net/mlx5: prefix all functions with mlx5
[dpdk.git] / drivers / net / mlx5 / mlx5_ethdev.c
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright 2015 6WIND S.A.
3  * Copyright 2015 Mellanox.
4  */
5
6 #define _GNU_SOURCE
7
8 #include <stddef.h>
9 #include <assert.h>
10 #include <inttypes.h>
11 #include <unistd.h>
12 #include <stdint.h>
13 #include <stdio.h>
14 #include <string.h>
15 #include <stdlib.h>
16 #include <errno.h>
17 #include <dirent.h>
18 #include <net/if.h>
19 #include <sys/ioctl.h>
20 #include <sys/socket.h>
21 #include <sys/utsname.h>
22 #include <netinet/in.h>
23 #include <linux/ethtool.h>
24 #include <linux/sockios.h>
25 #include <linux/version.h>
26 #include <fcntl.h>
27 #include <stdalign.h>
28 #include <sys/un.h>
29
30 #include <rte_atomic.h>
31 #include <rte_ethdev_driver.h>
32 #include <rte_bus_pci.h>
33 #include <rte_mbuf.h>
34 #include <rte_common.h>
35 #include <rte_interrupts.h>
36 #include <rte_alarm.h>
37 #include <rte_malloc.h>
38
39 #include "mlx5.h"
40 #include "mlx5_glue.h"
41 #include "mlx5_rxtx.h"
42 #include "mlx5_utils.h"
43
44 /* Add defines in case the running kernel is not the same as user headers. */
45 #ifndef ETHTOOL_GLINKSETTINGS
46 struct ethtool_link_settings {
47         uint32_t cmd;
48         uint32_t speed;
49         uint8_t duplex;
50         uint8_t port;
51         uint8_t phy_address;
52         uint8_t autoneg;
53         uint8_t mdio_support;
54         uint8_t eth_to_mdix;
55         uint8_t eth_tp_mdix_ctrl;
56         int8_t link_mode_masks_nwords;
57         uint32_t reserved[8];
58         uint32_t link_mode_masks[];
59 };
60
61 #define ETHTOOL_GLINKSETTINGS 0x0000004c
62 #define ETHTOOL_LINK_MODE_1000baseT_Full_BIT 5
63 #define ETHTOOL_LINK_MODE_Autoneg_BIT 6
64 #define ETHTOOL_LINK_MODE_1000baseKX_Full_BIT 17
65 #define ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT 18
66 #define ETHTOOL_LINK_MODE_10000baseKR_Full_BIT 19
67 #define ETHTOOL_LINK_MODE_10000baseR_FEC_BIT 20
68 #define ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT 21
69 #define ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT 22
70 #define ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT 23
71 #define ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT 24
72 #define ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT 25
73 #define ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT 26
74 #define ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT 27
75 #define ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT 28
76 #define ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT 29
77 #define ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT 30
78 #endif
79 #ifndef HAVE_ETHTOOL_LINK_MODE_25G
80 #define ETHTOOL_LINK_MODE_25000baseCR_Full_BIT 31
81 #define ETHTOOL_LINK_MODE_25000baseKR_Full_BIT 32
82 #define ETHTOOL_LINK_MODE_25000baseSR_Full_BIT 33
83 #endif
84 #ifndef HAVE_ETHTOOL_LINK_MODE_50G
85 #define ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT 34
86 #define ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT 35
87 #endif
88 #ifndef HAVE_ETHTOOL_LINK_MODE_100G
89 #define ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT 36
90 #define ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT 37
91 #define ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT 38
92 #define ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT 39
93 #endif
94
95 /**
96  * Get interface name from private structure.
97  *
98  * @param[in] dev
99  *   Pointer to Ethernet device.
100  * @param[out] ifname
101  *   Interface name output buffer.
102  *
103  * @return
104  *   0 on success, -1 on failure and errno is set.
105  */
106 int
107 mlx5_get_ifname(const struct rte_eth_dev *dev, char (*ifname)[IF_NAMESIZE])
108 {
109         struct priv *priv = dev->data->dev_private;
110         DIR *dir;
111         struct dirent *dent;
112         unsigned int dev_type = 0;
113         unsigned int dev_port_prev = ~0u;
114         char match[IF_NAMESIZE] = "";
115
116         {
117                 MKSTR(path, "%s/device/net", priv->ibdev_path);
118
119                 dir = opendir(path);
120                 if (dir == NULL)
121                         return -1;
122         }
123         while ((dent = readdir(dir)) != NULL) {
124                 char *name = dent->d_name;
125                 FILE *file;
126                 unsigned int dev_port;
127                 int r;
128
129                 if ((name[0] == '.') &&
130                     ((name[1] == '\0') ||
131                      ((name[1] == '.') && (name[2] == '\0'))))
132                         continue;
133
134                 MKSTR(path, "%s/device/net/%s/%s",
135                       priv->ibdev_path, name,
136                       (dev_type ? "dev_id" : "dev_port"));
137
138                 file = fopen(path, "rb");
139                 if (file == NULL) {
140                         if (errno != ENOENT)
141                                 continue;
142                         /*
143                          * Switch to dev_id when dev_port does not exist as
144                          * is the case with Linux kernel versions < 3.15.
145                          */
146 try_dev_id:
147                         match[0] = '\0';
148                         if (dev_type)
149                                 break;
150                         dev_type = 1;
151                         dev_port_prev = ~0u;
152                         rewinddir(dir);
153                         continue;
154                 }
155                 r = fscanf(file, (dev_type ? "%x" : "%u"), &dev_port);
156                 fclose(file);
157                 if (r != 1)
158                         continue;
159                 /*
160                  * Switch to dev_id when dev_port returns the same value for
161                  * all ports. May happen when using a MOFED release older than
162                  * 3.0 with a Linux kernel >= 3.15.
