Loading...
1/*
2 * Plugable TCP congestion control support and newReno
3 * congestion control.
4 * Based on ideas from I/O scheduler suport and Web100.
5 *
6 * Copyright (C) 2005 Stephen Hemminger <shemminger@osdl.org>
7 */
8
9#include <linux/module.h>
10#include <linux/mm.h>
11#include <linux/types.h>
12#include <linux/list.h>
13#include <linux/gfp.h>
14#include <net/tcp.h>
15
16int sysctl_tcp_max_ssthresh = 0;
17
18static DEFINE_SPINLOCK(tcp_cong_list_lock);
19static LIST_HEAD(tcp_cong_list);
20
21/* Simple linear search, don't expect many entries! */
22static struct tcp_congestion_ops *tcp_ca_find(const char *name)
23{
24 struct tcp_congestion_ops *e;
25
26 list_for_each_entry_rcu(e, &tcp_cong_list, list) {
27 if (strcmp(e->name, name) == 0)
28 return e;
29 }
30
31 return NULL;
32}
33
34/*
35 * Attach new congestion control algorithm to the list
36 * of available options.
37 */
38int tcp_register_congestion_control(struct tcp_congestion_ops *ca)
39{
40 int ret = 0;
41
42 /* all algorithms must implement ssthresh and cong_avoid ops */
43 if (!ca->ssthresh || !ca->cong_avoid) {
44 printk(KERN_ERR "TCP %s does not implement required ops\n",
45 ca->name);
46 return -EINVAL;
47 }
48
49 spin_lock(&tcp_cong_list_lock);
50 if (tcp_ca_find(ca->name)) {
51 printk(KERN_NOTICE "TCP %s already registered\n", ca->name);
52 ret = -EEXIST;
53 } else {
54 list_add_tail_rcu(&ca->list, &tcp_cong_list);
55 printk(KERN_INFO "TCP %s registered\n", ca->name);
56 }
57 spin_unlock(&tcp_cong_list_lock);
58
59 return ret;
60}
61EXPORT_SYMBOL_GPL(tcp_register_congestion_control);
62
63/*
64 * Remove congestion control algorithm, called from
65 * the module's remove function. Module ref counts are used
66 * to ensure that this can't be done till all sockets using
67 * that method are closed.
68 */
69void tcp_unregister_congestion_control(struct tcp_congestion_ops *ca)
70{
71 spin_lock(&tcp_cong_list_lock);
72 list_del_rcu(&ca->list);
73 spin_unlock(&tcp_cong_list_lock);
74}
75EXPORT_SYMBOL_GPL(tcp_unregister_congestion_control);
76
77/* Assign choice of congestion control. */
78void tcp_init_congestion_control(struct sock *sk)
79{
80 struct inet_connection_sock *icsk = inet_csk(sk);
81 struct tcp_congestion_ops *ca;
82
83 /* if no choice made yet assign the current value set as default */
84 if (icsk->icsk_ca_ops == &tcp_init_congestion_ops) {
85 rcu_read_lock();
86 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
87 if (try_module_get(ca->owner)) {
88 icsk->icsk_ca_ops = ca;
89 break;
90 }
91
92 /* fallback to next available */
93 }
94 rcu_read_unlock();
95 }
96
97 if (icsk->icsk_ca_ops->init)
98 icsk->icsk_ca_ops->init(sk);
99}
100
101/* Manage refcounts on socket close. */
102void tcp_cleanup_congestion_control(struct sock *sk)
103{
104 struct inet_connection_sock *icsk = inet_csk(sk);
105
106 if (icsk->icsk_ca_ops->release)
107 icsk->icsk_ca_ops->release(sk);
108 module_put(icsk->icsk_ca_ops->owner);
109}
110
111/* Used by sysctl to change default congestion control */
112int tcp_set_default_congestion_control(const char *name)
113{
114 struct tcp_congestion_ops *ca;
115 int ret = -ENOENT;
116
117 spin_lock(&tcp_cong_list_lock);
118 ca = tcp_ca_find(name);
119#ifdef CONFIG_MODULES
120 if (!ca && capable(CAP_NET_ADMIN)) {
121 spin_unlock(&tcp_cong_list_lock);
122
123 request_module("tcp_%s", name);
124 spin_lock(&tcp_cong_list_lock);
125 ca = tcp_ca_find(name);
126 }
127#endif
128
129 if (ca) {
130 ca->flags |= TCP_CONG_NON_RESTRICTED; /* default is always allowed */
131 list_move(&ca->list, &tcp_cong_list);
132 ret = 0;
133 }
134 spin_unlock(&tcp_cong_list_lock);
135
136 return ret;
137}
138
139/* Set default value from kernel configuration at bootup */
140static int __init tcp_congestion_default(void)
141{
142 return tcp_set_default_congestion_control(CONFIG_DEFAULT_TCP_CONG);
143}
144late_initcall(tcp_congestion_default);
145
146
