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. | |
53 | * Andi Kleen: Add tcp_measure_rcv_mss to make | |
54 | * connections with MSS<min(MTU,ann. MSS) | |
55 | * work without delayed acks. | |
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 LT |
74 | |
75 | int sysctl_tcp_timestamps = 1; | |
76 | int sysctl_tcp_window_scaling = 1; | |
77 | int sysctl_tcp_sack = 1; | |
78 | int sysctl_tcp_fack = 1; | |
79 | int sysctl_tcp_reordering = TCP_FASTRETRANS_THRESH; | |
80 | int sysctl_tcp_ecn; | |
81 | int sysctl_tcp_dsack = 1; | |
82 | int sysctl_tcp_app_win = 31; | |
83 | int sysctl_tcp_adv_win_scale = 2; | |
84 | ||
85 | int sysctl_tcp_stdurg; | |
86 | int sysctl_tcp_rfc1337; | |
87 | int sysctl_tcp_max_orphans = NR_FILE; | |
88 | int sysctl_tcp_frto; | |
89 | int sysctl_tcp_nometrics_save; | |
1da177e4 LT |
90 | |
91 | int sysctl_tcp_moderate_rcvbuf = 1; | |
b3a8a40d | 92 | int sysctl_tcp_abc; |
1da177e4 | 93 | |
1da177e4 LT |
94 | #define FLAG_DATA 0x01 /* Incoming frame contained data. */ |
95 | #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ | |
96 | #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ | |
97 | #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ | |
98 | #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ | |
99 | #define FLAG_DATA_SACKED 0x20 /* New SACK. */ | |
100 | #define FLAG_ECE 0x40 /* ECE in this ACK */ | |
101 | #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */ | |
102 | #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ | |
103 | ||
104 | #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) | |
105 | #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) | |
106 | #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) | |
107 | #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) | |
108 | ||
109 | #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0) | |
110 | #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2) | |
111 | #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4) | |
112 | ||
113 | #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) | |
114 | ||
115 | /* Adapt the MSS value used to make delayed ack decision to the | |
116 | * real world. | |
117 | */ | |
40efc6fa SH |
118 | static void tcp_measure_rcv_mss(struct sock *sk, |
119 | const struct sk_buff *skb) | |
1da177e4 | 120 | { |
463c84b9 ACM |
121 | struct inet_connection_sock *icsk = inet_csk(sk); |
122 | const unsigned int lss = icsk->icsk_ack.last_seg_size; | |
123 | unsigned int len; | |
1da177e4 | 124 | |
463c84b9 | 125 | icsk->icsk_ack.last_seg_size = 0; |
1da177e4 LT |
126 | |
127 | /* skb->len may jitter because of SACKs, even if peer | |
128 | * sends good full-sized frames. | |
129 | */ | |
130 | len = skb->len; | |
463c84b9 ACM |
131 | if (len >= icsk->icsk_ack.rcv_mss) { |
132 | icsk->icsk_ack.rcv_mss = len; | |
1da177e4 LT |
133 | } else { |
134 | /* Otherwise, we make more careful check taking into account, | |
135 | * that SACKs block is variable. | |
136 | * | |
137 | * "len" is invariant segment length, including TCP header. | |
138 | */ | |
139 | len += skb->data - skb->h.raw; | |
140 | if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) || | |
141 | /* If PSH is not set, packet should be | |
142 | * full sized, provided peer TCP is not badly broken. | |
143 | * This observation (if it is correct 8)) allows | |
144 | * to handle super-low mtu links fairly. | |
145 | */ | |
146 | (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && | |
147 | !(tcp_flag_word(skb->h.th)&TCP_REMNANT))) { | |
148 | /* Subtract also invariant (if peer is RFC compliant), | |
149 | * tcp header plus fixed timestamp option length. | |
150 | * Resulting "len" is MSS free of SACK jitter. | |
151 | */ | |
463c84b9 ACM |
152 | len -= tcp_sk(sk)->tcp_header_len; |
153 | icsk->icsk_ack.last_seg_size = len; | |
1da177e4 | 154 | if (len == lss) { |
463c84b9 | 155 | icsk->icsk_ack.rcv_mss = len; |
1da177e4 LT |
156 | return; |
157 | } | |
158 | } | |
463c84b9 | 159 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; |
1da177e4 LT |
160 | } |
161 | } | |
162 | ||
463c84b9 | 163 | static void tcp_incr_quickack(struct sock *sk) |
1da177e4 | 164 | { |
463c84b9 ACM |
165 | struct inet_connection_sock *icsk = inet_csk(sk); |
166 | unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); | |
1da177e4 LT |
167 | |
168 | if (quickacks==0) | |
169 | quickacks=2; | |
463c84b9 ACM |
170 | if (quickacks > icsk->icsk_ack.quick) |
171 | icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); | |
1da177e4 LT |
172 | } |
173 | ||
463c84b9 | 174 | void tcp_enter_quickack_mode(struct sock *sk) |
1da177e4 | 175 | { |
463c84b9 ACM |
176 | struct inet_connection_sock *icsk = inet_csk(sk); |
177 | tcp_incr_quickack(sk); | |
178 | icsk->icsk_ack.pingpong = 0; | |
179 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 LT |
180 | } |
181 | ||
182 | /* Send ACKs quickly, if "quick" count is not exhausted | |
183 | * and the session is not interactive. | |
184 | */ | |
185 | ||
463c84b9 | 186 | static inline int tcp_in_quickack_mode(const struct sock *sk) |
1da177e4 | 187 | { |
463c84b9 ACM |
188 | const struct inet_connection_sock *icsk = inet_csk(sk); |
189 | return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; | |
1da177e4 LT |
190 | } |
191 | ||
192 | /* Buffer size and advertised window tuning. | |
193 | * | |
194 | * 1. Tuning sk->sk_sndbuf, when connection enters established state. | |
195 | */ | |
196 | ||
197 | static void tcp_fixup_sndbuf(struct sock *sk) | |
198 | { | |
199 | int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 + | |
200 | sizeof(struct sk_buff); | |
201 | ||
202 | if (sk->sk_sndbuf < 3 * sndmem) | |
203 | sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]); | |
204 | } | |
205 | ||
206 | /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) | |
207 | * | |
208 | * All tcp_full_space() is split to two parts: "network" buffer, allocated | |
209 | * forward and advertised in receiver window (tp->rcv_wnd) and | |
210 | * "application buffer", required to isolate scheduling/application | |
211 | * latencies from network. | |
212 | * window_clamp is maximal advertised window. It can be less than | |
213 | * tcp_full_space(), in this case tcp_full_space() - window_clamp | |
214 | * is reserved for "application" buffer. The less window_clamp is | |
215 | * the smoother our behaviour from viewpoint of network, but the lower | |
216 | * throughput and the higher sensitivity of the connection to losses. 8) | |
217 | * | |
218 | * rcv_ssthresh is more strict window_clamp used at "slow start" | |
219 | * phase to predict further behaviour of this connection. | |
220 | * It is used for two goals: | |
221 | * - to enforce header prediction at sender, even when application | |
222 | * requires some significant "application buffer". It is check #1. | |
223 | * - to prevent pruning of receive queue because of misprediction | |
224 | * of receiver window. Check #2. | |
225 | * | |
226 | * The scheme does not work when sender sends good segments opening | |
caa20d9a | 227 | * window and then starts to feed us spaghetti. But it should work |
1da177e4 LT |
228 | * in common situations. Otherwise, we have to rely on queue collapsing. |
229 | */ | |
230 | ||
231 | /* Slow part of check#2. */ | |
463c84b9 ACM |
232 | static int __tcp_grow_window(const struct sock *sk, struct tcp_sock *tp, |
233 | const struct sk_buff *skb) | |
1da177e4 LT |
234 | { |
235 | /* Optimize this! */ | |
236 | int truesize = tcp_win_from_space(skb->truesize)/2; | |
326f36e9 | 237 | int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2; |
1da177e4 LT |
238 | |
239 | while (tp->rcv_ssthresh <= window) { | |
240 | if (truesize <= skb->len) | |
463c84b9 | 241 | return 2 * inet_csk(sk)->icsk_ack.rcv_mss; |
1da177e4 LT |
242 | |
243 | truesize >>= 1; | |
244 | window >>= 1; | |
245 | } | |
246 | return 0; | |
247 | } | |
248 | ||
40efc6fa SH |
249 | static void tcp_grow_window(struct sock *sk, struct tcp_sock *tp, |
250 | struct sk_buff *skb) | |
1da177e4 LT |
251 | { |
252 | /* Check #1 */ | |
253 | if (tp->rcv_ssthresh < tp->window_clamp && | |
254 | (int)tp->rcv_ssthresh < tcp_space(sk) && | |
255 | !tcp_memory_pressure) { | |
256 | int incr; | |
257 | ||
258 | /* Check #2. Increase window, if skb with such overhead | |
259 | * will fit to rcvbuf in future. | |
260 | */ | |
261 | if (tcp_win_from_space(skb->truesize) <= skb->len) | |
262 | incr = 2*tp->advmss; | |
263 | else | |
264 | incr = __tcp_grow_window(sk, tp, skb); | |
265 | ||
266 | if (incr) { | |
267 | tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp); | |
463c84b9 | 268 | inet_csk(sk)->icsk_ack.quick |= 1; |
1da177e4 LT |
269 | } |
270 | } | |
271 | } | |
272 | ||
273 | /* 3. Tuning rcvbuf, when connection enters established state. */ | |
274 | ||
275 | static void tcp_fixup_rcvbuf(struct sock *sk) | |
276 | { | |
277 | struct tcp_sock *tp = tcp_sk(sk); | |
278 | int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); | |
279 | ||
280 | /* Try to select rcvbuf so that 4 mss-sized segments | |
caa20d9a | 281 | * will fit to window and corresponding skbs will fit to our rcvbuf. |
1da177e4 LT |
282 | * (was 3; 4 is minimum to allow fast retransmit to work.) |
283 | */ | |
284 | while (tcp_win_from_space(rcvmem) < tp->advmss) | |
285 | rcvmem += 128; | |
286 | if (sk->sk_rcvbuf < 4 * rcvmem) | |
287 | sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]); | |
288 | } | |
289 | ||
caa20d9a | 290 | /* 4. Try to fixup all. It is made immediately after connection enters |
1da177e4 LT |
291 | * established state. |
292 | */ | |
293 | static void tcp_init_buffer_space(struct sock *sk) | |
294 | { | |
295 | struct tcp_sock *tp = tcp_sk(sk); | |
296 | int maxwin; | |
297 | ||
298 | if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) | |
299 | tcp_fixup_rcvbuf(sk); | |
300 | if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) | |
301 | tcp_fixup_sndbuf(sk); | |
302 | ||
303 | tp->rcvq_space.space = tp->rcv_wnd; | |
304 | ||
305 | maxwin = tcp_full_space(sk); | |
306 | ||
307 | if (tp->window_clamp >= maxwin) { | |
308 | tp->window_clamp = maxwin; | |
309 | ||
310 | if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) | |
311 | tp->window_clamp = max(maxwin - | |
312 | (maxwin >> sysctl_tcp_app_win), | |
313 | 4 * tp->advmss); | |
314 | } | |
315 | ||
316 | /* Force reservation of one segment. */ | |
317 | if (sysctl_tcp_app_win && | |
318 | tp->window_clamp > 2 * tp->advmss && | |
319 | tp->window_clamp + tp->advmss > maxwin) | |
320 | tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); | |
321 | ||
322 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); | |
323 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
324 | } | |
325 | ||
1da177e4 LT |
326 | /* 5. Recalculate window clamp after socket hit its memory bounds. */ |
327 | static void tcp_clamp_window(struct sock *sk, struct tcp_sock *tp) | |
328 | { | |
6687e988 | 329 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 330 | |
6687e988 | 331 | icsk->icsk_ack.quick = 0; |
1da177e4 | 332 | |
326f36e9 JH |
333 | if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && |
334 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && | |
335 | !tcp_memory_pressure && | |
336 | atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) { | |
337 | sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), | |
338 | sysctl_tcp_rmem[2]); | |
1da177e4 | 339 | } |
326f36e9 | 340 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) |
1da177e4 | 341 | tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss); |
1da177e4 LT |
342 | } |
343 | ||
40efc6fa SH |
344 | |
345 | /* Initialize RCV_MSS value. | |
346 | * RCV_MSS is an our guess about MSS used by the peer. | |
347 | * We haven't any direct information about the MSS. | |
348 | * It's better to underestimate the RCV_MSS rather than overestimate. | |
349 | * Overestimations make us ACKing less frequently than needed. | |
350 | * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). | |
351 | */ | |
352 | void tcp_initialize_rcv_mss(struct sock *sk) | |
353 | { | |
354 | struct tcp_sock *tp = tcp_sk(sk); | |
355 | unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); | |
356 | ||
357 | hint = min(hint, tp->rcv_wnd/2); | |
358 | hint = min(hint, TCP_MIN_RCVMSS); | |
359 | hint = max(hint, TCP_MIN_MSS); | |
360 | ||
361 | inet_csk(sk)->icsk_ack.rcv_mss = hint; | |
362 | } | |
363 | ||
1da177e4 LT |
364 | /* Receiver "autotuning" code. |
365 | * | |
366 | * The algorithm for RTT estimation w/o timestamps is based on | |
367 | * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. | |
368 | * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps> | |
369 | * | |
370 | * More detail on this code can be found at | |
371 | * <http://www.psc.edu/~jheffner/senior_thesis.ps>, | |
372 | * though this reference is out of date. A new paper | |
373 | * is pending. | |
374 | */ | |
375 | static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) | |
376 | { | |
377 | u32 new_sample = tp->rcv_rtt_est.rtt; | |
378 | long m = sample; | |
379 | ||
380 | if (m == 0) | |
381 | m = 1; | |
382 | ||
383 | if (new_sample != 0) { | |
384 | /* If we sample in larger samples in the non-timestamp | |
385 | * case, we could grossly overestimate the RTT especially | |
386 | * with chatty applications or bulk transfer apps which | |
387 | * are stalled on filesystem I/O. | |
388 | * | |
389 | * Also, since we are only going for a minimum in the | |
31f34269 | 390 | * non-timestamp case, we do not smooth things out |
caa20d9a | 391 | * else with timestamps disabled convergence takes too |
1da177e4 LT |
392 | * long. |
393 | */ | |
394 | if (!win_dep) { | |
395 | m -= (new_sample >> 3); | |
396 | new_sample += m; | |
397 | } else if (m < new_sample) | |
398 | new_sample = m << 3; | |
399 | } else { | |
caa20d9a | 400 | /* No previous measure. */ |
1da177e4 LT |
401 | new_sample = m << 3; |
402 | } | |
403 | ||
404 | if (tp->rcv_rtt_est.rtt != new_sample) | |
405 | tp->rcv_rtt_est.rtt = new_sample; | |
406 | } | |
407 | ||
408 | static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) | |
409 | { | |
410 | if (tp->rcv_rtt_est.time == 0) | |
411 | goto new_measure; | |
412 | if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) | |
413 | return; | |
414 | tcp_rcv_rtt_update(tp, | |
415 | jiffies - tp->rcv_rtt_est.time, | |
416 | 1); | |
417 | ||
418 | new_measure: | |
419 | tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; | |
420 | tp->rcv_rtt_est.time = tcp_time_stamp; | |
421 | } | |
422 | ||
463c84b9 | 423 | static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 424 | { |
463c84b9 | 425 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
426 | if (tp->rx_opt.rcv_tsecr && |
427 | (TCP_SKB_CB(skb)->end_seq - | |
463c84b9 | 428 | TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) |
1da177e4 LT |
429 | tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); |
430 | } | |
431 | ||
432 | /* | |
433 | * This function should be called every time data is copied to user space. | |
434 | * It calculates the appropriate TCP receive buffer space. | |
435 | */ | |
436 | void tcp_rcv_space_adjust(struct sock *sk) | |
437 | { | |
438 | struct tcp_sock *tp = tcp_sk(sk); | |
439 | int time; | |
440 | int space; | |
441 | ||
442 | if (tp->rcvq_space.time == 0) | |
443 | goto new_measure; | |
444 | ||
445 | time = tcp_time_stamp - tp->rcvq_space.time; | |
446 | if (time < (tp->rcv_rtt_est.rtt >> 3) || | |
447 | tp->rcv_rtt_est.rtt == 0) | |
448 | return; | |
449 | ||
450 | space = 2 * (tp->copied_seq - tp->rcvq_space.seq); | |
451 | ||
452 | space = max(tp->rcvq_space.space, space); | |
453 | ||
454 | if (tp->rcvq_space.space != space) { | |
455 | int rcvmem; | |
456 | ||
457 | tp->rcvq_space.space = space; | |
458 | ||
6fcf9412 JH |
459 | if (sysctl_tcp_moderate_rcvbuf && |
460 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { | |
1da177e4 LT |
461 | int new_clamp = space; |
462 | ||
463 | /* Receive space grows, normalize in order to | |
464 | * take into account packet headers and sk_buff | |
465 | * structure overhead. | |
466 | */ | |
467 | space /= tp->advmss; | |
468 | if (!space) | |
469 | space = 1; | |
470 | rcvmem = (tp->advmss + MAX_TCP_HEADER + | |
471 | 16 + sizeof(struct sk_buff)); | |
472 | while (tcp_win_from_space(rcvmem) < tp->advmss) | |
473 | rcvmem += 128; | |
474 | space *= rcvmem; | |
475 | space = min(space, sysctl_tcp_rmem[2]); | |
476 | if (space > sk->sk_rcvbuf) { | |
477 | sk->sk_rcvbuf = space; | |
478 | ||
479 | /* Make the window clamp follow along. */ | |
480 | tp->window_clamp = new_clamp; | |
481 | } | |
482 | } | |
483 | } | |
484 | ||
485 | new_measure: | |
486 | tp->rcvq_space.seq = tp->copied_seq; | |
487 | tp->rcvq_space.time = tcp_time_stamp; | |
488 | } | |
489 | ||
490 | /* There is something which you must keep in mind when you analyze the | |
491 | * behavior of the tp->ato delayed ack timeout interval. When a | |
492 | * connection starts up, we want to ack as quickly as possible. The | |
493 | * problem is that "good" TCP's do slow start at the beginning of data | |
494 | * transmission. The means that until we send the first few ACK's the | |
495 | * sender will sit on his end and only queue most of his data, because | |
496 | * he can only send snd_cwnd unacked packets at any given time. For | |
497 | * each ACK we send, he increments snd_cwnd and transmits more of his | |
498 | * queue. -DaveM | |
499 | */ | |
500 | static void tcp_event_data_recv(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb) | |
501 | { | |
463c84b9 | 502 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
503 | u32 now; |
504 | ||
463c84b9 | 505 | inet_csk_schedule_ack(sk); |
1da177e4 | 506 | |
463c84b9 | 507 | tcp_measure_rcv_mss(sk, skb); |
1da177e4 LT |
508 | |
509 | tcp_rcv_rtt_measure(tp); | |
510 | ||
511 | now = tcp_time_stamp; | |
512 | ||
463c84b9 | 513 | if (!icsk->icsk_ack.ato) { |
1da177e4 LT |
514 | /* The _first_ data packet received, initialize |
515 | * delayed ACK engine. | |
516 | */ | |
463c84b9 ACM |
517 | tcp_incr_quickack(sk); |
518 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 | 519 | } else { |
463c84b9 | 520 | int m = now - icsk->icsk_ack.lrcvtime; |
1da177e4 LT |
521 | |
522 | if (m <= TCP_ATO_MIN/2) { | |
523 | /* The fastest case is the first. */ | |
463c84b9 ACM |
524 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; |
525 | } else if (m < icsk->icsk_ack.ato) { | |
526 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; | |
527 | if (icsk->icsk_ack.ato > icsk->icsk_rto) | |
528 | icsk->icsk_ack.ato = icsk->icsk_rto; | |
529 | } else if (m > icsk->icsk_rto) { | |
caa20d9a | 530 | /* Too long gap. Apparently sender failed to |
1da177e4 LT |
531 | * restart window, so that we send ACKs quickly. |
532 | */ | |
463c84b9 | 533 | tcp_incr_quickack(sk); |
1da177e4 LT |
534 | sk_stream_mem_reclaim(sk); |
535 | } | |
536 | } | |
463c84b9 | 537 | icsk->icsk_ack.lrcvtime = now; |
1da177e4 LT |
538 | |
539 | TCP_ECN_check_ce(tp, skb); | |
540 | ||
541 | if (skb->len >= 128) | |
542 | tcp_grow_window(sk, tp, skb); | |
543 | } | |
544 | ||
1da177e4 LT |
545 | /* Called to compute a smoothed rtt estimate. The data fed to this |
546 | * routine either comes from timestamps, or from segments that were | |
547 | * known _not_ to have been retransmitted [see Karn/Partridge | |
548 | * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 | |
549 | * piece by Van Jacobson. | |
550 | * NOTE: the next three routines used to be one big routine. | |
551 | * To save cycles in the RFC 1323 implementation it was better to break | |
552 | * it up into three procedures. -- erics | |
553 | */ | |
2d2abbab | 554 | static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) |
1da177e4 | 555 | { |
6687e988 | 556 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
557 | long m = mrtt; /* RTT */ |
558 | ||
1da177e4 LT |
559 | /* The following amusing code comes from Jacobson's |
560 | * article in SIGCOMM '88. Note that rtt and mdev | |
561 | * are scaled versions of rtt and mean deviation. | |
562 | * This is designed to be as fast as possible | |
563 | * m stands for "measurement". | |
564 | * | |
565 | * On a 1990 paper the rto value is changed to: | |
566 | * RTO = rtt + 4 * mdev | |
567 | * | |
568 | * Funny. This algorithm seems to be very broken. | |
569 | * These formulae increase RTO, when it should be decreased, increase | |
31f34269 | 570 | * too slowly, when it should be increased quickly, decrease too quickly |
1da177e4 LT |
571 | * etc. I guess in BSD RTO takes ONE value, so that it is absolutely |
572 | * does not matter how to _calculate_ it. Seems, it was trap | |
573 | * that VJ failed to avoid. 8) | |
574 | */ | |
575 | if(m == 0) | |
576 | m = 1; | |
577 | if (tp->srtt != 0) { | |
578 | m -= (tp->srtt >> 3); /* m is now error in rtt est */ | |
579 | tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ | |
580 | if (m < 0) { | |
581 | m = -m; /* m is now abs(error) */ | |
582 | m -= (tp->mdev >> 2); /* similar update on mdev */ | |
583 | /* This is similar to one of Eifel findings. | |
584 | * Eifel blocks mdev updates when rtt decreases. | |
585 | * This solution is a bit different: we use finer gain | |
586 | * for mdev in this case (alpha*beta). | |
587 | * Like Eifel it also prevents growth of rto, | |
588 | * but also it limits too fast rto decreases, | |
589 | * happening in pure Eifel. | |
590 | */ | |
591 | if (m > 0) | |
592 | m >>= 3; | |
593 | } else { | |
594 | m -= (tp->mdev >> 2); /* similar update on mdev */ | |
595 | } | |
596 | tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ | |
597 | if (tp->mdev > tp->mdev_max) { | |
598 | tp->mdev_max = tp->mdev; | |
599 | if (tp->mdev_max > tp->rttvar) | |
600 | tp->rttvar = tp->mdev_max; | |
601 | } | |
602 | if (after(tp->snd_una, tp->rtt_seq)) { | |
603 | if (tp->mdev_max < tp->rttvar) | |
604 | tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2; | |
605 | tp->rtt_seq = tp->snd_nxt; | |
606 | tp->mdev_max = TCP_RTO_MIN; | |
607 | } | |
608 | } else { | |
609 | /* no previous measure. */ | |
610 | tp->srtt = m<<3; /* take the measured time to be rtt */ | |
