]>
Commit | Line | Data |
---|---|---|
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 | * Authors: Ross Biro | |
9 | * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> | |
10 | * Mark Evans, <evansmp@uhura.aston.ac.uk> | |
11 | * Corey Minyard <wf-rch!minyard@relay.EU.net> | |
12 | * Florian La Roche, <flla@stud.uni-sb.de> | |
13 | * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> | |
14 | * Linus Torvalds, <torvalds@cs.helsinki.fi> | |
15 | * Alan Cox, <gw4pts@gw4pts.ampr.org> | |
16 | * Matthew Dillon, <dillon@apollo.west.oic.com> | |
17 | * Arnt Gulbrandsen, <agulbra@nvg.unit.no> | |
18 | * Jorge Cwik, <jorge@laser.satlink.net> | |
19 | */ | |
20 | ||
21 | /* | |
22 | * Changes: | |
23 | * Pedro Roque : Fast Retransmit/Recovery. | |
24 | * Two receive queues. | |
25 | * Retransmit queue handled by TCP. | |
26 | * Better retransmit timer handling. | |
27 | * New congestion avoidance. | |
28 | * Header prediction. | |
29 | * Variable renaming. | |
30 | * | |
31 | * Eric : Fast Retransmit. | |
32 | * Randy Scott : MSS option defines. | |
33 | * Eric Schenk : Fixes to slow start algorithm. | |
34 | * Eric Schenk : Yet another double ACK bug. | |
35 | * Eric Schenk : Delayed ACK bug fixes. | |
36 | * Eric Schenk : Floyd style fast retrans war avoidance. | |
37 | * David S. Miller : Don't allow zero congestion window. | |
38 | * Eric Schenk : Fix retransmitter so that it sends | |
39 | * next packet on ack of previous packet. | |
40 | * Andi Kleen : Moved open_request checking here | |
41 | * and process RSTs for open_requests. | |
42 | * Andi Kleen : Better prune_queue, and other fixes. | |
43 | * Andrey Savochkin: Fix RTT measurements in the presence of | |
44 | * timestamps. | |
45 | * Andrey Savochkin: Check sequence numbers correctly when | |
46 | * removing SACKs due to in sequence incoming | |
47 | * data segments. | |
48 | * Andi Kleen: Make sure we never ack data there is not | |
49 | * enough room for. Also make this condition | |
50 | * a fatal error if it might still happen. | |
51 | * Andi Kleen: Add tcp_measure_rcv_mss to make | |
52 | * connections with MSS<min(MTU,ann. MSS) | |
53 | * work without delayed acks. | |
54 | * Andi Kleen: Process packets with PSH set in the | |
55 | * fast path. | |
56 | * J Hadi Salim: ECN support | |
57 | * Andrei Gurtov, | |
58 | * Pasi Sarolahti, | |
59 | * Panu Kuhlberg: Experimental audit of TCP (re)transmission | |
60 | * engine. Lots of bugs are found. | |
61 | * Pasi Sarolahti: F-RTO for dealing with spurious RTOs | |
62 | */ | |
63 | ||
64 | #include <linux/mm.h> | |
65 | #include <linux/module.h> | |
66 | #include <linux/sysctl.h> | |
67 | #include <net/dst.h> | |
68 | #include <net/tcp.h> | |
69 | #include <net/inet_common.h> | |
70 | #include <linux/ipsec.h> | |
71 | #include <asm/unaligned.h> | |
72 | #include <net/netdma.h> | |
73 | ||
74 | int sysctl_tcp_timestamps __read_mostly = 1; | |
75 | int sysctl_tcp_window_scaling __read_mostly = 1; | |
76 | int sysctl_tcp_sack __read_mostly = 1; | |
77 | int sysctl_tcp_fack __read_mostly = 1; | |
78 | int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH; | |
79 | int sysctl_tcp_ecn __read_mostly; | |
80 | int sysctl_tcp_dsack __read_mostly = 1; | |
81 | int sysctl_tcp_app_win __read_mostly = 31; | |
82 | int sysctl_tcp_adv_win_scale __read_mostly = 2; | |
83 | ||
84 | int sysctl_tcp_stdurg __read_mostly; | |
85 | int sysctl_tcp_rfc1337 __read_mostly; | |
86 | int sysctl_tcp_max_orphans __read_mostly = NR_FILE; | |
87 | int sysctl_tcp_frto __read_mostly = 2; | |
88 | int sysctl_tcp_frto_response __read_mostly; | |
89 | int sysctl_tcp_nometrics_save __read_mostly; | |
90 | ||
91 | int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; | |
92 | int sysctl_tcp_abc __read_mostly; | |
93 | ||
94 | #define FLAG_DATA 0x01 /* Incoming frame contained data. */ | |
95 | #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ | |
96 | #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ | |
97 | #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ | |
98 | #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ | |
99 | #define FLAG_DATA_SACKED 0x20 /* New SACK. */ | |
100 | #define FLAG_ECE 0x40 /* ECE in this ACK */ | |
101 | #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */ | |
102 | #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ | |
103 | #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */ | |
104 | #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ | |
105 | #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ | |
106 | #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */ | |
107 | #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ | |
108 | ||
109 | #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) | |
110 | #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) | |
111 | #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) | |
112 | #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) | |
113 | #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED) | |
114 | ||
115 | #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) | |
116 | #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) | |
117 | ||
118 | /* Adapt the MSS value used to make delayed ack decision to the | |
119 | * real world. | |
120 | */ | |
121 | static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) | |
122 | { | |
123 | struct inet_connection_sock *icsk = inet_csk(sk); | |
124 | const unsigned int lss = icsk->icsk_ack.last_seg_size; | |
125 | unsigned int len; | |
126 | ||
127 | icsk->icsk_ack.last_seg_size = 0; | |
128 | ||
129 | /* skb->len may jitter because of SACKs, even if peer | |
130 | * sends good full-sized frames. | |
131 | */ | |
132 | len = skb_shinfo(skb)->gso_size ? : skb->len; | |
133 | if (len >= icsk->icsk_ack.rcv_mss) { | |
134 | icsk->icsk_ack.rcv_mss = len; | |
135 | } else { | |
136 | /* Otherwise, we make more careful check taking into account, | |
137 | * that SACKs block is variable. | |
138 | * | |
139 | * "len" is invariant segment length, including TCP header. | |
140 | */ | |
141 | len += skb->data - skb_transport_header(skb); | |
142 | if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) || | |
143 | /* If PSH is not set, packet should be | |
144 | * full sized, provided peer TCP is not badly broken. | |
145 | * This observation (if it is correct 8)) allows | |
146 | * to handle super-low mtu links fairly. | |
147 | */ | |
148 | (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && | |
149 | !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { | |
150 | /* Subtract also invariant (if peer is RFC compliant), | |
151 | * tcp header plus fixed timestamp option length. | |
152 | * Resulting "len" is MSS free of SACK jitter. | |
153 | */ | |
154 | len -= tcp_sk(sk)->tcp_header_len; | |
155 | icsk->icsk_ack.last_seg_size = len; | |
156 | if (len == lss) { | |
157 | icsk->icsk_ack.rcv_mss = len; | |
158 | return; | |
159 | } | |
160 | } | |
161 | if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) | |
162 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; | |
163 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; | |
164 | } | |
165 | } | |
166 | ||
167 | static void tcp_incr_quickack(struct sock *sk) | |
168 | { | |
169 | struct inet_connection_sock *icsk = inet_csk(sk); | |
170 | unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); | |
171 | ||
172 | if (quickacks == 0) | |
173 | quickacks = 2; | |
174 | if (quickacks > icsk->icsk_ack.quick) | |
175 | icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); | |
176 | } | |
177 | ||
178 | void tcp_enter_quickack_mode(struct sock *sk) | |
179 | { | |
180 | struct inet_connection_sock *icsk = inet_csk(sk); | |
181 | tcp_incr_quickack(sk); | |
182 | icsk->icsk_ack.pingpong = 0; | |
183 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
184 | } | |
185 | ||
186 | /* Send ACKs quickly, if "quick" count is not exhausted | |
187 | * and the session is not interactive. | |
188 | */ | |
189 | ||
190 | static inline int tcp_in_quickack_mode(const struct sock *sk) | |
191 | { | |
192 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
193 | return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; | |
194 | } | |
195 | ||
196 | static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp) | |
197 | { | |
198 | if (tp->ecn_flags & TCP_ECN_OK) | |
199 | tp->ecn_flags |= TCP_ECN_QUEUE_CWR; | |
200 | } | |
201 | ||
202 | static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb) | |
203 | { | |
204 | if (tcp_hdr(skb)->cwr) | |
205 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
206 | } | |
207 | ||
208 | static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp) | |
209 | { | |
210 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
211 | } | |
212 | ||
213 | static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb) | |
214 | { | |
215 | if (tp->ecn_flags & TCP_ECN_OK) { | |
216 | if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags)) | |
217 | tp->ecn_flags |= TCP_ECN_DEMAND_CWR; | |
218 | /* Funny extension: if ECT is not set on a segment, | |
219 | * it is surely retransmit. It is not in ECN RFC, | |
220 | * but Linux follows this rule. */ | |
221 | else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags))) | |
222 | tcp_enter_quickack_mode((struct sock *)tp); | |
223 | } | |
224 | } | |
225 | ||
226 | static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th) | |
227 | { | |
228 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) | |
229 | tp->ecn_flags &= ~TCP_ECN_OK; | |
230 | } | |
231 | ||
232 | static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th) | |
233 | { | |
234 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) | |
235 | tp->ecn_flags &= ~TCP_ECN_OK; | |
236 | } | |
237 | ||
238 | static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th) | |
239 | { | |
240 | if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) | |
241 | return 1; | |
242 | return 0; | |
243 | } | |
244 | ||
245 | /* Buffer size and advertised window tuning. | |
246 | * | |
247 | * 1. Tuning sk->sk_sndbuf, when connection enters established state. | |
248 | */ | |
249 | ||
250 | static void tcp_fixup_sndbuf(struct sock *sk) | |
251 | { | |
252 | int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 + | |
253 | sizeof(struct sk_buff); | |
254 | ||
255 | if (sk->sk_sndbuf < 3 * sndmem) | |
256 | sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]); | |
257 | } | |
258 | ||
259 | /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) | |
260 | * | |
261 | * All tcp_full_space() is split to two parts: "network" buffer, allocated | |
262 | * forward and advertised in receiver window (tp->rcv_wnd) and | |
263 | * "application buffer", required to isolate scheduling/application | |
264 | * latencies from network. | |
265 | * window_clamp is maximal advertised window. It can be less than | |
266 | * tcp_full_space(), in this case tcp_full_space() - window_clamp | |
267 | * is reserved for "application" buffer. The less window_clamp is | |
268 | * the smoother our behaviour from viewpoint of network, but the lower | |
269 | * throughput and the higher sensitivity of the connection to losses. 8) | |
270 | * | |
271 | * rcv_ssthresh is more strict window_clamp used at "slow start" | |
272 | * phase to predict further behaviour of this connection. | |
273 | * It is used for two goals: | |
274 | * - to enforce header prediction at sender, even when application | |
275 | * requires some significant "application buffer". It is check #1. | |
276 | * - to prevent pruning of receive queue because of misprediction | |
277 | * of receiver window. Check #2. | |
278 | * | |
279 | * The scheme does not work when sender sends good segments opening | |
280 | * window and then starts to feed us spaghetti. But it should work | |
281 | * in common situations. Otherwise, we have to rely on queue collapsing. | |
282 | */ | |
283 | ||
284 | /* Slow part of check#2. */ | |
285 | static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) | |
286 | { | |
287 | struct tcp_sock *tp = tcp_sk(sk); | |
288 | /* Optimize this! */ | |
289 | int truesize = tcp_win_from_space(skb->truesize) >> 1; | |
290 | int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1; | |
291 | ||
292 | while (tp->rcv_ssthresh <= window) { | |
293 | if (truesize <= skb->len) | |
294 | return 2 * inet_csk(sk)->icsk_ack.rcv_mss; | |
295 | ||
296 | truesize >>= 1; | |
297 | window >>= 1; | |
298 | } | |
299 | return 0; | |
300 | } | |
301 | ||
302 | static void tcp_grow_window(struct sock *sk, struct sk_buff *skb) | |
303 | { | |
304 | struct tcp_sock *tp = tcp_sk(sk); | |
305 | ||
306 | /* Check #1 */ | |
307 | if (tp->rcv_ssthresh < tp->window_clamp && | |
308 | (int)tp->rcv_ssthresh < tcp_space(sk) && | |
309 | !tcp_memory_pressure) { | |
310 | int incr; | |
311 | ||
312 | /* Check #2. Increase window, if skb with such overhead | |
313 | * will fit to rcvbuf in future. | |
314 | */ | |
315 | if (tcp_win_from_space(skb->truesize) <= skb->len) | |
316 | incr = 2 * tp->advmss; | |
317 | else | |
318 | incr = __tcp_grow_window(sk, skb); | |
319 | ||
320 | if (incr) { | |
321 | tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, | |
322 | tp->window_clamp); | |
323 | inet_csk(sk)->icsk_ack.quick |= 1; | |
324 | } | |
325 | } | |
326 | } | |
327 | ||
328 | /* 3. Tuning rcvbuf, when connection enters established state. */ | |
329 | ||
330 | static void tcp_fixup_rcvbuf(struct sock *sk) | |
331 | { | |
332 | struct tcp_sock *tp = tcp_sk(sk); | |
333 | int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); | |
334 | ||
335 | /* Try to select rcvbuf so that 4 mss-sized segments | |
336 | * will fit to window and corresponding skbs will fit to our rcvbuf. | |
337 | * (was 3; 4 is minimum to allow fast retransmit to work.) | |
338 | */ | |
339 | while (tcp_win_from_space(rcvmem) < tp->advmss) | |
340 | rcvmem += 128; | |
341 | if (sk->sk_rcvbuf < 4 * rcvmem) | |
342 | sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]); | |
343 | } | |
344 | ||
345 | /* 4. Try to fixup all. It is made immediately after connection enters | |
346 | * established state. | |
347 | */ | |
348 | static void tcp_init_buffer_space(struct sock *sk) | |
349 | { | |
350 | struct tcp_sock *tp = tcp_sk(sk); | |
351 | int maxwin; | |
352 | ||
353 | if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) | |
354 | tcp_fixup_rcvbuf(sk); | |
355 | if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) | |
356 | tcp_fixup_sndbuf(sk); | |
357 | ||
358 | tp->rcvq_space.space = tp->rcv_wnd; | |
359 | ||
360 | maxwin = tcp_full_space(sk); | |
361 | ||
362 | if (tp->window_clamp >= maxwin) { | |
363 | tp->window_clamp = maxwin; | |
364 | ||
365 | if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) | |
366 | tp->window_clamp = max(maxwin - | |
367 | (maxwin >> sysctl_tcp_app_win), | |
368 | 4 * tp->advmss); | |
369 | } | |
370 | ||
371 | /* Force reservation of one segment. */ | |
372 | if (sysctl_tcp_app_win && | |
373 | tp->window_clamp > 2 * tp->advmss && | |
374 | tp->window_clamp + tp->advmss > maxwin) | |
375 | tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); | |
376 | ||
377 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); | |
378 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
379 | } | |
380 | ||
381 | /* 5. Recalculate window clamp after socket hit its memory bounds. */ | |
382 | static void tcp_clamp_window(struct sock *sk) | |
383 | { | |
384 | struct tcp_sock *tp = tcp_sk(sk); | |
385 | struct inet_connection_sock *icsk = inet_csk(sk); | |
386 | ||
387 | icsk->icsk_ack.quick = 0; | |
388 | ||
389 | if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && | |
390 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && | |
391 | !tcp_memory_pressure && | |
392 | atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) { | |
393 | sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), | |
394 | sysctl_tcp_rmem[2]); | |
395 | } | |
396 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) | |
397 | tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); | |
398 | } | |
399 | ||
400 | /* Initialize RCV_MSS value. | |
401 | * RCV_MSS is an our guess about MSS used by the peer. | |
402 | * We haven't any direct information about the MSS. | |
403 | * It's better to underestimate the RCV_MSS rather than overestimate. | |
404 | * Overestimations make us ACKing less frequently than needed. | |
405 | * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). | |
406 | */ | |
407 | void tcp_initialize_rcv_mss(struct sock *sk) | |
408 | { | |
409 | struct tcp_sock *tp = tcp_sk(sk); | |
410 | unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); | |
411 | ||
412 | hint = min(hint, tp->rcv_wnd / 2); | |
413 | hint = min(hint, TCP_MIN_RCVMSS); | |
414 | hint = max(hint, TCP_MIN_MSS); | |
415 | ||
416 | inet_csk(sk)->icsk_ack.rcv_mss = hint; | |
417 | } | |
418 | ||
419 | /* Receiver "autotuning" code. | |
420 | * | |
421 | * The algorithm for RTT estimation w/o timestamps is based on | |
422 | * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. | |
423 | * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps> | |
424 | * | |
425 | * More detail on this code can be found at | |
426 | * <http://www.psc.edu/~jheffner/senior_thesis.ps>, | |
427 | * though this reference is out of date. A new paper | |
428 | * is pending. | |
429 | */ | |
430 | static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) | |
431 | { | |
432 | u32 new_sample = tp->rcv_rtt_est.rtt; | |
433 | long m = sample; | |
434 | ||
435 | if (m == 0) | |
436 | m = 1; | |
437 | ||
438 | if (new_sample != 0) { | |
439 | /* If we sample in larger samples in the non-timestamp | |
440 | * case, we could grossly overestimate the RTT especially | |
441 | * with chatty applications or bulk transfer apps which | |
442 | * are stalled on filesystem I/O. | |
443 | * | |
444 | * Also, since we are only going for a minimum in the | |
445 | * non-timestamp case, we do not smooth things out | |
446 | * else with timestamps disabled convergence takes too | |
447 | * long. | |
448 | */ | |
449 | if (!win_dep) { | |
450 | m -= (new_sample >> 3); | |
451 | new_sample += m; | |
452 | } else if (m < new_sample) | |
453 | new_sample = m << 3; | |
454 | } else { | |
455 | /* No previous measure. */ | |
456 | new_sample = m << 3; | |
457 | } | |
458 | ||
459 | if (tp->rcv_rtt_est.rtt != new_sample) | |
460 | tp->rcv_rtt_est.rtt = new_sample; | |
461 | } | |
462 | ||
463 | static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) | |
464 | { | |
465 | if (tp->rcv_rtt_est.time == 0) | |
466 | goto new_measure; | |
467 | if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) | |
468 | return; | |
469 | tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1); | |
470 | ||
471 | new_measure: | |
472 | tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; | |
473 | tp->rcv_rtt_est.time = tcp_time_stamp; | |
474 | } | |
475 | ||
476 | static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, | |
477 | const struct sk_buff *skb) | |
478 | { | |
479 | struct tcp_sock *tp = tcp_sk(sk); | |
480 | if (tp->rx_opt.rcv_tsecr && | |
481 | (TCP_SKB_CB(skb)->end_seq - | |
482 | TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) | |
483 | tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); | |
484 | } | |
485 | ||
486 | /* | |
487 | * This function should be called every time data is copied to user space. | |
488 | * It calculates the appropriate TCP receive buffer space. | |
489 | */ | |
490 | void tcp_rcv_space_adjust(struct sock *sk) | |
491 | { | |
492 | struct tcp_sock *tp = tcp_sk(sk); | |
493 | int time; | |
494 | int space; | |
495 | ||
496 | if (tp->rcvq_space.time == 0) | |
497 | goto new_measure; | |
498 | ||
499 | time = tcp_time_stamp - tp->rcvq_space.time; | |
500 | if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0) | |
501 | return; | |
502 | ||
503 | space = 2 * (tp->copied_seq - tp->rcvq_space.seq); | |
504 | ||
505 | space = max(tp->rcvq_space.space, space); | |
506 | ||
507 | if (tp->rcvq_space.space != space) { | |
508 | int rcvmem; | |
509 | ||
510 | tp->rcvq_space.space = space; | |
511 | ||
512 | if (sysctl_tcp_moderate_rcvbuf && | |
513 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { | |
514 | int new_clamp = space; | |
515 | ||
516 | /* Receive space grows, normalize in order to | |
517 | * take into account packet headers and sk_buff | |
518 | * structure overhead. | |
519 | */ | |
520 | space /= tp->advmss; | |
521 | if (!space) | |
522 | space = 1; | |
523 | rcvmem = (tp->advmss + MAX_TCP_HEADER + | |
524 | 16 + sizeof(struct sk_buff)); | |
525 | while (tcp_win_from_space(rcvmem) < tp->advmss) | |
526 | rcvmem += 128; | |
527 | space *= rcvmem; | |
528 | space = min(space, sysctl_tcp_rmem[2]); | |
529 | if (space > sk->sk_rcvbuf) { | |
530 | sk->sk_rcvbuf = space; | |
531 | ||
532 | /* Make the window clamp follow along. */ | |
533 | tp->window_clamp = new_clamp; | |
534 | } | |
535 | } | |
536 | } | |
537 | ||
538 | new_measure: | |
539 | tp->rcvq_space.seq = tp->copied_seq; | |
540 | tp->rcvq_space.time = tcp_time_stamp; | |
541 | } | |
542 | ||
543 | /* There is something which you must keep in mind when you analyze the | |
544 | * behavior of the tp->ato delayed ack timeout interval. When a | |
545 | * connection starts up, we want to ack as quickly as possible. The | |
546 | * problem is that "good" TCP's do slow start at the beginning of data | |
547 | * transmission. The means that until we send the first few ACK's the | |
548 | * sender will sit on his end and only queue most of his data, because | |
549 | * he can only send snd_cwnd unacked packets at any given time. For | |
550 | * each ACK we send, he increments snd_cwnd and transmits more of his | |
551 | * queue. -DaveM | |
552 | */ | |
553 | static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) | |
554 | { | |
555 | struct tcp_sock *tp = tcp_sk(sk); | |
556 | struct inet_connection_sock *icsk = inet_csk(sk); | |
557 | u32 now; | |
558 | ||
559 | inet_csk_schedule_ack(sk); | |
560 | ||
561 | tcp_measure_rcv_mss(sk, skb); | |
562 | ||
563 | tcp_rcv_rtt_measure(tp); | |
564 | ||
565 | now = tcp_time_stamp; | |
566 | ||
567 | if (!icsk->icsk_ack.ato) { | |
568 | /* The _first_ data packet received, initialize | |
569 | * delayed ACK engine. | |
570 | */ | |
571 | tcp_incr_quickack(sk); | |
572 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
573 | } else { | |
574 | int m = now - icsk->icsk_ack.lrcvtime; | |
575 | ||
576 | if (m <= TCP_ATO_MIN / 2) { | |
577 | /* The fastest case is the first. */ | |
578 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; | |
579 | } else if (m < icsk->icsk_ack.ato) { | |
580 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; | |
581 | if (icsk->icsk_ack.ato > icsk->icsk_rto) | |
582 | icsk->icsk_ack.ato = icsk->icsk_rto; | |
583 | } else if (m > icsk->icsk_rto) { | |
584 | /* Too long gap. Apparently sender failed to | |
585 | * restart window, so that we send ACKs quickly. | |
586 | */ | |
587 | tcp_incr_quickack(sk); | |
588 | sk_mem_reclaim(sk); | |
589 | } | |
590 | } | |
591 | icsk->icsk_ack.lrcvtime = now; | |
592 | ||
593 | TCP_ECN_check_ce(tp, skb); | |
594 | ||
595 | if (skb->len >= 128) | |
596 | tcp_grow_window(sk, skb); | |
597 | } | |
598 | ||
599 | static u32 tcp_rto_min(struct sock *sk) | |
600 | { | |
601 | struct dst_entry *dst = __sk_dst_get(sk); | |
602 | u32 rto_min = TCP_RTO_MIN; | |
603 | ||
604 | if (dst && dst_metric_locked(dst, RTAX_RTO_MIN)) | |
605 | rto_min = dst_metric_rtt(dst, RTAX_RTO_MIN); | |
606 | return rto_min; | |
607 | } | |
608 | ||
609 | /* Called to compute a smoothed rtt estimate. The data fed to this | |
610 | * routine either comes from timestamps, or from segments that were | |
611 | * known _not_ to have been retransmitted [see Karn/Partridge | |
612 | * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 | |
613 | * piece by Van Jacobson. | |
614 | * NOTE: the next three routines used to be one big routine. | |
615 | * To save cycles in the RFC 1323 implementation it was better to break | |
616 | * it up into three procedures. -- erics | |
617 | */ | |
618 | static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) | |
619 | { | |
620 | struct tcp_sock *tp = tcp_sk(sk); | |
621 | long m = mrtt; /* RTT */ | |
622 | ||
623 | /* The following amusing code comes from Jacobson's | |
624 | * article in SIGCOMM '88. Note that rtt and mdev | |
625 | * are scaled versions of rtt and mean deviation. | |
626 | * This is designed to be as fast as possible | |
627 | * m stands for "measurement". | |
628 | * | |
629 | * On a 1990 paper the rto value is changed to: | |
630 | * RTO = rtt + 4 * mdev | |
631 | * | |
632 | * Funny. This algorithm seems to be very broken. | |
633 | * These formulae increase RTO, when it should be decreased, increase | |
634 | * too slowly, when it should be increased quickly, decrease too quickly | |
635 | * etc. I guess in BSD RTO takes ONE value, so that it is absolutely | |
636 | * does not matter how to _calculate_ it. Seems, it was trap | |
637 | * that VJ failed to avoid. 8) | |
638 | */ | |
639 | if (m == 0) | |
640 | m = 1; | |
641 | if (tp->srtt != 0) { | |
642 | m -= (tp->srtt >> 3); /* m is now error in rtt est */ | |
643 | tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ | |
644 | if (m < 0) { | |
645 | m = -m; /* m is now abs(error) */ | |
646 | m -= (tp->mdev >> 2); /* similar update on mdev */ | |
647 | /* This is similar to one of Eifel findings. | |
648 | * Eifel blocks mdev updates when rtt decreases. | |
649 | * This solution is a bit different: we use finer gain | |
650 | * for mdev in this case (alpha*beta). | |
651 | * Like Eifel it also prevents growth of rto, | |
652 | * but also it limits too fast rto decreases, | |
653 | * happening in pure Eifel. | |
654 | */ | |
655 | if (m > 0) | |
656 | m >>= 3; | |
657 | } else { | |
658 | m -= (tp->mdev >> 2); /* similar update on mdev */ | |
659 | } | |
660 | tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ | |
661 | if (tp->mdev > tp->mdev_max) { | |
662 | tp->mdev_max = tp->mdev; | |
663 | if (tp->mdev_max > tp->rttvar) | |
