]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * INET An implementation of the TCP/IP protocol suite for the LINUX | |
3 | * operating system. INET is implemented using the BSD Socket | |
4 | * interface as the means of communication with the user level. | |
5 | * | |
6 | * Implementation of the Transmission Control Protocol(TCP). | |
7 | * | |
8 | * Version: $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $ | |
9 | * | |
02c30a84 | 10 | * Authors: Ross Biro |
1da177e4 LT |
11 | * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> |
12 | * Mark Evans, <evansmp@uhura.aston.ac.uk> | |
13 | * Corey Minyard <wf-rch!minyard@relay.EU.net> | |
14 | * Florian La Roche, <flla@stud.uni-sb.de> | |
15 | * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> | |
16 | * Linus Torvalds, <torvalds@cs.helsinki.fi> | |
17 | * Alan Cox, <gw4pts@gw4pts.ampr.org> | |
18 | * Matthew Dillon, <dillon@apollo.west.oic.com> | |
19 | * Arnt Gulbrandsen, <agulbra@nvg.unit.no> | |
20 | * Jorge Cwik, <jorge@laser.satlink.net> | |
21 | */ | |
22 | ||
23 | /* | |
24 | * Changes: | |
25 | * Pedro Roque : Fast Retransmit/Recovery. | |
26 | * Two receive queues. | |
27 | * Retransmit queue handled by TCP. | |
28 | * Better retransmit timer handling. | |
29 | * New congestion avoidance. | |
30 | * Header prediction. | |
31 | * Variable renaming. | |
32 | * | |
33 | * Eric : Fast Retransmit. | |
34 | * Randy Scott : MSS option defines. | |
35 | * Eric Schenk : Fixes to slow start algorithm. | |
36 | * Eric Schenk : Yet another double ACK bug. | |
37 | * Eric Schenk : Delayed ACK bug fixes. | |
38 | * Eric Schenk : Floyd style fast retrans war avoidance. | |
39 | * David S. Miller : Don't allow zero congestion window. | |
40 | * Eric Schenk : Fix retransmitter so that it sends | |
41 | * next packet on ack of previous packet. | |
42 | * Andi Kleen : Moved open_request checking here | |
43 | * and process RSTs for open_requests. | |
44 | * Andi Kleen : Better prune_queue, and other fixes. | |
caa20d9a | 45 | * Andrey Savochkin: Fix RTT measurements in the presence of |
1da177e4 LT |
46 | * timestamps. |
47 | * Andrey Savochkin: Check sequence numbers correctly when | |
48 | * removing SACKs due to in sequence incoming | |
49 | * data segments. | |
50 | * Andi Kleen: Make sure we never ack data there is not | |
51 | * enough room for. Also make this condition | |
52 | * a fatal error if it might still happen. | |
e905a9ed | 53 | * Andi Kleen: Add tcp_measure_rcv_mss to make |
1da177e4 | 54 | * connections with MSS<min(MTU,ann. MSS) |
e905a9ed | 55 | * work without delayed acks. |
1da177e4 LT |
56 | * Andi Kleen: Process packets with PSH set in the |
57 | * fast path. | |
58 | * J Hadi Salim: ECN support | |
59 | * Andrei Gurtov, | |
60 | * Pasi Sarolahti, | |
61 | * Panu Kuhlberg: Experimental audit of TCP (re)transmission | |
62 | * engine. Lots of bugs are found. | |
63 | * Pasi Sarolahti: F-RTO for dealing with spurious RTOs | |
1da177e4 LT |
64 | */ |
65 | ||
1da177e4 LT |
66 | #include <linux/mm.h> |
67 | #include <linux/module.h> | |
68 | #include <linux/sysctl.h> | |
69 | #include <net/tcp.h> | |
70 | #include <net/inet_common.h> | |
71 | #include <linux/ipsec.h> | |
72 | #include <asm/unaligned.h> | |
1a2449a8 | 73 | #include <net/netdma.h> |
1da177e4 | 74 | |
ab32ea5d BH |
75 | int sysctl_tcp_timestamps __read_mostly = 1; |
76 | int sysctl_tcp_window_scaling __read_mostly = 1; | |
77 | int sysctl_tcp_sack __read_mostly = 1; | |
78 | int sysctl_tcp_fack __read_mostly = 1; | |
79 | int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH; | |
80 | int sysctl_tcp_ecn __read_mostly; | |
81 | int sysctl_tcp_dsack __read_mostly = 1; | |
82 | int sysctl_tcp_app_win __read_mostly = 31; | |
83 | int sysctl_tcp_adv_win_scale __read_mostly = 2; | |
1da177e4 | 84 | |
ab32ea5d BH |
85 | int sysctl_tcp_stdurg __read_mostly; |
86 | int sysctl_tcp_rfc1337 __read_mostly; | |
87 | int sysctl_tcp_max_orphans __read_mostly = NR_FILE; | |
88 | int sysctl_tcp_frto __read_mostly; | |
3cfe3baa | 89 | int sysctl_tcp_frto_response __read_mostly; |
ab32ea5d | 90 | int sysctl_tcp_nometrics_save __read_mostly; |
1da177e4 | 91 | |
ab32ea5d BH |
92 | int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; |
93 | int sysctl_tcp_abc __read_mostly; | |
1da177e4 | 94 | |
1da177e4 LT |
95 | #define FLAG_DATA 0x01 /* Incoming frame contained data. */ |
96 | #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ | |
97 | #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ | |
98 | #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ | |
99 | #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ | |
100 | #define FLAG_DATA_SACKED 0x20 /* New SACK. */ | |
101 | #define FLAG_ECE 0x40 /* ECE in this ACK */ | |
102 | #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */ | |
103 | #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ | |
4dc2665e | 104 | #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */ |
2e605294 | 105 | #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ |
49ff4bb4 | 106 | #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained DSACK info */ |
1da177e4 LT |
107 | |
108 | #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) | |
109 | #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) | |
110 | #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) | |
111 | #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) | |
2e605294 | 112 | #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED) |
1da177e4 LT |
113 | |
114 | #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0) | |
115 | #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2) | |
116 | #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4) | |
117 | ||
4dc2665e IJ |
118 | #define IsSackFrto() (sysctl_tcp_frto == 0x2) |
119 | ||
1da177e4 LT |
120 | #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) |
121 | ||
e905a9ed | 122 | /* Adapt the MSS value used to make delayed ack decision to the |
1da177e4 | 123 | * real world. |
e905a9ed | 124 | */ |
40efc6fa SH |
125 | static void tcp_measure_rcv_mss(struct sock *sk, |
126 | const struct sk_buff *skb) | |
1da177e4 | 127 | { |
463c84b9 | 128 | struct inet_connection_sock *icsk = inet_csk(sk); |
e905a9ed | 129 | const unsigned int lss = icsk->icsk_ack.last_seg_size; |
463c84b9 | 130 | unsigned int len; |
1da177e4 | 131 | |
e905a9ed | 132 | icsk->icsk_ack.last_seg_size = 0; |
1da177e4 LT |
133 | |
134 | /* skb->len may jitter because of SACKs, even if peer | |
135 | * sends good full-sized frames. | |
136 | */ | |
ff9b5e0f | 137 | len = skb_shinfo(skb)->gso_size ?: skb->len; |
463c84b9 ACM |
138 | if (len >= icsk->icsk_ack.rcv_mss) { |
139 | icsk->icsk_ack.rcv_mss = len; | |
1da177e4 LT |
140 | } else { |
141 | /* Otherwise, we make more careful check taking into account, | |
142 | * that SACKs block is variable. | |
143 | * | |
144 | * "len" is invariant segment length, including TCP header. | |
145 | */ | |
9c70220b | 146 | len += skb->data - skb_transport_header(skb); |
1da177e4 LT |
147 | if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) || |
148 | /* If PSH is not set, packet should be | |
149 | * full sized, provided peer TCP is not badly broken. | |
150 | * This observation (if it is correct 8)) allows | |
151 | * to handle super-low mtu links fairly. | |
152 | */ | |
153 | (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && | |
aa8223c7 | 154 | !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { |
1da177e4 LT |
155 | /* Subtract also invariant (if peer is RFC compliant), |
156 | * tcp header plus fixed timestamp option length. | |
157 | * Resulting "len" is MSS free of SACK jitter. | |
158 | */ | |
463c84b9 ACM |
159 | len -= tcp_sk(sk)->tcp_header_len; |
160 | icsk->icsk_ack.last_seg_size = len; | |
1da177e4 | 161 | if (len == lss) { |
463c84b9 | 162 | icsk->icsk_ack.rcv_mss = len; |
1da177e4 LT |
163 | return; |
164 | } | |
165 | } | |
1ef9696c AK |
166 | if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) |
167 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; | |
463c84b9 | 168 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; |
1da177e4 LT |
169 | } |
170 | } | |
171 | ||
463c84b9 | 172 | static void tcp_incr_quickack(struct sock *sk) |
1da177e4 | 173 | { |
463c84b9 ACM |
174 | struct inet_connection_sock *icsk = inet_csk(sk); |
175 | unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); | |
1da177e4 LT |
176 | |
177 | if (quickacks==0) | |
178 | quickacks=2; | |
463c84b9 ACM |
179 | if (quickacks > icsk->icsk_ack.quick) |
180 | icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); | |
1da177e4 LT |
181 | } |
182 | ||
463c84b9 | 183 | void tcp_enter_quickack_mode(struct sock *sk) |
1da177e4 | 184 | { |
463c84b9 ACM |
185 | struct inet_connection_sock *icsk = inet_csk(sk); |
186 | tcp_incr_quickack(sk); | |
187 | icsk->icsk_ack.pingpong = 0; | |
188 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 LT |
189 | } |
190 | ||
191 | /* Send ACKs quickly, if "quick" count is not exhausted | |
192 | * and the session is not interactive. | |
193 | */ | |
194 | ||
463c84b9 | 195 | static inline int tcp_in_quickack_mode(const struct sock *sk) |
1da177e4 | 196 | { |
463c84b9 ACM |
197 | const struct inet_connection_sock *icsk = inet_csk(sk); |
198 | return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; | |
1da177e4 LT |
199 | } |
200 | ||
201 | /* Buffer size and advertised window tuning. | |
202 | * | |
203 | * 1. Tuning sk->sk_sndbuf, when connection enters established state. | |
204 | */ | |
205 | ||
206 | static void tcp_fixup_sndbuf(struct sock *sk) | |
207 | { | |
208 | int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 + | |
209 | sizeof(struct sk_buff); | |
210 | ||
211 | if (sk->sk_sndbuf < 3 * sndmem) | |
212 | sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]); | |
213 | } | |
214 | ||
215 | /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) | |
216 | * | |
217 | * All tcp_full_space() is split to two parts: "network" buffer, allocated | |
218 | * forward and advertised in receiver window (tp->rcv_wnd) and | |
219 | * "application buffer", required to isolate scheduling/application | |
220 | * latencies from network. | |
221 | * window_clamp is maximal advertised window. It can be less than | |
222 | * tcp_full_space(), in this case tcp_full_space() - window_clamp | |
223 | * is reserved for "application" buffer. The less window_clamp is | |
224 | * the smoother our behaviour from viewpoint of network, but the lower | |
225 | * throughput and the higher sensitivity of the connection to losses. 8) | |
226 | * | |
227 | * rcv_ssthresh is more strict window_clamp used at "slow start" | |
228 | * phase to predict further behaviour of this connection. | |
229 | * It is used for two goals: | |
230 | * - to enforce header prediction at sender, even when application | |
231 | * requires some significant "application buffer". It is check #1. | |
232 | * - to prevent pruning of receive queue because of misprediction | |
233 | * of receiver window. Check #2. | |
234 | * | |
235 | * The scheme does not work when sender sends good segments opening | |
caa20d9a | 236 | * window and then starts to feed us spaghetti. But it should work |
1da177e4 LT |
237 | * in common situations. Otherwise, we have to rely on queue collapsing. |
238 | */ | |
239 | ||
240 | /* Slow part of check#2. */ | |
9e412ba7 | 241 | static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 242 | { |
9e412ba7 | 243 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
244 | /* Optimize this! */ |
245 | int truesize = tcp_win_from_space(skb->truesize)/2; | |
326f36e9 | 246 | int window = tcp_win_from_space(sysctl_tcp_rmem[2])/2; |
1da177e4 LT |
247 | |
248 | while (tp->rcv_ssthresh <= window) { | |
249 | if (truesize <= skb->len) | |
463c84b9 | 250 | return 2 * inet_csk(sk)->icsk_ack.rcv_mss; |
1da177e4 LT |
251 | |
252 | truesize >>= 1; | |
253 | window >>= 1; | |
254 | } | |
255 | return 0; | |
256 | } | |
257 | ||
9e412ba7 | 258 | static void tcp_grow_window(struct sock *sk, |
40efc6fa | 259 | struct sk_buff *skb) |
1da177e4 | 260 | { |
9e412ba7 IJ |
261 | struct tcp_sock *tp = tcp_sk(sk); |
262 | ||
1da177e4 LT |
263 | /* Check #1 */ |
264 | if (tp->rcv_ssthresh < tp->window_clamp && | |
265 | (int)tp->rcv_ssthresh < tcp_space(sk) && | |
266 | !tcp_memory_pressure) { | |
267 | int incr; | |
268 | ||
269 | /* Check #2. Increase window, if skb with such overhead | |
270 | * will fit to rcvbuf in future. | |
271 | */ | |
272 | if (tcp_win_from_space(skb->truesize) <= skb->len) | |
273 | incr = 2*tp->advmss; | |
274 | else | |
9e412ba7 | 275 | incr = __tcp_grow_window(sk, skb); |
1da177e4 LT |
276 | |
277 | if (incr) { | |
278 | tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp); | |
463c84b9 | 279 | inet_csk(sk)->icsk_ack.quick |= 1; |
1da177e4 LT |
280 | } |
281 | } | |
282 | } | |
283 | ||
284 | /* 3. Tuning rcvbuf, when connection enters established state. */ | |
285 | ||
286 | static void tcp_fixup_rcvbuf(struct sock *sk) | |
287 | { | |
288 | struct tcp_sock *tp = tcp_sk(sk); | |
289 | int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); | |
290 | ||
291 | /* Try to select rcvbuf so that 4 mss-sized segments | |
caa20d9a | 292 | * will fit to window and corresponding skbs will fit to our rcvbuf. |
1da177e4 LT |
293 | * (was 3; 4 is minimum to allow fast retransmit to work.) |
294 | */ | |
295 | while (tcp_win_from_space(rcvmem) < tp->advmss) | |
296 | rcvmem += 128; | |
297 | if (sk->sk_rcvbuf < 4 * rcvmem) | |
298 | sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]); | |
299 | } | |
300 | ||
caa20d9a | 301 | /* 4. Try to fixup all. It is made immediately after connection enters |
1da177e4 LT |
302 | * established state. |
303 | */ | |
304 | static void tcp_init_buffer_space(struct sock *sk) | |
305 | { | |
306 | struct tcp_sock *tp = tcp_sk(sk); | |
307 | int maxwin; | |
308 | ||
309 | if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) | |
310 | tcp_fixup_rcvbuf(sk); | |
311 | if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) | |
312 | tcp_fixup_sndbuf(sk); | |
313 | ||
314 | tp->rcvq_space.space = tp->rcv_wnd; | |
315 | ||
316 | maxwin = tcp_full_space(sk); | |
317 | ||
318 | if (tp->window_clamp >= maxwin) { | |
319 | tp->window_clamp = maxwin; | |
320 | ||
321 | if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) | |
322 | tp->window_clamp = max(maxwin - | |
323 | (maxwin >> sysctl_tcp_app_win), | |
324 | 4 * tp->advmss); | |
325 | } | |
326 | ||
327 | /* Force reservation of one segment. */ | |
328 | if (sysctl_tcp_app_win && | |
329 | tp->window_clamp > 2 * tp->advmss && | |
330 | tp->window_clamp + tp->advmss > maxwin) | |
331 | tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); | |
332 | ||
333 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); | |
334 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
335 | } | |
336 | ||
1da177e4 | 337 | /* 5. Recalculate window clamp after socket hit its memory bounds. */ |
9e412ba7 | 338 | static void tcp_clamp_window(struct sock *sk) |
1da177e4 | 339 | { |
9e412ba7 | 340 | struct tcp_sock *tp = tcp_sk(sk); |
6687e988 | 341 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 342 | |
6687e988 | 343 | icsk->icsk_ack.quick = 0; |
1da177e4 | 344 | |
326f36e9 JH |
345 | if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && |
346 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && | |
347 | !tcp_memory_pressure && | |
348 | atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) { | |
349 | sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), | |
350 | sysctl_tcp_rmem[2]); | |
1da177e4 | 351 | } |
326f36e9 | 352 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) |
1da177e4 | 353 | tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss); |
1da177e4 LT |
354 | } |
355 | ||
40efc6fa SH |
356 | |
357 | /* Initialize RCV_MSS value. | |
358 | * RCV_MSS is an our guess about MSS used by the peer. | |
359 | * We haven't any direct information about the MSS. | |
360 | * It's better to underestimate the RCV_MSS rather than overestimate. | |
361 | * Overestimations make us ACKing less frequently than needed. | |
362 | * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). | |
363 | */ | |
364 | void tcp_initialize_rcv_mss(struct sock *sk) | |
365 | { | |
366 | struct tcp_sock *tp = tcp_sk(sk); | |
367 | unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); | |
368 | ||
369 | hint = min(hint, tp->rcv_wnd/2); | |
370 | hint = min(hint, TCP_MIN_RCVMSS); | |
371 | hint = max(hint, TCP_MIN_MSS); | |
372 | ||
373 | inet_csk(sk)->icsk_ack.rcv_mss = hint; | |
374 | } | |
375 | ||
1da177e4 LT |
376 | /* Receiver "autotuning" code. |
377 | * | |
378 | * The algorithm for RTT estimation w/o timestamps is based on | |
379 | * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. | |
380 | * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps> | |
381 | * | |
382 | * More detail on this code can be found at | |
383 | * <http://www.psc.edu/~jheffner/senior_thesis.ps>, | |
384 | * though this reference is out of date. A new paper | |
385 | * is pending. | |
386 | */ | |
387 | static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) | |
388 | { | |
389 | u32 new_sample = tp->rcv_rtt_est.rtt; | |
390 | long m = sample; | |
391 | ||
392 | if (m == 0) | |
393 | m = 1; | |
394 | ||
395 | if (new_sample != 0) { | |
396 | /* If we sample in larger samples in the non-timestamp | |
397 | * case, we could grossly overestimate the RTT especially | |
398 | * with chatty applications or bulk transfer apps which | |
399 | * are stalled on filesystem I/O. | |
400 | * | |
401 | * Also, since we are only going for a minimum in the | |
31f34269 | 402 | * non-timestamp case, we do not smooth things out |
caa20d9a | 403 | * else with timestamps disabled convergence takes too |
1da177e4 LT |
404 | * long. |
405 | */ | |
406 | if (!win_dep) { | |
407 | m -= (new_sample >> 3); | |
408 | new_sample += m; | |
409 | } else if (m < new_sample) | |
410 | new_sample = m << 3; | |
411 | } else { | |
caa20d9a | 412 | /* No previous measure. */ |
1da177e4 LT |
413 | new_sample = m << 3; |
414 | } | |
415 | ||
416 | if (tp->rcv_rtt_est.rtt != new_sample) | |
417 | tp->rcv_rtt_est.rtt = new_sample; | |
418 | } | |
419 | ||
420 | static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) | |
421 | { | |
422 | if (tp->rcv_rtt_est.time == 0) | |
423 | goto new_measure; | |
424 | if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) | |
425 | return; | |
426 | tcp_rcv_rtt_update(tp, | |
427 | jiffies - tp->rcv_rtt_est.time, | |
428 | 1); | |
429 | ||
430 | new_measure: | |
431 | tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; | |
432 | tp->rcv_rtt_est.time = tcp_time_stamp; | |
433 | } | |
434 | ||
463c84b9 | 435 | static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 436 | { |
463c84b9 | 437 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
438 | if (tp->rx_opt.rcv_tsecr && |
439 | (TCP_SKB_CB(skb)->end_seq - | |
463c84b9 | 440 | TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) |
1da177e4 LT |
441 | tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); |
442 | } | |
443 | ||
444 | /* | |
445 | * This function should be called every time data is copied to user space. | |
446 | * It calculates the appropriate TCP receive buffer space. | |
447 | */ | |
448 | void tcp_rcv_space_adjust(struct sock *sk) | |
449 | { | |
450 | struct tcp_sock *tp = tcp_sk(sk); | |
451 | int time; | |
452 | int space; | |
e905a9ed | 453 | |
1da177e4 LT |
454 | if (tp->rcvq_space.time == 0) |
455 | goto new_measure; | |
e905a9ed | 456 | |
1da177e4 LT |
457 | time = tcp_time_stamp - tp->rcvq_space.time; |
458 | if (time < (tp->rcv_rtt_est.rtt >> 3) || | |
459 | tp->rcv_rtt_est.rtt == 0) | |
460 | return; | |
e905a9ed | 461 | |
1da177e4 LT |
462 | space = 2 * (tp->copied_seq - tp->rcvq_space.seq); |
463 | ||
464 | space = max(tp->rcvq_space.space, space); | |
465 | ||
466 | if (tp->rcvq_space.space != space) { | |
467 | int rcvmem; | |
468 | ||
469 | tp->rcvq_space.space = space; | |
470 | ||
6fcf9412 JH |
471 | if (sysctl_tcp_moderate_rcvbuf && |
472 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { | |
1da177e4 LT |
473 | int new_clamp = space; |
474 | ||
475 | /* Receive space grows, normalize in order to | |
476 | * take into account packet headers and sk_buff | |
477 | * structure overhead. | |
478 | */ | |
479 | space /= tp->advmss; | |
480 | if (!space) | |
481 | space = 1; | |
482 | rcvmem = (tp->advmss + MAX_TCP_HEADER + | |
483 | 16 + sizeof(struct sk_buff)); | |
484 | while (tcp_win_from_space(rcvmem) < tp->advmss) | |
485 | rcvmem += 128; | |
486 | space *= rcvmem; | |
487 | space = min(space, sysctl_tcp_rmem[2]); | |
488 | if (space > sk->sk_rcvbuf) { | |
489 | sk->sk_rcvbuf = space; | |
490 | ||
491 | /* Make the window clamp follow along. */ | |
492 | tp->window_clamp = new_clamp; | |
493 | } | |
494 | } | |
495 | } | |
e905a9ed | 496 | |
1da177e4 LT |
497 | new_measure: |
498 | tp->rcvq_space.seq = tp->copied_seq; | |
499 | tp->rcvq_space.time = tcp_time_stamp; | |
500 | } | |
501 | ||
502 | /* There is something which you must keep in mind when you analyze the | |
503 | * behavior of the tp->ato delayed ack timeout interval. When a | |
504 | * connection starts up, we want to ack as quickly as possible. The | |
505 | * problem is that "good" TCP's do slow start at the beginning of data | |
506 | * transmission. The means that until we send the first few ACK's the | |
507 | * sender will sit on his end and only queue most of his data, because | |
508 | * he can only send snd_cwnd unacked packets at any given time. For | |
509 | * each ACK we send, he increments snd_cwnd and transmits more of his | |
510 | * queue. -DaveM | |
511 | */ | |
9e412ba7 | 512 | static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 513 | { |
9e412ba7 | 514 | struct tcp_sock *tp = tcp_sk(sk); |
463c84b9 | 515 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
516 | u32 now; |
517 | ||
463c84b9 | 518 | inet_csk_schedule_ack(sk); |
1da177e4 | 519 | |
463c84b9 | 520 | tcp_measure_rcv_mss(sk, skb); |
1da177e4 LT |
521 | |
522 | tcp_rcv_rtt_measure(tp); | |
e905a9ed | 523 | |
1da177e4 LT |
524 | now = tcp_time_stamp; |
525 | ||
463c84b9 | 526 | if (!icsk->icsk_ack.ato) { |
1da177e4 LT |
527 | /* The _first_ data packet received, initialize |
528 | * delayed ACK engine. | |
529 | */ | |
463c84b9 ACM |
530 | tcp_incr_quickack(sk); |
531 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 | 532 | } else { |
463c84b9 | 533 | int m = now - icsk->icsk_ack.lrcvtime; |
1da177e4 LT |
534 | |
535 | if (m <= TCP_ATO_MIN/2) { | |
536 | /* The fastest case is the first. */ | |
463c84b9 ACM |
537 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; |
538 | } else if (m < icsk->icsk_ack.ato) { | |
539 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; | |
540 | if (icsk->icsk_ack.ato > icsk->icsk_rto) | |
541 | icsk->icsk_ack.ato = icsk->icsk_rto; | |
542 | } else if (m > icsk->icsk_rto) { | |
caa20d9a | 543 | /* Too long gap. Apparently sender failed to |
1da177e4 LT |
544 | * restart window, so that we send ACKs quickly. |
545 | */ | |
463c84b9 | 546 | tcp_incr_quickack(sk); |
1da177e4 LT |
547 | sk_stream_mem_reclaim(sk); |
548 | } | |
549 | } | |
463c84b9 | 550 | icsk->icsk_ack.lrcvtime = now; |
1da177e4 LT |
551 | |
552 | TCP_ECN_check_ce(tp, skb); | |
553 | ||
554 | if (skb->len >= 128) | |
9e412ba7 | 555 | tcp_grow_window(sk, skb); |
1da177e4 LT |
556 | } |
557 | ||
05bb1fad DM |
558 | static u32 tcp_rto_min(struct sock *sk) |
559 | { | |
560 | struct dst_entry *dst = __sk_dst_get(sk); | |
561 | u32 rto_min = TCP_RTO_MIN; | |
562 | ||
5c127c58 | 563 | if (dst && dst_metric_locked(dst, RTAX_RTO_MIN)) |
05bb1fad DM |
564 | rto_min = dst->metrics[RTAX_RTO_MIN-1]; |
565 | return rto_min; | |
566 | } | |
567 | ||
1da177e4 LT |
568 | /* Called to compute a smoothed rtt estimate. The data fed to this |
569 | * routine either comes from timestamps, or from segments that were | |
570 | * known _not_ to have been retransmitted [see Karn/Partridge | |
571 | * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 | |
572 | * piece by Van Jacobson. | |
573 | * NOTE: the next three routines used to be one big routine. | |
574 | * To save cycles in the RFC 1323 implementation it was better to break | |
575 | * it up into three procedures. -- erics | |
576 | */ | |
2d2abbab | 577 | static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) |
1da177e4 | 578 | { |
6687e988 | 579 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
580 | long m = mrtt; /* RTT */ |
581 | ||
1da177e4 LT |
582 | /* The following amusing code comes from Jacobson's |
583 | * article in SIGCOMM '88. Note that rtt and mdev | |
584 | * are scaled versions of rtt and mean deviation. | |
e905a9ed | 585 | * This is designed to be as fast as possible |
1da177e4 LT |
