1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * Copyright (c) 2005, 2006 Andrea Bittau <a.bittau@cs.ucl.ac.uk> 4 * 5 * Changes to meet Linux coding standards, and DCCP infrastructure fixes. 6 * 7 * Copyright (c) 2006 Arnaldo Carvalho de Melo <acme@conectiva.com.br> 8 */ 9 10 /* 11 * This implementation should follow RFC 4341 12 */ 13 #include <linux/slab.h> 14 #include "../feat.h" 15 #include "ccid2.h" 16 17 18 #ifdef CONFIG_IP_DCCP_CCID2_DEBUG 19 static bool ccid2_debug; 20 #define ccid2_pr_debug(format, a...) DCCP_PR_DEBUG(ccid2_debug, format, ##a) 21 #else 22 #define ccid2_pr_debug(format, a...) 23 #endif 24 25 static int ccid2_hc_tx_alloc_seq(struct ccid2_hc_tx_sock *hc) 26 { 27 struct ccid2_seq *seqp; 28 int i; 29 30 /* check if we have space to preserve the pointer to the buffer */ 31 if (hc->tx_seqbufc >= (sizeof(hc->tx_seqbuf) / 32 sizeof(struct ccid2_seq *))) 33 return -ENOMEM; 34 35 /* allocate buffer and initialize linked list */ 36 seqp = kmalloc_array(CCID2_SEQBUF_LEN, sizeof(struct ccid2_seq), 37 gfp_any()); 38 if (seqp == NULL) 39 return -ENOMEM; 40 41 for (i = 0; i < (CCID2_SEQBUF_LEN - 1); i++) { 42 seqp[i].ccid2s_next = &seqp[i + 1]; 43 seqp[i + 1].ccid2s_prev = &seqp[i]; 44 } 45 seqp[CCID2_SEQBUF_LEN - 1].ccid2s_next = seqp; 46 seqp->ccid2s_prev = &seqp[CCID2_SEQBUF_LEN - 1]; 47 48 /* This is the first allocation. Initiate the head and tail. */ 49 if (hc->tx_seqbufc == 0) 50 hc->tx_seqh = hc->tx_seqt = seqp; 51 else { 52 /* link the existing list with the one we just created */ 53 hc->tx_seqh->ccid2s_next = seqp; 54 seqp->ccid2s_prev = hc->tx_seqh; 55 56 hc->tx_seqt->ccid2s_prev = &seqp[CCID2_SEQBUF_LEN - 1]; 57 seqp[CCID2_SEQBUF_LEN - 1].ccid2s_next = hc->tx_seqt; 58 } 59 60 /* store the original pointer to the buffer so we can free it */ 61 hc->tx_seqbuf[hc->tx_seqbufc] = seqp; 62 hc->tx_seqbufc++; 63 64 return 0; 65 } 66 67 static int ccid2_hc_tx_send_packet(struct sock *sk, struct sk_buff *skb) 68 { 69 if (ccid2_cwnd_network_limited(ccid2_hc_tx_sk(sk))) 70 return CCID_PACKET_WILL_DEQUEUE_LATER; 71 return CCID_PACKET_SEND_AT_ONCE; 72 } 73 74 static void ccid2_change_l_ack_ratio(struct sock *sk, u32 val) 75 { 76 u32 max_ratio = DIV_ROUND_UP(ccid2_hc_tx_sk(sk)->tx_cwnd, 2); 77 78 /* 79 * Ensure that Ack Ratio does not exceed ceil(cwnd/2), which is (2) from 80 * RFC 4341, 6.1.2. We ignore the statement that Ack Ratio 2 is always 81 * acceptable since this causes starvation/deadlock whenever cwnd < 2. 82 * The same problem arises when Ack Ratio is 0 (ie. Ack Ratio disabled). 83 */ 84 if (val == 0 || val > max_ratio) { 85 DCCP_WARN("Limiting Ack Ratio (%u) to %u\n", val, max_ratio); 86 val = max_ratio; 87 } 88 dccp_feat_signal_nn_change(sk, DCCPF_ACK_RATIO, 89 min_t(u32, val, DCCPF_ACK_RATIO_MAX)); 90 } 91 92 static void ccid2_check_l_ack_ratio(struct sock *sk) 93 { 94 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk); 95 96 /* 97 * After a loss, idle period, application limited period, or RTO we 98 * need to check that the ack ratio is still less than the congestion 99 * window. Otherwise, we will send an entire congestion window of 