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 <signal.h>
   8 #include <sys/socket.h>
   9 #include <linux/netlink.h>
  10
  11 #include <rte_string_fns.h>
  12 #include <rte_log.h>
  13 #include <rte_dev.h>
  14 #include <rte_interrupts.h>
  15 #include <rte_alarm.h>
  16 #include <rte_bus.h>
  17 #include <rte_spinlock.h>
  18 #include <rte_errno.h>
  19
  20 #include "eal_private.h"
  21
  22 static struct rte_intr_handle *intr_handle;
  23 static rte_rwlock_t monitor_lock = RTE_RWLOCK_INITIALIZER;
  24 static uint32_t monitor_refcount;
  25 static bool hotplug_handle;
  26
  27 #define EAL_UEV_MSG_LEN 4096
  28 #define EAL_UEV_MSG_ELEM_LEN 128
  29
  30 /*
  31  * spinlock for device hot-unplug failure handling. If it try to access bus or
  32  * device, such as handle sigbus on bus or handle memory failure for device
  33  * just need to use this lock. It could protect the bus and the device to avoid
  34  * race condition.
  35  */
  36 static rte_spinlock_t failure_handle_lock = RTE_SPINLOCK_INITIALIZER;
  37
  38 static struct sigaction sigbus_action_old;
  39
  40 static int sigbus_need_recover;
  41
  42 static void dev_uev_handler(__rte_unused void *param);
  43
  44 /* identify the system layer which reports this event. */
  45 enum eal_dev_event_subsystem {
  46         EAL_DEV_EVENT_SUBSYSTEM_PCI, /* PCI bus device event */
  47         EAL_DEV_EVENT_SUBSYSTEM_UIO, /* UIO driver device event */
  48         EAL_DEV_EVENT_SUBSYSTEM_VFIO, /* VFIO driver device event */
  49         EAL_DEV_EVENT_SUBSYSTEM_MAX
  50 };
  51
  52 static void
  53 sigbus_action_recover(void)
  54 {
  55         if (sigbus_need_recover) {
  56                 sigaction(SIGBUS, &sigbus_action_old, NULL);
  57                 sigbus_need_recover = 0;
  58         }
  59 }
  60
  61 static void sigbus_handler(int signum, siginfo_t *info,
  62                                 void *ctx __rte_unused)
  63 {
  64         int ret;
  65
  66         RTE_LOG(DEBUG, EAL, "Thread catch SIGBUS, fault address:%p\n",
  67                 info->si_addr);
  68
  69         rte_spinlock_lock(&failure_handle_lock);
  70         ret = rte_bus_sigbus_handler(info->si_addr);
  71         rte_spinlock_unlock(&failure_handle_lock);
  72         if (ret == -1) {
  73                 rte_exit(EXIT_FAILURE,
  74                          "Failed to handle SIGBUS for hot-unplug, "
  75                          "(rte_errno: %s)!", strerror(rte_errno));
  76         } else if (ret == 1) {
  77                 if (sigbus_action_old.sa_flags == SA_SIGINFO
  78                     && sigbus_action_old.sa_sigaction) {
  79                         (*(sigbus_action_old.sa_sigaction))(signum,
  80                                                             info, ctx);
  81                 } else if (sigbus_action_old.sa_flags != SA_SIGINFO
  82                            && sigbus_action_old.sa_handler) {
  83                         (*(sigbus_action_old.sa_handler))(signum);
  84                 } else {
  85                         rte_exit(EXIT_FAILURE,
  86                                  "Failed to handle generic SIGBUS!");
  87                 }
  88         }
  89
  90         RTE_LOG(DEBUG, EAL, "Success to handle SIGBUS for hot-unplug!\n");
  91 }
  92
  93 static int cmp_dev_name(const struct rte_device *dev,
  94         const void *_name)
  95 {
  96         const char *name = _name;
  97
  98         return strcmp(dev->name, name);
  99 }
 100
 101 static int
 102 dev_uev_socket_fd_create(void)
 103 {
 104         struct sockaddr_nl addr;
 105         int ret, fd;
 106
 107         fd = socket(PF_NETLINK, SOCK_RAW | SOCK_CLOEXEC | SOCK_NONBLOCK,
 108                     NETLINK_KOBJECT_UEVENT);
 109         if (fd < 0) {
 110                 RTE_LOG(ERR, EAL, "create uevent fd failed.\n");
