xref: /dpdk/drivers/net/mlx5/mlx5_rxtx.c (revision 1cde1b9a9b4dbf31cb5e5ccdfc5da3cb079f43a2)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015-2019 Mellanox Technologies, Ltd
4  */
5 
6 #include <assert.h>
7 #include <stdint.h>
8 #include <string.h>
9 #include <stdlib.h>
10 
11 /* Verbs header. */
12 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
13 #ifdef PEDANTIC
14 #pragma GCC diagnostic ignored "-Wpedantic"
15 #endif
16 #include <infiniband/verbs.h>
17 #include <infiniband/mlx5dv.h>
18 #ifdef PEDANTIC
19 #pragma GCC diagnostic error "-Wpedantic"
20 #endif
21 
22 #include <rte_mbuf.h>
23 #include <rte_mempool.h>
24 #include <rte_prefetch.h>
25 #include <rte_common.h>
26 #include <rte_branch_prediction.h>
27 #include <rte_ether.h>
28 #include <rte_cycles.h>
29 
30 #include "mlx5.h"
31 #include "mlx5_utils.h"
32 #include "mlx5_rxtx.h"
33 #include "mlx5_autoconf.h"
34 #include "mlx5_defs.h"
35 #include "mlx5_prm.h"
36 
37 /* TX burst subroutines return codes. */
38 enum mlx5_txcmp_code {
39 	MLX5_TXCMP_CODE_EXIT = 0,
40 	MLX5_TXCMP_CODE_ERROR,
41 	MLX5_TXCMP_CODE_SINGLE,
42 	MLX5_TXCMP_CODE_MULTI,
43 	MLX5_TXCMP_CODE_TSO,
44 	MLX5_TXCMP_CODE_EMPW,
45 };
46 
47 /*
48  * These defines are used to configure Tx burst routine option set
49  * supported at compile time. The not specified options are optimized out
50  * out due to if conditions can be explicitly calculated at compile time.
51  * The offloads with bigger runtime check (require more CPU cycles to
52  * skip) overhead should have the bigger index - this is needed to
53  * select the better matching routine function if no exact match and
54  * some offloads are not actually requested.
55  */
56 #define MLX5_TXOFF_CONFIG_MULTI (1u << 0) /* Multi-segment packets.*/
57 #define MLX5_TXOFF_CONFIG_TSO (1u << 1) /* TCP send offload supported.*/
58 #define MLX5_TXOFF_CONFIG_SWP (1u << 2) /* Tunnels/SW Parser offloads.*/
59 #define MLX5_TXOFF_CONFIG_CSUM (1u << 3) /* Check Sums offloaded. */
60 #define MLX5_TXOFF_CONFIG_INLINE (1u << 4) /* Data inlining supported. */
61 #define MLX5_TXOFF_CONFIG_VLAN (1u << 5) /* VLAN insertion supported.*/
62 #define MLX5_TXOFF_CONFIG_METADATA (1u << 6) /* Flow metadata. */
63 #define MLX5_TXOFF_CONFIG_EMPW (1u << 8) /* Enhanced MPW supported.*/
64 
65 /* The most common offloads groups. */
66 #define MLX5_TXOFF_CONFIG_NONE 0
67 #define MLX5_TXOFF_CONFIG_FULL (MLX5_TXOFF_CONFIG_MULTI | \
68 				MLX5_TXOFF_CONFIG_TSO | \
69 				MLX5_TXOFF_CONFIG_SWP | \
70 				MLX5_TXOFF_CONFIG_CSUM | \
71 				MLX5_TXOFF_CONFIG_INLINE | \
72 				MLX5_TXOFF_CONFIG_VLAN | \
73 				MLX5_TXOFF_CONFIG_METADATA)
74 
75 #define MLX5_TXOFF_CONFIG(mask) (olx & MLX5_TXOFF_CONFIG_##mask)
76 
77 #define MLX5_TXOFF_DECL(func, olx) \
78 static uint16_t mlx5_tx_burst_##func(void *txq, \
79 				     struct rte_mbuf **pkts, \
80 				    uint16_t pkts_n) \
81 { \
82 	return mlx5_tx_burst_tmpl((struct mlx5_txq_data *)txq, \
83 		    pkts, pkts_n, (olx)); \
84 }
85 
86 #define MLX5_TXOFF_INFO(func, olx) {mlx5_tx_burst_##func, olx},
87 
88 static __rte_always_inline uint32_t
89 rxq_cq_to_pkt_type(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe);
90 
91 static __rte_always_inline int
92 mlx5_rx_poll_len(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
93 		 uint16_t cqe_cnt, volatile struct mlx5_mini_cqe8 **mcqe);
94 
95 static __rte_always_inline uint32_t
96 rxq_cq_to_ol_flags(volatile struct mlx5_cqe *cqe);
97 
98 static __rte_always_inline void
99 rxq_cq_to_mbuf(struct mlx5_rxq_data *rxq, struct rte_mbuf *pkt,
100 	       volatile struct mlx5_cqe *cqe, uint32_t rss_hash_res);
101 
102 static __rte_always_inline void
103 mprq_buf_replace(struct mlx5_rxq_data *rxq, uint16_t rq_idx,
104 		 const unsigned int strd_n);
105 
106 static int
107 mlx5_queue_state_modify(struct rte_eth_dev *dev,
108 			struct mlx5_mp_arg_queue_state_modify *sm);
109 
110 static inline void
111 mlx5_lro_update_tcp_hdr(struct rte_tcp_hdr *restrict tcp,
112 			volatile struct mlx5_cqe *restrict cqe,
113 			uint32_t phcsum);
114 
115 static inline void
116 mlx5_lro_update_hdr(uint8_t *restrict padd,
117 		    volatile struct mlx5_cqe *restrict cqe,
118 		    uint32_t len);
119 
120 uint32_t mlx5_ptype_table[] __rte_cache_aligned = {
121 	[0xff] = RTE_PTYPE_ALL_MASK, /* Last entry for errored packet. */
122 };
123 
124 uint8_t mlx5_cksum_table[1 << 10] __rte_cache_aligned;
125 uint8_t mlx5_swp_types_table[1 << 10] __rte_cache_aligned;
126 
127 /**
128  * Build a table to translate Rx completion flags to packet type.
129  *
130  * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
131  */
132 void
133 mlx5_set_ptype_table(void)
134 {
135 	unsigned int i;
136 	uint32_t (*p)[RTE_DIM(mlx5_ptype_table)] = &mlx5_ptype_table;
137 
138 	/* Last entry must not be overwritten, reserved for errored packet. */
139 	for (i = 0; i < RTE_DIM(mlx5_ptype_table) - 1; ++i)
140 		(*p)[i] = RTE_PTYPE_UNKNOWN;
141 	/*
142 	 * The index to the array should have:
143 	 * bit[1:0] = l3_hdr_type
144 	 * bit[4:2] = l4_hdr_type
145 	 * bit[5] = ip_frag
146 	 * bit[6] = tunneled
147 	 * bit[7] = outer_l3_type
148 	 */
149 	/* L2 */
150 	(*p)[0x00] = RTE_PTYPE_L2_ETHER;
151 	/* L3 */
152 	(*p)[0x01] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
153 		     RTE_PTYPE_L4_NONFRAG;
154 	(*p)[0x02] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
155 		     RTE_PTYPE_L4_NONFRAG;
156 	/* Fragmented */
157 	(*p)[0x21] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
158 		     RTE_PTYPE_L4_FRAG;
159 	(*p)[0x22] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
160 		     RTE_PTYPE_L4_FRAG;
161 	/* TCP */
162 	(*p)[0x05] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
163 		     RTE_PTYPE_L4_TCP;
164 	(*p)[0x06] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
165 		     RTE_PTYPE_L4_TCP;
166 	(*p)[0x0d] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
167 		     RTE_PTYPE_L4_TCP;
168 	(*p)[0x0e] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
169 		     RTE_PTYPE_L4_TCP;
170 	(*p)[0x11] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
171 		     RTE_PTYPE_L4_TCP;
172 	(*p)[0x12] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
173 		     RTE_PTYPE_L4_TCP;
174 	/* UDP */
175 	(*p)[0x09] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
176 		     RTE_PTYPE_L4_UDP;
177 	(*p)[0x0a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
178 		     RTE_PTYPE_L4_UDP;
179 	/* Repeat with outer_l3_type being set. Just in case. */
180 	(*p)[0x81] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
181 		     RTE_PTYPE_L4_NONFRAG;
182 	(*p)[0x82] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
183 		     RTE_PTYPE_L4_NONFRAG;
184 	(*p)[0xa1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
185 		     RTE_PTYPE_L4_FRAG;
186 	(*p)[0xa2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
187 		     RTE_PTYPE_L4_FRAG;
188 	(*p)[0x85] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
189 		     RTE_PTYPE_L4_TCP;
190 	(*p)[0x86] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
191 		     RTE_PTYPE_L4_TCP;
192 	(*p)[0x8d] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
193 		     RTE_PTYPE_L4_TCP;
194 	(*p)[0x8e] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
195 		     RTE_PTYPE_L4_TCP;
196 	(*p)[0x91] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
197 		     RTE_PTYPE_L4_TCP;
198 	(*p)[0x92] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
199 		     RTE_PTYPE_L4_TCP;
200 	(*p)[0x89] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
201 		     RTE_PTYPE_L4_UDP;
202 	(*p)[0x8a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
203 		     RTE_PTYPE_L4_UDP;
204 	/* Tunneled - L3 */
205 	(*p)[0x40] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN;
206 	(*p)[0x41] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
207 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
208 		     RTE_PTYPE_INNER_L4_NONFRAG;
209 	(*p)[0x42] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
210 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
211 		     RTE_PTYPE_INNER_L4_NONFRAG;
212 	(*p)[0xc0] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN;
213 	(*p)[0xc1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
214 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
215 		     RTE_PTYPE_INNER_L4_NONFRAG;
216 	(*p)[0xc2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
217 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
218 		     RTE_PTYPE_INNER_L4_NONFRAG;
219 	/* Tunneled - Fragmented */
220 	(*p)[0x61] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
221 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
222 		     RTE_PTYPE_INNER_L4_FRAG;
223 	(*p)[0x62] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
224 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
225 		     RTE_PTYPE_INNER_L4_FRAG;
226 	(*p)[0xe1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
227 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
228 		     RTE_PTYPE_INNER_L4_FRAG;
229 	(*p)[0xe2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
230 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
231 		     RTE_PTYPE_INNER_L4_FRAG;
232 	/* Tunneled - TCP */
233 	(*p)[0x45] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
234 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
235 		     RTE_PTYPE_INNER_L4_TCP;
236 	(*p)[0x46] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
237 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
238 		     RTE_PTYPE_INNER_L4_TCP;
239 	(*p)[0x4d] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
240 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
241 		     RTE_PTYPE_INNER_L4_TCP;
242 	(*p)[0x4e] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
243 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
244 		     RTE_PTYPE_INNER_L4_TCP;
245 	(*p)[0x51] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
246 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
247 		     RTE_PTYPE_INNER_L4_TCP;
248 	(*p)[0x52] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
249 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
250 		     RTE_PTYPE_INNER_L4_TCP;
251 	(*p)[0xc5] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
252 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
253 		     RTE_PTYPE_INNER_L4_TCP;
254 	(*p)[0xc6] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
255 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
256 		     RTE_PTYPE_INNER_L4_TCP;
257 	(*p)[0xcd] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
258 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
259 		     RTE_PTYPE_INNER_L4_TCP;
260 	(*p)[0xce] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
261 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
262 		     RTE_PTYPE_INNER_L4_TCP;
263 	(*p)[0xd1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
264 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
265 		     RTE_PTYPE_INNER_L4_TCP;
266 	(*p)[0xd2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
267 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
268 		     RTE_PTYPE_INNER_L4_TCP;
269 	/* Tunneled - UDP */
270 	(*p)[0x49] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
271 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
272 		     RTE_PTYPE_INNER_L4_UDP;
273 	(*p)[0x4a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
274 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
275 		     RTE_PTYPE_INNER_L4_UDP;
276 	(*p)[0xc9] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
277 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
278 		     RTE_PTYPE_INNER_L4_UDP;
279 	(*p)[0xca] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
280 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
281 		     RTE_PTYPE_INNER_L4_UDP;
282 }
283 
284 /**
285  * Build a table to translate packet to checksum type of Verbs.
286  */
287 void
288 mlx5_set_cksum_table(void)
289 {
290 	unsigned int i;
291 	uint8_t v;
292 
293 	/*
294 	 * The index should have:
295 	 * bit[0] = PKT_TX_TCP_SEG
296 	 * bit[2:3] = PKT_TX_UDP_CKSUM, PKT_TX_TCP_CKSUM
297 	 * bit[4] = PKT_TX_IP_CKSUM
298 	 * bit[8] = PKT_TX_OUTER_IP_CKSUM
299 	 * bit[9] = tunnel
300 	 */
301 	for (i = 0; i < RTE_DIM(mlx5_cksum_table); ++i) {
302 		v = 0;
303 		if (i & (1 << 9)) {
304 			/* Tunneled packet. */
305 			if (i & (1 << 8)) /* Outer IP. */
306 				v |= MLX5_ETH_WQE_L3_CSUM;
307 			if (i & (1 << 4)) /* Inner IP. */
308 				v |= MLX5_ETH_WQE_L3_INNER_CSUM;
309 			if (i & (3 << 2 | 1 << 0)) /* L4 or TSO. */
310 				v |= MLX5_ETH_WQE_L4_INNER_CSUM;
311 		} else {
312 			/* No tunnel. */
313 			if (i & (1 << 4)) /* IP. */
314 				v |= MLX5_ETH_WQE_L3_CSUM;
315 			if (i & (3 << 2 | 1 << 0)) /* L4 or TSO. */
316 				v |= MLX5_ETH_WQE_L4_CSUM;
317 		}
318 		mlx5_cksum_table[i] = v;
319 	}
320 }
321 
322 /**
323  * Build a table to translate packet type of mbuf to SWP type of Verbs.
324  */
325 void
326 mlx5_set_swp_types_table(void)
327 {
328 	unsigned int i;
329 	uint8_t v;
330 
331 	/*
332 	 * The index should have:
333 	 * bit[0:1] = PKT_TX_L4_MASK
334 	 * bit[4] = PKT_TX_IPV6
335 	 * bit[8] = PKT_TX_OUTER_IPV6
336 	 * bit[9] = PKT_TX_OUTER_UDP
337 	 */
338 	for (i = 0; i < RTE_DIM(mlx5_swp_types_table); ++i) {
339 		v = 0;
340 		if (i & (1 << 8))
341 			v |= MLX5_ETH_WQE_L3_OUTER_IPV6;
342 		if (i & (1 << 9))
343 			v |= MLX5_ETH_WQE_L4_OUTER_UDP;
344 		if (i & (1 << 4))
345 			v |= MLX5_ETH_WQE_L3_INNER_IPV6;
346 		if ((i & 3) == (PKT_TX_UDP_CKSUM >> 52))
347 			v |= MLX5_ETH_WQE_L4_INNER_UDP;
348 		mlx5_swp_types_table[i] = v;
349 	}
350 }
351 
352 /**
353  * Set Software Parser flags and offsets in Ethernet Segment of WQE.
354  * Flags must be preliminary initialized to zero.
355  *
356  * @param loc
357  *   Pointer to burst routine local context.
358  * @param swp_flags
359  *   Pointer to store Software Parser flags
360  * @param olx
361  *   Configured Tx offloads mask. It is fully defined at
362  *   compile time and may be used for optimization.
363  *
364  * @return
365  *   Software Parser offsets packed in dword.
366  *   Software Parser flags are set by pointer.
367  */
368 static __rte_always_inline uint32_t
369 txq_mbuf_to_swp(struct mlx5_txq_local *restrict loc,
370 		uint8_t *swp_flags,
371 		unsigned int olx)
372 {
373 	uint64_t ol, tunnel;
374 	unsigned int idx, off;
375 	uint32_t set;
376 
377 	if (!MLX5_TXOFF_CONFIG(SWP))
378 		return 0;
379 	ol = loc->mbuf->ol_flags;
380 	tunnel = ol & PKT_TX_TUNNEL_MASK;
381 	/*
382 	 * Check whether Software Parser is required.
383 	 * Only customized tunnels may ask for.
384 	 */
385 	if (likely(tunnel != PKT_TX_TUNNEL_UDP && tunnel != PKT_TX_TUNNEL_IP))
386 		return 0;
387 	/*
388 	 * The index should have:
389 	 * bit[0:1] = PKT_TX_L4_MASK
390 	 * bit[4] = PKT_TX_IPV6
391 	 * bit[8] = PKT_TX_OUTER_IPV6
392 	 * bit[9] = PKT_TX_OUTER_UDP
393 	 */
394 	idx = (ol & (PKT_TX_L4_MASK | PKT_TX_IPV6 | PKT_TX_OUTER_IPV6)) >> 52;
395 	idx |= (tunnel == PKT_TX_TUNNEL_UDP) ? (1 << 9) : 0;
396 	*swp_flags = mlx5_swp_types_table[idx];
397 	/*
398 	 * Set offsets for SW parser. Since ConnectX-5, SW parser just
399 	 * complements HW parser. SW parser starts to engage only if HW parser
400 	 * can't reach a header. For the older devices, HW parser will not kick
401 	 * in if any of SWP offsets is set. Therefore, all of the L3 offsets
402 	 * should be set regardless of HW offload.
403 	 */
404 	off = loc->mbuf->outer_l2_len;
405 	if (MLX5_TXOFF_CONFIG(VLAN) && ol & PKT_TX_VLAN_PKT)
406 		off += sizeof(struct rte_vlan_hdr);
407 	set = (off >> 1) << 8; /* Outer L3 offset. */
408 	off += loc->mbuf->outer_l3_len;
409 	if (tunnel == PKT_TX_TUNNEL_UDP)
410 		set |= off >> 1; /* Outer L4 offset. */
411 	if (ol & (PKT_TX_IPV4 | PKT_TX_IPV6)) { /* Inner IP. */
412 		const uint64_t csum = ol & PKT_TX_L4_MASK;
413 			off += loc->mbuf->l2_len;
414 		set |= (off >> 1) << 24; /* Inner L3 offset. */
415 		if (csum == PKT_TX_TCP_CKSUM ||
416 		    csum == PKT_TX_UDP_CKSUM ||
417 		    (MLX5_TXOFF_CONFIG(TSO) && ol & PKT_TX_TCP_SEG)) {
418 			off += loc->mbuf->l3_len;
419 			set |= (off >> 1) << 16; /* Inner L4 offset. */
420 		}
421 	}
422 	set = rte_cpu_to_le_32(set);
423 	return set;
424 }
425 
426 /**
427  * Convert the Checksum offloads to Verbs.
428  *
429  * @param buf
430  *   Pointer to the mbuf.
431  *
432  * @return
433  *   Converted checksum flags.
434  */
435 static __rte_always_inline uint8_t
436 txq_ol_cksum_to_cs(struct rte_mbuf *buf)
437 {
438 	uint32_t idx;
439 	uint8_t is_tunnel = !!(buf->ol_flags & PKT_TX_TUNNEL_MASK);
440 	const uint64_t ol_flags_mask = PKT_TX_TCP_SEG | PKT_TX_L4_MASK |
441 				       PKT_TX_IP_CKSUM | PKT_TX_OUTER_IP_CKSUM;
442 
443 	/*
444 	 * The index should have:
445 	 * bit[0] = PKT_TX_TCP_SEG
446 	 * bit[2:3] = PKT_TX_UDP_CKSUM, PKT_TX_TCP_CKSUM
447 	 * bit[4] = PKT_TX_IP_CKSUM
448 	 * bit[8] = PKT_TX_OUTER_IP_CKSUM
449 	 * bit[9] = tunnel
450 	 */
451 	idx = ((buf->ol_flags & ol_flags_mask) >> 50) | (!!is_tunnel << 9);
452 	return mlx5_cksum_table[idx];
453 }
454 
455 /**
456  * Internal function to compute the number of used descriptors in an RX queue
457  *
458  * @param rxq
459  *   The Rx queue.
460  *
461  * @return
462  *   The number of used rx descriptor.
463  */
464 static uint32_t
465 rx_queue_count(struct mlx5_rxq_data *rxq)
466 {
467 	struct rxq_zip *zip = &rxq->zip;
468 	volatile struct mlx5_cqe *cqe;
469 	const unsigned int cqe_n = (1 << rxq->cqe_n);
470 	const unsigned int cqe_cnt = cqe_n - 1;
471 	unsigned int cq_ci;
472 	unsigned int used;
473 
474 	/* if we are processing a compressed cqe */
475 	if (zip->ai) {
476 		used = zip->cqe_cnt - zip->ca;
477 		cq_ci = zip->cq_ci;
478 	} else {
479 		used = 0;
480 		cq_ci = rxq->cq_ci;
481 	}
482 	cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
483 	while (check_cqe(cqe, cqe_n, cq_ci) != MLX5_CQE_STATUS_HW_OWN) {
484 		int8_t op_own;
485 		unsigned int n;
486 
487 		op_own = cqe->op_own;
488 		if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED)
489 			n = rte_be_to_cpu_32(cqe->byte_cnt);
490 		else
491 			n = 1;
492 		cq_ci += n;
493 		used += n;
494 		cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
495 	}
496 	used = RTE_MIN(used, (1U << rxq->elts_n) - 1);
497 	return used;
498 }
499 
500 /**
501  * DPDK callback to check the status of a rx descriptor.
502  *
503  * @param rx_queue
504  *   The Rx queue.
505  * @param[in] offset
506  *   The index of the descriptor in the ring.
507  *
508  * @return
509  *   The status of the tx descriptor.
510  */
511 int
512 mlx5_rx_descriptor_status(void *rx_queue, uint16_t offset)
513 {
514 	struct mlx5_rxq_data *rxq = rx_queue;
515 	struct mlx5_rxq_ctrl *rxq_ctrl =
516 			container_of(rxq, struct mlx5_rxq_ctrl, rxq);
517 	struct rte_eth_dev *dev = ETH_DEV(rxq_ctrl->priv);
518 
519 	if (dev->rx_pkt_burst != mlx5_rx_burst) {
520 		rte_errno = ENOTSUP;
521 		return -rte_errno;
522 	}
523 	if (offset >= (1 << rxq->elts_n)) {
524 		rte_errno = EINVAL;
525 		return -rte_errno;
526 	}
527 	if (offset < rx_queue_count(rxq))
528 		return RTE_ETH_RX_DESC_DONE;
529 	return RTE_ETH_RX_DESC_AVAIL;
530 }
531 
532 /**
533  * DPDK callback to get the number of used descriptors in a RX queue
534  *
535  * @param dev
536  *   Pointer to the device structure.
537  *
538  * @param rx_queue_id
539  *   The Rx queue.
540  *
541  * @return
542  *   The number of used rx descriptor.
543  *   -EINVAL if the queue is invalid
544  */
545 uint32_t
546 mlx5_rx_queue_count(struct rte_eth_dev *dev, uint16_t rx_queue_id)
547 {
548 	struct mlx5_priv *priv = dev->data->dev_private;
549 	struct mlx5_rxq_data *rxq;
550 
551 	if (dev->rx_pkt_burst != mlx5_rx_burst) {
552 		rte_errno = ENOTSUP;
553 		return -rte_errno;
554 	}
555 	rxq = (*priv->rxqs)[rx_queue_id];
556 	if (!rxq) {
557 		rte_errno = EINVAL;
558 		return -rte_errno;
559 	}
560 	return rx_queue_count(rxq);
561 }
562 
563 #define MLX5_SYSTEM_LOG_DIR "/var/log"
564 /**
565  * Dump debug information to log file.
566  *
567  * @param fname
568  *   The file name.
569  * @param hex_title
570  *   If not NULL this string is printed as a header to the output
571  *   and the output will be in hexadecimal view.
572  * @param buf
573  *   This is the buffer address to print out.
574  * @param len
575  *   The number of bytes to dump out.
576  */
577 void
578 mlx5_dump_debug_information(const char *fname, const char *hex_title,
579 			    const void *buf, unsigned int hex_len)
580 {
581 	FILE *fd;
582 
583 	MKSTR(path, "%s/%s", MLX5_SYSTEM_LOG_DIR, fname);
584 	fd = fopen(path, "a+");
585 	if (!fd) {
586 		DRV_LOG(WARNING, "cannot open %s for debug dump\n",
587 			path);
588 		MKSTR(path2, "./%s", fname);
589 		fd = fopen(path2, "a+");
590 		if (!fd) {
591 			DRV_LOG(ERR, "cannot open %s for debug dump\n",
592 				path2);
593 			return;
594 		}
595 		DRV_LOG(INFO, "New debug dump in file %s\n", path2);
596 	} else {
597 		DRV_LOG(INFO, "New debug dump in file %s\n", path);
598 	}
599 	if (hex_title)
600 		rte_hexdump(fd, hex_title, buf, hex_len);
601 	else
602 		fprintf(fd, "%s", (const char *)buf);
603 	fprintf(fd, "\n\n\n");
604 	fclose(fd);
605 }
606 
607 /**
608  * Move QP from error state to running state and initialize indexes.
609  *
610  * @param txq_ctrl
611  *   Pointer to TX queue control structure.
612  *
613  * @return
614  *   0 on success, else -1.
615  */
616 static int
617 tx_recover_qp(struct mlx5_txq_ctrl *txq_ctrl)
618 {
619 	struct mlx5_mp_arg_queue_state_modify sm = {
620 			.is_wq = 0,
621 			.queue_id = txq_ctrl->txq.idx,
622 	};
623 
624 	if (mlx5_queue_state_modify(ETH_DEV(txq_ctrl->priv), &sm))
625 		return -1;
626 	txq_ctrl->txq.wqe_ci = 0;
627 	txq_ctrl->txq.wqe_pi = 0;
628 	txq_ctrl->txq.elts_comp = 0;
629 	return 0;
630 }
631 
632 /* Return 1 if the error CQE is signed otherwise, sign it and return 0. */
633 static int
634 check_err_cqe_seen(volatile struct mlx5_err_cqe *err_cqe)
635 {
636 	static const uint8_t magic[] = "seen";
637 	int ret = 1;
638 	unsigned int i;
639 
640 	for (i = 0; i < sizeof(magic); ++i)
641 		if (!ret || err_cqe->rsvd1[i] != magic[i]) {
642 			ret = 0;
643 			err_cqe->rsvd1[i] = magic[i];
644 		}
645 	return ret;
646 }
647 
648 /**
649  * Handle error CQE.
650  *
651  * @param txq
652  *   Pointer to TX queue structure.
653  * @param error_cqe
654  *   Pointer to the error CQE.
655  *
656  * @return
657  *   Negative value if queue recovery failed,
658  *   the last Tx buffer element to free otherwise.
659  */
660 int
661 mlx5_tx_error_cqe_handle(struct mlx5_txq_data *restrict txq,
662 			 volatile struct mlx5_err_cqe *err_cqe)
663 {
664 	if (err_cqe->syndrome != MLX5_CQE_SYNDROME_WR_FLUSH_ERR) {
665 		const uint16_t wqe_m = ((1 << txq->wqe_n) - 1);
666 		struct mlx5_txq_ctrl *txq_ctrl =
667 				container_of(txq, struct mlx5_txq_ctrl, txq);
668 		uint16_t new_wqe_pi = rte_be_to_cpu_16(err_cqe->wqe_counter);
669 		int seen = check_err_cqe_seen(err_cqe);
670 
671 		if (!seen && txq_ctrl->dump_file_n <
672 		    txq_ctrl->priv->config.max_dump_files_num) {
673 			MKSTR(err_str, "Unexpected CQE error syndrome "
674 			      "0x%02x CQN = %u SQN = %u wqe_counter = %u "
675 			      "wq_ci = %u cq_ci = %u", err_cqe->syndrome,
676 			      txq->cqe_s, txq->qp_num_8s >> 8,
677 			      rte_be_to_cpu_16(err_cqe->wqe_counter),
678 			      txq->wqe_ci, txq->cq_ci);
679 			MKSTR(name, "dpdk_mlx5_port_%u_txq_%u_index_%u_%u",
680 			      PORT_ID(txq_ctrl->priv), txq->idx,
681 			      txq_ctrl->dump_file_n, (uint32_t)rte_rdtsc());
682 			mlx5_dump_debug_information(name, NULL, err_str, 0);
683 			mlx5_dump_debug_information(name, "MLX5 Error CQ:",
684 						    (const void *)((uintptr_t)
685 						    txq->cqes),
686 						    sizeof(*err_cqe) *
687 						    (1 << txq->cqe_n));
688 			mlx5_dump_debug_information(name, "MLX5 Error SQ:",
689 						    (const void *)((uintptr_t)
690 						    txq->wqes),
691 						    MLX5_WQE_SIZE *
692 						    (1 << txq->wqe_n));
693 			txq_ctrl->dump_file_n++;
694 		}
695 		if (!seen)
696 			/*
697 			 * Count errors in WQEs units.