100 * packets and got no response because we haven't sent ack ratio 101 * packets yet. 102 * If the ack ratio does need to be reduced, we reduce it to half of 103 * the congestion window (or 1 if that's zero) instead of to the 104 * congestion window. This prevents problems if one ack is lost. 105 */ 106 if (dccp_feat_nn_get(sk, DCCPF_ACK_RATIO) > hc->tx_cwnd) 107 ccid2_change_l_ack_ratio(sk, hc->tx_cwnd/2 ? : 1U); 108 } 109 110 static void ccid2_change_l_seq_window(struct sock *sk, u64 val) 111 { 112 dccp_feat_signal_nn_change(sk, DCCPF_SEQUENCE_WINDOW, 113 clamp_val(val, DCCPF_SEQ_WMIN, 114 DCCPF_SEQ_WMAX)); 115 } 116 117 static void dccp_tasklet_schedule(struct sock *sk) 118 { 119 struct tasklet_struct *t = &dccp_sk(sk)->dccps_xmitlet; 120 121 if (!test_and_set_bit(TASKLET_STATE_SCHED, &t->state)) { 122 sock_hold(sk); 123 __tasklet_schedule(t); 124 } 125 } 126 127 static void ccid2_hc_tx_rto_expire(struct timer_list *t) 128 { 129 struct ccid2_hc_tx_sock *hc = from_timer(hc, t, tx_rtotimer); 130 struct sock *sk = hc->sk; 131 const bool sender_was_blocked = ccid2_cwnd_network_limited(hc); 132 133 bh_lock_sock(sk); 134 if (sock_owned_by_user(sk)) { 135 sk_reset_timer(sk, &hc->tx_rtotimer, jiffies + HZ / 5); 136 goto out; 137 } 138 139 ccid2_pr_debug("RTO_EXPIRE\n"); 140 141 if (sk->sk_state == DCCP_CLOSED) 142 goto out; 143 144 /* back-off timer */ 145 hc->tx_rto <<= 1; 146 if (hc->tx_rto > DCCP_RTO_MAX) 147 hc->tx_rto = DCCP_RTO_MAX; 148 149 /* adjust pipe, cwnd etc */ 150 hc->tx_ssthresh = hc->tx_cwnd / 2; 151 if (hc->tx_ssthresh < 2) 152 hc->tx_ssthresh = 2; 153 hc->tx_cwnd = 1; 154 hc->tx_pipe = 0; 155 156 /* clear state about stuff we sent */ 157 hc->tx_seqt = hc->tx_seqh; 158 hc->tx_packets_acked = 0; 159 160 /* clear ack ratio state. */ 161 hc->tx_rpseq = 0; 162 hc->tx_rpdupack = -1; 163 ccid2_change_l_ack_ratio(sk, 1); 164 165 /* if we were blocked before, we may now send cwnd=1 packet */ 166 if (sender_was_blocked) 167 dccp_tasklet_schedule(sk); 168 /* restart backed-off timer */ 169 sk_reset_timer(sk, &hc->tx_rtotimer, jiffies + hc->tx_rto); 170 out: 171 bh_unlock_sock(sk); 172 sock_put(sk); 173 } 174 175 /* 176 * Congestion window validation (RFC 2861). 177 */ 178 static bool ccid2_do_cwv = true; 179 module_param(ccid2_do_cwv, bool, 0644); 180 MODULE_PARM_DESC(ccid2_do_cwv, "Perform RFC2861 Congestion Window Validation"); 181 182 /** 183 * ccid2_update_used_window - Track how much of cwnd is actually used 184 * This is done in addition to CWV. The sender needs to have an idea of how many 185 * packets may be in flight, to set the local Sequence Window value accordingly 186 * (RFC 4340, 7.5.2). The CWV mechanism is exploited to keep track of the 187 * maximum-used window. We use an EWMA low-pass filter to filter out noise. 