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 #include <linux/netdevice.h>
   8 #include <linux/etherdevice.h>
   9 #include <linux/string.h>
  10 #include <linux/if_arp.h>
  11 #include <linux/inetdevice.h>
  12 #include <linux/inet.h>
  13 #include <linux/interrupt.h>
  14 #include <linux/netpoll.h>
  15 #include <linux/sched.h>
  16 #include <linux/delay.h>
  17 #include <linux/types.h>
  18 #include <linux/workqueue.h>
  19 #include <linux/netlink.h>
  20 #include <linux/net_dropmon.h>
  21 #include <linux/percpu.h>
  22 #include <linux/timer.h>
  23 #include <linux/bitops.h>
  24 #include <linux/slab.h>
  25 #include <net/genetlink.h>
  26 #include <net/netevent.h>
  27
  28 #include <trace/events/skb.h>
  29 #include <trace/events/napi.h>
  30
  31 #include <asm/unaligned.h>
  32
  33 #define TRACE_ON 1
  34 #define TRACE_OFF 0
  35
  36 static void send_dm_alert(struct work_struct *unused);
  37
  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         struct work_struct dm_alert_work;
  49         struct sk_buff __rcu *skb;
  50         atomic_t dm_hit_count;
  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         .maxattr        = NET_DM_CMD_MAX,
  68 };
  69
  70 static DEFINE_PER_CPU(struct per_cpu_dm_data, dm_cpu_data);
  71
  72 static int dm_hit_limit = 64;
  73 static int dm_delay = 1;
  74 static unsigned long dm_hw_check_delta = 2*HZ;
  75 static LIST_HEAD(hw_stats_list);
  76 static int initialized = 0;
  77
  78 static void reset_per_cpu_data(struct per_cpu_dm_data *data)
  79 {
  80         size_t al;
  81         struct net_dm_alert_msg *msg;
  82         struct nlattr *nla;
  83         struct sk_buff *skb;
  84         struct sk_buff *oskb = rcu_dereference_protected(data->skb, 1);
  85
  86         al = sizeof(struct net_dm_alert_msg);
  87         al += dm_hit_limit * sizeof(struct net_dm_drop_point);
  88         al += sizeof(struct nlattr);
  89
  90         skb = genlmsg_new(al, GFP_KERNEL);
  91
  92         if (skb) {
  93                 genlmsg_put(skb, 0, 0, &net_drop_monitor_family,
  94                                 0, NET_DM_CMD_ALERT);
  95                 nla = nla_reserve(skb, NLA_UNSPEC,
  96                                   sizeof(struct net_dm_alert_msg));
  97                 msg = nla_data(nla);
  98                 memset(msg, 0, al);
  99         } else if (initialized)
 100                 schedule_work_on(smp_processor_id(), &data->dm_alert_work);
 101
 102         /*
 103          * Don't need to lock this, since we are guaranteed to only
 104          * run this on a single cpu at a time.
 105          * Note also that we only update data->skb if the old and new skb
 106          * pointers don't match.  This ensures that we don't continually call
 107          * synchornize_rcu if we repeatedly fail to alloc a new netlink message.
 108          */
 109         if (skb != oskb) {
 110                 rcu_assign_pointer(data->skb, skb);
 111
 112                 synchronize_rcu();
 113
 114                 atomic_set(&data->dm_hit_count, dm_hit_limit);
 115         }
 116
 117 }
 118
 119 static void send_dm_alert(struct work_struct *unused)
 120 {
 121         struct sk_buff *skb;
 122         struct per_cpu_dm_data *data = &get_cpu_var(dm_cpu_data);
 123
 124         /*
 125          * Grab the skb we're about to send
 126          */
 127         skb = rcu_dereference_protected(data->skb, 1);
 128
 129         /*
 130          * Replace it with a new one
 131          */
 132         reset_per_cpu_data(data);
 133
 134         /*
 135          * Ship it!
