net/core/drop_monitor.c

   1 /*
   2  * Monitoring code for network dropped packet alerts
   3  *
   4  * Copyright (C) 2009 Neil Horman <nhorman@tuxdriver.com>
   5  */
   6
   7 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   8
   9 #include <linux/netdevice.h>
  10 #include <linux/etherdevice.h>
  11 #include <linux/string.h>
  12 #include <linux/if_arp.h>
  13 #include <linux/inetdevice.h>
  14 #include <linux/inet.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/netpoll.h>
  17 #include <linux/sched.h>
  18 #include <linux/delay.h>
  19 #include <linux/types.h>
  20 #include <linux/workqueue.h>
  21 #include <linux/netlink.h>
  22 #include <linux/net_dropmon.h>
  23 #include <linux/percpu.h>
  24 #include <linux/timer.h>
  25 #include <linux/bitops.h>
  26 #include <linux/slab.h>
  27 #include <linux/module.h>
  28 #include <net/genetlink.h>
  29 #include <net/netevent.h>
  30
  31 #include <trace/events/skb.h>
  32 #include <trace/events/napi.h>
  33
  34 #include <asm/unaligned.h>
  35
  36 #define TRACE_ON 1
  37 #define TRACE_OFF 0
  38
  39 /*
  40  * Globals, our netlink socket pointer
  41  * and the work handle that will send up
  42  * netlink alerts
  43  */
  44 static int trace_state = TRACE_OFF;
  45 static DEFINE_MUTEX(trace_state_mutex);
  46
  47 struct per_cpu_dm_data {
  48         spinlock_t              lock;
  49         struct sk_buff          *skb;
  50         struct work_struct      dm_alert_work;
  51         struct timer_list       send_timer;
  52 };
  53
  54 struct dm_hw_stat_delta {
  55         struct net_device *dev;
  56         unsigned long last_rx;
  57         struct list_head list;
  58         struct rcu_head rcu;
  59         unsigned long last_drop_val;
  60 };
  61
  62 static struct genl_family net_drop_monitor_family = {
  63         .id             = GENL_ID_GENERATE,
  64         .hdrsize        = 0,
  65         .name           = "NET_DM",
  66         .version        = 2,
  67 };
  68
  69 static DEFINE_PER_CPU(struct per_cpu_dm_data, dm_cpu_data);
  70
  71 static int dm_hit_limit = 64;
  72 static int dm_delay = 1;
  73 static unsigned long dm_hw_check_delta = 2*HZ;
  74 static LIST_HEAD(hw_stats_list);
  75
  76 static struct sk_buff *reset_per_cpu_data(struct per_cpu_dm_data *data)
  77 {
  78         size_t al;
  79         struct net_dm_alert_msg *msg;
  80         struct nlattr *nla;
  81         struct sk_buff *skb;
  82         unsigned long flags;
  83
  84         al = sizeof(struct net_dm_alert_msg);
  85         al += dm_hit_limit * sizeof(struct net_dm_drop_point);
  86         al += sizeof(struct nlattr);
  87
  88         skb = genlmsg_new(al, GFP_KERNEL);
  89
  90         if (skb) {
  91                 genlmsg_put(skb, 0, 0, &net_drop_monitor_family,
  92                                 0, NET_DM_CMD_ALERT);
  93                 nla = nla_reserve(skb, NLA_UNSPEC,
  94                                   sizeof(struct net_dm_alert_msg));
  95                 msg = nla_data(nla);
  96                 memset(msg, 0, al);
  97         } else {
  98                 mod_timer(&data->send_timer, jiffies + HZ / 10);
  99         }
 100
 101         spin_lock_irqsave(&data->lock, flags);
 102         swap(data->skb, skb);
 103         spin_unlock_irqrestore(&data->lock, flags);
 104
 105         return skb;
 106 }
 107
 108 static void send_dm_alert(struct work_struct *work)
 109 {
 110         struct sk_buff *skb;
 111         struct per_cpu_dm_data *data;
 112
 113         data = container_of(work, struct per_cpu_dm_data, dm_alert_work);
 114
 115         skb = reset_per_cpu_data(data);
 116
 117         if (skb)
 118                 genlmsg_multicast(skb, 0, NET_DM_GRP_ALERT, GFP_KERNEL);
 119 }
 120
 121 /*
 122  * This is the timer function to delay the sending of an alert
 123  * in the event that more drops will arrive during the
 124  * hysteresis period.
