xref: /freebsd/sys/netinet/cc/cc_dctcp.c (revision 1f4bcc459a76b7aa664f3fd557684cd0ba6da352)
1 /*-
2  * Copyright (c) 2007-2008
3  *	Swinburne University of Technology, Melbourne, Australia
4  * Copyright (c) 2009-2010 Lawrence Stewart <lstewart@freebsd.org>
5  * Copyright (c) 2014 Midori Kato <katoon@sfc.wide.ad.jp>
6  * Copyright (c) 2014 The FreeBSD Foundation
7  * All rights reserved.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  */
30 
31 /*
32  * An implementation of the DCTCP algorithm for FreeBSD, based on
33  * "Data Center TCP (DCTCP)" by M. Alizadeh, A. Greenberg, D. A. Maltz,
34  * J. Padhye, P. Patel, B. Prabhakar, S. Sengupta, and M. Sridharan.,
35  * in ACM Conference on SIGCOMM 2010, New York, USA,
36  * Originally released as the contribution of Microsoft Research project.
37  */
38 
39 #include <sys/cdefs.h>
40 __FBSDID("$FreeBSD$");
41 
42 #include <sys/param.h>
43 #include <sys/kernel.h>
44 #include <sys/malloc.h>
45 #include <sys/module.h>
46 #include <sys/socket.h>
47 #include <sys/socketvar.h>
48 #include <sys/sysctl.h>
49 #include <sys/systm.h>
50 
51 #include <net/vnet.h>
52 
53 #include <netinet/tcp.h>
54 #include <netinet/tcp_seq.h>
55 #include <netinet/tcp_var.h>
56 #include <netinet/cc/cc.h>
57 #include <netinet/cc/cc_module.h>
58 
59 #define	CAST_PTR_INT(X)	(*((int*)(X)))
60 
61 #define MAX_ALPHA_VALUE 1024
62 static VNET_DEFINE(uint32_t, dctcp_alpha) = 0;
63 #define V_dctcp_alpha	    VNET(dctcp_alpha)
64 static VNET_DEFINE(uint32_t, dctcp_shift_g) = 4;
65 #define	V_dctcp_shift_g	    VNET(dctcp_shift_g)
66 static VNET_DEFINE(uint32_t, dctcp_slowstart) = 0;
67 #define	V_dctcp_slowstart   VNET(dctcp_slowstart)
68 
69 struct dctcp {
70 	int     bytes_ecn;	/* # of marked bytes during a RTT */
71 	int     bytes_total;	/* # of acked bytes during a RTT */
72 	int     alpha;		/* the fraction of marked bytes */
73 	int     ce_prev;	/* CE state of the last segment */
74 	int     save_sndnxt;	/* end sequence number of the current window */
75 	int	ece_curr;	/* ECE flag in this segment */
76 	int	ece_prev;	/* ECE flag in the last segment */
77 	uint32_t    num_cong_events; /* # of congestion events */
78 };
79 
80 static MALLOC_DEFINE(M_dctcp, "dctcp data",
81     "Per connection data required for the dctcp algorithm");
82 
83 static void	dctcp_ack_received(struct cc_var *ccv, uint16_t type);
84 static void	dctcp_after_idle(struct cc_var *ccv);
85 static void	dctcp_cb_destroy(struct cc_var *ccv);
86 static int	dctcp_cb_init(struct cc_var *ccv);
87 static void	dctcp_cong_signal(struct cc_var *ccv, uint32_t type);
88 static void	dctcp_conn_init(struct cc_var *ccv);
89 static void	dctcp_post_recovery(struct cc_var *ccv);
90 static void	dctcp_ecnpkt_handler(struct cc_var *ccv);
91 static void	dctcp_update_alpha(struct cc_var *ccv);
92 
93 struct cc_algo dctcp_cc_algo = {
94 	.name = "dctcp",
95 	.ack_received = dctcp_ack_received,
96 	.cb_destroy = dctcp_cb_destroy,
97 	.cb_init = dctcp_cb_init,
98 	.cong_signal = dctcp_cong_signal,
99 	.conn_init = dctcp_conn_init,
100 	.post_recovery = dctcp_post_recovery,
101 	.ecnpkt_handler = dctcp_ecnpkt_handler,
102 	.after_idle = dctcp_after_idle,
103 };
104 
105 static void
106 dctcp_ack_received(struct cc_var *ccv, uint16_t type)
107 {
108 	struct dctcp *dctcp_data;
109 	int bytes_acked = 0;
110 
111 	dctcp_data = ccv->cc_data;
112 
113 	if (CCV(ccv, t_flags) & TF_ECN_PERMIT) {
114 		/*
115 		 * DCTCP doesn't treat receipt of ECN marked packet as a
116 		 * congestion event. Thus, DCTCP always executes the ACK
117 		 * processing out of congestion recovery.
