net/mlx5: fix link speed capability information
[dpdk.git] / drivers / net / mlx5 / mlx5_ethdev.c
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright 2015 6WIND S.A.
5  *   Copyright 2015 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 <stddef.h>
35 #include <assert.h>
36 #include <unistd.h>
37 #include <stdint.h>
38 #include <stdio.h>
39 #include <string.h>
40 #include <stdlib.h>
41 #include <errno.h>
42 #include <dirent.h>
43 #include <net/if.h>
44 #include <sys/ioctl.h>
45 #include <sys/socket.h>
46 #include <netinet/in.h>
47 #include <linux/ethtool.h>
48 #include <linux/sockios.h>
49 #include <fcntl.h>
50
51 /* DPDK headers don't like -pedantic. */
52 #ifdef PEDANTIC
53 #pragma GCC diagnostic ignored "-Wpedantic"
54 #endif
55 #include <rte_atomic.h>
56 #include <rte_ethdev.h>
57 #include <rte_mbuf.h>
58 #include <rte_common.h>
59 #include <rte_interrupts.h>
60 #include <rte_alarm.h>
61 #include <rte_malloc.h>
62 #ifdef PEDANTIC
63 #pragma GCC diagnostic error "-Wpedantic"
64 #endif
65
66 #include "mlx5.h"
67 #include "mlx5_rxtx.h"
68 #include "mlx5_utils.h"
69
70 /**
71  * Return private structure associated with an Ethernet device.
72  *
73  * @param dev
74  *   Pointer to Ethernet device structure.
75  *
76  * @return
77  *   Pointer to private structure.
78  */
79 struct priv *
80 mlx5_get_priv(struct rte_eth_dev *dev)
81 {
82         struct mlx5_secondary_data *sd;
83
84         if (!mlx5_is_secondary())
85                 return dev->data->dev_private;
86         sd = &mlx5_secondary_data[dev->data->port_id];
87         return sd->data.dev_private;
88 }
89
90 /**
91  * Check if running as a secondary process.
92  *
93  * @return
94  *   Nonzero if running as a secondary process.
95  */
96 inline int
97 mlx5_is_secondary(void)
98 {
99         return rte_eal_process_type() != RTE_PROC_PRIMARY;
100 }
101
102 /**
103  * Get interface name from private structure.
104  *
105  * @param[in] priv
106  *   Pointer to private structure.
107  * @param[out] ifname
108  *   Interface name output buffer.
109  *
110  * @return
111  *   0 on success, -1 on failure and errno is set.
112  */
113 int
114 priv_get_ifname(const struct priv *priv, char (*ifname)[IF_NAMESIZE])
115 {
116         DIR *dir;
117         struct dirent *dent;
118         unsigned int dev_type = 0;
119         unsigned int dev_port_prev = ~0u;
120         char match[IF_NAMESIZE] = "";
121
122         {
123                 MKSTR(path, "%s/device/net", priv->ctx->device->ibdev_path);
124
125                 dir = opendir(path);
126                 if (dir == NULL)
127                         return -1;
128         }
129         while ((dent = readdir(dir)) != NULL) {
130                 char *name = dent->d_name;
131                 FILE *file;
132                 unsigned int dev_port;
133                 int r;
134
135                 if ((name[0] == '.') &&
136                     ((name[1] == '\0') ||
137                      ((name[1] == '.') && (name[2] == '\0'))))
138                         continue;
139
140                 MKSTR(path, "%s/device/net/%s/%s",
141                       priv->ctx->device->ibdev_path, name,
142                       (dev_type ? "dev_id" : "dev_port"));
143
144                 file = fopen(path, "rb");
145                 if (file == NULL) {
146                         if (errno != ENOENT)
147                                 continue;
148                         /*
149                          * Switch to dev_id when dev_port does not exist as
150                          * is the case with Linux kernel versions < 3.15.
151                          */
152 try_dev_id:
153                         match[0] = '\0';
154                         if (dev_type)
155                                 break;
156                         dev_type = 1;
157                         dev_port_prev = ~0u;
158                         rewinddir(dir);
159                         continue;
160                 }
161                 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
162                 fclose(file);
163                 if (r != 1)
164                         continue;
165                 /*
166                  * Switch to dev_id when dev_port returns the same value for
167                  * all ports. May happen when using a MOFED release older than
168                  * 3.0 with a Linux kernel >= 3.15.
169                  */
170                 if (dev_port == dev_port_prev)
171                         goto try_dev_id;
172                 dev_port_prev = dev_port;
173                 if (dev_port == (priv->port - 1u))
174                         snprintf(match, sizeof(match), "%s", name);
175         }
176         closedir(dir);
177         if (match[0] == '\0')
178                 return -1;
179         strncpy(*ifname, match, sizeof(*ifname));
180         return 0;
181 }
182
183 /**
184  * Read from sysfs entry.
185  *
186  * @param[in] priv
187  *   Pointer to private structure.
188  * @param[in] entry
189  *   Entry name relative to sysfs path.
190  * @param[out] buf
191  *   Data output buffer.
192  * @param size
193  *   Buffer size.
194  *
195  * @return
196  *   0 on success, -1 on failure and errno is set.
197  */
198 static int
199 priv_sysfs_read(const struct priv *priv, const char *entry,
200                 char *buf, size_t size)
201 {
202         char ifname[IF_NAMESIZE];
203         FILE *file;
204         int ret;
205         int err;
206
207         if (priv_get_ifname(priv, &ifname))
208                 return -1;
209
210         MKSTR(path, "%s/device/net/%s/%s", priv->ctx->device->ibdev_path,
211               ifname, entry);
212
213         file = fopen(path, "rb");
214         if (file == NULL)
215                 return -1;
216         ret = fread(buf, 1, size, file);
217         err = errno;
218         if (((size_t)ret < size) && (ferror(file)))
219                 ret = -1;
220         else
221                 ret = size;
222         fclose(file);
223         errno = err;
224         return ret;
225 }
226
227 /**
228  * Write to sysfs entry.