611 | tp->mdev = m<<1; /* make sure rto = 3*rtt */ | |
612 | tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN); | |
613 | tp->rtt_seq = tp->snd_nxt; | |
614 | } | |
1da177e4 LT |
615 | } |
616 | ||
617 | /* Calculate rto without backoff. This is the second half of Van Jacobson's | |
618 | * routine referred to above. | |
619 | */ | |
463c84b9 | 620 | static inline void tcp_set_rto(struct sock *sk) |
1da177e4 | 621 | { |
463c84b9 | 622 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
623 | /* Old crap is replaced with new one. 8) |
624 | * | |
625 | * More seriously: | |
626 | * 1. If rtt variance happened to be less 50msec, it is hallucination. | |
627 | * It cannot be less due to utterly erratic ACK generation made | |
628 | * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ | |
629 | * to do with delayed acks, because at cwnd>2 true delack timeout | |
630 | * is invisible. Actually, Linux-2.4 also generates erratic | |
caa20d9a | 631 | * ACKs in some circumstances. |
1da177e4 | 632 | */ |
463c84b9 | 633 | inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar; |
1da177e4 LT |
634 | |
635 | /* 2. Fixups made earlier cannot be right. | |
636 | * If we do not estimate RTO correctly without them, | |
637 | * all the algo is pure shit and should be replaced | |
caa20d9a | 638 | * with correct one. It is exactly, which we pretend to do. |
1da177e4 LT |
639 | */ |
640 | } | |
641 | ||
642 | /* NOTE: clamping at TCP_RTO_MIN is not required, current algo | |
643 | * guarantees that rto is higher. | |
644 | */ | |
463c84b9 | 645 | static inline void tcp_bound_rto(struct sock *sk) |
1da177e4 | 646 | { |
463c84b9 ACM |
647 | if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX) |
648 | inet_csk(sk)->icsk_rto = TCP_RTO_MAX; | |
1da177e4 LT |
649 | } |
650 | ||
651 | /* Save metrics learned by this TCP session. | |
652 | This function is called only, when TCP finishes successfully | |
653 | i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. | |
654 | */ | |
655 | void tcp_update_metrics(struct sock *sk) | |
656 | { | |
657 | struct tcp_sock *tp = tcp_sk(sk); | |
658 | struct dst_entry *dst = __sk_dst_get(sk); | |
659 | ||
660 | if (sysctl_tcp_nometrics_save) | |
661 | return; | |
662 | ||
663 | dst_confirm(dst); | |
664 | ||
665 | if (dst && (dst->flags&DST_HOST)) { | |
6687e988 | 666 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
667 | int m; |
668 | ||
6687e988 | 669 | if (icsk->icsk_backoff || !tp->srtt) { |
1da177e4 LT |
670 | /* This session failed to estimate rtt. Why? |
671 | * Probably, no packets returned in time. | |
672 | * Reset our results. | |
673 | */ | |
674 | if (!(dst_metric_locked(dst, RTAX_RTT))) | |
675 | dst->metrics[RTAX_RTT-1] = 0; | |
676 | return; | |
677 | } | |
678 | ||
679 | m = dst_metric(dst, RTAX_RTT) - tp->srtt; | |
680 | ||
681 | /* If newly calculated rtt larger than stored one, | |
682 | * store new one. Otherwise, use EWMA. Remember, | |
683 | * rtt overestimation is always better than underestimation. | |
684 | */ | |
685 | if (!(dst_metric_locked(dst, RTAX_RTT))) { | |
686 | if (m <= 0) | |
687 | dst->metrics[RTAX_RTT-1] = tp->srtt; | |
688 | else | |
689 | dst->metrics[RTAX_RTT-1] -= (m>>3); | |
690 | } | |
691 | ||
692 | if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { | |
693 | if (m < 0) | |
694 | m = -m; | |
695 | ||
696 | /* Scale deviation to rttvar fixed point */ | |
697 | m >>= 1; | |
698 | if (m < tp->mdev) | |
699 | m = tp->mdev; | |
700 | ||
701 | if (m >= dst_metric(dst, RTAX_RTTVAR)) | |
702 | dst->metrics[RTAX_RTTVAR-1] = m; | |
703 | else | |
704 | dst->metrics[RTAX_RTTVAR-1] -= | |
705 | (dst->metrics[RTAX_RTTVAR-1] - m)>>2; | |
706 | } | |
707 | ||
708 | if (tp->snd_ssthresh >= 0xFFFF) { | |
709 | /* Slow start still did not finish. */ | |
710 | if (dst_metric(dst, RTAX_SSTHRESH) && | |
711 | !dst_metric_locked(dst, RTAX_SSTHRESH) && | |
712 | (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) | |
713 | dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1; | |
714 | if (!dst_metric_locked(dst, RTAX_CWND) && | |
715 | tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) | |
716 | dst->metrics[RTAX_CWND-1] = tp->snd_cwnd; | |
717 | } else if (tp->snd_cwnd > tp->snd_ssthresh && | |
6687e988 | 718 | icsk->icsk_ca_state == TCP_CA_Open) { |
1da177e4 LT |
719 | /* Cong. avoidance phase, cwnd is reliable. */ |
720 | if (!dst_metric_locked(dst, RTAX_SSTHRESH)) | |
721 | dst->metrics[RTAX_SSTHRESH-1] = | |
722 | max(tp->snd_cwnd >> 1, tp->snd_ssthresh); | |
723 | if (!dst_metric_locked(dst, RTAX_CWND)) | |
724 | dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1; | |
725 | } else { | |
726 | /* Else slow start did not finish, cwnd is non-sense, | |
727 | ssthresh may be also invalid. | |
728 | */ | |
729 | if (!dst_metric_locked(dst, RTAX_CWND)) | |
730 | dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1; | |
731 | if (dst->metrics[RTAX_SSTHRESH-1] && | |
732 | !dst_metric_locked(dst, RTAX_SSTHRESH) && | |
733 | tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1]) | |
734 | dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh; | |
735 | } | |
736 | ||
737 | if (!dst_metric_locked(dst, RTAX_REORDERING)) { | |
738 | if (dst->metrics[RTAX_REORDERING-1] < tp->reordering && | |
739 | tp->reordering != sysctl_tcp_reordering) | |
740 | dst->metrics[RTAX_REORDERING-1] = tp->reordering; | |
741 | } | |
742 | } | |
743 | } | |
744 | ||
745 | /* Numbers are taken from RFC2414. */ | |
746 | __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst) | |
747 | { | |
748 | __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); | |
749 | ||
750 | if (!cwnd) { | |
c1b4a7e6 | 751 | if (tp->mss_cache > 1460) |
1da177e4 LT |
752 | cwnd = 2; |
753 | else | |
c1b4a7e6 | 754 | cwnd = (tp->mss_cache > 1095) ? 3 : 4; |
1da177e4 LT |
755 | } |
756 | return min_t(__u32, cwnd, tp->snd_cwnd_clamp); | |
757 | } | |
758 | ||
40efc6fa SH |
759 | /* Set slow start threshold and cwnd not falling to slow start */ |
760 | void tcp_enter_cwr(struct sock *sk) | |
761 | { | |
762 | struct tcp_sock *tp = tcp_sk(sk); | |
763 | ||
764 | tp->prior_ssthresh = 0; | |
765 | tp->bytes_acked = 0; | |
766 | if (inet_csk(sk)->icsk_ca_state < TCP_CA_CWR) { | |
767 | tp->undo_marker = 0; | |
768 | tp->snd_ssthresh = inet_csk(sk)->icsk_ca_ops->ssthresh(sk); | |
769 | tp->snd_cwnd = min(tp->snd_cwnd, | |
770 | tcp_packets_in_flight(tp) + 1U); | |
771 | tp->snd_cwnd_cnt = 0; | |
772 | tp->high_seq = tp->snd_nxt; | |
773 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
774 | TCP_ECN_queue_cwr(tp); | |
775 | ||
776 | tcp_set_ca_state(sk, TCP_CA_CWR); | |
777 | } | |
778 | } | |
779 | ||
1da177e4 LT |
780 | /* Initialize metrics on socket. */ |
781 | ||
782 | static void tcp_init_metrics(struct sock *sk) | |
783 | { | |
784 | struct tcp_sock *tp = tcp_sk(sk); | |
785 | struct dst_entry *dst = __sk_dst_get(sk); | |
786 | ||
787 | if (dst == NULL) | |
788 | goto reset; | |
789 | ||
790 | dst_confirm(dst); | |
791 | ||
792 | if (dst_metric_locked(dst, RTAX_CWND)) | |
793 | tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); | |
794 | if (dst_metric(dst, RTAX_SSTHRESH)) { | |
795 | tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); | |
796 | if (tp->snd_ssthresh > tp->snd_cwnd_clamp) | |
797 | tp->snd_ssthresh = tp->snd_cwnd_clamp; | |
798 | } | |
799 | if (dst_metric(dst, RTAX_REORDERING) && | |
800 | tp->reordering != dst_metric(dst, RTAX_REORDERING)) { | |
801 | tp->rx_opt.sack_ok &= ~2; | |
802 | tp->reordering = dst_metric(dst, RTAX_REORDERING); | |
803 | } | |
804 | ||
805 | if (dst_metric(dst, RTAX_RTT) == 0) | |
806 | goto reset; | |
807 | ||
808 | if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3)) | |
809 | goto reset; | |
810 | ||
811 | /* Initial rtt is determined from SYN,SYN-ACK. | |
812 | * The segment is small and rtt may appear much | |
813 | * less than real one. Use per-dst memory | |
814 | * to make it more realistic. | |
815 | * | |
816 | * A bit of theory. RTT is time passed after "normal" sized packet | |
caa20d9a | 817 | * is sent until it is ACKed. In normal circumstances sending small |
1da177e4 LT |
818 | * packets force peer to delay ACKs and calculation is correct too. |
819 | * The algorithm is adaptive and, provided we follow specs, it | |
820 | * NEVER underestimate RTT. BUT! If peer tries to make some clever | |
821 | * tricks sort of "quick acks" for time long enough to decrease RTT | |
822 | * to low value, and then abruptly stops to do it and starts to delay | |
823 | * ACKs, wait for troubles. | |
824 | */ | |
825 | if (dst_metric(dst, RTAX_RTT) > tp->srtt) { | |
826 | tp->srtt = dst_metric(dst, RTAX_RTT); | |
827 | tp->rtt_seq = tp->snd_nxt; | |
828 | } | |
829 | if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) { | |
830 | tp->mdev = dst_metric(dst, RTAX_RTTVAR); | |
831 | tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN); | |
832 | } | |
463c84b9 ACM |
833 | tcp_set_rto(sk); |
834 | tcp_bound_rto(sk); | |
835 | if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp) | |
1da177e4 LT |
836 | goto reset; |
837 | tp->snd_cwnd = tcp_init_cwnd(tp, dst); | |
838 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
839 | return; | |
840 | ||
841 | reset: | |
842 | /* Play conservative. If timestamps are not | |
843 | * supported, TCP will fail to recalculate correct | |
844 | * rtt, if initial rto is too small. FORGET ALL AND RESET! | |
845 | */ | |
846 | if (!tp->rx_opt.saw_tstamp && tp->srtt) { | |
847 | tp->srtt = 0; | |
848 | tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT; | |
463c84b9 | 849 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
1da177e4 LT |
850 | } |
851 | } | |
852 | ||
6687e988 ACM |
853 | static void tcp_update_reordering(struct sock *sk, const int metric, |
854 | const int ts) | |
1da177e4 | 855 | { |
6687e988 | 856 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
857 | if (metric > tp->reordering) { |
858 | tp->reordering = min(TCP_MAX_REORDERING, metric); | |
859 | ||
860 | /* This exciting event is worth to be remembered. 8) */ | |
861 | if (ts) | |
862 | NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER); | |
863 | else if (IsReno(tp)) | |
864 | NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER); | |
865 | else if (IsFack(tp)) | |
866 | NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER); | |
867 | else | |
868 | NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER); | |
869 | #if FASTRETRANS_DEBUG > 1 | |
870 | printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", | |
6687e988 | 871 | tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, |
1da177e4 LT |
872 | tp->reordering, |
873 | tp->fackets_out, | |
874 | tp->sacked_out, | |
875 | tp->undo_marker ? tp->undo_retrans : 0); | |
876 | #endif | |
877 | /* Disable FACK yet. */ | |
878 | tp->rx_opt.sack_ok &= ~2; | |
879 | } | |
880 | } | |
881 | ||
882 | /* This procedure tags the retransmission queue when SACKs arrive. | |
883 | * | |
884 | * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). | |
885 | * Packets in queue with these bits set are counted in variables | |
886 | * sacked_out, retrans_out and lost_out, correspondingly. | |
887 | * | |
888 | * Valid combinations are: | |
889 | * Tag InFlight Description | |
890 | * 0 1 - orig segment is in flight. | |
891 | * S 0 - nothing flies, orig reached receiver. | |
892 | * L 0 - nothing flies, orig lost by net. | |
893 | * R 2 - both orig and retransmit are in flight. | |
894 | * L|R 1 - orig is lost, retransmit is in flight. | |
895 | * S|R 1 - orig reached receiver, retrans is still in flight. | |
896 | * (L|S|R is logically valid, it could occur when L|R is sacked, | |
897 | * but it is equivalent to plain S and code short-curcuits it to S. | |
898 | * L|S is logically invalid, it would mean -1 packet in flight 8)) | |
899 | * | |
900 | * These 6 states form finite state machine, controlled by the following events: | |
901 | * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) | |
902 | * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) | |
903 | * 3. Loss detection event of one of three flavors: | |
904 | * A. Scoreboard estimator decided the packet is lost. | |
905 | * A'. Reno "three dupacks" marks head of queue lost. | |
906 | * A''. Its FACK modfication, head until snd.fack is lost. | |
907 | * B. SACK arrives sacking data transmitted after never retransmitted | |
908 | * hole was sent out. | |
909 | * C. SACK arrives sacking SND.NXT at the moment, when the | |
910 | * segment was retransmitted. | |
911 | * 4. D-SACK added new rule: D-SACK changes any tag to S. | |
912 | * | |
913 | * It is pleasant to note, that state diagram turns out to be commutative, | |
914 | * so that we are allowed not to be bothered by order of our actions, | |
915 | * when multiple events arrive simultaneously. (see the function below). | |
916 | * | |
917 | * Reordering detection. | |
918 | * -------------------- | |
919 | * Reordering metric is maximal distance, which a packet can be displaced | |
920 | * in packet stream. With SACKs we can estimate it: | |
921 | * | |
922 | * 1. SACK fills old hole and the corresponding segment was not | |
923 | * ever retransmitted -> reordering. Alas, we cannot use it | |
924 | * when segment was retransmitted. | |
925 | * 2. The last flaw is solved with D-SACK. D-SACK arrives | |
926 | * for retransmitted and already SACKed segment -> reordering.. | |
927 | * Both of these heuristics are not used in Loss state, when we cannot | |
928 | * account for retransmits accurately. | |
929 | */ | |
930 | static int | |
931 | tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una) | |
932 | { | |
6687e988 | 933 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
934 | struct tcp_sock *tp = tcp_sk(sk); |
935 | unsigned char *ptr = ack_skb->h.raw + TCP_SKB_CB(ack_skb)->sacked; | |
936 | struct tcp_sack_block *sp = (struct tcp_sack_block *)(ptr+2); | |
937 | int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3; | |
938 | int reord = tp->packets_out; | |
939 | int prior_fackets; | |
940 | u32 lost_retrans = 0; | |
941 | int flag = 0; | |
6a438bbe | 942 | int dup_sack = 0; |
1da177e4 LT |
943 | int i; |
944 | ||
1da177e4 LT |
945 | if (!tp->sacked_out) |
946 | tp->fackets_out = 0; | |
947 | prior_fackets = tp->fackets_out; | |
948 | ||
6a438bbe SH |
949 | /* SACK fastpath: |
950 | * if the only SACK change is the increase of the end_seq of | |
951 | * the first block then only apply that SACK block | |
952 | * and use retrans queue hinting otherwise slowpath */ | |
953 | flag = 1; | |
954 | for (i = 0; i< num_sacks; i++) { | |
955 | __u32 start_seq = ntohl(sp[i].start_seq); | |
956 | __u32 end_seq = ntohl(sp[i].end_seq); | |
957 | ||
958 | if (i == 0){ | |
959 | if (tp->recv_sack_cache[i].start_seq != start_seq) | |
960 | flag = 0; | |
961 | } else { | |
962 | if ((tp->recv_sack_cache[i].start_seq != start_seq) || | |
963 | (tp->recv_sack_cache[i].end_seq != end_seq)) | |
964 | flag = 0; | |
965 | } | |
966 | tp->recv_sack_cache[i].start_seq = start_seq; | |
967 | tp->recv_sack_cache[i].end_seq = end_seq; | |
1da177e4 LT |
968 | |
969 | /* Check for D-SACK. */ | |
970 | if (i == 0) { | |
971 | u32 ack = TCP_SKB_CB(ack_skb)->ack_seq; | |
972 | ||
973 | if (before(start_seq, ack)) { | |
974 | dup_sack = 1; | |
975 | tp->rx_opt.sack_ok |= 4; | |
976 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV); | |
977 | } else if (num_sacks > 1 && | |
978 | !after(end_seq, ntohl(sp[1].end_seq)) && | |
979 | !before(start_seq, ntohl(sp[1].start_seq))) { | |
980 | dup_sack = 1; | |
981 | tp->rx_opt.sack_ok |= 4; | |
982 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV); | |
983 | } | |
984 | ||
985 | /* D-SACK for already forgotten data... | |
986 | * Do dumb counting. */ | |
987 | if (dup_sack && | |
988 | !after(end_seq, prior_snd_una) && | |
989 | after(end_seq, tp->undo_marker)) | |
990 | tp->undo_retrans--; | |
991 | ||
992 | /* Eliminate too old ACKs, but take into | |
993 | * account more or less fresh ones, they can | |
994 | * contain valid SACK info. | |
995 | */ | |
996 | if (before(ack, prior_snd_una - tp->max_window)) | |
997 | return 0; | |
998 | } | |
6a438bbe SH |
999 | } |
1000 | ||
1001 | if (flag) | |
1002 | num_sacks = 1; | |
1003 | else { | |
1004 | int j; | |
1005 | tp->fastpath_skb_hint = NULL; | |
1006 | ||
1007 | /* order SACK blocks to allow in order walk of the retrans queue */ | |
1008 | for (i = num_sacks-1; i > 0; i--) { | |
1009 | for (j = 0; j < i; j++){ | |
1010 | if (after(ntohl(sp[j].start_seq), | |
1011 | ntohl(sp[j+1].start_seq))){ | |
1012 | sp[j].start_seq = htonl(tp->recv_sack_cache[j+1].start_seq); | |
1013 | sp[j].end_seq = htonl(tp->recv_sack_cache[j+1].end_seq); | |
1014 | sp[j+1].start_seq = htonl(tp->recv_sack_cache[j].start_seq); | |
1015 | sp[j+1].end_seq = htonl(tp->recv_sack_cache[j].end_seq); | |
1016 | } | |
1017 | ||
1018 | } | |
1019 | } | |
1020 | } | |
1021 | ||
1022 | /* clear flag as used for different purpose in following code */ | |
1023 | flag = 0; | |
1024 | ||
1025 | for (i=0; i<num_sacks; i++, sp++) { | |
1026 | struct sk_buff *skb; | |
1027 | __u32 start_seq = ntohl(sp->start_seq); | |
1028 | __u32 end_seq = ntohl(sp->end_seq); | |
1029 | int fack_count; | |
1030 | ||
1031 | /* Use SACK fastpath hint if valid */ | |
1032 | if (tp->fastpath_skb_hint) { | |
1033 | skb = tp->fastpath_skb_hint; | |
1034 | fack_count = tp->fastpath_cnt_hint; | |
1035 | } else { | |
1036 | skb = sk->sk_write_queue.next; | |
1037 | fack_count = 0; | |
1038 | } | |
1da177e4 LT |
1039 | |
1040 | /* Event "B" in the comment above. */ | |
1041 | if (after(end_seq, tp->high_seq)) | |
1042 | flag |= FLAG_DATA_LOST; | |
1043 | ||
6a438bbe | 1044 | sk_stream_for_retrans_queue_from(skb, sk) { |
6475be16 DM |
1045 | int in_sack, pcount; |
1046 | u8 sacked; | |
1da177e4 | 1047 | |
6a438bbe SH |
1048 | tp->fastpath_skb_hint = skb; |
1049 | tp->fastpath_cnt_hint = fack_count; | |
1050 | ||
1da177e4 LT |
1051 | /* The retransmission queue is always in order, so |
1052 | * we can short-circuit the walk early. | |
1053 | */ | |
6475be16 | 1054 | if (!before(TCP_SKB_CB(skb)->seq, end_seq)) |
1da177e4 LT |
1055 | break; |
1056 | ||
3c05d92e HX |
1057 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && |
1058 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1059 | ||
6475be16 DM |
1060 | pcount = tcp_skb_pcount(skb); |
1061 | ||
3c05d92e HX |
1062 | if (pcount > 1 && !in_sack && |
1063 | after(TCP_SKB_CB(skb)->end_seq, start_seq)) { | |
6475be16 DM |
1064 | unsigned int pkt_len; |
1065 | ||
3c05d92e HX |
1066 | in_sack = !after(start_seq, |
1067 | TCP_SKB_CB(skb)->seq); | |
1068 | ||
1069 | if (!in_sack) | |
6475be16 DM |
1070 | pkt_len = (start_seq - |
1071 | TCP_SKB_CB(skb)->seq); | |
1072 | else | |
1073 | pkt_len = (end_seq - | |
1074 | TCP_SKB_CB(skb)->seq); | |
7967168c | 1075 | if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size)) |
6475be16 DM |
1076 | break; |
1077 | pcount = tcp_skb_pcount(skb); | |
1078 | } | |
1079 | ||
1080 | fack_count += pcount; | |
1da177e4 | 1081 | |
6475be16 DM |
1082 | sacked = TCP_SKB_CB(skb)->sacked; |
1083 | ||
1da177e4 LT |
1084 | /* Account D-SACK for retransmitted packet. */ |
1085 | if ((dup_sack && in_sack) && | |
1086 | (sacked & TCPCB_RETRANS) && | |
1087 | after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker)) | |
1088 | tp->undo_retrans--; | |
1089 | ||
1090 | /* The frame is ACKed. */ | |
1091 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) { | |
1092 | if (sacked&TCPCB_RETRANS) { | |
1093 | if ((dup_sack && in_sack) && | |
1094 | (sacked&TCPCB_SACKED_ACKED)) | |
1095 | reord = min(fack_count, reord); | |
1096 | } else { | |
1097 | /* If it was in a hole, we detected reordering. */ | |
1098 | if (fack_count < prior_fackets && | |
1099 | !(sacked&TCPCB_SACKED_ACKED)) | |
1100 | reord = min(fack_count, reord); | |
1101 | } | |
1102 | ||
1103 | /* Nothing to do; acked frame is about to be dropped. */ | |
1104 | continue; | |
1105 | } | |
1106 | ||
1107 | if ((sacked&TCPCB_SACKED_RETRANS) && | |
1108 | after(end_seq, TCP_SKB_CB(skb)->ack_seq) && | |
1109 | (!lost_retrans || after(end_seq, lost_retrans))) | |
1110 | lost_retrans = end_seq; | |
1111 | ||
1112 | if (!in_sack) | |
1113 | continue; | |
1114 | ||
1115 | if (!(sacked&TCPCB_SACKED_ACKED)) { | |
1116 | if (sacked & TCPCB_SACKED_RETRANS) { | |
1117 | /* If the segment is not tagged as lost, | |
1118 | * we do not clear RETRANS, believing | |
1119 | * that retransmission is still in flight. | |
1120 | */ | |
1121 | if (sacked & TCPCB_LOST) { | |
1122 | TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); | |
1123 | tp->lost_out -= tcp_skb_pcount(skb); | |
1124 | tp->retrans_out -= tcp_skb_pcount(skb); | |
6a438bbe SH |
1125 | |
1126 | /* clear lost hint */ | |
1127 | tp->retransmit_skb_hint = NULL; | |
1da177e4 LT |
1128 | } |
1129 | } else { | |
1130 | /* New sack for not retransmitted frame, | |
1131 | * which was in hole. It is reordering. | |
1132 | */ | |
1133 | if (!(sacked & TCPCB_RETRANS) && | |
1134 | fack_count < prior_fackets) | |
1135 | reord = min(fack_count, reord); | |
1136 | ||
1137 | if (sacked & TCPCB_LOST) { | |
1138 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; | |
1139 | tp->lost_out -= tcp_skb_pcount(skb); | |
6a438bbe SH |
1140 | |
1141 | /* clear lost hint */ | |
1142 | tp->retransmit_skb_hint = NULL; | |
1da177e4 LT |
1143 | } |
1144 | } | |
1145 | ||
1146 | TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED; | |
1147 | flag |= FLAG_DATA_SACKED; | |
1148 | tp->sacked_out += tcp_skb_pcount(skb); | |
1149 | ||
1150 | if (fack_count > tp->fackets_out) | |
1151 | tp->fackets_out = fack_count; | |
1152 | } else { | |
1153 | if (dup_sack && (sacked&TCPCB_RETRANS)) | |
1154 | reord = min(fack_count, reord); | |
1155 | } | |
1156 | ||
1157 | /* D-SACK. We can detect redundant retransmission | |
1158 | * in S|R and plain R frames and clear it. | |
1159 | * undo_retrans is decreased above, L|R frames | |
1160 | * are accounted above as well. | |
1161 | */ | |
1162 | if (dup_sack && | |
