1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2019 Intel Corporation
3 */
4
5 #include "ice_rxtx_vec_common.h"
6
7 #include <tmmintrin.h>
8
9 #ifndef __INTEL_COMPILER
10 #pragma GCC diagnostic ignored "-Wcast-qual"
11 #endif
12
13 static inline __m128i
ice_flex_rxd_to_fdir_flags_vec(const __m128i fdir_id0_3)14 ice_flex_rxd_to_fdir_flags_vec(const __m128i fdir_id0_3)
15 {
16 #define FDID_MIS_MAGIC 0xFFFFFFFF
17 RTE_BUILD_BUG_ON(RTE_MBUF_F_RX_FDIR != (1 << 2));
18 RTE_BUILD_BUG_ON(RTE_MBUF_F_RX_FDIR_ID != (1 << 13));
19 const __m128i pkt_fdir_bit = _mm_set1_epi32(RTE_MBUF_F_RX_FDIR |
20 RTE_MBUF_F_RX_FDIR_ID);
21 /* desc->flow_id field == 0xFFFFFFFF means fdir mismatch */
22 const __m128i fdir_mis_mask = _mm_set1_epi32(FDID_MIS_MAGIC);
23 __m128i fdir_mask = _mm_cmpeq_epi32(fdir_id0_3,
24 fdir_mis_mask);
25 /* this XOR op results to bit-reverse the fdir_mask */
26 fdir_mask = _mm_xor_si128(fdir_mask, fdir_mis_mask);
27 const __m128i fdir_flags = _mm_and_si128(fdir_mask, pkt_fdir_bit);
28
29 return fdir_flags;
30 }
31
32 static inline void
ice_rxq_rearm(struct ice_rx_queue * rxq)33 ice_rxq_rearm(struct ice_rx_queue *rxq)
34 {
35 int i;
36 uint16_t rx_id;
37 volatile union ice_rx_flex_desc *rxdp;
38 struct ice_rx_entry *rxep = &rxq->sw_ring[rxq->rxrearm_start];
39 struct rte_mbuf *mb0, *mb1;
40 __m128i hdr_room = _mm_set_epi64x(RTE_PKTMBUF_HEADROOM,
41 RTE_PKTMBUF_HEADROOM);
42 __m128i dma_addr0, dma_addr1;
43
44 rxdp = rxq->rx_ring + rxq->rxrearm_start;
45
46 /* Pull 'n' more MBUFs into the software ring */
47 if (rte_mempool_get_bulk(rxq->mp,
48 (void *)rxep,
49 ICE_RXQ_REARM_THRESH) < 0) {
50 if (rxq->rxrearm_nb + ICE_RXQ_REARM_THRESH >=
51 rxq->nb_rx_desc) {
52 dma_addr0 = _mm_setzero_si128();
53 for (i = 0; i < ICE_DESCS_PER_LOOP; i++) {
54 rxep[i].mbuf = &rxq->fake_mbuf;
55 _mm_store_si128((__m128i *)&rxdp[i].read,
56 dma_addr0);
57 }
58 }
59 rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
60 ICE_RXQ_REARM_THRESH;
61 return;
62 }
63
64 /* Initialize the mbufs in vector, process 2 mbufs in one loop */
65 for (i = 0; i < ICE_RXQ_REARM_THRESH; i += 2, rxep += 2) {
66 __m128i vaddr0, vaddr1;
67
68 mb0 = rxep[0].mbuf;
69 mb1 = rxep[1].mbuf;
70
71 /* load buf_addr(lo 64bit) and buf_iova(hi 64bit) */
72 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, buf_iova) !=
73 offsetof(struct rte_mbuf, buf_addr) + 8);
74 vaddr0 = _mm_loadu_si128((__m128i *)&mb0->buf_addr);
75 vaddr1 = _mm_loadu_si128((__m128i *)&mb1->buf_addr);
76
77 /* convert pa to dma_addr hdr/data */
78 dma_addr0 = _mm_unpackhi_epi64(vaddr0, vaddr0);
79 dma_addr1 = _mm_unpackhi_epi64(vaddr1, vaddr1);
80
81 /* add headroom to pa values */
82 dma_addr0 = _mm_add_epi64(dma_addr0, hdr_room);
83 dma_addr1 = _mm_add_epi64(dma_addr1, hdr_room);
84
85 /* flush desc with pa dma_addr */
86 _mm_store_si128((__m128i *)&rxdp++->read, dma_addr0);
87 _mm_store_si128((__m128i *)&rxdp++->read, dma_addr1);
88 }
89
90 rxq->rxrearm_start += ICE_RXQ_REARM_THRESH;
91 if (rxq->rxrearm_start >= rxq->nb_rx_desc)
92 rxq->rxrearm_start = 0;
93
94 rxq->rxrearm_nb -= ICE_RXQ_REARM_THRESH;
95
96 rx_id = (uint16_t)((rxq->rxrearm_start == 0) ?
