summaryrefslogtreecommitdiff
path: root/drivers/net/mlx4/mlx4_rxtx.c
blob: 391271a616ef1a05ddf6ef91881e4318f633efad (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
/* SPDX-License-Identifier: BSD-3-Clause
 * Copyright 2017 6WIND S.A.
 * Copyright 2017 Mellanox Technologies, Ltd
 */

/**
 * @file
 * Data plane functions for mlx4 driver.
 */

#include <assert.h>
#include <stdint.h>
#include <string.h>

/* Verbs headers do not support -pedantic. */
#ifdef PEDANTIC
#pragma GCC diagnostic ignored "-Wpedantic"
#endif
#include <infiniband/verbs.h>
#ifdef PEDANTIC
#pragma GCC diagnostic error "-Wpedantic"
#endif

#include <rte_branch_prediction.h>
#include <rte_common.h>
#include <rte_io.h>
#include <rte_mbuf.h>
#include <rte_mempool.h>
#include <rte_prefetch.h>

#include "mlx4.h"
#include "mlx4_prm.h"
#include "mlx4_rxtx.h"
#include "mlx4_utils.h"

/**
 * Pointer-value pair structure used in tx_post_send for saving the first
 * DWORD (32 byte) of a TXBB.
 */
struct pv {
	union {
		volatile struct mlx4_wqe_data_seg *dseg;
		volatile uint32_t *dst;
	};
	uint32_t val;
};

/** A helper structure for TSO packet handling. */
struct tso_info {
	/** Pointer to the array of saved first DWORD (32 byte) of a TXBB. */
	struct pv *pv;
	/** Current entry in the pv array. */
	int pv_counter;
	/** Total size of the WQE including padding. */
	uint32_t wqe_size;
	/** Size of TSO header to prepend to each packet to send. */
	uint16_t tso_header_size;
	/** Total size of the TSO segment in the WQE. */
	uint16_t wqe_tso_seg_size;
	/** Raw WQE size in units of 16 Bytes and without padding. */
	uint8_t fence_size;
};

/** A table to translate Rx completion flags to packet type. */
uint32_t mlx4_ptype_table[0x100] __rte_cache_aligned = {
	/*
	 * The index to the array should have:
	 *  bit[7] - MLX4_CQE_L2_TUNNEL
	 *  bit[6] - MLX4_CQE_L2_TUNNEL_IPV4
	 *  bit[5] - MLX4_CQE_STATUS_UDP
	 *  bit[4] - MLX4_CQE_STATUS_TCP
	 *  bit[3] - MLX4_CQE_STATUS_IPV4OPT
	 *  bit[2] - MLX4_CQE_STATUS_IPV6
	 *  bit[1] - MLX4_CQE_STATUS_IPF
	 *  bit[0] - MLX4_CQE_STATUS_IPV4
	 * giving a total of up to 256 entries.
	 */
	/* L2 */
	[0x00] = RTE_PTYPE_L2_ETHER,
	/* L3 */
	[0x01] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_L4_NONFRAG,
	[0x02] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_L4_FRAG,
	[0x03] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_L4_FRAG,
	[0x04] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_L4_NONFRAG,
	[0x06] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_L4_FRAG,
	[0x08] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT |
		     RTE_PTYPE_L4_NONFRAG,
	[0x09] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT |
		     RTE_PTYPE_L4_NONFRAG,
	[0x0a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT |
		     RTE_PTYPE_L4_FRAG,
	[0x0b] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT |
		     RTE_PTYPE_L4_FRAG,
	/* TCP */
	[0x11] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_L4_TCP,
	[0x14] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_L4_TCP,
	[0x16] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_L4_FRAG,
	[0x18] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT |
		     RTE_PTYPE_L4_TCP,
	[0x19] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT |
		     RTE_PTYPE_L4_TCP,
	/* UDP */
	[0x21] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_L4_UDP,
	[0x24] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_L4_UDP,
	[0x26] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_L4_FRAG,
	[0x28] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT |
		     RTE_PTYPE_L4_UDP,
	[0x29] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT |
		     RTE_PTYPE_L4_UDP,
	/* Tunneled - L3 IPV6 */
	[0x80] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
	[0x81] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L4_NONFRAG,
	[0x82] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L4_FRAG,
	[0x83] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L4_FRAG,
	[0x84] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L4_NONFRAG,
	[0x86] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L4_FRAG,
	[0x88] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT |
		     RTE_PTYPE_INNER_L4_NONFRAG,
	[0x89] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT |
		     RTE_PTYPE_INNER_L4_NONFRAG,
	[0x8a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT |
		     RTE_PTYPE_INNER_L4_FRAG,
	[0x8b] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT |
		     RTE_PTYPE_INNER_L4_FRAG,
	/* Tunneled - L3 IPV6, TCP */
	[0x91] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L4_TCP,
	[0x94] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L4_TCP,
	[0x96] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L4_FRAG,
	[0x98] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT | RTE_PTYPE_INNER_L4_TCP,
	[0x99] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT | RTE_PTYPE_INNER_L4_TCP,
	/* Tunneled - L3 IPV6, UDP */
	[0xa1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L4_UDP,
	[0xa4] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L4_UDP,
	[0xa6] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L4_FRAG,
	[0xa8] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT |
		     RTE_PTYPE_INNER_L4_UDP,
	[0xa9] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT |
		     RTE_PTYPE_INNER_L4_UDP,
	/* Tunneled - L3 IPV4 */
	[0xc0] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
	[0xc1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L4_NONFRAG,
	[0xc2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L4_FRAG,
	[0xc3] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L4_FRAG,
	[0xc4] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L4_NONFRAG,
	[0xc6] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L4_FRAG,
	[0xc8] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT |
		     RTE_PTYPE_INNER_L4_NONFRAG,
	[0xc9] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT |
		     RTE_PTYPE_INNER_L4_NONFRAG,
	[0xca] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT |
		     RTE_PTYPE_INNER_L4_FRAG,
	[0xcb] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT |
		     RTE_PTYPE_INNER_L4_FRAG,
	/* Tunneled - L3 IPV4, TCP */
	[0xd1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L4_TCP,
	[0xd4] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L4_TCP,
	[0xd6] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L4_FRAG,
	[0xd8] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT |
		     RTE_PTYPE_INNER_L4_TCP,
	[0xd9] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT |
		     RTE_PTYPE_INNER_L4_TCP,
	/* Tunneled - L3 IPV4, UDP */
	[0xe1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L4_UDP,
	[0xe4] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L4_UDP,
	[0xe6] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L4_FRAG,
	[0xe8] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT |
		     RTE_PTYPE_INNER_L4_UDP,
	[0xe9] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
		     RTE_PTYPE_INNER_L3_IPV4_EXT |
		     RTE_PTYPE_INNER_L4_UDP,
};