1163 | (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) { | |
1164 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; | |
1165 | tp->retrans_out -= tcp_skb_pcount(skb); | |
6a438bbe | 1166 | tp->retransmit_skb_hint = NULL; |
1da177e4 LT |
1167 | } |
1168 | } | |
1169 | } | |
1170 | ||
1171 | /* Check for lost retransmit. This superb idea is | |
1172 | * borrowed from "ratehalving". Event "C". | |
1173 | * Later note: FACK people cheated me again 8), | |
1174 | * we have to account for reordering! Ugly, | |
1175 | * but should help. | |
1176 | */ | |
6687e988 | 1177 | if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) { |
1da177e4 LT |
1178 | struct sk_buff *skb; |
1179 | ||
1180 | sk_stream_for_retrans_queue(skb, sk) { | |
1181 | if (after(TCP_SKB_CB(skb)->seq, lost_retrans)) | |
1182 | break; | |
1183 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1184 | continue; | |
1185 | if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) && | |
1186 | after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) && | |
1187 | (IsFack(tp) || | |
1188 | !before(lost_retrans, | |
1189 | TCP_SKB_CB(skb)->ack_seq + tp->reordering * | |
c1b4a7e6 | 1190 | tp->mss_cache))) { |
1da177e4 LT |
1191 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; |
1192 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1193 | ||
6a438bbe SH |
1194 | /* clear lost hint */ |
1195 | tp->retransmit_skb_hint = NULL; | |
1196 | ||
1da177e4 LT |
1197 | if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) { |
1198 | tp->lost_out += tcp_skb_pcount(skb); | |
1199 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1200 | flag |= FLAG_DATA_SACKED; | |
1201 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT); | |
1202 | } | |
1203 | } | |
1204 | } | |
1205 | } | |
1206 | ||
1207 | tp->left_out = tp->sacked_out + tp->lost_out; | |
1208 | ||
6687e988 ACM |
1209 | if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss) |
1210 | tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0); | |
1da177e4 LT |
1211 | |
1212 | #if FASTRETRANS_DEBUG > 0 | |
1213 | BUG_TRAP((int)tp->sacked_out >= 0); | |
1214 | BUG_TRAP((int)tp->lost_out >= 0); | |
1215 | BUG_TRAP((int)tp->retrans_out >= 0); | |
1216 | BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0); | |
1217 | #endif | |
1218 | return flag; | |
1219 | } | |
1220 | ||
1221 | /* RTO occurred, but do not yet enter loss state. Instead, transmit two new | |
1222 | * segments to see from the next ACKs whether any data was really missing. | |
1223 | * If the RTO was spurious, new ACKs should arrive. | |
1224 | */ | |
1225 | void tcp_enter_frto(struct sock *sk) | |
1226 | { | |
6687e988 | 1227 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
1228 | struct tcp_sock *tp = tcp_sk(sk); |
1229 | struct sk_buff *skb; | |
1230 | ||
1231 | tp->frto_counter = 1; | |
1232 | ||
6687e988 | 1233 | if (icsk->icsk_ca_state <= TCP_CA_Disorder || |
1da177e4 | 1234 | tp->snd_una == tp->high_seq || |
6687e988 ACM |
1235 | (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { |
1236 | tp->prior_ssthresh = tcp_current_ssthresh(sk); | |
1237 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
1238 | tcp_ca_event(sk, CA_EVENT_FRTO); | |
1da177e4 LT |
1239 | } |
1240 | ||
1241 | /* Have to clear retransmission markers here to keep the bookkeeping | |
1242 | * in shape, even though we are not yet in Loss state. | |
1243 | * If something was really lost, it is eventually caught up | |
1244 | * in tcp_enter_frto_loss. | |
1245 | */ | |
1246 | tp->retrans_out = 0; | |
1247 | tp->undo_marker = tp->snd_una; | |
1248 | tp->undo_retrans = 0; | |
1249 | ||
1250 | sk_stream_for_retrans_queue(skb, sk) { | |
1251 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_RETRANS; | |
1252 | } | |
1253 | tcp_sync_left_out(tp); | |
1254 | ||
6687e988 | 1255 | tcp_set_ca_state(sk, TCP_CA_Open); |
1da177e4 LT |
1256 | tp->frto_highmark = tp->snd_nxt; |
1257 | } | |
1258 | ||
1259 | /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, | |
1260 | * which indicates that we should follow the traditional RTO recovery, | |
1261 | * i.e. mark everything lost and do go-back-N retransmission. | |
1262 | */ | |
1263 | static void tcp_enter_frto_loss(struct sock *sk) | |
1264 | { | |
1265 | struct tcp_sock *tp = tcp_sk(sk); | |
1266 | struct sk_buff *skb; | |
1267 | int cnt = 0; | |
1268 | ||
1269 | tp->sacked_out = 0; | |
1270 | tp->lost_out = 0; | |
1271 | tp->fackets_out = 0; | |
1272 | ||
1273 | sk_stream_for_retrans_queue(skb, sk) { | |
1274 | cnt += tcp_skb_pcount(skb); | |
1275 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; | |
1276 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) { | |
1277 | ||
1278 | /* Do not mark those segments lost that were | |
1279 | * forward transmitted after RTO | |
1280 | */ | |
1281 | if (!after(TCP_SKB_CB(skb)->end_seq, | |
1282 | tp->frto_highmark)) { | |
1283 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1284 | tp->lost_out += tcp_skb_pcount(skb); | |
1285 | } | |
1286 | } else { | |
1287 | tp->sacked_out += tcp_skb_pcount(skb); | |
1288 | tp->fackets_out = cnt; | |
1289 | } | |
1290 | } | |
1291 | tcp_sync_left_out(tp); | |
1292 | ||
1293 | tp->snd_cwnd = tp->frto_counter + tcp_packets_in_flight(tp)+1; | |
1294 | tp->snd_cwnd_cnt = 0; | |
1295 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1296 | tp->undo_marker = 0; | |
1297 | tp->frto_counter = 0; | |
1298 | ||
1299 | tp->reordering = min_t(unsigned int, tp->reordering, | |
1300 | sysctl_tcp_reordering); | |
6687e988 | 1301 | tcp_set_ca_state(sk, TCP_CA_Loss); |
1da177e4 LT |
1302 | tp->high_seq = tp->frto_highmark; |
1303 | TCP_ECN_queue_cwr(tp); | |
6a438bbe SH |
1304 | |
1305 | clear_all_retrans_hints(tp); | |
1da177e4 LT |
1306 | } |
1307 | ||
1308 | void tcp_clear_retrans(struct tcp_sock *tp) | |
1309 | { | |
1310 | tp->left_out = 0; | |
1311 | tp->retrans_out = 0; | |
1312 | ||
1313 | tp->fackets_out = 0; | |
1314 | tp->sacked_out = 0; | |
1315 | tp->lost_out = 0; | |
1316 | ||
1317 | tp->undo_marker = 0; | |
1318 | tp->undo_retrans = 0; | |
1319 | } | |
1320 | ||
1321 | /* Enter Loss state. If "how" is not zero, forget all SACK information | |
1322 | * and reset tags completely, otherwise preserve SACKs. If receiver | |
1323 | * dropped its ofo queue, we will know this due to reneging detection. | |
1324 | */ | |
1325 | void tcp_enter_loss(struct sock *sk, int how) | |
1326 | { | |
6687e988 | 1327 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
1328 | struct tcp_sock *tp = tcp_sk(sk); |
1329 | struct sk_buff *skb; | |
1330 | int cnt = 0; | |
1331 | ||
1332 | /* Reduce ssthresh if it has not yet been made inside this window. */ | |
6687e988 ACM |
1333 | if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || |
1334 | (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { | |
1335 | tp->prior_ssthresh = tcp_current_ssthresh(sk); | |
1336 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
1337 | tcp_ca_event(sk, CA_EVENT_LOSS); | |
1da177e4 LT |
1338 | } |
1339 | tp->snd_cwnd = 1; | |
1340 | tp->snd_cwnd_cnt = 0; | |
1341 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1342 | ||
9772efb9 | 1343 | tp->bytes_acked = 0; |
1da177e4 LT |
1344 | tcp_clear_retrans(tp); |
1345 | ||
1346 | /* Push undo marker, if it was plain RTO and nothing | |
1347 | * was retransmitted. */ | |
1348 | if (!how) | |
1349 | tp->undo_marker = tp->snd_una; | |
1350 | ||
1351 | sk_stream_for_retrans_queue(skb, sk) { | |
1352 | cnt += tcp_skb_pcount(skb); | |
1353 | if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS) | |
1354 | tp->undo_marker = 0; | |
1355 | TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; | |
1356 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { | |
1357 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; | |
1358 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1359 | tp->lost_out += tcp_skb_pcount(skb); | |
1360 | } else { | |
1361 | tp->sacked_out += tcp_skb_pcount(skb); | |
1362 | tp->fackets_out = cnt; | |
1363 | } | |
1364 | } | |
1365 | tcp_sync_left_out(tp); | |
1366 | ||
1367 | tp->reordering = min_t(unsigned int, tp->reordering, | |
1368 | sysctl_tcp_reordering); | |
6687e988 | 1369 | tcp_set_ca_state(sk, TCP_CA_Loss); |
1da177e4 LT |
1370 | tp->high_seq = tp->snd_nxt; |
1371 | TCP_ECN_queue_cwr(tp); | |
6a438bbe SH |
1372 | |
1373 | clear_all_retrans_hints(tp); | |
1da177e4 LT |
1374 | } |
1375 | ||
463c84b9 | 1376 | static int tcp_check_sack_reneging(struct sock *sk) |
1da177e4 LT |
1377 | { |
1378 | struct sk_buff *skb; | |
1379 | ||
1380 | /* If ACK arrived pointing to a remembered SACK, | |
1381 | * it means that our remembered SACKs do not reflect | |
1382 | * real state of receiver i.e. | |
1383 | * receiver _host_ is heavily congested (or buggy). | |
1384 | * Do processing similar to RTO timeout. | |
1385 | */ | |
1386 | if ((skb = skb_peek(&sk->sk_write_queue)) != NULL && | |
1387 | (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { | |
6687e988 | 1388 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
1389 | NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING); |
1390 | ||
1391 | tcp_enter_loss(sk, 1); | |
6687e988 | 1392 | icsk->icsk_retransmits++; |
1da177e4 | 1393 | tcp_retransmit_skb(sk, skb_peek(&sk->sk_write_queue)); |
463c84b9 | 1394 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
6687e988 | 1395 | icsk->icsk_rto, TCP_RTO_MAX); |
1da177e4 LT |
1396 | return 1; |
1397 | } | |
1398 | return 0; | |
1399 | } | |
1400 | ||
1401 | static inline int tcp_fackets_out(struct tcp_sock *tp) | |
1402 | { | |
1403 | return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out; | |
1404 | } | |
1405 | ||
463c84b9 | 1406 | static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 1407 | { |
463c84b9 | 1408 | return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto); |
1da177e4 LT |
1409 | } |
1410 | ||
1411 | static inline int tcp_head_timedout(struct sock *sk, struct tcp_sock *tp) | |
1412 | { | |
1413 | return tp->packets_out && | |
463c84b9 | 1414 | tcp_skb_timedout(sk, skb_peek(&sk->sk_write_queue)); |
1da177e4 LT |
1415 | } |
1416 | ||
1417 | /* Linux NewReno/SACK/FACK/ECN state machine. | |
1418 | * -------------------------------------- | |
1419 | * | |
1420 | * "Open" Normal state, no dubious events, fast path. | |
1421 | * "Disorder" In all the respects it is "Open", | |
1422 | * but requires a bit more attention. It is entered when | |
1423 | * we see some SACKs or dupacks. It is split of "Open" | |
1424 | * mainly to move some processing from fast path to slow one. | |
1425 | * "CWR" CWND was reduced due to some Congestion Notification event. | |
1426 | * It can be ECN, ICMP source quench, local device congestion. | |
1427 | * "Recovery" CWND was reduced, we are fast-retransmitting. | |
1428 | * "Loss" CWND was reduced due to RTO timeout or SACK reneging. | |
1429 | * | |
1430 | * tcp_fastretrans_alert() is entered: | |
1431 | * - each incoming ACK, if state is not "Open" | |
1432 | * - when arrived ACK is unusual, namely: | |
1433 | * * SACK | |
1434 | * * Duplicate ACK. | |
1435 | * * ECN ECE. | |
1436 | * | |
1437 | * Counting packets in flight is pretty simple. | |
1438 | * | |
1439 | * in_flight = packets_out - left_out + retrans_out | |
1440 | * | |
1441 | * packets_out is SND.NXT-SND.UNA counted in packets. | |
1442 | * | |
1443 | * retrans_out is number of retransmitted segments. | |
1444 | * | |
1445 | * left_out is number of segments left network, but not ACKed yet. | |
1446 | * | |
1447 | * left_out = sacked_out + lost_out | |
1448 | * | |
1449 | * sacked_out: Packets, which arrived to receiver out of order | |
1450 | * and hence not ACKed. With SACKs this number is simply | |
1451 | * amount of SACKed data. Even without SACKs | |
1452 | * it is easy to give pretty reliable estimate of this number, | |
1453 | * counting duplicate ACKs. | |
1454 | * | |
1455 | * lost_out: Packets lost by network. TCP has no explicit | |
1456 | * "loss notification" feedback from network (for now). | |
1457 | * It means that this number can be only _guessed_. | |
1458 | * Actually, it is the heuristics to predict lossage that | |
1459 | * distinguishes different algorithms. | |
1460 | * | |
1461 | * F.e. after RTO, when all the queue is considered as lost, | |
1462 | * lost_out = packets_out and in_flight = retrans_out. | |
1463 | * | |
1464 | * Essentially, we have now two algorithms counting | |
1465 | * lost packets. | |
1466 | * | |
1467 | * FACK: It is the simplest heuristics. As soon as we decided | |
1468 | * that something is lost, we decide that _all_ not SACKed | |
1469 | * packets until the most forward SACK are lost. I.e. | |
1470 | * lost_out = fackets_out - sacked_out and left_out = fackets_out. | |
1471 | * It is absolutely correct estimate, if network does not reorder | |
1472 | * packets. And it loses any connection to reality when reordering | |
1473 | * takes place. We use FACK by default until reordering | |
1474 | * is suspected on the path to this destination. | |
1475 | * | |
1476 | * NewReno: when Recovery is entered, we assume that one segment | |
1477 | * is lost (classic Reno). While we are in Recovery and | |
1478 | * a partial ACK arrives, we assume that one more packet | |
1479 | * is lost (NewReno). This heuristics are the same in NewReno | |
1480 | * and SACK. | |
1481 | * | |
1482 | * Imagine, that's all! Forget about all this shamanism about CWND inflation | |
1483 | * deflation etc. CWND is real congestion window, never inflated, changes | |
1484 | * only according to classic VJ rules. | |
1485 | * | |
1486 | * Really tricky (and requiring careful tuning) part of algorithm | |
1487 | * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). | |
1488 | * The first determines the moment _when_ we should reduce CWND and, | |
1489 | * hence, slow down forward transmission. In fact, it determines the moment | |
1490 | * when we decide that hole is caused by loss, rather than by a reorder. | |
1491 | * | |
1492 | * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill | |
1493 | * holes, caused by lost packets. | |
1494 | * | |
1495 | * And the most logically complicated part of algorithm is undo | |
1496 | * heuristics. We detect false retransmits due to both too early | |
1497 | * fast retransmit (reordering) and underestimated RTO, analyzing | |
1498 | * timestamps and D-SACKs. When we detect that some segments were | |
1499 | * retransmitted by mistake and CWND reduction was wrong, we undo | |
1500 | * window reduction and abort recovery phase. This logic is hidden | |
1501 | * inside several functions named tcp_try_undo_<something>. | |
1502 | */ | |
1503 | ||
1504 | /* This function decides, when we should leave Disordered state | |
1505 | * and enter Recovery phase, reducing congestion window. | |
1506 | * | |
1507 | * Main question: may we further continue forward transmission | |
1508 | * with the same cwnd? | |
1509 | */ | |
1510 | static int tcp_time_to_recover(struct sock *sk, struct tcp_sock *tp) | |
1511 | { | |
1512 | __u32 packets_out; | |
1513 | ||
1514 | /* Trick#1: The loss is proven. */ | |
1515 | if (tp->lost_out) | |
1516 | return 1; | |
1517 | ||
1518 | /* Not-A-Trick#2 : Classic rule... */ | |
1519 | if (tcp_fackets_out(tp) > tp->reordering) | |
1520 | return 1; | |
1521 | ||
1522 | /* Trick#3 : when we use RFC2988 timer restart, fast | |
1523 | * retransmit can be triggered by timeout of queue head. | |
1524 | */ | |
1525 | if (tcp_head_timedout(sk, tp)) | |
1526 | return 1; | |
1527 | ||
1528 | /* Trick#4: It is still not OK... But will it be useful to delay | |
1529 | * recovery more? | |
1530 | */ | |
1531 | packets_out = tp->packets_out; | |
1532 | if (packets_out <= tp->reordering && | |
1533 | tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && | |
1534 | !tcp_may_send_now(sk, tp)) { | |
1535 | /* We have nothing to send. This connection is limited | |
1536 | * either by receiver window or by application. | |
1537 | */ | |
1538 | return 1; | |
1539 | } | |
1540 | ||
1541 | return 0; | |
1542 | } | |
1543 | ||
1544 | /* If we receive more dupacks than we expected counting segments | |
1545 | * in assumption of absent reordering, interpret this as reordering. | |
1546 | * The only another reason could be bug in receiver TCP. | |
1547 | */ | |
6687e988 | 1548 | static void tcp_check_reno_reordering(struct sock *sk, const int addend) |
1da177e4 | 1549 | { |
6687e988 | 1550 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
1551 | u32 holes; |
1552 | ||
1553 | holes = max(tp->lost_out, 1U); | |
1554 | holes = min(holes, tp->packets_out); | |
1555 | ||
1556 | if ((tp->sacked_out + holes) > tp->packets_out) { | |
1557 | tp->sacked_out = tp->packets_out - holes; | |
6687e988 | 1558 | tcp_update_reordering(sk, tp->packets_out + addend, 0); |
1da177e4 LT |
1559 | } |
1560 | } | |
1561 | ||
1562 | /* Emulate SACKs for SACKless connection: account for a new dupack. */ | |
1563 | ||
6687e988 | 1564 | static void tcp_add_reno_sack(struct sock *sk) |
1da177e4 | 1565 | { |
6687e988 | 1566 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 1567 | tp->sacked_out++; |
6687e988 | 1568 | tcp_check_reno_reordering(sk, 0); |
1da177e4 LT |
1569 | tcp_sync_left_out(tp); |
1570 | } | |
1571 | ||
1572 | /* Account for ACK, ACKing some data in Reno Recovery phase. */ | |
1573 | ||
1574 | static void tcp_remove_reno_sacks(struct sock *sk, struct tcp_sock *tp, int acked) | |
1575 | { | |
1576 | if (acked > 0) { | |
1577 | /* One ACK acked hole. The rest eat duplicate ACKs. */ | |
1578 | if (acked-1 >= tp->sacked_out) | |
1579 | tp->sacked_out = 0; | |
1580 | else | |
1581 | tp->sacked_out -= acked-1; | |
1582 | } | |
6687e988 | 1583 | tcp_check_reno_reordering(sk, acked); |
1da177e4 LT |
1584 | tcp_sync_left_out(tp); |
1585 | } | |
1586 | ||
1587 | static inline void tcp_reset_reno_sack(struct tcp_sock *tp) | |
1588 | { | |
1589 | tp->sacked_out = 0; | |
1590 | tp->left_out = tp->lost_out; | |
1591 | } | |
1592 | ||
1593 | /* Mark head of queue up as lost. */ | |
1594 | static void tcp_mark_head_lost(struct sock *sk, struct tcp_sock *tp, | |
1595 | int packets, u32 high_seq) | |
1596 | { | |
1597 | struct sk_buff *skb; | |
6a438bbe | 1598 | int cnt; |
1da177e4 | 1599 | |
6a438bbe SH |
1600 | BUG_TRAP(packets <= tp->packets_out); |
1601 | if (tp->lost_skb_hint) { | |
1602 | skb = tp->lost_skb_hint; | |
1603 | cnt = tp->lost_cnt_hint; | |
1604 | } else { | |
1605 | skb = sk->sk_write_queue.next; | |
1606 | cnt = 0; | |
1607 | } | |
1da177e4 | 1608 | |
6a438bbe SH |
1609 | sk_stream_for_retrans_queue_from(skb, sk) { |
1610 | /* TODO: do this better */ | |
1611 | /* this is not the most efficient way to do this... */ | |
1612 | tp->lost_skb_hint = skb; | |
1613 | tp->lost_cnt_hint = cnt; | |
1614 | cnt += tcp_skb_pcount(skb); | |
1615 | if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq)) | |
1da177e4 LT |
1616 | break; |
1617 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) { | |
1618 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1619 | tp->lost_out += tcp_skb_pcount(skb); | |
6a438bbe SH |
1620 | |
1621 | /* clear xmit_retransmit_queue hints | |
1622 | * if this is beyond hint */ | |
1623 | if(tp->retransmit_skb_hint != NULL && | |
1624 | before(TCP_SKB_CB(skb)->seq, | |
1625 | TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) { | |
1626 | ||
1627 | tp->retransmit_skb_hint = NULL; | |
1628 | } | |
1da177e4 LT |
1629 | } |
1630 | } | |
1631 | tcp_sync_left_out(tp); | |
1632 | } | |
1633 | ||
1634 | /* Account newly detected lost packet(s) */ | |
1635 | ||
1636 | static void tcp_update_scoreboard(struct sock *sk, struct tcp_sock *tp) | |
1637 | { | |
1638 | if (IsFack(tp)) { | |
1639 | int lost = tp->fackets_out - tp->reordering; | |
1640 | if (lost <= 0) | |
1641 | lost = 1; | |
1642 | tcp_mark_head_lost(sk, tp, lost, tp->high_seq); | |
1643 | } else { | |
1644 | tcp_mark_head_lost(sk, tp, 1, tp->high_seq); | |
1645 | } | |
1646 | ||
1647 | /* New heuristics: it is possible only after we switched | |
1648 | * to restart timer each time when something is ACKed. | |
1649 | * Hence, we can detect timed out packets during fast | |
1650 | * retransmit without falling to slow start. | |
1651 | */ | |
79320d7e | 1652 | if (!IsReno(tp) && tcp_head_timedout(sk, tp)) { |
1da177e4 LT |
1653 | struct sk_buff *skb; |
1654 | ||
6a438bbe SH |
1655 | skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint |
1656 | : sk->sk_write_queue.next; | |
1657 | ||
1658 | sk_stream_for_retrans_queue_from(skb, sk) { | |
1659 | if (!tcp_skb_timedout(sk, skb)) | |
1660 | break; | |
1661 | ||
1662 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) { | |
1da177e4 LT |
1663 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; |
1664 | tp->lost_out += tcp_skb_pcount(skb); | |
6a438bbe SH |
1665 | |
1666 | /* clear xmit_retrans hint */ | |
1667 | if (tp->retransmit_skb_hint && | |
1668 | before(TCP_SKB_CB(skb)->seq, | |
1669 | TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) | |
1670 | ||
1671 | tp->retransmit_skb_hint = NULL; | |
1da177e4 LT |
1672 | } |
1673 | } | |
6a438bbe SH |
1674 | |
1675 | tp->scoreboard_skb_hint = skb; | |
1676 | ||
1da177e4 LT |
1677 | tcp_sync_left_out(tp); |
1678 | } | |
1679 | } | |
1680 | ||
1681 | /* CWND moderation, preventing bursts due to too big ACKs | |
1682 | * in dubious situations. | |
1683 | */ | |
1684 | static inline void tcp_moderate_cwnd(struct tcp_sock *tp) | |
1685 | { | |
1686 | tp->snd_cwnd = min(tp->snd_cwnd, | |
1687 | tcp_packets_in_flight(tp)+tcp_max_burst(tp)); | |
1688 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1689 | } | |
1690 | ||
72dc5b92 SH |
1691 | /* Lower bound on congestion window is slow start threshold |
1692 | * unless congestion avoidance choice decides to overide it. | |
1693 | */ | |
1694 | static inline u32 tcp_cwnd_min(const struct sock *sk) | |
1695 | { | |
1696 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; | |
1697 | ||
1698 | return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh; | |
1699 | } | |
1700 | ||
1da177e4 | 1701 | /* Decrease cwnd each second ack. */ |
6687e988 | 1702 | static void tcp_cwnd_down(struct sock *sk) |
1da177e4 | 1703 | { |
