xref: /xnu-11215.1.10/bsd/netinet/tcp_cc.c (revision 8d741a5de7ff4191bf97d57b9f54c2f6d4a15585)
1 /*
2  * Copyright (c) 2013-2021 Apple Inc. All rights reserved.
3  *
4  * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
5  *
6  * This file contains Original Code and/or Modifications of Original Code
7  * as defined in and that are subject to the Apple Public Source License
8  * Version 2.0 (the 'License'). You may not use this file except in
9  * compliance with the License. The rights granted to you under the License
10  * may not be used to create, or enable the creation or redistribution of,
11  * unlawful or unlicensed copies of an Apple operating system, or to
12  * circumvent, violate, or enable the circumvention or violation of, any
13  * terms of an Apple operating system software license agreement.
14  *
15  * Please obtain a copy of the License at
16  * http://www.opensource.apple.com/apsl/ and read it before using this file.
17  *
18  * The Original Code and all software distributed under the License are
19  * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
20  * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
21  * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
22  * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
23  * Please see the License for the specific language governing rights and
24  * limitations under the License.
25  *
26  * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
27  */
28 
29 #include "tcp_includes.h"
30 
31 #include <sys/param.h>
32 #include <sys/kernel.h>
33 #include <sys/syslog.h>
34 #include <sys/kern_control.h>
35 #include <sys/domain.h>
36 
37 #include <netinet/in.h>
38 #include <mach/sdt.h>
39 #include <libkern/OSAtomic.h>
40 
41 #include <libkern/OSTypes.h>
42 
43 extern struct tcp_cc_algo tcp_cc_newreno;
44 extern struct tcp_cc_algo tcp_cc_ledbat;
45 extern struct tcp_cc_algo tcp_cc_cubic;
46 extern struct tcp_cc_algo tcp_cc_prague;
47 
48  #define SET_SNDSB_IDEAL_SIZE(sndsb, size) \
49 	sndsb->sb_idealsize = min(max(tcp_sendspace, tp->snd_ssthresh), \
50 	tcp_autosndbuf_max);
51 
52 /* Array containing pointers to currently implemented TCP CC algorithms */
53 struct tcp_cc_algo* tcp_cc_algo_list[TCP_CC_ALGO_COUNT];
54 
55 static struct tcp_cc_algo tcp_cc_algo_none;
56 /*
57  * Initialize TCP congestion control algorithms.
58  */
59 
60 void
tcp_cc_init(void)61 tcp_cc_init(void)
62 {
63 	bzero(&tcp_cc_algo_list, sizeof(tcp_cc_algo_list));
64 	bzero(&tcp_cc_algo_none, sizeof(tcp_cc_algo_none));
65 
66 	tcp_cc_algo_list[TCP_CC_ALGO_NONE] = &tcp_cc_algo_none;
67 	tcp_cc_algo_list[TCP_CC_ALGO_NEWRENO_INDEX] = &tcp_cc_newreno;
68 	tcp_cc_algo_list[TCP_CC_ALGO_BACKGROUND_INDEX] = &tcp_cc_ledbat;
69 	tcp_cc_algo_list[TCP_CC_ALGO_CUBIC_INDEX] = &tcp_cc_cubic;
70 	tcp_cc_algo_list[TCP_CC_ALGO_PRAGUE_INDEX] = &tcp_cc_prague;
71 
