4 * Copyright(c) 2010-2014 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 "ixgbe_ethdev.h"
39 #include "ixgbe_rxtx.h"
41 #include <tmmintrin.h>
43 #ifndef __INTEL_COMPILER
44 #pragma GCC diagnostic ignored "-Wcast-qual"
48 ixgbe_rxq_rearm(struct igb_rx_queue *rxq)
52 volatile union ixgbe_adv_rx_desc *rxdp;
53 struct igb_rx_entry *rxep = &rxq->sw_ring[rxq->rxrearm_start];
54 struct rte_mbuf *mb0, *mb1;
55 __m128i hdr_room = _mm_set_epi64x(RTE_PKTMBUF_HEADROOM,
56 RTE_PKTMBUF_HEADROOM);
57 __m128i dma_addr0, dma_addr1;
59 rxdp = rxq->rx_ring + rxq->rxrearm_start;
61 /* Pull 'n' more MBUFs into the software ring */
62 if (rte_mempool_get_bulk(rxq->mb_pool,
64 RTE_IXGBE_RXQ_REARM_THRESH) < 0) {
65 if (rxq->rxrearm_nb + RTE_IXGBE_RXQ_REARM_THRESH >=
67 dma_addr0 = _mm_setzero_si128();
68 for (i = 0; i < RTE_IXGBE_DESCS_PER_LOOP; i++) {
69 rxep[i].mbuf = &rxq->fake_mbuf;
70 _mm_store_si128((__m128i *)&rxdp[i].read,
74 rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
75 RTE_IXGBE_RXQ_REARM_THRESH;
79 /* Initialize the mbufs in vector, process 2 mbufs in one loop */
80 for (i = 0; i < RTE_IXGBE_RXQ_REARM_THRESH; i += 2, rxep += 2) {
81 __m128i vaddr0, vaddr1;
88 * Flush mbuf with pkt template.
89 * Data to be rearmed is 6 bytes long.
90 * Though, RX will overwrite ol_flags that are coming next
91 * anyway. So overwrite whole 8 bytes with one load:
92 * 6 bytes of rearm_data plus first 2 bytes of ol_flags.
94 p0 = (uintptr_t)&mb0->rearm_data;
95 *(uint64_t *)p0 = rxq->mbuf_initializer;
96 p1 = (uintptr_t)&mb1->rearm_data;
97 *(uint64_t *)p1 = rxq->mbuf_initializer;
99 /* load buf_addr(lo 64bit) and buf_physaddr(hi 64bit) */
100 vaddr0 = _mm_loadu_si128((__m128i *)&(mb0->buf_addr));
101 vaddr1 = _mm_loadu_si128((__m128i *)&(mb1->buf_addr));
103 /* convert pa to dma_addr hdr/data */
104 dma_addr0 = _mm_unpackhi_epi64(vaddr0, vaddr0);
105 dma_addr1 = _mm_unpackhi_epi64(vaddr1, vaddr1);
107 /* add headroom to pa values */
108 dma_addr0 = _mm_add_epi64(dma_addr0, hdr_room);
109 dma_addr1 = _mm_add_epi64(dma_addr1, hdr_room);
111 /* flush desc with pa dma_addr */
112 _mm_store_si128((__m128i *)&rxdp++->read, dma_addr0);
113 _mm_store_si128((__m128i *)&rxdp++->read, dma_addr1);
116 rxq->rxrearm_start += RTE_IXGBE_RXQ_REARM_THRESH;
117 if (rxq->rxrearm_start >= rxq->nb_rx_desc)
118 rxq->rxrearm_start = 0;
120 rxq->rxrearm_nb -= RTE_IXGBE_RXQ_REARM_THRESH;
122 rx_id = (uint16_t) ((rxq->rxrearm_start == 0) ?
