xref: /dpdk/drivers/net/intel/i40e/i40e_rxtx_vec_neon.c (revision 5cc9919fd443fbd3fce77a257601890a0ee6a247)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2010-2015 Intel Corporation.
3  * Copyright(c) 2016-2018, Linaro Limited.
4  */
5 
6 #include <stdint.h>
7 
8 #include <ethdev_driver.h>
9 #include <rte_bitops.h>
10 #include <rte_malloc.h>
11 #include <rte_vect.h>
12 
13 #include "base/i40e_prototype.h"
14 #include "base/i40e_type.h"
15 #include "i40e_ethdev.h"
16 #include "i40e_rxtx.h"
17 #include "i40e_rxtx_vec_common.h"
18 
19 static inline void
20 i40e_rxq_rearm(struct i40e_rx_queue *rxq)
21 {
22 	int i;
23 	uint16_t rx_id;
24 	volatile union i40e_rx_desc *rxdp;
25 	struct i40e_rx_entry *rxep = &rxq->sw_ring[rxq->rxrearm_start];
26 	struct rte_mbuf *mb0, *mb1;
27 	uint64x2_t dma_addr0, dma_addr1;
28 	uint64x2_t zero = vdupq_n_u64(0);
29 	uint64_t paddr;
30 
31 	rxdp = rxq->rx_ring + rxq->rxrearm_start;
32 
33 	/* Pull 'n' more MBUFs into the software ring */
34 	if (unlikely(rte_mempool_get_bulk(rxq->mp,
35 					  (void *)rxep,
36 					  RTE_I40E_RXQ_REARM_THRESH) < 0)) {
37 		if (rxq->rxrearm_nb + RTE_I40E_RXQ_REARM_THRESH >=
38 		    rxq->nb_rx_desc) {
39 			for (i = 0; i < RTE_I40E_DESCS_PER_LOOP; i++) {
40 				rxep[i].mbuf = &rxq->fake_mbuf;
41 				vst1q_u64(RTE_CAST_PTR(uint64_t *, &rxdp[i].read), zero);
42 			}
43 		}
44 		rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
45 			RTE_I40E_RXQ_REARM_THRESH;
46 		return;
47 	}
48 
49 	/* Initialize the mbufs in vector, process 2 mbufs in one loop */
50 	for (i = 0; i < RTE_I40E_RXQ_REARM_THRESH; i += 2, rxep += 2) {
51 		mb0 = rxep[0].mbuf;
52 		mb1 = rxep[1].mbuf;
53 
54 		paddr = mb0->buf_iova + RTE_PKTMBUF_HEADROOM;
55 		dma_addr0 = vdupq_n_u64(paddr);
56 
57 		/* flush desc with pa dma_addr */
58 		vst1q_u64(RTE_CAST_PTR(uint64_t *, &rxdp++->read), dma_addr0);
59 
60 		paddr = mb1->buf_iova + RTE_PKTMBUF_HEADROOM;
61 		dma_addr1 = vdupq_n_u64(paddr);
62 		vst1q_u64(RTE_CAST_PTR(uint64_t *, &rxdp++->read), dma_addr1);
63 	}
64 
65 	rxq->rxrearm_start += RTE_I40E_RXQ_REARM_THRESH;
66 	rx_id = rxq->rxrearm_start - 1;
67 
68 	if (unlikely(rxq->rxrearm_start >= rxq->nb_rx_desc)) {
69 		rxq->rxrearm_start = 0;
70 		rx_id = rxq->nb_rx_desc - 1;
71 	}
72 
73 	rxq->rxrearm_nb -= RTE_I40E_RXQ_REARM_THRESH;
74 
75 	rte_io_wmb();
76 	/* Update the tail pointer on the NIC */
77 	I40E_PCI_REG_WRITE_RELAXED(rxq->qrx_tail, rx_id);
78 }
79 
80 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
81 /* NEON version of FDIR mark extraction for 4 32B descriptors at a time */
82 static inline uint32x4_t
83 descs_to_fdir_32b(volatile union i40e_rx_desc *rxdp, struct rte_mbuf **rx_pkt)
84 {
85 	/* 32B descriptors: Load 2nd half of descriptors for FDIR ID data */
86 	uint64x2_t desc0_qw23, desc1_qw23, desc2_qw23, desc3_qw23;
87 	desc0_qw23 = vld1q_u64(RTE_CAST_PTR(uint64_t *, &(rxdp + 0)->wb.qword2));
88 	desc1_qw23 = vld1q_u64(RTE_CAST_PTR(uint64_t *, &(rxdp + 1)->wb.qword2));
89 	desc2_qw23 = vld1q_u64(RTE_CAST_PTR(uint64_t *, &(rxdp + 2)->wb.qword2));
90 	desc3_qw23 = vld1q_u64(RTE_CAST_PTR(uint64_t *, &(rxdp + 3)->wb.qword2));
91 
92 	/* FDIR ID data: move last u32 of each desc to 4 u32 lanes */
93 	uint32x4_t v_unpack_02, v_unpack_13;
94 	v_unpack_02 = vzipq_u32(vreinterpretq_u32_u64(desc0_qw23),
95 				vreinterpretq_u32_u64(desc2_qw23)).val[1];
96 	v_unpack_13 = vzipq_u32(vreinterpretq_u32_u64(desc1_qw23),
97 				vreinterpretq_u32_u64(desc3_qw23)).val[1];
98 	uint32x4_t v_fdir_ids = vzipq_u32(v_unpack_02, v_unpack_13).val[1];
99 
100 	/* Extended Status: extract from each lower 32 bits, to u32 lanes */
101 	v_unpack_02 = vzipq_u32(vreinterpretq_u32_u64(desc0_qw23),
102 				vreinterpretq_u32_u64(desc2_qw23)).val[0];
103 	v_unpack_13 = vzipq_u32(vreinterpretq_u32_u64(desc1_qw23),
104 				vreinterpretq_u32_u64(desc3_qw23)).val[0];
105 	uint32x4_t v_flt_status = vzipq_u32(v_unpack_02, v_unpack_13).val[0];
106 
107 	/* Shift u32 left and right to "mask away" bits not required.
