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 <net/genetlink.h>
  28 #include <net/netevent.h>
  29
  30 #include <trace/events/skb.h>
  31 #include <trace/events/napi.h>
  32
  33 #include <asm/unaligned.h>
  34
  35 #define TRACE_ON 1
  36 #define TRACE_OFF 0
  37
  38 static void send_dm_alert(struct work_struct *unused);
  39
  40
  41 /*
  42  * Globals, our netlink socket pointer
  43  * and the work handle that will send up
  44  * netlink alerts
  45  */
  46 static int trace_state = TRACE_OFF;
  47 static DEFINE_MUTEX(trace_state_mutex);
  48
  49 struct per_cpu_dm_data {
  50         struct work_struct dm_alert_work;
  51         struct sk_buff __rcu *skb;
  52         atomic_t dm_hit_count;
  53         struct timer_list send_timer;
  54         int cpu;
  55 };
  56
  57 struct dm_hw_stat_delta {
  58         struct net_device *dev;
  59         unsigned long last_rx;
  60         struct list_head list;
  61         struct rcu_head rcu;
  62         unsigned long last_drop_val;
  63 };
  64
  65 static struct genl_family net_drop_monitor_family = {
  66         .id             = GENL_ID_GENERATE,
  67         .hdrsize        = 0,
  68         .name           = "NET_DM",
  69         .version        = 2,
  70         .maxattr        = NET_DM_CMD_MAX,
  71 };
  72
  73 static DEFINE_PER_CPU(struct per_cpu_dm_data, dm_cpu_data);
  74
  75 static int dm_hit_limit = 64;
  76 static int dm_delay = 1;
  77 static unsigned long dm_hw_check_delta = 2*HZ;
  78 static LIST_HEAD(hw_stats_list);
  79
  80 static void reset_per_cpu_data(struct per_cpu_dm_data *data)
  81 {
  82         size_t al;
  83         struct net_dm_alert_msg *msg;
  84         struct nlattr *nla;
  85         struct sk_buff *skb;
  86         struct sk_buff *oskb = rcu_dereference_protected(data->skb, 1);
  87
  88         al = sizeof(struct net_dm_alert_msg);
  89         al += dm_hit_limit * sizeof(struct net_dm_drop_point);
  90         al += sizeof(struct nlattr);
  91
  92         skb = genlmsg_new(al, GFP_KERNEL);
  93
  94         if (skb) {
  95                 genlmsg_put(skb, 0, 0, &net_drop_monitor_family,
  96                                 0, NET_DM_CMD_ALERT);
  97                 nla = nla_reserve(skb, NLA_UNSPEC,
  98                                   sizeof(struct net_dm_alert_msg));
  99                 msg = nla_data(nla);
 100                 memset(msg, 0, al);
 101         } else
 102                 schedule_work_on(data->cpu, &data->dm_alert_work);
 103
 104         /*
 105          * Don't need to lock this, since we are guaranteed to only
 106          * run this on a single cpu at a time.
 107          * Note also that we only update data->skb if the old and new skb
 108          * pointers don't match.  This ensures that we don't continually call
 109          * synchornize_rcu if we repeatedly fail to alloc a new netlink message.
 110          */
 111         if (skb != oskb) {
 112                 rcu_assign_pointer(data->skb, skb);
 113
 114                 synchronize_rcu();
 115
 116                 atomic_set(&data->dm_hit_count, dm_hit_limit);
 117         }
 118
 119 }
 120
 121 static void send_dm_alert(struct work_struct *unused)
 122 {
 123         struct sk_buff *skb;
 124         struct per_cpu_dm_data *data = &get_cpu_var(dm_cpu_data);
 125
 126         WARN_ON_ONCE(data->cpu != smp_processor_id());
 127
 128         /*
 129          * Grab the skb we're about to send
 130          */
 131         skb = rcu_dereference_protected(data->skb, 1);
 132
 133         /*
 134          * Replace it with a new one
 135          */
 136         reset_per_cpu_data(data);
 137
 138         /*
 139          * Ship it!
