Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Pluggable TCP congestion control support and newReno
4 * congestion control.
5 * Based on ideas from I/O scheduler support and Web100.
6 *
7 * Copyright (C) 2005 Stephen Hemminger <shemminger@osdl.org>
8 */
9
10#define pr_fmt(fmt) "TCP: " fmt
11
12#include <linux/module.h>
13#include <linux/mm.h>
14#include <linux/types.h>
15#include <linux/list.h>
16#include <linux/gfp.h>
17#include <linux/jhash.h>
18#include <net/tcp.h>
19
20static DEFINE_SPINLOCK(tcp_cong_list_lock);
21static LIST_HEAD(tcp_cong_list);
22
23/* Simple linear search, don't expect many entries! */
24struct tcp_congestion_ops *tcp_ca_find(const char *name)
25{
26 struct tcp_congestion_ops *e;
27
28 list_for_each_entry_rcu(e, &tcp_cong_list, list) {
29 if (strcmp(e->name, name) == 0)
30 return e;
31 }
32
33 return NULL;
34}
35
36/* Must be called with rcu lock held */
37static struct tcp_congestion_ops *tcp_ca_find_autoload(struct net *net,
38 const char *name)
39{
40 struct tcp_congestion_ops *ca = tcp_ca_find(name);
41
42#ifdef CONFIG_MODULES
43 if (!ca && capable(CAP_NET_ADMIN)) {
44 rcu_read_unlock();
45 request_module("tcp_%s", name);
46 rcu_read_lock();
47 ca = tcp_ca_find(name);
48 }
49#endif
50 return ca;
51}
52
53/* Simple linear search, not much in here. */
54struct tcp_congestion_ops *tcp_ca_find_key(u32 key)
55{
56 struct tcp_congestion_ops *e;
57
58 list_for_each_entry_rcu(e, &tcp_cong_list, list) {
59 if (e->key == key)
60 return e;
61 }
62
63 return NULL;
64}
65
66/*
67 * Attach new congestion control algorithm to the list
68 * of available options.
69 */
70int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
71{
72 int ret = 0;
73
74 /* all algorithms must implement these */
75 if (!ca->ssthresh || !ca->undo_cwnd ||
76 !(ca->cong_avoid || ca->cong_control)) {
77 pr_err("%s does not implement required ops\n", ca->name);
78 return -EINVAL;
79 }
80
81 ca->key = jhash(ca->name, sizeof(ca->name), strlen(ca->name));
82
83 spin_lock(&tcp_cong_list_lock);
84 if (ca->key == TCP_CA_UNSPEC || tcp_ca_find_key(ca->key)) {
85 pr_notice("%s already registered or non-unique key\n",
86 ca->name);
87 ret = -EEXIST;
88 } else {
89 list_add_tail_rcu(&ca->list, &tcp_cong_list);
90 pr_debug("%s registered\n", ca->name);
91 }
92 spin_unlock(&tcp_cong_list_lock);
93
94 return ret;
95}
96EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
97
98/*
99 * Remove congestion control algorithm, called from
100 * the module's remove function. Module ref counts are used
101 * to ensure that this can't be done till all sockets using
102 * that method are closed.
103 */
104void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
105{
106 spin_lock(&tcp_cong_list_lock);
107 list_del_rcu(&ca->list);
108 spin_unlock(&tcp_cong_list_lock);
109
110 /* Wait for outstanding readers to complete before the
111 * module gets removed entirely.
112 *
113 * A try_module_get() should fail by now as our module is
114 * in "going" state since no refs are held anymore and
115 * module_exit() handler being called.
