net/netfilter/nf_conntrack_ecache.c

0001 // SPDX-License-Identifier: GPL-2.0-only
0002 /* Event cache for netfilter. */
0003
0004 /*
0005  * (C) 2005 Harald Welte <laforge@gnumonks.org>
0006  * (C) 2005 Patrick McHardy <kaber@trash.net>
0007  * (C) 2005-2006 Netfilter Core Team <coreteam@netfilter.org>
0008  * (C) 2005 USAGI/WIDE Project <http://www.linux-ipv6.org>
0009  */
0010
0011 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
0012
0013 #include <linux/types.h>
0014 #include <linux/netfilter.h>
0015 #include <linux/skbuff.h>
0016 #include <linux/vmalloc.h>
0017 #include <linux/stddef.h>
0018 #include <linux/err.h>
0019 #include <linux/kernel.h>
0020 #include <linux/netdevice.h>
0021 #include <linux/slab.h>
0022 #include <linux/export.h>
0023
0024 #include <net/netfilter/nf_conntrack.h>
0025 #include <net/netfilter/nf_conntrack_core.h>
0026 #include <net/netfilter/nf_conntrack_ecache.h>
0027 #include <net/netfilter/nf_conntrack_extend.h>
0028
0029 static DEFINE_MUTEX(nf_ct_ecache_mutex);
0030
0031 #define DYING_NULLS_VAL         ((1 << 30) + 1)
0032 #define ECACHE_MAX_JIFFIES      msecs_to_jiffies(10)
0033 #define ECACHE_RETRY_JIFFIES        msecs_to_jiffies(10)
0034
0035 enum retry_state {
0036     STATE_CONGESTED,
0037     STATE_RESTART,
0038     STATE_DONE,
0039 };
0040
0041 struct nf_conntrack_net_ecache *nf_conn_pernet_ecache(const struct net *net)
0042 {
0043     struct nf_conntrack_net *cnet = nf_ct_pernet(net);
0044
0045     return &cnet->ecache;
0046 }
0047 #if IS_MODULE(CONFIG_NF_CT_NETLINK)
0048 EXPORT_SYMBOL_GPL(nf_conn_pernet_ecache);
0049 #endif
0050
0051 static enum retry_state ecache_work_evict_list(struct nf_conntrack_net *cnet)
0052 {
0053     unsigned long stop = jiffies + ECACHE_MAX_JIFFIES;
0054     struct hlist_nulls_head evicted_list;
0055     enum retry_state ret = STATE_DONE;
0056     struct nf_conntrack_tuple_hash *h;
0057     struct hlist_nulls_node *n;
0058     unsigned int sent;
0059
0060     INIT_HLIST_NULLS_HEAD(&evicted_list, DYING_NULLS_VAL);
0061
0062 next:
0063     sent = 0;
0064     spin_lock_bh(&cnet->ecache.dying_lock);
0065
0066     hlist_nulls_for_each_entry_safe(h, n, &cnet->ecache.dying_list, hnnode) {
0067         struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
0068
0069         /* The worker owns all entries, ct remains valid until nf_ct_put
0070          * in the loop below.
0071          */
0072         if (nf_conntrack_event(IPCT_DESTROY, ct)) {
0073             ret = STATE_CONGESTED;
0074             break;
0075         }
0076
0077         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
0078         hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode, &evicted_list);
0079
0080         if (time_after(stop, jiffies)) {
0081             ret = STATE_RESTART;
0082             break;
0083         }
0084
0085         if (sent++ > 16) {
0086             spin_unlock_bh(&cnet->ecache.dying_lock);
0087             cond_resched();
0088             goto next;
0089         }
0090     }
0091
0092     spin_unlock_bh(&cnet->ecache.dying_lock);
0093
0094     hlist_nulls_for_each_entry_safe(h, n, &evicted_list, hnnode) {
0095         struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
0096
0097         hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode);
0098         nf_ct_put(ct);
0099
0100         cond_resched();
0101     }
0102
0103     return ret;
0104 }
0105
0106 static void ecache_work(struct work_struct *work)
0107 {
0108     struct nf_conntrack_net *cnet = container_of(work, struct nf_conntrack_net, ecache.dwork.work);
