drivers/net/failsafe/failsafe.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2017 6WIND S.A.
   3  * Copyright 2017 Mellanox Technologies, Ltd
   4  */
   5
   6 #include <stdbool.h>
   7
   8 #include <rte_alarm.h>
   9 #include <rte_malloc.h>
  10 #include <rte_ethdev_driver.h>
  11 #include <rte_ethdev_vdev.h>
  12 #include <rte_devargs.h>
  13 #include <rte_kvargs.h>
  14 #include <rte_bus_vdev.h>
  15
  16 #include "failsafe_private.h"
  17
  18 const char pmd_failsafe_driver_name[] = FAILSAFE_DRIVER_NAME;
  19 static const struct rte_eth_link eth_link = {
  20         .link_speed = ETH_SPEED_NUM_10G,
  21         .link_duplex = ETH_LINK_FULL_DUPLEX,
  22         .link_status = ETH_LINK_UP,
  23         .link_autoneg = ETH_LINK_AUTONEG,
  24 };
  25
  26 static int
  27 fs_sub_device_alloc(struct rte_eth_dev *dev,
  28                 const char *params)
  29 {
  30         uint8_t nb_subs;
  31         int ret;
  32         int i;
  33         struct sub_device *sdev;
  34         uint8_t sdev_iterator;
  35
  36         ret = failsafe_args_count_subdevice(dev, params);
  37         if (ret)
  38                 return ret;
  39         if (PRIV(dev)->subs_tail > FAILSAFE_MAX_ETHPORTS) {
  40                 ERROR("Cannot allocate more than %d ports",
  41                         FAILSAFE_MAX_ETHPORTS);
  42                 return -ENOSPC;
  43         }
  44         nb_subs = PRIV(dev)->subs_tail;
  45         PRIV(dev)->subs = rte_zmalloc(NULL,
  46                         sizeof(struct sub_device) * nb_subs,
  47                         RTE_CACHE_LINE_SIZE);
  48         if (PRIV(dev)->subs == NULL) {
  49                 ERROR("Could not allocate sub_devices");
  50                 return -ENOMEM;
  51         }
  52         /* Initiate static sub devices linked list. */
  53         for (i = 1; i < nb_subs; i++)
  54                 PRIV(dev)->subs[i - 1].next = PRIV(dev)->subs + i;
  55         PRIV(dev)->subs[i - 1].next = PRIV(dev)->subs;
  56
  57         FOREACH_SUBDEV(sdev, sdev_iterator, dev) {
  58                 sdev->sdev_port_id = RTE_MAX_ETHPORTS;
  59         }
  60         return 0;
  61 }
  62
  63 static void fs_hotplug_alarm(void *arg);
  64
  65 int
  66 failsafe_hotplug_alarm_install(struct rte_eth_dev *dev)
  67 {
  68         int ret;
  69
  70         if (dev == NULL)
  71                 return -EINVAL;
  72         if (PRIV(dev)->pending_alarm)
  73                 return 0;
  74         ret = rte_eal_alarm_set(failsafe_hotplug_poll * 1000,
  75                                 fs_hotplug_alarm,
  76                                 dev);
  77         if (ret) {
  78                 ERROR("Could not set up plug-in event detection");
  79                 return ret;
  80         }
  81         PRIV(dev)->pending_alarm = 1;
  82         return 0;
  83 }
  84
  85 int
  86 failsafe_hotplug_alarm_cancel(struct rte_eth_dev *dev)
  87 {
  88         int ret = 0;
  89
  90         rte_errno = 0;
  91         rte_eal_alarm_cancel(fs_hotplug_alarm, dev);
  92         if (rte_errno) {
  93                 ERROR("rte_eal_alarm_cancel failed (errno: %s)",
  94                       strerror(rte_errno));
  95                 ret = -rte_errno;
  96         } else {
  97                 PRIV(dev)->pending_alarm = 0;
  98         }
  99         return ret;
 100 }
 101
 102 static void
 103 fs_hotplug_alarm(void *arg)
 104 {
 105         struct rte_eth_dev *dev = arg;
 106         struct sub_device *sdev;
 107         int ret;
 108         uint8_t i;
 109
 110         if (!PRIV(dev)->pending_alarm)
 111                 return;
 112         PRIV(dev)->pending_alarm = 0;
 113         FOREACH_SUBDEV(sdev, i, dev)
 114                 if (sdev->state != PRIV(dev)->state)
 115                         break;
 116         /* if we have non-probed device */
