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 | * | |
8 | * Version: $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $ | |
9 | * | |
02c30a84 | 10 | * Authors: Ross Biro |
1da177e4 LT |
11 | * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> |
12 | * Mark Evans, <evansmp@uhura.aston.ac.uk> | |
13 | * Corey Minyard <wf-rch!minyard@relay.EU.net> | |
14 | * Florian La Roche, <flla@stud.uni-sb.de> | |
15 | * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> | |
16 | * Linus Torvalds, <torvalds@cs.helsinki.fi> | |
17 | * Alan Cox, <gw4pts@gw4pts.ampr.org> | |
18 | * Matthew Dillon, <dillon@apollo.west.oic.com> | |
19 | * Arnt Gulbrandsen, <agulbra@nvg.unit.no> | |
20 | * Jorge Cwik, <jorge@laser.satlink.net> | |
21 | */ | |
22 | ||
23 | /* | |
24 | * Changes: | |
25 | * Pedro Roque : Fast Retransmit/Recovery. | |
26 | * Two receive queues. | |
27 | * Retransmit queue handled by TCP. | |
28 | * Better retransmit timer handling. | |
29 | * New congestion avoidance. | |
30 | * Header prediction. | |
31 | * Variable renaming. | |
32 | * | |
33 | * Eric : Fast Retransmit. | |
34 | * Randy Scott : MSS option defines. | |
35 | * Eric Schenk : Fixes to slow start algorithm. | |
36 | * Eric Schenk : Yet another double ACK bug. | |
37 | * Eric Schenk : Delayed ACK bug fixes. | |
38 | * Eric Schenk : Floyd style fast retrans war avoidance. | |
39 | * David S. Miller : Don't allow zero congestion window. | |
40 | * Eric Schenk : Fix retransmitter so that it sends | |
41 | * next packet on ack of previous packet. | |
42 | * Andi Kleen : Moved open_request checking here | |
43 | * and process RSTs for open_requests. | |
44 | * Andi Kleen : Better prune_queue, and other fixes. | |
caa20d9a | 45 | * Andrey Savochkin: Fix RTT measurements in the presence of |
1da177e4 LT |
46 | * timestamps. |
47 | * Andrey Savochkin: Check sequence numbers correctly when | |
48 | * removing SACKs due to in sequence incoming | |
49 | * data segments. | |
50 | * Andi Kleen: Make sure we never ack data there is not | |
51 | * enough room for. Also make this condition | |
52 | * a fatal error if it might still happen. | |
e905a9ed | 53 | * Andi Kleen: Add tcp_measure_rcv_mss to make |
1da177e4 | 54 | * connections with MSS<min(MTU,ann. MSS) |
e905a9ed | 55 | * work without delayed acks. |
1da177e4 LT |
56 | * Andi Kleen: Process packets with PSH set in the |
57 | * fast path. | |
58 | * J Hadi Salim: ECN support | |
59 | * Andrei Gurtov, | |
60 | * Pasi Sarolahti, | |
61 | * Panu Kuhlberg: Experimental audit of TCP (re)transmission | |
62 | * engine. Lots of bugs are found. | |
63 | * Pasi Sarolahti: F-RTO for dealing with spurious RTOs | |
1da177e4 LT |
64 | */ |
65 | ||
1da177e4 LT |
66 | #include <linux/mm.h> |
67 | #include <linux/module.h> | |
68 | #include <linux/sysctl.h> | |
69 | #include <net/tcp.h> | |
70 | #include <net/inet_common.h> | |
71 | #include <linux/ipsec.h> | |
72 | #include <asm/unaligned.h> | |
1a2449a8 | 73 | #include <net/netdma.h> |
1da177e4 | 74 | |
ab32ea5d BH |
75 | int sysctl_tcp_timestamps __read_mostly = 1; |
76 | int sysctl_tcp_window_scaling __read_mostly = 1; | |
77 | int sysctl_tcp_sack __read_mostly = 1; | |
78 | int sysctl_tcp_fack __read_mostly = 1; | |
79 | int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH; | |
80 | int sysctl_tcp_ecn __read_mostly; | |
81 | int sysctl_tcp_dsack __read_mostly = 1; | |
82 | int sysctl_tcp_app_win __read_mostly = 31; | |
83 | int sysctl_tcp_adv_win_scale __read_mostly = 2; | |
1da177e4 | 84 | |
ab32ea5d BH |
85 | int sysctl_tcp_stdurg __read_mostly; |
86 | int sysctl_tcp_rfc1337 __read_mostly; | |
87 | int sysctl_tcp_max_orphans __read_mostly = NR_FILE; | |
c96fd3d4 | 88 | int sysctl_tcp_frto __read_mostly = 2; |
3cfe3baa | 89 | int sysctl_tcp_frto_response __read_mostly; |
ab32ea5d | 90 | int sysctl_tcp_nometrics_save __read_mostly; |
1da177e4 | 91 | |
ab32ea5d BH |
92 | int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; |
93 | int sysctl_tcp_abc __read_mostly; | |
1da177e4 | 94 | |
1da177e4 LT |
95 | #define FLAG_DATA 0x01 /* Incoming frame contained data. */ |
96 | #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ | |
97 | #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ | |
98 | #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ | |
99 | #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ | |
100 | #define FLAG_DATA_SACKED 0x20 /* New SACK. */ | |
101 | #define FLAG_ECE 0x40 /* ECE in this ACK */ | |
102 | #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */ | |
103 | #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ | |
4dc2665e | 104 | #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */ |
2e605294 | 105 | #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ |
564262c1 | 106 | #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ |
009a2e3e | 107 | #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */ |
1da177e4 LT |
108 | |
109 | #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) | |
110 | #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) | |
111 | #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) | |
112 | #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) | |
2e605294 | 113 | #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED) |
1da177e4 | 114 | |
4dc2665e IJ |
115 | #define IsSackFrto() (sysctl_tcp_frto == 0x2) |
116 | ||
1da177e4 | 117 | #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) |
bdf1ee5d | 118 | #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) |
1da177e4 | 119 | |
e905a9ed | 120 | /* Adapt the MSS value used to make delayed ack decision to the |
1da177e4 | 121 | * real world. |
e905a9ed | 122 | */ |
40efc6fa SH |
123 | static void tcp_measure_rcv_mss(struct sock *sk, |
124 | const struct sk_buff *skb) | |
1da177e4 | 125 | { |
463c84b9 | 126 | struct inet_connection_sock *icsk = inet_csk(sk); |
e905a9ed | 127 | const unsigned int lss = icsk->icsk_ack.last_seg_size; |
463c84b9 | 128 | unsigned int len; |
1da177e4 | 129 | |
e905a9ed | 130 | icsk->icsk_ack.last_seg_size = 0; |
1da177e4 LT |
131 | |
132 | /* skb->len may jitter because of SACKs, even if peer | |
133 | * sends good full-sized frames. | |
134 | */ | |
ff9b5e0f | 135 | len = skb_shinfo(skb)->gso_size ?: skb->len; |
463c84b9 ACM |
136 | if (len >= icsk->icsk_ack.rcv_mss) { |
137 | icsk->icsk_ack.rcv_mss = len; | |
1da177e4 LT |
138 | } else { |
139 | /* Otherwise, we make more careful check taking into account, | |
140 | * that SACKs block is variable. | |
141 | * | |
142 | * "len" is invariant segment length, including TCP header. | |
143 | */ | |
9c70220b | 144 | len += skb->data - skb_transport_header(skb); |
1da177e4 LT |
145 | if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) || |
146 | /* If PSH is not set, packet should be | |
147 | * full sized, provided peer TCP is not badly broken. | |
148 | * This observation (if it is correct 8)) allows | |
149 | * to handle super-low mtu links fairly. | |
150 | */ | |
151 | (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && | |
aa8223c7 | 152 | !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { |
1da177e4 LT |
153 | /* Subtract also invariant (if peer is RFC compliant), |
154 | * tcp header plus fixed timestamp option length. | |
155 | * Resulting "len" is MSS free of SACK jitter. | |
156 | */ | |
463c84b9 ACM |
