr_epi8(fdir_mask, fdir_mask, 4);
592 __m256i fdir_mb6_7 = _mm256_and_si256(mb6_7, tmp6_7);
593 tmp6_7 = _mm256_blend_epi32(fdir_zero_mask, tmp6_7,
595 mb6_7 = _mm256_andnot_si256(tmp6_7, mb6_7);
596 rx_pkts[i + 6]->hash.fdir.hi = _mm256_extract_epi32(fdir_mb6_7, 3);
597 rx_pkts[i + 7]->hash.fdir.hi = _mm256_extract_epi32(fdir_mb6_7, 7);
599 /* End of 16B descriptor handling */
601 /* 32B descriptor FDIR ID mark handling. Returns bits
602 * to be OR-ed into the mbuf olflags.
604 __m256i fdir_add_flags;
605 fdir_add_flags = desc_fdir_processing_32b(rxdp, rx_pkts, i, 0);
606 mbuf_flags = _mm256_or_si256(mbuf_flags, fdir_add_flags);
608 fdir_add_flags = desc_fdir_processing_32b(rxdp, rx_pkts, i, 2);
609 mbuf_flags = _mm256_or_si256(mbuf_flags, fdir_add_flags);
611 fdir_add_flags = desc_fdir_processing_32b(rxdp, rx_pkts, i, 4);
612 mbuf_flags = _mm256_or_si256(mbuf_flags, fdir_add_flags);
614 fdir_add_flags = desc_fdir_processing_32b(rxdp, rx_pkts, i, 6);
615 mbuf_flags = _mm256_or_si256(mbuf_flags, fdir_add_flags);
616 /* End 32B desc handling */
617 #endif /* RTE_LIBRTE_I40E_16BYTE_RX_DESC */
619 } /* if() on FDIR enabled */
622 * At this point, we have the 8 sets of flags in the low 16-bits
623 * of each 32-bit value in vlan0.
624 * We want to extract these, and merge them with the mbuf init data
625 * so we can do a single write to the mbuf to set the flags
626 * and all the other initialization fields. Extracting the
627 * appropriate flags means that we have to do a shift and blend for
628 * each mbuf before we do the write. However, we can also
629 * add in the previously computed rx_descriptor fields to
630 * make a single 256-bit write per mbuf
632 /* check the structure matches expectations */
633 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, ol_flags) !=
634 offsetof(struct rte_mbuf, rearm_data) + 8);
635 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, rearm_data) !=
636 RTE_ALIGN(offsetof(struct rte_mbuf, rearm_data), 16));
637 /* build up data and do writes */
638 __m256i rearm0, rearm1, rearm2, rearm3, rearm4, rearm5,
640 rearm6 = _mm256_blend_epi32(mbuf_init, _mm256_slli_si256(mbuf_flags, 8), 0x04);
641 rearm4 = _mm256_blend_epi32(mbuf_init, _mm256_slli_si256(mbuf_flags, 4), 0x04);
642 rearm2 = _mm256_blend_epi32(mbuf_init, mbuf_flags, 0x04);
643 rearm0 = _mm256_blend_epi32(mbuf_init, _mm256_srli_si256(mbuf_flags, 4), 0x04);
644 /* permute to add in the rx_descriptor e.g. rss fields */
645 rearm6 = _mm256_permute2f128_si256(rearm6, mb6_7, 0x20);
646 rearm4 = _mm256_permute2f128_si256(rearm4, mb4_5, 0x20);
647 rearm2 = _mm256_permute2f128_si256(rearm2, mb2_3, 0x20);
648 rearm0 = _mm256_permute2f128_si256(rearm0, mb0_1, 0x20);
650 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 6]->rearm_data, rearm6);
651 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 4]->rearm_data, rearm4);
652 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 2]->rearm_data, rearm2);
653 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 0]->rearm_data, rearm0);
655 /* repeat for the odd mbufs */
656 const __m256i odd_flags = _mm256_castsi128_si256(
657 _mm256_extracti128_si256(mbuf_flags, 1));
658 rearm7 = _mm256_blend_epi32(mbuf_init, _mm256_slli_si256(odd_flags, 8), 0x04);
659 rearm5 = _mm256_blend_epi32(mbuf_init, _mm256_slli_si256(odd_flags, 4), 0x04);
660 rearm3 = _mm256_blend_epi32(mbuf_init, odd_flags, 0x04);
661 rearm1 = _mm256_blend_epi32(mbuf_init, _mm256_srli_si256(odd_flags, 4), 0x04);
662 /* since odd mbufs are already in hi 128-bits use blend */
663 rearm7 = _mm256_blend_epi32(rearm7, mb6_7, 0xF0);
664 rearm5 = _mm256_blend_epi32(rearm5, mb4_5, 0xF0);
665 rearm3 = _mm256_blend_epi32(rearm3, mb2_3, 0xF0);
666 rearm1 = _mm256_blend_epi32(rearm1, mb0_1, 0xF0);
667 /* again write to mbufs */
668 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 7]->rearm_data, rearm7);
669 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 5]->rearm_data, rearm5);
670 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 3]->rearm_data, rearm3);
671 _mm256_storeu_si256((__m256i *)&rx_pkts[i + 1]->rearm_data, rearm1);
673 /* extract and record EOP bit */
675 const __m128i eop_mask = _mm_set1_epi16(
676 1 << I40E_RX_DESC_STATUS_EOF_SHIFT);
677 const __m256i eop_bits256 = _mm256_and_si256(status0_7,
679 /* pack status bits into a single 128-bit register */
680 const __m128i eop_bits = _mm_packus_epi32(
681 _mm256_castsi256_si128(eop_bits256),
682 _mm256_extractf128_si256(eop_bits256, 1));
684 * flip bits, and mask out the EOP bit, which is now
685 * a split-packet bit i.e. !EOP, rather than EOP one.