 117         if (i != PRIV(dev)->subs_tail) {
 118                 if (fs_lock(dev, 1) != 0)
 119                         goto reinstall;
 120                 ret = failsafe_eth_dev_state_sync(dev);
 121                 fs_unlock(dev, 1);
 122                 if (ret)
 123                         ERROR("Unable to synchronize sub_device state");
 124         }
 125         failsafe_dev_remove(dev);
 126 reinstall:
 127         ret = failsafe_hotplug_alarm_install(dev);
 128         if (ret)
 129                 ERROR("Unable to set up next alarm");
 130 }
 131
 132 static int
 133 fs_mutex_init(struct fs_priv *priv)
 134 {
 135         int ret;
 136         pthread_mutexattr_t attr;
 137
 138         ret = pthread_mutexattr_init(&attr);
 139         if (ret) {
 140                 ERROR("Cannot initiate mutex attributes - %s", strerror(ret));
 141                 return ret;
 142         }
 143         /* Allow mutex relocks for the thread holding the mutex. */
 144         ret = pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE);
 145         if (ret) {
 146                 ERROR("Cannot set mutex type - %s", strerror(ret));
 147                 return ret;
 148         }
 149         ret = pthread_mutex_init(&priv->hotplug_mutex, &attr);
 150         if (ret) {
 151                 ERROR("Cannot initiate mutex - %s", strerror(ret));
 152                 return ret;
 153         }
 154         return 0;
 155 }
 156
 157 static int
 158 fs_eth_dev_create(struct rte_vdev_device *vdev)
 159 {
 160         struct rte_eth_dev *dev;
 161         struct rte_ether_addr *mac;
 162         struct fs_priv *priv;
 163         struct sub_device *sdev;
 164         const char *params;
 165         unsigned int socket_id;
 166         uint8_t i;
 167         int ret;
 168
 169         dev = NULL;
 170         priv = NULL;
 171         socket_id = rte_socket_id();
 172         INFO("Creating fail-safe device on NUMA socket %u", socket_id);
 173         params = rte_vdev_device_args(vdev);
 174         if (params == NULL) {
 175                 ERROR("This PMD requires sub-devices, none provided");
 176                 return -1;
 177         }
 178         dev = rte_eth_vdev_allocate(vdev, sizeof(*priv));
 179         if (dev == NULL) {
 180                 ERROR("Unable to allocate rte_eth_dev");
 181                 return -1;
 182         }
 183         dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
 184         priv = PRIV(dev);
 185         priv->data = dev->data;
 186         priv->rxp = FS_RX_PROXY_INIT;
 187         dev->dev_ops = &failsafe_ops;
 188         dev->data->mac_addrs = &PRIV(dev)->mac_addrs[0];
 189         dev->data->dev_link = eth_link;
 190         PRIV(dev)->nb_mac_addr = 1;
 191         TAILQ_INIT(&PRIV(dev)->flow_list);
 192         dev->rx_pkt_burst = (eth_rx_burst_t)&failsafe_rx_burst;
 193         dev->tx_pkt_burst = (eth_tx_burst_t)&failsafe_tx_burst;
 194         ret = fs_sub_device_alloc(dev, params);
 195         if (ret) {
 196                 ERROR("Could not allocate sub_devices");
 197                 goto free_dev;
 198         }
 199         ret = failsafe_args_parse(dev, params);
 200         if (ret)
 201                 goto free_subs;
 202         ret = rte_eth_dev_owner_new(&priv->my_owner.id);
 203         if (ret) {
 204                 ERROR("Failed to get unique owner identifier");
 205                 goto free_args;
 206         }
 207         snprintf(priv->my_owner.name, sizeof(priv->my_owner.name),
 208                  FAILSAFE_OWNER_NAME);
 209         DEBUG("Failsafe port %u owner info: %s_%016"PRIX64, dev->data->port_id,
 210               priv->my_owner.name, priv->my_owner.id);
 211         ret = rte_eth_dev_callback_register(RTE_ETH_ALL, RTE_ETH_EVENT_NEW,
 212                                             failsafe_eth_new_event_callback,
