]>
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 */ |
1da177e4 LT |
107 | |
108 | #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) | |
109 | #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) | |
110 | #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) | |
111 | #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) | |
2e605294 | 112 | #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED) |
1da177e4 | 113 | |
4dc2665e IJ |
114 | #define IsSackFrto() (sysctl_tcp_frto == 0x2) |
115 | ||
1da177e4 | 116 | #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) |
bdf1ee5d | 117 | #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) |
1da177e4 | 118 | |
e905a9ed | 119 | /* Adapt the MSS value used to make delayed ack decision to the |
1da177e4 | 120 | * real world. |
e905a9ed | 121 | */ |
40efc6fa SH |
122 | static void tcp_measure_rcv_mss(struct sock *sk, |
123 | const struct sk_buff *skb) | |
1da177e4 | 124 | { |
463c84b9 | 125 | struct inet_connection_sock *icsk = inet_csk(sk); |
e905a9ed | 126 | const unsigned int lss = icsk->icsk_ack.last_seg_size; |
463c84b9 | 127 | unsigned int len; |
1da177e4 | 128 | |
e905a9ed | 129 | icsk->icsk_ack.last_seg_size = 0; |
1da177e4 LT |
130 | |
131 | /* skb->len may jitter because of SACKs, even if peer | |
132 | * sends good full-sized frames. | |
133 | */ | |
ff9b5e0f | 134 | len = skb_shinfo(skb)->gso_size ?: skb->len; |
463c84b9 ACM |
135 | if (len >= icsk->icsk_ack.rcv_mss) { |
136 | icsk->icsk_ack.rcv_mss = len; | |
1da177e4 LT |
137 | } else { |
138 | /* Otherwise, we make more careful check taking into account, | |
139 | * that SACKs block is variable. | |
140 | * | |
141 | * "len" is invariant segment length, including TCP header. | |
142 | */ | |
9c70220b | 143 | len += skb->data - skb_transport_header(skb); |
1da177e4 LT |
144 | if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) || |
145 | /* If PSH is not set, packet should be | |
146 | * full sized, provided peer TCP is not badly broken. | |
147 | * This observation (if it is correct 8)) allows | |
148 | * to handle super-low mtu links fairly. | |
149 | */ | |
150 | (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && | |
aa8223c7 | 151 | !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { |
1da177e4 LT |
152 | /* Subtract also invariant (if peer is RFC compliant), |
153 | * tcp header plus fixed timestamp option length. | |
154 | * Resulting "len" is MSS free of SACK jitter. | |
155 | */ | |
463c84b9 ACM |
156 | len -= tcp_sk(sk)->tcp_header_len; |
157 | icsk->icsk_ack.last_seg_size = len; | |
1da177e4 | 158 | if (len == lss) { |
463c84b9 | 159 | icsk->icsk_ack.rcv_mss = len; |
1da177e4 LT |
160 | return; |
161 | } | |
162 | } | |
1ef9696c AK |
163 | if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) |
164 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; | |
463c84b9 | 165 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; |
1da177e4 LT |
166 | } |
167 | } | |
168 | ||
463c84b9 | 169 | static void tcp_incr_quickack(struct sock *sk) |
1da177e4 | 170 | { |
463c84b9 ACM |
171 | struct inet_connection_sock *icsk = inet_csk(sk); |
172 | unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); | |
1da177e4 LT |
173 | |
174 | if (quickacks==0) | |
175 | quickacks=2; | |
463c84b9 ACM |
176 | if (quickacks > icsk->icsk_ack.quick) |
177 | icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); | |
1da177e4 LT |
178 | } |
179 | ||
463c84b9 | 180 | void tcp_enter_quickack_mode(struct sock *sk) |
1da177e4 | 181 | { |
463c84b9 ACM |
182 | struct inet_connection_sock *icsk = inet_csk(sk); |
183 | tcp_incr_quickack(sk); | |
184 | icsk->icsk_ack.pingpong = 0; | |
185 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 LT |
186 | } |
187 | ||
188 | /* Send ACKs quickly, if "quick" count is not exhausted | |
189 | * and the session is not interactive. | |
190 | */ | |
191 | ||
463c84b9 | 192 | static inline int tcp_in_quickack_mode(const struct sock *sk) |
1da177e4 | 193 | { |
463c84b9 ACM |
194 | const struct inet_connection_sock *icsk = inet_csk(sk); |
195 | return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; | |
1da177e4 LT |
196 | } |
197 | ||
bdf1ee5d IJ |
198 | static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp) |
199 | { | |
200 | if (tp->ecn_flags&TCP_ECN_OK) | |
201 | tp->ecn_flags |= TCP_ECN_QUEUE_CWR; | |
202 | } | |
203 | ||
204 | static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb) | |
205 | { | |
206 | if (tcp_hdr(skb)->cwr) | |
207 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
208 | } | |
209 | ||
210 | static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp) | |
211 | { | |
212 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
213 | } | |
214 | ||
215 | static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb) | |
216 | { | |
217 | if (tp->ecn_flags&TCP_ECN_OK) { | |
218 | if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags)) | |
219 | tp->ecn_flags |= TCP_ECN_DEMAND_CWR; | |
220 | /* Funny extension: if ECT is not set on a segment, | |
221 | * it is surely retransmit. It is not in ECN RFC, | |
222 | * but Linux follows this rule. */ | |
223 | else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags))) | |
224 | tcp_enter_quickack_mode((struct sock *)tp); | |
225 | } | |
226 | } | |
227 | ||
228 | static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th) | |
229 | { | |
230 | if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || th->cwr)) | |
231 | tp->ecn_flags &= ~TCP_ECN_OK; | |
232 | } | |
233 | ||
234 | static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th) | |
235 | { | |
236 | if ((tp->ecn_flags&TCP_ECN_OK) && (!th->ece || !th->cwr)) | |
237 | tp->ecn_flags &= ~TCP_ECN_OK; | |
238 | } | |
239 | ||
240 | static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th) | |
241 | { | |
242 | if (th->ece && !th->syn && (tp->ecn_flags&TCP_ECN_OK)) | |
243 | return 1; | |
244 | return 0; | |
245 | } | |
246 | ||
1da177e4 LT |
247 | /* Buffer size and advertised window tuning. |
248 | * | |
249 | * 1. Tuning sk->sk_sndbuf, when connection enters established state. | |
250 | */ | |
251 | ||
252 | static void tcp_fixup_sndbuf(struct sock *sk) | |
253 | { | |
254 | int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 + | |
255 | sizeof(struct sk_buff); | |
256 | ||
257 | if (sk->sk_sndbuf < 3 * sndmem) | |
258 | sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]); | |
259 | } | |
260 | ||
261 | /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) | |
262 | * | |
263 | * All tcp_full_space() is split to two parts: "network" buffer, allocated | |
264 | * forward and advertised in receiver window (tp->rcv_wnd) and | |
265 | * "application buffer", required to isolate scheduling/application | |
266 | * latencies from network. | |
267 | * window_clamp is maximal advertised window. It can be less than | |
268 | * tcp_full_space(), in this case tcp_full_space() - window_clamp | |
269 | * is reserved for "application" buffer. The less window_clamp is | |
270 | * the smoother our behaviour from viewpoint of network, but the lower | |
271 | * throughput and the higher sensitivity of the connection to losses. 8) | |
272 | * | |
273 | * rcv_ssthresh is more strict window_clamp used at "slow start" | |
274 | * phase to predict further behaviour of this connection. | |
275 | * It is used for two goals: | |
276 | * - to enforce header prediction at sender, even when application | |
277 | * requires some significant "application buffer". It is check #1. | |
278 | * - to prevent pruning of receive queue because of misprediction | |
279 | * of receiver window. Check #2. | |
280 | * | |
281 | * The scheme does not work when sender sends good segments opening | |
caa20d9a | 282 | * window and then starts to feed us spaghetti. But it should work |
1da177e4 LT |
283 | * in common situations. Otherwise, we have to rely on queue collapsing. |
284 | */ | |
285 | ||
286 | /* Slow part of check#2. */ | |
9e412ba7 | 287 | static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 288 | { |
9e412ba7 | 289 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
290 | /* Optimize this! */ |
291 | int truesize = tcp_win_from_space(skb->truesize)/2; | |
326f36e9 | 292 | int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2; |
1da177e4 LT |
293 | |
294 | while (tp->rcv_ssthresh <= window) { | |
295 | if (truesize <= skb->len) | |
463c84b9 | 296 | return 2 * inet_csk(sk)->icsk_ack.rcv_mss; |
1da177e4 LT |
297 | |
298 | truesize >>= 1; | |
299 | window >>= 1; | |
300 | } | |
301 | return 0; | |
302 | } | |
303 | ||
9e412ba7 | 304 | static void tcp_grow_window(struct sock *sk, |
40efc6fa | 305 | struct sk_buff *skb) |
1da177e4 | 306 | { |
9e412ba7 IJ |
307 | struct tcp_sock *tp = tcp_sk(sk); |
308 | ||
1da177e4 LT |
309 | /* Check #1 */ |
310 | if (tp->rcv_ssthresh < tp->window_clamp && | |
311 | (int)tp->rcv_ssthresh < tcp_space(sk) && | |
312 | !tcp_memory_pressure) { | |
313 | int incr; | |
314 | ||
315 | /* Check #2. Increase window, if skb with such overhead | |
316 | * will fit to rcvbuf in future. | |
317 | */ | |
318 | if (tcp_win_from_space(skb->truesize) <= skb->len) | |
319 | incr = 2*tp->advmss; | |
320 | else | |
9e412ba7 | 321 | incr = __tcp_grow_window(sk, skb); |
1da177e4 LT |
322 | |
323 | if (incr) { | |
324 | tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp); | |
463c84b9 | 325 | inet_csk(sk)->icsk_ack.quick |= 1; |
1da177e4 LT |
326 | } |
327 | } | |
328 | } | |
329 | ||
330 | /* 3. Tuning rcvbuf, when connection enters established state. */ | |
331 | ||
332 | static void tcp_fixup_rcvbuf(struct sock *sk) | |
333 | { | |
334 | struct tcp_sock *tp = tcp_sk(sk); | |
335 | int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); | |
336 | ||
337 | /* Try to select rcvbuf so that 4 mss-sized segments | |
caa20d9a | 338 | * will fit to window and corresponding skbs will fit to our rcvbuf. |
1da177e4 LT |
339 | * (was 3; 4 is minimum to allow fast retransmit to work.) |
340 | */ | |
341 | while (tcp_win_from_space(rcvmem) < tp->advmss) | |
342 | rcvmem += 128; | |
343 | if (sk->sk_rcvbuf < 4 * rcvmem) | |
344 | sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]); | |
345 | } | |
346 | ||
caa20d9a | 347 | /* 4. Try to fixup all. It is made immediately after connection enters |
1da177e4 LT |
348 | * established state. |
349 | */ | |
350 | static void tcp_init_buffer_space(struct sock *sk) | |
351 | { | |
352 | struct tcp_sock *tp = tcp_sk(sk); | |
353 | int maxwin; | |
354 | ||
355 | if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) | |
356 | tcp_fixup_rcvbuf(sk); | |
357 | if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) | |
358 | tcp_fixup_sndbuf(sk); | |
359 | ||
360 | tp->rcvq_space.space = tp->rcv_wnd; | |
361 | ||
362 | maxwin = tcp_full_space(sk); | |
363 | ||
364 | if (tp->window_clamp >= maxwin) { | |
365 | tp->window_clamp = maxwin; | |
366 | ||
367 | if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) | |
368 | tp->window_clamp = max(maxwin - | |
369 | (maxwin >> sysctl_tcp_app_win), | |
370 | 4 * tp->advmss); | |
371 | } | |
372 | ||
373 | /* Force reservation of one segment. */ | |
374 | if (sysctl_tcp_app_win && | |
375 | tp->window_clamp > 2 * tp->advmss && | |
376 | tp->window_clamp + tp->advmss > maxwin) | |
377 | tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); | |
378 | ||
379 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); | |
380 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
381 | } | |
382 | ||
1da177e4 | 383 | /* 5. Recalculate window clamp after socket hit its memory bounds. */ |
9e412ba7 | 384 | static void tcp_clamp_window(struct sock *sk) |
1da177e4 | 385 | { |
9e412ba7 | 386 | struct tcp_sock *tp = tcp_sk(sk); |
6687e988 | 387 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 388 | |
6687e988 | 389 | icsk->icsk_ack.quick = 0; |
1da177e4 | 390 | |
326f36e9 JH |
391 | if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && |
392 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && | |
393 | !tcp_memory_pressure && | |
394 | atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) { | |
395 | sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), | |
396 | sysctl_tcp_rmem[2]); | |
1da177e4 | 397 | } |
326f36e9 | 398 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) |
1da177e4 | 399 | tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss); |
1da177e4 LT |
400 | } |
401 | ||
40efc6fa SH |
402 | |
403 | /* Initialize RCV_MSS value. | |
404 | * RCV_MSS is an our guess about MSS used by the peer. | |
405 | * We haven't any direct information about the MSS. | |
406 | * It's better to underestimate the RCV_MSS rather than overestimate. | |
407 | * Overestimations make us ACKing less frequently than needed. | |
408 | * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). | |
409 | */ | |
410 | void tcp_initialize_rcv_mss(struct sock *sk) | |
411 | { | |
412 | struct tcp_sock *tp = tcp_sk(sk); | |
413 | unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); | |
414 | ||
415 | hint = min(hint, tp->rcv_wnd/2); | |
416 | hint = min(hint, TCP_MIN_RCVMSS); | |
417 | hint = max(hint, TCP_MIN_MSS); | |
418 | ||
419 | inet_csk(sk)->icsk_ack.rcv_mss = hint; | |
420 | } | |
421 | ||
1da177e4 LT |
422 | /* Receiver "autotuning" code. |
423 | * | |
424 | * The algorithm for RTT estimation w/o timestamps is based on | |
425 | * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. | |
426 | * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps> | |
427 | * | |
428 | * More detail on this code can be found at | |
429 | * <http://www.psc.edu/~jheffner/senior_thesis.ps>, | |
430 | * though this reference is out of date. A new paper | |
431 | * is pending. | |
432 | */ | |
433 | static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) | |
434 | { | |
435 | u32 new_sample = tp->rcv_rtt_est.rtt; | |
436 | long m = sample; | |
437 | ||
438 | if (m == 0) | |
439 | m = 1; | |
440 | ||
441 | if (new_sample != 0) { | |
442 | /* If we sample in larger samples in the non-timestamp | |
443 | * case, we could grossly overestimate the RTT especially | |
444 | * with chatty applications or bulk transfer apps which | |
445 | * are stalled on filesystem I/O. | |
446 | * | |
447 | * Also, since we are only going for a minimum in the | |
31f34269 | 448 | * non-timestamp case, we do not smooth things out |
caa20d9a | 449 | * else with timestamps disabled convergence takes too |
1da177e4 LT |
450 | * long. |
451 | */ | |
452 | if (!win_dep) { | |
453 | m -= (new_sample >> 3); | |
454 | new_sample += m; | |
455 | } else if (m < new_sample) | |
456 | new_sample = m << 3; | |
457 | } else { | |
caa20d9a | 458 | /* No previous measure. */ |
1da177e4 LT |
459 | new_sample = m << 3; |
460 | } | |
461 | ||
462 | if (tp->rcv_rtt_est.rtt != new_sample) | |
463 | tp->rcv_rtt_est.rtt = new_sample; | |
464 | } | |
465 | ||
466 | static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) | |
467 | { | |
468 | if (tp->rcv_rtt_est.time == 0) | |
469 | goto new_measure; | |
470 | if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) | |
471 | return; | |
472 | tcp_rcv_rtt_update(tp, | |
473 | jiffies - tp->rcv_rtt_est.time, | |
474 | 1); | |
475 | ||
476 | new_measure: | |
477 | tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; | |
478 | tp->rcv_rtt_est.time = tcp_time_stamp; | |
479 | } | |
480 | ||
463c84b9 | 481 | static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 482 | { |
463c84b9 | 483 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
484 | if (tp->rx_opt.rcv_tsecr && |
485 | (TCP_SKB_CB(skb)->end_seq - | |
463c84b9 | 486 | TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) |
1da177e4 LT |
487 | tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); |
488 | } | |
489 | ||
490 | /* | |
491 | * This function should be called every time data is copied to user space. | |
492 | * It calculates the appropriate TCP receive buffer space. | |
493 | */ | |
494 | void tcp_rcv_space_adjust(struct sock *sk) | |
495 | { | |
496 | struct tcp_sock *tp = tcp_sk(sk); | |
497 | int time; | |
498 | int space; | |
e905a9ed | 499 | |
1da177e4 LT |
500 | if (tp->rcvq_space.time == 0) |
501 | goto new_measure; | |
e905a9ed | 502 | |
1da177e4 LT |
503 | time = tcp_time_stamp - tp->rcvq_space.time; |
504 | if (time < (tp->rcv_rtt_est.rtt >> 3) || | |
505 | tp->rcv_rtt_est.rtt == 0) | |
506 | return; | |
e905a9ed | 507 | |
1da177e4 LT |
508 | space = 2 * (tp->copied_seq - tp->rcvq_space.seq); |
509 | ||
510 | space = max(tp->rcvq_space.space, space); | |
511 | ||
512 | if (tp->rcvq_space.space != space) { | |
513 | int rcvmem; | |
514 | ||
515 | tp->rcvq_space.space = space; | |
516 | ||
6fcf9412 JH |
517 | if (sysctl_tcp_moderate_rcvbuf && |
518 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { | |
1da177e4 LT |
519 | int new_clamp = space; |
520 | ||
521 | /* Receive space grows, normalize in order to | |
522 | * take into account packet headers and sk_buff | |
523 | * structure overhead. | |
524 | */ | |
525 | space /= tp->advmss; | |
526 | if (!space) | |
527 | space = 1; | |
528 | rcvmem = (tp->advmss + MAX_TCP_HEADER + | |
529 | 16 + sizeof(struct sk_buff)); | |
530 | while (tcp_win_from_space(rcvmem) < tp->advmss) | |
531 | rcvmem += 128; | |
532 | space *= rcvmem; | |
533 | space = min(space, sysctl_tcp_rmem[2]); | |
534 | if (space > sk->sk_rcvbuf) { | |
535 | sk->sk_rcvbuf = space; | |
536 | ||
537 | /* Make the window clamp follow along. */ | |
538 | tp->window_clamp = new_clamp; | |
539 | } | |
540 | } | |
541 | } | |
e905a9ed | 542 | |
1da177e4 LT |
543 | new_measure: |
544 | tp->rcvq_space.seq = tp->copied_seq; | |
545 | tp->rcvq_space.time = tcp_time_stamp; | |
546 | } | |
547 | ||
548 | /* There is something which you must keep in mind when you analyze the | |
549 | * behavior of the tp->ato delayed ack timeout interval. When a | |
550 | * connection starts up, we want to ack as quickly as possible. The | |
551 | * problem is that "good" TCP's do slow start at the beginning of data | |
552 | * transmission. The means that until we send the first few ACK's the | |
553 | * sender will sit on his end and only queue most of his data, because | |
554 | * he can only send snd_cwnd unacked packets at any given time. For | |
555 | * each ACK we send, he increments snd_cwnd and transmits more of his | |
556 | * queue. -DaveM | |
557 | */ | |
9e412ba7 | 558 | static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 559 | { |
9e412ba7 | 560 | struct tcp_sock *tp = tcp_sk(sk); |
463c84b9 | 561 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
562 | u32 now; |
563 | ||
463c84b9 | 564 | inet_csk_schedule_ack(sk); |
1da177e4 | 565 | |
463c84b9 | 566 | tcp_measure_rcv_mss(sk, skb); |
1da177e4 LT |
567 | |
568 | tcp_rcv_rtt_measure(tp); | |
e905a9ed | 569 | |
1da177e4 LT |
570 | now = tcp_time_stamp; |
571 | ||
463c84b9 | 572 | if (!icsk->icsk_ack.ato) { |
1da177e4 LT |
573 | /* The _first_ data packet received, initialize |
574 | * delayed ACK engine. | |
575 | */ | |
463c84b9 ACM |
576 | tcp_incr_quickack(sk); |
577 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 | 578 | } else { |
463c84b9 | 579 | int m = now - icsk->icsk_ack.lrcvtime; |
1da177e4 LT |
580 | |
581 | if (m <= TCP_ATO_MIN/2) { | |
582 | /* The fastest case is the first. */ | |
463c84b9 ACM |
583 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; |
584 | } else if (m < icsk->icsk_ack.ato) { | |
585 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; | |
586 | if (icsk->icsk_ack.ato > icsk->icsk_rto) | |
587 | icsk->icsk_ack.ato = icsk->icsk_rto; | |
588 | } else if (m > icsk->icsk_rto) { | |
caa20d9a | 589 | /* Too long gap. Apparently sender failed to |
1da177e4 LT |
590 | * restart window, so that we send ACKs quickly. |
591 | */ | |
463c84b9 | 592 | tcp_incr_quickack(sk); |
1da177e4 LT |
593 | sk_stream_mem_reclaim(sk); |
594 | } | |
595 | } | |
463c84b9 | 596 | icsk->icsk_ack.lrcvtime = now; |
1da177e4 LT |
597 | |
598 | TCP_ECN_check_ce(tp, skb); | |
599 | ||
600 | if (skb->len >= 128) | |
9e412ba7 | 601 | tcp_grow_window(sk, skb); |
1da177e4 LT |
602 | } |
603 | ||
05bb1fad DM |
604 | static u32 tcp_rto_min(struct sock *sk) |
605 | { | |
606 | struct dst_entry *dst = __sk_dst_get(sk); | |
607 | u32 rto_min = TCP_RTO_MIN; | |
608 | ||
5c127c58 | 609 | if (dst && dst_metric_locked(dst, RTAX_RTO_MIN)) |
05bb1fad DM |
610 | rto_min = dst->metrics[RTAX_RTO_MIN-1]; |
611 | return rto_min; | |
612 | } | |
613 | ||
1da177e4 LT |
614 | /* Called to compute a smoothed rtt estimate. The data fed to this |
615 | * routine either comes from timestamps, or from segments that were | |
616 | * known _not_ to have been retransmitted [see Karn/Partridge | |
617 | * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 | |
618 | * piece by Van Jacobson. | |
619 | * NOTE: the next three routines used to be one big routine. | |
620 | * To save cycles in the RFC 1323 implementation it was better to break | |
621 | * it up into three procedures. -- erics | |
622 | */ | |
2d2abbab | 623 | static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) |
1da177e4 | 624 | { |
6687e988 | 625 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
626 | long m = mrtt; /* RTT */ |
627 | ||
1da177e4 LT |
628 | /* The following amusing code comes from Jacobson's |
629 | * article in SIGCOMM '88. Note that rtt and mdev | |
630 | * are scaled versions of rtt and mean deviation. | |
e905a9ed | 631 | * This is designed to be as fast as possible |
1da177e4 LT |
632 | * m stands for "measurement". |
633 | * | |
634 | * On a 1990 paper the rto value is changed to: | |
635 | * RTO = rtt + 4 * mdev | |
636 | * | |
637 | * Funny. This algorithm seems to be very broken. | |
638 | * These formulae increase RTO, when it should be decreased, increase | |
31f34269 | 639 | * too slowly, when it should be increased quickly, decrease too quickly |
1da177e4 LT |
640 | * etc. I guess in BSD RTO takes ONE value, so that it is absolutely |
641 | * does not matter how to _calculate_ it. Seems, it was trap | |
642 | * that VJ failed to avoid. 8) | |
643 | */ | |
2de979bd | 644 | if (m == 0) |
1da177e4 LT |
645 | m = 1; |
646 | if (tp->srtt != 0) { | |
647 | m -= (tp->srtt >> 3); /* m is now error in rtt est */ | |
648 | tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ | |
649 | if (m < 0) { | |
650 | m = -m; /* m is now abs(error) */ | |
651 | m -= (tp->mdev >> 2); /* similar update on mdev */ | |
652 | /* This is similar to one of Eifel findings. | |
653 | * Eifel blocks mdev updates when rtt decreases. | |
654 | * This solution is a bit different: we use finer gain | |
655 | * for mdev in this case (alpha*beta). | |
656 | * Like Eifel it also prevents growth of rto, | |
657 | * but also it limits too fast rto decreases, | |
658 | * happening in pure Eifel. | |
659 | */ | |
660 | if (m > 0) | |
661 | m >>= 3; | |
662 | } else { | |
663 | m -= (tp->mdev >> 2); /* similar update on mdev */ | |
664 | } | |
665 | tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ | |
666 | if (tp->mdev > tp->mdev_max) { | |
667 | tp->mdev_max = tp->mdev; | |
668 | if (tp->mdev_max > tp->rttvar) | |
669 | tp->rttvar = tp->mdev_max; | |
670 | } | |
671 | if (after(tp->snd_una, tp->rtt_seq)) { | |
672 | if (tp->mdev_max < tp->rttvar) | |
673 | tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2; | |
674 | tp->rtt_seq = tp->snd_nxt; | |
05bb1fad | 675 | tp->mdev_max = tcp_rto_min(sk); |
1da177e4 LT |
676 | } |
677 | } else { | |
678 | /* no previous measure. */ | |
679 | tp->srtt = m<<3; /* take the measured time to be rtt */ | |
680 | tp->mdev = m<<1; /* make sure rto = 3*rtt */ | |
05bb1fad | 681 | tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); |
1da177e4 LT |
682 | tp->rtt_seq = tp->snd_nxt; |
683 | } | |
1da177e4 LT |
684 | } |
685 | ||
686 | /* Calculate rto without backoff. This is the second half of Van Jacobson's | |
687 | * routine referred to above. | |
688 | */ | |
463c84b9 | 689 | static inline void tcp_set_rto(struct sock *sk) |
1da177e4 | 690 | { |
463c84b9 | 691 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
692 | /* Old crap is replaced with new one. 8) |
693 | * | |
694 | * More seriously: | |
695 | * 1. If rtt variance happened to be less 50msec, it is hallucination. | |
696 | * It cannot be less due to utterly erratic ACK generation made | |
697 | * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ | |
698 | * to do with delayed acks, because at cwnd>2 true delack timeout | |
699 | * is invisible. Actually, Linux-2.4 also generates erratic | |
caa20d9a | 700 | * ACKs in some circumstances. |
