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 static void send_dm_alert(struct work_struct *unused);
  40
  41
  42 /*
  43  * Globals, our netlink socket pointer
  44  * and the work handle that will send up
  45  * netlink alerts
  46  */
  47 static int trace_state = TRACE_OFF;
  48 static DEFINE_MUTEX(trace_state_mutex);
  49
  50 struct per_cpu_dm_data {
  51         struct work_struct dm_alert_work;
  52         struct sk_buff __rcu *skb;
  53         atomic_t dm_hit_count;
  54         struct timer_list send_timer;
  55         int cpu;
  56 };
  57
  58 struct dm_hw_stat_delta {
  59         struct net_device *dev;
  60         unsigned long last_rx;
  61         struct list_head list;
  62         struct rcu_head rcu;
  63         unsigned long last_drop_val;
  64 };
  65
  66 static struct genl_family net_drop_monitor_family = {
  67         .id             = GENL_ID_GENERATE,
  68         .hdrsize        = 0,
  69         .name           = "NET_DM",
  70         .version        = 2,
  71         .maxattr        = NET_DM_CMD_MAX,
  72 };
  73
  74 static DEFINE_PER_CPU(struct per_cpu_dm_data, dm_cpu_data);
  75
  76 static int dm_hit_limit = 64;
  77 static int dm_delay = 1;
  78 static unsigned long dm_hw_check_delta = 2*HZ;
  79 static LIST_HEAD(hw_stats_list);
  80
  81 static void reset_per_cpu_data(struct per_cpu_dm_data *data)
  82 {
  83         size_t al;
  84         struct net_dm_alert_msg *msg;
  85         struct nlattr *nla;
  86         struct sk_buff *skb;
  87         struct sk_buff *oskb = rcu_dereference_protected(data->skb, 1);
  88
  89         al = sizeof(struct net_dm_alert_msg);
  90         al += dm_hit_limit * sizeof(struct net_dm_drop_point);
  91         al += sizeof(struct nlattr);
  92
  93         skb = genlmsg_new(al, GFP_KERNEL);
  94
  95         if (skb) {
  96                 genlmsg_put(skb, 0, 0, &net_drop_monitor_family,
  97                                 0, NET_DM_CMD_ALERT);
  98                 nla = nla_reserve(skb, NLA_UNSPEC,
  99                                   sizeof(struct net_dm_alert_msg));
 100                 msg = nla_data(nla);
 101                 memset(msg, 0, al);
 102         } else
 103                 schedule_work_on(data->cpu, &data->dm_alert_work);
 104
 105         /*
 106          * Don't need to lock this, since we are guaranteed to only
 107          * run this on a single cpu at a time.
 108          * Note also that we only update data->skb if the old and new skb
 109          * pointers don't match.  This ensures that we don't continually call
 110          * synchornize_rcu if we repeatedly fail to alloc a new netlink message.
 111          */
 112         if (skb != oskb) {
 113                 rcu_assign_pointer(data->skb, skb);
 114
 115                 synchronize_rcu();
 116
 117                 atomic_set(&data->dm_hit_count, dm_hit_limit);
 118         }
 119
 120 }
 121
 122 static void send_dm_alert(struct work_struct *unused)
 123 {
 124         struct sk_buff *skb;
 125         struct per_cpu_dm_data *data = &get_cpu_var(dm_cpu_data);
 126
 127         WARN_ON_ONCE(data->cpu != smp_processor_id());
 128
 129         /*
 130          * Grab the skb we're about to send
 131          */
 132         skb = rcu_dereference_protected(data->skb, 1);
 133
 134         /*
 135          * Replace it with a new one
 136          */
 137         reset_per_cpu_data(data);
 138
 139         /*
 140          * Ship it!
