xref: /freebsd/sys/netinet/cc/cc_dctcp.c (revision 62cfcf62f627e5093fb37026a6d8c98e4d2ef04c)
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 DCTCP_SHIFT 10
60 #define MAX_ALPHA_VALUE (1<<DCTCP_SHIFT)
61 VNET_DEFINE_STATIC(uint32_t, dctcp_alpha) = MAX_ALPHA_VALUE;
62 #define V_dctcp_alpha	    VNET(dctcp_alpha)
63 VNET_DEFINE_STATIC(uint32_t, dctcp_shift_g) = 4;
64 #define	V_dctcp_shift_g	    VNET(dctcp_shift_g)
65 VNET_DEFINE_STATIC(uint32_t, dctcp_slowstart) = 0;
66 #define	V_dctcp_slowstart   VNET(dctcp_slowstart)
67 
68 struct dctcp {
69 	uint32_t bytes_ecn;	  /* # of marked bytes during a RTT */
70 	uint32_t bytes_total;	  /* # of acked bytes during a RTT */
71 	int      alpha;		  /* the fraction of marked bytes */
72 	int      ce_prev;	  /* CE state of the last segment */
73 	tcp_seq  save_sndnxt;	  /* end sequence number of the current window */
74 	int      ece_curr;	  /* ECE flag in this segment */
75 	int      ece_prev;	  /* ECE flag in the last segment */
76 	uint32_t num_cong_events; /* # of congestion events */
77 };
78 
79 static MALLOC_DEFINE(M_dctcp, "dctcp data",
80     "Per connection data required for the dctcp algorithm");
81 
82 static void	dctcp_ack_received(struct cc_var *ccv, uint16_t type);
83 static void	dctcp_after_idle(struct cc_var *ccv);
84 static void	dctcp_cb_destroy(struct cc_var *ccv);
85 static int	dctcp_cb_init(struct cc_var *ccv);
86 static void	dctcp_cong_signal(struct cc_var *ccv, uint32_t type);
87 static void	dctcp_conn_init(struct cc_var *ccv);
88 static void	dctcp_post_recovery(struct cc_var *ccv);
89 static void	dctcp_ecnpkt_handler(struct cc_var *ccv);
90 static void	dctcp_update_alpha(struct cc_var *ccv);
91 
92 struct cc_algo dctcp_cc_algo = {
93 	.name = "dctcp",
94 	.ack_received = dctcp_ack_received,
95 	.cb_destroy = dctcp_cb_destroy,
96 	.cb_init = dctcp_cb_init,
97 	.cong_signal = dctcp_cong_signal,
98 	.conn_init = dctcp_conn_init,
99 	.post_recovery = dctcp_post_recovery,
100 	.ecnpkt_handler = dctcp_ecnpkt_handler,
101 	.after_idle = dctcp_after_idle,
102 };
103 
104 static void
105 dctcp_ack_received(struct cc_var *ccv, uint16_t type)
106 {
107 	struct dctcp *dctcp_data;
108 	int bytes_acked = 0;
109 
110 	dctcp_data = ccv->cc_data;
111 
112 	if (CCV(ccv, t_flags2) & TF2_ECN_PERMIT) {
113 		/*
114 		 * DCTCP doesn't treat receipt of ECN marked packet as a
115 		 * congestion event. Thus, DCTCP always executes the ACK
116 		 * processing out of congestion recovery.
117 		 */
118 		if (IN_CONGRECOVERY(CCV(ccv, t_flags))) {
119 			EXIT_CONGRECOVERY(CCV(ccv, t_flags));
120 			newreno_cc_algo.ack_received(ccv, type);
121 			ENTER_CONGRECOVERY(CCV(ccv, t_flags));
122 		} else
123 			newreno_cc_algo.ack_received(ccv, type);
124 
125 		if (type == CC_DUPACK)
126 			bytes_acked = min(ccv->bytes_this_ack, CCV(ccv, t_maxseg));
127 
128 		if (type == CC_ACK)
129 			bytes_acked = ccv->bytes_this_ack;
130 
131 		/* Update total bytes. */
132 		dctcp_data->bytes_total += bytes_acked;
133 
134 		/* Update total marked bytes. */
135 		if (dctcp_data->ece_curr) {
136 			//XXRMS: For fluid-model DCTCP, update
137 			//cwnd here during for RTT fairness
138 			if (!dctcp_data->ece_prev
139 			    && bytes_acked > CCV(ccv, t_maxseg)) {
140 				dctcp_data->bytes_ecn +=
141 				    (bytes_acked - CCV(ccv, t_maxseg));
142 			} else
143 				dctcp_data->bytes_ecn += bytes_acked;
144 			dctcp_data->ece_prev = 1;
145 		} else {
146 			if (dctcp_data->ece_prev
147 			    && bytes_acked > CCV(ccv, t_maxseg))
148 				dctcp_data->bytes_ecn += CCV(ccv, t_maxseg);
149 			dctcp_data->ece_prev = 0;
150 		}
151 		dctcp_data->ece_curr = 0;
152 
153 		/*
154 		 * Update the fraction of marked bytes at the end of
155 		 * current window size.
