Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * INET An implementation of the TCP/IP protocol suite for the LINUX | |
3 | * operating system. INET is implemented using the BSD Socket | |
4 | * interface as the means of communication with the user level. | |
5 | * | |
6 | * Implementation of the Transmission Control Protocol(TCP). | |
7 | * | |
02c30a84 | 8 | * Authors: Ross Biro |
1da177e4 LT |
9 | * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> |
10 | * Mark Evans, <evansmp@uhura.aston.ac.uk> | |
11 | * Corey Minyard <wf-rch!minyard@relay.EU.net> | |
12 | * Florian La Roche, <flla@stud.uni-sb.de> | |
13 | * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> | |
14 | * Linus Torvalds, <torvalds@cs.helsinki.fi> | |
15 | * Alan Cox, <gw4pts@gw4pts.ampr.org> | |
16 | * Matthew Dillon, <dillon@apollo.west.oic.com> | |
17 | * Arnt Gulbrandsen, <agulbra@nvg.unit.no> | |
18 | * Jorge Cwik, <jorge@laser.satlink.net> | |
19 | */ | |
20 | ||
21 | /* | |
22 | * Changes: | |
23 | * Pedro Roque : Fast Retransmit/Recovery. | |
24 | * Two receive queues. | |
25 | * Retransmit queue handled by TCP. | |
26 | * Better retransmit timer handling. | |
27 | * New congestion avoidance. | |
28 | * Header prediction. | |
29 | * Variable renaming. | |
30 | * | |
31 | * Eric : Fast Retransmit. | |
32 | * Randy Scott : MSS option defines. | |
33 | * Eric Schenk : Fixes to slow start algorithm. | |
34 | * Eric Schenk : Yet another double ACK bug. | |
35 | * Eric Schenk : Delayed ACK bug fixes. | |
36 | * Eric Schenk : Floyd style fast retrans war avoidance. | |
37 | * David S. Miller : Don't allow zero congestion window. | |
38 | * Eric Schenk : Fix retransmitter so that it sends | |
39 | * next packet on ack of previous packet. | |
40 | * Andi Kleen : Moved open_request checking here | |
41 | * and process RSTs for open_requests. | |
42 | * Andi Kleen : Better prune_queue, and other fixes. | |
caa20d9a | 43 | * Andrey Savochkin: Fix RTT measurements in the presence of |
1da177e4 LT |
44 | * timestamps. |
45 | * Andrey Savochkin: Check sequence numbers correctly when | |
46 | * removing SACKs due to in sequence incoming | |
47 | * data segments. | |
48 | * Andi Kleen: Make sure we never ack data there is not | |
49 | * enough room for. Also make this condition | |
50 | * a fatal error if it might still happen. | |
e905a9ed | 51 | * Andi Kleen: Add tcp_measure_rcv_mss to make |
1da177e4 | 52 | * connections with MSS<min(MTU,ann. MSS) |
e905a9ed | 53 | * work without delayed acks. |
1da177e4 LT |
54 | * Andi Kleen: Process packets with PSH set in the |
55 | * fast path. | |
56 | * J Hadi Salim: ECN support | |
57 | * Andrei Gurtov, | |
58 | * Pasi Sarolahti, | |
59 | * Panu Kuhlberg: Experimental audit of TCP (re)transmission | |
60 | * engine. Lots of bugs are found. | |
61 | * Pasi Sarolahti: F-RTO for dealing with spurious RTOs | |
1da177e4 LT |
62 | */ |
63 | ||
afd46503 JP |
64 | #define pr_fmt(fmt) "TCP: " fmt |
65 | ||
1da177e4 | 66 | #include <linux/mm.h> |
5a0e3ad6 | 67 | #include <linux/slab.h> |
1da177e4 LT |
68 | #include <linux/module.h> |
69 | #include <linux/sysctl.h> | |
a0bffffc | 70 | #include <linux/kernel.h> |
ad971f61 | 71 | #include <linux/prefetch.h> |
5ffc02a1 | 72 | #include <net/dst.h> |
1da177e4 LT |
73 | #include <net/tcp.h> |
74 | #include <net/inet_common.h> | |
75 | #include <linux/ipsec.h> | |
76 | #include <asm/unaligned.h> | |
e1c8a607 | 77 | #include <linux/errqueue.h> |
5941521c | 78 | #include <trace/events/tcp.h> |
60e2a778 UB |
79 | #include <linux/unaligned/access_ok.h> |
80 | #include <linux/static_key.h> | |
1da177e4 | 81 | |
ab32ea5d | 82 | int sysctl_tcp_max_orphans __read_mostly = NR_FILE; |
1da177e4 | 83 | |
1da177e4 LT |
84 | #define FLAG_DATA 0x01 /* Incoming frame contained data. */ |
85 | #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ | |
86 | #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ | |
87 | #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ | |
88 | #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ | |
89 | #define FLAG_DATA_SACKED 0x20 /* New SACK. */ | |
90 | #define FLAG_ECE 0x40 /* ECE in this ACK */ | |
291a00d1 | 91 | #define FLAG_LOST_RETRANS 0x80 /* This ACK marks some retransmission lost */ |
31770e34 | 92 | #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ |
e33099f9 | 93 | #define FLAG_ORIG_SACK_ACKED 0x200 /* Never retransmitted data are (s)acked */ |
2e605294 | 94 | #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ |
564262c1 | 95 | #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ |
df92c839 | 96 | #define FLAG_SET_XMIT_TIMER 0x1000 /* Set TLP or RTO timer */ |
cadbd031 | 97 | #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ |
12fb3dd9 | 98 | #define FLAG_UPDATE_TS_RECENT 0x4000 /* tcp_replace_ts_recent() */ |
d0e1a1b5 | 99 | #define FLAG_NO_CHALLENGE_ACK 0x8000 /* do not call tcp_send_challenge_ack() */ |
1da177e4 LT |
100 | |
101 | #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) | |
102 | #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) | |
103 | #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) | |
104 | #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) | |
105 | ||
1da177e4 | 106 | #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) |
bdf1ee5d | 107 | #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) |
1da177e4 | 108 | |
e662ca40 YC |
109 | #define REXMIT_NONE 0 /* no loss recovery to do */ |
110 | #define REXMIT_LOST 1 /* retransmit packets marked lost */ | |
111 | #define REXMIT_NEW 2 /* FRTO-style transmit of unsent/new packets */ | |
112 | ||
0b9aefea MRL |
113 | static void tcp_gro_dev_warn(struct sock *sk, const struct sk_buff *skb, |
114 | unsigned int len) | |
dcb17d22 MRL |
115 | { |
116 | static bool __once __read_mostly; | |
117 | ||
118 | if (!__once) { | |
119 | struct net_device *dev; | |
120 | ||
121 | __once = true; | |
122 | ||
123 | rcu_read_lock(); | |
124 | dev = dev_get_by_index_rcu(sock_net(sk), skb->skb_iif); | |
0b9aefea MRL |
125 | if (!dev || len >= dev->mtu) |
126 | pr_warn("%s: Driver has suspect GRO implementation, TCP performance may be compromised.\n", | |
127 | dev ? dev->name : "Unknown driver"); | |
dcb17d22 MRL |
128 | rcu_read_unlock(); |
129 | } | |
130 | } | |
131 | ||
e905a9ed | 132 | /* Adapt the MSS value used to make delayed ack decision to the |
1da177e4 | 133 | * real world. |
e905a9ed | 134 | */ |
056834d9 | 135 | static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 136 | { |
463c84b9 | 137 | struct inet_connection_sock *icsk = inet_csk(sk); |
e905a9ed | 138 | const unsigned int lss = icsk->icsk_ack.last_seg_size; |
463c84b9 | 139 | unsigned int len; |
1da177e4 | 140 | |
e905a9ed | 141 | icsk->icsk_ack.last_seg_size = 0; |
1da177e4 LT |
142 | |
143 | /* skb->len may jitter because of SACKs, even if peer | |
144 | * sends good full-sized frames. | |
145 | */ | |
056834d9 | 146 | len = skb_shinfo(skb)->gso_size ? : skb->len; |
463c84b9 | 147 | if (len >= icsk->icsk_ack.rcv_mss) { |
dcb17d22 MRL |
148 | icsk->icsk_ack.rcv_mss = min_t(unsigned int, len, |
149 | tcp_sk(sk)->advmss); | |
0b9aefea MRL |
150 | /* Account for possibly-removed options */ |
151 | if (unlikely(len > icsk->icsk_ack.rcv_mss + | |
152 | MAX_TCP_OPTION_SPACE)) | |
153 | tcp_gro_dev_warn(sk, skb, len); | |
1da177e4 LT |
154 | } else { |
155 | /* Otherwise, we make more careful check taking into account, | |
156 | * that SACKs block is variable. | |
157 | * | |
158 | * "len" is invariant segment length, including TCP header. | |
159 | */ | |
9c70220b | 160 | len += skb->data - skb_transport_header(skb); |
bee7ca9e | 161 | if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || |
1da177e4 LT |
162 | /* If PSH is not set, packet should be |
163 | * full sized, provided peer TCP is not badly broken. | |
164 | * This observation (if it is correct 8)) allows | |
165 | * to handle super-low mtu links fairly. | |
166 | */ | |
167 | (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && | |
aa8223c7 | 168 | !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { |
1da177e4 LT |
169 | /* Subtract also invariant (if peer is RFC compliant), |
170 | * tcp header plus fixed timestamp option length. | |
171 | * Resulting "len" is MSS free of SACK jitter. | |
172 | */ | |
463c84b9 ACM |
173 | len -= tcp_sk(sk)->tcp_header_len; |
174 | icsk->icsk_ack.last_seg_size = len; | |
1da177e4 | 175 | if (len == lss) { |
463c84b9 | 176 | icsk->icsk_ack.rcv_mss = len; |
1da177e4 LT |
177 | return; |
178 | } | |
179 | } | |
1ef9696c AK |
180 | if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) |
181 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; | |
463c84b9 | 182 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; |
1da177e4 LT |
183 | } |
184 | } | |
185 | ||
463c84b9 | 186 | static void tcp_incr_quickack(struct sock *sk) |
1da177e4 | 187 | { |
463c84b9 | 188 | struct inet_connection_sock *icsk = inet_csk(sk); |
95c96174 | 189 | unsigned int quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); |
1da177e4 | 190 | |
056834d9 IJ |
191 | if (quickacks == 0) |
192 | quickacks = 2; | |
463c84b9 ACM |
193 | if (quickacks > icsk->icsk_ack.quick) |
194 | icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); | |
1da177e4 LT |
