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 <net/genetlink.h>
  25 #include <net/netevent.h>
  26
  27 #include <trace/events/skb.h>
  28 #include <trace/events/napi.h>
  29
  30 #include <asm/unaligned.h>
  31
  32 #define TRACE_ON 1
  33 #define TRACE_OFF 0
  34
  35 static void send_dm_alert(struct work_struct *unused);
  36
  37
  38 /*
  39  * Globals, our netlink socket pointer
  40  * and the work handle that will send up
  41  * netlink alerts
  42  */
  43 static int trace_state = TRACE_OFF;
  44 static spinlock_t trace_state_lock = SPIN_LOCK_UNLOCKED;
  45
  46 struct per_cpu_dm_data {
  47         struct work_struct dm_alert_work;
  48         struct sk_buff *skb;
  49         atomic_t dm_hit_count;
  50         struct timer_list send_timer;
  51 };
  52
  53 struct dm_hw_stat_delta {
  54         struct net_device *dev;
  55         struct list_head list;
  56         struct rcu_head rcu;
  57         unsigned long last_drop_val;
  58 };
  59
  60 static struct genl_family net_drop_monitor_family = {
  61         .id             = GENL_ID_GENERATE,
  62         .hdrsize        = 0,
  63         .name           = "NET_DM",
  64         .version        = 2,
  65         .maxattr        = NET_DM_CMD_MAX,
  66 };
  67
  68 static DEFINE_PER_CPU(struct per_cpu_dm_data, dm_cpu_data);
  69
  70 static int dm_hit_limit = 64;
  71 static int dm_delay = 1;
  72 static unsigned long dm_hw_check_delta = 2*HZ;
  73 static LIST_HEAD(hw_stats_list);
  74
  75 static void reset_per_cpu_data(struct per_cpu_dm_data *data)
  76 {
  77         size_t al;
  78         struct net_dm_alert_msg *msg;
  79         struct nlattr *nla;
  80
  81         al = sizeof(struct net_dm_alert_msg);
  82         al += dm_hit_limit * sizeof(struct net_dm_drop_point);
  83         al += sizeof(struct nlattr);
  84
  85         data->skb = genlmsg_new(al, GFP_KERNEL);
  86         genlmsg_put(data->skb, 0, 0, &net_drop_monitor_family,
  87                         0, NET_DM_CMD_ALERT);
  88         nla = nla_reserve(data->skb, NLA_UNSPEC, sizeof(struct net_dm_alert_msg));
  89         msg = nla_data(nla);
  90         memset(msg, 0, al);
  91         atomic_set(&data->dm_hit_count, dm_hit_limit);
  92 }
  93
  94 static void send_dm_alert(struct work_struct *unused)
  95 {
  96         struct sk_buff *skb;
  97         struct per_cpu_dm_data *data = &__get_cpu_var(dm_cpu_data);
  98
  99         /*
 100          * Grab the skb we're about to send
 101          */
 102         skb = data->skb;
 103
 104         /*
 105          * Replace it with a new one
 106          */
 107         reset_per_cpu_data(data);
 108
 109         /*
 110          * Ship it!
