]>
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; | |
88 | int sysctl_tcp_frto __read_mostly; | |
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) */ |
49ff4bb4 | 106 | #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained DSACK 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 | { | |
866 | tp->rx_opt.sack_ok &= ~2; | |
867 | } | |
868 | ||
869 | /* Take a notice that peer is sending DSACKs */ | |
870 | static void tcp_dsack_seen(struct tcp_sock *tp) | |
871 | { | |
872 | tp->rx_opt.sack_ok |= 4; | |
873 | } | |
874 | ||
1da177e4 LT |
875 | /* Initialize metrics on socket. */ |
876 | ||
877 | static void tcp_init_metrics(struct sock *sk) | |
878 | { | |
879 | struct tcp_sock *tp = tcp_sk(sk); | |
880 | struct dst_entry *dst = __sk_dst_get(sk); | |
881 | ||
882 | if (dst == NULL) | |
883 | goto reset; | |
884 | ||
885 | dst_confirm(dst); | |
886 | ||
887 | if (dst_metric_locked(dst, RTAX_CWND)) | |
888 | tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); | |
889 | if (dst_metric(dst, RTAX_SSTHRESH)) { | |
890 | tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); | |
891 | if (tp->snd_ssthresh > tp->snd_cwnd_clamp) | |
892 | tp->snd_ssthresh = tp->snd_cwnd_clamp; | |
893 | } | |
894 | if (dst_metric(dst, RTAX_REORDERING) && | |
895 | tp->reordering != dst_metric(dst, RTAX_REORDERING)) { | |
e60402d0 | 896 | tcp_disable_fack(tp); |
1da177e4 LT |
897 | tp->reordering = dst_metric(dst, RTAX_REORDERING); |
898 | } | |
899 | ||
900 | if (dst_metric(dst, RTAX_RTT) == 0) | |
901 | goto reset; | |
902 | ||
903 | if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3)) | |
904 | goto reset; | |
905 | ||
906 | /* Initial rtt is determined from SYN,SYN-ACK. | |
907 | * The segment is small and rtt may appear much | |
908 | * less than real one. Use per-dst memory | |
909 | * to make it more realistic. | |
910 | * | |
911 | * A bit of theory. RTT is time passed after "normal" sized packet | |
caa20d9a | 912 | * is sent until it is ACKed. In normal circumstances sending small |
1da177e4 LT |
913 | * packets force peer to delay ACKs and calculation is correct too. |
914 | * The algorithm is adaptive and, provided we follow specs, it | |
915 | * NEVER underestimate RTT. BUT! If peer tries to make some clever | |
916 | * tricks sort of "quick acks" for time long enough to decrease RTT | |
917 | * to low value, and then abruptly stops to do it and starts to delay | |
918 | * ACKs, wait for troubles. | |
919 | */ | |
920 | if (dst_metric(dst, RTAX_RTT) > tp->srtt) { | |
921 | tp->srtt = dst_metric(dst, RTAX_RTT); | |
922 | tp->rtt_seq = tp->snd_nxt; | |
923 | } | |
924 | if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) { | |
925 | tp->mdev = dst_metric(dst, RTAX_RTTVAR); | |
926 | tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN); | |
927 | } | |
463c84b9 ACM |
928 | tcp_set_rto(sk); |
929 | tcp_bound_rto(sk); | |
930 | if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp) | |
1da177e4 LT |
931 | goto reset; |
932 | tp->snd_cwnd = tcp_init_cwnd(tp, dst); | |
933 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
934 | return; | |
935 | ||
936 | reset: | |
937 | /* Play conservative. If timestamps are not | |
938 | * supported, TCP will fail to recalculate correct | |
939 | * rtt, if initial rto is too small. FORGET ALL AND RESET! | |
940 | */ | |
941 | if (!tp->rx_opt.saw_tstamp && tp->srtt) { | |
942 | tp->srtt = 0; | |
943 | tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT; | |
463c84b9 | 944 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
1da177e4 LT |
945 | } |
946 | } | |
947 | ||
6687e988 ACM |
948 | static void tcp_update_reordering(struct sock *sk, const int metric, |
949 | const int ts) | |
1da177e4 | 950 | { |
6687e988 | 951 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
952 | if (metric > tp->reordering) { |
953 | tp->reordering = min(TCP_MAX_REORDERING, metric); | |
954 | ||
955 | /* This exciting event is worth to be remembered. 8) */ | |
956 | if (ts) | |
957 | NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER); | |
e60402d0 | 958 | else if (tcp_is_reno(tp)) |
1da177e4 | 959 | NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER); |
e60402d0 | 960 | else if (tcp_is_fack(tp)) |
1da177e4 LT |
961 | NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER); |
962 | else | |
963 | NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER); | |
964 | #if FASTRETRANS_DEBUG > 1 | |
965 | printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", | |
6687e988 | 966 | tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, |
1da177e4 LT |
967 | tp->reordering, |
968 | tp->fackets_out, | |
969 | tp->sacked_out, | |
970 | tp->undo_marker ? tp->undo_retrans : 0); | |
971 | #endif | |
e60402d0 | 972 | tcp_disable_fack(tp); |
1da177e4 LT |
973 | } |
974 | } | |
975 | ||
976 | /* This procedure tags the retransmission queue when SACKs arrive. | |
977 | * | |
978 | * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). | |
979 | * Packets in queue with these bits set are counted in variables | |
980 | * sacked_out, retrans_out and lost_out, correspondingly. | |
981 | * | |
982 | * Valid combinations are: | |
983 | * Tag InFlight Description | |
984 | * 0 1 - orig segment is in flight. | |
985 | * S 0 - nothing flies, orig reached receiver. | |
986 | * L 0 - nothing flies, orig lost by net. | |
987 | * R 2 - both orig and retransmit are in flight. | |
988 | * L|R 1 - orig is lost, retransmit is in flight. | |
989 | * S|R 1 - orig reached receiver, retrans is still in flight. | |
990 | * (L|S|R is logically valid, it could occur when L|R is sacked, | |
991 | * but it is equivalent to plain S and code short-curcuits it to S. | |
992 | * L|S is logically invalid, it would mean -1 packet in flight 8)) | |
993 | * | |
994 | * These 6 states form finite state machine, controlled by the following events: | |
995 | * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) | |
996 | * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) | |
997 | * 3. Loss detection event of one of three flavors: | |
998 | * A. Scoreboard estimator decided the packet is lost. | |
999 | * A'. Reno "three dupacks" marks head of queue lost. | |
1000 | * A''. Its FACK modfication, head until snd.fack is lost. | |
1001 | * B. SACK arrives sacking data transmitted after never retransmitted | |
1002 | * hole was sent out. | |
1003 | * C. SACK arrives sacking SND.NXT at the moment, when the | |
1004 | * segment was retransmitted. | |
1005 | * 4. D-SACK added new rule: D-SACK changes any tag to S. | |
1006 | * | |
1007 | * It is pleasant to note, that state diagram turns out to be commutative, | |
1008 | * so that we are allowed not to be bothered by order of our actions, | |
1009 | * when multiple events arrive simultaneously. (see the function below). | |
1010 | * | |
1011 | * Reordering detection. | |
1012 | * -------------------- | |
1013 | * Reordering metric is maximal distance, which a packet can be displaced | |
1014 | * in packet stream. With SACKs we can estimate it: | |
1015 | * | |
1016 | * 1. SACK fills old hole and the corresponding segment was not | |
1017 | * ever retransmitted -> reordering. Alas, we cannot use it | |
1018 | * when segment was retransmitted. | |
1019 | * 2. The last flaw is solved with D-SACK. D-SACK arrives | |
1020 | * for retransmitted and already SACKed segment -> reordering.. | |
1021 | * Both of these heuristics are not used in Loss state, when we cannot | |
1022 | * account for retransmits accurately. | |
5b3c9882 IJ |
1023 | * |
1024 | * SACK block validation. | |
1025 | * ---------------------- | |
1026 | * | |
1027 | * SACK block range validation checks that the received SACK block fits to | |
1028 | * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. | |
1029 | * Note that SND.UNA is not included to the range though being valid because | |
1030 | * it means that the receiver is rather inconsistent with itself (reports | |
1031 | * SACK reneging when it should advance SND.UNA). | |
1032 | * | |
1033 | * Implements also blockage to start_seq wrap-around. Problem lies in the | |
1034 | * fact that though start_seq (s) is before end_seq (i.e., not reversed), | |
1035 | * there's no guarantee that it will be before snd_nxt (n). The problem | |
1036 | * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt | |
1037 | * wrap (s_w): | |
1038 | * | |
1039 | * <- outs wnd -> <- wrapzone -> | |
1040 | * u e n u_w e_w s n_w | |
1041 | * | | | | | | | | |
1042 | * |<------------+------+----- TCP seqno space --------------+---------->| | |
1043 | * ...-- <2^31 ->| |<--------... | |
1044 | * ...---- >2^31 ------>| |<--------... | |
1045 | * | |
1046 | * Current code wouldn't be vulnerable but it's better still to discard such | |
1047 | * crazy SACK blocks. Doing this check for start_seq alone closes somewhat | |
1048 | * similar case (end_seq after snd_nxt wrap) as earlier reversed check in | |
1049 | * snd_nxt wrap -> snd_una region will then become "well defined", i.e., | |
1050 | * equal to the ideal case (infinite seqno space without wrap caused issues). | |
1051 | * | |
1052 | * With D-SACK the lower bound is extended to cover sequence space below | |
1053 | * SND.UNA down to undo_marker, which is the last point of interest. Yet | |
1054 | * again, DSACK block must not to go across snd_una (for the same reason as | |
1055 | * for the normal SACK blocks, explained above). But there all simplicity | |
1056 | * ends, TCP might receive valid D-SACKs below that. As long as they reside | |
1057 | * fully below undo_marker they do not affect behavior in anyway and can | |
1058 | * therefore be safely ignored. In rare cases (which are more or less | |
1059 | * theoretical ones), the D-SACK will nicely cross that boundary due to skb | |
1060 | * fragmentation and packet reordering past skb's retransmission. To consider | |
1061 | * them correctly, the acceptable range must be extended even more though | |
1062 | * the exact amount is rather hard to quantify. However, tp->max_window can | |
1063 | * be used as an exaggerated estimate. | |
1da177e4 | 1064 | */ |
5b3c9882 IJ |
1065 | static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack, |
1066 | u32 start_seq, u32 end_seq) | |
1067 | { | |
1068 | /* Too far in future, or reversed (interpretation is ambiguous) */ | |
1069 | if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) | |
1070 | return 0; | |
1071 | ||
1072 | /* Nasty start_seq wrap-around check (see comments above) */ | |
1073 | if (!before(start_seq, tp->snd_nxt)) | |
1074 | return 0; | |
1075 | ||
1076 | /* In outstanding window? ...This is valid exit for DSACKs too. | |
1077 | * start_seq == snd_una is non-sensical (see comments above) | |
1078 | */ | |
1079 | if (after(start_seq, tp->snd_una)) | |
1080 | return 1; | |
1081 | ||
1082 | if (!is_dsack || !tp->undo_marker) | |
1083 | return 0; | |
1084 | ||
1085 | /* ...Then it's D-SACK, and must reside below snd_una completely */ | |
1086 | if (!after(end_seq, tp->snd_una)) | |
1087 | return 0; | |
1088 | ||
1089 | if (!before(start_seq, tp->undo_marker)) | |
1090 | return 1; | |
1091 | ||
1092 | /* Too old */ | |
1093 | if (!after(end_seq, tp->undo_marker)) | |
1094 | return 0; | |
1095 | ||
1096 | /* Undo_marker boundary crossing (overestimates a lot). Known already: | |
1097 | * start_seq < undo_marker and end_seq >= undo_marker. | |
1098 | */ | |
1099 | return !before(start_seq, end_seq - tp->max_window); | |
1100 | } | |
1101 | ||
1102 | ||
d06e021d DM |
1103 | static int tcp_check_dsack(struct tcp_sock *tp, struct sk_buff *ack_skb, |
1104 | struct tcp_sack_block_wire *sp, int num_sacks, | |
1105 | u32 prior_snd_una) | |
1106 | { | |
1107 | u32 start_seq_0 = ntohl(get_unaligned(&sp[0].start_seq)); | |
1108 | u32 end_seq_0 = ntohl(get_unaligned(&sp[0].end_seq)); | |
1109 | int dup_sack = 0; | |
1110 | ||
1111 | if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { | |
1112 | dup_sack = 1; | |
e60402d0 | 1113 | tcp_dsack_seen(tp); |
d06e021d DM |
1114 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV); |
1115 | } else if (num_sacks > 1) { | |
1116 | u32 end_seq_1 = ntohl(get_unaligned(&sp[1].end_seq)); | |
1117 | u32 start_seq_1 = ntohl(get_unaligned(&sp[1].start_seq)); | |
1118 | ||
1119 | if (!after(end_seq_0, end_seq_1) && | |
1120 | !before(start_seq_0, start_seq_1)) { | |
1121 | dup_sack = 1; | |
e60402d0 | 1122 | tcp_dsack_seen(tp); |
d06e021d DM |
1123 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV); |
1124 | } | |
1125 | } | |
1126 | ||
1127 | /* D-SACK for already forgotten data... Do dumb counting. */ | |
1128 | if (dup_sack && | |
1129 | !after(end_seq_0, prior_snd_una) && | |
1130 | after(end_seq_0, tp->undo_marker)) | |
1131 | tp->undo_retrans--; | |
1132 | ||
1133 | return dup_sack; | |
1134 | } | |
1135 | ||
1da177e4 LT |
1136 | static int |
1137 | tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una) | |
1138 | { | |
6687e988 | 1139 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 1140 | struct tcp_sock *tp = tcp_sk(sk); |
9c70220b ACM |
1141 | unsigned char *ptr = (skb_transport_header(ack_skb) + |
1142 | TCP_SKB_CB(ack_skb)->sacked); | |
269bd27e | 1143 | struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2); |
fda03fbb | 1144 | struct sk_buff *cached_skb; |
1da177e4 LT |
1145 | int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3; |
1146 | int reord = tp->packets_out; | |
1147 | int prior_fackets; | |
1148 | u32 lost_retrans = 0; | |
1149 | int flag = 0; | |
7769f406 | 1150 | int found_dup_sack = 0; |
fda03fbb | 1151 | int cached_fack_count; |
1da177e4 | 1152 | int i; |
fda03fbb | 1153 | int first_sack_index; |
1da177e4 | 1154 | |
d738cd8f | 1155 | if (!tp->sacked_out) { |
1da177e4 | 1156 | tp->fackets_out = 0; |
d738cd8f IJ |
1157 | tp->highest_sack = tp->snd_una; |
1158 | } | |
1da177e4 LT |
1159 | prior_fackets = tp->fackets_out; |
1160 | ||
d06e021d DM |
1161 | found_dup_sack = tcp_check_dsack(tp, ack_skb, sp, |
1162 | num_sacks, prior_snd_una); | |
1163 | if (found_dup_sack) | |
49ff4bb4 | 1164 | flag |= FLAG_DSACKING_ACK; |
6f74651a BE |
1165 | |
1166 | /* Eliminate too old ACKs, but take into | |
1167 | * account more or less fresh ones, they can | |
1168 | * contain valid SACK info. | |
1169 | */ | |
1170 | if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) | |
1171 | return 0; | |
1172 | ||
6a438bbe SH |
1173 | /* SACK fastpath: |
1174 | * if the only SACK change is the increase of the end_seq of | |
1175 | * the first block then only apply that SACK block | |
1176 | * and use retrans queue hinting otherwise slowpath */ | |
1177 | flag = 1; | |
6f74651a BE |
1178 | for (i = 0; i < num_sacks; i++) { |
1179 | __be32 start_seq = sp[i].start_seq; | |
1180 | __be32 end_seq = sp[i].end_seq; | |
6a438bbe | 1181 | |
6f74651a | 1182 | if (i == 0) { |
6a438bbe SH |
1183 | if (tp->recv_sack_cache[i].start_seq != start_seq) |
1184 | flag = 0; | |
1185 | } else { | |
1186 | if ((tp->recv_sack_cache[i].start_seq != start_seq) || | |
1187 | (tp->recv_sack_cache[i].end_seq != end_seq)) | |
1188 | flag = 0; | |
1189 | } | |
1190 | tp->recv_sack_cache[i].start_seq = start_seq; | |
1191 | tp->recv_sack_cache[i].end_seq = end_seq; | |
6a438bbe | 1192 | } |
8a3c3a97 BE |
1193 | /* Clear the rest of the cache sack blocks so they won't match mistakenly. */ |
1194 | for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) { | |
1195 | tp->recv_sack_cache[i].start_seq = 0; | |
1196 | tp->recv_sack_cache[i].end_seq = 0; | |
1197 | } | |
6a438bbe | 1198 | |
fda03fbb | 1199 | first_sack_index = 0; |
6a438bbe SH |
1200 | if (flag) |
1201 | num_sacks = 1; | |
1202 | else { | |
1203 | int j; | |
1204 | tp->fastpath_skb_hint = NULL; | |
1205 | ||
1206 | /* order SACK blocks to allow in order walk of the retrans queue */ | |
1207 | for (i = num_sacks-1; i > 0; i--) { | |
1208 | for (j = 0; j < i; j++){ | |
1209 | if (after(ntohl(sp[j].start_seq), | |
1210 | ntohl(sp[j+1].start_seq))){ | |
db3ccdac BE |
1211 | struct tcp_sack_block_wire tmp; |
1212 | ||
1213 | tmp = sp[j]; | |
1214 | sp[j] = sp[j+1]; | |
1215 | sp[j+1] = tmp; | |
fda03fbb BE |
1216 | |
1217 | /* Track where the first SACK block goes to */ | |
1218 | if (j == first_sack_index) | |
1219 | first_sack_index = j+1; | |
6a438bbe SH |
1220 | } |
1221 | ||
1222 | } | |
1223 | } | |
1224 | } | |
1225 | ||
1226 | /* clear flag as used for different purpose in following code */ | |
1227 | flag = 0; | |
1228 | ||
fda03fbb BE |
1229 | /* Use SACK fastpath hint if valid */ |
1230 | cached_skb = tp->fastpath_skb_hint; | |
1231 | cached_fack_count = tp->fastpath_cnt_hint; | |
1232 | if (!cached_skb) { | |
fe067e8a | 1233 | cached_skb = tcp_write_queue_head(sk); |
fda03fbb BE |
1234 | cached_fack_count = 0; |
1235 | } | |
1236 | ||
6a438bbe SH |
1237 | for (i=0; i<num_sacks; i++, sp++) { |
1238 | struct sk_buff *skb; | |
1239 | __u32 start_seq = ntohl(sp->start_seq); | |
1240 | __u32 end_seq = ntohl(sp->end_seq); | |
1241 | int fack_count; | |
7769f406 | 1242 | int dup_sack = (found_dup_sack && (i == first_sack_index)); |
6a438bbe | 1243 | |
18f02545 IJ |
1244 | if (!tcp_is_sackblock_valid(tp, dup_sack, start_seq, end_seq)) { |
1245 | if (dup_sack) { | |
1246 | if (!tp->undo_marker) | |
1247 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDNOUNDO); | |
1248 | else | |
1249 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKIGNOREDOLD); | |
1250 | } else | |
1251 | NET_INC_STATS_BH(LINUX_MIB_TCPSACKDISCARD); | |
5b3c9882 | 1252 | continue; |
18f02545 | 1253 | } |
5b3c9882 | 1254 | |
fda03fbb BE |
1255 | skb = cached_skb; |
1256 | fack_count = cached_fack_count; | |
1da177e4 LT |
1257 | |
1258 | /* Event "B" in the comment above. */ | |
1259 | if (after(end_seq, tp->high_seq)) | |
1260 | flag |= FLAG_DATA_LOST; | |
1261 | ||
fe067e8a | 1262 | tcp_for_write_queue_from(skb, sk) { |
6475be16 DM |
1263 | int in_sack, pcount; |
1264 | u8 sacked; | |
1da177e4 | 1265 | |
fe067e8a DM |
1266 | if (skb == tcp_send_head(sk)) |
1267 | break; | |
1268 | ||
fda03fbb BE |
1269 | cached_skb = skb; |
1270 | cached_fack_count = fack_count; | |
1271 | if (i == first_sack_index) { | |
1272 | tp->fastpath_skb_hint = skb; | |
1273 | tp->fastpath_cnt_hint = fack_count; | |
1274 | } | |
6a438bbe | 1275 | |
1da177e4 LT |
1276 | /* The retransmission queue is always in order, so |
1277 | * we can short-circuit the walk early. | |
1278 | */ | |
6475be16 | 1279 | if (!before(TCP_SKB_CB(skb)->seq, end_seq)) |
1da177e4 LT |
1280 | break; |
1281 | ||
3c05d92e HX |
1282 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && |
1283 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1284 | ||
6475be16 DM |
1285 | pcount = tcp_skb_pcount(skb); |
1286 | ||
3c05d92e HX |
1287 | if (pcount > 1 && !in_sack && |
1288 | after(TCP_SKB_CB(skb)->end_seq, start_seq)) { | |
6475be16 DM |
1289 | unsigned int pkt_len; |
1290 | ||
3c05d92e HX |
1291 | in_sack = !after(start_seq, |
1292 | TCP_SKB_CB(skb)->seq); | |
1293 | ||
1294 | if (!in_sack) | |
6475be16 DM |
1295 | pkt_len = (start_seq - |
1296 | TCP_SKB_CB(skb)->seq); | |
1297 | else | |
1298 | pkt_len = (end_seq - | |
1299 | TCP_SKB_CB(skb)->seq); | |
7967168c | 1300 | if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size)) |
6475be16 DM |
1301 | break; |
1302 | pcount = tcp_skb_pcount(skb); | |
1303 | } | |
1304 | ||
1305 | fack_count += pcount; | |
1da177e4 | 1306 | |
6475be16 DM |
1307 | sacked = TCP_SKB_CB(skb)->sacked; |
1308 | ||
1da177e4 LT |
1309 | /* Account D-SACK for retransmitted packet. */ |
1310 | if ((dup_sack && in_sack) && | |
1311 | (sacked & TCPCB_RETRANS) && | |
1312 | after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker)) | |
1313 | tp->undo_retrans--; | |
1314 | ||
1315 | /* The frame is ACKed. */ | |
1316 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) { | |
1317 | if (sacked&TCPCB_RETRANS) { | |
1318 | if ((dup_sack && in_sack) && | |
1319 | (sacked&TCPCB_SACKED_ACKED)) | |
1320 | reord = min(fack_count, reord); | |
1321 | } else { | |
1322 | /* If it was in a hole, we detected reordering. */ | |
1323 | if (fack_count < prior_fackets && | |
1324 | !(sacked&TCPCB_SACKED_ACKED)) | |
1325 | reord = min(fack_count, reord); | |
1326 | } | |
1327 | ||
1328 | /* Nothing to do; acked frame is about to be dropped. */ | |
1329 | continue; | |
1330 | } | |
1331 | ||
1332 | if ((sacked&TCPCB_SACKED_RETRANS) && | |
1333 | after(end_seq, TCP_SKB_CB(skb)->ack_seq) && | |
1334 | (!lost_retrans || after(end_seq, lost_retrans))) | |
1335 | lost_retrans = end_seq; | |
1336 | ||
1337 | if (!in_sack) | |
1338 | continue; | |
1339 | ||
1340 | if (!(sacked&TCPCB_SACKED_ACKED)) { | |
1341 | if (sacked & TCPCB_SACKED_RETRANS) { | |
1342 | /* If the segment is not tagged as lost, | |
1343 | * we do not clear RETRANS, believing | |
1344 | * that retransmission is still in flight. | |
1345 | */ | |
1346 | if (sacked & TCPCB_LOST) { | |
1347 | TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); | |
1348 | tp->lost_out -= tcp_skb_pcount(skb); | |
1349 | tp->retrans_out -= tcp_skb_pcount(skb); | |
6a438bbe SH |
1350 | |
1351 | /* clear lost hint */ | |
1352 | tp->retransmit_skb_hint = NULL; | |
1da177e4 LT |
1353 | } |
1354 | } else { | |
1355 | /* New sack for not retransmitted frame, | |
1356 | * which was in hole. It is reordering. | |
