f6d23c9fb668fe36f9a446851a88b11997827425
[dpdk.git] / drivers / net / i40e / i40e_rxtx.c
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2010-2016 Intel Corporation
3  */
4
5 #include <stdio.h>
6 #include <stdlib.h>
7 #include <string.h>
8 #include <errno.h>
9 #include <stdint.h>
10 #include <stdarg.h>
11 #include <unistd.h>
12 #include <inttypes.h>
13 #include <sys/queue.h>
14
15 #include <rte_string_fns.h>
16 #include <rte_memzone.h>
17 #include <rte_mbuf.h>
18 #include <rte_malloc.h>
19 #include <rte_ether.h>
20 #include <rte_ethdev_driver.h>
21 #include <rte_tcp.h>
22 #include <rte_sctp.h>
23 #include <rte_udp.h>
24 #include <rte_ip.h>
25 #include <rte_net.h>
26
27 #include "i40e_logs.h"
28 #include "base/i40e_prototype.h"
29 #include "base/i40e_type.h"
30 #include "i40e_ethdev.h"
31 #include "i40e_rxtx.h"
32
33 #define DEFAULT_TX_RS_THRESH   32
34 #define DEFAULT_TX_FREE_THRESH 32
35
36 #define I40E_TX_MAX_BURST  32
37
38 #define I40E_DMA_MEM_ALIGN 4096
39
40 /* Base address of the HW descriptor ring should be 128B aligned. */
41 #define I40E_RING_BASE_ALIGN    128
42
43 #define I40E_TXD_CMD (I40E_TX_DESC_CMD_EOP | I40E_TX_DESC_CMD_RS)
44
45 #ifdef RTE_LIBRTE_IEEE1588
46 #define I40E_TX_IEEE1588_TMST PKT_TX_IEEE1588_TMST
47 #else
48 #define I40E_TX_IEEE1588_TMST 0
49 #endif
50
51 #define I40E_TX_CKSUM_OFFLOAD_MASK (             \
52                 PKT_TX_IP_CKSUM |                \
53                 PKT_TX_L4_MASK |                 \
54                 PKT_TX_TCP_SEG |                 \
55                 PKT_TX_OUTER_IP_CKSUM)
56
57 #define I40E_TX_OFFLOAD_MASK (  \
58                 PKT_TX_OUTER_IPV4 |     \
59                 PKT_TX_OUTER_IPV6 |     \
60                 PKT_TX_IPV4 |           \
61                 PKT_TX_IPV6 |           \
62                 PKT_TX_IP_CKSUM |       \
63                 PKT_TX_L4_MASK |        \
64                 PKT_TX_OUTER_IP_CKSUM | \
65                 PKT_TX_TCP_SEG |        \
66                 PKT_TX_QINQ_PKT |       \
67                 PKT_TX_VLAN_PKT |       \
68                 PKT_TX_TUNNEL_MASK |    \
69                 I40E_TX_IEEE1588_TMST)
70
71 #define I40E_TX_OFFLOAD_NOTSUP_MASK \
72                 (PKT_TX_OFFLOAD_MASK ^ I40E_TX_OFFLOAD_MASK)
73
74 static inline void
75 i40e_rxd_to_vlan_tci(struct rte_mbuf *mb, volatile union i40e_rx_desc *rxdp)
76 {
77         if (rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len) &
78                 (1 << I40E_RX_DESC_STATUS_L2TAG1P_SHIFT)) {
79                 mb->ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
80                 mb->vlan_tci =
81                         rte_le_to_cpu_16(rxdp->wb.qword0.lo_dword.l2tag1);
82                 PMD_RX_LOG(DEBUG, "Descriptor l2tag1: %u",
83                            rte_le_to_cpu_16(rxdp->wb.qword0.lo_dword.l2tag1));
84         } else {
85                 mb->vlan_tci = 0;
86         }
87 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
88         if (rte_le_to_cpu_16(rxdp->wb.qword2.ext_status) &
89                 (1 << I40E_RX_DESC_EXT_STATUS_L2TAG2P_SHIFT)) {
90                 mb->ol_flags |= PKT_RX_QINQ_STRIPPED | PKT_RX_QINQ |
91                         PKT_RX_VLAN_STRIPPED | PKT_RX_VLAN;
92                 mb->vlan_tci_outer = mb->vlan_tci;
93                 mb->vlan_tci = rte_le_to_cpu_16(rxdp->wb.qword2.l2tag2_2);
94                 PMD_RX_LOG(DEBUG, "Descriptor l2tag2_1: %u, l2tag2_2: %u",
95                            rte_le_to_cpu_16(rxdp->wb.qword2.l2tag2_1),
96                            rte_le_to_cpu_16(rxdp->wb.qword2.l2tag2_2));
97         } else {
98                 mb->vlan_tci_outer = 0;
99         }
100 #endif
101         PMD_RX_LOG(DEBUG, "Mbuf vlan_tci: %u, vlan_tci_outer: %u",
102                    mb->vlan_tci, mb->vlan_tci_outer);
103 }
104
105 /* Translate the rx descriptor status to pkt flags */
106 static inline uint64_t
107 i40e_rxd_status_to_pkt_flags(uint64_t qword)
108 {
109         uint64_t flags;
110
111         /* Check if RSS_HASH */
112         flags = (((qword >> I40E_RX_DESC_STATUS_FLTSTAT_SHIFT) &
113                                         I40E_RX_DESC_FLTSTAT_RSS_HASH) ==
114                         I40E_RX_DESC_FLTSTAT_RSS_HASH) ? PKT_RX_RSS_HASH : 0;
115
116         /* Check if FDIR Match */
117         flags |= (qword & (1 << I40E_RX_DESC_STATUS_FLM_SHIFT) ?
118                                                         PKT_RX_FDIR : 0);
119
120         return flags;
121 }
122
123 static inline uint64_t
124 i40e_rxd_error_to_pkt_flags(uint64_t qword)
125 {
126         uint64_t flags = 0;
127         uint64_t error_bits = (qword >> I40E_RXD_QW1_ERROR_SHIFT);
128
129 #define I40E_RX_ERR_BITS 0x3f
130         if (likely((error_bits & I40E_RX_ERR_BITS) == 0)) {
131                 flags |= (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD);
132                 return flags;
133         }
134
135         if (unlikely(error_bits & (1 << I40E_RX_DESC_ERROR_IPE_SHIFT)))
136                 flags |= PKT_RX_IP_CKSUM_BAD;
137         else
138                 flags |= PKT_RX_IP_CKSUM_GOOD;
139
140         if (unlikely(error_bits & (1 << I40E_RX_DESC_ERROR_L4E_SHIFT)))
141                 flags |= PKT_RX_L4_CKSUM_BAD;
142         else
143                 flags |= PKT_RX_L4_CKSUM_GOOD;
144
145         if (unlikely(error_bits & (1 << I40E_RX_DESC_ERROR_EIPE_SHIFT)))
146                 flags |= PKT_RX_EIP_CKSUM_BAD;
147
148         return flags;
149 }
150
151 /* Function to check and set the ieee1588 timesync index and get the
152  * appropriate flags.
153  */
154 #ifdef RTE_LIBRTE_IEEE1588
155 static inline uint64_t
156 i40e_get_iee15888_flags(struct rte_mbuf *mb, uint64_t qword)
157 {
158         uint64_t pkt_flags = 0;
159         uint16_t tsyn = (qword & (I40E_RXD_QW1_STATUS_TSYNVALID_MASK
160                                   | I40E_RXD_QW1_STATUS_TSYNINDX_MASK))
161                                     >> I40E_RX_DESC_STATUS_TSYNINDX_SHIFT;
162
163         if ((mb->packet_type & RTE_PTYPE_L2_MASK)
164                         == RTE_PTYPE_L2_ETHER_TIMESYNC)
165                 pkt_flags = PKT_RX_IEEE1588_PTP;
166         if (tsyn & 0x04) {
167                 pkt_flags |= PKT_RX_IEEE1588_TMST;
168                 mb->timesync = tsyn & 0x03;
169         }
170
171         return pkt_flags;
172 }
173 #endif
174
175 static inline uint64_t
176 i40e_rxd_build_fdir(volatile union i40e_rx_desc *rxdp, struct rte_mbuf *mb)
177 {
178         uint64_t flags = 0;
179 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
180         uint16_t flexbh, flexbl;
181
182         flexbh = (rte_le_to_cpu_32(rxdp->wb.qword2.ext_status) >>
183                 I40E_RX_DESC_EXT_STATUS_FLEXBH_SHIFT) &
184                 I40E_RX_DESC_EXT_STATUS_FLEXBH_MASK;
185         flexbl = (rte_le_to_cpu_32(rxdp->wb.qword2.ext_status) >>
186                 I40E_RX_DESC_EXT_STATUS_FLEXBL_SHIFT) &
187                 I40E_RX_DESC_EXT_STATUS_FLEXBL_MASK;
188
189
190         if (flexbh == I40E_RX_DESC_EXT_STATUS_FLEXBH_FD_ID) {
191                 mb->hash.fdir.hi =
192                         rte_le_to_cpu_32(rxdp->wb.qword3.hi_dword.fd_id);
193                 flags |= PKT_RX_FDIR_ID;
194         } else if (flexbh == I40E_RX_DESC_EXT_STATUS_FLEXBH_FLEX) {
195                 mb->hash.fdir.hi =
196                         rte_le_to_cpu_32(rxdp->wb.qword3.hi_dword.flex_bytes_hi);
197                 flags |= PKT_RX_FDIR_FLX;
198         }
199         if (flexbl == I40E_RX_DESC_EXT_STATUS_FLEXBL_FLEX) {
200                 mb->hash.fdir.lo =
201                         rte_le_to_cpu_32(rxdp->wb.qword3.lo_dword.flex_bytes_lo);
202                 flags |= PKT_RX_FDIR_FLX;
203         }
204 #else
205         mb->hash.fdir.hi =
206                 rte_le_to_cpu_32(rxdp->wb.qword0.hi_dword.fd_id);
207         flags |= PKT_RX_FDIR_ID;
208 #endif
209         return flags;
210 }
211
212 static inline void
213 i40e_parse_tunneling_params(uint64_t ol_flags,
214                             union i40e_tx_offload tx_offload,
215                             uint32_t *cd_tunneling)
216 {
217         /* EIPT: External (outer) IP header type */
218         if (ol_flags & PKT_TX_OUTER_IP_CKSUM)
219                 *cd_tunneling |= I40E_TX_CTX_EXT_IP_IPV4;
220         else if (ol_flags & PKT_TX_OUTER_IPV4)
221                 *cd_tunneling |= I40E_TX_CTX_EXT_IP_IPV4_NO_CSUM;
222         else if (ol_flags & PKT_TX_OUTER_IPV6)
223                 *cd_tunneling |= I40E_TX_CTX_EXT_IP_IPV6;
224
225         /* EIPLEN: External (outer) IP header length, in DWords */
226         *cd_tunneling |= (tx_offload.outer_l3_len >> 2) <<
227                 I40E_TXD_CTX_QW0_EXT_IPLEN_SHIFT;
228
229         /* L4TUNT: L4 Tunneling Type */
230         switch (ol_flags & PKT_TX_TUNNEL_MASK) {
231         case PKT_TX_TUNNEL_IPIP:
232                 /* for non UDP / GRE tunneling, set to 00b */
233                 break;
234         case PKT_TX_TUNNEL_VXLAN:
235         case PKT_TX_TUNNEL_GENEVE:
236                 *cd_tunneling |= I40E_TXD_CTX_UDP_TUNNELING;
237                 break;
238         case PKT_TX_TUNNEL_GRE:
239                 *cd_tunneling |= I40E_TXD_CTX_GRE_TUNNELING;
240                 break;
241         default:
242                 PMD_TX_LOG(ERR, "Tunnel type not supported");
243                 return;
244         }
245
246         /* L4TUNLEN: L4 Tunneling Length, in Words
247          *
248          * We depend on app to set rte_mbuf.l2_len correctly.
249          * For IP in GRE it should be set to the length of the GRE
250          * header;
251          * for MAC in GRE or MAC in UDP it should be set to the length
252          * of the GRE or UDP headers plus the inner MAC up to including
253          * its last Ethertype.
254          */
255         *cd_tunneling |= (tx_offload.l2_len >> 1) <<
256                 I40E_TXD_CTX_QW0_NATLEN_SHIFT;
257 }
258
259 static inline void
260 i40e_txd_enable_checksum(uint64_t ol_flags,
261                         uint32_t *td_cmd,
262                         uint32_t *td_offset,
263                         union i40e_tx_offload tx_offload)
264 {
265         /* Set MACLEN */
266         if (ol_flags & PKT_TX_TUNNEL_MASK)
267                 *td_offset |= (tx_offload.outer_l2_len >> 1)
268                                 << I40E_TX_DESC_LENGTH_MACLEN_SHIFT;
269         else
270                 *td_offset |= (tx_offload.l2_len >> 1)
271                         << I40E_TX_DESC_LENGTH_MACLEN_SHIFT;
272
273         /* Enable L3 checksum offloads */
274         if (ol_flags & PKT_TX_IP_CKSUM) {
275                 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV4_CSUM;
276                 *td_offset |= (tx_offload.l3_len >> 2)
277                                 << I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
278         } else if (ol_flags & PKT_TX_IPV4) {
279                 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV4;
280                 *td_offset |= (tx_offload.l3_len >> 2)
281                                 << I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
282         } else if (ol_flags & PKT_TX_IPV6) {
283                 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV6;
284                 *td_offset |= (tx_offload.l3_len >> 2)
285                                 << I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
286         }
287
288         if (ol_flags & PKT_TX_TCP_SEG) {
289                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_TCP;
290                 *td_offset |= (tx_offload.l4_len >> 2)
291                         << I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
292                 return;
293         }
294
295         /* Enable L4 checksum offloads */
296         switch (ol_flags & PKT_TX_L4_MASK) {
297         case PKT_TX_TCP_CKSUM:
298                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_TCP;
299                 *td_offset |= (sizeof(struct rte_tcp_hdr) >> 2) <<
300                                 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
301                 break;
302         case PKT_TX_SCTP_CKSUM:
303                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_SCTP;
304                 *td_offset |= (sizeof(struct rte_sctp_hdr) >> 2) <<
305                                 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
306                 break;
307         case PKT_TX_UDP_CKSUM:
308                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_UDP;
309                 *td_offset |= (sizeof(struct rte_udp_hdr) >> 2) <<
310                                 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
311                 break;
312         default:
313                 break;
314         }
315 }
316
317 /* Construct the tx flags */
318 static inline uint64_t
319 i40e_build_ctob(uint32_t td_cmd,
320                 uint32_t td_offset,
321                 unsigned int size,
322                 uint32_t td_tag)
323 {
324         return rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DATA |
325                         ((uint64_t)td_cmd  << I40E_TXD_QW1_CMD_SHIFT) |
326                         ((uint64_t)td_offset << I40E_TXD_QW1_OFFSET_SHIFT) |
327                         ((uint64_t)size  << I40E_TXD_QW1_TX_BUF_SZ_SHIFT) |
328                         ((uint64_t)td_tag  << I40E_TXD_QW1_L2TAG1_SHIFT));
329 }
330
331 static inline int
332 i40e_xmit_cleanup(struct i40e_tx_queue *txq)
333 {
334         struct i40e_tx_entry *sw_ring = txq->sw_ring;
335         volatile struct i40e_tx_desc *txd = txq->tx_ring;
336         uint16_t last_desc_cleaned = txq->last_desc_cleaned;
337         uint16_t nb_tx_desc = txq->nb_tx_desc;
338         uint16_t desc_to_clean_to;
339         uint16_t nb_tx_to_clean;
340
341         desc_to_clean_to = (uint16_t)(last_desc_cleaned + txq->tx_rs_thresh);
342         if (desc_to_clean_to >= nb_tx_desc)
343                 desc_to_clean_to = (uint16_t)(desc_to_clean_to - nb_tx_desc);
344
345         desc_to_clean_to = sw_ring[desc_to_clean_to].last_id;
346         if ((txd[desc_to_clean_to].cmd_type_offset_bsz &
347                         rte_cpu_to_le_64(I40E_TXD_QW1_DTYPE_MASK)) !=
348                         rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DESC_DONE)) {
349                 PMD_TX_FREE_LOG(DEBUG, "TX descriptor %4u is not done "
350                         "(port=%d queue=%d)", desc_to_clean_to,
351                                 txq->port_id, txq->queue_id);
352                 return -1;
353         }
354
355         if (last_desc_cleaned > desc_to_clean_to)
356                 nb_tx_to_clean = (uint16_t)((nb_tx_desc - last_desc_cleaned) +
357                                                         desc_to_clean_to);
358         else
359                 nb_tx_to_clean = (uint16_t)(desc_to_clean_to -
360                                         last_desc_cleaned);
361
362         txd[desc_to_clean_to].cmd_type_offset_bsz = 0;
363
364         txq->last_desc_cleaned = desc_to_clean_to;
365         txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + nb_tx_to_clean);
366
367         return 0;
368 }
369
370 static inline int
371 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
372 check_rx_burst_bulk_alloc_preconditions(struct i40e_rx_queue *rxq)
373 #else
374 check_rx_burst_bulk_alloc_preconditions(__rte_unused struct i40e_rx_queue *rxq)
375 #endif
376 {
377         int ret = 0;
378
379 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
380         if (!(rxq->rx_free_thresh >= RTE_PMD_I40E_RX_MAX_BURST)) {
381                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
382                              "rxq->rx_free_thresh=%d, "
383                              "RTE_PMD_I40E_RX_MAX_BURST=%d",
384                              rxq->rx_free_thresh, RTE_PMD_I40E_RX_MAX_BURST);
385                 ret = -EINVAL;
386         } else if (!(rxq->rx_free_thresh < rxq->nb_rx_desc)) {
387                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
388                              "rxq->rx_free_thresh=%d, "
389                              "rxq->nb_rx_desc=%d",
390                              rxq->rx_free_thresh, rxq->nb_rx_desc);
391                 ret = -EINVAL;
392         } else if (rxq->nb_rx_desc % rxq->rx_free_thresh != 0) {
393                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
394                              "rxq->nb_rx_desc=%d, "
395                              "rxq->rx_free_thresh=%d",
396                              rxq->nb_rx_desc, rxq->rx_free_thresh);
397                 ret = -EINVAL;
398         }
399 #else
400         ret = -EINVAL;
401 #endif
402
403         return ret;
404 }
405
406 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
407 #define I40E_LOOK_AHEAD 8
408 #if (I40E_LOOK_AHEAD != 8)
409 #error "PMD I40E: I40E_LOOK_AHEAD must be 8\n"
410 #endif
411 static inline int
412 i40e_rx_scan_hw_ring(struct i40e_rx_queue *rxq)
413 {
414         volatile union i40e_rx_desc *rxdp;
415         struct i40e_rx_entry *rxep;
416         struct rte_mbuf *mb;
417         uint16_t pkt_len;
418         uint64_t qword1;
419         uint32_t rx_status;
420         int32_t s[I40E_LOOK_AHEAD], nb_dd;
421         int32_t i, j, nb_rx = 0;
422         uint64_t pkt_flags;
423         uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
424
425         rxdp = &rxq->rx_ring[rxq->rx_tail];
426         rxep = &rxq->sw_ring[rxq->rx_tail];
427
428         qword1 = rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len);
429         rx_status = (qword1 & I40E_RXD_QW1_STATUS_MASK) >>
430                                 I40E_RXD_QW1_STATUS_SHIFT;
431
432         /* Make sure there is at least 1 packet to receive */
433         if (!(rx_status & (1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
434                 return 0;
435
436         /**
437          * Scan LOOK_AHEAD descriptors at a time to determine which
438          * descriptors reference packets that are ready to be received.