/**
 * Stamp TXBB burst so it won't be reused by the HW.
 *
 * Routine is used when freeing WQE used by the chip or when failing
 * building an WQ entry has failed leaving partial information on the queue.
 *
 * @param sq
 *   Pointer to the SQ structure.
 * @param start
 *   Pointer to the first TXBB to stamp.
 * @param end
 *   Pointer to the followed end TXBB to stamp.
 *
 * @return
 *   Stamping burst size in byte units.
 */
static uint32_t
mlx4_txq_stamp_freed_wqe(struct mlx4_sq *sq, volatile uint32_t *start,
			 volatile uint32_t *end)
{
	uint32_t stamp = sq->stamp;
	int32_t size = (intptr_t)end - (intptr_t)start;

	assert(start != end);
	/* Hold SQ ring wrap around. */
	if (size < 0) {
		size = (int32_t)sq->size + size;
		do {
			*start = stamp;
			start += MLX4_SQ_STAMP_DWORDS;
		} while (start != (volatile uint32_t *)sq->eob);
		start = (volatile uint32_t *)sq->buf;
		/* Flip invalid stamping ownership. */
		stamp ^= RTE_BE32(1u << MLX4_SQ_OWNER_BIT);
		sq->stamp = stamp;
		if (start == end)
			return size;
	}
	do {
		*start = stamp;
		start += MLX4_SQ_STAMP_DWORDS;
	} while (start != end);
	return (uint32_t)size;
}

/**
 * Manage Tx completions.
 *
 * When sending a burst, mlx4_tx_burst() posts several WRs.
 * To improve performance, a completion event is only required once every
 * MLX4_PMD_TX_PER_COMP_REQ sends. Doing so discards completion information
 * for other WRs, but this information would not be used anyway.
 *
 * @param txq
 *   Pointer to Tx queue structure.
 * @param elts_m
 *   Tx elements number mask.
 * @param sq
 *   Pointer to the SQ structure.
 */
static void
mlx4_txq_complete(struct txq *txq, const unsigned int elts_m,
		  struct mlx4_sq *sq)
{
	unsigned int elts_tail = txq->elts_tail;
	struct mlx4_cq *cq = &txq->mcq;
	volatile struct mlx4_cqe *cqe;
	uint32_t completed;
	uint32_t cons_index = cq->cons_index;
	volatile uint32_t *first_txbb;

	/*
	 * Traverse over all CQ entries reported and handle each WQ entry
	 * reported by them.
	 */
	do {
		cqe = (volatile struct mlx4_cqe *)mlx4_get_cqe(cq, cons_index);
		if (unlikely(!!(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK) ^
		    !!(cons_index & cq->cqe_cnt)))
			break;
#ifndef NDEBUG
		/*
		 * Make sure we read the CQE after we read the ownership bit.
		 */
		rte_io_rmb();
		if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
			     MLX4_CQE_OPCODE_ERROR)) {
			volatile struct mlx4_err_cqe *cqe_err =
				(volatile struct mlx4_err_cqe *)cqe;
			ERROR("%p CQE error - vendor syndrome: 0x%x"
			      " syndrome: 0x%x\n",
			      (void *)txq, cqe_err->vendor_err,
			      cqe_err->syndrome);
			break;
		}
#endif /* NDEBUG */
		cons_index++;
	} while (1);
	completed = (cons_index - cq->cons_index) * txq->elts_comp_cd_init;
	if (unlikely(!completed))
		return;
	/* First stamping address is the end of the last one. */
	first_txbb = (&(*txq->elts)[elts_tail & elts_m])->eocb;
	elts_tail += completed;
	/* The new tail element holds the end address. */
	sq->remain_size += mlx4_txq_stamp_freed_wqe(sq, first_txbb,
		(&(*txq->elts)[elts_tail & elts_m])->eocb);
	/* Update CQ consumer index. */
	cq->cons_index = cons_index;
	*cq->set_ci_db = rte_cpu_to_be_32(cons_index & MLX4_CQ_DB_CI_MASK);
	txq->elts_tail = elts_tail;
}

/**
 * Write Tx data segment to the SQ.
 *
 * @param dseg
 *   Pointer to data segment in SQ.
 * @param lkey
 *   Memory region lkey.
 * @param addr
 *   Data address.
 * @param byte_count
 *   Big endian bytes count of the data to send.
 */
static inline void
mlx4_fill_tx_data_seg(volatile struct mlx4_wqe_data_seg *dseg,
		       uint32_t lkey, uintptr_t addr, rte_be32_t  byte_count)
{
	dseg->addr = rte_cpu_to_be_64(addr);
	dseg->lkey = lkey;
#if RTE_CACHE_LINE_SIZE < 64
	/*
	 * Need a barrier here before writing the byte_count
	 * fields to make sure that all the data is visible
	 * before the byte_count field is set.
	 * Otherwise, if the segment begins a new cacheline,
	 * the HCA prefetcher could grab the 64-byte chunk and
	 * get a valid (!= 0xffffffff) byte count but stale
	 * data, and end up sending the wrong data.
	 */
	rte_io_wmb();
#endif /* RTE_CACHE_LINE_SIZE */
	dseg->byte_count = byte_count;
}

