xref: /linux/net/ipv4/inet_timewait_sock.c (revision 07fdad3a93756b872da7b53647715c48d0f4a2d0)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * INET		An implementation of the TCP/IP protocol suite for the LINUX
4  *		operating system.  INET is implemented using the  BSD Socket
5  *		interface as the means of communication with the user level.
6  *
7  *		Generic TIME_WAIT sockets functions
8  *
9  *		From code orinally in TCP
10  */
11 
12 #include <linux/kernel.h>
13 #include <linux/slab.h>
14 #include <linux/module.h>
15 #include <net/inet_hashtables.h>
16 #include <net/inet_timewait_sock.h>
17 #include <net/ip.h>
18 #include <net/tcp.h>
19 #include <net/psp.h>
20 
21 /**
22  *	inet_twsk_bind_unhash - unhash a timewait socket from bind hash
23  *	@tw: timewait socket
24  *	@hashinfo: hashinfo pointer
25  *
26  *	unhash a timewait socket from bind hash, if hashed.
27  *	bind hash lock must be held by caller.
28  *	Returns 1 if caller should call inet_twsk_put() after lock release.
29  */
30 void inet_twsk_bind_unhash(struct inet_timewait_sock *tw,
31 			  struct inet_hashinfo *hashinfo)
32 {
33 	struct inet_bind2_bucket *tb2 = tw->tw_tb2;
34 	struct inet_bind_bucket *tb = tw->tw_tb;
35 
36 	if (!tb)
37 		return;
38 
39 	__sk_del_bind_node((struct sock *)tw);
40 	tw->tw_tb = NULL;
41 	tw->tw_tb2 = NULL;
42 	inet_bind2_bucket_destroy(hashinfo->bind2_bucket_cachep, tb2);
43 	inet_bind_bucket_destroy(tb);
44 
45 	__sock_put((struct sock *)tw);
46 }
47 
48 /* Must be called with locally disabled BHs. */
49 static void inet_twsk_kill(struct inet_timewait_sock *tw)
50 {
51 	struct inet_hashinfo *hashinfo = tw->tw_dr->hashinfo;
52 	spinlock_t *lock = inet_ehash_lockp(hashinfo, tw->tw_hash);
53 	struct inet_bind_hashbucket *bhead, *bhead2;
54 
55 	spin_lock(lock);
56 	sk_nulls_del_node_init_rcu((struct sock *)tw);
57 	spin_unlock(lock);
58 
59 	/* Disassociate with bind bucket. */
60 	bhead = &hashinfo->bhash[inet_bhashfn(twsk_net(tw), tw->tw_num,
61 			hashinfo->bhash_size)];
62 	bhead2 = inet_bhashfn_portaddr(hashinfo, (struct sock *)tw,
63 				       twsk_net(tw), tw->tw_num);
64 
65 	spin_lock(&bhead->lock);
66 	spin_lock(&bhead2->lock);
67 	inet_twsk_bind_unhash(tw, hashinfo);
68 	spin_unlock(&bhead2->lock);
69 	spin_unlock(&bhead->lock);
70 
71 	refcount_dec(&tw->tw_dr->tw_refcount);
72 	inet_twsk_put(tw);
73 }
74 
75 void inet_twsk_free(struct inet_timewait_sock *tw)
76 {
77 	struct module *owner = tw->tw_prot->owner;
78 
79 	tcp_twsk_destructor((struct sock *)tw);
80 	kmem_cache_free(tw->tw_prot->twsk_prot->twsk_slab, tw);
81 	module_put(owner);
82 }
83 
84 void inet_twsk_put(struct inet_timewait_sock *tw)
85 {
86 	if (refcount_dec_and_test(&tw->tw_refcnt))
87 		inet_twsk_free(tw);
88 }
89 EXPORT_SYMBOL_GPL(inet_twsk_put);
90 
91 static void inet_twsk_add_node_rcu(struct inet_timewait_sock *tw,
92 				   struct hlist_nulls_head *list)
93 {
94 	hlist_nulls_add_head_rcu(&tw->tw_node, list);
95 }
96 
97 static void inet_twsk_schedule(struct inet_timewait_sock *tw, int timeo)
98 {
99 	__inet_twsk_schedule(tw, timeo, false);
100 }
101 
102 /*
103  * Enter the time wait state.
