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 /*
  37  * Globals, our netlink socket pointer
  38  * and the work handle that will send up
  39  * netlink alerts
  40  */
  41 static int trace_state = TRACE_OFF;
  42 static DEFINE_MUTEX(trace_state_mutex);
  43
  44 struct per_cpu_dm_data {
  45         spinlock_t              lock;
  46         struct sk_buff          *skb;
  47         struct work_struct      dm_alert_work;
  48         struct timer_list       send_timer;
  49 };
  50
  51 struct dm_hw_stat_delta {
  52         struct net_device *dev;
  53         unsigned long last_rx;
  54         struct list_head list;
  55         struct rcu_head rcu;
  56         unsigned long last_drop_val;
  57 };
  58
  59 static struct genl_family net_drop_monitor_family = {
  60         .id             = GENL_ID_GENERATE,
  61         .hdrsize        = 0,
  62         .name           = "NET_DM",
  63         .version        = 2,
  64 };
  65
  66 static DEFINE_PER_CPU(struct per_cpu_dm_data, dm_cpu_data);
  67
  68 static int dm_hit_limit = 64;
  69 static int dm_delay = 1;
  70 static unsigned long dm_hw_check_delta = 2*HZ;
  71 static LIST_HEAD(hw_stats_list);
  72
  73 static struct sk_buff *reset_per_cpu_data(struct per_cpu_dm_data *data)
  74 {
  75         size_t al;
  76         struct net_dm_alert_msg *msg;
  77         struct nlattr *nla;
  78         struct sk_buff *skb;
  79         unsigned long flags;
  80         void *msg_header;
  81
  82         al = sizeof(struct net_dm_alert_msg);
  83         al += dm_hit_limit * sizeof(struct net_dm_drop_point);
  84         al += sizeof(struct nlattr);
  85
  86         skb = genlmsg_new(al, GFP_KERNEL);
  87
  88         if (!skb)
  89                 goto err;
  90
  91         msg_header = genlmsg_put(skb, 0, 0, &net_drop_monitor_family,
  92                                  0, NET_DM_CMD_ALERT);
  93         if (!msg_header) {
  94                 nlmsg_free(skb);
  95                 skb = NULL;
  96                 goto err;
  97         }
  98         nla = nla_reserve(skb, NLA_UNSPEC,
  99                           sizeof(struct net_dm_alert_msg));
 100         if (!nla) {
 101                 nlmsg_free(skb);
 102                 skb = NULL;
 103                 goto err;
 104         }
 105         msg = nla_data(nla);
 106         memset(msg, 0, al);
 107         goto out;
 108
 109 err:
 110         mod_timer(&data->send_timer, jiffies + HZ / 10);
 111 out:
 112         spin_lock_irqsave(&data->lock, flags);
 113         swap(data->skb, skb);
 114         spin_unlock_irqrestore(&data->lock, flags);
 115
 116         if (skb) {
 117                 struct nlmsghdr *nlh = (struct nlmsghdr *)skb->data;
 118                 struct genlmsghdr *gnlh = (struct genlmsghdr *)nlmsg_data(nlh);
 119
 120                 genlmsg_end(skb, genlmsg_data(gnlh));
 121         }
 122
 123         return skb;
 124 }
 125
 126 static void send_dm_alert(struct work_struct *work)
 127 {
 128         struct sk_buff *skb;
 129         struct per_cpu_dm_data *data;
 130
 131         data = container_of(work, struct per_cpu_dm_data, dm_alert_work);
 132
 133         skb = reset_per_cpu_data(data);
 134
 135         if (skb)
 136                 genlmsg_multicast(skb, 0, NET_DM_GRP_ALERT, GFP_KERNEL);
 137 }
 138
 139 /*
 140  * This is the timer function to delay the sending of an alert
 141  * in the event that more drops will arrive during the
 142  * hysteresis period.
