xref: /dpdk/drivers/net/mlx5/mlx5_rxtx_vec.c (revision 3c2ddbd413e34698b66df65a57b0c760ffe0b3d6)
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright 2017 6WIND S.A.
5  *   Copyright 2017 Mellanox.
6  *
7  *   Redistribution and use in source and binary forms, with or without
8  *   modification, are permitted provided that the following conditions
9  *   are met:
10  *
11  *     * Redistributions of source code must retain the above copyright
12  *       notice, this list of conditions and the following disclaimer.
13  *     * Redistributions in binary form must reproduce the above copyright
14  *       notice, this list of conditions and the following disclaimer in
15  *       the documentation and/or other materials provided with the
16  *       distribution.
17  *     * Neither the name of 6WIND S.A. nor the names of its
18  *       contributors may be used to endorse or promote products derived
19  *       from this software without specific prior written permission.
20  *
21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  */
33 
34 #include <assert.h>
35 #include <stdint.h>
36 #include <string.h>
37 #include <stdlib.h>
38 
39 /* Verbs header. */
40 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
41 #ifdef PEDANTIC
42 #pragma GCC diagnostic ignored "-Wpedantic"
43 #endif
44 #include <infiniband/verbs.h>
45 #include <infiniband/mlx5dv.h>
46 #ifdef PEDANTIC
47 #pragma GCC diagnostic error "-Wpedantic"
48 #endif
49 
50 #include <rte_mbuf.h>
51 #include <rte_mempool.h>
52 #include <rte_prefetch.h>
53 
54 #include "mlx5.h"
55 #include "mlx5_utils.h"
56 #include "mlx5_rxtx.h"
57 #include "mlx5_rxtx_vec.h"
58 #include "mlx5_autoconf.h"
59 #include "mlx5_defs.h"
60 #include "mlx5_prm.h"
61 
62 #ifdef RTE_ARCH_X86_64
63 #include "mlx5_rxtx_vec_sse.h"
64 #else
65 #error "This should not be compiled if SIMD instructions are not supported."
66 #endif
67 
68 /**
69  * Count the number of continuous single segment packets.
70  *
71  * @param pkts
72  *   Pointer to array of packets.
73  * @param pkts_n
74  *   Number of packets.
75  *
76  * @return
77  *   Number of continuous single segment packets.
78  */
79 static inline unsigned int
80 txq_check_multiseg(struct rte_mbuf **pkts, uint16_t pkts_n)
81 {
82 	unsigned int pos;
83 
84 	if (!pkts_n)
85 		return 0;
86 	/* Count the number of continuous single segment packets. */
87 	for (pos = 0; pos < pkts_n; ++pos)
88 		if (NB_SEGS(pkts[pos]) > 1)
89 			break;
90 	return pos;
91 }
92 
93 /**
94  * Count the number of packets having same ol_flags and calculate cs_flags.
95  *
96  * @param txq
97  *   Pointer to TX queue structure.
98  * @param pkts
99  *   Pointer to array of packets.
100  * @param pkts_n
101  *   Number of packets.
102  * @param cs_flags
103  *   Pointer of flags to be returned.
104  *
105  * @return
106  *   Number of packets having same ol_flags.
107  */
108 static inline unsigned int
109 txq_calc_offload(struct mlx5_txq_data *txq, struct rte_mbuf **pkts,
110 		 uint16_t pkts_n, uint8_t *cs_flags)
111 {
112 	unsigned int pos;
113 	const uint64_t ol_mask =
114 		PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM |
115 		PKT_TX_UDP_CKSUM | PKT_TX_TUNNEL_GRE |
116 		PKT_TX_TUNNEL_VXLAN | PKT_TX_OUTER_IP_CKSUM;
117 
118 	if (!pkts_n)
119 		return 0;
120 	/* Count the number of packets having same ol_flags. */
121 	for (pos = 1; pos < pkts_n; ++pos)
122 		if ((pkts[pos]->ol_flags ^ pkts[0]->ol_flags) & ol_mask)
123 			break;
124 	/* Should open another MPW session for the rest. */
125 	if (pkts[0]->ol_flags &
126 	    (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
127 		const uint64_t is_tunneled =
128 			pkts[0]->ol_flags &
129 			(PKT_TX_TUNNEL_GRE |
130 			 PKT_TX_TUNNEL_VXLAN);
131 
132 		if (is_tunneled && txq->tunnel_en) {
133 			*cs_flags = MLX5_ETH_WQE_L3_INNER_CSUM |
134 				    MLX5_ETH_WQE_L4_INNER_CSUM;
135 			if (pkts[0]->ol_flags & PKT_TX_OUTER_IP_CKSUM)
136 				*cs_flags |= MLX5_ETH_WQE_L3_CSUM;
137 		} else {
138 			*cs_flags = MLX5_ETH_WQE_L3_CSUM |
139 				    MLX5_ETH_WQE_L4_CSUM;
140 		}
141 	}
142 	return pos;
143 }
144 
145 /**
146  * DPDK callback for vectorized TX.
