da1a3ca8150031267ef43dd39476815e8e649545
[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         info->pci_dev = dev->pci_dev;
566
567         priv_lock(priv);
568         /* FIXME: we should ask the device for these values. */
569         info->min_rx_bufsize = 32;
570         info->max_rx_pktlen = 65536;
571         /*
572          * Since we need one CQ per QP, the limit is the minimum number
573          * between the two values.
574          */
575         max = ((priv->device_attr.max_cq > priv->device_attr.max_qp) ?
576                priv->device_attr.max_qp : priv->device_attr.max_cq);
577         /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
578         if (max >= 65535)
579                 max = 65535;
580         info->max_rx_queues = max;
581         info->max_tx_queues = max;
582         info->max_mac_addrs = RTE_DIM(priv->mac);
583         info->rx_offload_capa =
584                 (priv->hw_csum ?
585                  (DEV_RX_OFFLOAD_IPV4_CKSUM |
586                   DEV_RX_OFFLOAD_UDP_CKSUM |
587                   DEV_RX_OFFLOAD_TCP_CKSUM) :
588                  0) |
589                 (priv->hw_vlan_strip ? DEV_RX_OFFLOAD_VLAN_STRIP : 0);
590         if (!priv->mps)
591                 info->tx_offload_capa = DEV_TX_OFFLOAD_VLAN_INSERT;
592         if (priv->hw_csum)
593                 info->tx_offload_capa |=
594                         (DEV_TX_OFFLOAD_IPV4_CKSUM |
595                          DEV_TX_OFFLOAD_UDP_CKSUM |
596                          DEV_TX_OFFLOAD_TCP_CKSUM);
597         if (priv_get_ifname(priv, &ifname) == 0)
598                 info->if_index = if_nametoindex(ifname);
599         /* FIXME: RETA update/query API expects the callee to know the size of
600          * the indirection table, for this PMD the size varies depending on
601          * the number of RX queues, it becomes impossible to find the correct
602          * size if it is not fixed.
603          * The API should be updated to solve this problem. */
604         info->reta_size = priv->ind_table_max_size;
605         info->hash_key_size = ((*priv->rss_conf) ?
606                                (*priv->rss_conf)[0]->rss_key_len :
607                                0);
608         info->speed_capa = priv->link_speed_capa;
609         priv_unlock(priv);
610 }
611
612 const uint32_t *
613 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
614 {
615         static const uint32_t ptypes[] = {
616                 /* refers to rxq_cq_to_pkt_type() */
617                 RTE_PTYPE_L3_IPV4,
618                 RTE_PTYPE_L3_IPV6,
619                 RTE_PTYPE_INNER_L3_IPV4,
620                 RTE_PTYPE_INNER_L3_IPV6,
621                 RTE_PTYPE_UNKNOWN
622
623         };
624
625         if (dev->rx_pkt_burst == mlx5_rx_burst)
626                 return ptypes;
627         return NULL;
628 }
629
630 /**
631  * Retrieve physical link information (unlocked version using legacy ioctl).
632  *
633  * @param dev
634  *   Pointer to Ethernet device structure.
635  * @param wait_to_complete
636  *   Wait for request completion (ignored).
637  */
638 static int
639 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev, int wait_to_complete)
640 {
641         struct priv *priv = mlx5_get_priv(dev);
642         struct ethtool_cmd edata = {
643                 .cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
644         };
645         struct ifreq ifr;
646         struct rte_eth_link dev_link;
647         int link_speed = 0;
648
649         (void)wait_to_complete;
650         if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
651                 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
652                 return -1;
653         }
654         memset(&dev_link, 0, sizeof(dev_link));
655         dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
656                                 (ifr.ifr_flags & IFF_RUNNING));
657         ifr.ifr_data = (void *)&edata;
658         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
659                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
660                      strerror(errno));
661                 return -1;
662         }
663         link_speed = ethtool_cmd_speed(&edata);
664         if (link_speed == -1)
665                 dev_link.link_speed = 0;
666         else
667                 dev_link.link_speed = link_speed;
668         priv->link_speed_capa = 0;
669         if (edata.supported & SUPPORTED_Autoneg)
670                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
671         if (edata.supported & (SUPPORTED_1000baseT_Full |
672                                SUPPORTED_1000baseKX_Full))
673                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
674         if (edata.supported & SUPPORTED_10000baseKR_Full)
675                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
676         if (edata.supported & (SUPPORTED_40000baseKR4_Full |
677                                SUPPORTED_40000baseCR4_Full |
678                                SUPPORTED_40000baseSR4_Full |
679                                SUPPORTED_40000baseLR4_Full))
680                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
681         dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
682                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
683         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
684                         ETH_LINK_SPEED_FIXED);
685         if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
686                 /* Link status changed. */
687                 dev->data->dev_link = dev_link;
688                 return 0;
689         }
690         /* Link status is still the same. */
691         return -1;
692 }
693
694 /**
695  * Retrieve physical link information (unlocked version using new ioctl from
696  * Linux 4.5).
