xref: /dpdk/drivers/net/mlx5/mlx5_rxtx.c (revision bb44fb6fe7713ddcd023d5b9bacadf074d68092e)
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015 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 
29 #include "mlx5.h"
30 #include "mlx5_utils.h"
31 #include "mlx5_rxtx.h"
32 #include "mlx5_autoconf.h"
33 #include "mlx5_defs.h"
34 #include "mlx5_prm.h"
35 
36 static __rte_always_inline uint32_t
37 rxq_cq_to_pkt_type(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe);
38 
39 static __rte_always_inline int
40 mlx5_rx_poll_len(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
41 		 uint16_t cqe_cnt, volatile struct mlx5_mini_cqe8 **mcqe);
42 
43 static __rte_always_inline uint32_t
44 rxq_cq_to_ol_flags(volatile struct mlx5_cqe *cqe);
45 
46 static __rte_always_inline void
47 rxq_cq_to_mbuf(struct mlx5_rxq_data *rxq, struct rte_mbuf *pkt,
48 	       volatile struct mlx5_cqe *cqe, uint32_t rss_hash_res);
49 
50 static __rte_always_inline void
51 mprq_buf_replace(struct mlx5_rxq_data *rxq, uint16_t rq_idx);
52 
53 uint32_t mlx5_ptype_table[] __rte_cache_aligned = {
54 	[0xff] = RTE_PTYPE_ALL_MASK, /* Last entry for errored packet. */
55 };
56 
57 uint8_t mlx5_cksum_table[1 << 10] __rte_cache_aligned;
58 uint8_t mlx5_swp_types_table[1 << 10] __rte_cache_aligned;
59 
60 /**
61  * Build a table to translate Rx completion flags to packet type.
62  *
63  * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
64  */
65 void
66 mlx5_set_ptype_table(void)
67 {
68 	unsigned int i;
69 	uint32_t (*p)[RTE_DIM(mlx5_ptype_table)] = &mlx5_ptype_table;
70 
71 	/* Last entry must not be overwritten, reserved for errored packet. */
72 	for (i = 0; i < RTE_DIM(mlx5_ptype_table) - 1; ++i)
73 		(*p)[i] = RTE_PTYPE_UNKNOWN;
74 	/*
75 	 * The index to the array should have:
76 	 * bit[1:0] = l3_hdr_type
77 	 * bit[4:2] = l4_hdr_type
78 	 * bit[5] = ip_frag
79 	 * bit[6] = tunneled
80 	 * bit[7] = outer_l3_type
81 	 */
82 	/* L2 */
83 	(*p)[0x00] = RTE_PTYPE_L2_ETHER;
84 	/* L3 */
85 	(*p)[0x01] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
86 		     RTE_PTYPE_L4_NONFRAG;
87 	(*p)[0x02] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
88 		     RTE_PTYPE_L4_NONFRAG;
89 	/* Fragmented */
90 	(*p)[0x21] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
91 		     RTE_PTYPE_L4_FRAG;
92 	(*p)[0x22] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
93 		     RTE_PTYPE_L4_FRAG;
94 	/* TCP */
95 	(*p)[0x05] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
96 		     RTE_PTYPE_L4_TCP;
97 	(*p)[0x06] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
98 		     RTE_PTYPE_L4_TCP;
99 	(*p)[0x0d] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
100 		     RTE_PTYPE_L4_TCP;
101 	(*p)[0x0e] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
102 		     RTE_PTYPE_L4_TCP;
103 	(*p)[0x11] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
104 		     RTE_PTYPE_L4_TCP;
105 	(*p)[0x12] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
106 		     RTE_PTYPE_L4_TCP;
107 	/* UDP */
108 	(*p)[0x09] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
109 		     RTE_PTYPE_L4_UDP;
110 	(*p)[0x0a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
111 		     RTE_PTYPE_L4_UDP;
112 	/* Repeat with outer_l3_type being set. Just in case. */
113 	(*p)[0x81] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
114 		     RTE_PTYPE_L4_NONFRAG;
115 	(*p)[0x82] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
116 		     RTE_PTYPE_L4_NONFRAG;
117 	(*p)[0xa1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
118 		     RTE_PTYPE_L4_FRAG;
119 	(*p)[0xa2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
120 		     RTE_PTYPE_L4_FRAG;
121 	(*p)[0x85] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
122 		     RTE_PTYPE_L4_TCP;
123 	(*p)[0x86] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
124 		     RTE_PTYPE_L4_TCP;
125 	(*p)[0x8d] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
126 		     RTE_PTYPE_L4_TCP;
127 	(*p)[0x8e] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
128 		     RTE_PTYPE_L4_TCP;
129 	(*p)[0x91] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
130 		     RTE_PTYPE_L4_TCP;
131 	(*p)[0x92] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
132 		     RTE_PTYPE_L4_TCP;
133 	(*p)[0x89] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
134 		     RTE_PTYPE_L4_UDP;
135 	(*p)[0x8a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
136 		     RTE_PTYPE_L4_UDP;
137 	/* Tunneled - L3 */
138 	(*p)[0x40] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN;
139 	(*p)[0x41] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
140 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
141 		     RTE_PTYPE_INNER_L4_NONFRAG;
142 	(*p)[0x42] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
143 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
144 		     RTE_PTYPE_INNER_L4_NONFRAG;
145 	(*p)[0xc0] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN;
146 	(*p)[0xc1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
147 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
148 		     RTE_PTYPE_INNER_L4_NONFRAG;
149 	(*p)[0xc2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
150 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
151 		     RTE_PTYPE_INNER_L4_NONFRAG;
152 	/* Tunneled - Fragmented */
153 	(*p)[0x61] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
154 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
155 		     RTE_PTYPE_INNER_L4_FRAG;
156 	(*p)[0x62] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
157 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
158 		     RTE_PTYPE_INNER_L4_FRAG;
159 	(*p)[0xe1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
160 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
161 		     RTE_PTYPE_INNER_L4_FRAG;
162 	(*p)[0xe2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
163 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
164 		     RTE_PTYPE_INNER_L4_FRAG;
165 	/* Tunneled - TCP */
166 	(*p)[0x45] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
167 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
168 		     RTE_PTYPE_INNER_L4_TCP;
169 	(*p)[0x46] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
170 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
171 		     RTE_PTYPE_INNER_L4_TCP;
172 	(*p)[0x4d] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
173 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
174 		     RTE_PTYPE_INNER_L4_TCP;
175 	(*p)[0x4e] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
176 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
177 		     RTE_PTYPE_INNER_L4_TCP;
178 	(*p)[0x51] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
179 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
180 		     RTE_PTYPE_INNER_L4_TCP;
181 	(*p)[0x52] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
182 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
183 		     RTE_PTYPE_INNER_L4_TCP;
184 	(*p)[0xc5] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
185 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
186 		     RTE_PTYPE_INNER_L4_TCP;
187 	(*p)[0xc6] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
188 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
189 		     RTE_PTYPE_INNER_L4_TCP;
190 	(*p)[0xcd] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
191 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
192 		     RTE_PTYPE_INNER_L4_TCP;
193 	(*p)[0xce] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
194 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
195 		     RTE_PTYPE_INNER_L4_TCP;
196 	(*p)[0xd1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
197 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
198 		     RTE_PTYPE_INNER_L4_TCP;
199 	(*p)[0xd2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
200 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
201 		     RTE_PTYPE_INNER_L4_TCP;
202 	/* Tunneled - UDP */
203 	(*p)[0x49] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
204 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
205 		     RTE_PTYPE_INNER_L4_UDP;
206 	(*p)[0x4a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
207 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
208 		     RTE_PTYPE_INNER_L4_UDP;
209 	(*p)[0xc9] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
210 		     RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
211 		     RTE_PTYPE_INNER_L4_UDP;
212 	(*p)[0xca] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
213 		     RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
214 		     RTE_PTYPE_INNER_L4_UDP;
215 }
216 
217 /**
218  * Build a table to translate packet to checksum type of Verbs.
219  */
220 void
221 mlx5_set_cksum_table(void)
222 {
223 	unsigned int i;
224 	uint8_t v;
225 
226 	/*
227 	 * The index should have:
228 	 * bit[0] = PKT_TX_TCP_SEG
229 	 * bit[2:3] = PKT_TX_UDP_CKSUM, PKT_TX_TCP_CKSUM
230 	 * bit[4] = PKT_TX_IP_CKSUM
231 	 * bit[8] = PKT_TX_OUTER_IP_CKSUM
232 	 * bit[9] = tunnel
233 	 */
234 	for (i = 0; i < RTE_DIM(mlx5_cksum_table); ++i) {
235 		v = 0;
236 		if (i & (1 << 9)) {
237 			/* Tunneled packet. */
238 			if (i & (1 << 8)) /* Outer IP. */
239 				v |= MLX5_ETH_WQE_L3_CSUM;
240 			if (i & (1 << 4)) /* Inner IP. */
241 				v |= MLX5_ETH_WQE_L3_INNER_CSUM;
242 			if (i & (3 << 2 | 1 << 0)) /* L4 or TSO. */
243 				v |= MLX5_ETH_WQE_L4_INNER_CSUM;
244 		} else {
245 			/* No tunnel. */
246 			if (i & (1 << 4)) /* IP. */
247 				v |= MLX5_ETH_WQE_L3_CSUM;
248 			if (i & (3 << 2 | 1 << 0)) /* L4 or TSO. */
249 				v |= MLX5_ETH_WQE_L4_CSUM;
250 		}
251 		mlx5_cksum_table[i] = v;
252 	}
253 }
254 
255 /**
256  * Build a table to translate packet type of mbuf to SWP type of Verbs.
257  */
258 void
259 mlx5_set_swp_types_table(void)
260 {
261 	unsigned int i;
262 	uint8_t v;
263 
264 	/*
265 	 * The index should have:
266 	 * bit[0:1] = PKT_TX_L4_MASK
267 	 * bit[4] = PKT_TX_IPV6
268 	 * bit[8] = PKT_TX_OUTER_IPV6
269 	 * bit[9] = PKT_TX_OUTER_UDP
270 	 */
271 	for (i = 0; i < RTE_DIM(mlx5_swp_types_table); ++i) {
272 		v = 0;
273 		if (i & (1 << 8))
274 			v |= MLX5_ETH_WQE_L3_OUTER_IPV6;
275 		if (i & (1 << 9))
276 			v |= MLX5_ETH_WQE_L4_OUTER_UDP;
277 		if (i & (1 << 4))
278 			v |= MLX5_ETH_WQE_L3_INNER_IPV6;
279 		if ((i & 3) == (PKT_TX_UDP_CKSUM >> 52))
280 			v |= MLX5_ETH_WQE_L4_INNER_UDP;
281 		mlx5_swp_types_table[i] = v;
282 	}
283 }
284 
285 /**
286  * Return the size of tailroom of WQ.
287  *
288  * @param txq
289  *   Pointer to TX queue structure.
290  * @param addr
291  *   Pointer to tail of WQ.
292  *
293  * @return
294  *   Size of tailroom.
295  */
296 static inline size_t
297 tx_mlx5_wq_tailroom(struct mlx5_txq_data *txq, void *addr)
298 {
299 	size_t tailroom;
300 	tailroom = (uintptr_t)(txq->wqes) +
301 		   (1 << txq->wqe_n) * MLX5_WQE_SIZE -
302 		   (uintptr_t)addr;
303 	return tailroom;
304 }
305 
306 /**
307  * Copy data to tailroom of circular queue.
308  *
309  * @param dst
310  *   Pointer to destination.
311  * @param src
312  *   Pointer to source.
313  * @param n
314  *   Number of bytes to copy.
315  * @param base
316  *   Pointer to head of queue.
317  * @param tailroom
318  *   Size of tailroom from dst.
319  *
320  * @return
321  *   Pointer after copied data.
322  */
323 static inline void *
324 mlx5_copy_to_wq(void *dst, const void *src, size_t n,
325 		void *base, size_t tailroom)
326 {
327 	void *ret;
328 
329 	if (n > tailroom) {
330 		rte_memcpy(dst, src, tailroom);
331 		rte_memcpy(base, (void *)((uintptr_t)src + tailroom),
332 			   n - tailroom);
333 		ret = (uint8_t *)base + n - tailroom;
334 	} else {
335 		rte_memcpy(dst, src, n);
336 		ret = (n == tailroom) ? base : (uint8_t *)dst + n;
337 	}
338 	return ret;
339 }
340 
341 /**
342  * Inline TSO headers into WQE.
343  *
344  * @return
345  *   0 on success, negative errno value on failure.