72 	tcp_ccdbg_control_register();
73 }
74 
75 void
tcp_cc_resize_sndbuf(struct tcpcb * tp)76 tcp_cc_resize_sndbuf(struct tcpcb *tp)
77 {
78 	struct sockbuf *sb;
79 	/*
80 	 * If the send socket buffer size is bigger than ssthresh,
81 	 * it is time to trim it because we do not want to hold
82 	 * too many mbufs in the socket buffer
83 	 */
84 	sb = &tp->t_inpcb->inp_socket->so_snd;
85 	if (sb->sb_hiwat > tp->snd_ssthresh &&
86 	    (sb->sb_flags & SB_AUTOSIZE)) {
87 		if (sb->sb_idealsize > tp->snd_ssthresh) {
88 			SET_SNDSB_IDEAL_SIZE(sb, tp->snd_ssthresh);
89 		}
90 		sb->sb_flags |= SB_TRIM;
91 	}
92 }
93 
94 void
tcp_bad_rexmt_fix_sndbuf(struct tcpcb * tp)95 tcp_bad_rexmt_fix_sndbuf(struct tcpcb *tp)
96 {
97 	struct sockbuf *sb;
98 	sb = &tp->t_inpcb->inp_socket->so_snd;
99 	if ((sb->sb_flags & (SB_TRIM | SB_AUTOSIZE)) == (SB_TRIM | SB_AUTOSIZE)) {
100 		/*
101 		 * If there was a retransmission that was not necessary
102 		 * then the size of socket buffer can be restored to
103 		 * what it was before
104 		 */
105 		SET_SNDSB_IDEAL_SIZE(sb, tp->snd_ssthresh);
106 		if (sb->sb_hiwat <= sb->sb_idealsize) {
107 			sbreserve(sb, sb->sb_idealsize);
108 			sb->sb_flags &= ~SB_TRIM;
109 		}
110 	}
111 }
112 
113 /*
114  * Calculate initial cwnd according to RFC3390.
115  */
116 void
tcp_cc_cwnd_init_or_reset(struct tcpcb * tp)117 tcp_cc_cwnd_init_or_reset(struct tcpcb *tp)
118 {
119 	if (tp->t_flags & TF_LOCAL) {
120 		tp->snd_cwnd = tp->t_maxseg * ss_fltsz_local;
121 	} else {
122 		if (tcp_cubic_minor_fixes) {
123 			tp->snd_cwnd = tcp_initial_cwnd(tp);
124 		} else {
125 			/* initial congestion window according to RFC 3390 */
126 			tp->snd_cwnd = min(4 * tp->t_maxseg,
127 			    max(2 * tp->t_maxseg, TCP_CC_CWND_INIT_BYTES));
128 		}
129 	}
130 }
131 
132 /*
133  * Indicate whether this ack should be delayed.
134  * Here is the explanation for different settings of tcp_delack_enabled:
135  *  - when set to 1, the behavior is same as when set to 2. We kept this
136  *    for binary compatibility.
137  *  - when set to 2, will "ack every other packet"
138  *      - if our last ack wasn't a 0-sized window.
139  *      - if the peer hasn't sent us a TH_PUSH data packet (radar 3649245).
140  *              If TH_PUSH is set, take this as a clue that we need to ACK
141  *              with no delay. This helps higher level protocols who
142  *              won't send us more data even if the window is open
143  *              because their last "segment" hasn't been ACKed
144  *  - when set to 3,  will do "streaming detection"
145  *      - if we receive more than "maxseg_unacked" full packets
146  *        in the last 100ms
147  *      - if the connection is not in slow-start or idle or
148  *        loss/recovery states
149  *      - if those criteria aren't met, it will ack every other packet.