195 | } |
196 | ||
1b9f4092 | 197 | static void tcp_enter_quickack_mode(struct sock *sk) |
1da177e4 | 198 | { |
463c84b9 ACM |
199 | struct inet_connection_sock *icsk = inet_csk(sk); |
200 | tcp_incr_quickack(sk); | |
201 | icsk->icsk_ack.pingpong = 0; | |
202 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 LT |
203 | } |
204 | ||
205 | /* Send ACKs quickly, if "quick" count is not exhausted | |
206 | * and the session is not interactive. | |
207 | */ | |
208 | ||
2251ae46 | 209 | static bool tcp_in_quickack_mode(struct sock *sk) |
1da177e4 | 210 | { |
463c84b9 | 211 | const struct inet_connection_sock *icsk = inet_csk(sk); |
2251ae46 | 212 | const struct dst_entry *dst = __sk_dst_get(sk); |
a2a385d6 | 213 | |
2251ae46 JM |
214 | return (dst && dst_metric(dst, RTAX_QUICKACK)) || |
215 | (icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong); | |
1da177e4 LT |
216 | } |
217 | ||
735d3831 | 218 | static void tcp_ecn_queue_cwr(struct tcp_sock *tp) |
bdf1ee5d | 219 | { |
056834d9 | 220 | if (tp->ecn_flags & TCP_ECN_OK) |
bdf1ee5d IJ |
221 | tp->ecn_flags |= TCP_ECN_QUEUE_CWR; |
222 | } | |
223 | ||
735d3831 | 224 | static void tcp_ecn_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb) |
bdf1ee5d IJ |
225 | { |
226 | if (tcp_hdr(skb)->cwr) | |
227 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
228 | } | |
229 | ||
735d3831 | 230 | static void tcp_ecn_withdraw_cwr(struct tcp_sock *tp) |
bdf1ee5d IJ |
231 | { |
232 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
233 | } | |
234 | ||
735d3831 | 235 | static void __tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb) |
bdf1ee5d | 236 | { |
b82d1bb4 | 237 | switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) { |
7a269ffa | 238 | case INET_ECN_NOT_ECT: |
bdf1ee5d | 239 | /* Funny extension: if ECT is not set on a segment, |
7a269ffa ED |
240 | * and we already seen ECT on a previous segment, |
241 | * it is probably a retransmit. | |
242 | */ | |
243 | if (tp->ecn_flags & TCP_ECN_SEEN) | |
bdf1ee5d | 244 | tcp_enter_quickack_mode((struct sock *)tp); |
7a269ffa ED |
245 | break; |
246 | case INET_ECN_CE: | |
9890092e FW |
247 | if (tcp_ca_needs_ecn((struct sock *)tp)) |
248 | tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_IS_CE); | |
249 | ||
aae06bf5 ED |
250 | if (!(tp->ecn_flags & TCP_ECN_DEMAND_CWR)) { |
251 | /* Better not delay acks, sender can have a very low cwnd */ | |
252 | tcp_enter_quickack_mode((struct sock *)tp); | |
253 | tp->ecn_flags |= TCP_ECN_DEMAND_CWR; | |
254 | } | |
9890092e FW |
255 | tp->ecn_flags |= TCP_ECN_SEEN; |
256 | break; | |
7a269ffa | 257 | default: |
9890092e FW |
258 | if (tcp_ca_needs_ecn((struct sock *)tp)) |
259 | tcp_ca_event((struct sock *)tp, CA_EVENT_ECN_NO_CE); | |
7a269ffa | 260 | tp->ecn_flags |= TCP_ECN_SEEN; |
9890092e | 261 | break; |
bdf1ee5d IJ |
262 | } |
263 | } | |
264 | ||
735d3831 FW |
265 | static void tcp_ecn_check_ce(struct tcp_sock *tp, const struct sk_buff *skb) |
266 | { | |
267 | if (tp->ecn_flags & TCP_ECN_OK) | |
268 | __tcp_ecn_check_ce(tp, skb); | |
269 | } | |
270 | ||
271 | static void tcp_ecn_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th) | |
bdf1ee5d | 272 | { |
056834d9 | 273 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) |
bdf1ee5d IJ |
274 | tp->ecn_flags &= ~TCP_ECN_OK; |
275 | } | |
276 | ||
735d3831 | 277 | static void tcp_ecn_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th) |
bdf1ee5d | 278 | { |
056834d9 | 279 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) |
bdf1ee5d IJ |
280 | tp->ecn_flags &= ~TCP_ECN_OK; |
281 | } | |
282 | ||
735d3831 | 283 | static bool tcp_ecn_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th) |
bdf1ee5d | 284 | { |
056834d9 | 285 | if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) |
a2a385d6 ED |
286 | return true; |
287 | return false; | |
bdf1ee5d IJ |
288 | } |
289 | ||
1da177e4 LT |
290 | /* Buffer size and advertised window tuning. |
291 | * | |
292 | * 1. Tuning sk->sk_sndbuf, when connection enters established state. | |
293 | */ | |
294 | ||
6ae70532 | 295 | static void tcp_sndbuf_expand(struct sock *sk) |
1da177e4 | 296 | { |
6ae70532 | 297 | const struct tcp_sock *tp = tcp_sk(sk); |
77bfc174 | 298 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; |
6ae70532 ED |
299 | int sndmem, per_mss; |
300 | u32 nr_segs; | |
301 | ||
302 | /* Worst case is non GSO/TSO : each frame consumes one skb | |
303 | * and skb->head is kmalloced using power of two area of memory | |
304 | */ | |
305 | per_mss = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + | |
306 | MAX_TCP_HEADER + | |
307 | SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); | |
308 | ||
309 | per_mss = roundup_pow_of_two(per_mss) + | |
310 | SKB_DATA_ALIGN(sizeof(struct sk_buff)); | |
311 | ||
312 | nr_segs = max_t(u32, TCP_INIT_CWND, tp->snd_cwnd); | |
313 | nr_segs = max_t(u32, nr_segs, tp->reordering + 1); | |
314 | ||
315 | /* Fast Recovery (RFC 5681 3.2) : | |
316 | * Cubic needs 1.7 factor, rounded to 2 to include | |
317 | * extra cushion (application might react slowly to POLLOUT) | |
318 | */ | |
77bfc174 YC |
319 | sndmem = ca_ops->sndbuf_expand ? ca_ops->sndbuf_expand(sk) : 2; |
320 | sndmem *= nr_segs * per_mss; | |
1da177e4 | 321 | |
06a59ecb ED |
322 | if (sk->sk_sndbuf < sndmem) |
323 | sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); | |
1da177e4 LT |
324 | } |
325 | ||
326 | /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) | |
327 | * | |
328 | * All tcp_full_space() is split to two parts: "network" buffer, allocated | |
329 | * forward and advertised in receiver window (tp->rcv_wnd) and | |
330 | * "application buffer", required to isolate scheduling/application | |
331 | * latencies from network. | |
332 | * window_clamp is maximal advertised window. It can be less than | |
333 | * tcp_full_space(), in this case tcp_full_space() - window_clamp | |
334 | * is reserved for "application" buffer. The less window_clamp is | |
335 | * the smoother our behaviour from viewpoint of network, but the lower | |
336 | * throughput and the higher sensitivity of the connection to losses. 8) | |
337 | * | |
338 | * rcv_ssthresh is more strict window_clamp used at "slow start" | |
339 | * phase to predict further behaviour of this connection. | |
340 | * It is used for two goals: | |
341 | * - to enforce header prediction at sender, even when application | |
342 | * requires some significant "application buffer". It is check #1. | |
343 | * - to prevent pruning of receive queue because of misprediction | |
344 | * of receiver window. Check #2. | |
345 | * | |
346 | * The scheme does not work when sender sends good segments opening | |
caa20d9a | 347 | * window and then starts to feed us spaghetti. But it should work |
1da177e4 LT |
348 | * in common situations. Otherwise, we have to rely on queue collapsing. |
349 | */ | |
350 | ||
351 | /* Slow part of check#2. */ | |
9e412ba7 | 352 | static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 353 | { |
9e412ba7 | 354 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 355 | /* Optimize this! */ |
94f0893e ED |
356 | int truesize = tcp_win_from_space(sk, skb->truesize) >> 1; |
357 | int window = tcp_win_from_space(sk, sysctl_tcp_rmem[2]) >> 1; | |
1da177e4 LT |
358 | |
359 | while (tp->rcv_ssthresh <= window) { | |
360 | if (truesize <= skb->len) | |
463c84b9 | 361 | return 2 * inet_csk(sk)->icsk_ack.rcv_mss; |
1da177e4 LT |
362 | |
363 | truesize >>= 1; | |
364 | window >>= 1; | |
365 | } | |
366 | return 0; | |
367 | } | |
368 | ||
cf533ea5 | 369 | static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 370 | { |
9e412ba7 IJ |
371 | struct tcp_sock *tp = tcp_sk(sk); |
372 | ||
1da177e4 LT |
373 | /* Check #1 */ |
374 | if (tp->rcv_ssthresh < tp->window_clamp && | |
375 | (int)tp->rcv_ssthresh < tcp_space(sk) && | |
b8da51eb | 376 | !tcp_under_memory_pressure(sk)) { |
1da177e4 LT |
377 | int incr; |
378 | ||
379 | /* Check #2. Increase window, if skb with such overhead | |
380 | * will fit to rcvbuf in future. | |
381 | */ | |
94f0893e | 382 | if (tcp_win_from_space(sk, skb->truesize) <= skb->len) |
056834d9 | 383 | incr = 2 * tp->advmss; |
1da177e4 | 384 | else |
9e412ba7 | 385 | incr = __tcp_grow_window(sk, skb); |
1da177e4 LT |
386 | |
387 | if (incr) { | |
4d846f02 | 388 | incr = max_t(int, incr, 2 * skb->len); |
056834d9 IJ |
389 | tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, |
390 | tp->window_clamp); | |
463c84b9 | 391 | inet_csk(sk)->icsk_ack.quick |= 1; |
1da177e4 LT |
392 | } |
393 | } | |
394 | } | |
395 | ||
396 | /* 3. Tuning rcvbuf, when connection enters established state. */ | |
1da177e4 LT |
397 | static void tcp_fixup_rcvbuf(struct sock *sk) |
398 | { | |
e9266a02 | 399 | u32 mss = tcp_sk(sk)->advmss; |
e9266a02 | 400 | int rcvmem; |
1da177e4 | 401 | |
85f16525 YC |
402 | rcvmem = 2 * SKB_TRUESIZE(mss + MAX_TCP_HEADER) * |
403 | tcp_default_init_rwnd(mss); | |
e9266a02 | 404 | |
b0983d3c ED |
405 | /* Dynamic Right Sizing (DRS) has 2 to 3 RTT latency |
406 | * Allow enough cushion so that sender is not limited by our window | |
407 | */ | |
4540c0cf | 408 | if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf) |
b0983d3c ED |
409 | rcvmem <<= 2; |
410 | ||
e9266a02 ED |