1357 | */ | |
1358 | if (!(sacked & TCPCB_RETRANS) && | |
1359 | fack_count < prior_fackets) | |
1360 | reord = min(fack_count, reord); | |
1361 | ||
1362 | if (sacked & TCPCB_LOST) { | |
1363 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; | |
1364 | tp->lost_out -= tcp_skb_pcount(skb); | |
6a438bbe SH |
1365 | |
1366 | /* clear lost hint */ | |
1367 | tp->retransmit_skb_hint = NULL; | |
1da177e4 | 1368 | } |
4dc2665e IJ |
1369 | /* SACK enhanced F-RTO detection. |
1370 | * Set flag if and only if non-rexmitted | |
1371 | * segments below frto_highmark are | |
1372 | * SACKed (RFC4138; Appendix B). | |
1373 | * Clearing correct due to in-order walk | |
1374 | */ | |
1375 | if (after(end_seq, tp->frto_highmark)) { | |
1376 | flag &= ~FLAG_ONLY_ORIG_SACKED; | |
1377 | } else { | |
1378 | if (!(sacked & TCPCB_RETRANS)) | |
1379 | flag |= FLAG_ONLY_ORIG_SACKED; | |
1380 | } | |
1da177e4 LT |
1381 | } |
1382 | ||
1383 | TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED; | |
1384 | flag |= FLAG_DATA_SACKED; | |
1385 | tp->sacked_out += tcp_skb_pcount(skb); | |
1386 | ||
1387 | if (fack_count > tp->fackets_out) | |
1388 | tp->fackets_out = fack_count; | |
d738cd8f IJ |
1389 | |
1390 | if (after(TCP_SKB_CB(skb)->seq, | |
1391 | tp->highest_sack)) | |
1392 | tp->highest_sack = TCP_SKB_CB(skb)->seq; | |
1da177e4 LT |
1393 | } else { |
1394 | if (dup_sack && (sacked&TCPCB_RETRANS)) | |
1395 | reord = min(fack_count, reord); | |
1396 | } | |
1397 | ||
1398 | /* D-SACK. We can detect redundant retransmission | |
1399 | * in S|R and plain R frames and clear it. | |
1400 | * undo_retrans is decreased above, L|R frames | |
1401 | * are accounted above as well. | |
1402 | */ | |
1403 | if (dup_sack && | |
1404 | (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) { | |
1405 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; | |
1406 | tp->retrans_out -= tcp_skb_pcount(skb); | |
6a438bbe | 1407 | tp->retransmit_skb_hint = NULL; |
1da177e4 LT |
1408 | } |
1409 | } | |
1410 | } | |
1411 | ||
1412 | /* Check for lost retransmit. This superb idea is | |
1413 | * borrowed from "ratehalving". Event "C". | |
1414 | * Later note: FACK people cheated me again 8), | |
1415 | * we have to account for reordering! Ugly, | |
1416 | * but should help. | |
1417 | */ | |
6687e988 | 1418 | if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) { |
1da177e4 LT |
1419 | struct sk_buff *skb; |
1420 | ||
fe067e8a DM |
1421 | tcp_for_write_queue(skb, sk) { |
1422 | if (skb == tcp_send_head(sk)) | |
1423 | break; | |
1da177e4 LT |
1424 | if (after(TCP_SKB_CB(skb)->seq, lost_retrans)) |
1425 | break; | |
1426 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1427 | continue; | |
1428 | if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) && | |
1429 | after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) && | |
e60402d0 | 1430 | (tcp_is_fack(tp) || |
1da177e4 LT |
1431 | !before(lost_retrans, |
1432 | TCP_SKB_CB(skb)->ack_seq + tp->reordering * | |
c1b4a7e6 | 1433 | tp->mss_cache))) { |
1da177e4 LT |
1434 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; |
1435 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1436 | ||
6a438bbe SH |
1437 | /* clear lost hint */ |
1438 | tp->retransmit_skb_hint = NULL; | |
1439 | ||
1da177e4 LT |
1440 | if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) { |
1441 | tp->lost_out += tcp_skb_pcount(skb); | |
1442 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1443 | flag |= FLAG_DATA_SACKED; | |
1444 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT); | |
1445 | } | |
1446 | } | |
1447 | } | |
1448 | } | |
1449 | ||
86426c22 IJ |
1450 | tcp_verify_left_out(tp); |
1451 | ||
288035f9 | 1452 | if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss && |
c5e7af0d | 1453 | (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark))) |
6687e988 | 1454 | tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0); |
1da177e4 LT |
1455 | |
1456 | #if FASTRETRANS_DEBUG > 0 | |
1457 | BUG_TRAP((int)tp->sacked_out >= 0); | |
1458 | BUG_TRAP((int)tp->lost_out >= 0); | |
1459 | BUG_TRAP((int)tp->retrans_out >= 0); | |
1460 | BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0); | |
1461 | #endif | |
1462 | return flag; | |
1463 | } | |
1464 | ||
575ee714 IJ |
1465 | /* F-RTO can only be used if TCP has never retransmitted anything other than |
1466 | * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here) | |
30935cf4 | 1467 | */ |
4ddf6676 IJ |
1468 | static void tcp_check_reno_reordering(struct sock *sk, const int addend) |
1469 | { | |
1470 | struct tcp_sock *tp = tcp_sk(sk); | |
1471 | u32 holes; | |
1472 | ||
1473 | holes = max(tp->lost_out, 1U); | |
1474 | holes = min(holes, tp->packets_out); | |
1475 | ||
1476 | if ((tp->sacked_out + holes) > tp->packets_out) { | |
1477 | tp->sacked_out = tp->packets_out - holes; | |
1478 | tcp_update_reordering(sk, tp->packets_out + addend, 0); | |
1479 | } | |
1480 | } | |
1481 | ||
1482 | /* Emulate SACKs for SACKless connection: account for a new dupack. */ | |
1483 | ||
1484 | static void tcp_add_reno_sack(struct sock *sk) | |
1485 | { | |
1486 | struct tcp_sock *tp = tcp_sk(sk); | |
1487 | tp->sacked_out++; | |
1488 | tcp_check_reno_reordering(sk, 0); | |
005903bc | 1489 | tcp_verify_left_out(tp); |
4ddf6676 IJ |
1490 | } |
1491 | ||
1492 | /* Account for ACK, ACKing some data in Reno Recovery phase. */ | |
1493 | ||
1494 | static void tcp_remove_reno_sacks(struct sock *sk, int acked) | |
1495 | { | |
1496 | struct tcp_sock *tp = tcp_sk(sk); | |
1497 | ||
1498 | if (acked > 0) { | |
1499 | /* One ACK acked hole. The rest eat duplicate ACKs. */ | |
1500 | if (acked-1 >= tp->sacked_out) | |
1501 | tp->sacked_out = 0; | |
1502 | else | |
1503 | tp->sacked_out -= acked-1; | |
1504 | } | |
1505 | tcp_check_reno_reordering(sk, acked); | |
005903bc | 1506 | tcp_verify_left_out(tp); |
4ddf6676 IJ |
1507 | } |
1508 | ||
1509 | static inline void tcp_reset_reno_sack(struct tcp_sock *tp) | |
1510 | { | |
1511 | tp->sacked_out = 0; | |
4ddf6676 IJ |
1512 | } |
1513 | ||
46d0de4e | 1514 | int tcp_use_frto(struct sock *sk) |
bdaae17d IJ |
1515 | { |
1516 | const struct tcp_sock *tp = tcp_sk(sk); | |
46d0de4e IJ |
1517 | struct sk_buff *skb; |
1518 | ||
575ee714 | 1519 | if (!sysctl_tcp_frto) |
46d0de4e | 1520 | return 0; |
bdaae17d | 1521 | |
4dc2665e IJ |
1522 | if (IsSackFrto()) |
1523 | return 1; | |
1524 | ||
46d0de4e IJ |
1525 | /* Avoid expensive walking of rexmit queue if possible */ |
1526 | if (tp->retrans_out > 1) | |
1527 | return 0; | |
1528 | ||
fe067e8a DM |
1529 | skb = tcp_write_queue_head(sk); |
1530 | skb = tcp_write_queue_next(sk, skb); /* Skips head */ | |
1531 | tcp_for_write_queue_from(skb, sk) { | |
1532 | if (skb == tcp_send_head(sk)) | |
1533 | break; | |
46d0de4e IJ |
1534 | if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS) |
1535 | return 0; | |
1536 | /* Short-circuit when first non-SACKed skb has been checked */ | |
1537 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) | |
1538 | break; | |
1539 | } | |
1540 | return 1; | |
bdaae17d IJ |
1541 | } |
1542 | ||
30935cf4 IJ |
1543 | /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO |
1544 | * recovery a bit and use heuristics in tcp_process_frto() to detect if | |
d1a54c6a IJ |
1545 | * the RTO was spurious. Only clear SACKED_RETRANS of the head here to |
1546 | * keep retrans_out counting accurate (with SACK F-RTO, other than head | |
1547 | * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS | |
1548 | * bits are handled if the Loss state is really to be entered (in | |
1549 | * tcp_enter_frto_loss). | |
7487c48c IJ |
1550 | * |
1551 | * Do like tcp_enter_loss() would; when RTO expires the second time it | |
1552 | * does: | |
1553 | * "Reduce ssthresh if it has not yet been made inside this window." | |
1da177e4 LT |
1554 | */ |
1555 | void tcp_enter_frto(struct sock *sk) | |
1556 | { | |
6687e988 | 1557 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
1558 | struct tcp_sock *tp = tcp_sk(sk); |
1559 | struct sk_buff *skb; | |
1560 | ||
7487c48c | 1561 | if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) || |
e905a9ed | 1562 | tp->snd_una == tp->high_seq || |
7487c48c IJ |
1563 | ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) && |
1564 | !icsk->icsk_retransmits)) { | |
6687e988 | 1565 | tp->prior_ssthresh = tcp_current_ssthresh(sk); |
66e93e45 IJ |
1566 | /* Our state is too optimistic in ssthresh() call because cwnd |
1567 | * is not reduced until tcp_enter_frto_loss() when previous FRTO | |
1568 | * recovery has not yet completed. Pattern would be this: RTO, | |
1569 | * Cumulative ACK, RTO (2xRTO for the same segment does not end | |
1570 | * up here twice). | |
1571 | * RFC4138 should be more specific on what to do, even though | |
1572 | * RTO is quite unlikely to occur after the first Cumulative ACK | |
1573 | * due to back-off and complexity of triggering events ... | |
1574 | */ | |
1575 | if (tp->frto_counter) { | |
1576 | u32 stored_cwnd; | |
1577 | stored_cwnd = tp->snd_cwnd; | |
1578 | tp->snd_cwnd = 2; | |
1579 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
1580 | tp->snd_cwnd = stored_cwnd; | |
1581 | } else { | |
1582 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
1583 | } | |
1584 | /* ... in theory, cong.control module could do "any tricks" in | |
1585 | * ssthresh(), which means that ca_state, lost bits and lost_out | |
1586 | * counter would have to be faked before the call occurs. We | |
1587 | * consider that too expensive, unlikely and hacky, so modules | |
1588 | * using these in ssthresh() must deal these incompatibility | |
1589 | * issues if they receives CA_EVENT_FRTO and frto_counter != 0 | |
1590 | */ | |
6687e988 | 1591 | tcp_ca_event(sk, CA_EVENT_FRTO); |
1da177e4 LT |
1592 | } |
1593 | ||
1da177e4 LT |
1594 | tp->undo_marker = tp->snd_una; |
1595 | tp->undo_retrans = 0; | |
1596 | ||
fe067e8a | 1597 | skb = tcp_write_queue_head(sk); |
009a2e3e IJ |
1598 | if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) |
1599 | tp->undo_marker = 0; | |
d1a54c6a | 1600 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
522e7548 | 1601 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; |
d1a54c6a | 1602 | tp->retrans_out -= tcp_skb_pcount(skb); |
1da177e4 | 1603 | } |
005903bc | 1604 | tcp_verify_left_out(tp); |
1da177e4 | 1605 | |
4dc2665e IJ |
1606 | /* Earlier loss recovery underway (see RFC4138; Appendix B). |
1607 | * The last condition is necessary at least in tp->frto_counter case. | |
1608 | */ | |
1609 | if (IsSackFrto() && (tp->frto_counter || | |
1610 | ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) && | |
1611 | after(tp->high_seq, tp->snd_una)) { | |
1612 | tp->frto_highmark = tp->high_seq; | |
1613 | } else { | |
1614 | tp->frto_highmark = tp->snd_nxt; | |
1615 | } | |
7b0eb22b IJ |
1616 | tcp_set_ca_state(sk, TCP_CA_Disorder); |
1617 | tp->high_seq = tp->snd_nxt; | |
7487c48c | 1618 | tp->frto_counter = 1; |
1da177e4 LT |
1619 | } |
1620 | ||
1621 | /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, | |
1622 | * which indicates that we should follow the traditional RTO recovery, | |
1623 | * i.e. mark everything lost and do go-back-N retransmission. | |
1624 | */ | |
d1a54c6a | 1625 | static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag) |
1da177e4 LT |
1626 | { |
1627 | struct tcp_sock *tp = tcp_sk(sk); | |
1628 | struct sk_buff *skb; | |
1da177e4 | 1629 | |
1da177e4 | 1630 | tp->lost_out = 0; |
d1a54c6a | 1631 | tp->retrans_out = 0; |
e60402d0 | 1632 | if (tcp_is_reno(tp)) |
9bff40fd | 1633 | tcp_reset_reno_sack(tp); |
1da177e4 | 1634 | |
fe067e8a DM |
1635 | tcp_for_write_queue(skb, sk) { |
1636 | if (skb == tcp_send_head(sk)) | |
1637 | break; | |
d1a54c6a IJ |
1638 | /* |
1639 | * Count the retransmission made on RTO correctly (only when | |
1640 | * waiting for the first ACK and did not get it)... | |
1641 | */ | |
1642 | if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) { | |
0a9f2a46 IJ |
1643 | /* For some reason this R-bit might get cleared? */ |
1644 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
1645 | tp->retrans_out += tcp_skb_pcount(skb); | |
d1a54c6a IJ |
1646 | /* ...enter this if branch just for the first segment */ |
1647 | flag |= FLAG_DATA_ACKED; | |
1648 | } else { | |
009a2e3e IJ |
1649 | if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) |
1650 | tp->undo_marker = 0; | |
d1a54c6a IJ |
1651 | TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); |
1652 | } | |
1da177e4 | 1653 | |
9bff40fd IJ |
1654 | /* Don't lost mark skbs that were fwd transmitted after RTO */ |
1655 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) && | |
1656 | !after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) { | |
1657 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1658 | tp->lost_out += tcp_skb_pcount(skb); | |
1da177e4 LT |
1659 | } |
1660 | } | |
005903bc | 1661 | tcp_verify_left_out(tp); |
1da177e4 | 1662 | |
95c4922b | 1663 | tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments; |
1da177e4 LT |
1664 | tp->snd_cwnd_cnt = 0; |
1665 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1da177e4 LT |
1666 | tp->frto_counter = 0; |
1667 | ||
1668 | tp->reordering = min_t(unsigned int, tp->reordering, | |
1669 | sysctl_tcp_reordering); | |
6687e988 | 1670 | tcp_set_ca_state(sk, TCP_CA_Loss); |
1da177e4 LT |
1671 | tp->high_seq = tp->frto_highmark; |
1672 | TCP_ECN_queue_cwr(tp); | |
6a438bbe | 1673 | |
5af4ec23 | 1674 | tcp_clear_all_retrans_hints(tp); |
1da177e4 LT |
1675 | } |
1676 | ||
1677 | void tcp_clear_retrans(struct tcp_sock *tp) | |
1678 | { | |
1da177e4 LT |
1679 | tp->retrans_out = 0; |
1680 | ||
1681 | tp->fackets_out = 0; | |
1682 | tp->sacked_out = 0; | |
1683 | tp->lost_out = 0; | |
1684 | ||
1685 | tp->undo_marker = 0; | |
1686 | tp->undo_retrans = 0; | |
1687 | } | |
1688 | ||
1689 | /* Enter Loss state. If "how" is not zero, forget all SACK information | |
1690 | * and reset tags completely, otherwise preserve SACKs. If receiver | |
1691 | * dropped its ofo queue, we will know this due to reneging detection. | |
1692 | */ | |
1693 | void tcp_enter_loss(struct sock *sk, int how) | |
1694 | { | |
6687e988 | 1695 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
1696 | struct tcp_sock *tp = tcp_sk(sk); |
1697 | struct sk_buff *skb; | |
1698 | int cnt = 0; | |
1699 | ||
1700 | /* Reduce ssthresh if it has not yet been made inside this window. */ | |
6687e988 ACM |
1701 | if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || |
1702 | (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { | |
1703 | tp->prior_ssthresh = tcp_current_ssthresh(sk); | |
1704 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
1705 | tcp_ca_event(sk, CA_EVENT_LOSS); | |
1da177e4 LT |
1706 | } |
1707 | tp->snd_cwnd = 1; | |
1708 | tp->snd_cwnd_cnt = 0; | |
1709 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1710 | ||
9772efb9 | 1711 | tp->bytes_acked = 0; |
1da177e4 LT |
1712 | tcp_clear_retrans(tp); |
1713 | ||
1714 | /* Push undo marker, if it was plain RTO and nothing | |
1715 | * was retransmitted. */ | |
1716 | if (!how) | |
1717 | tp->undo_marker = tp->snd_una; | |
1718 | ||
fe067e8a DM |
1719 | tcp_for_write_queue(skb, sk) { |
1720 | if (skb == tcp_send_head(sk)) | |
1721 | break; | |
1da177e4 LT |
1722 | cnt += tcp_skb_pcount(skb); |
1723 | if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS) | |
1724 | tp->undo_marker = 0; | |
1725 | TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; | |
1726 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { | |
1727 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; | |
1728 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1729 | tp->lost_out += tcp_skb_pcount(skb); | |
1730 | } else { | |
1731 | tp->sacked_out += tcp_skb_pcount(skb); | |
1732 | tp->fackets_out = cnt; | |
1733 | } | |
1734 | } | |
005903bc | 1735 | tcp_verify_left_out(tp); |
1da177e4 LT |
1736 | |
1737 | tp->reordering = min_t(unsigned int, tp->reordering, | |
1738 | sysctl_tcp_reordering); | |
6687e988 | 1739 | tcp_set_ca_state(sk, TCP_CA_Loss); |
1da177e4 LT |
1740 | tp->high_seq = tp->snd_nxt; |
1741 | TCP_ECN_queue_cwr(tp); | |
580e572a IJ |
1742 | /* Abort FRTO algorithm if one is in progress */ |
1743 | tp->frto_counter = 0; | |
6a438bbe | 1744 | |
5af4ec23 | 1745 | tcp_clear_all_retrans_hints(tp); |
1da177e4 LT |
1746 | } |
1747 | ||
463c84b9 | 1748 | static int tcp_check_sack_reneging(struct sock *sk) |
1da177e4 LT |
1749 | { |
1750 | struct sk_buff *skb; | |
1751 | ||
1752 | /* If ACK arrived pointing to a remembered SACK, | |
1753 | * it means that our remembered SACKs do not reflect | |
1754 | * real state of receiver i.e. | |
1755 | * receiver _host_ is heavily congested (or buggy). | |
1756 | * Do processing similar to RTO timeout. | |
1757 | */ | |
fe067e8a | 1758 | if ((skb = tcp_write_queue_head(sk)) != NULL && |
1da177e4 | 1759 | (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { |
6687e988 | 1760 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
1761 | NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING); |
1762 | ||
1763 | tcp_enter_loss(sk, 1); | |
6687e988 | 1764 | icsk->icsk_retransmits++; |
fe067e8a | 1765 | tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); |
463c84b9 | 1766 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
6687e988 | 1767 | icsk->icsk_rto, TCP_RTO_MAX); |
1da177e4 LT |
1768 | return 1; |
1769 | } | |
1770 | return 0; | |
1771 | } | |
1772 | ||
1773 | static inline int tcp_fackets_out(struct tcp_sock *tp) | |
1774 | { | |
e60402d0 | 1775 | return tcp_is_reno(tp) ? tp->sacked_out+1 : tp->fackets_out; |
1da177e4 LT |
1776 | } |
1777 | ||
463c84b9 | 1778 | static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 1779 | { |
463c84b9 | 1780 | return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto); |
1da177e4 LT |
1781 | } |
1782 | ||
9e412ba7 | 1783 | static inline int tcp_head_timedout(struct sock *sk) |
1da177e4 | 1784 | { |
9e412ba7 IJ |
1785 | struct tcp_sock *tp = tcp_sk(sk); |
1786 | ||
1da177e4 | 1787 | return tp->packets_out && |
fe067e8a | 1788 | tcp_skb_timedout(sk, tcp_write_queue_head(sk)); |
1da177e4 LT |
1789 | } |
1790 | ||
1791 | /* Linux NewReno/SACK/FACK/ECN state machine. | |
1792 | * -------------------------------------- | |
1793 | * | |
1794 | * "Open" Normal state, no dubious events, fast path. | |
1795 | * "Disorder" In all the respects it is "Open", | |
1796 | * but requires a bit more attention. It is entered when | |
1797 | * we see some SACKs or dupacks. It is split of "Open" | |
1798 | * mainly to move some processing from fast path to slow one. | |
1799 | * "CWR" CWND was reduced due to some Congestion Notification event. | |
1800 | * It can be ECN, ICMP source quench, local device congestion. | |
1801 | * "Recovery" CWND was reduced, we are fast-retransmitting. | |
1802 | * "Loss" CWND was reduced due to RTO timeout or SACK reneging. | |
1803 | * | |
1804 | * tcp_fastretrans_alert() is entered: | |
1805 | * - each incoming ACK, if state is not "Open" | |
1806 | * - when arrived ACK is unusual, namely: | |
1807 | * * SACK | |
1808 | * * Duplicate ACK. | |
1809 | * * ECN ECE. | |
1810 | * | |
1811 | * Counting packets in flight is pretty simple. | |
1812 | * | |
1813 | * in_flight = packets_out - left_out + retrans_out | |
1814 | * | |
1815 | * packets_out is SND.NXT-SND.UNA counted in packets. | |
1816 | * | |
1817 | * retrans_out is number of retransmitted segments. | |
1818 | * | |
1819 | * left_out is number of segments left network, but not ACKed yet. | |
1820 | * | |
1821 | * left_out = sacked_out + lost_out | |
1822 | * | |
1823 | * sacked_out: Packets, which arrived to receiver out of order | |
1824 | * and hence not ACKed. With SACKs this number is simply | |
1825 | * amount of SACKed data. Even without SACKs | |
1826 | * it is easy to give pretty reliable estimate of this number, | |
1827 | * counting duplicate ACKs. | |
1828 | * | |
1829 | * lost_out: Packets lost by network. TCP has no explicit | |
1830 | * "loss notification" feedback from network (for now). | |
1831 | * It means that this number can be only _guessed_. | |
1832 | * Actually, it is the heuristics to predict lossage that | |
1833 | * distinguishes different algorithms. | |
1834 | * | |
1835 | * F.e. after RTO, when all the queue is considered as lost, | |
1836 | * lost_out = packets_out and in_flight = retrans_out. | |
1837 | * | |
1838 | * Essentially, we have now two algorithms counting | |
1839 | * lost packets. | |
1840 | * | |
1841 | * FACK: It is the simplest heuristics. As soon as we decided | |
1842 | * that something is lost, we decide that _all_ not SACKed | |
1843 | * packets until the most forward SACK are lost. I.e. | |
1844 | * lost_out = fackets_out - sacked_out and left_out = fackets_out. | |
1845 | * It is absolutely correct estimate, if network does not reorder | |
1846 | * packets. And it loses any connection to reality when reordering | |
1847 | * takes place. We use FACK by default until reordering | |
1848 | * is suspected on the path to this destination. | |
1849 | * | |
1850 | * NewReno: when Recovery is entered, we assume that one segment | |
1851 | * is lost (classic Reno). While we are in Recovery and | |
1852 | * a partial ACK arrives, we assume that one more packet | |