698 			 * Later it can be improved to count error packets,
699 			 * for example, by SQ parsing to find how much packets
700 			 * should be counted for each WQE.
701 			 */
702 			txq->stats.oerrors += ((txq->wqe_ci & wqe_m) -
703 						new_wqe_pi) & wqe_m;
704 		if (tx_recover_qp(txq_ctrl) == 0) {
705 			txq->cq_ci++;
706 			/* Release all the remaining buffers. */
707 			return txq->elts_head;
708 		}
709 		/* Recovering failed - try again later on the same WQE. */
710 		return -1;
711 	} else {
712 		txq->cq_ci++;
713 	}
714 	/* Do not release buffers. */
715 	return txq->elts_tail;
716 }
717 
718 /**
719  * Translate RX completion flags to packet type.
720  *
721  * @param[in] rxq
722  *   Pointer to RX queue structure.
723  * @param[in] cqe
724  *   Pointer to CQE.
725  *
726  * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
727  *
728  * @return
729  *   Packet type for struct rte_mbuf.
730  */
731 static inline uint32_t
732 rxq_cq_to_pkt_type(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe)
733 {
734 	uint8_t idx;
735 	uint8_t pinfo = cqe->pkt_info;
736 	uint16_t ptype = cqe->hdr_type_etc;
737 
738 	/*
739 	 * The index to the array should have:
740 	 * bit[1:0] = l3_hdr_type
741 	 * bit[4:2] = l4_hdr_type
742 	 * bit[5] = ip_frag
743 	 * bit[6] = tunneled
744 	 * bit[7] = outer_l3_type
745 	 */
746 	idx = ((pinfo & 0x3) << 6) | ((ptype & 0xfc00) >> 10);
747 	return mlx5_ptype_table[idx] | rxq->tunnel * !!(idx & (1 << 6));
748 }
749 
750 /**
751  * Initialize Rx WQ and indexes.
752  *
753  * @param[in] rxq
754  *   Pointer to RX queue structure.
755  */
756 void
757 mlx5_rxq_initialize(struct mlx5_rxq_data *rxq)
758 {
759 	const unsigned int wqe_n = 1 << rxq->elts_n;
760 	unsigned int i;
761 
762 	for (i = 0; (i != wqe_n); ++i) {
763 		volatile struct mlx5_wqe_data_seg *scat;
764 		uintptr_t addr;
765 		uint32_t byte_count;
766 
767 		if (mlx5_rxq_mprq_enabled(rxq)) {
768 			struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[i];
769 
770 			scat = &((volatile struct mlx5_wqe_mprq *)
771 				rxq->wqes)[i].dseg;
772 			addr = (uintptr_t)mlx5_mprq_buf_addr(buf,
773 							 1 << rxq->strd_num_n);
774 			byte_count = (1 << rxq->strd_sz_n) *
775 					(1 << rxq->strd_num_n);
776 		} else {
777 			struct rte_mbuf *buf = (*rxq->elts)[i];
778 
779 			scat = &((volatile struct mlx5_wqe_data_seg *)
780 					rxq->wqes)[i];
781 			addr = rte_pktmbuf_mtod(buf, uintptr_t);
782 			byte_count = DATA_LEN(buf);
783 		}
784 		/* scat->addr must be able to store a pointer. */
785 		assert(sizeof(scat->addr) >= sizeof(uintptr_t));
786 		*scat = (struct mlx5_wqe_data_seg){
787 			.addr = rte_cpu_to_be_64(addr),
788 			.byte_count = rte_cpu_to_be_32(byte_count),
789 			.lkey = mlx5_rx_addr2mr(rxq, addr),
790 		};
791 	}
792 	rxq->consumed_strd = 0;
793 	rxq->decompressed = 0;
794 	rxq->rq_pi = 0;
795 	rxq->zip = (struct rxq_zip){
796 		.ai = 0,
797 	};
798 	/* Update doorbell counter. */
799 	rxq->rq_ci = wqe_n >> rxq->sges_n;
800 	rte_cio_wmb();
801 	*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
802 }
803 
804 /**
805  * Modify a Verbs/DevX queue state.
806  * This must be called from the primary process.
807  *
808  * @param dev
809  *   Pointer to Ethernet device.
810  * @param sm
811  *   State modify request parameters.
812  *
813  * @return
814  *   0 in case of success else non-zero value and rte_errno is set.
815  */
816 int
817 mlx5_queue_state_modify_primary(struct rte_eth_dev *dev,
818 			const struct mlx5_mp_arg_queue_state_modify *sm)
819 {
820 	int ret;
821 	struct mlx5_priv *priv = dev->data->dev_private;
822 
823 	if (sm->is_wq) {
824 		struct mlx5_rxq_data *rxq = (*priv->rxqs)[sm->queue_id];
825 		struct mlx5_rxq_ctrl *rxq_ctrl =
826 			container_of(rxq, struct mlx5_rxq_ctrl, rxq);
827 
828 		if (rxq_ctrl->obj->type == MLX5_RXQ_OBJ_TYPE_IBV) {
829 			struct ibv_wq_attr mod = {
830 				.attr_mask = IBV_WQ_ATTR_STATE,
831 				.wq_state = sm->state,
832 			};
833 
834 			ret = mlx5_glue->modify_wq(rxq_ctrl->obj->wq, &mod);
835 		} else { /* rxq_ctrl->obj->type == MLX5_RXQ_OBJ_TYPE_DEVX_RQ. */
836 			struct mlx5_devx_modify_rq_attr rq_attr;
837 
838 			memset(&rq_attr, 0, sizeof(rq_attr));
839 			if (sm->state == IBV_WQS_RESET) {
840 				rq_attr.rq_state = MLX5_RQC_STATE_ERR;
841 				rq_attr.state = MLX5_RQC_STATE_RST;
842 			} else if (sm->state == IBV_WQS_RDY) {
843 				rq_attr.rq_state = MLX5_RQC_STATE_RST;
844 				rq_attr.state = MLX5_RQC_STATE_RDY;
845 			} else if (sm->state == IBV_WQS_ERR) {
846 				rq_attr.rq_state = MLX5_RQC_STATE_RDY;
847 				rq_attr.state = MLX5_RQC_STATE_ERR;
848 			}
849 			ret = mlx5_devx_cmd_modify_rq(rxq_ctrl->obj->rq,
850 						      &rq_attr);
851 		}
852 		if (ret) {
853 			DRV_LOG(ERR, "Cannot change Rx WQ state to %u  - %s\n",
854 					sm->state, strerror(errno));
855 			rte_errno = errno;
856 			return ret;
857 		}
858 	} else {
859 		struct mlx5_txq_data *txq = (*priv->txqs)[sm->queue_id];
860 		struct mlx5_txq_ctrl *txq_ctrl =
861 			container_of(txq, struct mlx5_txq_ctrl, txq);
862 		struct ibv_qp_attr mod = {
863 			.qp_state = IBV_QPS_RESET,
864 			.port_num = (uint8_t)priv->ibv_port,
865 		};
866 		struct ibv_qp *qp = txq_ctrl->ibv->qp;
867 
868 		ret = mlx5_glue->modify_qp(qp, &mod, IBV_QP_STATE);
869 		if (ret) {
870 			DRV_LOG(ERR, "Cannot change the Tx QP state to RESET "
871 				"%s\n", strerror(errno));
872 			rte_errno = errno;
873 			return ret;
874 		}
875 		mod.qp_state = IBV_QPS_INIT;
876 		ret = mlx5_glue->modify_qp(qp, &mod,
877 					   (IBV_QP_STATE | IBV_QP_PORT));
878 		if (ret) {
879 			DRV_LOG(ERR, "Cannot change Tx QP state to INIT %s\n",
880 				strerror(errno));
881 			rte_errno = errno;
882 			return ret;
883 		}
884 		mod.qp_state = IBV_QPS_RTR;
885 		ret = mlx5_glue->modify_qp(qp, &mod, IBV_QP_STATE);
886 		if (ret) {
887 			DRV_LOG(ERR, "Cannot change Tx QP state to RTR %s\n",
888 				strerror(errno));
889 			rte_errno = errno;
890 			return ret;
891 		}
892 		mod.qp_state = IBV_QPS_RTS;
893 		ret = mlx5_glue->modify_qp(qp, &mod, IBV_QP_STATE);
894 		if (ret) {
895 			DRV_LOG(ERR, "Cannot change Tx QP state to RTS %s\n",
896 				strerror(errno));
897 			rte_errno = errno;
898 			return ret;
899 		}
900 	}
901 	return 0;
902 }
903 
904 /**
905  * Modify a Verbs queue state.
906  *
907  * @param dev
908  *   Pointer to Ethernet device.
909  * @param sm
910  *   State modify request parameters.
911  *
912  * @return
913  *   0 in case of success else non-zero value.
914  */
915 static int
916 mlx5_queue_state_modify(struct rte_eth_dev *dev,
917 			struct mlx5_mp_arg_queue_state_modify *sm)
918 {
919 	int ret = 0;
920 
921 	switch (rte_eal_process_type()) {
922 	case RTE_PROC_PRIMARY:
923 		ret = mlx5_queue_state_modify_primary(dev, sm);
924 		break;
925 	case RTE_PROC_SECONDARY:
926 		ret = mlx5_mp_req_queue_state_modify(dev, sm);
927 		break;
928 	default:
929 		break;
930 	}
931 	return ret;
932 }
933 
934 /**
935  * Handle a Rx error.
936  * The function inserts the RQ state to reset when the first error CQE is
937  * shown, then drains the CQ by the caller function loop. When the CQ is empty,
938  * it moves the RQ state to ready and initializes the RQ.
939  * Next CQE identification and error counting are in the caller responsibility.
940  *
941  * @param[in] rxq
942  *   Pointer to RX queue structure.
943  * @param[in] mbuf_prepare
944  *   Whether to prepare mbufs for the RQ.
945  *
946  * @return
947  *   -1 in case of recovery error, otherwise the CQE status.
948  */
949 int
950 mlx5_rx_err_handle(struct mlx5_rxq_data *rxq, uint8_t mbuf_prepare)
951 {
952 	const uint16_t cqe_n = 1 << rxq->cqe_n;
953 	const uint16_t cqe_mask = cqe_n - 1;
954 	const unsigned int wqe_n = 1 << rxq->elts_n;
955 	struct mlx5_rxq_ctrl *rxq_ctrl =
956 			container_of(rxq, struct mlx5_rxq_ctrl, rxq);
957 	union {
958 		volatile struct mlx5_cqe *cqe;
959 		volatile struct mlx5_err_cqe *err_cqe;
960 	} u = {
961 		.cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_mask],
962 	};
963 	struct mlx5_mp_arg_queue_state_modify sm;
964 	int ret;
965 
966 	switch (rxq->err_state) {
967 	case MLX5_RXQ_ERR_STATE_NO_ERROR:
968 		rxq->err_state = MLX5_RXQ_ERR_STATE_NEED_RESET;
969 		/* Fall-through */
970 	case MLX5_RXQ_ERR_STATE_NEED_RESET:
971 		sm.is_wq = 1;
972 		sm.queue_id = rxq->idx;
973 		sm.state = IBV_WQS_RESET;
974 		if (mlx5_queue_state_modify(ETH_DEV(rxq_ctrl->priv), &sm))
975 			return -1;
976 		if (rxq_ctrl->dump_file_n <
977 		    rxq_ctrl->priv->config.max_dump_files_num) {
978 			MKSTR(err_str, "Unexpected CQE error syndrome "
979 			      "0x%02x CQN = %u RQN = %u wqe_counter = %u"
980 			      " rq_ci = %u cq_ci = %u", u.err_cqe->syndrome,
981 			      rxq->cqn, rxq_ctrl->wqn,
982 			      rte_be_to_cpu_16(u.err_cqe->wqe_counter),
983 			      rxq->rq_ci << rxq->sges_n, rxq->cq_ci);
984 			MKSTR(name, "dpdk_mlx5_port_%u_rxq_%u_%u",
985 			      rxq->port_id, rxq->idx, (uint32_t)rte_rdtsc());
986 			mlx5_dump_debug_information(name, NULL, err_str, 0);
987 			mlx5_dump_debug_information(name, "MLX5 Error CQ:",
988 						    (const void *)((uintptr_t)
989 								    rxq->cqes),
990 						    sizeof(*u.cqe) * cqe_n);
991 			mlx5_dump_debug_information(name, "MLX5 Error RQ:",
992 						    (const void *)((uintptr_t)
993 								    rxq->wqes),
994 						    16 * wqe_n);
995 			rxq_ctrl->dump_file_n++;
996 		}
997 		rxq->err_state = MLX5_RXQ_ERR_STATE_NEED_READY;
998 		/* Fall-through */
999 	case MLX5_RXQ_ERR_STATE_NEED_READY:
1000 		ret = check_cqe(u.cqe, cqe_n, rxq->cq_ci);
1001 		if (ret == MLX5_CQE_STATUS_HW_OWN) {
1002 			rte_cio_wmb();
1003 			*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1004 			rte_cio_wmb();
1005 			/*
1006 			 * The RQ consumer index must be zeroed while moving
1007 			 * from RESET state to RDY state.
1008 			 */
1009 			*rxq->rq_db = rte_cpu_to_be_32(0);
1010 			rte_cio_wmb();
1011 			sm.is_wq = 1;
1012 			sm.queue_id = rxq->idx;
1013 			sm.state = IBV_WQS_RDY;
1014 			if (mlx5_queue_state_modify(ETH_DEV(rxq_ctrl->priv),
1015 						    &sm))
1016 				return -1;
1017 			if (mbuf_prepare) {
1018 				const uint16_t q_mask = wqe_n - 1;
1019 				uint16_t elt_idx;
1020 				struct rte_mbuf **elt;
1021 				int i;
1022 				unsigned int n = wqe_n - (rxq->rq_ci -
1023 							  rxq->rq_pi);
1024 
1025 				for (i = 0; i < (int)n; ++i) {
1026 					elt_idx = (rxq->rq_ci + i) & q_mask;
1027 					elt = &(*rxq->elts)[elt_idx];
1028 					*elt = rte_mbuf_raw_alloc(rxq->mp);
1029 					if (!*elt) {
1030 						for (i--; i >= 0; --i) {
1031 							elt_idx = (rxq->rq_ci +
1032 								   i) & q_mask;
1033 							elt = &(*rxq->elts)
1034 								[elt_idx];
1035 							rte_pktmbuf_free_seg
1036 								(*elt);
1037 						}
1038 						return -1;
1039 					}
1040 				}
1041 			}
1042 			mlx5_rxq_initialize(rxq);
1043 			rxq->err_state = MLX5_RXQ_ERR_STATE_NO_ERROR;
1044 		}
1045 		return ret;
1046 	default:
1047 		return -1;
1048 	}
1049 }
1050 
1051 /**
1052  * Get size of the next packet for a given CQE. For compressed CQEs, the
1053  * consumer index is updated only once all packets of the current one have
1054  * been processed.
1055  *
1056  * @param rxq
1057  *   Pointer to RX queue.
1058  * @param cqe
1059  *   CQE to process.
1060  * @param[out] mcqe
1061  *   Store pointer to mini-CQE if compressed. Otherwise, the pointer is not
1062  *   written.
1063  *
1064  * @return
1065  *   0 in case of empty CQE, otherwise the packet size in bytes.
1066  */
1067 static inline int
1068 mlx5_rx_poll_len(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
1069 		 uint16_t cqe_cnt, volatile struct mlx5_mini_cqe8 **mcqe)
1070 {
1071 	struct rxq_zip *zip = &rxq->zip;
1072 	uint16_t cqe_n = cqe_cnt + 1;
1073 	int len;
1074 	uint16_t idx, end;
1075 
1076 	do {
1077 		len = 0;
1078 		/* Process compressed data in the CQE and mini arrays. */
1079 		if (zip->ai) {
1080 			volatile struct mlx5_mini_cqe8 (*mc)[8] =
1081 				(volatile struct mlx5_mini_cqe8 (*)[8])
1082 				(uintptr_t)(&(*rxq->cqes)[zip->ca &
1083 							  cqe_cnt].pkt_info);
1084 
1085 			len = rte_be_to_cpu_32((*mc)[zip->ai & 7].byte_cnt);
1086 			*mcqe = &(*mc)[zip->ai & 7];
1087 			if ((++zip->ai & 7) == 0) {
1088 				/* Invalidate consumed CQEs */
1089 				idx = zip->ca;
1090 				end = zip->na;
1091 				while (idx != end) {
1092 					(*rxq->cqes)[idx & cqe_cnt].op_own =
1093 						MLX5_CQE_INVALIDATE;
1094 					++idx;
1095 				}
1096 				/*
1097 				 * Increment consumer index to skip the number
1098 				 * of CQEs consumed. Hardware leaves holes in
1099 				 * the CQ ring for software use.
1100 				 */
1101 				zip->ca = zip->na;
1102 				zip->na += 8;
1103 			}
1104 			if (unlikely(rxq->zip.ai == rxq->zip.cqe_cnt)) {
1105 				/* Invalidate the rest */
1106 				idx = zip->ca;
1107 				end = zip->cq_ci;
1108 
1109 				while (idx != end) {
1110 					(*rxq->cqes)[idx & cqe_cnt].op_own =
1111 						MLX5_CQE_INVALIDATE;
1112 					++idx;
1113 				}
1114 				rxq->cq_ci = zip->cq_ci;
1115 				zip->ai = 0;
1116 			}
1117 		/*
1118 		 * No compressed data, get next CQE and verify if it is
1119 		 * compressed.
1120 		 */
1121 		} else {
1122 			int ret;
1123 			int8_t op_own;
1124 
1125 			ret = check_cqe(cqe, cqe_n, rxq->cq_ci);
1126 			if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) {
1127 				if (unlikely(ret == MLX5_CQE_STATUS_ERR ||
1128 					     rxq->err_state)) {
1129 					ret = mlx5_rx_err_handle(rxq, 0);
1130 					if (ret == MLX5_CQE_STATUS_HW_OWN ||
1131 					    ret == -1)
1132 						return 0;
1133 				} else {
1134 					return 0;
1135 				}
1136 			}
1137 			++rxq->cq_ci;
1138 			op_own = cqe->op_own;
1139 			if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED) {
1140 				volatile struct mlx5_mini_cqe8 (*mc)[8] =
1141 					(volatile struct mlx5_mini_cqe8 (*)[8])
1142 					(uintptr_t)(&(*rxq->cqes)
1143 						[rxq->cq_ci &
1144 						 cqe_cnt].pkt_info);
1145 
1146 				/* Fix endianness. */
1147 				zip->cqe_cnt = rte_be_to_cpu_32(cqe->byte_cnt);
1148 				/*
1149 				 * Current mini array position is the one
1150 				 * returned by check_cqe64().
1151 				 *
1152 				 * If completion comprises several mini arrays,
1153 				 * as a special case the second one is located
1154 				 * 7 CQEs after the initial CQE instead of 8
1155 				 * for subsequent ones.
1156 				 */
1157 				zip->ca = rxq->cq_ci;
1158 				zip->na = zip->ca + 7;
1159 				/* Compute the next non compressed CQE. */
1160 				--rxq->cq_ci;
1161 				zip->cq_ci = rxq->cq_ci + zip->cqe_cnt;
1162 				/* Get packet size to return. */
1163 				len = rte_be_to_cpu_32((*mc)[0].byte_cnt);
1164 				*mcqe = &(*mc)[0];
1165 				zip->ai = 1;
1166 				/* Prefetch all to be invalidated */
1167 				idx = zip->ca;
1168 				end = zip->cq_ci;
1169 				while (idx != end) {
1170 					rte_prefetch0(&(*rxq->cqes)[(idx) &
1171 								    cqe_cnt]);
1172 					++idx;
1173 				}
1174 			} else {
1175 				len = rte_be_to_cpu_32(cqe->byte_cnt);
1176 			}
1177 		}
1178 		if (unlikely(rxq->err_state)) {
1179 			cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1180 			++rxq->stats.idropped;
1181 		} else {
1182 			return len;
1183 		}
1184 	} while (1);
1185 }
1186 
1187 /**
1188  * Translate RX completion flags to offload flags.
1189  *
1190  * @param[in] cqe
1191  *   Pointer to CQE.
1192  *
1193  * @return
1194  *   Offload flags (ol_flags) for struct rte_mbuf.
1195  */
1196 static inline uint32_t
1197 rxq_cq_to_ol_flags(volatile struct mlx5_cqe *cqe)
1198 {
1199 	uint32_t ol_flags = 0;
1200 	uint16_t flags = rte_be_to_cpu_16(cqe->hdr_type_etc);
1201 
1202 	ol_flags =
1203 		TRANSPOSE(flags,
1204 			  MLX5_CQE_RX_L3_HDR_VALID,
1205 			  PKT_RX_IP_CKSUM_GOOD) |
1206 		TRANSPOSE(flags,
1207 			  MLX5_CQE_RX_L4_HDR_VALID,
1208 			  PKT_RX_L4_CKSUM_GOOD);
1209 	return ol_flags;
1210 }
1211 
1212 /**
1213  * Fill in mbuf fields from RX completion flags.
1214  * Note that pkt->ol_flags should be initialized outside of this function.
1215  *
1216  * @param rxq
1217  *   Pointer to RX queue.
1218  * @param pkt
1219  *   mbuf to fill.
1220  * @param cqe
1221  *   CQE to process.
1222  * @param rss_hash_res
1223  *   Packet RSS Hash result.
1224  */
1225 static inline void
1226 rxq_cq_to_mbuf(struct mlx5_rxq_data *rxq, struct rte_mbuf *pkt,
1227 	       volatile struct mlx5_cqe *cqe, uint32_t rss_hash_res)
1228 {
1229 	/* Update packet information. */
1230 	pkt->packet_type = rxq_cq_to_pkt_type(rxq, cqe);
1231 	if (rss_hash_res && rxq->rss_hash) {
1232 		pkt->hash.rss = rss_hash_res;
1233 		pkt->ol_flags |= PKT_RX_RSS_HASH;
1234 	}
1235 	if (rxq->mark && MLX5_FLOW_MARK_IS_VALID(cqe->sop_drop_qpn)) {
1236 		pkt->ol_flags |= PKT_RX_FDIR;
1237 		if (cqe->sop_drop_qpn !=
1238 		    rte_cpu_to_be_32(MLX5_FLOW_MARK_DEFAULT)) {
1239 			uint32_t mark = cqe->sop_drop_qpn;
1240 
1241 			pkt->ol_flags |= PKT_RX_FDIR_ID;
1242 			pkt->hash.fdir.hi = mlx5_flow_mark_get(mark);
1243 		}
1244 	}
1245 	if (rxq->csum)
1246 		pkt->ol_flags |= rxq_cq_to_ol_flags(cqe);
1247 	if (rxq->vlan_strip &&
1248 	    (cqe->hdr_type_etc & rte_cpu_to_be_16(MLX5_CQE_VLAN_STRIPPED))) {
1249 		pkt->ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
1250 		pkt->vlan_tci = rte_be_to_cpu_16(cqe->vlan_info);
1251 	}
1252 	if (rxq->hw_timestamp) {
1253 		pkt->timestamp = rte_be_to_cpu_64(cqe->timestamp);
1254 		pkt->ol_flags |= PKT_RX_TIMESTAMP;
1255 	}
1256 }
1257 
1258 /**
1259  * DPDK callback for RX.
1260  *
1261  * @param dpdk_rxq
1262  *   Generic pointer to RX queue structure.
1263  * @param[out] pkts
1264  *   Array to store received packets.
1265  * @param pkts_n
1266  *   Maximum number of packets in array.
1267  *
1268  * @return
1269  *   Number of packets successfully received (<= pkts_n).
1270  */
1271 uint16_t
1272 mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1273 {
1274 	struct mlx5_rxq_data *rxq = dpdk_rxq;
1275 	const unsigned int wqe_cnt = (1 << rxq->elts_n) - 1;
1276 	const unsigned int cqe_cnt = (1 << rxq->cqe_n) - 1;
1277 	const unsigned int sges_n = rxq->sges_n;
1278 	struct rte_mbuf *pkt = NULL;
1279 	struct rte_mbuf *seg = NULL;
1280 	volatile struct mlx5_cqe *cqe =
1281 		&(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1282 	unsigned int i = 0;
1283 	unsigned int rq_ci = rxq->rq_ci << sges_n;
1284 	int len = 0; /* keep its value across iterations. */
1285 
1286 	while (pkts_n) {
1287 		unsigned int idx = rq_ci & wqe_cnt;
1288 		volatile struct mlx5_wqe_data_seg *wqe =
1289 			&((volatile struct mlx5_wqe_data_seg *)rxq->wqes)[idx];
1290 		struct rte_mbuf *rep = (*rxq->elts)[idx];
1291 		volatile struct mlx5_mini_cqe8 *mcqe = NULL;
1292 		uint32_t rss_hash_res;
1293 
1294 		if (pkt)
1295 			NEXT(seg) = rep;
1296 		seg = rep;
1297 		rte_prefetch0(seg);
1298 		rte_prefetch0(cqe);
1299 		rte_prefetch0(wqe);
1300 		rep = rte_mbuf_raw_alloc(rxq->mp);
1301 		if (unlikely(rep == NULL)) {
1302 			++rxq->stats.rx_nombuf;
1303 			if (!pkt) {
1304 				/*
1305 				 * no buffers before we even started,
1306 				 * bail out silently.
1307 				 */
1308 				break;
1309 			}
1310 			while (pkt != seg) {
1311 				assert(pkt != (*rxq->elts)[idx]);
1312 				rep = NEXT(pkt);
1313 				NEXT(pkt) = NULL;
1314 				NB_SEGS(pkt) = 1;
1315 				rte_mbuf_raw_free(pkt);
1316 				pkt = rep;
1317 			}
1318 			break;
1319 		}
1320 		if (!pkt) {
1321 			cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1322 			len = mlx5_rx_poll_len(rxq, cqe, cqe_cnt, &mcqe);
1323 			if (!len) {
1324 				rte_mbuf_raw_free(rep);
1325 				break;
1326 			}
1327 			pkt = seg;
1328 			assert(len >= (rxq->crc_present << 2));
1329 			pkt->ol_flags = 0;
1330 			/* If compressed, take hash result from mini-CQE. */
1331 			rss_hash_res = rte_be_to_cpu_32(mcqe == NULL ?
1332 							cqe->rx_hash_res :
1333 							mcqe->rx_hash_result);
1334 			rxq_cq_to_mbuf(rxq, pkt, cqe, rss_hash_res);
1335 			if (rxq->crc_present)
1336 				len -= RTE_ETHER_CRC_LEN;
1337 			PKT_LEN(pkt) = len;
1338 			if (cqe->lro_num_seg > 1) {
1339 				mlx5_lro_update_hdr
1340 					(rte_pktmbuf_mtod(pkt, uint8_t *), cqe,
1341 					 len);
1342 				pkt->ol_flags |= PKT_RX_LRO;
1343 				pkt->tso_segsz = len / cqe->lro_num_seg;
1344 			}
1345 		}
1346 		DATA_LEN(rep) = DATA_LEN(seg);
1347 		PKT_LEN(rep) = PKT_LEN(seg);
1348 		SET_DATA_OFF(rep, DATA_OFF(seg));
1349 		PORT(rep) = PORT(seg);
1350 		(*rxq->elts)[idx] = rep;
1351 		/*
1352 		 * Fill NIC descriptor with the new buffer.  The lkey and size
1353 		 * of the buffers are already known, only the buffer address
1354 		 * changes.