163                  */
164                 if (dev_port == dev_port_prev)
165                         goto try_dev_id;
166                 dev_port_prev = dev_port;
167                 if (dev_port == (priv->port - 1u))
168                         snprintf(match, sizeof(match), "%s", name);
169         }
170         closedir(dir);
171         if (match[0] == '\0')
172                 return -1;
173         strncpy(*ifname, match, sizeof(*ifname));
174         return 0;
175 }
176
177 /**
178  * Perform ifreq ioctl() on associated Ethernet device.
179  *
180  * @param[in] dev
181  *   Pointer to Ethernet device.
182  * @param req
183  *   Request number to pass to ioctl().
184  * @param[out] ifr
185  *   Interface request structure output buffer.
186  *
187  * @return
188  *   0 on success, -1 on failure and errno is set.
189  */
190 int
191 mlx5_ifreq(const struct rte_eth_dev *dev, int req, struct ifreq *ifr)
192 {
193         int sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
194         int ret = -1;
195
196         if (sock == -1)
197                 return ret;
198         if (mlx5_get_ifname(dev, &ifr->ifr_name) == 0)
199                 ret = ioctl(sock, req, ifr);
200         close(sock);
201         return ret;
202 }
203
204 /**
205  * Get device MTU.
206  *
207  * @param dev
208  *   Pointer to Ethernet device.
209  * @param[out] mtu
210  *   MTU value output buffer.
211  *
212  * @return
213  *   0 on success, -1 on failure and errno is set.
214  */
215 int
216 mlx5_get_mtu(struct rte_eth_dev *dev, uint16_t *mtu)
217 {
218         struct ifreq request;
219         int ret = mlx5_ifreq(dev, SIOCGIFMTU, &request);
220
221         if (ret)
222                 return ret;
223         *mtu = request.ifr_mtu;
224         return 0;
225 }
226
227 /**
228  * Set device MTU.
229  *
230  * @param dev
231  *   Pointer to Ethernet device.
232  * @param mtu
233  *   MTU value to set.
234  *
235  * @return
236  *   0 on success, -1 on failure and errno is set.
237  */
238 static int
239 mlx5_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
240 {
241         struct ifreq request = { .ifr_mtu = mtu, };
242
243         return mlx5_ifreq(dev, SIOCSIFMTU, &request);
244 }
245
246 /**
247  * Set device flags.
248  *
249  * @param dev
250  *   Pointer to Ethernet device.
251  * @param keep
252  *   Bitmask for flags that must remain untouched.
253  * @param flags
254  *   Bitmask for flags to modify.
255  *
256  * @return
257  *   0 on success, -1 on failure and errno is set.
258  */
259 int
260 mlx5_set_flags(struct rte_eth_dev *dev, unsigned int keep, unsigned int flags)
261 {
262         struct ifreq request;
263         int ret = mlx5_ifreq(dev, SIOCGIFFLAGS, &request);
264
265         if (ret)
266                 return ret;
267         request.ifr_flags &= keep;
268         request.ifr_flags |= flags & ~keep;
269         return mlx5_ifreq(dev, SIOCSIFFLAGS, &request);
270 }
271
272 /**
273  * DPDK callback for Ethernet device configuration.
274  *
275  * @param dev
276  *   Pointer to Ethernet device structure.
277  *
278  * @return
279  *   0 on success, negative errno value on failure.
280  */
281 int
282 mlx5_dev_configure(struct rte_eth_dev *dev)
283 {
284         struct priv *priv = dev->data->dev_private;
285         unsigned int rxqs_n = dev->data->nb_rx_queues;
286         unsigned int txqs_n = dev->data->nb_tx_queues;
287         unsigned int i;
288         unsigned int j;
289         unsigned int reta_idx_n;
290         const uint8_t use_app_rss_key =
291                 !!dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key;
292         uint64_t supp_tx_offloads = mlx5_get_tx_port_offloads(dev);
293         uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
294         uint64_t supp_rx_offloads =
295                 (mlx5_get_rx_port_offloads() |
296                  mlx5_get_rx_queue_offloads(dev));
297         uint64_t rx_offloads = dev->data->dev_conf.rxmode.offloads;
298
299         if ((tx_offloads & supp_tx_offloads) != tx_offloads) {
300                 ERROR("Some Tx offloads are not supported "
301                       "requested 0x%" PRIx64 " supported 0x%" PRIx64,
302                       tx_offloads, supp_tx_offloads);
303                 return ENOTSUP;
304         }
305         if ((rx_offloads & supp_rx_offloads) != rx_offloads) {
306                 ERROR("Some Rx offloads are not supported "
307                       "requested 0x%" PRIx64 " supported 0x%" PRIx64,
308                       rx_offloads, supp_rx_offloads);
309                 return ENOTSUP;
310         }
311         if (use_app_rss_key &&
312             (dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key_len !=
313              rss_hash_default_key_len)) {
314                 /* MLX5 RSS only support 40bytes key. */
315                 return EINVAL;
316         }
317         priv->rss_conf.rss_key =
318                 rte_realloc(priv->rss_conf.rss_key,
319                             rss_hash_default_key_len, 0);
320         if (!priv->rss_conf.rss_key) {
321                 ERROR("cannot allocate RSS hash key memory (%u)", rxqs_n);
322                 return ENOMEM;
323         }
324         memcpy(priv->rss_conf.rss_key,
325                use_app_rss_key ?
