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 VNET_DEFINE_STATIC(uint32_t, dctcp_alpha) = 0; 63 #define V_dctcp_alpha VNET(dctcp_alpha) 64 VNET_DEFINE_STATIC(uint32_t, dctcp_shift_g) = 4; 65 #define V_dctcp_shift_g VNET(dctcp_shift_g) 66 VNET_DEFINE_STATIC(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 free(ccv->cc_data, M_dctcp); 188 } 189 190 static int 191 dctcp_cb_init(struct cc_var *ccv) 192 { 193 struct dctcp *dctcp_data; 194 195 dctcp_data = malloc(sizeof(struct dctcp), M_dctcp, M_NOWAIT|M_ZERO); 196 197 if (dctcp_data == NULL) 198 return (ENOMEM); 199 200 /* Initialize some key variables with sensible defaults. */ 201 dctcp_data->bytes_ecn = 0; 202 dctcp_data->bytes_total = 0; 203 /* 204 * When alpha is set to 0 in the beginning, DCTCP sender transfers as 205 * much data as possible until the value converges which may expand the 206 * queueing delay at the switch. When alpha is set to 1, queueing delay 207 * is kept small. 208 * Throughput-sensitive applications should have alpha = 0 209 * Latency-sensitive applications should have alpha = 1 210 * 211 * Note: DCTCP draft suggests initial alpha to be 1 but we've decided to 212 * keep it 0 as default. 213 */ 214 dctcp_data->alpha = V_dctcp_alpha; 215 dctcp_data->save_sndnxt = 0; 216 dctcp_data->ce_prev = 0; 217 dctcp_data->ece_curr = 0; 218 dctcp_data->ece_prev = 0; 219 dctcp_data->num_cong_events = 0; 220 221 ccv->cc_data = dctcp_data; 222 return (0); 223 } 224 225 /* 226 * Perform any necessary tasks before we enter congestion recovery. 227 */ 228 static void 229 dctcp_cong_signal(struct cc_var *ccv, uint32_t type) 230 { 231 struct dctcp *dctcp_data; 232 u_int win, mss; 233 234 dctcp_data = ccv->cc_data; 235 win = CCV(ccv, snd_cwnd); 236 mss = CCV(ccv, t_maxseg); 237 238 switch (type) { 239 case CC_NDUPACK: 240 if (!IN_FASTRECOVERY(CCV(ccv, t_flags))) { 241 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { 242 CCV(ccv, snd_ssthresh) = mss * 243 max(win / 2 / mss, 2); 244 dctcp_data->num_cong_events++; 245 } else { 246 /* cwnd has already updated as congestion 247 * recovery. Reverse cwnd value using 248 * snd_cwnd_prev and recalculate snd_ssthresh 249 */ 250 win = CCV(ccv, snd_cwnd_prev); 251 CCV(ccv, snd_ssthresh) = 252 max(win / 2 / mss, 2) * mss; 253 } 254 ENTER_RECOVERY(CCV(ccv, t_flags)); 255 } 256 break; 257 case CC_ECN: 258 /* 259 * Save current snd_cwnd when the host encounters both 260 * congestion recovery and fast recovery. 261 */ 262 CCV(ccv, snd_cwnd_prev) = win; 263 if (!IN_CONGRECOVERY(CCV(ccv, t_flags))) { 264 if (V_dctcp_slowstart && 265 dctcp_data->num_cong_events++ == 0) { 266 CCV(ccv, snd_ssthresh) = 267 mss * max(win / 2 / mss, 2); 268 dctcp_data->alpha = MAX_ALPHA_VALUE; 269 dctcp_data->bytes_ecn = 0; 270 dctcp_data->bytes_total = 0; 271 dctcp_data->save_sndnxt = CCV(ccv, snd_nxt); 272 } else 273 CCV(ccv, snd_ssthresh) = max((win - ((win * 274 dctcp_data->alpha) >> 11)) / mss, 2) * mss; 275 CCV(ccv, snd_cwnd) = CCV(ccv, snd_ssthresh); 276 ENTER_CONGRECOVERY(CCV(ccv, t_flags)); 277 } 278 dctcp_data->ece_curr = 1; 279 break; 280 case CC_RTO: 281 if (CCV(ccv, t_flags) & TF_ECN_PERMIT) { 282 CCV(ccv, t_flags) |= TF_ECN_SND_CWR; 283 dctcp_update_alpha(ccv); 284 dctcp_data->save_sndnxt += CCV(ccv, t_maxseg); 285 dctcp_data->num_cong_events++; 286 } 287 break; 288 } 289 } 290 291 static void 292 dctcp_conn_init(struct cc_var *ccv) 293 { 294 struct dctcp *dctcp_data; 295 296 dctcp_data = ccv->cc_data; 297 298 if (CCV(ccv, t_flags) & TF_ECN_PERMIT) 299 dctcp_data->save_sndnxt = CCV(ccv, snd_nxt); 300 } 301 302 /* 303 * Perform any necessary tasks before we exit congestion recovery. 