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         int cpu;
  53 };
  54
  55 struct dm_hw_stat_delta {
  56         struct net_device *dev;
  57         unsigned long last_rx;
  58         struct list_head list;
  59         struct rcu_head rcu;
  60         unsigned long last_drop_val;
  61 };
  62
  63 static struct genl_family net_drop_monitor_family = {
  64         .id             = GENL_ID_GENERATE,
  65         .hdrsize        = 0,
  66         .name           = "NET_DM",
  67         .version        = 2,
  68         .maxattr        = NET_DM_CMD_MAX,
  69 };
  70
  71 static DEFINE_PER_CPU(struct per_cpu_dm_data, dm_cpu_data);
  72
  73 static int dm_hit_limit = 64;
  74 static int dm_delay = 1;
  75 static unsigned long dm_hw_check_delta = 2*HZ;
  76 static LIST_HEAD(hw_stats_list);
  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
 100                 schedule_work_on(data->cpu, &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         WARN_ON_ONCE(data->cpu != smp_processor_id());
 125
 126         /*
 127          * Grab the skb we're about to send
 128          */
 129         skb = rcu_dereference_protected(data->skb, 1);
 130
 131         /*
 132          * Replace it with a new one
 133          */
 134         reset_per_cpu_data(data);
 135
 136         /*
 137          * Ship it!
 138          */
 139         if (skb)
 140                 genlmsg_multicast(skb, 0, NET_DM_GRP_ALERT, GFP_KERNEL);
 141
 142         put_cpu_var(dm_cpu_data);
 143 }
 144
 145 /*
 146  * This is the timer function to delay the sending of an alert
 147  * in the event that more drops will arrive during the
 148  * hysteresis period.  Note that it operates under the timer interrupt
 149  * so we don't need to disable preemption here
 150  */
 151 static void sched_send_work(unsigned long unused)
 152 {
 153         struct per_cpu_dm_data *data =  &get_cpu_var(dm_cpu_data);
 154
 155         schedule_work_on(smp_processor_id(), &data->dm_alert_work);
 156
 157         put_cpu_var(dm_cpu_data);
 158 }
 159
 160 static void trace_drop_common(struct sk_buff *skb, void *location)
 161 {
 162         struct net_dm_alert_msg *msg;
 163         struct nlmsghdr *nlh;
 164         struct nlattr *nla;
 165         int i;
 166         struct sk_buff *dskb;
 167         struct per_cpu_dm_data *data = &get_cpu_var(dm_cpu_data);
 168
 169
 170         rcu_read_lock();
 171         dskb = rcu_dereference(data->skb);
 172
 173         if (!dskb)
 174                 goto out;
 175
 176         if (!atomic_add_unless(&data->dm_hit_count, -1, 0)) {
 177                 /*
 178                  * we're already at zero, discard this hit
 179                  */
 180                 goto out;
 181         }
 182
 183         nlh = (struct nlmsghdr *)dskb->data;
 184         nla = genlmsg_data(nlmsg_data(nlh));
 185         msg = nla_data(nla);
 186         for (i = 0; i < msg->entries; i++) {
 187                 if (!memcmp(&location, msg->points[i].pc, sizeof(void *))) {
 188                         msg->points[i].count++;
 189                         atomic_inc(&data->dm_hit_count);
 190                         goto out;
 191                 }
 192         }
 193
 194         /*
 195          * We need to create a new entry
 196          */
 197         __nla_reserve_nohdr(dskb, sizeof(struct net_dm_drop_point));
 198         nla->nla_len += NLA_ALIGN(sizeof(struct net_dm_drop_point));
 199         memcpy(msg->points[msg->entries].pc, &location, sizeof(void *));
 200         msg->points[msg->entries].count = 1;
 201         msg->entries++;
 202
 203         if (!timer_pending(&data->send_timer)) {
 204                 data->send_timer.expires = jiffies + dm_delay * HZ;
 205                 add_timer_on(&data->send_timer, smp_processor_id());
 206         }
 207