188 */ 189 static void ccid2_update_used_window(struct ccid2_hc_tx_sock *hc, u32 new_wnd) 190 { 191 hc->tx_expected_wnd = (3 * hc->tx_expected_wnd + new_wnd) / 4; 192 } 193 194 /* This borrows the code of tcp_cwnd_application_limited() */ 195 static void ccid2_cwnd_application_limited(struct sock *sk, const u32 now) 196 { 197 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk); 198 /* don't reduce cwnd below the initial window (IW) */ 199 u32 init_win = rfc3390_bytes_to_packets(dccp_sk(sk)->dccps_mss_cache), 200 win_used = max(hc->tx_cwnd_used, init_win); 201 202 if (win_used < hc->tx_cwnd) { 203 hc->tx_ssthresh = max(hc->tx_ssthresh, 204 (hc->tx_cwnd >> 1) + (hc->tx_cwnd >> 2)); 205 hc->tx_cwnd = (hc->tx_cwnd + win_used) >> 1; 206 } 207 hc->tx_cwnd_used = 0; 208 hc->tx_cwnd_stamp = now; 209 210 ccid2_check_l_ack_ratio(sk); 211 } 212 213 /* This borrows the code of tcp_cwnd_restart() */ 214 static void ccid2_cwnd_restart(struct sock *sk, const u32 now) 215 { 216 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk); 217 u32 cwnd = hc->tx_cwnd, restart_cwnd, 218 iwnd = rfc3390_bytes_to_packets(dccp_sk(sk)->dccps_mss_cache); 219 s32 delta = now - hc->tx_lsndtime; 220 221 hc->tx_ssthresh = max(hc->tx_ssthresh, (cwnd >> 1) + (cwnd >> 2)); 222 223 /* don't reduce cwnd below the initial window (IW) */ 224 restart_cwnd = min(cwnd, iwnd); 225 226 while ((delta -= hc->tx_rto) >= 0 && cwnd > restart_cwnd) 227 cwnd >>= 1; 228 hc->tx_cwnd = max(cwnd, restart_cwnd); 229 hc->tx_cwnd_stamp = now; 230 hc->tx_cwnd_used = 0; 231 232 ccid2_check_l_ack_ratio(sk); 233 } 234 235 static void ccid2_hc_tx_packet_sent(struct sock *sk, unsigned int len) 236 { 237 struct dccp_sock *dp = dccp_sk(sk); 238 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk); 239 const u32 now = ccid2_jiffies32; 240 struct ccid2_seq *next; 241 242 /* slow-start after idle periods (RFC 2581, RFC 2861) */ 243 if (ccid2_do_cwv && !hc->tx_pipe && 244 (s32)(now - hc->tx_lsndtime) >= hc->tx_rto) 245 ccid2_cwnd_restart(sk, now); 246 247 hc->tx_lsndtime = now; 248 hc->tx_pipe += 1; 249 250 /* see whether cwnd was fully used (RFC 2861), update expected window */ 251 if (ccid2_cwnd_network_limited(hc)) { 252 ccid2_update_used_window(hc, hc->tx_cwnd); 253 hc->tx_cwnd_used = 0; 254 hc->tx_cwnd_stamp = now; 255 } else { 256 if (hc->tx_pipe > hc->tx_cwnd_used) 257 hc->tx_cwnd_used = hc->tx_pipe; 258 259 ccid2_update_used_window(hc, hc->tx_cwnd_used); 260 261 if (ccid2_do_cwv && (s32)(now - hc->tx_cwnd_stamp) >= hc->tx_rto) 262 ccid2_cwnd_application_limited(sk, now); 263 } 264 265 hc->tx_seqh->ccid2s_seq = dp->dccps_gss; 266 hc->tx_seqh->ccid2s_acked = 0; 267 hc->tx_seqh->ccid2s_sent = now; 268 269 next = hc->tx_seqh->ccid2s_next; 270 /* check if we need to alloc more space */ 271 if (next == hc->tx_seqt) { 272 if (ccid2_hc_tx_alloc_seq(hc)) { 273 DCCP_CRIT("packet history - out of memory!"); 274 /* FIXME: find a more graceful way to bail out */ 275 return; 276 } 277 next = hc->tx_seqh->ccid2s_next; 278 BUG_ON(next == hc->tx_seqt); 279 } 280 hc->tx_seqh = next; 281 282 ccid2_pr_debug("cwnd=%d pipe=%d\n", hc->tx_cwnd, hc->tx_pipe); 283 284 /* 285 * FIXME: The code below is broken and the variables have been removed 286 * from the socket struct. The `ackloss' variable was always set to 0, 287 * and with arsent there are several problems: 288 * (i) it doesn't just count the number of Acks, but all sent packets; 289 * (ii) it is expressed in # of packets, not # of windows, so the 290 * comparison below uses the wrong formula: Appendix A of RFC 4341 291 * comes up with the number K = cwnd / (R^2 - R) of consecutive windows 292 * of data with no lost or marked Ack packets. If arsent were the # of 293 * consecutive Acks received without loss, then Ack Ratio needs to be 294 * decreased by 1 when 295 * arsent >= K * cwnd / R = cwnd^2 / (R^3 - R^2) 296 * where cwnd / R is the number of Acks received per window of data 297 * (cf. RFC 4341, App. A). The problems are that 298 * - arsent counts other packets as well; 299 * - the comparison uses a formula different from RFC 4341; 300 * - computing a cubic/quadratic equation each time is too complicated. 301 * Hence a different algorithm is needed. 302 */ 303 #if 0 304 /* Ack Ratio. Need to maintain a concept of how many windows we sent */ 305 hc->tx_arsent++; 306 /* We had an ack loss in this window... */ 307 if (hc->tx_ackloss) { 308 if (hc->tx_arsent >= hc->tx_cwnd) { 309 hc->tx_arsent = 0; 310 hc->tx_ackloss = 0; 311 } 312 } else { 313 /* No acks lost up to now... */ 314 /* decrease ack ratio if enough packets were sent */ 315 if (dp->dccps_l_ack_ratio > 1) { 316 /* XXX don't calculate denominator each time */ 317 int denom = dp->dccps_l_ack_ratio * dp->dccps_l_ack_ratio - 318 dp->dccps_l_ack_ratio; 319 320 denom = hc->tx_cwnd * hc->tx_cwnd / denom; 321 322 if (hc->tx_arsent >= denom) { 323 ccid2_change_l_ack_ratio(sk, dp->dccps_l_ack_ratio - 1); 324 hc->tx_arsent = 0; 325 } 326 } else { 327 /* we can't increase ack ratio further [1] */ 328 hc->tx_arsent = 0; /* or maybe set it to cwnd*/ 329 } 330 } 331 #endif 332 333 sk_reset_timer(sk, &hc->tx_rtotimer, jiffies + hc->tx_rto); 334 335 #ifdef CONFIG_IP_DCCP_CCID2_DEBUG 336 do { 337 struct ccid2_seq *seqp = hc->tx_seqt; 338 339 while (seqp != hc->tx_seqh) { 340 ccid2_pr_debug("out seq=%llu acked=%d time=%u\n", 341 (unsigned long long)seqp->ccid2s_seq, 342 seqp->ccid2s_acked, seqp->ccid2s_sent); 343 seqp = seqp->ccid2s_next; 344 } 345 } while (0); 346 ccid2_pr_debug("=========\n"); 347 #endif 348 } 349 350 /** 351 * ccid2_rtt_estimator - Sample RTT and compute RTO using RFC2988 algorithm 352 * This code is almost identical with TCP's tcp_rtt_estimator(), since 353 * - it has a higher sampling frequency (recommended by RFC 1323), 354 * - the RTO does not collapse into RTT due to RTTVAR going towards zero, 355 * - it is simple (cf. more complex proposals such as Eifel timer or research 356 * which suggests that the gain should be set according to window size), 357 * - in tests it was found to work well with CCID2 [gerrit]. 358 */ 359 static void ccid2_rtt_estimator(struct sock *sk, const long mrtt) 360 { 361 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk); 362 long m = mrtt ? : 1; 363 364 if (hc->tx_srtt == 0) { 365 /* First measurement m */ 366 hc->tx_srtt = m << 3; 367 hc->tx_mdev = m << 1; 368 369 hc->tx_mdev_max = max(hc->tx_mdev, tcp_rto_min(sk)); 370 hc->tx_rttvar = hc->tx_mdev_max; 371 372 hc->tx_rtt_seq = dccp_sk(sk)->dccps_gss; 373 } else { 374 /* Update scaled SRTT as SRTT += 1/8 * (m - SRTT) */ 375 m -= (hc->tx_srtt >> 3); 376 hc->tx_srtt += m; 377 378 /* Similarly, update scaled mdev with regard to |m| */ 379 if (m < 0) { 380 m = -m; 381 m -= (hc->tx_mdev >> 2); 382 /* 383 * This neutralises RTO increase when RTT < SRTT - mdev 384 * (see P. Sarolahti, A. Kuznetsov,"Congestion Control 385 * in Linux TCP", USENIX 2002, pp. 49-62). 