net/i40e: convert to new Rx offloads API
[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_SIMPLE_FLAGS ((uint32_t)ETH_TXQ_FLAGS_NOMULTSEGS | \
44                                         ETH_TXQ_FLAGS_NOOFFLOADS)
45
46 #define I40E_TXD_CMD (I40E_TX_DESC_CMD_EOP | I40E_TX_DESC_CMD_RS)
47
48 #ifdef RTE_LIBRTE_IEEE1588
49 #define I40E_TX_IEEE1588_TMST PKT_TX_IEEE1588_TMST
50 #else
51 #define I40E_TX_IEEE1588_TMST 0
52 #endif
53
54 #define I40E_TX_CKSUM_OFFLOAD_MASK (             \
55                 PKT_TX_IP_CKSUM |                \
56                 PKT_TX_L4_MASK |                 \
57                 PKT_TX_TCP_SEG |                 \
58                 PKT_TX_OUTER_IP_CKSUM)
59
60 #define I40E_TX_OFFLOAD_MASK (  \
61                 PKT_TX_IP_CKSUM |       \
62                 PKT_TX_L4_MASK |        \
63                 PKT_TX_OUTER_IP_CKSUM | \
64                 PKT_TX_TCP_SEG |        \
65                 PKT_TX_QINQ_PKT |       \
66                 PKT_TX_VLAN_PKT |       \
67                 PKT_TX_TUNNEL_MASK |    \
68                 I40E_TX_IEEE1588_TMST)
69
70 #define I40E_TX_OFFLOAD_NOTSUP_MASK \
71                 (PKT_TX_OFFLOAD_MASK ^ I40E_TX_OFFLOAD_MASK)
72
73 static inline void
74 i40e_rxd_to_vlan_tci(struct rte_mbuf *mb, volatile union i40e_rx_desc *rxdp)
75 {
76         if (rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len) &
77                 (1 << I40E_RX_DESC_STATUS_L2TAG1P_SHIFT)) {
78                 mb->ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
79                 mb->vlan_tci =
80                         rte_le_to_cpu_16(rxdp->wb.qword0.lo_dword.l2tag1);
81                 PMD_RX_LOG(DEBUG, "Descriptor l2tag1: %u",
82                            rte_le_to_cpu_16(rxdp->wb.qword0.lo_dword.l2tag1));
83         } else {
84                 mb->vlan_tci = 0;
85         }
86 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
87         if (rte_le_to_cpu_16(rxdp->wb.qword2.ext_status) &
88                 (1 << I40E_RX_DESC_EXT_STATUS_L2TAG2P_SHIFT)) {
89                 mb->ol_flags |= PKT_RX_QINQ_STRIPPED;
90                 mb->vlan_tci_outer = mb->vlan_tci;
91                 mb->vlan_tci = rte_le_to_cpu_16(rxdp->wb.qword2.l2tag2_2);
92                 PMD_RX_LOG(DEBUG, "Descriptor l2tag2_1: %u, l2tag2_2: %u",
93                            rte_le_to_cpu_16(rxdp->wb.qword2.l2tag2_1),
94                            rte_le_to_cpu_16(rxdp->wb.qword2.l2tag2_2));
95         } else {
96                 mb->vlan_tci_outer = 0;
97         }
98 #endif
99         PMD_RX_LOG(DEBUG, "Mbuf vlan_tci: %u, vlan_tci_outer: %u",
100                    mb->vlan_tci, mb->vlan_tci_outer);
101 }
102
103 /* Translate the rx descriptor status to pkt flags */
104 static inline uint64_t
105 i40e_rxd_status_to_pkt_flags(uint64_t qword)
106 {
107         uint64_t flags;
108
109         /* Check if RSS_HASH */
110         flags = (((qword >> I40E_RX_DESC_STATUS_FLTSTAT_SHIFT) &
111                                         I40E_RX_DESC_FLTSTAT_RSS_HASH) ==
112                         I40E_RX_DESC_FLTSTAT_RSS_HASH) ? PKT_RX_RSS_HASH : 0;
113
114         /* Check if FDIR Match */
115         flags |= (qword & (1 << I40E_RX_DESC_STATUS_FLM_SHIFT) ?
116                                                         PKT_RX_FDIR : 0);
117
118         return flags;
119 }
120
121 static inline uint64_t
122 i40e_rxd_error_to_pkt_flags(uint64_t qword)
123 {
124         uint64_t flags = 0;
125         uint64_t error_bits = (qword >> I40E_RXD_QW1_ERROR_SHIFT);
126
127 #define I40E_RX_ERR_BITS 0x3f
128         if (likely((error_bits & I40E_RX_ERR_BITS) == 0)) {
129                 flags |= (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD);
130                 return flags;
131         }
132
133         if (unlikely(error_bits & (1 << I40E_RX_DESC_ERROR_IPE_SHIFT)))
134                 flags |= PKT_RX_IP_CKSUM_BAD;
135         else
136                 flags |= PKT_RX_IP_CKSUM_GOOD;
137
138         if (unlikely(error_bits & (1 << I40E_RX_DESC_ERROR_L4E_SHIFT)))
139                 flags |= PKT_RX_L4_CKSUM_BAD;
140         else
141                 flags |= PKT_RX_L4_CKSUM_GOOD;
142
143         if (unlikely(error_bits & (1 << I40E_RX_DESC_ERROR_EIPE_SHIFT)))
144                 flags |= PKT_RX_EIP_CKSUM_BAD;
145
146         return flags;
147 }
148
149 /* Function to check and set the ieee1588 timesync index and get the
150  * appropriate flags.
151  */
152 #ifdef RTE_LIBRTE_IEEE1588
153 static inline uint64_t
154 i40e_get_iee15888_flags(struct rte_mbuf *mb, uint64_t qword)
155 {
156         uint64_t pkt_flags = 0;
157         uint16_t tsyn = (qword & (I40E_RXD_QW1_STATUS_TSYNVALID_MASK
158                                   | I40E_RXD_QW1_STATUS_TSYNINDX_MASK))
159                                     >> I40E_RX_DESC_STATUS_TSYNINDX_SHIFT;
160
161         if ((mb->packet_type & RTE_PTYPE_L2_MASK)
162                         == RTE_PTYPE_L2_ETHER_TIMESYNC)
163                 pkt_flags = PKT_RX_IEEE1588_PTP;
164         if (tsyn & 0x04) {
165                 pkt_flags |= PKT_RX_IEEE1588_TMST;
166                 mb->timesync = tsyn & 0x03;
167         }
168
169         return pkt_flags;
170 }
171 #endif
172
173 #define I40E_RX_DESC_EXT_STATUS_FLEXBH_MASK   0x03
174 #define I40E_RX_DESC_EXT_STATUS_FLEXBH_FD_ID  0x01
175 #define I40E_RX_DESC_EXT_STATUS_FLEXBH_FLEX   0x02
176 #define I40E_RX_DESC_EXT_STATUS_FLEXBL_MASK   0x03
177 #define I40E_RX_DESC_EXT_STATUS_FLEXBL_FLEX   0x01
178
179 static inline uint64_t
180 i40e_rxd_build_fdir(volatile union i40e_rx_desc *rxdp, struct rte_mbuf *mb)
181 {
182         uint64_t flags = 0;
183 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
184         uint16_t flexbh, flexbl;
185
186         flexbh = (rte_le_to_cpu_32(rxdp->wb.qword2.ext_status) >>
187                 I40E_RX_DESC_EXT_STATUS_FLEXBH_SHIFT) &
188                 I40E_RX_DESC_EXT_STATUS_FLEXBH_MASK;
189         flexbl = (rte_le_to_cpu_32(rxdp->wb.qword2.ext_status) >>
190                 I40E_RX_DESC_EXT_STATUS_FLEXBL_SHIFT) &
191                 I40E_RX_DESC_EXT_STATUS_FLEXBL_MASK;
192
193
194         if (flexbh == I40E_RX_DESC_EXT_STATUS_FLEXBH_FD_ID) {
195                 mb->hash.fdir.hi =
196                         rte_le_to_cpu_32(rxdp->wb.qword3.hi_dword.fd_id);
197                 flags |= PKT_RX_FDIR_ID;
198         } else if (flexbh == I40E_RX_DESC_EXT_STATUS_FLEXBH_FLEX) {
199                 mb->hash.fdir.hi =
200                         rte_le_to_cpu_32(rxdp->wb.qword3.hi_dword.flex_bytes_hi);
201                 flags |= PKT_RX_FDIR_FLX;
202         }
203         if (flexbl == I40E_RX_DESC_EXT_STATUS_FLEXBL_FLEX) {
204                 mb->hash.fdir.lo =
205                         rte_le_to_cpu_32(rxdp->wb.qword3.lo_dword.flex_bytes_lo);
206                 flags |= PKT_RX_FDIR_FLX;
207         }
208 #else
209         mb->hash.fdir.hi =
210                 rte_le_to_cpu_32(rxdp->wb.qword0.hi_dword.fd_id);
211         flags |= PKT_RX_FDIR_ID;
212 #endif
213         return flags;
214 }
215
216 static inline void
217 i40e_parse_tunneling_params(uint64_t ol_flags,
218                             union i40e_tx_offload tx_offload,
219                             uint32_t *cd_tunneling)
220 {
221         /* EIPT: External (outer) IP header type */
222         if (ol_flags & PKT_TX_OUTER_IP_CKSUM)
223                 *cd_tunneling |= I40E_TX_CTX_EXT_IP_IPV4;
224         else if (ol_flags & PKT_TX_OUTER_IPV4)
225                 *cd_tunneling |= I40E_TX_CTX_EXT_IP_IPV4_NO_CSUM;
226         else if (ol_flags & PKT_TX_OUTER_IPV6)
227                 *cd_tunneling |= I40E_TX_CTX_EXT_IP_IPV6;
228
229         /* EIPLEN: External (outer) IP header length, in DWords */
230         *cd_tunneling |= (tx_offload.outer_l3_len >> 2) <<
231                 I40E_TXD_CTX_QW0_EXT_IPLEN_SHIFT;
232
233         /* L4TUNT: L4 Tunneling Type */
234         switch (ol_flags & PKT_TX_TUNNEL_MASK) {
235         case PKT_TX_TUNNEL_IPIP:
236                 /* for non UDP / GRE tunneling, set to 00b */
237                 break;
238         case PKT_TX_TUNNEL_VXLAN:
239         case PKT_TX_TUNNEL_GENEVE:
240                 *cd_tunneling |= I40E_TXD_CTX_UDP_TUNNELING;
241                 break;
242         case PKT_TX_TUNNEL_GRE:
243                 *cd_tunneling |= I40E_TXD_CTX_GRE_TUNNELING;
244                 break;
245         default:
246                 PMD_TX_LOG(ERR, "Tunnel type not supported");
247                 return;
248         }
249
250         /* L4TUNLEN: L4 Tunneling Length, in Words
251          *
252          * We depend on app to set rte_mbuf.l2_len correctly.
253          * For IP in GRE it should be set to the length of the GRE
254          * header;
255          * for MAC in GRE or MAC in UDP it should be set to the length
256          * of the GRE or UDP headers plus the inner MAC up to including
257          * its last Ethertype.
258          */
259         *cd_tunneling |= (tx_offload.l2_len >> 1) <<
260                 I40E_TXD_CTX_QW0_NATLEN_SHIFT;
261 }
262
263 static inline void
264 i40e_txd_enable_checksum(uint64_t ol_flags,
265                         uint32_t *td_cmd,
266                         uint32_t *td_offset,
267                         union i40e_tx_offload tx_offload)
268 {
269         /* Set MACLEN */
270         if (ol_flags & PKT_TX_TUNNEL_MASK)
271                 *td_offset |= (tx_offload.outer_l2_len >> 1)
272                                 << I40E_TX_DESC_LENGTH_MACLEN_SHIFT;
273         else
274                 *td_offset |= (tx_offload.l2_len >> 1)
275                         << I40E_TX_DESC_LENGTH_MACLEN_SHIFT;
276
277         /* Enable L3 checksum offloads */
278         if (ol_flags & PKT_TX_IP_CKSUM) {
279                 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV4_CSUM;
280                 *td_offset |= (tx_offload.l3_len >> 2)
281                                 << I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
282         } else if (ol_flags & PKT_TX_IPV4) {
283                 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV4;
284                 *td_offset |= (tx_offload.l3_len >> 2)
285                                 << I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
286         } else if (ol_flags & PKT_TX_IPV6) {
287                 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV6;
288                 *td_offset |= (tx_offload.l3_len >> 2)
289                                 << I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
290         }
291
292         if (ol_flags & PKT_TX_TCP_SEG) {
293                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_TCP;
294                 *td_offset |= (tx_offload.l4_len >> 2)
295                         << I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
296                 return;
297         }
298
299         /* Enable L4 checksum offloads */
300         switch (ol_flags & PKT_TX_L4_MASK) {
301         case PKT_TX_TCP_CKSUM:
302                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_TCP;
303                 *td_offset |= (sizeof(struct tcp_hdr) >> 2) <<
304                                 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
305                 break;
306         case PKT_TX_SCTP_CKSUM:
307                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_SCTP;
308                 *td_offset |= (sizeof(struct sctp_hdr) >> 2) <<
309                                 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
310                 break;
311         case PKT_TX_UDP_CKSUM:
312                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_UDP;
313                 *td_offset |= (sizeof(struct udp_hdr) >> 2) <<
314                                 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
315                 break;
316         default:
317                 break;
318         }
319 }
320
321 /* Construct the tx flags */
322 static inline uint64_t
323 i40e_build_ctob(uint32_t td_cmd,
324                 uint32_t td_offset,
325                 unsigned int size,
326                 uint32_t td_tag)
327 {
328         return rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DATA |
329                         ((uint64_t)td_cmd  << I40E_TXD_QW1_CMD_SHIFT) |
330                         ((uint64_t)td_offset << I40E_TXD_QW1_OFFSET_SHIFT) |
331                         ((uint64_t)size  << I40E_TXD_QW1_TX_BUF_SZ_SHIFT) |
332                         ((uint64_t)td_tag  << I40E_TXD_QW1_L2TAG1_SHIFT));
333 }
334
335 static inline int
336 i40e_xmit_cleanup(struct i40e_tx_queue *txq)
337 {
338         struct i40e_tx_entry *sw_ring = txq->sw_ring;
339         volatile struct i40e_tx_desc *txd = txq->tx_ring;
340         uint16_t last_desc_cleaned = txq->last_desc_cleaned;
341         uint16_t nb_tx_desc = txq->nb_tx_desc;
342         uint16_t desc_to_clean_to;
343         uint16_t nb_tx_to_clean;
344
345         desc_to_clean_to = (uint16_t)(last_desc_cleaned + txq->tx_rs_thresh);
346         if (desc_to_clean_to >= nb_tx_desc)
347                 desc_to_clean_to = (uint16_t)(desc_to_clean_to - nb_tx_desc);
348
349         desc_to_clean_to = sw_ring[desc_to_clean_to].last_id;
350         if ((txd[desc_to_clean_to].cmd_type_offset_bsz &
351                         rte_cpu_to_le_64(I40E_TXD_QW1_DTYPE_MASK)) !=
352                         rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DESC_DONE)) {
353                 PMD_TX_FREE_LOG(DEBUG, "TX descriptor %4u is not done "
354                         "(port=%d queue=%d)", desc_to_clean_to,
355                                 txq->port_id, txq->queue_id);
356                 return -1;
357         }
358
359         if (last_desc_cleaned > desc_to_clean_to)
360                 nb_tx_to_clean = (uint16_t)((nb_tx_desc - last_desc_cleaned) +
361                                                         desc_to_clean_to);
362         else
363                 nb_tx_to_clean = (uint16_t)(desc_to_clean_to -
364                                         last_desc_cleaned);
365
366         txd[desc_to_clean_to].cmd_type_offset_bsz = 0;
367
368         txq->last_desc_cleaned = desc_to_clean_to;
369         txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + nb_tx_to_clean);
370
371         return 0;
372 }
373
374 static inline int
375 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
376 check_rx_burst_bulk_alloc_preconditions(struct i40e_rx_queue *rxq)
377 #else
378 check_rx_burst_bulk_alloc_preconditions(__rte_unused struct i40e_rx_queue *rxq)
379 #endif
380 {
381         int ret = 0;
382
383 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
384         if (!(rxq->rx_free_thresh >= RTE_PMD_I40E_RX_MAX_BURST)) {
385                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
386                              "rxq->rx_free_thresh=%d, "
387                              "RTE_PMD_I40E_RX_MAX_BURST=%d",
388                              rxq->rx_free_thresh, RTE_PMD_I40E_RX_MAX_BURST);
389                 ret = -EINVAL;
390         } else if (!(rxq->rx_free_thresh < rxq->nb_rx_desc)) {
391                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
392                              "rxq->rx_free_thresh=%d, "
393                              "rxq->nb_rx_desc=%d",
394                              rxq->rx_free_thresh, rxq->nb_rx_desc);
395                 ret = -EINVAL;
396         } else if (rxq->nb_rx_desc % rxq->rx_free_thresh != 0) {
397                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
398                              "rxq->nb_rx_desc=%d, "
399                              "rxq->rx_free_thresh=%d",
400                              rxq->nb_rx_desc, rxq->rx_free_thresh);
401                 ret = -EINVAL;
402         }
403 #else
404         ret = -EINVAL;
405 #endif
406
407         return ret;
408 }
409
410 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
411 #define I40E_LOOK_AHEAD 8
412 #if (I40E_LOOK_AHEAD != 8)
413 #error "PMD I40E: I40E_LOOK_AHEAD must be 8\n"
414 #endif
415 static inline int
416 i40e_rx_scan_hw_ring(struct i40e_rx_queue *rxq)
417 {
418         volatile union i40e_rx_desc *rxdp;
419         struct i40e_rx_entry *rxep;
420         struct rte_mbuf *mb;
421         uint16_t pkt_len;
422         uint64_t qword1;
423         uint32_t rx_status;
424         int32_t s[I40E_LOOK_AHEAD], nb_dd;
425         int32_t i, j, nb_rx = 0;
426         uint64_t pkt_flags;
427         uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
428
429         rxdp = &rxq->rx_ring[rxq->rx_tail];
430         rxep = &rxq->sw_ring[rxq->rx_tail];
431
432         qword1 = rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len);
433         rx_status = (qword1 & I40E_RXD_QW1_STATUS_MASK) >>
434                                 I40E_RXD_QW1_STATUS_SHIFT;
435
436         /* Make sure there is at least 1 packet to receive */
437         if (!(rx_status & (1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
438                 return 0;
439
440         /**
441          * Scan LOOK_AHEAD descriptors at a time to determine which
442          * descriptors reference packets that are ready to be received.
