9ccc35475850c839e3b6f5a19945d30afa01d8ca
[dpdk.git] / drivers / net / i40e / i40e_rxtx.c
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright(c) 2010-2016 Intel Corporation. All rights reserved.
5  *   All rights reserved.
6  *
7  *   Redistribution and use in source and binary forms, with or without
8  *   modification, are permitted provided that the following conditions
9  *   are met:
10  *
11  *     * Redistributions of source code must retain the above copyright
12  *       notice, this list of conditions and the following disclaimer.
13  *     * Redistributions in binary form must reproduce the above copyright
14  *       notice, this list of conditions and the following disclaimer in
15  *       the documentation and/or other materials provided with the
16  *       distribution.
17  *     * Neither the name of Intel Corporation nor the names of its
18  *       contributors may be used to endorse or promote products derived
19  *       from this software without specific prior written permission.
20  *
21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  */
33
34 #include <stdio.h>
35 #include <stdlib.h>
36 #include <string.h>
37 #include <errno.h>
38 #include <stdint.h>
39 #include <stdarg.h>
40 #include <unistd.h>
41 #include <inttypes.h>
42 #include <sys/queue.h>
43
44 #include <rte_string_fns.h>
45 #include <rte_memzone.h>
46 #include <rte_mbuf.h>
47 #include <rte_malloc.h>
48 #include <rte_ether.h>
49 #include <rte_ethdev.h>
50 #include <rte_tcp.h>
51 #include <rte_sctp.h>
52 #include <rte_udp.h>
53 #include <rte_ip.h>
54 #include <rte_net.h>
55
56 #include "i40e_logs.h"
57 #include "base/i40e_prototype.h"
58 #include "base/i40e_type.h"
59 #include "i40e_ethdev.h"
60 #include "i40e_rxtx.h"
61
62 #define DEFAULT_TX_RS_THRESH   32
63 #define DEFAULT_TX_FREE_THRESH 32
64 #define I40E_MAX_PKT_TYPE      256
65
66 #define I40E_TX_MAX_BURST  32
67
68 #define I40E_DMA_MEM_ALIGN 4096
69
70 /* Base address of the HW descriptor ring should be 128B aligned. */
71 #define I40E_RING_BASE_ALIGN    128
72
73 #define I40E_SIMPLE_FLAGS ((uint32_t)ETH_TXQ_FLAGS_NOMULTSEGS | \
74                                         ETH_TXQ_FLAGS_NOOFFLOADS)
75
76 #define I40E_TXD_CMD (I40E_TX_DESC_CMD_EOP | I40E_TX_DESC_CMD_RS)
77
78 #ifdef RTE_LIBRTE_IEEE1588
79 #define I40E_TX_IEEE1588_TMST PKT_TX_IEEE1588_TMST
80 #else
81 #define I40E_TX_IEEE1588_TMST 0
82 #endif
83
84 #define I40E_TX_CKSUM_OFFLOAD_MASK (             \
85                 PKT_TX_IP_CKSUM |                \
86                 PKT_TX_L4_MASK |                 \
87                 PKT_TX_TCP_SEG |                 \
88                 PKT_TX_OUTER_IP_CKSUM)
89
90 #define I40E_TX_OFFLOAD_MASK (  \
91                 PKT_TX_IP_CKSUM |       \
92                 PKT_TX_L4_MASK |        \
93                 PKT_TX_OUTER_IP_CKSUM | \
94                 PKT_TX_TCP_SEG |        \
95                 PKT_TX_QINQ_PKT |       \
96                 PKT_TX_VLAN_PKT |       \
97                 PKT_TX_TUNNEL_MASK |    \
98                 I40E_TX_IEEE1588_TMST)
99
100 #define I40E_TX_OFFLOAD_NOTSUP_MASK \
101                 (PKT_TX_OFFLOAD_MASK ^ I40E_TX_OFFLOAD_MASK)
102
103 static uint16_t i40e_xmit_pkts_simple(void *tx_queue,
104                                       struct rte_mbuf **tx_pkts,
105                                       uint16_t nb_pkts);
106
107 static inline void
108 i40e_rxd_to_vlan_tci(struct rte_mbuf *mb, volatile union i40e_rx_desc *rxdp)
109 {
110         if (rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len) &
111                 (1 << I40E_RX_DESC_STATUS_L2TAG1P_SHIFT)) {
112                 mb->ol_flags |= PKT_RX_VLAN_PKT | PKT_RX_VLAN_STRIPPED;
113                 mb->vlan_tci =
114                         rte_le_to_cpu_16(rxdp->wb.qword0.lo_dword.l2tag1);
115                 PMD_RX_LOG(DEBUG, "Descriptor l2tag1: %u",
116                            rte_le_to_cpu_16(rxdp->wb.qword0.lo_dword.l2tag1));
117         } else {
118                 mb->vlan_tci = 0;
119         }
120 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
121         if (rte_le_to_cpu_16(rxdp->wb.qword2.ext_status) &
122                 (1 << I40E_RX_DESC_EXT_STATUS_L2TAG2P_SHIFT)) {
123                 mb->ol_flags |= PKT_RX_QINQ_STRIPPED;
124                 mb->vlan_tci_outer = mb->vlan_tci;
125                 mb->vlan_tci = rte_le_to_cpu_16(rxdp->wb.qword2.l2tag2_2);
126                 PMD_RX_LOG(DEBUG, "Descriptor l2tag2_1: %u, l2tag2_2: %u",
127                            rte_le_to_cpu_16(rxdp->wb.qword2.l2tag2_1),
128                            rte_le_to_cpu_16(rxdp->wb.qword2.l2tag2_2));
129         } else {
130                 mb->vlan_tci_outer = 0;
131         }
132 #endif
133         PMD_RX_LOG(DEBUG, "Mbuf vlan_tci: %u, vlan_tci_outer: %u",
134                    mb->vlan_tci, mb->vlan_tci_outer);
135 }
136
137 /* Translate the rx descriptor status to pkt flags */
138 static inline uint64_t
139 i40e_rxd_status_to_pkt_flags(uint64_t qword)
140 {
141         uint64_t flags;
142
143         /* Check if RSS_HASH */
144         flags = (((qword >> I40E_RX_DESC_STATUS_FLTSTAT_SHIFT) &
145                                         I40E_RX_DESC_FLTSTAT_RSS_HASH) ==
146                         I40E_RX_DESC_FLTSTAT_RSS_HASH) ? PKT_RX_RSS_HASH : 0;
147
148         /* Check if FDIR Match */
149         flags |= (qword & (1 << I40E_RX_DESC_STATUS_FLM_SHIFT) ?
150                                                         PKT_RX_FDIR : 0);
151
152         return flags;
153 }
154
155 static inline uint64_t
156 i40e_rxd_error_to_pkt_flags(uint64_t qword)
157 {
158         uint64_t flags = 0;
159         uint64_t error_bits = (qword >> I40E_RXD_QW1_ERROR_SHIFT);
160
161 #define I40E_RX_ERR_BITS 0x3f
162         if (likely((error_bits & I40E_RX_ERR_BITS) == 0)) {
163                 flags |= (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD);
164                 return flags;
165         }
166
167         if (unlikely(error_bits & (1 << I40E_RX_DESC_ERROR_IPE_SHIFT)))
168                 flags |= PKT_RX_IP_CKSUM_BAD;
169         else
170                 flags |= PKT_RX_IP_CKSUM_GOOD;
171
172         if (unlikely(error_bits & (1 << I40E_RX_DESC_ERROR_L4E_SHIFT)))
173                 flags |= PKT_RX_L4_CKSUM_BAD;
174         else
175                 flags |= PKT_RX_L4_CKSUM_GOOD;
176
177         if (unlikely(error_bits & (1 << I40E_RX_DESC_ERROR_EIPE_SHIFT)))
178                 flags |= PKT_RX_EIP_CKSUM_BAD;
179
180         return flags;
181 }
182
183 /* Function to check and set the ieee1588 timesync index and get the
184  * appropriate flags.
185  */
186 #ifdef RTE_LIBRTE_IEEE1588
187 static inline uint64_t
188 i40e_get_iee15888_flags(struct rte_mbuf *mb, uint64_t qword)
189 {
190         uint64_t pkt_flags = 0;
191         uint16_t tsyn = (qword & (I40E_RXD_QW1_STATUS_TSYNVALID_MASK
192                                   | I40E_RXD_QW1_STATUS_TSYNINDX_MASK))
193                                     >> I40E_RX_DESC_STATUS_TSYNINDX_SHIFT;
194
195         if ((mb->packet_type & RTE_PTYPE_L2_MASK)
196                         == RTE_PTYPE_L2_ETHER_TIMESYNC)
197                 pkt_flags = PKT_RX_IEEE1588_PTP;
198         if (tsyn & 0x04) {
199                 pkt_flags |= PKT_RX_IEEE1588_TMST;
200                 mb->timesync = tsyn & 0x03;
201         }
202
203         return pkt_flags;
204 }
205 #endif
206
207 #define I40E_RX_DESC_EXT_STATUS_FLEXBH_MASK   0x03
208 #define I40E_RX_DESC_EXT_STATUS_FLEXBH_FD_ID  0x01
209 #define I40E_RX_DESC_EXT_STATUS_FLEXBH_FLEX   0x02
210 #define I40E_RX_DESC_EXT_STATUS_FLEXBL_MASK   0x03
211 #define I40E_RX_DESC_EXT_STATUS_FLEXBL_FLEX   0x01
212
213 static inline uint64_t
214 i40e_rxd_build_fdir(volatile union i40e_rx_desc *rxdp, struct rte_mbuf *mb)
215 {
216         uint64_t flags = 0;
217 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
218         uint16_t flexbh, flexbl;
219
220         flexbh = (rte_le_to_cpu_32(rxdp->wb.qword2.ext_status) >>
221                 I40E_RX_DESC_EXT_STATUS_FLEXBH_SHIFT) &
222                 I40E_RX_DESC_EXT_STATUS_FLEXBH_MASK;
223         flexbl = (rte_le_to_cpu_32(rxdp->wb.qword2.ext_status) >>
224                 I40E_RX_DESC_EXT_STATUS_FLEXBL_SHIFT) &
225                 I40E_RX_DESC_EXT_STATUS_FLEXBL_MASK;
226
227
228         if (flexbh == I40E_RX_DESC_EXT_STATUS_FLEXBH_FD_ID) {
229                 mb->hash.fdir.hi =
230                         rte_le_to_cpu_32(rxdp->wb.qword3.hi_dword.fd_id);
231                 flags |= PKT_RX_FDIR_ID;
232         } else if (flexbh == I40E_RX_DESC_EXT_STATUS_FLEXBH_FLEX) {
233                 mb->hash.fdir.hi =
234                         rte_le_to_cpu_32(rxdp->wb.qword3.hi_dword.flex_bytes_hi);
235                 flags |= PKT_RX_FDIR_FLX;
236         }
237         if (flexbl == I40E_RX_DESC_EXT_STATUS_FLEXBL_FLEX) {
238                 mb->hash.fdir.lo =
239                         rte_le_to_cpu_32(rxdp->wb.qword3.lo_dword.flex_bytes_lo);
240                 flags |= PKT_RX_FDIR_FLX;
241         }
242 #else
243         mb->hash.fdir.hi =
244                 rte_le_to_cpu_32(rxdp->wb.qword0.hi_dword.fd_id);
245         flags |= PKT_RX_FDIR_ID;
246 #endif
247         return flags;
248 }
249
250 static inline void
251 i40e_parse_tunneling_params(uint64_t ol_flags,
252                             union i40e_tx_offload tx_offload,
253                             uint32_t *cd_tunneling)
254 {
255         /* EIPT: External (outer) IP header type */
256         if (ol_flags & PKT_TX_OUTER_IP_CKSUM)
257                 *cd_tunneling |= I40E_TX_CTX_EXT_IP_IPV4;
258         else if (ol_flags & PKT_TX_OUTER_IPV4)
259                 *cd_tunneling |= I40E_TX_CTX_EXT_IP_IPV4_NO_CSUM;
260         else if (ol_flags & PKT_TX_OUTER_IPV6)
261                 *cd_tunneling |= I40E_TX_CTX_EXT_IP_IPV6;
262
263         /* EIPLEN: External (outer) IP header length, in DWords */
264         *cd_tunneling |= (tx_offload.outer_l3_len >> 2) <<
265                 I40E_TXD_CTX_QW0_EXT_IPLEN_SHIFT;
266
267         /* L4TUNT: L4 Tunneling Type */
268         switch (ol_flags & PKT_TX_TUNNEL_MASK) {
269         case PKT_TX_TUNNEL_IPIP:
270                 /* for non UDP / GRE tunneling, set to 00b */
271                 break;
272         case PKT_TX_TUNNEL_VXLAN:
273         case PKT_TX_TUNNEL_GENEVE:
274                 *cd_tunneling |= I40E_TXD_CTX_UDP_TUNNELING;
275                 break;
276         case PKT_TX_TUNNEL_GRE:
277                 *cd_tunneling |= I40E_TXD_CTX_GRE_TUNNELING;
278                 break;
279         default:
280                 PMD_TX_LOG(ERR, "Tunnel type not supported");
281                 return;
282         }
283
284         /* L4TUNLEN: L4 Tunneling Length, in Words
285          *
286          * We depend on app to set rte_mbuf.l2_len correctly.
287          * For IP in GRE it should be set to the length of the GRE
288          * header;
289          * for MAC in GRE or MAC in UDP it should be set to the length
290          * of the GRE or UDP headers plus the inner MAC up to including
291          * its last Ethertype.