157 | len -= tcp_sk(sk)->tcp_header_len; |
158 | icsk->icsk_ack.last_seg_size = len; | |
1da177e4 | 159 | if (len == lss) { |
463c84b9 | 160 | icsk->icsk_ack.rcv_mss = len; |
1da177e4 LT |
161 | return; |
162 | } | |
163 | } | |
1ef9696c AK |
164 | if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) |
165 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; | |
463c84b9 | 166 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; |
1da177e4 LT |
167 | } |
168 | } | |
169 | ||
463c84b9 | 170 | static void tcp_incr_quickack(struct sock *sk) |
1da177e4 | 171 | { |
463c84b9 ACM |
172 | struct inet_connection_sock *icsk = inet_csk(sk); |
173 | unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); | |
1da177e4 LT |
174 | |
175 | if (quickacks==0) | |
176 | quickacks=2; | |
463c84b9 ACM |
177 | if (quickacks > icsk->icsk_ack.quick) |
178 | icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); | |
1da177e4 LT |
179 | } |
180 | ||
463c84b9 | 181 | void tcp_enter_quickack_mode(struct sock *sk) |
1da177e4 | 182 | { |
463c84b9 ACM |
183 | struct inet_connection_sock *icsk = inet_csk(sk); |
184 | tcp_incr_quickack(sk); | |
185 | icsk->icsk_ack.pingpong = 0; | |
186 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 LT |
187 | } |
188 | ||
189 | /* Send ACKs quickly, if "quick" count is not exhausted | |
190 | * and the session is not interactive. | |
191 | */ | |
192 | ||
463c84b9 | 193 | static inline int tcp_in_quickack_mode(const struct sock *sk) |
1da177e4 | 194 | { |
463c84b9 ACM |
195 | const struct inet_connection_sock *icsk = inet_csk(sk); |
196 | return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; | |
1da177e4 LT |
197 | } |
198 | ||
bdf1ee5d IJ |
199 | static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp) |
200 | { | |
201 | if (tp->ecn_flags&TCP_ECN_OK) | |
202 | tp->ecn_flags |= TCP_ECN_QUEUE_CWR; | |
203 | } | |
204 | ||
205 | static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb) | |
206 | { | |
207 | if (tcp_hdr(skb)->cwr) | |
208 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
209 | } | |
210 | ||
211 | static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp) | |
212 | { | |
213 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
214 | } | |
215 | ||
216 | static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb) | |
217 | { | |
218 | if (tp->ecn_flags&TCP_ECN_OK) { | |
219 | if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags)) | |
220 | tp->ecn_flags |= TCP_ECN_DEMAND_CWR; | |
221 | /* Funny extension: if ECT is not set on a segment, | |
222 | * it is surely retransmit. It is not in ECN RFC, | |
223 | * but Linux follows this rule. */ | |
224 | else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags))) | |
225 | tcp_enter_quickack_mode((struct sock *)tp); | |
226 | } | |
227 | } | |
228 | ||
229 | static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th) | |
230 | { | |
231 | if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || th->cwr)) | |
232 | tp->ecn_flags &= ~TCP_ECN_OK; | |
233 | } | |
234 | ||
235 | static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th) | |
236 | { | |
237 | if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || !th->cwr)) | |
238 | tp->ecn_flags &= ~TCP_ECN_OK; | |
239 | } | |
240 | ||
241 | static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th) | |
242 | { | |
243 | if (th->ece && !th->syn && (tp->ecn_flags&TCP_ECN_OK)) | |
244 | return 1; | |
245 | return 0; | |
246 | } | |
247 | ||
1da177e4 LT |
248 | /* Buffer size and advertised window tuning. |
249 | * | |
250 | * 1. Tuning sk->sk_sndbuf, when connection enters established state. | |
251 | */ | |
252 | ||
253 | static void tcp_fixup_sndbuf(struct sock *sk) | |
254 | { | |
255 | int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 + | |
256 | sizeof(struct sk_buff); | |
257 | ||
258 | if (sk->sk_sndbuf < 3 * sndmem) | |
259 | sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]); | |
260 | } | |
261 | ||
262 | /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) | |
263 | * | |
264 | * All tcp_full_space() is split to two parts: "network" buffer, allocated | |
265 | * forward and advertised in receiver window (tp->rcv_wnd) and | |
266 | * "application buffer", required to isolate scheduling/application | |
267 | * latencies from network. | |
268 | * window_clamp is maximal advertised window. It can be less than | |
269 | * tcp_full_space(), in this case tcp_full_space() - window_clamp | |
270 | * is reserved for "application" buffer. The less window_clamp is | |
271 | * the smoother our behaviour from viewpoint of network, but the lower | |
272 | * throughput and the higher sensitivity of the connection to losses. 8) | |
273 | * | |
274 | * rcv_ssthresh is more strict window_clamp used at "slow start" | |
275 | * phase to predict further behaviour of this connection. | |
276 | * It is used for two goals: | |
277 | * - to enforce header prediction at sender, even when application | |
278 | * requires some significant "application buffer". It is check #1. | |
279 | * - to prevent pruning of receive queue because of misprediction | |
280 | * of receiver window. Check #2. | |
281 | * | |
282 | * The scheme does not work when sender sends good segments opening | |
caa20d9a | 283 | * window and then starts to feed us spaghetti. But it should work |
1da177e4 LT |
284 | * in common situations. Otherwise, we have to rely on queue collapsing. |
285 | */ | |
286 | ||
287 | /* Slow part of check#2. */ | |
9e412ba7 | 288 | static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 289 | { |
9e412ba7 | 290 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
291 | /* Optimize this! */ |
292 | int truesize = tcp_win_from_space(skb->truesize)/2; | |
326f36e9 | 293 | int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2; |
1da177e4 LT |
294 | |
295 | while (tp->rcv_ssthresh <= window) { | |
296 | if (truesize <= skb->len) | |
463c84b9 | 297 | return 2 * inet_csk(sk)->icsk_ack.rcv_mss; |
1da177e4 LT |
298 | |
299 | truesize >>= 1; | |
300 | window >>= 1; | |
301 | } | |
302 | return 0; | |
303 | } | |
304 | ||
9e412ba7 | 305 | static void tcp_grow_window(struct sock *sk, |
40efc6fa | 306 | struct sk_buff *skb) |
1da177e4 | 307 | { |
9e412ba7 IJ |
308 | struct tcp_sock *tp = tcp_sk(sk); |
309 | ||
1da177e4 LT |
310 | /* Check #1 */ |
311 | if (tp->rcv_ssthresh < tp->window_clamp && | |
312 | (int)tp->rcv_ssthresh < tcp_space(sk) && | |
313 | !tcp_memory_pressure) { | |
314 | int incr; | |
315 | ||
316 | /* Check #2. Increase window, if skb with such overhead | |
317 | * will fit to rcvbuf in future. | |
318 | */ | |
319 | if (tcp_win_from_space(skb->truesize) <= skb->len) | |
320 | incr = 2*tp->advmss; | |
321 | else | |
9e412ba7 | 322 | incr = __tcp_grow_window(sk, skb); |
1da177e4 LT |
323 | |
324 | if (incr) { | |
325 | tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp); | |
463c84b9 | 326 | inet_csk(sk)->icsk_ack.quick |= 1; |
1da177e4 LT |
327 | } |
328 | } | |
329 | } | |
330 | ||
331 | /* 3. Tuning rcvbuf, when connection enters established state. */ | |
332 | ||
333 | static void tcp_fixup_rcvbuf(struct sock *sk) | |
334 | { | |
335 | struct tcp_sock *tp = tcp_sk(sk); | |
336 | int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); | |
337 | ||
338 | /* Try to select rcvbuf so that 4 mss-sized segments | |