386 */ 387 if (m > 0) 388 m >>= 3; 389 } else { 390 m -= (hc->tx_mdev >> 2); 391 } 392 hc->tx_mdev += m; 393 394 if (hc->tx_mdev > hc->tx_mdev_max) { 395 hc->tx_mdev_max = hc->tx_mdev; 396 if (hc->tx_mdev_max > hc->tx_rttvar) 397 hc->tx_rttvar = hc->tx_mdev_max; 398 } 399 400 /* 401 * Decay RTTVAR at most once per flight, exploiting that 402 * 1) pipe <= cwnd <= Sequence_Window = W (RFC 4340, 7.5.2) 403 * 2) AWL = GSS-W+1 <= GAR <= GSS (RFC 4340, 7.5.1) 404 * GAR is a useful bound for FlightSize = pipe. 405 * AWL is probably too low here, as it over-estimates pipe. 406 */ 407 if (after48(dccp_sk(sk)->dccps_gar, hc->tx_rtt_seq)) { 408 if (hc->tx_mdev_max < hc->tx_rttvar) 409 hc->tx_rttvar -= (hc->tx_rttvar - 410 hc->tx_mdev_max) >> 2; 411 hc->tx_rtt_seq = dccp_sk(sk)->dccps_gss; 412 hc->tx_mdev_max = tcp_rto_min(sk); 413 } 414 } 415 416 /* 417 * Set RTO from SRTT and RTTVAR 418 * As in TCP, 4 * RTTVAR >= TCP_RTO_MIN, giving a minimum RTO of 200 ms. 419 * This agrees with RFC 4341, 5: 420 * "Because DCCP does not retransmit data, DCCP does not require 421 * TCP's recommended minimum timeout of one second". 422 */ 423 hc->tx_rto = (hc->tx_srtt >> 3) + hc->tx_rttvar; 424 425 if (hc->tx_rto > DCCP_RTO_MAX) 426 hc->tx_rto = DCCP_RTO_MAX; 427 } 428 429 static void ccid2_new_ack(struct sock *sk, struct ccid2_seq *seqp, 430 unsigned int *maxincr) 431 { 432 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk); 433 struct dccp_sock *dp = dccp_sk(sk); 434 int r_seq_used = hc->tx_cwnd / dp->dccps_l_ack_ratio; 435 436 if (hc->tx_cwnd < dp->dccps_l_seq_win && 437 r_seq_used < dp->dccps_r_seq_win) { 438 if (hc->tx_cwnd < hc->tx_ssthresh) { 439 if (*maxincr > 0 && ++hc->tx_packets_acked >= 2) { 440 hc->tx_cwnd += 1; 441 *maxincr -= 1; 442 hc->tx_packets_acked = 0; 443 } 444 } else if (++hc->tx_packets_acked >= hc->tx_cwnd) { 445 hc->tx_cwnd += 1; 446 hc->tx_packets_acked = 0; 447 } 448 } 449 450 /* 451 * Adjust the local sequence window and the ack ratio to allow about 452 * 5 times the number of packets in the network (RFC 4340 7.5.2) 453 */ 454 if (r_seq_used * CCID2_WIN_CHANGE_FACTOR >= dp->dccps_r_seq_win) 455 ccid2_change_l_ack_ratio(sk, dp->dccps_l_ack_ratio * 2); 456 else if (r_seq_used * CCID2_WIN_CHANGE_FACTOR < dp->dccps_r_seq_win/2) 457 ccid2_change_l_ack_ratio(sk, dp->dccps_l_ack_ratio / 2 ? : 1U); 458 459 if (hc->tx_cwnd * CCID2_WIN_CHANGE_FACTOR >= dp->dccps_l_seq_win) 460 ccid2_change_l_seq_window(sk, dp->dccps_l_seq_win * 2); 461 else if (hc->tx_cwnd * CCID2_WIN_CHANGE_FACTOR < dp->dccps_l_seq_win/2) 462 ccid2_change_l_seq_window(sk, dp->dccps_l_seq_win / 2); 463 464 /* 465 * FIXME: RTT is sampled several times per acknowledgment (for each 466 * entry in the Ack Vector), instead of once per Ack (as in TCP SACK). 