292          */
293         *cd_tunneling |= (tx_offload.l2_len >> 1) <<
294                 I40E_TXD_CTX_QW0_NATLEN_SHIFT;
295 }
296
297 static inline void
298 i40e_txd_enable_checksum(uint64_t ol_flags,
299                         uint32_t *td_cmd,
300                         uint32_t *td_offset,
301                         union i40e_tx_offload tx_offload)
302 {
303         /* Set MACLEN */
304         if (ol_flags & PKT_TX_TUNNEL_MASK)
305                 *td_offset |= (tx_offload.outer_l2_len >> 1)
306                                 << I40E_TX_DESC_LENGTH_MACLEN_SHIFT;
307         else
308                 *td_offset |= (tx_offload.l2_len >> 1)
309                         << I40E_TX_DESC_LENGTH_MACLEN_SHIFT;
310
311         /* Enable L3 checksum offloads */
312         if (ol_flags & PKT_TX_IP_CKSUM) {
313                 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV4_CSUM;
314                 *td_offset |= (tx_offload.l3_len >> 2)
315                                 << I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
316         } else if (ol_flags & PKT_TX_IPV4) {
317                 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV4;
318                 *td_offset |= (tx_offload.l3_len >> 2)
319                                 << I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
320         } else if (ol_flags & PKT_TX_IPV6) {
321                 *td_cmd |= I40E_TX_DESC_CMD_IIPT_IPV6;
322                 *td_offset |= (tx_offload.l3_len >> 2)
323                                 << I40E_TX_DESC_LENGTH_IPLEN_SHIFT;
324         }
325
326         if (ol_flags & PKT_TX_TCP_SEG) {
327                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_TCP;
328                 *td_offset |= (tx_offload.l4_len >> 2)
329                         << I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
330                 return;
331         }
332
333         /* Enable L4 checksum offloads */
334         switch (ol_flags & PKT_TX_L4_MASK) {
335         case PKT_TX_TCP_CKSUM:
336                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_TCP;
337                 *td_offset |= (sizeof(struct tcp_hdr) >> 2) <<
338                                 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
339                 break;
340         case PKT_TX_SCTP_CKSUM:
341                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_SCTP;
342                 *td_offset |= (sizeof(struct sctp_hdr) >> 2) <<
343                                 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
344                 break;
345         case PKT_TX_UDP_CKSUM:
346                 *td_cmd |= I40E_TX_DESC_CMD_L4T_EOFT_UDP;
347                 *td_offset |= (sizeof(struct udp_hdr) >> 2) <<
348                                 I40E_TX_DESC_LENGTH_L4_FC_LEN_SHIFT;
349                 break;
350         default:
351                 break;
352         }
353 }
354
355 /* Construct the tx flags */
356 static inline uint64_t
357 i40e_build_ctob(uint32_t td_cmd,
358                 uint32_t td_offset,
359                 unsigned int size,
360                 uint32_t td_tag)
361 {
362         return rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DATA |
363                         ((uint64_t)td_cmd  << I40E_TXD_QW1_CMD_SHIFT) |
364                         ((uint64_t)td_offset << I40E_TXD_QW1_OFFSET_SHIFT) |
365                         ((uint64_t)size  << I40E_TXD_QW1_TX_BUF_SZ_SHIFT) |
366                         ((uint64_t)td_tag  << I40E_TXD_QW1_L2TAG1_SHIFT));
367 }
368
369 static inline int
370 i40e_xmit_cleanup(struct i40e_tx_queue *txq)
371 {
372         struct i40e_tx_entry *sw_ring = txq->sw_ring;
373         volatile struct i40e_tx_desc *txd = txq->tx_ring;
374         uint16_t last_desc_cleaned = txq->last_desc_cleaned;
375         uint16_t nb_tx_desc = txq->nb_tx_desc;
376         uint16_t desc_to_clean_to;
377         uint16_t nb_tx_to_clean;
378
379         desc_to_clean_to = (uint16_t)(last_desc_cleaned + txq->tx_rs_thresh);
380         if (desc_to_clean_to >= nb_tx_desc)
381                 desc_to_clean_to = (uint16_t)(desc_to_clean_to - nb_tx_desc);
382
383         desc_to_clean_to = sw_ring[desc_to_clean_to].last_id;
384         if ((txd[desc_to_clean_to].cmd_type_offset_bsz &
385                         rte_cpu_to_le_64(I40E_TXD_QW1_DTYPE_MASK)) !=
386                         rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DESC_DONE)) {
387                 PMD_TX_FREE_LOG(DEBUG, "TX descriptor %4u is not done "
388                         "(port=%d queue=%d)", desc_to_clean_to,
389                                 txq->port_id, txq->queue_id);
390                 return -1;
391         }
392
393         if (last_desc_cleaned > desc_to_clean_to)
394                 nb_tx_to_clean = (uint16_t)((nb_tx_desc - last_desc_cleaned) +
395                                                         desc_to_clean_to);
396         else
397                 nb_tx_to_clean = (uint16_t)(desc_to_clean_to -
398                                         last_desc_cleaned);
399
400         txd[desc_to_clean_to].cmd_type_offset_bsz = 0;
401
402         txq->last_desc_cleaned = desc_to_clean_to;
403         txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + nb_tx_to_clean);
404
405         return 0;
406 }
407
408 static inline int
409 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
410 check_rx_burst_bulk_alloc_preconditions(struct i40e_rx_queue *rxq)
411 #else
412 check_rx_burst_bulk_alloc_preconditions(__rte_unused struct i40e_rx_queue *rxq)
413 #endif
414 {
415         int ret = 0;
416
417 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
418         if (!(rxq->rx_free_thresh >= RTE_PMD_I40E_RX_MAX_BURST)) {
419                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
420                              "rxq->rx_free_thresh=%d, "
421                              "RTE_PMD_I40E_RX_MAX_BURST=%d",
422                              rxq->rx_free_thresh, RTE_PMD_I40E_RX_MAX_BURST);
423                 ret = -EINVAL;
424         } else if (!(rxq->rx_free_thresh < rxq->nb_rx_desc)) {
425                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
426                              "rxq->rx_free_thresh=%d, "
427                              "rxq->nb_rx_desc=%d",
428                              rxq->rx_free_thresh, rxq->nb_rx_desc);
429                 ret = -EINVAL;
430         } else if (rxq->nb_rx_desc % rxq->rx_free_thresh != 0) {
431                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions: "
432                              "rxq->nb_rx_desc=%d, "
433                              "rxq->rx_free_thresh=%d",
434                              rxq->nb_rx_desc, rxq->rx_free_thresh);
435                 ret = -EINVAL;
436         }
437 #else
438         ret = -EINVAL;
439 #endif
440
441         return ret;
442 }
443
444 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
445 #define I40E_LOOK_AHEAD 8
446 #if (I40E_LOOK_AHEAD != 8)
447 #error "PMD I40E: I40E_LOOK_AHEAD must be 8\n"
448 #endif
449 static inline int
450 i40e_rx_scan_hw_ring(struct i40e_rx_queue *rxq)
451 {
452         volatile union i40e_rx_desc *rxdp;
453         struct i40e_rx_entry *rxep;
454         struct rte_mbuf *mb;
455         uint16_t pkt_len;
456         uint64_t qword1;
457         uint32_t rx_status;
458         int32_t s[I40E_LOOK_AHEAD], nb_dd;
459         int32_t i, j, nb_rx = 0;
460         uint64_t pkt_flags;
461
462         rxdp = &rxq->rx_ring[rxq->rx_tail];
463         rxep = &rxq->sw_ring[rxq->rx_tail];
464
465         qword1 = rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len);
466         rx_status = (qword1 & I40E_RXD_QW1_STATUS_MASK) >>
467                                 I40E_RXD_QW1_STATUS_SHIFT;
468
469         /* Make sure there is at least 1 packet to receive */
470         if (!(rx_status & (1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
471                 return 0;
472
473         /**
474          * Scan LOOK_AHEAD descriptors at a time to determine which
475          * descriptors reference packets that are ready to be received.
476          */
477         for (i = 0; i < RTE_PMD_I40E_RX_MAX_BURST; i+=I40E_LOOK_AHEAD,
478                         rxdp += I40E_LOOK_AHEAD, rxep += I40E_LOOK_AHEAD) {
479                 /* Read desc statuses backwards to avoid race condition */
480                 for (j = I40E_LOOK_AHEAD - 1; j >= 0; j--) {
481                         qword1 = rte_le_to_cpu_64(\
482                                 rxdp[j].wb.qword1.status_error_len);
483                         s[j] = (qword1 & I40E_RXD_QW1_STATUS_MASK) >>
484                                         I40E_RXD_QW1_STATUS_SHIFT;
485                 }
486
487                 rte_smp_rmb();
488
489                 /* Compute how many status bits were set */
490                 for (j = 0, nb_dd = 0; j < I40E_LOOK_AHEAD; j++)
491                         nb_dd += s[j] & (1 << I40E_RX_DESC_STATUS_DD_SHIFT);
492
493                 nb_rx += nb_dd;
494
495                 /* Translate descriptor info to mbuf parameters */
496                 for (j = 0; j < nb_dd; j++) {
497                         mb = rxep[j].mbuf;
498                         qword1 = rte_le_to_cpu_64(\
499                                 rxdp[j].wb.qword1.status_error_len);
500                         pkt_len = ((qword1 & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
501                                 I40E_RXD_QW1_LENGTH_PBUF_SHIFT) - rxq->crc_len;
502                         mb->data_len = pkt_len;
503                         mb->pkt_len = pkt_len;
504                         mb->ol_flags = 0;
505                         i40e_rxd_to_vlan_tci(mb, &rxdp[j]);
506                         pkt_flags = i40e_rxd_status_to_pkt_flags(qword1);
507                         pkt_flags |= i40e_rxd_error_to_pkt_flags(qword1);
508                         mb->packet_type =
509                                 i40e_rxd_pkt_type_mapping((uint8_t)((qword1 &
510                                                 I40E_RXD_QW1_PTYPE_MASK) >>
511                                                 I40E_RXD_QW1_PTYPE_SHIFT));
512                         if (pkt_flags & PKT_RX_RSS_HASH)
513                                 mb->hash.rss = rte_le_to_cpu_32(\
514                                         rxdp[j].wb.qword0.hi_dword.rss);
515                         if (pkt_flags & PKT_RX_FDIR)
516                                 pkt_flags |= i40e_rxd_build_fdir(&rxdp[j], mb);
517
518 #ifdef RTE_LIBRTE_IEEE1588
519                         pkt_flags |= i40e_get_iee15888_flags(mb, qword1);
520 #endif
521                         mb->ol_flags |= pkt_flags;
522
523                 }
524
525                 for (j = 0; j < I40E_LOOK_AHEAD; j++)
526                         rxq->rx_stage[i + j] = rxep[j].mbuf;
527
528                 if (nb_dd != I40E_LOOK_AHEAD)
529                         break;
530         }
531
532         /* Clear software ring entries */
533         for (i = 0; i < nb_rx; i++)
534                 rxq->sw_ring[rxq->rx_tail + i].mbuf = NULL;
535
536         return nb_rx;
537 }
538
539 static inline uint16_t
540 i40e_rx_fill_from_stage(struct i40e_rx_queue *rxq,
541                         struct rte_mbuf **rx_pkts,
542                         uint16_t nb_pkts)
543 {
544         uint16_t i;
545         struct rte_mbuf **stage = &rxq->rx_stage[rxq->rx_next_avail];
546
547         nb_pkts = (uint16_t)RTE_MIN(nb_pkts, rxq->rx_nb_avail);
548
549         for (i = 0; i < nb_pkts; i++)
550                 rx_pkts[i] = stage[i];
551
552         rxq->rx_nb_avail = (uint16_t)(rxq->rx_nb_avail - nb_pkts);
553         rxq->rx_next_avail = (uint16_t)(rxq->rx_next_avail + nb_pkts);
554
555         return nb_pkts;
556 }
557
558 static inline int
559 i40e_rx_alloc_bufs(struct i40e_rx_queue *rxq)
560 {
561         volatile union i40e_rx_desc *rxdp;
562         struct i40e_rx_entry *rxep;
563         struct rte_mbuf *mb;
564         uint16_t alloc_idx, i;
565         uint64_t dma_addr;
566         int diag;
567
568         /* Allocate buffers in bulk */
569         alloc_idx = (uint16_t)(rxq->rx_free_trigger -
570                                 (rxq->rx_free_thresh - 1));
571         rxep = &(rxq->sw_ring[alloc_idx]);
572         diag = rte_mempool_get_bulk(rxq->mp, (void *)rxep,
573                                         rxq->rx_free_thresh);
574         if (unlikely(diag != 0)) {
575                 PMD_DRV_LOG(ERR, "Failed to get mbufs in bulk");
576                 return -ENOMEM;
577         }
578
579         rxdp = &rxq->rx_ring[alloc_idx];
580         for (i = 0; i < rxq->rx_free_thresh; i++) {
581                 if (likely(i < (rxq->rx_free_thresh - 1)))
582                         /* Prefetch next mbuf */
583                         rte_prefetch0(rxep[i + 1].mbuf);
584
585                 mb = rxep[i].mbuf;
586                 rte_mbuf_refcnt_set(mb, 1);
587                 mb->next = NULL;
588                 mb->data_off = RTE_PKTMBUF_HEADROOM;
589                 mb->nb_segs = 1;
590                 mb->port = rxq->port_id;
591                 dma_addr = rte_cpu_to_le_64(\
592                         rte_mbuf_data_dma_addr_default(mb));
593                 rxdp[i].read.hdr_addr = 0;
594                 rxdp[i].read.pkt_addr = dma_addr;
595         }
596
597         /* Update rx tail regsiter */
598         rte_wmb();
599         I40E_PCI_REG_WRITE_RELAXED(rxq->qrx_tail, rxq->rx_free_trigger);
600
601         rxq->rx_free_trigger =
602                 (uint16_t)(rxq->rx_free_trigger + rxq->rx_free_thresh);
603         if (rxq->rx_free_trigger >= rxq->nb_rx_desc)
604                 rxq->rx_free_trigger = (uint16_t)(rxq->rx_free_thresh - 1);
605
606         return 0;
607 }
608
609 static inline uint16_t
610 rx_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
611 {
612         struct i40e_rx_queue *rxq = (struct i40e_rx_queue *)rx_queue;
613         uint16_t nb_rx = 0;
614
615         if (!nb_pkts)
616                 return 0;
617
618         if (rxq->rx_nb_avail)
619                 return i40e_rx_fill_from_stage(rxq, rx_pkts, nb_pkts);
620
621         nb_rx = (uint16_t)i40e_rx_scan_hw_ring(rxq);
622         rxq->rx_next_avail = 0;
623         rxq->rx_nb_avail = nb_rx;
624         rxq->rx_tail = (uint16_t)(rxq->rx_tail + nb_rx);
625
626         if (rxq->rx_tail > rxq->rx_free_trigger) {
627                 if (i40e_rx_alloc_bufs(rxq) != 0) {
628                         uint16_t i, j;
629
630                         PMD_RX_LOG(DEBUG, "Rx mbuf alloc failed for "
631                                    "port_id=%u, queue_id=%u",
632                                    rxq->port_id, rxq->queue_id);
633                         rxq->rx_nb_avail = 0;
634                         rxq->rx_tail = (uint16_t)(rxq->rx_tail - nb_rx);
635                         for (i = 0, j = rxq->rx_tail; i < nb_rx; i++, j++)
636                                 rxq->sw_ring[j].mbuf = rxq->rx_stage[i];
637
638                         return 0;
639                 }
640         }
641
642         if (rxq->rx_tail >= rxq->nb_rx_desc)
643                 rxq->rx_tail = 0;
644
645         if (rxq->rx_nb_avail)
646                 return i40e_rx_fill_from_stage(rxq, rx_pkts, nb_pkts);
647
648         return 0;
649 }
650
651 static uint16_t
652 i40e_recv_pkts_bulk_alloc(void *rx_queue,
653                           struct rte_mbuf **rx_pkts,
654                           uint16_t nb_pkts)
655 {
656         uint16_t nb_rx = 0, n, count;
657
658         if (unlikely(nb_pkts == 0))
659                 return 0;
660
661         if (likely(nb_pkts <= RTE_PMD_I40E_RX_MAX_BURST))
662                 return rx_recv_pkts(rx_queue, rx_pkts, nb_pkts);
663
664         while (nb_pkts) {
665                 n = RTE_MIN(nb_pkts, RTE_PMD_I40E_RX_MAX_BURST);
666                 count = rx_recv_pkts(rx_queue, &rx_pkts[nb_rx], n);
667                 nb_rx = (uint16_t)(nb_rx + count);
668                 nb_pkts = (uint16_t)(nb_pkts - count);
669                 if (count < n)
670                         break;
671         }
672
673         return nb_rx;
674 }
675 #else
676 static uint16_t
677 i40e_recv_pkts_bulk_alloc(void __rte_unused *rx_queue,
678                           struct rte_mbuf __rte_unused **rx_pkts,
679                           uint16_t __rte_unused nb_pkts)
680 {
681         return 0;
682 }
683 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
684
685 uint16_t
686 i40e_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
687 {
688         struct i40e_rx_queue *rxq;
689         volatile union i40e_rx_desc *rx_ring;
690         volatile union i40e_rx_desc *rxdp;
691         union i40e_rx_desc rxd;
692         struct i40e_rx_entry *sw_ring;
693         struct i40e_rx_entry *rxe;
694         struct rte_mbuf *rxm;
695         struct rte_mbuf *nmb;
696         uint16_t nb_rx;
697         uint32_t rx_status;
698         uint64_t qword1;
699         uint16_t rx_packet_len;
700         uint16_t rx_id, nb_hold;
701         uint64_t dma_addr;
702         uint64_t pkt_flags;
703
704         nb_rx = 0;
705         nb_hold = 0;
706         rxq = rx_queue;
707         rx_id = rxq->rx_tail;
708         rx_ring = rxq->rx_ring;
709         sw_ring = rxq->sw_ring;
710
711         while (nb_rx < nb_pkts) {
712                 rxdp = &rx_ring[rx_id];
713                 qword1 = rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len);
714                 rx_status = (qword1 & I40E_RXD_QW1_STATUS_MASK)
715                                 >> I40E_RXD_QW1_STATUS_SHIFT;
716
717                 /* Check the DD bit first */
718                 if (!(rx_status & (1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
719                         break;
720
721                 nmb = rte_mbuf_raw_alloc(rxq->mp);
722                 if (unlikely(!nmb))
723                         break;
724                 rxd = *rxdp;
725
726                 nb_hold++;
727                 rxe = &sw_ring[rx_id];
728                 rx_id++;
729                 if (unlikely(rx_id == rxq->nb_rx_desc))
730                         rx_id = 0;
731
732                 /* Prefetch next mbuf */
733                 rte_prefetch0(sw_ring[rx_id].mbuf);
734
735                 /**
736                  * When next RX descriptor is on a cache line boundary,
737                  * prefetch the next 4 RX descriptors and next 8 pointers
738                  * to mbufs.