123 (rxq->nb_rx_desc - 1) : (rxq->rxrearm_start - 1));
125 /* Update the tail pointer on the NIC */
126 IXGBE_PCI_REG_WRITE(rxq->rdt_reg_addr, rx_id);
129 /* Handling the offload flags (olflags) field takes computation
130 * time when receiving packets. Therefore we provide a flag to disable
131 * the processing of the olflags field when they are not needed. This
132 * gives improved performance, at the cost of losing the offload info
133 * in the received packet
135 #ifdef RTE_IXGBE_RX_OLFLAGS_ENABLE
137 #define OLFLAGS_MASK ((uint16_t)(PKT_RX_VLAN_PKT | PKT_RX_IPV4_HDR |\
138 PKT_RX_IPV4_HDR_EXT | PKT_RX_IPV6_HDR |\
139 PKT_RX_IPV6_HDR_EXT))
140 #define OLFLAGS_MASK_V (((uint64_t)OLFLAGS_MASK << 48) | \
141 ((uint64_t)OLFLAGS_MASK << 32) | \
142 ((uint64_t)OLFLAGS_MASK << 16) | \
143 ((uint64_t)OLFLAGS_MASK))
144 #define PTYPE_SHIFT (1)
145 #define VTAG_SHIFT (3)
148 desc_to_olflags_v(__m128i descs[4], struct rte_mbuf **rx_pkts)
150 __m128i ptype0, ptype1, vtag0, vtag1;
156 ptype0 = _mm_unpacklo_epi16(descs[0], descs[1]);
157 ptype1 = _mm_unpacklo_epi16(descs[2], descs[3]);
158 vtag0 = _mm_unpackhi_epi16(descs[0], descs[1]);
159 vtag1 = _mm_unpackhi_epi16(descs[2], descs[3]);
161 ptype1 = _mm_unpacklo_epi32(ptype0, ptype1);
162 vtag1 = _mm_unpacklo_epi32(vtag0, vtag1);
164 ptype1 = _mm_slli_epi16(ptype1, PTYPE_SHIFT);
165 vtag1 = _mm_srli_epi16(vtag1, VTAG_SHIFT);
167 ptype1 = _mm_or_si128(ptype1, vtag1);
168 vol.dword = _mm_cvtsi128_si64(ptype1) & OLFLAGS_MASK_V;
170 rx_pkts[0]->ol_flags = vol.e[0];
171 rx_pkts[1]->ol_flags = vol.e[1];
172 rx_pkts[2]->ol_flags = vol.e[2];
173 rx_pkts[3]->ol_flags = vol.e[3];
176 #define desc_to_olflags_v(desc, rx_pkts) do {} while (0)
180 * vPMD receive routine, now only accept (nb_pkts == RTE_IXGBE_VPMD_RX_BURST)
184 * - nb_pkts < RTE_IXGBE_VPMD_RX_BURST, just return no packet
185 * - nb_pkts > RTE_IXGBE_VPMD_RX_BURST, only scan RTE_IXGBE_VPMD_RX_BURST
187 * - don't support ol_flags for rss and csum err
189 static inline uint16_t
190 _recv_raw_pkts_vec(struct igb_rx_queue *rxq, struct rte_mbuf **rx_pkts,
191 uint16_t nb_pkts, uint8_t *split_packet)
193 volatile union ixgbe_adv_rx_desc *rxdp;
194 struct igb_rx_entry *sw_ring;
195 uint16_t nb_pkts_recd;
199 __m128i crc_adjust = _mm_set_epi16(
200 0, 0, 0, 0, /* ignore non-length fields */
201 0, /* ignore high-16bits of pkt_len */
202 -rxq->crc_len, /* sub crc on pkt_len */
203 -rxq->crc_len, /* sub crc on data_len */
204 0 /* ignore pkt_type field */
206 __m128i dd_check, eop_check;
208 if (unlikely(nb_pkts < RTE_IXGBE_VPMD_RX_BURST))
211 /* Just the act of getting into the function from the application is
212 * going to cost about 7 cycles */
213 rxdp = rxq->rx_ring + rxq->rx_tail;
215 _mm_prefetch((const void *)rxdp, _MM_HINT_T0);
217 /* See if we need to rearm the RX queue - gives the prefetch a bit
219 if (rxq->rxrearm_nb > RTE_IXGBE_RXQ_REARM_THRESH)
220 ixgbe_rxq_rearm(rxq);
222 /* Before we start moving massive data around, check to see if
223 * there is actually a packet available */