467 * This causes the RTT to be over-estimated, since the older entries 468 * in the Ack Vector have earlier sending times. 469 * The cleanest solution is to not use the ccid2s_sent field at all 470 * and instead use DCCP timestamps: requires changes in other places. 471 */ 472 ccid2_rtt_estimator(sk, ccid2_jiffies32 - seqp->ccid2s_sent); 473 } 474 475 static void ccid2_congestion_event(struct sock *sk, struct ccid2_seq *seqp) 476 { 477 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk); 478 479 if ((s32)(seqp->ccid2s_sent - hc->tx_last_cong) < 0) { 480 ccid2_pr_debug("Multiple losses in an RTT---treating as one\n"); 481 return; 482 } 483 484 hc->tx_last_cong = ccid2_jiffies32; 485 486 hc->tx_cwnd = hc->tx_cwnd / 2 ? : 1U; 487 hc->tx_ssthresh = max(hc->tx_cwnd, 2U); 488 489 ccid2_check_l_ack_ratio(sk); 490 } 491 492 static int ccid2_hc_tx_parse_options(struct sock *sk, u8 packet_type, 493 u8 option, u8 *optval, u8 optlen) 494 { 495 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk); 496 497 switch (option) { 498 case DCCPO_ACK_VECTOR_0: 499 case DCCPO_ACK_VECTOR_1: 500 return dccp_ackvec_parsed_add(&hc->tx_av_chunks, optval, optlen, 501 option - DCCPO_ACK_VECTOR_0); 502 } 503 return 0; 504 } 505 506 static void ccid2_hc_tx_packet_recv(struct sock *sk, struct sk_buff *skb) 507 { 508 struct dccp_sock *dp = dccp_sk(sk); 509 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk); 510 const bool sender_was_blocked = ccid2_cwnd_network_limited(hc); 511 struct dccp_ackvec_parsed *avp; 512 u64 ackno, seqno; 513 struct ccid2_seq *seqp; 514 int done = 0; 515 unsigned int maxincr = 0; 516 517 /* check reverse path congestion */ 518 seqno = DCCP_SKB_CB(skb)->dccpd_seq; 519 520 /* XXX this whole "algorithm" is broken. Need to fix it to keep track 521 * of the seqnos of the dupacks so that rpseq and rpdupack are correct 522 * -sorbo. 523 */ 524 /* need to bootstrap */ 525 if (hc->tx_rpdupack == -1) { 526 hc->tx_rpdupack = 0; 527 hc->tx_rpseq = seqno; 528 } else { 529 /* check if packet is consecutive */ 530 if (dccp_delta_seqno(hc->tx_rpseq, seqno) == 1) 531 hc->tx_rpseq = seqno; 532 /* it's a later packet */ 533 else if (after48(seqno, hc->tx_rpseq)) { 534 hc->tx_rpdupack++; 535 536 /* check if we got enough dupacks */ 537 if (hc->tx_rpdupack >= NUMDUPACK) { 538 hc->tx_rpdupack = -1; /* XXX lame */ 539 hc->tx_rpseq = 0; 540 #ifdef __CCID2_COPES_GRACEFULLY_WITH_ACK_CONGESTION_CONTROL__ 541 /* 542 * FIXME: Ack Congestion Control is broken; in 543 * the current state instabilities occurred with 544 * Ack Ratios greater than 1; causing hang-ups 545 * and long RTO timeouts. This needs to be fixed 546 * before opening up dynamic changes. -- gerrit 547 */ 548 ccid2_change_l_ack_ratio(sk, 2 * dp->dccps_l_ack_ratio); 549 #endif 550 } 551 } 552 } 553 554 /* check forward path congestion */ 555 if (dccp_packet_without_ack(skb)) 556 return; 557 558 /* still didn't send out new data packets */ 559 if (hc->tx_seqh == hc->tx_seqt) 560 goto done; 561 562 ackno = DCCP_SKB_CB(skb)->dccpd_ack_seq; 563 if (after48(ackno, hc->tx_high_ack)) 564 hc->tx_high_ack = ackno; 565 566 seqp = hc->tx_seqt; 567 while (before48(seqp->ccid2s_seq, ackno)) { 568 seqp = seqp->ccid2s_next; 569 if (seqp == hc->tx_seqh) { 570 seqp = hc->tx_seqh->ccid2s_prev; 571 break; 572 } 573 } 574 575 /* 576 * In slow-start, cwnd can increase up to a maximum of Ack Ratio/2 577 * packets per acknowledgement. Rounding up avoids that cwnd is not 578 * advanced when Ack Ratio is 1 and gives a slight edge otherwise. 