108 	 * Data required is 4:5 (zero based), so left shift by 26 (32-6)
109 	 * and then right shift by 30 (32 - 2 bits required).
110 	 */
111 	v_flt_status = vshlq_n_u32(v_flt_status, 26);
112 	v_flt_status = vshrq_n_u32(v_flt_status, 30);
113 
114 	/* Generate constant 1 in all u32 lanes */
115 	RTE_BUILD_BUG_ON(I40E_RX_DESC_EXT_STATUS_FLEXBH_FD_ID != 1);
116 	uint32x4_t v_u32_one = vdupq_n_u32(1);
117 
118 	/* Per desc mask, bits set if FDIR ID is valid */
119 	uint32x4_t v_fd_id_mask = vceqq_u32(v_flt_status, v_u32_one);
120 
121 	/* Mask ID data to zero if the FD_ID bit not set in desc */
122 	v_fdir_ids = vandq_u32(v_fdir_ids, v_fd_id_mask);
123 
124 	/* Store data to fdir.hi in mbuf */
125 	rx_pkt[0]->hash.fdir.hi = vgetq_lane_u32(v_fdir_ids, 0);
126 	rx_pkt[1]->hash.fdir.hi = vgetq_lane_u32(v_fdir_ids, 1);
127 	rx_pkt[2]->hash.fdir.hi = vgetq_lane_u32(v_fdir_ids, 2);
128 	rx_pkt[3]->hash.fdir.hi = vgetq_lane_u32(v_fdir_ids, 3);
129 
130 	/* Convert fdir_id_mask into a single bit, then shift as required for
131 	 * correct location in the mbuf->olflags
132 	 */
133 	RTE_BUILD_BUG_ON(RTE_MBUF_F_RX_FDIR_ID != (1 << 13));
134 	v_fd_id_mask = vshrq_n_u32(v_fd_id_mask, 31);
135 	v_fd_id_mask = vshlq_n_u32(v_fd_id_mask, 13);
136 
137 	/* The returned value must be combined into each mbuf. This is already
138 	 * being done for RSS and VLAN mbuf olflags, so return bits to OR in.
139 	 */
140 	return v_fd_id_mask;
141 }
142 
143 #else /* 32 or 16B FDIR ID handling */
144 
145 /* Handle 16B descriptor FDIR ID flag setting based on FLM(bit11). See scalar driver
146  * for scalar implementation of the same functionality.
147  */
148 static inline uint32x4_t
149 descs_to_fdir_16b(uint32x4_t fltstat, uint64x2_t descs[4], struct rte_mbuf **rx_pkt)
150 {
151 	/* Unpack filter-status data from descriptors */
152 	uint32x4_t v_tmp_02 = vzipq_u32(vreinterpretq_u32_u64(descs[0]),
153 					vreinterpretq_u32_u64(descs[2])).val[0];
154 	uint32x4_t v_tmp_13 = vzipq_u32(vreinterpretq_u32_u64(descs[1]),
155 					vreinterpretq_u32_u64(descs[3])).val[0];
156 	uint32x4_t v_fdir_ids = vzipq_u32(v_tmp_02, v_tmp_13).val[1];
157 
158 	/* Generate 111 and 11 in each u32 lane */
159 	uint32x4_t v_111_mask = vdupq_n_u32(7);
160 	uint32x4_t v_11_mask = vdupq_n_u32(3);
161 
162 	/* Compare and mask away FDIR ID data if bit not set */
163 	uint32x4_t v_u32_bits = vandq_u32(v_111_mask, fltstat);
164 	uint32x4_t v_fdir_id_mask = vceqq_u32(v_u32_bits, v_11_mask);
165 	v_fdir_ids = vandq_u32(v_fdir_id_mask, v_fdir_ids);
166 
167 	/* Store data to fdir.hi in mbuf */
168 	rx_pkt[0]->hash.fdir.hi = vgetq_lane_u32(v_fdir_ids, 0);
169 	rx_pkt[1]->hash.fdir.hi = vgetq_lane_u32(v_fdir_ids, 1);
170 	rx_pkt[2]->hash.fdir.hi = vgetq_lane_u32(v_fdir_ids, 2);
171 	rx_pkt[3]->hash.fdir.hi = vgetq_lane_u32(v_fdir_ids, 3);
172 
173 	/* Top lane ones mask for FDIR isolation */
174 	uint32x4_t v_desc_fdir_mask = {0, UINT32_MAX, 0, 0};
175 
176 	/* Move fdir_id_mask to correct lane, zero RSS in mbuf if fdir hits */
177 	uint32x4_t v_zeros = {0, 0, 0, 0};
178 	uint32x4_t v_desc3_shift = vextq_u32(v_fdir_id_mask, v_zeros, 2);
179 	uint32x4_t v_desc3_mask = vandq_u32(v_desc_fdir_mask, v_desc3_shift);
180 	descs[3] = vreinterpretq_u64_u32(vbslq_u32(v_desc3_mask, v_zeros,
181 				vreinterpretq_u32_u64(descs[3])));
182 
183 	uint32x4_t v_desc2_shift = vextq_u32(v_fdir_id_mask, v_zeros, 1);
184 	uint32x4_t v_desc2_mask = vandq_u32(v_desc_fdir_mask, v_desc2_shift);