326                dev->data->dev_conf.rx_adv_conf.rss_conf.rss_key :
327                rss_hash_default_key,
328                rss_hash_default_key_len);
329         priv->rss_conf.rss_key_len = rss_hash_default_key_len;
330         priv->rss_conf.rss_hf = dev->data->dev_conf.rx_adv_conf.rss_conf.rss_hf;
331         priv->rxqs = (void *)dev->data->rx_queues;
332         priv->txqs = (void *)dev->data->tx_queues;
333         if (txqs_n != priv->txqs_n) {
334                 INFO("%p: TX queues number update: %u -> %u",
335                      (void *)dev, priv->txqs_n, txqs_n);
336                 priv->txqs_n = txqs_n;
337         }
338         if (rxqs_n > priv->config.ind_table_max_size) {
339                 ERROR("cannot handle this many RX queues (%u)", rxqs_n);
340                 return EINVAL;
341         }
342         if (rxqs_n == priv->rxqs_n)
343                 return 0;
344         INFO("%p: RX queues number update: %u -> %u",
345              (void *)dev, priv->rxqs_n, rxqs_n);
346         priv->rxqs_n = rxqs_n;
347         /* If the requested number of RX queues is not a power of two, use the
348          * maximum indirection table size for better balancing.
349          * The result is always rounded to the next power of two. */
350         reta_idx_n = (1 << log2above((rxqs_n & (rxqs_n - 1)) ?
351                                      priv->config.ind_table_max_size :
352                                      rxqs_n));
353         if (mlx5_rss_reta_index_resize(dev, reta_idx_n))
354                 return ENOMEM;
355         /* When the number of RX queues is not a power of two, the remaining
356          * table entries are padded with reused WQs and hashes are not spread
357          * uniformly. */
358         for (i = 0, j = 0; (i != reta_idx_n); ++i) {
359                 (*priv->reta_idx)[i] = j;
360                 if (++j == rxqs_n)
361                         j = 0;
362         }
363         return 0;
364
365 }
366
367 /**
368  * DPDK callback to get information about the device.
369  *
370  * @param dev
371  *   Pointer to Ethernet device structure.
372  * @param[out] info
373  *   Info structure output buffer.
374  */
375 void
376 mlx5_dev_infos_get(struct rte_eth_dev *dev, struct rte_eth_dev_info *info)
377 {
378         struct priv *priv = dev->data->dev_private;
379         struct mlx5_dev_config *config = &priv->config;
380         unsigned int max;
381         char ifname[IF_NAMESIZE];
382
383         info->pci_dev = RTE_ETH_DEV_TO_PCI(dev);
384         /* FIXME: we should ask the device for these values. */
385         info->min_rx_bufsize = 32;
386         info->max_rx_pktlen = 65536;
387         /*
388          * Since we need one CQ per QP, the limit is the minimum number
389          * between the two values.
390          */
391         max = RTE_MIN(priv->device_attr.orig_attr.max_cq,
392                       priv->device_attr.orig_attr.max_qp);
393         /* If max >= 65535 then max = 0, max_rx_queues is uint16_t. */
394         if (max >= 65535)
395                 max = 65535;
396         info->max_rx_queues = max;
397         info->max_tx_queues = max;
398         info->max_mac_addrs = RTE_DIM(priv->mac);
399         info->rx_queue_offload_capa = mlx5_get_rx_queue_offloads(dev);
400         info->rx_offload_capa = (mlx5_get_rx_port_offloads() |
401                                  info->rx_queue_offload_capa);
402         info->tx_offload_capa = mlx5_get_tx_port_offloads(dev);
403         if (mlx5_get_ifname(dev, &ifname) == 0)
404                 info->if_index = if_nametoindex(ifname);
405         info->reta_size = priv->reta_idx_n ?
406                 priv->reta_idx_n : config->ind_table_max_size;
407         info->hash_key_size = priv->rss_conf.rss_key_len;
408         info->speed_capa = priv->link_speed_capa;
409         info->flow_type_rss_offloads = ~MLX5_RSS_HF_MASK;
410 }
411
412 /**
413  * Get supported packet types.
414  *
415  * @param dev
416  *   Pointer to Ethernet device structure.
417  *
418  * @return
419  *   A pointer to the supported Packet types array.
420  */
421 const uint32_t *
422 mlx5_dev_supported_ptypes_get(struct rte_eth_dev *dev)
423 {
424         static const uint32_t ptypes[] = {
425                 /* refers to rxq_cq_to_pkt_type() */
426                 RTE_PTYPE_L2_ETHER,
427                 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
428                 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
429                 RTE_PTYPE_L4_NONFRAG,
430                 RTE_PTYPE_L4_FRAG,
431                 RTE_PTYPE_L4_TCP,
432                 RTE_PTYPE_L4_UDP,
433                 RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
434                 RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
435                 RTE_PTYPE_INNER_L4_NONFRAG,
436                 RTE_PTYPE_INNER_L4_FRAG,
437                 RTE_PTYPE_INNER_L4_TCP,
438                 RTE_PTYPE_INNER_L4_UDP,
439                 RTE_PTYPE_UNKNOWN
440         };
441
442         if (dev->rx_pkt_burst == mlx5_rx_burst ||
443             dev->rx_pkt_burst == mlx5_rx_burst_vec)
444                 return ptypes;
445         return NULL;
446 }
447
448 /**
449  * DPDK callback to retrieve physical link information.
450  *
451  * @param dev
452  *   Pointer to Ethernet device structure.
453  *
454  * @return
455  *   0 on success, -1 on error.