/**
 * Obtain and calculate TSO information needed for assembling a TSO WQE.
 *
 * @param buf
 *   Pointer to the first packet mbuf.
 * @param txq
 *   Pointer to Tx queue structure.
 * @param tinfo
 *   Pointer to a structure to fill the info with.
 *
 * @return
 *   0 on success, negative value upon error.
 */
static inline int
mlx4_tx_burst_tso_get_params(struct rte_mbuf *buf,
			     struct txq *txq,
			     struct tso_info *tinfo)
{
	struct mlx4_sq *sq = &txq->msq;
	const uint8_t tunneled = txq->priv->hw_csum_l2tun &&
				 (buf->ol_flags & PKT_TX_TUNNEL_MASK);

	tinfo->tso_header_size = buf->l2_len + buf->l3_len + buf->l4_len;
	if (tunneled)
		tinfo->tso_header_size +=
				buf->outer_l2_len + buf->outer_l3_len;
	if (unlikely(buf->tso_segsz == 0 ||
		     tinfo->tso_header_size == 0 ||
		     tinfo->tso_header_size > MLX4_MAX_TSO_HEADER ||
		     tinfo->tso_header_size > buf->data_len))
		return -EINVAL;
	/*
	 * Calculate the WQE TSO segment size
	 * Note:
	 * 1. An LSO segment must be padded such that the subsequent data
	 *    segment is 16-byte aligned.
	 * 2. The start address of the TSO segment is always 16 Bytes aligned.
	 */
	tinfo->wqe_tso_seg_size = RTE_ALIGN(sizeof(struct mlx4_wqe_lso_seg) +
					    tinfo->tso_header_size,
					    sizeof(struct mlx4_wqe_data_seg));
	tinfo->fence_size = ((sizeof(struct mlx4_wqe_ctrl_seg) +
			     tinfo->wqe_tso_seg_size) >> MLX4_SEG_SHIFT) +
			     buf->nb_segs;
	tinfo->wqe_size =
		RTE_ALIGN((uint32_t)(tinfo->fence_size << MLX4_SEG_SHIFT),
			  MLX4_TXBB_SIZE);
	/* Validate WQE size and WQE space in the send queue. */
	if (sq->remain_size < tinfo->wqe_size ||
	    tinfo->wqe_size > MLX4_MAX_WQE_SIZE)
		return -ENOMEM;
	/* Init pv. */
	tinfo->pv = (struct pv *)txq->bounce_buf;
	tinfo->pv_counter = 0;
	return 0;
}

/**
 * Fill the TSO WQE data segments with info on buffers to transmit .
 *
 * @param buf
 *   Pointer to the first packet mbuf.
 * @param txq
 *   Pointer to Tx queue structure.
 * @param tinfo
 *   Pointer to TSO info to use.
 * @param dseg
 *   Pointer to the first data segment in the TSO WQE.
 * @param ctrl
 *   Pointer to the control segment in the TSO WQE.
 *
 * @return
 *   0 on success, negative value upon error.
 */
static inline volatile struct mlx4_wqe_ctrl_seg *
mlx4_tx_burst_fill_tso_dsegs(struct rte_mbuf *buf,
			     struct txq *txq,
			     struct tso_info *tinfo,
			     volatile struct mlx4_wqe_data_seg *dseg,
			     volatile struct mlx4_wqe_ctrl_seg *ctrl)
{
	uint32_t lkey;
	int nb_segs = buf->nb_segs;
	int nb_segs_txbb;
	struct mlx4_sq *sq = &txq->msq;
	struct rte_mbuf *sbuf = buf;
	struct pv *pv = tinfo->pv;
	int *pv_counter = &tinfo->pv_counter;
	volatile struct mlx4_wqe_ctrl_seg *ctrl_next =
			(volatile struct mlx4_wqe_ctrl_seg *)
				((volatile uint8_t *)ctrl + tinfo->wqe_size);
	uint16_t data_len = sbuf->data_len - tinfo->tso_header_size;
	uintptr_t data_addr = rte_pktmbuf_mtod_offset(sbuf, uintptr_t,
						      tinfo->tso_header_size);

	do {
		/* how many dseg entries do we have in the current TXBB ? */
		nb_segs_txbb = (MLX4_TXBB_SIZE -
				((uintptr_t)dseg & (MLX4_TXBB_SIZE - 1))) >>
			       MLX4_SEG_SHIFT;
		switch (nb_segs_txbb) {
#ifndef NDEBUG
		default:
			/* Should never happen. */
			rte_panic("%p: Invalid number of SGEs(%d) for a TXBB",
			(void *)txq, nb_segs_txbb);
			/* rte_panic never returns. */
			break;
#endif /* NDEBUG */
		case 4:
			/* Memory region key for this memory pool. */
			lkey = mlx4_tx_mb2mr(txq, sbuf);
			if (unlikely(lkey == (uint32_t)-1))
				goto err;
			dseg->addr = rte_cpu_to_be_64(data_addr);
			dseg->lkey = lkey;
			/*
			 * This data segment starts at the beginning of a new
			 * TXBB, so we need to postpone its byte_count writing
			 * for later.
			 */
			pv[*pv_counter].dseg = dseg;
			/*
			 * Zero length segment is treated as inline segment
			 * with zero data.
			 */
			pv[(*pv_counter)++].val =
				rte_cpu_to_be_32(data_len ?
						 data_len :
						 0x80000000);
			if (--nb_segs == 0)
				return ctrl_next;
			/* Prepare next buf info */
			sbuf = sbuf->next;
			dseg++;
			data_len = sbuf->data_len;
			data_addr = rte_pktmbuf_mtod(sbuf, uintptr_t);
			/* fallthrough */
		case 3:
			lkey = mlx4_tx_mb2mr(txq, sbuf);
			if (unlikely(lkey == (uint32_t)-1))
				goto err;
			mlx4_fill_tx_data_seg(dseg, lkey, data_addr,
					rte_cpu_to_be_32(data_len ?
							 data_len :
							 0x80000000));
			if (--nb_segs == 0)
				return ctrl_next;
			/* Prepare next buf info */
			sbuf = sbuf->next;
			dseg++;
			data_len = sbuf->data_len;
			data_addr = rte_pktmbuf_mtod(sbuf, uintptr_t);
			/* fallthrough */
		case 2:
			lkey = mlx4_tx_mb2mr(txq, sbuf);
			if (unlikely(lkey == (uint32_t)-1))
				goto err;
			mlx4_fill_tx_data_seg(dseg, lkey, data_addr,
					rte_cpu_to_be_32(data_len ?
							 data_len :
							 0x80000000));
			if (--nb_segs == 0)
				return ctrl_next;
			/* Prepare next buf info */
			sbuf = sbuf->next;
			dseg++;
			data_len = sbuf->data_len;
			data_addr = rte_pktmbuf_mtod(sbuf, uintptr_t);
			/* fallthrough */
		case 1:
			lkey = mlx4_tx_mb2mr(txq, sbuf);
			if (unlikely(lkey == (uint32_t)-1))
				goto err;
			mlx4_fill_tx_data_seg(dseg, lkey, data_addr,
					rte_cpu_to_be_32(data_len ?
							 data_len :
							 0x80000000));
			if (--nb_segs == 0)
				return ctrl_next;
			/* Prepare next buf info */
			sbuf = sbuf->next;
			dseg++;
			data_len = sbuf->data_len;
			data_addr = rte_pktmbuf_mtod(sbuf, uintptr_t);
			/* fallthrough */
		}
		/* Wrap dseg if it points at the end of the queue. */
		if ((volatile uint8_t *)dseg >= sq->eob)
			dseg = (volatile struct mlx4_wqe_data_seg *)
					((volatile uint8_t *)dseg - sq->size);
	} while (true);
err:
	return NULL;
}