1da177e4 | 701 | */ |
463c84b9 | 702 | inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar; |
1da177e4 LT |
703 | |
704 | /* 2. Fixups made earlier cannot be right. | |
705 | * If we do not estimate RTO correctly without them, | |
706 | * all the algo is pure shit and should be replaced | |
caa20d9a | 707 | * with correct one. It is exactly, which we pretend to do. |
1da177e4 LT |
708 | */ |
709 | } | |
710 | ||
711 | /* NOTE: clamping at TCP_RTO_MIN is not required, current algo | |
712 | * guarantees that rto is higher. | |
713 | */ | |
463c84b9 | 714 | static inline void tcp_bound_rto(struct sock *sk) |
1da177e4 | 715 | { |
463c84b9 ACM |
716 | if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX) |
717 | inet_csk(sk)->icsk_rto = TCP_RTO_MAX; | |
1da177e4 LT |
718 | } |
719 | ||
720 | /* Save metrics learned by this TCP session. | |
721 | This function is called only, when TCP finishes successfully | |
722 | i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. | |
723 | */ | |
724 | void tcp_update_metrics(struct sock *sk) | |
725 | { | |
726 | struct tcp_sock *tp = tcp_sk(sk); | |
727 | struct dst_entry *dst = __sk_dst_get(sk); | |
728 | ||
729 | if (sysctl_tcp_nometrics_save) | |
730 | return; | |
731 | ||
732 | dst_confirm(dst); | |
733 | ||
734 | if (dst && (dst->flags&DST_HOST)) { | |
6687e988 | 735 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
736 | int m; |
737 | ||
6687e988 | 738 | if (icsk->icsk_backoff || !tp->srtt) { |
1da177e4 LT |
739 | /* This session failed to estimate rtt. Why? |
740 | * Probably, no packets returned in time. | |
741 | * Reset our results. | |
742 | */ | |
743 | if (!(dst_metric_locked(dst, RTAX_RTT))) | |
744 | dst->metrics[RTAX_RTT-1] = 0; | |
745 | return; | |
746 | } | |
747 | ||
748 | m = dst_metric(dst, RTAX_RTT) - tp->srtt; | |
749 | ||
750 | /* If newly calculated rtt larger than stored one, | |
751 | * store new one. Otherwise, use EWMA. Remember, | |
752 | * rtt overestimation is always better than underestimation. | |
753 | */ | |
754 | if (!(dst_metric_locked(dst, RTAX_RTT))) { | |
755 | if (m <= 0) | |
756 | dst->metrics[RTAX_RTT-1] = tp->srtt; | |
757 | else | |
758 | dst->metrics[RTAX_RTT-1] -= (m>>3); | |
759 | } | |
760 | ||
761 | if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { | |
762 | if (m < 0) | |
763 | m = -m; | |
764 | ||
765 | /* Scale deviation to rttvar fixed point */ | |
766 | m >>= 1; | |
767 | if (m < tp->mdev) | |
768 | m = tp->mdev; | |
769 | ||
770 | if (m >= dst_metric(dst, RTAX_RTTVAR)) | |
771 | dst->metrics[RTAX_RTTVAR-1] = m; | |
772 | else | |
773 | dst->metrics[RTAX_RTTVAR-1] -= | |
774 | (dst->metrics[RTAX_RTTVAR-1] - m)>>2; | |
775 | } | |
776 | ||
777 | if (tp->snd_ssthresh >= 0xFFFF) { | |
778 | /* Slow start still did not finish. */ | |
779 | if (dst_metric(dst, RTAX_SSTHRESH) && | |
780 | !dst_metric_locked(dst, RTAX_SSTHRESH) && | |
781 | (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) | |
782 | dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1; | |
783 | if (!dst_metric_locked(dst, RTAX_CWND) && | |
784 | tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) | |
785 | dst->metrics[RTAX_CWND-1] = tp->snd_cwnd; | |
786 | } else if (tp->snd_cwnd > tp->snd_ssthresh && | |
6687e988 | 787 | icsk->icsk_ca_state == TCP_CA_Open) { |
1da177e4 LT |
788 | /* Cong. avoidance phase, cwnd is reliable. */ |
789 | if (!dst_metric_locked(dst, RTAX_SSTHRESH)) | |
790 | dst->metrics[RTAX_SSTHRESH-1] = | |
791 | max(tp->snd_cwnd >> 1, tp->snd_ssthresh); | |
792 | if (!dst_metric_locked(dst, RTAX_CWND)) | |
793 | dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1; | |
794 | } else { | |
795 | /* Else slow start did not finish, cwnd is non-sense, | |
796 | ssthresh may be also invalid. | |
797 | */ | |
798 | if (!dst_metric_locked(dst, RTAX_CWND)) | |
799 | dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1; | |
800 | if (dst->metrics[RTAX_SSTHRESH-1] && | |
801 | !dst_metric_locked(dst, RTAX_SSTHRESH) && | |
802 | tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1]) | |
803 | dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh; | |
804 | } | |
805 | ||
806 | if (!dst_metric_locked(dst, RTAX_REORDERING)) { | |
807 | if (dst->metrics[RTAX_REORDERING-1] < tp->reordering && | |
808 | tp->reordering != sysctl_tcp_reordering) | |
809 | dst->metrics[RTAX_REORDERING-1] = tp->reordering; | |
810 | } | |
811 | } | |
812 | } | |
813 | ||
26722873 DM |
814 | /* Numbers are taken from RFC3390. |
815 | * | |
816 | * John Heffner states: | |
817 | * | |
818 | * The RFC specifies a window of no more than 4380 bytes | |
819 | * unless 2*MSS > 4380. Reading the pseudocode in the RFC | |
820 | * is a bit misleading because they use a clamp at 4380 bytes | |
821 | * rather than use a multiplier in the relevant range. | |
822 | */ | |
1da177e4 LT |
823 | __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst) |
824 | { | |
825 | __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); | |
826 | ||
827 | if (!cwnd) { | |
c1b4a7e6 | 828 | if (tp->mss_cache > 1460) |
1da177e4 LT |
829 | cwnd = 2; |
830 | else | |
c1b4a7e6 | 831 | cwnd = (tp->mss_cache > 1095) ? 3 : 4; |
1da177e4 LT |
832 | } |
833 | return min_t(__u32, cwnd, tp->snd_cwnd_clamp); | |
834 | } | |
835 | ||
40efc6fa | 836 | /* Set slow start threshold and cwnd not falling to slow start */ |
3cfe3baa | 837 | void tcp_enter_cwr(struct sock *sk, const int set_ssthresh) |
40efc6fa SH |
838 | { |
839 | struct tcp_sock *tp = tcp_sk(sk); | |
3cfe3baa | 840 | const struct inet_connection_sock *icsk = inet_csk(sk); |
40efc6fa SH |
841 | |
842 | tp->prior_ssthresh = 0; | |
843 | tp->bytes_acked = 0; | |
e01f9d77 | 844 | if (icsk->icsk_ca_state < TCP_CA_CWR) { |
40efc6fa | 845 | tp->undo_marker = 0; |
3cfe3baa IJ |
846 | if (set_ssthresh) |
847 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
40efc6fa SH |
848 | tp->snd_cwnd = min(tp->snd_cwnd, |
849 | tcp_packets_in_flight(tp) + 1U); | |
850 | tp->snd_cwnd_cnt = 0; | |
851 | tp->high_seq = tp->snd_nxt; | |
852 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
853 | TCP_ECN_queue_cwr(tp); | |
854 | ||
855 | tcp_set_ca_state(sk, TCP_CA_CWR); | |
856 | } | |
857 | } | |
858 | ||
e60402d0 IJ |
859 | /* |
860 | * Packet counting of FACK is based on in-order assumptions, therefore TCP | |
861 | * disables it when reordering is detected | |
862 | */ | |
863 | static void tcp_disable_fack(struct tcp_sock *tp) | |
864 | { | |
865 | tp->rx_opt.sack_ok &= ~2; | |
866 | } | |
867 | ||
868 | /* Take a notice that peer is sending DSACKs */ | |
869 | static void tcp_dsack_seen(struct tcp_sock *tp) | |
870 | { | |
871 | tp->rx_opt.sack_ok |= 4; | |
872 | } | |
873 | ||
1da177e4 LT |
874 | /* Initialize metrics on socket. */ |
875 | ||
876 | static void tcp_init_metrics(struct sock *sk) | |
877 | { | |
878 | struct tcp_sock *tp = tcp_sk(sk); | |
879 | struct dst_entry *dst = __sk_dst_get(sk); | |
880 | ||
881 | if (dst == NULL) | |
882 | goto reset; | |
883 | ||
884 | dst_confirm(dst); | |
885 | ||
886 | if (dst_metric_locked(dst, RTAX_CWND)) | |
887 | tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); | |
888 | if (dst_metric(dst, RTAX_SSTHRESH)) { | |
889 | tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); | |
890 | if (tp->snd_ssthresh > tp->snd_cwnd_clamp) | |
891 | tp->snd_ssthresh = tp->snd_cwnd_clamp; | |
892 | } | |
893 | if (dst_metric(dst, RTAX_REORDERING) && | |
894 | tp->reordering != dst_metric(dst, RTAX_REORDERING)) { | |
e60402d0 | 895 | tcp_disable_fack(tp); |
1da177e4 LT |
896 | tp->reordering = dst_metric(dst, RTAX_REORDERING); |
897 | } | |
898 | ||
899 | if (dst_metric(dst, RTAX_RTT) == 0) | |
900 | goto reset; | |
901 | ||
902 | if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3)) | |
903 | goto reset; | |
904 | ||
905 | /* Initial rtt is determined from SYN,SYN-ACK. | |
906 | * The segment is small and rtt may appear much | |
907 | * less than real one. Use per-dst memory | |
908 | * to make it more realistic. | |
909 | * | |
910 | * A bit of theory. RTT is time passed after "normal" sized packet | |
caa20d9a | 911 | * is sent until it is ACKed. In normal circumstances sending small |
1da177e4 LT |
912 | * packets force peer to delay ACKs and calculation is correct too. |
913 | * The algorithm is adaptive and, provided we follow specs, it | |
914 | * NEVER underestimate RTT. BUT! If peer tries to make some clever | |
915 | * tricks sort of "quick acks" for time long enough to decrease RTT | |
916 | * to low value, and then abruptly stops to do it and starts to delay | |
917 | * ACKs, wait for troubles. | |
918 | */ | |
919 | if (dst_metric(dst, RTAX_RTT) > tp->srtt) { | |
920 | tp->srtt = dst_metric(dst, RTAX_RTT); | |
921 | tp->rtt_seq = tp->snd_nxt; | |
922 | } | |
923 | if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) { | |
924 | tp->mdev = dst_metric(dst, RTAX_RTTVAR); | |
925 | tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN); | |
926 | } | |
463c84b9 ACM |
927 | tcp_set_rto(sk); |
928 | tcp_bound_rto(sk); | |
929 | if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp) | |
1da177e4 LT |
930 | goto reset; |
931 | tp->snd_cwnd = tcp_init_cwnd(tp, dst); | |
932 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
933 | return; | |
934 | ||
935 | reset: | |
936 | /* Play conservative. If timestamps are not | |
937 | * supported, TCP will fail to recalculate correct | |
938 | * rtt, if initial rto is too small. FORGET ALL AND RESET! | |
939 | */ | |
940 | if (!tp->rx_opt.saw_tstamp && tp->srtt) { | |
941 | tp->srtt = 0; | |
942 | tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT; | |
463c84b9 | 943 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
1da177e4 LT |
944 | } |
945 | } | |
946 | ||
6687e988 ACM |
947 | static void tcp_update_reordering(struct sock *sk, const int metric, |
948 | const int ts) | |
1da177e4 | 949 | { |
6687e988 | 950 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
951 | if (metric > tp->reordering) { |
952 | tp->reordering = min(TCP_MAX_REORDERING, metric); | |
953 | ||
954 | /* This exciting event is worth to be remembered. 8) */ | |
955 | if (ts) | |
956 | NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER); | |
e60402d0 | 957 | else if (tcp_is_reno(tp)) |
1da177e4 | 958 | NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER); |
e60402d0 | 959 | else if (tcp_is_fack(tp)) |
1da177e4 LT |
960 | NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER); |
961 | else | |
962 | NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER); | |
963 | #if FASTRETRANS_DEBUG > 1 | |
964 | printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", | |
6687e988 | 965 | tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, |
1da177e4 LT |
966 | tp->reordering, |
967 | tp->fackets_out, | |
968 | tp->sacked_out, | |
969 | tp->undo_marker ? tp->undo_retrans : 0); | |
970 | #endif | |
e60402d0 | 971 | tcp_disable_fack(tp); |
1da177e4 LT |
972 | } |
973 | } | |
974 | ||
975 | /* This procedure tags the retransmission queue when SACKs arrive. | |
976 | * | |
977 | * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). | |
978 | * Packets in queue with these bits set are counted in variables | |
979 | * sacked_out, retrans_out and lost_out, correspondingly. | |
980 | * | |
981 | * Valid combinations are: | |
982 | * Tag InFlight Description | |
983 | * 0 1 - orig segment is in flight. | |
984 | * S 0 - nothing flies, orig reached receiver. | |
985 | * L 0 - nothing flies, orig lost by net. | |
986 | * R 2 - both orig and retransmit are in flight. | |
987 | * L|R 1 - orig is lost, retransmit is in flight. | |
988 | * S|R 1 - orig reached receiver, retrans is still in flight. | |
989 | * (L|S|R is logically valid, it could occur when L|R is sacked, | |
990 | * but it is equivalent to plain S and code short-curcuits it to S. | |
991 | * L|S is logically invalid, it would mean -1 packet in flight 8)) | |
992 | * | |
993 | * These 6 states form finite state machine, controlled by the following events: | |
994 | * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) | |
995 | * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) | |
996 | * 3. Loss detection event of one of three flavors: | |
997 | * A. Scoreboard estimator decided the packet is lost. | |
998 | * A'. Reno "three dupacks" marks head of queue lost. | |
999 | * A''. Its FACK modfication, head until snd.fack is lost. | |
1000 | * B. SACK arrives sacking data transmitted after never retransmitted | |
1001 | * hole was sent out. | |
1002 | * C. SACK arrives sacking SND.NXT at the moment, when the | |
1003 | * segment was retransmitted. | |
1004 | * 4. D-SACK added new rule: D-SACK changes any tag to S. | |
1005 | * | |
1006 | * It is pleasant to note, that state diagram turns out to be commutative, | |
1007 | * so that we are allowed not to be bothered by order of our actions, | |
1008 | * when multiple events arrive simultaneously. (see the function below). | |
1009 | * | |
1010 | * Reordering detection. | |
1011 | * -------------------- | |
1012 | * Reordering metric is maximal distance, which a packet can be displaced | |
1013 | * in packet stream. With SACKs we can estimate it: | |
1014 | * | |
1015 | * 1. SACK fills old hole and the corresponding segment was not | |
1016 | * ever retransmitted -> reordering. Alas, we cannot use it | |
1017 | * when segment was retransmitted. | |
1018 | * 2. The last flaw is solved with D-SACK. D-SACK arrives | |
1019 | * for retransmitted and already SACKed segment -> reordering.. | |
1020 | * Both of these heuristics are not used in Loss state, when we cannot | |
1021 | * account for retransmits accurately. | |
1022 | */ | |
d06e021d DM |
1023 | static int tcp_check_dsack(struct tcp_sock *tp, struct sk_buff *ack_skb, |
1024 | struct tcp_sack_block_wire *sp, int num_sacks, | |
1025 | u32 prior_snd_una) | |
1026 | { | |
1027 | u32 start_seq_0 = ntohl(get_unaligned(&sp[0].start_seq)); | |
1028 | u32 end_seq_0 = ntohl(get_unaligned(&sp[0].end_seq)); | |
1029 | int dup_sack = 0; | |
1030 | ||
1031 | if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { | |
1032 | dup_sack = 1; | |
e60402d0 | 1033 | tcp_dsack_seen(tp); |
d06e021d DM |
1034 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV); |
1035 | } else if (num_sacks > 1) { | |
1036 | u32 end_seq_1 = ntohl(get_unaligned(&sp[1].end_seq)); | |
1037 | u32 start_seq_1 = ntohl(get_unaligned(&sp[1].start_seq)); | |
1038 | ||
1039 | if (!after(end_seq_0, end_seq_1) && | |
1040 | !before(start_seq_0, start_seq_1)) { | |
1041 | dup_sack = 1; | |
e60402d0 | 1042 | tcp_dsack_seen(tp); |
d06e021d DM |
1043 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV); |
1044 | } | |
1045 | } | |
1046 | ||
1047 | /* D-SACK for already forgotten data... Do dumb counting. */ | |
1048 | if (dup_sack && | |
1049 | !after(end_seq_0, prior_snd_una) && | |
1050 | after(end_seq_0, tp->undo_marker)) | |
1051 | tp->undo_retrans--; | |
1052 | ||
1053 | return dup_sack; | |
1054 | } | |
1055 | ||
1da177e4 LT |
1056 | static int |
1057 | tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una) | |
1058 | { | |
6687e988 | 1059 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 1060 | struct tcp_sock *tp = tcp_sk(sk); |
9c70220b ACM |
1061 | unsigned char *ptr = (skb_transport_header(ack_skb) + |
1062 | TCP_SKB_CB(ack_skb)->sacked); | |
269bd27e | 1063 | struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2); |
fda03fbb | 1064 | struct sk_buff *cached_skb; |
1da177e4 LT |
1065 | int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3; |
1066 | int reord = tp->packets_out; | |
1067 | int prior_fackets; | |
1068 | u32 lost_retrans = 0; | |
1069 | int flag = 0; | |
7769f406 | 1070 | int found_dup_sack = 0; |
fda03fbb | 1071 | int cached_fack_count; |
1da177e4 | 1072 | int i; |
fda03fbb | 1073 | int first_sack_index; |
1da177e4 | 1074 | |
d738cd8f | 1075 | if (!tp->sacked_out) { |
1da177e4 | 1076 | tp->fackets_out = 0; |
d738cd8f IJ |
1077 | tp->highest_sack = tp->snd_una; |
1078 | } | |
1da177e4 LT |
1079 | prior_fackets = tp->fackets_out; |
1080 | ||
d06e021d DM |
1081 | found_dup_sack = tcp_check_dsack(tp, ack_skb, sp, |
1082 | num_sacks, prior_snd_una); | |
1083 | if (found_dup_sack) | |
49ff4bb4 | 1084 | flag |= FLAG_DSACKING_ACK; |
6f74651a BE |
1085 | |
1086 | /* Eliminate too old ACKs, but take into | |
1087 | * account more or less fresh ones, they can | |
1088 | * contain valid SACK info. | |
1089 | */ | |
1090 | if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) | |
1091 | return 0; | |
1092 | ||
6a438bbe SH |
1093 | /* SACK fastpath: |
1094 | * if the only SACK change is the increase of the end_seq of | |
1095 | * the first block then only apply that SACK block | |
1096 | * and use retrans queue hinting otherwise slowpath */ | |
1097 | flag = 1; | |
6f74651a BE |
1098 | for (i = 0; i < num_sacks; i++) { |
1099 | __be32 start_seq = sp[i].start_seq; | |
1100 | __be32 end_seq = sp[i].end_seq; | |
6a438bbe | 1101 | |
6f74651a | 1102 | if (i == 0) { |
6a438bbe SH |
1103 | if (tp->recv_sack_cache[i].start_seq != start_seq) |
1104 | flag = 0; | |
1105 | } else { | |
1106 | if ((tp->recv_sack_cache[i].start_seq != start_seq) || | |
1107 | (tp->recv_sack_cache[i].end_seq != end_seq)) | |
1108 | flag = 0; | |
1109 | } | |
1110 | tp->recv_sack_cache[i].start_seq = start_seq; | |
1111 | tp->recv_sack_cache[i].end_seq = end_seq; | |
6a438bbe | 1112 | } |
8a3c3a97 BE |
1113 | /* Clear the rest of the cache sack blocks so they won't match mistakenly. */ |
1114 | for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) { | |
1115 | tp->recv_sack_cache[i].start_seq = 0; | |
1116 | tp->recv_sack_cache[i].end_seq = 0; | |
1117 | } | |
6a438bbe | 1118 | |
fda03fbb | 1119 | first_sack_index = 0; |
6a438bbe SH |
1120 | if (flag) |
1121 | num_sacks = 1; | |
1122 | else { | |
1123 | int j; | |
1124 | tp->fastpath_skb_hint = NULL; | |
1125 | ||
1126 | /* order SACK blocks to allow in order walk of the retrans queue */ | |
1127 | for (i = num_sacks-1; i > 0; i--) { | |
1128 | for (j = 0; j < i; j++){ | |
1129 | if (after(ntohl(sp[j].start_seq), | |
1130 | ntohl(sp[j+1].start_seq))){ | |
db3ccdac BE |
1131 | struct tcp_sack_block_wire tmp; |
1132 | ||
1133 | tmp = sp[j]; | |
1134 | sp[j] = sp[j+1]; | |
1135 | sp[j+1] = tmp; | |
fda03fbb BE |
1136 | |
1137 | /* Track where the first SACK block goes to */ | |
1138 | if (j == first_sack_index) | |
1139 | first_sack_index = j+1; | |
6a438bbe SH |
1140 | } |
1141 | ||
1142 | } | |
1143 | } | |
1144 | } | |
1145 | ||
1146 | /* clear flag as used for different purpose in following code */ | |
1147 | flag = 0; | |
1148 | ||
fda03fbb BE |
1149 | /* Use SACK fastpath hint if valid */ |
1150 | cached_skb = tp->fastpath_skb_hint; | |
1151 | cached_fack_count = tp->fastpath_cnt_hint; | |
1152 | if (!cached_skb) { | |
fe067e8a | 1153 | cached_skb = tcp_write_queue_head(sk); |
fda03fbb BE |
1154 | cached_fack_count = 0; |
1155 | } | |
1156 | ||
6a438bbe SH |
1157 | for (i=0; i<num_sacks; i++, sp++) { |
1158 | struct sk_buff *skb; | |
1159 | __u32 start_seq = ntohl(sp->start_seq); | |
1160 | __u32 end_seq = ntohl(sp->end_seq); | |
1161 | int fack_count; | |
7769f406 | 1162 | int dup_sack = (found_dup_sack && (i == first_sack_index)); |
6a438bbe | 1163 | |
fda03fbb BE |
1164 | skb = cached_skb; |
1165 | fack_count = cached_fack_count; | |
1da177e4 LT |
1166 | |
1167 | /* Event "B" in the comment above. */ | |
1168 | if (after(end_seq, tp->high_seq)) | |
1169 | flag |= FLAG_DATA_LOST; | |
1170 | ||
fe067e8a | 1171 | tcp_for_write_queue_from(skb, sk) { |
6475be16 DM |
1172 | int in_sack, pcount; |
1173 | u8 sacked; | |
1da177e4 | 1174 | |
fe067e8a DM |
1175 | if (skb == tcp_send_head(sk)) |
1176 | break; | |
1177 | ||
fda03fbb BE |
1178 | cached_skb = skb; |
1179 | cached_fack_count = fack_count; | |
1180 | if (i == first_sack_index) { | |
1181 | tp->fastpath_skb_hint = skb; | |
1182 | tp->fastpath_cnt_hint = fack_count; | |
1183 | } | |
6a438bbe | 1184 | |
1da177e4 LT |
1185 | /* The retransmission queue is always in order, so |
1186 | * we can short-circuit the walk early. | |
1187 | */ | |
6475be16 | 1188 | if (!before(TCP_SKB_CB(skb)->seq, end_seq)) |
1da177e4 LT |
1189 | break; |
1190 | ||
3c05d92e HX |
1191 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && |
1192 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1193 | ||
6475be16 DM |
1194 | pcount = tcp_skb_pcount(skb); |
1195 | ||
3c05d92e HX |
1196 | if (pcount > 1 && !in_sack && |
1197 | after(TCP_SKB_CB(skb)->end_seq, start_seq)) { | |
6475be16 DM |
1198 | unsigned int pkt_len; |
1199 | ||
3c05d92e HX |
1200 | in_sack = !after(start_seq, |
1201 | TCP_SKB_CB(skb)->seq); | |
1202 | ||
1203 | if (!in_sack) | |
6475be16 DM |
1204 | pkt_len = (start_seq - |
1205 | TCP_SKB_CB(skb)->seq); | |
1206 | else | |
1207 | pkt_len = (end_seq - | |
1208 | TCP_SKB_CB(skb)->seq); | |
7967168c | 1209 | if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size)) |
6475be16 DM |
1210 | break; |
1211 | pcount = tcp_skb_pcount(skb); | |
1212 | } | |
1213 | ||
1214 | fack_count += pcount; | |
1da177e4 | 1215 | |
6475be16 DM |
1216 | sacked = TCP_SKB_CB(skb)->sacked; |
1217 | ||
1da177e4 LT |
1218 | /* Account D-SACK for retransmitted packet. */ |
1219 | if ((dup_sack && in_sack) && | |
1220 | (sacked & TCPCB_RETRANS) && | |
1221 | after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker)) | |
1222 | tp->undo_retrans--; | |
1223 | ||
1224 | /* The frame is ACKed. */ | |
1225 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) { | |
1226 | if (sacked&TCPCB_RETRANS) { | |
1227 | if ((dup_sack && in_sack) && | |
1228 | (sacked&TCPCB_SACKED_ACKED)) | |
1229 | reord = min(fack_count, reord); | |
1230 | } else { | |
1231 | /* If it was in a hole, we detected reordering. */ | |
1232 | if (fack_count < prior_fackets && | |
1233 | !(sacked&TCPCB_SACKED_ACKED)) | |
1234 | reord = min(fack_count, reord); | |
1235 | } | |
1236 | ||
1237 | /* Nothing to do; acked frame is about to be dropped. */ | |
1238 | continue; | |
1239 | } | |
1240 | ||
1241 | if ((sacked&TCPCB_SACKED_RETRANS) && | |
1242 | after(end_seq, TCP_SKB_CB(skb)->ack_seq) && | |
1243 | (!lost_retrans || after(end_seq, lost_retrans))) | |
1244 | lost_retrans = end_seq; | |
1245 | ||
1246 | if (!in_sack) | |
1247 | continue; | |
1248 | ||
1249 | if (!(sacked&TCPCB_SACKED_ACKED)) { | |
1250 | if (sacked & TCPCB_SACKED_RETRANS) { | |
1251 | /* If the segment is not tagged as lost, | |
1252 | * we do not clear RETRANS, believing | |
1253 | * that retransmission is still in flight. | |
1254 | */ | |
1255 | if (sacked & TCPCB_LOST) { | |
1256 | TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); | |
1257 | tp->lost_out -= tcp_skb_pcount(skb); | |
1258 | tp->retrans_out -= tcp_skb_pcount(skb); | |
6a438bbe SH |
1259 | |
1260 | /* clear lost hint */ | |
1261 | tp->retransmit_skb_hint = NULL; | |
1da177e4 LT |
1262 | } |
1263 | } else { | |
1264 | /* New sack for not retransmitted frame, | |
1265 | * which was in hole. It is reordering. | |
1266 | */ | |
1267 | if (!(sacked & TCPCB_RETRANS) && | |
1268 | fack_count < prior_fackets) | |
1269 | reord = min(fack_count, reord); | |
1270 | ||
1271 | if (sacked & TCPCB_LOST) { | |
1272 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; | |
1273 | tp->lost_out -= tcp_skb_pcount(skb); | |
6a438bbe SH |
1274 | |
1275 | /* clear lost hint */ | |
1276 | tp->retransmit_skb_hint = NULL; | |
1da177e4 | 1277 | } |
4dc2665e IJ |
1278 | /* SACK enhanced F-RTO detection. |
1279 | * Set flag if and only if non-rexmitted | |
1280 | * segments below frto_highmark are | |
1281 | * SACKed (RFC4138; Appendix B). | |
1282 | * Clearing correct due to in-order walk | |
1283 | */ | |
1284 | if (after(end_seq, tp->frto_highmark)) { | |
1285 | flag &= ~FLAG_ONLY_ORIG_SACKED; | |
1286 | } else { | |
1287 | if (!(sacked & TCPCB_RETRANS)) | |
1288 | flag |= FLAG_ONLY_ORIG_SACKED; | |
1289 | } | |
1da177e4 LT |
1290 | } |
1291 | ||
1292 | TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED; | |
1293 | flag |= FLAG_DATA_SACKED; | |
1294 | tp->sacked_out += tcp_skb_pcount(skb); | |
1295 | ||
1296 | if (fack_count > tp->fackets_out) | |
1297 | tp->fackets_out = fack_count; | |
d738cd8f IJ |
1298 | |
1299 | if (after(TCP_SKB_CB(skb)->seq, | |
1300 | tp->highest_sack)) | |
1301 | tp->highest_sack = TCP_SKB_CB(skb)->seq; | |
1da177e4 LT |
1302 | } else { |
1303 | if (dup_sack && (sacked&TCPCB_RETRANS)) | |
1304 | reord = min(fack_count, reord); | |
1305 | } | |
1306 | ||
1307 | /* D-SACK. We can detect redundant retransmission | |