439          */
440         for (i = 0; i < RTE_PMD_I40E_RX_MAX_BURST; i+=I40E_LOOK_AHEAD,
441                         rxdp += I40E_LOOK_AHEAD, rxep += I40E_LOOK_AHEAD) {
442                 /* Read desc statuses backwards to avoid race condition */
443                 for (j = I40E_LOOK_AHEAD - 1; j >= 0; j--) {
444                         qword1 = rte_le_to_cpu_64(\
445                                 rxdp[j].wb.qword1.status_error_len);
446                         s[j] = (qword1 & I40E_RXD_QW1_STATUS_MASK) >>
447                                         I40E_RXD_QW1_STATUS_SHIFT;
448                 }
449
450                 rte_smp_rmb();
451
452                 /* Compute how many status bits were set */
453                 for (j = 0, nb_dd = 0; j < I40E_LOOK_AHEAD; j++)
454                         nb_dd += s[j] & (1 << I40E_RX_DESC_STATUS_DD_SHIFT);
455
456                 nb_rx += nb_dd;
457
458                 /* Translate descriptor info to mbuf parameters */
459                 for (j = 0; j < nb_dd; j++) {
460                         mb = rxep[j].mbuf;
461                         qword1 = rte_le_to_cpu_64(\
462                                 rxdp[j].wb.qword1.status_error_len);
463                         pkt_len = ((qword1 & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
464                                 I40E_RXD_QW1_LENGTH_PBUF_SHIFT) - rxq->crc_len;
465                         mb->data_len = pkt_len;
466                         mb->pkt_len = pkt_len;
467                         mb->ol_flags = 0;
468                         i40e_rxd_to_vlan_tci(mb, &rxdp[j]);
469                         pkt_flags = i40e_rxd_status_to_pkt_flags(qword1);
470                         pkt_flags |= i40e_rxd_error_to_pkt_flags(qword1);
471                         mb->packet_type =
472                                 ptype_tbl[(uint8_t)((qword1 &
473                                 I40E_RXD_QW1_PTYPE_MASK) >>
474                                 I40E_RXD_QW1_PTYPE_SHIFT)];
475                         if (pkt_flags & PKT_RX_RSS_HASH)
476                                 mb->hash.rss = rte_le_to_cpu_32(\
477                                         rxdp[j].wb.qword0.hi_dword.rss);
478                         if (pkt_flags & PKT_RX_FDIR)
479                                 pkt_flags |= i40e_rxd_build_fdir(&rxdp[j], mb);
480
481 #ifdef RTE_LIBRTE_IEEE1588
482                         pkt_flags |= i40e_get_iee15888_flags(mb, qword1);
483 #endif
484                         mb->ol_flags |= pkt_flags;
485
486                 }
487
488                 for (j = 0; j < I40E_LOOK_AHEAD; j++)
489                         rxq->rx_stage[i + j] = rxep[j].mbuf;
490
491                 if (nb_dd != I40E_LOOK_AHEAD)
492                         break;
493         }
494
495         /* Clear software ring entries */
496         for (i = 0; i < nb_rx; i++)
497                 rxq->sw_ring[rxq->rx_tail + i].mbuf = NULL;
498
499         return nb_rx;
500 }
501
502 static inline uint16_t
503 i40e_rx_fill_from_stage(struct i40e_rx_queue *rxq,
504                         struct rte_mbuf **rx_pkts,
505                         uint16_t nb_pkts)
506 {
507         uint16_t i;
508         struct rte_mbuf **stage = &rxq->rx_stage[rxq->rx_next_avail];
509
510         nb_pkts = (uint16_t)RTE_MIN(nb_pkts, rxq->rx_nb_avail);
511
512         for (i = 0; i < nb_pkts; i++)
513                 rx_pkts[i] = stage[i];
514
515         rxq->rx_nb_avail = (uint16_t)(rxq->rx_nb_avail - nb_pkts);
516         rxq->rx_next_avail = (uint16_t)(rxq->rx_next_avail + nb_pkts);
517
518         return nb_pkts;
519 }
520
521 static inline int
522 i40e_rx_alloc_bufs(struct i40e_rx_queue *rxq)
523 {
524         volatile union i40e_rx_desc *rxdp;
525         struct i40e_rx_entry *rxep;
526         struct rte_mbuf *mb;
527         uint16_t alloc_idx, i;
528         uint64_t dma_addr;
529         int diag;
530
531         /* Allocate buffers in bulk */
532         alloc_idx = (uint16_t)(rxq->rx_free_trigger -
533                                 (rxq->rx_free_thresh - 1));
534         rxep = &(rxq->sw_ring[alloc_idx]);
535         diag = rte_mempool_get_bulk(rxq->mp, (void *)rxep,
536                                         rxq->rx_free_thresh);
537         if (unlikely(diag != 0)) {
538                 PMD_DRV_LOG(ERR, "Failed to get mbufs in bulk");
539                 return -ENOMEM;
540         }
541
542         rxdp = &rxq->rx_ring[alloc_idx];
543         for (i = 0; i < rxq->rx_free_thresh; i++) {
544                 if (likely(i < (rxq->rx_free_thresh - 1)))
545                         /* Prefetch next mbuf */
546                         rte_prefetch0(rxep[i + 1].mbuf);
547
548                 mb = rxep[i].mbuf;
549                 rte_mbuf_refcnt_set(mb, 1);
550                 mb->next = NULL;
551                 mb->data_off = RTE_PKTMBUF_HEADROOM;
552                 mb->nb_segs = 1;
553                 mb->port = rxq->port_id;
554                 dma_addr = rte_cpu_to_le_64(\
555                         rte_mbuf_data_iova_default(mb));
556                 rxdp[i].read.hdr_addr = 0;
557                 rxdp[i].read.pkt_addr = dma_addr;
558         }
559
560         /* Update rx tail regsiter */
561         I40E_PCI_REG_WRITE(rxq->qrx_tail, rxq->rx_free_trigger);
562
563         rxq->rx_free_trigger =
564                 (uint16_t)(rxq->rx_free_trigger + rxq->rx_free_thresh);
565         if (rxq->rx_free_trigger >= rxq->nb_rx_desc)
566                 rxq->rx_free_trigger = (uint16_t)(rxq->rx_free_thresh - 1);
567
568         return 0;
569 }
570
571 static inline uint16_t
572 rx_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
573 {
574         struct i40e_rx_queue *rxq = (struct i40e_rx_queue *)rx_queue;
575         struct rte_eth_dev *dev;
576         uint16_t nb_rx = 0;
577
578         if (!nb_pkts)
579                 return 0;
580
581         if (rxq->rx_nb_avail)
582                 return i40e_rx_fill_from_stage(rxq, rx_pkts, nb_pkts);
583
584         nb_rx = (uint16_t)i40e_rx_scan_hw_ring(rxq);
585         rxq->rx_next_avail = 0;
586         rxq->rx_nb_avail = nb_rx;
587         rxq->rx_tail = (uint16_t)(rxq->rx_tail + nb_rx);
588
589         if (rxq->rx_tail > rxq->rx_free_trigger) {
590                 if (i40e_rx_alloc_bufs(rxq) != 0) {
591                         uint16_t i, j;
592
593                         dev = I40E_VSI_TO_ETH_DEV(rxq->vsi);
594                         dev->data->rx_mbuf_alloc_failed +=
595                                 rxq->rx_free_thresh;
596
597                         rxq->rx_nb_avail = 0;
598                         rxq->rx_tail = (uint16_t)(rxq->rx_tail - nb_rx);
599                         for (i = 0, j = rxq->rx_tail; i < nb_rx; i++, j++)
600                                 rxq->sw_ring[j].mbuf = rxq->rx_stage[i];
601
602                         return 0;
603                 }
604         }
605
606         if (rxq->rx_tail >= rxq->nb_rx_desc)
607                 rxq->rx_tail = 0;
608
609         if (rxq->rx_nb_avail)
610                 return i40e_rx_fill_from_stage(rxq, rx_pkts, nb_pkts);
611
612         return 0;
613 }
614
615 static uint16_t
616 i40e_recv_pkts_bulk_alloc(void *rx_queue,
617                           struct rte_mbuf **rx_pkts,
618                           uint16_t nb_pkts)
619 {
620         uint16_t nb_rx = 0, n, count;
621
622         if (unlikely(nb_pkts == 0))
623                 return 0;
624
625         if (likely(nb_pkts <= RTE_PMD_I40E_RX_MAX_BURST))
626                 return rx_recv_pkts(rx_queue, rx_pkts, nb_pkts);
627
628         while (nb_pkts) {
629                 n = RTE_MIN(nb_pkts, RTE_PMD_I40E_RX_MAX_BURST);
630                 count = rx_recv_pkts(rx_queue, &rx_pkts[nb_rx], n);
631                 nb_rx = (uint16_t)(nb_rx + count);
632                 nb_pkts = (uint16_t)(nb_pkts - count);
633                 if (count < n)
634                         break;
635         }
636
637         return nb_rx;
638 }
639 #else
640 static uint16_t
641 i40e_recv_pkts_bulk_alloc(void __rte_unused *rx_queue,
642                           struct rte_mbuf __rte_unused **rx_pkts,
643                           uint16_t __rte_unused nb_pkts)
644 {
645         return 0;
646 }
647 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
648
649 uint16_t
650 i40e_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
651 {
652         struct i40e_rx_queue *rxq;
653         volatile union i40e_rx_desc *rx_ring;
654         volatile union i40e_rx_desc *rxdp;
655         union i40e_rx_desc rxd;
656         struct i40e_rx_entry *sw_ring;
657         struct i40e_rx_entry *rxe;
658         struct rte_eth_dev *dev;
659         struct rte_mbuf *rxm;
660         struct rte_mbuf *nmb;
661         uint16_t nb_rx;
662         uint32_t rx_status;
663         uint64_t qword1;
664         uint16_t rx_packet_len;
665         uint16_t rx_id, nb_hold;
666         uint64_t dma_addr;
667         uint64_t pkt_flags;
668         uint32_t *ptype_tbl;
669
670         nb_rx = 0;
671         nb_hold = 0;
672         rxq = rx_queue;
673         rx_id = rxq->rx_tail;
674         rx_ring = rxq->rx_ring;
675         sw_ring = rxq->sw_ring;
676         ptype_tbl = rxq->vsi->adapter->ptype_tbl;
677
678         while (nb_rx < nb_pkts) {
679                 rxdp = &rx_ring[rx_id];
680                 qword1 = rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len);
681                 rx_status = (qword1 & I40E_RXD_QW1_STATUS_MASK)
682                                 >> I40E_RXD_QW1_STATUS_SHIFT;
683
684                 /* Check the DD bit first */
685                 if (!(rx_status & (1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
686                         break;
687
688                 nmb = rte_mbuf_raw_alloc(rxq->mp);
689                 if (unlikely(!nmb)) {
690                         dev = I40E_VSI_TO_ETH_DEV(rxq->vsi);
691                         dev->data->rx_mbuf_alloc_failed++;
692                         break;
693                 }
694
695                 rxd = *rxdp;
696                 nb_hold++;
697                 rxe = &sw_ring[rx_id];
698                 rx_id++;
699                 if (unlikely(rx_id == rxq->nb_rx_desc))
700                         rx_id = 0;
701
702                 /* Prefetch next mbuf */
703                 rte_prefetch0(sw_ring[rx_id].mbuf);
704
705                 /**
706                  * When next RX descriptor is on a cache line boundary,
707                  * prefetch the next 4 RX descriptors and next 8 pointers
708                  * to mbufs.
709                  */
710                 if ((rx_id & 0x3) == 0) {
711                         rte_prefetch0(&rx_ring[rx_id]);
712                         rte_prefetch0(&sw_ring[rx_id]);
713                 }
714                 rxm = rxe->mbuf;
715                 rxe->mbuf = nmb;
716                 dma_addr =
717                         rte_cpu_to_le_64(rte_mbuf_data_iova_default(nmb));
718                 rxdp->read.hdr_addr = 0;
719                 rxdp->read.pkt_addr = dma_addr;
720
721                 rx_packet_len = ((qword1 & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
722                                 I40E_RXD_QW1_LENGTH_PBUF_SHIFT) - rxq->crc_len;
723
724                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
725                 rte_prefetch0(RTE_PTR_ADD(rxm->buf_addr, RTE_PKTMBUF_HEADROOM));
726                 rxm->nb_segs = 1;
727                 rxm->next = NULL;
728                 rxm->pkt_len = rx_packet_len;
729                 rxm->data_len = rx_packet_len;
730                 rxm->port = rxq->port_id;
731                 rxm->ol_flags = 0;
732                 i40e_rxd_to_vlan_tci(rxm, &rxd);
733                 pkt_flags = i40e_rxd_status_to_pkt_flags(qword1);
734                 pkt_flags |= i40e_rxd_error_to_pkt_flags(qword1);
735                 rxm->packet_type =
736                         ptype_tbl[(uint8_t)((qword1 &
737                         I40E_RXD_QW1_PTYPE_MASK) >> I40E_RXD_QW1_PTYPE_SHIFT)];
738                 if (pkt_flags & PKT_RX_RSS_HASH)
739                         rxm->hash.rss =
740                                 rte_le_to_cpu_32(rxd.wb.qword0.hi_dword.rss);
741                 if (pkt_flags & PKT_RX_FDIR)
742                         pkt_flags |= i40e_rxd_build_fdir(&rxd, rxm);
743
744 #ifdef RTE_LIBRTE_IEEE1588
745                 pkt_flags |= i40e_get_iee15888_flags(rxm, qword1);
746 #endif
747                 rxm->ol_flags |= pkt_flags;
748
749                 rx_pkts[nb_rx++] = rxm;
750         }
751         rxq->rx_tail = rx_id;
752
753         /**
754          * If the number of free RX descriptors is greater than the RX free
755          * threshold of the queue, advance the receive tail register of queue.
756          * Update that register with the value of the last processed RX
757          * descriptor minus 1.
758          */
759         nb_hold = (uint16_t)(nb_hold + rxq->nb_rx_hold);
760         if (nb_hold > rxq->rx_free_thresh) {
761                 rx_id = (uint16_t) ((rx_id == 0) ?
762                         (rxq->nb_rx_desc - 1) : (rx_id - 1));
763                 I40E_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
764                 nb_hold = 0;
765         }
766         rxq->nb_rx_hold = nb_hold;
767
768         return nb_rx;
769 }
770
771 uint16_t
772 i40e_recv_scattered_pkts(void *rx_queue,
773                          struct rte_mbuf **rx_pkts,
774                          uint16_t nb_pkts)
775 {
776         struct i40e_rx_queue *rxq = rx_queue;
777         volatile union i40e_rx_desc *rx_ring = rxq->rx_ring;
778         volatile union i40e_rx_desc *rxdp;
779         union i40e_rx_desc rxd;
780         struct i40e_rx_entry *sw_ring = rxq->sw_ring;
781         struct i40e_rx_entry *rxe;
782         struct rte_mbuf *first_seg = rxq->pkt_first_seg;
783         struct rte_mbuf *last_seg = rxq->pkt_last_seg;
784         struct rte_mbuf *nmb, *rxm;
785         uint16_t rx_id = rxq->rx_tail;
786         uint16_t nb_rx = 0, nb_hold = 0, rx_packet_len;
787         struct rte_eth_dev *dev;
788         uint32_t rx_status;
789         uint64_t qword1;
790         uint64_t dma_addr;
791         uint64_t pkt_flags;
792         uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
793
794         while (nb_rx < nb_pkts) {
795                 rxdp = &rx_ring[rx_id];
796                 qword1 = rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len);
797                 rx_status = (qword1 & I40E_RXD_QW1_STATUS_MASK) >>
798                                         I40E_RXD_QW1_STATUS_SHIFT;
799
800                 /* Check the DD bit */
801                 if (!(rx_status & (1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
802                         break;
803
804                 nmb = rte_mbuf_raw_alloc(rxq->mp);
805                 if (unlikely(!nmb)) {
806                         dev = I40E_VSI_TO_ETH_DEV(rxq->vsi);
807                         dev->data->rx_mbuf_alloc_failed++;
808                         break;
809                 }
810
811                 rxd = *rxdp;
812                 nb_hold++;
813                 rxe = &sw_ring[rx_id];
814                 rx_id++;
815                 if (rx_id == rxq->nb_rx_desc)
816                         rx_id = 0;
817
818                 /* Prefetch next mbuf */
819                 rte_prefetch0(sw_ring[rx_id].mbuf);
820
821                 /**
822                  * When next RX descriptor is on a cache line boundary,
823                  * prefetch the next 4 RX descriptors and next 8 pointers
824                  * to mbufs.
825                  */
826                 if ((rx_id & 0x3) == 0) {
827                         rte_prefetch0(&rx_ring[rx_id]);
828                         rte_prefetch0(&sw_ring[rx_id]);
829                 }
830
831                 rxm = rxe->mbuf;
832                 rxe->mbuf = nmb;
833                 dma_addr =
834                         rte_cpu_to_le_64(rte_mbuf_data_iova_default(nmb));
835
836                 /* Set data buffer address and data length of the mbuf */
837                 rxdp->read.hdr_addr = 0;
838                 rxdp->read.pkt_addr = dma_addr;
839                 rx_packet_len = (qword1 & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
840                                         I40E_RXD_QW1_LENGTH_PBUF_SHIFT;
841                 rxm->data_len = rx_packet_len;
842                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
843
844                 /**
845                  * If this is the first buffer of the received packet, set the
846                  * pointer to the first mbuf of the packet and initialize its
847                  * context. Otherwise, update the total length and the number
848                  * of segments of the current scattered packet, and update the
849                  * pointer to the last mbuf of the current packet.
850                  */
851                 if (!first_seg) {
852                         first_seg = rxm;
853                         first_seg->nb_segs = 1;
854                         first_seg->pkt_len = rx_packet_len;
855                 } else {
856                         first_seg->pkt_len =
857                                 (uint16_t)(first_seg->pkt_len +
858                                                 rx_packet_len);
859                         first_seg->nb_segs++;
860                         last_seg->next = rxm;
861                 }
862
863                 /**
864                  * If this is not the last buffer of the received packet,
865                  * update the pointer to the last mbuf of the current scattered
866                  * packet and continue to parse the RX ring.