/**
 * Fill the packet's l2, l3 and l4 headers to the WQE.
 *
 * This will be used as the header for each TSO segment that is transmitted.
 *
 * @param buf
 *   Pointer to the first packet mbuf.
 * @param txq
 *   Pointer to Tx queue structure.
 * @param tinfo
 *   Pointer to TSO info to use.
 * @param ctrl
 *   Pointer to the control segment in the TSO WQE.
 *
 * @return
 *   0 on success, negative value upon error.
 */
static inline volatile struct mlx4_wqe_data_seg *
mlx4_tx_burst_fill_tso_hdr(struct rte_mbuf *buf,
			   struct txq *txq,
			   struct tso_info *tinfo,
			   volatile struct mlx4_wqe_ctrl_seg *ctrl)
{
	volatile struct mlx4_wqe_lso_seg *tseg =
		(volatile struct mlx4_wqe_lso_seg *)(ctrl + 1);
	struct mlx4_sq *sq = &txq->msq;
	struct pv *pv = tinfo->pv;
	int *pv_counter = &tinfo->pv_counter;
	int remain_size = tinfo->tso_header_size;
	char *from = rte_pktmbuf_mtod(buf, char *);
	uint16_t txbb_avail_space;
	/* Union to overcome volatile constraints when copying TSO header. */
	union {
		volatile uint8_t *vto;
		uint8_t *to;
	} thdr = { .vto = (volatile uint8_t *)tseg->header, };

	/*
	 * TSO data always starts at offset 20 from the beginning of the TXBB
	 * (16 byte ctrl + 4byte TSO desc). Since each TXBB is 64Byte aligned
	 * we can write the first 44 TSO header bytes without worry for TxQ
	 * wrapping or overwriting the first TXBB 32bit word.
	 */
	txbb_avail_space = MLX4_TXBB_SIZE -
			   (sizeof(struct mlx4_wqe_ctrl_seg) +
			    sizeof(struct mlx4_wqe_lso_seg));
	while (remain_size >= (int)(txbb_avail_space + sizeof(uint32_t))) {
		/* Copy to end of txbb. */
		rte_memcpy(thdr.to, from, txbb_avail_space);
		from += txbb_avail_space;
		thdr.to += txbb_avail_space;
		/* New TXBB, Check for TxQ wrap. */
		if (thdr.to >= sq->eob)
			thdr.vto = sq->buf;
		/* New TXBB, stash the first 32bits for later use. */
		pv[*pv_counter].dst = (volatile uint32_t *)thdr.to;
		pv[(*pv_counter)++].val = *(uint32_t *)from,
		from += sizeof(uint32_t);
		thdr.to += sizeof(uint32_t);
		remain_size -= txbb_avail_space + sizeof(uint32_t);
		/* Avail space in new TXBB is TXBB size - 4 */
		txbb_avail_space = MLX4_TXBB_SIZE - sizeof(uint32_t);
	}
	if (remain_size > txbb_avail_space) {
		rte_memcpy(thdr.to, from, txbb_avail_space);
		from += txbb_avail_space;
		thdr.to += txbb_avail_space;
		remain_size -= txbb_avail_space;
		/* New TXBB, Check for TxQ wrap. */
		if (thdr.to >= sq->eob)
			thdr.vto = sq->buf;
		pv[*pv_counter].dst = (volatile uint32_t *)thdr.to;
		rte_memcpy(&pv[*pv_counter].val, from, remain_size);
		(*pv_counter)++;
	} else if (remain_size) {
		rte_memcpy(thdr.to, from, remain_size);
	}
	tseg->mss_hdr_size = rte_cpu_to_be_32((buf->tso_segsz << 16) |
					      tinfo->tso_header_size);
	/* Calculate data segment location */
	return (volatile struct mlx4_wqe_data_seg *)
				((uintptr_t)tseg + tinfo->wqe_tso_seg_size);
}

/**
 * Write data segments and header for TSO uni/multi segment packet.
 *
 * @param buf
 *   Pointer to the first packet mbuf.
 * @param txq
 *   Pointer to Tx queue structure.
 * @param ctrl
 *   Pointer to the WQE control segment.
 *
 * @return
 *   Pointer to the next WQE control segment on success, NULL otherwise.
 */
static volatile struct mlx4_wqe_ctrl_seg *
mlx4_tx_burst_tso(struct rte_mbuf *buf, struct txq *txq,
		  volatile struct mlx4_wqe_ctrl_seg *ctrl)
{
	volatile struct mlx4_wqe_data_seg *dseg;
	volatile struct mlx4_wqe_ctrl_seg *ctrl_next;
	struct mlx4_sq *sq = &txq->msq;
	struct tso_info tinfo;
	struct pv *pv;
	int pv_counter;
	int ret;

	ret = mlx4_tx_burst_tso_get_params(buf, txq, &tinfo);
	if (unlikely(ret))
		goto error;
	dseg = mlx4_tx_burst_fill_tso_hdr(buf, txq, &tinfo, ctrl);
	if (unlikely(dseg == NULL))
		goto error;
	if ((uintptr_t)dseg >= (uintptr_t)sq->eob)
		dseg = (volatile struct mlx4_wqe_data_seg *)
					((uintptr_t)dseg - sq->size);
	ctrl_next = mlx4_tx_burst_fill_tso_dsegs(buf, txq, &tinfo, dseg, ctrl);
	if (unlikely(ctrl_next == NULL))
		goto error;
	/* Write the first DWORD of each TXBB save earlier. */
	if (likely(tinfo.pv_counter)) {
		pv = tinfo.pv;
		pv_counter = tinfo.pv_counter;
		/* Need a barrier here before writing the first TXBB word. */
		rte_io_wmb();
		do {
			--pv_counter;
			*pv[pv_counter].dst = pv[pv_counter].val;
		} while (pv_counter > 0);
	}
	ctrl->fence_size = tinfo.fence_size;
	sq->remain_size -= tinfo.wqe_size;
	return ctrl_next;
error:
	txq->stats.odropped++;
	return NULL;
}