346  */
347 static int
348 inline_tso(struct mlx5_txq_data *txq, struct rte_mbuf *buf,
349 	   uint32_t *length,
350 	   uintptr_t *addr,
351 	   uint16_t *pkt_inline_sz,
352 	   uint8_t **raw,
353 	   uint16_t *max_wqe,
354 	   uint16_t *tso_segsz,
355 	   uint16_t *tso_header_sz)
356 {
357 	uintptr_t end = (uintptr_t)(((uintptr_t)txq->wqes) +
358 				    (1 << txq->wqe_n) * MLX5_WQE_SIZE);
359 	unsigned int copy_b;
360 	uint8_t vlan_sz = (buf->ol_flags & PKT_TX_VLAN_PKT) ? 4 : 0;
361 	const uint8_t tunneled = txq->tunnel_en && (buf->ol_flags &
362 				 PKT_TX_TUNNEL_MASK);
363 	uint16_t n_wqe;
364 
365 	*tso_segsz = buf->tso_segsz;
366 	*tso_header_sz = buf->l2_len + vlan_sz + buf->l3_len + buf->l4_len;
367 	if (unlikely(*tso_segsz == 0 || *tso_header_sz == 0)) {
368 		txq->stats.oerrors++;
369 		return -EINVAL;
370 	}
371 	if (tunneled)
372 		*tso_header_sz += buf->outer_l2_len + buf->outer_l3_len;
373 	/* First seg must contain all TSO headers. */
374 	if (unlikely(*tso_header_sz > MLX5_MAX_TSO_HEADER) ||
375 		     *tso_header_sz > DATA_LEN(buf)) {
376 		txq->stats.oerrors++;
377 		return -EINVAL;
378 	}
379 	copy_b = *tso_header_sz - *pkt_inline_sz;
380 	if (!copy_b || ((end - (uintptr_t)*raw) < copy_b))
381 		return -EAGAIN;
382 	n_wqe = (MLX5_WQE_DS(copy_b) - 1 + 3) / 4;
383 	if (unlikely(*max_wqe < n_wqe))
384 		return -EINVAL;
385 	*max_wqe -= n_wqe;
386 	rte_memcpy((void *)*raw, (void *)*addr, copy_b);
387 	*length -= copy_b;
388 	*addr += copy_b;
389 	copy_b = MLX5_WQE_DS(copy_b) * MLX5_WQE_DWORD_SIZE;
390 	*pkt_inline_sz += copy_b;
391 	*raw += copy_b;
392 	return 0;
393 }
394 
395 /**
396  * DPDK callback to check the status of a tx descriptor.
397  *
398  * @param tx_queue
399  *   The tx queue.
400  * @param[in] offset
401  *   The index of the descriptor in the ring.
402  *
403  * @return
404  *   The status of the tx descriptor.
405  */
406 int
407 mlx5_tx_descriptor_status(void *tx_queue, uint16_t offset)
408 {
409 	struct mlx5_txq_data *txq = tx_queue;
410 	uint16_t used;
411 
412 	mlx5_tx_complete(txq);
413 	used = txq->elts_head - txq->elts_tail;
414 	if (offset < used)
415 		return RTE_ETH_TX_DESC_FULL;
416 	return RTE_ETH_TX_DESC_DONE;
417 }
418 
419 /**
420  * DPDK callback to check the status of a rx descriptor.
421  *
422  * @param rx_queue
423  *   The rx queue.
424  * @param[in] offset
425  *   The index of the descriptor in the ring.
426  *
427  * @return
428  *   The status of the tx descriptor.
429  */
430 int
431 mlx5_rx_descriptor_status(void *rx_queue, uint16_t offset)
432 {
433 	struct mlx5_rxq_data *rxq = rx_queue;
434 	struct rxq_zip *zip = &rxq->zip;
435 	volatile struct mlx5_cqe *cqe;
436 	const unsigned int cqe_n = (1 << rxq->cqe_n);
437 	const unsigned int cqe_cnt = cqe_n - 1;
438 	unsigned int cq_ci;
439 	unsigned int used;
440 
441 	/* if we are processing a compressed cqe */
442 	if (zip->ai) {
443 		used = zip->cqe_cnt - zip->ca;
444 		cq_ci = zip->cq_ci;
445 	} else {
446 		used = 0;
447 		cq_ci = rxq->cq_ci;
448 	}
449 	cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
450 	while (check_cqe(cqe, cqe_n, cq_ci) == 0) {
451 		int8_t op_own;
452 		unsigned int n;
453 
454 		op_own = cqe->op_own;
455 		if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED)
456 			n = rte_be_to_cpu_32(cqe->byte_cnt);
457 		else
458 			n = 1;
459 		cq_ci += n;
460 		used += n;
461 		cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
462 	}
463 	used = RTE_MIN(used, (1U << rxq->elts_n) - 1);
464 	if (offset < used)
465 		return RTE_ETH_RX_DESC_DONE;
466 	return RTE_ETH_RX_DESC_AVAIL;
467 }
468 
469 /**
470  * DPDK callback for TX.
471  *
472  * @param dpdk_txq
473  *   Generic pointer to TX queue structure.
474  * @param[in] pkts
475  *   Packets to transmit.
476  * @param pkts_n
477  *   Number of packets in array.
478  *
479  * @return
480  *   Number of packets successfully transmitted (<= pkts_n).
481  */
482 uint16_t
483 mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
484 {
485 	struct mlx5_txq_data *txq = (struct mlx5_txq_data *)dpdk_txq;
486 	uint16_t elts_head = txq->elts_head;
487 	const uint16_t elts_n = 1 << txq->elts_n;
488 	const uint16_t elts_m = elts_n - 1;
489 	unsigned int i = 0;
490 	unsigned int j = 0;
491 	unsigned int k = 0;
492 	uint16_t max_elts;
493 	uint16_t max_wqe;
494 	unsigned int comp;
495 	volatile struct mlx5_wqe_ctrl *last_wqe = NULL;
496 	unsigned int segs_n = 0;
497 	const unsigned int max_inline = txq->max_inline;
498 	uint64_t addr_64;
499 
500 	if (unlikely(!pkts_n))
501 		return 0;
502 	/* Prefetch first packet cacheline. */
503 	rte_prefetch0(*pkts);
504 	/* Start processing. */
505 	mlx5_tx_complete(txq);
506 	max_elts = (elts_n - (elts_head - txq->elts_tail));
507 	/* A CQE slot must always be available. */
508 	assert((1u << txq->cqe_n) - (txq->cq_pi - txq->cq_ci));
509 	max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
510 	if (unlikely(!max_wqe))
511 		return 0;
512 	do {
513 		struct rte_mbuf *buf = *pkts; /* First_seg. */
514 		uint8_t *raw;
515 		volatile struct mlx5_wqe_v *wqe = NULL;
516 		volatile rte_v128u32_t *dseg = NULL;
517 		uint32_t length;
518 		unsigned int ds = 0;
519 		unsigned int sg = 0; /* counter of additional segs attached. */
520 		uintptr_t addr;
521 		uint16_t pkt_inline_sz = MLX5_WQE_DWORD_SIZE + 2;
522 		uint16_t tso_header_sz = 0;
523 		uint16_t ehdr;
524 		uint8_t cs_flags;
525 		uint8_t tso = txq->tso_en && (buf->ol_flags & PKT_TX_TCP_SEG);
526 		uint32_t swp_offsets = 0;
527 		uint8_t swp_types = 0;
528 		uint16_t tso_segsz = 0;
529 #ifdef MLX5_PMD_SOFT_COUNTERS
530 		uint32_t total_length = 0;
531 #endif
532 		int ret;
533 
534 		segs_n = buf->nb_segs;
535 		/*
536 		 * Make sure there is enough room to store this packet and
537 		 * that one ring entry remains unused.
538 		 */
539 		assert(segs_n);
540 		if (max_elts < segs_n)
541 			break;
542 		max_elts -= segs_n;
543 		sg = --segs_n;
544 		if (unlikely(--max_wqe == 0))
545 			break;
546 		wqe = (volatile struct mlx5_wqe_v *)
547 			tx_mlx5_wqe(txq, txq->wqe_ci);
548 		rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci + 1));
549 		if (pkts_n - i > 1)
550 			rte_prefetch0(*(pkts + 1));
551 		addr = rte_pktmbuf_mtod(buf, uintptr_t);
552 		length = DATA_LEN(buf);
553 		ehdr = (((uint8_t *)addr)[1] << 8) |
554 		       ((uint8_t *)addr)[0];
555 #ifdef MLX5_PMD_SOFT_COUNTERS
556 		total_length = length;
557 #endif
558 		if (length < (MLX5_WQE_DWORD_SIZE + 2)) {
559 			txq->stats.oerrors++;
560 			break;
561 		}
562 		/* Update element. */
563 		(*txq->elts)[elts_head & elts_m] = buf;
564 		/* Prefetch next buffer data. */
565 		if (pkts_n - i > 1)
566 			rte_prefetch0(
567 			    rte_pktmbuf_mtod(*(pkts + 1), volatile void *));
568 		cs_flags = txq_ol_cksum_to_cs(buf);
569 		txq_mbuf_to_swp(txq, buf, (uint8_t *)&swp_offsets, &swp_types);
570 		raw = ((uint8_t *)(uintptr_t)wqe) + 2 * MLX5_WQE_DWORD_SIZE;
571 		/* Replace the Ethernet type by the VLAN if necessary. */
572 		if (buf->ol_flags & PKT_TX_VLAN_PKT) {
573 			uint32_t vlan = rte_cpu_to_be_32(0x81000000 |
574 							 buf->vlan_tci);
575 			unsigned int len = 2 * ETHER_ADDR_LEN - 2;
576 
577 			addr += 2;
578 			length -= 2;
579 			/* Copy Destination and source mac address. */
580 			memcpy((uint8_t *)raw, ((uint8_t *)addr), len);
581 			/* Copy VLAN. */
582 			memcpy((uint8_t *)raw + len, &vlan, sizeof(vlan));
583 			/* Copy missing two bytes to end the DSeg. */
584 			memcpy((uint8_t *)raw + len + sizeof(vlan),
585 			       ((uint8_t *)addr) + len, 2);
586 			addr += len + 2;
587 			length -= (len + 2);
588 		} else {
589 			memcpy((uint8_t *)raw, ((uint8_t *)addr) + 2,
590 			       MLX5_WQE_DWORD_SIZE);
591 			length -= pkt_inline_sz;
592 			addr += pkt_inline_sz;
593 		}
594 		raw += MLX5_WQE_DWORD_SIZE;
595 		if (tso) {
596 			ret = inline_tso(txq, buf, &length,
597 					 &addr, &pkt_inline_sz,
598 					 &raw, &max_wqe,
599 					 &tso_segsz, &tso_header_sz);
600 			if (ret == -EINVAL) {
601 				break;
602 			} else if (ret == -EAGAIN) {
603 				/* NOP WQE. */
604 				wqe->ctrl = (rte_v128u32_t){
605 					rte_cpu_to_be_32(txq->wqe_ci << 8),
606 					rte_cpu_to_be_32(txq->qp_num_8s | 1),
607 					0,
608 					0,
609 				};
610 				ds = 1;
611 #ifdef MLX5_PMD_SOFT_COUNTERS
612 				total_length = 0;
613 #endif
614 				k++;
615 				goto next_wqe;
616 			}
617 		}
618 		/* Inline if enough room. */
619 		if (max_inline || tso) {
620 			uint32_t inl = 0;
621 			uintptr_t end = (uintptr_t)
622 				(((uintptr_t)txq->wqes) +
623 				 (1 << txq->wqe_n) * MLX5_WQE_SIZE);
624 			unsigned int inline_room = max_inline *
625 						   RTE_CACHE_LINE_SIZE -
626 						   (pkt_inline_sz - 2) -
627 						   !!tso * sizeof(inl);
628 			uintptr_t addr_end;
629 			unsigned int copy_b;
630 
631 pkt_inline:
632 			addr_end = RTE_ALIGN_FLOOR(addr + inline_room,
633 						   RTE_CACHE_LINE_SIZE);
634 			copy_b = (addr_end > addr) ?
635 				 RTE_MIN((addr_end - addr), length) : 0;
636 			if (copy_b && ((end - (uintptr_t)raw) > copy_b)) {
637 				/*
638 				 * One Dseg remains in the current WQE.  To
639 				 * keep the computation positive, it is
640 				 * removed after the bytes to Dseg conversion.