697  *
698  * @param dev
699  *   Pointer to Ethernet device structure.
700  * @param wait_to_complete
701  *   Wait for request completion (ignored).
702  */
703 static int
704 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev, int wait_to_complete)
705 {
706 #ifdef ETHTOOL_GLINKSETTINGS
707         struct priv *priv = mlx5_get_priv(dev);
708         struct ethtool_link_settings edata = {
709                 .cmd = ETHTOOL_GLINKSETTINGS,
710         };
711         struct ifreq ifr;
712         struct rte_eth_link dev_link;
713         uint64_t sc;
714
715         (void)wait_to_complete;
716         if (priv_ifreq(priv, SIOCGIFFLAGS, &ifr)) {
717                 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
718                 return -1;
719         }
720         memset(&dev_link, 0, sizeof(dev_link));
721         dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
722                                 (ifr.ifr_flags & IFF_RUNNING));
723         ifr.ifr_data = (void *)&edata;
724         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
725                 DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
726                       strerror(errno));
727                 return -1;
728         }
729         dev_link.link_speed = edata.speed;
730         sc = edata.link_mode_masks[0] |
731                 ((uint64_t)edata.link_mode_masks[1] << 32);
732         priv->link_speed_capa = 0;
733         /* Link speeds available in kernel v4.5. */
734         if (sc & ETHTOOL_LINK_MODE_Autoneg_BIT)
735                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
736         if (sc & (ETHTOOL_LINK_MODE_1000baseT_Full_BIT |
737                   ETHTOOL_LINK_MODE_1000baseKX_Full_BIT))
738                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
739         if (sc & (ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT |
740                   ETHTOOL_LINK_MODE_10000baseKR_Full_BIT |
741                   ETHTOOL_LINK_MODE_10000baseR_FEC_BIT))
742                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
743         if (sc & (ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT |
744                   ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT))
745                 priv->link_speed_capa |= ETH_LINK_SPEED_20G;
746         if (sc & (ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT |
747                   ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT |
748                   ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT |
749                   ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT))
750                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
751         if (sc & (ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT |
752                   ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT |
753                   ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT |
754                   ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT))
755                 priv->link_speed_capa |= ETH_LINK_SPEED_56G;
756         /* Link speeds available in kernel v4.6. */
757 #ifdef HAVE_ETHTOOL_LINK_MODE_25G
758         if (sc & (ETHTOOL_LINK_MODE_25000baseCR_Full_BIT |
759                   ETHTOOL_LINK_MODE_25000baseKR_Full_BIT |
760                   ETHTOOL_LINK_MODE_25000baseSR_Full_BIT))
761                 priv->link_speed_capa |= ETH_LINK_SPEED_25G;
762 #endif
763 #ifdef HAVE_ETHTOOL_LINK_MODE_50G
764         if (sc & (ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT |
765                   ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT))
766                 priv->link_speed_capa |= ETH_LINK_SPEED_50G;
767 #endif
768 #ifdef HAVE_ETHTOOL_LINK_MODE_100G
769         if (sc & (ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT |
770                   ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT |
771                   ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT |
772                   ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT))
773                 priv->link_speed_capa |= ETH_LINK_SPEED_100G;
774 #endif
775         dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
776                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
777         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
778                                   ETH_LINK_SPEED_FIXED);
779         if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
780                 /* Link status changed. */
781                 dev->data->dev_link = dev_link;
782                 return 0;
783         }
784 #else
785         (void)dev;
786         (void)wait_to_complete;
787 #endif
788         /* Link status is still the same. */
789         return -1;
790 }
791
792 /**
793  * DPDK callback to retrieve physical link information (unlocked version).
794  *
795  * @param dev
796  *   Pointer to Ethernet device structure.
797  * @param wait_to_complete
798  *   Wait for request completion (ignored).