 111                 return -1;
 112         }
 113
 114         memset(&addr, 0, sizeof(addr));
 115         addr.nl_family = AF_NETLINK;
 116         addr.nl_pid = 0;
 117         addr.nl_groups = 0xffffffff;
 118
 119         ret = bind(fd, (struct sockaddr *) &addr, sizeof(addr));
 120         if (ret < 0) {
 121                 RTE_LOG(ERR, EAL, "Failed to bind uevent socket.\n");
 122                 goto err;
 123         }
 124
 125         if (rte_intr_fd_set(intr_handle, fd))
 126                 goto err;
 127
 128         return 0;
 129 err:
 130         close(fd);
 131         fd = -1;
 132         return ret;
 133 }
 134
 135 struct rte_dev_event {
 136         enum rte_dev_event_type type;   /**< device event type */
 137         int subsystem;                  /**< subsystem id */
 138         char *devname;                  /**< device name */
 139 };
 140
 141 static int
 142 dev_uev_parse(const char *buf, struct rte_dev_event *event, int length)
 143 {
 144         char action[EAL_UEV_MSG_ELEM_LEN];
 145         char subsystem[EAL_UEV_MSG_ELEM_LEN];
 146         char pci_slot_name[EAL_UEV_MSG_ELEM_LEN];
 147         int i = 0;
 148
 149         memset(action, 0, EAL_UEV_MSG_ELEM_LEN);
 150         memset(subsystem, 0, EAL_UEV_MSG_ELEM_LEN);
 151         memset(pci_slot_name, 0, EAL_UEV_MSG_ELEM_LEN);
 152
 153         while (i < length) {
 154                 for (; i < length; i++) {
 155                         if (*buf)
 156                                 break;
 157                         buf++;
 158                 }
 159                 if (i >= length)
 160                         break;
 161
 162                 /**
 163                  * check device uevent from kernel side, no need to check
 164                  * uevent from udev.
 165                  */
 166                 if (!strncmp(buf, "libudev", 7)) {
 167                         buf += 7;
 168                         i += 7;
 169                         return -1;
 170                 }
 171                 if (!strncmp(buf, "ACTION=", 7)) {
 172                         buf += 7;
 173                         i += 7;
 174                         strlcpy(action, buf, sizeof(action));
 175                 } else if (!strncmp(buf, "SUBSYSTEM=", 10)) {
 176                         buf += 10;
 177                         i += 10;
 178                         strlcpy(subsystem, buf, sizeof(subsystem));
 179                 } else if (!strncmp(buf, "PCI_SLOT_NAME=", 14)) {
 180                         buf += 14;
 181                         i += 14;
 182                         strlcpy(pci_slot_name, buf, sizeof(subsystem));
 183                         event->devname = strdup(pci_slot_name);
 184                 }
 185                 for (; i < length; i++) {
 186                         if (*buf == '\0')
 187                                 break;
 188                         buf++;
 189                 }
 190         }
 191
 192         /* parse the subsystem layer */
 193         if (!strncmp(subsystem, "uio", 3))
 194                 event->subsystem = EAL_DEV_EVENT_SUBSYSTEM_UIO;
 195         else if (!strncmp(subsystem, "pci", 3))
 196                 event->subsystem = EAL_DEV_EVENT_SUBSYSTEM_PCI;
 197         else if (!strncmp(subsystem, "vfio", 4))
 198                 event->subsystem = EAL_DEV_EVENT_SUBSYSTEM_VFIO;
 199         else
 200                 goto err;
 201
 202         /* parse the action type */
 203         if (!strncmp(action, "add", 3))
 204                 event->type = RTE_DEV_EVENT_ADD;
 205         else if (!strncmp(action, "remove", 6))
 206                 event->type = RTE_DEV_EVENT_REMOVE;
 207         else
 208                 goto err;
 209         return 0;
 210 err:
 211         free(event->devname);
 212         return -1;