147  *
148  * @param dpdk_txq
149  *   Generic pointer to TX queue structure.
150  * @param[in] pkts
151  *   Packets to transmit.
152  * @param pkts_n
153  *   Number of packets in array.
154  *
155  * @return
156  *   Number of packets successfully transmitted (<= pkts_n).
157  */
158 uint16_t
159 mlx5_tx_burst_raw_vec(void *dpdk_txq, struct rte_mbuf **pkts,
160 		      uint16_t pkts_n)
161 {
162 	struct mlx5_txq_data *txq = (struct mlx5_txq_data *)dpdk_txq;
163 	uint16_t nb_tx = 0;
164 
165 	while (pkts_n > nb_tx) {
166 		uint16_t n;
167 		uint16_t ret;
168 
169 		n = RTE_MIN((uint16_t)(pkts_n - nb_tx), MLX5_VPMD_TX_MAX_BURST);
170 		ret = txq_burst_v(txq, &pkts[nb_tx], n, 0);
171 		nb_tx += ret;
172 		if (!ret)
173 			break;
174 	}
175 	return nb_tx;
176 }
177 
178 /**
179  * DPDK callback for vectorized TX with multi-seg packets and offload.
180  *
181  * @param dpdk_txq
182  *   Generic pointer to TX queue structure.
183  * @param[in] pkts
184  *   Packets to transmit.
185  * @param pkts_n
186  *   Number of packets in array.
187  *
188  * @return
189  *   Number of packets successfully transmitted (<= pkts_n).
190  */
191 uint16_t
192 mlx5_tx_burst_vec(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
193 {
194 	struct mlx5_txq_data *txq = (struct mlx5_txq_data *)dpdk_txq;
195 	uint16_t nb_tx = 0;
196 
197 	while (pkts_n > nb_tx) {
198 		uint8_t cs_flags = 0;
199 		uint16_t n;
200 		uint16_t ret;
201 
202 		/* Transmit multi-seg packets in the head of pkts list. */
203 		if (!(txq->flags & ETH_TXQ_FLAGS_NOMULTSEGS) &&
204 		    NB_SEGS(pkts[nb_tx]) > 1)
205 			nb_tx += txq_scatter_v(txq,
206 					       &pkts[nb_tx],
207 					       pkts_n - nb_tx);
208 		n = RTE_MIN((uint16_t)(pkts_n - nb_tx), MLX5_VPMD_TX_MAX_BURST);
209 		if (!(txq->flags & ETH_TXQ_FLAGS_NOMULTSEGS))
210 			n = txq_check_multiseg(&pkts[nb_tx], n);
211 		if (!(txq->flags & ETH_TXQ_FLAGS_NOOFFLOADS))
212 			n = txq_calc_offload(txq, &pkts[nb_tx], n, &cs_flags);
213 		ret = txq_burst_v(txq, &pkts[nb_tx], n, cs_flags);
214 		nb_tx += ret;
215 		if (!ret)
216 			break;
217 	}
218 	return nb_tx;
219 }
220 
221 /**
222  * Skip error packets.
223  *
224  * @param rxq
225  *   Pointer to RX queue structure.
226  * @param[out] pkts
227  *   Array to store received packets.
228  * @param pkts_n
229  *   Maximum number of packets in array.
230  *
231  * @return
232  *   Number of packets successfully received (<= pkts_n).
