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         void *msg_header;
  84
  85         al = sizeof(struct net_dm_alert_msg);
  86         al += dm_hit_limit * sizeof(struct net_dm_drop_point);
  87         al += sizeof(struct nlattr);
  88
  89         skb = genlmsg_new(al, GFP_KERNEL);
  90
  91         if (!skb)
  92                 goto err;
  93
  94         msg_header = genlmsg_put(skb, 0, 0, &net_drop_monitor_family,
  95                                  0, NET_DM_CMD_ALERT);
  96         if (!msg_header) {
  97                 nlmsg_free(skb);
  98                 skb = NULL;
  99                 goto err;
 100         }
 101         nla = nla_reserve(skb, NLA_UNSPEC,
 102                           sizeof(struct net_dm_alert_msg));
 103         if (!nla) {
 104                 nlmsg_free(skb);
 105                 skb = NULL;
 106                 goto err;
 107         }
 108         msg = nla_data(nla);
 109         memset(msg, 0, al);
 110         goto out;
 111
 112 err:
 113         mod_timer(&data->send_timer, jiffies + HZ / 10);
 114 out:
 115         spin_lock_irqsave(&data->lock, flags);
 116         swap(data->skb, skb);
 117         spin_unlock_irqrestore(&data->lock, flags);
 118
 119         if (skb) {
 120                 struct nlmsghdr *nlh = (struct nlmsghdr *)skb->data;
 121                 struct genlmsghdr *gnlh = (struct genlmsghdr *)nlmsg_data(nlh);
 122
 123                 genlmsg_end(skb, genlmsg_data(gnlh));
 124         }
 125
 126         return skb;
 127 }
 128
 129 static const struct genl_multicast_group dropmon_mcgrps[] = {
 130         { .name = "events", },
 131 };
 132
 133 static void send_dm_alert(struct work_struct *work)
 134 {
 135         struct sk_buff *skb;
 136         struct per_cpu_dm_data *data;
 137
 138         data = container_of(work, struct per_cpu_dm_data, dm_alert_work);
 139
 140         skb = reset_per_cpu_data(data);
 141
 142         if (skb)
 143                 genlmsg_multicast(&net_drop_monitor_family, skb, 0,
 144                                   0, GFP_KERNEL);
 145 }
 146
 147 /*
 148  * This is the timer function to delay the sending of an alert
 149  * in the event that more drops will arrive during the
 150  * hysteresis period.
 151  */
 152 static void sched_send_work(unsigned long _data)
 153 {
 154         struct per_cpu_dm_data *data = (struct per_cpu_dm_data *)_data;
 155
 156         schedule_work(&data->dm_alert_work);
 157 }
 158
 159 static void trace_drop_common(struct sk_buff *skb, void *location)
 160 {
 161         struct net_dm_alert_msg *msg;
 162         struct net_dm_drop_point *point;
 163         struct nlmsghdr *nlh;
 164         struct nlattr *nla;
 165         int i;
 166         struct sk_buff *dskb;
 167         struct per_cpu_dm_data *data;
 168         unsigned long flags;
 169
 170         local_irq_save(flags);
 171         data = this_cpu_ptr(&dm_cpu_data);
 172         spin_lock(&data->lock);
 173         dskb = data->skb;
 174
 175         if (!dskb)
 176                 goto out;
 177
 178         nlh = (struct nlmsghdr *)dskb->data;
 179         nla = genlmsg_data(nlmsg_data(nlh));
 180         msg = nla_data(nla);
 181         point = msg->points;
 182         for (i = 0; i < msg->entries; i++) {
 183                 if (!memcmp(&location, &point->pc, sizeof(void *))) {
 184                         point->count++;
 185                         goto out;
 186                 }
 187                 point++;
 188         }
 189         if (msg->entries == dm_hit_limit)
 190                 goto out;
 191         /*
 192          * We need to create a new entry
 193          */
 194         __nla_reserve_nohdr(dskb, sizeof(struct net_dm_drop_point));
 195         nla->nla_len += NLA_ALIGN(sizeof(struct net_dm_drop_point));
 196         memcpy(point->pc, &location, sizeof(void *));
 197         point->count = 1;
 198         msg->entries++;
 199
 200         if (!timer_pending(&data->send_timer)) {
 201                 data->send_timer.expires = jiffies + dm_delay * HZ;
 202                 add_timer(&data->send_timer);
 203         }
 204
 205 out:
 206         spin_unlock_irqrestore(&data->lock, flags);
 207 }
 208
 209 static void trace_kfree_skb_hit(void *ignore, struct sk_buff *skb, void *location)
 210 {
 211         trace_drop_common(skb, location);
 212 }
 213
 214 static void trace_napi_poll_hit(void *ignore, struct napi_struct *napi,
 215                                 int work, int budget)
 216 {
 217         struct dm_hw_stat_delta *new_stat;