687 __m128i split_bits = _mm_andnot_si128(eop_bits,
690 * eop bits are out of order, so we need to shuffle them
691 * back into order again. In doing so, only use low 8
692 * bits, which acts like another pack instruction
693 * The original order is (hi->lo): 1,3,5,7,0,2,4,6
694 * [Since we use epi8, the 16-bit positions are
695 * multiplied by 2 in the eop_shuffle value.]
697 __m128i eop_shuffle = _mm_set_epi8(
698 0xFF, 0xFF, 0xFF, 0xFF, /* zero hi 64b */
699 0xFF, 0xFF, 0xFF, 0xFF,
700 8, 0, 10, 2, /* move values to lo 64b */
702 split_bits = _mm_shuffle_epi8(split_bits, eop_shuffle);
703 *(uint64_t *)split_packet = _mm_cvtsi128_si64(split_bits);
704 split_packet += RTE_I40E_DESCS_PER_LOOP_AVX;
707 /* perform dd_check */
708 status0_7 = _mm256_and_si256(status0_7, dd_check);
709 status0_7 = _mm256_packs_epi32(status0_7,
710 _mm256_setzero_si256());
712 uint64_t burst = __builtin_popcountll(_mm_cvtsi128_si64(
713 _mm256_extracti128_si256(status0_7, 1)));
714 burst += __builtin_popcountll(_mm_cvtsi128_si64(
715 _mm256_castsi256_si128(status0_7)));
717 if (burst != RTE_I40E_DESCS_PER_LOOP_AVX)
721 /* update tail pointers */
722 rxq->rx_tail += received;
723 rxq->rx_tail &= (rxq->nb_rx_desc - 1);
724 if ((rxq->rx_tail & 1) == 1 && received > 1) { /* keep avx2 aligned */
728 rxq->rxrearm_nb += received;
734 * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
737 i40e_recv_pkts_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts,
740 return _recv_raw_pkts_vec_avx2(rx_queue, rx_pkts, nb_pkts, NULL);
744 * vPMD receive routine that reassembles single burst of 32 scattered packets
746 * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
749 i40e_recv_scattered_burst_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts,
752 struct i40e_rx_queue *rxq = rx_queue;
753 uint8_t split_flags[RTE_I40E_VPMD_RX_BURST] = {0};
755 /* get some new buffers */
756 uint16_t nb_bufs = _recv_raw_pkts_vec_avx2(rxq, rx_pkts, nb_pkts,
761 /* happy day case, full burst + no packets to be joined */
762 const uint64_t *split_fl64 = (uint64_t *)split_flags;
764 if (rxq->pkt_first_seg == NULL &&
765 split_fl64[0] == 0 && split_fl64[1] == 0 &&
766 split_fl64[2] == 0 && split_fl64[3] == 0)
769 /* reassemble any packets that need reassembly*/
772 if (rxq->pkt_first_seg == NULL) {
773 /* find the first split flag, and only reassemble then*/
774 while (i < nb_bufs && !split_flags[i])
778 rxq->pkt_first_seg = rx_pkts[i];
780 return i + reassemble_packets(rxq, &rx_pkts[i], nb_bufs - i,
785 * vPMD receive routine that reassembles scattered packets.