456  */
457 static int
458 mlx5_link_update_unlocked_gset(struct rte_eth_dev *dev)
459 {
460         struct priv *priv = dev->data->dev_private;
461         struct ethtool_cmd edata = {
462                 .cmd = ETHTOOL_GSET /* Deprecated since Linux v4.5. */
463         };
464         struct ifreq ifr;
465         struct rte_eth_link dev_link;
466         int link_speed = 0;
467
468         if (mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr)) {
469                 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
470                 return -1;
471         }
472         memset(&dev_link, 0, sizeof(dev_link));
473         dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
474                                 (ifr.ifr_flags & IFF_RUNNING));
475         ifr.ifr_data = (void *)&edata;
476         if (mlx5_ifreq(dev, SIOCETHTOOL, &ifr)) {
477                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GSET) failed: %s",
478                      strerror(errno));
479                 return -1;
480         }
481         link_speed = ethtool_cmd_speed(&edata);
482         if (link_speed == -1)
483                 dev_link.link_speed = 0;
484         else
485                 dev_link.link_speed = link_speed;
486         priv->link_speed_capa = 0;
487         if (edata.supported & SUPPORTED_Autoneg)
488                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
489         if (edata.supported & (SUPPORTED_1000baseT_Full |
490                                SUPPORTED_1000baseKX_Full))
491                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
492         if (edata.supported & SUPPORTED_10000baseKR_Full)
493                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
494         if (edata.supported & (SUPPORTED_40000baseKR4_Full |
495                                SUPPORTED_40000baseCR4_Full |
496                                SUPPORTED_40000baseSR4_Full |
497                                SUPPORTED_40000baseLR4_Full))
498                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
499         dev_link.link_duplex = ((edata.duplex == DUPLEX_HALF) ?
500                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
501         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
502                         ETH_LINK_SPEED_FIXED);
503         if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
504                 /* Link status changed. */
505                 dev->data->dev_link = dev_link;
506                 return 0;
507         }
508         /* Link status is still the same. */
509         return -1;
510 }
511
512 /**
513  * Retrieve physical link information (unlocked version using new ioctl).
514  *
515  * @param dev
516  *   Pointer to Ethernet device structure.
517  *
518  * @return
519  *   0 on success, -1 on error.
520  */
521 static int
522 mlx5_link_update_unlocked_gs(struct rte_eth_dev *dev)
523 {
524         struct priv *priv = dev->data->dev_private;
525         struct ethtool_link_settings gcmd = { .cmd = ETHTOOL_GLINKSETTINGS };
526         struct ifreq ifr;
527         struct rte_eth_link dev_link;
528         uint64_t sc;
529
530         if (mlx5_ifreq(dev, SIOCGIFFLAGS, &ifr)) {
531                 WARN("ioctl(SIOCGIFFLAGS) failed: %s", strerror(errno));
532                 return -1;
533         }
534         memset(&dev_link, 0, sizeof(dev_link));
535         dev_link.link_status = ((ifr.ifr_flags & IFF_UP) &&
536                                 (ifr.ifr_flags & IFF_RUNNING));
537         ifr.ifr_data = (void *)&gcmd;
538         if (mlx5_ifreq(dev, SIOCETHTOOL, &ifr)) {
539                 DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
540                       strerror(errno));
541                 return -1;
542         }
543         gcmd.link_mode_masks_nwords = -gcmd.link_mode_masks_nwords;
544
545         alignas(struct ethtool_link_settings)
546         uint8_t data[offsetof(struct ethtool_link_settings, link_mode_masks) +
547                      sizeof(uint32_t) * gcmd.link_mode_masks_nwords * 3];
548         struct ethtool_link_settings *ecmd = (void *)data;
549
550         *ecmd = gcmd;
551         ifr.ifr_data = (void *)ecmd;
552         if (mlx5_ifreq(dev, SIOCETHTOOL, &ifr)) {
553                 DEBUG("ioctl(SIOCETHTOOL, ETHTOOL_GLINKSETTINGS) failed: %s",
554                       strerror(errno));
555                 return -1;
556         }
557         dev_link.link_speed = ecmd->speed;
558         sc = ecmd->link_mode_masks[0] |
559                 ((uint64_t)ecmd->link_mode_masks[1] << 32);
560         priv->link_speed_capa = 0;
561         if (sc & MLX5_BITSHIFT(ETHTOOL_LINK_MODE_Autoneg_BIT))
562                 priv->link_speed_capa |= ETH_LINK_SPEED_AUTONEG;
563         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseT_Full_BIT) |
564                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_1000baseKX_Full_BIT)))
565                 priv->link_speed_capa |= ETH_LINK_SPEED_1G;
566         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKX4_Full_BIT) |
567                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseKR_Full_BIT) |
568                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_10000baseR_FEC_BIT)))
569                 priv->link_speed_capa |= ETH_LINK_SPEED_10G;
570         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseMLD2_Full_BIT) |
571                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_20000baseKR2_Full_BIT)))
572                 priv->link_speed_capa |= ETH_LINK_SPEED_20G;
573         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseKR4_Full_BIT) |
574                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseCR4_Full_BIT) |
575                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseSR4_Full_BIT) |
576                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_40000baseLR4_Full_BIT)))
577                 priv->link_speed_capa |= ETH_LINK_SPEED_40G;
578         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseKR4_Full_BIT) |
579                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseCR4_Full_BIT) |
580                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseSR4_Full_BIT) |
581                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_56000baseLR4_Full_BIT)))
582                 priv->link_speed_capa |= ETH_LINK_SPEED_56G;
583         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseCR_Full_BIT) |
584                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseKR_Full_BIT) |
585                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_25000baseSR_Full_BIT)))
586                 priv->link_speed_capa |= ETH_LINK_SPEED_25G;
587         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseCR2_Full_BIT) |
588                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_50000baseKR2_Full_BIT)))
589                 priv->link_speed_capa |= ETH_LINK_SPEED_50G;
590         if (sc & (MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseKR4_Full_BIT) |
591                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseSR4_Full_BIT) |
592                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseCR4_Full_BIT) |
593                   MLX5_BITSHIFT(ETHTOOL_LINK_MODE_100000baseLR4_ER4_Full_BIT)))
594                 priv->link_speed_capa |= ETH_LINK_SPEED_100G;
595         dev_link.link_duplex = ((ecmd->duplex == DUPLEX_HALF) ?
