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/param.h> 40 #include <sys/kernel.h> 41 #include <sys/malloc.h> 42 #include <sys/module.h> 43 #include <sys/socket.h> 44 #include <sys/socketvar.h> 45 #include <sys/sysctl.h> 46 #include <sys/systm.h> 47 48 #include <net/vnet.h> 49 50 #include <net/route.h> 51 #include <net/route/nhop.h> 52 53 #include <netinet/in_pcb.h> 54 #include <netinet/tcp.h> 55 #include <netinet/tcp_seq.h> 56 #include <netinet/tcp_var.h> 57 #include <netinet/cc/cc.h> 58 #include <netinet/cc/cc_module.h> 59 60 #define DCTCP_SHIFT 10 61 #define MAX_ALPHA_VALUE (1<<DCTCP_SHIFT) 62 VNET_DEFINE_STATIC(uint32_t, dctcp_alpha) = MAX_ALPHA_VALUE; 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 VNET_DEFINE_STATIC(uint32_t, dctcp_ect1) = 0; 69 #define V_dctcp_ect1 VNET(dctcp_ect1) 70 71 struct dctcp { 72 uint32_t bytes_ecn; /* # of marked bytes during a RTT */ 73 uint32_t bytes_total; /* # of acked bytes during a RTT */ 74 int alpha; /* the fraction of marked bytes */ 75 int ce_prev; /* CE state of the last segment */ 76 tcp_seq save_sndnxt; /* end sequence number of the current window */ 77 int ece_curr; /* ECE flag in this segment */ 78 int ece_prev; /* ECE flag in the last segment */ 79 uint32_t num_cong_events; /* # of congestion events */ 80 }; 81 82 static void dctcp_ack_received(struct cc_var *ccv, ccsignal_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, void *ptr); 86 static void dctcp_cong_signal(struct cc_var *ccv, ccsignal_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 static size_t dctcp_data_sz(void); 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 .cc_data_sz = dctcp_data_sz, 104 }; 105 106 static void 107 dctcp_ack_received(struct cc_var *ccv, ccsignal_t type) 108 { 109 struct dctcp *dctcp_data; 110 int bytes_acked = 0; 111 uint32_t mss = tcp_fixed_maxseg(ccv->tp); 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, mss); 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 > mss) { 143 dctcp_data->bytes_ecn += 144 (bytes_acked - mss); 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 > mss) 151 dctcp_data->bytes_ecn += mss; 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(tptoinpcb(ccv->tp)); 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, ccsignal_t type) 242 { 243 struct dctcp *dctcp_data; 244 uint32_t cwin, mss, pipe; 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_fixed_maxseg(ccv->tp); 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 if (CCV(ccv, t_rxtshift) == 1) { 297 pipe = tcp_compute_pipe(ccv->tp); 298 CCV(ccv, snd_ssthresh) = max(2, 299 min(CCV(ccv, snd_wnd), pipe) / 2 / mss) * mss; 300 } 301 CCV(ccv, snd_cwnd) = mss; 302 dctcp_update_alpha(ccv); 303 dctcp_data->save_sndnxt += mss; 304 dctcp_data->num_cong_events++; 305 break; 306 default: 307 break; 308 } 309 } else 310 newreno_cc_cong_signal(ccv, type); 311 } 312 313 static void 314 dctcp_conn_init(struct cc_var *ccv) 315 { 316 struct dctcp *dctcp_data; 317 318 dctcp_data = ccv->cc_data; 319 320 if (CCV(ccv, t_flags2) & TF2_ECN_PERMIT) { 321 dctcp_data->save_sndnxt = CCV(ccv, snd_nxt); 322 if (V_dctcp_ect1) 323 CCV(ccv, t_flags2) |= TF2_ECN_USE_ECT1; 324 } 325 } 326 327 /* 328 * Perform any necessary tasks before we exit congestion recovery. 329 */ 330 static void 331 dctcp_post_recovery(struct cc_var *ccv) 332 { 333 newreno_cc_post_recovery(ccv); 334 335 if (CCV(ccv, t_flags2) & TF2_ECN_PERMIT) 336 dctcp_update_alpha(ccv); 337 } 338 339 /* 340 * Execute an additional ECN processing using ECN field in IP header 341 * and the CWR bit in TCP header. 342 */ 343 static void 344 dctcp_ecnpkt_handler(struct cc_var *ccv) 345 { 346 struct dctcp *dctcp_data; 347 uint32_t ccflag; 348 int acknow; 349 350 dctcp_data = ccv->cc_data; 351 ccflag = ccv->flags; 352 acknow = 0; 353 354 /* 355 * DCTCP responds with an ACK immediately when the CE state 356 * in between this segment and the last segment has changed. 