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 #include <sys/param.h>
41 #include <sys/kernel.h>
42 #include <sys/malloc.h>
43 #include <sys/module.h>
44 #include <sys/socket.h>
45 #include <sys/socketvar.h>
46 #include <sys/sysctl.h>
47 #include <sys/systm.h>
48
49 #include <net/vnet.h>
50
51 #include <net/route.h>
52 #include <net/route/nhop.h>
53
54 #include <netinet/in_pcb.h>
55 #include <netinet/tcp.h>
56 #include <netinet/tcp_seq.h>
57 #include <netinet/tcp_var.h>
58 #include <netinet/cc/cc.h>
59 #include <netinet/cc/cc_module.h>
60
61 #define DCTCP_SHIFT 10
62 #define MAX_ALPHA_VALUE (1<<DCTCP_SHIFT)
63 VNET_DEFINE_STATIC(uint32_t, dctcp_alpha) = MAX_ALPHA_VALUE;
64 #define V_dctcp_alpha VNET(dctcp_alpha)
65 VNET_DEFINE_STATIC(uint32_t, dctcp_shift_g) = 4;
66 #define V_dctcp_shift_g VNET(dctcp_shift_g)
67 VNET_DEFINE_STATIC(uint32_t, dctcp_slowstart) = 0;
68 #define V_dctcp_slowstart VNET(dctcp_slowstart)
69 VNET_DEFINE_STATIC(uint32_t, dctcp_ect1) = 0;
70 #define V_dctcp_ect1 VNET(dctcp_ect1)
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
dctcp_ack_received(struct cc_var * ccv,uint16_t type)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
dctcp_data_sz(void)168 dctcp_data_sz(void)
169 {
170 return (sizeof(struct dctcp));
171 }
172
173 static void
dctcp_after_idle(struct cc_var * ccv)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
dctcp_cb_destroy(struct cc_var * ccv)195 dctcp_cb_destroy(struct cc_var *ccv)
196 {
197 free(ccv->cc_data, M_CC_MEM);
198 }
199
200 static int
dctcp_cb_init(struct cc_var * ccv,void * ptr)201 dctcp_cb_init(struct cc_var *ccv, void *ptr)
202 {
203 struct dctcp *dctcp_data;
204
205 INP_WLOCK_ASSERT(tptoinpcb(ccv->ccvc.tcp));
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
dctcp_cong_signal(struct cc_var * ccv,uint32_t type)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
dctcp_conn_init(struct cc_var * ccv)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 if (V_dctcp_ect1)
319 CCV(ccv, t_flags2) |= TF2_ECN_USE_ECT1;
320 }
321 }
322
323 /*
324 * Perform any necessary tasks before we exit congestion recovery.
325 */
326 static void
dctcp_post_recovery(struct cc_var * ccv)327 dctcp_post_recovery(struct cc_var *ccv)
328 {
329 newreno_cc_post_recovery(ccv);
330
331 if (CCV(ccv, t_flags2) & TF2_ECN_PERMIT)
332 dctcp_update_alpha(ccv);
333 }
334
335 /*
336 * Execute an additional ECN processing using ECN field in IP header
337 * and the CWR bit in TCP header.
338 */
339 static void
dctcp_ecnpkt_handler(struct cc_var * ccv)340 dctcp_ecnpkt_handler(struct cc_var *ccv)
341 {
342 struct dctcp *dctcp_data;
343 uint32_t ccflag;
344 int acknow;
345
346 dctcp_data = ccv->cc_data;
347 ccflag = ccv->flags;
348 acknow = 0;
349
350 /*
351 * DCTCP responds with an ACK immediately when the CE state
352 * in between this segment and the last segment has changed.
353 */
354 if (ccflag & CCF_IPHDR_CE) {
355 if (!dctcp_data->ce_prev) {
356 acknow = 1;
357 dctcp_data->ce_prev = 1;
358 CCV(ccv, t_flags2) |= TF2_ECN_SND_ECE;
359 }
360 } else {
361 if (dctcp_data->ce_prev) {
362 acknow = 1;
363 dctcp_data->ce_prev = 0;
364 CCV(ccv, t_flags2) &= ~TF2_ECN_SND_ECE;
365 }
366 }
367
368 if ((acknow) || (ccflag & CCF_TCPHDR_CWR)) {
369 ccv->flags |= CCF_ACKNOW;
370 } else {
371 ccv->flags &= ~CCF_ACKNOW;
372 }
373 }
374
375 /*
376 * Update the fraction of marked bytes represented as 'alpha'.