867                  */
868                 if (!(rx_status & (1 << I40E_RX_DESC_STATUS_EOF_SHIFT))) {
869                         last_seg = rxm;
870                         continue;
871                 }
872
873                 /**
874                  * This is the last buffer of the received packet. If the CRC
875                  * is not stripped by the hardware:
876                  *  - Subtract the CRC length from the total packet length.
877                  *  - If the last buffer only contains the whole CRC or a part
878                  *  of it, free the mbuf associated to the last buffer. If part
879                  *  of the CRC is also contained in the previous mbuf, subtract
880                  *  the length of that CRC part from the data length of the
881                  *  previous mbuf.
882                  */
883                 rxm->next = NULL;
884                 if (unlikely(rxq->crc_len > 0)) {
885                         first_seg->pkt_len -= RTE_ETHER_CRC_LEN;
886                         if (rx_packet_len <= RTE_ETHER_CRC_LEN) {
887                                 rte_pktmbuf_free_seg(rxm);
888                                 first_seg->nb_segs--;
889                                 last_seg->data_len =
890                                         (uint16_t)(last_seg->data_len -
891                                         (RTE_ETHER_CRC_LEN - rx_packet_len));
892                                 last_seg->next = NULL;
893                         } else
894                                 rxm->data_len = (uint16_t)(rx_packet_len -
895                                                         RTE_ETHER_CRC_LEN);
896                 }
897
898                 first_seg->port = rxq->port_id;
899                 first_seg->ol_flags = 0;
900                 i40e_rxd_to_vlan_tci(first_seg, &rxd);
901                 pkt_flags = i40e_rxd_status_to_pkt_flags(qword1);
902                 pkt_flags |= i40e_rxd_error_to_pkt_flags(qword1);
903                 first_seg->packet_type =
904                         ptype_tbl[(uint8_t)((qword1 &
905                         I40E_RXD_QW1_PTYPE_MASK) >> I40E_RXD_QW1_PTYPE_SHIFT)];
906                 if (pkt_flags & PKT_RX_RSS_HASH)
907                         first_seg->hash.rss =
908                                 rte_le_to_cpu_32(rxd.wb.qword0.hi_dword.rss);
909                 if (pkt_flags & PKT_RX_FDIR)
910                         pkt_flags |= i40e_rxd_build_fdir(&rxd, first_seg);
911
912 #ifdef RTE_LIBRTE_IEEE1588
913                 pkt_flags |= i40e_get_iee15888_flags(first_seg, qword1);
914 #endif
915                 first_seg->ol_flags |= pkt_flags;
916
917                 /* Prefetch data of first segment, if configured to do so. */
918                 rte_prefetch0(RTE_PTR_ADD(first_seg->buf_addr,
919                         first_seg->data_off));
920                 rx_pkts[nb_rx++] = first_seg;
921                 first_seg = NULL;
922         }
923
924         /* Record index of the next RX descriptor to probe. */
925         rxq->rx_tail = rx_id;
926         rxq->pkt_first_seg = first_seg;
927         rxq->pkt_last_seg = last_seg;
928
929         /**
930          * If the number of free RX descriptors is greater than the RX free
931          * threshold of the queue, advance the Receive Descriptor Tail (RDT)
932          * register. Update the RDT with the value of the last processed RX
933          * descriptor minus 1, to guarantee that the RDT register is never
934          * equal to the RDH register, which creates a "full" ring situtation
935          * from the hardware point of view.
936          */
937         nb_hold = (uint16_t)(nb_hold + rxq->nb_rx_hold);
938         if (nb_hold > rxq->rx_free_thresh) {
939                 rx_id = (uint16_t)(rx_id == 0 ?
940                         (rxq->nb_rx_desc - 1) : (rx_id - 1));
941                 I40E_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
942                 nb_hold = 0;
943         }
944         rxq->nb_rx_hold = nb_hold;
945
946         return nb_rx;
947 }
948
949 /* Check if the context descriptor is needed for TX offloading */
950 static inline uint16_t
951 i40e_calc_context_desc(uint64_t flags)
952 {
953         static uint64_t mask = PKT_TX_OUTER_IP_CKSUM |
954                 PKT_TX_TCP_SEG |
955                 PKT_TX_QINQ_PKT |
956                 PKT_TX_TUNNEL_MASK;
957
958 #ifdef RTE_LIBRTE_IEEE1588
959         mask |= PKT_TX_IEEE1588_TMST;
960 #endif
961
962         return (flags & mask) ? 1 : 0;
963 }
964
965 /* set i40e TSO context descriptor */
966 static inline uint64_t
967 i40e_set_tso_ctx(struct rte_mbuf *mbuf, union i40e_tx_offload tx_offload)
968 {
969         uint64_t ctx_desc = 0;
970         uint32_t cd_cmd, hdr_len, cd_tso_len;
971
972         if (!tx_offload.l4_len) {
973                 PMD_DRV_LOG(DEBUG, "L4 length set to 0");
974                 return ctx_desc;
975         }
976
977         hdr_len = tx_offload.l2_len + tx_offload.l3_len + tx_offload.l4_len;
978         hdr_len += (mbuf->ol_flags & PKT_TX_TUNNEL_MASK) ?
979                    tx_offload.outer_l2_len + tx_offload.outer_l3_len : 0;
980
981         cd_cmd = I40E_TX_CTX_DESC_TSO;
982         cd_tso_len = mbuf->pkt_len - hdr_len;
983         ctx_desc |= ((uint64_t)cd_cmd << I40E_TXD_CTX_QW1_CMD_SHIFT) |
984                 ((uint64_t)cd_tso_len <<
985                  I40E_TXD_CTX_QW1_TSO_LEN_SHIFT) |
986                 ((uint64_t)mbuf->tso_segsz <<
987                  I40E_TXD_CTX_QW1_MSS_SHIFT);
988
989         return ctx_desc;
990 }
991
992 /* HW requires that Tx buffer size ranges from 1B up to (16K-1)B. */
993 #define I40E_MAX_DATA_PER_TXD \
994         (I40E_TXD_QW1_TX_BUF_SZ_MASK >> I40E_TXD_QW1_TX_BUF_SZ_SHIFT)
995 /* Calculate the number of TX descriptors needed for each pkt */
996 static inline uint16_t
997 i40e_calc_pkt_desc(struct rte_mbuf *tx_pkt)
998 {
999         struct rte_mbuf *txd = tx_pkt;
1000         uint16_t count = 0;
1001
1002         while (txd != NULL) {
1003                 count += DIV_ROUND_UP(txd->data_len, I40E_MAX_DATA_PER_TXD);
1004                 txd = txd->next;
1005         }
1006
1007         return count;
1008 }
1009
1010 uint16_t
1011 i40e_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1012 {
1013         struct i40e_tx_queue *txq;
1014         struct i40e_tx_entry *sw_ring;
1015         struct i40e_tx_entry *txe, *txn;
1016         volatile struct i40e_tx_desc *txd;
1017         volatile struct i40e_tx_desc *txr;
1018         struct rte_mbuf *tx_pkt;
1019         struct rte_mbuf *m_seg;
1020         uint32_t cd_tunneling_params;
1021         uint16_t tx_id;
1022         uint16_t nb_tx;
1023         uint32_t td_cmd;
1024         uint32_t td_offset;
1025         uint32_t td_tag;
1026         uint64_t ol_flags;
1027         uint16_t nb_used;
1028         uint16_t nb_ctx;
1029         uint16_t tx_last;
1030         uint16_t slen;
1031         uint64_t buf_dma_addr;
1032         union i40e_tx_offload tx_offload = {0};
1033
1034         txq = tx_queue;
1035         sw_ring = txq->sw_ring;
1036         txr = txq->tx_ring;
1037         tx_id = txq->tx_tail;
1038         txe = &sw_ring[tx_id];
1039
1040         /* Check if the descriptor ring needs to be cleaned. */
1041         if (txq->nb_tx_free < txq->tx_free_thresh)
1042                 (void)i40e_xmit_cleanup(txq);
1043
1044         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1045                 td_cmd = 0;
1046                 td_tag = 0;
1047                 td_offset = 0;
1048
1049                 tx_pkt = *tx_pkts++;
1050                 RTE_MBUF_PREFETCH_TO_FREE(txe->mbuf);
1051
1052                 ol_flags = tx_pkt->ol_flags;
1053                 tx_offload.l2_len = tx_pkt->l2_len;
1054                 tx_offload.l3_len = tx_pkt->l3_len;
1055                 tx_offload.outer_l2_len = tx_pkt->outer_l2_len;
1056                 tx_offload.outer_l3_len = tx_pkt->outer_l3_len;
1057                 tx_offload.l4_len = tx_pkt->l4_len;
1058                 tx_offload.tso_segsz = tx_pkt->tso_segsz;
1059
1060                 /* Calculate the number of context descriptors needed. */
1061                 nb_ctx = i40e_calc_context_desc(ol_flags);
1062
1063                 /**
1064                  * The number of descriptors that must be allocated for
1065                  * a packet equals to the number of the segments of that
1066                  * packet plus 1 context descriptor if needed.
1067                  * Recalculate the needed tx descs when TSO enabled in case
1068                  * the mbuf data size exceeds max data size that hw allows
1069                  * per tx desc.
1070                  */
1071                 if (ol_flags & PKT_TX_TCP_SEG)
1072                         nb_used = (uint16_t)(i40e_calc_pkt_desc(tx_pkt) +
1073                                              nb_ctx);
1074                 else
1075                         nb_used = (uint16_t)(tx_pkt->nb_segs + nb_ctx);
1076                 tx_last = (uint16_t)(tx_id + nb_used - 1);
1077
1078                 /* Circular ring */
1079                 if (tx_last >= txq->nb_tx_desc)
1080                         tx_last = (uint16_t)(tx_last - txq->nb_tx_desc);
1081
1082                 if (nb_used > txq->nb_tx_free) {
1083                         if (i40e_xmit_cleanup(txq) != 0) {
1084                                 if (nb_tx == 0)
1085                                         return 0;
1086                                 goto end_of_tx;
1087                         }
1088                         if (unlikely(nb_used > txq->tx_rs_thresh)) {
1089                                 while (nb_used > txq->nb_tx_free) {
1090                                         if (i40e_xmit_cleanup(txq) != 0) {
1091                                                 if (nb_tx == 0)
1092                                                         return 0;
1093                                                 goto end_of_tx;
1094                                         }
1095                                 }
1096                         }
1097                 }
1098
1099                 /* Descriptor based VLAN insertion */
1100                 if (ol_flags & (PKT_TX_VLAN_PKT | PKT_TX_QINQ_PKT)) {
1101                         td_cmd |= I40E_TX_DESC_CMD_IL2TAG1;
1102                         td_tag = tx_pkt->vlan_tci;
1103                 }
1104
1105                 /* Always enable CRC offload insertion */
1106                 td_cmd |= I40E_TX_DESC_CMD_ICRC;
1107
1108                 /* Fill in tunneling parameters if necessary */
1109                 cd_tunneling_params = 0;
1110                 if (ol_flags & PKT_TX_TUNNEL_MASK)
1111                         i40e_parse_tunneling_params(ol_flags, tx_offload,
1112                                                     &cd_tunneling_params);
1113                 /* Enable checksum offloading */
1114                 if (ol_flags & I40E_TX_CKSUM_OFFLOAD_MASK)
1115                         i40e_txd_enable_checksum(ol_flags, &td_cmd,
1116                                                  &td_offset, tx_offload);
1117
1118                 if (nb_ctx) {
1119                         /* Setup TX context descriptor if required */
1120                         volatile struct i40e_tx_context_desc *ctx_txd =
1121                                 (volatile struct i40e_tx_context_desc *)\
1122                                                         &txr[tx_id];
1123                         uint16_t cd_l2tag2 = 0;
1124                         uint64_t cd_type_cmd_tso_mss =
1125                                 I40E_TX_DESC_DTYPE_CONTEXT;
1126
1127                         txn = &sw_ring[txe->next_id];
1128                         RTE_MBUF_PREFETCH_TO_FREE(txn->mbuf);
1129                         if (txe->mbuf != NULL) {
1130                                 rte_pktmbuf_free_seg(txe->mbuf);
1131                                 txe->mbuf = NULL;
1132                         }
1133
1134                         /* TSO enabled means no timestamp */
1135                         if (ol_flags & PKT_TX_TCP_SEG)
1136                                 cd_type_cmd_tso_mss |=
1137                                         i40e_set_tso_ctx(tx_pkt, tx_offload);
1138                         else {
1139 #ifdef RTE_LIBRTE_IEEE1588
1140                                 if (ol_flags & PKT_TX_IEEE1588_TMST)
1141                                         cd_type_cmd_tso_mss |=
1142                                                 ((uint64_t)I40E_TX_CTX_DESC_TSYN <<
1143                                                  I40E_TXD_CTX_QW1_CMD_SHIFT);
1144 #endif
1145                         }
1146
1147                         ctx_txd->tunneling_params =
1148                                 rte_cpu_to_le_32(cd_tunneling_params);
1149                         if (ol_flags & PKT_TX_QINQ_PKT) {
1150                                 cd_l2tag2 = tx_pkt->vlan_tci_outer;
1151                                 cd_type_cmd_tso_mss |=
1152                                         ((uint64_t)I40E_TX_CTX_DESC_IL2TAG2 <<
1153                                                 I40E_TXD_CTX_QW1_CMD_SHIFT);
1154                         }
1155                         ctx_txd->l2tag2 = rte_cpu_to_le_16(cd_l2tag2);
1156                         ctx_txd->type_cmd_tso_mss =
1157                                 rte_cpu_to_le_64(cd_type_cmd_tso_mss);
1158
1159                         PMD_TX_LOG(DEBUG, "mbuf: %p, TCD[%u]:\n"
1160                                 "tunneling_params: %#x;\n"
1161                                 "l2tag2: %#hx;\n"
1162                                 "rsvd: %#hx;\n"
1163                                 "type_cmd_tso_mss: %#"PRIx64";\n",
1164                                 tx_pkt, tx_id,
1165                                 ctx_txd->tunneling_params,
1166                                 ctx_txd->l2tag2,
1167                                 ctx_txd->rsvd,
1168                                 ctx_txd->type_cmd_tso_mss);
1169
1170                         txe->last_id = tx_last;
1171                         tx_id = txe->next_id;
1172                         txe = txn;
1173                 }
1174
1175                 m_seg = tx_pkt;
1176                 do {
1177                         txd = &txr[tx_id];
1178                         txn = &sw_ring[txe->next_id];
1179
1180                         if (txe->mbuf)
1181                                 rte_pktmbuf_free_seg(txe->mbuf);
1182                         txe->mbuf = m_seg;
1183
1184                         /* Setup TX Descriptor */
1185                         slen = m_seg->data_len;
1186                         buf_dma_addr = rte_mbuf_data_iova(m_seg);
1187
1188                         while ((ol_flags & PKT_TX_TCP_SEG) &&
1189                                 unlikely(slen > I40E_MAX_DATA_PER_TXD)) {
1190                                 txd->buffer_addr =
1191                                         rte_cpu_to_le_64(buf_dma_addr);
1192                                 txd->cmd_type_offset_bsz =
1193                                         i40e_build_ctob(td_cmd,
1194                                         td_offset, I40E_MAX_DATA_PER_TXD,
1195                                         td_tag);
1196
1197                                 buf_dma_addr += I40E_MAX_DATA_PER_TXD;
1198                                 slen -= I40E_MAX_DATA_PER_TXD;
1199
1200                                 txe->last_id = tx_last;
1201                                 tx_id = txe->next_id;
1202                                 txe = txn;
1203                                 txd = &txr[tx_id];
1204                                 txn = &sw_ring[txe->next_id];
1205                         }
1206                         PMD_TX_LOG(DEBUG, "mbuf: %p, TDD[%u]:\n"
1207                                 "buf_dma_addr: %#"PRIx64";\n"
1208                                 "td_cmd: %#x;\n"
1209                                 "td_offset: %#x;\n"
1210                                 "td_len: %u;\n"
1211                                 "td_tag: %#x;\n",
1212                                 tx_pkt, tx_id, buf_dma_addr,
1213                                 td_cmd, td_offset, slen, td_tag);
1214
1215                         txd->buffer_addr = rte_cpu_to_le_64(buf_dma_addr);
1216                         txd->cmd_type_offset_bsz = i40e_build_ctob(td_cmd,
1217                                                 td_offset, slen, td_tag);
1218                         txe->last_id = tx_last;
1219                         tx_id = txe->next_id;
1220                         txe = txn;
1221                         m_seg = m_seg->next;
1222                 } while (m_seg != NULL);
1223
1224                 /* The last packet data descriptor needs End Of Packet (EOP) */
1225                 td_cmd |= I40E_TX_DESC_CMD_EOP;
1226                 txq->nb_tx_used = (uint16_t)(txq->nb_tx_used + nb_used);
1227                 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_used);
1228
1229                 if (txq->nb_tx_used >= txq->tx_rs_thresh) {
1230                         PMD_TX_FREE_LOG(DEBUG,
1231                                         "Setting RS bit on TXD id="
1232                                         "%4u (port=%d queue=%d)",
1233                                         tx_last, txq->port_id, txq->queue_id);
1234
1235                         td_cmd |= I40E_TX_DESC_CMD_RS;
1236
1237                         /* Update txq RS bit counters */
1238                         txq->nb_tx_used = 0;
1239                 }
1240
1241                 txd->cmd_type_offset_bsz |=
1242                         rte_cpu_to_le_64(((uint64_t)td_cmd) <<
1243                                         I40E_TXD_QW1_CMD_SHIFT);
1244         }
1245
1246 end_of_tx:
1247         PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u tx_tail=%u nb_tx=%u",
1248                    (unsigned) txq->port_id, (unsigned) txq->queue_id,
1249                    (unsigned) tx_id, (unsigned) nb_tx);
1250
1251         I40E_PCI_REG_WRITE(txq->qtx_tail, tx_id);
1252         txq->tx_tail = tx_id;
1253
1254         return nb_tx;
1255 }
1256
1257 static __rte_always_inline int
1258 i40e_tx_free_bufs(struct i40e_tx_queue *txq)
1259 {
1260         struct i40e_tx_entry *txep;
1261         uint16_t i;
1262
1263         if ((txq->tx_ring[txq->tx_next_dd].cmd_type_offset_bsz &
1264                         rte_cpu_to_le_64(I40E_TXD_QW1_DTYPE_MASK)) !=
1265                         rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DESC_DONE))
1266                 return 0;
1267
1268         txep = &(txq->sw_ring[txq->tx_next_dd - (txq->tx_rs_thresh - 1)]);
1269
1270         for (i = 0; i < txq->tx_rs_thresh; i++)
1271                 rte_prefetch0((txep + i)->mbuf);
1272
1273         if (txq->offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE) {
1274                 for (i = 0; i < txq->tx_rs_thresh; ++i, ++txep) {
1275                         rte_mempool_put(txep->mbuf->pool, txep->mbuf);
1276                         txep->mbuf = NULL;
1277                 }
1278         } else {
1279                 for (i = 0; i < txq->tx_rs_thresh; ++i, ++txep) {
1280                         rte_pktmbuf_free_seg(txep->mbuf);
1281                         txep->mbuf = NULL;
1282                 }
1283         }
1284
1285         txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + txq->tx_rs_thresh);
1286         txq->tx_next_dd = (uint16_t)(txq->tx_next_dd + txq->tx_rs_thresh);
1287         if (txq->tx_next_dd >= txq->nb_tx_desc)
1288                 txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
1289
1290         return txq->tx_rs_thresh;
1291 }
1292
1293 /* Populate 4 descriptors with data from 4 mbufs */
1294 static inline void
1295 tx4(volatile struct i40e_tx_desc *txdp, struct rte_mbuf **pkts)
1296 {
1297         uint64_t dma_addr;
1298         uint32_t i;
1299
1300         for (i = 0; i < 4; i++, txdp++, pkts++) {
1301                 dma_addr = rte_mbuf_data_iova(*pkts);
1302                 txdp->buffer_addr = rte_cpu_to_le_64(dma_addr);
1303                 txdp->cmd_type_offset_bsz =
1304                         i40e_build_ctob((uint32_t)I40E_TD_CMD, 0,
1305                                         (*pkts)->data_len, 0);
1306         }
1307 }
1308
1309 /* Populate 1 descriptor with data from 1 mbuf */
1310 static inline void
1311 tx1(volatile struct i40e_tx_desc *txdp, struct rte_mbuf **pkts)
1312 {
1313         uint64_t dma_addr;
1314
1315         dma_addr = rte_mbuf_data_iova(*pkts);
1316         txdp->buffer_addr = rte_cpu_to_le_64(dma_addr);
1317         txdp->cmd_type_offset_bsz =
1318                 i40e_build_ctob((uint32_t)I40E_TD_CMD, 0,
1319                                 (*pkts)->data_len, 0);
1320 }
1321
1322 /* Fill hardware descriptor ring with mbuf data */
1323 static inline void
1324 i40e_tx_fill_hw_ring(struct i40e_tx_queue *txq,
1325                      struct rte_mbuf **pkts,
1326                      uint16_t nb_pkts)
1327 {
1328         volatile struct i40e_tx_desc *txdp = &(txq->tx_ring[txq->tx_tail]);
1329         struct i40e_tx_entry *txep = &(txq->sw_ring[txq->tx_tail]);
1330         const int N_PER_LOOP = 4;
1331         const int N_PER_LOOP_MASK = N_PER_LOOP - 1;
1332         int mainpart, leftover;
1333         int i, j;
1334
1335         mainpart = (nb_pkts & ((uint32_t) ~N_PER_LOOP_MASK));
1336         leftover = (nb_pkts & ((uint32_t)  N_PER_LOOP_MASK));
1337         for (i = 0; i < mainpart; i += N_PER_LOOP) {
1338                 for (j = 0; j < N_PER_LOOP; ++j) {
1339                         (txep + i + j)->mbuf = *(pkts + i + j);
1340                 }
1341                 tx4(txdp + i, pkts + i);
1342         }
1343         if (unlikely(leftover > 0)) {
1344                 for (i = 0; i < leftover; ++i) {
1345                         (txep + mainpart + i)->mbuf = *(pkts + mainpart + i);
1346                         tx1(txdp + mainpart + i, pkts + mainpart + i);
1347                 }
1348         }
1349 }
1350
1351 static inline uint16_t
1352 tx_xmit_pkts(struct i40e_tx_queue *txq,
1353              struct rte_mbuf **tx_pkts,
1354              uint16_t nb_pkts)
1355 {
1356         volatile struct i40e_tx_desc *txr = txq->tx_ring;
1357         uint16_t n = 0;
1358
1359         /**
1360          * Begin scanning the H/W ring for done descriptors when the number
1361          * of available descriptors drops below tx_free_thresh. For each done
1362          * descriptor, free the associated buffer.
