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/iavf_prototype.h"
10 #include "base/iavf_type.h"
12 #include "iavf_rxtx.h"
13 #include "iavf_rxtx_vec_common.h"
15 #include <tmmintrin.h>
17 #ifndef __INTEL_COMPILER
18 #pragma GCC diagnostic ignored "-Wcast-qual"
22 iavf_rxq_rearm(struct iavf_rx_queue *rxq)
27 volatile union iavf_rx_desc *rxdp;
28 struct rte_mbuf **rxp = &rxq->sw_ring[rxq->rxrearm_start];
29 struct rte_mbuf *mb0, *mb1;
30 __m128i hdr_room = _mm_set_epi64x(RTE_PKTMBUF_HEADROOM,
31 RTE_PKTMBUF_HEADROOM);
32 __m128i dma_addr0, dma_addr1;
34 rxdp = rxq->rx_ring + rxq->rxrearm_start;
36 /* Pull 'n' more MBUFs into the software ring */
37 if (rte_mempool_get_bulk(rxq->mp, (void *)rxp,
38 rxq->rx_free_thresh) < 0) {
39 if (rxq->rxrearm_nb + rxq->rx_free_thresh >= rxq->nb_rx_desc) {
40 dma_addr0 = _mm_setzero_si128();
41 for (i = 0; i < IAVF_VPMD_DESCS_PER_LOOP; i++) {
42 rxp[i] = &rxq->fake_mbuf;
43 _mm_store_si128((__m128i *)&rxdp[i].read,
47 rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
52 /* Initialize the mbufs in vector, process 2 mbufs in one loop */
53 for (i = 0; i < rxq->rx_free_thresh; i += 2, rxp += 2) {
54 __m128i vaddr0, vaddr1;
59 /* load buf_addr(lo 64bit) and buf_iova(hi 64bit) */
60 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, buf_iova) !=
61 offsetof(struct rte_mbuf, buf_addr) + 8);
62 vaddr0 = _mm_loadu_si128((__m128i *)&mb0->buf_addr);
63 vaddr1 = _mm_loadu_si128((__m128i *)&mb1->buf_addr);
65 /* convert pa to dma_addr hdr/data */
66 dma_addr0 = _mm_unpackhi_epi64(vaddr0, vaddr0);
67 dma_addr1 = _mm_unpackhi_epi64(vaddr1, vaddr1);
69 /* add headroom to pa values */
70 dma_addr0 = _mm_add_epi64(dma_addr0, hdr_room);
71 dma_addr1 = _mm_add_epi64(dma_addr1, hdr_room);
73 /* flush desc with pa dma_addr */
74 _mm_store_si128((__m128i *)&rxdp++->read, dma_addr0);
75 _mm_store_si128((__m128i *)&rxdp++->read, dma_addr1);
78 rxq->rxrearm_start += rxq->rx_free_thresh;
79 if (rxq->rxrearm_start >= rxq->nb_rx_desc)
80 rxq->rxrearm_start = 0;
82 rxq->rxrearm_nb -= rxq->rx_free_thresh;
84 rx_id = (uint16_t)((rxq->rxrearm_start == 0) ?
85 (rxq->nb_rx_desc - 1) : (rxq->rxrearm_start - 1));
87 PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_tail=%u "
88 "rearm_start=%u rearm_nb=%u",
89 rxq->port_id, rxq->queue_id,
90 rx_id, rxq->rxrearm_start, rxq->rxrearm_nb);
92 /* Update the tail pointer on the NIC */
93 IAVF_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
97 desc_to_olflags_v(struct iavf_rx_queue *rxq, __m128i descs[4],
98 struct rte_mbuf **rx_pkts)
100 const __m128i mbuf_init = _mm_set_epi64x(0, rxq->mbuf_initializer);
101 __m128i rearm0, rearm1, rearm2, rearm3;
103 __m128i vlan0, vlan1, rss, l3_l4e;
105 /* mask everything except RSS, flow director and VLAN flags
106 * bit2 is for VLAN tag, bit11 for flow director indication
107 * bit13:12 for RSS indication.