 125  */
 126 static void sched_send_work(unsigned long _data)
 127 {
 128         struct per_cpu_dm_data *data = (struct per_cpu_dm_data *)_data;
 129
 130         schedule_work(&data->dm_alert_work);
 131 }
 132
 133 static void trace_drop_common(struct sk_buff *skb, void *location)
 134 {
 135         struct net_dm_alert_msg *msg;
 136         struct nlmsghdr *nlh;
 137         struct nlattr *nla;
 138         int i;
 139         struct sk_buff *dskb;
 140         struct per_cpu_dm_data *data;
 141         unsigned long flags;
 142
 143         local_irq_save(flags);
 144         data = &__get_cpu_var(dm_cpu_data);
 145         spin_lock(&data->lock);
 146         dskb = data->skb;
 147
 148         if (!dskb)
 149                 goto out;
 150
 151         nlh = (struct nlmsghdr *)dskb->data;
 152         nla = genlmsg_data(nlmsg_data(nlh));
 153         msg = nla_data(nla);
 154         for (i = 0; i < msg->entries; i++) {
 155                 if (!memcmp(&location, msg->points[i].pc, sizeof(void *))) {
 156                         msg->points[i].count++;
 157                         goto out;
 158                 }
 159         }
 160         if (msg->entries == dm_hit_limit)
 161                 goto out;
 162         /*
 163          * We need to create a new entry
 164          */
 165         __nla_reserve_nohdr(dskb, sizeof(struct net_dm_drop_point));
 166         nla->nla_len += NLA_ALIGN(sizeof(struct net_dm_drop_point));
 167         memcpy(msg->points[msg->entries].pc, &location, sizeof(void *));
 168         msg->points[msg->entries].count = 1;
 169         msg->entries++;
 170
 171         if (!timer_pending(&data->send_timer)) {
 172                 data->send_timer.expires = jiffies + dm_delay * HZ;
 173                 add_timer(&data->send_timer);
 174         }
 175
 176 out:
 177         spin_unlock_irqrestore(&data->lock, flags);
 178 }
 179
 180 static void trace_kfree_skb_hit(void *ignore, struct sk_buff *skb, void *location)
 181 {
 182         trace_drop_common(skb, location);
 183 }
 184
 185 static void trace_napi_poll_hit(void *ignore, struct napi_struct *napi)
 186 {
 187         struct dm_hw_stat_delta *new_stat;
 188
 189         /*
 190          * Don't check napi structures with no associated device
 191          */
 192         if (!napi->dev)
 193                 return;
 194
 195         rcu_read_lock();
 196         list_for_each_entry_rcu(new_stat, &hw_stats_list, list) {
 197                 /*
 198                  * only add a note to our monitor buffer if:
 199                  * 1) this is the dev we received on
 200                  * 2) its after the last_rx delta
 201                  * 3) our rx_dropped count has gone up
 202                  */
 203                 if ((new_stat->dev == napi->dev)  &&
 204                     (time_after(jiffies, new_stat->last_rx + dm_hw_check_delta)) &&
 205                     (napi->dev->stats.rx_dropped != new_stat->last_drop_val)) {
 206                         trace_drop_common(NULL, NULL);
 207                         new_stat->last_drop_val = napi->dev->stats.rx_dropped;
 208                         new_stat->last_rx = jiffies;
 209                         break;
 210                 }
 211         }
 212         rcu_read_unlock();
 213 }
 214
 215 static int set_all_monitor_traces(int state)
 216 {
 217         int rc = 0;
 218         struct dm_hw_stat_delta *new_stat = NULL;
 219         struct dm_hw_stat_delta *temp;
 220
 221         mutex_lock(&trace_state_mutex);
 222
 223         if (state == trace_state) {
 224                 rc = -EAGAIN;
 225                 goto out_unlock;