185 	descs[2] = vreinterpretq_u64_u32(vbslq_u32(v_desc2_mask, v_zeros,
186 				vreinterpretq_u32_u64(descs[2])));
187 
188 	uint32x4_t v_desc1_shift = v_fdir_id_mask;
189 	uint32x4_t v_desc1_mask = vandq_u32(v_desc_fdir_mask, v_desc1_shift);
190 	descs[1] = vreinterpretq_u64_u32(vbslq_u32(v_desc1_mask, v_zeros,
191 				vreinterpretq_u32_u64(descs[1])));
192 
193 	uint32x4_t v_desc0_shift = vextq_u32(v_zeros, v_fdir_id_mask, 3);
194 	uint32x4_t v_desc0_mask = vandq_u32(v_desc_fdir_mask, v_desc0_shift);
195 	descs[0] = vreinterpretq_u64_u32(vbslq_u32(v_desc0_mask, v_zeros,
196 				vreinterpretq_u32_u64(descs[0])));
197 
198 	/* Shift to 1 or 0 bit per u32 lane, then to RTE_MBUF_F_RX_FDIR_ID offset */
199 	RTE_BUILD_BUG_ON(RTE_MBUF_F_RX_FDIR_ID != (1 << 13));
200 	uint32x4_t v_mask_one_bit = vshrq_n_u32(v_fdir_id_mask, 31);
201 	return vshlq_n_u32(v_mask_one_bit, 13);
202 }
203 #endif
204 
205 static inline void
206 desc_to_olflags_v(struct i40e_rx_queue *rxq, volatile union i40e_rx_desc *rxdp,
207 		  uint64x2_t descs[4], struct rte_mbuf **rx_pkts)
208 {
209 	uint32x4_t vlan0, vlan1, rss, l3_l4e;
210 	const uint64x2_t mbuf_init = {rxq->mbuf_initializer, 0};
211 	uint64x2_t rearm0, rearm1, rearm2, rearm3;
212 
213 	/* mask everything except RSS, flow director and VLAN flags
214 	 * bit2 is for VLAN tag, bit11 for flow director indication
215 	 * bit13:12 for RSS indication.
216 	 */
217 	const uint32x4_t rss_vlan_msk = {
218 			0x1c03804, 0x1c03804, 0x1c03804, 0x1c03804};
219 
220 	const uint32x4_t cksum_mask = {
221 			RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_BAD |
222 			RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
223 			RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD,
224 			RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_BAD |
225 			RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
226 			RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD,
227 			RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_BAD |
228 			RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
229 			RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD,
230 			RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_BAD |
231 			RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
232 			RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD};
233 
234 	/* map rss and vlan type to rss hash and vlan flag */
235 	const uint8x16_t vlan_flags = {
236 			0, 0, 0, 0,
237 			RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED, 0, 0, 0,
238 			0, 0, 0, 0,
239 			0, 0, 0, 0};
240 
241 	const uint8x16_t rss_flags = {
242 			0, RTE_MBUF_F_RX_FDIR, 0, 0,
243 			0, 0, RTE_MBUF_F_RX_RSS_HASH, RTE_MBUF_F_RX_RSS_HASH | RTE_MBUF_F_RX_FDIR,
244 			0, 0, 0, 0,
245 			0, 0, 0, 0};
246 
247 	const uint8x16_t l3_l4e_flags = {
248 			(RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_GOOD) >> 1,
249 			RTE_MBUF_F_RX_IP_CKSUM_BAD >> 1,
250 			(RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_BAD) >> 1,
251 			(RTE_MBUF_F_RX_L4_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
252 			(RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD) >> 1,
253 			(RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
254 			(RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
255 			 RTE_MBUF_F_RX_L4_CKSUM_BAD) >> 1,
256 			(RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
257 			 RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
258 			0, 0, 0, 0, 0, 0, 0, 0};