586 | * m stands for "measurement". |
587 | * | |
588 | * On a 1990 paper the rto value is changed to: | |
589 | * RTO = rtt + 4 * mdev | |
590 | * | |
591 | * Funny. This algorithm seems to be very broken. | |
592 | * These formulae increase RTO, when it should be decreased, increase | |
31f34269 | 593 | * too slowly, when it should be increased quickly, decrease too quickly |
1da177e4 LT |
594 | * etc. I guess in BSD RTO takes ONE value, so that it is absolutely |
595 | * does not matter how to _calculate_ it. Seems, it was trap | |
596 | * that VJ failed to avoid. 8) | |
597 | */ | |
2de979bd | 598 | if (m == 0) |
1da177e4 LT |
599 | m = 1; |
600 | if (tp->srtt != 0) { | |
601 | m -= (tp->srtt >> 3); /* m is now error in rtt est */ | |
602 | tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ | |
603 | if (m < 0) { | |
604 | m = -m; /* m is now abs(error) */ | |
605 | m -= (tp->mdev >> 2); /* similar update on mdev */ | |
606 | /* This is similar to one of Eifel findings. | |
607 | * Eifel blocks mdev updates when rtt decreases. | |
608 | * This solution is a bit different: we use finer gain | |
609 | * for mdev in this case (alpha*beta). | |
610 | * Like Eifel it also prevents growth of rto, | |
611 | * but also it limits too fast rto decreases, | |
612 | * happening in pure Eifel. | |
613 | */ | |
614 | if (m > 0) | |
615 | m >>= 3; | |
616 | } else { | |
617 | m -= (tp->mdev >> 2); /* similar update on mdev */ | |
618 | } | |
619 | tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ | |
620 | if (tp->mdev > tp->mdev_max) { | |
621 | tp->mdev_max = tp->mdev; | |
622 | if (tp->mdev_max > tp->rttvar) | |
623 | tp->rttvar = tp->mdev_max; | |
624 | } | |
625 | if (after(tp->snd_una, tp->rtt_seq)) { | |
626 | if (tp->mdev_max < tp->rttvar) | |
627 | tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2; | |
628 | tp->rtt_seq = tp->snd_nxt; | |
05bb1fad | 629 | tp->mdev_max = tcp_rto_min(sk); |
1da177e4 LT |
630 | } |
631 | } else { | |
632 | /* no previous measure. */ | |
633 | tp->srtt = m<<3; /* take the measured time to be rtt */ | |
634 | tp->mdev = m<<1; /* make sure rto = 3*rtt */ | |
05bb1fad | 635 | tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); |
1da177e4 LT |
636 | tp->rtt_seq = tp->snd_nxt; |
637 | } | |
1da177e4 LT |
638 | } |
639 | ||
640 | /* Calculate rto without backoff. This is the second half of Van Jacobson's | |
641 | * routine referred to above. | |
642 | */ | |
463c84b9 | 643 | static inline void tcp_set_rto(struct sock *sk) |
1da177e4 | 644 | { |
463c84b9 | 645 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
646 | /* Old crap is replaced with new one. 8) |
647 | * | |
648 | * More seriously: | |
649 | * 1. If rtt variance happened to be less 50msec, it is hallucination. | |
650 | * It cannot be less due to utterly erratic ACK generation made | |
651 | * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ | |
652 | * to do with delayed acks, because at cwnd>2 true delack timeout | |
653 | * is invisible. Actually, Linux-2.4 also generates erratic | |
caa20d9a | 654 | * ACKs in some circumstances. |
1da177e4 | 655 | */ |
463c84b9 | 656 | inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar; |
1da177e4 LT |
657 | |
658 | /* 2. Fixups made earlier cannot be right. | |
659 | * If we do not estimate RTO correctly without them, | |
660 | * all the algo is pure shit and should be replaced | |
caa20d9a | 661 | * with correct one. It is exactly, which we pretend to do. |
1da177e4 LT |
662 | */ |
663 | } | |
664 | ||
665 | /* NOTE: clamping at TCP_RTO_MIN is not required, current algo | |
666 | * guarantees that rto is higher. | |
667 | */ | |
463c84b9 | 668 | static inline void tcp_bound_rto(struct sock *sk) |
1da177e4 | 669 | { |
463c84b9 ACM |
670 | if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX) |
671 | inet_csk(sk)->icsk_rto = TCP_RTO_MAX; | |
1da177e4 LT |
672 | } |
673 | ||
674 | /* Save metrics learned by this TCP session. | |
675 | This function is called only, when TCP finishes successfully | |
676 | i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. | |
677 | */ | |
678 | void tcp_update_metrics(struct sock *sk) | |
679 | { | |
680 | struct tcp_sock *tp = tcp_sk(sk); | |
681 | struct dst_entry *dst = __sk_dst_get(sk); | |
682 | ||
683 | if (sysctl_tcp_nometrics_save) | |
684 | return; | |
685 | ||
686 | dst_confirm(dst); | |
687 | ||
688 | if (dst && (dst->flags&DST_HOST)) { | |
6687e988 | 689 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
690 | int m; |
691 | ||
6687e988 | 692 | if (icsk->icsk_backoff || !tp->srtt) { |
1da177e4 LT |
693 | /* This session failed to estimate rtt. Why? |
694 | * Probably, no packets returned in time. | |
695 | * Reset our results. | |
696 | */ | |
697 | if (!(dst_metric_locked(dst, RTAX_RTT))) | |
698 | dst->metrics[RTAX_RTT-1] = 0; | |
699 | return; | |
700 | } | |
701 | ||
702 | m = dst_metric(dst, RTAX_RTT) - tp->srtt; | |
703 | ||
704 | /* If newly calculated rtt larger than stored one, | |
705 | * store new one. Otherwise, use EWMA. Remember, | |
706 | * rtt overestimation is always better than underestimation. | |
707 | */ | |
708 | if (!(dst_metric_locked(dst, RTAX_RTT))) { | |
709 | if (m <= 0) | |
710 | dst->metrics[RTAX_RTT-1] = tp->srtt; | |
711 | else | |
712 | dst->metrics[RTAX_RTT-1] -= (m>>3); | |
713 | } | |
714 | ||
715 | if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { | |
716 | if (m < 0) | |
717 | m = -m; | |
718 | ||
719 | /* Scale deviation to rttvar fixed point */ | |
720 | m >>= 1; | |
721 | if (m < tp->mdev) | |
722 | m = tp->mdev; | |
723 | ||
724 | if (m >= dst_metric(dst, RTAX_RTTVAR)) | |
725 | dst->metrics[RTAX_RTTVAR-1] = m; | |
726 | else | |
727 | dst->metrics[RTAX_RTTVAR-1] -= | |
728 | (dst->metrics[RTAX_RTTVAR-1] - m)>>2; | |
729 | } | |
730 | ||
731 | if (tp->snd_ssthresh >= 0xFFFF) { | |
732 | /* Slow start still did not finish. */ | |
733 | if (dst_metric(dst, RTAX_SSTHRESH) && | |
734 | !dst_metric_locked(dst, RTAX_SSTHRESH) && | |
735 | (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) | |
736 | dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1; | |
737 | if (!dst_metric_locked(dst, RTAX_CWND) && | |
738 | tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) | |
739 | dst->metrics[RTAX_CWND-1] = tp->snd_cwnd; | |
740 | } else if (tp->snd_cwnd > tp->snd_ssthresh && | |
6687e988 | 741 | icsk->icsk_ca_state == TCP_CA_Open) { |
1da177e4 LT |
742 | /* Cong. avoidance phase, cwnd is reliable. */ |
743 | if (!dst_metric_locked(dst, RTAX_SSTHRESH)) | |
744 | dst->metrics[RTAX_SSTHRESH-1] = | |
745 | max(tp->snd_cwnd >> 1, tp->snd_ssthresh); | |
746 | if (!dst_metric_locked(dst, RTAX_CWND)) | |
747 | dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1; | |
748 | } else { | |
749 | /* Else slow start did not finish, cwnd is non-sense, | |
750 | ssthresh may be also invalid. | |
751 | */ | |
752 | if (!dst_metric_locked(dst, RTAX_CWND)) | |
753 | dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1; | |
754 | if (dst->metrics[RTAX_SSTHRESH-1] && | |
755 | !dst_metric_locked(dst, RTAX_SSTHRESH) && | |
756 | tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1]) | |
757 | dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh; | |
758 | } | |
759 | ||
760 | if (!dst_metric_locked(dst, RTAX_REORDERING)) { | |
761 | if (dst->metrics[RTAX_REORDERING-1] < tp->reordering && | |
762 | tp->reordering != sysctl_tcp_reordering) | |
763 | dst->metrics[RTAX_REORDERING-1] = tp->reordering; | |
764 | } | |
765 | } | |
766 | } | |
767 | ||
26722873 DM |
768 | /* Numbers are taken from RFC3390. |
769 | * | |
770 | * John Heffner states: | |
771 | * | |
772 | * The RFC specifies a window of no more than 4380 bytes | |
773 | * unless 2*MSS > 4380. Reading the pseudocode in the RFC | |
774 | * is a bit misleading because they use a clamp at 4380 bytes | |
775 | * rather than use a multiplier in the relevant range. | |
776 | */ | |
1da177e4 LT |
777 | __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst) |
778 | { | |
779 | __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); | |
780 | ||
781 | if (!cwnd) { | |
c1b4a7e6 | 782 | if (tp->mss_cache > 1460) |
1da177e4 LT |
783 | cwnd = 2; |
784 | else | |
c1b4a7e6 | 785 | cwnd = (tp->mss_cache > 1095) ? 3 : 4; |
1da177e4 LT |
786 | } |
787 | return min_t(__u32, cwnd, tp->snd_cwnd_clamp); | |
788 | } | |
789 | ||
40efc6fa | 790 | /* Set slow start threshold and cwnd not falling to slow start */ |
3cfe3baa | 791 | void tcp_enter_cwr(struct sock *sk, const int set_ssthresh) |
40efc6fa SH |
792 | { |
793 | struct tcp_sock *tp = tcp_sk(sk); | |
3cfe3baa | 794 | const struct inet_connection_sock *icsk = inet_csk(sk); |
40efc6fa SH |
795 | |
796 | tp->prior_ssthresh = 0; | |
797 | tp->bytes_acked = 0; | |
e01f9d77 | 798 | if (icsk->icsk_ca_state < TCP_CA_CWR) { |
40efc6fa | 799 | tp->undo_marker = 0; |
3cfe3baa IJ |
800 | if (set_ssthresh) |
801 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
40efc6fa SH |
802 | tp->snd_cwnd = min(tp->snd_cwnd, |
803 | tcp_packets_in_flight(tp) + 1U); | |
804 | tp->snd_cwnd_cnt = 0; | |
805 | tp->high_seq = tp->snd_nxt; | |
806 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
807 | TCP_ECN_queue_cwr(tp); | |
808 | ||
809 | tcp_set_ca_state(sk, TCP_CA_CWR); | |
810 | } | |
811 | } | |
812 | ||
1da177e4 LT |
813 | /* Initialize metrics on socket. */ |
814 | ||
815 | static void tcp_init_metrics(struct sock *sk) | |
816 | { | |
817 | struct tcp_sock *tp = tcp_sk(sk); | |
818 | struct dst_entry *dst = __sk_dst_get(sk); | |
819 | ||
820 | if (dst == NULL) | |
821 | goto reset; | |
822 | ||
823 | dst_confirm(dst); | |
824 | ||
825 | if (dst_metric_locked(dst, RTAX_CWND)) | |
826 | tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); | |
827 | if (dst_metric(dst, RTAX_SSTHRESH)) { | |
828 | tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); | |
829 | if (tp->snd_ssthresh > tp->snd_cwnd_clamp) | |
830 | tp->snd_ssthresh = tp->snd_cwnd_clamp; | |
831 | } | |
832 | if (dst_metric(dst, RTAX_REORDERING) && | |
833 | tp->reordering != dst_metric(dst, RTAX_REORDERING)) { | |
834 | tp->rx_opt.sack_ok &= ~2; | |
835 | tp->reordering = dst_metric(dst, RTAX_REORDERING); | |
836 | } | |
837 | ||
838 | if (dst_metric(dst, RTAX_RTT) == 0) | |
839 | goto reset; | |
840 | ||
841 | if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3)) | |
842 | goto reset; | |
843 | ||
844 | /* Initial rtt is determined from SYN,SYN-ACK. | |
845 | * The segment is small and rtt may appear much | |
846 | * less than real one. Use per-dst memory | |
847 | * to make it more realistic. | |
848 | * | |
849 | * A bit of theory. RTT is time passed after "normal" sized packet | |
caa20d9a | 850 | * is sent until it is ACKed. In normal circumstances sending small |
1da177e4 LT |
851 | * packets force peer to delay ACKs and calculation is correct too. |
852 | * The algorithm is adaptive and, provided we follow specs, it | |
853 | * NEVER underestimate RTT. BUT! If peer tries to make some clever | |
854 | * tricks sort of "quick acks" for time long enough to decrease RTT | |
855 | * to low value, and then abruptly stops to do it and starts to delay | |
856 | * ACKs, wait for troubles. | |
857 | */ | |
858 | if (dst_metric(dst, RTAX_RTT) > tp->srtt) { | |
859 | tp->srtt = dst_metric(dst, RTAX_RTT); | |
860 | tp->rtt_seq = tp->snd_nxt; | |
861 | } | |
862 | if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) { | |
863 | tp->mdev = dst_metric(dst, RTAX_RTTVAR); | |
864 | tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN); | |
865 | } | |
463c84b9 ACM |
866 | tcp_set_rto(sk); |
867 | tcp_bound_rto(sk); | |
868 | if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp) | |
1da177e4 LT |
869 | goto reset; |
870 | tp->snd_cwnd = tcp_init_cwnd(tp, dst); | |
871 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
872 | return; | |
873 | ||
874 | reset: | |
875 | /* Play conservative. If timestamps are not | |
876 | * supported, TCP will fail to recalculate correct | |
877 | * rtt, if initial rto is too small. FORGET ALL AND RESET! | |
878 | */ | |
879 | if (!tp->rx_opt.saw_tstamp && tp->srtt) { | |
880 | tp->srtt = 0; | |
881 | tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT; | |
463c84b9 | 882 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
1da177e4 LT |
883 | } |
884 | } | |
885 | ||
6687e988 ACM |
886 | static void tcp_update_reordering(struct sock *sk, const int metric, |
887 | const int ts) | |
1da177e4 | 888 | { |
6687e988 | 889 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
890 | if (metric > tp->reordering) { |
891 | tp->reordering = min(TCP_MAX_REORDERING, metric); | |
892 | ||
893 | /* This exciting event is worth to be remembered. 8) */ | |
894 | if (ts) | |
895 | NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER); | |
896 | else if (IsReno(tp)) | |
897 | NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER); | |
898 | else if (IsFack(tp)) | |
899 | NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER); | |
900 | else | |
901 | NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER); | |
902 | #if FASTRETRANS_DEBUG > 1 | |
903 | printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", | |
6687e988 | 904 | tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, |
1da177e4 LT |
905 | tp->reordering, |
906 | tp->fackets_out, | |
907 | tp->sacked_out, | |
908 | tp->undo_marker ? tp->undo_retrans : 0); | |
909 | #endif | |
910 | /* Disable FACK yet. */ | |
911 | tp->rx_opt.sack_ok &= ~2; | |
912 | } | |
913 | } | |
914 | ||
915 | /* This procedure tags the retransmission queue when SACKs arrive. | |
916 | * | |
917 | * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). | |
918 | * Packets in queue with these bits set are counted in variables | |
919 | * sacked_out, retrans_out and lost_out, correspondingly. | |
920 | * | |
921 | * Valid combinations are: | |
922 | * Tag InFlight Description | |
923 | * 0 1 - orig segment is in flight. | |
924 | * S 0 - nothing flies, orig reached receiver. | |
925 | * L 0 - nothing flies, orig lost by net. | |
926 | * R 2 - both orig and retransmit are in flight. | |
927 | * L|R 1 - orig is lost, retransmit is in flight. | |
928 | * S|R 1 - orig reached receiver, retrans is still in flight. | |
929 | * (L|S|R is logically valid, it could occur when L|R is sacked, | |
930 | * but it is equivalent to plain S and code short-curcuits it to S. | |
931 | * L|S is logically invalid, it would mean -1 packet in flight 8)) | |
932 | * | |
933 | * These 6 states form finite state machine, controlled by the following events: | |
934 | * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) | |
935 | * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) | |
936 | * 3. Loss detection event of one of three flavors: | |
937 | * A. Scoreboard estimator decided the packet is lost. | |
938 | * A'. Reno "three dupacks" marks head of queue lost. | |
939 | * A''. Its FACK modfication, head until snd.fack is lost. | |
940 | * B. SACK arrives sacking data transmitted after never retransmitted | |
941 | * hole was sent out. | |
942 | * C. SACK arrives sacking SND.NXT at the moment, when the | |
943 | * segment was retransmitted. | |
944 | * 4. D-SACK added new rule: D-SACK changes any tag to S. | |
945 | * | |
946 | * It is pleasant to note, that state diagram turns out to be commutative, | |
947 | * so that we are allowed not to be bothered by order of our actions, | |
948 | * when multiple events arrive simultaneously. (see the function below). | |
949 | * | |
950 | * Reordering detection. | |
951 | * -------------------- | |
952 | * Reordering metric is maximal distance, which a packet can be displaced | |
953 | * in packet stream. With SACKs we can estimate it: | |
954 | * | |
955 | * 1. SACK fills old hole and the corresponding segment was not | |
956 | * ever retransmitted -> reordering. Alas, we cannot use it | |
957 | * when segment was retransmitted. | |
958 | * 2. The last flaw is solved with D-SACK. D-SACK arrives | |
959 | * for retransmitted and already SACKed segment -> reordering.. | |
960 | * Both of these heuristics are not used in Loss state, when we cannot | |
961 | * account for retransmits accurately. | |
962 | */ | |
963 | static int | |
964 | tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una) | |
965 | { | |
6687e988 | 966 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 967 | struct tcp_sock *tp = tcp_sk(sk); |
9c70220b ACM |
968 | unsigned char *ptr = (skb_transport_header(ack_skb) + |
969 | TCP_SKB_CB(ack_skb)->sacked); | |
269bd27e | 970 | struct tcp_sack_block_wire *sp = (struct tcp_sack_block_wire *)(ptr+2); |
fda03fbb | 971 | struct sk_buff *cached_skb; |
1da177e4 LT |
972 | int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3; |
973 | int reord = tp->packets_out; | |
974 | int prior_fackets; | |
975 | u32 lost_retrans = 0; | |
976 | int flag = 0; | |
7769f406 | 977 | int found_dup_sack = 0; |
fda03fbb | 978 | int cached_fack_count; |
1da177e4 | 979 | int i; |
fda03fbb | 980 | int first_sack_index; |
1da177e4 | 981 | |
d738cd8f | 982 | if (!tp->sacked_out) { |
1da177e4 | 983 | tp->fackets_out = 0; |
d738cd8f IJ |
984 | tp->highest_sack = tp->snd_una; |
985 | } | |
1da177e4 LT |
986 | prior_fackets = tp->fackets_out; |
987 | ||
6f74651a BE |
988 | /* Check for D-SACK. */ |
989 | if (before(ntohl(sp[0].start_seq), TCP_SKB_CB(ack_skb)->ack_seq)) { | |
49ff4bb4 | 990 | flag |= FLAG_DSACKING_ACK; |
7769f406 | 991 | found_dup_sack = 1; |
6f74651a BE |
992 | tp->rx_opt.sack_ok |= 4; |
993 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV); | |
994 | } else if (num_sacks > 1 && | |
995 | !after(ntohl(sp[0].end_seq), ntohl(sp[1].end_seq)) && | |
996 | !before(ntohl(sp[0].start_seq), ntohl(sp[1].start_seq))) { | |
49ff4bb4 | 997 | flag |= FLAG_DSACKING_ACK; |
7769f406 | 998 | found_dup_sack = 1; |
6f74651a BE |
999 | tp->rx_opt.sack_ok |= 4; |
1000 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV); | |
1001 | } | |
1002 | ||
1003 | /* D-SACK for already forgotten data... | |
1004 | * Do dumb counting. */ | |
7769f406 | 1005 | if (found_dup_sack && |
6f74651a BE |
1006 | !after(ntohl(sp[0].end_seq), prior_snd_una) && |
1007 | after(ntohl(sp[0].end_seq), tp->undo_marker)) | |
1008 | tp->undo_retrans--; | |
1009 | ||
1010 | /* Eliminate too old ACKs, but take into | |
1011 | * account more or less fresh ones, they can | |
1012 | * contain valid SACK info. | |
1013 | */ | |
1014 | if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) | |
1015 | return 0; | |
1016 | ||
6a438bbe SH |
1017 | /* SACK fastpath: |
1018 | * if the only SACK change is the increase of the end_seq of | |
1019 | * the first block then only apply that SACK block | |
1020 | * and use retrans queue hinting otherwise slowpath */ | |
1021 | flag = 1; | |
6f74651a BE |
1022 | for (i = 0; i < num_sacks; i++) { |
1023 | __be32 start_seq = sp[i].start_seq; | |
1024 | __be32 end_seq = sp[i].end_seq; | |
6a438bbe | 1025 | |
6f74651a | 1026 | if (i == 0) { |
6a438bbe SH |
1027 | if (tp->recv_sack_cache[i].start_seq != start_seq) |
1028 | flag = 0; | |
1029 | } else { | |
1030 | if ((tp->recv_sack_cache[i].start_seq != start_seq) || | |
1031 | (tp->recv_sack_cache[i].end_seq != end_seq)) | |
1032 | flag = 0; | |
1033 | } | |
1034 | tp->recv_sack_cache[i].start_seq = start_seq; | |
1035 | tp->recv_sack_cache[i].end_seq = end_seq; | |
6a438bbe | 1036 | } |
8a3c3a97 BE |
1037 | /* Clear the rest of the cache sack blocks so they won't match mistakenly. */ |
1038 | for (; i < ARRAY_SIZE(tp->recv_sack_cache); i++) { | |
1039 | tp->recv_sack_cache[i].start_seq = 0; | |
1040 | tp->recv_sack_cache[i].end_seq = 0; | |
1041 | } | |
6a438bbe | 1042 | |
fda03fbb | 1043 | first_sack_index = 0; |
6a438bbe SH |
1044 | if (flag) |
1045 | num_sacks = 1; | |
1046 | else { | |
1047 | int j; | |
1048 | tp->fastpath_skb_hint = NULL; | |
1049 | ||
1050 | /* order SACK blocks to allow in order walk of the retrans queue */ | |
1051 | for (i = num_sacks-1; i > 0; i--) { | |
1052 | for (j = 0; j < i; j++){ | |
1053 | if (after(ntohl(sp[j].start_seq), | |
1054 | ntohl(sp[j+1].start_seq))){ | |
db3ccdac BE |
1055 | struct tcp_sack_block_wire tmp; |
1056 | ||
1057 | tmp = sp[j]; | |
1058 | sp[j] = sp[j+1]; | |
1059 | sp[j+1] = tmp; | |
fda03fbb BE |
1060 | |
1061 | /* Track where the first SACK block goes to */ | |
1062 | if (j == first_sack_index) | |
1063 | first_sack_index = j+1; | |
6a438bbe SH |
1064 | } |
1065 | ||
1066 | } | |
1067 | } | |
1068 | } | |
1069 | ||
1070 | /* clear flag as used for different purpose in following code */ | |
1071 | flag = 0; | |
1072 | ||
fda03fbb BE |
1073 | /* Use SACK fastpath hint if valid */ |
1074 | cached_skb = tp->fastpath_skb_hint; | |
1075 | cached_fack_count = tp->fastpath_cnt_hint; | |
1076 | if (!cached_skb) { | |
fe067e8a | 1077 | cached_skb = tcp_write_queue_head(sk); |
fda03fbb BE |
1078 | cached_fack_count = 0; |
1079 | } | |
1080 | ||
6a438bbe SH |
1081 | for (i=0; i<num_sacks; i++, sp++) { |
1082 | struct sk_buff *skb; | |
1083 | __u32 start_seq = ntohl(sp->start_seq); | |
1084 | __u32 end_seq = ntohl(sp->end_seq); | |
1085 | int fack_count; | |
7769f406 | 1086 | int dup_sack = (found_dup_sack && (i == first_sack_index)); |
6a438bbe | 1087 | |
fda03fbb BE |
1088 | skb = cached_skb; |
1089 | fack_count = cached_fack_count; | |
1da177e4 LT |
1090 | |
1091 | /* Event "B" in the comment above. */ | |
1092 | if (after(end_seq, tp->high_seq)) | |
1093 | flag |= FLAG_DATA_LOST; | |
1094 | ||
fe067e8a | 1095 | tcp_for_write_queue_from(skb, sk) { |
6475be16 DM |
1096 | int in_sack, pcount; |
1097 | u8 sacked; | |
1da177e4 | 1098 | |
fe067e8a DM |
1099 | if (skb == tcp_send_head(sk)) |
1100 | break; | |
1101 | ||
fda03fbb BE |
1102 | cached_skb = skb; |
1103 | cached_fack_count = fack_count; | |
1104 | if (i == first_sack_index) { | |
1105 | tp->fastpath_skb_hint = skb; | |
1106 | tp->fastpath_cnt_hint = fack_count; | |
1107 | } | |
6a438bbe | 1108 | |
1da177e4 LT |
1109 | /* The retransmission queue is always in order, so |
1110 | * we can short-circuit the walk early. | |
1111 | */ | |
6475be16 | 1112 | if (!before(TCP_SKB_CB(skb)->seq, end_seq)) |
1da177e4 LT |
1113 | break; |
1114 | ||
3c05d92e HX |
1115 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && |
1116 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1117 | ||
6475be16 DM |
1118 | pcount = tcp_skb_pcount(skb); |
1119 | ||
3c05d92e HX |
1120 | if (pcount > 1 && !in_sack && |
1121 | after(TCP_SKB_CB(skb)->end_seq, start_seq)) { | |
6475be16 DM |
1122 | unsigned int pkt_len; |
1123 | ||
3c05d92e HX |
1124 | in_sack = !after(start_seq, |
1125 | TCP_SKB_CB(skb)->seq); | |
1126 | ||
1127 | if (!in_sack) | |
6475be16 DM |
1128 | pkt_len = (start_seq - |
1129 | TCP_SKB_CB(skb)->seq); | |
1130 | else | |
1131 | pkt_len = (end_seq - | |
1132 | TCP_SKB_CB(skb)->seq); | |
7967168c | 1133 | if (tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size)) |
6475be16 DM |
1134 | break; |
1135 | pcount = tcp_skb_pcount(skb); | |
1136 | } | |
1137 | ||
1138 | fack_count += pcount; | |
1da177e4 | 1139 | |
6475be16 DM |
1140 | sacked = TCP_SKB_CB(skb)->sacked; |
1141 | ||
1da177e4 LT |
1142 | /* Account D-SACK for retransmitted packet. */ |
1143 | if ((dup_sack && in_sack) && | |
1144 | (sacked & TCPCB_RETRANS) && | |
1145 | after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker)) | |
1146 | tp->undo_retrans--; | |
1147 | ||
1148 | /* The frame is ACKed. */ | |
1149 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) { | |
1150 | if (sacked&TCPCB_RETRANS) { | |
1151 | if ((dup_sack && in_sack) && | |
1152 | (sacked&TCPCB_SACKED_ACKED)) | |
1153 | reord = min(fack_count, reord); | |
1154 | } else { | |
1155 | /* If it was in a hole, we detected reordering. */ | |
1156 | if (fack_count < prior_fackets && | |
1157 | !(sacked&TCPCB_SACKED_ACKED)) | |
1158 | reord = min(fack_count, reord); | |
1159 | } | |
1160 | ||
1161 | /* Nothing to do; acked frame is about to be dropped. */ | |
1162 | continue; | |
1163 | } | |
1164 | ||
1165 | if ((sacked&TCPCB_SACKED_RETRANS) && | |
1166 | after(end_seq, TCP_SKB_CB(skb)->ack_seq) && | |
1167 | (!lost_retrans || after(end_seq, lost_retrans))) | |
1168 | lost_retrans = end_seq; | |
1169 | ||
1170 | if (!in_sack) | |
1171 | continue; | |
1172 | ||
1173 | if (!(sacked&TCPCB_SACKED_ACKED)) { | |
1174 | if (sacked & TCPCB_SACKED_RETRANS) { | |
1175 | /* If the segment is not tagged as lost, | |
1176 | * we do not clear RETRANS, believing | |
1177 | * that retransmission is still in flight. | |
1178 | */ | |
1179 | if (sacked & TCPCB_LOST) { | |
1180 | TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); | |
1181 | tp->lost_out -= tcp_skb_pcount(skb); | |
1182 | tp->retrans_out -= tcp_skb_pcount(skb); | |
6a438bbe SH |
1183 | |
1184 | /* clear lost hint */ | |
1185 | tp->retransmit_skb_hint = NULL; | |
1da177e4 LT |
1186 | } |
1187 | } else { | |