664 | tp->rttvar = tp->mdev_max; | |
665 | } | |
666 | if (after(tp->snd_una, tp->rtt_seq)) { | |
667 | if (tp->mdev_max < tp->rttvar) | |
668 | tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2; | |
669 | tp->rtt_seq = tp->snd_nxt; | |
670 | tp->mdev_max = tcp_rto_min(sk); | |
671 | } | |
672 | } else { | |
673 | /* no previous measure. */ | |
674 | tp->srtt = m << 3; /* take the measured time to be rtt */ | |
675 | tp->mdev = m << 1; /* make sure rto = 3*rtt */ | |
676 | tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); | |
677 | tp->rtt_seq = tp->snd_nxt; | |
678 | } | |
679 | } | |
680 | ||
681 | /* Calculate rto without backoff. This is the second half of Van Jacobson's | |
682 | * routine referred to above. | |
683 | */ | |
684 | static inline void tcp_set_rto(struct sock *sk) | |
685 | { | |
686 | const struct tcp_sock *tp = tcp_sk(sk); | |
687 | /* Old crap is replaced with new one. 8) | |
688 | * | |
689 | * More seriously: | |
690 | * 1. If rtt variance happened to be less 50msec, it is hallucination. | |
691 | * It cannot be less due to utterly erratic ACK generation made | |
692 | * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ | |
693 | * to do with delayed acks, because at cwnd>2 true delack timeout | |
694 | * is invisible. Actually, Linux-2.4 also generates erratic | |
695 | * ACKs in some circumstances. | |
696 | */ | |
697 | inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar; | |
698 | ||
699 | /* 2. Fixups made earlier cannot be right. | |
700 | * If we do not estimate RTO correctly without them, | |
701 | * all the algo is pure shit and should be replaced | |
702 | * with correct one. It is exactly, which we pretend to do. | |
703 | */ | |
704 | } | |
705 | ||
706 | /* NOTE: clamping at TCP_RTO_MIN is not required, current algo | |
707 | * guarantees that rto is higher. | |
708 | */ | |
709 | static inline void tcp_bound_rto(struct sock *sk) | |
710 | { | |
711 | if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX) | |
712 | inet_csk(sk)->icsk_rto = TCP_RTO_MAX; | |
713 | } | |
714 | ||
715 | /* Save metrics learned by this TCP session. | |
716 | This function is called only, when TCP finishes successfully | |
717 | i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. | |
718 | */ | |
719 | void tcp_update_metrics(struct sock *sk) | |
720 | { | |
721 | struct tcp_sock *tp = tcp_sk(sk); | |
722 | struct dst_entry *dst = __sk_dst_get(sk); | |
723 | ||
724 | if (sysctl_tcp_nometrics_save) | |
725 | return; | |
726 | ||
727 | dst_confirm(dst); | |
728 | ||
729 | if (dst && (dst->flags & DST_HOST)) { | |
730 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
731 | int m; | |
732 | unsigned long rtt; | |
733 | ||
734 | if (icsk->icsk_backoff || !tp->srtt) { | |
735 | /* This session failed to estimate rtt. Why? | |
736 | * Probably, no packets returned in time. | |
737 | * Reset our results. | |
738 | */ | |
739 | if (!(dst_metric_locked(dst, RTAX_RTT))) | |
740 | dst->metrics[RTAX_RTT - 1] = 0; | |
741 | return; | |
742 | } | |
743 | ||
744 | rtt = dst_metric_rtt(dst, RTAX_RTT); | |
745 | m = rtt - tp->srtt; | |
746 | ||
747 | /* If newly calculated rtt larger than stored one, | |
748 | * store new one. Otherwise, use EWMA. Remember, | |
749 | * rtt overestimation is always better than underestimation. | |
750 | */ | |
751 | if (!(dst_metric_locked(dst, RTAX_RTT))) { | |
752 | if (m <= 0) | |
753 | set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt); | |
754 | else | |
755 | set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3)); | |
756 | } | |
757 | ||
758 | if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { | |
759 | unsigned long var; | |
760 | if (m < 0) | |
761 | m = -m; | |
762 | ||
763 | /* Scale deviation to rttvar fixed point */ | |
764 | m >>= 1; | |
765 | if (m < tp->mdev) | |
766 | m = tp->mdev; | |
767 | ||
768 | var = dst_metric_rtt(dst, RTAX_RTTVAR); | |
769 | if (m >= var) | |
770 | var = m; | |
771 | else | |
772 | var -= (var - m) >> 2; | |
773 | ||
774 | set_dst_metric_rtt(dst, RTAX_RTTVAR, var); | |
775 | } | |
776 | ||
777 | if (tp->snd_ssthresh >= 0xFFFF) { | |
778 | /* Slow start still did not finish. */ | |
779 | if (dst_metric(dst, RTAX_SSTHRESH) && | |
780 | !dst_metric_locked(dst, RTAX_SSTHRESH) && | |
781 | (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) | |
782 | dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1; | |
783 | if (!dst_metric_locked(dst, RTAX_CWND) && | |
784 | tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) | |
785 | dst->metrics[RTAX_CWND - 1] = tp->snd_cwnd; | |
786 | } else if (tp->snd_cwnd > tp->snd_ssthresh && | |
787 | icsk->icsk_ca_state == TCP_CA_Open) { | |
788 | /* Cong. avoidance phase, cwnd is reliable. */ | |
789 | if (!dst_metric_locked(dst, RTAX_SSTHRESH)) | |
790 | dst->metrics[RTAX_SSTHRESH-1] = | |
791 | max(tp->snd_cwnd >> 1, tp->snd_ssthresh); | |
792 | if (!dst_metric_locked(dst, RTAX_CWND)) | |
793 | dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_cwnd) >> 1; | |
794 | } else { | |
795 | /* Else slow start did not finish, cwnd is non-sense, | |
796 | ssthresh may be also invalid. | |
797 | */ | |
798 | if (!dst_metric_locked(dst, RTAX_CWND)) | |
799 | dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_ssthresh) >> 1; | |
800 | if (dst_metric(dst, RTAX_SSTHRESH) && | |
801 | !dst_metric_locked(dst, RTAX_SSTHRESH) && | |
802 | tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH)) | |
803 | dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh; | |
804 | } | |
805 | ||
806 | if (!dst_metric_locked(dst, RTAX_REORDERING)) { | |
807 | if (dst_metric(dst, RTAX_REORDERING) < tp->reordering && | |
808 | tp->reordering != sysctl_tcp_reordering) | |
809 | dst->metrics[RTAX_REORDERING-1] = tp->reordering; | |
810 | } | |
811 | } | |
812 | } | |
813 | ||
814 | /* Numbers are taken from RFC3390. | |
815 | * | |
816 | * John Heffner states: | |
817 | * | |
818 | * The RFC specifies a window of no more than 4380 bytes | |
819 | * unless 2*MSS > 4380. Reading the pseudocode in the RFC | |
820 | * is a bit misleading because they use a clamp at 4380 bytes | |
821 | * rather than use a multiplier in the relevant range. | |
822 | */ | |
823 | __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst) | |
824 | { | |
825 | __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); | |
826 | ||
827 | if (!cwnd) { | |
828 | if (tp->mss_cache > 1460) | |
829 | cwnd = 2; | |
830 | else | |
831 | cwnd = (tp->mss_cache > 1095) ? 3 : 4; | |
832 | } | |
833 | return min_t(__u32, cwnd, tp->snd_cwnd_clamp); | |
834 | } | |
835 | ||
836 | /* Set slow start threshold and cwnd not falling to slow start */ | |
837 | void tcp_enter_cwr(struct sock *sk, const int set_ssthresh) | |
838 | { | |
839 | struct tcp_sock *tp = tcp_sk(sk); | |
840 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
841 | ||
842 | tp->prior_ssthresh = 0; | |
843 | tp->bytes_acked = 0; | |
844 | if (icsk->icsk_ca_state < TCP_CA_CWR) { | |
845 | tp->undo_marker = 0; | |
846 | if (set_ssthresh) | |
847 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
848 | tp->snd_cwnd = min(tp->snd_cwnd, | |
849 | tcp_packets_in_flight(tp) + 1U); | |
850 | tp->snd_cwnd_cnt = 0; | |
851 | tp->high_seq = tp->snd_nxt; | |
852 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
853 | TCP_ECN_queue_cwr(tp); | |
854 | ||
855 | tcp_set_ca_state(sk, TCP_CA_CWR); | |
856 | } | |
857 | } | |
858 | ||
859 | /* | |
860 | * Packet counting of FACK is based on in-order assumptions, therefore TCP | |
861 | * disables it when reordering is detected | |
862 | */ | |
863 | static void tcp_disable_fack(struct tcp_sock *tp) | |
864 | { | |
865 | /* RFC3517 uses different metric in lost marker => reset on change */ | |
866 | if (tcp_is_fack(tp)) | |
867 | tp->lost_skb_hint = NULL; | |
868 | tp->rx_opt.sack_ok &= ~2; | |
869 | } | |
870 | ||
871 | /* Take a notice that peer is sending D-SACKs */ | |
872 | static void tcp_dsack_seen(struct tcp_sock *tp) | |
873 | { | |
874 | tp->rx_opt.sack_ok |= 4; | |
875 | } | |
876 | ||
877 | /* Initialize metrics on socket. */ | |
878 | ||
879 | static void tcp_init_metrics(struct sock *sk) | |
880 | { | |
881 | struct tcp_sock *tp = tcp_sk(sk); | |
882 | struct dst_entry *dst = __sk_dst_get(sk); | |
883 | ||
884 | if (dst == NULL) | |
885 | goto reset; | |
886 | ||
887 | dst_confirm(dst); | |
888 | ||
889 | if (dst_metric_locked(dst, RTAX_CWND)) | |
890 | tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); | |
891 | if (dst_metric(dst, RTAX_SSTHRESH)) { | |
892 | tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); | |
893 | if (tp->snd_ssthresh > tp->snd_cwnd_clamp) | |
894 | tp->snd_ssthresh = tp->snd_cwnd_clamp; | |
895 | } | |
896 | if (dst_metric(dst, RTAX_REORDERING) && | |
897 | tp->reordering != dst_metric(dst, RTAX_REORDERING)) { | |
898 | tcp_disable_fack(tp); | |
899 | tp->reordering = dst_metric(dst, RTAX_REORDERING); | |
900 | } | |
901 | ||
902 | if (dst_metric(dst, RTAX_RTT) == 0) | |
903 | goto reset; | |
904 | ||
905 | if (!tp->srtt && dst_metric_rtt(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3)) | |
906 | goto reset; | |
907 | ||
908 | /* Initial rtt is determined from SYN,SYN-ACK. | |
909 | * The segment is small and rtt may appear much | |
910 | * less than real one. Use per-dst memory | |
911 | * to make it more realistic. | |
912 | * | |
913 | * A bit of theory. RTT is time passed after "normal" sized packet | |
914 | * is sent until it is ACKed. In normal circumstances sending small | |
915 | * packets force peer to delay ACKs and calculation is correct too. | |
916 | * The algorithm is adaptive and, provided we follow specs, it | |
917 | * NEVER underestimate RTT. BUT! If peer tries to make some clever | |
918 | * tricks sort of "quick acks" for time long enough to decrease RTT | |
919 | * to low value, and then abruptly stops to do it and starts to delay | |
920 | * ACKs, wait for troubles. | |
921 | */ | |
922 | if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) { | |
923 | tp->srtt = dst_metric_rtt(dst, RTAX_RTT); | |
924 | tp->rtt_seq = tp->snd_nxt; | |
925 | } | |
926 | if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) { | |
927 | tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR); | |
928 | tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); | |
929 | } | |
930 | tcp_set_rto(sk); | |
931 | tcp_bound_rto(sk); | |
932 | if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp) | |
933 | goto reset; | |
934 | tp->snd_cwnd = tcp_init_cwnd(tp, dst); | |
935 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
936 | return; | |
937 | ||
938 | reset: | |
939 | /* Play conservative. If timestamps are not | |
940 | * supported, TCP will fail to recalculate correct | |
941 | * rtt, if initial rto is too small. FORGET ALL AND RESET! | |
942 | */ | |
943 | if (!tp->rx_opt.saw_tstamp && tp->srtt) { | |
944 | tp->srtt = 0; | |
945 | tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT; | |
946 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; | |
947 | } | |
948 | } | |
949 | ||
950 | static void tcp_update_reordering(struct sock *sk, const int metric, | |
951 | const int ts) | |
952 | { | |
953 | struct tcp_sock *tp = tcp_sk(sk); | |
954 | if (metric > tp->reordering) { | |
955 | int mib_idx; | |
956 | ||
957 | tp->reordering = min(TCP_MAX_REORDERING, metric); | |
958 | ||
959 | /* This exciting event is worth to be remembered. 8) */ | |
960 | if (ts) | |
961 | mib_idx = LINUX_MIB_TCPTSREORDER; | |
962 | else if (tcp_is_reno(tp)) | |
963 | mib_idx = LINUX_MIB_TCPRENOREORDER; | |
964 | else if (tcp_is_fack(tp)) | |
965 | mib_idx = LINUX_MIB_TCPFACKREORDER; | |
966 | else | |
967 | mib_idx = LINUX_MIB_TCPSACKREORDER; | |
968 | ||
969 | NET_INC_STATS_BH(sock_net(sk), mib_idx); | |
970 | #if FASTRETRANS_DEBUG > 1 | |
971 | printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", | |
972 | tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, | |
973 | tp->reordering, | |
974 | tp->fackets_out, | |
975 | tp->sacked_out, | |
976 | tp->undo_marker ? tp->undo_retrans : 0); | |
977 | #endif | |
978 | tcp_disable_fack(tp); | |
979 | } | |
980 | } | |
981 | ||
982 | /* This must be called before lost_out is incremented */ | |
983 | static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) | |
984 | { | |
985 | if ((tp->retransmit_skb_hint == NULL) || | |
986 | before(TCP_SKB_CB(skb)->seq, | |
987 | TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) | |
988 | tp->retransmit_skb_hint = skb; | |
989 | ||
990 | if (!tp->lost_out || | |
991 | after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high)) | |
992 | tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; | |
993 | } | |
994 | ||
995 | static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb) | |
996 | { | |
997 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { | |
998 | tcp_verify_retransmit_hint(tp, skb); | |
999 | ||
1000 | tp->lost_out += tcp_skb_pcount(skb); | |
1001 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1002 | } | |
1003 | } | |
1004 | ||
1005 | static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, | |
1006 | struct sk_buff *skb) | |
1007 | { | |
1008 | tcp_verify_retransmit_hint(tp, skb); | |
1009 | ||
1010 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { | |
1011 | tp->lost_out += tcp_skb_pcount(skb); | |
1012 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1013 | } | |
1014 | } | |
1015 | ||
1016 | /* This procedure tags the retransmission queue when SACKs arrive. | |
1017 | * | |
1018 | * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). | |
1019 | * Packets in queue with these bits set are counted in variables | |
1020 | * sacked_out, retrans_out and lost_out, correspondingly. | |
1021 | * | |
1022 | * Valid combinations are: | |
1023 | * Tag InFlight Description | |
1024 | * 0 1 - orig segment is in flight. | |
1025 | * S 0 - nothing flies, orig reached receiver. | |
1026 | * L 0 - nothing flies, orig lost by net. | |
1027 | * R 2 - both orig and retransmit are in flight. | |
1028 | * L|R 1 - orig is lost, retransmit is in flight. | |
1029 | * S|R 1 - orig reached receiver, retrans is still in flight. | |
1030 | * (L|S|R is logically valid, it could occur when L|R is sacked, | |
1031 | * but it is equivalent to plain S and code short-curcuits it to S. | |
1032 | * L|S is logically invalid, it would mean -1 packet in flight 8)) | |
1033 | * | |
1034 | * These 6 states form finite state machine, controlled by the following events: | |
1035 | * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) | |
1036 | * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) | |
1037 | * 3. Loss detection event of one of three flavors: | |
1038 | * A. Scoreboard estimator decided the packet is lost. | |
1039 | * A'. Reno "three dupacks" marks head of queue lost. | |
1040 | * A''. Its FACK modfication, head until snd.fack is lost. | |
1041 | * B. SACK arrives sacking data transmitted after never retransmitted | |
1042 | * hole was sent out. | |
1043 | * C. SACK arrives sacking SND.NXT at the moment, when the | |
1044 | * segment was retransmitted. | |
1045 | * 4. D-SACK added new rule: D-SACK changes any tag to S. | |
1046 | * | |
1047 | * It is pleasant to note, that state diagram turns out to be commutative, | |
1048 | * so that we are allowed not to be bothered by order of our actions, | |
1049 | * when multiple events arrive simultaneously. (see the function below). | |
1050 | * | |
1051 | * Reordering detection. | |
1052 | * -------------------- | |
1053 | * Reordering metric is maximal distance, which a packet can be displaced | |
1054 | * in packet stream. With SACKs we can estimate it: | |
1055 | * | |
1056 | * 1. SACK fills old hole and the corresponding segment was not | |
1057 | * ever retransmitted -> reordering. Alas, we cannot use it | |
1058 | * when segment was retransmitted. | |
1059 | * 2. The last flaw is solved with D-SACK. D-SACK arrives | |
1060 | * for retransmitted and already SACKed segment -> reordering.. | |
1061 | * Both of these heuristics are not used in Loss state, when we cannot | |
1062 | * account for retransmits accurately. | |
1063 | * | |
1064 | * SACK block validation. | |
1065 | * ---------------------- | |
1066 | * | |
1067 | * SACK block range validation checks that the received SACK block fits to | |
1068 | * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. | |
1069 | * Note that SND.UNA is not included to the range though being valid because | |
1070 | * it means that the receiver is rather inconsistent with itself reporting | |
1071 | * SACK reneging when it should advance SND.UNA. Such SACK block this is | |
1072 | * perfectly valid, however, in light of RFC2018 which explicitly states | |
1073 | * that "SACK block MUST reflect the newest segment. Even if the newest | |
1074 | * segment is going to be discarded ...", not that it looks very clever | |
1075 | * in case of head skb. Due to potentional receiver driven attacks, we | |
1076 | * choose to avoid immediate execution of a walk in write queue due to | |
1077 | * reneging and defer head skb's loss recovery to standard loss recovery | |
1078 | * procedure that will eventually trigger (nothing forbids us doing this). | |
1079 | * | |
1080 | * Implements also blockage to start_seq wrap-around. Problem lies in the | |
1081 | * fact that though start_seq (s) is before end_seq (i.e., not reversed), | |
1082 | * there's no guarantee that it will be before snd_nxt (n). The problem | |
1083 | * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt | |
1084 | * wrap (s_w): | |
1085 | * | |
1086 | * <- outs wnd -> <- wrapzone -> | |
1087 | * u e n u_w e_w s n_w | |
1088 | * | | | | | | | | |
1089 | * |<------------+------+----- TCP seqno space --------------+---------->| | |
1090 | * ...-- <2^31 ->| |<--------... | |
1091 | * ...---- >2^31 ------>| |<--------... | |
1092 | * | |
1093 | * Current code wouldn't be vulnerable but it's better still to discard such | |
1094 | * crazy SACK blocks. Doing this check for start_seq alone closes somewhat | |
1095 | * similar case (end_seq after snd_nxt wrap) as earlier reversed check in | |
1096 | * snd_nxt wrap -> snd_una region will then become "well defined", i.e., | |
1097 | * equal to the ideal case (infinite seqno space without wrap caused issues). | |
1098 | * | |
1099 | * With D-SACK the lower bound is extended to cover sequence space below | |
1100 | * SND.UNA down to undo_marker, which is the last point of interest. Yet | |
1101 | * again, D-SACK block must not to go across snd_una (for the same reason as | |
1102 | * for the normal SACK blocks, explained above). But there all simplicity | |
1103 | * ends, TCP might receive valid D-SACKs below that. As long as they reside | |
1104 | * fully below undo_marker they do not affect behavior in anyway and can | |
1105 | * therefore be safely ignored. In rare cases (which are more or less | |
1106 | * theoretical ones), the D-SACK will nicely cross that boundary due to skb | |
1107 | * fragmentation and packet reordering past skb's retransmission. To consider | |
1108 | * them correctly, the acceptable range must be extended even more though | |
1109 | * the exact amount is rather hard to quantify. However, tp->max_window can | |
1110 | * be used as an exaggerated estimate. | |
1111 | */ | |
1112 | static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack, | |
1113 | u32 start_seq, u32 end_seq) | |
1114 | { | |
1115 | /* Too far in future, or reversed (interpretation is ambiguous) */ | |
1116 | if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) | |
1117 | return 0; | |
1118 | ||
1119 | /* Nasty start_seq wrap-around check (see comments above) */ | |
1120 | if (!before(start_seq, tp->snd_nxt)) | |
1121 | return 0; | |
1122 | ||
1123 | /* In outstanding window? ...This is valid exit for D-SACKs too. | |
1124 | * start_seq == snd_una is non-sensical (see comments above) | |
1125 | */ | |
1126 | if (after(start_seq, tp->snd_una)) | |
1127 | return 1; | |
1128 | ||
1129 | if (!is_dsack || !tp->undo_marker) | |
1130 | return 0; | |
1131 | ||
1132 | /* ...Then it's D-SACK, and must reside below snd_una completely */ | |
1133 | if (!after(end_seq, tp->snd_una)) | |
1134 | return 0; | |
1135 | ||
1136 | if (!before(start_seq, tp->undo_marker)) | |
1137 | return 1; | |
1138 | ||
1139 | /* Too old */ | |
1140 | if (!after(end_seq, tp->undo_marker)) | |
1141 | return 0; | |
1142 | ||
1143 | /* Undo_marker boundary crossing (overestimates a lot). Known already: | |
1144 | * start_seq < undo_marker and end_seq >= undo_marker. | |
1145 | */ | |
1146 | return !before(start_seq, end_seq - tp->max_window); | |
1147 | } | |
1148 | ||
1149 | /* Check for lost retransmit. This superb idea is borrowed from "ratehalving". | |
1150 | * Event "C". Later note: FACK people cheated me again 8), we have to account | |
1151 | * for reordering! Ugly, but should help. | |
1152 | * | |
1153 | * Search retransmitted skbs from write_queue that were sent when snd_nxt was | |
1154 | * less than what is now known to be received by the other end (derived from | |
1155 | * highest SACK block). Also calculate the lowest snd_nxt among the remaining | |
1156 | * retransmitted skbs to avoid some costly processing per ACKs. | |
1157 | */ | |
1158 | static void tcp_mark_lost_retrans(struct sock *sk) | |
1159 | { | |
1160 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
1161 | struct tcp_sock *tp = tcp_sk(sk); | |
1162 | struct sk_buff *skb; | |
1163 | int cnt = 0; | |
1164 | u32 new_low_seq = tp->snd_nxt; | |
1165 | u32 received_upto = tcp_highest_sack_seq(tp); | |
1166 | ||
1167 | if (!tcp_is_fack(tp) || !tp->retrans_out || | |
1168 | !after(received_upto, tp->lost_retrans_low) || | |
1169 | icsk->icsk_ca_state != TCP_CA_Recovery) | |
1170 | return; | |
1171 | ||
1172 | tcp_for_write_queue(skb, sk) { | |
1173 | u32 ack_seq = TCP_SKB_CB(skb)->ack_seq; | |
1174 | ||
1175 | if (skb == tcp_send_head(sk)) | |
1176 | break; | |
1177 | if (cnt == tp->retrans_out) | |
1178 | break; | |
1179 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1180 | continue; | |
1181 | ||
1182 | if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) | |
1183 | continue; | |
1184 | ||
1185 | if (after(received_upto, ack_seq) && | |
1186 | (tcp_is_fack(tp) || | |
1187 | !before(received_upto, | |
1188 | ack_seq + tp->reordering * tp->mss_cache))) { | |
1189 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; | |
1190 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1191 | ||
1192 | tcp_skb_mark_lost_uncond_verify(tp, skb); | |
1193 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT); | |
1194 | } else { | |
1195 | if (before(ack_seq, new_low_seq)) | |
1196 | new_low_seq = ack_seq; | |
1197 | cnt += tcp_skb_pcount(skb); | |
1198 | } | |
1199 | } | |
1200 | ||
1201 | if (tp->retrans_out) | |
1202 | tp->lost_retrans_low = new_low_seq; | |
1203 | } | |
1204 | ||
1205 | static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb, | |
1206 | struct tcp_sack_block_wire *sp, int num_sacks, | |
1207 | u32 prior_snd_una) | |
1208 | { | |
1209 | struct tcp_sock *tp = tcp_sk(sk); | |
1210 | u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); | |
1211 | u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); | |
1212 | int dup_sack = 0; | |
1213 | ||
1214 | if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { | |
1215 | dup_sack = 1; | |
1216 | tcp_dsack_seen(tp); | |
1217 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV); | |
1218 | } else if (num_sacks > 1) { | |
1219 | u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); | |
1220 | u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); | |
1221 | ||
1222 | if (!after(end_seq_0, end_seq_1) && | |
1223 | !before(start_seq_0, start_seq_1)) { | |
1224 | dup_sack = 1; | |
1225 | tcp_dsack_seen(tp); | |
1226 | NET_INC_STATS_BH(sock_net(sk), | |
1227 | LINUX_MIB_TCPDSACKOFORECV); | |
1228 | } | |
1229 | } | |
1230 | ||
1231 | /* D-SACK for already forgotten data... Do dumb counting. */ | |
1232 | if (dup_sack && | |
1233 | !after(end_seq_0, prior_snd_una) && | |
1234 | after(end_seq_0, tp->undo_marker)) | |
1235 | tp->undo_retrans--; | |
1236 | ||
1237 | return dup_sack; | |
1238 | } | |
1239 | ||
1240 | /* Check if skb is fully within the SACK block. In presence of GSO skbs, | |
1241 | * the incoming SACK may not exactly match but we can find smaller MSS | |
1242 | * aligned portion of it that matches. Therefore we might need to fragment | |
1243 | * which may fail and creates some hassle (caller must handle error case | |
1244 | * returns). | |
1245 | */ | |
1246 | static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, | |
1247 | u32 start_seq, u32 end_seq) | |
1248 | { | |
1249 | int in_sack, err; | |
1250 | unsigned int pkt_len; | |
1251 | unsigned int mss; | |
1252 | ||
1253 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && | |
1254 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1255 | ||
1256 | if (tcp_skb_pcount(skb) > 1 && !in_sack && | |
1257 | after(TCP_SKB_CB(skb)->end_seq, start_seq)) { | |
1258 | mss = tcp_skb_mss(skb); | |
1259 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); | |
1260 | ||
1261 | if (!in_sack) { | |
1262 | pkt_len = start_seq - TCP_SKB_CB(skb)->seq; | |
1263 | if (pkt_len < mss) | |
1264 | pkt_len = mss; | |
1265 | } else { | |
1266 | pkt_len = end_seq - TCP_SKB_CB(skb)->seq; | |
1267 | if (pkt_len < mss) | |
1268 | return -EINVAL; | |
1269 | } | |
1270 | ||
1271 | /* Round if necessary so that SACKs cover only full MSSes | |
1272 | * and/or the remaining small portion (if present) | |
1273 | */ | |
1274 | if (pkt_len > mss) { | |
1275 | unsigned int new_len = (pkt_len / mss) * mss; | |
1276 | if (!in_sack && new_len < pkt_len) { | |
1277 | new_len += mss; | |
1278 | if (new_len > skb->len) | |
1279 | return 0; | |
1280 | } | |
1281 | pkt_len = new_len; | |
1282 | } | |
1283 | err = tcp_fragment(sk, skb, pkt_len, mss); | |
1284 | if (err < 0) | |
1285 | return err; | |
1286 | } | |
1287 | ||
1288 | return in_sack; | |
1289 | } | |
1290 | ||
1291 | static int tcp_sacktag_one(struct sk_buff *skb, struct sock *sk, | |