 141          */
 142         if (skb)
 143                 genlmsg_multicast(skb, 0, NET_DM_GRP_ALERT, GFP_KERNEL);
 144
 145         put_cpu_var(dm_cpu_data);
 146 }
 147
 148 /*
 149  * This is the timer function to delay the sending of an alert
 150  * in the event that more drops will arrive during the
 151  * hysteresis period.  Note that it operates under the timer interrupt
 152  * so we don't need to disable preemption here
 153  */
 154 static void sched_send_work(unsigned long unused)
 155 {
 156         struct per_cpu_dm_data *data =  &get_cpu_var(dm_cpu_data);
 157
 158         schedule_work_on(smp_processor_id(), &data->dm_alert_work);
 159
 160         put_cpu_var(dm_cpu_data);
 161 }
 162
 163 static void trace_drop_common(struct sk_buff *skb, void *location)
 164 {
 165         struct net_dm_alert_msg *msg;
 166         struct nlmsghdr *nlh;
 167         struct nlattr *nla;
 168         int i;
 169         struct sk_buff *dskb;
 170         struct per_cpu_dm_data *data = &get_cpu_var(dm_cpu_data);
 171
 172
 173         rcu_read_lock();
 174         dskb = rcu_dereference(data->skb);
 175
 176         if (!dskb)
 177                 goto out;
 178
 179         if (!atomic_add_unless(&data->dm_hit_count, -1, 0)) {
 180                 /*
 181                  * we're already at zero, discard this hit
 182                  */
 183                 goto out;
 184         }
 185
 186         nlh = (struct nlmsghdr *)dskb->data;
 187         nla = genlmsg_data(nlmsg_data(nlh));
 188         msg = nla_data(nla);
 189         for (i = 0; i < msg->entries; i++) {
 190                 if (!memcmp(&location, msg->points[i].pc, sizeof(void *))) {
 191                         msg->points[i].count++;
 192                         atomic_inc(&data->dm_hit_count);
 193                         goto out;
 194                 }
 195         }
 196
 197         /*
 198          * We need to create a new entry
 199          */
 200         __nla_reserve_nohdr(dskb, sizeof(struct net_dm_drop_point));
 201         nla->nla_len += NLA_ALIGN(sizeof(struct net_dm_drop_point));
 202         memcpy(msg->points[msg->entries].pc, &location, sizeof(void *));
 203         msg->points[msg->entries].count = 1;
 204         msg->entries++;
 205
 206         if (!timer_pending(&data->send_timer)) {
 207                 data->send_timer.expires = jiffies + dm_delay * HZ;
 208                 add_timer_on(&data->send_timer, smp_processor_id());
 209         }
 210
 211 out:
 212         rcu_read_unlock();
 213         put_cpu_var(dm_cpu_data);
 214         return;
 215 }
 216
 217 static void trace_kfree_skb_hit(void *ignore, struct sk_buff *skb, void *location)
 218 {
 219         trace_drop_common(skb, location);
 220 }
 221
 222 static void trace_napi_poll_hit(void *ignore, struct napi_struct *napi)
 223 {
 224         struct dm_hw_stat_delta *new_stat;
 225
 226         /*
 227          * Don't check napi structures with no associated device
 228          */
 229         if (!napi->dev)
 230                 return;
 231
 232         rcu_read_lock();
 233         list_for_each_entry_rcu(new_stat, &hw_stats_list, list) {
 234                 /*
 235                  * only add a note to our monitor buffer if:
 236                  * 1) this is the dev we received on
 237                  * 2) its after the last_rx delta
 238                  * 3) our rx_dropped count has gone up
 239                  */
 240                 if ((new_stat->dev == napi->dev)  &&
 241                     (time_after(jiffies, new_stat->last_rx + dm_hw_check_delta)) &&
 242                     (napi->dev->stats.rx_dropped != new_stat->last_drop_val)) {
 243                         trace_drop_common(NULL, NULL);
 244                         new_stat->last_drop_val = napi->dev->stats.rx_dropped;
 245                         new_stat->last_rx = jiffies;