641 				 */
642 				uint16_t n = (MLX5_WQE_DS(copy_b) - 1 + 3) / 4;
643 
644 				if (unlikely(max_wqe < n))
645 					break;
646 				max_wqe -= n;
647 				if (tso) {
648 					assert(inl == 0);
649 					inl = rte_cpu_to_be_32(copy_b |
650 							       MLX5_INLINE_SEG);
651 					rte_memcpy((void *)raw,
652 						   (void *)&inl, sizeof(inl));
653 					raw += sizeof(inl);
654 					pkt_inline_sz += sizeof(inl);
655 				}
656 				rte_memcpy((void *)raw, (void *)addr, copy_b);
657 				addr += copy_b;
658 				length -= copy_b;
659 				pkt_inline_sz += copy_b;
660 			}
661 			/*
662 			 * 2 DWORDs consumed by the WQE header + ETH segment +
663 			 * the size of the inline part of the packet.
664 			 */
665 			ds = 2 + MLX5_WQE_DS(pkt_inline_sz - 2);
666 			if (length > 0) {
667 				if (ds % (MLX5_WQE_SIZE /
668 					  MLX5_WQE_DWORD_SIZE) == 0) {
669 					if (unlikely(--max_wqe == 0))
670 						break;
671 					dseg = (volatile rte_v128u32_t *)
672 					       tx_mlx5_wqe(txq, txq->wqe_ci +
673 							   ds / 4);
674 				} else {
675 					dseg = (volatile rte_v128u32_t *)
676 						((uintptr_t)wqe +
677 						 (ds * MLX5_WQE_DWORD_SIZE));
678 				}
679 				goto use_dseg;
680 			} else if (!segs_n) {
681 				goto next_pkt;
682 			} else {
683 				/*
684 				 * Further inline the next segment only for
685 				 * non-TSO packets.
686 				 */
687 				if (!tso) {
688 					raw += copy_b;
689 					inline_room -= copy_b;
690 				} else {
691 					inline_room = 0;
692 				}
693 				/* Move to the next segment. */
694 				--segs_n;
695 				buf = buf->next;
696 				assert(buf);
697 				addr = rte_pktmbuf_mtod(buf, uintptr_t);
698 				length = DATA_LEN(buf);
699 #ifdef MLX5_PMD_SOFT_COUNTERS
700 				total_length += length;
701 #endif
702 				(*txq->elts)[++elts_head & elts_m] = buf;
703 				goto pkt_inline;
704 			}
705 		} else {
706 			/*
707 			 * No inline has been done in the packet, only the
708 			 * Ethernet Header as been stored.
709 			 */
710 			dseg = (volatile rte_v128u32_t *)
711 				((uintptr_t)wqe + (3 * MLX5_WQE_DWORD_SIZE));
712 			ds = 3;
713 use_dseg:
714 			/* Add the remaining packet as a simple ds. */
715 			addr_64 = rte_cpu_to_be_64(addr);
716 			*dseg = (rte_v128u32_t){
717 				rte_cpu_to_be_32(length),
718 				mlx5_tx_mb2mr(txq, buf),
719 				addr_64,
720 				addr_64 >> 32,
721 			};
722 			++ds;
723 			if (!segs_n)
724 				goto next_pkt;
725 		}
726 next_seg:
727 		assert(buf);
728 		assert(ds);
729 		assert(wqe);
730 		/*
731 		 * Spill on next WQE when the current one does not have
732 		 * enough room left. Size of WQE must a be a multiple
733 		 * of data segment size.
734 		 */
735 		assert(!(MLX5_WQE_SIZE % MLX5_WQE_DWORD_SIZE));
736 		if (!(ds % (MLX5_WQE_SIZE / MLX5_WQE_DWORD_SIZE))) {
737 			if (unlikely(--max_wqe == 0))
738 				break;
739 			dseg = (volatile rte_v128u32_t *)
740 			       tx_mlx5_wqe(txq, txq->wqe_ci + ds / 4);
741 			rte_prefetch0(tx_mlx5_wqe(txq,
742 						  txq->wqe_ci + ds / 4 + 1));
743 		} else {
744 			++dseg;
745 		}
746 		++ds;
747 		buf = buf->next;
748 		assert(buf);
749 		length = DATA_LEN(buf);
750 #ifdef MLX5_PMD_SOFT_COUNTERS
751 		total_length += length;
752 #endif
753 		/* Store segment information. */
754 		addr_64 = rte_cpu_to_be_64(rte_pktmbuf_mtod(buf, uintptr_t));
755 		*dseg = (rte_v128u32_t){
756 			rte_cpu_to_be_32(length),
757 			mlx5_tx_mb2mr(txq, buf),
758 			addr_64,
759 			addr_64 >> 32,
760 		};
761 		(*txq->elts)[++elts_head & elts_m] = buf;
762 		if (--segs_n)
763 			goto next_seg;
764 next_pkt:
765 		if (ds > MLX5_DSEG_MAX) {
766 			txq->stats.oerrors++;
767 			break;
768 		}
769 		++elts_head;
770 		++pkts;
771 		++i;
772 		j += sg;
773 		/* Initialize known and common part of the WQE structure. */
774 		if (tso) {
775 			wqe->ctrl = (rte_v128u32_t){
776 				rte_cpu_to_be_32((txq->wqe_ci << 8) |
777 						 MLX5_OPCODE_TSO),
778 				rte_cpu_to_be_32(txq->qp_num_8s | ds),
779 				0,
780 				0,
781 			};
782 			wqe->eseg = (rte_v128u32_t){
783 				swp_offsets,
784 				cs_flags | (swp_types << 8) |
785 				(rte_cpu_to_be_16(tso_segsz) << 16),
786 				0,
787 				(ehdr << 16) | rte_cpu_to_be_16(tso_header_sz),
788 			};
789 		} else {
790 			wqe->ctrl = (rte_v128u32_t){
791 				rte_cpu_to_be_32((txq->wqe_ci << 8) |
792 						 MLX5_OPCODE_SEND),
793 				rte_cpu_to_be_32(txq->qp_num_8s | ds),
794 				0,
795 				0,
796 			};
797 			wqe->eseg = (rte_v128u32_t){
798 				swp_offsets,
799 				cs_flags | (swp_types << 8),
800 				0,
801 				(ehdr << 16) | rte_cpu_to_be_16(pkt_inline_sz),
802 			};
803 		}
804 next_wqe:
805 		txq->wqe_ci += (ds + 3) / 4;
806 		/* Save the last successful WQE for completion request */
807 		last_wqe = (volatile struct mlx5_wqe_ctrl *)wqe;
808 #ifdef MLX5_PMD_SOFT_COUNTERS
809 		/* Increment sent bytes counter. */
810 		txq->stats.obytes += total_length;
811 #endif
812 	} while (i < pkts_n);
813 	/* Take a shortcut if nothing must be sent. */
814 	if (unlikely((i + k) == 0))
815 		return 0;
816 	txq->elts_head += (i + j);
817 	/* Check whether completion threshold has been reached. */
818 	comp = txq->elts_comp + i + j + k;
819 	if (comp >= MLX5_TX_COMP_THRESH) {
820 		/* Request completion on last WQE. */
821 		last_wqe->ctrl2 = rte_cpu_to_be_32(8);
822 		/* Save elts_head in unused "immediate" field of WQE. */
823 		last_wqe->ctrl3 = txq->elts_head;
824 		txq->elts_comp = 0;
825 #ifndef NDEBUG
826 		++txq->cq_pi;
827 #endif
828 	} else {
829 		txq->elts_comp = comp;
830 	}
831 #ifdef MLX5_PMD_SOFT_COUNTERS
832 	/* Increment sent packets counter. */
833 	txq->stats.opackets += i;
834 #endif
835 	/* Ring QP doorbell. */
836 	mlx5_tx_dbrec(txq, (volatile struct mlx5_wqe *)last_wqe);
837 	return i;
838 }
839 
840 /**
841  * Open a MPW session.
842  *
843  * @param txq
844  *   Pointer to TX queue structure.
845  * @param mpw
846  *   Pointer to MPW session structure.
847  * @param length
848  *   Packet length.
849  */
850 static inline void
851 mlx5_mpw_new(struct mlx5_txq_data *txq, struct mlx5_mpw *mpw, uint32_t length)
852 {
853 	uint16_t idx = txq->wqe_ci & ((1 << txq->wqe_n) - 1);
854 	volatile struct mlx5_wqe_data_seg (*dseg)[MLX5_MPW_DSEG_MAX] =
855 		(volatile struct mlx5_wqe_data_seg (*)[])
856 		tx_mlx5_wqe(txq, idx + 1);
857 
858 	mpw->state = MLX5_MPW_STATE_OPENED;
859 	mpw->pkts_n = 0;
860 	mpw->len = length;
861 	mpw->total_len = 0;
862 	mpw->wqe = (volatile struct mlx5_wqe *)tx_mlx5_wqe(txq, idx);
863 	mpw->wqe->eseg.mss = rte_cpu_to_be_16(length);
864 	mpw->wqe->eseg.inline_hdr_sz = 0;
865 	mpw->wqe->eseg.rsvd0 = 0;
866 	mpw->wqe->eseg.rsvd1 = 0;
867 	mpw->wqe->eseg.rsvd2 = 0;
868 	mpw->wqe->ctrl[0] = rte_cpu_to_be_32((MLX5_OPC_MOD_MPW << 24) |
869 					     (txq->wqe_ci << 8) |
870 					     MLX5_OPCODE_TSO);
871 	mpw->wqe->ctrl[2] = 0;
872 	mpw->wqe->ctrl[3] = 0;
873 	mpw->data.dseg[0] = (volatile struct mlx5_wqe_data_seg *)
874 		(((uintptr_t)mpw->wqe) + (2 * MLX5_WQE_DWORD_SIZE));
875 	mpw->data.dseg[1] = (volatile struct mlx5_wqe_data_seg *)
876 		(((uintptr_t)mpw->wqe) + (3 * MLX5_WQE_DWORD_SIZE));
877 	mpw->data.dseg[2] = &(*dseg)[0];
878 	mpw->data.dseg[3] = &(*dseg)[1];
879 	mpw->data.dseg[4] = &(*dseg)[2];
880 }
881 
882 /**
883  * Close a MPW session.
884  *
885  * @param txq
886  *   Pointer to TX queue structure.
887  * @param mpw
888  *   Pointer to MPW session structure.
889  */
890 static inline void
891 mlx5_mpw_close(struct mlx5_txq_data *txq, struct mlx5_mpw *mpw)
892 {
893 	unsigned int num = mpw->pkts_n;
894 
895 	/*
896 	 * Store size in multiple of 16 bytes. Control and Ethernet segments
897 	 * count as 2.
898 	 */
899 	mpw->wqe->ctrl[1] = rte_cpu_to_be_32(txq->qp_num_8s | (2 + num));
900 	mpw->state = MLX5_MPW_STATE_CLOSED;
901 	if (num < 3)
902 		++txq->wqe_ci;
903 	else
904 		txq->wqe_ci += 2;
905 	rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci));
906 	rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci + 1));
907 }
908 
909 /**
910  * DPDK callback for TX with MPW support.
911  *
912  * @param dpdk_txq
913  *   Generic pointer to TX queue structure.
914  * @param[in] pkts
915  *   Packets to transmit.
916  * @param pkts_n
917  *   Number of packets in array.
918  *
919  * @return
920  *   Number of packets successfully transmitted (<= pkts_n).
921  */
922 uint16_t
923 mlx5_tx_burst_mpw(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
924 {
925 	struct mlx5_txq_data *txq = (struct mlx5_txq_data *)dpdk_txq;
926 	uint16_t elts_head = txq->elts_head;
927 	const uint16_t elts_n = 1 << txq->elts_n;
928 	const uint16_t elts_m = elts_n - 1;
929 	unsigned int i = 0;
930 	unsigned int j = 0;
931 	uint16_t max_elts;
932 	uint16_t max_wqe;
933 	unsigned int comp;
934 	struct mlx5_mpw mpw = {
935 		.state = MLX5_MPW_STATE_CLOSED,
936 	};
937 
938 	if (unlikely(!pkts_n))
939 		return 0;
940 	/* Prefetch first packet cacheline. */
941 	rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci));
942 	rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci + 1));
943 	/* Start processing. */
944 	mlx5_tx_complete(txq);
945 	max_elts = (elts_n - (elts_head - txq->elts_tail));
946 	/* A CQE slot must always be available. */
947 	assert((1u << txq->cqe_n) - (txq->cq_pi - txq->cq_ci));
948 	max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
949 	if (unlikely(!max_wqe))
950 		return 0;
951 	do {
952 		struct rte_mbuf *buf = *(pkts++);
953 		uint32_t length;
954 		unsigned int segs_n = buf->nb_segs;
955 		uint32_t cs_flags;
956 
957 		/*
958 		 * Make sure there is enough room to store this packet and
959 		 * that one ring entry remains unused.