1355 		 */
1356 		wqe->addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(rep, uintptr_t));
1357 		/* If there's only one MR, no need to replace LKey in WQE. */
1358 		if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh) > 1))
1359 			wqe->lkey = mlx5_rx_mb2mr(rxq, rep);
1360 		if (len > DATA_LEN(seg)) {
1361 			len -= DATA_LEN(seg);
1362 			++NB_SEGS(pkt);
1363 			++rq_ci;
1364 			continue;
1365 		}
1366 		DATA_LEN(seg) = len;
1367 #ifdef MLX5_PMD_SOFT_COUNTERS
1368 		/* Increment bytes counter. */
1369 		rxq->stats.ibytes += PKT_LEN(pkt);
1370 #endif
1371 		/* Return packet. */
1372 		*(pkts++) = pkt;
1373 		pkt = NULL;
1374 		--pkts_n;
1375 		++i;
1376 		/* Align consumer index to the next stride. */
1377 		rq_ci >>= sges_n;
1378 		++rq_ci;
1379 		rq_ci <<= sges_n;
1380 	}
1381 	if (unlikely((i == 0) && ((rq_ci >> sges_n) == rxq->rq_ci)))
1382 		return 0;
1383 	/* Update the consumer index. */
1384 	rxq->rq_ci = rq_ci >> sges_n;
1385 	rte_cio_wmb();
1386 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1387 	rte_cio_wmb();
1388 	*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
1389 #ifdef MLX5_PMD_SOFT_COUNTERS
1390 	/* Increment packets counter. */
1391 	rxq->stats.ipackets += i;
1392 #endif
1393 	return i;
1394 }
1395 
1396 /**
1397  * Update LRO packet TCP header.
1398  * The HW LRO feature doesn't update the TCP header after coalescing the
1399  * TCP segments but supplies information in CQE to fill it by SW.
1400  *
1401  * @param tcp
1402  *   Pointer to the TCP header.
1403  * @param cqe
1404  *   Pointer to the completion entry..
1405  * @param phcsum
1406  *   The L3 pseudo-header checksum.
1407  */
1408 static inline void
1409 mlx5_lro_update_tcp_hdr(struct rte_tcp_hdr *restrict tcp,
1410 			volatile struct mlx5_cqe *restrict cqe,
1411 			uint32_t phcsum)
1412 {
1413 	uint8_t l4_type = (rte_be_to_cpu_16(cqe->hdr_type_etc) &
1414 			   MLX5_CQE_L4_TYPE_MASK) >> MLX5_CQE_L4_TYPE_SHIFT;
1415 	/*
1416 	 * The HW calculates only the TCP payload checksum, need to complete
1417 	 * the TCP header checksum and the L3 pseudo-header checksum.
1418 	 */
1419 	uint32_t csum = phcsum + cqe->csum;
1420 
1421 	if (l4_type == MLX5_L4_HDR_TYPE_TCP_EMPTY_ACK ||
1422 	    l4_type == MLX5_L4_HDR_TYPE_TCP_WITH_ACL) {
1423 		tcp->tcp_flags |= RTE_TCP_ACK_FLAG;
1424 		tcp->recv_ack = cqe->lro_ack_seq_num;
1425 		tcp->rx_win = cqe->lro_tcp_win;
1426 	}
1427 	if (cqe->lro_tcppsh_abort_dupack & MLX5_CQE_LRO_PUSH_MASK)
1428 		tcp->tcp_flags |= RTE_TCP_PSH_FLAG;
1429 	tcp->cksum = 0;
1430 	csum += rte_raw_cksum(tcp, (tcp->data_off & 0xF) * 4);
1431 	csum = ((csum & 0xffff0000) >> 16) + (csum & 0xffff);
1432 	csum = (~csum) & 0xffff;
1433 	if (csum == 0)
1434 		csum = 0xffff;
1435 	tcp->cksum = csum;
1436 }
1437 
1438 /**
1439  * Update LRO packet headers.
1440  * The HW LRO feature doesn't update the L3/TCP headers after coalescing the
1441  * TCP segments but supply information in CQE to fill it by SW.
1442  *
1443  * @param padd
1444  *   The packet address.
1445  * @param cqe
1446  *   Pointer to the completion entry..
1447  * @param len
1448  *   The packet length.
1449  */
1450 static inline void
1451 mlx5_lro_update_hdr(uint8_t *restrict padd,
1452 		    volatile struct mlx5_cqe *restrict cqe,
1453 		    uint32_t len)
1454 {
1455 	union {
1456 		struct rte_ether_hdr *eth;
1457 		struct rte_vlan_hdr *vlan;
1458 		struct rte_ipv4_hdr *ipv4;
1459 		struct rte_ipv6_hdr *ipv6;
1460 		struct rte_tcp_hdr *tcp;
1461 		uint8_t *hdr;
1462 	} h = {
1463 			.hdr = padd,
1464 	};
1465 	uint16_t proto = h.eth->ether_type;
1466 	uint32_t phcsum;
1467 
1468 	h.eth++;
1469 	while (proto == RTE_BE16(RTE_ETHER_TYPE_VLAN) ||
1470 	       proto == RTE_BE16(RTE_ETHER_TYPE_QINQ)) {
1471 		proto = h.vlan->eth_proto;
1472 		h.vlan++;
1473 	}
1474 	if (proto == RTE_BE16(RTE_ETHER_TYPE_IPV4)) {
1475 		h.ipv4->time_to_live = cqe->lro_min_ttl;
1476 		h.ipv4->total_length = rte_cpu_to_be_16(len - (h.hdr - padd));
1477 		h.ipv4->hdr_checksum = 0;
1478 		h.ipv4->hdr_checksum = rte_ipv4_cksum(h.ipv4);
1479 		phcsum = rte_ipv4_phdr_cksum(h.ipv4, 0);
1480 		h.ipv4++;
1481 	} else {
1482 		h.ipv6->hop_limits = cqe->lro_min_ttl;
1483 		h.ipv6->payload_len = rte_cpu_to_be_16(len - (h.hdr - padd) -
1484 						       sizeof(*h.ipv6));
1485 		phcsum = rte_ipv6_phdr_cksum(h.ipv6, 0);
1486 		h.ipv6++;
1487 	}
1488 	mlx5_lro_update_tcp_hdr(h.tcp, cqe, phcsum);
1489 }
1490 
1491 void
1492 mlx5_mprq_buf_free_cb(void *addr __rte_unused, void *opaque)
1493 {
1494 	struct mlx5_mprq_buf *buf = opaque;
1495 
1496 	if (rte_atomic16_read(&buf->refcnt) == 1) {
1497 		rte_mempool_put(buf->mp, buf);
1498 	} else if (rte_atomic16_add_return(&buf->refcnt, -1) == 0) {
1499 		rte_atomic16_set(&buf->refcnt, 1);
1500 		rte_mempool_put(buf->mp, buf);
1501 	}
1502 }
1503 
1504 void
1505 mlx5_mprq_buf_free(struct mlx5_mprq_buf *buf)
1506 {
1507 	mlx5_mprq_buf_free_cb(NULL, buf);
1508 }
1509 
1510 static inline void
1511 mprq_buf_replace(struct mlx5_rxq_data *rxq, uint16_t rq_idx,
1512 		 const unsigned int strd_n)
1513 {
1514 	struct mlx5_mprq_buf *rep = rxq->mprq_repl;
1515 	volatile struct mlx5_wqe_data_seg *wqe =
1516 		&((volatile struct mlx5_wqe_mprq *)rxq->wqes)[rq_idx].dseg;
1517 	void *addr;
1518 
1519 	assert(rep != NULL);
1520 	/* Replace MPRQ buf. */
1521 	(*rxq->mprq_bufs)[rq_idx] = rep;
1522 	/* Replace WQE. */
1523 	addr = mlx5_mprq_buf_addr(rep, strd_n);
1524 	wqe->addr = rte_cpu_to_be_64((uintptr_t)addr);
1525 	/* If there's only one MR, no need to replace LKey in WQE. */
1526 	if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh) > 1))
1527 		wqe->lkey = mlx5_rx_addr2mr(rxq, (uintptr_t)addr);
1528 	/* Stash a mbuf for next replacement. */
1529 	if (likely(!rte_mempool_get(rxq->mprq_mp, (void **)&rep)))
1530 		rxq->mprq_repl = rep;
1531 	else
1532 		rxq->mprq_repl = NULL;
1533 }
1534 
1535 /**
1536  * DPDK callback for RX with Multi-Packet RQ support.
1537  *
1538  * @param dpdk_rxq
1539  *   Generic pointer to RX queue structure.
1540  * @param[out] pkts
1541  *   Array to store received packets.
1542  * @param pkts_n
1543  *   Maximum number of packets in array.
1544  *
1545  * @return
1546  *   Number of packets successfully received (<= pkts_n).
1547  */
1548 uint16_t
1549 mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1550 {
1551 	struct mlx5_rxq_data *rxq = dpdk_rxq;
1552 	const unsigned int strd_n = 1 << rxq->strd_num_n;
1553 	const unsigned int strd_sz = 1 << rxq->strd_sz_n;
1554 	const unsigned int strd_shift =
1555 		MLX5_MPRQ_STRIDE_SHIFT_BYTE * rxq->strd_shift_en;
1556 	const unsigned int cq_mask = (1 << rxq->cqe_n) - 1;
1557 	const unsigned int wq_mask = (1 << rxq->elts_n) - 1;
1558 	volatile struct mlx5_cqe *cqe = &(*rxq->cqes)[rxq->cq_ci & cq_mask];
1559 	unsigned int i = 0;
1560 	uint32_t rq_ci = rxq->rq_ci;
1561 	uint16_t consumed_strd = rxq->consumed_strd;
1562 	uint16_t headroom_sz = rxq->strd_headroom_en * RTE_PKTMBUF_HEADROOM;
1563 	struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[rq_ci & wq_mask];
1564 
1565 	while (i < pkts_n) {
1566 		struct rte_mbuf *pkt;
1567 		void *addr;
1568 		int ret;
1569 		unsigned int len;
1570 		uint16_t strd_cnt;
1571 		uint16_t strd_idx;
1572 		uint32_t offset;
1573 		uint32_t byte_cnt;
1574 		volatile struct mlx5_mini_cqe8 *mcqe = NULL;
1575 		uint32_t rss_hash_res = 0;
1576 		uint8_t lro_num_seg;
1577 
1578 		if (consumed_strd == strd_n) {
1579 			/* Replace WQE only if the buffer is still in use. */
1580 			if (rte_atomic16_read(&buf->refcnt) > 1) {
1581 				mprq_buf_replace(rxq, rq_ci & wq_mask, strd_n);
1582 				/* Release the old buffer. */
1583 				mlx5_mprq_buf_free(buf);
1584 			} else if (unlikely(rxq->mprq_repl == NULL)) {
1585 				struct mlx5_mprq_buf *rep;
1586 
1587 				/*
1588 				 * Currently, the MPRQ mempool is out of buffer
1589 				 * and doing memcpy regardless of the size of Rx
1590 				 * packet. Retry allocation to get back to
1591 				 * normal.
1592 				 */
1593 				if (!rte_mempool_get(rxq->mprq_mp,
1594 						     (void **)&rep))
1595 					rxq->mprq_repl = rep;
1596 			}
1597 			/* Advance to the next WQE. */
1598 			consumed_strd = 0;
1599 			++rq_ci;
1600 			buf = (*rxq->mprq_bufs)[rq_ci & wq_mask];
1601 		}
1602 		cqe = &(*rxq->cqes)[rxq->cq_ci & cq_mask];
1603 		ret = mlx5_rx_poll_len(rxq, cqe, cq_mask, &mcqe);
1604 		if (!ret)
1605 			break;
1606 		byte_cnt = ret;
1607 		strd_cnt = (byte_cnt & MLX5_MPRQ_STRIDE_NUM_MASK) >>
1608 			   MLX5_MPRQ_STRIDE_NUM_SHIFT;
1609 		assert(strd_cnt);
1610 		consumed_strd += strd_cnt;
1611 		if (byte_cnt & MLX5_MPRQ_FILLER_MASK)
1612 			continue;
1613 		if (mcqe == NULL) {
1614 			rss_hash_res = rte_be_to_cpu_32(cqe->rx_hash_res);
1615 			strd_idx = rte_be_to_cpu_16(cqe->wqe_counter);
1616 		} else {
1617 			/* mini-CQE for MPRQ doesn't have hash result. */
1618 			strd_idx = rte_be_to_cpu_16(mcqe->stride_idx);
1619 		}
1620 		assert(strd_idx < strd_n);
1621 		assert(!((rte_be_to_cpu_16(cqe->wqe_id) ^ rq_ci) & wq_mask));
1622 		lro_num_seg = cqe->lro_num_seg;
1623 		/*
1624 		 * Currently configured to receive a packet per a stride. But if
1625 		 * MTU is adjusted through kernel interface, device could
1626 		 * consume multiple strides without raising an error. In this
1627 		 * case, the packet should be dropped because it is bigger than
1628 		 * the max_rx_pkt_len.
1629 		 */
1630 		if (unlikely(!lro_num_seg && strd_cnt > 1)) {
1631 			++rxq->stats.idropped;
1632 			continue;
1633 		}
1634 		pkt = rte_pktmbuf_alloc(rxq->mp);
1635 		if (unlikely(pkt == NULL)) {
1636 			++rxq->stats.rx_nombuf;
1637 			break;
1638 		}
1639 		len = (byte_cnt & MLX5_MPRQ_LEN_MASK) >> MLX5_MPRQ_LEN_SHIFT;
1640 		assert((int)len >= (rxq->crc_present << 2));
1641 		if (rxq->crc_present)
1642 			len -= RTE_ETHER_CRC_LEN;
1643 		offset = strd_idx * strd_sz + strd_shift;
1644 		addr = RTE_PTR_ADD(mlx5_mprq_buf_addr(buf, strd_n), offset);
1645 		/*
1646 		 * Memcpy packets to the target mbuf if:
1647 		 * - The size of packet is smaller than mprq_max_memcpy_len.
1648 		 * - Out of buffer in the Mempool for Multi-Packet RQ.
1649 		 */
1650 		if (len <= rxq->mprq_max_memcpy_len || rxq->mprq_repl == NULL) {
1651 			/*
1652 			 * When memcpy'ing packet due to out-of-buffer, the
1653 			 * packet must be smaller than the target mbuf.
1654 			 */
1655 			if (unlikely(rte_pktmbuf_tailroom(pkt) < len)) {
1656 				rte_pktmbuf_free_seg(pkt);
1657 				++rxq->stats.idropped;
1658 				continue;
1659 			}
1660 			rte_memcpy(rte_pktmbuf_mtod(pkt, void *), addr, len);
1661 			DATA_LEN(pkt) = len;
1662 		} else {
1663 			rte_iova_t buf_iova;
1664 			struct rte_mbuf_ext_shared_info *shinfo;
1665 			uint16_t buf_len = strd_cnt * strd_sz;
1666 			void *buf_addr;
1667 
1668 			/* Increment the refcnt of the whole chunk. */
1669 			rte_atomic16_add_return(&buf->refcnt, 1);
1670 			assert((uint16_t)rte_atomic16_read(&buf->refcnt) <=
1671 			       strd_n + 1);
1672 			buf_addr = RTE_PTR_SUB(addr, headroom_sz);
1673 			/*
1674 			 * MLX5 device doesn't use iova but it is necessary in a
1675 			 * case where the Rx packet is transmitted via a
1676 			 * different PMD.
1677 			 */
1678 			buf_iova = rte_mempool_virt2iova(buf) +
1679 				   RTE_PTR_DIFF(buf_addr, buf);
1680 			shinfo = &buf->shinfos[strd_idx];
1681 			rte_mbuf_ext_refcnt_set(shinfo, 1);
1682 			/*
1683 			 * EXT_ATTACHED_MBUF will be set to pkt->ol_flags when
1684 			 * attaching the stride to mbuf and more offload flags
1685 			 * will be added below by calling rxq_cq_to_mbuf().
1686 			 * Other fields will be overwritten.
1687 			 */
1688 			rte_pktmbuf_attach_extbuf(pkt, buf_addr, buf_iova,
1689 						  buf_len, shinfo);
1690 			/* Set mbuf head-room. */
1691 			pkt->data_off = headroom_sz;
1692 			assert(pkt->ol_flags == EXT_ATTACHED_MBUF);
1693 			/*
1694 			 * Prevent potential overflow due to MTU change through
1695 			 * kernel interface.
1696 			 */
1697 			if (unlikely(rte_pktmbuf_tailroom(pkt) < len)) {
1698 				rte_pktmbuf_free_seg(pkt);
1699 				++rxq->stats.idropped;
1700 				continue;
1701 			}
1702 			DATA_LEN(pkt) = len;
1703 			/*
1704 			 * LRO packet may consume all the stride memory, in this
1705 			 * case packet head-room space is not guaranteed so must
1706 			 * to add an empty mbuf for the head-room.
1707 			 */
1708 			if (!rxq->strd_headroom_en) {
1709 				struct rte_mbuf *headroom_mbuf =
1710 						rte_pktmbuf_alloc(rxq->mp);
1711 
1712 				if (unlikely(headroom_mbuf == NULL)) {
1713 					rte_pktmbuf_free_seg(pkt);
1714 					++rxq->stats.rx_nombuf;
1715 					break;
1716 				}
1717 				PORT(pkt) = rxq->port_id;
1718 				NEXT(headroom_mbuf) = pkt;
1719 				pkt = headroom_mbuf;
1720 				NB_SEGS(pkt) = 2;
1721 			}
1722 		}
1723 		rxq_cq_to_mbuf(rxq, pkt, cqe, rss_hash_res);
1724 		if (lro_num_seg > 1) {
1725 			mlx5_lro_update_hdr(addr, cqe, len);
1726 			pkt->ol_flags |= PKT_RX_LRO;
1727 			pkt->tso_segsz = strd_sz;
1728 		}
1729 		PKT_LEN(pkt) = len;
1730 		PORT(pkt) = rxq->port_id;
1731 #ifdef MLX5_PMD_SOFT_COUNTERS
1732 		/* Increment bytes counter. */
1733 		rxq->stats.ibytes += PKT_LEN(pkt);
1734 #endif
1735 		/* Return packet. */
1736 		*(pkts++) = pkt;
1737 		++i;
1738 	}
1739 	/* Update the consumer indexes. */
1740 	rxq->consumed_strd = consumed_strd;
1741 	rte_cio_wmb();
1742 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1743 	if (rq_ci != rxq->rq_ci) {
1744 		rxq->rq_ci = rq_ci;
1745 		rte_cio_wmb();
1746 		*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
1747 	}
1748 #ifdef MLX5_PMD_SOFT_COUNTERS
1749 	/* Increment packets counter. */
1750 	rxq->stats.ipackets += i;
1751 #endif
1752 	return i;
1753 }
1754 
1755 /**
1756  * Dummy DPDK callback for TX.
1757  *
1758  * This function is used to temporarily replace the real callback during
1759  * unsafe control operations on the queue, or in case of error.
1760  *
1761  * @param dpdk_txq
1762  *   Generic pointer to TX queue structure.
1763  * @param[in] pkts
1764  *   Packets to transmit.
1765  * @param pkts_n
1766  *   Number of packets in array.
1767  *
1768  * @return
1769  *   Number of packets successfully transmitted (<= pkts_n).
1770  */
1771 uint16_t
1772 removed_tx_burst(void *dpdk_txq __rte_unused,
1773 		 struct rte_mbuf **pkts __rte_unused,
1774 		 uint16_t pkts_n __rte_unused)
1775 {
1776 	rte_mb();
1777 	return 0;
1778 }
1779 
1780 /**
1781  * Dummy DPDK callback for RX.
1782  *
1783  * This function is used to temporarily replace the real callback during
1784  * unsafe control operations on the queue, or in case of error.
1785  *
1786  * @param dpdk_rxq
1787  *   Generic pointer to RX queue structure.
1788  * @param[out] pkts
1789  *   Array to store received packets.
1790  * @param pkts_n
1791  *   Maximum number of packets in array.
1792  *
1793  * @return
1794  *   Number of packets successfully received (<= pkts_n).
1795  */
1796 uint16_t
1797 removed_rx_burst(void *dpdk_txq __rte_unused,
1798 		 struct rte_mbuf **pkts __rte_unused,
1799 		 uint16_t pkts_n __rte_unused)
1800 {
1801 	rte_mb();
1802 	return 0;
1803 }
1804 
1805 /*
1806  * Vectorized Rx/Tx routines are not compiled in when required vector
1807  * instructions are not supported on a target architecture. The following null
1808  * stubs are needed for linkage when those are not included outside of this file
1809  * (e.g.  mlx5_rxtx_vec_sse.c for x86).
1810  */
1811 
1812 __rte_weak uint16_t
1813 mlx5_rx_burst_vec(void *dpdk_txq __rte_unused,
1814 		  struct rte_mbuf **pkts __rte_unused,
1815 		  uint16_t pkts_n __rte_unused)
1816 {
1817 	return 0;
1818 }
1819 
1820 __rte_weak int
1821 mlx5_rxq_check_vec_support(struct mlx5_rxq_data *rxq __rte_unused)
1822 {
1823 	return -ENOTSUP;
1824 }
1825 
1826 __rte_weak int
1827 mlx5_check_vec_rx_support(struct rte_eth_dev *dev __rte_unused)
1828 {
1829 	return -ENOTSUP;
1830 }
1831 
1832 /**
1833  * Free the mbufs from the linear array of pointers.
1834  *
1835  * @param pkts
1836  *   Pointer to array of packets to be free.
1837  * @param pkts_n
1838  *   Number of packets to be freed.
1839  * @param olx
1840  *   Configured Tx offloads mask. It is fully defined at
1841  *   compile time and may be used for optimization.
1842  */
1843 static __rte_always_inline void
1844 mlx5_tx_free_mbuf(struct rte_mbuf **restrict pkts,
1845 		  unsigned int pkts_n,
1846 		  unsigned int olx __rte_unused)
1847 {
1848 	struct rte_mempool *pool = NULL;
1849 	struct rte_mbuf **p_free = NULL;
1850 	struct rte_mbuf *mbuf;
1851 	unsigned int n_free = 0;
1852 
1853 	/*
1854 	 * The implemented algorithm eliminates
1855 	 * copying pointers to temporary array
1856 	 * for rte_mempool_put_bulk() calls.
1857 	 */
1858 	assert(pkts);
1859 	assert(pkts_n);
1860 	for (;;) {
1861 		for (;;) {
1862 			/*
1863 			 * Decrement mbuf reference counter, detach
1864 			 * indirect and external buffers if needed.
1865 			 */
1866 			mbuf = rte_pktmbuf_prefree_seg(*pkts);
1867 			if (likely(mbuf != NULL)) {
1868 				assert(mbuf == *pkts);
1869 				if (likely(n_free != 0)) {
1870 					if (unlikely(pool != mbuf->pool))
1871 						/* From different pool. */
1872 						break;
1873 				} else {
1874 					/* Start new scan array. */
1875 					pool = mbuf->pool;
1876 					p_free = pkts;
1877 				}
1878 				++n_free;
1879 				++pkts;
1880 				--pkts_n;
1881 				if (unlikely(pkts_n == 0)) {
1882 					mbuf = NULL;
1883 					break;
1884 				}
1885 			} else {
1886 				/*
1887 				 * This happens if mbuf is still referenced.
1888 				 * We can't put it back to the pool, skip.
1889 				 */
1890 				++pkts;
1891 				--pkts_n;
1892 				if (unlikely(n_free != 0))
1893 					/* There is some array to free.*/
1894 					break;
1895 				if (unlikely(pkts_n == 0))
1896 					/* Last mbuf, nothing to free. */
1897 					return;
1898 			}
1899 		}
1900 		for (;;) {
1901 			/*
1902 			 * This loop is implemented to avoid multiple
1903 			 * inlining of rte_mempool_put_bulk().
1904 			 */
1905 			assert(pool);
1906 			assert(p_free);
1907 			assert(n_free);
1908 			/*
1909 			 * Free the array of pre-freed mbufs
1910 			 * belonging to the same memory pool.
1911 			 */
1912 			rte_mempool_put_bulk(pool, (void *)p_free, n_free);
1913 			if (unlikely(mbuf != NULL)) {
1914 				/* There is the request to start new scan. */
1915 				pool = mbuf->pool;
1916 				p_free = pkts++;
1917 				n_free = 1;
1918 				--pkts_n;
1919 				if (likely(pkts_n != 0))
1920 					break;
1921 				/*
1922 				 * This is the last mbuf to be freed.
1923 				 * Do one more loop iteration to complete.
1924 				 * This is rare case of the last unique mbuf.
1925 				 */
1926 				mbuf = NULL;
1927 				continue;
1928 			}
1929 			if (likely(pkts_n == 0))
1930 				return;
1931 			n_free = 0;
1932 			break;
1933 		}
1934 	}
1935 }
1936 
1937 /**
1938  * Free the mbuf from the elts ring buffer till new tail.
1939  *
1940  * @param txq
1941  *   Pointer to Tx queue structure.
1942  * @param tail
1943  *   Index in elts to free up to, becomes new elts tail.
1944  * @param olx
1945  *   Configured Tx offloads mask. It is fully defined at
1946  *   compile time and may be used for optimization.
1947  */
1948 static __rte_always_inline void
1949 mlx5_tx_free_elts(struct mlx5_txq_data *restrict txq,
1950 		  uint16_t tail,
1951 		  unsigned int olx __rte_unused)
1952 {
1953 	uint16_t n_elts = tail - txq->elts_tail;
1954 
1955 	assert(n_elts);
1956 	assert(n_elts <= txq->elts_s);
1957 	/*
1958 	 * Implement a loop to support ring buffer wraparound
1959 	 * with single inlining of mlx5_tx_free_mbuf().
1960 	 */
1961 	do {
1962 		unsigned int part;
1963 
1964 		part = txq->elts_s - (txq->elts_tail & txq->elts_m);
1965 		part = RTE_MIN(part, n_elts);
1966 		assert(part);
1967 		assert(part <= txq->elts_s);
1968 		mlx5_tx_free_mbuf(&txq->elts[txq->elts_tail & txq->elts_m],
1969 				  part, olx);
1970 		txq->elts_tail += part;
1971 		n_elts -= part;
1972 	} while (n_elts);
1973 }
1974 
1975 /**
1976  * Store the mbuf being sent into elts ring buffer.
1977  * On Tx completion these mbufs will be freed.
1978  *
1979  * @param txq
1980  *   Pointer to Tx queue structure.
1981  * @param pkts
1982  *   Pointer to array of packets to be stored.
1983  * @param pkts_n
1984  *   Number of packets to be stored.
1985  * @param olx
1986  *   Configured Tx offloads mask. It is fully defined at
1987  *   compile time and may be used for optimization.
1988  */
1989 static __rte_always_inline void
1990 mlx5_tx_copy_elts(struct mlx5_txq_data *restrict txq,
1991 		  struct rte_mbuf **restrict pkts,
1992 		  unsigned int pkts_n,
1993 		  unsigned int olx __rte_unused)
1994 {
1995 	unsigned int part;
1996 	struct rte_mbuf **elts = (struct rte_mbuf **)txq->elts;
1997 
1998 	assert(pkts);
1999 	assert(pkts_n);
2000 	part = txq->elts_s - (txq->elts_head & txq->elts_m);
2001 	assert(part);
2002 	assert(part <= txq->elts_s);
2003 	/* This code is a good candidate for vectorizing with SIMD. */
2004 	rte_memcpy((void *)(elts + (txq->elts_head & txq->elts_m)),
2005 		   (void *)pkts,
2006 		   RTE_MIN(part, pkts_n) * sizeof(struct rte_mbuf *));
2007 	txq->elts_head += pkts_n;
2008 	if (unlikely(part < pkts_n))
2009 		/* The copy is wrapping around the elts array. */
2010 		rte_memcpy((void *)elts, (void *)(pkts + part),
2011 			   (pkts_n - part) * sizeof(struct rte_mbuf *));
2012 }
2013 
2014 /**
2015  * Update completion queue consuming index via doorbell
2016  * and flush the completed data buffers.
