mem: add thread unsafe version for DMA mask check
[dpdk.git] / lib / librte_eal / linuxapp / eal / eal_dev.c
index 9c8d1a0..d589c69 100644 (file)
  * Copyright(c) 2018 Intel Corporation
  */
 
+#include <string.h>
+#include <unistd.h>
+#include <fcntl.h>
+#include <signal.h>
+#include <sys/socket.h>
+#include <linux/netlink.h>
+
+#include <rte_string_fns.h>
 #include <rte_log.h>
 #include <rte_compat.h>
 #include <rte_dev.h>
+#include <rte_malloc.h>
+#include <rte_interrupts.h>
+#include <rte_alarm.h>
+#include <rte_bus.h>
+#include <rte_eal.h>
+#include <rte_spinlock.h>
+#include <rte_errno.h>
+
+#include "eal_private.h"
+
+static struct rte_intr_handle intr_handle = {.fd = -1 };
+static bool monitor_started;
+static bool hotplug_handle;
+
+#define EAL_UEV_MSG_LEN 4096
+#define EAL_UEV_MSG_ELEM_LEN 128
+
+/*
+ * spinlock for device hot-unplug failure handling. If it try to access bus or
+ * device, such as handle sigbus on bus or handle memory failure for device
+ * just need to use this lock. It could protect the bus and the device to avoid
+ * race condition.
+ */
+static rte_spinlock_t failure_handle_lock = RTE_SPINLOCK_INITIALIZER;
+
+static struct sigaction sigbus_action_old;
+
+static int sigbus_need_recover;
+
+static void dev_uev_handler(__rte_unused void *param);
+
+/* identify the system layer which reports this event. */
+enum eal_dev_event_subsystem {
+       EAL_DEV_EVENT_SUBSYSTEM_PCI, /* PCI bus device event */
+       EAL_DEV_EVENT_SUBSYSTEM_UIO, /* UIO driver device event */
+       EAL_DEV_EVENT_SUBSYSTEM_VFIO, /* VFIO driver device event */
+       EAL_DEV_EVENT_SUBSYSTEM_MAX
+};
+
+static void
+sigbus_action_recover(void)
+{
+       if (sigbus_need_recover) {
+               sigaction(SIGBUS, &sigbus_action_old, NULL);
+               sigbus_need_recover = 0;
+       }
+}
+
+static void sigbus_handler(int signum, siginfo_t *info,
+                               void *ctx __rte_unused)
+{
+       int ret;
+
+       RTE_LOG(DEBUG, EAL, "Thread[%d] catch SIGBUS, fault address:%p\n",
+               (int)pthread_self(), info->si_addr);
+
+       rte_spinlock_lock(&failure_handle_lock);
+       ret = rte_bus_sigbus_handler(info->si_addr);
+       rte_spinlock_unlock(&failure_handle_lock);
+       if (ret == -1) {
+               rte_exit(EXIT_FAILURE,
+                        "Failed to handle SIGBUS for hot-unplug, "
+                        "(rte_errno: %s)!", strerror(rte_errno));
+       } else if (ret == 1) {
+               if (sigbus_action_old.sa_flags == SA_SIGINFO
+                   && sigbus_action_old.sa_sigaction) {
+                       (*(sigbus_action_old.sa_sigaction))(signum,
+                                                           info, ctx);
+               } else if (sigbus_action_old.sa_flags != SA_SIGINFO
+                          && sigbus_action_old.sa_handler) {
+                       (*(sigbus_action_old.sa_handler))(signum);
+               } else {
+                       rte_exit(EXIT_FAILURE,
+                                "Failed to handle generic SIGBUS!");
+               }
+       }
+
+       RTE_LOG(DEBUG, EAL, "Success to handle SIGBUS for hot-unplug!\n");
+}
+
+static int cmp_dev_name(const struct rte_device *dev,
+       const void *_name)
+{
+       const char *name = _name;
+
+       return strcmp(dev->name, name);
+}
+
+static int
+dev_uev_socket_fd_create(void)
+{
+       struct sockaddr_nl addr;
+       int ret;
+
+       intr_handle.fd = socket(PF_NETLINK, SOCK_RAW | SOCK_CLOEXEC |
+                       SOCK_NONBLOCK,
+                       NETLINK_KOBJECT_UEVENT);
+       if (intr_handle.fd < 0) {
+               RTE_LOG(ERR, EAL, "create uevent fd failed.\n");
+               return -1;
+       }
+
+       memset(&addr, 0, sizeof(addr));
+       addr.nl_family = AF_NETLINK;
+       addr.nl_pid = 0;
+       addr.nl_groups = 0xffffffff;
+
+       ret = bind(intr_handle.fd, (struct sockaddr *) &addr, sizeof(addr));
+       if (ret < 0) {
+               RTE_LOG(ERR, EAL, "Failed to bind uevent socket.\n");