799  */
800 int
801 mlx5_link_update_unlocked(struct rte_eth_dev *dev, int wait_to_complete)
802 {
803         int ret;
804
805         ret = mlx5_link_update_unlocked_gs(dev, wait_to_complete);
806         if (ret < 0)
807                 ret = mlx5_link_update_unlocked_gset(dev, wait_to_complete);
808         return ret;
809 }
810
811 /**
812  * DPDK callback to retrieve physical link information.
813  *
814  * @param dev
815  *   Pointer to Ethernet device structure.
816  * @param wait_to_complete
817  *   Wait for request completion (ignored).
818  */
819 int
820 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete)
821 {
822         struct priv *priv = mlx5_get_priv(dev);
823         int ret;
824
825         priv_lock(priv);
826         ret = mlx5_link_update_unlocked(dev, wait_to_complete);
827         priv_unlock(priv);
828         return ret;
829 }
830
831 /**
832  * DPDK callback to change the MTU.
833  *
834  * Setting the MTU affects hardware MRU (packets larger than the MTU cannot be
835  * received). Use this as a hint to enable/disable scattered packets support
836  * and improve performance when not needed.
837  * Since failure is not an option, reconfiguring queues on the fly is not
838  * recommended.
839  *
840  * @param dev
841  *   Pointer to Ethernet device structure.
842  * @param in_mtu
843  *   New MTU.
844  *
845  * @return
846  *   0 on success, negative errno value on failure.
847  */
848 int
849 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
850 {
851         struct priv *priv = dev->data->dev_private;
852         int ret = 0;
853         unsigned int i;
854         uint16_t (*rx_func)(void *, struct rte_mbuf **, uint16_t) =
855                 mlx5_rx_burst;
856         unsigned int max_frame_len;
857         int rehash;
858         int restart = priv->started;
859
860         if (mlx5_is_secondary())
861                 return -E_RTE_SECONDARY;
862
863         priv_lock(priv);
864         /* Set kernel interface MTU first. */
865         if (priv_set_mtu(priv, mtu)) {
866                 ret = errno;
867                 WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
868                      strerror(ret));
869                 goto out;
870         } else
871                 DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
872         /* Temporarily replace RX handler with a fake one, assuming it has not
873          * been copied elsewhere. */
874         dev->rx_pkt_burst = removed_rx_burst;
875         /* Make sure everyone has left mlx5_rx_burst() and uses
876          * removed_rx_burst() instead. */
877         rte_wmb();
878         usleep(1000);
879         /* MTU does not include header and CRC. */
880         max_frame_len = ETHER_HDR_LEN + mtu + ETHER_CRC_LEN;
881         /* Check if at least one queue is going to need a SGE update. */
882         for (i = 0; i != priv->rxqs_n; ++i) {
883                 struct rxq *rxq = (*priv->rxqs)[i];
884                 unsigned int mb_len;
885                 unsigned int size = RTE_PKTMBUF_HEADROOM + max_frame_len;
886                 unsigned int sges_n;
887
888                 if (rxq == NULL)
889                         continue;
890                 mb_len = rte_pktmbuf_data_room_size(rxq->mp);
891                 assert(mb_len >= RTE_PKTMBUF_HEADROOM);
892                 /*
893                  * Determine the number of SGEs needed for a full packet
894                  * and round it to the next power of two.
895                  */
896                 sges_n = log2above((size / mb_len) + !!(size % mb_len));
897                 if (sges_n != rxq->sges_n)
898                         break;
899         }
900         /*
901          * If all queues have the right number of SGEs, a simple rehash
902          * of their buffers is enough, otherwise SGE information can only
903          * be updated in a queue by recreating it. All resources that depend
904          * on queues (flows, indirection tables) must be recreated as well in
905          * that case.