443          */
444         for (i = 0; i < RTE_PMD_I40E_RX_MAX_BURST; i+=I40E_LOOK_AHEAD,
445                         rxdp += I40E_LOOK_AHEAD, rxep += I40E_LOOK_AHEAD) {
446                 /* Read desc statuses backwards to avoid race condition */
447                 for (j = I40E_LOOK_AHEAD - 1; j >= 0; j--) {
448                         qword1 = rte_le_to_cpu_64(\
449                                 rxdp[j].wb.qword1.status_error_len);
450                         s[j] = (qword1 & I40E_RXD_QW1_STATUS_MASK) >>
451                                         I40E_RXD_QW1_STATUS_SHIFT;
452                 }
453
454                 rte_smp_rmb();
455
456                 /* Compute how many status bits were set */
457                 for (j = 0, nb_dd = 0; j < I40E_LOOK_AHEAD; j++)
458                         nb_dd += s[j] & (1 << I40E_RX_DESC_STATUS_DD_SHIFT);
459
460                 nb_rx += nb_dd;
461
462                 /* Translate descriptor info to mbuf parameters */
463                 for (j = 0; j < nb_dd; j++) {
464                         mb = rxep[j].mbuf;
465                         qword1 = rte_le_to_cpu_64(\
466                                 rxdp[j].wb.qword1.status_error_len);
467                         pkt_len = ((qword1 & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
468                                 I40E_RXD_QW1_LENGTH_PBUF_SHIFT) - rxq->crc_len;
469                         mb->data_len = pkt_len;
470                         mb->pkt_len = pkt_len;
471                         mb->ol_flags = 0;
472                         i40e_rxd_to_vlan_tci(mb, &rxdp[j]);
473                         pkt_flags = i40e_rxd_status_to_pkt_flags(qword1);
474                         pkt_flags |= i40e_rxd_error_to_pkt_flags(qword1);
475                         mb->packet_type =
476                                 ptype_tbl[(uint8_t)((qword1 &
477                                 I40E_RXD_QW1_PTYPE_MASK) >>
478                                 I40E_RXD_QW1_PTYPE_SHIFT)];
479                         if (pkt_flags & PKT_RX_RSS_HASH)
480                                 mb->hash.rss = rte_le_to_cpu_32(\
481                                         rxdp[j].wb.qword0.hi_dword.rss);
482                         if (pkt_flags & PKT_RX_FDIR)
483                                 pkt_flags |= i40e_rxd_build_fdir(&rxdp[j], mb);
484
485 #ifdef RTE_LIBRTE_IEEE1588
486                         pkt_flags |= i40e_get_iee15888_flags(mb, qword1);
487 #endif
488                         mb->ol_flags |= pkt_flags;
489
490                 }
491
492                 for (j = 0; j < I40E_LOOK_AHEAD; j++)
493                         rxq->rx_stage[i + j] = rxep[j].mbuf;
494
495                 if (nb_dd != I40E_LOOK_AHEAD)
496                         break;
497         }
498
499         /* Clear software ring entries */
500         for (i = 0; i < nb_rx; i++)
501                 rxq->sw_ring[rxq->rx_tail + i].mbuf = NULL;
502
503         return nb_rx;
504 }
505
506 static inline uint16_t
507 i40e_rx_fill_from_stage(struct i40e_rx_queue *rxq,
508                         struct rte_mbuf **rx_pkts,
509                         uint16_t nb_pkts)
510 {
511         uint16_t i;
512         struct rte_mbuf **stage = &rxq->rx_stage[rxq->rx_next_avail];
513
514         nb_pkts = (uint16_t)RTE_MIN(nb_pkts, rxq->rx_nb_avail);
515
516         for (i = 0; i < nb_pkts; i++)
517                 rx_pkts[i] = stage[i];
518
519         rxq->rx_nb_avail = (uint16_t)(rxq->rx_nb_avail - nb_pkts);
520         rxq->rx_next_avail = (uint16_t)(rxq->rx_next_avail + nb_pkts);
521
522         return nb_pkts;
523 }
524
525 static inline int
526 i40e_rx_alloc_bufs(struct i40e_rx_queue *rxq)
527 {
528         volatile union i40e_rx_desc *rxdp;
529         struct i40e_rx_entry *rxep;
530         struct rte_mbuf *mb;
531         uint16_t alloc_idx, i;
532         uint64_t dma_addr;
533         int diag;
534
535         /* Allocate buffers in bulk */
536         alloc_idx = (uint16_t)(rxq->rx_free_trigger -
537                                 (rxq->rx_free_thresh - 1));
538         rxep = &(rxq->sw_ring[alloc_idx]);
539         diag = rte_mempool_get_bulk(rxq->mp, (void *)rxep,
540                                         rxq->rx_free_thresh);
541         if (unlikely(diag != 0)) {
542                 PMD_DRV_LOG(ERR, "Failed to get mbufs in bulk");
543                 return -ENOMEM;
544         }
545
546         rxdp = &rxq->rx_ring[alloc_idx];
547         for (i = 0; i < rxq->rx_free_thresh; i++) {
548                 if (likely(i < (rxq->rx_free_thresh - 1)))
549                         /* Prefetch next mbuf */
550                         rte_prefetch0(rxep[i + 1].mbuf);
551
552                 mb = rxep[i].mbuf;
553                 rte_mbuf_refcnt_set(mb, 1);
554                 mb->next = NULL;
555                 mb->data_off = RTE_PKTMBUF_HEADROOM;
556                 mb->nb_segs = 1;
557                 mb->port = rxq->port_id;
558                 dma_addr = rte_cpu_to_le_64(\
559                         rte_mbuf_data_iova_default(mb));
560                 rxdp[i].read.hdr_addr = 0;
561                 rxdp[i].read.pkt_addr = dma_addr;
562         }
563
564         /* Update rx tail regsiter */
565         rte_wmb();
566         I40E_PCI_REG_WRITE_RELAXED(rxq->qrx_tail, rxq->rx_free_trigger);
567
568         rxq->rx_free_trigger =
569                 (uint16_t)(rxq->rx_free_trigger + rxq->rx_free_thresh);
570         if (rxq->rx_free_trigger >= rxq->nb_rx_desc)
571                 rxq->rx_free_trigger = (uint16_t)(rxq->rx_free_thresh - 1);
572
573         return 0;
574 }
575
576 static inline uint16_t
577 rx_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
578 {
579         struct i40e_rx_queue *rxq = (struct i40e_rx_queue *)rx_queue;
580         struct rte_eth_dev *dev;
581         uint16_t nb_rx = 0;
582
583         if (!nb_pkts)
584                 return 0;
585
586         if (rxq->rx_nb_avail)
587                 return i40e_rx_fill_from_stage(rxq, rx_pkts, nb_pkts);
588
589         nb_rx = (uint16_t)i40e_rx_scan_hw_ring(rxq);
590         rxq->rx_next_avail = 0;
591         rxq->rx_nb_avail = nb_rx;
592         rxq->rx_tail = (uint16_t)(rxq->rx_tail + nb_rx);
593
594         if (rxq->rx_tail > rxq->rx_free_trigger) {
595                 if (i40e_rx_alloc_bufs(rxq) != 0) {
596                         uint16_t i, j;
597
598                         dev = I40E_VSI_TO_ETH_DEV(rxq->vsi);
599                         dev->data->rx_mbuf_alloc_failed +=
600                                 rxq->rx_free_thresh;
601
602                         rxq->rx_nb_avail = 0;
603                         rxq->rx_tail = (uint16_t)(rxq->rx_tail - nb_rx);
604                         for (i = 0, j = rxq->rx_tail; i < nb_rx; i++, j++)
605                                 rxq->sw_ring[j].mbuf = rxq->rx_stage[i];
606
607                         return 0;
608                 }
609         }
610
611         if (rxq->rx_tail >= rxq->nb_rx_desc)
612                 rxq->rx_tail = 0;
613
614         if (rxq->rx_nb_avail)
615                 return i40e_rx_fill_from_stage(rxq, rx_pkts, nb_pkts);
616
617         return 0;
618 }
619
620 static uint16_t
621 i40e_recv_pkts_bulk_alloc(void *rx_queue,
622                           struct rte_mbuf **rx_pkts,
623                           uint16_t nb_pkts)
624 {
625         uint16_t nb_rx = 0, n, count;
626
627         if (unlikely(nb_pkts == 0))
628                 return 0;
629
630         if (likely(nb_pkts <= RTE_PMD_I40E_RX_MAX_BURST))
631                 return rx_recv_pkts(rx_queue, rx_pkts, nb_pkts);
632
633         while (nb_pkts) {
634                 n = RTE_MIN(nb_pkts, RTE_PMD_I40E_RX_MAX_BURST);
635                 count = rx_recv_pkts(rx_queue, &rx_pkts[nb_rx], n);
636                 nb_rx = (uint16_t)(nb_rx + count);
637                 nb_pkts = (uint16_t)(nb_pkts - count);
638                 if (count < n)
639                         break;
640         }
641
642         return nb_rx;
643 }
644 #else
645 static uint16_t
646 i40e_recv_pkts_bulk_alloc(void __rte_unused *rx_queue,
647                           struct rte_mbuf __rte_unused **rx_pkts,
648                           uint16_t __rte_unused nb_pkts)
649 {
650         return 0;
651 }
652 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
653
654 uint16_t
655 i40e_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
656 {
657         struct i40e_rx_queue *rxq;
658         volatile union i40e_rx_desc *rx_ring;
659         volatile union i40e_rx_desc *rxdp;
660         union i40e_rx_desc rxd;
661         struct i40e_rx_entry *sw_ring;
662         struct i40e_rx_entry *rxe;
663         struct rte_eth_dev *dev;
664         struct rte_mbuf *rxm;
665         struct rte_mbuf *nmb;
666         uint16_t nb_rx;
667         uint32_t rx_status;
668         uint64_t qword1;
669         uint16_t rx_packet_len;
670         uint16_t rx_id, nb_hold;
671         uint64_t dma_addr;
672         uint64_t pkt_flags;
673         uint32_t *ptype_tbl;
674
675         nb_rx = 0;
676         nb_hold = 0;
677         rxq = rx_queue;
678         rx_id = rxq->rx_tail;
679         rx_ring = rxq->rx_ring;
680         sw_ring = rxq->sw_ring;
681         ptype_tbl = rxq->vsi->adapter->ptype_tbl;
682
683         while (nb_rx < nb_pkts) {
684                 rxdp = &rx_ring[rx_id];
685                 qword1 = rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len);
686                 rx_status = (qword1 & I40E_RXD_QW1_STATUS_MASK)
687                                 >> I40E_RXD_QW1_STATUS_SHIFT;
688
689                 /* Check the DD bit first */
690                 if (!(rx_status & (1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
691                         break;
692
693                 nmb = rte_mbuf_raw_alloc(rxq->mp);
694                 if (unlikely(!nmb)) {
695                         dev = I40E_VSI_TO_ETH_DEV(rxq->vsi);
696                         dev->data->rx_mbuf_alloc_failed++;
697                         break;
698                 }
699
700                 rxd = *rxdp;
701                 nb_hold++;
702                 rxe = &sw_ring[rx_id];
703                 rx_id++;
704                 if (unlikely(rx_id == rxq->nb_rx_desc))
705                         rx_id = 0;
706
707                 /* Prefetch next mbuf */
708                 rte_prefetch0(sw_ring[rx_id].mbuf);
709
710                 /**
711                  * When next RX descriptor is on a cache line boundary,
712                  * prefetch the next 4 RX descriptors and next 8 pointers
713                  * to mbufs.
714                  */
715                 if ((rx_id & 0x3) == 0) {
716                         rte_prefetch0(&rx_ring[rx_id]);
717                         rte_prefetch0(&sw_ring[rx_id]);
718                 }
719                 rxm = rxe->mbuf;
720                 rxe->mbuf = nmb;
721                 dma_addr =
722                         rte_cpu_to_le_64(rte_mbuf_data_iova_default(nmb));
723                 rxdp->read.hdr_addr = 0;
724                 rxdp->read.pkt_addr = dma_addr;
725
726                 rx_packet_len = ((qword1 & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
727                                 I40E_RXD_QW1_LENGTH_PBUF_SHIFT) - rxq->crc_len;
728
729                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
730                 rte_prefetch0(RTE_PTR_ADD(rxm->buf_addr, RTE_PKTMBUF_HEADROOM));
731                 rxm->nb_segs = 1;
732                 rxm->next = NULL;
733                 rxm->pkt_len = rx_packet_len;
734                 rxm->data_len = rx_packet_len;
735                 rxm->port = rxq->port_id;
736                 rxm->ol_flags = 0;
737                 i40e_rxd_to_vlan_tci(rxm, &rxd);
738                 pkt_flags = i40e_rxd_status_to_pkt_flags(qword1);
739                 pkt_flags |= i40e_rxd_error_to_pkt_flags(qword1);
740                 rxm->packet_type =
741                         ptype_tbl[(uint8_t)((qword1 &
742                         I40E_RXD_QW1_PTYPE_MASK) >> I40E_RXD_QW1_PTYPE_SHIFT)];
743                 if (pkt_flags & PKT_RX_RSS_HASH)
744                         rxm->hash.rss =
745                                 rte_le_to_cpu_32(rxd.wb.qword0.hi_dword.rss);
746                 if (pkt_flags & PKT_RX_FDIR)
747                         pkt_flags |= i40e_rxd_build_fdir(&rxd, rxm);
748
749 #ifdef RTE_LIBRTE_IEEE1588
750                 pkt_flags |= i40e_get_iee15888_flags(rxm, qword1);
751 #endif
752                 rxm->ol_flags |= pkt_flags;
753
754                 rx_pkts[nb_rx++] = rxm;
755         }
756         rxq->rx_tail = rx_id;
757
758         /**
759          * If the number of free RX descriptors is greater than the RX free
760          * threshold of the queue, advance the receive tail register of queue.
761          * Update that register with the value of the last processed RX
762          * descriptor minus 1.
763          */
764         nb_hold = (uint16_t)(nb_hold + rxq->nb_rx_hold);
765         if (nb_hold > rxq->rx_free_thresh) {
766                 rx_id = (uint16_t) ((rx_id == 0) ?