104  * Essentially we whip up a timewait bucket, copy the relevant info into it
105  * from the SK, and mess with hash chains and list linkage.
106  *
107  * The caller must not access @tw anymore after this function returns.
108  */
109 void inet_twsk_hashdance_schedule(struct inet_timewait_sock *tw,
110 				  struct sock *sk,
111 				  struct inet_hashinfo *hashinfo,
112 				  int timeo)
113 {
114 	const struct inet_sock *inet = inet_sk(sk);
115 	const struct inet_connection_sock *icsk = inet_csk(sk);
116 	struct inet_ehash_bucket *ehead = inet_ehash_bucket(hashinfo, sk->sk_hash);
117 	spinlock_t *lock = inet_ehash_lockp(hashinfo, sk->sk_hash);
118 	struct inet_bind_hashbucket *bhead, *bhead2;
119 
120 	/* Step 1: Put TW into bind hash. Original socket stays there too.
121 	   Note, that any socket with inet->num != 0 MUST be bound in
122 	   binding cache, even if it is closed.
123 	 */
124 	bhead = &hashinfo->bhash[inet_bhashfn(twsk_net(tw), inet->inet_num,
125 			hashinfo->bhash_size)];
126 	bhead2 = inet_bhashfn_portaddr(hashinfo, sk, twsk_net(tw), inet->inet_num);
127 
128 	local_bh_disable();
129 	spin_lock(&bhead->lock);
130 	spin_lock(&bhead2->lock);
131 
132 	tw->tw_tb = icsk->icsk_bind_hash;
133 	WARN_ON(!icsk->icsk_bind_hash);
134 
135 	tw->tw_tb2 = icsk->icsk_bind2_hash;
136 	WARN_ON(!icsk->icsk_bind2_hash);
137 	sk_add_bind_node((struct sock *)tw, &tw->tw_tb2->owners);
138 
139 	spin_unlock(&bhead2->lock);
140 	spin_unlock(&bhead->lock);
141 
142 	spin_lock(lock);
143 
144 	/* Step 2: Hash TW into tcp ehash chain */
145 	inet_twsk_add_node_rcu(tw, &ehead->chain);
146 
147 	/* Step 3: Remove SK from hash chain */
148 	if (__sk_nulls_del_node_init_rcu(sk))
149 		sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
150 
151 
152 	/* Ensure above writes are committed into memory before updating the
153 	 * refcount.
154 	 * Provides ordering vs later refcount_inc().
155 	 */
156 	smp_wmb();
157 	/* tw_refcnt is set to 3 because we have :
158 	 * - one reference for bhash chain.
159 	 * - one reference for ehash chain.
160 	 * - one reference for timer.
161 	 * Also note that after this point, we lost our implicit reference
162 	 * so we are not allowed to use tw anymore.