596                                 ETH_LINK_HALF_DUPLEX : ETH_LINK_FULL_DUPLEX);
597         dev_link.link_autoneg = !(dev->data->dev_conf.link_speeds &
598                                   ETH_LINK_SPEED_FIXED);
599         if (memcmp(&dev_link, &dev->data->dev_link, sizeof(dev_link))) {
600                 /* Link status changed. */
601                 dev->data->dev_link = dev_link;
602                 return 0;
603         }
604         /* Link status is still the same. */
605         return -1;
606 }
607
608 /**
609  * Enable receiving and transmitting traffic.
610  *
611  * @param dev
612  *   Pointer to Ethernet device.
613  */
614 static void
615 mlx5_link_start(struct rte_eth_dev *dev)
616 {
617         struct priv *priv = dev->data->dev_private;
618         int err;
619
620         dev->tx_pkt_burst = mlx5_select_tx_function(dev);
621         dev->rx_pkt_burst = mlx5_select_rx_function(dev);
622         err = mlx5_traffic_enable(dev);
623         if (err)
624                 ERROR("%p: error occurred while configuring control flows: %s",
625                       (void *)dev, strerror(err));
626         err = mlx5_flow_start(dev, &priv->flows);
627         if (err)
628                 ERROR("%p: error occurred while configuring flows: %s",
629                       (void *)dev, strerror(err));
630 }
631
632 /**
633  * Disable receiving and transmitting traffic.
634  *
635  * @param dev
636  *   Pointer to Ethernet device.
637  */
638 static void
639 mlx5_link_stop(struct rte_eth_dev *dev)
640 {
641         struct priv *priv = dev->data->dev_private;
642
643         mlx5_flow_stop(dev, &priv->flows);
644         mlx5_traffic_disable(dev);
645         dev->rx_pkt_burst = removed_rx_burst;
646         dev->tx_pkt_burst = removed_tx_burst;
647 }
648
649 /**
650  * Querying the link status till it changes to the desired state.
651  * Number of query attempts is bounded by MLX5_MAX_LINK_QUERY_ATTEMPTS.
652  *
653  * @param dev
654  *   Pointer to Ethernet device.
655  * @param status
656  *   Link desired status.
657  *
658  * @return
659  *   0 on success, negative errno value on failure.
660  */
661 int
662 mlx5_force_link_status_change(struct rte_eth_dev *dev, int status)
663 {
664         int try = 0;
665
666         while (try < MLX5_MAX_LINK_QUERY_ATTEMPTS) {
667                 mlx5_link_update(dev, 0);
668                 if (dev->data->dev_link.link_status == status)
669                         return 0;
670                 try++;
671                 sleep(1);
672         }
673         return -EAGAIN;
674 }
675
676 /**
677  * DPDK callback to retrieve physical link information.
678  *
679  * @param dev
680  *   Pointer to Ethernet device structure.
681  * @param wait_to_complete
682  *   Wait for request completion (ignored).
683  *
684  * @return
685  *   0 on success, -1 on error.
686  */
687 int
688 mlx5_link_update(struct rte_eth_dev *dev, int wait_to_complete __rte_unused)
689 {
690         struct utsname utsname;
691         int ver[3];
692         int ret;
693         struct rte_eth_link dev_link = dev->data->dev_link;
694
695         if (uname(&utsname) == -1 ||
696             sscanf(utsname.release, "%d.%d.%d",
697                    &ver[0], &ver[1], &ver[2]) != 3 ||
698             KERNEL_VERSION(ver[0], ver[1], ver[2]) < KERNEL_VERSION(4, 9, 0))
699                 ret = mlx5_link_update_unlocked_gset(dev);
700         else
701                 ret = mlx5_link_update_unlocked_gs(dev);
702         /* If lsc interrupt is disabled, should always be ready for traffic. */
703         if (!dev->data->dev_conf.intr_conf.lsc) {
704                 mlx5_link_start(dev);
705                 return ret;
706         }
707         /* Re-select burst callbacks only if link status has been changed. */
708         if (!ret && dev_link.link_status != dev->data->dev_link.link_status) {
709                 if (dev->data->dev_link.link_status == ETH_LINK_UP)
710                         mlx5_link_start(dev);
711                 else
712                         mlx5_link_stop(dev);
713         }
714         return ret;
715 }
716
717 /**
718  * DPDK callback to change the MTU.
719  *
720  * @param dev
721  *   Pointer to Ethernet device structure.
722  * @param in_mtu
723  *   New MTU.
724  *
725  * @return
726  *   0 on success, negative errno value on failure.