767                         (rxq->nb_rx_desc - 1) : (rx_id - 1));
768                 I40E_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
769                 nb_hold = 0;
770         }
771         rxq->nb_rx_hold = nb_hold;
772
773         return nb_rx;
774 }
775
776 uint16_t
777 i40e_recv_scattered_pkts(void *rx_queue,
778                          struct rte_mbuf **rx_pkts,
779                          uint16_t nb_pkts)
780 {
781         struct i40e_rx_queue *rxq = rx_queue;
782         volatile union i40e_rx_desc *rx_ring = rxq->rx_ring;
783         volatile union i40e_rx_desc *rxdp;
784         union i40e_rx_desc rxd;
785         struct i40e_rx_entry *sw_ring = rxq->sw_ring;
786         struct i40e_rx_entry *rxe;
787         struct rte_mbuf *first_seg = rxq->pkt_first_seg;
788         struct rte_mbuf *last_seg = rxq->pkt_last_seg;
789         struct rte_mbuf *nmb, *rxm;
790         uint16_t rx_id = rxq->rx_tail;
791         uint16_t nb_rx = 0, nb_hold = 0, rx_packet_len;
792         struct rte_eth_dev *dev;
793         uint32_t rx_status;
794         uint64_t qword1;
795         uint64_t dma_addr;
796         uint64_t pkt_flags;
797         uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl;
798
799         while (nb_rx < nb_pkts) {
800                 rxdp = &rx_ring[rx_id];
801                 qword1 = rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len);
802                 rx_status = (qword1 & I40E_RXD_QW1_STATUS_MASK) >>
803                                         I40E_RXD_QW1_STATUS_SHIFT;
804
805                 /* Check the DD bit */
806                 if (!(rx_status & (1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
807                         break;
808
809                 nmb = rte_mbuf_raw_alloc(rxq->mp);
810                 if (unlikely(!nmb)) {
811                         dev = I40E_VSI_TO_ETH_DEV(rxq->vsi);
812                         dev->data->rx_mbuf_alloc_failed++;
813                         break;
814                 }
815
816                 rxd = *rxdp;
817                 nb_hold++;
818                 rxe = &sw_ring[rx_id];
819                 rx_id++;
820                 if (rx_id == rxq->nb_rx_desc)
821                         rx_id = 0;
822
823                 /* Prefetch next mbuf */
824                 rte_prefetch0(sw_ring[rx_id].mbuf);
825
826                 /**
827                  * When next RX descriptor is on a cache line boundary,
828                  * prefetch the next 4 RX descriptors and next 8 pointers
829                  * to mbufs.
830                  */
831                 if ((rx_id & 0x3) == 0) {
832                         rte_prefetch0(&rx_ring[rx_id]);
833                         rte_prefetch0(&sw_ring[rx_id]);
834                 }
835
836                 rxm = rxe->mbuf;
837                 rxe->mbuf = nmb;
838                 dma_addr =
839                         rte_cpu_to_le_64(rte_mbuf_data_iova_default(nmb));
840
841                 /* Set data buffer address and data length of the mbuf */
842                 rxdp->read.hdr_addr = 0;
843                 rxdp->read.pkt_addr = dma_addr;
844                 rx_packet_len = (qword1 & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
845                                         I40E_RXD_QW1_LENGTH_PBUF_SHIFT;
846                 rxm->data_len = rx_packet_len;
847                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
848
849                 /**
850                  * If this is the first buffer of the received packet, set the
851                  * pointer to the first mbuf of the packet and initialize its
852                  * context. Otherwise, update the total length and the number
853                  * of segments of the current scattered packet, and update the
854                  * pointer to the last mbuf of the current packet.
855                  */
856                 if (!first_seg) {
857                         first_seg = rxm;
858                         first_seg->nb_segs = 1;
859                         first_seg->pkt_len = rx_packet_len;
860                 } else {
861                         first_seg->pkt_len =
862                                 (uint16_t)(first_seg->pkt_len +
863                                                 rx_packet_len);
864                         first_seg->nb_segs++;
865                         last_seg->next = rxm;
866                 }
867
868                 /**
869                  * If this is not the last buffer of the received packet,
870                  * update the pointer to the last mbuf of the current scattered
871                  * packet and continue to parse the RX ring.
872                  */
873                 if (!(rx_status & (1 << I40E_RX_DESC_STATUS_EOF_SHIFT))) {
874                         last_seg = rxm;
875                         continue;
876                 }
877
878                 /**
879                  * This is the last buffer of the received packet. If the CRC
880                  * is not stripped by the hardware:
881                  *  - Subtract the CRC length from the total packet length.
882                  *  - If the last buffer only contains the whole CRC or a part
883                  *  of it, free the mbuf associated to the last buffer. If part
884                  *  of the CRC is also contained in the previous mbuf, subtract
885                  *  the length of that CRC part from the data length of the
886                  *  previous mbuf.
887                  */
888                 rxm->next = NULL;
889                 if (unlikely(rxq->crc_len > 0)) {
890                         first_seg->pkt_len -= ETHER_CRC_LEN;
891                         if (rx_packet_len <= ETHER_CRC_LEN) {
892                                 rte_pktmbuf_free_seg(rxm);
893                                 first_seg->nb_segs--;
894                                 last_seg->data_len =
895                                         (uint16_t)(last_seg->data_len -
896                                         (ETHER_CRC_LEN - rx_packet_len));
897                                 last_seg->next = NULL;
898                         } else
899                                 rxm->data_len = (uint16_t)(rx_packet_len -
900                                                                 ETHER_CRC_LEN);
901                 }
902
903                 first_seg->port = rxq->port_id;
904                 first_seg->ol_flags = 0;
905                 i40e_rxd_to_vlan_tci(first_seg, &rxd);
906                 pkt_flags = i40e_rxd_status_to_pkt_flags(qword1);
907                 pkt_flags |= i40e_rxd_error_to_pkt_flags(qword1);
908                 first_seg->packet_type =
909                         ptype_tbl[(uint8_t)((qword1 &
910                         I40E_RXD_QW1_PTYPE_MASK) >> I40E_RXD_QW1_PTYPE_SHIFT)];
911                 if (pkt_flags & PKT_RX_RSS_HASH)
912                         first_seg->hash.rss =
913                                 rte_le_to_cpu_32(rxd.wb.qword0.hi_dword.rss);
914                 if (pkt_flags & PKT_RX_FDIR)
915                         pkt_flags |= i40e_rxd_build_fdir(&rxd, first_seg);
916
917 #ifdef RTE_LIBRTE_IEEE1588
918                 pkt_flags |= i40e_get_iee15888_flags(first_seg, qword1);
919 #endif
920                 first_seg->ol_flags |= pkt_flags;
921
922                 /* Prefetch data of first segment, if configured to do so. */
923                 rte_prefetch0(RTE_PTR_ADD(first_seg->buf_addr,
924                         first_seg->data_off));
925                 rx_pkts[nb_rx++] = first_seg;
926                 first_seg = NULL;
927         }
928
929         /* Record index of the next RX descriptor to probe. */
930         rxq->rx_tail = rx_id;
931         rxq->pkt_first_seg = first_seg;
932         rxq->pkt_last_seg = last_seg;
933
934         /**
935          * If the number of free RX descriptors is greater than the RX free
936          * threshold of the queue, advance the Receive Descriptor Tail (RDT)
937          * register. Update the RDT with the value of the last processed RX
938          * descriptor minus 1, to guarantee that the RDT register is never
939          * equal to the RDH register, which creates a "full" ring situtation
940          * from the hardware point of view.
941          */
942         nb_hold = (uint16_t)(nb_hold + rxq->nb_rx_hold);
943         if (nb_hold > rxq->rx_free_thresh) {
944                 rx_id = (uint16_t)(rx_id == 0 ?
945                         (rxq->nb_rx_desc - 1) : (rx_id - 1));
946                 I40E_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
947                 nb_hold = 0;
948         }
949         rxq->nb_rx_hold = nb_hold;
950
951         return nb_rx;
952 }
953
954 /* Check if the context descriptor is needed for TX offloading */
955 static inline uint16_t
956 i40e_calc_context_desc(uint64_t flags)
957 {
958         static uint64_t mask = PKT_TX_OUTER_IP_CKSUM |
959                 PKT_TX_TCP_SEG |
960                 PKT_TX_QINQ_PKT |
961                 PKT_TX_TUNNEL_MASK;
962
963 #ifdef RTE_LIBRTE_IEEE1588
964         mask |= PKT_TX_IEEE1588_TMST;
965 #endif
966
967         return (flags & mask) ? 1 : 0;
968 }
969
970 /* set i40e TSO context descriptor */
971 static inline uint64_t
972 i40e_set_tso_ctx(struct rte_mbuf *mbuf, union i40e_tx_offload tx_offload)
973 {
974         uint64_t ctx_desc = 0;
975         uint32_t cd_cmd, hdr_len, cd_tso_len;
976
977         if (!tx_offload.l4_len) {
978                 PMD_DRV_LOG(DEBUG, "L4 length set to 0");
979                 return ctx_desc;
980         }
981
982         /**
983          * in case of non tunneling packet, the outer_l2_len and
984          * outer_l3_len must be 0.
985          */
986         hdr_len = tx_offload.outer_l2_len +
987                 tx_offload.outer_l3_len +
988                 tx_offload.l2_len +
989                 tx_offload.l3_len +
990                 tx_offload.l4_len;
991
992         cd_cmd = I40E_TX_CTX_DESC_TSO;
993         cd_tso_len = mbuf->pkt_len - hdr_len;
994         ctx_desc |= ((uint64_t)cd_cmd << I40E_TXD_CTX_QW1_CMD_SHIFT) |
995                 ((uint64_t)cd_tso_len <<
996                  I40E_TXD_CTX_QW1_TSO_LEN_SHIFT) |
997                 ((uint64_t)mbuf->tso_segsz <<
998                  I40E_TXD_CTX_QW1_MSS_SHIFT);
999
1000         return ctx_desc;
1001 }
1002
1003 uint16_t
1004 i40e_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1005 {
1006         struct i40e_tx_queue *txq;
1007         struct i40e_tx_entry *sw_ring;
1008         struct i40e_tx_entry *txe, *txn;
1009         volatile struct i40e_tx_desc *txd;
1010         volatile struct i40e_tx_desc *txr;
1011         struct rte_mbuf *tx_pkt;
1012         struct rte_mbuf *m_seg;
1013         uint32_t cd_tunneling_params;
1014         uint16_t tx_id;
1015         uint16_t nb_tx;
1016         uint32_t td_cmd;
1017         uint32_t td_offset;
1018         uint32_t td_tag;
1019         uint64_t ol_flags;
1020         uint16_t nb_used;
1021         uint16_t nb_ctx;
1022         uint16_t tx_last;
1023         uint16_t slen;
1024         uint64_t buf_dma_addr;
1025         union i40e_tx_offload tx_offload = {0};
1026
1027         txq = tx_queue;
1028         sw_ring = txq->sw_ring;
1029         txr = txq->tx_ring;
1030         tx_id = txq->tx_tail;
1031         txe = &sw_ring[tx_id];
1032
1033         /* Check if the descriptor ring needs to be cleaned. */
1034         if (txq->nb_tx_free < txq->tx_free_thresh)
1035                 i40e_xmit_cleanup(txq);
1036
1037         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1038                 td_cmd = 0;
1039                 td_tag = 0;
1040                 td_offset = 0;
1041
1042                 tx_pkt = *tx_pkts++;
1043                 RTE_MBUF_PREFETCH_TO_FREE(txe->mbuf);
1044
1045                 ol_flags = tx_pkt->ol_flags;
1046                 tx_offload.l2_len = tx_pkt->l2_len;
1047                 tx_offload.l3_len = tx_pkt->l3_len;
1048                 tx_offload.outer_l2_len = tx_pkt->outer_l2_len;
1049                 tx_offload.outer_l3_len = tx_pkt->outer_l3_len;
1050                 tx_offload.l4_len = tx_pkt->l4_len;
1051                 tx_offload.tso_segsz = tx_pkt->tso_segsz;
1052
1053                 /* Calculate the number of context descriptors needed. */
1054                 nb_ctx = i40e_calc_context_desc(ol_flags);
1055
1056                 /**
1057                  * The number of descriptors that must be allocated for
1058                  * a packet equals to the number of the segments of that
1059                  * packet plus 1 context descriptor if needed.
1060                  */
1061                 nb_used = (uint16_t)(tx_pkt->nb_segs + nb_ctx);
1062                 tx_last = (uint16_t)(tx_id + nb_used - 1);
1063
1064                 /* Circular ring */
1065                 if (tx_last >= txq->nb_tx_desc)
1066                         tx_last = (uint16_t)(tx_last - txq->nb_tx_desc);
1067
1068                 if (nb_used > txq->nb_tx_free) {
1069                         if (i40e_xmit_cleanup(txq) != 0) {
1070                                 if (nb_tx == 0)
1071                                         return 0;
1072                                 goto end_of_tx;
1073                         }
1074                         if (unlikely(nb_used > txq->tx_rs_thresh)) {
1075                                 while (nb_used > txq->nb_tx_free) {
1076                                         if (i40e_xmit_cleanup(txq) != 0) {
1077                                                 if (nb_tx == 0)
1078                                                         return 0;
1079                                                 goto end_of_tx;
1080                                         }
1081                                 }
1082                         }
1083                 }
1084
1085                 /* Descriptor based VLAN insertion */
1086                 if (ol_flags & (PKT_TX_VLAN_PKT | PKT_TX_QINQ_PKT)) {
1087                         td_cmd |= I40E_TX_DESC_CMD_IL2TAG1;
1088                         td_tag = tx_pkt->vlan_tci;
1089                 }
1090
1091                 /* Always enable CRC offload insertion */
1092                 td_cmd |= I40E_TX_DESC_CMD_ICRC;
1093
1094                 /* Fill in tunneling parameters if necessary */
1095                 cd_tunneling_params = 0;
1096                 if (ol_flags & PKT_TX_TUNNEL_MASK)
1097                         i40e_parse_tunneling_params(ol_flags, tx_offload,
1098                                                     &cd_tunneling_params);
1099                 /* Enable checksum offloading */
1100                 if (ol_flags & I40E_TX_CKSUM_OFFLOAD_MASK)
1101                         i40e_txd_enable_checksum(ol_flags, &td_cmd,
1102                                                  &td_offset, tx_offload);
1103
1104                 if (nb_ctx) {
1105                         /* Setup TX context descriptor if required */
1106                         volatile struct i40e_tx_context_desc *ctx_txd =
1107                                 (volatile struct i40e_tx_context_desc *)\
1108                                                         &txr[tx_id];
1109                         uint16_t cd_l2tag2 = 0;
1110                         uint64_t cd_type_cmd_tso_mss =
1111                                 I40E_TX_DESC_DTYPE_CONTEXT;
1112
1113                         txn = &sw_ring[txe->next_id];
1114                         RTE_MBUF_PREFETCH_TO_FREE(txn->mbuf);
1115                         if (txe->mbuf != NULL) {
1116                                 rte_pktmbuf_free_seg(txe->mbuf);
1117                                 txe->mbuf = NULL;
1118                         }
1119
1120                         /* TSO enabled means no timestamp */
1121                         if (ol_flags & PKT_TX_TCP_SEG)
1122                                 cd_type_cmd_tso_mss |=
1123                                         i40e_set_tso_ctx(tx_pkt, tx_offload);
1124                         else {
1125 #ifdef RTE_LIBRTE_IEEE1588
1126                                 if (ol_flags & PKT_TX_IEEE1588_TMST)
1127                                         cd_type_cmd_tso_mss |=
1128                                                 ((uint64_t)I40E_TX_CTX_DESC_TSYN <<
1129                                                  I40E_TXD_CTX_QW1_CMD_SHIFT);
1130 #endif
1131                         }
1132
1133                         ctx_txd->tunneling_params =
1134                                 rte_cpu_to_le_32(cd_tunneling_params);
1135                         if (ol_flags & PKT_TX_QINQ_PKT) {
1136                                 cd_l2tag2 = tx_pkt->vlan_tci_outer;
1137                                 cd_type_cmd_tso_mss |=
1138                                         ((uint64_t)I40E_TX_CTX_DESC_IL2TAG2 <<
1139                                                 I40E_TXD_CTX_QW1_CMD_SHIFT);
1140                         }
1141                         ctx_txd->l2tag2 = rte_cpu_to_le_16(cd_l2tag2);
1142                         ctx_txd->type_cmd_tso_mss =
1143                                 rte_cpu_to_le_64(cd_type_cmd_tso_mss);
1144
1145                         PMD_TX_LOG(DEBUG, "mbuf: %p, TCD[%u]:\n"
1146                                 "tunneling_params: %#x;\n"
1147                                 "l2tag2: %#hx;\n"
1148                                 "rsvd: %#hx;\n"