0109     int ret, delay = -1;
0110
0111     ret = ecache_work_evict_list(cnet);
0112     switch (ret) {
0113     case STATE_CONGESTED:
0114         delay = ECACHE_RETRY_JIFFIES;
0115         break;
0116     case STATE_RESTART:
0117         delay = 0;
0118         break;
0119     case STATE_DONE:
0120         break;
0121     }
0122
0123     if (delay >= 0)
0124         schedule_delayed_work(&cnet->ecache.dwork, delay);
0125 }
0126
0127 static int __nf_conntrack_eventmask_report(struct nf_conntrack_ecache *e,
0128                        const u32 events,
0129                        const u32 missed,
0130                        const struct nf_ct_event *item)
0131 {
0132     struct net *net = nf_ct_net(item->ct);
0133     struct nf_ct_event_notifier *notify;
0134     u32 old, want;
0135     int ret;
0136
0137     if (!((events | missed) & e->ctmask))
0138         return 0;
0139
0140     rcu_read_lock();
0141
0142     notify = rcu_dereference(net->ct.nf_conntrack_event_cb);
0143     if (!notify) {
0144         rcu_read_unlock();
0145         return 0;
0146     }
0147
0148     ret = notify->ct_event(events | missed, item);
0149     rcu_read_unlock();
0150
0151     if (likely(ret >= 0 && missed == 0))
0152         return 0;
0153
0154     do {
0155         old = READ_ONCE(e->missed);
0156         if (ret < 0)
0157             want = old | events;
0158         else
0159             want = old & ~missed;
0160     } while (cmpxchg(&e->missed, old, want) != old);
0161
0162     return ret;
0163 }
0164
0165 int nf_conntrack_eventmask_report(unsigned int events, struct nf_conn *ct,
0166                   u32 portid, int report)
0167 {
0168     struct nf_conntrack_ecache *e;
0169     struct nf_ct_event item;
0170     unsigned int missed;
0171     int ret;
0172
0173     if (!nf_ct_is_confirmed(ct))
0174         return 0;
0175
0176     e = nf_ct_ecache_find(ct);
0177     if (!e)
0178         return 0;
0179
0180     memset(&item, 0, sizeof(item));
0181
0182     item.ct = ct;
0183     item.portid = e->portid ? e->portid : portid;
0184     item.report = report;
0185
0186     /* This is a resent of a destroy event? If so, skip missed */
0187     missed = e->portid ? 0 : e->missed;
0188
0189     ret = __nf_conntrack_eventmask_report(e, events, missed, &item);
0190     if (unlikely(ret < 0 && (events & (1 << IPCT_DESTROY)))) {
0191         /* This is a destroy event that has been triggered by a process,
0192          * we store the PORTID to include it in the retransmission.
0193          */
0194         if (e->portid == 0 && portid != 0)
0195             e->portid = portid;
0196     }
0197
0198     return ret;
0199 }
0200 EXPORT_SYMBOL_GPL(nf_conntrack_eventmask_report);
0201
0202 /* deliver cached events and clear cache entry - must be called with locally
0203  * disabled softirqs */
0204 void nf_ct_deliver_cached_events(struct nf_conn *ct)
0205 {
0206     struct nf_conntrack_ecache *e;
0207     struct nf_ct_event item;
0208     unsigned int events;
0209
0210     if (!nf_ct_is_confirmed(ct) || nf_ct_is_dying(ct))
0211         return;
0212
0213     e = nf_ct_ecache_find(ct);
0214     if (e == NULL)
0215         return;
0216
0217     events = xchg(&e->cache, 0);
0218
0219     item.ct = ct;
0220     item.portid = 0;
0221     item.report = 0;
0222
0223     /* We make a copy of the missed event cache without taking
0224      * the lock, thus we may send missed events twice. However,
0225      * this does not harm and it happens very rarely.