960 		 */
961 		assert(segs_n);
962 		if (max_elts < segs_n)
963 			break;
964 		/* Do not bother with large packets MPW cannot handle. */
965 		if (segs_n > MLX5_MPW_DSEG_MAX) {
966 			txq->stats.oerrors++;
967 			break;
968 		}
969 		max_elts -= segs_n;
970 		--pkts_n;
971 		cs_flags = txq_ol_cksum_to_cs(buf);
972 		/* Retrieve packet information. */
973 		length = PKT_LEN(buf);
974 		assert(length);
975 		/* Start new session if packet differs. */
976 		if ((mpw.state == MLX5_MPW_STATE_OPENED) &&
977 		    ((mpw.len != length) ||
978 		     (segs_n != 1) ||
979 		     (mpw.wqe->eseg.cs_flags != cs_flags)))
980 			mlx5_mpw_close(txq, &mpw);
981 		if (mpw.state == MLX5_MPW_STATE_CLOSED) {
982 			/*
983 			 * Multi-Packet WQE consumes at most two WQE.
984 			 * mlx5_mpw_new() expects to be able to use such
985 			 * resources.
986 			 */
987 			if (unlikely(max_wqe < 2))
988 				break;
989 			max_wqe -= 2;
990 			mlx5_mpw_new(txq, &mpw, length);
991 			mpw.wqe->eseg.cs_flags = cs_flags;
992 		}
993 		/* Multi-segment packets must be alone in their MPW. */
994 		assert((segs_n == 1) || (mpw.pkts_n == 0));
995 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
996 		length = 0;
997 #endif
998 		do {
999 			volatile struct mlx5_wqe_data_seg *dseg;
1000 			uintptr_t addr;
1001 
1002 			assert(buf);
1003 			(*txq->elts)[elts_head++ & elts_m] = buf;
1004 			dseg = mpw.data.dseg[mpw.pkts_n];
1005 			addr = rte_pktmbuf_mtod(buf, uintptr_t);
1006 			*dseg = (struct mlx5_wqe_data_seg){
1007 				.byte_count = rte_cpu_to_be_32(DATA_LEN(buf)),
1008 				.lkey = mlx5_tx_mb2mr(txq, buf),
1009 				.addr = rte_cpu_to_be_64(addr),
1010 			};
1011 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
1012 			length += DATA_LEN(buf);
1013 #endif
1014 			buf = buf->next;
1015 			++mpw.pkts_n;
1016 			++j;
1017 		} while (--segs_n);
1018 		assert(length == mpw.len);
1019 		if (mpw.pkts_n == MLX5_MPW_DSEG_MAX)
1020 			mlx5_mpw_close(txq, &mpw);
1021 #ifdef MLX5_PMD_SOFT_COUNTERS
1022 		/* Increment sent bytes counter. */
1023 		txq->stats.obytes += length;
1024 #endif
1025 		++i;
1026 	} while (pkts_n);
1027 	/* Take a shortcut if nothing must be sent. */
1028 	if (unlikely(i == 0))
1029 		return 0;
1030 	/* Check whether completion threshold has been reached. */
1031 	/* "j" includes both packets and segments. */
1032 	comp = txq->elts_comp + j;
1033 	if (comp >= MLX5_TX_COMP_THRESH) {
1034 		volatile struct mlx5_wqe *wqe = mpw.wqe;
1035 
1036 		/* Request completion on last WQE. */
1037 		wqe->ctrl[2] = rte_cpu_to_be_32(8);
1038 		/* Save elts_head in unused "immediate" field of WQE. */
1039 		wqe->ctrl[3] = elts_head;
1040 		txq->elts_comp = 0;
1041 #ifndef NDEBUG
1042 		++txq->cq_pi;
1043 #endif
1044 	} else {
1045 		txq->elts_comp = comp;
1046 	}
1047 #ifdef MLX5_PMD_SOFT_COUNTERS
1048 	/* Increment sent packets counter. */
1049 	txq->stats.opackets += i;
1050 #endif
1051 	/* Ring QP doorbell. */
1052 	if (mpw.state == MLX5_MPW_STATE_OPENED)
1053 		mlx5_mpw_close(txq, &mpw);
1054 	mlx5_tx_dbrec(txq, mpw.wqe);
1055 	txq->elts_head = elts_head;
1056 	return i;
1057 }
1058 
1059 /**
1060  * Open a MPW inline session.
1061  *
1062  * @param txq
1063  *   Pointer to TX queue structure.
1064  * @param mpw
1065  *   Pointer to MPW session structure.
1066  * @param length
1067  *   Packet length.
1068  */
1069 static inline void
1070 mlx5_mpw_inline_new(struct mlx5_txq_data *txq, struct mlx5_mpw *mpw,
1071 		    uint32_t length)
1072 {
1073 	uint16_t idx = txq->wqe_ci & ((1 << txq->wqe_n) - 1);
1074 	struct mlx5_wqe_inl_small *inl;
1075 
1076 	mpw->state = MLX5_MPW_INL_STATE_OPENED;
1077 	mpw->pkts_n = 0;
1078 	mpw->len = length;
1079 	mpw->total_len = 0;
1080 	mpw->wqe = (volatile struct mlx5_wqe *)tx_mlx5_wqe(txq, idx);
1081 	mpw->wqe->ctrl[0] = rte_cpu_to_be_32((MLX5_OPC_MOD_MPW << 24) |
1082 					     (txq->wqe_ci << 8) |
1083 					     MLX5_OPCODE_TSO);
1084 	mpw->wqe->ctrl[2] = 0;
1085 	mpw->wqe->ctrl[3] = 0;
1086 	mpw->wqe->eseg.mss = rte_cpu_to_be_16(length);
1087 	mpw->wqe->eseg.inline_hdr_sz = 0;
1088 	mpw->wqe->eseg.cs_flags = 0;
1089 	mpw->wqe->eseg.rsvd0 = 0;
1090 	mpw->wqe->eseg.rsvd1 = 0;
1091 	mpw->wqe->eseg.rsvd2 = 0;
1092 	inl = (struct mlx5_wqe_inl_small *)
1093 		(((uintptr_t)mpw->wqe) + 2 * MLX5_WQE_DWORD_SIZE);
1094 	mpw->data.raw = (uint8_t *)&inl->raw;
1095 }
1096 
1097 /**
1098  * Close a MPW inline session.
1099  *
1100  * @param txq
1101  *   Pointer to TX queue structure.
1102  * @param mpw
1103  *   Pointer to MPW session structure.
1104  */
1105 static inline void
1106 mlx5_mpw_inline_close(struct mlx5_txq_data *txq, struct mlx5_mpw *mpw)
1107 {
1108 	unsigned int size;
1109 	struct mlx5_wqe_inl_small *inl = (struct mlx5_wqe_inl_small *)
1110 		(((uintptr_t)mpw->wqe) + (2 * MLX5_WQE_DWORD_SIZE));
1111 
1112 	size = MLX5_WQE_SIZE - MLX5_MWQE64_INL_DATA + mpw->total_len;
1113 	/*
1114 	 * Store size in multiple of 16 bytes. Control and Ethernet segments
1115 	 * count as 2.
1116 	 */
1117 	mpw->wqe->ctrl[1] = rte_cpu_to_be_32(txq->qp_num_8s |
1118 					     MLX5_WQE_DS(size));
1119 	mpw->state = MLX5_MPW_STATE_CLOSED;
1120 	inl->byte_cnt = rte_cpu_to_be_32(mpw->total_len | MLX5_INLINE_SEG);
1121 	txq->wqe_ci += (size + (MLX5_WQE_SIZE - 1)) / MLX5_WQE_SIZE;
1122 }
1123 
1124 /**
1125  * DPDK callback for TX with MPW inline support.
1126  *
1127  * @param dpdk_txq
1128  *   Generic pointer to TX queue structure.
1129  * @param[in] pkts
1130  *   Packets to transmit.
1131  * @param pkts_n
1132  *   Number of packets in array.
1133  *
1134  * @return
1135  *   Number of packets successfully transmitted (<= pkts_n).
1136  */
1137 uint16_t
1138 mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
1139 			 uint16_t pkts_n)
1140 {
1141 	struct mlx5_txq_data *txq = (struct mlx5_txq_data *)dpdk_txq;
1142 	uint16_t elts_head = txq->elts_head;
1143 	const uint16_t elts_n = 1 << txq->elts_n;
1144 	const uint16_t elts_m = elts_n - 1;
1145 	unsigned int i = 0;
1146 	unsigned int j = 0;
1147 	uint16_t max_elts;
1148 	uint16_t max_wqe;
1149 	unsigned int comp;
1150 	unsigned int inline_room = txq->max_inline * RTE_CACHE_LINE_SIZE;
1151 	struct mlx5_mpw mpw = {
1152 		.state = MLX5_MPW_STATE_CLOSED,
1153 	};
1154 	/*
1155 	 * Compute the maximum number of WQE which can be consumed by inline
1156 	 * code.
1157 	 * - 2 DSEG for:
1158 	 *   - 1 control segment,
1159 	 *   - 1 Ethernet segment,
1160 	 * - N Dseg from the inline request.
1161 	 */
1162 	const unsigned int wqe_inl_n =
1163 		((2 * MLX5_WQE_DWORD_SIZE +
1164 		  txq->max_inline * RTE_CACHE_LINE_SIZE) +
1165 		 RTE_CACHE_LINE_SIZE - 1) / RTE_CACHE_LINE_SIZE;
1166 
1167 	if (unlikely(!pkts_n))
1168 		return 0;
1169 	/* Prefetch first packet cacheline. */
1170 	rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci));
1171 	rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci + 1));
1172 	/* Start processing. */
1173 	mlx5_tx_complete(txq);
1174 	max_elts = (elts_n - (elts_head - txq->elts_tail));
1175 	/* A CQE slot must always be available. */
1176 	assert((1u << txq->cqe_n) - (txq->cq_pi - txq->cq_ci));
1177 	do {
1178 		struct rte_mbuf *buf = *(pkts++);
1179 		uintptr_t addr;
1180 		uint32_t length;
1181 		unsigned int segs_n = buf->nb_segs;
1182 		uint8_t cs_flags;
1183 
1184 		/*
1185 		 * Make sure there is enough room to store this packet and
1186 		 * that one ring entry remains unused.
1187 		 */
1188 		assert(segs_n);
1189 		if (max_elts < segs_n)
1190 			break;
1191 		/* Do not bother with large packets MPW cannot handle. */
1192 		if (segs_n > MLX5_MPW_DSEG_MAX) {
1193 			txq->stats.oerrors++;
1194 			break;
1195 		}
1196 		max_elts -= segs_n;
1197 		--pkts_n;
1198 		/*
1199 		 * Compute max_wqe in case less WQE were consumed in previous
1200 		 * iteration.
1201 		 */
1202 		max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
1203 		cs_flags = txq_ol_cksum_to_cs(buf);
1204 		/* Retrieve packet information. */
1205 		length = PKT_LEN(buf);
1206 		/* Start new session if packet differs. */
1207 		if (mpw.state == MLX5_MPW_STATE_OPENED) {
1208 			if ((mpw.len != length) ||
1209 			    (segs_n != 1) ||
1210 			    (mpw.wqe->eseg.cs_flags != cs_flags))
1211 				mlx5_mpw_close(txq, &mpw);
1212 		} else if (mpw.state == MLX5_MPW_INL_STATE_OPENED) {
1213 			if ((mpw.len != length) ||
1214 			    (segs_n != 1) ||
1215 			    (length > inline_room) ||
1216 			    (mpw.wqe->eseg.cs_flags != cs_flags)) {
1217 				mlx5_mpw_inline_close(txq, &mpw);
1218 				inline_room =
1219 					txq->max_inline * RTE_CACHE_LINE_SIZE;
1220 			}
1221 		}
1222 		if (mpw.state == MLX5_MPW_STATE_CLOSED) {
1223 			if ((segs_n != 1) ||
1224 			    (length > inline_room)) {
1225 				/*
1226 				 * Multi-Packet WQE consumes at most two WQE.
1227 				 * mlx5_mpw_new() expects to be able to use
1228 				 * such resources.