224 if (!(rxdp->wb.upper.status_error &
225 rte_cpu_to_le_32(IXGBE_RXDADV_STAT_DD)))
228 /* 4 packets DD mask */
229 dd_check = _mm_set_epi64x(0x0000000100000001LL, 0x0000000100000001LL);
231 /* 4 packets EOP mask */
232 eop_check = _mm_set_epi64x(0x0000000200000002LL, 0x0000000200000002LL);
234 /* mask to shuffle from desc. to mbuf */
235 shuf_msk = _mm_set_epi8(
236 7, 6, 5, 4, /* octet 4~7, 32bits rss */
237 0xFF, 0xFF, /* skip high 16 bits vlan_macip, zero out */
238 15, 14, /* octet 14~15, low 16 bits vlan_macip */
239 0xFF, 0xFF, /* skip high 16 bits pkt_len, zero out */
240 13, 12, /* octet 12~13, low 16 bits pkt_len */
241 13, 12, /* octet 12~13, 16 bits data_len */
242 0xFF, 0xFF /* skip pkt_type field */
245 /* Cache is empty -> need to scan the buffer rings, but first move
246 * the next 'n' mbufs into the cache */
247 sw_ring = &rxq->sw_ring[rxq->rx_tail];
250 * A. load 4 packet in one loop
251 * B. copy 4 mbuf point from swring to rx_pkts
252 * C. calc the number of DD bits among the 4 packets
253 * [C*. extract the end-of-packet bit, if requested]
254 * D. fill info. from desc to mbuf
256 for (pos = 0, nb_pkts_recd = 0; pos < RTE_IXGBE_VPMD_RX_BURST;
257 pos += RTE_IXGBE_DESCS_PER_LOOP,
258 rxdp += RTE_IXGBE_DESCS_PER_LOOP) {
259 __m128i descs[RTE_IXGBE_DESCS_PER_LOOP];
260 __m128i pkt_mb1, pkt_mb2, pkt_mb3, pkt_mb4;
261 __m128i zero, staterr, sterr_tmp1, sterr_tmp2;
262 __m128i mbp1, mbp2; /* two mbuf pointer in one XMM reg. */
265 rte_prefetch0(&rx_pkts[pos]->cacheline1);
266 rte_prefetch0(&rx_pkts[pos + 1]->cacheline1);
267 rte_prefetch0(&rx_pkts[pos + 2]->cacheline1);
268 rte_prefetch0(&rx_pkts[pos + 3]->cacheline1);
271 /* B.1 load 1 mbuf point */
272 mbp1 = _mm_loadu_si128((__m128i *)&sw_ring[pos]);
274 /* Read desc statuses backwards to avoid race condition */
275 /* A.1 load 4 pkts desc */
276 descs[3] = _mm_loadu_si128((__m128i *)(rxdp + 3));
278 /* B.2 copy 2 mbuf point into rx_pkts */
279 _mm_storeu_si128((__m128i *)&rx_pkts[pos], mbp1);
281 /* B.1 load 1 mbuf point */
282 mbp2 = _mm_loadu_si128((__m128i *)&sw_ring[pos+2]);
284 descs[2] = _mm_loadu_si128((__m128i *)(rxdp + 2));
285 /* B.1 load 2 mbuf point */
286 descs[1] = _mm_loadu_si128((__m128i *)(rxdp + 1));
287 descs[0] = _mm_loadu_si128((__m128i *)(rxdp));
289 /* B.2 copy 2 mbuf point into rx_pkts */
290 _mm_storeu_si128((__m128i *)&rx_pkts[pos+2], mbp2);
292 /* avoid compiler reorder optimization */
293 rte_compiler_barrier();
295 /* D.1 pkt 3,4 convert format from desc to pktmbuf */
296 pkt_mb4 = _mm_shuffle_epi8(descs[3], shuf_msk);
297 pkt_mb3 = _mm_shuffle_epi8(descs[2], shuf_msk);
299 /* C.1 4=>2 filter staterr info only */
300 sterr_tmp2 = _mm_unpackhi_epi32(descs[3], descs[2]);
301 /* C.1 4=>2 filter staterr info only */
302 sterr_tmp1 = _mm_unpackhi_epi32(descs[1], descs[0]);
304 /* set ol_flags with packet type and vlan tag */
305 desc_to_olflags_v(descs, &rx_pkts[pos]);
307 /* D.2 pkt 3,4 set in_port/nb_seg and remove crc */
308 pkt_mb4 = _mm_add_epi16(pkt_mb4, crc_adjust);
