lib/eal/linux/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;
  27 static rte_rwlock_t monitor_lock = RTE_RWLOCK_INITIALIZER;
  28 static uint32_t monitor_refcount;
  29 static bool hotplug_handle;
  30
  31 #define EAL_UEV_MSG_LEN 4096
  32 #define EAL_UEV_MSG_ELEM_LEN 128
  33
  34 /*
  35  * spinlock for device hot-unplug failure handling. If it try to access bus or
  36  * device, such as handle sigbus on bus or handle memory failure for device
  37  * just need to use this lock. It could protect the bus and the device to avoid
  38  * race condition.
  39  */
  40 static rte_spinlock_t failure_handle_lock = RTE_SPINLOCK_INITIALIZER;
  41
  42 static struct sigaction sigbus_action_old;
  43
  44 static int sigbus_need_recover;
  45
  46 static void dev_uev_handler(__rte_unused void *param);
  47
  48 /* identify the system layer which reports this event. */
  49 enum eal_dev_event_subsystem {
  50         EAL_DEV_EVENT_SUBSYSTEM_PCI, /* PCI bus device event */
  51         EAL_DEV_EVENT_SUBSYSTEM_UIO, /* UIO driver device event */
  52         EAL_DEV_EVENT_SUBSYSTEM_VFIO, /* VFIO driver device event */
  53         EAL_DEV_EVENT_SUBSYSTEM_MAX
  54 };
  55
  56 static void
  57 sigbus_action_recover(void)
  58 {
  59         if (sigbus_need_recover) {
  60                 sigaction(SIGBUS, &sigbus_action_old, NULL);
  61                 sigbus_need_recover = 0;
  62         }
  63 }
  64
  65 static void sigbus_handler(int signum, siginfo_t *info,
  66                                 void *ctx __rte_unused)
  67 {
  68         int ret;
  69
  70         RTE_LOG(DEBUG, EAL, "Thread catch SIGBUS, fault address:%p\n",
  71                 info->si_addr);
  72
  73         rte_spinlock_lock(&failure_handle_lock);
  74         ret = rte_bus_sigbus_handler(info->si_addr);
  75         rte_spinlock_unlock(&failure_handle_lock);
  76         if (ret == -1) {
  77                 rte_exit(EXIT_FAILURE,
  78                          "Failed to handle SIGBUS for hot-unplug, "
  79                          "(rte_errno: %s)!", strerror(rte_errno));
  80         } else if (ret == 1) {
  81                 if (sigbus_action_old.sa_flags == SA_SIGINFO
  82                     && sigbus_action_old.sa_sigaction) {
  83                         (*(sigbus_action_old.sa_sigaction))(signum,
  84                                                             info, ctx);
  85                 } else if (sigbus_action_old.sa_flags != SA_SIGINFO
  86                            && sigbus_action_old.sa_handler) {
  87                         (*(sigbus_action_old.sa_handler))(signum);
  88                 } else {
  89                         rte_exit(EXIT_FAILURE,
  90                                  "Failed to handle generic SIGBUS!");
  91                 }
  92         }
  93
  94         RTE_LOG(DEBUG, EAL, "Success to handle SIGBUS for hot-unplug!\n");
  95 }
  96
  97 static int cmp_dev_name(const struct rte_device *dev,
  98         const void *_name)
  99 {
 100         const char *name = _name;
 101
 102         return strcmp(dev->name, name);
 103 }
 104
 105 static int
 106 dev_uev_socket_fd_create(void)
 107 {
 108         struct sockaddr_nl addr;
 109         int ret, fd;
 110
 111         fd = socket(PF_NETLINK, SOCK_RAW | SOCK_CLOEXEC | SOCK_NONBLOCK,
 112                     NETLINK_KOBJECT_UEVENT);
 113         if (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(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         if (rte_intr_fd_set(intr_handle, fd))
 130                 goto err;
 131
 132         return 0;
 133 err:
 134         close(fd);
 135         fd = -1;
 136         return ret;
 137 }
 138
 139 struct rte_dev_event {
 140         enum rte_dev_event_type type;   /**< device event type */
 141         int subsystem;                  /**< subsystem id */
 142         char *devname;                  /**< device name */
 143 };
 144
 145 static int
 146 dev_uev_parse(const char *buf, struct rte_dev_event *event, int length)
 147 {
 148         char action[EAL_UEV_MSG_ELEM_LEN];
 149         char subsystem[EAL_UEV_MSG_ELEM_LEN];
 150         char pci_slot_name[EAL_UEV_MSG_ELEM_LEN];
 151         int i = 0;
 152
 153         memset(action, 0, EAL_UEV_MSG_ELEM_LEN);
 154         memset(subsystem, 0, EAL_UEV_MSG_ELEM_LEN);
 155         memset(pci_slot_name, 0, EAL_UEV_MSG_ELEM_LEN);
 156
 157         while (i < length) {
 158                 for (; i < length; i++) {
 159                         if (*buf)
 160                                 break;
 161                         buf++;
 162                 }
 163                 if (i >= length)
 164                         break;
 165
 166                 /**
 167                  * check device uevent from kernel side, no need to check
 168                  * uevent from udev.