163 	 */
164 	refcount_set(&tw->tw_refcnt, 3);
165 
166 	inet_twsk_schedule(tw, timeo);
167 
168 	spin_unlock(lock);
169 	local_bh_enable();
170 }
171 
172 static void tw_timer_handler(struct timer_list *t)
173 {
174 	struct inet_timewait_sock *tw = timer_container_of(tw, t, tw_timer);
175 
176 	inet_twsk_kill(tw);
177 }
178 
179 struct inet_timewait_sock *inet_twsk_alloc(const struct sock *sk,
180 					   struct inet_timewait_death_row *dr,
181 					   const int state)
182 {
183 	struct inet_timewait_sock *tw;
184 
185 	if (refcount_read(&dr->tw_refcount) - 1 >=
186 	    READ_ONCE(dr->sysctl_max_tw_buckets))
187 		return NULL;
188 
189 	tw = kmem_cache_alloc(sk->sk_prot_creator->twsk_prot->twsk_slab,
190 			      GFP_ATOMIC);
191 	if (tw) {
192 		const struct inet_sock *inet = inet_sk(sk);
193 
194 		tw->tw_dr	    = dr;
195 		/* Give us an identity. */
196 		tw->tw_daddr	    = inet->inet_daddr;
197 		tw->tw_rcv_saddr    = inet->inet_rcv_saddr;
198 		tw->tw_bound_dev_if = sk->sk_bound_dev_if;
199 		tw->tw_tos	    = inet->tos;
200 		tw->tw_num	    = inet->inet_num;
201 		tw->tw_state	    = TCP_TIME_WAIT;
202 		tw->tw_substate	    = state;
203 		tw->tw_sport	    = inet->inet_sport;
204 		tw->tw_dport	    = inet->inet_dport;
205 		tw->tw_family	    = sk->sk_family;
206 		tw->tw_reuse	    = sk->sk_reuse;
207 		tw->tw_reuseport    = sk->sk_reuseport;
208 		tw->tw_hash	    = sk->sk_hash;
209 		tw->tw_ipv6only	    = 0;
210 		tw->tw_transparent  = inet_test_bit(TRANSPARENT, sk);
211 		tw->tw_connect_bind = !!(sk->sk_userlocks & SOCK_CONNECT_BIND);
212 		tw->tw_prot	    = sk->sk_prot_creator;
213 		atomic64_set(&tw->tw_cookie, atomic64_read(&sk->sk_cookie));
214 		twsk_net_set(tw, sock_net(sk));
215 		timer_setup(&tw->tw_timer, tw_timer_handler, 0);
216 #ifdef CONFIG_SOCK_VALIDATE_XMIT
217 		tw->tw_validate_xmit_skb = NULL;
218 #endif
219 		/*
220 		 * Because we use RCU lookups, we should not set tw_refcnt
221 		 * to a non null value before everything is setup for this
222 		 * timewait socket.
223 		 */
224 		refcount_set(&tw->tw_refcnt, 0);
225 
226 		__module_get(tw->tw_prot->owner);
227 		psp_twsk_init(tw, sk);
228 	}
229 
230 	return tw;
231 }
232 
233 /* These are always called from BH context.  See callers in
234  * tcp_input.c to verify this.
235  */
236 
237 /* This is for handling early-kills of TIME_WAIT sockets.
238  * Warning : consume reference.
239  * Caller should not access tw anymore.
240  */
241 void inet_twsk_deschedule_put(struct inet_timewait_sock *tw)
242 {
243 	struct inet_hashinfo *hashinfo = tw->tw_dr->hashinfo;
244 	spinlock_t *lock = inet_ehash_lockp(hashinfo, tw->tw_hash);
245 
246 	/* inet_twsk_purge() walks over all sockets, including tw ones,
247 	 * and removes them via inet_twsk_deschedule_put() after a
248 	 * refcount_inc_not_zero().
249 	 *
250 	 * inet_twsk_hashdance_schedule() must (re)init the refcount before
251 	 * arming the timer, i.e. inet_twsk_purge can obtain a reference to
252 	 * a twsk that did not yet schedule the timer.
253 	 *
254 	 * The ehash lock synchronizes these two:
255 	 * After acquiring the lock, the timer is always scheduled (else
256 	 * timer_shutdown returns false), because hashdance_schedule releases
257 	 * the ehash lock only after completing the timer initialization.
258 	 *
259 	 * Without grabbing the ehash lock, we get:
260 	 * 1) cpu x sets twsk refcount to 3
261 	 * 2) cpu y bumps refcount to 4
262 	 * 3) cpu y calls inet_twsk_deschedule_put() and shuts timer down
263 	 * 4) cpu x tries to start timer, but mod_timer is a noop post-shutdown
264 	 * -> timer refcount is never decremented.
265 	 */
266 	spin_lock(lock);
267 	/*  Makes sure hashdance_schedule() has completed */
268 	spin_unlock(lock);
269 
270 	if (timer_shutdown_sync(&tw->tw_timer))
271 		inet_twsk_kill(tw);
272 	inet_twsk_put(tw);
273 }
274 EXPORT_SYMBOL(inet_twsk_deschedule_put);
275 
276 void __inet_twsk_schedule(struct inet_timewait_sock *tw, int timeo, bool rearm)
277 {
278 	/* timeout := RTO * 3.5
279 	 *
280 	 * 3.5 = 1+2+0.5 to wait for two retransmits.