1188 | /* New sack for not retransmitted frame, | |
1189 | * which was in hole. It is reordering. | |
1190 | */ | |
1191 | if (!(sacked & TCPCB_RETRANS) && | |
1192 | fack_count < prior_fackets) | |
1193 | reord = min(fack_count, reord); | |
1194 | ||
1195 | if (sacked & TCPCB_LOST) { | |
1196 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; | |
1197 | tp->lost_out -= tcp_skb_pcount(skb); | |
6a438bbe SH |
1198 | |
1199 | /* clear lost hint */ | |
1200 | tp->retransmit_skb_hint = NULL; | |
1da177e4 | 1201 | } |
4dc2665e IJ |
1202 | /* SACK enhanced F-RTO detection. |
1203 | * Set flag if and only if non-rexmitted | |
1204 | * segments below frto_highmark are | |
1205 | * SACKed (RFC4138; Appendix B). | |
1206 | * Clearing correct due to in-order walk | |
1207 | */ | |
1208 | if (after(end_seq, tp->frto_highmark)) { | |
1209 | flag &= ~FLAG_ONLY_ORIG_SACKED; | |
1210 | } else { | |
1211 | if (!(sacked & TCPCB_RETRANS)) | |
1212 | flag |= FLAG_ONLY_ORIG_SACKED; | |
1213 | } | |
1da177e4 LT |
1214 | } |
1215 | ||
1216 | TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED; | |
1217 | flag |= FLAG_DATA_SACKED; | |
1218 | tp->sacked_out += tcp_skb_pcount(skb); | |
1219 | ||
1220 | if (fack_count > tp->fackets_out) | |
1221 | tp->fackets_out = fack_count; | |
d738cd8f IJ |
1222 | |
1223 | if (after(TCP_SKB_CB(skb)->seq, | |
1224 | tp->highest_sack)) | |
1225 | tp->highest_sack = TCP_SKB_CB(skb)->seq; | |
1da177e4 LT |
1226 | } else { |
1227 | if (dup_sack && (sacked&TCPCB_RETRANS)) | |
1228 | reord = min(fack_count, reord); | |
1229 | } | |
1230 | ||
1231 | /* D-SACK. We can detect redundant retransmission | |
1232 | * in S|R and plain R frames and clear it. | |
1233 | * undo_retrans is decreased above, L|R frames | |
1234 | * are accounted above as well. | |
1235 | */ | |
1236 | if (dup_sack && | |
1237 | (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) { | |
1238 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; | |
1239 | tp->retrans_out -= tcp_skb_pcount(skb); | |
6a438bbe | 1240 | tp->retransmit_skb_hint = NULL; |
1da177e4 LT |
1241 | } |
1242 | } | |
1243 | } | |
1244 | ||
1245 | /* Check for lost retransmit. This superb idea is | |
1246 | * borrowed from "ratehalving". Event "C". | |
1247 | * Later note: FACK people cheated me again 8), | |
1248 | * we have to account for reordering! Ugly, | |
1249 | * but should help. | |
1250 | */ | |
6687e988 | 1251 | if (lost_retrans && icsk->icsk_ca_state == TCP_CA_Recovery) { |
1da177e4 LT |
1252 | struct sk_buff *skb; |
1253 | ||
fe067e8a DM |
1254 | tcp_for_write_queue(skb, sk) { |
1255 | if (skb == tcp_send_head(sk)) | |
1256 | break; | |
1da177e4 LT |
1257 | if (after(TCP_SKB_CB(skb)->seq, lost_retrans)) |
1258 | break; | |
1259 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1260 | continue; | |
1261 | if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) && | |
1262 | after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) && | |
1263 | (IsFack(tp) || | |
1264 | !before(lost_retrans, | |
1265 | TCP_SKB_CB(skb)->ack_seq + tp->reordering * | |
c1b4a7e6 | 1266 | tp->mss_cache))) { |
1da177e4 LT |
1267 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; |
1268 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1269 | ||
6a438bbe SH |
1270 | /* clear lost hint */ |
1271 | tp->retransmit_skb_hint = NULL; | |
1272 | ||
1da177e4 LT |
1273 | if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) { |
1274 | tp->lost_out += tcp_skb_pcount(skb); | |
1275 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1276 | flag |= FLAG_DATA_SACKED; | |
1277 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT); | |
1278 | } | |
1279 | } | |
1280 | } | |
1281 | } | |
1282 | ||
1283 | tp->left_out = tp->sacked_out + tp->lost_out; | |
1284 | ||
288035f9 | 1285 | if ((reord < tp->fackets_out) && icsk->icsk_ca_state != TCP_CA_Loss && |
c5e7af0d | 1286 | (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark))) |
6687e988 | 1287 | tcp_update_reordering(sk, ((tp->fackets_out + 1) - reord), 0); |
1da177e4 LT |
1288 | |
1289 | #if FASTRETRANS_DEBUG > 0 | |
1290 | BUG_TRAP((int)tp->sacked_out >= 0); | |
1291 | BUG_TRAP((int)tp->lost_out >= 0); | |
1292 | BUG_TRAP((int)tp->retrans_out >= 0); | |
1293 | BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0); | |
1294 | #endif | |
1295 | return flag; | |
1296 | } | |
1297 | ||
575ee714 IJ |
1298 | /* F-RTO can only be used if TCP has never retransmitted anything other than |
1299 | * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here) | |
30935cf4 | 1300 | */ |
46d0de4e | 1301 | int tcp_use_frto(struct sock *sk) |
bdaae17d IJ |
1302 | { |
1303 | const struct tcp_sock *tp = tcp_sk(sk); | |
46d0de4e IJ |
1304 | struct sk_buff *skb; |
1305 | ||
575ee714 | 1306 | if (!sysctl_tcp_frto) |
46d0de4e | 1307 | return 0; |
bdaae17d | 1308 | |
4dc2665e IJ |
1309 | if (IsSackFrto()) |
1310 | return 1; | |
1311 | ||
46d0de4e IJ |
1312 | /* Avoid expensive walking of rexmit queue if possible */ |
1313 | if (tp->retrans_out > 1) | |
1314 | return 0; | |
1315 | ||
fe067e8a DM |
1316 | skb = tcp_write_queue_head(sk); |
1317 | skb = tcp_write_queue_next(sk, skb); /* Skips head */ | |
1318 | tcp_for_write_queue_from(skb, sk) { | |
1319 | if (skb == tcp_send_head(sk)) | |
1320 | break; | |
46d0de4e IJ |
1321 | if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS) |
1322 | return 0; | |
1323 | /* Short-circuit when first non-SACKed skb has been checked */ | |
1324 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) | |
1325 | break; | |
1326 | } | |
1327 | return 1; | |
bdaae17d IJ |
1328 | } |
1329 | ||
30935cf4 IJ |
1330 | /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO |
1331 | * recovery a bit and use heuristics in tcp_process_frto() to detect if | |
d1a54c6a IJ |
1332 | * the RTO was spurious. Only clear SACKED_RETRANS of the head here to |
1333 | * keep retrans_out counting accurate (with SACK F-RTO, other than head | |
1334 | * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS | |
1335 | * bits are handled if the Loss state is really to be entered (in | |
1336 | * tcp_enter_frto_loss). | |
7487c48c IJ |
1337 | * |
1338 | * Do like tcp_enter_loss() would; when RTO expires the second time it | |
1339 | * does: | |
1340 | * "Reduce ssthresh if it has not yet been made inside this window." | |
1da177e4 LT |
1341 | */ |
1342 | void tcp_enter_frto(struct sock *sk) | |
1343 | { | |
6687e988 | 1344 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
1345 | struct tcp_sock *tp = tcp_sk(sk); |
1346 | struct sk_buff *skb; | |
1347 | ||
7487c48c | 1348 | if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) || |
e905a9ed | 1349 | tp->snd_una == tp->high_seq || |
7487c48c IJ |
1350 | ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) && |
1351 | !icsk->icsk_retransmits)) { | |
6687e988 | 1352 | tp->prior_ssthresh = tcp_current_ssthresh(sk); |
66e93e45 IJ |
1353 | /* Our state is too optimistic in ssthresh() call because cwnd |
1354 | * is not reduced until tcp_enter_frto_loss() when previous FRTO | |
1355 | * recovery has not yet completed. Pattern would be this: RTO, | |
1356 | * Cumulative ACK, RTO (2xRTO for the same segment does not end | |
1357 | * up here twice). | |
1358 | * RFC4138 should be more specific on what to do, even though | |
1359 | * RTO is quite unlikely to occur after the first Cumulative ACK | |
1360 | * due to back-off and complexity of triggering events ... | |
1361 | */ | |
1362 | if (tp->frto_counter) { | |
1363 | u32 stored_cwnd; | |
1364 | stored_cwnd = tp->snd_cwnd; | |
1365 | tp->snd_cwnd = 2; | |
1366 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
1367 | tp->snd_cwnd = stored_cwnd; | |
1368 | } else { | |
1369 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
1370 | } | |
1371 | /* ... in theory, cong.control module could do "any tricks" in | |
1372 | * ssthresh(), which means that ca_state, lost bits and lost_out | |
1373 | * counter would have to be faked before the call occurs. We | |
1374 | * consider that too expensive, unlikely and hacky, so modules | |
1375 | * using these in ssthresh() must deal these incompatibility | |
1376 | * issues if they receives CA_EVENT_FRTO and frto_counter != 0 | |
1377 | */ | |
6687e988 | 1378 | tcp_ca_event(sk, CA_EVENT_FRTO); |
1da177e4 LT |
1379 | } |
1380 | ||
1da177e4 LT |
1381 | tp->undo_marker = tp->snd_una; |
1382 | tp->undo_retrans = 0; | |
1383 | ||
fe067e8a | 1384 | skb = tcp_write_queue_head(sk); |
d1a54c6a | 1385 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
522e7548 | 1386 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; |
d1a54c6a | 1387 | tp->retrans_out -= tcp_skb_pcount(skb); |
1da177e4 LT |
1388 | } |
1389 | tcp_sync_left_out(tp); | |
1390 | ||
4dc2665e IJ |
1391 | /* Earlier loss recovery underway (see RFC4138; Appendix B). |
1392 | * The last condition is necessary at least in tp->frto_counter case. | |
1393 | */ | |
1394 | if (IsSackFrto() && (tp->frto_counter || | |
1395 | ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) && | |
1396 | after(tp->high_seq, tp->snd_una)) { | |
1397 | tp->frto_highmark = tp->high_seq; | |
1398 | } else { | |
1399 | tp->frto_highmark = tp->snd_nxt; | |
1400 | } | |
7b0eb22b IJ |
1401 | tcp_set_ca_state(sk, TCP_CA_Disorder); |
1402 | tp->high_seq = tp->snd_nxt; | |
7487c48c | 1403 | tp->frto_counter = 1; |
1da177e4 LT |
1404 | } |
1405 | ||
1406 | /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, | |
1407 | * which indicates that we should follow the traditional RTO recovery, | |
1408 | * i.e. mark everything lost and do go-back-N retransmission. | |
1409 | */ | |
d1a54c6a | 1410 | static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag) |
1da177e4 LT |
1411 | { |
1412 | struct tcp_sock *tp = tcp_sk(sk); | |
1413 | struct sk_buff *skb; | |
1414 | int cnt = 0; | |
1415 | ||
1416 | tp->sacked_out = 0; | |
1417 | tp->lost_out = 0; | |
1418 | tp->fackets_out = 0; | |
d1a54c6a | 1419 | tp->retrans_out = 0; |
1da177e4 | 1420 | |
fe067e8a DM |
1421 | tcp_for_write_queue(skb, sk) { |
1422 | if (skb == tcp_send_head(sk)) | |
1423 | break; | |
1da177e4 | 1424 | cnt += tcp_skb_pcount(skb); |
d1a54c6a IJ |
1425 | /* |
1426 | * Count the retransmission made on RTO correctly (only when | |
1427 | * waiting for the first ACK and did not get it)... | |
1428 | */ | |
1429 | if ((tp->frto_counter == 1) && !(flag&FLAG_DATA_ACKED)) { | |
0a9f2a46 IJ |
1430 | /* For some reason this R-bit might get cleared? */ |
1431 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
1432 | tp->retrans_out += tcp_skb_pcount(skb); | |
d1a54c6a IJ |
1433 | /* ...enter this if branch just for the first segment */ |
1434 | flag |= FLAG_DATA_ACKED; | |
1435 | } else { | |
1436 | TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); | |
1437 | } | |
1da177e4 LT |
1438 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) { |
1439 | ||
1440 | /* Do not mark those segments lost that were | |
1441 | * forward transmitted after RTO | |
1442 | */ | |
1443 | if (!after(TCP_SKB_CB(skb)->end_seq, | |
1444 | tp->frto_highmark)) { | |
1445 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1446 | tp->lost_out += tcp_skb_pcount(skb); | |
1447 | } | |
1448 | } else { | |
1449 | tp->sacked_out += tcp_skb_pcount(skb); | |
1450 | tp->fackets_out = cnt; | |
1451 | } | |
1452 | } | |
1453 | tcp_sync_left_out(tp); | |
1454 | ||
95c4922b | 1455 | tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments; |
1da177e4 LT |
1456 | tp->snd_cwnd_cnt = 0; |
1457 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1458 | tp->undo_marker = 0; | |
1459 | tp->frto_counter = 0; | |
1460 | ||
1461 | tp->reordering = min_t(unsigned int, tp->reordering, | |
1462 | sysctl_tcp_reordering); | |
6687e988 | 1463 | tcp_set_ca_state(sk, TCP_CA_Loss); |
1da177e4 LT |
1464 | tp->high_seq = tp->frto_highmark; |
1465 | TCP_ECN_queue_cwr(tp); | |
6a438bbe SH |
1466 | |
1467 | clear_all_retrans_hints(tp); | |
1da177e4 LT |
1468 | } |
1469 | ||
1470 | void tcp_clear_retrans(struct tcp_sock *tp) | |
1471 | { | |
1472 | tp->left_out = 0; | |
1473 | tp->retrans_out = 0; | |
1474 | ||
1475 | tp->fackets_out = 0; | |
1476 | tp->sacked_out = 0; | |
1477 | tp->lost_out = 0; | |
1478 | ||
1479 | tp->undo_marker = 0; | |
1480 | tp->undo_retrans = 0; | |
1481 | } | |
1482 | ||
1483 | /* Enter Loss state. If "how" is not zero, forget all SACK information | |
1484 | * and reset tags completely, otherwise preserve SACKs. If receiver | |
1485 | * dropped its ofo queue, we will know this due to reneging detection. | |
1486 | */ | |
1487 | void tcp_enter_loss(struct sock *sk, int how) | |
1488 | { | |
6687e988 | 1489 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
1490 | struct tcp_sock *tp = tcp_sk(sk); |
1491 | struct sk_buff *skb; | |
1492 | int cnt = 0; | |
1493 | ||
1494 | /* Reduce ssthresh if it has not yet been made inside this window. */ | |
6687e988 ACM |
1495 | if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || |
1496 | (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { | |
1497 | tp->prior_ssthresh = tcp_current_ssthresh(sk); | |
1498 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
1499 | tcp_ca_event(sk, CA_EVENT_LOSS); | |
1da177e4 LT |
1500 | } |
1501 | tp->snd_cwnd = 1; | |
1502 | tp->snd_cwnd_cnt = 0; | |
1503 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1504 | ||
9772efb9 | 1505 | tp->bytes_acked = 0; |
1da177e4 LT |
1506 | tcp_clear_retrans(tp); |
1507 | ||
1508 | /* Push undo marker, if it was plain RTO and nothing | |
1509 | * was retransmitted. */ | |
1510 | if (!how) | |
1511 | tp->undo_marker = tp->snd_una; | |
1512 | ||
fe067e8a DM |
1513 | tcp_for_write_queue(skb, sk) { |
1514 | if (skb == tcp_send_head(sk)) | |
1515 | break; | |
1da177e4 LT |
1516 | cnt += tcp_skb_pcount(skb); |
1517 | if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS) | |
1518 | tp->undo_marker = 0; | |
1519 | TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; | |
1520 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { | |
1521 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; | |
1522 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1523 | tp->lost_out += tcp_skb_pcount(skb); | |
1524 | } else { | |
1525 | tp->sacked_out += tcp_skb_pcount(skb); | |
1526 | tp->fackets_out = cnt; | |
1527 | } | |
1528 | } | |
1529 | tcp_sync_left_out(tp); | |
1530 | ||
1531 | tp->reordering = min_t(unsigned int, tp->reordering, | |
1532 | sysctl_tcp_reordering); | |
6687e988 | 1533 | tcp_set_ca_state(sk, TCP_CA_Loss); |
1da177e4 LT |
1534 | tp->high_seq = tp->snd_nxt; |
1535 | TCP_ECN_queue_cwr(tp); | |
580e572a IJ |
1536 | /* Abort FRTO algorithm if one is in progress */ |
1537 | tp->frto_counter = 0; | |
6a438bbe SH |
1538 | |
1539 | clear_all_retrans_hints(tp); | |
1da177e4 LT |
1540 | } |
1541 | ||
463c84b9 | 1542 | static int tcp_check_sack_reneging(struct sock *sk) |
1da177e4 LT |
1543 | { |
1544 | struct sk_buff *skb; | |
1545 | ||
1546 | /* If ACK arrived pointing to a remembered SACK, | |
1547 | * it means that our remembered SACKs do not reflect | |
1548 | * real state of receiver i.e. | |
1549 | * receiver _host_ is heavily congested (or buggy). | |
1550 | * Do processing similar to RTO timeout. | |
1551 | */ | |
fe067e8a | 1552 | if ((skb = tcp_write_queue_head(sk)) != NULL && |
1da177e4 | 1553 | (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { |
6687e988 | 1554 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
1555 | NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING); |
1556 | ||
1557 | tcp_enter_loss(sk, 1); | |
6687e988 | 1558 | icsk->icsk_retransmits++; |
fe067e8a | 1559 | tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); |
463c84b9 | 1560 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
6687e988 | 1561 | icsk->icsk_rto, TCP_RTO_MAX); |
1da177e4 LT |
1562 | return 1; |
1563 | } | |
1564 | return 0; | |
1565 | } | |
1566 | ||
1567 | static inline int tcp_fackets_out(struct tcp_sock *tp) | |
1568 | { | |
1569 | return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out; | |
1570 | } | |
1571 | ||
463c84b9 | 1572 | static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 1573 | { |
463c84b9 | 1574 | return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto); |
1da177e4 LT |
1575 | } |
1576 | ||
9e412ba7 | 1577 | static inline int tcp_head_timedout(struct sock *sk) |
1da177e4 | 1578 | { |
9e412ba7 IJ |
1579 | struct tcp_sock *tp = tcp_sk(sk); |
1580 | ||
1da177e4 | 1581 | return tp->packets_out && |
fe067e8a | 1582 | tcp_skb_timedout(sk, tcp_write_queue_head(sk)); |
1da177e4 LT |
1583 | } |
1584 | ||
1585 | /* Linux NewReno/SACK/FACK/ECN state machine. | |
1586 | * -------------------------------------- | |
1587 | * | |
1588 | * "Open" Normal state, no dubious events, fast path. | |
1589 | * "Disorder" In all the respects it is "Open", | |
1590 | * but requires a bit more attention. It is entered when | |
1591 | * we see some SACKs or dupacks. It is split of "Open" | |
1592 | * mainly to move some processing from fast path to slow one. | |
1593 | * "CWR" CWND was reduced due to some Congestion Notification event. | |
1594 | * It can be ECN, ICMP source quench, local device congestion. | |
1595 | * "Recovery" CWND was reduced, we are fast-retransmitting. | |
1596 | * "Loss" CWND was reduced due to RTO timeout or SACK reneging. | |
1597 | * | |
1598 | * tcp_fastretrans_alert() is entered: | |
1599 | * - each incoming ACK, if state is not "Open" | |
1600 | * - when arrived ACK is unusual, namely: | |
1601 | * * SACK | |
1602 | * * Duplicate ACK. | |
1603 | * * ECN ECE. | |
1604 | * | |
1605 | * Counting packets in flight is pretty simple. | |
1606 | * | |
1607 | * in_flight = packets_out - left_out + retrans_out | |
1608 | * | |
1609 | * packets_out is SND.NXT-SND.UNA counted in packets. | |
1610 | * | |
1611 | * retrans_out is number of retransmitted segments. | |
1612 | * | |
1613 | * left_out is number of segments left network, but not ACKed yet. | |
1614 | * | |
1615 | * left_out = sacked_out + lost_out | |
1616 | * | |
1617 | * sacked_out: Packets, which arrived to receiver out of order | |
1618 | * and hence not ACKed. With SACKs this number is simply | |
1619 | * amount of SACKed data. Even without SACKs | |
1620 | * it is easy to give pretty reliable estimate of this number, | |
1621 | * counting duplicate ACKs. | |
1622 | * | |
1623 | * lost_out: Packets lost by network. TCP has no explicit | |
1624 | * "loss notification" feedback from network (for now). | |
1625 | * It means that this number can be only _guessed_. | |
1626 | * Actually, it is the heuristics to predict lossage that | |
1627 | * distinguishes different algorithms. | |
1628 | * | |
1629 | * F.e. after RTO, when all the queue is considered as lost, | |
1630 | * lost_out = packets_out and in_flight = retrans_out. | |
1631 | * | |
1632 | * Essentially, we have now two algorithms counting | |
1633 | * lost packets. | |
1634 | * | |
1635 | * FACK: It is the simplest heuristics. As soon as we decided | |
1636 | * that something is lost, we decide that _all_ not SACKed | |
1637 | * packets until the most forward SACK are lost. I.e. | |
1638 | * lost_out = fackets_out - sacked_out and left_out = fackets_out. | |
1639 | * It is absolutely correct estimate, if network does not reorder | |
1640 | * packets. And it loses any connection to reality when reordering | |
1641 | * takes place. We use FACK by default until reordering | |
1642 | * is suspected on the path to this destination. | |
1643 | * | |
1644 | * NewReno: when Recovery is entered, we assume that one segment | |
1645 | * is lost (classic Reno). While we are in Recovery and | |
1646 | * a partial ACK arrives, we assume that one more packet | |
1647 | * is lost (NewReno). This heuristics are the same in NewReno | |
1648 | * and SACK. | |
1649 | * | |
1650 | * Imagine, that's all! Forget about all this shamanism about CWND inflation | |
1651 | * deflation etc. CWND is real congestion window, never inflated, changes | |
1652 | * only according to classic VJ rules. | |
1653 | * | |
1654 | * Really tricky (and requiring careful tuning) part of algorithm | |
1655 | * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). | |
1656 | * The first determines the moment _when_ we should reduce CWND and, | |
1657 | * hence, slow down forward transmission. In fact, it determines the moment | |
1658 | * when we decide that hole is caused by loss, rather than by a reorder. | |
1659 | * | |
1660 | * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill | |
1661 | * holes, caused by lost packets. | |
1662 | * | |
1663 | * And the most logically complicated part of algorithm is undo | |
1664 | * heuristics. We detect false retransmits due to both too early | |
1665 | * fast retransmit (reordering) and underestimated RTO, analyzing | |
1666 | * timestamps and D-SACKs. When we detect that some segments were | |
1667 | * retransmitted by mistake and CWND reduction was wrong, we undo | |
1668 | * window reduction and abort recovery phase. This logic is hidden | |
1669 | * inside several functions named tcp_try_undo_<something>. | |
1670 | */ | |
1671 | ||
1672 | /* This function decides, when we should leave Disordered state | |
1673 | * and enter Recovery phase, reducing congestion window. | |
1674 | * | |
1675 | * Main question: may we further continue forward transmission | |
1676 | * with the same cwnd? | |
1677 | */ | |
9e412ba7 | 1678 | static int tcp_time_to_recover(struct sock *sk) |
1da177e4 | 1679 | { |
9e412ba7 | 1680 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
1681 | __u32 packets_out; |
1682 | ||
52c63f1e IJ |
1683 | /* Do not perform any recovery during FRTO algorithm */ |
1684 | if (tp->frto_counter) | |
1685 | return 0; | |
1686 | ||
1da177e4 LT |
1687 | /* Trick#1: The loss is proven. */ |
1688 | if (tp->lost_out) | |
1689 | return 1; | |
1690 | ||
1691 | /* Not-A-Trick#2 : Classic rule... */ | |
1692 | if (tcp_fackets_out(tp) > tp->reordering) | |
1693 | return 1; | |
1694 | ||
1695 | /* Trick#3 : when we use RFC2988 timer restart, fast | |
1696 | * retransmit can be triggered by timeout of queue head. | |
1697 | */ | |
9e412ba7 | 1698 | if (tcp_head_timedout(sk)) |
1da177e4 LT |
1699 | return 1; |
1700 | ||
1701 | /* Trick#4: It is still not OK... But will it be useful to delay | |
1702 | * recovery more? | |
1703 | */ | |
1704 | packets_out = tp->packets_out; | |
1705 | if (packets_out <= tp->reordering && | |
1706 | tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && | |
9e412ba7 | 1707 | !tcp_may_send_now(sk)) { |
1da177e4 LT |
1708 | /* We have nothing to send. This connection is limited |
1709 | * either by receiver window or by application. | |
1710 | */ | |
1711 | return 1; | |
1712 | } | |
1713 | ||
1714 | return 0; | |
1715 | } | |
1716 | ||
1717 | /* If we receive more dupacks than we expected counting segments | |
1718 | * in assumption of absent reordering, interpret this as reordering. | |
1719 | * The only another reason could be bug in receiver TCP. | |
1720 | */ | |
6687e988 | 1721 | static void tcp_check_reno_reordering(struct sock *sk, const int addend) |
1da177e4 | 1722 | { |
6687e988 | 1723 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
1724 | u32 holes; |
1725 | ||
1726 | holes = max(tp->lost_out, 1U); | |
1727 | holes = min(holes, tp->packets_out); | |
1728 | ||
1729 | if ((tp->sacked_out + holes) > tp->packets_out) { | |
1730 | tp->sacked_out = tp->packets_out - holes; | |
6687e988 | 1731 | tcp_update_reordering(sk, tp->packets_out + addend, 0); |
1da177e4 LT |
1732 | } |
1733 | } | |
1734 | ||
1735 | /* Emulate SACKs for SACKless connection: account for a new dupack. */ | |
1736 | ||
6687e988 | 1737 | static void tcp_add_reno_sack(struct sock *sk) |
1da177e4 | 1738 | { |
6687e988 | 1739 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 1740 | tp->sacked_out++; |
6687e988 | 1741 | tcp_check_reno_reordering(sk, 0); |
1da177e4 LT |
1742 | tcp_sync_left_out(tp); |
1743 | } | |
1744 | ||
1745 | /* Account for ACK, ACKing some data in Reno Recovery phase. */ | |
1746 | ||
9e412ba7 | 1747 | static void tcp_remove_reno_sacks(struct sock *sk, int acked) |
1da177e4 | 1748 | { |
9e412ba7 IJ |
1749 | struct tcp_sock *tp = tcp_sk(sk); |
1750 | ||
1da177e4 LT |
1751 | if (acked > 0) { |
1752 | /* One ACK acked hole. The rest eat duplicate ACKs. */ | |
1753 | if (acked-1 >= tp->sacked_out) | |
1754 | tp->sacked_out = 0; | |
1755 | else | |
1756 | tp->sacked_out -= acked-1; | |
1757 | } | |
6687e988 | 1758 | tcp_check_reno_reordering(sk, acked); |
1da177e4 LT |
1759 | tcp_sync_left_out(tp); |
1760 | } | |
1761 | ||
1762 | static inline void tcp_reset_reno_sack(struct tcp_sock *tp) | |
1763 | { | |
1764 | tp->sacked_out = 0; | |
1765 | tp->left_out = tp->lost_out; | |
1766 | } | |
1767 | ||
d8f4f223 IJ |
1768 | /* RFC: This is from the original, I doubt that this is necessary at all: |