2017  *
2018  * @param txq
2019  *   Pointer to TX queue structure.
2020  * @param valid CQE pointer
2021  *   if not NULL update txq->wqe_pi and flush the buffers
2022  * @param itail
2023  *   if not negative - flush the buffers till this index.
2024  * @param olx
2025  *   Configured Tx offloads mask. It is fully defined at
2026  *   compile time and may be used for optimization.
2027  */
2028 static __rte_always_inline void
2029 mlx5_tx_comp_flush(struct mlx5_txq_data *restrict txq,
2030 		   volatile struct mlx5_cqe *last_cqe,
2031 		   int itail,
2032 		   unsigned int olx __rte_unused)
2033 {
2034 	uint16_t tail;
2035 
2036 	if (likely(last_cqe != NULL)) {
2037 		txq->wqe_pi = rte_be_to_cpu_16(last_cqe->wqe_counter);
2038 		tail = ((volatile struct mlx5_wqe_cseg *)
2039 			(txq->wqes + (txq->wqe_pi & txq->wqe_m)))->misc;
2040 	} else if (itail >= 0) {
2041 		tail = (uint16_t)itail;
2042 	} else {
2043 		return;
2044 	}
2045 	rte_compiler_barrier();
2046 	*txq->cq_db = rte_cpu_to_be_32(txq->cq_ci);
2047 	if (likely(tail != txq->elts_tail)) {
2048 		mlx5_tx_free_elts(txq, tail, olx);
2049 		assert(tail == txq->elts_tail);
2050 	}
2051 }
2052 
2053 /**
2054  * Manage TX completions. This routine checks the CQ for
2055  * arrived CQEs, deduces the last accomplished WQE in SQ,
2056  * updates SQ producing index and frees all completed mbufs.
2057  *
2058  * @param txq
2059  *   Pointer to TX queue structure.
2060  * @param olx
2061  *   Configured Tx offloads mask. It is fully defined at
2062  *   compile time and may be used for optimization.
2063  *
2064  * NOTE: not inlined intentionally, it makes tx_burst
2065  * routine smaller, simple and faster - from experiments.
2066  */
2067 static void
2068 mlx5_tx_handle_completion(struct mlx5_txq_data *restrict txq,
2069 			  unsigned int olx __rte_unused)
2070 {
2071 	unsigned int count = MLX5_TX_COMP_MAX_CQE;
2072 	volatile struct mlx5_cqe *last_cqe = NULL;
2073 	int ret;
2074 
2075 	static_assert(MLX5_CQE_STATUS_HW_OWN < 0, "Must be negative value");
2076 	static_assert(MLX5_CQE_STATUS_SW_OWN < 0, "Must be negative value");
2077 	do {
2078 		volatile struct mlx5_cqe *cqe;
2079 
2080 		cqe = &txq->cqes[txq->cq_ci & txq->cqe_m];
2081 		ret = check_cqe(cqe, txq->cqe_s, txq->cq_ci);
2082 		if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) {
2083 			if (likely(ret != MLX5_CQE_STATUS_ERR)) {
2084 				/* No new CQEs in completion queue. */
2085 				assert(ret == MLX5_CQE_STATUS_HW_OWN);
2086 				break;
2087 			}
2088 			/*
2089 			 * Some error occurred, try to restart.
2090 			 * We have no barrier after WQE related Doorbell
2091 			 * written, make sure all writes are completed
2092 			 * here, before we might perform SQ reset.
2093 			 */
2094 			rte_wmb();
2095 			ret = mlx5_tx_error_cqe_handle
2096 				(txq, (volatile struct mlx5_err_cqe *)cqe);
2097 			/*
2098 			 * Flush buffers, update consuming index
2099 			 * if recovery succeeded. Otherwise
2100 			 * just try to recover later.
2101 			 */
2102 			last_cqe = NULL;
2103 			break;
2104 		}
2105 		/* Normal transmit completion. */
2106 		++txq->cq_ci;
2107 		last_cqe = cqe;
2108 #ifndef NDEBUG
2109 		if (txq->cq_pi)
2110 			--txq->cq_pi;
2111 #endif
2112 	/*
2113 	 * We have to restrict the amount of processed CQEs
2114 	 * in one tx_burst routine call. The CQ may be large
2115 	 * and many CQEs may be updated by the NIC in one
2116 	 * transaction. Buffers freeing is time consuming,
2117 	 * multiple iterations may introduce significant
2118 	 * latency.
2119 	 */
2120 	} while (--count);
2121 	mlx5_tx_comp_flush(txq, last_cqe, ret, olx);
2122 }
2123 
2124 /**
2125  * Check if the completion request flag should be set in the last WQE.
2126  * Both pushed mbufs and WQEs are monitored and the completion request
2127  * flag is set if any of thresholds is reached.
2128  *
2129  * @param txq
2130  *   Pointer to TX queue structure.
2131  * @param loc
2132  *   Pointer to burst routine local context.
2133  * @param multi,
2134  *   Routine is called from multi-segment sending loop,
2135  *   do not correct the elts_head according to the pkts_copy.
2136  * @param olx
2137  *   Configured Tx offloads mask. It is fully defined at
2138  *   compile time and may be used for optimization.
2139  */
2140 static __rte_always_inline void
2141 mlx5_tx_request_completion(struct mlx5_txq_data *restrict txq,
2142 			   struct mlx5_txq_local *restrict loc,
2143 			   bool multi,
2144 			   unsigned int olx)
2145 {
2146 	uint16_t head = txq->elts_head;
2147 	unsigned int part;
2148 
2149 	part = (MLX5_TXOFF_CONFIG(INLINE) || multi) ?
2150 	       0 : loc->pkts_sent - loc->pkts_copy;
2151 	head += part;
2152 	if ((uint16_t)(head - txq->elts_comp) >= MLX5_TX_COMP_THRESH ||
2153 	     (MLX5_TXOFF_CONFIG(INLINE) &&
2154 	     (uint16_t)(txq->wqe_ci - txq->wqe_comp) >= txq->wqe_thres)) {
2155 		volatile struct mlx5_wqe *last = loc->wqe_last;
2156 
2157 		txq->elts_comp = head;
2158 		if (MLX5_TXOFF_CONFIG(INLINE))
2159 			txq->wqe_comp = txq->wqe_ci;
2160 		/* Request unconditional completion on last WQE. */
2161 		last->cseg.flags = RTE_BE32(MLX5_COMP_ALWAYS <<
2162 					    MLX5_COMP_MODE_OFFSET);
2163 		/* Save elts_head in unused "immediate" field of WQE. */
2164 		last->cseg.misc = head;
2165 		/*
2166 		 * A CQE slot must always be available. Count the
2167 		 * issued CEQ "always" request instead of production
2168 		 * index due to here can be CQE with errors and
2169 		 * difference with ci may become inconsistent.
2170 		 */
2171 		assert(txq->cqe_s > ++txq->cq_pi);
2172 	}
2173 }
2174 
2175 /**
2176  * DPDK callback to check the status of a tx descriptor.
2177  *
2178  * @param tx_queue
2179  *   The tx queue.
2180  * @param[in] offset
2181  *   The index of the descriptor in the ring.
2182  *
2183  * @return
2184  *   The status of the tx descriptor.
2185  */
2186 int
2187 mlx5_tx_descriptor_status(void *tx_queue, uint16_t offset)
2188 {
2189 	struct mlx5_txq_data *restrict txq = tx_queue;
2190 	uint16_t used;
2191 
2192 	mlx5_tx_handle_completion(txq, 0);
2193 	used = txq->elts_head - txq->elts_tail;
2194 	if (offset < used)
2195 		return RTE_ETH_TX_DESC_FULL;
2196 	return RTE_ETH_TX_DESC_DONE;
2197 }
2198 
2199 /**
2200  * Build the Control Segment with specified opcode:
2201  * - MLX5_OPCODE_SEND
2202  * - MLX5_OPCODE_ENHANCED_MPSW
2203  * - MLX5_OPCODE_TSO
2204  *
2205  * @param txq
2206  *   Pointer to TX queue structure.
2207  * @param loc
2208  *   Pointer to burst routine local context.
2209  * @param wqe
2210  *   Pointer to WQE to fill with built Control Segment.
2211  * @param ds
2212  *   Supposed length of WQE in segments.
2213  * @param opcode
2214  *   SQ WQE opcode to put into Control Segment.
2215  * @param olx
2216  *   Configured Tx offloads mask. It is fully defined at
2217  *   compile time and may be used for optimization.
2218  */
2219 static __rte_always_inline void
2220 mlx5_tx_cseg_init(struct mlx5_txq_data *restrict txq,
2221 		  struct mlx5_txq_local *restrict loc __rte_unused,
2222 		  struct mlx5_wqe *restrict wqe,
2223 		  unsigned int ds,
2224 		  unsigned int opcode,
2225 		  unsigned int olx __rte_unused)
2226 {
2227 	struct mlx5_wqe_cseg *restrict cs = &wqe->cseg;
2228 
2229 	cs->opcode = rte_cpu_to_be_32((txq->wqe_ci << 8) | opcode);
2230 	cs->sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
2231 	cs->flags = RTE_BE32(MLX5_COMP_ONLY_FIRST_ERR <<
2232 			     MLX5_COMP_MODE_OFFSET);
2233 	cs->misc = RTE_BE32(0);
2234 }
2235 
2236 /**
2237  * Build the Ethernet Segment without inlined data.
2238  * Supports Software Parser, Checksums and VLAN
2239  * insertion Tx offload features.
2240  *
2241  * @param txq
2242  *   Pointer to TX queue structure.
2243  * @param loc
2244  *   Pointer to burst routine local context.
2245  * @param wqe
2246  *   Pointer to WQE to fill with built Ethernet Segment.
2247  * @param olx
2248  *   Configured Tx offloads mask. It is fully defined at
2249  *   compile time and may be used for optimization.
2250  */
2251 static __rte_always_inline void
2252 mlx5_tx_eseg_none(struct mlx5_txq_data *restrict txq __rte_unused,
2253 		  struct mlx5_txq_local *restrict loc,
2254 		  struct mlx5_wqe *restrict wqe,
2255 		  unsigned int olx)
2256 {
2257 	struct mlx5_wqe_eseg *restrict es = &wqe->eseg;
2258 	uint32_t csum;
2259 
2260 	/*
2261 	 * Calculate and set check sum flags first, dword field
2262 	 * in segment may be shared with Software Parser flags.
2263 	 */
2264 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2265 	es->flags = rte_cpu_to_le_32(csum);
2266 	/*
2267 	 * Calculate and set Software Parser offsets and flags.
2268 	 * These flags a set for custom UDP and IP tunnel packets.
2269 	 */
2270 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
2271 	/* Fill metadata field if needed. */
2272 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
2273 		       loc->mbuf->ol_flags & PKT_TX_METADATA ?
2274 		       loc->mbuf->tx_metadata : 0 : 0;
2275 	/* Engage VLAN tag insertion feature if requested. */
2276 	if (MLX5_TXOFF_CONFIG(VLAN) &&
2277 	    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
2278 		/*
2279 		 * We should get here only if device support
2280 		 * this feature correctly.
2281 		 */
2282 		assert(txq->vlan_en);
2283 		es->inline_hdr = rte_cpu_to_be_32(MLX5_ETH_WQE_VLAN_INSERT |
2284 						  loc->mbuf->vlan_tci);
2285 	} else {
2286 		es->inline_hdr = RTE_BE32(0);
2287 	}
2288 }
2289 
2290 /**
2291  * Build the Ethernet Segment with minimal inlined data
2292  * of MLX5_ESEG_MIN_INLINE_SIZE bytes length. This is
2293  * used to fill the gap in single WQEBB WQEs.
2294  * Supports Software Parser, Checksums and VLAN
2295  * insertion Tx offload features.
2296  *
2297  * @param txq
2298  *   Pointer to TX queue structure.
2299  * @param loc
2300  *   Pointer to burst routine local context.
2301  * @param wqe
2302  *   Pointer to WQE to fill with built Ethernet Segment.
2303  * @param vlan
2304  *   Length of VLAN tag insertion if any.
2305  * @param olx
2306  *   Configured Tx offloads mask. It is fully defined at
2307  *   compile time and may be used for optimization.
2308  */
2309 static __rte_always_inline void
2310 mlx5_tx_eseg_dmin(struct mlx5_txq_data *restrict txq __rte_unused,
2311 		  struct mlx5_txq_local *restrict loc,
2312 		  struct mlx5_wqe *restrict wqe,
2313 		  unsigned int vlan,
2314 		  unsigned int olx)
2315 {
2316 	struct mlx5_wqe_eseg *restrict es = &wqe->eseg;
2317 	uint32_t csum;
2318 	uint8_t *psrc, *pdst;
2319 
2320 	/*
2321 	 * Calculate and set check sum flags first, dword field
2322 	 * in segment may be shared with Software Parser flags.
2323 	 */
2324 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2325 	es->flags = rte_cpu_to_le_32(csum);
2326 	/*
2327 	 * Calculate and set Software Parser offsets and flags.
2328 	 * These flags a set for custom UDP and IP tunnel packets.
2329 	 */
2330 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
2331 	/* Fill metadata field if needed. */
2332 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
2333 		       loc->mbuf->ol_flags & PKT_TX_METADATA ?
2334 		       loc->mbuf->tx_metadata : 0 : 0;
2335 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2336 				(sizeof(uint16_t) +
2337 				 sizeof(rte_v128u32_t)),
2338 		      "invalid Ethernet Segment data size");
2339 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2340 				(sizeof(uint16_t) +
2341 				 sizeof(struct rte_vlan_hdr) +
2342 				 2 * RTE_ETHER_ADDR_LEN),
2343 		      "invalid Ethernet Segment data size");
2344 	psrc = rte_pktmbuf_mtod(loc->mbuf, uint8_t *);
2345 	es->inline_hdr_sz = RTE_BE16(MLX5_ESEG_MIN_INLINE_SIZE);
2346 	es->inline_data = *(unaligned_uint16_t *)psrc;
2347 	psrc +=	sizeof(uint16_t);
2348 	pdst = (uint8_t *)(es + 1);
2349 	if (MLX5_TXOFF_CONFIG(VLAN) && vlan) {
2350 		/* Implement VLAN tag insertion as part inline data. */
2351 		memcpy(pdst, psrc, 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t));
2352 		pdst += 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
2353 		psrc +=	2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
2354 		/* Insert VLAN ethertype + VLAN tag. */
2355 		*(unaligned_uint32_t *)pdst = rte_cpu_to_be_32
2356 						((RTE_ETHER_TYPE_VLAN << 16) |
2357 						 loc->mbuf->vlan_tci);
2358 		pdst += sizeof(struct rte_vlan_hdr);
2359 		/* Copy the rest two bytes from packet data. */
2360 		assert(pdst == RTE_PTR_ALIGN(pdst, sizeof(uint16_t)));
2361 		*(uint16_t *)pdst = *(unaligned_uint16_t *)psrc;
2362 	} else {
2363 		/* Fill the gap in the title WQEBB with inline data. */
2364 		rte_mov16(pdst, psrc);
2365 	}
2366 }
2367 
2368 /**
2369  * Build the Ethernet Segment with entire packet
2370  * data inlining. Checks the boundary of WQEBB and
2371  * ring buffer wrapping, supports Software Parser,
2372  * Checksums and VLAN insertion Tx offload features.
2373  *
2374  * @param txq
2375  *   Pointer to TX queue structure.
2376  * @param loc
2377  *   Pointer to burst routine local context.
2378  * @param wqe
2379  *   Pointer to WQE to fill with built Ethernet Segment.
2380  * @param vlan
2381  *   Length of VLAN tag insertion if any.
2382  * @param inlen
2383  *   Length of data to inline (VLAN included, if any).
2384  * @param tso
2385  *   TSO flag, set mss field from the packet.
2386  * @param olx
2387  *   Configured Tx offloads mask. It is fully defined at
2388  *   compile time and may be used for optimization.
2389  *
2390  * @return
2391  *   Pointer to the next Data Segment (aligned and wrapped around).
2392  */
2393 static __rte_always_inline struct mlx5_wqe_dseg *
2394 mlx5_tx_eseg_data(struct mlx5_txq_data *restrict txq,
2395 		  struct mlx5_txq_local *restrict loc,
2396 		  struct mlx5_wqe *restrict wqe,
2397 		  unsigned int vlan,
2398 		  unsigned int inlen,
2399 		  unsigned int tso,
2400 		  unsigned int olx)
2401 {
2402 	struct mlx5_wqe_eseg *restrict es = &wqe->eseg;
2403 	uint32_t csum;
2404 	uint8_t *psrc, *pdst;
2405 	unsigned int part;
2406 
2407 	/*
2408 	 * Calculate and set check sum flags first, dword field
2409 	 * in segment may be shared with Software Parser flags.
2410 	 */
2411 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2412 	if (tso) {
2413 		csum <<= 24;
2414 		csum |= loc->mbuf->tso_segsz;
2415 		es->flags = rte_cpu_to_be_32(csum);
2416 	} else {
2417 		es->flags = rte_cpu_to_le_32(csum);
2418 	}
2419 	/*
2420 	 * Calculate and set Software Parser offsets and flags.
2421 	 * These flags a set for custom UDP and IP tunnel packets.
2422 	 */
2423 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
2424 	/* Fill metadata field if needed. */
2425 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
2426 		       loc->mbuf->ol_flags & PKT_TX_METADATA ?
2427 		       loc->mbuf->tx_metadata : 0 : 0;
2428 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2429 				(sizeof(uint16_t) +
2430 				 sizeof(rte_v128u32_t)),
2431 		      "invalid Ethernet Segment data size");
2432 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2433 				(sizeof(uint16_t) +
2434 				 sizeof(struct rte_vlan_hdr) +
2435 				 2 * RTE_ETHER_ADDR_LEN),
2436 		      "invalid Ethernet Segment data size");
2437 	psrc = rte_pktmbuf_mtod(loc->mbuf, uint8_t *);
2438 	es->inline_hdr_sz = rte_cpu_to_be_16(inlen);
2439 	es->inline_data = *(unaligned_uint16_t *)psrc;
2440 	psrc +=	sizeof(uint16_t);
2441 	pdst = (uint8_t *)(es + 1);
2442 	if (MLX5_TXOFF_CONFIG(VLAN) && vlan) {
2443 		/* Implement VLAN tag insertion as part inline data. */
2444 		memcpy(pdst, psrc, 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t));
2445 		pdst += 2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
2446 		psrc +=	2 * RTE_ETHER_ADDR_LEN - sizeof(uint16_t);
2447 		/* Insert VLAN ethertype + VLAN tag. */
2448 		*(unaligned_uint32_t *)pdst = rte_cpu_to_be_32
2449 						((RTE_ETHER_TYPE_VLAN << 16) |
2450 						 loc->mbuf->vlan_tci);
2451 		pdst += sizeof(struct rte_vlan_hdr);
2452 		/* Copy the rest two bytes from packet data. */
2453 		assert(pdst == RTE_PTR_ALIGN(pdst, sizeof(uint16_t)));
2454 		*(uint16_t *)pdst = *(unaligned_uint16_t *)psrc;
2455 		psrc += sizeof(uint16_t);
2456 	} else {
2457 		/* Fill the gap in the title WQEBB with inline data. */
2458 		rte_mov16(pdst, psrc);
2459 		psrc += sizeof(rte_v128u32_t);
2460 	}
2461 	pdst = (uint8_t *)(es + 2);
2462 	assert(inlen >= MLX5_ESEG_MIN_INLINE_SIZE);
2463 	assert(pdst < (uint8_t *)txq->wqes_end);
2464 	inlen -= MLX5_ESEG_MIN_INLINE_SIZE;
2465 	if (!inlen) {
2466 		assert(pdst == RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE));
2467 		return (struct mlx5_wqe_dseg *)pdst;
2468 	}
2469 	/*
2470 	 * The WQEBB space availability is checked by caller.
2471 	 * Here we should be aware of WQE ring buffer wraparound only.
2472 	 */
2473 	part = (uint8_t *)txq->wqes_end - pdst;
2474 	part = RTE_MIN(part, inlen);
2475 	do {
2476 		rte_memcpy(pdst, psrc, part);
2477 		inlen -= part;
2478 		if (likely(!inlen)) {
2479 			/*
2480 			 * If return value is not used by the caller
2481 			 * the code below will be optimized out.
2482 			 */
2483 			pdst += part;
2484 			pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
2485 			if (unlikely(pdst >= (uint8_t *)txq->wqes_end))
2486 				pdst = (uint8_t *)txq->wqes;
2487 			return (struct mlx5_wqe_dseg *)pdst;
2488 		}
2489 		pdst = (uint8_t *)txq->wqes;
2490 		psrc += part;
2491 		part = inlen;
2492 	} while (true);
2493 }
2494 
2495 /**
2496  * Copy data from chain of mbuf to the specified linear buffer.
2497  * Checksums and VLAN insertion Tx offload features. If data
2498  * from some mbuf copied completely this mbuf is freed. Local
2499  * structure is used to keep the byte stream state.
2500  *
2501  * @param pdst
2502  *   Pointer to the destination linear buffer.
2503  * @param loc
2504  *   Pointer to burst routine local context.
2505  * @param len
2506  *   Length of data to be copied.
2507  * @param olx
2508  *   Configured Tx offloads mask. It is fully defined at
2509  *   compile time and may be used for optimization.
2510  */
2511 static __rte_always_inline void
2512 mlx5_tx_mseg_memcpy(uint8_t *pdst,
2513 		    struct mlx5_txq_local *restrict loc,
2514 		    unsigned int len,
2515 		    unsigned int olx __rte_unused)
2516 {
2517 	struct rte_mbuf *mbuf;
2518 	unsigned int part, dlen;
2519 	uint8_t *psrc;
2520 
2521 	assert(len);
2522 	do {
2523 		/* Allow zero length packets, must check first. */
2524 		dlen = rte_pktmbuf_data_len(loc->mbuf);
2525 		if (dlen <= loc->mbuf_off) {
2526 			/* Exhausted packet, just free. */
2527 			mbuf = loc->mbuf;
2528 			loc->mbuf = mbuf->next;
2529 			rte_pktmbuf_free_seg(mbuf);
2530 			loc->mbuf_off = 0;
2531 			assert(loc->mbuf_nseg > 1);
2532 			assert(loc->mbuf);
2533 			--loc->mbuf_nseg;
2534 			continue;
2535 		}
2536 		dlen -= loc->mbuf_off;
2537 		psrc = rte_pktmbuf_mtod_offset(loc->mbuf, uint8_t *,
2538 					       loc->mbuf_off);
2539 		part = RTE_MIN(len, dlen);
2540 		rte_memcpy(pdst, psrc, part);
2541 		loc->mbuf_off += part;
2542 		len -= part;
2543 		if (!len) {
2544 			if (loc->mbuf_off >= rte_pktmbuf_data_len(loc->mbuf)) {
2545 				loc->mbuf_off = 0;
2546 				/* Exhausted packet, just free. */
2547 				mbuf = loc->mbuf;
2548 				loc->mbuf = mbuf->next;
2549 				rte_pktmbuf_free_seg(mbuf);
2550 				loc->mbuf_off = 0;
2551 				assert(loc->mbuf_nseg >= 1);
2552 				--loc->mbuf_nseg;
2553 			}
2554 			return;
2555 		}
2556 		pdst += part;
2557 	} while (true);
2558 }
2559 
2560 /**
2561  * Build the Ethernet Segment with inlined data from
2562  * multi-segment packet. Checks the boundary of WQEBB
2563  * and ring buffer wrapping, supports Software Parser,
2564  * Checksums and VLAN insertion Tx offload features.
2565  *
2566  * @param txq
2567  *   Pointer to TX queue structure.
2568  * @param loc
2569  *   Pointer to burst routine local context.
2570  * @param wqe
2571  *   Pointer to WQE to fill with built Ethernet Segment.
2572  * @param vlan
2573  *   Length of VLAN tag insertion if any.
2574  * @param inlen
2575  *   Length of data to inline (VLAN included, if any).
2576  * @param tso
2577  *   TSO flag, set mss field from the packet.
2578  * @param olx
2579  *   Configured Tx offloads mask. It is fully defined at
2580  *   compile time and may be used for optimization.
2581  *
2582  * @return
2583  *   Pointer to the next Data Segment (aligned and
2584  *   possible NOT wrapped around - caller should do
2585  *   wrapping check on its own).
2586  */
2587 static __rte_always_inline struct mlx5_wqe_dseg *
2588 mlx5_tx_eseg_mdat(struct mlx5_txq_data *restrict txq,
2589 		  struct mlx5_txq_local *restrict loc,
2590 		  struct mlx5_wqe *restrict wqe,
2591 		  unsigned int vlan,
2592 		  unsigned int inlen,
2593 		  unsigned int tso,
2594 		  unsigned int olx)
2595 {
2596 	struct mlx5_wqe_eseg *restrict es = &wqe->eseg;
2597 	uint32_t csum;
2598 	uint8_t *pdst;
2599 	unsigned int part;
2600 
2601 	/*
2602 	 * Calculate and set check sum flags first, uint32_t field
2603 	 * in segment may be shared with Software Parser flags.
2604 	 */
2605 	csum = MLX5_TXOFF_CONFIG(CSUM) ? txq_ol_cksum_to_cs(loc->mbuf) : 0;
2606 	if (tso) {
2607 		csum <<= 24;
2608 		csum |= loc->mbuf->tso_segsz;
2609 		es->flags = rte_cpu_to_be_32(csum);
2610 	} else {
2611 		es->flags = rte_cpu_to_le_32(csum);
2612 	}
2613 	/*
2614 	 * Calculate and set Software Parser offsets and flags.
2615 	 * These flags a set for custom UDP and IP tunnel packets.
2616 	 */
2617 	es->swp_offs = txq_mbuf_to_swp(loc, &es->swp_flags, olx);
2618 	/* Fill metadata field if needed. */
2619 	es->metadata = MLX5_TXOFF_CONFIG(METADATA) ?
2620 		       loc->mbuf->ol_flags & PKT_TX_METADATA ?
2621 		       loc->mbuf->tx_metadata : 0 : 0;
2622 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2623 				(sizeof(uint16_t) +
2624 				 sizeof(rte_v128u32_t)),
2625 		      "invalid Ethernet Segment data size");
2626 	static_assert(MLX5_ESEG_MIN_INLINE_SIZE ==
2627 				(sizeof(uint16_t) +
2628 				 sizeof(struct rte_vlan_hdr) +
2629 				 2 * RTE_ETHER_ADDR_LEN),
2630 		      "invalid Ethernet Segment data size");
2631 	assert(inlen >= MLX5_ESEG_MIN_INLINE_SIZE);
2632 	es->inline_hdr_sz = rte_cpu_to_be_16(inlen);
2633 	pdst = (uint8_t *)&es->inline_data;
2634 	if (MLX5_TXOFF_CONFIG(VLAN) && vlan) {
2635 		/* Implement VLAN tag insertion as part inline data. */
2636 		mlx5_tx_mseg_memcpy(pdst, loc, 2 * RTE_ETHER_ADDR_LEN, olx);
2637 		pdst += 2 * RTE_ETHER_ADDR_LEN;
2638 		*(unaligned_uint32_t *)pdst = rte_cpu_to_be_32
2639 						((RTE_ETHER_TYPE_VLAN << 16) |
2640 						 loc->mbuf->vlan_tci);
2641 		pdst += sizeof(struct rte_vlan_hdr);
2642 		inlen -= 2 * RTE_ETHER_ADDR_LEN + sizeof(struct rte_vlan_hdr);
2643 	}
2644 	assert(pdst < (uint8_t *)txq->wqes_end);
2645 	/*
2646 	 * The WQEBB space availability is checked by caller.
2647 	 * Here we should be aware of WQE ring buffer wraparound only.