229  *
230  * @param[in] priv
231  *   Pointer to private structure.
232  * @param[in] entry
233  *   Entry name relative to sysfs path.
234  * @param[in] buf
235  *   Data buffer.
236  * @param size
237  *   Buffer size.
238  *
239  * @return
240  *   0 on success, -1 on failure and errno is set.
241  */
242 static int
243 priv_sysfs_write(const struct priv *priv, const char *entry,
244                  char *buf, size_t size)
245 {
246         char ifname[IF_NAMESIZE];
247         FILE *file;
248         int ret;
249         int err;
250
251         if (priv_get_ifname(priv, &ifname))
252                 return -1;
253
254         MKSTR(path, "%s/device/net/%s/%s", priv->ctx->device->ibdev_path,
255               ifname, entry);
256
257         file = fopen(path, "wb");
258         if (file == NULL)
259                 return -1;
260         ret = fwrite(buf, 1, size, file);
261         err = errno;
262         if (((size_t)ret < size) || (ferror(file)))
263                 ret = -1;
264         else
265                 ret = size;
266         fclose(file);
267         errno = err;
268         return ret;
269 }
270
271 /**
272  * Get unsigned long sysfs property.
273  *
274  * @param priv
275  *   Pointer to private structure.
276  * @param[in] name
277  *   Entry name relative to sysfs path.
278  * @param[out] value
279  *   Value output buffer.
280  *
281  * @return
282  *   0 on success, -1 on failure and errno is set.
283  */
284 static int
285 priv_get_sysfs_ulong(struct priv *priv, const char *name, unsigned long *value)
286 {
287         int ret;
288         unsigned long value_ret;
289         char value_str[32];
290
291         ret = priv_sysfs_read(priv, name, value_str, (sizeof(value_str) - 1));
292         if (ret == -1) {
293                 DEBUG("cannot read %s value from sysfs: %s",
294                       name, strerror(errno));
295                 return -1;
296         }
297         value_str[ret] = '\0';
298         errno = 0;
299         value_ret = strtoul(value_str, NULL, 0);
300         if (errno) {
301                 DEBUG("invalid %s value `%s': %s", name, value_str,
302                       strerror(errno));
303                 return -1;
304         }
305         *value = value_ret;
306         return 0;
307 }
308
309 /**
310  * Set unsigned long sysfs property.
311  *
312  * @param priv
313  *   Pointer to private structure.
314  * @param[in] name
315  *   Entry name relative to sysfs path.
316  * @param value
317  *   Value to set.
318  *
319  * @return
320  *   0 on success, -1 on failure and errno is set.
321  */
322 static int
323 priv_set_sysfs_ulong(struct priv *priv, const char *name, unsigned long value)
324 {
325         int ret;
326         MKSTR(value_str, "%lu", value);
327
328         ret = priv_sysfs_write(priv, name, value_str, (sizeof(value_str) - 1));
329         if (ret == -1) {
330                 DEBUG("cannot write %s `%s' (%lu) to sysfs: %s",
331                       name, value_str, value, strerror(errno));
332                 return -1;
333         }
334         return 0;
335 }
336
337 /**
338  * Perform ifreq ioctl() on associated Ethernet device.
339  *
340  * @param[in] priv
341  *   Pointer to private structure.
342  * @param req
343  *   Request number to pass to ioctl().
344  * @param[out] ifr
345  *   Interface request structure output buffer.
346  *
347  * @return
348  *   0 on success, -1 on failure and errno is set.
349  */
350 int
351 priv_ifreq(const struct priv *priv, int req, struct ifreq *ifr)
352 {
353         int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
354         int ret = -1;
355
356         if (sock == -1)
357                 return ret;
358         if (priv_get_ifname(priv, &ifr->ifr_name) == 0)
359                 ret = ioctl(sock, req, ifr);
360         close(sock);
361         return ret;
362 }
363
364 /**
365  * Return the number of active VFs for the current device.
366  *
367  * @param[in] priv
368  *   Pointer to private structure.
369  * @param[out] num_vfs
370  *   Number of active VFs.
371  *
372  * @return
373  *   0 on success, -1 on failure and errno is set.
374  */
375 int
376 priv_get_num_vfs(struct priv *priv, uint16_t *num_vfs)
377 {
378         /* The sysfs entry name depends on the operating system. */
379         const char **name = (const char *[]){
380                 "device/sriov_numvfs",
381                 "device/mlx5_num_vfs",
382                 NULL,
383         };
384         int ret;
385
386         do {
387                 unsigned long ulong_num_vfs;
388
389                 ret = priv_get_sysfs_ulong(priv, *name, &ulong_num_vfs);
390                 if (!ret)
391                         *num_vfs = ulong_num_vfs;
392         } while (*(++name) && ret);
393         return ret;
394 }
395
396 /**
397  * Get device MTU.
398  *
399  * @param priv
400  *   Pointer to private structure.
401  * @param[out] mtu
402  *   MTU value output buffer.
403  *
404  * @return
405  *   0 on success, -1 on failure and errno is set.
406  */
407 int
408 priv_get_mtu(struct priv *priv, uint16_t *mtu)
409 {
410         unsigned long ulong_mtu;
411
412         if (priv_get_sysfs_ulong(priv, "mtu", &ulong_mtu) == -1)
413                 return -1;
414         *mtu = ulong_mtu;
415         return 0;
416 }
417
418 /**
419  * Set device MTU.
420  *
421  * @param priv
422  *   Pointer to private structure.
423  * @param mtu
424  *   MTU value to set.
425  *
426  * @return
427  *   0 on success, -1 on failure and errno is set.