109 const __m128i rss_vlan_msk = _mm_set_epi32(
110 0x1c03804, 0x1c03804, 0x1c03804, 0x1c03804);
112 const __m128i cksum_mask = _mm_set_epi32(
113 PKT_RX_IP_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD |
114 PKT_RX_L4_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD |
115 PKT_RX_EIP_CKSUM_BAD,
116 PKT_RX_IP_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD |
117 PKT_RX_L4_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD |
118 PKT_RX_EIP_CKSUM_BAD,
119 PKT_RX_IP_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD |
120 PKT_RX_L4_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD |
121 PKT_RX_EIP_CKSUM_BAD,
122 PKT_RX_IP_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD |
123 PKT_RX_L4_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD |
124 PKT_RX_EIP_CKSUM_BAD);
126 /* map rss and vlan type to rss hash and vlan flag */
127 const __m128i vlan_flags = _mm_set_epi8(0, 0, 0, 0,
129 0, 0, 0, PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
132 const __m128i rss_flags = _mm_set_epi8(0, 0, 0, 0,
134 PKT_RX_RSS_HASH | PKT_RX_FDIR, PKT_RX_RSS_HASH, 0, 0,
135 0, 0, PKT_RX_FDIR, 0);
137 const __m128i l3_l4e_flags = _mm_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
138 /* shift right 1 bit to make sure it not exceed 255 */
139 (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
140 PKT_RX_IP_CKSUM_BAD) >> 1,
141 (PKT_RX_IP_CKSUM_GOOD | PKT_RX_EIP_CKSUM_BAD |
142 PKT_RX_L4_CKSUM_BAD) >> 1,
143 (PKT_RX_EIP_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
144 (PKT_RX_IP_CKSUM_GOOD | PKT_RX_EIP_CKSUM_BAD) >> 1,
145 (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
146 (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD) >> 1,
147 PKT_RX_IP_CKSUM_BAD >> 1,
148 (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1);
150 vlan0 = _mm_unpackhi_epi32(descs[0], descs[1]);
151 vlan1 = _mm_unpackhi_epi32(descs[2], descs[3]);
152 vlan0 = _mm_unpacklo_epi64(vlan0, vlan1);
154 vlan1 = _mm_and_si128(vlan0, rss_vlan_msk);
155 vlan0 = _mm_shuffle_epi8(vlan_flags, vlan1);
157 rss = _mm_srli_epi32(vlan1, 11);
158 rss = _mm_shuffle_epi8(rss_flags, rss);
160 l3_l4e = _mm_srli_epi32(vlan1, 22);
161 l3_l4e = _mm_shuffle_epi8(l3_l4e_flags, l3_l4e);
162 /* then we shift left 1 bit */
163 l3_l4e = _mm_slli_epi32(l3_l4e, 1);
164 /* we need to mask out the reduntant bits */
165 l3_l4e = _mm_and_si128(l3_l4e, cksum_mask);
167 vlan0 = _mm_or_si128(vlan0, rss);
168 vlan0 = _mm_or_si128(vlan0, l3_l4e);
170 /* At this point, we have the 4 sets of flags in the low 16-bits
171 * of each 32-bit value in vlan0.
172 * We want to extract these, and merge them with the mbuf init data
173 * so we can do a single 16-byte write to the mbuf to set the flags
174 * and all the other initialization fields. Extracting the
175 * appropriate flags means that we have to do a shift and blend for
176 * each mbuf before we do the write.