0226      */
0227     __nf_conntrack_eventmask_report(e, events, e->missed, &item);
0228 }
0229 EXPORT_SYMBOL_GPL(nf_ct_deliver_cached_events);
0230
0231 void nf_ct_expect_event_report(enum ip_conntrack_expect_events event,
0232                    struct nf_conntrack_expect *exp,
0233                    u32 portid, int report)
0234
0235 {
0236     struct net *net = nf_ct_exp_net(exp);
0237     struct nf_ct_event_notifier *notify;
0238     struct nf_conntrack_ecache *e;
0239
0240     rcu_read_lock();
0241     notify = rcu_dereference(net->ct.nf_conntrack_event_cb);
0242     if (!notify)
0243         goto out_unlock;
0244
0245     e = nf_ct_ecache_find(exp->master);
0246     if (!e)
0247         goto out_unlock;
0248
0249     if (e->expmask & (1 << event)) {
0250         struct nf_exp_event item = {
0251             .exp    = exp,
0252             .portid = portid,
0253             .report = report
0254         };
0255         notify->exp_event(1 << event, &item);
0256     }
0257 out_unlock:
0258     rcu_read_unlock();
0259 }
0260
0261 void nf_conntrack_register_notifier(struct net *net,
0262                     const struct nf_ct_event_notifier *new)
0263 {
0264     struct nf_ct_event_notifier *notify;
0265
0266     mutex_lock(&nf_ct_ecache_mutex);
0267     notify = rcu_dereference_protected(net->ct.nf_conntrack_event_cb,
0268                        lockdep_is_held(&nf_ct_ecache_mutex));
0269     WARN_ON_ONCE(notify);
0270     rcu_assign_pointer(net->ct.nf_conntrack_event_cb, new);
0271     mutex_unlock(&nf_ct_ecache_mutex);
0272 }
0273 EXPORT_SYMBOL_GPL(nf_conntrack_register_notifier);
0274
0275 void nf_conntrack_unregister_notifier(struct net *net)
0276 {
0277     mutex_lock(&nf_ct_ecache_mutex);
0278     RCU_INIT_POINTER(net->ct.nf_conntrack_event_cb, NULL);
0279     mutex_unlock(&nf_ct_ecache_mutex);
0280     /* synchronize_rcu() is called after netns pre_exit */
0281 }
0282 EXPORT_SYMBOL_GPL(nf_conntrack_unregister_notifier);
0283
0284 void nf_conntrack_ecache_work(struct net *net, enum nf_ct_ecache_state state)
0285 {
0286     struct nf_conntrack_net *cnet = nf_ct_pernet(net);
0287
0288     if (state == NFCT_ECACHE_DESTROY_FAIL &&
0289         !delayed_work_pending(&cnet->ecache.dwork)) {
0290         schedule_delayed_work(&cnet->ecache.dwork, HZ);
0291         net->ct.ecache_dwork_pending = true;
0292     } else if (state == NFCT_ECACHE_DESTROY_SENT) {
0293         if (!hlist_nulls_empty(&cnet->ecache.dying_list))
0294             mod_delayed_work(system_wq, &cnet->ecache.dwork, 0);
0295         else
0296             net->ct.ecache_dwork_pending = false;
0297     }
0298 }
0299
0300 bool nf_ct_ecache_ext_add(struct nf_conn *ct, u16 ctmask, u16 expmask, gfp_t gfp)
0301 {
0302     struct net *net = nf_ct_net(ct);
0303     struct nf_conntrack_ecache *e;
0304
0305     switch (net->ct.sysctl_events) {
0306     case 0:
0307          /* assignment via template / ruleset? ignore sysctl. */
0308         if (ctmask || expmask)
0309             break;
0310         return true;
0311     case 2: /* autodetect: no event listener, don't allocate extension. */
0312         if (!READ_ONCE(net->ct.ctnetlink_has_listener))
0313             return true;
0314         fallthrough;
0315     case 1:
0316         /* always allocate an extension. */
0317         if (!ctmask && !expmask) {
0318             ctmask = ~0;
0319             expmask = ~0;
0320         }
0321         break;
0322     default:
0323         WARN_ON_ONCE(1);
0324         return true;
0325     }
0326
0327     e = nf_ct_ext_add(ct, NF_CT_EXT_ECACHE, gfp);
0328     if (e) {
0329         e->ctmask  = ctmask;
0330         e->expmask = expmask;
0331     }
0332
0333     return e != NULL;
0334 }
0335 EXPORT_SYMBOL_GPL(nf_ct_ecache_ext_add);
0336
0337 #define NF_CT_EVENTS_DEFAULT 2
0338 static int nf_ct_events __read_mostly = NF_CT_EVENTS_DEFAULT;
0339
0340 void nf_conntrack_ecache_pernet_init(struct net *net)
0341 {
0342     struct nf_conntrack_net *cnet = nf_ct_pernet(net);
0343
0344     net->ct.sysctl_events = nf_ct_events;
0345
0346     INIT_DELAYED_WORK(&cnet->ecache.dwork, ecache_work);
0347     INIT_HLIST_NULLS_HEAD(&cnet->ecache.dying_list, DYING_NULLS_VAL);
0348     spin_lock_init(&cnet->ecache.dying_lock);
0349
0350     BUILD_BUG_ON(__IPCT_MAX >= 16); /* e->ctmask is u16 */
0351 }
0352
0353 void nf_conntrack_ecache_pernet_fini(struct net *net)
0354 {
0355     struct nf_conntrack_net *cnet = nf_ct_pernet(net);
0356
0357     cancel_delayed_work_sync(&cnet->ecache.dwork);
0358 }