428  */
429 static int
430 priv_set_mtu(struct priv *priv, uint16_t mtu)
431 {
432         uint16_t new_mtu;
433
434         if (priv_set_sysfs_ulong(priv, "mtu", mtu) ||
435             priv_get_mtu(priv, &new_mtu))
436                 return -1;
437         if (new_mtu == mtu)
438                 return 0;
439         errno = EINVAL;
440         return -1;
441 }
442
443 /**
444  * Set device flags.
445  *
446  * @param priv
447  *   Pointer to private structure.
448  * @param keep
449  *   Bitmask for flags that must remain untouched.
450  * @param flags
451  *   Bitmask for flags to modify.
452  *
453  * @return
454  *   0 on success, -1 on failure and errno is set.
455  */
456 int
457 priv_set_flags(struct priv *priv, unsigned int keep, unsigned int flags)
458 {
459         unsigned long tmp;
460
461         if (priv_get_sysfs_ulong(priv, "flags", &tmp) == -1)
462                 return -1;
463         tmp &= keep;
464         tmp |= (flags & (~keep));
465         return priv_set_sysfs_ulong(priv, "flags", tmp);
466 }
467
468 /**
469  * Ethernet device configuration.
470  *
471  * Prepare the driver for a given number of TX and RX queues.
472  *
473  * @param dev
474  *   Pointer to Ethernet device structure.
475  *
476  * @return
477  *   0 on success, errno value on failure.
478  */
479 static int
480 dev_configure(struct rte_eth_dev *dev)
481 {
482         struct priv *priv = dev->data->dev_private;
483         unsigned int rxqs_n = dev->data->nb_rx_queues;
484         unsigned int txqs_n = dev->data->nb_tx_queues;
485         unsigned int i;
486         unsigned int j;
487         unsigned int reta_idx_n;
488
489         priv->rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
490         priv->rxqs = (void *)dev->data->rx_queues;
491         priv->txqs = (void *)dev->data->tx_queues;
492         if (txqs_n != priv->txqs_n) {
493                 INFO("%p: TX queues number update: %u -> %u",
494                      (void *)dev, priv->txqs_n, txqs_n);
495                 priv->txqs_n = txqs_n;
496         }
497         if (rxqs_n > priv->ind_table_max_size) {
498                 ERROR("cannot handle this many RX queues (%u)", rxqs_n);
499                 return EINVAL;
500         }
501         if (rxqs_n == priv->rxqs_n)
502                 return 0;
503         INFO("%p: RX queues number update: %u -> %u",
504              (void *)dev, priv->rxqs_n, rxqs_n);
505         priv->rxqs_n = rxqs_n;
506         /* If the requested number of RX queues is not a power of two, use the
507          * maximum indirection table size for better balancing.
508          * The result is always rounded to the next power of two. */
509         reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
510                                      priv->ind_table_max_size :
511                                      rxqs_n));
512         if (priv_rss_reta_index_resize(priv, reta_idx_n))
513                 return ENOMEM;
514         /* When the number of RX queues is not a power of two, the remaining
515          * table entries are padded with reused WQs and hashes are not spread
516          * uniformly. */
517         for (i = 0, j = 0; (i != reta_idx_n); ++i) {
518                 (*priv->reta_idx)[i] = j;
519                 if (++j == rxqs_n)
520                         j = 0;
521         }
522         return 0;
523 }
524
525 /**
526  * DPDK callback for Ethernet device configuration.
527  *
528  * @param dev
529  *   Pointer to Ethernet device structure.
530  *
531  * @return
532  *   0 on success, negative errno value on failure.
533  */
534 int
535 mlx5_dev_configure(struct rte_eth_dev *dev)
536 {
537         struct priv *priv = dev->data->dev_private;
538         int ret;
539
540         if (mlx5_is_secondary())
541                 return -E_RTE_SECONDARY;
542
543         priv_lock(priv);
544         ret = dev_configure(dev);
545         assert(ret >= 0);
546         priv_unlock(priv);
547         return -ret;
548 }
549
550 /**
551  * DPDK callback to get information about the device.
552  *
553  * @param dev
554  *   Pointer to Ethernet device structure.
555  * @param[out] info
556  *   Info structure output buffer.
557  */
558 void
559 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
560 {
561         struct priv *priv = mlx5_get_priv(dev);
562         unsigned int max;
563         char ifname[IF_NAMESIZE];
564
565         priv_lock(priv);
566         /* FIXME: we should ask the device for these values. */
567         info->min_rx_bufsize = 32;
568         info->max_rx_pktlen = 65536;
569         /*
570          * Since we need one CQ per QP, the limit is the minimum number
571          * between the two values.
572          */
573         max = ((priv->device_attr.max_cq > priv->device_attr.max_qp) ?
574                priv->device_attr.max_qp : priv->device_attr.max_cq);
575         /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
576         if (max >= 65535)
577                 max = 65535;
578         info->max_rx_queues = max;
579         info->max_tx_queues = max;
580         info->max_mac_addrs = RTE_DIM(priv->mac);
581         info->rx_offload_capa =
582                 (priv->hw_csum ?
583                  (DEV_RX_OFFLOAD_IPV4_CKSUM |
584                   DEV_RX_OFFLOAD_UDP_CKSUM |
585                   DEV_RX_OFFLOAD_TCP_CKSUM) :
586                  0) |
587                 (priv->hw_vlan_strip ? DEV_RX_OFFLOAD_VLAN_STRIP : 0);
588         if (!priv->mps)
589                 info->tx_offload_capa = DEV_TX_OFFLOAD_VLAN_INSERT;
590         if (priv->hw_csum)
591                 info->tx_offload_capa |=
592                         (DEV_TX_OFFLOAD_IPV4_CKSUM |
593                          DEV_TX_OFFLOAD_UDP_CKSUM |
594                          DEV_TX_OFFLOAD_TCP_CKSUM);
595         if (priv_get_ifname(priv, &ifname) == 0)
596                 info->if_index = if_nametoindex(ifname);
597         /* FIXME: RETA update/query API expects the callee to know the size of
598          * the indirection table, for this PMD the size varies depending on
599          * the number of RX queues, it becomes impossible to find the correct
600          * size if it is not fixed.
