net/core/drop_monitor.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Monitoring code for network dropped packet alerts
   4  *
   5  * Copyright (C) 2009 Neil Horman <nhorman@tuxdriver.com>
   6  */
   7
   8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  10 #include <linux/netdevice.h>
  11 #include <linux/etherdevice.h>
  12 #include <linux/string.h>
  13 #include <linux/if_arp.h>
  14 #include <linux/inetdevice.h>
  15 #include <linux/inet.h>
  16 #include <linux/interrupt.h>
  17 #include <linux/netpoll.h>
  18 #include <linux/sched.h>
  19 #include <linux/delay.h>
  20 #include <linux/types.h>
  21 #include <linux/workqueue.h>
  22 #include <linux/netlink.h>
  23 #include <linux/net_dropmon.h>
  24 #include <linux/percpu.h>
  25 #include <linux/timer.h>
  26 #include <linux/bitops.h>
  27 #include <linux/slab.h>
  28 #include <linux/module.h>
  29 #include <net/genetlink.h>
  30 #include <net/netevent.h>
  31
  32 #include <trace/events/skb.h>
  33 #include <trace/events/napi.h>
  34
  35 #include <asm/unaligned.h>
  36
  37 #define TRACE_ON 1
  38 #define TRACE_OFF 0
  39
  40 /*
  41  * Globals, our netlink socket pointer
  42  * and the work handle that will send up
  43  * netlink alerts
  44  */
  45 static int trace_state = TRACE_OFF;
  46
  47 /* net_dm_mutex
  48  *
  49  * An overall lock guarding every operation coming from userspace.
  50  * It also guards the global 'hw_stats_list' list.
  51  */
  52 static DEFINE_MUTEX(net_dm_mutex);
  53
  54 struct per_cpu_dm_data {
  55         spinlock_t              lock;   /* Protects 'skb' and 'send_timer' */
  56         struct sk_buff          *skb;
  57         struct work_struct      dm_alert_work;
  58         struct timer_list       send_timer;
  59 };
  60
  61 struct dm_hw_stat_delta {
  62         struct net_device *dev;
  63         unsigned long last_rx;
  64         struct list_head list;
  65         struct rcu_head rcu;
  66         unsigned long last_drop_val;
  67 };
  68
  69 static struct genl_family net_drop_monitor_family;
  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 struct sk_buff *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         unsigned long flags;
  85         void *msg_header;
  86
  87         al = sizeof(struct net_dm_alert_msg);
  88         al += dm_hit_limit * sizeof(struct net_dm_drop_point);
  89         al += sizeof(struct nlattr);
  90
  91         skb = genlmsg_new(al, GFP_KERNEL);
  92
  93         if (!skb)
  94                 goto err;
  95
  96         msg_header = genlmsg_put(skb, 0, 0, &net_drop_monitor_family,
  97                                  0, NET_DM_CMD_ALERT);
  98         if (!msg_header) {
  99                 nlmsg_free(skb);
 100                 skb = NULL;
 101                 goto err;
 102         }
 103         nla = nla_reserve(skb, NLA_UNSPEC,
 104                           sizeof(struct net_dm_alert_msg));
 105         if (!nla) {
 106                 nlmsg_free(skb);
 107                 skb = NULL;
 108                 goto err;
 109         }
 110         msg = nla_data(nla);
 111         memset(msg, 0, al);
 112         goto out;
 113
 114 err:
 115         mod_timer(&data->send_timer, jiffies + HZ / 10);
 116 out:
 117         spin_lock_irqsave(&data->lock, flags);
 118         swap(data->skb, skb);
 119         spin_unlock_irqrestore(&data->lock, flags);
 120
 121         if (skb) {
 122                 struct nlmsghdr *nlh = (struct nlmsghdr *)skb->data;
 123                 struct genlmsghdr *gnlh = (struct genlmsghdr *)nlmsg_data(nlh);
 124
 125                 genlmsg_end(skb, genlmsg_data(gnlh));
 126         }
 127
 128         return skb;
 129 }
 130
 131 static const struct genl_multicast_group dropmon_mcgrps[] = {
 132         { .name = "events", },
 133 };
 134
 135 static void send_dm_alert(struct work_struct *work)
 136 {
 137         struct sk_buff *skb;
 138         struct per_cpu_dm_data *data;
 139
 140         data = container_of(work, struct per_cpu_dm_data, dm_alert_work);
 141
 142         skb = reset_per_cpu_data(data);
 143
 144         if (skb)
 145                 genlmsg_multicast(&net_drop_monitor_family, skb, 0,
 146                                   0, GFP_KERNEL);
 147 }
 148
 149 /*
 150  * This is the timer function to delay the sending of an alert
 151  * in the event that more drops will arrive during the
 152  * hysteresis period.