 246                         break;
 247                 }
 248         }
 249         rcu_read_unlock();
 250 }
 251
 252 static int set_all_monitor_traces(int state)
 253 {
 254         int rc = 0;
 255         struct dm_hw_stat_delta *new_stat = NULL;
 256         struct dm_hw_stat_delta *temp;
 257
 258         mutex_lock(&trace_state_mutex);
 259
 260         if (state == trace_state) {
 261                 rc = -EAGAIN;
 262                 goto out_unlock;
 263         }
 264
 265         switch (state) {
 266         case TRACE_ON:
 267                 if (!try_module_get(THIS_MODULE)) {
 268                         rc = -ENODEV;
 269                         break;
 270                 }
 271
 272                 rc |= register_trace_kfree_skb(trace_kfree_skb_hit, NULL);
 273                 rc |= register_trace_napi_poll(trace_napi_poll_hit, NULL);
 274                 break;
 275
 276         case TRACE_OFF:
 277                 rc |= unregister_trace_kfree_skb(trace_kfree_skb_hit, NULL);
 278                 rc |= unregister_trace_napi_poll(trace_napi_poll_hit, NULL);
 279
 280                 tracepoint_synchronize_unregister();
 281
 282                 /*
 283                  * Clean the device list
 284                  */
 285                 list_for_each_entry_safe(new_stat, temp, &hw_stats_list, list) {
 286                         if (new_stat->dev == NULL) {
 287                                 list_del_rcu(&new_stat->list);
 288                                 kfree_rcu(new_stat, rcu);
 289                         }
 290                 }
 291
 292                 module_put(THIS_MODULE);
 293
 294                 break;
 295         default:
 296                 rc = 1;
 297                 break;
 298         }
 299
 300         if (!rc)
 301                 trace_state = state;
 302         else
 303                 rc = -EINPROGRESS;
 304
 305 out_unlock:
 306         mutex_unlock(&trace_state_mutex);
 307
 308         return rc;
 309 }
 310
 311
 312 static int net_dm_cmd_config(struct sk_buff *skb,
 313                         struct genl_info *info)
 314 {
 315         return -ENOTSUPP;
 316 }
 317
 318 static int net_dm_cmd_trace(struct sk_buff *skb,
 319                         struct genl_info *info)
 320 {
 321         switch (info->genlhdr->cmd) {
 322         case NET_DM_CMD_START:
 323                 return set_all_monitor_traces(TRACE_ON);
 324                 break;
 325         case NET_DM_CMD_STOP:
 326                 return set_all_monitor_traces(TRACE_OFF);
 327                 break;
 328         }
 329
 330         return -ENOTSUPP;
 331 }
 332
 333 static int dropmon_net_event(struct notifier_block *ev_block,
 334                         unsigned long event, void *ptr)
 335 {
 336         struct net_device *dev = ptr;
 337         struct dm_hw_stat_delta *new_stat = NULL;
 338         struct dm_hw_stat_delta *tmp;
 339
 340         switch (event) {
 341         case NETDEV_REGISTER:
 342                 new_stat = kzalloc(sizeof(struct dm_hw_stat_delta), GFP_KERNEL);
 343
 344                 if (!new_stat)
 345                         goto out;
 346
 347                 new_stat->dev = dev;
 348                 new_stat->last_rx = jiffies;
 349                 mutex_lock(&trace_state_mutex);
 350                 list_add_rcu(&new_stat->list, &hw_stats_list);
 351                 mutex_unlock(&trace_state_mutex);
 352                 break;
 353         case NETDEV_UNREGISTER:
 354                 mutex_lock(&trace_state_mutex);
 355                 list_for_each_entry_safe(new_stat, tmp, &hw_stats_list, list) {
 356                         if (new_stat->dev == dev) {
 357                                 new_stat->dev = NULL;
 358                                 if (trace_state == TRACE_OFF) {
 359                                         list_del_rcu(&new_stat->list);