259 
260 	vlan0 = vzipq_u32(vreinterpretq_u32_u64(descs[0]),
261 			  vreinterpretq_u32_u64(descs[2])).val[1];
262 	vlan1 = vzipq_u32(vreinterpretq_u32_u64(descs[1]),
263 			  vreinterpretq_u32_u64(descs[3])).val[1];
264 	vlan0 = vzipq_u32(vlan0, vlan1).val[0];
265 
266 	vlan1 = vandq_u32(vlan0, rss_vlan_msk);
267 	vlan0 = vreinterpretq_u32_u8(vqtbl1q_u8(vlan_flags,
268 						vreinterpretq_u8_u32(vlan1)));
269 
270 	const uint32x4_t desc_fltstat = vshrq_n_u32(vlan1, 11);
271 	rss = vreinterpretq_u32_u8(vqtbl1q_u8(rss_flags,
272 					      vreinterpretq_u8_u32(desc_fltstat)));
273 
274 	l3_l4e = vshrq_n_u32(vlan1, 22);
275 	l3_l4e = vreinterpretq_u32_u8(vqtbl1q_u8(l3_l4e_flags,
276 					      vreinterpretq_u8_u32(l3_l4e)));
277 	/* then we shift left 1 bit */
278 	l3_l4e = vshlq_n_u32(l3_l4e, 1);
279 	/* we need to mask out the redundant bits */
280 	l3_l4e = vandq_u32(l3_l4e, cksum_mask);
281 
282 	vlan0 = vorrq_u32(vlan0, rss);
283 	vlan0 = vorrq_u32(vlan0, l3_l4e);
284 
285 	/* Extract FDIR ID only if FDIR is enabled to avoid useless work */
286 	if (rxq->fdir_enabled) {
287 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
288 		uint32x4_t v_fdir_ol_flags = descs_to_fdir_32b(rxdp, rx_pkts);
289 #else
290 		(void)rxdp; /* rxdp not required for 16B desc mode */
291 		uint32x4_t v_fdir_ol_flags = descs_to_fdir_16b(desc_fltstat, descs, rx_pkts);
292 #endif
293 		/* OR in ol_flag bits after descriptor specific extraction */
294 		vlan0 = vorrq_u32(vlan0, v_fdir_ol_flags);
295 	}
296 
297 	rearm0 = vsetq_lane_u64(vgetq_lane_u32(vlan0, 0), mbuf_init, 1);
298 	rearm1 = vsetq_lane_u64(vgetq_lane_u32(vlan0, 1), mbuf_init, 1);
299 	rearm2 = vsetq_lane_u64(vgetq_lane_u32(vlan0, 2), mbuf_init, 1);
300 	rearm3 = vsetq_lane_u64(vgetq_lane_u32(vlan0, 3), mbuf_init, 1);
301 
302 	vst1q_u64((uint64_t *)&rx_pkts[0]->rearm_data, rearm0);
303 	vst1q_u64((uint64_t *)&rx_pkts[1]->rearm_data, rearm1);
304 	vst1q_u64((uint64_t *)&rx_pkts[2]->rearm_data, rearm2);
305 	vst1q_u64((uint64_t *)&rx_pkts[3]->rearm_data, rearm3);
306 }
307 
308 #define PKTLEN_SHIFT     10
309 #define I40E_UINT16_BIT (CHAR_BIT * sizeof(uint16_t))
310 
311 static inline void
312 desc_to_ptype_v(uint64x2_t descs[4], struct rte_mbuf **__rte_restrict rx_pkts,
313 		uint32_t *__rte_restrict ptype_tbl)
314 {
315 	int i;
316 	uint8_t ptype;
317 	uint8x16_t tmp;
318 
319 	for (i = 0; i < 4; i++) {
320 		tmp = vreinterpretq_u8_u64(vshrq_n_u64(descs[i], 30));
321 		ptype = vgetq_lane_u8(tmp, 8);
322 		rx_pkts[i]->packet_type = ptype_tbl[ptype];
323 	}
324 
325 }
326 
327 /**
328  * vPMD raw receive routine, only accept(nb_pkts >= RTE_I40E_DESCS_PER_LOOP)
329  *
330  * Notice:
331  * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
332  * - floor align nb_pkts to a RTE_I40E_DESCS_PER_LOOP power-of-two
333  */
334 static inline uint16_t
335 _recv_raw_pkts_vec(struct i40e_rx_queue *__rte_restrict rxq,
336 		   struct rte_mbuf **__rte_restrict rx_pkts,
337 		   uint16_t nb_pkts, uint8_t *split_packet)
338 {
339 	volatile union i40e_rx_desc *rxdp;
340 	struct i40e_rx_entry *sw_ring;
341 	uint16_t nb_pkts_recd;
342 	int pos;
343 	uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
344 
345 	/* mask to shuffle from desc. to mbuf */
346 	uint8x16_t shuf_msk = {
347 		0xFF, 0xFF,   /* pkt_type set as unknown */
348 		0xFF, 0xFF,   /* pkt_type set as unknown */
349 		14, 15,       /* octet 15~14, low 16 bits pkt_len */
350 		0xFF, 0xFF,   /* skip high 16 bits pkt_len, zero out */