727  */
728 int
729 mlx5_dev_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
730 {
731         struct priv *priv = dev->data->dev_private;
732         uint16_t kern_mtu;
733         int ret = 0;
734
735         ret = mlx5_get_mtu(dev, &kern_mtu);
736         if (ret)
737                 goto out;
738         /* Set kernel interface MTU first. */
739         ret = mlx5_set_mtu(dev, mtu);
740         if (ret)
741                 goto out;
742         ret = mlx5_get_mtu(dev, &kern_mtu);
743         if (ret)
744                 goto out;
745         if (kern_mtu == mtu) {
746                 priv->mtu = mtu;
747                 DEBUG("adapter port %u MTU set to %u", priv->port, mtu);
748         }
749         return 0;
750 out:
751         ret = errno;
752         WARN("cannot set port %u MTU to %u: %s", priv->port, mtu,
753              strerror(ret));
754         assert(ret >= 0);
755         return -ret;
756 }
757
758 /**
759  * DPDK callback to get flow control status.
760  *
761  * @param dev
762  *   Pointer to Ethernet device structure.
763  * @param[out] fc_conf
764  *   Flow control output buffer.
765  *
766  * @return
767  *   0 on success, negative errno value on failure.
768  */
769 int
770 mlx5_dev_get_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
771 {
772         struct ifreq ifr;
773         struct ethtool_pauseparam ethpause = {
774                 .cmd = ETHTOOL_GPAUSEPARAM
775         };
776         int ret;
777
778         ifr.ifr_data = (void *)&ethpause;
779         if (mlx5_ifreq(dev, SIOCETHTOOL, &ifr)) {
780                 ret = errno;
781                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_GPAUSEPARAM) failed: %s",
782                      strerror(ret));
783                 goto out;
784         }
785         fc_conf->autoneg = ethpause.autoneg;
786         if (ethpause.rx_pause && ethpause.tx_pause)
787                 fc_conf->mode = RTE_FC_FULL;
788         else if (ethpause.rx_pause)
789                 fc_conf->mode = RTE_FC_RX_PAUSE;
790         else if (ethpause.tx_pause)
791                 fc_conf->mode = RTE_FC_TX_PAUSE;
792         else
793                 fc_conf->mode = RTE_FC_NONE;
794         ret = 0;
795 out:
796         assert(ret >= 0);
797         return -ret;
798 }
799
800 /**
801  * DPDK callback to modify flow control parameters.
802  *
803  * @param dev
804  *   Pointer to Ethernet device structure.
805  * @param[in] fc_conf
806  *   Flow control parameters.
807  *
808  * @return
809  *   0 on success, negative errno value on failure.
810  */
811 int
812 mlx5_dev_set_flow_ctrl(struct rte_eth_dev *dev, struct rte_eth_fc_conf *fc_conf)
813 {
814         struct ifreq ifr;
815         struct ethtool_pauseparam ethpause = {
816                 .cmd = ETHTOOL_SPAUSEPARAM
817         };
818         int ret;
819
820         ifr.ifr_data = (void *)&ethpause;
821         ethpause.autoneg = fc_conf->autoneg;
822         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
823             (fc_conf->mode & RTE_FC_RX_PAUSE))
824                 ethpause.rx_pause = 1;
825         else
826                 ethpause.rx_pause = 0;
827
828         if (((fc_conf->mode & RTE_FC_FULL) == RTE_FC_FULL) ||
829             (fc_conf->mode & RTE_FC_TX_PAUSE))
830                 ethpause.tx_pause = 1;
831         else
832                 ethpause.tx_pause = 0;
833         if (mlx5_ifreq(dev, SIOCETHTOOL, &ifr)) {
834                 ret = errno;
835                 WARN("ioctl(SIOCETHTOOL, ETHTOOL_SPAUSEPARAM)"
836                      " failed: %s",
837                      strerror(ret));
838                 goto out;
839         }
840         ret = 0;
841 out:
842         assert(ret >= 0);
843         return -ret;
844 }
845
846 /**
847  * Get PCI information from struct ibv_device.
848  *
849  * @param device
850  *   Pointer to Ethernet device structure.
851  * @param[out] pci_addr
852  *   PCI bus address output buffer.
853  *
854  * @return
855  *   0 on success, -1 on failure and errno is set.
856  */
857 int
858 mlx5_ibv_device_to_pci_addr(const struct ibv_device *device,
859                             struct rte_pci_addr *pci_addr)
860 {
861         FILE *file;
862         char line[32];
863         MKSTR(path, "%s/device/uevent", device->ibdev_path);
864
865         file = fopen(path, "rb");
866         if (file == NULL)
867                 return -1;
868         while (fgets(line, sizeof(line), file) == line) {
869                 size_t len = strlen(line);
870                 int ret;
871
872                 /* Truncate long lines. */
873                 if (len == (sizeof(line) - 1))
874                         while (line[(len - 1)] != '\n') {
875                                 ret = fgetc(file);
876                                 if (ret == EOF)
877                                         break;
878                                 line[(len - 1)] = ret;
879                         }
880                 /* Extract information. */
881                 if (sscanf(line,
882                            "PCI_SLOT_NAME="
883                            "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 "\n",
884                            &pci_addr->domain,
885                            &pci_addr->bus,
886                            &pci_addr->devid,
887                            &pci_addr->function) == 4) {
888                         ret = 0;
889                         break;
890                 }
891         }
892         fclose(file);
893         return 0;
894 }
895
896 /**
897  * Update the link status.
898  *
899  * @param dev
900  *   Pointer to Ethernet device.
901  *
902  * @return
903  *   Zero if the callback process can be called immediately.
904  */
905 static int
906 mlx5_link_status_update(struct rte_eth_dev *dev)
907 {
908         struct priv *priv = dev->data->dev_private;
909         struct rte_eth_link *link = &dev->data->dev_link;
910
911         mlx5_link_update(dev, 0);
912         if (((link->link_speed == 0) && link->link_status) ||
913                 ((link->link_speed != 0) && !link->link_status)) {
914                 /*
915                  * Inconsistent status. Event likely occurred before the
916                  * kernel netdevice exposes the new status.