 226         }
 227
 228         switch (state) {
 229         case TRACE_ON:
 230                 if (!try_module_get(THIS_MODULE)) {
 231                         rc = -ENODEV;
 232                         break;
 233                 }
 234
 235                 rc |= register_trace_kfree_skb(trace_kfree_skb_hit, NULL);
 236                 rc |= register_trace_napi_poll(trace_napi_poll_hit, NULL);
 237                 break;
 238
 239         case TRACE_OFF:
 240                 rc |= unregister_trace_kfree_skb(trace_kfree_skb_hit, NULL);
 241                 rc |= unregister_trace_napi_poll(trace_napi_poll_hit, NULL);
 242
 243                 tracepoint_synchronize_unregister();
 244
 245                 /*
 246                  * Clean the device list
 247                  */
 248                 list_for_each_entry_safe(new_stat, temp, &hw_stats_list, list) {
 249                         if (new_stat->dev == NULL) {
 250                                 list_del_rcu(&new_stat->list);
 251                                 kfree_rcu(new_stat, rcu);
 252                         }
 253                 }
 254
 255                 module_put(THIS_MODULE);
 256
 257                 break;
 258         default:
 259                 rc = 1;
 260                 break;
 261         }
 262
 263         if (!rc)
 264                 trace_state = state;
 265         else
 266                 rc = -EINPROGRESS;
 267
 268 out_unlock:
 269         mutex_unlock(&trace_state_mutex);
 270
 271         return rc;
 272 }
 273
 274
 275 static int net_dm_cmd_config(struct sk_buff *skb,
 276                         struct genl_info *info)
 277 {
 278         return -ENOTSUPP;
 279 }
 280
 281 static int net_dm_cmd_trace(struct sk_buff *skb,
 282                         struct genl_info *info)
 283 {
 284         switch (info->genlhdr->cmd) {
 285         case NET_DM_CMD_START:
 286                 return set_all_monitor_traces(TRACE_ON);
 287                 break;
 288         case NET_DM_CMD_STOP:
 289                 return set_all_monitor_traces(TRACE_OFF);
 290                 break;
 291         }
 292
 293         return -ENOTSUPP;
 294 }
 295
 296 static int dropmon_net_event(struct notifier_block *ev_block,
 297                         unsigned long event, void *ptr)
 298 {
 299         struct net_device *dev = ptr;
 300         struct dm_hw_stat_delta *new_stat = NULL;
 301         struct dm_hw_stat_delta *tmp;
 302
 303         switch (event) {
 304         case NETDEV_REGISTER:
 305                 new_stat = kzalloc(sizeof(struct dm_hw_stat_delta), GFP_KERNEL);
 306
 307                 if (!new_stat)
 308                         goto out;
 309
 310                 new_stat->dev = dev;
 311                 new_stat->last_rx = jiffies;
 312                 mutex_lock(&trace_state_mutex);
 313                 list_add_rcu(&new_stat->list, &hw_stats_list);
 314                 mutex_unlock(&trace_state_mutex);
 315                 break;
 316         case NETDEV_UNREGISTER:
 317                 mutex_lock(&trace_state_mutex);
 318                 list_for_each_entry_safe(new_stat, tmp, &hw_stats_list, list) {
 319                         if (new_stat->dev == dev) {
 320                                 new_stat->dev = NULL;
 321                                 if (trace_state == TRACE_OFF) {
 322                                         list_del_rcu(&new_stat->list);
 323                                         kfree_rcu(new_stat, rcu);
 324                                         break;
 325                                 }
 326                         }
 327                 }