1853 | * is lost (NewReno). This heuristics are the same in NewReno | |
1854 | * and SACK. | |
1855 | * | |
1856 | * Imagine, that's all! Forget about all this shamanism about CWND inflation | |
1857 | * deflation etc. CWND is real congestion window, never inflated, changes | |
1858 | * only according to classic VJ rules. | |
1859 | * | |
1860 | * Really tricky (and requiring careful tuning) part of algorithm | |
1861 | * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). | |
1862 | * The first determines the moment _when_ we should reduce CWND and, | |
1863 | * hence, slow down forward transmission. In fact, it determines the moment | |
1864 | * when we decide that hole is caused by loss, rather than by a reorder. | |
1865 | * | |
1866 | * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill | |
1867 | * holes, caused by lost packets. | |
1868 | * | |
1869 | * And the most logically complicated part of algorithm is undo | |
1870 | * heuristics. We detect false retransmits due to both too early | |
1871 | * fast retransmit (reordering) and underestimated RTO, analyzing | |
1872 | * timestamps and D-SACKs. When we detect that some segments were | |
1873 | * retransmitted by mistake and CWND reduction was wrong, we undo | |
1874 | * window reduction and abort recovery phase. This logic is hidden | |
1875 | * inside several functions named tcp_try_undo_<something>. | |
1876 | */ | |
1877 | ||
1878 | /* This function decides, when we should leave Disordered state | |
1879 | * and enter Recovery phase, reducing congestion window. | |
1880 | * | |
1881 | * Main question: may we further continue forward transmission | |
1882 | * with the same cwnd? | |
1883 | */ | |
9e412ba7 | 1884 | static int tcp_time_to_recover(struct sock *sk) |
1da177e4 | 1885 | { |
9e412ba7 | 1886 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
1887 | __u32 packets_out; |
1888 | ||
52c63f1e IJ |
1889 | /* Do not perform any recovery during FRTO algorithm */ |
1890 | if (tp->frto_counter) | |
1891 | return 0; | |
1892 | ||
1da177e4 LT |
1893 | /* Trick#1: The loss is proven. */ |
1894 | if (tp->lost_out) | |
1895 | return 1; | |
1896 | ||
1897 | /* Not-A-Trick#2 : Classic rule... */ | |
1898 | if (tcp_fackets_out(tp) > tp->reordering) | |
1899 | return 1; | |
1900 | ||
1901 | /* Trick#3 : when we use RFC2988 timer restart, fast | |
1902 | * retransmit can be triggered by timeout of queue head. | |
1903 | */ | |
9e412ba7 | 1904 | if (tcp_head_timedout(sk)) |
1da177e4 LT |
1905 | return 1; |
1906 | ||
1907 | /* Trick#4: It is still not OK... But will it be useful to delay | |
1908 | * recovery more? | |
1909 | */ | |
1910 | packets_out = tp->packets_out; | |
1911 | if (packets_out <= tp->reordering && | |
1912 | tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && | |
9e412ba7 | 1913 | !tcp_may_send_now(sk)) { |
1da177e4 LT |
1914 | /* We have nothing to send. This connection is limited |
1915 | * either by receiver window or by application. | |
1916 | */ | |
1917 | return 1; | |
1918 | } | |
1919 | ||
1920 | return 0; | |
1921 | } | |
1922 | ||
d8f4f223 IJ |
1923 | /* RFC: This is from the original, I doubt that this is necessary at all: |
1924 | * clear xmit_retrans hint if seq of this skb is beyond hint. How could we | |
1925 | * retransmitted past LOST markings in the first place? I'm not fully sure | |
1926 | * about undo and end of connection cases, which can cause R without L? | |
1927 | */ | |
1928 | static void tcp_verify_retransmit_hint(struct tcp_sock *tp, | |
1929 | struct sk_buff *skb) | |
1930 | { | |
1931 | if ((tp->retransmit_skb_hint != NULL) && | |
1932 | before(TCP_SKB_CB(skb)->seq, | |
1933 | TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) | |
19b2b486 | 1934 | tp->retransmit_skb_hint = NULL; |
d8f4f223 IJ |
1935 | } |
1936 | ||
1da177e4 | 1937 | /* Mark head of queue up as lost. */ |
9e412ba7 | 1938 | static void tcp_mark_head_lost(struct sock *sk, |
1da177e4 LT |
1939 | int packets, u32 high_seq) |
1940 | { | |
9e412ba7 | 1941 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 1942 | struct sk_buff *skb; |
6a438bbe | 1943 | int cnt; |
1da177e4 | 1944 | |
6a438bbe SH |
1945 | BUG_TRAP(packets <= tp->packets_out); |
1946 | if (tp->lost_skb_hint) { | |
1947 | skb = tp->lost_skb_hint; | |
1948 | cnt = tp->lost_cnt_hint; | |
1949 | } else { | |
fe067e8a | 1950 | skb = tcp_write_queue_head(sk); |
6a438bbe SH |
1951 | cnt = 0; |
1952 | } | |
1da177e4 | 1953 | |
fe067e8a DM |
1954 | tcp_for_write_queue_from(skb, sk) { |
1955 | if (skb == tcp_send_head(sk)) | |
1956 | break; | |
6a438bbe SH |
1957 | /* TODO: do this better */ |
1958 | /* this is not the most efficient way to do this... */ | |
1959 | tp->lost_skb_hint = skb; | |
1960 | tp->lost_cnt_hint = cnt; | |
1961 | cnt += tcp_skb_pcount(skb); | |
1962 | if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq)) | |
1da177e4 LT |
1963 | break; |
1964 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) { | |
1965 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1966 | tp->lost_out += tcp_skb_pcount(skb); | |
d8f4f223 | 1967 | tcp_verify_retransmit_hint(tp, skb); |
1da177e4 LT |
1968 | } |
1969 | } | |
005903bc | 1970 | tcp_verify_left_out(tp); |
1da177e4 LT |
1971 | } |
1972 | ||
1973 | /* Account newly detected lost packet(s) */ | |
1974 | ||
9e412ba7 | 1975 | static void tcp_update_scoreboard(struct sock *sk) |
1da177e4 | 1976 | { |
9e412ba7 IJ |
1977 | struct tcp_sock *tp = tcp_sk(sk); |
1978 | ||
e60402d0 | 1979 | if (tcp_is_fack(tp)) { |
1da177e4 LT |
1980 | int lost = tp->fackets_out - tp->reordering; |
1981 | if (lost <= 0) | |
1982 | lost = 1; | |
9e412ba7 | 1983 | tcp_mark_head_lost(sk, lost, tp->high_seq); |
1da177e4 | 1984 | } else { |
9e412ba7 | 1985 | tcp_mark_head_lost(sk, 1, tp->high_seq); |
1da177e4 LT |
1986 | } |
1987 | ||
1988 | /* New heuristics: it is possible only after we switched | |
1989 | * to restart timer each time when something is ACKed. | |
1990 | * Hence, we can detect timed out packets during fast | |
1991 | * retransmit without falling to slow start. | |
1992 | */ | |
e60402d0 | 1993 | if (!tcp_is_reno(tp) && tcp_head_timedout(sk)) { |
1da177e4 LT |
1994 | struct sk_buff *skb; |
1995 | ||
6a438bbe | 1996 | skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint |
fe067e8a | 1997 | : tcp_write_queue_head(sk); |
6a438bbe | 1998 | |
fe067e8a DM |
1999 | tcp_for_write_queue_from(skb, sk) { |
2000 | if (skb == tcp_send_head(sk)) | |
2001 | break; | |
6a438bbe SH |
2002 | if (!tcp_skb_timedout(sk, skb)) |
2003 | break; | |
2004 | ||
2005 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) { | |
1da177e4 LT |
2006 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; |
2007 | tp->lost_out += tcp_skb_pcount(skb); | |
d8f4f223 | 2008 | tcp_verify_retransmit_hint(tp, skb); |
1da177e4 LT |
2009 | } |
2010 | } | |
6a438bbe SH |
2011 | |
2012 | tp->scoreboard_skb_hint = skb; | |
2013 | ||
005903bc | 2014 | tcp_verify_left_out(tp); |
1da177e4 LT |
2015 | } |
2016 | } | |
2017 | ||
2018 | /* CWND moderation, preventing bursts due to too big ACKs | |
2019 | * in dubious situations. | |
2020 | */ | |
2021 | static inline void tcp_moderate_cwnd(struct tcp_sock *tp) | |
2022 | { | |
2023 | tp->snd_cwnd = min(tp->snd_cwnd, | |
2024 | tcp_packets_in_flight(tp)+tcp_max_burst(tp)); | |
2025 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
2026 | } | |
2027 | ||
72dc5b92 SH |
2028 | /* Lower bound on congestion window is slow start threshold |
2029 | * unless congestion avoidance choice decides to overide it. | |
2030 | */ | |
2031 | static inline u32 tcp_cwnd_min(const struct sock *sk) | |
2032 | { | |
2033 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; | |
2034 | ||
2035 | return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh; | |
2036 | } | |
2037 | ||
1da177e4 | 2038 | /* Decrease cwnd each second ack. */ |
1e757f99 | 2039 | static void tcp_cwnd_down(struct sock *sk, int flag) |
1da177e4 | 2040 | { |
6687e988 | 2041 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 2042 | int decr = tp->snd_cwnd_cnt + 1; |
1da177e4 | 2043 | |
49ff4bb4 | 2044 | if ((flag&(FLAG_ANY_PROGRESS|FLAG_DSACKING_ACK)) || |
e60402d0 | 2045 | (tcp_is_reno(tp) && !(flag&FLAG_NOT_DUP))) { |
1e757f99 IJ |
2046 | tp->snd_cwnd_cnt = decr&1; |
2047 | decr >>= 1; | |
1da177e4 | 2048 | |
1e757f99 IJ |
2049 | if (decr && tp->snd_cwnd > tcp_cwnd_min(sk)) |
2050 | tp->snd_cwnd -= decr; | |
1da177e4 | 2051 | |
1e757f99 IJ |
2052 | tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1); |
2053 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
2054 | } | |
1da177e4 LT |
2055 | } |
2056 | ||
2057 | /* Nothing was retransmitted or returned timestamp is less | |
2058 | * than timestamp of the first retransmission. | |
2059 | */ | |
2060 | static inline int tcp_packet_delayed(struct tcp_sock *tp) | |
2061 | { | |
2062 | return !tp->retrans_stamp || | |
2063 | (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && | |
2064 | (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0); | |
2065 | } | |
2066 | ||
2067 | /* Undo procedures. */ | |
2068 | ||
2069 | #if FASTRETRANS_DEBUG > 1 | |
9e412ba7 | 2070 | static void DBGUNDO(struct sock *sk, const char *msg) |
1da177e4 | 2071 | { |
9e412ba7 | 2072 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 2073 | struct inet_sock *inet = inet_sk(sk); |
9e412ba7 | 2074 | |
1da177e4 LT |
2075 | printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n", |
2076 | msg, | |
2077 | NIPQUAD(inet->daddr), ntohs(inet->dport), | |
83ae4088 | 2078 | tp->snd_cwnd, tcp_left_out(tp), |
1da177e4 LT |
2079 | tp->snd_ssthresh, tp->prior_ssthresh, |
2080 | tp->packets_out); | |
2081 | } | |
2082 | #else | |
2083 | #define DBGUNDO(x...) do { } while (0) | |
2084 | #endif | |
2085 | ||
6687e988 | 2086 | static void tcp_undo_cwr(struct sock *sk, const int undo) |
1da177e4 | 2087 | { |
6687e988 ACM |
2088 | struct tcp_sock *tp = tcp_sk(sk); |
2089 | ||
1da177e4 | 2090 | if (tp->prior_ssthresh) { |
6687e988 ACM |
2091 | const struct inet_connection_sock *icsk = inet_csk(sk); |
2092 | ||
2093 | if (icsk->icsk_ca_ops->undo_cwnd) | |
2094 | tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); | |
1da177e4 LT |
2095 | else |
2096 | tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1); | |
2097 | ||
2098 | if (undo && tp->prior_ssthresh > tp->snd_ssthresh) { | |
2099 | tp->snd_ssthresh = tp->prior_ssthresh; | |
2100 | TCP_ECN_withdraw_cwr(tp); | |
2101 | } | |
2102 | } else { | |
2103 | tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); | |
2104 | } | |
2105 | tcp_moderate_cwnd(tp); | |
2106 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
6a438bbe SH |
2107 | |
2108 | /* There is something screwy going on with the retrans hints after | |
2109 | an undo */ | |
5af4ec23 | 2110 | tcp_clear_all_retrans_hints(tp); |
1da177e4 LT |
2111 | } |
2112 | ||
2113 | static inline int tcp_may_undo(struct tcp_sock *tp) | |
2114 | { | |
2115 | return tp->undo_marker && | |
2116 | (!tp->undo_retrans || tcp_packet_delayed(tp)); | |
2117 | } | |
2118 | ||
2119 | /* People celebrate: "We love our President!" */ | |
9e412ba7 | 2120 | static int tcp_try_undo_recovery(struct sock *sk) |
1da177e4 | 2121 | { |
9e412ba7 IJ |
2122 | struct tcp_sock *tp = tcp_sk(sk); |
2123 | ||
1da177e4 LT |
2124 | if (tcp_may_undo(tp)) { |
2125 | /* Happy end! We did not retransmit anything | |
2126 | * or our original transmission succeeded. | |
2127 | */ | |
9e412ba7 | 2128 | DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); |
6687e988 ACM |
2129 | tcp_undo_cwr(sk, 1); |
2130 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) | |
1da177e4 LT |
2131 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); |
2132 | else | |
2133 | NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO); | |
2134 | tp->undo_marker = 0; | |
2135 | } | |
e60402d0 | 2136 | if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { |
1da177e4 LT |
2137 | /* Hold old state until something *above* high_seq |
2138 | * is ACKed. For Reno it is MUST to prevent false | |
2139 | * fast retransmits (RFC2582). SACK TCP is safe. */ | |
2140 | tcp_moderate_cwnd(tp); | |
2141 | return 1; | |
2142 | } | |
6687e988 | 2143 | tcp_set_ca_state(sk, TCP_CA_Open); |
1da177e4 LT |
2144 | return 0; |
2145 | } | |
2146 | ||
2147 | /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ | |
9e412ba7 | 2148 | static void tcp_try_undo_dsack(struct sock *sk) |
1da177e4 | 2149 | { |
9e412ba7 IJ |
2150 | struct tcp_sock *tp = tcp_sk(sk); |
2151 | ||
1da177e4 | 2152 | if (tp->undo_marker && !tp->undo_retrans) { |
9e412ba7 | 2153 | DBGUNDO(sk, "D-SACK"); |
6687e988 | 2154 | tcp_undo_cwr(sk, 1); |
1da177e4 LT |
2155 | tp->undo_marker = 0; |
2156 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO); | |
2157 | } | |
2158 | } | |
2159 | ||
2160 | /* Undo during fast recovery after partial ACK. */ | |
2161 | ||
9e412ba7 | 2162 | static int tcp_try_undo_partial(struct sock *sk, int acked) |
1da177e4 | 2163 | { |
9e412ba7 | 2164 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 2165 | /* Partial ACK arrived. Force Hoe's retransmit. */ |
e60402d0 | 2166 | int failed = tcp_is_reno(tp) || tp->fackets_out>tp->reordering; |
1da177e4 LT |
2167 | |
2168 | if (tcp_may_undo(tp)) { | |
2169 | /* Plain luck! Hole if filled with delayed | |
2170 | * packet, rather than with a retransmit. | |
2171 | */ | |
2172 | if (tp->retrans_out == 0) | |
2173 | tp->retrans_stamp = 0; | |
2174 | ||
6687e988 | 2175 | tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); |
1da177e4 | 2176 | |
9e412ba7 | 2177 | DBGUNDO(sk, "Hoe"); |
6687e988 | 2178 | tcp_undo_cwr(sk, 0); |
1da177e4 LT |
2179 | NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO); |
2180 | ||
2181 | /* So... Do not make Hoe's retransmit yet. | |
2182 | * If the first packet was delayed, the rest | |
2183 | * ones are most probably delayed as well. | |
2184 | */ | |
2185 | failed = 0; | |
2186 | } | |
2187 | return failed; | |
2188 | } | |
2189 | ||
2190 | /* Undo during loss recovery after partial ACK. */ | |
9e412ba7 | 2191 | static int tcp_try_undo_loss(struct sock *sk) |
1da177e4 | 2192 | { |
9e412ba7 IJ |
2193 | struct tcp_sock *tp = tcp_sk(sk); |
2194 | ||
1da177e4 LT |
2195 | if (tcp_may_undo(tp)) { |
2196 | struct sk_buff *skb; | |
fe067e8a DM |
2197 | tcp_for_write_queue(skb, sk) { |
2198 | if (skb == tcp_send_head(sk)) | |
2199 | break; | |
1da177e4 LT |
2200 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; |
2201 | } | |
6a438bbe | 2202 | |
5af4ec23 | 2203 | tcp_clear_all_retrans_hints(tp); |
6a438bbe | 2204 | |
9e412ba7 | 2205 | DBGUNDO(sk, "partial loss"); |
1da177e4 | 2206 | tp->lost_out = 0; |
6687e988 | 2207 | tcp_undo_cwr(sk, 1); |
1da177e4 | 2208 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); |
463c84b9 | 2209 | inet_csk(sk)->icsk_retransmits = 0; |
1da177e4 | 2210 | tp->undo_marker = 0; |
e60402d0 | 2211 | if (tcp_is_sack(tp)) |
6687e988 | 2212 | tcp_set_ca_state(sk, TCP_CA_Open); |
1da177e4 LT |
2213 | return 1; |
2214 | } | |
2215 | return 0; | |
2216 | } | |
2217 | ||
6687e988 | 2218 | static inline void tcp_complete_cwr(struct sock *sk) |
1da177e4 | 2219 | { |
6687e988 | 2220 | struct tcp_sock *tp = tcp_sk(sk); |
317a76f9 | 2221 | tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); |
1da177e4 | 2222 | tp->snd_cwnd_stamp = tcp_time_stamp; |
6687e988 | 2223 | tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); |
1da177e4 LT |
2224 | } |
2225 | ||
9e412ba7 | 2226 | static void tcp_try_to_open(struct sock *sk, int flag) |
1da177e4 | 2227 | { |
9e412ba7 IJ |
2228 | struct tcp_sock *tp = tcp_sk(sk); |
2229 | ||
86426c22 IJ |
2230 | tcp_verify_left_out(tp); |
2231 | ||
1da177e4 LT |
2232 | if (tp->retrans_out == 0) |
2233 | tp->retrans_stamp = 0; | |
2234 | ||
2235 | if (flag&FLAG_ECE) | |
3cfe3baa | 2236 | tcp_enter_cwr(sk, 1); |
1da177e4 | 2237 | |
6687e988 | 2238 | if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { |
1da177e4 LT |
2239 | int state = TCP_CA_Open; |
2240 | ||
d02596e3 | 2241 | if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker) |
1da177e4 LT |
2242 | state = TCP_CA_Disorder; |
2243 | ||
6687e988 ACM |
2244 | if (inet_csk(sk)->icsk_ca_state != state) { |
2245 | tcp_set_ca_state(sk, state); | |
1da177e4 LT |
2246 | tp->high_seq = tp->snd_nxt; |
2247 | } | |
2248 | tcp_moderate_cwnd(tp); | |
2249 | } else { | |
1e757f99 | 2250 | tcp_cwnd_down(sk, flag); |
1da177e4 LT |
2251 | } |
2252 | } | |
2253 | ||
5d424d5a JH |
2254 | static void tcp_mtup_probe_failed(struct sock *sk) |
2255 | { | |
2256 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2257 | ||
2258 | icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; | |
2259 | icsk->icsk_mtup.probe_size = 0; | |
2260 | } | |
2261 | ||
2262 | static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb) | |
2263 | { | |
2264 | struct tcp_sock *tp = tcp_sk(sk); | |
2265 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2266 | ||
2267 | /* FIXME: breaks with very large cwnd */ | |
2268 | tp->prior_ssthresh = tcp_current_ssthresh(sk); | |
2269 | tp->snd_cwnd = tp->snd_cwnd * | |
2270 | tcp_mss_to_mtu(sk, tp->mss_cache) / | |
2271 | icsk->icsk_mtup.probe_size; | |
2272 | tp->snd_cwnd_cnt = 0; | |
2273 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
2274 | tp->rcv_ssthresh = tcp_current_ssthresh(sk); | |
2275 | ||
2276 | icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; | |
2277 | icsk->icsk_mtup.probe_size = 0; | |
2278 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); | |
2279 | } | |
2280 | ||
2281 | ||
1da177e4 LT |
2282 | /* Process an event, which can update packets-in-flight not trivially. |
2283 | * Main goal of this function is to calculate new estimate for left_out, | |
2284 | * taking into account both packets sitting in receiver's buffer and | |
2285 | * packets lost by network. | |
2286 | * | |
2287 | * Besides that it does CWND reduction, when packet loss is detected | |
2288 | * and changes state of machine. | |
2289 | * | |
2290 | * It does _not_ decide what to send, it is made in function | |
2291 | * tcp_xmit_retransmit_queue(). | |
2292 | */ | |
2293 | static void | |
1b6d427b | 2294 | tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag) |
1da177e4 | 2295 | { |
6687e988 | 2296 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 2297 | struct tcp_sock *tp = tcp_sk(sk); |
2e605294 IJ |
2298 | int is_dupack = !(flag&(FLAG_SND_UNA_ADVANCED|FLAG_NOT_DUP)); |
2299 | int do_lost = is_dupack || ((flag&FLAG_DATA_SACKED) && | |
2300 | (tp->fackets_out > tp->reordering)); | |
1da177e4 LT |
2301 | |
2302 | /* Some technical things: | |
2303 | * 1. Reno does not count dupacks (sacked_out) automatically. */ | |
2304 | if (!tp->packets_out) | |
2305 | tp->sacked_out = 0; | |
91fed7a1 IJ |
2306 | |
2307 | if (WARN_ON(!tp->sacked_out && tp->fackets_out)) | |
1da177e4 LT |
2308 | tp->fackets_out = 0; |
2309 | ||
e905a9ed | 2310 | /* Now state machine starts. |
1da177e4 LT |
2311 | * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ |
2312 | if (flag&FLAG_ECE) | |
2313 | tp->prior_ssthresh = 0; | |
2314 | ||
2315 | /* B. In all the states check for reneging SACKs. */ | |
463c84b9 | 2316 | if (tp->sacked_out && tcp_check_sack_reneging(sk)) |
1da177e4 LT |
2317 | return; |
2318 | ||
2319 | /* C. Process data loss notification, provided it is valid. */ | |
2320 | if ((flag&FLAG_DATA_LOST) && | |
2321 | before(tp->snd_una, tp->high_seq) && | |
6687e988 | 2322 | icsk->icsk_ca_state != TCP_CA_Open && |
1da177e4 | 2323 | tp->fackets_out > tp->reordering) { |
9e412ba7 | 2324 | tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, tp->high_seq); |
1da177e4 LT |
2325 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSS); |
2326 | } | |
2327 | ||
005903bc IJ |
2328 | /* D. Check consistency of the current state. */ |
2329 | tcp_verify_left_out(tp); | |
1da177e4 LT |
2330 | |
2331 | /* E. Check state exit conditions. State can be terminated | |
2332 | * when high_seq is ACKed. */ | |
6687e988 | 2333 | if (icsk->icsk_ca_state == TCP_CA_Open) { |
7b0eb22b | 2334 | BUG_TRAP(tp->retrans_out == 0); |
1da177e4 LT |
2335 | tp->retrans_stamp = 0; |
2336 | } else if (!before(tp->snd_una, tp->high_seq)) { | |
6687e988 | 2337 | switch (icsk->icsk_ca_state) { |
1da177e4 | 2338 | case TCP_CA_Loss: |
6687e988 | 2339 | icsk->icsk_retransmits = 0; |
9e412ba7 | 2340 | if (tcp_try_undo_recovery(sk)) |
1da177e4 LT |
2341 | return; |
2342 | break; | |
2343 | ||
2344 | case TCP_CA_CWR: | |
2345 | /* CWR is to be held something *above* high_seq | |
2346 | * is ACKed for CWR bit to reach receiver. */ | |
2347 | if (tp->snd_una != tp->high_seq) { | |
6687e988 ACM |
2348 | tcp_complete_cwr(sk); |
2349 | tcp_set_ca_state(sk, TCP_CA_Open); | |
1da177e4 LT |
2350 | } |
2351 | break; | |
2352 | ||
2353 | case TCP_CA_Disorder: | |
9e412ba7 | 2354 | tcp_try_undo_dsack(sk); |
1da177e4 LT |
2355 | if (!tp->undo_marker || |
2356 | /* For SACK case do not Open to allow to undo | |
2357 | * catching for all duplicate ACKs. */ | |
e60402d0 | 2358 | tcp_is_reno(tp) || tp->snd_una != tp->high_seq) { |
1da177e4 | 2359 | tp->undo_marker = 0; |
6687e988 | 2360 | tcp_set_ca_state(sk, TCP_CA_Open); |
1da177e4 LT |
2361 | } |
2362 | break; | |
2363 | ||
2364 | case TCP_CA_Recovery: | |
e60402d0 | 2365 | if (tcp_is_reno(tp)) |