caa20d9a | 339 | * will fit to window and corresponding skbs will fit to our rcvbuf. |
1da177e4 LT |
340 | * (was 3; 4 is minimum to allow fast retransmit to work.) |
341 | */ | |
342 | while (tcp_win_from_space(rcvmem) < tp->advmss) | |
343 | rcvmem += 128; | |
344 | if (sk->sk_rcvbuf < 4 * rcvmem) | |
345 | sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]); | |
346 | } | |
347 | ||
caa20d9a | 348 | /* 4. Try to fixup all. It is made immediately after connection enters |
1da177e4 LT |
349 | * established state. |
350 | */ | |
351 | static void tcp_init_buffer_space(struct sock *sk) | |
352 | { | |
353 | struct tcp_sock *tp = tcp_sk(sk); | |
354 | int maxwin; | |
355 | ||
356 | if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) | |
357 | tcp_fixup_rcvbuf(sk); | |
358 | if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) | |
359 | tcp_fixup_sndbuf(sk); | |
360 | ||
361 | tp->rcvq_space.space = tp->rcv_wnd; | |
362 | ||
363 | maxwin = tcp_full_space(sk); | |
364 | ||
365 | if (tp->window_clamp >= maxwin) { | |
366 | tp->window_clamp = maxwin; | |
367 | ||
368 | if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) | |
369 | tp->window_clamp = max(maxwin - | |
370 | (maxwin >> sysctl_tcp_app_win), | |
371 | 4 * tp->advmss); | |
372 | } | |
373 | ||
374 | /* Force reservation of one segment. */ | |
375 | if (sysctl_tcp_app_win && | |
376 | tp->window_clamp > 2 * tp->advmss && | |
377 | tp->window_clamp + tp->advmss > maxwin) | |
378 | tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); | |
379 | ||
380 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); | |
381 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
382 | } | |
383 | ||
1da177e4 | 384 | /* 5. Recalculate window clamp after socket hit its memory bounds. */ |
9e412ba7 | 385 | static void tcp_clamp_window(struct sock *sk) |
1da177e4 | 386 | { |
9e412ba7 | 387 | struct tcp_sock *tp = tcp_sk(sk); |
6687e988 | 388 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 389 | |
6687e988 | 390 | icsk->icsk_ack.quick = 0; |
1da177e4 | 391 | |
326f36e9 JH |
392 | if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && |
393 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && | |
394 | !tcp_memory_pressure && | |
395 | atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) { | |
396 | sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), | |
397 | sysctl_tcp_rmem[2]); | |
1da177e4 | 398 | } |
326f36e9 | 399 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) |
1da177e4 | 400 | tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss); |
1da177e4 LT |
401 | } |
402 | ||
40efc6fa SH |
403 | |
404 | /* Initialize RCV_MSS value. | |
405 | * RCV_MSS is an our guess about MSS used by the peer. | |
406 | * We haven't any direct information about the MSS. | |
407 | * It's better to underestimate the RCV_MSS rather than overestimate. | |
408 | * Overestimations make us ACKing less frequently than needed. | |
409 | * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). | |
410 | */ | |
411 | void tcp_initialize_rcv_mss(struct sock *sk) | |
412 | { | |
413 | struct tcp_sock *tp = tcp_sk(sk); | |
414 | unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); | |
415 | ||
416 | hint = min(hint, tp->rcv_wnd/2); | |
417 | hint = min(hint, TCP_MIN_RCVMSS); | |
418 | hint = max(hint, TCP_MIN_MSS); | |
419 | ||
420 | inet_csk(sk)->icsk_ack.rcv_mss = hint; | |
421 | } | |
422 | ||
1da177e4 LT |
423 | /* Receiver "autotuning" code. |
424 | * | |
425 | * The algorithm for RTT estimation w/o timestamps is based on | |
426 | * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. | |
427 | * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps> | |
428 | * | |
429 | * More detail on this code can be found at | |
430 | * <http://www.psc.edu/~jheffner/senior_thesis.ps>, | |
431 | * though this reference is out of date. A new paper | |
432 | * is pending. | |
433 | */ | |
434 | static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) | |
435 | { | |
436 | u32 new_sample = tp->rcv_rtt_est.rtt; | |
437 | long m = sample; | |
438 | ||
439 | if (m == 0) | |
440 | m = 1; | |
441 | ||
442 | if (new_sample != 0) { | |
443 | /* If we sample in larger samples in the non-timestamp | |
444 | * case, we could grossly overestimate the RTT especially | |
445 | * with chatty applications or bulk transfer apps which | |
446 | * are stalled on filesystem I/O. | |
447 | * | |
448 | * Also, since we are only going for a minimum in the | |
31f34269 | 449 | * non-timestamp case, we do not smooth things out |
caa20d9a | 450 | * else with timestamps disabled convergence takes too |
1da177e4 LT |
451 | * long. |
452 | */ | |
453 | if (!win_dep) { | |
454 | m -= (new_sample >> 3); | |
455 | new_sample += m; | |
456 | } else if (m < new_sample) | |
457 | new_sample = m << 3; | |
458 | } else { | |
caa20d9a | 459 | /* No previous measure. */ |
1da177e4 LT |
460 | new_sample = m << 3; |
461 | } | |
462 | ||
463 | if (tp->rcv_rtt_est.rtt != new_sample) | |
464 | tp->rcv_rtt_est.rtt = new_sample; | |
465 | } | |
466 | ||
467 | static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) | |
468 | { | |
469 | if (tp->rcv_rtt_est.time == 0) | |
470 | goto new_measure; | |
471 | if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) | |
472 | return; | |
473 | tcp_rcv_rtt_update(tp, | |
474 | jiffies - tp->rcv_rtt_est.time, | |
475 | 1); | |
476 | ||
477 | new_measure: | |
478 | tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; | |
479 | tp->rcv_rtt_est.time = tcp_time_stamp; | |
480 | } | |
481 | ||
463c84b9 | 482 | static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 483 | { |
463c84b9 | 484 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
485 | if (tp->rx_opt.rcv_tsecr && |
486 | (TCP_SKB_CB(skb)->end_seq - | |
463c84b9 | 487 | TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) |
1da177e4 LT |
488 | tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); |
489 | } | |
490 | ||
491 | /* | |
492 | * This function should be called every time data is copied to user space. | |
493 | * It calculates the appropriate TCP receive buffer space. | |
494 | */ | |
495 | void tcp_rcv_space_adjust(struct sock *sk) | |
496 | { | |
497 | struct tcp_sock *tp = tcp_sk(sk); | |
498 | int time; | |
499 | int space; | |
e905a9ed | 500 | |
1da177e4 LT |
501 | if (tp->rcvq_space.time == 0) |
502 | goto new_measure; | |
e905a9ed | 503 | |
1da177e4 LT |
504 | time = tcp_time_stamp - tp->rcvq_space.time; |
505 | if (time < (tp->rcv_rtt_est.rtt >> 3) || | |
506 | tp->rcv_rtt_est.rtt == 0) | |
507 | return; | |
e905a9ed | 508 | |
1da177e4 LT |
509 | space = 2 * (tp->copied_seq - tp->rcvq_space.seq); |
510 | ||
511 | space = max(tp->rcvq_space.space, space); | |
512 | ||
513 | if (tp->rcvq_space.space != space) { | |
514 | int rcvmem; | |
515 | ||
516 | tp->rcvq_space.space = space; | |
517 | ||
6fcf9412 JH |
518 | if (sysctl_tcp_moderate_rcvbuf && |
519 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { | |
1da177e4 LT |
520 | int new_clamp = space; |
521 | ||
522 | /* Receive space grows, normalize in order to | |
523 | * take into account packet headers and sk_buff | |
524 | * structure overhead. | |
525 | */ | |
526 | space /= tp->advmss; | |
527 | if (!space) | |
528 | space = 1; | |
529 | rcvmem = (tp->advmss + MAX_TCP_HEADER + | |