1308 | * in S|R and plain R frames and clear it. | |
1309 | * undo_retrans is decreased above, L|R frames | |
1310 | * are accounted above as well. | |
1311 | */ | |
1312 | if (dup_sack && | |
1313 | (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) { | |
1314 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; | |
1315 | tp->retrans_out -= tcp_skb_pcount(skb); | |
6a438bbe | 1316 | tp->retransmit_skb_hint = NULL; |
1da177e4 LT |
1317 | } |
1318 | } | |
1319 | } | |
1320 | ||
1321 | /* Check for lost retransmit. This superb idea is | |
1322 | * borrowed from "ratehalving". Event "C". | |
1323 | * Later note: FACK people cheated me again 8), | |
1324 | * we have to account for reordering! Ugly, | |
1325 | * but should help. | |
1326 | */ | |
6687e988 | 1327 | if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) { |
1da177e4 LT |
1328 | struct sk_buff *skb; |
1329 | ||
fe067e8a DM |
1330 | tcp_for_write_queue(skb, sk) { |
1331 | if (skb == tcp_send_head(sk)) | |
1332 | break; | |
1da177e4 LT |
1333 | if (after(TCP_SKB_CB(skb)->seq, lost_retrans)) |
1334 | break; | |
1335 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1336 | continue; | |
1337 | if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) && | |
1338 | after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) && | |
e60402d0 | 1339 | (tcp_is_fack(tp) || |
1da177e4 LT |
1340 | !before(lost_retrans, |
1341 | TCP_SKB_CB(skb)->ack_seq + tp->reordering * | |
c1b4a7e6 | 1342 | tp->mss_cache))) { |
1da177e4 LT |
1343 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; |
1344 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1345 | ||
6a438bbe SH |
1346 | /* clear lost hint */ |
1347 | tp->retransmit_skb_hint = NULL; | |
1348 | ||
1da177e4 LT |
1349 | if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) { |
1350 | tp->lost_out += tcp_skb_pcount(skb); | |
1351 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1352 | flag |= FLAG_DATA_SACKED; | |
1353 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT); | |
1354 | } | |
1355 | } | |
1356 | } | |
1357 | } | |
1358 | ||
86426c22 IJ |
1359 | tcp_verify_left_out(tp); |
1360 | ||
288035f9 | 1361 | if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss && |
c5e7af0d | 1362 | (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark))) |
6687e988 | 1363 | tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0); |
1da177e4 LT |
1364 | |
1365 | #if FASTRETRANS_DEBUG > 0 | |
1366 | BUG_TRAP((int)tp->sacked_out >= 0); | |
1367 | BUG_TRAP((int)tp->lost_out >= 0); | |
1368 | BUG_TRAP((int)tp->retrans_out >= 0); | |
1369 | BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0); | |
1370 | #endif | |
1371 | return flag; | |
1372 | } | |
1373 | ||
575ee714 IJ |
1374 | /* F-RTO can only be used if TCP has never retransmitted anything other than |
1375 | * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here) | |
30935cf4 | 1376 | */ |
4ddf6676 IJ |
1377 | static void tcp_check_reno_reordering(struct sock *sk, const int addend) |
1378 | { | |
1379 | struct tcp_sock *tp = tcp_sk(sk); | |
1380 | u32 holes; | |
1381 | ||
1382 | holes = max(tp->lost_out, 1U); | |
1383 | holes = min(holes, tp->packets_out); | |
1384 | ||
1385 | if ((tp->sacked_out + holes) > tp->packets_out) { | |
1386 | tp->sacked_out = tp->packets_out - holes; | |
1387 | tcp_update_reordering(sk, tp->packets_out + addend, 0); | |
1388 | } | |
1389 | } | |
1390 | ||
1391 | /* Emulate SACKs for SACKless connection: account for a new dupack. */ | |
1392 | ||
1393 | static void tcp_add_reno_sack(struct sock *sk) | |
1394 | { | |
1395 | struct tcp_sock *tp = tcp_sk(sk); | |
1396 | tp->sacked_out++; | |
1397 | tcp_check_reno_reordering(sk, 0); | |
005903bc | 1398 | tcp_verify_left_out(tp); |
4ddf6676 IJ |
1399 | } |
1400 | ||
1401 | /* Account for ACK, ACKing some data in Reno Recovery phase. */ | |
1402 | ||
1403 | static void tcp_remove_reno_sacks(struct sock *sk, int acked) | |
1404 | { | |
1405 | struct tcp_sock *tp = tcp_sk(sk); | |
1406 | ||
1407 | if (acked > 0) { | |
1408 | /* One ACK acked hole. The rest eat duplicate ACKs. */ | |
1409 | if (acked-1 >= tp->sacked_out) | |
1410 | tp->sacked_out = 0; | |
1411 | else | |
1412 | tp->sacked_out -= acked-1; | |
1413 | } | |
1414 | tcp_check_reno_reordering(sk, acked); | |
005903bc | 1415 | tcp_verify_left_out(tp); |
4ddf6676 IJ |
1416 | } |
1417 | ||
1418 | static inline void tcp_reset_reno_sack(struct tcp_sock *tp) | |
1419 | { | |
1420 | tp->sacked_out = 0; | |
4ddf6676 IJ |
1421 | } |
1422 | ||
46d0de4e | 1423 | int tcp_use_frto(struct sock *sk) |
bdaae17d IJ |
1424 | { |
1425 | const struct tcp_sock *tp = tcp_sk(sk); | |
46d0de4e IJ |
1426 | struct sk_buff *skb; |
1427 | ||
575ee714 | 1428 | if (!sysctl_tcp_frto) |
46d0de4e | 1429 | return 0; |
bdaae17d | 1430 | |
4dc2665e IJ |
1431 | if (IsSackFrto()) |
1432 | return 1; | |
1433 | ||
46d0de4e IJ |
1434 | /* Avoid expensive walking of rexmit queue if possible */ |
1435 | if (tp->retrans_out > 1) | |
1436 | return 0; | |
1437 | ||
fe067e8a DM |
1438 | skb = tcp_write_queue_head(sk); |
1439 | skb = tcp_write_queue_next(sk, skb); /* Skips head */ | |
1440 | tcp_for_write_queue_from(skb, sk) { | |
1441 | if (skb == tcp_send_head(sk)) | |
1442 | break; | |
46d0de4e IJ |
1443 | if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS) |
1444 | return 0; | |
1445 | /* Short-circuit when first non-SACKed skb has been checked */ | |
1446 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) | |
1447 | break; | |
1448 | } | |
1449 | return 1; | |
bdaae17d IJ |
1450 | } |
1451 | ||
30935cf4 IJ |
1452 | /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO |
1453 | * recovery a bit and use heuristics in tcp_process_frto() to detect if | |
d1a54c6a IJ |
1454 | * the RTO was spurious. Only clear SACKED_RETRANS of the head here to |
1455 | * keep retrans_out counting accurate (with SACK F-RTO, other than head | |
1456 | * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS | |
1457 | * bits are handled if the Loss state is really to be entered (in | |
1458 | * tcp_enter_frto_loss). | |
7487c48c IJ |
1459 | * |
1460 | * Do like tcp_enter_loss() would; when RTO expires the second time it | |
1461 | * does: | |
1462 | * "Reduce ssthresh if it has not yet been made inside this window." | |
1da177e4 LT |
1463 | */ |
1464 | void tcp_enter_frto(struct sock *sk) | |
1465 | { | |
6687e988 | 1466 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
1467 | struct tcp_sock *tp = tcp_sk(sk); |
1468 | struct sk_buff *skb; | |
1469 | ||
7487c48c | 1470 | if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) || |
e905a9ed | 1471 | tp->snd_una == tp->high_seq || |
7487c48c IJ |
1472 | ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) && |
1473 | !icsk->icsk_retransmits)) { | |
6687e988 | 1474 | tp->prior_ssthresh = tcp_current_ssthresh(sk); |
66e93e45 IJ |
1475 | /* Our state is too optimistic in ssthresh() call because cwnd |
1476 | * is not reduced until tcp_enter_frto_loss() when previous FRTO | |
1477 | * recovery has not yet completed. Pattern would be this: RTO, | |
1478 | * Cumulative ACK, RTO (2xRTO for the same segment does not end | |
1479 | * up here twice). | |
1480 | * RFC4138 should be more specific on what to do, even though | |
1481 | * RTO is quite unlikely to occur after the first Cumulative ACK | |
1482 | * due to back-off and complexity of triggering events ... | |
1483 | */ | |
1484 | if (tp->frto_counter) { | |
1485 | u32 stored_cwnd; | |
1486 | stored_cwnd = tp->snd_cwnd; | |
1487 | tp->snd_cwnd = 2; | |
1488 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
1489 | tp->snd_cwnd = stored_cwnd; | |
1490 | } else { | |
1491 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
1492 | } | |
1493 | /* ... in theory, cong.control module could do "any tricks" in | |
1494 | * ssthresh(), which means that ca_state, lost bits and lost_out | |
1495 | * counter would have to be faked before the call occurs. We | |
1496 | * consider that too expensive, unlikely and hacky, so modules | |
1497 | * using these in ssthresh() must deal these incompatibility | |
1498 | * issues if they receives CA_EVENT_FRTO and frto_counter != 0 | |
1499 | */ | |
6687e988 | 1500 | tcp_ca_event(sk, CA_EVENT_FRTO); |
1da177e4 LT |
1501 | } |
1502 | ||
1da177e4 LT |
1503 | tp->undo_marker = tp->snd_una; |
1504 | tp->undo_retrans = 0; | |
1505 | ||
fe067e8a | 1506 | skb = tcp_write_queue_head(sk); |
d1a54c6a | 1507 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
522e7548 | 1508 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; |
d1a54c6a | 1509 | tp->retrans_out -= tcp_skb_pcount(skb); |
1da177e4 | 1510 | } |
005903bc | 1511 | tcp_verify_left_out(tp); |
1da177e4 | 1512 | |
4dc2665e IJ |
1513 | /* Earlier loss recovery underway (see RFC4138; Appendix B). |
1514 | * The last condition is necessary at least in tp->frto_counter case. | |
1515 | */ | |
1516 | if (IsSackFrto() && (tp->frto_counter || | |
1517 | ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) && | |
1518 | after(tp->high_seq, tp->snd_una)) { | |
1519 | tp->frto_highmark = tp->high_seq; | |
1520 | } else { | |
1521 | tp->frto_highmark = tp->snd_nxt; | |
1522 | } | |
7b0eb22b IJ |
1523 | tcp_set_ca_state(sk, TCP_CA_Disorder); |
1524 | tp->high_seq = tp->snd_nxt; | |
7487c48c | 1525 | tp->frto_counter = 1; |
1da177e4 LT |
1526 | } |
1527 | ||
1528 | /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, | |
1529 | * which indicates that we should follow the traditional RTO recovery, | |
1530 | * i.e. mark everything lost and do go-back-N retransmission. | |
1531 | */ | |
d1a54c6a | 1532 | static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag) |
1da177e4 LT |
1533 | { |
1534 | struct tcp_sock *tp = tcp_sk(sk); | |
1535 | struct sk_buff *skb; | |
1da177e4 | 1536 | |
1da177e4 | 1537 | tp->lost_out = 0; |
d1a54c6a | 1538 | tp->retrans_out = 0; |
e60402d0 | 1539 | if (tcp_is_reno(tp)) |
9bff40fd | 1540 | tcp_reset_reno_sack(tp); |
1da177e4 | 1541 | |
fe067e8a DM |
1542 | tcp_for_write_queue(skb, sk) { |
1543 | if (skb == tcp_send_head(sk)) | |
1544 | break; | |
d1a54c6a IJ |
1545 | /* |
1546 | * Count the retransmission made on RTO correctly (only when | |
1547 | * waiting for the first ACK and did not get it)... | |
1548 | */ | |
1549 | if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) { | |
0a9f2a46 IJ |
1550 | /* For some reason this R-bit might get cleared? */ |
1551 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
1552 | tp->retrans_out += tcp_skb_pcount(skb); | |
d1a54c6a IJ |
1553 | /* ...enter this if branch just for the first segment */ |
1554 | flag |= FLAG_DATA_ACKED; | |
1555 | } else { | |
1556 | TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); | |
1557 | } | |
1da177e4 | 1558 | |
9bff40fd IJ |
1559 | /* Don't lost mark skbs that were fwd transmitted after RTO */ |
1560 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) && | |
1561 | !after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) { | |
1562 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1563 | tp->lost_out += tcp_skb_pcount(skb); | |
1da177e4 LT |
1564 | } |
1565 | } | |
005903bc | 1566 | tcp_verify_left_out(tp); |
1da177e4 | 1567 | |
95c4922b | 1568 | tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments; |
1da177e4 LT |
1569 | tp->snd_cwnd_cnt = 0; |
1570 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1571 | tp->undo_marker = 0; | |
1572 | tp->frto_counter = 0; | |
1573 | ||
1574 | tp->reordering = min_t(unsigned int, tp->reordering, | |
1575 | sysctl_tcp_reordering); | |
6687e988 | 1576 | tcp_set_ca_state(sk, TCP_CA_Loss); |
1da177e4 LT |
1577 | tp->high_seq = tp->frto_highmark; |
1578 | TCP_ECN_queue_cwr(tp); | |
6a438bbe SH |
1579 | |
1580 | clear_all_retrans_hints(tp); | |
1da177e4 LT |
1581 | } |
1582 | ||
1583 | void tcp_clear_retrans(struct tcp_sock *tp) | |
1584 | { | |
1da177e4 LT |
1585 | tp->retrans_out = 0; |
1586 | ||
1587 | tp->fackets_out = 0; | |
1588 | tp->sacked_out = 0; | |
1589 | tp->lost_out = 0; | |
1590 | ||
1591 | tp->undo_marker = 0; | |
1592 | tp->undo_retrans = 0; | |
1593 | } | |
1594 | ||
1595 | /* Enter Loss state. If "how" is not zero, forget all SACK information | |
1596 | * and reset tags completely, otherwise preserve SACKs. If receiver | |
1597 | * dropped its ofo queue, we will know this due to reneging detection. | |
1598 | */ | |
1599 | void tcp_enter_loss(struct sock *sk, int how) | |
1600 | { | |
6687e988 | 1601 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
1602 | struct tcp_sock *tp = tcp_sk(sk); |
1603 | struct sk_buff *skb; | |
1604 | int cnt = 0; | |
1605 | ||
1606 | /* Reduce ssthresh if it has not yet been made inside this window. */ | |
6687e988 ACM |
1607 | if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || |
1608 | (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { | |
1609 | tp->prior_ssthresh = tcp_current_ssthresh(sk); | |
1610 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
1611 | tcp_ca_event(sk, CA_EVENT_LOSS); | |
1da177e4 LT |
1612 | } |
1613 | tp->snd_cwnd = 1; | |
1614 | tp->snd_cwnd_cnt = 0; | |
1615 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1616 | ||
9772efb9 | 1617 | tp->bytes_acked = 0; |
1da177e4 LT |
1618 | tcp_clear_retrans(tp); |
1619 | ||
1620 | /* Push undo marker, if it was plain RTO and nothing | |
1621 | * was retransmitted. */ | |
1622 | if (!how) | |
1623 | tp->undo_marker = tp->snd_una; | |
1624 | ||
fe067e8a DM |
1625 | tcp_for_write_queue(skb, sk) { |
1626 | if (skb == tcp_send_head(sk)) | |
1627 | break; | |
1da177e4 LT |
1628 | cnt += tcp_skb_pcount(skb); |
1629 | if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS) | |
1630 | tp->undo_marker = 0; | |
1631 | TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; | |
1632 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { | |
1633 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; | |
1634 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1635 | tp->lost_out += tcp_skb_pcount(skb); | |
1636 | } else { | |
1637 | tp->sacked_out += tcp_skb_pcount(skb); | |
1638 | tp->fackets_out = cnt; | |
1639 | } | |
1640 | } | |
005903bc | 1641 | tcp_verify_left_out(tp); |
1da177e4 LT |
1642 | |
1643 | tp->reordering = min_t(unsigned int, tp->reordering, | |
1644 | sysctl_tcp_reordering); | |
6687e988 | 1645 | tcp_set_ca_state(sk, TCP_CA_Loss); |
1da177e4 LT |
1646 | tp->high_seq = tp->snd_nxt; |
1647 | TCP_ECN_queue_cwr(tp); | |
580e572a IJ |
1648 | /* Abort FRTO algorithm if one is in progress */ |
1649 | tp->frto_counter = 0; | |
6a438bbe SH |
1650 | |
1651 | clear_all_retrans_hints(tp); | |
1da177e4 LT |
1652 | } |
1653 | ||
463c84b9 | 1654 | static int tcp_check_sack_reneging(struct sock *sk) |
1da177e4 LT |
1655 | { |
1656 | struct sk_buff *skb; | |
1657 | ||
1658 | /* If ACK arrived pointing to a remembered SACK, | |
1659 | * it means that our remembered SACKs do not reflect | |
1660 | * real state of receiver i.e. | |
1661 | * receiver _host_ is heavily congested (or buggy). | |
1662 | * Do processing similar to RTO timeout. | |
1663 | */ | |
fe067e8a | 1664 | if ((skb = tcp_write_queue_head(sk)) != NULL && |
1da177e4 | 1665 | (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { |
6687e988 | 1666 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
1667 | NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING); |
1668 | ||
1669 | tcp_enter_loss(sk, 1); | |
6687e988 | 1670 | icsk->icsk_retransmits++; |
fe067e8a | 1671 | tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); |
463c84b9 | 1672 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
6687e988 | 1673 | icsk->icsk_rto, TCP_RTO_MAX); |
1da177e4 LT |
1674 | return 1; |
1675 | } | |
1676 | return 0; | |
1677 | } | |
1678 | ||
1679 | static inline int tcp_fackets_out(struct tcp_sock *tp) | |
1680 | { | |
e60402d0 | 1681 | return tcp_is_reno(tp) ? tp->sacked_out+1 : tp->fackets_out; |
1da177e4 LT |
1682 | } |
1683 | ||
463c84b9 | 1684 | static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 1685 | { |
463c84b9 | 1686 | return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto); |
1da177e4 LT |
1687 | } |
1688 | ||
9e412ba7 | 1689 | static inline int tcp_head_timedout(struct sock *sk) |
1da177e4 | 1690 | { |
9e412ba7 IJ |
1691 | struct tcp_sock *tp = tcp_sk(sk); |
1692 | ||
1da177e4 | 1693 | return tp->packets_out && |
fe067e8a | 1694 | tcp_skb_timedout(sk, tcp_write_queue_head(sk)); |
1da177e4 LT |
1695 | } |
1696 | ||
1697 | /* Linux NewReno/SACK/FACK/ECN state machine. | |
1698 | * -------------------------------------- | |
1699 | * | |
1700 | * "Open" Normal state, no dubious events, fast path. | |
1701 | * "Disorder" In all the respects it is "Open", | |
1702 | * but requires a bit more attention. It is entered when | |
1703 | * we see some SACKs or dupacks. It is split of "Open" | |
1704 | * mainly to move some processing from fast path to slow one. | |
1705 | * "CWR" CWND was reduced due to some Congestion Notification event. | |
1706 | * It can be ECN, ICMP source quench, local device congestion. | |
1707 | * "Recovery" CWND was reduced, we are fast-retransmitting. | |
1708 | * "Loss" CWND was reduced due to RTO timeout or SACK reneging. | |
1709 | * | |
1710 | * tcp_fastretrans_alert() is entered: | |
1711 | * - each incoming ACK, if state is not "Open" | |
1712 | * - when arrived ACK is unusual, namely: | |
1713 | * * SACK | |
1714 | * * Duplicate ACK. | |
1715 | * * ECN ECE. | |
1716 | * | |
1717 | * Counting packets in flight is pretty simple. | |
1718 | * | |
1719 | * in_flight = packets_out - left_out + retrans_out | |
1720 | * | |
1721 | * packets_out is SND.NXT-SND.UNA counted in packets. | |
1722 | * | |
1723 | * retrans_out is number of retransmitted segments. | |
1724 | * | |
1725 | * left_out is number of segments left network, but not ACKed yet. | |
1726 | * | |
1727 | * left_out = sacked_out + lost_out | |
1728 | * | |
1729 | * sacked_out: Packets, which arrived to receiver out of order | |
1730 | * and hence not ACKed. With SACKs this number is simply | |
1731 | * amount of SACKed data. Even without SACKs | |
1732 | * it is easy to give pretty reliable estimate of this number, | |
1733 | * counting duplicate ACKs. | |
1734 | * | |
1735 | * lost_out: Packets lost by network. TCP has no explicit | |
1736 | * "loss notification" feedback from network (for now). | |
1737 | * It means that this number can be only _guessed_. | |
1738 | * Actually, it is the heuristics to predict lossage that | |
1739 | * distinguishes different algorithms. | |
1740 | * | |
1741 | * F.e. after RTO, when all the queue is considered as lost, | |
1742 | * lost_out = packets_out and in_flight = retrans_out. | |
1743 | * | |
1744 | * Essentially, we have now two algorithms counting | |
1745 | * lost packets. | |
1746 | * | |
1747 | * FACK: It is the simplest heuristics. As soon as we decided | |
1748 | * that something is lost, we decide that _all_ not SACKed | |
1749 | * packets until the most forward SACK are lost. I.e. | |
1750 | * lost_out = fackets_out - sacked_out and left_out = fackets_out. | |
1751 | * It is absolutely correct estimate, if network does not reorder | |
1752 | * packets. And it loses any connection to reality when reordering | |
1753 | * takes place. We use FACK by default until reordering | |
1754 | * is suspected on the path to this destination. | |
1755 | * | |
1756 | * NewReno: when Recovery is entered, we assume that one segment | |
1757 | * is lost (classic Reno). While we are in Recovery and | |
1758 | * a partial ACK arrives, we assume that one more packet | |
1759 | * is lost (NewReno). This heuristics are the same in NewReno | |
1760 | * and SACK. | |
1761 | * | |
1762 | * Imagine, that's all! Forget about all this shamanism about CWND inflation | |
1763 | * deflation etc. CWND is real congestion window, never inflated, changes | |
1764 | * only according to classic VJ rules. | |
1765 | * | |
1766 | * Really tricky (and requiring careful tuning) part of algorithm | |
1767 | * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). | |
1768 | * The first determines the moment _when_ we should reduce CWND and, | |
1769 | * hence, slow down forward transmission. In fact, it determines the moment | |
1770 | * when we decide that hole is caused by loss, rather than by a reorder. | |
1771 | * | |
1772 | * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill | |
1773 | * holes, caused by lost packets. | |
1774 | * | |
1775 | * And the most logically complicated part of algorithm is undo | |
1776 | * heuristics. We detect false retransmits due to both too early | |
1777 | * fast retransmit (reordering) and underestimated RTO, analyzing | |
1778 | * timestamps and D-SACKs. When we detect that some segments were | |
1779 | * retransmitted by mistake and CWND reduction was wrong, we undo | |
1780 | * window reduction and abort recovery phase. This logic is hidden | |
1781 | * inside several functions named tcp_try_undo_<something>. | |
1782 | */ | |
1783 | ||
1784 | /* This function decides, when we should leave Disordered state | |
1785 | * and enter Recovery phase, reducing congestion window. | |
1786 | * | |
1787 | * Main question: may we further continue forward transmission | |
1788 | * with the same cwnd? | |
1789 | */ | |
9e412ba7 | 1790 | static int tcp_time_to_recover(struct sock *sk) |
1da177e4 | 1791 | { |
9e412ba7 | 1792 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
1793 | __u32 packets_out; |
1794 | ||
52c63f1e IJ |
1795 | /* Do not perform any recovery during FRTO algorithm */ |
1796 | if (tp->frto_counter) | |
1797 | return 0; | |
1798 | ||
1da177e4 LT |
1799 | /* Trick#1: The loss is proven. */ |
1800 | if (tp->lost_out) | |
1801 | return 1; | |
1802 | ||
1803 | /* Not-A-Trick#2 : Classic rule... */ | |
1804 | if (tcp_fackets_out(tp) > tp->reordering) | |
1805 | return 1; | |
1806 | ||
1807 | /* Trick#3 : when we use RFC2988 timer restart, fast | |
1808 | * retransmit can be triggered by timeout of queue head. | |
1809 | */ | |
9e412ba7 | 1810 | if (tcp_head_timedout(sk)) |
1da177e4 LT |
1811 | return 1; |
1812 | ||
1813 | /* Trick#4: It is still not OK... But will it be useful to delay | |
1814 | * recovery more? | |
1815 | */ | |
1816 | packets_out = tp->packets_out; | |
1817 | if (packets_out <= tp->reordering && | |
1818 | tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && | |
9e412ba7 | 1819 | !tcp_may_send_now(sk)) { |
1da177e4 LT |
1820 | /* We have nothing to send. This connection is limited |
1821 | * either by receiver window or by application. | |
1822 | */ | |
1823 | return 1; | |
1824 | } | |
1825 | ||
1826 | return 0; | |
1827 | } | |
1828 | ||
d8f4f223 IJ |
1829 | /* RFC: This is from the original, I doubt that this is necessary at all: |
1830 | * clear xmit_retrans hint if seq of this skb is beyond hint. How could we | |
1831 | * retransmitted past LOST markings in the first place? I'm not fully sure | |
1832 | * about undo and end of connection cases, which can cause R without L? | |
1833 | */ | |
1834 | static void tcp_verify_retransmit_hint(struct tcp_sock *tp, | |
1835 | struct sk_buff *skb) | |
1836 | { | |
1837 | if ((tp->retransmit_skb_hint != NULL) && | |
1838 | before(TCP_SKB_CB(skb)->seq, | |
1839 | TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) | |
19b2b486 | 1840 | tp->retransmit_skb_hint = NULL; |
d8f4f223 IJ |
1841 | } |
1842 | ||
1da177e4 | 1843 | /* Mark head of queue up as lost. */ |
9e412ba7 | 1844 | static void tcp_mark_head_lost(struct sock *sk, |
1da177e4 LT |
1845 | int packets, u32 high_seq) |
1846 | { | |
9e412ba7 | 1847 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 1848 | struct sk_buff *skb; |
6a438bbe | 1849 | int cnt; |
1da177e4 | 1850 | |
6a438bbe SH |
1851 | BUG_TRAP(packets <= tp->packets_out); |
1852 | if (tp->lost_skb_hint) { | |
1853 | skb = tp->lost_skb_hint; | |
1854 | cnt = tp->lost_cnt_hint; | |
1855 | } else { | |
fe067e8a | 1856 | skb = tcp_write_queue_head(sk); |
6a438bbe SH |
1857 | cnt = 0; |
1858 | } | |
1da177e4 | 1859 | |
fe067e8a DM |
1860 | tcp_for_write_queue_from(skb, sk) { |
1861 | if (skb == tcp_send_head(sk)) | |
1862 | break; | |
6a438bbe SH |
1863 | /* TODO: do this better */ |
1864 | /* this is not the most efficient way to do this... */ | |
1865 | tp->lost_skb_hint = skb; | |
1866 | tp->lost_cnt_hint = cnt; | |
1867 | cnt += tcp_skb_pcount(skb); | |
1868 | if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq)) | |
1da177e4 LT |
1869 | break; |
1870 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) { | |
1871 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1872 | tp->lost_out += tcp_skb_pcount(skb); | |
