diff options
Diffstat (limited to 'net/ipv4/tcp_compound.c')
-rw-r--r-- | net/ipv4/tcp_compound.c | 448 |
1 files changed, 448 insertions, 0 deletions
diff --git a/net/ipv4/tcp_compound.c b/net/ipv4/tcp_compound.c new file mode 100644 index 000000000000..bc54f7e9aea9 --- /dev/null +++ b/net/ipv4/tcp_compound.c | |||
@@ -0,0 +1,448 @@ | |||
1 | /* | ||
2 | * TCP Vegas congestion control | ||
3 | * | ||
4 | * This is based on the congestion detection/avoidance scheme described in | ||
5 | * Lawrence S. Brakmo and Larry L. Peterson. | ||
6 | * "TCP Vegas: End to end congestion avoidance on a global internet." | ||
7 | * IEEE Journal on Selected Areas in Communication, 13(8):1465--1480, | ||
8 | * October 1995. Available from: | ||
9 | * ftp://ftp.cs.arizona.edu/xkernel/Papers/jsac.ps | ||
10 | * | ||
11 | * See http://www.cs.arizona.edu/xkernel/ for their implementation. | ||
12 | * The main aspects that distinguish this implementation from the | ||
13 | * Arizona Vegas implementation are: | ||
14 | * o We do not change the loss detection or recovery mechanisms of | ||
15 | * Linux in any way. Linux already recovers from losses quite well, | ||
16 | * using fine-grained timers, NewReno, and FACK. | ||
17 | * o To avoid the performance penalty imposed by increasing cwnd | ||
18 | * only every-other RTT during slow start, we increase during | ||
19 | * every RTT during slow start, just like Reno. | ||
20 | * o Largely to allow continuous cwnd growth during slow start, | ||
21 | * we use the rate at which ACKs come back as the "actual" | ||
22 | * rate, rather than the rate at which data is sent. | ||
23 | * o To speed convergence to the right rate, we set the cwnd | ||
24 | * to achieve the right ("actual") rate when we exit slow start. | ||
25 | * o To filter out the noise caused by delayed ACKs, we use the | ||
26 | * minimum RTT sample observed during the last RTT to calculate | ||
27 | * the actual rate. | ||
28 | * o When the sender re-starts from idle, it waits until it has | ||
29 | * received ACKs for an entire flight of new data before making | ||
30 | * a cwnd adjustment decision. The original Vegas implementation | ||
31 | * assumed senders never went idle. | ||
32 | * | ||
33 | * | ||
34 | * TCP Compound based on TCP Vegas | ||
35 | * | ||
36 | * further details can be found here: | ||
37 | * ftp://ftp.research.microsoft.com/pub/tr/TR-2005-86.pdf | ||
38 | */ | ||
39 | |||
40 | #include <linux/config.h> | ||
41 | #include <linux/mm.h> | ||
42 | #include <linux/module.h> | ||
43 | #include <linux/skbuff.h> | ||
44 | #include <linux/inet_diag.h> | ||
45 | |||
46 | #include <net/tcp.h> | ||
47 | |||
48 | /* Default values of the Vegas variables, in fixed-point representation | ||
49 | * with V_PARAM_SHIFT bits to the right of the binary point. | ||
50 | */ | ||
51 | #define V_PARAM_SHIFT 1 | ||
52 | |||
53 | #define TCP_COMPOUND_ALPHA 3U | ||
54 | #define TCP_COMPOUND_BETA 1U | ||
55 | #define TCP_COMPOUND_GAMMA 30 | ||
56 | #define TCP_COMPOUND_ZETA 1 | ||
57 | |||
58 | /* TCP compound variables */ | ||
59 | struct compound { | ||
60 | u32 beg_snd_nxt; /* right edge during last RTT */ | ||
61 | u32 beg_snd_una; /* left edge during last RTT */ | ||
62 | u32 beg_snd_cwnd; /* saves the size of the cwnd */ | ||
63 | u8 doing_vegas_now; /* if true, do vegas for this RTT */ | ||
64 | u16 cntRTT; /* # of RTTs measured within last RTT */ | ||
