]>
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. | |
45 | * Andrey Savochkin: Fix RTT measurements in the presnce of | |
46 | * timestamps. | |
47 | * Andrey Savochkin: Check sequence numbers correctly when | |
48 | * removing SACKs due to in sequence incoming | |
49 | * data segments. | |
50 | * Andi Kleen: Make sure we never ack data there is not | |
51 | * enough room for. Also make this condition | |
52 | * a fatal error if it might still happen. | |
53 | * Andi Kleen: Add tcp_measure_rcv_mss to make | |
54 | * connections with MSS<min(MTU,ann. MSS) | |
55 | * work without delayed acks. | |
56 | * Andi Kleen: Process packets with PSH set in the | |
57 | * fast path. | |
58 | * J Hadi Salim: ECN support | |
59 | * Andrei Gurtov, | |
60 | * Pasi Sarolahti, | |
61 | * Panu Kuhlberg: Experimental audit of TCP (re)transmission | |
62 | * engine. Lots of bugs are found. | |
63 | * Pasi Sarolahti: F-RTO for dealing with spurious RTOs | |
64 | * Angelo Dell'Aera: TCP Westwood+ support | |
65 | */ | |
66 | ||
67 | #include <linux/config.h> | |
68 | #include <linux/mm.h> | |
69 | #include <linux/module.h> | |
70 | #include <linux/sysctl.h> | |
71 | #include <net/tcp.h> | |
72 | #include <net/inet_common.h> | |
73 | #include <linux/ipsec.h> | |
74 | #include <asm/unaligned.h> | |
75 | ||
76 | int sysctl_tcp_timestamps = 1; | |
77 | int sysctl_tcp_window_scaling = 1; | |
78 | int sysctl_tcp_sack = 1; | |
79 | int sysctl_tcp_fack = 1; | |
80 | int sysctl_tcp_reordering = TCP_FASTRETRANS_THRESH; | |
81 | int sysctl_tcp_ecn; | |
82 | int sysctl_tcp_dsack = 1; | |
83 | int sysctl_tcp_app_win = 31; | |
84 | int sysctl_tcp_adv_win_scale = 2; | |
85 | ||
86 | int sysctl_tcp_stdurg; | |
87 | int sysctl_tcp_rfc1337; | |
88 | int sysctl_tcp_max_orphans = NR_FILE; | |
89 | int sysctl_tcp_frto; | |
90 | int sysctl_tcp_nometrics_save; | |
91 | int sysctl_tcp_westwood; | |
92 | int sysctl_tcp_vegas_cong_avoid; | |
93 | ||
94 | int sysctl_tcp_moderate_rcvbuf = 1; | |
95 | ||
96 | /* Default values of the Vegas variables, in fixed-point representation | |
97 | * with V_PARAM_SHIFT bits to the right of the binary point. | |
98 | */ | |
99 | #define V_PARAM_SHIFT 1 | |
100 | int sysctl_tcp_vegas_alpha = 1<<V_PARAM_SHIFT; | |
101 | int sysctl_tcp_vegas_beta = 3<<V_PARAM_SHIFT; | |
102 | int sysctl_tcp_vegas_gamma = 1<<V_PARAM_SHIFT; | |
103 | int sysctl_tcp_bic = 1; | |
104 | int sysctl_tcp_bic_fast_convergence = 1; | |
105 | int sysctl_tcp_bic_low_window = 14; | |
106 | int sysctl_tcp_bic_beta = 819; /* = 819/1024 (BICTCP_BETA_SCALE) */ | |
107 | ||
108 | #define FLAG_DATA 0x01 /* Incoming frame contained data. */ | |
109 | #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ | |
110 | #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ | |
111 | #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ | |
112 | #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ | |
113 | #define FLAG_DATA_SACKED 0x20 /* New SACK. */ | |
114 | #define FLAG_ECE 0x40 /* ECE in this ACK */ | |
115 | #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */ | |
116 | #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ | |
117 | ||
118 | #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) | |
119 | #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) | |
120 | #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) | |
121 | #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) | |
122 | ||
123 | #define IsReno(tp) ((tp)->rx_opt.sack_ok == 0) | |
124 | #define IsFack(tp) ((tp)->rx_opt.sack_ok & 2) | |
125 | #define IsDSack(tp) ((tp)->rx_opt.sack_ok & 4) | |
126 | ||
127 | #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) | |
128 | ||
129 | /* Adapt the MSS value used to make delayed ack decision to the | |
130 | * real world. | |
131 | */ | |
132 | static inline void tcp_measure_rcv_mss(struct tcp_sock *tp, | |
133 | struct sk_buff *skb) | |
134 | { | |
135 | unsigned int len, lss; | |
136 | ||
137 | lss = tp->ack.last_seg_size; | |
138 | tp->ack.last_seg_size = 0; | |
139 | ||
140 | /* skb->len may jitter because of SACKs, even if peer | |
141 | * sends good full-sized frames. | |
142 | */ | |
143 | len = skb->len; | |
144 | if (len >= tp->ack.rcv_mss) { | |
145 | tp->ack.rcv_mss = len; | |
146 | } else { | |
147 | /* Otherwise, we make more careful check taking into account, | |
148 | * that SACKs block is variable. | |
149 | * | |
150 | * "len" is invariant segment length, including TCP header. | |
151 | */ | |
152 | len += skb->data - skb->h.raw; | |
153 | if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) || | |
154 | /* If PSH is not set, packet should be | |
155 | * full sized, provided peer TCP is not badly broken. | |
156 | * This observation (if it is correct 8)) allows | |
157 | * to handle super-low mtu links fairly. | |
158 | */ | |
159 | (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && | |
160 | !(tcp_flag_word(skb->h.th)&TCP_REMNANT))) { | |
161 | /* Subtract also invariant (if peer is RFC compliant), | |
162 | * tcp header plus fixed timestamp option length. | |
163 | * Resulting "len" is MSS free of SACK jitter. | |
164 | */ | |
165 | len -= tp->tcp_header_len; | |
166 | tp->ack.last_seg_size = len; | |
167 | if (len == lss) { | |
168 | tp->ack.rcv_mss = len; | |
169 | return; | |
170 | } | |
171 | } | |
172 | tp->ack.pending |= TCP_ACK_PUSHED; | |
173 | } | |
174 | } | |
175 | ||
176 | static void tcp_incr_quickack(struct tcp_sock *tp) | |
177 | { | |
178 | unsigned quickacks = tp->rcv_wnd/(2*tp->ack.rcv_mss); | |
179 | ||
180 | if (quickacks==0) | |
181 | quickacks=2; | |
182 | if (quickacks > tp->ack.quick) | |
183 | tp->ack.quick = min(quickacks, TCP_MAX_QUICKACKS); | |
184 | } | |
185 | ||
186 | void tcp_enter_quickack_mode(struct tcp_sock *tp) | |
187 | { | |
188 | tcp_incr_quickack(tp); | |
189 | tp->ack.pingpong = 0; | |
190 | tp->ack.ato = TCP_ATO_MIN; | |
191 | } | |
192 | ||
193 | /* Send ACKs quickly, if "quick" count is not exhausted | |
194 | * and the session is not interactive. | |
195 | */ | |
196 | ||
197 | static __inline__ int tcp_in_quickack_mode(struct tcp_sock *tp) | |
198 | { | |
199 | return (tp->ack.quick && !tp->ack.pingpong); | |
200 | } | |
201 | ||
202 | /* Buffer size and advertised window tuning. | |
203 | * | |
204 | * 1. Tuning sk->sk_sndbuf, when connection enters established state. | |
205 | */ | |
206 | ||
207 | static void tcp_fixup_sndbuf(struct sock *sk) | |
208 | { | |
209 | int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 + | |
210 | sizeof(struct sk_buff); | |
211 | ||
212 | if (sk->sk_sndbuf < 3 * sndmem) | |
213 | sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]); | |
214 | } | |
215 | ||
216 | /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) | |
217 | * | |
218 | * All tcp_full_space() is split to two parts: "network" buffer, allocated | |
219 | * forward and advertised in receiver window (tp->rcv_wnd) and | |
220 | * "application buffer", required to isolate scheduling/application | |
221 | * latencies from network. | |
222 | * window_clamp is maximal advertised window. It can be less than | |
223 | * tcp_full_space(), in this case tcp_full_space() - window_clamp | |
224 | * is reserved for "application" buffer. The less window_clamp is | |
225 | * the smoother our behaviour from viewpoint of network, but the lower | |
226 | * throughput and the higher sensitivity of the connection to losses. 8) | |
227 | * | |
228 | * rcv_ssthresh is more strict window_clamp used at "slow start" | |
229 | * phase to predict further behaviour of this connection. | |
230 | * It is used for two goals: | |
231 | * - to enforce header prediction at sender, even when application | |
232 | * requires some significant "application buffer". It is check #1. | |
233 | * - to prevent pruning of receive queue because of misprediction | |
234 | * of receiver window. Check #2. | |
235 | * | |
236 | * The scheme does not work when sender sends good segments opening | |
237 | * window and then starts to feed us spagetti. But it should work | |
238 | * in common situations. Otherwise, we have to rely on queue collapsing. | |
239 | */ | |
240 | ||
241 | /* Slow part of check#2. */ | |
242 | static int __tcp_grow_window(struct sock *sk, struct tcp_sock *tp, | |
243 | struct sk_buff *skb) | |
244 | { | |
245 | /* Optimize this! */ | |
246 | int truesize = tcp_win_from_space(skb->truesize)/2; | |
247 | int window = tcp_full_space(sk)/2; | |
248 | ||
249 | while (tp->rcv_ssthresh <= window) { | |
250 | if (truesize <= skb->len) | |
251 | return 2*tp->ack.rcv_mss; | |
252 | ||
253 | truesize >>= 1; | |
254 | window >>= 1; | |
255 | } | |
256 | return 0; | |
257 | } | |
258 | ||
259 | static inline void tcp_grow_window(struct sock *sk, struct tcp_sock *tp, | |
260 | struct sk_buff *skb) | |
261 | { | |
262 | /* Check #1 */ | |
263 | if (tp->rcv_ssthresh < tp->window_clamp && | |
264 | (int)tp->rcv_ssthresh < tcp_space(sk) && | |
265 | !tcp_memory_pressure) { | |
266 | int incr; | |
267 | ||
268 | /* Check #2. Increase window, if skb with such overhead | |
269 | * will fit to rcvbuf in future. | |
270 | */ | |
271 | if (tcp_win_from_space(skb->truesize) <= skb->len) | |
272 | incr = 2*tp->advmss; | |
273 | else | |
274 | incr = __tcp_grow_window(sk, tp, skb); | |
275 | ||
276 | if (incr) { | |
277 | tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, tp->window_clamp); | |
278 | tp->ack.quick |= 1; | |
279 | } | |
280 | } | |
281 | } | |
282 | ||
283 | /* 3. Tuning rcvbuf, when connection enters established state. */ | |
284 | ||
285 | static void tcp_fixup_rcvbuf(struct sock *sk) | |
286 | { | |
287 | struct tcp_sock *tp = tcp_sk(sk); | |
288 | int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); | |
289 | ||
290 | /* Try to select rcvbuf so that 4 mss-sized segments | |
291 | * will fit to window and correspoding skbs will fit to our rcvbuf. | |
292 | * (was 3; 4 is minimum to allow fast retransmit to work.) | |
293 | */ | |
294 | while (tcp_win_from_space(rcvmem) < tp->advmss) | |
295 | rcvmem += 128; | |
296 | if (sk->sk_rcvbuf < 4 * rcvmem) | |
297 | sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]); | |
298 | } | |
299 | ||
300 | /* 4. Try to fixup all. It is made iimediately after connection enters | |
301 | * established state. | |
302 | */ | |
303 | static void tcp_init_buffer_space(struct sock *sk) | |
304 | { | |
305 | struct tcp_sock *tp = tcp_sk(sk); | |
306 | int maxwin; | |
307 | ||
308 | if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) | |
309 | tcp_fixup_rcvbuf(sk); | |
310 | if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) | |
311 | tcp_fixup_sndbuf(sk); | |
312 | ||
313 | tp->rcvq_space.space = tp->rcv_wnd; | |
314 | ||
315 | maxwin = tcp_full_space(sk); | |
316 | ||
317 | if (tp->window_clamp >= maxwin) { | |
318 | tp->window_clamp = maxwin; | |
319 | ||
320 | if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) | |
321 | tp->window_clamp = max(maxwin - | |
322 | (maxwin >> sysctl_tcp_app_win), | |
323 | 4 * tp->advmss); | |
324 | } | |
325 | ||
326 | /* Force reservation of one segment. */ | |
327 | if (sysctl_tcp_app_win && | |
328 | tp->window_clamp > 2 * tp->advmss && | |
329 | tp->window_clamp + tp->advmss > maxwin) | |
330 | tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); | |
331 | ||
332 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); | |
333 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
334 | } | |
335 | ||
336 | static void init_bictcp(struct tcp_sock *tp) | |
337 | { | |
338 | tp->bictcp.cnt = 0; | |
339 | ||
340 | tp->bictcp.last_max_cwnd = 0; | |
341 | tp->bictcp.last_cwnd = 0; | |
342 | tp->bictcp.last_stamp = 0; | |
343 | } | |
344 | ||
345 | /* 5. Recalculate window clamp after socket hit its memory bounds. */ | |
346 | static void tcp_clamp_window(struct sock *sk, struct tcp_sock *tp) | |
347 | { | |
348 | struct sk_buff *skb; | |
349 | unsigned int app_win = tp->rcv_nxt - tp->copied_seq; | |
350 | int ofo_win = 0; | |
351 | ||
352 | tp->ack.quick = 0; | |
353 | ||
354 | skb_queue_walk(&tp->out_of_order_queue, skb) { | |
355 | ofo_win += skb->len; | |
356 | } | |
357 | ||
358 | /* If overcommit is due to out of order segments, | |
359 | * do not clamp window. Try to expand rcvbuf instead. | |
360 | */ | |
361 | if (ofo_win) { | |
362 | if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && | |
363 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && | |
364 | !tcp_memory_pressure && | |
365 | atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) | |
366 | sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), | |
367 | sysctl_tcp_rmem[2]); | |
368 | } | |
369 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) { | |
370 | app_win += ofo_win; | |
371 | if (atomic_read(&sk->sk_rmem_alloc) >= 2 * sk->sk_rcvbuf) | |
372 | app_win >>= 1; | |
373 | if (app_win > tp->ack.rcv_mss) | |
374 | app_win -= tp->ack.rcv_mss; | |
375 | app_win = max(app_win, 2U*tp->advmss); | |
376 | ||
377 | if (!ofo_win) | |
378 | tp->window_clamp = min(tp->window_clamp, app_win); | |
379 | tp->rcv_ssthresh = min(tp->window_clamp, 2U*tp->advmss); | |
380 | } | |
381 | } | |
382 | ||
383 | /* Receiver "autotuning" code. | |
384 | * | |
385 | * The algorithm for RTT estimation w/o timestamps is based on | |
386 | * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. | |
387 | * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps> | |
388 | * | |
389 | * More detail on this code can be found at | |
390 | * <http://www.psc.edu/~jheffner/senior_thesis.ps>, | |
391 | * though this reference is out of date. A new paper | |
392 | * is pending. | |
393 | */ | |
394 | static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) | |
395 | { | |
396 | u32 new_sample = tp->rcv_rtt_est.rtt; | |
397 | long m = sample; | |
398 | ||
399 | if (m == 0) | |
400 | m = 1; | |
401 | ||
402 | if (new_sample != 0) { | |
403 | /* If we sample in larger samples in the non-timestamp | |
404 | * case, we could grossly overestimate the RTT especially | |
405 | * with chatty applications or bulk transfer apps which | |
406 | * are stalled on filesystem I/O. | |
407 | * | |
408 | * Also, since we are only going for a minimum in the | |
409 | * non-timestamp case, we do not smoothe things out | |
410 | * else with timestamps disabled convergance takes too | |
411 | * long. | |
412 | */ | |
413 | if (!win_dep) { | |
414 | m -= (new_sample >> 3); | |
415 | new_sample += m; | |
416 | } else if (m < new_sample) | |
417 | new_sample = m << 3; | |
418 | } else { | |
419 | /* No previous mesaure. */ | |
420 | new_sample = m << 3; | |
421 | } | |
422 | ||
423 | if (tp->rcv_rtt_est.rtt != new_sample) | |
424 | tp->rcv_rtt_est.rtt = new_sample; | |
425 | } | |
426 | ||
427 | static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) | |
428 | { | |
429 | if (tp->rcv_rtt_est.time == 0) | |
430 | goto new_measure; | |
431 | if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) | |
432 | return; | |
433 | tcp_rcv_rtt_update(tp, | |
434 | jiffies - tp->rcv_rtt_est.time, | |
435 | 1); | |
436 | ||
437 | new_measure: | |
438 | tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; | |
439 | tp->rcv_rtt_est.time = tcp_time_stamp; | |
440 | } | |
441 | ||
442 | static inline void tcp_rcv_rtt_measure_ts(struct tcp_sock *tp, struct sk_buff *skb) | |
443 | { | |
444 | if (tp->rx_opt.rcv_tsecr && | |
445 | (TCP_SKB_CB(skb)->end_seq - | |
446 | TCP_SKB_CB(skb)->seq >= tp->ack.rcv_mss)) | |
447 | tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); | |
448 | } | |
449 | ||
450 | /* | |
451 | * This function should be called every time data is copied to user space. | |
452 | * It calculates the appropriate TCP receive buffer space. | |
453 | */ | |
454 | void tcp_rcv_space_adjust(struct sock *sk) | |
455 | { | |
456 | struct tcp_sock *tp = tcp_sk(sk); | |
457 | int time; | |
458 | int space; | |
459 | ||
460 | if (tp->rcvq_space.time == 0) | |
461 | goto new_measure; | |
462 | ||
463 | time = tcp_time_stamp - tp->rcvq_space.time; | |
464 | if (time < (tp->rcv_rtt_est.rtt >> 3) || | |
465 | tp->rcv_rtt_est.rtt == 0) | |
466 | return; | |
467 | ||
468 | space = 2 * (tp->copied_seq - tp->rcvq_space.seq); | |
469 | ||
470 | space = max(tp->rcvq_space.space, space); | |
471 | ||
472 | if (tp->rcvq_space.space != space) { | |
473 | int rcvmem; | |
474 | ||
475 | tp->rcvq_space.space = space; | |
476 | ||
477 | if (sysctl_tcp_moderate_rcvbuf) { | |
478 | int new_clamp = space; | |
479 | ||
480 | /* Receive space grows, normalize in order to | |
481 | * take into account packet headers and sk_buff | |
482 | * structure overhead. | |
483 | */ | |
484 | space /= tp->advmss; | |
485 | if (!space) | |
486 | space = 1; | |
487 | rcvmem = (tp->advmss + MAX_TCP_HEADER + | |
488 | 16 + sizeof(struct sk_buff)); | |
489 | while (tcp_win_from_space(rcvmem) < tp->advmss) | |
490 | rcvmem += 128; | |
491 | space *= rcvmem; | |
492 | space = min(space, sysctl_tcp_rmem[2]); | |
493 | if (space > sk->sk_rcvbuf) { | |
494 | sk->sk_rcvbuf = space; | |
495 | ||
496 | /* Make the window clamp follow along. */ | |
497 | tp->window_clamp = new_clamp; | |
498 | } | |
499 | } | |
500 | } | |
501 | ||
502 | new_measure: | |
503 | tp->rcvq_space.seq = tp->copied_seq; | |
504 | tp->rcvq_space.time = tcp_time_stamp; | |
505 | } | |
506 | ||
507 | /* There is something which you must keep in mind when you analyze the | |
508 | * behavior of the tp->ato delayed ack timeout interval. When a | |
509 | * connection starts up, we want to ack as quickly as possible. The | |
510 | * problem is that "good" TCP's do slow start at the beginning of data | |
511 | * transmission. The means that until we send the first few ACK's the | |
512 | * sender will sit on his end and only queue most of his data, because | |
513 | * he can only send snd_cwnd unacked packets at any given time. For | |
514 | * each ACK we send, he increments snd_cwnd and transmits more of his | |
515 | * queue. -DaveM | |
516 | */ | |
517 | static void tcp_event_data_recv(struct sock *sk, struct tcp_sock *tp, struct sk_buff *skb) | |
518 | { | |
519 | u32 now; | |
520 | ||
521 | tcp_schedule_ack(tp); | |
522 | ||
523 | tcp_measure_rcv_mss(tp, skb); | |
524 | ||
525 | tcp_rcv_rtt_measure(tp); | |
526 | ||
527 | now = tcp_time_stamp; | |
528 | ||
529 | if (!tp->ack.ato) { | |
530 | /* The _first_ data packet received, initialize | |
531 | * delayed ACK engine. | |
532 | */ | |
533 | tcp_incr_quickack(tp); | |
534 | tp->ack.ato = TCP_ATO_MIN; | |
535 | } else { | |
536 | int m = now - tp->ack.lrcvtime; | |
537 | ||
538 | if (m <= TCP_ATO_MIN/2) { | |
539 | /* The fastest case is the first. */ | |
540 | tp->ack.ato = (tp->ack.ato>>1) + TCP_ATO_MIN/2; | |
541 | } else if (m < tp->ack.ato) { | |
542 | tp->ack.ato = (tp->ack.ato>>1) + m; | |
543 | if (tp->ack.ato > tp->rto) | |
544 | tp->ack.ato = tp->rto; | |
545 | } else if (m > tp->rto) { | |
546 | /* Too long gap. Apparently sender falled to | |
547 | * restart window, so that we send ACKs quickly. | |
548 | */ | |
549 | tcp_incr_quickack(tp); | |
550 | sk_stream_mem_reclaim(sk); | |
551 | } | |
552 | } | |
553 | tp->ack.lrcvtime = now; | |
554 | ||
555 | TCP_ECN_check_ce(tp, skb); | |
556 | ||
557 | if (skb->len >= 128) | |
558 | tcp_grow_window(sk, tp, skb); | |
559 | } | |
560 | ||
561 | /* When starting a new connection, pin down the current choice of | |
562 | * congestion algorithm. | |
563 | */ | |
564 | void tcp_ca_init(struct tcp_sock *tp) | |
565 | { | |
566 | if (sysctl_tcp_westwood) | |
567 | tp->adv_cong = TCP_WESTWOOD; | |
568 | else if (sysctl_tcp_bic) | |
569 | tp->adv_cong = TCP_BIC; | |
570 | else if (sysctl_tcp_vegas_cong_avoid) { | |
571 | tp->adv_cong = TCP_VEGAS; | |
572 | tp->vegas.baseRTT = 0x7fffffff; | |
573 | tcp_vegas_enable(tp); | |
574 | } | |
575 | } | |
576 | ||
577 | /* Do RTT sampling needed for Vegas. | |
578 | * Basically we: | |
579 | * o min-filter RTT samples from within an RTT to get the current | |
580 | * propagation delay + queuing delay (we are min-filtering to try to | |
581 | * avoid the effects of delayed ACKs) | |
582 | * o min-filter RTT samples from a much longer window (forever for now) | |
583 | * to find the propagation delay (baseRTT) | |
584 | */ | |
585 | static inline void vegas_rtt_calc(struct tcp_sock *tp, __u32 rtt) | |
586 | { | |
587 | __u32 vrtt = rtt + 1; /* Never allow zero rtt or baseRTT */ | |
588 | ||
589 | /* Filter to find propagation delay: */ | |
590 | if (vrtt < tp->vegas.baseRTT) | |
591 | tp->vegas.baseRTT = vrtt; | |
592 | ||
593 | /* Find the min RTT during the last RTT to find | |
594 | * the current prop. delay + queuing delay: | |
595 | */ | |
596 | tp->vegas.minRTT = min(tp->vegas.minRTT, vrtt); | |
597 | tp->vegas.cntRTT++; | |
598 | } | |
599 | ||
600 | /* Called to compute a smoothed rtt estimate. The data fed to this | |
601 | * routine either comes from timestamps, or from segments that were | |
602 | * known _not_ to have been retransmitted [see Karn/Partridge | |
603 | * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 | |
604 | * piece by Van Jacobson. | |
605 | * NOTE: the next three routines used to be one big routine. | |
606 | * To save cycles in the RFC 1323 implementation it was better to break | |
607 | * it up into three procedures. -- erics | |
608 | */ | |
609 | static void tcp_rtt_estimator(struct tcp_sock *tp, __u32 mrtt) | |
610 | { | |
611 | long m = mrtt; /* RTT */ | |
612 | ||
613 | if (tcp_vegas_enabled(tp)) | |
614 | vegas_rtt_calc(tp, mrtt); | |
615 | ||
616 | /* The following amusing code comes from Jacobson's | |
617 | * article in SIGCOMM '88. Note that rtt and mdev | |
618 | * are scaled versions of rtt and mean deviation. | |
619 | * This is designed to be as fast as possible | |
620 | * m stands for "measurement". | |
621 | * | |
622 | * On a 1990 paper the rto value is changed to: | |
623 | * RTO = rtt + 4 * mdev | |
624 | * | |
625 | * Funny. This algorithm seems to be very broken. | |
626 | * These formulae increase RTO, when it should be decreased, increase | |
627 | * too slowly, when it should be incresed fastly, decrease too fastly | |
628 | * etc. I guess in BSD RTO takes ONE value, so that it is absolutely | |
629 | * does not matter how to _calculate_ it. Seems, it was trap | |
630 | * that VJ failed to avoid. 8) | |
631 | */ | |
632 | if(m == 0) | |
633 | m = 1; | |
634 | if (tp->srtt != 0) { | |
635 | m -= (tp->srtt >> 3); /* m is now error in rtt est */ | |
636 | tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ | |
637 | if (m < 0) { | |
638 | m = -m; /* m is now abs(error) */ | |
639 | m -= (tp->mdev >> 2); /* similar update on mdev */ | |
640 | /* This is similar to one of Eifel findings. | |
641 | * Eifel blocks mdev updates when rtt decreases. | |
642 | * This solution is a bit different: we use finer gain | |
643 | * for mdev in this case (alpha*beta). | |
644 | * Like Eifel it also prevents growth of rto, | |
