]>
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/slab.h> | |
66 | #include <linux/module.h> | |
67 | #include <linux/sysctl.h> | |
68 | #include <linux/kernel.h> | |
69 | #include <net/dst.h> | |
70 | #include <net/tcp.h> | |
71 | #include <net/inet_common.h> | |
72 | #include <linux/ipsec.h> | |
73 | #include <asm/unaligned.h> | |
74 | #include <net/netdma.h> | |
75 | ||
76 | int sysctl_tcp_timestamps __read_mostly = 1; | |
77 | int sysctl_tcp_window_scaling __read_mostly = 1; | |
78 | int sysctl_tcp_sack __read_mostly = 1; | |
79 | int sysctl_tcp_fack __read_mostly = 1; | |
80 | int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH; | |
81 | EXPORT_SYMBOL(sysctl_tcp_reordering); | |
82 | int sysctl_tcp_ecn __read_mostly = 2; | |
83 | EXPORT_SYMBOL(sysctl_tcp_ecn); | |
84 | int sysctl_tcp_dsack __read_mostly = 1; | |
85 | int sysctl_tcp_app_win __read_mostly = 31; | |
86 | int sysctl_tcp_adv_win_scale __read_mostly = 2; | |
87 | EXPORT_SYMBOL(sysctl_tcp_adv_win_scale); | |
88 | ||
89 | int sysctl_tcp_stdurg __read_mostly; | |
90 | int sysctl_tcp_rfc1337 __read_mostly; | |
91 | int sysctl_tcp_max_orphans __read_mostly = NR_FILE; | |
92 | int sysctl_tcp_frto __read_mostly = 2; | |
93 | int sysctl_tcp_frto_response __read_mostly; | |
94 | int sysctl_tcp_nometrics_save __read_mostly; | |
95 | ||
96 | int sysctl_tcp_thin_dupack __read_mostly; | |
97 | ||
98 | int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; | |
99 | int sysctl_tcp_abc __read_mostly; | |
100 | ||
101 | #define FLAG_DATA 0x01 /* Incoming frame contained data. */ | |
102 | #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ | |
103 | #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ | |
104 | #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ | |
105 | #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ | |
106 | #define FLAG_DATA_SACKED 0x20 /* New SACK. */ | |
107 | #define FLAG_ECE 0x40 /* ECE in this ACK */ | |
108 | #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */ | |
109 | #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ | |
110 | #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */ | |
111 | #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ | |
112 | #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ | |
113 | #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */ | |
114 | #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ | |
115 | ||
116 | #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) | |
117 | #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) | |
118 | #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) | |
119 | #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) | |
120 | #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED) | |
121 | ||
122 | #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) | |
123 | #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) | |
124 | ||
125 | /* Adapt the MSS value used to make delayed ack decision to the | |
126 | * real world. | |
127 | */ | |
128 | static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) | |
129 | { | |
130 | struct inet_connection_sock *icsk = inet_csk(sk); | |
131 | const unsigned int lss = icsk->icsk_ack.last_seg_size; | |
132 | unsigned int len; | |
133 | ||
134 | icsk->icsk_ack.last_seg_size = 0; | |
135 | ||
136 | /* skb->len may jitter because of SACKs, even if peer | |
137 | * sends good full-sized frames. | |
138 | */ | |
139 | len = skb_shinfo(skb)->gso_size ? : skb->len; | |
140 | if (len >= icsk->icsk_ack.rcv_mss) { | |
141 | icsk->icsk_ack.rcv_mss = len; | |
142 | } else { | |
143 | /* Otherwise, we make more careful check taking into account, | |
144 | * that SACKs block is variable. | |
145 | * | |
146 | * "len" is invariant segment length, including TCP header. | |
147 | */ | |
148 | len += skb->data - skb_transport_header(skb); | |
149 | if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || | |
150 | /* If PSH is not set, packet should be | |
151 | * full sized, provided peer TCP is not badly broken. | |
152 | * This observation (if it is correct 8)) allows | |
153 | * to handle super-low mtu links fairly. | |