 360                                         kfree_rcu(new_stat, rcu);
 361                                         break;
 362                                 }
 363                         }
 364                 }
 365                 mutex_unlock(&trace_state_mutex);
 366                 break;
 367         }
 368 out:
 369         return NOTIFY_DONE;
 370 }
 371
 372 static struct genl_ops dropmon_ops[] = {
 373         {
 374                 .cmd = NET_DM_CMD_CONFIG,
 375                 .doit = net_dm_cmd_config,
 376         },
 377         {
 378                 .cmd = NET_DM_CMD_START,
 379                 .doit = net_dm_cmd_trace,
 380         },
 381         {
 382                 .cmd = NET_DM_CMD_STOP,
 383                 .doit = net_dm_cmd_trace,
 384         },
 385 };
 386
 387 static struct notifier_block dropmon_net_notifier = {
 388         .notifier_call = dropmon_net_event
 389 };
 390
 391 static int __init init_net_drop_monitor(void)
 392 {
 393         struct per_cpu_dm_data *data;
 394         int cpu, rc;
 395
 396         pr_info("Initializing network drop monitor service\n");
 397
 398         if (sizeof(void *) > 8) {
 399                 pr_err("Unable to store program counters on this arch, Drop monitor failed\n");
 400                 return -ENOSPC;
 401         }
 402
 403         rc = genl_register_family_with_ops(&net_drop_monitor_family,
 404                                            dropmon_ops,
 405                                            ARRAY_SIZE(dropmon_ops));
 406         if (rc) {
 407                 pr_err("Could not create drop monitor netlink family\n");
 408                 return rc;
 409         }
 410
 411         rc = register_netdevice_notifier(&dropmon_net_notifier);
 412         if (rc < 0) {
 413                 pr_crit("Failed to register netdevice notifier\n");
 414                 goto out_unreg;
 415         }
 416
 417         rc = 0;
 418
 419         for_each_possible_cpu(cpu) {
 420                 data = &per_cpu(dm_cpu_data, cpu);
 421                 data->cpu = cpu;
 422                 INIT_WORK(&data->dm_alert_work, send_dm_alert);
 423                 init_timer(&data->send_timer);
 424                 data->send_timer.data = cpu;
 425                 data->send_timer.function = sched_send_work;
 426                 reset_per_cpu_data(data);
 427         }
 428
 429
 430         goto out;
 431
 432 out_unreg:
 433         genl_unregister_family(&net_drop_monitor_family);
 434 out:
 435         return rc;
 436 }
 437
 438 static void exit_net_drop_monitor(void)
 439 {
 440         struct per_cpu_dm_data *data;
 441         int cpu;
 442
 443         BUG_ON(unregister_netdevice_notifier(&dropmon_net_notifier));
 444
 445         /*
 446          * Because of the module_get/put we do in the trace state change path
 447          * we are guarnateed not to have any current users when we get here
 448          * all we need to do is make sure that we don't have any running timers
 449          * or pending schedule calls
 450          */
 451
 452         for_each_possible_cpu(cpu) {
 453                 data = &per_cpu(dm_cpu_data, cpu);
 454                 del_timer_sync(&data->send_timer);
 455                 cancel_work_sync(&data->dm_alert_work);
 456                 /*
 457                  * At this point, we should have exclusive access
 458                  * to this struct and can free the skb inside it
 459                  */
 460                 kfree_skb(data->skb);
 461         }
 462
 463         BUG_ON(genl_unregister_family(&net_drop_monitor_family));
 464 }
 465
 466 module_init(init_net_drop_monitor);
 467 module_exit(exit_net_drop_monitor);
 468
 469 MODULE_LICENSE("GPL v2");
 470 MODULE_AUTHOR("Neil Horman <nhorman@tuxdriver.com>");
 471 MODULE_ALIAS_GENL_FAMILY("NET_DM");