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_GEQ(ccv->curack, CCV(ccv, snd_recover))) || 159 (!IN_FASTRECOVERY(CCV(ccv, t_flags)) && 160 SEQ_GT(ccv->curack, dctcp_data->save_sndnxt))) 161 dctcp_update_alpha(ccv); 162 } else 163 newreno_cc_algo.ack_received(ccv, type); 164 } 165 166 static void 167 dctcp_after_idle(struct cc_var *ccv) 168 { 169 struct dctcp *dctcp_data; 170 171 if (CCV(ccv, t_flags2) & TF2_ECN_PERMIT) { 172 dctcp_data = ccv->cc_data; 173 174 /* Initialize internal parameters after idle time */ 175 dctcp_data->bytes_ecn = 0; 176 dctcp_data->bytes_total = 0; 177 dctcp_data->save_sndnxt = CCV(ccv, snd_nxt); 178 dctcp_data->alpha = V_dctcp_alpha; 179 dctcp_data->ece_curr = 0; 180 dctcp_data->ece_prev = 0; 181 dctcp_data->num_cong_events = 0; 182 } 183 184 newreno_cc_algo.after_idle(ccv); 185 } 186 187 static void 188 dctcp_cb_destroy(struct cc_var *ccv) 189 { 190 free(ccv->cc_data, M_dctcp); 191 } 192 193 static int 194 dctcp_cb_init(struct cc_var *ccv) 195 { 196 struct dctcp *dctcp_data; 197 198 dctcp_data = malloc(sizeof(struct dctcp), M_dctcp, M_NOWAIT|M_ZERO); 199 200 if (dctcp_data == NULL) 201 return (ENOMEM); 202 203 /* Initialize some key variables with sensible defaults. */ 204 dctcp_data->bytes_ecn = 0; 205 dctcp_data->bytes_total = 0; 206 /* 207 * When alpha is set to 0 in the beginning, DCTCP sender transfers as 208 * much data as possible until the value converges which may expand the 209 * queueing delay at the switch. When alpha is set to 1, queueing delay 210 * is kept small. 211 * Throughput-sensitive applications should have alpha = 0 212 * Latency-sensitive applications should have alpha = 1 213 * 214 * Note: DCTCP draft suggests initial alpha to be 1 but we've decided to 215 * keep it 0 as default. 216 */ 217 dctcp_data->alpha = V_dctcp_alpha; 218 dctcp_data->save_sndnxt = 0; 219 dctcp_data->ce_prev = 0; 220 dctcp_data->ece_curr = 0; 221 dctcp_data->ece_prev = 0; 222 dctcp_data->num_cong_events = 0; 223 224 ccv->cc_data = dctcp_data; 225 return (0); 226 } 227 228 /* 229 * Perform any necessary tasks before we enter congestion recovery. 230 */ 231 static void 232 dctcp_cong_signal(struct cc_var *ccv, uint32_t type) 233 { 234 struct dctcp *dctcp_data; 235 u_int cwin, mss; 236 237 if (CCV(ccv, t_flags2) & TF2_ECN_PERMIT) { 238 dctcp_data = ccv->cc_data; 239 cwin = CCV(ccv, snd_cwnd); 240 mss = CCV(ccv, t_maxseg); 241 242 switch (type) { 243 case CC_NDUPACK: 244 if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) { 245 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { 246 CCV(ccv, snd_ssthresh) = 247 max(cwin / 2, 2 * mss); 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) = 256 max(cwin / 2, 2 * mss); 257 } 258 ENTER_RECOVERY(CCV(ccv, t_flags)); 259 } 260 break; 261 case CC_ECN: 262 /* 263 * Save current snd_cwnd when the host encounters both 264 * congestion recovery and fast recovery. 265 */ 266 CCV(ccv, snd_cwnd_prev) = cwin; 267 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { 268 if (V_dctcp_slowstart && 269 dctcp_data->num_cong_events++ == 0) { 270 CCV(ccv, snd_ssthresh) = 271 max(cwin / 2, 2 * mss); 272 dctcp_data->alpha = MAX_ALPHA_VALUE; 273 dctcp_data->bytes_ecn = 0; 274 dctcp_data->bytes_total = 0; 275 dctcp_data->save_sndnxt = CCV(ccv, snd_nxt); 276 } else 277 CCV(ccv, snd_ssthresh) = 278 max((cwin - (((uint64_t)cwin * 279 dctcp_data->alpha) >> (DCTCP_SHIFT+1))), 280 2 * mss); 281 CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh); 282 ENTER_CONGRECOVERY(CCV(ccv, t_flags)); 283 } 284 dctcp_data->ece_curr = 1; 285 break; 286 case CC_RTO: 287 CCV(ccv, t_flags2) |= TF2_ECN_SND_CWR; 288 dctcp_update_alpha(ccv); 289 dctcp_data->save_sndnxt += CCV(ccv, t_maxseg); 290 dctcp_data->num_cong_events++; 291 break; 292 } 293 } else 294 newreno_cc_algo.cong_signal(ccv, type); 295 } 296 297 static void 298 dctcp_conn_init(struct cc_var *ccv) 299 { 300 struct dctcp *dctcp_data; 301 302 dctcp_data = ccv->cc_data; 303 304 if (CCV(ccv, t_flags2) & TF2_ECN_PERMIT) 305 dctcp_data->save_sndnxt = CCV(ccv, snd_nxt); 306 } 307 308 /* 309 * Perform any necessary tasks before we exit congestion recovery. 