xref: /openbmc/linux/net/ipv4/tcp_nv.c (revision 9a29f5fc)
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/module.h>
29 #include <linux/math64.h>
30 #include <net/tcp.h>
31 #include <linux/inet_diag.h>
32 
33 /* TCP NV parameters
34  *
35  * nv_pad		Max number of queued packets allowed in network
36  * nv_pad_buffer	Do not grow cwnd if this closed to nv_pad
37  * nv_reset_period	How often (in) seconds)to reset min_rtt
38  * nv_min_cwnd		Don't decrease cwnd below this if there are no losses
39  * nv_cong_dec_mult	Decrease cwnd by X% (30%) of congestion when detected
40  * nv_ssthresh_factor	On congestion set ssthresh to this * <desired cwnd> / 8
41  * nv_rtt_factor	RTT averaging factor
42  * nv_loss_dec_factor	Decrease cwnd to this (80%) when losses occur
43  * nv_dec_eval_min_calls	Wait this many RTT measurements before dec cwnd
44  * nv_inc_eval_min_calls	Wait this many RTT measurements before inc cwnd
45  * nv_ssthresh_eval_min_calls	Wait this many RTT measurements before stopping
46  *				slow-start due to congestion
47  * nv_stop_rtt_cnt	Only grow cwnd for this many RTTs after non-congestion
48  * nv_rtt_min_cnt	Wait these many RTTs before making congesion decision
49  * nv_cwnd_growth_rate_neg
50  * nv_cwnd_growth_rate_pos
51  *	How quickly to double growth rate (not rate) of cwnd when not
52  *	congested. One value (nv_cwnd_growth_rate_neg) for when
53  *	rate < 1 pkt/RTT (after losses). The other (nv_cwnd_growth_rate_pos)
54  *	otherwise.
55  */
56 
57 static int nv_pad __read_mostly = 10;
58 static int nv_pad_buffer __read_mostly = 2;
59 static int nv_reset_period __read_mostly = 5; /* in seconds */
60 static int nv_min_cwnd __read_mostly = 2;
61 static int nv_cong_dec_mult __read_mostly = 30 * 128 / 100; /* = 30% */
62 static int nv_ssthresh_factor __read_mostly = 8; /* = 1 */
63 static int nv_rtt_factor __read_mostly = 128; /* = 1/2*old + 1/2*new */
64 static int nv_loss_dec_factor __read_mostly = 819; /* => 80% */
65 static int nv_cwnd_growth_rate_neg __read_mostly = 8;
66 static int nv_cwnd_growth_rate_pos __read_mostly; /* 0 => fixed like Reno */
67 static int nv_dec_eval_min_calls __read_mostly = 60;
68 static int nv_inc_eval_min_calls __read_mostly = 20;
69 static int nv_ssthresh_eval_min_calls __read_mostly = 30;
70 static int nv_stop_rtt_cnt __read_mostly = 10;
71 static int nv_rtt_min_cnt __read_mostly = 2;
72 
73 module_param(nv_pad, int, 0644);
74 MODULE_PARM_DESC(nv_pad, "max queued packets allowed in network");
75 module_param(nv_reset_period, int, 0644);
76 MODULE_PARM_DESC(nv_reset_period, "nv_min_rtt reset period (secs)");
77 module_param(nv_min_cwnd, int, 0644);
78 MODULE_PARM_DESC(nv_min_cwnd, "NV will not decrease cwnd below this value"
79 		 " without losses");
80 
81 /* TCP NV Parameters */
82 struct tcpnv {
83 	unsigned long nv_min_rtt_reset_jiffies;  /* when to switch to
84 						  * nv_min_rtt_new */
85 	s8  cwnd_growth_factor;	/* Current cwnd growth factor,
86 				 * < 0 => less than 1 packet/RTT */
87 	u8  available8;