 140          */
 141         if (skb)
 142                 genlmsg_multicast(skb, 0, NET_DM_GRP_ALERT, GFP_KERNEL);
 143
 144         put_cpu_var(dm_cpu_data);
 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.  Note that it operates under the timer interrupt
 151  * so we don't need to disable preemption here
 152  */
 153 static void sched_send_work(unsigned long unused)
 154 {
 155         struct per_cpu_dm_data *data =  &get_cpu_var(dm_cpu_data);
 156
 157         schedule_work_on(smp_processor_id(), &data->dm_alert_work);
 158
 159         put_cpu_var(dm_cpu_data);
 160 }
 161
 162 static void trace_drop_common(struct sk_buff *skb, void *location)
 163 {
 164         struct net_dm_alert_msg *msg;
 165         struct nlmsghdr *nlh;
 166         struct nlattr *nla;
 167         int i;
 168         struct sk_buff *dskb;
 169         struct per_cpu_dm_data *data = &get_cpu_var(dm_cpu_data);
 170
 171
 172         rcu_read_lock();
 173         dskb = rcu_dereference(data->skb);
 174
 175         if (!dskb)
 176                 goto out;
 177
 178         if (!atomic_add_unless(&data->dm_hit_count, -1, 0)) {
 179                 /*
 180                  * we're already at zero, discard this hit
 181                  */
 182                 goto out;
 183         }
 184
 185         nlh = (struct nlmsghdr *)dskb->data;
 186         nla = genlmsg_data(nlmsg_data(nlh));
 187         msg = nla_data(nla);
 188         for (i = 0; i < msg->entries; i++) {
 189                 if (!memcmp(&location, msg->points[i].pc, sizeof(void *))) {
 190                         msg->points[i].count++;
 191                         atomic_inc(&data->dm_hit_count);
 192                         goto out;
 193                 }
 194         }
 195
 196         /*
 197          * We need to create a new entry
 198          */
 199         __nla_reserve_nohdr(dskb, sizeof(struct net_dm_drop_point));
 200         nla->nla_len += NLA_ALIGN(sizeof(struct net_dm_drop_point));
 201         memcpy(msg->points[msg->entries].pc, &location, sizeof(void *));
 202         msg->points[msg->entries].count = 1;
 203         msg->entries++;
 204
 205         if (!timer_pending(&data->send_timer)) {
 206                 data->send_timer.expires = jiffies + dm_delay * HZ;
 207                 add_timer_on(&data->send_timer, smp_processor_id());
 208         }
 209
 210 out:
 211         rcu_read_unlock();
 212         put_cpu_var(dm_cpu_data);
 213         return;
 214 }
 215
 216 static void trace_kfree_skb_hit(void *ignore, struct sk_buff *skb, void *location)
 217 {
 218         trace_drop_common(skb, location);
 219 }
 220
 221 static void trace_napi_poll_hit(void *ignore, struct napi_struct *napi)
 222 {
 223         struct dm_hw_stat_delta *new_stat;
 224
 225         /*
 226          * Don't check napi structures with no associated device
 227          */
 228         if (!napi->dev)
 229                 return;
 230
 231         rcu_read_lock();
 232         list_for_each_entry_rcu(new_stat, &hw_stats_list, list) {
 233                 /*
 234                  * only add a note to our monitor buffer if:
 235                  * 1) this is the dev we received on
 236                  * 2) its after the last_rx delta
 237                  * 3) our rx_dropped count has gone up
 238                  */
 239                 if ((new_stat->dev == napi->dev)  &&
 240                     (time_after(jiffies, new_stat->last_rx + dm_hw_check_delta)) &&
 241                     (napi->dev->stats.rx_dropped != new_stat->last_drop_val)) {
 242                         trace_drop_common(NULL, NULL);
 243                         new_stat->last_drop_val = napi->dev->stats.rx_dropped;
 244                         new_stat->last_rx = jiffies;
 245                         break;
 246                 }
 247         }
 248         rcu_read_unlock();
 249 }
 250
 251 static int set_all_monitor_traces(int state)
 252 {
 253         int rc = 0;
 254         struct dm_hw_stat_delta *new_stat = NULL;
 255         struct dm_hw_stat_delta *temp;
 256
 257         mutex_lock(&trace_state_mutex);
 258
 259         if (state == trace_state) {
 260                 rc = -EAGAIN;
 261                 goto out_unlock;
 262         }
 263
 264         switch (state) {
 265         case TRACE_ON:
 266                 rc |= register_trace_kfree_skb(trace_kfree_skb_hit, NULL);
 267                 rc |= register_trace_napi_poll(trace_napi_poll_hit, NULL);
 268                 break;
 269         case TRACE_OFF:
 270                 rc |= unregister_trace_kfree_skb(trace_kfree_skb_hit, NULL);
 271                 rc |= unregister_trace_napi_poll(trace_napi_poll_hit, NULL);
 272
 273                 tracepoint_synchronize_unregister();
 274
 275                 /*
 276                  * Clean the device list
 277                  */
 278                 list_for_each_entry_safe(new_stat, temp, &hw_stats_list, list) {
 279                         if (new_stat->dev == NULL) {