154 | */ | |
155 | (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && | |
156 | !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { | |
157 | /* Subtract also invariant (if peer is RFC compliant), | |
158 | * tcp header plus fixed timestamp option length. | |
159 | * Resulting "len" is MSS free of SACK jitter. | |
160 | */ | |
161 | len -= tcp_sk(sk)->tcp_header_len; | |
162 | icsk->icsk_ack.last_seg_size = len; | |
163 | if (len == lss) { | |
164 | icsk->icsk_ack.rcv_mss = len; | |
165 | return; | |
166 | } | |
167 | } | |
168 | if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) | |
169 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; | |
170 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; | |
171 | } | |
172 | } | |
173 | ||
174 | static void tcp_incr_quickack(struct sock *sk) | |
175 | { | |
176 | struct inet_connection_sock *icsk = inet_csk(sk); | |
177 | unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); | |
178 | ||
179 | if (quickacks == 0) | |
180 | quickacks = 2; | |
181 | if (quickacks > icsk->icsk_ack.quick) | |
182 | icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); | |
183 | } | |
184 | ||
185 | static void tcp_enter_quickack_mode(struct sock *sk) | |
186 | { | |
187 | struct inet_connection_sock *icsk = inet_csk(sk); | |
188 | tcp_incr_quickack(sk); | |
189 | icsk->icsk_ack.pingpong = 0; | |
190 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
191 | } | |
192 | ||
193 | /* Send ACKs quickly, if "quick" count is not exhausted | |
194 | * and the session is not interactive. | |
195 | */ | |
196 | ||
197 | static inline int tcp_in_quickack_mode(const struct sock *sk) | |
198 | { | |
199 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
200 | return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; | |
201 | } | |
202 | ||
203 | static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp) | |
204 | { | |
205 | if (tp->ecn_flags & TCP_ECN_OK) | |
206 | tp->ecn_flags |= TCP_ECN_QUEUE_CWR; | |
207 | } | |
208 | ||
209 | static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb) | |
210 | { | |
211 | if (tcp_hdr(skb)->cwr) | |
212 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
213 | } | |
214 | ||
215 | static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp) | |
216 | { | |
217 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
218 | } | |
219 | ||
220 | static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb) | |
221 | { | |
222 | if (tp->ecn_flags & TCP_ECN_OK) { | |
223 | if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags)) | |
224 | tp->ecn_flags |= TCP_ECN_DEMAND_CWR; | |
225 | /* Funny extension: if ECT is not set on a segment, | |
226 | * it is surely retransmit. It is not in ECN RFC, | |
227 | * but Linux follows this rule. */ | |
228 | else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags))) | |
229 | tcp_enter_quickack_mode((struct sock *)tp); | |
230 | } | |
231 | } | |
232 | ||
233 | static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th) | |
234 | { | |
235 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) | |
236 | tp->ecn_flags &= ~TCP_ECN_OK; | |
237 | } | |
238 | ||
239 | static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th) | |
240 | { | |
241 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) | |
242 | tp->ecn_flags &= ~TCP_ECN_OK; | |
243 | } | |
244 | ||
245 | static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th) | |
246 | { | |
247 | if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) | |
248 | return 1; | |
249 | return 0; | |
250 | } | |
251 | ||
252 | /* Buffer size and advertised window tuning. | |
253 | * | |
254 | * 1. Tuning sk->sk_sndbuf, when connection enters established state. | |
255 | */ | |
256 | ||
257 | static void tcp_fixup_sndbuf(struct sock *sk) | |
258 | { | |
259 | int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 + | |
260 | sizeof(struct sk_buff); | |
261 | ||
262 | if (sk->sk_sndbuf < 3 * sndmem) { | |
263 | sk->sk_sndbuf = 3 * sndmem; | |
264 | if (sk->sk_sndbuf > sysctl_tcp_wmem[2]) | |
265 | sk->sk_sndbuf = sysctl_tcp_wmem[2]; | |