1149                                 "type_cmd_tso_mss: %#"PRIx64";\n",
1150                                 tx_pkt, tx_id,
1151                                 ctx_txd->tunneling_params,
1152                                 ctx_txd->l2tag2,
1153                                 ctx_txd->rsvd,
1154                                 ctx_txd->type_cmd_tso_mss);
1155
1156                         txe->last_id = tx_last;
1157                         tx_id = txe->next_id;
1158                         txe = txn;
1159                 }
1160
1161                 m_seg = tx_pkt;
1162                 do {
1163                         txd = &txr[tx_id];
1164                         txn = &sw_ring[txe->next_id];
1165
1166                         if (txe->mbuf)
1167                                 rte_pktmbuf_free_seg(txe->mbuf);
1168                         txe->mbuf = m_seg;
1169
1170                         /* Setup TX Descriptor */
1171                         slen = m_seg->data_len;
1172                         buf_dma_addr = rte_mbuf_data_iova(m_seg);
1173
1174                         PMD_TX_LOG(DEBUG, "mbuf: %p, TDD[%u]:\n"
1175                                 "buf_dma_addr: %#"PRIx64";\n"
1176                                 "td_cmd: %#x;\n"
1177                                 "td_offset: %#x;\n"
1178                                 "td_len: %u;\n"
1179                                 "td_tag: %#x;\n",
1180                                 tx_pkt, tx_id, buf_dma_addr,
1181                                 td_cmd, td_offset, slen, td_tag);
1182
1183                         txd->buffer_addr = rte_cpu_to_le_64(buf_dma_addr);
1184                         txd->cmd_type_offset_bsz = i40e_build_ctob(td_cmd,
1185                                                 td_offset, slen, td_tag);
1186                         txe->last_id = tx_last;
1187                         tx_id = txe->next_id;
1188                         txe = txn;
1189                         m_seg = m_seg->next;
1190                 } while (m_seg != NULL);
1191
1192                 /* The last packet data descriptor needs End Of Packet (EOP) */
1193                 td_cmd |= I40E_TX_DESC_CMD_EOP;
1194                 txq->nb_tx_used = (uint16_t)(txq->nb_tx_used + nb_used);
1195                 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_used);
1196
1197                 if (txq->nb_tx_used >= txq->tx_rs_thresh) {
1198                         PMD_TX_FREE_LOG(DEBUG,
1199                                         "Setting RS bit on TXD id="
1200                                         "%4u (port=%d queue=%d)",
1201                                         tx_last, txq->port_id, txq->queue_id);
1202
1203                         td_cmd |= I40E_TX_DESC_CMD_RS;
1204
1205                         /* Update txq RS bit counters */
1206                         txq->nb_tx_used = 0;
1207                 }
1208
1209                 txd->cmd_type_offset_bsz |=
1210                         rte_cpu_to_le_64(((uint64_t)td_cmd) <<
1211                                         I40E_TXD_QW1_CMD_SHIFT);
1212         }
1213
1214 end_of_tx:
1215         rte_wmb();
1216
1217         PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u tx_tail=%u nb_tx=%u",
1218                    (unsigned) txq->port_id, (unsigned) txq->queue_id,
1219                    (unsigned) tx_id, (unsigned) nb_tx);
1220
1221         I40E_PCI_REG_WRITE_RELAXED(txq->qtx_tail, tx_id);
1222         txq->tx_tail = tx_id;
1223
1224         return nb_tx;
1225 }
1226
1227 static __rte_always_inline int
1228 i40e_tx_free_bufs(struct i40e_tx_queue *txq)
1229 {
1230         struct i40e_tx_entry *txep;
1231         uint16_t i;
1232
1233         if ((txq->tx_ring[txq->tx_next_dd].cmd_type_offset_bsz &
1234                         rte_cpu_to_le_64(I40E_TXD_QW1_DTYPE_MASK)) !=
1235                         rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DESC_DONE))
1236                 return 0;
1237
1238         txep = &(txq->sw_ring[txq->tx_next_dd - (txq->tx_rs_thresh - 1)]);
1239
1240         for (i = 0; i < txq->tx_rs_thresh; i++)
1241                 rte_prefetch0((txep + i)->mbuf);
1242
1243         if (txq->txq_flags & (uint32_t)ETH_TXQ_FLAGS_NOREFCOUNT) {
1244                 for (i = 0; i < txq->tx_rs_thresh; ++i, ++txep) {
1245                         rte_mempool_put(txep->mbuf->pool, txep->mbuf);
1246                         txep->mbuf = NULL;
1247                 }
1248         } else {
1249                 for (i = 0; i < txq->tx_rs_thresh; ++i, ++txep) {
1250                         rte_pktmbuf_free_seg(txep->mbuf);
1251                         txep->mbuf = NULL;
1252                 }
1253         }
1254
1255         txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + txq->tx_rs_thresh);
1256         txq->tx_next_dd = (uint16_t)(txq->tx_next_dd + txq->tx_rs_thresh);
1257         if (txq->tx_next_dd >= txq->nb_tx_desc)
1258                 txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
1259
1260         return txq->tx_rs_thresh;
1261 }
1262
1263 /* Populate 4 descriptors with data from 4 mbufs */
1264 static inline void
1265 tx4(volatile struct i40e_tx_desc *txdp, struct rte_mbuf **pkts)
1266 {
1267         uint64_t dma_addr;
1268         uint32_t i;
1269
1270         for (i = 0; i < 4; i++, txdp++, pkts++) {
1271                 dma_addr = rte_mbuf_data_iova(*pkts);
1272                 txdp->buffer_addr = rte_cpu_to_le_64(dma_addr);
1273                 txdp->cmd_type_offset_bsz =
1274                         i40e_build_ctob((uint32_t)I40E_TD_CMD, 0,
1275                                         (*pkts)->data_len, 0);
1276         }
1277 }
1278
1279 /* Populate 1 descriptor with data from 1 mbuf */
1280 static inline void
1281 tx1(volatile struct i40e_tx_desc *txdp, struct rte_mbuf **pkts)
1282 {
1283         uint64_t dma_addr;
1284
1285         dma_addr = rte_mbuf_data_iova(*pkts);
1286         txdp->buffer_addr = rte_cpu_to_le_64(dma_addr);
1287         txdp->cmd_type_offset_bsz =
1288                 i40e_build_ctob((uint32_t)I40E_TD_CMD, 0,
1289                                 (*pkts)->data_len, 0);
1290 }
1291
1292 /* Fill hardware descriptor ring with mbuf data */
1293 static inline void
1294 i40e_tx_fill_hw_ring(struct i40e_tx_queue *txq,
1295                      struct rte_mbuf **pkts,
1296                      uint16_t nb_pkts)
1297 {
1298         volatile struct i40e_tx_desc *txdp = &(txq->tx_ring[txq->tx_tail]);
1299         struct i40e_tx_entry *txep = &(txq->sw_ring[txq->tx_tail]);
1300         const int N_PER_LOOP = 4;
1301         const int N_PER_LOOP_MASK = N_PER_LOOP - 1;
1302         int mainpart, leftover;
1303         int i, j;
1304
1305         mainpart = (nb_pkts & ((uint32_t) ~N_PER_LOOP_MASK));
1306         leftover = (nb_pkts & ((uint32_t)  N_PER_LOOP_MASK));
1307         for (i = 0; i < mainpart; i += N_PER_LOOP) {
1308                 for (j = 0; j < N_PER_LOOP; ++j) {
1309                         (txep + i + j)->mbuf = *(pkts + i + j);
1310                 }
1311                 tx4(txdp + i, pkts + i);
1312         }
1313         if (unlikely(leftover > 0)) {
1314                 for (i = 0; i < leftover; ++i) {
1315                         (txep + mainpart + i)->mbuf = *(pkts + mainpart + i);
1316                         tx1(txdp + mainpart + i, pkts + mainpart + i);
1317                 }
1318         }
1319 }
1320
1321 static inline uint16_t
1322 tx_xmit_pkts(struct i40e_tx_queue *txq,
1323              struct rte_mbuf **tx_pkts,
1324              uint16_t nb_pkts)
1325 {
1326         volatile struct i40e_tx_desc *txr = txq->tx_ring;
1327         uint16_t n = 0;
1328
1329         /**
1330          * Begin scanning the H/W ring for done descriptors when the number
1331          * of available descriptors drops below tx_free_thresh. For each done
1332          * descriptor, free the associated buffer.
1333          */
1334         if (txq->nb_tx_free < txq->tx_free_thresh)
1335                 i40e_tx_free_bufs(txq);
1336
1337         /* Use available descriptor only */
1338         nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
1339         if (unlikely(!nb_pkts))
1340                 return 0;
1341
1342         txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
1343         if ((txq->tx_tail + nb_pkts) > txq->nb_tx_desc) {
1344                 n = (uint16_t)(txq->nb_tx_desc - txq->tx_tail);
1345                 i40e_tx_fill_hw_ring(txq, tx_pkts, n);
1346                 txr[txq->tx_next_rs].cmd_type_offset_bsz |=
1347                         rte_cpu_to_le_64(((uint64_t)I40E_TX_DESC_CMD_RS) <<
1348                                                 I40E_TXD_QW1_CMD_SHIFT);
1349                 txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
1350                 txq->tx_tail = 0;
1351         }
1352
1353         /* Fill hardware descriptor ring with mbuf data */
1354         i40e_tx_fill_hw_ring(txq, tx_pkts + n, (uint16_t)(nb_pkts - n));
1355         txq->tx_tail = (uint16_t)(txq->tx_tail + (nb_pkts - n));
1356
1357         /* Determin if RS bit needs to be set */
1358         if (txq->tx_tail > txq->tx_next_rs) {
1359                 txr[txq->tx_next_rs].cmd_type_offset_bsz |=
1360                         rte_cpu_to_le_64(((uint64_t)I40E_TX_DESC_CMD_RS) <<
1361                                                 I40E_TXD_QW1_CMD_SHIFT);
1362                 txq->tx_next_rs =
1363                         (uint16_t)(txq->tx_next_rs + txq->tx_rs_thresh);
1364                 if (txq->tx_next_rs >= txq->nb_tx_desc)
1365                         txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
1366         }
1367
1368         if (txq->tx_tail >= txq->nb_tx_desc)
1369                 txq->tx_tail = 0;
1370
1371         /* Update the tx tail register */
1372         rte_wmb();
1373         I40E_PCI_REG_WRITE_RELAXED(txq->qtx_tail, txq->tx_tail);
1374
1375         return nb_pkts;
1376 }
1377
1378 static uint16_t
1379 i40e_xmit_pkts_simple(void *tx_queue,
1380                       struct rte_mbuf **tx_pkts,
1381                       uint16_t nb_pkts)
1382 {
1383         uint16_t nb_tx = 0;
1384
1385         if (likely(nb_pkts <= I40E_TX_MAX_BURST))
1386                 return tx_xmit_pkts((struct i40e_tx_queue *)tx_queue,
1387                                                 tx_pkts, nb_pkts);
1388
1389         while (nb_pkts) {
1390                 uint16_t ret, num = (uint16_t)RTE_MIN(nb_pkts,
1391                                                 I40E_TX_MAX_BURST);
1392
1393                 ret = tx_xmit_pkts((struct i40e_tx_queue *)tx_queue,
1394                                                 &tx_pkts[nb_tx], num);
1395                 nb_tx = (uint16_t)(nb_tx + ret);
1396                 nb_pkts = (uint16_t)(nb_pkts - ret);
1397                 if (ret < num)
1398                         break;
1399         }
1400
1401         return nb_tx;
1402 }
1403
1404 static uint16_t
1405 i40e_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
1406                    uint16_t nb_pkts)
1407 {
1408         uint16_t nb_tx = 0;
1409         struct i40e_tx_queue *txq = (struct i40e_tx_queue *)tx_queue;
1410
1411         while (nb_pkts) {
1412                 uint16_t ret, num;
1413
1414                 num = (uint16_t)RTE_MIN(nb_pkts, txq->tx_rs_thresh);
1415                 ret = i40e_xmit_fixed_burst_vec(tx_queue, &tx_pkts[nb_tx],
1416                                                 num);
1417                 nb_tx += ret;
1418                 nb_pkts -= ret;
1419                 if (ret < num)
1420                         break;
1421         }
1422
1423         return nb_tx;
1424 }
1425
1426 /*********************************************************************
1427  *
1428  *  TX prep functions
1429  *
1430  **********************************************************************/
1431 uint16_t
1432 i40e_prep_pkts(__rte_unused void *tx_queue, struct rte_mbuf **tx_pkts,
1433                 uint16_t nb_pkts)
1434 {
1435         int i, ret;
1436         uint64_t ol_flags;
1437         struct rte_mbuf *m;
1438
1439         for (i = 0; i < nb_pkts; i++) {
1440                 m = tx_pkts[i];
1441                 ol_flags = m->ol_flags;
1442
1443                 /* Check for m->nb_segs to not exceed the limits. */
1444                 if (!(ol_flags & PKT_TX_TCP_SEG)) {
1445                         if (m->nb_segs > I40E_TX_MAX_SEG ||
1446                             m->nb_segs > I40E_TX_MAX_MTU_SEG) {
1447                                 rte_errno = -EINVAL;
1448                                 return i;
1449                         }
1450                 } else if ((m->tso_segsz < I40E_MIN_TSO_MSS) ||
1451                                 (m->tso_segsz > I40E_MAX_TSO_MSS)) {
1452                         /* MSS outside the range (256B - 9674B) are considered
1453                          * malicious
1454                          */
1455                         rte_errno = -EINVAL;
1456                         return i;
1457                 }
1458
1459                 if (ol_flags & I40E_TX_OFFLOAD_NOTSUP_MASK) {
1460                         rte_errno = -ENOTSUP;
1461                         return i;
1462                 }
1463
1464 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
1465                 ret = rte_validate_tx_offload(m);
1466                 if (ret != 0) {
1467                         rte_errno = ret;
1468                         return i;
1469                 }
1470 #endif
1471                 ret = rte_net_intel_cksum_prepare(m);
1472                 if (ret != 0) {
1473                         rte_errno = ret;
1474                         return i;
1475                 }
1476         }
1477         return i;
1478 }
1479
1480 /*
1481  * Find the VSI the queue belongs to. 'queue_idx' is the queue index
1482  * application used, which assume having sequential ones. But from driver's
1483  * perspective, it's different. For example, q0 belongs to FDIR VSI, q1-q64
1484  * to MAIN VSI, , q65-96 to SRIOV VSIs, q97-128 to VMDQ VSIs. For application
1485  * running on host, q1-64 and q97-128 can be used, total 96 queues. They can
1486  * use queue_idx from 0 to 95 to access queues, while real queue would be
1487  * different. This function will do a queue mapping to find VSI the queue
1488  * belongs to.
1489  */
1490 static struct i40e_vsi*
1491 i40e_pf_get_vsi_by_qindex(struct i40e_pf *pf, uint16_t queue_idx)
1492 {
1493         /* the queue in MAIN VSI range */
1494         if (queue_idx < pf->main_vsi->nb_qps)
1495                 return pf->main_vsi;
1496
1497         queue_idx -= pf->main_vsi->nb_qps;
1498
1499         /* queue_idx is greater than VMDQ VSIs range */
1500         if (queue_idx > pf->nb_cfg_vmdq_vsi * pf->vmdq_nb_qps - 1) {
1501                 PMD_INIT_LOG(ERR, "queue_idx out of range. VMDQ configured?");
1502                 return NULL;
1503         }
1504
1505         return pf->vmdq[queue_idx / pf->vmdq_nb_qps].vsi;
1506 }
1507
1508 static uint16_t
1509 i40e_get_queue_offset_by_qindex(struct i40e_pf *pf, uint16_t queue_idx)
1510 {
1511         /* the queue in MAIN VSI range */
1512         if (queue_idx < pf->main_vsi->nb_qps)
1513                 return queue_idx;
1514
1515         /* It's VMDQ queues */
1516         queue_idx -= pf->main_vsi->nb_qps;
1517
1518         if (pf->nb_cfg_vmdq_vsi)
1519                 return queue_idx % pf->vmdq_nb_qps;
1520         else {
1521                 PMD_INIT_LOG(ERR, "Fail to get queue offset");
1522                 return (uint16_t)(-1);
1523         }
1524 }
1525
1526 int
1527 i40e_dev_rx_queue_start(struct rte_eth_dev *dev, uint16_t rx_queue_id)
1528 {
1529         struct i40e_rx_queue *rxq;
1530         int err = -1;
1531         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1532
1533         PMD_INIT_FUNC_TRACE();
1534
1535         if (rx_queue_id < dev->data->nb_rx_queues) {
1536                 rxq = dev->data->rx_queues[rx_queue_id];
1537
1538                 err = i40e_alloc_rx_queue_mbufs(rxq);
1539                 if (err) {
1540                         PMD_DRV_LOG(ERR, "Failed to allocate RX queue mbuf");
1541                         return err;
1542                 }
1543
1544                 rte_wmb();
1545
1546                 /* Init the RX tail regieter. */
1547                 I40E_PCI_REG_WRITE(rxq->qrx_tail, rxq->nb_rx_desc - 1);
1548
1549                 err = i40e_switch_rx_queue(hw, rxq->reg_idx, TRUE);
1550
1551                 if (err) {
1552                         PMD_DRV_LOG(ERR, "Failed to switch RX queue %u on",
1553                                     rx_queue_id);
1554
1555                         i40e_rx_queue_release_mbufs(rxq);
1556                         i40e_reset_rx_queue(rxq);
1557                 } else
1558                         dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
1559         }
1560
1561         return err;
1562 }
1563
1564 int
1565 i40e_dev_rx_queue_stop(struct rte_eth_dev *dev, uint16_t rx_queue_id)
1566 {
1567         struct i40e_rx_queue *rxq;
1568         int err;
1569         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1570
1571         if (rx_queue_id < dev->data->nb_rx_queues) {
1572                 rxq = dev->data->rx_queues[rx_queue_id];
1573
1574                 /*
1575                 * rx_queue_id is queue id application refers to, while
1576                 * rxq->reg_idx is the real queue index.
