| 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/kernel.h> |
| 20 | #include <linux/netdevice.h> |
| 21 | #include <linux/slab.h> |
| 22 | #include <linux/export.h> |
| 23 | |
| 24 | #include <net/netfilter/nf_conntrack.h> |
| 25 | #include <net/netfilter/nf_conntrack_core.h> |
| 26 | #include <net/netfilter/nf_conntrack_ecache.h> |
| 27 | #include <net/netfilter/nf_conntrack_extend.h> |
| 28 | |
| 29 | static DEFINE_MUTEX(nf_ct_ecache_mutex); |
| 30 | |
| 31 | #define DYING_NULLS_VAL ((1 << 30) + 1) |
| 32 | #define ECACHE_MAX_JIFFIES msecs_to_jiffies(10) |
| 33 | #define ECACHE_RETRY_JIFFIES msecs_to_jiffies(10) |
| 34 | |
| 35 | enum retry_state { |
| 36 | STATE_CONGESTED, |
| 37 | STATE_RESTART, |
| 38 | STATE_DONE, |
| 39 | }; |
| 40 | |
| 41 | struct nf_conntrack_net_ecache *nf_conn_pernet_ecache(const struct net *net) |
| 42 | { |
| 43 | struct nf_conntrack_net *cnet = nf_ct_pernet(net); |
| 44 | |
| 45 | return &cnet->ecache; |
| 46 | } |
| 47 | #if IS_MODULE(CONFIG_NF_CT_NETLINK) |
| 48 | EXPORT_SYMBOL_GPL(nf_conn_pernet_ecache); |
| 49 | #endif |
| 50 | |
| 51 | static enum retry_state ecache_work_evict_list(struct nf_conntrack_net *cnet) |
| 52 | { |
| 53 | unsigned long stop = jiffies + ECACHE_MAX_JIFFIES; |
| 54 | struct hlist_nulls_head evicted_list; |
| 55 | enum retry_state ret = STATE_DONE; |
| 56 | struct nf_conntrack_tuple_hash *h; |
| 57 | struct hlist_nulls_node *n; |
| 58 | unsigned int sent; |
| 59 | |
| 60 | INIT_HLIST_NULLS_HEAD(&evicted_list, DYING_NULLS_VAL); |
| 61 | |
| 62 | next: |
| 63 | sent = 0; |
| 64 | spin_lock_bh(lock: &cnet->ecache.dying_lock); |
| 65 | |
| 66 | hlist_nulls_for_each_entry_safe(h, n, &cnet->ecache.dying_list, hnnode) { |
| 67 | struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(hash: h); |
| 68 | |
| 69 | /* The worker owns all entries, ct remains valid until nf_ct_put |
| 70 | * in the loop below. |
| 71 | */ |
| 72 | if (nf_conntrack_event(event: IPCT_DESTROY, ct)) { |
| 73 | ret = STATE_CONGESTED; |
| 74 | break; |
| 75 | } |
| 76 | |
| 77 | hlist_nulls_del_rcu(n: &ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode); |
| 78 | hlist_nulls_add_head(n: &ct->tuplehash[IP_CT_DIR_REPLY].hnnode, h: &evicted_list); |
| 79 | |
| 80 | if (time_after(stop, jiffies)) { |
| 81 | ret = STATE_RESTART; |
| 82 | break; |
| 83 | } |
| 84 | |
| 85 | if (sent++ > 16) { |
| 86 | spin_unlock_bh(lock: &cnet->ecache.dying_lock); |
| 87 | cond_resched(); |
| 88 | goto next; |
| 89 | } |
| 90 | } |
| 91 | |
| 92 | spin_unlock_bh(lock: &cnet->ecache.dying_lock); |
| 93 | |
| 94 | hlist_nulls_for_each_entry_safe(h, n, &evicted_list, hnnode) { |
| 95 | struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(hash: h); |
| 96 | |
| 97 | hlist_nulls_del_rcu(n: &ct->tuplehash[IP_CT_DIR_REPLY].hnnode); |
| 98 | nf_ct_put(ct); |
| 99 | |
| 100 | cond_resched(); |
| 101 | } |
| 102 | |
| 103 | return ret; |
| 104 | } |
| 105 | |
| 106 | static void ecache_work(struct work_struct *work) |
| 107 | { |