/**
 * Write data segments of multi-segment packet.
 *
 * @param buf
 *   Pointer to the first packet mbuf.
 * @param txq
 *   Pointer to Tx queue structure.
 * @param ctrl
 *   Pointer to the WQE control segment.
 *
 * @return
 *   Pointer to the next WQE control segment on success, NULL otherwise.
 */
static volatile struct mlx4_wqe_ctrl_seg *
mlx4_tx_burst_segs(struct rte_mbuf *buf, struct txq *txq,
		   volatile struct mlx4_wqe_ctrl_seg *ctrl)
{
	struct pv *pv = (struct pv *)txq->bounce_buf;
	struct mlx4_sq *sq = &txq->msq;
	struct rte_mbuf *sbuf = buf;
	uint32_t lkey;
	int pv_counter = 0;
	int nb_segs = buf->nb_segs;
	uint32_t wqe_size;
	volatile struct mlx4_wqe_data_seg *dseg =
		(volatile struct mlx4_wqe_data_seg *)(ctrl + 1);

	ctrl->fence_size = 1 + nb_segs;
	wqe_size = RTE_ALIGN((uint32_t)(ctrl->fence_size << MLX4_SEG_SHIFT),
			     MLX4_TXBB_SIZE);
	/* Validate WQE size and WQE space in the send queue. */
	if (sq->remain_size < wqe_size ||
	    wqe_size > MLX4_MAX_WQE_SIZE)
		return NULL;
	/*
	 * Fill the data segments with buffer information.
	 * First WQE TXBB head segment is always control segment,
	 * so jump to tail TXBB data segments code for the first
	 * WQE data segments filling.
	 */
	goto txbb_tail_segs;
txbb_head_seg:
	/* Memory region key (big endian) for this memory pool. */
	lkey = mlx4_tx_mb2mr(txq, sbuf);
	if (unlikely(lkey == (uint32_t)-1)) {
		DEBUG("%p: unable to get MP <-> MR association",
		      (void *)txq);
		return NULL;
	}
	/* Handle WQE wraparound. */
	if (dseg >=
		(volatile struct mlx4_wqe_data_seg *)sq->eob)
		dseg = (volatile struct mlx4_wqe_data_seg *)
			sq->buf;
	dseg->addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(sbuf, uintptr_t));
	dseg->lkey = lkey;
	/*
	 * This data segment starts at the beginning of a new
	 * TXBB, so we need to postpone its byte_count writing
	 * for later.
	 */
	pv[pv_counter].dseg = dseg;
	/*
	 * Zero length segment is treated as inline segment
	 * with zero data.
	 */
	pv[pv_counter++].val = rte_cpu_to_be_32(sbuf->data_len ?
						sbuf->data_len : 0x80000000);
	sbuf = sbuf->next;
	dseg++;
	nb_segs--;
txbb_tail_segs:
	/* Jump to default if there are more than two segments remaining. */
	switch (nb_segs) {
	default:
		lkey = mlx4_tx_mb2mr(txq, sbuf);
		if (unlikely(lkey == (uint32_t)-1)) {
			DEBUG("%p: unable to get MP <-> MR association",
			      (void *)txq);
			return NULL;
		}
		mlx4_fill_tx_data_seg(dseg, lkey,
				      rte_pktmbuf_mtod(sbuf, uintptr_t),
				      rte_cpu_to_be_32(sbuf->data_len ?
						       sbuf->data_len :
						       0x80000000));
		sbuf = sbuf->next;
		dseg++;
		nb_segs--;
		/* fallthrough */
	case 2:
		lkey = mlx4_tx_mb2mr(txq, sbuf);
		if (unlikely(lkey == (uint32_t)-1)) {
			DEBUG("%p: unable to get MP <-> MR association",
			      (void *)txq);
			return NULL;
		}
		mlx4_fill_tx_data_seg(dseg, lkey,
				      rte_pktmbuf_mtod(sbuf, uintptr_t),
				      rte_cpu_to_be_32(sbuf->data_len ?
						       sbuf->data_len :
						       0x80000000));
		sbuf = sbuf->next;
		dseg++;
		nb_segs--;
		/* fallthrough */
	case 1:
		lkey = mlx4_tx_mb2mr(txq, sbuf);
		if (unlikely(lkey == (uint32_t)-1)) {
			DEBUG("%p: unable to get MP <-> MR association",
			      (void *)txq);
			return NULL;
		}
		mlx4_fill_tx_data_seg(dseg, lkey,
				      rte_pktmbuf_mtod(sbuf, uintptr_t),
				      rte_cpu_to_be_32(sbuf->data_len ?
						       sbuf->data_len :
						       0x80000000));
		nb_segs--;
		if (nb_segs) {
			sbuf = sbuf->next;
			dseg++;
			goto txbb_head_seg;
		}
		/* fallthrough */
	case 0:
		break;
	}
	/* Write the first DWORD of each TXBB save earlier. */
	if (pv_counter) {
		/* Need a barrier here before writing the byte_count. */
		rte_io_wmb();
		for (--pv_counter; pv_counter  >= 0; pv_counter--)
			pv[pv_counter].dseg->byte_count = pv[pv_counter].val;
	}
	sq->remain_size -= wqe_size;
	/* Align next WQE address to the next TXBB. */
	return (volatile struct mlx4_wqe_ctrl_seg *)
		((volatile uint8_t *)ctrl + wqe_size);
}