309 pkt_mb3 = _mm_add_epi16(pkt_mb3, crc_adjust);
311 /* D.1 pkt 1,2 convert format from desc to pktmbuf */
312 pkt_mb2 = _mm_shuffle_epi8(descs[1], shuf_msk);
313 pkt_mb1 = _mm_shuffle_epi8(descs[0], shuf_msk);
315 /* C.2 get 4 pkts staterr value */
316 zero = _mm_xor_si128(dd_check, dd_check);
317 staterr = _mm_unpacklo_epi32(sterr_tmp1, sterr_tmp2);
319 /* D.3 copy final 3,4 data to rx_pkts */
320 _mm_storeu_si128((void *)&rx_pkts[pos+3]->rx_descriptor_fields1,
322 _mm_storeu_si128((void *)&rx_pkts[pos+2]->rx_descriptor_fields1,
325 /* D.2 pkt 1,2 set in_port/nb_seg and remove crc */
326 pkt_mb2 = _mm_add_epi16(pkt_mb2, crc_adjust);
327 pkt_mb1 = _mm_add_epi16(pkt_mb1, crc_adjust);
329 /* C* extract and record EOP bit */
331 __m128i eop_shuf_mask = _mm_set_epi8(
332 0xFF, 0xFF, 0xFF, 0xFF,
333 0xFF, 0xFF, 0xFF, 0xFF,
334 0xFF, 0xFF, 0xFF, 0xFF,
335 0x04, 0x0C, 0x00, 0x08
338 /* and with mask to extract bits, flipping 1-0 */
339 __m128i eop_bits = _mm_andnot_si128(staterr, eop_check);
340 /* the staterr values are not in order, as the count
341 * count of dd bits doesn't care. However, for end of
342 * packet tracking, we do care, so shuffle. This also
343 * compresses the 32-bit values to 8-bit */
344 eop_bits = _mm_shuffle_epi8(eop_bits, eop_shuf_mask);
345 /* store the resulting 32-bit value */
346 *(int *)split_packet = _mm_cvtsi128_si32(eop_bits);
347 split_packet += RTE_IXGBE_DESCS_PER_LOOP;
349 /* zero-out next pointers */
350 rx_pkts[pos]->next = NULL;
351 rx_pkts[pos + 1]->next = NULL;
352 rx_pkts[pos + 2]->next = NULL;
353 rx_pkts[pos + 3]->next = NULL;
356 /* C.3 calc available number of desc */
357 staterr = _mm_and_si128(staterr, dd_check);
358 staterr = _mm_packs_epi32(staterr, zero);
360 /* D.3 copy final 1,2 data to rx_pkts */
361 _mm_storeu_si128((void *)&rx_pkts[pos+1]->rx_descriptor_fields1,
363 _mm_storeu_si128((void *)&rx_pkts[pos]->rx_descriptor_fields1,
366 /* C.4 calc avaialbe number of desc */
367 var = __builtin_popcountll(_mm_cvtsi128_si64(staterr));
369 if (likely(var != RTE_IXGBE_DESCS_PER_LOOP))
373 /* Update our internal tail pointer */
374 rxq->rx_tail = (uint16_t)(rxq->rx_tail + nb_pkts_recd);
375 rxq->rx_tail = (uint16_t)(rxq->rx_tail & (rxq->nb_rx_desc - 1));
376 rxq->rxrearm_nb = (uint16_t)(rxq->rxrearm_nb + nb_pkts_recd);
382 * vPMD receive routine, now only accept (nb_pkts == RTE_IXGBE_VPMD_RX_BURST)
386 * - nb_pkts < RTE_IXGBE_VPMD_RX_BURST, just return no packet
387 * - nb_pkts > RTE_IXGBE_VPMD_RX_BURST, only scan RTE_IXGBE_VPMD_RX_BURST
389 * - don't support ol_flags for rss and csum err
392 ixgbe_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
395 return _recv_raw_pkts_vec(rx_queue, rx_pkts, nb_pkts, NULL);
398 static inline uint16_t
399 reassemble_packets(struct igb_rx_queue *rxq, struct rte_mbuf **rx_bufs,
400 uint16_t nb_bufs, uint8_t *split_flags)
402 struct rte_mbuf *pkts[RTE_IXGBE_VPMD_RX_BURST]; /*finished pkts*/
403 struct rte_mbuf *start = rxq->pkt_first_seg;
404 struct rte_mbuf *end = rxq->pkt_last_seg;