1363          */
1364         if (txq->nb_tx_free < txq->tx_free_thresh)
1365                 i40e_tx_free_bufs(txq);
1366
1367         /* Use available descriptor only */
1368         nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
1369         if (unlikely(!nb_pkts))
1370                 return 0;
1371
1372         txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
1373         if ((txq->tx_tail + nb_pkts) > txq->nb_tx_desc) {
1374                 n = (uint16_t)(txq->nb_tx_desc - txq->tx_tail);
1375                 i40e_tx_fill_hw_ring(txq, tx_pkts, n);
1376                 txr[txq->tx_next_rs].cmd_type_offset_bsz |=
1377                         rte_cpu_to_le_64(((uint64_t)I40E_TX_DESC_CMD_RS) <<
1378                                                 I40E_TXD_QW1_CMD_SHIFT);
1379                 txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
1380                 txq->tx_tail = 0;
1381         }
1382
1383         /* Fill hardware descriptor ring with mbuf data */
1384         i40e_tx_fill_hw_ring(txq, tx_pkts + n, (uint16_t)(nb_pkts - n));
1385         txq->tx_tail = (uint16_t)(txq->tx_tail + (nb_pkts - n));
1386
1387         /* Determin if RS bit needs to be set */
1388         if (txq->tx_tail > txq->tx_next_rs) {
1389                 txr[txq->tx_next_rs].cmd_type_offset_bsz |=
1390                         rte_cpu_to_le_64(((uint64_t)I40E_TX_DESC_CMD_RS) <<
1391                                                 I40E_TXD_QW1_CMD_SHIFT);
1392                 txq->tx_next_rs =
1393                         (uint16_t)(txq->tx_next_rs + txq->tx_rs_thresh);
1394                 if (txq->tx_next_rs >= txq->nb_tx_desc)
1395                         txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
1396         }
1397
1398         if (txq->tx_tail >= txq->nb_tx_desc)
1399                 txq->tx_tail = 0;
1400
1401         /* Update the tx tail register */
1402         I40E_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
1403
1404         return nb_pkts;
1405 }
1406
1407 static uint16_t
1408 i40e_xmit_pkts_simple(void *tx_queue,
1409                       struct rte_mbuf **tx_pkts,
1410                       uint16_t nb_pkts)
1411 {
1412         uint16_t nb_tx = 0;
1413
1414         if (likely(nb_pkts <= I40E_TX_MAX_BURST))
1415                 return tx_xmit_pkts((struct i40e_tx_queue *)tx_queue,
1416                                                 tx_pkts, nb_pkts);
1417
1418         while (nb_pkts) {
1419                 uint16_t ret, num = (uint16_t)RTE_MIN(nb_pkts,
1420                                                 I40E_TX_MAX_BURST);
1421
1422                 ret = tx_xmit_pkts((struct i40e_tx_queue *)tx_queue,
1423                                                 &tx_pkts[nb_tx], num);
1424                 nb_tx = (uint16_t)(nb_tx + ret);
1425                 nb_pkts = (uint16_t)(nb_pkts - ret);
1426                 if (ret < num)
1427                         break;
1428         }
1429
1430         return nb_tx;
1431 }
1432
1433 static uint16_t
1434 i40e_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
1435                    uint16_t nb_pkts)
1436 {
1437         uint16_t nb_tx = 0;
1438         struct i40e_tx_queue *txq = (struct i40e_tx_queue *)tx_queue;
1439
1440         while (nb_pkts) {
1441                 uint16_t ret, num;
1442
1443                 num = (uint16_t)RTE_MIN(nb_pkts, txq->tx_rs_thresh);
1444                 ret = i40e_xmit_fixed_burst_vec(tx_queue, &tx_pkts[nb_tx],
1445                                                 num);
1446                 nb_tx += ret;
1447                 nb_pkts -= ret;
1448                 if (ret < num)
1449                         break;
1450         }
1451
1452         return nb_tx;
1453 }
1454
1455 /*********************************************************************
1456  *
1457  *  TX prep functions
1458  *
1459  **********************************************************************/
1460 uint16_t
1461 i40e_prep_pkts(__rte_unused void *tx_queue, struct rte_mbuf **tx_pkts,
1462                 uint16_t nb_pkts)
1463 {
1464         int i, ret;
1465         uint64_t ol_flags;
1466         struct rte_mbuf *m;
1467
1468         for (i = 0; i < nb_pkts; i++) {
1469                 m = tx_pkts[i];
1470                 ol_flags = m->ol_flags;
1471
1472                 /* Check for m->nb_segs to not exceed the limits. */
1473                 if (!(ol_flags & PKT_TX_TCP_SEG)) {
1474                         if (m->nb_segs > I40E_TX_MAX_MTU_SEG ||
1475                             m->pkt_len > I40E_FRAME_SIZE_MAX) {
1476                                 rte_errno = EINVAL;
1477                                 return i;
1478                         }
1479                 } else if (m->nb_segs > I40E_TX_MAX_SEG ||
1480                            m->tso_segsz < I40E_MIN_TSO_MSS ||
1481                            m->tso_segsz > I40E_MAX_TSO_MSS ||
1482                            m->pkt_len > I40E_TSO_FRAME_SIZE_MAX) {
1483                         /* MSS outside the range (256B - 9674B) are considered
1484                          * malicious
1485                          */
1486                         rte_errno = EINVAL;
1487                         return i;
1488                 }
1489
1490                 if (ol_flags & I40E_TX_OFFLOAD_NOTSUP_MASK) {
1491                         rte_errno = ENOTSUP;
1492                         return i;
1493                 }
1494
1495                 /* check the size of packet */
1496                 if (m->pkt_len < I40E_TX_MIN_PKT_LEN) {
1497                         rte_errno = EINVAL;
1498                         return i;
1499                 }
1500
1501 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
1502                 ret = rte_validate_tx_offload(m);
1503                 if (ret != 0) {
1504                         rte_errno = -ret;
1505                         return i;
1506                 }
1507 #endif
1508                 ret = rte_net_intel_cksum_prepare(m);
1509                 if (ret != 0) {
1510                         rte_errno = -ret;
1511                         return i;
1512                 }
1513         }
1514         return i;
1515 }
1516
1517 /*
1518  * Find the VSI the queue belongs to. 'queue_idx' is the queue index
1519  * application used, which assume having sequential ones. But from driver's
1520  * perspective, it's different. For example, q0 belongs to FDIR VSI, q1-q64
1521  * to MAIN VSI, , q65-96 to SRIOV VSIs, q97-128 to VMDQ VSIs. For application
1522  * running on host, q1-64 and q97-128 can be used, total 96 queues. They can
1523  * use queue_idx from 0 to 95 to access queues, while real queue would be
1524  * different. This function will do a queue mapping to find VSI the queue
1525  * belongs to.
1526  */
1527 static struct i40e_vsi*
1528 i40e_pf_get_vsi_by_qindex(struct i40e_pf *pf, uint16_t queue_idx)
1529 {
1530         /* the queue in MAIN VSI range */
1531         if (queue_idx < pf->main_vsi->nb_qps)
1532                 return pf->main_vsi;
1533
1534         queue_idx -= pf->main_vsi->nb_qps;
1535
1536         /* queue_idx is greater than VMDQ VSIs range */
1537         if (queue_idx > pf->nb_cfg_vmdq_vsi * pf->vmdq_nb_qps - 1) {
1538                 PMD_INIT_LOG(ERR, "queue_idx out of range. VMDQ configured?");
1539                 return NULL;
1540         }
1541
1542         return pf->vmdq[queue_idx / pf->vmdq_nb_qps].vsi;
1543 }
1544
1545 static uint16_t
1546 i40e_get_queue_offset_by_qindex(struct i40e_pf *pf, uint16_t queue_idx)
1547 {
1548         /* the queue in MAIN VSI range */
1549         if (queue_idx < pf->main_vsi->nb_qps)
1550                 return queue_idx;
1551
1552         /* It's VMDQ queues */
1553         queue_idx -= pf->main_vsi->nb_qps;
1554
1555         if (pf->nb_cfg_vmdq_vsi)
1556                 return queue_idx % pf->vmdq_nb_qps;
1557         else {
1558                 PMD_INIT_LOG(ERR, "Fail to get queue offset");
1559                 return (uint16_t)(-1);
1560         }
1561 }
1562
1563 int
1564 i40e_dev_rx_queue_start(struct rte_eth_dev *dev, uint16_t rx_queue_id)
1565 {
1566         struct i40e_rx_queue *rxq;
1567         int err;
1568         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1569
1570         PMD_INIT_FUNC_TRACE();
1571
1572         rxq = dev->data->rx_queues[rx_queue_id];
1573
1574         err = i40e_alloc_rx_queue_mbufs(rxq);
1575         if (err) {
1576                 PMD_DRV_LOG(ERR, "Failed to allocate RX queue mbuf");
1577                 return err;
1578         }
1579
1580         /* Init the RX tail regieter. */
1581         I40E_PCI_REG_WRITE(rxq->qrx_tail, rxq->nb_rx_desc - 1);
1582
1583         err = i40e_switch_rx_queue(hw, rxq->reg_idx, TRUE);
1584         if (err) {
1585                 PMD_DRV_LOG(ERR, "Failed to switch RX queue %u on",
1586                             rx_queue_id);
1587
1588                 i40e_rx_queue_release_mbufs(rxq);
1589                 i40e_reset_rx_queue(rxq);
1590                 return err;
1591         }
1592         dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
1593
1594         return 0;
1595 }
1596
1597 int
1598 i40e_dev_rx_queue_stop(struct rte_eth_dev *dev, uint16_t rx_queue_id)
1599 {
1600         struct i40e_rx_queue *rxq;
1601         int err;
1602         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1603
1604         rxq = dev->data->rx_queues[rx_queue_id];
1605
1606         /*
1607          * rx_queue_id is queue id application refers to, while
1608          * rxq->reg_idx is the real queue index.
1609          */
1610         err = i40e_switch_rx_queue(hw, rxq->reg_idx, FALSE);
1611         if (err) {
1612                 PMD_DRV_LOG(ERR, "Failed to switch RX queue %u off",
1613                             rx_queue_id);
1614                 return err;
1615         }
1616         i40e_rx_queue_release_mbufs(rxq);
1617         i40e_reset_rx_queue(rxq);
1618         dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
1619
1620         return 0;
1621 }
1622
1623 int
1624 i40e_dev_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id)
1625 {
1626         int err;
1627         struct i40e_tx_queue *txq;
1628         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1629
1630         PMD_INIT_FUNC_TRACE();
1631
1632         txq = dev->data->tx_queues[tx_queue_id];
1633
1634         /*
1635          * tx_queue_id is queue id application refers to, while
1636          * rxq->reg_idx is the real queue index.
1637          */
1638         err = i40e_switch_tx_queue(hw, txq->reg_idx, TRUE);
1639         if (err) {
1640                 PMD_DRV_LOG(ERR, "Failed to switch TX queue %u on",
1641                             tx_queue_id);
1642                 return err;
1643         }
1644         dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
1645
1646         return 0;
1647 }
1648
1649 int
1650 i40e_dev_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id)
1651 {
1652         struct i40e_tx_queue *txq;
1653         int err;
1654         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1655
1656         txq = dev->data->tx_queues[tx_queue_id];
1657
1658         /*
1659          * tx_queue_id is queue id application refers to, while
1660          * txq->reg_idx is the real queue index.
1661          */
1662         err = i40e_switch_tx_queue(hw, txq->reg_idx, FALSE);
1663         if (err) {
1664                 PMD_DRV_LOG(ERR, "Failed to switch TX queue %u of",
1665                             tx_queue_id);
1666                 return err;
1667         }
1668
1669         i40e_tx_queue_release_mbufs(txq);
1670         i40e_reset_tx_queue(txq);
1671         dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
1672
1673         return 0;
1674 }
1675
1676 const uint32_t *
1677 i40e_dev_supported_ptypes_get(struct rte_eth_dev *dev)
1678 {
1679         static const uint32_t ptypes[] = {
1680                 /* refers to i40e_rxd_pkt_type_mapping() */
1681                 RTE_PTYPE_L2_ETHER,
1682                 RTE_PTYPE_L2_ETHER_TIMESYNC,
1683                 RTE_PTYPE_L2_ETHER_LLDP,
1684                 RTE_PTYPE_L2_ETHER_ARP,
1685                 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
1686                 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
1687                 RTE_PTYPE_L4_FRAG,
1688                 RTE_PTYPE_L4_ICMP,
1689                 RTE_PTYPE_L4_NONFRAG,
1690                 RTE_PTYPE_L4_SCTP,
1691                 RTE_PTYPE_L4_TCP,
1692                 RTE_PTYPE_L4_UDP,
1693                 RTE_PTYPE_TUNNEL_GRENAT,
1694                 RTE_PTYPE_TUNNEL_IP,
1695                 RTE_PTYPE_INNER_L2_ETHER,
1696                 RTE_PTYPE_INNER_L2_ETHER_VLAN,
1697                 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
1698                 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
1699                 RTE_PTYPE_INNER_L4_FRAG,
1700                 RTE_PTYPE_INNER_L4_ICMP,
1701                 RTE_PTYPE_INNER_L4_NONFRAG,
1702                 RTE_PTYPE_INNER_L4_SCTP,
1703                 RTE_PTYPE_INNER_L4_TCP,
1704                 RTE_PTYPE_INNER_L4_UDP,
1705                 RTE_PTYPE_UNKNOWN
1706         };
1707
1708         if (dev->rx_pkt_burst == i40e_recv_pkts ||
1709 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
1710             dev->rx_pkt_burst == i40e_recv_pkts_bulk_alloc ||
1711 #endif
1712             dev->rx_pkt_burst == i40e_recv_scattered_pkts ||
1713             dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec ||
1714             dev->rx_pkt_burst == i40e_recv_pkts_vec ||
1715             dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec_avx2 ||
1716             dev->rx_pkt_burst == i40e_recv_pkts_vec_avx2)
1717                 return ptypes;
1718         return NULL;
1719 }
1720
1721 static int
1722 i40e_dev_first_queue(uint16_t idx, void **queues, int num)
1723 {
1724         uint16_t i;
1725
1726         for (i = 0; i < num; i++) {
1727                 if (i != idx && queues[i])
1728                         return 0;
1729         }
1730
1731         return 1;
1732 }
1733
1734 static int
1735 i40e_dev_rx_queue_setup_runtime(struct rte_eth_dev *dev,
1736                                 struct i40e_rx_queue *rxq)
1737 {
1738         struct i40e_adapter *ad =
1739                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
1740         int use_def_burst_func =
1741                 check_rx_burst_bulk_alloc_preconditions(rxq);
1742         uint16_t buf_size =
1743                 (uint16_t)(rte_pktmbuf_data_room_size(rxq->mp) -
1744                            RTE_PKTMBUF_HEADROOM);
1745         int use_scattered_rx =
1746                 (rxq->max_pkt_len > buf_size);
1747
1748         if (i40e_rx_queue_init(rxq) != I40E_SUCCESS) {
1749                 PMD_DRV_LOG(ERR,
1750                             "Failed to do RX queue initialization");
1751                 return -EINVAL;
1752         }
1753
1754         if (i40e_dev_first_queue(rxq->queue_id,
1755                                  dev->data->rx_queues,
1756                                  dev->data->nb_rx_queues)) {
1757                 /**
1758                  * If it is the first queue to setup,
1759                  * set all flags to default and call
1760                  * i40e_set_rx_function.
