lib/librte_eal/linux/eal/eal_dev.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2018 Intel Corporation
   3  */
   4
   5 #include <string.h>
   6 #include <unistd.h>
   7 #include <fcntl.h>
   8 #include <signal.h>
   9 #include <sys/socket.h>
  10 #include <linux/netlink.h>
  11
  12 #include <rte_string_fns.h>
  13 #include <rte_log.h>
  14 #include <rte_compat.h>
  15 #include <rte_dev.h>
  16 #include <rte_malloc.h>
  17 #include <rte_interrupts.h>
  18 #include <rte_alarm.h>
  19 #include <rte_bus.h>
  20 #include <rte_eal.h>
  21 #include <rte_spinlock.h>
  22 #include <rte_errno.h>
  23
  24 #include "eal_private.h"
  25
  26 static struct rte_intr_handle intr_handle = {.fd = -1 };
  27 static bool monitor_started;
  28 static bool hotplug_handle;
  29
  30 #define EAL_UEV_MSG_LEN 4096
  31 #define EAL_UEV_MSG_ELEM_LEN 128
  32
  33 /*
  34  * spinlock for device hot-unplug failure handling. If it try to access bus or
  35  * device, such as handle sigbus on bus or handle memory failure for device
  36  * just need to use this lock. It could protect the bus and the device to avoid
  37  * race condition.
  38  */
  39 static rte_spinlock_t failure_handle_lock = RTE_SPINLOCK_INITIALIZER;
  40
  41 static struct sigaction sigbus_action_old;
  42
  43 static int sigbus_need_recover;
  44
  45 static void dev_uev_handler(__rte_unused void *param);
  46
  47 /* identify the system layer which reports this event. */
  48 enum eal_dev_event_subsystem {
  49         EAL_DEV_EVENT_SUBSYSTEM_PCI, /* PCI bus device event */
  50         EAL_DEV_EVENT_SUBSYSTEM_UIO, /* UIO driver device event */
  51         EAL_DEV_EVENT_SUBSYSTEM_VFIO, /* VFIO driver device event */
  52         EAL_DEV_EVENT_SUBSYSTEM_MAX
  53 };
  54
  55 static void
  56 sigbus_action_recover(void)
  57 {
  58         if (sigbus_need_recover) {
  59                 sigaction(SIGBUS, &sigbus_action_old, NULL);
  60                 sigbus_need_recover = 0;
  61         }
  62 }
  63
  64 static void sigbus_handler(int signum, siginfo_t *info,
  65                                 void *ctx __rte_unused)
  66 {
  67         int ret;
  68
  69         RTE_LOG(DEBUG, EAL, "Thread catch SIGBUS, fault address:%p\n",
  70                 info->si_addr);
  71
  72         rte_spinlock_lock(&failure_handle_lock);
  73         ret = rte_bus_sigbus_handler(info->si_addr);
  74         rte_spinlock_unlock(&failure_handle_lock);
  75         if (ret == -1) {
  76                 rte_exit(EXIT_FAILURE,
  77                          "Failed to handle SIGBUS for hot-unplug, "
  78                          "(rte_errno: %s)!", strerror(rte_errno));
  79         } else if (ret == 1) {
  80                 if (sigbus_action_old.sa_flags == SA_SIGINFO
  81                     && sigbus_action_old.sa_sigaction) {
  82                         (*(sigbus_action_old.sa_sigaction))(signum,
  83                                                             info, ctx);
  84                 } else if (sigbus_action_old.sa_flags != SA_SIGINFO
  85                            && sigbus_action_old.sa_handler) {
  86                         (*(sigbus_action_old.sa_handler))(signum);
  87                 } else {
  88                         rte_exit(EXIT_FAILURE,
  89                                  "Failed to handle generic SIGBUS!");
  90                 }
  91         }
  92
  93         RTE_LOG(DEBUG, EAL, "Success to handle SIGBUS for hot-unplug!\n");
  94 }
  95
  96 static int cmp_dev_name(const struct rte_device *dev,
  97         const void *_name)
  98 {
  99         const char *name = _name;
 100
 101         return strcmp(dev->name, name);
 102 }
 103
 104 static int
 105 dev_uev_socket_fd_create(void)
 106 {
 107         struct sockaddr_nl addr;
 108         int ret;
 109
 110         intr_handle.fd = socket(PF_NETLINK, SOCK_RAW | SOCK_CLOEXEC |
 111                         SOCK_NONBLOCK,
 112                         NETLINK_KOBJECT_UEVENT);
 113         if (intr_handle.fd < 0) {
 114                 RTE_LOG(ERR, EAL, "create uevent fd failed.\n");
 115                 return -1;
 116         }
 117
 118         memset(&addr, 0, sizeof(addr));
 119         addr.nl_family = AF_NETLINK;
 120         addr.nl_pid = 0;
 121         addr.nl_groups = 0xffffffff;
 122
 123         ret = bind(intr_handle.fd, (struct sockaddr *) &addr, sizeof(addr));
 124         if (ret < 0) {
 125                 RTE_LOG(ERR, EAL, "Failed to bind uevent socket.\n");
 126                 goto err;
 127         }
 128
 129         return 0;
 130 err:
 131         close(intr_handle.fd);
 132         intr_handle.fd = -1;
 133         return ret;
 134 }
 135
 136 static int
 137 dev_uev_parse(const char *buf, struct rte_dev_event *event, int length)
 138 {
 139         char action[EAL_UEV_MSG_ELEM_LEN];
 140         char subsystem[EAL_UEV_MSG_ELEM_LEN];
 141         char pci_slot_name[EAL_UEV_MSG_ELEM_LEN];
 142         int i = 0;
 143
 144         memset(action, 0, EAL_UEV_MSG_ELEM_LEN);
 145         memset(subsystem, 0, EAL_UEV_MSG_ELEM_LEN);
 146         memset(pci_slot_name, 0, EAL_UEV_MSG_ELEM_LEN);
 147
 148         while (i < length) {
 149                 for (; i < length; i++) {
 150                         if (*buf)
 151                                 break;
 152                         buf++;
 153                 }
 154                 /**
 155                  * check device uevent from kernel side, no need to check
 156                  * uevent from udev.