1da177e4 | 2366 | tcp_reset_reno_sack(tp); |
9e412ba7 | 2367 | if (tcp_try_undo_recovery(sk)) |
1da177e4 | 2368 | return; |
6687e988 | 2369 | tcp_complete_cwr(sk); |
1da177e4 LT |
2370 | break; |
2371 | } | |
2372 | } | |
2373 | ||
2374 | /* F. Process state. */ | |
6687e988 | 2375 | switch (icsk->icsk_ca_state) { |
1da177e4 | 2376 | case TCP_CA_Recovery: |
2e605294 | 2377 | if (!(flag & FLAG_SND_UNA_ADVANCED)) { |
e60402d0 | 2378 | if (tcp_is_reno(tp) && is_dupack) |
6687e988 | 2379 | tcp_add_reno_sack(sk); |
1b6d427b IJ |
2380 | } else |
2381 | do_lost = tcp_try_undo_partial(sk, pkts_acked); | |
1da177e4 LT |
2382 | break; |
2383 | case TCP_CA_Loss: | |
2384 | if (flag&FLAG_DATA_ACKED) | |
6687e988 | 2385 | icsk->icsk_retransmits = 0; |
9e412ba7 | 2386 | if (!tcp_try_undo_loss(sk)) { |
1da177e4 LT |
2387 | tcp_moderate_cwnd(tp); |
2388 | tcp_xmit_retransmit_queue(sk); | |
2389 | return; | |
2390 | } | |
6687e988 | 2391 | if (icsk->icsk_ca_state != TCP_CA_Open) |
1da177e4 LT |
2392 | return; |
2393 | /* Loss is undone; fall through to processing in Open state. */ | |
2394 | default: | |
e60402d0 | 2395 | if (tcp_is_reno(tp)) { |
2e605294 | 2396 | if (flag & FLAG_SND_UNA_ADVANCED) |
1da177e4 LT |
2397 | tcp_reset_reno_sack(tp); |
2398 | if (is_dupack) | |
6687e988 | 2399 | tcp_add_reno_sack(sk); |
1da177e4 LT |
2400 | } |
2401 | ||
6687e988 | 2402 | if (icsk->icsk_ca_state == TCP_CA_Disorder) |
9e412ba7 | 2403 | tcp_try_undo_dsack(sk); |
1da177e4 | 2404 | |
9e412ba7 IJ |
2405 | if (!tcp_time_to_recover(sk)) { |
2406 | tcp_try_to_open(sk, flag); | |
1da177e4 LT |
2407 | return; |
2408 | } | |
2409 | ||
5d424d5a JH |
2410 | /* MTU probe failure: don't reduce cwnd */ |
2411 | if (icsk->icsk_ca_state < TCP_CA_CWR && | |
2412 | icsk->icsk_mtup.probe_size && | |
0e7b1368 | 2413 | tp->snd_una == tp->mtu_probe.probe_seq_start) { |
5d424d5a JH |
2414 | tcp_mtup_probe_failed(sk); |
2415 | /* Restores the reduction we did in tcp_mtup_probe() */ | |
2416 | tp->snd_cwnd++; | |
2417 | tcp_simple_retransmit(sk); | |
2418 | return; | |
2419 | } | |
2420 | ||
1da177e4 LT |
2421 | /* Otherwise enter Recovery state */ |
2422 | ||
e60402d0 | 2423 | if (tcp_is_reno(tp)) |
1da177e4 LT |
2424 | NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY); |
2425 | else | |
2426 | NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY); | |
2427 | ||
2428 | tp->high_seq = tp->snd_nxt; | |
2429 | tp->prior_ssthresh = 0; | |
2430 | tp->undo_marker = tp->snd_una; | |
2431 | tp->undo_retrans = tp->retrans_out; | |
2432 | ||
6687e988 | 2433 | if (icsk->icsk_ca_state < TCP_CA_CWR) { |
1da177e4 | 2434 | if (!(flag&FLAG_ECE)) |
6687e988 ACM |
2435 | tp->prior_ssthresh = tcp_current_ssthresh(sk); |
2436 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
1da177e4 LT |
2437 | TCP_ECN_queue_cwr(tp); |
2438 | } | |
2439 | ||
9772efb9 | 2440 | tp->bytes_acked = 0; |
1da177e4 | 2441 | tp->snd_cwnd_cnt = 0; |
6687e988 | 2442 | tcp_set_ca_state(sk, TCP_CA_Recovery); |
1da177e4 LT |
2443 | } |
2444 | ||
2e605294 | 2445 | if (do_lost || tcp_head_timedout(sk)) |
9e412ba7 | 2446 | tcp_update_scoreboard(sk); |
1e757f99 | 2447 | tcp_cwnd_down(sk, flag); |
1da177e4 LT |
2448 | tcp_xmit_retransmit_queue(sk); |
2449 | } | |
2450 | ||
2451 | /* Read draft-ietf-tcplw-high-performance before mucking | |
caa20d9a | 2452 | * with this code. (Supersedes RFC1323) |
1da177e4 | 2453 | */ |
2d2abbab | 2454 | static void tcp_ack_saw_tstamp(struct sock *sk, int flag) |
1da177e4 | 2455 | { |
1da177e4 LT |
2456 | /* RTTM Rule: A TSecr value received in a segment is used to |
2457 | * update the averaged RTT measurement only if the segment | |
2458 | * acknowledges some new data, i.e., only if it advances the | |
2459 | * left edge of the send window. | |
2460 | * | |
2461 | * See draft-ietf-tcplw-high-performance-00, section 3.3. | |
2462 | * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> | |
2463 | * | |
2464 | * Changed: reset backoff as soon as we see the first valid sample. | |
caa20d9a | 2465 | * If we do not, we get strongly overestimated rto. With timestamps |
1da177e4 LT |
2466 | * samples are accepted even from very old segments: f.e., when rtt=1 |
2467 | * increases to 8, we retransmit 5 times and after 8 seconds delayed | |
2468 | * answer arrives rto becomes 120 seconds! If at least one of segments | |
2469 | * in window is lost... Voila. --ANK (010210) | |
2470 | */ | |
463c84b9 ACM |
2471 | struct tcp_sock *tp = tcp_sk(sk); |
2472 | const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr; | |
2d2abbab | 2473 | tcp_rtt_estimator(sk, seq_rtt); |
463c84b9 ACM |
2474 | tcp_set_rto(sk); |
2475 | inet_csk(sk)->icsk_backoff = 0; | |
2476 | tcp_bound_rto(sk); | |
1da177e4 LT |
2477 | } |
2478 | ||
2d2abbab | 2479 | static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) |
1da177e4 LT |
2480 | { |
2481 | /* We don't have a timestamp. Can only use | |
2482 | * packets that are not retransmitted to determine | |
2483 | * rtt estimates. Also, we must not reset the | |
2484 | * backoff for rto until we get a non-retransmitted | |
2485 | * packet. This allows us to deal with a situation | |
2486 | * where the network delay has increased suddenly. | |
2487 | * I.e. Karn's algorithm. (SIGCOMM '87, p5.) | |
2488 | */ | |
2489 | ||
2490 | if (flag & FLAG_RETRANS_DATA_ACKED) | |
2491 | return; | |
2492 | ||
2d2abbab | 2493 | tcp_rtt_estimator(sk, seq_rtt); |
463c84b9 ACM |
2494 | tcp_set_rto(sk); |
2495 | inet_csk(sk)->icsk_backoff = 0; | |
2496 | tcp_bound_rto(sk); | |
1da177e4 LT |
2497 | } |
2498 | ||
463c84b9 | 2499 | static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, |
2d2abbab | 2500 | const s32 seq_rtt) |
1da177e4 | 2501 | { |
463c84b9 | 2502 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
2503 | /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ |
2504 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) | |
2d2abbab | 2505 | tcp_ack_saw_tstamp(sk, flag); |
1da177e4 | 2506 | else if (seq_rtt >= 0) |
2d2abbab | 2507 | tcp_ack_no_tstamp(sk, seq_rtt, flag); |
1da177e4 LT |
2508 | } |
2509 | ||
16751347 | 2510 | static void tcp_cong_avoid(struct sock *sk, u32 ack, |
40efc6fa | 2511 | u32 in_flight, int good) |
1da177e4 | 2512 | { |
6687e988 | 2513 | const struct inet_connection_sock *icsk = inet_csk(sk); |
16751347 | 2514 | icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight, good); |
6687e988 | 2515 | tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; |
1da177e4 LT |
2516 | } |
2517 | ||
1da177e4 LT |
2518 | /* Restart timer after forward progress on connection. |
2519 | * RFC2988 recommends to restart timer to now+rto. | |
2520 | */ | |
6728e7dc | 2521 | static void tcp_rearm_rto(struct sock *sk) |
1da177e4 | 2522 | { |
9e412ba7 IJ |
2523 | struct tcp_sock *tp = tcp_sk(sk); |
2524 | ||
1da177e4 | 2525 | if (!tp->packets_out) { |
463c84b9 | 2526 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); |
1da177e4 | 2527 | } else { |
3f421baa | 2528 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX); |
1da177e4 LT |
2529 | } |
2530 | } | |
2531 | ||
7c46a03e | 2532 | /* If we get here, the whole TSO packet has not been acked. */ |
13fcf850 | 2533 | static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) |
1da177e4 LT |
2534 | { |
2535 | struct tcp_sock *tp = tcp_sk(sk); | |
7c46a03e | 2536 | u32 packets_acked; |
1da177e4 | 2537 | |
7c46a03e | 2538 | BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); |
1da177e4 LT |
2539 | |
2540 | packets_acked = tcp_skb_pcount(skb); | |
7c46a03e | 2541 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
1da177e4 LT |
2542 | return 0; |
2543 | packets_acked -= tcp_skb_pcount(skb); | |
2544 | ||
2545 | if (packets_acked) { | |
1da177e4 | 2546 | BUG_ON(tcp_skb_pcount(skb) == 0); |
7c46a03e | 2547 | BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); |
1da177e4 LT |
2548 | } |
2549 | ||
13fcf850 | 2550 | return packets_acked; |
1da177e4 LT |
2551 | } |
2552 | ||
7c46a03e IJ |
2553 | /* Remove acknowledged frames from the retransmission queue. If our packet |
2554 | * is before the ack sequence we can discard it as it's confirmed to have | |
2555 | * arrived at the other end. | |
2556 | */ | |
2557 | static int tcp_clean_rtx_queue(struct sock *sk, s32 *seq_rtt_p) | |
1da177e4 LT |
2558 | { |
2559 | struct tcp_sock *tp = tcp_sk(sk); | |
2d2abbab | 2560 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 2561 | struct sk_buff *skb; |
7c46a03e | 2562 | u32 now = tcp_time_stamp; |
13fcf850 | 2563 | int fully_acked = 1; |
7c46a03e | 2564 | int flag = 0; |
6418204f | 2565 | int prior_packets = tp->packets_out; |
7c46a03e | 2566 | s32 seq_rtt = -1; |
b9ce204f | 2567 | ktime_t last_ackt = net_invalid_timestamp(); |
1da177e4 | 2568 | |
7c46a03e | 2569 | while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) { |
e905a9ed | 2570 | struct tcp_skb_cb *scb = TCP_SKB_CB(skb); |
13fcf850 IJ |
2571 | u32 end_seq; |
2572 | u32 packets_acked; | |
7c46a03e | 2573 | u8 sacked = scb->sacked; |
1da177e4 | 2574 | |
1da177e4 | 2575 | if (after(scb->end_seq, tp->snd_una)) { |
13fcf850 IJ |
2576 | if (tcp_skb_pcount(skb) == 1 || |
2577 | !after(tp->snd_una, scb->seq)) | |
2578 | break; | |
2579 | ||
2580 | packets_acked = tcp_tso_acked(sk, skb); | |
2581 | if (!packets_acked) | |
2582 | break; | |
2583 | ||
2584 | fully_acked = 0; | |
2585 | end_seq = tp->snd_una; | |
2586 | } else { | |
2587 | packets_acked = tcp_skb_pcount(skb); | |
2588 | end_seq = scb->end_seq; | |
1da177e4 LT |
2589 | } |
2590 | ||
5d424d5a | 2591 | /* MTU probing checks */ |
7c46a03e IJ |
2592 | if (fully_acked && icsk->icsk_mtup.probe_size && |
2593 | !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) { | |
2594 | tcp_mtup_probe_success(sk, skb); | |
5d424d5a JH |
2595 | } |
2596 | ||
1da177e4 LT |
2597 | if (sacked) { |
2598 | if (sacked & TCPCB_RETRANS) { | |
2de979bd | 2599 | if (sacked & TCPCB_SACKED_RETRANS) |
13fcf850 | 2600 | tp->retrans_out -= packets_acked; |
7c46a03e | 2601 | flag |= FLAG_RETRANS_DATA_ACKED; |
1da177e4 | 2602 | seq_rtt = -1; |
009a2e3e IJ |
2603 | if ((flag & FLAG_DATA_ACKED) || |
2604 | (packets_acked > 1)) | |
2605 | flag |= FLAG_NONHEAD_RETRANS_ACKED; | |
2d2abbab | 2606 | } else if (seq_rtt < 0) { |
1da177e4 | 2607 | seq_rtt = now - scb->when; |
13fcf850 IJ |
2608 | if (fully_acked) |
2609 | last_ackt = skb->tstamp; | |
a61bbcf2 | 2610 | } |
7c46a03e | 2611 | |
1da177e4 | 2612 | if (sacked & TCPCB_SACKED_ACKED) |
13fcf850 | 2613 | tp->sacked_out -= packets_acked; |
1da177e4 | 2614 | if (sacked & TCPCB_LOST) |
13fcf850 | 2615 | tp->lost_out -= packets_acked; |
7c46a03e IJ |
2616 | |
2617 | if ((sacked & TCPCB_URG) && tp->urg_mode && | |
2618 | !before(end_seq, tp->snd_up)) | |
2619 | tp->urg_mode = 0; | |
2d2abbab | 2620 | } else if (seq_rtt < 0) { |
1da177e4 | 2621 | seq_rtt = now - scb->when; |
13fcf850 IJ |
2622 | if (fully_acked) |
2623 | last_ackt = skb->tstamp; | |
2d2abbab | 2624 | } |
13fcf850 IJ |
2625 | tp->packets_out -= packets_acked; |
2626 | ||
009a2e3e IJ |
2627 | /* Initial outgoing SYN's get put onto the write_queue |
2628 | * just like anything else we transmit. It is not | |
2629 | * true data, and if we misinform our callers that | |
2630 | * this ACK acks real data, we will erroneously exit | |
2631 | * connection startup slow start one packet too | |
2632 | * quickly. This is severely frowned upon behavior. | |
2633 | */ | |
2634 | if (!(scb->flags & TCPCB_FLAG_SYN)) { | |
2635 | flag |= FLAG_DATA_ACKED; | |
2636 | } else { | |
2637 | flag |= FLAG_SYN_ACKED; | |
2638 | tp->retrans_stamp = 0; | |
2639 | } | |
2640 | ||
13fcf850 IJ |
2641 | if (!fully_acked) |
2642 | break; | |
2643 | ||
fe067e8a | 2644 | tcp_unlink_write_queue(skb, sk); |
1da177e4 | 2645 | sk_stream_free_skb(sk, skb); |
5af4ec23 | 2646 | tcp_clear_all_retrans_hints(tp); |
1da177e4 LT |
2647 | } |
2648 | ||
7c46a03e | 2649 | if (flag & FLAG_ACKED) { |
6418204f | 2650 | u32 pkts_acked = prior_packets - tp->packets_out; |
164891aa SH |
2651 | const struct tcp_congestion_ops *ca_ops |
2652 | = inet_csk(sk)->icsk_ca_ops; | |
2653 | ||
7c46a03e | 2654 | tcp_ack_update_rtt(sk, flag, seq_rtt); |
6728e7dc | 2655 | tcp_rearm_rto(sk); |
317a76f9 | 2656 | |
91fed7a1 | 2657 | tp->fackets_out -= min(pkts_acked, tp->fackets_out); |
13fcf850 IJ |
2658 | /* hint's skb might be NULL but we don't need to care */ |
2659 | tp->fastpath_cnt_hint -= min_t(u32, pkts_acked, | |
2660 | tp->fastpath_cnt_hint); | |
e60402d0 | 2661 | if (tcp_is_reno(tp)) |
1b6d427b IJ |
2662 | tcp_remove_reno_sacks(sk, pkts_acked); |
2663 | ||
30cfd0ba SH |
2664 | if (ca_ops->pkts_acked) { |
2665 | s32 rtt_us = -1; | |
2666 | ||
2667 | /* Is the ACK triggering packet unambiguous? */ | |
7c46a03e | 2668 | if (!(flag & FLAG_RETRANS_DATA_ACKED)) { |
30cfd0ba SH |
2669 | /* High resolution needed and available? */ |
2670 | if (ca_ops->flags & TCP_CONG_RTT_STAMP && | |
2671 | !ktime_equal(last_ackt, | |
2672 | net_invalid_timestamp())) | |
2673 | rtt_us = ktime_us_delta(ktime_get_real(), | |
2674 | last_ackt); | |
2675 | else if (seq_rtt > 0) | |
2676 | rtt_us = jiffies_to_usecs(seq_rtt); | |
2677 | } | |
b9ce204f | 2678 | |
30cfd0ba SH |
2679 | ca_ops->pkts_acked(sk, pkts_acked, rtt_us); |
2680 | } | |
1da177e4 LT |
2681 | } |
2682 | ||
2683 | #if FASTRETRANS_DEBUG > 0 | |
2684 | BUG_TRAP((int)tp->sacked_out >= 0); | |
2685 | BUG_TRAP((int)tp->lost_out >= 0); | |
2686 | BUG_TRAP((int)tp->retrans_out >= 0); | |
e60402d0 | 2687 | if (!tp->packets_out && tcp_is_sack(tp)) { |
6687e988 | 2688 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2689 | if (tp->lost_out) { |
2690 | printk(KERN_DEBUG "Leak l=%u %d\n", | |
6687e988 | 2691 | tp->lost_out, icsk->icsk_ca_state); |
1da177e4 LT |
2692 | tp->lost_out = 0; |
2693 | } | |
2694 | if (tp->sacked_out) { | |
2695 | printk(KERN_DEBUG "Leak s=%u %d\n", | |
6687e988 | 2696 | tp->sacked_out, icsk->icsk_ca_state); |
1da177e4 LT |
2697 | tp->sacked_out = 0; |
2698 | } | |
2699 | if (tp->retrans_out) { | |
2700 | printk(KERN_DEBUG "Leak r=%u %d\n", | |
6687e988 | 2701 | tp->retrans_out, icsk->icsk_ca_state); |
1da177e4 LT |
2702 | tp->retrans_out = 0; |
2703 | } | |
2704 | } | |
2705 | #endif | |
2706 | *seq_rtt_p = seq_rtt; | |
7c46a03e | 2707 | return flag; |
1da177e4 LT |
2708 | } |
2709 | ||
2710 | static void tcp_ack_probe(struct sock *sk) | |
2711 | { | |
463c84b9 ACM |
2712 | const struct tcp_sock *tp = tcp_sk(sk); |
2713 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1da177e4 LT |
2714 | |
2715 | /* Was it a usable window open? */ | |
2716 | ||
fe067e8a | 2717 | if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, |
1da177e4 | 2718 | tp->snd_una + tp->snd_wnd)) { |
463c84b9 ACM |
2719 | icsk->icsk_backoff = 0; |
2720 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); | |
1da177e4 LT |
2721 | /* Socket must be waked up by subsequent tcp_data_snd_check(). |
2722 | * This function is not for random using! | |
2723 | */ | |
2724 | } else { | |
463c84b9 | 2725 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
3f421baa ACM |
2726 | min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), |
2727 | TCP_RTO_MAX); | |
1da177e4 LT |
2728 | } |
2729 | } | |
2730 | ||
6687e988 | 2731 | static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag) |
1da177e4 LT |
2732 | { |
2733 | return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || | |
6687e988 | 2734 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open); |
1da177e4 LT |
2735 | } |
2736 | ||
6687e988 | 2737 | static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag) |
1da177e4 | 2738 | { |
6687e988 | 2739 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 2740 | return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && |
6687e988 | 2741 | !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR)); |
1da177e4 LT |
2742 | } |
2743 | ||
2744 | /* Check that window update is acceptable. | |
2745 | * The function assumes that snd_una<=ack<=snd_next. | |
2746 | */ | |
463c84b9 ACM |
2747 | static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack, |
2748 | const u32 ack_seq, const u32 nwin) | |
1da177e4 LT |
2749 | { |
2750 | return (after(ack, tp->snd_una) || | |
2751 | after(ack_seq, tp->snd_wl1) || | |
2752 | (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd)); | |
2753 | } | |
2754 | ||
2755 | /* Update our send window. | |
2756 | * | |
2757 | * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 | |
2758 | * and in FreeBSD. NetBSD's one is even worse.) is wrong. | |
2759 | */ | |
9e412ba7 IJ |
2760 | static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack, |
2761 | u32 ack_seq) | |
1da177e4 | 2762 | { |
9e412ba7 | 2763 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 2764 | int flag = 0; |
aa8223c7 | 2765 | u32 nwin = ntohs(tcp_hdr(skb)->window); |
1da177e4 | 2766 | |
aa8223c7 | 2767 | if (likely(!tcp_hdr(skb)->syn)) |
1da177e4 LT |
2768 | nwin <<= tp->rx_opt.snd_wscale; |
2769 | ||
2770 | if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { | |
2771 | flag |= FLAG_WIN_UPDATE; | |
2772 | tcp_update_wl(tp, ack, ack_seq); | |
2773 | ||
2774 | if (tp->snd_wnd != nwin) { | |
2775 | tp->snd_wnd = nwin; | |
2776 | ||
2777 | /* Note, it is the only place, where | |
2778 | * fast path is recovered for sending TCP. | |
2779 | */ | |
2ad41065 | 2780 | tp->pred_flags = 0; |
9e412ba7 | 2781 | tcp_fast_path_check(sk); |
1da177e4 LT |
2782 | |
2783 | if (nwin > tp->max_window) { | |
2784 | tp->max_window = nwin; | |
d83d8461 | 2785 | tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); |
1da177e4 LT |
2786 | } |
2787 | } | |
2788 | } | |
2789 | ||
2790 | tp->snd_una = ack; | |
2791 | ||
2792 | return flag; | |
2793 | } | |
2794 | ||
9ead9a1d IJ |
2795 | /* A very conservative spurious RTO response algorithm: reduce cwnd and |
2796 | * continue in congestion avoidance. | |
2797 | */ | |
2798 | static void tcp_conservative_spur_to_response(struct tcp_sock *tp) | |
2799 | { | |
2800 | tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); | |
aa8b6a7a | 2801 | tp->snd_cwnd_cnt = 0; |
46323655 | 2802 | TCP_ECN_queue_cwr(tp); |
9ead9a1d IJ |
2803 | tcp_moderate_cwnd(tp); |
2804 | } | |
2805 | ||
3cfe3baa IJ |
2806 | /* A conservative spurious RTO response algorithm: reduce cwnd using |
2807 | * rate halving and continue in congestion avoidance. | |
2808 | */ | |
2809 | static void tcp_ratehalving_spur_to_response(struct sock *sk) | |
2810 | { | |
3cfe3baa | 2811 | tcp_enter_cwr(sk, 0); |
3cfe3baa IJ |
2812 | } |
2813 | ||
e317f6f6 | 2814 | static void tcp_undo_spur_to_response(struct sock *sk, int flag) |
3cfe3baa | 2815 | { |
e317f6f6 IJ |
2816 | if (flag&FLAG_ECE) |
2817 | tcp_ratehalving_spur_to_response(sk); | |
2818 | else | |
2819 | tcp_undo_cwr(sk, 1); | |
3cfe3baa IJ |
2820 | } |
2821 | ||
30935cf4 IJ |
2822 | /* F-RTO spurious RTO detection algorithm (RFC4138) |
2823 | * | |
6408d206 IJ |
2824 | * F-RTO affects during two new ACKs following RTO (well, almost, see inline |
2825 | * comments). State (ACK number) is kept in frto_counter. When ACK advances | |
2826 | * window (but not to or beyond highest sequence sent before RTO): | |
30935cf4 IJ |
2827 | * On First ACK, send two new segments out. |
2828 | * On Second ACK, RTO was likely spurious. Do spurious response (response | |
2829 | * algorithm is not part of the F-RTO detection algorithm | |
2830 | * given in RFC4138 but can be selected separately). | |
2831 | * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss | |
d551e454 IJ |
2832 | * and TCP falls back to conventional RTO recovery. F-RTO allows overriding |
2833 | * of Nagle, this is done using frto_counter states 2 and 3, when a new data | |
2834 | * segment of any size sent during F-RTO, state 2 is upgraded to 3. | |
30935cf4 IJ |
2835 | * |
2836 | * Rationale: if the RTO was spurious, new ACKs should arrive from the | |
2837 | * original window even after we transmit two new data segments. | |
2838 | * | |
4dc2665e IJ |
2839 | * SACK version: |
2840 | * on first step, wait until first cumulative ACK arrives, then move to | |
2841 | * the second step. In second step, the next ACK decides. | |
2842 | * | |
30935cf4 IJ |
2843 | * F-RTO is implemented (mainly) in four functions: |
2844 | * - tcp_use_frto() is used to determine if TCP is can use F-RTO | |
2845 | * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is | |
2846 | * called when tcp_use_frto() showed green light | |
2847 | * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm | |
2848 | * - tcp_enter_frto_loss() is called if there is not enough evidence | |
2849 | * to prove that the RTO is indeed spurious. It transfers the control | |
2850 | * from F-RTO to the conventional RTO recovery | |
2851 | */ | |
2e605294 | 2852 | static int tcp_process_frto(struct sock *sk, int flag) |
1da177e4 LT |
2853 | { |
2854 | struct tcp_sock *tp = tcp_sk(sk); | |
e905a9ed | 2855 | |
005903bc | 2856 | tcp_verify_left_out(tp); |
e905a9ed | 2857 | |
7487c48c IJ |
2858 | /* Duplicate the behavior from Loss state (fastretrans_alert) */ |
2859 | if (flag&FLAG_DATA_ACKED) | |
2860 | inet_csk(sk)->icsk_retransmits = 0; | |
2861 | ||
009a2e3e IJ |
2862 | if ((flag & FLAG_NONHEAD_RETRANS_ACKED) || |
2863 | ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED))) | |
2864 | tp->undo_marker = 0; | |
2865 | ||
95c4922b | 2866 | if (!before(tp->snd_una, tp->frto_highmark)) { |