1761                  */
1762                 ad->rx_bulk_alloc_allowed = true;
1763                 ad->rx_vec_allowed = true;
1764                 dev->data->scattered_rx = use_scattered_rx;
1765                 if (use_def_burst_func)
1766                         ad->rx_bulk_alloc_allowed = false;
1767                 i40e_set_rx_function(dev);
1768                 return 0;
1769         } else if (ad->rx_vec_allowed && !rte_is_power_of_2(rxq->nb_rx_desc)) {
1770                 PMD_DRV_LOG(ERR, "Vector mode is allowed, but descriptor"
1771                             " number %d of queue %d isn't power of 2",
1772                             rxq->nb_rx_desc, rxq->queue_id);
1773                 return -EINVAL;
1774         }
1775
1776         /* check bulk alloc conflict */
1777         if (ad->rx_bulk_alloc_allowed && use_def_burst_func) {
1778                 PMD_DRV_LOG(ERR, "Can't use default burst.");
1779                 return -EINVAL;
1780         }
1781         /* check scatterred conflict */
1782         if (!dev->data->scattered_rx && use_scattered_rx) {
1783                 PMD_DRV_LOG(ERR, "Scattered rx is required.");
1784                 return -EINVAL;
1785         }
1786         /* check vector conflict */
1787         if (ad->rx_vec_allowed && i40e_rxq_vec_setup(rxq)) {
1788                 PMD_DRV_LOG(ERR, "Failed vector rx setup.");
1789                 return -EINVAL;
1790         }
1791
1792         return 0;
1793 }
1794
1795 int
1796 i40e_dev_rx_queue_setup(struct rte_eth_dev *dev,
1797                         uint16_t queue_idx,
1798                         uint16_t nb_desc,
1799                         unsigned int socket_id,
1800                         const struct rte_eth_rxconf *rx_conf,
1801                         struct rte_mempool *mp)
1802 {
1803         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1804         struct i40e_adapter *ad =
1805                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
1806         struct i40e_vsi *vsi;
1807         struct i40e_pf *pf = NULL;
1808         struct i40e_vf *vf = NULL;
1809         struct i40e_rx_queue *rxq;
1810         const struct rte_memzone *rz;
1811         uint32_t ring_size;
1812         uint16_t len, i;
1813         uint16_t reg_idx, base, bsf, tc_mapping;
1814         int q_offset, use_def_burst_func = 1;
1815         uint64_t offloads;
1816
1817         offloads = rx_conf->offloads | dev->data->dev_conf.rxmode.offloads;
1818
1819         if (hw->mac.type == I40E_MAC_VF || hw->mac.type == I40E_MAC_X722_VF) {
1820                 vf = I40EVF_DEV_PRIVATE_TO_VF(dev->data->dev_private);
1821                 vsi = &vf->vsi;
1822                 if (!vsi)
1823                         return -EINVAL;
1824                 reg_idx = queue_idx;
1825         } else {
1826                 pf = I40E_DEV_PRIVATE_TO_PF(dev->data->dev_private);
1827                 vsi = i40e_pf_get_vsi_by_qindex(pf, queue_idx);
1828                 if (!vsi)
1829                         return -EINVAL;
1830                 q_offset = i40e_get_queue_offset_by_qindex(pf, queue_idx);
1831                 if (q_offset < 0)
1832                         return -EINVAL;
1833                 reg_idx = vsi->base_queue + q_offset;
1834         }
1835
1836         if (nb_desc % I40E_ALIGN_RING_DESC != 0 ||
1837             (nb_desc > I40E_MAX_RING_DESC) ||
1838             (nb_desc < I40E_MIN_RING_DESC)) {
1839                 PMD_DRV_LOG(ERR, "Number (%u) of receive descriptors is "
1840                             "invalid", nb_desc);
1841                 return -EINVAL;
1842         }
1843
1844         /* Free memory if needed */
1845         if (dev->data->rx_queues[queue_idx]) {
1846                 i40e_dev_rx_queue_release(dev->data->rx_queues[queue_idx]);
1847                 dev->data->rx_queues[queue_idx] = NULL;
1848         }
1849
1850         /* Allocate the rx queue data structure */
1851         rxq = rte_zmalloc_socket("i40e rx queue",
1852                                  sizeof(struct i40e_rx_queue),
1853                                  RTE_CACHE_LINE_SIZE,
1854                                  socket_id);
1855         if (!rxq) {
1856                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
1857                             "rx queue data structure");
1858                 return -ENOMEM;
1859         }
1860         rxq->mp = mp;
1861         rxq->nb_rx_desc = nb_desc;
1862         rxq->rx_free_thresh = rx_conf->rx_free_thresh;
1863         rxq->queue_id = queue_idx;
1864         rxq->reg_idx = reg_idx;
1865         rxq->port_id = dev->data->port_id;
1866         if (dev->data->dev_conf.rxmode.offloads & DEV_RX_OFFLOAD_KEEP_CRC)
1867                 rxq->crc_len = RTE_ETHER_CRC_LEN;
1868         else
1869                 rxq->crc_len = 0;
1870         rxq->drop_en = rx_conf->rx_drop_en;
1871         rxq->vsi = vsi;
1872         rxq->rx_deferred_start = rx_conf->rx_deferred_start;
1873         rxq->offloads = offloads;
1874
1875         /* Allocate the maximun number of RX ring hardware descriptor. */
1876         len = I40E_MAX_RING_DESC;
1877
1878         /**
1879          * Allocating a little more memory because vectorized/bulk_alloc Rx
1880          * functions doesn't check boundaries each time.
1881          */
1882         len += RTE_PMD_I40E_RX_MAX_BURST;
1883
1884         ring_size = RTE_ALIGN(len * sizeof(union i40e_rx_desc),
1885                               I40E_DMA_MEM_ALIGN);
1886
1887         rz = rte_eth_dma_zone_reserve(dev, "rx_ring", queue_idx,
1888                               ring_size, I40E_RING_BASE_ALIGN, socket_id);
1889         if (!rz) {
1890                 i40e_dev_rx_queue_release(rxq);
1891                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for RX");
1892                 return -ENOMEM;
1893         }
1894
1895         /* Zero all the descriptors in the ring. */
1896         memset(rz->addr, 0, ring_size);
1897
1898         rxq->rx_ring_phys_addr = rz->iova;
1899         rxq->rx_ring = (union i40e_rx_desc *)rz->addr;
1900
1901         len = (uint16_t)(nb_desc + RTE_PMD_I40E_RX_MAX_BURST);
1902
1903         /* Allocate the software ring. */
1904         rxq->sw_ring =
1905                 rte_zmalloc_socket("i40e rx sw ring",
1906                                    sizeof(struct i40e_rx_entry) * len,
1907                                    RTE_CACHE_LINE_SIZE,
1908                                    socket_id);
1909         if (!rxq->sw_ring) {
1910                 i40e_dev_rx_queue_release(rxq);
1911                 PMD_DRV_LOG(ERR, "Failed to allocate memory for SW ring");
1912                 return -ENOMEM;
1913         }
1914
1915         i40e_reset_rx_queue(rxq);
1916         rxq->q_set = TRUE;
1917
1918         for (i = 0; i < I40E_MAX_TRAFFIC_CLASS; i++) {
1919                 if (!(vsi->enabled_tc & (1 << i)))
1920                         continue;
1921                 tc_mapping = rte_le_to_cpu_16(vsi->info.tc_mapping[i]);
1922                 base = (tc_mapping & I40E_AQ_VSI_TC_QUE_OFFSET_MASK) >>
1923                         I40E_AQ_VSI_TC_QUE_OFFSET_SHIFT;
1924                 bsf = (tc_mapping & I40E_AQ_VSI_TC_QUE_NUMBER_MASK) >>
1925                         I40E_AQ_VSI_TC_QUE_NUMBER_SHIFT;
1926
1927                 if (queue_idx >= base && queue_idx < (base + BIT(bsf)))
1928                         rxq->dcb_tc = i;
1929         }
1930
1931         if (dev->data->dev_started) {
1932                 if (i40e_dev_rx_queue_setup_runtime(dev, rxq)) {
1933                         i40e_dev_rx_queue_release(rxq);
1934                         return -EINVAL;
1935                 }
1936         } else {
1937                 use_def_burst_func =
1938                         check_rx_burst_bulk_alloc_preconditions(rxq);
1939                 if (!use_def_burst_func) {
1940 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
1941                         PMD_INIT_LOG(DEBUG,
1942                           "Rx Burst Bulk Alloc Preconditions are "
1943                           "satisfied. Rx Burst Bulk Alloc function will be "
1944                           "used on port=%d, queue=%d.",
1945                           rxq->port_id, rxq->queue_id);
1946 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
1947                 } else {
1948                         PMD_INIT_LOG(DEBUG,
1949                           "Rx Burst Bulk Alloc Preconditions are "
1950                           "not satisfied, Scattered Rx is requested, "
1951                           "or RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC is "
1952                           "not enabled on port=%d, queue=%d.",
1953                           rxq->port_id, rxq->queue_id);
1954                         ad->rx_bulk_alloc_allowed = false;
1955                 }
1956         }
1957
1958         dev->data->rx_queues[queue_idx] = rxq;
1959         return 0;
1960 }
1961
1962 void
1963 i40e_dev_rx_queue_release(void *rxq)
1964 {
1965         struct i40e_rx_queue *q = (struct i40e_rx_queue *)rxq;
1966
1967         if (!q) {
1968                 PMD_DRV_LOG(DEBUG, "Pointer to rxq is NULL");
1969                 return;
1970         }
1971
1972         i40e_rx_queue_release_mbufs(q);
1973         rte_free(q->sw_ring);
1974         rte_free(q);
1975 }
1976
1977 uint32_t
1978 i40e_dev_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id)
1979 {
1980 #define I40E_RXQ_SCAN_INTERVAL 4
1981         volatile union i40e_rx_desc *rxdp;
1982         struct i40e_rx_queue *rxq;
1983         uint16_t desc = 0;
1984
1985         rxq = dev->data->rx_queues[rx_queue_id];
1986         rxdp = &(rxq->rx_ring[rxq->rx_tail]);
1987         while ((desc < rxq->nb_rx_desc) &&
1988                 ((rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len) &
1989                 I40E_RXD_QW1_STATUS_MASK) >> I40E_RXD_QW1_STATUS_SHIFT) &
1990                                 (1 << I40E_RX_DESC_STATUS_DD_SHIFT)) {
1991                 /**
1992                  * Check the DD bit of a rx descriptor of each 4 in a group,
1993                  * to avoid checking too frequently and downgrading performance
1994                  * too much.
1995                  */
1996                 desc += I40E_RXQ_SCAN_INTERVAL;
1997                 rxdp += I40E_RXQ_SCAN_INTERVAL;
1998                 if (rxq->rx_tail + desc >= rxq->nb_rx_desc)
1999                         rxdp = &(rxq->rx_ring[rxq->rx_tail +
2000                                         desc - rxq->nb_rx_desc]);
2001         }
2002
2003         return desc;
2004 }
2005
2006 int
2007 i40e_dev_rx_descriptor_done(void *rx_queue, uint16_t offset)
2008 {
2009         volatile union i40e_rx_desc *rxdp;
2010         struct i40e_rx_queue *rxq = rx_queue;
2011         uint16_t desc;
2012         int ret;
2013
2014         if (unlikely(offset >= rxq->nb_rx_desc)) {
2015                 PMD_DRV_LOG(ERR, "Invalid RX descriptor id %u", offset);
2016                 return 0;
2017         }
2018
2019         desc = rxq->rx_tail + offset;
2020         if (desc >= rxq->nb_rx_desc)
2021                 desc -= rxq->nb_rx_desc;
2022
2023         rxdp = &(rxq->rx_ring[desc]);
2024
2025         ret = !!(((rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len) &
2026                 I40E_RXD_QW1_STATUS_MASK) >> I40E_RXD_QW1_STATUS_SHIFT) &
2027                                 (1 << I40E_RX_DESC_STATUS_DD_SHIFT));
2028
2029         return ret;
2030 }
2031
2032 int
2033 i40e_dev_rx_descriptor_status(void *rx_queue, uint16_t offset)
2034 {
2035         struct i40e_rx_queue *rxq = rx_queue;
2036         volatile uint64_t *status;
2037         uint64_t mask;
2038         uint32_t desc;
2039
2040         if (unlikely(offset >= rxq->nb_rx_desc))
2041                 return -EINVAL;
2042
2043         if (offset >= rxq->nb_rx_desc - rxq->nb_rx_hold)
2044                 return RTE_ETH_RX_DESC_UNAVAIL;
2045
2046         desc = rxq->rx_tail + offset;
2047         if (desc >= rxq->nb_rx_desc)
2048                 desc -= rxq->nb_rx_desc;
2049
2050         status = &rxq->rx_ring[desc].wb.qword1.status_error_len;
2051         mask = rte_le_to_cpu_64((1ULL << I40E_RX_DESC_STATUS_DD_SHIFT)
2052                 << I40E_RXD_QW1_STATUS_SHIFT);
2053         if (*status & mask)
2054                 return RTE_ETH_RX_DESC_DONE;
2055
2056         return RTE_ETH_RX_DESC_AVAIL;
2057 }
2058
2059 int
2060 i40e_dev_tx_descriptor_status(void *tx_queue, uint16_t offset)
2061 {
2062         struct i40e_tx_queue *txq = tx_queue;
2063         volatile uint64_t *status;
2064         uint64_t mask, expect;
2065         uint32_t desc;
2066
2067         if (unlikely(offset >= txq->nb_tx_desc))
2068                 return -EINVAL;
2069
2070         desc = txq->tx_tail + offset;
2071         /* go to next desc that has the RS bit */
2072         desc = ((desc + txq->tx_rs_thresh - 1) / txq->tx_rs_thresh) *
2073                 txq->tx_rs_thresh;
2074         if (desc >= txq->nb_tx_desc) {
2075                 desc -= txq->nb_tx_desc;
2076                 if (desc >= txq->nb_tx_desc)
2077                         desc -= txq->nb_tx_desc;
2078         }
2079
2080         status = &txq->tx_ring[desc].cmd_type_offset_bsz;
2081         mask = rte_le_to_cpu_64(I40E_TXD_QW1_DTYPE_MASK);
2082         expect = rte_cpu_to_le_64(
2083                 I40E_TX_DESC_DTYPE_DESC_DONE << I40E_TXD_QW1_DTYPE_SHIFT);
2084         if ((*status & mask) == expect)
2085                 return RTE_ETH_TX_DESC_DONE;
2086
2087         return RTE_ETH_TX_DESC_FULL;
2088 }
2089
2090 static int
2091 i40e_dev_tx_queue_setup_runtime(struct rte_eth_dev *dev,
2092                                 struct i40e_tx_queue *txq)
2093 {
2094         struct i40e_adapter *ad =
2095                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
2096
2097         if (i40e_tx_queue_init(txq) != I40E_SUCCESS) {
2098                 PMD_DRV_LOG(ERR,
2099                             "Failed to do TX queue initialization");
2100                 return -EINVAL;
2101         }
2102
2103         if (i40e_dev_first_queue(txq->queue_id,
2104                                  dev->data->tx_queues,
2105                                  dev->data->nb_tx_queues)) {
2106                 /**
2107                  * If it is the first queue to setup,
2108                  * set all flags and call
2109                  * i40e_set_tx_function.
2110                  */
2111                 i40e_set_tx_function_flag(dev, txq);
2112                 i40e_set_tx_function(dev);
2113                 return 0;
2114         }
2115
2116         /* check vector conflict */
2117         if (ad->tx_vec_allowed) {
2118                 if (txq->tx_rs_thresh > RTE_I40E_TX_MAX_FREE_BUF_SZ ||
2119                     i40e_txq_vec_setup(txq)) {
2120                         PMD_DRV_LOG(ERR, "Failed vector tx setup.");
2121                         return -EINVAL;
2122                 }
2123         }
2124         /* check simple tx conflict */
2125         if (ad->tx_simple_allowed) {
2126                 if ((txq->offloads & ~DEV_TX_OFFLOAD_MBUF_FAST_FREE) != 0 ||
2127                                 txq->tx_rs_thresh < RTE_PMD_I40E_TX_MAX_BURST) {
2128                         PMD_DRV_LOG(ERR, "No-simple tx is required.");
2129                         return -EINVAL;
2130                 }
2131         }
2132
2133         return 0;
2134 }
2135
2136 int
2137 i40e_dev_tx_queue_setup(struct rte_eth_dev *dev,
2138                         uint16_t queue_idx,
2139                         uint16_t nb_desc,
2140                         unsigned int socket_id,
2141                         const struct rte_eth_txconf *tx_conf)
2142 {
2143         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2144         struct i40e_vsi *vsi;
2145         struct i40e_pf *pf = NULL;
2146         struct i40e_vf *vf = NULL;
2147         struct i40e_tx_queue *txq;
2148         const struct rte_memzone *tz;
2149         uint32_t ring_size;
2150         uint16_t tx_rs_thresh, tx_free_thresh;
2151         uint16_t reg_idx, i, base, bsf, tc_mapping;
2152         int q_offset;
2153         uint64_t offloads;
2154
2155         offloads = tx_conf->offloads | dev->data->dev_conf.txmode.offloads;
2156
2157         if (hw->mac.type == I40E_MAC_VF || hw->mac.type == I40E_MAC_X722_VF) {
2158                 vf = I40EVF_DEV_PRIVATE_TO_VF(dev->data->dev_private);
2159                 vsi = &vf->vsi;
2160                 if (!vsi)
2161                         return -EINVAL;
2162                 reg_idx = queue_idx;
2163         } else {
2164                 pf = I40E_DEV_PRIVATE_TO_PF(dev->data->dev_private);
2165                 vsi = i40e_pf_get_vsi_by_qindex(pf, queue_idx);
2166                 if (!vsi)
2167                         return -EINVAL;
2168                 q_offset = i40e_get_queue_offset_by_qindex(pf, queue_idx);
2169                 if (q_offset < 0)
2170                         return -EINVAL;
2171                 reg_idx = vsi->base_queue + q_offset;
2172         }
2173
2174         if (nb_desc % I40E_ALIGN_RING_DESC != 0 ||
2175             (nb_desc > I40E_MAX_RING_DESC) ||
2176             (nb_desc < I40E_MIN_RING_DESC)) {
2177                 PMD_DRV_LOG(ERR, "Number (%u) of transmit descriptors is "
2178                             "invalid", nb_desc);
2179                 return -EINVAL;
2180         }
2181
2182         /**
2183          * The following two parameters control the setting of the RS bit on
2184          * transmit descriptors. TX descriptors will have their RS bit set
2185          * after txq->tx_rs_thresh descriptors have been used. The TX
2186          * descriptor ring will be cleaned after txq->tx_free_thresh
2187          * descriptors are used or if the number of descriptors required to
2188          * transmit a packet is greater than the number of free TX descriptors.