 143  */
 144 static void sched_send_work(unsigned long _data)
 145 {
 146         struct per_cpu_dm_data *data = (struct per_cpu_dm_data *)_data;
 147
 148         schedule_work(&data->dm_alert_work);
 149 }
 150
 151 static void trace_drop_common(struct sk_buff *skb, void *location)
 152 {
 153         struct net_dm_alert_msg *msg;
 154         struct nlmsghdr *nlh;
 155         struct nlattr *nla;
 156         int i;
 157         struct sk_buff *dskb;
 158         struct per_cpu_dm_data *data;
 159         unsigned long flags;
 160
 161         local_irq_save(flags);
 162         data = &__get_cpu_var(dm_cpu_data);
 163         spin_lock(&data->lock);
 164         dskb = data->skb;
 165
 166         if (!dskb)
 167                 goto out;
 168
 169         nlh = (struct nlmsghdr *)dskb->data;
 170         nla = genlmsg_data(nlmsg_data(nlh));
 171         msg = nla_data(nla);
 172         for (i = 0; i < msg->entries; i++) {
 173                 if (!memcmp(&location, msg->points[i].pc, sizeof(void *))) {
 174                         msg->points[i].count++;
 175                         goto out;
 176                 }
 177         }
 178         if (msg->entries == dm_hit_limit)
 179                 goto out;
 180         /*
 181          * We need to create a new entry
 182          */
 183         __nla_reserve_nohdr(dskb, sizeof(struct net_dm_drop_point));
 184         nla->nla_len += NLA_ALIGN(sizeof(struct net_dm_drop_point));
 185         memcpy(msg->points[msg->entries].pc, &location, sizeof(void *));
 186         msg->points[msg->entries].count = 1;
 187         msg->entries++;
 188
 189         if (!timer_pending(&data->send_timer)) {
 190                 data->send_timer.expires = jiffies + dm_delay * HZ;
 191                 add_timer(&data->send_timer);
 192         }
 193
 194 out:
 195         spin_unlock_irqrestore(&data->lock, flags);
 196 }
 197
 198 static void trace_kfree_skb_hit(void *ignore, struct sk_buff *skb, void *location)
 199 {
 200         trace_drop_common(skb, location);
 201 }
 202
 203 static void trace_napi_poll_hit(void *ignore, struct napi_struct *napi)
 204 {
 205         struct dm_hw_stat_delta *new_stat;
 206
 207         /*
 208          * Don't check napi structures with no associated device
 209          */
 210         if (!napi->dev)
 211                 return;
 212
 213         rcu_read_lock();
 214         list_for_each_entry_rcu(new_stat, &hw_stats_list, list) {
 215                 /*
 216                  * only add a note to our monitor buffer if:
 217                  * 1) this is the dev we received on
 218                  * 2) its after the last_rx delta
 219                  * 3) our rx_dropped count has gone up
 220                  */
 221                 if ((new_stat->dev == napi->dev)  &&
 222                     (time_after(jiffies, new_stat->last_rx + dm_hw_check_delta)) &&
 223                     (napi->dev->stats.rx_dropped != new_stat->last_drop_val)) {
 224                         trace_drop_common(NULL, NULL);
 225                         new_stat->last_drop_val = napi->dev->stats.rx_dropped;
 226                         new_stat->last_rx = jiffies;
 227                         break;
 228                 }
 229         }
 230         rcu_read_unlock();
 231 }
 232
 233 static int set_all_monitor_traces(int state)
 234 {
 235         int rc = 0;
 236         struct dm_hw_stat_delta *new_stat = NULL;
 237         struct dm_hw_stat_delta *temp;
 238
 239         mutex_lock(&trace_state_mutex);
 240
 241         if (state == trace_state) {
 242                 rc = -EAGAIN;
 243                 goto out_unlock;
 244         }
 245
 246         switch (state) {
 247         case TRACE_ON:
 248                 rc |= register_trace_kfree_skb(trace_kfree_skb_hit, NULL);
 249                 rc |= register_trace_napi_poll(trace_napi_poll_hit, NULL);
 250                 break;
 251         case TRACE_OFF:
 252                 rc |= unregister_trace_kfree_skb(trace_kfree_skb_hit, NULL);
 253                 rc |= unregister_trace_napi_poll(trace_napi_poll_hit, NULL);
 254
 255                 tracepoint_synchronize_unregister();
 256
 257                 /*
 258                  * Clean the device list
 259                  */
 260                 list_for_each_entry_safe(new_stat, temp, &hw_stats_list, list) {
 261                         if (new_stat->dev == NULL) {
 262                                 list_del_rcu(&new_stat->list);