116 */
117 synchronize_rcu();
118}
119EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
120
121u32 tcp_ca_get_key_by_name(struct net *net, const char *name, bool *ecn_ca)
122{
123 const struct tcp_congestion_ops *ca;
124 u32 key = TCP_CA_UNSPEC;
125
126 might_sleep();
127
128 rcu_read_lock();
129 ca = tcp_ca_find_autoload(net, name);
130 if (ca) {
131 key = ca->key;
132 *ecn_ca = ca->flags & TCP_CONG_NEEDS_ECN;
133 }
134 rcu_read_unlock();
135
136 return key;
137}
138EXPORT_SYMBOL_GPL(tcp_ca_get_key_by_name);
139
140char *tcp_ca_get_name_by_key(u32 key, char *buffer)
141{
142 const struct tcp_congestion_ops *ca;
143 char *ret = NULL;
144
145 rcu_read_lock();
146 ca = tcp_ca_find_key(key);
147 if (ca)
148 ret = strncpy(buffer, ca->name,
149 TCP_CA_NAME_MAX);
150 rcu_read_unlock();
151
152 return ret;
153}
154EXPORT_SYMBOL_GPL(tcp_ca_get_name_by_key);
155
156/* Assign choice of congestion control. */
157void tcp_assign_congestion_control(struct sock *sk)
158{
159 struct net *net = sock_net(sk);
160 struct inet_connection_sock *icsk = inet_csk(sk);
161 const struct tcp_congestion_ops *ca;
162
163 rcu_read_lock();
164 ca = rcu_dereference(net->ipv4.tcp_congestion_control);
165 if (unlikely(!bpf_try_module_get(ca, ca->owner)))
166 ca = &tcp_reno;
167 icsk->icsk_ca_ops = ca;
168 rcu_read_unlock();
169
170 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
171 if (ca->flags & TCP_CONG_NEEDS_ECN)
172 INET_ECN_xmit(sk);
173 else
174 INET_ECN_dontxmit(sk);
175}
176
177void tcp_init_congestion_control(struct sock *sk)
178{
179 struct inet_connection_sock *icsk = inet_csk(sk);
180
181 tcp_sk(sk)->prior_ssthresh = 0;
182 if (icsk->icsk_ca_ops->init)
183 icsk->icsk_ca_ops->init(sk);
184 if (tcp_ca_needs_ecn(sk))
185 INET_ECN_xmit(sk);
186 else
187 INET_ECN_dontxmit(sk);
188 icsk->icsk_ca_initialized = 1;
189}
190
191static void tcp_reinit_congestion_control(struct sock *sk,
192 const struct tcp_congestion_ops *ca)
193{
194 struct inet_connection_sock *icsk = inet_csk(sk);
195
196 tcp_cleanup_congestion_control(sk);
197 icsk->icsk_ca_ops = ca;
198 icsk->icsk_ca_setsockopt = 1;
199 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
200
201 if (ca->flags & TCP_CONG_NEEDS_ECN)
202 INET_ECN_xmit(sk);
203 else
204 INET_ECN_dontxmit(sk);
205
206 if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
207 tcp_init_congestion_control(sk);
208}
209
210/* Manage refcounts on socket close. */
211void tcp_cleanup_congestion_control(struct sock *sk)
212{
213 struct inet_connection_sock *icsk = inet_csk(sk);
214
215 if (icsk->icsk_ca_ops->release)
216 icsk->icsk_ca_ops->release(sk);
217 bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
218}
219
220/* Used by sysctl to change default congestion control */
221int tcp_set_default_congestion_control(struct net *net, const char *name)
222{
223 struct tcp_congestion_ops *ca;
224 const struct tcp_congestion_ops *prev;
225 int ret;
226
227 rcu_read_lock();
228 ca = tcp_ca_find_autoload(net, name);
229 if (!ca) {
230 ret = -ENOENT;
231 } else if (!bpf_try_module_get(ca, ca->owner)) {
232 ret = -EBUSY;
233 } else if (!net_eq(net, &init_net) &&
234 !(ca->flags & TCP_CONG_NON_RESTRICTED)) {
235 /* Only init netns can set default to a restricted algorithm */
236 ret = -EPERM;
237 } else {
238 prev = xchg(&net->ipv4.tcp_congestion_control, ca);
239 if (prev)
240 bpf_module_put(prev, prev->owner);
241
242 ca->flags |= TCP_CONG_NON_RESTRICTED;
243 ret = 0;
244 }
245 rcu_read_unlock();
246
247 return ret;
248}
249
250/* Set default value from kernel configuration at bootup */
251static int __init tcp_congestion_default(void)
252{
253 return tcp_set_default_congestion_control(&init_net,
254 CONFIG_DEFAULT_TCP_CONG);
255}
256late_initcall(tcp_congestion_default);
257
258/* Build string with list of available congestion control values */
259void tcp_get_available_congestion_control(char *buf, size_t maxlen)
260{
261 struct tcp_congestion_ops *ca;
262 size_t offs = 0;
263
264 rcu_read_lock();
265 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
266 offs += snprintf(buf + offs, maxlen - offs,
