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