• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * TCP NV: TCP with Congestion Avoidance
4  *
5  * TCP-NV is a successor of TCP-Vegas that has been developed to
6  * deal with the issues that occur in modern networks.
7  * Like TCP-Vegas, TCP-NV supports true congestion avoidance,
8  * the ability to detect congestion before packet losses occur.
9  * When congestion (queue buildup) starts to occur, TCP-NV
10  * predicts what the cwnd size should be for the current
11  * throughput and it reduces the cwnd proportionally to
12  * the difference between the current cwnd and the predicted cwnd.
13  *
14  * NV is only recommeneded for traffic within a data center, and when
15  * all the flows are NV (at least those within the data center). This
16  * is due to the inherent unfairness between flows using losses to
17  * detect congestion (congestion control) and those that use queue
18  * buildup to detect congestion (congestion avoidance).
19  *
20  * Note: High NIC coalescence values may lower the performance of NV
21  * due to the increased noise in RTT values. In particular, we have
22  * seen issues with rx-frames values greater than 8.
23  *
24  * TODO:
25  * 1) Add mechanism to deal with reverse congestion.
26  */
27 
28 #include <linux/mm.h>
29 #include <linux/module.h>
30 #include <linux/math64.h>
31 #include <net/tcp.h>
32 #include <linux/inet_diag.h>
33 
34 /* TCP NV parameters
35  *
36  * nv_pad		Max number of queued packets allowed in network
37  * nv_pad_buffer	Do not grow cwnd if this closed to nv_pad
38  * nv_reset_period	How often (in) seconds)to reset min_rtt
39  * nv_min_cwnd		Don't decrease cwnd below this if there are no losses
40  * nv_cong_dec_mult	Decrease cwnd by X% (30%) of congestion when detected
41  * nv_ssthresh_factor	On congestion set ssthresh to this * <desired cwnd> / 8
42  * nv_rtt_factor	RTT averaging factor
43  * nv_loss_dec_factor	Decrease cwnd to this (80%) when losses occur
44  * nv_dec_eval_min_calls	Wait this many RTT measurements before dec cwnd
45  * nv_inc_eval_min_calls	Wait this many RTT measurements before inc cwnd
46  * nv_ssthresh_eval_min_calls	Wait this many RTT measurements before stopping
47  *				slow-start due to congestion
48  * nv_stop_rtt_cnt	Only grow cwnd for this many RTTs after non-congestion
49  * nv_rtt_min_cnt	Wait these many RTTs before making congesion decision
50  * nv_cwnd_growth_rate_neg
51  * nv_cwnd_growth_rate_pos
52  *	How quickly to double growth rate (not rate) of cwnd when not
53  *	congested. One value (nv_cwnd_growth_rate_neg) for when
54  *	rate < 1 pkt/RTT (after losses). The other (nv_cwnd_growth_rate_pos)
55  *	otherwise.
56  */
57 
58 static int nv_pad __read_mostly = 10;
59 static int nv_pad_buffer __read_mostly = 2;
60 static int nv_reset_period __read_mostly = 5; /* in seconds */
61 static int nv_min_cwnd __read_mostly = 2;
62 static int nv_cong_dec_mult __read_mostly = 30 * 128 / 100; /* = 30% */
63 static int nv_ssthresh_factor __read_mostly = 8; /* = 1 */
64 static int nv_rtt_factor __read_mostly = 128; /* = 1/2*old + 1/2*new */
65 static int nv_loss_dec_factor __read_mostly = 819; /* => 80% */
66 static int nv_cwnd_growth_rate_neg __read_mostly = 8;
67 static int nv_cwnd_growth_rate_pos __read_mostly; /* 0 => fixed like Reno */
