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