1292 | int *reord, int dup_sack, int fack_count) | |
1293 | { | |
1294 | struct tcp_sock *tp = tcp_sk(sk); | |
1295 | u8 sacked = TCP_SKB_CB(skb)->sacked; | |
1296 | int flag = 0; | |
1297 | ||
1298 | /* Account D-SACK for retransmitted packet. */ | |
1299 | if (dup_sack && (sacked & TCPCB_RETRANS)) { | |
1300 | if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker)) | |
1301 | tp->undo_retrans--; | |
1302 | if (sacked & TCPCB_SACKED_ACKED) | |
1303 | *reord = min(fack_count, *reord); | |
1304 | } | |
1305 | ||
1306 | /* Nothing to do; acked frame is about to be dropped (was ACKed). */ | |
1307 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1308 | return flag; | |
1309 | ||
1310 | if (!(sacked & TCPCB_SACKED_ACKED)) { | |
1311 | if (sacked & TCPCB_SACKED_RETRANS) { | |
1312 | /* If the segment is not tagged as lost, | |
1313 | * we do not clear RETRANS, believing | |
1314 | * that retransmission is still in flight. | |
1315 | */ | |
1316 | if (sacked & TCPCB_LOST) { | |
1317 | TCP_SKB_CB(skb)->sacked &= | |
1318 | ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); | |
1319 | tp->lost_out -= tcp_skb_pcount(skb); | |
1320 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1321 | } | |
1322 | } else { | |
1323 | if (!(sacked & TCPCB_RETRANS)) { | |
1324 | /* New sack for not retransmitted frame, | |
1325 | * which was in hole. It is reordering. | |
1326 | */ | |
1327 | if (before(TCP_SKB_CB(skb)->seq, | |
1328 | tcp_highest_sack_seq(tp))) | |
1329 | *reord = min(fack_count, *reord); | |
1330 | ||
1331 | /* SACK enhanced F-RTO (RFC4138; Appendix B) */ | |
1332 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) | |
1333 | flag |= FLAG_ONLY_ORIG_SACKED; | |
1334 | } | |
1335 | ||
1336 | if (sacked & TCPCB_LOST) { | |
1337 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; | |
1338 | tp->lost_out -= tcp_skb_pcount(skb); | |
1339 | } | |
1340 | } | |
1341 | ||
1342 | TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED; | |
1343 | flag |= FLAG_DATA_SACKED; | |
1344 | tp->sacked_out += tcp_skb_pcount(skb); | |
1345 | ||
1346 | fack_count += tcp_skb_pcount(skb); | |
1347 | ||
1348 | /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ | |
1349 | if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) && | |
1350 | before(TCP_SKB_CB(skb)->seq, | |
1351 | TCP_SKB_CB(tp->lost_skb_hint)->seq)) | |
1352 | tp->lost_cnt_hint += tcp_skb_pcount(skb); | |
1353 | ||
1354 | if (fack_count > tp->fackets_out) | |
1355 | tp->fackets_out = fack_count; | |
1356 | ||
1357 | if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) | |
1358 | tcp_advance_highest_sack(sk, skb); | |
1359 | } | |
1360 | ||
1361 | /* D-SACK. We can detect redundant retransmission in S|R and plain R | |
1362 | * frames and clear it. undo_retrans is decreased above, L|R frames | |
1363 | * are accounted above as well. | |
1364 | */ | |
1365 | if (dup_sack && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) { | |
1366 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; | |
1367 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1368 | } | |
1369 | ||
1370 | return flag; | |
1371 | } | |
1372 | ||
1373 | static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, | |
1374 | struct tcp_sack_block *next_dup, | |
1375 | u32 start_seq, u32 end_seq, | |
1376 | int dup_sack_in, int *fack_count, | |
1377 | int *reord, int *flag) | |
1378 | { | |
1379 | tcp_for_write_queue_from(skb, sk) { | |
1380 | int in_sack = 0; | |
1381 | int dup_sack = dup_sack_in; | |
1382 | ||
1383 | if (skb == tcp_send_head(sk)) | |
1384 | break; | |
1385 | ||
1386 | /* queue is in-order => we can short-circuit the walk early */ | |
1387 | if (!before(TCP_SKB_CB(skb)->seq, end_seq)) | |
1388 | break; | |
1389 | ||
1390 | if ((next_dup != NULL) && | |
1391 | before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { | |
1392 | in_sack = tcp_match_skb_to_sack(sk, skb, | |
1393 | next_dup->start_seq, | |
1394 | next_dup->end_seq); | |
1395 | if (in_sack > 0) | |
1396 | dup_sack = 1; | |
1397 | } | |
1398 | ||
1399 | if (in_sack <= 0) | |
1400 | in_sack = tcp_match_skb_to_sack(sk, skb, start_seq, | |
1401 | end_seq); | |
1402 | if (unlikely(in_sack < 0)) | |
1403 | break; | |
1404 | ||
1405 | if (in_sack) | |
1406 | *flag |= tcp_sacktag_one(skb, sk, reord, dup_sack, | |
1407 | *fack_count); | |
1408 | ||
1409 | *fack_count += tcp_skb_pcount(skb); | |
1410 | } | |
1411 | return skb; | |
1412 | } | |
1413 | ||
1414 | /* Avoid all extra work that is being done by sacktag while walking in | |
1415 | * a normal way | |
1416 | */ | |
1417 | static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, | |
1418 | u32 skip_to_seq, int *fack_count) | |
1419 | { | |
1420 | tcp_for_write_queue_from(skb, sk) { | |
1421 | if (skb == tcp_send_head(sk)) | |
1422 | break; | |
1423 | ||
1424 | if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq)) | |
1425 | break; | |
1426 | ||
1427 | *fack_count += tcp_skb_pcount(skb); | |
1428 | } | |
1429 | return skb; | |
1430 | } | |
1431 | ||
1432 | static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, | |
1433 | struct sock *sk, | |
1434 | struct tcp_sack_block *next_dup, | |
1435 | u32 skip_to_seq, | |
1436 | int *fack_count, int *reord, | |
1437 | int *flag) | |
1438 | { | |
1439 | if (next_dup == NULL) | |
1440 | return skb; | |
1441 | ||
1442 | if (before(next_dup->start_seq, skip_to_seq)) { | |
1443 | skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq, fack_count); | |
1444 | skb = tcp_sacktag_walk(skb, sk, NULL, | |
1445 | next_dup->start_seq, next_dup->end_seq, | |
1446 | 1, fack_count, reord, flag); | |
1447 | } | |
1448 | ||
1449 | return skb; | |
1450 | } | |
1451 | ||
1452 | static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache) | |
1453 | { | |
1454 | return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); | |
1455 | } | |
1456 | ||
1457 | static int | |
1458 | tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, | |
1459 | u32 prior_snd_una) | |
1460 | { | |
1461 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
1462 | struct tcp_sock *tp = tcp_sk(sk); | |
1463 | unsigned char *ptr = (skb_transport_header(ack_skb) + | |
1464 | TCP_SKB_CB(ack_skb)->sacked); | |
1465 | struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); | |
1466 | struct tcp_sack_block sp[TCP_NUM_SACKS]; | |
1467 | struct tcp_sack_block *cache; | |
1468 | struct sk_buff *skb; | |
1469 | int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); | |
1470 | int used_sacks; | |
1471 | int reord = tp->packets_out; | |
1472 | int flag = 0; | |
1473 | int found_dup_sack = 0; | |
1474 | int fack_count; | |
1475 | int i, j; | |
1476 | int first_sack_index; | |
1477 | ||
1478 | if (!tp->sacked_out) { | |
1479 | if (WARN_ON(tp->fackets_out)) | |
1480 | tp->fackets_out = 0; | |
1481 | tcp_highest_sack_reset(sk); | |
1482 | } | |
1483 | ||
1484 | found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, | |
1485 | num_sacks, prior_snd_una); | |
1486 | if (found_dup_sack) | |
1487 | flag |= FLAG_DSACKING_ACK; | |
1488 | ||
1489 | /* Eliminate too old ACKs, but take into | |
1490 | * account more or less fresh ones, they can | |
1491 | * contain valid SACK info. | |
1492 | */ | |
1493 | if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) | |
1494 | return 0; | |
1495 | ||
1496 | if (!tp->packets_out) | |
1497 | goto out; | |
1498 | ||
1499 | used_sacks = 0; | |
1500 | first_sack_index = 0; | |
1501 | for (i = 0; i < num_sacks; i++) { | |
1502 | int dup_sack = !i && found_dup_sack; | |
1503 | ||
1504 | sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq); | |
1505 | sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq); | |
1506 | ||
1507 | if (!tcp_is_sackblock_valid(tp, dup_sack, | |
1508 | sp[used_sacks].start_seq, | |
1509 | sp[used_sacks].end_seq)) { | |
1510 | int mib_idx; | |
1511 | ||
1512 | if (dup_sack) { | |
1513 | if (!tp->undo_marker) | |
1514 | mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO; | |
1515 | else | |
1516 | mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD; | |
1517 | } else { | |
1518 | /* Don't count olds caused by ACK reordering */ | |
1519 | if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && | |
1520 | !after(sp[used_sacks].end_seq, tp->snd_una)) | |
1521 | continue; | |
1522 | mib_idx = LINUX_MIB_TCPSACKDISCARD; | |
1523 | } | |
1524 | ||
1525 | NET_INC_STATS_BH(sock_net(sk), mib_idx); | |
1526 | if (i == 0) | |
1527 | first_sack_index = -1; | |
1528 | continue; | |
1529 | } | |
1530 | ||
1531 | /* Ignore very old stuff early */ | |
1532 | if (!after(sp[used_sacks].end_seq, prior_snd_una)) | |
1533 | continue; | |
1534 | ||
1535 | used_sacks++; | |
1536 | } | |
1537 | ||
1538 | /* order SACK blocks to allow in order walk of the retrans queue */ | |
1539 | for (i = used_sacks - 1; i > 0; i--) { | |
1540 | for (j = 0; j < i; j++) { | |
1541 | if (after(sp[j].start_seq, sp[j + 1].start_seq)) { | |
1542 | struct tcp_sack_block tmp; | |
1543 | ||
1544 | tmp = sp[j]; | |
1545 | sp[j] = sp[j + 1]; | |
1546 | sp[j + 1] = tmp; | |
1547 | ||
1548 | /* Track where the first SACK block goes to */ | |
1549 | if (j == first_sack_index) | |
1550 | first_sack_index = j + 1; | |
1551 | } | |
1552 | } | |
1553 | } | |
1554 | ||
1555 | skb = tcp_write_queue_head(sk); | |
1556 | fack_count = 0; | |
1557 | i = 0; | |
1558 | ||
1559 | if (!tp->sacked_out) { | |
1560 | /* It's already past, so skip checking against it */ | |
1561 | cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); | |
1562 | } else { | |
1563 | cache = tp->recv_sack_cache; | |
1564 | /* Skip empty blocks in at head of the cache */ | |
1565 | while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && | |
1566 | !cache->end_seq) | |
1567 | cache++; | |
1568 | } | |
1569 | ||
1570 | while (i < used_sacks) { | |
1571 | u32 start_seq = sp[i].start_seq; | |
1572 | u32 end_seq = sp[i].end_seq; | |
1573 | int dup_sack = (found_dup_sack && (i == first_sack_index)); | |
1574 | struct tcp_sack_block *next_dup = NULL; | |
1575 | ||
1576 | if (found_dup_sack && ((i + 1) == first_sack_index)) | |
1577 | next_dup = &sp[i + 1]; | |
1578 | ||
1579 | /* Event "B" in the comment above. */ | |
1580 | if (after(end_seq, tp->high_seq)) | |
1581 | flag |= FLAG_DATA_LOST; | |
1582 | ||
1583 | /* Skip too early cached blocks */ | |
1584 | while (tcp_sack_cache_ok(tp, cache) && | |
1585 | !before(start_seq, cache->end_seq)) | |
1586 | cache++; | |
1587 | ||
1588 | /* Can skip some work by looking recv_sack_cache? */ | |
1589 | if (tcp_sack_cache_ok(tp, cache) && !dup_sack && | |
1590 | after(end_seq, cache->start_seq)) { | |
1591 | ||
1592 | /* Head todo? */ | |
1593 | if (before(start_seq, cache->start_seq)) { | |
1594 | skb = tcp_sacktag_skip(skb, sk, start_seq, | |
1595 | &fack_count); | |
1596 | skb = tcp_sacktag_walk(skb, sk, next_dup, | |
1597 | start_seq, | |
1598 | cache->start_seq, | |
1599 | dup_sack, &fack_count, | |
1600 | &reord, &flag); | |
1601 | } | |
1602 | ||
1603 | /* Rest of the block already fully processed? */ | |
1604 | if (!after(end_seq, cache->end_seq)) | |
1605 | goto advance_sp; | |
1606 | ||
1607 | skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, | |
1608 | cache->end_seq, | |
1609 | &fack_count, &reord, | |
1610 | &flag); | |
1611 | ||
1612 | /* ...tail remains todo... */ | |
1613 | if (tcp_highest_sack_seq(tp) == cache->end_seq) { | |
1614 | /* ...but better entrypoint exists! */ | |
1615 | skb = tcp_highest_sack(sk); | |
1616 | if (skb == NULL) | |
1617 | break; | |
1618 | fack_count = tp->fackets_out; | |
1619 | cache++; | |
1620 | goto walk; | |
1621 | } | |
1622 | ||
1623 | skb = tcp_sacktag_skip(skb, sk, cache->end_seq, | |
1624 | &fack_count); | |
1625 | /* Check overlap against next cached too (past this one already) */ | |
1626 | cache++; | |
1627 | continue; | |
1628 | } | |
1629 | ||
1630 | if (!before(start_seq, tcp_highest_sack_seq(tp))) { | |
1631 | skb = tcp_highest_sack(sk); | |
1632 | if (skb == NULL) | |
1633 | break; | |
1634 | fack_count = tp->fackets_out; | |
1635 | } | |
1636 | skb = tcp_sacktag_skip(skb, sk, start_seq, &fack_count); | |
1637 | ||
1638 | walk: | |
1639 | skb = tcp_sacktag_walk(skb, sk, next_dup, start_seq, end_seq, | |
1640 | dup_sack, &fack_count, &reord, &flag); | |
1641 | ||
1642 | advance_sp: | |
1643 | /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct | |
1644 | * due to in-order walk | |
1645 | */ | |
1646 | if (after(end_seq, tp->frto_highmark)) | |
1647 | flag &= ~FLAG_ONLY_ORIG_SACKED; | |
1648 | ||
1649 | i++; | |
1650 | } | |
1651 | ||
1652 | /* Clear the head of the cache sack blocks so we can skip it next time */ | |
1653 | for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { | |
1654 | tp->recv_sack_cache[i].start_seq = 0; | |
1655 | tp->recv_sack_cache[i].end_seq = 0; | |
1656 | } | |
1657 | for (j = 0; j < used_sacks; j++) | |
1658 | tp->recv_sack_cache[i++] = sp[j]; | |
1659 | ||
1660 | tcp_mark_lost_retrans(sk); | |
1661 | ||
1662 | tcp_verify_left_out(tp); | |
1663 | ||
1664 | if ((reord < tp->fackets_out) && | |
1665 | ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) && | |
1666 | (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark))) | |
1667 | tcp_update_reordering(sk, tp->fackets_out - reord, 0); | |
1668 | ||
1669 | out: | |
1670 | ||
1671 | #if FASTRETRANS_DEBUG > 0 | |
1672 | WARN_ON((int)tp->sacked_out < 0); | |
1673 | WARN_ON((int)tp->lost_out < 0); | |
1674 | WARN_ON((int)tp->retrans_out < 0); | |
1675 | WARN_ON((int)tcp_packets_in_flight(tp) < 0); | |
1676 | #endif | |
1677 | return flag; | |
1678 | } | |
1679 | ||
1680 | /* Limits sacked_out so that sum with lost_out isn't ever larger than | |
1681 | * packets_out. Returns zero if sacked_out adjustement wasn't necessary. | |
1682 | */ | |
1683 | int tcp_limit_reno_sacked(struct tcp_sock *tp) | |
1684 | { | |
1685 | u32 holes; | |
1686 | ||
1687 | holes = max(tp->lost_out, 1U); | |
1688 | holes = min(holes, tp->packets_out); | |
1689 | ||
1690 | if ((tp->sacked_out + holes) > tp->packets_out) { | |
1691 | tp->sacked_out = tp->packets_out - holes; | |
1692 | return 1; | |
1693 | } | |
1694 | return 0; | |
1695 | } | |
1696 | ||
1697 | /* If we receive more dupacks than we expected counting segments | |
1698 | * in assumption of absent reordering, interpret this as reordering. | |
1699 | * The only another reason could be bug in receiver TCP. | |
1700 | */ | |
1701 | static void tcp_check_reno_reordering(struct sock *sk, const int addend) | |
1702 | { | |
1703 | struct tcp_sock *tp = tcp_sk(sk); | |
1704 | if (tcp_limit_reno_sacked(tp)) | |
1705 | tcp_update_reordering(sk, tp->packets_out + addend, 0); | |
1706 | } | |
1707 | ||
1708 | /* Emulate SACKs for SACKless connection: account for a new dupack. */ | |
1709 | ||
1710 | static void tcp_add_reno_sack(struct sock *sk) | |
1711 | { | |
1712 | struct tcp_sock *tp = tcp_sk(sk); | |
1713 | tp->sacked_out++; | |
1714 | tcp_check_reno_reordering(sk, 0); | |
1715 | tcp_verify_left_out(tp); | |
1716 | } | |
1717 | ||
1718 | /* Account for ACK, ACKing some data in Reno Recovery phase. */ | |
1719 | ||
1720 | static void tcp_remove_reno_sacks(struct sock *sk, int acked) | |
1721 | { | |
1722 | struct tcp_sock *tp = tcp_sk(sk); | |
1723 | ||
1724 | if (acked > 0) { | |
1725 | /* One ACK acked hole. The rest eat duplicate ACKs. */ | |
1726 | if (acked - 1 >= tp->sacked_out) | |
1727 | tp->sacked_out = 0; | |
1728 | else | |
1729 | tp->sacked_out -= acked - 1; | |
1730 | } | |
1731 | tcp_check_reno_reordering(sk, acked); | |
1732 | tcp_verify_left_out(tp); | |
1733 | } | |
1734 | ||
1735 | static inline void tcp_reset_reno_sack(struct tcp_sock *tp) | |
1736 | { | |
1737 | tp->sacked_out = 0; | |
1738 | } | |
1739 | ||
1740 | static int tcp_is_sackfrto(const struct tcp_sock *tp) | |
1741 | { | |
1742 | return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp); | |
1743 | } | |
1744 | ||
1745 | /* F-RTO can only be used if TCP has never retransmitted anything other than | |
1746 | * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here) | |
1747 | */ | |
1748 | int tcp_use_frto(struct sock *sk) | |
1749 | { | |
1750 | const struct tcp_sock *tp = tcp_sk(sk); | |
1751 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
1752 | struct sk_buff *skb; | |
1753 | ||
1754 | if (!sysctl_tcp_frto) | |
1755 | return 0; | |
1756 | ||
1757 | /* MTU probe and F-RTO won't really play nicely along currently */ | |
1758 | if (icsk->icsk_mtup.probe_size) | |
1759 | return 0; | |
1760 | ||
1761 | if (tcp_is_sackfrto(tp)) | |
1762 | return 1; | |
1763 | ||
1764 | /* Avoid expensive walking of rexmit queue if possible */ | |
1765 | if (tp->retrans_out > 1) | |
1766 | return 0; | |
1767 | ||
1768 | skb = tcp_write_queue_head(sk); | |
1769 | if (tcp_skb_is_last(sk, skb)) | |
1770 | return 1; | |
1771 | skb = tcp_write_queue_next(sk, skb); /* Skips head */ | |
1772 | tcp_for_write_queue_from(skb, sk) { | |
1773 | if (skb == tcp_send_head(sk)) | |
1774 | break; | |
1775 | if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) | |
1776 | return 0; | |
1777 | /* Short-circuit when first non-SACKed skb has been checked */ | |
1778 | if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) | |
1779 | break; | |
1780 | } | |
1781 | return 1; | |
1782 | } | |
1783 | ||
1784 | /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO | |
1785 | * recovery a bit and use heuristics in tcp_process_frto() to detect if | |
1786 | * the RTO was spurious. Only clear SACKED_RETRANS of the head here to | |
1787 | * keep retrans_out counting accurate (with SACK F-RTO, other than head | |
1788 | * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS | |
1789 | * bits are handled if the Loss state is really to be entered (in | |
1790 | * tcp_enter_frto_loss). | |
1791 | * | |
1792 | * Do like tcp_enter_loss() would; when RTO expires the second time it | |
1793 | * does: | |
1794 | * "Reduce ssthresh if it has not yet been made inside this window." | |
1795 | */ | |
1796 | void tcp_enter_frto(struct sock *sk) | |
1797 | { | |
1798 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
1799 | struct tcp_sock *tp = tcp_sk(sk); | |
1800 | struct sk_buff *skb; | |
1801 | ||
1802 | if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) || | |
1803 | tp->snd_una == tp->high_seq || | |
1804 | ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) && | |
1805 | !icsk->icsk_retransmits)) { | |
1806 | tp->prior_ssthresh = tcp_current_ssthresh(sk); | |
1807 | /* Our state is too optimistic in ssthresh() call because cwnd | |
1808 | * is not reduced until tcp_enter_frto_loss() when previous F-RTO | |
1809 | * recovery has not yet completed. Pattern would be this: RTO, | |
1810 | * Cumulative ACK, RTO (2xRTO for the same segment does not end | |
1811 | * up here twice). | |
1812 | * RFC4138 should be more specific on what to do, even though | |
1813 | * RTO is quite unlikely to occur after the first Cumulative ACK | |
1814 | * due to back-off and complexity of triggering events ... | |
1815 | */ | |
1816 | if (tp->frto_counter) { | |
1817 | u32 stored_cwnd; | |
1818 | stored_cwnd = tp->snd_cwnd; | |
1819 | tp->snd_cwnd = 2; | |
1820 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
1821 | tp->snd_cwnd = stored_cwnd; | |
1822 | } else { | |
1823 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
1824 | } | |
1825 | /* ... in theory, cong.control module could do "any tricks" in | |
1826 | * ssthresh(), which means that ca_state, lost bits and lost_out | |
1827 | * counter would have to be faked before the call occurs. We | |
1828 | * consider that too expensive, unlikely and hacky, so modules | |
1829 | * using these in ssthresh() must deal these incompatibility | |
1830 | * issues if they receives CA_EVENT_FRTO and frto_counter != 0 | |
1831 | */ | |
1832 | tcp_ca_event(sk, CA_EVENT_FRTO); | |
1833 | } | |
1834 | ||
1835 | tp->undo_marker = tp->snd_una; | |
1836 | tp->undo_retrans = 0; | |
1837 | ||
1838 | skb = tcp_write_queue_head(sk); | |
1839 | if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) | |
1840 | tp->undo_marker = 0; | |
1841 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { | |
1842 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; | |
1843 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1844 | } | |
1845 | tcp_verify_left_out(tp); | |
1846 | ||
1847 | /* Too bad if TCP was application limited */ | |
1848 | tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1); | |
1849 | ||
1850 | /* Earlier loss recovery underway (see RFC4138; Appendix B). | |
1851 | * The last condition is necessary at least in tp->frto_counter case. | |
1852 | */ | |
1853 | if (tcp_is_sackfrto(tp) && (tp->frto_counter || | |
1854 | ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) && | |
1855 | after(tp->high_seq, tp->snd_una)) { | |
1856 | tp->frto_highmark = tp->high_seq; | |
1857 | } else { | |
1858 | tp->frto_highmark = tp->snd_nxt; | |
1859 | } | |
1860 | tcp_set_ca_state(sk, TCP_CA_Disorder); | |
1861 | tp->high_seq = tp->snd_nxt; | |
1862 | tp->frto_counter = 1; | |
1863 | } | |
1864 | ||
1865 | /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, | |
1866 | * which indicates that we should follow the traditional RTO recovery, | |
1867 | * i.e. mark everything lost and do go-back-N retransmission. | |
1868 | */ | |
1869 | static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag) | |
1870 | { | |
1871 | struct tcp_sock *tp = tcp_sk(sk); | |
1872 | struct sk_buff *skb; | |
1873 | ||
1874 | tp->lost_out = 0; | |
1875 | tp->retrans_out = 0; | |
1876 | if (tcp_is_reno(tp)) | |
1877 | tcp_reset_reno_sack(tp); | |
1878 | ||
1879 | tcp_for_write_queue(skb, sk) { | |
1880 | if (skb == tcp_send_head(sk)) | |
1881 | break; | |
1882 | ||
1883 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; | |
1884 | /* | |
1885 | * Count the retransmission made on RTO correctly (only when | |
1886 | * waiting for the first ACK and did not get it)... | |
1887 | */ | |
1888 | if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) { | |
1889 | /* For some reason this R-bit might get cleared? */ | |
1890 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
1891 | tp->retrans_out += tcp_skb_pcount(skb); | |
1892 | /* ...enter this if branch just for the first segment */ | |
1893 | flag |= FLAG_DATA_ACKED; | |
1894 | } else { | |
1895 | if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) | |
1896 | tp->undo_marker = 0; | |
1897 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; | |
1898 | } | |
1899 | ||
1900 | /* Marking forward transmissions that were made after RTO lost | |
1901 | * can cause unnecessary retransmissions in some scenarios, | |
1902 | * SACK blocks will mitigate that in some but not in all cases. | |
1903 | * We used to not mark them but it was causing break-ups with | |
1904 | * receivers that do only in-order receival. | |
1905 | * | |
1906 | * TODO: we could detect presence of such receiver and select | |
1907 | * different behavior per flow. | |
1908 | */ | |
1909 | if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { | |
1910 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1911 | tp->lost_out += tcp_skb_pcount(skb); | |
1912 | tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; | |
1913 | } | |
1914 | } | |
1915 | tcp_verify_left_out(tp); | |
1916 | ||
1917 | tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments; | |
1918 | tp->snd_cwnd_cnt = 0; | |
1919 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1920 | tp->frto_counter = 0; | |
1921 | tp->bytes_acked = 0; | |
1922 | ||
1923 | tp->reordering = min_t(unsigned int, tp->reordering, | |
1924 | sysctl_tcp_reordering); | |
1925 | tcp_set_ca_state(sk, TCP_CA_Loss); | |
1926 | tp->high_seq = tp->snd_nxt; | |
1927 | TCP_ECN_queue_cwr(tp); | |
1928 | ||
1929 | tcp_clear_all_retrans_hints(tp); | |
1930 | } | |
1931 | ||
1932 | static void tcp_clear_retrans_partial(struct tcp_sock *tp) | |
1933 | { | |
1934 | tp->retrans_out = 0; | |
1935 | tp->lost_out = 0; | |
1936 | ||
1937 | tp->undo_marker = 0; | |
1938 | tp->undo_retrans = 0; | |
1939 | } | |
1940 | ||
1941 | void tcp_clear_retrans(struct tcp_sock *tp) | |
1942 | { | |
1943 | tcp_clear_retrans_partial(tp); | |
1944 | ||
1945 | tp->fackets_out = 0; | |
1946 | tp->sacked_out = 0; | |
1947 | } | |
1948 | ||
1949 | /* Enter Loss state. If "how" is not zero, forget all SACK information | |
1950 | * and reset tags completely, otherwise preserve SACKs. If receiver | |
1951 | * dropped its ofo queue, we will know this due to reneging detection. | |
1952 | */ | |
1953 | void tcp_enter_loss(struct sock *sk, int how) | |
1954 | { | |
1955 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
1956 | struct tcp_sock *tp = tcp_sk(sk); | |
1957 | struct sk_buff *skb; | |
1958 | ||
1959 | /* Reduce ssthresh if it has not yet been made inside this window. */ | |
1960 | if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || | |
1961 | (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { | |
1962 | tp->prior_ssthresh = tcp_current_ssthresh(sk); | |
1963 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
1964 | tcp_ca_event(sk, CA_EVENT_LOSS); | |
1965 | } | |
1966 | tp->snd_cwnd = 1; | |
1967 | tp->snd_cwnd_cnt = 0; | |
1968 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1969 | ||
1970 | tp->bytes_acked = 0; | |
1971 | tcp_clear_retrans_partial(tp); | |
1972 | ||
1973 | if (tcp_is_reno(tp)) | |
1974 | tcp_reset_reno_sack(tp); | |
1975 | ||
1976 | if (!how) { | |
1977 | /* Push undo marker, if it was plain RTO and nothing | |
1978 | * was retransmitted. */ | |
1979 | tp->undo_marker = tp->snd_una; | |
1980 | } else { | |
1981 | tp->sacked_out = 0; | |
1982 | tp->fackets_out = 0; | |
1983 | } | |
1984 | tcp_clear_all_retrans_hints(tp); | |
1985 | ||
1986 | tcp_for_write_queue(skb, sk) { | |
1987 | if (skb == tcp_send_head(sk)) | |
1988 | break; | |
1989 | ||
1990 | if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) | |
1991 | tp->undo_marker = 0; | |
1992 | TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; | |
1993 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { | |
1994 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; | |