 169                  */
 170                 if (!strncmp(buf, "libudev", 7)) {
 171                         buf += 7;
 172                         i += 7;
 173                         return -1;
 174                 }
 175                 if (!strncmp(buf, "ACTION=", 7)) {
 176                         buf += 7;
 177                         i += 7;
 178                         strlcpy(action, buf, sizeof(action));
 179                 } else if (!strncmp(buf, "SUBSYSTEM=", 10)) {
 180                         buf += 10;
 181                         i += 10;
 182                         strlcpy(subsystem, buf, sizeof(subsystem));
 183                 } else if (!strncmp(buf, "PCI_SLOT_NAME=", 14)) {
 184                         buf += 14;
 185                         i += 14;
 186                         strlcpy(pci_slot_name, buf, sizeof(subsystem));
 187                         event->devname = strdup(pci_slot_name);
 188                 }
 189                 for (; i < length; i++) {
 190                         if (*buf == '\0')
 191                                 break;
 192                         buf++;
 193                 }
 194         }
 195
 196         /* parse the subsystem layer */
 197         if (!strncmp(subsystem, "uio", 3))
 198                 event->subsystem = EAL_DEV_EVENT_SUBSYSTEM_UIO;
 199         else if (!strncmp(subsystem, "pci", 3))
 200                 event->subsystem = EAL_DEV_EVENT_SUBSYSTEM_PCI;
 201         else if (!strncmp(subsystem, "vfio", 4))
 202                 event->subsystem = EAL_DEV_EVENT_SUBSYSTEM_VFIO;
 203         else
 204                 goto err;
 205
 206         /* parse the action type */
 207         if (!strncmp(action, "add", 3))
 208                 event->type = RTE_DEV_EVENT_ADD;
 209         else if (!strncmp(action, "remove", 6))
 210                 event->type = RTE_DEV_EVENT_REMOVE;
 211         else
 212                 goto err;
 213         return 0;
 214 err:
 215         free(event->devname);
 216         return -1;
 217 }
 218
 219 static void
 220 dev_delayed_unregister(void *param)
 221 {
 222         rte_intr_callback_unregister(intr_handle, dev_uev_handler, param);
 223         if (rte_intr_fd_get(intr_handle) >= 0) {
 224                 close(rte_intr_fd_get(intr_handle));
 225                 rte_intr_fd_set(intr_handle, -1);
 226         }
 227 }
 228
 229 static void
 230 dev_uev_handler(__rte_unused void *param)
 231 {
 232         struct rte_dev_event uevent;
 233         int ret;
 234         char buf[EAL_UEV_MSG_LEN];
 235         struct rte_bus *bus;
 236         struct rte_device *dev;
 237         const char *busname = "";
 238
 239         memset(&uevent, 0, sizeof(struct rte_dev_event));
 240         memset(buf, 0, EAL_UEV_MSG_LEN);
 241
 242         if (rte_intr_fd_get(intr_handle) < 0)
 243                 return;
 244
 245         ret = recv(rte_intr_fd_get(intr_handle), buf, EAL_UEV_MSG_LEN,
 246                    MSG_DONTWAIT);
 247         if (ret < 0 && errno == EAGAIN)
 248                 return;
 249         else if (ret <= 0) {
 250                 /* connection is closed or broken, can not up again. */
 251                 RTE_LOG(ERR, EAL, "uevent socket connection is broken.\n");
 252                 rte_eal_alarm_set(1, dev_delayed_unregister, NULL);
 253                 return;
 254         }
 255
 256         ret = dev_uev_parse(buf, &uevent, EAL_UEV_MSG_LEN);
 257         if (ret < 0) {
 258                 RTE_LOG(DEBUG, EAL, "Ignoring uevent '%s'\n", buf);
 259                 return;
 260         }
 261
 262         RTE_LOG(DEBUG, EAL, "receive uevent(name:%s, type:%d, subsystem:%d)\n",
 263                 uevent.devname, uevent.type, uevent.subsystem);
 264
 265         switch (uevent.subsystem) {
 266         case EAL_DEV_EVENT_SUBSYSTEM_PCI:
 267         case EAL_DEV_EVENT_SUBSYSTEM_UIO:
 268                 busname = "pci";
 269                 break;
 270         default:
 271                 break;
 272         }
 273
 274         if (uevent.devname) {
 275                 if (uevent.type == RTE_DEV_EVENT_REMOVE && hotplug_handle) {
 276                         rte_spinlock_lock(&failure_handle_lock);
 277                         bus = rte_bus_find_by_name(busname);
 278                         if (bus == NULL) {
 279                                 RTE_LOG(ERR, EAL, "Cannot find bus (%s)\n",
 280                                         busname);
 281                                 goto failure_handle_err;
 282                         }
 283
 284                         dev = bus->find_device(NULL, cmp_dev_name,
 285                                                uevent.devname);
 286                         if (dev == NULL) {
 287                                 RTE_LOG(ERR, EAL, "Cannot find device (%s) on "