1229 				 */
1230 				if (unlikely(max_wqe < 2))
1231 					break;
1232 				max_wqe -= 2;
1233 				mlx5_mpw_new(txq, &mpw, length);
1234 				mpw.wqe->eseg.cs_flags = cs_flags;
1235 			} else {
1236 				if (unlikely(max_wqe < wqe_inl_n))
1237 					break;
1238 				max_wqe -= wqe_inl_n;
1239 				mlx5_mpw_inline_new(txq, &mpw, length);
1240 				mpw.wqe->eseg.cs_flags = cs_flags;
1241 			}
1242 		}
1243 		/* Multi-segment packets must be alone in their MPW. */
1244 		assert((segs_n == 1) || (mpw.pkts_n == 0));
1245 		if (mpw.state == MLX5_MPW_STATE_OPENED) {
1246 			assert(inline_room ==
1247 			       txq->max_inline * RTE_CACHE_LINE_SIZE);
1248 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
1249 			length = 0;
1250 #endif
1251 			do {
1252 				volatile struct mlx5_wqe_data_seg *dseg;
1253 
1254 				assert(buf);
1255 				(*txq->elts)[elts_head++ & elts_m] = buf;
1256 				dseg = mpw.data.dseg[mpw.pkts_n];
1257 				addr = rte_pktmbuf_mtod(buf, uintptr_t);
1258 				*dseg = (struct mlx5_wqe_data_seg){
1259 					.byte_count =
1260 					       rte_cpu_to_be_32(DATA_LEN(buf)),
1261 					.lkey = mlx5_tx_mb2mr(txq, buf),
1262 					.addr = rte_cpu_to_be_64(addr),
1263 				};
1264 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
1265 				length += DATA_LEN(buf);
1266 #endif
1267 				buf = buf->next;
1268 				++mpw.pkts_n;
1269 				++j;
1270 			} while (--segs_n);
1271 			assert(length == mpw.len);
1272 			if (mpw.pkts_n == MLX5_MPW_DSEG_MAX)
1273 				mlx5_mpw_close(txq, &mpw);
1274 		} else {
1275 			unsigned int max;
1276 
1277 			assert(mpw.state == MLX5_MPW_INL_STATE_OPENED);
1278 			assert(length <= inline_room);
1279 			assert(length == DATA_LEN(buf));
1280 			addr = rte_pktmbuf_mtod(buf, uintptr_t);
1281 			(*txq->elts)[elts_head++ & elts_m] = buf;
1282 			/* Maximum number of bytes before wrapping. */
1283 			max = ((((uintptr_t)(txq->wqes)) +
1284 				(1 << txq->wqe_n) *
1285 				MLX5_WQE_SIZE) -
1286 			       (uintptr_t)mpw.data.raw);
1287 			if (length > max) {
1288 				rte_memcpy((void *)(uintptr_t)mpw.data.raw,
1289 					   (void *)addr,
1290 					   max);
1291 				mpw.data.raw = (volatile void *)txq->wqes;
1292 				rte_memcpy((void *)(uintptr_t)mpw.data.raw,
1293 					   (void *)(addr + max),
1294 					   length - max);
1295 				mpw.data.raw += length - max;
1296 			} else {
1297 				rte_memcpy((void *)(uintptr_t)mpw.data.raw,
1298 					   (void *)addr,
1299 					   length);
1300 
1301 				if (length == max)
1302 					mpw.data.raw =
1303 						(volatile void *)txq->wqes;
1304 				else
1305 					mpw.data.raw += length;
1306 			}
1307 			++mpw.pkts_n;
1308 			mpw.total_len += length;
1309 			++j;
1310 			if (mpw.pkts_n == MLX5_MPW_DSEG_MAX) {
1311 				mlx5_mpw_inline_close(txq, &mpw);
1312 				inline_room =
1313 					txq->max_inline * RTE_CACHE_LINE_SIZE;
1314 			} else {
1315 				inline_room -= length;
1316 			}
1317 		}
1318 #ifdef MLX5_PMD_SOFT_COUNTERS
1319 		/* Increment sent bytes counter. */
1320 		txq->stats.obytes += length;
1321 #endif
1322 		++i;
1323 	} while (pkts_n);
1324 	/* Take a shortcut if nothing must be sent. */
1325 	if (unlikely(i == 0))
1326 		return 0;
1327 	/* Check whether completion threshold has been reached. */
1328 	/* "j" includes both packets and segments. */
1329 	comp = txq->elts_comp + j;
1330 	if (comp >= MLX5_TX_COMP_THRESH) {
1331 		volatile struct mlx5_wqe *wqe = mpw.wqe;
1332 
1333 		/* Request completion on last WQE. */
1334 		wqe->ctrl[2] = rte_cpu_to_be_32(8);
1335 		/* Save elts_head in unused "immediate" field of WQE. */
1336 		wqe->ctrl[3] = elts_head;
1337 		txq->elts_comp = 0;
1338 #ifndef NDEBUG
1339 		++txq->cq_pi;
1340 #endif
1341 	} else {
1342 		txq->elts_comp = comp;
1343 	}
1344 #ifdef MLX5_PMD_SOFT_COUNTERS
1345 	/* Increment sent packets counter. */
1346 	txq->stats.opackets += i;
1347 #endif
1348 	/* Ring QP doorbell. */
1349 	if (mpw.state == MLX5_MPW_INL_STATE_OPENED)
1350 		mlx5_mpw_inline_close(txq, &mpw);
1351 	else if (mpw.state == MLX5_MPW_STATE_OPENED)
1352 		mlx5_mpw_close(txq, &mpw);
1353 	mlx5_tx_dbrec(txq, mpw.wqe);
1354 	txq->elts_head = elts_head;
1355 	return i;
1356 }
1357 
1358 /**
1359  * Open an Enhanced MPW session.
1360  *
1361  * @param txq
1362  *   Pointer to TX queue structure.
1363  * @param mpw
1364  *   Pointer to MPW session structure.
1365  * @param length
1366  *   Packet length.
1367  */
1368 static inline void
1369 mlx5_empw_new(struct mlx5_txq_data *txq, struct mlx5_mpw *mpw, int padding)
1370 {
1371 	uint16_t idx = txq->wqe_ci & ((1 << txq->wqe_n) - 1);
1372 
1373 	mpw->state = MLX5_MPW_ENHANCED_STATE_OPENED;
1374 	mpw->pkts_n = 0;
1375 	mpw->total_len = sizeof(struct mlx5_wqe);
1376 	mpw->wqe = (volatile struct mlx5_wqe *)tx_mlx5_wqe(txq, idx);
1377 	mpw->wqe->ctrl[0] =
1378 		rte_cpu_to_be_32((MLX5_OPC_MOD_ENHANCED_MPSW << 24) |
1379 				 (txq->wqe_ci << 8) |
1380 				 MLX5_OPCODE_ENHANCED_MPSW);
1381 	mpw->wqe->ctrl[2] = 0;
1382 	mpw->wqe->ctrl[3] = 0;
1383 	memset((void *)(uintptr_t)&mpw->wqe->eseg, 0, MLX5_WQE_DWORD_SIZE);
1384 	if (unlikely(padding)) {
1385 		uintptr_t addr = (uintptr_t)(mpw->wqe + 1);
1386 
1387 		/* Pad the first 2 DWORDs with zero-length inline header. */
1388 		*(volatile uint32_t *)addr = rte_cpu_to_be_32(MLX5_INLINE_SEG);
1389 		*(volatile uint32_t *)(addr + MLX5_WQE_DWORD_SIZE) =
1390 			rte_cpu_to_be_32(MLX5_INLINE_SEG);
1391 		mpw->total_len += 2 * MLX5_WQE_DWORD_SIZE;
1392 		/* Start from the next WQEBB. */
1393 		mpw->data.raw = (volatile void *)(tx_mlx5_wqe(txq, idx + 1));
1394 	} else {
1395 		mpw->data.raw = (volatile void *)(mpw->wqe + 1);
1396 	}
1397 }
1398 
1399 /**
1400  * Close an Enhanced MPW session.
1401  *
1402  * @param txq
1403  *   Pointer to TX queue structure.
1404  * @param mpw
1405  *   Pointer to MPW session structure.
1406  *
1407  * @return
1408  *   Number of consumed WQEs.
1409  */
1410 static inline uint16_t
1411 mlx5_empw_close(struct mlx5_txq_data *txq, struct mlx5_mpw *mpw)
1412 {
1413 	uint16_t ret;
1414 
1415 	/* Store size in multiple of 16 bytes. Control and Ethernet segments
1416 	 * count as 2.
1417 	 */
1418 	mpw->wqe->ctrl[1] = rte_cpu_to_be_32(txq->qp_num_8s |
1419 					     MLX5_WQE_DS(mpw->total_len));
1420 	mpw->state = MLX5_MPW_STATE_CLOSED;
1421 	ret = (mpw->total_len + (MLX5_WQE_SIZE - 1)) / MLX5_WQE_SIZE;
1422 	txq->wqe_ci += ret;
1423 	return ret;
1424 }
1425 
1426 /**
1427  * TX with Enhanced MPW support.
1428  *
1429  * @param txq
1430  *   Pointer to TX queue structure.
1431  * @param[in] pkts
1432  *   Packets to transmit.
1433  * @param pkts_n
1434  *   Number of packets in array.
1435  *
1436  * @return
1437  *   Number of packets successfully transmitted (<= pkts_n).
1438  */
1439 static inline uint16_t
1440 txq_burst_empw(struct mlx5_txq_data *txq, struct rte_mbuf **pkts,
1441 	       uint16_t pkts_n)
1442 {
1443 	uint16_t elts_head = txq->elts_head;
1444 	const uint16_t elts_n = 1 << txq->elts_n;
1445 	const uint16_t elts_m = elts_n - 1;
1446 	unsigned int i = 0;
1447 	unsigned int j = 0;
1448 	uint16_t max_elts;
1449 	uint16_t max_wqe;
1450 	unsigned int max_inline = txq->max_inline * RTE_CACHE_LINE_SIZE;
1451 	unsigned int mpw_room = 0;
1452 	unsigned int inl_pad = 0;
1453 	uint32_t inl_hdr;
1454 	uint64_t addr_64;
1455 	struct mlx5_mpw mpw = {
1456 		.state = MLX5_MPW_STATE_CLOSED,
1457 	};
1458 
1459 	if (unlikely(!pkts_n))
1460 		return 0;
1461 	/* Start processing. */
1462 	mlx5_tx_complete(txq);
1463 	max_elts = (elts_n - (elts_head - txq->elts_tail));
1464 	/* A CQE slot must always be available. */
1465 	assert((1u << txq->cqe_n) - (txq->cq_pi - txq->cq_ci));
1466 	max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
1467 	if (unlikely(!max_wqe))
1468 		return 0;
1469 	do {
1470 		struct rte_mbuf *buf = *(pkts++);
1471 		uintptr_t addr;
1472 		unsigned int do_inline = 0; /* Whether inline is possible. */
1473 		uint32_t length;
1474 		uint8_t cs_flags;
1475 
1476 		/* Multi-segmented packet is handled in slow-path outside. */
1477 		assert(NB_SEGS(buf) == 1);
1478 		/* Make sure there is enough room to store this packet. */
1479 		if (max_elts - j == 0)
1480 			break;
1481 		cs_flags = txq_ol_cksum_to_cs(buf);
1482 		/* Retrieve packet information. */
1483 		length = PKT_LEN(buf);
1484 		/* Start new session if:
1485 		 * - multi-segment packet
1486 		 * - no space left even for a dseg
1487 		 * - next packet can be inlined with a new WQE
1488 		 * - cs_flag differs
1489 		 */
1490 		if (mpw.state == MLX5_MPW_ENHANCED_STATE_OPENED) {
1491 			if ((inl_pad + sizeof(struct mlx5_wqe_data_seg) >
1492 			     mpw_room) ||
1493 			    (length <= txq->inline_max_packet_sz &&
1494 			     inl_pad + sizeof(inl_hdr) + length >
1495 			     mpw_room) ||
1496 			    (mpw.wqe->eseg.cs_flags != cs_flags))
1497 				max_wqe -= mlx5_empw_close(txq, &mpw);
1498 		}
1499 		if (unlikely(mpw.state == MLX5_MPW_STATE_CLOSED)) {
1500 			/* In Enhanced MPW, inline as much as the budget is
1501 			 * allowed. The remaining space is to be filled with
1502 			 * dsegs. If the title WQEBB isn't padded, it will have
1503 			 * 2 dsegs there.
1504 			 */
1505 			mpw_room = RTE_MIN(MLX5_WQE_SIZE_MAX,
1506 					   (max_inline ? max_inline :
1507 					    pkts_n * MLX5_WQE_DWORD_SIZE) +
1508 					   MLX5_WQE_SIZE);
1509 			if (unlikely(max_wqe * MLX5_WQE_SIZE < mpw_room))
1510 				break;
1511 			/* Don't pad the title WQEBB to not waste WQ. */
1512 			mlx5_empw_new(txq, &mpw, 0);
1513 			mpw_room -= mpw.total_len;
1514 			inl_pad = 0;
1515 			do_inline = length <= txq->inline_max_packet_sz &&
1516 				    sizeof(inl_hdr) + length <= mpw_room &&
1517 				    !txq->mpw_hdr_dseg;
1518 			mpw.wqe->eseg.cs_flags = cs_flags;
1519 		} else {
1520 			/* Evaluate whether the next packet can be inlined.