601          * The API should be updated to solve this problem. */
602         info->reta_size = priv->ind_table_max_size;
603         info->hash_key_size = ((*priv->rss_conf) ?
604                                (*priv->rss_conf)[0]->rss_key_len :
605                                0);
606         info->speed_capa = priv->link_speed_capa;
607         priv_unlock(priv);
608 }
609
610 const uint32_t *
611 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
612 {
613         static const uint32_t ptypes[] = {
614                 /* refers to rxq_cq_to_pkt_type() */
615                 RTE_PTYPE_L3_IPV4,
616                 RTE_PTYPE_L3_IPV6,
617                 RTE_PTYPE_INNER_L3_IPV4,
618                 RTE_PTYPE_INNER_L3_IPV6,
619                 RTE_PTYPE_UNKNOWN
620
621         };
622
623         if (dev->rx_pkt_burst == mlx5_rx_burst)
624                 return ptypes;
625         return NULL;
626 }
627
628 /**
629  * DPDK callback to retrieve physical link information (unlocked version).
630  *
631  * @param dev
632  *   Pointer to Ethernet device structure.
633  * @param wait_to_complete
634  *   Wait for request completion (ignored).
635  */
636 int
637 mlx5_link_update_unlocked(struct rte_eth_dev *dev, int wait_to_complete)
638 {
639         struct priv *priv = mlx5_get_priv(dev);
640         struct ethtool_cmd edata = {
641                 .cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
642         };
643         struct ifreq ifr;
644         struct rte_eth_link dev_link;
645         int link_speed = 0;
646
647         (void)wait_to_complete;
648         if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
649                 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
650                 return -1;
651         }
652         memset(&dev_link, 0, sizeof(dev_link));
653         dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
654                                 (ifr.ifr_flags & IFF_RUNNING));
655         ifr.ifr_data = (void *)&edata;
656         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
657                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
658                      strerror(errno));
659                 return -1;
660         }
661         link_speed = ethtool_cmd_speed(&edata);
662         if (link_speed == -1)
663                 dev_link.link_speed = 0;
664         else
665                 dev_link.link_speed = link_speed;
666         priv->link_speed_capa = 0;
667         if (edata.supported & SUPPORTED_Autoneg)
668                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
669         if (edata.supported & (SUPPORTED_1000baseT_Full |
670                                SUPPORTED_1000baseKX_Full))
671                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
672         if (edata.supported & SUPPORTED_10000baseKR_Full)
673                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
674         if (edata.supported & (SUPPORTED_40000baseKR4_Full |
675                                SUPPORTED_40000baseCR4_Full |
676                                SUPPORTED_40000baseSR4_Full |
677                                SUPPORTED_40000baseLR4_Full))
678                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
679         dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
680                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
681         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
682                         ETH_LINK_SPEED_FIXED);
683         if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
684                 /* Link status changed. */
685                 dev->data->dev_link = dev_link;
686                 return 0;
687         }
688         /* Link status is still the same. */
689         return -1;
690 }
691
692 /**
693  * DPDK callback to retrieve physical link information.
694  *
695  * @param dev
696  *   Pointer to Ethernet device structure.
697  * @param wait_to_complete
698  *   Wait for request completion (ignored).
699  */
700 int
701 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
702 {
703         struct priv *priv = mlx5_get_priv(dev);
704         int ret;
705
706         priv_lock(priv);
707         ret = mlx5_link_update_unlocked(dev, wait_to_complete);
708         priv_unlock(priv);
709         return ret;
710 }
711
712 /**
713  * DPDK callback to change the MTU.
714  *
715  * Setting the MTU affects hardware MRU (packets larger than the MTU cannot be
716  * received). Use this as a hint to enable/disable scattered packets support
717  * and improve performance when not needed.
718  * Since failure is not an option, reconfiguring queues on the fly is not
719  * recommended.
720  *
721  * @param dev
722  *   Pointer to Ethernet device structure.
723  * @param in_mtu
724  *   New MTU.
725  *
726  * @return
727  *   0 on success, negative errno value on failure.
728  */
729 int
730 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
731 {
732         struct priv *priv = dev->data->dev_private;
733         int ret = 0;
734         unsigned int i;
735         uint16_t (*rx_func)(void *, struct rte_mbuf **, uint16_t) =
736                 mlx5_rx_burst;
737         unsigned int max_frame_len;
738         int rehash;
739         int restart = priv->started;
740
741         if (mlx5_is_secondary())
742                 return -E_RTE_SECONDARY;
743
744         priv_lock(priv);
745         /* Set kernel interface MTU first. */
746         if (priv_set_mtu(priv, mtu)) {
747                 ret = errno;
748                 WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
749                      strerror(ret));
750                 goto out;
751         } else
752                 DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
753         /* Temporarily replace RX handler with a fake one, assuming it has not
754          * been copied elsewhere. */
755         dev->rx_pkt_burst = removed_rx_burst;
756         /* Make sure everyone has left mlx5_rx_burst() and uses
757          * removed_rx_burst() instead. */
758         rte_wmb();
759         usleep(1000);
760         /* MTU does not include header and CRC. */
761         max_frame_len = ETHER_HDR_LEN + mtu + ETHER_CRC_LEN;
762         /* Check if at least one queue is going to need a SGE update. */
763         for (i = 0; i != priv->rxqs_n; ++i) {
764                 struct rxq *rxq = (*priv->rxqs)[i];
765                 unsigned int mb_len;
766                 unsigned int size = RTE_PKTMBUF_HEADROOM + max_frame_len;
767                 unsigned int sges_n;
768
769                 if (rxq == NULL)
770                         continue;
771                 mb_len = rte_pktmbuf_data_room_size(rxq->mp);
772                 assert(mb_len >= RTE_PKTMBUF_HEADROOM);
773                 /*
774                  * Determine the number of SGEs needed for a full packet
775                  * and round it to the next power of two.
