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