1521 			 * Inlininig is possible when:
1522 			 * - length is less than configured value
1523 			 * - length fits for remaining space
1524 			 * - not required to fill the title WQEBB with dsegs
1525 			 */
1526 			do_inline =
1527 				length <= txq->inline_max_packet_sz &&
1528 				inl_pad + sizeof(inl_hdr) + length <=
1529 				 mpw_room &&
1530 				(!txq->mpw_hdr_dseg ||
1531 				 mpw.total_len >= MLX5_WQE_SIZE);
1532 		}
1533 		if (max_inline && do_inline) {
1534 			/* Inline packet into WQE. */
1535 			unsigned int max;
1536 
1537 			assert(mpw.state == MLX5_MPW_ENHANCED_STATE_OPENED);
1538 			assert(length == DATA_LEN(buf));
1539 			inl_hdr = rte_cpu_to_be_32(length | MLX5_INLINE_SEG);
1540 			addr = rte_pktmbuf_mtod(buf, uintptr_t);
1541 			mpw.data.raw = (volatile void *)
1542 				((uintptr_t)mpw.data.raw + inl_pad);
1543 			max = tx_mlx5_wq_tailroom(txq,
1544 					(void *)(uintptr_t)mpw.data.raw);
1545 			/* Copy inline header. */
1546 			mpw.data.raw = (volatile void *)
1547 				mlx5_copy_to_wq(
1548 					  (void *)(uintptr_t)mpw.data.raw,
1549 					  &inl_hdr,
1550 					  sizeof(inl_hdr),
1551 					  (void *)(uintptr_t)txq->wqes,
1552 					  max);
1553 			max = tx_mlx5_wq_tailroom(txq,
1554 					(void *)(uintptr_t)mpw.data.raw);
1555 			/* Copy packet data. */
1556 			mpw.data.raw = (volatile void *)
1557 				mlx5_copy_to_wq(
1558 					  (void *)(uintptr_t)mpw.data.raw,
1559 					  (void *)addr,
1560 					  length,
1561 					  (void *)(uintptr_t)txq->wqes,
1562 					  max);
1563 			++mpw.pkts_n;
1564 			mpw.total_len += (inl_pad + sizeof(inl_hdr) + length);
1565 			/* No need to get completion as the entire packet is
1566 			 * copied to WQ. Free the buf right away.
1567 			 */
1568 			rte_pktmbuf_free_seg(buf);
1569 			mpw_room -= (inl_pad + sizeof(inl_hdr) + length);
1570 			/* Add pad in the next packet if any. */
1571 			inl_pad = (((uintptr_t)mpw.data.raw +
1572 					(MLX5_WQE_DWORD_SIZE - 1)) &
1573 					~(MLX5_WQE_DWORD_SIZE - 1)) -
1574 				  (uintptr_t)mpw.data.raw;
1575 		} else {
1576 			/* No inline. Load a dseg of packet pointer. */
1577 			volatile rte_v128u32_t *dseg;
1578 
1579 			assert(mpw.state == MLX5_MPW_ENHANCED_STATE_OPENED);
1580 			assert((inl_pad + sizeof(*dseg)) <= mpw_room);
1581 			assert(length == DATA_LEN(buf));
1582 			if (!tx_mlx5_wq_tailroom(txq,
1583 					(void *)((uintptr_t)mpw.data.raw
1584 						+ inl_pad)))
1585 				dseg = (volatile void *)txq->wqes;
1586 			else
1587 				dseg = (volatile void *)
1588 					((uintptr_t)mpw.data.raw +
1589 					 inl_pad);
1590 			(*txq->elts)[elts_head++ & elts_m] = buf;
1591 			addr_64 = rte_cpu_to_be_64(rte_pktmbuf_mtod(buf,
1592 								    uintptr_t));
1593 			*dseg = (rte_v128u32_t) {
1594 				rte_cpu_to_be_32(length),
1595 				mlx5_tx_mb2mr(txq, buf),
1596 				addr_64,
1597 				addr_64 >> 32,
1598 			};
1599 			mpw.data.raw = (volatile void *)(dseg + 1);
1600 			mpw.total_len += (inl_pad + sizeof(*dseg));
1601 			++j;
1602 			++mpw.pkts_n;
1603 			mpw_room -= (inl_pad + sizeof(*dseg));
1604 			inl_pad = 0;
1605 		}
1606 #ifdef MLX5_PMD_SOFT_COUNTERS
1607 		/* Increment sent bytes counter. */
1608 		txq->stats.obytes += length;
1609 #endif
1610 		++i;
1611 	} while (i < pkts_n);
1612 	/* Take a shortcut if nothing must be sent. */
1613 	if (unlikely(i == 0))
1614 		return 0;
1615 	/* Check whether completion threshold has been reached. */
1616 	if (txq->elts_comp + j >= MLX5_TX_COMP_THRESH ||
1617 			(uint16_t)(txq->wqe_ci - txq->mpw_comp) >=
1618 			 (1 << txq->wqe_n) / MLX5_TX_COMP_THRESH_INLINE_DIV) {
1619 		volatile struct mlx5_wqe *wqe = mpw.wqe;
1620 
1621 		/* Request completion on last WQE. */
1622 		wqe->ctrl[2] = rte_cpu_to_be_32(8);
1623 		/* Save elts_head in unused "immediate" field of WQE. */
1624 		wqe->ctrl[3] = elts_head;
1625 		txq->elts_comp = 0;
1626 		txq->mpw_comp = txq->wqe_ci;
1627 #ifndef NDEBUG
1628 		++txq->cq_pi;
1629 #endif
1630 	} else {
1631 		txq->elts_comp += j;
1632 	}
1633 #ifdef MLX5_PMD_SOFT_COUNTERS
1634 	/* Increment sent packets counter. */
1635 	txq->stats.opackets += i;
1636 #endif
1637 	if (mpw.state == MLX5_MPW_ENHANCED_STATE_OPENED)
1638 		mlx5_empw_close(txq, &mpw);
1639 	/* Ring QP doorbell. */
1640 	mlx5_tx_dbrec(txq, mpw.wqe);
1641 	txq->elts_head = elts_head;
1642 	return i;
1643 }
1644 
1645 /**
1646  * DPDK callback for TX with Enhanced MPW support.
1647  *
1648  * @param dpdk_txq
1649  *   Generic pointer to TX queue structure.
1650  * @param[in] pkts
1651  *   Packets to transmit.
1652  * @param pkts_n
1653  *   Number of packets in array.
1654  *
1655  * @return
1656  *   Number of packets successfully transmitted (<= pkts_n).
1657  */
1658 uint16_t
1659 mlx5_tx_burst_empw(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
1660 {
1661 	struct mlx5_txq_data *txq = (struct mlx5_txq_data *)dpdk_txq;
1662 	uint16_t nb_tx = 0;
1663 
1664 	while (pkts_n > nb_tx) {
1665 		uint16_t n;
1666 		uint16_t ret;
1667 
1668 		n = txq_count_contig_multi_seg(&pkts[nb_tx], pkts_n - nb_tx);
1669 		if (n) {
1670 			ret = mlx5_tx_burst(dpdk_txq, &pkts[nb_tx], n);
1671 			if (!ret)
1672 				break;
1673 			nb_tx += ret;
1674 		}
1675 		n = txq_count_contig_single_seg(&pkts[nb_tx], pkts_n - nb_tx);
1676 		if (n) {
1677 			ret = txq_burst_empw(txq, &pkts[nb_tx], n);
1678 			if (!ret)
1679 				break;
1680 			nb_tx += ret;
1681 		}
1682 	}
1683 	return nb_tx;
1684 }
1685 
1686 /**
1687  * Translate RX completion flags to packet type.
1688  *
1689  * @param[in] rxq
1690  *   Pointer to RX queue structure.
1691  * @param[in] cqe
1692  *   Pointer to CQE.
1693  *
1694  * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
1695  *
1696  * @return
1697  *   Packet type for struct rte_mbuf.
1698  */
1699 static inline uint32_t
1700 rxq_cq_to_pkt_type(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe)
1701 {
1702 	uint8_t idx;
1703 	uint8_t pinfo = cqe->pkt_info;
1704 	uint16_t ptype = cqe->hdr_type_etc;
1705 
1706 	/*
1707 	 * The index to the array should have:
1708 	 * bit[1:0] = l3_hdr_type
1709 	 * bit[4:2] = l4_hdr_type
1710 	 * bit[5] = ip_frag
1711 	 * bit[6] = tunneled
1712 	 * bit[7] = outer_l3_type
1713 	 */
1714 	idx = ((pinfo & 0x3) << 6) | ((ptype & 0xfc00) >> 10);
1715 	return mlx5_ptype_table[idx] | rxq->tunnel * !!(idx & (1 << 6));
1716 }
1717 
1718 /**
1719  * Get size of the next packet for a given CQE. For compressed CQEs, the
1720  * consumer index is updated only once all packets of the current one have
1721  * been processed.
1722  *
1723  * @param rxq
1724  *   Pointer to RX queue.
1725  * @param cqe
1726  *   CQE to process.
1727  * @param[out] mcqe
1728  *   Store pointer to mini-CQE if compressed. Otherwise, the pointer is not
1729  *   written.
1730  *
1731  * @return
1732  *   Packet size in bytes (0 if there is none), -1 in case of completion
1733  *   with error.
1734  */
1735 static inline int
1736 mlx5_rx_poll_len(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
1737 		 uint16_t cqe_cnt, volatile struct mlx5_mini_cqe8 **mcqe)
1738 {
1739 	struct rxq_zip *zip = &rxq->zip;
1740 	uint16_t cqe_n = cqe_cnt + 1;
1741 	int len = 0;
1742 	uint16_t idx, end;
1743 
1744 	/* Process compressed data in the CQE and mini arrays. */
1745 	if (zip->ai) {
1746 		volatile struct mlx5_mini_cqe8 (*mc)[8] =
1747 			(volatile struct mlx5_mini_cqe8 (*)[8])
1748 			(uintptr_t)(&(*rxq->cqes)[zip->ca & cqe_cnt].pkt_info);
1749 
1750 		len = rte_be_to_cpu_32((*mc)[zip->ai & 7].byte_cnt);
1751 		*mcqe = &(*mc)[zip->ai & 7];
1752 		if ((++zip->ai & 7) == 0) {
1753 			/* Invalidate consumed CQEs */
1754 			idx = zip->ca;
1755 			end = zip->na;
1756 			while (idx != end) {
1757 				(*rxq->cqes)[idx & cqe_cnt].op_own =
1758 					MLX5_CQE_INVALIDATE;
1759 				++idx;
1760 			}
1761 			/*
1762 			 * Increment consumer index to skip the number of
1763 			 * CQEs consumed. Hardware leaves holes in the CQ
1764 			 * ring for software use.
1765 			 */
1766 			zip->ca = zip->na;
1767 			zip->na += 8;
1768 		}
1769 		if (unlikely(rxq->zip.ai == rxq->zip.cqe_cnt)) {
1770 			/* Invalidate the rest */
1771 			idx = zip->ca;
1772 			end = zip->cq_ci;
1773 
1774 			while (idx != end) {
1775 				(*rxq->cqes)[idx & cqe_cnt].op_own =
1776 					MLX5_CQE_INVALIDATE;
1777 				++idx;
1778 			}
1779 			rxq->cq_ci = zip->cq_ci;
1780 			zip->ai = 0;
1781 		}
1782 	/* No compressed data, get next CQE and verify if it is compressed. */
1783 	} else {
1784 		int ret;
1785 		int8_t op_own;
1786 
1787 		ret = check_cqe(cqe, cqe_n, rxq->cq_ci);
1788 		if (unlikely(ret == 1))
1789 			return 0;
1790 		++rxq->cq_ci;
1791 		op_own = cqe->op_own;
1792 		rte_cio_rmb();
1793 		if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED) {
1794 			volatile struct mlx5_mini_cqe8 (*mc)[8] =
1795 				(volatile struct mlx5_mini_cqe8 (*)[8])
1796 				(uintptr_t)(&(*rxq->cqes)[rxq->cq_ci &
1797 							  cqe_cnt].pkt_info);
1798 
1799 			/* Fix endianness. */
1800 			zip->cqe_cnt = rte_be_to_cpu_32(cqe->byte_cnt);
1801 			/*
1802 			 * Current mini array position is the one returned by
1803 			 * check_cqe64().
1804 			 *
1805 			 * If completion comprises several mini arrays, as a
1806 			 * special case the second one is located 7 CQEs after
1807 			 * the initial CQE instead of 8 for subsequent ones.