267 "%s%s",
268 offs == 0 ? "" : " ", ca->name);
269
270 if (WARN_ON_ONCE(offs >= maxlen))
271 break;
272 }
273 rcu_read_unlock();
274}
275
276/* Get current default congestion control */
277void tcp_get_default_congestion_control(struct net *net, char *name)
278{
279 const struct tcp_congestion_ops *ca;
280
281 rcu_read_lock();
282 ca = rcu_dereference(net->ipv4.tcp_congestion_control);
283 strncpy(name, ca->name, TCP_CA_NAME_MAX);
284 rcu_read_unlock();
285}
286
287/* Built list of non-restricted congestion control values */
288void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
289{
290 struct tcp_congestion_ops *ca;
291 size_t offs = 0;
292
293 *buf = '\0';
294 rcu_read_lock();
295 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
296 if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
297 continue;
298 offs += snprintf(buf + offs, maxlen - offs,
299 "%s%s",
300 offs == 0 ? "" : " ", ca->name);
301
302 if (WARN_ON_ONCE(offs >= maxlen))
303 break;
304 }
305 rcu_read_unlock();
306}
307
308/* Change list of non-restricted congestion control */
309int tcp_set_allowed_congestion_control(char *val)
310{
311 struct tcp_congestion_ops *ca;
312 char *saved_clone, *clone, *name;
313 int ret = 0;
314
315 saved_clone = clone = kstrdup(val, GFP_USER);
316 if (!clone)
317 return -ENOMEM;
318
319 spin_lock(&tcp_cong_list_lock);
320 /* pass 1 check for bad entries */
321 while ((name = strsep(&clone, " ")) && *name) {
322 ca = tcp_ca_find(name);
323 if (!ca) {
324 ret = -ENOENT;
325 goto out;
326 }
327 }
328
329 /* pass 2 clear old values */
330 list_for_each_entry_rcu(ca, &tcp_cong_list, list)
331 ca->flags &= ~TCP_CONG_NON_RESTRICTED;
332
333 /* pass 3 mark as allowed */
334 while ((name = strsep(&val, " ")) && *name) {
335 ca = tcp_ca_find(name);
336 WARN_ON(!ca);
337 if (ca)
338 ca->flags |= TCP_CONG_NON_RESTRICTED;
339 }
340out:
341 spin_unlock(&tcp_cong_list_lock);
342 kfree(saved_clone);
343
344 return ret;
345}
346
347/* Change congestion control for socket. If load is false, then it is the
348 * responsibility of the caller to call tcp_init_congestion_control or
349 * tcp_reinit_congestion_control (if the current congestion control was
350 * already initialized.
351 */
352int tcp_set_congestion_control(struct sock *sk, const char *name, bool load,
353 bool cap_net_admin)
354{
355 struct inet_connection_sock *icsk = inet_csk(sk);
356 const struct tcp_congestion_ops *ca;
357 int err = 0;
358
359 if (icsk->icsk_ca_dst_locked)
360 return -EPERM;
361
362 rcu_read_lock();
363 if (!load)
364 ca = tcp_ca_find(name);
365 else
366 ca = tcp_ca_find_autoload(sock_net(sk), name);
367
368 /* No change asking for existing value */
369 if (ca == icsk->icsk_ca_ops) {
370 icsk->icsk_ca_setsockopt = 1;
371 goto out;
372 }
373
374 if (!ca)
375 err = -ENOENT;
376 else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || cap_net_admin))
377 err = -EPERM;
378 else if (!bpf_try_module_get(ca, ca->owner))
379 err = -EBUSY;
380 else
381 tcp_reinit_congestion_control(sk, ca);
382 out:
383 rcu_read_unlock();
384 return err;
385}
386
387/* Slow start is used when congestion window is no greater than the slow start
388 * threshold. We base on RFC2581 and also handle stretch ACKs properly.
389 * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
390 * something better;) a packet is only considered (s)acked in its entirety to
391 * defend the ACK attacks described in the RFC. Slow start processes a stretch
392 * ACK of degree N as if N acks of degree 1 are received back to back except
393 * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
394 * returns the leftover acks to adjust cwnd in congestion avoidance mode.
395 */
396u32 tcp_slow_start(struct tcp_sock *tp, u32 acked)
397{
398 u32 cwnd = min(tp->snd_cwnd + acked, tp->snd_ssthresh);
399
400 acked -= cwnd - tp->snd_cwnd;
401 tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);
402
403 return acked;
404}
405EXPORT_SYMBOL_GPL(tcp_slow_start);
406
407/* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w),
408 * for every packet that was ACKed.