178 rearm0 = _mm_blend_epi16(mbuf_init, _mm_slli_si128(vlan0, 8), 0x10);
179 rearm1 = _mm_blend_epi16(mbuf_init, _mm_slli_si128(vlan0, 4), 0x10);
180 rearm2 = _mm_blend_epi16(mbuf_init, vlan0, 0x10);
181 rearm3 = _mm_blend_epi16(mbuf_init, _mm_srli_si128(vlan0, 4), 0x10);
183 /* write the rearm data and the olflags in one write */
184 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, ol_flags) !=
185 offsetof(struct rte_mbuf, rearm_data) + 8);
186 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, rearm_data) !=
187 RTE_ALIGN(offsetof(struct rte_mbuf, rearm_data), 16));
188 _mm_store_si128((__m128i *)&rx_pkts[0]->rearm_data, rearm0);
189 _mm_store_si128((__m128i *)&rx_pkts[1]->rearm_data, rearm1);
190 _mm_store_si128((__m128i *)&rx_pkts[2]->rearm_data, rearm2);
191 _mm_store_si128((__m128i *)&rx_pkts[3]->rearm_data, rearm3);
194 #define PKTLEN_SHIFT 10
197 desc_to_ptype_v(__m128i descs[4], struct rte_mbuf **rx_pkts)
199 __m128i ptype0 = _mm_unpackhi_epi64(descs[0], descs[1]);
200 __m128i ptype1 = _mm_unpackhi_epi64(descs[2], descs[3]);
201 static const uint32_t type_table[UINT8_MAX + 1] __rte_cache_aligned = {
203 [1] = RTE_PTYPE_L2_ETHER,
204 /* [2] - [21] reserved */
205 [22] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
207 [23] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
208 RTE_PTYPE_L4_NONFRAG,
209 [24] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
212 [26] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
214 [27] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
216 [28] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
218 /* All others reserved */
221 ptype0 = _mm_srli_epi64(ptype0, 30);
222 ptype1 = _mm_srli_epi64(ptype1, 30);
224 rx_pkts[0]->packet_type = type_table[_mm_extract_epi8(ptype0, 0)];
225 rx_pkts[1]->packet_type = type_table[_mm_extract_epi8(ptype0, 8)];
226 rx_pkts[2]->packet_type = type_table[_mm_extract_epi8(ptype1, 0)];
227 rx_pkts[3]->packet_type = type_table[_mm_extract_epi8(ptype1, 8)];
231 * - nb_pkts < IAVF_VPMD_DESCS_PER_LOOP, just return no packet
232 * - nb_pkts > IAVF_VPMD_RX_MAX_BURST, only scan IAVF_VPMD_RX_MAX_BURST
235 static inline uint16_t
236 _recv_raw_pkts_vec(struct iavf_rx_queue *rxq, struct rte_mbuf **rx_pkts,
237 uint16_t nb_pkts, uint8_t *split_packet)
239 volatile union iavf_rx_desc *rxdp;
240 struct rte_mbuf **sw_ring;
241 uint16_t nb_pkts_recd;
246 __m128i crc_adjust = _mm_set_epi16(
247 0, 0, 0, /* ignore non-length fields */
248 -rxq->crc_len, /* sub crc on data_len */
249 0, /* ignore high-16bits of pkt_len */
250 -rxq->crc_len, /* sub crc on pkt_len */
251 0, 0 /* ignore pkt_type field */
253 /* compile-time check the above crc_adjust layout is correct.