776                  */
777                 sges_n = log2above((size / mb_len) + !!(size % mb_len));
778                 if (sges_n != rxq->sges_n)
779                         break;
780         }
781         /*
782          * If all queues have the right number of SGEs, a simple rehash
783          * of their buffers is enough, otherwise SGE information can only
784          * be updated in a queue by recreating it. All resources that depend
785          * on queues (flows, indirection tables) must be recreated as well in
786          * that case.
787          */
788         rehash = (i == priv->rxqs_n);
789         if (!rehash) {
790                 /* Clean up everything as with mlx5_dev_stop(). */
791                 priv_special_flow_disable_all(priv);
792                 priv_mac_addrs_disable(priv);
793                 priv_destroy_hash_rxqs(priv);
794                 priv_fdir_disable(priv);
795                 priv_dev_interrupt_handler_uninstall(priv, dev);
796         }
797 recover:
798         /* Reconfigure each RX queue. */
799         for (i = 0; (i != priv->rxqs_n); ++i) {
800                 struct rxq *rxq = (*priv->rxqs)[i];
801                 struct rxq_ctrl *rxq_ctrl =
802                         container_of(rxq, struct rxq_ctrl, rxq);
803                 int sp;
804                 unsigned int mb_len;
805                 unsigned int tmp;
806
807                 if (rxq == NULL)
808                         continue;
809                 mb_len = rte_pktmbuf_data_room_size(rxq->mp);
810                 assert(mb_len >= RTE_PKTMBUF_HEADROOM);
811                 /* Toggle scattered support (sp) if necessary. */
812                 sp = (max_frame_len > (mb_len - RTE_PKTMBUF_HEADROOM));
813                 /* Provide new values to rxq_setup(). */
814                 dev->data->dev_conf.rxmode.jumbo_frame = sp;
815                 dev->data->dev_conf.rxmode.max_rx_pkt_len = max_frame_len;
816                 if (rehash)
817                         ret = rxq_rehash(dev, rxq_ctrl);
818                 else
819                         ret = rxq_ctrl_setup(dev, rxq_ctrl, 1 << rxq->elts_n,
820                                              rxq_ctrl->socket, NULL, rxq->mp);
821                 if (!ret)
822                         continue;
823                 /* Attempt to roll back in case of error. */
824                 tmp = (mb_len << rxq->sges_n) - RTE_PKTMBUF_HEADROOM;
825                 if (max_frame_len != tmp) {
826                         max_frame_len = tmp;
827                         goto recover;
828                 }
829                 /* Double fault, disable RX. */
830                 break;
831         }
832         /*
833          * Use a safe RX burst function in case of error, otherwise mimic
834          * mlx5_dev_start().
835          */
836         if (ret) {
837                 ERROR("unable to reconfigure RX queues, RX disabled");
838                 rx_func = removed_rx_burst;
839         } else if (restart &&
840                  !rehash &&
841                  !priv_create_hash_rxqs(priv) &&
842                  !priv_rehash_flows(priv)) {
843                 if (dev->data->dev_conf.fdir_conf.mode == RTE_FDIR_MODE_NONE)
844                         priv_fdir_enable(priv);
845                 priv_dev_interrupt_handler_install(priv, dev);
846         }
847         priv->mtu = mtu;
848         /* Burst functions can now be called again. */
849         rte_wmb();
850         dev->rx_pkt_burst = rx_func;
851 out:
852         priv_unlock(priv);
853         assert(ret >= 0);
854         return -ret;
855 }
856
857 /**
858  * DPDK callback to get flow control status.
859  *
860  * @param dev
861  *   Pointer to Ethernet device structure.
862  * @param[out] fc_conf
863  *   Flow control output buffer.
864  *
865  * @return
866  *   0 on success, negative errno value on failure.
867  */
868 int
869 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
870 {
871         struct priv *priv = dev->data->dev_private;
872         struct ifreq ifr;
873         struct ethtool_pauseparam ethpause = {
874                 .cmd = ETHTOOL_GPAUSEPARAM
875         };
876         int ret;
877
878         if (mlx5_is_secondary())
879                 return -E_RTE_SECONDARY;
880
881         ifr.ifr_data = (void *)&ethpause;
882         priv_lock(priv);
883         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
884                 ret = errno;
885                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
886                      " failed: %s",
887                      strerror(ret));
888                 goto out;
889         }
890
891         fc_conf->autoneg = ethpause.autoneg;
892         if (ethpause.rx_pause && ethpause.tx_pause)
893                 fc_conf->mode = RTE_FC_FULL;
894         else if (ethpause.rx_pause)
895                 fc_conf->mode = RTE_FC_RX_PAUSE;
896         else if (ethpause.tx_pause)
897                 fc_conf->mode = RTE_FC_TX_PAUSE;
898         else
899                 fc_conf->mode = RTE_FC_NONE;
900         ret = 0;
901
902 out:
903         priv_unlock(priv);
904         assert(ret >= 0);
905         return -ret;
906 }
907
908 /**
909  * DPDK callback to modify flow control parameters.
910  *
911  * @param dev
912  *   Pointer to Ethernet device structure.
913  * @param[in] fc_conf
914  *   Flow control parameters.
915  *
916  * @return
917  *   0 on success, negative errno value on failure.