739                  */
740                 if ((rx_id & 0x3) == 0) {
741                         rte_prefetch0(&rx_ring[rx_id]);
742                         rte_prefetch0(&sw_ring[rx_id]);
743                 }
744                 rxm = rxe->mbuf;
745                 rxe->mbuf = nmb;
746                 dma_addr =
747                         rte_cpu_to_le_64(rte_mbuf_data_dma_addr_default(nmb));
748                 rxdp->read.hdr_addr = 0;
749                 rxdp->read.pkt_addr = dma_addr;
750
751                 rx_packet_len = ((qword1 & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
752                                 I40E_RXD_QW1_LENGTH_PBUF_SHIFT) - rxq->crc_len;
753
754                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
755                 rte_prefetch0(RTE_PTR_ADD(rxm->buf_addr, RTE_PKTMBUF_HEADROOM));
756                 rxm->nb_segs = 1;
757                 rxm->next = NULL;
758                 rxm->pkt_len = rx_packet_len;
759                 rxm->data_len = rx_packet_len;
760                 rxm->port = rxq->port_id;
761                 rxm->ol_flags = 0;
762                 i40e_rxd_to_vlan_tci(rxm, &rxd);
763                 pkt_flags = i40e_rxd_status_to_pkt_flags(qword1);
764                 pkt_flags |= i40e_rxd_error_to_pkt_flags(qword1);
765                 rxm->packet_type =
766                         i40e_rxd_pkt_type_mapping((uint8_t)((qword1 &
767                         I40E_RXD_QW1_PTYPE_MASK) >> I40E_RXD_QW1_PTYPE_SHIFT));
768                 if (pkt_flags & PKT_RX_RSS_HASH)
769                         rxm->hash.rss =
770                                 rte_le_to_cpu_32(rxd.wb.qword0.hi_dword.rss);
771                 if (pkt_flags & PKT_RX_FDIR)
772                         pkt_flags |= i40e_rxd_build_fdir(&rxd, rxm);
773
774 #ifdef RTE_LIBRTE_IEEE1588
775                 pkt_flags |= i40e_get_iee15888_flags(rxm, qword1);
776 #endif
777                 rxm->ol_flags |= pkt_flags;
778
779                 rx_pkts[nb_rx++] = rxm;
780         }
781         rxq->rx_tail = rx_id;
782
783         /**
784          * If the number of free RX descriptors is greater than the RX free
785          * threshold of the queue, advance the receive tail register of queue.
786          * Update that register with the value of the last processed RX
787          * descriptor minus 1.
788          */
789         nb_hold = (uint16_t)(nb_hold + rxq->nb_rx_hold);
790         if (nb_hold > rxq->rx_free_thresh) {
791                 rx_id = (uint16_t) ((rx_id == 0) ?
792                         (rxq->nb_rx_desc - 1) : (rx_id - 1));
793                 I40E_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
794                 nb_hold = 0;
795         }
796         rxq->nb_rx_hold = nb_hold;
797
798         return nb_rx;
799 }
800
801 uint16_t
802 i40e_recv_scattered_pkts(void *rx_queue,
803                          struct rte_mbuf **rx_pkts,
804                          uint16_t nb_pkts)
805 {
806         struct i40e_rx_queue *rxq = rx_queue;
807         volatile union i40e_rx_desc *rx_ring = rxq->rx_ring;
808         volatile union i40e_rx_desc *rxdp;
809         union i40e_rx_desc rxd;
810         struct i40e_rx_entry *sw_ring = rxq->sw_ring;
811         struct i40e_rx_entry *rxe;
812         struct rte_mbuf *first_seg = rxq->pkt_first_seg;
813         struct rte_mbuf *last_seg = rxq->pkt_last_seg;
814         struct rte_mbuf *nmb, *rxm;
815         uint16_t rx_id = rxq->rx_tail;
816         uint16_t nb_rx = 0, nb_hold = 0, rx_packet_len;
817         uint32_t rx_status;
818         uint64_t qword1;
819         uint64_t dma_addr;
820         uint64_t pkt_flags;
821
822         while (nb_rx < nb_pkts) {
823                 rxdp = &rx_ring[rx_id];
824                 qword1 = rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len);
825                 rx_status = (qword1 & I40E_RXD_QW1_STATUS_MASK) >>
826                                         I40E_RXD_QW1_STATUS_SHIFT;
827
828                 /* Check the DD bit */
829                 if (!(rx_status & (1 << I40E_RX_DESC_STATUS_DD_SHIFT)))
830                         break;
831
832                 nmb = rte_mbuf_raw_alloc(rxq->mp);
833                 if (unlikely(!nmb))
834                         break;
835                 rxd = *rxdp;
836                 nb_hold++;
837                 rxe = &sw_ring[rx_id];
838                 rx_id++;
839                 if (rx_id == rxq->nb_rx_desc)
840                         rx_id = 0;
841
842                 /* Prefetch next mbuf */
843                 rte_prefetch0(sw_ring[rx_id].mbuf);
844
845                 /**
846                  * When next RX descriptor is on a cache line boundary,
847                  * prefetch the next 4 RX descriptors and next 8 pointers
848                  * to mbufs.
849                  */
850                 if ((rx_id & 0x3) == 0) {
851                         rte_prefetch0(&rx_ring[rx_id]);
852                         rte_prefetch0(&sw_ring[rx_id]);
853                 }
854
855                 rxm = rxe->mbuf;
856                 rxe->mbuf = nmb;
857                 dma_addr =
858                         rte_cpu_to_le_64(rte_mbuf_data_dma_addr_default(nmb));
859
860                 /* Set data buffer address and data length of the mbuf */
861                 rxdp->read.hdr_addr = 0;
862                 rxdp->read.pkt_addr = dma_addr;
863                 rx_packet_len = (qword1 & I40E_RXD_QW1_LENGTH_PBUF_MASK) >>
864                                         I40E_RXD_QW1_LENGTH_PBUF_SHIFT;
865                 rxm->data_len = rx_packet_len;
866                 rxm->data_off = RTE_PKTMBUF_HEADROOM;
867
868                 /**
869                  * If this is the first buffer of the received packet, set the
870                  * pointer to the first mbuf of the packet and initialize its
871                  * context. Otherwise, update the total length and the number
872                  * of segments of the current scattered packet, and update the
873                  * pointer to the last mbuf of the current packet.
874                  */
875                 if (!first_seg) {
876                         first_seg = rxm;
877                         first_seg->nb_segs = 1;
878                         first_seg->pkt_len = rx_packet_len;
879                 } else {
880                         first_seg->pkt_len =
881                                 (uint16_t)(first_seg->pkt_len +
882                                                 rx_packet_len);
883                         first_seg->nb_segs++;
884                         last_seg->next = rxm;
885                 }
886
887                 /**
888                  * If this is not the last buffer of the received packet,
889                  * update the pointer to the last mbuf of the current scattered
890                  * packet and continue to parse the RX ring.
891                  */
892                 if (!(rx_status & (1 << I40E_RX_DESC_STATUS_EOF_SHIFT))) {
893                         last_seg = rxm;
894                         continue;
895                 }
896
897                 /**
898                  * This is the last buffer of the received packet. If the CRC
899                  * is not stripped by the hardware:
900                  *  - Subtract the CRC length from the total packet length.
901                  *  - If the last buffer only contains the whole CRC or a part
902                  *  of it, free the mbuf associated to the last buffer. If part
903                  *  of the CRC is also contained in the previous mbuf, subtract
904                  *  the length of that CRC part from the data length of the
905                  *  previous mbuf.
906                  */
907                 rxm->next = NULL;
908                 if (unlikely(rxq->crc_len > 0)) {
909                         first_seg->pkt_len -= ETHER_CRC_LEN;
910                         if (rx_packet_len <= ETHER_CRC_LEN) {
911                                 rte_pktmbuf_free_seg(rxm);
912                                 first_seg->nb_segs--;
913                                 last_seg->data_len =
914                                         (uint16_t)(last_seg->data_len -
915                                         (ETHER_CRC_LEN - rx_packet_len));
916                                 last_seg->next = NULL;
917                         } else
918                                 rxm->data_len = (uint16_t)(rx_packet_len -
919                                                                 ETHER_CRC_LEN);
920                 }
921
922                 first_seg->port = rxq->port_id;
923                 first_seg->ol_flags = 0;
924                 i40e_rxd_to_vlan_tci(first_seg, &rxd);
925                 pkt_flags = i40e_rxd_status_to_pkt_flags(qword1);
926                 pkt_flags |= i40e_rxd_error_to_pkt_flags(qword1);
927                 first_seg->packet_type =
928                         i40e_rxd_pkt_type_mapping((uint8_t)((qword1 &
929                         I40E_RXD_QW1_PTYPE_MASK) >> I40E_RXD_QW1_PTYPE_SHIFT));
930                 if (pkt_flags & PKT_RX_RSS_HASH)
931                         first_seg->hash.rss =
932                                 rte_le_to_cpu_32(rxd.wb.qword0.hi_dword.rss);
933                 if (pkt_flags & PKT_RX_FDIR)
934                         pkt_flags |= i40e_rxd_build_fdir(&rxd, first_seg);
935
936 #ifdef RTE_LIBRTE_IEEE1588
937                 pkt_flags |= i40e_get_iee15888_flags(first_seg, qword1);
938 #endif
939                 first_seg->ol_flags |= pkt_flags;
940
941                 /* Prefetch data of first segment, if configured to do so. */
942                 rte_prefetch0(RTE_PTR_ADD(first_seg->buf_addr,
943                         first_seg->data_off));
944                 rx_pkts[nb_rx++] = first_seg;
945                 first_seg = NULL;
946         }
947
948         /* Record index of the next RX descriptor to probe. */
949         rxq->rx_tail = rx_id;
950         rxq->pkt_first_seg = first_seg;
951         rxq->pkt_last_seg = last_seg;
952
953         /**
954          * If the number of free RX descriptors is greater than the RX free
955          * threshold of the queue, advance the Receive Descriptor Tail (RDT)
956          * register. Update the RDT with the value of the last processed RX
957          * descriptor minus 1, to guarantee that the RDT register is never
958          * equal to the RDH register, which creates a "full" ring situtation
959          * from the hardware point of view.
960          */
961         nb_hold = (uint16_t)(nb_hold + rxq->nb_rx_hold);
962         if (nb_hold > rxq->rx_free_thresh) {
963                 rx_id = (uint16_t)(rx_id == 0 ?
964                         (rxq->nb_rx_desc - 1) : (rx_id - 1));
965                 I40E_PCI_REG_WRITE(rxq->qrx_tail, rx_id);
966                 nb_hold = 0;
967         }
968         rxq->nb_rx_hold = nb_hold;
969
970         return nb_rx;
971 }
972
973 /* Check if the context descriptor is needed for TX offloading */
974 static inline uint16_t
975 i40e_calc_context_desc(uint64_t flags)
976 {
977         static uint64_t mask = PKT_TX_OUTER_IP_CKSUM |
978                 PKT_TX_TCP_SEG |
979                 PKT_TX_QINQ_PKT |
980                 PKT_TX_TUNNEL_MASK;
981
982 #ifdef RTE_LIBRTE_IEEE1588
983         mask |= PKT_TX_IEEE1588_TMST;
984 #endif
985
986         return (flags & mask) ? 1 : 0;
987 }
988
989 /* set i40e TSO context descriptor */
990 static inline uint64_t
991 i40e_set_tso_ctx(struct rte_mbuf *mbuf, union i40e_tx_offload tx_offload)
992 {
993         uint64_t ctx_desc = 0;
994         uint32_t cd_cmd, hdr_len, cd_tso_len;
995
996         if (!tx_offload.l4_len) {
997                 PMD_DRV_LOG(DEBUG, "L4 length set to 0");
998                 return ctx_desc;
999         }
1000
1001         /**
1002          * in case of non tunneling packet, the outer_l2_len and
1003          * outer_l3_len must be 0.
1004          */
1005         hdr_len = tx_offload.outer_l2_len +
1006                 tx_offload.outer_l3_len +
1007                 tx_offload.l2_len +
1008                 tx_offload.l3_len +
1009                 tx_offload.l4_len;
1010
1011         cd_cmd = I40E_TX_CTX_DESC_TSO;
1012         cd_tso_len = mbuf->pkt_len - hdr_len;
1013         ctx_desc |= ((uint64_t)cd_cmd << I40E_TXD_CTX_QW1_CMD_SHIFT) |
1014                 ((uint64_t)cd_tso_len <<
1015                  I40E_TXD_CTX_QW1_TSO_LEN_SHIFT) |
1016                 ((uint64_t)mbuf->tso_segsz <<
1017                  I40E_TXD_CTX_QW1_MSS_SHIFT);
1018
1019         return ctx_desc;
1020 }
1021
1022 uint16_t
1023 i40e_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
1024 {
1025         struct i40e_tx_queue *txq;
1026         struct i40e_tx_entry *sw_ring;
1027         struct i40e_tx_entry *txe, *txn;
1028         volatile struct i40e_tx_desc *txd;
1029         volatile struct i40e_tx_desc *txr;
1030         struct rte_mbuf *tx_pkt;
1031         struct rte_mbuf *m_seg;
1032         uint32_t cd_tunneling_params;
1033         uint16_t tx_id;
1034         uint16_t nb_tx;
1035         uint32_t td_cmd;
1036         uint32_t td_offset;
1037         uint32_t tx_flags;
1038         uint32_t td_tag;
1039         uint64_t ol_flags;
1040         uint16_t nb_used;
1041         uint16_t nb_ctx;
1042         uint16_t tx_last;
1043         uint16_t slen;
1044         uint64_t buf_dma_addr;
1045         union i40e_tx_offload tx_offload = {0};
1046
1047         txq = tx_queue;
1048         sw_ring = txq->sw_ring;
1049         txr = txq->tx_ring;
1050         tx_id = txq->tx_tail;
1051         txe = &sw_ring[tx_id];
1052
1053         /* Check if the descriptor ring needs to be cleaned. */
1054         if (txq->nb_tx_free < txq->tx_free_thresh)
1055                 i40e_xmit_cleanup(txq);
1056
1057         for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
1058                 td_cmd = 0;
1059                 td_tag = 0;
1060                 td_offset = 0;
1061                 tx_flags = 0;
1062
1063                 tx_pkt = *tx_pkts++;
1064                 RTE_MBUF_PREFETCH_TO_FREE(txe->mbuf);
1065
1066                 ol_flags = tx_pkt->ol_flags;
1067                 tx_offload.l2_len = tx_pkt->l2_len;
1068                 tx_offload.l3_len = tx_pkt->l3_len;
1069                 tx_offload.outer_l2_len = tx_pkt->outer_l2_len;
1070                 tx_offload.outer_l3_len = tx_pkt->outer_l3_len;
1071                 tx_offload.l4_len = tx_pkt->l4_len;
1072                 tx_offload.tso_segsz = tx_pkt->tso_segsz;
1073
1074                 /* Calculate the number of context descriptors needed. */
1075                 nb_ctx = i40e_calc_context_desc(ol_flags);
1076
1077                 /**
1078                  * The number of descriptors that must be allocated for
1079                  * a packet equals to the number of the segments of that
1080                  * packet plus 1 context descriptor if needed.