906          */
907         rehash = (i == priv->rxqs_n);
908         if (!rehash) {
909                 /* Clean up everything as with mlx5_dev_stop(). */
910                 priv_special_flow_disable_all(priv);
911                 priv_mac_addrs_disable(priv);
912                 priv_destroy_hash_rxqs(priv);
913                 priv_fdir_disable(priv);
914                 priv_dev_interrupt_handler_uninstall(priv, dev);
915         }
916 recover:
917         /* Reconfigure each RX queue. */
918         for (i = 0; (i != priv->rxqs_n); ++i) {
919                 struct rxq *rxq = (*priv->rxqs)[i];
920                 struct rxq_ctrl *rxq_ctrl =
921                         container_of(rxq, struct rxq_ctrl, rxq);
922                 int sp;
923                 unsigned int mb_len;
924                 unsigned int tmp;
925
926                 if (rxq == NULL)
927                         continue;
928                 mb_len = rte_pktmbuf_data_room_size(rxq->mp);
929                 assert(mb_len >= RTE_PKTMBUF_HEADROOM);
930                 /* Toggle scattered support (sp) if necessary. */
931                 sp = (max_frame_len > (mb_len - RTE_PKTMBUF_HEADROOM));
932                 /* Provide new values to rxq_setup(). */
933                 dev->data->dev_conf.rxmode.jumbo_frame = sp;
934                 dev->data->dev_conf.rxmode.max_rx_pkt_len = max_frame_len;
935                 if (rehash)
936                         ret = rxq_rehash(dev, rxq_ctrl);
937                 else
938                         ret = rxq_ctrl_setup(dev, rxq_ctrl, 1 << rxq->elts_n,
939                                              rxq_ctrl->socket, NULL, rxq->mp);
940                 if (!ret)
941                         continue;
942                 /* Attempt to roll back in case of error. */
943                 tmp = (mb_len << rxq->sges_n) - RTE_PKTMBUF_HEADROOM;
944                 if (max_frame_len != tmp) {
945                         max_frame_len = tmp;
946                         goto recover;
947                 }
948                 /* Double fault, disable RX. */
949                 break;
950         }
951         /*
952          * Use a safe RX burst function in case of error, otherwise mimic
953          * mlx5_dev_start().
954          */
955         if (ret) {
956                 ERROR("unable to reconfigure RX queues, RX disabled");
957                 rx_func = removed_rx_burst;
958         } else if (restart &&
959                  !rehash &&
960                  !priv_create_hash_rxqs(priv) &&
961                  !priv_rehash_flows(priv)) {
962                 if (dev->data->dev_conf.fdir_conf.mode == RTE_FDIR_MODE_NONE)
963                         priv_fdir_enable(priv);
964                 priv_dev_interrupt_handler_install(priv, dev);
965         }
966         priv->mtu = mtu;
967         /* Burst functions can now be called again. */
968         rte_wmb();
969         dev->rx_pkt_burst = rx_func;
970 out:
971         priv_unlock(priv);
972         assert(ret >= 0);
973         return -ret;
974 }
975
976 /**
977  * DPDK callback to get flow control status.
978  *
979  * @param dev
980  *   Pointer to Ethernet device structure.
981  * @param[out] fc_conf
982  *   Flow control output buffer.
983  *
984  * @return
985  *   0 on success, negative errno value on failure.
986  */
987 int
988 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
989 {
990         struct priv *priv = dev->data->dev_private;
991         struct ifreq ifr;
992         struct ethtool_pauseparam ethpause = {
993                 .cmd = ETHTOOL_GPAUSEPARAM
994         };
995         int ret;
996
997         if (mlx5_is_secondary())
998                 return -E_RTE_SECONDARY;
999
1000         ifr.ifr_data = (void *)&ethpause;
1001         priv_lock(priv);
1002         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
1003                 ret = errno;
1004                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM)"
1005                      " failed: %s",
1006                      strerror(ret));
1007                 goto out;
1008         }
1009
1010         fc_conf->autoneg = ethpause.autoneg;
1011         if (ethpause.rx_pause && ethpause.tx_pause)
1012                 fc_conf->mode = RTE_FC_FULL;
1013         else if (ethpause.rx_pause)
1014                 fc_conf->mode = RTE_FC_RX_PAUSE;
1015         else if (ethpause.tx_pause)
1016                 fc_conf->mode = RTE_FC_TX_PAUSE;
1017         else
1018                 fc_conf->mode = RTE_FC_NONE;
1019         ret = 0;
1020
1021 out:
1022         priv_unlock(priv);
1023         assert(ret >= 0);
1024         return -ret;
1025 }
1026
1027 /**
1028  * DPDK callback to modify flow control parameters.
1029  *
1030  * @param dev
1031  *   Pointer to Ethernet device structure.
1032  * @param[in] fc_conf
1033  *   Flow control parameters.
1034  *
1035  * @return
1036  *   0 on success, negative errno value on failure.
