ceph/src/spdk/dpdk/lib/librte_eal/linuxapp/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 <sys/socket.h>
   8 #include <linux/netlink.h>
   9
  10 #include <rte_string_fns.h>
  11 #include <rte_log.h>
  12 #include <rte_compat.h>
  13 #include <rte_dev.h>
  14 #include <rte_malloc.h>
  15 #include <rte_interrupts.h>
  16 #include <rte_alarm.h>
  17
  18 #include "eal_private.h"
  19
  20 static struct rte_intr_handle intr_handle = {.fd = -1 };
  21 static bool monitor_started;
  22
  23 #define EAL_UEV_MSG_LEN 4096
  24 #define EAL_UEV_MSG_ELEM_LEN 128
  25
  26 static void dev_uev_handler(__rte_unused void *param);
  27
  28 /* identify the system layer which reports this event. */
  29 enum eal_dev_event_subsystem {
  30         EAL_DEV_EVENT_SUBSYSTEM_PCI, /* PCI bus device event */
  31         EAL_DEV_EVENT_SUBSYSTEM_UIO, /* UIO driver device event */
  32         EAL_DEV_EVENT_SUBSYSTEM_VFIO, /* VFIO driver device event */
  33         EAL_DEV_EVENT_SUBSYSTEM_MAX
  34 };
  35
  36 static int
  37 dev_uev_socket_fd_create(void)
  38 {
  39         struct sockaddr_nl addr;
  40         int ret;
  41
  42         intr_handle.fd = socket(PF_NETLINK, SOCK_RAW | SOCK_CLOEXEC |
  43                         SOCK_NONBLOCK,
  44                         NETLINK_KOBJECT_UEVENT);
  45         if (intr_handle.fd < 0) {
  46                 RTE_LOG(ERR, EAL, "create uevent fd failed.\n");
  47                 return -1;
  48         }
  49
  50         memset(&addr, 0, sizeof(addr));
  51         addr.nl_family = AF_NETLINK;
  52         addr.nl_pid = 0;
  53         addr.nl_groups = 0xffffffff;
  54
  55         ret = bind(intr_handle.fd, (struct sockaddr *) &addr, sizeof(addr));
  56         if (ret < 0) {
  57                 RTE_LOG(ERR, EAL, "Failed to bind uevent socket.\n");
  58                 goto err;
  59         }
  60
  61         return 0;
  62 err:
  63         close(intr_handle.fd);
  64         intr_handle.fd = -1;
  65         return ret;
  66 }
  67
  68 static int
  69 dev_uev_parse(const char *buf, struct rte_dev_event *event, int length)
  70 {
  71         char action[EAL_UEV_MSG_ELEM_LEN];
  72         char subsystem[EAL_UEV_MSG_ELEM_LEN];
  73         char pci_slot_name[EAL_UEV_MSG_ELEM_LEN];
  74         int i = 0;
  75
  76         memset(action, 0, EAL_UEV_MSG_ELEM_LEN);
  77         memset(subsystem, 0, EAL_UEV_MSG_ELEM_LEN);
  78         memset(pci_slot_name, 0, EAL_UEV_MSG_ELEM_LEN);
  79
  80         while (i < length) {
  81                 for (; i < length; i++) {
  82                         if (*buf)
  83                                 break;
  84                         buf++;
  85                 }
  86                 /**
  87                  * check device uevent from kernel side, no need to check
  88                  * uevent from udev.