147/* Build string with list of available congestion control values */
148void tcp_get_available_congestion_control(char *buf, size_t maxlen)
149{
150 struct tcp_congestion_ops *ca;
151 size_t offs = 0;
152
153 rcu_read_lock();
154 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
155 offs += snprintf(buf + offs, maxlen - offs,
156 "%s%s",
157 offs == 0 ? "" : " ", ca->name);
158
159 }
160 rcu_read_unlock();
161}
162
163/* Get current default congestion control */
164void tcp_get_default_congestion_control(char *name)
165{
166 struct tcp_congestion_ops *ca;
167 /* We will always have reno... */
168 BUG_ON(list_empty(&tcp_cong_list));
169
170 rcu_read_lock();
171 ca = list_entry(tcp_cong_list.next, struct tcp_congestion_ops, list);
172 strncpy(name, ca->name, TCP_CA_NAME_MAX);
173 rcu_read_unlock();
174}
175
176/* Built list of non-restricted congestion control values */
177void tcp_get_allowed_congestion_control(char *buf, size_t maxlen)
178{
179 struct tcp_congestion_ops *ca;
180 size_t offs = 0;
181
182 *buf = '\0';
183 rcu_read_lock();
184 list_for_each_entry_rcu(ca, &tcp_cong_list, list) {
185 if (!(ca->flags & TCP_CONG_NON_RESTRICTED))
186 continue;
187 offs += snprintf(buf + offs, maxlen - offs,
188 "%s%s",
189 offs == 0 ? "" : " ", ca->name);
190
191 }
192 rcu_read_unlock();
193}
194
195/* Change list of non-restricted congestion control */
196int tcp_set_allowed_congestion_control(char *val)
197{
198 struct tcp_congestion_ops *ca;
199 char *saved_clone, *clone, *name;
200 int ret = 0;
201
202 saved_clone = clone = kstrdup(val, GFP_USER);
203 if (!clone)
204 return -ENOMEM;
205
206 spin_lock(&tcp_cong_list_lock);
207 /* pass 1 check for bad entries */
208 while ((name = strsep(&clone, " ")) && *name) {
209 ca = tcp_ca_find(name);
210 if (!ca) {
211 ret = -ENOENT;
212 goto out;
213 }
214 }
215
216 /* pass 2 clear old values */
217 list_for_each_entry_rcu(ca, &tcp_cong_list, list)
218 ca->flags &= ~TCP_CONG_NON_RESTRICTED;
219
220 /* pass 3 mark as allowed */
221 while ((name = strsep(&val, " ")) && *name) {
222 ca = tcp_ca_find(name);
223 WARN_ON(!ca);
224 if (ca)
225 ca->flags |= TCP_CONG_NON_RESTRICTED;
226 }
227out:
228 spin_unlock(&tcp_cong_list_lock);
229 kfree(saved_clone);
230
231 return ret;
232}
233
234
235/* Change congestion control for socket */
236int tcp_set_congestion_control(struct sock *sk, const char *name)
237{
238 struct inet_connection_sock *icsk = inet_csk(sk);
239 struct tcp_congestion_ops *ca;
240 int err = 0;
241
242 rcu_read_lock();
243 ca = tcp_ca_find(name);
244
245 /* no change asking for existing value */
246 if (ca == icsk->icsk_ca_ops)
247 goto out;
248
249#ifdef CONFIG_MODULES
250 /* not found attempt to autoload module */
251 if (!ca && capable(CAP_NET_ADMIN)) {
252 rcu_read_unlock();
253 request_module("tcp_%s", name);
254 rcu_read_lock();
255 ca = tcp_ca_find(name);
256 }
257#endif
258 if (!ca)
259 err = -ENOENT;
260
261 else if (!((ca->flags & TCP_CONG_NON_RESTRICTED) || capable(CAP_NET_ADMIN)))
262 err = -EPERM;
263
264 else if (!try_module_get(ca->owner))
265 err = -EBUSY;
266
267 else {
268 tcp_cleanup_congestion_control(sk);
269 icsk->icsk_ca_ops = ca;
270
271 if (sk->sk_state != TCP_CLOSE && icsk->icsk_ca_ops->init)
272 icsk->icsk_ca_ops->init(sk);
273 }
274 out:
275 rcu_read_unlock();
276 return err;
277}
278
279/* RFC2861 Check whether we are limited by application or congestion window
280 * This is the inverse of cwnd check in tcp_tso_should_defer
281 */
282int tcp_is_cwnd_limited(const struct sock *sk, u32 in_flight)
283{
284 const struct tcp_sock *tp = tcp_sk(sk);
285 u32 left;
286
287 if (in_flight >= tp->snd_cwnd)
288 return 1;
289
290 left = tp->snd_cwnd - in_flight;
291 if (sk_can_gso(sk) &&
292 left * sysctl_tcp_tso_win_divisor < tp->snd_cwnd &&
293 left * tp->mss_cache < sk->sk_gso_max_size)
294 return 1;
295 return left <= tcp_max_burst(tp);
296}
297EXPORT_SYMBOL_GPL(tcp_is_cwnd_limited);
298
299/*