310 */ 311 static void 312 dctcp_post_recovery(struct cc_var *ccv) 313 { 314 newreno_cc_algo.post_recovery(ccv); 315 316 if (CCV(ccv, t_flags2) & TF2_ECN_PERMIT) 317 dctcp_update_alpha(ccv); 318 } 319 320 /* 321 * Execute an additional ECN processing using ECN field in IP header and the CWR 322 * bit in TCP header. 323 * 324 * delay_ack == 0 - Delayed ACK disabled 325 * delay_ack == 1 - Delayed ACK enabled 326 */ 327 328 static void 329 dctcp_ecnpkt_handler(struct cc_var *ccv) 330 { 331 struct dctcp *dctcp_data; 332 uint32_t ccflag; 333 int delay_ack; 334 335 dctcp_data = ccv->cc_data; 336 ccflag = ccv->flags; 337 delay_ack = 1; 338 339 /* 340 * DCTCP responds with an ACK immediately when the CE state 341 * in between this segment and the last segment has changed. 342 */ 343 if (ccflag & CCF_IPHDR_CE) { 344 if (!dctcp_data->ce_prev && (ccflag & CCF_DELACK)) 345 delay_ack = 0; 346 dctcp_data->ce_prev = 1; 347 CCV(ccv, t_flags2) |= TF2_ECN_SND_ECE; 348 } else { 349 if (dctcp_data->ce_prev && (ccflag & CCF_DELACK)) 350 delay_ack = 0; 351 dctcp_data->ce_prev = 0; 352 CCV(ccv, t_flags2) &= ~TF2_ECN_SND_ECE; 353 } 354 355 /* DCTCP sets delayed ack when this segment sets the CWR flag. */ 356 if ((ccflag & CCF_DELACK) && (ccflag & CCF_TCPHDR_CWR)) 357 delay_ack = 1; 358 359 if (delay_ack == 0) 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 * M. 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 * M 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 = ulmin(alpha_prev - (alpha_prev >> V_dctcp_shift_g) + 389 ((uint64_t)dctcp_data->bytes_ecn << (DCTCP_SHIFT - 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 (new > MAX_ALPHA_VALUE) 408 error = EINVAL; 409 else 410 V_dctcp_alpha = new; 411 } 412 413 return (error); 414 } 415 416 static int 417 dctcp_shift_g_handler(SYSCTL_HANDLER_ARGS) 418 { 419 uint32_t new; 420 int error; 421 422 new = V_dctcp_shift_g; 423 error = sysctl_handle_int(oidp, &new, 0, req); 424 if (error == 0 && req->newptr != NULL) { 425 if (new > DCTCP_SHIFT) 426 error = EINVAL; 427 else 428 V_dctcp_shift_g = new; 429 } 430 431 return (error); 432 } 433 434 static int 435 dctcp_slowstart_handler(SYSCTL_HANDLER_ARGS) 436 { 437 uint32_t new; 438 int error; 439 440 new = V_dctcp_slowstart; 441 error = sysctl_handle_int(oidp, &new, 0, req); 442 if (error == 0 && req->newptr != NULL) { 443 if (new > 1) 444 error = EINVAL; 445 else 446 V_dctcp_slowstart = new; 447 } 448 449 return (error); 450 } 451 452 SYSCTL_DECL(_net_inet_tcp_cc_dctcp); 453 SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, dctcp, CTLFLAG_RW, NULL, 454 "dctcp congestion control related settings"); 455 456 SYSCTL_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, alpha, 457 CTLFLAG_VNET|CTLTYPE_UINT|CTLFLAG_RW, &VNET_NAME(dctcp_alpha), 0, 458 &dctcp_alpha_handler, 459 "IU", "dctcp alpha parameter at start of session"); 460 461 SYSCTL_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, shift_g, 462 CTLFLAG_VNET|CTLTYPE_UINT|CTLFLAG_RW, &VNET_NAME(dctcp_shift_g), 4, 463 &dctcp_shift_g_handler, 464 "IU", "dctcp shift parameter"); 465 466 SYSCTL_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, slowstart, 467 CTLFLAG_VNET|CTLTYPE_UINT|CTLFLAG_RW, &VNET_NAME(dctcp_slowstart), 0, 468 &dctcp_slowstart_handler, 469 "IU", "half CWND reduction after the first slow start"); 470 471 DECLARE_CC_MODULE(dctcp, &dctcp_cc_algo); 472