118 		 */
119 		if (IN_CONGRECOVERY(CCV(ccv, t_flags))) {
120 			EXIT_CONGRECOVERY(CCV(ccv, t_flags));
121 			newreno_cc_algo.ack_received(ccv, type);
122 			ENTER_CONGRECOVERY(CCV(ccv, t_flags));
123 		} else
124 			newreno_cc_algo.ack_received(ccv, type);
125 
126 		if (type == CC_DUPACK)
127 			bytes_acked = CCV(ccv, t_maxseg);
128 
129 		if (type == CC_ACK)
130 			bytes_acked = ccv->bytes_this_ack;
131 
132 		/* Update total bytes. */
133 		dctcp_data->bytes_total += bytes_acked;
134 
135 		/* Update total marked bytes. */
136 		if (dctcp_data->ece_curr) {
137 			if (!dctcp_data->ece_prev
138 			    && bytes_acked > CCV(ccv, t_maxseg)) {
139 				dctcp_data->bytes_ecn +=
140 				    (bytes_acked - CCV(ccv, t_maxseg));
141 			} else
142 				dctcp_data->bytes_ecn += bytes_acked;
143 			dctcp_data->ece_prev = 1;
144 		} else {
145 			if (dctcp_data->ece_prev
146 			    && bytes_acked > CCV(ccv, t_maxseg))
147 				dctcp_data->bytes_ecn += CCV(ccv, t_maxseg);
148 			dctcp_data->ece_prev = 0;
149 		}
150 		dctcp_data->ece_curr = 0;
151 
152 		/*
153 		 * Update the fraction of marked bytes at the end of
154 		 * current window size.
155 		 */
156 		if ((IN_FASTRECOVERY(CCV(ccv, t_flags)) &&
157 		    SEQ_GEQ(ccv->curack, CCV(ccv, snd_recover))) ||
158 		    (!IN_FASTRECOVERY(CCV(ccv, t_flags)) &&
159 		    SEQ_GT(ccv->curack, dctcp_data->save_sndnxt)))
160 			dctcp_update_alpha(ccv);
161 	} else
162 		newreno_cc_algo.ack_received(ccv, type);
163 }
164 
165 static void
166 dctcp_after_idle(struct cc_var *ccv)
167 {
168 	struct dctcp *dctcp_data;
169 
170 	dctcp_data = ccv->cc_data;
171 
172 	/* Initialize internal parameters after idle time */
173 	dctcp_data->bytes_ecn = 0;
174 	dctcp_data->bytes_total = 0;
175 	dctcp_data->save_sndnxt = CCV(ccv, snd_nxt);
176 	dctcp_data->alpha = V_dctcp_alpha;
177 	dctcp_data->ece_curr = 0;
178 	dctcp_data->ece_prev = 0;
179 	dctcp_data->num_cong_events = 0;
180 
181 	dctcp_cc_algo.after_idle = newreno_cc_algo.after_idle;
182 }
183 
184 static void
185 dctcp_cb_destroy(struct cc_var *ccv)
186 {
187 	if (ccv->cc_data != NULL)
188 		free(ccv->cc_data, M_dctcp);
189 }
190 
191 static int
192 dctcp_cb_init(struct cc_var *ccv)
193 {
194 	struct dctcp *dctcp_data;
195 
196 	dctcp_data = malloc(sizeof(struct dctcp), M_dctcp, M_NOWAIT|M_ZERO);
197 
198 	if (dctcp_data == NULL)
199 		return (ENOMEM);
200 
201 	/* Initialize some key variables with sensible defaults. */
202 	dctcp_data->bytes_ecn = 0;
203 	dctcp_data->bytes_total = 0;
204 	/*
205 	 * When alpha is set to 0 in the beggining, DCTCP sender transfers as
206 	 * much data as possible until the value converges which may expand the
207 	 * queueing delay at the switch. When alpha is set to 1, queueing delay
208 	 * is kept small.
209 	 * Throughput-sensitive applications should have alpha = 0
210 	 * Latency-sensitive applications should have alpha = 1
211 	 *
212 	 * Note: DCTCP draft suggests initial alpha to be 1 but we've decided to
213 	 * keep it 0 as default.
214 	 */
215 	dctcp_data->alpha = V_dctcp_alpha;
216 	dctcp_data->save_sndnxt = 0;
217 	dctcp_data->ce_prev = 0;
218 	dctcp_data->ece_curr = 0;
219 	dctcp_data->ece_prev = 0;
220 	dctcp_data->num_cong_events = 0;
221 
222 	ccv->cc_data = dctcp_data;
223 	return (0);
224 }
225 
226 /*
227  * Perform any necessary tasks before we enter congestion recovery.