1995 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1996 | tp->lost_out += tcp_skb_pcount(skb); | |
1997 | tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; | |
1998 | } | |
1999 | } | |
2000 | tcp_verify_left_out(tp); | |
2001 | ||
2002 | tp->reordering = min_t(unsigned int, tp->reordering, | |
2003 | sysctl_tcp_reordering); | |
2004 | tcp_set_ca_state(sk, TCP_CA_Loss); | |
2005 | tp->high_seq = tp->snd_nxt; | |
2006 | TCP_ECN_queue_cwr(tp); | |
2007 | /* Abort F-RTO algorithm if one is in progress */ | |
2008 | tp->frto_counter = 0; | |
2009 | } | |
2010 | ||
2011 | /* If ACK arrived pointing to a remembered SACK, it means that our | |
2012 | * remembered SACKs do not reflect real state of receiver i.e. | |
2013 | * receiver _host_ is heavily congested (or buggy). | |
2014 | * | |
2015 | * Do processing similar to RTO timeout. | |
2016 | */ | |
2017 | static int tcp_check_sack_reneging(struct sock *sk, int flag) | |
2018 | { | |
2019 | if (flag & FLAG_SACK_RENEGING) { | |
2020 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2021 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING); | |
2022 | ||
2023 | tcp_enter_loss(sk, 1); | |
2024 | icsk->icsk_retransmits++; | |
2025 | tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); | |
2026 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, | |
2027 | icsk->icsk_rto, TCP_RTO_MAX); | |
2028 | return 1; | |
2029 | } | |
2030 | return 0; | |
2031 | } | |
2032 | ||
2033 | static inline int tcp_fackets_out(struct tcp_sock *tp) | |
2034 | { | |
2035 | return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out; | |
2036 | } | |
2037 | ||
2038 | /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs | |
2039 | * counter when SACK is enabled (without SACK, sacked_out is used for | |
2040 | * that purpose). | |
2041 | * | |
2042 | * Instead, with FACK TCP uses fackets_out that includes both SACKed | |
2043 | * segments up to the highest received SACK block so far and holes in | |
2044 | * between them. | |
2045 | * | |
2046 | * With reordering, holes may still be in flight, so RFC3517 recovery | |
2047 | * uses pure sacked_out (total number of SACKed segments) even though | |
2048 | * it violates the RFC that uses duplicate ACKs, often these are equal | |
2049 | * but when e.g. out-of-window ACKs or packet duplication occurs, | |
2050 | * they differ. Since neither occurs due to loss, TCP should really | |
2051 | * ignore them. | |
2052 | */ | |
2053 | static inline int tcp_dupack_heurestics(struct tcp_sock *tp) | |
2054 | { | |
2055 | return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1; | |
2056 | } | |
2057 | ||
2058 | static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb) | |
2059 | { | |
2060 | return (tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto); | |
2061 | } | |
2062 | ||
2063 | static inline int tcp_head_timedout(struct sock *sk) | |
2064 | { | |
2065 | struct tcp_sock *tp = tcp_sk(sk); | |
2066 | ||
2067 | return tp->packets_out && | |
2068 | tcp_skb_timedout(sk, tcp_write_queue_head(sk)); | |
2069 | } | |
2070 | ||
2071 | /* Linux NewReno/SACK/FACK/ECN state machine. | |
2072 | * -------------------------------------- | |
2073 | * | |
2074 | * "Open" Normal state, no dubious events, fast path. | |
2075 | * "Disorder" In all the respects it is "Open", | |
2076 | * but requires a bit more attention. It is entered when | |
2077 | * we see some SACKs or dupacks. It is split of "Open" | |
2078 | * mainly to move some processing from fast path to slow one. | |
2079 | * "CWR" CWND was reduced due to some Congestion Notification event. | |
2080 | * It can be ECN, ICMP source quench, local device congestion. | |
2081 | * "Recovery" CWND was reduced, we are fast-retransmitting. | |
2082 | * "Loss" CWND was reduced due to RTO timeout or SACK reneging. | |
2083 | * | |
2084 | * tcp_fastretrans_alert() is entered: | |
2085 | * - each incoming ACK, if state is not "Open" | |
2086 | * - when arrived ACK is unusual, namely: | |
2087 | * * SACK | |
2088 | * * Duplicate ACK. | |
2089 | * * ECN ECE. | |
2090 | * | |
2091 | * Counting packets in flight is pretty simple. | |
2092 | * | |
2093 | * in_flight = packets_out - left_out + retrans_out | |
2094 | * | |
2095 | * packets_out is SND.NXT-SND.UNA counted in packets. | |
2096 | * | |
2097 | * retrans_out is number of retransmitted segments. | |
2098 | * | |
2099 | * left_out is number of segments left network, but not ACKed yet. | |
2100 | * | |
2101 | * left_out = sacked_out + lost_out | |
2102 | * | |
2103 | * sacked_out: Packets, which arrived to receiver out of order | |
2104 | * and hence not ACKed. With SACKs this number is simply | |
2105 | * amount of SACKed data. Even without SACKs | |
2106 | * it is easy to give pretty reliable estimate of this number, | |
2107 | * counting duplicate ACKs. | |
2108 | * | |
2109 | * lost_out: Packets lost by network. TCP has no explicit | |
2110 | * "loss notification" feedback from network (for now). | |
2111 | * It means that this number can be only _guessed_. | |
2112 | * Actually, it is the heuristics to predict lossage that | |
2113 | * distinguishes different algorithms. | |
2114 | * | |
2115 | * F.e. after RTO, when all the queue is considered as lost, | |
2116 | * lost_out = packets_out and in_flight = retrans_out. | |
2117 | * | |
2118 | * Essentially, we have now two algorithms counting | |
2119 | * lost packets. | |
2120 | * | |
2121 | * FACK: It is the simplest heuristics. As soon as we decided | |
2122 | * that something is lost, we decide that _all_ not SACKed | |
2123 | * packets until the most forward SACK are lost. I.e. | |
2124 | * lost_out = fackets_out - sacked_out and left_out = fackets_out. | |
2125 | * It is absolutely correct estimate, if network does not reorder | |
2126 | * packets. And it loses any connection to reality when reordering | |
2127 | * takes place. We use FACK by default until reordering | |
2128 | * is suspected on the path to this destination. | |
2129 | * | |
2130 | * NewReno: when Recovery is entered, we assume that one segment | |
2131 | * is lost (classic Reno). While we are in Recovery and | |
2132 | * a partial ACK arrives, we assume that one more packet | |
2133 | * is lost (NewReno). This heuristics are the same in NewReno | |
2134 | * and SACK. | |
2135 | * | |
2136 | * Imagine, that's all! Forget about all this shamanism about CWND inflation | |
2137 | * deflation etc. CWND is real congestion window, never inflated, changes | |
2138 | * only according to classic VJ rules. | |
2139 | * | |
2140 | * Really tricky (and requiring careful tuning) part of algorithm | |
2141 | * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). | |
2142 | * The first determines the moment _when_ we should reduce CWND and, | |
2143 | * hence, slow down forward transmission. In fact, it determines the moment | |
2144 | * when we decide that hole is caused by loss, rather than by a reorder. | |
2145 | * | |
2146 | * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill | |
2147 | * holes, caused by lost packets. | |
2148 | * | |
2149 | * And the most logically complicated part of algorithm is undo | |
2150 | * heuristics. We detect false retransmits due to both too early | |
2151 | * fast retransmit (reordering) and underestimated RTO, analyzing | |
2152 | * timestamps and D-SACKs. When we detect that some segments were | |
2153 | * retransmitted by mistake and CWND reduction was wrong, we undo | |
2154 | * window reduction and abort recovery phase. This logic is hidden | |
2155 | * inside several functions named tcp_try_undo_<something>. | |
2156 | */ | |
2157 | ||
2158 | /* This function decides, when we should leave Disordered state | |
2159 | * and enter Recovery phase, reducing congestion window. | |
2160 | * | |
2161 | * Main question: may we further continue forward transmission | |
2162 | * with the same cwnd? | |
2163 | */ | |
2164 | static int tcp_time_to_recover(struct sock *sk) | |
2165 | { | |
2166 | struct tcp_sock *tp = tcp_sk(sk); | |
2167 | __u32 packets_out; | |
2168 | ||
2169 | /* Do not perform any recovery during F-RTO algorithm */ | |
2170 | if (tp->frto_counter) | |
2171 | return 0; | |
2172 | ||
2173 | /* Trick#1: The loss is proven. */ | |
2174 | if (tp->lost_out) | |
2175 | return 1; | |
2176 | ||
2177 | /* Not-A-Trick#2 : Classic rule... */ | |
2178 | if (tcp_dupack_heurestics(tp) > tp->reordering) | |
2179 | return 1; | |
2180 | ||
2181 | /* Trick#3 : when we use RFC2988 timer restart, fast | |
2182 | * retransmit can be triggered by timeout of queue head. | |
2183 | */ | |
2184 | if (tcp_is_fack(tp) && tcp_head_timedout(sk)) | |
2185 | return 1; | |
2186 | ||
2187 | /* Trick#4: It is still not OK... But will it be useful to delay | |
2188 | * recovery more? | |
2189 | */ | |
2190 | packets_out = tp->packets_out; | |
2191 | if (packets_out <= tp->reordering && | |
2192 | tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && | |
2193 | !tcp_may_send_now(sk)) { | |
2194 | /* We have nothing to send. This connection is limited | |
2195 | * either by receiver window or by application. | |
2196 | */ | |
2197 | return 1; | |
2198 | } | |
2199 | ||
2200 | return 0; | |
2201 | } | |
2202 | ||
2203 | /* Mark head of queue up as lost. With RFC3517 SACK, the packets is | |
2204 | * is against sacked "cnt", otherwise it's against facked "cnt" | |
2205 | */ | |
2206 | static void tcp_mark_head_lost(struct sock *sk, int packets) | |
2207 | { | |
2208 | struct tcp_sock *tp = tcp_sk(sk); | |
2209 | struct sk_buff *skb; | |
2210 | int cnt, oldcnt; | |
2211 | int err; | |
2212 | unsigned int mss; | |
2213 | ||
2214 | WARN_ON(packets > tp->packets_out); | |
2215 | if (tp->lost_skb_hint) { | |
2216 | skb = tp->lost_skb_hint; | |
2217 | cnt = tp->lost_cnt_hint; | |
2218 | } else { | |
2219 | skb = tcp_write_queue_head(sk); | |
2220 | cnt = 0; | |
2221 | } | |
2222 | ||
2223 | tcp_for_write_queue_from(skb, sk) { | |
2224 | if (skb == tcp_send_head(sk)) | |
2225 | break; | |
2226 | /* TODO: do this better */ | |
2227 | /* this is not the most efficient way to do this... */ | |
2228 | tp->lost_skb_hint = skb; | |
2229 | tp->lost_cnt_hint = cnt; | |
2230 | ||
2231 | if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq)) | |
2232 | break; | |
2233 | ||
2234 | oldcnt = cnt; | |
2235 | if (tcp_is_fack(tp) || tcp_is_reno(tp) || | |
2236 | (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) | |
2237 | cnt += tcp_skb_pcount(skb); | |
2238 | ||
2239 | if (cnt > packets) { | |
2240 | if (tcp_is_sack(tp) || (oldcnt >= packets)) | |
2241 | break; | |
2242 | ||
2243 | mss = skb_shinfo(skb)->gso_size; | |
2244 | err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss); | |
2245 | if (err < 0) | |
2246 | break; | |
2247 | cnt = packets; | |
2248 | } | |
2249 | ||
2250 | tcp_skb_mark_lost(tp, skb); | |
2251 | } | |
2252 | tcp_verify_left_out(tp); | |
2253 | } | |
2254 | ||
2255 | /* Account newly detected lost packet(s) */ | |
2256 | ||
2257 | static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit) | |
2258 | { | |
2259 | struct tcp_sock *tp = tcp_sk(sk); | |
2260 | ||
2261 | if (tcp_is_reno(tp)) { | |
2262 | tcp_mark_head_lost(sk, 1); | |
2263 | } else if (tcp_is_fack(tp)) { | |
2264 | int lost = tp->fackets_out - tp->reordering; | |
2265 | if (lost <= 0) | |
2266 | lost = 1; | |
2267 | tcp_mark_head_lost(sk, lost); | |
2268 | } else { | |
2269 | int sacked_upto = tp->sacked_out - tp->reordering; | |
2270 | if (sacked_upto < fast_rexmit) | |
2271 | sacked_upto = fast_rexmit; | |
2272 | tcp_mark_head_lost(sk, sacked_upto); | |
2273 | } | |
2274 | ||
2275 | /* New heuristics: it is possible only after we switched | |
2276 | * to restart timer each time when something is ACKed. | |
2277 | * Hence, we can detect timed out packets during fast | |
2278 | * retransmit without falling to slow start. | |
2279 | */ | |
2280 | if (tcp_is_fack(tp) && tcp_head_timedout(sk)) { | |
2281 | struct sk_buff *skb; | |
2282 | ||
2283 | skb = tp->scoreboard_skb_hint ? tp->scoreboard_skb_hint | |
2284 | : tcp_write_queue_head(sk); | |
2285 | ||
2286 | tcp_for_write_queue_from(skb, sk) { | |
2287 | if (skb == tcp_send_head(sk)) | |
2288 | break; | |
2289 | if (!tcp_skb_timedout(sk, skb)) | |
2290 | break; | |
2291 | ||
2292 | tcp_skb_mark_lost(tp, skb); | |
2293 | } | |
2294 | ||
2295 | tp->scoreboard_skb_hint = skb; | |
2296 | ||
2297 | tcp_verify_left_out(tp); | |
2298 | } | |
2299 | } | |
2300 | ||
2301 | /* CWND moderation, preventing bursts due to too big ACKs | |
2302 | * in dubious situations. | |
2303 | */ | |
2304 | static inline void tcp_moderate_cwnd(struct tcp_sock *tp) | |
2305 | { | |
2306 | tp->snd_cwnd = min(tp->snd_cwnd, | |
2307 | tcp_packets_in_flight(tp) + tcp_max_burst(tp)); | |
2308 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
2309 | } | |
2310 | ||
2311 | /* Lower bound on congestion window is slow start threshold | |
2312 | * unless congestion avoidance choice decides to overide it. | |
2313 | */ | |
2314 | static inline u32 tcp_cwnd_min(const struct sock *sk) | |
2315 | { | |
2316 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; | |
2317 | ||
2318 | return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh; | |
2319 | } | |
2320 | ||
2321 | /* Decrease cwnd each second ack. */ | |
2322 | static void tcp_cwnd_down(struct sock *sk, int flag) | |
2323 | { | |
2324 | struct tcp_sock *tp = tcp_sk(sk); | |
2325 | int decr = tp->snd_cwnd_cnt + 1; | |
2326 | ||
2327 | if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) || | |
2328 | (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) { | |
2329 | tp->snd_cwnd_cnt = decr & 1; | |
2330 | decr >>= 1; | |
2331 | ||
2332 | if (decr && tp->snd_cwnd > tcp_cwnd_min(sk)) | |
2333 | tp->snd_cwnd -= decr; | |
2334 | ||
2335 | tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1); | |
2336 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
2337 | } | |
2338 | } | |
2339 | ||
2340 | /* Nothing was retransmitted or returned timestamp is less | |
2341 | * than timestamp of the first retransmission. | |
2342 | */ | |
2343 | static inline int tcp_packet_delayed(struct tcp_sock *tp) | |
2344 | { | |
2345 | return !tp->retrans_stamp || | |
2346 | (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && | |
2347 | before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp)); | |
2348 | } | |
2349 | ||
2350 | /* Undo procedures. */ | |
2351 | ||
2352 | #if FASTRETRANS_DEBUG > 1 | |
2353 | static void DBGUNDO(struct sock *sk, const char *msg) | |
2354 | { | |
2355 | struct tcp_sock *tp = tcp_sk(sk); | |
2356 | struct inet_sock *inet = inet_sk(sk); | |
2357 | ||
2358 | if (sk->sk_family == AF_INET) { | |
2359 | printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n", | |
2360 | msg, | |
2361 | &inet->daddr, ntohs(inet->dport), | |
2362 | tp->snd_cwnd, tcp_left_out(tp), | |
2363 | tp->snd_ssthresh, tp->prior_ssthresh, | |
2364 | tp->packets_out); | |
2365 | } | |
2366 | #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) | |
2367 | else if (sk->sk_family == AF_INET6) { | |
2368 | struct ipv6_pinfo *np = inet6_sk(sk); | |
2369 | printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n", | |
2370 | msg, | |
2371 | &np->daddr, ntohs(inet->dport), | |
2372 | tp->snd_cwnd, tcp_left_out(tp), | |
2373 | tp->snd_ssthresh, tp->prior_ssthresh, | |
2374 | tp->packets_out); | |
2375 | } | |
2376 | #endif | |
2377 | } | |
2378 | #else | |
2379 | #define DBGUNDO(x...) do { } while (0) | |
2380 | #endif | |
2381 | ||
2382 | static void tcp_undo_cwr(struct sock *sk, const int undo) | |
2383 | { | |
2384 | struct tcp_sock *tp = tcp_sk(sk); | |
2385 | ||
2386 | if (tp->prior_ssthresh) { | |
2387 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
2388 | ||
2389 | if (icsk->icsk_ca_ops->undo_cwnd) | |
2390 | tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); | |
2391 | else | |
2392 | tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1); | |
2393 | ||
2394 | if (undo && tp->prior_ssthresh > tp->snd_ssthresh) { | |
2395 | tp->snd_ssthresh = tp->prior_ssthresh; | |
2396 | TCP_ECN_withdraw_cwr(tp); | |
2397 | } | |
2398 | } else { | |
2399 | tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); | |
2400 | } | |
2401 | tcp_moderate_cwnd(tp); | |
2402 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
2403 | } | |
2404 | ||
2405 | static inline int tcp_may_undo(struct tcp_sock *tp) | |
2406 | { | |
2407 | return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); | |
2408 | } | |
2409 | ||
2410 | /* People celebrate: "We love our President!" */ | |
2411 | static int tcp_try_undo_recovery(struct sock *sk) | |
2412 | { | |
2413 | struct tcp_sock *tp = tcp_sk(sk); | |
2414 | ||
2415 | if (tcp_may_undo(tp)) { | |
2416 | int mib_idx; | |
2417 | ||
2418 | /* Happy end! We did not retransmit anything | |
2419 | * or our original transmission succeeded. | |
2420 | */ | |
2421 | DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); | |
2422 | tcp_undo_cwr(sk, 1); | |
2423 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) | |
2424 | mib_idx = LINUX_MIB_TCPLOSSUNDO; | |
2425 | else | |
2426 | mib_idx = LINUX_MIB_TCPFULLUNDO; | |
2427 | ||
2428 | NET_INC_STATS_BH(sock_net(sk), mib_idx); | |
2429 | tp->undo_marker = 0; | |
2430 | } | |
2431 | if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { | |
2432 | /* Hold old state until something *above* high_seq | |
2433 | * is ACKed. For Reno it is MUST to prevent false | |
2434 | * fast retransmits (RFC2582). SACK TCP is safe. */ | |
2435 | tcp_moderate_cwnd(tp); | |
2436 | return 1; | |
2437 | } | |
2438 | tcp_set_ca_state(sk, TCP_CA_Open); | |
2439 | return 0; | |
2440 | } | |
2441 | ||
2442 | /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ | |
2443 | static void tcp_try_undo_dsack(struct sock *sk) | |
2444 | { | |
2445 | struct tcp_sock *tp = tcp_sk(sk); | |
2446 | ||
2447 | if (tp->undo_marker && !tp->undo_retrans) { | |
2448 | DBGUNDO(sk, "D-SACK"); | |
2449 | tcp_undo_cwr(sk, 1); | |
2450 | tp->undo_marker = 0; | |
2451 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO); | |
2452 | } | |
2453 | } | |
2454 | ||
2455 | /* Undo during fast recovery after partial ACK. */ | |
2456 | ||
2457 | static int tcp_try_undo_partial(struct sock *sk, int acked) | |
2458 | { | |
2459 | struct tcp_sock *tp = tcp_sk(sk); | |
2460 | /* Partial ACK arrived. Force Hoe's retransmit. */ | |
2461 | int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering); | |
2462 | ||
2463 | if (tcp_may_undo(tp)) { | |
2464 | /* Plain luck! Hole if filled with delayed | |
2465 | * packet, rather than with a retransmit. | |
2466 | */ | |
2467 | if (tp->retrans_out == 0) | |
2468 | tp->retrans_stamp = 0; | |
2469 | ||
2470 | tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); | |
2471 | ||
2472 | DBGUNDO(sk, "Hoe"); | |
2473 | tcp_undo_cwr(sk, 0); | |
2474 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO); | |
2475 | ||
2476 | /* So... Do not make Hoe's retransmit yet. | |
2477 | * If the first packet was delayed, the rest | |
2478 | * ones are most probably delayed as well. | |
2479 | */ | |
2480 | failed = 0; | |
2481 | } | |
2482 | return failed; | |
2483 | } | |
2484 | ||
2485 | /* Undo during loss recovery after partial ACK. */ | |
2486 | static int tcp_try_undo_loss(struct sock *sk) | |
2487 | { | |
2488 | struct tcp_sock *tp = tcp_sk(sk); | |
2489 | ||
2490 | if (tcp_may_undo(tp)) { | |
2491 | struct sk_buff *skb; | |
2492 | tcp_for_write_queue(skb, sk) { | |
2493 | if (skb == tcp_send_head(sk)) | |
2494 | break; | |
2495 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; | |
2496 | } | |
2497 | ||
2498 | tcp_clear_all_retrans_hints(tp); | |
2499 | ||
2500 | DBGUNDO(sk, "partial loss"); | |
2501 | tp->lost_out = 0; | |
2502 | tcp_undo_cwr(sk, 1); | |
2503 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO); | |
2504 | inet_csk(sk)->icsk_retransmits = 0; | |
2505 | tp->undo_marker = 0; | |
2506 | if (tcp_is_sack(tp)) | |
2507 | tcp_set_ca_state(sk, TCP_CA_Open); | |
2508 | return 1; | |
2509 | } | |
2510 | return 0; | |
2511 | } | |
2512 | ||
2513 | static inline void tcp_complete_cwr(struct sock *sk) | |
2514 | { | |
2515 | struct tcp_sock *tp = tcp_sk(sk); | |
2516 | tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); | |
2517 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
2518 | tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); | |
2519 | } | |
2520 | ||
2521 | static void tcp_try_keep_open(struct sock *sk) | |
2522 | { | |
2523 | struct tcp_sock *tp = tcp_sk(sk); | |
2524 | int state = TCP_CA_Open; | |
2525 | ||
2526 | if (tcp_left_out(tp) || tp->retrans_out || tp->undo_marker) | |
2527 | state = TCP_CA_Disorder; | |
2528 | ||
2529 | if (inet_csk(sk)->icsk_ca_state != state) { | |
2530 | tcp_set_ca_state(sk, state); | |
2531 | tp->high_seq = tp->snd_nxt; | |
2532 | } | |
2533 | } | |
2534 | ||
2535 | static void tcp_try_to_open(struct sock *sk, int flag) | |
2536 | { | |
2537 | struct tcp_sock *tp = tcp_sk(sk); | |
2538 | ||
2539 | tcp_verify_left_out(tp); | |
2540 | ||
2541 | if (!tp->frto_counter && tp->retrans_out == 0) | |
2542 | tp->retrans_stamp = 0; | |
2543 | ||
2544 | if (flag & FLAG_ECE) | |
2545 | tcp_enter_cwr(sk, 1); | |
2546 | ||
2547 | if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { | |
2548 | tcp_try_keep_open(sk); | |
2549 | tcp_moderate_cwnd(tp); | |
2550 | } else { | |
2551 | tcp_cwnd_down(sk, flag); | |
2552 | } | |
2553 | } | |
2554 | ||
2555 | static void tcp_mtup_probe_failed(struct sock *sk) | |
2556 | { | |
2557 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2558 | ||
2559 | icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; | |
2560 | icsk->icsk_mtup.probe_size = 0; | |
2561 | } | |
2562 | ||
2563 | static void tcp_mtup_probe_success(struct sock *sk, struct sk_buff *skb) | |
2564 | { | |
2565 | struct tcp_sock *tp = tcp_sk(sk); | |
2566 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2567 | ||
2568 | /* FIXME: breaks with very large cwnd */ | |
2569 | tp->prior_ssthresh = tcp_current_ssthresh(sk); | |
2570 | tp->snd_cwnd = tp->snd_cwnd * | |
2571 | tcp_mss_to_mtu(sk, tp->mss_cache) / | |
2572 | icsk->icsk_mtup.probe_size; | |
2573 | tp->snd_cwnd_cnt = 0; | |
2574 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
2575 | tp->rcv_ssthresh = tcp_current_ssthresh(sk); | |
2576 | ||
2577 | icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; | |
2578 | icsk->icsk_mtup.probe_size = 0; | |
2579 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); | |
2580 | } | |
2581 | ||
2582 | /* Do a simple retransmit without using the backoff mechanisms in | |
2583 | * tcp_timer. This is used for path mtu discovery. | |
2584 | * The socket is already locked here. | |
2585 | */ | |
2586 | void tcp_simple_retransmit(struct sock *sk) | |
2587 | { | |
2588 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
2589 | struct tcp_sock *tp = tcp_sk(sk); | |
2590 | struct sk_buff *skb; | |
2591 | unsigned int mss = tcp_current_mss(sk, 0); | |
2592 | u32 prior_lost = tp->lost_out; | |
2593 | ||
2594 | tcp_for_write_queue(skb, sk) { | |
2595 | if (skb == tcp_send_head(sk)) | |
2596 | break; | |
2597 | if (skb->len > mss && | |
2598 | !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { | |
2599 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { | |
2600 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; | |
2601 | tp->retrans_out -= tcp_skb_pcount(skb); | |
2602 | } | |
2603 | tcp_skb_mark_lost_uncond_verify(tp, skb); | |
2604 | } | |
2605 | } | |
2606 | ||
2607 | tcp_clear_retrans_hints_partial(tp); | |
2608 | ||
2609 | if (prior_lost == tp->lost_out) | |
2610 | return; | |
2611 | ||
2612 | if (tcp_is_reno(tp)) | |
2613 | tcp_limit_reno_sacked(tp); | |
2614 | ||
2615 | tcp_verify_left_out(tp); | |
2616 | ||
2617 | /* Don't muck with the congestion window here. | |
2618 | * Reason is that we do not increase amount of _data_ | |
2619 | * in network, but units changed and effective | |
2620 | * cwnd/ssthresh really reduced now. | |
2621 | */ | |
2622 | if (icsk->icsk_ca_state != TCP_CA_Loss) { | |
2623 | tp->high_seq = tp->snd_nxt; | |
2624 | tp->snd_ssthresh = tcp_current_ssthresh(sk); | |
2625 | tp->prior_ssthresh = 0; | |
2626 | tp->undo_marker = 0; | |
2627 | tcp_set_ca_state(sk, TCP_CA_Loss); | |
2628 | } | |
2629 | tcp_xmit_retransmit_queue(sk); | |
2630 | } | |
2631 | ||
2632 | /* Process an event, which can update packets-in-flight not trivially. | |
2633 | * Main goal of this function is to calculate new estimate for left_out, | |
2634 | * taking into account both packets sitting in receiver's buffer and | |
2635 | * packets lost by network. | |
2636 | * | |
2637 | * Besides that it does CWND reduction, when packet loss is detected | |
2638 | * and changes state of machine. | |
2639 | * | |
2640 | * It does _not_ decide what to send, it is made in function | |
2641 | * tcp_xmit_retransmit_queue(). | |
2642 | */ | |
2643 | static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag) | |
2644 | { | |
2645 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2646 | struct tcp_sock *tp = tcp_sk(sk); | |
2647 | int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP)); | |
2648 | int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) && | |
2649 | (tcp_fackets_out(tp) > tp->reordering)); | |
2650 | int fast_rexmit = 0, mib_idx; | |
2651 | ||
2652 | if (WARN_ON(!tp->packets_out && tp->sacked_out)) | |
2653 | tp->sacked_out = 0; | |
2654 | if (WARN_ON(!tp->sacked_out && tp->fackets_out)) | |
2655 | tp->fackets_out = 0; | |
2656 | ||
2657 | /* Now state machine starts. | |
2658 | * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ | |
2659 | if (flag & FLAG_ECE) | |
2660 | tp->prior_ssthresh = 0; | |
2661 | ||
2662 | /* B. In all the states check for reneging SACKs. */ | |
2663 | if (tcp_check_sack_reneging(sk, flag)) | |
2664 | return; | |
2665 | ||
2666 | /* C. Process data loss notification, provided it is valid. */ | |
2667 | if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) && | |
2668 | before(tp->snd_una, tp->high_seq) && | |
2669 | icsk->icsk_ca_state != TCP_CA_Open && | |
2670 | tp->fackets_out > tp->reordering) { | |
2671 | tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering); | |