 208 out:
 209         rcu_read_unlock();
 210         put_cpu_var(dm_cpu_data);
 211         return;
 212 }
 213
 214 static void trace_kfree_skb_hit(void *ignore, struct sk_buff *skb, void *location)
 215 {
 216         trace_drop_common(skb, location);
 217 }
 218
 219 static void trace_napi_poll_hit(void *ignore, struct napi_struct *napi)
 220 {
 221         struct dm_hw_stat_delta *new_stat;
 222
 223         /*
 224          * Don't check napi structures with no associated device
 225          */
 226         if (!napi->dev)
 227                 return;
 228
 229         rcu_read_lock();
 230         list_for_each_entry_rcu(new_stat, &hw_stats_list, list) {
 231                 /*
 232                  * only add a note to our monitor buffer if:
 233                  * 1) this is the dev we received on
 234                  * 2) its after the last_rx delta
 235                  * 3) our rx_dropped count has gone up
 236                  */
 237                 if ((new_stat->dev == napi->dev)  &&
 238                     (time_after(jiffies, new_stat->last_rx + dm_hw_check_delta)) &&
 239                     (napi->dev->stats.rx_dropped != new_stat->last_drop_val)) {
 240                         trace_drop_common(NULL, NULL);
 241                         new_stat->last_drop_val = napi->dev->stats.rx_dropped;
 242                         new_stat->last_rx = jiffies;
 243                         break;
 244                 }
 245         }
 246         rcu_read_unlock();
 247 }
 248
 249 static int set_all_monitor_traces(int state)
 250 {
 251         int rc = 0;
 252         struct dm_hw_stat_delta *new_stat = NULL;
 253         struct dm_hw_stat_delta *temp;
 254
 255         mutex_lock(&trace_state_mutex);
 256
 257         if (state == trace_state) {
 258                 rc = -EAGAIN;
 259                 goto out_unlock;
 260         }
 261
 262         switch (state) {
 263         case TRACE_ON:
 264                 rc |= register_trace_kfree_skb(trace_kfree_skb_hit, NULL);
 265                 rc |= register_trace_napi_poll(trace_napi_poll_hit, NULL);
 266                 break;
 267         case TRACE_OFF:
 268                 rc |= unregister_trace_kfree_skb(trace_kfree_skb_hit, NULL);
 269                 rc |= unregister_trace_napi_poll(trace_napi_poll_hit, NULL);
 270
 271                 tracepoint_synchronize_unregister();
 272
 273                 /*
 274                  * Clean the device list
 275                  */
 276                 list_for_each_entry_safe(new_stat, temp, &hw_stats_list, list) {
 277                         if (new_stat->dev == NULL) {
 278                                 list_del_rcu(&new_stat->list);
 279                                 kfree_rcu(new_stat, rcu);
 280                         }
 281                 }
 282                 break;
 283         default:
 284                 rc = 1;
 285                 break;
 286         }
 287
 288         if (!rc)
 289                 trace_state = state;
 290         else
 291                 rc = -EINPROGRESS;
 292
 293 out_unlock:
 294         mutex_unlock(&trace_state_mutex);
 295
 296         return rc;
 297 }
 298
 299
 300 static int net_dm_cmd_config(struct sk_buff *skb,
 301                         struct genl_info *info)
 302 {
 303         return -ENOTSUPP;
 304 }
 305
 306 static int net_dm_cmd_trace(struct sk_buff *skb,
 307                         struct genl_info *info)
 308 {
 309         switch (info->genlhdr->cmd) {
 310         case NET_DM_CMD_START:
 311                 return set_all_monitor_traces(TRACE_ON);
 312                 break;
 313         case NET_DM_CMD_STOP:
 314                 return set_all_monitor_traces(TRACE_OFF);
 315                 break;
 316         }
 317
 318         return -ENOTSUPP;
 319 }
 320