917                  */
918                 if (!priv->pending_alarm) {
919                         priv->pending_alarm = 1;
920                         rte_eal_alarm_set(MLX5_ALARM_TIMEOUT_US,
921                                           mlx5_dev_link_status_handler,
922                                           priv->dev);
923                 }
924                 return 1;
925         } else if (unlikely(priv->pending_alarm)) {
926                 /* Link interrupt occurred while alarm is already scheduled. */
927                 priv->pending_alarm = 0;
928                 rte_eal_alarm_cancel(mlx5_dev_link_status_handler, priv->dev);
929         }
930         return 0;
931 }
932
933 /**
934  * Device status handler.
935  *
936  * @param dev
937  *   Pointer to Ethernet device.
938  * @param events
939  *   Pointer to event flags holder.
940  *
941  * @return
942  *   Events bitmap of callback process which can be called immediately.
943  */
944 static uint32_t
945 mlx5_dev_status_handler(struct rte_eth_dev *dev)
946 {
947         struct priv *priv = dev->data->dev_private;
948         struct ibv_async_event event;
949         uint32_t ret = 0;
950
951         /* Read all message and acknowledge them. */
952         for (;;) {
953                 if (mlx5_glue->get_async_event(priv->ctx, &event))
954                         break;
955                 if ((event.event_type == IBV_EVENT_PORT_ACTIVE ||
956                         event.event_type == IBV_EVENT_PORT_ERR) &&
957                         (dev->data->dev_conf.intr_conf.lsc == 1))
958                         ret |= (1 << RTE_ETH_EVENT_INTR_LSC);
959                 else if (event.event_type == IBV_EVENT_DEVICE_FATAL &&
960                         dev->data->dev_conf.intr_conf.rmv == 1)
961                         ret |= (1 << RTE_ETH_EVENT_INTR_RMV);
962                 else
963                         DEBUG("event type %d on port %d not handled",
964                               event.event_type, event.element.port_num);
965                 mlx5_glue->ack_async_event(&event);
966         }
967         if (ret & (1 << RTE_ETH_EVENT_INTR_LSC))
968                 if (mlx5_link_status_update(dev))
969                         ret &= ~(1 << RTE_ETH_EVENT_INTR_LSC);
970         return ret;
971 }
972
973 /**
974  * Handle delayed link status event.
975  *
976  * @param arg
977  *   Registered argument.
978  */
979 void
980 mlx5_dev_link_status_handler(void *arg)
981 {
982         struct rte_eth_dev *dev = arg;
983         struct priv *priv = dev->data->dev_private;
984         int ret;
985
986         priv->pending_alarm = 0;
987         ret = mlx5_link_status_update(dev);
988         if (!ret)
989                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
990 }
991
992 /**
993  * Handle interrupts from the NIC.
994  *
995  * @param[in] intr_handle
996  *   Interrupt handler.
997  * @param cb_arg
998  *   Callback argument.
999  */
1000 void
1001 mlx5_dev_interrupt_handler(void *cb_arg)
1002 {
1003         struct rte_eth_dev *dev = cb_arg;
1004         uint32_t events;
1005
1006         events = mlx5_dev_status_handler(dev);
1007         if (events & (1 << RTE_ETH_EVENT_INTR_LSC))
1008                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_LSC, NULL);
1009         if (events & (1 << RTE_ETH_EVENT_INTR_RMV))
1010                 _rte_eth_dev_callback_process(dev, RTE_ETH_EVENT_INTR_RMV, NULL);
1011 }
1012
1013 /**
1014  * Handle interrupts from the socket.
1015  *
1016  * @param cb_arg
1017  *   Callback argument.
1018  */
1019 static void
1020 mlx5_dev_handler_socket(void *cb_arg)
1021 {
1022         struct rte_eth_dev *dev = cb_arg;
1023
1024         mlx5_socket_handle(dev);
1025 }
1026
1027 /**
1028  * Uninstall interrupt handler.
1029  *
1030  * @param dev
1031  *   Pointer to Ethernet device.
1032  */
1033 void
1034 mlx5_dev_interrupt_handler_uninstall(struct rte_eth_dev *dev)
1035 {
1036         struct priv *priv = dev->data->dev_private;
1037
1038         if (dev->data->dev_conf.intr_conf.lsc ||
1039             dev->data->dev_conf.intr_conf.rmv)
1040                 rte_intr_callback_unregister(&priv->intr_handle,
1041                                              mlx5_dev_interrupt_handler, dev);
1042         if (priv->primary_socket)
1043                 rte_intr_callback_unregister(&priv->intr_handle_socket,
1044                                              mlx5_dev_handler_socket, dev);
1045         if (priv->pending_alarm) {
1046                 priv->pending_alarm = 0;
1047                 rte_eal_alarm_cancel(mlx5_dev_link_status_handler, dev);
1048         }
1049         priv->intr_handle.fd = 0;
1050         priv->intr_handle.type = RTE_INTR_HANDLE_UNKNOWN;
1051         priv->intr_handle_socket.fd = 0;
1052         priv->intr_handle_socket.type = RTE_INTR_HANDLE_UNKNOWN;
1053 }
1054
1055 /**
1056  * Install interrupt handler.
1057  *
1058  * @param dev
1059  *   Pointer to Ethernet device.