 157                  */
 158                 if (!strncmp(buf, "libudev", 7)) {
 159                         buf += 7;
 160                         i += 7;
 161                         return -1;
 162                 }
 163                 if (!strncmp(buf, "ACTION=", 7)) {
 164                         buf += 7;
 165                         i += 7;
 166                         strlcpy(action, buf, sizeof(action));
 167                 } else if (!strncmp(buf, "SUBSYSTEM=", 10)) {
 168                         buf += 10;
 169                         i += 10;
 170                         strlcpy(subsystem, buf, sizeof(subsystem));
 171                 } else if (!strncmp(buf, "PCI_SLOT_NAME=", 14)) {
 172                         buf += 14;
 173                         i += 14;
 174                         strlcpy(pci_slot_name, buf, sizeof(subsystem));
 175                         event->devname = strdup(pci_slot_name);
 176                 }
 177                 for (; i < length; i++) {
 178                         if (*buf == '\0')
 179                                 break;
 180                         buf++;
 181                 }
 182         }
 183
 184         /* parse the subsystem layer */
 185         if (!strncmp(subsystem, "uio", 3))
 186                 event->subsystem = EAL_DEV_EVENT_SUBSYSTEM_UIO;
 187         else if (!strncmp(subsystem, "pci", 3))
 188                 event->subsystem = EAL_DEV_EVENT_SUBSYSTEM_PCI;
 189         else if (!strncmp(subsystem, "vfio", 4))
 190                 event->subsystem = EAL_DEV_EVENT_SUBSYSTEM_VFIO;
 191         else
 192                 return -1;
 193
 194         /* parse the action type */
 195         if (!strncmp(action, "add", 3))
 196                 event->type = RTE_DEV_EVENT_ADD;
 197         else if (!strncmp(action, "remove", 6))
 198                 event->type = RTE_DEV_EVENT_REMOVE;
 199         else
 200                 return -1;
 201         return 0;
 202 }
 203
 204 static void
 205 dev_delayed_unregister(void *param)
 206 {
 207         rte_intr_callback_unregister(&intr_handle, dev_uev_handler, param);
 208         close(intr_handle.fd);
 209         intr_handle.fd = -1;
 210 }
 211
 212 static void
 213 dev_uev_handler(__rte_unused void *param)
 214 {
 215         struct rte_dev_event uevent;
 216         int ret;
 217         char buf[EAL_UEV_MSG_LEN];
 218         struct rte_bus *bus;
 219         struct rte_device *dev;
 220         const char *busname = "";
 221
 222         memset(&uevent, 0, sizeof(struct rte_dev_event));
 223         memset(buf, 0, EAL_UEV_MSG_LEN);
 224
 225         ret = recv(intr_handle.fd, buf, EAL_UEV_MSG_LEN, MSG_DONTWAIT);
 226         if (ret < 0 && errno == EAGAIN)
 227                 return;
 228         else if (ret <= 0) {
 229                 /* connection is closed or broken, can not up again. */
 230                 RTE_LOG(ERR, EAL, "uevent socket connection is broken.\n");
 231                 rte_eal_alarm_set(1, dev_delayed_unregister, NULL);
 232                 return;
 233         }
 234
 235         ret = dev_uev_parse(buf, &uevent, EAL_UEV_MSG_LEN);
 236         if (ret < 0) {
 237                 RTE_LOG(DEBUG, EAL, "It is not an valid event "
 238                         "that need to be handle.\n");
 239                 return;
 240         }
 241
 242         RTE_LOG(DEBUG, EAL, "receive uevent(name:%s, type:%d, subsystem:%d)\n",
 243                 uevent.devname, uevent.type, uevent.subsystem);
 244
 245         switch (uevent.subsystem) {
 246         case EAL_DEV_EVENT_SUBSYSTEM_PCI:
 247         case EAL_DEV_EVENT_SUBSYSTEM_UIO:
 248                 busname = "pci";
 249                 break;
 250         default:
 251                 break;
 252         }
 253
 254         if (uevent.devname) {
 255                 if (uevent.type == RTE_DEV_EVENT_REMOVE && hotplug_handle) {
 256                         rte_spinlock_lock(&failure_handle_lock);
 257                         bus = rte_bus_find_by_name(busname);
 258                         if (bus == NULL) {
 259                                 RTE_LOG(ERR, EAL, "Cannot find bus (%s)\n",
