net/netfilter/nf_conntrack_ecache.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /* Event cache for netfilter. */
   3
   4 /*
   5  * (C) 2005 Harald Welte <laforge@gnumonks.org>
   6  * (C) 2005 Patrick McHardy <kaber@trash.net>
   7  * (C) 2005-2006 Netfilter Core Team <coreteam@netfilter.org>
   8  * (C) 2005 USAGI/WIDE Project <http://www.linux-ipv6.org>
   9  */
  10
  11 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  12
  13 #include <linux/types.h>
  14 #include <linux/netfilter.h>
  15 #include <linux/skbuff.h>
  16 #include <linux/vmalloc.h>
  17 #include <linux/stddef.h>
  18 #include <linux/err.h>
  19 #include <linux/percpu.h>
  20 #include <linux/kernel.h>
  21 #include <linux/netdevice.h>
  22 #include <linux/slab.h>
  23 #include <linux/export.h>
  24
  25 #include <net/netfilter/nf_conntrack.h>
  26 #include <net/netfilter/nf_conntrack_core.h>
  27 #include <net/netfilter/nf_conntrack_ecache.h>
  28 #include <net/netfilter/nf_conntrack_extend.h>
  29
  30 extern unsigned int nf_conntrack_net_id;
  31
  32 static DEFINE_MUTEX(nf_ct_ecache_mutex);
  33
  34 #define ECACHE_RETRY_WAIT (HZ/10)
  35 #define ECACHE_STACK_ALLOC (256 / sizeof(void *))
  36
  37 enum retry_state {
  38         STATE_CONGESTED,
  39         STATE_RESTART,
  40         STATE_DONE,
  41 };
  42
  43 static enum retry_state ecache_work_evict_list(struct ct_pcpu *pcpu)
  44 {
  45         struct nf_conn *refs[ECACHE_STACK_ALLOC];
  46         enum retry_state ret = STATE_DONE;
  47         struct nf_conntrack_tuple_hash *h;
  48         struct hlist_nulls_node *n;
  49         unsigned int evicted = 0;
  50
  51         spin_lock(&pcpu->lock);
  52
  53         hlist_nulls_for_each_entry(h, n, &pcpu->dying, hnnode) {
  54                 struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
  55                 struct nf_conntrack_ecache *e;
  56
  57                 if (!nf_ct_is_confirmed(ct))
  58                         continue;
  59
  60                 /* This ecache access is safe because the ct is on the
  61                  * pcpu dying list and we hold the spinlock -- the entry
  62                  * cannot be free'd until after the lock is released.
  63                  *
  64                  * This is true even if ct has a refcount of 0: the
  65                  * cpu that is about to free the entry must remove it
  66                  * from the dying list and needs the lock to do so.
  67                  */
  68                 e = nf_ct_ecache_find(ct);
  69                 if (!e || e->state != NFCT_ECACHE_DESTROY_FAIL)
  70                         continue;
  71
  72                 /* ct is in NFCT_ECACHE_DESTROY_FAIL state, this means
  73                  * the worker owns this entry: the ct will remain valid
  74                  * until the worker puts its ct reference.