1060  */
1061 void
1062 mlx5_dev_interrupt_handler_install(struct rte_eth_dev *dev)
1063 {
1064         struct priv *priv = dev->data->dev_private;
1065         int rc, flags;
1066
1067         assert(priv->ctx->async_fd > 0);
1068         flags = fcntl(priv->ctx->async_fd, F_GETFL);
1069         rc = fcntl(priv->ctx->async_fd, F_SETFL, flags | O_NONBLOCK);
1070         if (rc < 0) {
1071                 INFO("failed to change file descriptor async event queue");
1072                 dev->data->dev_conf.intr_conf.lsc = 0;
1073                 dev->data->dev_conf.intr_conf.rmv = 0;
1074         }
1075         if (dev->data->dev_conf.intr_conf.lsc ||
1076             dev->data->dev_conf.intr_conf.rmv) {
1077                 priv->intr_handle.fd = priv->ctx->async_fd;
1078                 priv->intr_handle.type = RTE_INTR_HANDLE_EXT;
1079                 rte_intr_callback_register(&priv->intr_handle,
1080                                            mlx5_dev_interrupt_handler, dev);
1081         }
1082         rc = mlx5_socket_init(dev);
1083         if (!rc && priv->primary_socket) {
1084                 priv->intr_handle_socket.fd = priv->primary_socket;
1085                 priv->intr_handle_socket.type = RTE_INTR_HANDLE_EXT;
1086                 rte_intr_callback_register(&priv->intr_handle_socket,
1087                                            mlx5_dev_handler_socket, dev);
1088         }
1089 }
1090
1091 /**
1092  * DPDK callback to bring the link DOWN.
1093  *
1094  * @param dev
1095  *   Pointer to Ethernet device structure.
1096  *
1097  * @return
1098  *   0 on success, errno value on failure.
1099  */
1100 int
1101 mlx5_set_link_down(struct rte_eth_dev *dev)
1102 {
1103         return mlx5_set_flags(dev, ~IFF_UP, ~IFF_UP);
1104 }
1105
1106 /**
1107  * DPDK callback to bring the link UP.
1108  *
1109  * @param dev
1110  *   Pointer to Ethernet device structure.
1111  *
1112  * @return
1113  *   0 on success, errno value on failure.
1114  */
1115 int
1116 mlx5_set_link_up(struct rte_eth_dev *dev)
1117 {
1118         return mlx5_set_flags(dev, ~IFF_UP, IFF_UP);
1119 }
1120
1121 /**
1122  * Configure the TX function to use.
1123  *
1124  * @param dev
1125  *   Pointer to private data structure.
1126  *
1127  * @return
1128  *   Pointer to selected Tx burst function.
1129  */
1130 eth_tx_burst_t
1131 mlx5_select_tx_function(struct rte_eth_dev *dev)
1132 {
1133         struct priv *priv = dev->data->dev_private;
1134         eth_tx_burst_t tx_pkt_burst = mlx5_tx_burst;
1135         struct mlx5_dev_config *config = &priv->config;
1136         uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
1137         int tso = !!(tx_offloads & (DEV_TX_OFFLOAD_TCP_TSO |
1138                                     DEV_TX_OFFLOAD_VXLAN_TNL_TSO |
1139                                     DEV_TX_OFFLOAD_GRE_TNL_TSO));
1140         int vlan_insert = !!(tx_offloads & DEV_TX_OFFLOAD_VLAN_INSERT);
1141
1142         assert(priv != NULL);
1143         /* Select appropriate TX function. */
1144         if (vlan_insert || tso)
1145                 return tx_pkt_burst;
1146         if (config->mps == MLX5_MPW_ENHANCED) {
1147                 if (mlx5_check_vec_tx_support(dev) > 0) {
1148                         if (mlx5_check_raw_vec_tx_support(dev) > 0)
1149                                 tx_pkt_burst = mlx5_tx_burst_raw_vec;
1150                         else
1151                                 tx_pkt_burst = mlx5_tx_burst_vec;
1152                         DEBUG("selected Enhanced MPW TX vectorized function");
1153                 } else {
1154                         tx_pkt_burst = mlx5_tx_burst_empw;
1155                         DEBUG("selected Enhanced MPW TX function");
1156                 }
1157         } else if (config->mps && (config->txq_inline > 0)) {
1158                 tx_pkt_burst = mlx5_tx_burst_mpw_inline;
1159                 DEBUG("selected MPW inline TX function");
1160         } else if (config->mps) {
1161                 tx_pkt_burst = mlx5_tx_burst_mpw;
1162                 DEBUG("selected MPW TX function");
1163         }
1164         return tx_pkt_burst;
1165 }
1166
1167 /**
1168  * Configure the RX function to use.
1169  *
1170  * @param dev
1171  *   Pointer to private data structure.
1172  *
1173  * @return
1174  *   Pointer to selected Rx burst function.
1175  */
1176 eth_rx_burst_t
1177 mlx5_select_rx_function(struct rte_eth_dev *dev)
1178 {
1179         eth_rx_burst_t rx_pkt_burst = mlx5_rx_burst;
1180
1181         assert(dev != NULL);
1182         if (mlx5_check_vec_rx_support(dev) > 0) {
1183                 rx_pkt_burst = mlx5_rx_burst_vec;
1184                 DEBUG("selected RX vectorized function");
1185         }
1186         return rx_pkt_burst;
1187 }
1188
1189 /**
1190  * Check if mlx5 device was removed.
1191  *
1192  * @param dev
1193  *   Pointer to Ethernet device structure.
1194  *
1195  * @return
1196  *   1 when device is removed, otherwise 0.
1197  */
1198 int
1199 mlx5_is_removed(struct rte_eth_dev *dev)
1200 {
1201         struct ibv_device_attr device_attr;
1202         struct priv *priv = dev->data->dev_private;
1203
1204         if (mlx5_glue->query_device(priv->ctx, &device_attr) == EIO)
1205                 return 1;
1206         return 0;
1207 }