6687e988 | 1704 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 1705 | int decr = tp->snd_cwnd_cnt + 1; |
1da177e4 LT |
1706 | |
1707 | tp->snd_cwnd_cnt = decr&1; | |
1708 | decr >>= 1; | |
1709 | ||
72dc5b92 | 1710 | if (decr && tp->snd_cwnd > tcp_cwnd_min(sk)) |
1da177e4 LT |
1711 | tp->snd_cwnd -= decr; |
1712 | ||
1713 | tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1); | |
1714 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1715 | } | |
1716 | ||
1717 | /* Nothing was retransmitted or returned timestamp is less | |
1718 | * than timestamp of the first retransmission. | |
1719 | */ | |
1720 | static inline int tcp_packet_delayed(struct tcp_sock *tp) | |
1721 | { | |
1722 | return !tp->retrans_stamp || | |
1723 | (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && | |
1724 | (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0); | |
1725 | } | |
1726 | ||
1727 | /* Undo procedures. */ | |
1728 | ||
1729 | #if FASTRETRANS_DEBUG > 1 | |
1730 | static void DBGUNDO(struct sock *sk, struct tcp_sock *tp, const char *msg) | |
1731 | { | |
1732 | struct inet_sock *inet = inet_sk(sk); | |
1733 | printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n", | |
1734 | msg, | |
1735 | NIPQUAD(inet->daddr), ntohs(inet->dport), | |
1736 | tp->snd_cwnd, tp->left_out, | |
1737 | tp->snd_ssthresh, tp->prior_ssthresh, | |
1738 | tp->packets_out); | |
1739 | } | |
1740 | #else | |
1741 | #define DBGUNDO(x...) do { } while (0) | |
1742 | #endif | |
1743 | ||
6687e988 | 1744 | static void tcp_undo_cwr(struct sock *sk, const int undo) |
1da177e4 | 1745 | { |
6687e988 ACM |
1746 | struct tcp_sock *tp = tcp_sk(sk); |
1747 | ||
1da177e4 | 1748 | if (tp->prior_ssthresh) { |
6687e988 ACM |
1749 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1750 | ||
1751 | if (icsk->icsk_ca_ops->undo_cwnd) | |
1752 | tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); | |
1da177e4 LT |
1753 | else |
1754 | tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1); | |
1755 | ||
1756 | if (undo && tp->prior_ssthresh > tp->snd_ssthresh) { | |
1757 | tp->snd_ssthresh = tp->prior_ssthresh; | |
1758 | TCP_ECN_withdraw_cwr(tp); | |
1759 | } | |
1760 | } else { | |
1761 | tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); | |
1762 | } | |
1763 | tcp_moderate_cwnd(tp); | |
1764 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
6a438bbe SH |
1765 | |
1766 | /* There is something screwy going on with the retrans hints after | |
1767 | an undo */ | |
1768 | clear_all_retrans_hints(tp); | |
1da177e4 LT |
1769 | } |
1770 | ||
1771 | static inline int tcp_may_undo(struct tcp_sock *tp) | |
1772 | { | |
1773 | return tp->undo_marker && | |
1774 | (!tp->undo_retrans || tcp_packet_delayed(tp)); | |
1775 | } | |
1776 | ||
1777 | /* People celebrate: "We love our President!" */ | |
1778 | static int tcp_try_undo_recovery(struct sock *sk, struct tcp_sock *tp) | |
1779 | { | |
1780 | if (tcp_may_undo(tp)) { | |
1781 | /* Happy end! We did not retransmit anything | |
1782 | * or our original transmission succeeded. | |
1783 | */ | |
6687e988 ACM |
1784 | DBGUNDO(sk, tp, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); |
1785 | tcp_undo_cwr(sk, 1); | |
1786 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) | |
1da177e4 LT |
1787 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); |
1788 | else | |
1789 | NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO); | |
1790 | tp->undo_marker = 0; | |
1791 | } | |
1792 | if (tp->snd_una == tp->high_seq && IsReno(tp)) { | |
1793 | /* Hold old state until something *above* high_seq | |
1794 | * is ACKed. For Reno it is MUST to prevent false | |
1795 | * fast retransmits (RFC2582). SACK TCP is safe. */ | |
1796 | tcp_moderate_cwnd(tp); | |
1797 | return 1; | |
1798 | } | |
6687e988 | 1799 | tcp_set_ca_state(sk, TCP_CA_Open); |
1da177e4 LT |
1800 | return 0; |
1801 | } | |
1802 | ||
1803 | /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ | |
1804 | static void tcp_try_undo_dsack(struct sock *sk, struct tcp_sock *tp) | |
1805 | { | |
1806 | if (tp->undo_marker && !tp->undo_retrans) { | |
1807 | DBGUNDO(sk, tp, "D-SACK"); | |
6687e988 | 1808 | tcp_undo_cwr(sk, 1); |
1da177e4 LT |
1809 | tp->undo_marker = 0; |
1810 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO); | |
1811 | } | |
1812 | } | |
1813 | ||
1814 | /* Undo during fast recovery after partial ACK. */ | |
1815 | ||
1816 | static int tcp_try_undo_partial(struct sock *sk, struct tcp_sock *tp, | |
1817 | int acked) | |
1818 | { | |
1819 | /* Partial ACK arrived. Force Hoe's retransmit. */ | |
1820 | int failed = IsReno(tp) || tp->fackets_out>tp->reordering; | |
1821 | ||
1822 | if (tcp_may_undo(tp)) { | |
1823 | /* Plain luck! Hole if filled with delayed | |
1824 | * packet, rather than with a retransmit. | |
1825 | */ | |
1826 | if (tp->retrans_out == 0) | |
1827 | tp->retrans_stamp = 0; | |
1828 | ||
6687e988 | 1829 | tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); |
1da177e4 LT |
1830 | |
1831 | DBGUNDO(sk, tp, "Hoe"); | |
6687e988 | 1832 | tcp_undo_cwr(sk, 0); |
1da177e4 LT |
1833 | NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO); |
1834 | ||
1835 | /* So... Do not make Hoe's retransmit yet. | |
1836 | * If the first packet was delayed, the rest | |
1837 | * ones are most probably delayed as well. | |
1838 | */ | |
1839 | failed = 0; | |
1840 | } | |
1841 | return failed; | |
1842 | } | |
1843 | ||
1844 | /* Undo during loss recovery after partial ACK. */ | |
1845 | static int tcp_try_undo_loss(struct sock *sk, struct tcp_sock *tp) | |
1846 | { | |
1847 | if (tcp_may_undo(tp)) { | |
1848 | struct sk_buff *skb; | |
1849 | sk_stream_for_retrans_queue(skb, sk) { | |
1850 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; | |
1851 | } | |
6a438bbe SH |
1852 | |
1853 | clear_all_retrans_hints(tp); | |
1854 | ||
1da177e4 LT |
1855 | DBGUNDO(sk, tp, "partial loss"); |
1856 | tp->lost_out = 0; | |
1857 | tp->left_out = tp->sacked_out; | |
6687e988 | 1858 | tcp_undo_cwr(sk, 1); |
1da177e4 | 1859 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); |
463c84b9 | 1860 | inet_csk(sk)->icsk_retransmits = 0; |
1da177e4 LT |
1861 | tp->undo_marker = 0; |
1862 | if (!IsReno(tp)) | |
6687e988 | 1863 | tcp_set_ca_state(sk, TCP_CA_Open); |
1da177e4 LT |
1864 | return 1; |
1865 | } | |
1866 | return 0; | |
1867 | } | |
1868 | ||
6687e988 | 1869 | static inline void tcp_complete_cwr(struct sock *sk) |
1da177e4 | 1870 | { |
6687e988 | 1871 | struct tcp_sock *tp = tcp_sk(sk); |
317a76f9 | 1872 | tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); |
1da177e4 | 1873 | tp->snd_cwnd_stamp = tcp_time_stamp; |
6687e988 | 1874 | tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); |
1da177e4 LT |
1875 | } |
1876 | ||
1877 | static void tcp_try_to_open(struct sock *sk, struct tcp_sock *tp, int flag) | |
1878 | { | |
1879 | tp->left_out = tp->sacked_out; | |
1880 | ||
1881 | if (tp->retrans_out == 0) | |
1882 | tp->retrans_stamp = 0; | |
1883 | ||
1884 | if (flag&FLAG_ECE) | |
6687e988 | 1885 | tcp_enter_cwr(sk); |
1da177e4 | 1886 | |
6687e988 | 1887 | if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { |
1da177e4 LT |
1888 | int state = TCP_CA_Open; |
1889 | ||
1890 | if (tp->left_out || tp->retrans_out || tp->undo_marker) | |
1891 | state = TCP_CA_Disorder; | |
1892 | ||
6687e988 ACM |
1893 | if (inet_csk(sk)->icsk_ca_state != state) { |
1894 | tcp_set_ca_state(sk, state); | |
1da177e4 LT |
1895 | tp->high_seq = tp->snd_nxt; |
1896 | } | |
1897 | tcp_moderate_cwnd(tp); | |
1898 | } else { | |
6687e988 | 1899 | tcp_cwnd_down(sk); |
1da177e4 LT |
1900 | } |
1901 | } | |
1902 | ||
5d424d5a JH |
1903 | static void tcp_mtup_probe_failed(struct sock *sk) |
1904 | { | |
1905 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1906 | ||
1907 | icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; | |
1908 | icsk->icsk_mtup.probe_size = 0; | |
1909 | } | |
1910 | ||
1911 | static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb) | |
1912 | { | |
1913 | struct tcp_sock *tp = tcp_sk(sk); | |
1914 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1915 | ||
1916 | /* FIXME: breaks with very large cwnd */ | |
1917 | tp->prior_ssthresh = tcp_current_ssthresh(sk); | |
1918 | tp->snd_cwnd = tp->snd_cwnd * | |
1919 | tcp_mss_to_mtu(sk, tp->mss_cache) / | |
1920 | icsk->icsk_mtup.probe_size; | |
1921 | tp->snd_cwnd_cnt = 0; | |
1922 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1923 | tp->rcv_ssthresh = tcp_current_ssthresh(sk); | |
1924 | ||
1925 | icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; | |
1926 | icsk->icsk_mtup.probe_size = 0; | |
1927 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); | |
1928 | } | |
1929 | ||
1930 | ||
1da177e4 LT |
1931 | /* Process an event, which can update packets-in-flight not trivially. |
1932 | * Main goal of this function is to calculate new estimate for left_out, | |
1933 | * taking into account both packets sitting in receiver's buffer and | |
1934 | * packets lost by network. | |
1935 | * | |
1936 | * Besides that it does CWND reduction, when packet loss is detected | |
1937 | * and changes state of machine. | |
1938 | * | |
1939 | * It does _not_ decide what to send, it is made in function | |
1940 | * tcp_xmit_retransmit_queue(). | |
1941 | */ | |
1942 | static void | |
1943 | tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una, | |
1944 | int prior_packets, int flag) | |
1945 | { | |
6687e988 | 1946 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
1947 | struct tcp_sock *tp = tcp_sk(sk); |
1948 | int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP)); | |
1949 | ||
1950 | /* Some technical things: | |
1951 | * 1. Reno does not count dupacks (sacked_out) automatically. */ | |
1952 | if (!tp->packets_out) | |
1953 | tp->sacked_out = 0; | |
1954 | /* 2. SACK counts snd_fack in packets inaccurately. */ | |
1955 | if (tp->sacked_out == 0) | |
1956 | tp->fackets_out = 0; | |
1957 | ||
1958 | /* Now state machine starts. | |
1959 | * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ | |
1960 | if (flag&FLAG_ECE) | |
1961 | tp->prior_ssthresh = 0; | |
1962 | ||
1963 | /* B. In all the states check for reneging SACKs. */ | |
463c84b9 | 1964 | if (tp->sacked_out && tcp_check_sack_reneging(sk)) |
1da177e4 LT |
1965 | return; |
1966 | ||
1967 | /* C. Process data loss notification, provided it is valid. */ | |
1968 | if ((flag&FLAG_DATA_LOST) && | |
1969 | before(tp->snd_una, tp->high_seq) && | |
6687e988 | 1970 | icsk->icsk_ca_state != TCP_CA_Open && |
1da177e4 LT |
1971 | tp->fackets_out > tp->reordering) { |
1972 | tcp_mark_head_lost(sk, tp, tp->fackets_out-tp->reordering, tp->high_seq); | |
1973 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSS); | |
1974 | } | |
1975 | ||
1976 | /* D. Synchronize left_out to current state. */ | |
1977 | tcp_sync_left_out(tp); | |
1978 | ||
1979 | /* E. Check state exit conditions. State can be terminated | |
1980 | * when high_seq is ACKed. */ | |
6687e988 | 1981 | if (icsk->icsk_ca_state == TCP_CA_Open) { |
1da177e4 LT |
1982 | if (!sysctl_tcp_frto) |
1983 | BUG_TRAP(tp->retrans_out == 0); | |
1984 | tp->retrans_stamp = 0; | |
1985 | } else if (!before(tp->snd_una, tp->high_seq)) { | |
6687e988 | 1986 | switch (icsk->icsk_ca_state) { |
1da177e4 | 1987 | case TCP_CA_Loss: |
6687e988 | 1988 | icsk->icsk_retransmits = 0; |
1da177e4 LT |
1989 | if (tcp_try_undo_recovery(sk, tp)) |
1990 | return; | |
1991 | break; | |
1992 | ||
1993 | case TCP_CA_CWR: | |
1994 | /* CWR is to be held something *above* high_seq | |
1995 | * is ACKed for CWR bit to reach receiver. */ | |
1996 | if (tp->snd_una != tp->high_seq) { | |
6687e988 ACM |
1997 | tcp_complete_cwr(sk); |
1998 | tcp_set_ca_state(sk, TCP_CA_Open); | |
1da177e4 LT |
1999 | } |
2000 | break; | |
2001 | ||
2002 | case TCP_CA_Disorder: | |
2003 | tcp_try_undo_dsack(sk, tp); | |
2004 | if (!tp->undo_marker || | |
2005 | /* For SACK case do not Open to allow to undo | |
2006 | * catching for all duplicate ACKs. */ | |
2007 | IsReno(tp) || tp->snd_una != tp->high_seq) { | |
2008 | tp->undo_marker = 0; | |
6687e988 | 2009 | tcp_set_ca_state(sk, TCP_CA_Open); |
1da177e4 LT |
2010 | } |
2011 | break; | |
2012 | ||
2013 | case TCP_CA_Recovery: | |
2014 | if (IsReno(tp)) | |
2015 | tcp_reset_reno_sack(tp); | |
2016 | if (tcp_try_undo_recovery(sk, tp)) | |
2017 | return; | |
6687e988 | 2018 | tcp_complete_cwr(sk); |
1da177e4 LT |
2019 | break; |
2020 | } | |
2021 | } | |
2022 | ||
2023 | /* F. Process state. */ | |
6687e988 | 2024 | switch (icsk->icsk_ca_state) { |
1da177e4 LT |
2025 | case TCP_CA_Recovery: |
2026 | if (prior_snd_una == tp->snd_una) { | |
2027 | if (IsReno(tp) && is_dupack) | |
6687e988 | 2028 | tcp_add_reno_sack(sk); |
1da177e4 LT |
2029 | } else { |
2030 | int acked = prior_packets - tp->packets_out; | |
2031 | if (IsReno(tp)) | |
2032 | tcp_remove_reno_sacks(sk, tp, acked); | |
2033 | is_dupack = tcp_try_undo_partial(sk, tp, acked); | |
2034 | } | |
2035 | break; | |
2036 | case TCP_CA_Loss: | |
2037 | if (flag&FLAG_DATA_ACKED) | |
6687e988 | 2038 | icsk->icsk_retransmits = 0; |
1da177e4 LT |
2039 | if (!tcp_try_undo_loss(sk, tp)) { |
2040 | tcp_moderate_cwnd(tp); | |
2041 | tcp_xmit_retransmit_queue(sk); | |
2042 | return; | |
2043 | } | |
6687e988 | 2044 | if (icsk->icsk_ca_state != TCP_CA_Open) |
1da177e4 LT |
2045 | return; |
2046 | /* Loss is undone; fall through to processing in Open state. */ | |
2047 | default: | |
2048 | if (IsReno(tp)) { | |
2049 | if (tp->snd_una != prior_snd_una) | |
2050 | tcp_reset_reno_sack(tp); | |
2051 | if (is_dupack) | |
6687e988 | 2052 | tcp_add_reno_sack(sk); |
1da177e4 LT |
2053 | } |
2054 | ||
6687e988 | 2055 | if (icsk->icsk_ca_state == TCP_CA_Disorder) |
1da177e4 LT |
2056 | tcp_try_undo_dsack(sk, tp); |
2057 | ||
2058 | if (!tcp_time_to_recover(sk, tp)) { | |
2059 | tcp_try_to_open(sk, tp, flag); | |
2060 | return; | |
2061 | } | |
2062 | ||
5d424d5a JH |
2063 | /* MTU probe failure: don't reduce cwnd */ |
2064 | if (icsk->icsk_ca_state < TCP_CA_CWR && | |
2065 | icsk->icsk_mtup.probe_size && | |
0e7b1368 | 2066 | tp->snd_una == tp->mtu_probe.probe_seq_start) { |
5d424d5a JH |
2067 | tcp_mtup_probe_failed(sk); |
2068 | /* Restores the reduction we did in tcp_mtup_probe() */ | |
2069 | tp->snd_cwnd++; | |
2070 | tcp_simple_retransmit(sk); | |
2071 | return; | |
2072 | } | |
2073 | ||
1da177e4 LT |
2074 | /* Otherwise enter Recovery state */ |
2075 | ||
2076 | if (IsReno(tp)) | |
2077 | NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY); | |
2078 | else | |
2079 | NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY); | |
2080 | ||
2081 | tp->high_seq = tp->snd_nxt; | |
2082 | tp->prior_ssthresh = 0; | |
2083 | tp->undo_marker = tp->snd_una; | |
2084 | tp->undo_retrans = tp->retrans_out; | |
2085 | ||
6687e988 | 2086 | if (icsk->icsk_ca_state < TCP_CA_CWR) { |
1da177e4 | 2087 | if (!(flag&FLAG_ECE)) |
6687e988 ACM |
2088 | tp->prior_ssthresh = tcp_current_ssthresh(sk); |
2089 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
1da177e4 LT |
2090 | TCP_ECN_queue_cwr(tp); |
2091 | } | |
2092 | ||
9772efb9 | 2093 | tp->bytes_acked = 0; |
1da177e4 | 2094 | tp->snd_cwnd_cnt = 0; |
6687e988 | 2095 | tcp_set_ca_state(sk, TCP_CA_Recovery); |
1da177e4 LT |
2096 | } |
2097 | ||
2098 | if (is_dupack || tcp_head_timedout(sk, tp)) | |
2099 | tcp_update_scoreboard(sk, tp); | |
6687e988 | 2100 | tcp_cwnd_down(sk); |
1da177e4 LT |
2101 | tcp_xmit_retransmit_queue(sk); |
2102 | } | |
2103 | ||
2104 | /* Read draft-ietf-tcplw-high-performance before mucking | |
caa20d9a | 2105 | * with this code. (Supersedes RFC1323) |
1da177e4 | 2106 | */ |
2d2abbab | 2107 | static void tcp_ack_saw_tstamp(struct sock *sk, int flag) |
1da177e4 | 2108 | { |
1da177e4 LT |
2109 | /* RTTM Rule: A TSecr value received in a segment is used to |
2110 | * update the averaged RTT measurement only if the segment | |
2111 | * acknowledges some new data, i.e., only if it advances the | |
2112 | * left edge of the send window. | |
2113 | * | |
2114 | * See draft-ietf-tcplw-high-performance-00, section 3.3. | |
2115 | * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> | |
2116 | * | |
2117 | * Changed: reset backoff as soon as we see the first valid sample. | |
caa20d9a | 2118 | * If we do not, we get strongly overestimated rto. With timestamps |
1da177e4 LT |
2119 | * samples are accepted even from very old segments: f.e., when rtt=1 |
2120 | * increases to 8, we retransmit 5 times and after 8 seconds delayed | |
2121 | * answer arrives rto becomes 120 seconds! If at least one of segments | |
2122 | * in window is lost... Voila. --ANK (010210) | |
2123 | */ | |
463c84b9 ACM |
2124 | struct tcp_sock *tp = tcp_sk(sk); |
2125 | const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr; | |
2d2abbab | 2126 | tcp_rtt_estimator(sk, seq_rtt); |
463c84b9 ACM |
2127 | tcp_set_rto(sk); |
2128 | inet_csk(sk)->icsk_backoff = 0; | |
2129 | tcp_bound_rto(sk); | |
1da177e4 LT |
2130 | } |
2131 | ||
2d2abbab | 2132 | static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) |
1da177e4 LT |
2133 | { |
2134 | /* We don't have a timestamp. Can only use | |
2135 | * packets that are not retransmitted to determine | |
2136 | * rtt estimates. Also, we must not reset the | |
2137 | * backoff for rto until we get a non-retransmitted | |
2138 | * packet. This allows us to deal with a situation | |
2139 | * where the network delay has increased suddenly. | |
2140 | * I.e. Karn's algorithm. (SIGCOMM '87, p5.) | |
2141 | */ | |
2142 | ||
2143 | if (flag & FLAG_RETRANS_DATA_ACKED) | |
2144 | return; | |
2145 | ||
2d2abbab | 2146 | tcp_rtt_estimator(sk, seq_rtt); |
463c84b9 ACM |
2147 | tcp_set_rto(sk); |
2148 | inet_csk(sk)->icsk_backoff = 0; | |
2149 | tcp_bound_rto(sk); | |
1da177e4 LT |
2150 | } |
2151 | ||
463c84b9 | 2152 | static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, |
2d2abbab | 2153 | const s32 seq_rtt) |
1da177e4 | 2154 | { |
463c84b9 | 2155 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
2156 | /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ |
2157 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) | |
2d2abbab | 2158 | tcp_ack_saw_tstamp(sk, flag); |
1da177e4 | 2159 | else if (seq_rtt >= 0) |
2d2abbab | 2160 | tcp_ack_no_tstamp(sk, seq_rtt, flag); |
1da177e4 LT |
2161 | } |
2162 | ||
40efc6fa SH |
2163 | static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 rtt, |
2164 | u32 in_flight, int good) | |
1da177e4 | 2165 | { |
6687e988 ACM |
2166 | const struct inet_connection_sock *icsk = inet_csk(sk); |
2167 | icsk->icsk_ca_ops->cong_avoid(sk, ack, rtt, in_flight, good); | |
2168 | tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; | |
1da177e4 LT |
2169 | } |
2170 | ||
1da177e4 LT |
2171 | /* Restart timer after forward progress on connection. |
2172 | * RFC2988 recommends to restart timer to now+rto. | |
2173 | */ | |
2174 | ||
40efc6fa | 2175 | static void tcp_ack_packets_out(struct sock *sk, struct tcp_sock *tp) |
1da177e4 LT |
2176 | { |
2177 | if (!tp->packets_out) { | |
463c84b9 | 2178 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); |
1da177e4 | 2179 | } else { |
3f421baa | 2180 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX); |
1da177e4 LT |
2181 | } |
2182 | } | |
2183 | ||
1da177e4 LT |
2184 | static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb, |
2185 | __u32 now, __s32 *seq_rtt) | |
2186 | { | |
2187 | struct tcp_sock *tp = tcp_sk(sk); | |
2188 | struct tcp_skb_cb *scb = TCP_SKB_CB(skb); | |
2189 | __u32 seq = tp->snd_una; | |
2190 | __u32 packets_acked; | |
2191 | int acked = 0; | |
2192 | ||
2193 | /* If we get here, the whole TSO packet has not been | |
2194 | * acked. | |
2195 | */ | |
2196 | BUG_ON(!after(scb->end_seq, seq)); | |
2197 | ||
2198 | packets_acked = tcp_skb_pcount(skb); | |
2199 | if (tcp_trim_head(sk, skb, seq - scb->seq)) | |
2200 | return 0; | |
2201 | packets_acked -= tcp_skb_pcount(skb); | |
2202 | ||
2203 | if (packets_acked) { | |
2204 | __u8 sacked = scb->sacked; | |
2205 | ||
2206 | acked |= FLAG_DATA_ACKED; | |
2207 | if (sacked) { | |
2208 | if (sacked & TCPCB_RETRANS) { | |
2209 | if (sacked & TCPCB_SACKED_RETRANS) | |
2210 | tp->retrans_out -= packets_acked; | |
2211 | acked |= FLAG_RETRANS_DATA_ACKED; | |
2212 | *seq_rtt = -1; | |
2213 | } else if (*seq_rtt < 0) | |
2214 | *seq_rtt = now - scb->when; | |
2215 | if (sacked & TCPCB_SACKED_ACKED) | |
2216 | tp->sacked_out -= packets_acked; | |
2217 | if (sacked & TCPCB_LOST) | |
2218 | tp->lost_out -= packets_acked; | |
2219 | if (sacked & TCPCB_URG) { | |
2220 | if (tp->urg_mode && | |
2221 | !before(seq, tp->snd_up)) | |
2222 | tp->urg_mode = 0; | |
2223 | } | |
2224 | } else if (*seq_rtt < 0) | |
2225 | *seq_rtt = now - scb->when; | |
2226 | ||
2227 | if (tp->fackets_out) { | |
2228 | __u32 dval = min(tp->fackets_out, packets_acked); | |
2229 | tp->fackets_out -= dval; | |
2230 | } | |
2231 | tp->packets_out -= packets_acked; | |
2232 | ||
2233 | BUG_ON(tcp_skb_pcount(skb) == 0); | |
2234 | BUG_ON(!before(scb->seq, scb->end_seq)); | |
2235 | } | |
2236 | ||
2237 | return acked; | |
2238 | } | |
2239 | ||
40efc6fa | 2240 | static u32 tcp_usrtt(const struct sk_buff *skb) |
2d2abbab SH |
2241 | { |
2242 | struct timeval tv, now; | |
2243 | ||
2244 | do_gettimeofday(&now); | |
2245 | skb_get_timestamp(skb, &tv); | |
2246 | return (now.tv_sec - tv.tv_sec) * 1000000 + (now.tv_usec - tv.tv_usec); | |
2247 | } | |
1da177e4 LT |
2248 | |
2249 | /* Remove acknowledged frames from the retransmission queue. */ | |
2d2abbab | 2250 | static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p) |
1da177e4 LT |
2251 | { |
2252 | struct tcp_sock *tp = tcp_sk(sk); | |
2d2abbab | 2253 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2254 | struct sk_buff *skb; |
2255 | __u32 now = tcp_time_stamp; | |
2256 | int acked = 0; | |
2257 | __s32 seq_rtt = -1; | |
317a76f9 | 2258 | u32 pkts_acked = 0; |
2d2abbab SH |
2259 | void (*rtt_sample)(struct sock *sk, u32 usrtt) |
2260 | = icsk->icsk_ca_ops->rtt_sample; | |
1da177e4 LT |
2261 | |
2262 | while ((skb = skb_peek(&sk->sk_write_queue)) && | |
2263 | skb != sk->sk_send_head) { | |