2672 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS); | |
2673 | } | |
2674 | ||
2675 | /* D. Check consistency of the current state. */ | |
2676 | tcp_verify_left_out(tp); | |
2677 | ||
2678 | /* E. Check state exit conditions. State can be terminated | |
2679 | * when high_seq is ACKed. */ | |
2680 | if (icsk->icsk_ca_state == TCP_CA_Open) { | |
2681 | WARN_ON(tp->retrans_out != 0); | |
2682 | tp->retrans_stamp = 0; | |
2683 | } else if (!before(tp->snd_una, tp->high_seq)) { | |
2684 | switch (icsk->icsk_ca_state) { | |
2685 | case TCP_CA_Loss: | |
2686 | icsk->icsk_retransmits = 0; | |
2687 | if (tcp_try_undo_recovery(sk)) | |
2688 | return; | |
2689 | break; | |
2690 | ||
2691 | case TCP_CA_CWR: | |
2692 | /* CWR is to be held something *above* high_seq | |
2693 | * is ACKed for CWR bit to reach receiver. */ | |
2694 | if (tp->snd_una != tp->high_seq) { | |
2695 | tcp_complete_cwr(sk); | |
2696 | tcp_set_ca_state(sk, TCP_CA_Open); | |
2697 | } | |
2698 | break; | |
2699 | ||
2700 | case TCP_CA_Disorder: | |
2701 | tcp_try_undo_dsack(sk); | |
2702 | if (!tp->undo_marker || | |
2703 | /* For SACK case do not Open to allow to undo | |
2704 | * catching for all duplicate ACKs. */ | |
2705 | tcp_is_reno(tp) || tp->snd_una != tp->high_seq) { | |
2706 | tp->undo_marker = 0; | |
2707 | tcp_set_ca_state(sk, TCP_CA_Open); | |
2708 | } | |
2709 | break; | |
2710 | ||
2711 | case TCP_CA_Recovery: | |
2712 | if (tcp_is_reno(tp)) | |
2713 | tcp_reset_reno_sack(tp); | |
2714 | if (tcp_try_undo_recovery(sk)) | |
2715 | return; | |
2716 | tcp_complete_cwr(sk); | |
2717 | break; | |
2718 | } | |
2719 | } | |
2720 | ||
2721 | /* F. Process state. */ | |
2722 | switch (icsk->icsk_ca_state) { | |
2723 | case TCP_CA_Recovery: | |
2724 | if (!(flag & FLAG_SND_UNA_ADVANCED)) { | |
2725 | if (tcp_is_reno(tp) && is_dupack) | |
2726 | tcp_add_reno_sack(sk); | |
2727 | } else | |
2728 | do_lost = tcp_try_undo_partial(sk, pkts_acked); | |
2729 | break; | |
2730 | case TCP_CA_Loss: | |
2731 | if (flag & FLAG_DATA_ACKED) | |
2732 | icsk->icsk_retransmits = 0; | |
2733 | if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED) | |
2734 | tcp_reset_reno_sack(tp); | |
2735 | if (!tcp_try_undo_loss(sk)) { | |
2736 | tcp_moderate_cwnd(tp); | |
2737 | tcp_xmit_retransmit_queue(sk); | |
2738 | return; | |
2739 | } | |
2740 | if (icsk->icsk_ca_state != TCP_CA_Open) | |
2741 | return; | |
2742 | /* Loss is undone; fall through to processing in Open state. */ | |
2743 | default: | |
2744 | if (tcp_is_reno(tp)) { | |
2745 | if (flag & FLAG_SND_UNA_ADVANCED) | |
2746 | tcp_reset_reno_sack(tp); | |
2747 | if (is_dupack) | |
2748 | tcp_add_reno_sack(sk); | |
2749 | } | |
2750 | ||
2751 | if (icsk->icsk_ca_state == TCP_CA_Disorder) | |
2752 | tcp_try_undo_dsack(sk); | |
2753 | ||
2754 | if (!tcp_time_to_recover(sk)) { | |
2755 | tcp_try_to_open(sk, flag); | |
2756 | return; | |
2757 | } | |
2758 | ||
2759 | /* MTU probe failure: don't reduce cwnd */ | |
2760 | if (icsk->icsk_ca_state < TCP_CA_CWR && | |
2761 | icsk->icsk_mtup.probe_size && | |
2762 | tp->snd_una == tp->mtu_probe.probe_seq_start) { | |
2763 | tcp_mtup_probe_failed(sk); | |
2764 | /* Restores the reduction we did in tcp_mtup_probe() */ | |
2765 | tp->snd_cwnd++; | |
2766 | tcp_simple_retransmit(sk); | |
2767 | return; | |
2768 | } | |
2769 | ||
2770 | /* Otherwise enter Recovery state */ | |
2771 | ||
2772 | if (tcp_is_reno(tp)) | |
2773 | mib_idx = LINUX_MIB_TCPRENORECOVERY; | |
2774 | else | |
2775 | mib_idx = LINUX_MIB_TCPSACKRECOVERY; | |
2776 | ||
2777 | NET_INC_STATS_BH(sock_net(sk), mib_idx); | |
2778 | ||
2779 | tp->high_seq = tp->snd_nxt; | |
2780 | tp->prior_ssthresh = 0; | |
2781 | tp->undo_marker = tp->snd_una; | |
2782 | tp->undo_retrans = tp->retrans_out; | |
2783 | ||
2784 | if (icsk->icsk_ca_state < TCP_CA_CWR) { | |
2785 | if (!(flag & FLAG_ECE)) | |
2786 | tp->prior_ssthresh = tcp_current_ssthresh(sk); | |
2787 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
2788 | TCP_ECN_queue_cwr(tp); | |
2789 | } | |
2790 | ||
2791 | tp->bytes_acked = 0; | |
2792 | tp->snd_cwnd_cnt = 0; | |
2793 | tcp_set_ca_state(sk, TCP_CA_Recovery); | |
2794 | fast_rexmit = 1; | |
2795 | } | |
2796 | ||
2797 | if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk))) | |
2798 | tcp_update_scoreboard(sk, fast_rexmit); | |
2799 | tcp_cwnd_down(sk, flag); | |
2800 | tcp_xmit_retransmit_queue(sk); | |
2801 | } | |
2802 | ||
2803 | /* Read draft-ietf-tcplw-high-performance before mucking | |
2804 | * with this code. (Supersedes RFC1323) | |
2805 | */ | |
2806 | static void tcp_ack_saw_tstamp(struct sock *sk, int flag) | |
2807 | { | |
2808 | /* RTTM Rule: A TSecr value received in a segment is used to | |
2809 | * update the averaged RTT measurement only if the segment | |
2810 | * acknowledges some new data, i.e., only if it advances the | |
2811 | * left edge of the send window. | |
2812 | * | |
2813 | * See draft-ietf-tcplw-high-performance-00, section 3.3. | |
2814 | * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> | |
2815 | * | |
2816 | * Changed: reset backoff as soon as we see the first valid sample. | |
2817 | * If we do not, we get strongly overestimated rto. With timestamps | |
2818 | * samples are accepted even from very old segments: f.e., when rtt=1 | |
2819 | * increases to 8, we retransmit 5 times and after 8 seconds delayed | |
2820 | * answer arrives rto becomes 120 seconds! If at least one of segments | |
2821 | * in window is lost... Voila. --ANK (010210) | |
2822 | */ | |
2823 | struct tcp_sock *tp = tcp_sk(sk); | |
2824 | const __u32 seq_rtt = tcp_time_stamp - tp->rx_opt.rcv_tsecr; | |
2825 | tcp_rtt_estimator(sk, seq_rtt); | |
2826 | tcp_set_rto(sk); | |
2827 | inet_csk(sk)->icsk_backoff = 0; | |
2828 | tcp_bound_rto(sk); | |
2829 | } | |
2830 | ||
2831 | static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) | |
2832 | { | |
2833 | /* We don't have a timestamp. Can only use | |
2834 | * packets that are not retransmitted to determine | |
2835 | * rtt estimates. Also, we must not reset the | |
2836 | * backoff for rto until we get a non-retransmitted | |
2837 | * packet. This allows us to deal with a situation | |
2838 | * where the network delay has increased suddenly. | |
2839 | * I.e. Karn's algorithm. (SIGCOMM '87, p5.) | |
2840 | */ | |
2841 | ||
2842 | if (flag & FLAG_RETRANS_DATA_ACKED) | |
2843 | return; | |
2844 | ||
2845 | tcp_rtt_estimator(sk, seq_rtt); | |
2846 | tcp_set_rto(sk); | |
2847 | inet_csk(sk)->icsk_backoff = 0; | |
2848 | tcp_bound_rto(sk); | |
2849 | } | |
2850 | ||
2851 | static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, | |
2852 | const s32 seq_rtt) | |
2853 | { | |
2854 | const struct tcp_sock *tp = tcp_sk(sk); | |
2855 | /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ | |
2856 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) | |
2857 | tcp_ack_saw_tstamp(sk, flag); | |
2858 | else if (seq_rtt >= 0) | |
2859 | tcp_ack_no_tstamp(sk, seq_rtt, flag); | |
2860 | } | |
2861 | ||
2862 | static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight) | |
2863 | { | |
2864 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
2865 | icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight); | |
2866 | tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; | |
2867 | } | |
2868 | ||
2869 | /* Restart timer after forward progress on connection. | |
2870 | * RFC2988 recommends to restart timer to now+rto. | |
2871 | */ | |
2872 | static void tcp_rearm_rto(struct sock *sk) | |
2873 | { | |
2874 | struct tcp_sock *tp = tcp_sk(sk); | |
2875 | ||
2876 | if (!tp->packets_out) { | |
2877 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); | |
2878 | } else { | |
2879 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, | |
2880 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
2881 | } | |
2882 | } | |
2883 | ||
2884 | /* If we get here, the whole TSO packet has not been acked. */ | |
2885 | static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) | |
2886 | { | |
2887 | struct tcp_sock *tp = tcp_sk(sk); | |
2888 | u32 packets_acked; | |
2889 | ||
2890 | BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); | |
2891 | ||
2892 | packets_acked = tcp_skb_pcount(skb); | |
2893 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) | |
2894 | return 0; | |
2895 | packets_acked -= tcp_skb_pcount(skb); | |
2896 | ||
2897 | if (packets_acked) { | |
2898 | BUG_ON(tcp_skb_pcount(skb) == 0); | |
2899 | BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); | |
2900 | } | |
2901 | ||
2902 | return packets_acked; | |
2903 | } | |
2904 | ||
2905 | /* Remove acknowledged frames from the retransmission queue. If our packet | |
2906 | * is before the ack sequence we can discard it as it's confirmed to have | |
2907 | * arrived at the other end. | |
2908 | */ | |
2909 | static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets, | |
2910 | u32 prior_snd_una) | |
2911 | { | |
2912 | struct tcp_sock *tp = tcp_sk(sk); | |
2913 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
2914 | struct sk_buff *skb; | |
2915 | u32 now = tcp_time_stamp; | |
2916 | int fully_acked = 1; | |
2917 | int flag = 0; | |
2918 | u32 pkts_acked = 0; | |
2919 | u32 reord = tp->packets_out; | |
2920 | u32 prior_sacked = tp->sacked_out; | |
2921 | s32 seq_rtt = -1; | |
2922 | s32 ca_seq_rtt = -1; | |
2923 | ktime_t last_ackt = net_invalid_timestamp(); | |
2924 | ||
2925 | while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) { | |
2926 | struct tcp_skb_cb *scb = TCP_SKB_CB(skb); | |
2927 | u32 end_seq; | |
2928 | u32 acked_pcount; | |
2929 | u8 sacked = scb->sacked; | |
2930 | ||
2931 | /* Determine how many packets and what bytes were acked, tso and else */ | |
2932 | if (after(scb->end_seq, tp->snd_una)) { | |
2933 | if (tcp_skb_pcount(skb) == 1 || | |
2934 | !after(tp->snd_una, scb->seq)) | |
2935 | break; | |
2936 | ||
2937 | acked_pcount = tcp_tso_acked(sk, skb); | |
2938 | if (!acked_pcount) | |
2939 | break; | |
2940 | ||
2941 | fully_acked = 0; | |
2942 | end_seq = tp->snd_una; | |
2943 | } else { | |
2944 | acked_pcount = tcp_skb_pcount(skb); | |
2945 | end_seq = scb->end_seq; | |
2946 | } | |
2947 | ||
2948 | /* MTU probing checks */ | |
2949 | if (fully_acked && icsk->icsk_mtup.probe_size && | |
2950 | !after(tp->mtu_probe.probe_seq_end, scb->end_seq)) { | |
2951 | tcp_mtup_probe_success(sk, skb); | |
2952 | } | |
2953 | ||
2954 | if (sacked & TCPCB_RETRANS) { | |
2955 | if (sacked & TCPCB_SACKED_RETRANS) | |
2956 | tp->retrans_out -= acked_pcount; | |
2957 | flag |= FLAG_RETRANS_DATA_ACKED; | |
2958 | ca_seq_rtt = -1; | |
2959 | seq_rtt = -1; | |
2960 | if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1)) | |
2961 | flag |= FLAG_NONHEAD_RETRANS_ACKED; | |
2962 | } else { | |
2963 | ca_seq_rtt = now - scb->when; | |
2964 | last_ackt = skb->tstamp; | |
2965 | if (seq_rtt < 0) { | |
2966 | seq_rtt = ca_seq_rtt; | |
2967 | } | |
2968 | if (!(sacked & TCPCB_SACKED_ACKED)) | |
2969 | reord = min(pkts_acked, reord); | |
2970 | } | |
2971 | ||
2972 | if (sacked & TCPCB_SACKED_ACKED) | |
2973 | tp->sacked_out -= acked_pcount; | |
2974 | if (sacked & TCPCB_LOST) | |
2975 | tp->lost_out -= acked_pcount; | |
2976 | ||
2977 | tp->packets_out -= acked_pcount; | |
2978 | pkts_acked += acked_pcount; | |
2979 | ||
2980 | /* Initial outgoing SYN's get put onto the write_queue | |
2981 | * just like anything else we transmit. It is not | |
2982 | * true data, and if we misinform our callers that | |
2983 | * this ACK acks real data, we will erroneously exit | |
2984 | * connection startup slow start one packet too | |
2985 | * quickly. This is severely frowned upon behavior. | |
2986 | */ | |
2987 | if (!(scb->flags & TCPCB_FLAG_SYN)) { | |
2988 | flag |= FLAG_DATA_ACKED; | |
2989 | } else { | |
2990 | flag |= FLAG_SYN_ACKED; | |
2991 | tp->retrans_stamp = 0; | |
2992 | } | |
2993 | ||
2994 | if (!fully_acked) | |
2995 | break; | |
2996 | ||
2997 | tcp_unlink_write_queue(skb, sk); | |
2998 | sk_wmem_free_skb(sk, skb); | |
2999 | tp->scoreboard_skb_hint = NULL; | |
3000 | if (skb == tp->retransmit_skb_hint) | |
3001 | tp->retransmit_skb_hint = NULL; | |
3002 | if (skb == tp->lost_skb_hint) | |
3003 | tp->lost_skb_hint = NULL; | |
3004 | } | |
3005 | ||
3006 | if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una))) | |
3007 | tp->snd_up = tp->snd_una; | |
3008 | ||
3009 | if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) | |
3010 | flag |= FLAG_SACK_RENEGING; | |
3011 | ||
3012 | if (flag & FLAG_ACKED) { | |
3013 | const struct tcp_congestion_ops *ca_ops | |
3014 | = inet_csk(sk)->icsk_ca_ops; | |
3015 | ||
3016 | tcp_ack_update_rtt(sk, flag, seq_rtt); | |
3017 | tcp_rearm_rto(sk); | |
3018 | ||
3019 | if (tcp_is_reno(tp)) { | |
3020 | tcp_remove_reno_sacks(sk, pkts_acked); | |
3021 | } else { | |
3022 | /* Non-retransmitted hole got filled? That's reordering */ | |
3023 | if (reord < prior_fackets) | |
3024 | tcp_update_reordering(sk, tp->fackets_out - reord, 0); | |
3025 | ||
3026 | /* No need to care for underflows here because | |
3027 | * the lost_skb_hint gets NULLed if we're past it | |
3028 | * (or something non-trivial happened) | |
3029 | */ | |
3030 | if (tcp_is_fack(tp)) | |
3031 | tp->lost_cnt_hint -= pkts_acked; | |
3032 | else | |
3033 | tp->lost_cnt_hint -= prior_sacked - tp->sacked_out; | |
3034 | } | |
3035 | ||
3036 | tp->fackets_out -= min(pkts_acked, tp->fackets_out); | |
3037 | ||
3038 | if (ca_ops->pkts_acked) { | |
3039 | s32 rtt_us = -1; | |
3040 | ||
3041 | /* Is the ACK triggering packet unambiguous? */ | |
3042 | if (!(flag & FLAG_RETRANS_DATA_ACKED)) { | |
3043 | /* High resolution needed and available? */ | |
3044 | if (ca_ops->flags & TCP_CONG_RTT_STAMP && | |
3045 | !ktime_equal(last_ackt, | |
3046 | net_invalid_timestamp())) | |
3047 | rtt_us = ktime_us_delta(ktime_get_real(), | |
3048 | last_ackt); | |
3049 | else if (ca_seq_rtt > 0) | |
3050 | rtt_us = jiffies_to_usecs(ca_seq_rtt); | |
3051 | } | |
3052 | ||
3053 | ca_ops->pkts_acked(sk, pkts_acked, rtt_us); | |
3054 | } | |
3055 | } | |
3056 | ||
3057 | #if FASTRETRANS_DEBUG > 0 | |
3058 | WARN_ON((int)tp->sacked_out < 0); | |
3059 | WARN_ON((int)tp->lost_out < 0); | |
3060 | WARN_ON((int)tp->retrans_out < 0); | |
3061 | if (!tp->packets_out && tcp_is_sack(tp)) { | |
3062 | icsk = inet_csk(sk); | |
3063 | if (tp->lost_out) { | |
3064 | printk(KERN_DEBUG "Leak l=%u %d\n", | |
3065 | tp->lost_out, icsk->icsk_ca_state); | |
3066 | tp->lost_out = 0; | |
3067 | } | |
3068 | if (tp->sacked_out) { | |
3069 | printk(KERN_DEBUG "Leak s=%u %d\n", | |
3070 | tp->sacked_out, icsk->icsk_ca_state); | |
3071 | tp->sacked_out = 0; | |
3072 | } | |
3073 | if (tp->retrans_out) { | |
3074 | printk(KERN_DEBUG "Leak r=%u %d\n", | |
3075 | tp->retrans_out, icsk->icsk_ca_state); | |
3076 | tp->retrans_out = 0; | |
3077 | } | |
3078 | } | |
3079 | #endif | |
3080 | return flag; | |
3081 | } | |
3082 | ||
3083 | static void tcp_ack_probe(struct sock *sk) | |
3084 | { | |
3085 | const struct tcp_sock *tp = tcp_sk(sk); | |
3086 | struct inet_connection_sock *icsk = inet_csk(sk); | |
3087 | ||
3088 | /* Was it a usable window open? */ | |
3089 | ||
3090 | if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) { | |
3091 | icsk->icsk_backoff = 0; | |
3092 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); | |
3093 | /* Socket must be waked up by subsequent tcp_data_snd_check(). | |
3094 | * This function is not for random using! | |
3095 | */ | |
3096 | } else { | |
3097 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, | |
3098 | min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), | |
3099 | TCP_RTO_MAX); | |
3100 | } | |
3101 | } | |
3102 | ||
3103 | static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag) | |
3104 | { | |
3105 | return (!(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || | |
3106 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open); | |
3107 | } | |
3108 | ||
3109 | static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag) | |
3110 | { | |
3111 | const struct tcp_sock *tp = tcp_sk(sk); | |
3112 | return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && | |
3113 | !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR)); | |
3114 | } | |
3115 | ||
3116 | /* Check that window update is acceptable. | |
3117 | * The function assumes that snd_una<=ack<=snd_next. | |
3118 | */ | |
3119 | static inline int tcp_may_update_window(const struct tcp_sock *tp, | |
3120 | const u32 ack, const u32 ack_seq, | |
3121 | const u32 nwin) | |
3122 | { | |
3123 | return (after(ack, tp->snd_una) || | |
3124 | after(ack_seq, tp->snd_wl1) || | |
3125 | (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd)); | |
3126 | } | |
3127 | ||
3128 | /* Update our send window. | |
3129 | * | |
3130 | * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 | |
3131 | * and in FreeBSD. NetBSD's one is even worse.) is wrong. | |
3132 | */ | |
3133 | static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack, | |
3134 | u32 ack_seq) | |
3135 | { | |
3136 | struct tcp_sock *tp = tcp_sk(sk); | |
3137 | int flag = 0; | |
3138 | u32 nwin = ntohs(tcp_hdr(skb)->window); | |
3139 | ||
3140 | if (likely(!tcp_hdr(skb)->syn)) | |
3141 | nwin <<= tp->rx_opt.snd_wscale; | |
3142 | ||
3143 | if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { | |
3144 | flag |= FLAG_WIN_UPDATE; | |
3145 | tcp_update_wl(tp, ack, ack_seq); | |
3146 | ||
3147 | if (tp->snd_wnd != nwin) { | |
3148 | tp->snd_wnd = nwin; | |
3149 | ||
3150 | /* Note, it is the only place, where | |
3151 | * fast path is recovered for sending TCP. | |
3152 | */ | |
3153 | tp->pred_flags = 0; | |
3154 | tcp_fast_path_check(sk); | |
3155 | ||
3156 | if (nwin > tp->max_window) { | |
3157 | tp->max_window = nwin; | |
3158 | tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); | |
3159 | } | |
3160 | } | |
3161 | } | |
3162 | ||
3163 | tp->snd_una = ack; | |
3164 | ||
3165 | return flag; | |
3166 | } | |
3167 | ||
3168 | /* A very conservative spurious RTO response algorithm: reduce cwnd and | |
3169 | * continue in congestion avoidance. | |
3170 | */ | |
3171 | static void tcp_conservative_spur_to_response(struct tcp_sock *tp) | |
3172 | { | |
3173 | tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); | |
3174 | tp->snd_cwnd_cnt = 0; | |
3175 | tp->bytes_acked = 0; | |
3176 | TCP_ECN_queue_cwr(tp); | |
3177 | tcp_moderate_cwnd(tp); | |
3178 | } | |
3179 | ||
3180 | /* A conservative spurious RTO response algorithm: reduce cwnd using | |
3181 | * rate halving and continue in congestion avoidance. | |
3182 | */ | |
3183 | static void tcp_ratehalving_spur_to_response(struct sock *sk) | |
3184 | { | |
3185 | tcp_enter_cwr(sk, 0); | |
3186 | } | |
3187 | ||
3188 | static void tcp_undo_spur_to_response(struct sock *sk, int flag) | |
3189 | { | |
3190 | if (flag & FLAG_ECE) | |
3191 | tcp_ratehalving_spur_to_response(sk); | |
3192 | else | |
3193 | tcp_undo_cwr(sk, 1); | |
3194 | } | |
3195 | ||
3196 | /* F-RTO spurious RTO detection algorithm (RFC4138) | |
3197 | * | |
3198 | * F-RTO affects during two new ACKs following RTO (well, almost, see inline | |
3199 | * comments). State (ACK number) is kept in frto_counter. When ACK advances | |
3200 | * window (but not to or beyond highest sequence sent before RTO): | |
3201 | * On First ACK, send two new segments out. | |
3202 | * On Second ACK, RTO was likely spurious. Do spurious response (response | |
3203 | * algorithm is not part of the F-RTO detection algorithm | |
3204 | * given in RFC4138 but can be selected separately). | |
3205 | * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss | |
3206 | * and TCP falls back to conventional RTO recovery. F-RTO allows overriding | |
3207 | * of Nagle, this is done using frto_counter states 2 and 3, when a new data | |
3208 | * segment of any size sent during F-RTO, state 2 is upgraded to 3. | |
3209 | * | |
3210 | * Rationale: if the RTO was spurious, new ACKs should arrive from the | |
3211 | * original window even after we transmit two new data segments. | |
3212 | * | |
3213 | * SACK version: | |
3214 | * on first step, wait until first cumulative ACK arrives, then move to | |
3215 | * the second step. In second step, the next ACK decides. | |
3216 | * | |
3217 | * F-RTO is implemented (mainly) in four functions: | |
3218 | * - tcp_use_frto() is used to determine if TCP is can use F-RTO | |
3219 | * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is | |
3220 | * called when tcp_use_frto() showed green light | |
3221 | * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm | |
3222 | * - tcp_enter_frto_loss() is called if there is not enough evidence | |
3223 | * to prove that the RTO is indeed spurious. It transfers the control | |
3224 | * from F-RTO to the conventional RTO recovery | |
3225 | */ | |
3226 | static int tcp_process_frto(struct sock *sk, int flag) | |
3227 | { | |
3228 | struct tcp_sock *tp = tcp_sk(sk); | |
3229 | ||
3230 | tcp_verify_left_out(tp); | |
3231 | ||
3232 | /* Duplicate the behavior from Loss state (fastretrans_alert) */ | |
3233 | if (flag & FLAG_DATA_ACKED) | |
3234 | inet_csk(sk)->icsk_retransmits = 0; | |
3235 | ||
3236 | if ((flag & FLAG_NONHEAD_RETRANS_ACKED) || | |
3237 | ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED))) | |
3238 | tp->undo_marker = 0; | |
3239 | ||
3240 | if (!before(tp->snd_una, tp->frto_highmark)) { | |
3241 | tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag); | |
3242 | return 1; | |
3243 | } | |
3244 | ||
3245 | if (!tcp_is_sackfrto(tp)) { | |
3246 | /* RFC4138 shortcoming in step 2; should also have case c): | |
3247 | * ACK isn't duplicate nor advances window, e.g., opposite dir | |
3248 | * data, winupdate | |
3249 | */ | |
3250 | if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP)) | |
3251 | return 1; | |
3252 | ||
3253 | if (!(flag & FLAG_DATA_ACKED)) { | |
3254 | tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3), | |
3255 | flag); | |
3256 | return 1; | |
3257 | } | |
3258 | } else { | |
3259 | if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) { | |
3260 | /* Prevent sending of new data. */ | |
3261 | tp->snd_cwnd = min(tp->snd_cwnd, | |
3262 | tcp_packets_in_flight(tp)); | |
3263 | return 1; | |
3264 | } | |
3265 | ||
3266 | if ((tp->frto_counter >= 2) && | |
3267 | (!(flag & FLAG_FORWARD_PROGRESS) || | |
3268 | ((flag & FLAG_DATA_SACKED) && | |
3269 | !(flag & FLAG_ONLY_ORIG_SACKED)))) { | |
3270 | /* RFC4138 shortcoming (see comment above) */ | |
3271 | if (!(flag & FLAG_FORWARD_PROGRESS) && | |
3272 | (flag & FLAG_NOT_DUP)) | |
3273 | return 1; | |
3274 | ||
3275 | tcp_enter_frto_loss(sk, 3, flag); | |
3276 | return 1; | |
3277 | } | |
3278 | } | |
3279 | ||
3280 | if (tp->frto_counter == 1) { | |
3281 | /* tcp_may_send_now needs to see updated state */ | |
3282 | tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; | |
3283 | tp->frto_counter = 2; | |
3284 | ||
3285 | if (!tcp_may_send_now(sk)) | |
3286 | tcp_enter_frto_loss(sk, 2, flag); | |
3287 | ||
3288 | return 1; | |
3289 | } else { | |
3290 | switch (sysctl_tcp_frto_response) { | |
3291 | case 2: | |
3292 | tcp_undo_spur_to_response(sk, flag); | |
3293 | break; | |
3294 | case 1: | |
3295 | tcp_conservative_spur_to_response(tp); | |
3296 | break; | |
3297 | default: | |
3298 | tcp_ratehalving_spur_to_response(sk); | |
3299 | break; | |
3300 | } | |
3301 | tp->frto_counter = 0; | |
3302 | tp->undo_marker = 0; | |
3303 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS); | |
3304 | } | |
3305 | return 0; | |
3306 | } | |
3307 | ||
3308 | /* This routine deals with incoming acks, but not outgoing ones. */ | |
3309 | static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag) | |
3310 | { | |
3311 | struct inet_connection_sock *icsk = inet_csk(sk); | |
3312 | struct tcp_sock *tp = tcp_sk(sk); | |
3313 | u32 prior_snd_una = tp->snd_una; | |
3314 | u32 ack_seq = TCP_SKB_CB(skb)->seq; | |
3315 | u32 ack = TCP_SKB_CB(skb)->ack_seq; | |
3316 | u32 prior_in_flight; | |
3317 | u32 prior_fackets; | |
3318 | int prior_packets; | |
3319 | int frto_cwnd = 0; | |
3320 | ||
3321 | /* If the ack is newer than sent or older than previous acks | |
3322 | * then we can probably ignore it. | |
3323 | */ | |
3324 | if (after(ack, tp->snd_nxt)) | |
3325 | goto uninteresting_ack; | |
3326 | ||
3327 | if (before(ack, prior_snd_una)) | |
3328 | goto old_ack; | |
3329 | ||
3330 | if (after(ack, prior_snd_una)) | |
3331 | flag |= FLAG_SND_UNA_ADVANCED; | |
3332 | ||
3333 | if (sysctl_tcp_abc) { | |
3334 | if (icsk->icsk_ca_state < TCP_CA_CWR) | |
3335 | tp->bytes_acked += ack - prior_snd_una; | |
3336 | else if (icsk->icsk_ca_state == TCP_CA_Loss) | |
3337 | /* we assume just one segment left network */ | |
3338 | tp->bytes_acked += min(ack - prior_snd_una, | |
3339 | tp->mss_cache); | |
3340 | } | |
3341 | ||
3342 | prior_fackets = tp->fackets_out; | |
3343 | prior_in_flight = tcp_packets_in_flight(tp); | |
3344 | ||
3345 | if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) { | |
3346 | /* Window is constant, pure forward advance. | |
3347 | * No more checks are required. | |
3348 | * Note, we use the fact that SND.UNA>=SND.WL2. | |
3349 | */ | |
3350 | tcp_update_wl(tp, ack, ack_seq); | |
3351 | tp->snd_una = ack; | |
3352 | flag |= FLAG_WIN_UPDATE; | |
3353 | ||
3354 | tcp_ca_event(sk, CA_EVENT_FAST_ACK); | |
3355 | ||
3356 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS); | |
3357 | } else { | |
3358 | if (ack_seq != TCP_SKB_CB(skb)->end_seq) | |
3359 | flag |= FLAG_DATA; | |
3360 | else | |
3361 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS); | |
3362 | ||
3363 | flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); | |
3364 | ||
3365 | if (TCP_SKB_CB(skb)->sacked) | |
3366 | flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); | |
3367 | ||
3368 | if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb))) | |
3369 | flag |= FLAG_ECE; | |
3370 | ||
3371 | tcp_ca_event(sk, CA_EVENT_SLOW_ACK); | |
3372 | } | |
3373 | ||
3374 | /* We passed data and got it acked, remove any soft error | |
3375 | * log. Something worked... | |
3376 | */ | |