1577                 */
1578                 err = i40e_switch_rx_queue(hw, rxq->reg_idx, FALSE);
1579
1580                 if (err) {
1581                         PMD_DRV_LOG(ERR, "Failed to switch RX queue %u off",
1582                                     rx_queue_id);
1583                         return err;
1584                 }
1585                 i40e_rx_queue_release_mbufs(rxq);
1586                 i40e_reset_rx_queue(rxq);
1587                 dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
1588         }
1589
1590         return 0;
1591 }
1592
1593 int
1594 i40e_dev_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id)
1595 {
1596         int err = -1;
1597         struct i40e_tx_queue *txq;
1598         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1599
1600         PMD_INIT_FUNC_TRACE();
1601
1602         if (tx_queue_id < dev->data->nb_tx_queues) {
1603                 txq = dev->data->tx_queues[tx_queue_id];
1604
1605                 /*
1606                 * tx_queue_id is queue id application refers to, while
1607                 * rxq->reg_idx is the real queue index.
1608                 */
1609                 err = i40e_switch_tx_queue(hw, txq->reg_idx, TRUE);
1610                 if (err)
1611                         PMD_DRV_LOG(ERR, "Failed to switch TX queue %u on",
1612                                     tx_queue_id);
1613                 else
1614                         dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
1615         }
1616
1617         return err;
1618 }
1619
1620 int
1621 i40e_dev_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id)
1622 {
1623         struct i40e_tx_queue *txq;
1624         int err;
1625         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1626
1627         if (tx_queue_id < dev->data->nb_tx_queues) {
1628                 txq = dev->data->tx_queues[tx_queue_id];
1629
1630                 /*
1631                 * tx_queue_id is queue id application refers to, while
1632                 * txq->reg_idx is the real queue index.
1633                 */
1634                 err = i40e_switch_tx_queue(hw, txq->reg_idx, FALSE);
1635
1636                 if (err) {
1637                         PMD_DRV_LOG(ERR, "Failed to switch TX queue %u of",
1638                                     tx_queue_id);
1639                         return err;
1640                 }
1641
1642                 i40e_tx_queue_release_mbufs(txq);
1643                 i40e_reset_tx_queue(txq);
1644                 dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
1645         }
1646
1647         return 0;
1648 }
1649
1650 const uint32_t *
1651 i40e_dev_supported_ptypes_get(struct rte_eth_dev *dev)
1652 {
1653         static const uint32_t ptypes[] = {
1654                 /* refers to i40e_rxd_pkt_type_mapping() */
1655                 RTE_PTYPE_L2_ETHER,
1656                 RTE_PTYPE_L2_ETHER_TIMESYNC,
1657                 RTE_PTYPE_L2_ETHER_LLDP,
1658                 RTE_PTYPE_L2_ETHER_ARP,
1659                 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
1660                 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
1661                 RTE_PTYPE_L4_FRAG,
1662                 RTE_PTYPE_L4_ICMP,
1663                 RTE_PTYPE_L4_NONFRAG,
1664                 RTE_PTYPE_L4_SCTP,
1665                 RTE_PTYPE_L4_TCP,
1666                 RTE_PTYPE_L4_UDP,
1667                 RTE_PTYPE_TUNNEL_GRENAT,
1668                 RTE_PTYPE_TUNNEL_IP,
1669                 RTE_PTYPE_INNER_L2_ETHER,
1670                 RTE_PTYPE_INNER_L2_ETHER_VLAN,
1671                 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
1672                 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
1673                 RTE_PTYPE_INNER_L4_FRAG,
1674                 RTE_PTYPE_INNER_L4_ICMP,
1675                 RTE_PTYPE_INNER_L4_NONFRAG,
1676                 RTE_PTYPE_INNER_L4_SCTP,
1677                 RTE_PTYPE_INNER_L4_TCP,
1678                 RTE_PTYPE_INNER_L4_UDP,
1679                 RTE_PTYPE_UNKNOWN
1680         };
1681
1682         if (dev->rx_pkt_burst == i40e_recv_pkts ||
1683 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
1684             dev->rx_pkt_burst == i40e_recv_pkts_bulk_alloc ||
1685 #endif
1686             dev->rx_pkt_burst == i40e_recv_scattered_pkts ||
1687             dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec ||
1688             dev->rx_pkt_burst == i40e_recv_pkts_vec ||
1689             dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec_avx2 ||
1690             dev->rx_pkt_burst == i40e_recv_pkts_vec_avx2)
1691                 return ptypes;
1692         return NULL;
1693 }
1694
1695 static int
1696 i40e_check_rx_queue_offloads(struct rte_eth_dev *dev, uint64_t requested)
1697 {
1698         struct rte_eth_dev_info dev_info;
1699         uint64_t mandatory = dev->data->dev_conf.rxmode.offloads;
1700         uint64_t supported; /* All per port offloads */
1701
1702         dev->dev_ops->dev_infos_get(dev, &dev_info);
1703         supported = dev_info.rx_offload_capa ^ dev_info.rx_queue_offload_capa;
1704         if ((requested & dev_info.rx_offload_capa) != requested)
1705                 return 0; /* requested range check */
1706         return !((mandatory ^ requested) & supported);
1707 }
1708
1709 int
1710 i40e_dev_rx_queue_setup(struct rte_eth_dev *dev,
1711                         uint16_t queue_idx,
1712                         uint16_t nb_desc,
1713                         unsigned int socket_id,
1714                         const struct rte_eth_rxconf *rx_conf,
1715                         struct rte_mempool *mp)
1716 {
1717         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1718         struct i40e_adapter *ad =
1719                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
1720         struct i40e_vsi *vsi;
1721         struct i40e_pf *pf = NULL;
1722         struct i40e_vf *vf = NULL;
1723         struct i40e_rx_queue *rxq;
1724         const struct rte_memzone *rz;
1725         uint32_t ring_size;
1726         uint16_t len, i;
1727         uint16_t reg_idx, base, bsf, tc_mapping;
1728         int q_offset, use_def_burst_func = 1;
1729         struct rte_eth_dev_info dev_info;
1730
1731         if (!i40e_check_rx_queue_offloads(dev, rx_conf->offloads)) {
1732                 dev->dev_ops->dev_infos_get(dev, &dev_info);
1733                 PMD_INIT_LOG(ERR, "%p: Rx queue offloads 0x%" PRIx64
1734                         " don't match port  offloads 0x%" PRIx64
1735                         " or supported offloads 0x%" PRIx64,
1736                         (void *)dev, rx_conf->offloads,
1737                         dev->data->dev_conf.rxmode.offloads,
1738                         dev_info.rx_offload_capa);
1739                 return -ENOTSUP;
1740         }
1741
1742         if (hw->mac.type == I40E_MAC_VF || hw->mac.type == I40E_MAC_X722_VF) {
1743                 vf = I40EVF_DEV_PRIVATE_TO_VF(dev->data->dev_private);
1744                 vsi = &vf->vsi;
1745                 if (!vsi)
1746                         return -EINVAL;
1747                 reg_idx = queue_idx;
1748         } else {
1749                 pf = I40E_DEV_PRIVATE_TO_PF(dev->data->dev_private);
1750                 vsi = i40e_pf_get_vsi_by_qindex(pf, queue_idx);
1751                 if (!vsi)
1752                         return -EINVAL;
1753                 q_offset = i40e_get_queue_offset_by_qindex(pf, queue_idx);
1754                 if (q_offset < 0)
1755                         return -EINVAL;
1756                 reg_idx = vsi->base_queue + q_offset;
1757         }
1758
1759         if (nb_desc % I40E_ALIGN_RING_DESC != 0 ||
1760             (nb_desc > I40E_MAX_RING_DESC) ||
1761             (nb_desc < I40E_MIN_RING_DESC)) {
1762                 PMD_DRV_LOG(ERR, "Number (%u) of receive descriptors is "
1763                             "invalid", nb_desc);
1764                 return -EINVAL;
1765         }
1766
1767         /* Free memory if needed */
1768         if (dev->data->rx_queues[queue_idx]) {
1769                 i40e_dev_rx_queue_release(dev->data->rx_queues[queue_idx]);
1770                 dev->data->rx_queues[queue_idx] = NULL;
1771         }
1772
1773         /* Allocate the rx queue data structure */
1774         rxq = rte_zmalloc_socket("i40e rx queue",
1775                                  sizeof(struct i40e_rx_queue),
1776                                  RTE_CACHE_LINE_SIZE,
1777                                  socket_id);
1778         if (!rxq) {
1779                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
1780                             "rx queue data structure");
1781                 return -ENOMEM;
1782         }
1783         rxq->mp = mp;
1784         rxq->nb_rx_desc = nb_desc;
1785         rxq->rx_free_thresh = rx_conf->rx_free_thresh;
1786         rxq->queue_id = queue_idx;
1787         rxq->reg_idx = reg_idx;
1788         rxq->port_id = dev->data->port_id;
1789         rxq->crc_len = (uint8_t)((dev->data->dev_conf.rxmode.offloads &
1790                         DEV_RX_OFFLOAD_CRC_STRIP) ? 0 : ETHER_CRC_LEN);
1791         rxq->drop_en = rx_conf->rx_drop_en;
1792         rxq->vsi = vsi;
1793         rxq->rx_deferred_start = rx_conf->rx_deferred_start;
1794
1795         /* Allocate the maximun number of RX ring hardware descriptor. */
1796         len = I40E_MAX_RING_DESC;
1797
1798         /**
1799          * Allocating a little more memory because vectorized/bulk_alloc Rx
1800          * functions doesn't check boundaries each time.
1801          */
1802         len += RTE_PMD_I40E_RX_MAX_BURST;
1803
1804         ring_size = RTE_ALIGN(len * sizeof(union i40e_rx_desc),
1805                               I40E_DMA_MEM_ALIGN);
1806
1807         rz = rte_eth_dma_zone_reserve(dev, "rx_ring", queue_idx,
1808                               ring_size, I40E_RING_BASE_ALIGN, socket_id);
1809         if (!rz) {
1810                 i40e_dev_rx_queue_release(rxq);
1811                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for RX");
1812                 return -ENOMEM;
1813         }
1814
1815         /* Zero all the descriptors in the ring. */
1816         memset(rz->addr, 0, ring_size);
1817
1818         rxq->rx_ring_phys_addr = rz->iova;
1819         rxq->rx_ring = (union i40e_rx_desc *)rz->addr;
1820
1821         len = (uint16_t)(nb_desc + RTE_PMD_I40E_RX_MAX_BURST);
1822
1823         /* Allocate the software ring. */
1824         rxq->sw_ring =
1825                 rte_zmalloc_socket("i40e rx sw ring",
1826                                    sizeof(struct i40e_rx_entry) * len,
1827                                    RTE_CACHE_LINE_SIZE,
1828                                    socket_id);
1829         if (!rxq->sw_ring) {
1830                 i40e_dev_rx_queue_release(rxq);
1831                 PMD_DRV_LOG(ERR, "Failed to allocate memory for SW ring");
1832                 return -ENOMEM;
1833         }
1834
1835         i40e_reset_rx_queue(rxq);
1836         rxq->q_set = TRUE;
1837         dev->data->rx_queues[queue_idx] = rxq;
1838
1839         use_def_burst_func = check_rx_burst_bulk_alloc_preconditions(rxq);
1840
1841         if (!use_def_burst_func) {
1842 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
1843                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions are "
1844                              "satisfied. Rx Burst Bulk Alloc function will be "
1845                              "used on port=%d, queue=%d.",
1846                              rxq->port_id, rxq->queue_id);
1847 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
1848         } else {
1849                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions are "
1850                              "not satisfied, Scattered Rx is requested, "
1851                              "or RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC is "
1852                              "not enabled on port=%d, queue=%d.",
1853                              rxq->port_id, rxq->queue_id);
1854                 ad->rx_bulk_alloc_allowed = false;
1855         }
1856
1857         for (i = 0; i < I40E_MAX_TRAFFIC_CLASS; i++) {
1858                 if (!(vsi->enabled_tc & (1 << i)))
1859                         continue;
1860                 tc_mapping = rte_le_to_cpu_16(vsi->info.tc_mapping[i]);
1861                 base = (tc_mapping & I40E_AQ_VSI_TC_QUE_OFFSET_MASK) >>
1862                         I40E_AQ_VSI_TC_QUE_OFFSET_SHIFT;
1863                 bsf = (tc_mapping & I40E_AQ_VSI_TC_QUE_NUMBER_MASK) >>
1864                         I40E_AQ_VSI_TC_QUE_NUMBER_SHIFT;
1865
1866                 if (queue_idx >= base && queue_idx < (base + BIT(bsf)))
1867                         rxq->dcb_tc = i;
1868         }
1869
1870         return 0;
1871 }
1872
1873 void
1874 i40e_dev_rx_queue_release(void *rxq)
1875 {
1876         struct i40e_rx_queue *q = (struct i40e_rx_queue *)rxq;
1877
1878         if (!q) {
1879                 PMD_DRV_LOG(DEBUG, "Pointer to rxq is NULL");
1880                 return;
1881         }
1882
1883         i40e_rx_queue_release_mbufs(q);
1884         rte_free(q->sw_ring);
1885         rte_free(q);
1886 }
1887
1888 uint32_t
1889 i40e_dev_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id)
1890 {
1891 #define I40E_RXQ_SCAN_INTERVAL 4
1892         volatile union i40e_rx_desc *rxdp;
1893         struct i40e_rx_queue *rxq;
1894         uint16_t desc = 0;
1895
1896         rxq = dev->data->rx_queues[rx_queue_id];
1897         rxdp = &(rxq->rx_ring[rxq->rx_tail]);
1898         while ((desc < rxq->nb_rx_desc) &&
1899                 ((rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len) &
1900                 I40E_RXD_QW1_STATUS_MASK) >> I40E_RXD_QW1_STATUS_SHIFT) &
1901                                 (1 << I40E_RX_DESC_STATUS_DD_SHIFT)) {
1902                 /**
1903                  * Check the DD bit of a rx descriptor of each 4 in a group,
1904                  * to avoid checking too frequently and downgrading performance
1905                  * too much.