411 | if (sk->sk_rcvbuf < rcvmem) |
412 | sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]); | |
1da177e4 LT |
413 | } |
414 | ||
caa20d9a | 415 | /* 4. Try to fixup all. It is made immediately after connection enters |
1da177e4 LT |
416 | * established state. |
417 | */ | |
10467163 | 418 | void tcp_init_buffer_space(struct sock *sk) |
1da177e4 | 419 | { |
0c12654a | 420 | int tcp_app_win = sock_net(sk)->ipv4.sysctl_tcp_app_win; |
1da177e4 LT |
421 | struct tcp_sock *tp = tcp_sk(sk); |
422 | int maxwin; | |
423 | ||
424 | if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) | |
425 | tcp_fixup_rcvbuf(sk); | |
426 | if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) | |
6ae70532 | 427 | tcp_sndbuf_expand(sk); |
1da177e4 LT |
428 | |
429 | tp->rcvq_space.space = tp->rcv_wnd; | |
9a568de4 | 430 | tcp_mstamp_refresh(tp); |
645f4c6f | 431 | tp->rcvq_space.time = tp->tcp_mstamp; |
b0983d3c | 432 | tp->rcvq_space.seq = tp->copied_seq; |
1da177e4 LT |
433 | |
434 | maxwin = tcp_full_space(sk); | |
435 | ||
436 | if (tp->window_clamp >= maxwin) { | |
437 | tp->window_clamp = maxwin; | |
438 | ||
0c12654a | 439 | if (tcp_app_win && maxwin > 4 * tp->advmss) |
1da177e4 | 440 | tp->window_clamp = max(maxwin - |
0c12654a | 441 | (maxwin >> tcp_app_win), |
1da177e4 LT |
442 | 4 * tp->advmss); |
443 | } | |
444 | ||
445 | /* Force reservation of one segment. */ | |
0c12654a | 446 | if (tcp_app_win && |
1da177e4 LT |
447 | tp->window_clamp > 2 * tp->advmss && |
448 | tp->window_clamp + tp->advmss > maxwin) | |
449 | tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); | |
450 | ||
451 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); | |
c2203cf7 | 452 | tp->snd_cwnd_stamp = tcp_jiffies32; |
1da177e4 LT |
453 | } |
454 | ||
1da177e4 | 455 | /* 5. Recalculate window clamp after socket hit its memory bounds. */ |
9e412ba7 | 456 | static void tcp_clamp_window(struct sock *sk) |
1da177e4 | 457 | { |
9e412ba7 | 458 | struct tcp_sock *tp = tcp_sk(sk); |
6687e988 | 459 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 460 | |
6687e988 | 461 | icsk->icsk_ack.quick = 0; |
1da177e4 | 462 | |
326f36e9 JH |
463 | if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && |
464 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && | |
b8da51eb | 465 | !tcp_under_memory_pressure(sk) && |
180d8cd9 | 466 | sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) { |
326f36e9 JH |
467 | sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), |
468 | sysctl_tcp_rmem[2]); | |
1da177e4 | 469 | } |
326f36e9 | 470 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) |
056834d9 | 471 | tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); |
1da177e4 LT |
472 | } |
473 | ||
40efc6fa SH |
474 | /* Initialize RCV_MSS value. |
475 | * RCV_MSS is an our guess about MSS used by the peer. | |
476 | * We haven't any direct information about the MSS. | |
477 | * It's better to underestimate the RCV_MSS rather than overestimate. | |
478 | * Overestimations make us ACKing less frequently than needed. | |
479 | * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). | |
480 | */ | |
481 | void tcp_initialize_rcv_mss(struct sock *sk) | |
482 | { | |
cf533ea5 | 483 | const struct tcp_sock *tp = tcp_sk(sk); |
40efc6fa SH |
484 | unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); |
485 | ||
056834d9 | 486 | hint = min(hint, tp->rcv_wnd / 2); |
bee7ca9e | 487 | hint = min(hint, TCP_MSS_DEFAULT); |
40efc6fa SH |
488 | hint = max(hint, TCP_MIN_MSS); |
489 | ||
490 | inet_csk(sk)->icsk_ack.rcv_mss = hint; | |
491 | } | |
4bc2f18b | 492 | EXPORT_SYMBOL(tcp_initialize_rcv_mss); |
40efc6fa | 493 | |
1da177e4 LT |
494 | /* Receiver "autotuning" code. |
495 | * | |
496 | * The algorithm for RTT estimation w/o timestamps is based on | |
497 | * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. | |
631dd1a8 | 498 | * <http://public.lanl.gov/radiant/pubs.html#DRS> |
1da177e4 LT |
499 | * |
500 | * More detail on this code can be found at | |
631dd1a8 | 501 | * <http://staff.psc.edu/jheffner/>, |
1da177e4 LT |
502 | * though this reference is out of date. A new paper |
503 | * is pending. | |
504 | */ | |
505 | static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) | |
506 | { | |
645f4c6f | 507 | u32 new_sample = tp->rcv_rtt_est.rtt_us; |
1da177e4 LT |
508 | long m = sample; |
509 | ||
510 | if (m == 0) | |
511 | m = 1; | |
512 | ||
513 | if (new_sample != 0) { | |
514 | /* If we sample in larger samples in the non-timestamp | |
515 | * case, we could grossly overestimate the RTT especially | |
516 | * with chatty applications or bulk transfer apps which | |
517 | * are stalled on filesystem I/O. | |
518 | * | |
519 | * Also, since we are only going for a minimum in the | |
31f34269 | 520 | * non-timestamp case, we do not smooth things out |
caa20d9a | 521 | * else with timestamps disabled convergence takes too |
1da177e4 LT |
522 | * long. |
523 | */ | |
524 | if (!win_dep) { | |
525 | m -= (new_sample >> 3); | |
526 | new_sample += m; | |
18a223e0 NC |
527 | } else { |
528 | m <<= 3; | |
529 | if (m < new_sample) | |
530 | new_sample = m; | |
531 | } | |
1da177e4 | 532 | } else { |
caa20d9a | 533 | /* No previous measure. */ |
1da177e4 LT |
534 | new_sample = m << 3; |
535 | } | |
536 | ||
645f4c6f | 537 | tp->rcv_rtt_est.rtt_us = new_sample; |
1da177e4 LT |
538 | } |
539 | ||
540 | static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) | |
541 | { | |
645f4c6f ED |
542 | u32 delta_us; |
543 | ||
9a568de4 | 544 | if (tp->rcv_rtt_est.time == 0) |
1da177e4 LT |
545 | goto new_measure; |
546 | if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) | |
547 | return; | |
9a568de4 | 548 | delta_us = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcv_rtt_est.time); |
645f4c6f | 549 | tcp_rcv_rtt_update(tp, delta_us, 1); |
1da177e4 LT |
550 | |
551 | new_measure: | |
552 | tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; | |
645f4c6f | 553 | tp->rcv_rtt_est.time = tp->tcp_mstamp; |
1da177e4 LT |
554 | } |
555 | ||
056834d9 IJ |
556 | static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, |
557 | const struct sk_buff *skb) | |
1da177e4 | 558 | { |
463c84b9 | 559 | struct tcp_sock *tp = tcp_sk(sk); |
9a568de4 | 560 | |
1da177e4 LT |
561 | if (tp->rx_opt.rcv_tsecr && |
562 | (TCP_SKB_CB(skb)->end_seq - | |
9a568de4 ED |
563 | TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) { |
564 | u32 delta = tcp_time_stamp(tp) - tp->rx_opt.rcv_tsecr; | |
565 | u32 delta_us = delta * (USEC_PER_SEC / TCP_TS_HZ); | |
566 | ||
567 | tcp_rcv_rtt_update(tp, delta_us, 0); | |
568 | } | |
1da177e4 LT |
569 | } |
570 | ||
571 | /* | |
572 | * This function should be called every time data is copied to user space. | |
573 | * It calculates the appropriate TCP receive buffer space. | |
574 | */ | |
575 | void tcp_rcv_space_adjust(struct sock *sk) | |
576 | { | |
577 | struct tcp_sock *tp = tcp_sk(sk); | |
578 | int time; | |
b0983d3c | 579 | int copied; |
e905a9ed | 580 | |
9a568de4 | 581 | time = tcp_stamp_us_delta(tp->tcp_mstamp, tp->rcvq_space.time); |
645f4c6f | 582 | if (time < (tp->rcv_rtt_est.rtt_us >> 3) || tp->rcv_rtt_est.rtt_us == 0) |
1da177e4 | 583 | return; |
e905a9ed | 584 | |
b0983d3c ED |
585 | /* Number of bytes copied to user in last RTT */ |
586 | copied = tp->copied_seq - tp->rcvq_space.seq; | |
587 | if (copied <= tp->rcvq_space.space) | |
588 | goto new_measure; | |
589 | ||
590 | /* A bit of theory : | |
591 | * copied = bytes received in previous RTT, our base window | |
592 | * To cope with packet losses, we need a 2x factor | |
593 | * To cope with slow start, and sender growing its cwin by 100 % | |
594 | * every RTT, we need a 4x factor, because the ACK we are sending | |
595 | * now is for the next RTT, not the current one : | |
596 | * <prev RTT . ><current RTT .. ><next RTT .... > | |
597 | */ | |
598 | ||
4540c0cf | 599 | if (sock_net(sk)->ipv4.sysctl_tcp_moderate_rcvbuf && |
b0983d3c ED |
600 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { |
601 | int rcvwin, rcvmem, rcvbuf; | |
1da177e4 | 602 | |
b0983d3c ED |
603 | /* minimal window to cope with packet losses, assuming |
604 | * steady state. Add some cushion because of small variations. | |
605 | */ | |
606 | rcvwin = (copied << 1) + 16 * tp->advmss; | |
1da177e4 | 607 | |
b0983d3c ED |
608 | /* If rate increased by 25%, |
609 | * assume slow start, rcvwin = 3 * copied | |
610 | * If rate increased by 50%, | |
611 | * assume sender can use 2x growth, rcvwin = 4 * copied | |
612 | */ | |
613 | if (copied >= | |
614 | tp->rcvq_space.space + (tp->rcvq_space.space >> 2)) { | |
615 | if (copied >= | |
616 | tp->rcvq_space.space + (tp->rcvq_space.space >> 1)) | |
617 | rcvwin <<= 1; | |
618 | else | |
619 | rcvwin += (rcvwin >> 1); | |
620 | } | |
1da177e4 | 621 | |
b0983d3c | 622 | rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER); |
94f0893e | 623 | while (tcp_win_from_space(sk, rcvmem) < tp->advmss) |
b0983d3c | 624 | rcvmem += 128; |
1da177e4 | 625 | |
b0983d3c ED |
626 | rcvbuf = min(rcvwin / tp->advmss * rcvmem, sysctl_tcp_rmem[2]); |
627 | if (rcvbuf > sk->sk_rcvbuf) { | |
628 | sk->sk_rcvbuf = rcvbuf; | |
1da177e4 | 629 | |
b0983d3c ED |
630 | /* Make the window clamp follow along. */ |