d8f4f223 | 1873 | tcp_verify_retransmit_hint(tp, skb); |
1da177e4 LT |
1874 | } |
1875 | } | |
005903bc | 1876 | tcp_verify_left_out(tp); |
1da177e4 LT |
1877 | } |
1878 | ||
1879 | /* Account newly detected lost packet(s) */ | |
1880 | ||
9e412ba7 | 1881 | static void tcp_update_scoreboard(struct sock *sk) |
1da177e4 | 1882 | { |
9e412ba7 IJ |
1883 | struct tcp_sock *tp = tcp_sk(sk); |
1884 | ||
e60402d0 | 1885 | if (tcp_is_fack(tp)) { |
1da177e4 LT |
1886 | int lost = tp->fackets_out - tp->reordering; |
1887 | if (lost <= 0) | |
1888 | lost = 1; | |
9e412ba7 | 1889 | tcp_mark_head_lost(sk, lost, tp->high_seq); |
1da177e4 | 1890 | } else { |
9e412ba7 | 1891 | tcp_mark_head_lost(sk, 1, tp->high_seq); |
1da177e4 LT |
1892 | } |
1893 | ||
1894 | /* New heuristics: it is possible only after we switched | |
1895 | * to restart timer each time when something is ACKed. | |
1896 | * Hence, we can detect timed out packets during fast | |
1897 | * retransmit without falling to slow start. | |
1898 | */ | |
e60402d0 | 1899 | if (!tcp_is_reno(tp) && tcp_head_timedout(sk)) { |
1da177e4 LT |
1900 | struct sk_buff *skb; |
1901 | ||
6a438bbe | 1902 | skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint |
fe067e8a | 1903 | : tcp_write_queue_head(sk); |
6a438bbe | 1904 | |
fe067e8a DM |
1905 | tcp_for_write_queue_from(skb, sk) { |
1906 | if (skb == tcp_send_head(sk)) | |
1907 | break; | |
6a438bbe SH |
1908 | if (!tcp_skb_timedout(sk, skb)) |
1909 | break; | |
1910 | ||
1911 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) { | |
1da177e4 LT |
1912 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; |
1913 | tp->lost_out += tcp_skb_pcount(skb); | |
d8f4f223 | 1914 | tcp_verify_retransmit_hint(tp, skb); |
1da177e4 LT |
1915 | } |
1916 | } | |
6a438bbe SH |
1917 | |
1918 | tp->scoreboard_skb_hint = skb; | |
1919 | ||
005903bc | 1920 | tcp_verify_left_out(tp); |
1da177e4 LT |
1921 | } |
1922 | } | |
1923 | ||
1924 | /* CWND moderation, preventing bursts due to too big ACKs | |
1925 | * in dubious situations. | |
1926 | */ | |
1927 | static inline void tcp_moderate_cwnd(struct tcp_sock *tp) | |
1928 | { | |
1929 | tp->snd_cwnd = min(tp->snd_cwnd, | |
1930 | tcp_packets_in_flight(tp)+tcp_max_burst(tp)); | |
1931 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1932 | } | |
1933 | ||
72dc5b92 SH |
1934 | /* Lower bound on congestion window is slow start threshold |
1935 | * unless congestion avoidance choice decides to overide it. | |
1936 | */ | |
1937 | static inline u32 tcp_cwnd_min(const struct sock *sk) | |
1938 | { | |
1939 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; | |
1940 | ||
1941 | return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh; | |
1942 | } | |
1943 | ||
1da177e4 | 1944 | /* Decrease cwnd each second ack. */ |
1e757f99 | 1945 | static void tcp_cwnd_down(struct sock *sk, int flag) |
1da177e4 | 1946 | { |
6687e988 | 1947 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 1948 | int decr = tp->snd_cwnd_cnt + 1; |
1da177e4 | 1949 | |
49ff4bb4 | 1950 | if ((flag&(FLAG_ANY_PROGRESS|FLAG_DSACKING_ACK)) || |
e60402d0 | 1951 | (tcp_is_reno(tp) && !(flag&FLAG_NOT_DUP))) { |
1e757f99 IJ |
1952 | tp->snd_cwnd_cnt = decr&1; |
1953 | decr >>= 1; | |
1da177e4 | 1954 | |
1e757f99 IJ |
1955 | if (decr && tp->snd_cwnd > tcp_cwnd_min(sk)) |
1956 | tp->snd_cwnd -= decr; | |
1da177e4 | 1957 | |
1e757f99 IJ |
1958 | tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1); |
1959 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1960 | } | |
1da177e4 LT |
1961 | } |
1962 | ||
1963 | /* Nothing was retransmitted or returned timestamp is less | |
1964 | * than timestamp of the first retransmission. | |
1965 | */ | |
1966 | static inline int tcp_packet_delayed(struct tcp_sock *tp) | |
1967 | { | |
1968 | return !tp->retrans_stamp || | |
1969 | (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && | |
1970 | (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0); | |
1971 | } | |
1972 | ||
1973 | /* Undo procedures. */ | |
1974 | ||
1975 | #if FASTRETRANS_DEBUG > 1 | |
9e412ba7 | 1976 | static void DBGUNDO(struct sock *sk, const char *msg) |
1da177e4 | 1977 | { |
9e412ba7 | 1978 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 1979 | struct inet_sock *inet = inet_sk(sk); |
9e412ba7 | 1980 | |
1da177e4 LT |
1981 | printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n", |
1982 | msg, | |
1983 | NIPQUAD(inet->daddr), ntohs(inet->dport), | |
83ae4088 | 1984 | tp->snd_cwnd, tcp_left_out(tp), |
1da177e4 LT |
1985 | tp->snd_ssthresh, tp->prior_ssthresh, |
1986 | tp->packets_out); | |
1987 | } | |
1988 | #else | |
1989 | #define DBGUNDO(x...) do { } while (0) | |
1990 | #endif | |
1991 | ||
6687e988 | 1992 | static void tcp_undo_cwr(struct sock *sk, const int undo) |
1da177e4 | 1993 | { |
6687e988 ACM |
1994 | struct tcp_sock *tp = tcp_sk(sk); |
1995 | ||
1da177e4 | 1996 | if (tp->prior_ssthresh) { |
6687e988 ACM |
1997 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1998 | ||
1999 | if (icsk->icsk_ca_ops->undo_cwnd) | |
2000 | tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); | |
1da177e4 LT |
2001 | else |
2002 | tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1); | |
2003 | ||
2004 | if (undo && tp->prior_ssthresh > tp->snd_ssthresh) { | |
2005 | tp->snd_ssthresh = tp->prior_ssthresh; | |
2006 | TCP_ECN_withdraw_cwr(tp); | |
2007 | } | |
2008 | } else { | |
2009 | tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); | |
2010 | } | |
2011 | tcp_moderate_cwnd(tp); | |
2012 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
6a438bbe SH |
2013 | |
2014 | /* There is something screwy going on with the retrans hints after | |
2015 | an undo */ | |
2016 | clear_all_retrans_hints(tp); | |
1da177e4 LT |
2017 | } |
2018 | ||
2019 | static inline int tcp_may_undo(struct tcp_sock *tp) | |
2020 | { | |
2021 | return tp->undo_marker && | |
2022 | (!tp->undo_retrans || tcp_packet_delayed(tp)); | |
2023 | } | |
2024 | ||
2025 | /* People celebrate: "We love our President!" */ | |
9e412ba7 | 2026 | static int tcp_try_undo_recovery(struct sock *sk) |
1da177e4 | 2027 | { |
9e412ba7 IJ |
2028 | struct tcp_sock *tp = tcp_sk(sk); |
2029 | ||
1da177e4 LT |
2030 | if (tcp_may_undo(tp)) { |
2031 | /* Happy end! We did not retransmit anything | |
2032 | * or our original transmission succeeded. | |
2033 | */ | |
9e412ba7 | 2034 | DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); |
6687e988 ACM |
2035 | tcp_undo_cwr(sk, 1); |
2036 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) | |
1da177e4 LT |
2037 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); |
2038 | else | |
2039 | NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO); | |
2040 | tp->undo_marker = 0; | |
2041 | } | |
e60402d0 | 2042 | if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { |
1da177e4 LT |
2043 | /* Hold old state until something *above* high_seq |
2044 | * is ACKed. For Reno it is MUST to prevent false | |
2045 | * fast retransmits (RFC2582). SACK TCP is safe. */ | |
2046 | tcp_moderate_cwnd(tp); | |
2047 | return 1; | |
2048 | } | |
6687e988 | 2049 | tcp_set_ca_state(sk, TCP_CA_Open); |
1da177e4 LT |
2050 | return 0; |
2051 | } | |
2052 | ||
2053 | /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ | |
9e412ba7 | 2054 | static void tcp_try_undo_dsack(struct sock *sk) |
1da177e4 | 2055 | { |
9e412ba7 IJ |
2056 | struct tcp_sock *tp = tcp_sk(sk); |
2057 | ||
1da177e4 | 2058 | if (tp->undo_marker && !tp->undo_retrans) { |
9e412ba7 | 2059 | DBGUNDO(sk, "D-SACK"); |
6687e988 | 2060 | tcp_undo_cwr(sk, 1); |
1da177e4 LT |
2061 | tp->undo_marker = 0; |
2062 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO); | |
2063 | } | |
2064 | } | |
2065 | ||
2066 | /* Undo during fast recovery after partial ACK. */ | |
2067 | ||
9e412ba7 | 2068 | static int tcp_try_undo_partial(struct sock *sk, int acked) |
1da177e4 | 2069 | { |
9e412ba7 | 2070 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 2071 | /* Partial ACK arrived. Force Hoe's retransmit. */ |
e60402d0 | 2072 | int failed = tcp_is_reno(tp) || tp->fackets_out>tp->reordering; |
1da177e4 LT |
2073 | |
2074 | if (tcp_may_undo(tp)) { | |
2075 | /* Plain luck! Hole if filled with delayed | |
2076 | * packet, rather than with a retransmit. | |
2077 | */ | |
2078 | if (tp->retrans_out == 0) | |
2079 | tp->retrans_stamp = 0; | |
2080 | ||
6687e988 | 2081 | tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); |
1da177e4 | 2082 | |
9e412ba7 | 2083 | DBGUNDO(sk, "Hoe"); |
6687e988 | 2084 | tcp_undo_cwr(sk, 0); |
1da177e4 LT |
2085 | NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO); |
2086 | ||
2087 | /* So... Do not make Hoe's retransmit yet. | |
2088 | * If the first packet was delayed, the rest | |
2089 | * ones are most probably delayed as well. | |
2090 | */ | |
2091 | failed = 0; | |
2092 | } | |
2093 | return failed; | |
2094 | } | |
2095 | ||
2096 | /* Undo during loss recovery after partial ACK. */ | |
9e412ba7 | 2097 | static int tcp_try_undo_loss(struct sock *sk) |
1da177e4 | 2098 | { |
9e412ba7 IJ |
2099 | struct tcp_sock *tp = tcp_sk(sk); |
2100 | ||
1da177e4 LT |
2101 | if (tcp_may_undo(tp)) { |
2102 | struct sk_buff *skb; | |
fe067e8a DM |
2103 | tcp_for_write_queue(skb, sk) { |
2104 | if (skb == tcp_send_head(sk)) | |
2105 | break; | |
1da177e4 LT |
2106 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; |
2107 | } | |
6a438bbe SH |
2108 | |
2109 | clear_all_retrans_hints(tp); | |
2110 | ||
9e412ba7 | 2111 | DBGUNDO(sk, "partial loss"); |
1da177e4 | 2112 | tp->lost_out = 0; |
6687e988 | 2113 | tcp_undo_cwr(sk, 1); |
1da177e4 | 2114 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); |
463c84b9 | 2115 | inet_csk(sk)->icsk_retransmits = 0; |
1da177e4 | 2116 | tp->undo_marker = 0; |
e60402d0 | 2117 | if (tcp_is_sack(tp)) |
6687e988 | 2118 | tcp_set_ca_state(sk, TCP_CA_Open); |
1da177e4 LT |
2119 | return 1; |
2120 | } | |
2121 | return 0; | |
2122 | } | |
2123 | ||
6687e988 | 2124 | static inline void tcp_complete_cwr(struct sock *sk) |
1da177e4 | 2125 | { |
6687e988 | 2126 | struct tcp_sock *tp = tcp_sk(sk); |
317a76f9 | 2127 | tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); |
1da177e4 | 2128 | tp->snd_cwnd_stamp = tcp_time_stamp; |
6687e988 | 2129 | tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); |
1da177e4 LT |
2130 | } |
2131 | ||
9e412ba7 | 2132 | static void tcp_try_to_open(struct sock *sk, int flag) |
1da177e4 | 2133 | { |
9e412ba7 IJ |
2134 | struct tcp_sock *tp = tcp_sk(sk); |
2135 | ||
86426c22 IJ |
2136 | tcp_verify_left_out(tp); |
2137 | ||
1da177e4 LT |
2138 | if (tp->retrans_out == 0) |
2139 | tp->retrans_stamp = 0; | |
2140 | ||
2141 | if (flag&FLAG_ECE) | |
3cfe3baa | 2142 | tcp_enter_cwr(sk, 1); |
1da177e4 | 2143 | |
6687e988 | 2144 | if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { |
1da177e4 LT |
2145 | int state = TCP_CA_Open; |
2146 | ||
d02596e3 | 2147 | if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker) |
1da177e4 LT |
2148 | state = TCP_CA_Disorder; |
2149 | ||
6687e988 ACM |
2150 | if (inet_csk(sk)->icsk_ca_state != state) { |
2151 | tcp_set_ca_state(sk, state); | |
1da177e4 LT |
2152 | tp->high_seq = tp->snd_nxt; |
2153 | } | |
2154 | tcp_moderate_cwnd(tp); | |
2155 | } else { | |
1e757f99 | 2156 | tcp_cwnd_down(sk, flag); |
1da177e4 LT |
2157 | } |
2158 | } | |
2159 | ||
5d424d5a JH |
2160 | static void tcp_mtup_probe_failed(struct sock *sk) |
2161 | { | |
2162 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2163 | ||
2164 | icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; | |
2165 | icsk->icsk_mtup.probe_size = 0; | |
2166 | } | |
2167 | ||
2168 | static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb) | |
2169 | { | |
2170 | struct tcp_sock *tp = tcp_sk(sk); | |
2171 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2172 | ||
2173 | /* FIXME: breaks with very large cwnd */ | |
2174 | tp->prior_ssthresh = tcp_current_ssthresh(sk); | |
2175 | tp->snd_cwnd = tp->snd_cwnd * | |
2176 | tcp_mss_to_mtu(sk, tp->mss_cache) / | |
2177 | icsk->icsk_mtup.probe_size; | |
2178 | tp->snd_cwnd_cnt = 0; | |
2179 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
2180 | tp->rcv_ssthresh = tcp_current_ssthresh(sk); | |
2181 | ||
2182 | icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; | |
2183 | icsk->icsk_mtup.probe_size = 0; | |
2184 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); | |
2185 | } | |
2186 | ||
2187 | ||
1da177e4 LT |
2188 | /* Process an event, which can update packets-in-flight not trivially. |
2189 | * Main goal of this function is to calculate new estimate for left_out, | |
2190 | * taking into account both packets sitting in receiver's buffer and | |
2191 | * packets lost by network. | |
2192 | * | |
2193 | * Besides that it does CWND reduction, when packet loss is detected | |
2194 | * and changes state of machine. | |
2195 | * | |
2196 | * It does _not_ decide what to send, it is made in function | |
2197 | * tcp_xmit_retransmit_queue(). | |
2198 | */ | |
2199 | static void | |
1b6d427b | 2200 | tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag) |
1da177e4 | 2201 | { |
6687e988 | 2202 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 2203 | struct tcp_sock *tp = tcp_sk(sk); |
2e605294 IJ |
2204 | int is_dupack = !(flag&(FLAG_SND_UNA_ADVANCED|FLAG_NOT_DUP)); |
2205 | int do_lost = is_dupack || ((flag&FLAG_DATA_SACKED) && | |
2206 | (tp->fackets_out > tp->reordering)); | |
1da177e4 LT |
2207 | |
2208 | /* Some technical things: | |
2209 | * 1. Reno does not count dupacks (sacked_out) automatically. */ | |
2210 | if (!tp->packets_out) | |
2211 | tp->sacked_out = 0; | |
e905a9ed | 2212 | /* 2. SACK counts snd_fack in packets inaccurately. */ |
1da177e4 LT |
2213 | if (tp->sacked_out == 0) |
2214 | tp->fackets_out = 0; | |
2215 | ||
e905a9ed | 2216 | /* Now state machine starts. |
1da177e4 LT |
2217 | * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ |
2218 | if (flag&FLAG_ECE) | |
2219 | tp->prior_ssthresh = 0; | |
2220 | ||
2221 | /* B. In all the states check for reneging SACKs. */ | |
463c84b9 | 2222 | if (tp->sacked_out && tcp_check_sack_reneging(sk)) |
1da177e4 LT |
2223 | return; |
2224 | ||
2225 | /* C. Process data loss notification, provided it is valid. */ | |
2226 | if ((flag&FLAG_DATA_LOST) && | |
2227 | before(tp->snd_una, tp->high_seq) && | |
6687e988 | 2228 | icsk->icsk_ca_state != TCP_CA_Open && |
1da177e4 | 2229 | tp->fackets_out > tp->reordering) { |
9e412ba7 | 2230 | tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, tp->high_seq); |
1da177e4 LT |
2231 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSS); |
2232 | } | |
2233 | ||
005903bc IJ |
2234 | /* D. Check consistency of the current state. */ |
2235 | tcp_verify_left_out(tp); | |
1da177e4 LT |
2236 | |
2237 | /* E. Check state exit conditions. State can be terminated | |
2238 | * when high_seq is ACKed. */ | |
6687e988 | 2239 | if (icsk->icsk_ca_state == TCP_CA_Open) { |
7b0eb22b | 2240 | BUG_TRAP(tp->retrans_out == 0); |
1da177e4 LT |
2241 | tp->retrans_stamp = 0; |
2242 | } else if (!before(tp->snd_una, tp->high_seq)) { | |
6687e988 | 2243 | switch (icsk->icsk_ca_state) { |
1da177e4 | 2244 | case TCP_CA_Loss: |
6687e988 | 2245 | icsk->icsk_retransmits = 0; |
9e412ba7 | 2246 | if (tcp_try_undo_recovery(sk)) |
1da177e4 LT |
2247 | return; |
2248 | break; | |
2249 | ||
2250 | case TCP_CA_CWR: | |
2251 | /* CWR is to be held something *above* high_seq | |
2252 | * is ACKed for CWR bit to reach receiver. */ | |
2253 | if (tp->snd_una != tp->high_seq) { | |
6687e988 ACM |
2254 | tcp_complete_cwr(sk); |
2255 | tcp_set_ca_state(sk, TCP_CA_Open); | |
1da177e4 LT |
2256 | } |
2257 | break; | |
2258 | ||
2259 | case TCP_CA_Disorder: | |
9e412ba7 | 2260 | tcp_try_undo_dsack(sk); |
1da177e4 LT |
2261 | if (!tp->undo_marker || |
2262 | /* For SACK case do not Open to allow to undo | |
2263 | * catching for all duplicate ACKs. */ | |
e60402d0 | 2264 | tcp_is_reno(tp) || tp->snd_una != tp->high_seq) { |
1da177e4 | 2265 | tp->undo_marker = 0; |
6687e988 | 2266 | tcp_set_ca_state(sk, TCP_CA_Open); |
1da177e4 LT |
2267 | } |
2268 | break; | |
2269 | ||
2270 | case TCP_CA_Recovery: | |
e60402d0 | 2271 | if (tcp_is_reno(tp)) |
1da177e4 | 2272 | tcp_reset_reno_sack(tp); |
9e412ba7 | 2273 | if (tcp_try_undo_recovery(sk)) |
1da177e4 | 2274 | return; |
6687e988 | 2275 | tcp_complete_cwr(sk); |
1da177e4 LT |
2276 | break; |
2277 | } | |
2278 | } | |
2279 | ||
2280 | /* F. Process state. */ | |
6687e988 | 2281 | switch (icsk->icsk_ca_state) { |
1da177e4 | 2282 | case TCP_CA_Recovery: |
2e605294 | 2283 | if (!(flag & FLAG_SND_UNA_ADVANCED)) { |
e60402d0 | 2284 | if (tcp_is_reno(tp) && is_dupack) |
6687e988 | 2285 | tcp_add_reno_sack(sk); |
1b6d427b IJ |
2286 | } else |
2287 | do_lost = tcp_try_undo_partial(sk, pkts_acked); | |
1da177e4 LT |
2288 | break; |
2289 | case TCP_CA_Loss: | |
2290 | if (flag&FLAG_DATA_ACKED) | |
6687e988 | 2291 | icsk->icsk_retransmits = 0; |
9e412ba7 | 2292 | if (!tcp_try_undo_loss(sk)) { |
1da177e4 LT |
2293 | tcp_moderate_cwnd(tp); |
2294 | tcp_xmit_retransmit_queue(sk); | |
2295 | return; | |
2296 | } | |
6687e988 | 2297 | if (icsk->icsk_ca_state != TCP_CA_Open) |
1da177e4 LT |
2298 | return; |
2299 | /* Loss is undone; fall through to processing in Open state. */ | |
2300 | default: | |
e60402d0 | 2301 | if (tcp_is_reno(tp)) { |
2e605294 | 2302 | if (flag & FLAG_SND_UNA_ADVANCED) |
1da177e4 LT |
2303 | tcp_reset_reno_sack(tp); |
2304 | if (is_dupack) | |
6687e988 | 2305 | tcp_add_reno_sack(sk); |
1da177e4 LT |
2306 | } |
2307 | ||
6687e988 | 2308 | if (icsk->icsk_ca_state == TCP_CA_Disorder) |
9e412ba7 | 2309 | tcp_try_undo_dsack(sk); |
1da177e4 | 2310 | |
9e412ba7 IJ |
2311 | if (!tcp_time_to_recover(sk)) { |
2312 | tcp_try_to_open(sk, flag); | |
1da177e4 LT |
2313 | return; |
2314 | } | |
2315 | ||
5d424d5a JH |
2316 | /* MTU probe failure: don't reduce cwnd */ |
2317 | if (icsk->icsk_ca_state < TCP_CA_CWR && | |
2318 | icsk->icsk_mtup.probe_size && | |
0e7b1368 | 2319 | tp->snd_una == tp->mtu_probe.probe_seq_start) { |
5d424d5a JH |
2320 | tcp_mtup_probe_failed(sk); |
2321 | /* Restores the reduction we did in tcp_mtup_probe() */ | |
2322 | tp->snd_cwnd++; | |
2323 | tcp_simple_retransmit(sk); | |
2324 | return; | |
2325 | } | |
2326 | ||
1da177e4 LT |
2327 | /* Otherwise enter Recovery state */ |
2328 | ||
e60402d0 | 2329 | if (tcp_is_reno(tp)) |
1da177e4 LT |
2330 | NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY); |
2331 | else | |
2332 | NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY); | |
2333 | ||
2334 | tp->high_seq = tp->snd_nxt; | |
2335 | tp->prior_ssthresh = 0; | |
2336 | tp->undo_marker = tp->snd_una; | |
2337 | tp->undo_retrans = tp->retrans_out; | |
2338 | ||
6687e988 | 2339 | if (icsk->icsk_ca_state < TCP_CA_CWR) { |
1da177e4 | 2340 | if (!(flag&FLAG_ECE)) |
6687e988 ACM |
2341 | tp->prior_ssthresh = tcp_current_ssthresh(sk); |
2342 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
1da177e4 LT |
2343 | TCP_ECN_queue_cwr(tp); |
2344 | } | |
2345 | ||
9772efb9 | 2346 | tp->bytes_acked = 0; |
1da177e4 | 2347 | tp->snd_cwnd_cnt = 0; |
6687e988 | 2348 | tcp_set_ca_state(sk, TCP_CA_Recovery); |
1da177e4 LT |
2349 | } |
2350 | ||
2e605294 | 2351 | if (do_lost || tcp_head_timedout(sk)) |
9e412ba7 | 2352 | tcp_update_scoreboard(sk); |
1e757f99 | 2353 | tcp_cwnd_down(sk, flag); |
1da177e4 LT |
2354 | tcp_xmit_retransmit_queue(sk); |
2355 | } | |
2356 | ||
2357 | /* Read draft-ietf-tcplw-high-performance before mucking | |
caa20d9a | 2358 | * with this code. (Supersedes RFC1323) |
1da177e4 | 2359 | */ |
2d2abbab | 2360 | static void tcp_ack_saw_tstamp(struct sock *sk, int flag) |
1da177e4 | 2361 | { |
1da177e4 LT |
2362 | /* RTTM Rule: A TSecr value received in a segment is used to |
2363 | * update the averaged RTT measurement only if the segment | |
2364 | * acknowledges some new data, i.e., only if it advances the | |
2365 | * left edge of the send window. | |
2366 | * | |
2367 | * See draft-ietf-tcplw-high-performance-00, section 3.3. | |
2368 | * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> | |
2369 | * | |
2370 | * Changed: reset backoff as soon as we see the first valid sample. | |
caa20d9a | 2371 | * If we do not, we get strongly overestimated rto. With timestamps |
1da177e4 LT |
2372 | * samples are accepted even from very old segments: f.e., when rtt=1 |
2373 | * increases to 8, we retransmit 5 times and after 8 seconds delayed | |
2374 | * answer arrives rto becomes 120 seconds! If at least one of segments | |
2375 | * in window is lost... Voila. --ANK (010210) | |
2376 | */ | |
463c84b9 ACM |
2377 | struct tcp_sock *tp = tcp_sk(sk); |
2378 | const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr; | |
2d2abbab | 2379 | tcp_rtt_estimator(sk, seq_rtt); |
463c84b9 ACM |
2380 | tcp_set_rto(sk); |
2381 | inet_csk(sk)->icsk_backoff = 0; | |
2382 | tcp_bound_rto(sk); | |
1da177e4 LT |
2383 | } |
2384 | ||
2d2abbab | 2385 | static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) |
1da177e4 LT |
2386 | { |
2387 | /* We don't have a timestamp. Can only use | |
2388 | * packets that are not retransmitted to determine | |
2389 | * rtt estimates. Also, we must not reset the | |
2390 | * backoff for rto until we get a non-retransmitted | |
2391 | * packet. This allows us to deal with a situation | |
2392 | * where the network delay has increased suddenly. | |
2393 | * I.e. Karn's algorithm. (SIGCOMM '87, p5.) | |
2394 | */ | |
2395 | ||
2396 | if (flag & FLAG_RETRANS_DATA_ACKED) | |
2397 | return; | |
2398 | ||
2d2abbab | 2399 | tcp_rtt_estimator(sk, seq_rtt); |
463c84b9 ACM |
2400 | tcp_set_rto(sk); |
2401 | inet_csk(sk)->icsk_backoff = 0; | |
2402 | tcp_bound_rto(sk); | |
1da177e4 LT |
2403 | } |
2404 | ||
463c84b9 | 2405 | static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, |
2d2abbab | 2406 | const s32 seq_rtt) |
1da177e4 | 2407 | { |
463c84b9 | 2408 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
2409 | /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ |
2410 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) | |
2d2abbab | 2411 | tcp_ack_saw_tstamp(sk, flag); |
1da177e4 | 2412 | else if (seq_rtt >= 0) |
2d2abbab | 2413 | tcp_ack_no_tstamp(sk, seq_rtt, flag); |
1da177e4 LT |
2414 | } |
2415 | ||
16751347 | 2416 | static void tcp_cong_avoid(struct sock *sk, u32 ack, |
40efc6fa | 2417 | u32 in_flight, int good) |
1da177e4 | 2418 | { |
6687e988 | 2419 | const struct inet_connection_sock *icsk = inet_csk(sk); |
16751347 | 2420 | icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight, good); |
6687e988 | 2421 | tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; |
1da177e4 LT |
2422 | } |
2423 | ||
1da177e4 LT |
2424 | /* Restart timer after forward progress on connection. |
2425 | * RFC2988 recommends to restart timer to now+rto. | |
2426 | */ | |
6728e7dc | 2427 | static void tcp_rearm_rto(struct sock *sk) |
1da177e4 | 2428 | { |
9e412ba7 IJ |
2429 | struct tcp_sock *tp = tcp_sk(sk); |
2430 | ||
1da177e4 | 2431 | if (!tp->packets_out) { |
463c84b9 | 2432 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); |
1da177e4 | 2433 | } else { |
3f421baa | 2434 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX); |
1da177e4 LT |
2435 | } |
2436 | } | |
2437 | ||
1da177e4 LT |
2438 | static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb, |
2439 | __u32 now, __s32 *seq_rtt) | |
2440 | { | |
2441 | struct tcp_sock *tp = tcp_sk(sk); | |
e905a9ed | 2442 | struct tcp_skb_cb *scb = TCP_SKB_CB(skb); |
1da177e4 LT |
2443 | __u32 seq = tp->snd_una; |
2444 | __u32 packets_acked; | |
2445 | int acked = 0; | |
2446 | ||
2447 | /* If we get here, the whole TSO packet has not been | |
2448 | * acked. | |
2449 | */ | |
2450 | BUG_ON(!after(scb->end_seq, seq)); | |
2451 | ||
2452 | packets_acked = tcp_skb_pcount(skb); | |
2453 | if (tcp_trim_head(sk, skb, seq - scb->seq)) | |
2454 | return 0; | |
2455 | packets_acked -= tcp_skb_pcount(skb); | |
2456 | ||
2457 | if (packets_acked) { | |
2458 | __u8 sacked = scb->sacked; | |
2459 | ||
2460 | acked |= FLAG_DATA_ACKED; | |
2461 | if (sacked) { | |
2462 | if (sacked & TCPCB_RETRANS) { | |
2463 | if (sacked & TCPCB_SACKED_RETRANS) | |
2464 | tp->retrans_out -= packets_acked; | |
2465 | acked |= FLAG_RETRANS_DATA_ACKED; | |
2466 | *seq_rtt = -1; | |
2467 | } else if (*seq_rtt < 0) | |
2468 | *seq_rtt = now - scb->when; | |
2469 | if (sacked & TCPCB_SACKED_ACKED) | |
2470 | tp->sacked_out -= packets_acked; | |
2471 | if (sacked & TCPCB_LOST) | |
2472 | tp->lost_out -= packets_acked; | |
