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_extend.h>
  28
  29 static DEFINE_MUTEX(nf_ct_ecache_mutex);
  30
  31 #define ECACHE_RETRY_WAIT (HZ/10)
  32
  33 enum retry_state {
  34         STATE_CONGESTED,
  35         STATE_RESTART,
  36         STATE_DONE,
  37 };
  38
  39 static enum retry_state ecache_work_evict_list(struct ct_pcpu *pcpu)
  40 {
  41         struct nf_conn *refs[16];
  42         struct nf_conntrack_tuple_hash *h;
  43         struct hlist_nulls_node *n;
  44         unsigned int evicted = 0;
  45         enum retry_state ret = STATE_DONE;
  46
  47         spin_lock(&pcpu->lock);
  48
  49         hlist_nulls_for_each_entry(h, n, &pcpu->dying, hnnode) {
  50                 struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
  51                 struct nf_conntrack_ecache *e;
  52
  53                 if (!nf_ct_is_confirmed(ct))
  54                         continue;
  55
  56                 e = nf_ct_ecache_find(ct);
  57                 if (!e || e->state != NFCT_ECACHE_DESTROY_FAIL)
  58                         continue;
  59
  60                 if (nf_conntrack_event(IPCT_DESTROY, ct)) {
  61                         ret = STATE_CONGESTED;
  62                         break;
  63                 }
  64
  65                 e->state = NFCT_ECACHE_DESTROY_SENT;
  66                 refs[evicted] = ct;
  67
  68                 if (++evicted >= ARRAY_SIZE(refs)) {
  69                         ret = STATE_RESTART;
  70                         break;
  71                 }
  72         }
  73
  74         spin_unlock(&pcpu->lock);
  75
  76         /* can't _put while holding lock */
  77         while (evicted)
  78                 nf_ct_put(refs[--evicted]);
  79
  80         return ret;
  81 }
  82
  83 static void ecache_work(struct work_struct *work)
  84 {
  85         struct netns_ct *ctnet =
  86                 container_of(work, struct netns_ct, ecache_dwork.work);
  87         int cpu, delay = -1;
  88         struct ct_pcpu *pcpu;
  89
  90         local_bh_disable();
  91
  92         for_each_possible_cpu(cpu) {
  93                 enum retry_state ret;
  94
  95                 pcpu = per_cpu_ptr(ctnet->pcpu_lists, cpu);
  96
  97                 ret = ecache_work_evict_list(pcpu);
  98
  99                 switch (ret) {
 100                 case STATE_CONGESTED:
 101                         delay = ECACHE_RETRY_WAIT;
 102                         goto out;
 103                 case STATE_RESTART:
 104                         delay = 0;
 105                         break;
 106                 case STATE_DONE:
 107                         break;
 108                 }
 109         }
 110
 111  out:
 112         local_bh_enable();
 113
 114         ctnet->ecache_dwork_pending = delay > 0;
 115         if (delay >= 0)
 116                 schedule_delayed_work(&ctnet->ecache_dwork, delay);
 117 }
 118
 119 int nf_conntrack_eventmask_report(unsigned int eventmask, struct nf_conn *ct,
 120                                   u32 portid, int report)
 121 {
 122         int ret = 0;
 123         struct net *net = nf_ct_net(ct);
 124         struct nf_ct_event_notifier *notify;
 125         struct nf_conntrack_ecache *e;
 126
 127         rcu_read_lock();
 128         notify = rcu_dereference(net->ct.nf_conntrack_event_cb);
 129         if (!notify)
 130                 goto out_unlock;
 131
 132         e = nf_ct_ecache_find(ct);
 133         if (!e)
 134                 goto out_unlock;
 135
 136         if (nf_ct_is_confirmed(ct)) {
 137                 struct nf_ct_event item = {
 138                         .ct     = ct,
 139                         .portid = e->portid ? e->portid : portid,
 140                         .report = report
 141                 };
 142                 /* This is a resent of a destroy event? If so, skip missed */
 143                 unsigned long missed = e->portid ? 0 : e->missed;
 144
 145                 if (!((eventmask | missed) & e->ctmask))
 146                         goto out_unlock;
 147
 148                 ret = notify->fcn(eventmask | missed, &item);
 149                 if (unlikely(ret < 0 || missed)) {
 150                         spin_lock_bh(&ct->lock);
 151                         if (ret < 0) {
 152                                 /* This is a destroy event that has been
 153                                  * triggered by a process, we store the PORTID
 154                                  * to include it in the retransmission.