88 	u16 available16;
89 	u8  nv_allow_cwnd_growth:1, /* whether cwnd can grow */
90 		nv_reset:1,	    /* whether to reset values */
91 		nv_catchup:1;	    /* whether we are growing because
92 				     * of temporary cwnd decrease */
93 	u8  nv_eval_call_cnt;	/* call count since last eval */
94 	u8  nv_min_cwnd;	/* nv won't make a ca decision if cwnd is
95 				 * smaller than this. It may grow to handle
96 				 * TSO, LRO and interrupt coalescence because
97 				 * with these a small cwnd cannot saturate
98 				 * the link. Note that this is different from
99 				 * the file local nv_min_cwnd */
100 	u8  nv_rtt_cnt;		/* RTTs without making ca decision */;
101 	u32 nv_last_rtt;	/* last rtt */
102 	u32 nv_min_rtt;		/* active min rtt. Used to determine slope */
103 	u32 nv_min_rtt_new;	/* min rtt for future use */
104 	u32 nv_base_rtt;        /* If non-zero it represents the threshold for
105 				 * congestion */
106 	u32 nv_lower_bound_rtt; /* Used in conjunction with nv_base_rtt. It is
107 				 * set to 80% of nv_base_rtt. It helps reduce
108 				 * unfairness between flows */
109 	u32 nv_rtt_max_rate;	/* max rate seen during current RTT */
110 	u32 nv_rtt_start_seq;	/* current RTT ends when packet arrives
111 				 * acking beyond nv_rtt_start_seq */
112 	u32 nv_last_snd_una;	/* Previous value of tp->snd_una. It is
113 				 * used to determine bytes acked since last
114 				 * call to bictcp_acked */
115 	u32 nv_no_cong_cnt;	/* Consecutive no congestion decisions */
116 };
117 
118 #define NV_INIT_RTT	  U32_MAX
119 #define NV_MIN_CWND	  4
120 #define NV_MIN_CWND_GROW  2
121 #define NV_TSO_CWND_BOUND 80
122 
123 static inline void tcpnv_reset(struct tcpnv *ca, struct sock *sk)
124 {
125 	struct tcp_sock *tp = tcp_sk(sk);
126 
127 	ca->nv_reset = 0;
128 	ca->nv_no_cong_cnt = 0;
129 	ca->nv_rtt_cnt = 0;
130 	ca->nv_last_rtt = 0;
131 	ca->nv_rtt_max_rate = 0;
132 	ca->nv_rtt_start_seq = tp->snd_una;
133 	ca->nv_eval_call_cnt = 0;
134 	ca->nv_last_snd_una = tp->snd_una;
135 }
136 
137 static void tcpnv_init(struct sock *sk)
138 {
139 	struct tcpnv *ca = inet_csk_ca(sk);
140 	int base_rtt;
141 
142 	tcpnv_reset(ca, sk);
143 
144 	/* See if base_rtt is available from socket_ops bpf program.
145 	 * It is meant to be used in environments, such as communication
146 	 * within a datacenter, where we have reasonable estimates of
147 	 * RTTs
148 	 */
149 	base_rtt = tcp_call_bpf(sk, BPF_SOCK_OPS_BASE_RTT, 0, NULL);
150 	if (base_rtt > 0) {
151 		ca->nv_base_rtt = base_rtt;
152 		ca->nv_lower_bound_rtt = (base_rtt * 205) >> 8; /* 80% */
153 	} else {
154 		ca->nv_base_rtt = 0;
155 		ca->nv_lower_bound_rtt = 0;
156 	}
157 
158 	ca->nv_allow_cwnd_growth = 1;
159 	ca->nv_min_rtt_reset_jiffies = jiffies + 2 * HZ;
160 	ca->nv_min_rtt = NV_INIT_RTT;
161 	ca->nv_min_rtt_new = NV_INIT_RTT;
162 	ca->nv_min_cwnd = NV_MIN_CWND;
163 	ca->nv_catchup = 0;
164 	ca->cwnd_growth_factor = 0;
165 }
166 
167 /* If provided, apply upper (base_rtt) and lower (lower_bound_rtt)
168  * bounds to RTT.