68 static int nv_dec_eval_min_calls __read_mostly = 60;
69 static int nv_inc_eval_min_calls __read_mostly = 20;
70 static int nv_ssthresh_eval_min_calls __read_mostly = 30;
71 static int nv_stop_rtt_cnt __read_mostly = 10;
72 static int nv_rtt_min_cnt __read_mostly = 2;
73 
74 module_param(nv_pad, int, 0644);
75 MODULE_PARM_DESC(nv_pad, "max queued packets allowed in network");
76 module_param(nv_reset_period, int, 0644);
77 MODULE_PARM_DESC(nv_reset_period, "nv_min_rtt reset period (secs)");
78 module_param(nv_min_cwnd, int, 0644);
79 MODULE_PARM_DESC(nv_min_cwnd, "NV will not decrease cwnd below this value"
80 		 " without losses");
81 
82 /* TCP NV Parameters */
83 struct tcpnv {
84 	unsigned long nv_min_rtt_reset_jiffies;  /* when to switch to
85 						  * nv_min_rtt_new */
86 	s8  cwnd_growth_factor;	/* Current cwnd growth factor,
87 				 * < 0 => less than 1 packet/RTT */
88 	u8  available8;
89 	u16 available16;
90 	u8  nv_allow_cwnd_growth:1, /* whether cwnd can grow */
91 		nv_reset:1,	    /* whether to reset values */
92 		nv_catchup:1;	    /* whether we are growing because
93 				     * of temporary cwnd decrease */
94 	u8  nv_eval_call_cnt;	/* call count since last eval */
95 	u8  nv_min_cwnd;	/* nv won't make a ca decision if cwnd is
96 				 * smaller than this. It may grow to handle
97 				 * TSO, LRO and interrupt coalescence because
98 				 * with these a small cwnd cannot saturate
99 				 * the link. Note that this is different from
100 				 * the file local nv_min_cwnd */
101 	u8  nv_rtt_cnt;		/* RTTs without making ca decision */;
102 	u32 nv_last_rtt;	/* last rtt */
103 	u32 nv_min_rtt;		/* active min rtt. Used to determine slope */
104 	u32 nv_min_rtt_new;	/* min rtt for future use */
105 	u32 nv_base_rtt;        /* If non-zero it represents the threshold for
106 				 * congestion */
107 	u32 nv_lower_bound_rtt; /* Used in conjunction with nv_base_rtt. It is
108 				 * set to 80% of nv_base_rtt. It helps reduce
109 				 * unfairness between flows */
110 	u32 nv_rtt_max_rate;	/* max rate seen during current RTT */
111 	u32 nv_rtt_start_seq;	/* current RTT ends when packet arrives
112 				 * acking beyond nv_rtt_start_seq */
113 	u32 nv_last_snd_una;	/* Previous value of tp->snd_una. It is
114 				 * used to determine bytes acked since last
115 				 * call to bictcp_acked */
116 	u32 nv_no_cong_cnt;	/* Consecutive no congestion decisions */
117 };
118 
119 #define NV_INIT_RTT	  U32_MAX
120 #define NV_MIN_CWND	  4
121 #define NV_MIN_CWND_GROW  2
122 #define NV_TSO_CWND_BOUND 80
123 
tcpnv_reset(struct tcpnv * ca,struct sock * sk)124 static inline void tcpnv_reset(struct tcpnv *ca, struct sock *sk)
125 {
126 	struct tcp_sock *tp = tcp_sk(sk);
127 
128 	ca->nv_reset = 0;
129 	ca->nv_no_cong_cnt = 0;
130 	ca->nv_rtt_cnt = 0;
131 	ca->nv_last_rtt = 0;
132 	ca->nv_rtt_max_rate = 0;
133 	ca->nv_rtt_start_seq = tp->snd_una;
134 	ca->nv_eval_call_cnt = 0;
135 	ca->nv_last_snd_una = tp->snd_una;
136 }
137 
tcpnv_init(struct sock * sk)138 static void tcpnv_init(struct sock *sk)
139 {
140 	struct tcpnv *ca = inet_csk_ca(sk);
141 	int base_rtt;
142 
143 	tcpnv_reset(ca, sk);
144 
145 	/* See if base_rtt is available from socket_ops bpf program.