1769 | * clear xmit_retrans hint if seq of this skb is beyond hint. How could we | |
1770 | * retransmitted past LOST markings in the first place? I'm not fully sure | |
1771 | * about undo and end of connection cases, which can cause R without L? | |
1772 | */ | |
1773 | static void tcp_verify_retransmit_hint(struct tcp_sock *tp, | |
1774 | struct sk_buff *skb) | |
1775 | { | |
1776 | if ((tp->retransmit_skb_hint != NULL) && | |
1777 | before(TCP_SKB_CB(skb)->seq, | |
1778 | TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) | |
19b2b486 | 1779 | tp->retransmit_skb_hint = NULL; |
d8f4f223 IJ |
1780 | } |
1781 | ||
1da177e4 | 1782 | /* Mark head of queue up as lost. */ |
9e412ba7 | 1783 | static void tcp_mark_head_lost(struct sock *sk, |
1da177e4 LT |
1784 | int packets, u32 high_seq) |
1785 | { | |
9e412ba7 | 1786 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 1787 | struct sk_buff *skb; |
6a438bbe | 1788 | int cnt; |
1da177e4 | 1789 | |
6a438bbe SH |
1790 | BUG_TRAP(packets <= tp->packets_out); |
1791 | if (tp->lost_skb_hint) { | |
1792 | skb = tp->lost_skb_hint; | |
1793 | cnt = tp->lost_cnt_hint; | |
1794 | } else { | |
fe067e8a | 1795 | skb = tcp_write_queue_head(sk); |
6a438bbe SH |
1796 | cnt = 0; |
1797 | } | |
1da177e4 | 1798 | |
fe067e8a DM |
1799 | tcp_for_write_queue_from(skb, sk) { |
1800 | if (skb == tcp_send_head(sk)) | |
1801 | break; | |
6a438bbe SH |
1802 | /* TODO: do this better */ |
1803 | /* this is not the most efficient way to do this... */ | |
1804 | tp->lost_skb_hint = skb; | |
1805 | tp->lost_cnt_hint = cnt; | |
1806 | cnt += tcp_skb_pcount(skb); | |
1807 | if (cnt > packets || after(TCP_SKB_CB(skb)->end_seq, high_seq)) | |
1da177e4 LT |
1808 | break; |
1809 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) { | |
1810 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1811 | tp->lost_out += tcp_skb_pcount(skb); | |
d8f4f223 | 1812 | tcp_verify_retransmit_hint(tp, skb); |
1da177e4 LT |
1813 | } |
1814 | } | |
1815 | tcp_sync_left_out(tp); | |
1816 | } | |
1817 | ||
1818 | /* Account newly detected lost packet(s) */ | |
1819 | ||
9e412ba7 | 1820 | static void tcp_update_scoreboard(struct sock *sk) |
1da177e4 | 1821 | { |
9e412ba7 IJ |
1822 | struct tcp_sock *tp = tcp_sk(sk); |
1823 | ||
1da177e4 LT |
1824 | if (IsFack(tp)) { |
1825 | int lost = tp->fackets_out - tp->reordering; | |
1826 | if (lost <= 0) | |
1827 | lost = 1; | |
9e412ba7 | 1828 | tcp_mark_head_lost(sk, lost, tp->high_seq); |
1da177e4 | 1829 | } else { |
9e412ba7 | 1830 | tcp_mark_head_lost(sk, 1, tp->high_seq); |
1da177e4 LT |
1831 | } |
1832 | ||
1833 | /* New heuristics: it is possible only after we switched | |
1834 | * to restart timer each time when something is ACKed. | |
1835 | * Hence, we can detect timed out packets during fast | |
1836 | * retransmit without falling to slow start. | |
1837 | */ | |
9e412ba7 | 1838 | if (!IsReno(tp) && tcp_head_timedout(sk)) { |
1da177e4 LT |
1839 | struct sk_buff *skb; |
1840 | ||
6a438bbe | 1841 | skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint |
fe067e8a | 1842 | : tcp_write_queue_head(sk); |
6a438bbe | 1843 | |
fe067e8a DM |
1844 | tcp_for_write_queue_from(skb, sk) { |
1845 | if (skb == tcp_send_head(sk)) | |
1846 | break; | |
6a438bbe SH |
1847 | if (!tcp_skb_timedout(sk, skb)) |
1848 | break; | |
1849 | ||
1850 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) { | |
1da177e4 LT |
1851 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; |
1852 | tp->lost_out += tcp_skb_pcount(skb); | |
d8f4f223 | 1853 | tcp_verify_retransmit_hint(tp, skb); |
1da177e4 LT |
1854 | } |
1855 | } | |
6a438bbe SH |
1856 | |
1857 | tp->scoreboard_skb_hint = skb; | |
1858 | ||
1da177e4 LT |
1859 | tcp_sync_left_out(tp); |
1860 | } | |
1861 | } | |
1862 | ||
1863 | /* CWND moderation, preventing bursts due to too big ACKs | |
1864 | * in dubious situations. | |
1865 | */ | |
1866 | static inline void tcp_moderate_cwnd(struct tcp_sock *tp) | |
1867 | { | |
1868 | tp->snd_cwnd = min(tp->snd_cwnd, | |
1869 | tcp_packets_in_flight(tp)+tcp_max_burst(tp)); | |
1870 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1871 | } | |
1872 | ||
72dc5b92 SH |
1873 | /* Lower bound on congestion window is slow start threshold |
1874 | * unless congestion avoidance choice decides to overide it. | |
1875 | */ | |
1876 | static inline u32 tcp_cwnd_min(const struct sock *sk) | |
1877 | { | |
1878 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; | |
1879 | ||
1880 | return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh; | |
1881 | } | |
1882 | ||
1da177e4 | 1883 | /* Decrease cwnd each second ack. */ |
1e757f99 | 1884 | static void tcp_cwnd_down(struct sock *sk, int flag) |
1da177e4 | 1885 | { |
6687e988 | 1886 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 1887 | int decr = tp->snd_cwnd_cnt + 1; |
1da177e4 | 1888 | |
49ff4bb4 | 1889 | if ((flag&(FLAG_ANY_PROGRESS|FLAG_DSACKING_ACK)) || |
1e757f99 IJ |
1890 | (IsReno(tp) && !(flag&FLAG_NOT_DUP))) { |
1891 | tp->snd_cwnd_cnt = decr&1; | |
1892 | decr >>= 1; | |
1da177e4 | 1893 | |
1e757f99 IJ |
1894 | if (decr && tp->snd_cwnd > tcp_cwnd_min(sk)) |
1895 | tp->snd_cwnd -= decr; | |
1da177e4 | 1896 | |
1e757f99 IJ |
1897 | tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1); |
1898 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1899 | } | |
1da177e4 LT |
1900 | } |
1901 | ||
1902 | /* Nothing was retransmitted or returned timestamp is less | |
1903 | * than timestamp of the first retransmission. | |
1904 | */ | |
1905 | static inline int tcp_packet_delayed(struct tcp_sock *tp) | |
1906 | { | |
1907 | return !tp->retrans_stamp || | |
1908 | (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && | |
1909 | (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0); | |
1910 | } | |
1911 | ||
1912 | /* Undo procedures. */ | |
1913 | ||
1914 | #if FASTRETRANS_DEBUG > 1 | |
9e412ba7 | 1915 | static void DBGUNDO(struct sock *sk, const char *msg) |
1da177e4 | 1916 | { |
9e412ba7 | 1917 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 1918 | struct inet_sock *inet = inet_sk(sk); |
9e412ba7 | 1919 | |
1da177e4 LT |
1920 | printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n", |
1921 | msg, | |
1922 | NIPQUAD(inet->daddr), ntohs(inet->dport), | |
1923 | tp->snd_cwnd, tp->left_out, | |
1924 | tp->snd_ssthresh, tp->prior_ssthresh, | |
1925 | tp->packets_out); | |
1926 | } | |
1927 | #else | |
1928 | #define DBGUNDO(x...) do { } while (0) | |
1929 | #endif | |
1930 | ||
6687e988 | 1931 | static void tcp_undo_cwr(struct sock *sk, const int undo) |
1da177e4 | 1932 | { |
6687e988 ACM |
1933 | struct tcp_sock *tp = tcp_sk(sk); |
1934 | ||
1da177e4 | 1935 | if (tp->prior_ssthresh) { |
6687e988 ACM |
1936 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1937 | ||
1938 | if (icsk->icsk_ca_ops->undo_cwnd) | |
1939 | tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); | |
1da177e4 LT |
1940 | else |
1941 | tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1); | |
1942 | ||
1943 | if (undo && tp->prior_ssthresh > tp->snd_ssthresh) { | |
1944 | tp->snd_ssthresh = tp->prior_ssthresh; | |
1945 | TCP_ECN_withdraw_cwr(tp); | |
1946 | } | |
1947 | } else { | |
1948 | tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); | |
1949 | } | |
1950 | tcp_moderate_cwnd(tp); | |
1951 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
6a438bbe SH |
1952 | |
1953 | /* There is something screwy going on with the retrans hints after | |
1954 | an undo */ | |
1955 | clear_all_retrans_hints(tp); | |
1da177e4 LT |
1956 | } |
1957 | ||
1958 | static inline int tcp_may_undo(struct tcp_sock *tp) | |
1959 | { | |
1960 | return tp->undo_marker && | |
1961 | (!tp->undo_retrans || tcp_packet_delayed(tp)); | |
1962 | } | |
1963 | ||
1964 | /* People celebrate: "We love our President!" */ | |
9e412ba7 | 1965 | static int tcp_try_undo_recovery(struct sock *sk) |
1da177e4 | 1966 | { |
9e412ba7 IJ |
1967 | struct tcp_sock *tp = tcp_sk(sk); |
1968 | ||
1da177e4 LT |
1969 | if (tcp_may_undo(tp)) { |
1970 | /* Happy end! We did not retransmit anything | |
1971 | * or our original transmission succeeded. | |
1972 | */ | |
9e412ba7 | 1973 | DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); |
6687e988 ACM |
1974 | tcp_undo_cwr(sk, 1); |
1975 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) | |
1da177e4 LT |
1976 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); |
1977 | else | |
1978 | NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO); | |
1979 | tp->undo_marker = 0; | |
1980 | } | |
1981 | if (tp->snd_una == tp->high_seq && IsReno(tp)) { | |
1982 | /* Hold old state until something *above* high_seq | |
1983 | * is ACKed. For Reno it is MUST to prevent false | |
1984 | * fast retransmits (RFC2582). SACK TCP is safe. */ | |
1985 | tcp_moderate_cwnd(tp); | |
1986 | return 1; | |
1987 | } | |
6687e988 | 1988 | tcp_set_ca_state(sk, TCP_CA_Open); |
1da177e4 LT |
1989 | return 0; |
1990 | } | |
1991 | ||
1992 | /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ | |
9e412ba7 | 1993 | static void tcp_try_undo_dsack(struct sock *sk) |
1da177e4 | 1994 | { |
9e412ba7 IJ |
1995 | struct tcp_sock *tp = tcp_sk(sk); |
1996 | ||
1da177e4 | 1997 | if (tp->undo_marker && !tp->undo_retrans) { |
9e412ba7 | 1998 | DBGUNDO(sk, "D-SACK"); |
6687e988 | 1999 | tcp_undo_cwr(sk, 1); |
1da177e4 LT |
2000 | tp->undo_marker = 0; |
2001 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO); | |
2002 | } | |
2003 | } | |
2004 | ||
2005 | /* Undo during fast recovery after partial ACK. */ | |
2006 | ||
9e412ba7 | 2007 | static int tcp_try_undo_partial(struct sock *sk, int acked) |
1da177e4 | 2008 | { |
9e412ba7 | 2009 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
2010 | /* Partial ACK arrived. Force Hoe's retransmit. */ |
2011 | int failed = IsReno(tp) || tp->fackets_out>tp->reordering; | |
2012 | ||
2013 | if (tcp_may_undo(tp)) { | |
2014 | /* Plain luck! Hole if filled with delayed | |
2015 | * packet, rather than with a retransmit. | |
2016 | */ | |
2017 | if (tp->retrans_out == 0) | |
2018 | tp->retrans_stamp = 0; | |
2019 | ||
6687e988 | 2020 | tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); |
1da177e4 | 2021 | |
9e412ba7 | 2022 | DBGUNDO(sk, "Hoe"); |
6687e988 | 2023 | tcp_undo_cwr(sk, 0); |
1da177e4 LT |
2024 | NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO); |
2025 | ||
2026 | /* So... Do not make Hoe's retransmit yet. | |
2027 | * If the first packet was delayed, the rest | |
2028 | * ones are most probably delayed as well. | |
2029 | */ | |
2030 | failed = 0; | |
2031 | } | |
2032 | return failed; | |
2033 | } | |
2034 | ||
2035 | /* Undo during loss recovery after partial ACK. */ | |
9e412ba7 | 2036 | static int tcp_try_undo_loss(struct sock *sk) |
1da177e4 | 2037 | { |
9e412ba7 IJ |
2038 | struct tcp_sock *tp = tcp_sk(sk); |
2039 | ||
1da177e4 LT |
2040 | if (tcp_may_undo(tp)) { |
2041 | struct sk_buff *skb; | |
fe067e8a DM |
2042 | tcp_for_write_queue(skb, sk) { |
2043 | if (skb == tcp_send_head(sk)) | |
2044 | break; | |
1da177e4 LT |
2045 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; |
2046 | } | |
6a438bbe SH |
2047 | |
2048 | clear_all_retrans_hints(tp); | |
2049 | ||
9e412ba7 | 2050 | DBGUNDO(sk, "partial loss"); |
1da177e4 LT |
2051 | tp->lost_out = 0; |
2052 | tp->left_out = tp->sacked_out; | |
6687e988 | 2053 | tcp_undo_cwr(sk, 1); |
1da177e4 | 2054 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); |
463c84b9 | 2055 | inet_csk(sk)->icsk_retransmits = 0; |
1da177e4 LT |
2056 | tp->undo_marker = 0; |
2057 | if (!IsReno(tp)) | |
6687e988 | 2058 | tcp_set_ca_state(sk, TCP_CA_Open); |
1da177e4 LT |
2059 | return 1; |
2060 | } | |
2061 | return 0; | |
2062 | } | |
2063 | ||
6687e988 | 2064 | static inline void tcp_complete_cwr(struct sock *sk) |
1da177e4 | 2065 | { |
6687e988 | 2066 | struct tcp_sock *tp = tcp_sk(sk); |
317a76f9 | 2067 | tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); |
1da177e4 | 2068 | tp->snd_cwnd_stamp = tcp_time_stamp; |
6687e988 | 2069 | tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); |
1da177e4 LT |
2070 | } |
2071 | ||
9e412ba7 | 2072 | static void tcp_try_to_open(struct sock *sk, int flag) |
1da177e4 | 2073 | { |
9e412ba7 IJ |
2074 | struct tcp_sock *tp = tcp_sk(sk); |
2075 | ||
af15cc7b | 2076 | tcp_sync_left_out(tp); |
1da177e4 LT |
2077 | |
2078 | if (tp->retrans_out == 0) | |
2079 | tp->retrans_stamp = 0; | |
2080 | ||
2081 | if (flag&FLAG_ECE) | |
3cfe3baa | 2082 | tcp_enter_cwr(sk, 1); |
1da177e4 | 2083 | |
6687e988 | 2084 | if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { |
1da177e4 LT |
2085 | int state = TCP_CA_Open; |
2086 | ||
2087 | if (tp->left_out || tp->retrans_out || tp->undo_marker) | |
2088 | state = TCP_CA_Disorder; | |
2089 | ||
6687e988 ACM |
2090 | if (inet_csk(sk)->icsk_ca_state != state) { |
2091 | tcp_set_ca_state(sk, state); | |
1da177e4 LT |
2092 | tp->high_seq = tp->snd_nxt; |
2093 | } | |
2094 | tcp_moderate_cwnd(tp); | |
2095 | } else { | |
1e757f99 | 2096 | tcp_cwnd_down(sk, flag); |
1da177e4 LT |
2097 | } |
2098 | } | |
2099 | ||
5d424d5a JH |
2100 | static void tcp_mtup_probe_failed(struct sock *sk) |
2101 | { | |
2102 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2103 | ||
2104 | icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; | |
2105 | icsk->icsk_mtup.probe_size = 0; | |
2106 | } | |
2107 | ||
2108 | static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb) | |
2109 | { | |
2110 | struct tcp_sock *tp = tcp_sk(sk); | |
2111 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2112 | ||
2113 | /* FIXME: breaks with very large cwnd */ | |
2114 | tp->prior_ssthresh = tcp_current_ssthresh(sk); | |
2115 | tp->snd_cwnd = tp->snd_cwnd * | |
2116 | tcp_mss_to_mtu(sk, tp->mss_cache) / | |
2117 | icsk->icsk_mtup.probe_size; | |
2118 | tp->snd_cwnd_cnt = 0; | |
2119 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
2120 | tp->rcv_ssthresh = tcp_current_ssthresh(sk); | |
2121 | ||
2122 | icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; | |
2123 | icsk->icsk_mtup.probe_size = 0; | |
2124 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); | |
2125 | } | |
2126 | ||
2127 | ||
1da177e4 LT |
2128 | /* Process an event, which can update packets-in-flight not trivially. |
2129 | * Main goal of this function is to calculate new estimate for left_out, | |
2130 | * taking into account both packets sitting in receiver's buffer and | |
2131 | * packets lost by network. | |
2132 | * | |
2133 | * Besides that it does CWND reduction, when packet loss is detected | |
2134 | * and changes state of machine. | |
2135 | * | |
2136 | * It does _not_ decide what to send, it is made in function | |
2137 | * tcp_xmit_retransmit_queue(). | |
2138 | */ | |
2139 | static void | |
2e605294 | 2140 | tcp_fastretrans_alert(struct sock *sk, int prior_packets, int flag) |
1da177e4 | 2141 | { |
6687e988 | 2142 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 2143 | struct tcp_sock *tp = tcp_sk(sk); |
2e605294 IJ |
2144 | int is_dupack = !(flag&(FLAG_SND_UNA_ADVANCED|FLAG_NOT_DUP)); |
2145 | int do_lost = is_dupack || ((flag&FLAG_DATA_SACKED) && | |
2146 | (tp->fackets_out > tp->reordering)); | |
1da177e4 LT |
2147 | |
2148 | /* Some technical things: | |
2149 | * 1. Reno does not count dupacks (sacked_out) automatically. */ | |
2150 | if (!tp->packets_out) | |
2151 | tp->sacked_out = 0; | |
e905a9ed | 2152 | /* 2. SACK counts snd_fack in packets inaccurately. */ |
1da177e4 LT |
2153 | if (tp->sacked_out == 0) |
2154 | tp->fackets_out = 0; | |
2155 | ||
e905a9ed | 2156 | /* Now state machine starts. |
1da177e4 LT |
2157 | * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ |
2158 | if (flag&FLAG_ECE) | |
2159 | tp->prior_ssthresh = 0; | |
2160 | ||
2161 | /* B. In all the states check for reneging SACKs. */ | |
463c84b9 | 2162 | if (tp->sacked_out && tcp_check_sack_reneging(sk)) |
1da177e4 LT |
2163 | return; |
2164 | ||
2165 | /* C. Process data loss notification, provided it is valid. */ | |
2166 | if ((flag&FLAG_DATA_LOST) && | |
2167 | before(tp->snd_una, tp->high_seq) && | |
6687e988 | 2168 | icsk->icsk_ca_state != TCP_CA_Open && |
1da177e4 | 2169 | tp->fackets_out > tp->reordering) { |
9e412ba7 | 2170 | tcp_mark_head_lost(sk, tp->fackets_out-tp->reordering, tp->high_seq); |
1da177e4 LT |
2171 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSS); |
2172 | } | |
2173 | ||
2174 | /* D. Synchronize left_out to current state. */ | |
2175 | tcp_sync_left_out(tp); | |
2176 | ||
2177 | /* E. Check state exit conditions. State can be terminated | |
2178 | * when high_seq is ACKed. */ | |
6687e988 | 2179 | if (icsk->icsk_ca_state == TCP_CA_Open) { |
7b0eb22b | 2180 | BUG_TRAP(tp->retrans_out == 0); |
1da177e4 LT |
2181 | tp->retrans_stamp = 0; |
2182 | } else if (!before(tp->snd_una, tp->high_seq)) { | |
6687e988 | 2183 | switch (icsk->icsk_ca_state) { |
1da177e4 | 2184 | case TCP_CA_Loss: |
6687e988 | 2185 | icsk->icsk_retransmits = 0; |
9e412ba7 | 2186 | if (tcp_try_undo_recovery(sk)) |
1da177e4 LT |
2187 | return; |
2188 | break; | |
2189 | ||
2190 | case TCP_CA_CWR: | |
2191 | /* CWR is to be held something *above* high_seq | |
2192 | * is ACKed for CWR bit to reach receiver. */ | |
2193 | if (tp->snd_una != tp->high_seq) { | |
6687e988 ACM |
2194 | tcp_complete_cwr(sk); |
2195 | tcp_set_ca_state(sk, TCP_CA_Open); | |
1da177e4 LT |
2196 | } |
2197 | break; | |
2198 | ||
2199 | case TCP_CA_Disorder: | |
9e412ba7 | 2200 | tcp_try_undo_dsack(sk); |
1da177e4 LT |
2201 | if (!tp->undo_marker || |
2202 | /* For SACK case do not Open to allow to undo | |
2203 | * catching for all duplicate ACKs. */ | |
2204 | IsReno(tp) || tp->snd_una != tp->high_seq) { | |
2205 | tp->undo_marker = 0; | |
6687e988 | 2206 | tcp_set_ca_state(sk, TCP_CA_Open); |
1da177e4 LT |
2207 | } |
2208 | break; | |
2209 | ||
2210 | case TCP_CA_Recovery: | |
2211 | if (IsReno(tp)) | |
2212 | tcp_reset_reno_sack(tp); | |
9e412ba7 | 2213 | if (tcp_try_undo_recovery(sk)) |
1da177e4 | 2214 | return; |
6687e988 | 2215 | tcp_complete_cwr(sk); |
1da177e4 LT |
2216 | break; |
2217 | } | |
2218 | } | |
2219 | ||
2220 | /* F. Process state. */ | |
6687e988 | 2221 | switch (icsk->icsk_ca_state) { |
1da177e4 | 2222 | case TCP_CA_Recovery: |
2e605294 | 2223 | if (!(flag & FLAG_SND_UNA_ADVANCED)) { |
1da177e4 | 2224 | if (IsReno(tp) && is_dupack) |
6687e988 | 2225 | tcp_add_reno_sack(sk); |
1da177e4 LT |
2226 | } else { |
2227 | int acked = prior_packets - tp->packets_out; | |
2228 | if (IsReno(tp)) | |
9e412ba7 | 2229 | tcp_remove_reno_sacks(sk, acked); |
2e605294 | 2230 | do_lost = tcp_try_undo_partial(sk, acked); |
1da177e4 LT |
2231 | } |
2232 | break; | |
2233 | case TCP_CA_Loss: | |
2234 | if (flag&FLAG_DATA_ACKED) | |
6687e988 | 2235 | icsk->icsk_retransmits = 0; |
9e412ba7 | 2236 | if (!tcp_try_undo_loss(sk)) { |
1da177e4 LT |
2237 | tcp_moderate_cwnd(tp); |
2238 | tcp_xmit_retransmit_queue(sk); | |
2239 | return; | |
2240 | } | |
6687e988 | 2241 | if (icsk->icsk_ca_state != TCP_CA_Open) |
1da177e4 LT |
2242 | return; |
2243 | /* Loss is undone; fall through to processing in Open state. */ | |
2244 | default: | |
2245 | if (IsReno(tp)) { | |
2e605294 | 2246 | if (flag & FLAG_SND_UNA_ADVANCED) |
1da177e4 LT |
2247 | tcp_reset_reno_sack(tp); |
2248 | if (is_dupack) | |
6687e988 | 2249 | tcp_add_reno_sack(sk); |
1da177e4 LT |
2250 | } |
2251 | ||
6687e988 | 2252 | if (icsk->icsk_ca_state == TCP_CA_Disorder) |
9e412ba7 | 2253 | tcp_try_undo_dsack(sk); |
1da177e4 | 2254 | |
9e412ba7 IJ |
2255 | if (!tcp_time_to_recover(sk)) { |
2256 | tcp_try_to_open(sk, flag); | |
1da177e4 LT |
2257 | return; |
2258 | } | |
2259 | ||
5d424d5a JH |
2260 | /* MTU probe failure: don't reduce cwnd */ |
2261 | if (icsk->icsk_ca_state < TCP_CA_CWR && | |
2262 | icsk->icsk_mtup.probe_size && | |
0e7b1368 | 2263 | tp->snd_una == tp->mtu_probe.probe_seq_start) { |
5d424d5a JH |
2264 | tcp_mtup_probe_failed(sk); |
2265 | /* Restores the reduction we did in tcp_mtup_probe() */ | |
2266 | tp->snd_cwnd++; | |
2267 | tcp_simple_retransmit(sk); | |
2268 | return; | |
2269 | } | |
2270 | ||
1da177e4 LT |
2271 | /* Otherwise enter Recovery state */ |
2272 | ||
2273 | if (IsReno(tp)) | |
2274 | NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY); | |
2275 | else | |
2276 | NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY); | |
2277 | ||
2278 | tp->high_seq = tp->snd_nxt; | |
2279 | tp->prior_ssthresh = 0; | |
2280 | tp->undo_marker = tp->snd_una; | |
2281 | tp->undo_retrans = tp->retrans_out; | |
2282 | ||
6687e988 | 2283 | if (icsk->icsk_ca_state < TCP_CA_CWR) { |
1da177e4 | 2284 | if (!(flag&FLAG_ECE)) |
6687e988 ACM |
2285 | tp->prior_ssthresh = tcp_current_ssthresh(sk); |
2286 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
1da177e4 LT |
2287 | TCP_ECN_queue_cwr(tp); |
2288 | } | |
2289 | ||
9772efb9 | 2290 | tp->bytes_acked = 0; |
1da177e4 | 2291 | tp->snd_cwnd_cnt = 0; |
6687e988 | 2292 | tcp_set_ca_state(sk, TCP_CA_Recovery); |
1da177e4 LT |
2293 | } |
2294 | ||
2e605294 | 2295 | if (do_lost || tcp_head_timedout(sk)) |
9e412ba7 | 2296 | tcp_update_scoreboard(sk); |
1e757f99 | 2297 | tcp_cwnd_down(sk, flag); |
1da177e4 LT |
2298 | tcp_xmit_retransmit_queue(sk); |
2299 | } | |
2300 | ||
2301 | /* Read draft-ietf-tcplw-high-performance before mucking | |
caa20d9a | 2302 | * with this code. (Supersedes RFC1323) |
1da177e4 | 2303 | */ |
2d2abbab | 2304 | static void tcp_ack_saw_tstamp(struct sock *sk, int flag) |
1da177e4 | 2305 | { |
1da177e4 LT |
2306 | /* RTTM Rule: A TSecr value received in a segment is used to |
2307 | * update the averaged RTT measurement only if the segment | |
2308 | * acknowledges some new data, i.e., only if it advances the | |
2309 | * left edge of the send window. | |
2310 | * | |
2311 | * See draft-ietf-tcplw-high-performance-00, section 3.3. | |
2312 | * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> | |
2313 | * | |
2314 | * Changed: reset backoff as soon as we see the first valid sample. | |
caa20d9a | 2315 | * If we do not, we get strongly overestimated rto. With timestamps |
1da177e4 LT |
2316 | * samples are accepted even from very old segments: f.e., when rtt=1 |
2317 | * increases to 8, we retransmit 5 times and after 8 seconds delayed | |
2318 | * answer arrives rto becomes 120 seconds! If at least one of segments | |
2319 | * in window is lost... Voila. --ANK (010210) | |
2320 | */ | |
463c84b9 ACM |
2321 | struct tcp_sock *tp = tcp_sk(sk); |
2322 | const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr; | |
2d2abbab | 2323 | tcp_rtt_estimator(sk, seq_rtt); |
463c84b9 ACM |
2324 | tcp_set_rto(sk); |
2325 | inet_csk(sk)->icsk_backoff = 0; | |
2326 | tcp_bound_rto(sk); | |
1da177e4 LT |
2327 | } |
2328 | ||
2d2abbab | 2329 | static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) |
1da177e4 LT |
2330 | { |
2331 | /* We don't have a timestamp. Can only use | |
2332 | * packets that are not retransmitted to determine | |
2333 | * rtt estimates. Also, we must not reset the | |
2334 | * backoff for rto until we get a non-retransmitted | |
2335 | * packet. This allows us to deal with a situation | |
2336 | * where the network delay has increased suddenly. | |
2337 | * I.e. Karn's algorithm. (SIGCOMM '87, p5.) | |
2338 | */ | |
2339 | ||
2340 | if (flag & FLAG_RETRANS_DATA_ACKED) | |
2341 | return; | |
2342 | ||
2d2abbab | 2343 | tcp_rtt_estimator(sk, seq_rtt); |
463c84b9 ACM |
2344 | tcp_set_rto(sk); |
2345 | inet_csk(sk)->icsk_backoff = 0; | |
2346 | tcp_bound_rto(sk); | |
1da177e4 LT |
2347 | } |
2348 | ||
463c84b9 | 2349 | static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, |
2d2abbab | 2350 | const s32 seq_rtt) |
1da177e4 | 2351 | { |
463c84b9 | 2352 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
2353 | /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ |
2354 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) | |
2d2abbab | 2355 | tcp_ack_saw_tstamp(sk, flag); |
1da177e4 | 2356 | else if (seq_rtt >= 0) |
2d2abbab | 2357 | tcp_ack_no_tstamp(sk, seq_rtt, flag); |
1da177e4 LT |
2358 | } |
2359 | ||
16751347 | 2360 | static void tcp_cong_avoid(struct sock *sk, u32 ack, |
40efc6fa | 2361 | u32 in_flight, int good) |
1da177e4 | 2362 | { |
6687e988 | 2363 | const struct inet_connection_sock *icsk = inet_csk(sk); |
16751347 | 2364 | icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight, good); |
6687e988 | 2365 | tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; |
1da177e4 LT |
2366 | } |
2367 | ||
1da177e4 LT |
2368 | /* Restart timer after forward progress on connection. |
2369 | * RFC2988 recommends to restart timer to now+rto. | |
2370 | */ | |
2371 | ||
9e412ba7 | 2372 | static void tcp_ack_packets_out(struct sock *sk) |
1da177e4 | 2373 | { |
9e412ba7 IJ |
2374 | struct tcp_sock *tp = tcp_sk(sk); |
2375 | ||
1da177e4 | 2376 | if (!tp->packets_out) { |
463c84b9 | 2377 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); |
