4 * Copyright(c) 2010-2015 Intel Corporation. All rights reserved.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
11 * * Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * * Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in
15 * the documentation and/or other materials provided with the
17 * * Neither the name of Intel Corporation nor the names of its
18 * contributors may be used to endorse or promote products derived
19 * from this software without specific prior written permission.
21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
35 #include <rte_ethdev.h>
36 #include <rte_malloc.h>
38 #include "base/i40e_prototype.h"
39 #include "base/i40e_type.h"
40 #include "i40e_ethdev.h"
41 #include "i40e_rxtx.h"
43 #include <tmmintrin.h>
45 #ifndef __INTEL_COMPILER
46 #pragma GCC diagnostic ignored "-Wcast-qual"
50 i40e_rxq_rearm(struct i40e_rx_queue *rxq)
54 volatile union i40e_rx_desc *rxdp;
55 struct i40e_rx_entry *rxep = &rxq->sw_ring[rxq->rxrearm_start];
56 struct rte_mbuf *mb0, *mb1;
57 __m128i hdr_room = _mm_set_epi64x(RTE_PKTMBUF_HEADROOM,
58 RTE_PKTMBUF_HEADROOM);
59 __m128i dma_addr0, dma_addr1;
61 rxdp = rxq->rx_ring + rxq->rxrearm_start;
63 /* Pull 'n' more MBUFs into the software ring */
64 if (rte_mempool_get_bulk(rxq->mp,
66 RTE_I40E_RXQ_REARM_THRESH) < 0) {
67 if (rxq->rxrearm_nb + RTE_I40E_RXQ_REARM_THRESH >=
69 dma_addr0 = _mm_setzero_si128();
70 for (i = 0; i < RTE_I40E_DESCS_PER_LOOP; i++) {
71 rxep[i].mbuf = &rxq->fake_mbuf;
72 _mm_store_si128((__m128i *)&rxdp[i].read,
76 rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
77 RTE_I40E_RXQ_REARM_THRESH;
81 /* Initialize the mbufs in vector, process 2 mbufs in one loop */
82 for (i = 0; i < RTE_I40E_RXQ_REARM_THRESH; i += 2, rxep += 2) {
83 __m128i vaddr0, vaddr1;
89 /* Flush mbuf with pkt template.
90 * Data to be rearmed is 6 bytes long.
91 * Though, RX will overwrite ol_flags that are coming next
92 * anyway. So overwrite whole 8 bytes with one load:
93 * 6 bytes of rearm_data plus first 2 bytes of ol_flags.
95 p0 = (uintptr_t)&mb0->rearm_data;
96 *(uint64_t *)p0 = rxq->mbuf_initializer;
97 p1 = (uintptr_t)&mb1->rearm_data;
98 *(uint64_t *)p1 = rxq->mbuf_initializer;
100 /* load buf_addr(lo 64bit) and buf_physaddr(hi 64bit) */
101 vaddr0 = _mm_loadu_si128((__m128i *)&mb0->buf_addr);
102 vaddr1 = _mm_loadu_si128((__m128i *)&mb1->buf_addr);
104 /* convert pa to dma_addr hdr/data */
105 dma_addr0 = _mm_unpackhi_epi64(vaddr0, vaddr0);
106 dma_addr1 = _mm_unpackhi_epi64(vaddr1, vaddr1);
108 /* add headroom to pa values */
109 dma_addr0 = _mm_add_epi64(dma_addr0, hdr_room);
110 dma_addr1 = _mm_add_epi64(dma_addr1, hdr_room);
112 /* flush desc with pa dma_addr */
113 _mm_store_si128((__m128i *)&rxdp++->read, dma_addr0);
114 _mm_store_si128((__m128i *)&rxdp++->read, dma_addr1);
117 rxq->rxrearm_start += RTE_I40E_RXQ_REARM_THRESH;
118 if (rxq->rxrearm_start >= rxq->nb_rx_desc)
119 rxq->rxrearm_start = 0;
121 rxq->rxrearm_nb -= RTE_I40E_RXQ_REARM_THRESH;
123 rx_id = (uint16_t)((rxq->rxrearm_start == 0) ?