146 	 * It is meant to be used in environments, such as communication
147 	 * within a datacenter, where we have reasonable estimates of
148 	 * RTTs
149 	 */
150 	base_rtt = tcp_call_bpf(sk, BPF_SOCK_OPS_BASE_RTT, 0, NULL);
151 	if (base_rtt > 0) {
152 		ca->nv_base_rtt = base_rtt;
153 		ca->nv_lower_bound_rtt = (base_rtt * 205) >> 8; /* 80% */
154 	} else {
155 		ca->nv_base_rtt = 0;
156 		ca->nv_lower_bound_rtt = 0;
157 	}
158 
159 	ca->nv_allow_cwnd_growth = 1;
160 	ca->nv_min_rtt_reset_jiffies = jiffies + 2 * HZ;
161 	ca->nv_min_rtt = NV_INIT_RTT;
162 	ca->nv_min_rtt_new = NV_INIT_RTT;
163 	ca->nv_min_cwnd = NV_MIN_CWND;
164 	ca->nv_catchup = 0;
165 	ca->cwnd_growth_factor = 0;
166 }
167 
168 /* If provided, apply upper (base_rtt) and lower (lower_bound_rtt)
169  * bounds to RTT.
170  */
nv_get_bounded_rtt(struct tcpnv * ca,u32 val)171 inline u32 nv_get_bounded_rtt(struct tcpnv *ca, u32 val)
172 {
173 	if (ca->nv_lower_bound_rtt > 0 && val < ca->nv_lower_bound_rtt)
174 		return ca->nv_lower_bound_rtt;
175 	else if (ca->nv_base_rtt > 0 && val > ca->nv_base_rtt)
176 		return ca->nv_base_rtt;
177 	else
178 		return val;
179 }
180 
tcpnv_cong_avoid(struct sock * sk,u32 ack,u32 acked)181 static void tcpnv_cong_avoid(struct sock *sk, u32 ack, u32 acked)
182 {
183 	struct tcp_sock *tp = tcp_sk(sk);
184 	struct tcpnv *ca = inet_csk_ca(sk);
185 	u32 cnt;
186 
187 	if (!tcp_is_cwnd_limited(sk))
188 		return;
189 
190 	/* Only grow cwnd if NV has not detected congestion */
191 	if (!ca->nv_allow_cwnd_growth)
192 		return;
193 
194 	if (tcp_in_slow_start(tp)) {
195 		acked = tcp_slow_start(tp, acked);
196 		if (!acked)
197 			return;
198 	}
199 
200 	if (ca->cwnd_growth_factor < 0) {
201 		cnt = tp->snd_cwnd << -ca->cwnd_growth_factor;
202 		tcp_cong_avoid_ai(tp, cnt, acked);
203 	} else {
204 		cnt = max(4U, tp->snd_cwnd >> ca->cwnd_growth_factor);
205 		tcp_cong_avoid_ai(tp, cnt, acked);
206 	}
207 }
208 
tcpnv_recalc_ssthresh(struct sock * sk)209 static u32 tcpnv_recalc_ssthresh(struct sock *sk)
210 {
211 	const struct tcp_sock *tp = tcp_sk(sk);
212 
213 	return max((tp->snd_cwnd * nv_loss_dec_factor) >> 10, 2U);
214 }
215 
tcpnv_state(struct sock * sk,u8 new_state)216 static void tcpnv_state(struct sock *sk, u8 new_state)
217 {
218 	struct tcpnv *ca = inet_csk_ca(sk);
219 
220 	if (new_state == TCP_CA_Open && ca->nv_reset) {
221 		tcpnv_reset(ca, sk);
222 	} else if (new_state == TCP_CA_Loss || new_state == TCP_CA_CWR ||
223 		new_state == TCP_CA_Recovery) {
224 		ca->nv_reset = 1;
225 		ca->nv_allow_cwnd_growth = 0;
226 		if (new_state == TCP_CA_Loss) {
227 			/* Reset cwnd growth factor to Reno value */
228 			if (ca->cwnd_growth_factor > 0)
229 				ca->cwnd_growth_factor = 0;
230 			/* Decrease growth rate if allowed */
231 			if (nv_cwnd_growth_rate_neg > 0 &&
232 			    ca->cwnd_growth_factor > -8)
233 				ca->cwnd_growth_factor--;
234 		}
235 	}
236 }
237 
238 /* Do congestion avoidance calculations for TCP-NV
239  */
tcpnv_acked(struct sock * sk,const struct ack_sample * sample)240 static void tcpnv_acked(struct sock *sk, const struct ack_sample *sample)
241 {