 136          */
 137         if (skb)
 138                 genlmsg_multicast(skb, 0, NET_DM_GRP_ALERT, GFP_KERNEL);
 139
 140         put_cpu_var(dm_cpu_data);
 141 }
 142
 143 /*
 144  * This is the timer function to delay the sending of an alert
 145  * in the event that more drops will arrive during the
 146  * hysteresis period.  Note that it operates under the timer interrupt
 147  * so we don't need to disable preemption here
 148  */
 149 static void sched_send_work(unsigned long unused)
 150 {
 151         struct per_cpu_dm_data *data =  &get_cpu_var(dm_cpu_data);
 152
 153         schedule_work_on(smp_processor_id(), &data->dm_alert_work);
 154
 155         put_cpu_var(dm_cpu_data);
 156 }
 157
 158 static void trace_drop_common(struct sk_buff *skb, void *location)
 159 {
 160         struct net_dm_alert_msg *msg;
 161         struct nlmsghdr *nlh;
 162         struct nlattr *nla;
 163         int i;
 164         struct sk_buff *dskb;
 165         struct per_cpu_dm_data *data = &get_cpu_var(dm_cpu_data);
 166
 167
 168         rcu_read_lock();
 169         dskb = rcu_dereference(data->skb);
 170
 171         if (!dskb)
 172                 goto out;
 173
 174         if (!atomic_add_unless(&data->dm_hit_count, -1, 0)) {
 175                 /*
 176                  * we're already at zero, discard this hit
 177                  */
 178                 goto out;
 179         }
 180
 181         nlh = (struct nlmsghdr *)dskb->data;
 182         nla = genlmsg_data(nlmsg_data(nlh));
 183         msg = nla_data(nla);
 184         for (i = 0; i < msg->entries; i++) {
 185                 if (!memcmp(&location, msg->points[i].pc, sizeof(void *))) {
 186                         msg->points[i].count++;
 187                         atomic_inc(&data->dm_hit_count);
 188                         goto out;
 189                 }
 190         }
 191
 192         /*
 193          * We need to create a new entry
 194          */
 195         __nla_reserve_nohdr(dskb, sizeof(struct net_dm_drop_point));
 196         nla->nla_len += NLA_ALIGN(sizeof(struct net_dm_drop_point));
 197         memcpy(msg->points[msg->entries].pc, &location, sizeof(void *));
 198         msg->points[msg->entries].count = 1;
 199         msg->entries++;
 200
 201         if (!timer_pending(&data->send_timer)) {
 202                 data->send_timer.expires = jiffies + dm_delay * HZ;
 203                 add_timer_on(&data->send_timer, smp_processor_id());
 204         }
 205
 206 out:
 207         rcu_read_unlock();
 208         put_cpu_var(dm_cpu_data);
 209         return;
 210 }
 211
 212 static void trace_kfree_skb_hit(void *ignore, struct sk_buff *skb, void *location)
 213 {
 214         trace_drop_common(skb, location);
 215 }
 216
 217 static void trace_napi_poll_hit(void *ignore, struct napi_struct *napi)
 218 {
 219         struct dm_hw_stat_delta *new_stat;
 220
 221         /*
 222          * Don't check napi structures with no associated device
 223          */
 224         if (!napi->dev)
 225                 return;
 226
 227         rcu_read_lock();
 228         list_for_each_entry_rcu(new_stat, &hw_stats_list, list) {
 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                 if ((new_stat->dev == napi->dev)  &&
 236                     (time_after(jiffies, new_stat->last_rx + dm_hw_check_delta)) &&
 237                     (napi->dev->stats.rx_dropped != new_stat->last_drop_val)) {
 238                         trace_drop_common(NULL, NULL);
 239                         new_stat->last_drop_val = napi->dev->stats.rx_dropped;
 240                         new_stat->last_rx = jiffies;
 241                         break;
 242                 }
 243         }
 244         rcu_read_unlock();
 245 }
 246
 247 static int set_all_monitor_traces(int state)
 248 {
 249         int rc = 0;
 250         struct dm_hw_stat_delta *new_stat = NULL;
 251         struct dm_hw_stat_delta *temp;
 252
 253         mutex_lock(&trace_state_mutex);
 254
 255         if (state == trace_state) {
 256                 rc = -EAGAIN;
 257                 goto out_unlock;
 258         }
 259
 260         switch (state) {
 261         case TRACE_ON:
 262                 rc |= register_trace_kfree_skb(trace_kfree_skb_hit, NULL);
 263                 rc |= register_trace_napi_poll(trace_napi_poll_hit, NULL);
 264                 break;
 265         case TRACE_OFF:
 266                 rc |= unregister_trace_kfree_skb(trace_kfree_skb_hit, NULL);
 267                 rc |= unregister_trace_napi_poll(trace_napi_poll_hit, NULL);
 268
 269                 tracepoint_synchronize_unregister();
 270
 271                 /*
 272                  * Clean the device list
 273                  */
 274                 list_for_each_entry_safe(new_stat, temp, &hw_stats_list, list) {
 275                         if (new_stat->dev == NULL) {
 276                                 list_del_rcu(&new_stat->list);