351 		14, 15,       /* octet 15~14, 16 bits data_len */
352 		2, 3,         /* octet 2~3, low 16 bits vlan_macip */
353 		4, 5, 6, 7    /* octet 4~7, 32bits rss */
354 		};
355 
356 	uint8x16_t eop_check = {
357 		0x02, 0x00, 0x02, 0x00,
358 		0x02, 0x00, 0x02, 0x00,
359 		0x00, 0x00, 0x00, 0x00,
360 		0x00, 0x00, 0x00, 0x00
361 		};
362 
363 	uint16x8_t crc_adjust = {
364 		0, 0,         /* ignore pkt_type field */
365 		rxq->crc_len, /* sub crc on pkt_len */
366 		0,            /* ignore high-16bits of pkt_len */
367 		rxq->crc_len, /* sub crc on data_len */
368 		0, 0, 0       /* ignore non-length fields */
369 		};
370 
371 	/* nb_pkts has to be floor-aligned to RTE_I40E_DESCS_PER_LOOP */
372 	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_I40E_DESCS_PER_LOOP);
373 
374 	/* Just the act of getting into the function from the application is
375 	 * going to cost about 7 cycles
376 	 */
377 	rxdp = rxq->rx_ring + rxq->rx_tail;
378 
379 	rte_prefetch_non_temporal(rxdp);
380 
381 	/* See if we need to rearm the RX queue - gives the prefetch a bit
382 	 * of time to act
383 	 */
384 	if (rxq->rxrearm_nb > RTE_I40E_RXQ_REARM_THRESH)
385 		i40e_rxq_rearm(rxq);
386 
387 	/* Before we start moving massive data around, check to see if
388 	 * there is actually a packet available
389 	 */
390 	if (!(rxdp->wb.qword1.status_error_len &
391 			rte_cpu_to_le_32(1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
392 		return 0;
393 
394 	/* Cache is empty -> need to scan the buffer rings, but first move
395 	 * the next 'n' mbufs into the cache
396 	 */
397 	sw_ring = &rxq->sw_ring[rxq->rx_tail];
398 
399 	/* A. load 4 packet in one loop
400 	 * [A*. mask out 4 unused dirty field in desc]
401 	 * B. copy 4 mbuf point from swring to rx_pkts
402 	 * C. calc the number of DD bits among the 4 packets
403 	 * [C*. extract the end-of-packet bit, if requested]
404 	 * D. fill info. from desc to mbuf
405 	 */
406 
407 	for (pos = 0, nb_pkts_recd = 0; pos < nb_pkts;
408 			pos += RTE_I40E_DESCS_PER_LOOP,
409 			rxdp += RTE_I40E_DESCS_PER_LOOP) {
410 		uint64x2_t descs[RTE_I40E_DESCS_PER_LOOP];
411 		uint8x16_t pkt_mb1, pkt_mb2, pkt_mb3, pkt_mb4;
412 		uint16x8x2_t sterr_tmp1, sterr_tmp2;
413 		uint64x2_t mbp1, mbp2;
414 		uint16x8_t staterr;
415 		uint16x8_t tmp;
416 		uint64_t stat;
417 
418 		int32x4_t len_shl = {0, 0, 0, PKTLEN_SHIFT};
419 
420 		/* A.1 load desc[3-0] */
421 		descs[3] =  vld1q_u64(RTE_CAST_PTR(uint64_t *, rxdp + 3));
422 		descs[2] =  vld1q_u64(RTE_CAST_PTR(uint64_t *, rxdp + 2));
423 		descs[1] =  vld1q_u64(RTE_CAST_PTR(uint64_t *, rxdp + 1));
424 		descs[0] =  vld1q_u64(RTE_CAST_PTR(uint64_t *, rxdp));
425 
426 		/* Use acquire fence to order loads of descriptor qwords */
427 		rte_atomic_thread_fence(rte_memory_order_acquire);
428 		/* A.2 reload qword0 to make it ordered after qword1 load */
429 		descs[3] = vld1q_lane_u64(RTE_CAST_PTR(uint64_t *, rxdp + 3), descs[3], 0);
430 		descs[2] = vld1q_lane_u64(RTE_CAST_PTR(uint64_t *, rxdp + 2), descs[2], 0);
431 		descs[1] = vld1q_lane_u64(RTE_CAST_PTR(uint64_t *, rxdp + 1), descs[1], 0);
432 		descs[0] = vld1q_lane_u64(RTE_CAST_PTR(uint64_t *, rxdp), descs[0], 0);
433 
434 		/* B.1 load 4 mbuf point */
435 		mbp1 = vld1q_u64((uint64_t *)&sw_ring[pos]);
436 		mbp2 = vld1q_u64((uint64_t *)&sw_ring[pos + 2]);
437 
438 		/* B.2 copy 4 mbuf point into rx_pkts  */
439 		vst1q_u64((uint64_t *)&rx_pkts[pos], mbp1);