/**
 * DPDK callback for Tx.
 *
 * @param dpdk_txq
 *   Generic pointer to Tx queue structure.
 * @param[in] pkts
 *   Packets to transmit.
 * @param pkts_n
 *   Number of packets in array.
 *
 * @return
 *   Number of packets successfully transmitted (<= pkts_n).
 */
uint16_t
mlx4_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
{
	struct txq *txq = (struct txq *)dpdk_txq;
	unsigned int elts_head = txq->elts_head;
	const unsigned int elts_n = txq->elts_n;
	const unsigned int elts_m = elts_n - 1;
	unsigned int bytes_sent = 0;
	unsigned int i;
	unsigned int max = elts_head - txq->elts_tail;
	struct mlx4_sq *sq = &txq->msq;
	volatile struct mlx4_wqe_ctrl_seg *ctrl;
	struct txq_elt *elt;

	assert(txq->elts_comp_cd != 0);
	if (likely(max >= txq->elts_comp_cd_init))
		mlx4_txq_complete(txq, elts_m, sq);
	max = elts_n - max;
	assert(max >= 1);
	assert(max <= elts_n);
	/* Always leave one free entry in the ring. */
	--max;
	if (max > pkts_n)
		max = pkts_n;
	elt = &(*txq->elts)[elts_head & elts_m];
	/* First Tx burst element saves the next WQE control segment. */
	ctrl = elt->wqe;
	for (i = 0; (i != max); ++i) {
		struct rte_mbuf *buf = pkts[i];
		struct txq_elt *elt_next = &(*txq->elts)[++elts_head & elts_m];
		uint32_t owner_opcode = sq->owner_opcode;
		volatile struct mlx4_wqe_data_seg *dseg =
				(volatile struct mlx4_wqe_data_seg *)(ctrl + 1);
		volatile struct mlx4_wqe_ctrl_seg *ctrl_next;
		union {
			uint32_t flags;
			uint16_t flags16[2];
		} srcrb;
		uint32_t lkey;
		bool tso = txq->priv->tso && (buf->ol_flags & PKT_TX_TCP_SEG);

		/* Clean up old buffer. */
		if (likely(elt->buf != NULL)) {
			struct rte_mbuf *tmp = elt->buf;

#ifndef NDEBUG
			/* Poisoning. */
			memset(&elt->buf, 0x66, sizeof(struct rte_mbuf *));
#endif
			/* Faster than rte_pktmbuf_free(). */
			do {
				struct rte_mbuf *next = tmp->next;

				rte_pktmbuf_free_seg(tmp);
				tmp = next;
			} while (tmp != NULL);
		}
		RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
		if (tso) {
			/* Change opcode to TSO */
			owner_opcode &= ~MLX4_OPCODE_CONFIG_CMD;
			owner_opcode |= MLX4_OPCODE_LSO | MLX4_WQE_CTRL_RR;
			ctrl_next = mlx4_tx_burst_tso(buf, txq, ctrl);
			if (!ctrl_next) {
				elt->buf = NULL;
				break;
			}
		} else if (buf->nb_segs == 1) {
			/* Validate WQE space in the send queue. */
			if (sq->remain_size < MLX4_TXBB_SIZE) {
				elt->buf = NULL;
				break;
			}
			lkey = mlx4_tx_mb2mr(txq, buf);
			if (unlikely(lkey == (uint32_t)-1)) {
				/* MR does not exist. */
				DEBUG("%p: unable to get MP <-> MR association",
				      (void *)txq);
				elt->buf = NULL;
				break;
			}
			mlx4_fill_tx_data_seg(dseg++, lkey,
					      rte_pktmbuf_mtod(buf, uintptr_t),
					      rte_cpu_to_be_32(buf->data_len));
			/* Set WQE size in 16-byte units. */
			ctrl->fence_size = 0x2;
			sq->remain_size -= MLX4_TXBB_SIZE;
			/* Align next WQE address to the next TXBB. */
			ctrl_next = ctrl + 0x4;
		} else {
			ctrl_next = mlx4_tx_burst_segs(buf, txq, ctrl);
			if (!ctrl_next) {
				elt->buf = NULL;
				break;
			}
		}
		/* Hold SQ ring wrap around. */
		if ((volatile uint8_t *)ctrl_next >= sq->eob) {
			ctrl_next = (volatile struct mlx4_wqe_ctrl_seg *)
				((volatile uint8_t *)ctrl_next - sq->size);
			/* Flip HW valid ownership. */
			sq->owner_opcode ^= 1u << MLX4_SQ_OWNER_BIT;
		}
		/*
		 * For raw Ethernet, the SOLICIT flag is used to indicate
		 * that no ICRC should be calculated.
		 */
		if (--txq->elts_comp_cd == 0) {
			/* Save the completion burst end address. */
			elt_next->eocb = (volatile uint32_t *)ctrl_next;
			txq->elts_comp_cd = txq->elts_comp_cd_init;
			srcrb.flags = RTE_BE32(MLX4_WQE_CTRL_SOLICIT |
					       MLX4_WQE_CTRL_CQ_UPDATE);
		} else {
			srcrb.flags = RTE_BE32(MLX4_WQE_CTRL_SOLICIT);
		}
		/* Enable HW checksum offload if requested */
		if (txq->csum &&
		    (buf->ol_flags &
		     (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM))) {
			const uint64_t is_tunneled = (buf->ol_flags &
						      (PKT_TX_TUNNEL_GRE |
						       PKT_TX_TUNNEL_VXLAN));

			if (is_tunneled && txq->csum_l2tun) {
				owner_opcode |= MLX4_WQE_CTRL_IIP_HDR_CSUM |
						MLX4_WQE_CTRL_IL4_HDR_CSUM;
				if (buf->ol_flags & PKT_TX_OUTER_IP_CKSUM)
					srcrb.flags |=
					    RTE_BE32(MLX4_WQE_CTRL_IP_HDR_CSUM);
			} else {
				srcrb.flags |=
					RTE_BE32(MLX4_WQE_CTRL_IP_HDR_CSUM |
						MLX4_WQE_CTRL_TCP_UDP_CSUM);
			}
		}
		if (txq->lb) {
			/*
			 * Copy destination MAC address to the WQE, this allows
			 * loopback in eSwitch, so that VFs and PF can
			 * communicate with each other.
			 */
			srcrb.flags16[0] = *(rte_pktmbuf_mtod(buf, uint16_t *));
			ctrl->imm = *(rte_pktmbuf_mtod_offset(buf, uint32_t *,
					      sizeof(uint16_t)));
		} else {
			ctrl->imm = 0;
		}
		ctrl->srcrb_flags = srcrb.flags;
		/*
		 * Make sure descriptor is fully written before
		 * setting ownership bit (because HW can start
		 * executing as soon as we do).
		 */
		rte_io_wmb();
		ctrl->owner_opcode = rte_cpu_to_be_32(owner_opcode);
		elt->buf = buf;
		bytes_sent += buf->pkt_len;
		ctrl = ctrl_next;
		elt = elt_next;
	}
	/* Take a shortcut if nothing must be sent. */
	if (unlikely(i == 0))
		return 0;
	/* Save WQE address of the next Tx burst element. */
	elt->wqe = ctrl;
	/* Increment send statistics counters. */
	txq->stats.opackets += i;
	txq->stats.obytes += bytes_sent;
	/* Make sure that descriptors are written before doorbell record. */
	rte_wmb();
	/* Ring QP doorbell. */
	rte_write32(txq->msq.doorbell_qpn, MLX4_TX_BFREG(txq));
	txq->elts_head += i;
	return i;
}