918  */
919 int
920 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
921 {
922         struct priv *priv = dev->data->dev_private;
923         struct ifreq ifr;
924         struct ethtool_pauseparam ethpause = {
925                 .cmd = ETHTOOL_SPAUSEPARAM
926         };
927         int ret;
928
929         if (mlx5_is_secondary())
930                 return -E_RTE_SECONDARY;
931
932         ifr.ifr_data = (void *)&ethpause;
933         ethpause.autoneg = fc_conf->autoneg;
934         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
935             (fc_conf->mode & RTE_FC_RX_PAUSE))
936                 ethpause.rx_pause = 1;
937         else
938                 ethpause.rx_pause = 0;
939
940         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
941             (fc_conf->mode & RTE_FC_TX_PAUSE))
942                 ethpause.tx_pause = 1;
943         else
944                 ethpause.tx_pause = 0;
945
946         priv_lock(priv);
947         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
948                 ret = errno;
949                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
950                      " failed: %s",
951                      strerror(ret));
952                 goto out;
953         }
954         ret = 0;
955
956 out:
957         priv_unlock(priv);
958         assert(ret >= 0);
959         return -ret;
960 }
961
962 /**
963  * Get PCI information from struct ibv_device.
964  *
965  * @param device
966  *   Pointer to Ethernet device structure.
967  * @param[out] pci_addr
968  *   PCI bus address output buffer.
969  *
970  * @return
971  *   0 on success, -1 on failure and errno is set.
972  */
973 int
974 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
975                             struct rte_pci_addr *pci_addr)
976 {
977         FILE *file;
978         char line[32];
979         MKSTR(path, "%s/device/uevent", device->ibdev_path);
980
981         file = fopen(path, "rb");
982         if (file == NULL)
983                 return -1;
984         while (fgets(line, sizeof(line), file) == line) {
985                 size_t len = strlen(line);
986                 int ret;
987
988                 /* Truncate long lines. */
989                 if (len == (sizeof(line) - 1))
990                         while (line[(len - 1)] != '\n') {
991                                 ret = fgetc(file);
992                                 if (ret == EOF)
993                                         break;
994                                 line[(len - 1)] = ret;
995                         }
996                 /* Extract information. */
997                 if (sscanf(line,
998                            "PCI_SLOT_NAME="
999                            "%" SCNx16 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
1000                            &pci_addr->domain,
1001                            &pci_addr->bus,
1002                            &pci_addr->devid,
1003                            &pci_addr->function) == 4) {
1004                         ret = 0;
1005                         break;
1006                 }
1007         }
1008         fclose(file);
1009         return 0;
1010 }
1011
1012 /**
1013  * Link status handler.
1014  *
1015  * @param priv
1016  *   Pointer to private structure.
1017  * @param dev
1018  *   Pointer to the rte_eth_dev structure.
1019  *
1020  * @return
1021  *   Nonzero if the callback process can be called immediately.
1022  */
1023 static int
1024 priv_dev_link_status_handler(struct priv *priv, struct rte_eth_dev *dev)
1025 {
1026         struct ibv_async_event event;
1027         int port_change = 0;
1028         int ret = 0;
1029
1030         /* Read all message and acknowledge them. */
1031         for (;;) {
1032                 if (ibv_get_async_event(priv->ctx, &event))
1033                         break;
1034
1035                 if (event.event_type == IBV_EVENT_PORT_ACTIVE ||
1036                     event.event_type == IBV_EVENT_PORT_ERR)
1037                         port_change = 1;
1038                 else
1039                         DEBUG("event type %d on port %d not handled",
1040                               event.event_type, event.element.port_num);
1041                 ibv_ack_async_event(&event);
1042         }
1043
1044         if (port_change ^ priv->pending_alarm) {
1045                 struct rte_eth_link *link = &dev->data->dev_link;
1046
1047                 priv->pending_alarm = 0;
1048                 mlx5_link_update_unlocked(dev, 0);
1049                 if (((link->link_speed == 0) && link->link_status) ||
1050                     ((link->link_speed != 0) && !link->link_status)) {
1051                         /* Inconsistent status, check again later. */
1052                         priv->pending_alarm = 1;
1053                         rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
1054                                           mlx5_dev_link_status_handler,
1055                                           dev);
1056                 } else
1057                         ret = 1;
1058         }
1059         return ret;
1060 }
1061
1062 /**
1063  * Handle delayed link status event.
1064  *
1065  * @param arg
1066  *   Registered argument.
1067  */
1068 void
1069 mlx5_dev_link_status_handler(void *arg)
1070 {
1071         struct rte_eth_dev *dev = arg;
1072         struct priv *priv = dev->data->dev_private;
1073         int ret;
1074
1075         priv_lock(priv);
1076         assert(priv->pending_alarm == 1);
1077         ret = priv_dev_link_status_handler(priv, dev);
1078         priv_unlock(priv);
1079         if (ret)
1080                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1081 }
1082
1083 /**
1084  * Handle interrupts from the NIC.
1085  *
1086  * @param[in] intr_handle
1087  *   Interrupt handler.
1088  * @param cb_arg
1089  *   Callback argument.
1090  */
1091 void
1092 mlx5_dev_interrupt_handler(struct rte_intr_handle *intr_handle, void *cb_arg)
1093 {
1094         struct rte_eth_dev *dev = cb_arg;
1095         struct priv *priv = dev->data->dev_private;
1096         int ret;
1097
1098         (void)intr_handle;
1099         priv_lock(priv);
1100         ret = priv_dev_link_status_handler(priv, dev);
1101         priv_unlock(priv);
1102         if (ret)
1103                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1104 }
1105
1106 /**
1107  * Uninstall interrupt handler.
1108  *
1109  * @param priv
1110  *   Pointer to private structure.
1111  * @param dev
1112  *   Pointer to the rte_eth_dev structure.
1113  */
1114 void
1115 priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
1116 {
1117         if (!dev->data->dev_conf.intr_conf.lsc)
1118                 return;
1119         rte_intr_callback_unregister(&priv->intr_handle,
1120                                      mlx5_dev_interrupt_handler,
1121                                      dev);
1122         if (priv->pending_alarm)
1123                 rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev);
1124         priv->pending_alarm = 0;
1125         priv->intr_handle.fd = 0;
1126         priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1127 }
1128
1129 /**
1130  * Install interrupt handler.
