xref: /freebsd/sys/netinet/cc/cc_dctcp.c (revision 814bd1ed438f7dfc5bedcb1f3e772a46fe7026bb)
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 <net/route.h>
54 #include <net/route/nhop.h>
55 
56 #include <netinet/in_pcb.h>
57 #include <netinet/tcp.h>
58 #include <netinet/tcp_seq.h>
59 #include <netinet/tcp_var.h>
60 #include <netinet/cc/cc.h>
61 #include <netinet/cc/cc_module.h>
62 
63 #define DCTCP_SHIFT 10
64 #define MAX_ALPHA_VALUE (1<<DCTCP_SHIFT)
65 VNET_DEFINE_STATIC(uint32_t, dctcp_alpha) = MAX_ALPHA_VALUE;
66 #define V_dctcp_alpha	    VNET(dctcp_alpha)
67 VNET_DEFINE_STATIC(uint32_t, dctcp_shift_g) = 4;
68 #define	V_dctcp_shift_g	    VNET(dctcp_shift_g)
69 VNET_DEFINE_STATIC(uint32_t, dctcp_slowstart) = 0;
70 #define	V_dctcp_slowstart   VNET(dctcp_slowstart)
71 
72 struct dctcp {
73 	uint32_t bytes_ecn;	  /* # of marked bytes during a RTT */
74 	uint32_t bytes_total;	  /* # of acked bytes during a RTT */
75 	int      alpha;		  /* the fraction of marked bytes */
76 	int      ce_prev;	  /* CE state of the last segment */
77 	tcp_seq  save_sndnxt;	  /* end sequence number of the current window */
78 	int      ece_curr;	  /* ECE flag in this segment */
79 	int      ece_prev;	  /* ECE flag in the last segment */
80 	uint32_t num_cong_events; /* # of congestion events */
81 };
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, void *ptr);
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 static size_t	dctcp_data_sz(void);
93 
94 struct cc_algo dctcp_cc_algo = {
95 	.name = "dctcp",
96 	.ack_received = dctcp_ack_received,
97 	.cb_destroy = dctcp_cb_destroy,
98 	.cb_init = dctcp_cb_init,
99 	.cong_signal = dctcp_cong_signal,
100 	.conn_init = dctcp_conn_init,
101 	.post_recovery = dctcp_post_recovery,
102 	.ecnpkt_handler = dctcp_ecnpkt_handler,
103 	.after_idle = dctcp_after_idle,
104 	.cc_data_sz = dctcp_data_sz,
105 };
106 
107 static void
108 dctcp_ack_received(struct cc_var *ccv, uint16_t type)
109 {
110 	struct dctcp *dctcp_data;
111 	int bytes_acked = 0;
112 
113 	dctcp_data = ccv->cc_data;
114 
115 	if (CCV(ccv, t_flags2) & TF2_ECN_PERMIT) {
116 		/*
117 		 * DCTCP doesn't treat receipt of ECN marked packet as a
118 		 * congestion event. Thus, DCTCP always executes the ACK
119 		 * processing out of congestion recovery.
120 		 */
121 		if (IN_CONGRECOVERY(CCV(ccv, t_flags))) {
122 			EXIT_CONGRECOVERY(CCV(ccv, t_flags));
123 			newreno_cc_ack_received(ccv, type);
124 			ENTER_CONGRECOVERY(CCV(ccv, t_flags));
125 		} else
126 			newreno_cc_ack_received(ccv, type);
127 
128 		if (type == CC_DUPACK)
129 			bytes_acked = min(ccv->bytes_this_ack, CCV(ccv, t_maxseg));
130 
131 		if (type == CC_ACK)
132 			bytes_acked = ccv->bytes_this_ack;
133 
134 		/* Update total bytes. */
135 		dctcp_data->bytes_total += bytes_acked;
136 
137 		/* Update total marked bytes. */
138 		if (dctcp_data->ece_curr) {
139 			//XXRMS: For fluid-model DCTCP, update
140 			//cwnd here during for RTT fairness
141 			if (!dctcp_data->ece_prev
142 			    && bytes_acked > CCV(ccv, t_maxseg)) {
143 				dctcp_data->bytes_ecn +=
144 				    (bytes_acked - CCV(ccv, t_maxseg));
145 			} else
146 				dctcp_data->bytes_ecn += bytes_acked;
147 			dctcp_data->ece_prev = 1;
148 		} else {
149 			if (dctcp_data->ece_prev
150 			    && bytes_acked > CCV(ccv, t_maxseg))
151 				dctcp_data->bytes_ecn += CCV(ccv, t_maxseg);
152 			dctcp_data->ece_prev = 0;
153 		}
154 		dctcp_data->ece_curr = 0;
155 
156 		/*
157 		 * Update the fraction of marked bytes at the end of
158 		 * current window size.