405 unsigned pkt_idx = 0, buf_idx = 0;
408 while (buf_idx < nb_bufs) {
410 /* processing a split packet */
411 end->next = rx_bufs[buf_idx];
412 rx_bufs[buf_idx]->data_len += rxq->crc_len;
415 start->pkt_len += rx_bufs[buf_idx]->data_len;
418 if (!split_flags[buf_idx]) {
419 /* it's the last packet of the set */
420 start->hash = end->hash;
421 start->ol_flags = end->ol_flags;
422 /* we need to strip crc for the whole packet */
423 start->pkt_len -= rxq->crc_len;
424 if (end->data_len > rxq->crc_len)
425 end->data_len -= rxq->crc_len;
427 /* free up last mbuf */
428 struct rte_mbuf *secondlast = start;
429 while (secondlast->next != end)
430 secondlast = secondlast->next;
431 secondlast->data_len -= (rxq->crc_len -
433 secondlast->next = NULL;
434 rte_pktmbuf_free_seg(end);
437 pkts[pkt_idx++] = start;
441 /* not processing a split packet */
442 if (!split_flags[buf_idx]) {
443 /* not a split packet, save and skip */
444 pkts[pkt_idx++] = rx_bufs[buf_idx];
447 end = start = rx_bufs[buf_idx];
448 rx_bufs[buf_idx]->data_len += rxq->crc_len;
449 rx_bufs[buf_idx]->pkt_len += rxq->crc_len;
454 /* save the partial packet for next time */
455 rxq->pkt_first_seg = start;
456 rxq->pkt_last_seg = end;
457 memcpy(rx_bufs, pkts, pkt_idx * (sizeof(*pkts)));
462 * vPMD receive routine that reassembles scattered packets
465 * - don't support ol_flags for rss and csum err
466 * - now only accept (nb_pkts == RTE_IXGBE_VPMD_RX_BURST)
469 ixgbe_recv_scattered_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
472 struct igb_rx_queue *rxq = rx_queue;
473 uint8_t split_flags[RTE_IXGBE_VPMD_RX_BURST] = {0};
475 /* get some new buffers */
476 uint16_t nb_bufs = _recv_raw_pkts_vec(rxq, rx_pkts, nb_pkts,
481 /* happy day case, full burst + no packets to be joined */
482 const uint32_t *split_fl32 = (uint32_t *)split_flags;
483 if (rxq->pkt_first_seg == NULL &&
484 split_fl32[0] == 0 && split_fl32[1] == 0 &&
485 split_fl32[2] == 0 && split_fl32[3] == 0)
488 /* reassemble any packets that need reassembly*/
490 if (rxq->pkt_first_seg == NULL) {
491 /* find the first split flag, and only reassemble then*/
492 while (!split_flags[i] && i < nb_bufs)
497 return i + reassemble_packets(rxq, &rx_pkts[i], nb_bufs - i,
502 vtx1(volatile union ixgbe_adv_tx_desc *txdp,
503 struct rte_mbuf *pkt, uint64_t flags)
505 __m128i descriptor = _mm_set_epi64x((uint64_t)pkt->pkt_len << 46 |
506 flags | pkt->data_len,
507 pkt->buf_physaddr + pkt->data_off);
508 _mm_store_si128((__m128i *)&txdp->read, descriptor);
512 vtx(volatile union ixgbe_adv_tx_desc *txdp,
513 struct rte_mbuf **pkt, uint16_t nb_pkts, uint64_t flags)
516 for (i = 0; i < nb_pkts; ++i, ++txdp, ++pkt)
517 vtx1(txdp, *pkt, flags);
520 static inline int __attribute__((always_inline))
521 ixgbe_tx_free_bufs(struct igb_tx_queue *txq)
523 struct igb_tx_entry_v *txep;
528 struct rte_mbuf *m, *free[RTE_IXGBE_TX_MAX_FREE_BUF_SZ];
530 /* check DD bit on threshold descriptor */
531 status = txq->tx_ring[txq->tx_next_dd].wb.status;