1131  *
1132  * @param priv
1133  *   Pointer to private structure.
1134  * @param dev
1135  *   Pointer to the rte_eth_dev structure.
1136  */
1137 void
1138 priv_dev_interrupt_handler_install(struct priv *priv, struct rte_eth_dev *dev)
1139 {
1140         int rc, flags;
1141
1142         if (!dev->data->dev_conf.intr_conf.lsc)
1143                 return;
1144         assert(priv->ctx->async_fd > 0);
1145         flags = fcntl(priv->ctx->async_fd, F_GETFL);
1146         rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1147         if (rc < 0) {
1148                 INFO("failed to change file descriptor async event queue");
1149                 dev->data->dev_conf.intr_conf.lsc = 0;
1150         } else {
1151                 priv->intr_handle.fd = priv->ctx->async_fd;
1152                 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1153                 rte_intr_callback_register(&priv->intr_handle,
1154                                            mlx5_dev_interrupt_handler,
1155                                            dev);
1156         }
1157 }
1158
1159 /**
1160  * Change the link state (UP / DOWN).
1161  *
1162  * @param priv
1163  *   Pointer to Ethernet device structure.
1164  * @param up
1165  *   Nonzero for link up, otherwise link down.
1166  *
1167  * @return
1168  *   0 on success, errno value on failure.
1169  */
1170 static int
1171 priv_set_link(struct priv *priv, int up)
1172 {
1173         struct rte_eth_dev *dev = priv->dev;
1174         int err;
1175
1176         if (up) {
1177                 err = priv_set_flags(priv, ~IFF_UP, IFF_UP);
1178                 if (err)
1179                         return err;
1180                 priv_select_tx_function(priv);
1181                 priv_select_rx_function(priv);
1182         } else {
1183                 err = priv_set_flags(priv, ~IFF_UP, ~IFF_UP);
1184                 if (err)
1185                         return err;
1186                 dev->rx_pkt_burst = removed_rx_burst;
1187                 dev->tx_pkt_burst = removed_tx_burst;
1188         }
1189         return 0;
1190 }
1191
1192 /**
1193  * DPDK callback to bring the link DOWN.
1194  *
1195  * @param dev
1196  *   Pointer to Ethernet device structure.
1197  *
1198  * @return
1199  *   0 on success, errno value on failure.
1200  */
1201 int
1202 mlx5_set_link_down(struct rte_eth_dev *dev)
1203 {
1204         struct priv *priv = dev->data->dev_private;
1205         int err;
1206
1207         priv_lock(priv);
1208         err = priv_set_link(priv, 0);
1209         priv_unlock(priv);
1210         return err;
1211 }
1212
1213 /**
1214  * DPDK callback to bring the link UP.
1215  *
1216  * @param dev
1217  *   Pointer to Ethernet device structure.
1218  *
1219  * @return
1220  *   0 on success, errno value on failure.
1221  */
1222 int
1223 mlx5_set_link_up(struct rte_eth_dev *dev)
1224 {
1225         struct priv *priv = dev->data->dev_private;
1226         int err;
1227
1228         priv_lock(priv);
1229         err = priv_set_link(priv, 1);
1230         priv_unlock(priv);
1231         return err;
1232 }
1233
1234 /**
1235  * Configure secondary process queues from a private data pointer (primary
1236  * or secondary) and update burst callbacks. Can take place only once.
1237  *
1238  * All queues must have been previously created by the primary process to
1239  * avoid undefined behavior.
1240  *
1241  * @param priv
1242  *   Private data pointer from either primary or secondary process.
1243  *
1244  * @return
1245  *   Private data pointer from secondary process, NULL in case of error.
1246  */
1247 struct priv *
1248 mlx5_secondary_data_setup(struct priv *priv)
1249 {
1250         unsigned int port_id = 0;
1251         struct mlx5_secondary_data *sd;
1252         void **tx_queues;
1253         void **rx_queues;
1254         unsigned int nb_tx_queues;
1255         unsigned int nb_rx_queues;
1256         unsigned int i;
1257
1258         /* priv must be valid at this point. */
1259         assert(priv != NULL);
1260         /* priv->dev must also be valid but may point to local memory from
1261          * another process, possibly with the same address and must not
1262          * be dereferenced yet. */
1263         assert(priv->dev != NULL);
1264         /* Determine port ID by finding out where priv comes from. */
1265         while (1) {
1266                 sd = &mlx5_secondary_data[port_id];
1267                 rte_spinlock_lock(&sd->lock);
1268                 /* Primary process? */
1269                 if (sd->primary_priv == priv)
1270                         break;
1271                 /* Secondary process? */
1272                 if (sd->data.dev_private == priv)
1273                         break;
1274                 rte_spinlock_unlock(&sd->lock);
1275                 if (++port_id == RTE_DIM(mlx5_secondary_data))
1276                         port_id = 0;
1277         }
1278         /* Switch to secondary private structure. If private data has already
1279          * been updated by another thread, there is nothing else to do. */
1280         priv = sd->data.dev_private;
1281         if (priv->dev->data == &sd->data)
1282                 goto end;
1283         /* Sanity checks. Secondary private structure is supposed to point
1284          * to local eth_dev, itself still pointing to the shared device data
1285          * structure allocated by the primary process. */
1286         assert(sd->shared_dev_data != &sd->data);
1287         assert(sd->data.nb_tx_queues == 0);
1288         assert(sd->data.tx_queues == NULL);
1289         assert(sd->data.nb_rx_queues == 0);
1290         assert(sd->data.rx_queues == NULL);
1291         assert(priv != sd->primary_priv);
1292         assert(priv->dev->data == sd->shared_dev_data);
1293         assert(priv->txqs_n == 0);