1906                  */
1907                 desc += I40E_RXQ_SCAN_INTERVAL;
1908                 rxdp += I40E_RXQ_SCAN_INTERVAL;
1909                 if (rxq->rx_tail + desc >= rxq->nb_rx_desc)
1910                         rxdp = &(rxq->rx_ring[rxq->rx_tail +
1911                                         desc - rxq->nb_rx_desc]);
1912         }
1913
1914         return desc;
1915 }
1916
1917 int
1918 i40e_dev_rx_descriptor_done(void *rx_queue, uint16_t offset)
1919 {
1920         volatile union i40e_rx_desc *rxdp;
1921         struct i40e_rx_queue *rxq = rx_queue;
1922         uint16_t desc;
1923         int ret;
1924
1925         if (unlikely(offset >= rxq->nb_rx_desc)) {
1926                 PMD_DRV_LOG(ERR, "Invalid RX descriptor id %u", offset);
1927                 return 0;
1928         }
1929
1930         desc = rxq->rx_tail + offset;
1931         if (desc >= rxq->nb_rx_desc)
1932                 desc -= rxq->nb_rx_desc;
1933
1934         rxdp = &(rxq->rx_ring[desc]);
1935
1936         ret = !!(((rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len) &
1937                 I40E_RXD_QW1_STATUS_MASK) >> I40E_RXD_QW1_STATUS_SHIFT) &
1938                                 (1 << I40E_RX_DESC_STATUS_DD_SHIFT));
1939
1940         return ret;
1941 }
1942
1943 int
1944 i40e_dev_rx_descriptor_status(void *rx_queue, uint16_t offset)
1945 {
1946         struct i40e_rx_queue *rxq = rx_queue;
1947         volatile uint64_t *status;
1948         uint64_t mask;
1949         uint32_t desc;
1950
1951         if (unlikely(offset >= rxq->nb_rx_desc))
1952                 return -EINVAL;
1953
1954         if (offset >= rxq->nb_rx_desc - rxq->nb_rx_hold)
1955                 return RTE_ETH_RX_DESC_UNAVAIL;
1956
1957         desc = rxq->rx_tail + offset;
1958         if (desc >= rxq->nb_rx_desc)
1959                 desc -= rxq->nb_rx_desc;
1960
1961         status = &rxq->rx_ring[desc].wb.qword1.status_error_len;
1962         mask = rte_le_to_cpu_64((1ULL << I40E_RX_DESC_STATUS_DD_SHIFT)
1963                 << I40E_RXD_QW1_STATUS_SHIFT);
1964         if (*status & mask)
1965                 return RTE_ETH_RX_DESC_DONE;
1966
1967         return RTE_ETH_RX_DESC_AVAIL;
1968 }
1969
1970 int
1971 i40e_dev_tx_descriptor_status(void *tx_queue, uint16_t offset)
1972 {
1973         struct i40e_tx_queue *txq = tx_queue;
1974         volatile uint64_t *status;
1975         uint64_t mask, expect;
1976         uint32_t desc;
1977
1978         if (unlikely(offset >= txq->nb_tx_desc))
1979                 return -EINVAL;
1980
1981         desc = txq->tx_tail + offset;
1982         /* go to next desc that has the RS bit */
1983         desc = ((desc + txq->tx_rs_thresh - 1) / txq->tx_rs_thresh) *
1984                 txq->tx_rs_thresh;
1985         if (desc >= txq->nb_tx_desc) {
1986                 desc -= txq->nb_tx_desc;
1987                 if (desc >= txq->nb_tx_desc)
1988                         desc -= txq->nb_tx_desc;
1989         }
1990
1991         status = &txq->tx_ring[desc].cmd_type_offset_bsz;
1992         mask = rte_le_to_cpu_64(I40E_TXD_QW1_DTYPE_MASK);
1993         expect = rte_cpu_to_le_64(
1994                 I40E_TX_DESC_DTYPE_DESC_DONE << I40E_TXD_QW1_DTYPE_SHIFT);
1995         if ((*status & mask) == expect)
1996                 return RTE_ETH_TX_DESC_DONE;
1997
1998         return RTE_ETH_TX_DESC_FULL;
1999 }
2000
2001 int
2002 i40e_dev_tx_queue_setup(struct rte_eth_dev *dev,
2003                         uint16_t queue_idx,
2004                         uint16_t nb_desc,
2005                         unsigned int socket_id,
2006                         const struct rte_eth_txconf *tx_conf)
2007 {
2008         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2009         struct i40e_vsi *vsi;
2010         struct i40e_pf *pf = NULL;
2011         struct i40e_vf *vf = NULL;
2012         struct i40e_tx_queue *txq;
2013         const struct rte_memzone *tz;
2014         uint32_t ring_size;
2015         uint16_t tx_rs_thresh, tx_free_thresh;
2016         uint16_t reg_idx, i, base, bsf, tc_mapping;
2017         int q_offset;
2018
2019         if (hw->mac.type == I40E_MAC_VF || hw->mac.type == I40E_MAC_X722_VF) {
2020                 vf = I40EVF_DEV_PRIVATE_TO_VF(dev->data->dev_private);
2021                 vsi = &vf->vsi;
2022                 if (!vsi)
2023                         return -EINVAL;
2024                 reg_idx = queue_idx;
2025         } else {
2026                 pf = I40E_DEV_PRIVATE_TO_PF(dev->data->dev_private);
2027                 vsi = i40e_pf_get_vsi_by_qindex(pf, queue_idx);
2028                 if (!vsi)
2029                         return -EINVAL;
2030                 q_offset = i40e_get_queue_offset_by_qindex(pf, queue_idx);
2031                 if (q_offset < 0)
2032                         return -EINVAL;
2033                 reg_idx = vsi->base_queue + q_offset;
2034         }
2035
2036         if (nb_desc % I40E_ALIGN_RING_DESC != 0 ||
2037             (nb_desc > I40E_MAX_RING_DESC) ||
2038             (nb_desc < I40E_MIN_RING_DESC)) {
2039                 PMD_DRV_LOG(ERR, "Number (%u) of transmit descriptors is "
2040                             "invalid", nb_desc);
2041                 return -EINVAL;
2042         }
2043
2044         /**
2045          * The following two parameters control the setting of the RS bit on
2046          * transmit descriptors. TX descriptors will have their RS bit set
2047          * after txq->tx_rs_thresh descriptors have been used. The TX
2048          * descriptor ring will be cleaned after txq->tx_free_thresh
2049          * descriptors are used or if the number of descriptors required to
2050          * transmit a packet is greater than the number of free TX descriptors.
2051          *
2052          * The following constraints must be satisfied:
2053          *  - tx_rs_thresh must be greater than 0.
2054          *  - tx_rs_thresh must be less than the size of the ring minus 2.
2055          *  - tx_rs_thresh must be less than or equal to tx_free_thresh.
2056          *  - tx_rs_thresh must be a divisor of the ring size.
2057          *  - tx_free_thresh must be greater than 0.
2058          *  - tx_free_thresh must be less than the size of the ring minus 3.
2059          *
2060          * One descriptor in the TX ring is used as a sentinel to avoid a H/W
2061          * race condition, hence the maximum threshold constraints. When set
2062          * to zero use default values.
2063          */
2064         tx_rs_thresh = (uint16_t)((tx_conf->tx_rs_thresh) ?
2065                 tx_conf->tx_rs_thresh : DEFAULT_TX_RS_THRESH);
2066         tx_free_thresh = (uint16_t)((tx_conf->tx_free_thresh) ?
2067                 tx_conf->tx_free_thresh : DEFAULT_TX_FREE_THRESH);
2068         if (tx_rs_thresh >= (nb_desc - 2)) {
2069                 PMD_INIT_LOG(ERR, "tx_rs_thresh must be less than the "
2070                              "number of TX descriptors minus 2. "
2071                              "(tx_rs_thresh=%u port=%d queue=%d)",
2072                              (unsigned int)tx_rs_thresh,
2073                              (int)dev->data->port_id,
2074                              (int)queue_idx);
2075                 return I40E_ERR_PARAM;
2076         }
2077         if (tx_free_thresh >= (nb_desc - 3)) {
2078                 PMD_INIT_LOG(ERR, "tx_free_thresh must be less than the "
2079                              "number of TX descriptors minus 3. "
2080                              "(tx_free_thresh=%u port=%d queue=%d)",
2081                              (unsigned int)tx_free_thresh,
2082                              (int)dev->data->port_id,
2083                              (int)queue_idx);
2084                 return I40E_ERR_PARAM;
2085         }
2086         if (tx_rs_thresh > tx_free_thresh) {
2087                 PMD_INIT_LOG(ERR, "tx_rs_thresh must be less than or "
2088                              "equal to tx_free_thresh. (tx_free_thresh=%u"
2089                              " tx_rs_thresh=%u port=%d queue=%d)",
2090                              (unsigned int)tx_free_thresh,
2091                              (unsigned int)tx_rs_thresh,
2092                              (int)dev->data->port_id,
2093                              (int)queue_idx);
2094                 return I40E_ERR_PARAM;
2095         }
2096         if ((nb_desc % tx_rs_thresh) != 0) {
2097                 PMD_INIT_LOG(ERR, "tx_rs_thresh must be a divisor of the "
2098                              "number of TX descriptors. (tx_rs_thresh=%u"
2099                              " port=%d queue=%d)",
2100                              (unsigned int)tx_rs_thresh,
2101                              (int)dev->data->port_id,
2102                              (int)queue_idx);
2103                 return I40E_ERR_PARAM;
2104         }
2105         if ((tx_rs_thresh > 1) && (tx_conf->tx_thresh.wthresh != 0)) {
2106                 PMD_INIT_LOG(ERR, "TX WTHRESH must be set to 0 if "
2107                              "tx_rs_thresh is greater than 1. "
2108                              "(tx_rs_thresh=%u port=%d queue=%d)",
2109                              (unsigned int)tx_rs_thresh,
2110                              (int)dev->data->port_id,
2111                              (int)queue_idx);
2112                 return I40E_ERR_PARAM;
2113         }
2114
2115         /* Free memory if needed. */
2116         if (dev->data->tx_queues[queue_idx]) {
2117                 i40e_dev_tx_queue_release(dev->data->tx_queues[queue_idx]);
2118                 dev->data->tx_queues[queue_idx] = NULL;
2119         }
2120
2121         /* Allocate the TX queue data structure. */
2122         txq = rte_zmalloc_socket("i40e tx queue",
2123                                   sizeof(struct i40e_tx_queue),
2124                                   RTE_CACHE_LINE_SIZE,
2125                                   socket_id);
2126         if (!txq) {
2127                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
2128                             "tx queue structure");
2129                 return -ENOMEM;
2130         }
2131
2132         /* Allocate TX hardware ring descriptors. */
2133         ring_size = sizeof(struct i40e_tx_desc) * I40E_MAX_RING_DESC;
2134         ring_size = RTE_ALIGN(ring_size, I40E_DMA_MEM_ALIGN);
2135         tz = rte_eth_dma_zone_reserve(dev, "tx_ring", queue_idx,
2136                               ring_size, I40E_RING_BASE_ALIGN, socket_id);
2137         if (!tz) {
2138                 i40e_dev_tx_queue_release(txq);
2139                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for TX");
2140                 return -ENOMEM;
2141         }
2142
2143         txq->nb_tx_desc = nb_desc;
2144         txq->tx_rs_thresh = tx_rs_thresh;
2145         txq->tx_free_thresh = tx_free_thresh;
2146         txq->pthresh = tx_conf->tx_thresh.pthresh;
2147         txq->hthresh = tx_conf->tx_thresh.hthresh;
2148         txq->wthresh = tx_conf->tx_thresh.wthresh;
2149         txq->queue_id = queue_idx;
2150         txq->reg_idx = reg_idx;
2151         txq->port_id = dev->data->port_id;
2152         txq->txq_flags = tx_conf->txq_flags;
2153         txq->vsi = vsi;
2154         txq->tx_deferred_start = tx_conf->tx_deferred_start;
2155
2156         txq->tx_ring_phys_addr = tz->iova;
2157         txq->tx_ring = (struct i40e_tx_desc *)tz->addr;
2158
2159         /* Allocate software ring */
2160         txq->sw_ring =
2161                 rte_zmalloc_socket("i40e tx sw ring",
2162                                    sizeof(struct i40e_tx_entry) * nb_desc,
2163                                    RTE_CACHE_LINE_SIZE,
2164                                    socket_id);
2165         if (!txq->sw_ring) {
2166                 i40e_dev_tx_queue_release(txq);
2167                 PMD_DRV_LOG(ERR, "Failed to allocate memory for SW TX ring");
2168                 return -ENOMEM;
2169         }
2170
2171         i40e_reset_tx_queue(txq);
2172         txq->q_set = TRUE;
2173         dev->data->tx_queues[queue_idx] = txq;
2174
2175         /* Use a simple TX queue without offloads or multi segs if possible */
2176         i40e_set_tx_function_flag(dev, txq);
2177
2178         for (i = 0; i < I40E_MAX_TRAFFIC_CLASS; i++) {
2179                 if (!(vsi->enabled_tc & (1 << i)))
2180                         continue;
2181                 tc_mapping = rte_le_to_cpu_16(vsi->info.tc_mapping[i]);
2182                 base = (tc_mapping & I40E_AQ_VSI_TC_QUE_OFFSET_MASK) >>
2183                         I40E_AQ_VSI_TC_QUE_OFFSET_SHIFT;
2184                 bsf = (tc_mapping & I40E_AQ_VSI_TC_QUE_NUMBER_MASK) >>
2185                         I40E_AQ_VSI_TC_QUE_NUMBER_SHIFT;
2186
2187                 if (queue_idx >= base && queue_idx < (base + BIT(bsf)))
2188                         txq->dcb_tc = i;
2189         }
2190
2191         return 0;
2192 }
2193
2194 void
2195 i40e_dev_tx_queue_release(void *txq)
2196 {
2197         struct i40e_tx_queue *q = (struct i40e_tx_queue *)txq;
2198
2199         if (!q) {
2200                 PMD_DRV_LOG(DEBUG, "Pointer to TX queue is NULL");
2201                 return;
2202         }
2203
2204         i40e_tx_queue_release_mbufs(q);
2205         rte_free(q->sw_ring);
2206         rte_free(q);
2207 }
2208
2209 const struct rte_memzone *
2210 i40e_memzone_reserve(const char *name, uint32_t len, int socket_id)
2211 {
2212         const struct rte_memzone *mz;
2213
2214         mz = rte_memzone_lookup(name);
2215         if (mz)
2216                 return mz;
2217
2218         mz = rte_memzone_reserve_aligned(name, len, socket_id,
2219                         RTE_MEMZONE_IOVA_CONTIG, I40E_RING_BASE_ALIGN);
2220         return mz;
2221 }
2222
2223 void
2224 i40e_rx_queue_release_mbufs(struct i40e_rx_queue *rxq)
2225 {
2226         uint16_t i;
2227
2228         /* SSE Vector driver has a different way of releasing mbufs. */
2229         if (rxq->rx_using_sse) {
2230                 i40e_rx_queue_release_mbufs_vec(rxq);
2231                 return;
2232         }
2233
2234         if (!rxq->sw_ring) {
2235                 PMD_DRV_LOG(DEBUG, "Pointer to sw_ring is NULL");
2236                 return;
2237         }
2238
2239         for (i = 0; i < rxq->nb_rx_desc; i++) {
2240                 if (rxq->sw_ring[i].mbuf) {
2241                         rte_pktmbuf_free_seg(rxq->sw_ring[i].mbuf);
2242                         rxq->sw_ring[i].mbuf = NULL;
2243                 }
2244         }
2245 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
2246         if (rxq->rx_nb_avail == 0)
2247                 return;
2248         for (i = 0; i < rxq->rx_nb_avail; i++) {
2249                 struct rte_mbuf *mbuf;
2250
2251                 mbuf = rxq->rx_stage[rxq->rx_next_avail + i];
2252                 rte_pktmbuf_free_seg(mbuf);
2253         }
2254         rxq->rx_nb_avail = 0;
2255 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
2256 }
2257
2258 void
2259 i40e_reset_rx_queue(struct i40e_rx_queue *rxq)
2260 {
2261         unsigned i;
2262         uint16_t len;
2263
2264         if (!rxq) {
2265                 PMD_DRV_LOG(DEBUG, "Pointer to rxq is NULL");
2266                 return;
2267         }
2268
2269 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
2270         if (check_rx_burst_bulk_alloc_preconditions(rxq) == 0)
2271                 len = (uint16_t)(rxq->nb_rx_desc + RTE_PMD_I40E_RX_MAX_BURST);
2272         else
2273 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
2274                 len = rxq->nb_rx_desc;
2275
2276         for (i = 0; i < len * sizeof(union i40e_rx_desc); i++)
2277                 ((volatile char *)rxq->rx_ring)[i] = 0;
2278
2279         memset(&rxq->fake_mbuf, 0x0, sizeof(rxq->fake_mbuf));
2280         for (i = 0; i < RTE_PMD_I40E_RX_MAX_BURST; ++i)
2281                 rxq->sw_ring[rxq->nb_rx_desc + i].mbuf = &rxq->fake_mbuf;
2282
2283 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
2284         rxq->rx_nb_avail = 0;
2285         rxq->rx_next_avail = 0;
2286         rxq->rx_free_trigger = (uint16_t)(rxq->rx_free_thresh - 1);
2287 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
2288         rxq->rx_tail = 0;
2289         rxq->nb_rx_hold = 0;
2290         rxq->pkt_first_seg = NULL;
2291         rxq->pkt_last_seg = NULL;
2292
2293         rxq->rxrearm_start = 0;
2294         rxq->rxrearm_nb = 0;
2295 }
2296
2297 void
2298 i40e_tx_queue_release_mbufs(struct i40e_tx_queue *txq)
2299 {
2300         struct rte_eth_dev *dev;
2301         uint16_t i;
2302
2303         dev = &rte_eth_devices[txq->port_id];
2304
2305         if (!txq || !txq->sw_ring) {
2306                 PMD_DRV_LOG(DEBUG, "Pointer to rxq or sw_ring is NULL");
2307                 return;
2308         }
2309
2310         /**
2311          *  vPMD tx will not set sw_ring's mbuf to NULL after free,
2312          *  so need to free remains more carefully.