169  */
170 inline u32 nv_get_bounded_rtt(struct tcpnv *ca, u32 val)
171 {
172 	if (ca->nv_lower_bound_rtt > 0 && val < ca->nv_lower_bound_rtt)
173 		return ca->nv_lower_bound_rtt;
174 	else if (ca->nv_base_rtt > 0 && val > ca->nv_base_rtt)
175 		return ca->nv_base_rtt;
176 	else
177 		return val;
178 }
179 
180 static void tcpnv_cong_avoid(struct sock *sk, u32 ack, u32 acked)
181 {
182 	struct tcp_sock *tp = tcp_sk(sk);
183 	struct tcpnv *ca = inet_csk_ca(sk);
184 	u32 cnt;
185 
186 	if (!tcp_is_cwnd_limited(sk))
187 		return;
188 
189 	/* Only grow cwnd if NV has not detected congestion */
190 	if (!ca->nv_allow_cwnd_growth)
191 		return;
192 
193 	if (tcp_in_slow_start(tp)) {
194 		acked = tcp_slow_start(tp, acked);
195 		if (!acked)
196 			return;
197 	}
198 
199 	if (ca->cwnd_growth_factor < 0) {
200 		cnt = tcp_snd_cwnd(tp) << -ca->cwnd_growth_factor;
201 		tcp_cong_avoid_ai(tp, cnt, acked);
202 	} else {
203 		cnt = max(4U, tcp_snd_cwnd(tp) >> ca->cwnd_growth_factor);
204 		tcp_cong_avoid_ai(tp, cnt, acked);
205 	}
206 }
207 
208 static u32 tcpnv_recalc_ssthresh(struct sock *sk)
209 {
210 	const struct tcp_sock *tp = tcp_sk(sk);
211 
212 	return max((tcp_snd_cwnd(tp) * nv_loss_dec_factor) >> 10, 2U);
213 }
214 
215 static void tcpnv_state(struct sock *sk, u8 new_state)
216 {
217 	struct tcpnv *ca = inet_csk_ca(sk);
218 
219 	if (new_state == TCP_CA_Open && ca->nv_reset) {
220 		tcpnv_reset(ca, sk);
221 	} else if (new_state == TCP_CA_Loss || new_state == TCP_CA_CWR ||
222 		new_state == TCP_CA_Recovery) {
223 		ca->nv_reset = 1;
224 		ca->nv_allow_cwnd_growth = 0;
225 		if (new_state == TCP_CA_Loss) {
226 			/* Reset cwnd growth factor to Reno value */
227 			if (ca->cwnd_growth_factor > 0)
228 				ca->cwnd_growth_factor = 0;
229 			/* Decrease growth rate if allowed */
230 			if (nv_cwnd_growth_rate_neg > 0 &&
231 			    ca->cwnd_growth_factor > -8)
232 				ca->cwnd_growth_factor--;
233 		}
234 	}
235 }
236 
237 /* Do congestion avoidance calculations for TCP-NV
238  */
239 static void tcpnv_acked(struct sock *sk, const struct ack_sample *sample)
240 {
241 	const struct inet_connection_sock *icsk = inet_csk(sk);
242 	struct tcp_sock *tp = tcp_sk(sk);
243 	struct tcpnv *ca = inet_csk_ca(sk);
244 	unsigned long now = jiffies;
245 	u64 rate64;
246 	u32 rate, max_win, cwnd_by_slope;
247 	u32 avg_rtt;
248 	u32 bytes_acked = 0;
249 
250 	/* Some calls are for duplicates without timetamps */
251 	if (sample->rtt_us < 0)
252 		return;
253 
254 	/* If not in TCP_CA_Open or TCP_CA_Disorder states, skip. */
255 	if (icsk->icsk_ca_state != TCP_CA_Open &&
256 	    icsk->icsk_ca_state != TCP_CA_Disorder)
