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// 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#include <trace/events/tcp.h>
20
21static DEFINE_SPINLOCK(tcp_cong_list_lock);
22static LIST_HEAD(tcp_cong_list);
23
24/* Simple linear search, don't expect many entries! */
25struct tcp_congestion_ops *tcp_ca_find(const char *name)
26{
27 struct tcp_congestion_ops *e;
28
29 list_for_each_entry_rcu(e, &tcp_cong_list, list) {
30 if (strcmp(e->name, name) == 0)
31 return e;
32 }
33
34 return NULL;
35}
36
37void tcp_set_ca_state(struct sock *sk, const u8 ca_state)
38{
39 struct inet_connection_sock *icsk = inet_csk(sk);
40
41 trace_tcp_cong_state_set(sk, ca_state);
42
43 if (icsk->icsk_ca_ops->set_state)
44 icsk->icsk_ca_ops->set_state(sk, ca_state);
45 icsk->icsk_ca_state = ca_state;
46}
47
48/* Must be called with rcu lock held */
49static struct tcp_congestion_ops *tcp_ca_find_autoload(struct net *net,
50 const char *name)
51{
52 struct tcp_congestion_ops *ca = tcp_ca_find(name);
53
54#ifdef CONFIG_MODULES
55 if (!ca && capable(CAP_NET_ADMIN)) {
56 rcu_read_unlock();
57 request_module("tcp_%s", name);
58 rcu_read_lock();
59 ca = tcp_ca_find(name);
60 }
61#endif
62 return ca;
63}
64
65/* Simple linear search, not much in here. */
66struct tcp_congestion_ops *tcp_ca_find_key(u32 key)
67{
68 struct tcp_congestion_ops *e;
69
70 list_for_each_entry_rcu(e, &tcp_cong_list, list) {
71 if (e->key == key)
72 return e;
73 }
74
75 return NULL;
76}
77
78/*
79 * Attach new congestion control algorithm to the list
80 * of available options.
81 */
82int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
83{
84 int ret = 0;
85
86 /* all algorithms must implement these */
87 if (!ca->ssthresh || !ca->undo_cwnd ||
88 !(ca->cong_avoid || ca->cong_control)) {
89 pr_err("%s does not implement required ops\n", ca->name);
90 return -EINVAL;
91 }
92
93 ca->key = jhash(ca->name, sizeof(ca->name), strlen(ca->name));
94
95 spin_lock(&tcp_cong_list_lock);
96 if (ca->key == TCP_CA_UNSPEC || tcp_ca_find_key(ca->key)) {
97 pr_notice("%s already registered or non-unique key\n",
98 ca->name);
99 ret = -EEXIST;
100 } else {
101 list_add_tail_rcu(&ca->list, &tcp_cong_list);
102 pr_debug("%s registered\n", ca->name);
103 }
104 spin_unlock(&tcp_cong_list_lock);
105
106 return ret;
107}
108EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
109
110/*
111 * Remove congestion control algorithm, called from
112 * the module's remove function. Module ref counts are used
113 * to ensure that this can't be done till all sockets using
114 * that method are closed.
115 */
116void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
117{
118 spin_lock(&tcp_cong_list_lock);
119 list_del_rcu(&ca->list);
120 spin_unlock(&tcp_cong_list_lock);
121
122 /* Wait for outstanding readers to complete before the
123 * module gets removed entirely.
124 *
125 * A try_module_get() should fail by now as our module is
126 * in "going" state since no refs are held anymore and
127 * module_exit() handler being called.