579 */ 580 if (hc->tx_cwnd < hc->tx_ssthresh) 581 maxincr = DIV_ROUND_UP(dp->dccps_l_ack_ratio, 2); 582 583 /* go through all ack vectors */ 584 list_for_each_entry(avp, &hc->tx_av_chunks, node) { 585 /* go through this ack vector */ 586 for (; avp->len--; avp->vec++) { 587 u64 ackno_end_rl = SUB48(ackno, 588 dccp_ackvec_runlen(avp->vec)); 589 590 ccid2_pr_debug("ackvec %llu |%u,%u|\n", 591 (unsigned long long)ackno, 592 dccp_ackvec_state(avp->vec) >> 6, 593 dccp_ackvec_runlen(avp->vec)); 594 /* if the seqno we are analyzing is larger than the 595 * current ackno, then move towards the tail of our 596 * seqnos. 597 */ 598 while (after48(seqp->ccid2s_seq, ackno)) { 599 if (seqp == hc->tx_seqt) { 600 done = 1; 601 break; 602 } 603 seqp = seqp->ccid2s_prev; 604 } 605 if (done) 606 break; 607 608 /* check all seqnos in the range of the vector 609 * run length 610 */ 611 while (between48(seqp->ccid2s_seq,ackno_end_rl,ackno)) { 612 const u8 state = dccp_ackvec_state(avp->vec); 613 614 /* new packet received or marked */ 615 if (state != DCCPAV_NOT_RECEIVED && 616 !seqp->ccid2s_acked) { 617 if (state == DCCPAV_ECN_MARKED) 618 ccid2_congestion_event(sk, 619 seqp); 620 else 621 ccid2_new_ack(sk, seqp, 622 &maxincr); 623 624 seqp->ccid2s_acked = 1; 625 ccid2_pr_debug("Got ack for %llu\n", 626 (unsigned long long)seqp->ccid2s_seq); 627 hc->tx_pipe--; 628 } 629 if (seqp == hc->tx_seqt) { 630 done = 1; 631 break; 632 } 633 seqp = seqp->ccid2s_prev; 634 } 635 if (done) 636 break; 637 638 ackno = SUB48(ackno_end_rl, 1); 639 } 640 if (done) 641 break; 642 } 643 644 /* The state about what is acked should be correct now 645 * Check for NUMDUPACK 646 */ 647 seqp = hc->tx_seqt; 648 while (before48(seqp->ccid2s_seq, hc->tx_high_ack)) { 649 seqp = seqp->ccid2s_next; 650 if (seqp == hc->tx_seqh) { 651 seqp = hc->tx_seqh->ccid2s_prev; 652 break; 653 } 654 } 655 done = 0; 656 while (1) { 657 if (seqp->ccid2s_acked) { 658 done++; 659 if (done == NUMDUPACK) 660 break; 661 } 662 if (seqp == hc->tx_seqt) 663 break; 664 seqp = seqp->ccid2s_prev; 665 } 666 667 /* If there are at least 3 acknowledgements, anything unacknowledged 668 * below the last sequence number is considered lost 669 */ 670 if (done == NUMDUPACK) { 671 struct ccid2_seq *last_acked = seqp; 672 673 /* check for lost packets */ 674 while (1) { 675 if (!seqp->ccid2s_acked) { 676 ccid2_pr_debug("Packet lost: %llu\n", 677 (unsigned long long)seqp->ccid2s_seq); 678 /* XXX need to traverse from tail -> head in 679 * order to detect multiple congestion events in 680 * one ack vector. 