257 		return;
258 
259 	/* Stop cwnd growth if we were in catch up mode */
260 	if (ca->nv_catchup && tcp_snd_cwnd(tp) >= nv_min_cwnd) {
261 		ca->nv_catchup = 0;
262 		ca->nv_allow_cwnd_growth = 0;
263 	}
264 
265 	bytes_acked = tp->snd_una - ca->nv_last_snd_una;
266 	ca->nv_last_snd_una = tp->snd_una;
267 
268 	if (sample->in_flight == 0)
269 		return;
270 
271 	/* Calculate moving average of RTT */
272 	if (nv_rtt_factor > 0) {
273 		if (ca->nv_last_rtt > 0) {
274 			avg_rtt = (((u64)sample->rtt_us) * nv_rtt_factor +
275 				   ((u64)ca->nv_last_rtt)
276 				   * (256 - nv_rtt_factor)) >> 8;
277 		} else {
278 			avg_rtt = sample->rtt_us;
279 			ca->nv_min_rtt = avg_rtt << 1;
280 		}
281 		ca->nv_last_rtt = avg_rtt;
282 	} else {
283 		avg_rtt = sample->rtt_us;
284 	}
285 
286 	/* rate in 100's bits per second */
287 	rate64 = ((u64)sample->in_flight) * 80000;
288 	do_div(rate64, avg_rtt ?: 1);
289 	rate = (u32)rate64;
290 
291 	/* Remember the maximum rate seen during this RTT
292 	 * Note: It may be more than one RTT. This function should be
293 	 *       called at least nv_dec_eval_min_calls times.
294 	 */
295 	if (ca->nv_rtt_max_rate < rate)
296 		ca->nv_rtt_max_rate = rate;
297 
298 	/* We have valid information, increment counter */
299 	if (ca->nv_eval_call_cnt < 255)
300 		ca->nv_eval_call_cnt++;
301 
302 	/* Apply bounds to rtt. Only used to update min_rtt */
303 	avg_rtt = nv_get_bounded_rtt(ca, avg_rtt);
304 
305 	/* update min rtt if necessary */
306 	if (avg_rtt < ca->nv_min_rtt)
307 		ca->nv_min_rtt = avg_rtt;
308 
309 	/* update future min_rtt if necessary */
310 	if (avg_rtt < ca->nv_min_rtt_new)
311 		ca->nv_min_rtt_new = avg_rtt;
312 
313 	/* nv_min_rtt is updated with the minimum (possibley averaged) rtt
314 	 * seen in the last sysctl_tcp_nv_reset_period seconds (i.e. a
315 	 * warm reset). This new nv_min_rtt will be continued to be updated
316 	 * and be used for another sysctl_tcp_nv_reset_period seconds,
317 	 * when it will be updated again.
318 	 * In practice we introduce some randomness, so the actual period used
319 	 * is chosen randomly from the range:
320 	 *   [sysctl_tcp_nv_reset_period*3/4, sysctl_tcp_nv_reset_period*5/4)
321 	 */
322 	if (time_after_eq(now, ca->nv_min_rtt_reset_jiffies)) {
323 		unsigned char rand;
324 
325 		ca->nv_min_rtt = ca->nv_min_rtt_new;
326 		ca->nv_min_rtt_new = NV_INIT_RTT;
327 		get_random_bytes(&rand, 1);
328 		ca->nv_min_rtt_reset_jiffies =
329 			now + ((nv_reset_period * (384 + rand) * HZ) >> 9);
330 		/* Every so often we decrease ca->nv_min_cwnd in case previous
331 		 *  value is no longer accurate.