 277                                 kfree_rcu(new_stat, rcu);
 278                         }
 279                 }
 280                 break;
 281         default:
 282                 rc = 1;
 283                 break;
 284         }
 285
 286         if (!rc)
 287                 trace_state = state;
 288         else
 289                 rc = -EINPROGRESS;
 290
 291 out_unlock:
 292         mutex_unlock(&trace_state_mutex);
 293
 294         return rc;
 295 }
 296
 297
 298 static int net_dm_cmd_config(struct sk_buff *skb,
 299                         struct genl_info *info)
 300 {
 301         return -ENOTSUPP;
 302 }
 303
 304 static int net_dm_cmd_trace(struct sk_buff *skb,
 305                         struct genl_info *info)
 306 {
 307         switch (info->genlhdr->cmd) {
 308         case NET_DM_CMD_START:
 309                 return set_all_monitor_traces(TRACE_ON);
 310                 break;
 311         case NET_DM_CMD_STOP:
 312                 return set_all_monitor_traces(TRACE_OFF);
 313                 break;
 314         }
 315
 316         return -ENOTSUPP;
 317 }
 318
 319 static int dropmon_net_event(struct notifier_block *ev_block,
 320                         unsigned long event, void *ptr)
 321 {
 322         struct net_device *dev = ptr;
 323         struct dm_hw_stat_delta *new_stat = NULL;
 324         struct dm_hw_stat_delta *tmp;
 325
 326         switch (event) {
 327         case NETDEV_REGISTER:
 328                 new_stat = kzalloc(sizeof(struct dm_hw_stat_delta), GFP_KERNEL);
 329
 330                 if (!new_stat)
 331                         goto out;
 332
 333                 new_stat->dev = dev;
 334                 new_stat->last_rx = jiffies;
 335                 mutex_lock(&trace_state_mutex);
 336                 list_add_rcu(&new_stat->list, &hw_stats_list);
 337                 mutex_unlock(&trace_state_mutex);
 338                 break;
 339         case NETDEV_UNREGISTER:
 340                 mutex_lock(&trace_state_mutex);
 341                 list_for_each_entry_safe(new_stat, tmp, &hw_stats_list, list) {
 342                         if (new_stat->dev == dev) {
 343                                 new_stat->dev = NULL;
 344                                 if (trace_state == TRACE_OFF) {
 345                                         list_del_rcu(&new_stat->list);
 346                                         kfree_rcu(new_stat, rcu);
 347                                         break;
 348                                 }
 349                         }
 350                 }
 351                 mutex_unlock(&trace_state_mutex);
 352                 break;
 353         }
 354 out:
 355         return NOTIFY_DONE;
 356 }
 357
 358 static struct genl_ops dropmon_ops[] = {
 359         {
 360                 .cmd = NET_DM_CMD_CONFIG,
 361                 .doit = net_dm_cmd_config,
 362         },
 363         {
 364                 .cmd = NET_DM_CMD_START,
 365                 .doit = net_dm_cmd_trace,
 366         },
 367         {
 368                 .cmd = NET_DM_CMD_STOP,
 369                 .doit = net_dm_cmd_trace,
 370         },
 371 };
 372
 373 static struct notifier_block dropmon_net_notifier = {
 374         .notifier_call = dropmon_net_event
 375 };
 376
 377 static int __init init_net_drop_monitor(void)
 378 {
 379         struct per_cpu_dm_data *data;
 380         int cpu, rc;
 381
 382         printk(KERN_INFO "Initializing network drop monitor service\n");
 383
 384         if (sizeof(void *) > 8) {
 385                 printk(KERN_ERR "Unable to store program counters on this arch, Drop monitor failed\n");
 386                 return -ENOSPC;
 387         }
 388
 389         rc = genl_register_family_with_ops(&net_drop_monitor_family,
 390                                            dropmon_ops,
 391                                            ARRAY_SIZE(dropmon_ops));
 392         if (rc) {
 393                 printk(KERN_ERR "Could not create drop monitor netlink family\n");
 394                 return rc;
 395         }
 396
 397         rc = register_netdevice_notifier(&dropmon_net_notifier);
 398         if (rc < 0) {
 399                 printk(KERN_CRIT "Failed to register netdevice notifier\n");
 400                 goto out_unreg;
 401         }
 402
 403         rc = 0;
 404
 405         for_each_present_cpu(cpu) {
 406                 data = &per_cpu(dm_cpu_data, cpu);
 407                 reset_per_cpu_data(data);
 408                 INIT_WORK(&data->dm_alert_work, send_dm_alert);
 409                 init_timer(&data->send_timer);
 410                 data->send_timer.data = cpu;
 411                 data->send_timer.function = sched_send_work;
 412         }
 413
 414         initialized = 1;
 415
 416         goto out;
 417
 418 out_unreg:
 419         genl_unregister_family(&net_drop_monitor_family);
 420 out:
 421         return rc;
 422 }
 423
 424 late_initcall(init_net_drop_monitor);