228  */
229 static void
230 dctcp_cong_signal(struct cc_var *ccv, uint32_t type)
231 {
232 	struct dctcp *dctcp_data;
233 	u_int win, mss;
234 
235 	dctcp_data = ccv->cc_data;
236 	win = CCV(ccv, snd_cwnd);
237 	mss = CCV(ccv, t_maxseg);
238 
239 	switch (type) {
240 	case CC_NDUPACK:
241 		if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) {
242 			if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
243 				CCV(ccv, snd_ssthresh) = mss *
244 				    max(win / 2 / mss, 2);
245 				dctcp_data->num_cong_events++;
246 			} else {
247 				/* cwnd has already updated as congestion
248 				 * recovery. Reverse cwnd value using
249 				 * snd_cwnd_prev and recalculate snd_ssthresh
250 				 */
251 				win = CCV(ccv, snd_cwnd_prev);
252 				CCV(ccv, snd_ssthresh) =
253 				    max(win / 2 / mss, 2) * mss;
254 			}
255 			ENTER_RECOVERY(CCV(ccv, t_flags));
256 		}
257 		break;
258 	case CC_ECN:
259 		/*
260 		 * Save current snd_cwnd when the host encounters both
261 		 * congestion recovery and fast recovery.
262 		 */
263 		CCV(ccv, snd_cwnd_prev) = win;
264 		if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
265 			if (V_dctcp_slowstart &&
266 			    dctcp_data->num_cong_events++ == 0) {
267 				CCV(ccv, snd_ssthresh) =
268 				    mss * max(win / 2 / mss, 2);
269 				dctcp_data->alpha = MAX_ALPHA_VALUE;
270 				dctcp_data->bytes_ecn = 0;
271 				dctcp_data->bytes_total = 0;
272 				dctcp_data->save_sndnxt = CCV(ccv, snd_nxt);
273 			} else
274 				CCV(ccv, snd_ssthresh) = max((win - ((win *
275 				    dctcp_data->alpha) >> 11)) / mss, 2) * mss;
276 			CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh);
277 			ENTER_CONGRECOVERY(CCV(ccv, t_flags));
278 		}
279 		dctcp_data->ece_curr = 1;
280 		break;
281 	case CC_RTO:
282 		if (CCV(ccv, t_flags) & TF_ECN_PERMIT) {
283 			CCV(ccv, t_flags) |= TF_ECN_SND_CWR;
284 			dctcp_update_alpha(ccv);
285 			dctcp_data->save_sndnxt += CCV(ccv, t_maxseg);
286 			dctcp_data->num_cong_events++;
287 		}
288 		break;
289 	}
290 }
291 
292 static void
293 dctcp_conn_init(struct cc_var *ccv)
294 {
295 	struct dctcp *dctcp_data;
296 
297 	dctcp_data = ccv->cc_data;
298 
299 	if (CCV(ccv, t_flags) & TF_ECN_PERMIT)
300 		dctcp_data->save_sndnxt = CCV(ccv, snd_nxt);
301 }
302 
303 /*
304  * Perform any necessary tasks before we exit congestion recovery.
305  */
306 static void
307 dctcp_post_recovery(struct cc_var *ccv)
308 {
309 	dctcp_cc_algo.post_recovery = newreno_cc_algo.post_recovery;
310 
311 	if (CCV(ccv, t_flags) & TF_ECN_PERMIT)
312 		dctcp_update_alpha(ccv);
313 }
314 
315 /*
316  * Execute an additional ECN processing using ECN field in IP header and the CWR
317  * bit in TCP header.
318  *
319  * delay_ack == 0 - Delayed ACK disabled
320  * delay_ack == 1 - Delayed ACK enabled
321  */
322 
323 static void
324 dctcp_ecnpkt_handler(struct cc_var *ccv)
325 {
326 	struct dctcp *dctcp_data;
327 	uint32_t ccflag;
328 	int delay_ack;
329 
330 	dctcp_data = ccv->cc_data;
331 	ccflag = ccv->flags;
332 	delay_ack = 1;
333 
334 	/*
335 	 * DCTCP responses an ACK immediately when the CE state
336 	 * in between this segment and the last segment is not same.
337 	 */
338 	if (ccflag & CCF_IPHDR_CE) {
339 		if (!dctcp_data->ce_prev && (ccflag & CCF_DELACK))
340 			delay_ack = 0;
341 		dctcp_data->ce_prev = 1;
342 		CCV(ccv, t_flags) |= TF_ECN_SND_ECE;
343 	} else {
344 		if (dctcp_data->ce_prev && (ccflag & CCF_DELACK))
345 			delay_ack = 0;
346 		dctcp_data->ce_prev = 0;
347 		CCV(ccv, t_flags) &= ~TF_ECN_SND_ECE;
348 	}
349 
350 	/* DCTCP sets delayed ack when this segment sets the CWR flag. */
351 	if ((ccflag & CCF_DELACK) && (ccflag & CCF_TCPHDR_CWR))
352 		delay_ack = 1;
353 
354 	if (delay_ack == 0)
355 		ccv->flags |= CCF_ACKNOW;
356 	else
357 		ccv->flags &= ~CCF_ACKNOW;
358 }
359 
360 /*
361  * Update the fraction of marked bytes represented as 'alpha'.