254 * NOTE: the first field (lowest address) is given last in set_epi16
257 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
258 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
259 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
260 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
261 __m128i dd_check, eop_check;
263 /* nb_pkts shall be less equal than IAVF_VPMD_RX_MAX_BURST */
264 nb_pkts = RTE_MIN(nb_pkts, IAVF_VPMD_RX_MAX_BURST);
266 /* nb_pkts has to be floor-aligned to IAVF_VPMD_DESCS_PER_LOOP */
267 nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, IAVF_VPMD_DESCS_PER_LOOP);
269 /* Just the act of getting into the function from the application is
270 * going to cost about 7 cycles
272 rxdp = rxq->rx_ring + rxq->rx_tail;
276 /* See if we need to rearm the RX queue - gives the prefetch a bit
279 if (rxq->rxrearm_nb > rxq->rx_free_thresh)
282 /* Before we start moving massive data around, check to see if
283 * there is actually a packet available
285 if (!(rxdp->wb.qword1.status_error_len &
286 rte_cpu_to_le_32(1 << IAVF_RX_DESC_STATUS_DD_SHIFT)))
289 /* 4 packets DD mask */
290 dd_check = _mm_set_epi64x(0x0000000100000001LL, 0x0000000100000001LL);
292 /* 4 packets EOP mask */
293 eop_check = _mm_set_epi64x(0x0000000200000002LL, 0x0000000200000002LL);
295 /* mask to shuffle from desc. to mbuf */
296 shuf_msk = _mm_set_epi8(
297 7, 6, 5, 4, /* octet 4~7, 32bits rss */
298 3, 2, /* octet 2~3, low 16 bits vlan_macip */
299 15, 14, /* octet 15~14, 16 bits data_len */
300 0xFF, 0xFF, /* skip high 16 bits pkt_len, zero out */
301 15, 14, /* octet 15~14, low 16 bits pkt_len */
302 0xFF, 0xFF, 0xFF, 0xFF /* pkt_type set as unknown */
304 /* Compile-time verify the shuffle mask
305 * NOTE: some field positions already verified above, but duplicated
306 * here for completeness in case of future modifications.
308 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, pkt_len) !=
309 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 4);
310 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, data_len) !=
311 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 8);
312 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, vlan_tci) !=
313 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 10);
314 RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, hash) !=
315 offsetof(struct rte_mbuf, rx_descriptor_fields1) + 12);
317 /* Cache is empty -> need to scan the buffer rings, but first move
318 * the next 'n' mbufs into the cache
320 sw_ring = &rxq->sw_ring[rxq->rx_tail];
322 /* A. load 4 packet in one loop
323 * [A*. mask out 4 unused dirty field in desc]
324 * B. copy 4 mbuf point from swring to rx_pkts
325 * C. calc the number of DD bits among the 4 packets
326 * [C*. extract the end-of-packet bit, if requested]
327 * D. fill info. from desc to mbuf
330 for (pos = 0, nb_pkts_recd = 0; pos < nb_pkts;
331 pos += IAVF_VPMD_DESCS_PER_LOOP,
332 rxdp += IAVF_VPMD_DESCS_PER_LOOP) {
333 __m128i descs[IAVF_VPMD_DESCS_PER_LOOP];