65 | u32 minRTT; /* min of RTTs measured within last RTT (in usec) */ | ||
66 | u32 baseRTT; /* the min of all Vegas RTT measurements seen (in usec) */ | ||
67 | |||
68 | u32 cwnd; | ||
69 | u32 dwnd; | ||
70 | }; | ||
71 | |||
72 | /* There are several situations when we must "re-start" Vegas: | ||
73 | * | ||
74 | * o when a connection is established | ||
75 | * o after an RTO | ||
76 | * o after fast recovery | ||
77 | * o when we send a packet and there is no outstanding | ||
78 | * unacknowledged data (restarting an idle connection) | ||
79 | * | ||
80 | * In these circumstances we cannot do a Vegas calculation at the | ||
81 | * end of the first RTT, because any calculation we do is using | ||
82 | * stale info -- both the saved cwnd and congestion feedback are | ||
83 | * stale. | ||
84 | * | ||
85 | * Instead we must wait until the completion of an RTT during | ||
86 | * which we actually receive ACKs. | ||
87 | */ | ||
88 | static inline void vegas_enable(struct sock *sk) | ||
89 | { | ||
90 | const struct tcp_sock *tp = tcp_sk(sk); | ||
91 | struct compound *vegas = inet_csk_ca(sk); | ||
92 | |||
93 | /* Begin taking Vegas samples next time we send something. */ | ||
94 | vegas->doing_vegas_now = 1; | ||
95 | |||
96 | /* Set the beginning of the next send window. */ | ||
97 | vegas->beg_snd_nxt = tp->snd_nxt; | ||
98 | |||
99 | vegas->cntRTT = 0; | ||
100 | vegas->minRTT = 0x7fffffff; | ||
101 | } | ||
102 | |||
103 | /* Stop taking Vegas samples for now. */ | ||
104 | static inline void vegas_disable(struct sock *sk) | ||
105 | { | ||
106 | struct compound *vegas = inet_csk_ca(sk); | ||
107 | |||
108 | vegas->doing_vegas_now = 0; | ||
109 | } | ||
110 | |||
111 | static void tcp_compound_init(struct sock *sk) | ||
112 | { | ||
113 | struct compound *vegas = inet_csk_ca(sk); | ||
114 | const struct tcp_sock *tp = tcp_sk(sk); | ||
115 | |||
116 | vegas->baseRTT = 0x7fffffff; | ||
117 | vegas_enable(sk); | ||
118 | |||
119 | vegas->dwnd = 0; | ||
120 | vegas->cwnd = tp->snd_cwnd; | ||
121 | } | ||
122 | |||
123 | /* Do RTT sampling needed for Vegas. | ||
124 | * Basically we: | ||
125 | * o min-filter RTT samples from within an RTT to get the current | ||
126 | * propagation delay + queuing delay (we are min-filtering to try to | ||
127 | * avoid the effects of delayed ACKs) | ||
128 | * o min-filter RTT samples from a much longer window (forever for now) | ||
129 | * to find the propagation delay (baseRTT) | ||
130 | */ | ||
131 | static void tcp_compound_rtt_calc(struct sock *sk, u32 usrtt) | ||
132 | { | ||
133 | struct compound *vegas = inet_csk_ca(sk); | ||
134 | u32 vrtt = usrtt + 1; /* Never allow zero rtt or baseRTT */ | ||
135 | |||
136 | /* Filter to find propagation delay: */ | ||
137 | if (vrtt < vegas->baseRTT) | ||
138 | vegas->baseRTT = vrtt; | ||
139 | |||
140 | /* Find the min RTT during the last RTT to find | ||
141 | * the current prop. delay + queuing delay: | ||
142 | */ | ||
143 | |||
144 | vegas->minRTT = min(vegas->minRTT, vrtt); | ||
145 | vegas->cntRTT++; | ||
146 | } | ||
147 | |||
148 | static void tcp_compound_state(struct sock *sk, u8 ca_state) | ||
149 | { | ||
150 | |||
151 | if (ca_state == TCP_CA_Open) | ||
152 | vegas_enable(sk); | ||
153 | else | ||
154 | vegas_disable(sk); | ||
155 | } | ||
156 | |||
157 | |||
158 | /* 64bit divisor, dividend and result. dynamic precision */ | ||