3377 | sk->sk_err_soft = 0; | |
3378 | icsk->icsk_probes_out = 0; | |
3379 | tp->rcv_tstamp = tcp_time_stamp; | |
3380 | prior_packets = tp->packets_out; | |
3381 | if (!prior_packets) | |
3382 | goto no_queue; | |
3383 | ||
3384 | /* See if we can take anything off of the retransmit queue. */ | |
3385 | flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una); | |
3386 | ||
3387 | if (tp->frto_counter) | |
3388 | frto_cwnd = tcp_process_frto(sk, flag); | |
3389 | /* Guarantee sacktag reordering detection against wrap-arounds */ | |
3390 | if (before(tp->frto_highmark, tp->snd_una)) | |
3391 | tp->frto_highmark = 0; | |
3392 | ||
3393 | if (tcp_ack_is_dubious(sk, flag)) { | |
3394 | /* Advance CWND, if state allows this. */ | |
3395 | if ((flag & FLAG_DATA_ACKED) && !frto_cwnd && | |
3396 | tcp_may_raise_cwnd(sk, flag)) | |
3397 | tcp_cong_avoid(sk, ack, prior_in_flight); | |
3398 | tcp_fastretrans_alert(sk, prior_packets - tp->packets_out, | |
3399 | flag); | |
3400 | } else { | |
3401 | if ((flag & FLAG_DATA_ACKED) && !frto_cwnd) | |
3402 | tcp_cong_avoid(sk, ack, prior_in_flight); | |
3403 | } | |
3404 | ||
3405 | if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) | |
3406 | dst_confirm(sk->sk_dst_cache); | |
3407 | ||
3408 | return 1; | |
3409 | ||
3410 | no_queue: | |
3411 | /* If this ack opens up a zero window, clear backoff. It was | |
3412 | * being used to time the probes, and is probably far higher than | |
3413 | * it needs to be for normal retransmission. | |
3414 | */ | |
3415 | if (tcp_send_head(sk)) | |
3416 | tcp_ack_probe(sk); | |
3417 | return 1; | |
3418 | ||
3419 | old_ack: | |
3420 | if (TCP_SKB_CB(skb)->sacked) { | |
3421 | tcp_sacktag_write_queue(sk, skb, prior_snd_una); | |
3422 | if (icsk->icsk_ca_state == TCP_CA_Open) | |
3423 | tcp_try_keep_open(sk); | |
3424 | } | |
3425 | ||
3426 | uninteresting_ack: | |
3427 | SOCK_DEBUG(sk, "Ack %u out of %u:%u\n", ack, tp->snd_una, tp->snd_nxt); | |
3428 | return 0; | |
3429 | } | |
3430 | ||
3431 | /* Look for tcp options. Normally only called on SYN and SYNACK packets. | |
3432 | * But, this can also be called on packets in the established flow when | |
3433 | * the fast version below fails. | |
3434 | */ | |
3435 | void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, | |
3436 | int estab) | |
3437 | { | |
3438 | unsigned char *ptr; | |
3439 | struct tcphdr *th = tcp_hdr(skb); | |
3440 | int length = (th->doff * 4) - sizeof(struct tcphdr); | |
3441 | ||
3442 | ptr = (unsigned char *)(th + 1); | |
3443 | opt_rx->saw_tstamp = 0; | |
3444 | ||
3445 | while (length > 0) { | |
3446 | int opcode = *ptr++; | |
3447 | int opsize; | |
3448 | ||
3449 | switch (opcode) { | |
3450 | case TCPOPT_EOL: | |
3451 | return; | |
3452 | case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ | |
3453 | length--; | |
3454 | continue; | |
3455 | default: | |
3456 | opsize = *ptr++; | |
3457 | if (opsize < 2) /* "silly options" */ | |
3458 | return; | |
3459 | if (opsize > length) | |
3460 | return; /* don't parse partial options */ | |
3461 | switch (opcode) { | |
3462 | case TCPOPT_MSS: | |
3463 | if (opsize == TCPOLEN_MSS && th->syn && !estab) { | |
3464 | u16 in_mss = get_unaligned_be16(ptr); | |
3465 | if (in_mss) { | |
3466 | if (opt_rx->user_mss && | |
3467 | opt_rx->user_mss < in_mss) | |
3468 | in_mss = opt_rx->user_mss; | |
3469 | opt_rx->mss_clamp = in_mss; | |
3470 | } | |
3471 | } | |
3472 | break; | |
3473 | case TCPOPT_WINDOW: | |
3474 | if (opsize == TCPOLEN_WINDOW && th->syn && | |
3475 | !estab && sysctl_tcp_window_scaling) { | |
3476 | __u8 snd_wscale = *(__u8 *)ptr; | |
3477 | opt_rx->wscale_ok = 1; | |
3478 | if (snd_wscale > 14) { | |
3479 | if (net_ratelimit()) | |
3480 | printk(KERN_INFO "tcp_parse_options: Illegal window " | |
3481 | "scaling value %d >14 received.\n", | |
3482 | snd_wscale); | |
3483 | snd_wscale = 14; | |
3484 | } | |
3485 | opt_rx->snd_wscale = snd_wscale; | |
3486 | } | |
3487 | break; | |
3488 | case TCPOPT_TIMESTAMP: | |
3489 | if ((opsize == TCPOLEN_TIMESTAMP) && | |
3490 | ((estab && opt_rx->tstamp_ok) || | |
3491 | (!estab && sysctl_tcp_timestamps))) { | |
3492 | opt_rx->saw_tstamp = 1; | |
3493 | opt_rx->rcv_tsval = get_unaligned_be32(ptr); | |
3494 | opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4); | |
3495 | } | |
3496 | break; | |
3497 | case TCPOPT_SACK_PERM: | |
3498 | if (opsize == TCPOLEN_SACK_PERM && th->syn && | |
3499 | !estab && sysctl_tcp_sack) { | |
3500 | opt_rx->sack_ok = 1; | |
3501 | tcp_sack_reset(opt_rx); | |
3502 | } | |
3503 | break; | |
3504 | ||
3505 | case TCPOPT_SACK: | |
3506 | if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && | |
3507 | !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && | |
3508 | opt_rx->sack_ok) { | |
3509 | TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; | |
3510 | } | |
3511 | break; | |
3512 | #ifdef CONFIG_TCP_MD5SIG | |
3513 | case TCPOPT_MD5SIG: | |
3514 | /* | |
3515 | * The MD5 Hash has already been | |
3516 | * checked (see tcp_v{4,6}_do_rcv()). | |
3517 | */ | |
3518 | break; | |
3519 | #endif | |
3520 | } | |
3521 | ||
3522 | ptr += opsize-2; | |
3523 | length -= opsize; | |
3524 | } | |
3525 | } | |
3526 | } | |
3527 | ||
3528 | static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, struct tcphdr *th) | |
3529 | { | |
3530 | __be32 *ptr = (__be32 *)(th + 1); | |
3531 | ||
3532 | if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | |
3533 | | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { | |
3534 | tp->rx_opt.saw_tstamp = 1; | |
3535 | ++ptr; | |
3536 | tp->rx_opt.rcv_tsval = ntohl(*ptr); | |
3537 | ++ptr; | |
3538 | tp->rx_opt.rcv_tsecr = ntohl(*ptr); | |
3539 | return 1; | |
3540 | } | |
3541 | return 0; | |
3542 | } | |
3543 | ||
3544 | /* Fast parse options. This hopes to only see timestamps. | |
3545 | * If it is wrong it falls back on tcp_parse_options(). | |
3546 | */ | |
3547 | static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, | |
3548 | struct tcp_sock *tp) | |
3549 | { | |
3550 | if (th->doff == sizeof(struct tcphdr) >> 2) { | |
3551 | tp->rx_opt.saw_tstamp = 0; | |
3552 | return 0; | |
3553 | } else if (tp->rx_opt.tstamp_ok && | |
3554 | th->doff == (sizeof(struct tcphdr)>>2)+(TCPOLEN_TSTAMP_ALIGNED>>2)) { | |
3555 | if (tcp_parse_aligned_timestamp(tp, th)) | |
3556 | return 1; | |
3557 | } | |
3558 | tcp_parse_options(skb, &tp->rx_opt, 1); | |
3559 | return 1; | |
3560 | } | |
3561 | ||
3562 | #ifdef CONFIG_TCP_MD5SIG | |
3563 | /* | |
3564 | * Parse MD5 Signature option | |
3565 | */ | |
3566 | u8 *tcp_parse_md5sig_option(struct tcphdr *th) | |
3567 | { | |
3568 | int length = (th->doff << 2) - sizeof (*th); | |
3569 | u8 *ptr = (u8*)(th + 1); | |
3570 | ||
3571 | /* If the TCP option is too short, we can short cut */ | |
3572 | if (length < TCPOLEN_MD5SIG) | |
3573 | return NULL; | |
3574 | ||
3575 | while (length > 0) { | |
3576 | int opcode = *ptr++; | |
3577 | int opsize; | |
3578 | ||
3579 | switch(opcode) { | |
3580 | case TCPOPT_EOL: | |
3581 | return NULL; | |
3582 | case TCPOPT_NOP: | |
3583 | length--; | |
3584 | continue; | |
3585 | default: | |
3586 | opsize = *ptr++; | |
3587 | if (opsize < 2 || opsize > length) | |
3588 | return NULL; | |
3589 | if (opcode == TCPOPT_MD5SIG) | |
3590 | return ptr; | |
3591 | } | |
3592 | ptr += opsize - 2; | |
3593 | length -= opsize; | |
3594 | } | |
3595 | return NULL; | |
3596 | } | |
3597 | #endif | |
3598 | ||
3599 | static inline void tcp_store_ts_recent(struct tcp_sock *tp) | |
3600 | { | |
3601 | tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; | |
3602 | tp->rx_opt.ts_recent_stamp = get_seconds(); | |
3603 | } | |
3604 | ||
3605 | static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) | |
3606 | { | |
3607 | if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { | |
3608 | /* PAWS bug workaround wrt. ACK frames, the PAWS discard | |
3609 | * extra check below makes sure this can only happen | |
3610 | * for pure ACK frames. -DaveM | |
3611 | * | |
3612 | * Not only, also it occurs for expired timestamps. | |
3613 | */ | |
3614 | ||
3615 | if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) >= 0 || | |
3616 | get_seconds() >= tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS) | |
3617 | tcp_store_ts_recent(tp); | |
3618 | } | |
3619 | } | |
3620 | ||
3621 | /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM | |
3622 | * | |
3623 | * It is not fatal. If this ACK does _not_ change critical state (seqs, window) | |
3624 | * it can pass through stack. So, the following predicate verifies that | |
3625 | * this segment is not used for anything but congestion avoidance or | |
3626 | * fast retransmit. Moreover, we even are able to eliminate most of such | |
3627 | * second order effects, if we apply some small "replay" window (~RTO) | |
3628 | * to timestamp space. | |
3629 | * | |
3630 | * All these measures still do not guarantee that we reject wrapped ACKs | |
3631 | * on networks with high bandwidth, when sequence space is recycled fastly, | |
3632 | * but it guarantees that such events will be very rare and do not affect | |
3633 | * connection seriously. This doesn't look nice, but alas, PAWS is really | |
3634 | * buggy extension. | |
3635 | * | |
3636 | * [ Later note. Even worse! It is buggy for segments _with_ data. RFC | |
3637 | * states that events when retransmit arrives after original data are rare. | |
3638 | * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is | |
3639 | * the biggest problem on large power networks even with minor reordering. | |
3640 | * OK, let's give it small replay window. If peer clock is even 1hz, it is safe | |
3641 | * up to bandwidth of 18Gigabit/sec. 8) ] | |
3642 | */ | |
3643 | ||
3644 | static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) | |
3645 | { | |
3646 | struct tcp_sock *tp = tcp_sk(sk); | |
3647 | struct tcphdr *th = tcp_hdr(skb); | |
3648 | u32 seq = TCP_SKB_CB(skb)->seq; | |
3649 | u32 ack = TCP_SKB_CB(skb)->ack_seq; | |
3650 | ||
3651 | return (/* 1. Pure ACK with correct sequence number. */ | |
3652 | (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && | |
3653 | ||
3654 | /* 2. ... and duplicate ACK. */ | |
3655 | ack == tp->snd_una && | |
3656 | ||
3657 | /* 3. ... and does not update window. */ | |
3658 | !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && | |
3659 | ||
3660 | /* 4. ... and sits in replay window. */ | |
3661 | (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); | |
3662 | } | |
3663 | ||
3664 | static inline int tcp_paws_discard(const struct sock *sk, | |
3665 | const struct sk_buff *skb) | |
3666 | { | |
3667 | const struct tcp_sock *tp = tcp_sk(sk); | |
3668 | return ((s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) > TCP_PAWS_WINDOW && | |
3669 | get_seconds() < tp->rx_opt.ts_recent_stamp + TCP_PAWS_24DAYS && | |
3670 | !tcp_disordered_ack(sk, skb)); | |
3671 | } | |
3672 | ||
3673 | /* Check segment sequence number for validity. | |
3674 | * | |
3675 | * Segment controls are considered valid, if the segment | |
3676 | * fits to the window after truncation to the window. Acceptability | |
3677 | * of data (and SYN, FIN, of course) is checked separately. | |
3678 | * See tcp_data_queue(), for example. | |
3679 | * | |
3680 | * Also, controls (RST is main one) are accepted using RCV.WUP instead | |
3681 | * of RCV.NXT. Peer still did not advance his SND.UNA when we | |
3682 | * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. | |
3683 | * (borrowed from freebsd) | |
3684 | */ | |
3685 | ||
3686 | static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq) | |
3687 | { | |
3688 | return !before(end_seq, tp->rcv_wup) && | |
3689 | !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); | |
3690 | } | |
3691 | ||
3692 | /* When we get a reset we do this. */ | |
3693 | static void tcp_reset(struct sock *sk) | |
3694 | { | |
3695 | /* We want the right error as BSD sees it (and indeed as we do). */ | |
3696 | switch (sk->sk_state) { | |
3697 | case TCP_SYN_SENT: | |
3698 | sk->sk_err = ECONNREFUSED; | |
3699 | break; | |
3700 | case TCP_CLOSE_WAIT: | |
3701 | sk->sk_err = EPIPE; | |
3702 | break; | |
3703 | case TCP_CLOSE: | |
3704 | return; | |
3705 | default: | |
3706 | sk->sk_err = ECONNRESET; | |
3707 | } | |
3708 | ||
3709 | if (!sock_flag(sk, SOCK_DEAD)) | |
3710 | sk->sk_error_report(sk); | |
3711 | ||
3712 | tcp_done(sk); | |
3713 | } | |
3714 | ||
3715 | /* | |
3716 | * Process the FIN bit. This now behaves as it is supposed to work | |
3717 | * and the FIN takes effect when it is validly part of sequence | |
3718 | * space. Not before when we get holes. | |
3719 | * | |
3720 | * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT | |
3721 | * (and thence onto LAST-ACK and finally, CLOSE, we never enter | |
3722 | * TIME-WAIT) | |
3723 | * | |
3724 | * If we are in FINWAIT-1, a received FIN indicates simultaneous | |
3725 | * close and we go into CLOSING (and later onto TIME-WAIT) | |
3726 | * | |
3727 | * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. | |
3728 | */ | |
3729 | static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th) | |
3730 | { | |
3731 | struct tcp_sock *tp = tcp_sk(sk); | |
3732 | ||
3733 | inet_csk_schedule_ack(sk); | |
3734 | ||
3735 | sk->sk_shutdown |= RCV_SHUTDOWN; | |
3736 | sock_set_flag(sk, SOCK_DONE); | |
3737 | ||
3738 | switch (sk->sk_state) { | |
3739 | case TCP_SYN_RECV: | |
3740 | case TCP_ESTABLISHED: | |
3741 | /* Move to CLOSE_WAIT */ | |
3742 | tcp_set_state(sk, TCP_CLOSE_WAIT); | |
3743 | inet_csk(sk)->icsk_ack.pingpong = 1; | |
3744 | break; | |
3745 | ||
3746 | case TCP_CLOSE_WAIT: | |
3747 | case TCP_CLOSING: | |
3748 | /* Received a retransmission of the FIN, do | |
3749 | * nothing. | |
3750 | */ | |
3751 | break; | |
3752 | case TCP_LAST_ACK: | |
3753 | /* RFC793: Remain in the LAST-ACK state. */ | |
3754 | break; | |
3755 | ||
3756 | case TCP_FIN_WAIT1: | |
3757 | /* This case occurs when a simultaneous close | |
3758 | * happens, we must ack the received FIN and | |
3759 | * enter the CLOSING state. | |
3760 | */ | |
3761 | tcp_send_ack(sk); | |
3762 | tcp_set_state(sk, TCP_CLOSING); | |
3763 | break; | |
3764 | case TCP_FIN_WAIT2: | |
3765 | /* Received a FIN -- send ACK and enter TIME_WAIT. */ | |
3766 | tcp_send_ack(sk); | |
3767 | tcp_time_wait(sk, TCP_TIME_WAIT, 0); | |
3768 | break; | |
3769 | default: | |
3770 | /* Only TCP_LISTEN and TCP_CLOSE are left, in these | |
3771 | * cases we should never reach this piece of code. | |
3772 | */ | |
3773 | printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n", | |
3774 | __func__, sk->sk_state); | |
3775 | break; | |
3776 | } | |
3777 | ||
3778 | /* It _is_ possible, that we have something out-of-order _after_ FIN. | |
3779 | * Probably, we should reset in this case. For now drop them. | |
3780 | */ | |
3781 | __skb_queue_purge(&tp->out_of_order_queue); | |
3782 | if (tcp_is_sack(tp)) | |
3783 | tcp_sack_reset(&tp->rx_opt); | |
3784 | sk_mem_reclaim(sk); | |
3785 | ||
3786 | if (!sock_flag(sk, SOCK_DEAD)) { | |
3787 | sk->sk_state_change(sk); | |
3788 | ||
3789 | /* Do not send POLL_HUP for half duplex close. */ | |
3790 | if (sk->sk_shutdown == SHUTDOWN_MASK || | |
3791 | sk->sk_state == TCP_CLOSE) | |
3792 | sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); | |
3793 | else | |
3794 | sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); | |
3795 | } | |
3796 | } | |
3797 | ||
3798 | static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, | |
3799 | u32 end_seq) | |
3800 | { | |
3801 | if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { | |
3802 | if (before(seq, sp->start_seq)) | |
3803 | sp->start_seq = seq; | |
3804 | if (after(end_seq, sp->end_seq)) | |
3805 | sp->end_seq = end_seq; | |
3806 | return 1; | |
3807 | } | |
3808 | return 0; | |
3809 | } | |
3810 | ||
3811 | static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq) | |
3812 | { | |
3813 | struct tcp_sock *tp = tcp_sk(sk); | |
3814 | ||
3815 | if (tcp_is_sack(tp) && sysctl_tcp_dsack) { | |
3816 | int mib_idx; | |
3817 | ||
3818 | if (before(seq, tp->rcv_nxt)) | |
3819 | mib_idx = LINUX_MIB_TCPDSACKOLDSENT; | |
3820 | else | |
3821 | mib_idx = LINUX_MIB_TCPDSACKOFOSENT; | |
3822 | ||
3823 | NET_INC_STATS_BH(sock_net(sk), mib_idx); | |
3824 | ||
3825 | tp->rx_opt.dsack = 1; | |
3826 | tp->duplicate_sack[0].start_seq = seq; | |
3827 | tp->duplicate_sack[0].end_seq = end_seq; | |
3828 | tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks + 1; | |
3829 | } | |
3830 | } | |
3831 | ||
3832 | static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq) | |
3833 | { | |
3834 | struct tcp_sock *tp = tcp_sk(sk); | |
3835 | ||
3836 | if (!tp->rx_opt.dsack) | |
3837 | tcp_dsack_set(sk, seq, end_seq); | |
3838 | else | |
3839 | tcp_sack_extend(tp->duplicate_sack, seq, end_seq); | |
3840 | } | |
3841 | ||
3842 | static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb) | |
3843 | { | |
3844 | struct tcp_sock *tp = tcp_sk(sk); | |
3845 | ||
3846 | if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && | |
3847 | before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
3848 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); | |
3849 | tcp_enter_quickack_mode(sk); | |
3850 | ||
3851 | if (tcp_is_sack(tp) && sysctl_tcp_dsack) { | |
3852 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
3853 | ||
3854 | if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) | |
3855 | end_seq = tp->rcv_nxt; | |
3856 | tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq); | |
3857 | } | |
3858 | } | |
3859 | ||
3860 | tcp_send_ack(sk); | |
3861 | } | |
3862 | ||
3863 | /* These routines update the SACK block as out-of-order packets arrive or | |
3864 | * in-order packets close up the sequence space. | |
3865 | */ | |
3866 | static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) | |
3867 | { | |
3868 | int this_sack; | |
3869 | struct tcp_sack_block *sp = &tp->selective_acks[0]; | |
3870 | struct tcp_sack_block *swalk = sp + 1; | |
3871 | ||
3872 | /* See if the recent change to the first SACK eats into | |
3873 | * or hits the sequence space of other SACK blocks, if so coalesce. | |
3874 | */ | |
3875 | for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { | |
3876 | if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { | |
3877 | int i; | |
3878 | ||
3879 | /* Zap SWALK, by moving every further SACK up by one slot. | |
3880 | * Decrease num_sacks. | |
3881 | */ | |
3882 | tp->rx_opt.num_sacks--; | |
3883 | tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks + | |
3884 | tp->rx_opt.dsack; | |
3885 | for (i = this_sack; i < tp->rx_opt.num_sacks; i++) | |
3886 | sp[i] = sp[i + 1]; | |
3887 | continue; | |
3888 | } | |
3889 | this_sack++, swalk++; | |
3890 | } | |
3891 | } | |
3892 | ||
3893 | static inline void tcp_sack_swap(struct tcp_sack_block *sack1, | |
3894 | struct tcp_sack_block *sack2) | |
3895 | { | |
3896 | __u32 tmp; | |
3897 | ||
3898 | tmp = sack1->start_seq; | |
3899 | sack1->start_seq = sack2->start_seq; | |
3900 | sack2->start_seq = tmp; | |
3901 | ||
3902 | tmp = sack1->end_seq; | |
3903 | sack1->end_seq = sack2->end_seq; | |
3904 | sack2->end_seq = tmp; | |
3905 | } | |
3906 | ||
3907 | static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) | |
3908 | { | |
3909 | struct tcp_sock *tp = tcp_sk(sk); | |
3910 | struct tcp_sack_block *sp = &tp->selective_acks[0]; | |
3911 | int cur_sacks = tp->rx_opt.num_sacks; | |
3912 | int this_sack; | |
3913 | ||
3914 | if (!cur_sacks) | |
3915 | goto new_sack; | |
3916 | ||
3917 | for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { | |
3918 | if (tcp_sack_extend(sp, seq, end_seq)) { | |
3919 | /* Rotate this_sack to the first one. */ | |
3920 | for (; this_sack > 0; this_sack--, sp--) | |
3921 | tcp_sack_swap(sp, sp - 1); | |
3922 | if (cur_sacks > 1) | |
3923 | tcp_sack_maybe_coalesce(tp); | |
3924 | return; | |
3925 | } | |
3926 | } | |
3927 | ||
3928 | /* Could not find an adjacent existing SACK, build a new one, | |
3929 | * put it at the front, and shift everyone else down. We | |
3930 | * always know there is at least one SACK present already here. | |
3931 | * | |
3932 | * If the sack array is full, forget about the last one. | |
3933 | */ | |
3934 | if (this_sack >= TCP_NUM_SACKS) { | |
3935 | this_sack--; | |
3936 | tp->rx_opt.num_sacks--; | |
3937 | sp--; | |
3938 | } | |
3939 | for (; this_sack > 0; this_sack--, sp--) | |
3940 | *sp = *(sp - 1); | |
3941 | ||
3942 | new_sack: | |
3943 | /* Build the new head SACK, and we're done. */ | |
3944 | sp->start_seq = seq; | |
3945 | sp->end_seq = end_seq; | |
3946 | tp->rx_opt.num_sacks++; | |
3947 | tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; | |
3948 | } | |
3949 | ||
3950 | /* RCV.NXT advances, some SACKs should be eaten. */ | |
3951 | ||
3952 | static void tcp_sack_remove(struct tcp_sock *tp) | |
3953 | { | |
3954 | struct tcp_sack_block *sp = &tp->selective_acks[0]; | |
3955 | int num_sacks = tp->rx_opt.num_sacks; | |
3956 | int this_sack; | |
3957 | ||
3958 | /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ | |
3959 | if (skb_queue_empty(&tp->out_of_order_queue)) { | |
3960 | tp->rx_opt.num_sacks = 0; | |
3961 | tp->rx_opt.eff_sacks = tp->rx_opt.dsack; | |
3962 | return; | |
3963 | } | |
3964 | ||
3965 | for (this_sack = 0; this_sack < num_sacks;) { | |
3966 | /* Check if the start of the sack is covered by RCV.NXT. */ | |
3967 | if (!before(tp->rcv_nxt, sp->start_seq)) { | |
3968 | int i; | |
3969 | ||
3970 | /* RCV.NXT must cover all the block! */ | |
3971 | WARN_ON(before(tp->rcv_nxt, sp->end_seq)); | |
3972 | ||
3973 | /* Zap this SACK, by moving forward any other SACKS. */ | |
3974 | for (i=this_sack+1; i < num_sacks; i++) | |
3975 | tp->selective_acks[i-1] = tp->selective_acks[i]; | |
3976 | num_sacks--; | |
3977 | continue; | |
3978 | } | |
3979 | this_sack++; | |
3980 | sp++; | |
3981 | } | |
3982 | if (num_sacks != tp->rx_opt.num_sacks) { | |
3983 | tp->rx_opt.num_sacks = num_sacks; | |
3984 | tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks + | |
3985 | tp->rx_opt.dsack; | |
3986 | } | |
3987 | } | |
3988 | ||
3989 | /* This one checks to see if we can put data from the | |
3990 | * out_of_order queue into the receive_queue. | |
3991 | */ | |
3992 | static void tcp_ofo_queue(struct sock *sk) | |
3993 | { | |
3994 | struct tcp_sock *tp = tcp_sk(sk); | |
3995 | __u32 dsack_high = tp->rcv_nxt; | |
3996 | struct sk_buff *skb; | |
3997 | ||
3998 | while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { | |
3999 | if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) | |
4000 | break; | |
4001 | ||
4002 | if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { | |
4003 | __u32 dsack = dsack_high; | |
4004 | if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) | |
4005 | dsack_high = TCP_SKB_CB(skb)->end_seq; | |
4006 | tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack); | |
4007 | } | |
4008 | ||
4009 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { | |
4010 | SOCK_DEBUG(sk, "ofo packet was already received \n"); | |
4011 | __skb_unlink(skb, &tp->out_of_order_queue); | |
4012 | __kfree_skb(skb); | |
4013 | continue; | |
4014 | } | |
4015 | SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", | |
4016 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, | |
4017 | TCP_SKB_CB(skb)->end_seq); | |
4018 | ||
4019 | __skb_unlink(skb, &tp->out_of_order_queue); | |
4020 | __skb_queue_tail(&sk->sk_receive_queue, skb); | |
4021 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
4022 | if (tcp_hdr(skb)->fin) | |
4023 | tcp_fin(skb, sk, tcp_hdr(skb)); | |
4024 | } | |
4025 | } | |
4026 | ||
4027 | static int tcp_prune_ofo_queue(struct sock *sk); | |
4028 | static int tcp_prune_queue(struct sock *sk); | |
4029 | ||
4030 | static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size) | |
4031 | { | |
4032 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || | |
4033 | !sk_rmem_schedule(sk, size)) { | |
4034 | ||
4035 | if (tcp_prune_queue(sk) < 0) | |
4036 | return -1; | |
4037 | ||
4038 | if (!sk_rmem_schedule(sk, size)) { | |
4039 | if (!tcp_prune_ofo_queue(sk)) | |
4040 | return -1; | |
4041 | ||
4042 | if (!sk_rmem_schedule(sk, size)) | |
4043 | return -1; | |
4044 | } | |
4045 | } | |
4046 | return 0; | |
4047 | } | |
4048 | ||
4049 | static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) | |
4050 | { | |
4051 | struct tcphdr *th = tcp_hdr(skb); | |
4052 | struct tcp_sock *tp = tcp_sk(sk); | |
4053 | int eaten = -1; | |
4054 | ||
4055 | if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) | |
4056 | goto drop; | |
4057 | ||
4058 | __skb_pull(skb, th->doff * 4); | |
4059 | ||
4060 | TCP_ECN_accept_cwr(tp, skb); | |
4061 | ||
4062 | if (tp->rx_opt.dsack) { | |
4063 | tp->rx_opt.dsack = 0; | |
4064 | tp->rx_opt.eff_sacks = tp->rx_opt.num_sacks; | |
4065 | } | |
4066 | ||
4067 | /* Queue data for delivery to the user. | |
4068 | * Packets in sequence go to the receive queue. | |
4069 | * Out of sequence packets to the out_of_order_queue. | |
4070 | */ | |
4071 | if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { | |
4072 | if (tcp_receive_window(tp) == 0) | |
4073 | goto out_of_window; | |
4074 | ||
4075 | /* Ok. In sequence. In window. */ | |
4076 | if (tp->ucopy.task == current && | |
4077 | tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && | |
4078 | sock_owned_by_user(sk) && !tp->urg_data) { | |
4079 | int chunk = min_t(unsigned int, skb->len, | |
4080 | tp->ucopy.len); | |
4081 | ||
4082 | __set_current_state(TASK_RUNNING); | |
4083 | ||
4084 | local_bh_enable(); | |
4085 | if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { | |
4086 | tp->ucopy.len -= chunk; | |
4087 | tp->copied_seq += chunk; | |
4088 | eaten = (chunk == skb->len && !th->fin); | |
4089 | tcp_rcv_space_adjust(sk); | |
4090 | } | |
4091 | local_bh_disable(); | |
4092 | } | |
4093 | ||
4094 | if (eaten <= 0) { | |
4095 | queue_and_out: | |
4096 | if (eaten < 0 && | |
4097 | tcp_try_rmem_schedule(sk, skb->truesize)) | |
4098 | goto drop; | |
4099 | ||
4100 | skb_set_owner_r(skb, sk); | |
4101 | __skb_queue_tail(&sk->sk_receive_queue, skb); | |
4102 | } | |