 153  */
 154 static void sched_send_work(struct timer_list *t)
 155 {
 156         struct per_cpu_dm_data *data = from_timer(data, t, send_timer);
 157
 158         schedule_work(&data->dm_alert_work);
 159 }
 160
 161 static void trace_drop_common(struct sk_buff *skb, void *location)
 162 {
 163         struct net_dm_alert_msg *msg;
 164         struct nlmsghdr *nlh;
 165         struct nlattr *nla;
 166         int i;
 167         struct sk_buff *dskb;
 168         struct per_cpu_dm_data *data;
 169         unsigned long flags;
 170
 171         local_irq_save(flags);
 172         data = this_cpu_ptr(&dm_cpu_data);
 173         spin_lock(&data->lock);
 174         dskb = data->skb;
 175
 176         if (!dskb)
 177                 goto out;
 178
 179         nlh = (struct nlmsghdr *)dskb->data;
 180         nla = genlmsg_data(nlmsg_data(nlh));
 181         msg = nla_data(nla);
 182         for (i = 0; i < msg->entries; i++) {
 183                 if (!memcmp(&location, msg->points[i].pc, sizeof(void *))) {
 184                         msg->points[i].count++;
 185                         goto out;
 186                 }
 187         }
 188         if (msg->entries == dm_hit_limit)
 189                 goto out;
 190         /*
 191          * We need to create a new entry
 192          */
 193         __nla_reserve_nohdr(dskb, sizeof(struct net_dm_drop_point));
 194         nla->nla_len += NLA_ALIGN(sizeof(struct net_dm_drop_point));
 195         memcpy(msg->points[msg->entries].pc, &location, sizeof(void *));
 196         msg->points[msg->entries].count = 1;
 197         msg->entries++;
 198
 199         if (!timer_pending(&data->send_timer)) {
 200                 data->send_timer.expires = jiffies + dm_delay * HZ;
 201                 add_timer(&data->send_timer);
 202         }
 203
 204 out:
 205         spin_unlock_irqrestore(&data->lock, flags);
 206 }
 207
 208 static void trace_kfree_skb_hit(void *ignore, struct sk_buff *skb, void *location)
 209 {
 210         trace_drop_common(skb, location);
 211 }
 212
 213 static void trace_napi_poll_hit(void *ignore, struct napi_struct *napi,
 214                                 int work, int budget)
 215 {
 216         struct dm_hw_stat_delta *new_stat;
 217
 218         /*
 219          * Don't check napi structures with no associated device
 220          */
 221         if (!napi->dev)
 222                 return;
 223
 224         rcu_read_lock();
 225         list_for_each_entry_rcu(new_stat, &hw_stats_list, list) {
 226                 /*
 227                  * only add a note to our monitor buffer if:
 228                  * 1) this is the dev we received on
 229                  * 2) its after the last_rx delta
 230                  * 3) our rx_dropped count has gone up
 231                  */
 232                 if ((new_stat->dev == napi->dev)  &&
 233                     (time_after(jiffies, new_stat->last_rx + dm_hw_check_delta)) &&
 234                     (napi->dev->stats.rx_dropped != new_stat->last_drop_val)) {
 235                         trace_drop_common(NULL, NULL);
 236                         new_stat->last_drop_val = napi->dev->stats.rx_dropped;
 237                         new_stat->last_rx = jiffies;
 238                         break;
 239                 }
 240         }
 241         rcu_read_unlock();
 242 }
 243
 244 static int net_dm_trace_on_set(struct netlink_ext_ack *extack)
 245 {
 246         int cpu, rc;
 247
 248         if (!try_module_get(THIS_MODULE)) {
 249                 NL_SET_ERR_MSG_MOD(extack, "Failed to take reference on module");
 250                 return -ENODEV;
 251         }
 252
 253         for_each_possible_cpu(cpu) {
 254                 struct per_cpu_dm_data *data = &per_cpu(dm_cpu_data, cpu);
 255
 256                 INIT_WORK(&data->dm_alert_work, send_dm_alert);
 257                 timer_setup(&data->send_timer, sched_send_work, 0);
 258         }
 259
 260         rc = register_trace_kfree_skb(trace_kfree_skb_hit, NULL);
 261         if (rc) {
 262                 NL_SET_ERR_MSG_MOD(extack, "Failed to connect probe to kfree_skb() tracepoint");
 263                 goto err_module_put;
 264         }
 265
 266         rc = register_trace_napi_poll(trace_napi_poll_hit, NULL);
 267         if (rc) {
 268                 NL_SET_ERR_MSG_MOD(extack, "Failed to connect probe to napi_poll() tracepoint");
 269                 goto err_unregister_trace;
 270         }
 271
 272         return 0;
 273
 274 err_unregister_trace:
 275         unregister_trace_kfree_skb(trace_kfree_skb_hit, NULL);
 276 err_module_put:
 277         module_put(THIS_MODULE);
 278         return rc;
 279 }
 280
 281 static void net_dm_trace_off_set(void)
 282 {
 283         struct dm_hw_stat_delta *new_stat, *temp;
 284         int cpu;
 285
 286         unregister_trace_napi_poll(trace_napi_poll_hit, NULL);
 287         unregister_trace_kfree_skb(trace_kfree_skb_hit, NULL);
 288
 289         tracepoint_synchronize_unregister();
 290
 291         /* Make sure we do not send notifications to user space after request
 292          * to stop tracing returns.