128 */
129 synchronize_rcu();
130}
131EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
132
133u32 tcp_ca_get_key_by_name(struct net *net, const char *name, bool *ecn_ca)
134{
135 const struct tcp_congestion_ops *ca;
136 u32 key = TCP_CA_UNSPEC;
137
138 might_sleep();
139
140 rcu_read_lock();
141 ca = tcp_ca_find_autoload(net, name);
142 if (ca) {
143 key = ca->key;
144 *ecn_ca = ca->flags & TCP_CONG_NEEDS_ECN;
145 }
146 rcu_read_unlock();
147
148 return key;
149}
150
151char *tcp_ca_get_name_by_key(u32 key, char *buffer)
152{
153 const struct tcp_congestion_ops *ca;
154 char *ret = NULL;
155
156 rcu_read_lock();
157 ca = tcp_ca_find_key(key);
158 if (ca)
159 ret = strncpy(buffer, ca->name,
160 TCP_CA_NAME_MAX);
161 rcu_read_unlock();
162
163 return ret;
164}
165
166/* Assign choice of congestion control. */
167void tcp_assign_congestion_control(struct sock *sk)
168{
169 struct net *net = sock_net(sk);
170 struct inet_connection_sock *icsk = inet_csk(sk);
171 const struct tcp_congestion_ops *ca;
172
173 rcu_read_lock();
174 ca = rcu_dereference(net->ipv4.tcp_congestion_control);
175 if (unlikely(!bpf_try_module_get(ca, ca->owner)))
176 ca = &tcp_reno;
177 icsk->icsk_ca_ops = ca;
178 rcu_read_unlock();
179
180 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
181 if (ca->flags & TCP_CONG_NEEDS_ECN)
182 INET_ECN_xmit(sk);
183 else
184 INET_ECN_dontxmit(sk);
185}
186
187void tcp_init_congestion_control(struct sock *sk)
188{
189 struct inet_connection_sock *icsk = inet_csk(sk);
190
191 tcp_sk(sk)->prior_ssthresh = 0;
192 if (icsk->icsk_ca_ops->init)
193 icsk->icsk_ca_ops->init(sk);
194 if (tcp_ca_needs_ecn(sk))
195 INET_ECN_xmit(sk);
196 else
197 INET_ECN_dontxmit(sk);
198 icsk->icsk_ca_initialized = 1;
199}
200
201static void tcp_reinit_congestion_control(struct sock *sk,
202 const struct tcp_congestion_ops *ca)
203{
204 struct inet_connection_sock *icsk = inet_csk(sk);
205
206 tcp_cleanup_congestion_control(sk);
207 icsk->icsk_ca_ops = ca;
208 icsk->icsk_ca_setsockopt = 1;
209 memset(icsk->icsk_ca_priv, 0, sizeof(icsk->icsk_ca_priv));
210
211 if (ca->flags & TCP_CONG_NEEDS_ECN)
212 INET_ECN_xmit(sk);
213 else
214 INET_ECN_dontxmit(sk);
215
216 if (!((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)))
217 tcp_init_congestion_control(sk);
218}
219
220/* Manage refcounts on socket close. */
221void tcp_cleanup_congestion_control(struct sock *sk)
222{
223 struct inet_connection_sock *icsk = inet_csk(sk);
224
225 if (icsk->icsk_ca_ops->release)
226 icsk->icsk_ca_ops->release(sk);
227 bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
228}
229
230/* Used by sysctl to change default congestion control */
231int tcp_set_default_congestion_control(struct net *net, const char *name)
232{
233 struct tcp_congestion_ops *ca;
234 const struct tcp_congestion_ops *prev;
235 int ret;
236
237 rcu_read_lock();
238 ca = tcp_ca_find_autoload(net, name);
239 if (!ca) {
240 ret = -ENOENT;
241 } else if (!bpf_try_module_get(ca, ca->owner)) {
242 ret = -EBUSY;
243 } else if (!net_eq(net, &init_net) &&
244 !(ca->flags & TCP_CONG_NON_RESTRICTED)) {
245 /* Only init netns can set default to a restricted algorithm */
246 ret = -EPERM;
247 } else {
248 prev = xchg(&net->ipv4.tcp_congestion_control, ca);
249 if (prev)
250 bpf_module_put(prev, prev->owner);
251
252 ca->flags |= TCP_CONG_NON_RESTRICTED;
253 ret = 0;
254 }
255 rcu_read_unlock();
256
257 return ret;
258}
259
260/* Set default value from kernel configuration at bootup */
261static int __init tcp_congestion_default(void)
262{
263 return tcp_set_default_congestion_control(&init_net,
264 CONFIG_DEFAULT_TCP_CONG);