4103 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
4104 | if (skb->len) | |
4105 | tcp_event_data_recv(sk, skb); | |
4106 | if (th->fin) | |
4107 | tcp_fin(skb, sk, th); | |
4108 | ||
4109 | if (!skb_queue_empty(&tp->out_of_order_queue)) { | |
4110 | tcp_ofo_queue(sk); | |
4111 | ||
4112 | /* RFC2581. 4.2. SHOULD send immediate ACK, when | |
4113 | * gap in queue is filled. | |
4114 | */ | |
4115 | if (skb_queue_empty(&tp->out_of_order_queue)) | |
4116 | inet_csk(sk)->icsk_ack.pingpong = 0; | |
4117 | } | |
4118 | ||
4119 | if (tp->rx_opt.num_sacks) | |
4120 | tcp_sack_remove(tp); | |
4121 | ||
4122 | tcp_fast_path_check(sk); | |
4123 | ||
4124 | if (eaten > 0) | |
4125 | __kfree_skb(skb); | |
4126 | else if (!sock_flag(sk, SOCK_DEAD)) | |
4127 | sk->sk_data_ready(sk, 0); | |
4128 | return; | |
4129 | } | |
4130 | ||
4131 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { | |
4132 | /* A retransmit, 2nd most common case. Force an immediate ack. */ | |
4133 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); | |
4134 | tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); | |
4135 | ||
4136 | out_of_window: | |
4137 | tcp_enter_quickack_mode(sk); | |
4138 | inet_csk_schedule_ack(sk); | |
4139 | drop: | |
4140 | __kfree_skb(skb); | |
4141 | return; | |
4142 | } | |
4143 | ||
4144 | /* Out of window. F.e. zero window probe. */ | |
4145 | if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) | |
4146 | goto out_of_window; | |
4147 | ||
4148 | tcp_enter_quickack_mode(sk); | |
4149 | ||
4150 | if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
4151 | /* Partial packet, seq < rcv_next < end_seq */ | |
4152 | SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", | |
4153 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, | |
4154 | TCP_SKB_CB(skb)->end_seq); | |
4155 | ||
4156 | tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); | |
4157 | ||
4158 | /* If window is closed, drop tail of packet. But after | |
4159 | * remembering D-SACK for its head made in previous line. | |
4160 | */ | |
4161 | if (!tcp_receive_window(tp)) | |
4162 | goto out_of_window; | |
4163 | goto queue_and_out; | |
4164 | } | |
4165 | ||
4166 | TCP_ECN_check_ce(tp, skb); | |
4167 | ||
4168 | if (tcp_try_rmem_schedule(sk, skb->truesize)) | |
4169 | goto drop; | |
4170 | ||
4171 | /* Disable header prediction. */ | |
4172 | tp->pred_flags = 0; | |
4173 | inet_csk_schedule_ack(sk); | |
4174 | ||
4175 | SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", | |
4176 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); | |
4177 | ||
4178 | skb_set_owner_r(skb, sk); | |
4179 | ||
4180 | if (!skb_peek(&tp->out_of_order_queue)) { | |
4181 | /* Initial out of order segment, build 1 SACK. */ | |
4182 | if (tcp_is_sack(tp)) { | |
4183 | tp->rx_opt.num_sacks = 1; | |
4184 | tp->rx_opt.dsack = 0; | |
4185 | tp->rx_opt.eff_sacks = 1; | |
4186 | tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; | |
4187 | tp->selective_acks[0].end_seq = | |
4188 | TCP_SKB_CB(skb)->end_seq; | |
4189 | } | |
4190 | __skb_queue_head(&tp->out_of_order_queue, skb); | |
4191 | } else { | |
4192 | struct sk_buff *skb1 = tp->out_of_order_queue.prev; | |
4193 | u32 seq = TCP_SKB_CB(skb)->seq; | |
4194 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
4195 | ||
4196 | if (seq == TCP_SKB_CB(skb1)->end_seq) { | |
4197 | __skb_queue_after(&tp->out_of_order_queue, skb1, skb); | |
4198 | ||
4199 | if (!tp->rx_opt.num_sacks || | |
4200 | tp->selective_acks[0].end_seq != seq) | |
4201 | goto add_sack; | |
4202 | ||
4203 | /* Common case: data arrive in order after hole. */ | |
4204 | tp->selective_acks[0].end_seq = end_seq; | |
4205 | return; | |
4206 | } | |
4207 | ||
4208 | /* Find place to insert this segment. */ | |
4209 | do { | |
4210 | if (!after(TCP_SKB_CB(skb1)->seq, seq)) | |
4211 | break; | |
4212 | } while ((skb1 = skb1->prev) != | |
4213 | (struct sk_buff *)&tp->out_of_order_queue); | |
4214 | ||
4215 | /* Do skb overlap to previous one? */ | |
4216 | if (skb1 != (struct sk_buff *)&tp->out_of_order_queue && | |
4217 | before(seq, TCP_SKB_CB(skb1)->end_seq)) { | |
4218 | if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { | |
4219 | /* All the bits are present. Drop. */ | |
4220 | __kfree_skb(skb); | |
4221 | tcp_dsack_set(sk, seq, end_seq); | |
4222 | goto add_sack; | |
4223 | } | |
4224 | if (after(seq, TCP_SKB_CB(skb1)->seq)) { | |
4225 | /* Partial overlap. */ | |
4226 | tcp_dsack_set(sk, seq, | |
4227 | TCP_SKB_CB(skb1)->end_seq); | |
4228 | } else { | |
4229 | skb1 = skb1->prev; | |
4230 | } | |
4231 | } | |
4232 | __skb_queue_after(&tp->out_of_order_queue, skb1, skb); | |
4233 | ||
4234 | /* And clean segments covered by new one as whole. */ | |
4235 | while ((skb1 = skb->next) != | |
4236 | (struct sk_buff *)&tp->out_of_order_queue && | |
4237 | after(end_seq, TCP_SKB_CB(skb1)->seq)) { | |
4238 | if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { | |
4239 | tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, | |
4240 | end_seq); | |
4241 | break; | |
4242 | } | |
4243 | __skb_unlink(skb1, &tp->out_of_order_queue); | |
4244 | tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, | |
4245 | TCP_SKB_CB(skb1)->end_seq); | |
4246 | __kfree_skb(skb1); | |
4247 | } | |
4248 | ||
4249 | add_sack: | |
4250 | if (tcp_is_sack(tp)) | |
4251 | tcp_sack_new_ofo_skb(sk, seq, end_seq); | |
4252 | } | |
4253 | } | |
4254 | ||
4255 | static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb, | |
4256 | struct sk_buff_head *list) | |
4257 | { | |
4258 | struct sk_buff *next = skb->next; | |
4259 | ||
4260 | __skb_unlink(skb, list); | |
4261 | __kfree_skb(skb); | |
4262 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED); | |
4263 | ||
4264 | return next; | |
4265 | } | |
4266 | ||
4267 | /* Collapse contiguous sequence of skbs head..tail with | |
4268 | * sequence numbers start..end. | |
4269 | * Segments with FIN/SYN are not collapsed (only because this | |
4270 | * simplifies code) | |
4271 | */ | |
4272 | static void | |
4273 | tcp_collapse(struct sock *sk, struct sk_buff_head *list, | |
4274 | struct sk_buff *head, struct sk_buff *tail, | |
4275 | u32 start, u32 end) | |
4276 | { | |
4277 | struct sk_buff *skb; | |
4278 | ||
4279 | /* First, check that queue is collapsible and find | |
4280 | * the point where collapsing can be useful. */ | |
4281 | for (skb = head; skb != tail;) { | |
4282 | /* No new bits? It is possible on ofo queue. */ | |
4283 | if (!before(start, TCP_SKB_CB(skb)->end_seq)) { | |
4284 | skb = tcp_collapse_one(sk, skb, list); | |
4285 | continue; | |
4286 | } | |
4287 | ||
4288 | /* The first skb to collapse is: | |
4289 | * - not SYN/FIN and | |
4290 | * - bloated or contains data before "start" or | |
4291 | * overlaps to the next one. | |
4292 | */ | |
4293 | if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin && | |
4294 | (tcp_win_from_space(skb->truesize) > skb->len || | |
4295 | before(TCP_SKB_CB(skb)->seq, start) || | |
4296 | (skb->next != tail && | |
4297 | TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb->next)->seq))) | |
4298 | break; | |
4299 | ||
4300 | /* Decided to skip this, advance start seq. */ | |
4301 | start = TCP_SKB_CB(skb)->end_seq; | |
4302 | skb = skb->next; | |
4303 | } | |
4304 | if (skb == tail || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin) | |
4305 | return; | |
4306 | ||
4307 | while (before(start, end)) { | |
4308 | struct sk_buff *nskb; | |
4309 | unsigned int header = skb_headroom(skb); | |
4310 | int copy = SKB_MAX_ORDER(header, 0); | |
4311 | ||
4312 | /* Too big header? This can happen with IPv6. */ | |
4313 | if (copy < 0) | |
4314 | return; | |
4315 | if (end - start < copy) | |
4316 | copy = end - start; | |
4317 | nskb = alloc_skb(copy + header, GFP_ATOMIC); | |
4318 | if (!nskb) | |
4319 | return; | |
4320 | ||
4321 | skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head); | |
4322 | skb_set_network_header(nskb, (skb_network_header(skb) - | |
4323 | skb->head)); | |
4324 | skb_set_transport_header(nskb, (skb_transport_header(skb) - | |
4325 | skb->head)); | |
4326 | skb_reserve(nskb, header); | |
4327 | memcpy(nskb->head, skb->head, header); | |
4328 | memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); | |
4329 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; | |
4330 | __skb_queue_before(list, skb, nskb); | |
4331 | skb_set_owner_r(nskb, sk); | |
4332 | ||
4333 | /* Copy data, releasing collapsed skbs. */ | |
4334 | while (copy > 0) { | |
4335 | int offset = start - TCP_SKB_CB(skb)->seq; | |
4336 | int size = TCP_SKB_CB(skb)->end_seq - start; | |
4337 | ||
4338 | BUG_ON(offset < 0); | |
4339 | if (size > 0) { | |
4340 | size = min(copy, size); | |
4341 | if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) | |
4342 | BUG(); | |
4343 | TCP_SKB_CB(nskb)->end_seq += size; | |
4344 | copy -= size; | |
4345 | start += size; | |
4346 | } | |
4347 | if (!before(start, TCP_SKB_CB(skb)->end_seq)) { | |
4348 | skb = tcp_collapse_one(sk, skb, list); | |
4349 | if (skb == tail || | |
4350 | tcp_hdr(skb)->syn || | |
4351 | tcp_hdr(skb)->fin) | |
4352 | return; | |
4353 | } | |
4354 | } | |
4355 | } | |
4356 | } | |
4357 | ||
4358 | /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs | |
4359 | * and tcp_collapse() them until all the queue is collapsed. | |
4360 | */ | |
4361 | static void tcp_collapse_ofo_queue(struct sock *sk) | |
4362 | { | |
4363 | struct tcp_sock *tp = tcp_sk(sk); | |
4364 | struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); | |
4365 | struct sk_buff *head; | |
4366 | u32 start, end; | |
4367 | ||
4368 | if (skb == NULL) | |
4369 | return; | |
4370 | ||
4371 | start = TCP_SKB_CB(skb)->seq; | |
4372 | end = TCP_SKB_CB(skb)->end_seq; | |
4373 | head = skb; | |
4374 | ||
4375 | for (;;) { | |
4376 | skb = skb->next; | |
4377 | ||
4378 | /* Segment is terminated when we see gap or when | |
4379 | * we are at the end of all the queue. */ | |
4380 | if (skb == (struct sk_buff *)&tp->out_of_order_queue || | |
4381 | after(TCP_SKB_CB(skb)->seq, end) || | |
4382 | before(TCP_SKB_CB(skb)->end_seq, start)) { | |
4383 | tcp_collapse(sk, &tp->out_of_order_queue, | |
4384 | head, skb, start, end); | |
4385 | head = skb; | |
4386 | if (skb == (struct sk_buff *)&tp->out_of_order_queue) | |
4387 | break; | |
4388 | /* Start new segment */ | |
4389 | start = TCP_SKB_CB(skb)->seq; | |
4390 | end = TCP_SKB_CB(skb)->end_seq; | |
4391 | } else { | |
4392 | if (before(TCP_SKB_CB(skb)->seq, start)) | |
4393 | start = TCP_SKB_CB(skb)->seq; | |
4394 | if (after(TCP_SKB_CB(skb)->end_seq, end)) | |
4395 | end = TCP_SKB_CB(skb)->end_seq; | |
4396 | } | |
4397 | } | |
4398 | } | |
4399 | ||
4400 | /* | |
4401 | * Purge the out-of-order queue. | |
4402 | * Return true if queue was pruned. | |
4403 | */ | |
4404 | static int tcp_prune_ofo_queue(struct sock *sk) | |
4405 | { | |
4406 | struct tcp_sock *tp = tcp_sk(sk); | |
4407 | int res = 0; | |
4408 | ||
4409 | if (!skb_queue_empty(&tp->out_of_order_queue)) { | |
4410 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED); | |
4411 | __skb_queue_purge(&tp->out_of_order_queue); | |
4412 | ||
4413 | /* Reset SACK state. A conforming SACK implementation will | |
4414 | * do the same at a timeout based retransmit. When a connection | |
4415 | * is in a sad state like this, we care only about integrity | |
4416 | * of the connection not performance. | |
4417 | */ | |
4418 | if (tp->rx_opt.sack_ok) | |
4419 | tcp_sack_reset(&tp->rx_opt); | |
4420 | sk_mem_reclaim(sk); | |
4421 | res = 1; | |
4422 | } | |
4423 | return res; | |
4424 | } | |
4425 | ||
4426 | /* Reduce allocated memory if we can, trying to get | |
4427 | * the socket within its memory limits again. | |
4428 | * | |
4429 | * Return less than zero if we should start dropping frames | |
4430 | * until the socket owning process reads some of the data | |
4431 | * to stabilize the situation. | |
4432 | */ | |
4433 | static int tcp_prune_queue(struct sock *sk) | |
4434 | { | |
4435 | struct tcp_sock *tp = tcp_sk(sk); | |
4436 | ||
4437 | SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); | |
4438 | ||
4439 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED); | |
4440 | ||
4441 | if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) | |
4442 | tcp_clamp_window(sk); | |
4443 | else if (tcp_memory_pressure) | |
4444 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); | |
4445 | ||
4446 | tcp_collapse_ofo_queue(sk); | |
4447 | tcp_collapse(sk, &sk->sk_receive_queue, | |
4448 | sk->sk_receive_queue.next, | |
4449 | (struct sk_buff *)&sk->sk_receive_queue, | |
4450 | tp->copied_seq, tp->rcv_nxt); | |
4451 | sk_mem_reclaim(sk); | |
4452 | ||
4453 | if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) | |
4454 | return 0; | |
4455 | ||
4456 | /* Collapsing did not help, destructive actions follow. | |
4457 | * This must not ever occur. */ | |
4458 | ||
4459 | tcp_prune_ofo_queue(sk); | |
4460 | ||
4461 | if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) | |
4462 | return 0; | |
4463 | ||
4464 | /* If we are really being abused, tell the caller to silently | |
4465 | * drop receive data on the floor. It will get retransmitted | |
4466 | * and hopefully then we'll have sufficient space. | |
4467 | */ | |
4468 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED); | |
4469 | ||
4470 | /* Massive buffer overcommit. */ | |
4471 | tp->pred_flags = 0; | |
4472 | return -1; | |
4473 | } | |
4474 | ||
4475 | /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. | |
4476 | * As additional protections, we do not touch cwnd in retransmission phases, | |
4477 | * and if application hit its sndbuf limit recently. | |
4478 | */ | |
4479 | void tcp_cwnd_application_limited(struct sock *sk) | |
4480 | { | |
4481 | struct tcp_sock *tp = tcp_sk(sk); | |
4482 | ||
4483 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && | |
4484 | sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { | |
4485 | /* Limited by application or receiver window. */ | |
4486 | u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); | |
4487 | u32 win_used = max(tp->snd_cwnd_used, init_win); | |
4488 | if (win_used < tp->snd_cwnd) { | |
4489 | tp->snd_ssthresh = tcp_current_ssthresh(sk); | |
4490 | tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; | |
4491 | } | |
4492 | tp->snd_cwnd_used = 0; | |
4493 | } | |
4494 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
4495 | } | |
4496 | ||
4497 | static int tcp_should_expand_sndbuf(struct sock *sk) | |
4498 | { | |
4499 | struct tcp_sock *tp = tcp_sk(sk); | |
4500 | ||
4501 | /* If the user specified a specific send buffer setting, do | |
4502 | * not modify it. | |
4503 | */ | |
4504 | if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) | |
4505 | return 0; | |
4506 | ||
4507 | /* If we are under global TCP memory pressure, do not expand. */ | |
4508 | if (tcp_memory_pressure) | |
4509 | return 0; | |
4510 | ||
4511 | /* If we are under soft global TCP memory pressure, do not expand. */ | |
4512 | if (atomic_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0]) | |
4513 | return 0; | |
4514 | ||
4515 | /* If we filled the congestion window, do not expand. */ | |
4516 | if (tp->packets_out >= tp->snd_cwnd) | |
4517 | return 0; | |
4518 | ||
4519 | return 1; | |
4520 | } | |
4521 | ||
4522 | /* When incoming ACK allowed to free some skb from write_queue, | |
4523 | * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket | |
4524 | * on the exit from tcp input handler. | |
4525 | * | |
4526 | * PROBLEM: sndbuf expansion does not work well with largesend. | |
4527 | */ | |
4528 | static void tcp_new_space(struct sock *sk) | |
4529 | { | |
4530 | struct tcp_sock *tp = tcp_sk(sk); | |
4531 | ||
4532 | if (tcp_should_expand_sndbuf(sk)) { | |
4533 | int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + | |
4534 | MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); | |
4535 | int demanded = max_t(unsigned int, tp->snd_cwnd, | |
4536 | tp->reordering + 1); | |
4537 | sndmem *= 2 * demanded; | |
4538 | if (sndmem > sk->sk_sndbuf) | |
4539 | sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); | |
4540 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
4541 | } | |
4542 | ||
4543 | sk->sk_write_space(sk); | |
4544 | } | |
4545 | ||
4546 | static void tcp_check_space(struct sock *sk) | |
4547 | { | |
4548 | if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { | |
4549 | sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); | |
4550 | if (sk->sk_socket && | |
4551 | test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) | |
4552 | tcp_new_space(sk); | |
4553 | } | |
4554 | } | |
4555 | ||
4556 | static inline void tcp_data_snd_check(struct sock *sk) | |
4557 | { | |
4558 | tcp_push_pending_frames(sk); | |
4559 | tcp_check_space(sk); | |
4560 | } | |
4561 | ||
4562 | /* | |
4563 | * Check if sending an ack is needed. | |
4564 | */ | |
4565 | static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) | |
4566 | { | |
4567 | struct tcp_sock *tp = tcp_sk(sk); | |
4568 | ||
4569 | /* More than one full frame received... */ | |
4570 | if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss | |
4571 | /* ... and right edge of window advances far enough. | |
4572 | * (tcp_recvmsg() will send ACK otherwise). Or... | |
4573 | */ | |
4574 | && __tcp_select_window(sk) >= tp->rcv_wnd) || | |
4575 | /* We ACK each frame or... */ | |
4576 | tcp_in_quickack_mode(sk) || | |
4577 | /* We have out of order data. */ | |
4578 | (ofo_possible && skb_peek(&tp->out_of_order_queue))) { | |
4579 | /* Then ack it now */ | |
4580 | tcp_send_ack(sk); | |
4581 | } else { | |
4582 | /* Else, send delayed ack. */ | |
4583 | tcp_send_delayed_ack(sk); | |
4584 | } | |
4585 | } | |
4586 | ||
4587 | static inline void tcp_ack_snd_check(struct sock *sk) | |
4588 | { | |
4589 | if (!inet_csk_ack_scheduled(sk)) { | |
4590 | /* We sent a data segment already. */ | |
4591 | return; | |
4592 | } | |
4593 | __tcp_ack_snd_check(sk, 1); | |
4594 | } | |
4595 | ||
4596 | /* | |
4597 | * This routine is only called when we have urgent data | |
4598 | * signaled. Its the 'slow' part of tcp_urg. It could be | |
4599 | * moved inline now as tcp_urg is only called from one | |
4600 | * place. We handle URGent data wrong. We have to - as | |
4601 | * BSD still doesn't use the correction from RFC961. | |
4602 | * For 1003.1g we should support a new option TCP_STDURG to permit | |
4603 | * either form (or just set the sysctl tcp_stdurg). | |
4604 | */ | |
4605 | ||
4606 | static void tcp_check_urg(struct sock *sk, struct tcphdr *th) | |
4607 | { | |
4608 | struct tcp_sock *tp = tcp_sk(sk); | |
4609 | u32 ptr = ntohs(th->urg_ptr); | |
4610 | ||
4611 | if (ptr && !sysctl_tcp_stdurg) | |
4612 | ptr--; | |
4613 | ptr += ntohl(th->seq); | |
4614 | ||
4615 | /* Ignore urgent data that we've already seen and read. */ | |
4616 | if (after(tp->copied_seq, ptr)) | |
4617 | return; | |
4618 | ||
4619 | /* Do not replay urg ptr. | |
4620 | * | |
4621 | * NOTE: interesting situation not covered by specs. | |
4622 | * Misbehaving sender may send urg ptr, pointing to segment, | |
4623 | * which we already have in ofo queue. We are not able to fetch | |
4624 | * such data and will stay in TCP_URG_NOTYET until will be eaten | |
4625 | * by recvmsg(). Seems, we are not obliged to handle such wicked | |
4626 | * situations. But it is worth to think about possibility of some | |
4627 | * DoSes using some hypothetical application level deadlock. | |
4628 | */ | |
4629 | if (before(ptr, tp->rcv_nxt)) | |
4630 | return; | |
4631 | ||
4632 | /* Do we already have a newer (or duplicate) urgent pointer? */ | |
4633 | if (tp->urg_data && !after(ptr, tp->urg_seq)) | |
4634 | return; | |
4635 | ||
4636 | /* Tell the world about our new urgent pointer. */ | |
4637 | sk_send_sigurg(sk); | |
4638 | ||
4639 | /* We may be adding urgent data when the last byte read was | |
4640 | * urgent. To do this requires some care. We cannot just ignore | |
4641 | * tp->copied_seq since we would read the last urgent byte again | |
4642 | * as data, nor can we alter copied_seq until this data arrives | |
4643 | * or we break the semantics of SIOCATMARK (and thus sockatmark()) | |
4644 | * | |
4645 | * NOTE. Double Dutch. Rendering to plain English: author of comment | |
4646 | * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); | |
4647 | * and expect that both A and B disappear from stream. This is _wrong_. | |
4648 | * Though this happens in BSD with high probability, this is occasional. | |
4649 | * Any application relying on this is buggy. Note also, that fix "works" | |
4650 | * only in this artificial test. Insert some normal data between A and B and we will | |
4651 | * decline of BSD again. Verdict: it is better to remove to trap | |
4652 | * buggy users. | |
4653 | */ | |
4654 | if (tp->urg_seq == tp->copied_seq && tp->urg_data && | |
4655 | !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { | |
4656 | struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); | |
4657 | tp->copied_seq++; | |
4658 | if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { | |
4659 | __skb_unlink(skb, &sk->sk_receive_queue); | |
4660 | __kfree_skb(skb); | |
4661 | } | |
4662 | } | |
4663 | ||
4664 | tp->urg_data = TCP_URG_NOTYET; | |
4665 | tp->urg_seq = ptr; | |
4666 | ||
4667 | /* Disable header prediction. */ | |
4668 | tp->pred_flags = 0; | |
4669 | } | |
4670 | ||
4671 | /* This is the 'fast' part of urgent handling. */ | |
4672 | static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th) | |
4673 | { | |
4674 | struct tcp_sock *tp = tcp_sk(sk); | |
4675 | ||
4676 | /* Check if we get a new urgent pointer - normally not. */ | |
4677 | if (th->urg) | |
4678 | tcp_check_urg(sk, th); | |
4679 | ||
4680 | /* Do we wait for any urgent data? - normally not... */ | |
4681 | if (tp->urg_data == TCP_URG_NOTYET) { | |
4682 | u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - | |
4683 | th->syn; | |
4684 | ||
4685 | /* Is the urgent pointer pointing into this packet? */ | |
4686 | if (ptr < skb->len) { | |
4687 | u8 tmp; | |
4688 | if (skb_copy_bits(skb, ptr, &tmp, 1)) | |
4689 | BUG(); | |
4690 | tp->urg_data = TCP_URG_VALID | tmp; | |
4691 | if (!sock_flag(sk, SOCK_DEAD)) | |
4692 | sk->sk_data_ready(sk, 0); | |
4693 | } | |
4694 | } | |
4695 | } | |
4696 | ||
4697 | static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) | |
4698 | { | |
4699 | struct tcp_sock *tp = tcp_sk(sk); | |
4700 | int chunk = skb->len - hlen; | |
4701 | int err; | |
4702 | ||
4703 | local_bh_enable(); | |
4704 | if (skb_csum_unnecessary(skb)) | |
4705 | err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); | |
4706 | else | |
4707 | err = skb_copy_and_csum_datagram_iovec(skb, hlen, | |
4708 | tp->ucopy.iov); | |
4709 | ||
4710 | if (!err) { | |
4711 | tp->ucopy.len -= chunk; | |
4712 | tp->copied_seq += chunk; | |
4713 | tcp_rcv_space_adjust(sk); | |
4714 | } | |
4715 | ||
4716 | local_bh_disable(); | |
4717 | return err; | |
4718 | } | |
4719 | ||
4720 | static __sum16 __tcp_checksum_complete_user(struct sock *sk, | |
4721 | struct sk_buff *skb) | |
4722 | { | |
4723 | __sum16 result; | |
4724 | ||
4725 | if (sock_owned_by_user(sk)) { | |
4726 | local_bh_enable(); | |
4727 | result = __tcp_checksum_complete(skb); | |
4728 | local_bh_disable(); | |
4729 | } else { | |
4730 | result = __tcp_checksum_complete(skb); | |
4731 | } | |
4732 | return result; | |
4733 | } | |
4734 | ||
4735 | static inline int tcp_checksum_complete_user(struct sock *sk, | |
4736 | struct sk_buff *skb) | |
4737 | { | |
4738 | return !skb_csum_unnecessary(skb) && | |
4739 | __tcp_checksum_complete_user(sk, skb); | |
4740 | } | |
4741 | ||
4742 | #ifdef CONFIG_NET_DMA | |
4743 | static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, | |
4744 | int hlen) | |
4745 | { | |
4746 | struct tcp_sock *tp = tcp_sk(sk); | |
4747 | int chunk = skb->len - hlen; | |
4748 | int dma_cookie; | |
4749 | int copied_early = 0; | |
4750 | ||
4751 | if (tp->ucopy.wakeup) | |
4752 | return 0; | |
4753 | ||
4754 | if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) | |
4755 | tp->ucopy.dma_chan = get_softnet_dma(); | |
4756 | ||
4757 | if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) { | |
4758 | ||
4759 | dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, | |
4760 | skb, hlen, | |
4761 | tp->ucopy.iov, chunk, | |
4762 | tp->ucopy.pinned_list); | |
4763 | ||
4764 | if (dma_cookie < 0) | |
4765 | goto out; | |
4766 | ||
4767 | tp->ucopy.dma_cookie = dma_cookie; | |
4768 | copied_early = 1; | |
4769 | ||
4770 | tp->ucopy.len -= chunk; | |
4771 | tp->copied_seq += chunk; | |
4772 | tcp_rcv_space_adjust(sk); | |
4773 | ||
4774 | if ((tp->ucopy.len == 0) || | |
4775 | (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) || | |
4776 | (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { | |
4777 | tp->ucopy.wakeup = 1; | |
4778 | sk->sk_data_ready(sk, 0); | |
4779 | } | |
4780 | } else if (chunk > 0) { | |
4781 | tp->ucopy.wakeup = 1; | |
4782 | sk->sk_data_ready(sk, 0); | |
4783 | } | |
4784 | out: | |
4785 | return copied_early; | |
4786 | } | |
4787 | #endif /* CONFIG_NET_DMA */ | |
4788 | ||
4789 | /* Does PAWS and seqno based validation of an incoming segment, flags will | |
4790 | * play significant role here. | |
4791 | */ | |
4792 | static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb, | |
4793 | struct tcphdr *th, int syn_inerr) | |
4794 | { | |
4795 | struct tcp_sock *tp = tcp_sk(sk); | |
4796 | ||
4797 | /* RFC1323: H1. Apply PAWS check first. */ | |
4798 | if (tcp_fast_parse_options(skb, th, tp) && tp->rx_opt.saw_tstamp && | |
4799 | tcp_paws_discard(sk, skb)) { | |
4800 | if (!th->rst) { | |
4801 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); | |
4802 | tcp_send_dupack(sk, skb); | |
4803 | goto discard; | |
4804 | } | |
4805 | /* Reset is accepted even if it did not pass PAWS. */ | |
4806 | } | |
4807 | ||
4808 | /* Step 1: check sequence number */ | |
4809 | if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { | |
4810 | /* RFC793, page 37: "In all states except SYN-SENT, all reset | |