300 * Slow start is used when congestion window is less than slow start
301 * threshold. This version implements the basic RFC2581 version
302 * and optionally supports:
303 * RFC3742 Limited Slow Start - growth limited to max_ssthresh
304 * RFC3465 Appropriate Byte Counting - growth limited by bytes acknowledged
305 */
306void tcp_slow_start(struct tcp_sock *tp)
307{
308 int cnt; /* increase in packets */
309
310 /* RFC3465: ABC Slow start
311 * Increase only after a full MSS of bytes is acked
312 *
313 * TCP sender SHOULD increase cwnd by the number of
314 * previously unacknowledged bytes ACKed by each incoming
315 * acknowledgment, provided the increase is not more than L
316 */
317 if (sysctl_tcp_abc && tp->bytes_acked < tp->mss_cache)
318 return;
319
320 if (sysctl_tcp_max_ssthresh > 0 && tp->snd_cwnd > sysctl_tcp_max_ssthresh)
321 cnt = sysctl_tcp_max_ssthresh >> 1; /* limited slow start */
322 else
323 cnt = tp->snd_cwnd; /* exponential increase */
324
325 /* RFC3465: ABC
326 * We MAY increase by 2 if discovered delayed ack
327 */
328 if (sysctl_tcp_abc > 1 && tp->bytes_acked >= 2*tp->mss_cache)
329 cnt <<= 1;
330 tp->bytes_acked = 0;
331
332 tp->snd_cwnd_cnt += cnt;
333 while (tp->snd_cwnd_cnt >= tp->snd_cwnd) {
334 tp->snd_cwnd_cnt -= tp->snd_cwnd;
335 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
336 tp->snd_cwnd++;
337 }
338}
339EXPORT_SYMBOL_GPL(tcp_slow_start);
340
341/* In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd (or alternative w) */
342void tcp_cong_avoid_ai(struct tcp_sock *tp, u32 w)
343{
344 if (tp->snd_cwnd_cnt >= w) {
345 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
346 tp->snd_cwnd++;
347 tp->snd_cwnd_cnt = 0;
348 } else {
349 tp->snd_cwnd_cnt++;
350 }
351}
352EXPORT_SYMBOL_GPL(tcp_cong_avoid_ai);
353
354/*
355 * TCP Reno congestion control
356 * This is special case used for fallback as well.
357 */
358/* This is Jacobson's slow start and congestion avoidance.
359 * SIGCOMM '88, p. 328.
360 */
361void tcp_reno_cong_avoid(struct sock *sk, u32 ack, u32 in_flight)
362{
363 struct tcp_sock *tp = tcp_sk(sk);
364
365 if (!tcp_is_cwnd_limited(sk, in_flight))
366 return;
367
368 /* In "safe" area, increase. */
369 if (tp->snd_cwnd <= tp->snd_ssthresh)
370 tcp_slow_start(tp);
371
372 /* In dangerous area, increase slowly. */
373 else if (sysctl_tcp_abc) {
374 /* RFC3465: Appropriate Byte Count
375 * increase once for each full cwnd acked
376 */
377 if (tp->bytes_acked >= tp->snd_cwnd*tp->mss_cache) {
378 tp->bytes_acked -= tp->snd_cwnd*tp->mss_cache;
379 if (tp->snd_cwnd < tp->snd_cwnd_clamp)
380 tp->snd_cwnd++;
381 }
382 } else {
383 tcp_cong_avoid_ai(tp, tp->snd_cwnd);
384 }
385}
386EXPORT_SYMBOL_GPL(tcp_reno_cong_avoid);
387
388/* Slow start threshold is half the congestion window (min 2) */
389u32 tcp_reno_ssthresh(struct sock *sk)
390{
391 const struct tcp_sock *tp = tcp_sk(sk);
392 return max(tp->snd_cwnd >> 1U, 2U);
393}
394EXPORT_SYMBOL_GPL(tcp_reno_ssthresh);
395
396/* Lower bound on congestion window with halving. */
397u32 tcp_reno_min_cwnd(const struct sock *sk)
398{
399 const struct tcp_sock *tp = tcp_sk(sk);
400 return tp->snd_ssthresh/2;
401}
402EXPORT_SYMBOL_GPL(tcp_reno_min_cwnd);
403
404struct tcp_congestion_ops tcp_reno = {
405 .flags = TCP_CONG_NON_RESTRICTED,
406 .name = "reno",
407 .owner = THIS_MODULE,
408 .ssthresh = tcp_reno_ssthresh,
409 .cong_avoid = tcp_reno_cong_avoid,
410 .min_cwnd = tcp_reno_min_cwnd,
411};
412
413/* Initial congestion control used (until SYN)
414 * really reno under another name so we can tell difference
415 * during tcp_set_default_congestion_control
416 */
417struct tcp_congestion_ops tcp_init_congestion_ops = {
418 .name = "",
419 .owner = THIS_MODULE,
420 .ssthresh = tcp_reno_ssthresh,
421 .cong_avoid = tcp_reno_cong_avoid,
422 .min_cwnd = tcp_reno_min_cwnd,
423};
424EXPORT_SYMBOL_GPL(tcp_init_congestion_ops);
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};