1294         assert(priv->txqs == NULL);
1295         assert(priv->rxqs_n == 0);
1296         assert(priv->rxqs == NULL);
1297         nb_tx_queues = sd->shared_dev_data->nb_tx_queues;
1298         nb_rx_queues = sd->shared_dev_data->nb_rx_queues;
1299         /* Allocate local storage for queues. */
1300         tx_queues = rte_zmalloc("secondary ethdev->tx_queues",
1301                                 sizeof(sd->data.tx_queues[0]) * nb_tx_queues,
1302                                 RTE_CACHE_LINE_SIZE);
1303         rx_queues = rte_zmalloc("secondary ethdev->rx_queues",
1304                                 sizeof(sd->data.rx_queues[0]) * nb_rx_queues,
1305                                 RTE_CACHE_LINE_SIZE);
1306         if (tx_queues == NULL || rx_queues == NULL)
1307                 goto error;
1308         /* Lock to prevent control operations during setup. */
1309         priv_lock(priv);
1310         /* TX queues. */
1311         for (i = 0; i != nb_tx_queues; ++i) {
1312                 struct txq *primary_txq = (*sd->primary_priv->txqs)[i];
1313                 struct txq_ctrl *primary_txq_ctrl;
1314                 struct txq_ctrl *txq_ctrl;
1315
1316                 if (primary_txq == NULL)
1317                         continue;
1318                 primary_txq_ctrl = container_of(primary_txq,
1319                                                 struct txq_ctrl, txq);
1320                 txq_ctrl = rte_calloc_socket("TXQ", 1, sizeof(*txq_ctrl) +
1321                                              (1 << primary_txq->elts_n) *
1322                                              sizeof(struct rte_mbuf *), 0,
1323                                              primary_txq_ctrl->socket);
1324                 if (txq_ctrl != NULL) {
1325                         if (txq_ctrl_setup(priv->dev,
1326                                            txq_ctrl,
1327                                            1 << primary_txq->elts_n,
1328                                            primary_txq_ctrl->socket,
1329                                            NULL) == 0) {
1330                                 txq_ctrl->txq.stats.idx =
1331                                         primary_txq->stats.idx;
1332                                 tx_queues[i] = &txq_ctrl->txq;
1333                                 continue;
1334                         }
1335                         rte_free(txq_ctrl);
1336                 }
1337                 while (i) {
1338                         txq_ctrl = tx_queues[--i];
1339                         txq_cleanup(txq_ctrl);
1340                         rte_free(txq_ctrl);
1341                 }
1342                 goto error;
1343         }
1344         /* RX queues. */
1345         for (i = 0; i != nb_rx_queues; ++i) {
1346                 struct rxq_ctrl *primary_rxq =
1347                         container_of((*sd->primary_priv->rxqs)[i],
1348                                      struct rxq_ctrl, rxq);
1349
1350                 if (primary_rxq == NULL)
1351                         continue;
1352                 /* Not supported yet. */
1353                 rx_queues[i] = NULL;
1354         }
1355         /* Update everything. */
1356         priv->txqs = (void *)tx_queues;
1357         priv->txqs_n = nb_tx_queues;
1358         priv->rxqs = (void *)rx_queues;
1359         priv->rxqs_n = nb_rx_queues;
1360         sd->data.rx_queues = rx_queues;
1361         sd->data.tx_queues = tx_queues;
1362         sd->data.nb_rx_queues = nb_rx_queues;
1363         sd->data.nb_tx_queues = nb_tx_queues;
1364         sd->data.dev_link = sd->shared_dev_data->dev_link;
1365         sd->data.mtu = sd->shared_dev_data->mtu;
1366         memcpy(sd->data.rx_queue_state, sd->shared_dev_data->rx_queue_state,
1367                sizeof(sd->data.rx_queue_state));
1368         memcpy(sd->data.tx_queue_state, sd->shared_dev_data->tx_queue_state,
1369                sizeof(sd->data.tx_queue_state));
1370         sd->data.dev_flags = sd->shared_dev_data->dev_flags;
1371         /* Use local data from now on. */
1372         rte_mb();
1373         priv->dev->data = &sd->data;
1374         rte_mb();
1375         priv_select_tx_function(priv);
1376         priv_select_rx_function(priv);
1377         priv_unlock(priv);
1378 end:
1379         /* More sanity checks. */
1380         assert(priv->dev->data == &sd->data);
1381         rte_spinlock_unlock(&sd->lock);
1382         return priv;
1383 error:
1384         priv_unlock(priv);
1385         rte_free(tx_queues);
1386         rte_free(rx_queues);
1387         rte_spinlock_unlock(&sd->lock);
1388         return NULL;
1389 }
1390
1391 /**
1392  * Configure the TX function to use.
1393  *
1394  * @param priv
1395  *   Pointer to private structure.
1396  */
1397 void
1398 priv_select_tx_function(struct priv *priv)
1399 {
1400         priv->dev->tx_pkt_burst = mlx5_tx_burst;
1401         /* Display warning for unsupported configurations. */
1402         if (priv->sriov && priv->mps)
1403                 WARN("multi-packet send WQE cannot be used on a SR-IOV setup");
1404         /* Select appropriate TX function. */
1405         if ((priv->sriov == 0) && priv->mps && priv->txq_inline) {
1406                 priv->dev->tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1407                 DEBUG("selected MPW inline TX function");
1408         } else if ((priv->sriov == 0) && priv->mps) {
1409                 priv->dev->tx_pkt_burst = mlx5_tx_burst_mpw;
1410                 DEBUG("selected MPW TX function");
1411         }
1412 }
1413
1414 /**
1415  * Configure the RX function to use.
1416  *
1417  * @param priv
1418  *   Pointer to private structure.
1419  */
1420 void
1421 priv_select_rx_function(struct priv *priv)
1422 {
1423         priv->dev->rx_pkt_burst = mlx5_rx_burst;
1424 }