2648 	 */
2649 	part = (uint8_t *)txq->wqes_end - pdst;
2650 	part = RTE_MIN(part, inlen);
2651 	assert(part);
2652 	do {
2653 		mlx5_tx_mseg_memcpy(pdst, loc, part, olx);
2654 		inlen -= part;
2655 		if (likely(!inlen)) {
2656 			pdst += part;
2657 			pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
2658 			return (struct mlx5_wqe_dseg *)pdst;
2659 		}
2660 		pdst = (uint8_t *)txq->wqes;
2661 		part = inlen;
2662 	} while (true);
2663 }
2664 
2665 /**
2666  * Build the Data Segment of pointer type.
2667  *
2668  * @param txq
2669  *   Pointer to TX queue structure.
2670  * @param loc
2671  *   Pointer to burst routine local context.
2672  * @param dseg
2673  *   Pointer to WQE to fill with built Data Segment.
2674  * @param buf
2675  *   Data buffer to point.
2676  * @param len
2677  *   Data buffer length.
2678  * @param olx
2679  *   Configured Tx offloads mask. It is fully defined at
2680  *   compile time and may be used for optimization.
2681  */
2682 static __rte_always_inline void
2683 mlx5_tx_dseg_ptr(struct mlx5_txq_data *restrict txq,
2684 		 struct mlx5_txq_local *restrict loc,
2685 		 struct mlx5_wqe_dseg *restrict dseg,
2686 		 uint8_t *buf,
2687 		 unsigned int len,
2688 		 unsigned int olx __rte_unused)
2689 
2690 {
2691 	assert(len);
2692 	dseg->bcount = rte_cpu_to_be_32(len);
2693 	dseg->lkey = mlx5_tx_mb2mr(txq, loc->mbuf);
2694 	dseg->pbuf = rte_cpu_to_be_64((uintptr_t)buf);
2695 }
2696 
2697 /**
2698  * Build the Data Segment of pointer type or inline
2699  * if data length is less than buffer in minimal
2700  * Data Segment size.
2701  *
2702  * @param txq
2703  *   Pointer to TX queue structure.
2704  * @param loc
2705  *   Pointer to burst routine local context.
2706  * @param dseg
2707  *   Pointer to WQE to fill with built Data Segment.
2708  * @param buf
2709  *   Data buffer to point.
2710  * @param len
2711  *   Data buffer length.
2712  * @param olx
2713  *   Configured Tx offloads mask. It is fully defined at
2714  *   compile time and may be used for optimization.
2715  */
2716 static __rte_always_inline void
2717 mlx5_tx_dseg_iptr(struct mlx5_txq_data *restrict txq,
2718 		  struct mlx5_txq_local *restrict loc,
2719 		  struct mlx5_wqe_dseg *restrict dseg,
2720 		  uint8_t *buf,
2721 		  unsigned int len,
2722 		  unsigned int olx __rte_unused)
2723 
2724 {
2725 	uintptr_t dst, src;
2726 
2727 	assert(len);
2728 	if (len > MLX5_DSEG_MIN_INLINE_SIZE) {
2729 		dseg->bcount = rte_cpu_to_be_32(len);
2730 		dseg->lkey = mlx5_tx_mb2mr(txq, loc->mbuf);
2731 		dseg->pbuf = rte_cpu_to_be_64((uintptr_t)buf);
2732 
2733 		return;
2734 	}
2735 	dseg->bcount = rte_cpu_to_be_32(len | MLX5_ETH_WQE_DATA_INLINE);
2736 	/* Unrolled implementation of generic rte_memcpy. */
2737 	dst = (uintptr_t)&dseg->inline_data[0];
2738 	src = (uintptr_t)buf;
2739 #ifdef RTE_ARCH_STRICT_ALIGN
2740 	memcpy(dst, src, len);
2741 #else
2742 	if (len & 0x08) {
2743 		*(uint64_t *)dst = *(uint64_t *)src;
2744 		dst += sizeof(uint64_t);
2745 		src += sizeof(uint64_t);
2746 	}
2747 	if (len & 0x04) {
2748 		*(uint32_t *)dst = *(uint32_t *)src;
2749 		dst += sizeof(uint32_t);
2750 		src += sizeof(uint32_t);
2751 	}
2752 	if (len & 0x02) {
2753 		*(uint16_t *)dst = *(uint16_t *)src;
2754 		dst += sizeof(uint16_t);
2755 		src += sizeof(uint16_t);
2756 	}
2757 	if (len & 0x01)
2758 		*(uint8_t *)dst = *(uint8_t *)src;
2759 #endif
2760 }
2761 
2762 /**
2763  * Build the Data Segment of inlined data from single
2764  * segment packet, no VLAN insertion.
2765  *
2766  * @param txq
2767  *   Pointer to TX queue structure.
2768  * @param loc
2769  *   Pointer to burst routine local context.
2770  * @param dseg
2771  *   Pointer to WQE to fill with built Data Segment.
2772  * @param buf
2773  *   Data buffer to point.
2774  * @param len
2775  *   Data buffer length.
2776  * @param olx
2777  *   Configured Tx offloads mask. It is fully defined at
2778  *   compile time and may be used for optimization.
2779  *
2780  * @return
2781  *   Pointer to the next Data Segment after inlined data.
2782  *   Ring buffer wraparound check is needed. We do not
2783  *   do it here because it may not be needed for the
2784  *   last packet in the eMPW session.
2785  */
2786 static __rte_always_inline struct mlx5_wqe_dseg *
2787 mlx5_tx_dseg_empw(struct mlx5_txq_data *restrict txq,
2788 		  struct mlx5_txq_local *restrict loc __rte_unused,
2789 		  struct mlx5_wqe_dseg *restrict dseg,
2790 		  uint8_t *buf,
2791 		  unsigned int len,
2792 		  unsigned int olx __rte_unused)
2793 {
2794 	unsigned int part;
2795 	uint8_t *pdst;
2796 
2797 	dseg->bcount = rte_cpu_to_be_32(len | MLX5_ETH_WQE_DATA_INLINE);
2798 	pdst = &dseg->inline_data[0];
2799 	/*
2800 	 * The WQEBB space availability is checked by caller.
2801 	 * Here we should be aware of WQE ring buffer wraparound only.
2802 	 */
2803 	part = (uint8_t *)txq->wqes_end - pdst;
2804 	part = RTE_MIN(part, len);
2805 	do {
2806 		rte_memcpy(pdst, buf, part);
2807 		len -= part;
2808 		if (likely(!len)) {
2809 			pdst += part;
2810 			pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
2811 			/* Note: no final wraparound check here. */
2812 			return (struct mlx5_wqe_dseg *)pdst;
2813 		}
2814 		pdst = (uint8_t *)txq->wqes;
2815 		buf += part;
2816 		part = len;
2817 	} while (true);
2818 }
2819 
2820 /**
2821  * Build the Data Segment of inlined data from single
2822  * segment packet with VLAN insertion.
2823  *
2824  * @param txq
2825  *   Pointer to TX queue structure.
2826  * @param loc
2827  *   Pointer to burst routine local context.
2828  * @param dseg
2829  *   Pointer to the dseg fill with built Data Segment.
2830  * @param buf
2831  *   Data buffer to point.
2832  * @param len
2833  *   Data buffer length.
2834  * @param olx
2835  *   Configured Tx offloads mask. It is fully defined at
2836  *   compile time and may be used for optimization.
2837  *
2838  * @return
2839  *   Pointer to the next Data Segment after inlined data.
2840  *   Ring buffer wraparound check is needed.
2841  */
2842 static __rte_always_inline struct mlx5_wqe_dseg *
2843 mlx5_tx_dseg_vlan(struct mlx5_txq_data *restrict txq,
2844 		  struct mlx5_txq_local *restrict loc __rte_unused,
2845 		  struct mlx5_wqe_dseg *restrict dseg,
2846 		  uint8_t *buf,
2847 		  unsigned int len,
2848 		  unsigned int olx __rte_unused)
2849 
2850 {
2851 	unsigned int part;
2852 	uint8_t *pdst;
2853 
2854 	assert(len > MLX5_ESEG_MIN_INLINE_SIZE);
2855 	static_assert(MLX5_DSEG_MIN_INLINE_SIZE ==
2856 				 (2 * RTE_ETHER_ADDR_LEN),
2857 		      "invalid Data Segment data size");
2858 	dseg->bcount = rte_cpu_to_be_32((len + sizeof(struct rte_vlan_hdr)) |
2859 					MLX5_ETH_WQE_DATA_INLINE);
2860 	pdst = &dseg->inline_data[0];
2861 	memcpy(pdst, buf, MLX5_DSEG_MIN_INLINE_SIZE);
2862 	buf += MLX5_DSEG_MIN_INLINE_SIZE;
2863 	pdst += MLX5_DSEG_MIN_INLINE_SIZE;
2864 	len -= MLX5_DSEG_MIN_INLINE_SIZE;
2865 	/* Insert VLAN ethertype + VLAN tag. Pointer is aligned. */
2866 	assert(pdst == RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE));
2867 	if (unlikely(pdst >= (uint8_t *)txq->wqes_end))
2868 		pdst = (uint8_t *)txq->wqes;
2869 	*(uint32_t *)pdst = rte_cpu_to_be_32((RTE_ETHER_TYPE_VLAN << 16) |
2870 					      loc->mbuf->vlan_tci);
2871 	pdst += sizeof(struct rte_vlan_hdr);
2872 	/*
2873 	 * The WQEBB space availability is checked by caller.
2874 	 * Here we should be aware of WQE ring buffer wraparound only.
2875 	 */
2876 	part = (uint8_t *)txq->wqes_end - pdst;
2877 	part = RTE_MIN(part, len);
2878 	do {
2879 		rte_memcpy(pdst, buf, part);
2880 		len -= part;
2881 		if (likely(!len)) {
2882 			pdst += part;
2883 			pdst = RTE_PTR_ALIGN(pdst, MLX5_WSEG_SIZE);
2884 			/* Note: no final wraparound check here. */
2885 			return (struct mlx5_wqe_dseg *)pdst;
2886 		}
2887 		pdst = (uint8_t *)txq->wqes;
2888 		buf += part;
2889 		part = len;
2890 	} while (true);
2891 }
2892 
2893 /**
2894  * Build the Ethernet Segment with optionally inlined data with
2895  * VLAN insertion and following Data Segments (if any) from
2896  * multi-segment packet. Used by ordinary send and TSO.
2897  *
2898  * @param txq
2899  *   Pointer to TX queue structure.
2900  * @param loc
2901  *   Pointer to burst routine local context.
2902  * @param wqe
2903  *   Pointer to WQE to fill with built Ethernet/Data Segments.
2904  * @param vlan
2905  *   Length of VLAN header to insert, 0 means no VLAN insertion.
2906  * @param inlen
2907  *   Data length to inline. For TSO this parameter specifies
2908  *   exact value, for ordinary send routine can be aligned by
2909  *   caller to provide better WQE space saving and data buffer
2910  *   start address alignment. This length includes VLAN header
2911  *   being inserted.
2912  * @param tso
2913  *   Zero means ordinary send, inlined data can be extended,
2914  *   otherwise this is TSO, inlined data length is fixed.
2915  * @param olx
2916  *   Configured Tx offloads mask. It is fully defined at
2917  *   compile time and may be used for optimization.
2918  *
2919  * @return
2920  *   Actual size of built WQE in segments.
2921  */
2922 static __rte_always_inline unsigned int
2923 mlx5_tx_mseg_build(struct mlx5_txq_data *restrict txq,
2924 		   struct mlx5_txq_local *restrict loc,
2925 		   struct mlx5_wqe *restrict wqe,
2926 		   unsigned int vlan,
2927 		   unsigned int inlen,
2928 		   unsigned int tso,
2929 		   unsigned int olx __rte_unused)
2930 {
2931 	struct mlx5_wqe_dseg *restrict dseg;
2932 	unsigned int ds;
2933 
2934 	assert((rte_pktmbuf_pkt_len(loc->mbuf) + vlan) >= inlen);
2935 	loc->mbuf_nseg = NB_SEGS(loc->mbuf);
2936 	loc->mbuf_off = 0;
2937 
2938 	dseg = mlx5_tx_eseg_mdat(txq, loc, wqe, vlan, inlen, tso, olx);
2939 	if (!loc->mbuf_nseg)
2940 		goto dseg_done;
2941 	/*
2942 	 * There are still some mbuf remaining, not inlined.
2943 	 * The first mbuf may be partially inlined and we
2944 	 * must process the possible non-zero data offset.
2945 	 */
2946 	if (loc->mbuf_off) {
2947 		unsigned int dlen;
2948 		uint8_t *dptr;
2949 
2950 		/*
2951 		 * Exhausted packets must be dropped before.
2952 		 * Non-zero offset means there are some data
2953 		 * remained in the packet.
2954 		 */
2955 		assert(loc->mbuf_off < rte_pktmbuf_data_len(loc->mbuf));
2956 		assert(rte_pktmbuf_data_len(loc->mbuf));
2957 		dptr = rte_pktmbuf_mtod_offset(loc->mbuf, uint8_t *,
2958 					       loc->mbuf_off);
2959 		dlen = rte_pktmbuf_data_len(loc->mbuf) - loc->mbuf_off;
2960 		/*
2961 		 * Build the pointer/minimal data Data Segment.
2962 		 * Do ring buffer wrapping check in advance.
2963 		 */
2964 		if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
2965 			dseg = (struct mlx5_wqe_dseg *)txq->wqes;
2966 		mlx5_tx_dseg_iptr(txq, loc, dseg, dptr, dlen, olx);
2967 		/* Store the mbuf to be freed on completion. */
2968 		assert(loc->elts_free);
2969 		txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
2970 		--loc->elts_free;
2971 		++dseg;
2972 		if (--loc->mbuf_nseg == 0)
2973 			goto dseg_done;
2974 		loc->mbuf = loc->mbuf->next;
2975 		loc->mbuf_off = 0;
2976 	}
2977 	do {
2978 		if (unlikely(!rte_pktmbuf_data_len(loc->mbuf))) {
2979 			struct rte_mbuf *mbuf;
2980 
2981 			/* Zero length segment found, just skip. */
2982 			mbuf = loc->mbuf;
2983 			loc->mbuf = loc->mbuf->next;
2984 			rte_pktmbuf_free_seg(mbuf);
2985 			if (--loc->mbuf_nseg == 0)
2986 				break;
2987 		} else {
2988 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
2989 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
2990 			mlx5_tx_dseg_iptr
2991 				(txq, loc, dseg,
2992 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
2993 				 rte_pktmbuf_data_len(loc->mbuf), olx);
2994 			assert(loc->elts_free);
2995 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
2996 			--loc->elts_free;
2997 			++dseg;
2998 			if (--loc->mbuf_nseg == 0)
2999 				break;
3000 			loc->mbuf = loc->mbuf->next;
3001 		}
3002 	} while (true);
3003 
3004 dseg_done:
3005 	/* Calculate actual segments used from the dseg pointer. */
3006 	if ((uintptr_t)wqe < (uintptr_t)dseg)
3007 		ds = ((uintptr_t)dseg - (uintptr_t)wqe) / MLX5_WSEG_SIZE;
3008 	else
3009 		ds = (((uintptr_t)dseg - (uintptr_t)wqe) +
3010 		      txq->wqe_s * MLX5_WQE_SIZE) / MLX5_WSEG_SIZE;
3011 	return ds;
3012 }
3013 
3014 /**
3015  * Tx one packet function for multi-segment TSO. Supports all
3016  * types of Tx offloads, uses MLX5_OPCODE_TSO to build WQEs,
3017  * sends one packet per WQE.
3018  *
3019  * This routine is responsible for storing processed mbuf
3020  * into elts ring buffer and update elts_head.
3021  *
3022  * @param txq
3023  *   Pointer to TX queue structure.
3024  * @param loc
3025  *   Pointer to burst routine local context.
3026  * @param olx
3027  *   Configured Tx offloads mask. It is fully defined at
3028  *   compile time and may be used for optimization.
3029  *
3030  * @return
3031  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3032  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3033  * Local context variables partially updated.
3034  */
3035 static __rte_always_inline enum mlx5_txcmp_code
3036 mlx5_tx_packet_multi_tso(struct mlx5_txq_data *restrict txq,
3037 			struct mlx5_txq_local *restrict loc,
3038 			unsigned int olx)
3039 {
3040 	struct mlx5_wqe *restrict wqe;
3041 	unsigned int ds, dlen, inlen, ntcp, vlan = 0;
3042 
3043 	/*
3044 	 * Calculate data length to be inlined to estimate
3045 	 * the required space in WQE ring buffer.
3046 	 */
3047 	dlen = rte_pktmbuf_pkt_len(loc->mbuf);
3048 	if (MLX5_TXOFF_CONFIG(VLAN) && loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
3049 		vlan = sizeof(struct rte_vlan_hdr);
3050 	inlen = loc->mbuf->l2_len + vlan +
3051 		loc->mbuf->l3_len + loc->mbuf->l4_len;
3052 	if (unlikely((!inlen || !loc->mbuf->tso_segsz)))
3053 		return MLX5_TXCMP_CODE_ERROR;
3054 	if (loc->mbuf->ol_flags & PKT_TX_TUNNEL_MASK)
3055 		inlen += loc->mbuf->outer_l2_len + loc->mbuf->outer_l3_len;
3056 	/* Packet must contain all TSO headers. */
3057 	if (unlikely(inlen > MLX5_MAX_TSO_HEADER ||
3058 		     inlen <= MLX5_ESEG_MIN_INLINE_SIZE ||
3059 		     inlen > (dlen + vlan)))
3060 		return MLX5_TXCMP_CODE_ERROR;
3061 	assert(inlen >= txq->inlen_mode);
3062 	/*
3063 	 * Check whether there are enough free WQEBBs:
3064 	 * - Control Segment
3065 	 * - Ethernet Segment
3066 	 * - First Segment of inlined Ethernet data
3067 	 * - ... data continued ...
3068 	 * - Data Segments of pointer/min inline type
3069 	 */
3070 	ds = NB_SEGS(loc->mbuf) + 2 + (inlen -
3071 				       MLX5_ESEG_MIN_INLINE_SIZE +
3072 				       MLX5_WSEG_SIZE +
3073 				       MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
3074 	if (unlikely(loc->wqe_free < ((ds + 3) / 4)))
3075 		return MLX5_TXCMP_CODE_EXIT;
3076 	/* Check for maximal WQE size. */
3077 	if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4)))
3078 		return MLX5_TXCMP_CODE_ERROR;
3079 #ifdef MLX5_PMD_SOFT_COUNTERS
3080 	/* Update sent data bytes/packets counters. */
3081 	ntcp = (dlen - (inlen - vlan) + loc->mbuf->tso_segsz - 1) /
3082 		loc->mbuf->tso_segsz;
3083 	/*
3084 	 * One will be added for mbuf itself
3085 	 * at the end of the mlx5_tx_burst from
3086 	 * loc->pkts_sent field.
3087 	 */
3088 	--ntcp;
3089 	txq->stats.opackets += ntcp;
3090 	txq->stats.obytes += dlen + vlan + ntcp * inlen;
3091 #endif
3092 	wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3093 	loc->wqe_last = wqe;
3094 	mlx5_tx_cseg_init(txq, loc, wqe, 0, MLX5_OPCODE_TSO, olx);
3095 	ds = mlx5_tx_mseg_build(txq, loc, wqe, vlan, inlen, 1, olx);
3096 	wqe->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
3097 	txq->wqe_ci += (ds + 3) / 4;
3098 	loc->wqe_free -= (ds + 3) / 4;
3099 	/* Request CQE generation if limits are reached. */
3100 	mlx5_tx_request_completion(txq, loc, true, olx);
3101 	return MLX5_TXCMP_CODE_MULTI;
3102 }
3103 
3104 /**
3105  * Tx one packet function for multi-segment SEND. Supports all
3106  * types of Tx offloads, uses MLX5_OPCODE_SEND to build WQEs,
3107  * sends one packet per WQE, without any data inlining in
3108  * Ethernet Segment.
3109  *
3110  * This routine is responsible for storing processed mbuf
3111  * into elts ring buffer and update elts_head.
3112  *
3113  * @param txq
3114  *   Pointer to TX queue structure.
3115  * @param loc
3116  *   Pointer to burst routine local context.
3117  * @param olx
3118  *   Configured Tx offloads mask. It is fully defined at
3119  *   compile time and may be used for optimization.
3120  *
3121  * @return
3122  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3123  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3124  * Local context variables partially updated.
3125  */
3126 static __rte_always_inline enum mlx5_txcmp_code
3127 mlx5_tx_packet_multi_send(struct mlx5_txq_data *restrict txq,
3128 			  struct mlx5_txq_local *restrict loc,
3129 			  unsigned int olx)
3130 {
3131 	struct mlx5_wqe_dseg *restrict dseg;
3132 	struct mlx5_wqe *restrict wqe;
3133 	unsigned int ds, nseg;
3134 
3135 	assert(NB_SEGS(loc->mbuf) > 1);
3136 	/*
3137 	 * No inline at all, it means the CPU cycles saving
3138 	 * is prioritized at configuration, we should not
3139 	 * copy any packet data to WQE.
3140 	 */
3141 	nseg = NB_SEGS(loc->mbuf);
3142 	ds = 2 + nseg;
3143 	if (unlikely(loc->wqe_free < ((ds + 3) / 4)))
3144 		return MLX5_TXCMP_CODE_EXIT;
3145 	/* Check for maximal WQE size. */
3146 	if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4)))
3147 		return MLX5_TXCMP_CODE_ERROR;
3148 	/*
3149 	 * Some Tx offloads may cause an error if
3150 	 * packet is not long enough, check against
3151 	 * assumed minimal length.
3152 	 */
3153 	if (rte_pktmbuf_pkt_len(loc->mbuf) <= MLX5_ESEG_MIN_INLINE_SIZE)
3154 		return MLX5_TXCMP_CODE_ERROR;
3155 #ifdef MLX5_PMD_SOFT_COUNTERS
3156 	/* Update sent data bytes counter. */
3157 	txq->stats.obytes += rte_pktmbuf_pkt_len(loc->mbuf);
3158 	if (MLX5_TXOFF_CONFIG(VLAN) &&
3159 	    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
3160 		txq->stats.obytes += sizeof(struct rte_vlan_hdr);
3161 #endif
3162 	/*
3163 	 * SEND WQE, one WQEBB:
3164 	 * - Control Segment, SEND opcode
3165 	 * - Ethernet Segment, optional VLAN, no inline
3166 	 * - Data Segments, pointer only type
3167 	 */
3168 	wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3169 	loc->wqe_last = wqe;
3170 	mlx5_tx_cseg_init(txq, loc, wqe, ds, MLX5_OPCODE_SEND, olx);
3171 	mlx5_tx_eseg_none(txq, loc, wqe, olx);
3172 	dseg = &wqe->dseg[0];
3173 	do {
3174 		if (unlikely(!rte_pktmbuf_data_len(loc->mbuf))) {
3175 			struct rte_mbuf *mbuf;
3176 
3177 			/*
3178 			 * Zero length segment found, have to
3179 			 * correct total size of WQE in segments.
3180 			 * It is supposed to be rare occasion, so
3181 			 * in normal case (no zero length segments)
3182 			 * we avoid extra writing to the Control
3183 			 * Segment.
3184 			 */
3185 			--ds;
3186 			wqe->cseg.sq_ds -= RTE_BE32(1);
3187 			mbuf = loc->mbuf;
3188 			loc->mbuf = mbuf->next;
3189 			rte_pktmbuf_free_seg(mbuf);
3190 			if (--nseg == 0)
3191 				break;
3192 		} else {
3193 			mlx5_tx_dseg_ptr
3194 				(txq, loc, dseg,
3195 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
3196 				 rte_pktmbuf_data_len(loc->mbuf), olx);
3197 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
3198 			--loc->elts_free;
3199 			if (--nseg == 0)
3200 				break;
3201 			++dseg;
3202 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
3203 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
3204 			loc->mbuf = loc->mbuf->next;
3205 		}
3206 	} while (true);
3207 	txq->wqe_ci += (ds + 3) / 4;
3208 	loc->wqe_free -= (ds + 3) / 4;
3209 	/* Request CQE generation if limits are reached. */
3210 	mlx5_tx_request_completion(txq, loc, true, olx);
3211 	return MLX5_TXCMP_CODE_MULTI;
3212 }
3213 
3214 /**
3215  * Tx one packet function for multi-segment SEND. Supports all
3216  * types of Tx offloads, uses MLX5_OPCODE_SEND to build WQEs,
3217  * sends one packet per WQE, with data inlining in
3218  * Ethernet Segment and minimal Data Segments.
3219  *
3220  * This routine is responsible for storing processed mbuf
3221  * into elts ring buffer and update elts_head.
3222  *
3223  * @param txq
3224  *   Pointer to TX queue structure.
3225  * @param loc
3226  *   Pointer to burst routine local context.
3227  * @param olx
3228  *   Configured Tx offloads mask. It is fully defined at
3229  *   compile time and may be used for optimization.
3230  *
3231  * @return
3232  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3233  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3234  * Local context variables partially updated.
3235  */
3236 static __rte_always_inline enum mlx5_txcmp_code
3237 mlx5_tx_packet_multi_inline(struct mlx5_txq_data *restrict txq,
3238 			    struct mlx5_txq_local *restrict loc,
3239 			    unsigned int olx)
3240 {
3241 	struct mlx5_wqe *restrict wqe;
3242 	unsigned int ds, inlen, dlen, vlan = 0;
3243 
3244 	assert(MLX5_TXOFF_CONFIG(INLINE));
3245 	assert(NB_SEGS(loc->mbuf) > 1);
3246 	/*
3247 	 * First calculate data length to be inlined
3248 	 * to estimate the required space for WQE.
3249 	 */
3250 	dlen = rte_pktmbuf_pkt_len(loc->mbuf);
3251 	if (MLX5_TXOFF_CONFIG(VLAN) && loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
3252 		vlan = sizeof(struct rte_vlan_hdr);
3253 	inlen = dlen + vlan;
3254 	/* Check against minimal length. */
3255 	if (inlen <= MLX5_ESEG_MIN_INLINE_SIZE)
3256 		return MLX5_TXCMP_CODE_ERROR;
3257 	assert(txq->inlen_send >= MLX5_ESEG_MIN_INLINE_SIZE);
3258 	if (inlen > txq->inlen_send) {
3259 		struct rte_mbuf *mbuf;
3260 		unsigned int nxlen;
3261 		uintptr_t start;
3262 
3263 		/*
3264 		 * Packet length exceeds the allowed inline
3265 		 * data length, check whether the minimal
3266 		 * inlining is required.
3267 		 */
3268 		if (txq->inlen_mode) {
3269 			assert(txq->inlen_mode >= MLX5_ESEG_MIN_INLINE_SIZE);
3270 			assert(txq->inlen_mode <= txq->inlen_send);
3271 			inlen = txq->inlen_mode;
3272 		} else {
3273 			if (!vlan || txq->vlan_en) {
3274 				/*
3275 				 * VLAN insertion will be done inside by HW.
3276 				 * It is not utmost effective - VLAN flag is
3277 				 * checked twice, but we should proceed the
3278 				 * inlining length correctly and take into
3279 				 * account the VLAN header being inserted.
3280 				 */
3281 				return mlx5_tx_packet_multi_send
3282 							(txq, loc, olx);
3283 			}
3284 			inlen = MLX5_ESEG_MIN_INLINE_SIZE;
3285 		}
3286 		/*
3287 		 * Now we know the minimal amount of data is requested
3288 		 * to inline. Check whether we should inline the buffers
3289 		 * from the chain beginning to eliminate some mbufs.