150  */
151 int
tcp_cc_delay_ack(struct tcpcb * tp,struct tcphdr * th)152 tcp_cc_delay_ack(struct tcpcb *tp, struct tcphdr *th)
153 {
154 	switch (tcp_delack_enabled) {
155 	case 1:
156 	case 2:
157 		if ((tp->t_flags & TF_RXWIN0SENT) == 0 &&
158 		    (th->th_flags & TH_PUSH) == 0 &&
159 		    (tp->t_unacksegs == 1)) {
160 			return 1;
161 		}
162 		break;
163 	case 3:
164 		if (tcp_ack_strategy == TCP_ACK_STRATEGY_LEGACY) {
165 			if ((tp->t_flags & TF_RXWIN0SENT) == 0 &&
166 			    (th->th_flags & TH_PUSH) == 0 &&
167 			    ((tp->t_unacksegs == 1) ||
168 			    ((tp->t_flags & TF_STRETCHACK) &&
169 			    tp->t_unacksegs < maxseg_unacked))) {
170 				return 1;
171 			}
172 		} else {
173 			uint32_t recwin;
174 
175 			/* Get the receive-window we would announce */
176 			recwin = tcp_sbspace(tp);
177 			if (recwin > (uint32_t)(TCP_MAXWIN << tp->rcv_scale)) {
178 				recwin = (uint32_t)(TCP_MAXWIN << tp->rcv_scale);
179 			}
180 
181 			/* Delay ACK, if:
182 			 *
183 			 * 1. We are not sending a zero-window
184 			 * 2. We are not forcing fast ACKs
185 			 * 3. We have more than the low-water mark in receive-buffer
186 			 * 4. The receive-window is not increasing
187 			 * 5. We have less than or equal of an MSS unacked or
188 			 *    Window actually has been growing larger than the initial value by half of it.
189 			 *    (this makes sure that during ramp-up we ACK every second MSS
190 			 *    until we pass the tcp_recvspace * 1.5-threshold)
191 			 * 6. We haven't waited for half a BDP
192 			 * 7. The amount of unacked data is less than the maximum ACK-burst (256 MSS)
193 			 *    We try to avoid having the sender end up hitting huge ACK-ranges.
194 			 *
195 			 * (a note on 6: The receive-window is
196 			 * roughly 2 BDP. Thus, recwin / 4 means half a BDP and
197 			 * thus we enforce an ACK roughly twice per RTT - even
198 			 * if the app does not read)
199 			 */
200 			if ((tp->t_flags & TF_RXWIN0SENT) == 0 &&
201 			    tp->t_forced_acks == 0 &&
202 			    tp->t_inpcb->inp_socket->so_rcv.sb_cc > tp->t_inpcb->inp_socket->so_rcv.sb_lowat &&
203 			    recwin <= tp->t_last_recwin &&
204 			    (tp->rcv_nxt - tp->last_ack_sent <= tp->t_maxseg ||
205 			    recwin > (uint32_t)(tcp_recvspace + (tcp_recvspace >> 1))) &&
206 			    (tp->rcv_nxt - tp->last_ack_sent) < (recwin >> 2) &&
207 			    (tp->rcv_nxt - tp->last_ack_sent) < 256 * tp->t_maxseg) {
208 				tp->t_stat.acks_delayed++;
209 				return 1;
210 			}
211 		}
212 		break;
213 	}
214 	return 0;
215 }
216 
217 void
tcp_cc_allocate_state(struct tcpcb * tp)218 tcp_cc_allocate_state(struct tcpcb *tp)
219 {
220 	if ((tp->tcp_cc_index == TCP_CC_ALGO_CUBIC_INDEX ||
221 	    tp->tcp_cc_index == TCP_CC_ALGO_PRAGUE_INDEX ||
222 	    tp->tcp_cc_index == TCP_CC_ALGO_BACKGROUND_INDEX) &&
223 	    tp->t_ccstate == NULL) {
224 		tp->t_ccstate = &tp->_t_ccstate;
225 
226 		bzero(tp->t_ccstate, sizeof(*tp->t_ccstate));
227 	}
228 }
229 
230 /*
231  * If stretch ack was disabled automatically on long standing connections,
232  * re-evaluate the situation after 15 minutes to enable it.