 213                                             dev);
 214         if (ret) {
 215                 ERROR("Failed to register NEW callback");
 216                 goto free_args;
 217         }
 218         ret = failsafe_eal_init(dev);
 219         if (ret)
 220                 goto unregister_new_callback;
 221         ret = fs_mutex_init(priv);
 222         if (ret)
 223                 goto unregister_new_callback;
 224         ret = failsafe_hotplug_alarm_install(dev);
 225         if (ret) {
 226                 ERROR("Could not set up plug-in event detection");
 227                 goto unregister_new_callback;
 228         }
 229         mac = &dev->data->mac_addrs[0];
 230         if (failsafe_mac_from_arg) {
 231                 /*
 232                  * If MAC address was provided as a parameter,
 233                  * apply to all probed slaves.
 234                  */
 235                 FOREACH_SUBDEV_STATE(sdev, i, dev, DEV_PROBED) {
 236                         ret = rte_eth_dev_default_mac_addr_set(PORT_ID(sdev),
 237                                                                mac);
 238                         if (ret) {
 239                                 ERROR("Failed to set default MAC address");
 240                                 goto cancel_alarm;
 241                         }
 242                 }
 243         } else {
 244                 /*
 245                  * Use the ether_addr from first probed
 246                  * device, either preferred or fallback.
 247                  */
 248                 FOREACH_SUBDEV(sdev, i, dev)
 249                         if (sdev->state >= DEV_PROBED) {
 250                                 rte_ether_addr_copy(
 251                                         &ETH(sdev)->data->mac_addrs[0], mac);
 252                                 break;
 253                         }
 254                 /*
 255                  * If no device has been probed and no ether_addr
 256                  * has been provided on the command line, use a random
 257                  * valid one.
 258                  * It will be applied during future slave state syncs to
 259                  * probed slaves.
 260                  */
 261                 if (i == priv->subs_tail)
 262                         rte_eth_random_addr(&mac->addr_bytes[0]);
 263         }
 264         INFO("MAC address is %02x:%02x:%02x:%02x:%02x:%02x",
 265                 mac->addr_bytes[0], mac->addr_bytes[1],
 266                 mac->addr_bytes[2], mac->addr_bytes[3],
 267                 mac->addr_bytes[4], mac->addr_bytes[5]);
 268         dev->data->dev_flags |= RTE_ETH_DEV_INTR_LSC;
 269         PRIV(dev)->intr_handle = (struct rte_intr_handle){
 270                 .fd = -1,
 271                 .type = RTE_INTR_HANDLE_EXT,
 272         };
 273         rte_eth_dev_probing_finish(dev);
 274         return 0;
 275 cancel_alarm:
 276         failsafe_hotplug_alarm_cancel(dev);
 277 unregister_new_callback:
 278         rte_eth_dev_callback_unregister(RTE_ETH_ALL, RTE_ETH_EVENT_NEW,
 279                                         failsafe_eth_new_event_callback, dev);
 280 free_args:
 281         failsafe_args_free(dev);
 282 free_subs:
 283         rte_free(PRIV(dev)->subs);
 284 free_dev:
 285         /* mac_addrs must not be freed alone because part of dev_private */
 286         dev->data->mac_addrs = NULL;
 287         rte_eth_dev_release_port(dev);
 288         return -1;
 289 }
 290
 291 static int
 292 fs_rte_eth_free(const char *name)
 293 {
 294         struct rte_eth_dev *dev;
 295         int ret;
 296
 297         dev = rte_eth_dev_allocated(name);
 298         if (dev == NULL)
 299                 return 0; /* port already released */
 300         ret = failsafe_eth_dev_close(dev);