530 | 16 + sizeof(struct sk_buff)); | |
531 | while (tcp_win_from_space(rcvmem) < tp->advmss) | |
532 | rcvmem += 128; | |
533 | space *= rcvmem; | |
534 | space = min(space, sysctl_tcp_rmem[2]); | |
535 | if (space > sk->sk_rcvbuf) { | |
536 | sk->sk_rcvbuf = space; | |
537 | ||
538 | /* Make the window clamp follow along. */ | |
539 | tp->window_clamp = new_clamp; | |
540 | } | |
541 | } | |
542 | } | |
e905a9ed | 543 | |
1da177e4 LT |
544 | new_measure: |
545 | tp->rcvq_space.seq = tp->copied_seq; | |
546 | tp->rcvq_space.time = tcp_time_stamp; | |
547 | } | |
548 | ||
549 | /* There is something which you must keep in mind when you analyze the | |
550 | * behavior of the tp->ato delayed ack timeout interval. When a | |
551 | * connection starts up, we want to ack as quickly as possible. The | |
552 | * problem is that "good" TCP's do slow start at the beginning of data | |
553 | * transmission. The means that until we send the first few ACK's the | |
554 | * sender will sit on his end and only queue most of his data, because | |
555 | * he can only send snd_cwnd unacked packets at any given time. For | |
556 | * each ACK we send, he increments snd_cwnd and transmits more of his | |
557 | * queue. -DaveM | |
558 | */ | |
9e412ba7 | 559 | static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 560 | { |
9e412ba7 | 561 | struct tcp_sock *tp = tcp_sk(sk); |
463c84b9 | 562 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
563 | u32 now; |
564 | ||
463c84b9 | 565 | inet_csk_schedule_ack(sk); |
1da177e4 | 566 | |
463c84b9 | 567 | tcp_measure_rcv_mss(sk, skb); |
1da177e4 LT |
568 | |
569 | tcp_rcv_rtt_measure(tp); | |
e905a9ed | 570 | |
1da177e4 LT |
571 | now = tcp_time_stamp; |
572 | ||
463c84b9 | 573 | if (!icsk->icsk_ack.ato) { |
1da177e4 LT |
574 | /* The _first_ data packet received, initialize |
575 | * delayed ACK engine. | |
576 | */ | |
463c84b9 ACM |
577 | tcp_incr_quickack(sk); |
578 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 | 579 | } else { |
463c84b9 | 580 | int m = now - icsk->icsk_ack.lrcvtime; |
1da177e4 LT |
581 | |
582 | if (m <= TCP_ATO_MIN/2) { | |
583 | /* The fastest case is the first. */ | |
463c84b9 ACM |
584 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; |
585 | } else if (m < icsk->icsk_ack.ato) { | |
586 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; | |
587 | if (icsk->icsk_ack.ato > icsk->icsk_rto) | |
588 | icsk->icsk_ack.ato = icsk->icsk_rto; | |
589 | } else if (m > icsk->icsk_rto) { | |
caa20d9a | 590 | /* Too long gap. Apparently sender failed to |
1da177e4 LT |
591 | * restart window, so that we send ACKs quickly. |
592 | */ | |
463c84b9 | 593 | tcp_incr_quickack(sk); |
1da177e4 LT |
594 | sk_stream_mem_reclaim(sk); |
595 | } | |
596 | } | |
463c84b9 | 597 | icsk->icsk_ack.lrcvtime = now; |
1da177e4 LT |
598 | |
599 | TCP_ECN_check_ce(tp, skb); | |
600 | ||
601 | if (skb->len >= 128) | |
9e412ba7 | 602 | tcp_grow_window(sk, skb); |
1da177e4 LT |
603 | } |
604 | ||
05bb1fad DM |
605 | static u32 tcp_rto_min(struct sock *sk) |
606 | { | |
607 | struct dst_entry *dst = __sk_dst_get(sk); | |
608 | u32 rto_min = TCP_RTO_MIN; | |
609 | ||
5c127c58 | 610 | if (dst && dst_metric_locked(dst, RTAX_RTO_MIN)) |
05bb1fad DM |
611 | rto_min = dst->metrics[RTAX_RTO_MIN-1]; |
612 | return rto_min; | |
613 | } | |
614 | ||
1da177e4 LT |
615 | /* Called to compute a smoothed rtt estimate. The data fed to this |
616 | * routine either comes from timestamps, or from segments that were | |
617 | * known _not_ to have been retransmitted [see Karn/Partridge | |
618 | * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 | |
619 | * piece by Van Jacobson. | |
620 | * NOTE: the next three routines used to be one big routine. | |
621 | * To save cycles in the RFC 1323 implementation it was better to break | |
622 | * it up into three procedures. -- erics | |
623 | */ | |
2d2abbab | 624 | static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) |
1da177e4 | 625 | { |
6687e988 | 626 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
627 | long m = mrtt; /* RTT */ |
628 | ||
1da177e4 LT |
629 | /* The following amusing code comes from Jacobson's |
630 | * article in SIGCOMM '88. Note that rtt and mdev | |
631 | * are scaled versions of rtt and mean deviation. | |
e905a9ed | 632 | * This is designed to be as fast as possible |
1da177e4 LT |
633 | * m stands for "measurement". |
634 | * | |
635 | * On a 1990 paper the rto value is changed to: | |
636 | * RTO = rtt + 4 * mdev | |
637 | * | |
638 | * Funny. This algorithm seems to be very broken. | |
639 | * These formulae increase RTO, when it should be decreased, increase | |
31f34269 | 640 | * too slowly, when it should be increased quickly, decrease too quickly |
1da177e4 LT |
641 | * etc. I guess in BSD RTO takes ONE value, so that it is absolutely |
642 | * does not matter how to _calculate_ it. Seems, it was trap | |
643 | * that VJ failed to avoid. 8) | |
644 | */ | |
2de979bd | 645 | if (m == 0) |
1da177e4 LT |
646 | m = 1; |
647 | if (tp->srtt != 0) { | |
648 | m -= (tp->srtt >> 3); /* m is now error in rtt est */ | |
649 | tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ | |
650 | if (m < 0) { | |
651 | m = -m; /* m is now abs(error) */ | |
652 | m -= (tp->mdev >> 2); /* similar update on mdev */ | |
653 | /* This is similar to one of Eifel findings. | |
654 | * Eifel blocks mdev updates when rtt decreases. | |
655 | * This solution is a bit different: we use finer gain | |
656 | * for mdev in this case (alpha*beta). | |
657 | * Like Eifel it also prevents growth of rto, | |
658 | * but also it limits too fast rto decreases, | |
659 | * happening in pure Eifel. | |
660 | */ | |
661 | if (m > 0) | |
662 | m >>= 3; | |
663 | } else { | |
664 | m -= (tp->mdev >> 2); /* similar update on mdev */ | |
665 | } | |
666 | tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ | |
667 | if (tp->mdev > tp->mdev_max) { | |
668 | tp->mdev_max = tp->mdev; | |
669 | if (tp->mdev_max > tp->rttvar) | |
670 | tp->rttvar = tp->mdev_max; | |
671 | } | |
672 | if (after(tp->snd_una, tp->rtt_seq)) { | |
673 | if (tp->mdev_max < tp->rttvar) | |
674 | tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2; | |
675 | tp->rtt_seq = tp->snd_nxt; | |
05bb1fad | 676 | tp->mdev_max = tcp_rto_min(sk); |
1da177e4 LT |
677 | } |
678 | } else { | |
679 | /* no previous measure. */ | |
680 | tp->srtt = m<<3; /* take the measured time to be rtt */ | |
681 | tp->mdev = m<<1; /* make sure rto = 3*rtt */ | |
05bb1fad | 682 | tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); |
1da177e4 LT |
683 | tp->rtt_seq = tp->snd_nxt; |
684 | } | |
1da177e4 LT |
685 | } |
686 | ||
687 | /* Calculate rto without backoff. This is the second half of Van Jacobson's | |
688 | * routine referred to above. | |
689 | */ | |
463c84b9 | 690 | static inline void tcp_set_rto(struct sock *sk) |
1da177e4 | 691 | { |
463c84b9 | 692 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
693 | /* Old crap is replaced with new one. 8) |
694 | * | |
695 | * More seriously: | |
696 | * 1. If rtt variance happened to be less 50msec, it is hallucination. | |
697 | * It cannot be less due to utterly erratic ACK generation made | |
698 | * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ | |
699 | * to do with delayed acks, because at cwnd>2 true delack timeout | |