3290 		 */
3291 		mbuf = loc->mbuf;
3292 		nxlen = rte_pktmbuf_data_len(mbuf);
3293 		if (unlikely(nxlen <= txq->inlen_send)) {
3294 			/* We can inline first mbuf at least. */
3295 			if (nxlen < inlen) {
3296 				unsigned int smlen;
3297 
3298 				/* Scan mbufs till inlen filled. */
3299 				do {
3300 					smlen = nxlen;
3301 					mbuf = NEXT(mbuf);
3302 					assert(mbuf);
3303 					nxlen = rte_pktmbuf_data_len(mbuf);
3304 					nxlen += smlen;
3305 				} while (unlikely(nxlen < inlen));
3306 				if (unlikely(nxlen > txq->inlen_send)) {
3307 					/* We cannot inline entire mbuf. */
3308 					smlen = inlen - smlen;
3309 					start = rte_pktmbuf_mtod_offset
3310 						    (mbuf, uintptr_t, smlen);
3311 					goto do_align;
3312 				}
3313 			}
3314 			do {
3315 				inlen = nxlen;
3316 				mbuf = NEXT(mbuf);
3317 				/* There should be not end of packet. */
3318 				assert(mbuf);
3319 				nxlen = inlen + rte_pktmbuf_data_len(mbuf);
3320 			} while (unlikely(nxlen < txq->inlen_send));
3321 		}
3322 		start = rte_pktmbuf_mtod(mbuf, uintptr_t);
3323 		/*
3324 		 * Check whether we can do inline to align start
3325 		 * address of data buffer to cacheline.
3326 		 */
3327 do_align:
3328 		start = (~start + 1) & (RTE_CACHE_LINE_SIZE - 1);
3329 		if (unlikely(start)) {
3330 			start += inlen;
3331 			if (start <= txq->inlen_send)
3332 				inlen = start;
3333 		}
3334 	}
3335 	/*
3336 	 * Check whether there are enough free WQEBBs:
3337 	 * - Control Segment
3338 	 * - Ethernet Segment
3339 	 * - First Segment of inlined Ethernet data
3340 	 * - ... data continued ...
3341 	 * - Data Segments of pointer/min inline type
3342 	 *
3343 	 * Estimate the number of Data Segments conservatively,
3344 	 * supposing no any mbufs is being freed during inlining.
3345 	 */
3346 	assert(inlen <= txq->inlen_send);
3347 	ds = NB_SEGS(loc->mbuf) + 2 + (inlen -
3348 				       MLX5_ESEG_MIN_INLINE_SIZE +
3349 				       MLX5_WSEG_SIZE +
3350 				       MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
3351 	if (unlikely(loc->wqe_free < ((ds + 3) / 4)))
3352 		return MLX5_TXCMP_CODE_EXIT;
3353 	/* Check for maximal WQE size. */
3354 	if (unlikely((MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE) < ((ds + 3) / 4)))
3355 		return MLX5_TXCMP_CODE_ERROR;
3356 #ifdef MLX5_PMD_SOFT_COUNTERS
3357 	/* Update sent data bytes/packets counters. */
3358 	txq->stats.obytes += dlen + vlan;
3359 #endif
3360 	wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3361 	loc->wqe_last = wqe;
3362 	mlx5_tx_cseg_init(txq, loc, wqe, 0, MLX5_OPCODE_SEND, olx);
3363 	ds = mlx5_tx_mseg_build(txq, loc, wqe, vlan, inlen, 0, olx);
3364 	wqe->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
3365 	txq->wqe_ci += (ds + 3) / 4;
3366 	loc->wqe_free -= (ds + 3) / 4;
3367 	/* Request CQE generation if limits are reached. */
3368 	mlx5_tx_request_completion(txq, loc, true, olx);
3369 	return MLX5_TXCMP_CODE_MULTI;
3370 }
3371 
3372 /**
3373  * Tx burst function for multi-segment packets. Supports all
3374  * types of Tx offloads, uses MLX5_OPCODE_SEND/TSO to build WQEs,
3375  * sends one packet per WQE. Function stops sending if it
3376  * encounters the single-segment packet.
3377  *
3378  * This routine is responsible for storing processed mbuf
3379  * into elts ring buffer and update elts_head.
3380  *
3381  * @param txq
3382  *   Pointer to TX queue structure.
3383  * @param[in] pkts
3384  *   Packets to transmit.
3385  * @param pkts_n
3386  *   Number of packets in array.
3387  * @param loc
3388  *   Pointer to burst routine local context.
3389  * @param olx
3390  *   Configured Tx offloads mask. It is fully defined at
3391  *   compile time and may be used for optimization.
3392  *
3393  * @return
3394  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3395  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3396  *   MLX5_TXCMP_CODE_SINGLE - single-segment packet encountered.
3397  *   MLX5_TXCMP_CODE_TSO - TSO single-segment packet encountered.
3398  * Local context variables updated.
3399  */
3400 static __rte_always_inline enum mlx5_txcmp_code
3401 mlx5_tx_burst_mseg(struct mlx5_txq_data *restrict txq,
3402 		   struct rte_mbuf **restrict pkts,
3403 		   unsigned int pkts_n,
3404 		   struct mlx5_txq_local *restrict loc,
3405 		   unsigned int olx)
3406 {
3407 	assert(loc->elts_free && loc->wqe_free);
3408 	assert(pkts_n > loc->pkts_sent);
3409 	pkts += loc->pkts_sent + 1;
3410 	pkts_n -= loc->pkts_sent;
3411 	for (;;) {
3412 		enum mlx5_txcmp_code ret;
3413 
3414 		assert(NB_SEGS(loc->mbuf) > 1);
3415 		/*
3416 		 * Estimate the number of free elts quickly but
3417 		 * conservatively. Some segment may be fully inlined
3418 		 * and freed, ignore this here - precise estimation
3419 		 * is costly.
3420 		 */
3421 		if (loc->elts_free < NB_SEGS(loc->mbuf))
3422 			return MLX5_TXCMP_CODE_EXIT;
3423 		if (MLX5_TXOFF_CONFIG(TSO) &&
3424 		    unlikely(loc->mbuf->ol_flags & PKT_TX_TCP_SEG)) {
3425 			/* Proceed with multi-segment TSO. */
3426 			ret = mlx5_tx_packet_multi_tso(txq, loc, olx);
3427 		} else if (MLX5_TXOFF_CONFIG(INLINE)) {
3428 			/* Proceed with multi-segment SEND with inlining. */
3429 			ret = mlx5_tx_packet_multi_inline(txq, loc, olx);
3430 		} else {
3431 			/* Proceed with multi-segment SEND w/o inlining. */
3432 			ret = mlx5_tx_packet_multi_send(txq, loc, olx);
3433 		}
3434 		if (ret == MLX5_TXCMP_CODE_EXIT)
3435 			return MLX5_TXCMP_CODE_EXIT;
3436 		if (ret == MLX5_TXCMP_CODE_ERROR)
3437 			return MLX5_TXCMP_CODE_ERROR;
3438 		/* WQE is built, go to the next packet. */
3439 		++loc->pkts_sent;
3440 		--pkts_n;
3441 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
3442 			return MLX5_TXCMP_CODE_EXIT;
3443 		loc->mbuf = *pkts++;
3444 		if (pkts_n > 1)
3445 			rte_prefetch0(*pkts);
3446 		if (likely(NB_SEGS(loc->mbuf) > 1))
3447 			continue;
3448 		/* Here ends the series of multi-segment packets. */
3449 		if (MLX5_TXOFF_CONFIG(TSO) &&
3450 		    unlikely(loc->mbuf->ol_flags & PKT_TX_TCP_SEG))
3451 			return MLX5_TXCMP_CODE_TSO;
3452 		return MLX5_TXCMP_CODE_SINGLE;
3453 	}
3454 	assert(false);
3455 }
3456 
3457 /**
3458  * Tx burst function for single-segment packets with TSO.
3459  * Supports all types of Tx offloads, except multi-packets.
3460  * Uses MLX5_OPCODE_TSO to build WQEs, sends one packet per WQE.
3461  * Function stops sending if it encounters the multi-segment
3462  * packet or packet without TSO requested.
3463  *
3464  * The routine is responsible for storing processed mbuf
3465  * into elts ring buffer and update elts_head if inline
3466  * offloads is requested due to possible early freeing
3467  * of the inlined mbufs (can not store pkts array in elts
3468  * as a batch).
3469  *
3470  * @param txq
3471  *   Pointer to TX queue structure.
3472  * @param[in] pkts
3473  *   Packets to transmit.
3474  * @param pkts_n
3475  *   Number of packets in array.
3476  * @param loc
3477  *   Pointer to burst routine local context.
3478  * @param olx
3479  *   Configured Tx offloads mask. It is fully defined at
3480  *   compile time and may be used for optimization.
3481  *
3482  * @return
3483  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3484  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3485  *   MLX5_TXCMP_CODE_SINGLE - single-segment packet encountered.
3486  *   MLX5_TXCMP_CODE_MULTI - multi-segment packet encountered.
3487  * Local context variables updated.
3488  */
3489 static __rte_always_inline enum mlx5_txcmp_code
3490 mlx5_tx_burst_tso(struct mlx5_txq_data *restrict txq,
3491 		  struct rte_mbuf **restrict pkts,
3492 		  unsigned int pkts_n,
3493 		  struct mlx5_txq_local *restrict loc,
3494 		  unsigned int olx)
3495 {
3496 	assert(loc->elts_free && loc->wqe_free);
3497 	assert(pkts_n > loc->pkts_sent);
3498 	pkts += loc->pkts_sent + 1;
3499 	pkts_n -= loc->pkts_sent;
3500 	for (;;) {
3501 		struct mlx5_wqe_dseg *restrict dseg;
3502 		struct mlx5_wqe *restrict wqe;
3503 		unsigned int ds, dlen, hlen, ntcp, vlan = 0;
3504 		uint8_t *dptr;
3505 
3506 		assert(NB_SEGS(loc->mbuf) == 1);
3507 		dlen = rte_pktmbuf_data_len(loc->mbuf);
3508 		if (MLX5_TXOFF_CONFIG(VLAN) &&
3509 		    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
3510 			vlan = sizeof(struct rte_vlan_hdr);
3511 		}
3512 		/*
3513 		 * First calculate the WQE size to check
3514 		 * whether we have enough space in ring buffer.
3515 		 */
3516 		hlen = loc->mbuf->l2_len + vlan +
3517 		       loc->mbuf->l3_len + loc->mbuf->l4_len;
3518 		if (unlikely((!hlen || !loc->mbuf->tso_segsz)))
3519 			return MLX5_TXCMP_CODE_ERROR;
3520 		if (loc->mbuf->ol_flags & PKT_TX_TUNNEL_MASK)
3521 			hlen += loc->mbuf->outer_l2_len +
3522 				loc->mbuf->outer_l3_len;
3523 		/* Segment must contain all TSO headers. */
3524 		if (unlikely(hlen > MLX5_MAX_TSO_HEADER ||
3525 			     hlen <= MLX5_ESEG_MIN_INLINE_SIZE ||
3526 			     hlen > (dlen + vlan)))
3527 			return MLX5_TXCMP_CODE_ERROR;
3528 		/*
3529 		 * Check whether there are enough free WQEBBs:
3530 		 * - Control Segment
3531 		 * - Ethernet Segment
3532 		 * - First Segment of inlined Ethernet data
3533 		 * - ... data continued ...
3534 		 * - Finishing Data Segment of pointer type
3535 		 */
3536 		ds = 4 + (hlen - MLX5_ESEG_MIN_INLINE_SIZE +
3537 			  MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
3538 		if (loc->wqe_free < ((ds + 3) / 4))
3539 			return MLX5_TXCMP_CODE_EXIT;
3540 #ifdef MLX5_PMD_SOFT_COUNTERS
3541 		/* Update sent data bytes/packets counters. */
3542 		ntcp = (dlen + vlan - hlen +
3543 			loc->mbuf->tso_segsz - 1) /
3544 			loc->mbuf->tso_segsz;
3545 		/*
3546 		 * One will be added for mbuf itself at the end
3547 		 * of the mlx5_tx_burst from loc->pkts_sent field.
3548 		 */
3549 		--ntcp;
3550 		txq->stats.opackets += ntcp;
3551 		txq->stats.obytes += dlen + vlan + ntcp * hlen;
3552 #endif
3553 		/*
3554 		 * Build the TSO WQE:
3555 		 * - Control Segment
3556 		 * - Ethernet Segment with hlen bytes inlined
3557 		 * - Data Segment of pointer type
3558 		 */
3559 		wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3560 		loc->wqe_last = wqe;
3561 		mlx5_tx_cseg_init(txq, loc, wqe, ds,
3562 				  MLX5_OPCODE_TSO, olx);
3563 		dseg = mlx5_tx_eseg_data(txq, loc, wqe, vlan, hlen, 1, olx);
3564 		dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *) + hlen - vlan;
3565 		dlen -= hlen - vlan;
3566 		mlx5_tx_dseg_ptr(txq, loc, dseg, dptr, dlen, olx);
3567 		/*
3568 		 * WQE is built, update the loop parameters
3569 		 * and go to the next packet.
3570 		 */
3571 		txq->wqe_ci += (ds + 3) / 4;
3572 		loc->wqe_free -= (ds + 3) / 4;
3573 		if (MLX5_TXOFF_CONFIG(INLINE))
3574 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
3575 		--loc->elts_free;
3576 		++loc->pkts_sent;
3577 		--pkts_n;
3578 		/* Request CQE generation if limits are reached. */
3579 		mlx5_tx_request_completion(txq, loc, false, olx);
3580 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
3581 			return MLX5_TXCMP_CODE_EXIT;
3582 		loc->mbuf = *pkts++;
3583 		if (pkts_n > 1)
3584 			rte_prefetch0(*pkts);
3585 		if (MLX5_TXOFF_CONFIG(MULTI) &&
3586 		    unlikely(NB_SEGS(loc->mbuf) > 1))
3587 			return MLX5_TXCMP_CODE_MULTI;
3588 		if (likely(!(loc->mbuf->ol_flags & PKT_TX_TCP_SEG)))
3589 			return MLX5_TXCMP_CODE_SINGLE;
3590 		/* Continue with the next TSO packet. */
3591 	}
3592 	assert(false);
3593 }
3594 
3595 /**
3596  * Analyze the packet and select the best method to send.
3597  *
3598  * @param txq
3599  *   Pointer to TX queue structure.
3600  * @param loc
3601  *   Pointer to burst routine local context.
3602  * @param olx
3603  *   Configured Tx offloads mask. It is fully defined at
3604  *   compile time and may be used for optimization.
3605  * @param newp
3606  *   The predefined flag whether do complete check for
3607  *   multi-segment packets and TSO.
3608  *
3609  * @return
3610  *  MLX5_TXCMP_CODE_MULTI - multi-segment packet encountered.
3611  *  MLX5_TXCMP_CODE_TSO - TSO required, use TSO/LSO.
3612  *  MLX5_TXCMP_CODE_SINGLE - single-segment packet, use SEND.
3613  *  MLX5_TXCMP_CODE_EMPW - single-segment packet, use MPW.
3614  */
3615 static __rte_always_inline enum mlx5_txcmp_code
3616 mlx5_tx_able_to_empw(struct mlx5_txq_data *restrict txq,
3617 		     struct mlx5_txq_local *restrict loc,
3618 		     unsigned int olx,
3619 		     bool newp)
3620 {
3621 	/* Check for multi-segment packet. */
3622 	if (newp &&
3623 	    MLX5_TXOFF_CONFIG(MULTI) &&
3624 	    unlikely(NB_SEGS(loc->mbuf) > 1))
3625 		return MLX5_TXCMP_CODE_MULTI;
3626 	/* Check for TSO packet. */
3627 	if (newp &&
3628 	    MLX5_TXOFF_CONFIG(TSO) &&
3629 	    unlikely(loc->mbuf->ol_flags & PKT_TX_TCP_SEG))
3630 		return MLX5_TXCMP_CODE_TSO;
3631 	/* Check if eMPW is enabled at all. */
3632 	if (!MLX5_TXOFF_CONFIG(EMPW))
3633 		return MLX5_TXCMP_CODE_SINGLE;
3634 	/* Check if eMPW can be engaged. */
3635 	if (MLX5_TXOFF_CONFIG(VLAN) &&
3636 	    unlikely(loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) &&
3637 		(!MLX5_TXOFF_CONFIG(INLINE) ||
3638 		 unlikely((rte_pktmbuf_data_len(loc->mbuf) +
3639 			   sizeof(struct rte_vlan_hdr)) > txq->inlen_empw))) {
3640 		/*
3641 		 * eMPW does not support VLAN insertion offload,
3642 		 * we have to inline the entire packet but
3643 		 * packet is too long for inlining.
3644 		 */
3645 		return MLX5_TXCMP_CODE_SINGLE;
3646 	}
3647 	return MLX5_TXCMP_CODE_EMPW;
3648 }
3649 
3650 /**
3651  * Check the next packet attributes to match with the eMPW batch ones.
3652  *
3653  * @param txq
3654  *   Pointer to TX queue structure.
3655  * @param es
3656  *   Pointer to Ethernet Segment of eMPW batch.
3657  * @param loc
3658  *   Pointer to burst routine local context.
3659  * @param olx
3660  *   Configured Tx offloads mask. It is fully defined at
3661  *   compile time and may be used for optimization.
3662  *
3663  * @return
3664  *  true - packet match with eMPW batch attributes.
3665  *  false - no match, eMPW should be restarted.
3666  */
3667 static __rte_always_inline bool
3668 mlx5_tx_match_empw(struct mlx5_txq_data *restrict txq __rte_unused,
3669 		   struct mlx5_wqe_eseg *restrict es,
3670 		   struct mlx5_txq_local *restrict loc,
3671 		   unsigned int olx)
3672 {
3673 	uint8_t swp_flags = 0;
3674 
3675 	/* Compare the checksum flags, if any. */
3676 	if (MLX5_TXOFF_CONFIG(CSUM) &&
3677 	    txq_ol_cksum_to_cs(loc->mbuf) != es->cs_flags)
3678 		return false;
3679 	/* Compare the Software Parser offsets and flags. */
3680 	if (MLX5_TXOFF_CONFIG(SWP) &&
3681 	    (es->swp_offs != txq_mbuf_to_swp(loc, &swp_flags, olx) ||
3682 	     es->swp_flags != swp_flags))
3683 		return false;
3684 	/* Fill metadata field if needed. */
3685 	if (MLX5_TXOFF_CONFIG(METADATA) &&
3686 		es->metadata != (loc->mbuf->ol_flags & PKT_TX_METADATA ?
3687 				 loc->mbuf->tx_metadata : 0))
3688 		return false;
3689 	/* There must be no VLAN packets in eMPW loop. */
3690 	if (MLX5_TXOFF_CONFIG(VLAN))
3691 		assert(!(loc->mbuf->ol_flags & PKT_TX_VLAN_PKT));
3692 	return true;
3693 }
3694 
3695 /*
3696  * Update send loop variables and WQE for eMPW loop
3697  * without data inlining. Number of Data Segments is
3698  * equal to the number of sent packets.
3699  *
3700  * @param txq
3701  *   Pointer to TX queue structure.
3702  * @param loc
3703  *   Pointer to burst routine local context.
3704  * @param ds
3705  *   Number of packets/Data Segments/Packets.
3706  * @param slen
3707  *   Accumulated statistics, bytes sent
3708  * @param olx
3709  *   Configured Tx offloads mask. It is fully defined at
3710  *   compile time and may be used for optimization.
3711  *
3712  * @return
3713  *  true - packet match with eMPW batch attributes.
3714  *  false - no match, eMPW should be restarted.
3715  */
3716 static __rte_always_inline void
3717 mlx5_tx_sdone_empw(struct mlx5_txq_data *restrict txq,
3718 		   struct mlx5_txq_local *restrict loc,
3719 		   unsigned int ds,
3720 		   unsigned int slen,
3721 		   unsigned int olx)
3722 {
3723 	assert(!MLX5_TXOFF_CONFIG(INLINE));
3724 #ifdef MLX5_PMD_SOFT_COUNTERS
3725 	/* Update sent data bytes counter. */
3726 	 txq->stats.obytes += slen;
3727 #else
3728 	(void)slen;
3729 #endif
3730 	loc->elts_free -= ds;
3731 	loc->pkts_sent += ds;
3732 	ds += 2;
3733 	loc->wqe_last->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | ds);
3734 	txq->wqe_ci += (ds + 3) / 4;
3735 	loc->wqe_free -= (ds + 3) / 4;
3736 	/* Request CQE generation if limits are reached. */
3737 	mlx5_tx_request_completion(txq, loc, false, olx);
3738 }
3739 
3740 /*
3741  * Update send loop variables and WQE for eMPW loop
3742  * with data inlining. Gets the size of pushed descriptors
3743  * and data to the WQE.
3744  *
3745  * @param txq
3746  *   Pointer to TX queue structure.
3747  * @param loc
3748  *   Pointer to burst routine local context.
3749  * @param len
3750  *   Total size of descriptor/data in bytes.
3751  * @param slen
3752  *   Accumulated statistics, data bytes sent.
3753  * @param olx
3754  *   Configured Tx offloads mask. It is fully defined at
3755  *   compile time and may be used for optimization.
3756  *
3757  * @return
3758  *  true - packet match with eMPW batch attributes.
3759  *  false - no match, eMPW should be restarted.
3760  */
3761 static __rte_always_inline void
3762 mlx5_tx_idone_empw(struct mlx5_txq_data *restrict txq,
3763 		   struct mlx5_txq_local *restrict loc,
3764 		   unsigned int len,
3765 		   unsigned int slen,
3766 		   unsigned int olx __rte_unused)
3767 {
3768 	assert(MLX5_TXOFF_CONFIG(INLINE));
3769 	assert((len % MLX5_WSEG_SIZE) == 0);
3770 #ifdef MLX5_PMD_SOFT_COUNTERS
3771 	/* Update sent data bytes counter. */
3772 	 txq->stats.obytes += slen;
3773 #else
3774 	(void)slen;
3775 #endif
3776 	len = len / MLX5_WSEG_SIZE + 2;
3777 	loc->wqe_last->cseg.sq_ds = rte_cpu_to_be_32(txq->qp_num_8s | len);
3778 	txq->wqe_ci += (len + 3) / 4;
3779 	loc->wqe_free -= (len + 3) / 4;
3780 	/* Request CQE generation if limits are reached. */
3781 	mlx5_tx_request_completion(txq, loc, false, olx);
3782 }
3783 
3784 /**
3785  * The set of Tx burst functions for single-segment packets
3786  * without TSO and with Multi-Packet Writing feature support.
3787  * Supports all types of Tx offloads, except multi-packets
3788  * and TSO.
3789  *
3790  * Uses MLX5_OPCODE_EMPW to build WQEs if possible and sends
3791  * as many packet per WQE as it can. If eMPW is not configured
3792  * or packet can not be sent with eMPW (VLAN insertion) the
3793  * ordinary SEND opcode is used and only one packet placed
3794  * in WQE.
3795  *
3796  * Functions stop sending if it encounters the multi-segment
3797  * packet or packet with TSO requested.
3798  *
3799  * The routines are responsible for storing processed mbuf
3800  * into elts ring buffer and update elts_head if inlining
3801  * offload is requested. Otherwise the copying mbufs to elts
3802  * can be postponed and completed at the end of burst routine.
3803  *
3804  * @param txq
3805  *   Pointer to TX queue structure.
3806  * @param[in] pkts
3807  *   Packets to transmit.
3808  * @param pkts_n
3809  *   Number of packets in array.
3810  * @param loc
3811  *   Pointer to burst routine local context.
3812  * @param olx
3813  *   Configured Tx offloads mask. It is fully defined at
3814  *   compile time and may be used for optimization.
3815  *
3816  * @return
3817  *   MLX5_TXCMP_CODE_EXIT - sending is done or impossible.
3818  *   MLX5_TXCMP_CODE_ERROR - some unrecoverable error occurred.
3819  *   MLX5_TXCMP_CODE_MULTI - multi-segment packet encountered.
3820  *   MLX5_TXCMP_CODE_TSO - TSO packet encountered.
3821  *   MLX5_TXCMP_CODE_SINGLE - used inside functions set.
3822  *   MLX5_TXCMP_CODE_EMPW - used inside functions set.
3823  *
3824  * Local context variables updated.
3825  *
3826  *
3827  * The routine sends packets with MLX5_OPCODE_EMPW
3828  * without inlining, this is dedicated optimized branch.
3829  * No VLAN insertion is supported.
3830  */
3831 static __rte_always_inline enum mlx5_txcmp_code
3832 mlx5_tx_burst_empw_simple(struct mlx5_txq_data *restrict txq,
3833 			  struct rte_mbuf **restrict pkts,
3834 			  unsigned int pkts_n,
3835 			  struct mlx5_txq_local *restrict loc,
3836 			  unsigned int olx)
3837 {
3838 	/*
3839 	 * Subroutine is the part of mlx5_tx_burst_single()
3840 	 * and sends single-segment packet with eMPW opcode
3841 	 * without data inlining.
3842 	 */
3843 	assert(!MLX5_TXOFF_CONFIG(INLINE));
3844 	assert(MLX5_TXOFF_CONFIG(EMPW));
3845 	assert(loc->elts_free && loc->wqe_free);
3846 	assert(pkts_n > loc->pkts_sent);
3847 	static_assert(MLX5_EMPW_MIN_PACKETS >= 2, "invalid min size");
3848 	pkts += loc->pkts_sent + 1;
3849 	pkts_n -= loc->pkts_sent;
3850 	for (;;) {
3851 		struct mlx5_wqe_dseg *restrict dseg;
3852 		struct mlx5_wqe_eseg *restrict eseg;
3853 		enum mlx5_txcmp_code ret;
3854 		unsigned int part, loop;
3855 		unsigned int slen = 0;
3856 
3857 next_empw:
3858 		part = RTE_MIN(pkts_n, MLX5_EMPW_MAX_PACKETS);
3859 		if (unlikely(loc->elts_free < part)) {
3860 			/* We have no enough elts to save all mbufs. */
3861 			if (unlikely(loc->elts_free < MLX5_EMPW_MIN_PACKETS))
3862 				return MLX5_TXCMP_CODE_EXIT;
3863 			/* But we still able to send at least minimal eMPW. */
3864 			part = loc->elts_free;
3865 		}
3866 		/* Check whether we have enough WQEs */
3867 		if (unlikely(loc->wqe_free < ((2 + part + 3) / 4))) {
3868 			if (unlikely(loc->wqe_free <
3869 				((2 + MLX5_EMPW_MIN_PACKETS + 3) / 4)))
3870 				return MLX5_TXCMP_CODE_EXIT;
3871 			part = (loc->wqe_free * 4) - 2;
3872 		}
3873 		if (likely(part > 1))
3874 			rte_prefetch0(*pkts);
3875 		loc->wqe_last = txq->wqes + (txq->wqe_ci & txq->wqe_m);
3876 		/*
3877 		 * Build eMPW title WQEBB:
3878 		 * - Control Segment, eMPW opcode
3879 		 * - Ethernet Segment, no inline
3880 		 */
3881 		mlx5_tx_cseg_init(txq, loc, loc->wqe_last, part + 2,
3882 				  MLX5_OPCODE_ENHANCED_MPSW, olx);
3883 		mlx5_tx_eseg_none(txq, loc, loc->wqe_last,
3884 				  olx & ~MLX5_TXOFF_CONFIG_VLAN);
3885 		eseg = &loc->wqe_last->eseg;
3886 		dseg = &loc->wqe_last->dseg[0];
3887 		loop = part;
3888 		for (;;) {
3889 			uint32_t dlen = rte_pktmbuf_data_len(loc->mbuf);
3890 #ifdef MLX5_PMD_SOFT_COUNTERS
3891 			/* Update sent data bytes counter. */
3892 			slen += dlen;
3893 #endif
3894 			mlx5_tx_dseg_ptr
3895 				(txq, loc, dseg,
3896 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
3897 				 dlen, olx);
3898 			if (unlikely(--loop == 0))
3899 				break;
3900 			loc->mbuf = *pkts++;
3901 			if (likely(loop > 1))
3902 				rte_prefetch0(*pkts);
3903 			ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
3904 			/*
3905 			 * Unroll the completion code to avoid
3906 			 * returning variable value - it results in
3907 			 * unoptimized sequent checking in caller.