124 (rxq->nb_rx_desc - 1) : (rxq->rxrearm_start - 1));
126 /* Update the tail pointer on the NIC */
127 I40E_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
130 /* Handling the offload flags (olflags) field takes computation
131 * time when receiving packets. Therefore we provide a flag to disable
132 * the processing of the olflags field when they are not needed. This
133 * gives improved performance, at the cost of losing the offload info
134 * in the received packet
136 #ifdef RTE_LIBRTE_I40E_RX_OLFLAGS_ENABLE
139 desc_to_olflags_v(__m128i descs[4], struct rte_mbuf **rx_pkts)
141 __m128i vlan0, vlan1, rss;
147 /* mask everything except rss and vlan flags
148 *bit2 is for vlan tag, bits 13:12 for rss
150 const __m128i rss_vlan_msk = _mm_set_epi16(
151 0x0000, 0x0000, 0x0000, 0x0000,
152 0x3004, 0x3004, 0x3004, 0x3004);
154 /* map rss and vlan type to rss hash and vlan flag */
155 const __m128i vlan_flags = _mm_set_epi8(0, 0, 0, 0,
157 0, 0, 0, PKT_RX_VLAN_PKT,
160 const __m128i rss_flags = _mm_set_epi8(0, 0, 0, 0,
163 PKT_RX_FDIR, 0, PKT_RX_RSS_HASH, 0);
165 vlan0 = _mm_unpackhi_epi16(descs[0], descs[1]);
166 vlan1 = _mm_unpackhi_epi16(descs[2], descs[3]);
167 vlan0 = _mm_unpacklo_epi32(vlan0, vlan1);
169 vlan1 = _mm_and_si128(vlan0, rss_vlan_msk);
170 vlan0 = _mm_shuffle_epi8(vlan_flags, vlan1);
172 rss = _mm_srli_epi16(vlan1, 12);
173 rss = _mm_shuffle_epi8(rss_flags, rss);
175 vlan0 = _mm_or_si128(vlan0, rss);
176 vol.dword = _mm_cvtsi128_si64(vlan0);
178 rx_pkts[0]->ol_flags = vol.e[0];
179 rx_pkts[1]->ol_flags = vol.e[1];
180 rx_pkts[2]->ol_flags = vol.e[2];
181 rx_pkts[3]->ol_flags = vol.e[3];
184 #define desc_to_olflags_v(desc, rx_pkts) do {} while (0)
187 #define PKTLEN_SHIFT (6)
188 #define PKTLEN_MASK (0x3FFF)
189 /* Handling the pkt len field is not aligned with 1byte, so shift is
190 * needed to let it align
193 desc_pktlen_align(__m128i descs[4])
195 __m128i pktlen0, pktlen1, zero;
201 /* mask everything except pktlen field*/
202 const __m128i pktlen_msk = _mm_set_epi32(PKTLEN_MASK, PKTLEN_MASK,
203 PKTLEN_MASK, PKTLEN_MASK);
205 pktlen0 = _mm_unpackhi_epi32(descs[0], descs[2]);
206 pktlen1 = _mm_unpackhi_epi32(descs[1], descs[3]);
207 pktlen0 = _mm_unpackhi_epi32(pktlen0, pktlen1);
209 zero = _mm_xor_si128(pktlen0, pktlen0);
211 pktlen0 = _mm_srli_epi32(pktlen0, PKTLEN_SHIFT);
212 pktlen0 = _mm_and_si128(pktlen0, pktlen_msk);
214 pktlen0 = _mm_packs_epi32(pktlen0, zero);
215 vol.dword = _mm_cvtsi128_si64(pktlen0);
216 /* let the descriptor byte 15-14 store the pkt len */
217 *((uint16_t *)&descs[0]+7) = vol.e[0];
218 *((uint16_t *)&descs[1]+7) = vol.e[1];
219 *((uint16_t *)&descs[2]+7) = vol.e[2];
220 *((uint16_t *)&descs[3]+7) = vol.e[3];
225 * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
226 * - nb_pkts > RTE_I40E_VPMD_RX_BURST, only scan RTE_I40E_VPMD_RX_BURST
229 static inline uint16_t
230 _recv_raw_pkts_vec(struct i40e_rx_queue *rxq, struct rte_mbuf **rx_pkts,
231 uint16_t nb_pkts, uint8_t *split_packet)