242 	const struct inet_connection_sock *icsk = inet_csk(sk);
243 	struct tcp_sock *tp = tcp_sk(sk);
244 	struct tcpnv *ca = inet_csk_ca(sk);
245 	unsigned long now = jiffies;
246 	u64 rate64;
247 	u32 rate, max_win, cwnd_by_slope;
248 	u32 avg_rtt;
249 	u32 bytes_acked = 0;
250 
251 	/* Some calls are for duplicates without timetamps */
252 	if (sample->rtt_us < 0)
253 		return;
254 
255 	/* If not in TCP_CA_Open or TCP_CA_Disorder states, skip. */
256 	if (icsk->icsk_ca_state != TCP_CA_Open &&
257 	    icsk->icsk_ca_state != TCP_CA_Disorder)
258 		return;
259 
260 	/* Stop cwnd growth if we were in catch up mode */
261 	if (ca->nv_catchup && tp->snd_cwnd >= nv_min_cwnd) {
262 		ca->nv_catchup = 0;
263 		ca->nv_allow_cwnd_growth = 0;
264 	}
265 
266 	bytes_acked = tp->snd_una - ca->nv_last_snd_una;
267 	ca->nv_last_snd_una = tp->snd_una;
268 
269 	if (sample->in_flight == 0)
270 		return;
271 
272 	/* Calculate moving average of RTT */
273 	if (nv_rtt_factor > 0) {
274 		if (ca->nv_last_rtt > 0) {
275 			avg_rtt = (((u64)sample->rtt_us) * nv_rtt_factor +
276 				   ((u64)ca->nv_last_rtt)
277 				   * (256 - nv_rtt_factor)) >> 8;
278 		} else {
279 			avg_rtt = sample->rtt_us;
280 			ca->nv_min_rtt = avg_rtt << 1;
281 		}
282 		ca->nv_last_rtt = avg_rtt;
283 	} else {
284 		avg_rtt = sample->rtt_us;
285 	}
286 
287 	/* rate in 100's bits per second */
288 	rate64 = ((u64)sample->in_flight) * 80000;
289 	do_div(rate64, avg_rtt ?: 1);
290 	rate = (u32)rate64;
291 
292 	/* Remember the maximum rate seen during this RTT
293 	 * Note: It may be more than one RTT. This function should be
294 	 *       called at least nv_dec_eval_min_calls times.
295 	 */
296 	if (ca->nv_rtt_max_rate < rate)
297 		ca->nv_rtt_max_rate = rate;
298 
299 	/* We have valid information, increment counter */
300 	if (ca->nv_eval_call_cnt < 255)
301 		ca->nv_eval_call_cnt++;
302 
303 	/* Apply bounds to rtt. Only used to update min_rtt */
304 	avg_rtt = nv_get_bounded_rtt(ca, avg_rtt);
305 
306 	/* update min rtt if necessary */
307 	if (avg_rtt < ca->nv_min_rtt)
308 		ca->nv_min_rtt = avg_rtt;
309 
310 	/* update future min_rtt if necessary */
311 	if (avg_rtt < ca->nv_min_rtt_new)
312 		ca->nv_min_rtt_new = avg_rtt;
313 
314 	/* nv_min_rtt is updated with the minimum (possibley averaged) rtt
315 	 * seen in the last sysctl_tcp_nv_reset_period seconds (i.e. a
316 	 * warm reset). This new nv_min_rtt will be continued to be updated
317 	 * and be used for another sysctl_tcp_nv_reset_period seconds,
318 	 * when it will be updated again.
319 	 * In practice we introduce some randomness, so the actual period used
320 	 * is chosen randomly from the range:
321 	 *   [sysctl_tcp_nv_reset_period*3/4, sysctl_tcp_nv_reset_period*5/4)
322 	 */
323 	if (time_after_eq(now, ca->nv_min_rtt_reset_jiffies)) {
324 		unsigned char rand;
325 
326 		ca->nv_min_rtt = ca->nv_min_rtt_new;
327 		ca->nv_min_rtt_new = NV_INIT_RTT;
328 		get_random_bytes(&rand, 1);
329 		ca->nv_min_rtt_reset_jiffies =
330 			now + ((nv_reset_period * (384 + rand) * HZ) >> 9);
331 		/* Every so often we decrease ca->nv_min_cwnd in case previous
332 		 *  value is no longer accurate.