265}
266late_initcall(tcp_congestion_default);
267
268/* Build string with list of available congestion control values */
269void tcp_get_available_congestion_control(char *buf, size_t maxlen)
270{
271 struct tcp_congestion_ops *ca;
272 size_t offs = 0;
273
274 rcu_read_lock();
275 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
276 offs += snprintf(buf + offs, maxlen - offs,
277 "%s%s",
278 offs == 0 ? "" : " ", ca->name);
279
280 if (WARN_ON_ONCE(offs >= maxlen))
281 break;
282 }
283 rcu_read_unlock();
284}
285
286/* Get current default congestion control */
287void tcp_get_default_congestion_control(struct net *net, char *name)
288{
289 const struct tcp_congestion_ops *ca;
290
291 rcu_read_lock();
292 ca = rcu_dereference(net->ipv4.tcp_congestion_control);
293 strncpy(name, ca->name, TCP_CA_NAME_MAX);
294 rcu_read_unlock();
295}
296
297/* Built list of non-restricted congestion control values */
298void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
299{
300 struct tcp_congestion_ops *ca;
301 size_t offs = 0;
302
303 *buf = '\0';
304 rcu_read_lock();
305 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
306 if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
307 continue;
308 offs += snprintf(buf + offs, maxlen - offs,
309 "%s%s",
310 offs == 0 ? "" : " ", ca->name);
311
312 if (WARN_ON_ONCE(offs >= maxlen))
313 break;
314 }
315 rcu_read_unlock();
316}
317
318/* Change list of non-restricted congestion control */
319int tcp_set_allowed_congestion_control(char *val)
320{
321 struct tcp_congestion_ops *ca;
322 char *saved_clone, *clone, *name;
323 int ret = 0;
324
325 saved_clone = clone = kstrdup(val, GFP_USER);
326 if (!clone)
327 return -ENOMEM;
328
329 spin_lock(&tcp_cong_list_lock);
330 /* pass 1 check for bad entries */
331 while ((name = strsep(&clone, " ")) && *name) {
332 ca = tcp_ca_find(name);
333 if (!ca) {
334 ret = -ENOENT;
335 goto out;
336 }
337 }
338
339 /* pass 2 clear old values */
340 list_for_each_entry_rcu(ca, &tcp_cong_list, list)
341 ca->flags &= ~TCP_CONG_NON_RESTRICTED;
342
343 /* pass 3 mark as allowed */
344 while ((name = strsep(&val, " ")) && *name) {
345 ca = tcp_ca_find(name);
346 WARN_ON(!ca);
347 if (ca)
348 ca->flags |= TCP_CONG_NON_RESTRICTED;
349 }
350out:
351 spin_unlock(&tcp_cong_list_lock);
352 kfree(saved_clone);
353
354 return ret;
355}
356
357/* Change congestion control for socket. If load is false, then it is the
358 * responsibility of the caller to call tcp_init_congestion_control or
359 * tcp_reinit_congestion_control (if the current congestion control was
360 * already initialized.
361 */
362int tcp_set_congestion_control(struct sock *sk, const char *name, bool load,
363 bool cap_net_admin)
364{
365 struct inet_connection_sock *icsk = inet_csk(sk);
366 const struct tcp_congestion_ops *ca;
367 int err = 0;
368
369 if (icsk->icsk_ca_dst_locked)
370 return -EPERM;
371
372 rcu_read_lock();
373 if (!load)
374 ca = tcp_ca_find(name);
375 else
376 ca = tcp_ca_find_autoload(sock_net(sk), name);
377
378 /* No change asking for existing value */
379 if (ca == icsk->icsk_ca_ops) {
380 icsk->icsk_ca_setsockopt = 1;
381 goto out;
382 }
383
384 if (!ca)
385 err = -ENOENT;
386 else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || cap_net_admin))
387 err = -EPERM;
388 else if (!bpf_try_module_get(ca, ca->owner))
389 err = -EBUSY;
390 else
391 tcp_reinit_congestion_control(sk, ca);
392 out:
393 rcu_read_unlock();
394 return err;
395}
396
397/* Slow start is used when congestion window is no greater than the slow start
398 * threshold. We base on RFC2581 and also handle stretch ACKs properly.