233 volatile union i40e_rx_desc *rxdp;
234 struct i40e_rx_entry *sw_ring;
235 uint16_t nb_pkts_recd;
240 __m128i crc_adjust = _mm_set_epi16(
241 0, 0, 0, /* ignore non-length fields */
242 -rxq->crc_len, /* sub crc on data_len */
243 0, /* ignore high-16bits of pkt_len */
244 -rxq->crc_len, /* sub crc on pkt_len */
245 0, 0 /* ignore pkt_type field */
247 __m128i dd_check, eop_check;
249 /* nb_pkts shall be less equal than RTE_I40E_MAX_RX_BURST */
250 nb_pkts = RTE_MIN(nb_pkts, RTE_I40E_MAX_RX_BURST);
252 /* nb_pkts has to be floor-aligned to RTE_I40E_DESCS_PER_LOOP */
253 nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_I40E_DESCS_PER_LOOP);
255 /* Just the act of getting into the function from the application is
256 * going to cost about 7 cycles
258 rxdp = rxq->rx_ring + rxq->rx_tail;
260 _mm_prefetch((const void *)rxdp, _MM_HINT_T0);
262 /* See if we need to rearm the RX queue - gives the prefetch a bit
265 if (rxq->rxrearm_nb > RTE_I40E_RXQ_REARM_THRESH)
268 /* Before we start moving massive data around, check to see if
269 * there is actually a packet available
271 if (!(rxdp->wb.qword1.status_error_len &
272 rte_cpu_to_le_32(1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
275 /* 4 packets DD mask */
276 dd_check = _mm_set_epi64x(0x0000000100000001LL, 0x0000000100000001LL);
278 /* 4 packets EOP mask */
279 eop_check = _mm_set_epi64x(0x0000000200000002LL, 0x0000000200000002LL);
281 /* mask to shuffle from desc. to mbuf */
282 shuf_msk = _mm_set_epi8(
283 7, 6, 5, 4, /* octet 4~7, 32bits rss */
284 3, 2, /* octet 2~3, low 16 bits vlan_macip */
285 15, 14, /* octet 15~14, 16 bits data_len */
286 0xFF, 0xFF, /* skip high 16 bits pkt_len, zero out */
287 15, 14, /* octet 15~14, low 16 bits pkt_len */
288 0xFF, 0xFF, /* pkt_type set as unknown */
289 0xFF, 0xFF /*pkt_type set as unknown */
292 /* Cache is empty -> need to scan the buffer rings, but first move
293 * the next 'n' mbufs into the cache
295 sw_ring = &rxq->sw_ring[rxq->rx_tail];
297 /* A. load 4 packet in one loop
298 * [A*. mask out 4 unused dirty field in desc]
299 * B. copy 4 mbuf point from swring to rx_pkts
300 * C. calc the number of DD bits among the 4 packets
301 * [C*. extract the end-of-packet bit, if requested]
302 * D. fill info. from desc to mbuf
305 for (pos = 0, nb_pkts_recd = 0; pos < RTE_I40E_VPMD_RX_BURST;
306 pos += RTE_I40E_DESCS_PER_LOOP,
307 rxdp += RTE_I40E_DESCS_PER_LOOP) {
308 __m128i descs[RTE_I40E_DESCS_PER_LOOP];
309 __m128i pkt_mb1, pkt_mb2, pkt_mb3, pkt_mb4;
310 __m128i zero, staterr, sterr_tmp1, sterr_tmp2;
311 __m128i mbp1, mbp2; /* two mbuf pointer in one XMM reg. */
313 /* B.1 load 1 mbuf point */
314 mbp1 = _mm_loadu_si128((__m128i *)&sw_ring[pos]);
315 /* Read desc statuses backwards to avoid race condition */
316 /* A.1 load 4 pkts desc */
317 descs[3] = _mm_loadu_si128((__m128i *)(rxdp + 3));
319 /* B.2 copy 2 mbuf point into rx_pkts */
320 _mm_storeu_si128((__m128i *)&rx_pkts[pos], mbp1);