156 		 */
157 		if (!IN_FASTRECOVERY(CCV(ccv, t_flags)) &&
158 		    SEQ_GT(ccv->curack, dctcp_data->save_sndnxt))
159 			dctcp_update_alpha(ccv);
160 	} else
161 		newreno_cc_algo.ack_received(ccv, type);
162 }
163 
164 static void
165 dctcp_after_idle(struct cc_var *ccv)
166 {
167 	struct dctcp *dctcp_data;
168 
169 	if (CCV(ccv, t_flags2) & TF2_ECN_PERMIT) {
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 
182 	newreno_cc_algo.after_idle(ccv);
183 }
184 
185 static void
186 dctcp_cb_destroy(struct cc_var *ccv)
187 {
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 	u_int cwin, mss;
234 
235 	if (CCV(ccv, t_flags2) & TF2_ECN_PERMIT) {
236 		dctcp_data = ccv->cc_data;
237 		cwin = CCV(ccv, snd_cwnd);
238 		mss = CCV(ccv, t_maxseg);
239 
240 		switch (type) {
241 		case CC_NDUPACK:
242 			if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) {
243 				if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
244 					CCV(ccv, snd_ssthresh) =
245 					    max(cwin / 2, 2 * mss);
246 					dctcp_data->num_cong_events++;
247 				} else {
248 					/* cwnd has already updated as congestion
249 					 * recovery. Reverse cwnd value using
250 					 * snd_cwnd_prev and recalculate snd_ssthresh
251 					 */
252 					cwin = CCV(ccv, snd_cwnd_prev);
253 					CCV(ccv, snd_ssthresh) =
254 					    max(cwin / 2, 2 * mss);
255 				}
256 				ENTER_RECOVERY(CCV(ccv, t_flags));
257 			}
258 			break;
259 		case CC_ECN:
260 			/*
261 			 * Save current snd_cwnd when the host encounters both
262 			 * congestion recovery and fast recovery.
263 			 */
264 			CCV(ccv, snd_cwnd_prev) = cwin;
265 			if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
266 				if (V_dctcp_slowstart &&
267 				    dctcp_data->num_cong_events++ == 0) {
268 					CCV(ccv, snd_ssthresh) =
269 					    max(cwin / 2, 2 * mss);
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) =
276 					    max((cwin - (((uint64_t)cwin *
277 					    dctcp_data->alpha) >> (DCTCP_SHIFT+1))),
278 					    2 * mss);
279 				CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh);
280 				ENTER_CONGRECOVERY(CCV(ccv, t_flags));
281 			}
282 			dctcp_data->ece_curr = 1;
283 			break;
284 		case CC_RTO:
285 			dctcp_update_alpha(ccv);
286 			dctcp_data->save_sndnxt += CCV(ccv, t_maxseg);
287 			dctcp_data->num_cong_events++;
288 			break;
289 		}
290 	} else
291 		newreno_cc_algo.cong_signal(ccv, type);
292 }
293 
294 static void
295 dctcp_conn_init(struct cc_var *ccv)
296 {
297 	struct dctcp *dctcp_data;
298 
299 	dctcp_data = ccv->cc_data;
300 
301 	if (CCV(ccv, t_flags2) & TF2_ECN_PERMIT)
302 		dctcp_data->save_sndnxt = CCV(ccv, snd_nxt);
303 }
304 
305 /*
306  * Perform any necessary tasks before we exit congestion recovery.
307  */
308 static void
309 dctcp_post_recovery(struct cc_var *ccv)
310 {
311 	newreno_cc_algo.post_recovery(ccv);
312 
313 	if (CCV(ccv, t_flags2) & TF2_ECN_PERMIT)
314 		dctcp_update_alpha(ccv);
315 }
316 
317 /*
318  * Execute an additional ECN processing using ECN field in IP header
319  * and the CWR bit in TCP header.
320  */
321 static void
322 dctcp_ecnpkt_handler(struct cc_var *ccv)
323 {
324 	struct dctcp *dctcp_data;
325 	uint32_t ccflag;
326 	int acknow;
327 
328 	dctcp_data = ccv->cc_data;
329 	ccflag = ccv->flags;
330 	acknow = 0;
331 
332 	/*
333 	 * DCTCP responds with an ACK immediately when the CE state
334 	 * in between this segment and the last segment has changed.