266 | } | |
267 | } | |
268 | ||
269 | /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) | |
270 | * | |
271 | * All tcp_full_space() is split to two parts: "network" buffer, allocated | |
272 | * forward and advertised in receiver window (tp->rcv_wnd) and | |
273 | * "application buffer", required to isolate scheduling/application | |
274 | * latencies from network. | |
275 | * window_clamp is maximal advertised window. It can be less than | |
276 | * tcp_full_space(), in this case tcp_full_space() - window_clamp | |
277 | * is reserved for "application" buffer. The less window_clamp is | |
278 | * the smoother our behaviour from viewpoint of network, but the lower | |
279 | * throughput and the higher sensitivity of the connection to losses. 8) | |
280 | * | |
281 | * rcv_ssthresh is more strict window_clamp used at "slow start" | |
282 | * phase to predict further behaviour of this connection. | |
283 | * It is used for two goals: | |
284 | * - to enforce header prediction at sender, even when application | |
285 | * requires some significant "application buffer". It is check #1. | |
286 | * - to prevent pruning of receive queue because of misprediction | |
287 | * of receiver window. Check #2. | |
288 | * | |
289 | * The scheme does not work when sender sends good segments opening | |
290 | * window and then starts to feed us spaghetti. But it should work | |
291 | * in common situations. Otherwise, we have to rely on queue collapsing. | |
292 | */ | |
293 | ||
294 | /* Slow part of check#2. */ | |
295 | static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) | |
296 | { | |
297 | struct tcp_sock *tp = tcp_sk(sk); | |
298 | /* Optimize this! */ | |
299 | int truesize = tcp_win_from_space(skb->truesize) >> 1; | |
300 | int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1; | |
301 | ||
302 | while (tp->rcv_ssthresh <= window) { | |
303 | if (truesize <= skb->len) | |
304 | return 2 * inet_csk(sk)->icsk_ack.rcv_mss; | |
305 | ||
306 | truesize >>= 1; | |
307 | window >>= 1; | |
308 | } | |
309 | return 0; | |
310 | } | |
311 | ||
312 | static void tcp_grow_window(struct sock *sk, struct sk_buff *skb) | |
313 | { | |
314 | struct tcp_sock *tp = tcp_sk(sk); | |
315 | ||
316 | /* Check #1 */ | |
317 | if (tp->rcv_ssthresh < tp->window_clamp && | |
318 | (int)tp->rcv_ssthresh < tcp_space(sk) && | |
319 | !tcp_memory_pressure) { | |
320 | int incr; | |
321 | ||
322 | /* Check #2. Increase window, if skb with such overhead | |
323 | * will fit to rcvbuf in future. | |
324 | */ | |
325 | if (tcp_win_from_space(skb->truesize) <= skb->len) | |
326 | incr = 2 * tp->advmss; | |
327 | else | |
328 | incr = __tcp_grow_window(sk, skb); | |
329 | ||
330 | if (incr) { | |
331 | tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, | |
332 | tp->window_clamp); | |
333 | inet_csk(sk)->icsk_ack.quick |= 1; | |
334 | } | |
335 | } | |
336 | } | |
337 | ||
338 | /* 3. Tuning rcvbuf, when connection enters established state. */ | |
339 | ||
340 | static void tcp_fixup_rcvbuf(struct sock *sk) | |
341 | { | |
342 | struct tcp_sock *tp = tcp_sk(sk); | |
343 | int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); | |
344 | ||
345 | /* Try to select rcvbuf so that 4 mss-sized segments | |
346 | * will fit to window and corresponding skbs will fit to our rcvbuf. | |
347 | * (was 3; 4 is minimum to allow fast retransmit to work.) | |
348 | */ | |
349 | while (tcp_win_from_space(rcvmem) < tp->advmss) | |
350 | rcvmem += 128; | |
351 | if (sk->sk_rcvbuf < 4 * rcvmem) | |
352 | sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]); | |
353 | } | |
354 | ||
355 | /* 4. Try to fixup all. It is made immediately after connection enters | |
356 | * established state. | |
357 | */ | |
358 | static void tcp_init_buffer_space(struct sock *sk) | |
359 | { | |
360 | struct tcp_sock *tp = tcp_sk(sk); | |
361 | int maxwin; | |
362 | ||
363 | if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) | |
364 | tcp_fixup_rcvbuf(sk); | |
365 | if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) | |
366 | tcp_fixup_sndbuf(sk); | |
367 | ||
368 | tp->rcvq_space.space = tp->rcv_wnd; | |