645 | * but also it limits too fast rto decreases, | |
646 | * happening in pure Eifel. | |
647 | */ | |
648 | if (m > 0) | |
649 | m >>= 3; | |
650 | } else { | |
651 | m -= (tp->mdev >> 2); /* similar update on mdev */ | |
652 | } | |
653 | tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ | |
654 | if (tp->mdev > tp->mdev_max) { | |
655 | tp->mdev_max = tp->mdev; | |
656 | if (tp->mdev_max > tp->rttvar) | |
657 | tp->rttvar = tp->mdev_max; | |
658 | } | |
659 | if (after(tp->snd_una, tp->rtt_seq)) { | |
660 | if (tp->mdev_max < tp->rttvar) | |
661 | tp->rttvar -= (tp->rttvar-tp->mdev_max)>>2; | |
662 | tp->rtt_seq = tp->snd_nxt; | |
663 | tp->mdev_max = TCP_RTO_MIN; | |
664 | } | |
665 | } else { | |
666 | /* no previous measure. */ | |
667 | tp->srtt = m<<3; /* take the measured time to be rtt */ | |
668 | tp->mdev = m<<1; /* make sure rto = 3*rtt */ | |
669 | tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN); | |
670 | tp->rtt_seq = tp->snd_nxt; | |
671 | } | |
672 | ||
673 | tcp_westwood_update_rtt(tp, tp->srtt >> 3); | |
674 | } | |
675 | ||
676 | /* Calculate rto without backoff. This is the second half of Van Jacobson's | |
677 | * routine referred to above. | |
678 | */ | |
679 | static inline void tcp_set_rto(struct tcp_sock *tp) | |
680 | { | |
681 | /* Old crap is replaced with new one. 8) | |
682 | * | |
683 | * More seriously: | |
684 | * 1. If rtt variance happened to be less 50msec, it is hallucination. | |
685 | * It cannot be less due to utterly erratic ACK generation made | |
686 | * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ | |
687 | * to do with delayed acks, because at cwnd>2 true delack timeout | |
688 | * is invisible. Actually, Linux-2.4 also generates erratic | |
689 | * ACKs in some curcumstances. | |
690 | */ | |
691 | tp->rto = (tp->srtt >> 3) + tp->rttvar; | |
692 | ||
693 | /* 2. Fixups made earlier cannot be right. | |
694 | * If we do not estimate RTO correctly without them, | |
695 | * all the algo is pure shit and should be replaced | |
696 | * with correct one. It is exaclty, which we pretend to do. | |
697 | */ | |
698 | } | |
699 | ||
700 | /* NOTE: clamping at TCP_RTO_MIN is not required, current algo | |
701 | * guarantees that rto is higher. | |
702 | */ | |
703 | static inline void tcp_bound_rto(struct tcp_sock *tp) | |
704 | { | |
705 | if (tp->rto > TCP_RTO_MAX) | |
706 | tp->rto = TCP_RTO_MAX; | |
707 | } | |
708 | ||
709 | /* Save metrics learned by this TCP session. | |
710 | This function is called only, when TCP finishes successfully | |
711 | i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. | |
712 | */ | |
713 | void tcp_update_metrics(struct sock *sk) | |
714 | { | |
715 | struct tcp_sock *tp = tcp_sk(sk); | |
716 | struct dst_entry *dst = __sk_dst_get(sk); | |
717 | ||
718 | if (sysctl_tcp_nometrics_save) | |
719 | return; | |
720 | ||
721 | dst_confirm(dst); | |
722 | ||
723 | if (dst && (dst->flags&DST_HOST)) { | |
724 | int m; | |
725 | ||
726 | if (tp->backoff || !tp->srtt) { | |
727 | /* This session failed to estimate rtt. Why? | |
728 | * Probably, no packets returned in time. | |
729 | * Reset our results. | |
730 | */ | |
731 | if (!(dst_metric_locked(dst, RTAX_RTT))) | |
732 | dst->metrics[RTAX_RTT-1] = 0; | |
733 | return; | |
734 | } | |
735 | ||
736 | m = dst_metric(dst, RTAX_RTT) - tp->srtt; | |
737 | ||
738 | /* If newly calculated rtt larger than stored one, | |
739 | * store new one. Otherwise, use EWMA. Remember, | |
740 | * rtt overestimation is always better than underestimation. | |
741 | */ | |
742 | if (!(dst_metric_locked(dst, RTAX_RTT))) { | |
743 | if (m <= 0) | |
744 | dst->metrics[RTAX_RTT-1] = tp->srtt; | |
745 | else | |
746 | dst->metrics[RTAX_RTT-1] -= (m>>3); | |
747 | } | |
748 | ||
749 | if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { | |
750 | if (m < 0) | |
751 | m = -m; | |
752 | ||
753 | /* Scale deviation to rttvar fixed point */ | |
754 | m >>= 1; | |
755 | if (m < tp->mdev) | |
756 | m = tp->mdev; | |
757 | ||
758 | if (m >= dst_metric(dst, RTAX_RTTVAR)) | |
759 | dst->metrics[RTAX_RTTVAR-1] = m; | |
760 | else | |
761 | dst->metrics[RTAX_RTTVAR-1] -= | |
762 | (dst->metrics[RTAX_RTTVAR-1] - m)>>2; | |
763 | } | |
764 | ||
765 | if (tp->snd_ssthresh >= 0xFFFF) { | |
766 | /* Slow start still did not finish. */ | |
767 | if (dst_metric(dst, RTAX_SSTHRESH) && | |
768 | !dst_metric_locked(dst, RTAX_SSTHRESH) && | |
769 | (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) | |
770 | dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1; | |
771 | if (!dst_metric_locked(dst, RTAX_CWND) && | |
772 | tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) | |
773 | dst->metrics[RTAX_CWND-1] = tp->snd_cwnd; | |
774 | } else if (tp->snd_cwnd > tp->snd_ssthresh && | |
775 | tp->ca_state == TCP_CA_Open) { | |
776 | /* Cong. avoidance phase, cwnd is reliable. */ | |
777 | if (!dst_metric_locked(dst, RTAX_SSTHRESH)) | |
778 | dst->metrics[RTAX_SSTHRESH-1] = | |
779 | max(tp->snd_cwnd >> 1, tp->snd_ssthresh); | |
780 | if (!dst_metric_locked(dst, RTAX_CWND)) | |
781 | dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1; | |
782 | } else { | |
783 | /* Else slow start did not finish, cwnd is non-sense, | |
784 | ssthresh may be also invalid. | |
785 | */ | |
786 | if (!dst_metric_locked(dst, RTAX_CWND)) | |
787 | dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1; | |
788 | if (dst->metrics[RTAX_SSTHRESH-1] && | |
789 | !dst_metric_locked(dst, RTAX_SSTHRESH) && | |
790 | tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1]) | |
791 | dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh; | |
792 | } | |
793 | ||
794 | if (!dst_metric_locked(dst, RTAX_REORDERING)) { | |
795 | if (dst->metrics[RTAX_REORDERING-1] < tp->reordering && | |
796 | tp->reordering != sysctl_tcp_reordering) | |
797 | dst->metrics[RTAX_REORDERING-1] = tp->reordering; | |
798 | } | |
799 | } | |
800 | } | |
801 | ||
802 | /* Numbers are taken from RFC2414. */ | |
803 | __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst) | |
804 | { | |
805 | __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); | |
806 | ||
807 | if (!cwnd) { | |
808 | if (tp->mss_cache_std > 1460) | |
809 | cwnd = 2; | |
810 | else | |
811 | cwnd = (tp->mss_cache_std > 1095) ? 3 : 4; | |
812 | } | |
813 | return min_t(__u32, cwnd, tp->snd_cwnd_clamp); | |
814 | } | |
815 | ||
816 | /* Initialize metrics on socket. */ | |
817 | ||
818 | static void tcp_init_metrics(struct sock *sk) | |
819 | { | |
820 | struct tcp_sock *tp = tcp_sk(sk); | |
821 | struct dst_entry *dst = __sk_dst_get(sk); | |
822 | ||
823 | if (dst == NULL) | |
824 | goto reset; | |
825 | ||
826 | dst_confirm(dst); | |
827 | ||
828 | if (dst_metric_locked(dst, RTAX_CWND)) | |
829 | tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); | |
830 | if (dst_metric(dst, RTAX_SSTHRESH)) { | |
831 | tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); | |
832 | if (tp->snd_ssthresh > tp->snd_cwnd_clamp) | |
833 | tp->snd_ssthresh = tp->snd_cwnd_clamp; | |
834 | } | |
835 | if (dst_metric(dst, RTAX_REORDERING) && | |
836 | tp->reordering != dst_metric(dst, RTAX_REORDERING)) { | |
837 | tp->rx_opt.sack_ok &= ~2; | |
838 | tp->reordering = dst_metric(dst, RTAX_REORDERING); | |
839 | } | |
840 | ||
841 | if (dst_metric(dst, RTAX_RTT) == 0) | |
842 | goto reset; | |
843 | ||
844 | if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3)) | |
845 | goto reset; | |
846 | ||
847 | /* Initial rtt is determined from SYN,SYN-ACK. | |
848 | * The segment is small and rtt may appear much | |
849 | * less than real one. Use per-dst memory | |
850 | * to make it more realistic. | |
851 | * | |
852 | * A bit of theory. RTT is time passed after "normal" sized packet | |
853 | * is sent until it is ACKed. In normal curcumstances sending small | |
854 | * packets force peer to delay ACKs and calculation is correct too. | |
855 | * The algorithm is adaptive and, provided we follow specs, it | |
856 | * NEVER underestimate RTT. BUT! If peer tries to make some clever | |
857 | * tricks sort of "quick acks" for time long enough to decrease RTT | |
858 | * to low value, and then abruptly stops to do it and starts to delay | |
859 | * ACKs, wait for troubles. | |
860 | */ | |
861 | if (dst_metric(dst, RTAX_RTT) > tp->srtt) { | |
862 | tp->srtt = dst_metric(dst, RTAX_RTT); | |
863 | tp->rtt_seq = tp->snd_nxt; | |
864 | } | |
865 | if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) { | |
866 | tp->mdev = dst_metric(dst, RTAX_RTTVAR); | |
867 | tp->mdev_max = tp->rttvar = max(tp->mdev, TCP_RTO_MIN); | |
868 | } | |
869 | tcp_set_rto(tp); | |
870 | tcp_bound_rto(tp); | |
871 | if (tp->rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp) | |
872 | goto reset; | |
873 | tp->snd_cwnd = tcp_init_cwnd(tp, dst); | |
874 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
875 | return; | |
876 | ||
877 | reset: | |
878 | /* Play conservative. If timestamps are not | |
879 | * supported, TCP will fail to recalculate correct | |
880 | * rtt, if initial rto is too small. FORGET ALL AND RESET! | |
881 | */ | |
882 | if (!tp->rx_opt.saw_tstamp && tp->srtt) { | |
883 | tp->srtt = 0; | |
884 | tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT; | |
885 | tp->rto = TCP_TIMEOUT_INIT; | |
886 | } | |
887 | } | |
888 | ||
889 | static void tcp_update_reordering(struct tcp_sock *tp, int metric, int ts) | |
890 | { | |
891 | if (metric > tp->reordering) { | |
892 | tp->reordering = min(TCP_MAX_REORDERING, metric); | |
893 | ||
894 | /* This exciting event is worth to be remembered. 8) */ | |
895 | if (ts) | |
896 | NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER); | |
897 | else if (IsReno(tp)) | |
898 | NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER); | |
899 | else if (IsFack(tp)) | |
900 | NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER); | |
901 | else | |
902 | NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER); | |
903 | #if FASTRETRANS_DEBUG > 1 | |
904 | printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", | |
905 | tp->rx_opt.sack_ok, tp->ca_state, | |
906 | tp->reordering, | |
907 | tp->fackets_out, | |
908 | tp->sacked_out, | |
909 | tp->undo_marker ? tp->undo_retrans : 0); | |
910 | #endif | |
911 | /* Disable FACK yet. */ | |
912 | tp->rx_opt.sack_ok &= ~2; | |
913 | } | |
914 | } | |
915 | ||
916 | /* This procedure tags the retransmission queue when SACKs arrive. | |
917 | * | |
918 | * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). | |
919 | * Packets in queue with these bits set are counted in variables | |
920 | * sacked_out, retrans_out and lost_out, correspondingly. | |
921 | * | |
922 | * Valid combinations are: | |
923 | * Tag InFlight Description | |
924 | * 0 1 - orig segment is in flight. | |
925 | * S 0 - nothing flies, orig reached receiver. | |
926 | * L 0 - nothing flies, orig lost by net. | |
927 | * R 2 - both orig and retransmit are in flight. | |
928 | * L|R 1 - orig is lost, retransmit is in flight. | |
929 | * S|R 1 - orig reached receiver, retrans is still in flight. | |
930 | * (L|S|R is logically valid, it could occur when L|R is sacked, | |
931 | * but it is equivalent to plain S and code short-curcuits it to S. | |
932 | * L|S is logically invalid, it would mean -1 packet in flight 8)) | |
933 | * | |
934 | * These 6 states form finite state machine, controlled by the following events: | |
935 | * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) | |
936 | * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) | |
937 | * 3. Loss detection event of one of three flavors: | |
938 | * A. Scoreboard estimator decided the packet is lost. | |
939 | * A'. Reno "three dupacks" marks head of queue lost. | |
940 | * A''. Its FACK modfication, head until snd.fack is lost. | |
941 | * B. SACK arrives sacking data transmitted after never retransmitted | |
942 | * hole was sent out. | |
943 | * C. SACK arrives sacking SND.NXT at the moment, when the | |
944 | * segment was retransmitted. | |
945 | * 4. D-SACK added new rule: D-SACK changes any tag to S. | |
946 | * | |
947 | * It is pleasant to note, that state diagram turns out to be commutative, | |
948 | * so that we are allowed not to be bothered by order of our actions, | |
949 | * when multiple events arrive simultaneously. (see the function below). | |
950 | * | |
951 | * Reordering detection. | |
952 | * -------------------- | |
953 | * Reordering metric is maximal distance, which a packet can be displaced | |
954 | * in packet stream. With SACKs we can estimate it: | |
955 | * | |
956 | * 1. SACK fills old hole and the corresponding segment was not | |
957 | * ever retransmitted -> reordering. Alas, we cannot use it | |
958 | * when segment was retransmitted. | |
959 | * 2. The last flaw is solved with D-SACK. D-SACK arrives | |
960 | * for retransmitted and already SACKed segment -> reordering.. | |
961 | * Both of these heuristics are not used in Loss state, when we cannot | |
962 | * account for retransmits accurately. | |
963 | */ | |
964 | static int | |
965 | tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, u32 prior_snd_una) | |
966 | { | |
967 | struct tcp_sock *tp = tcp_sk(sk); | |
968 | unsigned char *ptr = ack_skb->h.raw + TCP_SKB_CB(ack_skb)->sacked; | |
969 | struct tcp_sack_block *sp = (struct tcp_sack_block *)(ptr+2); | |
970 | int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE)>>3; | |
971 | int reord = tp->packets_out; | |
972 | int prior_fackets; | |
973 | u32 lost_retrans = 0; | |
974 | int flag = 0; | |
975 | int i; | |
976 | ||
977 | /* So, SACKs for already sent large segments will be lost. | |
978 | * Not good, but alternative is to resegment the queue. */ | |
979 | if (sk->sk_route_caps & NETIF_F_TSO) { | |
980 | sk->sk_route_caps &= ~NETIF_F_TSO; | |
981 | sock_set_flag(sk, SOCK_NO_LARGESEND); | |
982 | tp->mss_cache = tp->mss_cache_std; | |
983 | } | |
984 | ||
985 | if (!tp->sacked_out) | |
986 | tp->fackets_out = 0; | |
987 | prior_fackets = tp->fackets_out; | |
988 | ||
989 | for (i=0; i<num_sacks; i++, sp++) { | |
990 | struct sk_buff *skb; | |
991 | __u32 start_seq = ntohl(sp->start_seq); | |
992 | __u32 end_seq = ntohl(sp->end_seq); | |
993 | int fack_count = 0; | |
994 | int dup_sack = 0; | |
995 | ||
996 | /* Check for D-SACK. */ | |
997 | if (i == 0) { | |
998 | u32 ack = TCP_SKB_CB(ack_skb)->ack_seq; | |
999 | ||
1000 | if (before(start_seq, ack)) { | |
1001 | dup_sack = 1; | |
1002 | tp->rx_opt.sack_ok |= 4; | |
1003 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV); | |
1004 | } else if (num_sacks > 1 && | |
1005 | !after(end_seq, ntohl(sp[1].end_seq)) && | |
1006 | !before(start_seq, ntohl(sp[1].start_seq))) { | |
1007 | dup_sack = 1; | |
1008 | tp->rx_opt.sack_ok |= 4; | |
1009 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV); | |
1010 | } | |
1011 | ||
1012 | /* D-SACK for already forgotten data... | |
1013 | * Do dumb counting. */ | |
1014 | if (dup_sack && | |
1015 | !after(end_seq, prior_snd_una) && | |
1016 | after(end_seq, tp->undo_marker)) | |
1017 | tp->undo_retrans--; | |
1018 | ||
1019 | /* Eliminate too old ACKs, but take into | |
1020 | * account more or less fresh ones, they can | |
1021 | * contain valid SACK info. | |
1022 | */ | |
1023 | if (before(ack, prior_snd_una - tp->max_window)) | |
1024 | return 0; | |
1025 | } | |
1026 | ||
1027 | /* Event "B" in the comment above. */ | |
1028 | if (after(end_seq, tp->high_seq)) | |
1029 | flag |= FLAG_DATA_LOST; | |
1030 | ||
1031 | sk_stream_for_retrans_queue(skb, sk) { | |
1032 | u8 sacked = TCP_SKB_CB(skb)->sacked; | |
1033 | int in_sack; | |
1034 | ||
1035 | /* The retransmission queue is always in order, so | |
1036 | * we can short-circuit the walk early. | |
1037 | */ | |
1038 | if(!before(TCP_SKB_CB(skb)->seq, end_seq)) | |
1039 | break; | |
1040 | ||
1041 | fack_count += tcp_skb_pcount(skb); | |
1042 | ||
1043 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && | |
1044 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1045 | ||
1046 | /* Account D-SACK for retransmitted packet. */ | |
1047 | if ((dup_sack && in_sack) && | |
1048 | (sacked & TCPCB_RETRANS) && | |
1049 | after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker)) | |
1050 | tp->undo_retrans--; | |
1051 | ||
1052 | /* The frame is ACKed. */ | |
1053 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) { | |
1054 | if (sacked&TCPCB_RETRANS) { | |
1055 | if ((dup_sack && in_sack) && | |
1056 | (sacked&TCPCB_SACKED_ACKED)) | |
1057 | reord = min(fack_count, reord); | |
1058 | } else { | |
1059 | /* If it was in a hole, we detected reordering. */ | |
1060 | if (fack_count < prior_fackets && | |
1061 | !(sacked&TCPCB_SACKED_ACKED)) | |
1062 | reord = min(fack_count, reord); | |
1063 | } | |
1064 | ||
1065 | /* Nothing to do; acked frame is about to be dropped. */ | |
1066 | continue; | |
1067 | } | |
1068 | ||
1069 | if ((sacked&TCPCB_SACKED_RETRANS) && | |
1070 | after(end_seq, TCP_SKB_CB(skb)->ack_seq) && | |
1071 | (!lost_retrans || after(end_seq, lost_retrans))) | |
1072 | lost_retrans = end_seq; | |
1073 | ||
1074 | if (!in_sack) | |
1075 | continue; | |
1076 | ||
1077 | if (!(sacked&TCPCB_SACKED_ACKED)) { | |
1078 | if (sacked & TCPCB_SACKED_RETRANS) { | |
1079 | /* If the segment is not tagged as lost, | |
1080 | * we do not clear RETRANS, believing | |
1081 | * that retransmission is still in flight. | |
1082 | */ | |
1083 | if (sacked & TCPCB_LOST) { | |
1084 | TCP_SKB_CB(skb)->sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); | |
1085 | tp->lost_out -= tcp_skb_pcount(skb); | |
1086 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1087 | } | |
1088 | } else { | |
1089 | /* New sack for not retransmitted frame, | |
1090 | * which was in hole. It is reordering. | |
1091 | */ | |
1092 | if (!(sacked & TCPCB_RETRANS) && | |
1093 | fack_count < prior_fackets) | |
1094 | reord = min(fack_count, reord); | |
1095 | ||
1096 | if (sacked & TCPCB_LOST) { | |
1097 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; | |
1098 | tp->lost_out -= tcp_skb_pcount(skb); | |
1099 | } | |
1100 | } | |
1101 | ||
1102 | TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED; | |
1103 | flag |= FLAG_DATA_SACKED; | |
1104 | tp->sacked_out += tcp_skb_pcount(skb); | |
1105 | ||
1106 | if (fack_count > tp->fackets_out) | |
1107 | tp->fackets_out = fack_count; | |
1108 | } else { | |
1109 | if (dup_sack && (sacked&TCPCB_RETRANS)) | |
1110 | reord = min(fack_count, reord); | |
1111 | } | |
1112 | ||
1113 | /* D-SACK. We can detect redundant retransmission | |
1114 | * in S|R and plain R frames and clear it. | |
1115 | * undo_retrans is decreased above, L|R frames | |
1116 | * are accounted above as well. | |
1117 | */ | |
1118 | if (dup_sack && | |
1119 | (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS)) { | |
1120 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; | |
1121 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1122 | } | |
1123 | } | |
1124 | } | |
1125 | ||
1126 | /* Check for lost retransmit. This superb idea is | |
1127 | * borrowed from "ratehalving". Event "C". | |
1128 | * Later note: FACK people cheated me again 8), | |
1129 | * we have to account for reordering! Ugly, | |
1130 | * but should help. | |
1131 | */ | |
1132 | if (lost_retrans && tp->ca_state == TCP_CA_Recovery) { | |
1133 | struct sk_buff *skb; | |
1134 | ||
1135 | sk_stream_for_retrans_queue(skb, sk) { | |
1136 | if (after(TCP_SKB_CB(skb)->seq, lost_retrans)) | |
1137 | break; | |
1138 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1139 | continue; | |
1140 | if ((TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) && | |
1141 | after(lost_retrans, TCP_SKB_CB(skb)->ack_seq) && | |
1142 | (IsFack(tp) || | |
1143 | !before(lost_retrans, | |
1144 | TCP_SKB_CB(skb)->ack_seq + tp->reordering * | |
1145 | tp->mss_cache_std))) { | |
1146 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; | |
1147 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1148 | ||
1149 | if (!(TCP_SKB_CB(skb)->sacked&(TCPCB_LOST|TCPCB_SACKED_ACKED))) { | |
1150 | tp->lost_out += tcp_skb_pcount(skb); | |
1151 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1152 | flag |= FLAG_DATA_SACKED; | |
1153 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT); | |
1154 | } | |
1155 | } | |
1156 | } | |
1157 | } | |
1158 | ||
1159 | tp->left_out = tp->sacked_out + tp->lost_out; | |
1160 | ||
1161 | if ((reord < tp->fackets_out) && tp->ca_state != TCP_CA_Loss) | |
1162 | tcp_update_reordering(tp, ((tp->fackets_out + 1) - reord), 0); | |
1163 | ||
1164 | #if FASTRETRANS_DEBUG > 0 | |
1165 | BUG_TRAP((int)tp->sacked_out >= 0); | |
1166 | BUG_TRAP((int)tp->lost_out >= 0); | |
1167 | BUG_TRAP((int)tp->retrans_out >= 0); | |
1168 | BUG_TRAP((int)tcp_packets_in_flight(tp) >= 0); | |
1169 | #endif | |
1170 | return flag; | |
1171 | } | |
1172 | ||
1173 | /* RTO occurred, but do not yet enter loss state. Instead, transmit two new | |
1174 | * segments to see from the next ACKs whether any data was really missing. | |
1175 | * If the RTO was spurious, new ACKs should arrive. | |
1176 | */ | |
1177 | void tcp_enter_frto(struct sock *sk) | |
1178 | { | |
1179 | struct tcp_sock *tp = tcp_sk(sk); | |
1180 | struct sk_buff *skb; | |
1181 | ||
1182 | tp->frto_counter = 1; | |
1183 | ||
1184 | if (tp->ca_state <= TCP_CA_Disorder || | |
1185 | tp->snd_una == tp->high_seq || | |
1186 | (tp->ca_state == TCP_CA_Loss && !tp->retransmits)) { | |
1187 | tp->prior_ssthresh = tcp_current_ssthresh(tp); | |
1188 | if (!tcp_westwood_ssthresh(tp)) | |
1189 | tp->snd_ssthresh = tcp_recalc_ssthresh(tp); | |
1190 | } | |
1191 | ||
1192 | /* Have to clear retransmission markers here to keep the bookkeeping | |
1193 | * in shape, even though we are not yet in Loss state. | |
1194 | * If something was really lost, it is eventually caught up | |
1195 | * in tcp_enter_frto_loss. | |
1196 | */ | |
1197 | tp->retrans_out = 0; | |
1198 | tp->undo_marker = tp->snd_una; | |
1199 | tp->undo_retrans = 0; | |
1200 | ||
1201 | sk_stream_for_retrans_queue(skb, sk) { | |
1202 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_RETRANS; | |
1203 | } | |
1204 | tcp_sync_left_out(tp); | |
1205 | ||
1206 | tcp_set_ca_state(tp, TCP_CA_Open); | |
1207 | tp->frto_highmark = tp->snd_nxt; | |
1208 | } | |
1209 | ||
1210 | /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, | |
1211 | * which indicates that we should follow the traditional RTO recovery, | |
1212 | * i.e. mark everything lost and do go-back-N retransmission. | |
1213 | */ | |
1214 | static void tcp_enter_frto_loss(struct sock *sk) | |
1215 | { | |
1216 | struct tcp_sock *tp = tcp_sk(sk); | |
1217 | struct sk_buff *skb; | |
1218 | int cnt = 0; | |
1219 | ||
1220 | tp->sacked_out = 0; | |
1221 | tp->lost_out = 0; | |
1222 | tp->fackets_out = 0; | |
1223 | ||
1224 | sk_stream_for_retrans_queue(skb, sk) { | |
1225 | cnt += tcp_skb_pcount(skb); | |