233  */
234 #define TCP_STRETCHACK_DISABLE_WIN (15 * 60 * TCP_RETRANSHZ)
235 void
tcp_cc_after_idle_stretchack(struct tcpcb * tp)236 tcp_cc_after_idle_stretchack(struct tcpcb *tp)
237 {
238 	struct tcp_globals *globals;
239 	int32_t tdiff;
240 
241 	if (!(tp->t_flagsext & TF_DISABLE_STRETCHACK)) {
242 		return;
243 	}
244 
245 	globals = tcp_get_globals(tp);
246 	tdiff = timer_diff(tcp_globals_now(globals), 0, tp->rcv_nostrack_ts, 0);
247 	if (tdiff < 0) {
248 		tdiff = -tdiff;
249 	}
250 
251 	if (tdiff > TCP_STRETCHACK_DISABLE_WIN) {
252 		tp->t_flagsext &= ~TF_DISABLE_STRETCHACK;
253 		tp->t_stretchack_delayed = 0;
254 
255 		tcp_reset_stretch_ack(tp);
256 	}
257 }
258 
259 /*
260  * Detect if the congestion window is non-validated according to
261  * draft-ietf-tcpm-newcwv-07
262  */
263 inline uint32_t
tcp_cc_is_cwnd_nonvalidated(struct tcpcb * tp)264 tcp_cc_is_cwnd_nonvalidated(struct tcpcb *tp)
265 {
266 	struct socket *so = tp->t_inpcb->inp_socket;
267 
268 	if (tp->t_pipeack == 0) {
269 		tp->t_flagsext &= ~TF_CWND_NONVALIDATED;
270 		return 0;
271 	}
272 
273 	/*
274 	 * The congestion window is validated if the number of bytes acked
275 	 * is more than half of the current window or if there is more
276 	 * data to send in the send socket buffer
277 	 */
278 	if (tp->t_pipeack >= (tp->snd_cwnd >> 1) ||
279 	    (so != NULL && so->so_snd.sb_cc > tp->snd_cwnd)) {
280 		tp->t_flagsext &= ~TF_CWND_NONVALIDATED;
281 	} else {
282 		tp->t_flagsext |= TF_CWND_NONVALIDATED;
283 	}
284 
285 	return tp->t_flagsext & TF_CWND_NONVALIDATED;
286 }
287 
288 /*
289  * Adjust congestion window in response to congestion in non-validated
290  * phase.
291  */
292 inline void
tcp_cc_adjust_nonvalidated_cwnd(struct tcpcb * tp)293 tcp_cc_adjust_nonvalidated_cwnd(struct tcpcb *tp)
294 {
295 	tp->t_pipeack = tcp_get_max_pipeack(tp);
296 	tcp_clear_pipeack_state(tp);
297 	tp->snd_cwnd = (max(tp->t_pipeack, tp->t_lossflightsize) >> 1);
298 	if (tcp_cubic_minor_fixes) {
299 		tp->snd_cwnd = max(tp->snd_cwnd, tp->t_maxseg);
300 	} else {
301 		tp->snd_cwnd = max(tp->snd_cwnd, TCP_CC_CWND_INIT_BYTES);
302 	}
303 	tp->snd_cwnd += tp->t_maxseg * tcprexmtthresh;
304 	tp->t_flagsext &= ~TF_CWND_NONVALIDATED;
305 }
306 
307 /*
308  * Return maximum of all the pipeack samples. Since the number of samples
309  * TCP_PIPEACK_SAMPLE_COUNT is 3 at this time, it will be simpler to do
310  * a comparision. We should change ths if the number of samples increases.
311  */
312 inline uint32_t
tcp_get_max_pipeack(struct tcpcb * tp)313 tcp_get_max_pipeack(struct tcpcb *tp)
314 {
315 	uint32_t max_pipeack = 0;
316 	max_pipeack = (tp->t_pipeack_sample[0] > tp->t_pipeack_sample[1]) ?
317 	    tp->t_pipeack_sample[0] : tp->t_pipeack_sample[1];
318 	max_pipeack = (tp->t_pipeack_sample[2] > max_pipeack) ?
319 	    tp->t_pipeack_sample[2] : max_pipeack;
320 
321 	return max_pipeack;
322 }
323 
324 inline void
tcp_clear_pipeack_state(struct tcpcb * tp)325 tcp_clear_pipeack_state(struct tcpcb *tp)
326 {
327 	bzero(tp->t_pipeack_sample, sizeof(tp->t_pipeack_sample));
328 	tp->t_pipeack_ind = 0;
329 	tp->t_lossflightsize = 0;
330 }
331