2473 | if (sacked & TCPCB_URG) { | |
2474 | if (tp->urg_mode && | |
2475 | !before(seq, tp->snd_up)) | |
2476 | tp->urg_mode = 0; | |
2477 | } | |
2478 | } else if (*seq_rtt < 0) | |
2479 | *seq_rtt = now - scb->when; | |
2480 | ||
2481 | if (tp->fackets_out) { | |
2482 | __u32 dval = min(tp->fackets_out, packets_acked); | |
2483 | tp->fackets_out -= dval; | |
2484 | } | |
48611c47 IJ |
2485 | /* hint's skb might be NULL but we don't need to care */ |
2486 | tp->fastpath_cnt_hint -= min_t(u32, packets_acked, | |
2487 | tp->fastpath_cnt_hint); | |
1da177e4 LT |
2488 | tp->packets_out -= packets_acked; |
2489 | ||
2490 | BUG_ON(tcp_skb_pcount(skb) == 0); | |
2491 | BUG_ON(!before(scb->seq, scb->end_seq)); | |
2492 | } | |
2493 | ||
2494 | return acked; | |
2495 | } | |
2496 | ||
1da177e4 | 2497 | /* Remove acknowledged frames from the retransmission queue. */ |
2d2abbab | 2498 | static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p) |
1da177e4 LT |
2499 | { |
2500 | struct tcp_sock *tp = tcp_sk(sk); | |
2d2abbab | 2501 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2502 | struct sk_buff *skb; |
2503 | __u32 now = tcp_time_stamp; | |
2504 | int acked = 0; | |
6418204f | 2505 | int prior_packets = tp->packets_out; |
1da177e4 | 2506 | __s32 seq_rtt = -1; |
b9ce204f | 2507 | ktime_t last_ackt = net_invalid_timestamp(); |
1da177e4 | 2508 | |
fe067e8a DM |
2509 | while ((skb = tcp_write_queue_head(sk)) && |
2510 | skb != tcp_send_head(sk)) { | |
e905a9ed | 2511 | struct tcp_skb_cb *scb = TCP_SKB_CB(skb); |
1da177e4 LT |
2512 | __u8 sacked = scb->sacked; |
2513 | ||
2514 | /* If our packet is before the ack sequence we can | |
2515 | * discard it as it's confirmed to have arrived at | |
2516 | * the other end. | |
2517 | */ | |
2518 | if (after(scb->end_seq, tp->snd_una)) { | |
cb83199a DM |
2519 | if (tcp_skb_pcount(skb) > 1 && |
2520 | after(tp->snd_una, scb->seq)) | |
1da177e4 LT |
2521 | acked |= tcp_tso_acked(sk, skb, |
2522 | now, &seq_rtt); | |
2523 | break; | |
2524 | } | |
2525 | ||
2526 | /* Initial outgoing SYN's get put onto the write_queue | |
2527 | * just like anything else we transmit. It is not | |
2528 | * true data, and if we misinform our callers that | |
2529 | * this ACK acks real data, we will erroneously exit | |
2530 | * connection startup slow start one packet too | |
2531 | * quickly. This is severely frowned upon behavior. | |
2532 | */ | |
2533 | if (!(scb->flags & TCPCB_FLAG_SYN)) { | |
2534 | acked |= FLAG_DATA_ACKED; | |
2535 | } else { | |
2536 | acked |= FLAG_SYN_ACKED; | |
2537 | tp->retrans_stamp = 0; | |
2538 | } | |
2539 | ||
5d424d5a JH |
2540 | /* MTU probing checks */ |
2541 | if (icsk->icsk_mtup.probe_size) { | |
0e7b1368 | 2542 | if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) { |
5d424d5a JH |
2543 | tcp_mtup_probe_success(sk, skb); |
2544 | } | |
2545 | } | |
2546 | ||
1da177e4 LT |
2547 | if (sacked) { |
2548 | if (sacked & TCPCB_RETRANS) { | |
2de979bd | 2549 | if (sacked & TCPCB_SACKED_RETRANS) |
1da177e4 LT |
2550 | tp->retrans_out -= tcp_skb_pcount(skb); |
2551 | acked |= FLAG_RETRANS_DATA_ACKED; | |
2552 | seq_rtt = -1; | |
2d2abbab | 2553 | } else if (seq_rtt < 0) { |
1da177e4 | 2554 | seq_rtt = now - scb->when; |
164891aa | 2555 | last_ackt = skb->tstamp; |
a61bbcf2 | 2556 | } |
1da177e4 LT |
2557 | if (sacked & TCPCB_SACKED_ACKED) |
2558 | tp->sacked_out -= tcp_skb_pcount(skb); | |
2559 | if (sacked & TCPCB_LOST) | |
2560 | tp->lost_out -= tcp_skb_pcount(skb); | |
2561 | if (sacked & TCPCB_URG) { | |
2562 | if (tp->urg_mode && | |
2563 | !before(scb->end_seq, tp->snd_up)) | |
2564 | tp->urg_mode = 0; | |
2565 | } | |
2d2abbab | 2566 | } else if (seq_rtt < 0) { |
1da177e4 | 2567 | seq_rtt = now - scb->when; |
164891aa | 2568 | last_ackt = skb->tstamp; |
2d2abbab | 2569 | } |
1da177e4 | 2570 | tcp_dec_pcount_approx(&tp->fackets_out, skb); |
e9144bd8 | 2571 | tp->packets_out -= tcp_skb_pcount(skb); |
fe067e8a | 2572 | tcp_unlink_write_queue(skb, sk); |
1da177e4 | 2573 | sk_stream_free_skb(sk, skb); |
6a438bbe | 2574 | clear_all_retrans_hints(tp); |
1da177e4 LT |
2575 | } |
2576 | ||
2577 | if (acked&FLAG_ACKED) { | |
6418204f | 2578 | u32 pkts_acked = prior_packets - tp->packets_out; |
164891aa SH |
2579 | const struct tcp_congestion_ops *ca_ops |
2580 | = inet_csk(sk)->icsk_ca_ops; | |
2581 | ||
2d2abbab | 2582 | tcp_ack_update_rtt(sk, acked, seq_rtt); |
6728e7dc | 2583 | tcp_rearm_rto(sk); |
317a76f9 | 2584 | |
e60402d0 | 2585 | if (tcp_is_reno(tp)) |
1b6d427b IJ |
2586 | tcp_remove_reno_sacks(sk, pkts_acked); |
2587 | ||
30cfd0ba SH |
2588 | if (ca_ops->pkts_acked) { |
2589 | s32 rtt_us = -1; | |
2590 | ||
2591 | /* Is the ACK triggering packet unambiguous? */ | |
2592 | if (!(acked & FLAG_RETRANS_DATA_ACKED)) { | |
2593 | /* High resolution needed and available? */ | |
2594 | if (ca_ops->flags & TCP_CONG_RTT_STAMP && | |
2595 | !ktime_equal(last_ackt, | |
2596 | net_invalid_timestamp())) | |
2597 | rtt_us = ktime_us_delta(ktime_get_real(), | |
2598 | last_ackt); | |
2599 | else if (seq_rtt > 0) | |
2600 | rtt_us = jiffies_to_usecs(seq_rtt); | |
2601 | } | |
b9ce204f | 2602 | |
30cfd0ba SH |
2603 | ca_ops->pkts_acked(sk, pkts_acked, rtt_us); |
2604 | } | |
1da177e4 LT |
2605 | } |
2606 | ||
2607 | #if FASTRETRANS_DEBUG > 0 | |
2608 | BUG_TRAP((int)tp->sacked_out >= 0); | |
2609 | BUG_TRAP((int)tp->lost_out >= 0); | |
2610 | BUG_TRAP((int)tp->retrans_out >= 0); | |
e60402d0 | 2611 | if (!tp->packets_out && tcp_is_sack(tp)) { |
6687e988 | 2612 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2613 | if (tp->lost_out) { |
2614 | printk(KERN_DEBUG "Leak l=%u %d\n", | |
6687e988 | 2615 | tp->lost_out, icsk->icsk_ca_state); |
1da177e4 LT |
2616 | tp->lost_out = 0; |
2617 | } | |
2618 | if (tp->sacked_out) { | |
2619 | printk(KERN_DEBUG "Leak s=%u %d\n", | |
6687e988 | 2620 | tp->sacked_out, icsk->icsk_ca_state); |
1da177e4 LT |
2621 | tp->sacked_out = 0; |
2622 | } | |
2623 | if (tp->retrans_out) { | |
2624 | printk(KERN_DEBUG "Leak r=%u %d\n", | |
6687e988 | 2625 | tp->retrans_out, icsk->icsk_ca_state); |
1da177e4 LT |
2626 | tp->retrans_out = 0; |
2627 | } | |
2628 | } | |
2629 | #endif | |
2630 | *seq_rtt_p = seq_rtt; | |
2631 | return acked; | |
2632 | } | |
2633 | ||
2634 | static void tcp_ack_probe(struct sock *sk) | |
2635 | { | |
463c84b9 ACM |
2636 | const struct tcp_sock *tp = tcp_sk(sk); |
2637 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1da177e4 LT |
2638 | |
2639 | /* Was it a usable window open? */ | |
2640 | ||
fe067e8a | 2641 | if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, |
1da177e4 | 2642 | tp->snd_una + tp->snd_wnd)) { |
463c84b9 ACM |
2643 | icsk->icsk_backoff = 0; |
2644 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); | |
1da177e4 LT |
2645 | /* Socket must be waked up by subsequent tcp_data_snd_check(). |
2646 | * This function is not for random using! | |
2647 | */ | |
2648 | } else { | |
463c84b9 | 2649 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
3f421baa ACM |
2650 | min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), |
2651 | TCP_RTO_MAX); | |
1da177e4 LT |
2652 | } |
2653 | } | |
2654 | ||
6687e988 | 2655 | static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag) |
1da177e4 LT |
2656 | { |
2657 | return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || | |
6687e988 | 2658 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open); |
1da177e4 LT |
2659 | } |
2660 | ||
6687e988 | 2661 | static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag) |
1da177e4 | 2662 | { |
6687e988 | 2663 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 2664 | return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && |
6687e988 | 2665 | !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR)); |
1da177e4 LT |
2666 | } |
2667 | ||
2668 | /* Check that window update is acceptable. | |
2669 | * The function assumes that snd_una<=ack<=snd_next. | |
2670 | */ | |
463c84b9 ACM |
2671 | static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack, |
2672 | const u32 ack_seq, const u32 nwin) | |
1da177e4 LT |
2673 | { |
2674 | return (after(ack, tp->snd_una) || | |
2675 | after(ack_seq, tp->snd_wl1) || | |
2676 | (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd)); | |
2677 | } | |
2678 | ||
2679 | /* Update our send window. | |
2680 | * | |
2681 | * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 | |
2682 | * and in FreeBSD. NetBSD's one is even worse.) is wrong. | |
2683 | */ | |
9e412ba7 IJ |
2684 | static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack, |
2685 | u32 ack_seq) | |
1da177e4 | 2686 | { |
9e412ba7 | 2687 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 2688 | int flag = 0; |
aa8223c7 | 2689 | u32 nwin = ntohs(tcp_hdr(skb)->window); |
1da177e4 | 2690 | |
aa8223c7 | 2691 | if (likely(!tcp_hdr(skb)->syn)) |
1da177e4 LT |
2692 | nwin <<= tp->rx_opt.snd_wscale; |
2693 | ||
2694 | if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { | |
2695 | flag |= FLAG_WIN_UPDATE; | |
2696 | tcp_update_wl(tp, ack, ack_seq); | |
2697 | ||
2698 | if (tp->snd_wnd != nwin) { | |
2699 | tp->snd_wnd = nwin; | |
2700 | ||
2701 | /* Note, it is the only place, where | |
2702 | * fast path is recovered for sending TCP. | |
2703 | */ | |
2ad41065 | 2704 | tp->pred_flags = 0; |
9e412ba7 | 2705 | tcp_fast_path_check(sk); |
1da177e4 LT |
2706 | |
2707 | if (nwin > tp->max_window) { | |
2708 | tp->max_window = nwin; | |
d83d8461 | 2709 | tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); |
1da177e4 LT |
2710 | } |
2711 | } | |
2712 | } | |
2713 | ||
2714 | tp->snd_una = ack; | |
2715 | ||
2716 | return flag; | |
2717 | } | |
2718 | ||
9ead9a1d IJ |
2719 | /* A very conservative spurious RTO response algorithm: reduce cwnd and |
2720 | * continue in congestion avoidance. | |
2721 | */ | |
2722 | static void tcp_conservative_spur_to_response(struct tcp_sock *tp) | |
2723 | { | |
2724 | tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); | |
aa8b6a7a | 2725 | tp->snd_cwnd_cnt = 0; |
46323655 | 2726 | TCP_ECN_queue_cwr(tp); |
9ead9a1d IJ |
2727 | tcp_moderate_cwnd(tp); |
2728 | } | |
2729 | ||
3cfe3baa IJ |
2730 | /* A conservative spurious RTO response algorithm: reduce cwnd using |
2731 | * rate halving and continue in congestion avoidance. | |
2732 | */ | |
2733 | static void tcp_ratehalving_spur_to_response(struct sock *sk) | |
2734 | { | |
3cfe3baa | 2735 | tcp_enter_cwr(sk, 0); |
3cfe3baa IJ |
2736 | } |
2737 | ||
e317f6f6 | 2738 | static void tcp_undo_spur_to_response(struct sock *sk, int flag) |
3cfe3baa | 2739 | { |
e317f6f6 IJ |
2740 | if (flag&FLAG_ECE) |
2741 | tcp_ratehalving_spur_to_response(sk); | |
2742 | else | |
2743 | tcp_undo_cwr(sk, 1); | |
3cfe3baa IJ |
2744 | } |
2745 | ||
30935cf4 IJ |
2746 | /* F-RTO spurious RTO detection algorithm (RFC4138) |
2747 | * | |
6408d206 IJ |
2748 | * F-RTO affects during two new ACKs following RTO (well, almost, see inline |
2749 | * comments). State (ACK number) is kept in frto_counter. When ACK advances | |
2750 | * window (but not to or beyond highest sequence sent before RTO): | |
30935cf4 IJ |
2751 | * On First ACK, send two new segments out. |
2752 | * On Second ACK, RTO was likely spurious. Do spurious response (response | |
2753 | * algorithm is not part of the F-RTO detection algorithm | |
2754 | * given in RFC4138 but can be selected separately). | |
2755 | * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss | |
d551e454 IJ |
2756 | * and TCP falls back to conventional RTO recovery. F-RTO allows overriding |
2757 | * of Nagle, this is done using frto_counter states 2 and 3, when a new data | |
2758 | * segment of any size sent during F-RTO, state 2 is upgraded to 3. | |
30935cf4 IJ |
2759 | * |
2760 | * Rationale: if the RTO was spurious, new ACKs should arrive from the | |
2761 | * original window even after we transmit two new data segments. | |
2762 | * | |
4dc2665e IJ |
2763 | * SACK version: |
2764 | * on first step, wait until first cumulative ACK arrives, then move to | |
2765 | * the second step. In second step, the next ACK decides. | |
2766 | * | |
30935cf4 IJ |
2767 | * F-RTO is implemented (mainly) in four functions: |
2768 | * - tcp_use_frto() is used to determine if TCP is can use F-RTO | |
2769 | * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is | |
2770 | * called when tcp_use_frto() showed green light | |
2771 | * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm | |
2772 | * - tcp_enter_frto_loss() is called if there is not enough evidence | |
2773 | * to prove that the RTO is indeed spurious. It transfers the control | |
2774 | * from F-RTO to the conventional RTO recovery | |
2775 | */ | |
2e605294 | 2776 | static int tcp_process_frto(struct sock *sk, int flag) |
1da177e4 LT |
2777 | { |
2778 | struct tcp_sock *tp = tcp_sk(sk); | |
e905a9ed | 2779 | |
005903bc | 2780 | tcp_verify_left_out(tp); |
e905a9ed | 2781 | |
7487c48c IJ |
2782 | /* Duplicate the behavior from Loss state (fastretrans_alert) */ |
2783 | if (flag&FLAG_DATA_ACKED) | |
2784 | inet_csk(sk)->icsk_retransmits = 0; | |
2785 | ||
95c4922b | 2786 | if (!before(tp->snd_una, tp->frto_highmark)) { |
d551e454 | 2787 | tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag); |
7c9a4a5b | 2788 | return 1; |
95c4922b IJ |
2789 | } |
2790 | ||
e60402d0 | 2791 | if (!IsSackFrto() || tcp_is_reno(tp)) { |
4dc2665e IJ |
2792 | /* RFC4138 shortcoming in step 2; should also have case c): |
2793 | * ACK isn't duplicate nor advances window, e.g., opposite dir | |
2794 | * data, winupdate | |
2795 | */ | |
2e605294 | 2796 | if (!(flag&FLAG_ANY_PROGRESS) && (flag&FLAG_NOT_DUP)) |
4dc2665e IJ |
2797 | return 1; |
2798 | ||
2799 | if (!(flag&FLAG_DATA_ACKED)) { | |
2800 | tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3), | |
2801 | flag); | |
2802 | return 1; | |
2803 | } | |
2804 | } else { | |
2805 | if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) { | |
2806 | /* Prevent sending of new data. */ | |
2807 | tp->snd_cwnd = min(tp->snd_cwnd, | |
2808 | tcp_packets_in_flight(tp)); | |
2809 | return 1; | |
2810 | } | |
6408d206 | 2811 | |
d551e454 | 2812 | if ((tp->frto_counter >= 2) && |
4dc2665e IJ |
2813 | (!(flag&FLAG_FORWARD_PROGRESS) || |
2814 | ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) { | |
2815 | /* RFC4138 shortcoming (see comment above) */ | |
2816 | if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP)) | |
2817 | return 1; | |
2818 | ||
2819 | tcp_enter_frto_loss(sk, 3, flag); | |
2820 | return 1; | |
2821 | } | |
1da177e4 LT |
2822 | } |
2823 | ||
2824 | if (tp->frto_counter == 1) { | |
575ee714 IJ |
2825 | /* Sending of the next skb must be allowed or no FRTO */ |
2826 | if (!tcp_send_head(sk) || | |
2827 | after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, | |
2828 | tp->snd_una + tp->snd_wnd)) { | |
d551e454 IJ |
2829 | tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), |
2830 | flag); | |
575ee714 IJ |
2831 | return 1; |
2832 | } | |
2833 | ||
1da177e4 | 2834 | tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; |
94d0ea77 | 2835 | tp->frto_counter = 2; |
7c9a4a5b | 2836 | return 1; |
d551e454 | 2837 | } else { |
3cfe3baa IJ |
2838 | switch (sysctl_tcp_frto_response) { |
2839 | case 2: | |
e317f6f6 | 2840 | tcp_undo_spur_to_response(sk, flag); |
3cfe3baa IJ |
2841 | break; |
2842 | case 1: | |
2843 | tcp_conservative_spur_to_response(tp); | |
2844 | break; | |
2845 | default: | |
2846 | tcp_ratehalving_spur_to_response(sk); | |
2847 | break; | |
3ff50b79 | 2848 | } |
94d0ea77 | 2849 | tp->frto_counter = 0; |
1da177e4 | 2850 | } |
7c9a4a5b | 2851 | return 0; |
1da177e4 LT |
2852 | } |
2853 | ||
1da177e4 LT |
2854 | /* This routine deals with incoming acks, but not outgoing ones. */ |
2855 | static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag) | |
2856 | { | |
6687e988 | 2857 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2858 | struct tcp_sock *tp = tcp_sk(sk); |
2859 | u32 prior_snd_una = tp->snd_una; | |
2860 | u32 ack_seq = TCP_SKB_CB(skb)->seq; | |
2861 | u32 ack = TCP_SKB_CB(skb)->ack_seq; | |
2862 | u32 prior_in_flight; | |
2863 | s32 seq_rtt; | |
2864 | int prior_packets; | |
7c9a4a5b | 2865 | int frto_cwnd = 0; |
1da177e4 LT |
2866 | |
2867 | /* If the ack is newer than sent or older than previous acks | |
2868 | * then we can probably ignore it. | |
2869 | */ | |
2870 | if (after(ack, tp->snd_nxt)) | |
2871 | goto uninteresting_ack; | |
2872 | ||
2873 | if (before(ack, prior_snd_una)) | |
2874 | goto old_ack; | |
2875 | ||
2e605294 IJ |
2876 | if (after(ack, prior_snd_una)) |
2877 | flag |= FLAG_SND_UNA_ADVANCED; | |
2878 | ||
3fdf3f0c DO |
2879 | if (sysctl_tcp_abc) { |
2880 | if (icsk->icsk_ca_state < TCP_CA_CWR) | |
2881 | tp->bytes_acked += ack - prior_snd_una; | |
2882 | else if (icsk->icsk_ca_state == TCP_CA_Loss) | |
2883 | /* we assume just one segment left network */ | |
2884 | tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache); | |
2885 | } | |
9772efb9 | 2886 | |
1da177e4 LT |
2887 | if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) { |
2888 | /* Window is constant, pure forward advance. | |
2889 | * No more checks are required. | |
2890 | * Note, we use the fact that SND.UNA>=SND.WL2. | |
2891 | */ | |
2892 | tcp_update_wl(tp, ack, ack_seq); | |
2893 | tp->snd_una = ack; | |
1da177e4 LT |
2894 | flag |= FLAG_WIN_UPDATE; |
2895 | ||
6687e988 | 2896 | tcp_ca_event(sk, CA_EVENT_FAST_ACK); |
317a76f9 | 2897 | |
1da177e4 LT |
2898 | NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS); |
2899 | } else { | |
2900 | if (ack_seq != TCP_SKB_CB(skb)->end_seq) | |
2901 | flag |= FLAG_DATA; | |
2902 | else | |
2903 | NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS); | |
2904 | ||
9e412ba7 | 2905 | flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); |
1da177e4 LT |
2906 | |
2907 | if (TCP_SKB_CB(skb)->sacked) | |
2908 | flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); | |
2909 | ||
aa8223c7 | 2910 | if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb))) |
1da177e4 LT |
2911 | flag |= FLAG_ECE; |
2912 | ||
6687e988 | 2913 | tcp_ca_event(sk, CA_EVENT_SLOW_ACK); |
1da177e4 LT |
2914 | } |
2915 | ||
2916 | /* We passed data and got it acked, remove any soft error | |
2917 | * log. Something worked... | |
2918 | */ | |
2919 | sk->sk_err_soft = 0; | |
2920 | tp->rcv_tstamp = tcp_time_stamp; | |
2921 | prior_packets = tp->packets_out; | |
2922 | if (!prior_packets) | |
2923 | goto no_queue; | |
2924 | ||
2925 | prior_in_flight = tcp_packets_in_flight(tp); | |
2926 | ||
2927 | /* See if we can take anything off of the retransmit queue. */ | |
2d2abbab | 2928 | flag |= tcp_clean_rtx_queue(sk, &seq_rtt); |
1da177e4 LT |
2929 | |
2930 | if (tp->frto_counter) | |
2e605294 | 2931 | frto_cwnd = tcp_process_frto(sk, flag); |
1da177e4 | 2932 | |
6687e988 | 2933 | if (tcp_ack_is_dubious(sk, flag)) { |
caa20d9a | 2934 | /* Advance CWND, if state allows this. */ |
7c9a4a5b IJ |
2935 | if ((flag & FLAG_DATA_ACKED) && !frto_cwnd && |
2936 | tcp_may_raise_cwnd(sk, flag)) | |
16751347 | 2937 | tcp_cong_avoid(sk, ack, prior_in_flight, 0); |
1b6d427b | 2938 | tcp_fastretrans_alert(sk, prior_packets - tp->packets_out, flag); |
1da177e4 | 2939 | } else { |
7c9a4a5b | 2940 | if ((flag & FLAG_DATA_ACKED) && !frto_cwnd) |
16751347 | 2941 | tcp_cong_avoid(sk, ack, prior_in_flight, 1); |
1da177e4 LT |
2942 | } |
2943 | ||
2944 | if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP)) | |
2945 | dst_confirm(sk->sk_dst_cache); | |
2946 | ||
2947 | return 1; | |
2948 | ||
2949 | no_queue: | |
6687e988 | 2950 | icsk->icsk_probes_out = 0; |
1da177e4 LT |
2951 | |
2952 | /* If this ack opens up a zero window, clear backoff. It was | |
2953 | * being used to time the probes, and is probably far higher than | |
2954 | * it needs to be for normal retransmission. | |
2955 | */ | |
fe067e8a | 2956 | if (tcp_send_head(sk)) |
1da177e4 LT |
2957 | tcp_ack_probe(sk); |
2958 | return 1; | |
2959 | ||
2960 | old_ack: | |
2961 | if (TCP_SKB_CB(skb)->sacked) | |
2962 | tcp_sacktag_write_queue(sk, skb, prior_snd_una); | |
2963 | ||
2964 | uninteresting_ack: | |
2965 | SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt); | |
2966 | return 0; | |
2967 | } | |
2968 | ||
2969 | ||
2970 | /* Look for tcp options. Normally only called on SYN and SYNACK packets. | |
2971 | * But, this can also be called on packets in the established flow when | |
2972 | * the fast version below fails. | |
2973 | */ | |
2974 | void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab) | |
2975 | { | |
2976 | unsigned char *ptr; | |
aa8223c7 | 2977 | struct tcphdr *th = tcp_hdr(skb); |
1da177e4 LT |
2978 | int length=(th->doff*4)-sizeof(struct tcphdr); |
2979 | ||
2980 | ptr = (unsigned char *)(th + 1); | |
2981 | opt_rx->saw_tstamp = 0; | |
2982 | ||
2de979bd | 2983 | while (length > 0) { |
e905a9ed | 2984 | int opcode=*ptr++; |
1da177e4 LT |
2985 | int opsize; |
2986 | ||
2987 | switch (opcode) { | |
2988 | case TCPOPT_EOL: | |
2989 | return; | |
2990 | case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ | |
2991 | length--; | |
2992 | continue; | |
2993 | default: | |
2994 | opsize=*ptr++; | |
2995 | if (opsize < 2) /* "silly options" */ | |
2996 | return; | |
2997 | if (opsize > length) | |
2998 | return; /* don't parse partial options */ | |
2de979bd | 2999 | switch (opcode) { |
1da177e4 | 3000 | case TCPOPT_MSS: |
2de979bd | 3001 | if (opsize==TCPOLEN_MSS && th->syn && !estab) { |
4f3608b7 | 3002 | u16 in_mss = ntohs(get_unaligned((__be16 *)ptr)); |
1da177e4 LT |
3003 | if (in_mss) { |
3004 | if (opt_rx->user_mss && opt_rx->user_mss < in_mss) | |
3005 | in_mss = opt_rx->user_mss; | |
3006 | opt_rx->mss_clamp = in_mss; | |
3007 | } | |
3008 | } | |
3009 | break; | |
3010 | case TCPOPT_WINDOW: | |
2de979bd | 3011 | if (opsize==TCPOLEN_WINDOW && th->syn && !estab) |
1da177e4 LT |
3012 | if (sysctl_tcp_window_scaling) { |
3013 | __u8 snd_wscale = *(__u8 *) ptr; | |
3014 | opt_rx->wscale_ok = 1; | |
3015 | if (snd_wscale > 14) { | |
2de979bd | 3016 | if (net_ratelimit()) |
1da177e4 LT |
3017 | printk(KERN_INFO "tcp_parse_options: Illegal window " |
3018 | "scaling value %d >14 received.\n", | |
3019 | snd_wscale); | |
3020 | snd_wscale = 14; | |
3021 | } | |
3022 | opt_rx->snd_wscale = snd_wscale; | |
3023 | } | |
3024 | break; | |
3025 | case TCPOPT_TIMESTAMP: | |
2de979bd | 3026 | if (opsize==TCPOLEN_TIMESTAMP) { |
1da177e4 LT |
3027 | if ((estab && opt_rx->tstamp_ok) || |
3028 | (!estab && sysctl_tcp_timestamps)) { | |
3029 | opt_rx->saw_tstamp = 1; | |
4f3608b7 AV |
3030 | opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr)); |
3031 | opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4))); | |
1da177e4 LT |
3032 | } |
3033 | } | |
3034 | break; | |
3035 | case TCPOPT_SACK_PERM: | |
2de979bd | 3036 | if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) { |
1da177e4 LT |
3037 | if (sysctl_tcp_sack) { |
3038 | opt_rx->sack_ok = 1; | |
3039 | tcp_sack_reset(opt_rx); | |
3040 | } | |
3041 | } | |
3042 | break; | |
3043 | ||
3044 | case TCPOPT_SACK: | |
2de979bd | 3045 | if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && |
1da177e4 LT |
3046 | !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && |
3047 | opt_rx->sack_ok) { | |
3048 | TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; | |
3049 | } | |
d7ea5b91 | 3050 | break; |
cfb6eeb4 YH |
3051 | #ifdef CONFIG_TCP_MD5SIG |
3052 | case TCPOPT_MD5SIG: | |
3053 | /* | |
3054 | * The MD5 Hash has already been | |
3055 | * checked (see tcp_v{4,6}_do_rcv()). | |
3056 | */ | |
3057 | break; | |
3058 | #endif | |
3ff50b79 SH |
3059 | } |
3060 | ||
e905a9ed YH |
3061 | ptr+=opsize-2; |
3062 | length-=opsize; | |
3ff50b79 | 3063 | } |
1da177e4 LT |
3064 | } |
3065 | } | |
3066 | ||
3067 | /* Fast parse options. This hopes to only see timestamps. | |
3068 | * If it is wrong it falls back on tcp_parse_options(). | |
3069 | */ | |
40efc6fa SH |
3070 | static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, |
3071 | struct tcp_sock *tp) | |
1da177e4 LT |
3072 | { |
3073 | if (th->doff == sizeof(struct tcphdr)>>2) { | |
3074 | tp->rx_opt.saw_tstamp = 0; | |
3075 | return 0; | |
3076 | } else if (tp->rx_opt.tstamp_ok && | |
3077 | th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) { | |
4f3608b7 AV |
3078 | __be32 *ptr = (__be32 *)(th + 1); |
3079 | if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | |
1da177e4 LT |
3080 | | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { |
3081 | tp->rx_opt.saw_tstamp = 1; | |
3082 | ++ptr; | |