1037  */
1038 int
1039 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
1040 {
1041         struct priv *priv = dev->data->dev_private;
1042         struct ifreq ifr;
1043         struct ethtool_pauseparam ethpause = {
1044                 .cmd = ETHTOOL_SPAUSEPARAM
1045         };
1046         int ret;
1047
1048         if (mlx5_is_secondary())
1049                 return -E_RTE_SECONDARY;
1050
1051         ifr.ifr_data = (void *)&ethpause;
1052         ethpause.autoneg = fc_conf->autoneg;
1053         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1054             (fc_conf->mode & RTE_FC_RX_PAUSE))
1055                 ethpause.rx_pause = 1;
1056         else
1057                 ethpause.rx_pause = 0;
1058
1059         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
1060             (fc_conf->mode & RTE_FC_TX_PAUSE))
1061                 ethpause.tx_pause = 1;
1062         else
1063                 ethpause.tx_pause = 0;
1064
1065         priv_lock(priv);
1066         if (priv_ifreq(priv, SIOCETHTOOL, &ifr)) {
1067                 ret = errno;
1068                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
1069                      " failed: %s",
1070                      strerror(ret));
1071                 goto out;
1072         }
1073         ret = 0;
1074
1075 out:
1076         priv_unlock(priv);
1077         assert(ret >= 0);
1078         return -ret;
1079 }
1080
1081 /**
1082  * Get PCI information from struct ibv_device.
1083  *
1084  * @param device
1085  *   Pointer to Ethernet device structure.
1086  * @param[out] pci_addr
1087  *   PCI bus address output buffer.
1088  *
1089  * @return
1090  *   0 on success, -1 on failure and errno is set.
1091  */
1092 int
1093 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
1094                             struct rte_pci_addr *pci_addr)
1095 {
1096         FILE *file;
1097         char line[32];
1098         MKSTR(path, "%s/device/uevent", device->ibdev_path);
1099
1100         file = fopen(path, "rb");
1101         if (file == NULL)
1102                 return -1;
1103         while (fgets(line, sizeof(line), file) == line) {
1104                 size_t len = strlen(line);
1105                 int ret;
1106
1107                 /* Truncate long lines. */
1108                 if (len == (sizeof(line) - 1))
1109                         while (line[(len - 1)] != '\n') {
1110                                 ret = fgetc(file);
1111                                 if (ret == EOF)
1112                                         break;
1113                                 line[(len - 1)] = ret;
1114                         }
1115                 /* Extract information. */
1116                 if (sscanf(line,
1117                            "PCI_SLOT_NAME="
1118                            "%" SCNx16 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
1119                            &pci_addr->domain,
1120                            &pci_addr->bus,
1121                            &pci_addr->devid,
1122                            &pci_addr->function) == 4) {
1123                         ret = 0;
1124                         break;
1125                 }
1126         }
1127         fclose(file);
1128         return 0;
1129 }
1130
1131 /**
1132  * Link status handler.
1133  *
1134  * @param priv
1135  *   Pointer to private structure.
1136  * @param dev
1137  *   Pointer to the rte_eth_dev structure.
1138  *
1139  * @return
1140  *   Nonzero if the callback process can be called immediately.
1141  */
1142 static int
1143 priv_dev_link_status_handler(struct priv *priv, struct rte_eth_dev *dev)
1144 {
1145         struct ibv_async_event event;
1146         int port_change = 0;
1147         int ret = 0;
1148
1149         /* Read all message and acknowledge them. */
1150         for (;;) {
1151                 if (ibv_get_async_event(priv->ctx, &event))
1152                         break;
1153
1154                 if (event.event_type == IBV_EVENT_PORT_ACTIVE ||
1155                     event.event_type == IBV_EVENT_PORT_ERR)
1156                         port_change = 1;
1157                 else
1158                         DEBUG("event type %d on port %d not handled",
1159                               event.event_type, event.element.port_num);
1160                 ibv_ack_async_event(&event);
1161         }
1162
1163         if (port_change ^ priv->pending_alarm) {
1164                 struct rte_eth_link *link = &dev->data->dev_link;
1165
1166                 priv->pending_alarm = 0;
1167                 mlx5_link_update_unlocked(dev, 0);
1168                 if (((link->link_speed == 0) && link->link_status) ||
1169                     ((link->link_speed != 0) && !link->link_status)) {
1170                         /* Inconsistent status, check again later. */
1171                         priv->pending_alarm = 1;
1172                         rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
1173                                           mlx5_dev_link_status_handler,
1174                                           dev);
1175                 } else
1176                         ret = 1;
1177         }
1178         return ret;
1179 }
1180
1181 /**
1182  * Handle delayed link status event.
1183  *
1184  * @param arg
1185  *   Registered argument.