 111          */
 112         genlmsg_multicast(skb, 0, NET_DM_GRP_ALERT, GFP_KERNEL);
 113
 114 }
 115
 116 /*
 117  * This is the timer function to delay the sending of an alert
 118  * in the event that more drops will arrive during the
 119  * hysteresis period.  Note that it operates under the timer interrupt
 120  * so we don't need to disable preemption here
 121  */
 122 static void sched_send_work(unsigned long unused)
 123 {
 124         struct per_cpu_dm_data *data =  &__get_cpu_var(dm_cpu_data);
 125
 126         schedule_work(&data->dm_alert_work);
 127 }
 128
 129 static void trace_drop_common(struct sk_buff *skb, void *location)
 130 {
 131         struct net_dm_alert_msg *msg;
 132         struct nlmsghdr *nlh;
 133         struct nlattr *nla;
 134         int i;
 135         struct per_cpu_dm_data *data = &__get_cpu_var(dm_cpu_data);
 136
 137
 138         if (!atomic_add_unless(&data->dm_hit_count, -1, 0)) {
 139                 /*
 140                  * we're already at zero, discard this hit
 141                  */
 142                 goto out;
 143         }
 144
 145         nlh = (struct nlmsghdr *)data->skb->data;
 146         nla = genlmsg_data(nlmsg_data(nlh));
 147         msg = nla_data(nla);
 148         for (i = 0; i < msg->entries; i++) {
 149                 if (!memcmp(&location, msg->points[i].pc, sizeof(void *))) {
 150                         msg->points[i].count++;
 151                         goto out;
 152                 }
 153         }
 154
 155         /*
 156          * We need to create a new entry
 157          */
 158         __nla_reserve_nohdr(data->skb, sizeof(struct net_dm_drop_point));
 159         nla->nla_len += NLA_ALIGN(sizeof(struct net_dm_drop_point));
 160         memcpy(msg->points[msg->entries].pc, &location, sizeof(void *));
 161         msg->points[msg->entries].count = 1;
 162         msg->entries++;
 163
 164         if (!timer_pending(&data->send_timer)) {
 165                 data->send_timer.expires = jiffies + dm_delay * HZ;
 166                 add_timer_on(&data->send_timer, smp_processor_id());
 167         }
 168
 169 out:
 170         return;
 171 }
 172
 173 static void trace_kfree_skb_hit(struct sk_buff *skb, void *location)
 174 {
 175         trace_drop_common(skb, location);
 176 }
 177
 178 static void trace_napi_poll_hit(struct napi_struct *napi)
 179 {
 180         struct dm_hw_stat_delta *new_stat;
 181
 182         /*
 183          * Ratelimit our check time to dm_hw_check_delta jiffies
 184          */
 185         if (!napi->dev ||
 186             !time_after(jiffies, napi->dev->last_rx + dm_hw_check_delta))
 187                 return;
 188
 189         rcu_read_lock();
 190         list_for_each_entry_rcu(new_stat, &hw_stats_list, list) {
 191                 if ((new_stat->dev == napi->dev)  &&
 192                     (napi->dev->stats.rx_dropped != new_stat->last_drop_val)) {
 193                         trace_drop_common(NULL, NULL);
 194                         new_stat->last_drop_val = napi->dev->stats.rx_dropped;
 195                         break;
 196                 }
 197         }
 198         rcu_read_unlock();
 199 }
 200
 201
 202 static void free_dm_hw_stat(struct rcu_head *head)
 203 {
 204         struct dm_hw_stat_delta *n;
 205         n = container_of(head, struct dm_hw_stat_delta, rcu);
 206         kfree(n);
 207 }
 208
 209 static int set_all_monitor_traces(int state)
 210 {
 211         int rc = 0;
 212         struct dm_hw_stat_delta *new_stat = NULL;
 213         struct dm_hw_stat_delta *temp;
 214
 215         spin_lock(&trace_state_lock);
 216
 217         switch (state) {
 218         case TRACE_ON:
 219                 rc |= register_trace_kfree_skb(trace_kfree_skb_hit);
 220                 rc |= register_trace_napi_poll(trace_napi_poll_hit);
 221                 break;
 222         case TRACE_OFF:
 223                 rc |= unregister_trace_kfree_skb(trace_kfree_skb_hit);
 224                 rc |= unregister_trace_napi_poll(trace_napi_poll_hit);
 225
 226                 tracepoint_synchronize_unregister();
 227
 228                 /*
 229                  * Clean the device list
 230                  */
 231                 list_for_each_entry_safe(new_stat, temp, &hw_stats_list, list) {
 232                         if (new_stat->dev == NULL) {
 233                                 list_del_rcu(&new_stat->list);
 234                                 call_rcu(&new_stat->rcu, free_dm_hw_stat);
 235                         }
 236                 }
 237                 break;
 238         default:
 239                 rc = 1;
 240                 break;
 241         }
 242
 243         if (!rc)
 244                 trace_state = state;
 245
 246         spin_unlock(&trace_state_lock);
 247
 248         if (rc)
 249                 return -EINPROGRESS;