2189          *
2190          * The following constraints must be satisfied:
2191          *  - tx_rs_thresh must be greater than 0.
2192          *  - tx_rs_thresh must be less than the size of the ring minus 2.
2193          *  - tx_rs_thresh must be less than or equal to tx_free_thresh.
2194          *  - tx_rs_thresh must be a divisor of the ring size.
2195          *  - tx_free_thresh must be greater than 0.
2196          *  - tx_free_thresh must be less than the size of the ring minus 3.
2197          *  - tx_free_thresh + tx_rs_thresh must not exceed nb_desc.
2198          *
2199          * One descriptor in the TX ring is used as a sentinel to avoid a H/W
2200          * race condition, hence the maximum threshold constraints. When set
2201          * to zero use default values.
2202          */
2203         tx_free_thresh = (uint16_t)((tx_conf->tx_free_thresh) ?
2204                 tx_conf->tx_free_thresh : DEFAULT_TX_FREE_THRESH);
2205         /* force tx_rs_thresh to adapt an aggresive tx_free_thresh */
2206         tx_rs_thresh = (DEFAULT_TX_RS_THRESH + tx_free_thresh > nb_desc) ?
2207                 nb_desc - tx_free_thresh : DEFAULT_TX_RS_THRESH;
2208         if (tx_conf->tx_rs_thresh > 0)
2209                 tx_rs_thresh = tx_conf->tx_rs_thresh;
2210         if (tx_rs_thresh + tx_free_thresh > nb_desc) {
2211                 PMD_INIT_LOG(ERR, "tx_rs_thresh + tx_free_thresh must not "
2212                                 "exceed nb_desc. (tx_rs_thresh=%u "
2213                                 "tx_free_thresh=%u nb_desc=%u port=%d queue=%d)",
2214                                 (unsigned int)tx_rs_thresh,
2215                                 (unsigned int)tx_free_thresh,
2216                                 (unsigned int)nb_desc,
2217                                 (int)dev->data->port_id,
2218                                 (int)queue_idx);
2219                 return I40E_ERR_PARAM;
2220         }
2221         if (tx_rs_thresh >= (nb_desc - 2)) {
2222                 PMD_INIT_LOG(ERR, "tx_rs_thresh must be less than the "
2223                              "number of TX descriptors minus 2. "
2224                              "(tx_rs_thresh=%u port=%d queue=%d)",
2225                              (unsigned int)tx_rs_thresh,
2226                              (int)dev->data->port_id,
2227                              (int)queue_idx);
2228                 return I40E_ERR_PARAM;
2229         }
2230         if (tx_free_thresh >= (nb_desc - 3)) {
2231                 PMD_INIT_LOG(ERR, "tx_free_thresh must be less than the "
2232                              "number of TX descriptors minus 3. "
2233                              "(tx_free_thresh=%u port=%d queue=%d)",
2234                              (unsigned int)tx_free_thresh,
2235                              (int)dev->data->port_id,
2236                              (int)queue_idx);
2237                 return I40E_ERR_PARAM;
2238         }
2239         if (tx_rs_thresh > tx_free_thresh) {
2240                 PMD_INIT_LOG(ERR, "tx_rs_thresh must be less than or "
2241                              "equal to tx_free_thresh. (tx_free_thresh=%u"
2242                              " tx_rs_thresh=%u port=%d queue=%d)",
2243                              (unsigned int)tx_free_thresh,
2244                              (unsigned int)tx_rs_thresh,
2245                              (int)dev->data->port_id,
2246                              (int)queue_idx);
2247                 return I40E_ERR_PARAM;
2248         }
2249         if ((nb_desc % tx_rs_thresh) != 0) {
2250                 PMD_INIT_LOG(ERR, "tx_rs_thresh must be a divisor of the "
2251                              "number of TX descriptors. (tx_rs_thresh=%u"
2252                              " port=%d queue=%d)",
2253                              (unsigned int)tx_rs_thresh,
2254                              (int)dev->data->port_id,
2255                              (int)queue_idx);
2256                 return I40E_ERR_PARAM;
2257         }
2258         if ((tx_rs_thresh > 1) && (tx_conf->tx_thresh.wthresh != 0)) {
2259                 PMD_INIT_LOG(ERR, "TX WTHRESH must be set to 0 if "
2260                              "tx_rs_thresh is greater than 1. "
2261                              "(tx_rs_thresh=%u port=%d queue=%d)",
2262                              (unsigned int)tx_rs_thresh,
2263                              (int)dev->data->port_id,
2264                              (int)queue_idx);
2265                 return I40E_ERR_PARAM;
2266         }
2267
2268         /* Free memory if needed. */
2269         if (dev->data->tx_queues[queue_idx]) {
2270                 i40e_dev_tx_queue_release(dev->data->tx_queues[queue_idx]);
2271                 dev->data->tx_queues[queue_idx] = NULL;
2272         }
2273
2274         /* Allocate the TX queue data structure. */
2275         txq = rte_zmalloc_socket("i40e tx queue",
2276                                   sizeof(struct i40e_tx_queue),
2277                                   RTE_CACHE_LINE_SIZE,
2278                                   socket_id);
2279         if (!txq) {
2280                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
2281                             "tx queue structure");
2282                 return -ENOMEM;
2283         }
2284
2285         /* Allocate TX hardware ring descriptors. */
2286         ring_size = sizeof(struct i40e_tx_desc) * I40E_MAX_RING_DESC;
2287         ring_size = RTE_ALIGN(ring_size, I40E_DMA_MEM_ALIGN);
2288         tz = rte_eth_dma_zone_reserve(dev, "tx_ring", queue_idx,
2289                               ring_size, I40E_RING_BASE_ALIGN, socket_id);
2290         if (!tz) {
2291                 i40e_dev_tx_queue_release(txq);
2292                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for TX");
2293                 return -ENOMEM;
2294         }
2295
2296         txq->nb_tx_desc = nb_desc;
2297         txq->tx_rs_thresh = tx_rs_thresh;
2298         txq->tx_free_thresh = tx_free_thresh;
2299         txq->pthresh = tx_conf->tx_thresh.pthresh;
2300         txq->hthresh = tx_conf->tx_thresh.hthresh;
2301         txq->wthresh = tx_conf->tx_thresh.wthresh;
2302         txq->queue_id = queue_idx;
2303         txq->reg_idx = reg_idx;
2304         txq->port_id = dev->data->port_id;
2305         txq->offloads = offloads;
2306         txq->vsi = vsi;
2307         txq->tx_deferred_start = tx_conf->tx_deferred_start;
2308
2309         txq->tx_ring_phys_addr = tz->iova;
2310         txq->tx_ring = (struct i40e_tx_desc *)tz->addr;
2311
2312         /* Allocate software ring */
2313         txq->sw_ring =
2314                 rte_zmalloc_socket("i40e tx sw ring",
2315                                    sizeof(struct i40e_tx_entry) * nb_desc,
2316                                    RTE_CACHE_LINE_SIZE,
2317                                    socket_id);
2318         if (!txq->sw_ring) {
2319                 i40e_dev_tx_queue_release(txq);
2320                 PMD_DRV_LOG(ERR, "Failed to allocate memory for SW TX ring");
2321                 return -ENOMEM;
2322         }
2323
2324         i40e_reset_tx_queue(txq);
2325         txq->q_set = TRUE;
2326
2327         for (i = 0; i < I40E_MAX_TRAFFIC_CLASS; i++) {
2328                 if (!(vsi->enabled_tc & (1 << i)))
2329                         continue;
2330                 tc_mapping = rte_le_to_cpu_16(vsi->info.tc_mapping[i]);
2331                 base = (tc_mapping & I40E_AQ_VSI_TC_QUE_OFFSET_MASK) >>
2332                         I40E_AQ_VSI_TC_QUE_OFFSET_SHIFT;
2333                 bsf = (tc_mapping & I40E_AQ_VSI_TC_QUE_NUMBER_MASK) >>
2334                         I40E_AQ_VSI_TC_QUE_NUMBER_SHIFT;
2335
2336                 if (queue_idx >= base && queue_idx < (base + BIT(bsf)))
2337                         txq->dcb_tc = i;
2338         }
2339
2340         if (dev->data->dev_started) {
2341                 if (i40e_dev_tx_queue_setup_runtime(dev, txq)) {
2342                         i40e_dev_tx_queue_release(txq);
2343                         return -EINVAL;
2344                 }
2345         } else {
2346                 /**
2347                  * Use a simple TX queue without offloads or
2348                  * multi segs if possible
2349                  */
2350                 i40e_set_tx_function_flag(dev, txq);
2351         }
2352         dev->data->tx_queues[queue_idx] = txq;
2353
2354         return 0;
2355 }
2356
2357 void
2358 i40e_dev_tx_queue_release(void *txq)
2359 {
2360         struct i40e_tx_queue *q = (struct i40e_tx_queue *)txq;
2361
2362         if (!q) {
2363                 PMD_DRV_LOG(DEBUG, "Pointer to TX queue is NULL");
2364                 return;
2365         }
2366
2367         i40e_tx_queue_release_mbufs(q);
2368         rte_free(q->sw_ring);
2369         rte_free(q);
2370 }
2371
2372 const struct rte_memzone *
2373 i40e_memzone_reserve(const char *name, uint32_t len, int socket_id)
2374 {
2375         const struct rte_memzone *mz;
2376
2377         mz = rte_memzone_lookup(name);
2378         if (mz)
2379                 return mz;
2380
2381         mz = rte_memzone_reserve_aligned(name, len, socket_id,
2382                         RTE_MEMZONE_IOVA_CONTIG, I40E_RING_BASE_ALIGN);
2383         return mz;
2384 }
2385
2386 void
2387 i40e_rx_queue_release_mbufs(struct i40e_rx_queue *rxq)
2388 {
2389         uint16_t i;
2390
2391         /* SSE Vector driver has a different way of releasing mbufs. */
2392         if (rxq->rx_using_sse) {
2393                 i40e_rx_queue_release_mbufs_vec(rxq);
2394                 return;
2395         }
2396
2397         if (!rxq->sw_ring) {
2398                 PMD_DRV_LOG(DEBUG, "Pointer to sw_ring is NULL");
2399                 return;
2400         }
2401
2402         for (i = 0; i < rxq->nb_rx_desc; i++) {
2403                 if (rxq->sw_ring[i].mbuf) {
2404                         rte_pktmbuf_free_seg(rxq->sw_ring[i].mbuf);
2405                         rxq->sw_ring[i].mbuf = NULL;
2406                 }
2407         }
2408 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
2409         if (rxq->rx_nb_avail == 0)
2410                 return;
2411         for (i = 0; i < rxq->rx_nb_avail; i++) {
2412                 struct rte_mbuf *mbuf;
2413
2414                 mbuf = rxq->rx_stage[rxq->rx_next_avail + i];
2415                 rte_pktmbuf_free_seg(mbuf);
2416         }
2417         rxq->rx_nb_avail = 0;
2418 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
2419 }
2420
2421 void
2422 i40e_reset_rx_queue(struct i40e_rx_queue *rxq)
2423 {
2424         unsigned i;
2425         uint16_t len;
2426
2427         if (!rxq) {
2428                 PMD_DRV_LOG(DEBUG, "Pointer to rxq is NULL");
2429                 return;
2430         }
2431
2432 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
2433         if (check_rx_burst_bulk_alloc_preconditions(rxq) == 0)
2434                 len = (uint16_t)(rxq->nb_rx_desc + RTE_PMD_I40E_RX_MAX_BURST);
2435         else
2436 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
2437                 len = rxq->nb_rx_desc;
2438
2439         for (i = 0; i < len * sizeof(union i40e_rx_desc); i++)
2440                 ((volatile char *)rxq->rx_ring)[i] = 0;
2441
2442         memset(&rxq->fake_mbuf, 0x0, sizeof(rxq->fake_mbuf));
2443         for (i = 0; i < RTE_PMD_I40E_RX_MAX_BURST; ++i)
2444                 rxq->sw_ring[rxq->nb_rx_desc + i].mbuf = &rxq->fake_mbuf;
2445
2446 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
2447         rxq->rx_nb_avail = 0;
2448         rxq->rx_next_avail = 0;
2449         rxq->rx_free_trigger = (uint16_t)(rxq->rx_free_thresh - 1);
2450 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
2451         rxq->rx_tail = 0;
2452         rxq->nb_rx_hold = 0;
2453         rxq->pkt_first_seg = NULL;
2454         rxq->pkt_last_seg = NULL;
2455
2456         rxq->rxrearm_start = 0;
2457         rxq->rxrearm_nb = 0;
2458 }
2459
2460 void
2461 i40e_tx_queue_release_mbufs(struct i40e_tx_queue *txq)
2462 {
2463         struct rte_eth_dev *dev;
2464         uint16_t i;
2465
2466         if (!txq || !txq->sw_ring) {
2467                 PMD_DRV_LOG(DEBUG, "Pointer to txq or sw_ring is NULL");
2468                 return;
2469         }
2470
2471         dev = &rte_eth_devices[txq->port_id];
2472
2473         /**
2474          *  vPMD tx will not set sw_ring's mbuf to NULL after free,
2475          *  so need to free remains more carefully.
2476          */
2477         if (dev->tx_pkt_burst == i40e_xmit_pkts_vec_avx2 ||
2478                         dev->tx_pkt_burst == i40e_xmit_pkts_vec) {
2479                 i = txq->tx_next_dd - txq->tx_rs_thresh + 1;
2480                 if (txq->tx_tail < i) {
2481                         for (; i < txq->nb_tx_desc; i++) {
2482                                 rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
2483                                 txq->sw_ring[i].mbuf = NULL;
2484                         }
2485                         i = 0;
2486                 }
2487                 for (; i < txq->tx_tail; i++) {
2488                         rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
2489                         txq->sw_ring[i].mbuf = NULL;
2490                 }
2491         } else {
2492                 for (i = 0; i < txq->nb_tx_desc; i++) {
2493                         if (txq->sw_ring[i].mbuf) {
2494                                 rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
2495                                 txq->sw_ring[i].mbuf = NULL;
2496                         }
2497                 }
2498         }
2499 }
2500
2501 static int
2502 i40e_tx_done_cleanup_full(struct i40e_tx_queue *txq,
2503                         uint32_t free_cnt)
2504 {
2505         struct i40e_tx_entry *swr_ring = txq->sw_ring;
2506         uint16_t i, tx_last, tx_id;
2507         uint16_t nb_tx_free_last;
2508         uint16_t nb_tx_to_clean;
2509         uint32_t pkt_cnt;
2510
2511         /* Start free mbuf from the next of tx_tail */
2512         tx_last = txq->tx_tail;
2513         tx_id  = swr_ring[tx_last].next_id;
2514
2515         if (txq->nb_tx_free == 0 && i40e_xmit_cleanup(txq))
2516                 return 0;
2517
2518         nb_tx_to_clean = txq->nb_tx_free;
2519         nb_tx_free_last = txq->nb_tx_free;
2520         if (!free_cnt)
2521                 free_cnt = txq->nb_tx_desc;
2522
2523         /* Loop through swr_ring to count the amount of
2524          * freeable mubfs and packets.