 321 static int dropmon_net_event(struct notifier_block *ev_block,
 322                         unsigned long event, void *ptr)
 323 {
 324         struct net_device *dev = ptr;
 325         struct dm_hw_stat_delta *new_stat = NULL;
 326         struct dm_hw_stat_delta *tmp;
 327
 328         switch (event) {
 329         case NETDEV_REGISTER:
 330                 new_stat = kzalloc(sizeof(struct dm_hw_stat_delta), GFP_KERNEL);
 331
 332                 if (!new_stat)
 333                         goto out;
 334
 335                 new_stat->dev = dev;
 336                 new_stat->last_rx = jiffies;
 337                 mutex_lock(&trace_state_mutex);
 338                 list_add_rcu(&new_stat->list, &hw_stats_list);
 339                 mutex_unlock(&trace_state_mutex);
 340                 break;
 341         case NETDEV_UNREGISTER:
 342                 mutex_lock(&trace_state_mutex);
 343                 list_for_each_entry_safe(new_stat, tmp, &hw_stats_list, list) {
 344                         if (new_stat->dev == dev) {
 345                                 new_stat->dev = NULL;
 346                                 if (trace_state == TRACE_OFF) {
 347                                         list_del_rcu(&new_stat->list);
 348                                         kfree_rcu(new_stat, rcu);
 349                                         break;
 350                                 }
 351                         }
 352                 }
 353                 mutex_unlock(&trace_state_mutex);
 354                 break;
 355         }
 356 out:
 357         return NOTIFY_DONE;
 358 }
 359
 360 static struct genl_ops dropmon_ops[] = {
 361         {
 362                 .cmd = NET_DM_CMD_CONFIG,
 363                 .doit = net_dm_cmd_config,
 364         },
 365         {
 366                 .cmd = NET_DM_CMD_START,
 367                 .doit = net_dm_cmd_trace,
 368         },
 369         {
 370                 .cmd = NET_DM_CMD_STOP,
 371                 .doit = net_dm_cmd_trace,
 372         },
 373 };
 374
 375 static struct notifier_block dropmon_net_notifier = {
 376         .notifier_call = dropmon_net_event
 377 };
 378
 379 static int __init init_net_drop_monitor(void)
 380 {
 381         struct per_cpu_dm_data *data;
 382         int cpu, rc;
 383
 384         printk(KERN_INFO "Initializing network drop monitor service\n");
 385
 386         if (sizeof(void *) > 8) {
 387                 printk(KERN_ERR "Unable to store program counters on this arch, Drop monitor failed\n");
 388                 return -ENOSPC;
 389         }
 390
 391         rc = genl_register_family_with_ops(&net_drop_monitor_family,
 392                                            dropmon_ops,
 393                                            ARRAY_SIZE(dropmon_ops));
 394         if (rc) {
 395                 printk(KERN_ERR "Could not create drop monitor netlink family\n");
 396                 return rc;
 397         }
 398
 399         rc = register_netdevice_notifier(&dropmon_net_notifier);
 400         if (rc < 0) {
 401                 printk(KERN_CRIT "Failed to register netdevice notifier\n");
 402                 goto out_unreg;
 403         }
 404
 405         rc = 0;
 406
 407         for_each_present_cpu(cpu) {
 408                 data = &per_cpu(dm_cpu_data, cpu);
 409                 data->cpu = cpu;
 410                 INIT_WORK(&data->dm_alert_work, send_dm_alert);
 411                 init_timer(&data->send_timer);
 412                 data->send_timer.data = cpu;
 413                 data->send_timer.function = sched_send_work;
 414                 reset_per_cpu_data(data);
 415         }
 416
 417
 418         goto out;
 419
 420 out_unreg:
 421         genl_unregister_family(&net_drop_monitor_family);
 422 out:
 423         return rc;
 424 }
 425
 426 late_initcall(init_net_drop_monitor);