 250         return rc;
 251 }
 252
 253
 254 static int net_dm_cmd_config(struct sk_buff *skb,
 255                         struct genl_info *info)
 256 {
 257         return -ENOTSUPP;
 258 }
 259
 260 static int net_dm_cmd_trace(struct sk_buff *skb,
 261                         struct genl_info *info)
 262 {
 263         switch (info->genlhdr->cmd) {
 264         case NET_DM_CMD_START:
 265                 return set_all_monitor_traces(TRACE_ON);
 266                 break;
 267         case NET_DM_CMD_STOP:
 268                 return set_all_monitor_traces(TRACE_OFF);
 269                 break;
 270         }
 271
 272         return -ENOTSUPP;
 273 }
 274
 275 static int dropmon_net_event(struct notifier_block *ev_block,
 276                         unsigned long event, void *ptr)
 277 {
 278         struct net_device *dev = ptr;
 279         struct dm_hw_stat_delta *new_stat = NULL;
 280         struct dm_hw_stat_delta *tmp;
 281
 282         switch (event) {
 283         case NETDEV_REGISTER:
 284                 new_stat = kzalloc(sizeof(struct dm_hw_stat_delta), GFP_KERNEL);
 285
 286                 if (!new_stat)
 287                         goto out;
 288
 289                 new_stat->dev = dev;
 290                 INIT_RCU_HEAD(&new_stat->rcu);
 291                 spin_lock(&trace_state_lock);
 292                 list_add_rcu(&new_stat->list, &hw_stats_list);
 293                 spin_unlock(&trace_state_lock);
 294                 break;
 295         case NETDEV_UNREGISTER:
 296                 spin_lock(&trace_state_lock);
 297                 list_for_each_entry_safe(new_stat, tmp, &hw_stats_list, list) {
 298                         if (new_stat->dev == dev) {
 299                                 new_stat->dev = NULL;
 300                                 if (trace_state == TRACE_OFF) {
 301                                         list_del_rcu(&new_stat->list);
 302                                         call_rcu(&new_stat->rcu, free_dm_hw_stat);
 303                                         break;
 304                                 }
 305                         }
 306                 }
 307                 spin_unlock(&trace_state_lock);
 308                 break;
 309         }
 310 out:
 311         return NOTIFY_DONE;
 312 }
 313
 314 static struct genl_ops dropmon_ops[] = {
 315         {
 316                 .cmd = NET_DM_CMD_CONFIG,
 317                 .doit = net_dm_cmd_config,
 318         },
 319         {
 320                 .cmd = NET_DM_CMD_START,
 321                 .doit = net_dm_cmd_trace,
 322         },
 323         {
 324                 .cmd = NET_DM_CMD_STOP,
 325                 .doit = net_dm_cmd_trace,
 326         },
 327 };
 328
 329 static struct notifier_block dropmon_net_notifier = {
 330         .notifier_call = dropmon_net_event
 331 };
 332
 333 static int __init init_net_drop_monitor(void)
 334 {
 335         int cpu;
 336         int rc, i, ret;
 337         struct per_cpu_dm_data *data;
 338         printk(KERN_INFO "Initalizing network drop monitor service\n");
 339
 340         if (sizeof(void *) > 8) {
 341                 printk(KERN_ERR "Unable to store program counters on this arch, Drop monitor failed\n");
 342                 return -ENOSPC;
 343         }
 344
 345         if (genl_register_family(&net_drop_monitor_family) < 0) {
 346                 printk(KERN_ERR "Could not create drop monitor netlink family\n");
 347                 return -EFAULT;
 348         }
 349
 350         rc = -EFAULT;
 351
 352         for (i = 0; i < ARRAY_SIZE(dropmon_ops); i++) {
 353                 ret = genl_register_ops(&net_drop_monitor_family,
 354                                         &dropmon_ops[i]);
 355                 if (ret) {
 356                         printk(KERN_CRIT "Failed to register operation %d\n",
 357                                 dropmon_ops[i].cmd);
 358                         goto out_unreg;
 359                 }
 360         }
 361
 362         rc = register_netdevice_notifier(&dropmon_net_notifier);
 363         if (rc < 0) {
 364                 printk(KERN_CRIT "Failed to register netdevice notifier\n");
 365                 goto out_unreg;
 366         }
 367
 368         rc = 0;
 369
 370         for_each_present_cpu(cpu) {
 371                 data = &per_cpu(dm_cpu_data, cpu);
 372                 reset_per_cpu_data(data);
 373                 INIT_WORK(&data->dm_alert_work, send_dm_alert);
 374                 init_timer(&data->send_timer);
 375                 data->send_timer.data = cpu;
 376                 data->send_timer.function = sched_send_work;
 377         }
 378
 379         goto out;
 380
 381 out_unreg:
 382         genl_unregister_family(&net_drop_monitor_family);
 383 out:
 384         return rc;
 385 }
 386
 387 late_initcall(init_net_drop_monitor);