1081                  */
1082                 nb_used = (uint16_t)(tx_pkt->nb_segs + nb_ctx);
1083                 tx_last = (uint16_t)(tx_id + nb_used - 1);
1084
1085                 /* Circular ring */
1086                 if (tx_last >= txq->nb_tx_desc)
1087                         tx_last = (uint16_t)(tx_last - txq->nb_tx_desc);
1088
1089                 if (nb_used > txq->nb_tx_free) {
1090                         if (i40e_xmit_cleanup(txq) != 0) {
1091                                 if (nb_tx == 0)
1092                                         return 0;
1093                                 goto end_of_tx;
1094                         }
1095                         if (unlikely(nb_used > txq->tx_rs_thresh)) {
1096                                 while (nb_used > txq->nb_tx_free) {
1097                                         if (i40e_xmit_cleanup(txq) != 0) {
1098                                                 if (nb_tx == 0)
1099                                                         return 0;
1100                                                 goto end_of_tx;
1101                                         }
1102                                 }
1103                         }
1104                 }
1105
1106                 /* Descriptor based VLAN insertion */
1107                 if (ol_flags & (PKT_TX_VLAN_PKT | PKT_TX_QINQ_PKT)) {
1108                         tx_flags |= tx_pkt->vlan_tci <<
1109                                 I40E_TX_FLAG_L2TAG1_SHIFT;
1110                         tx_flags |= I40E_TX_FLAG_INSERT_VLAN;
1111                         td_cmd |= I40E_TX_DESC_CMD_IL2TAG1;
1112                         td_tag = (tx_flags & I40E_TX_FLAG_L2TAG1_MASK) >>
1113                                                 I40E_TX_FLAG_L2TAG1_SHIFT;
1114                 }
1115
1116                 /* Always enable CRC offload insertion */
1117                 td_cmd |= I40E_TX_DESC_CMD_ICRC;
1118
1119                 /* Fill in tunneling parameters if necessary */
1120                 cd_tunneling_params = 0;
1121                 if (ol_flags & PKT_TX_TUNNEL_MASK)
1122                         i40e_parse_tunneling_params(ol_flags, tx_offload,
1123                                                     &cd_tunneling_params);
1124                 /* Enable checksum offloading */
1125                 if (ol_flags & I40E_TX_CKSUM_OFFLOAD_MASK)
1126                         i40e_txd_enable_checksum(ol_flags, &td_cmd,
1127                                                  &td_offset, tx_offload);
1128
1129                 if (nb_ctx) {
1130                         /* Setup TX context descriptor if required */
1131                         volatile struct i40e_tx_context_desc *ctx_txd =
1132                                 (volatile struct i40e_tx_context_desc *)\
1133                                                         &txr[tx_id];
1134                         uint16_t cd_l2tag2 = 0;
1135                         uint64_t cd_type_cmd_tso_mss =
1136                                 I40E_TX_DESC_DTYPE_CONTEXT;
1137
1138                         txn = &sw_ring[txe->next_id];
1139                         RTE_MBUF_PREFETCH_TO_FREE(txn->mbuf);
1140                         if (txe->mbuf != NULL) {
1141                                 rte_pktmbuf_free_seg(txe->mbuf);
1142                                 txe->mbuf = NULL;
1143                         }
1144
1145                         /* TSO enabled means no timestamp */
1146                         if (ol_flags & PKT_TX_TCP_SEG)
1147                                 cd_type_cmd_tso_mss |=
1148                                         i40e_set_tso_ctx(tx_pkt, tx_offload);
1149                         else {
1150 #ifdef RTE_LIBRTE_IEEE1588
1151                                 if (ol_flags & PKT_TX_IEEE1588_TMST)
1152                                         cd_type_cmd_tso_mss |=
1153                                                 ((uint64_t)I40E_TX_CTX_DESC_TSYN <<
1154                                                  I40E_TXD_CTX_QW1_CMD_SHIFT);
1155 #endif
1156                         }
1157
1158                         ctx_txd->tunneling_params =
1159                                 rte_cpu_to_le_32(cd_tunneling_params);
1160                         if (ol_flags & PKT_TX_QINQ_PKT) {
1161                                 cd_l2tag2 = tx_pkt->vlan_tci_outer;
1162                                 cd_type_cmd_tso_mss |=
1163                                         ((uint64_t)I40E_TX_CTX_DESC_IL2TAG2 <<
1164                                                 I40E_TXD_CTX_QW1_CMD_SHIFT);
1165                         }
1166                         ctx_txd->l2tag2 = rte_cpu_to_le_16(cd_l2tag2);
1167                         ctx_txd->type_cmd_tso_mss =
1168                                 rte_cpu_to_le_64(cd_type_cmd_tso_mss);
1169
1170                         PMD_TX_LOG(DEBUG, "mbuf: %p, TCD[%u]:\n"
1171                                 "tunneling_params: %#x;\n"
1172                                 "l2tag2: %#hx;\n"
1173                                 "rsvd: %#hx;\n"
1174                                 "type_cmd_tso_mss: %#"PRIx64";\n",
1175                                 tx_pkt, tx_id,
1176                                 ctx_txd->tunneling_params,
1177                                 ctx_txd->l2tag2,
1178                                 ctx_txd->rsvd,
1179                                 ctx_txd->type_cmd_tso_mss);
1180
1181                         txe->last_id = tx_last;
1182                         tx_id = txe->next_id;
1183                         txe = txn;
1184                 }
1185
1186                 m_seg = tx_pkt;
1187                 do {
1188                         txd = &txr[tx_id];
1189                         txn = &sw_ring[txe->next_id];
1190
1191                         if (txe->mbuf)
1192                                 rte_pktmbuf_free_seg(txe->mbuf);
1193                         txe->mbuf = m_seg;
1194
1195                         /* Setup TX Descriptor */
1196                         slen = m_seg->data_len;
1197                         buf_dma_addr = rte_mbuf_data_dma_addr(m_seg);
1198
1199                         PMD_TX_LOG(DEBUG, "mbuf: %p, TDD[%u]:\n"
1200                                 "buf_dma_addr: %#"PRIx64";\n"
1201                                 "td_cmd: %#x;\n"
1202                                 "td_offset: %#x;\n"
1203                                 "td_len: %u;\n"
1204                                 "td_tag: %#x;\n",
1205                                 tx_pkt, tx_id, buf_dma_addr,
1206                                 td_cmd, td_offset, slen, td_tag);
1207
1208                         txd->buffer_addr = rte_cpu_to_le_64(buf_dma_addr);
1209                         txd->cmd_type_offset_bsz = i40e_build_ctob(td_cmd,
1210                                                 td_offset, slen, td_tag);
1211                         txe->last_id = tx_last;
1212                         tx_id = txe->next_id;
1213                         txe = txn;
1214                         m_seg = m_seg->next;
1215                 } while (m_seg != NULL);
1216
1217                 /* The last packet data descriptor needs End Of Packet (EOP) */
1218                 td_cmd |= I40E_TX_DESC_CMD_EOP;
1219                 txq->nb_tx_used = (uint16_t)(txq->nb_tx_used + nb_used);
1220                 txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_used);
1221
1222                 if (txq->nb_tx_used >= txq->tx_rs_thresh) {
1223                         PMD_TX_FREE_LOG(DEBUG,
1224                                         "Setting RS bit on TXD id="
1225                                         "%4u (port=%d queue=%d)",
1226                                         tx_last, txq->port_id, txq->queue_id);
1227
1228                         td_cmd |= I40E_TX_DESC_CMD_RS;
1229
1230                         /* Update txq RS bit counters */
1231                         txq->nb_tx_used = 0;
1232                 }
1233
1234                 txd->cmd_type_offset_bsz |=
1235                         rte_cpu_to_le_64(((uint64_t)td_cmd) <<
1236                                         I40E_TXD_QW1_CMD_SHIFT);
1237         }
1238
1239 end_of_tx:
1240         rte_wmb();
1241
1242         PMD_TX_LOG(DEBUG, "port_id=%u queue_id=%u tx_tail=%u nb_tx=%u",
1243                    (unsigned) txq->port_id, (unsigned) txq->queue_id,
1244                    (unsigned) tx_id, (unsigned) nb_tx);
1245
1246         I40E_PCI_REG_WRITE_RELAXED(txq->qtx_tail, tx_id);
1247         txq->tx_tail = tx_id;
1248
1249         return nb_tx;
1250 }
1251
1252 static inline int __attribute__((always_inline))
1253 i40e_tx_free_bufs(struct i40e_tx_queue *txq)
1254 {
1255         struct i40e_tx_entry *txep;
1256         uint16_t i;
1257
1258         if ((txq->tx_ring[txq->tx_next_dd].cmd_type_offset_bsz &
1259                         rte_cpu_to_le_64(I40E_TXD_QW1_DTYPE_MASK)) !=
1260                         rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DESC_DONE))
1261                 return 0;
1262
1263         txep = &(txq->sw_ring[txq->tx_next_dd - (txq->tx_rs_thresh - 1)]);
1264
1265         for (i = 0; i < txq->tx_rs_thresh; i++)
1266                 rte_prefetch0((txep + i)->mbuf);
1267
1268         if (txq->txq_flags & (uint32_t)ETH_TXQ_FLAGS_NOREFCOUNT) {
1269                 for (i = 0; i < txq->tx_rs_thresh; ++i, ++txep) {
1270                         rte_mempool_put(txep->mbuf->pool, txep->mbuf);
1271                         txep->mbuf = NULL;
1272                 }
1273         } else {
1274                 for (i = 0; i < txq->tx_rs_thresh; ++i, ++txep) {
1275                         rte_pktmbuf_free_seg(txep->mbuf);
1276                         txep->mbuf = NULL;
1277                 }
1278         }
1279
1280         txq->nb_tx_free = (uint16_t)(txq->nb_tx_free + txq->tx_rs_thresh);
1281         txq->tx_next_dd = (uint16_t)(txq->tx_next_dd + txq->tx_rs_thresh);
1282         if (txq->tx_next_dd >= txq->nb_tx_desc)
1283                 txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
1284
1285         return txq->tx_rs_thresh;
1286 }
1287
1288 /* Populate 4 descriptors with data from 4 mbufs */
1289 static inline void
1290 tx4(volatile struct i40e_tx_desc *txdp, struct rte_mbuf **pkts)
1291 {
1292         uint64_t dma_addr;
1293         uint32_t i;
1294
1295         for (i = 0; i < 4; i++, txdp++, pkts++) {
1296                 dma_addr = rte_mbuf_data_dma_addr(*pkts);
1297                 txdp->buffer_addr = rte_cpu_to_le_64(dma_addr);
1298                 txdp->cmd_type_offset_bsz =
1299                         i40e_build_ctob((uint32_t)I40E_TD_CMD, 0,
1300                                         (*pkts)->data_len, 0);
1301         }
1302 }
1303
1304 /* Populate 1 descriptor with data from 1 mbuf */
1305 static inline void
1306 tx1(volatile struct i40e_tx_desc *txdp, struct rte_mbuf **pkts)
1307 {
1308         uint64_t dma_addr;
1309
1310         dma_addr = rte_mbuf_data_dma_addr(*pkts);
1311         txdp->buffer_addr = rte_cpu_to_le_64(dma_addr);
1312         txdp->cmd_type_offset_bsz =
1313                 i40e_build_ctob((uint32_t)I40E_TD_CMD, 0,
1314                                 (*pkts)->data_len, 0);
1315 }
1316
1317 /* Fill hardware descriptor ring with mbuf data */
1318 static inline void
1319 i40e_tx_fill_hw_ring(struct i40e_tx_queue *txq,
1320                      struct rte_mbuf **pkts,
1321                      uint16_t nb_pkts)
1322 {
1323         volatile struct i40e_tx_desc *txdp = &(txq->tx_ring[txq->tx_tail]);
1324         struct i40e_tx_entry *txep = &(txq->sw_ring[txq->tx_tail]);
1325         const int N_PER_LOOP = 4;
1326         const int N_PER_LOOP_MASK = N_PER_LOOP - 1;
1327         int mainpart, leftover;
1328         int i, j;
1329
1330         mainpart = (nb_pkts & ((uint32_t) ~N_PER_LOOP_MASK));
1331         leftover = (nb_pkts & ((uint32_t)  N_PER_LOOP_MASK));
1332         for (i = 0; i < mainpart; i += N_PER_LOOP) {
1333                 for (j = 0; j < N_PER_LOOP; ++j) {
1334                         (txep + i + j)->mbuf = *(pkts + i + j);
1335                 }
1336                 tx4(txdp + i, pkts + i);
1337         }
1338         if (unlikely(leftover > 0)) {
1339                 for (i = 0; i < leftover; ++i) {
1340                         (txep + mainpart + i)->mbuf = *(pkts + mainpart + i);
1341                         tx1(txdp + mainpart + i, pkts + mainpart + i);
1342                 }
1343         }
1344 }
1345
1346 static inline uint16_t
1347 tx_xmit_pkts(struct i40e_tx_queue *txq,
1348              struct rte_mbuf **tx_pkts,
1349              uint16_t nb_pkts)
1350 {
1351         volatile struct i40e_tx_desc *txr = txq->tx_ring;
1352         uint16_t n = 0;
1353
1354         /**
1355          * Begin scanning the H/W ring for done descriptors when the number
1356          * of available descriptors drops below tx_free_thresh. For each done
1357          * descriptor, free the associated buffer.
1358          */
1359         if (txq->nb_tx_free < txq->tx_free_thresh)
1360                 i40e_tx_free_bufs(txq);
1361
1362         /* Use available descriptor only */
1363         nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts);
1364         if (unlikely(!nb_pkts))
1365                 return 0;
1366
1367         txq->nb_tx_free = (uint16_t)(txq->nb_tx_free - nb_pkts);
1368         if ((txq->tx_tail + nb_pkts) > txq->nb_tx_desc) {
1369                 n = (uint16_t)(txq->nb_tx_desc - txq->tx_tail);
1370                 i40e_tx_fill_hw_ring(txq, tx_pkts, n);
1371                 txr[txq->tx_next_rs].cmd_type_offset_bsz |=
1372                         rte_cpu_to_le_64(((uint64_t)I40E_TX_DESC_CMD_RS) <<
1373                                                 I40E_TXD_QW1_CMD_SHIFT);
1374                 txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
1375                 txq->tx_tail = 0;
1376         }
1377
1378         /* Fill hardware descriptor ring with mbuf data */
1379         i40e_tx_fill_hw_ring(txq, tx_pkts + n, (uint16_t)(nb_pkts - n));
1380         txq->tx_tail = (uint16_t)(txq->tx_tail + (nb_pkts - n));
1381
1382         /* Determin if RS bit needs to be set */
1383         if (txq->tx_tail > txq->tx_next_rs) {
1384                 txr[txq->tx_next_rs].cmd_type_offset_bsz |=
1385                         rte_cpu_to_le_64(((uint64_t)I40E_TX_DESC_CMD_RS) <<
1386                                                 I40E_TXD_QW1_CMD_SHIFT);
1387                 txq->tx_next_rs =
1388                         (uint16_t)(txq->tx_next_rs + txq->tx_rs_thresh);
1389                 if (txq->tx_next_rs >= txq->nb_tx_desc)
1390                         txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
1391         }
1392
1393         if (txq->tx_tail >= txq->nb_tx_desc)
1394                 txq->tx_tail = 0;
1395
1396         /* Update the tx tail register */
1397         rte_wmb();
1398         I40E_PCI_REG_WRITE_RELAXED(txq->qtx_tail, txq->tx_tail);
1399
1400         return nb_pkts;
1401 }
1402
1403 static uint16_t
1404 i40e_xmit_pkts_simple(void *tx_queue,
1405                       struct rte_mbuf **tx_pkts,
1406                       uint16_t nb_pkts)
1407 {
1408         uint16_t nb_tx = 0;
1409
1410         if (likely(nb_pkts <= I40E_TX_MAX_BURST))
1411                 return tx_xmit_pkts((struct i40e_tx_queue *)tx_queue,
1412                                                 tx_pkts, nb_pkts);
1413
1414         while (nb_pkts) {
1415                 uint16_t ret, num = (uint16_t)RTE_MIN(nb_pkts,
1416                                                 I40E_TX_MAX_BURST);
1417
1418                 ret = tx_xmit_pkts((struct i40e_tx_queue *)tx_queue,
1419                                                 &tx_pkts[nb_tx], num);
1420                 nb_tx = (uint16_t)(nb_tx + ret);
1421                 nb_pkts = (uint16_t)(nb_pkts - ret);
1422                 if (ret < num)
1423                         break;
1424         }
1425
1426         return nb_tx;
1427 }
1428
1429 /*********************************************************************
1430  *
1431  *  TX prep functions
1432  *
1433  **********************************************************************/
1434 uint16_t
1435 i40e_prep_pkts(__rte_unused void *tx_queue, struct rte_mbuf **tx_pkts,
1436                 uint16_t nb_pkts)
1437 {
1438         int i, ret;
1439         uint64_t ol_flags;
1440         struct rte_mbuf *m;
1441
1442         for (i = 0; i < nb_pkts; i++) {
1443                 m = tx_pkts[i];
1444                 ol_flags = m->ol_flags;
1445
1446                 /**
1447                  * m->nb_segs is uint8_t, so nb_segs is always less than
1448                  * I40E_TX_MAX_SEG.
1449                  * We check only a condition for nb_segs > I40E_TX_MAX_MTU_SEG.