1808 			 */
1809 			zip->ca = rxq->cq_ci;
1810 			zip->na = zip->ca + 7;
1811 			/* Compute the next non compressed CQE. */
1812 			--rxq->cq_ci;
1813 			zip->cq_ci = rxq->cq_ci + zip->cqe_cnt;
1814 			/* Get packet size to return. */
1815 			len = rte_be_to_cpu_32((*mc)[0].byte_cnt);
1816 			*mcqe = &(*mc)[0];
1817 			zip->ai = 1;
1818 			/* Prefetch all the entries to be invalidated */
1819 			idx = zip->ca;
1820 			end = zip->cq_ci;
1821 			while (idx != end) {
1822 				rte_prefetch0(&(*rxq->cqes)[(idx) & cqe_cnt]);
1823 				++idx;
1824 			}
1825 		} else {
1826 			len = rte_be_to_cpu_32(cqe->byte_cnt);
1827 		}
1828 		/* Error while receiving packet. */
1829 		if (unlikely(MLX5_CQE_OPCODE(op_own) == MLX5_CQE_RESP_ERR))
1830 			return -1;
1831 	}
1832 	return len;
1833 }
1834 
1835 /**
1836  * Translate RX completion flags to offload flags.
1837  *
1838  * @param[in] cqe
1839  *   Pointer to CQE.
1840  *
1841  * @return
1842  *   Offload flags (ol_flags) for struct rte_mbuf.
1843  */
1844 static inline uint32_t
1845 rxq_cq_to_ol_flags(volatile struct mlx5_cqe *cqe)
1846 {
1847 	uint32_t ol_flags = 0;
1848 	uint16_t flags = rte_be_to_cpu_16(cqe->hdr_type_etc);
1849 
1850 	ol_flags =
1851 		TRANSPOSE(flags,
1852 			  MLX5_CQE_RX_L3_HDR_VALID,
1853 			  PKT_RX_IP_CKSUM_GOOD) |
1854 		TRANSPOSE(flags,
1855 			  MLX5_CQE_RX_L4_HDR_VALID,
1856 			  PKT_RX_L4_CKSUM_GOOD);
1857 	return ol_flags;
1858 }
1859 
1860 /**
1861  * Fill in mbuf fields from RX completion flags.
1862  * Note that pkt->ol_flags should be initialized outside of this function.
1863  *
1864  * @param rxq
1865  *   Pointer to RX queue.
1866  * @param pkt
1867  *   mbuf to fill.
1868  * @param cqe
1869  *   CQE to process.
1870  * @param rss_hash_res
1871  *   Packet RSS Hash result.
1872  */
1873 static inline void
1874 rxq_cq_to_mbuf(struct mlx5_rxq_data *rxq, struct rte_mbuf *pkt,
1875 	       volatile struct mlx5_cqe *cqe, uint32_t rss_hash_res)
1876 {
1877 	/* Update packet information. */
1878 	pkt->packet_type = rxq_cq_to_pkt_type(rxq, cqe);
1879 	if (rss_hash_res && rxq->rss_hash) {
1880 		pkt->hash.rss = rss_hash_res;
1881 		pkt->ol_flags |= PKT_RX_RSS_HASH;
1882 	}
1883 	if (rxq->mark && MLX5_FLOW_MARK_IS_VALID(cqe->sop_drop_qpn)) {
1884 		pkt->ol_flags |= PKT_RX_FDIR;
1885 		if (cqe->sop_drop_qpn !=
1886 		    rte_cpu_to_be_32(MLX5_FLOW_MARK_DEFAULT)) {
1887 			uint32_t mark = cqe->sop_drop_qpn;
1888 
1889 			pkt->ol_flags |= PKT_RX_FDIR_ID;
1890 			pkt->hash.fdir.hi = mlx5_flow_mark_get(mark);
1891 		}
1892 	}
1893 	if (rxq->csum)
1894 		pkt->ol_flags |= rxq_cq_to_ol_flags(cqe);
1895 	if (rxq->vlan_strip &&
1896 	    (cqe->hdr_type_etc & rte_cpu_to_be_16(MLX5_CQE_VLAN_STRIPPED))) {
1897 		pkt->ol_flags |= PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
1898 		pkt->vlan_tci = rte_be_to_cpu_16(cqe->vlan_info);
1899 	}
1900 	if (rxq->hw_timestamp) {
1901 		pkt->timestamp = rte_be_to_cpu_64(cqe->timestamp);
1902 		pkt->ol_flags |= PKT_RX_TIMESTAMP;
1903 	}
1904 }
1905 
1906 /**
1907  * DPDK callback for RX.
1908  *
1909  * @param dpdk_rxq
1910  *   Generic pointer to RX queue structure.
1911  * @param[out] pkts
1912  *   Array to store received packets.
1913  * @param pkts_n
1914  *   Maximum number of packets in array.
1915  *
1916  * @return
1917  *   Number of packets successfully received (<= pkts_n).
1918  */
1919 uint16_t
1920 mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1921 {
1922 	struct mlx5_rxq_data *rxq = dpdk_rxq;
1923 	const unsigned int wqe_cnt = (1 << rxq->elts_n) - 1;
1924 	const unsigned int cqe_cnt = (1 << rxq->cqe_n) - 1;
1925 	const unsigned int sges_n = rxq->sges_n;
1926 	struct rte_mbuf *pkt = NULL;
1927 	struct rte_mbuf *seg = NULL;
1928 	volatile struct mlx5_cqe *cqe =
1929 		&(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1930 	unsigned int i = 0;
1931 	unsigned int rq_ci = rxq->rq_ci << sges_n;
1932 	int len = 0; /* keep its value across iterations. */
1933 
1934 	while (pkts_n) {
1935 		unsigned int idx = rq_ci & wqe_cnt;
1936 		volatile struct mlx5_wqe_data_seg *wqe =
1937 			&((volatile struct mlx5_wqe_data_seg *)rxq->wqes)[idx];
1938 		struct rte_mbuf *rep = (*rxq->elts)[idx];
1939 		volatile struct mlx5_mini_cqe8 *mcqe = NULL;
1940 		uint32_t rss_hash_res;
1941 
1942 		if (pkt)
1943 			NEXT(seg) = rep;
1944 		seg = rep;
1945 		rte_prefetch0(seg);
1946 		rte_prefetch0(cqe);
1947 		rte_prefetch0(wqe);
1948 		rep = rte_mbuf_raw_alloc(rxq->mp);
1949 		if (unlikely(rep == NULL)) {
1950 			++rxq->stats.rx_nombuf;
1951 			if (!pkt) {
1952 				/*
1953 				 * no buffers before we even started,
1954 				 * bail out silently.
1955 				 */
1956 				break;
1957 			}
1958 			while (pkt != seg) {
1959 				assert(pkt != (*rxq->elts)[idx]);
1960 				rep = NEXT(pkt);
1961 				NEXT(pkt) = NULL;
1962 				NB_SEGS(pkt) = 1;
1963 				rte_mbuf_raw_free(pkt);
1964 				pkt = rep;
1965 			}
1966 			break;
1967 		}
1968 		if (!pkt) {
1969 			cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1970 			len = mlx5_rx_poll_len(rxq, cqe, cqe_cnt, &mcqe);
1971 			if (!len) {
1972 				rte_mbuf_raw_free(rep);
1973 				break;
1974 			}
1975 			if (unlikely(len == -1)) {
1976 				/* RX error, packet is likely too large. */
1977 				rte_mbuf_raw_free(rep);
1978 				++rxq->stats.idropped;
1979 				goto skip;
1980 			}
1981 			pkt = seg;
1982 			assert(len >= (rxq->crc_present << 2));
1983 			pkt->ol_flags = 0;
1984 			/* If compressed, take hash result from mini-CQE. */
1985 			rss_hash_res = rte_be_to_cpu_32(mcqe == NULL ?
1986 							cqe->rx_hash_res :
1987 							mcqe->rx_hash_result);
1988 			rxq_cq_to_mbuf(rxq, pkt, cqe, rss_hash_res);
1989 			if (rxq->crc_present)
1990 				len -= ETHER_CRC_LEN;
1991 			PKT_LEN(pkt) = len;
1992 		}
1993 		DATA_LEN(rep) = DATA_LEN(seg);
1994 		PKT_LEN(rep) = PKT_LEN(seg);
1995 		SET_DATA_OFF(rep, DATA_OFF(seg));
1996 		PORT(rep) = PORT(seg);
1997 		(*rxq->elts)[idx] = rep;
1998 		/*
1999 		 * Fill NIC descriptor with the new buffer.  The lkey and size
2000 		 * of the buffers are already known, only the buffer address
2001 		 * changes.
2002 		 */
2003 		wqe->addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(rep, uintptr_t));
2004 		/* If there's only one MR, no need to replace LKey in WQE. */
2005 		if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh) > 1))
2006 			wqe->lkey = mlx5_rx_mb2mr(rxq, rep);
2007 		if (len > DATA_LEN(seg)) {
2008 			len -= DATA_LEN(seg);
2009 			++NB_SEGS(pkt);
2010 			++rq_ci;
2011 			continue;
2012 		}
2013 		DATA_LEN(seg) = len;
2014 #ifdef MLX5_PMD_SOFT_COUNTERS
2015 		/* Increment bytes counter. */
2016 		rxq->stats.ibytes += PKT_LEN(pkt);
2017 #endif
2018 		/* Return packet. */
2019 		*(pkts++) = pkt;
2020 		pkt = NULL;
2021 		--pkts_n;
2022 		++i;
2023 skip:
2024 		/* Align consumer index to the next stride. */
2025 		rq_ci >>= sges_n;
2026 		++rq_ci;
2027 		rq_ci <<= sges_n;
2028 	}
2029 	if (unlikely((i == 0) && ((rq_ci >> sges_n) == rxq->rq_ci)))
2030 		return 0;
2031 	/* Update the consumer index. */
2032 	rxq->rq_ci = rq_ci >> sges_n;
2033 	rte_cio_wmb();
2034 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
2035 	rte_cio_wmb();
2036 	*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
2037 #ifdef MLX5_PMD_SOFT_COUNTERS
2038 	/* Increment packets counter. */
2039 	rxq->stats.ipackets += i;
2040 #endif
2041 	return i;
2042 }
2043 
2044 void
2045 mlx5_mprq_buf_free_cb(void *addr __rte_unused, void *opaque)
2046 {
2047 	struct mlx5_mprq_buf *buf = opaque;
2048 
2049 	if (rte_atomic16_read(&buf->refcnt) == 1) {
2050 		rte_mempool_put(buf->mp, buf);
2051 	} else if (rte_atomic16_add_return(&buf->refcnt, -1) == 0) {
2052 		rte_atomic16_set(&buf->refcnt, 1);
2053 		rte_mempool_put(buf->mp, buf);
2054 	}
2055 }
2056 
2057 void
2058 mlx5_mprq_buf_free(struct mlx5_mprq_buf *buf)
2059 {
2060 	mlx5_mprq_buf_free_cb(NULL, buf);
2061 }
2062 
2063 static inline void
2064 mprq_buf_replace(struct mlx5_rxq_data *rxq, uint16_t rq_idx)
2065 {
2066 	struct mlx5_mprq_buf *rep = rxq->mprq_repl;
2067 	volatile struct mlx5_wqe_data_seg *wqe =
2068 		&((volatile struct mlx5_wqe_mprq *)rxq->wqes)[rq_idx].dseg;
2069 	void *addr;
2070 
2071 	assert(rep != NULL);
2072 	/* Replace MPRQ buf. */
2073 	(*rxq->mprq_bufs)[rq_idx] = rep;
2074 	/* Replace WQE. */
2075 	addr = mlx5_mprq_buf_addr(rep);
2076 	wqe->addr = rte_cpu_to_be_64((uintptr_t)addr);
2077 	/* If there's only one MR, no need to replace LKey in WQE. */
2078 	if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh) > 1))
2079 		wqe->lkey = mlx5_rx_addr2mr(rxq, (uintptr_t)addr);
2080 	/* Stash a mbuf for next replacement. */
2081 	if (likely(!rte_mempool_get(rxq->mprq_mp, (void **)&rep)))
2082 		rxq->mprq_repl = rep;
2083 	else
2084 		rxq->mprq_repl = NULL;
2085 }
2086 
2087 /**
2088  * DPDK callback for RX with Multi-Packet RQ support.
2089  *
2090  * @param dpdk_rxq
2091  *   Generic pointer to RX queue structure.
2092  * @param[out] pkts
2093  *   Array to store received packets.
2094  * @param pkts_n
2095  *   Maximum number of packets in array.
2096  *
2097  * @return
2098  *   Number of packets successfully received (<= pkts_n).