631 | tp->window_clamp = rcvwin; | |
1da177e4 LT |
632 | } |
633 | } | |
b0983d3c | 634 | tp->rcvq_space.space = copied; |
e905a9ed | 635 | |
1da177e4 LT |
636 | new_measure: |
637 | tp->rcvq_space.seq = tp->copied_seq; | |
645f4c6f | 638 | tp->rcvq_space.time = tp->tcp_mstamp; |
1da177e4 LT |
639 | } |
640 | ||
641 | /* There is something which you must keep in mind when you analyze the | |
642 | * behavior of the tp->ato delayed ack timeout interval. When a | |
643 | * connection starts up, we want to ack as quickly as possible. The | |
644 | * problem is that "good" TCP's do slow start at the beginning of data | |
645 | * transmission. The means that until we send the first few ACK's the | |
646 | * sender will sit on his end and only queue most of his data, because | |
647 | * he can only send snd_cwnd unacked packets at any given time. For | |
648 | * each ACK we send, he increments snd_cwnd and transmits more of his | |
649 | * queue. -DaveM | |
650 | */ | |
9e412ba7 | 651 | static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 652 | { |
9e412ba7 | 653 | struct tcp_sock *tp = tcp_sk(sk); |
463c84b9 | 654 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
655 | u32 now; |
656 | ||
463c84b9 | 657 | inet_csk_schedule_ack(sk); |
1da177e4 | 658 | |
463c84b9 | 659 | tcp_measure_rcv_mss(sk, skb); |
1da177e4 LT |
660 | |
661 | tcp_rcv_rtt_measure(tp); | |
e905a9ed | 662 | |
70eabf0e | 663 | now = tcp_jiffies32; |
1da177e4 | 664 | |
463c84b9 | 665 | if (!icsk->icsk_ack.ato) { |
1da177e4 LT |
666 | /* The _first_ data packet received, initialize |
667 | * delayed ACK engine. | |
668 | */ | |
463c84b9 ACM |
669 | tcp_incr_quickack(sk); |
670 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 | 671 | } else { |
463c84b9 | 672 | int m = now - icsk->icsk_ack.lrcvtime; |
1da177e4 | 673 | |
056834d9 | 674 | if (m <= TCP_ATO_MIN / 2) { |
1da177e4 | 675 | /* The fastest case is the first. */ |
463c84b9 ACM |
676 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; |
677 | } else if (m < icsk->icsk_ack.ato) { | |
678 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; | |
679 | if (icsk->icsk_ack.ato > icsk->icsk_rto) | |
680 | icsk->icsk_ack.ato = icsk->icsk_rto; | |
681 | } else if (m > icsk->icsk_rto) { | |
caa20d9a | 682 | /* Too long gap. Apparently sender failed to |
1da177e4 LT |
683 | * restart window, so that we send ACKs quickly. |
684 | */ | |
463c84b9 | 685 | tcp_incr_quickack(sk); |
3ab224be | 686 | sk_mem_reclaim(sk); |
1da177e4 LT |
687 | } |
688 | } | |
463c84b9 | 689 | icsk->icsk_ack.lrcvtime = now; |
1da177e4 | 690 | |
735d3831 | 691 | tcp_ecn_check_ce(tp, skb); |
1da177e4 LT |
692 | |
693 | if (skb->len >= 128) | |
9e412ba7 | 694 | tcp_grow_window(sk, skb); |
1da177e4 LT |
695 | } |
696 | ||
1da177e4 LT |
697 | /* Called to compute a smoothed rtt estimate. The data fed to this |
698 | * routine either comes from timestamps, or from segments that were | |
699 | * known _not_ to have been retransmitted [see Karn/Partridge | |
700 | * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 | |
701 | * piece by Van Jacobson. | |
702 | * NOTE: the next three routines used to be one big routine. | |
703 | * To save cycles in the RFC 1323 implementation it was better to break | |
704 | * it up into three procedures. -- erics | |
705 | */ | |
740b0f18 | 706 | static void tcp_rtt_estimator(struct sock *sk, long mrtt_us) |
1da177e4 | 707 | { |
6687e988 | 708 | struct tcp_sock *tp = tcp_sk(sk); |
740b0f18 ED |
709 | long m = mrtt_us; /* RTT */ |
710 | u32 srtt = tp->srtt_us; | |
1da177e4 | 711 | |
1da177e4 LT |
712 | /* The following amusing code comes from Jacobson's |
713 | * article in SIGCOMM '88. Note that rtt and mdev | |
714 | * are scaled versions of rtt and mean deviation. | |
e905a9ed | 715 | * This is designed to be as fast as possible |
1da177e4 LT |
716 | * m stands for "measurement". |
717 | * | |
718 | * On a 1990 paper the rto value is changed to: | |
719 | * RTO = rtt + 4 * mdev | |
720 | * | |
721 | * Funny. This algorithm seems to be very broken. | |
722 | * These formulae increase RTO, when it should be decreased, increase | |
31f34269 | 723 | * too slowly, when it should be increased quickly, decrease too quickly |
1da177e4 LT |
724 | * etc. I guess in BSD RTO takes ONE value, so that it is absolutely |
725 | * does not matter how to _calculate_ it. Seems, it was trap | |
726 | * that VJ failed to avoid. 8) | |
727 | */ | |
4a5ab4e2 ED |
728 | if (srtt != 0) { |
729 | m -= (srtt >> 3); /* m is now error in rtt est */ | |
730 | srtt += m; /* rtt = 7/8 rtt + 1/8 new */ | |
1da177e4 LT |
731 | if (m < 0) { |
732 | m = -m; /* m is now abs(error) */ | |
740b0f18 | 733 | m -= (tp->mdev_us >> 2); /* similar update on mdev */ |
1da177e4 LT |
734 | /* This is similar to one of Eifel findings. |
735 | * Eifel blocks mdev updates when rtt decreases. | |
736 | * This solution is a bit different: we use finer gain | |
737 | * for mdev in this case (alpha*beta). | |
738 | * Like Eifel it also prevents growth of rto, | |
739 | * but also it limits too fast rto decreases, | |
740 | * happening in pure Eifel. | |
741 | */ | |
742 | if (m > 0) | |
743 | m >>= 3; | |
744 | } else { | |
740b0f18 | 745 | m -= (tp->mdev_us >> 2); /* similar update on mdev */ |
1da177e4 | 746 | } |
740b0f18 ED |
747 | tp->mdev_us += m; /* mdev = 3/4 mdev + 1/4 new */ |
748 | if (tp->mdev_us > tp->mdev_max_us) { | |
749 | tp->mdev_max_us = tp->mdev_us; | |
750 | if (tp->mdev_max_us > tp->rttvar_us) | |
751 | tp->rttvar_us = tp->mdev_max_us; | |
1da177e4 LT |
752 | } |
753 | if (after(tp->snd_una, tp->rtt_seq)) { | |
740b0f18 ED |
754 | if (tp->mdev_max_us < tp->rttvar_us) |
755 | tp->rttvar_us -= (tp->rttvar_us - tp->mdev_max_us) >> 2; | |
1da177e4 | 756 | tp->rtt_seq = tp->snd_nxt; |
740b0f18 | 757 | tp->mdev_max_us = tcp_rto_min_us(sk); |
1da177e4 LT |
758 | } |
759 | } else { | |
760 | /* no previous measure. */ | |
4a5ab4e2 | 761 | srtt = m << 3; /* take the measured time to be rtt */ |
740b0f18 ED |
762 | tp->mdev_us = m << 1; /* make sure rto = 3*rtt */ |
763 | tp->rttvar_us = max(tp->mdev_us, tcp_rto_min_us(sk)); | |
764 | tp->mdev_max_us = tp->rttvar_us; | |
1da177e4 LT |
765 | tp->rtt_seq = tp->snd_nxt; |
766 | } | |
740b0f18 | 767 | tp->srtt_us = max(1U, srtt); |
1da177e4 LT |
768 | } |
769 | ||
95bd09eb ED |
770 | /* Set the sk_pacing_rate to allow proper sizing of TSO packets. |
771 | * Note: TCP stack does not yet implement pacing. | |
772 | * FQ packet scheduler can be used to implement cheap but effective | |
773 | * TCP pacing, to smooth the burst on large writes when packets | |
774 | * in flight is significantly lower than cwnd (or rwin) | |
775 | */ | |
43e122b0 ED |
776 | int sysctl_tcp_pacing_ss_ratio __read_mostly = 200; |
777 | int sysctl_tcp_pacing_ca_ratio __read_mostly = 120; | |
778 | ||
95bd09eb ED |
779 | static void tcp_update_pacing_rate(struct sock *sk) |
780 | { | |
781 | const struct tcp_sock *tp = tcp_sk(sk); | |
782 | u64 rate; | |
783 | ||
784 | /* set sk_pacing_rate to 200 % of current rate (mss * cwnd / srtt) */ | |
43e122b0 ED |
785 | rate = (u64)tp->mss_cache * ((USEC_PER_SEC / 100) << 3); |
786 | ||
787 | /* current rate is (cwnd * mss) / srtt | |
788 | * In Slow Start [1], set sk_pacing_rate to 200 % the current rate. | |
789 | * In Congestion Avoidance phase, set it to 120 % the current rate. | |
790 | * | |
791 | * [1] : Normal Slow Start condition is (tp->snd_cwnd < tp->snd_ssthresh) | |
792 | * If snd_cwnd >= (tp->snd_ssthresh / 2), we are approaching | |
793 | * end of slow start and should slow down. | |
794 | */ | |
795 | if (tp->snd_cwnd < tp->snd_ssthresh / 2) | |
796 | rate *= sysctl_tcp_pacing_ss_ratio; | |
797 | else | |
798 | rate *= sysctl_tcp_pacing_ca_ratio; | |
95bd09eb ED |
799 | |
800 | rate *= max(tp->snd_cwnd, tp->packets_out); | |
801 | ||
740b0f18 ED |
802 | if (likely(tp->srtt_us)) |
803 | do_div(rate, tp->srtt_us); | |
95bd09eb | 804 | |
ba537427 ED |
805 | /* ACCESS_ONCE() is needed because sch_fq fetches sk_pacing_rate |
806 | * without any lock. We want to make sure compiler wont store | |
807 | * intermediate values in this location. | |
808 | */ | |
809 | ACCESS_ONCE(sk->sk_pacing_rate) = min_t(u64, rate, | |
810 | sk->sk_max_pacing_rate); | |
95bd09eb ED |
811 | } |
812 | ||
1da177e4 LT |
813 | /* Calculate rto without backoff. This is the second half of Van Jacobson's |
814 | * routine referred to above. | |
815 | */ | |
f7e56a76 | 816 | static void tcp_set_rto(struct sock *sk) |
1da177e4 | 817 | { |
463c84b9 | 818 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
819 | /* Old crap is replaced with new one. 8) |
820 | * | |
821 | * More seriously: | |
822 | * 1. If rtt variance happened to be less 50msec, it is hallucination. | |
823 | * It cannot be less due to utterly erratic ACK generation made | |
824 | * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ | |
825 | * to do with delayed acks, because at cwnd>2 true delack timeout | |
826 | * is invisible. Actually, Linux-2.4 also generates erratic | |
caa20d9a | 827 | * ACKs in some circumstances. |
1da177e4 | 828 | */ |
f1ecd5d9 | 829 | inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); |
1da177e4 LT |
830 | |
831 | /* 2. Fixups made earlier cannot be right. | |