335 	 */
336 	if (ccflag & CCF_IPHDR_CE) {
337 		if (!dctcp_data->ce_prev) {
338 			acknow = 1;
339 			dctcp_data->ce_prev = 1;
340 			CCV(ccv, t_flags2) |= TF2_ECN_SND_ECE;
341 		}
342 	} else {
343 		if (dctcp_data->ce_prev) {
344 			acknow = 1;
345 			dctcp_data->ce_prev = 0;
346 			CCV(ccv, t_flags2) &= ~TF2_ECN_SND_ECE;
347 		}
348 	}
349 
350 	if ((acknow) || (ccflag & CCF_TCPHDR_CWR)) {
351 		ccv->flags |= CCF_ACKNOW;
352 	} else {
353 		ccv->flags &= ~CCF_ACKNOW;
354 	}
355 }
356 
357 /*
358  * Update the fraction of marked bytes represented as 'alpha'.
359  * Also initialize several internal parameters at the end of this function.
360  */
361 static void
362 dctcp_update_alpha(struct cc_var *ccv)
363 {
364 	struct dctcp *dctcp_data;
365 	int alpha_prev;
366 
367 	dctcp_data = ccv->cc_data;
368 	alpha_prev = dctcp_data->alpha;
369 	dctcp_data->bytes_total = max(dctcp_data->bytes_total, 1);
370 
371 	/*
372 	 * Update alpha: alpha = (1 - g) * alpha + g * M.
373 	 * Here:
374 	 * g is weight factor
375 	 *	recommaded to be set to 1/16
376 	 *	small g = slow convergence between competitive DCTCP flows
377 	 *	large g = impacts low utilization of bandwidth at switches
378 	 * M is fraction of marked segments in last RTT
379 	 *	updated every RTT
380 	 * Alpha must be round to 0 - MAX_ALPHA_VALUE.
381 	 */
382 	dctcp_data->alpha = ulmin(alpha_prev - (alpha_prev >> V_dctcp_shift_g) +
383 	    ((uint64_t)dctcp_data->bytes_ecn << (DCTCP_SHIFT - V_dctcp_shift_g)) /
384 	    dctcp_data->bytes_total, MAX_ALPHA_VALUE);
385 
386 	/* Initialize internal parameters for next alpha calculation */
387 	dctcp_data->bytes_ecn = 0;
388 	dctcp_data->bytes_total = 0;
389 	dctcp_data->save_sndnxt = CCV(ccv, snd_nxt);
390 }
391 
392 static int
393 dctcp_alpha_handler(SYSCTL_HANDLER_ARGS)
394 {
395 	uint32_t new;
396 	int error;
397 
398 	new = V_dctcp_alpha;
399 	error = sysctl_handle_int(oidp, &new, 0, req);
400 	if (error == 0 && req->newptr != NULL) {
401 		if (new > MAX_ALPHA_VALUE)
402 			error = EINVAL;
403 		else
404 			V_dctcp_alpha = new;
405 	}
406 
407 	return (error);
408 }
409 
410 static int
411 dctcp_shift_g_handler(SYSCTL_HANDLER_ARGS)
412 {
413 	uint32_t new;
414 	int error;
415 
416 	new = V_dctcp_shift_g;
417 	error = sysctl_handle_int(oidp, &new, 0, req);
418 	if (error == 0 && req->newptr != NULL) {
419 		if (new > DCTCP_SHIFT)
420 			error = EINVAL;
421 		else
422 			V_dctcp_shift_g = new;
423 	}
424 
425 	return (error);
426 }
427 
428 static int
429 dctcp_slowstart_handler(SYSCTL_HANDLER_ARGS)
430 {
431 	uint32_t new;
432 	int error;
433 
434 	new = V_dctcp_slowstart;
435 	error = sysctl_handle_int(oidp, &new, 0, req);
436 	if (error == 0 && req->newptr != NULL) {
437 		if (new > 1)
438 			error = EINVAL;
439 		else
440 			V_dctcp_slowstart = new;
441 	}
442 
443 	return (error);
444 }
445 
446 SYSCTL_DECL(_net_inet_tcp_cc_dctcp);
447 SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, dctcp,
448     CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
449     "dctcp congestion control related settings");
450 
451 SYSCTL_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, alpha,
452     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
453     &VNET_NAME(dctcp_alpha), 0, &dctcp_alpha_handler, "IU",
454     "dctcp alpha parameter at start of session");
455 
456 SYSCTL_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, shift_g,
457     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
458     &VNET_NAME(dctcp_shift_g), 4, &dctcp_shift_g_handler, "IU",
459     "dctcp shift parameter");
460 
461 SYSCTL_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, slowstart,
462     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
463     &VNET_NAME(dctcp_slowstart), 0, &dctcp_slowstart_handler, "IU",
464     "half CWND reduction after the first slow start");
465 
466 DECLARE_CC_MODULE(dctcp, &dctcp_cc_algo);
467