| 108 | struct nf_conntrack_net *cnet = container_of(work, struct nf_conntrack_net, ecache.dwork.work); |
| 109 | int ret, delay = -1; |
| 110 | |
| 111 | ret = ecache_work_evict_list(cnet); |
| 112 | switch (ret) { |
| 113 | case STATE_CONGESTED: |
| 114 | delay = ECACHE_RETRY_JIFFIES; |
| 115 | break; |
| 116 | case STATE_RESTART: |
| 117 | delay = 0; |
| 118 | break; |
| 119 | case STATE_DONE: |
| 120 | break; |
| 121 | } |
| 122 | |
| 123 | if (delay >= 0) |
| 124 | schedule_delayed_work(dwork: &cnet->ecache.dwork, delay); |
| 125 | } |
| 126 | |
| 127 | static int __nf_conntrack_eventmask_report(struct nf_conntrack_ecache *e, |
| 128 | const u32 events, |
| 129 | const u32 missed, |
| 130 | const struct nf_ct_event *item) |
| 131 | { |
| 132 | struct net *net = nf_ct_net(ct: item->ct); |
| 133 | struct nf_ct_event_notifier *notify; |
| 134 | u32 old, want; |
| 135 | int ret; |
| 136 | |
| 137 | if (!((events | missed) & e->ctmask)) |
| 138 | return 0; |
| 139 | |
| 140 | rcu_read_lock(); |
| 141 | |
| 142 | notify = rcu_dereference(net->ct.nf_conntrack_event_cb); |
| 143 | if (!notify) { |
| 144 | rcu_read_unlock(); |
| 145 | return 0; |
| 146 | } |
| 147 | |
| 148 | ret = notify->ct_event(events | missed, item); |
| 149 | rcu_read_unlock(); |
| 150 | |
| 151 | if (likely(ret >= 0 && missed == 0)) |
| 152 | return 0; |
| 153 | |
| 154 | do { |
| 155 | old = READ_ONCE(e->missed); |
| 156 | if (ret < 0) |
| 157 | want = old | events; |
| 158 | else |
| 159 | want = old & ~missed; |
| 160 | } while (cmpxchg(&e->missed, old, want) != old); |
| 161 | |
| 162 | return ret; |
| 163 | } |
| 164 | |
| 165 | static void nf_ct_ecache_tstamp_refresh(struct nf_conntrack_ecache *e) |
| 166 | { |
| 167 | #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP |
| 168 | if (local64_read(&e->timestamp)) |
| 169 | local64_set(&e->timestamp, ktime_get_real_ns()); |
| 170 | #endif |
| 171 | } |
| 172 | |
| 173 | int nf_conntrack_eventmask_report(unsigned int events, struct nf_conn *ct, |
| 174 | u32 portid, int report) |
| 175 | { |
| 176 | struct nf_conntrack_ecache *e; |
| 177 | struct nf_ct_event item; |
| 178 | unsigned int missed; |
| 179 | int ret; |
| 180 | |
| 181 | if (!nf_ct_is_confirmed(ct)) |
| 182 | return 0; |
| 183 | |
| 184 | e = nf_ct_ecache_find(ct); |
| 185 | if (!e) |
| 186 | return 0; |
| 187 | |
| 188 | memset(&item, 0, sizeof(item)); |
| 189 | |
| 190 | item.ct = ct; |
| 191 | item.portid = e->portid ? e->portid : portid; |
| 192 | item.report = report; |
| 193 | |
| 194 | /* This is a resent of a destroy event? If so, skip missed */ |
| 195 | missed = e->portid ? 0 : e->missed; |
| 196 | |
| 197 | nf_ct_ecache_tstamp_refresh(e); |
| 198 | |
| 199 | ret = __nf_conntrack_eventmask_report(e, events, missed, item: &item); |
| 200 | if (unlikely(ret < 0 && (events & (1 << IPCT_DESTROY)))) { |
| 201 | /* This is a destroy event that has been triggered by a process, |
| 202 | * we store the PORTID to include it in the retransmission. |
| 203 | */ |
| 204 | if (e->portid == 0 && portid != 0) |
| 205 | e->portid = portid; |