357 */ 358 if (ccflag & CCF_IPHDR_CE) { 359 if (!dctcp_data->ce_prev) { 360 acknow = 1; 361 dctcp_data->ce_prev = 1; 362 CCV(ccv, t_flags2) |= TF2_ECN_SND_ECE; 363 } 364 } else { 365 if (dctcp_data->ce_prev) { 366 acknow = 1; 367 dctcp_data->ce_prev = 0; 368 CCV(ccv, t_flags2) &= ~TF2_ECN_SND_ECE; 369 } 370 } 371 372 if ((acknow) || (ccflag & CCF_TCPHDR_CWR)) { 373 ccv->flags |= CCF_ACKNOW; 374 } else { 375 ccv->flags &= ~CCF_ACKNOW; 376 } 377 } 378 379 /* 380 * Update the fraction of marked bytes represented as 'alpha'. 381 * Also initialize several internal parameters at the end of this function. 382 */ 383 static void 384 dctcp_update_alpha(struct cc_var *ccv) 385 { 386 struct dctcp *dctcp_data; 387 int alpha_prev; 388 389 dctcp_data = ccv->cc_data; 390 alpha_prev = dctcp_data->alpha; 391 dctcp_data->bytes_total = max(dctcp_data->bytes_total, 1); 392 393 /* 394 * Update alpha: alpha = (1 - g) * alpha + g * M. 395 * Here: 396 * g is weight factor 397 * recommaded to be set to 1/16 398 * small g = slow convergence between competitive DCTCP flows 399 * large g = impacts low utilization of bandwidth at switches 400 * M is fraction of marked segments in last RTT 401 * updated every RTT 402 * Alpha must be round to 0 - MAX_ALPHA_VALUE. 403 */ 404 dctcp_data->alpha = ulmin(alpha_prev - (alpha_prev >> V_dctcp_shift_g) + 405 ((uint64_t)dctcp_data->bytes_ecn << (DCTCP_SHIFT - V_dctcp_shift_g)) / 406 dctcp_data->bytes_total, MAX_ALPHA_VALUE); 407 408 /* Initialize internal parameters for next alpha calculation */ 409 dctcp_data->bytes_ecn = 0; 410 dctcp_data->bytes_total = 0; 411 dctcp_data->save_sndnxt = CCV(ccv, snd_nxt); 412 } 413 414 static int 415 dctcp_alpha_handler(SYSCTL_HANDLER_ARGS) 416 { 417 uint32_t new; 418 int error; 419 420 new = V_dctcp_alpha; 421 error = sysctl_handle_int(oidp, &new, 0, req); 422 if (error == 0 && req->newptr != NULL) { 423 if (new > MAX_ALPHA_VALUE) 424 error = EINVAL; 425 else 426 V_dctcp_alpha = new; 427 } 428 429 return (error); 430 } 431 432 static int 433 dctcp_shift_g_handler(SYSCTL_HANDLER_ARGS) 434 { 435 uint32_t new; 436 int error; 437 438 new = V_dctcp_shift_g; 439 error = sysctl_handle_int(oidp, &new, 0, req); 440 if (error == 0 && req->newptr != NULL) { 441 if (new > DCTCP_SHIFT) 442 error = EINVAL; 443 else 444 V_dctcp_shift_g = new; 445 } 446 447 return (error); 448 } 449 450 static int 451 dctcp_slowstart_handler(SYSCTL_HANDLER_ARGS) 452 { 453 uint32_t new; 454 int error; 455 456 new = V_dctcp_slowstart; 457 error = sysctl_handle_int(oidp, &new, 0, req); 458 if (error == 0 && req->newptr != NULL) { 459 if (new > 1) 460 error = EINVAL; 461 else 462 V_dctcp_slowstart = new; 463 } 464 465 return (error); 466 } 467 468 SYSCTL_DECL(_net_inet_tcp_cc_dctcp); 469 SYSCTL_NODE(_net_inet_tcp_cc, OID_AUTO, dctcp, 470 CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 471 "dctcp congestion control related settings"); 472 473 SYSCTL_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, alpha, 474 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 475 &VNET_NAME(dctcp_alpha), 0, &dctcp_alpha_handler, "IU", 476 "dctcp alpha parameter at start of session"); 477 478 SYSCTL_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, shift_g, 479 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 480 &VNET_NAME(dctcp_shift_g), 4, &dctcp_shift_g_handler, "IU", 481 "dctcp shift parameter"); 482 483 SYSCTL_PROC(_net_inet_tcp_cc_dctcp, OID_AUTO, slowstart, 484 CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 485 &VNET_NAME(dctcp_slowstart), 0, &dctcp_slowstart_handler, "IU", 486 "half CWND reduction after the first slow start"); 487 488 SYSCTL_UINT(_net_inet_tcp_cc_dctcp, OID_AUTO, ect1, 489 CTLFLAG_VNET | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 490 &VNET_NAME(dctcp_ect1), 0, 491 "Send DCTCP segments with ÍP ECT(0) or ECT(1)"); 492 493 DECLARE_CC_MODULE(dctcp, &dctcp_cc_algo); 494 MODULE_VERSION(dctcp, 2); 495