4811 | * (RST) segments are validated by checking their SEQ-fields." | |
4812 | * And page 69: "If an incoming segment is not acceptable, | |
4813 | * an acknowledgment should be sent in reply (unless the RST | |
4814 | * bit is set, if so drop the segment and return)". | |
4815 | */ | |
4816 | if (!th->rst) | |
4817 | tcp_send_dupack(sk, skb); | |
4818 | goto discard; | |
4819 | } | |
4820 | ||
4821 | /* Step 2: check RST bit */ | |
4822 | if (th->rst) { | |
4823 | tcp_reset(sk); | |
4824 | goto discard; | |
4825 | } | |
4826 | ||
4827 | /* ts_recent update must be made after we are sure that the packet | |
4828 | * is in window. | |
4829 | */ | |
4830 | tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); | |
4831 | ||
4832 | /* step 3: check security and precedence [ignored] */ | |
4833 | ||
4834 | /* step 4: Check for a SYN in window. */ | |
4835 | if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
4836 | if (syn_inerr) | |
4837 | TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); | |
4838 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN); | |
4839 | tcp_reset(sk); | |
4840 | return -1; | |
4841 | } | |
4842 | ||
4843 | return 1; | |
4844 | ||
4845 | discard: | |
4846 | __kfree_skb(skb); | |
4847 | return 0; | |
4848 | } | |
4849 | ||
4850 | /* | |
4851 | * TCP receive function for the ESTABLISHED state. | |
4852 | * | |
4853 | * It is split into a fast path and a slow path. The fast path is | |
4854 | * disabled when: | |
4855 | * - A zero window was announced from us - zero window probing | |
4856 | * is only handled properly in the slow path. | |
4857 | * - Out of order segments arrived. | |
4858 | * - Urgent data is expected. | |
4859 | * - There is no buffer space left | |
4860 | * - Unexpected TCP flags/window values/header lengths are received | |
4861 | * (detected by checking the TCP header against pred_flags) | |
4862 | * - Data is sent in both directions. Fast path only supports pure senders | |
4863 | * or pure receivers (this means either the sequence number or the ack | |
4864 | * value must stay constant) | |
4865 | * - Unexpected TCP option. | |
4866 | * | |
4867 | * When these conditions are not satisfied it drops into a standard | |
4868 | * receive procedure patterned after RFC793 to handle all cases. | |
4869 | * The first three cases are guaranteed by proper pred_flags setting, | |
4870 | * the rest is checked inline. Fast processing is turned on in | |
4871 | * tcp_data_queue when everything is OK. | |
4872 | */ | |
4873 | int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, | |
4874 | struct tcphdr *th, unsigned len) | |
4875 | { | |
4876 | struct tcp_sock *tp = tcp_sk(sk); | |
4877 | int res; | |
4878 | ||
4879 | /* | |
4880 | * Header prediction. | |
4881 | * The code loosely follows the one in the famous | |
4882 | * "30 instruction TCP receive" Van Jacobson mail. | |
4883 | * | |
4884 | * Van's trick is to deposit buffers into socket queue | |
4885 | * on a device interrupt, to call tcp_recv function | |
4886 | * on the receive process context and checksum and copy | |
4887 | * the buffer to user space. smart... | |
4888 | * | |
4889 | * Our current scheme is not silly either but we take the | |
4890 | * extra cost of the net_bh soft interrupt processing... | |
4891 | * We do checksum and copy also but from device to kernel. | |
4892 | */ | |
4893 | ||
4894 | tp->rx_opt.saw_tstamp = 0; | |
4895 | ||
4896 | /* pred_flags is 0xS?10 << 16 + snd_wnd | |
4897 | * if header_prediction is to be made | |
4898 | * 'S' will always be tp->tcp_header_len >> 2 | |
4899 | * '?' will be 0 for the fast path, otherwise pred_flags is 0 to | |
4900 | * turn it off (when there are holes in the receive | |
4901 | * space for instance) | |
4902 | * PSH flag is ignored. | |
4903 | */ | |
4904 | ||
4905 | if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && | |
4906 | TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { | |
4907 | int tcp_header_len = tp->tcp_header_len; | |
4908 | ||
4909 | /* Timestamp header prediction: tcp_header_len | |
4910 | * is automatically equal to th->doff*4 due to pred_flags | |
4911 | * match. | |
4912 | */ | |
4913 | ||
4914 | /* Check timestamp */ | |
4915 | if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { | |
4916 | /* No? Slow path! */ | |
4917 | if (!tcp_parse_aligned_timestamp(tp, th)) | |
4918 | goto slow_path; | |
4919 | ||
4920 | /* If PAWS failed, check it more carefully in slow path */ | |
4921 | if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) | |
4922 | goto slow_path; | |
4923 | ||
4924 | /* DO NOT update ts_recent here, if checksum fails | |
4925 | * and timestamp was corrupted part, it will result | |
4926 | * in a hung connection since we will drop all | |
4927 | * future packets due to the PAWS test. | |
4928 | */ | |
4929 | } | |
4930 | ||
4931 | if (len <= tcp_header_len) { | |
4932 | /* Bulk data transfer: sender */ | |
4933 | if (len == tcp_header_len) { | |
4934 | /* Predicted packet is in window by definition. | |
4935 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. | |
4936 | * Hence, check seq<=rcv_wup reduces to: | |
4937 | */ | |
4938 | if (tcp_header_len == | |
4939 | (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && | |
4940 | tp->rcv_nxt == tp->rcv_wup) | |
4941 | tcp_store_ts_recent(tp); | |
4942 | ||
4943 | /* We know that such packets are checksummed | |
4944 | * on entry. | |
4945 | */ | |
4946 | tcp_ack(sk, skb, 0); | |
4947 | __kfree_skb(skb); | |
4948 | tcp_data_snd_check(sk); | |
4949 | return 0; | |
4950 | } else { /* Header too small */ | |
4951 | TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); | |
4952 | goto discard; | |
4953 | } | |
4954 | } else { | |
4955 | int eaten = 0; | |
4956 | int copied_early = 0; | |
4957 | ||
4958 | if (tp->copied_seq == tp->rcv_nxt && | |
4959 | len - tcp_header_len <= tp->ucopy.len) { | |
4960 | #ifdef CONFIG_NET_DMA | |
4961 | if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { | |
4962 | copied_early = 1; | |
4963 | eaten = 1; | |
4964 | } | |
4965 | #endif | |
4966 | if (tp->ucopy.task == current && | |
4967 | sock_owned_by_user(sk) && !copied_early) { | |
4968 | __set_current_state(TASK_RUNNING); | |
4969 | ||
4970 | if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) | |
4971 | eaten = 1; | |
4972 | } | |
4973 | if (eaten) { | |
4974 | /* Predicted packet is in window by definition. | |
4975 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. | |
4976 | * Hence, check seq<=rcv_wup reduces to: | |
4977 | */ | |
4978 | if (tcp_header_len == | |
4979 | (sizeof(struct tcphdr) + | |
4980 | TCPOLEN_TSTAMP_ALIGNED) && | |
4981 | tp->rcv_nxt == tp->rcv_wup) | |
4982 | tcp_store_ts_recent(tp); | |
4983 | ||
4984 | tcp_rcv_rtt_measure_ts(sk, skb); | |
4985 | ||
4986 | __skb_pull(skb, tcp_header_len); | |
4987 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
4988 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER); | |
4989 | } | |
4990 | if (copied_early) | |
4991 | tcp_cleanup_rbuf(sk, skb->len); | |
4992 | } | |
4993 | if (!eaten) { | |
4994 | if (tcp_checksum_complete_user(sk, skb)) | |
4995 | goto csum_error; | |
4996 | ||
4997 | /* Predicted packet is in window by definition. | |
4998 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. | |
4999 | * Hence, check seq<=rcv_wup reduces to: | |
5000 | */ | |
5001 | if (tcp_header_len == | |
5002 | (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && | |
5003 | tp->rcv_nxt == tp->rcv_wup) | |
5004 | tcp_store_ts_recent(tp); | |
5005 | ||
5006 | tcp_rcv_rtt_measure_ts(sk, skb); | |
5007 | ||
5008 | if ((int)skb->truesize > sk->sk_forward_alloc) | |
5009 | goto step5; | |
5010 | ||
5011 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS); | |
5012 | ||
5013 | /* Bulk data transfer: receiver */ | |
5014 | __skb_pull(skb, tcp_header_len); | |
5015 | __skb_queue_tail(&sk->sk_receive_queue, skb); | |
5016 | skb_set_owner_r(skb, sk); | |
5017 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
5018 | } | |
5019 | ||
5020 | tcp_event_data_recv(sk, skb); | |
5021 | ||
5022 | if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { | |
5023 | /* Well, only one small jumplet in fast path... */ | |
5024 | tcp_ack(sk, skb, FLAG_DATA); | |
5025 | tcp_data_snd_check(sk); | |
5026 | if (!inet_csk_ack_scheduled(sk)) | |
5027 | goto no_ack; | |
5028 | } | |
5029 | ||
5030 | if (!copied_early || tp->rcv_nxt != tp->rcv_wup) | |
5031 | __tcp_ack_snd_check(sk, 0); | |
5032 | no_ack: | |
5033 | #ifdef CONFIG_NET_DMA | |
5034 | if (copied_early) | |
5035 | __skb_queue_tail(&sk->sk_async_wait_queue, skb); | |
5036 | else | |
5037 | #endif | |
5038 | if (eaten) | |
5039 | __kfree_skb(skb); | |
5040 | else | |
5041 | sk->sk_data_ready(sk, 0); | |
5042 | return 0; | |
5043 | } | |
5044 | } | |
5045 | ||
5046 | slow_path: | |
5047 | if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb)) | |
5048 | goto csum_error; | |
5049 | ||
5050 | /* | |
5051 | * Standard slow path. | |
5052 | */ | |
5053 | ||
5054 | res = tcp_validate_incoming(sk, skb, th, 1); | |
5055 | if (res <= 0) | |
5056 | return -res; | |
5057 | ||
5058 | step5: | |
5059 | if (th->ack) | |
5060 | tcp_ack(sk, skb, FLAG_SLOWPATH); | |
5061 | ||
5062 | tcp_rcv_rtt_measure_ts(sk, skb); | |
5063 | ||
5064 | /* Process urgent data. */ | |
5065 | tcp_urg(sk, skb, th); | |
5066 | ||
5067 | /* step 7: process the segment text */ | |
5068 | tcp_data_queue(sk, skb); | |
5069 | ||
5070 | tcp_data_snd_check(sk); | |
5071 | tcp_ack_snd_check(sk); | |
5072 | return 0; | |
5073 | ||
5074 | csum_error: | |
5075 | TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); | |
5076 | ||
5077 | discard: | |
5078 | __kfree_skb(skb); | |
5079 | return 0; | |
5080 | } | |
5081 | ||
5082 | static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, | |
5083 | struct tcphdr *th, unsigned len) | |
5084 | { | |
5085 | struct tcp_sock *tp = tcp_sk(sk); | |
5086 | struct inet_connection_sock *icsk = inet_csk(sk); | |
5087 | int saved_clamp = tp->rx_opt.mss_clamp; | |
5088 | ||
5089 | tcp_parse_options(skb, &tp->rx_opt, 0); | |
5090 | ||
5091 | if (th->ack) { | |
5092 | /* rfc793: | |
5093 | * "If the state is SYN-SENT then | |
5094 | * first check the ACK bit | |
5095 | * If the ACK bit is set | |
5096 | * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send | |
5097 | * a reset (unless the RST bit is set, if so drop | |
5098 | * the segment and return)" | |
5099 | * | |
5100 | * We do not send data with SYN, so that RFC-correct | |
5101 | * test reduces to: | |
5102 | */ | |
5103 | if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) | |
5104 | goto reset_and_undo; | |
5105 | ||
5106 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && | |
5107 | !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, | |
5108 | tcp_time_stamp)) { | |
5109 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED); | |
5110 | goto reset_and_undo; | |
5111 | } | |
5112 | ||
5113 | /* Now ACK is acceptable. | |
5114 | * | |
5115 | * "If the RST bit is set | |
5116 | * If the ACK was acceptable then signal the user "error: | |
5117 | * connection reset", drop the segment, enter CLOSED state, | |
5118 | * delete TCB, and return." | |
5119 | */ | |
5120 | ||
5121 | if (th->rst) { | |
5122 | tcp_reset(sk); | |
5123 | goto discard; | |
5124 | } | |
5125 | ||
5126 | /* rfc793: | |
5127 | * "fifth, if neither of the SYN or RST bits is set then | |
5128 | * drop the segment and return." | |
5129 | * | |
5130 | * See note below! | |
5131 | * --ANK(990513) | |
5132 | */ | |
5133 | if (!th->syn) | |
5134 | goto discard_and_undo; | |
5135 | ||
5136 | /* rfc793: | |
5137 | * "If the SYN bit is on ... | |
5138 | * are acceptable then ... | |
5139 | * (our SYN has been ACKed), change the connection | |
5140 | * state to ESTABLISHED..." | |
5141 | */ | |
5142 | ||
5143 | TCP_ECN_rcv_synack(tp, th); | |
5144 | ||
5145 | tp->snd_wl1 = TCP_SKB_CB(skb)->seq; | |
5146 | tcp_ack(sk, skb, FLAG_SLOWPATH); | |
5147 | ||
5148 | /* Ok.. it's good. Set up sequence numbers and | |
5149 | * move to established. | |
5150 | */ | |
5151 | tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; | |
5152 | tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; | |
5153 | ||
5154 | /* RFC1323: The window in SYN & SYN/ACK segments is | |
5155 | * never scaled. | |
5156 | */ | |
5157 | tp->snd_wnd = ntohs(th->window); | |
5158 | tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, TCP_SKB_CB(skb)->seq); | |
5159 | ||
5160 | if (!tp->rx_opt.wscale_ok) { | |
5161 | tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; | |
5162 | tp->window_clamp = min(tp->window_clamp, 65535U); | |
5163 | } | |
5164 | ||
5165 | if (tp->rx_opt.saw_tstamp) { | |
5166 | tp->rx_opt.tstamp_ok = 1; | |
5167 | tp->tcp_header_len = | |
5168 | sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; | |
5169 | tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; | |
5170 | tcp_store_ts_recent(tp); | |
5171 | } else { | |
5172 | tp->tcp_header_len = sizeof(struct tcphdr); | |
5173 | } | |
5174 | ||
5175 | if (tcp_is_sack(tp) && sysctl_tcp_fack) | |
5176 | tcp_enable_fack(tp); | |
5177 | ||
5178 | tcp_mtup_init(sk); | |
5179 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); | |
5180 | tcp_initialize_rcv_mss(sk); | |
5181 | ||
5182 | /* Remember, tcp_poll() does not lock socket! | |
5183 | * Change state from SYN-SENT only after copied_seq | |
5184 | * is initialized. */ | |
5185 | tp->copied_seq = tp->rcv_nxt; | |
5186 | smp_mb(); | |
5187 | tcp_set_state(sk, TCP_ESTABLISHED); | |
5188 | ||
5189 | security_inet_conn_established(sk, skb); | |
5190 | ||
5191 | /* Make sure socket is routed, for correct metrics. */ | |
5192 | icsk->icsk_af_ops->rebuild_header(sk); | |
5193 | ||
5194 | tcp_init_metrics(sk); | |
5195 | ||
5196 | tcp_init_congestion_control(sk); | |
5197 | ||
5198 | /* Prevent spurious tcp_cwnd_restart() on first data | |
5199 | * packet. | |
5200 | */ | |
5201 | tp->lsndtime = tcp_time_stamp; | |
5202 | ||
5203 | tcp_init_buffer_space(sk); | |
5204 | ||
5205 | if (sock_flag(sk, SOCK_KEEPOPEN)) | |
5206 | inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); | |
5207 | ||
5208 | if (!tp->rx_opt.snd_wscale) | |
5209 | __tcp_fast_path_on(tp, tp->snd_wnd); | |
5210 | else | |
5211 | tp->pred_flags = 0; | |
5212 | ||
5213 | if (!sock_flag(sk, SOCK_DEAD)) { | |
5214 | sk->sk_state_change(sk); | |
5215 | sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); | |
5216 | } | |
5217 | ||
5218 | if (sk->sk_write_pending || | |
5219 | icsk->icsk_accept_queue.rskq_defer_accept || | |
5220 | icsk->icsk_ack.pingpong) { | |
5221 | /* Save one ACK. Data will be ready after | |
5222 | * several ticks, if write_pending is set. | |
5223 | * | |
5224 | * It may be deleted, but with this feature tcpdumps | |
5225 | * look so _wonderfully_ clever, that I was not able | |
5226 | * to stand against the temptation 8) --ANK | |
5227 | */ | |
5228 | inet_csk_schedule_ack(sk); | |
5229 | icsk->icsk_ack.lrcvtime = tcp_time_stamp; | |
5230 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
5231 | tcp_incr_quickack(sk); | |
5232 | tcp_enter_quickack_mode(sk); | |
5233 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, | |
5234 | TCP_DELACK_MAX, TCP_RTO_MAX); | |
5235 | ||
5236 | discard: | |
5237 | __kfree_skb(skb); | |
5238 | return 0; | |
5239 | } else { | |
5240 | tcp_send_ack(sk); | |
5241 | } | |
5242 | return -1; | |
5243 | } | |
5244 | ||
5245 | /* No ACK in the segment */ | |
5246 | ||
5247 | if (th->rst) { | |
5248 | /* rfc793: | |
5249 | * "If the RST bit is set | |
5250 | * | |
5251 | * Otherwise (no ACK) drop the segment and return." | |
5252 | */ | |
5253 | ||
5254 | goto discard_and_undo; | |
5255 | } | |
5256 | ||
5257 | /* PAWS check. */ | |
5258 | if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && | |
5259 | tcp_paws_check(&tp->rx_opt, 0)) | |
5260 | goto discard_and_undo; | |
5261 | ||
5262 | if (th->syn) { | |
5263 | /* We see SYN without ACK. It is attempt of | |
5264 | * simultaneous connect with crossed SYNs. | |
5265 | * Particularly, it can be connect to self. | |
5266 | */ | |
5267 | tcp_set_state(sk, TCP_SYN_RECV); | |
5268 | ||
5269 | if (tp->rx_opt.saw_tstamp) { | |
5270 | tp->rx_opt.tstamp_ok = 1; | |
5271 | tcp_store_ts_recent(tp); | |
5272 | tp->tcp_header_len = | |
5273 | sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; | |
5274 | } else { | |
5275 | tp->tcp_header_len = sizeof(struct tcphdr); | |
5276 | } | |
5277 | ||
5278 | tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; | |
5279 | tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; | |
5280 | ||
5281 | /* RFC1323: The window in SYN & SYN/ACK segments is | |
5282 | * never scaled. | |
5283 | */ | |
5284 | tp->snd_wnd = ntohs(th->window); | |
5285 | tp->snd_wl1 = TCP_SKB_CB(skb)->seq; | |
5286 | tp->max_window = tp->snd_wnd; | |
5287 | ||
5288 | TCP_ECN_rcv_syn(tp, th); | |
5289 | ||
5290 | tcp_mtup_init(sk); | |
5291 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); | |
5292 | tcp_initialize_rcv_mss(sk); | |
5293 | ||
5294 | tcp_send_synack(sk); | |
5295 | #if 0 | |
5296 | /* Note, we could accept data and URG from this segment. | |
5297 | * There are no obstacles to make this. | |
5298 | * | |
5299 | * However, if we ignore data in ACKless segments sometimes, | |
5300 | * we have no reasons to accept it sometimes. | |
5301 | * Also, seems the code doing it in step6 of tcp_rcv_state_process | |
5302 | * is not flawless. So, discard packet for sanity. | |
5303 | * Uncomment this return to process the data. | |
5304 | */ | |
5305 | return -1; | |
5306 | #else | |
5307 | goto discard; | |
5308 | #endif | |
5309 | } | |
5310 | /* "fifth, if neither of the SYN or RST bits is set then | |
5311 | * drop the segment and return." | |
5312 | */ | |
5313 | ||
5314 | discard_and_undo: | |
5315 | tcp_clear_options(&tp->rx_opt); | |
5316 | tp->rx_opt.mss_clamp = saved_clamp; | |
5317 | goto discard; | |
5318 | ||
5319 | reset_and_undo: | |
5320 | tcp_clear_options(&tp->rx_opt); | |
5321 | tp->rx_opt.mss_clamp = saved_clamp; | |
5322 | return 1; | |
5323 | } | |
5324 | ||
5325 | /* | |
5326 | * This function implements the receiving procedure of RFC 793 for | |
5327 | * all states except ESTABLISHED and TIME_WAIT. | |
5328 | * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be | |
5329 | * address independent. | |
5330 | */ | |
5331 | ||
5332 | int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, | |
5333 | struct tcphdr *th, unsigned len) | |
5334 | { | |
5335 | struct tcp_sock *tp = tcp_sk(sk); | |
5336 | struct inet_connection_sock *icsk = inet_csk(sk); | |
5337 | int queued = 0; | |
5338 | int res; | |
5339 | ||
5340 | tp->rx_opt.saw_tstamp = 0; | |
5341 | ||
5342 | switch (sk->sk_state) { | |
5343 | case TCP_CLOSE: | |
5344 | goto discard; | |
5345 | ||
5346 | case TCP_LISTEN: | |
5347 | if (th->ack) | |
5348 | return 1; | |
5349 | ||
5350 | if (th->rst) | |
5351 | goto discard; | |
5352 | ||
5353 | if (th->syn) { | |
5354 | if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) | |
5355 | return 1; | |
5356 | ||
5357 | /* Now we have several options: In theory there is | |
5358 | * nothing else in the frame. KA9Q has an option to | |
5359 | * send data with the syn, BSD accepts data with the | |
5360 | * syn up to the [to be] advertised window and | |
5361 | * Solaris 2.1 gives you a protocol error. For now | |
5362 | * we just ignore it, that fits the spec precisely | |
5363 | * and avoids incompatibilities. It would be nice in | |
5364 | * future to drop through and process the data. | |
5365 | * | |
5366 | * Now that TTCP is starting to be used we ought to | |
5367 | * queue this data. | |
5368 | * But, this leaves one open to an easy denial of | |
5369 | * service attack, and SYN cookies can't defend | |
5370 | * against this problem. So, we drop the data | |
5371 | * in the interest of security over speed unless | |
5372 | * it's still in use. | |
5373 | */ | |
5374 | kfree_skb(skb); | |
5375 | return 0; | |
5376 | } | |
5377 | goto discard; | |
5378 | ||
5379 | case TCP_SYN_SENT: | |
5380 | queued = tcp_rcv_synsent_state_process(sk, skb, th, len); | |
5381 | if (queued >= 0) | |
5382 | return queued; | |
5383 | ||
5384 | /* Do step6 onward by hand. */ | |
5385 | tcp_urg(sk, skb, th); | |
5386 | __kfree_skb(skb); | |
5387 | tcp_data_snd_check(sk); | |
5388 | return 0; | |
5389 | } | |
5390 | ||
5391 | res = tcp_validate_incoming(sk, skb, th, 0); | |
5392 | if (res <= 0) | |
5393 | return -res; | |
5394 | ||
5395 | /* step 5: check the ACK field */ | |
5396 | if (th->ack) { | |
5397 | int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH); | |
5398 | ||
5399 | switch (sk->sk_state) { | |
5400 | case TCP_SYN_RECV: | |
5401 | if (acceptable) { | |
5402 | tp->copied_seq = tp->rcv_nxt; | |
5403 | smp_mb(); | |
5404 | tcp_set_state(sk, TCP_ESTABLISHED); | |
5405 | sk->sk_state_change(sk); | |
5406 | ||
5407 | /* Note, that this wakeup is only for marginal | |
5408 | * crossed SYN case. Passively open sockets | |
5409 | * are not waked up, because sk->sk_sleep == | |
5410 | * NULL and sk->sk_socket == NULL. | |
5411 | */ | |
5412 | if (sk->sk_socket) | |
5413 | sk_wake_async(sk, | |
5414 | SOCK_WAKE_IO, POLL_OUT); | |
5415 | ||
5416 | tp->snd_una = TCP_SKB_CB(skb)->ack_seq; | |
5417 | tp->snd_wnd = ntohs(th->window) << | |
5418 | tp->rx_opt.snd_wscale; | |
5419 | tcp_init_wl(tp, TCP_SKB_CB(skb)->ack_seq, | |
5420 | TCP_SKB_CB(skb)->seq); | |
5421 | ||
5422 | /* tcp_ack considers this ACK as duplicate | |
5423 | * and does not calculate rtt. | |
5424 | * Fix it at least with timestamps. | |
5425 | */ | |
5426 | if (tp->rx_opt.saw_tstamp && | |
5427 | tp->rx_opt.rcv_tsecr && !tp->srtt) | |
5428 | tcp_ack_saw_tstamp(sk, 0); | |
5429 | ||
5430 | if (tp->rx_opt.tstamp_ok) | |
5431 | tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; | |
5432 | ||
5433 | /* Make sure socket is routed, for | |
5434 | * correct metrics. | |
5435 | */ | |
5436 | icsk->icsk_af_ops->rebuild_header(sk); | |
5437 | ||
5438 | tcp_init_metrics(sk); | |
5439 | ||
5440 | tcp_init_congestion_control(sk); | |
5441 | ||
5442 | /* Prevent spurious tcp_cwnd_restart() on | |
5443 | * first data packet. | |
5444 | */ | |
5445 | tp->lsndtime = tcp_time_stamp; | |
5446 | ||
5447 | tcp_mtup_init(sk); | |
5448 | tcp_initialize_rcv_mss(sk); | |
5449 | tcp_init_buffer_space(sk); | |
5450 | tcp_fast_path_on(tp); | |
5451 | } else { | |
5452 | return 1; | |
5453 | } | |
5454 | break; | |
5455 | ||
5456 | case TCP_FIN_WAIT1: | |
5457 | if (tp->snd_una == tp->write_seq) { | |
5458 | tcp_set_state(sk, TCP_FIN_WAIT2); | |
5459 | sk->sk_shutdown |= SEND_SHUTDOWN; | |
5460 | dst_confirm(sk->sk_dst_cache); | |
5461 | ||
5462 | if (!sock_flag(sk, SOCK_DEAD)) | |
5463 | /* Wake up lingering close() */ | |
5464 | sk->sk_state_change(sk); | |
5465 | else { | |
5466 | int tmo; | |
5467 | ||
5468 | if (tp->linger2 < 0 || | |
5469 | (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && | |
5470 | after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { | |
5471 | tcp_done(sk); | |
5472 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA); | |
5473 | return 1; | |
5474 | } | |
5475 | ||
5476 | tmo = tcp_fin_time(sk); | |
5477 | if (tmo > TCP_TIMEWAIT_LEN) { | |
5478 | inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); | |
5479 | } else if (th->fin || sock_owned_by_user(sk)) { | |
5480 | /* Bad case. We could lose such FIN otherwise. | |
5481 | * It is not a big problem, but it looks confusing | |
5482 | * and not so rare event. We still can lose it now, | |
5483 | * if it spins in bh_lock_sock(), but it is really | |
5484 | * marginal case. | |
5485 | */ | |
5486 | inet_csk_reset_keepalive_timer(sk, tmo); | |
5487 | } else { | |
5488 | tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); | |
5489 | goto discard; | |
5490 | } | |
5491 | } | |
5492 | } | |
5493 | break; | |
5494 | ||
5495 | case TCP_CLOSING: | |
5496 | if (tp->snd_una == tp->write_seq) { | |
5497 | tcp_time_wait(sk, TCP_TIME_WAIT, 0); | |
5498 | goto discard; | |
5499 | } | |
5500 | break; | |
5501 | ||
5502 | case TCP_LAST_ACK: | |
5503 | if (tp->snd_una == tp->write_seq) { | |
5504 | tcp_update_metrics(sk); | |
5505 | tcp_done(sk); | |
5506 | goto discard; | |
5507 | } | |
5508 | break; | |
5509 | } | |
5510 | } else | |
5511 | goto discard; | |
5512 | ||
5513 | /* step 6: check the URG bit */ | |
5514 | tcp_urg(sk, skb, th); | |
5515 | ||
5516 | /* step 7: process the segment text */ | |
5517 | switch (sk->sk_state) { | |
5518 | case TCP_CLOSE_WAIT: | |
5519 | case TCP_CLOSING: | |
5520 | case TCP_LAST_ACK: | |
5521 | if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) | |
5522 | break; | |
5523 | case TCP_FIN_WAIT1: | |
5524 | case TCP_FIN_WAIT2: | |
5525 | /* RFC 793 says to queue data in these states, | |
5526 | * RFC 1122 says we MUST send a reset. | |
5527 | * BSD 4.4 also does reset. | |
5528 | */ | |
5529 | if (sk->sk_shutdown & RCV_SHUTDOWN) { | |
5530 | if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && | |
5531 | after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { | |
5532 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA); | |
5533 | tcp_reset(sk); | |
5534 | return 1; | |
5535 | } | |
5536 | } | |
5537 | /* Fall through */ | |
5538 | case TCP_ESTABLISHED: | |
5539 | tcp_data_queue(sk, skb); | |
5540 | queued = 1; | |
5541 | break; | |
5542 | } | |
5543 | ||
5544 | /* tcp_data could move socket to TIME-WAIT */ | |
5545 | if (sk->sk_state != TCP_CLOSE) { | |
5546 | tcp_data_snd_check(sk); | |
5547 | tcp_ack_snd_check(sk); | |
5548 | } | |
5549 | ||
5550 | if (!queued) { | |
5551 | discard: | |
5552 | __kfree_skb(skb); | |
5553 | } | |
5554 | return 0; | |
5555 | } | |
5556 | ||
5557 | EXPORT_SYMBOL(sysctl_tcp_ecn); | |
5558 | EXPORT_SYMBOL(sysctl_tcp_reordering); | |
5559 | EXPORT_SYMBOL(sysctl_tcp_adv_win_scale); | |
5560 | EXPORT_SYMBOL(tcp_parse_options); | |
5561 | #ifdef CONFIG_TCP_MD5SIG | |
5562 | EXPORT_SYMBOL(tcp_parse_md5sig_option); | |
5563 | #endif | |
5564 | EXPORT_SYMBOL(tcp_rcv_established); | |
5565 | EXPORT_SYMBOL(tcp_rcv_state_process); | |
5566 | EXPORT_SYMBOL(tcp_initialize_rcv_mss); |