 213 }
 214
 215 static void
 216 dev_delayed_unregister(void *param)
 217 {
 218         rte_intr_callback_unregister(intr_handle, dev_uev_handler, param);
 219         if (rte_intr_fd_get(intr_handle) >= 0) {
 220                 close(rte_intr_fd_get(intr_handle));
 221                 rte_intr_fd_set(intr_handle, -1);
 222         }
 223 }
 224
 225 static void
 226 dev_uev_handler(__rte_unused void *param)
 227 {
 228         struct rte_dev_event uevent;
 229         int ret;
 230         char buf[EAL_UEV_MSG_LEN + 1];
 231         struct rte_bus *bus;
 232         struct rte_device *dev;
 233         const char *busname = "";
 234
 235         memset(&uevent, 0, sizeof(struct rte_dev_event));
 236         memset(buf, 0, EAL_UEV_MSG_LEN + 1);
 237
 238         if (rte_intr_fd_get(intr_handle) < 0)
 239                 return;
 240
 241         ret = recv(rte_intr_fd_get(intr_handle), buf, EAL_UEV_MSG_LEN,
 242                    MSG_DONTWAIT);
 243         if (ret < 0 && errno == EAGAIN)
 244                 return;
 245         else if (ret <= 0) {
 246                 /* connection is closed or broken, can not up again. */
 247                 RTE_LOG(ERR, EAL, "uevent socket connection is broken.\n");
 248                 rte_eal_alarm_set(1, dev_delayed_unregister, NULL);
 249                 return;
 250         }
 251
 252         ret = dev_uev_parse(buf, &uevent, EAL_UEV_MSG_LEN);
 253         if (ret < 0) {
 254                 RTE_LOG(DEBUG, EAL, "Ignoring uevent '%s'\n", buf);
 255                 return;
 256         }
 257
 258         RTE_LOG(DEBUG, EAL, "receive uevent(name:%s, type:%d, subsystem:%d)\n",
 259                 uevent.devname, uevent.type, uevent.subsystem);
 260
 261         switch (uevent.subsystem) {
 262         case EAL_DEV_EVENT_SUBSYSTEM_PCI:
 263         case EAL_DEV_EVENT_SUBSYSTEM_UIO:
 264                 busname = "pci";
 265                 break;
 266         default:
 267                 break;
 268         }
 269
 270         if (uevent.devname) {
 271                 if (uevent.type == RTE_DEV_EVENT_REMOVE && hotplug_handle) {
 272                         rte_spinlock_lock(&failure_handle_lock);
 273                         bus = rte_bus_find_by_name(busname);
 274                         if (bus == NULL) {
 275                                 RTE_LOG(ERR, EAL, "Cannot find bus (%s)\n",
 276                                         busname);
 277                                 goto failure_handle_err;
 278                         }
 279
 280                         dev = bus->find_device(NULL, cmp_dev_name,
 281                                                uevent.devname);
 282                         if (dev == NULL) {
 283                                 RTE_LOG(ERR, EAL, "Cannot find device (%s) on "
 284                                         "bus (%s)\n", uevent.devname, busname);
 285                                 goto failure_handle_err;
 286                         }
 287
 288                         ret = bus->hot_unplug_handler(dev);
 289                         if (ret) {
 290                                 RTE_LOG(ERR, EAL, "Can not handle hot-unplug "
 291                                         "for device (%s)\n", dev->name);
 292                         }
 293                         rte_spinlock_unlock(&failure_handle_lock);
 294                 }
 295                 rte_dev_event_callback_process(uevent.devname, uevent.type);
 296                 free(uevent.devname);
 297         }
 298
 299         return;
 300
 301 failure_handle_err:
 302         rte_spinlock_unlock(&failure_handle_lock);
 303         free(uevent.devname);
 304 }
 305
 306 int
 307 rte_dev_event_monitor_start(void)
 308 {
 309         int ret = 0;
 310
 311         rte_rwlock_write_lock(&monitor_lock);
 312