786 * Main receive routine that can handle arbitrary burst sizes
788 * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
791 i40e_recv_scattered_pkts_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts,
795 while (nb_pkts > RTE_I40E_VPMD_RX_BURST) {
796 uint16_t burst = i40e_recv_scattered_burst_vec_avx2(rx_queue,
797 rx_pkts + retval, RTE_I40E_VPMD_RX_BURST);
800 if (burst < RTE_I40E_VPMD_RX_BURST)
803 return retval + i40e_recv_scattered_burst_vec_avx2(rx_queue,
804 rx_pkts + retval, nb_pkts);
809 vtx1(volatile struct i40e_tx_desc *txdp,
810 struct rte_mbuf *pkt, uint64_t flags)
812 uint64_t high_qw = (I40E_TX_DESC_DTYPE_DATA |
813 ((uint64_t)flags << I40E_TXD_QW1_CMD_SHIFT) |
814 ((uint64_t)pkt->data_len << I40E_TXD_QW1_TX_BUF_SZ_SHIFT));
816 __m128i descriptor = _mm_set_epi64x(high_qw,
817 pkt->buf_physaddr + pkt->data_off);
818 _mm_store_si128((__m128i *)txdp, descriptor);
822 vtx(volatile struct i40e_tx_desc *txdp,
823 struct rte_mbuf **pkt, uint16_t nb_pkts, uint64_t flags)
825 const uint64_t hi_qw_tmpl = (I40E_TX_DESC_DTYPE_DATA |
826 ((uint64_t)flags << I40E_TXD_QW1_CMD_SHIFT));
828 /* if unaligned on 32-bit boundary, do one to align */
829 if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) {
830 vtx1(txdp, *pkt, flags);
831 nb_pkts--, txdp++, pkt++;
834 /* do two at a time while possible, in bursts */
835 for (; nb_pkts > 3; txdp += 4, pkt += 4, nb_pkts -= 4) {
836 uint64_t hi_qw3 = hi_qw_tmpl |
837 ((uint64_t)pkt[3]->data_len << I40E_TXD_QW1_TX_BUF_SZ_SHIFT);
838 uint64_t hi_qw2 = hi_qw_tmpl |
839 ((uint64_t)pkt[2]->data_len << I40E_TXD_QW1_TX_BUF_SZ_SHIFT);
840 uint64_t hi_qw1 = hi_qw_tmpl |
841 ((uint64_t)pkt[1]->data_len << I40E_TXD_QW1_TX_BUF_SZ_SHIFT);
842 uint64_t hi_qw0 = hi_qw_tmpl |
843 ((uint64_t)pkt[0]->data_len << I40E_TXD_QW1_TX_BUF_SZ_SHIFT);
845 __m256i desc2_3 = _mm256_set_epi64x(
846 hi_qw3, pkt[3]->buf_physaddr + pkt[3]->data_off,
847 hi_qw2, pkt[2]->buf_physaddr + pkt[2]->data_off);
848 __m256i desc0_1 = _mm256_set_epi64x(
849 hi_qw1, pkt[1]->buf_physaddr + pkt[1]->data_off,
850 hi_qw0, pkt[0]->buf_physaddr + pkt[0]->data_off);
851 _mm256_store_si256((void *)(txdp + 2), desc2_3);
852 _mm256_store_si256((void *)txdp, desc0_1);
855 /* do any last ones */
857 vtx1(txdp, *pkt, flags);
858 txdp++, pkt++, nb_pkts--;
862 static inline uint16_t
863 i40e_xmit_fixed_burst_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
866 struct i40e_tx_queue *txq = (struct i40e_tx_queue *)tx_queue;
867 volatile struct i40e_tx_desc *txdp;
868 struct i40e_tx_entry *txep;
869 uint16_t n, nb_commit, tx_id;
870 uint64_t flags = I40E_TD_CMD;
871 uint64_t rs = I40E_TX_DESC_CMD_RS | I40E_TD_CMD;
873 /* cross rx_thresh boundary is not allowed */
874 nb_pkts = RTE_MIN(nb_pkts, txq->tx_rs_thresh);
876 if (txq->nb_tx_free < txq->tx_free_thresh)
877 i40e_tx_free_bufs(txq);
879 nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
880 if (unlikely(nb_pkts == 0))
883 tx_id = txq->tx_tail;
884 txdp = &txq->tx_ring[tx_id];
885 txep = &txq->sw_ring[tx_id];
887 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
889 n = (uint16_t)(txq->nb_tx_desc - tx_id);
890 if (nb_commit >= n) {
891 tx_backlog_entry(txep, tx_pkts, n);
893 vtx(txdp, tx_pkts, n - 1, flags);
897 vtx1(txdp, *tx_pkts++, rs);
899 nb_commit = (uint16_t)(nb_commit - n);
902 txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
904 /* avoid reach the end of ring */
905 txdp = &txq->tx_ring[tx_id];
906 txep = &txq->sw_ring[tx_id];
909 tx_backlog_entry(txep, tx_pkts, nb_commit);
911 vtx(txdp, tx_pkts, nb_commit, flags);
913 tx_id = (uint16_t)(tx_id + nb_commit);
914 if (tx_id > txq->tx_next_rs) {
915 txq->tx_ring[txq->tx_next_rs].cmd_type_offset_bsz |=
916 rte_cpu_to_le_64(((uint64_t)I40E_TX_DESC_CMD_RS) <<
917 I40E_TXD_QW1_CMD_SHIFT);
919 (uint16_t)(txq->tx_next_rs + txq->tx_rs_thresh);
922 txq->tx_tail = tx_id;
924 I40E_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
930 i40e_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
934 struct i40e_tx_queue *txq = (struct i40e_tx_queue *)tx_queue;
939 num = (uint16_t)RTE_MIN(nb_pkts, txq->tx_rs_thresh);
940 ret = i40e_xmit_fixed_burst_vec_avx2(tx_queue, &tx_pkts[nb_tx],