97 (rxq->nb_rx_desc - 1) : (rxq->rxrearm_start - 1));
98
99 /* Update the tail pointer on the NIC */
100 ICE_PCI_REG_WC_WRITE(rxq->qrx_tail, rx_id);
101 }
102
103 static inline void
ice_rx_desc_to_olflags_v(struct ice_rx_queue * rxq,__m128i descs[4],struct rte_mbuf ** rx_pkts)104 ice_rx_desc_to_olflags_v(struct ice_rx_queue *rxq, __m128i descs[4],
105 struct rte_mbuf **rx_pkts)
106 {
107 const __m128i mbuf_init = _mm_set_epi64x(0, rxq->mbuf_initializer);
108 __m128i rearm0, rearm1, rearm2, rearm3;
109
110 __m128i tmp_desc, flags, rss_vlan;
111
112 /* mask everything except checksum, RSS and VLAN flags.
113 * bit6:4 for checksum.
114 * bit12 for RSS indication.
115 * bit13 for VLAN indication.
116 */
117 const __m128i desc_mask = _mm_set_epi32(0x30f0, 0x30f0,
118 0x30f0, 0x30f0);
119 const __m128i cksum_mask = _mm_set_epi32(RTE_MBUF_F_RX_IP_CKSUM_MASK |
120 RTE_MBUF_F_RX_L4_CKSUM_MASK |
121 RTE_MBUF_F_RX_OUTER_L4_CKSUM_MASK |
122 RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD,
123 RTE_MBUF_F_RX_IP_CKSUM_MASK |
124 RTE_MBUF_F_RX_L4_CKSUM_MASK |
125 RTE_MBUF_F_RX_OUTER_L4_CKSUM_MASK |
126 RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD,
127 RTE_MBUF_F_RX_IP_CKSUM_MASK |
128 RTE_MBUF_F_RX_L4_CKSUM_MASK |
129 RTE_MBUF_F_RX_OUTER_L4_CKSUM_MASK |
130 RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD,
131 RTE_MBUF_F_RX_IP_CKSUM_MASK |
132 RTE_MBUF_F_RX_L4_CKSUM_MASK |
133 RTE_MBUF_F_RX_OUTER_L4_CKSUM_MASK |
134 RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD);
135
136 /* map the checksum, rss and vlan fields to the checksum, rss
137 * and vlan flag
138 */
139 const __m128i cksum_flags =
140 _mm_set_epi8((RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 |
141 RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
142 RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
143 (RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
144 RTE_MBUF_F_RX_L4_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
145 (RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
146 RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
147 (RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
148 RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
149 (RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_BAD |
150 RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
151 (RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_BAD |
152 RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
153 (RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_GOOD |
154 RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
155 (RTE_MBUF_F_RX_OUTER_L4_CKSUM_BAD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_GOOD |
156 RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
157 /**
158 * shift right 20 bits to use the low two bits to indicate
159 * outer checksum status
160 * shift right 1 bit to make sure it not exceed 255
161 */
162 (RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
163 RTE_MBUF_F_RX_L4_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
164 (RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
165 RTE_MBUF_F_RX_L4_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
166 (RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
167 RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
168 (RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
169 RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
170 (RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_BAD |
171 RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
172 (RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_BAD |
173 RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
174 (RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_GOOD |
175 RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
176 (RTE_MBUF_F_RX_OUTER_L4_CKSUM_GOOD >> 20 | RTE_MBUF_F_RX_L4_CKSUM_GOOD |
177 RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1);
178
179 const __m128i rss_vlan_flags = _mm_set_epi8(0, 0, 0, 0,
180 0, 0, 0, 0,
181 0, 0, 0, 0,
182 RTE_MBUF_F_RX_RSS_HASH | RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED,
183 RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED,
184 RTE_MBUF_F_RX_RSS_HASH, 0);
185
186 /* merge 4 descriptors */
187 flags = _mm_unpackhi_epi32(descs[0], descs[1]);
188 tmp_desc = _mm_unpackhi_epi32(descs[2], descs[3]);
189 tmp_desc = _mm_unpacklo_epi64(flags, tmp_desc);
190 tmp_desc = _mm_and_si128(tmp_desc, desc_mask);
191
192 /* checksum flags */
193 tmp_desc = _mm_srli_epi32(tmp_desc, 4);
194 flags = _mm_shuffle_epi8(cksum_flags, tmp_desc);
195 /* then we shift left 1 bit */
196 flags = _mm_slli_epi32(flags, 1);
197
198 __m128i l4_outer_mask = _mm_set_epi32(0x6, 0x6, 0x6, 0x6);
199 __m128i l4_outer_flags = _mm_and_si128(flags, l4_outer_mask);
200 l4_outer_flags = _mm_slli_epi32(l4_outer_flags, 20);
201
202 __m128i l3_l4_mask = _mm_set_epi32(~0x6, ~0x6, ~0x6, ~0x6);
203 __m128i l3_l4_flags = _mm_and_si128(flags, l3_l4_mask);
204 flags = _mm_or_si128(l3_l4_flags, l4_outer_flags);
205 /* we need to mask out the redundant bits introduced by RSS or
206 * VLAN fields.