700 | * is invisible. Actually, Linux-2.4 also generates erratic | |
caa20d9a | 701 | * ACKs in some circumstances. |
1da177e4 | 702 | */ |
463c84b9 | 703 | inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar; |
1da177e4 LT |
704 | |
705 | /* 2. Fixups made earlier cannot be right. | |
706 | * If we do not estimate RTO correctly without them, | |
707 | * all the algo is pure shit and should be replaced | |
caa20d9a | 708 | * with correct one. It is exactly, which we pretend to do. |
1da177e4 LT |
709 | */ |
710 | } | |
711 | ||
712 | /* NOTE: clamping at TCP_RTO_MIN is not required, current algo | |
713 | * guarantees that rto is higher. | |
714 | */ | |
463c84b9 | 715 | static inline void tcp_bound_rto(struct sock *sk) |
1da177e4 | 716 | { |
463c84b9 ACM |
717 | if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX) |
718 | inet_csk(sk)->icsk_rto = TCP_RTO_MAX; | |
1da177e4 LT |
719 | } |
720 | ||
721 | /* Save metrics learned by this TCP session. | |
722 | This function is called only, when TCP finishes successfully | |
723 | i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. | |
724 | */ | |
725 | void tcp_update_metrics(struct sock *sk) | |
726 | { | |
727 | struct tcp_sock *tp = tcp_sk(sk); | |
728 | struct dst_entry *dst = __sk_dst_get(sk); | |
729 | ||
730 | if (sysctl_tcp_nometrics_save) | |
731 | return; | |
732 | ||
733 | dst_confirm(dst); | |
734 | ||
735 | if (dst && (dst->flags&DST_HOST)) { | |
6687e988 | 736 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
737 | int m; |
738 | ||
6687e988 | 739 | if (icsk->icsk_backoff || !tp->srtt) { |
1da177e4 LT |
740 | /* This session failed to estimate rtt. Why? |
741 | * Probably, no packets returned in time. | |
742 | * Reset our results. | |
743 | */ | |
744 | if (!(dst_metric_locked(dst, RTAX_RTT))) | |
745 | dst->metrics[RTAX_RTT-1] = 0; | |
746 | return; | |
747 | } | |
748 | ||
749 | m = dst_metric(dst, RTAX_RTT) - tp->srtt; | |
750 | ||
751 | /* If newly calculated rtt larger than stored one, | |
752 | * store new one. Otherwise, use EWMA. Remember, | |
753 | * rtt overestimation is always better than underestimation. | |
754 | */ | |
755 | if (!(dst_metric_locked(dst, RTAX_RTT))) { | |
756 | if (m <= 0) | |
757 | dst->metrics[RTAX_RTT-1] = tp->srtt; | |
758 | else | |
759 | dst->metrics[RTAX_RTT-1] -= (m>>3); | |
760 | } | |
761 | ||
762 | if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { | |
763 | if (m < 0) | |
764 | m = -m; | |
765 | ||
766 | /* Scale deviation to rttvar fixed point */ | |
767 | m >>= 1; | |
768 | if (m < tp->mdev) | |
769 | m = tp->mdev; | |
770 | ||
771 | if (m >= dst_metric(dst, RTAX_RTTVAR)) | |
772 | dst->metrics[RTAX_RTTVAR-1] = m; | |
773 | else | |
774 | dst->metrics[RTAX_RTTVAR-1] -= | |
775 | (dst->metrics[RTAX_RTTVAR-1] - m)>>2; | |
776 | } | |
777 | ||
778 | if (tp->snd_ssthresh >= 0xFFFF) { | |
779 | /* Slow start still did not finish. */ | |
780 | if (dst_metric(dst, RTAX_SSTHRESH) && | |
781 | !dst_metric_locked(dst, RTAX_SSTHRESH) && | |
782 | (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) | |
783 | dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1; | |
784 | if (!dst_metric_locked(dst, RTAX_CWND) && | |
785 | tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) | |
786 | dst->metrics[RTAX_CWND-1] = tp->snd_cwnd; | |
787 | } else if (tp->snd_cwnd > tp->snd_ssthresh && | |
6687e988 | 788 | icsk->icsk_ca_state == TCP_CA_Open) { |
1da177e4 LT |
789 | /* Cong. avoidance phase, cwnd is reliable. */ |
790 | if (!dst_metric_locked(dst, RTAX_SSTHRESH)) | |
791 | dst->metrics[RTAX_SSTHRESH-1] = | |
792 | max(tp->snd_cwnd >> 1, tp->snd_ssthresh); | |
793 | if (!dst_metric_locked(dst, RTAX_CWND)) | |
794 | dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1; | |
795 | } else { | |
796 | /* Else slow start did not finish, cwnd is non-sense, | |
797 | ssthresh may be also invalid. | |
798 | */ | |
799 | if (!dst_metric_locked(dst, RTAX_CWND)) | |
800 | dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1; | |
801 | if (dst->metrics[RTAX_SSTHRESH-1] && | |
802 | !dst_metric_locked(dst, RTAX_SSTHRESH) && | |
803 | tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1]) | |
804 | dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh; | |
805 | } | |
806 | ||
807 | if (!dst_metric_locked(dst, RTAX_REORDERING)) { | |
808 | if (dst->metrics[RTAX_REORDERING-1] < tp->reordering && | |
809 | tp->reordering != sysctl_tcp_reordering) | |
810 | dst->metrics[RTAX_REORDERING-1] = tp->reordering; | |
811 | } | |
812 | } | |
813 | } | |
814 | ||
26722873 DM |
815 | /* Numbers are taken from RFC3390. |
816 | * | |
817 | * John Heffner states: | |
818 | * | |
819 | * The RFC specifies a window of no more than 4380 bytes | |
820 | * unless 2*MSS > 4380. Reading the pseudocode in the RFC | |
821 | * is a bit misleading because they use a clamp at 4380 bytes | |
822 | * rather than use a multiplier in the relevant range. | |
823 | */ | |
1da177e4 LT |
824 | __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst) |
825 | { | |
826 | __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); | |
827 | ||
828 | if (!cwnd) { | |
c1b4a7e6 | 829 | if (tp->mss_cache > 1460) |
1da177e4 LT |
830 | cwnd = 2; |
831 | else | |
c1b4a7e6 | 832 | cwnd = (tp->mss_cache > 1095) ? 3 : 4; |
1da177e4 LT |
833 | } |
834 | return min_t(__u32, cwnd, tp->snd_cwnd_clamp); | |
835 | } | |
836 | ||
40efc6fa | 837 | /* Set slow start threshold and cwnd not falling to slow start */ |
3cfe3baa | 838 | void tcp_enter_cwr(struct sock *sk, const int set_ssthresh) |
40efc6fa SH |
839 | { |
840 | struct tcp_sock *tp = tcp_sk(sk); | |
3cfe3baa | 841 | const struct inet_connection_sock *icsk = inet_csk(sk); |
40efc6fa SH |
842 | |
843 | tp->prior_ssthresh = 0; | |
844 | tp->bytes_acked = 0; | |
e01f9d77 | 845 | if (icsk->icsk_ca_state < TCP_CA_CWR) { |
40efc6fa | 846 | tp->undo_marker = 0; |
3cfe3baa IJ |
847 | if (set_ssthresh) |
848 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
40efc6fa SH |
849 | tp->snd_cwnd = min(tp->snd_cwnd, |
850 | tcp_packets_in_flight(tp) + 1U); | |
851 | tp->snd_cwnd_cnt = 0; | |
852 | tp->high_seq = tp->snd_nxt; | |
853 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
854 | TCP_ECN_queue_cwr(tp); | |
855 | ||
856 | tcp_set_ca_state(sk, TCP_CA_CWR); | |
857 | } | |
858 | } | |
859 | ||
e60402d0 IJ |
860 | /* |
861 | * Packet counting of FACK is based on in-order assumptions, therefore TCP | |
862 | * disables it when reordering is detected | |
863 | */ | |
864 | static void tcp_disable_fack(struct tcp_sock *tp) | |
865 | { | |
85cc391c IJ |
866 | /* RFC3517 uses different metric in lost marker => reset on change */ |
867 | if (tcp_is_fack(tp)) | |
868 | tp->lost_skb_hint = NULL; | |
e60402d0 IJ |
869 | tp->rx_opt.sack_ok &= ~2; |
870 | } | |
871 | ||
564262c1 | 872 | /* Take a notice that peer is sending D-SACKs */ |
e60402d0 IJ |
873 | static void tcp_dsack_seen(struct tcp_sock *tp) |
874 | { | |
875 | tp->rx_opt.sack_ok |= 4; | |
876 | } | |
877 | ||
1da177e4 LT |
878 | /* Initialize metrics on socket. */ |
879 | ||
880 | static void tcp_init_metrics(struct sock *sk) | |
881 | { | |
882 | struct tcp_sock *tp = tcp_sk(sk); | |
883 | struct dst_entry *dst = __sk_dst_get(sk); | |
884 | ||
885 | if (dst == NULL) | |
886 | goto reset; | |
887 | ||