369 | ||
370 | maxwin = tcp_full_space(sk); | |
371 | ||
372 | if (tp->window_clamp >= maxwin) { | |
373 | tp->window_clamp = maxwin; | |
374 | ||
375 | if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) | |
376 | tp->window_clamp = max(maxwin - | |
377 | (maxwin >> sysctl_tcp_app_win), | |
378 | 4 * tp->advmss); | |
379 | } | |
380 | ||
381 | /* Force reservation of one segment. */ | |
382 | if (sysctl_tcp_app_win && | |
383 | tp->window_clamp > 2 * tp->advmss && | |
384 | tp->window_clamp + tp->advmss > maxwin) | |
385 | tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); | |
386 | ||
387 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); | |
388 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
389 | } | |
390 | ||
391 | /* 5. Recalculate window clamp after socket hit its memory bounds. */ | |
392 | static void tcp_clamp_window(struct sock *sk) | |
393 | { | |
394 | struct tcp_sock *tp = tcp_sk(sk); | |
395 | struct inet_connection_sock *icsk = inet_csk(sk); | |
396 | ||
397 | icsk->icsk_ack.quick = 0; | |
398 | ||
399 | if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && | |
400 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && | |
401 | !tcp_memory_pressure && | |
402 | atomic_long_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) { | |
403 | sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), | |
404 | sysctl_tcp_rmem[2]); | |
405 | } | |
406 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) | |
407 | tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); | |
408 | } | |
409 | ||
410 | /* Initialize RCV_MSS value. | |
411 | * RCV_MSS is an our guess about MSS used by the peer. | |
412 | * We haven't any direct information about the MSS. | |
413 | * It's better to underestimate the RCV_MSS rather than overestimate. | |
414 | * Overestimations make us ACKing less frequently than needed. | |
415 | * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). | |
416 | */ | |
417 | void tcp_initialize_rcv_mss(struct sock *sk) | |
418 | { | |
419 | struct tcp_sock *tp = tcp_sk(sk); | |
420 | unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); | |
421 | ||
422 | hint = min(hint, tp->rcv_wnd / 2); | |
423 | hint = min(hint, TCP_MSS_DEFAULT); | |
424 | hint = max(hint, TCP_MIN_MSS); | |
425 | ||
426 | inet_csk(sk)->icsk_ack.rcv_mss = hint; | |
427 | } | |
428 | EXPORT_SYMBOL(tcp_initialize_rcv_mss); | |
429 | ||
430 | /* Receiver "autotuning" code. | |
431 | * | |
432 | * The algorithm for RTT estimation w/o timestamps is based on | |
433 | * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. | |
434 | * <http://public.lanl.gov/radiant/pubs.html#DRS> | |
435 | * | |
436 | * More detail on this code can be found at | |
437 | * <http://staff.psc.edu/jheffner/>, | |
438 | * though this reference is out of date. A new paper | |
439 | * is pending. | |
440 | */ | |
441 | static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) | |
442 | { | |
443 | u32 new_sample = tp->rcv_rtt_est.rtt; | |
444 | long m = sample; | |
445 | ||
446 | if (m == 0) | |
447 | m = 1; | |
448 | ||
449 | if (new_sample != 0) { | |
450 | /* If we sample in larger samples in the non-timestamp | |
451 | * case, we could grossly overestimate the RTT especially | |
452 | * with chatty applications or bulk transfer apps which | |
453 | * are stalled on filesystem I/O. | |
454 | * | |
455 | * Also, since we are only going for a minimum in the | |
456 | * non-timestamp case, we do not smooth things out | |
457 | * else with timestamps disabled convergence takes too | |
458 | * long. | |
459 | */ | |
460 | if (!win_dep) { | |
461 | m -= (new_sample >> 3); | |
462 | new_sample += m; | |
463 | } else if (m < new_sample) | |
464 | new_sample = m << 3; | |
465 | } else { | |
466 | /* No previous measure. */ | |
467 | new_sample = m << 3; | |
468 | } | |
469 | ||
470 | if (tp->rcv_rtt_est.rtt != new_sample) | |
471 | tp->rcv_rtt_est.rtt = new_sample; | |
472 | } | |
473 | ||
474 | static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) | |
475 | { | |
476 | if (tp->rcv_rtt_est.time == 0) | |
477 | goto new_measure; | |