207 */
208 flags = _mm_and_si128(flags, cksum_mask);
209
210 /* RSS, VLAN flag */
211 tmp_desc = _mm_srli_epi32(tmp_desc, 8);
212 rss_vlan = _mm_shuffle_epi8(rss_vlan_flags, tmp_desc);
213
214 /* merge the flags */
215 flags = _mm_or_si128(flags, rss_vlan);
216
217 if (rxq->fdir_enabled) {
218 const __m128i fdir_id0_1 =
219 _mm_unpackhi_epi32(descs[0], descs[1]);
220
221 const __m128i fdir_id2_3 =
222 _mm_unpackhi_epi32(descs[2], descs[3]);
223
224 const __m128i fdir_id0_3 =
225 _mm_unpackhi_epi64(fdir_id0_1, fdir_id2_3);
226
227 const __m128i fdir_flags =
228 ice_flex_rxd_to_fdir_flags_vec(fdir_id0_3);
229
230 /* merge with fdir_flags */
231 flags = _mm_or_si128(flags, fdir_flags);
232
233 /* write fdir_id to mbuf */
234 rx_pkts[0]->hash.fdir.hi =
235 _mm_extract_epi32(fdir_id0_3, 0);
236
237 rx_pkts[1]->hash.fdir.hi =
238 _mm_extract_epi32(fdir_id0_3, 1);
239
240 rx_pkts[2]->hash.fdir.hi =
241 _mm_extract_epi32(fdir_id0_3, 2);
242
243 rx_pkts[3]->hash.fdir.hi =
244 _mm_extract_epi32(fdir_id0_3, 3);
245 } /* if() on fdir_enabled */
246
247 /**
248 * At this point, we have the 4 sets of flags in the low 16-bits
249 * of each 32-bit value in flags.
250 * We want to extract these, and merge them with the mbuf init data
251 * so we can do a single 16-byte write to the mbuf to set the flags
252 * and all the other initialization fields. Extracting the
253 * appropriate flags means that we have to do a shift and blend for
254 * each mbuf before we do the write.
255 */
256 rearm0 = _mm_blend_epi16(mbuf_init, _mm_slli_si128(flags, 8), 0x30);
257 rearm1 = _mm_blend_epi16(mbuf_init, _mm_slli_si128(flags, 4), 0x30);
258 rearm2 = _mm_blend_epi16(mbuf_init, flags, 0x30);
259 rearm3 = _mm_blend_epi16(mbuf_init, _mm_srli_si128(flags, 4), 0x30);
260
261 /* write the rearm data and the olflags in one write */
262 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, ol_flags) !=
263 offsetof(struct rte_mbuf, rearm_data) + 8);
264 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, rearm_data) !=
265 RTE_ALIGN(offsetof(struct rte_mbuf, rearm_data), 16));
266 _mm_store_si128((__m128i *)&rx_pkts[0]->rearm_data, rearm0);
267 _mm_store_si128((__m128i *)&rx_pkts[1]->rearm_data, rearm1);
268 _mm_store_si128((__m128i *)&rx_pkts[2]->rearm_data, rearm2);
269 _mm_store_si128((__m128i *)&rx_pkts[3]->rearm_data, rearm3);
270 }
271
272 static inline void
ice_rx_desc_to_ptype_v(__m128i descs[4],struct rte_mbuf ** rx_pkts,uint32_t * ptype_tbl)273 ice_rx_desc_to_ptype_v(__m128i descs[4], struct rte_mbuf **rx_pkts,
274 uint32_t *ptype_tbl)
275 {
276 const __m128i ptype_mask = _mm_set_epi16(ICE_RX_FLEX_DESC_PTYPE_M, 0,
277 ICE_RX_FLEX_DESC_PTYPE_M, 0,
278 ICE_RX_FLEX_DESC_PTYPE_M, 0,
279 ICE_RX_FLEX_DESC_PTYPE_M, 0);
280 __m128i ptype_01 = _mm_unpacklo_epi32(descs[0], descs[1]);
281 __m128i ptype_23 = _mm_unpacklo_epi32(descs[2], descs[3]);
282 __m128i ptype_all = _mm_unpacklo_epi64(ptype_01, ptype_23);
283
284 ptype_all = _mm_and_si128(ptype_all, ptype_mask);
285
286 rx_pkts[0]->packet_type = ptype_tbl[_mm_extract_epi16(ptype_all, 1)];
287 rx_pkts[1]->packet_type = ptype_tbl[_mm_extract_epi16(ptype_all, 3)];
288 rx_pkts[2]->packet_type = ptype_tbl[_mm_extract_epi16(ptype_all, 5)];
289 rx_pkts[3]->packet_type = ptype_tbl[_mm_extract_epi16(ptype_all, 7)];