440 		vst1q_u64((uint64_t *)&rx_pkts[pos + 2], mbp2);
441 
442 		if (split_packet) {
443 			rte_mbuf_prefetch_part2(rx_pkts[pos]);
444 			rte_mbuf_prefetch_part2(rx_pkts[pos + 1]);
445 			rte_mbuf_prefetch_part2(rx_pkts[pos + 2]);
446 			rte_mbuf_prefetch_part2(rx_pkts[pos + 3]);
447 		}
448 
449 		/* pkts shift the pktlen field to be 16-bit aligned*/
450 		uint32x4_t len3 = vshlq_u32(vreinterpretq_u32_u64(descs[3]),
451 					    len_shl);
452 		descs[3] = vreinterpretq_u64_u16(vsetq_lane_u16
453 				(vgetq_lane_u16(vreinterpretq_u16_u32(len3), 7),
454 				 vreinterpretq_u16_u64(descs[3]),
455 				 7));
456 		uint32x4_t len2 = vshlq_u32(vreinterpretq_u32_u64(descs[2]),
457 					    len_shl);
458 		descs[2] = vreinterpretq_u64_u16(vsetq_lane_u16
459 				(vgetq_lane_u16(vreinterpretq_u16_u32(len2), 7),
460 				 vreinterpretq_u16_u64(descs[2]),
461 				 7));
462 		uint32x4_t len1 = vshlq_u32(vreinterpretq_u32_u64(descs[1]),
463 					    len_shl);
464 		descs[1] = vreinterpretq_u64_u16(vsetq_lane_u16
465 				(vgetq_lane_u16(vreinterpretq_u16_u32(len1), 7),
466 				 vreinterpretq_u16_u64(descs[1]),
467 				 7));
468 		uint32x4_t len0 = vshlq_u32(vreinterpretq_u32_u64(descs[0]),
469 					    len_shl);
470 		descs[0] = vreinterpretq_u64_u16(vsetq_lane_u16
471 				(vgetq_lane_u16(vreinterpretq_u16_u32(len0), 7),
472 				 vreinterpretq_u16_u64(descs[0]),
473 				 7));
474 
475 		desc_to_olflags_v(rxq, rxdp, descs, &rx_pkts[pos]);
476 
477 		/* D.1 pkts convert format from desc to pktmbuf */
478 		pkt_mb4 = vqtbl1q_u8(vreinterpretq_u8_u64(descs[3]), shuf_msk);
479 		pkt_mb3 = vqtbl1q_u8(vreinterpretq_u8_u64(descs[2]), shuf_msk);
480 		pkt_mb2 = vqtbl1q_u8(vreinterpretq_u8_u64(descs[1]), shuf_msk);
481 		pkt_mb1 = vqtbl1q_u8(vreinterpretq_u8_u64(descs[0]), shuf_msk);
482 
483 		/* D.2 pkts set in_port/nb_seg and remove crc */
484 		tmp = vsubq_u16(vreinterpretq_u16_u8(pkt_mb4), crc_adjust);
485 		pkt_mb4 = vreinterpretq_u8_u16(tmp);
486 		tmp = vsubq_u16(vreinterpretq_u16_u8(pkt_mb3), crc_adjust);
487 		pkt_mb3 = vreinterpretq_u8_u16(tmp);
488 		tmp = vsubq_u16(vreinterpretq_u16_u8(pkt_mb2), crc_adjust);
489 		pkt_mb2 = vreinterpretq_u8_u16(tmp);
490 		tmp = vsubq_u16(vreinterpretq_u16_u8(pkt_mb1), crc_adjust);
491 		pkt_mb1 = vreinterpretq_u8_u16(tmp);
492 
493 		/* D.3 copy final data to rx_pkts */
494 		vst1q_u8((void *)&rx_pkts[pos + 3]->rx_descriptor_fields1,
495 				pkt_mb4);
496 		vst1q_u8((void *)&rx_pkts[pos + 2]->rx_descriptor_fields1,
497 				pkt_mb3);
498 		vst1q_u8((void *)&rx_pkts[pos + 1]->rx_descriptor_fields1,
499 				pkt_mb2);
500 		vst1q_u8((void *)&rx_pkts[pos]->rx_descriptor_fields1,
501 				pkt_mb1);
502 
503 		desc_to_ptype_v(descs, &rx_pkts[pos], ptype_tbl);
504 
505 		if (likely(pos + RTE_I40E_DESCS_PER_LOOP < nb_pkts)) {
506 			rte_prefetch_non_temporal(rxdp + RTE_I40E_DESCS_PER_LOOP);
507 		}
508 
509 		/* C.1 4=>2 filter staterr info only */
510 		sterr_tmp2 = vzipq_u16(vreinterpretq_u16_u64(descs[1]),
511 				       vreinterpretq_u16_u64(descs[3]));
512 		sterr_tmp1 = vzipq_u16(vreinterpretq_u16_u64(descs[0]),
513 				       vreinterpretq_u16_u64(descs[2]));
514 
515 		/* C.2 get 4 pkts staterr value  */
516 		staterr = vzipq_u16(sterr_tmp1.val[1],
517 				    sterr_tmp2.val[1]).val[0];
518 
519 		/* C* extract and record EOP bit */
520 		if (split_packet) {
521 			uint8x16_t eop_shuf_mask = {
522 					0x00, 0x02, 0x04, 0x06,
523 					0xFF, 0xFF, 0xFF, 0xFF,
524 					0xFF, 0xFF, 0xFF, 0xFF,