2264 | struct tcp_skb_cb *scb = TCP_SKB_CB(skb); | |
2265 | __u8 sacked = scb->sacked; | |
2266 | ||
2267 | /* If our packet is before the ack sequence we can | |
2268 | * discard it as it's confirmed to have arrived at | |
2269 | * the other end. | |
2270 | */ | |
2271 | if (after(scb->end_seq, tp->snd_una)) { | |
cb83199a DM |
2272 | if (tcp_skb_pcount(skb) > 1 && |
2273 | after(tp->snd_una, scb->seq)) | |
1da177e4 LT |
2274 | acked |= tcp_tso_acked(sk, skb, |
2275 | now, &seq_rtt); | |
2276 | break; | |
2277 | } | |
2278 | ||
2279 | /* Initial outgoing SYN's get put onto the write_queue | |
2280 | * just like anything else we transmit. It is not | |
2281 | * true data, and if we misinform our callers that | |
2282 | * this ACK acks real data, we will erroneously exit | |
2283 | * connection startup slow start one packet too | |
2284 | * quickly. This is severely frowned upon behavior. | |
2285 | */ | |
2286 | if (!(scb->flags & TCPCB_FLAG_SYN)) { | |
2287 | acked |= FLAG_DATA_ACKED; | |
317a76f9 | 2288 | ++pkts_acked; |
1da177e4 LT |
2289 | } else { |
2290 | acked |= FLAG_SYN_ACKED; | |
2291 | tp->retrans_stamp = 0; | |
2292 | } | |
2293 | ||
5d424d5a JH |
2294 | /* MTU probing checks */ |
2295 | if (icsk->icsk_mtup.probe_size) { | |
0e7b1368 | 2296 | if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) { |
5d424d5a JH |
2297 | tcp_mtup_probe_success(sk, skb); |
2298 | } | |
2299 | } | |
2300 | ||
1da177e4 LT |
2301 | if (sacked) { |
2302 | if (sacked & TCPCB_RETRANS) { | |
2303 | if(sacked & TCPCB_SACKED_RETRANS) | |
2304 | tp->retrans_out -= tcp_skb_pcount(skb); | |
2305 | acked |= FLAG_RETRANS_DATA_ACKED; | |
2306 | seq_rtt = -1; | |
2d2abbab | 2307 | } else if (seq_rtt < 0) { |
1da177e4 | 2308 | seq_rtt = now - scb->when; |
2d2abbab SH |
2309 | if (rtt_sample) |
2310 | (*rtt_sample)(sk, tcp_usrtt(skb)); | |
a61bbcf2 | 2311 | } |
1da177e4 LT |
2312 | if (sacked & TCPCB_SACKED_ACKED) |
2313 | tp->sacked_out -= tcp_skb_pcount(skb); | |
2314 | if (sacked & TCPCB_LOST) | |
2315 | tp->lost_out -= tcp_skb_pcount(skb); | |
2316 | if (sacked & TCPCB_URG) { | |
2317 | if (tp->urg_mode && | |
2318 | !before(scb->end_seq, tp->snd_up)) | |
2319 | tp->urg_mode = 0; | |
2320 | } | |
2d2abbab | 2321 | } else if (seq_rtt < 0) { |
1da177e4 | 2322 | seq_rtt = now - scb->when; |
2d2abbab SH |
2323 | if (rtt_sample) |
2324 | (*rtt_sample)(sk, tcp_usrtt(skb)); | |
2325 | } | |
1da177e4 LT |
2326 | tcp_dec_pcount_approx(&tp->fackets_out, skb); |
2327 | tcp_packets_out_dec(tp, skb); | |
8728b834 | 2328 | __skb_unlink(skb, &sk->sk_write_queue); |
1da177e4 | 2329 | sk_stream_free_skb(sk, skb); |
6a438bbe | 2330 | clear_all_retrans_hints(tp); |
1da177e4 LT |
2331 | } |
2332 | ||
2333 | if (acked&FLAG_ACKED) { | |
2d2abbab | 2334 | tcp_ack_update_rtt(sk, acked, seq_rtt); |
1da177e4 | 2335 | tcp_ack_packets_out(sk, tp); |
317a76f9 | 2336 | |
6687e988 ACM |
2337 | if (icsk->icsk_ca_ops->pkts_acked) |
2338 | icsk->icsk_ca_ops->pkts_acked(sk, pkts_acked); | |
1da177e4 LT |
2339 | } |
2340 | ||
2341 | #if FASTRETRANS_DEBUG > 0 | |
2342 | BUG_TRAP((int)tp->sacked_out >= 0); | |
2343 | BUG_TRAP((int)tp->lost_out >= 0); | |
2344 | BUG_TRAP((int)tp->retrans_out >= 0); | |
2345 | if (!tp->packets_out && tp->rx_opt.sack_ok) { | |
6687e988 | 2346 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2347 | if (tp->lost_out) { |
2348 | printk(KERN_DEBUG "Leak l=%u %d\n", | |
6687e988 | 2349 | tp->lost_out, icsk->icsk_ca_state); |
1da177e4 LT |
2350 | tp->lost_out = 0; |
2351 | } | |
2352 | if (tp->sacked_out) { | |
2353 | printk(KERN_DEBUG "Leak s=%u %d\n", | |
6687e988 | 2354 | tp->sacked_out, icsk->icsk_ca_state); |
1da177e4 LT |
2355 | tp->sacked_out = 0; |
2356 | } | |
2357 | if (tp->retrans_out) { | |
2358 | printk(KERN_DEBUG "Leak r=%u %d\n", | |
6687e988 | 2359 | tp->retrans_out, icsk->icsk_ca_state); |
1da177e4 LT |
2360 | tp->retrans_out = 0; |
2361 | } | |
2362 | } | |
2363 | #endif | |
2364 | *seq_rtt_p = seq_rtt; | |
2365 | return acked; | |
2366 | } | |
2367 | ||
2368 | static void tcp_ack_probe(struct sock *sk) | |
2369 | { | |
463c84b9 ACM |
2370 | const struct tcp_sock *tp = tcp_sk(sk); |
2371 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1da177e4 LT |
2372 | |
2373 | /* Was it a usable window open? */ | |
2374 | ||
2375 | if (!after(TCP_SKB_CB(sk->sk_send_head)->end_seq, | |
2376 | tp->snd_una + tp->snd_wnd)) { | |
463c84b9 ACM |
2377 | icsk->icsk_backoff = 0; |
2378 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); | |
1da177e4 LT |
2379 | /* Socket must be waked up by subsequent tcp_data_snd_check(). |
2380 | * This function is not for random using! | |
2381 | */ | |
2382 | } else { | |
463c84b9 | 2383 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
3f421baa ACM |
2384 | min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), |
2385 | TCP_RTO_MAX); | |
1da177e4 LT |
2386 | } |
2387 | } | |
2388 | ||
6687e988 | 2389 | static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag) |
1da177e4 LT |
2390 | { |
2391 | return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || | |
6687e988 | 2392 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open); |
1da177e4 LT |
2393 | } |
2394 | ||
6687e988 | 2395 | static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag) |
1da177e4 | 2396 | { |
6687e988 | 2397 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 2398 | return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && |
6687e988 | 2399 | !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR)); |
1da177e4 LT |
2400 | } |
2401 | ||
2402 | /* Check that window update is acceptable. | |
2403 | * The function assumes that snd_una<=ack<=snd_next. | |
2404 | */ | |
463c84b9 ACM |
2405 | static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack, |
2406 | const u32 ack_seq, const u32 nwin) | |
1da177e4 LT |
2407 | { |
2408 | return (after(ack, tp->snd_una) || | |
2409 | after(ack_seq, tp->snd_wl1) || | |
2410 | (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd)); | |
2411 | } | |
2412 | ||
2413 | /* Update our send window. | |
2414 | * | |
2415 | * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 | |
2416 | * and in FreeBSD. NetBSD's one is even worse.) is wrong. | |
2417 | */ | |
2418 | static int tcp_ack_update_window(struct sock *sk, struct tcp_sock *tp, | |
2419 | struct sk_buff *skb, u32 ack, u32 ack_seq) | |
2420 | { | |
2421 | int flag = 0; | |
2422 | u32 nwin = ntohs(skb->h.th->window); | |
2423 | ||
2424 | if (likely(!skb->h.th->syn)) | |
2425 | nwin <<= tp->rx_opt.snd_wscale; | |
2426 | ||
2427 | if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { | |
2428 | flag |= FLAG_WIN_UPDATE; | |
2429 | tcp_update_wl(tp, ack, ack_seq); | |
2430 | ||
2431 | if (tp->snd_wnd != nwin) { | |
2432 | tp->snd_wnd = nwin; | |
2433 | ||
2434 | /* Note, it is the only place, where | |
2435 | * fast path is recovered for sending TCP. | |
2436 | */ | |
2ad41065 | 2437 | tp->pred_flags = 0; |
1da177e4 LT |
2438 | tcp_fast_path_check(sk, tp); |
2439 | ||
2440 | if (nwin > tp->max_window) { | |
2441 | tp->max_window = nwin; | |
d83d8461 | 2442 | tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); |
1da177e4 LT |
2443 | } |
2444 | } | |
2445 | } | |
2446 | ||
2447 | tp->snd_una = ack; | |
2448 | ||
2449 | return flag; | |
2450 | } | |
2451 | ||
2452 | static void tcp_process_frto(struct sock *sk, u32 prior_snd_una) | |
2453 | { | |
2454 | struct tcp_sock *tp = tcp_sk(sk); | |
2455 | ||
2456 | tcp_sync_left_out(tp); | |
2457 | ||
2458 | if (tp->snd_una == prior_snd_una || | |
2459 | !before(tp->snd_una, tp->frto_highmark)) { | |
2460 | /* RTO was caused by loss, start retransmitting in | |
2461 | * go-back-N slow start | |
2462 | */ | |
2463 | tcp_enter_frto_loss(sk); | |
2464 | return; | |
2465 | } | |
2466 | ||
2467 | if (tp->frto_counter == 1) { | |
2468 | /* First ACK after RTO advances the window: allow two new | |
2469 | * segments out. | |
2470 | */ | |
2471 | tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; | |
2472 | } else { | |
2473 | /* Also the second ACK after RTO advances the window. | |
2474 | * The RTO was likely spurious. Reduce cwnd and continue | |
2475 | * in congestion avoidance | |
2476 | */ | |
2477 | tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); | |
2478 | tcp_moderate_cwnd(tp); | |
2479 | } | |
2480 | ||
2481 | /* F-RTO affects on two new ACKs following RTO. | |
caa20d9a | 2482 | * At latest on third ACK the TCP behavior is back to normal. |
1da177e4 LT |
2483 | */ |
2484 | tp->frto_counter = (tp->frto_counter + 1) % 3; | |
2485 | } | |
2486 | ||
1da177e4 LT |
2487 | /* This routine deals with incoming acks, but not outgoing ones. */ |
2488 | static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag) | |
2489 | { | |
6687e988 | 2490 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2491 | struct tcp_sock *tp = tcp_sk(sk); |
2492 | u32 prior_snd_una = tp->snd_una; | |
2493 | u32 ack_seq = TCP_SKB_CB(skb)->seq; | |
2494 | u32 ack = TCP_SKB_CB(skb)->ack_seq; | |
2495 | u32 prior_in_flight; | |
2496 | s32 seq_rtt; | |
2497 | int prior_packets; | |
2498 | ||
2499 | /* If the ack is newer than sent or older than previous acks | |
2500 | * then we can probably ignore it. | |
2501 | */ | |
2502 | if (after(ack, tp->snd_nxt)) | |
2503 | goto uninteresting_ack; | |
2504 | ||
2505 | if (before(ack, prior_snd_una)) | |
2506 | goto old_ack; | |
2507 | ||
3fdf3f0c DO |
2508 | if (sysctl_tcp_abc) { |
2509 | if (icsk->icsk_ca_state < TCP_CA_CWR) | |
2510 | tp->bytes_acked += ack - prior_snd_una; | |
2511 | else if (icsk->icsk_ca_state == TCP_CA_Loss) | |
2512 | /* we assume just one segment left network */ | |
2513 | tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache); | |
2514 | } | |
9772efb9 | 2515 | |
1da177e4 LT |
2516 | if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) { |
2517 | /* Window is constant, pure forward advance. | |
2518 | * No more checks are required. | |
2519 | * Note, we use the fact that SND.UNA>=SND.WL2. | |
2520 | */ | |
2521 | tcp_update_wl(tp, ack, ack_seq); | |
2522 | tp->snd_una = ack; | |
1da177e4 LT |
2523 | flag |= FLAG_WIN_UPDATE; |
2524 | ||
6687e988 | 2525 | tcp_ca_event(sk, CA_EVENT_FAST_ACK); |
317a76f9 | 2526 | |
1da177e4 LT |
2527 | NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS); |
2528 | } else { | |
2529 | if (ack_seq != TCP_SKB_CB(skb)->end_seq) | |
2530 | flag |= FLAG_DATA; | |
2531 | else | |
2532 | NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS); | |
2533 | ||
2534 | flag |= tcp_ack_update_window(sk, tp, skb, ack, ack_seq); | |
2535 | ||
2536 | if (TCP_SKB_CB(skb)->sacked) | |
2537 | flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); | |
2538 | ||
2539 | if (TCP_ECN_rcv_ecn_echo(tp, skb->h.th)) | |
2540 | flag |= FLAG_ECE; | |
2541 | ||
6687e988 | 2542 | tcp_ca_event(sk, CA_EVENT_SLOW_ACK); |
1da177e4 LT |
2543 | } |
2544 | ||
2545 | /* We passed data and got it acked, remove any soft error | |
2546 | * log. Something worked... | |
2547 | */ | |
2548 | sk->sk_err_soft = 0; | |
2549 | tp->rcv_tstamp = tcp_time_stamp; | |
2550 | prior_packets = tp->packets_out; | |
2551 | if (!prior_packets) | |
2552 | goto no_queue; | |
2553 | ||
2554 | prior_in_flight = tcp_packets_in_flight(tp); | |
2555 | ||
2556 | /* See if we can take anything off of the retransmit queue. */ | |
2d2abbab | 2557 | flag |= tcp_clean_rtx_queue(sk, &seq_rtt); |
1da177e4 LT |
2558 | |
2559 | if (tp->frto_counter) | |
2560 | tcp_process_frto(sk, prior_snd_una); | |
2561 | ||
6687e988 | 2562 | if (tcp_ack_is_dubious(sk, flag)) { |
caa20d9a | 2563 | /* Advance CWND, if state allows this. */ |
6687e988 ACM |
2564 | if ((flag & FLAG_DATA_ACKED) && tcp_may_raise_cwnd(sk, flag)) |
2565 | tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 0); | |
1da177e4 LT |
2566 | tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag); |
2567 | } else { | |
317a76f9 | 2568 | if ((flag & FLAG_DATA_ACKED)) |
6687e988 | 2569 | tcp_cong_avoid(sk, ack, seq_rtt, prior_in_flight, 1); |
1da177e4 LT |
2570 | } |
2571 | ||
2572 | if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP)) | |
2573 | dst_confirm(sk->sk_dst_cache); | |
2574 | ||
2575 | return 1; | |
2576 | ||
2577 | no_queue: | |
6687e988 | 2578 | icsk->icsk_probes_out = 0; |
1da177e4 LT |
2579 | |
2580 | /* If this ack opens up a zero window, clear backoff. It was | |
2581 | * being used to time the probes, and is probably far higher than | |
2582 | * it needs to be for normal retransmission. | |
2583 | */ | |
2584 | if (sk->sk_send_head) | |
2585 | tcp_ack_probe(sk); | |
2586 | return 1; | |
2587 | ||
2588 | old_ack: | |
2589 | if (TCP_SKB_CB(skb)->sacked) | |
2590 | tcp_sacktag_write_queue(sk, skb, prior_snd_una); | |
2591 | ||
2592 | uninteresting_ack: | |
2593 | SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt); | |
2594 | return 0; | |
2595 | } | |
2596 | ||
2597 | ||
2598 | /* Look for tcp options. Normally only called on SYN and SYNACK packets. | |
2599 | * But, this can also be called on packets in the established flow when | |
2600 | * the fast version below fails. | |
2601 | */ | |
2602 | void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab) | |
2603 | { | |
2604 | unsigned char *ptr; | |
2605 | struct tcphdr *th = skb->h.th; | |
2606 | int length=(th->doff*4)-sizeof(struct tcphdr); | |
2607 | ||
2608 | ptr = (unsigned char *)(th + 1); | |
2609 | opt_rx->saw_tstamp = 0; | |
2610 | ||
2611 | while(length>0) { | |
2612 | int opcode=*ptr++; | |
2613 | int opsize; | |
2614 | ||
2615 | switch (opcode) { | |
2616 | case TCPOPT_EOL: | |
2617 | return; | |
2618 | case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ | |
2619 | length--; | |
2620 | continue; | |
2621 | default: | |
2622 | opsize=*ptr++; | |
2623 | if (opsize < 2) /* "silly options" */ | |
2624 | return; | |
2625 | if (opsize > length) | |
2626 | return; /* don't parse partial options */ | |
2627 | switch(opcode) { | |
2628 | case TCPOPT_MSS: | |
2629 | if(opsize==TCPOLEN_MSS && th->syn && !estab) { | |
2630 | u16 in_mss = ntohs(get_unaligned((__u16 *)ptr)); | |
2631 | if (in_mss) { | |
2632 | if (opt_rx->user_mss && opt_rx->user_mss < in_mss) | |
2633 | in_mss = opt_rx->user_mss; | |
2634 | opt_rx->mss_clamp = in_mss; | |
2635 | } | |
2636 | } | |
2637 | break; | |
2638 | case TCPOPT_WINDOW: | |
2639 | if(opsize==TCPOLEN_WINDOW && th->syn && !estab) | |
2640 | if (sysctl_tcp_window_scaling) { | |
2641 | __u8 snd_wscale = *(__u8 *) ptr; | |
2642 | opt_rx->wscale_ok = 1; | |
2643 | if (snd_wscale > 14) { | |
2644 | if(net_ratelimit()) | |
2645 | printk(KERN_INFO "tcp_parse_options: Illegal window " | |
2646 | "scaling value %d >14 received.\n", | |
2647 | snd_wscale); | |
2648 | snd_wscale = 14; | |
2649 | } | |
2650 | opt_rx->snd_wscale = snd_wscale; | |
2651 | } | |
2652 | break; | |
2653 | case TCPOPT_TIMESTAMP: | |
2654 | if(opsize==TCPOLEN_TIMESTAMP) { | |
2655 | if ((estab && opt_rx->tstamp_ok) || | |
2656 | (!estab && sysctl_tcp_timestamps)) { | |
2657 | opt_rx->saw_tstamp = 1; | |
2658 | opt_rx->rcv_tsval = ntohl(get_unaligned((__u32 *)ptr)); | |
2659 | opt_rx->rcv_tsecr = ntohl(get_unaligned((__u32 *)(ptr+4))); | |
2660 | } | |
2661 | } | |
2662 | break; | |
2663 | case TCPOPT_SACK_PERM: | |
2664 | if(opsize==TCPOLEN_SACK_PERM && th->syn && !estab) { | |
2665 | if (sysctl_tcp_sack) { | |
2666 | opt_rx->sack_ok = 1; | |
2667 | tcp_sack_reset(opt_rx); | |
2668 | } | |
2669 | } | |
2670 | break; | |
2671 | ||
2672 | case TCPOPT_SACK: | |
2673 | if((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && | |
2674 | !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && | |
2675 | opt_rx->sack_ok) { | |
2676 | TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; | |
2677 | } | |
2678 | }; | |
2679 | ptr+=opsize-2; | |
2680 | length-=opsize; | |
2681 | }; | |
2682 | } | |
2683 | } | |
2684 | ||
2685 | /* Fast parse options. This hopes to only see timestamps. | |
2686 | * If it is wrong it falls back on tcp_parse_options(). | |
2687 | */ | |
40efc6fa SH |
2688 | static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, |
2689 | struct tcp_sock *tp) | |
1da177e4 LT |
2690 | { |
2691 | if (th->doff == sizeof(struct tcphdr)>>2) { | |
2692 | tp->rx_opt.saw_tstamp = 0; | |
2693 | return 0; | |
2694 | } else if (tp->rx_opt.tstamp_ok && | |
2695 | th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) { | |
2696 | __u32 *ptr = (__u32 *)(th + 1); | |
2697 | if (*ptr == ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | |
2698 | | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { | |
2699 | tp->rx_opt.saw_tstamp = 1; | |
2700 | ++ptr; | |
2701 | tp->rx_opt.rcv_tsval = ntohl(*ptr); | |
2702 | ++ptr; | |
2703 | tp->rx_opt.rcv_tsecr = ntohl(*ptr); | |
2704 | return 1; | |
2705 | } | |
2706 | } | |
2707 | tcp_parse_options(skb, &tp->rx_opt, 1); | |
2708 | return 1; | |
2709 | } | |
2710 | ||
2711 | static inline void tcp_store_ts_recent(struct tcp_sock *tp) | |
2712 | { | |
2713 | tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; | |
2714 | tp->rx_opt.ts_recent_stamp = xtime.tv_sec; | |
2715 | } | |
2716 | ||
2717 | static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) | |
2718 | { | |
2719 | if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { | |
2720 | /* PAWS bug workaround wrt. ACK frames, the PAWS discard | |
2721 | * extra check below makes sure this can only happen | |
2722 | * for pure ACK frames. -DaveM | |
2723 | * | |
2724 | * Not only, also it occurs for expired timestamps. | |
2725 | */ | |
2726 | ||
2727 | if((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 || | |
2728 | xtime.tv_sec >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS) | |
2729 | tcp_store_ts_recent(tp); | |
2730 | } | |
2731 | } | |
2732 | ||
2733 | /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM | |
2734 | * | |
2735 | * It is not fatal. If this ACK does _not_ change critical state (seqs, window) | |
2736 | * it can pass through stack. So, the following predicate verifies that | |
2737 | * this segment is not used for anything but congestion avoidance or | |
2738 | * fast retransmit. Moreover, we even are able to eliminate most of such | |
2739 | * second order effects, if we apply some small "replay" window (~RTO) | |
2740 | * to timestamp space. | |
2741 | * | |
2742 | * All these measures still do not guarantee that we reject wrapped ACKs | |
2743 | * on networks with high bandwidth, when sequence space is recycled fastly, | |
2744 | * but it guarantees that such events will be very rare and do not affect | |
2745 | * connection seriously. This doesn't look nice, but alas, PAWS is really | |
2746 | * buggy extension. | |
2747 | * | |
2748 | * [ Later note. Even worse! It is buggy for segments _with_ data. RFC | |
2749 | * states that events when retransmit arrives after original data are rare. | |
2750 | * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is | |
2751 | * the biggest problem on large power networks even with minor reordering. | |
2752 | * OK, let's give it small replay window. If peer clock is even 1hz, it is safe | |
2753 | * up to bandwidth of 18Gigabit/sec. 8) ] | |
2754 | */ | |
2755 | ||
463c84b9 | 2756 | static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 2757 | { |
463c84b9 | 2758 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
2759 | struct tcphdr *th = skb->h.th; |
2760 | u32 seq = TCP_SKB_CB(skb)->seq; | |
2761 | u32 ack = TCP_SKB_CB(skb)->ack_seq; | |
2762 | ||
2763 | return (/* 1. Pure ACK with correct sequence number. */ | |
2764 | (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && | |
2765 | ||
2766 | /* 2. ... and duplicate ACK. */ | |
2767 | ack == tp->snd_una && | |
2768 | ||
2769 | /* 3. ... and does not update window. */ | |
2770 | !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && | |
2771 | ||
2772 | /* 4. ... and sits in replay window. */ | |
463c84b9 | 2773 | (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); |
1da177e4 LT |
2774 | } |
2775 | ||
463c84b9 | 2776 | static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 2777 | { |
463c84b9 | 2778 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
2779 | return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW && |
2780 | xtime.tv_sec < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS && | |
463c84b9 | 2781 | !tcp_disordered_ack(sk, skb)); |
1da177e4 LT |
2782 | } |
2783 | ||
2784 | /* Check segment sequence number for validity. | |
2785 | * | |
2786 | * Segment controls are considered valid, if the segment | |
2787 | * fits to the window after truncation to the window. Acceptability | |
2788 | * of data (and SYN, FIN, of course) is checked separately. | |
2789 | * See tcp_data_queue(), for example. | |
2790 | * | |
2791 | * Also, controls (RST is main one) are accepted using RCV.WUP instead | |
2792 | * of RCV.NXT. Peer still did not advance his SND.UNA when we | |
2793 | * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. | |
2794 | * (borrowed from freebsd) | |
2795 | */ | |
2796 | ||
2797 | static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq) | |
2798 | { | |
2799 | return !before(end_seq, tp->rcv_wup) && | |
2800 | !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); | |
2801 | } | |
2802 | ||
2803 | /* When we get a reset we do this. */ | |
2804 | static void tcp_reset(struct sock *sk) | |
2805 | { | |
2806 | /* We want the right error as BSD sees it (and indeed as we do). */ | |
2807 | switch (sk->sk_state) { | |
2808 | case TCP_SYN_SENT: | |
2809 | sk->sk_err = ECONNREFUSED; | |
2810 | break; | |
2811 | case TCP_CLOSE_WAIT: | |
2812 | sk->sk_err = EPIPE; | |
2813 | break; | |
2814 | case TCP_CLOSE: | |
2815 | return; | |
2816 | default: | |
2817 | sk->sk_err = ECONNRESET; | |
2818 | } | |
2819 | ||
2820 | if (!sock_flag(sk, SOCK_DEAD)) | |
2821 | sk->sk_error_report(sk); | |
2822 | ||
2823 | tcp_done(sk); | |
2824 | } | |
2825 | ||
2826 | /* | |
2827 | * Process the FIN bit. This now behaves as it is supposed to work | |
2828 | * and the FIN takes effect when it is validly part of sequence | |
2829 | * space. Not before when we get holes. | |
2830 | * | |
2831 | * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT | |
2832 | * (and thence onto LAST-ACK and finally, CLOSE, we never enter | |
2833 | * TIME-WAIT) | |
2834 | * | |
2835 | * If we are in FINWAIT-1, a received FIN indicates simultaneous | |
2836 | * close and we go into CLOSING (and later onto TIME-WAIT) | |
2837 | * | |
2838 | * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. | |
2839 | */ | |
2840 | static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th) | |
2841 | { | |
2842 | struct tcp_sock *tp = tcp_sk(sk); | |
2843 | ||
463c84b9 | 2844 | inet_csk_schedule_ack(sk); |
1da177e4 LT |
2845 | |
2846 | sk->sk_shutdown |= RCV_SHUTDOWN; | |
2847 | sock_set_flag(sk, SOCK_DONE); | |
2848 | ||
2849 | switch (sk->sk_state) { | |
2850 | case TCP_SYN_RECV: | |
2851 | case TCP_ESTABLISHED: | |
2852 | /* Move to CLOSE_WAIT */ | |
2853 | tcp_set_state(sk, TCP_CLOSE_WAIT); | |
463c84b9 | 2854 | inet_csk(sk)->icsk_ack.pingpong = 1; |
1da177e4 LT |
2855 | break; |
2856 | ||
2857 | case TCP_CLOSE_WAIT: | |
2858 | case TCP_CLOSING: | |
2859 | /* Received a retransmission of the FIN, do | |
2860 | * nothing. | |
2861 | */ | |
2862 | break; | |
2863 | case TCP_LAST_ACK: | |
2864 | /* RFC793: Remain in the LAST-ACK state. */ | |
2865 | break; | |
2866 | ||
2867 | case TCP_FIN_WAIT1: | |
2868 | /* This case occurs when a simultaneous close | |
2869 | * happens, we must ack the received FIN and | |
2870 | * enter the CLOSING state. | |
2871 | */ | |
2872 | tcp_send_ack(sk); | |
2873 | tcp_set_state(sk, TCP_CLOSING); | |
2874 | break; | |
2875 | case TCP_FIN_WAIT2: | |
2876 | /* Received a FIN -- send ACK and enter TIME_WAIT. */ | |
2877 | tcp_send_ack(sk); | |
2878 | tcp_time_wait(sk, TCP_TIME_WAIT, 0); | |
2879 | break; | |
2880 | default: | |
2881 | /* Only TCP_LISTEN and TCP_CLOSE are left, in these | |
2882 | * cases we should never reach this piece of code. | |
2883 | */ | |
2884 | printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n", | |
2885 | __FUNCTION__, sk->sk_state); | |
2886 | break; | |
2887 | }; | |
2888 | ||
2889 | /* It _is_ possible, that we have something out-of-order _after_ FIN. | |
2890 | * Probably, we should reset in this case. For now drop them. | |
2891 | */ | |
2892 | __skb_queue_purge(&tp->out_of_order_queue); | |
2893 | if (tp->rx_opt.sack_ok) | |
2894 | tcp_sack_reset(&tp->rx_opt); | |
2895 | sk_stream_mem_reclaim(sk); | |
2896 | ||
2897 | if (!sock_flag(sk, SOCK_DEAD)) { | |
2898 | sk->sk_state_change(sk); | |
2899 | ||
2900 | /* Do not send POLL_HUP for half duplex close. */ | |
2901 | if (sk->sk_shutdown == SHUTDOWN_MASK || | |
2902 | sk->sk_state == TCP_CLOSE) | |
2903 | sk_wake_async(sk, 1, POLL_HUP); | |
2904 | else | |
2905 | sk_wake_async(sk, 1, POLL_IN); | |
2906 | } | |
2907 | } | |
2908 | ||
40efc6fa | 2909 | static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq) |
1da177e4 LT |
2910 | { |
2911 | if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { | |
2912 | if (before(seq, sp->start_seq)) | |
2913 | sp->start_seq = seq; | |
2914 | if (after(end_seq, sp->end_seq)) | |
2915 | sp->end_seq = end_seq; | |
2916 | return 1; | |
2917 | } | |
2918 | return 0; | |
2919 | } | |
2920 | ||
40efc6fa | 2921 | static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq) |
1da177e4 LT |
2922 | { |
2923 | if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) { | |
2924 | if (before(seq, tp->rcv_nxt)) | |
2925 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT); | |
2926 | else | |
2927 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT); | |
2928 | ||
2929 | tp->rx_opt.dsack = 1; | |
2930 | tp->duplicate_sack[0].start_seq = seq; | |
2931 | tp->duplicate_sack[0].end_seq = end_seq; | |
2932 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok); | |
2933 | } | |
2934 | } | |
2935 | ||
40efc6fa | 2936 | static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq) |
1da177e4 LT |
2937 | { |
2938 | if (!tp->rx_opt.dsack) | |
2939 | tcp_dsack_set(tp, seq, end_seq); | |
2940 | else | |
2941 | tcp_sack_extend(tp->duplicate_sack, seq, end_seq); | |
2942 | } | |
2943 | ||
2944 | static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb) | |
2945 | { | |
2946 | struct tcp_sock *tp = tcp_sk(sk); | |
2947 | ||
2948 | if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && | |
2949 | before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
2950 | NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); | |
463c84b9 | 2951 | tcp_enter_quickack_mode(sk); |
1da177e4 LT |
2952 | |
2953 | if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) { | |
2954 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
2955 | ||
2956 | if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) | |
2957 | end_seq = tp->rcv_nxt; | |
2958 | tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq); | |
2959 | } | |
2960 | } | |
2961 | ||
2962 | tcp_send_ack(sk); | |
2963 | } | |
2964 | ||
2965 | /* These routines update the SACK block as out-of-order packets arrive or | |
2966 | * in-order packets close up the sequence space. | |
2967 | */ | |
2968 | static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) | |
2969 | { | |
2970 | int this_sack; | |
2971 | struct tcp_sack_block *sp = &tp->selective_acks[0]; | |
2972 | struct tcp_sack_block *swalk = sp+1; | |
2973 | ||
2974 | /* See if the recent change to the first SACK eats into | |
2975 | * or hits the sequence space of other SACK blocks, if so coalesce. | |
2976 | */ | |
2977 | for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) { | |
2978 | if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { | |
2979 | int i; | |
2980 | ||
2981 | /* Zap SWALK, by moving every further SACK up by one slot. | |
2982 | * Decrease num_sacks. | |
2983 | */ | |
2984 | tp->rx_opt.num_sacks--; | |
2985 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); | |
2986 | for(i=this_sack; i < tp->rx_opt.num_sacks; i++) | |
2987 | sp[i] = sp[i+1]; | |
2988 | continue; | |
2989 | } | |
2990 | this_sack++, swalk++; | |
2991 | } | |
2992 | } | |
2993 | ||
40efc6fa | 2994 | static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2) |
1da177e4 LT |
2995 | { |
2996 | __u32 tmp; | |
2997 | ||
2998 | tmp = sack1->start_seq; | |
2999 | sack1->start_seq = sack2->start_seq; | |
3000 | sack2->start_seq = tmp; | |
3001 | ||
3002 | tmp = sack1->end_seq; | |
3003 | sack1->end_seq = sack2->end_seq; | |
3004 | sack2->end_seq = tmp; | |
3005 | } | |
3006 | ||
3007 | static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) | |
3008 | { | |
3009 | struct tcp_sock *tp = tcp_sk(sk); | |
3010 | struct tcp_sack_block *sp = &tp->selective_acks[0]; | |
3011 | int cur_sacks = tp->rx_opt.num_sacks; | |
3012 | int this_sack; | |
3013 | ||
3014 | if (!cur_sacks) | |
3015 | goto new_sack; | |
3016 | ||
3017 | for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) { | |
3018 | if (tcp_sack_extend(sp, seq, end_seq)) { | |
3019 | /* Rotate this_sack to the first one. */ | |
3020 | for (; this_sack>0; this_sack--, sp--) | |
3021 | tcp_sack_swap(sp, sp-1); | |
3022 | if (cur_sacks > 1) | |
3023 | tcp_sack_maybe_coalesce(tp); | |
3024 | return; | |
3025 | } | |
3026 | } | |
3027 | ||
3028 | /* Could not find an adjacent existing SACK, build a new one, | |
3029 | * put it at the front, and shift everyone else down. We | |
3030 | * always know there is at least one SACK present already here. | |
3031 | * | |
3032 | * If the sack array is full, forget about the last one. | |
3033 | */ | |
3034 | if (this_sack >= 4) { | |
3035 | this_sack--; | |
3036 | tp->rx_opt.num_sacks--; | |
3037 | sp--; | |
3038 | } | |
3039 | for(; this_sack > 0; this_sack--, sp--) | |
3040 | *sp = *(sp-1); | |
3041 | ||
3042 | new_sack: | |
3043 | /* Build the new head SACK, and we're done. */ | |
3044 | sp->start_seq = seq; | |
3045 | sp->end_seq = end_seq; | |
3046 | tp->rx_opt.num_sacks++; | |
3047 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); | |
3048 | } | |
3049 | ||
3050 | /* RCV.NXT advances, some SACKs should be eaten. */ | |
3051 | ||
3052 | static void tcp_sack_remove(struct tcp_sock *tp) | |
3053 | { | |
3054 | struct tcp_sack_block *sp = &tp->selective_acks[0]; | |
3055 | int num_sacks = tp->rx_opt.num_sacks; | |
3056 | int this_sack; | |
3057 | ||
3058 | /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ | |
b03efcfb | 3059 | if (skb_queue_empty(&tp->out_of_order_queue)) { |
1da177e4 LT |
3060 | tp->rx_opt.num_sacks = 0; |
3061 | tp->rx_opt.eff_sacks = tp->rx_opt.dsack; | |
3062 | return; | |
3063 | } | |
3064 | ||
3065 | for(this_sack = 0; this_sack < num_sacks; ) { | |
3066 | /* Check if the start of the sack is covered by RCV.NXT. */ | |
3067 | if (!before(tp->rcv_nxt, sp->start_seq)) { | |
3068 | int i; | |
3069 | ||
3070 | /* RCV.NXT must cover all the block! */ | |
3071 | BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq)); | |
3072 | ||
3073 | /* Zap this SACK, by moving forward any other SACKS. */ | |
3074 | for (i=this_sack+1; i < num_sacks; i++) | |
3075 | tp->selective_acks[i-1] = tp->selective_acks[i]; | |
3076 | num_sacks--; | |
3077 | continue; | |
3078 | } | |
3079 | this_sack++; | |
3080 | sp++; | |
3081 | } | |
3082 | if (num_sacks != tp->rx_opt.num_sacks) { | |
3083 | tp->rx_opt.num_sacks = num_sacks; | |
3084 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); | |
3085 | } | |
3086 | } | |
3087 | ||
3088 | /* This one checks to see if we can put data from the | |
3089 | * out_of_order queue into the receive_queue. | |
3090 | */ | |
3091 | static void tcp_ofo_queue(struct sock *sk) | |
3092 | { | |
3093 | struct tcp_sock *tp = tcp_sk(sk); | |
3094 | __u32 dsack_high = tp->rcv_nxt; | |
3095 | struct sk_buff *skb; | |
3096 | ||
3097 | while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { | |
3098 | if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) | |
3099 | break; | |
3100 | ||
3101 | if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { | |
3102 | __u32 dsack = dsack_high; | |
3103 | if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) | |
3104 | dsack_high = TCP_SKB_CB(skb)->end_seq; | |
3105 | tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack); | |
3106 | } | |
3107 | ||
3108 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { | |
3109 | SOCK_DEBUG(sk, "ofo packet was already received \n"); | |
8728b834 | 3110 | __skb_unlink(skb, &tp->out_of_order_queue); |
1da177e4 LT |
3111 | __kfree_skb(skb); |
3112 | continue; | |
3113 | } | |
3114 | SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", | |
3115 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, | |
3116 | TCP_SKB_CB(skb)->end_seq); | |
3117 | ||
8728b834 | 3118 | __skb_unlink(skb, &tp->out_of_order_queue); |
1da177e4 LT |
3119 | __skb_queue_tail(&sk->sk_receive_queue, skb); |
3120 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
3121 | if(skb->h.th->fin) | |
3122 | tcp_fin(skb, sk, skb->h.th); | |
3123 | } | |
3124 | } | |
3125 | ||
3126 | static int tcp_prune_queue(struct sock *sk); | |
3127 | ||
3128 | static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) | |
3129 | { | |
3130 | struct tcphdr *th = skb->h.th; | |
3131 | struct tcp_sock *tp = tcp_sk(sk); | |
3132 | int eaten = -1; | |
3133 | ||
3134 | if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) | |
3135 | goto drop; | |
3136 | ||
1da177e4 LT |
3137 | __skb_pull(skb, th->doff*4); |
3138 | ||
3139 | TCP_ECN_accept_cwr(tp, skb); | |
3140 | ||
3141 | if (tp->rx_opt.dsack) { | |
3142 | tp->rx_opt.dsack = 0; | |
3143 | tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks, | |
3144 | 4 - tp->rx_opt.tstamp_ok); | |
3145 | } | |
3146 | ||
3147 | /* Queue data for delivery to the user. | |
3148 | * Packets in sequence go to the receive queue. | |
3149 | * Out of sequence packets to the out_of_order_queue. | |
3150 | */ | |
3151 | if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { | |
3152 | if (tcp_receive_window(tp) == 0) | |
3153 | goto out_of_window; | |
3154 | ||
3155 | /* Ok. In sequence. In window. */ | |
3156 | if (tp->ucopy.task == current && | |
3157 | tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && | |
3158 | sock_owned_by_user(sk) && !tp->urg_data) { | |
3159 | int chunk = min_t(unsigned int, skb->len, | |
3160 | tp->ucopy.len); | |
3161 | ||
3162 | __set_current_state(TASK_RUNNING); | |
3163 | ||
3164 | local_bh_enable(); | |
3165 | if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { | |
3166 | tp->ucopy.len -= chunk; | |
3167 | tp->copied_seq += chunk; | |
3168 | eaten = (chunk == skb->len && !th->fin); | |
3169 | tcp_rcv_space_adjust(sk); | |
3170 | } | |
3171 | local_bh_disable(); | |
3172 | } | |
3173 | ||
3174 | if (eaten <= 0) { | |
3175 | queue_and_out: | |
3176 | if (eaten < 0 && | |
3177 | (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || | |
3178 | !sk_stream_rmem_schedule(sk, skb))) { | |
3179 | if (tcp_prune_queue(sk) < 0 || | |
3180 | !sk_stream_rmem_schedule(sk, skb)) | |
3181 | goto drop; | |
3182 | } | |
3183 | sk_stream_set_owner_r(skb, sk); | |
3184 | __skb_queue_tail(&sk->sk_receive_queue, skb); | |
3185 | } | |
3186 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
3187 | if(skb->len) | |
3188 | tcp_event_data_recv(sk, tp, skb); | |
3189 | if(th->fin) | |
3190 | tcp_fin(skb, sk, th); | |
3191 | ||
b03efcfb | 3192 | if (!skb_queue_empty(&tp->out_of_order_queue)) { |
1da177e4 LT |
3193 | tcp_ofo_queue(sk); |
3194 | ||
3195 | /* RFC2581. 4.2. SHOULD send immediate ACK, when | |
3196 | * gap in queue is filled. | |
3197 | */ | |
b03efcfb | 3198 | if (skb_queue_empty(&tp->out_of_order_queue)) |
463c84b9 | 3199 | inet_csk(sk)->icsk_ack.pingpong = 0; |
1da177e4 LT |
3200 | } |
3201 | ||
3202 | if (tp->rx_opt.num_sacks) | |
3203 | tcp_sack_remove(tp); | |
3204 | ||
3205 | tcp_fast_path_check(sk, tp); | |
3206 | ||
3207 | if (eaten > 0) | |
3208 | __kfree_skb(skb); | |
3209 | else if (!sock_flag(sk, SOCK_DEAD)) | |
3210 | sk->sk_data_ready(sk, 0); | |
3211 | return; | |
3212 | } | |
3213 | ||
3214 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { | |
3215 | /* A retransmit, 2nd most common case. Force an immediate ack. */ | |
3216 | NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); | |
3217 | tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); | |
3218 | ||
3219 | out_of_window: | |
463c84b9 ACM |
3220 | tcp_enter_quickack_mode(sk); |
3221 | inet_csk_schedule_ack(sk); | |
1da177e4 LT |
3222 | drop: |
3223 | __kfree_skb(skb); | |
3224 | return; | |
3225 | } | |
3226 | ||
3227 | /* Out of window. F.e. zero window probe. */ | |
3228 | if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) | |
3229 | goto out_of_window; | |
3230 | ||
463c84b9 | 3231 | tcp_enter_quickack_mode(sk); |
1da177e4 LT |
3232 | |
3233 | if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
3234 | /* Partial packet, seq < rcv_next < end_seq */ | |
3235 | SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", | |
3236 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, | |
3237 | TCP_SKB_CB(skb)->end_seq); | |
3238 | ||
3239 | tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); | |
3240 | ||
3241 | /* If window is closed, drop tail of packet. But after | |
3242 | * remembering D-SACK for its head made in previous line. | |
3243 | */ | |
3244 | if (!tcp_receive_window(tp)) | |
3245 | goto out_of_window; | |
3246 | goto queue_and_out; | |
3247 | } | |
3248 | ||
3249 | TCP_ECN_check_ce(tp, skb); | |
3250 | ||
3251 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || | |
3252 | !sk_stream_rmem_schedule(sk, skb)) { | |
3253 | if (tcp_prune_queue(sk) < 0 || | |
3254 | !sk_stream_rmem_schedule(sk, skb)) | |
3255 | goto drop; | |
3256 | } | |
3257 | ||
3258 | /* Disable header prediction. */ | |
3259 | tp->pred_flags = 0; | |
463c84b9 | 3260 | inet_csk_schedule_ack(sk); |
1da177e4 LT |
3261 | |
3262 | SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", | |
3263 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); | |
3264 | ||
3265 | sk_stream_set_owner_r(skb, sk); | |
3266 | ||
3267 | if (!skb_peek(&tp->out_of_order_queue)) { | |
3268 | /* Initial out of order segment, build 1 SACK. */ | |
3269 | if (tp->rx_opt.sack_ok) { | |
3270 | tp->rx_opt.num_sacks = 1; | |
3271 | tp->rx_opt.dsack = 0; | |
3272 | tp->rx_opt.eff_sacks = 1; | |
3273 | tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; | |
3274 | tp->selective_acks[0].end_seq = | |
3275 | TCP_SKB_CB(skb)->end_seq; | |
3276 | } | |
3277 | __skb_queue_head(&tp->out_of_order_queue,skb); | |
3278 | } else { | |
3279 | struct sk_buff *skb1 = tp->out_of_order_queue.prev; | |
3280 | u32 seq = TCP_SKB_CB(skb)->seq; | |
3281 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
3282 | ||
3283 | if (seq == TCP_SKB_CB(skb1)->end_seq) { | |
8728b834 | 3284 | __skb_append(skb1, skb, &tp->out_of_order_queue); |
1da177e4 LT |
3285 | |
3286 | if (!tp->rx_opt.num_sacks || | |
3287 | tp->selective_acks[0].end_seq != seq) | |
3288 | goto add_sack; | |
3289 | ||
3290 | /* Common case: data arrive in order after hole. */ | |
3291 | tp->selective_acks[0].end_seq = end_seq; | |
3292 | return; | |
3293 | } | |
3294 | ||
3295 | /* Find place to insert this segment. */ | |
3296 | do { | |
3297 | if (!after(TCP_SKB_CB(skb1)->seq, seq)) | |
3298 | break; | |
3299 | } while ((skb1 = skb1->prev) != | |
3300 | (struct sk_buff*)&tp->out_of_order_queue); | |
3301 | ||
3302 | /* Do skb overlap to previous one? */ | |
3303 | if (skb1 != (struct sk_buff*)&tp->out_of_order_queue && | |
3304 | before(seq, TCP_SKB_CB(skb1)->end_seq)) { | |
3305 | if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { | |
3306 | /* All the bits are present. Drop. */ | |
3307 | __kfree_skb(skb); | |
3308 | tcp_dsack_set(tp, seq, end_seq); | |
3309 | goto add_sack; | |
3310 | } | |
3311 | if (after(seq, TCP_SKB_CB(skb1)->seq)) { | |
3312 | /* Partial overlap. */ | |
3313 | tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq); | |
3314 | } else { | |
3315 | skb1 = skb1->prev; | |
3316 | } | |
3317 | } | |
3318 | __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue); | |
3319 | ||
3320 | /* And clean segments covered by new one as whole. */ | |
3321 | while ((skb1 = skb->next) != | |
3322 | (struct sk_buff*)&tp->out_of_order_queue && | |
3323 | after(end_seq, TCP_SKB_CB(skb1)->seq)) { | |
3324 | if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { | |
3325 | tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq); | |
3326 | break; | |
3327 | } | |
8728b834 | 3328 | __skb_unlink(skb1, &tp->out_of_order_queue); |
1da177e4 LT |
3329 | tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq); |
3330 | __kfree_skb(skb1); | |
3331 | } | |
3332 | ||
3333 | add_sack: | |
3334 | if (tp->rx_opt.sack_ok) | |
3335 | tcp_sack_new_ofo_skb(sk, seq, end_seq); | |
3336 | } | |
3337 | } | |
3338 | ||
3339 | /* Collapse contiguous sequence of skbs head..tail with | |
3340 | * sequence numbers start..end. | |
3341 | * Segments with FIN/SYN are not collapsed (only because this | |
3342 | * simplifies code) | |
3343 | */ | |
3344 | static void | |
8728b834 DM |
3345 | tcp_collapse(struct sock *sk, struct sk_buff_head *list, |
3346 | struct sk_buff *head, struct sk_buff *tail, | |
3347 | u32 start, u32 end) | |
1da177e4 LT |
3348 | { |
3349 | struct sk_buff *skb; | |
3350 | ||
caa20d9a | 3351 | /* First, check that queue is collapsible and find |
1da177e4 LT |
3352 | * the point where collapsing can be useful. */ |
3353 | for (skb = head; skb != tail; ) { | |
3354 | /* No new bits? It is possible on ofo queue. */ | |
3355 | if (!before(start, TCP_SKB_CB(skb)->end_seq)) { | |
3356 | struct sk_buff *next = skb->next; | |
8728b834 | 3357 | __skb_unlink(skb, list); |
1da177e4 LT |
3358 | __kfree_skb(skb); |
3359 | NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); | |
3360 | skb = next; | |
3361 | continue; | |
3362 | } | |
3363 | ||
3364 | /* The first skb to collapse is: | |
3365 | * - not SYN/FIN and | |
3366 | * - bloated or contains data before "start" or | |
3367 | * overlaps to the next one. | |
3368 | */ | |
3369 | if (!skb->h.th->syn && !skb->h.th->fin && | |
3370 | (tcp_win_from_space(skb->truesize) > skb->len || | |
3371 | before(TCP_SKB_CB(skb)->seq, start) || | |
3372 | (skb->next != tail && | |
3373 | TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq))) | |
3374 | break; | |
3375 | ||
3376 | /* Decided to skip this, advance start seq. */ | |
3377 | start = TCP_SKB_CB(skb)->end_seq; | |
3378 | skb = skb->next; | |
3379 | } | |
3380 | if (skb == tail || skb->h.th->syn || skb->h.th->fin) | |
3381 | return; | |
3382 | ||
3383 | while (before(start, end)) { | |
3384 | struct sk_buff *nskb; | |
3385 | int header = skb_headroom(skb); | |
3386 | int copy = SKB_MAX_ORDER(header, 0); | |
3387 | ||
3388 | /* Too big header? This can happen with IPv6. */ | |
3389 | if (copy < 0) | |
3390 | return; | |
3391 | if (end-start < copy) | |
3392 | copy = end-start; | |
3393 | nskb = alloc_skb(copy+header, GFP_ATOMIC); | |
3394 | if (!nskb) | |
3395 | return; | |
3396 | skb_reserve(nskb, header); | |
3397 | memcpy(nskb->head, skb->head, header); | |
3398 | nskb->nh.raw = nskb->head + (skb->nh.raw-skb->head); | |
3399 | nskb->h.raw = nskb->head + (skb->h.raw-skb->head); | |
3400 | nskb->mac.raw = nskb->head + (skb->mac.raw-skb->head); | |
3401 | memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); | |
3402 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; | |
8728b834 | 3403 | __skb_insert(nskb, skb->prev, skb, list); |
1da177e4 LT |
3404 | sk_stream_set_owner_r(nskb, sk); |
3405 | ||
3406 | /* Copy data, releasing collapsed skbs. */ | |
3407 | while (copy > 0) { | |
3408 | int offset = start - TCP_SKB_CB(skb)->seq; | |
3409 | int size = TCP_SKB_CB(skb)->end_seq - start; | |
3410 | ||
09a62660 | 3411 | BUG_ON(offset < 0); |
1da177e4 LT |
3412 | if (size > 0) { |
3413 | size = min(copy, size); | |
3414 | if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) | |
3415 | BUG(); | |
3416 | TCP_SKB_CB(nskb)->end_seq += size; | |
3417 | copy -= size; | |
3418 | start += size; | |
3419 | } | |
3420 | if (!before(start, TCP_SKB_CB(skb)->end_seq)) { | |
3421 | struct sk_buff *next = skb->next; | |
8728b834 | 3422 | __skb_unlink(skb, list); |
1da177e4 LT |
3423 | __kfree_skb(skb); |
3424 | NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); | |