888 | dst_confirm(dst); | |
889 | ||
890 | if (dst_metric_locked(dst, RTAX_CWND)) | |
891 | tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); | |
892 | if (dst_metric(dst, RTAX_SSTHRESH)) { | |
893 | tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); | |
894 | if (tp->snd_ssthresh > tp->snd_cwnd_clamp) | |
895 | tp->snd_ssthresh = tp->snd_cwnd_clamp; | |
896 | } | |
897 | if (dst_metric(dst, RTAX_REORDERING) && | |
898 | tp->reordering != dst_metric(dst, RTAX_REORDERING)) { | |
e60402d0 | 899 | tcp_disable_fack(tp); |
1da177e4 LT |
900 | tp->reordering = dst_metric(dst, RTAX_REORDERING); |
901 | } | |
902 | ||
903 | if (dst_metric(dst, RTAX_RTT) == 0) | |
904 | goto reset; | |
905 | ||
906 | if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3)) | |
907 | goto reset; | |
908 | ||
909 | /* Initial rtt is determined from SYN,SYN-ACK. | |
910 | * The segment is small and rtt may appear much | |
911 | * less than real one. Use per-dst memory | |
912 | * to make it more realistic. | |
913 | * | |
914 | * A bit of theory. RTT is time passed after "normal" sized packet | |
caa20d9a | 915 | * is sent until it is ACKed. In normal circumstances sending small |
1da177e4 LT |
916 | * packets force peer to delay ACKs and calculation is correct too. |
917 | * The algorithm is adaptive and, provided we follow specs, it | |
918 | * NEVER underestimate RTT. BUT! If peer tries to make some clever | |
919 | * tricks sort of "quick acks" for time long enough to decrease RTT | |
920 | * to low value, and then abruptly stops to do it and starts to delay | |
921 | * ACKs, wait for troubles. | |
922 | */ | |
923 | if (dst_metric(dst, RTAX_RTT) > tp->srtt) { | |
924 | tp->srtt = dst_metric(dst, RTAX_RTT); | |
925 | tp->rtt_seq = tp->snd_nxt; | |
926 | } | |
927 | if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) { | |
928 | tp->mdev = dst_metric(dst, RTAX_RTTVAR); | |
488faa2a | 929 | tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); |
1da177e4 | 930 | } |
463c84b9 ACM |
931 | tcp_set_rto(sk); |
932 | tcp_bound_rto(sk); | |
933 | if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp) | |
1da177e4 LT |
934 | goto reset; |
935 | tp->snd_cwnd = tcp_init_cwnd(tp, dst); | |
936 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
937 | return; | |
938 | ||
939 | reset: | |
940 | /* Play conservative. If timestamps are not | |
941 | * supported, TCP will fail to recalculate correct | |
942 | * rtt, if initial rto is too small. FORGET ALL AND RESET! | |
943 | */ | |
944 | if (!tp->rx_opt.saw_tstamp && tp->srtt) { | |
945 | tp->srtt = 0; | |
946 | tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT; | |
463c84b9 | 947 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
1da177e4 LT |
948 | } |
949 | } | |
950 | ||
6687e988 ACM |
951 | static void tcp_update_reordering(struct sock *sk, const int metric, |
952 | const int ts) | |
1da177e4 | 953 | { |
6687e988 | 954 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
955 | if (metric > tp->reordering) { |
956 | tp->reordering = min(TCP_MAX_REORDERING, metric); | |
957 | ||
958 | /* This exciting event is worth to be remembered. 8) */ | |
959 | if (ts) | |
960 | NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER); | |
e60402d0 | 961 | else if (tcp_is_reno(tp)) |
1da177e4 | 962 | NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER); |
e60402d0 | 963 | else if (tcp_is_fack(tp)) |
1da177e4 LT |
964 | NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER); |
965 | else | |
966 | NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER); | |
967 | #if FASTRETRANS_DEBUG > 1 | |
968 | printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", | |
6687e988 | 969 | tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, |
1da177e4 LT |
970 | tp->reordering, |
971 | tp->fackets_out, | |
972 | tp->sacked_out, | |
973 | tp->undo_marker ? tp->undo_retrans : 0); | |
974 | #endif | |
e60402d0 | 975 | tcp_disable_fack(tp); |
1da177e4 LT |
976 | } |
977 | } | |
978 | ||
979 | /* This procedure tags the retransmission queue when SACKs arrive. | |
980 | * | |
981 | * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). | |
982 | * Packets in queue with these bits set are counted in variables | |
983 | * sacked_out, retrans_out and lost_out, correspondingly. | |
984 | * | |
985 | * Valid combinations are: | |
986 | * Tag InFlight Description | |
987 | * 0 1 - orig segment is in flight. | |
988 | * S 0 - nothing flies, orig reached receiver. | |
989 | * L 0 - nothing flies, orig lost by net. | |
990 | * R 2 - both orig and retransmit are in flight. | |
991 | * L|R 1 - orig is lost, retransmit is in flight. | |
992 | * S|R 1 - orig reached receiver, retrans is still in flight. | |
993 | * (L|S|R is logically valid, it could occur when L|R is sacked, | |
994 | * but it is equivalent to plain S and code short-curcuits it to S. | |
995 | * L|S is logically invalid, it would mean -1 packet in flight 8)) | |
996 | * | |
997 | * These 6 states form finite state machine, controlled by the following events: | |
998 | * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) | |
999 | * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) | |
1000 | * 3. Loss detection event of one of three flavors: | |
1001 | * A. Scoreboard estimator decided the packet is lost. | |
1002 | * A'. Reno "three dupacks" marks head of queue lost. | |
1003 | * A''. Its FACK modfication, head until snd.fack is lost. | |
1004 | * B. SACK arrives sacking data transmitted after never retransmitted | |
1005 | * hole was sent out. | |
1006 | * C. SACK arrives sacking SND.NXT at the moment, when the | |
1007 | * segment was retransmitted. | |
1008 | * 4. D-SACK added new rule: D-SACK changes any tag to S. | |
1009 | * | |
1010 | * It is pleasant to note, that state diagram turns out to be commutative, | |
1011 | * so that we are allowed not to be bothered by order of our actions, | |
1012 | * when multiple events arrive simultaneously. (see the function below). | |
1013 | * | |
1014 | * Reordering detection. | |
1015 | * -------------------- | |
1016 | * Reordering metric is maximal distance, which a packet can be displaced | |
1017 | * in packet stream. With SACKs we can estimate it: | |
1018 | * | |
1019 | * 1. SACK fills old hole and the corresponding segment was not | |
1020 | * ever retransmitted -> reordering. Alas, we cannot use it | |
1021 | * when segment was retransmitted. | |
1022 | * 2. The last flaw is solved with D-SACK. D-SACK arrives | |
1023 | * for retransmitted and already SACKed segment -> reordering.. | |
1024 | * Both of these heuristics are not used in Loss state, when we cannot | |
1025 | * account for retransmits accurately. | |
5b3c9882 IJ |
1026 | * |
1027 | * SACK block validation. | |
1028 | * ---------------------- | |
1029 | * | |
1030 | * SACK block range validation checks that the received SACK block fits to | |
1031 | * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. | |
1032 | * Note that SND.UNA is not included to the range though being valid because | |
0e835331 IJ |
1033 | * it means that the receiver is rather inconsistent with itself reporting |
1034 | * SACK reneging when it should advance SND.UNA. Such SACK block this is | |
1035 | * perfectly valid, however, in light of RFC2018 which explicitly states | |
1036 | * that "SACK block MUST reflect the newest segment. Even if the newest | |