478 | if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) | |
479 | return; | |
480 | tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1); | |
481 | ||
482 | new_measure: | |
483 | tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; | |
484 | tp->rcv_rtt_est.time = tcp_time_stamp; | |
485 | } | |
486 | ||
487 | static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, | |
488 | const struct sk_buff *skb) | |
489 | { | |
490 | struct tcp_sock *tp = tcp_sk(sk); | |
491 | if (tp->rx_opt.rcv_tsecr && | |
492 | (TCP_SKB_CB(skb)->end_seq - | |
493 | TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) | |
494 | tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); | |
495 | } | |
496 | ||
497 | /* | |
498 | * This function should be called every time data is copied to user space. | |
499 | * It calculates the appropriate TCP receive buffer space. | |
500 | */ | |
501 | void tcp_rcv_space_adjust(struct sock *sk) | |
502 | { | |
503 | struct tcp_sock *tp = tcp_sk(sk); | |
504 | int time; | |
505 | int space; | |
506 | ||
507 | if (tp->rcvq_space.time == 0) | |
508 | goto new_measure; | |
509 | ||
510 | time = tcp_time_stamp - tp->rcvq_space.time; | |
511 | if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0) | |
512 | return; | |
513 | ||
514 | space = 2 * (tp->copied_seq - tp->rcvq_space.seq); | |
515 | ||
516 | space = max(tp->rcvq_space.space, space); | |
517 | ||
518 | if (tp->rcvq_space.space != space) { | |
519 | int rcvmem; | |
520 | ||
521 | tp->rcvq_space.space = space; | |
522 | ||
523 | if (sysctl_tcp_moderate_rcvbuf && | |
524 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { | |
525 | int new_clamp = space; | |
526 | ||
527 | /* Receive space grows, normalize in order to | |
528 | * take into account packet headers and sk_buff | |
529 | * structure overhead. | |
530 | */ | |
531 | space /= tp->advmss; | |
532 | if (!space) | |
533 | space = 1; | |
534 | rcvmem = (tp->advmss + MAX_TCP_HEADER + | |
535 | 16 + sizeof(struct sk_buff)); | |
536 | while (tcp_win_from_space(rcvmem) < tp->advmss) | |
537 | rcvmem += 128; | |
538 | space *= rcvmem; | |
539 | space = min(space, sysctl_tcp_rmem[2]); | |
540 | if (space > sk->sk_rcvbuf) { | |
541 | sk->sk_rcvbuf = space; | |
542 | ||
543 | /* Make the window clamp follow along. */ | |
544 | tp->window_clamp = new_clamp; | |
545 | } | |
546 | } | |
547 | } | |
548 | ||
549 | new_measure: | |
550 | tp->rcvq_space.seq = tp->copied_seq; | |
551 | tp->rcvq_space.time = tcp_time_stamp; | |
552 | } | |
553 | ||
554 | /* There is something which you must keep in mind when you analyze the | |
555 | * behavior of the tp->ato delayed ack timeout interval. When a | |
556 | * connection starts up, we want to ack as quickly as possible. The | |
557 | * problem is that "good" TCP's do slow start at the beginning of data | |
558 | * transmission. The means that until we send the first few ACK's the | |
559 | * sender will sit on his end and only queue most of his data, because | |
560 | * he can only send snd_cwnd unacked packets at any given time. For | |
561 | * each ACK we send, he increments snd_cwnd and transmits more of his | |
562 | * queue. -DaveM | |
563 | */ | |
564 | static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) | |
565 | { | |
566 | struct tcp_sock *tp = tcp_sk(sk); | |
567 | struct inet_connection_sock *icsk = inet_csk(sk); | |
568 | u32 now; | |
569 | ||
570 | inet_csk_schedule_ack(sk); | |
571 | ||
572 | tcp_measure_rcv_mss(sk, skb); | |
573 | ||
574 | tcp_rcv_rtt_measure(tp); | |
575 | ||
576 | now = tcp_time_stamp; | |
577 | ||
578 | if (!icsk->icsk_ack.ato) { | |
579 | /* The _first_ data packet received, initialize | |
580 | * delayed ACK engine. | |
581 | */ | |
582 | tcp_incr_quickack(sk); | |
583 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
584 | } else { | |
585 | int m = now - icsk->icsk_ack.lrcvtime; | |
586 | ||
587 | if (m <= TCP_ATO_MIN / 2) { | |
588 | /* The fastest case is the first. */ | |
589 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; | |
590 | } else if (m < icsk->icsk_ack.ato) { | |