333 		 */
334 		ca->nv_min_cwnd = max(ca->nv_min_cwnd / 2, NV_MIN_CWND);
335 	}
336 
337 	/* Once per RTT check if we need to do congestion avoidance */
338 	if (before(ca->nv_rtt_start_seq, tp->snd_una)) {
339 		ca->nv_rtt_start_seq = tp->snd_nxt;
340 		if (ca->nv_rtt_cnt < 0xff)
341 			/* Increase counter for RTTs without CA decision */
342 			ca->nv_rtt_cnt++;
343 
344 		/* If this function is only called once within an RTT
345 		 * the cwnd is probably too small (in some cases due to
346 		 * tso, lro or interrupt coalescence), so we increase
347 		 * ca->nv_min_cwnd.
348 		 */
349 		if (ca->nv_eval_call_cnt == 1 &&
350 		    bytes_acked >= (ca->nv_min_cwnd - 1) * tp->mss_cache &&
351 		    ca->nv_min_cwnd < (NV_TSO_CWND_BOUND + 1)) {
352 			ca->nv_min_cwnd = min(ca->nv_min_cwnd
353 					      + NV_MIN_CWND_GROW,
354 					      NV_TSO_CWND_BOUND + 1);
355 			ca->nv_rtt_start_seq = tp->snd_nxt +
356 				ca->nv_min_cwnd * tp->mss_cache;
357 			ca->nv_eval_call_cnt = 0;
358 			ca->nv_allow_cwnd_growth = 1;
359 			return;
360 		}
361 
362 		/* Find the ideal cwnd for current rate from slope
363 		 * slope = 80000.0 * mss / nv_min_rtt
364 		 * cwnd_by_slope = nv_rtt_max_rate / slope
365 		 */
366 		cwnd_by_slope = (u32)
367 			div64_u64(((u64)ca->nv_rtt_max_rate) * ca->nv_min_rtt,
368 				  80000ULL * tp->mss_cache);
369 		max_win = cwnd_by_slope + nv_pad;
370 
371 		/* If cwnd > max_win, decrease cwnd
372 		 * if cwnd < max_win, grow cwnd
373 		 * else leave the same
374 		 */
375 		if (tp->snd_cwnd > max_win) {
376 			/* there is congestion, check that it is ok
377 			 * to make a CA decision
378 			 * 1. We should have at least nv_dec_eval_min_calls
379 			 *    data points before making a CA  decision
380 			 * 2. We only make a congesion decision after
381 			 *    nv_rtt_min_cnt RTTs
382 			 */
383 			if (ca->nv_rtt_cnt < nv_rtt_min_cnt) {
384 				return;
385 			} else if (tp->snd_ssthresh == TCP_INFINITE_SSTHRESH) {
386 				if (ca->nv_eval_call_cnt <
387 				    nv_ssthresh_eval_min_calls)
388 					return;
389 				/* otherwise we will decrease cwnd */
390 			} else if (ca->nv_eval_call_cnt <
391 				   nv_dec_eval_min_calls) {
392 				if (ca->nv_allow_cwnd_growth &&
393 				    ca->nv_rtt_cnt > nv_stop_rtt_cnt)
394 					ca->nv_allow_cwnd_growth = 0;
395 				return;
396 			}
397 
398 			/* We have enough data to determine we are congested */
399 			ca->nv_allow_cwnd_growth = 0;
400 			tp->snd_ssthresh =
401 				(nv_ssthresh_factor * max_win) >> 3;
402 			if (tp->snd_cwnd - max_win > 2) {
403 				/* gap > 2, we do exponential cwnd decrease */
404 				int dec;
405 
406 				dec = max(2U, ((tp->snd_cwnd - max_win) *
407 					       nv_cong_dec_mult) >> 7);
408 				tp->snd_cwnd -= dec;
409 			} else if (nv_cong_dec_mult > 0) {
410 				tp->snd_cwnd = max_win;