525 					0xFF, 0xFF, 0xFF, 0xFF};
526 			uint8x16_t eop_bits;
527 
528 			/* and with mask to extract bits, flipping 1-0 */
529 			eop_bits = vmvnq_u8(vreinterpretq_u8_u16(staterr));
530 			eop_bits = vandq_u8(eop_bits, eop_check);
531 			/* the staterr values are not in order, as the count
532 			 * of dd bits doesn't care. However, for end of
533 			 * packet tracking, we do care, so shuffle. This also
534 			 * compresses the 32-bit values to 8-bit
535 			 */
536 			eop_bits = vqtbl1q_u8(eop_bits, eop_shuf_mask);
537 
538 			/* store the resulting 32-bit value */
539 			vst1q_lane_u32((uint32_t *)split_packet,
540 				       vreinterpretq_u32_u8(eop_bits), 0);
541 			split_packet += RTE_I40E_DESCS_PER_LOOP;
542 
543 			/* zero-out next pointers */
544 			rx_pkts[pos]->next = NULL;
545 			rx_pkts[pos + 1]->next = NULL;
546 			rx_pkts[pos + 2]->next = NULL;
547 			rx_pkts[pos + 3]->next = NULL;
548 		}
549 
550 		staterr = vshlq_n_u16(staterr, I40E_UINT16_BIT - 1);
551 		staterr = vreinterpretq_u16_s16(
552 				vshrq_n_s16(vreinterpretq_s16_u16(staterr),
553 					    I40E_UINT16_BIT - 1));
554 		stat = ~vgetq_lane_u64(vreinterpretq_u64_u16(staterr), 0);
555 
556 		/* C.4 calc available number of desc */
557 		if (unlikely(stat == 0)) {
558 			nb_pkts_recd += RTE_I40E_DESCS_PER_LOOP;
559 		} else {
560 			nb_pkts_recd += rte_ctz64(stat) / I40E_UINT16_BIT;
561 			break;
562 		}
563 	}
564 
565 	/* Update our internal tail pointer */
566 	rxq->rx_tail = (uint16_t)(rxq->rx_tail + nb_pkts_recd);
567 	rxq->rx_tail = (uint16_t)(rxq->rx_tail & (rxq->nb_rx_desc - 1));
568 	rxq->rxrearm_nb = (uint16_t)(rxq->rxrearm_nb + nb_pkts_recd);
569 
570 	return nb_pkts_recd;
571 }
572 
573  /*
574  * Notice:
575  * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
576  */
577 uint16_t
578 i40e_recv_pkts_vec(void *__rte_restrict rx_queue,
579 		struct rte_mbuf **__rte_restrict rx_pkts, uint16_t nb_pkts)
580 {
581 	return _recv_raw_pkts_vec(rx_queue, rx_pkts, nb_pkts, NULL);
582 }
583 
584 /**
585  * vPMD receive routine that reassembles single burst of 32 scattered packets
586  *
587  * Notice:
588  * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
589  */
590 static uint16_t
591 i40e_recv_scattered_burst_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
592 			      uint16_t nb_pkts)
593 {
594 
595 	struct i40e_rx_queue *rxq = rx_queue;
596 	uint8_t split_flags[RTE_I40E_VPMD_RX_BURST] = {0};
597 
598 	/* get some new buffers */
599 	uint16_t nb_bufs = _recv_raw_pkts_vec(rxq, rx_pkts, nb_pkts,
600 			split_flags);
601 	if (nb_bufs == 0)
602 		return 0;
603 
604 	/* happy day case, full burst + no packets to be joined */
605 	const uint64_t *split_fl64 = (uint64_t *)split_flags;
606 
607 	if (rxq->pkt_first_seg == NULL &&
608 			split_fl64[0] == 0 && split_fl64[1] == 0 &&
609 			split_fl64[2] == 0 && split_fl64[3] == 0)
610 		return nb_bufs;
611 
612 	/* reassemble any packets that need reassembly*/
613 	unsigned i = 0;
614 
615 	if (rxq->pkt_first_seg == NULL) {
616 		/* find the first split flag, and only reassemble then*/
617 		while (i < nb_bufs && !split_flags[i])
618 			i++;
619 		if (i == nb_bufs)
620 			return nb_bufs;
621 		rxq->pkt_first_seg = rx_pkts[i];
622 	}
623 	return i + ci_rx_reassemble_packets(&rx_pkts[i], nb_bufs - i, &split_flags[i],
624 		&rxq->pkt_first_seg, &rxq->pkt_last_seg, rxq->crc_len);
625 }
626 
627 /**
628  * vPMD receive routine that reassembles scattered packets.