681 */ 682 ccid2_congestion_event(sk, seqp); 683 hc->tx_pipe--; 684 } 685 if (seqp == hc->tx_seqt) 686 break; 687 seqp = seqp->ccid2s_prev; 688 } 689 690 hc->tx_seqt = last_acked; 691 } 692 693 /* trim acked packets in tail */ 694 while (hc->tx_seqt != hc->tx_seqh) { 695 if (!hc->tx_seqt->ccid2s_acked) 696 break; 697 698 hc->tx_seqt = hc->tx_seqt->ccid2s_next; 699 } 700 701 /* restart RTO timer if not all outstanding data has been acked */ 702 if (hc->tx_pipe == 0) 703 sk_stop_timer(sk, &hc->tx_rtotimer); 704 else 705 sk_reset_timer(sk, &hc->tx_rtotimer, jiffies + hc->tx_rto); 706 done: 707 /* check if incoming Acks allow pending packets to be sent */ 708 if (sender_was_blocked && !ccid2_cwnd_network_limited(hc)) 709 dccp_tasklet_schedule(sk); 710 dccp_ackvec_parsed_cleanup(&hc->tx_av_chunks); 711 } 712 713 static int ccid2_hc_tx_init(struct ccid *ccid, struct sock *sk) 714 { 715 struct ccid2_hc_tx_sock *hc = ccid_priv(ccid); 716 struct dccp_sock *dp = dccp_sk(sk); 717 u32 max_ratio; 718 719 /* RFC 4341, 5: initialise ssthresh to arbitrarily high (max) value */ 720 hc->tx_ssthresh = ~0U; 721 722 /* Use larger initial windows (RFC 4341, section 5). */ 723 hc->tx_cwnd = rfc3390_bytes_to_packets(dp->dccps_mss_cache); 724 hc->tx_expected_wnd = hc->tx_cwnd; 725 726 /* Make sure that Ack Ratio is enabled and within bounds. */ 727 max_ratio = DIV_ROUND_UP(hc->tx_cwnd, 2); 728 if (dp->dccps_l_ack_ratio == 0 || dp->dccps_l_ack_ratio > max_ratio) 729 dp->dccps_l_ack_ratio = max_ratio; 730 731 /* XXX init ~ to window size... */ 732 if (ccid2_hc_tx_alloc_seq(hc)) 733 return -ENOMEM; 734 735 hc->tx_rto = DCCP_TIMEOUT_INIT; 736 hc->tx_rpdupack = -1; 737 hc->tx_last_cong = hc->tx_lsndtime = hc->tx_cwnd_stamp = ccid2_jiffies32; 738 hc->tx_cwnd_used = 0; 739 hc->sk = sk; 740 timer_setup(&hc->tx_rtotimer, ccid2_hc_tx_rto_expire, 0); 741 INIT_LIST_HEAD(&hc->tx_av_chunks); 742 return 0; 743 } 744 745 static void ccid2_hc_tx_exit(struct sock *sk) 746 { 747 struct ccid2_hc_tx_sock *hc = ccid2_hc_tx_sk(sk); 748 int i; 749 750 sk_stop_timer(sk, &hc->tx_rtotimer); 751 752 for (i = 0; i < hc->tx_seqbufc; i++) 753 kfree(hc->tx_seqbuf[i]); 754 hc->tx_seqbufc = 0; 755 dccp_ackvec_parsed_cleanup(&hc->tx_av_chunks); 756 } 757 758 static void ccid2_hc_rx_packet_recv(struct sock *sk, struct sk_buff *skb) 759 { 760 struct ccid2_hc_rx_sock *hc = ccid2_hc_rx_sk(sk); 761 762 if (!dccp_data_packet(skb)) 763 return; 764 765 if (++hc->rx_num_data_pkts >= dccp_sk(sk)->dccps_r_ack_ratio) { 766 dccp_send_ack(sk); 767 hc->rx_num_data_pkts = 0; 768 } 769 } 770 771 struct ccid_operations ccid2_ops = { 772 .ccid_id = DCCPC_CCID2, 773 .ccid_name = "TCP-like", 774 .ccid_hc_tx_obj_size = sizeof(struct ccid2_hc_tx_sock), 775 .ccid_hc_tx_init = ccid2_hc_tx_init, 776 .ccid_hc_tx_exit = ccid2_hc_tx_exit, 777 .ccid_hc_tx_send_packet = ccid2_hc_tx_send_packet, 778 .ccid_hc_tx_packet_sent = ccid2_hc_tx_packet_sent, 779 .ccid_hc_tx_parse_options = ccid2_hc_tx_parse_options, 780 .ccid_hc_tx_packet_recv = ccid2_hc_tx_packet_recv, 781 .ccid_hc_rx_obj_size = sizeof(struct ccid2_hc_rx_sock), 782 .ccid_hc_rx_packet_recv = ccid2_hc_rx_packet_recv, 783 }; 784 785 #ifdef CONFIG_IP_DCCP_CCID2_DEBUG 786 module_param(ccid2_debug, bool, 0644); 787 MODULE_PARM_DESC(ccid2_debug, "Enable CCID-2 debug messages"); 788 #endif 789