159 		 */
160 		if (!IN_FASTRECOVERY(CCV(ccv, t_flags)) &&
161 		    SEQ_GT(ccv->curack, dctcp_data->save_sndnxt))
162 			dctcp_update_alpha(ccv);
163 	} else
164 		newreno_cc_ack_received(ccv, type);
165 }
166 
167 static size_t
168 dctcp_data_sz(void)
169 {
170 	return (sizeof(struct dctcp));
171 }
172 
173 static void
174 dctcp_after_idle(struct cc_var *ccv)
175 {
176 	struct dctcp *dctcp_data;
177 
178 	if (CCV(ccv, t_flags2) & TF2_ECN_PERMIT) {
179 		dctcp_data = ccv->cc_data;
180 
181 		/* Initialize internal parameters after idle time */
182 		dctcp_data->bytes_ecn = 0;
183 		dctcp_data->bytes_total = 0;
184 		dctcp_data->save_sndnxt = CCV(ccv, snd_nxt);
185 		dctcp_data->alpha = V_dctcp_alpha;
186 		dctcp_data->ece_curr = 0;
187 		dctcp_data->ece_prev = 0;
188 		dctcp_data->num_cong_events = 0;
189 	}
190 
191 	newreno_cc_after_idle(ccv);
192 }
193 
194 static void
195 dctcp_cb_destroy(struct cc_var *ccv)
196 {
197 	free(ccv->cc_data, M_CC_MEM);
198 }
199 
200 static int
201 dctcp_cb_init(struct cc_var *ccv, void *ptr)
202 {
203 	struct dctcp *dctcp_data;
204 
205 	INP_WLOCK_ASSERT(ccv->ccvc.tcp->t_inpcb);
206 	if (ptr == NULL) {
207 		dctcp_data = malloc(sizeof(struct dctcp), M_CC_MEM, M_NOWAIT|M_ZERO);
208 		if (dctcp_data == NULL)
209 			return (ENOMEM);
210 	} else
211 		dctcp_data = ptr;
212 	/* Initialize some key variables with sensible defaults. */
213 	dctcp_data->bytes_ecn = 0;
214 	dctcp_data->bytes_total = 0;
215 	/*
216 	 * When alpha is set to 0 in the beginning, DCTCP sender transfers as
217 	 * much data as possible until the value converges which may expand the
218 	 * queueing delay at the switch. When alpha is set to 1, queueing delay
219 	 * is kept small.
220 	 * Throughput-sensitive applications should have alpha = 0
221 	 * Latency-sensitive applications should have alpha = 1
222 	 *
223 	 * Note: DCTCP draft suggests initial alpha to be 1 but we've decided to
224 	 * keep it 0 as default.
225 	 */
226 	dctcp_data->alpha = V_dctcp_alpha;
227 	dctcp_data->save_sndnxt = 0;
228 	dctcp_data->ce_prev = 0;
229 	dctcp_data->ece_curr = 0;
230 	dctcp_data->ece_prev = 0;
231 	dctcp_data->num_cong_events = 0;
232 
233 	ccv->cc_data = dctcp_data;
234 	return (0);
235 }
236 
237 /*
238  * Perform any necessary tasks before we enter congestion recovery.