3908 			 */
3909 			if (ret == MLX5_TXCMP_CODE_MULTI) {
3910 				part -= loop;
3911 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
3912 				if (unlikely(!loc->elts_free ||
3913 					     !loc->wqe_free))
3914 					return MLX5_TXCMP_CODE_EXIT;
3915 				return MLX5_TXCMP_CODE_MULTI;
3916 			}
3917 			if (ret == MLX5_TXCMP_CODE_TSO) {
3918 				part -= loop;
3919 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
3920 				if (unlikely(!loc->elts_free ||
3921 					     !loc->wqe_free))
3922 					return MLX5_TXCMP_CODE_EXIT;
3923 				return MLX5_TXCMP_CODE_TSO;
3924 			}
3925 			if (ret == MLX5_TXCMP_CODE_SINGLE) {
3926 				part -= loop;
3927 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
3928 				if (unlikely(!loc->elts_free ||
3929 					     !loc->wqe_free))
3930 					return MLX5_TXCMP_CODE_EXIT;
3931 				return MLX5_TXCMP_CODE_SINGLE;
3932 			}
3933 			if (ret != MLX5_TXCMP_CODE_EMPW) {
3934 				assert(false);
3935 				part -= loop;
3936 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
3937 				return MLX5_TXCMP_CODE_ERROR;
3938 			}
3939 			/*
3940 			 * Check whether packet parameters coincide
3941 			 * within assumed eMPW batch:
3942 			 * - check sum settings
3943 			 * - metadata value
3944 			 * - software parser settings
3945 			 */
3946 			if (!mlx5_tx_match_empw(txq, eseg, loc, olx)) {
3947 				assert(loop);
3948 				part -= loop;
3949 				mlx5_tx_sdone_empw(txq, loc, part, slen, olx);
3950 				if (unlikely(!loc->elts_free ||
3951 					     !loc->wqe_free))
3952 					return MLX5_TXCMP_CODE_EXIT;
3953 				pkts_n -= part;
3954 				goto next_empw;
3955 			}
3956 			/* Packet attributes match, continue the same eMPW. */
3957 			++dseg;
3958 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
3959 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
3960 		}
3961 		/* eMPW is built successfully, update loop parameters. */
3962 		assert(!loop);
3963 		assert(pkts_n >= part);
3964 #ifdef MLX5_PMD_SOFT_COUNTERS
3965 		/* Update sent data bytes counter. */
3966 		txq->stats.obytes += slen;
3967 #endif
3968 		loc->elts_free -= part;
3969 		loc->pkts_sent += part;
3970 		txq->wqe_ci += (2 + part + 3) / 4;
3971 		loc->wqe_free -= (2 + part + 3) / 4;
3972 		pkts_n -= part;
3973 		/* Request CQE generation if limits are reached. */
3974 		mlx5_tx_request_completion(txq, loc, false, olx);
3975 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
3976 			return MLX5_TXCMP_CODE_EXIT;
3977 		loc->mbuf = *pkts++;
3978 		ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
3979 		if (unlikely(ret != MLX5_TXCMP_CODE_EMPW))
3980 			return ret;
3981 		/* Continue sending eMPW batches. */
3982 	}
3983 	assert(false);
3984 }
3985 
3986 /**
3987  * The routine sends packets with MLX5_OPCODE_EMPW
3988  * with inlining, optionally supports VLAN insertion.
3989  */
3990 static __rte_always_inline enum mlx5_txcmp_code
3991 mlx5_tx_burst_empw_inline(struct mlx5_txq_data *restrict txq,
3992 			  struct rte_mbuf **restrict pkts,
3993 			  unsigned int pkts_n,
3994 			  struct mlx5_txq_local *restrict loc,
3995 			  unsigned int olx)
3996 {
3997 	/*
3998 	 * Subroutine is the part of mlx5_tx_burst_single()
3999 	 * and sends single-segment packet with eMPW opcode
4000 	 * with data inlining.
4001 	 */
4002 	assert(MLX5_TXOFF_CONFIG(INLINE));
4003 	assert(MLX5_TXOFF_CONFIG(EMPW));
4004 	assert(loc->elts_free && loc->wqe_free);
4005 	assert(pkts_n > loc->pkts_sent);
4006 	static_assert(MLX5_EMPW_MIN_PACKETS >= 2, "invalid min size");
4007 	pkts += loc->pkts_sent + 1;
4008 	pkts_n -= loc->pkts_sent;
4009 	for (;;) {
4010 		struct mlx5_wqe_dseg *restrict dseg;
4011 		struct mlx5_wqe_eseg *restrict eseg;
4012 		enum mlx5_txcmp_code ret;
4013 		unsigned int room, part, nlim;
4014 		unsigned int slen = 0;
4015 
4016 		/*
4017 		 * Limits the amount of packets in one WQE
4018 		 * to improve CQE latency generation.
4019 		 */
4020 		nlim = RTE_MIN(pkts_n, MLX5_EMPW_MAX_PACKETS);
4021 		/* Check whether we have minimal amount WQEs */
4022 		if (unlikely(loc->wqe_free <
4023 			    ((2 + MLX5_EMPW_MIN_PACKETS + 3) / 4)))
4024 			return MLX5_TXCMP_CODE_EXIT;
4025 		if (likely(pkts_n > 1))
4026 			rte_prefetch0(*pkts);
4027 		loc->wqe_last = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4028 		/*
4029 		 * Build eMPW title WQEBB:
4030 		 * - Control Segment, eMPW opcode, zero DS
4031 		 * - Ethernet Segment, no inline
4032 		 */
4033 		mlx5_tx_cseg_init(txq, loc, loc->wqe_last, 0,
4034 				  MLX5_OPCODE_ENHANCED_MPSW, olx);
4035 		mlx5_tx_eseg_none(txq, loc, loc->wqe_last,
4036 				  olx & ~MLX5_TXOFF_CONFIG_VLAN);
4037 		eseg = &loc->wqe_last->eseg;
4038 		dseg = &loc->wqe_last->dseg[0];
4039 		room = RTE_MIN(MLX5_WQE_SIZE_MAX / MLX5_WQE_SIZE,
4040 			       loc->wqe_free) * MLX5_WQE_SIZE -
4041 					MLX5_WQE_CSEG_SIZE -
4042 					MLX5_WQE_ESEG_SIZE;
4043 		/* Build WQE till we have space, packets and resources. */
4044 		part = room;
4045 		for (;;) {
4046 			uint32_t dlen = rte_pktmbuf_data_len(loc->mbuf);
4047 			uint8_t *dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *);
4048 			unsigned int tlen;
4049 
4050 			assert(room >= MLX5_WQE_DSEG_SIZE);
4051 			assert((room % MLX5_WQE_DSEG_SIZE) == 0);
4052 			assert((uintptr_t)dseg < (uintptr_t)txq->wqes_end);
4053 			/*
4054 			 * Some Tx offloads may cause an error if
4055 			 * packet is not long enough, check against
4056 			 * assumed minimal length.
4057 			 */
4058 			if (unlikely(dlen <= MLX5_ESEG_MIN_INLINE_SIZE)) {
4059 				part -= room;
4060 				if (unlikely(!part))
4061 					return MLX5_TXCMP_CODE_ERROR;
4062 				/*
4063 				 * We have some successfully built
4064 				 * packet Data Segments to send.
4065 				 */
4066 				mlx5_tx_idone_empw(txq, loc, part, slen, olx);
4067 				return MLX5_TXCMP_CODE_ERROR;
4068 			}
4069 			/* Inline or not inline - that's the Question. */
4070 			if (dlen > txq->inlen_empw)
4071 				goto pointer_empw;
4072 			/* Inline entire packet, optional VLAN insertion. */
4073 			tlen = sizeof(dseg->bcount) + dlen;
4074 			if (MLX5_TXOFF_CONFIG(VLAN) &&
4075 			    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
4076 				/*
4077 				 * The packet length must be checked in
4078 				 * mlx5_tx_able_to_empw() and packet
4079 				 * fits into inline length guaranteed.
4080 				 */
4081 				assert((dlen + sizeof(struct rte_vlan_hdr)) <=
4082 					txq->inlen_empw);
4083 				tlen += sizeof(struct rte_vlan_hdr);
4084 				if (room < tlen)
4085 					break;
4086 				dseg = mlx5_tx_dseg_vlan(txq, loc, dseg,
4087 							 dptr, dlen, olx);
4088 #ifdef MLX5_PMD_SOFT_COUNTERS
4089 				/* Update sent data bytes counter. */
4090 				slen +=	sizeof(struct rte_vlan_hdr);
4091 #endif
4092 			} else {
4093 				if (room < tlen)
4094 					break;
4095 				dseg = mlx5_tx_dseg_empw(txq, loc, dseg,
4096 							 dptr, dlen, olx);
4097 			}
4098 			tlen = RTE_ALIGN(tlen, MLX5_WSEG_SIZE);
4099 			assert(room >= tlen);
4100 			room -= tlen;
4101 			/*
4102 			 * Packet data are completely inlined,
4103 			 * free the packet immediately.
4104 			 */
4105 			rte_pktmbuf_free_seg(loc->mbuf);
4106 			goto next_mbuf;
4107 pointer_empw:
4108 			/*
4109 			 * Not inlinable VLAN packets are
4110 			 * proceeded outside of this routine.
4111 			 */
4112 			assert(room >= MLX5_WQE_DSEG_SIZE);
4113 			if (MLX5_TXOFF_CONFIG(VLAN))
4114 				assert(!(loc->mbuf->ol_flags &
4115 					 PKT_TX_VLAN_PKT));
4116 			mlx5_tx_dseg_ptr(txq, loc, dseg, dptr, dlen, olx);
4117 			/* We have to store mbuf in elts.*/
4118 			txq->elts[txq->elts_head++ & txq->elts_m] = loc->mbuf;
4119 			room -= MLX5_WQE_DSEG_SIZE;
4120 			/* Ring buffer wraparound is checked at the loop end.*/
4121 			++dseg;
4122 next_mbuf:
4123 #ifdef MLX5_PMD_SOFT_COUNTERS
4124 			/* Update sent data bytes counter. */
4125 			slen += dlen;
4126 #endif
4127 			loc->pkts_sent++;
4128 			loc->elts_free--;
4129 			pkts_n--;
4130 			if (unlikely(!pkts_n || !loc->elts_free)) {
4131 				/*
4132 				 * We have no resources/packets to
4133 				 * continue build descriptors.
4134 				 */
4135 				part -= room;
4136 				mlx5_tx_idone_empw(txq, loc, part, slen, olx);
4137 				return MLX5_TXCMP_CODE_EXIT;
4138 			}
4139 			loc->mbuf = *pkts++;
4140 			if (likely(pkts_n > 1))
4141 				rte_prefetch0(*pkts);
4142 			ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
4143 			/*
4144 			 * Unroll the completion code to avoid
4145 			 * returning variable value - it results in
4146 			 * unoptimized sequent checking in caller.
4147 			 */
4148 			if (ret == MLX5_TXCMP_CODE_MULTI) {
4149 				part -= room;
4150 				mlx5_tx_idone_empw(txq, loc, part, slen, olx);
4151 				if (unlikely(!loc->elts_free ||
4152 					     !loc->wqe_free))
4153 					return MLX5_TXCMP_CODE_EXIT;
4154 				return MLX5_TXCMP_CODE_MULTI;
4155 			}
4156 			if (ret == MLX5_TXCMP_CODE_TSO) {
4157 				part -= room;
4158 				mlx5_tx_idone_empw(txq, loc, part, slen, olx);
4159 				if (unlikely(!loc->elts_free ||
4160 					     !loc->wqe_free))
4161 					return MLX5_TXCMP_CODE_EXIT;
4162 				return MLX5_TXCMP_CODE_TSO;
4163 			}
4164 			if (ret == MLX5_TXCMP_CODE_SINGLE) {
4165 				part -= room;
4166 				mlx5_tx_idone_empw(txq, loc, part, slen, olx);
4167 				if (unlikely(!loc->elts_free ||
4168 					     !loc->wqe_free))
4169 					return MLX5_TXCMP_CODE_EXIT;
4170 				return MLX5_TXCMP_CODE_SINGLE;
4171 			}
4172 			if (ret != MLX5_TXCMP_CODE_EMPW) {
4173 				assert(false);
4174 				part -= room;
4175 				mlx5_tx_idone_empw(txq, loc, part, slen, olx);
4176 				return MLX5_TXCMP_CODE_ERROR;
4177 			}
4178 			/* Check if we have minimal room left. */
4179 			nlim--;
4180 			if (unlikely(!nlim || room < MLX5_WQE_DSEG_SIZE))
4181 				break;
4182 			/*
4183 			 * Check whether packet parameters coincide
4184 			 * within assumed eMPW batch:
4185 			 * - check sum settings
4186 			 * - metadata value
4187 			 * - software parser settings
4188 			 */
4189 			if (!mlx5_tx_match_empw(txq, eseg, loc, olx))
4190 				break;
4191 			/* Packet attributes match, continue the same eMPW. */
4192 			if ((uintptr_t)dseg >= (uintptr_t)txq->wqes_end)
4193 				dseg = (struct mlx5_wqe_dseg *)txq->wqes;
4194 		}
4195 		/*
4196 		 * We get here to close an existing eMPW
4197 		 * session and start the new one.
4198 		 */
4199 		assert(pkts_n);
4200 		part -= room;
4201 		if (unlikely(!part))
4202 			return MLX5_TXCMP_CODE_EXIT;
4203 		mlx5_tx_idone_empw(txq, loc, part, slen, olx);
4204 		if (unlikely(!loc->elts_free ||
4205 			     !loc->wqe_free))
4206 			return MLX5_TXCMP_CODE_EXIT;
4207 		/* Continue the loop with new eMPW session. */
4208 	}
4209 	assert(false);
4210 }
4211 
4212 /**
4213  * The routine sends packets with ordinary MLX5_OPCODE_SEND.
4214  * Data inlining and VLAN insertion are supported.
4215  */
4216 static __rte_always_inline enum mlx5_txcmp_code
4217 mlx5_tx_burst_single_send(struct mlx5_txq_data *restrict txq,
4218 			  struct rte_mbuf **restrict pkts,
4219 			  unsigned int pkts_n,
4220 			  struct mlx5_txq_local *restrict loc,
4221 			  unsigned int olx)
4222 {
4223 	/*
4224 	 * Subroutine is the part of mlx5_tx_burst_single()
4225 	 * and sends single-segment packet with SEND opcode.
4226 	 */
4227 	assert(loc->elts_free && loc->wqe_free);
4228 	assert(pkts_n > loc->pkts_sent);
4229 	pkts += loc->pkts_sent + 1;
4230 	pkts_n -= loc->pkts_sent;
4231 	for (;;) {
4232 		struct mlx5_wqe *restrict wqe;
4233 		enum mlx5_txcmp_code ret;
4234 
4235 		assert(NB_SEGS(loc->mbuf) == 1);
4236 		if (MLX5_TXOFF_CONFIG(INLINE)) {
4237 			unsigned int inlen, vlan = 0;
4238 
4239 			inlen = rte_pktmbuf_data_len(loc->mbuf);
4240 			if (MLX5_TXOFF_CONFIG(VLAN) &&
4241 			    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT) {
4242 				vlan = sizeof(struct rte_vlan_hdr);
4243 				inlen += vlan;
4244 				static_assert((sizeof(struct rte_vlan_hdr) +
4245 					       sizeof(struct rte_ether_hdr)) ==
4246 					       MLX5_ESEG_MIN_INLINE_SIZE,
4247 					       "invalid min inline data size");
4248 			}
4249 			/*
4250 			 * If inlining is enabled at configuration time
4251 			 * the limit must be not less than minimal size.
4252 			 * Otherwise we would do extra check for data
4253 			 * size to avoid crashes due to length overflow.
4254 			 */
4255 			assert(txq->inlen_send >= MLX5_ESEG_MIN_INLINE_SIZE);
4256 			if (inlen <= txq->inlen_send) {
4257 				unsigned int seg_n, wqe_n;
4258 
4259 				rte_prefetch0(rte_pktmbuf_mtod
4260 						(loc->mbuf, uint8_t *));
4261 				/* Check against minimal length. */
4262 				if (inlen <= MLX5_ESEG_MIN_INLINE_SIZE)
4263 					return MLX5_TXCMP_CODE_ERROR;
4264 				/*
4265 				 * Completely inlined packet data WQE:
4266 				 * - Control Segment, SEND opcode
4267 				 * - Ethernet Segment, no VLAN insertion
4268 				 * - Data inlined, VLAN optionally inserted
4269 				 * - Alignment to MLX5_WSEG_SIZE
4270 				 * Have to estimate amount of WQEBBs
4271 				 */
4272 				seg_n = (inlen + 3 * MLX5_WSEG_SIZE -
4273 					 MLX5_ESEG_MIN_INLINE_SIZE +
4274 					 MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
4275 				/* Check if there are enough WQEBBs. */
4276 				wqe_n = (seg_n + 3) / 4;
4277 				if (wqe_n > loc->wqe_free)
4278 					return MLX5_TXCMP_CODE_EXIT;
4279 				wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4280 				loc->wqe_last = wqe;
4281 				mlx5_tx_cseg_init(txq, loc, wqe, seg_n,
4282 						  MLX5_OPCODE_SEND, olx);
4283 				mlx5_tx_eseg_data(txq, loc, wqe,
4284 						  vlan, inlen, 0, olx);
4285 				txq->wqe_ci += wqe_n;
4286 				loc->wqe_free -= wqe_n;
4287 				/*
4288 				 * Packet data are completely inlined,
4289 				 * free the packet immediately.
4290 				 */
4291 				rte_pktmbuf_free_seg(loc->mbuf);
4292 			} else if (!MLX5_TXOFF_CONFIG(EMPW) &&
4293 				   txq->inlen_mode) {
4294 				/*
4295 				 * If minimal inlining is requested the eMPW
4296 				 * feature should be disabled due to data is
4297 				 * inlined into Ethernet Segment, which can
4298 				 * not contain inlined data for eMPW due to
4299 				 * segment shared for all packets.
4300 				 */
4301 				struct mlx5_wqe_dseg *restrict dseg;
4302 				unsigned int ds;
4303 				uint8_t *dptr;
4304 
4305 				/*
4306 				 * The inline-mode settings require
4307 				 * to inline the specified amount of
4308 				 * data bytes to the Ethernet Segment.
4309 				 * We should check the free space in
4310 				 * WQE ring buffer to inline partially.
4311 				 */
4312 				assert(txq->inlen_send >= txq->inlen_mode);
4313 				assert(inlen > txq->inlen_mode);
4314 				assert(txq->inlen_mode >=
4315 						MLX5_ESEG_MIN_INLINE_SIZE);
4316 				/*
4317 				 * Check whether there are enough free WQEBBs:
4318 				 * - Control Segment
4319 				 * - Ethernet Segment
4320 				 * - First Segment of inlined Ethernet data
4321 				 * - ... data continued ...
4322 				 * - Finishing Data Segment of pointer type
4323 				 */
4324 				ds = (MLX5_WQE_CSEG_SIZE +
4325 				      MLX5_WQE_ESEG_SIZE +
4326 				      MLX5_WQE_DSEG_SIZE +
4327 				      txq->inlen_mode -
4328 				      MLX5_ESEG_MIN_INLINE_SIZE +
4329 				      MLX5_WQE_DSEG_SIZE +
4330 				      MLX5_WSEG_SIZE - 1) / MLX5_WSEG_SIZE;
4331 				if (loc->wqe_free < ((ds + 3) / 4))
4332 					return MLX5_TXCMP_CODE_EXIT;
4333 				/*
4334 				 * Build the ordinary SEND WQE:
4335 				 * - Control Segment
4336 				 * - Ethernet Segment, inline inlen_mode bytes
4337 				 * - Data Segment of pointer type
4338 				 */
4339 				wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4340 				loc->wqe_last = wqe;
4341 				mlx5_tx_cseg_init(txq, loc, wqe, ds,
4342 						  MLX5_OPCODE_SEND, olx);
4343 				dseg = mlx5_tx_eseg_data(txq, loc, wqe, vlan,
4344 							 txq->inlen_mode,
4345 							 0, olx);
4346 				dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *) +
4347 				       txq->inlen_mode - vlan;
4348 				inlen -= txq->inlen_mode;
4349 				mlx5_tx_dseg_ptr(txq, loc, dseg,
4350 						 dptr, inlen, olx);
4351 				/*
4352 				 * WQE is built, update the loop parameters
4353 				 * and got to the next packet.
4354 				 */
4355 				txq->wqe_ci += (ds + 3) / 4;
4356 				loc->wqe_free -= (ds + 3) / 4;
4357 				/* We have to store mbuf in elts.*/
4358 				assert(MLX5_TXOFF_CONFIG(INLINE));
4359 				txq->elts[txq->elts_head++ & txq->elts_m] =
4360 						loc->mbuf;
4361 				--loc->elts_free;
4362 			} else {
4363 				uint8_t *dptr;
4364 				unsigned int dlen;
4365 
4366 				/*
4367 				 * Partially inlined packet data WQE, we have
4368 				 * some space in title WQEBB, we can fill it
4369 				 * with some packet data. It takes one WQEBB,
4370 				 * it is available, no extra space check:
4371 				 * - Control Segment, SEND opcode
4372 				 * - Ethernet Segment, no VLAN insertion
4373 				 * - MLX5_ESEG_MIN_INLINE_SIZE bytes of Data
4374 				 * - Data Segment, pointer type
4375 				 *
4376 				 * We also get here if VLAN insertion is not
4377 				 * supported by HW, the inline is enabled.
4378 				 */
4379 				wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4380 				loc->wqe_last = wqe;
4381 				mlx5_tx_cseg_init(txq, loc, wqe, 4,
4382 						  MLX5_OPCODE_SEND, olx);
4383 				mlx5_tx_eseg_dmin(txq, loc, wqe, vlan, olx);
4384 				dptr = rte_pktmbuf_mtod(loc->mbuf, uint8_t *) +
4385 				       MLX5_ESEG_MIN_INLINE_SIZE - vlan;
4386 				/*
4387 				 * The length check is performed above, by
4388 				 * comparing with txq->inlen_send. We should
4389 				 * not get overflow here.
4390 				 */
4391 				assert(inlen > MLX5_ESEG_MIN_INLINE_SIZE);
4392 				dlen = inlen - MLX5_ESEG_MIN_INLINE_SIZE;
4393 				mlx5_tx_dseg_ptr(txq, loc, &wqe->dseg[1],
4394 						 dptr, dlen, olx);
4395 				++txq->wqe_ci;
4396 				--loc->wqe_free;
4397 				/* We have to store mbuf in elts.*/
4398 				assert(MLX5_TXOFF_CONFIG(INLINE));
4399 				txq->elts[txq->elts_head++ & txq->elts_m] =
4400 						loc->mbuf;
4401 				--loc->elts_free;
4402 			}
4403 #ifdef MLX5_PMD_SOFT_COUNTERS
4404 			/* Update sent data bytes counter. */
4405 			txq->stats.obytes += vlan +
4406 					rte_pktmbuf_data_len(loc->mbuf);
4407 #endif
4408 		} else {
4409 			/*
4410 			 * No inline at all, it means the CPU cycles saving
4411 			 * is prioritized at configuration, we should not
4412 			 * copy any packet data to WQE.
4413 			 *
4414 			 * SEND WQE, one WQEBB:
4415 			 * - Control Segment, SEND opcode
4416 			 * - Ethernet Segment, optional VLAN, no inline
4417 			 * - Data Segment, pointer type
4418 			 */
4419 			wqe = txq->wqes + (txq->wqe_ci & txq->wqe_m);
4420 			loc->wqe_last = wqe;
4421 			mlx5_tx_cseg_init(txq, loc, wqe, 3,
4422 					  MLX5_OPCODE_SEND, olx);
4423 			mlx5_tx_eseg_none(txq, loc, wqe, olx);
4424 			mlx5_tx_dseg_ptr
4425 				(txq, loc, &wqe->dseg[0],
4426 				 rte_pktmbuf_mtod(loc->mbuf, uint8_t *),
4427 				 rte_pktmbuf_data_len(loc->mbuf), olx);
4428 			++txq->wqe_ci;
4429 			--loc->wqe_free;
4430 			/*
4431 			 * We should not store mbuf pointer in elts
4432 			 * if no inlining is configured, this is done
4433 			 * by calling routine in a batch copy.
4434 			 */
4435 			assert(!MLX5_TXOFF_CONFIG(INLINE));
4436 			--loc->elts_free;
4437 #ifdef MLX5_PMD_SOFT_COUNTERS
4438 			/* Update sent data bytes counter. */
4439 			txq->stats.obytes += rte_pktmbuf_data_len(loc->mbuf);
4440 			if (MLX5_TXOFF_CONFIG(VLAN) &&
4441 			    loc->mbuf->ol_flags & PKT_TX_VLAN_PKT)
4442 				txq->stats.obytes +=
4443 					sizeof(struct rte_vlan_hdr);
4444 #endif
4445 		}
4446 		++loc->pkts_sent;
4447 		--pkts_n;
4448 		/* Request CQE generation if limits are reached. */
4449 		mlx5_tx_request_completion(txq, loc, false, olx);
4450 		if (unlikely(!pkts_n || !loc->elts_free || !loc->wqe_free))
4451 			return MLX5_TXCMP_CODE_EXIT;
4452 		loc->mbuf = *pkts++;
4453 		if (pkts_n > 1)
4454 			rte_prefetch0(*pkts);
4455 		ret = mlx5_tx_able_to_empw(txq, loc, olx, true);
4456 		if (unlikely(ret != MLX5_TXCMP_CODE_SINGLE))
4457 			return ret;
4458 	}
4459 	assert(false);
4460 }
4461 
4462 static __rte_always_inline enum mlx5_txcmp_code
4463 mlx5_tx_burst_single(struct mlx5_txq_data *restrict txq,
4464 		     struct rte_mbuf **restrict pkts,
4465 		     unsigned int pkts_n,
4466 		     struct mlx5_txq_local *restrict loc,
4467 		     unsigned int olx)
4468 {
4469 	enum mlx5_txcmp_code ret;
4470 
4471 	ret = mlx5_tx_able_to_empw(txq, loc, olx, false);
4472 	if (ret == MLX5_TXCMP_CODE_SINGLE)
4473 		goto ordinary_send;
4474 	assert(ret == MLX5_TXCMP_CODE_EMPW);
4475 	for (;;) {
4476 		/* Optimize for inline/no inline eMPW send. */
4477 		ret = (MLX5_TXOFF_CONFIG(INLINE)) ?
4478 			mlx5_tx_burst_empw_inline
4479 				(txq, pkts, pkts_n, loc, olx) :
4480 			mlx5_tx_burst_empw_simple
4481 				(txq, pkts, pkts_n, loc, olx);
4482 		if (ret != MLX5_TXCMP_CODE_SINGLE)
4483 			return ret;
4484 		/* The resources to send one packet should remain. */
4485 		assert(loc->elts_free && loc->wqe_free);
4486 ordinary_send:
4487 		ret = mlx5_tx_burst_single_send(txq, pkts, pkts_n, loc, olx);
4488 		assert(ret != MLX5_TXCMP_CODE_SINGLE);
4489 		if (ret != MLX5_TXCMP_CODE_EMPW)
4490 			return ret;
4491 		/* The resources to send one packet should remain. */
4492 		assert(loc->elts_free && loc->wqe_free);
4493 	}
4494 }
4495 
4496 /**
4497  * DPDK Tx callback template. This is configured template
4498  * used to generate routines optimized for specified offload setup.
4499  * One of this generated functions is chosen at SQ configuration
4500  * time.
4501  *
4502  * @param txq
4503  *   Generic pointer to TX queue structure.
4504  * @param[in] pkts
4505  *   Packets to transmit.
4506  * @param pkts_n
4507  *   Number of packets in array.
4508  * @param olx
4509  *   Configured offloads mask, presents the bits of MLX5_TXOFF_CONFIG_xxx
4510  *   values. Should be static to take compile time static configuration
4511  *   advantages.
4512  *
4513  * @return
4514  *   Number of packets successfully transmitted (<= pkts_n).