  89                  */
  90                 if (!strncmp(buf, "libudev", 7)) {
  91                         buf += 7;
  92                         i += 7;
  93                         return -1;
  94                 }
  95                 if (!strncmp(buf, "ACTION=", 7)) {
  96                         buf += 7;
  97                         i += 7;
  98                         strlcpy(action, buf, sizeof(action));
  99                 } else if (!strncmp(buf, "SUBSYSTEM=", 10)) {
 100                         buf += 10;
 101                         i += 10;
 102                         strlcpy(subsystem, buf, sizeof(subsystem));
 103                 } else if (!strncmp(buf, "PCI_SLOT_NAME=", 14)) {
 104                         buf += 14;
 105                         i += 14;
 106                         strlcpy(pci_slot_name, buf, sizeof(subsystem));
 107                         event->devname = strdup(pci_slot_name);
 108                 }
 109                 for (; i < length; i++) {
 110                         if (*buf == '\0')
 111                                 break;
 112                         buf++;
 113                 }
 114         }
 115
 116         /* parse the subsystem layer */
 117         if (!strncmp(subsystem, "uio", 3))
 118                 event->subsystem = EAL_DEV_EVENT_SUBSYSTEM_UIO;
 119         else if (!strncmp(subsystem, "pci", 3))
 120                 event->subsystem = EAL_DEV_EVENT_SUBSYSTEM_PCI;
 121         else if (!strncmp(subsystem, "vfio", 4))
 122                 event->subsystem = EAL_DEV_EVENT_SUBSYSTEM_VFIO;
 123         else
 124                 return -1;
 125
 126         /* parse the action type */
 127         if (!strncmp(action, "add", 3))
 128                 event->type = RTE_DEV_EVENT_ADD;
 129         else if (!strncmp(action, "remove", 6))
 130                 event->type = RTE_DEV_EVENT_REMOVE;
 131         else
 132                 return -1;
 133         return 0;
 134 }
 135
 136 static void
 137 dev_delayed_unregister(void *param)
 138 {
 139         rte_intr_callback_unregister(&intr_handle, dev_uev_handler, param);
 140         close(intr_handle.fd);
 141         intr_handle.fd = -1;
 142 }
 143
 144 static void
 145 dev_uev_handler(__rte_unused void *param)
 146 {
 147         struct rte_dev_event uevent;
 148         int ret;
 149         char buf[EAL_UEV_MSG_LEN];
 150
 151         memset(&uevent, 0, sizeof(struct rte_dev_event));
 152         memset(buf, 0, EAL_UEV_MSG_LEN);
 153
 154         ret = recv(intr_handle.fd, buf, EAL_UEV_MSG_LEN, MSG_DONTWAIT);
 155         if (ret < 0 && errno == EAGAIN)
 156                 return;
 157         else if (ret <= 0) {
 158                 /* connection is closed or broken, can not up again. */
 159                 RTE_LOG(ERR, EAL, "uevent socket connection is broken.\n");
 160                 rte_eal_alarm_set(1, dev_delayed_unregister, NULL);
 161                 return;
 162         }
 163
 164         ret = dev_uev_parse(buf, &uevent, EAL_UEV_MSG_LEN);
 165         if (ret < 0) {
 166                 RTE_LOG(DEBUG, EAL, "It is not an valid event "
 167                         "that need to be handle.\n");
 168                 return;
 169         }
 170
 171         RTE_LOG(DEBUG, EAL, "receive uevent(name:%s, type:%d, subsystem:%d)\n",
 172                 uevent.devname, uevent.type, uevent.subsystem);
 173
 174         if (uevent.devname)
 175                 dev_callback_process(uevent.devname, uevent.type);
 176 }
 177
 178 int __rte_experimental
 179 rte_dev_event_monitor_start(void)
 180 {
 181         int ret;
 182
 183         if (monitor_started)
 184                 return 0;
 185
 186         ret = dev_uev_socket_fd_create();
 187         if (ret) {
 188                 RTE_LOG(ERR, EAL, "error create device event fd.\n");
 189                 return -1;
 190         }
 191
 192         intr_handle.type = RTE_INTR_HANDLE_DEV_EVENT;
 193         ret = rte_intr_callback_register(&intr_handle, dev_uev_handler, NULL);
 194
 195         if (ret) {
 196                 RTE_LOG(ERR, EAL, "fail to register uevent callback.\n");
 197                 return -1;
 198         }
 199
 200         monitor_started = true;
 201
 202         return 0;
 203 }
 204
 205 int __rte_experimental
 206 rte_dev_event_monitor_stop(void)
 207 {
 208         int ret;
 209
 210         if (!monitor_started)
 211                 return 0;
 212
 213         ret = rte_intr_callback_unregister(&intr_handle, dev_uev_handler,
 214                                            (void *)-1);
 215         if (ret < 0) {
 216                 RTE_LOG(ERR, EAL, "fail to unregister uevent callback.\n");
 217                 return ret;
 218         }
 219
 220         close(intr_handle.fd);
 221         intr_handle.fd = -1;
 222         monitor_started = false;
 223         return 0;
 224 }