1450                  */
1451                 if (!(ol_flags & PKT_TX_TCP_SEG)) {
1452                         if (m->nb_segs > I40E_TX_MAX_MTU_SEG) {
1453                                 rte_errno = -EINVAL;
1454                                 return i;
1455                         }
1456                 } else if ((m->tso_segsz < I40E_MIN_TSO_MSS) ||
1457                                 (m->tso_segsz > I40E_MAX_TSO_MSS)) {
1458                         /* MSS outside the range (256B - 9674B) are considered
1459                          * malicious
1460                          */
1461                         rte_errno = -EINVAL;
1462                         return i;
1463                 }
1464
1465                 if (ol_flags & I40E_TX_OFFLOAD_NOTSUP_MASK) {
1466                         rte_errno = -ENOTSUP;
1467                         return i;
1468                 }
1469
1470 #ifdef RTE_LIBRTE_ETHDEV_DEBUG
1471                 ret = rte_validate_tx_offload(m);
1472                 if (ret != 0) {
1473                         rte_errno = ret;
1474                         return i;
1475                 }
1476 #endif
1477                 ret = rte_net_intel_cksum_prepare(m);
1478                 if (ret != 0) {
1479                         rte_errno = ret;
1480                         return i;
1481                 }
1482         }
1483         return i;
1484 }
1485
1486 /*
1487  * Find the VSI the queue belongs to. 'queue_idx' is the queue index
1488  * application used, which assume having sequential ones. But from driver's
1489  * perspective, it's different. For example, q0 belongs to FDIR VSI, q1-q64
1490  * to MAIN VSI, , q65-96 to SRIOV VSIs, q97-128 to VMDQ VSIs. For application
1491  * running on host, q1-64 and q97-128 can be used, total 96 queues. They can
1492  * use queue_idx from 0 to 95 to access queues, while real queue would be
1493  * different. This function will do a queue mapping to find VSI the queue
1494  * belongs to.
1495  */
1496 static struct i40e_vsi*
1497 i40e_pf_get_vsi_by_qindex(struct i40e_pf *pf, uint16_t queue_idx)
1498 {
1499         /* the queue in MAIN VSI range */
1500         if (queue_idx < pf->main_vsi->nb_qps)
1501                 return pf->main_vsi;
1502
1503         queue_idx -= pf->main_vsi->nb_qps;
1504
1505         /* queue_idx is greater than VMDQ VSIs range */
1506         if (queue_idx > pf->nb_cfg_vmdq_vsi * pf->vmdq_nb_qps - 1) {
1507                 PMD_INIT_LOG(ERR, "queue_idx out of range. VMDQ configured?");
1508                 return NULL;
1509         }
1510
1511         return pf->vmdq[queue_idx / pf->vmdq_nb_qps].vsi;
1512 }
1513
1514 static uint16_t
1515 i40e_get_queue_offset_by_qindex(struct i40e_pf *pf, uint16_t queue_idx)
1516 {
1517         /* the queue in MAIN VSI range */
1518         if (queue_idx < pf->main_vsi->nb_qps)
1519                 return queue_idx;
1520
1521         /* It's VMDQ queues */
1522         queue_idx -= pf->main_vsi->nb_qps;
1523
1524         if (pf->nb_cfg_vmdq_vsi)
1525                 return queue_idx % pf->vmdq_nb_qps;
1526         else {
1527                 PMD_INIT_LOG(ERR, "Fail to get queue offset");
1528                 return (uint16_t)(-1);
1529         }
1530 }
1531
1532 int
1533 i40e_dev_rx_queue_start(struct rte_eth_dev *dev, uint16_t rx_queue_id)
1534 {
1535         struct i40e_rx_queue *rxq;
1536         int err = -1;
1537         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1538
1539         PMD_INIT_FUNC_TRACE();
1540
1541         if (rx_queue_id < dev->data->nb_rx_queues) {
1542                 rxq = dev->data->rx_queues[rx_queue_id];
1543
1544                 err = i40e_alloc_rx_queue_mbufs(rxq);
1545                 if (err) {
1546                         PMD_DRV_LOG(ERR, "Failed to allocate RX queue mbuf");
1547                         return err;
1548                 }
1549
1550                 rte_wmb();
1551
1552                 /* Init the RX tail regieter. */
1553                 I40E_PCI_REG_WRITE(rxq->qrx_tail, rxq->nb_rx_desc - 1);
1554
1555                 err = i40e_switch_rx_queue(hw, rxq->reg_idx, TRUE);
1556
1557                 if (err) {
1558                         PMD_DRV_LOG(ERR, "Failed to switch RX queue %u on",
1559                                     rx_queue_id);
1560
1561                         i40e_rx_queue_release_mbufs(rxq);
1562                         i40e_reset_rx_queue(rxq);
1563                 } else
1564                         dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
1565         }
1566
1567         return err;
1568 }
1569
1570 int
1571 i40e_dev_rx_queue_stop(struct rte_eth_dev *dev, uint16_t rx_queue_id)
1572 {
1573         struct i40e_rx_queue *rxq;
1574         int err;
1575         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1576
1577         if (rx_queue_id < dev->data->nb_rx_queues) {
1578                 rxq = dev->data->rx_queues[rx_queue_id];
1579
1580                 /*
1581                 * rx_queue_id is queue id aplication refers to, while
1582                 * rxq->reg_idx is the real queue index.
1583                 */
1584                 err = i40e_switch_rx_queue(hw, rxq->reg_idx, FALSE);
1585
1586                 if (err) {
1587                         PMD_DRV_LOG(ERR, "Failed to switch RX queue %u off",
1588                                     rx_queue_id);
1589                         return err;
1590                 }
1591                 i40e_rx_queue_release_mbufs(rxq);
1592                 i40e_reset_rx_queue(rxq);
1593                 dev->data->rx_queue_state[rx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
1594         }
1595
1596         return 0;
1597 }
1598
1599 int
1600 i40e_dev_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id)
1601 {
1602         int err = -1;
1603         struct i40e_tx_queue *txq;
1604         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1605
1606         PMD_INIT_FUNC_TRACE();
1607
1608         if (tx_queue_id < dev->data->nb_tx_queues) {
1609                 txq = dev->data->tx_queues[tx_queue_id];
1610
1611                 /*
1612                 * tx_queue_id is queue id aplication refers to, while
1613                 * rxq->reg_idx is the real queue index.
1614                 */
1615                 err = i40e_switch_tx_queue(hw, txq->reg_idx, TRUE);
1616                 if (err)
1617                         PMD_DRV_LOG(ERR, "Failed to switch TX queue %u on",
1618                                     tx_queue_id);
1619                 else
1620                         dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STARTED;
1621         }
1622
1623         return err;
1624 }
1625
1626 int
1627 i40e_dev_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id)
1628 {
1629         struct i40e_tx_queue *txq;
1630         int err;
1631         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1632
1633         if (tx_queue_id < dev->data->nb_tx_queues) {
1634                 txq = dev->data->tx_queues[tx_queue_id];
1635
1636                 /*
1637                 * tx_queue_id is queue id aplication refers to, while
1638                 * txq->reg_idx is the real queue index.
1639                 */
1640                 err = i40e_switch_tx_queue(hw, txq->reg_idx, FALSE);
1641
1642                 if (err) {
1643                         PMD_DRV_LOG(ERR, "Failed to switch TX queue %u of",
1644                                     tx_queue_id);
1645                         return err;
1646                 }
1647
1648                 i40e_tx_queue_release_mbufs(txq);
1649                 i40e_reset_tx_queue(txq);
1650                 dev->data->tx_queue_state[tx_queue_id] = RTE_ETH_QUEUE_STATE_STOPPED;
1651         }
1652
1653         return 0;
1654 }
1655
1656 const uint32_t *
1657 i40e_dev_supported_ptypes_get(struct rte_eth_dev *dev)
1658 {
1659         static const uint32_t ptypes[] = {
1660                 /* refers to i40e_rxd_pkt_type_mapping() */
1661                 RTE_PTYPE_L2_ETHER,
1662                 RTE_PTYPE_L2_ETHER_TIMESYNC,
1663                 RTE_PTYPE_L2_ETHER_LLDP,
1664                 RTE_PTYPE_L2_ETHER_ARP,
1665                 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
1666                 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
1667                 RTE_PTYPE_L4_FRAG,
1668                 RTE_PTYPE_L4_ICMP,
1669                 RTE_PTYPE_L4_NONFRAG,
1670                 RTE_PTYPE_L4_SCTP,
1671                 RTE_PTYPE_L4_TCP,
1672                 RTE_PTYPE_L4_UDP,
1673                 RTE_PTYPE_TUNNEL_GRENAT,
1674                 RTE_PTYPE_TUNNEL_IP,
1675                 RTE_PTYPE_INNER_L2_ETHER,
1676                 RTE_PTYPE_INNER_L2_ETHER_VLAN,
1677                 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
1678                 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
1679                 RTE_PTYPE_INNER_L4_FRAG,
1680                 RTE_PTYPE_INNER_L4_ICMP,
1681                 RTE_PTYPE_INNER_L4_NONFRAG,
1682                 RTE_PTYPE_INNER_L4_SCTP,
1683                 RTE_PTYPE_INNER_L4_TCP,
1684                 RTE_PTYPE_INNER_L4_UDP,
1685                 RTE_PTYPE_UNKNOWN
1686         };
1687
1688         if (dev->rx_pkt_burst == i40e_recv_pkts ||
1689 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
1690             dev->rx_pkt_burst == i40e_recv_pkts_bulk_alloc ||
1691 #endif
1692             dev->rx_pkt_burst == i40e_recv_scattered_pkts ||
1693             dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec ||
1694             dev->rx_pkt_burst == i40e_recv_pkts_vec)
1695                 return ptypes;
1696         return NULL;
1697 }
1698
1699 int
1700 i40e_dev_rx_queue_setup(struct rte_eth_dev *dev,
1701                         uint16_t queue_idx,
1702                         uint16_t nb_desc,
1703                         unsigned int socket_id,
1704                         const struct rte_eth_rxconf *rx_conf,
1705                         struct rte_mempool *mp)
1706 {
1707         struct i40e_vsi *vsi;
1708         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1709         struct i40e_pf *pf = I40E_DEV_PRIVATE_TO_PF(dev->data->dev_private);
1710         struct i40e_adapter *ad =
1711                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
1712         struct i40e_rx_queue *rxq;
1713         const struct rte_memzone *rz;
1714         uint32_t ring_size;
1715         uint16_t len, i;
1716         uint16_t base, bsf, tc_mapping;
1717         int use_def_burst_func = 1;
1718
1719         if (hw->mac.type == I40E_MAC_VF || hw->mac.type == I40E_MAC_X722_VF) {
1720                 struct i40e_vf *vf =
1721                         I40EVF_DEV_PRIVATE_TO_VF(dev->data->dev_private);
1722                 vsi = &vf->vsi;
1723         } else
1724                 vsi = i40e_pf_get_vsi_by_qindex(pf, queue_idx);
1725
1726         if (vsi == NULL) {
1727                 PMD_DRV_LOG(ERR, "VSI not available or queue "
1728                             "index exceeds the maximum");
1729                 return I40E_ERR_PARAM;
1730         }
1731         if (nb_desc % I40E_ALIGN_RING_DESC != 0 ||
1732                         (nb_desc > I40E_MAX_RING_DESC) ||
1733                         (nb_desc < I40E_MIN_RING_DESC)) {
1734                 PMD_DRV_LOG(ERR, "Number (%u) of receive descriptors is "
1735                             "invalid", nb_desc);
1736                 return I40E_ERR_PARAM;
1737         }
1738
1739         /* Free memory if needed */
1740         if (dev->data->rx_queues[queue_idx]) {
1741                 i40e_dev_rx_queue_release(dev->data->rx_queues[queue_idx]);
1742                 dev->data->rx_queues[queue_idx] = NULL;
1743         }
1744
1745         /* Allocate the rx queue data structure */
1746         rxq = rte_zmalloc_socket("i40e rx queue",
1747                                  sizeof(struct i40e_rx_queue),
1748                                  RTE_CACHE_LINE_SIZE,
1749                                  socket_id);
1750         if (!rxq) {
1751                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
1752                             "rx queue data structure");
1753                 return -ENOMEM;
1754         }
1755         rxq->mp = mp;
1756         rxq->nb_rx_desc = nb_desc;
1757         rxq->rx_free_thresh = rx_conf->rx_free_thresh;
1758         rxq->queue_id = queue_idx;
1759         if (hw->mac.type == I40E_MAC_VF || hw->mac.type == I40E_MAC_X722_VF)
1760                 rxq->reg_idx = queue_idx;
1761         else /* PF device */
1762                 rxq->reg_idx = vsi->base_queue +
1763                         i40e_get_queue_offset_by_qindex(pf, queue_idx);
1764
1765         rxq->port_id = dev->data->port_id;
1766         rxq->crc_len = (uint8_t) ((dev->data->dev_conf.rxmode.hw_strip_crc) ?
1767                                                         0 : ETHER_CRC_LEN);
1768         rxq->drop_en = rx_conf->rx_drop_en;
1769         rxq->vsi = vsi;
1770         rxq->rx_deferred_start = rx_conf->rx_deferred_start;
1771
1772         /* Allocate the maximun number of RX ring hardware descriptor. */
1773         len = I40E_MAX_RING_DESC;
1774
1775 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
1776         /**
1777          * Allocating a little more memory because vectorized/bulk_alloc Rx
1778          * functions doesn't check boundaries each time.
1779          */
1780         len += RTE_PMD_I40E_RX_MAX_BURST;
1781 #endif
1782
1783         ring_size = RTE_ALIGN(len * sizeof(union i40e_rx_desc),
1784                               I40E_DMA_MEM_ALIGN);
1785
1786         rz = rte_eth_dma_zone_reserve(dev, "rx_ring", queue_idx,
1787                               ring_size, I40E_RING_BASE_ALIGN, socket_id);
1788         if (!rz) {
1789                 i40e_dev_rx_queue_release(rxq);
1790                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for RX");
1791                 return -ENOMEM;
1792         }
1793
1794         /* Zero all the descriptors in the ring. */
1795         memset(rz->addr, 0, ring_size);
1796
1797         rxq->rx_ring_phys_addr = rte_mem_phy2mch(rz->memseg_id, rz->phys_addr);
1798         rxq->rx_ring = (union i40e_rx_desc *)rz->addr;
1799
1800 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
1801         len = (uint16_t)(nb_desc + RTE_PMD_I40E_RX_MAX_BURST);
1802 #else
1803         len = nb_desc;
1804 #endif
1805
1806         /* Allocate the software ring. */
1807         rxq->sw_ring =
1808                 rte_zmalloc_socket("i40e rx sw ring",
1809                                    sizeof(struct i40e_rx_entry) * len,
1810                                    RTE_CACHE_LINE_SIZE,
1811                                    socket_id);
1812         if (!rxq->sw_ring) {
1813                 i40e_dev_rx_queue_release(rxq);
1814                 PMD_DRV_LOG(ERR, "Failed to allocate memory for SW ring");
1815                 return -ENOMEM;
1816         }
1817
1818         i40e_reset_rx_queue(rxq);
1819         rxq->q_set = TRUE;
1820         dev->data->rx_queues[queue_idx] = rxq;
1821
1822         use_def_burst_func = check_rx_burst_bulk_alloc_preconditions(rxq);
1823
1824         if (!use_def_burst_func) {
1825 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
1826                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions are "
1827                              "satisfied. Rx Burst Bulk Alloc function will be "
1828                              "used on port=%d, queue=%d.",
1829                              rxq->port_id, rxq->queue_id);
1830 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
1831         } else {
1832                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions are "
1833                              "not satisfied, Scattered Rx is requested, "
1834                              "or RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC is "
1835                              "not enabled on port=%d, queue=%d.",
1836                              rxq->port_id, rxq->queue_id);
1837                 ad->rx_bulk_alloc_allowed = false;
1838         }
1839
1840         for (i = 0; i < I40E_MAX_TRAFFIC_CLASS; i++) {
1841                 if (!(vsi->enabled_tc & (1 << i)))
1842                         continue;
1843                 tc_mapping = rte_le_to_cpu_16(vsi->info.tc_mapping[i]);
1844                 base = (tc_mapping & I40E_AQ_VSI_TC_QUE_OFFSET_MASK) >>
1845                         I40E_AQ_VSI_TC_QUE_OFFSET_SHIFT;
1846                 bsf = (tc_mapping & I40E_AQ_VSI_TC_QUE_NUMBER_MASK) >>
1847                         I40E_AQ_VSI_TC_QUE_NUMBER_SHIFT;
1848
1849                 if (queue_idx >= base && queue_idx < (base + BIT(bsf)))
1850                         rxq->dcb_tc = i;
1851         }
1852
1853         return 0;
1854 }
1855
1856 void
1857 i40e_dev_rx_queue_release(void *rxq)
1858 {
1859         struct i40e_rx_queue *q = (struct i40e_rx_queue *)rxq;
1860
1861         if (!q) {
1862                 PMD_DRV_LOG(DEBUG, "Pointer to rxq is NULL");
1863                 return;
1864         }
1865
1866         i40e_rx_queue_release_mbufs(q);
1867         rte_free(q->sw_ring);
1868         rte_free(q);
1869 }
1870
1871 uint32_t
1872 i40e_dev_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id)
1873 {
1874 #define I40E_RXQ_SCAN_INTERVAL 4
1875         volatile union i40e_rx_desc *rxdp;
1876         struct i40e_rx_queue *rxq;
1877         uint16_t desc = 0;
1878
1879         rxq = dev->data->rx_queues[rx_queue_id];
1880         rxdp = &(rxq->rx_ring[rxq->rx_tail]);
1881         while ((desc < rxq->nb_rx_desc) &&
1882                 ((rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len) &
1883                 I40E_RXD_QW1_STATUS_MASK) >> I40E_RXD_QW1_STATUS_SHIFT) &
1884                                 (1 << I40E_RX_DESC_STATUS_DD_SHIFT)) {
1885                 /**
1886                  * Check the DD bit of a rx descriptor of each 4 in a group,
1887                  * to avoid checking too frequently and downgrading performance
1888                  * too much.