  75                  */
  76                 if (nf_conntrack_event(IPCT_DESTROY, ct)) {
  77                         ret = STATE_CONGESTED;
  78                         break;
  79                 }
  80
  81                 e->state = NFCT_ECACHE_DESTROY_SENT;
  82                 refs[evicted] = ct;
  83
  84                 if (++evicted >= ARRAY_SIZE(refs)) {
  85                         ret = STATE_RESTART;
  86                         break;
  87                 }
  88         }
  89
  90         spin_unlock(&pcpu->lock);
  91
  92         /* can't _put while holding lock */
  93         while (evicted)
  94                 nf_ct_put(refs[--evicted]);
  95
  96         return ret;
  97 }
  98
  99 static void ecache_work(struct work_struct *work)
 100 {
 101         struct nf_conntrack_net *cnet = container_of(work, struct nf_conntrack_net, ecache_dwork.work);
 102         struct netns_ct *ctnet = cnet->ct_net;
 103         int cpu, delay = -1;
 104         struct ct_pcpu *pcpu;
 105
 106         local_bh_disable();
 107
 108         for_each_possible_cpu(cpu) {
 109                 enum retry_state ret;
 110
 111                 pcpu = per_cpu_ptr(ctnet->pcpu_lists, cpu);
 112
 113                 ret = ecache_work_evict_list(pcpu);
 114
 115                 switch (ret) {
 116                 case STATE_CONGESTED:
 117                         delay = ECACHE_RETRY_WAIT;
 118                         goto out;
 119                 case STATE_RESTART:
 120                         delay = 0;
 121                         break;
 122                 case STATE_DONE:
 123                         break;
 124                 }
 125         }
 126
 127  out:
 128         local_bh_enable();
 129
 130         ctnet->ecache_dwork_pending = delay > 0;
 131         if (delay >= 0)
 132                 schedule_delayed_work(&cnet->ecache_dwork, delay);
 133 }
 134
 135 int nf_conntrack_eventmask_report(unsigned int eventmask, struct nf_conn *ct,
 136                                   u32 portid, int report)
 137 {
 138         int ret = 0;
 139         struct net *net = nf_ct_net(ct);
 140         struct nf_ct_event_notifier *notify;
 141         struct nf_conntrack_ecache *e;
 142
 143         rcu_read_lock();
 144         notify = rcu_dereference(net->ct.nf_conntrack_event_cb);
 145         if (!notify)
 146                 goto out_unlock;
 147
 148         e = nf_ct_ecache_find(ct);
 149         if (!e)
 150                 goto out_unlock;
 151
 152         if (nf_ct_is_confirmed(ct)) {
 153                 struct nf_ct_event item = {
 154                         .ct     = ct,
 155                         .portid = e->portid ? e->portid : portid,
 156                         .report = report
 157                 };
 158                 /* This is a resent of a destroy event? If so, skip missed */
 159                 unsigned long missed = e->portid ? 0 : e->missed;
 160
 161                 if (!((eventmask | missed) & e->ctmask))
 162                         goto out_unlock;
 163
 164                 ret = notify->fcn(eventmask | missed, &item);
 165                 if (unlikely(ret < 0 || missed)) {
 166                         spin_lock_bh(&ct->lock);
 167                         if (ret < 0) {
 168                                 /* This is a destroy event that has been
 169                                  * triggered by a process, we store the PORTID
 170                                  * to include it in the retransmission.
 171                                  */
 172                                 if (eventmask & (1 << IPCT_DESTROY)) {
 173                                         if (e->portid == 0 && portid != 0)
 174                                                 e->portid = portid;
 175                                         e->state = NFCT_ECACHE_DESTROY_FAIL;
 176                                 } else {
 177                                         e->missed |= eventmask;
 178                                 }
 179                         } else {