159 | static inline u64 div64_64(u64 dividend, u64 divisor) | ||
160 | { | ||
161 | u32 d = divisor; | ||
162 | |||
163 | if (divisor > 0xffffffffULL) { | ||
164 | unsigned int shift = fls(divisor >> 32); | ||
165 | |||
166 | d = divisor >> shift; | ||
167 | dividend >>= shift; | ||
168 | } | ||
169 | |||
170 | /* avoid 64 bit division if possible */ | ||
171 | if (dividend >> 32) | ||
172 | do_div(dividend, d); | ||
173 | else | ||
174 | dividend = (u32) dividend / d; | ||
175 | |||
176 | return dividend; | ||
177 | } | ||
178 | |||
179 | /* calculate the quartic root of "a" using Newton-Raphson */ | ||
180 | static u32 qroot(u64 a) | ||
181 | { | ||
182 | u32 x, x1; | ||
183 | |||
184 | /* Initial estimate is based on: | ||
185 | * qrt(x) = exp(log(x) / 4) | ||
186 | */ | ||
187 | x = 1u << (fls64(a) >> 2); | ||
188 | |||
189 | /* | ||
190 | * Iteration based on: | ||
191 | * 3 | ||
192 | * x = ( 3 * x + a / x ) / 4 | ||
193 | * k+1 k k | ||
194 | */ | ||
195 | do { | ||
196 | u64 x3 = x; | ||
197 | |||
198 | x1 = x; | ||
199 | x3 *= x; | ||
200 | x3 *= x; | ||
201 | |||
202 | x = (3 * x + (u32) div64_64(a, x3)) / 4; | ||
203 | } while (abs(x1 - x) > 1); | ||
204 | |||
205 | return x; | ||
206 | } | ||
207 | |||
208 | |||
209 | /* | ||
210 | * If the connection is idle and we are restarting, | ||
211 | * then we don't want to do any Vegas calculations | ||
212 | * until we get fresh RTT samples. So when we | ||
213 | * restart, we reset our Vegas state to a clean | ||
214 | * slate. After we get acks for this flight of | ||
215 | * packets, _then_ we can make Vegas calculations | ||
216 | * again. | ||
217 | */ | ||
218 | static void tcp_compound_cwnd_event(struct sock *sk, enum tcp_ca_event event) | ||
219 | { | ||
220 | if (event == CA_EVENT_CWND_RESTART || event == CA_EVENT_TX_START) | ||
221 | tcp_compound_init(sk); | ||
222 | } | ||
223 | |||
224 | static void tcp_compound_cong_avoid(struct sock *sk, u32 ack, | ||
225 | u32 seq_rtt, u32 in_flight, int flag) | ||
226 | { | ||
227 | struct tcp_sock *tp = tcp_sk(sk); | ||
228 | struct compound *vegas = inet_csk_ca(sk); | ||
229 | u8 inc = 0; | ||
230 | |||
231 | if (vegas->cwnd + vegas->dwnd > tp->snd_cwnd) { | ||
232 | if (vegas->cwnd > tp->snd_cwnd || vegas->dwnd > tp->snd_cwnd) { | ||
233 | vegas->cwnd = tp->snd_cwnd; | ||
234 | vegas->dwnd = 0; | ||
235 | } else | ||
236 | vegas->cwnd = tp->snd_cwnd - vegas->dwnd; | ||
237 | |||
238 | } | ||
239 | |||
240 | if (!tcp_is_cwnd_limited(sk, in_flight)) | ||
241 | return; | ||
242 | |||
243 | if (vegas->cwnd <= tp->snd_ssthresh) | ||
244 | inc = 1; | ||
245 | else if (tp->snd_cwnd_cnt < tp->snd_cwnd) | ||
246 | tp->snd_cwnd_cnt++; | ||
247 | |||
248 | if (tp->snd_cwnd_cnt >= tp->snd_cwnd) { | ||
249 | inc = 1; | ||
250 | tp->snd_cwnd_cnt = 0; | ||
251 | } | ||
252 | |||
253 | if (inc && tp->snd_cwnd < tp->snd_cwnd_clamp) | ||
254 | vegas->cwnd++; | ||
255 | |||
256 | /* The key players are v_beg_snd_una and v_beg_snd_nxt. | ||
257 | * | ||
258 | * These are so named because they represent the approximate values | ||
259 | * of snd_una and snd_nxt at the beginning of the current RTT. More | ||
260 | * precisely, they represent the amount of data sent during the RTT. | ||
261 | * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt, | ||
262 | * we will calculate that (v_beg_snd_nxt - v_beg_snd_una) outstanding | ||