332 		 */
333 		ca->nv_min_cwnd = max(ca->nv_min_cwnd / 2, NV_MIN_CWND);
334 	}
335 
336 	/* Once per RTT check if we need to do congestion avoidance */
337 	if (before(ca->nv_rtt_start_seq, tp->snd_una)) {
338 		ca->nv_rtt_start_seq = tp->snd_nxt;
339 		if (ca->nv_rtt_cnt < 0xff)
340 			/* Increase counter for RTTs without CA decision */
341 			ca->nv_rtt_cnt++;
342 
343 		/* If this function is only called once within an RTT
344 		 * the cwnd is probably too small (in some cases due to
345 		 * tso, lro or interrupt coalescence), so we increase
346 		 * ca->nv_min_cwnd.
347 		 */
348 		if (ca->nv_eval_call_cnt == 1 &&
349 		    bytes_acked >= (ca->nv_min_cwnd - 1) * tp->mss_cache &&
350 		    ca->nv_min_cwnd < (NV_TSO_CWND_BOUND + 1)) {
351 			ca->nv_min_cwnd = min(ca->nv_min_cwnd
352 					      + NV_MIN_CWND_GROW,
353 					      NV_TSO_CWND_BOUND + 1);
354 			ca->nv_rtt_start_seq = tp->snd_nxt +
355 				ca->nv_min_cwnd * tp->mss_cache;
356 			ca->nv_eval_call_cnt = 0;
357 			ca->nv_allow_cwnd_growth = 1;
358 			return;
359 		}
360 
361 		/* Find the ideal cwnd for current rate from slope
362 		 * slope = 80000.0 * mss / nv_min_rtt
363 		 * cwnd_by_slope = nv_rtt_max_rate / slope
364 		 */
365 		cwnd_by_slope = (u32)
366 			div64_u64(((u64)ca->nv_rtt_max_rate) * ca->nv_min_rtt,
367 				  80000ULL * tp->mss_cache);
368 		max_win = cwnd_by_slope + nv_pad;
369 
370 		/* If cwnd > max_win, decrease cwnd
371 		 * if cwnd < max_win, grow cwnd
372 		 * else leave the same
373 		 */
374 		if (tcp_snd_cwnd(tp) > max_win) {
375 			/* there is congestion, check that it is ok
376 			 * to make a CA decision
377 			 * 1. We should have at least nv_dec_eval_min_calls
378 			 *    data points before making a CA  decision
379 			 * 2. We only make a congesion decision after
380 			 *    nv_rtt_min_cnt RTTs
381 			 */
382 			if (ca->nv_rtt_cnt < nv_rtt_min_cnt) {
383 				return;
384 			} else if (tp->snd_ssthresh == TCP_INFINITE_SSTHRESH) {
385 				if (ca->nv_eval_call_cnt <
386 				    nv_ssthresh_eval_min_calls)
387 					return;
388 				/* otherwise we will decrease cwnd */
389 			} else if (ca->nv_eval_call_cnt <
390 				   nv_dec_eval_min_calls) {
391 				if (ca->nv_allow_cwnd_growth &&
392 				    ca->nv_rtt_cnt > nv_stop_rtt_cnt)
393 					ca->nv_allow_cwnd_growth = 0;
394 				return;
395 			}
396 
397 			/* We have enough data to determine we are congested */
398 			ca->nv_allow_cwnd_growth = 0;
399 			tp->snd_ssthresh =
400 				(nv_ssthresh_factor * max_win) >> 3;
401 			if (tcp_snd_cwnd(tp) - max_win > 2) {
402 				/* gap > 2, we do exponential cwnd decrease */
403 				int dec;
404 
405 				dec = max(2U, ((tcp_snd_cwnd(tp) - max_win) *
406 					       nv_cong_dec_mult) >> 7);
407 				tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - dec);