409 */
410void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked)
411{
412 /* If credits accumulated at a higher w, apply them gently now. */
413 if (tp->snd_cwnd_cnt >= w) {
414 tp->snd_cwnd_cnt = 0;
415 tp->snd_cwnd++;
416 }
417
418 tp->snd_cwnd_cnt += acked;
419 if (tp->snd_cwnd_cnt >= w) {
420 u32 delta = tp->snd_cwnd_cnt / w;
421
422 tp->snd_cwnd_cnt -= delta * w;
423 tp->snd_cwnd += delta;
424 }
425 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_cwnd_clamp);
426}
427EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
428
429/*
430 * TCP Reno congestion control
431 * This is special case used for fallback as well.
432 */
433/* This is Jacobson's slow start and congestion avoidance.
434 * SIGCOMM '88, p. 328.
435 */
436void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
437{
438 struct tcp_sock *tp = tcp_sk(sk);
439
440 if (!tcp_is_cwnd_limited(sk))
441 return;
442
443 /* In "safe" area, increase. */
444 if (tcp_in_slow_start(tp)) {
445 acked = tcp_slow_start(tp, acked);
446 if (!acked)
447 return;
448 }
449 /* In dangerous area, increase slowly. */
450 tcp_cong_avoid_ai(tp, tp->snd_cwnd, acked);
451}
452EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
453
454/* Slow start threshold is half the congestion window (min 2) */
455u32 tcp_reno_ssthresh(struct sock *sk)
456{
457 const struct tcp_sock *tp = tcp_sk(sk);
458
459 return max(tp->snd_cwnd >> 1U, 2U);
460}
461EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
462
463u32 tcp_reno_undo_cwnd(struct sock *sk)
464{
465 const struct tcp_sock *tp = tcp_sk(sk);
466
467 return max(tp->snd_cwnd, tp->prior_cwnd);
468}
469EXPORT_SYMBOL_GPL(tcp_reno_undo_cwnd);
470
471struct tcp_congestion_ops tcp_reno = {
472 .flags = TCP_CONG_NON_RESTRICTED,
473 .name = "reno",
474 .owner = THIS_MODULE,
475 .ssthresh = tcp_reno_ssthresh,
476 .cong_avoid = tcp_reno_cong_avoid,
477 .undo_cwnd = tcp_reno_undo_cwnd,
478};
1/*
2 * Pluggable TCP congestion control support and newReno
3 * congestion control.
4 * Based on ideas from I/O scheduler support and Web100.
5 *
6 * Copyright (C) 2005 Stephen Hemminger <shemminger@osdl.org>
7 */
8
9#define pr_fmt(fmt) "TCP: " fmt
10
11#include <linux/module.h>
12#include <linux/mm.h>
13#include <linux/types.h>
14#include <linux/list.h>
15#include <linux/gfp.h>
16#include <linux/jhash.h>
17#include <net/tcp.h>
18
19static DEFINE_SPINLOCK(tcp_cong_list_lock);
20static LIST_HEAD(tcp_cong_list);
21
22/* Simple linear search, don't expect many entries! */
23static struct tcp_congestion_ops *tcp_ca_find(const char *name)
24{
25 struct tcp_congestion_ops *e;
26
27 list_for_each_entry_rcu(e, &tcp_cong_list, list) {
28 if (strcmp(e->name, name) == 0)
29 return e;
30 }
31
32 return NULL;
33}
34
35/* Must be called with rcu lock held */
36static const struct tcp_congestion_ops *__tcp_ca_find_autoload(const char *name)
37{
38 const struct tcp_congestion_ops *ca = tcp_ca_find(name);
39#ifdef CONFIG_MODULES
40 if (!ca && capable(CAP_NET_ADMIN)) {
41 rcu_read_unlock();
42 request_module("tcp_%s", name);
43 rcu_read_lock();
44 ca = tcp_ca_find(name);
45 }
46#endif
47 return ca;
48}
49
50/* Simple linear search, not much in here. */
51struct tcp_congestion_ops *tcp_ca_find_key(u32 key)
52{
53 struct tcp_congestion_ops *e;
54
55 list_for_each_entry_rcu(e, &tcp_cong_list, list) {
56 if (e->key == key)
57 return e;
58 }
59
60 return NULL;
61}
62
63/*
64 * Attach new congestion control algorithm to the list
65 * of available options.