399 * We do not implement RFC3465 Appropriate Byte Counting (ABC) per se but
400 * something better;) a packet is only considered (s)acked in its entirety to
401 * defend the ACK attacks described in the RFC. Slow start processes a stretch
402 * ACK of degree N as if N acks of degree 1 are received back to back except
403 * ABC caps N to 2. Slow start exits when cwnd grows over ssthresh and
404 * returns the leftover acks to adjust cwnd in congestion avoidance mode.
405 */
406u32 tcp_slow_start(struct tcp_sock *tp, u32 acked)
407{
408 u32 cwnd = min(tcp_snd_cwnd(tp) + acked, tp->snd_ssthresh);
409
410 acked -= cwnd - tcp_snd_cwnd(tp);
411 tcp_snd_cwnd_set(tp, min(cwnd, tp->snd_cwnd_clamp));
412
413 return acked;
414}
415EXPORT_SYMBOL_GPL(tcp_slow_start);
416
417/* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w),
418 * for every packet that was ACKed.
419 */
420void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w, u32 acked)
421{
422 /* If credits accumulated at a higher w, apply them gently now. */
423 if (tp->snd_cwnd_cnt >= w) {
424 tp->snd_cwnd_cnt = 0;
425 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + 1);
426 }
427
428 tp->snd_cwnd_cnt += acked;
429 if (tp->snd_cwnd_cnt >= w) {
430 u32 delta = tp->snd_cwnd_cnt / w;
431
432 tp->snd_cwnd_cnt -= delta * w;
433 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) + delta);
434 }
435 tcp_snd_cwnd_set(tp, min(tcp_snd_cwnd(tp), tp->snd_cwnd_clamp));
436}
437EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
438
439/*
440 * TCP Reno congestion control
441 * This is special case used for fallback as well.
442 */
443/* This is Jacobson's slow start and congestion avoidance.
444 * SIGCOMM '88, p. 328.
445 */
446void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 acked)
447{
448 struct tcp_sock *tp = tcp_sk(sk);
449
450 if (!tcp_is_cwnd_limited(sk))
451 return;
452
453 /* In "safe" area, increase. */
454 if (tcp_in_slow_start(tp)) {
455 acked = tcp_slow_start(tp, acked);
456 if (!acked)
457 return;
458 }
459 /* In dangerous area, increase slowly. */
460 tcp_cong_avoid_ai(tp, tcp_snd_cwnd(tp), acked);
461}
462EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
463
464/* Slow start threshold is half the congestion window (min 2) */
465u32 tcp_reno_ssthresh(struct sock *sk)
466{
467 const struct tcp_sock *tp = tcp_sk(sk);
468
469 return max(tcp_snd_cwnd(tp) >> 1U, 2U);
470}
471EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
472
473u32 tcp_reno_undo_cwnd(struct sock *sk)
474{
475 const struct tcp_sock *tp = tcp_sk(sk);
476
477 return max(tcp_snd_cwnd(tp), tp->prior_cwnd);
478}
479EXPORT_SYMBOL_GPL(tcp_reno_undo_cwnd);
480
481struct tcp_congestion_ops tcp_reno = {
482 .flags = TCP_CONG_NON_RESTRICTED,
483 .name = "reno",
484 .owner = THIS_MODULE,
485 .ssthresh = tcp_reno_ssthresh,
486 .cong_avoid = tcp_reno_cong_avoid,
487 .undo_cwnd = tcp_reno_undo_cwnd,
488};