322 /* B.1 load 1 mbuf point */
323 mbp2 = _mm_loadu_si128((__m128i *)&sw_ring[pos+2]);
325 descs[2] = _mm_loadu_si128((__m128i *)(rxdp + 2));
326 /* B.1 load 2 mbuf point */
327 descs[1] = _mm_loadu_si128((__m128i *)(rxdp + 1));
328 descs[0] = _mm_loadu_si128((__m128i *)(rxdp));
330 /* B.2 copy 2 mbuf point into rx_pkts */
331 _mm_storeu_si128((__m128i *)&rx_pkts[pos+2], mbp2);
334 rte_prefetch0(&rx_pkts[pos]->cacheline1);
335 rte_prefetch0(&rx_pkts[pos + 1]->cacheline1);
336 rte_prefetch0(&rx_pkts[pos + 2]->cacheline1);
337 rte_prefetch0(&rx_pkts[pos + 3]->cacheline1);
340 /*shift the pktlen field*/
341 desc_pktlen_align(descs);
343 /* avoid compiler reorder optimization */
344 rte_compiler_barrier();
346 /* D.1 pkt 3,4 convert format from desc to pktmbuf */
347 pkt_mb4 = _mm_shuffle_epi8(descs[3], shuf_msk);
348 pkt_mb3 = _mm_shuffle_epi8(descs[2], shuf_msk);
350 /* C.1 4=>2 filter staterr info only */
351 sterr_tmp2 = _mm_unpackhi_epi32(descs[3], descs[2]);
352 /* C.1 4=>2 filter staterr info only */
353 sterr_tmp1 = _mm_unpackhi_epi32(descs[1], descs[0]);
355 desc_to_olflags_v(descs, &rx_pkts[pos]);
357 /* D.2 pkt 3,4 set in_port/nb_seg and remove crc */
358 pkt_mb4 = _mm_add_epi16(pkt_mb4, crc_adjust);
359 pkt_mb3 = _mm_add_epi16(pkt_mb3, crc_adjust);
361 /* D.1 pkt 1,2 convert format from desc to pktmbuf */
362 pkt_mb2 = _mm_shuffle_epi8(descs[1], shuf_msk);
363 pkt_mb1 = _mm_shuffle_epi8(descs[0], shuf_msk);
365 /* C.2 get 4 pkts staterr value */
366 zero = _mm_xor_si128(dd_check, dd_check);
367 staterr = _mm_unpacklo_epi32(sterr_tmp1, sterr_tmp2);
369 /* D.3 copy final 3,4 data to rx_pkts */
370 _mm_storeu_si128((void *)&rx_pkts[pos+3]->rx_descriptor_fields1,
372 _mm_storeu_si128((void *)&rx_pkts[pos+2]->rx_descriptor_fields1,
375 /* D.2 pkt 1,2 set in_port/nb_seg and remove crc */
376 pkt_mb2 = _mm_add_epi16(pkt_mb2, crc_adjust);
377 pkt_mb1 = _mm_add_epi16(pkt_mb1, crc_adjust);
379 /* C* extract and record EOP bit */
381 __m128i eop_shuf_mask = _mm_set_epi8(
382 0xFF, 0xFF, 0xFF, 0xFF,
383 0xFF, 0xFF, 0xFF, 0xFF,
384 0xFF, 0xFF, 0xFF, 0xFF,
385 0x04, 0x0C, 0x00, 0x08
388 /* and with mask to extract bits, flipping 1-0 */
389 __m128i eop_bits = _mm_andnot_si128(staterr, eop_check);
390 /* the staterr values are not in order, as the count
391 * count of dd bits doesn't care. However, for end of
392 * packet tracking, we do care, so shuffle. This also
393 * compresses the 32-bit values to 8-bit
395 eop_bits = _mm_shuffle_epi8(eop_bits, eop_shuf_mask);
396 /* store the resulting 32-bit value */
397 *(int *)split_packet = _mm_cvtsi128_si32(eop_bits);
398 split_packet += RTE_I40E_DESCS_PER_LOOP;
400 /* zero-out next pointers */
401 rx_pkts[pos]->next = NULL;
402 rx_pkts[pos + 1]->next = NULL;
403 rx_pkts[pos + 2]->next = NULL;
404 rx_pkts[pos + 3]->next = NULL;
407 /* C.3 calc available number of desc */
408 staterr = _mm_and_si128(staterr, dd_check);
409 staterr = _mm_packs_epi32(staterr, zero);