408 			} else if (nv_cong_dec_mult > 0) {
409 				tcp_snd_cwnd_set(tp, max_win);
410 			}
411 			if (ca->cwnd_growth_factor > 0)
412 				ca->cwnd_growth_factor = 0;
413 			ca->nv_no_cong_cnt = 0;
414 		} else if (tcp_snd_cwnd(tp) <= max_win - nv_pad_buffer) {
415 			/* There is no congestion, grow cwnd if allowed*/
416 			if (ca->nv_eval_call_cnt < nv_inc_eval_min_calls)
417 				return;
418 
419 			ca->nv_allow_cwnd_growth = 1;
420 			ca->nv_no_cong_cnt++;
421 			if (ca->cwnd_growth_factor < 0 &&
422 			    nv_cwnd_growth_rate_neg > 0 &&
423 			    ca->nv_no_cong_cnt > nv_cwnd_growth_rate_neg) {
424 				ca->cwnd_growth_factor++;
425 				ca->nv_no_cong_cnt = 0;
426 			} else if (ca->cwnd_growth_factor >= 0 &&
427 				   nv_cwnd_growth_rate_pos > 0 &&
428 				   ca->nv_no_cong_cnt >
429 				   nv_cwnd_growth_rate_pos) {
430 				ca->cwnd_growth_factor++;
431 				ca->nv_no_cong_cnt = 0;
432 			}
433 		} else {
434 			/* cwnd is in-between, so do nothing */
435 			return;
436 		}
437 
438 		/* update state */
439 		ca->nv_eval_call_cnt = 0;
440 		ca->nv_rtt_cnt = 0;
441 		ca->nv_rtt_max_rate = 0;
442 
443 		/* Don't want to make cwnd < nv_min_cwnd
444 		 * (it wasn't before, if it is now is because nv
445 		 *  decreased it).
446 		 */
447 		if (tcp_snd_cwnd(tp) < nv_min_cwnd)
448 			tcp_snd_cwnd_set(tp, nv_min_cwnd);
449 	}
450 }
451 
452 /* Extract info for Tcp socket info provided via netlink */
453 static size_t tcpnv_get_info(struct sock *sk, u32 ext, int *attr,
454 			     union tcp_cc_info *info)
455 {
456 	const struct tcpnv *ca = inet_csk_ca(sk);
457 
458 	if (ext & (1 << (INET_DIAG_VEGASINFO - 1))) {
459 		info->vegas.tcpv_enabled = 1;
460 		info->vegas.tcpv_rttcnt = ca->nv_rtt_cnt;
461 		info->vegas.tcpv_rtt = ca->nv_last_rtt;
462 		info->vegas.tcpv_minrtt = ca->nv_min_rtt;
463 
464 		*attr = INET_DIAG_VEGASINFO;
465 		return sizeof(struct tcpvegas_info);
466 	}
467 	return 0;
468 }
469 
470 static struct tcp_congestion_ops tcpnv __read_mostly = {
471 	.init		= tcpnv_init,
472 	.ssthresh	= tcpnv_recalc_ssthresh,
473 	.cong_avoid	= tcpnv_cong_avoid,
474 	.set_state	= tcpnv_state,
475 	.undo_cwnd	= tcp_reno_undo_cwnd,
476 	.pkts_acked     = tcpnv_acked,
477 	.get_info	= tcpnv_get_info,
478 
479 	.owner		= THIS_MODULE,
480 	.name		= "nv",
481 };
482 
483 static int __init tcpnv_register(void)
484 {
485 	BUILD_BUG_ON(sizeof(struct tcpnv) > ICSK_CA_PRIV_SIZE);
486 
487 	return tcp_register_congestion_control(&tcpnv);
488 }
489 
490 static void __exit tcpnv_unregister(void)
491 {
492 	tcp_unregister_congestion_control(&tcpnv);
493 }
494 
495 module_init(tcpnv_register);
496 module_exit(tcpnv_unregister);
497 
498 MODULE_AUTHOR("Lawrence Brakmo");
499 MODULE_LICENSE("GPL");
500 MODULE_DESCRIPTION("TCP NV");
501 MODULE_VERSION("1.0");
502