2099  */
2100 uint16_t
2101 mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
2102 {
2103 	struct mlx5_rxq_data *rxq = dpdk_rxq;
2104 	const unsigned int strd_n = 1 << rxq->strd_num_n;
2105 	const unsigned int strd_sz = 1 << rxq->strd_sz_n;
2106 	const unsigned int strd_shift =
2107 		MLX5_MPRQ_STRIDE_SHIFT_BYTE * rxq->strd_shift_en;
2108 	const unsigned int cq_mask = (1 << rxq->cqe_n) - 1;
2109 	const unsigned int wq_mask = (1 << rxq->elts_n) - 1;
2110 	volatile struct mlx5_cqe *cqe = &(*rxq->cqes)[rxq->cq_ci & cq_mask];
2111 	unsigned int i = 0;
2112 	uint16_t rq_ci = rxq->rq_ci;
2113 	uint16_t consumed_strd = rxq->consumed_strd;
2114 	struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[rq_ci & wq_mask];
2115 
2116 	while (i < pkts_n) {
2117 		struct rte_mbuf *pkt;
2118 		void *addr;
2119 		int ret;
2120 		unsigned int len;
2121 		uint16_t strd_cnt;
2122 		uint16_t strd_idx;
2123 		uint32_t offset;
2124 		uint32_t byte_cnt;
2125 		volatile struct mlx5_mini_cqe8 *mcqe = NULL;
2126 		uint32_t rss_hash_res = 0;
2127 
2128 		if (consumed_strd == strd_n) {
2129 			/* Replace WQE only if the buffer is still in use. */
2130 			if (rte_atomic16_read(&buf->refcnt) > 1) {
2131 				mprq_buf_replace(rxq, rq_ci & wq_mask);
2132 				/* Release the old buffer. */
2133 				mlx5_mprq_buf_free(buf);
2134 			} else if (unlikely(rxq->mprq_repl == NULL)) {
2135 				struct mlx5_mprq_buf *rep;
2136 
2137 				/*
2138 				 * Currently, the MPRQ mempool is out of buffer
2139 				 * and doing memcpy regardless of the size of Rx
2140 				 * packet. Retry allocation to get back to
2141 				 * normal.
2142 				 */
2143 				if (!rte_mempool_get(rxq->mprq_mp,
2144 						     (void **)&rep))
2145 					rxq->mprq_repl = rep;
2146 			}
2147 			/* Advance to the next WQE. */
2148 			consumed_strd = 0;
2149 			++rq_ci;
2150 			buf = (*rxq->mprq_bufs)[rq_ci & wq_mask];
2151 		}
2152 		cqe = &(*rxq->cqes)[rxq->cq_ci & cq_mask];
2153 		ret = mlx5_rx_poll_len(rxq, cqe, cq_mask, &mcqe);
2154 		if (!ret)
2155 			break;
2156 		if (unlikely(ret == -1)) {
2157 			/* RX error, packet is likely too large. */
2158 			++rxq->stats.idropped;
2159 			continue;
2160 		}
2161 		byte_cnt = ret;
2162 		strd_cnt = (byte_cnt & MLX5_MPRQ_STRIDE_NUM_MASK) >>
2163 			   MLX5_MPRQ_STRIDE_NUM_SHIFT;
2164 		assert(strd_cnt);
2165 		consumed_strd += strd_cnt;
2166 		if (byte_cnt & MLX5_MPRQ_FILLER_MASK)
2167 			continue;
2168 		if (mcqe == NULL) {
2169 			rss_hash_res = rte_be_to_cpu_32(cqe->rx_hash_res);
2170 			strd_idx = rte_be_to_cpu_16(cqe->wqe_counter);
2171 		} else {
2172 			/* mini-CQE for MPRQ doesn't have hash result. */
2173 			strd_idx = rte_be_to_cpu_16(mcqe->stride_idx);
2174 		}
2175 		assert(strd_idx < strd_n);
2176 		assert(!((rte_be_to_cpu_16(cqe->wqe_id) ^ rq_ci) & wq_mask));
2177 		/*
2178 		 * Currently configured to receive a packet per a stride. But if
2179 		 * MTU is adjusted through kernel interface, device could
2180 		 * consume multiple strides without raising an error. In this
2181 		 * case, the packet should be dropped because it is bigger than
2182 		 * the max_rx_pkt_len.
2183 		 */
2184 		if (unlikely(strd_cnt > 1)) {
2185 			++rxq->stats.idropped;
2186 			continue;
2187 		}
2188 		pkt = rte_pktmbuf_alloc(rxq->mp);
2189 		if (unlikely(pkt == NULL)) {
2190 			++rxq->stats.rx_nombuf;
2191 			break;
2192 		}
2193 		len = (byte_cnt & MLX5_MPRQ_LEN_MASK) >> MLX5_MPRQ_LEN_SHIFT;
2194 		assert((int)len >= (rxq->crc_present << 2));
2195 		if (rxq->crc_present)
2196 			len -= ETHER_CRC_LEN;
2197 		offset = strd_idx * strd_sz + strd_shift;
2198 		addr = RTE_PTR_ADD(mlx5_mprq_buf_addr(buf), offset);
2199 		/* Initialize the offload flag. */
2200 		pkt->ol_flags = 0;
2201 		/*
2202 		 * Memcpy packets to the target mbuf if:
2203 		 * - The size of packet is smaller than mprq_max_memcpy_len.
2204 		 * - Out of buffer in the Mempool for Multi-Packet RQ.
2205 		 */
2206 		if (len <= rxq->mprq_max_memcpy_len || rxq->mprq_repl == NULL) {
2207 			/*
2208 			 * When memcpy'ing packet due to out-of-buffer, the
2209 			 * packet must be smaller than the target mbuf.
2210 			 */
2211 			if (unlikely(rte_pktmbuf_tailroom(pkt) < len)) {
2212 				rte_pktmbuf_free_seg(pkt);
2213 				++rxq->stats.idropped;
2214 				continue;
2215 			}
2216 			rte_memcpy(rte_pktmbuf_mtod(pkt, void *), addr, len);
2217 		} else {
2218 			rte_iova_t buf_iova;
2219 			struct rte_mbuf_ext_shared_info *shinfo;
2220 			uint16_t buf_len = strd_cnt * strd_sz;
2221 
2222 			/* Increment the refcnt of the whole chunk. */
2223 			rte_atomic16_add_return(&buf->refcnt, 1);
2224 			assert((uint16_t)rte_atomic16_read(&buf->refcnt) <=
2225 			       strd_n + 1);
2226 			addr = RTE_PTR_SUB(addr, RTE_PKTMBUF_HEADROOM);
2227 			/*
2228 			 * MLX5 device doesn't use iova but it is necessary in a
2229 			 * case where the Rx packet is transmitted via a
2230 			 * different PMD.
2231 			 */
2232 			buf_iova = rte_mempool_virt2iova(buf) +
2233 				   RTE_PTR_DIFF(addr, buf);
2234 			shinfo = rte_pktmbuf_ext_shinfo_init_helper(addr,
2235 					&buf_len, mlx5_mprq_buf_free_cb, buf);
2236 			/*
2237 			 * EXT_ATTACHED_MBUF will be set to pkt->ol_flags when
2238 			 * attaching the stride to mbuf and more offload flags
2239 			 * will be added below by calling rxq_cq_to_mbuf().
2240 			 * Other fields will be overwritten.
2241 			 */
2242 			rte_pktmbuf_attach_extbuf(pkt, addr, buf_iova, buf_len,
2243 						  shinfo);
2244 			rte_pktmbuf_reset_headroom(pkt);
2245 			assert(pkt->ol_flags == EXT_ATTACHED_MBUF);
2246 			/*
2247 			 * Prevent potential overflow due to MTU change through
2248 			 * kernel interface.
2249 			 */
2250 			if (unlikely(rte_pktmbuf_tailroom(pkt) < len)) {
2251 				rte_pktmbuf_free_seg(pkt);
2252 				++rxq->stats.idropped;
2253 				continue;
2254 			}
2255 		}
2256 		rxq_cq_to_mbuf(rxq, pkt, cqe, rss_hash_res);
2257 		PKT_LEN(pkt) = len;
2258 		DATA_LEN(pkt) = len;
2259 		PORT(pkt) = rxq->port_id;
2260 #ifdef MLX5_PMD_SOFT_COUNTERS
2261 		/* Increment bytes counter. */
2262 		rxq->stats.ibytes += PKT_LEN(pkt);
2263 #endif
2264 		/* Return packet. */
2265 		*(pkts++) = pkt;
2266 		++i;
2267 	}
2268 	/* Update the consumer indexes. */
2269 	rxq->consumed_strd = consumed_strd;
2270 	rte_cio_wmb();
2271 	*rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
2272 	if (rq_ci != rxq->rq_ci) {
2273 		rxq->rq_ci = rq_ci;
2274 		rte_cio_wmb();
2275 		*rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
2276 	}
2277 #ifdef MLX5_PMD_SOFT_COUNTERS
2278 	/* Increment packets counter. */
2279 	rxq->stats.ipackets += i;
2280 #endif
2281 	return i;
2282 }
2283 
2284 /**
2285  * Dummy DPDK callback for TX.
2286  *
2287  * This function is used to temporarily replace the real callback during
2288  * unsafe control operations on the queue, or in case of error.
2289  *
2290  * @param dpdk_txq
2291  *   Generic pointer to TX queue structure.
2292  * @param[in] pkts
2293  *   Packets to transmit.
2294  * @param pkts_n
2295  *   Number of packets in array.
2296  *
2297  * @return
2298  *   Number of packets successfully transmitted (<= pkts_n).
2299  */
2300 uint16_t
2301 removed_tx_burst(void *dpdk_txq __rte_unused,
2302 		 struct rte_mbuf **pkts __rte_unused,
2303 		 uint16_t pkts_n __rte_unused)
2304 {
2305 	return 0;
2306 }
2307 
2308 /**
2309  * Dummy DPDK callback for RX.
2310  *
2311  * This function is used to temporarily replace the real callback during
2312  * unsafe control operations on the queue, or in case of error.
2313  *
2314  * @param dpdk_rxq
2315  *   Generic pointer to RX queue structure.
2316  * @param[out] pkts
2317  *   Array to store received packets.
2318  * @param pkts_n
2319  *   Maximum number of packets in array.
2320  *
2321  * @return
2322  *   Number of packets successfully received (<= pkts_n).
2323  */
2324 uint16_t
2325 removed_rx_burst(void *dpdk_txq __rte_unused,
2326 		 struct rte_mbuf **pkts __rte_unused,
2327 		 uint16_t pkts_n __rte_unused)
2328 {
2329 	return 0;
2330 }
2331 
2332 /*
2333  * Vectorized Rx/Tx routines are not compiled in when required vector
2334  * instructions are not supported on a target architecture. The following null
2335  * stubs are needed for linkage when those are not included outside of this file
2336  * (e.g.  mlx5_rxtx_vec_sse.c for x86).
2337  */
2338 
2339 uint16_t __attribute__((weak))
2340 mlx5_tx_burst_raw_vec(void *dpdk_txq __rte_unused,
2341 		      struct rte_mbuf **pkts __rte_unused,
2342 		      uint16_t pkts_n __rte_unused)
2343 {
2344 	return 0;
2345 }
2346 
2347 uint16_t __attribute__((weak))
2348 mlx5_tx_burst_vec(void *dpdk_txq __rte_unused,
2349 		  struct rte_mbuf **pkts __rte_unused,
2350 		  uint16_t pkts_n __rte_unused)
2351 {
2352 	return 0;
2353 }
2354 
2355 uint16_t __attribute__((weak))
2356 mlx5_rx_burst_vec(void *dpdk_txq __rte_unused,
2357 		  struct rte_mbuf **pkts __rte_unused,
2358 		  uint16_t pkts_n __rte_unused)
2359 {
2360 	return 0;
2361 }
2362 
2363 int __attribute__((weak))
2364 mlx5_check_raw_vec_tx_support(struct rte_eth_dev *dev __rte_unused)
2365 {
2366 	return -ENOTSUP;
2367 }
2368 
2369 int __attribute__((weak))
2370 mlx5_check_vec_tx_support(struct rte_eth_dev *dev __rte_unused)
2371 {
2372 	return -ENOTSUP;
2373 }
2374 
2375 int __attribute__((weak))
2376 mlx5_rxq_check_vec_support(struct mlx5_rxq_data *rxq __rte_unused)
2377 {
2378 	return -ENOTSUP;
2379 }
2380 
2381 int __attribute__((weak))
2382 mlx5_check_vec_rx_support(struct rte_eth_dev *dev __rte_unused)
2383 {
2384 	return -ENOTSUP;
2385 }
2386