 218
 219         /*
 220          * Don't check napi structures with no associated device
 221          */
 222         if (!napi->dev)
 223                 return;
 224
 225         rcu_read_lock();
 226         list_for_each_entry_rcu(new_stat, &hw_stats_list, list) {
 227                 struct net_device *dev;
 228
 229                 /*
 230                  * only add a note to our monitor buffer if:
 231                  * 1) this is the dev we received on
 232                  * 2) its after the last_rx delta
 233                  * 3) our rx_dropped count has gone up
 234                  */
 235                 /* Paired with WRITE_ONCE() in dropmon_net_event() */
 236                 dev = READ_ONCE(new_stat->dev);
 237                 if ((dev == napi->dev)  &&
 238                     (time_after(jiffies, new_stat->last_rx + dm_hw_check_delta)) &&
 239                     (napi->dev->stats.rx_dropped != new_stat->last_drop_val)) {
 240                         trace_drop_common(NULL, NULL);
 241                         new_stat->last_drop_val = napi->dev->stats.rx_dropped;
 242                         new_stat->last_rx = jiffies;
 243                         break;
 244                 }
 245         }
 246         rcu_read_unlock();
 247 }
 248
 249 static int set_all_monitor_traces(int state)
 250 {
 251         int rc = 0;
 252         struct dm_hw_stat_delta *new_stat = NULL;
 253         struct dm_hw_stat_delta *temp;
 254
 255         mutex_lock(&trace_state_mutex);
 256
 257         if (state == trace_state) {
 258                 rc = -EAGAIN;
 259                 goto out_unlock;
 260         }
 261
 262         switch (state) {
 263         case TRACE_ON:
 264                 if (!try_module_get(THIS_MODULE)) {
 265                         rc = -ENODEV;
 266                         break;
 267                 }
 268
 269                 rc |= register_trace_kfree_skb(trace_kfree_skb_hit, NULL);
 270                 rc |= register_trace_napi_poll(trace_napi_poll_hit, NULL);
 271                 break;
 272
 273         case TRACE_OFF:
 274                 rc |= unregister_trace_kfree_skb(trace_kfree_skb_hit, NULL);
 275                 rc |= unregister_trace_napi_poll(trace_napi_poll_hit, NULL);
 276
 277                 tracepoint_synchronize_unregister();
 278
 279                 /*
 280                  * Clean the device list
 281                  */
 282                 list_for_each_entry_safe(new_stat, temp, &hw_stats_list, list) {
 283                         if (new_stat->dev == NULL) {
 284                                 list_del_rcu(&new_stat->list);
 285                                 kfree_rcu(new_stat, rcu);
 286                         }
 287                 }
 288
 289                 module_put(THIS_MODULE);
 290
 291                 break;
 292         default:
 293                 rc = 1;
 294                 break;
 295         }
 296
 297         if (!rc)
 298                 trace_state = state;
 299         else
 300                 rc = -EINPROGRESS;
 301
 302 out_unlock:
 303         mutex_unlock(&trace_state_mutex);
 304
 305         return rc;
 306 }
 307
 308
 309 static int net_dm_cmd_config(struct sk_buff *skb,
 310                         struct genl_info *info)
 311 {
 312         return -ENOTSUPP;
 313 }
 314
 315 static int net_dm_cmd_trace(struct sk_buff *skb,
 316                         struct genl_info *info)
 317 {
 318         switch (info->genlhdr->cmd) {
 319         case NET_DM_CMD_START:
 320                 return set_all_monitor_traces(TRACE_ON);
 321         case NET_DM_CMD_STOP:
 322                 return set_all_monitor_traces(TRACE_OFF);
 323         }
 324
 325         return -ENOTSUPP;
 326 }
 327
 328 static int dropmon_net_event(struct notifier_block *ev_block,
 329                              unsigned long event, void *ptr)
 330 {
 331         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
 332         struct dm_hw_stat_delta *new_stat = NULL;
 333         struct dm_hw_stat_delta *tmp;
 334
 335         switch (event) {
 336         case NETDEV_REGISTER:
 337                 new_stat = kzalloc(sizeof(struct dm_hw_stat_delta), GFP_KERNEL);
 338
 339                 if (!new_stat)
 340                         goto out;
 341
 342                 new_stat->dev = dev;
 343                 new_stat->last_rx = jiffies;
 344                 mutex_lock(&trace_state_mutex);
 345                 list_add_rcu(&new_stat->list, &hw_stats_list);