1037 | * segment is going to be discarded ...", not that it looks very clever | |
1038 | * in case of head skb. Due to potentional receiver driven attacks, we | |
1039 | * choose to avoid immediate execution of a walk in write queue due to | |
1040 | * reneging and defer head skb's loss recovery to standard loss recovery | |
1041 | * procedure that will eventually trigger (nothing forbids us doing this). | |
5b3c9882 IJ |
1042 | * |
1043 | * Implements also blockage to start_seq wrap-around. Problem lies in the | |
1044 | * fact that though start_seq (s) is before end_seq (i.e., not reversed), | |
1045 | * there's no guarantee that it will be before snd_nxt (n). The problem | |
1046 | * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt | |
1047 | * wrap (s_w): | |
1048 | * | |
1049 | * <- outs wnd -> <- wrapzone -> | |
1050 | * u e n u_w e_w s n_w | |
1051 | * | | | | | | | | |
1052 | * |<------------+------+----- TCP seqno space --------------+---------->| | |
1053 | * ...-- <2^31 ->| |<--------... | |
1054 | * ...---- >2^31 ------>| |<--------... | |
1055 | * | |
1056 | * Current code wouldn't be vulnerable but it's better still to discard such | |
1057 | * crazy SACK blocks. Doing this check for start_seq alone closes somewhat | |
1058 | * similar case (end_seq after snd_nxt wrap) as earlier reversed check in | |
1059 | * snd_nxt wrap -> snd_una region will then become "well defined", i.e., | |
1060 | * equal to the ideal case (infinite seqno space without wrap caused issues). | |
1061 | * | |
1062 | * With D-SACK the lower bound is extended to cover sequence space below | |
1063 | * SND.UNA down to undo_marker, which is the last point of interest. Yet | |
564262c1 | 1064 | * again, D-SACK block must not to go across snd_una (for the same reason as |
5b3c9882 IJ |
1065 | * for the normal SACK blocks, explained above). But there all simplicity |
1066 | * ends, TCP might receive valid D-SACKs below that. As long as they reside | |
1067 | * fully below undo_marker they do not affect behavior in anyway and can | |
1068 | * therefore be safely ignored. In rare cases (which are more or less | |
1069 | * theoretical ones), the D-SACK will nicely cross that boundary due to skb | |
1070 | * fragmentation and packet reordering past skb's retransmission. To consider | |
1071 | * them correctly, the acceptable range must be extended even more though | |
1072 | * the exact amount is rather hard to quantify. However, tp->max_window can | |
1073 | * be used as an exaggerated estimate. | |
1da177e4 | 1074 | */ |
5b3c9882 IJ |
1075 | static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack, |
1076 | u32 start_seq, u32 end_seq) | |
1077 | { | |
1078 | /* Too far in future, or reversed (interpretation is ambiguous) */ | |
1079 | if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) | |
1080 | return 0; | |
1081 | ||
1082 | /* Nasty start_seq wrap-around check (see comments above) */ | |
1083 | if (!before(start_seq, tp->snd_nxt)) | |
1084 | return 0; | |
1085 | ||
564262c1 | 1086 | /* In outstanding window? ...This is valid exit for D-SACKs too. |
5b3c9882 IJ |
1087 | * start_seq == snd_una is non-sensical (see comments above) |
1088 | */ | |
1089 | if (after(start_seq, tp->snd_una)) | |
1090 | return 1; | |
1091 | ||
1092 | if (!is_dsack || !tp->undo_marker) | |
1093 | return 0; | |
1094 | ||
1095 | /* ...Then it's D-SACK, and must reside below snd_una completely */ | |
1096 | if (!after(end_seq, tp->snd_una)) | |
1097 | return 0; | |
1098 | ||
1099 | if (!before(start_seq, tp->undo_marker)) | |
1100 | return 1; | |
1101 | ||
1102 | /* Too old */ | |
1103 | if (!after(end_seq, tp->undo_marker)) | |
1104 | return 0; | |
1105 | ||
1106 | /* Undo_marker boundary crossing (overestimates a lot). Known already: | |
1107 | * start_seq < undo_marker and end_seq >= undo_marker. | |
1108 | */ | |
1109 | return !before(start_seq, end_seq - tp->max_window); | |
1110 | } | |
1111 | ||
1c1e87ed IJ |
1112 | /* Check for lost retransmit. This superb idea is borrowed from "ratehalving". |
1113 | * Event "C". Later note: FACK people cheated me again 8), we have to account | |
1114 | * for reordering! Ugly, but should help. | |
f785a8e2 IJ |
1115 | * |
1116 | * Search retransmitted skbs from write_queue that were sent when snd_nxt was | |
1117 | * less than what is now known to be received by the other end (derived from | |
9f58f3b7 IJ |
1118 | * highest SACK block). Also calculate the lowest snd_nxt among the remaining |
1119 | * retransmitted skbs to avoid some costly processing per ACKs. | |
1c1e87ed | 1120 | */ |
9f58f3b7 | 1121 | static int tcp_mark_lost_retrans(struct sock *sk) |
1c1e87ed | 1122 | { |
9f58f3b7 | 1123 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1c1e87ed IJ |
1124 | struct tcp_sock *tp = tcp_sk(sk); |
1125 | struct sk_buff *skb; | |
1126 | int flag = 0; | |
f785a8e2 | 1127 | int cnt = 0; |
df2e014b | 1128 | u32 new_low_seq = tp->snd_nxt; |
9f58f3b7 IJ |
1129 | u32 received_upto = TCP_SKB_CB(tp->highest_sack)->end_seq; |
1130 | ||
1131 | if (!tcp_is_fack(tp) || !tp->retrans_out || | |
1132 | !after(received_upto, tp->lost_retrans_low) || | |
1133 | icsk->icsk_ca_state != TCP_CA_Recovery) | |
1134 | return flag; | |
1c1e87ed IJ |
1135 | |
1136 | tcp_for_write_queue(skb, sk) { | |
1137 | u32 ack_seq = TCP_SKB_CB(skb)->ack_seq; | |
1138 | ||
1139 | if (skb == tcp_send_head(sk)) | |
1140 | break; | |
f785a8e2 | 1141 | if (cnt == tp->retrans_out) |
1c1e87ed IJ |
1142 | break; |
1143 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1144 | continue; | |
1145 | ||
f785a8e2 IJ |
1146 | if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) |
1147 | continue; | |
1148 | ||
1149 | if (after(received_upto, ack_seq) && | |
1c1e87ed | 1150 | (tcp_is_fack(tp) || |
f785a8e2 | 1151 | !before(received_upto, |
1c1e87ed IJ |
1152 | ack_seq + tp->reordering * tp->mss_cache))) { |
1153 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; | |
1154 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1155 | ||
1156 | /* clear lost hint */ | |
1157 | tp->retransmit_skb_hint = NULL; | |
1158 | ||
1159 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { | |
1160 | tp->lost_out += tcp_skb_pcount(skb); | |
1161 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1162 | flag |= FLAG_DATA_SACKED; | |
1163 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT); | |
1164 | } | |
f785a8e2 | 1165 | } else { |
df2e014b | 1166 | if (before(ack_seq, new_low_seq)) |
b08d6cb2 | 1167 | new_low_seq = ack_seq; |
f785a8e2 | 1168 | cnt += tcp_skb_pcount(skb); |
1c1e87ed IJ |
1169 | } |
1170 | } | |
b08d6cb2 IJ |
1171 | |
1172 | if (tp->retrans_out) | |
1173 | tp->lost_retrans_low = new_low_seq; | |
1174 | ||
1c1e87ed IJ |
1175 | return flag; |
1176 | } | |
5b3c9882 | 1177 | |
d06e021d DM |
1178 | static int tcp_check_dsack(struct tcp_sock *tp, struct sk_buff *ack_skb, |
1179 | struct tcp_sack_block_wire *sp, int num_sacks, | |
1180 | u32 prior_snd_una) | |
1181 | { | |
1182 | u32 start_seq_0 = ntohl(get_unaligned(&sp[0].start_seq)); | |
1183 | u32 end_seq_0 = ntohl(get_unaligned(&sp[0].end_seq)); | |
1184 | int dup_sack = 0; | |
1185 | ||
1186 | if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { | |
1187 | dup_sack = 1; | |
e60402d0 | 1188 | tcp_dsack_seen(tp); |
d06e021d DM |
