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