1226 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; | |
1227 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) { | |
1228 | ||
1229 | /* Do not mark those segments lost that were | |
1230 | * forward transmitted after RTO | |
1231 | */ | |
1232 | if (!after(TCP_SKB_CB(skb)->end_seq, | |
1233 | tp->frto_highmark)) { | |
1234 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1235 | tp->lost_out += tcp_skb_pcount(skb); | |
1236 | } | |
1237 | } else { | |
1238 | tp->sacked_out += tcp_skb_pcount(skb); | |
1239 | tp->fackets_out = cnt; | |
1240 | } | |
1241 | } | |
1242 | tcp_sync_left_out(tp); | |
1243 | ||
1244 | tp->snd_cwnd = tp->frto_counter + tcp_packets_in_flight(tp)+1; | |
1245 | tp->snd_cwnd_cnt = 0; | |
1246 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1247 | tp->undo_marker = 0; | |
1248 | tp->frto_counter = 0; | |
1249 | ||
1250 | tp->reordering = min_t(unsigned int, tp->reordering, | |
1251 | sysctl_tcp_reordering); | |
1252 | tcp_set_ca_state(tp, TCP_CA_Loss); | |
1253 | tp->high_seq = tp->frto_highmark; | |
1254 | TCP_ECN_queue_cwr(tp); | |
1255 | ||
1256 | init_bictcp(tp); | |
1257 | } | |
1258 | ||
1259 | void tcp_clear_retrans(struct tcp_sock *tp) | |
1260 | { | |
1261 | tp->left_out = 0; | |
1262 | tp->retrans_out = 0; | |
1263 | ||
1264 | tp->fackets_out = 0; | |
1265 | tp->sacked_out = 0; | |
1266 | tp->lost_out = 0; | |
1267 | ||
1268 | tp->undo_marker = 0; | |
1269 | tp->undo_retrans = 0; | |
1270 | } | |
1271 | ||
1272 | /* Enter Loss state. If "how" is not zero, forget all SACK information | |
1273 | * and reset tags completely, otherwise preserve SACKs. If receiver | |
1274 | * dropped its ofo queue, we will know this due to reneging detection. | |
1275 | */ | |
1276 | void tcp_enter_loss(struct sock *sk, int how) | |
1277 | { | |
1278 | struct tcp_sock *tp = tcp_sk(sk); | |
1279 | struct sk_buff *skb; | |
1280 | int cnt = 0; | |
1281 | ||
1282 | /* Reduce ssthresh if it has not yet been made inside this window. */ | |
1283 | if (tp->ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || | |
1284 | (tp->ca_state == TCP_CA_Loss && !tp->retransmits)) { | |
1285 | tp->prior_ssthresh = tcp_current_ssthresh(tp); | |
1286 | tp->snd_ssthresh = tcp_recalc_ssthresh(tp); | |
1287 | } | |
1288 | tp->snd_cwnd = 1; | |
1289 | tp->snd_cwnd_cnt = 0; | |
1290 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1291 | ||
1292 | tcp_clear_retrans(tp); | |
1293 | ||
1294 | /* Push undo marker, if it was plain RTO and nothing | |
1295 | * was retransmitted. */ | |
1296 | if (!how) | |
1297 | tp->undo_marker = tp->snd_una; | |
1298 | ||
1299 | sk_stream_for_retrans_queue(skb, sk) { | |
1300 | cnt += tcp_skb_pcount(skb); | |
1301 | if (TCP_SKB_CB(skb)->sacked&TCPCB_RETRANS) | |
1302 | tp->undo_marker = 0; | |
1303 | TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; | |
1304 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { | |
1305 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; | |
1306 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1307 | tp->lost_out += tcp_skb_pcount(skb); | |
1308 | } else { | |
1309 | tp->sacked_out += tcp_skb_pcount(skb); | |
1310 | tp->fackets_out = cnt; | |
1311 | } | |
1312 | } | |
1313 | tcp_sync_left_out(tp); | |
1314 | ||
1315 | tp->reordering = min_t(unsigned int, tp->reordering, | |
1316 | sysctl_tcp_reordering); | |
1317 | tcp_set_ca_state(tp, TCP_CA_Loss); | |
1318 | tp->high_seq = tp->snd_nxt; | |
1319 | TCP_ECN_queue_cwr(tp); | |
1320 | } | |
1321 | ||
1322 | static int tcp_check_sack_reneging(struct sock *sk, struct tcp_sock *tp) | |
1323 | { | |
1324 | struct sk_buff *skb; | |
1325 | ||
1326 | /* If ACK arrived pointing to a remembered SACK, | |
1327 | * it means that our remembered SACKs do not reflect | |
1328 | * real state of receiver i.e. | |
1329 | * receiver _host_ is heavily congested (or buggy). | |
1330 | * Do processing similar to RTO timeout. | |
1331 | */ | |
1332 | if ((skb = skb_peek(&sk->sk_write_queue)) != NULL && | |
1333 | (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { | |
1334 | NET_INC_STATS_BH(LINUX_MIB_TCPSACKRENEGING); | |
1335 | ||
1336 | tcp_enter_loss(sk, 1); | |
1337 | tp->retransmits++; | |
1338 | tcp_retransmit_skb(sk, skb_peek(&sk->sk_write_queue)); | |
1339 | tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto); | |
1340 | return 1; | |
1341 | } | |
1342 | return 0; | |
1343 | } | |
1344 | ||
1345 | static inline int tcp_fackets_out(struct tcp_sock *tp) | |
1346 | { | |
1347 | return IsReno(tp) ? tp->sacked_out+1 : tp->fackets_out; | |
1348 | } | |
1349 | ||
1350 | static inline int tcp_skb_timedout(struct tcp_sock *tp, struct sk_buff *skb) | |
1351 | { | |
1352 | return (tcp_time_stamp - TCP_SKB_CB(skb)->when > tp->rto); | |
1353 | } | |
1354 | ||
1355 | static inline int tcp_head_timedout(struct sock *sk, struct tcp_sock *tp) | |
1356 | { | |
1357 | return tp->packets_out && | |
1358 | tcp_skb_timedout(tp, skb_peek(&sk->sk_write_queue)); | |
1359 | } | |
1360 | ||
1361 | /* Linux NewReno/SACK/FACK/ECN state machine. | |
1362 | * -------------------------------------- | |
1363 | * | |
1364 | * "Open" Normal state, no dubious events, fast path. | |
1365 | * "Disorder" In all the respects it is "Open", | |
1366 | * but requires a bit more attention. It is entered when | |
1367 | * we see some SACKs or dupacks. It is split of "Open" | |
1368 | * mainly to move some processing from fast path to slow one. | |
1369 | * "CWR" CWND was reduced due to some Congestion Notification event. | |
1370 | * It can be ECN, ICMP source quench, local device congestion. | |
1371 | * "Recovery" CWND was reduced, we are fast-retransmitting. | |
1372 | * "Loss" CWND was reduced due to RTO timeout or SACK reneging. | |
1373 | * | |
1374 | * tcp_fastretrans_alert() is entered: | |
1375 | * - each incoming ACK, if state is not "Open" | |
1376 | * - when arrived ACK is unusual, namely: | |
1377 | * * SACK | |
1378 | * * Duplicate ACK. | |
1379 | * * ECN ECE. | |
1380 | * | |
1381 | * Counting packets in flight is pretty simple. | |
1382 | * | |
1383 | * in_flight = packets_out - left_out + retrans_out | |
1384 | * | |
1385 | * packets_out is SND.NXT-SND.UNA counted in packets. | |
1386 | * | |
1387 | * retrans_out is number of retransmitted segments. | |
1388 | * | |
1389 | * left_out is number of segments left network, but not ACKed yet. | |
1390 | * | |
1391 | * left_out = sacked_out + lost_out | |
1392 | * | |
1393 | * sacked_out: Packets, which arrived to receiver out of order | |
1394 | * and hence not ACKed. With SACKs this number is simply | |
1395 | * amount of SACKed data. Even without SACKs | |
1396 | * it is easy to give pretty reliable estimate of this number, | |
1397 | * counting duplicate ACKs. | |
1398 | * | |
1399 | * lost_out: Packets lost by network. TCP has no explicit | |
1400 | * "loss notification" feedback from network (for now). | |
1401 | * It means that this number can be only _guessed_. | |
1402 | * Actually, it is the heuristics to predict lossage that | |
1403 | * distinguishes different algorithms. | |
1404 | * | |
1405 | * F.e. after RTO, when all the queue is considered as lost, | |
1406 | * lost_out = packets_out and in_flight = retrans_out. | |
1407 | * | |
1408 | * Essentially, we have now two algorithms counting | |
1409 | * lost packets. | |
1410 | * | |
1411 | * FACK: It is the simplest heuristics. As soon as we decided | |
1412 | * that something is lost, we decide that _all_ not SACKed | |
1413 | * packets until the most forward SACK are lost. I.e. | |
1414 | * lost_out = fackets_out - sacked_out and left_out = fackets_out. | |
1415 | * It is absolutely correct estimate, if network does not reorder | |
1416 | * packets. And it loses any connection to reality when reordering | |
1417 | * takes place. We use FACK by default until reordering | |
1418 | * is suspected on the path to this destination. | |
1419 | * | |
1420 | * NewReno: when Recovery is entered, we assume that one segment | |
1421 | * is lost (classic Reno). While we are in Recovery and | |
1422 | * a partial ACK arrives, we assume that one more packet | |
1423 | * is lost (NewReno). This heuristics are the same in NewReno | |
1424 | * and SACK. | |
1425 | * | |
1426 | * Imagine, that's all! Forget about all this shamanism about CWND inflation | |
1427 | * deflation etc. CWND is real congestion window, never inflated, changes | |
1428 | * only according to classic VJ rules. | |
1429 | * | |
1430 | * Really tricky (and requiring careful tuning) part of algorithm | |
1431 | * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). | |
1432 | * The first determines the moment _when_ we should reduce CWND and, | |
1433 | * hence, slow down forward transmission. In fact, it determines the moment | |
1434 | * when we decide that hole is caused by loss, rather than by a reorder. | |
1435 | * | |
1436 | * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill | |
1437 | * holes, caused by lost packets. | |
1438 | * | |
1439 | * And the most logically complicated part of algorithm is undo | |
1440 | * heuristics. We detect false retransmits due to both too early | |
1441 | * fast retransmit (reordering) and underestimated RTO, analyzing | |
1442 | * timestamps and D-SACKs. When we detect that some segments were | |
1443 | * retransmitted by mistake and CWND reduction was wrong, we undo | |
1444 | * window reduction and abort recovery phase. This logic is hidden | |
1445 | * inside several functions named tcp_try_undo_<something>. | |
1446 | */ | |
1447 | ||
1448 | /* This function decides, when we should leave Disordered state | |
1449 | * and enter Recovery phase, reducing congestion window. | |
1450 | * | |
1451 | * Main question: may we further continue forward transmission | |
1452 | * with the same cwnd? | |
1453 | */ | |
1454 | static int tcp_time_to_recover(struct sock *sk, struct tcp_sock *tp) | |
1455 | { | |
1456 | __u32 packets_out; | |
1457 | ||
1458 | /* Trick#1: The loss is proven. */ | |
1459 | if (tp->lost_out) | |
1460 | return 1; | |
1461 | ||
1462 | /* Not-A-Trick#2 : Classic rule... */ | |
1463 | if (tcp_fackets_out(tp) > tp->reordering) | |
1464 | return 1; | |
1465 | ||
1466 | /* Trick#3 : when we use RFC2988 timer restart, fast | |
1467 | * retransmit can be triggered by timeout of queue head. | |
1468 | */ | |
1469 | if (tcp_head_timedout(sk, tp)) | |
1470 | return 1; | |
1471 | ||
1472 | /* Trick#4: It is still not OK... But will it be useful to delay | |
1473 | * recovery more? | |
1474 | */ | |
1475 | packets_out = tp->packets_out; | |
1476 | if (packets_out <= tp->reordering && | |
1477 | tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && | |
1478 | !tcp_may_send_now(sk, tp)) { | |
1479 | /* We have nothing to send. This connection is limited | |
1480 | * either by receiver window or by application. | |
1481 | */ | |
1482 | return 1; | |
1483 | } | |
1484 | ||
1485 | return 0; | |
1486 | } | |
1487 | ||
1488 | /* If we receive more dupacks than we expected counting segments | |
1489 | * in assumption of absent reordering, interpret this as reordering. | |
1490 | * The only another reason could be bug in receiver TCP. | |
1491 | */ | |
1492 | static void tcp_check_reno_reordering(struct tcp_sock *tp, int addend) | |
1493 | { | |
1494 | u32 holes; | |
1495 | ||
1496 | holes = max(tp->lost_out, 1U); | |
1497 | holes = min(holes, tp->packets_out); | |
1498 | ||
1499 | if ((tp->sacked_out + holes) > tp->packets_out) { | |
1500 | tp->sacked_out = tp->packets_out - holes; | |
1501 | tcp_update_reordering(tp, tp->packets_out+addend, 0); | |
1502 | } | |
1503 | } | |
1504 | ||
1505 | /* Emulate SACKs for SACKless connection: account for a new dupack. */ | |
1506 | ||
1507 | static void tcp_add_reno_sack(struct tcp_sock *tp) | |
1508 | { | |
1509 | tp->sacked_out++; | |
1510 | tcp_check_reno_reordering(tp, 0); | |
1511 | tcp_sync_left_out(tp); | |
1512 | } | |
1513 | ||
1514 | /* Account for ACK, ACKing some data in Reno Recovery phase. */ | |
1515 | ||
1516 | static void tcp_remove_reno_sacks(struct sock *sk, struct tcp_sock *tp, int acked) | |
1517 | { | |
1518 | if (acked > 0) { | |
1519 | /* One ACK acked hole. The rest eat duplicate ACKs. */ | |
1520 | if (acked-1 >= tp->sacked_out) | |
1521 | tp->sacked_out = 0; | |
1522 | else | |
1523 | tp->sacked_out -= acked-1; | |
1524 | } | |
1525 | tcp_check_reno_reordering(tp, acked); | |
1526 | tcp_sync_left_out(tp); | |
1527 | } | |
1528 | ||
1529 | static inline void tcp_reset_reno_sack(struct tcp_sock *tp) | |
1530 | { | |
1531 | tp->sacked_out = 0; | |
1532 | tp->left_out = tp->lost_out; | |
1533 | } | |
1534 | ||
1535 | /* Mark head of queue up as lost. */ | |
1536 | static void tcp_mark_head_lost(struct sock *sk, struct tcp_sock *tp, | |
1537 | int packets, u32 high_seq) | |
1538 | { | |
1539 | struct sk_buff *skb; | |
1540 | int cnt = packets; | |
1541 | ||
1542 | BUG_TRAP(cnt <= tp->packets_out); | |
1543 | ||
1544 | sk_stream_for_retrans_queue(skb, sk) { | |
1545 | cnt -= tcp_skb_pcount(skb); | |
1546 | if (cnt < 0 || after(TCP_SKB_CB(skb)->end_seq, high_seq)) | |
1547 | break; | |
1548 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) { | |
1549 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1550 | tp->lost_out += tcp_skb_pcount(skb); | |
1551 | } | |
1552 | } | |
1553 | tcp_sync_left_out(tp); | |
1554 | } | |
1555 | ||
1556 | /* Account newly detected lost packet(s) */ | |
1557 | ||
1558 | static void tcp_update_scoreboard(struct sock *sk, struct tcp_sock *tp) | |
1559 | { | |
1560 | if (IsFack(tp)) { | |
1561 | int lost = tp->fackets_out - tp->reordering; | |
1562 | if (lost <= 0) | |
1563 | lost = 1; | |
1564 | tcp_mark_head_lost(sk, tp, lost, tp->high_seq); | |
1565 | } else { | |
1566 | tcp_mark_head_lost(sk, tp, 1, tp->high_seq); | |
1567 | } | |
1568 | ||
1569 | /* New heuristics: it is possible only after we switched | |
1570 | * to restart timer each time when something is ACKed. | |
1571 | * Hence, we can detect timed out packets during fast | |
1572 | * retransmit without falling to slow start. | |
1573 | */ | |
1574 | if (tcp_head_timedout(sk, tp)) { | |
1575 | struct sk_buff *skb; | |
1576 | ||
1577 | sk_stream_for_retrans_queue(skb, sk) { | |
1578 | if (tcp_skb_timedout(tp, skb) && | |
1579 | !(TCP_SKB_CB(skb)->sacked&TCPCB_TAGBITS)) { | |
1580 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1581 | tp->lost_out += tcp_skb_pcount(skb); | |
1582 | } | |
1583 | } | |
1584 | tcp_sync_left_out(tp); | |
1585 | } | |
1586 | } | |
1587 | ||
1588 | /* CWND moderation, preventing bursts due to too big ACKs | |
1589 | * in dubious situations. | |
1590 | */ | |
1591 | static inline void tcp_moderate_cwnd(struct tcp_sock *tp) | |
1592 | { | |
1593 | tp->snd_cwnd = min(tp->snd_cwnd, | |
1594 | tcp_packets_in_flight(tp)+tcp_max_burst(tp)); | |
1595 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1596 | } | |
1597 | ||
1598 | /* Decrease cwnd each second ack. */ | |
1599 | ||
1600 | static void tcp_cwnd_down(struct tcp_sock *tp) | |
1601 | { | |
1602 | int decr = tp->snd_cwnd_cnt + 1; | |
1603 | __u32 limit; | |
1604 | ||
1605 | /* | |
1606 | * TCP Westwood | |
1607 | * Here limit is evaluated as BWestimation*RTTmin (for obtaining it | |
1608 | * in packets we use mss_cache). If sysctl_tcp_westwood is off | |
1609 | * tcp_westwood_bw_rttmin() returns 0. In such case snd_ssthresh is | |
1610 | * still used as usual. It prevents other strange cases in which | |
1611 | * BWE*RTTmin could assume value 0. It should not happen but... | |
1612 | */ | |
1613 | ||
1614 | if (!(limit = tcp_westwood_bw_rttmin(tp))) | |
1615 | limit = tp->snd_ssthresh/2; | |
1616 | ||
1617 | tp->snd_cwnd_cnt = decr&1; | |
1618 | decr >>= 1; | |
1619 | ||
1620 | if (decr && tp->snd_cwnd > limit) | |
1621 | tp->snd_cwnd -= decr; | |
1622 | ||
1623 | tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp)+1); | |
1624 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1625 | } | |
1626 | ||
1627 | /* Nothing was retransmitted or returned timestamp is less | |
1628 | * than timestamp of the first retransmission. | |
1629 | */ | |
1630 | static inline int tcp_packet_delayed(struct tcp_sock *tp) | |
1631 | { | |
1632 | return !tp->retrans_stamp || | |
1633 | (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && | |
1634 | (__s32)(tp->rx_opt.rcv_tsecr - tp->retrans_stamp) < 0); | |
1635 | } | |
1636 | ||
1637 | /* Undo procedures. */ | |
1638 | ||
1639 | #if FASTRETRANS_DEBUG > 1 | |
1640 | static void DBGUNDO(struct sock *sk, struct tcp_sock *tp, const char *msg) | |
1641 | { | |
1642 | struct inet_sock *inet = inet_sk(sk); | |
1643 | printk(KERN_DEBUG "Undo %s %u.%u.%u.%u/%u c%u l%u ss%u/%u p%u\n", | |
1644 | msg, | |
1645 | NIPQUAD(inet->daddr), ntohs(inet->dport), | |
1646 | tp->snd_cwnd, tp->left_out, | |
1647 | tp->snd_ssthresh, tp->prior_ssthresh, | |
1648 | tp->packets_out); | |
1649 | } | |
1650 | #else | |
1651 | #define DBGUNDO(x...) do { } while (0) | |
1652 | #endif | |
1653 | ||
1654 | static void tcp_undo_cwr(struct tcp_sock *tp, int undo) | |
1655 | { | |
1656 | if (tp->prior_ssthresh) { | |
1657 | if (tcp_is_bic(tp)) | |
1658 | tp->snd_cwnd = max(tp->snd_cwnd, tp->bictcp.last_max_cwnd); | |
1659 | else | |
1660 | tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh<<1); | |
1661 | ||
1662 | if (undo && tp->prior_ssthresh > tp->snd_ssthresh) { | |
1663 | tp->snd_ssthresh = tp->prior_ssthresh; | |
1664 | TCP_ECN_withdraw_cwr(tp); | |
1665 | } | |
1666 | } else { | |
1667 | tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); | |
1668 | } | |
1669 | tcp_moderate_cwnd(tp); | |
1670 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1671 | } | |
1672 | ||
1673 | static inline int tcp_may_undo(struct tcp_sock *tp) | |
1674 | { | |
1675 | return tp->undo_marker && | |
1676 | (!tp->undo_retrans || tcp_packet_delayed(tp)); | |
1677 | } | |
1678 | ||
1679 | /* People celebrate: "We love our President!" */ | |
1680 | static int tcp_try_undo_recovery(struct sock *sk, struct tcp_sock *tp) | |
1681 | { | |
1682 | if (tcp_may_undo(tp)) { | |
1683 | /* Happy end! We did not retransmit anything | |
1684 | * or our original transmission succeeded. | |
1685 | */ | |
1686 | DBGUNDO(sk, tp, tp->ca_state == TCP_CA_Loss ? "loss" : "retrans"); | |
1687 | tcp_undo_cwr(tp, 1); | |
1688 | if (tp->ca_state == TCP_CA_Loss) | |
1689 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); | |
1690 | else | |
1691 | NET_INC_STATS_BH(LINUX_MIB_TCPFULLUNDO); | |
1692 | tp->undo_marker = 0; | |
1693 | } | |
1694 | if (tp->snd_una == tp->high_seq && IsReno(tp)) { | |
1695 | /* Hold old state until something *above* high_seq | |
1696 | * is ACKed. For Reno it is MUST to prevent false | |
1697 | * fast retransmits (RFC2582). SACK TCP is safe. */ | |
1698 | tcp_moderate_cwnd(tp); | |
1699 | return 1; | |
1700 | } | |
1701 | tcp_set_ca_state(tp, TCP_CA_Open); | |
1702 | return 0; | |
1703 | } | |
1704 | ||
1705 | /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ | |
1706 | static void tcp_try_undo_dsack(struct sock *sk, struct tcp_sock *tp) | |
1707 | { | |
1708 | if (tp->undo_marker && !tp->undo_retrans) { | |
1709 | DBGUNDO(sk, tp, "D-SACK"); | |
1710 | tcp_undo_cwr(tp, 1); | |
1711 | tp->undo_marker = 0; | |
1712 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKUNDO); | |
1713 | } | |
1714 | } | |
1715 | ||
1716 | /* Undo during fast recovery after partial ACK. */ | |
1717 | ||
1718 | static int tcp_try_undo_partial(struct sock *sk, struct tcp_sock *tp, | |
1719 | int acked) | |
1720 | { | |
1721 | /* Partial ACK arrived. Force Hoe's retransmit. */ | |
1722 | int failed = IsReno(tp) || tp->fackets_out>tp->reordering; | |
1723 | ||
1724 | if (tcp_may_undo(tp)) { | |
1725 | /* Plain luck! Hole if filled with delayed | |
1726 | * packet, rather than with a retransmit. | |
1727 | */ | |
1728 | if (tp->retrans_out == 0) | |
1729 | tp->retrans_stamp = 0; | |
1730 | ||
1731 | tcp_update_reordering(tp, tcp_fackets_out(tp)+acked, 1); | |
1732 | ||
1733 | DBGUNDO(sk, tp, "Hoe"); | |
1734 | tcp_undo_cwr(tp, 0); | |
1735 | NET_INC_STATS_BH(LINUX_MIB_TCPPARTIALUNDO); | |
1736 | ||
1737 | /* So... Do not make Hoe's retransmit yet. | |
1738 | * If the first packet was delayed, the rest | |
1739 | * ones are most probably delayed as well. | |
1740 | */ | |
1741 | failed = 0; | |
1742 | } | |
1743 | return failed; | |
1744 | } | |
1745 | ||
1746 | /* Undo during loss recovery after partial ACK. */ | |
1747 | static int tcp_try_undo_loss(struct sock *sk, struct tcp_sock *tp) | |
1748 | { | |
1749 | if (tcp_may_undo(tp)) { | |
1750 | struct sk_buff *skb; | |
1751 | sk_stream_for_retrans_queue(skb, sk) { | |
1752 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; | |
1753 | } | |
1754 | DBGUNDO(sk, tp, "partial loss"); | |
1755 | tp->lost_out = 0; | |
1756 | tp->left_out = tp->sacked_out; | |
1757 | tcp_undo_cwr(tp, 1); | |
1758 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSSUNDO); | |
1759 | tp->retransmits = 0; | |
1760 | tp->undo_marker = 0; | |
1761 | if (!IsReno(tp)) | |
1762 | tcp_set_ca_state(tp, TCP_CA_Open); | |
1763 | return 1; | |
1764 | } | |
1765 | return 0; | |
1766 | } | |
1767 | ||
1768 | static inline void tcp_complete_cwr(struct tcp_sock *tp) | |
1769 | { | |
1770 | if (tcp_westwood_cwnd(tp)) | |
1771 | tp->snd_ssthresh = tp->snd_cwnd; | |
1772 | else | |
1773 | tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); | |
1774 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1775 | } | |
1776 | ||
1777 | static void tcp_try_to_open(struct sock *sk, struct tcp_sock *tp, int flag) | |
1778 | { | |
1779 | tp->left_out = tp->sacked_out; | |
1780 | ||
1781 | if (tp->retrans_out == 0) | |
1782 | tp->retrans_stamp = 0; | |
1783 | ||
1784 | if (flag&FLAG_ECE) | |
1785 | tcp_enter_cwr(tp); | |
1786 | ||
1787 | if (tp->ca_state != TCP_CA_CWR) { | |
1788 | int state = TCP_CA_Open; | |
1789 | ||
1790 | if (tp->left_out || tp->retrans_out || tp->undo_marker) | |
1791 | state = TCP_CA_Disorder; | |
1792 | ||
1793 | if (tp->ca_state != state) { | |
1794 | tcp_set_ca_state(tp, state); | |
1795 | tp->high_seq = tp->snd_nxt; | |
1796 | } | |
1797 | tcp_moderate_cwnd(tp); | |
1798 | } else { | |
1799 | tcp_cwnd_down(tp); | |
1800 | } | |
1801 | } | |
1802 | ||
1803 | /* Process an event, which can update packets-in-flight not trivially. | |
1804 | * Main goal of this function is to calculate new estimate for left_out, | |
1805 | * taking into account both packets sitting in receiver's buffer and | |
1806 | * packets lost by network. | |
1807 | * | |
1808 | * Besides that it does CWND reduction, when packet loss is detected | |
1809 | * and changes state of machine. | |
1810 | * | |
1811 | * It does _not_ decide what to send, it is made in function | |
1812 | * tcp_xmit_retransmit_queue(). | |
1813 | */ | |
1814 | static void | |
1815 | tcp_fastretrans_alert(struct sock *sk, u32 prior_snd_una, | |
1816 | int prior_packets, int flag) | |