| 206 | } |
| 207 | |
| 208 | return ret; |
| 209 | } |
| 210 | EXPORT_SYMBOL_GPL(nf_conntrack_eventmask_report); |
| 211 | |
| 212 | /* deliver cached events and clear cache entry - must be called with locally |
| 213 | * disabled softirqs */ |
| 214 | void nf_ct_deliver_cached_events(struct nf_conn *ct) |
| 215 | { |
| 216 | struct nf_conntrack_ecache *e; |
| 217 | struct nf_ct_event item; |
| 218 | unsigned int events; |
| 219 | |
| 220 | if (!nf_ct_is_confirmed(ct) || nf_ct_is_dying(ct)) |
| 221 | return; |
| 222 | |
| 223 | e = nf_ct_ecache_find(ct); |
| 224 | if (e == NULL) |
| 225 | return; |
| 226 | |
| 227 | events = xchg(&e->cache, 0); |
| 228 | |
| 229 | item.ct = ct; |
| 230 | item.portid = 0; |
| 231 | item.report = 0; |
| 232 | |
| 233 | /* We make a copy of the missed event cache without taking |
| 234 | * the lock, thus we may send missed events twice. However, |
| 235 | * this does not harm and it happens very rarely. |
| 236 | */ |
| 237 | __nf_conntrack_eventmask_report(e, events, missed: e->missed, item: &item); |
| 238 | } |
| 239 | EXPORT_SYMBOL_GPL(nf_ct_deliver_cached_events); |
| 240 | |
| 241 | void nf_ct_expect_event_report(enum ip_conntrack_expect_events event, |
| 242 | struct nf_conntrack_expect *exp, |
| 243 | u32 portid, int report) |
| 244 | |
| 245 | { |
| 246 | struct net *net = nf_ct_exp_net(exp); |
| 247 | struct nf_ct_event_notifier *notify; |
| 248 | struct nf_conntrack_ecache *e; |
| 249 | |
| 250 | rcu_read_lock(); |
| 251 | notify = rcu_dereference(net->ct.nf_conntrack_event_cb); |
| 252 | if (!notify) |
| 253 | goto out_unlock; |
| 254 | |
| 255 | e = nf_ct_ecache_find(ct: exp->master); |
| 256 | if (!e) |
| 257 | goto out_unlock; |
| 258 | |
| 259 | if (e->expmask & (1 << event)) { |
| 260 | struct nf_exp_event item = { |
| 261 | .exp = exp, |
| 262 | .portid = portid, |
| 263 | .report = report |
| 264 | }; |
| 265 | notify->exp_event(1 << event, &item); |
| 266 | } |
| 267 | out_unlock: |
| 268 | rcu_read_unlock(); |
| 269 | } |
| 270 | |
| 271 | void nf_conntrack_register_notifier(struct net *net, |
| 272 | const struct nf_ct_event_notifier *new) |
| 273 | { |
| 274 | struct nf_ct_event_notifier *notify; |
| 275 | |
| 276 | mutex_lock(&nf_ct_ecache_mutex); |
| 277 | notify = rcu_dereference_protected(net->ct.nf_conntrack_event_cb, |
| 278 | lockdep_is_held(&nf_ct_ecache_mutex)); |
| 279 | WARN_ON_ONCE(notify); |
| 280 | rcu_assign_pointer(net->ct.nf_conntrack_event_cb, new); |
| 281 | mutex_unlock(lock: &nf_ct_ecache_mutex); |
| 282 | } |
| 283 | EXPORT_SYMBOL_GPL(nf_conntrack_register_notifier); |
| 284 | |
| 285 | void nf_conntrack_unregister_notifier(struct net *net) |
| 286 | { |
| 287 | mutex_lock(&nf_ct_ecache_mutex); |
| 288 | RCU_INIT_POINTER(net->ct.nf_conntrack_event_cb, NULL); |
| 289 | mutex_unlock(lock: &nf_ct_ecache_mutex); |
| 290 | /* synchronize_rcu() is called after netns pre_exit */ |
| 291 | } |
| 292 | EXPORT_SYMBOL_GPL(nf_conntrack_unregister_notifier); |
| 293 | |
| 294 | void nf_conntrack_ecache_work(struct net *net, enum nf_ct_ecache_state state) |