 280                                 list_del_rcu(&new_stat->list);
 281                                 kfree_rcu(new_stat, rcu);
 282                         }
 283                 }
 284                 break;
 285         default:
 286                 rc = 1;
 287                 break;
 288         }
 289
 290         if (!rc)
 291                 trace_state = state;
 292         else
 293                 rc = -EINPROGRESS;
 294
 295 out_unlock:
 296         mutex_unlock(&trace_state_mutex);
 297
 298         return rc;
 299 }
 300
 301
 302 static int net_dm_cmd_config(struct sk_buff *skb,
 303                         struct genl_info *info)
 304 {
 305         return -ENOTSUPP;
 306 }
 307
 308 static int net_dm_cmd_trace(struct sk_buff *skb,
 309                         struct genl_info *info)
 310 {
 311         switch (info->genlhdr->cmd) {
 312         case NET_DM_CMD_START:
 313                 return set_all_monitor_traces(TRACE_ON);
 314                 break;
 315         case NET_DM_CMD_STOP:
 316                 return set_all_monitor_traces(TRACE_OFF);
 317                 break;
 318         }
 319
 320         return -ENOTSUPP;
 321 }
 322
 323 static int dropmon_net_event(struct notifier_block *ev_block,
 324                         unsigned long event, void *ptr)
 325 {
 326         struct net_device *dev = ptr;
 327         struct dm_hw_stat_delta *new_stat = NULL;
 328         struct dm_hw_stat_delta *tmp;
 329
 330         switch (event) {
 331         case NETDEV_REGISTER:
 332                 new_stat = kzalloc(sizeof(struct dm_hw_stat_delta), GFP_KERNEL);
 333
 334                 if (!new_stat)
 335                         goto out;
 336
 337                 new_stat->dev = dev;
 338                 new_stat->last_rx = jiffies;
 339                 mutex_lock(&trace_state_mutex);
 340                 list_add_rcu(&new_stat->list, &hw_stats_list);
 341                 mutex_unlock(&trace_state_mutex);
 342                 break;
 343         case NETDEV_UNREGISTER:
 344                 mutex_lock(&trace_state_mutex);
 345                 list_for_each_entry_safe(new_stat, tmp, &hw_stats_list, list) {
 346                         if (new_stat->dev == dev) {
 347                                 new_stat->dev = NULL;
 348                                 if (trace_state == TRACE_OFF) {
 349                                         list_del_rcu(&new_stat->list);
 350                                         kfree_rcu(new_stat, rcu);
 351                                         break;
 352                                 }
 353                         }
 354                 }
 355                 mutex_unlock(&trace_state_mutex);
 356                 break;
 357         }
 358 out:
 359         return NOTIFY_DONE;
 360 }
 361
 362 static struct genl_ops dropmon_ops[] = {
 363         {
 364                 .cmd = NET_DM_CMD_CONFIG,
 365                 .doit = net_dm_cmd_config,
 366         },
 367         {
 368                 .cmd = NET_DM_CMD_START,
 369                 .doit = net_dm_cmd_trace,
 370         },
 371         {
 372                 .cmd = NET_DM_CMD_STOP,
 373                 .doit = net_dm_cmd_trace,
 374         },
 375 };
 376
 377 static struct notifier_block dropmon_net_notifier = {
 378         .notifier_call = dropmon_net_event
 379 };
 380
 381 static int __init init_net_drop_monitor(void)
 382 {
 383         struct per_cpu_dm_data *data;
 384         int cpu, rc;
 385
 386         pr_info("Initializing network drop monitor service\n");
 387
 388         if (sizeof(void *) > 8) {
 389                 pr_err("Unable to store program counters on this arch, Drop monitor failed\n");
 390                 return -ENOSPC;
 391         }
 392
 393         rc = genl_register_family_with_ops(&net_drop_monitor_family,
 394                                            dropmon_ops,
 395                                            ARRAY_SIZE(dropmon_ops));
 396         if (rc) {
 397                 pr_err("Could not create drop monitor netlink family\n");
 398                 return rc;
 399         }
 400
 401         rc = register_netdevice_notifier(&dropmon_net_notifier);
 402         if (rc < 0) {
 403                 pr_crit("Failed to register netdevice notifier\n");
 404                 goto out_unreg;
 405         }
 406
 407         rc = 0;
 408
 409         for_each_present_cpu(cpu) {
 410                 data = &per_cpu(dm_cpu_data, cpu);
 411                 data->cpu = cpu;
 412                 INIT_WORK(&data->dm_alert_work, send_dm_alert);
 413                 init_timer(&data->send_timer);
 414                 data->send_timer.data = cpu;
 415                 data->send_timer.function = sched_send_work;
 416                 reset_per_cpu_data(data);
 417         }
 418
 419
 420         goto out;
 421
 422 out_unreg:
 423         genl_unregister_family(&net_drop_monitor_family);
 424 out:
 425         return rc;
 426 }
 427
 428 late_initcall(init_net_drop_monitor);