3083 | tp->rx_opt.rcv_tsval = ntohl(*ptr); | |
3084 | ++ptr; | |
3085 | tp->rx_opt.rcv_tsecr = ntohl(*ptr); | |
3086 | return 1; | |
3087 | } | |
3088 | } | |
3089 | tcp_parse_options(skb, &tp->rx_opt, 1); | |
3090 | return 1; | |
3091 | } | |
3092 | ||
3093 | static inline void tcp_store_ts_recent(struct tcp_sock *tp) | |
3094 | { | |
3095 | tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; | |
9d729f72 | 3096 | tp->rx_opt.ts_recent_stamp = get_seconds(); |
1da177e4 LT |
3097 | } |
3098 | ||
3099 | static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) | |
3100 | { | |
3101 | if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { | |
3102 | /* PAWS bug workaround wrt. ACK frames, the PAWS discard | |
3103 | * extra check below makes sure this can only happen | |
3104 | * for pure ACK frames. -DaveM | |
3105 | * | |
3106 | * Not only, also it occurs for expired timestamps. | |
3107 | */ | |
3108 | ||
2de979bd | 3109 | if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 || |
9d729f72 | 3110 | get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS) |
1da177e4 LT |
3111 | tcp_store_ts_recent(tp); |
3112 | } | |
3113 | } | |
3114 | ||
3115 | /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM | |
3116 | * | |
3117 | * It is not fatal. If this ACK does _not_ change critical state (seqs, window) | |
3118 | * it can pass through stack. So, the following predicate verifies that | |
3119 | * this segment is not used for anything but congestion avoidance or | |
3120 | * fast retransmit. Moreover, we even are able to eliminate most of such | |
3121 | * second order effects, if we apply some small "replay" window (~RTO) | |
3122 | * to timestamp space. | |
3123 | * | |
3124 | * All these measures still do not guarantee that we reject wrapped ACKs | |
3125 | * on networks with high bandwidth, when sequence space is recycled fastly, | |
3126 | * but it guarantees that such events will be very rare and do not affect | |
3127 | * connection seriously. This doesn't look nice, but alas, PAWS is really | |
3128 | * buggy extension. | |
3129 | * | |
3130 | * [ Later note. Even worse! It is buggy for segments _with_ data. RFC | |
3131 | * states that events when retransmit arrives after original data are rare. | |
3132 | * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is | |
3133 | * the biggest problem on large power networks even with minor reordering. | |
3134 | * OK, let's give it small replay window. If peer clock is even 1hz, it is safe | |
3135 | * up to bandwidth of 18Gigabit/sec. 8) ] | |
3136 | */ | |
3137 | ||
463c84b9 | 3138 | static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 3139 | { |
463c84b9 | 3140 | struct tcp_sock *tp = tcp_sk(sk); |
aa8223c7 | 3141 | struct tcphdr *th = tcp_hdr(skb); |
1da177e4 LT |
3142 | u32 seq = TCP_SKB_CB(skb)->seq; |
3143 | u32 ack = TCP_SKB_CB(skb)->ack_seq; | |
3144 | ||
3145 | return (/* 1. Pure ACK with correct sequence number. */ | |
3146 | (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && | |
3147 | ||
3148 | /* 2. ... and duplicate ACK. */ | |
3149 | ack == tp->snd_una && | |
3150 | ||
3151 | /* 3. ... and does not update window. */ | |
3152 | !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && | |
3153 | ||
3154 | /* 4. ... and sits in replay window. */ | |
463c84b9 | 3155 | (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); |
1da177e4 LT |
3156 | } |
3157 | ||
463c84b9 | 3158 | static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 3159 | { |
463c84b9 | 3160 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 3161 | return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW && |
9d729f72 | 3162 | get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS && |
463c84b9 | 3163 | !tcp_disordered_ack(sk, skb)); |
1da177e4 LT |
3164 | } |
3165 | ||
3166 | /* Check segment sequence number for validity. | |
3167 | * | |
3168 | * Segment controls are considered valid, if the segment | |
3169 | * fits to the window after truncation to the window. Acceptability | |
3170 | * of data (and SYN, FIN, of course) is checked separately. | |
3171 | * See tcp_data_queue(), for example. | |
3172 | * | |
3173 | * Also, controls (RST is main one) are accepted using RCV.WUP instead | |
3174 | * of RCV.NXT. Peer still did not advance his SND.UNA when we | |
3175 | * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. | |
3176 | * (borrowed from freebsd) | |
3177 | */ | |
3178 | ||
3179 | static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq) | |
3180 | { | |
3181 | return !before(end_seq, tp->rcv_wup) && | |
3182 | !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); | |
3183 | } | |
3184 | ||
3185 | /* When we get a reset we do this. */ | |
3186 | static void tcp_reset(struct sock *sk) | |
3187 | { | |
3188 | /* We want the right error as BSD sees it (and indeed as we do). */ | |
3189 | switch (sk->sk_state) { | |
3190 | case TCP_SYN_SENT: | |
3191 | sk->sk_err = ECONNREFUSED; | |
3192 | break; | |
3193 | case TCP_CLOSE_WAIT: | |
3194 | sk->sk_err = EPIPE; | |
3195 | break; | |
3196 | case TCP_CLOSE: | |
3197 | return; | |
3198 | default: | |
3199 | sk->sk_err = ECONNRESET; | |
3200 | } | |
3201 | ||
3202 | if (!sock_flag(sk, SOCK_DEAD)) | |
3203 | sk->sk_error_report(sk); | |
3204 | ||
3205 | tcp_done(sk); | |
3206 | } | |
3207 | ||
3208 | /* | |
3209 | * Process the FIN bit. This now behaves as it is supposed to work | |
3210 | * and the FIN takes effect when it is validly part of sequence | |
3211 | * space. Not before when we get holes. | |
3212 | * | |
3213 | * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT | |
3214 | * (and thence onto LAST-ACK and finally, CLOSE, we never enter | |
3215 | * TIME-WAIT) | |
3216 | * | |
3217 | * If we are in FINWAIT-1, a received FIN indicates simultaneous | |
3218 | * close and we go into CLOSING (and later onto TIME-WAIT) | |
3219 | * | |
3220 | * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. | |
3221 | */ | |
3222 | static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th) | |
3223 | { | |
3224 | struct tcp_sock *tp = tcp_sk(sk); | |
3225 | ||
463c84b9 | 3226 | inet_csk_schedule_ack(sk); |
1da177e4 LT |
3227 | |
3228 | sk->sk_shutdown |= RCV_SHUTDOWN; | |
3229 | sock_set_flag(sk, SOCK_DONE); | |
3230 | ||
3231 | switch (sk->sk_state) { | |
3232 | case TCP_SYN_RECV: | |
3233 | case TCP_ESTABLISHED: | |
3234 | /* Move to CLOSE_WAIT */ | |
3235 | tcp_set_state(sk, TCP_CLOSE_WAIT); | |
463c84b9 | 3236 | inet_csk(sk)->icsk_ack.pingpong = 1; |
1da177e4 LT |
3237 | break; |
3238 | ||
3239 | case TCP_CLOSE_WAIT: | |
3240 | case TCP_CLOSING: | |
3241 | /* Received a retransmission of the FIN, do | |
3242 | * nothing. | |
3243 | */ | |
3244 | break; | |
3245 | case TCP_LAST_ACK: | |
3246 | /* RFC793: Remain in the LAST-ACK state. */ | |
3247 | break; | |
3248 | ||
3249 | case TCP_FIN_WAIT1: | |
3250 | /* This case occurs when a simultaneous close | |
3251 | * happens, we must ack the received FIN and | |
3252 | * enter the CLOSING state. | |
3253 | */ | |
3254 | tcp_send_ack(sk); | |
3255 | tcp_set_state(sk, TCP_CLOSING); | |
3256 | break; | |
3257 | case TCP_FIN_WAIT2: | |
3258 | /* Received a FIN -- send ACK and enter TIME_WAIT. */ | |
3259 | tcp_send_ack(sk); | |
3260 | tcp_time_wait(sk, TCP_TIME_WAIT, 0); | |
3261 | break; | |
3262 | default: | |
3263 | /* Only TCP_LISTEN and TCP_CLOSE are left, in these | |
3264 | * cases we should never reach this piece of code. | |
3265 | */ | |
3266 | printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n", | |
3267 | __FUNCTION__, sk->sk_state); | |
3268 | break; | |
3ff50b79 | 3269 | } |
1da177e4 LT |
3270 | |
3271 | /* It _is_ possible, that we have something out-of-order _after_ FIN. | |
3272 | * Probably, we should reset in this case. For now drop them. | |
3273 | */ | |
3274 | __skb_queue_purge(&tp->out_of_order_queue); | |
e60402d0 | 3275 | if (tcp_is_sack(tp)) |
1da177e4 LT |
3276 | tcp_sack_reset(&tp->rx_opt); |
3277 | sk_stream_mem_reclaim(sk); | |
3278 | ||
3279 | if (!sock_flag(sk, SOCK_DEAD)) { | |
3280 | sk->sk_state_change(sk); | |
3281 | ||
3282 | /* Do not send POLL_HUP for half duplex close. */ | |
3283 | if (sk->sk_shutdown == SHUTDOWN_MASK || | |
3284 | sk->sk_state == TCP_CLOSE) | |
3285 | sk_wake_async(sk, 1, POLL_HUP); | |
3286 | else | |
3287 | sk_wake_async(sk, 1, POLL_IN); | |
3288 | } | |
3289 | } | |
3290 | ||
40efc6fa | 3291 | static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq) |
1da177e4 LT |
3292 | { |
3293 | if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { | |
3294 | if (before(seq, sp->start_seq)) | |
3295 | sp->start_seq = seq; | |
3296 | if (after(end_seq, sp->end_seq)) | |
3297 | sp->end_seq = end_seq; | |
3298 | return 1; | |
3299 | } | |
3300 | return 0; | |
3301 | } | |
3302 | ||
40efc6fa | 3303 | static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq) |
1da177e4 | 3304 | { |
e60402d0 | 3305 | if (tcp_is_sack(tp) && sysctl_tcp_dsack) { |
1da177e4 LT |
3306 | if (before(seq, tp->rcv_nxt)) |
3307 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT); | |
3308 | else | |
3309 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT); | |
3310 | ||
3311 | tp->rx_opt.dsack = 1; | |
3312 | tp->duplicate_sack[0].start_seq = seq; | |
3313 | tp->duplicate_sack[0].end_seq = end_seq; | |
3314 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok); | |
3315 | } | |
3316 | } | |
3317 | ||
40efc6fa | 3318 | static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq) |
1da177e4 LT |
3319 | { |
3320 | if (!tp->rx_opt.dsack) | |
3321 | tcp_dsack_set(tp, seq, end_seq); | |
3322 | else | |
3323 | tcp_sack_extend(tp->duplicate_sack, seq, end_seq); | |
3324 | } | |
3325 | ||
3326 | static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb) | |
3327 | { | |
3328 | struct tcp_sock *tp = tcp_sk(sk); | |
3329 | ||
3330 | if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && | |
3331 | before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
3332 | NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); | |
463c84b9 | 3333 | tcp_enter_quickack_mode(sk); |
1da177e4 | 3334 | |
e60402d0 | 3335 | if (tcp_is_sack(tp) && sysctl_tcp_dsack) { |
1da177e4 LT |
3336 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; |
3337 | ||
3338 | if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) | |
3339 | end_seq = tp->rcv_nxt; | |
3340 | tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq); | |
3341 | } | |
3342 | } | |
3343 | ||
3344 | tcp_send_ack(sk); | |
3345 | } | |
3346 | ||
3347 | /* These routines update the SACK block as out-of-order packets arrive or | |
3348 | * in-order packets close up the sequence space. | |
3349 | */ | |
3350 | static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) | |
3351 | { | |
3352 | int this_sack; | |
3353 | struct tcp_sack_block *sp = &tp->selective_acks[0]; | |
3354 | struct tcp_sack_block *swalk = sp+1; | |
3355 | ||
3356 | /* See if the recent change to the first SACK eats into | |
3357 | * or hits the sequence space of other SACK blocks, if so coalesce. | |
3358 | */ | |
3359 | for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) { | |
3360 | if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { | |
3361 | int i; | |
3362 | ||
3363 | /* Zap SWALK, by moving every further SACK up by one slot. | |
3364 | * Decrease num_sacks. | |
3365 | */ | |
3366 | tp->rx_opt.num_sacks--; | |
3367 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); | |
2de979bd | 3368 | for (i=this_sack; i < tp->rx_opt.num_sacks; i++) |
1da177e4 LT |
3369 | sp[i] = sp[i+1]; |
3370 | continue; | |
3371 | } | |
3372 | this_sack++, swalk++; | |
3373 | } | |
3374 | } | |
3375 | ||
40efc6fa | 3376 | static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2) |
1da177e4 LT |
3377 | { |
3378 | __u32 tmp; | |
3379 | ||
3380 | tmp = sack1->start_seq; | |
3381 | sack1->start_seq = sack2->start_seq; | |
3382 | sack2->start_seq = tmp; | |
3383 | ||
3384 | tmp = sack1->end_seq; | |
3385 | sack1->end_seq = sack2->end_seq; | |
3386 | sack2->end_seq = tmp; | |
3387 | } | |
3388 | ||
3389 | static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) | |
3390 | { | |
3391 | struct tcp_sock *tp = tcp_sk(sk); | |
3392 | struct tcp_sack_block *sp = &tp->selective_acks[0]; | |
3393 | int cur_sacks = tp->rx_opt.num_sacks; | |
3394 | int this_sack; | |
3395 | ||
3396 | if (!cur_sacks) | |
3397 | goto new_sack; | |
3398 | ||
3399 | for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) { | |
3400 | if (tcp_sack_extend(sp, seq, end_seq)) { | |
3401 | /* Rotate this_sack to the first one. */ | |
3402 | for (; this_sack>0; this_sack--, sp--) | |
3403 | tcp_sack_swap(sp, sp-1); | |
3404 | if (cur_sacks > 1) | |
3405 | tcp_sack_maybe_coalesce(tp); | |
3406 | return; | |
3407 | } | |
3408 | } | |
3409 | ||
3410 | /* Could not find an adjacent existing SACK, build a new one, | |
3411 | * put it at the front, and shift everyone else down. We | |
3412 | * always know there is at least one SACK present already here. | |
3413 | * | |
3414 | * If the sack array is full, forget about the last one. | |
3415 | */ | |
3416 | if (this_sack >= 4) { | |
3417 | this_sack--; | |
3418 | tp->rx_opt.num_sacks--; | |
3419 | sp--; | |
3420 | } | |
2de979bd | 3421 | for (; this_sack > 0; this_sack--, sp--) |
1da177e4 LT |
3422 | *sp = *(sp-1); |
3423 | ||
3424 | new_sack: | |
3425 | /* Build the new head SACK, and we're done. */ | |
3426 | sp->start_seq = seq; | |
3427 | sp->end_seq = end_seq; | |
3428 | tp->rx_opt.num_sacks++; | |
3429 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); | |
3430 | } | |
3431 | ||
3432 | /* RCV.NXT advances, some SACKs should be eaten. */ | |
3433 | ||
3434 | static void tcp_sack_remove(struct tcp_sock *tp) | |
3435 | { | |
3436 | struct tcp_sack_block *sp = &tp->selective_acks[0]; | |
3437 | int num_sacks = tp->rx_opt.num_sacks; | |
3438 | int this_sack; | |
3439 | ||
3440 | /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ | |
b03efcfb | 3441 | if (skb_queue_empty(&tp->out_of_order_queue)) { |
1da177e4 LT |
3442 | tp->rx_opt.num_sacks = 0; |
3443 | tp->rx_opt.eff_sacks = tp->rx_opt.dsack; | |
3444 | return; | |
3445 | } | |
3446 | ||
2de979bd | 3447 | for (this_sack = 0; this_sack < num_sacks; ) { |
1da177e4 LT |
3448 | /* Check if the start of the sack is covered by RCV.NXT. */ |
3449 | if (!before(tp->rcv_nxt, sp->start_seq)) { | |
3450 | int i; | |
3451 | ||
3452 | /* RCV.NXT must cover all the block! */ | |
3453 | BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq)); | |
3454 | ||
3455 | /* Zap this SACK, by moving forward any other SACKS. */ | |
3456 | for (i=this_sack+1; i < num_sacks; i++) | |
3457 | tp->selective_acks[i-1] = tp->selective_acks[i]; | |
3458 | num_sacks--; | |
3459 | continue; | |
3460 | } | |
3461 | this_sack++; | |
3462 | sp++; | |
3463 | } | |
3464 | if (num_sacks != tp->rx_opt.num_sacks) { | |
3465 | tp->rx_opt.num_sacks = num_sacks; | |
3466 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); | |
3467 | } | |
3468 | } | |
3469 | ||
3470 | /* This one checks to see if we can put data from the | |
3471 | * out_of_order queue into the receive_queue. | |
3472 | */ | |
3473 | static void tcp_ofo_queue(struct sock *sk) | |
3474 | { | |
3475 | struct tcp_sock *tp = tcp_sk(sk); | |
3476 | __u32 dsack_high = tp->rcv_nxt; | |
3477 | struct sk_buff *skb; | |
3478 | ||
3479 | while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { | |
3480 | if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) | |
3481 | break; | |
3482 | ||
3483 | if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { | |
3484 | __u32 dsack = dsack_high; | |
3485 | if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) | |
3486 | dsack_high = TCP_SKB_CB(skb)->end_seq; | |
3487 | tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack); | |
3488 | } | |
3489 | ||
3490 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { | |
3491 | SOCK_DEBUG(sk, "ofo packet was already received \n"); | |
8728b834 | 3492 | __skb_unlink(skb, &tp->out_of_order_queue); |
1da177e4 LT |
3493 | __kfree_skb(skb); |
3494 | continue; | |
3495 | } | |
3496 | SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", | |
3497 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, | |
3498 | TCP_SKB_CB(skb)->end_seq); | |
3499 | ||
8728b834 | 3500 | __skb_unlink(skb, &tp->out_of_order_queue); |
1da177e4 LT |
3501 | __skb_queue_tail(&sk->sk_receive_queue, skb); |
3502 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
aa8223c7 ACM |
3503 | if (tcp_hdr(skb)->fin) |
3504 | tcp_fin(skb, sk, tcp_hdr(skb)); | |
1da177e4 LT |
3505 | } |
3506 | } | |
3507 | ||
3508 | static int tcp_prune_queue(struct sock *sk); | |
3509 | ||
3510 | static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) | |
3511 | { | |
aa8223c7 | 3512 | struct tcphdr *th = tcp_hdr(skb); |
1da177e4 LT |
3513 | struct tcp_sock *tp = tcp_sk(sk); |
3514 | int eaten = -1; | |
3515 | ||
3516 | if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) | |
3517 | goto drop; | |
3518 | ||
1da177e4 LT |
3519 | __skb_pull(skb, th->doff*4); |
3520 | ||
3521 | TCP_ECN_accept_cwr(tp, skb); | |
3522 | ||
3523 | if (tp->rx_opt.dsack) { | |
3524 | tp->rx_opt.dsack = 0; | |
3525 | tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks, | |
3526 | 4 - tp->rx_opt.tstamp_ok); | |
3527 | } | |
3528 | ||
3529 | /* Queue data for delivery to the user. | |
3530 | * Packets in sequence go to the receive queue. | |
3531 | * Out of sequence packets to the out_of_order_queue. | |
3532 | */ | |
3533 | if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { | |
3534 | if (tcp_receive_window(tp) == 0) | |
3535 | goto out_of_window; | |
3536 | ||
3537 | /* Ok. In sequence. In window. */ | |
3538 | if (tp->ucopy.task == current && | |
3539 | tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && | |
3540 | sock_owned_by_user(sk) && !tp->urg_data) { | |
3541 | int chunk = min_t(unsigned int, skb->len, | |
3542 | tp->ucopy.len); | |
3543 | ||
3544 | __set_current_state(TASK_RUNNING); | |
3545 | ||
3546 | local_bh_enable(); | |
3547 | if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { | |
3548 | tp->ucopy.len -= chunk; | |
3549 | tp->copied_seq += chunk; | |
3550 | eaten = (chunk == skb->len && !th->fin); | |
3551 | tcp_rcv_space_adjust(sk); | |
3552 | } | |
3553 | local_bh_disable(); | |
3554 | } | |
3555 | ||
3556 | if (eaten <= 0) { | |
3557 | queue_and_out: | |
3558 | if (eaten < 0 && | |
3559 | (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || | |
3560 | !sk_stream_rmem_schedule(sk, skb))) { | |
3561 | if (tcp_prune_queue(sk) < 0 || | |
3562 | !sk_stream_rmem_schedule(sk, skb)) | |
3563 | goto drop; | |
3564 | } | |
3565 | sk_stream_set_owner_r(skb, sk); | |
3566 | __skb_queue_tail(&sk->sk_receive_queue, skb); | |
3567 | } | |
3568 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
2de979bd | 3569 | if (skb->len) |
9e412ba7 | 3570 | tcp_event_data_recv(sk, skb); |
2de979bd | 3571 | if (th->fin) |
1da177e4 LT |
3572 | tcp_fin(skb, sk, th); |
3573 | ||
b03efcfb | 3574 | if (!skb_queue_empty(&tp->out_of_order_queue)) { |
1da177e4 LT |
3575 | tcp_ofo_queue(sk); |
3576 | ||
3577 | /* RFC2581. 4.2. SHOULD send immediate ACK, when | |
3578 | * gap in queue is filled. | |
3579 | */ | |
b03efcfb | 3580 | if (skb_queue_empty(&tp->out_of_order_queue)) |
463c84b9 | 3581 | inet_csk(sk)->icsk_ack.pingpong = 0; |
1da177e4 LT |
3582 | } |
3583 | ||
3584 | if (tp->rx_opt.num_sacks) | |
3585 | tcp_sack_remove(tp); | |
3586 | ||
9e412ba7 | 3587 | tcp_fast_path_check(sk); |
1da177e4 LT |
3588 | |
3589 | if (eaten > 0) | |
3590 | __kfree_skb(skb); | |
3591 | else if (!sock_flag(sk, SOCK_DEAD)) | |
3592 | sk->sk_data_ready(sk, 0); | |
3593 | return; | |
3594 | } | |
3595 | ||
3596 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { | |
3597 | /* A retransmit, 2nd most common case. Force an immediate ack. */ | |
3598 | NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); | |
3599 | tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); | |
3600 | ||
3601 | out_of_window: | |
463c84b9 ACM |
3602 | tcp_enter_quickack_mode(sk); |
3603 | inet_csk_schedule_ack(sk); | |
1da177e4 LT |
3604 | drop: |
3605 | __kfree_skb(skb); | |
3606 | return; | |
3607 | } | |
3608 | ||
3609 | /* Out of window. F.e. zero window probe. */ | |
3610 | if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) | |
3611 | goto out_of_window; | |
3612 | ||
463c84b9 | 3613 | tcp_enter_quickack_mode(sk); |
1da177e4 LT |
3614 | |
3615 | if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
3616 | /* Partial packet, seq < rcv_next < end_seq */ | |
3617 | SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", | |
3618 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, | |
3619 | TCP_SKB_CB(skb)->end_seq); | |
3620 | ||
3621 | tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); | |
e905a9ed | 3622 | |
1da177e4 LT |
3623 | /* If window is closed, drop tail of packet. But after |
3624 | * remembering D-SACK for its head made in previous line. | |
3625 | */ | |
3626 | if (!tcp_receive_window(tp)) | |
3627 | goto out_of_window; | |
3628 | goto queue_and_out; | |
3629 | } | |
3630 | ||
3631 | TCP_ECN_check_ce(tp, skb); | |
3632 | ||
3633 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || | |
3634 | !sk_stream_rmem_schedule(sk, skb)) { | |
3635 | if (tcp_prune_queue(sk) < 0 || | |
3636 | !sk_stream_rmem_schedule(sk, skb)) | |
3637 | goto drop; | |
3638 | } | |
3639 | ||
3640 | /* Disable header prediction. */ | |
3641 | tp->pred_flags = 0; | |
463c84b9 | 3642 | inet_csk_schedule_ack(sk); |
1da177e4 LT |
3643 | |
3644 | SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", | |
3645 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); | |
3646 | ||
3647 | sk_stream_set_owner_r(skb, sk); | |
3648 | ||
3649 | if (!skb_peek(&tp->out_of_order_queue)) { | |
3650 | /* Initial out of order segment, build 1 SACK. */ | |
e60402d0 | 3651 | if (tcp_is_sack(tp)) { |
1da177e4 LT |
3652 | tp->rx_opt.num_sacks = 1; |
3653 | tp->rx_opt.dsack = 0; | |
3654 | tp->rx_opt.eff_sacks = 1; | |
3655 | tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; | |
3656 | tp->selective_acks[0].end_seq = | |
3657 | TCP_SKB_CB(skb)->end_seq; | |
3658 | } | |
3659 | __skb_queue_head(&tp->out_of_order_queue,skb); | |
3660 | } else { | |
3661 | struct sk_buff *skb1 = tp->out_of_order_queue.prev; | |
3662 | u32 seq = TCP_SKB_CB(skb)->seq; | |
3663 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
3664 | ||
3665 | if (seq == TCP_SKB_CB(skb1)->end_seq) { | |
8728b834 | 3666 | __skb_append(skb1, skb, &tp->out_of_order_queue); |
1da177e4 LT |
3667 | |
3668 | if (!tp->rx_opt.num_sacks || | |
3669 | tp->selective_acks[0].end_seq != seq) | |
3670 | goto add_sack; | |
3671 | ||
3672 | /* Common case: data arrive in order after hole. */ | |
3673 | tp->selective_acks[0].end_seq = end_seq; | |
3674 | return; | |
3675 | } | |
3676 | ||
3677 | /* Find place to insert this segment. */ | |
3678 | do { | |
3679 | if (!after(TCP_SKB_CB(skb1)->seq, seq)) | |
3680 | break; | |
3681 | } while ((skb1 = skb1->prev) != | |
3682 | (struct sk_buff*)&tp->out_of_order_queue); | |
3683 | ||
3684 | /* Do skb overlap to previous one? */ | |
3685 | if (skb1 != (struct sk_buff*)&tp->out_of_order_queue && | |
3686 | before(seq, TCP_SKB_CB(skb1)->end_seq)) { | |
3687 | if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { | |
3688 | /* All the bits are present. Drop. */ | |
3689 | __kfree_skb(skb); | |
3690 | tcp_dsack_set(tp, seq, end_seq); | |
3691 | goto add_sack; | |
3692 | } | |
3693 | if (after(seq, TCP_SKB_CB(skb1)->seq)) { | |
3694 | /* Partial overlap. */ | |
3695 | tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq); | |
3696 | } else { | |
3697 | skb1 = skb1->prev; | |
3698 | } | |
3699 | } | |
3700 | __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue); | |
e905a9ed | 3701 | |
1da177e4 LT |
3702 | /* And clean segments covered by new one as whole. */ |
3703 | while ((skb1 = skb->next) != | |
3704 | (struct sk_buff*)&tp->out_of_order_queue && | |
3705 | after(end_seq, TCP_SKB_CB(skb1)->seq)) { | |
3706 | if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { | |
3707 | tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq); | |
3708 | break; | |
3709 | } | |
8728b834 | 3710 | __skb_unlink(skb1, &tp->out_of_order_queue); |
1da177e4 LT |
3711 | tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq); |
3712 | __kfree_skb(skb1); | |
3713 | } | |
3714 | ||
3715 | add_sack: | |
e60402d0 | 3716 | if (tcp_is_sack(tp)) |
1da177e4 LT |
3717 | tcp_sack_new_ofo_skb(sk, seq, end_seq); |
3718 | } | |
3719 | } | |
3720 | ||