 328                 mutex_unlock(&trace_state_mutex);
 329                 break;
 330         }
 331 out:
 332         return NOTIFY_DONE;
 333 }
 334
 335 static struct genl_ops dropmon_ops[] = {
 336         {
 337                 .cmd = NET_DM_CMD_CONFIG,
 338                 .doit = net_dm_cmd_config,
 339         },
 340         {
 341                 .cmd = NET_DM_CMD_START,
 342                 .doit = net_dm_cmd_trace,
 343         },
 344         {
 345                 .cmd = NET_DM_CMD_STOP,
 346                 .doit = net_dm_cmd_trace,
 347         },
 348 };
 349
 350 static struct notifier_block dropmon_net_notifier = {
 351         .notifier_call = dropmon_net_event
 352 };
 353
 354 static int __init init_net_drop_monitor(void)
 355 {
 356         struct per_cpu_dm_data *data;
 357         int cpu, rc;
 358
 359         pr_info("Initializing network drop monitor service\n");
 360
 361         if (sizeof(void *) > 8) {
 362                 pr_err("Unable to store program counters on this arch, Drop monitor failed\n");
 363                 return -ENOSPC;
 364         }
 365
 366         rc = genl_register_family_with_ops(&net_drop_monitor_family,
 367                                            dropmon_ops,
 368                                            ARRAY_SIZE(dropmon_ops));
 369         if (rc) {
 370                 pr_err("Could not create drop monitor netlink family\n");
 371                 return rc;
 372         }
 373
 374         rc = register_netdevice_notifier(&dropmon_net_notifier);
 375         if (rc < 0) {
 376                 pr_crit("Failed to register netdevice notifier\n");
 377                 goto out_unreg;
 378         }
 379
 380         rc = 0;
 381
 382         for_each_possible_cpu(cpu) {
 383                 data = &per_cpu(dm_cpu_data, cpu);
 384                 INIT_WORK(&data->dm_alert_work, send_dm_alert);
 385                 init_timer(&data->send_timer);
 386                 data->send_timer.data = (unsigned long)data;
 387                 data->send_timer.function = sched_send_work;
 388                 spin_lock_init(&data->lock);
 389                 reset_per_cpu_data(data);
 390         }
 391
 392
 393         goto out;
 394
 395 out_unreg:
 396         genl_unregister_family(&net_drop_monitor_family);
 397 out:
 398         return rc;
 399 }
 400
 401 static void exit_net_drop_monitor(void)
 402 {
 403         struct per_cpu_dm_data *data;
 404         int cpu;
 405
 406         BUG_ON(unregister_netdevice_notifier(&dropmon_net_notifier));
 407
 408         /*
 409          * Because of the module_get/put we do in the trace state change path
 410          * we are guarnateed not to have any current users when we get here
 411          * all we need to do is make sure that we don't have any running timers
 412          * or pending schedule calls
 413          */
 414
 415         for_each_possible_cpu(cpu) {
 416                 data = &per_cpu(dm_cpu_data, cpu);
 417                 del_timer_sync(&data->send_timer);
 418                 cancel_work_sync(&data->dm_alert_work);
 419                 /*
 420                  * At this point, we should have exclusive access
 421                  * to this struct and can free the skb inside it
 422                  */
 423                 kfree_skb(data->skb);
 424         }
 425
 426         BUG_ON(genl_unregister_family(&net_drop_monitor_family));
 427 }
 428
 429 module_init(init_net_drop_monitor);
 430 module_exit(exit_net_drop_monitor);
 431
 432 MODULE_LICENSE("GPL v2");
 433 MODULE_AUTHOR("Neil Horman <nhorman@tuxdriver.com>");
 434 MODULE_ALIAS_GENL_FAMILY("NET_DM");