xref: /freebsd/sys/netinet/cc/cc_dctcp.c (revision 3416500aef140042c64bc149cb1ec6620483bc44)
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 beginning, 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 	uint32_t cwin, ssthresh_on_loss;
234 	u_int mss;
235 
236 	dctcp_data = ccv->cc_data;
237 	cwin = CCV(ccv, snd_cwnd);
238 	mss = CCV(ccv, t_maxseg);
239 	ssthresh_on_loss =
240 	    max((CCV(ccv, snd_max) - CCV(ccv, snd_una)) / 2 / mss, 2)
241 		* mss;
242 
243 	switch (type) {
244 	case CC_NDUPACK:
245 		if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) {
246 			if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
247 				CCV(ccv, snd_ssthresh) = ssthresh_on_loss;
248 				dctcp_data->num_cong_events++;
249 			} else {
250 				/* cwnd has already updated as congestion
251 				 * recovery. Reverse cwnd value using
252 				 * snd_cwnd_prev and recalculate snd_ssthresh
253 				 */
254 				cwin = CCV(ccv, snd_cwnd_prev);
255 				CCV(ccv, snd_ssthresh) = ssthresh_on_loss;
256 			}
257 			ENTER_RECOVERY(CCV(ccv, t_flags));
258 		}
259 		break;
260 	case CC_ECN:
261 		/*
262 		 * Save current snd_cwnd when the host encounters both
263 		 * congestion recovery and fast recovery.
264 		 */
265 		CCV(ccv, snd_cwnd_prev) = cwin;
266 		if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
267 			if (V_dctcp_slowstart &&
268 			    dctcp_data->num_cong_events++ == 0) {
269 				CCV(ccv, snd_ssthresh) = ssthresh_on_loss;
270 				dctcp_data->alpha = MAX_ALPHA_VALUE;
271 				dctcp_data->bytes_ecn = 0;
272 				dctcp_data->bytes_total = 0;
273 				dctcp_data->save_sndnxt = CCV(ccv, snd_nxt);
274 			} else
275 				CCV(ccv, snd_ssthresh) = max((cwin - ((cwin *
276 				    dctcp_data->alpha) >> 11)) / mss, 2) * mss;
277 			CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh);
278 			ENTER_CONGRECOVERY(CCV(ccv, t_flags));
279 		}
280 		dctcp_data->ece_curr = 1;
281 		break;
282 	case CC_RTO:
283 		if (CCV(ccv, t_flags) & TF_ECN_PERMIT) {
284 			CCV(ccv, t_flags) |= TF_ECN_SND_CWR;
285 			dctcp_update_alpha(ccv);
286 			dctcp_data->save_sndnxt += CCV(ccv, t_maxseg);
287 			dctcp_data->num_cong_events++;
288 			CCV(ccv, snd_ssthresh) = ssthresh_on_loss;
289 			CCV(ccv, snd_cwnd) = mss;
290 		}
291 		break;
292 	}
293 }
294 
295 static void
296 dctcp_conn_init(struct cc_var *ccv)
297 {
298 	struct dctcp *dctcp_data;
299 
300 	dctcp_data = ccv->cc_data;
301 
302 	if (CCV(ccv, t_flags) & TF_ECN_PERMIT)
303 		dctcp_data->save_sndnxt = CCV(ccv, snd_nxt);
304 }
305 
306 /*
307  * Perform any necessary tasks before we exit congestion recovery.
308  */
309 static void
310 dctcp_post_recovery(struct cc_var *ccv)
311 {
312 	dctcp_cc_algo.post_recovery = newreno_cc_algo.post_recovery;
313 
314 	if (CCV(ccv, t_flags) & TF_ECN_PERMIT)
315 		dctcp_update_alpha(ccv);
316 }
317 
318 /*
319  * Execute an additional ECN processing using ECN field in IP header and the CWR
320  * bit in TCP header.
321  *
322  * delay_ack == 0 - Delayed ACK disabled
323  * delay_ack == 1 - Delayed ACK enabled
324  */
325 
326 static void
327 dctcp_ecnpkt_handler(struct cc_var *ccv)
328 {
329 	struct dctcp *dctcp_data;
330 	uint32_t ccflag;
331 	int delay_ack;
332 
333 	dctcp_data = ccv->cc_data;
334 	ccflag = ccv->flags;
335 	delay_ack = 1;
336 
337 	/*
338 	 * DCTCP responses an ACK immediately when the CE state
339 	 * in between this segment and the last segment is not same.