2525          */
2526         for (pkt_cnt = 0; pkt_cnt < free_cnt; ) {
2527                 for (i = 0; i < nb_tx_to_clean &&
2528                         pkt_cnt < free_cnt &&
2529                         tx_id != tx_last; i++) {
2530                         if (swr_ring[tx_id].mbuf != NULL) {
2531                                 rte_pktmbuf_free_seg(swr_ring[tx_id].mbuf);
2532                                 swr_ring[tx_id].mbuf = NULL;
2533
2534                                 /*
2535                                  * last segment in the packet,
2536                                  * increment packet count
2537                                  */
2538                                 pkt_cnt += (swr_ring[tx_id].last_id == tx_id);
2539                         }
2540
2541                         tx_id = swr_ring[tx_id].next_id;
2542                 }
2543
2544                 if (txq->tx_rs_thresh > txq->nb_tx_desc -
2545                         txq->nb_tx_free || tx_id == tx_last)
2546                         break;
2547
2548                 if (pkt_cnt < free_cnt) {
2549                         if (i40e_xmit_cleanup(txq))
2550                                 break;
2551
2552                         nb_tx_to_clean = txq->nb_tx_free - nb_tx_free_last;
2553                         nb_tx_free_last = txq->nb_tx_free;
2554                 }
2555         }
2556
2557         return (int)pkt_cnt;
2558 }
2559
2560 static int
2561 i40e_tx_done_cleanup_simple(struct i40e_tx_queue *txq,
2562                         uint32_t free_cnt)
2563 {
2564         int i, n, cnt;
2565
2566         if (free_cnt == 0 || free_cnt > txq->nb_tx_desc)
2567                 free_cnt = txq->nb_tx_desc;
2568
2569         cnt = free_cnt - free_cnt % txq->tx_rs_thresh;
2570
2571         for (i = 0; i < cnt; i += n) {
2572                 if (txq->nb_tx_desc - txq->nb_tx_free < txq->tx_rs_thresh)
2573                         break;
2574
2575                 n = i40e_tx_free_bufs(txq);
2576
2577                 if (n == 0)
2578                         break;
2579         }
2580
2581         return i;
2582 }
2583
2584 static int
2585 i40e_tx_done_cleanup_vec(struct i40e_tx_queue *txq __rte_unused,
2586                         uint32_t free_cnt __rte_unused)
2587 {
2588         return -ENOTSUP;
2589 }
2590 int
2591 i40e_tx_done_cleanup(void *txq, uint32_t free_cnt)
2592 {
2593         struct i40e_tx_queue *q = (struct i40e_tx_queue *)txq;
2594         struct rte_eth_dev *dev = &rte_eth_devices[q->port_id];
2595         struct i40e_adapter *ad =
2596                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
2597
2598         if (ad->tx_simple_allowed) {
2599                 if (ad->tx_vec_allowed)
2600                         return i40e_tx_done_cleanup_vec(q, free_cnt);
2601                 else
2602                         return i40e_tx_done_cleanup_simple(q, free_cnt);
2603         } else {
2604                 return i40e_tx_done_cleanup_full(q, free_cnt);
2605         }
2606 }
2607
2608 void
2609 i40e_reset_tx_queue(struct i40e_tx_queue *txq)
2610 {
2611         struct i40e_tx_entry *txe;
2612         uint16_t i, prev, size;
2613
2614         if (!txq) {
2615                 PMD_DRV_LOG(DEBUG, "Pointer to txq is NULL");
2616                 return;
2617         }
2618
2619         txe = txq->sw_ring;
2620         size = sizeof(struct i40e_tx_desc) * txq->nb_tx_desc;
2621         for (i = 0; i < size; i++)
2622                 ((volatile char *)txq->tx_ring)[i] = 0;
2623
2624         prev = (uint16_t)(txq->nb_tx_desc - 1);
2625         for (i = 0; i < txq->nb_tx_desc; i++) {
2626                 volatile struct i40e_tx_desc *txd = &txq->tx_ring[i];
2627
2628                 txd->cmd_type_offset_bsz =
2629                         rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DESC_DONE);
2630                 txe[i].mbuf =  NULL;
2631                 txe[i].last_id = i;
2632                 txe[prev].next_id = i;
2633                 prev = i;
2634         }
2635
2636         txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
2637         txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
2638
2639         txq->tx_tail = 0;
2640         txq->nb_tx_used = 0;
2641
2642         txq->last_desc_cleaned = (uint16_t)(txq->nb_tx_desc - 1);
2643         txq->nb_tx_free = (uint16_t)(txq->nb_tx_desc - 1);
2644 }
2645
2646 /* Init the TX queue in hardware */
2647 int
2648 i40e_tx_queue_init(struct i40e_tx_queue *txq)
2649 {
2650         enum i40e_status_code err = I40E_SUCCESS;
2651         struct i40e_vsi *vsi = txq->vsi;
2652         struct i40e_hw *hw = I40E_VSI_TO_HW(vsi);
2653         uint16_t pf_q = txq->reg_idx;
2654         struct i40e_hmc_obj_txq tx_ctx;
2655         uint32_t qtx_ctl;
2656
2657         /* clear the context structure first */
2658         memset(&tx_ctx, 0, sizeof(tx_ctx));
2659         tx_ctx.new_context = 1;
2660         tx_ctx.base = txq->tx_ring_phys_addr / I40E_QUEUE_BASE_ADDR_UNIT;
2661         tx_ctx.qlen = txq->nb_tx_desc;
2662
2663 #ifdef RTE_LIBRTE_IEEE1588
2664         tx_ctx.timesync_ena = 1;
2665 #endif
2666         tx_ctx.rdylist = rte_le_to_cpu_16(vsi->info.qs_handle[txq->dcb_tc]);
2667         if (vsi->type == I40E_VSI_FDIR)
2668                 tx_ctx.fd_ena = TRUE;
2669
2670         err = i40e_clear_lan_tx_queue_context(hw, pf_q);
2671         if (err != I40E_SUCCESS) {
2672                 PMD_DRV_LOG(ERR, "Failure of clean lan tx queue context");
2673                 return err;
2674         }
2675
2676         err = i40e_set_lan_tx_queue_context(hw, pf_q, &tx_ctx);
2677         if (err != I40E_SUCCESS) {
2678                 PMD_DRV_LOG(ERR, "Failure of set lan tx queue context");
2679                 return err;
2680         }
2681
2682         /* Now associate this queue with this PCI function */
2683         qtx_ctl = I40E_QTX_CTL_PF_QUEUE;
2684         qtx_ctl |= ((hw->pf_id << I40E_QTX_CTL_PF_INDX_SHIFT) &
2685                                         I40E_QTX_CTL_PF_INDX_MASK);
2686         I40E_WRITE_REG(hw, I40E_QTX_CTL(pf_q), qtx_ctl);
2687         I40E_WRITE_FLUSH(hw);
2688
2689         txq->qtx_tail = hw->hw_addr + I40E_QTX_TAIL(pf_q);
2690
2691         return err;
2692 }
2693
2694 int
2695 i40e_alloc_rx_queue_mbufs(struct i40e_rx_queue *rxq)
2696 {
2697         struct i40e_rx_entry *rxe = rxq->sw_ring;
2698         uint64_t dma_addr;
2699         uint16_t i;
2700
2701         for (i = 0; i < rxq->nb_rx_desc; i++) {
2702                 volatile union i40e_rx_desc *rxd;
2703                 struct rte_mbuf *mbuf = rte_mbuf_raw_alloc(rxq->mp);
2704
2705                 if (unlikely(!mbuf)) {
2706                         PMD_DRV_LOG(ERR, "Failed to allocate mbuf for RX");
2707                         return -ENOMEM;
2708                 }
2709
2710                 rte_mbuf_refcnt_set(mbuf, 1);
2711                 mbuf->next = NULL;
2712                 mbuf->data_off = RTE_PKTMBUF_HEADROOM;
2713                 mbuf->nb_segs = 1;
2714                 mbuf->port = rxq->port_id;
2715
2716                 dma_addr =
2717                         rte_cpu_to_le_64(rte_mbuf_data_iova_default(mbuf));
2718
2719                 rxd = &rxq->rx_ring[i];
2720                 rxd->read.pkt_addr = dma_addr;
2721                 rxd->read.hdr_addr = 0;
2722 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
2723                 rxd->read.rsvd1 = 0;
2724                 rxd->read.rsvd2 = 0;
2725 #endif /* RTE_LIBRTE_I40E_16BYTE_RX_DESC */
2726
2727                 rxe[i].mbuf = mbuf;
2728         }
2729
2730         return 0;
2731 }
2732
2733 /*
2734  * Calculate the buffer length, and check the jumbo frame
2735  * and maximum packet length.
2736  */
2737 static int
2738 i40e_rx_queue_config(struct i40e_rx_queue *rxq)
2739 {
2740         struct i40e_pf *pf = I40E_VSI_TO_PF(rxq->vsi);
2741         struct i40e_hw *hw = I40E_VSI_TO_HW(rxq->vsi);
2742         struct rte_eth_dev_data *data = pf->dev_data;
2743         uint16_t buf_size;
2744
2745         buf_size = (uint16_t)(rte_pktmbuf_data_room_size(rxq->mp) -
2746                 RTE_PKTMBUF_HEADROOM);
2747
2748         switch (pf->flags & (I40E_FLAG_HEADER_SPLIT_DISABLED |
2749                         I40E_FLAG_HEADER_SPLIT_ENABLED)) {
2750         case I40E_FLAG_HEADER_SPLIT_ENABLED: /* Not supported */
2751                 rxq->rx_hdr_len = RTE_ALIGN(I40E_RXBUF_SZ_1024,
2752                                 (1 << I40E_RXQ_CTX_HBUFF_SHIFT));
2753                 rxq->rx_buf_len = RTE_ALIGN(I40E_RXBUF_SZ_2048,
2754                                 (1 << I40E_RXQ_CTX_DBUFF_SHIFT));
2755                 rxq->hs_mode = i40e_header_split_enabled;
2756                 break;
2757         case I40E_FLAG_HEADER_SPLIT_DISABLED:
2758         default:
2759                 rxq->rx_hdr_len = 0;
2760                 rxq->rx_buf_len = RTE_ALIGN_FLOOR(buf_size,
2761                         (1 << I40E_RXQ_CTX_DBUFF_SHIFT));
2762                 rxq->hs_mode = i40e_header_split_none;
2763                 break;
2764         }
2765
2766         rxq->max_pkt_len =
2767                 RTE_MIN((uint32_t)(hw->func_caps.rx_buf_chain_len *
2768                         rxq->rx_buf_len), data->dev_conf.rxmode.max_rx_pkt_len);
2769         if (data->dev_conf.rxmode.offloads & DEV_RX_OFFLOAD_JUMBO_FRAME) {
2770                 if (rxq->max_pkt_len <= RTE_ETHER_MAX_LEN ||
2771                         rxq->max_pkt_len > I40E_FRAME_SIZE_MAX) {
2772                         PMD_DRV_LOG(ERR, "maximum packet length must "
2773                                     "be larger than %u and smaller than %u,"
2774                                     "as jumbo frame is enabled",
2775                                     (uint32_t)RTE_ETHER_MAX_LEN,
2776                                     (uint32_t)I40E_FRAME_SIZE_MAX);
2777                         return I40E_ERR_CONFIG;
2778                 }
2779         } else {
2780                 if (rxq->max_pkt_len < RTE_ETHER_MIN_LEN ||
2781                         rxq->max_pkt_len > RTE_ETHER_MAX_LEN) {
2782                         PMD_DRV_LOG(ERR, "maximum packet length must be "
2783                                     "larger than %u and smaller than %u, "
2784                                     "as jumbo frame is disabled",
2785                                     (uint32_t)RTE_ETHER_MIN_LEN,
2786                                     (uint32_t)RTE_ETHER_MAX_LEN);
2787                         return I40E_ERR_CONFIG;
2788                 }
2789         }
2790
2791         return 0;
2792 }
2793
2794 /* Init the RX queue in hardware */
2795 int
2796 i40e_rx_queue_init(struct i40e_rx_queue *rxq)
2797 {
2798         int err = I40E_SUCCESS;
2799         struct i40e_hw *hw = I40E_VSI_TO_HW(rxq->vsi);
2800         struct rte_eth_dev_data *dev_data = I40E_VSI_TO_DEV_DATA(rxq->vsi);
2801         uint16_t pf_q = rxq->reg_idx;
2802         uint16_t buf_size;
2803         struct i40e_hmc_obj_rxq rx_ctx;
2804
2805         err = i40e_rx_queue_config(rxq);
2806         if (err < 0) {
2807                 PMD_DRV_LOG(ERR, "Failed to config RX queue");
2808                 return err;
2809         }
2810
2811         /* Clear the context structure first */
2812         memset(&rx_ctx, 0, sizeof(struct i40e_hmc_obj_rxq));
2813         rx_ctx.dbuff = rxq->rx_buf_len >> I40E_RXQ_CTX_DBUFF_SHIFT;
2814         rx_ctx.hbuff = rxq->rx_hdr_len >> I40E_RXQ_CTX_HBUFF_SHIFT;
2815
2816         rx_ctx.base = rxq->rx_ring_phys_addr / I40E_QUEUE_BASE_ADDR_UNIT;
2817         rx_ctx.qlen = rxq->nb_rx_desc;
2818 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
2819         rx_ctx.dsize = 1;
2820 #endif
2821         rx_ctx.dtype = rxq->hs_mode;
2822         if (rxq->hs_mode)
2823                 rx_ctx.hsplit_0 = I40E_HEADER_SPLIT_ALL;
2824         else
2825                 rx_ctx.hsplit_0 = I40E_HEADER_SPLIT_NONE;
2826         rx_ctx.rxmax = rxq->max_pkt_len;
2827         rx_ctx.tphrdesc_ena = 1;
2828         rx_ctx.tphwdesc_ena = 1;
2829         rx_ctx.tphdata_ena = 1;
2830         rx_ctx.tphhead_ena = 1;
2831         rx_ctx.lrxqthresh = 2;
2832         rx_ctx.crcstrip = (rxq->crc_len == 0) ? 1 : 0;
2833         rx_ctx.l2tsel = 1;
2834         /* showiv indicates if inner VLAN is stripped inside of tunnel
2835          * packet. When set it to 1, vlan information is stripped from
2836          * the inner header, but the hardware does not put it in the
2837          * descriptor. So set it zero by default.
2838          */
2839         rx_ctx.showiv = 0;
2840         rx_ctx.prefena = 1;
2841
2842         err = i40e_clear_lan_rx_queue_context(hw, pf_q);
2843         if (err != I40E_SUCCESS) {
2844                 PMD_DRV_LOG(ERR, "Failed to clear LAN RX queue context");
2845                 return err;
2846         }
2847         err = i40e_set_lan_rx_queue_context(hw, pf_q, &rx_ctx);
2848         if (err != I40E_SUCCESS) {
2849                 PMD_DRV_LOG(ERR, "Failed to set LAN RX queue context");
2850                 return err;
2851         }
2852
2853         rxq->qrx_tail = hw->hw_addr + I40E_QRX_TAIL(pf_q);
2854
2855         buf_size = (uint16_t)(rte_pktmbuf_data_room_size(rxq->mp) -
2856                 RTE_PKTMBUF_HEADROOM);
2857
2858         /* Check if scattered RX needs to be used. */
2859         if (rxq->max_pkt_len > buf_size)
2860                 dev_data->scattered_rx = 1;
2861
2862         /* Init the RX tail regieter. */
2863         I40E_PCI_REG_WRITE(rxq->qrx_tail, rxq->nb_rx_desc - 1);
2864
2865         return 0;
2866 }
2867
2868 void
2869 i40e_dev_clear_queues(struct rte_eth_dev *dev)
2870 {
2871         uint16_t i;
2872
2873         PMD_INIT_FUNC_TRACE();
2874
2875         for (i = 0; i < dev->data->nb_tx_queues; i++) {
2876                 if (!dev->data->tx_queues[i])
2877                         continue;
2878                 i40e_tx_queue_release_mbufs(dev->data->tx_queues[i]);
2879                 i40e_reset_tx_queue(dev->data->tx_queues[i]);
2880         }
2881
2882         for (i = 0; i < dev->data->nb_rx_queues; i++) {
2883                 if (!dev->data->rx_queues[i])
2884                         continue;
2885                 i40e_rx_queue_release_mbufs(dev->data->rx_queues[i]);
2886                 i40e_reset_rx_queue(dev->data->rx_queues[i]);
2887         }
2888 }
2889
2890 void
2891 i40e_dev_free_queues(struct rte_eth_dev *dev)
2892 {
2893         uint16_t i;
2894
2895         PMD_INIT_FUNC_TRACE();
2896
2897         for (i = 0; i < dev->data->nb_rx_queues; i++) {
2898                 if (!dev->data->rx_queues[i])
2899                         continue;
2900                 i40e_dev_rx_queue_release(dev->data->rx_queues[i]);
2901                 dev->data->rx_queues[i] = NULL;
2902         }
2903
2904         for (i = 0; i < dev->data->nb_tx_queues; i++) {
2905                 if (!dev->data->tx_queues[i])
2906                         continue;
2907                 i40e_dev_tx_queue_release(dev->data->tx_queues[i]);
2908                 dev->data->tx_queues[i] = NULL;
2909         }
2910 }
2911
2912 #define I40E_FDIR_NUM_TX_DESC  I40E_MIN_RING_DESC
2913 #define I40E_FDIR_NUM_RX_DESC  I40E_MIN_RING_DESC
2914
2915 enum i40e_status_code
2916 i40e_fdir_setup_tx_resources(struct i40e_pf *pf)
2917 {
2918         struct i40e_tx_queue *txq;
2919         const struct rte_memzone *tz = NULL;
2920         uint32_t ring_size;
2921         struct rte_eth_dev *dev;
2922
2923         if (!pf) {
2924                 PMD_DRV_LOG(ERR, "PF is not available");
2925                 return I40E_ERR_BAD_PTR;
2926         }
2927
2928         dev = pf->adapter->eth_dev;
2929
2930         /* Allocate the TX queue data structure. */
2931         txq = rte_zmalloc_socket("i40e fdir tx queue",
2932                                   sizeof(struct i40e_tx_queue),
2933                                   RTE_CACHE_LINE_SIZE,
2934                                   SOCKET_ID_ANY);
2935         if (!txq) {
2936                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
2937                                         "tx queue structure.");
2938                 return I40E_ERR_NO_MEMORY;
2939         }
2940
2941         /* Allocate TX hardware ring descriptors. */
2942         ring_size = sizeof(struct i40e_tx_desc) * I40E_FDIR_NUM_TX_DESC;
2943         ring_size = RTE_ALIGN(ring_size, I40E_DMA_MEM_ALIGN);
2944
2945         tz = rte_eth_dma_zone_reserve(dev, "fdir_tx_ring",
2946                                       I40E_FDIR_QUEUE_ID, ring_size,
2947                                       I40E_RING_BASE_ALIGN, SOCKET_ID_ANY);
2948         if (!tz) {
2949                 i40e_dev_tx_queue_release(txq);
2950                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for TX.");
2951                 return I40E_ERR_NO_MEMORY;
2952         }
2953
2954         txq->nb_tx_desc = I40E_FDIR_NUM_TX_DESC;
2955         txq->queue_id = I40E_FDIR_QUEUE_ID;
2956         txq->reg_idx = pf->fdir.fdir_vsi->base_queue;
2957         txq->vsi = pf->fdir.fdir_vsi;
2958
2959         txq->tx_ring_phys_addr = tz->iova;
2960         txq->tx_ring = (struct i40e_tx_desc *)tz->addr;
2961         /*
2962          * don't need to allocate software ring and reset for the fdir
2963          * program queue just set the queue has been configured.
2964          */
2965         txq->q_set = TRUE;
2966         pf->fdir.txq = txq;
2967
2968         return I40E_SUCCESS;
2969 }
2970
2971 enum i40e_status_code
2972 i40e_fdir_setup_rx_resources(struct i40e_pf *pf)
2973 {
2974         struct i40e_rx_queue *rxq;
2975         const struct rte_memzone *rz = NULL;
2976         uint32_t ring_size;
2977         struct rte_eth_dev *dev;
2978
2979         if (!pf) {
2980                 PMD_DRV_LOG(ERR, "PF is not available");
2981                 return I40E_ERR_BAD_PTR;
2982         }
2983
2984         dev = pf->adapter->eth_dev;
2985
2986         /* Allocate the RX queue data structure. */
2987         rxq = rte_zmalloc_socket("i40e fdir rx queue",
2988                                   sizeof(struct i40e_rx_queue),
2989                                   RTE_CACHE_LINE_SIZE,
2990                                   SOCKET_ID_ANY);
2991         if (!rxq) {
2992                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
2993                                         "rx queue structure.");
2994                 return I40E_ERR_NO_MEMORY;
2995         }
2996
2997         /* Allocate RX hardware ring descriptors. */
2998         ring_size = sizeof(union i40e_rx_desc) * I40E_FDIR_NUM_RX_DESC;
2999         ring_size = RTE_ALIGN(ring_size, I40E_DMA_MEM_ALIGN);
3000
3001         rz = rte_eth_dma_zone_reserve(dev, "fdir_rx_ring",
3002                                       I40E_FDIR_QUEUE_ID, ring_size,
3003                                       I40E_RING_BASE_ALIGN, SOCKET_ID_ANY);
3004         if (!rz) {
3005                 i40e_dev_rx_queue_release(rxq);
3006                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for RX.");
3007                 return I40E_ERR_NO_MEMORY;
3008         }
3009
3010         rxq->nb_rx_desc = I40E_FDIR_NUM_RX_DESC;
3011         rxq->queue_id = I40E_FDIR_QUEUE_ID;
3012         rxq->reg_idx = pf->fdir.fdir_vsi->base_queue;
3013         rxq->vsi = pf->fdir.fdir_vsi;
3014
3015         rxq->rx_ring_phys_addr = rz->iova;
3016         memset(rz->addr, 0, I40E_FDIR_NUM_RX_DESC * sizeof(union i40e_rx_desc));
3017         rxq->rx_ring = (union i40e_rx_desc *)rz->addr;
3018
3019         /*
3020          * Don't need to allocate software ring and reset for the fdir
3021          * rx queue, just set the queue has been configured.