3721 | /* Collapse contiguous sequence of skbs head..tail with | |
3722 | * sequence numbers start..end. | |
3723 | * Segments with FIN/SYN are not collapsed (only because this | |
3724 | * simplifies code) | |
3725 | */ | |
3726 | static void | |
8728b834 DM |
3727 | tcp_collapse(struct sock *sk, struct sk_buff_head *list, |
3728 | struct sk_buff *head, struct sk_buff *tail, | |
3729 | u32 start, u32 end) | |
1da177e4 LT |
3730 | { |
3731 | struct sk_buff *skb; | |
3732 | ||
caa20d9a | 3733 | /* First, check that queue is collapsible and find |
1da177e4 LT |
3734 | * the point where collapsing can be useful. */ |
3735 | for (skb = head; skb != tail; ) { | |
3736 | /* No new bits? It is possible on ofo queue. */ | |
3737 | if (!before(start, TCP_SKB_CB(skb)->end_seq)) { | |
3738 | struct sk_buff *next = skb->next; | |
8728b834 | 3739 | __skb_unlink(skb, list); |
1da177e4 LT |
3740 | __kfree_skb(skb); |
3741 | NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); | |
3742 | skb = next; | |
3743 | continue; | |
3744 | } | |
3745 | ||
3746 | /* The first skb to collapse is: | |
3747 | * - not SYN/FIN and | |
3748 | * - bloated or contains data before "start" or | |
3749 | * overlaps to the next one. | |
3750 | */ | |
aa8223c7 | 3751 | if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin && |
1da177e4 LT |
3752 | (tcp_win_from_space(skb->truesize) > skb->len || |
3753 | before(TCP_SKB_CB(skb)->seq, start) || | |
3754 | (skb->next != tail && | |
3755 | TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq))) | |
3756 | break; | |
3757 | ||
3758 | /* Decided to skip this, advance start seq. */ | |
3759 | start = TCP_SKB_CB(skb)->end_seq; | |
3760 | skb = skb->next; | |
3761 | } | |
aa8223c7 | 3762 | if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin) |
1da177e4 LT |
3763 | return; |
3764 | ||
3765 | while (before(start, end)) { | |
3766 | struct sk_buff *nskb; | |
3767 | int header = skb_headroom(skb); | |
3768 | int copy = SKB_MAX_ORDER(header, 0); | |
3769 | ||
3770 | /* Too big header? This can happen with IPv6. */ | |
3771 | if (copy < 0) | |
3772 | return; | |
3773 | if (end-start < copy) | |
3774 | copy = end-start; | |
3775 | nskb = alloc_skb(copy+header, GFP_ATOMIC); | |
3776 | if (!nskb) | |
3777 | return; | |
c51957da | 3778 | |
98e399f8 | 3779 | skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head); |
9c70220b ACM |
3780 | skb_set_network_header(nskb, (skb_network_header(skb) - |
3781 | skb->head)); | |
3782 | skb_set_transport_header(nskb, (skb_transport_header(skb) - | |
3783 | skb->head)); | |
1da177e4 LT |
3784 | skb_reserve(nskb, header); |
3785 | memcpy(nskb->head, skb->head, header); | |
1da177e4 LT |
3786 | memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); |
3787 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; | |
8728b834 | 3788 | __skb_insert(nskb, skb->prev, skb, list); |
1da177e4 LT |
3789 | sk_stream_set_owner_r(nskb, sk); |
3790 | ||
3791 | /* Copy data, releasing collapsed skbs. */ | |
3792 | while (copy > 0) { | |
3793 | int offset = start - TCP_SKB_CB(skb)->seq; | |
3794 | int size = TCP_SKB_CB(skb)->end_seq - start; | |
3795 | ||
09a62660 | 3796 | BUG_ON(offset < 0); |
1da177e4 LT |
3797 | if (size > 0) { |
3798 | size = min(copy, size); | |
3799 | if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) | |
3800 | BUG(); | |
3801 | TCP_SKB_CB(nskb)->end_seq += size; | |
3802 | copy -= size; | |
3803 | start += size; | |
3804 | } | |
3805 | if (!before(start, TCP_SKB_CB(skb)->end_seq)) { | |
3806 | struct sk_buff *next = skb->next; | |
8728b834 | 3807 | __skb_unlink(skb, list); |
1da177e4 LT |
3808 | __kfree_skb(skb); |
3809 | NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); | |
3810 | skb = next; | |
aa8223c7 ACM |
3811 | if (skb == tail || |
3812 | tcp_hdr(skb)->syn || | |
3813 | tcp_hdr(skb)->fin) | |
1da177e4 LT |
3814 | return; |
3815 | } | |
3816 | } | |
3817 | } | |
3818 | } | |
3819 | ||
3820 | /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs | |
3821 | * and tcp_collapse() them until all the queue is collapsed. | |
3822 | */ | |
3823 | static void tcp_collapse_ofo_queue(struct sock *sk) | |
3824 | { | |
3825 | struct tcp_sock *tp = tcp_sk(sk); | |
3826 | struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); | |
3827 | struct sk_buff *head; | |
3828 | u32 start, end; | |
3829 | ||
3830 | if (skb == NULL) | |
3831 | return; | |
3832 | ||
3833 | start = TCP_SKB_CB(skb)->seq; | |
3834 | end = TCP_SKB_CB(skb)->end_seq; | |
3835 | head = skb; | |
3836 | ||
3837 | for (;;) { | |
3838 | skb = skb->next; | |
3839 | ||
3840 | /* Segment is terminated when we see gap or when | |
3841 | * we are at the end of all the queue. */ | |
3842 | if (skb == (struct sk_buff *)&tp->out_of_order_queue || | |
3843 | after(TCP_SKB_CB(skb)->seq, end) || | |
3844 | before(TCP_SKB_CB(skb)->end_seq, start)) { | |
8728b834 DM |
3845 | tcp_collapse(sk, &tp->out_of_order_queue, |
3846 | head, skb, start, end); | |
1da177e4 LT |
3847 | head = skb; |
3848 | if (skb == (struct sk_buff *)&tp->out_of_order_queue) | |
3849 | break; | |
3850 | /* Start new segment */ | |
3851 | start = TCP_SKB_CB(skb)->seq; | |
3852 | end = TCP_SKB_CB(skb)->end_seq; | |
3853 | } else { | |
3854 | if (before(TCP_SKB_CB(skb)->seq, start)) | |
3855 | start = TCP_SKB_CB(skb)->seq; | |
3856 | if (after(TCP_SKB_CB(skb)->end_seq, end)) | |
3857 | end = TCP_SKB_CB(skb)->end_seq; | |
3858 | } | |
3859 | } | |
3860 | } | |
3861 | ||
3862 | /* Reduce allocated memory if we can, trying to get | |
3863 | * the socket within its memory limits again. | |
3864 | * | |
3865 | * Return less than zero if we should start dropping frames | |
3866 | * until the socket owning process reads some of the data | |
3867 | * to stabilize the situation. | |
3868 | */ | |
3869 | static int tcp_prune_queue(struct sock *sk) | |
3870 | { | |
e905a9ed | 3871 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
3872 | |
3873 | SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); | |
3874 | ||
3875 | NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED); | |
3876 | ||
3877 | if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) | |
9e412ba7 | 3878 | tcp_clamp_window(sk); |
1da177e4 LT |
3879 | else if (tcp_memory_pressure) |
3880 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); | |
3881 | ||
3882 | tcp_collapse_ofo_queue(sk); | |
8728b834 DM |
3883 | tcp_collapse(sk, &sk->sk_receive_queue, |
3884 | sk->sk_receive_queue.next, | |
1da177e4 LT |
3885 | (struct sk_buff*)&sk->sk_receive_queue, |
3886 | tp->copied_seq, tp->rcv_nxt); | |
3887 | sk_stream_mem_reclaim(sk); | |
3888 | ||
3889 | if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) | |
3890 | return 0; | |
3891 | ||
3892 | /* Collapsing did not help, destructive actions follow. | |
3893 | * This must not ever occur. */ | |
3894 | ||
3895 | /* First, purge the out_of_order queue. */ | |
b03efcfb DM |
3896 | if (!skb_queue_empty(&tp->out_of_order_queue)) { |
3897 | NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED); | |
1da177e4 LT |
3898 | __skb_queue_purge(&tp->out_of_order_queue); |
3899 | ||
3900 | /* Reset SACK state. A conforming SACK implementation will | |
3901 | * do the same at a timeout based retransmit. When a connection | |
3902 | * is in a sad state like this, we care only about integrity | |
3903 | * of the connection not performance. | |
3904 | */ | |
e60402d0 | 3905 | if (tcp_is_sack(tp)) |
1da177e4 LT |
3906 | tcp_sack_reset(&tp->rx_opt); |
3907 | sk_stream_mem_reclaim(sk); | |
3908 | } | |
3909 | ||
3910 | if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) | |
3911 | return 0; | |
3912 | ||
3913 | /* If we are really being abused, tell the caller to silently | |
3914 | * drop receive data on the floor. It will get retransmitted | |
3915 | * and hopefully then we'll have sufficient space. | |
3916 | */ | |
3917 | NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED); | |
3918 | ||
3919 | /* Massive buffer overcommit. */ | |
3920 | tp->pred_flags = 0; | |
3921 | return -1; | |
3922 | } | |
3923 | ||
3924 | ||
3925 | /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. | |
3926 | * As additional protections, we do not touch cwnd in retransmission phases, | |
3927 | * and if application hit its sndbuf limit recently. | |
3928 | */ | |
3929 | void tcp_cwnd_application_limited(struct sock *sk) | |
3930 | { | |
3931 | struct tcp_sock *tp = tcp_sk(sk); | |
3932 | ||
6687e988 | 3933 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && |
1da177e4 LT |
3934 | sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { |
3935 | /* Limited by application or receiver window. */ | |
d254bcdb IJ |
3936 | u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); |
3937 | u32 win_used = max(tp->snd_cwnd_used, init_win); | |
1da177e4 | 3938 | if (win_used < tp->snd_cwnd) { |
6687e988 | 3939 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
1da177e4 LT |
3940 | tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; |
3941 | } | |
3942 | tp->snd_cwnd_used = 0; | |
3943 | } | |
3944 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
3945 | } | |
3946 | ||
9e412ba7 | 3947 | static int tcp_should_expand_sndbuf(struct sock *sk) |
0d9901df | 3948 | { |
9e412ba7 IJ |
3949 | struct tcp_sock *tp = tcp_sk(sk); |
3950 | ||
0d9901df DM |
3951 | /* If the user specified a specific send buffer setting, do |
3952 | * not modify it. | |
3953 | */ | |
3954 | if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) | |
3955 | return 0; | |
3956 | ||
3957 | /* If we are under global TCP memory pressure, do not expand. */ | |
3958 | if (tcp_memory_pressure) | |
3959 | return 0; | |
3960 | ||
3961 | /* If we are under soft global TCP memory pressure, do not expand. */ | |
3962 | if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0]) | |
3963 | return 0; | |
3964 | ||
3965 | /* If we filled the congestion window, do not expand. */ | |
3966 | if (tp->packets_out >= tp->snd_cwnd) | |
3967 | return 0; | |
3968 | ||
3969 | return 1; | |
3970 | } | |
1da177e4 LT |
3971 | |
3972 | /* When incoming ACK allowed to free some skb from write_queue, | |
3973 | * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket | |
3974 | * on the exit from tcp input handler. | |
3975 | * | |
3976 | * PROBLEM: sndbuf expansion does not work well with largesend. | |
3977 | */ | |
3978 | static void tcp_new_space(struct sock *sk) | |
3979 | { | |
3980 | struct tcp_sock *tp = tcp_sk(sk); | |
3981 | ||
9e412ba7 | 3982 | if (tcp_should_expand_sndbuf(sk)) { |
e905a9ed | 3983 | int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + |
1da177e4 LT |
3984 | MAX_TCP_HEADER + 16 + sizeof(struct sk_buff), |
3985 | demanded = max_t(unsigned int, tp->snd_cwnd, | |
3986 | tp->reordering + 1); | |
3987 | sndmem *= 2*demanded; | |
3988 | if (sndmem > sk->sk_sndbuf) | |
3989 | sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); | |
3990 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
3991 | } | |
3992 | ||
3993 | sk->sk_write_space(sk); | |
3994 | } | |
3995 | ||
40efc6fa | 3996 | static void tcp_check_space(struct sock *sk) |
1da177e4 LT |
3997 | { |
3998 | if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { | |
3999 | sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); | |
4000 | if (sk->sk_socket && | |
4001 | test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) | |
4002 | tcp_new_space(sk); | |
4003 | } | |
4004 | } | |
4005 | ||
9e412ba7 | 4006 | static inline void tcp_data_snd_check(struct sock *sk) |
1da177e4 | 4007 | { |
9e412ba7 | 4008 | tcp_push_pending_frames(sk); |
1da177e4 LT |
4009 | tcp_check_space(sk); |
4010 | } | |
4011 | ||
4012 | /* | |
4013 | * Check if sending an ack is needed. | |
4014 | */ | |
4015 | static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) | |
4016 | { | |
4017 | struct tcp_sock *tp = tcp_sk(sk); | |
4018 | ||
4019 | /* More than one full frame received... */ | |
463c84b9 | 4020 | if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss |
1da177e4 LT |
4021 | /* ... and right edge of window advances far enough. |
4022 | * (tcp_recvmsg() will send ACK otherwise). Or... | |
4023 | */ | |
4024 | && __tcp_select_window(sk) >= tp->rcv_wnd) || | |
4025 | /* We ACK each frame or... */ | |
463c84b9 | 4026 | tcp_in_quickack_mode(sk) || |
1da177e4 LT |
4027 | /* We have out of order data. */ |
4028 | (ofo_possible && | |
4029 | skb_peek(&tp->out_of_order_queue))) { | |
4030 | /* Then ack it now */ | |
4031 | tcp_send_ack(sk); | |
4032 | } else { | |
4033 | /* Else, send delayed ack. */ | |
4034 | tcp_send_delayed_ack(sk); | |
4035 | } | |
4036 | } | |
4037 | ||
40efc6fa | 4038 | static inline void tcp_ack_snd_check(struct sock *sk) |
1da177e4 | 4039 | { |
463c84b9 | 4040 | if (!inet_csk_ack_scheduled(sk)) { |
1da177e4 LT |
4041 | /* We sent a data segment already. */ |
4042 | return; | |
4043 | } | |
4044 | __tcp_ack_snd_check(sk, 1); | |
4045 | } | |
4046 | ||
4047 | /* | |
4048 | * This routine is only called when we have urgent data | |
caa20d9a | 4049 | * signaled. Its the 'slow' part of tcp_urg. It could be |
1da177e4 LT |
4050 | * moved inline now as tcp_urg is only called from one |
4051 | * place. We handle URGent data wrong. We have to - as | |
4052 | * BSD still doesn't use the correction from RFC961. | |
4053 | * For 1003.1g we should support a new option TCP_STDURG to permit | |
4054 | * either form (or just set the sysctl tcp_stdurg). | |
4055 | */ | |
e905a9ed | 4056 | |
1da177e4 LT |
4057 | static void tcp_check_urg(struct sock * sk, struct tcphdr * th) |
4058 | { | |
4059 | struct tcp_sock *tp = tcp_sk(sk); | |
4060 | u32 ptr = ntohs(th->urg_ptr); | |
4061 | ||
4062 | if (ptr && !sysctl_tcp_stdurg) | |
4063 | ptr--; | |
4064 | ptr += ntohl(th->seq); | |
4065 | ||
4066 | /* Ignore urgent data that we've already seen and read. */ | |
4067 | if (after(tp->copied_seq, ptr)) | |
4068 | return; | |
4069 | ||
4070 | /* Do not replay urg ptr. | |
4071 | * | |
4072 | * NOTE: interesting situation not covered by specs. | |
4073 | * Misbehaving sender may send urg ptr, pointing to segment, | |
4074 | * which we already have in ofo queue. We are not able to fetch | |
4075 | * such data and will stay in TCP_URG_NOTYET until will be eaten | |
4076 | * by recvmsg(). Seems, we are not obliged to handle such wicked | |
4077 | * situations. But it is worth to think about possibility of some | |
4078 | * DoSes using some hypothetical application level deadlock. | |
4079 | */ | |
4080 | if (before(ptr, tp->rcv_nxt)) | |
4081 | return; | |
4082 | ||
4083 | /* Do we already have a newer (or duplicate) urgent pointer? */ | |
4084 | if (tp->urg_data && !after(ptr, tp->urg_seq)) | |
4085 | return; | |
4086 | ||
4087 | /* Tell the world about our new urgent pointer. */ | |
4088 | sk_send_sigurg(sk); | |
4089 | ||
4090 | /* We may be adding urgent data when the last byte read was | |
4091 | * urgent. To do this requires some care. We cannot just ignore | |
4092 | * tp->copied_seq since we would read the last urgent byte again | |
4093 | * as data, nor can we alter copied_seq until this data arrives | |
caa20d9a | 4094 | * or we break the semantics of SIOCATMARK (and thus sockatmark()) |
1da177e4 LT |
4095 | * |
4096 | * NOTE. Double Dutch. Rendering to plain English: author of comment | |
4097 | * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); | |
4098 | * and expect that both A and B disappear from stream. This is _wrong_. | |
4099 | * Though this happens in BSD with high probability, this is occasional. | |
4100 | * Any application relying on this is buggy. Note also, that fix "works" | |
4101 | * only in this artificial test. Insert some normal data between A and B and we will | |
4102 | * decline of BSD again. Verdict: it is better to remove to trap | |
4103 | * buggy users. | |
4104 | */ | |
4105 | if (tp->urg_seq == tp->copied_seq && tp->urg_data && | |
4106 | !sock_flag(sk, SOCK_URGINLINE) && | |
4107 | tp->copied_seq != tp->rcv_nxt) { | |
4108 | struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); | |
4109 | tp->copied_seq++; | |
4110 | if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { | |
8728b834 | 4111 | __skb_unlink(skb, &sk->sk_receive_queue); |
1da177e4 LT |
4112 | __kfree_skb(skb); |
4113 | } | |
4114 | } | |
4115 | ||
4116 | tp->urg_data = TCP_URG_NOTYET; | |
4117 | tp->urg_seq = ptr; | |
4118 | ||
4119 | /* Disable header prediction. */ | |
4120 | tp->pred_flags = 0; | |
4121 | } | |
4122 | ||
4123 | /* This is the 'fast' part of urgent handling. */ | |
4124 | static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th) | |
4125 | { | |
4126 | struct tcp_sock *tp = tcp_sk(sk); | |
4127 | ||
4128 | /* Check if we get a new urgent pointer - normally not. */ | |
4129 | if (th->urg) | |
4130 | tcp_check_urg(sk,th); | |
4131 | ||
4132 | /* Do we wait for any urgent data? - normally not... */ | |
4133 | if (tp->urg_data == TCP_URG_NOTYET) { | |
4134 | u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - | |
4135 | th->syn; | |
4136 | ||
e905a9ed | 4137 | /* Is the urgent pointer pointing into this packet? */ |
1da177e4 LT |
4138 | if (ptr < skb->len) { |
4139 | u8 tmp; | |
4140 | if (skb_copy_bits(skb, ptr, &tmp, 1)) | |
4141 | BUG(); | |
4142 | tp->urg_data = TCP_URG_VALID | tmp; | |
4143 | if (!sock_flag(sk, SOCK_DEAD)) | |
4144 | sk->sk_data_ready(sk, 0); | |
4145 | } | |
4146 | } | |
4147 | } | |
4148 | ||
4149 | static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) | |
4150 | { | |
4151 | struct tcp_sock *tp = tcp_sk(sk); | |
4152 | int chunk = skb->len - hlen; | |
4153 | int err; | |
4154 | ||
4155 | local_bh_enable(); | |
60476372 | 4156 | if (skb_csum_unnecessary(skb)) |
1da177e4 LT |
4157 | err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); |
4158 | else | |
4159 | err = skb_copy_and_csum_datagram_iovec(skb, hlen, | |
4160 | tp->ucopy.iov); | |
4161 | ||
4162 | if (!err) { | |
4163 | tp->ucopy.len -= chunk; | |
4164 | tp->copied_seq += chunk; | |
4165 | tcp_rcv_space_adjust(sk); | |
4166 | } | |
4167 | ||
4168 | local_bh_disable(); | |
4169 | return err; | |
4170 | } | |
4171 | ||
b51655b9 | 4172 | static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 4173 | { |
b51655b9 | 4174 | __sum16 result; |
1da177e4 LT |
4175 | |
4176 | if (sock_owned_by_user(sk)) { | |
4177 | local_bh_enable(); | |
4178 | result = __tcp_checksum_complete(skb); | |
4179 | local_bh_disable(); | |
4180 | } else { | |
4181 | result = __tcp_checksum_complete(skb); | |
4182 | } | |
4183 | return result; | |
4184 | } | |
4185 | ||
40efc6fa | 4186 | static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 4187 | { |
60476372 | 4188 | return !skb_csum_unnecessary(skb) && |
1da177e4 LT |
4189 | __tcp_checksum_complete_user(sk, skb); |
4190 | } | |
4191 | ||
1a2449a8 CL |
4192 | #ifdef CONFIG_NET_DMA |
4193 | static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen) | |
4194 | { | |
4195 | struct tcp_sock *tp = tcp_sk(sk); | |
4196 | int chunk = skb->len - hlen; | |
4197 | int dma_cookie; | |
4198 | int copied_early = 0; | |
4199 | ||
4200 | if (tp->ucopy.wakeup) | |
e905a9ed | 4201 | return 0; |
1a2449a8 CL |
4202 | |
4203 | if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) | |
4204 | tp->ucopy.dma_chan = get_softnet_dma(); | |
4205 | ||
60476372 | 4206 | if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) { |
1a2449a8 CL |
4207 | |
4208 | dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, | |
4209 | skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list); | |
4210 | ||
4211 | if (dma_cookie < 0) | |
4212 | goto out; | |
4213 | ||
4214 | tp->ucopy.dma_cookie = dma_cookie; | |
4215 | copied_early = 1; | |
4216 | ||
4217 | tp->ucopy.len -= chunk; | |
4218 | tp->copied_seq += chunk; | |
4219 | tcp_rcv_space_adjust(sk); | |
4220 | ||
4221 | if ((tp->ucopy.len == 0) || | |
aa8223c7 | 4222 | (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) || |
1a2449a8 CL |
4223 | (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { |
4224 | tp->ucopy.wakeup = 1; | |
4225 | sk->sk_data_ready(sk, 0); | |
4226 | } | |
4227 | } else if (chunk > 0) { | |
4228 | tp->ucopy.wakeup = 1; | |
4229 | sk->sk_data_ready(sk, 0); | |
4230 | } | |
4231 | out: | |
4232 | return copied_early; | |
4233 | } | |
4234 | #endif /* CONFIG_NET_DMA */ | |
4235 | ||
1da177e4 | 4236 | /* |
e905a9ed | 4237 | * TCP receive function for the ESTABLISHED state. |
1da177e4 | 4238 | * |
e905a9ed | 4239 | * It is split into a fast path and a slow path. The fast path is |
1da177e4 LT |
4240 | * disabled when: |
4241 | * - A zero window was announced from us - zero window probing | |
e905a9ed | 4242 | * is only handled properly in the slow path. |
1da177e4 LT |
4243 | * - Out of order segments arrived. |
4244 | * - Urgent data is expected. | |
4245 | * - There is no buffer space left | |
4246 | * - Unexpected TCP flags/window values/header lengths are received | |
e905a9ed | 4247 | * (detected by checking the TCP header against pred_flags) |
1da177e4 LT |
4248 | * - Data is sent in both directions. Fast path only supports pure senders |
4249 | * or pure receivers (this means either the sequence number or the ack | |
4250 | * value must stay constant) | |
4251 | * - Unexpected TCP option. | |
4252 | * | |
e905a9ed | 4253 | * When these conditions are not satisfied it drops into a standard |
1da177e4 LT |
4254 | * receive procedure patterned after RFC793 to handle all cases. |
4255 | * The first three cases are guaranteed by proper pred_flags setting, | |
e905a9ed | 4256 | * the rest is checked inline. Fast processing is turned on in |
1da177e4 LT |
4257 | * tcp_data_queue when everything is OK. |
4258 | */ | |
4259 | int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, | |
4260 | struct tcphdr *th, unsigned len) | |
4261 | { | |
4262 | struct tcp_sock *tp = tcp_sk(sk); | |
4263 | ||
4264 | /* | |
4265 | * Header prediction. | |
e905a9ed | 4266 | * The code loosely follows the one in the famous |
1da177e4 | 4267 | * "30 instruction TCP receive" Van Jacobson mail. |
e905a9ed YH |
4268 | * |
4269 | * Van's trick is to deposit buffers into socket queue | |
1da177e4 LT |
4270 | * on a device interrupt, to call tcp_recv function |
4271 | * on the receive process context and checksum and copy | |
4272 | * the buffer to user space. smart... | |
4273 | * | |
e905a9ed | 4274 | * Our current scheme is not silly either but we take the |
1da177e4 LT |
4275 | * extra cost of the net_bh soft interrupt processing... |
4276 | * We do checksum and copy also but from device to kernel. | |
4277 | */ | |
4278 | ||
4279 | tp->rx_opt.saw_tstamp = 0; | |
4280 | ||
4281 | /* pred_flags is 0xS?10 << 16 + snd_wnd | |
caa20d9a | 4282 | * if header_prediction is to be made |
1da177e4 LT |
4283 | * 'S' will always be tp->tcp_header_len >> 2 |
4284 | * '?' will be 0 for the fast path, otherwise pred_flags is 0 to | |
e905a9ed | 4285 | * turn it off (when there are holes in the receive |
1da177e4 LT |
4286 | * space for instance) |
4287 | * PSH flag is ignored. | |
4288 | */ | |
4289 | ||
4290 | if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && | |
4291 | TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { | |
4292 | int tcp_header_len = tp->tcp_header_len; | |
4293 | ||
4294 | /* Timestamp header prediction: tcp_header_len | |
4295 | * is automatically equal to th->doff*4 due to pred_flags | |
4296 | * match. | |
4297 | */ | |
4298 | ||
4299 | /* Check timestamp */ | |
4300 | if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { | |
4f3608b7 | 4301 | __be32 *ptr = (__be32 *)(th + 1); |
1da177e4 LT |
4302 | |
4303 | /* No? Slow path! */ | |
4f3608b7 | 4304 | if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
1da177e4 LT |
4305 | | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) |
4306 | goto slow_path; | |
4307 | ||
4308 | tp->rx_opt.saw_tstamp = 1; | |
e905a9ed | 4309 | ++ptr; |
1da177e4 LT |
4310 | tp->rx_opt.rcv_tsval = ntohl(*ptr); |
4311 | ++ptr; | |
4312 | tp->rx_opt.rcv_tsecr = ntohl(*ptr); | |
4313 | ||
4314 | /* If PAWS failed, check it more carefully in slow path */ | |
4315 | if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) | |
4316 | goto slow_path; | |
4317 | ||
4318 | /* DO NOT update ts_recent here, if checksum fails | |
4319 | * and timestamp was corrupted part, it will result | |
4320 | * in a hung connection since we will drop all | |
4321 | * future packets due to the PAWS test. | |
4322 | */ | |
4323 | } | |
4324 | ||
4325 | if (len <= tcp_header_len) { | |
4326 | /* Bulk data transfer: sender */ | |
4327 | if (len == tcp_header_len) { | |
4328 | /* Predicted packet is in window by definition. | |