66 */
67int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
68{
69 int ret = 0;
70
71 /* all algorithms must implement these */
72 if (!ca->ssthresh || !ca->undo_cwnd ||
73 !(ca->cong_avoid || ca->cong_control)) {
74 pr_err("%s does not implement required ops\n", ca->name);
75 return -EINVAL;
76 }
77
78 ca->key = jhash(ca->name, sizeof(ca->name), strlen(ca->name));
79
80 spin_lock(&tcp_cong_list_lock);
81 if (ca->key == TCP_CA_UNSPEC || tcp_ca_find_key(ca->key)) {
82 pr_notice("%s already registered or non-unique key\n",
83 ca->name);
84 ret = -EEXIST;
85 } else {
86 list_add_tail_rcu(&ca->list, &tcp_cong_list);
87 pr_debug("%s registered\n", ca->name);
88 }
89 spin_unlock(&tcp_cong_list_lock);
90
91 return ret;
92}
93EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
94
95/*
96 * Remove congestion control algorithm, called from
97 * the module's remove function. Module ref counts are used
98 * to ensure that this can't be done till all sockets using
99 * that method are closed.
100 */
101void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
102{
103 spin_lock(&tcp_cong_list_lock);
104 list_del_rcu(&ca->list);
105 spin_unlock(&tcp_cong_list_lock);
106
107 /* Wait for outstanding readers to complete before the
108 * module gets removed entirely.
109 *
110 * A try_module_get() should fail by now as our module is
111 * in "going" state since no refs are held anymore and
112 * module_exit() handler being called.
113 */
114 synchronize_rcu();
115}
116EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
117
118u32 tcp_ca_get_key_by_name(const char *name, bool *ecn_ca)
119{
120 const struct tcp_congestion_ops *ca;
121 u32 key = TCP_CA_UNSPEC;
122
123 might_sleep();
124
125 rcu_read_lock();
126 ca = __tcp_ca_find_autoload(name);
127 if (ca) {
128 key = ca->key;
129 *ecn_ca = ca->flags & TCP_CONG_NEEDS_ECN;
130 }
131 rcu_read_unlock();
132
133 return key;
134}
135EXPORT_SYMBOL_GPL(tcp_ca_get_key_by_name);
136
137char *tcp_ca_get_name_by_key(u32 key, char *buffer)
138{
139 const struct tcp_congestion_ops *ca;
140 char *ret = NULL;
141
142 rcu_read_lock();
143 ca = tcp_ca_find_key(key);
144 if (ca)
145 ret = strncpy(buffer, ca->name,
146 TCP_CA_NAME_MAX);
147 rcu_read_unlock();
148
149 return ret;
150}
151EXPORT_SYMBOL_GPL(tcp_ca_get_name_by_key);
152
153/* Assign choice of congestion control. */
154void tcp_assign_congestion_control(struct sock *sk)
155{
156 struct inet_connection_sock *icsk = inet_csk(sk);
157 struct tcp_congestion_ops *ca;
158
159 rcu_read_lock();
160 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
161 if (likely(try_module_get(ca->owner))) {
162 icsk->icsk_ca_ops = ca;
163 goto out;
164 }
165 /* Fallback to next available. The last really
166 * guaranteed fallback is Reno from this list.
167 */
168 }
169out:
170 rcu_read_unlock();
171
172 /* Clear out private data before diag gets it and
173 * the ca has not been initialized.