290 }
291
292 /**
293 * vPMD raw receive routine, only accept(nb_pkts >= ICE_DESCS_PER_LOOP)
294 *
295 * Notice:
296 * - nb_pkts < ICE_DESCS_PER_LOOP, just return no packet
297 * - floor align nb_pkts to a ICE_DESCS_PER_LOOP power-of-two
298 */
299 static inline uint16_t
_ice_recv_raw_pkts_vec(struct ice_rx_queue * rxq,struct rte_mbuf ** rx_pkts,uint16_t nb_pkts,uint8_t * split_packet)300 _ice_recv_raw_pkts_vec(struct ice_rx_queue *rxq, struct rte_mbuf **rx_pkts,
301 uint16_t nb_pkts, uint8_t *split_packet)
302 {
303 volatile union ice_rx_flex_desc *rxdp;
304 struct ice_rx_entry *sw_ring;
305 uint16_t nb_pkts_recd;
306 int pos;
307 uint64_t var;
308 uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
309 __m128i crc_adjust = _mm_set_epi16
310 (0, 0, 0, /* ignore non-length fields */
311 -rxq->crc_len, /* sub crc on data_len */
312 0, /* ignore high-16bits of pkt_len */
313 -rxq->crc_len, /* sub crc on pkt_len */
314 0, 0 /* ignore pkt_type field */
315 );
316 const __m128i zero = _mm_setzero_si128();
317 /* mask to shuffle from desc. to mbuf */
318 const __m128i shuf_msk = _mm_set_epi8
319 (0xFF, 0xFF,
320 0xFF, 0xFF, /* rss hash parsed separately */
321 11, 10, /* octet 10~11, 16 bits vlan_macip */
322 5, 4, /* octet 4~5, 16 bits data_len */
323 0xFF, 0xFF, /* skip high 16 bits pkt_len, zero out */
324 5, 4, /* octet 4~5, low 16 bits pkt_len */
325 0xFF, 0xFF, /* pkt_type set as unknown */
326 0xFF, 0xFF /* pkt_type set as unknown */
327 );
328 const __m128i eop_shuf_mask = _mm_set_epi8(0xFF, 0xFF,
329 0xFF, 0xFF,
330 0xFF, 0xFF,
331 0xFF, 0xFF,
332 0xFF, 0xFF,
333 0xFF, 0xFF,
334 0x04, 0x0C,
335 0x00, 0x08);
336
337 /**
338 * compile-time check the above crc_adjust layout is correct.
339 * NOTE: the first field (lowest address) is given last in set_epi16
340 * call above.
341 */
342 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
343 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
344 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
345 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
346
347 /* 4 packets DD mask */
348 const __m128i dd_check = _mm_set_epi64x(0x0000000100000001LL,
349 0x0000000100000001LL);
350 /* 4 packets EOP mask */
351 const __m128i eop_check = _mm_set_epi64x(0x0000000200000002LL,
352 0x0000000200000002LL);
353
354 /* nb_pkts has to be floor-aligned to ICE_DESCS_PER_LOOP */
355 nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, ICE_DESCS_PER_LOOP);
356
357 /* Just the act of getting into the function from the application is
358 * going to cost about 7 cycles
359 */
360 rxdp = rxq->rx_ring + rxq->rx_tail;
361
362 rte_prefetch0(rxdp);
363
364 /* See if we need to rearm the RX queue - gives the prefetch a bit
365 * of time to act
366 */
367 if (rxq->rxrearm_nb > ICE_RXQ_REARM_THRESH)
368 ice_rxq_rearm(rxq);
369
370 /* Before we start moving massive data around, check to see if
371 * there is actually a packet available
372 */
373 if (!(rxdp->wb.status_error0 &
374 rte_cpu_to_le_32(1 << ICE_RX_FLEX_DESC_STATUS0_DD_S)))
375 return 0;
376
377 /**
378 * Compile-time verify the shuffle mask
379 * NOTE: some field positions already verified above, but duplicated
380 * here for completeness in case of future modifications.