3022          */
3023         rxq->q_set = TRUE;
3024         pf->fdir.rxq = rxq;
3025
3026         return I40E_SUCCESS;
3027 }
3028
3029 void
3030 i40e_rxq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
3031         struct rte_eth_rxq_info *qinfo)
3032 {
3033         struct i40e_rx_queue *rxq;
3034
3035         rxq = dev->data->rx_queues[queue_id];
3036
3037         qinfo->mp = rxq->mp;
3038         qinfo->scattered_rx = dev->data->scattered_rx;
3039         qinfo->nb_desc = rxq->nb_rx_desc;
3040
3041         qinfo->conf.rx_free_thresh = rxq->rx_free_thresh;
3042         qinfo->conf.rx_drop_en = rxq->drop_en;
3043         qinfo->conf.rx_deferred_start = rxq->rx_deferred_start;
3044         qinfo->conf.offloads = rxq->offloads;
3045 }
3046
3047 void
3048 i40e_txq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
3049         struct rte_eth_txq_info *qinfo)
3050 {
3051         struct i40e_tx_queue *txq;
3052
3053         txq = dev->data->tx_queues[queue_id];
3054
3055         qinfo->nb_desc = txq->nb_tx_desc;
3056
3057         qinfo->conf.tx_thresh.pthresh = txq->pthresh;
3058         qinfo->conf.tx_thresh.hthresh = txq->hthresh;
3059         qinfo->conf.tx_thresh.wthresh = txq->wthresh;
3060
3061         qinfo->conf.tx_free_thresh = txq->tx_free_thresh;
3062         qinfo->conf.tx_rs_thresh = txq->tx_rs_thresh;
3063         qinfo->conf.tx_deferred_start = txq->tx_deferred_start;
3064         qinfo->conf.offloads = txq->offloads;
3065 }
3066
3067 static eth_rx_burst_t
3068 i40e_get_latest_rx_vec(bool scatter)
3069 {
3070 #if defined(RTE_ARCH_X86) && defined(CC_AVX2_SUPPORT)
3071         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2))
3072                 return scatter ? i40e_recv_scattered_pkts_vec_avx2 :
3073                                  i40e_recv_pkts_vec_avx2;
3074 #endif
3075         return scatter ? i40e_recv_scattered_pkts_vec :
3076                          i40e_recv_pkts_vec;
3077 }
3078
3079 static eth_rx_burst_t
3080 i40e_get_recommend_rx_vec(bool scatter)
3081 {
3082 #if defined(RTE_ARCH_X86) && defined(CC_AVX2_SUPPORT)
3083         /*
3084          * since AVX frequency can be different to base frequency, limit
3085          * use of AVX2 version to later plaforms, not all those that could
3086          * theoretically run it.
3087          */
3088         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F))
3089                 return scatter ? i40e_recv_scattered_pkts_vec_avx2 :
3090                                  i40e_recv_pkts_vec_avx2;
3091 #endif
3092         return scatter ? i40e_recv_scattered_pkts_vec :
3093                          i40e_recv_pkts_vec;
3094 }
3095
3096 void __attribute__((cold))
3097 i40e_set_rx_function(struct rte_eth_dev *dev)
3098 {
3099         struct i40e_adapter *ad =
3100                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
3101         uint16_t rx_using_sse, i;
3102         /* In order to allow Vector Rx there are a few configuration
3103          * conditions to be met and Rx Bulk Allocation should be allowed.
3104          */
3105         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
3106                 if (i40e_rx_vec_dev_conf_condition_check(dev) ||
3107                     !ad->rx_bulk_alloc_allowed) {
3108                         PMD_INIT_LOG(DEBUG, "Port[%d] doesn't meet"
3109                                      " Vector Rx preconditions",
3110                                      dev->data->port_id);
3111
3112                         ad->rx_vec_allowed = false;
3113                 }
3114                 if (ad->rx_vec_allowed) {
3115                         for (i = 0; i < dev->data->nb_rx_queues; i++) {
3116                                 struct i40e_rx_queue *rxq =
3117                                         dev->data->rx_queues[i];
3118
3119                                 if (rxq && i40e_rxq_vec_setup(rxq)) {
3120                                         ad->rx_vec_allowed = false;
3121                                         break;
3122                                 }
3123                         }
3124                 }
3125         }
3126
3127         if (ad->rx_vec_allowed) {
3128                 /* Vec Rx path */
3129                 PMD_INIT_LOG(DEBUG, "Vector Rx path will be used on port=%d.",
3130                                 dev->data->port_id);
3131                 if (ad->use_latest_vec)
3132                         dev->rx_pkt_burst =
3133                         i40e_get_latest_rx_vec(dev->data->scattered_rx);
3134                 else
3135                         dev->rx_pkt_burst =
3136                         i40e_get_recommend_rx_vec(dev->data->scattered_rx);
3137         } else if (!dev->data->scattered_rx && ad->rx_bulk_alloc_allowed) {
3138                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions are "
3139                                     "satisfied. Rx Burst Bulk Alloc function "
3140                                     "will be used on port=%d.",
3141                              dev->data->port_id);
3142
3143                 dev->rx_pkt_burst = i40e_recv_pkts_bulk_alloc;
3144         } else {
3145                 /* Simple Rx Path. */
3146                 PMD_INIT_LOG(DEBUG, "Simple Rx path will be used on port=%d.",
3147                              dev->data->port_id);
3148                 dev->rx_pkt_burst = dev->data->scattered_rx ?
3149                                         i40e_recv_scattered_pkts :
3150                                         i40e_recv_pkts;
3151         }
3152
3153         /* Propagate information about RX function choice through all queues. */
3154         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
3155                 rx_using_sse =
3156                         (dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec ||
3157                          dev->rx_pkt_burst == i40e_recv_pkts_vec ||
3158                          dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec_avx2 ||
3159                          dev->rx_pkt_burst == i40e_recv_pkts_vec_avx2);
3160
3161                 for (i = 0; i < dev->data->nb_rx_queues; i++) {
3162                         struct i40e_rx_queue *rxq = dev->data->rx_queues[i];
3163
3164                         if (rxq)
3165                                 rxq->rx_using_sse = rx_using_sse;
3166                 }
3167         }
3168 }
3169
3170 static const struct {
3171         eth_rx_burst_t pkt_burst;
3172         const char *info;
3173 } i40e_rx_burst_infos[] = {
3174         { i40e_recv_scattered_pkts,          "Scalar Scattered" },
3175         { i40e_recv_pkts_bulk_alloc,         "Scalar Bulk Alloc" },
3176         { i40e_recv_pkts,                    "Scalar" },
3177 #ifdef RTE_ARCH_X86
3178         { i40e_recv_scattered_pkts_vec_avx2, "Vector AVX2 Scattered" },
3179         { i40e_recv_pkts_vec_avx2,           "Vector AVX2" },
3180         { i40e_recv_scattered_pkts_vec,      "Vector SSE Scattered" },
3181         { i40e_recv_pkts_vec,                "Vector SSE" },
3182 #elif defined(RTE_ARCH_ARM64)
3183         { i40e_recv_scattered_pkts_vec,      "Vector Neon Scattered" },
3184         { i40e_recv_pkts_vec,                "Vector Neon" },
3185 #elif defined(RTE_ARCH_PPC_64)
3186         { i40e_recv_scattered_pkts_vec,      "Vector AltiVec Scattered" },
3187         { i40e_recv_pkts_vec,                "Vector AltiVec" },
3188 #endif
3189 };
3190
3191 int
3192 i40e_rx_burst_mode_get(struct rte_eth_dev *dev, __rte_unused uint16_t queue_id,
3193                        struct rte_eth_burst_mode *mode)
3194 {
3195         eth_rx_burst_t pkt_burst = dev->rx_pkt_burst;
3196         int ret = -EINVAL;
3197         unsigned int i;
3198
3199         for (i = 0; i < RTE_DIM(i40e_rx_burst_infos); ++i) {
3200                 if (pkt_burst == i40e_rx_burst_infos[i].pkt_burst) {
3201                         snprintf(mode->info, sizeof(mode->info), "%s",
3202                                  i40e_rx_burst_infos[i].info);
3203                         ret = 0;
3204                         break;
3205                 }
3206         }
3207
3208         return ret;
3209 }
3210
3211 void __attribute__((cold))
3212 i40e_set_tx_function_flag(struct rte_eth_dev *dev, struct i40e_tx_queue *txq)
3213 {
3214         struct i40e_adapter *ad =
3215                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
3216
3217         /* Use a simple Tx queue if possible (only fast free is allowed) */
3218         ad->tx_simple_allowed =
3219                 (txq->offloads ==
3220                  (txq->offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE) &&
3221                  txq->tx_rs_thresh >= RTE_PMD_I40E_TX_MAX_BURST);
3222         ad->tx_vec_allowed = (ad->tx_simple_allowed &&
3223                         txq->tx_rs_thresh <= RTE_I40E_TX_MAX_FREE_BUF_SZ);
3224
3225         if (ad->tx_vec_allowed)
3226                 PMD_INIT_LOG(DEBUG, "Vector Tx can be enabled on Tx queue %u.",
3227                                 txq->queue_id);
3228         else if (ad->tx_simple_allowed)
3229                 PMD_INIT_LOG(DEBUG, "Simple Tx can be enabled on Tx queue %u.",
3230                                 txq->queue_id);
3231         else
3232                 PMD_INIT_LOG(DEBUG,
3233                                 "Neither simple nor vector Tx enabled on Tx queue %u\n",
3234                                 txq->queue_id);
3235 }
3236
3237 static eth_tx_burst_t
3238 i40e_get_latest_tx_vec(void)
3239 {
3240 #if defined(RTE_ARCH_X86) && defined(CC_AVX2_SUPPORT)
3241         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX2))
3242                 return i40e_xmit_pkts_vec_avx2;
3243 #endif
3244         return i40e_xmit_pkts_vec;
3245 }
3246
3247 static eth_tx_burst_t
3248 i40e_get_recommend_tx_vec(void)
3249 {
3250 #if defined(RTE_ARCH_X86) && defined(CC_AVX2_SUPPORT)
3251         /*
3252          * since AVX frequency can be different to base frequency, limit
3253          * use of AVX2 version to later plaforms, not all those that could
3254          * theoretically run it.
3255          */
3256         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F))
3257                 return i40e_xmit_pkts_vec_avx2;
3258 #endif
3259         return i40e_xmit_pkts_vec;
3260 }
3261
3262 void __attribute__((cold))
3263 i40e_set_tx_function(struct rte_eth_dev *dev)
3264 {
3265         struct i40e_adapter *ad =
3266                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
3267         int i;
3268
3269         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
3270                 if (ad->tx_vec_allowed) {
3271                         for (i = 0; i < dev->data->nb_tx_queues; i++) {
3272                                 struct i40e_tx_queue *txq =
3273                                         dev->data->tx_queues[i];
3274
3275                                 if (txq && i40e_txq_vec_setup(txq)) {
3276                                         ad->tx_vec_allowed = false;
3277                                         break;
3278                                 }
3279                         }
3280                 }
3281         }
3282
3283         if (ad->tx_simple_allowed) {
3284                 if (ad->tx_vec_allowed) {
3285                         PMD_INIT_LOG(DEBUG, "Vector tx finally be used.");
3286                         if (ad->use_latest_vec)
3287                                 dev->tx_pkt_burst =
3288                                         i40e_get_latest_tx_vec();
3289                         else
3290                                 dev->tx_pkt_burst =
3291                                         i40e_get_recommend_tx_vec();
3292                 } else {
3293                         PMD_INIT_LOG(DEBUG, "Simple tx finally be used.");
3294                         dev->tx_pkt_burst = i40e_xmit_pkts_simple;
3295                 }
3296                 dev->tx_pkt_prepare = NULL;
3297         } else {
3298                 PMD_INIT_LOG(DEBUG, "Xmit tx finally be used.");
3299                 dev->tx_pkt_burst = i40e_xmit_pkts;
3300                 dev->tx_pkt_prepare = i40e_prep_pkts;
3301         }
3302 }
3303
3304 static const struct {
3305         eth_tx_burst_t pkt_burst;
3306         const char *info;
3307 } i40e_tx_burst_infos[] = {
3308         { i40e_xmit_pkts_simple,   "Scalar Simple" },
3309         { i40e_xmit_pkts,          "Scalar" },
3310 #ifdef RTE_ARCH_X86
3311         { i40e_xmit_pkts_vec_avx2, "Vector AVX2" },
3312         { i40e_xmit_pkts_vec,      "Vector SSE" },
3313 #elif defined(RTE_ARCH_ARM64)
3314         { i40e_xmit_pkts_vec,      "Vector Neon" },
3315 #elif defined(RTE_ARCH_PPC_64)
3316         { i40e_xmit_pkts_vec,      "Vector AltiVec" },
3317 #endif
3318 };
3319
3320 int
3321 i40e_tx_burst_mode_get(struct rte_eth_dev *dev, __rte_unused uint16_t queue_id,
3322                        struct rte_eth_burst_mode *mode)
3323 {
3324         eth_tx_burst_t pkt_burst = dev->tx_pkt_burst;
3325         int ret = -EINVAL;
3326         unsigned int i;
3327
3328         for (i = 0; i < RTE_DIM(i40e_tx_burst_infos); ++i) {
3329                 if (pkt_burst == i40e_tx_burst_infos[i].pkt_burst) {
3330                         snprintf(mode->info, sizeof(mode->info), "%s",
3331                                  i40e_tx_burst_infos[i].info);
3332                         ret = 0;
3333                         break;
3334                 }
3335         }
3336
3337         return ret;
3338 }
3339
3340 void __attribute__((cold))
3341 i40e_set_default_ptype_table(struct rte_eth_dev *dev)
3342 {
3343         struct i40e_adapter *ad =
3344                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
3345         int i;
3346
3347         for (i = 0; i < I40E_MAX_PKT_TYPE; i++)
3348                 ad->ptype_tbl[i] = i40e_get_default_pkt_type(i);
3349 }
3350
3351 void __attribute__((cold))
3352 i40e_set_default_pctype_table(struct rte_eth_dev *dev)
3353 {
3354         struct i40e_adapter *ad =
3355                         I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
3356         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
3357         int i;
3358
3359         for (i = 0; i < I40E_FLOW_TYPE_MAX; i++)
3360                 ad->pctypes_tbl[i] = 0ULL;
3361         ad->flow_types_mask = 0ULL;
3362         ad->pctypes_mask = 0ULL;
3363
3364         ad->pctypes_tbl[RTE_ETH_FLOW_FRAG_IPV4] =
3365                                 (1ULL << I40E_FILTER_PCTYPE_FRAG_IPV4);
3366         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_UDP] =
3367                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_UDP);
3368         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_TCP] =
3369                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_TCP);
3370         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_SCTP] =
3371                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_SCTP);
3372         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_OTHER] =
3373                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_OTHER);
3374         ad->pctypes_tbl[RTE_ETH_FLOW_FRAG_IPV6] =
3375                                 (1ULL << I40E_FILTER_PCTYPE_FRAG_IPV6);
3376         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_UDP] =
3377                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_UDP);
3378         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_TCP] =
3379                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_TCP);
3380         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_SCTP] =
3381                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_SCTP);
3382         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_OTHER] =
3383                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_OTHER);
3384         ad->pctypes_tbl[RTE_ETH_FLOW_L2_PAYLOAD] =
3385                                 (1ULL << I40E_FILTER_PCTYPE_L2_PAYLOAD);
3386
3387         if (hw->mac.type == I40E_MAC_X722 ||
3388                 hw->mac.type == I40E_MAC_X722_VF) {
3389                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_UDP] |=
3390                         (1ULL << I40E_FILTER_PCTYPE_NONF_UNICAST_IPV4_UDP);
3391                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_UDP] |=
3392                         (1ULL << I40E_FILTER_PCTYPE_NONF_MULTICAST_IPV4_UDP);
3393                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_TCP] |=
3394                         (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_TCP_SYN_NO_ACK);
3395                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_UDP] |=
3396                         (1ULL << I40E_FILTER_PCTYPE_NONF_UNICAST_IPV6_UDP);
3397                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_UDP] |=
3398                         (1ULL << I40E_FILTER_PCTYPE_NONF_MULTICAST_IPV6_UDP);
3399                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_TCP] |=
3400                         (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_TCP_SYN_NO_ACK);
3401         }
3402
3403         for (i = 0; i < I40E_FLOW_TYPE_MAX; i++) {
3404                 if (ad->pctypes_tbl[i])
3405                         ad->flow_types_mask |= (1ULL << i);
3406                 ad->pctypes_mask |= ad->pctypes_tbl[i];
3407         }
3408 }
3409
3410 #ifndef RTE_LIBRTE_I40E_INC_VECTOR
3411 /* Stubs needed for linkage when CONFIG_RTE_LIBRTE_I40E_INC_VECTOR is set to 'n' */
3412 int
3413 i40e_rx_vec_dev_conf_condition_check(struct rte_eth_dev __rte_unused *dev)
3414 {
3415         return -1;
3416 }
3417
3418 uint16_t
3419 i40e_recv_pkts_vec(
3420         void __rte_unused *rx_queue,
3421         struct rte_mbuf __rte_unused **rx_pkts,
3422         uint16_t __rte_unused nb_pkts)
3423 {
3424         return 0;
3425 }
3426
3427 uint16_t
3428 i40e_recv_scattered_pkts_vec(
3429         void __rte_unused *rx_queue,
3430         struct rte_mbuf __rte_unused **rx_pkts,
3431         uint16_t __rte_unused nb_pkts)
3432 {
3433         return 0;
3434 }
3435
3436 int
3437 i40e_rxq_vec_setup(struct i40e_rx_queue __rte_unused *rxq)
3438 {
3439         return -1;
3440 }
3441
3442 int
3443 i40e_txq_vec_setup(struct i40e_tx_queue __rte_unused *txq)
3444 {
3445         return -1;
3446 }
3447
3448 void
3449 i40e_rx_queue_release_mbufs_vec(struct i40e_rx_queue __rte_unused*rxq)
3450 {
3451         return;
3452 }
3453
3454 uint16_t
3455 i40e_xmit_fixed_burst_vec(void __rte_unused * tx_queue,
3456                           struct rte_mbuf __rte_unused **tx_pkts,
3457                           uint16_t __rte_unused nb_pkts)
3458 {
3459         return 0;
3460 }
3461 #endif /* ifndef RTE_LIBRTE_I40E_INC_VECTOR */
3462
3463 #ifndef CC_AVX2_SUPPORT
3464 uint16_t
3465 i40e_recv_pkts_vec_avx2(void __rte_unused *rx_queue,
3466                         struct rte_mbuf __rte_unused **rx_pkts,
3467                         uint16_t __rte_unused nb_pkts)
3468 {
3469         return 0;
3470 }
3471
3472 uint16_t
3473 i40e_recv_scattered_pkts_vec_avx2(void __rte_unused *rx_queue,
3474                         struct rte_mbuf __rte_unused **rx_pkts,
3475                         uint16_t __rte_unused nb_pkts)
3476 {
3477         return 0;
3478 }
3479
3480 uint16_t
3481 i40e_xmit_pkts_vec_avx2(void __rte_unused * tx_queue,
3482                           struct rte_mbuf __rte_unused **tx_pkts,
3483                           uint16_t __rte_unused nb_pkts)
3484 {
3485         return 0;
3486 }
3487 #endif /* ifndef CC_AVX2_SUPPORT */