263 | * bytes of data have been ACKed during the course of the RTT, giving | ||
264 | * an "actual" rate of: | ||
265 | * | ||
266 | * (v_beg_snd_nxt - v_beg_snd_una) / (rtt duration) | ||
267 | * | ||
268 | * Unfortunately, v_beg_snd_una is not exactly equal to snd_una, | ||
269 | * because delayed ACKs can cover more than one segment, so they | ||
270 | * don't line up nicely with the boundaries of RTTs. | ||
271 | * | ||
272 | * Another unfortunate fact of life is that delayed ACKs delay the | ||
273 | * advance of the left edge of our send window, so that the number | ||
274 | * of bytes we send in an RTT is often less than our cwnd will allow. | ||
275 | * So we keep track of our cwnd separately, in v_beg_snd_cwnd. | ||
276 | */ | ||
277 | |||
278 | if (after(ack, vegas->beg_snd_nxt)) { | ||
279 | /* Do the Vegas once-per-RTT cwnd adjustment. */ | ||
280 | u32 old_wnd, old_snd_cwnd; | ||
281 | |||
282 | /* Here old_wnd is essentially the window of data that was | ||
283 | * sent during the previous RTT, and has all | ||
284 | * been acknowledged in the course of the RTT that ended | ||
285 | * with the ACK we just received. Likewise, old_snd_cwnd | ||
286 | * is the cwnd during the previous RTT. | ||
287 | */ | ||
288 | if (!tp->mss_cache) | ||
289 | return; | ||
290 | |||
291 | old_wnd = (vegas->beg_snd_nxt - vegas->beg_snd_una) / | ||
292 | tp->mss_cache; | ||
293 | old_snd_cwnd = vegas->beg_snd_cwnd; | ||
294 | |||
295 | /* Save the extent of the current window so we can use this | ||
296 | * at the end of the next RTT. | ||
297 | */ | ||
298 | vegas->beg_snd_una = vegas->beg_snd_nxt; | ||
299 | vegas->beg_snd_nxt = tp->snd_nxt; | ||
300 | vegas->beg_snd_cwnd = tp->snd_cwnd; | ||
301 | |||
302 | /* We do the Vegas calculations only if we got enough RTT | ||
303 | * samples that we can be reasonably sure that we got | ||
304 | * at least one RTT sample that wasn't from a delayed ACK. | ||
305 | * If we only had 2 samples total, | ||
306 | * then that means we're getting only 1 ACK per RTT, which | ||
307 | * means they're almost certainly delayed ACKs. | ||
308 | * If we have 3 samples, we should be OK. | ||
309 | */ | ||
310 | |||
311 | if (vegas->cntRTT > 2) { | ||
312 | u32 rtt, target_cwnd, diff; | ||
313 | u32 brtt, dwnd; | ||
314 | |||
315 | /* We have enough RTT samples, so, using the Vegas | ||
316 | * algorithm, we determine if we should increase or | ||
317 | * decrease cwnd, and by how much. | ||
318 | */ | ||
319 | |||
320 | /* Pluck out the RTT we are using for the Vegas | ||
321 | * calculations. This is the min RTT seen during the | ||
322 | * last RTT. Taking the min filters out the effects | ||
323 | * of delayed ACKs, at the cost of noticing congestion | ||
324 | * a bit later. | ||
325 | */ | ||
326 | rtt = vegas->minRTT; | ||
327 | |||
328 | /* Calculate the cwnd we should have, if we weren't | ||
329 | * going too fast. | ||
330 | * | ||
331 | * This is: | ||
332 | * (actual rate in segments) * baseRTT | ||
333 | * We keep it as a fixed point number with | ||
334 | * V_PARAM_SHIFT bits to the right of the binary point. | ||
335 | */ | ||
336 | if (!rtt) | ||
337 | return; | ||
338 | |||
339 | brtt = vegas->baseRTT; | ||
340 | target_cwnd = ((old_wnd * brtt) | ||
341 | << V_PARAM_SHIFT) / rtt; | ||
342 | |||
343 | /* Calculate the difference between the window we had, | ||
344 | * and the window we would like to have. This quantity | ||