/**
 * Translate Rx completion flags to packet type.
 *
 * @param[in] cqe
 *   Pointer to CQE.
 *
 * @return
 *   Packet type for struct rte_mbuf.
 */
static inline uint32_t
rxq_cq_to_pkt_type(volatile struct mlx4_cqe *cqe,
		   uint32_t l2tun_offload)
{
	uint8_t idx = 0;
	uint32_t pinfo = rte_be_to_cpu_32(cqe->vlan_my_qpn);
	uint32_t status = rte_be_to_cpu_32(cqe->status);

	/*
	 * The index to the array should have:
	 *  bit[7] - MLX4_CQE_L2_TUNNEL
	 *  bit[6] - MLX4_CQE_L2_TUNNEL_IPV4
	 */
	if (l2tun_offload && (pinfo & MLX4_CQE_L2_TUNNEL))
		idx |= ((pinfo & MLX4_CQE_L2_TUNNEL) >> 20) |
		       ((pinfo & MLX4_CQE_L2_TUNNEL_IPV4) >> 19);
	/*
	 * The index to the array should have:
	 *  bit[5] - MLX4_CQE_STATUS_UDP
	 *  bit[4] - MLX4_CQE_STATUS_TCP
	 *  bit[3] - MLX4_CQE_STATUS_IPV4OPT
	 *  bit[2] - MLX4_CQE_STATUS_IPV6
	 *  bit[1] - MLX4_CQE_STATUS_IPF
	 *  bit[0] - MLX4_CQE_STATUS_IPV4
	 * giving a total of up to 256 entries.
	 */
	idx |= ((status & MLX4_CQE_STATUS_PTYPE_MASK) >> 22);
	if (status & MLX4_CQE_STATUS_IPV6)
		idx |= ((status & MLX4_CQE_STATUS_IPV6F) >> 11);
	return mlx4_ptype_table[idx];
}

/**
 * Translate Rx completion flags to offload flags.
 *
 * @param flags
 *   Rx completion flags returned by mlx4_cqe_flags().
 * @param csum
 *   Whether Rx checksums are enabled.
 * @param csum_l2tun
 *   Whether Rx L2 tunnel checksums are enabled.
 *
 * @return
 *   Offload flags (ol_flags) in mbuf format.
 */
static inline uint32_t
rxq_cq_to_ol_flags(uint32_t flags, int csum, int csum_l2tun)
{
	uint32_t ol_flags = 0;

	if (csum)
		ol_flags |=
			mlx4_transpose(flags,
				       MLX4_CQE_STATUS_IP_HDR_CSUM_OK,
				       PKT_RX_IP_CKSUM_GOOD) |
			mlx4_transpose(flags,
				       MLX4_CQE_STATUS_TCP_UDP_CSUM_OK,
				       PKT_RX_L4_CKSUM_GOOD);
	if ((flags & MLX4_CQE_L2_TUNNEL) && csum_l2tun)
		ol_flags |=
			mlx4_transpose(flags,
				       MLX4_CQE_L2_TUNNEL_IPOK,
				       PKT_RX_IP_CKSUM_GOOD) |
			mlx4_transpose(flags,
				       MLX4_CQE_L2_TUNNEL_L4_CSUM,
				       PKT_RX_L4_CKSUM_GOOD);
	return ol_flags;
}

/**
 * Extract checksum information from CQE flags.
 *
 * @param cqe
 *   Pointer to CQE structure.
 * @param csum
 *   Whether Rx checksums are enabled.
 * @param csum_l2tun
 *   Whether Rx L2 tunnel checksums are enabled.
 *
 * @return
 *   CQE checksum information.
 */
static inline uint32_t
mlx4_cqe_flags(volatile struct mlx4_cqe *cqe, int csum, int csum_l2tun)
{
	uint32_t flags = 0;

	/*
	 * The relevant bits are in different locations on their
	 * CQE fields therefore we can join them in one 32bit
	 * variable.
	 */
	if (csum)
		flags = (rte_be_to_cpu_32(cqe->status) &
			 MLX4_CQE_STATUS_IPV4_CSUM_OK);
	if (csum_l2tun)
		flags |= (rte_be_to_cpu_32(cqe->vlan_my_qpn) &
			  (MLX4_CQE_L2_TUNNEL |
			   MLX4_CQE_L2_TUNNEL_IPOK |
			   MLX4_CQE_L2_TUNNEL_L4_CSUM |
			   MLX4_CQE_L2_TUNNEL_IPV4));
	return flags;
}

/**
 * Poll one CQE from CQ.
 *
 * @param rxq
 *   Pointer to the receive queue structure.
 * @param[out] out
 *   Just polled CQE.
 *
 * @return
 *   Number of bytes of the CQE, 0 in case there is no completion.
 */
static unsigned int
mlx4_cq_poll_one(struct rxq *rxq, volatile struct mlx4_cqe **out)
{
	int ret = 0;
	volatile struct mlx4_cqe *cqe = NULL;
	struct mlx4_cq *cq = &rxq->mcq;

	cqe = (volatile struct mlx4_cqe *)mlx4_get_cqe(cq, cq->cons_index);
	if (!!(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK) ^
	    !!(cq->cons_index & cq->cqe_cnt))
		goto out;
	/*
	 * Make sure we read CQ entry contents after we've checked the
	 * ownership bit.
	 */
	rte_rmb();
	assert(!(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK));
	assert((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) !=
	       MLX4_CQE_OPCODE_ERROR);
	ret = rte_be_to_cpu_32(cqe->byte_cnt);
	++cq->cons_index;
out:
	*out = cqe;
	return ret;
}