1817 | { | |
1818 | struct tcp_sock *tp = tcp_sk(sk); | |
1819 | int is_dupack = (tp->snd_una == prior_snd_una && !(flag&FLAG_NOT_DUP)); | |
1820 | ||
1821 | /* Some technical things: | |
1822 | * 1. Reno does not count dupacks (sacked_out) automatically. */ | |
1823 | if (!tp->packets_out) | |
1824 | tp->sacked_out = 0; | |
1825 | /* 2. SACK counts snd_fack in packets inaccurately. */ | |
1826 | if (tp->sacked_out == 0) | |
1827 | tp->fackets_out = 0; | |
1828 | ||
1829 | /* Now state machine starts. | |
1830 | * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ | |
1831 | if (flag&FLAG_ECE) | |
1832 | tp->prior_ssthresh = 0; | |
1833 | ||
1834 | /* B. In all the states check for reneging SACKs. */ | |
1835 | if (tp->sacked_out && tcp_check_sack_reneging(sk, tp)) | |
1836 | return; | |
1837 | ||
1838 | /* C. Process data loss notification, provided it is valid. */ | |
1839 | if ((flag&FLAG_DATA_LOST) && | |
1840 | before(tp->snd_una, tp->high_seq) && | |
1841 | tp->ca_state != TCP_CA_Open && | |
1842 | tp->fackets_out > tp->reordering) { | |
1843 | tcp_mark_head_lost(sk, tp, tp->fackets_out-tp->reordering, tp->high_seq); | |
1844 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSS); | |
1845 | } | |
1846 | ||
1847 | /* D. Synchronize left_out to current state. */ | |
1848 | tcp_sync_left_out(tp); | |
1849 | ||
1850 | /* E. Check state exit conditions. State can be terminated | |
1851 | * when high_seq is ACKed. */ | |
1852 | if (tp->ca_state == TCP_CA_Open) { | |
1853 | if (!sysctl_tcp_frto) | |
1854 | BUG_TRAP(tp->retrans_out == 0); | |
1855 | tp->retrans_stamp = 0; | |
1856 | } else if (!before(tp->snd_una, tp->high_seq)) { | |
1857 | switch (tp->ca_state) { | |
1858 | case TCP_CA_Loss: | |
1859 | tp->retransmits = 0; | |
1860 | if (tcp_try_undo_recovery(sk, tp)) | |
1861 | return; | |
1862 | break; | |
1863 | ||
1864 | case TCP_CA_CWR: | |
1865 | /* CWR is to be held something *above* high_seq | |
1866 | * is ACKed for CWR bit to reach receiver. */ | |
1867 | if (tp->snd_una != tp->high_seq) { | |
1868 | tcp_complete_cwr(tp); | |
1869 | tcp_set_ca_state(tp, TCP_CA_Open); | |
1870 | } | |
1871 | break; | |
1872 | ||
1873 | case TCP_CA_Disorder: | |
1874 | tcp_try_undo_dsack(sk, tp); | |
1875 | if (!tp->undo_marker || | |
1876 | /* For SACK case do not Open to allow to undo | |
1877 | * catching for all duplicate ACKs. */ | |
1878 | IsReno(tp) || tp->snd_una != tp->high_seq) { | |
1879 | tp->undo_marker = 0; | |
1880 | tcp_set_ca_state(tp, TCP_CA_Open); | |
1881 | } | |
1882 | break; | |
1883 | ||
1884 | case TCP_CA_Recovery: | |
1885 | if (IsReno(tp)) | |
1886 | tcp_reset_reno_sack(tp); | |
1887 | if (tcp_try_undo_recovery(sk, tp)) | |
1888 | return; | |
1889 | tcp_complete_cwr(tp); | |
1890 | break; | |
1891 | } | |
1892 | } | |
1893 | ||
1894 | /* F. Process state. */ | |
1895 | switch (tp->ca_state) { | |
1896 | case TCP_CA_Recovery: | |
1897 | if (prior_snd_una == tp->snd_una) { | |
1898 | if (IsReno(tp) && is_dupack) | |
1899 | tcp_add_reno_sack(tp); | |
1900 | } else { | |
1901 | int acked = prior_packets - tp->packets_out; | |
1902 | if (IsReno(tp)) | |
1903 | tcp_remove_reno_sacks(sk, tp, acked); | |
1904 | is_dupack = tcp_try_undo_partial(sk, tp, acked); | |
1905 | } | |
1906 | break; | |
1907 | case TCP_CA_Loss: | |
1908 | if (flag&FLAG_DATA_ACKED) | |
1909 | tp->retransmits = 0; | |
1910 | if (!tcp_try_undo_loss(sk, tp)) { | |
1911 | tcp_moderate_cwnd(tp); | |
1912 | tcp_xmit_retransmit_queue(sk); | |
1913 | return; | |
1914 | } | |
1915 | if (tp->ca_state != TCP_CA_Open) | |
1916 | return; | |
1917 | /* Loss is undone; fall through to processing in Open state. */ | |
1918 | default: | |
1919 | if (IsReno(tp)) { | |
1920 | if (tp->snd_una != prior_snd_una) | |
1921 | tcp_reset_reno_sack(tp); | |
1922 | if (is_dupack) | |
1923 | tcp_add_reno_sack(tp); | |
1924 | } | |
1925 | ||
1926 | if (tp->ca_state == TCP_CA_Disorder) | |
1927 | tcp_try_undo_dsack(sk, tp); | |
1928 | ||
1929 | if (!tcp_time_to_recover(sk, tp)) { | |
1930 | tcp_try_to_open(sk, tp, flag); | |
1931 | return; | |
1932 | } | |
1933 | ||
1934 | /* Otherwise enter Recovery state */ | |
1935 | ||
1936 | if (IsReno(tp)) | |
1937 | NET_INC_STATS_BH(LINUX_MIB_TCPRENORECOVERY); | |
1938 | else | |
1939 | NET_INC_STATS_BH(LINUX_MIB_TCPSACKRECOVERY); | |
1940 | ||
1941 | tp->high_seq = tp->snd_nxt; | |
1942 | tp->prior_ssthresh = 0; | |
1943 | tp->undo_marker = tp->snd_una; | |
1944 | tp->undo_retrans = tp->retrans_out; | |
1945 | ||
1946 | if (tp->ca_state < TCP_CA_CWR) { | |
1947 | if (!(flag&FLAG_ECE)) | |
1948 | tp->prior_ssthresh = tcp_current_ssthresh(tp); | |
1949 | tp->snd_ssthresh = tcp_recalc_ssthresh(tp); | |
1950 | TCP_ECN_queue_cwr(tp); | |
1951 | } | |
1952 | ||
1953 | tp->snd_cwnd_cnt = 0; | |
1954 | tcp_set_ca_state(tp, TCP_CA_Recovery); | |
1955 | } | |
1956 | ||
1957 | if (is_dupack || tcp_head_timedout(sk, tp)) | |
1958 | tcp_update_scoreboard(sk, tp); | |
1959 | tcp_cwnd_down(tp); | |
1960 | tcp_xmit_retransmit_queue(sk); | |
1961 | } | |
1962 | ||
1963 | /* Read draft-ietf-tcplw-high-performance before mucking | |
1964 | * with this code. (Superceeds RFC1323) | |
1965 | */ | |
1966 | static void tcp_ack_saw_tstamp(struct tcp_sock *tp, int flag) | |
1967 | { | |
1968 | __u32 seq_rtt; | |
1969 | ||
1970 | /* RTTM Rule: A TSecr value received in a segment is used to | |
1971 | * update the averaged RTT measurement only if the segment | |
1972 | * acknowledges some new data, i.e., only if it advances the | |
1973 | * left edge of the send window. | |
1974 | * | |
1975 | * See draft-ietf-tcplw-high-performance-00, section 3.3. | |
1976 | * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> | |
1977 | * | |
1978 | * Changed: reset backoff as soon as we see the first valid sample. | |
1979 | * If we do not, we get strongly overstimated rto. With timestamps | |
1980 | * samples are accepted even from very old segments: f.e., when rtt=1 | |
1981 | * increases to 8, we retransmit 5 times and after 8 seconds delayed | |
1982 | * answer arrives rto becomes 120 seconds! If at least one of segments | |
1983 | * in window is lost... Voila. --ANK (010210) | |
1984 | */ | |
1985 | seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr; | |
1986 | tcp_rtt_estimator(tp, seq_rtt); | |
1987 | tcp_set_rto(tp); | |
1988 | tp->backoff = 0; | |
1989 | tcp_bound_rto(tp); | |
1990 | } | |
1991 | ||
1992 | static void tcp_ack_no_tstamp(struct tcp_sock *tp, u32 seq_rtt, int flag) | |
1993 | { | |
1994 | /* We don't have a timestamp. Can only use | |
1995 | * packets that are not retransmitted to determine | |
1996 | * rtt estimates. Also, we must not reset the | |
1997 | * backoff for rto until we get a non-retransmitted | |
1998 | * packet. This allows us to deal with a situation | |
1999 | * where the network delay has increased suddenly. | |
2000 | * I.e. Karn's algorithm. (SIGCOMM '87, p5.) | |
2001 | */ | |
2002 | ||
2003 | if (flag & FLAG_RETRANS_DATA_ACKED) | |
2004 | return; | |
2005 | ||
2006 | tcp_rtt_estimator(tp, seq_rtt); | |
2007 | tcp_set_rto(tp); | |
2008 | tp->backoff = 0; | |
2009 | tcp_bound_rto(tp); | |
2010 | } | |
2011 | ||
2012 | static inline void tcp_ack_update_rtt(struct tcp_sock *tp, | |
2013 | int flag, s32 seq_rtt) | |
2014 | { | |
2015 | /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ | |
2016 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) | |
2017 | tcp_ack_saw_tstamp(tp, flag); | |
2018 | else if (seq_rtt >= 0) | |
2019 | tcp_ack_no_tstamp(tp, seq_rtt, flag); | |
2020 | } | |
2021 | ||
2022 | /* | |
2023 | * Compute congestion window to use. | |
2024 | * | |
2025 | * This is from the implementation of BICTCP in | |
2026 | * Lison-Xu, Kahaled Harfoush, and Injog Rhee. | |
2027 | * "Binary Increase Congestion Control for Fast, Long Distance | |
2028 | * Networks" in InfoComm 2004 | |
2029 | * Available from: | |
2030 | * http://www.csc.ncsu.edu/faculty/rhee/export/bitcp.pdf | |
2031 | * | |
2032 | * Unless BIC is enabled and congestion window is large | |
2033 | * this behaves the same as the original Reno. | |
2034 | */ | |
2035 | static inline __u32 bictcp_cwnd(struct tcp_sock *tp) | |
2036 | { | |
2037 | /* orignal Reno behaviour */ | |
2038 | if (!tcp_is_bic(tp)) | |
2039 | return tp->snd_cwnd; | |
2040 | ||
2041 | if (tp->bictcp.last_cwnd == tp->snd_cwnd && | |
2042 | (s32)(tcp_time_stamp - tp->bictcp.last_stamp) <= (HZ>>5)) | |
2043 | return tp->bictcp.cnt; | |
2044 | ||
2045 | tp->bictcp.last_cwnd = tp->snd_cwnd; | |
2046 | tp->bictcp.last_stamp = tcp_time_stamp; | |
2047 | ||
2048 | /* start off normal */ | |
2049 | if (tp->snd_cwnd <= sysctl_tcp_bic_low_window) | |
2050 | tp->bictcp.cnt = tp->snd_cwnd; | |
2051 | ||
2052 | /* binary increase */ | |
2053 | else if (tp->snd_cwnd < tp->bictcp.last_max_cwnd) { | |
2054 | __u32 dist = (tp->bictcp.last_max_cwnd - tp->snd_cwnd) | |
2055 | / BICTCP_B; | |
2056 | ||
2057 | if (dist > BICTCP_MAX_INCREMENT) | |
2058 | /* linear increase */ | |
2059 | tp->bictcp.cnt = tp->snd_cwnd / BICTCP_MAX_INCREMENT; | |
2060 | else if (dist <= 1U) | |
2061 | /* binary search increase */ | |
2062 | tp->bictcp.cnt = tp->snd_cwnd * BICTCP_FUNC_OF_MIN_INCR | |
2063 | / BICTCP_B; | |
2064 | else | |
2065 | /* binary search increase */ | |
2066 | tp->bictcp.cnt = tp->snd_cwnd / dist; | |
2067 | } else { | |
2068 | /* slow start amd linear increase */ | |
2069 | if (tp->snd_cwnd < tp->bictcp.last_max_cwnd + BICTCP_B) | |
2070 | /* slow start */ | |
2071 | tp->bictcp.cnt = tp->snd_cwnd * BICTCP_FUNC_OF_MIN_INCR | |
2072 | / BICTCP_B; | |
2073 | else if (tp->snd_cwnd < tp->bictcp.last_max_cwnd | |
2074 | + BICTCP_MAX_INCREMENT*(BICTCP_B-1)) | |
2075 | /* slow start */ | |
2076 | tp->bictcp.cnt = tp->snd_cwnd * (BICTCP_B-1) | |
2077 | / (tp->snd_cwnd-tp->bictcp.last_max_cwnd); | |
2078 | else | |
2079 | /* linear increase */ | |
2080 | tp->bictcp.cnt = tp->snd_cwnd / BICTCP_MAX_INCREMENT; | |
2081 | } | |
2082 | return tp->bictcp.cnt; | |
2083 | } | |
2084 | ||
2085 | /* This is Jacobson's slow start and congestion avoidance. | |
2086 | * SIGCOMM '88, p. 328. | |
2087 | */ | |
2088 | static inline void reno_cong_avoid(struct tcp_sock *tp) | |
2089 | { | |
2090 | if (tp->snd_cwnd <= tp->snd_ssthresh) { | |
2091 | /* In "safe" area, increase. */ | |
2092 | if (tp->snd_cwnd < tp->snd_cwnd_clamp) | |
2093 | tp->snd_cwnd++; | |
2094 | } else { | |
2095 | /* In dangerous area, increase slowly. | |
2096 | * In theory this is tp->snd_cwnd += 1 / tp->snd_cwnd | |
2097 | */ | |
2098 | if (tp->snd_cwnd_cnt >= bictcp_cwnd(tp)) { | |
2099 | if (tp->snd_cwnd < tp->snd_cwnd_clamp) | |
2100 | tp->snd_cwnd++; | |
2101 | tp->snd_cwnd_cnt=0; | |
2102 | } else | |
2103 | tp->snd_cwnd_cnt++; | |
2104 | } | |
2105 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
2106 | } | |
2107 | ||
2108 | /* This is based on the congestion detection/avoidance scheme described in | |
2109 | * Lawrence S. Brakmo and Larry L. Peterson. | |
2110 | * "TCP Vegas: End to end congestion avoidance on a global internet." | |
2111 | * IEEE Journal on Selected Areas in Communication, 13(8):1465--1480, | |
2112 | * October 1995. Available from: | |
2113 | * ftp://ftp.cs.arizona.edu/xkernel/Papers/jsac.ps | |
2114 | * | |
2115 | * See http://www.cs.arizona.edu/xkernel/ for their implementation. | |
2116 | * The main aspects that distinguish this implementation from the | |
2117 | * Arizona Vegas implementation are: | |
2118 | * o We do not change the loss detection or recovery mechanisms of | |
2119 | * Linux in any way. Linux already recovers from losses quite well, | |
2120 | * using fine-grained timers, NewReno, and FACK. | |
2121 | * o To avoid the performance penalty imposed by increasing cwnd | |
2122 | * only every-other RTT during slow start, we increase during | |
2123 | * every RTT during slow start, just like Reno. | |
2124 | * o Largely to allow continuous cwnd growth during slow start, | |
2125 | * we use the rate at which ACKs come back as the "actual" | |
2126 | * rate, rather than the rate at which data is sent. | |
2127 | * o To speed convergence to the right rate, we set the cwnd | |
2128 | * to achieve the right ("actual") rate when we exit slow start. | |
2129 | * o To filter out the noise caused by delayed ACKs, we use the | |
2130 | * minimum RTT sample observed during the last RTT to calculate | |
2131 | * the actual rate. | |
2132 | * o When the sender re-starts from idle, it waits until it has | |
2133 | * received ACKs for an entire flight of new data before making | |
2134 | * a cwnd adjustment decision. The original Vegas implementation | |
2135 | * assumed senders never went idle. | |
2136 | */ | |
2137 | static void vegas_cong_avoid(struct tcp_sock *tp, u32 ack, u32 seq_rtt) | |
2138 | { | |
2139 | /* The key players are v_beg_snd_una and v_beg_snd_nxt. | |
2140 | * | |
2141 | * These are so named because they represent the approximate values | |
2142 | * of snd_una and snd_nxt at the beginning of the current RTT. More | |
2143 | * precisely, they represent the amount of data sent during the RTT. | |
2144 | * At the end of the RTT, when we receive an ACK for v_beg_snd_nxt, | |
2145 | * we will calculate that (v_beg_snd_nxt - v_beg_snd_una) outstanding | |
2146 | * bytes of data have been ACKed during the course of the RTT, giving | |
2147 | * an "actual" rate of: | |
2148 | * | |
2149 | * (v_beg_snd_nxt - v_beg_snd_una) / (rtt duration) | |
2150 | * | |
2151 | * Unfortunately, v_beg_snd_una is not exactly equal to snd_una, | |
2152 | * because delayed ACKs can cover more than one segment, so they | |
2153 | * don't line up nicely with the boundaries of RTTs. | |
2154 | * | |
2155 | * Another unfortunate fact of life is that delayed ACKs delay the | |
2156 | * advance of the left edge of our send window, so that the number | |
2157 | * of bytes we send in an RTT is often less than our cwnd will allow. | |
2158 | * So we keep track of our cwnd separately, in v_beg_snd_cwnd. | |
2159 | */ | |
2160 | ||
2161 | if (after(ack, tp->vegas.beg_snd_nxt)) { | |
2162 | /* Do the Vegas once-per-RTT cwnd adjustment. */ | |
2163 | u32 old_wnd, old_snd_cwnd; | |
2164 | ||
2165 | ||
2166 | /* Here old_wnd is essentially the window of data that was | |
2167 | * sent during the previous RTT, and has all | |
2168 | * been acknowledged in the course of the RTT that ended | |
2169 | * with the ACK we just received. Likewise, old_snd_cwnd | |
2170 | * is the cwnd during the previous RTT. | |
2171 | */ | |
2172 | old_wnd = (tp->vegas.beg_snd_nxt - tp->vegas.beg_snd_una) / | |
2173 | tp->mss_cache_std; | |
2174 | old_snd_cwnd = tp->vegas.beg_snd_cwnd; | |
2175 | ||
2176 | /* Save the extent of the current window so we can use this | |
2177 | * at the end of the next RTT. | |
2178 | */ | |
2179 | tp->vegas.beg_snd_una = tp->vegas.beg_snd_nxt; | |
2180 | tp->vegas.beg_snd_nxt = tp->snd_nxt; | |
2181 | tp->vegas.beg_snd_cwnd = tp->snd_cwnd; | |
2182 | ||
2183 | /* Take into account the current RTT sample too, to | |
2184 | * decrease the impact of delayed acks. This double counts | |
2185 | * this sample since we count it for the next window as well, | |
2186 | * but that's not too awful, since we're taking the min, | |
2187 | * rather than averaging. | |
2188 | */ | |
2189 | vegas_rtt_calc(tp, seq_rtt); | |
2190 | ||
2191 | /* We do the Vegas calculations only if we got enough RTT | |
2192 | * samples that we can be reasonably sure that we got | |
2193 | * at least one RTT sample that wasn't from a delayed ACK. | |
2194 | * If we only had 2 samples total, | |
2195 | * then that means we're getting only 1 ACK per RTT, which | |
2196 | * means they're almost certainly delayed ACKs. | |
2197 | * If we have 3 samples, we should be OK. | |
2198 | */ | |
2199 | ||
2200 | if (tp->vegas.cntRTT <= 2) { | |
2201 | /* We don't have enough RTT samples to do the Vegas | |
2202 | * calculation, so we'll behave like Reno. | |
2203 | */ | |
2204 | if (tp->snd_cwnd > tp->snd_ssthresh) | |
2205 | tp->snd_cwnd++; | |
2206 | } else { | |
2207 | u32 rtt, target_cwnd, diff; | |
2208 | ||
2209 | /* We have enough RTT samples, so, using the Vegas | |
2210 | * algorithm, we determine if we should increase or | |
2211 | * decrease cwnd, and by how much. | |
2212 | */ | |
2213 | ||
2214 | /* Pluck out the RTT we are using for the Vegas | |
2215 | * calculations. This is the min RTT seen during the | |
2216 | * last RTT. Taking the min filters out the effects | |
2217 | * of delayed ACKs, at the cost of noticing congestion | |
2218 | * a bit later. | |
2219 | */ | |
2220 | rtt = tp->vegas.minRTT; | |
2221 | ||
2222 | /* Calculate the cwnd we should have, if we weren't | |
2223 | * going too fast. | |
2224 | * | |
2225 | * This is: | |
2226 | * (actual rate in segments) * baseRTT | |
2227 | * We keep it as a fixed point number with | |
2228 | * V_PARAM_SHIFT bits to the right of the binary point. | |
2229 | */ | |
2230 | target_cwnd = ((old_wnd * tp->vegas.baseRTT) | |
2231 | << V_PARAM_SHIFT) / rtt; | |
2232 | ||
2233 | /* Calculate the difference between the window we had, | |
2234 | * and the window we would like to have. This quantity | |
2235 | * is the "Diff" from the Arizona Vegas papers. | |
2236 | * | |
2237 | * Again, this is a fixed point number with | |
2238 | * V_PARAM_SHIFT bits to the right of the binary | |
2239 | * point. | |
2240 | */ | |
2241 | diff = (old_wnd << V_PARAM_SHIFT) - target_cwnd; | |
2242 | ||
2243 | if (tp->snd_cwnd < tp->snd_ssthresh) { | |
2244 | /* Slow start. */ | |
2245 | if (diff > sysctl_tcp_vegas_gamma) { | |
2246 | /* Going too fast. Time to slow down | |
2247 | * and switch to congestion avoidance. | |
2248 | */ | |
2249 | tp->snd_ssthresh = 2; | |
2250 | ||
2251 | /* Set cwnd to match the actual rate | |
2252 | * exactly: | |
2253 | * cwnd = (actual rate) * baseRTT | |
2254 | * Then we add 1 because the integer | |
2255 | * truncation robs us of full link | |
2256 | * utilization. | |
2257 | */ | |
2258 | tp->snd_cwnd = min(tp->snd_cwnd, | |
2259 | (target_cwnd >> | |
2260 | V_PARAM_SHIFT)+1); | |
2261 | ||
2262 | } | |
2263 | } else { | |
2264 | /* Congestion avoidance. */ | |
2265 | u32 next_snd_cwnd; | |
2266 | ||
2267 | /* Figure out where we would like cwnd | |
2268 | * to be. | |
2269 | */ | |
2270 | if (diff > sysctl_tcp_vegas_beta) { | |
2271 | /* The old window was too fast, so | |
2272 | * we slow down. | |
2273 | */ | |
2274 | next_snd_cwnd = old_snd_cwnd - 1; | |
2275 | } else if (diff < sysctl_tcp_vegas_alpha) { | |
2276 | /* We don't have enough extra packets | |
2277 | * in the network, so speed up. | |
2278 | */ | |
2279 | next_snd_cwnd = old_snd_cwnd + 1; | |
2280 | } else { | |
2281 | /* Sending just as fast as we | |
2282 | * should be. | |
2283 | */ | |
2284 | next_snd_cwnd = old_snd_cwnd; | |
2285 | } | |
2286 | ||
2287 | /* Adjust cwnd upward or downward, toward the | |
2288 | * desired value. | |
2289 | */ | |
2290 | if (next_snd_cwnd > tp->snd_cwnd) | |
2291 | tp->snd_cwnd++; | |
2292 | else if (next_snd_cwnd < tp->snd_cwnd) | |
2293 | tp->snd_cwnd--; | |
2294 | } | |
2295 | } | |
2296 | ||
2297 | /* Wipe the slate clean for the next RTT. */ | |
2298 | tp->vegas.cntRTT = 0; | |
2299 | tp->vegas.minRTT = 0x7fffffff; | |
2300 | } | |
2301 | ||
2302 | /* The following code is executed for every ack we receive, | |
2303 | * except for conditions checked in should_advance_cwnd() | |
2304 | * before the call to tcp_cong_avoid(). Mainly this means that | |
2305 | * we only execute this code if the ack actually acked some | |
2306 | * data. | |
2307 | */ | |
2308 | ||
2309 | /* If we are in slow start, increase our cwnd in response to this ACK. | |
2310 | * (If we are not in slow start then we are in congestion avoidance, | |
2311 | * and adjust our congestion window only once per RTT. See the code | |
2312 | * above.) | |
2313 | */ | |
2314 | if (tp->snd_cwnd <= tp->snd_ssthresh) | |
2315 | tp->snd_cwnd++; | |
2316 | ||
2317 | /* to keep cwnd from growing without bound */ | |
2318 | tp->snd_cwnd = min_t(u32, tp->snd_cwnd, tp->snd_cwnd_clamp); | |
2319 | ||
2320 | /* Make sure that we are never so timid as to reduce our cwnd below | |
2321 | * 2 MSS. | |
2322 | * | |
2323 | * Going below 2 MSS would risk huge delayed ACKs from our receiver. | |
2324 | */ | |
2325 | tp->snd_cwnd = max(tp->snd_cwnd, 2U); | |
2326 | ||
2327 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
2328 | } | |
2329 | ||
2330 | static inline void tcp_cong_avoid(struct tcp_sock *tp, u32 ack, u32 seq_rtt) | |
2331 | { | |
2332 | if (tcp_vegas_enabled(tp)) | |
2333 | vegas_cong_avoid(tp, ack, seq_rtt); | |
2334 | else | |
2335 | reno_cong_avoid(tp); | |
2336 | } | |
2337 | ||
2338 | /* Restart timer after forward progress on connection. | |
2339 | * RFC2988 recommends to restart timer to now+rto. | |
2340 | */ | |
2341 | ||
2342 | static inline void tcp_ack_packets_out(struct sock *sk, struct tcp_sock *tp) | |
2343 | { | |
2344 | if (!tp->packets_out) { | |
2345 | tcp_clear_xmit_timer(sk, TCP_TIME_RETRANS); | |
2346 | } else { | |
2347 | tcp_reset_xmit_timer(sk, TCP_TIME_RETRANS, tp->rto); | |
2348 | } | |
2349 | } | |
2350 | ||
2351 | /* There is one downside to this scheme. Although we keep the | |
2352 | * ACK clock ticking, adjusting packet counters and advancing | |
2353 | * congestion window, we do not liberate socket send buffer | |
2354 | * space. | |
2355 | * | |
2356 | * Mucking with skb->truesize and sk->sk_wmem_alloc et al. | |
2357 | * then making a write space wakeup callback is a possible | |
2358 | * future enhancement. WARNING: it is not trivial to make. | |
2359 | */ | |
2360 | static int tcp_tso_acked(struct sock *sk, struct sk_buff *skb, | |
2361 | __u32 now, __s32 *seq_rtt) | |
2362 | { | |
2363 | struct tcp_sock *tp = tcp_sk(sk); | |
2364 | struct tcp_skb_cb *scb = TCP_SKB_CB(skb); | |
2365 | __u32 seq = tp->snd_una; | |
2366 | __u32 packets_acked; | |
2367 | int acked = 0; | |
2368 | ||
2369 | /* If we get here, the whole TSO packet has not been | |
2370 | * acked. | |
2371 | */ | |
2372 | BUG_ON(!after(scb->end_seq, seq)); | |
2373 | ||
2374 | packets_acked = tcp_skb_pcount(skb); | |
2375 | if (tcp_trim_head(sk, skb, seq - scb->seq)) | |
2376 | return 0; | |
2377 | packets_acked -= tcp_skb_pcount(skb); | |
2378 | ||
2379 | if (packets_acked) { | |
2380 | __u8 sacked = scb->sacked; | |
2381 | ||
2382 | acked |= FLAG_DATA_ACKED; | |
2383 | if (sacked) { | |
2384 | if (sacked & TCPCB_RETRANS) { | |
2385 | if (sacked & TCPCB_SACKED_RETRANS) | |
2386 | tp->retrans_out -= packets_acked; | |
2387 | acked |= FLAG_RETRANS_DATA_ACKED; | |
2388 | *seq_rtt = -1; | |
2389 | } else if (*seq_rtt < 0) | |
2390 | *seq_rtt = now - scb->when; | |
2391 | if (sacked & TCPCB_SACKED_ACKED) | |
2392 | tp->sacked_out -= packets_acked; | |
2393 | if (sacked & TCPCB_LOST) | |
2394 | tp->lost_out -= packets_acked; | |
2395 | if (sacked & TCPCB_URG) { | |
2396 | if (tp->urg_mode && | |