334 __m128i pkt_mb1, pkt_mb2, pkt_mb3, pkt_mb4;
335 __m128i zero, staterr, sterr_tmp1, sterr_tmp2;
336 /* 2 64 bit or 4 32 bit mbuf pointers in one XMM reg. */
338 #if defined(RTE_ARCH_X86_64)
342 /* B.1 load 2 (64 bit) or 4 (32 bit) mbuf points */
343 mbp1 = _mm_loadu_si128((__m128i *)&sw_ring[pos]);
344 /* Read desc statuses backwards to avoid race condition */
345 /* A.1 load 4 pkts desc */
346 descs[3] = _mm_loadu_si128((__m128i *)(rxdp + 3));
347 rte_compiler_barrier();
349 /* B.2 copy 2 64 bit or 4 32 bit mbuf point into rx_pkts */
350 _mm_storeu_si128((__m128i *)&rx_pkts[pos], mbp1);
352 #if defined(RTE_ARCH_X86_64)
353 /* B.1 load 2 64 bit mbuf points */
354 mbp2 = _mm_loadu_si128((__m128i *)&sw_ring[pos + 2]);
357 descs[2] = _mm_loadu_si128((__m128i *)(rxdp + 2));
358 rte_compiler_barrier();
359 /* B.1 load 2 mbuf point */
360 descs[1] = _mm_loadu_si128((__m128i *)(rxdp + 1));
361 rte_compiler_barrier();
362 descs[0] = _mm_loadu_si128((__m128i *)(rxdp));
364 #if defined(RTE_ARCH_X86_64)
365 /* B.2 copy 2 mbuf point into rx_pkts */
366 _mm_storeu_si128((__m128i *)&rx_pkts[pos + 2], mbp2);
370 rte_mbuf_prefetch_part2(rx_pkts[pos]);
371 rte_mbuf_prefetch_part2(rx_pkts[pos + 1]);
372 rte_mbuf_prefetch_part2(rx_pkts[pos + 2]);
373 rte_mbuf_prefetch_part2(rx_pkts[pos + 3]);
376 /* avoid compiler reorder optimization */
377 rte_compiler_barrier();
379 /* pkt 3,4 shift the pktlen field to be 16-bit aligned*/
380 const __m128i len3 = _mm_slli_epi32(descs[3], PKTLEN_SHIFT);
381 const __m128i len2 = _mm_slli_epi32(descs[2], PKTLEN_SHIFT);
383 /* merge the now-aligned packet length fields back in */
384 descs[3] = _mm_blend_epi16(descs[3], len3, 0x80);
385 descs[2] = _mm_blend_epi16(descs[2], len2, 0x80);
387 /* D.1 pkt 3,4 convert format from desc to pktmbuf */
388 pkt_mb4 = _mm_shuffle_epi8(descs[3], shuf_msk);
389 pkt_mb3 = _mm_shuffle_epi8(descs[2], shuf_msk);
391 /* C.1 4=>2 status err info only */
392 sterr_tmp2 = _mm_unpackhi_epi32(descs[3], descs[2]);
393 sterr_tmp1 = _mm_unpackhi_epi32(descs[1], descs[0]);
395 desc_to_olflags_v(rxq, descs, &rx_pkts[pos]);
397 /* D.2 pkt 3,4 set in_port/nb_seg and remove crc */
398 pkt_mb4 = _mm_add_epi16(pkt_mb4, crc_adjust);
399 pkt_mb3 = _mm_add_epi16(pkt_mb3, crc_adjust);
401 /* pkt 1,2 shift the pktlen field to be 16-bit aligned*/
402 const __m128i len1 = _mm_slli_epi32(descs[1], PKTLEN_SHIFT);
403 const __m128i len0 = _mm_slli_epi32(descs[0], PKTLEN_SHIFT);
405 /* merge the now-aligned packet length fields back in */
406 descs[1] = _mm_blend_epi16(descs[1], len1, 0x80);
407 descs[0] = _mm_blend_epi16(descs[0], len0, 0x80);
409 /* D.1 pkt 1,2 convert format from desc to pktmbuf */
410 pkt_mb2 = _mm_shuffle_epi8(descs[1], shuf_msk);
411 pkt_mb1 = _mm_shuffle_epi8(descs[0], shuf_msk);
413 /* C.2 get 4 pkts status err value */
414 zero = _mm_xor_si128(dd_check, dd_check);
415 staterr = _mm_unpacklo_epi32(sterr_tmp1, sterr_tmp2);
417 /* D.3 copy final 3,4 data to rx_pkts */
419 (void *)&rx_pkts[pos + 3]->rx_descriptor_fields1,