1889                  */
1890                 desc += I40E_RXQ_SCAN_INTERVAL;
1891                 rxdp += I40E_RXQ_SCAN_INTERVAL;
1892                 if (rxq->rx_tail + desc >= rxq->nb_rx_desc)
1893                         rxdp = &(rxq->rx_ring[rxq->rx_tail +
1894                                         desc - rxq->nb_rx_desc]);
1895         }
1896
1897         return desc;
1898 }
1899
1900 int
1901 i40e_dev_rx_descriptor_done(void *rx_queue, uint16_t offset)
1902 {
1903         volatile union i40e_rx_desc *rxdp;
1904         struct i40e_rx_queue *rxq = rx_queue;
1905         uint16_t desc;
1906         int ret;
1907
1908         if (unlikely(offset >= rxq->nb_rx_desc)) {
1909                 PMD_DRV_LOG(ERR, "Invalid RX queue id %u", offset);
1910                 return 0;
1911         }
1912
1913         desc = rxq->rx_tail + offset;
1914         if (desc >= rxq->nb_rx_desc)
1915                 desc -= rxq->nb_rx_desc;
1916
1917         rxdp = &(rxq->rx_ring[desc]);
1918
1919         ret = !!(((rte_le_to_cpu_64(rxdp->wb.qword1.status_error_len) &
1920                 I40E_RXD_QW1_STATUS_MASK) >> I40E_RXD_QW1_STATUS_SHIFT) &
1921                                 (1 << I40E_RX_DESC_STATUS_DD_SHIFT));
1922
1923         return ret;
1924 }
1925
1926 int
1927 i40e_dev_rx_descriptor_status(void *rx_queue, uint16_t offset)
1928 {
1929         struct i40e_rx_queue *rxq = rx_queue;
1930         volatile uint64_t *status;
1931         uint64_t mask;
1932         uint32_t desc;
1933
1934         if (unlikely(offset >= rxq->nb_rx_desc))
1935                 return -EINVAL;
1936
1937         if (offset >= rxq->nb_rx_desc - rxq->nb_rx_hold)
1938                 return RTE_ETH_RX_DESC_UNAVAIL;
1939
1940         desc = rxq->rx_tail + offset;
1941         if (desc >= rxq->nb_rx_desc)
1942                 desc -= rxq->nb_rx_desc;
1943
1944         status = &rxq->rx_ring[desc].wb.qword1.status_error_len;
1945         mask = rte_le_to_cpu_64((1ULL << I40E_RX_DESC_STATUS_DD_SHIFT)
1946                 << I40E_RXD_QW1_STATUS_SHIFT);
1947         if (*status & mask)
1948                 return RTE_ETH_RX_DESC_DONE;
1949
1950         return RTE_ETH_RX_DESC_AVAIL;
1951 }
1952
1953 int
1954 i40e_dev_tx_descriptor_status(void *tx_queue, uint16_t offset)
1955 {
1956         struct i40e_tx_queue *txq = tx_queue;
1957         volatile uint64_t *status;
1958         uint64_t mask, expect;
1959         uint32_t desc;
1960
1961         if (unlikely(offset >= txq->nb_tx_desc))
1962                 return -EINVAL;
1963
1964         desc = txq->tx_tail + offset;
1965         /* go to next desc that has the RS bit */
1966         desc = ((desc + txq->tx_rs_thresh - 1) / txq->tx_rs_thresh) *
1967                 txq->tx_rs_thresh;
1968         if (desc >= txq->nb_tx_desc) {
1969                 desc -= txq->nb_tx_desc;
1970                 if (desc >= txq->nb_tx_desc)
1971                         desc -= txq->nb_tx_desc;
1972         }
1973
1974         status = &txq->tx_ring[desc].cmd_type_offset_bsz;
1975         mask = rte_le_to_cpu_64(I40E_TXD_QW1_DTYPE_MASK);
1976         expect = rte_cpu_to_le_64(
1977                 I40E_TX_DESC_DTYPE_DESC_DONE << I40E_TXD_QW1_DTYPE_SHIFT);
1978         if ((*status & mask) == expect)
1979                 return RTE_ETH_TX_DESC_DONE;
1980
1981         return RTE_ETH_TX_DESC_FULL;
1982 }
1983
1984 int
1985 i40e_dev_tx_queue_setup(struct rte_eth_dev *dev,
1986                         uint16_t queue_idx,
1987                         uint16_t nb_desc,
1988                         unsigned int socket_id,
1989                         const struct rte_eth_txconf *tx_conf)
1990 {
1991         struct i40e_vsi *vsi;
1992         struct i40e_hw *hw = I40E_DEV_PRIVATE_TO_HW(dev->data->dev_private);
1993         struct i40e_pf *pf = I40E_DEV_PRIVATE_TO_PF(dev->data->dev_private);
1994         struct i40e_tx_queue *txq;
1995         const struct rte_memzone *tz;
1996         uint32_t ring_size;
1997         uint16_t tx_rs_thresh, tx_free_thresh;
1998         uint16_t i, base, bsf, tc_mapping;
1999
2000         if (hw->mac.type == I40E_MAC_VF || hw->mac.type == I40E_MAC_X722_VF) {
2001                 struct i40e_vf *vf =
2002                         I40EVF_DEV_PRIVATE_TO_VF(dev->data->dev_private);
2003                 vsi = &vf->vsi;
2004         } else
2005                 vsi = i40e_pf_get_vsi_by_qindex(pf, queue_idx);
2006
2007         if (vsi == NULL) {
2008                 PMD_DRV_LOG(ERR, "VSI is NULL, or queue index (%u) "
2009                             "exceeds the maximum", queue_idx);
2010                 return I40E_ERR_PARAM;
2011         }
2012
2013         if (nb_desc % I40E_ALIGN_RING_DESC != 0 ||
2014                         (nb_desc > I40E_MAX_RING_DESC) ||
2015                         (nb_desc < I40E_MIN_RING_DESC)) {
2016                 PMD_DRV_LOG(ERR, "Number (%u) of transmit descriptors is "
2017                             "invalid", nb_desc);
2018                 return I40E_ERR_PARAM;
2019         }
2020
2021         /**
2022          * The following two parameters control the setting of the RS bit on
2023          * transmit descriptors. TX descriptors will have their RS bit set
2024          * after txq->tx_rs_thresh descriptors have been used. The TX
2025          * descriptor ring will be cleaned after txq->tx_free_thresh
2026          * descriptors are used or if the number of descriptors required to
2027          * transmit a packet is greater than the number of free TX descriptors.
2028          *
2029          * The following constraints must be satisfied:
2030          *  - tx_rs_thresh must be greater than 0.
2031          *  - tx_rs_thresh must be less than the size of the ring minus 2.
2032          *  - tx_rs_thresh must be less than or equal to tx_free_thresh.
2033          *  - tx_rs_thresh must be a divisor of the ring size.
2034          *  - tx_free_thresh must be greater than 0.
2035          *  - tx_free_thresh must be less than the size of the ring minus 3.
2036          *
2037          * One descriptor in the TX ring is used as a sentinel to avoid a H/W
2038          * race condition, hence the maximum threshold constraints. When set
2039          * to zero use default values.
2040          */
2041         tx_rs_thresh = (uint16_t)((tx_conf->tx_rs_thresh) ?
2042                 tx_conf->tx_rs_thresh : DEFAULT_TX_RS_THRESH);
2043         tx_free_thresh = (uint16_t)((tx_conf->tx_free_thresh) ?
2044                 tx_conf->tx_free_thresh : DEFAULT_TX_FREE_THRESH);
2045         if (tx_rs_thresh >= (nb_desc - 2)) {
2046                 PMD_INIT_LOG(ERR, "tx_rs_thresh must be less than the "
2047                              "number of TX descriptors minus 2. "
2048                              "(tx_rs_thresh=%u port=%d queue=%d)",
2049                              (unsigned int)tx_rs_thresh,
2050                              (int)dev->data->port_id,
2051                              (int)queue_idx);
2052                 return I40E_ERR_PARAM;
2053         }
2054         if (tx_free_thresh >= (nb_desc - 3)) {
2055                 PMD_INIT_LOG(ERR, "tx_free_thresh must be less than the "
2056                              "number of TX descriptors minus 3. "
2057                              "(tx_free_thresh=%u port=%d queue=%d)",
2058                              (unsigned int)tx_free_thresh,
2059                              (int)dev->data->port_id,
2060                              (int)queue_idx);
2061                 return I40E_ERR_PARAM;
2062         }
2063         if (tx_rs_thresh > tx_free_thresh) {
2064                 PMD_INIT_LOG(ERR, "tx_rs_thresh must be less than or "
2065                              "equal to tx_free_thresh. (tx_free_thresh=%u"
2066                              " tx_rs_thresh=%u port=%d queue=%d)",
2067                              (unsigned int)tx_free_thresh,
2068                              (unsigned int)tx_rs_thresh,
2069                              (int)dev->data->port_id,
2070                              (int)queue_idx);
2071                 return I40E_ERR_PARAM;
2072         }
2073         if ((nb_desc % tx_rs_thresh) != 0) {
2074                 PMD_INIT_LOG(ERR, "tx_rs_thresh must be a divisor of the "
2075                              "number of TX descriptors. (tx_rs_thresh=%u"
2076                              " port=%d queue=%d)",
2077                              (unsigned int)tx_rs_thresh,
2078                              (int)dev->data->port_id,
2079                              (int)queue_idx);
2080                 return I40E_ERR_PARAM;
2081         }
2082         if ((tx_rs_thresh > 1) && (tx_conf->tx_thresh.wthresh != 0)) {
2083                 PMD_INIT_LOG(ERR, "TX WTHRESH must be set to 0 if "
2084                              "tx_rs_thresh is greater than 1. "
2085                              "(tx_rs_thresh=%u port=%d queue=%d)",
2086                              (unsigned int)tx_rs_thresh,
2087                              (int)dev->data->port_id,
2088                              (int)queue_idx);
2089                 return I40E_ERR_PARAM;
2090         }
2091
2092         /* Free memory if needed. */
2093         if (dev->data->tx_queues[queue_idx]) {
2094                 i40e_dev_tx_queue_release(dev->data->tx_queues[queue_idx]);
2095                 dev->data->tx_queues[queue_idx] = NULL;
2096         }
2097
2098         /* Allocate the TX queue data structure. */
2099         txq = rte_zmalloc_socket("i40e tx queue",
2100                                   sizeof(struct i40e_tx_queue),
2101                                   RTE_CACHE_LINE_SIZE,
2102                                   socket_id);
2103         if (!txq) {
2104                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
2105                             "tx queue structure");
2106                 return -ENOMEM;
2107         }
2108
2109         /* Allocate TX hardware ring descriptors. */
2110         ring_size = sizeof(struct i40e_tx_desc) * I40E_MAX_RING_DESC;
2111         ring_size = RTE_ALIGN(ring_size, I40E_DMA_MEM_ALIGN);
2112         tz = rte_eth_dma_zone_reserve(dev, "tx_ring", queue_idx,
2113                               ring_size, I40E_RING_BASE_ALIGN, socket_id);
2114         if (!tz) {
2115                 i40e_dev_tx_queue_release(txq);
2116                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for TX");
2117                 return -ENOMEM;
2118         }
2119
2120         txq->nb_tx_desc = nb_desc;
2121         txq->tx_rs_thresh = tx_rs_thresh;
2122         txq->tx_free_thresh = tx_free_thresh;
2123         txq->pthresh = tx_conf->tx_thresh.pthresh;
2124         txq->hthresh = tx_conf->tx_thresh.hthresh;
2125         txq->wthresh = tx_conf->tx_thresh.wthresh;
2126         txq->queue_id = queue_idx;
2127         if (hw->mac.type == I40E_MAC_VF || hw->mac.type == I40E_MAC_X722_VF)
2128                 txq->reg_idx = queue_idx;
2129         else /* PF device */
2130                 txq->reg_idx = vsi->base_queue +
2131                         i40e_get_queue_offset_by_qindex(pf, queue_idx);
2132
2133         txq->port_id = dev->data->port_id;
2134         txq->txq_flags = tx_conf->txq_flags;
2135         txq->vsi = vsi;
2136         txq->tx_deferred_start = tx_conf->tx_deferred_start;
2137
2138         txq->tx_ring_phys_addr = rte_mem_phy2mch(tz->memseg_id, tz->phys_addr);
2139         txq->tx_ring = (struct i40e_tx_desc *)tz->addr;
2140
2141         /* Allocate software ring */
2142         txq->sw_ring =
2143                 rte_zmalloc_socket("i40e tx sw ring",
2144                                    sizeof(struct i40e_tx_entry) * nb_desc,
2145                                    RTE_CACHE_LINE_SIZE,
2146                                    socket_id);
2147         if (!txq->sw_ring) {
2148                 i40e_dev_tx_queue_release(txq);
2149                 PMD_DRV_LOG(ERR, "Failed to allocate memory for SW TX ring");
2150                 return -ENOMEM;
2151         }
2152
2153         i40e_reset_tx_queue(txq);
2154         txq->q_set = TRUE;
2155         dev->data->tx_queues[queue_idx] = txq;
2156
2157         /* Use a simple TX queue without offloads or multi segs if possible */
2158         i40e_set_tx_function_flag(dev, txq);
2159
2160         for (i = 0; i < I40E_MAX_TRAFFIC_CLASS; i++) {
2161                 if (!(vsi->enabled_tc & (1 << i)))
2162                         continue;
2163                 tc_mapping = rte_le_to_cpu_16(vsi->info.tc_mapping[i]);
2164                 base = (tc_mapping & I40E_AQ_VSI_TC_QUE_OFFSET_MASK) >>
2165                         I40E_AQ_VSI_TC_QUE_OFFSET_SHIFT;
2166                 bsf = (tc_mapping & I40E_AQ_VSI_TC_QUE_NUMBER_MASK) >>
2167                         I40E_AQ_VSI_TC_QUE_NUMBER_SHIFT;
2168
2169                 if (queue_idx >= base && queue_idx < (base + BIT(bsf)))
2170                         txq->dcb_tc = i;
2171         }
2172
2173         return 0;
2174 }
2175
2176 void
2177 i40e_dev_tx_queue_release(void *txq)
2178 {
2179         struct i40e_tx_queue *q = (struct i40e_tx_queue *)txq;
2180
2181         if (!q) {
2182                 PMD_DRV_LOG(DEBUG, "Pointer to TX queue is NULL");
2183                 return;
2184         }
2185
2186         i40e_tx_queue_release_mbufs(q);
2187         rte_free(q->sw_ring);