1189 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV); |
1190 | } else if (num_sacks > 1) { | |
1191 | u32 end_seq_1 = ntohl(get_unaligned(&sp[1].end_seq)); | |
1192 | u32 start_seq_1 = ntohl(get_unaligned(&sp[1].start_seq)); | |
1193 | ||
1194 | if (!after(end_seq_0, end_seq_1) && | |
1195 | !before(start_seq_0, start_seq_1)) { | |
1196 | dup_sack = 1; | |
e60402d0 | 1197 | tcp_dsack_seen(tp); |
d06e021d DM |
1198 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV); |
1199 | } | |
1200 | } | |
1201 | ||
1202 | /* D-SACK for already forgotten data... Do dumb counting. */ | |
1203 | if (dup_sack && | |
1204 | !after(end_seq_0, prior_snd_una) && | |
1205 | after(end_seq_0, tp->undo_marker)) | |
1206 | tp->undo_retrans--; | |
1207 | ||
1208 | return dup_sack; | |
1209 | } | |
1210 | ||
d1935942 IJ |
1211 | /* Check if skb is fully within the SACK block. In presence of GSO skbs, |
1212 | * the incoming SACK may not exactly match but we can find smaller MSS | |
1213 | * aligned portion of it that matches. Therefore we might need to fragment | |
1214 | * which may fail and creates some hassle (caller must handle error case | |
1215 | * returns). | |
1216 | */ | |
0f79efdc AB |
1217 | static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, |
1218 | u32 start_seq, u32 end_seq) | |
d1935942 IJ |
1219 | { |
1220 | int in_sack, err; | |
1221 | unsigned int pkt_len; | |
1222 | ||
1223 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && | |
1224 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1225 | ||
1226 | if (tcp_skb_pcount(skb) > 1 && !in_sack && | |
1227 | after(TCP_SKB_CB(skb)->end_seq, start_seq)) { | |
1228 | ||
1229 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); | |
1230 | ||
1231 | if (!in_sack) | |
1232 | pkt_len = start_seq - TCP_SKB_CB(skb)->seq; | |
1233 | else | |
1234 | pkt_len = end_seq - TCP_SKB_CB(skb)->seq; | |
1235 | err = tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size); | |
1236 | if (err < 0) | |
1237 | return err; | |
1238 | } | |
1239 | ||
1240 | return in_sack; | |
1241 | } | |
1242 | ||
9e10c47c IJ |
1243 | static int tcp_sacktag_one(struct sk_buff *skb, struct tcp_sock *tp, |
1244 | int *reord, int dup_sack, int fack_count) | |
1245 | { | |
1246 | u8 sacked = TCP_SKB_CB(skb)->sacked; | |
1247 | int flag = 0; | |
1248 | ||
1249 | /* Account D-SACK for retransmitted packet. */ | |
1250 | if (dup_sack && (sacked & TCPCB_RETRANS)) { | |
1251 | if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker)) | |
1252 | tp->undo_retrans--; | |
1253 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una) && | |
1254 | (sacked & TCPCB_SACKED_ACKED)) | |
1255 | *reord = min(fack_count, *reord); | |
1256 | } | |
1257 | ||
1258 | /* Nothing to do; acked frame is about to be dropped (was ACKed). */ | |
1259 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1260 | return flag; | |
1261 | ||
1262 | if (!(sacked & TCPCB_SACKED_ACKED)) { | |
1263 | if (sacked & TCPCB_SACKED_RETRANS) { | |
1264 | /* If the segment is not tagged as lost, | |
1265 | * we do not clear RETRANS, believing | |
1266 | * that retransmission is still in flight. | |
1267 | */ | |
1268 | if (sacked & TCPCB_LOST) { | |
1269 | TCP_SKB_CB(skb)->sacked &= | |
1270 | ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); | |
1271 | tp->lost_out -= tcp_skb_pcount(skb); | |
1272 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1273 | ||
1274 | /* clear lost hint */ | |
1275 | tp->retransmit_skb_hint = NULL; | |
1276 | } | |
1277 | } else { | |
1278 | if (!(sacked & TCPCB_RETRANS)) { | |
1279 | /* New sack for not retransmitted frame, | |
1280 | * which was in hole. It is reordering. | |
1281 | */ | |
1282 | if (before(TCP_SKB_CB(skb)->seq, | |
1283 | tcp_highest_sack_seq(tp))) | |
1284 | *reord = min(fack_count, *reord); | |
1285 | ||
1286 | /* SACK enhanced F-RTO (RFC4138; Appendix B) */ | |
1287 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) | |
1288 | flag |= FLAG_ONLY_ORIG_SACKED; | |
1289 | } | |
1290 | ||
1291 | if (sacked & TCPCB_LOST) { | |
1292 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; | |
1293 | tp->lost_out -= tcp_skb_pcount(skb); | |
1294 | ||
1295 | /* clear lost hint */ | |
1296 | tp->retransmit_skb_hint = NULL; | |
1297 | } | |
1298 | } | |
1299 | ||
1300 | TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED; | |
1301 | flag |= FLAG_DATA_SACKED; | |
1302 | tp->sacked_out += tcp_skb_pcount(skb); | |
1303 | ||
1304 | fack_count += tcp_skb_pcount(skb); | |
1305 | ||
1306 | /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ | |
1307 | if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) && | |
1308 | before(TCP_SKB_CB(skb)->seq, | |
1309 | TCP_SKB_CB(tp->lost_skb_hint)->seq)) | |
1310 | tp->lost_cnt_hint += tcp_skb_pcount(skb); | |
1311 | ||
1312 | if (fack_count > tp->fackets_out) | |
1313 | tp->fackets_out = fack_count; | |
1314 | ||
1315 | if (after(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) | |
1316 | tp->highest_sack = skb; | |
1317 | ||
1318 | } else { | |
1319 | if (dup_sack && (sacked & TCPCB_RETRANS)) | |
1320 | *reord = min(fack_count, *reord); | |
1321 | } | |
1322 | ||
1323 | /* D-SACK. We can detect redundant retransmission in S|R and plain R | |
1324 | * frames and clear it. undo_retrans is decreased above, L|R frames | |
1325 | * are accounted above as well. | |
1326 | */ | |
1327 | if (dup_sack && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) { | |
1328 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; | |
1329 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1330 | tp->retransmit_skb_hint = NULL; | |
1331 | } | |
1332 | ||
1333 | return flag; | |
1334 | } | |
1335 | ||
1da177e4 LT |
1336 | static int |
1337 | tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una) | |
1338 | { | |
6687e988 | 1339 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 1340 | struct tcp_sock *tp = tcp_sk(sk); |
9c70220b ACM |
1341 | unsigned char *ptr = (skb_transport_header(ack_skb) + |
1342 | TCP_SKB_CB(ack_skb)->sacked); | |
fd6dad61 IJ |
1343 | struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); |
1344 | struct tcp_sack_block sp[4]; | |
fda03fbb | 1345 | struct sk_buff *cached_skb; |
1da177e4 | 1346 | int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3; |
fd6dad61 | 1347 | int used_sacks; |
1da177e4 | 1348 | int reord = tp->packets_out; |
1da177e4 | 1349 | int flag = 0; |
7769f406 | 1350 | int found_dup_sack = 0; |
fda03fbb | 1351 | int cached_fack_count; |
1da177e4 | 1352 | int i; |
fda03fbb | 1353 | int first_sack_index; |
94d3b1e5 | 1354 | int force_one_sack; |
1da177e4 | 1355 | |
d738cd8f | 1356 | if (!tp->sacked_out) { |
de83c058 IJ |
1357 | if (WARN_ON(tp->fackets_out)) |
1358 | tp->fackets_out = 0; | |
a47e5a98 | 1359 | tp->highest_sack = tcp_write_queue_head(sk); |
d738cd8f | 1360 | } |
1da177e4 | 1361 | |
fd6dad61 | 1362 | found_dup_sack = tcp_check_dsack(tp, ack_skb, sp_wire, |
d06e021d DM |
1363 | num_sacks, prior_snd_una); |
1364 | if (found_dup_sack) | |
49ff4bb4 | 1365 | flag |= FLAG_DSACKING_ACK; |
6f74651a BE |
1366 | |
1367 | /* Eliminate too old ACKs, but take into | |
1368 | * account more or less fresh ones, they can | |
1369 | * contain valid SACK info. | |
1370 | */ | |
1371 | if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) | |
1372 | return 0; | |
1373 |