 155                                  */
 156                                 if (eventmask & (1 << IPCT_DESTROY)) {
 157                                         if (e->portid == 0 && portid != 0)
 158                                                 e->portid = portid;
 159                                         e->state = NFCT_ECACHE_DESTROY_FAIL;
 160                                 } else {
 161                                         e->missed |= eventmask;
 162                                 }
 163                         } else {
 164                                 e->missed &= ~missed;
 165                         }
 166                         spin_unlock_bh(&ct->lock);
 167                 }
 168         }
 169 out_unlock:
 170         rcu_read_unlock();
 171         return ret;
 172 }
 173 EXPORT_SYMBOL_GPL(nf_conntrack_eventmask_report);
 174
 175 /* deliver cached events and clear cache entry - must be called with locally
 176  * disabled softirqs */
 177 void nf_ct_deliver_cached_events(struct nf_conn *ct)
 178 {
 179         struct net *net = nf_ct_net(ct);
 180         unsigned long events, missed;
 181         struct nf_ct_event_notifier *notify;
 182         struct nf_conntrack_ecache *e;
 183         struct nf_ct_event item;
 184         int ret;
 185
 186         rcu_read_lock();
 187         notify = rcu_dereference(net->ct.nf_conntrack_event_cb);
 188         if (notify == NULL)
 189                 goto out_unlock;
 190
 191         e = nf_ct_ecache_find(ct);
 192         if (e == NULL)
 193                 goto out_unlock;
 194
 195         events = xchg(&e->cache, 0);
 196
 197         if (!nf_ct_is_confirmed(ct) || nf_ct_is_dying(ct))
 198                 goto out_unlock;
 199
 200         /* We make a copy of the missed event cache without taking
 201          * the lock, thus we may send missed events twice. However,
 202          * this does not harm and it happens very rarely. */
 203         missed = e->missed;
 204
 205         if (!((events | missed) & e->ctmask))
 206                 goto out_unlock;
 207
 208         item.ct = ct;
 209         item.portid = 0;
 210         item.report = 0;
 211
 212         ret = notify->fcn(events | missed, &item);
 213
 214         if (likely(ret == 0 && !missed))
 215                 goto out_unlock;
 216
 217         spin_lock_bh(&ct->lock);
 218         if (ret < 0)
 219                 e->missed |= events;
 220         else
 221                 e->missed &= ~missed;
 222         spin_unlock_bh(&ct->lock);
 223
 224 out_unlock:
 225         rcu_read_unlock();
 226 }
 227 EXPORT_SYMBOL_GPL(nf_ct_deliver_cached_events);
 228
 229 void nf_ct_expect_event_report(enum ip_conntrack_expect_events event,
 230                                struct nf_conntrack_expect *exp,
 231                                u32 portid, int report)
 232
 233 {
 234         struct net *net = nf_ct_exp_net(exp);
 235         struct nf_exp_event_notifier *notify;
 236         struct nf_conntrack_ecache *e;
 237
 238         rcu_read_lock();
 239         notify = rcu_dereference(net->ct.nf_expect_event_cb);
 240         if (!notify)
 241                 goto out_unlock;
 242
 243         e = nf_ct_ecache_find(exp->master);
 244         if (!e)
 245                 goto out_unlock;
 246
 247         if (e->expmask & (1 << event)) {
 248                 struct nf_exp_event item = {
 249                         .exp    = exp,
 250                         .portid = portid,
 251                         .report = report
 252                 };
 253                 notify->fcn(1 << event, &item);
 254         }
 255 out_unlock:
 256         rcu_read_unlock();
 257 }
 258
 259 int nf_conntrack_register_notifier(struct net *net,
 260                                    struct nf_ct_event_notifier *new)
 261 {
 262         int ret;
 263         struct nf_ct_event_notifier *notify;
 264