/**
 * DPDK callback for Rx with scattered packets support.
 *
 * @param dpdk_rxq
 *   Generic pointer to Rx queue structure.
 * @param[out] pkts
 *   Array to store received packets.
 * @param pkts_n
 *   Maximum number of packets in array.
 *
 * @return
 *   Number of packets successfully received (<= pkts_n).
 */
uint16_t
mlx4_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
{
	struct rxq *rxq = dpdk_rxq;
	const uint32_t wr_cnt = (1 << rxq->elts_n) - 1;
	const uint16_t sges_n = rxq->sges_n;
	struct rte_mbuf *pkt = NULL;
	struct rte_mbuf *seg = NULL;
	unsigned int i = 0;
	uint32_t rq_ci = rxq->rq_ci << sges_n;
	int len = 0;

	while (pkts_n) {
		volatile struct mlx4_cqe *cqe;
		uint32_t idx = rq_ci & wr_cnt;
		struct rte_mbuf *rep = (*rxq->elts)[idx];
		volatile struct mlx4_wqe_data_seg *scat = &(*rxq->wqes)[idx];

		/* Update the 'next' pointer of the previous segment. */
		if (pkt)
			seg->next = rep;
		seg = rep;
		rte_prefetch0(seg);
		rte_prefetch0(scat);
		rep = rte_mbuf_raw_alloc(rxq->mp);
		if (unlikely(rep == NULL)) {
			++rxq->stats.rx_nombuf;
			if (!pkt) {
				/*
				 * No buffers before we even started,
				 * bail out silently.
				 */
				break;
			}
			while (pkt != seg) {
				assert(pkt != (*rxq->elts)[idx]);
				rep = pkt->next;
				pkt->next = NULL;
				pkt->nb_segs = 1;
				rte_mbuf_raw_free(pkt);
				pkt = rep;
			}
			break;
		}
		if (!pkt) {
			/* Looking for the new packet. */
			len = mlx4_cq_poll_one(rxq, &cqe);
			if (!len) {
				rte_mbuf_raw_free(rep);
				break;
			}
			if (unlikely(len < 0)) {
				/* Rx error, packet is likely too large. */
				rte_mbuf_raw_free(rep);
				++rxq->stats.idropped;
				goto skip;
			}
			pkt = seg;
			assert(len >= (rxq->crc_present << 2));
			/* Update packet information. */
			pkt->packet_type =
				rxq_cq_to_pkt_type(cqe, rxq->l2tun_offload);
			pkt->ol_flags = PKT_RX_RSS_HASH;
			pkt->hash.rss = cqe->immed_rss_invalid;
			if (rxq->crc_present)
				len -= ETHER_CRC_LEN;
			pkt->pkt_len = len;
			if (rxq->csum | rxq->csum_l2tun) {
				uint32_t flags =
					mlx4_cqe_flags(cqe,
						       rxq->csum,
						       rxq->csum_l2tun);

				pkt->ol_flags =
					rxq_cq_to_ol_flags(flags,
							   rxq->csum,
							   rxq->csum_l2tun);
			}
		}
		rep->nb_segs = 1;
		rep->port = rxq->port_id;
		rep->data_len = seg->data_len;
		rep->data_off = seg->data_off;
		(*rxq->elts)[idx] = rep;
		/*
		 * Fill NIC descriptor with the new buffer. The lkey and size
		 * of the buffers are already known, only the buffer address
		 * changes.
		 */
		scat->addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(rep, uintptr_t));
		/* If there's only one MR, no need to replace LKey in WQE. */
		if (unlikely(mlx4_mr_btree_len(&rxq->mr_ctrl.cache_bh) > 1))
			scat->lkey = mlx4_rx_mb2mr(rxq, rep);
		if (len > seg->data_len) {
			len -= seg->data_len;
			++pkt->nb_segs;
			++rq_ci;
			continue;
		}
		/* The last segment. */
		seg->data_len = len;
		/* Increment bytes counter. */
		rxq->stats.ibytes += pkt->pkt_len;
		/* Return packet. */
		*(pkts++) = pkt;
		pkt = NULL;
		--pkts_n;
		++i;
skip:
		/* Align consumer index to the next stride. */
		rq_ci >>= sges_n;
		++rq_ci;
		rq_ci <<= sges_n;
	}
	if (unlikely(i == 0 && (rq_ci >> sges_n) == rxq->rq_ci))
		return 0;
	/* Update the consumer index. */
	rxq->rq_ci = rq_ci >> sges_n;
	rte_wmb();
	*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
	*rxq->mcq.set_ci_db =
		rte_cpu_to_be_32(rxq->mcq.cons_index & MLX4_CQ_DB_CI_MASK);
	/* Increment packets counter. */
	rxq->stats.ipackets += i;
	return i;
}

/**
 * Dummy DPDK callback for Tx.
 *
 * This function is used to temporarily replace the real callback during
 * unsafe control operations on the queue, or in case of error.
 *
 * @param dpdk_txq
 *   Generic pointer to Tx queue structure.
 * @param[in] pkts
 *   Packets to transmit.
 * @param pkts_n
 *   Number of packets in array.
 *
 * @return
 *   Number of packets successfully transmitted (<= pkts_n).
 */
uint16_t
mlx4_tx_burst_removed(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
{
	(void)dpdk_txq;
	(void)pkts;
	(void)pkts_n;
	rte_mb();
	return 0;
}

/**
 * Dummy DPDK callback for Rx.
 *
 * This function is used to temporarily replace the real callback during
 * unsafe control operations on the queue, or in case of error.
 *
 * @param dpdk_rxq
 *   Generic pointer to Rx queue structure.
 * @param[out] pkts
 *   Array to store received packets.
 * @param pkts_n
 *   Maximum number of packets in array.
 *
 * @return
 *   Number of packets successfully received (<= pkts_n).
 */
uint16_t
mlx4_rx_burst_removed(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
{
	(void)dpdk_rxq;
	(void)pkts;
	(void)pkts_n;
	rte_mb();
	return 0;
}