1186  */
1187 void
1188 mlx5_dev_link_status_handler(void *arg)
1189 {
1190         struct rte_eth_dev *dev = arg;
1191         struct priv *priv = dev->data->dev_private;
1192         int ret;
1193
1194         priv_lock(priv);
1195         assert(priv->pending_alarm == 1);
1196         ret = priv_dev_link_status_handler(priv, dev);
1197         priv_unlock(priv);
1198         if (ret)
1199                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1200 }
1201
1202 /**
1203  * Handle interrupts from the NIC.
1204  *
1205  * @param[in] intr_handle
1206  *   Interrupt handler.
1207  * @param cb_arg
1208  *   Callback argument.
1209  */
1210 void
1211 mlx5_dev_interrupt_handler(struct rte_intr_handle *intr_handle, void *cb_arg)
1212 {
1213         struct rte_eth_dev *dev = cb_arg;
1214         struct priv *priv = dev->data->dev_private;
1215         int ret;
1216
1217         (void)intr_handle;
1218         priv_lock(priv);
1219         ret = priv_dev_link_status_handler(priv, dev);
1220         priv_unlock(priv);
1221         if (ret)
1222                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1223 }
1224
1225 /**
1226  * Uninstall interrupt handler.
1227  *
1228  * @param priv
1229  *   Pointer to private structure.
1230  * @param dev
1231  *   Pointer to the rte_eth_dev structure.
1232  */
1233 void
1234 priv_dev_interrupt_handler_uninstall(struct priv *priv, struct rte_eth_dev *dev)
1235 {
1236         if (!dev->data->dev_conf.intr_conf.lsc)
1237                 return;
1238         rte_intr_callback_unregister(&priv->intr_handle,
1239                                      mlx5_dev_interrupt_handler,
1240                                      dev);
1241         if (priv->pending_alarm)
1242                 rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev);
1243         priv->pending_alarm = 0;
1244         priv->intr_handle.fd = 0;
1245         priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1246 }
1247
1248 /**
1249  * Install interrupt handler.
1250  *
1251  * @param priv
1252  *   Pointer to private structure.
1253  * @param dev
1254  *   Pointer to the rte_eth_dev structure.
1255  */
1256 void
1257 priv_dev_interrupt_handler_install(struct priv *priv, struct rte_eth_dev *dev)
1258 {
1259         int rc, flags;
1260
1261         if (!dev->data->dev_conf.intr_conf.lsc)
1262                 return;
1263         assert(priv->ctx->async_fd > 0);
1264         flags = fcntl(priv->ctx->async_fd, F_GETFL);
1265         rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1266         if (rc < 0) {
1267                 INFO("failed to change file descriptor async event queue");
1268                 dev->data->dev_conf.intr_conf.lsc = 0;
1269         } else {
1270                 priv->intr_handle.fd = priv->ctx->async_fd;
1271                 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1272                 rte_intr_callback_register(&priv->intr_handle,
1273                                            mlx5_dev_interrupt_handler,
1274                                            dev);
1275         }
1276 }
1277
1278 /**
1279  * Change the link state (UP / DOWN).
1280  *
1281  * @param priv
1282  *   Pointer to Ethernet device structure.
1283  * @param up
1284  *   Nonzero for link up, otherwise link down.
1285  *
1286  * @return
1287  *   0 on success, errno value on failure.
1288  */
1289 static int
1290 priv_set_link(struct priv *priv, int up)
1291 {
1292         struct rte_eth_dev *dev = priv->dev;
1293         int err;
1294
1295         if (up) {
1296                 err = priv_set_flags(priv, ~IFF_UP, IFF_UP);
1297                 if (err)
1298                         return err;
1299                 priv_select_tx_function(priv);
1300                 priv_select_rx_function(priv);
1301         } else {
1302                 err = priv_set_flags(priv, ~IFF_UP, ~IFF_UP);
1303                 if (err)
1304                         return err;
1305                 dev->rx_pkt_burst = removed_rx_burst;
1306                 dev->tx_pkt_burst = removed_tx_burst;
1307         }
1308         return 0;
1309 }
1310
1311 /**
1312  * DPDK callback to bring the link DOWN.
1313  *
1314  * @param dev
1315  *   Pointer to Ethernet device structure.
1316  *
1317  * @return
1318  *   0 on success, errno value on failure.
1319  */
1320 int
1321 mlx5_set_link_down(struct rte_eth_dev *dev)
1322 {
1323         struct priv *priv = dev->data->dev_private;
1324         int err;
1325
1326         priv_lock(priv);
1327         err = priv_set_link(priv, 0);
1328         priv_unlock(priv);
1329         return err;
1330 }
1331
1332 /**
1333  * DPDK callback to bring the link UP.