 180                                 e->missed &= ~missed;
 181                         }
 182                         spin_unlock_bh(&ct->lock);
 183                 }
 184         }
 185 out_unlock:
 186         rcu_read_unlock();
 187         return ret;
 188 }
 189 EXPORT_SYMBOL_GPL(nf_conntrack_eventmask_report);
 190
 191 /* deliver cached events and clear cache entry - must be called with locally
 192  * disabled softirqs */
 193 void nf_ct_deliver_cached_events(struct nf_conn *ct)
 194 {
 195         struct net *net = nf_ct_net(ct);
 196         unsigned long events, missed;
 197         struct nf_ct_event_notifier *notify;
 198         struct nf_conntrack_ecache *e;
 199         struct nf_ct_event item;
 200         int ret;
 201
 202         rcu_read_lock();
 203         notify = rcu_dereference(net->ct.nf_conntrack_event_cb);
 204         if (notify == NULL)
 205                 goto out_unlock;
 206
 207         if (!nf_ct_is_confirmed(ct) || nf_ct_is_dying(ct))
 208                 goto out_unlock;
 209
 210         e = nf_ct_ecache_find(ct);
 211         if (e == NULL)
 212                 goto out_unlock;
 213
 214         events = xchg(&e->cache, 0);
 215
 216         /* We make a copy of the missed event cache without taking
 217          * the lock, thus we may send missed events twice. However,
 218          * this does not harm and it happens very rarely. */
 219         missed = e->missed;
 220
 221         if (!((events | missed) & e->ctmask))
 222                 goto out_unlock;
 223
 224         item.ct = ct;
 225         item.portid = 0;
 226         item.report = 0;
 227
 228         ret = notify->fcn(events | missed, &item);
 229
 230         if (likely(ret == 0 && !missed))
 231                 goto out_unlock;
 232
 233         spin_lock_bh(&ct->lock);
 234         if (ret < 0)
 235                 e->missed |= events;
 236         else
 237                 e->missed &= ~missed;
 238         spin_unlock_bh(&ct->lock);
 239
 240 out_unlock:
 241         rcu_read_unlock();
 242 }
 243 EXPORT_SYMBOL_GPL(nf_ct_deliver_cached_events);
 244
 245 void nf_ct_expect_event_report(enum ip_conntrack_expect_events event,
 246                                struct nf_conntrack_expect *exp,
 247                                u32 portid, int report)
 248
 249 {
 250         struct net *net = nf_ct_exp_net(exp);
 251         struct nf_exp_event_notifier *notify;
 252         struct nf_conntrack_ecache *e;
 253
 254         rcu_read_lock();
 255         notify = rcu_dereference(net->ct.nf_expect_event_cb);
 256         if (!notify)
 257                 goto out_unlock;
 258
 259         e = nf_ct_ecache_find(exp->master);
 260         if (!e)
 261                 goto out_unlock;
 262
 263         if (e->expmask & (1 << event)) {
 264                 struct nf_exp_event item = {
 265                         .exp    = exp,
 266                         .portid = portid,
 267                         .report = report
 268                 };
 269                 notify->fcn(1 << event, &item);
 270         }
 271 out_unlock:
 272         rcu_read_unlock();
 273 }
 274
 275 int nf_conntrack_register_notifier(struct net *net,
 276                                    struct nf_ct_event_notifier *new)
 277 {
 278         int ret;
 279         struct nf_ct_event_notifier *notify;
 280
 281         mutex_lock(&nf_ct_ecache_mutex);
 282         notify = rcu_dereference_protected(net->ct.nf_conntrack_event_cb,
 283                                            lockdep_is_held(&nf_ct_ecache_mutex));
 284         if (notify != NULL) {
 285                 ret = -EBUSY;
 286                 goto out_unlock;
 287         }
 288         rcu_assign_pointer(net->ct.nf_conntrack_event_cb, new);
 289         ret = 0;
 290
 291 out_unlock:
 292         mutex_unlock(&nf_ct_ecache_mutex);
 293         return ret;
 294 }
 295 EXPORT_SYMBOL_GPL(nf_conntrack_register_notifier);
 296
 297 void nf_conntrack_unregister_notifier(struct net *net,
 298                                       struct nf_ct_event_notifier *new)