+               goto err;
+       }
+
+       return 0;
+err:
+       close(intr_handle.fd);
+       intr_handle.fd = -1;
+       return ret;
+}
+
+static int
+dev_uev_parse(const char *buf, struct rte_dev_event *event, int length)
+{
+       char action[EAL_UEV_MSG_ELEM_LEN];
+       char subsystem[EAL_UEV_MSG_ELEM_LEN];
+       char pci_slot_name[EAL_UEV_MSG_ELEM_LEN];
+       int i = 0;
+
+       memset(action, 0, EAL_UEV_MSG_ELEM_LEN);
+       memset(subsystem, 0, EAL_UEV_MSG_ELEM_LEN);
+       memset(pci_slot_name, 0, EAL_UEV_MSG_ELEM_LEN);
+
+       while (i < length) {
+               for (; i < length; i++) {
+                       if (*buf)
+                               break;
+                       buf++;
+               }
+               /**
+                * check device uevent from kernel side, no need to check
+                * uevent from udev.
+                */
+               if (!strncmp(buf, "libudev", 7)) {
+                       buf += 7;
+                       i += 7;
+                       return -1;
+               }
+               if (!strncmp(buf, "ACTION=", 7)) {
+                       buf += 7;
+                       i += 7;
+                       strlcpy(action, buf, sizeof(action));
+               } else if (!strncmp(buf, "SUBSYSTEM=", 10)) {
+                       buf += 10;
+                       i += 10;
+                       strlcpy(subsystem, buf, sizeof(subsystem));
+               } else if (!strncmp(buf, "PCI_SLOT_NAME=", 14)) {
+                       buf += 14;
+                       i += 14;
+                       strlcpy(pci_slot_name, buf, sizeof(subsystem));
+                       event->devname = strdup(pci_slot_name);
+               }
+               for (; i < length; i++) {
+                       if (*buf == '\0')
+                               break;
+                       buf++;
+               }
+       }
+
+       /* parse the subsystem layer */
+       if (!strncmp(subsystem, "uio", 3))
+               event->subsystem = EAL_DEV_EVENT_SUBSYSTEM_UIO;
+       else if (!strncmp(subsystem, "pci", 3))
+               event->subsystem = EAL_DEV_EVENT_SUBSYSTEM_PCI;
+       else if (!strncmp(subsystem, "vfio", 4))
+               event->subsystem = EAL_DEV_EVENT_SUBSYSTEM_VFIO;
+       else
+               return -1;
+
+       /* parse the action type */
+       if (!strncmp(action, "add", 3))
+               event->type = RTE_DEV_EVENT_ADD;
+       else if (!strncmp(action, "remove", 6))
+               event->type = RTE_DEV_EVENT_REMOVE;
+       else
+               return -1;
+       return 0;
+}
+
+static void
+dev_delayed_unregister(void *param)
+{
+       rte_intr_callback_unregister(&intr_handle, dev_uev_handler, param);
+       close(intr_handle.fd);
+       intr_handle.fd = -1;
+}
+
+static void
+dev_uev_handler(__rte_unused void *param)
+{
+       struct rte_dev_event uevent;
+       int ret;
+       char buf[EAL_UEV_MSG_LEN];
+       struct rte_bus *bus;
+       struct rte_device *dev;
+       const char *busname = "";
 
+       memset(&uevent, 0, sizeof(struct rte_dev_event));
+       memset(buf, 0, EAL_UEV_MSG_LEN);
+
+       ret = recv(intr_handle.fd, buf, EAL_UEV_MSG_LEN, MSG_DONTWAIT);
+       if (ret < 0 && errno == EAGAIN)
+               return;
+       else if (ret <= 0) {
+               /* connection is closed or broken, can not up again. */
+               RTE_LOG(ERR, EAL, "uevent socket connection is broken.\n");
+               rte_eal_alarm_set(1, dev_delayed_unregister, NULL);
+               return;
+       }
+
+       ret = dev_uev_parse(buf, &uevent, EAL_UEV_MSG_LEN);
+       if (ret < 0) {
+               RTE_LOG(DEBUG, EAL, "It is not an valid event "
+                       "that need to be handle.\n");
+               return;
+       }
+
+       RTE_LOG(DEBUG, EAL, "receive uevent(name:%s, type:%d, subsystem:%d)\n",
+               uevent.devname, uevent.type, uevent.subsystem);
+
+       switch (uevent.subsystem) {
+       case EAL_DEV_EVENT_SUBSYSTEM_PCI:
+       case EAL_DEV_EVENT_SUBSYSTEM_UIO:
+               busname = "pci";
+               break;
+       default:
+               break;
+       }
+
+       if (uevent.devname) {
+               if (uevent.type == RTE_DEV_EVENT_REMOVE && hotplug_handle) {
+                       rte_spinlock_lock(&failure_handle_lock);
+                       bus = rte_bus_find_by_name(busname);
+                       if (bus == NULL) {
+                               RTE_LOG(ERR, EAL, "Cannot find bus (%s)\n",
+                                       busname);
+                               return;
+                       }
+
+                       dev = bus->find_device(NULL, cmp_dev_name,
+                                              uevent.devname);
+                       if (dev == NULL) {
+                               RTE_LOG(ERR, EAL, "Cannot find device (%s) on "
+                                       "bus (%s)\n", uevent.devname, busname);
+                               return;
+                       }
+
+                       ret = bus->hot_unplug_handler(dev);
+                       rte_spinlock_unlock(&failure_handle_lock);
+                       if (ret) {
+                               RTE_LOG(ERR, EAL, "Can not handle hot-unplug "
+                                       "for device (%s)\n", dev->name);
+                               return;
+                       }
+               }
+               rte_dev_event_callback_process(uevent.devname, uevent.type);
+       }
+}
 
 int __rte_experimental
 rte_dev_event_monitor_start(void)
 {
-       /* TODO: start uevent monitor for linux */
+       int ret;
+
+       if (monitor_started)
+               return 0;
+
+       ret = dev_uev_socket_fd_create();
+       if (ret) {
+               RTE_LOG(ERR, EAL, "error create device event fd.\n");
+               return -1;
+       }
+
+       intr_handle.type = RTE_INTR_HANDLE_DEV_EVENT;
+       ret = rte_intr_callback_register(&intr_handle, dev_uev_handler, NULL);
+
+       if (ret) {
+               RTE_LOG(ERR, EAL, "fail to register uevent callback.\n");
+               return -1;
+       }
+
+       monitor_started = true;
+
        return 0;
 }
 
 int __rte_experimental
 rte_dev_event_monitor_stop(void)
 {
-       /* TODO: stop uevent monitor for linux */
+       int ret;
+
+       if (!monitor_started)
+               return 0;
+
+       ret = rte_intr_callback_unregister(&intr_handle, dev_uev_handler,
+                                          (void *)-1);
+       if (ret < 0) {
+               RTE_LOG(ERR, EAL, "fail to unregister uevent callback.\n");
+               return ret;
+       }
+
+       close(intr_handle.fd);
+       intr_handle.fd = -1;
+       monitor_started = false;
+
        return 0;
 }
+
+int
+dev_sigbus_handler_register(void)
+{
+       sigset_t mask;
+       struct sigaction action;
+
+       rte_errno = 0;
+
+       if (sigbus_need_recover)
+               return 0;
+
+       sigemptyset(&mask);
+       sigaddset(&mask, SIGBUS);
+       action.sa_flags = SA_SIGINFO;
+       action.sa_mask = mask;
+       action.sa_sigaction = sigbus_handler;
+       sigbus_need_recover = !sigaction(SIGBUS, &action, &sigbus_action_old);
+
+       return rte_errno;
+}
+
+int
+dev_sigbus_handler_unregister(void)
+{
+       rte_errno = 0;
+
+       sigbus_action_recover();
+
+       return rte_errno;
+}
+
+int __rte_experimental
+rte_dev_hotplug_handle_enable(void)
+{
+       int ret = 0;
+
+       ret = dev_sigbus_handler_register();
+       if (ret < 0)
+               RTE_LOG(ERR, EAL,
+                       "fail to register sigbus handler for devices.\n");
+
+       hotplug_handle = true;
+
+       return ret;
+}
+
+int __rte_experimental
+rte_dev_hotplug_handle_disable(void)
+{
+       int ret = 0;
+
+       ret = dev_sigbus_handler_unregister();
+       if (ret < 0)
+               RTE_LOG(ERR, EAL,
+                       "fail to unregister sigbus handler for devices.\n");
+
+       hotplug_handle = false;
+
+       return ret;
+}