1da177e4 | 2378 | } else { |
3f421baa | 2379 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, inet_csk(sk)->icsk_rto, TCP_RTO_MAX); |
1da177e4 LT |
2380 | } |
2381 | } | |
2382 | ||
1da177e4 LT |
2383 | static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb, |
2384 | __u32 now, __s32 *seq_rtt) | |
2385 | { | |
2386 | struct tcp_sock *tp = tcp_sk(sk); | |
e905a9ed | 2387 | struct tcp_skb_cb *scb = TCP_SKB_CB(skb); |
1da177e4 LT |
2388 | __u32 seq = tp->snd_una; |
2389 | __u32 packets_acked; | |
2390 | int acked = 0; | |
2391 | ||
2392 | /* If we get here, the whole TSO packet has not been | |
2393 | * acked. | |
2394 | */ | |
2395 | BUG_ON(!after(scb->end_seq, seq)); | |
2396 | ||
2397 | packets_acked = tcp_skb_pcount(skb); | |
2398 | if (tcp_trim_head(sk, skb, seq - scb->seq)) | |
2399 | return 0; | |
2400 | packets_acked -= tcp_skb_pcount(skb); | |
2401 | ||
2402 | if (packets_acked) { | |
2403 | __u8 sacked = scb->sacked; | |
2404 | ||
2405 | acked |= FLAG_DATA_ACKED; | |
2406 | if (sacked) { | |
2407 | if (sacked & TCPCB_RETRANS) { | |
2408 | if (sacked & TCPCB_SACKED_RETRANS) | |
2409 | tp->retrans_out -= packets_acked; | |
2410 | acked |= FLAG_RETRANS_DATA_ACKED; | |
2411 | *seq_rtt = -1; | |
2412 | } else if (*seq_rtt < 0) | |
2413 | *seq_rtt = now - scb->when; | |
2414 | if (sacked & TCPCB_SACKED_ACKED) | |
2415 | tp->sacked_out -= packets_acked; | |
2416 | if (sacked & TCPCB_LOST) | |
2417 | tp->lost_out -= packets_acked; | |
2418 | if (sacked & TCPCB_URG) { | |
2419 | if (tp->urg_mode && | |
2420 | !before(seq, tp->snd_up)) | |
2421 | tp->urg_mode = 0; | |
2422 | } | |
2423 | } else if (*seq_rtt < 0) | |
2424 | *seq_rtt = now - scb->when; | |
2425 | ||
2426 | if (tp->fackets_out) { | |
2427 | __u32 dval = min(tp->fackets_out, packets_acked); | |
2428 | tp->fackets_out -= dval; | |
2429 | } | |
48611c47 IJ |
2430 | /* hint's skb might be NULL but we don't need to care */ |
2431 | tp->fastpath_cnt_hint -= min_t(u32, packets_acked, | |
2432 | tp->fastpath_cnt_hint); | |
1da177e4 LT |
2433 | tp->packets_out -= packets_acked; |
2434 | ||
2435 | BUG_ON(tcp_skb_pcount(skb) == 0); | |
2436 | BUG_ON(!before(scb->seq, scb->end_seq)); | |
2437 | } | |
2438 | ||
2439 | return acked; | |
2440 | } | |
2441 | ||
1da177e4 | 2442 | /* Remove acknowledged frames from the retransmission queue. */ |
2d2abbab | 2443 | static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p) |
1da177e4 LT |
2444 | { |
2445 | struct tcp_sock *tp = tcp_sk(sk); | |
2d2abbab | 2446 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2447 | struct sk_buff *skb; |
2448 | __u32 now = tcp_time_stamp; | |
2449 | int acked = 0; | |
6418204f | 2450 | int prior_packets = tp->packets_out; |
1da177e4 | 2451 | __s32 seq_rtt = -1; |
b9ce204f | 2452 | ktime_t last_ackt = net_invalid_timestamp(); |
1da177e4 | 2453 | |
fe067e8a DM |
2454 | while ((skb = tcp_write_queue_head(sk)) && |
2455 | skb != tcp_send_head(sk)) { | |
e905a9ed | 2456 | struct tcp_skb_cb *scb = TCP_SKB_CB(skb); |
1da177e4 LT |
2457 | __u8 sacked = scb->sacked; |
2458 | ||
2459 | /* If our packet is before the ack sequence we can | |
2460 | * discard it as it's confirmed to have arrived at | |
2461 | * the other end. | |
2462 | */ | |
2463 | if (after(scb->end_seq, tp->snd_una)) { | |
cb83199a DM |
2464 | if (tcp_skb_pcount(skb) > 1 && |
2465 | after(tp->snd_una, scb->seq)) | |
1da177e4 LT |
2466 | acked |= tcp_tso_acked(sk, skb, |
2467 | now, &seq_rtt); | |
2468 | break; | |
2469 | } | |
2470 | ||
2471 | /* Initial outgoing SYN's get put onto the write_queue | |
2472 | * just like anything else we transmit. It is not | |
2473 | * true data, and if we misinform our callers that | |
2474 | * this ACK acks real data, we will erroneously exit | |
2475 | * connection startup slow start one packet too | |
2476 | * quickly. This is severely frowned upon behavior. | |
2477 | */ | |
2478 | if (!(scb->flags & TCPCB_FLAG_SYN)) { | |
2479 | acked |= FLAG_DATA_ACKED; | |
2480 | } else { | |
2481 | acked |= FLAG_SYN_ACKED; | |
2482 | tp->retrans_stamp = 0; | |
2483 | } | |
2484 | ||
5d424d5a JH |
2485 | /* MTU probing checks */ |
2486 | if (icsk->icsk_mtup.probe_size) { | |
0e7b1368 | 2487 | if (!after(tp->mtu_probe.probe_seq_end, TCP_SKB_CB(skb)->end_seq)) { |
5d424d5a JH |
2488 | tcp_mtup_probe_success(sk, skb); |
2489 | } | |
2490 | } | |
2491 | ||
1da177e4 LT |
2492 | if (sacked) { |
2493 | if (sacked & TCPCB_RETRANS) { | |
2de979bd | 2494 | if (sacked & TCPCB_SACKED_RETRANS) |
1da177e4 LT |
2495 | tp->retrans_out -= tcp_skb_pcount(skb); |
2496 | acked |= FLAG_RETRANS_DATA_ACKED; | |
2497 | seq_rtt = -1; | |
2d2abbab | 2498 | } else if (seq_rtt < 0) { |
1da177e4 | 2499 | seq_rtt = now - scb->when; |
164891aa | 2500 | last_ackt = skb->tstamp; |
a61bbcf2 | 2501 | } |
1da177e4 LT |
2502 | if (sacked & TCPCB_SACKED_ACKED) |
2503 | tp->sacked_out -= tcp_skb_pcount(skb); | |
2504 | if (sacked & TCPCB_LOST) | |
2505 | tp->lost_out -= tcp_skb_pcount(skb); | |
2506 | if (sacked & TCPCB_URG) { | |
2507 | if (tp->urg_mode && | |
2508 | !before(scb->end_seq, tp->snd_up)) | |
2509 | tp->urg_mode = 0; | |
2510 | } | |
2d2abbab | 2511 | } else if (seq_rtt < 0) { |
1da177e4 | 2512 | seq_rtt = now - scb->when; |
164891aa | 2513 | last_ackt = skb->tstamp; |
2d2abbab | 2514 | } |
1da177e4 LT |
2515 | tcp_dec_pcount_approx(&tp->fackets_out, skb); |
2516 | tcp_packets_out_dec(tp, skb); | |
fe067e8a | 2517 | tcp_unlink_write_queue(skb, sk); |
1da177e4 | 2518 | sk_stream_free_skb(sk, skb); |
6a438bbe | 2519 | clear_all_retrans_hints(tp); |
1da177e4 LT |
2520 | } |
2521 | ||
2522 | if (acked&FLAG_ACKED) { | |
6418204f | 2523 | u32 pkts_acked = prior_packets - tp->packets_out; |
164891aa SH |
2524 | const struct tcp_congestion_ops *ca_ops |
2525 | = inet_csk(sk)->icsk_ca_ops; | |
2526 | ||
2d2abbab | 2527 | tcp_ack_update_rtt(sk, acked, seq_rtt); |
9e412ba7 | 2528 | tcp_ack_packets_out(sk); |
317a76f9 | 2529 | |
30cfd0ba SH |
2530 | if (ca_ops->pkts_acked) { |
2531 | s32 rtt_us = -1; | |
2532 | ||
2533 | /* Is the ACK triggering packet unambiguous? */ | |
2534 | if (!(acked & FLAG_RETRANS_DATA_ACKED)) { | |
2535 | /* High resolution needed and available? */ | |
2536 | if (ca_ops->flags & TCP_CONG_RTT_STAMP && | |
2537 | !ktime_equal(last_ackt, | |
2538 | net_invalid_timestamp())) | |
2539 | rtt_us = ktime_us_delta(ktime_get_real(), | |
2540 | last_ackt); | |
2541 | else if (seq_rtt > 0) | |
2542 | rtt_us = jiffies_to_usecs(seq_rtt); | |
2543 | } | |
b9ce204f | 2544 | |
30cfd0ba SH |
2545 | ca_ops->pkts_acked(sk, pkts_acked, rtt_us); |
2546 | } | |
1da177e4 LT |
2547 | } |
2548 | ||
2549 | #if FASTRETRANS_DEBUG > 0 | |
2550 | BUG_TRAP((int)tp->sacked_out >= 0); | |
2551 | BUG_TRAP((int)tp->lost_out >= 0); | |
2552 | BUG_TRAP((int)tp->retrans_out >= 0); | |
2553 | if (!tp->packets_out && tp->rx_opt.sack_ok) { | |
6687e988 | 2554 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2555 | if (tp->lost_out) { |
2556 | printk(KERN_DEBUG "Leak l=%u %d\n", | |
6687e988 | 2557 | tp->lost_out, icsk->icsk_ca_state); |
1da177e4 LT |
2558 | tp->lost_out = 0; |
2559 | } | |
2560 | if (tp->sacked_out) { | |
2561 | printk(KERN_DEBUG "Leak s=%u %d\n", | |
6687e988 | 2562 | tp->sacked_out, icsk->icsk_ca_state); |
1da177e4 LT |
2563 | tp->sacked_out = 0; |
2564 | } | |
2565 | if (tp->retrans_out) { | |
2566 | printk(KERN_DEBUG "Leak r=%u %d\n", | |
6687e988 | 2567 | tp->retrans_out, icsk->icsk_ca_state); |
1da177e4 LT |
2568 | tp->retrans_out = 0; |
2569 | } | |
2570 | } | |
2571 | #endif | |
2572 | *seq_rtt_p = seq_rtt; | |
2573 | return acked; | |
2574 | } | |
2575 | ||
2576 | static void tcp_ack_probe(struct sock *sk) | |
2577 | { | |
463c84b9 ACM |
2578 | const struct tcp_sock *tp = tcp_sk(sk); |
2579 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1da177e4 LT |
2580 | |
2581 | /* Was it a usable window open? */ | |
2582 | ||
fe067e8a | 2583 | if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, |
1da177e4 | 2584 | tp->snd_una + tp->snd_wnd)) { |
463c84b9 ACM |
2585 | icsk->icsk_backoff = 0; |
2586 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); | |
1da177e4 LT |
2587 | /* Socket must be waked up by subsequent tcp_data_snd_check(). |
2588 | * This function is not for random using! | |
2589 | */ | |
2590 | } else { | |
463c84b9 | 2591 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
3f421baa ACM |
2592 | min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), |
2593 | TCP_RTO_MAX); | |
1da177e4 LT |
2594 | } |
2595 | } | |
2596 | ||
6687e988 | 2597 | static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag) |
1da177e4 LT |
2598 | { |
2599 | return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || | |
6687e988 | 2600 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open); |
1da177e4 LT |
2601 | } |
2602 | ||
6687e988 | 2603 | static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag) |
1da177e4 | 2604 | { |
6687e988 | 2605 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 2606 | return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && |
6687e988 | 2607 | !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR)); |
1da177e4 LT |
2608 | } |
2609 | ||
2610 | /* Check that window update is acceptable. | |
2611 | * The function assumes that snd_una<=ack<=snd_next. | |
2612 | */ | |
463c84b9 ACM |
2613 | static inline int tcp_may_update_window(const struct tcp_sock *tp, const u32 ack, |
2614 | const u32 ack_seq, const u32 nwin) | |
1da177e4 LT |
2615 | { |
2616 | return (after(ack, tp->snd_una) || | |
2617 | after(ack_seq, tp->snd_wl1) || | |
2618 | (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd)); | |
2619 | } | |
2620 | ||
2621 | /* Update our send window. | |
2622 | * | |
2623 | * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 | |
2624 | * and in FreeBSD. NetBSD's one is even worse.) is wrong. | |
2625 | */ | |
9e412ba7 IJ |
2626 | static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack, |
2627 | u32 ack_seq) | |
1da177e4 | 2628 | { |
9e412ba7 | 2629 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 2630 | int flag = 0; |
aa8223c7 | 2631 | u32 nwin = ntohs(tcp_hdr(skb)->window); |
1da177e4 | 2632 | |
aa8223c7 | 2633 | if (likely(!tcp_hdr(skb)->syn)) |
1da177e4 LT |
2634 | nwin <<= tp->rx_opt.snd_wscale; |
2635 | ||
2636 | if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { | |
2637 | flag |= FLAG_WIN_UPDATE; | |
2638 | tcp_update_wl(tp, ack, ack_seq); | |
2639 | ||
2640 | if (tp->snd_wnd != nwin) { | |
2641 | tp->snd_wnd = nwin; | |
2642 | ||
2643 | /* Note, it is the only place, where | |
2644 | * fast path is recovered for sending TCP. | |
2645 | */ | |
2ad41065 | 2646 | tp->pred_flags = 0; |
9e412ba7 | 2647 | tcp_fast_path_check(sk); |
1da177e4 LT |
2648 | |
2649 | if (nwin > tp->max_window) { | |
2650 | tp->max_window = nwin; | |
d83d8461 | 2651 | tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); |
1da177e4 LT |
2652 | } |
2653 | } | |
2654 | } | |
2655 | ||
2656 | tp->snd_una = ack; | |
2657 | ||
2658 | return flag; | |
2659 | } | |
2660 | ||
9ead9a1d IJ |
2661 | /* A very conservative spurious RTO response algorithm: reduce cwnd and |
2662 | * continue in congestion avoidance. | |
2663 | */ | |
2664 | static void tcp_conservative_spur_to_response(struct tcp_sock *tp) | |
2665 | { | |
2666 | tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); | |
aa8b6a7a | 2667 | tp->snd_cwnd_cnt = 0; |
46323655 | 2668 | TCP_ECN_queue_cwr(tp); |
9ead9a1d IJ |
2669 | tcp_moderate_cwnd(tp); |
2670 | } | |
2671 | ||
3cfe3baa IJ |
2672 | /* A conservative spurious RTO response algorithm: reduce cwnd using |
2673 | * rate halving and continue in congestion avoidance. | |
2674 | */ | |
2675 | static void tcp_ratehalving_spur_to_response(struct sock *sk) | |
2676 | { | |
3cfe3baa | 2677 | tcp_enter_cwr(sk, 0); |
3cfe3baa IJ |
2678 | } |
2679 | ||
e317f6f6 | 2680 | static void tcp_undo_spur_to_response(struct sock *sk, int flag) |
3cfe3baa | 2681 | { |
e317f6f6 IJ |
2682 | if (flag&FLAG_ECE) |
2683 | tcp_ratehalving_spur_to_response(sk); | |
2684 | else | |
2685 | tcp_undo_cwr(sk, 1); | |
3cfe3baa IJ |
2686 | } |
2687 | ||
30935cf4 IJ |
2688 | /* F-RTO spurious RTO detection algorithm (RFC4138) |
2689 | * | |
6408d206 IJ |
2690 | * F-RTO affects during two new ACKs following RTO (well, almost, see inline |
2691 | * comments). State (ACK number) is kept in frto_counter. When ACK advances | |
2692 | * window (but not to or beyond highest sequence sent before RTO): | |
30935cf4 IJ |
2693 | * On First ACK, send two new segments out. |
2694 | * On Second ACK, RTO was likely spurious. Do spurious response (response | |
2695 | * algorithm is not part of the F-RTO detection algorithm | |
2696 | * given in RFC4138 but can be selected separately). | |
2697 | * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss | |
d551e454 IJ |
2698 | * and TCP falls back to conventional RTO recovery. F-RTO allows overriding |
2699 | * of Nagle, this is done using frto_counter states 2 and 3, when a new data | |
2700 | * segment of any size sent during F-RTO, state 2 is upgraded to 3. | |
30935cf4 IJ |
2701 | * |
2702 | * Rationale: if the RTO was spurious, new ACKs should arrive from the | |
2703 | * original window even after we transmit two new data segments. | |
2704 | * | |
4dc2665e IJ |
2705 | * SACK version: |
2706 | * on first step, wait until first cumulative ACK arrives, then move to | |
2707 | * the second step. In second step, the next ACK decides. | |
2708 | * | |
30935cf4 IJ |
2709 | * F-RTO is implemented (mainly) in four functions: |
2710 | * - tcp_use_frto() is used to determine if TCP is can use F-RTO | |
2711 | * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is | |
2712 | * called when tcp_use_frto() showed green light | |
2713 | * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm | |
2714 | * - tcp_enter_frto_loss() is called if there is not enough evidence | |
2715 | * to prove that the RTO is indeed spurious. It transfers the control | |
2716 | * from F-RTO to the conventional RTO recovery | |
2717 | */ | |
2e605294 | 2718 | static int tcp_process_frto(struct sock *sk, int flag) |
1da177e4 LT |
2719 | { |
2720 | struct tcp_sock *tp = tcp_sk(sk); | |
e905a9ed | 2721 | |
1da177e4 | 2722 | tcp_sync_left_out(tp); |
e905a9ed | 2723 | |
7487c48c IJ |
2724 | /* Duplicate the behavior from Loss state (fastretrans_alert) */ |
2725 | if (flag&FLAG_DATA_ACKED) | |
2726 | inet_csk(sk)->icsk_retransmits = 0; | |
2727 | ||
95c4922b | 2728 | if (!before(tp->snd_una, tp->frto_highmark)) { |
d551e454 | 2729 | tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag); |
7c9a4a5b | 2730 | return 1; |
95c4922b IJ |
2731 | } |
2732 | ||
4dc2665e IJ |
2733 | if (!IsSackFrto() || IsReno(tp)) { |
2734 | /* RFC4138 shortcoming in step 2; should also have case c): | |
2735 | * ACK isn't duplicate nor advances window, e.g., opposite dir | |
2736 | * data, winupdate | |
2737 | */ | |
2e605294 | 2738 | if (!(flag&FLAG_ANY_PROGRESS) && (flag&FLAG_NOT_DUP)) |
4dc2665e IJ |
2739 | return 1; |
2740 | ||
2741 | if (!(flag&FLAG_DATA_ACKED)) { | |
2742 | tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3), | |
2743 | flag); | |
2744 | return 1; | |
2745 | } | |
2746 | } else { | |
2747 | if (!(flag&FLAG_DATA_ACKED) && (tp->frto_counter == 1)) { | |
2748 | /* Prevent sending of new data. */ | |
2749 | tp->snd_cwnd = min(tp->snd_cwnd, | |
2750 | tcp_packets_in_flight(tp)); | |
2751 | return 1; | |
2752 | } | |
6408d206 | 2753 | |
d551e454 | 2754 | if ((tp->frto_counter >= 2) && |
4dc2665e IJ |
2755 | (!(flag&FLAG_FORWARD_PROGRESS) || |
2756 | ((flag&FLAG_DATA_SACKED) && !(flag&FLAG_ONLY_ORIG_SACKED)))) { | |
2757 | /* RFC4138 shortcoming (see comment above) */ | |
2758 | if (!(flag&FLAG_FORWARD_PROGRESS) && (flag&FLAG_NOT_DUP)) | |
2759 | return 1; | |
2760 | ||
2761 | tcp_enter_frto_loss(sk, 3, flag); | |
2762 | return 1; | |
2763 | } | |
1da177e4 LT |
2764 | } |
2765 | ||
2766 | if (tp->frto_counter == 1) { | |
575ee714 IJ |
2767 | /* Sending of the next skb must be allowed or no FRTO */ |
2768 | if (!tcp_send_head(sk) || | |
2769 | after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, | |
2770 | tp->snd_una + tp->snd_wnd)) { | |
d551e454 IJ |
2771 | tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), |
2772 | flag); | |
575ee714 IJ |
2773 | return 1; |
2774 | } | |
2775 | ||
1da177e4 | 2776 | tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; |
94d0ea77 | 2777 | tp->frto_counter = 2; |
7c9a4a5b | 2778 | return 1; |
d551e454 | 2779 | } else { |
3cfe3baa IJ |
2780 | switch (sysctl_tcp_frto_response) { |
2781 | case 2: | |
e317f6f6 | 2782 | tcp_undo_spur_to_response(sk, flag); |
3cfe3baa IJ |
2783 | break; |
2784 | case 1: | |
2785 | tcp_conservative_spur_to_response(tp); | |
2786 | break; | |
2787 | default: | |
2788 | tcp_ratehalving_spur_to_response(sk); | |
2789 | break; | |
3ff50b79 | 2790 | } |
94d0ea77 | 2791 | tp->frto_counter = 0; |
1da177e4 | 2792 | } |
7c9a4a5b | 2793 | return 0; |
1da177e4 LT |
2794 | } |
2795 | ||
1da177e4 LT |
2796 | /* This routine deals with incoming acks, but not outgoing ones. */ |
2797 | static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag) | |
2798 | { | |
6687e988 | 2799 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2800 | struct tcp_sock *tp = tcp_sk(sk); |
2801 | u32 prior_snd_una = tp->snd_una; | |
2802 | u32 ack_seq = TCP_SKB_CB(skb)->seq; | |
2803 | u32 ack = TCP_SKB_CB(skb)->ack_seq; | |
2804 | u32 prior_in_flight; | |
2805 | s32 seq_rtt; | |
2806 | int prior_packets; | |
7c9a4a5b | 2807 | int frto_cwnd = 0; |
1da177e4 LT |
2808 | |
2809 | /* If the ack is newer than sent or older than previous acks | |
2810 | * then we can probably ignore it. | |
2811 | */ | |
2812 | if (after(ack, tp->snd_nxt)) | |
2813 | goto uninteresting_ack; | |
2814 | ||
2815 | if (before(ack, prior_snd_una)) | |
2816 | goto old_ack; | |
2817 | ||
2e605294 IJ |
2818 | if (after(ack, prior_snd_una)) |
2819 | flag |= FLAG_SND_UNA_ADVANCED; | |
2820 | ||
3fdf3f0c DO |
2821 | if (sysctl_tcp_abc) { |
2822 | if (icsk->icsk_ca_state < TCP_CA_CWR) | |
2823 | tp->bytes_acked += ack - prior_snd_una; | |
2824 | else if (icsk->icsk_ca_state == TCP_CA_Loss) | |
2825 | /* we assume just one segment left network */ | |
2826 | tp->bytes_acked += min(ack - prior_snd_una, tp->mss_cache); | |
2827 | } | |
9772efb9 | 2828 | |
1da177e4 LT |
2829 | if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) { |
2830 | /* Window is constant, pure forward advance. | |
2831 | * No more checks are required. | |
2832 | * Note, we use the fact that SND.UNA>=SND.WL2. | |
2833 | */ | |
2834 | tcp_update_wl(tp, ack, ack_seq); | |
2835 | tp->snd_una = ack; | |
1da177e4 LT |
2836 | flag |= FLAG_WIN_UPDATE; |
2837 | ||
6687e988 | 2838 | tcp_ca_event(sk, CA_EVENT_FAST_ACK); |
317a76f9 | 2839 | |
1da177e4 LT |
2840 | NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS); |
2841 | } else { | |
2842 | if (ack_seq != TCP_SKB_CB(skb)->end_seq) | |
2843 | flag |= FLAG_DATA; | |
2844 | else | |
2845 | NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS); | |
2846 | ||
9e412ba7 | 2847 | flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); |
1da177e4 LT |
2848 | |
2849 | if (TCP_SKB_CB(skb)->sacked) | |
2850 | flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); | |
2851 | ||
aa8223c7 | 2852 | if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb))) |
1da177e4 LT |
2853 | flag |= FLAG_ECE; |
2854 | ||
6687e988 | 2855 | tcp_ca_event(sk, CA_EVENT_SLOW_ACK); |
1da177e4 LT |
2856 | } |
2857 | ||
2858 | /* We passed data and got it acked, remove any soft error | |
2859 | * log. Something worked... | |
2860 | */ | |
2861 | sk->sk_err_soft = 0; | |
2862 | tp->rcv_tstamp = tcp_time_stamp; | |
2863 | prior_packets = tp->packets_out; | |
2864 | if (!prior_packets) | |
2865 | goto no_queue; | |
2866 | ||
2867 | prior_in_flight = tcp_packets_in_flight(tp); | |
2868 | ||
2869 | /* See if we can take anything off of the retransmit queue. */ | |
2d2abbab | 2870 | flag |= tcp_clean_rtx_queue(sk, &seq_rtt); |
1da177e4 LT |
2871 | |
2872 | if (tp->frto_counter) | |
2e605294 | 2873 | frto_cwnd = tcp_process_frto(sk, flag); |
1da177e4 | 2874 | |
6687e988 | 2875 | if (tcp_ack_is_dubious(sk, flag)) { |
caa20d9a | 2876 | /* Advance CWND, if state allows this. */ |
7c9a4a5b IJ |
2877 | if ((flag & FLAG_DATA_ACKED) && !frto_cwnd && |
2878 | tcp_may_raise_cwnd(sk, flag)) | |
16751347 | 2879 | tcp_cong_avoid(sk, ack, prior_in_flight, 0); |
2e605294 | 2880 | tcp_fastretrans_alert(sk, prior_packets, flag); |
1da177e4 | 2881 | } else { |
7c9a4a5b | 2882 | if ((flag & FLAG_DATA_ACKED) && !frto_cwnd) |
16751347 | 2883 | tcp_cong_avoid(sk, ack, prior_in_flight, 1); |
1da177e4 LT |
2884 | } |
2885 | ||
2886 | if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP)) | |
2887 | dst_confirm(sk->sk_dst_cache); | |
2888 | ||
2889 | return 1; | |
2890 | ||
2891 | no_queue: | |
6687e988 | 2892 | icsk->icsk_probes_out = 0; |
1da177e4 LT |
2893 | |
2894 | /* If this ack opens up a zero window, clear backoff. It was | |
2895 | * being used to time the probes, and is probably far higher than | |
2896 | * it needs to be for normal retransmission. | |
2897 | */ | |
fe067e8a | 2898 | if (tcp_send_head(sk)) |
1da177e4 LT |
2899 | tcp_ack_probe(sk); |
2900 | return 1; | |
2901 | ||
2902 | old_ack: | |
2903 | if (TCP_SKB_CB(skb)->sacked) | |
2904 | tcp_sacktag_write_queue(sk, skb, prior_snd_una); | |
2905 | ||
2906 | uninteresting_ack: | |
2907 | SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt); | |
2908 | return 0; | |
2909 | } | |
2910 | ||
2911 | ||
2912 | /* Look for tcp options. Normally only called on SYN and SYNACK packets. | |
2913 | * But, this can also be called on packets in the established flow when | |
2914 | * the fast version below fails. | |
2915 | */ | |
2916 | void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab) | |
2917 | { | |
2918 | unsigned char *ptr; | |
aa8223c7 | 2919 | struct tcphdr *th = tcp_hdr(skb); |
1da177e4 LT |
2920 | int length=(th->doff*4)-sizeof(struct tcphdr); |
2921 | ||
2922 | ptr = (unsigned char *)(th + 1); | |
2923 | opt_rx->saw_tstamp = 0; | |
2924 | ||
2de979bd | 2925 | while (length > 0) { |
e905a9ed | 2926 | int opcode=*ptr++; |
1da177e4 LT |
2927 | int opsize; |
2928 | ||
2929 | switch (opcode) { | |
2930 | case TCPOPT_EOL: | |
2931 | return; | |
2932 | case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ | |
2933 | length--; | |
2934 | continue; | |
2935 | default: | |
2936 | opsize=*ptr++; | |
2937 | if (opsize < 2) /* "silly options" */ | |
2938 | return; | |
2939 | if (opsize > length) | |
2940 | return; /* don't parse partial options */ | |
2de979bd | 2941 | switch (opcode) { |
1da177e4 | 2942 | case TCPOPT_MSS: |
2de979bd | 2943 | if (opsize==TCPOLEN_MSS && th->syn && !estab) { |
4f3608b7 | 2944 | u16 in_mss = ntohs(get_unaligned((__be16 *)ptr)); |
1da177e4 LT |
2945 | if (in_mss) { |
2946 | if (opt_rx->user_mss && opt_rx->user_mss < in_mss) | |
2947 | in_mss = opt_rx->user_mss; | |
2948 | opt_rx->mss_clamp = in_mss; | |
2949 | } | |
2950 | } | |
2951 | break; | |
2952 | case TCPOPT_WINDOW: | |
2de979bd | 2953 | if (opsize==TCPOLEN_WINDOW && th->syn && !estab) |
1da177e4 LT |
2954 | if (sysctl_tcp_window_scaling) { |
2955 | __u8 snd_wscale = *(__u8 *) ptr; | |
2956 | opt_rx->wscale_ok = 1; | |
2957 | if (snd_wscale > 14) { | |
2de979bd | 2958 | if (net_ratelimit()) |
1da177e4 LT |
2959 | printk(KERN_INFO "tcp_parse_options: Illegal window " |
2960 | "scaling value %d >14 received.\n", | |
2961 | snd_wscale); | |
2962 | snd_wscale = 14; | |
2963 | } | |
2964 | opt_rx->snd_wscale = snd_wscale; | |
2965 | } | |
2966 | break; | |
2967 | case TCPOPT_TIMESTAMP: | |
2de979bd | 2968 | if (opsize==TCPOLEN_TIMESTAMP) { |
1da177e4 LT |
2969 | if ((estab && opt_rx->tstamp_ok) || |
2970 | (!estab && sysctl_tcp_timestamps)) { | |
2971 | opt_rx->saw_tstamp = 1; | |
4f3608b7 AV |
2972 | opt_rx->rcv_tsval = ntohl(get_unaligned((__be32 *)ptr)); |
2973 | opt_rx->rcv_tsecr = ntohl(get_unaligned((__be32 *)(ptr+4))); | |
1da177e4 LT |
2974 | } |
2975 | } | |
2976 | break; | |
2977 | case TCPOPT_SACK_PERM: | |
2de979bd | 2978 | if (opsize==TCPOLEN_SACK_PERM && th->syn && !estab) { |
1da177e4 LT |
2979 | if (sysctl_tcp_sack) { |
2980 | opt_rx->sack_ok = 1; | |
2981 | tcp_sack_reset(opt_rx); | |
2982 | } | |
2983 | } | |
2984 | break; | |
2985 | ||
2986 | case TCPOPT_SACK: | |
2de979bd | 2987 | if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && |
1da177e4 LT |
2988 | !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && |
2989 | opt_rx->sack_ok) { | |
2990 | TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; | |
2991 | } | |
d7ea5b91 | 2992 | break; |
cfb6eeb4 YH |
2993 | #ifdef CONFIG_TCP_MD5SIG |
2994 | case TCPOPT_MD5SIG: | |
2995 | /* | |
2996 | * The MD5 Hash has already been | |
2997 | * checked (see tcp_v{4,6}_do_rcv()). | |
2998 | */ | |
2999 | break; | |
3000 | #endif | |