340 	 */
341 	if (ccflag & CCF_IPHDR_CE) {
342 		if (!dctcp_data->ce_prev && (ccflag & CCF_DELACK))
343 			delay_ack = 0;
344 		dctcp_data->ce_prev = 1;
345 		CCV(ccv, t_flags) |= TF_ECN_SND_ECE;
346 	} else {
347 		if (dctcp_data->ce_prev && (ccflag & CCF_DELACK))
348 			delay_ack = 0;
349 		dctcp_data->ce_prev = 0;
350 		CCV(ccv, t_flags) &= ~TF_ECN_SND_ECE;
351 	}
352 
353 	/* DCTCP sets delayed ack when this segment sets the CWR flag. */
354 	if ((ccflag & CCF_DELACK) && (ccflag & CCF_TCPHDR_CWR))
355 		delay_ack = 1;
356 
357 	if (delay_ack == 0)
358 		ccv->flags |= CCF_ACKNOW;
359 	else
360 		ccv->flags &= ~CCF_ACKNOW;
361 }
362 
363 /*
364  * Update the fraction of marked bytes represented as 'alpha'.
365  * Also initialize several internal parameters at the end of this function.
366  */
367 static void
368 dctcp_update_alpha(struct cc_var *ccv)
369 {
370 	struct dctcp *dctcp_data;
371 	int alpha_prev;
372 
373 	dctcp_data = ccv->cc_data;
374 	alpha_prev = dctcp_data->alpha;
375 	dctcp_data->bytes_total = max(dctcp_data->bytes_total, 1);
376 
377 	/*
378 	 * Update alpha: alpha = (1 - g) * alpha + g * F.
379 	 * Here:
380 	 * g is weight factor
381 	 *	recommaded to be set to 1/16
382 	 *	small g = slow convergence between competitive DCTCP flows
383 	 *	large g = impacts low utilization of bandwidth at switches
384 	 * F is fraction of marked segments in last RTT
385 	 *	updated every RTT
386 	 * Alpha must be round to 0 - MAX_ALPHA_VALUE.
387 	 */
388 	dctcp_data->alpha = min(alpha_prev - (alpha_prev >> V_dctcp_shift_g) +
389 	    (dctcp_data->bytes_ecn << (10 - V_dctcp_shift_g)) /
390 	    dctcp_data->bytes_total, MAX_ALPHA_VALUE);
391 
392 	/* Initialize internal parameters for next alpha calculation */
393 	dctcp_data->bytes_ecn = 0;
394 	dctcp_data->bytes_total = 0;
395 	dctcp_data->save_sndnxt = CCV(ccv, snd_nxt);
396 }
397 
398 static int
399 dctcp_alpha_handler(SYSCTL_HANDLER_ARGS)
400 {
401 	uint32_t new;
402 	int error;
403 
404 	new = V_dctcp_alpha;
405 	error = sysctl_handle_int(oidp, &new, 0, req);
406 	if (error == 0 && req->newptr != NULL) {
407 		if (CAST_PTR_INT(req->newptr) > 1)
408 			error = EINVAL;
409 		else {
410 			if (new > MAX_ALPHA_VALUE)
411 				V_dctcp_alpha = MAX_ALPHA_VALUE;
412 			else
413 				V_dctcp_alpha = new;
414 		}
415 	}
416 
417 	return (error);
418 }
419 
420 static int
421 dctcp_shift_g_handler(SYSCTL_HANDLER_ARGS)
422 {
423 	uint32_t new;
424 	int error;
425 
426 	new = V_dctcp_shift_g;
427 	error = sysctl_handle_int(oidp, &new, 0, req);
428 	if (error == 0 && req->newptr != NULL) {
429 		if (CAST_PTR_INT(req->newptr) > 1)
430 			error = EINVAL;
431 		else
432 			V_dctcp_shift_g = new;
433 	}
434 
435 	return (error);
436 }
437 
438 static int
439 dctcp_slowstart_handler(SYSCTL_HANDLER_ARGS)
440 {
441 	uint32_t new;
442 	int error;
443 
444 	new = V_dctcp_slowstart;
445 	error = sysctl_handle_int(oidp, &new, 0, req);
446 	if (error == 0 && req->newptr != NULL) {
447 		if (CAST_PTR_INT(req->newptr) > 1)
448 			error = EINVAL;
449 		else
450 			V_dctcp_slowstart = new;
451 	}
452 
453 	return (error);
454 }
455 
456 SYSCTL_DECL(_net_inet_tcp_cc_dctcp);
457 SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, dctcp, CTLFLAG_RW, NULL,
458     "dctcp congestion control related settings");
459 
460 SYSCTL_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, alpha,
461     CTLFLAG_VNET|CTLTYPE_UINT|CTLFLAG_RW, &VNET_NAME(dctcp_alpha), 0,
462     &dctcp_alpha_handler,
463     "IU", "dctcp alpha parameter");
464 
465 SYSCTL_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, shift_g,
466     CTLFLAG_VNET|CTLTYPE_UINT|CTLFLAG_RW, &VNET_NAME(dctcp_shift_g), 4,
467     &dctcp_shift_g_handler,
468     "IU", "dctcp shift parameter");
469 
470 SYSCTL_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, slowstart,
471     CTLFLAG_VNET|CTLTYPE_UINT|CTLFLAG_RW, &VNET_NAME(dctcp_slowstart), 0,
472     &dctcp_slowstart_handler,
473     "IU", "half CWND reduction after the first slow start");
474 
475 DECLARE_CC_MODULE(dctcp, &dctcp_cc_algo);
476