532 if (!(status & IXGBE_ADVTXD_STAT_DD))
535 n = txq->tx_rs_thresh;
538 * first buffer to free from S/W ring is at index
539 * tx_next_dd - (tx_rs_thresh-1)
541 txep = &((struct igb_tx_entry_v *)txq->sw_ring)[txq->tx_next_dd -
543 #ifdef RTE_MBUF_REFCNT
544 m = __rte_pktmbuf_prefree_seg(txep[0].mbuf);
548 if (likely(m != NULL)) {
551 for (i = 1; i < n; i++) {
552 #ifdef RTE_MBUF_REFCNT
553 m = __rte_pktmbuf_prefree_seg(txep[i].mbuf);
557 if (likely(m != NULL)) {
558 if (likely(m->pool == free[0]->pool))
561 rte_mempool_put_bulk(free[0]->pool,
562 (void *)free, nb_free);
568 rte_mempool_put_bulk(free[0]->pool, (void **)free, nb_free);
570 for (i = 1; i < n; i++) {
571 m = __rte_pktmbuf_prefree_seg(txep[i].mbuf);
573 rte_mempool_put(m->pool, m);
577 /* buffers were freed, update counters */
578 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + txq->tx_rs_thresh);
579 txq->tx_next_dd = (uint16_t)(txq->tx_next_dd + txq->tx_rs_thresh);
580 if (txq->tx_next_dd >= txq->nb_tx_desc)
581 txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
583 return txq->tx_rs_thresh;
586 static inline void __attribute__((always_inline))
587 tx_backlog_entry(struct igb_tx_entry_v *txep,
588 struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
591 for (i = 0; i < (int)nb_pkts; ++i)
592 txep[i].mbuf = tx_pkts[i];
596 ixgbe_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
599 struct igb_tx_queue *txq = (struct igb_tx_queue *)tx_queue;
600 volatile union ixgbe_adv_tx_desc *txdp;
601 struct igb_tx_entry_v *txep;
602 uint16_t n, nb_commit, tx_id;
603 uint64_t flags = DCMD_DTYP_FLAGS;
604 uint64_t rs = IXGBE_ADVTXD_DCMD_RS|DCMD_DTYP_FLAGS;
607 if (unlikely(nb_pkts > RTE_IXGBE_VPMD_TX_BURST))
608 nb_pkts = RTE_IXGBE_VPMD_TX_BURST;
610 if (txq->nb_tx_free < txq->tx_free_thresh)
611 ixgbe_tx_free_bufs(txq);
613 nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
614 if (unlikely(nb_pkts == 0))
617 tx_id = txq->tx_tail;
618 txdp = &txq->tx_ring[tx_id];
619 txep = &((struct igb_tx_entry_v *)txq->sw_ring)[tx_id];
621 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
623 n = (uint16_t)(txq->nb_tx_desc - tx_id);
624 if (nb_commit >= n) {
626 tx_backlog_entry(txep, tx_pkts, n);
628 for (i = 0; i < n - 1; ++i, ++tx_pkts, ++txdp)
629 vtx1(txdp, *tx_pkts, flags);
631 vtx1(txdp, *tx_pkts++, rs);
633 nb_commit = (uint16_t)(nb_commit - n);
636 txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
638 /* avoid reach the end of ring */
639 txdp = &(txq->tx_ring[tx_id]);
640 txep = &(((struct igb_tx_entry_v *)txq->sw_ring)[tx_id]);
643 tx_backlog_entry(txep, tx_pkts, nb_commit);
645 vtx(txdp, tx_pkts, nb_commit, flags);
647 tx_id = (uint16_t)(tx_id + nb_commit);
648 if (tx_id > txq->tx_next_rs) {
649 txq->tx_ring[txq->tx_next_rs].read.cmd_type_len |=
650 rte_cpu_to_le_32(IXGBE_ADVTXD_DCMD_RS);
651 txq->tx_next_rs = (uint16_t)(txq->tx_next_rs +
655 txq->tx_tail = tx_id;
657 IXGBE_PCI_REG_WRITE(txq->tdt_reg_addr, txq->tx_tail);
663 ixgbe_tx_queue_release_mbufs(struct igb_tx_queue *txq)