3ff50b79 SH |
3001 | } |
3002 | ||
e905a9ed YH |
3003 | ptr+=opsize-2; |
3004 | length-=opsize; | |
3ff50b79 | 3005 | } |
1da177e4 LT |
3006 | } |
3007 | } | |
3008 | ||
3009 | /* Fast parse options. This hopes to only see timestamps. | |
3010 | * If it is wrong it falls back on tcp_parse_options(). | |
3011 | */ | |
40efc6fa SH |
3012 | static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, |
3013 | struct tcp_sock *tp) | |
1da177e4 LT |
3014 | { |
3015 | if (th->doff == sizeof(struct tcphdr)>>2) { | |
3016 | tp->rx_opt.saw_tstamp = 0; | |
3017 | return 0; | |
3018 | } else if (tp->rx_opt.tstamp_ok && | |
3019 | th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) { | |
4f3608b7 AV |
3020 | __be32 *ptr = (__be32 *)(th + 1); |
3021 | if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | |
1da177e4 LT |
3022 | | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { |
3023 | tp->rx_opt.saw_tstamp = 1; | |
3024 | ++ptr; | |
3025 | tp->rx_opt.rcv_tsval = ntohl(*ptr); | |
3026 | ++ptr; | |
3027 | tp->rx_opt.rcv_tsecr = ntohl(*ptr); | |
3028 | return 1; | |
3029 | } | |
3030 | } | |
3031 | tcp_parse_options(skb, &tp->rx_opt, 1); | |
3032 | return 1; | |
3033 | } | |
3034 | ||
3035 | static inline void tcp_store_ts_recent(struct tcp_sock *tp) | |
3036 | { | |
3037 | tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; | |
9d729f72 | 3038 | tp->rx_opt.ts_recent_stamp = get_seconds(); |
1da177e4 LT |
3039 | } |
3040 | ||
3041 | static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) | |
3042 | { | |
3043 | if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { | |
3044 | /* PAWS bug workaround wrt. ACK frames, the PAWS discard | |
3045 | * extra check below makes sure this can only happen | |
3046 | * for pure ACK frames. -DaveM | |
3047 | * | |
3048 | * Not only, also it occurs for expired timestamps. | |
3049 | */ | |
3050 | ||
2de979bd | 3051 | if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 || |
9d729f72 | 3052 | get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS) |
1da177e4 LT |
3053 | tcp_store_ts_recent(tp); |
3054 | } | |
3055 | } | |
3056 | ||
3057 | /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM | |
3058 | * | |
3059 | * It is not fatal. If this ACK does _not_ change critical state (seqs, window) | |
3060 | * it can pass through stack. So, the following predicate verifies that | |
3061 | * this segment is not used for anything but congestion avoidance or | |
3062 | * fast retransmit. Moreover, we even are able to eliminate most of such | |
3063 | * second order effects, if we apply some small "replay" window (~RTO) | |
3064 | * to timestamp space. | |
3065 | * | |
3066 | * All these measures still do not guarantee that we reject wrapped ACKs | |
3067 | * on networks with high bandwidth, when sequence space is recycled fastly, | |
3068 | * but it guarantees that such events will be very rare and do not affect | |
3069 | * connection seriously. This doesn't look nice, but alas, PAWS is really | |
3070 | * buggy extension. | |
3071 | * | |
3072 | * [ Later note. Even worse! It is buggy for segments _with_ data. RFC | |
3073 | * states that events when retransmit arrives after original data are rare. | |
3074 | * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is | |
3075 | * the biggest problem on large power networks even with minor reordering. | |
3076 | * OK, let's give it small replay window. If peer clock is even 1hz, it is safe | |
3077 | * up to bandwidth of 18Gigabit/sec. 8) ] | |
3078 | */ | |
3079 | ||
463c84b9 | 3080 | static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 3081 | { |
463c84b9 | 3082 | struct tcp_sock *tp = tcp_sk(sk); |
aa8223c7 | 3083 | struct tcphdr *th = tcp_hdr(skb); |
1da177e4 LT |
3084 | u32 seq = TCP_SKB_CB(skb)->seq; |
3085 | u32 ack = TCP_SKB_CB(skb)->ack_seq; | |
3086 | ||
3087 | return (/* 1. Pure ACK with correct sequence number. */ | |
3088 | (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && | |
3089 | ||
3090 | /* 2. ... and duplicate ACK. */ | |
3091 | ack == tp->snd_una && | |
3092 | ||
3093 | /* 3. ... and does not update window. */ | |
3094 | !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && | |
3095 | ||
3096 | /* 4. ... and sits in replay window. */ | |
463c84b9 | 3097 | (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); |
1da177e4 LT |
3098 | } |
3099 | ||
463c84b9 | 3100 | static inline int tcp_paws_discard(const struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 3101 | { |
463c84b9 | 3102 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 3103 | return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW && |
9d729f72 | 3104 | get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS && |
463c84b9 | 3105 | !tcp_disordered_ack(sk, skb)); |
1da177e4 LT |
3106 | } |
3107 | ||
3108 | /* Check segment sequence number for validity. | |
3109 | * | |
3110 | * Segment controls are considered valid, if the segment | |
3111 | * fits to the window after truncation to the window. Acceptability | |
3112 | * of data (and SYN, FIN, of course) is checked separately. | |
3113 | * See tcp_data_queue(), for example. | |
3114 | * | |
3115 | * Also, controls (RST is main one) are accepted using RCV.WUP instead | |
3116 | * of RCV.NXT. Peer still did not advance his SND.UNA when we | |
3117 | * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. | |
3118 | * (borrowed from freebsd) | |
3119 | */ | |
3120 | ||
3121 | static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq) | |
3122 | { | |
3123 | return !before(end_seq, tp->rcv_wup) && | |
3124 | !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); | |
3125 | } | |
3126 | ||
3127 | /* When we get a reset we do this. */ | |
3128 | static void tcp_reset(struct sock *sk) | |
3129 | { | |
3130 | /* We want the right error as BSD sees it (and indeed as we do). */ | |
3131 | switch (sk->sk_state) { | |
3132 | case TCP_SYN_SENT: | |
3133 | sk->sk_err = ECONNREFUSED; | |
3134 | break; | |
3135 | case TCP_CLOSE_WAIT: | |
3136 | sk->sk_err = EPIPE; | |
3137 | break; | |
3138 | case TCP_CLOSE: | |
3139 | return; | |
3140 | default: | |
3141 | sk->sk_err = ECONNRESET; | |
3142 | } | |
3143 | ||
3144 | if (!sock_flag(sk, SOCK_DEAD)) | |
3145 | sk->sk_error_report(sk); | |
3146 | ||
3147 | tcp_done(sk); | |
3148 | } | |
3149 | ||
3150 | /* | |
3151 | * Process the FIN bit. This now behaves as it is supposed to work | |
3152 | * and the FIN takes effect when it is validly part of sequence | |
3153 | * space. Not before when we get holes. | |
3154 | * | |
3155 | * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT | |
3156 | * (and thence onto LAST-ACK and finally, CLOSE, we never enter | |
3157 | * TIME-WAIT) | |
3158 | * | |
3159 | * If we are in FINWAIT-1, a received FIN indicates simultaneous | |
3160 | * close and we go into CLOSING (and later onto TIME-WAIT) | |
3161 | * | |
3162 | * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. | |
3163 | */ | |
3164 | static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th) | |
3165 | { | |
3166 | struct tcp_sock *tp = tcp_sk(sk); | |
3167 | ||
463c84b9 | 3168 | inet_csk_schedule_ack(sk); |
1da177e4 LT |
3169 | |
3170 | sk->sk_shutdown |= RCV_SHUTDOWN; | |
3171 | sock_set_flag(sk, SOCK_DONE); | |
3172 | ||
3173 | switch (sk->sk_state) { | |
3174 | case TCP_SYN_RECV: | |
3175 | case TCP_ESTABLISHED: | |
3176 | /* Move to CLOSE_WAIT */ | |
3177 | tcp_set_state(sk, TCP_CLOSE_WAIT); | |
463c84b9 | 3178 | inet_csk(sk)->icsk_ack.pingpong = 1; |
1da177e4 LT |
3179 | break; |
3180 | ||
3181 | case TCP_CLOSE_WAIT: | |
3182 | case TCP_CLOSING: | |
3183 | /* Received a retransmission of the FIN, do | |
3184 | * nothing. | |
3185 | */ | |
3186 | break; | |
3187 | case TCP_LAST_ACK: | |
3188 | /* RFC793: Remain in the LAST-ACK state. */ | |
3189 | break; | |
3190 | ||
3191 | case TCP_FIN_WAIT1: | |
3192 | /* This case occurs when a simultaneous close | |
3193 | * happens, we must ack the received FIN and | |
3194 | * enter the CLOSING state. | |
3195 | */ | |
3196 | tcp_send_ack(sk); | |
3197 | tcp_set_state(sk, TCP_CLOSING); | |
3198 | break; | |
3199 | case TCP_FIN_WAIT2: | |
3200 | /* Received a FIN -- send ACK and enter TIME_WAIT. */ | |
3201 | tcp_send_ack(sk); | |
3202 | tcp_time_wait(sk, TCP_TIME_WAIT, 0); | |
3203 | break; | |
3204 | default: | |
3205 | /* Only TCP_LISTEN and TCP_CLOSE are left, in these | |
3206 | * cases we should never reach this piece of code. | |
3207 | */ | |
3208 | printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n", | |
3209 | __FUNCTION__, sk->sk_state); | |
3210 | break; | |
3ff50b79 | 3211 | } |
1da177e4 LT |
3212 | |
3213 | /* It _is_ possible, that we have something out-of-order _after_ FIN. | |
3214 | * Probably, we should reset in this case. For now drop them. | |
3215 | */ | |
3216 | __skb_queue_purge(&tp->out_of_order_queue); | |
3217 | if (tp->rx_opt.sack_ok) | |
3218 | tcp_sack_reset(&tp->rx_opt); | |
3219 | sk_stream_mem_reclaim(sk); | |
3220 | ||
3221 | if (!sock_flag(sk, SOCK_DEAD)) { | |
3222 | sk->sk_state_change(sk); | |
3223 | ||
3224 | /* Do not send POLL_HUP for half duplex close. */ | |
3225 | if (sk->sk_shutdown == SHUTDOWN_MASK || | |
3226 | sk->sk_state == TCP_CLOSE) | |
3227 | sk_wake_async(sk, 1, POLL_HUP); | |
3228 | else | |
3229 | sk_wake_async(sk, 1, POLL_IN); | |
3230 | } | |
3231 | } | |
3232 | ||
40efc6fa | 3233 | static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq) |
1da177e4 LT |
3234 | { |
3235 | if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { | |
3236 | if (before(seq, sp->start_seq)) | |
3237 | sp->start_seq = seq; | |
3238 | if (after(end_seq, sp->end_seq)) | |
3239 | sp->end_seq = end_seq; | |
3240 | return 1; | |
3241 | } | |
3242 | return 0; | |
3243 | } | |
3244 | ||
40efc6fa | 3245 | static void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq) |
1da177e4 LT |
3246 | { |
3247 | if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) { | |
3248 | if (before(seq, tp->rcv_nxt)) | |
3249 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT); | |
3250 | else | |
3251 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT); | |
3252 | ||
3253 | tp->rx_opt.dsack = 1; | |
3254 | tp->duplicate_sack[0].start_seq = seq; | |
3255 | tp->duplicate_sack[0].end_seq = end_seq; | |
3256 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok); | |
3257 | } | |
3258 | } | |
3259 | ||
40efc6fa | 3260 | static void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq) |
1da177e4 LT |
3261 | { |
3262 | if (!tp->rx_opt.dsack) | |
3263 | tcp_dsack_set(tp, seq, end_seq); | |
3264 | else | |
3265 | tcp_sack_extend(tp->duplicate_sack, seq, end_seq); | |
3266 | } | |
3267 | ||
3268 | static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb) | |
3269 | { | |
3270 | struct tcp_sock *tp = tcp_sk(sk); | |
3271 | ||
3272 | if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && | |
3273 | before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
3274 | NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); | |
463c84b9 | 3275 | tcp_enter_quickack_mode(sk); |
1da177e4 LT |
3276 | |
3277 | if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) { | |
3278 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
3279 | ||
3280 | if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) | |
3281 | end_seq = tp->rcv_nxt; | |
3282 | tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq); | |
3283 | } | |
3284 | } | |
3285 | ||
3286 | tcp_send_ack(sk); | |
3287 | } | |
3288 | ||
3289 | /* These routines update the SACK block as out-of-order packets arrive or | |
3290 | * in-order packets close up the sequence space. | |
3291 | */ | |
3292 | static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) | |
3293 | { | |
3294 | int this_sack; | |
3295 | struct tcp_sack_block *sp = &tp->selective_acks[0]; | |
3296 | struct tcp_sack_block *swalk = sp+1; | |
3297 | ||
3298 | /* See if the recent change to the first SACK eats into | |
3299 | * or hits the sequence space of other SACK blocks, if so coalesce. | |
3300 | */ | |
3301 | for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) { | |
3302 | if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { | |
3303 | int i; | |
3304 | ||
3305 | /* Zap SWALK, by moving every further SACK up by one slot. | |
3306 | * Decrease num_sacks. | |
3307 | */ | |
3308 | tp->rx_opt.num_sacks--; | |
3309 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); | |
2de979bd | 3310 | for (i=this_sack; i < tp->rx_opt.num_sacks; i++) |
1da177e4 LT |
3311 | sp[i] = sp[i+1]; |
3312 | continue; | |
3313 | } | |
3314 | this_sack++, swalk++; | |
3315 | } | |
3316 | } | |
3317 | ||
40efc6fa | 3318 | static inline void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2) |
1da177e4 LT |
3319 | { |
3320 | __u32 tmp; | |
3321 | ||
3322 | tmp = sack1->start_seq; | |
3323 | sack1->start_seq = sack2->start_seq; | |
3324 | sack2->start_seq = tmp; | |
3325 | ||
3326 | tmp = sack1->end_seq; | |
3327 | sack1->end_seq = sack2->end_seq; | |
3328 | sack2->end_seq = tmp; | |
3329 | } | |
3330 | ||
3331 | static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) | |
3332 | { | |
3333 | struct tcp_sock *tp = tcp_sk(sk); | |
3334 | struct tcp_sack_block *sp = &tp->selective_acks[0]; | |
3335 | int cur_sacks = tp->rx_opt.num_sacks; | |
3336 | int this_sack; | |
3337 | ||
3338 | if (!cur_sacks) | |
3339 | goto new_sack; | |
3340 | ||
3341 | for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) { | |
3342 | if (tcp_sack_extend(sp, seq, end_seq)) { | |
3343 | /* Rotate this_sack to the first one. */ | |
3344 | for (; this_sack>0; this_sack--, sp--) | |
3345 | tcp_sack_swap(sp, sp-1); | |
3346 | if (cur_sacks > 1) | |
3347 | tcp_sack_maybe_coalesce(tp); | |
3348 | return; | |
3349 | } | |
3350 | } | |
3351 | ||
3352 | /* Could not find an adjacent existing SACK, build a new one, | |
3353 | * put it at the front, and shift everyone else down. We | |
3354 | * always know there is at least one SACK present already here. | |
3355 | * | |
3356 | * If the sack array is full, forget about the last one. | |
3357 | */ | |
3358 | if (this_sack >= 4) { | |
3359 | this_sack--; | |
3360 | tp->rx_opt.num_sacks--; | |
3361 | sp--; | |
3362 | } | |
2de979bd | 3363 | for (; this_sack > 0; this_sack--, sp--) |
1da177e4 LT |
3364 | *sp = *(sp-1); |
3365 | ||
3366 | new_sack: | |
3367 | /* Build the new head SACK, and we're done. */ | |
3368 | sp->start_seq = seq; | |
3369 | sp->end_seq = end_seq; | |
3370 | tp->rx_opt.num_sacks++; | |
3371 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); | |
3372 | } | |
3373 | ||
3374 | /* RCV.NXT advances, some SACKs should be eaten. */ | |
3375 | ||
3376 | static void tcp_sack_remove(struct tcp_sock *tp) | |
3377 | { | |
3378 | struct tcp_sack_block *sp = &tp->selective_acks[0]; | |
3379 | int num_sacks = tp->rx_opt.num_sacks; | |
3380 | int this_sack; | |
3381 | ||
3382 | /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ | |
b03efcfb | 3383 | if (skb_queue_empty(&tp->out_of_order_queue)) { |
1da177e4 LT |
3384 | tp->rx_opt.num_sacks = 0; |
3385 | tp->rx_opt.eff_sacks = tp->rx_opt.dsack; | |
3386 | return; | |
3387 | } | |
3388 | ||
2de979bd | 3389 | for (this_sack = 0; this_sack < num_sacks; ) { |
1da177e4 LT |
3390 | /* Check if the start of the sack is covered by RCV.NXT. */ |
3391 | if (!before(tp->rcv_nxt, sp->start_seq)) { | |
3392 | int i; | |
3393 | ||
3394 | /* RCV.NXT must cover all the block! */ | |
3395 | BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq)); | |
3396 | ||
3397 | /* Zap this SACK, by moving forward any other SACKS. */ | |
3398 | for (i=this_sack+1; i < num_sacks; i++) | |
3399 | tp->selective_acks[i-1] = tp->selective_acks[i]; | |
3400 | num_sacks--; | |
3401 | continue; | |
3402 | } | |
3403 | this_sack++; | |
3404 | sp++; | |
3405 | } | |
3406 | if (num_sacks != tp->rx_opt.num_sacks) { | |
3407 | tp->rx_opt.num_sacks = num_sacks; | |
3408 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); | |
3409 | } | |
3410 | } | |
3411 | ||
3412 | /* This one checks to see if we can put data from the | |
3413 | * out_of_order queue into the receive_queue. | |
3414 | */ | |
3415 | static void tcp_ofo_queue(struct sock *sk) | |
3416 | { | |
3417 | struct tcp_sock *tp = tcp_sk(sk); | |
3418 | __u32 dsack_high = tp->rcv_nxt; | |
3419 | struct sk_buff *skb; | |
3420 | ||
3421 | while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { | |
3422 | if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) | |
3423 | break; | |
3424 | ||
3425 | if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { | |
3426 | __u32 dsack = dsack_high; | |
3427 | if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) | |
3428 | dsack_high = TCP_SKB_CB(skb)->end_seq; | |
3429 | tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack); | |
3430 | } | |
3431 | ||
3432 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { | |
3433 | SOCK_DEBUG(sk, "ofo packet was already received \n"); | |
8728b834 | 3434 | __skb_unlink(skb, &tp->out_of_order_queue); |
1da177e4 LT |
3435 | __kfree_skb(skb); |
3436 | continue; | |
3437 | } | |
3438 | SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", | |
3439 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, | |
3440 | TCP_SKB_CB(skb)->end_seq); | |
3441 | ||
8728b834 | 3442 | __skb_unlink(skb, &tp->out_of_order_queue); |
1da177e4 LT |
3443 | __skb_queue_tail(&sk->sk_receive_queue, skb); |
3444 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
aa8223c7 ACM |
3445 | if (tcp_hdr(skb)->fin) |
3446 | tcp_fin(skb, sk, tcp_hdr(skb)); | |
1da177e4 LT |
3447 | } |
3448 | } | |
3449 | ||
3450 | static int tcp_prune_queue(struct sock *sk); | |
3451 | ||
3452 | static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) | |
3453 | { | |
aa8223c7 | 3454 | struct tcphdr *th = tcp_hdr(skb); |
1da177e4 LT |
3455 | struct tcp_sock *tp = tcp_sk(sk); |
3456 | int eaten = -1; | |
3457 | ||
3458 | if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) | |
3459 | goto drop; | |
3460 | ||
1da177e4 LT |
3461 | __skb_pull(skb, th->doff*4); |
3462 | ||
3463 | TCP_ECN_accept_cwr(tp, skb); | |
3464 | ||
3465 | if (tp->rx_opt.dsack) { | |
3466 | tp->rx_opt.dsack = 0; | |
3467 | tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks, | |
3468 | 4 - tp->rx_opt.tstamp_ok); | |
3469 | } | |
3470 | ||
3471 | /* Queue data for delivery to the user. | |
3472 | * Packets in sequence go to the receive queue. | |
3473 | * Out of sequence packets to the out_of_order_queue. | |
3474 | */ | |
3475 | if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { | |
3476 | if (tcp_receive_window(tp) == 0) | |
3477 | goto out_of_window; | |
3478 | ||
3479 | /* Ok. In sequence. In window. */ | |
3480 | if (tp->ucopy.task == current && | |
3481 | tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && | |
3482 | sock_owned_by_user(sk) && !tp->urg_data) { | |
3483 | int chunk = min_t(unsigned int, skb->len, | |
3484 | tp->ucopy.len); | |
3485 | ||
3486 | __set_current_state(TASK_RUNNING); | |
3487 | ||
3488 | local_bh_enable(); | |
3489 | if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { | |
3490 | tp->ucopy.len -= chunk; | |
3491 | tp->copied_seq += chunk; | |
3492 | eaten = (chunk == skb->len && !th->fin); | |
3493 | tcp_rcv_space_adjust(sk); | |
3494 | } | |
3495 | local_bh_disable(); | |
3496 | } | |
3497 | ||
3498 | if (eaten <= 0) { | |
3499 | queue_and_out: | |
3500 | if (eaten < 0 && | |
3501 | (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || | |
3502 | !sk_stream_rmem_schedule(sk, skb))) { | |
3503 | if (tcp_prune_queue(sk) < 0 || | |
3504 | !sk_stream_rmem_schedule(sk, skb)) | |
3505 | goto drop; | |
3506 | } | |
3507 | sk_stream_set_owner_r(skb, sk); | |
3508 | __skb_queue_tail(&sk->sk_receive_queue, skb); | |
3509 | } | |
3510 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
2de979bd | 3511 | if (skb->len) |
9e412ba7 | 3512 | tcp_event_data_recv(sk, skb); |
2de979bd | 3513 | if (th->fin) |
1da177e4 LT |
3514 | tcp_fin(skb, sk, th); |
3515 | ||
b03efcfb | 3516 | if (!skb_queue_empty(&tp->out_of_order_queue)) { |
1da177e4 LT |
3517 | tcp_ofo_queue(sk); |
3518 | ||
3519 | /* RFC2581. 4.2. SHOULD send immediate ACK, when | |
3520 | * gap in queue is filled. | |
3521 | */ | |
b03efcfb | 3522 | if (skb_queue_empty(&tp->out_of_order_queue)) |
463c84b9 | 3523 | inet_csk(sk)->icsk_ack.pingpong = 0; |
1da177e4 LT |
3524 | } |
3525 | ||
3526 | if (tp->rx_opt.num_sacks) | |
3527 | tcp_sack_remove(tp); | |
3528 | ||
9e412ba7 | 3529 | tcp_fast_path_check(sk); |
1da177e4 LT |
3530 | |
3531 | if (eaten > 0) | |
3532 | __kfree_skb(skb); | |
3533 | else if (!sock_flag(sk, SOCK_DEAD)) | |
3534 | sk->sk_data_ready(sk, 0); | |
3535 | return; | |
3536 | } | |
3537 | ||
3538 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { | |
3539 | /* A retransmit, 2nd most common case. Force an immediate ack. */ | |
3540 | NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); | |
3541 | tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); | |
3542 | ||
3543 | out_of_window: | |
463c84b9 ACM |
3544 | tcp_enter_quickack_mode(sk); |
3545 | inet_csk_schedule_ack(sk); | |
1da177e4 LT |
3546 | drop: |
3547 | __kfree_skb(skb); | |
3548 | return; | |
3549 | } | |
3550 | ||
3551 | /* Out of window. F.e. zero window probe. */ | |
3552 | if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) | |
3553 | goto out_of_window; | |
3554 | ||
463c84b9 | 3555 | tcp_enter_quickack_mode(sk); |
1da177e4 LT |
3556 | |
3557 | if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
3558 | /* Partial packet, seq < rcv_next < end_seq */ | |
3559 | SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", | |
3560 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, | |
3561 | TCP_SKB_CB(skb)->end_seq); | |
3562 | ||
3563 | tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); | |
e905a9ed | 3564 | |
1da177e4 LT |
3565 | /* If window is closed, drop tail of packet. But after |
3566 | * remembering D-SACK for its head made in previous line. | |
3567 | */ | |
3568 | if (!tcp_receive_window(tp)) | |
3569 | goto out_of_window; | |
3570 | goto queue_and_out; | |
3571 | } | |
3572 | ||
3573 | TCP_ECN_check_ce(tp, skb); | |
3574 | ||
3575 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || | |
3576 | !sk_stream_rmem_schedule(sk, skb)) { | |
3577 | if (tcp_prune_queue(sk) < 0 || | |
3578 | !sk_stream_rmem_schedule(sk, skb)) | |
3579 | goto drop; | |
3580 | } | |
3581 | ||
3582 | /* Disable header prediction. */ | |
3583 | tp->pred_flags = 0; | |
463c84b9 | 3584 | inet_csk_schedule_ack(sk); |
1da177e4 LT |
3585 | |
3586 | SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", | |
3587 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); | |
3588 | ||
3589 | sk_stream_set_owner_r(skb, sk); | |
3590 | ||
3591 | if (!skb_peek(&tp->out_of_order_queue)) { | |
3592 | /* Initial out of order segment, build 1 SACK. */ | |
3593 | if (tp->rx_opt.sack_ok) { | |
3594 | tp->rx_opt.num_sacks = 1; | |
3595 | tp->rx_opt.dsack = 0; | |
3596 | tp->rx_opt.eff_sacks = 1; | |
3597 | tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; | |
3598 | tp->selective_acks[0].end_seq = | |
3599 | TCP_SKB_CB(skb)->end_seq; | |
3600 | } | |
3601 | __skb_queue_head(&tp->out_of_order_queue,skb); | |
3602 | } else { | |
3603 | struct sk_buff *skb1 = tp->out_of_order_queue.prev; | |
3604 | u32 seq = TCP_SKB_CB(skb)->seq; | |
3605 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
3606 | ||
3607 | if (seq == TCP_SKB_CB(skb1)->end_seq) { | |
8728b834 | 3608 | __skb_append(skb1, skb, &tp->out_of_order_queue); |
1da177e4 LT |
3609 | |
3610 | if (!tp->rx_opt.num_sacks || | |
3611 | tp->selective_acks[0].end_seq != seq) | |
3612 | goto add_sack; | |
3613 | ||
3614 | /* Common case: data arrive in order after hole. */ | |
3615 | tp->selective_acks[0].end_seq = end_seq; | |
3616 | return; | |
3617 | } | |
3618 | ||
3619 | /* Find place to insert this segment. */ | |
3620 | do { | |
3621 | if (!after(TCP_SKB_CB(skb1)->seq, seq)) | |
3622 | break; | |
3623 | } while ((skb1 = skb1->prev) != | |
3624 | (struct sk_buff*)&tp->out_of_order_queue); | |
3625 | ||
3626 | /* Do skb overlap to previous one? */ | |
3627 | if (skb1 != (struct sk_buff*)&tp->out_of_order_queue && | |
3628 | before(seq, TCP_SKB_CB(skb1)->end_seq)) { | |
3629 | if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { | |
3630 | /* All the bits are present. Drop. */ | |
3631 | __kfree_skb(skb); | |
3632 | tcp_dsack_set(tp, seq, end_seq); | |
3633 | goto add_sack; | |
3634 | } | |
3635 | if (after(seq, TCP_SKB_CB(skb1)->seq)) { | |
3636 | /* Partial overlap. */ | |
3637 | tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq); | |
3638 | } else { | |
3639 | skb1 = skb1->prev; | |
3640 | } | |
3641 | } | |
3642 | __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue); | |
e905a9ed | 3643 | |
1da177e4 LT |
3644 | /* And clean segments covered by new one as whole. */ |
3645 | while ((skb1 = skb->next) != | |