2397 | !before(seq, tp->snd_up)) | |
2398 | tp->urg_mode = 0; | |
2399 | } | |
2400 | } else if (*seq_rtt < 0) | |
2401 | *seq_rtt = now - scb->when; | |
2402 | ||
2403 | if (tp->fackets_out) { | |
2404 | __u32 dval = min(tp->fackets_out, packets_acked); | |
2405 | tp->fackets_out -= dval; | |
2406 | } | |
2407 | tp->packets_out -= packets_acked; | |
2408 | ||
2409 | BUG_ON(tcp_skb_pcount(skb) == 0); | |
2410 | BUG_ON(!before(scb->seq, scb->end_seq)); | |
2411 | } | |
2412 | ||
2413 | return acked; | |
2414 | } | |
2415 | ||
2416 | ||
2417 | /* Remove acknowledged frames from the retransmission queue. */ | |
2418 | static int tcp_clean_rtx_queue(struct sock *sk, __s32 *seq_rtt_p) | |
2419 | { | |
2420 | struct tcp_sock *tp = tcp_sk(sk); | |
2421 | struct sk_buff *skb; | |
2422 | __u32 now = tcp_time_stamp; | |
2423 | int acked = 0; | |
2424 | __s32 seq_rtt = -1; | |
2425 | ||
2426 | while ((skb = skb_peek(&sk->sk_write_queue)) && | |
2427 | skb != sk->sk_send_head) { | |
2428 | struct tcp_skb_cb *scb = TCP_SKB_CB(skb); | |
2429 | __u8 sacked = scb->sacked; | |
2430 | ||
2431 | /* If our packet is before the ack sequence we can | |
2432 | * discard it as it's confirmed to have arrived at | |
2433 | * the other end. | |
2434 | */ | |
2435 | if (after(scb->end_seq, tp->snd_una)) { | |
2436 | if (tcp_skb_pcount(skb) > 1) | |
2437 | acked |= tcp_tso_acked(sk, skb, | |
2438 | now, &seq_rtt); | |
2439 | break; | |
2440 | } | |
2441 | ||
2442 | /* Initial outgoing SYN's get put onto the write_queue | |
2443 | * just like anything else we transmit. It is not | |
2444 | * true data, and if we misinform our callers that | |
2445 | * this ACK acks real data, we will erroneously exit | |
2446 | * connection startup slow start one packet too | |
2447 | * quickly. This is severely frowned upon behavior. | |
2448 | */ | |
2449 | if (!(scb->flags & TCPCB_FLAG_SYN)) { | |
2450 | acked |= FLAG_DATA_ACKED; | |
2451 | } else { | |
2452 | acked |= FLAG_SYN_ACKED; | |
2453 | tp->retrans_stamp = 0; | |
2454 | } | |
2455 | ||
2456 | if (sacked) { | |
2457 | if (sacked & TCPCB_RETRANS) { | |
2458 | if(sacked & TCPCB_SACKED_RETRANS) | |
2459 | tp->retrans_out -= tcp_skb_pcount(skb); | |
2460 | acked |= FLAG_RETRANS_DATA_ACKED; | |
2461 | seq_rtt = -1; | |
2462 | } else if (seq_rtt < 0) | |
2463 | seq_rtt = now - scb->when; | |
2464 | if (sacked & TCPCB_SACKED_ACKED) | |
2465 | tp->sacked_out -= tcp_skb_pcount(skb); | |
2466 | if (sacked & TCPCB_LOST) | |
2467 | tp->lost_out -= tcp_skb_pcount(skb); | |
2468 | if (sacked & TCPCB_URG) { | |
2469 | if (tp->urg_mode && | |
2470 | !before(scb->end_seq, tp->snd_up)) | |
2471 | tp->urg_mode = 0; | |
2472 | } | |
2473 | } else if (seq_rtt < 0) | |
2474 | seq_rtt = now - scb->when; | |
2475 | tcp_dec_pcount_approx(&tp->fackets_out, skb); | |
2476 | tcp_packets_out_dec(tp, skb); | |
2477 | __skb_unlink(skb, skb->list); | |
2478 | sk_stream_free_skb(sk, skb); | |
2479 | } | |
2480 | ||
2481 | if (acked&FLAG_ACKED) { | |
2482 | tcp_ack_update_rtt(tp, acked, seq_rtt); | |
2483 | tcp_ack_packets_out(sk, tp); | |
2484 | } | |
2485 | ||
2486 | #if FASTRETRANS_DEBUG > 0 | |
2487 | BUG_TRAP((int)tp->sacked_out >= 0); | |
2488 | BUG_TRAP((int)tp->lost_out >= 0); | |
2489 | BUG_TRAP((int)tp->retrans_out >= 0); | |
2490 | if (!tp->packets_out && tp->rx_opt.sack_ok) { | |
2491 | if (tp->lost_out) { | |
2492 | printk(KERN_DEBUG "Leak l=%u %d\n", | |
2493 | tp->lost_out, tp->ca_state); | |
2494 | tp->lost_out = 0; | |
2495 | } | |
2496 | if (tp->sacked_out) { | |
2497 | printk(KERN_DEBUG "Leak s=%u %d\n", | |
2498 | tp->sacked_out, tp->ca_state); | |
2499 | tp->sacked_out = 0; | |
2500 | } | |
2501 | if (tp->retrans_out) { | |
2502 | printk(KERN_DEBUG "Leak r=%u %d\n", | |
2503 | tp->retrans_out, tp->ca_state); | |
2504 | tp->retrans_out = 0; | |
2505 | } | |
2506 | } | |
2507 | #endif | |
2508 | *seq_rtt_p = seq_rtt; | |
2509 | return acked; | |
2510 | } | |
2511 | ||
2512 | static void tcp_ack_probe(struct sock *sk) | |
2513 | { | |
2514 | struct tcp_sock *tp = tcp_sk(sk); | |
2515 | ||
2516 | /* Was it a usable window open? */ | |
2517 | ||
2518 | if (!after(TCP_SKB_CB(sk->sk_send_head)->end_seq, | |
2519 | tp->snd_una + tp->snd_wnd)) { | |
2520 | tp->backoff = 0; | |
2521 | tcp_clear_xmit_timer(sk, TCP_TIME_PROBE0); | |
2522 | /* Socket must be waked up by subsequent tcp_data_snd_check(). | |
2523 | * This function is not for random using! | |
2524 | */ | |
2525 | } else { | |
2526 | tcp_reset_xmit_timer(sk, TCP_TIME_PROBE0, | |
2527 | min(tp->rto << tp->backoff, TCP_RTO_MAX)); | |
2528 | } | |
2529 | } | |
2530 | ||
2531 | static inline int tcp_ack_is_dubious(struct tcp_sock *tp, int flag) | |
2532 | { | |
2533 | return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || | |
2534 | tp->ca_state != TCP_CA_Open); | |
2535 | } | |
2536 | ||
2537 | static inline int tcp_may_raise_cwnd(struct tcp_sock *tp, int flag) | |
2538 | { | |
2539 | return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && | |
2540 | !((1<<tp->ca_state)&(TCPF_CA_Recovery|TCPF_CA_CWR)); | |
2541 | } | |
2542 | ||
2543 | /* Check that window update is acceptable. | |
2544 | * The function assumes that snd_una<=ack<=snd_next. | |
2545 | */ | |
2546 | static inline int tcp_may_update_window(struct tcp_sock *tp, u32 ack, | |
2547 | u32 ack_seq, u32 nwin) | |
2548 | { | |
2549 | return (after(ack, tp->snd_una) || | |
2550 | after(ack_seq, tp->snd_wl1) || | |
2551 | (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd)); | |
2552 | } | |
2553 | ||
2554 | /* Update our send window. | |
2555 | * | |
2556 | * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 | |
2557 | * and in FreeBSD. NetBSD's one is even worse.) is wrong. | |
2558 | */ | |
2559 | static int tcp_ack_update_window(struct sock *sk, struct tcp_sock *tp, | |
2560 | struct sk_buff *skb, u32 ack, u32 ack_seq) | |
2561 | { | |
2562 | int flag = 0; | |
2563 | u32 nwin = ntohs(skb->h.th->window); | |
2564 | ||
2565 | if (likely(!skb->h.th->syn)) | |
2566 | nwin <<= tp->rx_opt.snd_wscale; | |
2567 | ||
2568 | if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { | |
2569 | flag |= FLAG_WIN_UPDATE; | |
2570 | tcp_update_wl(tp, ack, ack_seq); | |
2571 | ||
2572 | if (tp->snd_wnd != nwin) { | |
2573 | tp->snd_wnd = nwin; | |
2574 | ||
2575 | /* Note, it is the only place, where | |
2576 | * fast path is recovered for sending TCP. | |
2577 | */ | |
2578 | tcp_fast_path_check(sk, tp); | |
2579 | ||
2580 | if (nwin > tp->max_window) { | |
2581 | tp->max_window = nwin; | |
2582 | tcp_sync_mss(sk, tp->pmtu_cookie); | |
2583 | } | |
2584 | } | |
2585 | } | |
2586 | ||
2587 | tp->snd_una = ack; | |
2588 | ||
2589 | return flag; | |
2590 | } | |
2591 | ||
2592 | static void tcp_process_frto(struct sock *sk, u32 prior_snd_una) | |
2593 | { | |
2594 | struct tcp_sock *tp = tcp_sk(sk); | |
2595 | ||
2596 | tcp_sync_left_out(tp); | |
2597 | ||
2598 | if (tp->snd_una == prior_snd_una || | |
2599 | !before(tp->snd_una, tp->frto_highmark)) { | |
2600 | /* RTO was caused by loss, start retransmitting in | |
2601 | * go-back-N slow start | |
2602 | */ | |
2603 | tcp_enter_frto_loss(sk); | |
2604 | return; | |
2605 | } | |
2606 | ||
2607 | if (tp->frto_counter == 1) { | |
2608 | /* First ACK after RTO advances the window: allow two new | |
2609 | * segments out. | |
2610 | */ | |
2611 | tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; | |
2612 | } else { | |
2613 | /* Also the second ACK after RTO advances the window. | |
2614 | * The RTO was likely spurious. Reduce cwnd and continue | |
2615 | * in congestion avoidance | |
2616 | */ | |
2617 | tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); | |
2618 | tcp_moderate_cwnd(tp); | |
2619 | } | |
2620 | ||
2621 | /* F-RTO affects on two new ACKs following RTO. | |
2622 | * At latest on third ACK the TCP behavor is back to normal. | |
2623 | */ | |
2624 | tp->frto_counter = (tp->frto_counter + 1) % 3; | |
2625 | } | |
2626 | ||
2627 | /* | |
2628 | * TCP Westwood+ | |
2629 | */ | |
2630 | ||
2631 | /* | |
2632 | * @init_westwood | |
2633 | * This function initializes fields used in TCP Westwood+. We can't | |
2634 | * get no information about RTTmin at this time so we simply set it to | |
2635 | * TCP_WESTWOOD_INIT_RTT. This value was chosen to be too conservative | |
2636 | * since in this way we're sure it will be updated in a consistent | |
2637 | * way as soon as possible. It will reasonably happen within the first | |
2638 | * RTT period of the connection lifetime. | |
2639 | */ | |
2640 | ||
2641 | static void init_westwood(struct sock *sk) | |
2642 | { | |
2643 | struct tcp_sock *tp = tcp_sk(sk); | |
2644 | ||
2645 | tp->westwood.bw_ns_est = 0; | |
2646 | tp->westwood.bw_est = 0; | |
2647 | tp->westwood.accounted = 0; | |
2648 | tp->westwood.cumul_ack = 0; | |
2649 | tp->westwood.rtt_win_sx = tcp_time_stamp; | |
2650 | tp->westwood.rtt = TCP_WESTWOOD_INIT_RTT; | |
2651 | tp->westwood.rtt_min = TCP_WESTWOOD_INIT_RTT; | |
2652 | tp->westwood.snd_una = tp->snd_una; | |
2653 | } | |
2654 | ||
2655 | /* | |
2656 | * @westwood_do_filter | |
2657 | * Low-pass filter. Implemented using constant coeffients. | |
2658 | */ | |
2659 | ||
2660 | static inline __u32 westwood_do_filter(__u32 a, __u32 b) | |
2661 | { | |
2662 | return (((7 * a) + b) >> 3); | |
2663 | } | |
2664 | ||
2665 | static void westwood_filter(struct sock *sk, __u32 delta) | |
2666 | { | |
2667 | struct tcp_sock *tp = tcp_sk(sk); | |
2668 | ||
2669 | tp->westwood.bw_ns_est = | |
2670 | westwood_do_filter(tp->westwood.bw_ns_est, | |
2671 | tp->westwood.bk / delta); | |
2672 | tp->westwood.bw_est = | |
2673 | westwood_do_filter(tp->westwood.bw_est, | |
2674 | tp->westwood.bw_ns_est); | |
2675 | } | |
2676 | ||
2677 | /* | |
2678 | * @westwood_update_rttmin | |
2679 | * It is used to update RTTmin. In this case we MUST NOT use | |
2680 | * WESTWOOD_RTT_MIN minimum bound since we could be on a LAN! | |
2681 | */ | |
2682 | ||
2683 | static inline __u32 westwood_update_rttmin(const struct sock *sk) | |
2684 | { | |
2685 | const struct tcp_sock *tp = tcp_sk(sk); | |
2686 | __u32 rttmin = tp->westwood.rtt_min; | |
2687 | ||
2688 | if (tp->westwood.rtt != 0 && | |
2689 | (tp->westwood.rtt < tp->westwood.rtt_min || !rttmin)) | |
2690 | rttmin = tp->westwood.rtt; | |
2691 | ||
2692 | return rttmin; | |
2693 | } | |
2694 | ||
2695 | /* | |
2696 | * @westwood_acked | |
2697 | * Evaluate increases for dk. | |
2698 | */ | |
2699 | ||
2700 | static inline __u32 westwood_acked(const struct sock *sk) | |
2701 | { | |
2702 | const struct tcp_sock *tp = tcp_sk(sk); | |
2703 | ||
2704 | return tp->snd_una - tp->westwood.snd_una; | |
2705 | } | |
2706 | ||
2707 | /* | |
2708 | * @westwood_new_window | |
2709 | * It evaluates if we are receiving data inside the same RTT window as | |
2710 | * when we started. | |
2711 | * Return value: | |
2712 | * It returns 0 if we are still evaluating samples in the same RTT | |
2713 | * window, 1 if the sample has to be considered in the next window. | |
2714 | */ | |
2715 | ||
2716 | static int westwood_new_window(const struct sock *sk) | |
2717 | { | |
2718 | const struct tcp_sock *tp = tcp_sk(sk); | |
2719 | __u32 left_bound; | |
2720 | __u32 rtt; | |
2721 | int ret = 0; | |
2722 | ||
2723 | left_bound = tp->westwood.rtt_win_sx; | |
2724 | rtt = max(tp->westwood.rtt, (u32) TCP_WESTWOOD_RTT_MIN); | |
2725 | ||
2726 | /* | |
2727 | * A RTT-window has passed. Be careful since if RTT is less than | |
2728 | * 50ms we don't filter but we continue 'building the sample'. | |
2729 | * This minimum limit was choosen since an estimation on small | |
2730 | * time intervals is better to avoid... | |
2731 | * Obvioulsy on a LAN we reasonably will always have | |
2732 | * right_bound = left_bound + WESTWOOD_RTT_MIN | |
2733 | */ | |
2734 | ||
2735 | if ((left_bound + rtt) < tcp_time_stamp) | |
2736 | ret = 1; | |
2737 | ||
2738 | return ret; | |
2739 | } | |
2740 | ||
2741 | /* | |
2742 | * @westwood_update_window | |
2743 | * It updates RTT evaluation window if it is the right moment to do | |
2744 | * it. If so it calls filter for evaluating bandwidth. | |
2745 | */ | |
2746 | ||
2747 | static void __westwood_update_window(struct sock *sk, __u32 now) | |
2748 | { | |
2749 | struct tcp_sock *tp = tcp_sk(sk); | |
2750 | __u32 delta = now - tp->westwood.rtt_win_sx; | |
2751 | ||
2752 | if (delta) { | |
2753 | if (tp->westwood.rtt) | |
2754 | westwood_filter(sk, delta); | |
2755 | ||
2756 | tp->westwood.bk = 0; | |
2757 | tp->westwood.rtt_win_sx = tcp_time_stamp; | |
2758 | } | |
2759 | } | |
2760 | ||
2761 | ||
2762 | static void westwood_update_window(struct sock *sk, __u32 now) | |
2763 | { | |
2764 | if (westwood_new_window(sk)) | |
2765 | __westwood_update_window(sk, now); | |
2766 | } | |
2767 | ||
2768 | /* | |
2769 | * @__tcp_westwood_fast_bw | |
2770 | * It is called when we are in fast path. In particular it is called when | |
2771 | * header prediction is successfull. In such case infact update is | |
2772 | * straight forward and doesn't need any particular care. | |
2773 | */ | |
2774 | ||
2775 | static void __tcp_westwood_fast_bw(struct sock *sk, struct sk_buff *skb) | |
2776 | { | |
2777 | struct tcp_sock *tp = tcp_sk(sk); | |
2778 | ||
2779 | westwood_update_window(sk, tcp_time_stamp); | |
2780 | ||
2781 | tp->westwood.bk += westwood_acked(sk); | |
2782 | tp->westwood.snd_una = tp->snd_una; | |
2783 | tp->westwood.rtt_min = westwood_update_rttmin(sk); | |
2784 | } | |
2785 | ||
2786 | static inline void tcp_westwood_fast_bw(struct sock *sk, struct sk_buff *skb) | |
2787 | { | |
2788 | if (tcp_is_westwood(tcp_sk(sk))) | |
2789 | __tcp_westwood_fast_bw(sk, skb); | |
2790 | } | |
2791 | ||
2792 | ||
2793 | /* | |
2794 | * @westwood_dupack_update | |
2795 | * It updates accounted and cumul_ack when receiving a dupack. | |
2796 | */ | |
2797 | ||
2798 | static void westwood_dupack_update(struct sock *sk) | |
2799 | { | |
2800 | struct tcp_sock *tp = tcp_sk(sk); | |
2801 | ||
2802 | tp->westwood.accounted += tp->mss_cache_std; | |
2803 | tp->westwood.cumul_ack = tp->mss_cache_std; | |
2804 | } | |
2805 | ||
2806 | static inline int westwood_may_change_cumul(struct tcp_sock *tp) | |
2807 | { | |
2808 | return (tp->westwood.cumul_ack > tp->mss_cache_std); | |
2809 | } | |
2810 | ||
2811 | static inline void westwood_partial_update(struct tcp_sock *tp) | |
2812 | { | |
2813 | tp->westwood.accounted -= tp->westwood.cumul_ack; | |
2814 | tp->westwood.cumul_ack = tp->mss_cache_std; | |
2815 | } | |
2816 | ||
2817 | static inline void westwood_complete_update(struct tcp_sock *tp) | |
2818 | { | |
2819 | tp->westwood.cumul_ack -= tp->westwood.accounted; | |
2820 | tp->westwood.accounted = 0; | |
2821 | } | |
2822 | ||
2823 | /* | |
2824 | * @westwood_acked_count | |
2825 | * This function evaluates cumul_ack for evaluating dk in case of | |
2826 | * delayed or partial acks. | |
2827 | */ | |
2828 | ||
2829 | static inline __u32 westwood_acked_count(struct sock *sk) | |
2830 | { | |
2831 | struct tcp_sock *tp = tcp_sk(sk); | |
2832 | ||
2833 | tp->westwood.cumul_ack = westwood_acked(sk); | |
2834 | ||
2835 | /* If cumul_ack is 0 this is a dupack since it's not moving | |
2836 | * tp->snd_una. | |
2837 | */ | |
2838 | if (!(tp->westwood.cumul_ack)) | |
2839 | westwood_dupack_update(sk); | |
2840 | ||
2841 | if (westwood_may_change_cumul(tp)) { | |
2842 | /* Partial or delayed ack */ | |
2843 | if (tp->westwood.accounted >= tp->westwood.cumul_ack) | |
2844 | westwood_partial_update(tp); | |
2845 | else | |
2846 | westwood_complete_update(tp); | |
2847 | } | |
2848 | ||
2849 | tp->westwood.snd_una = tp->snd_una; | |
2850 | ||
2851 | return tp->westwood.cumul_ack; | |
2852 | } | |
2853 | ||
2854 | ||
2855 | /* | |
2856 | * @__tcp_westwood_slow_bw | |
2857 | * It is called when something is going wrong..even if there could | |
2858 | * be no problems! Infact a simple delayed packet may trigger a | |
2859 | * dupack. But we need to be careful in such case. | |
2860 | */ | |
2861 | ||
2862 | static void __tcp_westwood_slow_bw(struct sock *sk, struct sk_buff *skb) | |
2863 | { | |
2864 | struct tcp_sock *tp = tcp_sk(sk); | |
2865 | ||
2866 | westwood_update_window(sk, tcp_time_stamp); | |
2867 | ||
2868 | tp->westwood.bk += westwood_acked_count(sk); | |
2869 | tp->westwood.rtt_min = westwood_update_rttmin(sk); | |
2870 | } | |
2871 | ||
2872 | static inline void tcp_westwood_slow_bw(struct sock *sk, struct sk_buff *skb) | |
2873 | { | |
2874 | if (tcp_is_westwood(tcp_sk(sk))) | |
2875 | __tcp_westwood_slow_bw(sk, skb); | |
2876 | } | |
2877 | ||
2878 | /* This routine deals with incoming acks, but not outgoing ones. */ | |
2879 | static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag) | |
2880 | { | |
2881 | struct tcp_sock *tp = tcp_sk(sk); | |
2882 | u32 prior_snd_una = tp->snd_una; | |
2883 | u32 ack_seq = TCP_SKB_CB(skb)->seq; | |
2884 | u32 ack = TCP_SKB_CB(skb)->ack_seq; | |
2885 | u32 prior_in_flight; | |
2886 | s32 seq_rtt; | |
2887 | int prior_packets; | |
2888 | ||
2889 | /* If the ack is newer than sent or older than previous acks | |
2890 | * then we can probably ignore it. | |
2891 | */ | |
2892 | if (after(ack, tp->snd_nxt)) | |
2893 | goto uninteresting_ack; | |
2894 | ||
2895 | if (before(ack, prior_snd_una)) | |
2896 | goto old_ack; | |
2897 | ||
2898 | if (!(flag&FLAG_SLOWPATH) && after(ack, prior_snd_una)) { | |
2899 | /* Window is constant, pure forward advance. | |
2900 | * No more checks are required. | |
2901 | * Note, we use the fact that SND.UNA>=SND.WL2. | |
2902 | */ | |
2903 | tcp_update_wl(tp, ack, ack_seq); | |
2904 | tp->snd_una = ack; | |
2905 | tcp_westwood_fast_bw(sk, skb); | |
2906 | flag |= FLAG_WIN_UPDATE; | |
2907 | ||
2908 | NET_INC_STATS_BH(LINUX_MIB_TCPHPACKS); | |
2909 | } else { | |
2910 | if (ack_seq != TCP_SKB_CB(skb)->end_seq) | |
2911 | flag |= FLAG_DATA; | |
2912 | else | |
2913 | NET_INC_STATS_BH(LINUX_MIB_TCPPUREACKS); | |
2914 | ||
2915 | flag |= tcp_ack_update_window(sk, tp, skb, ack, ack_seq); | |
2916 | ||
2917 | if (TCP_SKB_CB(skb)->sacked) | |
2918 | flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); | |
2919 | ||
2920 | if (TCP_ECN_rcv_ecn_echo(tp, skb->h.th)) | |
2921 | flag |= FLAG_ECE; | |
2922 | ||
2923 | tcp_westwood_slow_bw(sk,skb); | |
2924 | } | |
2925 | ||
2926 | /* We passed data and got it acked, remove any soft error | |
2927 | * log. Something worked... | |
2928 | */ | |
2929 | sk->sk_err_soft = 0; | |
2930 | tp->rcv_tstamp = tcp_time_stamp; | |
2931 | prior_packets = tp->packets_out; | |
2932 | if (!prior_packets) | |
2933 | goto no_queue; | |
2934 | ||
2935 | prior_in_flight = tcp_packets_in_flight(tp); | |
2936 | ||
2937 | /* See if we can take anything off of the retransmit queue. */ | |
2938 | flag |= tcp_clean_rtx_queue(sk, &seq_rtt); | |
2939 | ||
2940 | if (tp->frto_counter) | |
2941 | tcp_process_frto(sk, prior_snd_una); | |
2942 | ||
2943 | if (tcp_ack_is_dubious(tp, flag)) { | |
2944 | /* Advanve CWND, if state allows this. */ | |
2945 | if ((flag & FLAG_DATA_ACKED) && | |
2946 | (tcp_vegas_enabled(tp) || prior_in_flight >= tp->snd_cwnd) && | |
2947 | tcp_may_raise_cwnd(tp, flag)) | |
2948 | tcp_cong_avoid(tp, ack, seq_rtt); | |
2949 | tcp_fastretrans_alert(sk, prior_snd_una, prior_packets, flag); | |
2950 | } else { | |
2951 | if ((flag & FLAG_DATA_ACKED) && | |
2952 | (tcp_vegas_enabled(tp) || prior_in_flight >= tp->snd_cwnd)) | |
2953 | tcp_cong_avoid(tp, ack, seq_rtt); | |
2954 | } | |
2955 | ||
2956 | if ((flag & FLAG_FORWARD_PROGRESS) || !(flag&FLAG_NOT_DUP)) | |
2957 | dst_confirm(sk->sk_dst_cache); | |
2958 | ||
2959 | return 1; | |
2960 | ||
2961 | no_queue: | |
2962 | tp->probes_out = 0; | |
2963 | ||
2964 | /* If this ack opens up a zero window, clear backoff. It was | |
2965 | * being used to time the probes, and is probably far higher than | |
2966 | * it needs to be for normal retransmission. | |
2967 | */ | |
2968 | if (sk->sk_send_head) | |
2969 | tcp_ack_probe(sk); | |
2970 | return 1; | |
2971 | ||
2972 | old_ack: | |
2973 | if (TCP_SKB_CB(skb)->sacked) | |
2974 | tcp_sacktag_write_queue(sk, skb, prior_snd_una); | |
2975 | ||
2976 | uninteresting_ack: | |
2977 | SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt); | |
2978 | return 0; | |
2979 | } | |
2980 | ||
2981 | ||
2982 | /* Look for tcp options. Normally only called on SYN and SYNACK packets. | |
2983 | * But, this can also be called on packets in the established flow when | |
2984 | * the fast version below fails. | |
2985 | */ | |
2986 | void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, int estab) | |
2987 | { | |
2988 | unsigned char *ptr; | |
2989 | struct tcphdr *th = skb->h.th; | |
2990 | int length=(th->doff*4)-sizeof(struct tcphdr); | |
2991 | ||
2992 | ptr = (unsigned char *)(th + 1); | |
2993 | opt_rx->saw_tstamp = 0; | |
2994 | ||
2995 | while(length>0) { | |
2996 | int opcode=*ptr++; | |
2997 | int opsize; | |
2998 | ||
2999 | switch (opcode) { | |
3000 | case TCPOPT_EOL: | |
3001 | return; | |
3002 | case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ | |
3003 | length--; | |
3004 | continue; | |
3005 | default: | |
3006 | opsize=*ptr++; | |
3007 | if (opsize < 2) /* "silly options" */ | |
3008 | return; | |
3009 | if (opsize > length) | |
3010 | return; /* don't parse partial options */ | |
3011 | switch(opcode) { | |
3012 | case TCPOPT_MSS: | |
3013 | if(opsize==TCPOLEN_MSS && th->syn && !estab) { | |
3014 | u16 in_mss = ntohs(get_unaligned((__u16 *)ptr)); | |
3015 | if (in_mss) { | |
3016 | if (opt_rx->user_mss && opt_rx->user_mss < in_mss) | |
3017 | in_mss = opt_rx->user_mss; | |
3018 | opt_rx->mss_clamp = in_mss; | |
3019 | } | |
3020 | } | |
3021 | break; | |
3022 | case TCPOPT_WINDOW: | |
3023 | if(opsize==TCPOLEN_WINDOW && th->syn && !estab) | |
3024 | if (sysctl_tcp_window_scaling) { | |
3025 | __u8 snd_wscale = *(__u8 *) ptr; | |
3026 | opt_rx->wscale_ok = 1; | |
3027 | if (snd_wscale > 14) { | |
3028 | if(net_ratelimit()) | |
3029 | printk(KERN_INFO "tcp_parse_options: Illegal window " | |
3030 | "scaling value %d >14 received.\n", | |
3031 | snd_wscale); | |
3032 | snd_wscale = 14; | |
3033 | } | |
3034 | opt_rx->snd_wscale = snd_wscale; | |
3035 | } | |
3036 | break; | |
3037 | case TCPOPT_TIMESTAMP: | |
3038 | if(opsize==TCPOLEN_TIMESTAMP) { | |
3039 | if ((estab && opt_rx->tstamp_ok) || | |
3040 | (!estab && sysctl_tcp_timestamps)) { | |
3041 | opt_rx->saw_tstamp = 1; | |
3042 | opt_rx->rcv_tsval = ntohl(get_unaligned((__u32 *)ptr)); | |
3043 | opt_rx->rcv_tsecr = ntohl(get_unaligned((__u32 *)(ptr+4))); | |
3044 | } | |
3045 | } | |
3046 | break; | |
3047 | case TCPOPT_SACK_PERM: | |
3048 | if(opsize==TCPOLEN_SACK_PERM && th->syn && !estab) { | |
3049 | if (sysctl_tcp_sack) { | |
3050 | opt_rx->sack_ok = 1; | |
3051 | tcp_sack_reset(opt_rx); | |
3052 | } | |
3053 | } | |
3054 | break; | |
3055 | ||
3056 | case TCPOPT_SACK: | |