629  */
630 uint16_t
631 i40e_recv_scattered_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
632 			     uint16_t nb_pkts)
633 {
634 	uint16_t retval = 0;
635 
636 	while (nb_pkts > RTE_I40E_VPMD_RX_BURST) {
637 		uint16_t burst;
638 
639 		burst = i40e_recv_scattered_burst_vec(rx_queue,
640 						      rx_pkts + retval,
641 						      RTE_I40E_VPMD_RX_BURST);
642 		retval += burst;
643 		nb_pkts -= burst;
644 		if (burst < RTE_I40E_VPMD_RX_BURST)
645 			return retval;
646 	}
647 
648 	return retval + i40e_recv_scattered_burst_vec(rx_queue,
649 						      rx_pkts + retval,
650 						      nb_pkts);
651 }
652 
653 static inline void
654 vtx1(volatile struct i40e_tx_desc *txdp,
655 		struct rte_mbuf *pkt, uint64_t flags)
656 {
657 	uint64_t high_qw = (I40E_TX_DESC_DTYPE_DATA |
658 			((uint64_t)flags  << I40E_TXD_QW1_CMD_SHIFT) |
659 			((uint64_t)pkt->data_len << I40E_TXD_QW1_TX_BUF_SZ_SHIFT));
660 
661 	uint64x2_t descriptor = {pkt->buf_iova + pkt->data_off, high_qw};
662 	vst1q_u64(RTE_CAST_PTR(uint64_t *, txdp), descriptor);
663 }
664 
665 static inline void
666 vtx(volatile struct i40e_tx_desc *txdp, struct rte_mbuf **pkt,
667 		uint16_t nb_pkts,  uint64_t flags)
668 {
669 	int i;
670 
671 	for (i = 0; i < nb_pkts; ++i, ++txdp, ++pkt)
672 		vtx1(txdp, *pkt, flags);
673 }
674 
675 uint16_t
676 i40e_xmit_fixed_burst_vec(void *__rte_restrict tx_queue,
677 	struct rte_mbuf **__rte_restrict tx_pkts, uint16_t nb_pkts)
678 {
679 	struct i40e_tx_queue *txq = (struct i40e_tx_queue *)tx_queue;
680 	volatile struct i40e_tx_desc *txdp;
681 	struct ci_tx_entry *txep;
682 	uint16_t n, nb_commit, tx_id;
683 	uint64_t flags = I40E_TD_CMD;
684 	uint64_t rs = I40E_TX_DESC_CMD_RS | I40E_TD_CMD;
685 	int i;
686 
687 	if (txq->nb_tx_free < txq->tx_free_thresh)
688 		i40e_tx_free_bufs(txq);
689 
690 	nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
691 	if (unlikely(nb_pkts == 0))
692 		return 0;
693 
694 	tx_id = txq->tx_tail;
695 	txdp = &txq->tx_ring[tx_id];
696 	txep = &txq->sw_ring[tx_id];
697 
698 	txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
699 
700 	n = (uint16_t)(txq->nb_tx_desc - tx_id);
701 	if (nb_commit >= n) {
702 		tx_backlog_entry(txep, tx_pkts, n);
703 
704 		for (i = 0; i < n - 1; ++i, ++tx_pkts, ++txdp)
705 			vtx1(txdp, *tx_pkts, flags);
706 
707 		vtx1(txdp, *tx_pkts++, rs);
708 
709 		nb_commit = (uint16_t)(nb_commit - n);
710 
711 		tx_id = 0;
712 		txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
713 
714 		/* avoid reach the end of ring */
715 		txdp = &txq->tx_ring[tx_id];
716 		txep = &txq->sw_ring[tx_id];
717 	}
718 
719 	tx_backlog_entry(txep, tx_pkts, nb_commit);
720 
721 	vtx(txdp, tx_pkts, nb_commit, flags);
722 
723 	tx_id = (uint16_t)(tx_id + nb_commit);
724 	if (tx_id > txq->tx_next_rs) {
725 		txq->tx_ring[txq->tx_next_rs].cmd_type_offset_bsz |=
726 			rte_cpu_to_le_64(((uint64_t)I40E_TX_DESC_CMD_RS) <<
727 						I40E_TXD_QW1_CMD_SHIFT);
728 		txq->tx_next_rs =
729 			(uint16_t)(txq->tx_next_rs + txq->tx_rs_thresh);
730 	}
731 
732 	txq->tx_tail = tx_id;
733 
734 	rte_io_wmb();
735 	I40E_PCI_REG_WRITE_RELAXED(txq->qtx_tail, tx_id);
736 
737 	return nb_pkts;
738 }
739 
740 void __rte_cold
741 i40e_rx_queue_release_mbufs_vec(struct i40e_rx_queue *rxq)
742 {
743 	_i40e_rx_queue_release_mbufs_vec(rxq);
744 }
745 
746 int __rte_cold
747 i40e_rxq_vec_setup(struct i40e_rx_queue *rxq)
748 {
749 	return i40e_rxq_vec_setup_default(rxq);
750 }
751 
752 int __rte_cold
753 i40e_txq_vec_setup(struct i40e_tx_queue __rte_unused *txq)
754 {
755 	return 0;
756 }
757 
758 int __rte_cold
759 i40e_rx_vec_dev_conf_condition_check(struct rte_eth_dev *dev)
760 {
761 	return i40e_rx_vec_dev_conf_condition_check_default(dev);
762 }
763