 299 {
 300         struct nf_ct_event_notifier *notify;
 301
 302         mutex_lock(&nf_ct_ecache_mutex);
 303         notify = rcu_dereference_protected(net->ct.nf_conntrack_event_cb,
 304                                            lockdep_is_held(&nf_ct_ecache_mutex));
 305         BUG_ON(notify != new);
 306         RCU_INIT_POINTER(net->ct.nf_conntrack_event_cb, NULL);
 307         mutex_unlock(&nf_ct_ecache_mutex);
 308         /* synchronize_rcu() is called from ctnetlink_exit. */
 309 }
 310 EXPORT_SYMBOL_GPL(nf_conntrack_unregister_notifier);
 311
 312 int nf_ct_expect_register_notifier(struct net *net,
 313                                    struct nf_exp_event_notifier *new)
 314 {
 315         int ret;
 316         struct nf_exp_event_notifier *notify;
 317
 318         mutex_lock(&nf_ct_ecache_mutex);
 319         notify = rcu_dereference_protected(net->ct.nf_expect_event_cb,
 320                                            lockdep_is_held(&nf_ct_ecache_mutex));
 321         if (notify != NULL) {
 322                 ret = -EBUSY;
 323                 goto out_unlock;
 324         }
 325         rcu_assign_pointer(net->ct.nf_expect_event_cb, new);
 326         ret = 0;
 327
 328 out_unlock:
 329         mutex_unlock(&nf_ct_ecache_mutex);
 330         return ret;
 331 }
 332 EXPORT_SYMBOL_GPL(nf_ct_expect_register_notifier);
 333
 334 void nf_ct_expect_unregister_notifier(struct net *net,
 335                                       struct nf_exp_event_notifier *new)
 336 {
 337         struct nf_exp_event_notifier *notify;
 338
 339         mutex_lock(&nf_ct_ecache_mutex);
 340         notify = rcu_dereference_protected(net->ct.nf_expect_event_cb,
 341                                            lockdep_is_held(&nf_ct_ecache_mutex));
 342         BUG_ON(notify != new);
 343         RCU_INIT_POINTER(net->ct.nf_expect_event_cb, NULL);
 344         mutex_unlock(&nf_ct_ecache_mutex);
 345         /* synchronize_rcu() is called from ctnetlink_exit. */
 346 }
 347 EXPORT_SYMBOL_GPL(nf_ct_expect_unregister_notifier);
 348
 349 void nf_conntrack_ecache_work(struct net *net, enum nf_ct_ecache_state state)
 350 {
 351         struct nf_conntrack_net *cnet = net_generic(net, nf_conntrack_net_id);
 352
 353         if (state == NFCT_ECACHE_DESTROY_FAIL &&
 354             !delayed_work_pending(&cnet->ecache_dwork)) {
 355                 schedule_delayed_work(&cnet->ecache_dwork, HZ);
 356                 net->ct.ecache_dwork_pending = true;
 357         } else if (state == NFCT_ECACHE_DESTROY_SENT) {
 358                 net->ct.ecache_dwork_pending = false;
 359                 mod_delayed_work(system_wq, &cnet->ecache_dwork, 0);
 360         }
 361 }
 362
 363 #define NF_CT_EVENTS_DEFAULT 1
 364 static int nf_ct_events __read_mostly = NF_CT_EVENTS_DEFAULT;
 365
 366 static const struct nf_ct_ext_type event_extend = {
 367         .len    = sizeof(struct nf_conntrack_ecache),
 368         .align  = __alignof__(struct nf_conntrack_ecache),
 369         .id     = NF_CT_EXT_ECACHE,
 370 };
 371
 372 void nf_conntrack_ecache_pernet_init(struct net *net)
 373 {
 374         struct nf_conntrack_net *cnet = net_generic(net, nf_conntrack_net_id);
 375
 376         net->ct.sysctl_events = nf_ct_events;
 377         cnet->ct_net = &net->ct;
 378         INIT_DELAYED_WORK(&cnet->ecache_dwork, ecache_work);
 379 }
 380
 381 void nf_conntrack_ecache_pernet_fini(struct net *net)
 382 {
 383         struct nf_conntrack_net *cnet = net_generic(net, nf_conntrack_net_id);
 384
 385         cancel_delayed_work_sync(&cnet->ecache_dwork);
 386 }
 387
 388 int nf_conntrack_ecache_init(void)
 389 {
 390         int ret = nf_ct_extend_register(&event_extend);
 391         if (ret < 0)
 392                 pr_err("Unable to register event extension\n");
 393
 394         BUILD_BUG_ON(__IPCT_MAX >= 16); /* ctmask, missed use u16 */
 395
 396         return ret;
 397 }
 398
 399 void nf_conntrack_ecache_fini(void)
 400 {
 401         nf_ct_extend_unregister(&event_extend);
 402 }