411 			}
412 			if (ca->cwnd_growth_factor > 0)
413 				ca->cwnd_growth_factor = 0;
414 			ca->nv_no_cong_cnt = 0;
415 		} else if (tp->snd_cwnd <= max_win - nv_pad_buffer) {
416 			/* There is no congestion, grow cwnd if allowed*/
417 			if (ca->nv_eval_call_cnt < nv_inc_eval_min_calls)
418 				return;
419 
420 			ca->nv_allow_cwnd_growth = 1;
421 			ca->nv_no_cong_cnt++;
422 			if (ca->cwnd_growth_factor < 0 &&
423 			    nv_cwnd_growth_rate_neg > 0 &&
424 			    ca->nv_no_cong_cnt > nv_cwnd_growth_rate_neg) {
425 				ca->cwnd_growth_factor++;
426 				ca->nv_no_cong_cnt = 0;
427 			} else if (ca->cwnd_growth_factor >= 0 &&
428 				   nv_cwnd_growth_rate_pos > 0 &&
429 				   ca->nv_no_cong_cnt >
430 				   nv_cwnd_growth_rate_pos) {
431 				ca->cwnd_growth_factor++;
432 				ca->nv_no_cong_cnt = 0;
433 			}
434 		} else {
435 			/* cwnd is in-between, so do nothing */
436 			return;
437 		}
438 
439 		/* update state */
440 		ca->nv_eval_call_cnt = 0;
441 		ca->nv_rtt_cnt = 0;
442 		ca->nv_rtt_max_rate = 0;
443 
444 		/* Don't want to make cwnd < nv_min_cwnd
445 		 * (it wasn't before, if it is now is because nv
446 		 *  decreased it).
447 		 */
448 		if (tp->snd_cwnd < nv_min_cwnd)
449 			tp->snd_cwnd = nv_min_cwnd;
450 	}
451 }
452 
453 /* Extract info for Tcp socket info provided via netlink */
tcpnv_get_info(struct sock * sk,u32 ext,int * attr,union tcp_cc_info * info)454 static size_t tcpnv_get_info(struct sock *sk, u32 ext, int *attr,
455 			     union tcp_cc_info *info)
456 {
457 	const struct tcpnv *ca = inet_csk_ca(sk);
458 
459 	if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
460 		info->vegas.tcpv_enabled = 1;
461 		info->vegas.tcpv_rttcnt = ca->nv_rtt_cnt;
462 		info->vegas.tcpv_rtt = ca->nv_last_rtt;
463 		info->vegas.tcpv_minrtt = ca->nv_min_rtt;
464 
465 		*attr = INET_DIAG_VEGASINFO;
466 		return sizeof(struct tcpvegas_info);
467 	}
468 	return 0;
469 }
470 
471 static struct tcp_congestion_ops tcpnv __read_mostly = {
472 	.init		= tcpnv_init,
473 	.ssthresh	= tcpnv_recalc_ssthresh,
474 	.cong_avoid	= tcpnv_cong_avoid,
475 	.set_state	= tcpnv_state,
476 	.undo_cwnd	= tcp_reno_undo_cwnd,
477 	.pkts_acked     = tcpnv_acked,
478 	.get_info	= tcpnv_get_info,
479 
480 	.owner		= THIS_MODULE,
481 	.name		= "nv",
482 };
483 
tcpnv_register(void)484 static int __init tcpnv_register(void)
485 {
486 	BUILD_BUG_ON(sizeof(struct tcpnv) > ICSK_CA_PRIV_SIZE);
487 
488 	return tcp_register_congestion_control(&tcpnv);
489 }
490 
tcpnv_unregister(void)491 static void __exit tcpnv_unregister(void)
492 {
493 	tcp_unregister_congestion_control(&tcpnv);
494 }
495 
496 module_init(tcpnv_register);
497 module_exit(tcpnv_unregister);
498 
499 MODULE_AUTHOR("Lawrence Brakmo");
500 MODULE_LICENSE("GPL");
501 MODULE_DESCRIPTION("TCP NV");
502 MODULE_VERSION("1.0");
503