d551e454 | 2867 | tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag); |
7c9a4a5b | 2868 | return 1; |
95c4922b IJ |
2869 | } |
2870 | ||
e60402d0 | 2871 | if (!IsSackFrto() || tcp_is_reno(tp)) { |
4dc2665e IJ |
2872 | /* RFC4138 shortcoming in step 2; should also have case c): |
2873 | * ACK isn't duplicate nor advances window, e.g., opposite dir | |
2874 | * data, winupdate | |
2875 | */ | |
2e605294 | 2876 | if (!(flag&FLAG_ANY_PROGRESS) && (flag&FLAG_NOT_DUP)) |
4dc2665e IJ |
2877 | return 1; |
2878 | ||
2879 | if (!(flag&FLAG_DATA_ACKED)) { | |
2880 | tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3), | |
2881 | flag); | |
2882 | return 1; | |
2883 | } | |
2884 | } else { | |
2885 | if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) { | |
2886 | /* Prevent sending of new data. */ | |
2887 | tp->snd_cwnd = min(tp->snd_cwnd, | |
2888 | tcp_packets_in_flight(tp)); | |
2889 | return 1; | |
2890 | } | |
6408d206 | 2891 | |
d551e454 | 2892 | if ((tp->frto_counter >= 2) && |
4dc2665e IJ |
2893 | (!(flag&FLAG_FORWARD_PROGRESS) || |
2894 | ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) { | |
2895 | /* RFC4138 shortcoming (see comment above) */ | |
2896 | if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP)) | |
2897 | return 1; | |
2898 | ||
2899 | tcp_enter_frto_loss(sk, 3, flag); | |
2900 | return 1; | |
2901 | } | |
1da177e4 LT |
2902 | } |
2903 | ||
2904 | if (tp->frto_counter == 1) { | |
575ee714 IJ |
2905 | /* Sending of the next skb must be allowed or no FRTO */ |
2906 | if (!tcp_send_head(sk) || | |
2907 | after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, | |
2908 | tp->snd_una + tp->snd_wnd)) { | |
d551e454 IJ |
2909 | tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), |
2910 | flag); | |
575ee714 IJ |
2911 | return 1; |
2912 | } | |
2913 | ||
1da177e4 | 2914 | tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; |
94d0ea77 | 2915 | tp->frto_counter = 2; |
7c9a4a5b | 2916 | return 1; |
d551e454 | 2917 | } else { |
3cfe3baa IJ |
2918 | switch (sysctl_tcp_frto_response) { |
2919 | case 2: | |
e317f6f6 | 2920 | tcp_undo_spur_to_response(sk, flag); |
3cfe3baa IJ |
2921 | break; |
2922 | case 1: | |
2923 | tcp_conservative_spur_to_response(tp); | |
2924 | break; | |
2925 | default: | |
2926 | tcp_ratehalving_spur_to_response(sk); | |
2927 | break; | |
3ff50b79 | 2928 | } |
94d0ea77 | 2929 | tp->frto_counter = 0; |
009a2e3e | 2930 | tp->undo_marker = 0; |
1da177e4 | 2931 | } |
7c9a4a5b | 2932 | return 0; |
1da177e4 LT |
2933 | } |
2934 | ||
1da177e4 LT |
2935 | /* This routine deals with incoming acks, but not outgoing ones. */ |
2936 | static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag) | |
2937 | { | |
6687e988 | 2938 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2939 | struct tcp_sock *tp = tcp_sk(sk); |
2940 | u32 prior_snd_una = tp->snd_una; | |
2941 | u32 ack_seq = TCP_SKB_CB(skb)->seq; | |
2942 | u32 ack = TCP_SKB_CB(skb)->ack_seq; | |
2943 | u32 prior_in_flight; | |
2944 | s32 seq_rtt; | |
2945 | int prior_packets; | |
7c9a4a5b | 2946 | int frto_cwnd = 0; |
1da177e4 LT |
2947 | |
2948 | /* If the ack is newer than sent or older than previous acks | |
2949 | * then we can probably ignore it. | |
2950 | */ | |
2951 | if (after(ack, tp->snd_nxt)) | |
2952 | goto uninteresting_ack; | |
2953 | ||
2954 | if (before(ack, prior_snd_una)) | |
2955 | goto old_ack; | |
2956 | ||
2e605294 IJ |
2957 | if (after(ack, prior_snd_una)) |
2958 | flag |= FLAG_SND_UNA_ADVANCED; | |
2959 | ||
3fdf3f0c DO |
2960 | if (sysctl_tcp_abc) { |
2961 | if (icsk->icsk_ca_state < TCP_CA_CWR) | |
2962 | tp->bytes_acked += ack - prior_snd_una; | |
2963 | else if (icsk->icsk_ca_state == TCP_CA_Loss) | |
2964 | /* we assume just one segment left network */ | |
2965 | tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache); | |
2966 | } | |
9772efb9 | 2967 | |
1da177e4 LT |
2968 | if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) { |
2969 | /* Window is constant, pure forward advance. | |
2970 | * No more checks are required. | |
2971 | * Note, we use the fact that SND.UNA>=SND.WL2. | |
2972 | */ | |
2973 | tcp_update_wl(tp, ack, ack_seq); | |
2974 | tp->snd_una = ack; | |
1da177e4 LT |
2975 | flag |= FLAG_WIN_UPDATE; |
2976 | ||
6687e988 | 2977 | tcp_ca_event(sk, CA_EVENT_FAST_ACK); |
317a76f9 | 2978 | |
1da177e4 LT |
2979 | NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS); |
2980 | } else { | |
2981 | if (ack_seq != TCP_SKB_CB(skb)->end_seq) | |
2982 | flag |= FLAG_DATA; | |
2983 | else | |
2984 | NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS); | |
2985 | ||
9e412ba7 | 2986 | flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); |
1da177e4 LT |
2987 | |
2988 | if (TCP_SKB_CB(skb)->sacked) | |
2989 | flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); | |
2990 | ||
aa8223c7 | 2991 | if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb))) |
1da177e4 LT |
2992 | flag |= FLAG_ECE; |
2993 | ||
6687e988 | 2994 | tcp_ca_event(sk, CA_EVENT_SLOW_ACK); |
1da177e4 LT |
2995 | } |
2996 | ||
2997 | /* We passed data and got it acked, remove any soft error | |
2998 | * log. Something worked... | |
2999 | */ | |
3000 | sk->sk_err_soft = 0; | |
3001 | tp->rcv_tstamp = tcp_time_stamp; | |
3002 | prior_packets = tp->packets_out; | |
3003 | if (!prior_packets) | |
3004 | goto no_queue; | |
3005 | ||
3006 | prior_in_flight = tcp_packets_in_flight(tp); | |
3007 | ||
3008 | /* See if we can take anything off of the retransmit queue. */ | |
2d2abbab | 3009 | flag |= tcp_clean_rtx_queue(sk, &seq_rtt); |
1da177e4 LT |
3010 | |
3011 | if (tp->frto_counter) | |
2e605294 | 3012 | frto_cwnd = tcp_process_frto(sk, flag); |
1da177e4 | 3013 | |
6687e988 | 3014 | if (tcp_ack_is_dubious(sk, flag)) { |
caa20d9a | 3015 | /* Advance CWND, if state allows this. */ |
7c9a4a5b IJ |
3016 | if ((flag & FLAG_DATA_ACKED) && !frto_cwnd && |
3017 | tcp_may_raise_cwnd(sk, flag)) | |
16751347 | 3018 | tcp_cong_avoid(sk, ack, prior_in_flight, 0); |
1b6d427b | 3019 | tcp_fastretrans_alert(sk, prior_packets - tp->packets_out, flag); |
1da177e4 | 3020 | } else { |
7c9a4a5b | 3021 | if ((flag & FLAG_DATA_ACKED) && !frto_cwnd) |
16751347 | 3022 | tcp_cong_avoid(sk, ack, prior_in_flight, 1); |
1da177e4 LT |
3023 | } |
3024 | ||
3025 | if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP)) | |
3026 | dst_confirm(sk->sk_dst_cache); | |
3027 | ||
3028 | return 1; | |
3029 | ||
3030 | no_queue: | |
6687e988 | 3031 | icsk->icsk_probes_out = 0; |
1da177e4 LT |
3032 | |
3033 | /* If this ack opens up a zero window, clear backoff. It was | |
3034 | * being used to time the probes, and is probably far higher than | |
3035 | * it needs to be for normal retransmission. | |
3036 | */ | |
fe067e8a | 3037 | if (tcp_send_head(sk)) |
1da177e4 LT |
3038 | tcp_ack_probe(sk); |
3039 | return 1; | |
3040 | ||
3041 | old_ack: | |
3042 | if (TCP_SKB_CB(skb)->sacked) | |
3043 | tcp_sacktag_write_queue(sk, skb, prior_snd_una); | |
3044 | ||
3045 | uninteresting_ack: | |
3046 | SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt); | |
3047 | return 0; | |
3048 | } | |
3049 | ||
3050 | ||
3051 | /* Look for tcp options. Normally only called on SYN and SYNACK packets. | |
3052 | * But, this can also be called on packets in the established flow when | |
3053 | * the fast version below fails. | |
3054 | */ | |
3055 | void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab) | |
3056 | { | |
3057 | unsigned char *ptr; | |
aa8223c7 | 3058 | struct tcphdr *th = tcp_hdr(skb); |
1da177e4 LT |
3059 | int length=(th->doff*4)-sizeof(struct tcphdr); |
3060 | ||
3061 | ptr = (unsigned char *)(th + 1); | |
3062 | opt_rx->saw_tstamp = 0; | |
3063 | ||
2de979bd | 3064 | while (length > 0) { |
e905a9ed | 3065 | int opcode=*ptr++; |
1da177e4 LT |
3066 | int opsize; |
3067 | ||
3068 | switch (opcode) { | |
3069 | case TCPOPT_EOL: | |
3070 | return; | |
3071 | case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ | |
3072 | length--; | |
3073 | continue; | |
3074 | default: | |
3075 | opsize=*ptr++; | |
3076 | if (opsize < 2) /* "silly options" */ | |
3077 | return; | |
3078 | if (opsize > length) | |
3079 | return; /* don't parse partial options */ | |
2de979bd | 3080 | switch (opcode) { |
1da177e4 | 3081 | case TCPOPT_MSS: |
2de979bd | 3082 | if (opsize==TCPOLEN_MSS && th->syn && !estab) { |
4f3608b7 | 3083 | u16 in_mss = ntohs(get_unaligned((__be16 *)ptr)); |
1da177e4 LT |
3084 | if (in_mss) { |
3085 | if (opt_rx->user_mss && opt_rx->user_mss < in_mss) | |
3086 | in_mss = opt_rx->user_mss; | |
3087 | opt_rx->mss_clamp = in_mss; | |
3088 | } | |
3089 | } | |
3090 | break; | |
3091 | case TCPOPT_WINDOW: | |
2de979bd | 3092 | if (opsize==TCPOLEN_WINDOW && th->syn && !estab) |
1da177e4 LT |
3093 | if (sysctl_tcp_window_scaling) { |
3094 | __u8 snd_wscale = *(__u8 *) ptr; | |
3095 | opt_rx->wscale_ok = 1; | |
3096 | if (snd_wscale > 14) { | |
2de979bd | 3097 | if (net_ratelimit()) |
1da177e4 LT |
3098 | printk(KERN_INFO "tcp_parse_options: Illegal window " |
3099 | "scaling value %d >14 received.\n", | |
3100 | snd_wscale); | |
3101 | snd_wscale = 14; | |
3102 | } | |
3103 | opt_rx->snd_wscale = snd_wscale; | |
3104 | } | |
3105 | break; | |
3106 | case TCPOPT_TIMESTAMP: | |
2de979bd | 3107 | if (opsize==TCPOLEN_TIMESTAMP) { |
1da177e4 LT |
3108 | if ((estab && opt_rx->tstamp_ok) || |
3109 | (!estab && sysctl_tcp_timestamps)) { | |
3110 | opt_rx->saw_tstamp = 1; | |
4f3608b7 AV |
3111 | opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr)); |
3112 | opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4))); | |
1da177e4 LT |
3113 | } |
3114 | } | |
3115 | break; | |
3116 | case TCPOPT_SACK_PERM: | |
2de979bd | 3117 | if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) { |
1da177e4 LT |
3118 | if (sysctl_tcp_sack) { |
3119 | opt_rx->sack_ok = 1; | |
3120 | tcp_sack_reset(opt_rx); | |
3121 | } | |
3122 | } | |
3123 | break; | |
3124 | ||
3125 | case TCPOPT_SACK: | |
2de979bd | 3126 | if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && |
1da177e4 LT |
3127 | !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && |
3128 | opt_rx->sack_ok) { | |
3129 | TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; | |
3130 | } | |
d7ea5b91 | 3131 | break; |
cfb6eeb4 YH |
3132 | #ifdef CONFIG_TCP_MD5SIG |
3133 | case TCPOPT_MD5SIG: | |
3134 | /* | |
3135 | * The MD5 Hash has already been | |
3136 | * checked (see tcp_v{4,6}_do_rcv()). | |
3137 | */ | |
3138 | break; | |
3139 | #endif | |
3ff50b79 SH |
3140 | } |
3141 | ||
e905a9ed YH |
3142 | ptr+=opsize-2; |
3143 | length-=opsize; | |
3ff50b79 | 3144 | } |
1da177e4 LT |
3145 | } |
3146 | } | |
3147 | ||
3148 | /* Fast parse options. This hopes to only see timestamps. | |
3149 | * If it is wrong it falls back on tcp_parse_options(). | |
3150 | */ | |
40efc6fa SH |
3151 | static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, |
3152 | struct tcp_sock *tp) | |
1da177e4 LT |
3153 | { |
3154 | if (th->doff == sizeof(struct tcphdr)>>2) { | |
3155 | tp->rx_opt.saw_tstamp = 0; | |
3156 | return 0; | |
3157 | } else if (tp->rx_opt.tstamp_ok && | |
3158 | th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) { | |
4f3608b7 AV |
3159 | __be32 *ptr = (__be32 *)(th + 1); |
3160 | if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | |
1da177e4 LT |
3161 | | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { |
3162 | tp->rx_opt.saw_tstamp = 1; | |
3163 | ++ptr; | |
3164 | tp->rx_opt.rcv_tsval = ntohl(*ptr); | |
3165 | ++ptr; | |
3166 | tp->rx_opt.rcv_tsecr = ntohl(*ptr); | |
3167 | return 1; | |
3168 | } | |
3169 | } | |
3170 | tcp_parse_options(skb, &tp->rx_opt, 1); | |
3171 | return 1; | |
3172 | } | |
3173 | ||
3174 | static inline void tcp_store_ts_recent(struct tcp_sock *tp) | |
3175 | { | |
3176 | tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; | |
9d729f72 | 3177 | tp->rx_opt.ts_recent_stamp = get_seconds(); |
1da177e4 LT |
3178 | } |
3179 | ||
3180 | static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) | |
3181 | { | |
3182 | if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { | |
3183 | /* PAWS bug workaround wrt. ACK frames, the PAWS discard | |
3184 | * extra check below makes sure this can only happen | |
3185 | * for pure ACK frames. -DaveM | |
3186 | * | |
3187 | * Not only, also it occurs for expired timestamps. | |
3188 | */ | |
3189 | ||
2de979bd | 3190 | if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 || |
9d729f72 | 3191 | get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS) |
1da177e4 LT |
3192 | tcp_store_ts_recent(tp); |
3193 | } | |
3194 | } | |
3195 | ||
3196 | /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM | |
3197 | * | |
3198 | * It is not fatal. If this ACK does _not_ change critical state (seqs, window) | |
3199 | * it can pass through stack. So, the following predicate verifies that | |
3200 | * this segment is not used for anything but congestion avoidance or | |
3201 | * fast retransmit. Moreover, we even are able to eliminate most of such | |
3202 | * second order effects, if we apply some small "replay" window (~RTO) | |
3203 | * to timestamp space. | |
3204 | * | |
3205 | * All these measures still do not guarantee that we reject wrapped ACKs | |
3206 | * on networks with high bandwidth, when sequence space is recycled fastly, | |
3207 | * but it guarantees that such events will be very rare and do not affect | |
3208 | * connection seriously. This doesn't look nice, but alas, PAWS is really | |
3209 | * buggy extension. | |
3210 | * | |
3211 | * [ Later note. Even worse! It is buggy for segments _with_ data. RFC | |
3212 | * states that events when retransmit arrives after original data are rare. | |
3213 | * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is | |
3214 | * the biggest problem on large power networks even with minor reordering. | |
3215 | * OK, let's give it small replay window. If peer clock is even 1hz, it is safe | |
3216 | * up to bandwidth of 18Gigabit/sec. 8) ] | |
3217 | */ | |
3218 | ||
463c84b9 | 3219 | static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 3220 | { |
463c84b9 | 3221 | struct tcp_sock *tp = tcp_sk(sk); |
aa8223c7 | 3222 | struct tcphdr *th = tcp_hdr(skb); |
1da177e4 LT |
3223 | u32 seq = TCP_SKB_CB(skb)->seq; |
3224 | u32 ack = TCP_SKB_CB(skb)->ack_seq; | |
3225 | ||
3226 | return (/* 1. Pure ACK with correct sequence number. */ | |
3227 | (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && | |
3228 | ||
3229 | /* 2. ... and duplicate ACK. */ | |
3230 | ack == tp->snd_una && | |
3231 | ||
3232 | /* 3. ... and does not update window. */ | |
3233 | !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && | |
3234 | ||
3235 | /* 4. ... and sits in replay window. */ | |
463c84b9 | 3236 | (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); |
1da177e4 LT |
3237 | } |
3238 | ||
463c84b9 | 3239 | static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 3240 | { |
463c84b9 | 3241 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 3242 | return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW && |
9d729f72 | 3243 | get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS && |
463c84b9 | 3244 | !tcp_disordered_ack(sk, skb)); |
1da177e4 LT |
3245 | } |
3246 | ||
3247 | /* Check segment sequence number for validity. | |
3248 | * | |
3249 | * Segment controls are considered valid, if the segment | |
3250 | * fits to the window after truncation to the window. Acceptability | |
3251 | * of data (and SYN, FIN, of course) is checked separately. | |
3252 | * See tcp_data_queue(), for example. | |
3253 | * | |
3254 | * Also, controls (RST is main one) are accepted using RCV.WUP instead | |
3255 | * of RCV.NXT. Peer still did not advance his SND.UNA when we | |
3256 | * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. | |
3257 | * (borrowed from freebsd) | |
3258 | */ | |
3259 | ||
3260 | static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq) | |
3261 | { | |
3262 | return !before(end_seq, tp->rcv_wup) && | |
3263 | !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); | |
3264 | } | |
3265 | ||
3266 | /* When we get a reset we do this. */ | |
3267 | static void tcp_reset(struct sock *sk) | |
3268 | { | |
3269 | /* We want the right error as BSD sees it (and indeed as we do). */ | |
3270 | switch (sk->sk_state) { | |
3271 | case TCP_SYN_SENT: | |
3272 | sk->sk_err = ECONNREFUSED; | |
3273 | break; | |
3274 | case TCP_CLOSE_WAIT: | |
3275 | sk->sk_err = EPIPE; | |
3276 | break; | |
3277 | case TCP_CLOSE: | |
3278 | return; | |
3279 | default: | |
3280 | sk->sk_err = ECONNRESET; | |
3281 | } | |
3282 | ||
3283 | if (!sock_flag(sk, SOCK_DEAD)) | |
3284 | sk->sk_error_report(sk); | |
3285 | ||
3286 | tcp_done(sk); | |
3287 | } | |
3288 | ||
3289 | /* | |
3290 | * Process the FIN bit. This now behaves as it is supposed to work | |
3291 | * and the FIN takes effect when it is validly part of sequence | |
3292 | * space. Not before when we get holes. | |
3293 | * | |
3294 | * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT | |
3295 | * (and thence onto LAST-ACK and finally, CLOSE, we never enter | |
3296 | * TIME-WAIT) | |
3297 | * | |
3298 | * If we are in FINWAIT-1, a received FIN indicates simultaneous | |
3299 | * close and we go into CLOSING (and later onto TIME-WAIT) | |
3300 | * | |
3301 | * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. | |
3302 | */ | |
3303 | static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th) | |
3304 | { | |
3305 | struct tcp_sock *tp = tcp_sk(sk); | |
3306 | ||
463c84b9 | 3307 | inet_csk_schedule_ack(sk); |
1da177e4 LT |
3308 | |
3309 | sk->sk_shutdown |= RCV_SHUTDOWN; | |
3310 | sock_set_flag(sk, SOCK_DONE); | |
3311 | ||
3312 | switch (sk->sk_state) { | |
3313 | case TCP_SYN_RECV: | |
3314 | case TCP_ESTABLISHED: | |
3315 | /* Move to CLOSE_WAIT */ | |
3316 | tcp_set_state(sk, TCP_CLOSE_WAIT); | |
463c84b9 | 3317 | inet_csk(sk)->icsk_ack.pingpong = 1; |
1da177e4 LT |
3318 | break; |
3319 | ||
3320 | case TCP_CLOSE_WAIT: | |
3321 | case TCP_CLOSING: | |
3322 | /* Received a retransmission of the FIN, do | |
3323 | * nothing. | |
3324 | */ | |
3325 | break; | |
3326 | case TCP_LAST_ACK: | |
3327 | /* RFC793: Remain in the LAST-ACK state. */ | |
3328 | break; | |
3329 | ||
3330 | case TCP_FIN_WAIT1: | |
3331 | /* This case occurs when a simultaneous close | |
3332 | * happens, we must ack the received FIN and | |
3333 | * enter the CLOSING state. | |
3334 | */ | |
3335 | tcp_send_ack(sk); | |
3336 | tcp_set_state(sk, TCP_CLOSING); | |
3337 | break; | |
3338 | case TCP_FIN_WAIT2: | |
3339 | /* Received a FIN -- send ACK and enter TIME_WAIT. */ | |
3340 | tcp_send_ack(sk); | |
3341 | tcp_time_wait(sk, TCP_TIME_WAIT, 0); | |
3342 | break; | |
3343 | default: | |
3344 | /* Only TCP_LISTEN and TCP_CLOSE are left, in these | |
3345 | * cases we should never reach this piece of code. | |
3346 | */ | |
3347 | printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n", | |
3348 | __FUNCTION__, sk->sk_state); | |
3349 | break; | |
3ff50b79 | 3350 | } |
1da177e4 LT |
3351 | |
3352 | /* It _is_ possible, that we have something out-of-order _after_ FIN. | |
3353 | * Probably, we should reset in this case. For now drop them. | |
3354 | */ | |
3355 | __skb_queue_purge(&tp->out_of_order_queue); | |
e60402d0 | 3356 | if (tcp_is_sack(tp)) |
1da177e4 LT |
3357 | tcp_sack_reset(&tp->rx_opt); |
3358 | sk_stream_mem_reclaim(sk); | |
3359 | ||
3360 | if (!sock_flag(sk, SOCK_DEAD)) { | |
3361 | sk->sk_state_change(sk); | |
3362 | ||
3363 | /* Do not send POLL_HUP for half duplex close. */ | |
3364 | if (sk->sk_shutdown == SHUTDOWN_MASK || | |
3365 | sk->sk_state == TCP_CLOSE) | |
3366 | sk_wake_async(sk, 1, POLL_HUP); | |
3367 | else | |
3368 | sk_wake_async(sk, 1, POLL_IN); | |
3369 | } | |
3370 | } | |
3371 | ||
40efc6fa | 3372 | static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq) |
1da177e4 LT |
3373 | { |
3374 | if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { | |
3375 | if (before(seq, sp->start_seq)) | |
3376 | sp->start_seq = seq; | |
3377 | if (after(end_seq, sp->end_seq)) | |
3378 | sp->end_seq = end_seq; | |
3379 | return 1; | |
3380 | } | |
3381 | return 0; | |
3382 | } | |
3383 | ||
40efc6fa | 3384 | static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq) |
1da177e4 | 3385 | { |
e60402d0 | 3386 | if (tcp_is_sack(tp) && sysctl_tcp_dsack) { |
1da177e4 LT |
3387 | if (before(seq, tp->rcv_nxt)) |
3388 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT); | |
3389 | else | |
3390 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT); | |
3391 | ||
3392 | tp->rx_opt.dsack = 1; | |
3393 | tp->duplicate_sack[0].start_seq = seq; | |
3394 | tp->duplicate_sack[0].end_seq = end_seq; | |
3395 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok); | |
3396 | } | |
3397 | } | |
3398 | ||
40efc6fa | 3399 | static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq) |
1da177e4 LT |
3400 | { |
3401 | if (!tp->rx_opt.dsack) | |
3402 | tcp_dsack_set(tp, seq, end_seq); | |
3403 | else | |
3404 | tcp_sack_extend(tp->duplicate_sack, seq, end_seq); | |
3405 | } | |
3406 | ||
3407 | static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb) | |
3408 | { | |
3409 | struct tcp_sock *tp = tcp_sk(sk); | |
3410 | ||