233  */
234 static uint16_t
235 rxq_handle_pending_error(struct mlx5_rxq_data *rxq, struct rte_mbuf **pkts,
236 			 uint16_t pkts_n)
237 {
238 	uint16_t n = 0;
239 	unsigned int i;
240 #ifdef MLX5_PMD_SOFT_COUNTERS
241 	uint32_t err_bytes = 0;
242 #endif
243 
244 	for (i = 0; i < pkts_n; ++i) {
245 		struct rte_mbuf *pkt = pkts[i];
246 
247 		if (pkt->packet_type == RTE_PTYPE_ALL_MASK) {
248 #ifdef MLX5_PMD_SOFT_COUNTERS
249 			err_bytes += PKT_LEN(pkt);
250 #endif
251 			rte_pktmbuf_free_seg(pkt);
252 		} else {
253 			pkts[n++] = pkt;
254 		}
255 	}
256 	rxq->stats.idropped += (pkts_n - n);
257 #ifdef MLX5_PMD_SOFT_COUNTERS
258 	/* Correct counters of errored completions. */
259 	rxq->stats.ipackets -= (pkts_n - n);
260 	rxq->stats.ibytes -= err_bytes;
261 #endif
262 	rxq->pending_err = 0;
263 	return n;
264 }
265 
266 /**
267  * DPDK callback for vectorized RX.
268  *
269  * @param dpdk_rxq
270  *   Generic pointer to RX queue structure.
271  * @param[out] pkts
272  *   Array to store received packets.
273  * @param pkts_n
274  *   Maximum number of packets in array.
275  *
276  * @return
277  *   Number of packets successfully received (<= pkts_n).
278  */
279 uint16_t
280 mlx5_rx_burst_vec(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
281 {
282 	struct mlx5_rxq_data *rxq = dpdk_rxq;
283 	uint16_t nb_rx;
284 
285 	nb_rx = rxq_burst_v(rxq, pkts, pkts_n);
286 	if (unlikely(rxq->pending_err))
287 		nb_rx = rxq_handle_pending_error(rxq, pkts, nb_rx);
288 	return nb_rx;
289 }
290 
291 /**
292  * Check Tx queue flags are set for raw vectorized Tx.
293  *
294  * @param priv
295  *   Pointer to private structure.
296  *
297  * @return
298  *   1 if supported, negative errno value if not.
299  */
300 int __attribute__((cold))
301 priv_check_raw_vec_tx_support(struct priv *priv)
302 {
303 	uint16_t i;
304 
305 	/* All the configured queues should support. */
306 	for (i = 0; i < priv->txqs_n; ++i) {
307 		struct mlx5_txq_data *txq = (*priv->txqs)[i];
308 
309 		if (!(txq->flags & ETH_TXQ_FLAGS_NOMULTSEGS) ||
310 		    !(txq->flags & ETH_TXQ_FLAGS_NOOFFLOADS))
311 			break;
312 	}
313 	if (i != priv->txqs_n)
314 		return -ENOTSUP;
315 	return 1;
316 }
317 
318 /**
319  * Check a device can support vectorized TX.
320  *
321  * @param priv
322  *   Pointer to private structure.
323  *
324  * @return
325  *   1 if supported, negative errno value if not.
326  */
327 int __attribute__((cold))
328 priv_check_vec_tx_support(struct priv *priv)
329 {
330 	if (!priv->tx_vec_en ||
331 	    priv->txqs_n > MLX5_VPMD_MIN_TXQS ||
332 	    priv->mps != MLX5_MPW_ENHANCED ||
333 	    priv->tso)
334 		return -ENOTSUP;
335 	return 1;
336 }
337 
338 /**
339  * Check a RX queue can support vectorized RX.
340  *
341  * @param rxq
342  *   Pointer to RX queue.
343  *
344  * @return
345  *   1 if supported, negative errno value if not.
346  */
347 int __attribute__((cold))
348 rxq_check_vec_support(struct mlx5_rxq_data *rxq)
349 {
350 	struct mlx5_rxq_ctrl *ctrl =
351 		container_of(rxq, struct mlx5_rxq_ctrl, rxq);
352 
353 	if (!ctrl->priv->rx_vec_en || rxq->sges_n != 0)
354 		return -ENOTSUP;
355 	return 1;
356 }
357 
358 /**
359  * Check a device can support vectorized RX.
360  *
361  * @param priv
362  *   Pointer to private structure.
363  *
364  * @return
365  *   1 if supported, negative errno value if not.
366  */
367 int __attribute__((cold))
368 priv_check_vec_rx_support(struct priv *priv)
369 {
370 	uint16_t i;
371 
372 	if (!priv->rx_vec_en)
373 		return -ENOTSUP;
374 	/* All the configured queues should support. */
375 	for (i = 0; i < priv->rxqs_n; ++i) {
376 		struct mlx5_rxq_data *rxq = (*priv->rxqs)[i];
377 
378 		if (!rxq)
379 			continue;
380 		if (rxq_check_vec_support(rxq) < 0)
381 			break;
382 	}
383 	if (i != priv->rxqs_n)
384 		return -ENOTSUP;
385 	return 1;
386 }
387