381 */
382 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
383 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
384 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
385 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
386 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, vlan_tci) !=
387 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 10);
388 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, hash) !=
389 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 12);
390
391 /* Cache is empty -> need to scan the buffer rings, but first move
392 * the next 'n' mbufs into the cache
393 */
394 sw_ring = &rxq->sw_ring[rxq->rx_tail];
395
396 /* A. load 4 packet in one loop
397 * [A*. mask out 4 unused dirty field in desc]
398 * B. copy 4 mbuf point from swring to rx_pkts
399 * C. calc the number of DD bits among the 4 packets
400 * [C*. extract the end-of-packet bit, if requested]
401 * D. fill info. from desc to mbuf
402 */
403
404 for (pos = 0, nb_pkts_recd = 0; pos < nb_pkts;
405 pos += ICE_DESCS_PER_LOOP,
406 rxdp += ICE_DESCS_PER_LOOP) {
407 __m128i descs[ICE_DESCS_PER_LOOP];
408 __m128i pkt_mb0, pkt_mb1, pkt_mb2, pkt_mb3;
409 __m128i staterr, sterr_tmp1, sterr_tmp2;
410 /* 2 64 bit or 4 32 bit mbuf pointers in one XMM reg. */
411 __m128i mbp1;
412 #if defined(RTE_ARCH_X86_64)
413 __m128i mbp2;
414 #endif
415
416 /* B.1 load 2 (64 bit) or 4 (32 bit) mbuf points */
417 mbp1 = _mm_loadu_si128((__m128i *)&sw_ring[pos]);
418 /* Read desc statuses backwards to avoid race condition */
419 /* A.1 load desc[3] */
420 descs[3] = _mm_loadu_si128((__m128i *)(rxdp + 3));
421 rte_compiler_barrier();
422
423 /* B.2 copy 2 64 bit or 4 32 bit mbuf point into rx_pkts */
424 _mm_storeu_si128((__m128i *)&rx_pkts[pos], mbp1);
425
426 #if defined(RTE_ARCH_X86_64)
427 /* B.1 load 2 64 bit mbuf points */
428 mbp2 = _mm_loadu_si128((__m128i *)&sw_ring[pos + 2]);
429 #endif
430
431 /* A.1 load desc[2-0] */
432 descs[2] = _mm_loadu_si128((__m128i *)(rxdp + 2));
433 rte_compiler_barrier();
434 descs[1] = _mm_loadu_si128((__m128i *)(rxdp + 1));
435 rte_compiler_barrier();
436 descs[0] = _mm_loadu_si128((__m128i *)(rxdp));
437
438 #if defined(RTE_ARCH_X86_64)
439 /* B.2 copy 2 mbuf point into rx_pkts */
440 _mm_storeu_si128((__m128i *)&rx_pkts[pos + 2], mbp2);
441 #endif
442
443 if (split_packet) {
444 rte_mbuf_prefetch_part2(rx_pkts[pos]);
445 rte_mbuf_prefetch_part2(rx_pkts[pos + 1]);
446 rte_mbuf_prefetch_part2(rx_pkts[pos + 2]);
447 rte_mbuf_prefetch_part2(rx_pkts[pos + 3]);
448 }
449
450 /* avoid compiler reorder optimization */
451 rte_compiler_barrier();
452
453 /* D.1 pkt 3,4 convert format from desc to pktmbuf */
454 pkt_mb3 = _mm_shuffle_epi8(descs[3], shuf_msk);
455 pkt_mb2 = _mm_shuffle_epi8(descs[2], shuf_msk);
456
457 /* D.1 pkt 1,2 convert format from desc to pktmbuf */
458 pkt_mb1 = _mm_shuffle_epi8(descs[1], shuf_msk);
459 pkt_mb0 = _mm_shuffle_epi8(descs[0], shuf_msk);
460
461 /* C.1 4=>2 filter staterr info only */
462 sterr_tmp2 = _mm_unpackhi_epi32(descs[3], descs[2]);
463 /* C.1 4=>2 filter staterr info only */
464 sterr_tmp1 = _mm_unpackhi_epi32(descs[1], descs[0]);
465
466 ice_rx_desc_to_olflags_v(rxq, descs, &rx_pkts[pos]);
467
468 /* D.2 pkt 3,4 set in_port/nb_seg and remove crc */
469 pkt_mb3 = _mm_add_epi16(pkt_mb3, crc_adjust);
470 pkt_mb2 = _mm_add_epi16(pkt_mb2, crc_adjust);
471
472 /* D.2 pkt 1,2 set in_port/nb_seg and remove crc */
473 pkt_mb1 = _mm_add_epi16(pkt_mb1, crc_adjust);
474 pkt_mb0 = _mm_add_epi16(pkt_mb0, crc_adjust);
475
476 #ifndef RTE_LIBRTE_ICE_16BYTE_RX_DESC
477 /**
478 * needs to load 2nd 16B of each desc for RSS hash parsing,
479 * will cause performance drop to get into this context.