 265         mutex_lock(&nf_ct_ecache_mutex);
 266         notify = rcu_dereference_protected(net->ct.nf_conntrack_event_cb,
 267                                            lockdep_is_held(&nf_ct_ecache_mutex));
 268         if (notify != NULL) {
 269                 ret = -EBUSY;
 270                 goto out_unlock;
 271         }
 272         rcu_assign_pointer(net->ct.nf_conntrack_event_cb, new);
 273         ret = 0;
 274
 275 out_unlock:
 276         mutex_unlock(&nf_ct_ecache_mutex);
 277         return ret;
 278 }
 279 EXPORT_SYMBOL_GPL(nf_conntrack_register_notifier);
 280
 281 void nf_conntrack_unregister_notifier(struct net *net,
 282                                       struct nf_ct_event_notifier *new)
 283 {
 284         struct nf_ct_event_notifier *notify;
 285
 286         mutex_lock(&nf_ct_ecache_mutex);
 287         notify = rcu_dereference_protected(net->ct.nf_conntrack_event_cb,
 288                                            lockdep_is_held(&nf_ct_ecache_mutex));
 289         BUG_ON(notify != new);
 290         RCU_INIT_POINTER(net->ct.nf_conntrack_event_cb, NULL);
 291         mutex_unlock(&nf_ct_ecache_mutex);
 292         /* synchronize_rcu() is called from ctnetlink_exit. */
 293 }
 294 EXPORT_SYMBOL_GPL(nf_conntrack_unregister_notifier);
 295
 296 int nf_ct_expect_register_notifier(struct net *net,
 297                                    struct nf_exp_event_notifier *new)
 298 {
 299         int ret;
 300         struct nf_exp_event_notifier *notify;
 301
 302         mutex_lock(&nf_ct_ecache_mutex);
 303         notify = rcu_dereference_protected(net->ct.nf_expect_event_cb,
 304                                            lockdep_is_held(&nf_ct_ecache_mutex));
 305         if (notify != NULL) {
 306                 ret = -EBUSY;
 307                 goto out_unlock;
 308         }
 309         rcu_assign_pointer(net->ct.nf_expect_event_cb, new);
 310         ret = 0;
 311
 312 out_unlock:
 313         mutex_unlock(&nf_ct_ecache_mutex);
 314         return ret;
 315 }
 316 EXPORT_SYMBOL_GPL(nf_ct_expect_register_notifier);
 317
 318 void nf_ct_expect_unregister_notifier(struct net *net,
 319                                       struct nf_exp_event_notifier *new)
 320 {
 321         struct nf_exp_event_notifier *notify;
 322
 323         mutex_lock(&nf_ct_ecache_mutex);
 324         notify = rcu_dereference_protected(net->ct.nf_expect_event_cb,
 325                                            lockdep_is_held(&nf_ct_ecache_mutex));
 326         BUG_ON(notify != new);
 327         RCU_INIT_POINTER(net->ct.nf_expect_event_cb, NULL);
 328         mutex_unlock(&nf_ct_ecache_mutex);
 329         /* synchronize_rcu() is called from ctnetlink_exit. */
 330 }
 331 EXPORT_SYMBOL_GPL(nf_ct_expect_unregister_notifier);
 332
 333 #define NF_CT_EVENTS_DEFAULT 1
 334 static int nf_ct_events __read_mostly = NF_CT_EVENTS_DEFAULT;
 335
 336 static const struct nf_ct_ext_type event_extend = {
 337         .len    = sizeof(struct nf_conntrack_ecache),
 338         .align  = __alignof__(struct nf_conntrack_ecache),
 339         .id     = NF_CT_EXT_ECACHE,
 340 };
 341
 342 void nf_conntrack_ecache_pernet_init(struct net *net)
 343 {
 344         net->ct.sysctl_events = nf_ct_events;
 345         INIT_DELAYED_WORK(&net->ct.ecache_dwork, ecache_work);
 346 }
 347
 348 void nf_conntrack_ecache_pernet_fini(struct net *net)
 349 {
 350         cancel_delayed_work_sync(&net->ct.ecache_dwork);
 351 }
 352
 353 int nf_conntrack_ecache_init(void)
 354 {
 355         int ret = nf_ct_extend_register(&event_extend);
 356         if (ret < 0)
 357                 pr_err("Unable to register event extension\n");
 358
 359         BUILD_BUG_ON(__IPCT_MAX >= 16); /* ctmask, missed use u16 */
 360
 361         return ret;
 362 }
 363
 364 void nf_conntrack_ecache_fini(void)
 365 {
 366         nf_ct_extend_unregister(&event_extend);
 367 }