 288                                         "bus (%s)\n", uevent.devname, busname);
 289                                 goto failure_handle_err;
 290                         }
 291
 292                         ret = bus->hot_unplug_handler(dev);
 293                         if (ret) {
 294                                 RTE_LOG(ERR, EAL, "Can not handle hot-unplug "
 295                                         "for device (%s)\n", dev->name);
 296                         }
 297                         rte_spinlock_unlock(&failure_handle_lock);
 298                 }
 299                 rte_dev_event_callback_process(uevent.devname, uevent.type);
 300                 free(uevent.devname);
 301         }
 302
 303         return;
 304
 305 failure_handle_err:
 306         rte_spinlock_unlock(&failure_handle_lock);
 307         free(uevent.devname);
 308 }
 309
 310 int
 311 rte_dev_event_monitor_start(void)
 312 {
 313         int ret = 0;
 314
 315         rte_rwlock_write_lock(&monitor_lock);
 316
 317         if (monitor_refcount) {
 318                 monitor_refcount++;
 319                 goto exit;
 320         }
 321
 322         intr_handle = rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_PRIVATE);
 323         if (intr_handle == NULL) {
 324                 RTE_LOG(ERR, EAL, "Fail to allocate intr_handle\n");
 325                 goto exit;
 326         }
 327
 328         ret = rte_intr_type_set(intr_handle, RTE_INTR_HANDLE_DEV_EVENT);
 329         if (ret)
 330                 goto exit;
 331
 332         ret = rte_intr_fd_set(intr_handle, -1);
 333         if (ret)
 334                 goto exit;
 335
 336         ret = dev_uev_socket_fd_create();
 337         if (ret) {
 338                 RTE_LOG(ERR, EAL, "error create device event fd.\n");
 339                 goto exit;
 340         }
 341
 342         ret = rte_intr_callback_register(intr_handle, dev_uev_handler, NULL);
 343
 344         if (ret) {
 345                 close(rte_intr_fd_get(intr_handle));
 346                 goto exit;
 347         }
 348
 349         monitor_refcount++;
 350
 351 exit:
 352         if (ret) {
 353                 rte_intr_instance_free(intr_handle);
 354                 intr_handle = NULL;
 355         }
 356         rte_rwlock_write_unlock(&monitor_lock);
 357         return ret;
 358 }
 359
 360 int
 361 rte_dev_event_monitor_stop(void)
 362 {
 363         int ret = 0;
 364
 365         rte_rwlock_write_lock(&monitor_lock);
 366
 367         if (!monitor_refcount) {
 368                 RTE_LOG(ERR, EAL, "device event monitor already stopped\n");
 369                 goto exit;
 370         }
 371
 372         if (monitor_refcount > 1) {
 373                 monitor_refcount--;
 374                 goto exit;
 375         }
 376
 377         ret = rte_intr_callback_unregister(intr_handle, dev_uev_handler,
 378                                            (void *)-1);
 379         if (ret < 0) {
 380                 RTE_LOG(ERR, EAL, "fail to unregister uevent callback.\n");
 381                 goto exit;
 382         }
 383
 384         close(rte_intr_fd_get(intr_handle));
 385         rte_intr_instance_free(intr_handle);
 386         intr_handle = NULL;
 387
 388         monitor_refcount--;
 389
 390 exit:
 391         rte_rwlock_write_unlock(&monitor_lock);
 392
 393         return ret;
 394 }
 395
 396 int
 397 dev_sigbus_handler_register(void)
 398 {
 399         sigset_t mask;
 400         struct sigaction action;
 401
 402         rte_errno = 0;
 403
 404         if (sigbus_need_recover)
 405                 return 0;
 406
 407         sigemptyset(&mask);
 408         sigaddset(&mask, SIGBUS);
 409         action.sa_flags = SA_SIGINFO;
 410         action.sa_mask = mask;
 411         action.sa_sigaction = sigbus_handler;
 412         sigbus_need_recover = !sigaction(SIGBUS, &action, &sigbus_action_old);
 413
 414         return rte_errno;
 415 }
 416
 417 int
 418 dev_sigbus_handler_unregister(void)
 419 {
 420         rte_errno = 0;
 421
 422         sigbus_action_recover();
 423
 424         return rte_errno;
 425 }
 426
 427 int
 428 rte_dev_hotplug_handle_enable(void)
 429 {
 430         int ret = 0;
 431
 432         ret = dev_sigbus_handler_register();
 433         if (ret < 0)
 434                 RTE_LOG(ERR, EAL,
 435                         "fail to register sigbus handler for devices.\n");
 436
 437         hotplug_handle = true;
 438
 439         return ret;
 440 }
 441
 442 int
 443 rte_dev_hotplug_handle_disable(void)
 444 {
 445         int ret = 0;
 446
 447         ret = dev_sigbus_handler_unregister();
 448         if (ret < 0)
 449                 RTE_LOG(ERR, EAL,
 450                         "fail to unregister sigbus handler for devices.\n");
 451
 452         hotplug_handle = false;
 453
 454         return ret;
 455 }