3411 | if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && | |
3412 | before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
3413 | NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); | |
463c84b9 | 3414 | tcp_enter_quickack_mode(sk); |
1da177e4 | 3415 | |
e60402d0 | 3416 | if (tcp_is_sack(tp) && sysctl_tcp_dsack) { |
1da177e4 LT |
3417 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; |
3418 | ||
3419 | if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) | |
3420 | end_seq = tp->rcv_nxt; | |
3421 | tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq); | |
3422 | } | |
3423 | } | |
3424 | ||
3425 | tcp_send_ack(sk); | |
3426 | } | |
3427 | ||
3428 | /* These routines update the SACK block as out-of-order packets arrive or | |
3429 | * in-order packets close up the sequence space. | |
3430 | */ | |
3431 | static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) | |
3432 | { | |
3433 | int this_sack; | |
3434 | struct tcp_sack_block *sp = &tp->selective_acks[0]; | |
3435 | struct tcp_sack_block *swalk = sp+1; | |
3436 | ||
3437 | /* See if the recent change to the first SACK eats into | |
3438 | * or hits the sequence space of other SACK blocks, if so coalesce. | |
3439 | */ | |
3440 | for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) { | |
3441 | if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { | |
3442 | int i; | |
3443 | ||
3444 | /* Zap SWALK, by moving every further SACK up by one slot. | |
3445 | * Decrease num_sacks. | |
3446 | */ | |
3447 | tp->rx_opt.num_sacks--; | |
3448 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); | |
2de979bd | 3449 | for (i=this_sack; i < tp->rx_opt.num_sacks; i++) |
1da177e4 LT |
3450 | sp[i] = sp[i+1]; |
3451 | continue; | |
3452 | } | |
3453 | this_sack++, swalk++; | |
3454 | } | |
3455 | } | |
3456 | ||
40efc6fa | 3457 | static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2) |
1da177e4 LT |
3458 | { |
3459 | __u32 tmp; | |
3460 | ||
3461 | tmp = sack1->start_seq; | |
3462 | sack1->start_seq = sack2->start_seq; | |
3463 | sack2->start_seq = tmp; | |
3464 | ||
3465 | tmp = sack1->end_seq; | |
3466 | sack1->end_seq = sack2->end_seq; | |
3467 | sack2->end_seq = tmp; | |
3468 | } | |
3469 | ||
3470 | static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) | |
3471 | { | |
3472 | struct tcp_sock *tp = tcp_sk(sk); | |
3473 | struct tcp_sack_block *sp = &tp->selective_acks[0]; | |
3474 | int cur_sacks = tp->rx_opt.num_sacks; | |
3475 | int this_sack; | |
3476 | ||
3477 | if (!cur_sacks) | |
3478 | goto new_sack; | |
3479 | ||
3480 | for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) { | |
3481 | if (tcp_sack_extend(sp, seq, end_seq)) { | |
3482 | /* Rotate this_sack to the first one. */ | |
3483 | for (; this_sack>0; this_sack--, sp--) | |
3484 | tcp_sack_swap(sp, sp-1); | |
3485 | if (cur_sacks > 1) | |
3486 | tcp_sack_maybe_coalesce(tp); | |
3487 | return; | |
3488 | } | |
3489 | } | |
3490 | ||
3491 | /* Could not find an adjacent existing SACK, build a new one, | |
3492 | * put it at the front, and shift everyone else down. We | |
3493 | * always know there is at least one SACK present already here. | |
3494 | * | |
3495 | * If the sack array is full, forget about the last one. | |
3496 | */ | |
3497 | if (this_sack >= 4) { | |
3498 | this_sack--; | |
3499 | tp->rx_opt.num_sacks--; | |
3500 | sp--; | |
3501 | } | |
2de979bd | 3502 | for (; this_sack > 0; this_sack--, sp--) |
1da177e4 LT |
3503 | *sp = *(sp-1); |
3504 | ||
3505 | new_sack: | |
3506 | /* Build the new head SACK, and we're done. */ | |
3507 | sp->start_seq = seq; | |
3508 | sp->end_seq = end_seq; | |
3509 | tp->rx_opt.num_sacks++; | |
3510 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); | |
3511 | } | |
3512 | ||
3513 | /* RCV.NXT advances, some SACKs should be eaten. */ | |
3514 | ||
3515 | static void tcp_sack_remove(struct tcp_sock *tp) | |
3516 | { | |
3517 | struct tcp_sack_block *sp = &tp->selective_acks[0]; | |
3518 | int num_sacks = tp->rx_opt.num_sacks; | |
3519 | int this_sack; | |
3520 | ||
3521 | /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ | |
b03efcfb | 3522 | if (skb_queue_empty(&tp->out_of_order_queue)) { |
1da177e4 LT |
3523 | tp->rx_opt.num_sacks = 0; |
3524 | tp->rx_opt.eff_sacks = tp->rx_opt.dsack; | |
3525 | return; | |
3526 | } | |
3527 | ||
2de979bd | 3528 | for (this_sack = 0; this_sack < num_sacks; ) { |
1da177e4 LT |
3529 | /* Check if the start of the sack is covered by RCV.NXT. */ |
3530 | if (!before(tp->rcv_nxt, sp->start_seq)) { | |
3531 | int i; | |
3532 | ||
3533 | /* RCV.NXT must cover all the block! */ | |
3534 | BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq)); | |
3535 | ||
3536 | /* Zap this SACK, by moving forward any other SACKS. */ | |
3537 | for (i=this_sack+1; i < num_sacks; i++) | |
3538 | tp->selective_acks[i-1] = tp->selective_acks[i]; | |
3539 | num_sacks--; | |
3540 | continue; | |
3541 | } | |
3542 | this_sack++; | |
3543 | sp++; | |
3544 | } | |
3545 | if (num_sacks != tp->rx_opt.num_sacks) { | |
3546 | tp->rx_opt.num_sacks = num_sacks; | |
3547 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); | |
3548 | } | |
3549 | } | |
3550 | ||
3551 | /* This one checks to see if we can put data from the | |
3552 | * out_of_order queue into the receive_queue. | |
3553 | */ | |
3554 | static void tcp_ofo_queue(struct sock *sk) | |
3555 | { | |
3556 | struct tcp_sock *tp = tcp_sk(sk); | |
3557 | __u32 dsack_high = tp->rcv_nxt; | |
3558 | struct sk_buff *skb; | |
3559 | ||
3560 | while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { | |
3561 | if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) | |
3562 | break; | |
3563 | ||
3564 | if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { | |
3565 | __u32 dsack = dsack_high; | |
3566 | if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) | |
3567 | dsack_high = TCP_SKB_CB(skb)->end_seq; | |
3568 | tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack); | |
3569 | } | |
3570 | ||
3571 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { | |
3572 | SOCK_DEBUG(sk, "ofo packet was already received \n"); | |
8728b834 | 3573 | __skb_unlink(skb, &tp->out_of_order_queue); |
1da177e4 LT |
3574 | __kfree_skb(skb); |
3575 | continue; | |
3576 | } | |
3577 | SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", | |
3578 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, | |
3579 | TCP_SKB_CB(skb)->end_seq); | |
3580 | ||
8728b834 | 3581 | __skb_unlink(skb, &tp->out_of_order_queue); |
1da177e4 LT |
3582 | __skb_queue_tail(&sk->sk_receive_queue, skb); |
3583 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
aa8223c7 ACM |
3584 | if (tcp_hdr(skb)->fin) |
3585 | tcp_fin(skb, sk, tcp_hdr(skb)); | |
1da177e4 LT |
3586 | } |
3587 | } | |
3588 | ||
3589 | static int tcp_prune_queue(struct sock *sk); | |
3590 | ||
3591 | static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) | |
3592 | { | |
aa8223c7 | 3593 | struct tcphdr *th = tcp_hdr(skb); |
1da177e4 LT |
3594 | struct tcp_sock *tp = tcp_sk(sk); |
3595 | int eaten = -1; | |
3596 | ||
3597 | if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) | |
3598 | goto drop; | |
3599 | ||
1da177e4 LT |
3600 | __skb_pull(skb, th->doff*4); |
3601 | ||
3602 | TCP_ECN_accept_cwr(tp, skb); | |
3603 | ||
3604 | if (tp->rx_opt.dsack) { | |
3605 | tp->rx_opt.dsack = 0; | |
3606 | tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks, | |
3607 | 4 - tp->rx_opt.tstamp_ok); | |
3608 | } | |
3609 | ||
3610 | /* Queue data for delivery to the user. | |
3611 | * Packets in sequence go to the receive queue. | |
3612 | * Out of sequence packets to the out_of_order_queue. | |
3613 | */ | |
3614 | if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { | |
3615 | if (tcp_receive_window(tp) == 0) | |
3616 | goto out_of_window; | |
3617 | ||
3618 | /* Ok. In sequence. In window. */ | |
3619 | if (tp->ucopy.task == current && | |
3620 | tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && | |
3621 | sock_owned_by_user(sk) && !tp->urg_data) { | |
3622 | int chunk = min_t(unsigned int, skb->len, | |
3623 | tp->ucopy.len); | |
3624 | ||
3625 | __set_current_state(TASK_RUNNING); | |
3626 | ||
3627 | local_bh_enable(); | |
3628 | if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { | |
3629 | tp->ucopy.len -= chunk; | |
3630 | tp->copied_seq += chunk; | |
3631 | eaten = (chunk == skb->len && !th->fin); | |
3632 | tcp_rcv_space_adjust(sk); | |
3633 | } | |
3634 | local_bh_disable(); | |
3635 | } | |
3636 | ||
3637 | if (eaten <= 0) { | |
3638 | queue_and_out: | |
3639 | if (eaten < 0 && | |
3640 | (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || | |
3641 | !sk_stream_rmem_schedule(sk, skb))) { | |
3642 | if (tcp_prune_queue(sk) < 0 || | |
3643 | !sk_stream_rmem_schedule(sk, skb)) | |
3644 | goto drop; | |
3645 | } | |
3646 | sk_stream_set_owner_r(skb, sk); | |
3647 | __skb_queue_tail(&sk->sk_receive_queue, skb); | |
3648 | } | |
3649 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
2de979bd | 3650 | if (skb->len) |
9e412ba7 | 3651 | tcp_event_data_recv(sk, skb); |
2de979bd | 3652 | if (th->fin) |
1da177e4 LT |
3653 | tcp_fin(skb, sk, th); |
3654 | ||
b03efcfb | 3655 | if (!skb_queue_empty(&tp->out_of_order_queue)) { |
1da177e4 LT |
3656 | tcp_ofo_queue(sk); |
3657 | ||
3658 | /* RFC2581. 4.2. SHOULD send immediate ACK, when | |
3659 | * gap in queue is filled. | |
3660 | */ | |
b03efcfb | 3661 | if (skb_queue_empty(&tp->out_of_order_queue)) |
463c84b9 | 3662 | inet_csk(sk)->icsk_ack.pingpong = 0; |
1da177e4 LT |
3663 | } |
3664 | ||
3665 | if (tp->rx_opt.num_sacks) | |
3666 | tcp_sack_remove(tp); | |
3667 | ||
9e412ba7 | 3668 | tcp_fast_path_check(sk); |
1da177e4 LT |
3669 | |
3670 | if (eaten > 0) | |
3671 | __kfree_skb(skb); | |
3672 | else if (!sock_flag(sk, SOCK_DEAD)) | |
3673 | sk->sk_data_ready(sk, 0); | |
3674 | return; | |
3675 | } | |
3676 | ||
3677 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { | |
3678 | /* A retransmit, 2nd most common case. Force an immediate ack. */ | |
3679 | NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); | |
3680 | tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); | |
3681 | ||
3682 | out_of_window: | |
463c84b9 ACM |
3683 | tcp_enter_quickack_mode(sk); |
3684 | inet_csk_schedule_ack(sk); | |
1da177e4 LT |
3685 | drop: |
3686 | __kfree_skb(skb); | |
3687 | return; | |
3688 | } | |
3689 | ||
3690 | /* Out of window. F.e. zero window probe. */ | |
3691 | if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) | |
3692 | goto out_of_window; | |
3693 | ||
463c84b9 | 3694 | tcp_enter_quickack_mode(sk); |
1da177e4 LT |
3695 | |
3696 | if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
3697 | /* Partial packet, seq < rcv_next < end_seq */ | |
3698 | SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", | |
3699 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, | |
3700 | TCP_SKB_CB(skb)->end_seq); | |
3701 | ||
3702 | tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); | |
e905a9ed | 3703 | |
1da177e4 LT |
3704 | /* If window is closed, drop tail of packet. But after |
3705 | * remembering D-SACK for its head made in previous line. | |
3706 | */ | |
3707 | if (!tcp_receive_window(tp)) | |
3708 | goto out_of_window; | |
3709 | goto queue_and_out; | |
3710 | } | |
3711 | ||
3712 | TCP_ECN_check_ce(tp, skb); | |
3713 | ||
3714 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || | |
3715 | !sk_stream_rmem_schedule(sk, skb)) { | |
3716 | if (tcp_prune_queue(sk) < 0 || | |
3717 | !sk_stream_rmem_schedule(sk, skb)) | |
3718 | goto drop; | |
3719 | } | |
3720 | ||
3721 | /* Disable header prediction. */ | |
3722 | tp->pred_flags = 0; | |
463c84b9 | 3723 | inet_csk_schedule_ack(sk); |
1da177e4 LT |
3724 | |
3725 | SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", | |
3726 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); | |
3727 | ||
3728 | sk_stream_set_owner_r(skb, sk); | |
3729 | ||
3730 | if (!skb_peek(&tp->out_of_order_queue)) { | |
3731 | /* Initial out of order segment, build 1 SACK. */ | |
e60402d0 | 3732 | if (tcp_is_sack(tp)) { |
1da177e4 LT |
3733 | tp->rx_opt.num_sacks = 1; |
3734 | tp->rx_opt.dsack = 0; | |
3735 | tp->rx_opt.eff_sacks = 1; | |
3736 | tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; | |
3737 | tp->selective_acks[0].end_seq = | |
3738 | TCP_SKB_CB(skb)->end_seq; | |
3739 | } | |
3740 | __skb_queue_head(&tp->out_of_order_queue,skb); | |
3741 | } else { | |
3742 | struct sk_buff *skb1 = tp->out_of_order_queue.prev; | |
3743 | u32 seq = TCP_SKB_CB(skb)->seq; | |
3744 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
3745 | ||
3746 | if (seq == TCP_SKB_CB(skb1)->end_seq) { | |
8728b834 | 3747 | __skb_append(skb1, skb, &tp->out_of_order_queue); |
1da177e4 LT |
3748 | |
3749 | if (!tp->rx_opt.num_sacks || | |
3750 | tp->selective_acks[0].end_seq != seq) | |
3751 | goto add_sack; | |
3752 | ||
3753 | /* Common case: data arrive in order after hole. */ | |
3754 | tp->selective_acks[0].end_seq = end_seq; | |
3755 | return; | |
3756 | } | |
3757 | ||
3758 | /* Find place to insert this segment. */ | |
3759 | do { | |
3760 | if (!after(TCP_SKB_CB(skb1)->seq, seq)) | |
3761 | break; | |
3762 | } while ((skb1 = skb1->prev) != | |
3763 | (struct sk_buff*)&tp->out_of_order_queue); | |
3764 | ||
3765 | /* Do skb overlap to previous one? */ | |
3766 | if (skb1 != (struct sk_buff*)&tp->out_of_order_queue && | |
3767 | before(seq, TCP_SKB_CB(skb1)->end_seq)) { | |
3768 | if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { | |
3769 | /* All the bits are present. Drop. */ | |
3770 | __kfree_skb(skb); | |
3771 | tcp_dsack_set(tp, seq, end_seq); | |
3772 | goto add_sack; | |
3773 | } | |
3774 | if (after(seq, TCP_SKB_CB(skb1)->seq)) { | |
3775 | /* Partial overlap. */ | |
3776 | tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq); | |
3777 | } else { | |
3778 | skb1 = skb1->prev; | |
3779 | } | |
3780 | } | |
3781 | __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue); | |
e905a9ed | 3782 | |
1da177e4 LT |
3783 | /* And clean segments covered by new one as whole. */ |
3784 | while ((skb1 = skb->next) != | |
3785 | (struct sk_buff*)&tp->out_of_order_queue && | |
3786 | after(end_seq, TCP_SKB_CB(skb1)->seq)) { | |
3787 | if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { | |
3788 | tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq); | |
3789 | break; | |
3790 | } | |
8728b834 | 3791 | __skb_unlink(skb1, &tp->out_of_order_queue); |
1da177e4 LT |
3792 | tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq); |
3793 | __kfree_skb(skb1); | |
3794 | } | |
3795 | ||
3796 | add_sack: | |
e60402d0 | 3797 | if (tcp_is_sack(tp)) |
1da177e4 LT |
3798 | tcp_sack_new_ofo_skb(sk, seq, end_seq); |
3799 | } | |
3800 | } | |
3801 | ||
3802 | /* Collapse contiguous sequence of skbs head..tail with | |
3803 | * sequence numbers start..end. | |
3804 | * Segments with FIN/SYN are not collapsed (only because this | |
3805 | * simplifies code) | |
3806 | */ | |
3807 | static void | |
8728b834 DM |
3808 | tcp_collapse(struct sock *sk, struct sk_buff_head *list, |
3809 | struct sk_buff *head, struct sk_buff *tail, | |
3810 | u32 start, u32 end) | |
1da177e4 LT |
3811 | { |
3812 | struct sk_buff *skb; | |
3813 | ||
caa20d9a | 3814 | /* First, check that queue is collapsible and find |
1da177e4 LT |
3815 | * the point where collapsing can be useful. */ |
3816 | for (skb = head; skb != tail; ) { | |
3817 | /* No new bits? It is possible on ofo queue. */ | |
3818 | if (!before(start, TCP_SKB_CB(skb)->end_seq)) { | |
3819 | struct sk_buff *next = skb->next; | |
8728b834 | 3820 | __skb_unlink(skb, list); |
1da177e4 LT |
3821 | __kfree_skb(skb); |
3822 | NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); | |
3823 | skb = next; | |
3824 | continue; | |
3825 | } | |
3826 | ||
3827 | /* The first skb to collapse is: | |
3828 | * - not SYN/FIN and | |
3829 | * - bloated or contains data before "start" or | |
3830 | * overlaps to the next one. | |
3831 | */ | |
aa8223c7 | 3832 | if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin && |
1da177e4 LT |
3833 | (tcp_win_from_space(skb->truesize) > skb->len || |
3834 | before(TCP_SKB_CB(skb)->seq, start) || | |
3835 | (skb->next != tail && | |
3836 | TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq))) | |
3837 | break; | |
3838 | ||
3839 | /* Decided to skip this, advance start seq. */ | |
3840 | start = TCP_SKB_CB(skb)->end_seq; | |
3841 | skb = skb->next; | |
3842 | } | |
aa8223c7 | 3843 | if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin) |
1da177e4 LT |
3844 | return; |
3845 | ||
3846 | while (before(start, end)) { | |
3847 | struct sk_buff *nskb; | |
3848 | int header = skb_headroom(skb); | |
3849 | int copy = SKB_MAX_ORDER(header, 0); | |
3850 | ||
3851 | /* Too big header? This can happen with IPv6. */ | |
3852 | if (copy < 0) | |
3853 | return; | |
3854 | if (end-start < copy) | |
3855 | copy = end-start; | |
3856 | nskb = alloc_skb(copy+header, GFP_ATOMIC); | |
3857 | if (!nskb) | |
3858 | return; | |
c51957da | 3859 | |
98e399f8 | 3860 | skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head); |
9c70220b ACM |
3861 | skb_set_network_header(nskb, (skb_network_header(skb) - |
3862 | skb->head)); | |
3863 | skb_set_transport_header(nskb, (skb_transport_header(skb) - | |
3864 | skb->head)); | |
1da177e4 LT |
3865 | skb_reserve(nskb, header); |
3866 | memcpy(nskb->head, skb->head, header); | |
1da177e4 LT |
3867 | memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); |
3868 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; | |
8728b834 | 3869 | __skb_insert(nskb, skb->prev, skb, list); |
1da177e4 LT |
3870 | sk_stream_set_owner_r(nskb, sk); |
3871 | ||
3872 | /* Copy data, releasing collapsed skbs. */ | |
3873 | while (copy > 0) { | |
3874 | int offset = start - TCP_SKB_CB(skb)->seq; | |
3875 | int size = TCP_SKB_CB(skb)->end_seq - start; | |
3876 | ||
09a62660 | 3877 | BUG_ON(offset < 0); |
1da177e4 LT |
3878 | if (size > 0) { |
3879 | size = min(copy, size); | |
3880 | if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) | |
3881 | BUG(); | |
3882 | TCP_SKB_CB(nskb)->end_seq += size; | |
3883 | copy -= size; | |
3884 | start += size; | |
3885 | } | |
3886 | if (!before(start, TCP_SKB_CB(skb)->end_seq)) { | |
3887 | struct sk_buff *next = skb->next; | |
8728b834 | 3888 | __skb_unlink(skb, list); |
1da177e4 LT |
3889 | __kfree_skb(skb); |
3890 | NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); | |
3891 | skb = next; | |
aa8223c7 ACM |
3892 | if (skb == tail || |
3893 | tcp_hdr(skb)->syn || | |
3894 | tcp_hdr(skb)->fin) | |
1da177e4 LT |
3895 | return; |
3896 | } | |
3897 | } | |
3898 | } | |
3899 | } | |
3900 | ||
3901 | /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs | |
3902 | * and tcp_collapse() them until all the queue is collapsed. | |
3903 | */ | |
3904 | static void tcp_collapse_ofo_queue(struct sock *sk) | |
3905 | { | |
3906 | struct tcp_sock *tp = tcp_sk(sk); | |
3907 | struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); | |
3908 | struct sk_buff *head; | |
3909 | u32 start, end; | |
3910 | ||
3911 | if (skb == NULL) | |
3912 | return; | |
3913 | ||
3914 | start = TCP_SKB_CB(skb)->seq; | |
3915 | end = TCP_SKB_CB(skb)->end_seq; | |
3916 | head = skb; | |
3917 | ||
3918 | for (;;) { | |
3919 | skb = skb->next; | |
3920 | ||
3921 | /* Segment is terminated when we see gap or when | |
3922 | * we are at the end of all the queue. */ | |
3923 | if (skb == (struct sk_buff *)&tp->out_of_order_queue || | |
3924 | after(TCP_SKB_CB(skb)->seq, end) || | |
3925 | before(TCP_SKB_CB(skb)->end_seq, start)) { | |
8728b834 DM |
3926 | tcp_collapse(sk, &tp->out_of_order_queue, |
3927 | head, skb, start, end); | |
1da177e4 LT |
3928 | head = skb; |
3929 | if (skb == (struct sk_buff *)&tp->out_of_order_queue) | |
3930 | break; | |
3931 | /* Start new segment */ | |
3932 | start = TCP_SKB_CB(skb)->seq; | |
3933 | end = TCP_SKB_CB(skb)->end_seq; | |
3934 | } else { | |
3935 | if (before(TCP_SKB_CB(skb)->seq, start)) | |
3936 | start = TCP_SKB_CB(skb)->seq; | |
3937 | if (after(TCP_SKB_CB(skb)->end_seq, end)) | |
3938 | end = TCP_SKB_CB(skb)->end_seq; | |
3939 | } | |
3940 | } | |
3941 | } | |
3942 | ||
3943 | /* Reduce allocated memory if we can, trying to get | |
3944 | * the socket within its memory limits again. | |
3945 | * | |
3946 | * Return less than zero if we should start dropping frames | |
3947 | * until the socket owning process reads some of the data | |
3948 | * to stabilize the situation. | |
3949 | */ | |
3950 | static int tcp_prune_queue(struct sock *sk) | |
3951 | { | |
e905a9ed | 3952 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
3953 | |
3954 | SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); | |
3955 | ||
3956 | NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED); | |
3957 | ||
3958 | if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) | |
9e412ba7 | 3959 | tcp_clamp_window(sk); |
1da177e4 LT |
3960 | else if (tcp_memory_pressure) |
3961 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); | |
3962 | ||
3963 | tcp_collapse_ofo_queue(sk); | |
8728b834 DM |
3964 | tcp_collapse(sk, &sk->sk_receive_queue, |
3965 | sk->sk_receive_queue.next, | |
1da177e4 LT |
3966 | (struct sk_buff*)&sk->sk_receive_queue, |
3967 | tp->copied_seq, tp->rcv_nxt); | |