832 | * If we do not estimate RTO correctly without them, | |
833 | * all the algo is pure shit and should be replaced | |
caa20d9a | 834 | * with correct one. It is exactly, which we pretend to do. |
1da177e4 | 835 | */ |
1da177e4 | 836 | |
ee6aac59 IJ |
837 | /* NOTE: clamping at TCP_RTO_MIN is not required, current algo |
838 | * guarantees that rto is higher. | |
839 | */ | |
f1ecd5d9 | 840 | tcp_bound_rto(sk); |
1da177e4 LT |
841 | } |
842 | ||
cf533ea5 | 843 | __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) |
1da177e4 LT |
844 | { |
845 | __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); | |
846 | ||
22b71c8f | 847 | if (!cwnd) |
442b9635 | 848 | cwnd = TCP_INIT_CWND; |
1da177e4 LT |
849 | return min_t(__u32, cwnd, tp->snd_cwnd_clamp); |
850 | } | |
851 | ||
e60402d0 IJ |
852 | /* |
853 | * Packet counting of FACK is based on in-order assumptions, therefore TCP | |
854 | * disables it when reordering is detected | |
855 | */ | |
4aabd8ef | 856 | void tcp_disable_fack(struct tcp_sock *tp) |
e60402d0 | 857 | { |
85cc391c IJ |
858 | /* RFC3517 uses different metric in lost marker => reset on change */ |
859 | if (tcp_is_fack(tp)) | |
860 | tp->lost_skb_hint = NULL; | |
ab56222a | 861 | tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED; |
e60402d0 IJ |
862 | } |
863 | ||
564262c1 | 864 | /* Take a notice that peer is sending D-SACKs */ |
e60402d0 IJ |
865 | static void tcp_dsack_seen(struct tcp_sock *tp) |
866 | { | |
ab56222a | 867 | tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; |
e60402d0 IJ |
868 | } |
869 | ||
6687e988 ACM |
870 | static void tcp_update_reordering(struct sock *sk, const int metric, |
871 | const int ts) | |
1da177e4 | 872 | { |
6687e988 | 873 | struct tcp_sock *tp = tcp_sk(sk); |
2d2517ee | 874 | int mib_idx; |
40b215e5 | 875 | |
6f5b24ee SHY |
876 | if (WARN_ON_ONCE(metric < 0)) |
877 | return; | |
878 | ||
2d2517ee | 879 | if (metric > tp->reordering) { |
c6e21803 | 880 | tp->reordering = min(sock_net(sk)->ipv4.sysctl_tcp_max_reordering, metric); |
1da177e4 | 881 | |
1da177e4 | 882 | #if FASTRETRANS_DEBUG > 1 |
91df42be JP |
883 | pr_debug("Disorder%d %d %u f%u s%u rr%d\n", |
884 | tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, | |
885 | tp->reordering, | |
886 | tp->fackets_out, | |
887 | tp->sacked_out, | |
888 | tp->undo_marker ? tp->undo_retrans : 0); | |
1da177e4 | 889 | #endif |
e60402d0 | 890 | tcp_disable_fack(tp); |
1da177e4 | 891 | } |
eed530b6 | 892 | |
4f41b1c5 | 893 | tp->rack.reord = 1; |
2d2517ee YC |
894 | |
895 | /* This exciting event is worth to be remembered. 8) */ | |
896 | if (ts) | |
897 | mib_idx = LINUX_MIB_TCPTSREORDER; | |
898 | else if (tcp_is_reno(tp)) | |
899 | mib_idx = LINUX_MIB_TCPRENOREORDER; | |
900 | else if (tcp_is_fack(tp)) | |
901 | mib_idx = LINUX_MIB_TCPFACKREORDER; | |
902 | else | |
903 | mib_idx = LINUX_MIB_TCPSACKREORDER; | |
904 | ||
905 | NET_INC_STATS(sock_net(sk), mib_idx); | |
1da177e4 LT |
906 | } |
907 | ||
006f582c | 908 | /* This must be called before lost_out is incremented */ |
c8c213f2 IJ |
909 | static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) |
910 | { | |
51456b29 | 911 | if (!tp->retransmit_skb_hint || |
c8c213f2 IJ |
912 | before(TCP_SKB_CB(skb)->seq, |
913 | TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) | |
006f582c | 914 | tp->retransmit_skb_hint = skb; |
c8c213f2 IJ |
915 | } |
916 | ||
0682e690 NC |
917 | /* Sum the number of packets on the wire we have marked as lost. |
918 | * There are two cases we care about here: | |
919 | * a) Packet hasn't been marked lost (nor retransmitted), | |
920 | * and this is the first loss. | |
921 | * b) Packet has been marked both lost and retransmitted, | |
922 | * and this means we think it was lost again. | |
923 | */ | |
924 | static void tcp_sum_lost(struct tcp_sock *tp, struct sk_buff *skb) | |
925 | { | |
926 | __u8 sacked = TCP_SKB_CB(skb)->sacked; | |
927 | ||
928 | if (!(sacked & TCPCB_LOST) || | |
929 | ((sacked & TCPCB_LOST) && (sacked & TCPCB_SACKED_RETRANS))) | |
930 | tp->lost += tcp_skb_pcount(skb); | |
931 | } | |
932 | ||
41ea36e3 IJ |
933 | static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb) |
934 | { | |
935 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { | |
936 | tcp_verify_retransmit_hint(tp, skb); | |
937 | ||
938 | tp->lost_out += tcp_skb_pcount(skb); | |
0682e690 | 939 | tcp_sum_lost(tp, skb); |
41ea36e3 IJ |
940 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; |
941 | } | |
942 | } | |
943 | ||
4f41b1c5 | 944 | void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, struct sk_buff *skb) |
006f582c IJ |
945 | { |
946 | tcp_verify_retransmit_hint(tp, skb); | |
947 | ||
0682e690 | 948 | tcp_sum_lost(tp, skb); |
006f582c IJ |
949 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { |
950 | tp->lost_out += tcp_skb_pcount(skb); | |
951 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
952 | } | |
953 | } | |
954 | ||
1da177e4 LT |
955 | /* This procedure tags the retransmission queue when SACKs arrive. |
956 | * | |
957 | * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). | |
958 | * Packets in queue with these bits set are counted in variables | |
959 | * sacked_out, retrans_out and lost_out, correspondingly. | |
960 | * | |
961 | * Valid combinations are: | |
962 | * Tag InFlight Description | |
963 | * 0 1 - orig segment is in flight. | |
964 | * S 0 - nothing flies, orig reached receiver. | |
965 | * L 0 - nothing flies, orig lost by net. | |
966 | * R 2 - both orig and retransmit are in flight. | |
967 | * L|R 1 - orig is lost, retransmit is in flight. | |
968 | * S|R 1 - orig reached receiver, retrans is still in flight. | |
969 | * (L|S|R is logically valid, it could occur when L|R is sacked, | |
970 | * but it is equivalent to plain S and code short-curcuits it to S. | |
971 | * L|S is logically invalid, it would mean -1 packet in flight 8)) | |
972 | * | |
973 | * These 6 states form finite state machine, controlled by the following events: | |
974 | * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) | |
975 | * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) | |
974c1236 | 976 | * 3. Loss detection event of two flavors: |
1da177e4 LT |
977 | * A. Scoreboard estimator decided the packet is lost. |
978 | * A'. Reno "three dupacks" marks head of queue lost. | |
974c1236 YC |
979 | * A''. Its FACK modification, head until snd.fack is lost. |
980 | * B. SACK arrives sacking SND.NXT at the moment, when the | |
1da177e4 LT |
981 | * segment was retransmitted. |
982 | * 4. D-SACK added new rule: D-SACK changes any tag to S. | |
983 | * | |
984 | * It is pleasant to note, that state diagram turns out to be commutative, | |
985 | * so that we are allowed not to be bothered by order of our actions, | |
986 | * when multiple events arrive simultaneously. (see the function below). | |
987 | * | |
988 | * Reordering detection. | |
989 | * -------------------- | |
990 | * Reordering metric is maximal distance, which a packet can be displaced | |
991 | * in packet stream. With SACKs we can estimate it: | |
992 | * | |
993 | * 1. SACK fills old hole and the corresponding segment was not | |
994 | * ever retransmitted -> reordering. Alas, we cannot use it | |
995 | * when segment was retransmitted. | |
996 | * 2. The last flaw is solved with D-SACK. D-SACK arrives | |
997 | * for retransmitted and already SACKed segment -> reordering.. | |
998 | * Both of these heuristics are not used in Loss state, when we cannot | |
999 | * account for retransmits accurately. | |
5b3c9882 IJ |
1000 | * |
1001 | * SACK block validation. | |
1002 | * ---------------------- | |
1003 | * | |
1004 | * SACK block range validation checks that the received SACK block fits to | |
1005 | * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. | |
1006 | * Note that SND.UNA is not included to the range though being valid because | |
0e835331 IJ |
1007 | * it means that the receiver is rather inconsistent with itself reporting |
1008 | * SACK reneging when it should advance SND.UNA. Such SACK block this is | |
1009 | * perfectly valid, however, in light of RFC2018 which explicitly states | |
1010 | * that "SACK block MUST reflect the newest segment. Even if the newest | |
1011 | * segment is going to be discarded ...", not that it looks very clever | |
1012 | * in case of head skb. Due to potentional receiver driven attacks, we | |
1013 | * choose to avoid immediate execution of a walk in write queue due to | |
1014 | * reneging and defer head skb's loss recovery to standard loss recovery | |
1015 | * procedure that will eventually trigger (nothing forbids us doing this). | |
5b3c9882 IJ |
1016 | * |
1017 | * Implements also blockage to start_seq wrap-around. Problem lies in the | |
1018 | * fact that though start_seq (s) is before end_seq (i.e., not reversed), | |
1019 | * there's no guarantee that it will be before snd_nxt (n). The problem | |
1020 | * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt | |
1021 | * wrap (s_w): | |
1022 | * | |
1023 | * <- outs wnd -> <- wrapzone -> | |
1024 | * u e n u_w e_w s n_w | |
1025 | * | | | | | | | | |
1026 | * |<------------+------+----- TCP seqno space --------------+---------->| | |