 313         if (monitor_refcount) {
 314                 monitor_refcount++;
 315                 goto exit;
 316         }
 317
 318         intr_handle = rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_PRIVATE);
 319         if (intr_handle == NULL) {
 320                 RTE_LOG(ERR, EAL, "Fail to allocate intr_handle\n");
 321                 goto exit;
 322         }
 323
 324         ret = rte_intr_type_set(intr_handle, RTE_INTR_HANDLE_DEV_EVENT);
 325         if (ret)
 326                 goto exit;
 327
 328         ret = rte_intr_fd_set(intr_handle, -1);
 329         if (ret)
 330                 goto exit;
 331
 332         ret = dev_uev_socket_fd_create();
 333         if (ret) {
 334                 RTE_LOG(ERR, EAL, "error create device event fd.\n");
 335                 goto exit;
 336         }
 337
 338         ret = rte_intr_callback_register(intr_handle, dev_uev_handler, NULL);
 339
 340         if (ret) {
 341                 close(rte_intr_fd_get(intr_handle));
 342                 goto exit;
 343         }
 344
 345         monitor_refcount++;
 346
 347 exit:
 348         if (ret) {
 349                 rte_intr_instance_free(intr_handle);
 350                 intr_handle = NULL;
 351         }
 352         rte_rwlock_write_unlock(&monitor_lock);
 353         return ret;
 354 }
 355
 356 int
 357 rte_dev_event_monitor_stop(void)
 358 {
 359         int ret = 0;
 360
 361         rte_rwlock_write_lock(&monitor_lock);
 362
 363         if (!monitor_refcount) {
 364                 RTE_LOG(ERR, EAL, "device event monitor already stopped\n");
 365                 goto exit;
 366         }
 367
 368         if (monitor_refcount > 1) {
 369                 monitor_refcount--;
 370                 goto exit;
 371         }
 372
 373         ret = rte_intr_callback_unregister(intr_handle, dev_uev_handler,
 374                                            (void *)-1);
 375         if (ret < 0) {
 376                 RTE_LOG(ERR, EAL, "fail to unregister uevent callback.\n");
 377                 goto exit;
 378         }
 379
 380         close(rte_intr_fd_get(intr_handle));
 381         rte_intr_instance_free(intr_handle);
 382         intr_handle = NULL;
 383
 384         monitor_refcount--;
 385
 386 exit:
 387         rte_rwlock_write_unlock(&monitor_lock);
 388
 389         return ret;
 390 }
 391
 392 int
 393 dev_sigbus_handler_register(void)
 394 {
 395         sigset_t mask;
 396         struct sigaction action;
 397
 398         rte_errno = 0;
 399
 400         if (sigbus_need_recover)
 401                 return 0;
 402
 403         sigemptyset(&mask);
 404         sigaddset(&mask, SIGBUS);
 405         action.sa_flags = SA_SIGINFO;
 406         action.sa_mask = mask;
 407         action.sa_sigaction = sigbus_handler;
 408         sigbus_need_recover = !sigaction(SIGBUS, &action, &sigbus_action_old);
 409
 410         return rte_errno;
 411 }
 412
 413 int
 414 dev_sigbus_handler_unregister(void)
 415 {
 416         rte_errno = 0;
 417
 418         sigbus_action_recover();
 419
 420         return rte_errno;
 421 }
 422
 423 int
 424 rte_dev_hotplug_handle_enable(void)
 425 {
 426         int ret = 0;
 427
 428         ret = dev_sigbus_handler_register();
 429         if (ret < 0)
 430                 RTE_LOG(ERR, EAL,
 431                         "fail to register sigbus handler for devices.\n");
 432
 433         hotplug_handle = true;
 434
 435         return ret;
 436 }
 437
 438 int
 439 rte_dev_hotplug_handle_disable(void)
 440 {
 441         int ret = 0;
 442
 443         ret = dev_sigbus_handler_unregister();
 444         if (ret < 0)
 445                 RTE_LOG(ERR, EAL,
 446                         "fail to unregister sigbus handler for devices.\n");
 447
 448         hotplug_handle = false;
 449
 450         return ret;
 451 }