2188         rte_free(q);
2189 }
2190
2191 const struct rte_memzone *
2192 i40e_memzone_reserve(const char *name, uint32_t len, int socket_id)
2193 {
2194         const struct rte_memzone *mz;
2195
2196         mz = rte_memzone_lookup(name);
2197         if (mz)
2198                 return mz;
2199
2200         if (rte_xen_dom0_supported())
2201                 mz = rte_memzone_reserve_bounded(name, len,
2202                                 socket_id, 0, I40E_RING_BASE_ALIGN, RTE_PGSIZE_2M);
2203         else
2204                 mz = rte_memzone_reserve_aligned(name, len,
2205                                 socket_id, 0, I40E_RING_BASE_ALIGN);
2206         return mz;
2207 }
2208
2209 void
2210 i40e_rx_queue_release_mbufs(struct i40e_rx_queue *rxq)
2211 {
2212         uint16_t i;
2213
2214         /* SSE Vector driver has a different way of releasing mbufs. */
2215         if (rxq->rx_using_sse) {
2216                 i40e_rx_queue_release_mbufs_vec(rxq);
2217                 return;
2218         }
2219
2220         if (!rxq->sw_ring) {
2221                 PMD_DRV_LOG(DEBUG, "Pointer to sw_ring is NULL");
2222                 return;
2223         }
2224
2225         for (i = 0; i < rxq->nb_rx_desc; i++) {
2226                 if (rxq->sw_ring[i].mbuf) {
2227                         rte_pktmbuf_free_seg(rxq->sw_ring[i].mbuf);
2228                         rxq->sw_ring[i].mbuf = NULL;
2229                 }
2230         }
2231 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
2232         if (rxq->rx_nb_avail == 0)
2233                 return;
2234         for (i = 0; i < rxq->rx_nb_avail; i++) {
2235                 struct rte_mbuf *mbuf;
2236
2237                 mbuf = rxq->rx_stage[rxq->rx_next_avail + i];
2238                 rte_pktmbuf_free_seg(mbuf);
2239         }
2240         rxq->rx_nb_avail = 0;
2241 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
2242 }
2243
2244 void
2245 i40e_reset_rx_queue(struct i40e_rx_queue *rxq)
2246 {
2247         unsigned i;
2248         uint16_t len;
2249
2250         if (!rxq) {
2251                 PMD_DRV_LOG(DEBUG, "Pointer to rxq is NULL");
2252                 return;
2253         }
2254
2255 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
2256         if (check_rx_burst_bulk_alloc_preconditions(rxq) == 0)
2257                 len = (uint16_t)(rxq->nb_rx_desc + RTE_PMD_I40E_RX_MAX_BURST);
2258         else
2259 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
2260                 len = rxq->nb_rx_desc;
2261
2262         for (i = 0; i < len * sizeof(union i40e_rx_desc); i++)
2263                 ((volatile char *)rxq->rx_ring)[i] = 0;
2264
2265 #ifdef RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC
2266         memset(&rxq->fake_mbuf, 0x0, sizeof(rxq->fake_mbuf));
2267         for (i = 0; i < RTE_PMD_I40E_RX_MAX_BURST; ++i)
2268                 rxq->sw_ring[rxq->nb_rx_desc + i].mbuf = &rxq->fake_mbuf;
2269
2270         rxq->rx_nb_avail = 0;
2271         rxq->rx_next_avail = 0;
2272         rxq->rx_free_trigger = (uint16_t)(rxq->rx_free_thresh - 1);
2273 #endif /* RTE_LIBRTE_I40E_RX_ALLOW_BULK_ALLOC */
2274         rxq->rx_tail = 0;
2275         rxq->nb_rx_hold = 0;
2276         rxq->pkt_first_seg = NULL;
2277         rxq->pkt_last_seg = NULL;
2278
2279         rxq->rxrearm_start = 0;
2280         rxq->rxrearm_nb = 0;
2281 }
2282
2283 void
2284 i40e_tx_queue_release_mbufs(struct i40e_tx_queue *txq)
2285 {
2286         uint16_t i;
2287
2288         if (!txq || !txq->sw_ring) {
2289                 PMD_DRV_LOG(DEBUG, "Pointer to rxq or sw_ring is NULL");
2290                 return;
2291         }
2292
2293         for (i = 0; i < txq->nb_tx_desc; i++) {
2294                 if (txq->sw_ring[i].mbuf) {
2295                         rte_pktmbuf_free_seg(txq->sw_ring[i].mbuf);
2296                         txq->sw_ring[i].mbuf = NULL;
2297                 }
2298         }
2299 }
2300
2301 void
2302 i40e_reset_tx_queue(struct i40e_tx_queue *txq)
2303 {
2304         struct i40e_tx_entry *txe;
2305         uint16_t i, prev, size;
2306
2307         if (!txq) {
2308                 PMD_DRV_LOG(DEBUG, "Pointer to txq is NULL");
2309                 return;
2310         }
2311
2312         txe = txq->sw_ring;
2313         size = sizeof(struct i40e_tx_desc) * txq->nb_tx_desc;
2314         for (i = 0; i < size; i++)
2315                 ((volatile char *)txq->tx_ring)[i] = 0;
2316
2317         prev = (uint16_t)(txq->nb_tx_desc - 1);
2318         for (i = 0; i < txq->nb_tx_desc; i++) {
2319                 volatile struct i40e_tx_desc *txd = &txq->tx_ring[i];
2320
2321                 txd->cmd_type_offset_bsz =
2322                         rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DESC_DONE);
2323                 txe[i].mbuf =  NULL;
2324                 txe[i].last_id = i;
2325                 txe[prev].next_id = i;
2326                 prev = i;
2327         }
2328
2329         txq->tx_next_dd = (uint16_t)(txq->tx_rs_thresh - 1);
2330         txq->tx_next_rs = (uint16_t)(txq->tx_rs_thresh - 1);
2331
2332         txq->tx_tail = 0;
2333         txq->nb_tx_used = 0;
2334
2335         txq->last_desc_cleaned = (uint16_t)(txq->nb_tx_desc - 1);
2336         txq->nb_tx_free = (uint16_t)(txq->nb_tx_desc - 1);
2337 }
2338
2339 /* Init the TX queue in hardware */
2340 int
2341 i40e_tx_queue_init(struct i40e_tx_queue *txq)
2342 {
2343         enum i40e_status_code err = I40E_SUCCESS;
2344         struct i40e_vsi *vsi = txq->vsi;
2345         struct i40e_hw *hw = I40E_VSI_TO_HW(vsi);
2346         uint16_t pf_q = txq->reg_idx;
2347         struct i40e_hmc_obj_txq tx_ctx;
2348         uint32_t qtx_ctl;
2349
2350         /* clear the context structure first */
2351         memset(&tx_ctx, 0, sizeof(tx_ctx));
2352         tx_ctx.new_context = 1;
2353         tx_ctx.base = txq->tx_ring_phys_addr / I40E_QUEUE_BASE_ADDR_UNIT;
2354         tx_ctx.qlen = txq->nb_tx_desc;
2355
2356 #ifdef RTE_LIBRTE_IEEE1588
2357         tx_ctx.timesync_ena = 1;
2358 #endif
2359         tx_ctx.rdylist = rte_le_to_cpu_16(vsi->info.qs_handle[txq->dcb_tc]);
2360         if (vsi->type == I40E_VSI_FDIR)
2361                 tx_ctx.fd_ena = TRUE;
2362
2363         err = i40e_clear_lan_tx_queue_context(hw, pf_q);
2364         if (err != I40E_SUCCESS) {
2365                 PMD_DRV_LOG(ERR, "Failure of clean lan tx queue context");
2366                 return err;
2367         }
2368
2369         err = i40e_set_lan_tx_queue_context(hw, pf_q, &tx_ctx);
2370         if (err != I40E_SUCCESS) {
2371                 PMD_DRV_LOG(ERR, "Failure of set lan tx queue context");
2372                 return err;
2373         }
2374
2375         /* Now associate this queue with this PCI function */
2376         qtx_ctl = I40E_QTX_CTL_PF_QUEUE;
2377         qtx_ctl |= ((hw->pf_id << I40E_QTX_CTL_PF_INDX_SHIFT) &
2378                                         I40E_QTX_CTL_PF_INDX_MASK);
2379         I40E_WRITE_REG(hw, I40E_QTX_CTL(pf_q), qtx_ctl);
2380         I40E_WRITE_FLUSH(hw);
2381
2382         txq->qtx_tail = hw->hw_addr + I40E_QTX_TAIL(pf_q);
2383
2384         return err;
2385 }
2386
2387 int
2388 i40e_alloc_rx_queue_mbufs(struct i40e_rx_queue *rxq)
2389 {
2390         struct i40e_rx_entry *rxe = rxq->sw_ring;
2391         uint64_t dma_addr;
2392         uint16_t i;
2393
2394         for (i = 0; i < rxq->nb_rx_desc; i++) {
2395                 volatile union i40e_rx_desc *rxd;
2396                 struct rte_mbuf *mbuf = rte_mbuf_raw_alloc(rxq->mp);
2397
2398                 if (unlikely(!mbuf)) {
2399                         PMD_DRV_LOG(ERR, "Failed to allocate mbuf for RX");
2400                         return -ENOMEM;
2401                 }
2402
2403                 rte_mbuf_refcnt_set(mbuf, 1);
2404                 mbuf->next = NULL;
2405                 mbuf->data_off = RTE_PKTMBUF_HEADROOM;
2406                 mbuf->nb_segs = 1;
2407                 mbuf->port = rxq->port_id;
2408
2409                 dma_addr =
2410                         rte_cpu_to_le_64(rte_mbuf_data_dma_addr_default(mbuf));
2411
2412                 rxd = &rxq->rx_ring[i];
2413                 rxd->read.pkt_addr = dma_addr;
2414                 rxd->read.hdr_addr = 0;
2415 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
2416                 rxd->read.rsvd1 = 0;
2417                 rxd->read.rsvd2 = 0;
2418 #endif /* RTE_LIBRTE_I40E_16BYTE_RX_DESC */
2419
2420                 rxe[i].mbuf = mbuf;
2421         }
2422
2423         return 0;
2424 }
2425
2426 /*
2427  * Calculate the buffer length, and check the jumbo frame
2428  * and maximum packet length.
2429  */
2430 static int
2431 i40e_rx_queue_config(struct i40e_rx_queue *rxq)
2432 {
2433         struct i40e_pf *pf = I40E_VSI_TO_PF(rxq->vsi);
2434         struct i40e_hw *hw = I40E_VSI_TO_HW(rxq->vsi);
2435         struct rte_eth_dev_data *data = pf->dev_data;
2436         uint16_t buf_size, len;
2437
2438         buf_size = (uint16_t)(rte_pktmbuf_data_room_size(rxq->mp) -
2439                 RTE_PKTMBUF_HEADROOM);
2440
2441         switch (pf->flags & (I40E_FLAG_HEADER_SPLIT_DISABLED |
2442                         I40E_FLAG_HEADER_SPLIT_ENABLED)) {
2443         case I40E_FLAG_HEADER_SPLIT_ENABLED: /* Not supported */
2444                 rxq->rx_hdr_len = RTE_ALIGN(I40E_RXBUF_SZ_1024,
2445                                 (1 << I40E_RXQ_CTX_HBUFF_SHIFT));
2446                 rxq->rx_buf_len = RTE_ALIGN(I40E_RXBUF_SZ_2048,
2447                                 (1 << I40E_RXQ_CTX_DBUFF_SHIFT));
2448                 rxq->hs_mode = i40e_header_split_enabled;
2449                 break;
2450         case I40E_FLAG_HEADER_SPLIT_DISABLED:
2451         default:
2452                 rxq->rx_hdr_len = 0;
2453                 rxq->rx_buf_len = RTE_ALIGN(buf_size,
2454                         (1 << I40E_RXQ_CTX_DBUFF_SHIFT));
2455                 rxq->hs_mode = i40e_header_split_none;
2456                 break;
2457         }
2458
2459         len = hw->func_caps.rx_buf_chain_len * rxq->rx_buf_len;
2460         rxq->max_pkt_len = RTE_MIN(len, data->dev_conf.rxmode.max_rx_pkt_len);
2461         if (data->dev_conf.rxmode.jumbo_frame == 1) {
2462                 if (rxq->max_pkt_len <= ETHER_MAX_LEN ||
2463                         rxq->max_pkt_len > I40E_FRAME_SIZE_MAX) {
2464                         PMD_DRV_LOG(ERR, "maximum packet length must "
2465                                     "be larger than %u and smaller than %u,"
2466                                     "as jumbo frame is enabled",
2467                                     (uint32_t)ETHER_MAX_LEN,
2468                                     (uint32_t)I40E_FRAME_SIZE_MAX);
2469                         return I40E_ERR_CONFIG;
2470                 }
2471         } else {
2472                 if (rxq->max_pkt_len < ETHER_MIN_LEN ||
2473                         rxq->max_pkt_len > ETHER_MAX_LEN) {
2474                         PMD_DRV_LOG(ERR, "maximum packet length must be "
2475                                     "larger than %u and smaller than %u, "
2476                                     "as jumbo frame is disabled",
2477                                     (uint32_t)ETHER_MIN_LEN,
2478                                     (uint32_t)ETHER_MAX_LEN);
2479                         return I40E_ERR_CONFIG;
2480                 }
2481         }
2482
2483         return 0;
2484 }
2485
2486 /* Init the RX queue in hardware */
2487 int
2488 i40e_rx_queue_init(struct i40e_rx_queue *rxq)
2489 {
2490         int err = I40E_SUCCESS;
2491         struct i40e_hw *hw = I40E_VSI_TO_HW(rxq->vsi);
2492         struct rte_eth_dev_data *dev_data = I40E_VSI_TO_DEV_DATA(rxq->vsi);
2493         uint16_t pf_q = rxq->reg_idx;
2494         uint16_t buf_size;
2495         struct i40e_hmc_obj_rxq rx_ctx;
2496
2497         err = i40e_rx_queue_config(rxq);
2498         if (err < 0) {
2499                 PMD_DRV_LOG(ERR, "Failed to config RX queue");
2500                 return err;
2501         }
2502
2503         /* Clear the context structure first */
2504         memset(&rx_ctx, 0, sizeof(struct i40e_hmc_obj_rxq));
2505         rx_ctx.dbuff = rxq->rx_buf_len >> I40E_RXQ_CTX_DBUFF_SHIFT;
2506         rx_ctx.hbuff = rxq->rx_hdr_len >> I40E_RXQ_CTX_HBUFF_SHIFT;
2507
2508         rx_ctx.base = rxq->rx_ring_phys_addr / I40E_QUEUE_BASE_ADDR_UNIT;
2509         rx_ctx.qlen = rxq->nb_rx_desc;
2510 #ifndef RTE_LIBRTE_I40E_16BYTE_RX_DESC
2511         rx_ctx.dsize = 1;
2512 #endif
2513         rx_ctx.dtype = rxq->hs_mode;
2514         if (rxq->hs_mode)
2515                 rx_ctx.hsplit_0 = I40E_HEADER_SPLIT_ALL;
2516         else
2517                 rx_ctx.hsplit_0 = I40E_HEADER_SPLIT_NONE;
2518         rx_ctx.rxmax = rxq->max_pkt_len;
2519         rx_ctx.tphrdesc_ena = 1;
2520         rx_ctx.tphwdesc_ena = 1;
2521         rx_ctx.tphdata_ena = 1;
2522         rx_ctx.tphhead_ena = 1;
2523         rx_ctx.lrxqthresh = 2;
2524         rx_ctx.crcstrip = (rxq->crc_len == 0) ? 1 : 0;
2525         rx_ctx.l2tsel = 1;
2526         /* showiv indicates if inner VLAN is stripped inside of tunnel
2527          * packet. When set it to 1, vlan information is stripped from
2528          * the inner header, but the hardware does not put it in the
2529          * descriptor. So set it zero by default.