377 * Also initialize several internal parameters at the end of this function.
378 */
379 static void
dctcp_update_alpha(struct cc_var * ccv)380 dctcp_update_alpha(struct cc_var *ccv)
381 {
382 struct dctcp *dctcp_data;
383 int alpha_prev;
384
385 dctcp_data = ccv->cc_data;
386 alpha_prev = dctcp_data->alpha;
387 dctcp_data->bytes_total = max(dctcp_data->bytes_total, 1);
388
389 /*
390 * Update alpha: alpha = (1 - g) * alpha + g * M.
391 * Here:
392 * g is weight factor
393 * recommaded to be set to 1/16
394 * small g = slow convergence between competitive DCTCP flows
395 * large g = impacts low utilization of bandwidth at switches
396 * M is fraction of marked segments in last RTT
397 * updated every RTT
398 * Alpha must be round to 0 - MAX_ALPHA_VALUE.
399 */
400 dctcp_data->alpha = ulmin(alpha_prev - (alpha_prev >> V_dctcp_shift_g) +
401 ((uint64_t)dctcp_data->bytes_ecn << (DCTCP_SHIFT - V_dctcp_shift_g)) /
402 dctcp_data->bytes_total, MAX_ALPHA_VALUE);
403
404 /* Initialize internal parameters for next alpha calculation */
405 dctcp_data->bytes_ecn = 0;
406 dctcp_data->bytes_total = 0;
407 dctcp_data->save_sndnxt = CCV(ccv, snd_nxt);
408 }
409
410 static int
dctcp_alpha_handler(SYSCTL_HANDLER_ARGS)411 dctcp_alpha_handler(SYSCTL_HANDLER_ARGS)
412 {
413 uint32_t new;
414 int error;
415
416 new = V_dctcp_alpha;
417 error = sysctl_handle_int(oidp, &new, 0, req);
418 if (error == 0 && req->newptr != NULL) {
419 if (new > MAX_ALPHA_VALUE)
420 error = EINVAL;
421 else
422 V_dctcp_alpha = new;
423 }
424
425 return (error);
426 }
427
428 static int
dctcp_shift_g_handler(SYSCTL_HANDLER_ARGS)429 dctcp_shift_g_handler(SYSCTL_HANDLER_ARGS)
430 {
431 uint32_t new;
432 int error;
433
434 new = V_dctcp_shift_g;
435 error = sysctl_handle_int(oidp, &new, 0, req);
436 if (error == 0 && req->newptr != NULL) {
437 if (new > DCTCP_SHIFT)
438 error = EINVAL;
439 else
440 V_dctcp_shift_g = new;
441 }
442
443 return (error);
444 }
445
446 static int
dctcp_slowstart_handler(SYSCTL_HANDLER_ARGS)447 dctcp_slowstart_handler(SYSCTL_HANDLER_ARGS)
448 {
449 uint32_t new;
450 int error;
451
452 new = V_dctcp_slowstart;
453 error = sysctl_handle_int(oidp, &new, 0, req);
454 if (error == 0 && req->newptr != NULL) {
455 if (new > 1)
456 error = EINVAL;
457 else
458 V_dctcp_slowstart = new;
459 }
460
461 return (error);
462 }
463
464 SYSCTL_DECL(_net_inet_tcp_cc_dctcp);
465 SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, dctcp,
466 CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
467 "dctcp congestion control related settings");
468
469 SYSCTL_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, alpha,
470 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
471 &VNET_NAME(dctcp_alpha), 0, &dctcp_alpha_handler, "IU",
472 "dctcp alpha parameter at start of session");
473
474 SYSCTL_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, shift_g,
475 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
476 &VNET_NAME(dctcp_shift_g), 4, &dctcp_shift_g_handler, "IU",
477 "dctcp shift parameter");
478
479 SYSCTL_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, slowstart,
480 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
481 &VNET_NAME(dctcp_slowstart), 0, &dctcp_slowstart_handler, "IU",
482 "half CWND reduction after the first slow start");
483
484 SYSCTL_UINT(_net_inet_tcp_cc_dctcp, OID_AUTO, ect1,
485 CTLFLAG_VNET | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
486 &VNET_NAME(dctcp_ect1), 0,
487 "Send DCTCP segments with ÍP ECT(0) or ECT(1)");
488
489 DECLARE_CC_MODULE(dctcp, &dctcp_cc_algo);
490 MODULE_VERSION(dctcp, 2);
491