411 /* D.3 copy final 1,2 data to rx_pkts */
412 _mm_storeu_si128((void *)&rx_pkts[pos+1]->rx_descriptor_fields1,
414 _mm_storeu_si128((void *)&rx_pkts[pos]->rx_descriptor_fields1,
416 /* C.4 calc avaialbe number of desc */
417 var = __builtin_popcountll(_mm_cvtsi128_si64(staterr));
419 if (likely(var != RTE_I40E_DESCS_PER_LOOP))
423 /* Update our internal tail pointer */
424 rxq->rx_tail = (uint16_t)(rxq->rx_tail + nb_pkts_recd);
425 rxq->rx_tail = (uint16_t)(rxq->rx_tail & (rxq->nb_rx_desc - 1));
426 rxq->rxrearm_nb = (uint16_t)(rxq->rxrearm_nb + nb_pkts_recd);
433 * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
434 * - nb_pkts > RTE_I40E_VPMD_RX_BURST, only scan RTE_I40E_VPMD_RX_BURST
438 i40e_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
441 return _recv_raw_pkts_vec(rx_queue, rx_pkts, nb_pkts, NULL);
444 static inline uint16_t
445 reassemble_packets(struct i40e_rx_queue *rxq, struct rte_mbuf **rx_bufs,
446 uint16_t nb_bufs, uint8_t *split_flags)
448 struct rte_mbuf *pkts[RTE_I40E_VPMD_RX_BURST]; /*finished pkts*/
449 struct rte_mbuf *start = rxq->pkt_first_seg;
450 struct rte_mbuf *end = rxq->pkt_last_seg;
451 unsigned pkt_idx, buf_idx;
453 for (buf_idx = 0, pkt_idx = 0; buf_idx < nb_bufs; buf_idx++) {
455 /* processing a split packet */
456 end->next = rx_bufs[buf_idx];
457 rx_bufs[buf_idx]->data_len += rxq->crc_len;
460 start->pkt_len += rx_bufs[buf_idx]->data_len;
463 if (!split_flags[buf_idx]) {
464 /* it's the last packet of the set */
465 start->hash = end->hash;
466 start->ol_flags = end->ol_flags;
467 /* we need to strip crc for the whole packet */
468 start->pkt_len -= rxq->crc_len;
469 if (end->data_len > rxq->crc_len) {
470 end->data_len -= rxq->crc_len;
472 /* free up last mbuf */
473 struct rte_mbuf *secondlast = start;
475 while (secondlast->next != end)
476 secondlast = secondlast->next;
477 secondlast->data_len -= (rxq->crc_len -
479 secondlast->next = NULL;
480 rte_pktmbuf_free_seg(end);
483 pkts[pkt_idx++] = start;
487 /* not processing a split packet */
488 if (!split_flags[buf_idx]) {
489 /* not a split packet, save and skip */
490 pkts[pkt_idx++] = rx_bufs[buf_idx];
493 end = start = rx_bufs[buf_idx];
494 rx_bufs[buf_idx]->data_len += rxq->crc_len;
495 rx_bufs[buf_idx]->pkt_len += rxq->crc_len;
499 /* save the partial packet for next time */
500 rxq->pkt_first_seg = start;
501 rxq->pkt_last_seg = end;
502 memcpy(rx_bufs, pkts, pkt_idx * (sizeof(*pkts)));
506 /* vPMD receive routine that reassembles scattered packets
508 * - nb_pkts < RTE_I40E_DESCS_PER_LOOP, just return no packet
509 * - nb_pkts > RTE_I40E_VPMD_RX_BURST, only scan RTE_I40E_VPMD_RX_BURST
513 i40e_recv_scattered_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
517 struct i40e_rx_queue *rxq = rx_queue;
518 uint8_t split_flags[RTE_I40E_VPMD_RX_BURST] = {0};
520 /* get some new buffers */
521 uint16_t nb_bufs = _recv_raw_pkts_vec(rxq, rx_pkts, nb_pkts,
526 /* happy day case, full burst + no packets to be joined */