 293          */
 294         for_each_possible_cpu(cpu) {
 295                 struct per_cpu_dm_data *data = &per_cpu(dm_cpu_data, cpu);
 296
 297                 del_timer_sync(&data->send_timer);
 298                 cancel_work_sync(&data->dm_alert_work);
 299         }
 300
 301         list_for_each_entry_safe(new_stat, temp, &hw_stats_list, list) {
 302                 if (new_stat->dev == NULL) {
 303                         list_del_rcu(&new_stat->list);
 304                         kfree_rcu(new_stat, rcu);
 305                 }
 306         }
 307
 308         module_put(THIS_MODULE);
 309 }
 310
 311 static int set_all_monitor_traces(int state, struct netlink_ext_ack *extack)
 312 {
 313         int rc = 0;
 314
 315         if (state == trace_state) {
 316                 NL_SET_ERR_MSG_MOD(extack, "Trace state already set to requested state");
 317                 return -EAGAIN;
 318         }
 319
 320         switch (state) {
 321         case TRACE_ON:
 322                 rc = net_dm_trace_on_set(extack);
 323                 break;
 324         case TRACE_OFF:
 325                 net_dm_trace_off_set();
 326                 break;
 327         default:
 328                 rc = 1;
 329                 break;
 330         }
 331
 332         if (!rc)
 333                 trace_state = state;
 334         else
 335                 rc = -EINPROGRESS;
 336
 337         return rc;
 338 }
 339
 340 static int net_dm_cmd_config(struct sk_buff *skb,
 341                         struct genl_info *info)
 342 {
 343         NL_SET_ERR_MSG_MOD(info->extack, "Command not supported");
 344
 345         return -EOPNOTSUPP;
 346 }
 347
 348 static int net_dm_cmd_trace(struct sk_buff *skb,
 349                         struct genl_info *info)
 350 {
 351         switch (info->genlhdr->cmd) {
 352         case NET_DM_CMD_START:
 353                 return set_all_monitor_traces(TRACE_ON, info->extack);
 354         case NET_DM_CMD_STOP:
 355                 return set_all_monitor_traces(TRACE_OFF, info->extack);
 356         }
 357
 358         return -EOPNOTSUPP;
 359 }
 360
 361 static int dropmon_net_event(struct notifier_block *ev_block,
 362                              unsigned long event, void *ptr)
 363 {
 364         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
 365         struct dm_hw_stat_delta *new_stat = NULL;
 366         struct dm_hw_stat_delta *tmp;
 367
 368         switch (event) {
 369         case NETDEV_REGISTER:
 370                 new_stat = kzalloc(sizeof(struct dm_hw_stat_delta), GFP_KERNEL);
 371
 372                 if (!new_stat)
 373                         goto out;
 374
 375                 new_stat->dev = dev;
 376                 new_stat->last_rx = jiffies;
 377                 mutex_lock(&net_dm_mutex);
 378                 list_add_rcu(&new_stat->list, &hw_stats_list);
 379                 mutex_unlock(&net_dm_mutex);
 380                 break;
 381         case NETDEV_UNREGISTER:
 382                 mutex_lock(&net_dm_mutex);
 383                 list_for_each_entry_safe(new_stat, tmp, &hw_stats_list, list) {
 384                         if (new_stat->dev == dev) {
 385                                 new_stat->dev = NULL;
 386                                 if (trace_state == TRACE_OFF) {
 387                                         list_del_rcu(&new_stat->list);
 388                                         kfree_rcu(new_stat, rcu);
 389                                         break;
 390                                 }
 391                         }
 392                 }
 393                 mutex_unlock(&net_dm_mutex);
 394                 break;
 395         }
 396 out:
 397         return NOTIFY_DONE;
 398 }
 399
 400 static const struct genl_ops dropmon_ops[] = {
 401         {
 402                 .cmd = NET_DM_CMD_CONFIG,
 403                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 404                 .doit = net_dm_cmd_config,
 405         },
 406         {