 301         rte_eth_dev_release_port(dev);
 302         return ret;
 303 }
 304
 305 static bool
 306 devargs_already_listed(struct rte_devargs *devargs)
 307 {
 308         struct rte_devargs *list_da;
 309
 310         RTE_EAL_DEVARGS_FOREACH(devargs->bus->name, list_da) {
 311                 if (strcmp(list_da->name, devargs->name) == 0)
 312                         /* devargs already in the list */
 313                         return true;
 314         }
 315         return false;
 316 }
 317
 318 static int
 319 rte_pmd_failsafe_probe(struct rte_vdev_device *vdev)
 320 {
 321         const char *name;
 322         struct rte_eth_dev *eth_dev;
 323         struct sub_device  *sdev;
 324         struct rte_devargs devargs;
 325         uint8_t i;
 326         int ret;
 327
 328         name = rte_vdev_device_name(vdev);
 329         INFO("Initializing " FAILSAFE_DRIVER_NAME " for %s",
 330                         name);
 331
 332         if (rte_eal_process_type() == RTE_PROC_SECONDARY &&
 333             strlen(rte_vdev_device_args(vdev)) == 0) {
 334                 eth_dev = rte_eth_dev_attach_secondary(name);
 335                 if (!eth_dev) {
 336                         ERROR("Failed to probe %s", name);
 337                         return -1;
 338                 }
 339                 eth_dev->dev_ops = &failsafe_ops;
 340                 eth_dev->device = &vdev->device;
 341                 eth_dev->rx_pkt_burst = (eth_rx_burst_t)&failsafe_rx_burst;
 342                 eth_dev->tx_pkt_burst = (eth_tx_burst_t)&failsafe_tx_burst;
 343                 /*
 344                  * Failsafe will attempt to probe all of its sub-devices.
 345                  * Any failure in sub-devices is not a fatal error.
 346                  * A sub-device can be plugged later.
 347                  */
 348                 FOREACH_SUBDEV(sdev, i, eth_dev) {
 349                         /* skip empty devargs */
 350                         if (sdev->devargs.name[0] == '\0')
 351                                 continue;
 352
 353                         /* rebuild devargs to be able to get the bus name. */
 354                         ret = rte_devargs_parse(&devargs,
 355                                                 sdev->devargs.name);
 356                         if (ret != 0) {
 357                                 ERROR("Failed to parse devargs %s",
 358                                         devargs.name);
 359                                 continue;
 360                         }
 361                         if (!devargs_already_listed(&devargs)) {
 362                                 ret = rte_dev_probe(devargs.name);
 363                                 if (ret < 0) {
 364                                         ERROR("Failed to probe devargs %s",
 365                                               devargs.name);
 366                                         continue;
 367                                 }
 368                         }
 369                 }
 370                 rte_eth_dev_probing_finish(eth_dev);
 371                 return 0;
 372         }
 373
 374         return fs_eth_dev_create(vdev);
 375 }
 376
 377 static int
 378 rte_pmd_failsafe_remove(struct rte_vdev_device *vdev)
 379 {
 380         const char *name;
 381
 382         name = rte_vdev_device_name(vdev);
 383         INFO("Uninitializing " FAILSAFE_DRIVER_NAME " for %s", name);
 384         return fs_rte_eth_free(name);
 385 }
 386
 387 static struct rte_vdev_driver failsafe_drv = {
 388         .probe = rte_pmd_failsafe_probe,
 389         .remove = rte_pmd_failsafe_remove,
 390 };
 391
 392 RTE_PMD_REGISTER_VDEV(net_failsafe, failsafe_drv);
 393 RTE_PMD_REGISTER_PARAM_STRING(net_failsafe, PMD_FAILSAFE_PARAM_STRING);
 394 RTE_LOG_REGISTER(failsafe_logtype, pmd.net.failsafe, NOTICE)