1027 | * ...-- <2^31 ->| |<--------... | |
1028 | * ...---- >2^31 ------>| |<--------... | |
1029 | * | |
1030 | * Current code wouldn't be vulnerable but it's better still to discard such | |
1031 | * crazy SACK blocks. Doing this check for start_seq alone closes somewhat | |
1032 | * similar case (end_seq after snd_nxt wrap) as earlier reversed check in | |
1033 | * snd_nxt wrap -> snd_una region will then become "well defined", i.e., | |
1034 | * equal to the ideal case (infinite seqno space without wrap caused issues). | |
1035 | * | |
1036 | * With D-SACK the lower bound is extended to cover sequence space below | |
1037 | * SND.UNA down to undo_marker, which is the last point of interest. Yet | |
564262c1 | 1038 | * again, D-SACK block must not to go across snd_una (for the same reason as |
5b3c9882 IJ |
1039 | * for the normal SACK blocks, explained above). But there all simplicity |
1040 | * ends, TCP might receive valid D-SACKs below that. As long as they reside | |
1041 | * fully below undo_marker they do not affect behavior in anyway and can | |
1042 | * therefore be safely ignored. In rare cases (which are more or less | |
1043 | * theoretical ones), the D-SACK will nicely cross that boundary due to skb | |
1044 | * fragmentation and packet reordering past skb's retransmission. To consider | |
1045 | * them correctly, the acceptable range must be extended even more though | |
1046 | * the exact amount is rather hard to quantify. However, tp->max_window can | |
1047 | * be used as an exaggerated estimate. | |
1da177e4 | 1048 | */ |
a2a385d6 ED |
1049 | static bool tcp_is_sackblock_valid(struct tcp_sock *tp, bool is_dsack, |
1050 | u32 start_seq, u32 end_seq) | |
5b3c9882 IJ |
1051 | { |
1052 | /* Too far in future, or reversed (interpretation is ambiguous) */ | |
1053 | if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) | |
a2a385d6 | 1054 | return false; |
5b3c9882 IJ |
1055 | |
1056 | /* Nasty start_seq wrap-around check (see comments above) */ | |
1057 | if (!before(start_seq, tp->snd_nxt)) | |
a2a385d6 | 1058 | return false; |
5b3c9882 | 1059 | |
564262c1 | 1060 | /* In outstanding window? ...This is valid exit for D-SACKs too. |
5b3c9882 IJ |
1061 | * start_seq == snd_una is non-sensical (see comments above) |
1062 | */ | |
1063 | if (after(start_seq, tp->snd_una)) | |
a2a385d6 | 1064 | return true; |
5b3c9882 IJ |
1065 | |
1066 | if (!is_dsack || !tp->undo_marker) | |
a2a385d6 | 1067 | return false; |
5b3c9882 IJ |
1068 | |
1069 | /* ...Then it's D-SACK, and must reside below snd_una completely */ | |
f779b2d6 | 1070 | if (after(end_seq, tp->snd_una)) |
a2a385d6 | 1071 | return false; |
5b3c9882 IJ |
1072 | |
1073 | if (!before(start_seq, tp->undo_marker)) | |
a2a385d6 | 1074 | return true; |
5b3c9882 IJ |
1075 | |
1076 | /* Too old */ | |
1077 | if (!after(end_seq, tp->undo_marker)) | |
a2a385d6 | 1078 | return false; |
5b3c9882 IJ |
1079 | |
1080 | /* Undo_marker boundary crossing (overestimates a lot). Known already: | |
1081 | * start_seq < undo_marker and end_seq >= undo_marker. | |
1082 | */ | |
1083 | return !before(start_seq, end_seq - tp->max_window); | |
1084 | } | |
1085 | ||
a2a385d6 ED |
1086 | static bool tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, |
1087 | struct tcp_sack_block_wire *sp, int num_sacks, | |
1088 | u32 prior_snd_una) | |
d06e021d | 1089 | { |
1ed83465 | 1090 | struct tcp_sock *tp = tcp_sk(sk); |
d3e2ce3b HH |
1091 | u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); |
1092 | u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); | |
a2a385d6 | 1093 | bool dup_sack = false; |
d06e021d DM |
1094 | |
1095 | if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { | |
a2a385d6 | 1096 | dup_sack = true; |
e60402d0 | 1097 | tcp_dsack_seen(tp); |
c10d9310 | 1098 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPDSACKRECV); |
d06e021d | 1099 | } else if (num_sacks > 1) { |
d3e2ce3b HH |
1100 | u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); |
1101 | u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); | |
d06e021d DM |
1102 | |
1103 | if (!after(end_seq_0, end_seq_1) && | |
1104 | !before(start_seq_0, start_seq_1)) { | |
a2a385d6 | 1105 | dup_sack = true; |
e60402d0 | 1106 | tcp_dsack_seen(tp); |
c10d9310 | 1107 | NET_INC_STATS(sock_net(sk), |
de0744af | 1108 | LINUX_MIB_TCPDSACKOFORECV); |
d06e021d DM |
1109 | } |
1110 | } | |
1111 | ||
1112 | /* D-SACK for already forgotten data... Do dumb counting. */ | |
6e08d5e3 | 1113 | if (dup_sack && tp->undo_marker && tp->undo_retrans > 0 && |
d06e021d DM |
1114 | !after(end_seq_0, prior_snd_una) && |
1115 | after(end_seq_0, tp->undo_marker)) | |
1116 | tp->undo_retrans--; | |
1117 | ||
1118 | return dup_sack; | |
1119 | } | |
1120 | ||
a1197f5a | 1121 | struct tcp_sacktag_state { |
740b0f18 ED |
1122 | int reord; |
1123 | int fack_count; | |
31231a8a KKJ |
1124 | /* Timestamps for earliest and latest never-retransmitted segment |
1125 | * that was SACKed. RTO needs the earliest RTT to stay conservative, | |
1126 | * but congestion control should still get an accurate delay signal. | |
1127 | */ | |
9a568de4 ED |
1128 | u64 first_sackt; |
1129 | u64 last_sackt; | |
b9f64820 | 1130 | struct rate_sample *rate; |
740b0f18 | 1131 | int flag; |
75c119af | 1132 | unsigned int mss_now; |
a1197f5a IJ |
1133 | }; |
1134 | ||
d1935942 IJ |
1135 | /* Check if skb is fully within the SACK block. In presence of GSO skbs, |
1136 | * the incoming SACK may not exactly match but we can find smaller MSS | |
1137 | * aligned portion of it that matches. Therefore we might need to fragment | |
1138 | * which may fail and creates some hassle (caller must handle error case | |
1139 | * returns). | |
832d11c5 IJ |
1140 | * |
1141 | * FIXME: this could be merged to shift decision code | |
d1935942 | 1142 | */ |
0f79efdc | 1143 | static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, |
a2a385d6 | 1144 | u32 start_seq, u32 end_seq) |
d1935942 | 1145 | { |
a2a385d6 ED |
1146 | int err; |
1147 | bool in_sack; | |
d1935942 | 1148 | unsigned int pkt_len; |
adb92db8 | 1149 | unsigned int mss; |
d1935942 IJ |
1150 | |
1151 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && | |
1152 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1153 | ||
1154 | if (tcp_skb_pcount(skb) > 1 && !in_sack && | |
1155 | after(TCP_SKB_CB(skb)->end_seq, start_seq)) { | |
adb92db8 | 1156 | mss = tcp_skb_mss(skb); |
d1935942 IJ |
1157 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); |
1158 | ||
adb92db8 | 1159 | if (!in_sack) { |
d1935942 | 1160 | pkt_len = start_seq - TCP_SKB_CB(skb)->seq; |
adb92db8 IJ |
1161 | if (pkt_len < mss) |
1162 | pkt_len = mss; | |
1163 | } else { | |
d1935942 | 1164 | pkt_len = end_seq - TCP_SKB_CB(skb)->seq; |
adb92db8 IJ |
1165 | if (pkt_len < mss) |
1166 | return -EINVAL; | |
1167 | } | |
1168 | ||
1169 | /* Round if necessary so that SACKs cover only full MSSes | |
1170 | * and/or the remaining small portion (if present) | |
1171 | */ | |
1172 | if (pkt_len > mss) { | |
1173 | unsigned int new_len = (pkt_len / mss) * mss; | |
b451e5d2 | 1174 | if (!in_sack && new_len < pkt_len) |
adb92db8 | 1175 | new_len += mss; |
adb92db8 IJ |
1176 | pkt_len = new_len; |
1177 | } | |
b451e5d2 YC |
1178 | |
1179 | if (pkt_len >= skb->len && !in_sack) | |
1180 | return 0; | |
1181 | ||
75c119af ED |
1182 | err = tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, |
1183 | pkt_len, mss, GFP_ATOMIC); | |
d1935942 IJ |
1184 | if (err < 0) |
1185 | return err; | |
1186 | } | |
1187 | ||
1188 | return in_sack; | |
1189 | } | |
1190 | ||
cc9a672e NC |
1191 | /* Mark the given newly-SACKed range as such, adjusting counters and hints. */ |
1192 | static u8 tcp_sacktag_one(struct sock *sk, | |
1193 | struct tcp_sacktag_state *state, u8 sacked, | |
1194 | u32 start_seq, u32 end_seq, | |
740b0f18 | 1195 | int dup_sack, int pcount, |
9a568de4 | 1196 | u64 xmit_time) |
9e10c47c | 1197 | { |
6859d494 | 1198 | struct tcp_sock *tp = tcp_sk(sk); |
a1197f5a | 1199 | int fack_count = state->fack_count; |
9e10c47c IJ |
1200 | |
1201 | /* Account D-SACK for retransmitted packet. */ | |
1202 | if (dup_sack && (sacked & TCPCB_RETRANS)) { | |
6e08d5e3 | 1203 | if (tp->undo_marker && tp->undo_retrans > 0 && |
cc9a672e | 1204 | after(end_seq, tp->undo_marker)) |
9e10c47c | 1205 | tp->undo_retrans--; |
ede9f3b1 | 1206 | if (sacked & TCPCB_SACKED_ACKED) |
a1197f5a | 1207 | state->reord = min(fack_count, state->reord); |
9e10c47c IJ |
1208 | } |
1209 | ||
1210 | /* Nothing to do; acked frame is about to be dropped (was ACKed). */ | |
cc9a672e | 1211 | if (!after(end_seq, tp->snd_una)) |
a1197f5a | 1212 | return sacked; |
9e10c47c IJ |
1213 | |
1214 | if (!(sacked & TCPCB_SACKED_ACKED)) { | |
d2329f10 | 1215 | tcp_rack_advance(tp, sacked, end_seq, xmit_time); |
659a8ad5 | 1216 | |
9e10c47c IJ |
1217 | if (sacked & TCPCB_SACKED_RETRANS) { |
1218 | /* If the segment is not tagged as lost, | |
1219 | * we do not clear RETRANS, believing | |
1220 | * that retransmission is still in flight. | |
1221 | */ | |
1222 | if (sacked & TCPCB_LOST) { | |
a1197f5a | 1223 | sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); |
f58b22fd IJ |
1224 | tp->lost_out -= pcount; |
1225 | tp->retrans_out -= pcount; | |
9e10c47c IJ |