3057 | if((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && | |
3058 | !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && | |
3059 | opt_rx->sack_ok) { | |
3060 | TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; | |
3061 | } | |
3062 | }; | |
3063 | ptr+=opsize-2; | |
3064 | length-=opsize; | |
3065 | }; | |
3066 | } | |
3067 | } | |
3068 | ||
3069 | /* Fast parse options. This hopes to only see timestamps. | |
3070 | * If it is wrong it falls back on tcp_parse_options(). | |
3071 | */ | |
3072 | static inline int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, | |
3073 | struct tcp_sock *tp) | |
3074 | { | |
3075 | if (th->doff == sizeof(struct tcphdr)>>2) { | |
3076 | tp->rx_opt.saw_tstamp = 0; | |
3077 | return 0; | |
3078 | } else if (tp->rx_opt.tstamp_ok && | |
3079 | th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) { | |
3080 | __u32 *ptr = (__u32 *)(th + 1); | |
3081 | if (*ptr == ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | |
3082 | | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { | |
3083 | tp->rx_opt.saw_tstamp = 1; | |
3084 | ++ptr; | |
3085 | tp->rx_opt.rcv_tsval = ntohl(*ptr); | |
3086 | ++ptr; | |
3087 | tp->rx_opt.rcv_tsecr = ntohl(*ptr); | |
3088 | return 1; | |
3089 | } | |
3090 | } | |
3091 | tcp_parse_options(skb, &tp->rx_opt, 1); | |
3092 | return 1; | |
3093 | } | |
3094 | ||
3095 | static inline void tcp_store_ts_recent(struct tcp_sock *tp) | |
3096 | { | |
3097 | tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; | |
3098 | tp->rx_opt.ts_recent_stamp = xtime.tv_sec; | |
3099 | } | |
3100 | ||
3101 | static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) | |
3102 | { | |
3103 | if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { | |
3104 | /* PAWS bug workaround wrt. ACK frames, the PAWS discard | |
3105 | * extra check below makes sure this can only happen | |
3106 | * for pure ACK frames. -DaveM | |
3107 | * | |
3108 | * Not only, also it occurs for expired timestamps. | |
3109 | */ | |
3110 | ||
3111 | if((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 || | |
3112 | xtime.tv_sec >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS) | |
3113 | tcp_store_ts_recent(tp); | |
3114 | } | |
3115 | } | |
3116 | ||
3117 | /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM | |
3118 | * | |
3119 | * It is not fatal. If this ACK does _not_ change critical state (seqs, window) | |
3120 | * it can pass through stack. So, the following predicate verifies that | |
3121 | * this segment is not used for anything but congestion avoidance or | |
3122 | * fast retransmit. Moreover, we even are able to eliminate most of such | |
3123 | * second order effects, if we apply some small "replay" window (~RTO) | |
3124 | * to timestamp space. | |
3125 | * | |
3126 | * All these measures still do not guarantee that we reject wrapped ACKs | |
3127 | * on networks with high bandwidth, when sequence space is recycled fastly, | |
3128 | * but it guarantees that such events will be very rare and do not affect | |
3129 | * connection seriously. This doesn't look nice, but alas, PAWS is really | |
3130 | * buggy extension. | |
3131 | * | |
3132 | * [ Later note. Even worse! It is buggy for segments _with_ data. RFC | |
3133 | * states that events when retransmit arrives after original data are rare. | |
3134 | * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is | |
3135 | * the biggest problem on large power networks even with minor reordering. | |
3136 | * OK, let's give it small replay window. If peer clock is even 1hz, it is safe | |
3137 | * up to bandwidth of 18Gigabit/sec. 8) ] | |
3138 | */ | |
3139 | ||
3140 | static int tcp_disordered_ack(struct tcp_sock *tp, struct sk_buff *skb) | |
3141 | { | |
3142 | struct tcphdr *th = skb->h.th; | |
3143 | u32 seq = TCP_SKB_CB(skb)->seq; | |
3144 | u32 ack = TCP_SKB_CB(skb)->ack_seq; | |
3145 | ||
3146 | return (/* 1. Pure ACK with correct sequence number. */ | |
3147 | (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && | |
3148 | ||
3149 | /* 2. ... and duplicate ACK. */ | |
3150 | ack == tp->snd_una && | |
3151 | ||
3152 | /* 3. ... and does not update window. */ | |
3153 | !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && | |
3154 | ||
3155 | /* 4. ... and sits in replay window. */ | |
3156 | (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (tp->rto*1024)/HZ); | |
3157 | } | |
3158 | ||
3159 | static inline int tcp_paws_discard(struct tcp_sock *tp, struct sk_buff *skb) | |
3160 | { | |
3161 | return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW && | |
3162 | xtime.tv_sec < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS && | |
3163 | !tcp_disordered_ack(tp, skb)); | |
3164 | } | |
3165 | ||
3166 | /* Check segment sequence number for validity. | |
3167 | * | |
3168 | * Segment controls are considered valid, if the segment | |
3169 | * fits to the window after truncation to the window. Acceptability | |
3170 | * of data (and SYN, FIN, of course) is checked separately. | |
3171 | * See tcp_data_queue(), for example. | |
3172 | * | |
3173 | * Also, controls (RST is main one) are accepted using RCV.WUP instead | |
3174 | * of RCV.NXT. Peer still did not advance his SND.UNA when we | |
3175 | * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. | |
3176 | * (borrowed from freebsd) | |
3177 | */ | |
3178 | ||
3179 | static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq) | |
3180 | { | |
3181 | return !before(end_seq, tp->rcv_wup) && | |
3182 | !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); | |
3183 | } | |
3184 | ||
3185 | /* When we get a reset we do this. */ | |
3186 | static void tcp_reset(struct sock *sk) | |
3187 | { | |
3188 | /* We want the right error as BSD sees it (and indeed as we do). */ | |
3189 | switch (sk->sk_state) { | |
3190 | case TCP_SYN_SENT: | |
3191 | sk->sk_err = ECONNREFUSED; | |
3192 | break; | |
3193 | case TCP_CLOSE_WAIT: | |
3194 | sk->sk_err = EPIPE; | |
3195 | break; | |
3196 | case TCP_CLOSE: | |
3197 | return; | |
3198 | default: | |
3199 | sk->sk_err = ECONNRESET; | |
3200 | } | |
3201 | ||
3202 | if (!sock_flag(sk, SOCK_DEAD)) | |
3203 | sk->sk_error_report(sk); | |
3204 | ||
3205 | tcp_done(sk); | |
3206 | } | |
3207 | ||
3208 | /* | |
3209 | * Process the FIN bit. This now behaves as it is supposed to work | |
3210 | * and the FIN takes effect when it is validly part of sequence | |
3211 | * space. Not before when we get holes. | |
3212 | * | |
3213 | * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT | |
3214 | * (and thence onto LAST-ACK and finally, CLOSE, we never enter | |
3215 | * TIME-WAIT) | |
3216 | * | |
3217 | * If we are in FINWAIT-1, a received FIN indicates simultaneous | |
3218 | * close and we go into CLOSING (and later onto TIME-WAIT) | |
3219 | * | |
3220 | * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. | |
3221 | */ | |
3222 | static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th) | |
3223 | { | |
3224 | struct tcp_sock *tp = tcp_sk(sk); | |
3225 | ||
3226 | tcp_schedule_ack(tp); | |
3227 | ||
3228 | sk->sk_shutdown |= RCV_SHUTDOWN; | |
3229 | sock_set_flag(sk, SOCK_DONE); | |
3230 | ||
3231 | switch (sk->sk_state) { | |
3232 | case TCP_SYN_RECV: | |
3233 | case TCP_ESTABLISHED: | |
3234 | /* Move to CLOSE_WAIT */ | |
3235 | tcp_set_state(sk, TCP_CLOSE_WAIT); | |
3236 | tp->ack.pingpong = 1; | |
3237 | break; | |
3238 | ||
3239 | case TCP_CLOSE_WAIT: | |
3240 | case TCP_CLOSING: | |
3241 | /* Received a retransmission of the FIN, do | |
3242 | * nothing. | |
3243 | */ | |
3244 | break; | |
3245 | case TCP_LAST_ACK: | |
3246 | /* RFC793: Remain in the LAST-ACK state. */ | |
3247 | break; | |
3248 | ||
3249 | case TCP_FIN_WAIT1: | |
3250 | /* This case occurs when a simultaneous close | |
3251 | * happens, we must ack the received FIN and | |
3252 | * enter the CLOSING state. | |
3253 | */ | |
3254 | tcp_send_ack(sk); | |
3255 | tcp_set_state(sk, TCP_CLOSING); | |
3256 | break; | |
3257 | case TCP_FIN_WAIT2: | |
3258 | /* Received a FIN -- send ACK and enter TIME_WAIT. */ | |
3259 | tcp_send_ack(sk); | |
3260 | tcp_time_wait(sk, TCP_TIME_WAIT, 0); | |
3261 | break; | |
3262 | default: | |
3263 | /* Only TCP_LISTEN and TCP_CLOSE are left, in these | |
3264 | * cases we should never reach this piece of code. | |
3265 | */ | |
3266 | printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n", | |
3267 | __FUNCTION__, sk->sk_state); | |
3268 | break; | |
3269 | }; | |
3270 | ||
3271 | /* It _is_ possible, that we have something out-of-order _after_ FIN. | |
3272 | * Probably, we should reset in this case. For now drop them. | |
3273 | */ | |
3274 | __skb_queue_purge(&tp->out_of_order_queue); | |
3275 | if (tp->rx_opt.sack_ok) | |
3276 | tcp_sack_reset(&tp->rx_opt); | |
3277 | sk_stream_mem_reclaim(sk); | |
3278 | ||
3279 | if (!sock_flag(sk, SOCK_DEAD)) { | |
3280 | sk->sk_state_change(sk); | |
3281 | ||
3282 | /* Do not send POLL_HUP for half duplex close. */ | |
3283 | if (sk->sk_shutdown == SHUTDOWN_MASK || | |
3284 | sk->sk_state == TCP_CLOSE) | |
3285 | sk_wake_async(sk, 1, POLL_HUP); | |
3286 | else | |
3287 | sk_wake_async(sk, 1, POLL_IN); | |
3288 | } | |
3289 | } | |
3290 | ||
3291 | static __inline__ int | |
3292 | tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, u32 end_seq) | |
3293 | { | |
3294 | if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { | |
3295 | if (before(seq, sp->start_seq)) | |
3296 | sp->start_seq = seq; | |
3297 | if (after(end_seq, sp->end_seq)) | |
3298 | sp->end_seq = end_seq; | |
3299 | return 1; | |
3300 | } | |
3301 | return 0; | |
3302 | } | |
3303 | ||
3304 | static inline void tcp_dsack_set(struct tcp_sock *tp, u32 seq, u32 end_seq) | |
3305 | { | |
3306 | if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) { | |
3307 | if (before(seq, tp->rcv_nxt)) | |
3308 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOLDSENT); | |
3309 | else | |
3310 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFOSENT); | |
3311 | ||
3312 | tp->rx_opt.dsack = 1; | |
3313 | tp->duplicate_sack[0].start_seq = seq; | |
3314 | tp->duplicate_sack[0].end_seq = end_seq; | |
3315 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + 1, 4 - tp->rx_opt.tstamp_ok); | |
3316 | } | |
3317 | } | |
3318 | ||
3319 | static inline void tcp_dsack_extend(struct tcp_sock *tp, u32 seq, u32 end_seq) | |
3320 | { | |
3321 | if (!tp->rx_opt.dsack) | |
3322 | tcp_dsack_set(tp, seq, end_seq); | |
3323 | else | |
3324 | tcp_sack_extend(tp->duplicate_sack, seq, end_seq); | |
3325 | } | |
3326 | ||
3327 | static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb) | |
3328 | { | |
3329 | struct tcp_sock *tp = tcp_sk(sk); | |
3330 | ||
3331 | if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && | |
3332 | before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
3333 | NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); | |
3334 | tcp_enter_quickack_mode(tp); | |
3335 | ||
3336 | if (tp->rx_opt.sack_ok && sysctl_tcp_dsack) { | |
3337 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
3338 | ||
3339 | if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) | |
3340 | end_seq = tp->rcv_nxt; | |
3341 | tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, end_seq); | |
3342 | } | |
3343 | } | |
3344 | ||
3345 | tcp_send_ack(sk); | |
3346 | } | |
3347 | ||
3348 | /* These routines update the SACK block as out-of-order packets arrive or | |
3349 | * in-order packets close up the sequence space. | |
3350 | */ | |
3351 | static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) | |
3352 | { | |
3353 | int this_sack; | |
3354 | struct tcp_sack_block *sp = &tp->selective_acks[0]; | |
3355 | struct tcp_sack_block *swalk = sp+1; | |
3356 | ||
3357 | /* See if the recent change to the first SACK eats into | |
3358 | * or hits the sequence space of other SACK blocks, if so coalesce. | |
3359 | */ | |
3360 | for (this_sack = 1; this_sack < tp->rx_opt.num_sacks; ) { | |
3361 | if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { | |
3362 | int i; | |
3363 | ||
3364 | /* Zap SWALK, by moving every further SACK up by one slot. | |
3365 | * Decrease num_sacks. | |
3366 | */ | |
3367 | tp->rx_opt.num_sacks--; | |
3368 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); | |
3369 | for(i=this_sack; i < tp->rx_opt.num_sacks; i++) | |
3370 | sp[i] = sp[i+1]; | |
3371 | continue; | |
3372 | } | |
3373 | this_sack++, swalk++; | |
3374 | } | |
3375 | } | |
3376 | ||
3377 | static __inline__ void tcp_sack_swap(struct tcp_sack_block *sack1, struct tcp_sack_block *sack2) | |
3378 | { | |
3379 | __u32 tmp; | |
3380 | ||
3381 | tmp = sack1->start_seq; | |
3382 | sack1->start_seq = sack2->start_seq; | |
3383 | sack2->start_seq = tmp; | |
3384 | ||
3385 | tmp = sack1->end_seq; | |
3386 | sack1->end_seq = sack2->end_seq; | |
3387 | sack2->end_seq = tmp; | |
3388 | } | |
3389 | ||
3390 | static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) | |
3391 | { | |
3392 | struct tcp_sock *tp = tcp_sk(sk); | |
3393 | struct tcp_sack_block *sp = &tp->selective_acks[0]; | |
3394 | int cur_sacks = tp->rx_opt.num_sacks; | |
3395 | int this_sack; | |
3396 | ||
3397 | if (!cur_sacks) | |
3398 | goto new_sack; | |
3399 | ||
3400 | for (this_sack=0; this_sack<cur_sacks; this_sack++, sp++) { | |
3401 | if (tcp_sack_extend(sp, seq, end_seq)) { | |
3402 | /* Rotate this_sack to the first one. */ | |
3403 | for (; this_sack>0; this_sack--, sp--) | |
3404 | tcp_sack_swap(sp, sp-1); | |
3405 | if (cur_sacks > 1) | |
3406 | tcp_sack_maybe_coalesce(tp); | |
3407 | return; | |
3408 | } | |
3409 | } | |
3410 | ||
3411 | /* Could not find an adjacent existing SACK, build a new one, | |
3412 | * put it at the front, and shift everyone else down. We | |
3413 | * always know there is at least one SACK present already here. | |
3414 | * | |
3415 | * If the sack array is full, forget about the last one. | |
3416 | */ | |
3417 | if (this_sack >= 4) { | |
3418 | this_sack--; | |
3419 | tp->rx_opt.num_sacks--; | |
3420 | sp--; | |
3421 | } | |
3422 | for(; this_sack > 0; this_sack--, sp--) | |
3423 | *sp = *(sp-1); | |
3424 | ||
3425 | new_sack: | |
3426 | /* Build the new head SACK, and we're done. */ | |
3427 | sp->start_seq = seq; | |
3428 | sp->end_seq = end_seq; | |
3429 | tp->rx_opt.num_sacks++; | |
3430 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); | |
3431 | } | |
3432 | ||
3433 | /* RCV.NXT advances, some SACKs should be eaten. */ | |
3434 | ||
3435 | static void tcp_sack_remove(struct tcp_sock *tp) | |
3436 | { | |
3437 | struct tcp_sack_block *sp = &tp->selective_acks[0]; | |
3438 | int num_sacks = tp->rx_opt.num_sacks; | |
3439 | int this_sack; | |
3440 | ||
3441 | /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ | |
3442 | if (skb_queue_len(&tp->out_of_order_queue) == 0) { | |
3443 | tp->rx_opt.num_sacks = 0; | |
3444 | tp->rx_opt.eff_sacks = tp->rx_opt.dsack; | |
3445 | return; | |
3446 | } | |
3447 | ||
3448 | for(this_sack = 0; this_sack < num_sacks; ) { | |
3449 | /* Check if the start of the sack is covered by RCV.NXT. */ | |
3450 | if (!before(tp->rcv_nxt, sp->start_seq)) { | |
3451 | int i; | |
3452 | ||
3453 | /* RCV.NXT must cover all the block! */ | |
3454 | BUG_TRAP(!before(tp->rcv_nxt, sp->end_seq)); | |
3455 | ||
3456 | /* Zap this SACK, by moving forward any other SACKS. */ | |
3457 | for (i=this_sack+1; i < num_sacks; i++) | |
3458 | tp->selective_acks[i-1] = tp->selective_acks[i]; | |
3459 | num_sacks--; | |
3460 | continue; | |
3461 | } | |
3462 | this_sack++; | |
3463 | sp++; | |
3464 | } | |
3465 | if (num_sacks != tp->rx_opt.num_sacks) { | |
3466 | tp->rx_opt.num_sacks = num_sacks; | |
3467 | tp->rx_opt.eff_sacks = min(tp->rx_opt.num_sacks + tp->rx_opt.dsack, 4 - tp->rx_opt.tstamp_ok); | |
3468 | } | |
3469 | } | |
3470 | ||
3471 | /* This one checks to see if we can put data from the | |
3472 | * out_of_order queue into the receive_queue. | |
3473 | */ | |
3474 | static void tcp_ofo_queue(struct sock *sk) | |
3475 | { | |
3476 | struct tcp_sock *tp = tcp_sk(sk); | |
3477 | __u32 dsack_high = tp->rcv_nxt; | |
3478 | struct sk_buff *skb; | |
3479 | ||
3480 | while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { | |
3481 | if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) | |
3482 | break; | |
3483 | ||
3484 | if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { | |
3485 | __u32 dsack = dsack_high; | |
3486 | if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) | |
3487 | dsack_high = TCP_SKB_CB(skb)->end_seq; | |
3488 | tcp_dsack_extend(tp, TCP_SKB_CB(skb)->seq, dsack); | |
3489 | } | |
3490 | ||
3491 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { | |
3492 | SOCK_DEBUG(sk, "ofo packet was already received \n"); | |
3493 | __skb_unlink(skb, skb->list); | |
3494 | __kfree_skb(skb); | |
3495 | continue; | |
3496 | } | |
3497 | SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", | |
3498 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, | |
3499 | TCP_SKB_CB(skb)->end_seq); | |
3500 | ||
3501 | __skb_unlink(skb, skb->list); | |
3502 | __skb_queue_tail(&sk->sk_receive_queue, skb); | |
3503 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
3504 | if(skb->h.th->fin) | |
3505 | tcp_fin(skb, sk, skb->h.th); | |
3506 | } | |
3507 | } | |
3508 | ||
3509 | static int tcp_prune_queue(struct sock *sk); | |
3510 | ||
3511 | static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) | |
3512 | { | |
3513 | struct tcphdr *th = skb->h.th; | |
3514 | struct tcp_sock *tp = tcp_sk(sk); | |
3515 | int eaten = -1; | |
3516 | ||
3517 | if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) | |
3518 | goto drop; | |
3519 | ||
1da177e4 LT |
3520 | __skb_pull(skb, th->doff*4); |
3521 | ||
3522 | TCP_ECN_accept_cwr(tp, skb); | |
3523 | ||
3524 | if (tp->rx_opt.dsack) { | |
3525 | tp->rx_opt.dsack = 0; | |
3526 | tp->rx_opt.eff_sacks = min_t(unsigned int, tp->rx_opt.num_sacks, | |
3527 | 4 - tp->rx_opt.tstamp_ok); | |
3528 | } | |
3529 | ||
3530 | /* Queue data for delivery to the user. | |
3531 | * Packets in sequence go to the receive queue. | |
3532 | * Out of sequence packets to the out_of_order_queue. | |
3533 | */ | |
3534 | if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { | |
3535 | if (tcp_receive_window(tp) == 0) | |
3536 | goto out_of_window; | |
3537 | ||
3538 | /* Ok. In sequence. In window. */ | |
3539 | if (tp->ucopy.task == current && | |
3540 | tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && | |
3541 | sock_owned_by_user(sk) && !tp->urg_data) { | |
3542 | int chunk = min_t(unsigned int, skb->len, | |
3543 | tp->ucopy.len); | |
3544 | ||
3545 | __set_current_state(TASK_RUNNING); | |
3546 | ||
3547 | local_bh_enable(); | |
3548 | if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { | |
3549 | tp->ucopy.len -= chunk; | |
3550 | tp->copied_seq += chunk; | |
3551 | eaten = (chunk == skb->len && !th->fin); | |
3552 | tcp_rcv_space_adjust(sk); | |
3553 | } | |
3554 | local_bh_disable(); | |
3555 | } | |
3556 | ||
3557 | if (eaten <= 0) { | |
3558 | queue_and_out: | |
3559 | if (eaten < 0 && | |
3560 | (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || | |
3561 | !sk_stream_rmem_schedule(sk, skb))) { | |
3562 | if (tcp_prune_queue(sk) < 0 || | |
3563 | !sk_stream_rmem_schedule(sk, skb)) | |
3564 | goto drop; | |
3565 | } | |
3566 | sk_stream_set_owner_r(skb, sk); | |
3567 | __skb_queue_tail(&sk->sk_receive_queue, skb); | |
3568 | } | |
3569 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
3570 | if(skb->len) | |
3571 | tcp_event_data_recv(sk, tp, skb); | |
3572 | if(th->fin) | |
3573 | tcp_fin(skb, sk, th); | |
3574 | ||
3575 | if (skb_queue_len(&tp->out_of_order_queue)) { | |
3576 | tcp_ofo_queue(sk); | |
3577 | ||
3578 | /* RFC2581. 4.2. SHOULD send immediate ACK, when | |
3579 | * gap in queue is filled. | |
3580 | */ | |
3581 | if (!skb_queue_len(&tp->out_of_order_queue)) | |
3582 | tp->ack.pingpong = 0; | |
3583 | } | |
3584 | ||
3585 | if (tp->rx_opt.num_sacks) | |
3586 | tcp_sack_remove(tp); | |
3587 | ||
3588 | tcp_fast_path_check(sk, tp); | |
3589 | ||
3590 | if (eaten > 0) | |
3591 | __kfree_skb(skb); | |
3592 | else if (!sock_flag(sk, SOCK_DEAD)) | |
3593 | sk->sk_data_ready(sk, 0); | |
3594 | return; | |
3595 | } | |
3596 | ||
3597 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { | |
3598 | /* A retransmit, 2nd most common case. Force an immediate ack. */ | |
3599 | NET_INC_STATS_BH(LINUX_MIB_DELAYEDACKLOST); | |
3600 | tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); | |
3601 | ||
3602 | out_of_window: | |
3603 | tcp_enter_quickack_mode(tp); | |
3604 | tcp_schedule_ack(tp); | |
3605 | drop: | |
3606 | __kfree_skb(skb); | |
3607 | return; | |
3608 | } | |
3609 | ||
3610 | /* Out of window. F.e. zero window probe. */ | |
3611 | if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) | |
3612 | goto out_of_window; | |
3613 | ||
3614 | tcp_enter_quickack_mode(tp); | |
3615 | ||
3616 | if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
3617 | /* Partial packet, seq < rcv_next < end_seq */ | |
3618 | SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", | |
3619 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, | |
3620 | TCP_SKB_CB(skb)->end_seq); | |
3621 | ||
3622 | tcp_dsack_set(tp, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); | |
3623 | ||
3624 | /* If window is closed, drop tail of packet. But after | |
3625 | * remembering D-SACK for its head made in previous line. | |
3626 | */ | |
3627 | if (!tcp_receive_window(tp)) | |
3628 | goto out_of_window; | |
3629 | goto queue_and_out; | |
3630 | } | |
3631 | ||
3632 | TCP_ECN_check_ce(tp, skb); | |
3633 | ||
3634 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || | |
3635 | !sk_stream_rmem_schedule(sk, skb)) { | |
3636 | if (tcp_prune_queue(sk) < 0 || | |
3637 | !sk_stream_rmem_schedule(sk, skb)) | |
3638 | goto drop; | |
3639 | } | |
3640 | ||
3641 | /* Disable header prediction. */ | |
3642 | tp->pred_flags = 0; | |
3643 | tcp_schedule_ack(tp); | |
3644 | ||
3645 | SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", | |
3646 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); | |
3647 | ||
3648 | sk_stream_set_owner_r(skb, sk); | |
3649 | ||
3650 | if (!skb_peek(&tp->out_of_order_queue)) { | |
3651 | /* Initial out of order segment, build 1 SACK. */ | |
3652 | if (tp->rx_opt.sack_ok) { | |
3653 | tp->rx_opt.num_sacks = 1; | |
3654 | tp->rx_opt.dsack = 0; | |
3655 | tp->rx_opt.eff_sacks = 1; | |
3656 | tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; | |
3657 | tp->selective_acks[0].end_seq = | |
3658 | TCP_SKB_CB(skb)->end_seq; | |
3659 | } | |
3660 | __skb_queue_head(&tp->out_of_order_queue,skb); | |
3661 | } else { | |
3662 | struct sk_buff *skb1 = tp->out_of_order_queue.prev; | |
3663 | u32 seq = TCP_SKB_CB(skb)->seq; | |
3664 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
3665 | ||
3666 | if (seq == TCP_SKB_CB(skb1)->end_seq) { | |