281 	 *
282 	 * RATIONALE: if FIN arrived and we entered TIME-WAIT state,
283 	 * our ACK acking that FIN can be lost. If N subsequent retransmitted
284 	 * FINs (or previous seqments) are lost (probability of such event
285 	 * is p^(N+1), where p is probability to lose single packet and
286 	 * time to detect the loss is about RTO*(2^N - 1) with exponential
287 	 * backoff). Normal timewait length is calculated so, that we
288 	 * waited at least for one retransmitted FIN (maximal RTO is 120sec).
289 	 * [ BTW Linux. following BSD, violates this requirement waiting
290 	 *   only for 60sec, we should wait at least for 240 secs.
291 	 *   Well, 240 consumes too much of resources 8)
292 	 * ]
293 	 * This interval is not reduced to catch old duplicate and
294 	 * responces to our wandering segments living for two MSLs.
295 	 * However, if we use PAWS to detect
296 	 * old duplicates, we can reduce the interval to bounds required
297 	 * by RTO, rather than MSL. So, if peer understands PAWS, we
298 	 * kill tw bucket after 3.5*RTO (it is important that this number
299 	 * is greater than TS tick!) and detect old duplicates with help
300 	 * of PAWS.
301 	 */
302 
303 	if (!rearm) {
304 		bool kill = timeo <= 4*HZ;
305 
306 		__NET_INC_STATS(twsk_net(tw), kill ? LINUX_MIB_TIMEWAITKILLED :
307 						     LINUX_MIB_TIMEWAITED);
308 		BUG_ON(mod_timer(&tw->tw_timer, jiffies + timeo));
309 		refcount_inc(&tw->tw_dr->tw_refcount);
310 	} else {
311 		mod_timer_pending(&tw->tw_timer, jiffies + timeo);
312 	}
313 }
314 
315 /* Remove all non full sockets (TIME_WAIT and NEW_SYN_RECV) for dead netns */
316 void inet_twsk_purge(struct inet_hashinfo *hashinfo)
317 {
318 	struct inet_ehash_bucket *head = &hashinfo->ehash[0];
319 	unsigned int ehash_mask = hashinfo->ehash_mask;
320 	struct hlist_nulls_node *node;
321 	unsigned int slot;
322 	struct sock *sk;
323 
324 	for (slot = 0; slot <= ehash_mask; slot++, head++) {
325 		if (hlist_nulls_empty(&head->chain))
326 			continue;
327 
328 restart_rcu:
329 		cond_resched();
330 		rcu_read_lock();
331 restart:
332 		sk_nulls_for_each_rcu(sk, node, &head->chain) {
333 			int state = inet_sk_state_load(sk);
334 
335 			if ((1 << state) & ~(TCPF_TIME_WAIT |
336 					     TCPF_NEW_SYN_RECV))
337 				continue;
338 
339 			if (check_net(sock_net(sk)))
340 				continue;
341 
342 			if (unlikely(!refcount_inc_not_zero(&sk->sk_refcnt)))
343 				continue;
344 
345 			if (check_net(sock_net(sk))) {
346 				sock_gen_put(sk);
347 				goto restart;
348 			}
349 
350 			rcu_read_unlock();
351 			local_bh_disable();
352 			if (state == TCP_TIME_WAIT) {
353 				inet_twsk_deschedule_put(inet_twsk(sk));
354 			} else {
355 				struct request_sock *req = inet_reqsk(sk);
356 
357 				inet_csk_reqsk_queue_drop_and_put(req->rsk_listener,
358 								  req);
359 			}
360 			local_bh_enable();
361 			goto restart_rcu;
362 		}
363 		/* If the nulls value we got at the end of this lookup is
364 		 * not the expected one, we must restart lookup.
365 		 * We probably met an item that was moved to another chain.
366 		 */
367 		if (get_nulls_value(node) != slot)
368 			goto restart;
369 		rcu_read_unlock();
370 	}
371 }
372