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
  64 static DEFINE_PER_CPU(struct per_cpu_dm_data, dm_cpu_data);
  65
  66 static int dm_hit_limit = 64;
  67 static int dm_delay = 1;
  68 static unsigned long dm_hw_check_delta = 2*HZ;
  69 static LIST_HEAD(hw_stats_list);
  70
  71 static struct sk_buff *reset_per_cpu_data(struct per_cpu_dm_data *data)
  72 {
  73         size_t al;
  74         struct net_dm_alert_msg *msg;
  75         struct nlattr *nla;
  76         struct sk_buff *skb;
  77         unsigned long flags;
  78
  79         al = sizeof(struct net_dm_alert_msg);
  80         al += dm_hit_limit * sizeof(struct net_dm_drop_point);
  81         al += sizeof(struct nlattr);
  82
  83         skb = genlmsg_new(al, GFP_KERNEL);
  84
  85         if (skb) {
  86                 genlmsg_put(skb, 0, 0, &net_drop_monitor_family,
  87                                 0, NET_DM_CMD_ALERT);
  88                 nla = nla_reserve(skb, NLA_UNSPEC,
  89                                   sizeof(struct net_dm_alert_msg));
  90                 msg = nla_data(nla);
  91                 memset(msg, 0, al);
  92         } else {
  93                 mod_timer(&data->send_timer, jiffies + HZ / 10);
  94         }
  95
  96         spin_lock_irqsave(&data->lock, flags);
  97         swap(data->skb, skb);
  98         spin_unlock_irqrestore(&data->lock, flags);
  99
 100         return skb;
 101 }
 102
 103 static const struct genl_multicast_group dropmon_mcgrps[] = {
 104         { .name = "events", },
 105 };
 106
 107 static void send_dm_alert(struct work_struct *work)
 108 {
 109         struct sk_buff *skb;
 110         struct per_cpu_dm_data *data;
 111
 112         data = container_of(work, struct per_cpu_dm_data, dm_alert_work);
 113
 114         skb = reset_per_cpu_data(data);
 115
 116         if (skb)
 117                 genlmsg_multicast(&net_drop_monitor_family, skb, 0,
 118                                   0, 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 = this_cpu_ptr(&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                                 int work, int budget)
 187 {
 188         struct dm_hw_stat_delta *new_stat;
 189
 190         /*
 191          * Don't check napi structures with no associated device
 192          */
 193         if (!napi->dev)
 194                 return;
 195
 196         rcu_read_lock();
 197         list_for_each_entry_rcu(new_stat, &hw_stats_list, list) {
 198                 /*
 199                  * only add a note to our monitor buffer if:
 200                  * 1) this is the dev we received on
 201                  * 2) its after the last_rx delta
 202                  * 3) our rx_dropped count has gone up
 203                  */
 204                 if ((new_stat->dev == napi->dev)  &&
 205                     (time_after(jiffies, new_stat->last_rx + dm_hw_check_delta)) &&
 206                     (napi->dev->stats.rx_dropped != new_stat->last_drop_val)) {
 207                         trace_drop_common(NULL, NULL);
 208                         new_stat->last_drop_val = napi->dev->stats.rx_dropped;
 209                         new_stat->last_rx = jiffies;
 210                         break;
 211                 }
 212         }
 213         rcu_read_unlock();
 214 }
 215
 216 static int set_all_monitor_traces(int state)
 217 {
 218         int rc = 0;
 219         struct dm_hw_stat_delta *new_stat = NULL;
 220         struct dm_hw_stat_delta *temp;
 221
 222         mutex_lock(&trace_state_mutex);
 223
 224         if (state == trace_state) {
 225                 rc = -EAGAIN;
 226                 goto out_unlock;
 227         }
 228
 229         switch (state) {
 230         case TRACE_ON:
 231                 if (!try_module_get(THIS_MODULE)) {
 232                         rc = -ENODEV;
 233                         break;
 234                 }
 235
 236                 rc |= register_trace_kfree_skb(trace_kfree_skb_hit, NULL);
 237                 rc |= register_trace_napi_poll(trace_napi_poll_hit, NULL);
 238                 break;
 239
 240         case TRACE_OFF:
 241                 rc |= unregister_trace_kfree_skb(trace_kfree_skb_hit, NULL);
 242                 rc |= unregister_trace_napi_poll(trace_napi_poll_hit, NULL);
 243
 244                 tracepoint_synchronize_unregister();
 245
 246                 /*
 247                  * Clean the device list
 248                  */
 249                 list_for_each_entry_safe(new_stat, temp, &hw_stats_list, list) {
 250                         if (new_stat->dev == NULL) {
 251                                 list_del_rcu(&new_stat->list);
 252                                 kfree_rcu(new_stat, rcu);
 253                         }
 254                 }
 255
 256                 module_put(THIS_MODULE);
 257
 258                 break;
 259         default:
 260                 rc = 1;
 261                 break;
 262         }
 263
 264         if (!rc)
 265                 trace_state = state;
 266         else
 267                 rc = -EINPROGRESS;
 268
 269 out_unlock:
 270         mutex_unlock(&trace_state_mutex);
 271
 272         return rc;
 273 }
 274
 275
 276 static int net_dm_cmd_config(struct sk_buff *skb,
 277                         struct genl_info *info)
 278 {
 279         return -ENOTSUPP;
 280 }
 281
 282 static int net_dm_cmd_trace(struct sk_buff *skb,
 283                         struct genl_info *info)
 284 {
 285         switch (info->genlhdr->cmd) {
 286         case NET_DM_CMD_START:
 287                 return set_all_monitor_traces(TRACE_ON);
 288         case NET_DM_CMD_STOP:
 289                 return set_all_monitor_traces(TRACE_OFF);
 290         }
 291
 292         return -ENOTSUPP;
 293 }
 294
 295 static int dropmon_net_event(struct notifier_block *ev_block,
 296                              unsigned long event, void *ptr)
 297 {
 298         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
 299         struct dm_hw_stat_delta *new_stat = NULL;
 300         struct dm_hw_stat_delta *tmp;
 301
 302         switch (event) {
 303         case NETDEV_REGISTER:
 304                 new_stat = kzalloc(sizeof(struct dm_hw_stat_delta), GFP_KERNEL);
 305
 306                 if (!new_stat)
 307                         goto out;
 308
 309                 new_stat->dev = dev;
 310                 new_stat->last_rx = jiffies;
 311                 mutex_lock(&trace_state_mutex);
 312                 list_add_rcu(&new_stat->list, &hw_stats_list);
 313                 mutex_unlock(&trace_state_mutex);
 314                 break;
 315         case NETDEV_UNREGISTER:
 316                 mutex_lock(&trace_state_mutex);
 317                 list_for_each_entry_safe(new_stat, tmp, &hw_stats_list, list) {
 318                         if (new_stat->dev == dev) {
 319                                 new_stat->dev = NULL;
 320                                 if (trace_state == TRACE_OFF) {
 321                                         list_del_rcu(&new_stat->list);
 322                                         kfree_rcu(new_stat, rcu);
 323                                         break;
 324                                 }
 325                         }
 326                 }
 327                 mutex_unlock(&trace_state_mutex);
 328                 break;
 329         }
 330 out:
 331         return NOTIFY_DONE;
 332 }
 333
 334 static const struct genl_ops dropmon_ops[] = {
 335         {
 336                 .cmd = NET_DM_CMD_CONFIG,
 337                 .doit = net_dm_cmd_config,
 338         },
 339         {
 340                 .cmd = NET_DM_CMD_START,
 341                 .doit = net_dm_cmd_trace,
 342         },
 343         {
 344                 .cmd = NET_DM_CMD_STOP,
 345                 .doit = net_dm_cmd_trace,
 346         },
 347 };
 348
 349 static struct genl_family net_drop_monitor_family __ro_after_init = {
 350         .hdrsize        = 0,
 351         .name           = "NET_DM",
 352         .version        = 2,
 353         .module         = THIS_MODULE,
 354         .ops            = dropmon_ops,
 355         .n_ops          = ARRAY_SIZE(dropmon_ops),
 356         .mcgrps         = dropmon_mcgrps,
 357         .n_mcgrps       = ARRAY_SIZE(dropmon_mcgrps),
 358 };
 359
 360 static struct notifier_block dropmon_net_notifier = {
 361         .notifier_call = dropmon_net_event
 362 };
 363
 364 static int __init init_net_drop_monitor(void)
 365 {
 366         struct per_cpu_dm_data *data;
 367         int cpu, rc;
 368
 369         pr_info("Initializing network drop monitor service\n");
 370
 371         if (sizeof(void *) > 8) {
 372                 pr_err("Unable to store program counters on this arch, Drop monitor failed\n");
 373                 return -ENOSPC;
 374         }
 375
 376         rc = genl_register_family(&net_drop_monitor_family);
 377         if (rc) {
 378                 pr_err("Could not create drop monitor netlink family\n");
 379                 return rc;
 380         }
 381         WARN_ON(net_drop_monitor_family.mcgrp_offset != NET_DM_GRP_ALERT);
 382
 383         rc = register_netdevice_notifier(&dropmon_net_notifier);
 384         if (rc < 0) {
 385                 pr_crit("Failed to register netdevice notifier\n");
 386                 goto out_unreg;
 387         }
 388
 389         rc = 0;
 390
 391         for_each_possible_cpu(cpu) {
 392                 data = &per_cpu(dm_cpu_data, cpu);
 393                 INIT_WORK(&data->dm_alert_work, send_dm_alert);
 394                 init_timer(&data->send_timer);
 395                 data->send_timer.data = (unsigned long)data;
 396                 data->send_timer.function = sched_send_work;
 397                 spin_lock_init(&data->lock);
 398                 reset_per_cpu_data(data);
 399         }
 400
 401
 402         goto out;
 403
 404 out_unreg:
 405         genl_unregister_family(&net_drop_monitor_family);
 406 out:
 407         return rc;
 408 }
 409
 410 static void exit_net_drop_monitor(void)
 411 {
 412         struct per_cpu_dm_data *data;
 413         int cpu;
 414
 415         BUG_ON(unregister_netdevice_notifier(&dropmon_net_notifier));
 416
 417         /*
 418          * Because of the module_get/put we do in the trace state change path
 419          * we are guarnateed not to have any current users when we get here
 420          * all we need to do is make sure that we don't have any running timers
 421          * or pending schedule calls
 422          */
 423
 424         for_each_possible_cpu(cpu) {
 425                 data = &per_cpu(dm_cpu_data, cpu);
 426                 del_timer_sync(&data->send_timer);
 427                 cancel_work_sync(&data->dm_alert_work);
 428                 /*
 429                  * At this point, we should have exclusive access
 430                  * to this struct and can free the skb inside it
 431                  */
 432                 kfree_skb(data->skb);
 433         }
 434
 435         BUG_ON(genl_unregister_family(&net_drop_monitor_family));
 436 }
 437
 438 module_init(init_net_drop_monitor);
 439 module_exit(exit_net_drop_monitor);
 440
 441 MODULE_LICENSE("GPL v2");
 442 MODULE_AUTHOR("Neil Horman <nhorman@tuxdriver.com>");
 443 MODULE_ALIAS_GENL_FAMILY("NET_DM");