239  */
240 static void
241 dctcp_cong_signal(struct cc_var *ccv, uint32_t type)
242 {
243 	struct dctcp *dctcp_data;
244 	u_int cwin, mss;
245 
246 	if (CCV(ccv, t_flags2) & TF2_ECN_PERMIT) {
247 		dctcp_data = ccv->cc_data;
248 		cwin = CCV(ccv, snd_cwnd);
249 		mss = tcp_maxseg(ccv->ccvc.tcp);
250 
251 		switch (type) {
252 		case CC_NDUPACK:
253 			if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) {
254 				if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
255 					CCV(ccv, snd_ssthresh) =
256 					    max(cwin / 2, 2 * mss);
257 					dctcp_data->num_cong_events++;
258 				} else {
259 					/* cwnd has already updated as congestion
260 					 * recovery. Reverse cwnd value using
261 					 * snd_cwnd_prev and recalculate snd_ssthresh
262 					 */
263 					cwin = CCV(ccv, snd_cwnd_prev);
264 					CCV(ccv, snd_ssthresh) =
265 					    max(cwin / 2, 2 * mss);
266 				}
267 				ENTER_RECOVERY(CCV(ccv, t_flags));
268 			}
269 			break;
270 		case CC_ECN:
271 			/*
272 			 * Save current snd_cwnd when the host encounters both
273 			 * congestion recovery and fast recovery.
274 			 */
275 			CCV(ccv, snd_cwnd_prev) = cwin;
276 			if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) {
277 				if (V_dctcp_slowstart &&
278 				    dctcp_data->num_cong_events++ == 0) {
279 					CCV(ccv, snd_ssthresh) =
280 					    max(cwin / 2, 2 * mss);
281 					dctcp_data->alpha = MAX_ALPHA_VALUE;
282 					dctcp_data->bytes_ecn = 0;
283 					dctcp_data->bytes_total = 0;
284 					dctcp_data->save_sndnxt = CCV(ccv, snd_nxt);
285 				} else
286 					CCV(ccv, snd_ssthresh) =
287 					    max((cwin - (((uint64_t)cwin *
288 					    dctcp_data->alpha) >> (DCTCP_SHIFT+1))),
289 					    2 * mss);
290 				CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh);
291 				ENTER_CONGRECOVERY(CCV(ccv, t_flags));
292 			}
293 			dctcp_data->ece_curr = 1;
294 			break;
295 		case CC_RTO:
296 			CCV(ccv, snd_ssthresh) = max(min(CCV(ccv, snd_wnd),
297 							 CCV(ccv, snd_cwnd)) / 2 / mss,
298 						     2) * mss;
299 			CCV(ccv, snd_cwnd) = mss;
300 			dctcp_update_alpha(ccv);
301 			dctcp_data->save_sndnxt += CCV(ccv, t_maxseg);
302 			dctcp_data->num_cong_events++;
303 			break;
304 		}
305 	} else
306 		newreno_cc_cong_signal(ccv, type);
307 }
308 
309 static void
310 dctcp_conn_init(struct cc_var *ccv)
311 {
312 	struct dctcp *dctcp_data;
313 
314 	dctcp_data = ccv->cc_data;
315 
316 	if (CCV(ccv, t_flags2) & TF2_ECN_PERMIT)
317 		dctcp_data->save_sndnxt = CCV(ccv, snd_nxt);
318 }
319 
320 /*
321  * Perform any necessary tasks before we exit congestion recovery.
322  */
323 static void
324 dctcp_post_recovery(struct cc_var *ccv)
325 {
326 	newreno_cc_post_recovery(ccv);
327 
328 	if (CCV(ccv, t_flags2) & TF2_ECN_PERMIT)
329 		dctcp_update_alpha(ccv);
330 }
331 
332 /*
333  * Execute an additional ECN processing using ECN field in IP header
334  * and the CWR bit in TCP header.
335  */
336 static void
337 dctcp_ecnpkt_handler(struct cc_var *ccv)
338 {
339 	struct dctcp *dctcp_data;
340 	uint32_t ccflag;
341 	int acknow;
342 
343 	dctcp_data = ccv->cc_data;
344 	ccflag = ccv->flags;
345 	acknow = 0;
346 
347 	/*
348 	 * DCTCP responds with an ACK immediately when the CE state
349 	 * in between this segment and the last segment has changed.
350 	 */
351 	if (ccflag & CCF_IPHDR_CE) {
352 		if (!dctcp_data->ce_prev) {
353 			acknow = 1;
354 			dctcp_data->ce_prev = 1;
355 			CCV(ccv, t_flags2) |= TF2_ECN_SND_ECE;
356 		}
357 	} else {
358 		if (dctcp_data->ce_prev) {
359 			acknow = 1;
360 			dctcp_data->ce_prev = 0;
361 			CCV(ccv, t_flags2) &= ~TF2_ECN_SND_ECE;
362 		}
363 	}
364 
365 	if ((acknow) || (ccflag & CCF_TCPHDR_CWR)) {
366 		ccv->flags |= CCF_ACKNOW;
367 	} else {
368 		ccv->flags &= ~CCF_ACKNOW;
369 	}
370 }
371 
372 /*
373  * Update the fraction of marked bytes represented as 'alpha'.