2313          */
2314         if (dev->tx_pkt_burst == i40e_xmit_pkts_vec_avx2 ||
2315                         dev->tx_pkt_burst == i40e_xmit_pkts_vec) {
2316                 i = txq->tx_next_dd - txq->tx_rs_thresh + 1;
2317                 if (txq->tx_tail < i) {
2318                         for (; i < txq->nb_tx_desc; i++) {
2319                                 rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
2320                                 txq->sw_ring[i].mbuf = NULL;
2321                         }
2322                         i = 0;
2323                 }
2324                 for (; i < txq->tx_tail; i++) {
2325                         rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
2326                         txq->sw_ring[i].mbuf = NULL;
2327                 }
2328         } else {
2329                 for (i = 0; i < txq->nb_tx_desc; i++) {
2330                         if (txq->sw_ring[i].mbuf) {
2331                                 rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
2332                                 txq->sw_ring[i].mbuf = NULL;
2333                         }
2334                 }
2335         }
2336 }
2337
2338 void
2339 i40e_reset_tx_queue(struct i40e_tx_queue *txq)
2340 {
2341         struct i40e_tx_entry *txe;
2342         uint16_t i, prev, size;
2343
2344         if (!txq) {
2345                 PMD_DRV_LOG(DEBUG, "Pointer to txq is NULL");
2346                 return;
2347         }
2348
2349         txe = txq->sw_ring;
2350         size = sizeof(struct i40e_tx_desc) * txq->nb_tx_desc;
2351         for (i = 0; i < size; i++)
2352                 ((volatile char *)txq->tx_ring)[i] = 0;
2353
2354         prev = (uint16_t)(txq->nb_tx_desc - 1);
2355         for (i = 0; i < txq->nb_tx_desc; i++) {
2356                 volatile struct i40e_tx_desc *txd = &txq->tx_ring[i];
2357
2358                 txd->cmd_type_offset_bsz =
2359                         rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DESC_DONE);
2360                 txe[i].mbuf =  NULL;
2361                 txe[i].last_id = i;
2362                 txe[prev].next_id = i;
2363                 prev = i;
2364         }
2365
2366         txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
2367         txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
2368
2369         txq->tx_tail = 0;
2370         txq->nb_tx_used = 0;
2371
2372         txq->last_desc_cleaned = (uint16_t)(txq->nb_tx_desc - 1);
2373         txq->nb_tx_free = (uint16_t)(txq->nb_tx_desc - 1);
2374 }
2375
2376 /* Init the TX queue in hardware */
2377 int
2378 i40e_tx_queue_init(struct i40e_tx_queue *txq)
2379 {
2380         enum i40e_status_code err = I40E_SUCCESS;
2381         struct i40e_vsi *vsi = txq->vsi;
2382         struct i40e_hw *hw = I40E_VSI_TO_HW(vsi);
2383         uint16_t pf_q = txq->reg_idx;
2384         struct i40e_hmc_obj_txq tx_ctx;
2385         uint32_t qtx_ctl;
2386
2387         /* clear the context structure first */
2388         memset(&tx_ctx, 0, sizeof(tx_ctx));
2389         tx_ctx.new_context = 1;
2390         tx_ctx.base = txq->tx_ring_phys_addr / I40E_QUEUE_BASE_ADDR_UNIT;
2391         tx_ctx.qlen = txq->nb_tx_desc;
2392
2393 #ifdef RTE_LIBRTE_IEEE1588
2394         tx_ctx.timesync_ena = 1;
2395 #endif
2396         tx_ctx.rdylist = rte_le_to_cpu_16(vsi->info.qs_handle[txq->dcb_tc]);
2397         if (vsi->type == I40E_VSI_FDIR)
2398                 tx_ctx.fd_ena = TRUE;
2399
2400         err = i40e_clear_lan_tx_queue_context(hw, pf_q);
2401         if (err != I40E_SUCCESS) {
2402                 PMD_DRV_LOG(ERR, "Failure of clean lan tx queue context");
2403                 return err;
2404         }
2405
2406         err = i40e_set_lan_tx_queue_context(hw, pf_q, &tx_ctx);
2407         if (err != I40E_SUCCESS) {
2408                 PMD_DRV_LOG(ERR, "Failure of set lan tx queue context");
2409                 return err;
2410         }
2411
2412         /* Now associate this queue with this PCI function */
2413         qtx_ctl = I40E_QTX_CTL_PF_QUEUE;
2414         qtx_ctl |= ((hw->pf_id << I40E_QTX_CTL_PF_INDX_SHIFT) &
2415                                         I40E_QTX_CTL_PF_INDX_MASK);
2416         I40E_WRITE_REG(hw, I40E_QTX_CTL(pf_q), qtx_ctl);
2417         I40E_WRITE_FLUSH(hw);
2418
2419         txq->qtx_tail = hw->hw_addr + I40E_QTX_TAIL(pf_q);
2420
2421         return err;
2422 }
2423
2424 int
2425 i40e_alloc_rx_queue_mbufs(struct i40e_rx_queue *rxq)
2426 {
2427         struct i40e_rx_entry *rxe = rxq->sw_ring;
2428         uint64_t dma_addr;
2429         uint16_t i;
2430
2431         for (i = 0; i < rxq->nb_rx_desc; i++) {
2432                 volatile union i40e_rx_desc *rxd;
2433                 struct rte_mbuf *mbuf = rte_mbuf_raw_alloc(rxq->mp);
2434
2435                 if (unlikely(!mbuf)) {
2436                         PMD_DRV_LOG(ERR, "Failed to allocate mbuf for RX");
2437                         return -ENOMEM;
2438                 }
2439
2440                 rte_mbuf_refcnt_set(mbuf, 1);
2441                 mbuf->next = NULL;
2442                 mbuf->data_off = RTE_PKTMBUF_HEADROOM;
2443                 mbuf->nb_segs = 1;
2444                 mbuf->port = rxq->port_id;
2445
2446                 dma_addr =
2447                         rte_cpu_to_le_64(rte_mbuf_data_iova_default(mbuf));
2448
2449                 rxd = &rxq->rx_ring[i];
2450                 rxd->read.pkt_addr = dma_addr;
2451                 rxd->read.hdr_addr = 0;
2452 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
2453                 rxd->read.rsvd1 = 0;
2454                 rxd->read.rsvd2 = 0;
2455 #endif /* RTE_LIBRTE_I40E_16BYTE_RX_DESC */
2456
2457                 rxe[i].mbuf = mbuf;
2458         }
2459
2460         return 0;
2461 }
2462
2463 /*
2464  * Calculate the buffer length, and check the jumbo frame
2465  * and maximum packet length.
2466  */
2467 static int
2468 i40e_rx_queue_config(struct i40e_rx_queue *rxq)
2469 {
2470         struct i40e_pf *pf = I40E_VSI_TO_PF(rxq->vsi);
2471         struct i40e_hw *hw = I40E_VSI_TO_HW(rxq->vsi);
2472         struct rte_eth_dev_data *data = pf->dev_data;
2473         uint16_t buf_size, len;
2474
2475         buf_size = (uint16_t)(rte_pktmbuf_data_room_size(rxq->mp) -
2476                 RTE_PKTMBUF_HEADROOM);
2477
2478         switch (pf->flags & (I40E_FLAG_HEADER_SPLIT_DISABLED |
2479                         I40E_FLAG_HEADER_SPLIT_ENABLED)) {
2480         case I40E_FLAG_HEADER_SPLIT_ENABLED: /* Not supported */
2481                 rxq->rx_hdr_len = RTE_ALIGN(I40E_RXBUF_SZ_1024,
2482                                 (1 << I40E_RXQ_CTX_HBUFF_SHIFT));
2483                 rxq->rx_buf_len = RTE_ALIGN(I40E_RXBUF_SZ_2048,
2484                                 (1 << I40E_RXQ_CTX_DBUFF_SHIFT));
2485                 rxq->hs_mode = i40e_header_split_enabled;
2486                 break;
2487         case I40E_FLAG_HEADER_SPLIT_DISABLED:
2488         default:
2489                 rxq->rx_hdr_len = 0;
2490                 rxq->rx_buf_len = RTE_ALIGN_FLOOR(buf_size,
2491                         (1 << I40E_RXQ_CTX_DBUFF_SHIFT));
2492                 rxq->hs_mode = i40e_header_split_none;
2493                 break;
2494         }
2495
2496         len = hw->func_caps.rx_buf_chain_len * rxq->rx_buf_len;
2497         rxq->max_pkt_len = RTE_MIN(len, data->dev_conf.rxmode.max_rx_pkt_len);
2498         if (data->dev_conf.rxmode.offloads & DEV_RX_OFFLOAD_JUMBO_FRAME) {
2499                 if (rxq->max_pkt_len <= ETHER_MAX_LEN ||
2500                         rxq->max_pkt_len > I40E_FRAME_SIZE_MAX) {
2501                         PMD_DRV_LOG(ERR, "maximum packet length must "
2502                                     "be larger than %u and smaller than %u,"
2503                                     "as jumbo frame is enabled",
2504                                     (uint32_t)ETHER_MAX_LEN,
2505                                     (uint32_t)I40E_FRAME_SIZE_MAX);
2506                         return I40E_ERR_CONFIG;
2507                 }
2508         } else {
2509                 if (rxq->max_pkt_len < ETHER_MIN_LEN ||
2510                         rxq->max_pkt_len > ETHER_MAX_LEN) {
2511                         PMD_DRV_LOG(ERR, "maximum packet length must be "
2512                                     "larger than %u and smaller than %u, "
2513                                     "as jumbo frame is disabled",
2514                                     (uint32_t)ETHER_MIN_LEN,
2515                                     (uint32_t)ETHER_MAX_LEN);
2516                         return I40E_ERR_CONFIG;
2517                 }
2518         }
2519
2520         return 0;
2521 }
2522
2523 /* Init the RX queue in hardware */
2524 int
2525 i40e_rx_queue_init(struct i40e_rx_queue *rxq)
2526 {
2527         int err = I40E_SUCCESS;
2528         struct i40e_hw *hw = I40E_VSI_TO_HW(rxq->vsi);
2529         struct rte_eth_dev_data *dev_data = I40E_VSI_TO_DEV_DATA(rxq->vsi);
2530         uint16_t pf_q = rxq->reg_idx;
2531         uint16_t buf_size;
2532         struct i40e_hmc_obj_rxq rx_ctx;
2533
2534         err = i40e_rx_queue_config(rxq);
2535         if (err < 0) {
2536                 PMD_DRV_LOG(ERR, "Failed to config RX queue");
2537                 return err;
2538         }
2539
2540         /* Clear the context structure first */
2541         memset(&rx_ctx, 0, sizeof(struct i40e_hmc_obj_rxq));
2542         rx_ctx.dbuff = rxq->rx_buf_len >> I40E_RXQ_CTX_DBUFF_SHIFT;
2543         rx_ctx.hbuff = rxq->rx_hdr_len >> I40E_RXQ_CTX_HBUFF_SHIFT;
2544
2545         rx_ctx.base = rxq->rx_ring_phys_addr / I40E_QUEUE_BASE_ADDR_UNIT;
2546         rx_ctx.qlen = rxq->nb_rx_desc;
2547 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
2548         rx_ctx.dsize = 1;
2549 #endif
2550         rx_ctx.dtype = rxq->hs_mode;
2551         if (rxq->hs_mode)
2552                 rx_ctx.hsplit_0 = I40E_HEADER_SPLIT_ALL;
2553         else
2554                 rx_ctx.hsplit_0 = I40E_HEADER_SPLIT_NONE;
2555         rx_ctx.rxmax = rxq->max_pkt_len;
2556         rx_ctx.tphrdesc_ena = 1;
2557         rx_ctx.tphwdesc_ena = 1;
2558         rx_ctx.tphdata_ena = 1;
2559         rx_ctx.tphhead_ena = 1;
2560         rx_ctx.lrxqthresh = 2;
2561         rx_ctx.crcstrip = (rxq->crc_len == 0) ? 1 : 0;
2562         rx_ctx.l2tsel = 1;
2563         /* showiv indicates if inner VLAN is stripped inside of tunnel
2564          * packet. When set it to 1, vlan information is stripped from
2565          * the inner header, but the hardware does not put it in the
2566          * descriptor. So set it zero by default.
2567          */
2568         rx_ctx.showiv = 0;
2569         rx_ctx.prefena = 1;
2570
2571         err = i40e_clear_lan_rx_queue_context(hw, pf_q);
2572         if (err != I40E_SUCCESS) {
2573                 PMD_DRV_LOG(ERR, "Failed to clear LAN RX queue context");
2574                 return err;
2575         }
2576         err = i40e_set_lan_rx_queue_context(hw, pf_q, &rx_ctx);
2577         if (err != I40E_SUCCESS) {
2578                 PMD_DRV_LOG(ERR, "Failed to set LAN RX queue context");
2579                 return err;
2580         }
2581
2582         rxq->qrx_tail = hw->hw_addr + I40E_QRX_TAIL(pf_q);
2583
2584         buf_size = (uint16_t)(rte_pktmbuf_data_room_size(rxq->mp) -
2585                 RTE_PKTMBUF_HEADROOM);
2586
2587         /* Check if scattered RX needs to be used. */
2588         if ((rxq->max_pkt_len + 2 * I40E_VLAN_TAG_SIZE) > buf_size) {
2589                 dev_data->scattered_rx = 1;
2590         }
2591
2592         /* Init the RX tail regieter. */
2593         I40E_PCI_REG_WRITE(rxq->qrx_tail, rxq->nb_rx_desc - 1);
2594
2595         return 0;
2596 }
2597
2598 void
2599 i40e_dev_clear_queues(struct rte_eth_dev *dev)
2600 {
2601         uint16_t i;
2602
2603         PMD_INIT_FUNC_TRACE();
2604
2605         for (i = 0; i < dev->data->nb_tx_queues; i++) {
2606                 if (!dev->data->tx_queues[i])
2607                         continue;
2608                 i40e_tx_queue_release_mbufs(dev->data->tx_queues[i]);
2609                 i40e_reset_tx_queue(dev->data->tx_queues[i]);
2610         }
2611
2612         for (i = 0; i < dev->data->nb_rx_queues; i++) {
2613                 if (!dev->data->rx_queues[i])
2614                         continue;
2615                 i40e_rx_queue_release_mbufs(dev->data->rx_queues[i]);
2616                 i40e_reset_rx_queue(dev->data->rx_queues[i]);
2617         }
2618 }
2619
2620 void
2621 i40e_dev_free_queues(struct rte_eth_dev *dev)
2622 {
2623         uint16_t i;
2624
2625         PMD_INIT_FUNC_TRACE();
2626
2627         for (i = 0; i < dev->data->nb_rx_queues; i++) {
2628                 if (!dev->data->rx_queues[i])
2629                         continue;
2630                 i40e_dev_rx_queue_release(dev->data->rx_queues[i]);
2631                 dev->data->rx_queues[i] = NULL;
2632         }
2633         dev->data->nb_rx_queues = 0;
2634
2635         for (i = 0; i < dev->data->nb_tx_queues; i++) {
2636                 if (!dev->data->tx_queues[i])
2637                         continue;
2638                 i40e_dev_tx_queue_release(dev->data->tx_queues[i]);
2639                 dev->data->tx_queues[i] = NULL;
2640         }
2641         dev->data->nb_tx_queues = 0;
2642 }
2643
2644 #define I40E_FDIR_NUM_TX_DESC  I40E_MIN_RING_DESC
2645 #define I40E_FDIR_NUM_RX_DESC  I40E_MIN_RING_DESC
2646
2647 enum i40e_status_code
2648 i40e_fdir_setup_tx_resources(struct i40e_pf *pf)
2649 {
2650         struct i40e_tx_queue *txq;
2651         const struct rte_memzone *tz = NULL;
2652         uint32_t ring_size;
2653         struct rte_eth_dev *dev;
2654
2655         if (!pf) {
2656                 PMD_DRV_LOG(ERR, "PF is not available");
2657                 return I40E_ERR_BAD_PTR;
2658         }
2659
2660         dev = pf->adapter->eth_dev;
2661
2662         /* Allocate the TX queue data structure. */
2663         txq = rte_zmalloc_socket("i40e fdir tx queue",
2664                                   sizeof(struct i40e_tx_queue),
2665                                   RTE_CACHE_LINE_SIZE,
2666                                   SOCKET_ID_ANY);
2667         if (!txq) {
2668                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
2669                                         "tx queue structure.");
2670                 return I40E_ERR_NO_MEMORY;
2671         }
2672
2673         /* Allocate TX hardware ring descriptors. */
2674         ring_size = sizeof(struct i40e_tx_desc) * I40E_FDIR_NUM_TX_DESC;
2675         ring_size = RTE_ALIGN(ring_size, I40E_DMA_MEM_ALIGN);
2676
2677         tz = rte_eth_dma_zone_reserve(dev, "fdir_tx_ring",
2678                                       I40E_FDIR_QUEUE_ID, ring_size,
2679                                       I40E_RING_BASE_ALIGN, SOCKET_ID_ANY);
2680         if (!tz) {
2681                 i40e_dev_tx_queue_release(txq);
2682                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for TX.");
2683                 return I40E_ERR_NO_MEMORY;
2684         }
2685
2686         txq->nb_tx_desc = I40E_FDIR_NUM_TX_DESC;
2687         txq->queue_id = I40E_FDIR_QUEUE_ID;
2688         txq->reg_idx = pf->fdir.fdir_vsi->base_queue;
2689         txq->vsi = pf->fdir.fdir_vsi;
2690
2691         txq->tx_ring_phys_addr = tz->iova;
2692         txq->tx_ring = (struct i40e_tx_desc *)tz->addr;
2693         /*
2694          * don't need to allocate software ring and reset for the fdir
2695          * program queue just set the queue has been configured.
