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