 346                 mutex_unlock(&trace_state_mutex);
 347                 break;
 348         case NETDEV_UNREGISTER:
 349                 mutex_lock(&trace_state_mutex);
 350                 list_for_each_entry_safe(new_stat, tmp, &hw_stats_list, list) {
 351                         if (new_stat->dev == dev) {
 352
 353                                 /* Paired with READ_ONCE() in trace_napi_poll_hit() */
 354                                 WRITE_ONCE(new_stat->dev, NULL);
 355
 356                                 if (trace_state == TRACE_OFF) {
 357                                         list_del_rcu(&new_stat->list);
 358                                         kfree_rcu(new_stat, rcu);
 359                                         break;
 360                                 }
 361                         }
 362                 }
 363                 mutex_unlock(&trace_state_mutex);
 364                 break;
 365         }
 366 out:
 367         return NOTIFY_DONE;
 368 }
 369
 370 static const struct genl_ops dropmon_ops[] = {
 371         {
 372                 .cmd = NET_DM_CMD_CONFIG,
 373                 .doit = net_dm_cmd_config,
 374         },
 375         {
 376                 .cmd = NET_DM_CMD_START,
 377                 .doit = net_dm_cmd_trace,
 378         },
 379         {
 380                 .cmd = NET_DM_CMD_STOP,
 381                 .doit = net_dm_cmd_trace,
 382         },
 383 };
 384
 385 static struct notifier_block dropmon_net_notifier = {
 386         .notifier_call = dropmon_net_event
 387 };
 388
 389 static int __init init_net_drop_monitor(void)
 390 {
 391         struct per_cpu_dm_data *data;
 392         int cpu, rc;
 393
 394         pr_info("Initializing network drop monitor service\n");
 395
 396         if (sizeof(void *) > 8) {
 397                 pr_err("Unable to store program counters on this arch, Drop monitor failed\n");
 398                 return -ENOSPC;
 399         }
 400
 401         rc = genl_register_family_with_ops_groups(&net_drop_monitor_family,
 402                                                   dropmon_ops, dropmon_mcgrps);
 403         if (rc) {
 404                 pr_err("Could not create drop monitor netlink family\n");
 405                 return rc;
 406         }
 407         WARN_ON(net_drop_monitor_family.mcgrp_offset != NET_DM_GRP_ALERT);
 408
 409         rc = register_netdevice_notifier(&dropmon_net_notifier);
 410         if (rc < 0) {
 411                 pr_crit("Failed to register netdevice notifier\n");
 412                 goto out_unreg;
 413         }
 414
 415         rc = 0;
 416
 417         for_each_possible_cpu(cpu) {
 418                 data = &per_cpu(dm_cpu_data, cpu);
 419                 INIT_WORK(&data->dm_alert_work, send_dm_alert);
 420                 init_timer(&data->send_timer);
 421                 data->send_timer.data = (unsigned long)data;
 422                 data->send_timer.function = sched_send_work;
 423                 spin_lock_init(&data->lock);
 424                 reset_per_cpu_data(data);
 425         }
 426
 427
 428         goto out;
 429
 430 out_unreg:
 431         genl_unregister_family(&net_drop_monitor_family);
 432 out:
 433         return rc;
 434 }
 435
 436 static void exit_net_drop_monitor(void)
 437 {
 438         struct per_cpu_dm_data *data;
 439         int cpu;
 440
 441         BUG_ON(unregister_netdevice_notifier(&dropmon_net_notifier));
 442
 443         /*
 444          * Because of the module_get/put we do in the trace state change path
 445          * we are guarnateed not to have any current users when we get here
 446          * all we need to do is make sure that we don't have any running timers
 447          * or pending schedule calls
 448          */
 449
 450         for_each_possible_cpu(cpu) {
 451                 data = &per_cpu(dm_cpu_data, cpu);
 452                 del_timer_sync(&data->send_timer);
 453                 cancel_work_sync(&data->dm_alert_work);
 454                 /*
 455                  * At this point, we should have exclusive access
 456                  * to this struct and can free the skb inside it
 457                  */
 458                 kfree_skb(data->skb);
 459         }
 460
 461         BUG_ON(genl_unregister_family(&net_drop_monitor_family));
 462 }
 463
 464 module_init(init_net_drop_monitor);
 465 module_exit(exit_net_drop_monitor);
 466
 467 MODULE_LICENSE("GPL v2");
 468 MODULE_AUTHOR("Neil Horman <nhorman@tuxdriver.com>");
 469 MODULE_ALIAS_GENL_FAMILY("NET_DM");