304 */ 305 static void 306 dctcp_post_recovery(struct cc_var *ccv) 307 { 308 dctcp_cc_algo.post_recovery = newreno_cc_algo.post_recovery; 309 310 if (CCV(ccv, t_flags) & TF_ECN_PERMIT) 311 dctcp_update_alpha(ccv); 312 } 313 314 /* 315 * Execute an additional ECN processing using ECN field in IP header and the CWR 316 * bit in TCP header. 317 * 318 * delay_ack == 0 - Delayed ACK disabled 319 * delay_ack == 1 - Delayed ACK enabled 320 */ 321 322 static void 323 dctcp_ecnpkt_handler(struct cc_var *ccv) 324 { 325 struct dctcp *dctcp_data; 326 uint32_t ccflag; 327 int delay_ack; 328 329 dctcp_data = ccv->cc_data; 330 ccflag = ccv->flags; 331 delay_ack = 1; 332 333 /* 334 * DCTCP responses an ACK immediately when the CE state 335 * in between this segment and the last segment is not same. 336 */ 337 if (ccflag & CCF_IPHDR_CE) { 338 if (!dctcp_data->ce_prev && (ccflag & CCF_DELACK)) 339 delay_ack = 0; 340 dctcp_data->ce_prev = 1; 341 CCV(ccv, t_flags) |= TF_ECN_SND_ECE; 342 } else { 343 if (dctcp_data->ce_prev && (ccflag & CCF_DELACK)) 344 delay_ack = 0; 345 dctcp_data->ce_prev = 0; 346 CCV(ccv, t_flags) &= ~TF_ECN_SND_ECE; 347 } 348 349 /* DCTCP sets delayed ack when this segment sets the CWR flag. */ 350 if ((ccflag & CCF_DELACK) && (ccflag & CCF_TCPHDR_CWR)) 351 delay_ack = 1; 352 353 if (delay_ack == 0) 354 ccv->flags |= CCF_ACKNOW; 355 else 356 ccv->flags &= ~CCF_ACKNOW; 357 } 358 359 /* 360 * Update the fraction of marked bytes represented as 'alpha'. 361 * Also initialize several internal parameters at the end of this function. 362 */ 363 static void 364 dctcp_update_alpha(struct cc_var *ccv) 365 { 366 struct dctcp *dctcp_data; 367 int alpha_prev; 368 369 dctcp_data = ccv->cc_data; 370 alpha_prev = dctcp_data->alpha; 371 dctcp_data->bytes_total = max(dctcp_data->bytes_total, 1); 372 373 /* 374 * Update alpha: alpha = (1 - g) * alpha + g * F. 375 * Here: 376 * g is weight factor 377 * recommaded to be set to 1/16 378 * small g = slow convergence between competitive DCTCP flows 379 * large g = impacts low utilization of bandwidth at switches 380 * F is fraction of marked segments in last RTT 381 * updated every RTT 382 * Alpha must be round to 0 - MAX_ALPHA_VALUE. 383 */ 384 dctcp_data->alpha = min(alpha_prev - (alpha_prev >> V_dctcp_shift_g) + 385 (dctcp_data->bytes_ecn << (10 - V_dctcp_shift_g)) / 386 dctcp_data->bytes_total, MAX_ALPHA_VALUE); 387 388 /* Initialize internal parameters for next alpha calculation */ 389 dctcp_data->bytes_ecn = 0; 390 dctcp_data->bytes_total = 0; 391 dctcp_data->save_sndnxt = CCV(ccv, snd_nxt); 392 } 393 394 static int 395 dctcp_alpha_handler(SYSCTL_HANDLER_ARGS) 396 { 397 uint32_t new; 398 int error; 399 400 new = V_dctcp_alpha; 401 error = sysctl_handle_int(oidp, &new, 0, req); 402 if (error == 0 && req->newptr != NULL) { 403 if (CAST_PTR_INT(req->newptr) > 1) 404 error = EINVAL; 405 else { 406 if (new > MAX_ALPHA_VALUE) 407 V_dctcp_alpha = MAX_ALPHA_VALUE; 408 else 409 V_dctcp_alpha = new; 410 } 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 (CAST_PTR_INT(req->newptr) > 1) 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 (CAST_PTR_INT(req->newptr) > 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"); 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