3968 | sk_stream_mem_reclaim(sk); | |
3969 | ||
3970 | if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) | |
3971 | return 0; | |
3972 | ||
3973 | /* Collapsing did not help, destructive actions follow. | |
3974 | * This must not ever occur. */ | |
3975 | ||
3976 | /* First, purge the out_of_order queue. */ | |
b03efcfb DM |
3977 | if (!skb_queue_empty(&tp->out_of_order_queue)) { |
3978 | NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED); | |
1da177e4 LT |
3979 | __skb_queue_purge(&tp->out_of_order_queue); |
3980 | ||
3981 | /* Reset SACK state. A conforming SACK implementation will | |
3982 | * do the same at a timeout based retransmit. When a connection | |
3983 | * is in a sad state like this, we care only about integrity | |
3984 | * of the connection not performance. | |
3985 | */ | |
e60402d0 | 3986 | if (tcp_is_sack(tp)) |
1da177e4 LT |
3987 | tcp_sack_reset(&tp->rx_opt); |
3988 | sk_stream_mem_reclaim(sk); | |
3989 | } | |
3990 | ||
3991 | if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) | |
3992 | return 0; | |
3993 | ||
3994 | /* If we are really being abused, tell the caller to silently | |
3995 | * drop receive data on the floor. It will get retransmitted | |
3996 | * and hopefully then we'll have sufficient space. | |
3997 | */ | |
3998 | NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED); | |
3999 | ||
4000 | /* Massive buffer overcommit. */ | |
4001 | tp->pred_flags = 0; | |
4002 | return -1; | |
4003 | } | |
4004 | ||
4005 | ||
4006 | /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. | |
4007 | * As additional protections, we do not touch cwnd in retransmission phases, | |
4008 | * and if application hit its sndbuf limit recently. | |
4009 | */ | |
4010 | void tcp_cwnd_application_limited(struct sock *sk) | |
4011 | { | |
4012 | struct tcp_sock *tp = tcp_sk(sk); | |
4013 | ||
6687e988 | 4014 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && |
1da177e4 LT |
4015 | sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { |
4016 | /* Limited by application or receiver window. */ | |
d254bcdb IJ |
4017 | u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); |
4018 | u32 win_used = max(tp->snd_cwnd_used, init_win); | |
1da177e4 | 4019 | if (win_used < tp->snd_cwnd) { |
6687e988 | 4020 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
1da177e4 LT |
4021 | tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; |
4022 | } | |
4023 | tp->snd_cwnd_used = 0; | |
4024 | } | |
4025 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
4026 | } | |
4027 | ||
9e412ba7 | 4028 | static int tcp_should_expand_sndbuf(struct sock *sk) |
0d9901df | 4029 | { |
9e412ba7 IJ |
4030 | struct tcp_sock *tp = tcp_sk(sk); |
4031 | ||
0d9901df DM |
4032 | /* If the user specified a specific send buffer setting, do |
4033 | * not modify it. | |
4034 | */ | |
4035 | if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) | |
4036 | return 0; | |
4037 | ||
4038 | /* If we are under global TCP memory pressure, do not expand. */ | |
4039 | if (tcp_memory_pressure) | |
4040 | return 0; | |
4041 | ||
4042 | /* If we are under soft global TCP memory pressure, do not expand. */ | |
4043 | if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0]) | |
4044 | return 0; | |
4045 | ||
4046 | /* If we filled the congestion window, do not expand. */ | |
4047 | if (tp->packets_out >= tp->snd_cwnd) | |
4048 | return 0; | |
4049 | ||
4050 | return 1; | |
4051 | } | |
1da177e4 LT |
4052 | |
4053 | /* When incoming ACK allowed to free some skb from write_queue, | |
4054 | * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket | |
4055 | * on the exit from tcp input handler. | |
4056 | * | |
4057 | * PROBLEM: sndbuf expansion does not work well with largesend. | |
4058 | */ | |
4059 | static void tcp_new_space(struct sock *sk) | |
4060 | { | |
4061 | struct tcp_sock *tp = tcp_sk(sk); | |
4062 | ||
9e412ba7 | 4063 | if (tcp_should_expand_sndbuf(sk)) { |
e905a9ed | 4064 | int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + |
1da177e4 LT |
4065 | MAX_TCP_HEADER + 16 + sizeof(struct sk_buff), |
4066 | demanded = max_t(unsigned int, tp->snd_cwnd, | |
4067 | tp->reordering + 1); | |
4068 | sndmem *= 2*demanded; | |
4069 | if (sndmem > sk->sk_sndbuf) | |
4070 | sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); | |
4071 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
4072 | } | |
4073 | ||
4074 | sk->sk_write_space(sk); | |
4075 | } | |
4076 | ||
40efc6fa | 4077 | static void tcp_check_space(struct sock *sk) |
1da177e4 LT |
4078 | { |
4079 | if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { | |
4080 | sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); | |
4081 | if (sk->sk_socket && | |
4082 | test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) | |
4083 | tcp_new_space(sk); | |
4084 | } | |
4085 | } | |
4086 | ||
9e412ba7 | 4087 | static inline void tcp_data_snd_check(struct sock *sk) |
1da177e4 | 4088 | { |
9e412ba7 | 4089 | tcp_push_pending_frames(sk); |
1da177e4 LT |
4090 | tcp_check_space(sk); |
4091 | } | |
4092 | ||
4093 | /* | |
4094 | * Check if sending an ack is needed. | |
4095 | */ | |
4096 | static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) | |
4097 | { | |
4098 | struct tcp_sock *tp = tcp_sk(sk); | |
4099 | ||
4100 | /* More than one full frame received... */ | |
463c84b9 | 4101 | if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss |
1da177e4 LT |
4102 | /* ... and right edge of window advances far enough. |
4103 | * (tcp_recvmsg() will send ACK otherwise). Or... | |
4104 | */ | |
4105 | && __tcp_select_window(sk) >= tp->rcv_wnd) || | |
4106 | /* We ACK each frame or... */ | |
463c84b9 | 4107 | tcp_in_quickack_mode(sk) || |
1da177e4 LT |
4108 | /* We have out of order data. */ |
4109 | (ofo_possible && | |
4110 | skb_peek(&tp->out_of_order_queue))) { | |
4111 | /* Then ack it now */ | |
4112 | tcp_send_ack(sk); | |
4113 | } else { | |
4114 | /* Else, send delayed ack. */ | |
4115 | tcp_send_delayed_ack(sk); | |
4116 | } | |
4117 | } | |
4118 | ||
40efc6fa | 4119 | static inline void tcp_ack_snd_check(struct sock *sk) |
1da177e4 | 4120 | { |
463c84b9 | 4121 | if (!inet_csk_ack_scheduled(sk)) { |
1da177e4 LT |
4122 | /* We sent a data segment already. */ |
4123 | return; | |
4124 | } | |
4125 | __tcp_ack_snd_check(sk, 1); | |
4126 | } | |
4127 | ||
4128 | /* | |
4129 | * This routine is only called when we have urgent data | |
caa20d9a | 4130 | * signaled. Its the 'slow' part of tcp_urg. It could be |
1da177e4 LT |
4131 | * moved inline now as tcp_urg is only called from one |
4132 | * place. We handle URGent data wrong. We have to - as | |
4133 | * BSD still doesn't use the correction from RFC961. | |
4134 | * For 1003.1g we should support a new option TCP_STDURG to permit | |
4135 | * either form (or just set the sysctl tcp_stdurg). | |
4136 | */ | |
e905a9ed | 4137 | |
1da177e4 LT |
4138 | static void tcp_check_urg(struct sock * sk, struct tcphdr * th) |
4139 | { | |
4140 | struct tcp_sock *tp = tcp_sk(sk); | |
4141 | u32 ptr = ntohs(th->urg_ptr); | |
4142 | ||
4143 | if (ptr && !sysctl_tcp_stdurg) | |
4144 | ptr--; | |
4145 | ptr += ntohl(th->seq); | |
4146 | ||
4147 | /* Ignore urgent data that we've already seen and read. */ | |
4148 | if (after(tp->copied_seq, ptr)) | |
4149 | return; | |
4150 | ||
4151 | /* Do not replay urg ptr. | |
4152 | * | |
4153 | * NOTE: interesting situation not covered by specs. | |
4154 | * Misbehaving sender may send urg ptr, pointing to segment, | |
4155 | * which we already have in ofo queue. We are not able to fetch | |
4156 | * such data and will stay in TCP_URG_NOTYET until will be eaten | |
4157 | * by recvmsg(). Seems, we are not obliged to handle such wicked | |
4158 | * situations. But it is worth to think about possibility of some | |
4159 | * DoSes using some hypothetical application level deadlock. | |
4160 | */ | |
4161 | if (before(ptr, tp->rcv_nxt)) | |
4162 | return; | |
4163 | ||
4164 | /* Do we already have a newer (or duplicate) urgent pointer? */ | |
4165 | if (tp->urg_data && !after(ptr, tp->urg_seq)) | |
4166 | return; | |
4167 | ||
4168 | /* Tell the world about our new urgent pointer. */ | |
4169 | sk_send_sigurg(sk); | |
4170 | ||
4171 | /* We may be adding urgent data when the last byte read was | |
4172 | * urgent. To do this requires some care. We cannot just ignore | |
4173 | * tp->copied_seq since we would read the last urgent byte again | |
4174 | * as data, nor can we alter copied_seq until this data arrives | |
caa20d9a | 4175 | * or we break the semantics of SIOCATMARK (and thus sockatmark()) |
1da177e4 LT |
4176 | * |
4177 | * NOTE. Double Dutch. Rendering to plain English: author of comment | |
4178 | * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); | |
4179 | * and expect that both A and B disappear from stream. This is _wrong_. | |
4180 | * Though this happens in BSD with high probability, this is occasional. | |
4181 | * Any application relying on this is buggy. Note also, that fix "works" | |
4182 | * only in this artificial test. Insert some normal data between A and B and we will | |
4183 | * decline of BSD again. Verdict: it is better to remove to trap | |
4184 | * buggy users. | |
4185 | */ | |
4186 | if (tp->urg_seq == tp->copied_seq && tp->urg_data && | |
4187 | !sock_flag(sk, SOCK_URGINLINE) && | |
4188 | tp->copied_seq != tp->rcv_nxt) { | |
4189 | struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); | |
4190 | tp->copied_seq++; | |
4191 | if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { | |
8728b834 | 4192 | __skb_unlink(skb, &sk->sk_receive_queue); |
1da177e4 LT |
4193 | __kfree_skb(skb); |
4194 | } | |
4195 | } | |
4196 | ||
4197 | tp->urg_data = TCP_URG_NOTYET; | |
4198 | tp->urg_seq = ptr; | |
4199 | ||
4200 | /* Disable header prediction. */ | |
4201 | tp->pred_flags = 0; | |
4202 | } | |
4203 | ||
4204 | /* This is the 'fast' part of urgent handling. */ | |
4205 | static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th) | |
4206 | { | |
4207 | struct tcp_sock *tp = tcp_sk(sk); | |
4208 | ||
4209 | /* Check if we get a new urgent pointer - normally not. */ | |
4210 | if (th->urg) | |
4211 | tcp_check_urg(sk,th); | |
4212 | ||
4213 | /* Do we wait for any urgent data? - normally not... */ | |
4214 | if (tp->urg_data == TCP_URG_NOTYET) { | |
4215 | u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - | |
4216 | th->syn; | |
4217 | ||
e905a9ed | 4218 | /* Is the urgent pointer pointing into this packet? */ |
1da177e4 LT |
4219 | if (ptr < skb->len) { |
4220 | u8 tmp; | |
4221 | if (skb_copy_bits(skb, ptr, &tmp, 1)) | |
4222 | BUG(); | |
4223 | tp->urg_data = TCP_URG_VALID | tmp; | |
4224 | if (!sock_flag(sk, SOCK_DEAD)) | |
4225 | sk->sk_data_ready(sk, 0); | |
4226 | } | |
4227 | } | |
4228 | } | |
4229 | ||
4230 | static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) | |
4231 | { | |
4232 | struct tcp_sock *tp = tcp_sk(sk); | |
4233 | int chunk = skb->len - hlen; | |
4234 | int err; | |
4235 | ||
4236 | local_bh_enable(); | |
60476372 | 4237 | if (skb_csum_unnecessary(skb)) |
1da177e4 LT |
4238 | err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); |
4239 | else | |
4240 | err = skb_copy_and_csum_datagram_iovec(skb, hlen, | |
4241 | tp->ucopy.iov); | |
4242 | ||
4243 | if (!err) { | |
4244 | tp->ucopy.len -= chunk; | |
4245 | tp->copied_seq += chunk; | |
4246 | tcp_rcv_space_adjust(sk); | |
4247 | } | |
4248 | ||
4249 | local_bh_disable(); | |
4250 | return err; | |
4251 | } | |
4252 | ||
b51655b9 | 4253 | static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 4254 | { |
b51655b9 | 4255 | __sum16 result; |
1da177e4 LT |
4256 | |
4257 | if (sock_owned_by_user(sk)) { | |
4258 | local_bh_enable(); | |
4259 | result = __tcp_checksum_complete(skb); | |
4260 | local_bh_disable(); | |
4261 | } else { | |
4262 | result = __tcp_checksum_complete(skb); | |
4263 | } | |
4264 | return result; | |
4265 | } | |
4266 | ||
40efc6fa | 4267 | static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 4268 | { |
60476372 | 4269 | return !skb_csum_unnecessary(skb) && |
1da177e4 LT |
4270 | __tcp_checksum_complete_user(sk, skb); |
4271 | } | |
4272 | ||
1a2449a8 CL |
4273 | #ifdef CONFIG_NET_DMA |
4274 | static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen) | |
4275 | { | |
4276 | struct tcp_sock *tp = tcp_sk(sk); | |
4277 | int chunk = skb->len - hlen; | |
4278 | int dma_cookie; | |
4279 | int copied_early = 0; | |
4280 | ||
4281 | if (tp->ucopy.wakeup) | |
e905a9ed | 4282 | return 0; |
1a2449a8 CL |
4283 | |
4284 | if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) | |
4285 | tp->ucopy.dma_chan = get_softnet_dma(); | |
4286 | ||
60476372 | 4287 | if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) { |
1a2449a8 CL |
4288 | |
4289 | dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, | |
4290 | skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list); | |
4291 | ||
4292 | if (dma_cookie < 0) | |
4293 | goto out; | |
4294 | ||
4295 | tp->ucopy.dma_cookie = dma_cookie; | |
4296 | copied_early = 1; | |
4297 | ||
4298 | tp->ucopy.len -= chunk; | |
4299 | tp->copied_seq += chunk; | |
4300 | tcp_rcv_space_adjust(sk); | |
4301 | ||
4302 | if ((tp->ucopy.len == 0) || | |
aa8223c7 | 4303 | (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) || |
1a2449a8 CL |
4304 | (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { |
4305 | tp->ucopy.wakeup = 1; | |
4306 | sk->sk_data_ready(sk, 0); | |
4307 | } | |
4308 | } else if (chunk > 0) { | |
4309 | tp->ucopy.wakeup = 1; | |
4310 | sk->sk_data_ready(sk, 0); | |
4311 | } | |
4312 | out: | |
4313 | return copied_early; | |
4314 | } | |
4315 | #endif /* CONFIG_NET_DMA */ | |
4316 | ||
1da177e4 | 4317 | /* |
e905a9ed | 4318 | * TCP receive function for the ESTABLISHED state. |
1da177e4 | 4319 | * |
e905a9ed | 4320 | * It is split into a fast path and a slow path. The fast path is |
1da177e4 LT |
4321 | * disabled when: |
4322 | * - A zero window was announced from us - zero window probing | |
e905a9ed | 4323 | * is only handled properly in the slow path. |
1da177e4 LT |
4324 | * - Out of order segments arrived. |
4325 | * - Urgent data is expected. | |
4326 | * - There is no buffer space left | |
4327 | * - Unexpected TCP flags/window values/header lengths are received | |
e905a9ed | 4328 | * (detected by checking the TCP header against pred_flags) |
1da177e4 LT |
4329 | * - Data is sent in both directions. Fast path only supports pure senders |
4330 | * or pure receivers (this means either the sequence number or the ack | |
4331 | * value must stay constant) | |
4332 | * - Unexpected TCP option. | |
4333 | * | |
e905a9ed | 4334 | * When these conditions are not satisfied it drops into a standard |
1da177e4 LT |
4335 | * receive procedure patterned after RFC793 to handle all cases. |
4336 | * The first three cases are guaranteed by proper pred_flags setting, | |
e905a9ed | 4337 | * the rest is checked inline. Fast processing is turned on in |
1da177e4 LT |
4338 | * tcp_data_queue when everything is OK. |
4339 | */ | |
4340 | int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, | |
4341 | struct tcphdr *th, unsigned len) | |
4342 | { | |
4343 | struct tcp_sock *tp = tcp_sk(sk); | |
4344 | ||
4345 | /* | |
4346 | * Header prediction. | |
e905a9ed | 4347 | * The code loosely follows the one in the famous |
1da177e4 | 4348 | * "30 instruction TCP receive" Van Jacobson mail. |
e905a9ed YH |
4349 | * |
4350 | * Van's trick is to deposit buffers into socket queue | |
1da177e4 LT |
4351 | * on a device interrupt, to call tcp_recv function |
4352 | * on the receive process context and checksum and copy | |
4353 | * the buffer to user space. smart... | |
4354 | * | |
e905a9ed | 4355 | * Our current scheme is not silly either but we take the |
1da177e4 LT |
4356 | * extra cost of the net_bh soft interrupt processing... |
4357 | * We do checksum and copy also but from device to kernel. | |
4358 | */ | |
4359 | ||
4360 | tp->rx_opt.saw_tstamp = 0; | |
4361 | ||
4362 | /* pred_flags is 0xS?10 << 16 + snd_wnd | |
caa20d9a | 4363 | * if header_prediction is to be made |
1da177e4 LT |
4364 | * 'S' will always be tp->tcp_header_len >> 2 |
4365 | * '?' will be 0 for the fast path, otherwise pred_flags is 0 to | |
e905a9ed | 4366 | * turn it off (when there are holes in the receive |
1da177e4 LT |
4367 | * space for instance) |
4368 | * PSH flag is ignored. | |
4369 | */ | |
4370 | ||
4371 | if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && | |
4372 | TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { | |
4373 | int tcp_header_len = tp->tcp_header_len; | |
4374 | ||
4375 | /* Timestamp header prediction: tcp_header_len | |
4376 | * is automatically equal to th->doff*4 due to pred_flags | |
4377 | * match. | |
4378 | */ | |
4379 | ||
4380 | /* Check timestamp */ | |
4381 | if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { | |
4f3608b7 | 4382 | __be32 *ptr = (__be32 *)(th + 1); |
1da177e4 LT |
4383 | |
4384 | /* No? Slow path! */ | |
4f3608b7 | 4385 | if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
1da177e4 LT |
4386 | | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) |
4387 | goto slow_path; | |
4388 | ||
4389 | tp->rx_opt.saw_tstamp = 1; | |
e905a9ed | 4390 | ++ptr; |
1da177e4 LT |
4391 | tp->rx_opt.rcv_tsval = ntohl(*ptr); |
4392 | ++ptr; | |
4393 | tp->rx_opt.rcv_tsecr = ntohl(*ptr); | |
4394 | ||
4395 | /* If PAWS failed, check it more carefully in slow path */ | |
4396 | if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) | |
4397 | goto slow_path; | |
4398 | ||
4399 | /* DO NOT update ts_recent here, if checksum fails | |
4400 | * and timestamp was corrupted part, it will result | |
4401 | * in a hung connection since we will drop all | |
4402 | * future packets due to the PAWS test. | |
4403 | */ | |
4404 | } | |
4405 | ||
4406 | if (len <= tcp_header_len) { | |
4407 | /* Bulk data transfer: sender */ | |
4408 | if (len == tcp_header_len) { | |
4409 | /* Predicted packet is in window by definition. | |
4410 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. | |
4411 | * Hence, check seq<=rcv_wup reduces to: | |
4412 | */ | |
4413 | if (tcp_header_len == | |
4414 | (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && | |
4415 | tp->rcv_nxt == tp->rcv_wup) | |
4416 | tcp_store_ts_recent(tp); | |
4417 | ||
1da177e4 LT |
4418 | /* We know that such packets are checksummed |
4419 | * on entry. | |
4420 | */ | |
4421 | tcp_ack(sk, skb, 0); | |
e905a9ed | 4422 | __kfree_skb(skb); |
9e412ba7 | 4423 | tcp_data_snd_check(sk); |
1da177e4 LT |
4424 | return 0; |
4425 | } else { /* Header too small */ | |
4426 | TCP_INC_STATS_BH(TCP_MIB_INERRS); | |
4427 | goto discard; | |
4428 | } | |
4429 | } else { | |
4430 | int eaten = 0; | |
1a2449a8 | 4431 | int copied_early = 0; |
1da177e4 | 4432 | |
1a2449a8 CL |
4433 | if (tp->copied_seq == tp->rcv_nxt && |
4434 | len - tcp_header_len <= tp->ucopy.len) { | |
4435 | #ifdef CONFIG_NET_DMA | |
4436 | if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { | |
4437 | copied_early = 1; | |
4438 | eaten = 1; | |
4439 | } | |
4440 | #endif | |
4441 | if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) { | |
4442 | __set_current_state(TASK_RUNNING); | |
1da177e4 | 4443 | |
1a2449a8 CL |
4444 | if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) |
4445 | eaten = 1; | |
4446 | } | |
4447 | if (eaten) { | |
1da177e4 LT |
4448 | /* Predicted packet is in window by definition. |
4449 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. | |
4450 | * Hence, check seq<=rcv_wup reduces to: | |
4451 | */ | |
4452 | if (tcp_header_len == | |
4453 | (sizeof(struct tcphdr) + | |
4454 | TCPOLEN_TSTAMP_ALIGNED) && | |
4455 | tp->rcv_nxt == tp->rcv_wup) | |
4456 | tcp_store_ts_recent(tp); | |
4457 | ||
463c84b9 | 4458 | tcp_rcv_rtt_measure_ts(sk, skb); |
1da177e4 LT |
4459 | |
4460 | __skb_pull(skb, tcp_header_len); | |
4461 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
4462 | NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER); | |
1da177e4 | 4463 | } |
1a2449a8 CL |
4464 | if (copied_early) |
4465 | tcp_cleanup_rbuf(sk, skb->len); | |
1da177e4 LT |
4466 | } |
4467 | if (!eaten) { | |
4468 | if (tcp_checksum_complete_user(sk, skb)) | |
4469 | goto csum_error; | |
4470 | ||
4471 | /* Predicted packet is in window by definition. | |
4472 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. | |
4473 | * Hence, check seq<=rcv_wup reduces to: | |
4474 | */ | |
4475 | if (tcp_header_len == | |
4476 | (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && | |
4477 | tp->rcv_nxt == tp->rcv_wup) | |
4478 | tcp_store_ts_recent(tp); | |
4479 | ||
463c84b9 | 4480 | tcp_rcv_rtt_measure_ts(sk, skb); |
1da177e4 LT |
4481 | |
4482 | if ((int)skb->truesize > sk->sk_forward_alloc) | |
4483 | goto step5; | |
4484 | ||
4485 | NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS); | |
4486 | ||
4487 | /* Bulk data transfer: receiver */ | |
4488 | __skb_pull(skb,tcp_header_len); | |
4489 | __skb_queue_tail(&sk->sk_receive_queue, skb); | |
4490 | sk_stream_set_owner_r(skb, sk); | |
4491 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
4492 | } | |
4493 | ||
9e412ba7 | 4494 | tcp_event_data_recv(sk, skb); |
1da177e4 LT |
4495 | |
4496 | if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { | |
4497 | /* Well, only one small jumplet in fast path... */ | |
4498 | tcp_ack(sk, skb, FLAG_DATA); | |
9e412ba7 | 4499 | tcp_data_snd_check(sk); |
463c84b9 | 4500 | if (!inet_csk_ack_scheduled(sk)) |
1da177e4 LT |
4501 | goto no_ack; |
4502 | } | |
4503 | ||
31432412 | 4504 | __tcp_ack_snd_check(sk, 0); |
1da177e4 | 4505 | no_ack: |
1a2449a8 CL |
4506 | #ifdef CONFIG_NET_DMA |
4507 | if (copied_early) | |
4508 | __skb_queue_tail(&sk->sk_async_wait_queue, skb); | |
4509 | else | |
4510 | #endif | |
1da177e4 LT |
4511 | if (eaten) |
4512 | __kfree_skb(skb); | |
4513 | else | |
4514 | sk->sk_data_ready(sk, 0); | |
4515 | return 0; | |
4516 | } | |
4517 | } | |
4518 | ||
4519 | slow_path: | |
4520 | if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb)) | |
4521 | goto csum_error; | |
4522 | ||
4523 | /* | |
4524 | * RFC1323: H1. Apply PAWS check first. | |
4525 | */ | |
4526 | if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && | |
463c84b9 | 4527 | tcp_paws_discard(sk, skb)) { |
1da177e4 LT |
4528 | if (!th->rst) { |
4529 | NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); | |
4530 | tcp_send_dupack(sk, skb); | |
4531 | goto discard; | |
4532 | } | |
4533 | /* Resets are accepted even if PAWS failed. | |
4534 | ||
4535 | ts_recent update must be made after we are sure | |
4536 | that the packet is in window. | |
4537 | */ | |
4538 | } | |
4539 | ||
4540 | /* | |
4541 | * Standard slow path. | |
4542 | */ | |
4543 | ||
4544 | if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { | |
4545 | /* RFC793, page 37: "In all states except SYN-SENT, all reset | |
4546 | * (RST) segments are validated by checking their SEQ-fields." | |
4547 | * And page 69: "If an incoming segment is not acceptable, | |
4548 | * an acknowledgment should be sent in reply (unless the RST bit | |
4549 | * is set, if so drop the segment and return)". | |
4550 | */ | |
4551 | if (!th->rst) | |
4552 | tcp_send_dupack(sk, skb); | |
4553 | goto discard; | |
4554 | } | |
4555 | ||
2de979bd | 4556 | if (th->rst) { |
1da177e4 LT |
4557 | tcp_reset(sk); |
4558 | goto discard; | |
4559 | } | |
4560 | ||
4561 | tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); | |
4562 | ||
4563 | if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
4564 | TCP_INC_STATS_BH(TCP_MIB_INERRS); | |
4565 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); | |
4566 | tcp_reset(sk); | |
4567 | return 1; | |
4568 | } | |
4569 | ||
4570 | step5: | |
2de979bd | 4571 | if (th->ack) |
1da177e4 LT |
4572 | tcp_ack(sk, skb, FLAG_SLOWPATH); |
4573 | ||
463c84b9 | 4574 | tcp_rcv_rtt_measure_ts(sk, skb); |
1da177e4 LT |
4575 | |
4576 | /* Process urgent data. */ | |
4577 | tcp_urg(sk, skb, th); | |
4578 | ||
4579 | /* step 7: process the segment text */ | |
4580 | tcp_data_queue(sk, skb); | |
4581 | ||
9e412ba7 | 4582 | tcp_data_snd_check(sk); |
1da177e4 LT |
4583 | tcp_ack_snd_check(sk); |
4584 | return 0; | |
4585 | ||
4586 | csum_error: | |
4587 | TCP_INC_STATS_BH(TCP_MIB_INERRS); | |
4588 | ||
4589 | discard: | |
4590 | __kfree_skb(skb); | |
4591 | return 0; | |
4592 | } | |
4593 | ||
4594 | static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, | |
4595 | struct tcphdr *th, unsigned len) | |
4596 | { | |
4597 | struct tcp_sock *tp = tcp_sk(sk); | |
d83d8461 | 4598 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
4599 | int saved_clamp = tp->rx_opt.mss_clamp; |
4600 | ||
4601 | tcp_parse_options(skb, &tp->rx_opt, 0); | |
4602 | ||
4603 | if (th->ack) { | |
4604 | /* rfc793: | |
4605 | * "If the state is SYN-SENT then | |
4606 | * first check the ACK bit | |
4607 | * If the ACK bit is set | |
4608 | * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send | |
4609 | * a reset (unless the RST bit is set, if so drop | |
4610 | * the segment and return)" | |
4611 | * | |
4612 | * We do not send data with SYN, so that RFC-correct | |
4613 | * test reduces to: | |
4614 | */ | |
4615 | if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) | |
4616 | goto reset_and_undo; | |
4617 | ||
4618 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && | |
4619 | !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, | |
4620 | tcp_time_stamp)) { | |
4621 | NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED); | |
4622 | goto reset_and_undo; | |
4623 | } | |
4624 | ||
4625 | /* Now ACK is acceptable. | |
4626 | * | |
4627 | * "If the RST bit is set | |
4628 | * If the ACK was acceptable then signal the user "error: | |
4629 | * connection reset", drop the segment, enter CLOSED state, | |
4630 | * delete TCB, and return." | |
4631 | */ | |
4632 | ||
4633 | if (th->rst) { | |
4634 | tcp_reset(sk); | |
4635 | goto discard; | |
4636 | } | |
4637 | ||
4638 | /* rfc793: | |
4639 | * "fifth, if neither of the SYN or RST bits is set then | |
4640 | * drop the segment and return." | |
4641 | * | |
4642 | * See note below! | |
4643 | * --ANK(990513) | |
4644 | */ | |
4645 | if (!th->syn) | |
4646 | goto discard_and_undo; | |
4647 | ||
4648 | /* rfc793: | |
4649 | * "If the SYN bit is on ... | |
4650 | * are acceptable then ... | |
4651 | * (our SYN has been ACKed), change the connection | |
4652 | * state to ESTABLISHED..." | |
4653 | */ | |
4654 | ||
4655 | TCP_ECN_rcv_synack(tp, th); | |
1da177e4 LT |
4656 | |
4657 | tp->snd_wl1 = TCP_SKB_CB(skb)->seq; | |
4658 | tcp_ack(sk, skb, FLAG_SLOWPATH); | |
4659 | ||
4660 | /* Ok.. it's good. Set up sequence numbers and | |
4661 | * move to established. | |
4662 | */ | |
4663 | tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; | |
4664 | tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; | |
4665 | ||
4666 | /* RFC1323: The window in SYN & SYN/ACK segments is | |
4667 | * never scaled. | |
4668 | */ | |
4669 | tp->snd_wnd = ntohs(th->window); | |
4670 | tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq); | |
4671 | ||
4672 | if (!tp->rx_opt.wscale_ok) { | |
4673 | tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; | |
4674 | tp->window_clamp = min(tp->window_clamp, 65535U); | |
4675 | } | |
4676 | ||
4677 | if (tp->rx_opt.saw_tstamp) { | |
4678 | tp->rx_opt.tstamp_ok = 1; | |
4679 | tp->tcp_header_len = | |
4680 | sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; | |
4681 | tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; | |
4682 | tcp_store_ts_recent(tp); | |
4683 | } else { | |
4684 | tp->tcp_header_len = sizeof(struct tcphdr); | |
4685 | } | |
4686 | ||
e60402d0 IJ |
4687 | if (tcp_is_sack(tp) && sysctl_tcp_fack) |
4688 | tcp_enable_fack(tp); | |
1da177e4 | 4689 | |
5d424d5a | 4690 | tcp_mtup_init(sk); |
d83d8461 | 4691 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); |
1da177e4 LT |
4692 | tcp_initialize_rcv_mss(sk); |
4693 | ||
4694 | /* Remember, tcp_poll() does not lock socket! | |
4695 | * Change state from SYN-SENT only after copied_seq | |
4696 | * is initialized. */ | |
4697 | tp->copied_seq = tp->rcv_nxt; | |
e16aa207 | 4698 | smp_mb(); |
1da177e4 LT |
4699 | tcp_set_state(sk, TCP_ESTABLISHED); |
4700 | ||
6b877699 VY |
4701 | security_inet_conn_established(sk, skb); |
4702 | ||
1da177e4 | 4703 | /* Make sure socket is routed, for correct metrics. */ |
8292a17a | 4704 | icsk->icsk_af_ops->rebuild_header(sk); |
1da177e4 LT |
4705 | |
4706 | tcp_init_metrics(sk); | |
4707 | ||
6687e988 | 4708 | tcp_init_congestion_control(sk); |
317a76f9 | 4709 | |
1da177e4 LT |
4710 | /* Prevent spurious tcp_cwnd_restart() on first data |
4711 | * packet. | |
4712 | */ | |
4713 | tp->lsndtime = tcp_time_stamp; | |
4714 | ||
4715 | tcp_init_buffer_space(sk); | |
4716 | ||
4717 | if (sock_flag(sk, SOCK_KEEPOPEN)) | |
463c84b9 | 4718 | inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); |
1da177e4 LT |
4719 | |
4720 | if (!tp->rx_opt.snd_wscale) | |
4721 | __tcp_fast_path_on(tp, tp->snd_wnd); | |
4722 | else | |
4723 | tp->pred_flags = 0; | |
4724 | ||
4725 | if (!sock_flag(sk, SOCK_DEAD)) { | |
4726 | sk->sk_state_change(sk); | |
4727 | sk_wake_async(sk, 0, POLL_OUT); | |
4728 | } | |
4729 | ||
295f7324 ACM |
4730 | if (sk->sk_write_pending || |
4731 | icsk->icsk_accept_queue.rskq_defer_accept || | |
4732 | icsk->icsk_ack.pingpong) { | |
1da177e4 LT |
4733 | /* Save one ACK. Data will be ready after |
4734 | * several ticks, if write_pending is set. | |
4735 | * | |
4736 | * It may be deleted, but with this feature tcpdumps | |
4737 | * look so _wonderfully_ clever, that I was not able | |
4738 | * to stand against the temptation 8) --ANK | |
4739 | */ | |
463c84b9 | 4740 | inet_csk_schedule_ack(sk); |
295f7324 ACM |
4741 | icsk->icsk_ack.lrcvtime = tcp_time_stamp; |
4742 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
463c84b9 ACM |
4743 | tcp_incr_quickack(sk); |
4744 | tcp_enter_quickack_mode(sk); | |
3f421baa ACM |
4745 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, |
4746 | TCP_DELACK_MAX, TCP_RTO_MAX); | |
1da177e4 LT |
4747 | |
4748 | discard: | |
4749 | __kfree_skb(skb); | |
4750 | return 0; | |
4751 | } else { | |
4752 | tcp_send_ack(sk); | |
4753 | } | |
4754 | return -1; | |
4755 | } | |
4756 | ||
4757 | /* No ACK in the segment */ | |
4758 | ||
4759 | if (th->rst) { | |
4760 | /* rfc793: | |
4761 | * "If the RST bit is set | |
4762 | * | |
4763 | * Otherwise (no ACK) drop the segment and return." | |
4764 | */ | |
4765 | ||
4766 | goto discard_and_undo; | |
4767 | } | |
4768 | ||
4769 | /* PAWS check. */ | |
4770 | if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0)) | |
4771 | goto discard_and_undo; | |
4772 | ||
4773 | if (th->syn) { | |
4774 | /* We see SYN without ACK. It is attempt of | |
4775 | * simultaneous connect with crossed SYNs. | |
4776 | * Particularly, it can be connect to self. | |
4777 | */ | |
4778 | tcp_set_state(sk, TCP_SYN_RECV); | |
4779 | ||
4780 | if (tp->rx_opt.saw_tstamp) { | |
4781 | tp->rx_opt.tstamp_ok = 1; | |
4782 | tcp_store_ts_recent(tp); | |
4783 | tp->tcp_header_len = | |
4784 | sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; | |
4785 | } else { | |
4786 | tp->tcp_header_len = sizeof(struct tcphdr); | |
4787 | } | |
4788 | ||
4789 | tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; | |
4790 | tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; | |
4791 | ||
4792 | /* RFC1323: The window in SYN & SYN/ACK segments is | |
4793 | * never scaled. | |
4794 | */ | |
4795 | tp->snd_wnd = ntohs(th->window); | |
4796 | tp->snd_wl1 = TCP_SKB_CB(skb)->seq; | |
4797 | tp->max_window = tp->snd_wnd; | |
4798 | ||
4799 | TCP_ECN_rcv_syn(tp, th); | |
1da177e4 | 4800 | |
5d424d5a | 4801 | tcp_mtup_init(sk); |
d83d8461 | 4802 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); |
1da177e4 LT |
4803 | tcp_initialize_rcv_mss(sk); |
4804 | ||
4805 | ||
4806 | tcp_send_synack(sk); | |
4807 | #if 0 | |
4808 | /* Note, we could accept data and URG from this segment. | |
4809 | * There are no obstacles to make this. | |
4810 | * | |
4811 | * However, if we ignore data in ACKless segments sometimes, | |
4812 | * we have no reasons to accept it sometimes. | |
4813 | * Also, seems the code doing it in step6 of tcp_rcv_state_process | |
4814 | * is not flawless. So, discard packet for sanity. | |
4815 | * Uncomment this return to process the data. | |
4816 | */ | |
4817 | return -1; | |
4818 | #else | |
4819 | goto discard; | |
4820 | #endif | |
4821 | } | |
4822 | /* "fifth, if neither of the SYN or RST bits is set then | |
4823 | * drop the segment and return." | |
4824 | */ | |
4825 | ||
4826 | discard_and_undo: | |
4827 | tcp_clear_options(&tp->rx_opt); | |
4828 | tp->rx_opt.mss_clamp = saved_clamp; | |
4829 | goto discard; | |
4830 | ||
4831 | reset_and_undo: | |
4832 | tcp_clear_options(&tp->rx_opt); | |
4833 | tp->rx_opt.mss_clamp = saved_clamp; | |
4834 | return 1; | |
4835 | } | |
4836 | ||
4837 | ||
4838 | /* | |
4839 | * This function implements the receiving procedure of RFC 793 for | |
e905a9ed | 4840 | * all states except ESTABLISHED and TIME_WAIT. |
1da177e4 LT |
4841 | * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be |
4842 | * address independent. | |
4843 | */ | |
e905a9ed | 4844 | |
1da177e4 LT |
4845 | int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, |
4846 | struct tcphdr *th, unsigned len) | |
4847 | { | |
4848 | struct tcp_sock *tp = tcp_sk(sk); | |
8292a17a | 4849 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
4850 | int queued = 0; |
4851 | ||
4852 | tp->rx_opt.saw_tstamp = 0; | |
4853 | ||
4854 | switch (sk->sk_state) { | |
4855 | case TCP_CLOSE: | |
4856 | goto discard; | |
4857 | ||
4858 | case TCP_LISTEN: | |
2de979bd | 4859 | if (th->ack) |
1da177e4 LT |
4860 | return 1; |
4861 | ||
2de979bd | 4862 | if (th->rst) |
1da177e4 LT |
4863 | goto discard; |
4864 | ||
2de979bd | 4865 | if (th->syn) { |
8292a17a | 4866 | if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) |
1da177e4 LT |
4867 | return 1; |
4868 | ||
e905a9ed YH |
4869 | /* Now we have several options: In theory there is |
4870 | * nothing else in the frame. KA9Q has an option to | |
1da177e4 | 4871 | * send data with the syn, BSD accepts data with the |
e905a9ed YH |
4872 | * syn up to the [to be] advertised window and |
4873 | * Solaris 2.1 gives you a protocol error. For now | |
4874 | * we just ignore it, that fits the spec precisely | |
1da177e4 LT |
4875 | * and avoids incompatibilities. It would be nice in |
4876 | * future to drop through and process the data. | |
4877 | * | |
e905a9ed | 4878 | * Now that TTCP is starting to be used we ought to |
1da177e4 LT |
4879 | * queue this data. |
4880 | * But, this leaves one open to an easy denial of | |
e905a9ed | 4881 | * service attack, and SYN cookies can't defend |
1da177e4 | 4882 | * against this problem. So, we drop the data |
fb7e2399 MN |
4883 | * in the interest of security over speed unless |
4884 | * it's still in use. | |
1da177e4 | 4885 | */ |
fb7e2399 MN |
4886 | kfree_skb(skb); |
4887 | return 0; | |
1da177e4 LT |
4888 | } |
4889 | goto discard; | |
4890 | ||
4891 | case TCP_SYN_SENT: | |
1da177e4 LT |
4892 | queued = tcp_rcv_synsent_state_process(sk, skb, th, len); |
4893 | if (queued >= 0) | |
4894 | return queued; | |
4895 | ||
4896 | /* Do step6 onward by hand. */ | |
4897 | tcp_urg(sk, skb, th); | |
4898 | __kfree_skb(skb); | |
9e412ba7 | 4899 | tcp_data_snd_check(sk); |
1da177e4 LT |
4900 | return 0; |
4901 | } | |
4902 | ||
4903 | if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && | |
463c84b9 | 4904 | tcp_paws_discard(sk, skb)) { |
1da177e4 LT |
4905 | if (!th->rst) { |
4906 | NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); | |
4907 | tcp_send_dupack(sk, skb); | |
4908 | goto discard; | |
4909 | } | |
4910 | /* Reset is accepted even if it did not pass PAWS. */ | |
4911 | } | |
4912 | ||
4913 | /* step 1: check sequence number */ | |
4914 | if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { | |
4915 | if (!th->rst) | |
4916 | tcp_send_dupack(sk, skb); | |
4917 | goto discard; | |
4918 | } | |
4919 | ||
4920 | /* step 2: check RST bit */ | |
2de979bd | 4921 | if (th->rst) { |
1da177e4 LT |
4922 | tcp_reset(sk); |
4923 | goto discard; | |
4924 | } | |
4925 | ||
4926 | tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); | |
4927 | ||
4928 | /* step 3: check security and precedence [ignored] */ | |
4929 | ||
4930 | /* step 4: | |
4931 | * | |
4932 | * Check for a SYN in window. | |
4933 | */ | |
4934 | if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
4935 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); | |
4936 | tcp_reset(sk); | |
4937 | return 1; | |
4938 | } | |
4939 | ||
4940 | /* step 5: check the ACK field */ | |
4941 | if (th->ack) { | |
4942 | int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH); | |
4943 | ||
2de979bd | 4944 | switch (sk->sk_state) { |
1da177e4 LT |
4945 | case TCP_SYN_RECV: |
4946 | if (acceptable) { | |
4947 | tp->copied_seq = tp->rcv_nxt; | |
e16aa207 | 4948 | smp_mb(); |
1da177e4 LT |
4949 | tcp_set_state(sk, TCP_ESTABLISHED); |
4950 | sk->sk_state_change(sk); | |
4951 | ||
4952 | /* Note, that this wakeup is only for marginal | |
4953 | * crossed SYN case. Passively open sockets | |
4954 | * are not waked up, because sk->sk_sleep == | |
4955 | * NULL and sk->sk_socket == NULL. | |
4956 | */ | |
4957 | if (sk->sk_socket) { | |
4958 | sk_wake_async(sk,0,POLL_OUT); | |
4959 | } | |
4960 | ||
4961 | tp->snd_una = TCP_SKB_CB(skb)->ack_seq; | |
4962 | tp->snd_wnd = ntohs(th->window) << | |
4963 | tp->rx_opt.snd_wscale; | |
4964 | tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, | |
4965 | TCP_SKB_CB(skb)->seq); | |
4966 | ||
4967 | /* tcp_ack considers this ACK as duplicate | |
4968 | * and does not calculate rtt. | |
4969 | * Fix it at least with timestamps. | |
4970 | */ | |
4971 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && | |
4972 | !tp->srtt) | |
2d2abbab | 4973 | tcp_ack_saw_tstamp(sk, 0); |
1da177e4 LT |
4974 | |
4975 | if (tp->rx_opt.tstamp_ok) | |
4976 | tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; | |
4977 | ||
4978 | /* Make sure socket is routed, for | |
4979 | * correct metrics. | |
4980 | */ | |
8292a17a | 4981 | icsk->icsk_af_ops->rebuild_header(sk); |
1da177e4 LT |
4982 | |
4983 | tcp_init_metrics(sk); | |
4984 | ||
6687e988 | 4985 | tcp_init_congestion_control(sk); |
317a76f9 | 4986 | |
1da177e4 LT |
4987 | /* Prevent spurious tcp_cwnd_restart() on |
4988 | * first data packet. | |
4989 | */ | |
4990 | tp->lsndtime = tcp_time_stamp; | |
4991 | ||
5d424d5a | 4992 | tcp_mtup_init(sk); |
1da177e4 LT |
4993 | tcp_initialize_rcv_mss(sk); |
4994 | tcp_init_buffer_space(sk); | |
4995 | tcp_fast_path_on(tp); | |
4996 | } else { | |
4997 | return 1; | |
4998 | } | |
4999 | break; | |
5000 | ||
5001 | case TCP_FIN_WAIT1: | |
5002 | if (tp->snd_una == tp->write_seq) { | |
5003 | tcp_set_state(sk, TCP_FIN_WAIT2); | |
5004 | sk->sk_shutdown |= SEND_SHUTDOWN; | |
5005 | dst_confirm(sk->sk_dst_cache); | |
5006 | ||
5007 | if (!sock_flag(sk, SOCK_DEAD)) | |
5008 | /* Wake up lingering close() */ | |
5009 | sk->sk_state_change(sk); | |
5010 | else { | |
5011 | int tmo; | |
5012 | ||
5013 | if (tp->linger2 < 0 || | |
5014 | (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && | |
5015 | after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { | |
5016 | tcp_done(sk); | |
5017 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); | |
5018 | return 1; | |
5019 | } | |
5020 | ||
463c84b9 | 5021 | tmo = tcp_fin_time(sk); |
1da177e4 | 5022 | if (tmo > TCP_TIMEWAIT_LEN) { |
463c84b9 | 5023 | inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); |
1da177e4 LT |
5024 | } else if (th->fin || sock_owned_by_user(sk)) { |
5025 | /* Bad case. We could lose such FIN otherwise. | |
5026 | * It is not a big problem, but it looks confusing | |
5027 | * and not so rare event. We still can lose it now, | |
5028 | * if it spins in bh_lock_sock(), but it is really | |
5029 | * marginal case. | |
5030 | */ | |
463c84b9 | 5031 | inet_csk_reset_keepalive_timer(sk, tmo); |
1da177e4 LT |
5032 | } else { |
5033 | tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); | |
5034 | goto discard; | |
5035 | } | |
5036 | } | |
5037 | } | |
5038 | break; | |
5039 | ||
5040 | case TCP_CLOSING: | |
5041 | if (tp->snd_una == tp->write_seq) { | |
5042 | tcp_time_wait(sk, TCP_TIME_WAIT, 0); | |
5043 | goto discard; | |
5044 | } | |
5045 | break; | |
5046 | ||
5047 | case TCP_LAST_ACK: | |
5048 | if (tp->snd_una == tp->write_seq) { | |
5049 | tcp_update_metrics(sk); | |
5050 | tcp_done(sk); | |
5051 | goto discard; | |
5052 | } | |
5053 | break; | |
5054 | } | |
5055 | } else | |
5056 | goto discard; | |
5057 | ||
5058 | /* step 6: check the URG bit */ | |
5059 | tcp_urg(sk, skb, th); | |
5060 | ||
5061 | /* step 7: process the segment text */ | |
5062 | switch (sk->sk_state) { | |
5063 | case TCP_CLOSE_WAIT: | |
5064 | case TCP_CLOSING: | |
5065 | case TCP_LAST_ACK: | |
5066 | if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) | |
5067 | break; | |
5068 | case TCP_FIN_WAIT1: | |
5069 | case TCP_FIN_WAIT2: | |
5070 | /* RFC 793 says to queue data in these states, | |
e905a9ed | 5071 | * RFC 1122 says we MUST send a reset. |
1da177e4 LT |
5072 | * BSD 4.4 also does reset. |
5073 | */ | |
5074 | if (sk->sk_shutdown & RCV_SHUTDOWN) { | |
5075 | if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && | |
5076 | after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { | |
5077 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); | |
5078 | tcp_reset(sk); | |
5079 | return 1; | |
5080 | } | |
5081 | } | |
5082 | /* Fall through */ | |
e905a9ed | 5083 | case TCP_ESTABLISHED: |
1da177e4 LT |
5084 | tcp_data_queue(sk, skb); |
5085 | queued = 1; | |
5086 | break; | |
5087 | } | |
5088 | ||
5089 | /* tcp_data could move socket to TIME-WAIT */ | |
5090 | if (sk->sk_state != TCP_CLOSE) { | |
9e412ba7 | 5091 | tcp_data_snd_check(sk); |
1da177e4 LT |
5092 | tcp_ack_snd_check(sk); |
5093 | } | |
5094 | ||
e905a9ed | 5095 | if (!queued) { |
1da177e4 LT |
5096 | discard: |
5097 | __kfree_skb(skb); | |
5098 | } | |
5099 | return 0; | |
5100 | } | |
5101 | ||
5102 | EXPORT_SYMBOL(sysctl_tcp_ecn); | |
5103 | EXPORT_SYMBOL(sysctl_tcp_reordering); | |
5104 | EXPORT_SYMBOL(tcp_parse_options); | |
5105 | EXPORT_SYMBOL(tcp_rcv_established); | |
5106 | EXPORT_SYMBOL(tcp_rcv_state_process); | |
40efc6fa | 5107 | EXPORT_SYMBOL(tcp_initialize_rcv_mss); |