527 const uint64_t *split_fl64 = (uint64_t *)split_flags;
529 if (rxq->pkt_first_seg == NULL &&
530 split_fl64[0] == 0 && split_fl64[1] == 0 &&
531 split_fl64[2] == 0 && split_fl64[3] == 0)
534 /* reassemble any packets that need reassembly*/
537 if (rxq->pkt_first_seg == NULL) {
538 /* find the first split flag, and only reassemble then*/
539 while (i < nb_bufs && !split_flags[i])
544 return i + reassemble_packets(rxq, &rx_pkts[i], nb_bufs - i,
549 vtx1(volatile struct i40e_tx_desc *txdp,
550 struct rte_mbuf *pkt, uint64_t flags)
552 uint64_t high_qw = (I40E_TX_DESC_DTYPE_DATA |
553 ((uint64_t)flags << I40E_TXD_QW1_CMD_SHIFT) |
554 ((uint64_t)pkt->data_len << I40E_TXD_QW1_TX_BUF_SZ_SHIFT));
556 __m128i descriptor = _mm_set_epi64x(high_qw,
557 pkt->buf_physaddr + pkt->data_off);
558 _mm_store_si128((__m128i *)txdp, descriptor);
562 vtx(volatile struct i40e_tx_desc *txdp,
563 struct rte_mbuf **pkt, uint16_t nb_pkts, uint64_t flags)
567 for (i = 0; i < nb_pkts; ++i, ++txdp, ++pkt)
568 vtx1(txdp, *pkt, flags);
571 static inline int __attribute__((always_inline))
572 i40e_tx_free_bufs(struct i40e_tx_queue *txq)
574 struct i40e_tx_entry *txep;
578 struct rte_mbuf *m, *free[RTE_I40E_TX_MAX_FREE_BUF_SZ];
580 /* check DD bits on threshold descriptor */
581 if ((txq->tx_ring[txq->tx_next_dd].cmd_type_offset_bsz &
582 rte_cpu_to_le_64(I40E_TXD_QW1_DTYPE_MASK)) !=
583 rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DESC_DONE))
586 n = txq->tx_rs_thresh;
588 /* first buffer to free from S/W ring is at index
589 * tx_next_dd - (tx_rs_thresh-1)
591 txep = &txq->sw_ring[txq->tx_next_dd - (n - 1)];
592 m = __rte_pktmbuf_prefree_seg(txep[0].mbuf);
593 if (likely(m != NULL)) {
596 for (i = 1; i < n; i++) {
597 m = __rte_pktmbuf_prefree_seg(txep[i].mbuf);
598 if (likely(m != NULL)) {
599 if (likely(m->pool == free[0]->pool)) {
602 rte_mempool_put_bulk(free[0]->pool,
610 rte_mempool_put_bulk(free[0]->pool, (void **)free, nb_free);
612 for (i = 1; i < n; i++) {
613 m = __rte_pktmbuf_prefree_seg(txep[i].mbuf);
615 rte_mempool_put(m->pool, m);
619 /* buffers were freed, update counters */
620 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + txq->tx_rs_thresh);
621 txq->tx_next_dd = (uint16_t)(txq->tx_next_dd + txq->tx_rs_thresh);
622 if (txq->tx_next_dd >= txq->nb_tx_desc)
623 txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
625 return txq->tx_rs_thresh;
628 static inline void __attribute__((always_inline))
629 tx_backlog_entry(struct i40e_tx_entry *txep,
630 struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
634 for (i = 0; i < (int)nb_pkts; ++i)
635 txep[i].mbuf = tx_pkts[i];
639 i40e_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
642 struct i40e_tx_queue *txq = (struct i40e_tx_queue *)tx_queue;
643 volatile struct i40e_tx_desc *txdp;
644 struct i40e_tx_entry *txep;
645 uint16_t n, nb_commit, tx_id;
646 uint64_t flags = I40E_TD_CMD;
647 uint64_t rs = I40E_TX_DESC_CMD_RS | I40E_TD_CMD;
650 /* cross rx_thresh boundary is not allowed */