1226 | } |
1227 | } else { | |
1228 | if (!(sacked & TCPCB_RETRANS)) { | |
1229 | /* New sack for not retransmitted frame, | |
1230 | * which was in hole. It is reordering. | |
1231 | */ | |
cc9a672e | 1232 | if (before(start_seq, |
9e10c47c | 1233 | tcp_highest_sack_seq(tp))) |
a1197f5a IJ |
1234 | state->reord = min(fack_count, |
1235 | state->reord); | |
e33099f9 YC |
1236 | if (!after(end_seq, tp->high_seq)) |
1237 | state->flag |= FLAG_ORIG_SACK_ACKED; | |
9a568de4 ED |
1238 | if (state->first_sackt == 0) |
1239 | state->first_sackt = xmit_time; | |
1240 | state->last_sackt = xmit_time; | |
9e10c47c IJ |
1241 | } |
1242 | ||
1243 | if (sacked & TCPCB_LOST) { | |
a1197f5a | 1244 | sacked &= ~TCPCB_LOST; |
f58b22fd | 1245 | tp->lost_out -= pcount; |
9e10c47c IJ |
1246 | } |
1247 | } | |
1248 | ||
a1197f5a IJ |
1249 | sacked |= TCPCB_SACKED_ACKED; |
1250 | state->flag |= FLAG_DATA_SACKED; | |
f58b22fd | 1251 | tp->sacked_out += pcount; |
ddf1af6f | 1252 | tp->delivered += pcount; /* Out-of-order packets delivered */ |
9e10c47c | 1253 | |
f58b22fd | 1254 | fack_count += pcount; |
9e10c47c IJ |
1255 | |
1256 | /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ | |
00db4124 | 1257 | if (!tcp_is_fack(tp) && tp->lost_skb_hint && |
cc9a672e | 1258 | before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq)) |
f58b22fd | 1259 | tp->lost_cnt_hint += pcount; |
9e10c47c IJ |
1260 | |
1261 | if (fack_count > tp->fackets_out) | |
1262 | tp->fackets_out = fack_count; | |
9e10c47c IJ |
1263 | } |
1264 | ||
1265 | /* D-SACK. We can detect redundant retransmission in S|R and plain R | |
1266 | * frames and clear it. undo_retrans is decreased above, L|R frames | |
1267 | * are accounted above as well. | |
1268 | */ | |
a1197f5a IJ |
1269 | if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { |
1270 | sacked &= ~TCPCB_SACKED_RETRANS; | |
f58b22fd | 1271 | tp->retrans_out -= pcount; |
9e10c47c IJ |
1272 | } |
1273 | ||
a1197f5a | 1274 | return sacked; |
9e10c47c IJ |
1275 | } |
1276 | ||
daef52ba NC |
1277 | /* Shift newly-SACKed bytes from this skb to the immediately previous |
1278 | * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. | |
1279 | */ | |
f3319816 ED |
1280 | static bool tcp_shifted_skb(struct sock *sk, struct sk_buff *prev, |
1281 | struct sk_buff *skb, | |
a2a385d6 ED |
1282 | struct tcp_sacktag_state *state, |
1283 | unsigned int pcount, int shifted, int mss, | |
1284 | bool dup_sack) | |
832d11c5 IJ |
1285 | { |
1286 | struct tcp_sock *tp = tcp_sk(sk); | |
daef52ba NC |
1287 | u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ |
1288 | u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ | |
832d11c5 IJ |
1289 | |
1290 | BUG_ON(!pcount); | |
1291 | ||
4c90d3b3 NC |
1292 | /* Adjust counters and hints for the newly sacked sequence |
1293 | * range but discard the return value since prev is already | |
1294 | * marked. We must tag the range first because the seq | |
1295 | * advancement below implicitly advances | |
1296 | * tcp_highest_sack_seq() when skb is highest_sack. | |
1297 | */ | |
1298 | tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, | |
59c9af42 | 1299 | start_seq, end_seq, dup_sack, pcount, |
9a568de4 | 1300 | skb->skb_mstamp); |
b9f64820 | 1301 | tcp_rate_skb_delivered(sk, skb, state->rate); |
4c90d3b3 NC |
1302 | |
1303 | if (skb == tp->lost_skb_hint) | |
0af2a0d0 NC |
1304 | tp->lost_cnt_hint += pcount; |
1305 | ||
832d11c5 IJ |
1306 | TCP_SKB_CB(prev)->end_seq += shifted; |
1307 | TCP_SKB_CB(skb)->seq += shifted; | |
1308 | ||
cd7d8498 ED |
1309 | tcp_skb_pcount_add(prev, pcount); |
1310 | BUG_ON(tcp_skb_pcount(skb) < pcount); | |
1311 | tcp_skb_pcount_add(skb, -pcount); | |
832d11c5 IJ |
1312 | |
1313 | /* When we're adding to gso_segs == 1, gso_size will be zero, | |
1314 | * in theory this shouldn't be necessary but as long as DSACK | |
1315 | * code can come after this skb later on it's better to keep | |
1316 | * setting gso_size to something. | |
1317 | */ | |
f69ad292 ED |
1318 | if (!TCP_SKB_CB(prev)->tcp_gso_size) |
1319 | TCP_SKB_CB(prev)->tcp_gso_size = mss; | |
832d11c5 IJ |
1320 | |
1321 | /* CHECKME: To clear or not to clear? Mimics normal skb currently */ | |
51466a75 | 1322 | if (tcp_skb_pcount(skb) <= 1) |
f69ad292 | 1323 | TCP_SKB_CB(skb)->tcp_gso_size = 0; |
832d11c5 | 1324 | |
832d11c5 IJ |
1325 | /* Difference in this won't matter, both ACKed by the same cumul. ACK */ |
1326 | TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); | |
1327 | ||
832d11c5 IJ |
1328 | if (skb->len > 0) { |
1329 | BUG_ON(!tcp_skb_pcount(skb)); | |
c10d9310 | 1330 | NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTED); |
a2a385d6 | 1331 | return false; |
832d11c5 IJ |
1332 | } |
1333 | ||
1334 | /* Whole SKB was eaten :-) */ | |
1335 | ||
92ee76b6 IJ |
1336 | if (skb == tp->retransmit_skb_hint) |
1337 | tp->retransmit_skb_hint = prev; | |
92ee76b6 IJ |
1338 | if (skb == tp->lost_skb_hint) { |
1339 | tp->lost_skb_hint = prev; | |
1340 | tp->lost_cnt_hint -= tcp_skb_pcount(prev); | |
1341 | } | |
1342 | ||
5e8a402f | 1343 | TCP_SKB_CB(prev)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; |
a643b5d4 | 1344 | TCP_SKB_CB(prev)->eor = TCP_SKB_CB(skb)->eor; |
5e8a402f ED |
1345 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) |
1346 | TCP_SKB_CB(prev)->end_seq++; | |
1347 | ||
832d11c5 IJ |
1348 | if (skb == tcp_highest_sack(sk)) |
1349 | tcp_advance_highest_sack(sk, skb); | |
1350 | ||
cfea5a68 | 1351 | tcp_skb_collapse_tstamp(prev, skb); |
9a568de4 ED |
1352 | if (unlikely(TCP_SKB_CB(prev)->tx.delivered_mstamp)) |
1353 | TCP_SKB_CB(prev)->tx.delivered_mstamp = 0; | |
b9f64820 | 1354 | |
75c119af | 1355 | tcp_rtx_queue_unlink_and_free(skb, sk); |
832d11c5 | 1356 | |
c10d9310 | 1357 | NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKMERGED); |
111cc8b9 | 1358 | |
a2a385d6 | 1359 | return true; |
832d11c5 IJ |
1360 | } |
1361 | ||
1362 | /* I wish gso_size would have a bit more sane initialization than | |
1363 | * something-or-zero which complicates things | |
1364 | */ | |
cf533ea5 | 1365 | static int tcp_skb_seglen(const struct sk_buff *skb) |
832d11c5 | 1366 | { |
775ffabf | 1367 | return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); |
832d11c5 IJ |
1368 | } |
1369 | ||
1370 | /* Shifting pages past head area doesn't work */ | |
cf533ea5 | 1371 | static int skb_can_shift(const struct sk_buff *skb) |
832d11c5 IJ |
1372 | { |
1373 | return !skb_headlen(skb) && skb_is_nonlinear(skb); | |
1374 | } | |
1375 | ||
1376 | /* Try collapsing SACK blocks spanning across multiple skbs to a single | |
1377 | * skb. | |
1378 | */ | |
1379 | static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, | |
a1197f5a | 1380 | struct tcp_sacktag_state *state, |
832d11c5 | 1381 | u32 start_seq, u32 end_seq, |
a2a385d6 | 1382 | bool dup_sack) |
832d11c5 IJ |
1383 | { |
1384 | struct tcp_sock *tp = tcp_sk(sk); | |
1385 | struct sk_buff *prev; | |
1386 | int mss; | |
1387 | int pcount = 0; | |
1388 | int len; | |
1389 | int in_sack; | |
1390 | ||
1391 | if (!sk_can_gso(sk)) | |
1392 | goto fallback; | |
1393 | ||
1394 | /* Normally R but no L won't result in plain S */ | |
1395 | if (!dup_sack && | |
9969ca5f | 1396 | (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) |
832d11c5 IJ |
1397 | goto fallback; |
1398 | if (!skb_can_shift(skb)) | |
1399 | goto fallback; | |
1400 | /* This frame is about to be dropped (was ACKed). */ | |
1401 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1402 | goto fallback; | |
1403 | ||
1404 | /* Can only happen with delayed DSACK + discard craziness */ | |
75c119af ED |
1405 | prev = skb_rb_prev(skb); |
1406 | if (!prev) | |
832d11c5 | 1407 | goto fallback; |
832d11c5 IJ |
1408 | |
1409 | if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) | |
1410 | goto fallback; | |
1411 | ||
a643b5d4 MKL |
1412 | if (!tcp_skb_can_collapse_to(prev)) |
1413 | goto fallback; | |
1414 | ||
832d11c5 IJ |
1415 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && |
1416 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1417 | ||
1418 | if (in_sack) { | |
1419 | len = skb->len; | |
1420 | pcount = tcp_skb_pcount(skb); | |
775ffabf | 1421 | mss = tcp_skb_seglen(skb); |
832d11c5 IJ |
1422 | |
1423 | /* TODO: Fix DSACKs to not fragment already SACKed and we can | |
1424 | * drop this restriction as unnecessary | |
1425 | */ | |
775ffabf | 1426 | if (mss != tcp_skb_seglen(prev)) |
832d11c5 IJ |
1427 | goto fallback; |
1428 | } else { | |
1429 | if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) | |
1430 | goto noop; | |
1431 | /* CHECKME: This is non-MSS split case only?, this will | |
1432 | * cause skipped skbs due to advancing loop btw, original | |
1433 | * has that feature too | |
1434 | */ | |
1435 | if (tcp_skb_pcount(skb) <= 1) | |
1436 | goto noop; | |
1437 | ||
1438 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); | |
1439 | if (!in_sack) { | |
1440 | /* TODO: head merge to next could be attempted here | |
1441 | * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), | |
1442 | * though it might not be worth of the additional hassle | |
1443 | * | |
1444 | * ...we can probably just fallback to what was done | |
1445 | * previously. We could try merging non-SACKed ones | |