1334  *
1335  * @param dev
1336  *   Pointer to Ethernet device structure.
1337  *
1338  * @return
1339  *   0 on success, errno value on failure.
1340  */
1341 int
1342 mlx5_set_link_up(struct rte_eth_dev *dev)
1343 {
1344         struct priv *priv = dev->data->dev_private;
1345         int err;
1346
1347         priv_lock(priv);
1348         err = priv_set_link(priv, 1);
1349         priv_unlock(priv);
1350         return err;
1351 }
1352
1353 /**
1354  * Configure secondary process queues from a private data pointer (primary
1355  * or secondary) and update burst callbacks. Can take place only once.
1356  *
1357  * All queues must have been previously created by the primary process to
1358  * avoid undefined behavior.
1359  *
1360  * @param priv
1361  *   Private data pointer from either primary or secondary process.
1362  *
1363  * @return
1364  *   Private data pointer from secondary process, NULL in case of error.
1365  */
1366 struct priv *
1367 mlx5_secondary_data_setup(struct priv *priv)
1368 {
1369         unsigned int port_id = 0;
1370         struct mlx5_secondary_data *sd;
1371         void **tx_queues;
1372         void **rx_queues;
1373         unsigned int nb_tx_queues;
1374         unsigned int nb_rx_queues;
1375         unsigned int i;
1376
1377         /* priv must be valid at this point. */
1378         assert(priv != NULL);
1379         /* priv->dev must also be valid but may point to local memory from
1380          * another process, possibly with the same address and must not
1381          * be dereferenced yet. */
1382         assert(priv->dev != NULL);
1383         /* Determine port ID by finding out where priv comes from. */
1384         while (1) {
1385                 sd = &mlx5_secondary_data[port_id];
1386                 rte_spinlock_lock(&sd->lock);
1387                 /* Primary process? */
1388                 if (sd->primary_priv == priv)
1389                         break;
1390                 /* Secondary process? */
1391                 if (sd->data.dev_private == priv)
1392                         break;
1393                 rte_spinlock_unlock(&sd->lock);
1394                 if (++port_id == RTE_DIM(mlx5_secondary_data))
1395                         port_id = 0;
1396         }
1397         /* Switch to secondary private structure. If private data has already
1398          * been updated by another thread, there is nothing else to do. */
1399         priv = sd->data.dev_private;
1400         if (priv->dev->data == &sd->data)
1401                 goto end;
1402         /* Sanity checks. Secondary private structure is supposed to point
1403          * to local eth_dev, itself still pointing to the shared device data
1404          * structure allocated by the primary process. */
1405         assert(sd->shared_dev_data != &sd->data);
1406         assert(sd->data.nb_tx_queues == 0);
1407         assert(sd->data.tx_queues == NULL);
1408         assert(sd->data.nb_rx_queues == 0);
1409         assert(sd->data.rx_queues == NULL);
1410         assert(priv != sd->primary_priv);
1411         assert(priv->dev->data == sd->shared_dev_data);
1412         assert(priv->txqs_n == 0);
1413         assert(priv->txqs == NULL);
1414         assert(priv->rxqs_n == 0);
1415         assert(priv->rxqs == NULL);
1416         nb_tx_queues = sd->shared_dev_data->nb_tx_queues;
1417         nb_rx_queues = sd->shared_dev_data->nb_rx_queues;
1418         /* Allocate local storage for queues. */
1419         tx_queues = rte_zmalloc("secondary ethdev->tx_queues",
1420                                 sizeof(sd->data.tx_queues[0]) * nb_tx_queues,
1421                                 RTE_CACHE_LINE_SIZE);
1422         rx_queues = rte_zmalloc("secondary ethdev->rx_queues",
1423                                 sizeof(sd->data.rx_queues[0]) * nb_rx_queues,
1424                                 RTE_CACHE_LINE_SIZE);
1425         if (tx_queues == NULL || rx_queues == NULL)
1426                 goto error;
1427         /* Lock to prevent control operations during setup. */
1428         priv_lock(priv);
1429         /* TX queues. */
1430         for (i = 0; i != nb_tx_queues; ++i) {
1431                 struct txq *primary_txq = (*sd->primary_priv->txqs)[i];
1432                 struct txq_ctrl *primary_txq_ctrl;
1433                 struct txq_ctrl *txq_ctrl;
1434
1435                 if (primary_txq == NULL)
1436                         continue;
1437                 primary_txq_ctrl = container_of(primary_txq,