651 nb_pkts = RTE_MIN(nb_pkts, txq->tx_rs_thresh);
653 if (txq->nb_tx_free < txq->tx_free_thresh)
654 i40e_tx_free_bufs(txq);
656 nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
657 if (unlikely(nb_pkts == 0))
660 tx_id = txq->tx_tail;
661 txdp = &txq->tx_ring[tx_id];
662 txep = &txq->sw_ring[tx_id];
664 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
666 n = (uint16_t)(txq->nb_tx_desc - tx_id);
667 if (nb_commit >= n) {
668 tx_backlog_entry(txep, tx_pkts, n);
670 for (i = 0; i < n - 1; ++i, ++tx_pkts, ++txdp)
671 vtx1(txdp, *tx_pkts, flags);
673 vtx1(txdp, *tx_pkts++, rs);
675 nb_commit = (uint16_t)(nb_commit - n);
678 txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
680 /* avoid reach the end of ring */
681 txdp = &txq->tx_ring[tx_id];
682 txep = &txq->sw_ring[tx_id];
685 tx_backlog_entry(txep, tx_pkts, nb_commit);
687 vtx(txdp, tx_pkts, nb_commit, flags);
689 tx_id = (uint16_t)(tx_id + nb_commit);
690 if (tx_id > txq->tx_next_rs) {
691 txq->tx_ring[txq->tx_next_rs].cmd_type_offset_bsz |=
692 rte_cpu_to_le_64(((uint64_t)I40E_TX_DESC_CMD_RS) <<
693 I40E_TXD_QW1_CMD_SHIFT);
695 (uint16_t)(txq->tx_next_rs + txq->tx_rs_thresh);
698 txq->tx_tail = tx_id;
700 I40E_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
705 void __attribute__((cold))
706 i40e_rx_queue_release_mbufs_vec(struct i40e_rx_queue *rxq)
708 const unsigned mask = rxq->nb_rx_desc - 1;
711 if (rxq->sw_ring == NULL || rxq->rxrearm_nb >= rxq->nb_rx_desc)
714 /* free all mbufs that are valid in the ring */
715 for (i = rxq->rx_tail; i != rxq->rxrearm_start; i = (i + 1) & mask)
716 rte_pktmbuf_free_seg(rxq->sw_ring[i].mbuf);
717 rxq->rxrearm_nb = rxq->nb_rx_desc;
719 /* set all entries to NULL */
720 memset(rxq->sw_ring, 0, sizeof(rxq->sw_ring[0]) * rxq->nb_rx_desc);
723 int __attribute__((cold))
724 i40e_rxq_vec_setup(struct i40e_rx_queue *rxq)
727 struct rte_mbuf mb_def = { .buf_addr = 0 }; /* zeroed mbuf */
730 mb_def.data_off = RTE_PKTMBUF_HEADROOM;
731 mb_def.port = rxq->port_id;
732 rte_mbuf_refcnt_set(&mb_def, 1);
734 /* prevent compiler reordering: rearm_data covers previous fields */
735 rte_compiler_barrier();
736 p = (uintptr_t)&mb_def.rearm_data;
737 rxq->mbuf_initializer = *(uint64_t *)p;
741 int __attribute__((cold))
742 i40e_txq_vec_setup(struct i40e_tx_queue __rte_unused *txq)
747 int __attribute__((cold))
748 i40e_rx_vec_dev_conf_condition_check(struct rte_eth_dev *dev)
750 #ifndef RTE_LIBRTE_IEEE1588
751 struct rte_eth_rxmode *rxmode = &dev->data->dev_conf.rxmode;
752 struct rte_fdir_conf *fconf = &dev->data->dev_conf.fdir_conf;
754 #ifndef RTE_LIBRTE_I40E_RX_OLFLAGS_ENABLE
755 /* whithout rx ol_flags, no VP flag report */
756 if (rxmode->hw_vlan_strip != 0 ||
757 rxmode->hw_vlan_extend != 0)
761 /* no fdir support */
762 if (fconf->mode != RTE_FDIR_MODE_NONE)
765 /* - no csum error report support
766 * - no header split support
768 if (rxmode->hw_ip_checksum == 1 ||
769 rxmode->header_split == 1)