3646 | (struct sk_buff*)&tp->out_of_order_queue && | |
3647 | after(end_seq, TCP_SKB_CB(skb1)->seq)) { | |
3648 | if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { | |
3649 | tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq); | |
3650 | break; | |
3651 | } | |
8728b834 | 3652 | __skb_unlink(skb1, &tp->out_of_order_queue); |
1da177e4 LT |
3653 | tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq); |
3654 | __kfree_skb(skb1); | |
3655 | } | |
3656 | ||
3657 | add_sack: | |
3658 | if (tp->rx_opt.sack_ok) | |
3659 | tcp_sack_new_ofo_skb(sk, seq, end_seq); | |
3660 | } | |
3661 | } | |
3662 | ||
3663 | /* Collapse contiguous sequence of skbs head..tail with | |
3664 | * sequence numbers start..end. | |
3665 | * Segments with FIN/SYN are not collapsed (only because this | |
3666 | * simplifies code) | |
3667 | */ | |
3668 | static void | |
8728b834 DM |
3669 | tcp_collapse(struct sock *sk, struct sk_buff_head *list, |
3670 | struct sk_buff *head, struct sk_buff *tail, | |
3671 | u32 start, u32 end) | |
1da177e4 LT |
3672 | { |
3673 | struct sk_buff *skb; | |
3674 | ||
caa20d9a | 3675 | /* First, check that queue is collapsible and find |
1da177e4 LT |
3676 | * the point where collapsing can be useful. */ |
3677 | for (skb = head; skb != tail; ) { | |
3678 | /* No new bits? It is possible on ofo queue. */ | |
3679 | if (!before(start, TCP_SKB_CB(skb)->end_seq)) { | |
3680 | struct sk_buff *next = skb->next; | |
8728b834 | 3681 | __skb_unlink(skb, list); |
1da177e4 LT |
3682 | __kfree_skb(skb); |
3683 | NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); | |
3684 | skb = next; | |
3685 | continue; | |
3686 | } | |
3687 | ||
3688 | /* The first skb to collapse is: | |
3689 | * - not SYN/FIN and | |
3690 | * - bloated or contains data before "start" or | |
3691 | * overlaps to the next one. | |
3692 | */ | |
aa8223c7 | 3693 | if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin && |
1da177e4 LT |
3694 | (tcp_win_from_space(skb->truesize) > skb->len || |
3695 | before(TCP_SKB_CB(skb)->seq, start) || | |
3696 | (skb->next != tail && | |
3697 | TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq))) | |
3698 | break; | |
3699 | ||
3700 | /* Decided to skip this, advance start seq. */ | |
3701 | start = TCP_SKB_CB(skb)->end_seq; | |
3702 | skb = skb->next; | |
3703 | } | |
aa8223c7 | 3704 | if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin) |
1da177e4 LT |
3705 | return; |
3706 | ||
3707 | while (before(start, end)) { | |
3708 | struct sk_buff *nskb; | |
3709 | int header = skb_headroom(skb); | |
3710 | int copy = SKB_MAX_ORDER(header, 0); | |
3711 | ||
3712 | /* Too big header? This can happen with IPv6. */ | |
3713 | if (copy < 0) | |
3714 | return; | |
3715 | if (end-start < copy) | |
3716 | copy = end-start; | |
3717 | nskb = alloc_skb(copy+header, GFP_ATOMIC); | |
3718 | if (!nskb) | |
3719 | return; | |
c51957da | 3720 | |
98e399f8 | 3721 | skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head); |
9c70220b ACM |
3722 | skb_set_network_header(nskb, (skb_network_header(skb) - |
3723 | skb->head)); | |
3724 | skb_set_transport_header(nskb, (skb_transport_header(skb) - | |
3725 | skb->head)); | |
1da177e4 LT |
3726 | skb_reserve(nskb, header); |
3727 | memcpy(nskb->head, skb->head, header); | |
1da177e4 LT |
3728 | memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); |
3729 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; | |
8728b834 | 3730 | __skb_insert(nskb, skb->prev, skb, list); |
1da177e4 LT |
3731 | sk_stream_set_owner_r(nskb, sk); |
3732 | ||
3733 | /* Copy data, releasing collapsed skbs. */ | |
3734 | while (copy > 0) { | |
3735 | int offset = start - TCP_SKB_CB(skb)->seq; | |
3736 | int size = TCP_SKB_CB(skb)->end_seq - start; | |
3737 | ||
09a62660 | 3738 | BUG_ON(offset < 0); |
1da177e4 LT |
3739 | if (size > 0) { |
3740 | size = min(copy, size); | |
3741 | if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) | |
3742 | BUG(); | |
3743 | TCP_SKB_CB(nskb)->end_seq += size; | |
3744 | copy -= size; | |
3745 | start += size; | |
3746 | } | |
3747 | if (!before(start, TCP_SKB_CB(skb)->end_seq)) { | |
3748 | struct sk_buff *next = skb->next; | |
8728b834 | 3749 | __skb_unlink(skb, list); |
1da177e4 LT |
3750 | __kfree_skb(skb); |
3751 | NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); | |
3752 | skb = next; | |
aa8223c7 ACM |
3753 | if (skb == tail || |
3754 | tcp_hdr(skb)->syn || | |
3755 | tcp_hdr(skb)->fin) | |
1da177e4 LT |
3756 | return; |
3757 | } | |
3758 | } | |
3759 | } | |
3760 | } | |
3761 | ||
3762 | /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs | |
3763 | * and tcp_collapse() them until all the queue is collapsed. | |
3764 | */ | |
3765 | static void tcp_collapse_ofo_queue(struct sock *sk) | |
3766 | { | |
3767 | struct tcp_sock *tp = tcp_sk(sk); | |
3768 | struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); | |
3769 | struct sk_buff *head; | |
3770 | u32 start, end; | |
3771 | ||
3772 | if (skb == NULL) | |
3773 | return; | |
3774 | ||
3775 | start = TCP_SKB_CB(skb)->seq; | |
3776 | end = TCP_SKB_CB(skb)->end_seq; | |
3777 | head = skb; | |
3778 | ||
3779 | for (;;) { | |
3780 | skb = skb->next; | |
3781 | ||
3782 | /* Segment is terminated when we see gap or when | |
3783 | * we are at the end of all the queue. */ | |
3784 | if (skb == (struct sk_buff *)&tp->out_of_order_queue || | |
3785 | after(TCP_SKB_CB(skb)->seq, end) || | |
3786 | before(TCP_SKB_CB(skb)->end_seq, start)) { | |
8728b834 DM |
3787 | tcp_collapse(sk, &tp->out_of_order_queue, |
3788 | head, skb, start, end); | |
1da177e4 LT |
3789 | head = skb; |
3790 | if (skb == (struct sk_buff *)&tp->out_of_order_queue) | |
3791 | break; | |
3792 | /* Start new segment */ | |
3793 | start = TCP_SKB_CB(skb)->seq; | |
3794 | end = TCP_SKB_CB(skb)->end_seq; | |
3795 | } else { | |
3796 | if (before(TCP_SKB_CB(skb)->seq, start)) | |
3797 | start = TCP_SKB_CB(skb)->seq; | |
3798 | if (after(TCP_SKB_CB(skb)->end_seq, end)) | |
3799 | end = TCP_SKB_CB(skb)->end_seq; | |
3800 | } | |
3801 | } | |
3802 | } | |
3803 | ||
3804 | /* Reduce allocated memory if we can, trying to get | |
3805 | * the socket within its memory limits again. | |
3806 | * | |
3807 | * Return less than zero if we should start dropping frames | |
3808 | * until the socket owning process reads some of the data | |
3809 | * to stabilize the situation. | |
3810 | */ | |
3811 | static int tcp_prune_queue(struct sock *sk) | |
3812 | { | |
e905a9ed | 3813 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
3814 | |
3815 | SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); | |
3816 | ||
3817 | NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED); | |
3818 | ||
3819 | if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) | |
9e412ba7 | 3820 | tcp_clamp_window(sk); |
1da177e4 LT |
3821 | else if (tcp_memory_pressure) |
3822 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); | |
3823 | ||
3824 | tcp_collapse_ofo_queue(sk); | |
8728b834 DM |
3825 | tcp_collapse(sk, &sk->sk_receive_queue, |
3826 | sk->sk_receive_queue.next, | |
1da177e4 LT |
3827 | (struct sk_buff*)&sk->sk_receive_queue, |
3828 | tp->copied_seq, tp->rcv_nxt); | |
3829 | sk_stream_mem_reclaim(sk); | |
3830 | ||
3831 | if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) | |
3832 | return 0; | |
3833 | ||
3834 | /* Collapsing did not help, destructive actions follow. | |
3835 | * This must not ever occur. */ | |
3836 | ||
3837 | /* First, purge the out_of_order queue. */ | |
b03efcfb DM |
3838 | if (!skb_queue_empty(&tp->out_of_order_queue)) { |
3839 | NET_INC_STATS_BH(LINUX_MIB_OFOPRUNED); | |
1da177e4 LT |
3840 | __skb_queue_purge(&tp->out_of_order_queue); |
3841 | ||
3842 | /* Reset SACK state. A conforming SACK implementation will | |
3843 | * do the same at a timeout based retransmit. When a connection | |
3844 | * is in a sad state like this, we care only about integrity | |
3845 | * of the connection not performance. | |
3846 | */ | |
3847 | if (tp->rx_opt.sack_ok) | |
3848 | tcp_sack_reset(&tp->rx_opt); | |
3849 | sk_stream_mem_reclaim(sk); | |
3850 | } | |
3851 | ||
3852 | if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) | |
3853 | return 0; | |
3854 | ||
3855 | /* If we are really being abused, tell the caller to silently | |
3856 | * drop receive data on the floor. It will get retransmitted | |
3857 | * and hopefully then we'll have sufficient space. | |
3858 | */ | |
3859 | NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED); | |
3860 | ||
3861 | /* Massive buffer overcommit. */ | |
3862 | tp->pred_flags = 0; | |
3863 | return -1; | |
3864 | } | |
3865 | ||
3866 | ||
3867 | /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. | |
3868 | * As additional protections, we do not touch cwnd in retransmission phases, | |
3869 | * and if application hit its sndbuf limit recently. | |
3870 | */ | |
3871 | void tcp_cwnd_application_limited(struct sock *sk) | |
3872 | { | |
3873 | struct tcp_sock *tp = tcp_sk(sk); | |
3874 | ||
6687e988 | 3875 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && |
1da177e4 LT |
3876 | sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { |
3877 | /* Limited by application or receiver window. */ | |
d254bcdb IJ |
3878 | u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); |
3879 | u32 win_used = max(tp->snd_cwnd_used, init_win); | |
1da177e4 | 3880 | if (win_used < tp->snd_cwnd) { |
6687e988 | 3881 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
1da177e4 LT |
3882 | tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; |
3883 | } | |
3884 | tp->snd_cwnd_used = 0; | |
3885 | } | |
3886 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
3887 | } | |
3888 | ||
9e412ba7 | 3889 | static int tcp_should_expand_sndbuf(struct sock *sk) |
0d9901df | 3890 | { |
9e412ba7 IJ |
3891 | struct tcp_sock *tp = tcp_sk(sk); |
3892 | ||
0d9901df DM |
3893 | /* If the user specified a specific send buffer setting, do |
3894 | * not modify it. | |
3895 | */ | |
3896 | if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) | |
3897 | return 0; | |
3898 | ||
3899 | /* If we are under global TCP memory pressure, do not expand. */ | |
3900 | if (tcp_memory_pressure) | |
3901 | return 0; | |
3902 | ||
3903 | /* If we are under soft global TCP memory pressure, do not expand. */ | |
3904 | if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0]) | |
3905 | return 0; | |
3906 | ||
3907 | /* If we filled the congestion window, do not expand. */ | |
3908 | if (tp->packets_out >= tp->snd_cwnd) | |
3909 | return 0; | |
3910 | ||
3911 | return 1; | |
3912 | } | |
1da177e4 LT |
3913 | |
3914 | /* When incoming ACK allowed to free some skb from write_queue, | |
3915 | * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket | |
3916 | * on the exit from tcp input handler. | |
3917 | * | |
3918 | * PROBLEM: sndbuf expansion does not work well with largesend. | |
3919 | */ | |
3920 | static void tcp_new_space(struct sock *sk) | |
3921 | { | |
3922 | struct tcp_sock *tp = tcp_sk(sk); | |
3923 | ||
9e412ba7 | 3924 | if (tcp_should_expand_sndbuf(sk)) { |
e905a9ed | 3925 | int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + |
1da177e4 LT |
3926 | MAX_TCP_HEADER + 16 + sizeof(struct sk_buff), |
3927 | demanded = max_t(unsigned int, tp->snd_cwnd, | |
3928 | tp->reordering + 1); | |
3929 | sndmem *= 2*demanded; | |
3930 | if (sndmem > sk->sk_sndbuf) | |
3931 | sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); | |
3932 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
3933 | } | |
3934 | ||
3935 | sk->sk_write_space(sk); | |
3936 | } | |
3937 | ||
40efc6fa | 3938 | static void tcp_check_space(struct sock *sk) |
1da177e4 LT |
3939 | { |
3940 | if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { | |
3941 | sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); | |
3942 | if (sk->sk_socket && | |
3943 | test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) | |
3944 | tcp_new_space(sk); | |
3945 | } | |
3946 | } | |
3947 | ||
9e412ba7 | 3948 | static inline void tcp_data_snd_check(struct sock *sk) |
1da177e4 | 3949 | { |
9e412ba7 | 3950 | tcp_push_pending_frames(sk); |
1da177e4 LT |
3951 | tcp_check_space(sk); |
3952 | } | |
3953 | ||
3954 | /* | |
3955 | * Check if sending an ack is needed. | |
3956 | */ | |
3957 | static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) | |
3958 | { | |
3959 | struct tcp_sock *tp = tcp_sk(sk); | |
3960 | ||
3961 | /* More than one full frame received... */ | |
463c84b9 | 3962 | if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss |
1da177e4 LT |
3963 | /* ... and right edge of window advances far enough. |
3964 | * (tcp_recvmsg() will send ACK otherwise). Or... | |
3965 | */ | |
3966 | && __tcp_select_window(sk) >= tp->rcv_wnd) || | |
3967 | /* We ACK each frame or... */ | |
463c84b9 | 3968 | tcp_in_quickack_mode(sk) || |
1da177e4 LT |
3969 | /* We have out of order data. */ |
3970 | (ofo_possible && | |
3971 | skb_peek(&tp->out_of_order_queue))) { | |
3972 | /* Then ack it now */ | |
3973 | tcp_send_ack(sk); | |
3974 | } else { | |
3975 | /* Else, send delayed ack. */ | |
3976 | tcp_send_delayed_ack(sk); | |
3977 | } | |
3978 | } | |
3979 | ||
40efc6fa | 3980 | static inline void tcp_ack_snd_check(struct sock *sk) |
1da177e4 | 3981 | { |
463c84b9 | 3982 | if (!inet_csk_ack_scheduled(sk)) { |
1da177e4 LT |
3983 | /* We sent a data segment already. */ |
3984 | return; | |
3985 | } | |
3986 | __tcp_ack_snd_check(sk, 1); | |
3987 | } | |
3988 | ||
3989 | /* | |
3990 | * This routine is only called when we have urgent data | |
caa20d9a | 3991 | * signaled. Its the 'slow' part of tcp_urg. It could be |
1da177e4 LT |
3992 | * moved inline now as tcp_urg is only called from one |
3993 | * place. We handle URGent data wrong. We have to - as | |
3994 | * BSD still doesn't use the correction from RFC961. | |
3995 | * For 1003.1g we should support a new option TCP_STDURG to permit | |
3996 | * either form (or just set the sysctl tcp_stdurg). | |
3997 | */ | |
e905a9ed | 3998 | |
1da177e4 LT |
3999 | static void tcp_check_urg(struct sock * sk, struct tcphdr * th) |
4000 | { | |
4001 | struct tcp_sock *tp = tcp_sk(sk); | |
4002 | u32 ptr = ntohs(th->urg_ptr); | |
4003 | ||
4004 | if (ptr && !sysctl_tcp_stdurg) | |
4005 | ptr--; | |
4006 | ptr += ntohl(th->seq); | |
4007 | ||
4008 | /* Ignore urgent data that we've already seen and read. */ | |
4009 | if (after(tp->copied_seq, ptr)) | |
4010 | return; | |
4011 | ||
4012 | /* Do not replay urg ptr. | |
4013 | * | |
4014 | * NOTE: interesting situation not covered by specs. | |
4015 | * Misbehaving sender may send urg ptr, pointing to segment, | |
4016 | * which we already have in ofo queue. We are not able to fetch | |
4017 | * such data and will stay in TCP_URG_NOTYET until will be eaten | |
4018 | * by recvmsg(). Seems, we are not obliged to handle such wicked | |
4019 | * situations. But it is worth to think about possibility of some | |
4020 | * DoSes using some hypothetical application level deadlock. | |
4021 | */ | |
4022 | if (before(ptr, tp->rcv_nxt)) | |
4023 | return; | |
4024 | ||
4025 | /* Do we already have a newer (or duplicate) urgent pointer? */ | |
4026 | if (tp->urg_data && !after(ptr, tp->urg_seq)) | |
4027 | return; | |
4028 | ||
4029 | /* Tell the world about our new urgent pointer. */ | |
4030 | sk_send_sigurg(sk); | |
4031 | ||
4032 | /* We may be adding urgent data when the last byte read was | |
4033 | * urgent. To do this requires some care. We cannot just ignore | |
4034 | * tp->copied_seq since we would read the last urgent byte again | |
4035 | * as data, nor can we alter copied_seq until this data arrives | |
caa20d9a | 4036 | * or we break the semantics of SIOCATMARK (and thus sockatmark()) |
1da177e4 LT |
4037 | * |
4038 | * NOTE. Double Dutch. Rendering to plain English: author of comment | |
4039 | * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); | |
4040 | * and expect that both A and B disappear from stream. This is _wrong_. | |
4041 | * Though this happens in BSD with high probability, this is occasional. | |
4042 | * Any application relying on this is buggy. Note also, that fix "works" | |
4043 | * only in this artificial test. Insert some normal data between A and B and we will | |
4044 | * decline of BSD again. Verdict: it is better to remove to trap | |
4045 | * buggy users. | |
4046 | */ | |
4047 | if (tp->urg_seq == tp->copied_seq && tp->urg_data && | |
4048 | !sock_flag(sk, SOCK_URGINLINE) && | |
4049 | tp->copied_seq != tp->rcv_nxt) { | |
4050 | struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); | |
4051 | tp->copied_seq++; | |
4052 | if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { | |
8728b834 | 4053 | __skb_unlink(skb, &sk->sk_receive_queue); |
1da177e4 LT |
4054 | __kfree_skb(skb); |
4055 | } | |
4056 | } | |
4057 | ||
4058 | tp->urg_data = TCP_URG_NOTYET; | |
4059 | tp->urg_seq = ptr; | |
4060 | ||
4061 | /* Disable header prediction. */ | |
4062 | tp->pred_flags = 0; | |
4063 | } | |
4064 | ||
4065 | /* This is the 'fast' part of urgent handling. */ | |
4066 | static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th) | |
4067 | { | |
4068 | struct tcp_sock *tp = tcp_sk(sk); | |
4069 | ||
4070 | /* Check if we get a new urgent pointer - normally not. */ | |
4071 | if (th->urg) | |
4072 | tcp_check_urg(sk,th); | |
4073 | ||
4074 | /* Do we wait for any urgent data? - normally not... */ | |
4075 | if (tp->urg_data == TCP_URG_NOTYET) { | |
4076 | u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - | |
4077 | th->syn; | |
4078 | ||
e905a9ed | 4079 | /* Is the urgent pointer pointing into this packet? */ |
1da177e4 LT |
4080 | if (ptr < skb->len) { |
4081 | u8 tmp; | |
4082 | if (skb_copy_bits(skb, ptr, &tmp, 1)) | |
4083 | BUG(); | |
4084 | tp->urg_data = TCP_URG_VALID | tmp; | |
4085 | if (!sock_flag(sk, SOCK_DEAD)) | |
4086 | sk->sk_data_ready(sk, 0); | |
4087 | } | |
4088 | } | |
4089 | } | |
4090 | ||
4091 | static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) | |
4092 | { | |
4093 | struct tcp_sock *tp = tcp_sk(sk); | |
4094 | int chunk = skb->len - hlen; | |
4095 | int err; | |
4096 | ||
4097 | local_bh_enable(); | |
60476372 | 4098 | if (skb_csum_unnecessary(skb)) |
1da177e4 LT |
4099 | err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); |
4100 | else | |
4101 | err = skb_copy_and_csum_datagram_iovec(skb, hlen, | |
4102 | tp->ucopy.iov); | |
4103 | ||
4104 | if (!err) { | |
4105 | tp->ucopy.len -= chunk; | |
4106 | tp->copied_seq += chunk; | |
4107 | tcp_rcv_space_adjust(sk); | |
4108 | } | |
4109 | ||
4110 | local_bh_disable(); | |
4111 | return err; | |
4112 | } | |
4113 | ||
b51655b9 | 4114 | static __sum16 __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 4115 | { |
b51655b9 | 4116 | __sum16 result; |
1da177e4 LT |
4117 | |
4118 | if (sock_owned_by_user(sk)) { | |
4119 | local_bh_enable(); | |
4120 | result = __tcp_checksum_complete(skb); | |
4121 | local_bh_disable(); | |
4122 | } else { | |
4123 | result = __tcp_checksum_complete(skb); | |
4124 | } | |
4125 | return result; | |
4126 | } | |
4127 | ||
40efc6fa | 4128 | static inline int tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 4129 | { |
60476372 | 4130 | return !skb_csum_unnecessary(skb) && |
1da177e4 LT |
4131 | __tcp_checksum_complete_user(sk, skb); |
4132 | } | |
4133 | ||
1a2449a8 CL |
4134 | #ifdef CONFIG_NET_DMA |
4135 | static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, int hlen) | |
4136 | { | |
4137 | struct tcp_sock *tp = tcp_sk(sk); | |
4138 | int chunk = skb->len - hlen; | |
4139 | int dma_cookie; | |
4140 | int copied_early = 0; | |
4141 | ||
4142 | if (tp->ucopy.wakeup) | |
e905a9ed | 4143 | return 0; |
1a2449a8 CL |
4144 | |
4145 | if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) | |
4146 | tp->ucopy.dma_chan = get_softnet_dma(); | |
4147 | ||
60476372 | 4148 | if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) { |
1a2449a8 CL |
4149 | |
4150 | dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, | |
4151 | skb, hlen, tp->ucopy.iov, chunk, tp->ucopy.pinned_list); | |
4152 | ||
4153 | if (dma_cookie < 0) | |
4154 | goto out; | |
4155 | ||
4156 | tp->ucopy.dma_cookie = dma_cookie; | |
4157 | copied_early = 1; | |
4158 | ||
4159 | tp->ucopy.len -= chunk; | |
4160 | tp->copied_seq += chunk; | |
4161 | tcp_rcv_space_adjust(sk); | |
4162 | ||
4163 | if ((tp->ucopy.len == 0) || | |
aa8223c7 | 4164 | (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) || |
1a2449a8 CL |
4165 | (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { |
4166 | tp->ucopy.wakeup = 1; | |
4167 | sk->sk_data_ready(sk, 0); | |
4168 | } | |
4169 | } else if (chunk > 0) { | |
4170 | tp->ucopy.wakeup = 1; | |
4171 | sk->sk_data_ready(sk, 0); | |
4172 | } | |
4173 | out: | |
4174 | return copied_early; | |
4175 | } | |
4176 | #endif /* CONFIG_NET_DMA */ | |
4177 | ||
1da177e4 | 4178 | /* |
e905a9ed | 4179 | * TCP receive function for the ESTABLISHED state. |
1da177e4 | 4180 | * |
e905a9ed | 4181 | * It is split into a fast path and a slow path. The fast path is |
1da177e4 LT |
4182 | * disabled when: |
4183 | * - A zero window was announced from us - zero window probing | |
e905a9ed | 4184 | * is only handled properly in the slow path. |
1da177e4 LT |
4185 | * - Out of order segments arrived. |
4186 | * - Urgent data is expected. | |
4187 | * - There is no buffer space left | |
4188 | * - Unexpected TCP flags/window values/header lengths are received | |
e905a9ed | 4189 | * (detected by checking the TCP header against pred_flags) |
1da177e4 LT |
4190 | * - Data is sent in both directions. Fast path only supports pure senders |
4191 | * or pure receivers (this means either the sequence number or the ack | |
4192 | * value must stay constant) | |
4193 | * - Unexpected TCP option. | |
4194 | * | |
e905a9ed | 4195 | * When these conditions are not satisfied it drops into a standard |
1da177e4 LT |
4196 | * receive procedure patterned after RFC793 to handle all cases. |
4197 | * The first three cases are guaranteed by proper pred_flags setting, | |
e905a9ed | 4198 | * the rest is checked inline. Fast processing is turned on in |
1da177e4 LT |
4199 | * tcp_data_queue when everything is OK. |
4200 | */ | |
4201 | int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, | |
4202 | struct tcphdr *th, unsigned len) | |
4203 | { | |
4204 | struct tcp_sock *tp = tcp_sk(sk); | |
4205 | ||
4206 | /* | |
4207 | * Header prediction. | |
e905a9ed | 4208 | * The code loosely follows the one in the famous |
1da177e4 | 4209 | * "30 instruction TCP receive" Van Jacobson mail. |
e905a9ed YH |
4210 | * |
4211 | * Van's trick is to deposit buffers into socket queue | |
1da177e4 LT |
4212 | * on a device interrupt, to call tcp_recv function |
4213 | * on the receive process context and checksum and copy | |
4214 | * the buffer to user space. smart... | |
4215 | * | |
e905a9ed | 4216 | * Our current scheme is not silly either but we take the |
1da177e4 LT |
4217 | * extra cost of the net_bh soft interrupt processing... |
4218 | * We do checksum and copy also but from device to kernel. | |
4219 | */ | |
4220 | ||
4221 | tp->rx_opt.saw_tstamp = 0; | |
4222 | ||
4223 | /* pred_flags is 0xS?10 << 16 + snd_wnd | |
caa20d9a | 4224 | * if header_prediction is to be made |
1da177e4 LT |
4225 | * 'S' will always be tp->tcp_header_len >> 2 |
4226 | * '?' will be 0 for the fast path, otherwise pred_flags is 0 to | |
e905a9ed | 4227 | * turn it off (when there are holes in the receive |
1da177e4 LT |
4228 | * space for instance) |
4229 | * PSH flag is ignored. | |
4230 | */ | |
4231 | ||
4232 | if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && | |
4233 | TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { | |
4234 | int tcp_header_len = tp->tcp_header_len; | |
4235 | ||
4236 | /* Timestamp header prediction: tcp_header_len | |
4237 | * is automatically equal to th->doff*4 due to pred_flags | |
4238 | * match. | |
4239 | */ | |
4240 | ||
4241 | /* Check timestamp */ | |
4242 | if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { | |
4f3608b7 | 4243 | __be32 *ptr = (__be32 *)(th + 1); |
1da177e4 LT |
4244 | |
4245 | /* No? Slow path! */ | |
4f3608b7 | 4246 | if (*ptr != htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
1da177e4 LT |
4247 | | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) |
4248 | goto slow_path; | |
4249 | ||
4250 | tp->rx_opt.saw_tstamp = 1; | |
e905a9ed | 4251 | ++ptr; |
1da177e4 LT |
4252 | tp->rx_opt.rcv_tsval = ntohl(*ptr); |
4253 | ++ptr; | |
4254 | tp->rx_opt.rcv_tsecr = ntohl(*ptr); | |
4255 | ||
4256 | /* If PAWS failed, check it more carefully in slow path */ | |
4257 | if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) | |
4258 | goto slow_path; | |
4259 | ||
4260 | /* DO NOT update ts_recent here, if checksum fails | |
4261 | * and timestamp was corrupted part, it will result | |