4329 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. | |
4330 | * Hence, check seq<=rcv_wup reduces to: | |
4331 | */ | |
4332 | if (tcp_header_len == | |
4333 | (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && | |
4334 | tp->rcv_nxt == tp->rcv_wup) | |
4335 | tcp_store_ts_recent(tp); | |
4336 | ||
1da177e4 LT |
4337 | /* We know that such packets are checksummed |
4338 | * on entry. | |
4339 | */ | |
4340 | tcp_ack(sk, skb, 0); | |
e905a9ed | 4341 | __kfree_skb(skb); |
9e412ba7 | 4342 | tcp_data_snd_check(sk); |
1da177e4 LT |
4343 | return 0; |
4344 | } else { /* Header too small */ | |
4345 | TCP_INC_STATS_BH(TCP_MIB_INERRS); | |
4346 | goto discard; | |
4347 | } | |
4348 | } else { | |
4349 | int eaten = 0; | |
1a2449a8 | 4350 | int copied_early = 0; |
1da177e4 | 4351 | |
1a2449a8 CL |
4352 | if (tp->copied_seq == tp->rcv_nxt && |
4353 | len - tcp_header_len <= tp->ucopy.len) { | |
4354 | #ifdef CONFIG_NET_DMA | |
4355 | if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { | |
4356 | copied_early = 1; | |
4357 | eaten = 1; | |
4358 | } | |
4359 | #endif | |
4360 | if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) { | |
4361 | __set_current_state(TASK_RUNNING); | |
1da177e4 | 4362 | |
1a2449a8 CL |
4363 | if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) |
4364 | eaten = 1; | |
4365 | } | |
4366 | if (eaten) { | |
1da177e4 LT |
4367 | /* Predicted packet is in window by definition. |
4368 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. | |
4369 | * Hence, check seq<=rcv_wup reduces to: | |
4370 | */ | |
4371 | if (tcp_header_len == | |
4372 | (sizeof(struct tcphdr) + | |
4373 | TCPOLEN_TSTAMP_ALIGNED) && | |
4374 | tp->rcv_nxt == tp->rcv_wup) | |
4375 | tcp_store_ts_recent(tp); | |
4376 | ||
463c84b9 | 4377 | tcp_rcv_rtt_measure_ts(sk, skb); |
1da177e4 LT |
4378 | |
4379 | __skb_pull(skb, tcp_header_len); | |
4380 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
4381 | NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER); | |
1da177e4 | 4382 | } |
1a2449a8 CL |
4383 | if (copied_early) |
4384 | tcp_cleanup_rbuf(sk, skb->len); | |
1da177e4 LT |
4385 | } |
4386 | if (!eaten) { | |
4387 | if (tcp_checksum_complete_user(sk, skb)) | |
4388 | goto csum_error; | |
4389 | ||
4390 | /* Predicted packet is in window by definition. | |
4391 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. | |
4392 | * Hence, check seq<=rcv_wup reduces to: | |
4393 | */ | |
4394 | if (tcp_header_len == | |
4395 | (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && | |
4396 | tp->rcv_nxt == tp->rcv_wup) | |
4397 | tcp_store_ts_recent(tp); | |
4398 | ||
463c84b9 | 4399 | tcp_rcv_rtt_measure_ts(sk, skb); |
1da177e4 LT |
4400 | |
4401 | if ((int)skb->truesize > sk->sk_forward_alloc) | |
4402 | goto step5; | |
4403 | ||
4404 | NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS); | |
4405 | ||
4406 | /* Bulk data transfer: receiver */ | |
4407 | __skb_pull(skb,tcp_header_len); | |
4408 | __skb_queue_tail(&sk->sk_receive_queue, skb); | |
4409 | sk_stream_set_owner_r(skb, sk); | |
4410 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
4411 | } | |
4412 | ||
9e412ba7 | 4413 | tcp_event_data_recv(sk, skb); |
1da177e4 LT |
4414 | |
4415 | if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { | |
4416 | /* Well, only one small jumplet in fast path... */ | |
4417 | tcp_ack(sk, skb, FLAG_DATA); | |
9e412ba7 | 4418 | tcp_data_snd_check(sk); |
463c84b9 | 4419 | if (!inet_csk_ack_scheduled(sk)) |
1da177e4 LT |
4420 | goto no_ack; |
4421 | } | |
4422 | ||
31432412 | 4423 | __tcp_ack_snd_check(sk, 0); |
1da177e4 | 4424 | no_ack: |
1a2449a8 CL |
4425 | #ifdef CONFIG_NET_DMA |
4426 | if (copied_early) | |
4427 | __skb_queue_tail(&sk->sk_async_wait_queue, skb); | |
4428 | else | |
4429 | #endif | |
1da177e4 LT |
4430 | if (eaten) |
4431 | __kfree_skb(skb); | |
4432 | else | |
4433 | sk->sk_data_ready(sk, 0); | |
4434 | return 0; | |
4435 | } | |
4436 | } | |
4437 | ||
4438 | slow_path: | |
4439 | if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb)) | |
4440 | goto csum_error; | |
4441 | ||
4442 | /* | |
4443 | * RFC1323: H1. Apply PAWS check first. | |
4444 | */ | |
4445 | if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && | |
463c84b9 | 4446 | tcp_paws_discard(sk, skb)) { |
1da177e4 LT |
4447 | if (!th->rst) { |
4448 | NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); | |
4449 | tcp_send_dupack(sk, skb); | |
4450 | goto discard; | |
4451 | } | |
4452 | /* Resets are accepted even if PAWS failed. | |
4453 | ||
4454 | ts_recent update must be made after we are sure | |
4455 | that the packet is in window. | |
4456 | */ | |
4457 | } | |
4458 | ||
4459 | /* | |
4460 | * Standard slow path. | |
4461 | */ | |
4462 | ||
4463 | if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { | |
4464 | /* RFC793, page 37: "In all states except SYN-SENT, all reset | |
4465 | * (RST) segments are validated by checking their SEQ-fields." | |
4466 | * And page 69: "If an incoming segment is not acceptable, | |
4467 | * an acknowledgment should be sent in reply (unless the RST bit | |
4468 | * is set, if so drop the segment and return)". | |
4469 | */ | |
4470 | if (!th->rst) | |
4471 | tcp_send_dupack(sk, skb); | |
4472 | goto discard; | |
4473 | } | |
4474 | ||
2de979bd | 4475 | if (th->rst) { |
1da177e4 LT |
4476 | tcp_reset(sk); |
4477 | goto discard; | |
4478 | } | |
4479 | ||
4480 | tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); | |
4481 | ||
4482 | if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
4483 | TCP_INC_STATS_BH(TCP_MIB_INERRS); | |
4484 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); | |
4485 | tcp_reset(sk); | |
4486 | return 1; | |
4487 | } | |
4488 | ||
4489 | step5: | |
2de979bd | 4490 | if (th->ack) |
1da177e4 LT |
4491 | tcp_ack(sk, skb, FLAG_SLOWPATH); |
4492 | ||
463c84b9 | 4493 | tcp_rcv_rtt_measure_ts(sk, skb); |
1da177e4 LT |
4494 | |
4495 | /* Process urgent data. */ | |
4496 | tcp_urg(sk, skb, th); | |
4497 | ||
4498 | /* step 7: process the segment text */ | |
4499 | tcp_data_queue(sk, skb); | |
4500 | ||
9e412ba7 | 4501 | tcp_data_snd_check(sk); |
1da177e4 LT |
4502 | tcp_ack_snd_check(sk); |
4503 | return 0; | |
4504 | ||
4505 | csum_error: | |
4506 | TCP_INC_STATS_BH(TCP_MIB_INERRS); | |
4507 | ||
4508 | discard: | |
4509 | __kfree_skb(skb); | |
4510 | return 0; | |
4511 | } | |
4512 | ||
4513 | static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, | |
4514 | struct tcphdr *th, unsigned len) | |
4515 | { | |
4516 | struct tcp_sock *tp = tcp_sk(sk); | |
d83d8461 | 4517 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
4518 | int saved_clamp = tp->rx_opt.mss_clamp; |
4519 | ||
4520 | tcp_parse_options(skb, &tp->rx_opt, 0); | |
4521 | ||
4522 | if (th->ack) { | |
4523 | /* rfc793: | |
4524 | * "If the state is SYN-SENT then | |
4525 | * first check the ACK bit | |
4526 | * If the ACK bit is set | |
4527 | * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send | |
4528 | * a reset (unless the RST bit is set, if so drop | |
4529 | * the segment and return)" | |
4530 | * | |
4531 | * We do not send data with SYN, so that RFC-correct | |
4532 | * test reduces to: | |
4533 | */ | |
4534 | if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) | |
4535 | goto reset_and_undo; | |
4536 | ||
4537 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && | |
4538 | !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, | |
4539 | tcp_time_stamp)) { | |
4540 | NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED); | |
4541 | goto reset_and_undo; | |
4542 | } | |
4543 | ||
4544 | /* Now ACK is acceptable. | |
4545 | * | |
4546 | * "If the RST bit is set | |
4547 | * If the ACK was acceptable then signal the user "error: | |
4548 | * connection reset", drop the segment, enter CLOSED state, | |
4549 | * delete TCB, and return." | |
4550 | */ | |
4551 | ||
4552 | if (th->rst) { | |
4553 | tcp_reset(sk); | |
4554 | goto discard; | |
4555 | } | |
4556 | ||
4557 | /* rfc793: | |
4558 | * "fifth, if neither of the SYN or RST bits is set then | |
4559 | * drop the segment and return." | |
4560 | * | |
4561 | * See note below! | |
4562 | * --ANK(990513) | |
4563 | */ | |
4564 | if (!th->syn) | |
4565 | goto discard_and_undo; | |
4566 | ||
4567 | /* rfc793: | |
4568 | * "If the SYN bit is on ... | |
4569 | * are acceptable then ... | |
4570 | * (our SYN has been ACKed), change the connection | |
4571 | * state to ESTABLISHED..." | |
4572 | */ | |
4573 | ||
4574 | TCP_ECN_rcv_synack(tp, th); | |
1da177e4 LT |
4575 | |
4576 | tp->snd_wl1 = TCP_SKB_CB(skb)->seq; | |
4577 | tcp_ack(sk, skb, FLAG_SLOWPATH); | |
4578 | ||
4579 | /* Ok.. it's good. Set up sequence numbers and | |
4580 | * move to established. | |
4581 | */ | |
4582 | tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; | |
4583 | tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; | |
4584 | ||
4585 | /* RFC1323: The window in SYN & SYN/ACK segments is | |
4586 | * never scaled. | |
4587 | */ | |
4588 | tp->snd_wnd = ntohs(th->window); | |
4589 | tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq); | |
4590 | ||
4591 | if (!tp->rx_opt.wscale_ok) { | |
4592 | tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; | |
4593 | tp->window_clamp = min(tp->window_clamp, 65535U); | |
4594 | } | |
4595 | ||
4596 | if (tp->rx_opt.saw_tstamp) { | |
4597 | tp->rx_opt.tstamp_ok = 1; | |
4598 | tp->tcp_header_len = | |
4599 | sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; | |
4600 | tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; | |
4601 | tcp_store_ts_recent(tp); | |
4602 | } else { | |
4603 | tp->tcp_header_len = sizeof(struct tcphdr); | |
4604 | } | |
4605 | ||
e60402d0 IJ |
4606 | if (tcp_is_sack(tp) && sysctl_tcp_fack) |
4607 | tcp_enable_fack(tp); | |
1da177e4 | 4608 | |
5d424d5a | 4609 | tcp_mtup_init(sk); |
d83d8461 | 4610 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); |
1da177e4 LT |
4611 | tcp_initialize_rcv_mss(sk); |
4612 | ||
4613 | /* Remember, tcp_poll() does not lock socket! | |
4614 | * Change state from SYN-SENT only after copied_seq | |
4615 | * is initialized. */ | |
4616 | tp->copied_seq = tp->rcv_nxt; | |
e16aa207 | 4617 | smp_mb(); |
1da177e4 LT |
4618 | tcp_set_state(sk, TCP_ESTABLISHED); |
4619 | ||
6b877699 VY |
4620 | security_inet_conn_established(sk, skb); |
4621 | ||
1da177e4 | 4622 | /* Make sure socket is routed, for correct metrics. */ |
8292a17a | 4623 | icsk->icsk_af_ops->rebuild_header(sk); |
1da177e4 LT |
4624 | |
4625 | tcp_init_metrics(sk); | |
4626 | ||
6687e988 | 4627 | tcp_init_congestion_control(sk); |
317a76f9 | 4628 | |
1da177e4 LT |
4629 | /* Prevent spurious tcp_cwnd_restart() on first data |
4630 | * packet. | |
4631 | */ | |
4632 | tp->lsndtime = tcp_time_stamp; | |
4633 | ||
4634 | tcp_init_buffer_space(sk); | |
4635 | ||
4636 | if (sock_flag(sk, SOCK_KEEPOPEN)) | |
463c84b9 | 4637 | inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); |
1da177e4 LT |
4638 | |
4639 | if (!tp->rx_opt.snd_wscale) | |
4640 | __tcp_fast_path_on(tp, tp->snd_wnd); | |
4641 | else | |
4642 | tp->pred_flags = 0; | |
4643 | ||
4644 | if (!sock_flag(sk, SOCK_DEAD)) { | |
4645 | sk->sk_state_change(sk); | |
4646 | sk_wake_async(sk, 0, POLL_OUT); | |
4647 | } | |
4648 | ||
295f7324 ACM |
4649 | if (sk->sk_write_pending || |
4650 | icsk->icsk_accept_queue.rskq_defer_accept || | |
4651 | icsk->icsk_ack.pingpong) { | |
1da177e4 LT |
4652 | /* Save one ACK. Data will be ready after |
4653 | * several ticks, if write_pending is set. | |
4654 | * | |
4655 | * It may be deleted, but with this feature tcpdumps | |
4656 | * look so _wonderfully_ clever, that I was not able | |
4657 | * to stand against the temptation 8) --ANK | |
4658 | */ | |
463c84b9 | 4659 | inet_csk_schedule_ack(sk); |
295f7324 ACM |
4660 | icsk->icsk_ack.lrcvtime = tcp_time_stamp; |
4661 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
463c84b9 ACM |
4662 | tcp_incr_quickack(sk); |
4663 | tcp_enter_quickack_mode(sk); | |
3f421baa ACM |
4664 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, |
4665 | TCP_DELACK_MAX, TCP_RTO_MAX); | |
1da177e4 LT |
4666 | |
4667 | discard: | |
4668 | __kfree_skb(skb); | |
4669 | return 0; | |
4670 | } else { | |
4671 | tcp_send_ack(sk); | |
4672 | } | |
4673 | return -1; | |
4674 | } | |
4675 | ||
4676 | /* No ACK in the segment */ | |
4677 | ||
4678 | if (th->rst) { | |
4679 | /* rfc793: | |
4680 | * "If the RST bit is set | |
4681 | * | |
4682 | * Otherwise (no ACK) drop the segment and return." | |
4683 | */ | |
4684 | ||
4685 | goto discard_and_undo; | |
4686 | } | |
4687 | ||
4688 | /* PAWS check. */ | |
4689 | if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0)) | |
4690 | goto discard_and_undo; | |
4691 | ||
4692 | if (th->syn) { | |
4693 | /* We see SYN without ACK. It is attempt of | |
4694 | * simultaneous connect with crossed SYNs. | |
4695 | * Particularly, it can be connect to self. | |
4696 | */ | |
4697 | tcp_set_state(sk, TCP_SYN_RECV); | |
4698 | ||
4699 | if (tp->rx_opt.saw_tstamp) { | |
4700 | tp->rx_opt.tstamp_ok = 1; | |
4701 | tcp_store_ts_recent(tp); | |
4702 | tp->tcp_header_len = | |
4703 | sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; | |
4704 | } else { | |
4705 | tp->tcp_header_len = sizeof(struct tcphdr); | |
4706 | } | |
4707 | ||
4708 | tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; | |
4709 | tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; | |
4710 | ||
4711 | /* RFC1323: The window in SYN & SYN/ACK segments is | |
4712 | * never scaled. | |
4713 | */ | |
4714 | tp->snd_wnd = ntohs(th->window); | |
4715 | tp->snd_wl1 = TCP_SKB_CB(skb)->seq; | |
4716 | tp->max_window = tp->snd_wnd; | |
4717 | ||
4718 | TCP_ECN_rcv_syn(tp, th); | |
1da177e4 | 4719 | |
5d424d5a | 4720 | tcp_mtup_init(sk); |
d83d8461 | 4721 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); |
1da177e4 LT |
4722 | tcp_initialize_rcv_mss(sk); |
4723 | ||
4724 | ||
4725 | tcp_send_synack(sk); | |
4726 | #if 0 | |
4727 | /* Note, we could accept data and URG from this segment. | |
4728 | * There are no obstacles to make this. | |
4729 | * | |
4730 | * However, if we ignore data in ACKless segments sometimes, | |
4731 | * we have no reasons to accept it sometimes. | |
4732 | * Also, seems the code doing it in step6 of tcp_rcv_state_process | |
4733 | * is not flawless. So, discard packet for sanity. | |
4734 | * Uncomment this return to process the data. | |
4735 | */ | |
4736 | return -1; | |
4737 | #else | |
4738 | goto discard; | |
4739 | #endif | |
4740 | } | |
4741 | /* "fifth, if neither of the SYN or RST bits is set then | |
4742 | * drop the segment and return." | |
4743 | */ | |
4744 | ||
4745 | discard_and_undo: | |
4746 | tcp_clear_options(&tp->rx_opt); | |
4747 | tp->rx_opt.mss_clamp = saved_clamp; | |
4748 | goto discard; | |
4749 | ||
4750 | reset_and_undo: | |
4751 | tcp_clear_options(&tp->rx_opt); | |
4752 | tp->rx_opt.mss_clamp = saved_clamp; | |
4753 | return 1; | |
4754 | } | |
4755 | ||
4756 | ||
4757 | /* | |
4758 | * This function implements the receiving procedure of RFC 793 for | |
e905a9ed | 4759 | * all states except ESTABLISHED and TIME_WAIT. |
1da177e4 LT |
4760 | * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be |
4761 | * address independent. | |
4762 | */ | |
e905a9ed | 4763 | |
1da177e4 LT |
4764 | int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, |
4765 | struct tcphdr *th, unsigned len) | |
4766 | { | |
4767 | struct tcp_sock *tp = tcp_sk(sk); | |
8292a17a | 4768 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
4769 | int queued = 0; |
4770 | ||
4771 | tp->rx_opt.saw_tstamp = 0; | |
4772 | ||
4773 | switch (sk->sk_state) { | |
4774 | case TCP_CLOSE: | |
4775 | goto discard; | |
4776 | ||
4777 | case TCP_LISTEN: | |
2de979bd | 4778 | if (th->ack) |
1da177e4 LT |
4779 | return 1; |
4780 | ||
2de979bd | 4781 | if (th->rst) |
1da177e4 LT |
4782 | goto discard; |
4783 | ||
2de979bd | 4784 | if (th->syn) { |
8292a17a | 4785 | if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) |
1da177e4 LT |
4786 | return 1; |
4787 | ||
e905a9ed YH |
4788 | /* Now we have several options: In theory there is |
4789 | * nothing else in the frame. KA9Q has an option to | |
1da177e4 | 4790 | * send data with the syn, BSD accepts data with the |
e905a9ed YH |
4791 | * syn up to the [to be] advertised window and |
4792 | * Solaris 2.1 gives you a protocol error. For now | |
4793 | * we just ignore it, that fits the spec precisely | |
1da177e4 LT |
4794 | * and avoids incompatibilities. It would be nice in |
4795 | * future to drop through and process the data. | |
4796 | * | |
e905a9ed | 4797 | * Now that TTCP is starting to be used we ought to |
1da177e4 LT |
4798 | * queue this data. |
4799 | * But, this leaves one open to an easy denial of | |
e905a9ed | 4800 | * service attack, and SYN cookies can't defend |
1da177e4 | 4801 | * against this problem. So, we drop the data |
fb7e2399 MN |
4802 | * in the interest of security over speed unless |
4803 | * it's still in use. | |
1da177e4 | 4804 | */ |
fb7e2399 MN |
4805 | kfree_skb(skb); |
4806 | return 0; | |
1da177e4 LT |
4807 | } |
4808 | goto discard; | |
4809 | ||
4810 | case TCP_SYN_SENT: | |
1da177e4 LT |
4811 | queued = tcp_rcv_synsent_state_process(sk, skb, th, len); |
4812 | if (queued >= 0) | |
4813 | return queued; | |
4814 | ||
4815 | /* Do step6 onward by hand. */ | |
4816 | tcp_urg(sk, skb, th); | |
4817 | __kfree_skb(skb); | |
9e412ba7 | 4818 | tcp_data_snd_check(sk); |
1da177e4 LT |
4819 | return 0; |
4820 | } | |
4821 | ||
4822 | if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && | |
463c84b9 | 4823 | tcp_paws_discard(sk, skb)) { |
1da177e4 LT |
4824 | if (!th->rst) { |
4825 | NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); | |
4826 | tcp_send_dupack(sk, skb); | |
4827 | goto discard; | |
4828 | } | |
4829 | /* Reset is accepted even if it did not pass PAWS. */ | |
4830 | } | |
4831 | ||
4832 | /* step 1: check sequence number */ | |
4833 | if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { | |
4834 | if (!th->rst) | |
4835 | tcp_send_dupack(sk, skb); | |
4836 | goto discard; | |
4837 | } | |
4838 | ||
4839 | /* step 2: check RST bit */ | |
2de979bd | 4840 | if (th->rst) { |
1da177e4 LT |
4841 | tcp_reset(sk); |
4842 | goto discard; | |
4843 | } | |
4844 | ||
4845 | tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); | |
4846 | ||
4847 | /* step 3: check security and precedence [ignored] */ | |
4848 | ||
4849 | /* step 4: | |
4850 | * | |
4851 | * Check for a SYN in window. | |
4852 | */ | |
4853 | if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
4854 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); | |
4855 | tcp_reset(sk); | |
4856 | return 1; | |
4857 | } | |
4858 | ||
4859 | /* step 5: check the ACK field */ | |
4860 | if (th->ack) { | |
4861 | int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH); | |
4862 | ||
2de979bd | 4863 | switch (sk->sk_state) { |
1da177e4 LT |
4864 | case TCP_SYN_RECV: |
4865 | if (acceptable) { | |
4866 | tp->copied_seq = tp->rcv_nxt; | |
e16aa207 | 4867 | smp_mb(); |
1da177e4 LT |
4868 | tcp_set_state(sk, TCP_ESTABLISHED); |
4869 | sk->sk_state_change(sk); | |
4870 | ||
4871 | /* Note, that this wakeup is only for marginal | |
4872 | * crossed SYN case. Passively open sockets | |
4873 | * are not waked up, because sk->sk_sleep == | |
4874 | * NULL and sk->sk_socket == NULL. | |
4875 | */ | |
4876 | if (sk->sk_socket) { | |
4877 | sk_wake_async(sk,0,POLL_OUT); | |
4878 | } | |
4879 | ||
4880 | tp->snd_una = TCP_SKB_CB(skb)->ack_seq; | |
4881 | tp->snd_wnd = ntohs(th->window) << | |
4882 | tp->rx_opt.snd_wscale; | |
4883 | tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, | |
4884 | TCP_SKB_CB(skb)->seq); | |
4885 | ||
4886 | /* tcp_ack considers this ACK as duplicate | |
4887 | * and does not calculate rtt. | |
4888 | * Fix it at least with timestamps. | |
4889 | */ | |
4890 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && | |
4891 | !tp->srtt) | |
2d2abbab | 4892 | tcp_ack_saw_tstamp(sk, 0); |
1da177e4 LT |
4893 | |
4894 | if (tp->rx_opt.tstamp_ok) | |
4895 | tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; | |
4896 | ||
4897 | /* Make sure socket is routed, for | |
4898 | * correct metrics. | |
4899 | */ | |
8292a17a | 4900 | icsk->icsk_af_ops->rebuild_header(sk); |
1da177e4 LT |
4901 | |
4902 | tcp_init_metrics(sk); | |
4903 | ||
6687e988 | 4904 | tcp_init_congestion_control(sk); |
317a76f9 | 4905 | |
1da177e4 LT |
4906 | /* Prevent spurious tcp_cwnd_restart() on |
4907 | * first data packet. | |
4908 | */ | |
4909 | tp->lsndtime = tcp_time_stamp; | |
4910 | ||
5d424d5a | 4911 | tcp_mtup_init(sk); |
1da177e4 LT |
4912 | tcp_initialize_rcv_mss(sk); |
4913 | tcp_init_buffer_space(sk); | |
4914 | tcp_fast_path_on(tp); | |
4915 | } else { | |
4916 | return 1; | |
4917 | } | |
4918 | break; | |
4919 | ||
4920 | case TCP_FIN_WAIT1: | |
4921 | if (tp->snd_una == tp->write_seq) { | |
4922 | tcp_set_state(sk, TCP_FIN_WAIT2); | |
4923 | sk->sk_shutdown |= SEND_SHUTDOWN; | |
4924 | dst_confirm(sk->sk_dst_cache); | |
4925 | ||
4926 | if (!sock_flag(sk, SOCK_DEAD)) | |
4927 | /* Wake up lingering close() */ | |
4928 | sk->sk_state_change(sk); | |
4929 | else { | |
4930 | int tmo; | |
4931 | ||
4932 | if (tp->linger2 < 0 || | |
4933 | (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && | |
4934 | after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { | |
4935 | tcp_done(sk); | |
4936 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); | |
4937 | return 1; | |
4938 | } | |
4939 | ||
463c84b9 | 4940 | tmo = tcp_fin_time(sk); |
1da177e4 | 4941 | if (tmo > TCP_TIMEWAIT_LEN) { |
463c84b9 | 4942 | inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); |
1da177e4 LT |
4943 | } else if (th->fin || sock_owned_by_user(sk)) { |
4944 | /* Bad case. We could lose such FIN otherwise. | |
4945 | * It is not a big problem, but it looks confusing | |
4946 | * and not so rare event. We still can lose it now, | |
4947 | * if it spins in bh_lock_sock(), but it is really | |
4948 | * marginal case. | |
4949 | */ | |
463c84b9 | 4950 | inet_csk_reset_keepalive_timer(sk, tmo); |
1da177e4 LT |
4951 | } else { |
4952 | tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); | |
4953 | goto discard; | |
4954 | } | |
4955 | } | |
4956 | } | |
4957 | break; | |
4958 | ||
4959 | case TCP_CLOSING: | |
4960 | if (tp->snd_una == tp->write_seq) { | |
4961 | tcp_time_wait(sk, TCP_TIME_WAIT, 0); | |
4962 | goto discard; | |
4963 | } | |
4964 | break; | |
4965 | ||
4966 | case TCP_LAST_ACK: | |
4967 | if (tp->snd_una == tp->write_seq) { | |
4968 | tcp_update_metrics(sk); | |
4969 | tcp_done(sk); | |
4970 | goto discard; | |
4971 | } | |
4972 | break; | |
4973 | } | |
4974 | } else | |
4975 | goto discard; | |
4976 | ||
4977 | /* step 6: check the URG bit */ | |
4978 | tcp_urg(sk, skb, th); | |
4979 | ||
4980 | /* step 7: process the segment text */ | |
4981 | switch (sk->sk_state) { | |
4982 | case TCP_CLOSE_WAIT: | |
4983 | case TCP_CLOSING: | |
4984 | case TCP_LAST_ACK: | |
4985 | if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) | |
4986 | break; | |
4987 | case TCP_FIN_WAIT1: | |
4988 | case TCP_FIN_WAIT2: | |
4989 | /* RFC 793 says to queue data in these states, | |
e905a9ed | 4990 | * RFC 1122 says we MUST send a reset. |
1da177e4 LT |
4991 | * BSD 4.4 also does reset. |
4992 | */ | |
4993 | if (sk->sk_shutdown & RCV_SHUTDOWN) { | |
4994 | if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && | |
4995 | after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { | |
4996 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); | |
4997 | tcp_reset(sk); | |
4998 | return 1; | |
4999 | } | |
5000 | } | |
5001 | /* Fall through */ | |
e905a9ed | 5002 | case TCP_ESTABLISHED: |
1da177e4 LT |
5003 | tcp_data_queue(sk, skb); |
5004 | queued = 1; | |
5005 | break; | |
5006 | } | |
5007 | ||
5008 | /* tcp_data could move socket to TIME-WAIT */ | |
5009 | if (sk->sk_state != TCP_CLOSE) { | |
9e412ba7 | 5010 | tcp_data_snd_check(sk); |
1da177e4 LT |
5011 | tcp_ack_snd_check(sk); |
5012 | } | |
5013 | ||
e905a9ed | 5014 | if (!queued) { |
1da177e4 LT |
5015 | discard: |
5016 | __kfree_skb(skb); | |
5017 | } | |
5018 | return 0; | |
5019 | } | |
5020 | ||
5021 | EXPORT_SYMBOL(sysctl_tcp_ecn); | |
5022 | EXPORT_SYMBOL(sysctl_tcp_reordering); | |
5023 | EXPORT_SYMBOL(tcp_parse_options); | |
5024 | EXPORT_SYMBOL(tcp_rcv_established); | |
5025 | EXPORT_SYMBOL(tcp_rcv_state_process); | |
40efc6fa | 5026 | EXPORT_SYMBOL(tcp_initialize_rcv_mss); |