1438                                                 struct txq_ctrl, txq);
1439                 txq_ctrl = rte_calloc_socket("TXQ", 1, sizeof(*txq_ctrl) +
1440                                              (1 << primary_txq->elts_n) *
1441                                              sizeof(struct rte_mbuf *), 0,
1442                                              primary_txq_ctrl->socket);
1443                 if (txq_ctrl != NULL) {
1444                         if (txq_ctrl_setup(priv->dev,
1445                                            txq_ctrl,
1446                                            1 << primary_txq->elts_n,
1447                                            primary_txq_ctrl->socket,
1448                                            NULL) == 0) {
1449                                 txq_ctrl->txq.stats.idx =
1450                                         primary_txq->stats.idx;
1451                                 tx_queues[i] = &txq_ctrl->txq;
1452                                 continue;
1453                         }
1454                         rte_free(txq_ctrl);
1455                 }
1456                 while (i) {
1457                         txq_ctrl = tx_queues[--i];
1458                         txq_cleanup(txq_ctrl);
1459                         rte_free(txq_ctrl);
1460                 }
1461                 goto error;
1462         }
1463         /* RX queues. */
1464         for (i = 0; i != nb_rx_queues; ++i) {
1465                 struct rxq_ctrl *primary_rxq =
1466                         container_of((*sd->primary_priv->rxqs)[i],
1467                                      struct rxq_ctrl, rxq);
1468
1469                 if (primary_rxq == NULL)
1470                         continue;
1471                 /* Not supported yet. */
1472                 rx_queues[i] = NULL;
1473         }
1474         /* Update everything. */
1475         priv->txqs = (void *)tx_queues;
1476         priv->txqs_n = nb_tx_queues;
1477         priv->rxqs = (void *)rx_queues;
1478         priv->rxqs_n = nb_rx_queues;
1479         sd->data.rx_queues = rx_queues;
1480         sd->data.tx_queues = tx_queues;
1481         sd->data.nb_rx_queues = nb_rx_queues;
1482         sd->data.nb_tx_queues = nb_tx_queues;
1483         sd->data.dev_link = sd->shared_dev_data->dev_link;
1484         sd->data.mtu = sd->shared_dev_data->mtu;
1485         memcpy(sd->data.rx_queue_state, sd->shared_dev_data->rx_queue_state,
1486                sizeof(sd->data.rx_queue_state));
1487         memcpy(sd->data.tx_queue_state, sd->shared_dev_data->tx_queue_state,
1488                sizeof(sd->data.tx_queue_state));
1489         sd->data.dev_flags = sd->shared_dev_data->dev_flags;
1490         /* Use local data from now on. */
1491         rte_mb();
1492         priv->dev->data = &sd->data;
1493         rte_mb();
1494         priv_select_tx_function(priv);
1495         priv_select_rx_function(priv);
1496         priv_unlock(priv);
1497 end:
1498         /* More sanity checks. */
1499         assert(priv->dev->data == &sd->data);
1500         rte_spinlock_unlock(&sd->lock);
1501         return priv;
1502 error:
1503         priv_unlock(priv);
1504         rte_free(tx_queues);
1505         rte_free(rx_queues);
1506         rte_spinlock_unlock(&sd->lock);
1507         return NULL;
1508 }
1509
1510 /**
1511  * Configure the TX function to use.
1512  *
1513  * @param priv
1514  *   Pointer to private structure.
1515  */
1516 void
1517 priv_select_tx_function(struct priv *priv)
1518 {
1519         priv->dev->tx_pkt_burst = mlx5_tx_burst;
1520         /* Display warning for unsupported configurations. */
1521         if (priv->sriov && priv->mps)
1522                 WARN("multi-packet send WQE cannot be used on a SR-IOV setup");
1523         /* Select appropriate TX function. */
1524         if ((priv->sriov == 0) && priv->mps && priv->txq_inline) {
1525                 priv->dev->tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1526                 DEBUG("selected MPW inline TX function");
1527         } else if ((priv->sriov == 0) && priv->mps) {
1528                 priv->dev->tx_pkt_burst = mlx5_tx_burst_mpw;
1529                 DEBUG("selected MPW TX function");
1530         }
1531 }
1532
1533 /**
1534  * Configure the RX function to use.
1535  *
1536  * @param priv
1537  *   Pointer to private structure.
1538  */
1539 void
1540 priv_select_rx_function(struct priv *priv)
1541 {
1542         priv->dev->rx_pkt_burst = mlx5_rx_burst;
1543 }