374  * Also initialize several internal parameters at the end of this function.
375  */
376 static void
377 dctcp_update_alpha(struct cc_var *ccv)
378 {
379 	struct dctcp *dctcp_data;
380 	int alpha_prev;
381 
382 	dctcp_data = ccv->cc_data;
383 	alpha_prev = dctcp_data->alpha;
384 	dctcp_data->bytes_total = max(dctcp_data->bytes_total, 1);
385 
386 	/*
387 	 * Update alpha: alpha = (1 - g) * alpha + g * M.
388 	 * Here:
389 	 * g is weight factor
390 	 *	recommaded to be set to 1/16
391 	 *	small g = slow convergence between competitive DCTCP flows
392 	 *	large g = impacts low utilization of bandwidth at switches
393 	 * M is fraction of marked segments in last RTT
394 	 *	updated every RTT
395 	 * Alpha must be round to 0 - MAX_ALPHA_VALUE.
396 	 */
397 	dctcp_data->alpha = ulmin(alpha_prev - (alpha_prev >> V_dctcp_shift_g) +
398 	    ((uint64_t)dctcp_data->bytes_ecn << (DCTCP_SHIFT - V_dctcp_shift_g)) /
399 	    dctcp_data->bytes_total, MAX_ALPHA_VALUE);
400 
401 	/* Initialize internal parameters for next alpha calculation */
402 	dctcp_data->bytes_ecn = 0;
403 	dctcp_data->bytes_total = 0;
404 	dctcp_data->save_sndnxt = CCV(ccv, snd_nxt);
405 }
406 
407 static int
408 dctcp_alpha_handler(SYSCTL_HANDLER_ARGS)
409 {
410 	uint32_t new;
411 	int error;
412 
413 	new = V_dctcp_alpha;
414 	error = sysctl_handle_int(oidp, &new, 0, req);
415 	if (error == 0 && req->newptr != NULL) {
416 		if (new > MAX_ALPHA_VALUE)
417 			error = EINVAL;
418 		else
419 			V_dctcp_alpha = new;
420 	}
421 
422 	return (error);
423 }
424 
425 static int
426 dctcp_shift_g_handler(SYSCTL_HANDLER_ARGS)
427 {
428 	uint32_t new;
429 	int error;
430 
431 	new = V_dctcp_shift_g;
432 	error = sysctl_handle_int(oidp, &new, 0, req);
433 	if (error == 0 && req->newptr != NULL) {
434 		if (new > DCTCP_SHIFT)
435 			error = EINVAL;
436 		else
437 			V_dctcp_shift_g = new;
438 	}
439 
440 	return (error);
441 }
442 
443 static int
444 dctcp_slowstart_handler(SYSCTL_HANDLER_ARGS)
445 {
446 	uint32_t new;
447 	int error;
448 
449 	new = V_dctcp_slowstart;
450 	error = sysctl_handle_int(oidp, &new, 0, req);
451 	if (error == 0 && req->newptr != NULL) {
452 		if (new > 1)
453 			error = EINVAL;
454 		else
455 			V_dctcp_slowstart = new;
456 	}
457 
458 	return (error);
459 }
460 
461 SYSCTL_DECL(_net_inet_tcp_cc_dctcp);
462 SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, dctcp,
463     CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
464     "dctcp congestion control related settings");
465 
466 SYSCTL_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, alpha,
467     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
468     &VNET_NAME(dctcp_alpha), 0, &dctcp_alpha_handler, "IU",
469     "dctcp alpha parameter at start of session");
470 
471 SYSCTL_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, shift_g,
472     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
473     &VNET_NAME(dctcp_shift_g), 4, &dctcp_shift_g_handler, "IU",
474     "dctcp shift parameter");
475 
476 SYSCTL_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, slowstart,
477     CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
478     &VNET_NAME(dctcp_slowstart), 0, &dctcp_slowstart_handler, "IU",
479     "half CWND reduction after the first slow start");
480 
481 DECLARE_CC_MODULE(dctcp, &dctcp_cc_algo);
482 MODULE_VERSION(dctcp, 2);
483