174 */
175 if (ca->get_info)
176 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
177 if (ca->flags & TCP_CONG_NEEDS_ECN)
178 INET_ECN_xmit(sk);
179 else
180 INET_ECN_dontxmit(sk);
181}
182
183void tcp_init_congestion_control(struct sock *sk)
184{
185 const struct inet_connection_sock *icsk = inet_csk(sk);
186
187 if (icsk->icsk_ca_ops->init)
188 icsk->icsk_ca_ops->init(sk);
189 if (tcp_ca_needs_ecn(sk))
190 INET_ECN_xmit(sk);
191 else
192 INET_ECN_dontxmit(sk);
193}
194
195static void tcp_reinit_congestion_control(struct sock *sk,
196 const struct tcp_congestion_ops *ca)
197{
198 struct inet_connection_sock *icsk = inet_csk(sk);
199
200 tcp_cleanup_congestion_control(sk);
201 icsk->icsk_ca_ops = ca;
202 icsk->icsk_ca_setsockopt = 1;
203
204 if (sk->sk_state != TCP_CLOSE) {
205 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
206 tcp_init_congestion_control(sk);
207 }
208}
209
210/* Manage refcounts on socket close. */
211void tcp_cleanup_congestion_control(struct sock *sk)
212{
213 struct inet_connection_sock *icsk = inet_csk(sk);
214
215 if (icsk->icsk_ca_ops->release)
216 icsk->icsk_ca_ops->release(sk);
217 module_put(icsk->icsk_ca_ops->owner);
218}
219
220/* Used by sysctl to change default congestion control */
221int tcp_set_default_congestion_control(const char *name)
222{
223 struct tcp_congestion_ops *ca;
224 int ret = -ENOENT;
225
226 spin_lock(&tcp_cong_list_lock);
227 ca = tcp_ca_find(name);
228#ifdef CONFIG_MODULES
229 if (!ca && capable(CAP_NET_ADMIN)) {
230 spin_unlock(&tcp_cong_list_lock);
231
232 request_module("tcp_%s", name);
233 spin_lock(&tcp_cong_list_lock);
234 ca = tcp_ca_find(name);
235 }
236#endif
237
238 if (ca) {
239 ca->flags |= TCP_CONG_NON_RESTRICTED; /* default is always allowed */
240 list_move(&ca->list, &tcp_cong_list);
241 ret = 0;
242 }
243 spin_unlock(&tcp_cong_list_lock);
244
245 return ret;
246}
247
248/* Set default value from kernel configuration at bootup */
249static int __init tcp_congestion_default(void)
250{
251 return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG);
252}
253late_initcall(tcp_congestion_default);
254
255/* Build string with list of available congestion control values */
256void tcp_get_available_congestion_control(char *buf, size_t maxlen)
257{
258 struct tcp_congestion_ops *ca;
259 size_t offs = 0;
260
261 rcu_read_lock();
262 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
263 offs += snprintf(buf + offs, maxlen - offs,
264 "%s%s",
265 offs == 0 ? "" : " ", ca->name);
266 }
267 rcu_read_unlock();
268}
269
270/* Get current default congestion control */
271void tcp_get_default_congestion_control(char *name)
272{
273 struct tcp_congestion_ops *ca;
274 /* We will always have reno... */
275 BUG_ON(list_empty(&tcp_cong_list));
276
277 rcu_read_lock();
278 ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
279 strncpy(name, ca->name, TCP_CA_NAME_MAX);
280 rcu_read_unlock();
281}
282
283/* Built list of non-restricted congestion control values */
284void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
285{
286 struct tcp_congestion_ops *ca;
287 size_t offs = 0;
288
289 *buf = '\0';
290 rcu_read_lock();
291 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
292 if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
293 continue;
294 offs += snprintf(buf + offs, maxlen - offs,
295 "%s%s",
296 offs == 0 ? "" : " ", ca->name);
297 }
298 rcu_read_unlock();
299}
300
301/* Change list of non-restricted congestion control */
302int tcp_set_allowed_congestion_control(char *val)
303{
304 struct tcp_congestion_ops *ca;
305 char *saved_clone, *clone, *name;
306 int ret = 0;
307
308 saved_clone = clone = kstrdup(val, GFP_USER);
309 if (!clone)
310 return -ENOMEM;
311
312 spin_lock(&tcp_cong_list_lock);
313 /* pass 1 check for bad entries */
314 while ((name = strsep(&clone, " ")) && *name) {
315 ca = tcp_ca_find(name);
316 if (!ca) {
317 ret = -ENOENT;
318 goto out;
319 }
320 }
321
322 /* pass 2 clear old values */
323 list_for_each_entry_rcu(ca, &tcp_cong_list, list)
324 ca->flags &= ~TCP_CONG_NON_RESTRICTED;
325
326 /* pass 3 mark as allowed */
327 while ((name = strsep(&val, " ")) && *name) {
328 ca = tcp_ca_find(name);
329 WARN_ON(!ca);
330 if (ca)
331 ca->flags |= TCP_CONG_NON_RESTRICTED;
332 }
333out:
334 spin_unlock(&tcp_cong_list_lock);
335 kfree(saved_clone);
336
337 return ret;
338}
339
340/* Change congestion control for socket */
341int tcp_set_congestion_control(struct sock *sk, const char *name)
342{
343 struct inet_connection_sock *icsk = inet_csk(sk);
344 const struct tcp_congestion_ops *ca;
345 int err = 0;
346
347 if (icsk->icsk_ca_dst_locked)
348 return -EPERM;
349
350 rcu_read_lock();
351 ca = __tcp_ca_find_autoload(name);
352 /* No change asking for existing value */
353 if (ca == icsk->icsk_ca_ops) {
354 icsk->icsk_ca_setsockopt = 1;
355 goto out;
356 }
357 if (!ca)
358 err = -ENOENT;
359 else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) ||
360 ns_capable(sock_net(sk)->user_ns, CAP_NET_ADMIN)))
361 err = -EPERM;
362 else if (!try_module_get(ca->owner))
363 err = -EBUSY;
364 else
365 tcp_reinit_congestion_control(sk, ca);
366 out:
367 rcu_read_unlock();
368 return err;
369}
370
371/* Slow start is used when congestion window is no greater than the slow start
372 * threshold. We base on RFC2581 and also handle stretch ACKs properly.