4515  */
4516 static __rte_always_inline uint16_t
4517 mlx5_tx_burst_tmpl(struct mlx5_txq_data *restrict txq,
4518 		   struct rte_mbuf **restrict pkts,
4519 		   uint16_t pkts_n,
4520 		   unsigned int olx)
4521 {
4522 	struct mlx5_txq_local loc;
4523 	enum mlx5_txcmp_code ret;
4524 	unsigned int part;
4525 
4526 	assert(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail));
4527 	assert(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi));
4528 	if (unlikely(!pkts_n))
4529 		return 0;
4530 	loc.pkts_sent = 0;
4531 	loc.pkts_copy = 0;
4532 	loc.wqe_last = NULL;
4533 
4534 send_loop:
4535 	loc.pkts_loop = loc.pkts_sent;
4536 	/*
4537 	 * Check if there are some CQEs, if any:
4538 	 * - process an encountered errors
4539 	 * - process the completed WQEs
4540 	 * - free related mbufs
4541 	 * - doorbell the NIC about processed CQEs
4542 	 */
4543 	rte_prefetch0(*(pkts + loc.pkts_sent));
4544 	mlx5_tx_handle_completion(txq, olx);
4545 	/*
4546 	 * Calculate the number of available resources - elts and WQEs.
4547 	 * There are two possible different scenarios:
4548 	 * - no data inlining into WQEs, one WQEBB may contains upto
4549 	 *   four packets, in this case elts become scarce resource
4550 	 * - data inlining into WQEs, one packet may require multiple
4551 	 *   WQEBBs, the WQEs become the limiting factor.
4552 	 */
4553 	assert(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail));
4554 	loc.elts_free = txq->elts_s -
4555 				(uint16_t)(txq->elts_head - txq->elts_tail);
4556 	assert(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi));
4557 	loc.wqe_free = txq->wqe_s -
4558 				(uint16_t)(txq->wqe_ci - txq->wqe_pi);
4559 	if (unlikely(!loc.elts_free || !loc.wqe_free))
4560 		return loc.pkts_sent;
4561 	for (;;) {
4562 		/*
4563 		 * Fetch the packet from array. Usually this is
4564 		 * the first packet in series of multi/single
4565 		 * segment packets.
4566 		 */
4567 		loc.mbuf = *(pkts + loc.pkts_sent);
4568 		/* Dedicated branch for multi-segment packets. */
4569 		if (MLX5_TXOFF_CONFIG(MULTI) &&
4570 		    unlikely(NB_SEGS(loc.mbuf) > 1)) {
4571 			/*
4572 			 * Multi-segment packet encountered.
4573 			 * Hardware is able to process it only
4574 			 * with SEND/TSO opcodes, one packet
4575 			 * per WQE, do it in dedicated routine.
4576 			 */
4577 enter_send_multi:
4578 			assert(loc.pkts_sent >= loc.pkts_copy);
4579 			part = loc.pkts_sent - loc.pkts_copy;
4580 			if (!MLX5_TXOFF_CONFIG(INLINE) && part) {
4581 				/*
4582 				 * There are some single-segment mbufs not
4583 				 * stored in elts. The mbufs must be in the
4584 				 * same order as WQEs, so we must copy the
4585 				 * mbufs to elts here, before the coming
4586 				 * multi-segment packet mbufs is appended.
4587 				 */
4588 				mlx5_tx_copy_elts(txq, pkts + loc.pkts_copy,
4589 						  part, olx);
4590 				loc.pkts_copy = loc.pkts_sent;
4591 			}
4592 			assert(pkts_n > loc.pkts_sent);
4593 			ret = mlx5_tx_burst_mseg(txq, pkts, pkts_n, &loc, olx);
4594 			if (!MLX5_TXOFF_CONFIG(INLINE))
4595 				loc.pkts_copy = loc.pkts_sent;
4596 			/*
4597 			 * These returned code checks are supposed
4598 			 * to be optimized out due to routine inlining.
4599 			 */
4600 			if (ret == MLX5_TXCMP_CODE_EXIT) {
4601 				/*
4602 				 * The routine returns this code when
4603 				 * all packets are sent or there is no
4604 				 * enough resources to complete request.
4605 				 */
4606 				break;
4607 			}
4608 			if (ret == MLX5_TXCMP_CODE_ERROR) {
4609 				/*
4610 				 * The routine returns this code when
4611 				 * some error in the incoming packets
4612 				 * format occurred.
4613 				 */
4614 				txq->stats.oerrors++;
4615 				break;
4616 			}
4617 			if (ret == MLX5_TXCMP_CODE_SINGLE) {
4618 				/*
4619 				 * The single-segment packet was encountered
4620 				 * in the array, try to send it with the
4621 				 * best optimized way, possible engaging eMPW.
4622 				 */
4623 				goto enter_send_single;
4624 			}
4625 			if (MLX5_TXOFF_CONFIG(TSO) &&
4626 			    ret == MLX5_TXCMP_CODE_TSO) {
4627 				/*
4628 				 * The single-segment TSO packet was
4629 				 * encountered in the array.
4630 				 */
4631 				goto enter_send_tso;
4632 			}
4633 			/* We must not get here. Something is going wrong. */
4634 			assert(false);
4635 			txq->stats.oerrors++;
4636 			break;
4637 		}
4638 		/* Dedicated branch for single-segment TSO packets. */
4639 		if (MLX5_TXOFF_CONFIG(TSO) &&
4640 		    unlikely(loc.mbuf->ol_flags & PKT_TX_TCP_SEG)) {
4641 			/*
4642 			 * TSO might require special way for inlining
4643 			 * (dedicated parameters) and is sent with
4644 			 * MLX5_OPCODE_TSO opcode only, provide this
4645 			 * in dedicated branch.
4646 			 */
4647 enter_send_tso:
4648 			assert(NB_SEGS(loc.mbuf) == 1);
4649 			assert(pkts_n > loc.pkts_sent);
4650 			ret = mlx5_tx_burst_tso(txq, pkts, pkts_n, &loc, olx);
4651 			/*
4652 			 * These returned code checks are supposed
4653 			 * to be optimized out due to routine inlining.
4654 			 */
4655 			if (ret == MLX5_TXCMP_CODE_EXIT)
4656 				break;
4657 			if (ret == MLX5_TXCMP_CODE_ERROR) {
4658 				txq->stats.oerrors++;
4659 				break;
4660 			}
4661 			if (ret == MLX5_TXCMP_CODE_SINGLE)
4662 				goto enter_send_single;
4663 			if (MLX5_TXOFF_CONFIG(MULTI) &&
4664 			    ret == MLX5_TXCMP_CODE_MULTI) {
4665 				/*
4666 				 * The multi-segment packet was
4667 				 * encountered in the array.
4668 				 */
4669 				goto enter_send_multi;
4670 			}
4671 			/* We must not get here. Something is going wrong. */
4672 			assert(false);
4673 			txq->stats.oerrors++;
4674 			break;
4675 		}
4676 		/*
4677 		 * The dedicated branch for the single-segment packets
4678 		 * without TSO. Often these ones can be sent using
4679 		 * MLX5_OPCODE_EMPW with multiple packets in one WQE.
4680 		 * The routine builds the WQEs till it encounters
4681 		 * the TSO or multi-segment packet (in case if these
4682 		 * offloads are requested at SQ configuration time).
4683 		 */
4684 enter_send_single:
4685 		assert(pkts_n > loc.pkts_sent);
4686 		ret = mlx5_tx_burst_single(txq, pkts, pkts_n, &loc, olx);
4687 		/*
4688 		 * These returned code checks are supposed
4689 		 * to be optimized out due to routine inlining.
4690 		 */
4691 		if (ret == MLX5_TXCMP_CODE_EXIT)
4692 			break;
4693 		if (ret == MLX5_TXCMP_CODE_ERROR) {
4694 			txq->stats.oerrors++;
4695 			break;
4696 		}
4697 		if (MLX5_TXOFF_CONFIG(MULTI) &&
4698 		    ret == MLX5_TXCMP_CODE_MULTI) {
4699 			/*
4700 			 * The multi-segment packet was
4701 			 * encountered in the array.
4702 			 */
4703 			goto enter_send_multi;
4704 		}
4705 		if (MLX5_TXOFF_CONFIG(TSO) &&
4706 		    ret == MLX5_TXCMP_CODE_TSO) {
4707 			/*
4708 			 * The single-segment TSO packet was
4709 			 * encountered in the array.
4710 			 */
4711 			goto enter_send_tso;
4712 		}
4713 		/* We must not get here. Something is going wrong. */
4714 		assert(false);
4715 		txq->stats.oerrors++;
4716 		break;
4717 	}
4718 	/*
4719 	 * Main Tx loop is completed, do the rest:
4720 	 * - set completion request if thresholds are reached
4721 	 * - doorbell the hardware
4722 	 * - copy the rest of mbufs to elts (if any)
4723 	 */
4724 	assert(MLX5_TXOFF_CONFIG(INLINE) || loc.pkts_sent >= loc.pkts_copy);
4725 	/* Take a shortcut if nothing is sent. */
4726 	if (unlikely(loc.pkts_sent == loc.pkts_loop))
4727 		return loc.pkts_sent;
4728 	/*
4729 	 * Ring QP doorbell immediately after WQE building completion
4730 	 * to improve latencies. The pure software related data treatment
4731 	 * can be completed after doorbell. Tx CQEs for this SQ are
4732 	 * processed in this thread only by the polling.
4733 	 */
4734 	mlx5_tx_dbrec_cond_wmb(txq, loc.wqe_last, 0);
4735 	/* Not all of the mbufs may be stored into elts yet. */
4736 	part = MLX5_TXOFF_CONFIG(INLINE) ? 0 : loc.pkts_sent - loc.pkts_copy;
4737 	if (!MLX5_TXOFF_CONFIG(INLINE) && part) {
4738 		/*
4739 		 * There are some single-segment mbufs not stored in elts.
4740 		 * It can be only if the last packet was single-segment.
4741 		 * The copying is gathered into one place due to it is
4742 		 * a good opportunity to optimize that with SIMD.
4743 		 * Unfortunately if inlining is enabled the gaps in
4744 		 * pointer array may happen due to early freeing of the
4745 		 * inlined mbufs.
4746 		 */
4747 		mlx5_tx_copy_elts(txq, pkts + loc.pkts_copy, part, olx);
4748 		loc.pkts_copy = loc.pkts_sent;
4749 	}
4750 #ifdef MLX5_PMD_SOFT_COUNTERS
4751 	/* Increment sent packets counter. */
4752 	txq->stats.opackets += loc.pkts_sent;
4753 #endif
4754 	assert(txq->elts_s >= (uint16_t)(txq->elts_head - txq->elts_tail));
4755 	assert(txq->wqe_s >= (uint16_t)(txq->wqe_ci - txq->wqe_pi));
4756 	if (pkts_n > loc.pkts_sent) {
4757 		/*
4758 		 * If burst size is large there might be no enough CQE
4759 		 * fetched from completion queue and no enough resources
4760 		 * freed to send all the packets.
4761 		 */
4762 		goto send_loop;
4763 	}
4764 	return loc.pkts_sent;
4765 }
4766 
4767 /* Generate routines with Enhanced Multi-Packet Write support. */
4768 MLX5_TXOFF_DECL(full_empw,
4769 		MLX5_TXOFF_CONFIG_FULL | MLX5_TXOFF_CONFIG_EMPW)
4770 
4771 MLX5_TXOFF_DECL(none_empw,
4772 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW)
4773 
4774 MLX5_TXOFF_DECL(md_empw,
4775 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4776 
4777 MLX5_TXOFF_DECL(mt_empw,
4778 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4779 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4780 
4781 MLX5_TXOFF_DECL(mtsc_empw,
4782 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4783 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4784 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4785 
4786 MLX5_TXOFF_DECL(mti_empw,
4787 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4788 		MLX5_TXOFF_CONFIG_INLINE |
4789 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4790 
4791 MLX5_TXOFF_DECL(mtv_empw,
4792 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4793 		MLX5_TXOFF_CONFIG_VLAN |
4794 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4795 
4796 MLX5_TXOFF_DECL(mtiv_empw,
4797 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4798 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
4799 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4800 
4801 MLX5_TXOFF_DECL(sc_empw,
4802 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4803 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4804 
4805 MLX5_TXOFF_DECL(sci_empw,
4806 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4807 		MLX5_TXOFF_CONFIG_INLINE |
4808 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4809 
4810 MLX5_TXOFF_DECL(scv_empw,
4811 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4812 		MLX5_TXOFF_CONFIG_VLAN |
4813 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4814 
4815 MLX5_TXOFF_DECL(sciv_empw,
4816 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4817 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
4818 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4819 
4820 MLX5_TXOFF_DECL(i_empw,
4821 		MLX5_TXOFF_CONFIG_INLINE |
4822 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4823 
4824 MLX5_TXOFF_DECL(v_empw,
4825 		MLX5_TXOFF_CONFIG_VLAN |
4826 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4827 
4828 MLX5_TXOFF_DECL(iv_empw,
4829 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
4830 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4831 
4832 /* Generate routines without Enhanced Multi-Packet Write support. */
4833 MLX5_TXOFF_DECL(full,
4834 		MLX5_TXOFF_CONFIG_FULL)
4835 
4836 MLX5_TXOFF_DECL(none,
4837 		MLX5_TXOFF_CONFIG_NONE)
4838 
4839 MLX5_TXOFF_DECL(md,
4840 		MLX5_TXOFF_CONFIG_METADATA)
4841 
4842 MLX5_TXOFF_DECL(mt,
4843 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4844 		MLX5_TXOFF_CONFIG_METADATA)
4845 
4846 MLX5_TXOFF_DECL(mtsc,
4847 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4848 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4849 		MLX5_TXOFF_CONFIG_METADATA)
4850 
4851 MLX5_TXOFF_DECL(mti,
4852 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4853 		MLX5_TXOFF_CONFIG_INLINE |
4854 		MLX5_TXOFF_CONFIG_METADATA)
4855 
4856 
4857 MLX5_TXOFF_DECL(mtv,
4858 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4859 		MLX5_TXOFF_CONFIG_VLAN |
4860 		MLX5_TXOFF_CONFIG_METADATA)
4861 
4862 
4863 MLX5_TXOFF_DECL(mtiv,
4864 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4865 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
4866 		MLX5_TXOFF_CONFIG_METADATA)
4867 
4868 MLX5_TXOFF_DECL(sc,
4869 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4870 		MLX5_TXOFF_CONFIG_METADATA)
4871 
4872 MLX5_TXOFF_DECL(sci,
4873 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4874 		MLX5_TXOFF_CONFIG_INLINE |
4875 		MLX5_TXOFF_CONFIG_METADATA)
4876 
4877 
4878 MLX5_TXOFF_DECL(scv,
4879 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4880 		MLX5_TXOFF_CONFIG_VLAN |
4881 		MLX5_TXOFF_CONFIG_METADATA)
4882 
4883 
4884 MLX5_TXOFF_DECL(sciv,
4885 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4886 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
4887 		MLX5_TXOFF_CONFIG_METADATA)
4888 
4889 MLX5_TXOFF_DECL(i,
4890 		MLX5_TXOFF_CONFIG_INLINE |
4891 		MLX5_TXOFF_CONFIG_METADATA)
4892 
4893 MLX5_TXOFF_DECL(v,
4894 		MLX5_TXOFF_CONFIG_VLAN |
4895 		MLX5_TXOFF_CONFIG_METADATA)
4896 
4897 MLX5_TXOFF_DECL(iv,
4898 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
4899 		MLX5_TXOFF_CONFIG_METADATA)
4900 
4901 /*
4902  * Array of declared and compiled Tx burst function and corresponding
4903  * supported offloads set. The array is used to select the Tx burst
4904  * function for specified offloads set at Tx queue configuration time.
4905  */
4906 const struct {
4907 	eth_tx_burst_t func;
4908 	unsigned int olx;
4909 } txoff_func[] = {
4910 MLX5_TXOFF_INFO(full_empw,
4911 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4912 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4913 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
4914 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4915 
4916 MLX5_TXOFF_INFO(none_empw,
4917 		MLX5_TXOFF_CONFIG_NONE | MLX5_TXOFF_CONFIG_EMPW)
4918 
4919 MLX5_TXOFF_INFO(md_empw,
4920 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4921 
4922 MLX5_TXOFF_INFO(mt_empw,
4923 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4924 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4925 
4926 MLX5_TXOFF_INFO(mtsc_empw,
4927 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4928 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4929 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4930 
4931 MLX5_TXOFF_INFO(mti_empw,
4932 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4933 		MLX5_TXOFF_CONFIG_INLINE |
4934 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4935 
4936 MLX5_TXOFF_INFO(mtv_empw,
4937 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4938 		MLX5_TXOFF_CONFIG_VLAN |
4939 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4940 
4941 MLX5_TXOFF_INFO(mtiv_empw,
4942 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4943 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
4944 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4945 
4946 MLX5_TXOFF_INFO(sc_empw,
4947 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4948 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4949 
4950 MLX5_TXOFF_INFO(sci_empw,
4951 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4952 		MLX5_TXOFF_CONFIG_INLINE |
4953 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4954 
4955 MLX5_TXOFF_INFO(scv_empw,
4956 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4957 		MLX5_TXOFF_CONFIG_VLAN |
4958 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4959 
4960 MLX5_TXOFF_INFO(sciv_empw,
4961 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4962 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
4963 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4964 
4965 MLX5_TXOFF_INFO(i_empw,
4966 		MLX5_TXOFF_CONFIG_INLINE |
4967 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4968 
4969 MLX5_TXOFF_INFO(v_empw,
4970 		MLX5_TXOFF_CONFIG_VLAN |
4971 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4972 
4973 MLX5_TXOFF_INFO(iv_empw,
4974 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
4975 		MLX5_TXOFF_CONFIG_METADATA | MLX5_TXOFF_CONFIG_EMPW)
4976 
4977 MLX5_TXOFF_INFO(full,
4978 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4979 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4980 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
4981 		MLX5_TXOFF_CONFIG_METADATA)
4982 
4983 MLX5_TXOFF_INFO(none,
4984 		MLX5_TXOFF_CONFIG_NONE)
4985 
4986 MLX5_TXOFF_INFO(md,
4987 		MLX5_TXOFF_CONFIG_METADATA)
4988 
4989 MLX5_TXOFF_INFO(mt,
4990 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4991 		MLX5_TXOFF_CONFIG_METADATA)
4992 
4993 MLX5_TXOFF_INFO(mtsc,
4994 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
4995 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
4996 		MLX5_TXOFF_CONFIG_METADATA)
4997 
4998 MLX5_TXOFF_INFO(mti,
4999 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5000 		MLX5_TXOFF_CONFIG_INLINE |
5001 		MLX5_TXOFF_CONFIG_METADATA)
5002 
5003 
5004 MLX5_TXOFF_INFO(mtv,
5005 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5006 		MLX5_TXOFF_CONFIG_VLAN |
5007 		MLX5_TXOFF_CONFIG_METADATA)
5008 
5009 MLX5_TXOFF_INFO(mtiv,
5010 		MLX5_TXOFF_CONFIG_MULTI | MLX5_TXOFF_CONFIG_TSO |
5011 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5012 		MLX5_TXOFF_CONFIG_METADATA)
5013 
5014 MLX5_TXOFF_INFO(sc,
5015 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5016 		MLX5_TXOFF_CONFIG_METADATA)
5017 
5018 MLX5_TXOFF_INFO(sci,
5019 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5020 		MLX5_TXOFF_CONFIG_INLINE |
5021 		MLX5_TXOFF_CONFIG_METADATA)
5022 
5023 MLX5_TXOFF_INFO(scv,
5024 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5025 		MLX5_TXOFF_CONFIG_VLAN |
5026 		MLX5_TXOFF_CONFIG_METADATA)
5027 
5028 MLX5_TXOFF_INFO(sciv,
5029 		MLX5_TXOFF_CONFIG_SWP |	MLX5_TXOFF_CONFIG_CSUM |
5030 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5031 		MLX5_TXOFF_CONFIG_METADATA)
5032 
5033 MLX5_TXOFF_INFO(i,
5034 		MLX5_TXOFF_CONFIG_INLINE |
5035 		MLX5_TXOFF_CONFIG_METADATA)
5036 
5037 MLX5_TXOFF_INFO(v,
5038 		MLX5_TXOFF_CONFIG_VLAN |
5039 		MLX5_TXOFF_CONFIG_METADATA)
5040 
5041 MLX5_TXOFF_INFO(iv,
5042 		MLX5_TXOFF_CONFIG_INLINE | MLX5_TXOFF_CONFIG_VLAN |
5043 		MLX5_TXOFF_CONFIG_METADATA)
5044 };
5045 
5046 /**
5047  * Configure the Tx function to use. The routine checks configured
5048  * Tx offloads for the device and selects appropriate Tx burst
5049  * routine. There are multiple Tx burst routines compiled from
5050  * the same template in the most optimal way for the dedicated
5051  * Tx offloads set.
5052  *
5053  * @param dev
5054  *   Pointer to private data structure.
5055  *
5056  * @return
5057  *   Pointer to selected Tx burst function.
5058  */
5059 eth_tx_burst_t
5060 mlx5_select_tx_function(struct rte_eth_dev *dev)
5061 {
5062 	struct mlx5_priv *priv = dev->data->dev_private;
5063 	struct mlx5_dev_config *config = &priv->config;
5064 	uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
5065 	unsigned int diff = 0, olx = 0, i, m;
5066 
5067 	static_assert(MLX5_WQE_SIZE_MAX / MLX5_WSEG_SIZE <=
5068 		      MLX5_DSEG_MAX, "invalid WQE max size");
5069 	static_assert(MLX5_WQE_CSEG_SIZE == MLX5_WSEG_SIZE,
5070 		      "invalid WQE Control Segment size");
5071 	static_assert(MLX5_WQE_ESEG_SIZE == MLX5_WSEG_SIZE,
5072 		      "invalid WQE Ethernet Segment size");
5073 	static_assert(MLX5_WQE_DSEG_SIZE == MLX5_WSEG_SIZE,
5074 		      "invalid WQE Data Segment size");
5075 	static_assert(MLX5_WQE_SIZE == 4 * MLX5_WSEG_SIZE,
5076 		      "invalid WQE size");
5077 	assert(priv);
5078 	if (tx_offloads & DEV_TX_OFFLOAD_MULTI_SEGS) {
5079 		/* We should support Multi-Segment Packets. */
5080 		olx |= MLX5_TXOFF_CONFIG_MULTI;
5081 	}
5082 	if (tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO |
5083 			   DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
5084 			   DEV_TX_OFFLOAD_GRE_TNL_TSO |
5085 			   DEV_TX_OFFLOAD_IP_TNL_TSO |
5086 			   DEV_TX_OFFLOAD_UDP_TNL_TSO)) {
5087 		/* We should support TCP Send Offload. */
5088 		olx |= MLX5_TXOFF_CONFIG_TSO;
5089 	}
5090 	if (tx_offloads & (DEV_TX_OFFLOAD_IP_TNL_TSO |
5091 			   DEV_TX_OFFLOAD_UDP_TNL_TSO |
5092 			   DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)) {
5093 		/* We should support Software Parser for Tunnels. */
5094 		olx |= MLX5_TXOFF_CONFIG_SWP;
5095 	}
5096 	if (tx_offloads & (DEV_TX_OFFLOAD_IPV4_CKSUM |
5097 			   DEV_TX_OFFLOAD_UDP_CKSUM |
5098 			   DEV_TX_OFFLOAD_TCP_CKSUM |
5099 			   DEV_TX_OFFLOAD_OUTER_IPV4_CKSUM)) {
5100 		/* We should support IP/TCP/UDP Checksums. */
5101 		olx |= MLX5_TXOFF_CONFIG_CSUM;
5102 	}
5103 	if (tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT) {
5104 		/* We should support VLAN insertion. */
5105 		olx |= MLX5_TXOFF_CONFIG_VLAN;
5106 	}
5107 	if (priv->txqs_n && (*priv->txqs)[0]) {
5108 		struct mlx5_txq_data *txd = (*priv->txqs)[0];
5109 
5110 		if (txd->inlen_send) {
5111 			/*
5112 			 * Check the data inline requirements. Data inline
5113 			 * is enabled on per device basis, we can check
5114 			 * the first Tx queue only.
5115 			 *
5116 			 * If device does not support VLAN insertion in WQE
5117 			 * and some queues are requested to perform VLAN
5118 			 * insertion offload than inline must be enabled.
5119 			 */
5120 			olx |= MLX5_TXOFF_CONFIG_INLINE;
5121 		}
5122 	}
5123 	if (config->mps == MLX5_MPW_ENHANCED &&
5124 	    config->txq_inline_min <= 0) {
5125 		/*
5126 		 * The NIC supports Enhanced Multi-Packet Write.
5127 		 * We do not support legacy MPW due to its
5128 		 * hardware related problems, so we just ignore
5129 		 * legacy MLX5_MPW settings. There should be no
5130 		 * minimal required inline data.
5131 		 */
5132 		olx |= MLX5_TXOFF_CONFIG_EMPW;
5133 	}
5134 	if (tx_offloads & DEV_TX_OFFLOAD_MATCH_METADATA) {
5135 		/* We should support Flow metadata. */
5136 		olx |= MLX5_TXOFF_CONFIG_METADATA;
5137 	}
5138 	/*
5139 	 * Scan the routines table to find the minimal
5140 	 * satisfying routine with requested offloads.
5141 	 */
5142 	m = RTE_DIM(txoff_func);
5143 	for (i = 0; i < RTE_DIM(txoff_func); i++) {
5144 		unsigned int tmp;
5145 
5146 		tmp = txoff_func[i].olx;
5147 		if (tmp == olx) {
5148 			/* Meets requested offloads exactly.*/
5149 			m = i;
5150 			break;
5151 		}
5152 		if ((tmp & olx) != olx) {
5153 			/* Does not meet requested offloads at all. */
5154 			continue;
5155 		}
5156 		if ((olx ^ tmp) & MLX5_TXOFF_CONFIG_EMPW)
5157 			/* Do not enable eMPW if not configured. */
5158 			continue;
5159 		if ((olx ^ tmp) & MLX5_TXOFF_CONFIG_INLINE)
5160 			/* Do not enable inlining if not configured. */
5161 			continue;
5162 		/*
5163 		 * Some routine meets the requirements.
5164 		 * Check whether it has minimal amount
5165 		 * of not requested offloads.
5166 		 */
5167 		tmp = __builtin_popcountl(tmp & ~olx);
5168 		if (m >= RTE_DIM(txoff_func) || tmp < diff) {
5169 			/* First or better match, save and continue. */
5170 			m = i;
5171 			diff = tmp;
5172 			continue;
5173 		}
5174 		if (tmp == diff) {
5175 			tmp = txoff_func[i].olx ^ txoff_func[m].olx;
5176 			if (__builtin_ffsl(txoff_func[i].olx & ~tmp) <
5177 			    __builtin_ffsl(txoff_func[m].olx & ~tmp)) {
5178 				/* Lighter not requested offload. */
5179 				m = i;
5180 			}
5181 		}
5182 	}
5183 	if (m >= RTE_DIM(txoff_func)) {
5184 		DRV_LOG(DEBUG, "port %u has no selected Tx function"
5185 			       " for requested offloads %04X",
5186 				dev->data->port_id, olx);
5187 		return NULL;
5188 	}
5189 	DRV_LOG(DEBUG, "port %u has selected Tx function"
5190 		       " supporting offloads %04X/%04X",
5191 			dev->data->port_id, olx, txoff_func[m].olx);
5192 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_MULTI)
5193 		DRV_LOG(DEBUG, "\tMULTI (multi segment)");
5194 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_TSO)
5195 		DRV_LOG(DEBUG, "\tTSO   (TCP send offload)");
5196 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_SWP)
5197 		DRV_LOG(DEBUG, "\tSWP   (software parser)");
5198 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_CSUM)
5199 		DRV_LOG(DEBUG, "\tCSUM  (checksum offload)");
5200 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_INLINE)
5201 		DRV_LOG(DEBUG, "\tINLIN (inline data)");
5202 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_VLAN)
5203 		DRV_LOG(DEBUG, "\tVLANI (VLAN insertion)");
5204 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_METADATA)
5205 		DRV_LOG(DEBUG, "\tMETAD (tx Flow metadata)");
5206 	if (txoff_func[m].olx & MLX5_TXOFF_CONFIG_EMPW)
5207 		DRV_LOG(DEBUG, "\tEMPW  (Enhanced MPW)");
5208 	return txoff_func[m].func;
5209 }
5210