480 */
481 if (rxq->vsi->adapter->pf.dev_data->dev_conf.rxmode.offloads &
482 RTE_ETH_RX_OFFLOAD_RSS_HASH) {
483 /* load bottom half of every 32B desc */
484 const __m128i raw_desc_bh3 =
485 _mm_load_si128
486 ((void *)(&rxdp[3].wb.status_error1));
487 rte_compiler_barrier();
488 const __m128i raw_desc_bh2 =
489 _mm_load_si128
490 ((void *)(&rxdp[2].wb.status_error1));
491 rte_compiler_barrier();
492 const __m128i raw_desc_bh1 =
493 _mm_load_si128
494 ((void *)(&rxdp[1].wb.status_error1));
495 rte_compiler_barrier();
496 const __m128i raw_desc_bh0 =
497 _mm_load_si128
498 ((void *)(&rxdp[0].wb.status_error1));
499
500 /**
501 * to shift the 32b RSS hash value to the
502 * highest 32b of each 128b before mask
503 */
504 __m128i rss_hash3 =
505 _mm_slli_epi64(raw_desc_bh3, 32);
506 __m128i rss_hash2 =
507 _mm_slli_epi64(raw_desc_bh2, 32);
508 __m128i rss_hash1 =
509 _mm_slli_epi64(raw_desc_bh1, 32);
510 __m128i rss_hash0 =
511 _mm_slli_epi64(raw_desc_bh0, 32);
512
513 __m128i rss_hash_msk =
514 _mm_set_epi32(0xFFFFFFFF, 0, 0, 0);
515
516 rss_hash3 = _mm_and_si128
517 (rss_hash3, rss_hash_msk);
518 rss_hash2 = _mm_and_si128
519 (rss_hash2, rss_hash_msk);
520 rss_hash1 = _mm_and_si128
521 (rss_hash1, rss_hash_msk);
522 rss_hash0 = _mm_and_si128
523 (rss_hash0, rss_hash_msk);
524
525 pkt_mb3 = _mm_or_si128(pkt_mb3, rss_hash3);
526 pkt_mb2 = _mm_or_si128(pkt_mb2, rss_hash2);
527 pkt_mb1 = _mm_or_si128(pkt_mb1, rss_hash1);
528 pkt_mb0 = _mm_or_si128(pkt_mb0, rss_hash0);
529 } /* if() on RSS hash parsing */
530 #endif
531
532 /* C.2 get 4 pkts staterr value */
533 staterr = _mm_unpacklo_epi32(sterr_tmp1, sterr_tmp2);
534
535 /* D.3 copy final 3,4 data to rx_pkts */
536 _mm_storeu_si128
537 ((void *)&rx_pkts[pos + 3]->rx_descriptor_fields1,
538 pkt_mb3);
539 _mm_storeu_si128
540 ((void *)&rx_pkts[pos + 2]->rx_descriptor_fields1,
541 pkt_mb2);
542
543 /* C* extract and record EOP bit */
544 if (split_packet) {
545 /* and with mask to extract bits, flipping 1-0 */
546 __m128i eop_bits = _mm_andnot_si128(staterr, eop_check);
547 /* the staterr values are not in order, as the count
548 * of dd bits doesn't care. However, for end of
549 * packet tracking, we do care, so shuffle. This also
550 * compresses the 32-bit values to 8-bit
551 */
552 eop_bits = _mm_shuffle_epi8(eop_bits, eop_shuf_mask);
553 /* store the resulting 32-bit value */
554 *(int *)split_packet = _mm_cvtsi128_si32(eop_bits);
555 split_packet += ICE_DESCS_PER_LOOP;
556 }
557
558 /* C.3 calc available number of desc */
559 staterr = _mm_and_si128(staterr, dd_check);
560 staterr = _mm_packs_epi32(staterr, zero);
561
562 /* D.3 copy final 1,2 data to rx_pkts */
563 _mm_storeu_si128
564 ((void *)&rx_pkts[pos + 1]->rx_descriptor_fields1,
565 pkt_mb1);
566 _mm_storeu_si128((void *)&rx_pkts[pos]->rx_descriptor_fields1,
567 pkt_mb0);
568 ice_rx_desc_to_ptype_v(descs, &rx_pkts[pos], ptype_tbl);
569 /* C.4 calc available number of desc */
570 var = __builtin_popcountll(_mm_cvtsi128_si64(staterr));
571 nb_pkts_recd += var;
572 if (likely(var != ICE_DESCS_PER_LOOP))
573 break;
574 }
575
576 /* Update our internal tail pointer */
577 rxq->rx_tail = (uint16_t)(rxq->rx_tail + nb_pkts_recd);
578 rxq->rx_tail = (uint16_t)(rxq->rx_tail & (rxq->nb_rx_desc - 1));
579 rxq->rxrearm_nb = (uint16_t)(rxq->rxrearm_nb + nb_pkts_recd);
580
581 return nb_pkts_recd;
582 }