3425 | skb = next; | |
3426 | if (skb == tail || skb->h.th->syn || skb->h.th->fin) | |
3427 | return; | |
3428 | } | |
3429 | } | |
3430 | } | |
3431 | } | |
3432 | ||
3433 | /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs | |
3434 | * and tcp_collapse() them until all the queue is collapsed. | |
3435 | */ | |
3436 | static void tcp_collapse_ofo_queue(struct sock *sk) | |
3437 | { | |
3438 | struct tcp_sock *tp = tcp_sk(sk); | |
3439 | struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); | |
3440 | struct sk_buff *head; | |
3441 | u32 start, end; | |
3442 | ||
3443 | if (skb == NULL) | |
3444 | return; | |
3445 | ||
3446 | start = TCP_SKB_CB(skb)->seq; | |
3447 | end = TCP_SKB_CB(skb)->end_seq; | |
3448 | head = skb; | |
3449 | ||
3450 | for (;;) { | |
3451 | skb = skb->next; | |
3452 | ||
3453 | /* Segment is terminated when we see gap or when | |
3454 | * we are at the end of all the queue. */ | |
3455 | if (skb == (struct sk_buff *)&tp->out_of_order_queue || | |
3456 | after(TCP_SKB_CB(skb)->seq, end) || | |
3457 | before(TCP_SKB_CB(skb)->end_seq, start)) { | |
8728b834 DM |
3458 | tcp_collapse(sk, &tp->out_of_order_queue, |
3459 | head, skb, start, end); | |
1da177e4 LT |
3460 | head = skb; |
3461 | if (skb == (struct sk_buff *)&tp->out_of_order_queue) | |
3462 | break; | |
3463 | /* Start new segment */ | |
3464 | start = TCP_SKB_CB(skb)->seq; | |
3465 | end = TCP_SKB_CB(skb)->end_seq; | |
3466 | } else { | |
3467 | if (before(TCP_SKB_CB(skb)->seq, start)) | |
3468 | start = TCP_SKB_CB(skb)->seq; | |
3469 | if (after(TCP_SKB_CB(skb)->end_seq, end)) | |
3470 | end = TCP_SKB_CB(skb)->end_seq; | |
3471 | } | |
3472 | } | |
3473 | } | |
3474 | ||
3475 | /* Reduce allocated memory if we can, trying to get | |
3476 | * the socket within its memory limits again. | |
3477 | * | |
3478 | * Return less than zero if we should start dropping frames | |
3479 | * until the socket owning process reads some of the data | |
3480 | * to stabilize the situation. | |
3481 | */ | |
3482 | static int tcp_prune_queue(struct sock *sk) | |
3483 | { | |
3484 | struct tcp_sock *tp = tcp_sk(sk); | |
3485 | ||
3486 | SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); | |
3487 | ||
3488 | NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED); | |
3489 | ||
3490 | if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) | |
3491 | tcp_clamp_window(sk, tp); | |
3492 | else if (tcp_memory_pressure) | |
3493 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); | |
3494 | ||
3495 | tcp_collapse_ofo_queue(sk); | |
8728b834 DM |
3496 | tcp_collapse(sk, &sk->sk_receive_queue, |
3497 | sk->sk_receive_queue.next, | |
1da177e4 LT |
3498 | (struct sk_buff*)&sk->sk_receive_queue, |
3499 | tp->copied_seq, tp->rcv_nxt); | |
3500 | sk_stream_mem_reclaim(sk); | |
3501 | ||
3502 | if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) | |
3503 | return 0; | |
3504 | ||
3505 | /* Collapsing did not help, destructive actions follow. | |
3506 | * This must not ever occur. */ | |
3507 | ||
3508 | /* First, purge the out_of_order queue. */ | |
b03efcfb DM |
3509 | if (!skb_queue_empty(&tp->out_of_order_queue)) { |
3510 | NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED); | |
1da177e4 LT |
3511 | __skb_queue_purge(&tp->out_of_order_queue); |
3512 | ||
3513 | /* Reset SACK state. A conforming SACK implementation will | |
3514 | * do the same at a timeout based retransmit. When a connection | |
3515 | * is in a sad state like this, we care only about integrity | |
3516 | * of the connection not performance. | |
3517 | */ | |
3518 | if (tp->rx_opt.sack_ok) | |
3519 | tcp_sack_reset(&tp->rx_opt); | |
3520 | sk_stream_mem_reclaim(sk); | |
3521 | } | |
3522 | ||
3523 | if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) | |
3524 | return 0; | |
3525 | ||
3526 | /* If we are really being abused, tell the caller to silently | |
3527 | * drop receive data on the floor. It will get retransmitted | |
3528 | * and hopefully then we'll have sufficient space. | |
3529 | */ | |
3530 | NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED); | |
3531 | ||
3532 | /* Massive buffer overcommit. */ | |
3533 | tp->pred_flags = 0; | |
3534 | return -1; | |
3535 | } | |
3536 | ||
3537 | ||
3538 | /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. | |
3539 | * As additional protections, we do not touch cwnd in retransmission phases, | |
3540 | * and if application hit its sndbuf limit recently. | |
3541 | */ | |
3542 | void tcp_cwnd_application_limited(struct sock *sk) | |
3543 | { | |
3544 | struct tcp_sock *tp = tcp_sk(sk); | |
3545 | ||
6687e988 | 3546 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && |
1da177e4 LT |
3547 | sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { |
3548 | /* Limited by application or receiver window. */ | |
d254bcdb IJ |
3549 | u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); |
3550 | u32 win_used = max(tp->snd_cwnd_used, init_win); | |
1da177e4 | 3551 | if (win_used < tp->snd_cwnd) { |
6687e988 | 3552 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
1da177e4 LT |
3553 | tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; |
3554 | } | |
3555 | tp->snd_cwnd_used = 0; | |
3556 | } | |
3557 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
3558 | } | |
3559 | ||
40efc6fa | 3560 | static int tcp_should_expand_sndbuf(struct sock *sk, struct tcp_sock *tp) |
0d9901df DM |
3561 | { |
3562 | /* If the user specified a specific send buffer setting, do | |
3563 | * not modify it. | |
3564 | */ | |
3565 | if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) | |
3566 | return 0; | |
3567 | ||
3568 | /* If we are under global TCP memory pressure, do not expand. */ | |
3569 | if (tcp_memory_pressure) | |
3570 | return 0; | |
3571 | ||
3572 | /* If we are under soft global TCP memory pressure, do not expand. */ | |
3573 | if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0]) | |
3574 | return 0; | |
3575 | ||
3576 | /* If we filled the congestion window, do not expand. */ | |
3577 | if (tp->packets_out >= tp->snd_cwnd) | |
3578 | return 0; | |
3579 | ||
3580 | return 1; | |
3581 | } | |
1da177e4 LT |
3582 | |
3583 | /* When incoming ACK allowed to free some skb from write_queue, | |
3584 | * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket | |
3585 | * on the exit from tcp input handler. | |
3586 | * | |
3587 | * PROBLEM: sndbuf expansion does not work well with largesend. | |
3588 | */ | |
3589 | static void tcp_new_space(struct sock *sk) | |
3590 | { | |
3591 | struct tcp_sock *tp = tcp_sk(sk); | |
3592 | ||
0d9901df | 3593 | if (tcp_should_expand_sndbuf(sk, tp)) { |
c1b4a7e6 | 3594 | int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + |
1da177e4 LT |
3595 | MAX_TCP_HEADER + 16 + sizeof(struct sk_buff), |
3596 | demanded = max_t(unsigned int, tp->snd_cwnd, | |
3597 | tp->reordering + 1); | |
3598 | sndmem *= 2*demanded; | |
3599 | if (sndmem > sk->sk_sndbuf) | |
3600 | sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); | |
3601 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
3602 | } | |
3603 | ||
3604 | sk->sk_write_space(sk); | |
3605 | } | |
3606 | ||
40efc6fa | 3607 | static void tcp_check_space(struct sock *sk) |
1da177e4 LT |
3608 | { |
3609 | if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { | |
3610 | sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); | |
3611 | if (sk->sk_socket && | |
3612 | test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) | |
3613 | tcp_new_space(sk); | |
3614 | } | |
3615 | } | |
3616 | ||
40efc6fa | 3617 | static inline void tcp_data_snd_check(struct sock *sk, struct tcp_sock *tp) |
1da177e4 | 3618 | { |
55c97f3e | 3619 | tcp_push_pending_frames(sk, tp); |
1da177e4 LT |
3620 | tcp_check_space(sk); |
3621 | } | |
3622 | ||
3623 | /* | |
3624 | * Check if sending an ack is needed. | |
3625 | */ | |
3626 | static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) | |
3627 | { | |
3628 | struct tcp_sock *tp = tcp_sk(sk); | |
3629 | ||
3630 | /* More than one full frame received... */ | |
463c84b9 | 3631 | if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss |
1da177e4 LT |
3632 | /* ... and right edge of window advances far enough. |
3633 | * (tcp_recvmsg() will send ACK otherwise). Or... | |
3634 | */ | |
3635 | && __tcp_select_window(sk) >= tp->rcv_wnd) || | |
3636 | /* We ACK each frame or... */ | |
463c84b9 | 3637 | tcp_in_quickack_mode(sk) || |
1da177e4 LT |
3638 | /* We have out of order data. */ |
3639 | (ofo_possible && | |
3640 | skb_peek(&tp->out_of_order_queue))) { | |
3641 | /* Then ack it now */ | |
3642 | tcp_send_ack(sk); | |
3643 | } else { | |
3644 | /* Else, send delayed ack. */ | |
3645 | tcp_send_delayed_ack(sk); | |
3646 | } | |
3647 | } | |
3648 | ||
40efc6fa | 3649 | static inline void tcp_ack_snd_check(struct sock *sk) |
1da177e4 | 3650 | { |
463c84b9 | 3651 | if (!inet_csk_ack_scheduled(sk)) { |
1da177e4 LT |
3652 | /* We sent a data segment already. */ |
3653 | return; | |
3654 | } | |
3655 | __tcp_ack_snd_check(sk, 1); | |
3656 | } | |
3657 | ||
3658 | /* | |
3659 | * This routine is only called when we have urgent data | |
caa20d9a | 3660 | * signaled. Its the 'slow' part of tcp_urg. It could be |
1da177e4 LT |
3661 | * moved inline now as tcp_urg is only called from one |
3662 | * place. We handle URGent data wrong. We have to - as | |
3663 | * BSD still doesn't use the correction from RFC961. | |
3664 | * For 1003.1g we should support a new option TCP_STDURG to permit | |
3665 | * either form (or just set the sysctl tcp_stdurg). | |
3666 | */ | |
3667 | ||
3668 | static void tcp_check_urg(struct sock * sk, struct tcphdr * th) | |
3669 | { | |
3670 | struct tcp_sock *tp = tcp_sk(sk); | |
3671 | u32 ptr = ntohs(th->urg_ptr); | |
3672 | ||
3673 | if (ptr && !sysctl_tcp_stdurg) | |
3674 | ptr--; | |
3675 | ptr += ntohl(th->seq); | |
3676 | ||
3677 | /* Ignore urgent data that we've already seen and read. */ | |
3678 | if (after(tp->copied_seq, ptr)) | |
3679 | return; | |
3680 | ||
3681 | /* Do not replay urg ptr. | |
3682 | * | |
3683 | * NOTE: interesting situation not covered by specs. | |
3684 | * Misbehaving sender may send urg ptr, pointing to segment, | |
3685 | * which we already have in ofo queue. We are not able to fetch | |
3686 | * such data and will stay in TCP_URG_NOTYET until will be eaten | |
3687 | * by recvmsg(). Seems, we are not obliged to handle such wicked | |
3688 | * situations. But it is worth to think about possibility of some | |
3689 | * DoSes using some hypothetical application level deadlock. | |
3690 | */ | |
3691 | if (before(ptr, tp->rcv_nxt)) | |
3692 | return; | |
3693 | ||
3694 | /* Do we already have a newer (or duplicate) urgent pointer? */ | |
3695 | if (tp->urg_data && !after(ptr, tp->urg_seq)) | |
3696 | return; | |
3697 | ||
3698 | /* Tell the world about our new urgent pointer. */ | |
3699 | sk_send_sigurg(sk); | |
3700 | ||
3701 | /* We may be adding urgent data when the last byte read was | |
3702 | * urgent. To do this requires some care. We cannot just ignore | |
3703 | * tp->copied_seq since we would read the last urgent byte again | |
3704 | * as data, nor can we alter copied_seq until this data arrives | |
caa20d9a | 3705 | * or we break the semantics of SIOCATMARK (and thus sockatmark()) |
1da177e4 LT |
3706 | * |
3707 | * NOTE. Double Dutch. Rendering to plain English: author of comment | |
3708 | * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); | |
3709 | * and expect that both A and B disappear from stream. This is _wrong_. | |
3710 | * Though this happens in BSD with high probability, this is occasional. | |
3711 | * Any application relying on this is buggy. Note also, that fix "works" | |
3712 | * only in this artificial test. Insert some normal data between A and B and we will | |
3713 | * decline of BSD again. Verdict: it is better to remove to trap | |
3714 | * buggy users. | |
3715 | */ | |
3716 | if (tp->urg_seq == tp->copied_seq && tp->urg_data && | |
3717 | !sock_flag(sk, SOCK_URGINLINE) && | |
3718 | tp->copied_seq != tp->rcv_nxt) { | |
3719 | struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); | |
3720 | tp->copied_seq++; | |
3721 | if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { | |
8728b834 | 3722 | __skb_unlink(skb, &sk->sk_receive_queue); |
1da177e4 LT |
3723 | __kfree_skb(skb); |
3724 | } | |
3725 | } | |
3726 | ||
3727 | tp->urg_data = TCP_URG_NOTYET; | |
3728 | tp->urg_seq = ptr; | |
3729 | ||
3730 | /* Disable header prediction. */ | |
3731 | tp->pred_flags = 0; | |
3732 | } | |
3733 | ||
3734 | /* This is the 'fast' part of urgent handling. */ | |
3735 | static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th) | |
3736 | { | |
3737 | struct tcp_sock *tp = tcp_sk(sk); | |
3738 | ||
3739 | /* Check if we get a new urgent pointer - normally not. */ | |
3740 | if (th->urg) | |
3741 | tcp_check_urg(sk,th); | |
3742 | ||
3743 | /* Do we wait for any urgent data? - normally not... */ | |
3744 | if (tp->urg_data == TCP_URG_NOTYET) { | |
3745 | u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - | |
3746 | th->syn; | |
3747 | ||
3748 | /* Is the urgent pointer pointing into this packet? */ | |
3749 | if (ptr < skb->len) { | |
3750 | u8 tmp; | |
3751 | if (skb_copy_bits(skb, ptr, &tmp, 1)) | |
3752 | BUG(); | |
3753 | tp->urg_data = TCP_URG_VALID | tmp; | |
3754 | if (!sock_flag(sk, SOCK_DEAD)) | |
3755 | sk->sk_data_ready(sk, 0); | |
3756 | } | |
3757 | } | |
3758 | } | |
3759 | ||
3760 | static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) | |
3761 | { | |
3762 | struct tcp_sock *tp = tcp_sk(sk); | |
3763 | int chunk = skb->len - hlen; | |
3764 | int err; | |
3765 | ||
3766 | local_bh_enable(); | |
3767 | if (skb->ip_summed==CHECKSUM_UNNECESSARY) | |
3768 | err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); | |
3769 | else | |
3770 | err = skb_copy_and_csum_datagram_iovec(skb, hlen, | |
3771 | tp->ucopy.iov); | |
3772 | ||
3773 | if (!err) { | |
3774 | tp->ucopy.len -= chunk; | |
3775 | tp->copied_seq += chunk; | |
3776 | tcp_rcv_space_adjust(sk); | |
3777 | } | |
3778 | ||
3779 | local_bh_disable(); | |
3780 | return err; | |
3781 | } | |
3782 | ||
3783 | static int __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) | |
3784 | { | |
3785 | int result; | |
3786 | ||
3787 | if (sock_owned_by_user(sk)) { | |
3788 | local_bh_enable(); | |
3789 | result = __tcp_checksum_complete(skb); | |
3790 | local_bh_disable(); | |
3791 | } else { | |
3792 | result = __tcp_checksum_complete(skb); | |
3793 | } | |
3794 | return result; | |
3795 | } | |
3796 | ||
40efc6fa | 3797 | static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) |
1da177e4 LT |
3798 | { |
3799 | return skb->ip_summed != CHECKSUM_UNNECESSARY && | |
3800 | __tcp_checksum_complete_user(sk, skb); | |
3801 | } | |
3802 | ||
1a2449a8 CL |
3803 | #ifdef CONFIG_NET_DMA |
3804 | static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen) | |
3805 | { | |
3806 | struct tcp_sock *tp = tcp_sk(sk); | |
3807 | int chunk = skb->len - hlen; | |
3808 | int dma_cookie; | |
3809 | int copied_early = 0; | |
3810 | ||
3811 | if (tp->ucopy.wakeup) | |
3812 | return 0; | |
3813 | ||
3814 | if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) | |
3815 | tp->ucopy.dma_chan = get_softnet_dma(); | |
3816 | ||
3817 | if (tp->ucopy.dma_chan && skb->ip_summed == CHECKSUM_UNNECESSARY) { | |
3818 | ||
3819 | dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, | |
3820 | skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list); | |
3821 | ||
3822 | if (dma_cookie < 0) | |
3823 | goto out; | |
3824 | ||
3825 | tp->ucopy.dma_cookie = dma_cookie; | |
3826 | copied_early = 1; | |
3827 | ||
3828 | tp->ucopy.len -= chunk; | |
3829 | tp->copied_seq += chunk; | |
3830 | tcp_rcv_space_adjust(sk); | |
3831 | ||
3832 | if ((tp->ucopy.len == 0) || | |
3833 | (tcp_flag_word(skb->h.th) & TCP_FLAG_PSH) || | |
3834 | (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { | |
3835 | tp->ucopy.wakeup = 1; | |
3836 | sk->sk_data_ready(sk, 0); | |
3837 | } | |
3838 | } else if (chunk > 0) { | |
3839 | tp->ucopy.wakeup = 1; | |
3840 | sk->sk_data_ready(sk, 0); | |
3841 | } | |
3842 | out: | |
3843 | return copied_early; | |
3844 | } | |
3845 | #endif /* CONFIG_NET_DMA */ | |
3846 | ||
1da177e4 LT |
3847 | /* |
3848 | * TCP receive function for the ESTABLISHED state. | |
3849 | * | |
3850 | * It is split into a fast path and a slow path. The fast path is | |
3851 | * disabled when: | |
3852 | * - A zero window was announced from us - zero window probing | |
3853 | * is only handled properly in the slow path. | |
3854 | * - Out of order segments arrived. | |
3855 | * - Urgent data is expected. | |
3856 | * - There is no buffer space left | |
3857 | * - Unexpected TCP flags/window values/header lengths are received | |
3858 | * (detected by checking the TCP header against pred_flags) | |
3859 | * - Data is sent in both directions. Fast path only supports pure senders | |
3860 | * or pure receivers (this means either the sequence number or the ack | |
3861 | * value must stay constant) | |
3862 | * - Unexpected TCP option. | |
3863 | * | |
3864 | * When these conditions are not satisfied it drops into a standard | |
3865 | * receive procedure patterned after RFC793 to handle all cases. | |
3866 | * The first three cases are guaranteed by proper pred_flags setting, | |
3867 | * the rest is checked inline. Fast processing is turned on in | |
3868 | * tcp_data_queue when everything is OK. | |
3869 | */ | |
3870 | int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, | |
3871 | struct tcphdr *th, unsigned len) | |
3872 | { | |
3873 | struct tcp_sock *tp = tcp_sk(sk); | |
3874 | ||
3875 | /* | |
3876 | * Header prediction. | |
3877 | * The code loosely follows the one in the famous | |
3878 | * "30 instruction TCP receive" Van Jacobson mail. | |
3879 | * | |
3880 | * Van's trick is to deposit buffers into socket queue | |
3881 | * on a device interrupt, to call tcp_recv function | |
3882 | * on the receive process context and checksum and copy | |
3883 | * the buffer to user space. smart... | |
3884 | * | |
3885 | * Our current scheme is not silly either but we take the | |
3886 | * extra cost of the net_bh soft interrupt processing... | |
3887 | * We do checksum and copy also but from device to kernel. | |
3888 | */ | |
3889 | ||
3890 | tp->rx_opt.saw_tstamp = 0; | |
3891 | ||
3892 | /* pred_flags is 0xS?10 << 16 + snd_wnd | |
caa20d9a | 3893 | * if header_prediction is to be made |
1da177e4 LT |
3894 | * 'S' will always be tp->tcp_header_len >> 2 |
3895 | * '?' will be 0 for the fast path, otherwise pred_flags is 0 to | |
3896 | * turn it off (when there are holes in the receive | |
3897 | * space for instance) | |
3898 | * PSH flag is ignored. | |
3899 | */ | |
3900 | ||
3901 | if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && | |
3902 | TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { | |
3903 | int tcp_header_len = tp->tcp_header_len; | |
3904 | ||
3905 | /* Timestamp header prediction: tcp_header_len | |
3906 | * is automatically equal to th->doff*4 due to pred_flags | |
3907 | * match. | |
3908 | */ | |
3909 | ||
3910 | /* Check timestamp */ | |
3911 | if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { | |
3912 | __u32 *ptr = (__u32 *)(th + 1); | |
3913 | ||
3914 | /* No? Slow path! */ | |
3915 | if (*ptr != ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | |
3916 | | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) | |
3917 | goto slow_path; | |
3918 | ||
3919 | tp->rx_opt.saw_tstamp = 1; | |
3920 | ++ptr; | |
3921 | tp->rx_opt.rcv_tsval = ntohl(*ptr); | |
3922 | ++ptr; | |
3923 | tp->rx_opt.rcv_tsecr = ntohl(*ptr); | |
3924 | ||
3925 | /* If PAWS failed, check it more carefully in slow path */ | |
3926 | if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) | |
3927 | goto slow_path; | |
3928 | ||
3929 | /* DO NOT update ts_recent here, if checksum fails | |
3930 | * and timestamp was corrupted part, it will result | |
3931 | * in a hung connection since we will drop all | |
3932 | * future packets due to the PAWS test. | |
3933 | */ | |
3934 | } | |
3935 | ||
3936 | if (len <= tcp_header_len) { | |
3937 | /* Bulk data transfer: sender */ | |
3938 | if (len == tcp_header_len) { | |
3939 | /* Predicted packet is in window by definition. | |
3940 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. | |
3941 | * Hence, check seq<=rcv_wup reduces to: | |
3942 | */ | |
3943 | if (tcp_header_len == | |
3944 | (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && | |
3945 | tp->rcv_nxt == tp->rcv_wup) | |
3946 | tcp_store_ts_recent(tp); | |
3947 | ||
1da177e4 LT |
3948 | /* We know that such packets are checksummed |
3949 | * on entry. | |
3950 | */ | |
3951 | tcp_ack(sk, skb, 0); | |
3952 | __kfree_skb(skb); | |
55c97f3e | 3953 | tcp_data_snd_check(sk, tp); |
1da177e4 LT |
3954 | return 0; |
3955 | } else { /* Header too small */ | |
3956 | TCP_INC_STATS_BH(TCP_MIB_INERRS); | |
3957 | goto discard; | |
3958 | } | |
3959 | } else { | |
3960 | int eaten = 0; | |
1a2449a8 | 3961 | int copied_early = 0; |
1da177e4 | 3962 | |
1a2449a8 CL |
3963 | if (tp->copied_seq == tp->rcv_nxt && |
3964 | len - tcp_header_len <= tp->ucopy.len) { | |
3965 | #ifdef CONFIG_NET_DMA | |
3966 | if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { | |
3967 | copied_early = 1; | |
3968 | eaten = 1; | |
3969 | } | |
3970 | #endif | |
3971 | if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) { | |
3972 | __set_current_state(TASK_RUNNING); | |
1da177e4 | 3973 | |
1a2449a8 CL |
3974 | if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) |
3975 | eaten = 1; | |
3976 | } | |
3977 | if (eaten) { | |
1da177e4 LT |
3978 | /* Predicted packet is in window by definition. |
3979 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. | |
3980 | * Hence, check seq<=rcv_wup reduces to: | |
3981 | */ | |
3982 | if (tcp_header_len == | |
3983 | (sizeof(struct tcphdr) + | |
3984 | TCPOLEN_TSTAMP_ALIGNED) && | |
3985 | tp->rcv_nxt == tp->rcv_wup) | |
3986 | tcp_store_ts_recent(tp); | |
3987 | ||
463c84b9 | 3988 | tcp_rcv_rtt_measure_ts(sk, skb); |
1da177e4 LT |
3989 | |