2696          */
2697         txq->q_set = TRUE;
2698         pf->fdir.txq = txq;
2699
2700         return I40E_SUCCESS;
2701 }
2702
2703 enum i40e_status_code
2704 i40e_fdir_setup_rx_resources(struct i40e_pf *pf)
2705 {
2706         struct i40e_rx_queue *rxq;
2707         const struct rte_memzone *rz = NULL;
2708         uint32_t ring_size;
2709         struct rte_eth_dev *dev;
2710
2711         if (!pf) {
2712                 PMD_DRV_LOG(ERR, "PF is not available");
2713                 return I40E_ERR_BAD_PTR;
2714         }
2715
2716         dev = pf->adapter->eth_dev;
2717
2718         /* Allocate the RX queue data structure. */
2719         rxq = rte_zmalloc_socket("i40e fdir rx queue",
2720                                   sizeof(struct i40e_rx_queue),
2721                                   RTE_CACHE_LINE_SIZE,
2722                                   SOCKET_ID_ANY);
2723         if (!rxq) {
2724                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
2725                                         "rx queue structure.");
2726                 return I40E_ERR_NO_MEMORY;
2727         }
2728
2729         /* Allocate RX hardware ring descriptors. */
2730         ring_size = sizeof(union i40e_rx_desc) * I40E_FDIR_NUM_RX_DESC;
2731         ring_size = RTE_ALIGN(ring_size, I40E_DMA_MEM_ALIGN);
2732
2733         rz = rte_eth_dma_zone_reserve(dev, "fdir_rx_ring",
2734                                       I40E_FDIR_QUEUE_ID, ring_size,
2735                                       I40E_RING_BASE_ALIGN, SOCKET_ID_ANY);
2736         if (!rz) {
2737                 i40e_dev_rx_queue_release(rxq);
2738                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for RX.");
2739                 return I40E_ERR_NO_MEMORY;
2740         }
2741
2742         rxq->nb_rx_desc = I40E_FDIR_NUM_RX_DESC;
2743         rxq->queue_id = I40E_FDIR_QUEUE_ID;
2744         rxq->reg_idx = pf->fdir.fdir_vsi->base_queue;
2745         rxq->vsi = pf->fdir.fdir_vsi;
2746
2747         rxq->rx_ring_phys_addr = rz->iova;
2748         memset(rz->addr, 0, I40E_FDIR_NUM_RX_DESC * sizeof(union i40e_rx_desc));
2749         rxq->rx_ring = (union i40e_rx_desc *)rz->addr;
2750
2751         /*
2752          * Don't need to allocate software ring and reset for the fdir
2753          * rx queue, just set the queue has been configured.
2754          */
2755         rxq->q_set = TRUE;
2756         pf->fdir.rxq = rxq;
2757
2758         return I40E_SUCCESS;
2759 }
2760
2761 void
2762 i40e_rxq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
2763         struct rte_eth_rxq_info *qinfo)
2764 {
2765         struct i40e_rx_queue *rxq;
2766
2767         rxq = dev->data->rx_queues[queue_id];
2768
2769         qinfo->mp = rxq->mp;
2770         qinfo->scattered_rx = dev->data->scattered_rx;
2771         qinfo->nb_desc = rxq->nb_rx_desc;
2772
2773         qinfo->conf.rx_free_thresh = rxq->rx_free_thresh;
2774         qinfo->conf.rx_drop_en = rxq->drop_en;
2775         qinfo->conf.rx_deferred_start = rxq->rx_deferred_start;
2776         qinfo->conf.offloads = rxq->offloads;
2777 }
2778
2779 void
2780 i40e_txq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
2781         struct rte_eth_txq_info *qinfo)
2782 {
2783         struct i40e_tx_queue *txq;
2784
2785         txq = dev->data->tx_queues[queue_id];
2786
2787         qinfo->nb_desc = txq->nb_tx_desc;
2788
2789         qinfo->conf.tx_thresh.pthresh = txq->pthresh;
2790         qinfo->conf.tx_thresh.hthresh = txq->hthresh;
2791         qinfo->conf.tx_thresh.wthresh = txq->wthresh;
2792
2793         qinfo->conf.tx_free_thresh = txq->tx_free_thresh;
2794         qinfo->conf.tx_rs_thresh = txq->tx_rs_thresh;
2795         qinfo->conf.txq_flags = txq->txq_flags;
2796         qinfo->conf.tx_deferred_start = txq->tx_deferred_start;
2797 }
2798
2799 void __attribute__((cold))
2800 i40e_set_rx_function(struct rte_eth_dev *dev)
2801 {
2802         struct i40e_adapter *ad =
2803                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
2804         uint16_t rx_using_sse, i;
2805         /* In order to allow Vector Rx there are a few configuration
2806          * conditions to be met and Rx Bulk Allocation should be allowed.
2807          */
2808         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
2809                 if (i40e_rx_vec_dev_conf_condition_check(dev) ||
2810                     !ad->rx_bulk_alloc_allowed) {
2811                         PMD_INIT_LOG(DEBUG, "Port[%d] doesn't meet"
2812                                      " Vector Rx preconditions",
2813                                      dev->data->port_id);
2814
2815                         ad->rx_vec_allowed = false;
2816                 }
2817                 if (ad->rx_vec_allowed) {
2818                         for (i = 0; i < dev->data->nb_rx_queues; i++) {
2819                                 struct i40e_rx_queue *rxq =
2820                                         dev->data->rx_queues[i];
2821
2822                                 if (rxq && i40e_rxq_vec_setup(rxq)) {
2823                                         ad->rx_vec_allowed = false;
2824                                         break;
2825                                 }
2826                         }
2827                 }
2828         }
2829
2830         if (dev->data->scattered_rx) {
2831                 /* Set the non-LRO scattered callback: there are Vector and
2832                  * single allocation versions.
2833                  */
2834                 if (ad->rx_vec_allowed) {
2835                         PMD_INIT_LOG(DEBUG, "Using Vector Scattered Rx "
2836                                             "callback (port=%d).",
2837                                      dev->data->port_id);
2838
2839                         dev->rx_pkt_burst = i40e_recv_scattered_pkts_vec;
2840 #ifdef RTE_ARCH_X86
2841                         /*
2842                          * since AVX frequency can be different to base
2843                          * frequency, limit use of AVX2 version to later
2844                          * plaforms, not all those that could theoretically
2845                          * run it.
2846                          */
2847                         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F))
2848                                 dev->rx_pkt_burst =
2849                                         i40e_recv_scattered_pkts_vec_avx2;
2850 #endif
2851                 } else {
2852                         PMD_INIT_LOG(DEBUG, "Using a Scattered with bulk "
2853                                            "allocation callback (port=%d).",
2854                                      dev->data->port_id);
2855                         dev->rx_pkt_burst = i40e_recv_scattered_pkts;
2856                 }
2857         /* If parameters allow we are going to choose between the following
2858          * callbacks:
2859          *    - Vector
2860          *    - Bulk Allocation
2861          *    - Single buffer allocation (the simplest one)
2862          */
2863         } else if (ad->rx_vec_allowed) {
2864                 PMD_INIT_LOG(DEBUG, "Vector rx enabled, please make sure RX "
2865                                     "burst size no less than %d (port=%d).",
2866                              RTE_I40E_DESCS_PER_LOOP,
2867                              dev->data->port_id);
2868
2869                 dev->rx_pkt_burst = i40e_recv_pkts_vec;
2870 #ifdef RTE_ARCH_X86
2871                 /*
2872                  * since AVX frequency can be different to base
2873                  * frequency, limit use of AVX2 version to later
2874                  * plaforms, not all those that could theoretically
2875                  * run it.
2876                  */
2877                 if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F))
2878                         dev->rx_pkt_burst = i40e_recv_pkts_vec_avx2;
2879 #endif
2880         } else if (ad->rx_bulk_alloc_allowed) {
2881                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions are "
2882                                     "satisfied. Rx Burst Bulk Alloc function "
2883                                     "will be used on port=%d.",
2884                              dev->data->port_id);
2885
2886                 dev->rx_pkt_burst = i40e_recv_pkts_bulk_alloc;
2887         } else {
2888                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions are not "
2889                                     "satisfied, or Scattered Rx is requested "
2890                                     "(port=%d).",
2891                              dev->data->port_id);
2892
2893                 dev->rx_pkt_burst = i40e_recv_pkts;
2894         }
2895
2896         /* Propagate information about RX function choice through all queues. */
2897         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
2898                 rx_using_sse =
2899                         (dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec ||
2900                          dev->rx_pkt_burst == i40e_recv_pkts_vec ||
2901                          dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec_avx2 ||
2902                          dev->rx_pkt_burst == i40e_recv_pkts_vec_avx2);
2903
2904                 for (i = 0; i < dev->data->nb_rx_queues; i++) {
2905                         struct i40e_rx_queue *rxq = dev->data->rx_queues[i];
2906
2907                         if (rxq)
2908                                 rxq->rx_using_sse = rx_using_sse;
2909                 }
2910         }
2911 }
2912
2913 void __attribute__((cold))
2914 i40e_set_tx_function_flag(struct rte_eth_dev *dev, struct i40e_tx_queue *txq)
2915 {
2916         struct i40e_adapter *ad =
2917                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
2918
2919         /* Use a simple Tx queue (no offloads, no multi segs) if possible */
2920         if (((txq->txq_flags & I40E_SIMPLE_FLAGS) == I40E_SIMPLE_FLAGS)
2921                         && (txq->tx_rs_thresh >= RTE_PMD_I40E_TX_MAX_BURST)) {
2922                 if (txq->tx_rs_thresh <= RTE_I40E_TX_MAX_FREE_BUF_SZ) {
2923                         PMD_INIT_LOG(DEBUG, "Vector tx"
2924                                      " can be enabled on this txq.");
2925
2926                 } else {
2927                         ad->tx_vec_allowed = false;
2928                 }
2929         } else {
2930                 ad->tx_simple_allowed = false;
2931         }
2932 }
2933
2934 void __attribute__((cold))
2935 i40e_set_tx_function(struct rte_eth_dev *dev)
2936 {
2937         struct i40e_adapter *ad =
2938                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
2939         int i;
2940
2941         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
2942                 if (ad->tx_vec_allowed) {
2943                         for (i = 0; i < dev->data->nb_tx_queues; i++) {
2944                                 struct i40e_tx_queue *txq =
2945                                         dev->data->tx_queues[i];
2946
2947                                 if (txq && i40e_txq_vec_setup(txq)) {
2948                                         ad->tx_vec_allowed = false;
2949                                         break;
2950                                 }
2951                         }
2952                 }
2953         }
2954
2955         if (ad->tx_simple_allowed) {
2956                 if (ad->tx_vec_allowed) {
2957                         PMD_INIT_LOG(DEBUG, "Vector tx finally be used.");
2958                         dev->tx_pkt_burst = i40e_xmit_pkts_vec;
2959 #ifdef RTE_ARCH_X86
2960                         /*
2961                          * since AVX frequency can be different to base
2962                          * frequency, limit use of AVX2 version to later
2963                          * plaforms, not all those that could theoretically
2964                          * run it.
2965                          */
2966                         if (rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F))
2967                                 dev->tx_pkt_burst = i40e_xmit_pkts_vec_avx2;
2968 #endif
2969                 } else {
2970                         PMD_INIT_LOG(DEBUG, "Simple tx finally be used.");
2971                         dev->tx_pkt_burst = i40e_xmit_pkts_simple;
2972                 }
2973                 dev->tx_pkt_prepare = NULL;
2974         } else {
2975                 PMD_INIT_LOG(DEBUG, "Xmit tx finally be used.");
2976                 dev->tx_pkt_burst = i40e_xmit_pkts;
2977                 dev->tx_pkt_prepare = i40e_prep_pkts;
2978         }
2979 }
2980
2981 void __attribute__((cold))
2982 i40e_set_default_ptype_table(struct rte_eth_dev *dev)
2983 {
2984         struct i40e_adapter *ad =
2985                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
2986         int i;
2987
2988         for (i = 0; i < I40E_MAX_PKT_TYPE; i++)
2989                 ad->ptype_tbl[i] = i40e_get_default_pkt_type(i);
2990 }
2991
2992 void __attribute__((cold))
2993 i40e_set_default_pctype_table(struct rte_eth_dev *dev)
2994 {
2995         struct i40e_adapter *ad =
2996                         I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
2997         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
2998         int i;
2999
3000         for (i = 0; i < I40E_FLOW_TYPE_MAX; i++)
3001                 ad->pctypes_tbl[i] = 0ULL;
3002         ad->flow_types_mask = 0ULL;
3003         ad->pctypes_mask = 0ULL;
3004
3005         ad->pctypes_tbl[RTE_ETH_FLOW_FRAG_IPV4] =
3006                                 (1ULL << I40E_FILTER_PCTYPE_FRAG_IPV4);
3007         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_UDP] =
3008                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_UDP);
3009         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_TCP] =
3010                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_TCP);
3011         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_SCTP] =
3012                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_SCTP);
3013         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_OTHER] =
3014                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_OTHER);
3015         ad->pctypes_tbl[RTE_ETH_FLOW_FRAG_IPV6] =
3016                                 (1ULL << I40E_FILTER_PCTYPE_FRAG_IPV6);
3017         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_UDP] =
3018                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_UDP);
3019         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_TCP] =
3020                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_TCP);
3021         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_SCTP] =
3022                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_SCTP);
3023         ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_OTHER] =
3024                                 (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_OTHER);
3025         ad->pctypes_tbl[RTE_ETH_FLOW_L2_PAYLOAD] =
3026                                 (1ULL << I40E_FILTER_PCTYPE_L2_PAYLOAD);
3027
3028         if (hw->mac.type == I40E_MAC_X722) {
3029                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_UDP] |=
3030                         (1ULL << I40E_FILTER_PCTYPE_NONF_UNICAST_IPV4_UDP);
3031                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_UDP] |=
3032                         (1ULL << I40E_FILTER_PCTYPE_NONF_MULTICAST_IPV4_UDP);
3033                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV4_TCP] |=
3034                         (1ULL << I40E_FILTER_PCTYPE_NONF_IPV4_TCP_SYN_NO_ACK);
3035                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_UDP] |=
3036                         (1ULL << I40E_FILTER_PCTYPE_NONF_UNICAST_IPV6_UDP);
3037                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_UDP] |=
3038                         (1ULL << I40E_FILTER_PCTYPE_NONF_MULTICAST_IPV6_UDP);
3039                 ad->pctypes_tbl[RTE_ETH_FLOW_NONFRAG_IPV6_TCP] |=
3040                         (1ULL << I40E_FILTER_PCTYPE_NONF_IPV6_TCP_SYN_NO_ACK);
3041         }
3042
3043         for (i = 0; i < I40E_FLOW_TYPE_MAX; i++) {
3044                 if (ad->pctypes_tbl[i])
3045                         ad->flow_types_mask |= (1ULL << i);
3046                 ad->pctypes_mask |= ad->pctypes_tbl[i];
3047         }
3048 }
3049
3050 /* Stubs needed for linkage when CONFIG_RTE_I40E_INC_VECTOR is set to 'n' */
3051 int __attribute__((weak))
3052 i40e_rx_vec_dev_conf_condition_check(struct rte_eth_dev __rte_unused *dev)
3053 {
3054         return -1;
3055 }
3056
3057 uint16_t __attribute__((weak))
3058 i40e_recv_pkts_vec(
3059         void __rte_unused *rx_queue,
3060         struct rte_mbuf __rte_unused **rx_pkts,
3061         uint16_t __rte_unused nb_pkts)
3062 {
3063         return 0;
3064 }
3065
3066 uint16_t __attribute__((weak))
3067 i40e_recv_scattered_pkts_vec(
3068         void __rte_unused *rx_queue,
3069         struct rte_mbuf __rte_unused **rx_pkts,
3070         uint16_t __rte_unused nb_pkts)
3071 {
3072         return 0;
3073 }
3074
3075 uint16_t __attribute__((weak))
3076 i40e_recv_pkts_vec_avx2(void __rte_unused *rx_queue,
3077                         struct rte_mbuf __rte_unused **rx_pkts,
3078                         uint16_t __rte_unused nb_pkts)
3079 {
3080         return 0;
3081 }
3082
3083 uint16_t __attribute__((weak))
3084 i40e_recv_scattered_pkts_vec_avx2(void __rte_unused *rx_queue,
3085                         struct rte_mbuf __rte_unused **rx_pkts,
3086                         uint16_t __rte_unused nb_pkts)
3087 {
3088         return 0;
3089 }
3090
3091 int __attribute__((weak))
3092 i40e_rxq_vec_setup(struct i40e_rx_queue __rte_unused *rxq)
3093 {
3094         return -1;
3095 }
3096
3097 int __attribute__((weak))
3098 i40e_txq_vec_setup(struct i40e_tx_queue __rte_unused *txq)
3099 {
3100         return -1;
3101 }
3102
3103 void __attribute__((weak))
3104 i40e_rx_queue_release_mbufs_vec(struct i40e_rx_queue __rte_unused*rxq)
3105 {
3106         return;
3107 }
3108
3109 uint16_t __attribute__((weak))
3110 i40e_xmit_fixed_burst_vec(void __rte_unused * tx_queue,
3111                           struct rte_mbuf __rte_unused **tx_pkts,
3112                           uint16_t __rte_unused nb_pkts)
3113 {
3114         return 0;
3115 }
3116
3117 uint16_t __attribute__((weak))
3118 i40e_xmit_pkts_vec_avx2(void __rte_unused * tx_queue,
3119                           struct rte_mbuf __rte_unused **tx_pkts,
3120                           uint16_t __rte_unused nb_pkts)
3121 {
3122         return 0;
3123 }