3667 | __skb_append(skb1, skb); | |
3668 | ||
3669 | if (!tp->rx_opt.num_sacks || | |
3670 | tp->selective_acks[0].end_seq != seq) | |
3671 | goto add_sack; | |
3672 | ||
3673 | /* Common case: data arrive in order after hole. */ | |
3674 | tp->selective_acks[0].end_seq = end_seq; | |
3675 | return; | |
3676 | } | |
3677 | ||
3678 | /* Find place to insert this segment. */ | |
3679 | do { | |
3680 | if (!after(TCP_SKB_CB(skb1)->seq, seq)) | |
3681 | break; | |
3682 | } while ((skb1 = skb1->prev) != | |
3683 | (struct sk_buff*)&tp->out_of_order_queue); | |
3684 | ||
3685 | /* Do skb overlap to previous one? */ | |
3686 | if (skb1 != (struct sk_buff*)&tp->out_of_order_queue && | |
3687 | before(seq, TCP_SKB_CB(skb1)->end_seq)) { | |
3688 | if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { | |
3689 | /* All the bits are present. Drop. */ | |
3690 | __kfree_skb(skb); | |
3691 | tcp_dsack_set(tp, seq, end_seq); | |
3692 | goto add_sack; | |
3693 | } | |
3694 | if (after(seq, TCP_SKB_CB(skb1)->seq)) { | |
3695 | /* Partial overlap. */ | |
3696 | tcp_dsack_set(tp, seq, TCP_SKB_CB(skb1)->end_seq); | |
3697 | } else { | |
3698 | skb1 = skb1->prev; | |
3699 | } | |
3700 | } | |
3701 | __skb_insert(skb, skb1, skb1->next, &tp->out_of_order_queue); | |
3702 | ||
3703 | /* And clean segments covered by new one as whole. */ | |
3704 | while ((skb1 = skb->next) != | |
3705 | (struct sk_buff*)&tp->out_of_order_queue && | |
3706 | after(end_seq, TCP_SKB_CB(skb1)->seq)) { | |
3707 | if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { | |
3708 | tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, end_seq); | |
3709 | break; | |
3710 | } | |
3711 | __skb_unlink(skb1, skb1->list); | |
3712 | tcp_dsack_extend(tp, TCP_SKB_CB(skb1)->seq, TCP_SKB_CB(skb1)->end_seq); | |
3713 | __kfree_skb(skb1); | |
3714 | } | |
3715 | ||
3716 | add_sack: | |
3717 | if (tp->rx_opt.sack_ok) | |
3718 | tcp_sack_new_ofo_skb(sk, seq, end_seq); | |
3719 | } | |
3720 | } | |
3721 | ||
3722 | /* Collapse contiguous sequence of skbs head..tail with | |
3723 | * sequence numbers start..end. | |
3724 | * Segments with FIN/SYN are not collapsed (only because this | |
3725 | * simplifies code) | |
3726 | */ | |
3727 | static void | |
3728 | tcp_collapse(struct sock *sk, struct sk_buff *head, | |
3729 | struct sk_buff *tail, u32 start, u32 end) | |
3730 | { | |
3731 | struct sk_buff *skb; | |
3732 | ||
3733 | /* First, check that queue is collapsable and find | |
3734 | * the point where collapsing can be useful. */ | |
3735 | for (skb = head; skb != tail; ) { | |
3736 | /* No new bits? It is possible on ofo queue. */ | |
3737 | if (!before(start, TCP_SKB_CB(skb)->end_seq)) { | |
3738 | struct sk_buff *next = skb->next; | |
3739 | __skb_unlink(skb, skb->list); | |
3740 | __kfree_skb(skb); | |
3741 | NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); | |
3742 | skb = next; | |
3743 | continue; | |
3744 | } | |
3745 | ||
3746 | /* The first skb to collapse is: | |
3747 | * - not SYN/FIN and | |
3748 | * - bloated or contains data before "start" or | |
3749 | * overlaps to the next one. | |
3750 | */ | |
3751 | if (!skb->h.th->syn && !skb->h.th->fin && | |
3752 | (tcp_win_from_space(skb->truesize) > skb->len || | |
3753 | before(TCP_SKB_CB(skb)->seq, start) || | |
3754 | (skb->next != tail && | |
3755 | TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq))) | |
3756 | break; | |
3757 | ||
3758 | /* Decided to skip this, advance start seq. */ | |
3759 | start = TCP_SKB_CB(skb)->end_seq; | |
3760 | skb = skb->next; | |
3761 | } | |
3762 | if (skb == tail || skb->h.th->syn || skb->h.th->fin) | |
3763 | return; | |
3764 | ||
3765 | while (before(start, end)) { | |
3766 | struct sk_buff *nskb; | |
3767 | int header = skb_headroom(skb); | |
3768 | int copy = SKB_MAX_ORDER(header, 0); | |
3769 | ||
3770 | /* Too big header? This can happen with IPv6. */ | |
3771 | if (copy < 0) | |
3772 | return; | |
3773 | if (end-start < copy) | |
3774 | copy = end-start; | |
3775 | nskb = alloc_skb(copy+header, GFP_ATOMIC); | |
3776 | if (!nskb) | |
3777 | return; | |
3778 | skb_reserve(nskb, header); | |
3779 | memcpy(nskb->head, skb->head, header); | |
3780 | nskb->nh.raw = nskb->head + (skb->nh.raw-skb->head); | |
3781 | nskb->h.raw = nskb->head + (skb->h.raw-skb->head); | |
3782 | nskb->mac.raw = nskb->head + (skb->mac.raw-skb->head); | |
3783 | memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); | |
3784 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; | |
3785 | __skb_insert(nskb, skb->prev, skb, skb->list); | |
3786 | sk_stream_set_owner_r(nskb, sk); | |
3787 | ||
3788 | /* Copy data, releasing collapsed skbs. */ | |
3789 | while (copy > 0) { | |
3790 | int offset = start - TCP_SKB_CB(skb)->seq; | |
3791 | int size = TCP_SKB_CB(skb)->end_seq - start; | |
3792 | ||
3793 | if (offset < 0) BUG(); | |
3794 | if (size > 0) { | |
3795 | size = min(copy, size); | |
3796 | if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) | |
3797 | BUG(); | |
3798 | TCP_SKB_CB(nskb)->end_seq += size; | |
3799 | copy -= size; | |
3800 | start += size; | |
3801 | } | |
3802 | if (!before(start, TCP_SKB_CB(skb)->end_seq)) { | |
3803 | struct sk_buff *next = skb->next; | |
3804 | __skb_unlink(skb, skb->list); | |
3805 | __kfree_skb(skb); | |
3806 | NET_INC_STATS_BH(LINUX_MIB_TCPRCVCOLLAPSED); | |
3807 | skb = next; | |
3808 | if (skb == tail || skb->h.th->syn || skb->h.th->fin) | |
3809 | return; | |
3810 | } | |
3811 | } | |
3812 | } | |
3813 | } | |
3814 | ||
3815 | /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs | |
3816 | * and tcp_collapse() them until all the queue is collapsed. | |
3817 | */ | |
3818 | static void tcp_collapse_ofo_queue(struct sock *sk) | |
3819 | { | |
3820 | struct tcp_sock *tp = tcp_sk(sk); | |
3821 | struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); | |
3822 | struct sk_buff *head; | |
3823 | u32 start, end; | |
3824 | ||
3825 | if (skb == NULL) | |
3826 | return; | |
3827 | ||
3828 | start = TCP_SKB_CB(skb)->seq; | |
3829 | end = TCP_SKB_CB(skb)->end_seq; | |
3830 | head = skb; | |
3831 | ||
3832 | for (;;) { | |
3833 | skb = skb->next; | |
3834 | ||
3835 | /* Segment is terminated when we see gap or when | |
3836 | * we are at the end of all the queue. */ | |
3837 | if (skb == (struct sk_buff *)&tp->out_of_order_queue || | |
3838 | after(TCP_SKB_CB(skb)->seq, end) || | |
3839 | before(TCP_SKB_CB(skb)->end_seq, start)) { | |
3840 | tcp_collapse(sk, head, skb, start, end); | |
3841 | head = skb; | |
3842 | if (skb == (struct sk_buff *)&tp->out_of_order_queue) | |
3843 | break; | |
3844 | /* Start new segment */ | |
3845 | start = TCP_SKB_CB(skb)->seq; | |
3846 | end = TCP_SKB_CB(skb)->end_seq; | |
3847 | } else { | |
3848 | if (before(TCP_SKB_CB(skb)->seq, start)) | |
3849 | start = TCP_SKB_CB(skb)->seq; | |
3850 | if (after(TCP_SKB_CB(skb)->end_seq, end)) | |
3851 | end = TCP_SKB_CB(skb)->end_seq; | |
3852 | } | |
3853 | } | |
3854 | } | |
3855 | ||
3856 | /* Reduce allocated memory if we can, trying to get | |
3857 | * the socket within its memory limits again. | |
3858 | * | |
3859 | * Return less than zero if we should start dropping frames | |
3860 | * until the socket owning process reads some of the data | |
3861 | * to stabilize the situation. | |
3862 | */ | |
3863 | static int tcp_prune_queue(struct sock *sk) | |
3864 | { | |
3865 | struct tcp_sock *tp = tcp_sk(sk); | |
3866 | ||
3867 | SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); | |
3868 | ||
3869 | NET_INC_STATS_BH(LINUX_MIB_PRUNECALLED); | |
3870 | ||
3871 | if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) | |
3872 | tcp_clamp_window(sk, tp); | |
3873 | else if (tcp_memory_pressure) | |
3874 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); | |
3875 | ||
3876 | tcp_collapse_ofo_queue(sk); | |
3877 | tcp_collapse(sk, sk->sk_receive_queue.next, | |
3878 | (struct sk_buff*)&sk->sk_receive_queue, | |
3879 | tp->copied_seq, tp->rcv_nxt); | |
3880 | sk_stream_mem_reclaim(sk); | |
3881 | ||
3882 | if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) | |
3883 | return 0; | |
3884 | ||
3885 | /* Collapsing did not help, destructive actions follow. | |
3886 | * This must not ever occur. */ | |
3887 | ||
3888 | /* First, purge the out_of_order queue. */ | |
3889 | if (skb_queue_len(&tp->out_of_order_queue)) { | |
3890 | NET_ADD_STATS_BH(LINUX_MIB_OFOPRUNED, | |
3891 | skb_queue_len(&tp->out_of_order_queue)); | |
3892 | __skb_queue_purge(&tp->out_of_order_queue); | |
3893 | ||
3894 | /* Reset SACK state. A conforming SACK implementation will | |
3895 | * do the same at a timeout based retransmit. When a connection | |
3896 | * is in a sad state like this, we care only about integrity | |
3897 | * of the connection not performance. | |
3898 | */ | |
3899 | if (tp->rx_opt.sack_ok) | |
3900 | tcp_sack_reset(&tp->rx_opt); | |
3901 | sk_stream_mem_reclaim(sk); | |
3902 | } | |
3903 | ||
3904 | if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) | |
3905 | return 0; | |
3906 | ||
3907 | /* If we are really being abused, tell the caller to silently | |
3908 | * drop receive data on the floor. It will get retransmitted | |
3909 | * and hopefully then we'll have sufficient space. | |
3910 | */ | |
3911 | NET_INC_STATS_BH(LINUX_MIB_RCVPRUNED); | |
3912 | ||
3913 | /* Massive buffer overcommit. */ | |
3914 | tp->pred_flags = 0; | |
3915 | return -1; | |
3916 | } | |
3917 | ||
3918 | ||
3919 | /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. | |
3920 | * As additional protections, we do not touch cwnd in retransmission phases, | |
3921 | * and if application hit its sndbuf limit recently. | |
3922 | */ | |
3923 | void tcp_cwnd_application_limited(struct sock *sk) | |
3924 | { | |
3925 | struct tcp_sock *tp = tcp_sk(sk); | |
3926 | ||
3927 | if (tp->ca_state == TCP_CA_Open && | |
3928 | sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { | |
3929 | /* Limited by application or receiver window. */ | |
3930 | u32 win_used = max(tp->snd_cwnd_used, 2U); | |
3931 | if (win_used < tp->snd_cwnd) { | |
3932 | tp->snd_ssthresh = tcp_current_ssthresh(tp); | |
3933 | tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; | |
3934 | } | |
3935 | tp->snd_cwnd_used = 0; | |
3936 | } | |
3937 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
3938 | } | |
3939 | ||
3940 | ||
3941 | /* When incoming ACK allowed to free some skb from write_queue, | |
3942 | * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket | |
3943 | * on the exit from tcp input handler. | |
3944 | * | |
3945 | * PROBLEM: sndbuf expansion does not work well with largesend. | |
3946 | */ | |
3947 | static void tcp_new_space(struct sock *sk) | |
3948 | { | |
3949 | struct tcp_sock *tp = tcp_sk(sk); | |
3950 | ||
3951 | if (tp->packets_out < tp->snd_cwnd && | |
3952 | !(sk->sk_userlocks & SOCK_SNDBUF_LOCK) && | |
3953 | !tcp_memory_pressure && | |
3954 | atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) { | |
3955 | int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache_std) + | |
3956 | MAX_TCP_HEADER + 16 + sizeof(struct sk_buff), | |
3957 | demanded = max_t(unsigned int, tp->snd_cwnd, | |
3958 | tp->reordering + 1); | |
3959 | sndmem *= 2*demanded; | |
3960 | if (sndmem > sk->sk_sndbuf) | |
3961 | sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); | |
3962 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
3963 | } | |
3964 | ||
3965 | sk->sk_write_space(sk); | |
3966 | } | |
3967 | ||
3968 | static inline void tcp_check_space(struct sock *sk) | |
3969 | { | |
3970 | if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { | |
3971 | sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); | |
3972 | if (sk->sk_socket && | |
3973 | test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) | |
3974 | tcp_new_space(sk); | |
3975 | } | |
3976 | } | |
3977 | ||
3978 | static void __tcp_data_snd_check(struct sock *sk, struct sk_buff *skb) | |
3979 | { | |
3980 | struct tcp_sock *tp = tcp_sk(sk); | |
3981 | ||
3982 | if (after(TCP_SKB_CB(skb)->end_seq, tp->snd_una + tp->snd_wnd) || | |
3983 | tcp_packets_in_flight(tp) >= tp->snd_cwnd || | |
3984 | tcp_write_xmit(sk, tp->nonagle)) | |
3985 | tcp_check_probe_timer(sk, tp); | |
3986 | } | |
3987 | ||
3988 | static __inline__ void tcp_data_snd_check(struct sock *sk) | |
3989 | { | |
3990 | struct sk_buff *skb = sk->sk_send_head; | |
3991 | ||
3992 | if (skb != NULL) | |
3993 | __tcp_data_snd_check(sk, skb); | |
3994 | tcp_check_space(sk); | |
3995 | } | |
3996 | ||
3997 | /* | |
3998 | * Check if sending an ack is needed. | |
3999 | */ | |
4000 | static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) | |
4001 | { | |
4002 | struct tcp_sock *tp = tcp_sk(sk); | |
4003 | ||
4004 | /* More than one full frame received... */ | |
4005 | if (((tp->rcv_nxt - tp->rcv_wup) > tp->ack.rcv_mss | |
4006 | /* ... and right edge of window advances far enough. | |
4007 | * (tcp_recvmsg() will send ACK otherwise). Or... | |
4008 | */ | |
4009 | && __tcp_select_window(sk) >= tp->rcv_wnd) || | |
4010 | /* We ACK each frame or... */ | |
4011 | tcp_in_quickack_mode(tp) || | |
4012 | /* We have out of order data. */ | |
4013 | (ofo_possible && | |
4014 | skb_peek(&tp->out_of_order_queue))) { | |
4015 | /* Then ack it now */ | |
4016 | tcp_send_ack(sk); | |
4017 | } else { | |
4018 | /* Else, send delayed ack. */ | |
4019 | tcp_send_delayed_ack(sk); | |
4020 | } | |
4021 | } | |
4022 | ||
4023 | static __inline__ void tcp_ack_snd_check(struct sock *sk) | |
4024 | { | |
4025 | struct tcp_sock *tp = tcp_sk(sk); | |
4026 | if (!tcp_ack_scheduled(tp)) { | |
4027 | /* We sent a data segment already. */ | |
4028 | return; | |
4029 | } | |
4030 | __tcp_ack_snd_check(sk, 1); | |
4031 | } | |
4032 | ||
4033 | /* | |
4034 | * This routine is only called when we have urgent data | |
4035 | * signalled. Its the 'slow' part of tcp_urg. It could be | |
4036 | * moved inline now as tcp_urg is only called from one | |
4037 | * place. We handle URGent data wrong. We have to - as | |
4038 | * BSD still doesn't use the correction from RFC961. | |
4039 | * For 1003.1g we should support a new option TCP_STDURG to permit | |
4040 | * either form (or just set the sysctl tcp_stdurg). | |
4041 | */ | |
4042 | ||
4043 | static void tcp_check_urg(struct sock * sk, struct tcphdr * th) | |
4044 | { | |
4045 | struct tcp_sock *tp = tcp_sk(sk); | |
4046 | u32 ptr = ntohs(th->urg_ptr); | |
4047 | ||
4048 | if (ptr && !sysctl_tcp_stdurg) | |
4049 | ptr--; | |
4050 | ptr += ntohl(th->seq); | |
4051 | ||
4052 | /* Ignore urgent data that we've already seen and read. */ | |
4053 | if (after(tp->copied_seq, ptr)) | |
4054 | return; | |
4055 | ||
4056 | /* Do not replay urg ptr. | |
4057 | * | |
4058 | * NOTE: interesting situation not covered by specs. | |
4059 | * Misbehaving sender may send urg ptr, pointing to segment, | |
4060 | * which we already have in ofo queue. We are not able to fetch | |
4061 | * such data and will stay in TCP_URG_NOTYET until will be eaten | |
4062 | * by recvmsg(). Seems, we are not obliged to handle such wicked | |
4063 | * situations. But it is worth to think about possibility of some | |
4064 | * DoSes using some hypothetical application level deadlock. | |
4065 | */ | |
4066 | if (before(ptr, tp->rcv_nxt)) | |
4067 | return; | |
4068 | ||
4069 | /* Do we already have a newer (or duplicate) urgent pointer? */ | |
4070 | if (tp->urg_data && !after(ptr, tp->urg_seq)) | |
4071 | return; | |
4072 | ||
4073 | /* Tell the world about our new urgent pointer. */ | |
4074 | sk_send_sigurg(sk); | |
4075 | ||
4076 | /* We may be adding urgent data when the last byte read was | |
4077 | * urgent. To do this requires some care. We cannot just ignore | |
4078 | * tp->copied_seq since we would read the last urgent byte again | |
4079 | * as data, nor can we alter copied_seq until this data arrives | |
4080 | * or we break the sematics of SIOCATMARK (and thus sockatmark()) | |
4081 | * | |
4082 | * NOTE. Double Dutch. Rendering to plain English: author of comment | |
4083 | * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); | |
4084 | * and expect that both A and B disappear from stream. This is _wrong_. | |
4085 | * Though this happens in BSD with high probability, this is occasional. | |
4086 | * Any application relying on this is buggy. Note also, that fix "works" | |
4087 | * only in this artificial test. Insert some normal data between A and B and we will | |
4088 | * decline of BSD again. Verdict: it is better to remove to trap | |
4089 | * buggy users. | |
4090 | */ | |
4091 | if (tp->urg_seq == tp->copied_seq && tp->urg_data && | |
4092 | !sock_flag(sk, SOCK_URGINLINE) && | |
4093 | tp->copied_seq != tp->rcv_nxt) { | |
4094 | struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); | |
4095 | tp->copied_seq++; | |
4096 | if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { | |
4097 | __skb_unlink(skb, skb->list); | |
4098 | __kfree_skb(skb); | |
4099 | } | |
4100 | } | |
4101 | ||
4102 | tp->urg_data = TCP_URG_NOTYET; | |
4103 | tp->urg_seq = ptr; | |
4104 | ||
4105 | /* Disable header prediction. */ | |
4106 | tp->pred_flags = 0; | |
4107 | } | |
4108 | ||
4109 | /* This is the 'fast' part of urgent handling. */ | |
4110 | static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th) | |
4111 | { | |
4112 | struct tcp_sock *tp = tcp_sk(sk); | |
4113 | ||
4114 | /* Check if we get a new urgent pointer - normally not. */ | |
4115 | if (th->urg) | |
4116 | tcp_check_urg(sk,th); | |
4117 | ||
4118 | /* Do we wait for any urgent data? - normally not... */ | |
4119 | if (tp->urg_data == TCP_URG_NOTYET) { | |
4120 | u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - | |
4121 | th->syn; | |
4122 | ||
4123 | /* Is the urgent pointer pointing into this packet? */ | |
4124 | if (ptr < skb->len) { | |
4125 | u8 tmp; | |
4126 | if (skb_copy_bits(skb, ptr, &tmp, 1)) | |
4127 | BUG(); | |
4128 | tp->urg_data = TCP_URG_VALID | tmp; | |
4129 | if (!sock_flag(sk, SOCK_DEAD)) | |
4130 | sk->sk_data_ready(sk, 0); | |
4131 | } | |
4132 | } | |
4133 | } | |
4134 | ||
4135 | static int tcp_copy_to_iovec(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 err; | |
4140 | ||
4141 | local_bh_enable(); | |
4142 | if (skb->ip_summed==CHECKSUM_UNNECESSARY) | |
4143 | err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); | |
4144 | else | |
4145 | err = skb_copy_and_csum_datagram_iovec(skb, hlen, | |
4146 | tp->ucopy.iov); | |
4147 | ||
4148 | if (!err) { | |
4149 | tp->ucopy.len -= chunk; | |
4150 | tp->copied_seq += chunk; | |
4151 | tcp_rcv_space_adjust(sk); | |
4152 | } | |
4153 | ||
4154 | local_bh_disable(); | |
4155 | return err; | |
4156 | } | |
4157 | ||
4158 | static int __tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) | |
4159 | { | |
4160 | int result; | |
4161 | ||
4162 | if (sock_owned_by_user(sk)) { | |
4163 | local_bh_enable(); | |
4164 | result = __tcp_checksum_complete(skb); | |
4165 | local_bh_disable(); | |
4166 | } else { | |
4167 | result = __tcp_checksum_complete(skb); | |
4168 | } | |
4169 | return result; | |
4170 | } | |
4171 | ||
4172 | static __inline__ int | |
4173 | tcp_checksum_complete_user(struct sock *sk, struct sk_buff *skb) | |
4174 | { | |
4175 | return skb->ip_summed != CHECKSUM_UNNECESSARY && | |
4176 | __tcp_checksum_complete_user(sk, skb); | |
4177 | } | |
4178 | ||
4179 | /* | |
4180 | * TCP receive function for the ESTABLISHED state. | |
4181 | * | |
4182 | * It is split into a fast path and a slow path. The fast path is | |
4183 | * disabled when: | |
4184 | * - A zero window was announced from us - zero window probing | |
4185 | * is only handled properly in the slow path. | |
4186 | * - Out of order segments arrived. | |
4187 | * - Urgent data is expected. | |
4188 | * - There is no buffer space left | |
4189 | * - Unexpected TCP flags/window values/header lengths are received | |
4190 | * (detected by checking the TCP header against pred_flags) | |
4191 | * - Data is sent in both directions. Fast path only supports pure senders | |
4192 | * or pure receivers (this means either the sequence number or the ack | |
4193 | * value must stay constant) | |
4194 | * - Unexpected TCP option. | |
4195 | * | |
4196 | * When these conditions are not satisfied it drops into a standard | |
4197 | * receive procedure patterned after RFC793 to handle all cases. | |
4198 | * The first three cases are guaranteed by proper pred_flags setting, | |
4199 | * the rest is checked inline. Fast processing is turned on in | |
4200 | * tcp_data_queue when everything is OK. | |
4201 | */ | |
4202 | int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, | |
4203 | struct tcphdr *th, unsigned len) | |
4204 | { | |
4205 | struct tcp_sock *tp = tcp_sk(sk); | |
4206 | ||
4207 | /* | |
4208 | * Header prediction. | |
4209 | * The code loosely follows the one in the famous | |
4210 | * "30 instruction TCP receive" Van Jacobson mail. | |
4211 | * | |
4212 | * Van's trick is to deposit buffers into socket queue | |
4213 | * on a device interrupt, to call tcp_recv function | |
4214 | * on the receive process context and checksum and copy | |
4215 | * the buffer to user space. smart... | |
4216 | * | |
4217 | * Our current scheme is not silly either but we take the | |
4218 | * extra cost of the net_bh soft interrupt processing... | |
4219 | * We do checksum and copy also but from device to kernel. | |
4220 | */ | |
4221 | ||
4222 | tp->rx_opt.saw_tstamp = 0; | |
4223 | ||
4224 | /* pred_flags is 0xS?10 << 16 + snd_wnd | |
4225 | * if header_predition is to be made | |
4226 | * 'S' will always be tp->tcp_header_len >> 2 | |
4227 | * '?' will be 0 for the fast path, otherwise pred_flags is 0 to | |
4228 | * turn it off (when there are holes in the receive | |
4229 | * space for instance) | |
4230 | * PSH flag is ignored. | |
4231 | */ | |
4232 | ||
4233 | if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && | |
4234 | TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { | |
4235 | int tcp_header_len = tp->tcp_header_len; | |
4236 | ||
4237 | /* Timestamp header prediction: tcp_header_len | |
4238 | * is automatically equal to th->doff*4 due to pred_flags | |
4239 | * match. | |
4240 | */ | |
4241 | ||
4242 | /* Check timestamp */ | |
4243 | if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { | |
4244 | __u32 *ptr = (__u32 *)(th + 1); | |
4245 | ||
4246 | /* No? Slow path! */ | |
4247 | if (*ptr != ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | |
4248 | | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) | |
4249 | goto slow_path; | |
4250 | ||
4251 | tp->rx_opt.saw_tstamp = 1; | |
4252 | ++ptr; | |
4253 | tp->rx_opt.rcv_tsval = ntohl(*ptr); | |
4254 | ++ptr; | |
4255 | tp->rx_opt.rcv_tsecr = ntohl(*ptr); | |
4256 | ||
4257 | /* If PAWS failed, check it more carefully in slow path */ | |
4258 | if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) | |
4259 | goto slow_path; | |
4260 | ||