422 (void *)&rx_pkts[pos + 2]->rx_descriptor_fields1,
425 /* D.2 pkt 1,2 remove crc */
426 pkt_mb2 = _mm_add_epi16(pkt_mb2, crc_adjust);
427 pkt_mb1 = _mm_add_epi16(pkt_mb1, crc_adjust);
429 /* C* extract and record EOP bit */
431 __m128i eop_shuf_mask = _mm_set_epi8(
432 0xFF, 0xFF, 0xFF, 0xFF,
433 0xFF, 0xFF, 0xFF, 0xFF,
434 0xFF, 0xFF, 0xFF, 0xFF,
435 0x04, 0x0C, 0x00, 0x08
438 /* and with mask to extract bits, flipping 1-0 */
439 __m128i eop_bits = _mm_andnot_si128(staterr, eop_check);
440 /* the staterr values are not in order, as the count
441 * count of dd bits doesn't care. However, for end of
442 * packet tracking, we do care, so shuffle. This also
443 * compresses the 32-bit values to 8-bit
445 eop_bits = _mm_shuffle_epi8(eop_bits, eop_shuf_mask);
446 /* store the resulting 32-bit value */
447 *(int *)split_packet = _mm_cvtsi128_si32(eop_bits);
448 split_packet += IAVF_VPMD_DESCS_PER_LOOP;
451 /* C.3 calc available number of desc */
452 staterr = _mm_and_si128(staterr, dd_check);
453 staterr = _mm_packs_epi32(staterr, zero);
455 /* D.3 copy final 1,2 data to rx_pkts */
457 (void *)&rx_pkts[pos + 1]->rx_descriptor_fields1,
459 _mm_storeu_si128((void *)&rx_pkts[pos]->rx_descriptor_fields1,
461 desc_to_ptype_v(descs, &rx_pkts[pos]);
462 /* C.4 calc avaialbe number of desc */
463 var = __builtin_popcountll(_mm_cvtsi128_si64(staterr));
465 if (likely(var != IAVF_VPMD_DESCS_PER_LOOP))
469 /* Update our internal tail pointer */
470 rxq->rx_tail = (uint16_t)(rxq->rx_tail + nb_pkts_recd);
471 rxq->rx_tail = (uint16_t)(rxq->rx_tail & (rxq->nb_rx_desc - 1));
472 rxq->rxrearm_nb = (uint16_t)(rxq->rxrearm_nb + nb_pkts_recd);
478 * - nb_pkts < IAVF_DESCS_PER_LOOP, just return no packet
479 * - nb_pkts > IAVF_VPMD_RX_MAX_BURST, only scan IAVF_VPMD_RX_MAX_BURST
483 iavf_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
486 return _recv_raw_pkts_vec(rx_queue, rx_pkts, nb_pkts, NULL);
489 /* vPMD receive routine that reassembles scattered packets
491 * - nb_pkts < IAVF_VPMD_DESCS_PER_LOOP, just return no packet
492 * - nb_pkts > VPMD_RX_MAX_BURST, only scan IAVF_VPMD_RX_MAX_BURST
496 iavf_recv_scattered_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
499 struct iavf_rx_queue *rxq = rx_queue;
500 uint8_t split_flags[IAVF_VPMD_RX_MAX_BURST] = {0};
503 /* get some new buffers */
504 uint16_t nb_bufs = _recv_raw_pkts_vec(rxq, rx_pkts, nb_pkts,
509 /* happy day case, full burst + no packets to be joined */
510 const uint64_t *split_fl64 = (uint64_t *)split_flags;
512 if (!rxq->pkt_first_seg &&
513 split_fl64[0] == 0 && split_fl64[1] == 0 &&
514 split_fl64[2] == 0 && split_fl64[3] == 0)
517 /* reassemble any packets that need reassembly*/
518 if (!rxq->pkt_first_seg) {
519 /* find the first split flag, and only reassemble then*/
520 while (i < nb_bufs && !split_flags[i])
524 rxq->pkt_first_seg = rx_pkts[i];
526 return i + reassemble_packets(rxq, &rx_pkts[i], nb_bufs - i,
531 vtx1(volatile struct iavf_tx_desc *txdp, struct rte_mbuf *pkt, uint64_t flags)