345 | * is the "Diff" from the Arizona Vegas papers. | ||
346 | * | ||
347 | * Again, this is a fixed point number with | ||
348 | * V_PARAM_SHIFT bits to the right of the binary | ||
349 | * point. | ||
350 | */ | ||
351 | |||
352 | diff = (old_wnd << V_PARAM_SHIFT) - target_cwnd; | ||
353 | |||
354 | dwnd = vegas->dwnd; | ||
355 | |||
356 | if (diff < (TCP_COMPOUND_GAMMA << V_PARAM_SHIFT)) { | ||
357 | u64 v; | ||
358 | u32 x; | ||
359 | |||
360 | /* | ||
361 | * The TCP Compound paper describes the choice | ||
362 | * of "k" determines the agressiveness, | ||
363 | * ie. slope of the response function. | ||
364 | * | ||
365 | * For same value as HSTCP would be 0.8 | ||
366 | * but for computaional reasons, both the | ||
367 | * original authors and this implementation | ||
368 | * use 0.75. | ||
369 | */ | ||
370 | v = old_wnd; | ||
371 | x = qroot(v * v * v) >> TCP_COMPOUND_ALPHA; | ||
372 | if (x > 1) | ||
373 | dwnd = x - 1; | ||
374 | else | ||
375 | dwnd = 0; | ||
376 | |||
377 | dwnd += vegas->dwnd; | ||
378 | |||
379 | } else if ((dwnd << V_PARAM_SHIFT) < | ||
380 | (diff * TCP_COMPOUND_BETA)) | ||
381 | dwnd = 0; | ||
382 | else | ||
383 | dwnd = | ||
384 | ((dwnd << V_PARAM_SHIFT) - | ||
385 | (diff * | ||
386 | TCP_COMPOUND_BETA)) >> V_PARAM_SHIFT; | ||
387 | |||
388 | vegas->dwnd = dwnd; | ||
389 | |||
390 | } | ||
391 | |||
392 | /* Wipe the slate clean for the next RTT. */ | ||
393 | vegas->cntRTT = 0; | ||
394 | vegas->minRTT = 0x7fffffff; | ||
395 | } | ||
396 | |||
397 | tp->snd_cwnd = vegas->cwnd + vegas->dwnd; | ||
398 | } | ||
399 | |||
400 | /* Extract info for Tcp socket info provided via netlink. */ | ||
401 | static void tcp_compound_get_info(struct sock *sk, u32 ext, struct sk_buff *skb) | ||
402 | { | ||
403 | const struct compound *ca = inet_csk_ca(sk); | ||
404 | if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) { | ||
405 | struct tcpvegas_info *info; | ||
406 | |||
407 | info = RTA_DATA(__RTA_PUT(skb, INET_DIAG_VEGASINFO, | ||
408 | sizeof(*info))); | ||
409 | |||
410 | info->tcpv_enabled = ca->doing_vegas_now; | ||
411 | info->tcpv_rttcnt = ca->cntRTT; | ||
412 | info->tcpv_rtt = ca->baseRTT; | ||
413 | info->tcpv_minrtt = ca->minRTT; | ||
414 | rtattr_failure:; | ||
415 | } | ||
416 | } | ||
417 | |||
418 | static struct tcp_congestion_ops tcp_compound = { | ||
419 | .init = tcp_compound_init, | ||
420 | .ssthresh = tcp_reno_ssthresh, | ||
421 | .cong_avoid = tcp_compound_cong_avoid, | ||
422 | .rtt_sample = tcp_compound_rtt_calc, | ||
423 | .set_state = tcp_compound_state, | ||
424 | .cwnd_event = tcp_compound_cwnd_event, | ||
425 | .get_info = tcp_compound_get_info, | ||
426 | |||
427 | .owner = THIS_MODULE, | ||
428 | .name = "compound", | ||
429 | }; | ||
430 | |||
431 | static int __init tcp_compound_register(void) | ||
432 | { | ||
433 | BUG_ON(sizeof(struct compound) > ICSK_CA_PRIV_SIZE); | ||
434 | tcp_register_congestion_control(&tcp_compound); | ||
435 | return 0; | ||
436 | } | ||
437 | |||
438 | static void __exit tcp_compound_unregister(void) | ||
439 | { | ||
440 | tcp_unregister_congestion_control(&tcp_compound); | ||
441 | } | ||
442 | |||
443 | module_init(tcp_compound_register); | ||
444 | module_exit(tcp_compound_unregister); | ||
445 | |||
446 | MODULE_AUTHOR("Angelo P. Castellani, Stephen Hemminger"); | ||
447 | MODULE_LICENSE("GPL"); | ||
448 | MODULE_DESCRIPTION("TCP Compound"); | ||