2530          */
2531         rx_ctx.showiv = 0;
2532         rx_ctx.prefena = 1;
2533
2534         err = i40e_clear_lan_rx_queue_context(hw, pf_q);
2535         if (err != I40E_SUCCESS) {
2536                 PMD_DRV_LOG(ERR, "Failed to clear LAN RX queue context");
2537                 return err;
2538         }
2539         err = i40e_set_lan_rx_queue_context(hw, pf_q, &rx_ctx);
2540         if (err != I40E_SUCCESS) {
2541                 PMD_DRV_LOG(ERR, "Failed to set LAN RX queue context");
2542                 return err;
2543         }
2544
2545         rxq->qrx_tail = hw->hw_addr + I40E_QRX_TAIL(pf_q);
2546
2547         buf_size = (uint16_t)(rte_pktmbuf_data_room_size(rxq->mp) -
2548                 RTE_PKTMBUF_HEADROOM);
2549
2550         /* Check if scattered RX needs to be used. */
2551         if ((rxq->max_pkt_len + 2 * I40E_VLAN_TAG_SIZE) > buf_size) {
2552                 dev_data->scattered_rx = 1;
2553         }
2554
2555         /* Init the RX tail regieter. */
2556         I40E_PCI_REG_WRITE(rxq->qrx_tail, rxq->nb_rx_desc - 1);
2557
2558         return 0;
2559 }
2560
2561 void
2562 i40e_dev_clear_queues(struct rte_eth_dev *dev)
2563 {
2564         uint16_t i;
2565
2566         PMD_INIT_FUNC_TRACE();
2567
2568         for (i = 0; i < dev->data->nb_tx_queues; i++) {
2569                 if (!dev->data->tx_queues[i])
2570                         continue;
2571                 i40e_tx_queue_release_mbufs(dev->data->tx_queues[i]);
2572                 i40e_reset_tx_queue(dev->data->tx_queues[i]);
2573         }
2574
2575         for (i = 0; i < dev->data->nb_rx_queues; i++) {
2576                 if (!dev->data->rx_queues[i])
2577                         continue;
2578                 i40e_rx_queue_release_mbufs(dev->data->rx_queues[i]);
2579                 i40e_reset_rx_queue(dev->data->rx_queues[i]);
2580         }
2581 }
2582
2583 void
2584 i40e_dev_free_queues(struct rte_eth_dev *dev)
2585 {
2586         uint16_t i;
2587
2588         PMD_INIT_FUNC_TRACE();
2589
2590         for (i = 0; i < dev->data->nb_rx_queues; i++) {
2591                 if (!dev->data->rx_queues[i])
2592                         continue;
2593                 i40e_dev_rx_queue_release(dev->data->rx_queues[i]);
2594                 dev->data->rx_queues[i] = NULL;
2595         }
2596         dev->data->nb_rx_queues = 0;
2597
2598         for (i = 0; i < dev->data->nb_tx_queues; i++) {
2599                 if (!dev->data->tx_queues[i])
2600                         continue;
2601                 i40e_dev_tx_queue_release(dev->data->tx_queues[i]);
2602                 dev->data->tx_queues[i] = NULL;
2603         }
2604         dev->data->nb_tx_queues = 0;
2605 }
2606
2607 #define I40E_FDIR_NUM_TX_DESC  I40E_MIN_RING_DESC
2608 #define I40E_FDIR_NUM_RX_DESC  I40E_MIN_RING_DESC
2609
2610 enum i40e_status_code
2611 i40e_fdir_setup_tx_resources(struct i40e_pf *pf)
2612 {
2613         struct i40e_tx_queue *txq;
2614         const struct rte_memzone *tz = NULL;
2615         uint32_t ring_size;
2616         struct rte_eth_dev *dev;
2617
2618         if (!pf) {
2619                 PMD_DRV_LOG(ERR, "PF is not available");
2620                 return I40E_ERR_BAD_PTR;
2621         }
2622
2623         dev = pf->adapter->eth_dev;
2624
2625         /* Allocate the TX queue data structure. */
2626         txq = rte_zmalloc_socket("i40e fdir tx queue",
2627                                   sizeof(struct i40e_tx_queue),
2628                                   RTE_CACHE_LINE_SIZE,
2629                                   SOCKET_ID_ANY);
2630         if (!txq) {
2631                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
2632                                         "tx queue structure.");
2633                 return I40E_ERR_NO_MEMORY;
2634         }
2635
2636         /* Allocate TX hardware ring descriptors. */
2637         ring_size = sizeof(struct i40e_tx_desc) * I40E_FDIR_NUM_TX_DESC;
2638         ring_size = RTE_ALIGN(ring_size, I40E_DMA_MEM_ALIGN);
2639
2640         tz = rte_eth_dma_zone_reserve(dev, "fdir_tx_ring",
2641                                       I40E_FDIR_QUEUE_ID, ring_size,
2642                                       I40E_RING_BASE_ALIGN, SOCKET_ID_ANY);
2643         if (!tz) {
2644                 i40e_dev_tx_queue_release(txq);
2645                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for TX.");
2646                 return I40E_ERR_NO_MEMORY;
2647         }
2648
2649         txq->nb_tx_desc = I40E_FDIR_NUM_TX_DESC;
2650         txq->queue_id = I40E_FDIR_QUEUE_ID;
2651         txq->reg_idx = pf->fdir.fdir_vsi->base_queue;
2652         txq->vsi = pf->fdir.fdir_vsi;
2653
2654         txq->tx_ring_phys_addr = rte_mem_phy2mch(tz->memseg_id, tz->phys_addr);
2655         txq->tx_ring = (struct i40e_tx_desc *)tz->addr;
2656         /*
2657          * don't need to allocate software ring and reset for the fdir
2658          * program queue just set the queue has been configured.
2659          */
2660         txq->q_set = TRUE;
2661         pf->fdir.txq = txq;
2662
2663         return I40E_SUCCESS;
2664 }
2665
2666 enum i40e_status_code
2667 i40e_fdir_setup_rx_resources(struct i40e_pf *pf)
2668 {
2669         struct i40e_rx_queue *rxq;
2670         const struct rte_memzone *rz = NULL;
2671         uint32_t ring_size;
2672         struct rte_eth_dev *dev;
2673
2674         if (!pf) {
2675                 PMD_DRV_LOG(ERR, "PF is not available");
2676                 return I40E_ERR_BAD_PTR;
2677         }
2678
2679         dev = pf->adapter->eth_dev;
2680
2681         /* Allocate the RX queue data structure. */
2682         rxq = rte_zmalloc_socket("i40e fdir rx queue",
2683                                   sizeof(struct i40e_rx_queue),
2684                                   RTE_CACHE_LINE_SIZE,
2685                                   SOCKET_ID_ANY);
2686         if (!rxq) {
2687                 PMD_DRV_LOG(ERR, "Failed to allocate memory for "
2688                                         "rx queue structure.");
2689                 return I40E_ERR_NO_MEMORY;
2690         }
2691
2692         /* Allocate RX hardware ring descriptors. */
2693         ring_size = sizeof(union i40e_rx_desc) * I40E_FDIR_NUM_RX_DESC;
2694         ring_size = RTE_ALIGN(ring_size, I40E_DMA_MEM_ALIGN);
2695
2696         rz = rte_eth_dma_zone_reserve(dev, "fdir_rx_ring",
2697                                       I40E_FDIR_QUEUE_ID, ring_size,
2698                                       I40E_RING_BASE_ALIGN, SOCKET_ID_ANY);
2699         if (!rz) {
2700                 i40e_dev_rx_queue_release(rxq);
2701                 PMD_DRV_LOG(ERR, "Failed to reserve DMA memory for RX.");
2702                 return I40E_ERR_NO_MEMORY;
2703         }
2704
2705         rxq->nb_rx_desc = I40E_FDIR_NUM_RX_DESC;
2706         rxq->queue_id = I40E_FDIR_QUEUE_ID;
2707         rxq->reg_idx = pf->fdir.fdir_vsi->base_queue;
2708         rxq->vsi = pf->fdir.fdir_vsi;
2709
2710         rxq->rx_ring_phys_addr = rte_mem_phy2mch(rz->memseg_id, rz->phys_addr);
2711         rxq->rx_ring = (union i40e_rx_desc *)rz->addr;
2712
2713         /*
2714          * Don't need to allocate software ring and reset for the fdir
2715          * rx queue, just set the queue has been configured.
2716          */
2717         rxq->q_set = TRUE;
2718         pf->fdir.rxq = rxq;
2719
2720         return I40E_SUCCESS;
2721 }
2722
2723 void
2724 i40e_rxq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
2725         struct rte_eth_rxq_info *qinfo)
2726 {
2727         struct i40e_rx_queue *rxq;
2728
2729         rxq = dev->data->rx_queues[queue_id];
2730
2731         qinfo->mp = rxq->mp;
2732         qinfo->scattered_rx = dev->data->scattered_rx;
2733         qinfo->nb_desc = rxq->nb_rx_desc;
2734
2735         qinfo->conf.rx_free_thresh = rxq->rx_free_thresh;
2736         qinfo->conf.rx_drop_en = rxq->drop_en;
2737         qinfo->conf.rx_deferred_start = rxq->rx_deferred_start;
2738 }
2739
2740 void
2741 i40e_txq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
2742         struct rte_eth_txq_info *qinfo)
2743 {
2744         struct i40e_tx_queue *txq;
2745
2746         txq = dev->data->tx_queues[queue_id];
2747
2748         qinfo->nb_desc = txq->nb_tx_desc;
2749
2750         qinfo->conf.tx_thresh.pthresh = txq->pthresh;
2751         qinfo->conf.tx_thresh.hthresh = txq->hthresh;
2752         qinfo->conf.tx_thresh.wthresh = txq->wthresh;
2753
2754         qinfo->conf.tx_free_thresh = txq->tx_free_thresh;
2755         qinfo->conf.tx_rs_thresh = txq->tx_rs_thresh;
2756         qinfo->conf.txq_flags = txq->txq_flags;
2757         qinfo->conf.tx_deferred_start = txq->tx_deferred_start;
2758 }
2759
2760 void __attribute__((cold))
2761 i40e_set_rx_function(struct rte_eth_dev *dev)
2762 {
2763         struct i40e_adapter *ad =
2764                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
2765         uint16_t rx_using_sse, i;
2766         /* In order to allow Vector Rx there are a few configuration
2767          * conditions to be met and Rx Bulk Allocation should be allowed.
2768          */
2769         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
2770                 if (i40e_rx_vec_dev_conf_condition_check(dev) ||
2771                     !ad->rx_bulk_alloc_allowed) {
2772                         PMD_INIT_LOG(DEBUG, "Port[%d] doesn't meet"
2773                                      " Vector Rx preconditions",
2774                                      dev->data->port_id);
2775
2776                         ad->rx_vec_allowed = false;
2777                 }
2778                 if (ad->rx_vec_allowed) {
2779                         for (i = 0; i < dev->data->nb_rx_queues; i++) {
2780                                 struct i40e_rx_queue *rxq =
2781                                         dev->data->rx_queues[i];
2782
2783                                 if (rxq && i40e_rxq_vec_setup(rxq)) {
2784                                         ad->rx_vec_allowed = false;
2785                                         break;
2786                                 }
2787                         }
2788                 }
2789         }
2790
2791         if (dev->data->scattered_rx) {
2792                 /* Set the non-LRO scattered callback: there are Vector and
2793                  * single allocation versions.
2794                  */
2795                 if (ad->rx_vec_allowed) {
2796                         PMD_INIT_LOG(DEBUG, "Using Vector Scattered Rx "
2797                                             "callback (port=%d).",
2798                                      dev->data->port_id);
2799
2800                         dev->rx_pkt_burst = i40e_recv_scattered_pkts_vec;
2801                 } else {
2802                         PMD_INIT_LOG(DEBUG, "Using a Scattered with bulk "
2803                                            "allocation callback (port=%d).",
2804                                      dev->data->port_id);
2805                         dev->rx_pkt_burst = i40e_recv_scattered_pkts;
2806                 }
2807         /* If parameters allow we are going to choose between the following
2808          * callbacks:
2809          *    - Vector
2810          *    - Bulk Allocation
2811          *    - Single buffer allocation (the simplest one)
2812          */
2813         } else if (ad->rx_vec_allowed) {
2814                 PMD_INIT_LOG(DEBUG, "Vector rx enabled, please make sure RX "
2815                                     "burst size no less than %d (port=%d).",
2816                              RTE_I40E_DESCS_PER_LOOP,
2817                              dev->data->port_id);
2818
2819                 dev->rx_pkt_burst = i40e_recv_pkts_vec;
2820         } else if (ad->rx_bulk_alloc_allowed) {
2821                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions are "
2822                                     "satisfied. Rx Burst Bulk Alloc function "
2823                                     "will be used on port=%d.",
2824                              dev->data->port_id);
2825
2826                 dev->rx_pkt_burst = i40e_recv_pkts_bulk_alloc;
2827         } else {
2828                 PMD_INIT_LOG(DEBUG, "Rx Burst Bulk Alloc Preconditions are not "
2829                                     "satisfied, or Scattered Rx is requested "
2830                                     "(port=%d).",
2831                              dev->data->port_id);
2832
2833                 dev->rx_pkt_burst = i40e_recv_pkts;
2834         }
2835
2836         /* Propagate information about RX function choice through all queues. */
2837         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
2838                 rx_using_sse =
2839                         (dev->rx_pkt_burst == i40e_recv_scattered_pkts_vec ||
2840                          dev->rx_pkt_burst == i40e_recv_pkts_vec);
2841
2842                 for (i = 0; i < dev->data->nb_rx_queues; i++) {
2843                         struct i40e_rx_queue *rxq = dev->data->rx_queues[i];
2844
2845                         if (rxq)
2846                                 rxq->rx_using_sse = rx_using_sse;
2847                 }
2848         }
2849 }
2850
2851 void __attribute__((cold))
2852 i40e_set_tx_function_flag(struct rte_eth_dev *dev, struct i40e_tx_queue *txq)
2853 {
2854         struct i40e_adapter *ad =
2855                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
2856
2857         /* Use a simple Tx queue (no offloads, no multi segs) if possible */
2858         if (((txq->txq_flags & I40E_SIMPLE_FLAGS) == I40E_SIMPLE_FLAGS)
2859                         && (txq->tx_rs_thresh >= RTE_PMD_I40E_TX_MAX_BURST)) {
2860                 if (txq->tx_rs_thresh <= RTE_I40E_TX_MAX_FREE_BUF_SZ) {
2861                         PMD_INIT_LOG(DEBUG, "Vector tx"
2862                                      " can be enabled on this txq.");
2863
2864                 } else {
2865                         ad->tx_vec_allowed = false;
2866                 }
2867         } else {
2868                 ad->tx_simple_allowed = false;
2869         }
2870 }
2871
2872 void __attribute__((cold))
2873 i40e_set_tx_function(struct rte_eth_dev *dev)
2874 {
2875         struct i40e_adapter *ad =
2876                 I40E_DEV_PRIVATE_TO_ADAPTER(dev->data->dev_private);
2877         int i;
2878
2879         if (rte_eal_process_type() == RTE_PROC_PRIMARY) {
2880                 if (ad->tx_vec_allowed) {
2881                         for (i = 0; i < dev->data->nb_tx_queues; i++) {
2882                                 struct i40e_tx_queue *txq =
2883                                         dev->data->tx_queues[i];
2884
2885                                 if (txq && i40e_txq_vec_setup(txq)) {
2886                                         ad->tx_vec_allowed = false;
2887                                         break;
2888                                 }
2889                         }
2890                 }
2891         }
2892
2893         if (ad->tx_simple_allowed) {
2894                 if (ad->tx_vec_allowed) {
2895                         PMD_INIT_LOG(DEBUG, "Vector tx finally be used.");
2896                         dev->tx_pkt_burst = i40e_xmit_pkts_vec;
2897                 } else {
2898                         PMD_INIT_LOG(DEBUG, "Simple tx finally be used.");
2899                         dev->tx_pkt_burst = i40e_xmit_pkts_simple;
2900                 }
2901                 dev->tx_pkt_prepare = NULL;
2902         } else {
2903                 PMD_INIT_LOG(DEBUG, "Xmit tx finally be used.");
2904                 dev->tx_pkt_burst = i40e_xmit_pkts;
2905                 dev->tx_pkt_prepare = i40e_prep_pkts;
2906         }
2907 }
2908
2909 /* Stubs needed for linkage when CONFIG_RTE_I40E_INC_VECTOR is set to 'n' */
2910 int __attribute__((weak))
2911 i40e_rx_vec_dev_conf_condition_check(struct rte_eth_dev __rte_unused *dev)
2912 {
2913         return -1;
2914 }
2915
2916 uint16_t __attribute__((weak))
2917 i40e_recv_pkts_vec(
2918         void __rte_unused *rx_queue,
2919         struct rte_mbuf __rte_unused **rx_pkts,
2920         uint16_t __rte_unused nb_pkts)
2921 {
2922         return 0;
2923 }
2924
2925 uint16_t __attribute__((weak))
2926 i40e_recv_scattered_pkts_vec(
2927         void __rte_unused *rx_queue,
2928         struct rte_mbuf __rte_unused **rx_pkts,
2929         uint16_t __rte_unused nb_pkts)
2930 {
2931         return 0;
2932 }
2933
2934 int __attribute__((weak))
2935 i40e_rxq_vec_setup(struct i40e_rx_queue __rte_unused *rxq)
2936 {
2937         return -1;
2938 }
2939
2940 int __attribute__((weak))
2941 i40e_txq_vec_setup(struct i40e_tx_queue __rte_unused *txq)
2942 {
2943         return -1;
2944 }
2945
2946 void __attribute__((weak))
2947 i40e_rx_queue_release_mbufs_vec(struct i40e_rx_queue __rte_unused*rxq)
2948 {
2949         return;
2950 }
2951
2952 uint16_t __attribute__((weak))
2953 i40e_xmit_pkts_vec(void __rte_unused *tx_queue,
2954                    struct rte_mbuf __rte_unused **tx_pkts,
2955                    uint16_t __rte_unused nb_pkts)
2956 {
2957         return 0;
2958 }