583
584 /**
585 * Notice:
586 * - nb_pkts < ICE_DESCS_PER_LOOP, just return no packet
587 * - nb_pkts > ICE_VPMD_RX_BURST, only scan ICE_VPMD_RX_BURST
588 * numbers of DD bits
589 */
590 uint16_t
ice_recv_pkts_vec(void * rx_queue,struct rte_mbuf ** rx_pkts,uint16_t nb_pkts)591 ice_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
592 uint16_t nb_pkts)
593 {
594 return _ice_recv_raw_pkts_vec(rx_queue, rx_pkts, nb_pkts, NULL);
595 }
596
597 /**
598 * vPMD receive routine that reassembles single burst of 32 scattered packets
599 *
600 * Notice:
601 * - nb_pkts < ICE_DESCS_PER_LOOP, just return no packet
602 */
603 static uint16_t
ice_recv_scattered_burst_vec(void * rx_queue,struct rte_mbuf ** rx_pkts,uint16_t nb_pkts)604 ice_recv_scattered_burst_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
605 uint16_t nb_pkts)
606 {
607 struct ice_rx_queue *rxq = rx_queue;
608 uint8_t split_flags[ICE_VPMD_RX_BURST] = {0};
609
610 /* get some new buffers */
611 uint16_t nb_bufs = _ice_recv_raw_pkts_vec(rxq, rx_pkts, nb_pkts,
612 split_flags);
613 if (nb_bufs == 0)
614 return 0;
615
616 /* happy day case, full burst + no packets to be joined */
617 const uint64_t *split_fl64 = (uint64_t *)split_flags;
618
619 if (!rxq->pkt_first_seg &&
620 split_fl64[0] == 0 && split_fl64[1] == 0 &&
621 split_fl64[2] == 0 && split_fl64[3] == 0)
622 return nb_bufs;
623
624 /* reassemble any packets that need reassembly*/
625 unsigned int i = 0;
626
627 if (!rxq->pkt_first_seg) {
628 /* find the first split flag, and only reassemble then*/
629 while (i < nb_bufs && !split_flags[i])
630 i++;
631 if (i == nb_bufs)
632 return nb_bufs;
633 rxq->pkt_first_seg = rx_pkts[i];
634 }
635 return i + ice_rx_reassemble_packets(rxq, &rx_pkts[i], nb_bufs - i,
636 &split_flags[i]);
637 }
638
639 /**
640 * vPMD receive routine that reassembles scattered packets.
641 */
642 uint16_t
ice_recv_scattered_pkts_vec(void * rx_queue,struct rte_mbuf ** rx_pkts,uint16_t nb_pkts)643 ice_recv_scattered_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
644 uint16_t nb_pkts)
645 {
646 uint16_t retval = 0;
647
648 while (nb_pkts > ICE_VPMD_RX_BURST) {
649 uint16_t burst;
650
651 burst = ice_recv_scattered_burst_vec(rx_queue,
652 rx_pkts + retval,
653 ICE_VPMD_RX_BURST);
654 retval += burst;
655 nb_pkts -= burst;
656 if (burst < ICE_VPMD_RX_BURST)
657 return retval;
658 }
659
660 return retval + ice_recv_scattered_burst_vec(rx_queue,
661 rx_pkts + retval,
662 nb_pkts);
663 }
664
665 static inline void
ice_vtx1(volatile struct ice_tx_desc * txdp,struct rte_mbuf * pkt,uint64_t flags)666 ice_vtx1(volatile struct ice_tx_desc *txdp, struct rte_mbuf *pkt,
667 uint64_t flags)
668 {
669 uint64_t high_qw =
670 (ICE_TX_DESC_DTYPE_DATA |
671 ((uint64_t)flags << ICE_TXD_QW1_CMD_S) |
672 ((uint64_t)pkt->data_len << ICE_TXD_QW1_TX_BUF_SZ_S));
673
674 __m128i descriptor = _mm_set_epi64x(high_qw,
675 pkt->buf_iova + pkt->data_off);
676 _mm_store_si128((__m128i *)txdp, descriptor);
677 }
678
679 static inline void
ice_vtx(volatile struct ice_tx_desc * txdp,struct rte_mbuf ** pkt,uint16_t nb_pkts,uint64_t flags)680 ice_vtx(volatile struct ice_tx_desc *txdp, struct rte_mbuf **pkt,
681 uint16_t nb_pkts, uint64_t flags)
682 {
683 int i;
684
685 for (i = 0; i < nb_pkts; ++i, ++txdp, ++pkt)
686 ice_vtx1(txdp, *pkt, flags);
687 }
688
689 static uint16_t