1446 | * as well but it probably isn't going to buy off | |
1447 | * because later SACKs might again split them, and | |
1448 | * it would make skb timestamp tracking considerably | |
1449 | * harder problem. | |
1450 | */ | |
1451 | goto fallback; | |
1452 | } | |
1453 | ||
1454 | len = end_seq - TCP_SKB_CB(skb)->seq; | |
1455 | BUG_ON(len < 0); | |
1456 | BUG_ON(len > skb->len); | |
1457 | ||
1458 | /* MSS boundaries should be honoured or else pcount will | |
1459 | * severely break even though it makes things bit trickier. | |
1460 | * Optimize common case to avoid most of the divides | |
1461 | */ | |
1462 | mss = tcp_skb_mss(skb); | |
1463 | ||
1464 | /* TODO: Fix DSACKs to not fragment already SACKed and we can | |
1465 | * drop this restriction as unnecessary | |
1466 | */ | |
775ffabf | 1467 | if (mss != tcp_skb_seglen(prev)) |
832d11c5 IJ |
1468 | goto fallback; |
1469 | ||
1470 | if (len == mss) { | |
1471 | pcount = 1; | |
1472 | } else if (len < mss) { | |
1473 | goto noop; | |
1474 | } else { | |
1475 | pcount = len / mss; | |
1476 | len = pcount * mss; | |
1477 | } | |
1478 | } | |
1479 | ||
4648dc97 NC |
1480 | /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */ |
1481 | if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una)) | |
1482 | goto fallback; | |
1483 | ||
832d11c5 IJ |
1484 | if (!skb_shift(prev, skb, len)) |
1485 | goto fallback; | |
f3319816 | 1486 | if (!tcp_shifted_skb(sk, prev, skb, state, pcount, len, mss, dup_sack)) |
832d11c5 IJ |
1487 | goto out; |
1488 | ||
1489 | /* Hole filled allows collapsing with the next as well, this is very | |
1490 | * useful when hole on every nth skb pattern happens | |
1491 | */ | |
75c119af ED |
1492 | skb = skb_rb_next(prev); |
1493 | if (!skb) | |
832d11c5 | 1494 | goto out; |
832d11c5 | 1495 | |
f0bc52f3 | 1496 | if (!skb_can_shift(skb) || |
f0bc52f3 | 1497 | ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || |
775ffabf | 1498 | (mss != tcp_skb_seglen(skb))) |
832d11c5 IJ |
1499 | goto out; |
1500 | ||
1501 | len = skb->len; | |
1502 | if (skb_shift(prev, skb, len)) { | |
1503 | pcount += tcp_skb_pcount(skb); | |
f3319816 ED |
1504 | tcp_shifted_skb(sk, prev, skb, state, tcp_skb_pcount(skb), |
1505 | len, mss, 0); | |
832d11c5 IJ |
1506 | } |
1507 | ||
1508 | out: | |
a1197f5a | 1509 | state->fack_count += pcount; |
832d11c5 IJ |
1510 | return prev; |
1511 | ||
1512 | noop: | |
1513 | return skb; | |
1514 | ||
1515 | fallback: | |
c10d9310 | 1516 | NET_INC_STATS(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); |
832d11c5 IJ |
1517 | return NULL; |
1518 | } | |
1519 | ||
68f8353b IJ |
1520 | static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, |
1521 | struct tcp_sack_block *next_dup, | |
a1197f5a | 1522 | struct tcp_sacktag_state *state, |
68f8353b | 1523 | u32 start_seq, u32 end_seq, |
a2a385d6 | 1524 | bool dup_sack_in) |
68f8353b | 1525 | { |
832d11c5 IJ |
1526 | struct tcp_sock *tp = tcp_sk(sk); |
1527 | struct sk_buff *tmp; | |
1528 | ||
75c119af | 1529 | skb_rbtree_walk_from(skb) { |
68f8353b | 1530 | int in_sack = 0; |
a2a385d6 | 1531 | bool dup_sack = dup_sack_in; |
68f8353b | 1532 | |
68f8353b IJ |
1533 | /* queue is in-order => we can short-circuit the walk early */ |
1534 | if (!before(TCP_SKB_CB(skb)->seq, end_seq)) | |
1535 | break; | |
1536 | ||
00db4124 | 1537 | if (next_dup && |
68f8353b IJ |
1538 | before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { |
1539 | in_sack = tcp_match_skb_to_sack(sk, skb, | |
1540 | next_dup->start_seq, | |
1541 | next_dup->end_seq); | |
1542 | if (in_sack > 0) | |
a2a385d6 | 1543 | dup_sack = true; |
68f8353b IJ |
1544 | } |
1545 | ||
832d11c5 IJ |
1546 | /* skb reference here is a bit tricky to get right, since |
1547 | * shifting can eat and free both this skb and the next, | |
1548 | * so not even _safe variant of the loop is enough. | |
1549 | */ | |
1550 | if (in_sack <= 0) { | |
a1197f5a IJ |
1551 | tmp = tcp_shift_skb_data(sk, skb, state, |
1552 | start_seq, end_seq, dup_sack); | |
00db4124 | 1553 | if (tmp) { |
832d11c5 IJ |
1554 | if (tmp != skb) { |
1555 | skb = tmp; | |
1556 | continue; | |
1557 | } | |
1558 | ||
1559 | in_sack = 0; | |
1560 | } else { | |
1561 | in_sack = tcp_match_skb_to_sack(sk, skb, | |
1562 | start_seq, | |
1563 | end_seq); | |
1564 | } | |
1565 | } | |
1566 | ||
68f8353b IJ |
1567 | if (unlikely(in_sack < 0)) |
1568 | break; | |
1569 | ||
832d11c5 | 1570 | if (in_sack) { |
cc9a672e NC |
1571 | TCP_SKB_CB(skb)->sacked = |
1572 | tcp_sacktag_one(sk, | |
1573 | state, | |
1574 | TCP_SKB_CB(skb)->sacked, | |
1575 | TCP_SKB_CB(skb)->seq, | |
1576 | TCP_SKB_CB(skb)->end_seq, | |
1577 | dup_sack, | |
59c9af42 | 1578 | tcp_skb_pcount(skb), |
9a568de4 | 1579 | skb->skb_mstamp); |
b9f64820 | 1580 | tcp_rate_skb_delivered(sk, skb, state->rate); |
e2080072 ED |
1581 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) |
1582 | list_del_init(&skb->tcp_tsorted_anchor); | |
68f8353b | 1583 | |
832d11c5 IJ |
1584 | if (!before(TCP_SKB_CB(skb)->seq, |
1585 | tcp_highest_sack_seq(tp))) | |
1586 | tcp_advance_highest_sack(sk, skb); | |
1587 | } | |
1588 | ||
a1197f5a | 1589 | state->fack_count += tcp_skb_pcount(skb); |
68f8353b IJ |
1590 | } |
1591 | return skb; | |
1592 | } | |
1593 | ||
75c119af ED |
1594 | static struct sk_buff *tcp_sacktag_bsearch(struct sock *sk, |
1595 | struct tcp_sacktag_state *state, | |
1596 | u32 seq) | |
1597 | { | |
1598 | struct rb_node *parent, **p = &sk->tcp_rtx_queue.rb_node; | |
1599 | struct sk_buff *skb; | |
1600 | int unack_bytes; | |
1601 | ||
1602 | while (*p) { | |
1603 | parent = *p; | |
1604 | skb = rb_to_skb(parent); | |
1605 | if (before(seq, TCP_SKB_CB(skb)->seq)) { | |
1606 | p = &parent->rb_left; | |
1607 | continue; | |
1608 | } | |
1609 | if (!before(seq, TCP_SKB_CB(skb)->end_seq)) { | |
1610 | p = &parent->rb_right; | |
1611 | continue; | |
1612 | } | |
1613 | ||
1614 | state->fack_count = 0; | |
1615 | unack_bytes = TCP_SKB_CB(skb)->seq - tcp_sk(sk)->snd_una; | |
1616 | if (state->mss_now && unack_bytes > 0) | |
1617 | state->fack_count = unack_bytes / state->mss_now; | |
1618 | ||
1619 | return skb; | |
1620 | } | |
1621 | return NULL; | |
1622 | } | |
1623 | ||
68f8353b | 1624 | static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, |
a1197f5a IJ |
1625 | struct tcp_sacktag_state *state, |
1626 | u32 skip_to_seq) | |
68f8353b | 1627 | { |
75c119af ED |
1628 | if (skb && after(TCP_SKB_CB(skb)->seq, skip_to_seq)) |
1629 | return skb; | |
d152a7d8 | 1630 | |
75c119af | 1631 | return tcp_sacktag_bsearch(sk, state, skip_to_seq); |
68f8353b IJ |
1632 | } |
1633 | ||
1634 | static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, | |
1635 | struct sock *sk, | |
1636 | struct tcp_sack_block *next_dup, | |
a1197f5a IJ |
1637 | struct tcp_sacktag_state *state, |
1638 | u32 skip_to_seq) | |
68f8353b | 1639 | { |
51456b29 | 1640 | if (!next_dup) |
68f8353b IJ |
1641 | return skb; |
1642 | ||
1643 | if (before(next_dup->start_seq, skip_to_seq)) { | |
a1197f5a IJ |
1644 | skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq); |
1645 | skb = tcp_sacktag_walk(skb, sk, NULL, state, | |
1646 | next_dup->start_seq, next_dup->end_seq, | |
1647 | 1); | |
68f8353b IJ |
1648 | } |
1649 | ||
1650 | return skb; | |
1651 | } | |
1652 | ||
cf533ea5 | 1653 | static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) |
68f8353b IJ |
1654 | { |
1655 | return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); | |
1656 | } | |
1657 | ||
1da177e4 | 1658 | static int |
cf533ea5 | 1659 | tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, |
196da974 | 1660 | u32 prior_snd_una, struct tcp_sacktag_state *state) |
1da177e4 LT |
1661 | { |
1662 | struct tcp_sock *tp = tcp_sk(sk); | |
cf533ea5 ED |
1663 | const unsigned char *ptr = (skb_transport_header(ack_skb) + |
1664 | TCP_SKB_CB(ack_skb)->sacked); | |
fd6dad61 | 1665 | struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); |
4389dded | 1666 | struct tcp_sack_block sp[TCP_NUM_SACKS]; |
68f8353b IJ |
1667 | struct tcp_sack_block *cache; |
1668 | struct sk_buff *skb; | |
4389dded | 1669 | int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); |
fd6dad61 | 1670 | int used_sacks; |
a2a385d6 | 1671 | bool found_dup_sack = false; |
68f8353b | 1672 | int i, j; |
fda03fbb | 1673 | int first_sack_index; |
1da177e4 | 1674 | |
196da974 KKJ |
1675 | state->flag = 0; |
1676 | state->reord = tp->packets_out; | |
a1197f5a | 1677 | |
d738cd8f | 1678 | if (!tp->sacked_out) { |
de83c058 IJ |
1679 | if (WARN_ON(tp->fackets_out)) |
1680 | tp->fackets_out = 0; | |
6859d494 | 1681 | tcp_highest_sack_reset(sk); |
d738cd8f | 1682 | } |
1da177e4 | 1683 | |
1ed83465 | 1684 | found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, |
d06e021d | 1685 | num_sacks, prior_snd_una); |
b9f64820 | 1686 | if (found_dup_sack) { |
196da974 | 1687 | state->flag |= FLAG_DSACKING_ACK; |
b9f64820 YC |
1688 | tp->delivered++; /* A spurious retransmission is delivered */ |
1689 | } | |
6f74651a BE |
1690 | |
1691 | /* Eliminate too old ACKs, but take into | |
1692 | * account more or less fresh ones, they can | |
1693 | * contain valid SACK info. | |
1694 | */ | |
1695 | if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) | |
1696 | return 0; | |
1697 | ||