3990 | __skb_pull(skb, tcp_header_len); | |
3991 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
3992 | NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER); | |
1da177e4 | 3993 | } |
1a2449a8 CL |
3994 | if (copied_early) |
3995 | tcp_cleanup_rbuf(sk, skb->len); | |
1da177e4 LT |
3996 | } |
3997 | if (!eaten) { | |
3998 | if (tcp_checksum_complete_user(sk, skb)) | |
3999 | goto csum_error; | |
4000 | ||
4001 | /* Predicted packet is in window by definition. | |
4002 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. | |
4003 | * Hence, check seq<=rcv_wup reduces to: | |
4004 | */ | |
4005 | if (tcp_header_len == | |
4006 | (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && | |
4007 | tp->rcv_nxt == tp->rcv_wup) | |
4008 | tcp_store_ts_recent(tp); | |
4009 | ||
463c84b9 | 4010 | tcp_rcv_rtt_measure_ts(sk, skb); |
1da177e4 LT |
4011 | |
4012 | if ((int)skb->truesize > sk->sk_forward_alloc) | |
4013 | goto step5; | |
4014 | ||
4015 | NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS); | |
4016 | ||
4017 | /* Bulk data transfer: receiver */ | |
4018 | __skb_pull(skb,tcp_header_len); | |
4019 | __skb_queue_tail(&sk->sk_receive_queue, skb); | |
4020 | sk_stream_set_owner_r(skb, sk); | |
4021 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
4022 | } | |
4023 | ||
4024 | tcp_event_data_recv(sk, tp, skb); | |
4025 | ||
4026 | if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { | |
4027 | /* Well, only one small jumplet in fast path... */ | |
4028 | tcp_ack(sk, skb, FLAG_DATA); | |
55c97f3e | 4029 | tcp_data_snd_check(sk, tp); |
463c84b9 | 4030 | if (!inet_csk_ack_scheduled(sk)) |
1da177e4 LT |
4031 | goto no_ack; |
4032 | } | |
4033 | ||
31432412 | 4034 | __tcp_ack_snd_check(sk, 0); |
1da177e4 | 4035 | no_ack: |
1a2449a8 CL |
4036 | #ifdef CONFIG_NET_DMA |
4037 | if (copied_early) | |
4038 | __skb_queue_tail(&sk->sk_async_wait_queue, skb); | |
4039 | else | |
4040 | #endif | |
1da177e4 LT |
4041 | if (eaten) |
4042 | __kfree_skb(skb); | |
4043 | else | |
4044 | sk->sk_data_ready(sk, 0); | |
4045 | return 0; | |
4046 | } | |
4047 | } | |
4048 | ||
4049 | slow_path: | |
4050 | if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb)) | |
4051 | goto csum_error; | |
4052 | ||
4053 | /* | |
4054 | * RFC1323: H1. Apply PAWS check first. | |
4055 | */ | |
4056 | if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && | |
463c84b9 | 4057 | tcp_paws_discard(sk, skb)) { |
1da177e4 LT |
4058 | if (!th->rst) { |
4059 | NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); | |
4060 | tcp_send_dupack(sk, skb); | |
4061 | goto discard; | |
4062 | } | |
4063 | /* Resets are accepted even if PAWS failed. | |
4064 | ||
4065 | ts_recent update must be made after we are sure | |
4066 | that the packet is in window. | |
4067 | */ | |
4068 | } | |
4069 | ||
4070 | /* | |
4071 | * Standard slow path. | |
4072 | */ | |
4073 | ||
4074 | if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { | |
4075 | /* RFC793, page 37: "In all states except SYN-SENT, all reset | |
4076 | * (RST) segments are validated by checking their SEQ-fields." | |
4077 | * And page 69: "If an incoming segment is not acceptable, | |
4078 | * an acknowledgment should be sent in reply (unless the RST bit | |
4079 | * is set, if so drop the segment and return)". | |
4080 | */ | |
4081 | if (!th->rst) | |
4082 | tcp_send_dupack(sk, skb); | |
4083 | goto discard; | |
4084 | } | |
4085 | ||
4086 | if(th->rst) { | |
4087 | tcp_reset(sk); | |
4088 | goto discard; | |
4089 | } | |
4090 | ||
4091 | tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); | |
4092 | ||
4093 | if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
4094 | TCP_INC_STATS_BH(TCP_MIB_INERRS); | |
4095 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); | |
4096 | tcp_reset(sk); | |
4097 | return 1; | |
4098 | } | |
4099 | ||
4100 | step5: | |
4101 | if(th->ack) | |
4102 | tcp_ack(sk, skb, FLAG_SLOWPATH); | |
4103 | ||
463c84b9 | 4104 | tcp_rcv_rtt_measure_ts(sk, skb); |
1da177e4 LT |
4105 | |
4106 | /* Process urgent data. */ | |
4107 | tcp_urg(sk, skb, th); | |
4108 | ||
4109 | /* step 7: process the segment text */ | |
4110 | tcp_data_queue(sk, skb); | |
4111 | ||
55c97f3e | 4112 | tcp_data_snd_check(sk, tp); |
1da177e4 LT |
4113 | tcp_ack_snd_check(sk); |
4114 | return 0; | |
4115 | ||
4116 | csum_error: | |
4117 | TCP_INC_STATS_BH(TCP_MIB_INERRS); | |
4118 | ||
4119 | discard: | |
4120 | __kfree_skb(skb); | |
4121 | return 0; | |
4122 | } | |
4123 | ||
4124 | static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, | |
4125 | struct tcphdr *th, unsigned len) | |
4126 | { | |
4127 | struct tcp_sock *tp = tcp_sk(sk); | |
d83d8461 | 4128 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
4129 | int saved_clamp = tp->rx_opt.mss_clamp; |
4130 | ||
4131 | tcp_parse_options(skb, &tp->rx_opt, 0); | |
4132 | ||
4133 | if (th->ack) { | |
4134 | /* rfc793: | |
4135 | * "If the state is SYN-SENT then | |
4136 | * first check the ACK bit | |
4137 | * If the ACK bit is set | |
4138 | * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send | |
4139 | * a reset (unless the RST bit is set, if so drop | |
4140 | * the segment and return)" | |
4141 | * | |
4142 | * We do not send data with SYN, so that RFC-correct | |
4143 | * test reduces to: | |
4144 | */ | |
4145 | if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) | |
4146 | goto reset_and_undo; | |
4147 | ||
4148 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && | |
4149 | !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, | |
4150 | tcp_time_stamp)) { | |
4151 | NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED); | |
4152 | goto reset_and_undo; | |
4153 | } | |
4154 | ||
4155 | /* Now ACK is acceptable. | |
4156 | * | |
4157 | * "If the RST bit is set | |
4158 | * If the ACK was acceptable then signal the user "error: | |
4159 | * connection reset", drop the segment, enter CLOSED state, | |
4160 | * delete TCB, and return." | |
4161 | */ | |
4162 | ||
4163 | if (th->rst) { | |
4164 | tcp_reset(sk); | |
4165 | goto discard; | |
4166 | } | |
4167 | ||
4168 | /* rfc793: | |
4169 | * "fifth, if neither of the SYN or RST bits is set then | |
4170 | * drop the segment and return." | |
4171 | * | |
4172 | * See note below! | |
4173 | * --ANK(990513) | |
4174 | */ | |
4175 | if (!th->syn) | |
4176 | goto discard_and_undo; | |
4177 | ||
4178 | /* rfc793: | |
4179 | * "If the SYN bit is on ... | |
4180 | * are acceptable then ... | |
4181 | * (our SYN has been ACKed), change the connection | |
4182 | * state to ESTABLISHED..." | |
4183 | */ | |
4184 | ||
4185 | TCP_ECN_rcv_synack(tp, th); | |
1da177e4 LT |
4186 | |
4187 | tp->snd_wl1 = TCP_SKB_CB(skb)->seq; | |
4188 | tcp_ack(sk, skb, FLAG_SLOWPATH); | |
4189 | ||
4190 | /* Ok.. it's good. Set up sequence numbers and | |
4191 | * move to established. | |
4192 | */ | |
4193 | tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; | |
4194 | tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; | |
4195 | ||
4196 | /* RFC1323: The window in SYN & SYN/ACK segments is | |
4197 | * never scaled. | |
4198 | */ | |
4199 | tp->snd_wnd = ntohs(th->window); | |
4200 | tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq); | |
4201 | ||
4202 | if (!tp->rx_opt.wscale_ok) { | |
4203 | tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; | |
4204 | tp->window_clamp = min(tp->window_clamp, 65535U); | |
4205 | } | |
4206 | ||
4207 | if (tp->rx_opt.saw_tstamp) { | |
4208 | tp->rx_opt.tstamp_ok = 1; | |
4209 | tp->tcp_header_len = | |
4210 | sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; | |
4211 | tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; | |
4212 | tcp_store_ts_recent(tp); | |
4213 | } else { | |
4214 | tp->tcp_header_len = sizeof(struct tcphdr); | |
4215 | } | |
4216 | ||
4217 | if (tp->rx_opt.sack_ok && sysctl_tcp_fack) | |
4218 | tp->rx_opt.sack_ok |= 2; | |
4219 | ||
5d424d5a | 4220 | tcp_mtup_init(sk); |
d83d8461 | 4221 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); |
1da177e4 LT |
4222 | tcp_initialize_rcv_mss(sk); |
4223 | ||
4224 | /* Remember, tcp_poll() does not lock socket! | |
4225 | * Change state from SYN-SENT only after copied_seq | |
4226 | * is initialized. */ | |
4227 | tp->copied_seq = tp->rcv_nxt; | |
4228 | mb(); | |
4229 | tcp_set_state(sk, TCP_ESTABLISHED); | |
4230 | ||
4231 | /* Make sure socket is routed, for correct metrics. */ | |
8292a17a | 4232 | icsk->icsk_af_ops->rebuild_header(sk); |
1da177e4 LT |
4233 | |
4234 | tcp_init_metrics(sk); | |
4235 | ||
6687e988 | 4236 | tcp_init_congestion_control(sk); |
317a76f9 | 4237 | |
1da177e4 LT |
4238 | /* Prevent spurious tcp_cwnd_restart() on first data |
4239 | * packet. | |
4240 | */ | |
4241 | tp->lsndtime = tcp_time_stamp; | |
4242 | ||
4243 | tcp_init_buffer_space(sk); | |
4244 | ||
4245 | if (sock_flag(sk, SOCK_KEEPOPEN)) | |
463c84b9 | 4246 | inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); |
1da177e4 LT |
4247 | |
4248 | if (!tp->rx_opt.snd_wscale) | |
4249 | __tcp_fast_path_on(tp, tp->snd_wnd); | |
4250 | else | |
4251 | tp->pred_flags = 0; | |
4252 | ||
4253 | if (!sock_flag(sk, SOCK_DEAD)) { | |
4254 | sk->sk_state_change(sk); | |
4255 | sk_wake_async(sk, 0, POLL_OUT); | |
4256 | } | |
4257 | ||
295f7324 ACM |
4258 | if (sk->sk_write_pending || |
4259 | icsk->icsk_accept_queue.rskq_defer_accept || | |
4260 | icsk->icsk_ack.pingpong) { | |
1da177e4 LT |
4261 | /* Save one ACK. Data will be ready after |
4262 | * several ticks, if write_pending is set. | |
4263 | * | |
4264 | * It may be deleted, but with this feature tcpdumps | |
4265 | * look so _wonderfully_ clever, that I was not able | |
4266 | * to stand against the temptation 8) --ANK | |
4267 | */ | |
463c84b9 | 4268 | inet_csk_schedule_ack(sk); |
295f7324 ACM |
4269 | icsk->icsk_ack.lrcvtime = tcp_time_stamp; |
4270 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
463c84b9 ACM |
4271 | tcp_incr_quickack(sk); |
4272 | tcp_enter_quickack_mode(sk); | |
3f421baa ACM |
4273 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, |
4274 | TCP_DELACK_MAX, TCP_RTO_MAX); | |
1da177e4 LT |
4275 | |
4276 | discard: | |
4277 | __kfree_skb(skb); | |
4278 | return 0; | |
4279 | } else { | |
4280 | tcp_send_ack(sk); | |
4281 | } | |
4282 | return -1; | |
4283 | } | |
4284 | ||
4285 | /* No ACK in the segment */ | |
4286 | ||
4287 | if (th->rst) { | |
4288 | /* rfc793: | |
4289 | * "If the RST bit is set | |
4290 | * | |
4291 | * Otherwise (no ACK) drop the segment and return." | |
4292 | */ | |
4293 | ||
4294 | goto discard_and_undo; | |
4295 | } | |
4296 | ||
4297 | /* PAWS check. */ | |
4298 | if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0)) | |
4299 | goto discard_and_undo; | |
4300 | ||
4301 | if (th->syn) { | |
4302 | /* We see SYN without ACK. It is attempt of | |
4303 | * simultaneous connect with crossed SYNs. | |
4304 | * Particularly, it can be connect to self. | |
4305 | */ | |
4306 | tcp_set_state(sk, TCP_SYN_RECV); | |
4307 | ||
4308 | if (tp->rx_opt.saw_tstamp) { | |
4309 | tp->rx_opt.tstamp_ok = 1; | |
4310 | tcp_store_ts_recent(tp); | |
4311 | tp->tcp_header_len = | |
4312 | sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; | |
4313 | } else { | |
4314 | tp->tcp_header_len = sizeof(struct tcphdr); | |
4315 | } | |
4316 | ||
4317 | tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; | |
4318 | tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; | |
4319 | ||
4320 | /* RFC1323: The window in SYN & SYN/ACK segments is | |
4321 | * never scaled. | |
4322 | */ | |
4323 | tp->snd_wnd = ntohs(th->window); | |
4324 | tp->snd_wl1 = TCP_SKB_CB(skb)->seq; | |
4325 | tp->max_window = tp->snd_wnd; | |
4326 | ||
4327 | TCP_ECN_rcv_syn(tp, th); | |
1da177e4 | 4328 | |
5d424d5a | 4329 | tcp_mtup_init(sk); |
d83d8461 | 4330 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); |
1da177e4 LT |
4331 | tcp_initialize_rcv_mss(sk); |
4332 | ||
4333 | ||
4334 | tcp_send_synack(sk); | |
4335 | #if 0 | |
4336 | /* Note, we could accept data and URG from this segment. | |
4337 | * There are no obstacles to make this. | |
4338 | * | |
4339 | * However, if we ignore data in ACKless segments sometimes, | |
4340 | * we have no reasons to accept it sometimes. | |
4341 | * Also, seems the code doing it in step6 of tcp_rcv_state_process | |
4342 | * is not flawless. So, discard packet for sanity. | |
4343 | * Uncomment this return to process the data. | |
4344 | */ | |
4345 | return -1; | |
4346 | #else | |
4347 | goto discard; | |
4348 | #endif | |
4349 | } | |
4350 | /* "fifth, if neither of the SYN or RST bits is set then | |
4351 | * drop the segment and return." | |
4352 | */ | |
4353 | ||
4354 | discard_and_undo: | |
4355 | tcp_clear_options(&tp->rx_opt); | |
4356 | tp->rx_opt.mss_clamp = saved_clamp; | |
4357 | goto discard; | |
4358 | ||
4359 | reset_and_undo: | |
4360 | tcp_clear_options(&tp->rx_opt); | |
4361 | tp->rx_opt.mss_clamp = saved_clamp; | |
4362 | return 1; | |
4363 | } | |
4364 | ||
4365 | ||
4366 | /* | |
4367 | * This function implements the receiving procedure of RFC 793 for | |
4368 | * all states except ESTABLISHED and TIME_WAIT. | |
4369 | * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be | |
4370 | * address independent. | |
4371 | */ | |
4372 | ||
4373 | int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, | |
4374 | struct tcphdr *th, unsigned len) | |
4375 | { | |
4376 | struct tcp_sock *tp = tcp_sk(sk); | |
8292a17a | 4377 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
4378 | int queued = 0; |
4379 | ||
4380 | tp->rx_opt.saw_tstamp = 0; | |
4381 | ||
4382 | switch (sk->sk_state) { | |
4383 | case TCP_CLOSE: | |
4384 | goto discard; | |
4385 | ||
4386 | case TCP_LISTEN: | |
4387 | if(th->ack) | |
4388 | return 1; | |
4389 | ||
4390 | if(th->rst) | |
4391 | goto discard; | |
4392 | ||
4393 | if(th->syn) { | |
8292a17a | 4394 | if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) |
1da177e4 LT |
4395 | return 1; |
4396 | ||
1da177e4 LT |
4397 | /* Now we have several options: In theory there is |
4398 | * nothing else in the frame. KA9Q has an option to | |
4399 | * send data with the syn, BSD accepts data with the | |
4400 | * syn up to the [to be] advertised window and | |
4401 | * Solaris 2.1 gives you a protocol error. For now | |
4402 | * we just ignore it, that fits the spec precisely | |
4403 | * and avoids incompatibilities. It would be nice in | |
4404 | * future to drop through and process the data. | |
4405 | * | |
4406 | * Now that TTCP is starting to be used we ought to | |
4407 | * queue this data. | |
4408 | * But, this leaves one open to an easy denial of | |
4409 | * service attack, and SYN cookies can't defend | |
4410 | * against this problem. So, we drop the data | |
4411 | * in the interest of security over speed. | |
4412 | */ | |
4413 | goto discard; | |
4414 | } | |
4415 | goto discard; | |
4416 | ||
4417 | case TCP_SYN_SENT: | |
1da177e4 LT |
4418 | queued = tcp_rcv_synsent_state_process(sk, skb, th, len); |
4419 | if (queued >= 0) | |
4420 | return queued; | |
4421 | ||
4422 | /* Do step6 onward by hand. */ | |
4423 | tcp_urg(sk, skb, th); | |
4424 | __kfree_skb(skb); | |
55c97f3e | 4425 | tcp_data_snd_check(sk, tp); |
1da177e4 LT |
4426 | return 0; |
4427 | } | |
4428 | ||
4429 | if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && | |
463c84b9 | 4430 | tcp_paws_discard(sk, skb)) { |
1da177e4 LT |
4431 | if (!th->rst) { |
4432 | NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); | |
4433 | tcp_send_dupack(sk, skb); | |
4434 | goto discard; | |
4435 | } | |
4436 | /* Reset is accepted even if it did not pass PAWS. */ | |
4437 | } | |
4438 | ||
4439 | /* step 1: check sequence number */ | |
4440 | if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { | |
4441 | if (!th->rst) | |
4442 | tcp_send_dupack(sk, skb); | |
4443 | goto discard; | |
4444 | } | |
4445 | ||
4446 | /* step 2: check RST bit */ | |
4447 | if(th->rst) { | |
4448 | tcp_reset(sk); | |
4449 | goto discard; | |
4450 | } | |
4451 | ||
4452 | tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); | |
4453 | ||
4454 | /* step 3: check security and precedence [ignored] */ | |
4455 | ||
4456 | /* step 4: | |
4457 | * | |
4458 | * Check for a SYN in window. | |
4459 | */ | |
4460 | if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
4461 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); | |
4462 | tcp_reset(sk); | |
4463 | return 1; | |
4464 | } | |
4465 | ||
4466 | /* step 5: check the ACK field */ | |
4467 | if (th->ack) { | |
4468 | int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH); | |
4469 | ||
4470 | switch(sk->sk_state) { | |
4471 | case TCP_SYN_RECV: | |
4472 | if (acceptable) { | |
4473 | tp->copied_seq = tp->rcv_nxt; | |
4474 | mb(); | |
4475 | tcp_set_state(sk, TCP_ESTABLISHED); | |
4476 | sk->sk_state_change(sk); | |
4477 | ||
4478 | /* Note, that this wakeup is only for marginal | |
4479 | * crossed SYN case. Passively open sockets | |
4480 | * are not waked up, because sk->sk_sleep == | |
4481 | * NULL and sk->sk_socket == NULL. | |
4482 | */ | |
4483 | if (sk->sk_socket) { | |
4484 | sk_wake_async(sk,0,POLL_OUT); | |
4485 | } | |
4486 | ||
4487 | tp->snd_una = TCP_SKB_CB(skb)->ack_seq; | |
4488 | tp->snd_wnd = ntohs(th->window) << | |
4489 | tp->rx_opt.snd_wscale; | |
4490 | tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, | |
4491 | TCP_SKB_CB(skb)->seq); | |
4492 | ||
4493 | /* tcp_ack considers this ACK as duplicate | |
4494 | * and does not calculate rtt. | |
4495 | * Fix it at least with timestamps. | |
4496 | */ | |
4497 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && | |
4498 | !tp->srtt) | |
2d2abbab | 4499 | tcp_ack_saw_tstamp(sk, 0); |
1da177e4 LT |
4500 | |
4501 | if (tp->rx_opt.tstamp_ok) | |
4502 | tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; | |
4503 | ||
4504 | /* Make sure socket is routed, for | |
4505 | * correct metrics. | |
4506 | */ | |
8292a17a | 4507 | icsk->icsk_af_ops->rebuild_header(sk); |
1da177e4 LT |
4508 | |
4509 | tcp_init_metrics(sk); | |
4510 | ||
6687e988 | 4511 | tcp_init_congestion_control(sk); |
317a76f9 | 4512 | |
1da177e4 LT |
4513 | /* Prevent spurious tcp_cwnd_restart() on |
4514 | * first data packet. | |
4515 | */ | |
4516 | tp->lsndtime = tcp_time_stamp; | |
4517 | ||
5d424d5a | 4518 | tcp_mtup_init(sk); |
1da177e4 LT |
4519 | tcp_initialize_rcv_mss(sk); |
4520 | tcp_init_buffer_space(sk); | |
4521 | tcp_fast_path_on(tp); | |
4522 | } else { | |
4523 | return 1; | |
4524 | } | |
4525 | break; | |
4526 | ||
4527 | case TCP_FIN_WAIT1: | |
4528 | if (tp->snd_una == tp->write_seq) { | |
4529 | tcp_set_state(sk, TCP_FIN_WAIT2); | |
4530 | sk->sk_shutdown |= SEND_SHUTDOWN; | |
4531 | dst_confirm(sk->sk_dst_cache); | |
4532 | ||
4533 | if (!sock_flag(sk, SOCK_DEAD)) | |
4534 | /* Wake up lingering close() */ | |
4535 | sk->sk_state_change(sk); | |
4536 | else { | |
4537 | int tmo; | |
4538 | ||
4539 | if (tp->linger2 < 0 || | |
4540 | (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && | |
4541 | after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { | |
4542 | tcp_done(sk); | |
4543 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); | |
4544 | return 1; | |
4545 | } | |
4546 | ||
463c84b9 | 4547 | tmo = tcp_fin_time(sk); |
1da177e4 | 4548 | if (tmo > TCP_TIMEWAIT_LEN) { |
463c84b9 | 4549 | inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); |
1da177e4 LT |
4550 | } else if (th->fin || sock_owned_by_user(sk)) { |
4551 | /* Bad case. We could lose such FIN otherwise. | |
4552 | * It is not a big problem, but it looks confusing | |
4553 | * and not so rare event. We still can lose it now, | |
4554 | * if it spins in bh_lock_sock(), but it is really | |
4555 | * marginal case. | |
4556 | */ | |
463c84b9 | 4557 | inet_csk_reset_keepalive_timer(sk, tmo); |
1da177e4 LT |
4558 | } else { |
4559 | tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); | |
4560 | goto discard; | |
4561 | } | |
4562 | } | |
4563 | } | |
4564 | break; | |
4565 | ||
4566 | case TCP_CLOSING: | |
4567 | if (tp->snd_una == tp->write_seq) { | |
4568 | tcp_time_wait(sk, TCP_TIME_WAIT, 0); | |
4569 | goto discard; | |
4570 | } | |
4571 | break; | |
4572 | ||
4573 | case TCP_LAST_ACK: | |
4574 | if (tp->snd_una == tp->write_seq) { | |
4575 | tcp_update_metrics(sk); | |
4576 | tcp_done(sk); | |
4577 | goto discard; | |
4578 | } | |
4579 | break; | |
4580 | } | |
4581 | } else | |
4582 | goto discard; | |
4583 | ||
4584 | /* step 6: check the URG bit */ | |
4585 | tcp_urg(sk, skb, th); | |
4586 | ||
4587 | /* step 7: process the segment text */ | |
4588 | switch (sk->sk_state) { | |
4589 | case TCP_CLOSE_WAIT: | |
4590 | case TCP_CLOSING: | |
4591 | case TCP_LAST_ACK: | |
4592 | if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) | |
4593 | break; | |
4594 | case TCP_FIN_WAIT1: | |
4595 | case TCP_FIN_WAIT2: | |
4596 | /* RFC 793 says to queue data in these states, | |
4597 | * RFC 1122 says we MUST send a reset. | |
4598 | * BSD 4.4 also does reset. | |
4599 | */ | |
4600 | if (sk->sk_shutdown & RCV_SHUTDOWN) { | |
4601 | if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && | |
4602 | after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { | |
4603 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); | |
4604 | tcp_reset(sk); | |
4605 | return 1; | |
4606 | } | |
4607 | } | |
4608 | /* Fall through */ | |
4609 | case TCP_ESTABLISHED: | |
4610 | tcp_data_queue(sk, skb); | |
4611 | queued = 1; | |
4612 | break; | |
4613 | } | |
4614 | ||
4615 | /* tcp_data could move socket to TIME-WAIT */ | |
4616 | if (sk->sk_state != TCP_CLOSE) { | |
55c97f3e | 4617 | tcp_data_snd_check(sk, tp); |
1da177e4 LT |
4618 | tcp_ack_snd_check(sk); |
4619 | } | |
4620 | ||
4621 | if (!queued) { | |
4622 | discard: | |
4623 | __kfree_skb(skb); | |
4624 | } | |
4625 | return 0; | |
4626 | } | |
4627 | ||
4628 | EXPORT_SYMBOL(sysctl_tcp_ecn); | |
4629 | EXPORT_SYMBOL(sysctl_tcp_reordering); | |
4630 | EXPORT_SYMBOL(tcp_parse_options); | |
4631 | EXPORT_SYMBOL(tcp_rcv_established); | |
4632 | EXPORT_SYMBOL(tcp_rcv_state_process); | |
40efc6fa | 4633 | EXPORT_SYMBOL(tcp_initialize_rcv_mss); |