ice_xmit_fixed_burst_vec(void * tx_queue,struct rte_mbuf ** tx_pkts,uint16_t nb_pkts)690 ice_xmit_fixed_burst_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
691 uint16_t nb_pkts)
692 {
693 struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue;
694 volatile struct ice_tx_desc *txdp;
695 struct ice_tx_entry *txep;
696 uint16_t n, nb_commit, tx_id;
697 uint64_t flags = ICE_TD_CMD;
698 uint64_t rs = ICE_TX_DESC_CMD_RS | ICE_TD_CMD;
699 int i;
700
701 /* cross rx_thresh boundary is not allowed */
702 nb_pkts = RTE_MIN(nb_pkts, txq->tx_rs_thresh);
703
704 if (txq->nb_tx_free < txq->tx_free_thresh)
705 ice_tx_free_bufs_vec(txq);
706
707 nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
708 nb_commit = nb_pkts;
709 if (unlikely(nb_pkts == 0))
710 return 0;
711
712 tx_id = txq->tx_tail;
713 txdp = &txq->tx_ring[tx_id];
714 txep = &txq->sw_ring[tx_id];
715
716 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
717
718 n = (uint16_t)(txq->nb_tx_desc - tx_id);
719 if (nb_commit >= n) {
720 ice_tx_backlog_entry(txep, tx_pkts, n);
721
722 for (i = 0; i < n - 1; ++i, ++tx_pkts, ++txdp)
723 ice_vtx1(txdp, *tx_pkts, flags);
724
725 ice_vtx1(txdp, *tx_pkts++, rs);
726
727 nb_commit = (uint16_t)(nb_commit - n);
728
729 tx_id = 0;
730 txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
731
732 /* avoid reach the end of ring */
733 txdp = &txq->tx_ring[tx_id];
734 txep = &txq->sw_ring[tx_id];
735 }
736
737 ice_tx_backlog_entry(txep, tx_pkts, nb_commit);
738
739 ice_vtx(txdp, tx_pkts, nb_commit, flags);
740
741 tx_id = (uint16_t)(tx_id + nb_commit);
742 if (tx_id > txq->tx_next_rs) {
743 txq->tx_ring[txq->tx_next_rs].cmd_type_offset_bsz |=
744 rte_cpu_to_le_64(((uint64_t)ICE_TX_DESC_CMD_RS) <<
745 ICE_TXD_QW1_CMD_S);
746 txq->tx_next_rs =
747 (uint16_t)(txq->tx_next_rs + txq->tx_rs_thresh);
748 }
749
750 txq->tx_tail = tx_id;
751
752 ICE_PCI_REG_WC_WRITE(txq->qtx_tail, txq->tx_tail);
753
754 return nb_pkts;
755 }
756
757 uint16_t
ice_xmit_pkts_vec(void * tx_queue,struct rte_mbuf ** tx_pkts,uint16_t nb_pkts)758 ice_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
759 uint16_t nb_pkts)
760 {
761 uint16_t nb_tx = 0;
762 struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue;
763
764 while (nb_pkts) {
765 uint16_t ret, num;
766
767 num = (uint16_t)RTE_MIN(nb_pkts, txq->tx_rs_thresh);
768 ret = ice_xmit_fixed_burst_vec(tx_queue, &tx_pkts[nb_tx], num);
769 nb_tx += ret;
770 nb_pkts -= ret;
771 if (ret < num)
772 break;
773 }
774
775 return nb_tx;
776 }
777
778 int __rte_cold
ice_rxq_vec_setup(struct ice_rx_queue * rxq)779 ice_rxq_vec_setup(struct ice_rx_queue *rxq)
780 {
781 if (!rxq)
782 return -1;
783
784 rxq->rx_rel_mbufs = _ice_rx_queue_release_mbufs_vec;
785 return ice_rxq_vec_setup_default(rxq);
786 }
787
788 int __rte_cold
ice_txq_vec_setup(struct ice_tx_queue __rte_unused * txq)789 ice_txq_vec_setup(struct ice_tx_queue __rte_unused *txq)
790 {
791 if (!txq)
792 return -1;
793
794 txq->tx_rel_mbufs = _ice_tx_queue_release_mbufs_vec;
795 return 0;
796 }
797
798 int __rte_cold
ice_rx_vec_dev_check(struct rte_eth_dev * dev)799 ice_rx_vec_dev_check(struct rte_eth_dev *dev)
800 {
801 return ice_rx_vec_dev_check_default(dev);
802 }
803
804 int __rte_cold
ice_tx_vec_dev_check(struct rte_eth_dev * dev)805 ice_tx_vec_dev_check(struct rte_eth_dev *dev)
806 {
807 return ice_tx_vec_dev_check_default(dev);
808 }
809