373 * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
374 * something better;) a packet is only considered (s)acked in its entirety to
375 * defend the ACK attacks described in the RFC. Slow start processes a stretch
376 * ACK of degree N as if N acks of degree 1 are received back to back except
377 * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
378 * returns the leftover acks to adjust cwnd in congestion avoidance mode.
379 */
380u32 tcp_slow_start(struct tcp_sock *tp, u32 acked)
381{
382 u32 cwnd = min(tp->snd_cwnd + acked, tp->snd_ssthresh);
383
384 acked -= cwnd - tp->snd_cwnd;
385 tp->snd_cwnd = min(cwnd, tp->snd_cwnd_clamp);
386
387 return acked;
388}
389EXPORT_SYMBOL_GPL(tcp_slow_start);
390
391/* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w),
392 * for every packet that was ACKed.
393 */
394void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked)
395{
396 /* If credits accumulated at a higher w, apply them gently now. */
397 if (tp->snd_cwnd_cnt >= w) {
398 tp->snd_cwnd_cnt = 0;
399 tp->snd_cwnd++;
400 }
401
402 tp->snd_cwnd_cnt += acked;
403 if (tp->snd_cwnd_cnt >= w) {
404 u32 delta = tp->snd_cwnd_cnt / w;
405
406 tp->snd_cwnd_cnt -= delta * w;
407 tp->snd_cwnd += delta;
408 }
409 tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_cwnd_clamp);
410}
411EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
412
413/*
414 * TCP Reno congestion control
415 * This is special case used for fallback as well.
416 */
417/* This is Jacobson's slow start and congestion avoidance.
418 * SIGCOMM '88, p. 328.
419 */
420void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
421{
422 struct tcp_sock *tp = tcp_sk(sk);
423
424 if (!tcp_is_cwnd_limited(sk))
425 return;
426
427 /* In "safe" area, increase. */
428 if (tcp_in_slow_start(tp)) {
429 acked = tcp_slow_start(tp, acked);
430 if (!acked)
431 return;
432 }
433 /* In dangerous area, increase slowly. */
434 tcp_cong_avoid_ai(tp, tp->snd_cwnd, acked);
435}
436EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
437
438/* Slow start threshold is half the congestion window (min 2) */
439u32 tcp_reno_ssthresh(struct sock *sk)
440{
441 const struct tcp_sock *tp = tcp_sk(sk);
442
443 return max(tp->snd_cwnd >> 1U, 2U);
444}
445EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
446
447u32 tcp_reno_undo_cwnd(struct sock *sk)
448{
449 const struct tcp_sock *tp = tcp_sk(sk);
450
451 return max(tp->snd_cwnd, tp->snd_ssthresh << 1);
452}
453EXPORT_SYMBOL_GPL(tcp_reno_undo_cwnd);
454
455struct tcp_congestion_ops tcp_reno = {
456 .flags = TCP_CONG_NON_RESTRICTED,
457 .name = "reno",
458 .owner = THIS_MODULE,
459 .ssthresh = tcp_reno_ssthresh,
460 .cong_avoid = tcp_reno_cong_avoid,
461 .undo_cwnd = tcp_reno_undo_cwnd,
462};