4261 | /* DO NOT update ts_recent here, if checksum fails | |
4262 | * and timestamp was corrupted part, it will result | |
4263 | * in a hung connection since we will drop all | |
4264 | * future packets due to the PAWS test. | |
4265 | */ | |
4266 | } | |
4267 | ||
4268 | if (len <= tcp_header_len) { | |
4269 | /* Bulk data transfer: sender */ | |
4270 | if (len == tcp_header_len) { | |
4271 | /* Predicted packet is in window by definition. | |
4272 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. | |
4273 | * Hence, check seq<=rcv_wup reduces to: | |
4274 | */ | |
4275 | if (tcp_header_len == | |
4276 | (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && | |
4277 | tp->rcv_nxt == tp->rcv_wup) | |
4278 | tcp_store_ts_recent(tp); | |
4279 | ||
4280 | tcp_rcv_rtt_measure_ts(tp, skb); | |
4281 | ||
4282 | /* We know that such packets are checksummed | |
4283 | * on entry. | |
4284 | */ | |
4285 | tcp_ack(sk, skb, 0); | |
4286 | __kfree_skb(skb); | |
4287 | tcp_data_snd_check(sk); | |
4288 | return 0; | |
4289 | } else { /* Header too small */ | |
4290 | TCP_INC_STATS_BH(TCP_MIB_INERRS); | |
4291 | goto discard; | |
4292 | } | |
4293 | } else { | |
4294 | int eaten = 0; | |
4295 | ||
4296 | if (tp->ucopy.task == current && | |
4297 | tp->copied_seq == tp->rcv_nxt && | |
4298 | len - tcp_header_len <= tp->ucopy.len && | |
4299 | sock_owned_by_user(sk)) { | |
4300 | __set_current_state(TASK_RUNNING); | |
4301 | ||
4302 | if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) { | |
4303 | /* Predicted packet is in window by definition. | |
4304 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. | |
4305 | * Hence, check seq<=rcv_wup reduces to: | |
4306 | */ | |
4307 | if (tcp_header_len == | |
4308 | (sizeof(struct tcphdr) + | |
4309 | TCPOLEN_TSTAMP_ALIGNED) && | |
4310 | tp->rcv_nxt == tp->rcv_wup) | |
4311 | tcp_store_ts_recent(tp); | |
4312 | ||
4313 | tcp_rcv_rtt_measure_ts(tp, skb); | |
4314 | ||
4315 | __skb_pull(skb, tcp_header_len); | |
4316 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
4317 | NET_INC_STATS_BH(LINUX_MIB_TCPHPHITSTOUSER); | |
4318 | eaten = 1; | |
4319 | } | |
4320 | } | |
4321 | if (!eaten) { | |
4322 | if (tcp_checksum_complete_user(sk, skb)) | |
4323 | goto csum_error; | |
4324 | ||
4325 | /* Predicted packet is in window by definition. | |
4326 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. | |
4327 | * Hence, check seq<=rcv_wup reduces to: | |
4328 | */ | |
4329 | if (tcp_header_len == | |
4330 | (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && | |
4331 | tp->rcv_nxt == tp->rcv_wup) | |
4332 | tcp_store_ts_recent(tp); | |
4333 | ||
4334 | tcp_rcv_rtt_measure_ts(tp, skb); | |
4335 | ||
4336 | if ((int)skb->truesize > sk->sk_forward_alloc) | |
4337 | goto step5; | |
4338 | ||
4339 | NET_INC_STATS_BH(LINUX_MIB_TCPHPHITS); | |
4340 | ||
4341 | /* Bulk data transfer: receiver */ | |
4342 | __skb_pull(skb,tcp_header_len); | |
4343 | __skb_queue_tail(&sk->sk_receive_queue, skb); | |
4344 | sk_stream_set_owner_r(skb, sk); | |
4345 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
4346 | } | |
4347 | ||
4348 | tcp_event_data_recv(sk, tp, skb); | |
4349 | ||
4350 | if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { | |
4351 | /* Well, only one small jumplet in fast path... */ | |
4352 | tcp_ack(sk, skb, FLAG_DATA); | |
4353 | tcp_data_snd_check(sk); | |
4354 | if (!tcp_ack_scheduled(tp)) | |
4355 | goto no_ack; | |
4356 | } | |
4357 | ||
31432412 | 4358 | __tcp_ack_snd_check(sk, 0); |
1da177e4 LT |
4359 | no_ack: |
4360 | if (eaten) | |
4361 | __kfree_skb(skb); | |
4362 | else | |
4363 | sk->sk_data_ready(sk, 0); | |
4364 | return 0; | |
4365 | } | |
4366 | } | |
4367 | ||
4368 | slow_path: | |
4369 | if (len < (th->doff<<2) || tcp_checksum_complete_user(sk, skb)) | |
4370 | goto csum_error; | |
4371 | ||
4372 | /* | |
4373 | * RFC1323: H1. Apply PAWS check first. | |
4374 | */ | |
4375 | if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && | |
4376 | tcp_paws_discard(tp, skb)) { | |
4377 | if (!th->rst) { | |
4378 | NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); | |
4379 | tcp_send_dupack(sk, skb); | |
4380 | goto discard; | |
4381 | } | |
4382 | /* Resets are accepted even if PAWS failed. | |
4383 | ||
4384 | ts_recent update must be made after we are sure | |
4385 | that the packet is in window. | |
4386 | */ | |
4387 | } | |
4388 | ||
4389 | /* | |
4390 | * Standard slow path. | |
4391 | */ | |
4392 | ||
4393 | if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { | |
4394 | /* RFC793, page 37: "In all states except SYN-SENT, all reset | |
4395 | * (RST) segments are validated by checking their SEQ-fields." | |
4396 | * And page 69: "If an incoming segment is not acceptable, | |
4397 | * an acknowledgment should be sent in reply (unless the RST bit | |
4398 | * is set, if so drop the segment and return)". | |
4399 | */ | |
4400 | if (!th->rst) | |
4401 | tcp_send_dupack(sk, skb); | |
4402 | goto discard; | |
4403 | } | |
4404 | ||
4405 | if(th->rst) { | |
4406 | tcp_reset(sk); | |
4407 | goto discard; | |
4408 | } | |
4409 | ||
4410 | tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); | |
4411 | ||
4412 | if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
4413 | TCP_INC_STATS_BH(TCP_MIB_INERRS); | |
4414 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); | |
4415 | tcp_reset(sk); | |
4416 | return 1; | |
4417 | } | |
4418 | ||
4419 | step5: | |
4420 | if(th->ack) | |
4421 | tcp_ack(sk, skb, FLAG_SLOWPATH); | |
4422 | ||
4423 | tcp_rcv_rtt_measure_ts(tp, skb); | |
4424 | ||
4425 | /* Process urgent data. */ | |
4426 | tcp_urg(sk, skb, th); | |
4427 | ||
4428 | /* step 7: process the segment text */ | |
4429 | tcp_data_queue(sk, skb); | |
4430 | ||
4431 | tcp_data_snd_check(sk); | |
4432 | tcp_ack_snd_check(sk); | |
4433 | return 0; | |
4434 | ||
4435 | csum_error: | |
4436 | TCP_INC_STATS_BH(TCP_MIB_INERRS); | |
4437 | ||
4438 | discard: | |
4439 | __kfree_skb(skb); | |
4440 | return 0; | |
4441 | } | |
4442 | ||
4443 | static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, | |
4444 | struct tcphdr *th, unsigned len) | |
4445 | { | |
4446 | struct tcp_sock *tp = tcp_sk(sk); | |
4447 | int saved_clamp = tp->rx_opt.mss_clamp; | |
4448 | ||
4449 | tcp_parse_options(skb, &tp->rx_opt, 0); | |
4450 | ||
4451 | if (th->ack) { | |
4452 | /* rfc793: | |
4453 | * "If the state is SYN-SENT then | |
4454 | * first check the ACK bit | |
4455 | * If the ACK bit is set | |
4456 | * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send | |
4457 | * a reset (unless the RST bit is set, if so drop | |
4458 | * the segment and return)" | |
4459 | * | |
4460 | * We do not send data with SYN, so that RFC-correct | |
4461 | * test reduces to: | |
4462 | */ | |
4463 | if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) | |
4464 | goto reset_and_undo; | |
4465 | ||
4466 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && | |
4467 | !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, | |
4468 | tcp_time_stamp)) { | |
4469 | NET_INC_STATS_BH(LINUX_MIB_PAWSACTIVEREJECTED); | |
4470 | goto reset_and_undo; | |
4471 | } | |
4472 | ||
4473 | /* Now ACK is acceptable. | |
4474 | * | |
4475 | * "If the RST bit is set | |
4476 | * If the ACK was acceptable then signal the user "error: | |
4477 | * connection reset", drop the segment, enter CLOSED state, | |
4478 | * delete TCB, and return." | |
4479 | */ | |
4480 | ||
4481 | if (th->rst) { | |
4482 | tcp_reset(sk); | |
4483 | goto discard; | |
4484 | } | |
4485 | ||
4486 | /* rfc793: | |
4487 | * "fifth, if neither of the SYN or RST bits is set then | |
4488 | * drop the segment and return." | |
4489 | * | |
4490 | * See note below! | |
4491 | * --ANK(990513) | |
4492 | */ | |
4493 | if (!th->syn) | |
4494 | goto discard_and_undo; | |
4495 | ||
4496 | /* rfc793: | |
4497 | * "If the SYN bit is on ... | |
4498 | * are acceptable then ... | |
4499 | * (our SYN has been ACKed), change the connection | |
4500 | * state to ESTABLISHED..." | |
4501 | */ | |
4502 | ||
4503 | TCP_ECN_rcv_synack(tp, th); | |
4504 | if (tp->ecn_flags&TCP_ECN_OK) | |
4505 | sock_set_flag(sk, SOCK_NO_LARGESEND); | |
4506 | ||
4507 | tp->snd_wl1 = TCP_SKB_CB(skb)->seq; | |
4508 | tcp_ack(sk, skb, FLAG_SLOWPATH); | |
4509 | ||
4510 | /* Ok.. it's good. Set up sequence numbers and | |
4511 | * move to established. | |
4512 | */ | |
4513 | tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; | |
4514 | tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; | |
4515 | ||
4516 | /* RFC1323: The window in SYN & SYN/ACK segments is | |
4517 | * never scaled. | |
4518 | */ | |
4519 | tp->snd_wnd = ntohs(th->window); | |
4520 | tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq); | |
4521 | ||
4522 | if (!tp->rx_opt.wscale_ok) { | |
4523 | tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; | |
4524 | tp->window_clamp = min(tp->window_clamp, 65535U); | |
4525 | } | |
4526 | ||
4527 | if (tp->rx_opt.saw_tstamp) { | |
4528 | tp->rx_opt.tstamp_ok = 1; | |
4529 | tp->tcp_header_len = | |
4530 | sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; | |
4531 | tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; | |
4532 | tcp_store_ts_recent(tp); | |
4533 | } else { | |
4534 | tp->tcp_header_len = sizeof(struct tcphdr); | |
4535 | } | |
4536 | ||
4537 | if (tp->rx_opt.sack_ok && sysctl_tcp_fack) | |
4538 | tp->rx_opt.sack_ok |= 2; | |
4539 | ||
4540 | tcp_sync_mss(sk, tp->pmtu_cookie); | |
4541 | tcp_initialize_rcv_mss(sk); | |
4542 | ||
4543 | /* Remember, tcp_poll() does not lock socket! | |
4544 | * Change state from SYN-SENT only after copied_seq | |
4545 | * is initialized. */ | |
4546 | tp->copied_seq = tp->rcv_nxt; | |
4547 | mb(); | |
4548 | tcp_set_state(sk, TCP_ESTABLISHED); | |
4549 | ||
4550 | /* Make sure socket is routed, for correct metrics. */ | |
4551 | tp->af_specific->rebuild_header(sk); | |
4552 | ||
4553 | tcp_init_metrics(sk); | |
4554 | ||
4555 | /* Prevent spurious tcp_cwnd_restart() on first data | |
4556 | * packet. | |
4557 | */ | |
4558 | tp->lsndtime = tcp_time_stamp; | |
4559 | ||
4560 | tcp_init_buffer_space(sk); | |
4561 | ||
4562 | if (sock_flag(sk, SOCK_KEEPOPEN)) | |
4563 | tcp_reset_keepalive_timer(sk, keepalive_time_when(tp)); | |
4564 | ||
4565 | if (!tp->rx_opt.snd_wscale) | |
4566 | __tcp_fast_path_on(tp, tp->snd_wnd); | |
4567 | else | |
4568 | tp->pred_flags = 0; | |
4569 | ||
4570 | if (!sock_flag(sk, SOCK_DEAD)) { | |
4571 | sk->sk_state_change(sk); | |
4572 | sk_wake_async(sk, 0, POLL_OUT); | |
4573 | } | |
4574 | ||
4575 | if (sk->sk_write_pending || tp->defer_accept || tp->ack.pingpong) { | |
4576 | /* Save one ACK. Data will be ready after | |
4577 | * several ticks, if write_pending is set. | |
4578 | * | |
4579 | * It may be deleted, but with this feature tcpdumps | |
4580 | * look so _wonderfully_ clever, that I was not able | |
4581 | * to stand against the temptation 8) --ANK | |
4582 | */ | |
4583 | tcp_schedule_ack(tp); | |
4584 | tp->ack.lrcvtime = tcp_time_stamp; | |
4585 | tp->ack.ato = TCP_ATO_MIN; | |
4586 | tcp_incr_quickack(tp); | |
4587 | tcp_enter_quickack_mode(tp); | |
4588 | tcp_reset_xmit_timer(sk, TCP_TIME_DACK, TCP_DELACK_MAX); | |
4589 | ||
4590 | discard: | |
4591 | __kfree_skb(skb); | |
4592 | return 0; | |
4593 | } else { | |
4594 | tcp_send_ack(sk); | |
4595 | } | |
4596 | return -1; | |
4597 | } | |
4598 | ||
4599 | /* No ACK in the segment */ | |
4600 | ||
4601 | if (th->rst) { | |
4602 | /* rfc793: | |
4603 | * "If the RST bit is set | |
4604 | * | |
4605 | * Otherwise (no ACK) drop the segment and return." | |
4606 | */ | |
4607 | ||
4608 | goto discard_and_undo; | |
4609 | } | |
4610 | ||
4611 | /* PAWS check. */ | |
4612 | if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && tcp_paws_check(&tp->rx_opt, 0)) | |
4613 | goto discard_and_undo; | |
4614 | ||
4615 | if (th->syn) { | |
4616 | /* We see SYN without ACK. It is attempt of | |
4617 | * simultaneous connect with crossed SYNs. | |
4618 | * Particularly, it can be connect to self. | |
4619 | */ | |
4620 | tcp_set_state(sk, TCP_SYN_RECV); | |
4621 | ||
4622 | if (tp->rx_opt.saw_tstamp) { | |
4623 | tp->rx_opt.tstamp_ok = 1; | |
4624 | tcp_store_ts_recent(tp); | |
4625 | tp->tcp_header_len = | |
4626 | sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; | |
4627 | } else { | |
4628 | tp->tcp_header_len = sizeof(struct tcphdr); | |
4629 | } | |
4630 | ||
4631 | tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; | |
4632 | tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; | |
4633 | ||
4634 | /* RFC1323: The window in SYN & SYN/ACK segments is | |
4635 | * never scaled. | |
4636 | */ | |
4637 | tp->snd_wnd = ntohs(th->window); | |
4638 | tp->snd_wl1 = TCP_SKB_CB(skb)->seq; | |
4639 | tp->max_window = tp->snd_wnd; | |
4640 | ||
4641 | TCP_ECN_rcv_syn(tp, th); | |
4642 | if (tp->ecn_flags&TCP_ECN_OK) | |
4643 | sock_set_flag(sk, SOCK_NO_LARGESEND); | |
4644 | ||
4645 | tcp_sync_mss(sk, tp->pmtu_cookie); | |
4646 | tcp_initialize_rcv_mss(sk); | |
4647 | ||
4648 | ||
4649 | tcp_send_synack(sk); | |
4650 | #if 0 | |
4651 | /* Note, we could accept data and URG from this segment. | |
4652 | * There are no obstacles to make this. | |
4653 | * | |
4654 | * However, if we ignore data in ACKless segments sometimes, | |
4655 | * we have no reasons to accept it sometimes. | |
4656 | * Also, seems the code doing it in step6 of tcp_rcv_state_process | |
4657 | * is not flawless. So, discard packet for sanity. | |
4658 | * Uncomment this return to process the data. | |
4659 | */ | |
4660 | return -1; | |
4661 | #else | |
4662 | goto discard; | |
4663 | #endif | |
4664 | } | |
4665 | /* "fifth, if neither of the SYN or RST bits is set then | |
4666 | * drop the segment and return." | |
4667 | */ | |
4668 | ||
4669 | discard_and_undo: | |
4670 | tcp_clear_options(&tp->rx_opt); | |
4671 | tp->rx_opt.mss_clamp = saved_clamp; | |
4672 | goto discard; | |
4673 | ||
4674 | reset_and_undo: | |
4675 | tcp_clear_options(&tp->rx_opt); | |
4676 | tp->rx_opt.mss_clamp = saved_clamp; | |
4677 | return 1; | |
4678 | } | |
4679 | ||
4680 | ||
4681 | /* | |
4682 | * This function implements the receiving procedure of RFC 793 for | |
4683 | * all states except ESTABLISHED and TIME_WAIT. | |
4684 | * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be | |
4685 | * address independent. | |
4686 | */ | |
4687 | ||
4688 | int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, | |
4689 | struct tcphdr *th, unsigned len) | |
4690 | { | |
4691 | struct tcp_sock *tp = tcp_sk(sk); | |
4692 | int queued = 0; | |
4693 | ||
4694 | tp->rx_opt.saw_tstamp = 0; | |
4695 | ||
4696 | switch (sk->sk_state) { | |
4697 | case TCP_CLOSE: | |
4698 | goto discard; | |
4699 | ||
4700 | case TCP_LISTEN: | |
4701 | if(th->ack) | |
4702 | return 1; | |
4703 | ||
4704 | if(th->rst) | |
4705 | goto discard; | |
4706 | ||
4707 | if(th->syn) { | |
4708 | if(tp->af_specific->conn_request(sk, skb) < 0) | |
4709 | return 1; | |
4710 | ||
4711 | init_westwood(sk); | |
4712 | init_bictcp(tp); | |
4713 | ||
4714 | /* Now we have several options: In theory there is | |
4715 | * nothing else in the frame. KA9Q has an option to | |
4716 | * send data with the syn, BSD accepts data with the | |
4717 | * syn up to the [to be] advertised window and | |
4718 | * Solaris 2.1 gives you a protocol error. For now | |
4719 | * we just ignore it, that fits the spec precisely | |
4720 | * and avoids incompatibilities. It would be nice in | |
4721 | * future to drop through and process the data. | |
4722 | * | |
4723 | * Now that TTCP is starting to be used we ought to | |
4724 | * queue this data. | |
4725 | * But, this leaves one open to an easy denial of | |
4726 | * service attack, and SYN cookies can't defend | |
4727 | * against this problem. So, we drop the data | |
4728 | * in the interest of security over speed. | |
4729 | */ | |
4730 | goto discard; | |
4731 | } | |
4732 | goto discard; | |
4733 | ||
4734 | case TCP_SYN_SENT: | |
4735 | init_westwood(sk); | |
4736 | init_bictcp(tp); | |
4737 | ||
4738 | queued = tcp_rcv_synsent_state_process(sk, skb, th, len); | |
4739 | if (queued >= 0) | |
4740 | return queued; | |
4741 | ||
4742 | /* Do step6 onward by hand. */ | |
4743 | tcp_urg(sk, skb, th); | |
4744 | __kfree_skb(skb); | |
4745 | tcp_data_snd_check(sk); | |
4746 | return 0; | |
4747 | } | |
4748 | ||
4749 | if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && | |
4750 | tcp_paws_discard(tp, skb)) { | |
4751 | if (!th->rst) { | |
4752 | NET_INC_STATS_BH(LINUX_MIB_PAWSESTABREJECTED); | |
4753 | tcp_send_dupack(sk, skb); | |
4754 | goto discard; | |
4755 | } | |
4756 | /* Reset is accepted even if it did not pass PAWS. */ | |
4757 | } | |
4758 | ||
4759 | /* step 1: check sequence number */ | |
4760 | if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { | |
4761 | if (!th->rst) | |
4762 | tcp_send_dupack(sk, skb); | |
4763 | goto discard; | |
4764 | } | |
4765 | ||
4766 | /* step 2: check RST bit */ | |
4767 | if(th->rst) { | |
4768 | tcp_reset(sk); | |
4769 | goto discard; | |
4770 | } | |
4771 | ||
4772 | tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); | |
4773 | ||
4774 | /* step 3: check security and precedence [ignored] */ | |
4775 | ||
4776 | /* step 4: | |
4777 | * | |
4778 | * Check for a SYN in window. | |
4779 | */ | |
4780 | if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
4781 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONSYN); | |
4782 | tcp_reset(sk); | |
4783 | return 1; | |
4784 | } | |
4785 | ||
4786 | /* step 5: check the ACK field */ | |
4787 | if (th->ack) { | |
4788 | int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH); | |
4789 | ||
4790 | switch(sk->sk_state) { | |
4791 | case TCP_SYN_RECV: | |
4792 | if (acceptable) { | |
4793 | tp->copied_seq = tp->rcv_nxt; | |
4794 | mb(); | |
4795 | tcp_set_state(sk, TCP_ESTABLISHED); | |
4796 | sk->sk_state_change(sk); | |
4797 | ||
4798 | /* Note, that this wakeup is only for marginal | |
4799 | * crossed SYN case. Passively open sockets | |
4800 | * are not waked up, because sk->sk_sleep == | |
4801 | * NULL and sk->sk_socket == NULL. | |
4802 | */ | |
4803 | if (sk->sk_socket) { | |
4804 | sk_wake_async(sk,0,POLL_OUT); | |
4805 | } | |
4806 | ||
4807 | tp->snd_una = TCP_SKB_CB(skb)->ack_seq; | |
4808 | tp->snd_wnd = ntohs(th->window) << | |
4809 | tp->rx_opt.snd_wscale; | |
4810 | tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, | |
4811 | TCP_SKB_CB(skb)->seq); | |
4812 | ||
4813 | /* tcp_ack considers this ACK as duplicate | |
4814 | * and does not calculate rtt. | |
4815 | * Fix it at least with timestamps. | |
4816 | */ | |
4817 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && | |
4818 | !tp->srtt) | |
4819 | tcp_ack_saw_tstamp(tp, 0); | |
4820 | ||
4821 | if (tp->rx_opt.tstamp_ok) | |
4822 | tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; | |
4823 | ||
4824 | /* Make sure socket is routed, for | |
4825 | * correct metrics. | |
4826 | */ | |
4827 | tp->af_specific->rebuild_header(sk); | |
4828 | ||
4829 | tcp_init_metrics(sk); | |
4830 | ||
4831 | /* Prevent spurious tcp_cwnd_restart() on | |
4832 | * first data packet. | |
4833 | */ | |
4834 | tp->lsndtime = tcp_time_stamp; | |
4835 | ||
4836 | tcp_initialize_rcv_mss(sk); | |
4837 | tcp_init_buffer_space(sk); | |
4838 | tcp_fast_path_on(tp); | |
4839 | } else { | |
4840 | return 1; | |
4841 | } | |
4842 | break; | |
4843 | ||
4844 | case TCP_FIN_WAIT1: | |
4845 | if (tp->snd_una == tp->write_seq) { | |
4846 | tcp_set_state(sk, TCP_FIN_WAIT2); | |
4847 | sk->sk_shutdown |= SEND_SHUTDOWN; | |
4848 | dst_confirm(sk->sk_dst_cache); | |
4849 | ||
4850 | if (!sock_flag(sk, SOCK_DEAD)) | |
4851 | /* Wake up lingering close() */ | |
4852 | sk->sk_state_change(sk); | |
4853 | else { | |
4854 | int tmo; | |
4855 | ||
4856 | if (tp->linger2 < 0 || | |
4857 | (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && | |
4858 | after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { | |
4859 | tcp_done(sk); | |
4860 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); | |
4861 | return 1; | |
4862 | } | |
4863 | ||
4864 | tmo = tcp_fin_time(tp); | |
4865 | if (tmo > TCP_TIMEWAIT_LEN) { | |
4866 | tcp_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); | |
4867 | } else if (th->fin || sock_owned_by_user(sk)) { | |
4868 | /* Bad case. We could lose such FIN otherwise. | |
4869 | * It is not a big problem, but it looks confusing | |
4870 | * and not so rare event. We still can lose it now, | |
4871 | * if it spins in bh_lock_sock(), but it is really | |
4872 | * marginal case. | |
4873 | */ | |
4874 | tcp_reset_keepalive_timer(sk, tmo); | |
4875 | } else { | |
4876 | tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); | |
4877 | goto discard; | |
4878 | } | |
4879 | } | |
4880 | } | |
4881 | break; | |
4882 | ||
4883 | case TCP_CLOSING: | |
4884 | if (tp->snd_una == tp->write_seq) { | |
4885 | tcp_time_wait(sk, TCP_TIME_WAIT, 0); | |
4886 | goto discard; | |
4887 | } | |
4888 | break; | |
4889 | ||
4890 | case TCP_LAST_ACK: | |
4891 | if (tp->snd_una == tp->write_seq) { | |
4892 | tcp_update_metrics(sk); | |
4893 | tcp_done(sk); | |
4894 | goto discard; | |
4895 | } | |
4896 | break; | |
4897 | } | |
4898 | } else | |
4899 | goto discard; | |
4900 | ||
4901 | /* step 6: check the URG bit */ | |
4902 | tcp_urg(sk, skb, th); | |
4903 | ||
4904 | /* step 7: process the segment text */ | |
4905 | switch (sk->sk_state) { | |
4906 | case TCP_CLOSE_WAIT: | |
4907 | case TCP_CLOSING: | |
4908 | case TCP_LAST_ACK: | |
4909 | if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) | |
4910 | break; | |
4911 | case TCP_FIN_WAIT1: | |
4912 | case TCP_FIN_WAIT2: | |
4913 | /* RFC 793 says to queue data in these states, | |
4914 | * RFC 1122 says we MUST send a reset. | |
4915 | * BSD 4.4 also does reset. | |
4916 | */ | |
4917 | if (sk->sk_shutdown & RCV_SHUTDOWN) { | |
4918 | if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && | |
4919 | after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { | |
4920 | NET_INC_STATS_BH(LINUX_MIB_TCPABORTONDATA); | |
4921 | tcp_reset(sk); | |
4922 | return 1; | |
4923 | } | |
4924 | } | |
4925 | /* Fall through */ | |
4926 | case TCP_ESTABLISHED: | |
4927 | tcp_data_queue(sk, skb); | |
4928 | queued = 1; | |
4929 | break; | |
4930 | } | |
4931 | ||
4932 | /* tcp_data could move socket to TIME-WAIT */ | |
4933 | if (sk->sk_state != TCP_CLOSE) { | |
4934 | tcp_data_snd_check(sk); | |
4935 | tcp_ack_snd_check(sk); | |
4936 | } | |
4937 | ||
4938 | if (!queued) { | |
4939 | discard: | |
4940 | __kfree_skb(skb); | |
4941 | } | |
4942 | return 0; | |
4943 | } | |
4944 | ||
4945 | EXPORT_SYMBOL(sysctl_tcp_ecn); | |
4946 | EXPORT_SYMBOL(sysctl_tcp_reordering); | |
4947 | EXPORT_SYMBOL(tcp_parse_options); | |
4948 | EXPORT_SYMBOL(tcp_rcv_established); | |
4949 | EXPORT_SYMBOL(tcp_rcv_state_process); |