534 (IAVF_TX_DESC_DTYPE_DATA |
535 ((uint64_t)flags << IAVF_TXD_QW1_CMD_SHIFT) |
536 ((uint64_t)pkt->data_len <<
537 IAVF_TXD_QW1_TX_BUF_SZ_SHIFT));
539 __m128i descriptor = _mm_set_epi64x(high_qw,
540 pkt->buf_iova + pkt->data_off);
541 _mm_store_si128((__m128i *)txdp, descriptor);
545 iavf_vtx(volatile struct iavf_tx_desc *txdp, struct rte_mbuf **pkt,
546 uint16_t nb_pkts, uint64_t flags)
550 for (i = 0; i < nb_pkts; ++i, ++txdp, ++pkt)
551 vtx1(txdp, *pkt, flags);
555 iavf_xmit_fixed_burst_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
558 struct iavf_tx_queue *txq = (struct iavf_tx_queue *)tx_queue;
559 volatile struct iavf_tx_desc *txdp;
560 struct iavf_tx_entry *txep;
561 uint16_t n, nb_commit, tx_id;
562 uint64_t flags = IAVF_TX_DESC_CMD_EOP | 0x04; /* bit 2 must be set */
563 uint64_t rs = IAVF_TX_DESC_CMD_RS | flags;
566 /* cross rx_thresh boundary is not allowed */
567 nb_pkts = RTE_MIN(nb_pkts, txq->rs_thresh);
569 if (txq->nb_free < txq->free_thresh)
570 iavf_tx_free_bufs(txq);
572 nb_pkts = (uint16_t)RTE_MIN(txq->nb_free, nb_pkts);
573 if (unlikely(nb_pkts == 0))
577 tx_id = txq->tx_tail;
578 txdp = &txq->tx_ring[tx_id];
579 txep = &txq->sw_ring[tx_id];
581 txq->nb_free = (uint16_t)(txq->nb_free - nb_pkts);
583 n = (uint16_t)(txq->nb_tx_desc - tx_id);
584 if (nb_commit >= n) {
585 tx_backlog_entry(txep, tx_pkts, n);
587 for (i = 0; i < n - 1; ++i, ++tx_pkts, ++txdp)
588 vtx1(txdp, *tx_pkts, flags);
590 vtx1(txdp, *tx_pkts++, rs);
592 nb_commit = (uint16_t)(nb_commit - n);
595 txq->next_rs = (uint16_t)(txq->rs_thresh - 1);
597 /* avoid reach the end of ring */
598 txdp = &txq->tx_ring[tx_id];
599 txep = &txq->sw_ring[tx_id];
602 tx_backlog_entry(txep, tx_pkts, nb_commit);
604 iavf_vtx(txdp, tx_pkts, nb_commit, flags);
606 tx_id = (uint16_t)(tx_id + nb_commit);
607 if (tx_id > txq->next_rs) {
608 txq->tx_ring[txq->next_rs].cmd_type_offset_bsz |=
609 rte_cpu_to_le_64(((uint64_t)IAVF_TX_DESC_CMD_RS) <<
610 IAVF_TXD_QW1_CMD_SHIFT);
612 (uint16_t)(txq->next_rs + txq->rs_thresh);
615 txq->tx_tail = tx_id;
617 PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u tx_tail=%u nb_pkts=%u",
618 txq->port_id, txq->queue_id, tx_id, nb_pkts);
620 IAVF_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
625 static void __attribute__((cold))
626 iavf_rx_queue_release_mbufs_sse(struct iavf_rx_queue *rxq)
628 _iavf_rx_queue_release_mbufs_vec(rxq);
631 static void __attribute__((cold))
632 iavf_tx_queue_release_mbufs_sse(struct iavf_tx_queue *txq)
634 _iavf_tx_queue_release_mbufs_vec(txq);
637 static const struct iavf_rxq_ops sse_vec_rxq_ops = {
638 .release_mbufs = iavf_rx_queue_release_mbufs_sse,
641 static const struct iavf_txq_ops sse_vec_txq_ops = {
642 .release_mbufs = iavf_tx_queue_release_mbufs_sse,
645 int __attribute__((cold))
646 iavf_txq_vec_setup(struct iavf_tx_queue *txq)
648 txq->ops = &sse_vec_txq_ops;
652 int __attribute__((cold))
653 iavf_rxq_vec_setup(struct iavf_rx_queue *rxq)
655 rxq->ops = &sse_vec_rxq_ops;
656 return iavf_rxq_vec_setup_default(rxq);