362  * Also initialize several internal parameters at the end of this function.
363  */
364 static void
365 dctcp_update_alpha(struct cc_var *ccv)
366 {
367 	struct dctcp *dctcp_data;
368 	int alpha_prev;
369 
370 	dctcp_data = ccv->cc_data;
371 	alpha_prev = dctcp_data->alpha;
372 	dctcp_data->bytes_total = max(dctcp_data->bytes_total, 1);
373 
374 	/*
375 	 * Update alpha: alpha = (1 - g) * alpha + g * F.
376 	 * Here:
377 	 * g is weight factor
378 	 *	recommaded to be set to 1/16
379 	 *	small g = slow convergence between competitive DCTCP flows
380 	 *	large g = impacts low utilization of bandwidth at switches
381 	 * F is fraction of marked segments in last RTT
382 	 *	updated every RTT
383 	 * Alpha must be round to 0 - MAX_ALPHA_VALUE.
384 	 */
385 	dctcp_data->alpha = min(alpha_prev - (alpha_prev >> V_dctcp_shift_g) +
386 	    (dctcp_data->bytes_ecn << (10 - V_dctcp_shift_g)) /
387 	    dctcp_data->bytes_total, MAX_ALPHA_VALUE);
388 
389 	/* Initialize internal parameters for next alpha calculation */
390 	dctcp_data->bytes_ecn = 0;
391 	dctcp_data->bytes_total = 0;
392 	dctcp_data->save_sndnxt = CCV(ccv, snd_nxt);
393 }
394 
395 static int
396 dctcp_alpha_handler(SYSCTL_HANDLER_ARGS)
397 {
398 	uint32_t new;
399 	int error;
400 
401 	new = V_dctcp_alpha;
402 	error = sysctl_handle_int(oidp, &new, 0, req);
403 	if (error == 0 && req->newptr != NULL) {
404 		if (CAST_PTR_INT(req->newptr) > 1)
405 			error = EINVAL;
406 		else {
407 			if (new > MAX_ALPHA_VALUE)
408 				V_dctcp_alpha = MAX_ALPHA_VALUE;
409 			else
410 				V_dctcp_alpha = new;
411 		}
412 	}
413 
414 	return (error);
415 }
416 
417 static int
418 dctcp_shift_g_handler(SYSCTL_HANDLER_ARGS)
419 {
420 	uint32_t new;
421 	int error;
422 
423 	new = V_dctcp_shift_g;
424 	error = sysctl_handle_int(oidp, &new, 0, req);
425 	if (error == 0 && req->newptr != NULL) {
426 		if (CAST_PTR_INT(req->newptr) > 1)
427 			error = EINVAL;
428 		else
429 			V_dctcp_shift_g = new;
430 	}
431 
432 	return (error);
433 }
434 
435 static int
436 dctcp_slowstart_handler(SYSCTL_HANDLER_ARGS)
437 {
438 	uint32_t new;
439 	int error;
440 
441 	new = V_dctcp_slowstart;
442 	error = sysctl_handle_int(oidp, &new, 0, req);
443 	if (error == 0 && req->newptr != NULL) {
444 		if (CAST_PTR_INT(req->newptr) > 1)
445 			error = EINVAL;
446 		else
447 			V_dctcp_slowstart = new;
448 	}
449 
450 	return (error);
451 }
452 
453 SYSCTL_DECL(_net_inet_tcp_cc_dctcp);
454 SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, dctcp, CTLFLAG_RW, NULL,
455     "dctcp congestion control related settings");
456 
457 SYSCTL_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, alpha,
458     CTLFLAG_VNET|CTLTYPE_UINT|CTLFLAG_RW, &VNET_NAME(dctcp_alpha), 0,
459     &dctcp_alpha_handler,
460     "IU", "dctcp alpha parameter");
461 
462 SYSCTL_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, shift_g,
463     CTLFLAG_VNET|CTLTYPE_UINT|CTLFLAG_RW, &VNET_NAME(dctcp_shift_g), 4,
464     &dctcp_shift_g_handler,
465     "IU", "dctcp shift parameter");
466 
467 SYSCTL_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, slowstart,
468     CTLFLAG_VNET|CTLTYPE_UINT|CTLFLAG_RW, &VNET_NAME(dctcp_slowstart), 0,
469     &dctcp_slowstart_handler,
470     "IU", "half CWND reduction after the first slow start");
471 
472 DECLARE_CC_MODULE(dctcp, &dctcp_cc_algo);
473