| 295 | { |
| 296 | struct nf_conntrack_net *cnet = nf_ct_pernet(net); |
| 297 | |
| 298 | if (state == NFCT_ECACHE_DESTROY_FAIL && |
| 299 | !delayed_work_pending(&cnet->ecache.dwork)) { |
| 300 | schedule_delayed_work(dwork: &cnet->ecache.dwork, HZ); |
| 301 | net->ct.ecache_dwork_pending = true; |
| 302 | } else if (state == NFCT_ECACHE_DESTROY_SENT) { |
| 303 | if (!hlist_nulls_empty(h: &cnet->ecache.dying_list)) |
| 304 | mod_delayed_work(wq: system_percpu_wq, dwork: &cnet->ecache.dwork, delay: 0); |
| 305 | else |
| 306 | net->ct.ecache_dwork_pending = false; |
| 307 | } |
| 308 | } |
| 309 | |
| 310 | static void nf_ct_ecache_tstamp_new(const struct nf_conn *ct, struct nf_conntrack_ecache *e) |
| 311 | { |
| 312 | #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP |
| 313 | u64 ts = 0; |
| 314 | |
| 315 | if (nf_ct_ext_exist(ct, id: NF_CT_EXT_TSTAMP)) |
| 316 | ts = ktime_get_real_ns(); |
| 317 | |
| 318 | local64_set(&e->timestamp, ts); |
| 319 | #endif |
| 320 | } |
| 321 | |
| 322 | bool nf_ct_ecache_ext_add(struct nf_conn *ct, u16 ctmask, u16 expmask, gfp_t gfp) |
| 323 | { |
| 324 | struct net *net = nf_ct_net(ct); |
| 325 | struct nf_conntrack_ecache *e; |
| 326 | |
| 327 | switch (net->ct.sysctl_events) { |
| 328 | case 0: |
| 329 | /* assignment via template / ruleset? ignore sysctl. */ |
| 330 | if (ctmask || expmask) |
| 331 | break; |
| 332 | return true; |
| 333 | case 2: /* autodetect: no event listener, don't allocate extension. */ |
| 334 | if (!READ_ONCE(nf_ctnetlink_has_listener)) |
| 335 | return true; |
| 336 | fallthrough; |
| 337 | case 1: |
| 338 | /* always allocate an extension. */ |
| 339 | if (!ctmask && !expmask) { |
| 340 | ctmask = ~0; |
| 341 | expmask = ~0; |
| 342 | } |
| 343 | break; |
| 344 | default: |
| 345 | WARN_ON_ONCE(1); |
| 346 | return true; |
| 347 | } |
| 348 | |
| 349 | e = nf_ct_ext_add(ct, id: NF_CT_EXT_ECACHE, gfp); |
| 350 | if (e) { |
| 351 | nf_ct_ecache_tstamp_new(ct, e); |
| 352 | e->ctmask = ctmask; |
| 353 | e->expmask = expmask; |
| 354 | } |
| 355 | |
| 356 | return e != NULL; |
| 357 | } |
| 358 | EXPORT_SYMBOL_GPL(nf_ct_ecache_ext_add); |
| 359 | |
| 360 | #define NF_CT_EVENTS_DEFAULT 2 |
| 361 | static int nf_ct_events __read_mostly = NF_CT_EVENTS_DEFAULT; |
| 362 | |
| 363 | void nf_conntrack_ecache_pernet_init(struct net *net) |
| 364 | { |
| 365 | struct nf_conntrack_net *cnet = nf_ct_pernet(net); |
| 366 | |
| 367 | net->ct.sysctl_events = nf_ct_events; |
| 368 | |
| 369 | INIT_DELAYED_WORK(&cnet->ecache.dwork, ecache_work); |
| 370 | INIT_HLIST_NULLS_HEAD(&cnet->ecache.dying_list, DYING_NULLS_VAL); |
| 371 | spin_lock_init(&cnet->ecache.dying_lock); |
| 372 | |
| 373 | BUILD_BUG_ON(__IPCT_MAX >= 16); /* e->ctmask is u16 */ |
| 374 | } |
| 375 | |
| 376 | void nf_conntrack_ecache_pernet_fini(struct net *net) |
| 377 | { |
| 378 | struct nf_conntrack_net *cnet = nf_ct_pernet(net); |
| 379 | |
| 380 | cancel_delayed_work_sync(dwork: &cnet->ecache.dwork); |
| 381 | } |
| 382 | |