4262 | * in a hung connection since we will drop all | |
4263 | * future packets due to the PAWS test. | |
4264 | */ | |
4265 | } | |
4266 | ||
4267 | if (len <= tcp_header_len) { | |
4268 | /* Bulk data transfer: sender */ | |
4269 | if (len == tcp_header_len) { | |
4270 | /* Predicted packet is in window by definition. | |
4271 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. | |
4272 | * Hence, check seq<=rcv_wup reduces to: | |
4273 | */ | |
4274 | if (tcp_header_len == | |
4275 | (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && | |
4276 | tp->rcv_nxt == tp->rcv_wup) | |
4277 | tcp_store_ts_recent(tp); | |
4278 | ||
1da177e4 LT |
4279 | /* We know that such packets are checksummed |
4280 | * on entry. | |
4281 | */ | |
4282 | tcp_ack(sk, skb, 0); | |
e905a9ed | 4283 | __kfree_skb(skb); |
9e412ba7 | 4284 | tcp_data_snd_check(sk); |
1da177e4 LT |
4285 | return 0; |
4286 | } else { /* Header too small */ | |
4287 | TCP_INC_STATS_BH(TCP_MIB_INERRS); | |
4288 | goto discard; | |
4289 | } | |
4290 | } else { | |
4291 | int eaten = 0; | |
1a2449a8 | 4292 | int copied_early = 0; |
1da177e4 | 4293 | |
1a2449a8 CL |
4294 | if (tp->copied_seq == tp->rcv_nxt && |
4295 | len - tcp_header_len <= tp->ucopy.len) { | |
4296 | #ifdef CONFIG_NET_DMA | |
4297 | if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { | |
4298 | copied_early = 1; | |
4299 | eaten = 1; | |
4300 | } | |
4301 | #endif | |
4302 | if (tp->ucopy.task == current && sock_owned_by_user(sk) && !copied_early) { | |
4303 | __set_current_state(TASK_RUNNING); | |
1da177e4 | 4304 | |
1a2449a8 CL |
4305 | if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) |
4306 | eaten = 1; | |
4307 | } | |
4308 | if (eaten) { | |
1da177e4 LT |
4309 | /* Predicted packet is in window by definition. |
4310 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. | |
4311 | * Hence, check seq<=rcv_wup reduces to: | |
4312 | */ | |
4313 | if (tcp_header_len == | |
4314 | (sizeof(struct tcphdr) + | |
4315 | TCPOLEN_TSTAMP_ALIGNED) && | |
4316 | tp->rcv_nxt == tp->rcv_wup) | |
4317 | tcp_store_ts_recent(tp); | |
4318 | ||
463c84b9 | 4319 | tcp_rcv_rtt_measure_ts(sk, skb); |
1da177e4 LT |
4320 | |
4321 | __skb_pull(skb, tcp_header_len); | |
4322 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
4323 | NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER); | |
1da177e4 | 4324 | } |
1a2449a8 CL |
4325 | if (copied_early) |
4326 | tcp_cleanup_rbuf(sk, skb->len); | |
1da177e4 LT |
4327 | } |
4328 | if (!eaten) { | |
4329 | if (tcp_checksum_complete_user(sk, skb)) | |
4330 | goto csum_error; | |
4331 | ||
4332 | /* Predicted packet is in window by definition. | |
4333 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. | |
4334 | * Hence, check seq<=rcv_wup reduces to: | |
4335 | */ | |
4336 | if (tcp_header_len == | |
4337 | (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && | |
4338 | tp->rcv_nxt == tp->rcv_wup) | |
4339 | tcp_store_ts_recent(tp); | |
4340 | ||
463c84b9 | 4341 | tcp_rcv_rtt_measure_ts(sk, skb); |
1da177e4 LT |
4342 | |
4343 | if ((int)skb->truesize > sk->sk_forward_alloc) | |
4344 | goto step5; | |
4345 | ||
4346 | NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS); | |
4347 | ||
4348 | /* Bulk data transfer: receiver */ | |
4349 | __skb_pull(skb,tcp_header_len); | |
4350 | __skb_queue_tail(&sk->sk_receive_queue, skb); | |
4351 | sk_stream_set_owner_r(skb, sk); | |
4352 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
4353 | } | |
4354 | ||
9e412ba7 | 4355 | tcp_event_data_recv(sk, skb); |
1da177e4 LT |
4356 | |
4357 | if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { | |
4358 | /* Well, only one small jumplet in fast path... */ | |
4359 | tcp_ack(sk, skb, FLAG_DATA); | |
9e412ba7 | 4360 | tcp_data_snd_check(sk); |
463c84b9 | 4361 | if (!inet_csk_ack_scheduled(sk)) |
1da177e4 LT |
4362 | goto no_ack; |
4363 | } | |
4364 | ||
31432412 | 4365 | __tcp_ack_snd_check(sk, 0); |
1da177e4 | 4366 | no_ack: |
1a2449a8 CL |
4367 | #ifdef CONFIG_NET_DMA |
4368 | if (copied_early) | |
4369 | __skb_queue_tail(&sk->sk_async_wait_queue, skb); | |
4370 | else | |
4371 | #endif | |
1da177e4 LT |
4372 | if (eaten) |
4373 | __kfree_skb(skb); | |
4374 | else | |
4375 | sk->sk_data_ready(sk, 0); | |
4376 | return 0; | |
4377 | } | |
4378 | } | |
4379 | ||
4380 | slow_path: | |
4381 | if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb)) | |
4382 | goto csum_error; | |
4383 | ||
4384 | /* | |
4385 | * RFC1323: H1. Apply PAWS check first. | |
4386 | */ | |
4387 | if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && | |
463c84b9 | 4388 | tcp_paws_discard(sk, skb)) { |
1da177e4 LT |
4389 | if (!th->rst) { |
4390 | NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); | |
4391 | tcp_send_dupack(sk, skb); | |
4392 | goto discard; | |
4393 | } | |
4394 | /* Resets are accepted even if PAWS failed. | |
4395 | ||
4396 | ts_recent update must be made after we are sure | |
4397 | that the packet is in window. | |
4398 | */ | |
4399 | } | |
4400 | ||
4401 | /* | |
4402 | * Standard slow path. | |
4403 | */ | |
4404 | ||
4405 | if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { | |
4406 | /* RFC793, page 37: "In all states except SYN-SENT, all reset | |
4407 | * (RST) segments are validated by checking their SEQ-fields." | |
4408 | * And page 69: "If an incoming segment is not acceptable, | |
4409 | * an acknowledgment should be sent in reply (unless the RST bit | |
4410 | * is set, if so drop the segment and return)". | |
4411 | */ | |
4412 | if (!th->rst) | |
4413 | tcp_send_dupack(sk, skb); | |
4414 | goto discard; | |
4415 | } | |
4416 | ||
2de979bd | 4417 | if (th->rst) { |
1da177e4 LT |
4418 | tcp_reset(sk); |
4419 | goto discard; | |
4420 | } | |
4421 | ||
4422 | tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); | |
4423 | ||
4424 | if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
4425 | TCP_INC_STATS_BH(TCP_MIB_INERRS); | |
4426 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); | |
4427 | tcp_reset(sk); | |
4428 | return 1; | |
4429 | } | |
4430 | ||
4431 | step5: | |
2de979bd | 4432 | if (th->ack) |
1da177e4 LT |
4433 | tcp_ack(sk, skb, FLAG_SLOWPATH); |
4434 | ||
463c84b9 | 4435 | tcp_rcv_rtt_measure_ts(sk, skb); |
1da177e4 LT |
4436 | |
4437 | /* Process urgent data. */ | |
4438 | tcp_urg(sk, skb, th); | |
4439 | ||
4440 | /* step 7: process the segment text */ | |
4441 | tcp_data_queue(sk, skb); | |
4442 | ||
9e412ba7 | 4443 | tcp_data_snd_check(sk); |
1da177e4 LT |
4444 | tcp_ack_snd_check(sk); |
4445 | return 0; | |
4446 | ||
4447 | csum_error: | |
4448 | TCP_INC_STATS_BH(TCP_MIB_INERRS); | |
4449 | ||
4450 | discard: | |
4451 | __kfree_skb(skb); | |
4452 | return 0; | |
4453 | } | |
4454 | ||
4455 | static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, | |
4456 | struct tcphdr *th, unsigned len) | |
4457 | { | |
4458 | struct tcp_sock *tp = tcp_sk(sk); | |
d83d8461 | 4459 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
4460 | int saved_clamp = tp->rx_opt.mss_clamp; |
4461 | ||
4462 | tcp_parse_options(skb, &tp->rx_opt, 0); | |
4463 | ||
4464 | if (th->ack) { | |
4465 | /* rfc793: | |
4466 | * "If the state is SYN-SENT then | |
4467 | * first check the ACK bit | |
4468 | * If the ACK bit is set | |
4469 | * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send | |
4470 | * a reset (unless the RST bit is set, if so drop | |
4471 | * the segment and return)" | |
4472 | * | |
4473 | * We do not send data with SYN, so that RFC-correct | |
4474 | * test reduces to: | |
4475 | */ | |
4476 | if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) | |
4477 | goto reset_and_undo; | |
4478 | ||
4479 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && | |
4480 | !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, | |
4481 | tcp_time_stamp)) { | |
4482 | NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED); | |
4483 | goto reset_and_undo; | |
4484 | } | |
4485 | ||
4486 | /* Now ACK is acceptable. | |
4487 | * | |
4488 | * "If the RST bit is set | |
4489 | * If the ACK was acceptable then signal the user "error: | |
4490 | * connection reset", drop the segment, enter CLOSED state, | |
4491 | * delete TCB, and return." | |
4492 | */ | |
4493 | ||
4494 | if (th->rst) { | |
4495 | tcp_reset(sk); | |
4496 | goto discard; | |
4497 | } | |
4498 | ||
4499 | /* rfc793: | |
4500 | * "fifth, if neither of the SYN or RST bits is set then | |
4501 | * drop the segment and return." | |
4502 | * | |
4503 | * See note below! | |
4504 | * --ANK(990513) | |
4505 | */ | |
4506 | if (!th->syn) | |
4507 | goto discard_and_undo; | |
4508 | ||
4509 | /* rfc793: | |
4510 | * "If the SYN bit is on ... | |
4511 | * are acceptable then ... | |
4512 | * (our SYN has been ACKed), change the connection | |
4513 | * state to ESTABLISHED..." | |
4514 | */ | |
4515 | ||
4516 | TCP_ECN_rcv_synack(tp, th); | |
1da177e4 LT |
4517 | |
4518 | tp->snd_wl1 = TCP_SKB_CB(skb)->seq; | |
4519 | tcp_ack(sk, skb, FLAG_SLOWPATH); | |
4520 | ||
4521 | /* Ok.. it's good. Set up sequence numbers and | |
4522 | * move to established. | |
4523 | */ | |
4524 | tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; | |
4525 | tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; | |
4526 | ||
4527 | /* RFC1323: The window in SYN & SYN/ACK segments is | |
4528 | * never scaled. | |
4529 | */ | |
4530 | tp->snd_wnd = ntohs(th->window); | |
4531 | tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq); | |
4532 | ||
4533 | if (!tp->rx_opt.wscale_ok) { | |
4534 | tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; | |
4535 | tp->window_clamp = min(tp->window_clamp, 65535U); | |
4536 | } | |
4537 | ||
4538 | if (tp->rx_opt.saw_tstamp) { | |
4539 | tp->rx_opt.tstamp_ok = 1; | |
4540 | tp->tcp_header_len = | |
4541 | sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; | |
4542 | tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; | |
4543 | tcp_store_ts_recent(tp); | |
4544 | } else { | |
4545 | tp->tcp_header_len = sizeof(struct tcphdr); | |
4546 | } | |
4547 | ||
4548 | if (tp->rx_opt.sack_ok && sysctl_tcp_fack) | |
4549 | tp->rx_opt.sack_ok |= 2; | |
4550 | ||
5d424d5a | 4551 | tcp_mtup_init(sk); |
d83d8461 | 4552 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); |
1da177e4 LT |
4553 | tcp_initialize_rcv_mss(sk); |
4554 | ||
4555 | /* Remember, tcp_poll() does not lock socket! | |
4556 | * Change state from SYN-SENT only after copied_seq | |
4557 | * is initialized. */ | |
4558 | tp->copied_seq = tp->rcv_nxt; | |
e16aa207 | 4559 | smp_mb(); |
1da177e4 LT |
4560 | tcp_set_state(sk, TCP_ESTABLISHED); |
4561 | ||
6b877699 VY |
4562 | security_inet_conn_established(sk, skb); |
4563 | ||
1da177e4 | 4564 | /* Make sure socket is routed, for correct metrics. */ |
8292a17a | 4565 | icsk->icsk_af_ops->rebuild_header(sk); |
1da177e4 LT |
4566 | |
4567 | tcp_init_metrics(sk); | |
4568 | ||
6687e988 | 4569 | tcp_init_congestion_control(sk); |
317a76f9 | 4570 | |
1da177e4 LT |
4571 | /* Prevent spurious tcp_cwnd_restart() on first data |
4572 | * packet. | |
4573 | */ | |
4574 | tp->lsndtime = tcp_time_stamp; | |
4575 | ||
4576 | tcp_init_buffer_space(sk); | |
4577 | ||
4578 | if (sock_flag(sk, SOCK_KEEPOPEN)) | |
463c84b9 | 4579 | inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); |
1da177e4 LT |
4580 | |
4581 | if (!tp->rx_opt.snd_wscale) | |
4582 | __tcp_fast_path_on(tp, tp->snd_wnd); | |
4583 | else | |
4584 | tp->pred_flags = 0; | |
4585 | ||
4586 | if (!sock_flag(sk, SOCK_DEAD)) { | |
4587 | sk->sk_state_change(sk); | |
4588 | sk_wake_async(sk, 0, POLL_OUT); | |
4589 | } | |
4590 | ||
295f7324 ACM |
4591 | if (sk->sk_write_pending || |
4592 | icsk->icsk_accept_queue.rskq_defer_accept || | |
4593 | icsk->icsk_ack.pingpong) { | |
1da177e4 LT |
4594 | /* Save one ACK. Data will be ready after |
4595 | * several ticks, if write_pending is set. | |
4596 | * | |
4597 | * It may be deleted, but with this feature tcpdumps | |
4598 | * look so _wonderfully_ clever, that I was not able | |
4599 | * to stand against the temptation 8) --ANK | |
4600 | */ | |
463c84b9 | 4601 | inet_csk_schedule_ack(sk); |
295f7324 ACM |
4602 | icsk->icsk_ack.lrcvtime = tcp_time_stamp; |
4603 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
463c84b9 ACM |
4604 | tcp_incr_quickack(sk); |
4605 | tcp_enter_quickack_mode(sk); | |
3f421baa ACM |
4606 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, |
4607 | TCP_DELACK_MAX, TCP_RTO_MAX); | |
1da177e4 LT |
4608 | |
4609 | discard: | |
4610 | __kfree_skb(skb); | |
4611 | return 0; | |
4612 | } else { | |
4613 | tcp_send_ack(sk); | |
4614 | } | |
4615 | return -1; | |
4616 | } | |
4617 | ||
4618 | /* No ACK in the segment */ | |
4619 | ||
4620 | if (th->rst) { | |
4621 | /* rfc793: | |
4622 | * "If the RST bit is set | |
4623 | * | |
4624 | * Otherwise (no ACK) drop the segment and return." | |
4625 | */ | |
4626 | ||
4627 | goto discard_and_undo; | |
4628 | } | |
4629 | ||
4630 | /* PAWS check. */ | |
4631 | if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0)) | |
4632 | goto discard_and_undo; | |
4633 | ||
4634 | if (th->syn) { | |
4635 | /* We see SYN without ACK. It is attempt of | |
4636 | * simultaneous connect with crossed SYNs. | |
4637 | * Particularly, it can be connect to self. | |
4638 | */ | |
4639 | tcp_set_state(sk, TCP_SYN_RECV); | |
4640 | ||
4641 | if (tp->rx_opt.saw_tstamp) { | |
4642 | tp->rx_opt.tstamp_ok = 1; | |
4643 | tcp_store_ts_recent(tp); | |
4644 | tp->tcp_header_len = | |
4645 | sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; | |
4646 | } else { | |
4647 | tp->tcp_header_len = sizeof(struct tcphdr); | |
4648 | } | |
4649 | ||
4650 | tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; | |
4651 | tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; | |
4652 | ||
4653 | /* RFC1323: The window in SYN & SYN/ACK segments is | |
4654 | * never scaled. | |
4655 | */ | |
4656 | tp->snd_wnd = ntohs(th->window); | |
4657 | tp->snd_wl1 = TCP_SKB_CB(skb)->seq; | |
4658 | tp->max_window = tp->snd_wnd; | |
4659 | ||
4660 | TCP_ECN_rcv_syn(tp, th); | |
1da177e4 | 4661 | |
5d424d5a | 4662 | tcp_mtup_init(sk); |
d83d8461 | 4663 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); |
1da177e4 LT |
4664 | tcp_initialize_rcv_mss(sk); |
4665 | ||
4666 | ||
4667 | tcp_send_synack(sk); | |
4668 | #if 0 | |
4669 | /* Note, we could accept data and URG from this segment. | |
4670 | * There are no obstacles to make this. | |
4671 | * | |
4672 | * However, if we ignore data in ACKless segments sometimes, | |
4673 | * we have no reasons to accept it sometimes. | |
4674 | * Also, seems the code doing it in step6 of tcp_rcv_state_process | |
4675 | * is not flawless. So, discard packet for sanity. | |
4676 | * Uncomment this return to process the data. | |
4677 | */ | |
4678 | return -1; | |
4679 | #else | |
4680 | goto discard; | |
4681 | #endif | |
4682 | } | |
4683 | /* "fifth, if neither of the SYN or RST bits is set then | |
4684 | * drop the segment and return." | |
4685 | */ | |
4686 | ||
4687 | discard_and_undo: | |
4688 | tcp_clear_options(&tp->rx_opt); | |
4689 | tp->rx_opt.mss_clamp = saved_clamp; | |
4690 | goto discard; | |
4691 | ||
4692 | reset_and_undo: | |
4693 | tcp_clear_options(&tp->rx_opt); | |
4694 | tp->rx_opt.mss_clamp = saved_clamp; | |
4695 | return 1; | |
4696 | } | |
4697 | ||
4698 | ||
4699 | /* | |
4700 | * This function implements the receiving procedure of RFC 793 for | |
e905a9ed | 4701 | * all states except ESTABLISHED and TIME_WAIT. |
1da177e4 LT |
4702 | * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be |
4703 | * address independent. | |
4704 | */ | |
e905a9ed | 4705 | |
1da177e4 LT |
4706 | int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, |
4707 | struct tcphdr *th, unsigned len) | |
4708 | { | |
4709 | struct tcp_sock *tp = tcp_sk(sk); | |
8292a17a | 4710 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
4711 | int queued = 0; |
4712 | ||
4713 | tp->rx_opt.saw_tstamp = 0; | |
4714 | ||
4715 | switch (sk->sk_state) { | |
4716 | case TCP_CLOSE: | |
4717 | goto discard; | |
4718 | ||
4719 | case TCP_LISTEN: | |
2de979bd | 4720 | if (th->ack) |
1da177e4 LT |
4721 | return 1; |
4722 | ||
2de979bd | 4723 | if (th->rst) |
1da177e4 LT |
4724 | goto discard; |
4725 | ||
2de979bd | 4726 | if (th->syn) { |
8292a17a | 4727 | if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) |
1da177e4 LT |
4728 | return 1; |
4729 | ||
e905a9ed YH |
4730 | /* Now we have several options: In theory there is |
4731 | * nothing else in the frame. KA9Q has an option to | |
1da177e4 | 4732 | * send data with the syn, BSD accepts data with the |
e905a9ed YH |
4733 | * syn up to the [to be] advertised window and |
4734 | * Solaris 2.1 gives you a protocol error. For now | |
4735 | * we just ignore it, that fits the spec precisely | |
1da177e4 LT |
4736 | * and avoids incompatibilities. It would be nice in |
4737 | * future to drop through and process the data. | |
4738 | * | |
e905a9ed | 4739 | * Now that TTCP is starting to be used we ought to |
1da177e4 LT |
4740 | * queue this data. |
4741 | * But, this leaves one open to an easy denial of | |
e905a9ed | 4742 | * service attack, and SYN cookies can't defend |
1da177e4 | 4743 | * against this problem. So, we drop the data |
fb7e2399 MN |
4744 | * in the interest of security over speed unless |
4745 | * it's still in use. | |
1da177e4 | 4746 | */ |
fb7e2399 MN |
4747 | kfree_skb(skb); |
4748 | return 0; | |
1da177e4 LT |
4749 | } |
4750 | goto discard; | |
4751 | ||
4752 | case TCP_SYN_SENT: | |
1da177e4 LT |
4753 | queued = tcp_rcv_synsent_state_process(sk, skb, th, len); |
4754 | if (queued >= 0) | |
4755 | return queued; | |
4756 | ||
4757 | /* Do step6 onward by hand. */ | |
4758 | tcp_urg(sk, skb, th); | |
4759 | __kfree_skb(skb); | |
9e412ba7 | 4760 | tcp_data_snd_check(sk); |
1da177e4 LT |
4761 | return 0; |
4762 | } | |
4763 | ||
4764 | if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && | |
463c84b9 | 4765 | tcp_paws_discard(sk, skb)) { |
1da177e4 LT |
4766 | if (!th->rst) { |
4767 | NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); | |
4768 | tcp_send_dupack(sk, skb); | |
4769 | goto discard; | |
4770 | } | |
4771 | /* Reset is accepted even if it did not pass PAWS. */ | |
4772 | } | |
4773 | ||
4774 | /* step 1: check sequence number */ | |
4775 | if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { | |
4776 | if (!th->rst) | |
4777 | tcp_send_dupack(sk, skb); | |
4778 | goto discard; | |
4779 | } | |
4780 | ||
4781 | /* step 2: check RST bit */ | |
2de979bd | 4782 | if (th->rst) { |
1da177e4 LT |
4783 | tcp_reset(sk); |
4784 | goto discard; | |
4785 | } | |
4786 | ||
4787 | tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); | |
4788 | ||
4789 | /* step 3: check security and precedence [ignored] */ | |
4790 | ||
4791 | /* step 4: | |
4792 | * | |
4793 | * Check for a SYN in window. | |
4794 | */ | |
4795 | if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
4796 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); | |
4797 | tcp_reset(sk); | |
4798 | return 1; | |
4799 | } | |
4800 | ||
4801 | /* step 5: check the ACK field */ | |
4802 | if (th->ack) { | |
4803 | int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH); | |
4804 | ||
2de979bd | 4805 | switch (sk->sk_state) { |
1da177e4 LT |
4806 | case TCP_SYN_RECV: |
4807 | if (acceptable) { | |
4808 | tp->copied_seq = tp->rcv_nxt; | |
e16aa207 | 4809 | smp_mb(); |
1da177e4 LT |
4810 | tcp_set_state(sk, TCP_ESTABLISHED); |
4811 | sk->sk_state_change(sk); | |
4812 | ||
4813 | /* Note, that this wakeup is only for marginal | |
4814 | * crossed SYN case. Passively open sockets | |
4815 | * are not waked up, because sk->sk_sleep == | |
4816 | * NULL and sk->sk_socket == NULL. | |
4817 | */ | |
4818 | if (sk->sk_socket) { | |
4819 | sk_wake_async(sk,0,POLL_OUT); | |
4820 | } | |
4821 | ||
4822 | tp->snd_una = TCP_SKB_CB(skb)->ack_seq; | |
4823 | tp->snd_wnd = ntohs(th->window) << | |
4824 | tp->rx_opt.snd_wscale; | |
4825 | tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, | |
4826 | TCP_SKB_CB(skb)->seq); | |
4827 | ||
4828 | /* tcp_ack considers this ACK as duplicate | |
4829 | * and does not calculate rtt. | |
4830 | * Fix it at least with timestamps. | |
4831 | */ | |
4832 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && | |
4833 | !tp->srtt) | |
2d2abbab | 4834 | tcp_ack_saw_tstamp(sk, 0); |
1da177e4 LT |
4835 | |
4836 | if (tp->rx_opt.tstamp_ok) | |
4837 | tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; | |
4838 | ||
4839 | /* Make sure socket is routed, for | |
4840 | * correct metrics. | |
4841 | */ | |
8292a17a | 4842 | icsk->icsk_af_ops->rebuild_header(sk); |
1da177e4 LT |
4843 | |
4844 | tcp_init_metrics(sk); | |
4845 | ||
6687e988 | 4846 | tcp_init_congestion_control(sk); |
317a76f9 | 4847 | |
1da177e4 LT |
4848 | /* Prevent spurious tcp_cwnd_restart() on |
4849 | * first data packet. | |
4850 | */ | |
4851 | tp->lsndtime = tcp_time_stamp; | |
4852 | ||
5d424d5a | 4853 | tcp_mtup_init(sk); |
1da177e4 LT |
4854 | tcp_initialize_rcv_mss(sk); |
4855 | tcp_init_buffer_space(sk); | |
4856 | tcp_fast_path_on(tp); | |
4857 | } else { | |
4858 | return 1; | |
4859 | } | |
4860 | break; | |
4861 | ||
4862 | case TCP_FIN_WAIT1: | |
4863 | if (tp->snd_una == tp->write_seq) { | |
4864 | tcp_set_state(sk, TCP_FIN_WAIT2); | |
4865 | sk->sk_shutdown |= SEND_SHUTDOWN; | |
4866 | dst_confirm(sk->sk_dst_cache); | |
4867 | ||
4868 | if (!sock_flag(sk, SOCK_DEAD)) | |
4869 | /* Wake up lingering close() */ | |
4870 | sk->sk_state_change(sk); | |
4871 | else { | |
4872 | int tmo; | |
4873 | ||
4874 | if (tp->linger2 < 0 || | |
4875 | (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && | |
4876 | after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { | |
4877 | tcp_done(sk); | |
4878 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); | |
4879 | return 1; | |
4880 | } | |
4881 | ||
463c84b9 | 4882 | tmo = tcp_fin_time(sk); |
1da177e4 | 4883 | if (tmo > TCP_TIMEWAIT_LEN) { |
463c84b9 | 4884 | inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); |
1da177e4 LT |
4885 | } else if (th->fin || sock_owned_by_user(sk)) { |
4886 | /* Bad case. We could lose such FIN otherwise. | |
4887 | * It is not a big problem, but it looks confusing | |
4888 | * and not so rare event. We still can lose it now, | |
4889 | * if it spins in bh_lock_sock(), but it is really | |
4890 | * marginal case. | |
4891 | */ | |
463c84b9 | 4892 | inet_csk_reset_keepalive_timer(sk, tmo); |
1da177e4 LT |
4893 | } else { |
4894 | tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); | |
4895 | goto discard; | |
4896 | } | |
4897 | } | |
4898 | } | |
4899 | break; | |
4900 | ||
4901 | case TCP_CLOSING: | |
4902 | if (tp->snd_una == tp->write_seq) { | |
4903 | tcp_time_wait(sk, TCP_TIME_WAIT, 0); | |
4904 | goto discard; | |
4905 | } | |
4906 | break; | |
4907 | ||
4908 | case TCP_LAST_ACK: | |
4909 | if (tp->snd_una == tp->write_seq) { | |
4910 | tcp_update_metrics(sk); | |
4911 | tcp_done(sk); | |
4912 | goto discard; | |
4913 | } | |
4914 | break; | |
4915 | } | |
4916 | } else | |
4917 | goto discard; | |
4918 | ||
4919 | /* step 6: check the URG bit */ | |
4920 | tcp_urg(sk, skb, th); | |
4921 | ||
4922 | /* step 7: process the segment text */ | |
4923 | switch (sk->sk_state) { | |
4924 | case TCP_CLOSE_WAIT: | |
4925 | case TCP_CLOSING: | |
4926 | case TCP_LAST_ACK: | |
4927 | if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) | |
4928 | break; | |
4929 | case TCP_FIN_WAIT1: | |
4930 | case TCP_FIN_WAIT2: | |
4931 | /* RFC 793 says to queue data in these states, | |
e905a9ed | 4932 | * RFC 1122 says we MUST send a reset. |
1da177e4 LT |
4933 | * BSD 4.4 also does reset. |
4934 | */ | |
4935 | if (sk->sk_shutdown & RCV_SHUTDOWN) { | |
4936 | if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && | |
4937 | after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { | |
4938 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); | |
4939 | tcp_reset(sk); | |
4940 | return 1; | |
4941 | } | |
4942 | } | |
4943 | /* Fall through */ | |
e905a9ed | 4944 | case TCP_ESTABLISHED: |
1da177e4 LT |
4945 | tcp_data_queue(sk, skb); |
4946 | queued = 1; | |
4947 | break; | |
4948 | } | |
4949 | ||
4950 | /* tcp_data could move socket to TIME-WAIT */ | |
4951 | if (sk->sk_state != TCP_CLOSE) { | |
9e412ba7 | 4952 | tcp_data_snd_check(sk); |
1da177e4 LT |
4953 | tcp_ack_snd_check(sk); |
4954 | } | |
4955 | ||
e905a9ed | 4956 | if (!queued) { |
1da177e4 LT |
4957 | discard: |
4958 | __kfree_skb(skb); | |
4959 | } | |
4960 | return 0; | |
4961 | } | |
4962 | ||
4963 | EXPORT_SYMBOL(sysctl_tcp_ecn); | |
4964 | EXPORT_SYMBOL(sysctl_tcp_reordering); | |
4965 | EXPORT_SYMBOL(tcp_parse_options); | |
4966 | EXPORT_SYMBOL(tcp_rcv_established); | |
4967 | EXPORT_SYMBOL(tcp_rcv_state_process); | |
40efc6fa | 4968 | EXPORT_SYMBOL(tcp_initialize_rcv_mss); |