 263                                 kfree_rcu(new_stat, rcu);
 264                         }
 265                 }
 266                 break;
 267         default:
 268                 rc = 1;
 269                 break;
 270         }
 271
 272         if (!rc)
 273                 trace_state = state;
 274         else
 275                 rc = -EINPROGRESS;
 276
 277 out_unlock:
 278         mutex_unlock(&trace_state_mutex);
 279
 280         return rc;
 281 }
 282
 283
 284 static int net_dm_cmd_config(struct sk_buff *skb,
 285                         struct genl_info *info)
 286 {
 287         return -ENOTSUPP;
 288 }
 289
 290 static int net_dm_cmd_trace(struct sk_buff *skb,
 291                         struct genl_info *info)
 292 {
 293         switch (info->genlhdr->cmd) {
 294         case NET_DM_CMD_START:
 295                 return set_all_monitor_traces(TRACE_ON);
 296                 break;
 297         case NET_DM_CMD_STOP:
 298                 return set_all_monitor_traces(TRACE_OFF);
 299                 break;
 300         }
 301
 302         return -ENOTSUPP;
 303 }
 304
 305 static int dropmon_net_event(struct notifier_block *ev_block,
 306                         unsigned long event, void *ptr)
 307 {
 308         struct net_device *dev = ptr;
 309         struct dm_hw_stat_delta *new_stat = NULL;
 310         struct dm_hw_stat_delta *tmp;
 311
 312         switch (event) {
 313         case NETDEV_REGISTER:
 314                 new_stat = kzalloc(sizeof(struct dm_hw_stat_delta), GFP_KERNEL);
 315
 316                 if (!new_stat)
 317                         goto out;
 318
 319                 new_stat->dev = dev;
 320                 new_stat->last_rx = jiffies;
 321                 mutex_lock(&trace_state_mutex);
 322                 list_add_rcu(&new_stat->list, &hw_stats_list);
 323                 mutex_unlock(&trace_state_mutex);
 324                 break;
 325         case NETDEV_UNREGISTER:
 326                 mutex_lock(&trace_state_mutex);
 327                 list_for_each_entry_safe(new_stat, tmp, &hw_stats_list, list) {
 328                         if (new_stat->dev == dev) {
 329                                 new_stat->dev = NULL;
 330                                 if (trace_state == TRACE_OFF) {
 331                                         list_del_rcu(&new_stat->list);
 332                                         kfree_rcu(new_stat, rcu);
 333                                         break;
 334                                 }
 335                         }
 336                 }
 337                 mutex_unlock(&trace_state_mutex);
 338                 break;
 339         }
 340 out:
 341         return NOTIFY_DONE;
 342 }
 343
 344 static struct genl_ops dropmon_ops[] = {
 345         {
 346                 .cmd = NET_DM_CMD_CONFIG,
 347                 .doit = net_dm_cmd_config,
 348         },
 349         {
 350                 .cmd = NET_DM_CMD_START,
 351                 .doit = net_dm_cmd_trace,
 352         },
 353         {
 354                 .cmd = NET_DM_CMD_STOP,
 355                 .doit = net_dm_cmd_trace,
 356         },
 357 };
 358
 359 static struct notifier_block dropmon_net_notifier = {
 360         .notifier_call = dropmon_net_event
 361 };
 362
 363 static int __init init_net_drop_monitor(void)
 364 {
 365         struct per_cpu_dm_data *data;
 366         int cpu, rc;
 367
 368         printk(KERN_INFO "Initializing network drop monitor service\n");
 369
 370         if (sizeof(void *) > 8) {
 371                 printk(KERN_ERR "Unable to store program counters on this arch, Drop monitor failed\n");
 372                 return -ENOSPC;
 373         }
 374
 375         rc = genl_register_family_with_ops(&net_drop_monitor_family,
 376                                            dropmon_ops,
 377                                            ARRAY_SIZE(dropmon_ops));
 378         if (rc) {
 379                 printk(KERN_ERR "Could not create drop monitor netlink family\n");
 380                 return rc;
 381         }
 382
 383         rc = register_netdevice_notifier(&dropmon_net_notifier);
 384         if (rc < 0) {
 385                 printk(KERN_CRIT "Failed to register netdevice notifier\n");
 386                 goto out_unreg;
 387         }
 388
 389         rc = 0;
 390
 391         for_each_present_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 late_initcall(init_net_drop_monitor);