 407                 .cmd = NET_DM_CMD_START,
 408                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 409                 .doit = net_dm_cmd_trace,
 410         },
 411         {
 412                 .cmd = NET_DM_CMD_STOP,
 413                 .validate = GENL_DONT_VALIDATE_STRICT | GENL_DONT_VALIDATE_DUMP,
 414                 .doit = net_dm_cmd_trace,
 415         },
 416 };
 417
 418 static int net_dm_nl_pre_doit(const struct genl_ops *ops,
 419                               struct sk_buff *skb, struct genl_info *info)
 420 {
 421         mutex_lock(&net_dm_mutex);
 422
 423         return 0;
 424 }
 425
 426 static void net_dm_nl_post_doit(const struct genl_ops *ops,
 427                                 struct sk_buff *skb, struct genl_info *info)
 428 {
 429         mutex_unlock(&net_dm_mutex);
 430 }
 431
 432 static struct genl_family net_drop_monitor_family __ro_after_init = {
 433         .hdrsize        = 0,
 434         .name           = "NET_DM",
 435         .version        = 2,
 436         .pre_doit       = net_dm_nl_pre_doit,
 437         .post_doit      = net_dm_nl_post_doit,
 438         .module         = THIS_MODULE,
 439         .ops            = dropmon_ops,
 440         .n_ops          = ARRAY_SIZE(dropmon_ops),
 441         .mcgrps         = dropmon_mcgrps,
 442         .n_mcgrps       = ARRAY_SIZE(dropmon_mcgrps),
 443 };
 444
 445 static struct notifier_block dropmon_net_notifier = {
 446         .notifier_call = dropmon_net_event
 447 };
 448
 449 static int __init init_net_drop_monitor(void)
 450 {
 451         struct per_cpu_dm_data *data;
 452         int cpu, rc;
 453
 454         pr_info("Initializing network drop monitor service\n");
 455
 456         if (sizeof(void *) > 8) {
 457                 pr_err("Unable to store program counters on this arch, Drop monitor failed\n");
 458                 return -ENOSPC;
 459         }
 460
 461         rc = genl_register_family(&net_drop_monitor_family);
 462         if (rc) {
 463                 pr_err("Could not create drop monitor netlink family\n");
 464                 return rc;
 465         }
 466         WARN_ON(net_drop_monitor_family.mcgrp_offset != NET_DM_GRP_ALERT);
 467
 468         rc = register_netdevice_notifier(&dropmon_net_notifier);
 469         if (rc < 0) {
 470                 pr_crit("Failed to register netdevice notifier\n");
 471                 goto out_unreg;
 472         }
 473
 474         rc = 0;
 475
 476         for_each_possible_cpu(cpu) {
 477                 data = &per_cpu(dm_cpu_data, cpu);
 478                 INIT_WORK(&data->dm_alert_work, send_dm_alert);
 479                 timer_setup(&data->send_timer, sched_send_work, 0);
 480                 spin_lock_init(&data->lock);
 481                 reset_per_cpu_data(data);
 482         }
 483
 484         goto out;
 485
 486 out_unreg:
 487         genl_unregister_family(&net_drop_monitor_family);
 488 out:
 489         return rc;
 490 }
 491
 492 static void exit_net_drop_monitor(void)
 493 {
 494         struct per_cpu_dm_data *data;
 495         int cpu;
 496
 497         BUG_ON(unregister_netdevice_notifier(&dropmon_net_notifier));
 498
 499         /*
 500          * Because of the module_get/put we do in the trace state change path
 501          * we are guarnateed not to have any current users when we get here
 502          */
 503
 504         for_each_possible_cpu(cpu) {
 505                 data = &per_cpu(dm_cpu_data, cpu);
 506                 /*
 507                  * At this point, we should have exclusive access
 508                  * to this struct and can free the skb inside it
 509                  */
 510                 kfree_skb(data->skb);
 511         }
 512
 513         BUG_ON(genl_unregister_family(&net_drop_monitor_family));
 514 }
 515
 516 module_init(init_net_drop_monitor);
 517 module_exit(exit_net_drop_monitor);
 518
 519 MODULE_LICENSE("GPL v2");
 520 MODULE_AUTHOR("Neil Horman <nhorman@tuxdriver.com>");
 521 MODULE_ALIAS_GENL_FAMILY("NET_DM");