 260                                         busname);
 261                                 goto failure_handle_err;
 262                         }
 263
 264                         dev = bus->find_device(NULL, cmp_dev_name,
 265                                                uevent.devname);
 266                         if (dev == NULL) {
 267                                 RTE_LOG(ERR, EAL, "Cannot find device (%s) on "
 268                                         "bus (%s)\n", uevent.devname, busname);
 269                                 goto failure_handle_err;
 270                         }
 271
 272                         ret = bus->hot_unplug_handler(dev);
 273                         if (ret) {
 274                                 RTE_LOG(ERR, EAL, "Can not handle hot-unplug "
 275                                         "for device (%s)\n", dev->name);
 276                         }
 277                         rte_spinlock_unlock(&failure_handle_lock);
 278                 }
 279                 rte_dev_event_callback_process(uevent.devname, uevent.type);
 280         }
 281
 282         return;
 283
 284 failure_handle_err:
 285         rte_spinlock_unlock(&failure_handle_lock);
 286 }
 287
 288 int __rte_experimental
 289 rte_dev_event_monitor_start(void)
 290 {
 291         int ret;
 292
 293         if (monitor_started)
 294                 return 0;
 295
 296         ret = dev_uev_socket_fd_create();
 297         if (ret) {
 298                 RTE_LOG(ERR, EAL, "error create device event fd.\n");
 299                 return -1;
 300         }
 301
 302         intr_handle.type = RTE_INTR_HANDLE_DEV_EVENT;
 303         ret = rte_intr_callback_register(&intr_handle, dev_uev_handler, NULL);
 304
 305         if (ret) {
 306                 RTE_LOG(ERR, EAL, "fail to register uevent callback.\n");
 307                 return -1;
 308         }
 309
 310         monitor_started = true;
 311
 312         return 0;
 313 }
 314
 315 int __rte_experimental
 316 rte_dev_event_monitor_stop(void)
 317 {
 318         int ret;
 319
 320         if (!monitor_started)
 321                 return 0;
 322
 323         ret = rte_intr_callback_unregister(&intr_handle, dev_uev_handler,
 324                                            (void *)-1);
 325         if (ret < 0) {
 326                 RTE_LOG(ERR, EAL, "fail to unregister uevent callback.\n");
 327                 return ret;
 328         }
 329
 330         close(intr_handle.fd);
 331         intr_handle.fd = -1;
 332         monitor_started = false;
 333
 334         return 0;
 335 }
 336
 337 int
 338 dev_sigbus_handler_register(void)
 339 {
 340         sigset_t mask;
 341         struct sigaction action;
 342
 343         rte_errno = 0;
 344
 345         if (sigbus_need_recover)
 346                 return 0;
 347
 348         sigemptyset(&mask);
 349         sigaddset(&mask, SIGBUS);
 350         action.sa_flags = SA_SIGINFO;
 351         action.sa_mask = mask;
 352         action.sa_sigaction = sigbus_handler;
 353         sigbus_need_recover = !sigaction(SIGBUS, &action, &sigbus_action_old);
 354
 355         return rte_errno;
 356 }
 357
 358 int
 359 dev_sigbus_handler_unregister(void)
 360 {
 361         rte_errno = 0;
 362
 363         sigbus_action_recover();
 364
 365         return rte_errno;
 366 }
 367
 368 int __rte_experimental
 369 rte_dev_hotplug_handle_enable(void)
 370 {
 371         int ret = 0;
 372
 373         ret = dev_sigbus_handler_register();
 374         if (ret < 0)
 375                 RTE_LOG(ERR, EAL,
 376                         "fail to register sigbus handler for devices.\n");
 377
 378         hotplug_handle = true;
 379
 380         return ret;
 381 }
 382
 383 int __rte_experimental
 384 rte_dev_hotplug_handle_disable(void)
 385 {
 386         int ret = 0;
 387
 388         ret = dev_sigbus_handler_unregister();
 389         if (ret < 0)
 390                 RTE_LOG(ERR, EAL,
 391                         "fail to unregister sigbus handler for devices.\n");
 392
 393         hotplug_handle = false;
 394
 395         return ret;
 396 }