666 struct igb_tx_entry_v *txe;
667 uint16_t nb_free, max_desc;
669 if (txq->sw_ring != NULL) {
670 /* release the used mbufs in sw_ring */
671 nb_free = txq->nb_tx_free;
672 max_desc = (uint16_t)(txq->nb_tx_desc - 1);
673 for (i = txq->tx_next_dd - (txq->tx_rs_thresh - 1);
674 nb_free < max_desc && i != txq->tx_tail;
675 i = (i + 1) & max_desc) {
676 txe = (struct igb_tx_entry_v *)&txq->sw_ring[i];
677 if (txe->mbuf != NULL)
678 rte_pktmbuf_free_seg(txe->mbuf);
681 for (i = 0; i < txq->nb_tx_desc; i++) {
682 txe = (struct igb_tx_entry_v *)&txq->sw_ring[i];
689 ixgbe_tx_free_swring(struct igb_tx_queue *txq)
694 if (txq->sw_ring != NULL) {
695 rte_free((struct igb_rx_entry *)txq->sw_ring - 1);
701 ixgbe_reset_tx_queue(struct igb_tx_queue *txq)
703 static const union ixgbe_adv_tx_desc zeroed_desc = { .read = {
705 struct igb_tx_entry_v *txe = (struct igb_tx_entry_v *)txq->sw_ring;
708 /* Zero out HW ring memory */
709 for (i = 0; i < txq->nb_tx_desc; i++)
710 txq->tx_ring[i] = zeroed_desc;
712 /* Initialize SW ring entries */
713 for (i = 0; i < txq->nb_tx_desc; i++) {
714 volatile union ixgbe_adv_tx_desc *txd = &txq->tx_ring[i];
715 txd->wb.status = IXGBE_TXD_STAT_DD;
719 txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
720 txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
725 * Always allow 1 descriptor to be un-allocated to avoid
726 * a H/W race condition
728 txq->last_desc_cleaned = (uint16_t)(txq->nb_tx_desc - 1);
729 txq->nb_tx_free = (uint16_t)(txq->nb_tx_desc - 1);
731 memset((void *)&txq->ctx_cache, 0,
732 IXGBE_CTX_NUM * sizeof(struct ixgbe_advctx_info));
735 static struct ixgbe_txq_ops vec_txq_ops = {
736 .release_mbufs = ixgbe_tx_queue_release_mbufs,
737 .free_swring = ixgbe_tx_free_swring,
738 .reset = ixgbe_reset_tx_queue,
742 ixgbe_rxq_vec_setup(struct igb_rx_queue *rxq)
745 struct rte_mbuf mb_def = { .buf_addr = 0 }; /* zeroed mbuf */
748 mb_def.data_off = RTE_PKTMBUF_HEADROOM;
749 mb_def.port = rxq->port_id;
750 rte_mbuf_refcnt_set(&mb_def, 1);
752 /* prevent compiler reordering: rearm_data covers previous fields */
753 rte_compiler_barrier();
754 p = (uintptr_t)&mb_def.rearm_data;
755 rxq->mbuf_initializer = *(uint64_t *)p;
759 int ixgbe_txq_vec_setup(struct igb_tx_queue *txq)
761 if (txq->sw_ring == NULL)
764 /* leave the first one for overflow */
765 txq->sw_ring = (struct igb_tx_entry *)
766 ((struct igb_tx_entry_v *)txq->sw_ring + 1);
767 txq->ops = &vec_txq_ops;
772 int ixgbe_rx_vec_condition_check(struct rte_eth_dev *dev)
774 #ifndef RTE_LIBRTE_IEEE1588
775 struct rte_eth_rxmode *rxmode = &dev->data->dev_conf.rxmode;
776 struct rte_fdir_conf *fconf = &dev->data->dev_conf.fdir_conf;
778 #ifndef RTE_IXGBE_RX_OLFLAGS_ENABLE
779 /* whithout rx ol_flags, no VP flag report */
780 if (rxmode->hw_vlan_strip != 0 ||
781 rxmode->hw_vlan_extend != 0)
785 /* no fdir support */
786 if (fconf->mode != RTE_FDIR_MODE_NONE)
790 * - no csum error report support
791 * - no header split support
793 if (rxmode->hw_ip_checksum == 1 ||
794 rxmode->header_split == 1)