591 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; | |
592 | if (icsk->icsk_ack.ato > icsk->icsk_rto) | |
593 | icsk->icsk_ack.ato = icsk->icsk_rto; | |
594 | } else if (m > icsk->icsk_rto) { | |
595 | /* Too long gap. Apparently sender failed to | |
596 | * restart window, so that we send ACKs quickly. | |
597 | */ | |
598 | tcp_incr_quickack(sk); | |
599 | sk_mem_reclaim(sk); | |
600 | } | |
601 | } | |
602 | icsk->icsk_ack.lrcvtime = now; | |
603 | ||
604 | TCP_ECN_check_ce(tp, skb); | |
605 | ||
606 | if (skb->len >= 128) | |
607 | tcp_grow_window(sk, skb); | |
608 | } | |
609 | ||
610 | /* Called to compute a smoothed rtt estimate. The data fed to this | |
611 | * routine either comes from timestamps, or from segments that were | |
612 | * known _not_ to have been retransmitted [see Karn/Partridge | |
613 | * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 | |
614 | * piece by Van Jacobson. | |
615 | * NOTE: the next three routines used to be one big routine. | |
616 | * To save cycles in the RFC 1323 implementation it was better to break | |
617 | * it up into three procedures. -- erics | |
618 | */ | |
619 | static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) | |
620 | { | |
621 | struct tcp_sock *tp = tcp_sk(sk); | |
622 | long m = mrtt; /* RTT */ | |
623 | ||
624 | /* The following amusing code comes from Jacobson's | |
625 | * article in SIGCOMM '88. Note that rtt and mdev | |
626 | * are scaled versions of rtt and mean deviation. | |
627 | * This is designed to be as fast as possible | |
628 | * m stands for "measurement". | |
629 | * | |
630 | * On a 1990 paper the rto value is changed to: | |
631 | * RTO = rtt + 4 * mdev | |
632 | * | |
633 | * Funny. This algorithm seems to be very broken. | |
634 | * These formulae increase RTO, when it should be decreased, increase | |
635 | * too slowly, when it should be increased quickly, decrease too quickly | |
636 | * etc. I guess in BSD RTO takes ONE value, so that it is absolutely | |
637 | * does not matter how to _calculate_ it. Seems, it was trap | |
638 | * that VJ failed to avoid. 8) | |
639 | */ | |
640 | if (m == 0) | |
641 | m = 1; | |
642 | if (tp->srtt != 0) { | |
643 | m -= (tp->srtt >> 3); /* m is now error in rtt est */ | |
644 | tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ | |
645 | if (m < 0) { | |
646 | m = -m; /* m is now abs(error) */ | |
647 | m -= (tp->mdev >> 2); /* similar update on mdev */ | |
648 | /* This is similar to one of Eifel findings. | |
649 | * Eifel blocks mdev updates when rtt decreases. | |
650 | * This solution is a bit different: we use finer gain | |
651 | * for mdev in this case (alpha*beta). | |
652 | * Like Eifel it also prevents growth of rto, | |
653 | * but also it limits too fast rto decreases, | |
654 | * happening in pure Eifel. | |
655 | */ | |
656 | if (m > 0) | |
657 | m >>= 3; | |
658 | } else { | |
659 | m -= (tp->mdev >> 2); /* similar update on mdev */ | |
660 | } | |
661 | tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ | |
662 | if (tp->mdev > tp->mdev_max) { | |
663 | tp->mdev_max = tp->mdev; | |
664 | if (tp->mdev_max > tp->rttvar) | |
665 | tp->rttvar = tp->mdev_max; | |
666 | } | |
667 | if (after(tp->snd_una, tp->rtt_seq)) { | |
668 | if (tp->mdev_max < tp->rttvar) | |
669 | tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2; | |
670 | tp->rtt_seq = tp->snd_nxt; | |
671 | tp->mdev_max = tcp_rto_min(sk); | |
672 | } | |
673 | } else { | |
674 | /* no previous measure. */ | |
675 | tp->srtt = m << 3; /* take the measured time to be rtt */ | |
676 | tp->mdev = m << 1; /* make sure rto = 3*rtt */ | |
677 | tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); | |
678 | tp->rtt_seq = tp->snd_nxt; | |
679 | } | |
680 | } | |
681 | ||
682 | /* Calculate rto without backoff. This is the second half of Van Jacobson's | |
683 | * routine referred to above. | |
684 | */ | |
685 | static inline void tcp_set_rto(struct sock *sk) | |
686 | { | |
687 | const struct tcp_sock *tp = tcp_sk(sk); | |
688 | /* Old crap is replaced with new one. 8) | |
689 | * | |
690 | * More seriously: | |
691 | * 1. If rtt variance happened to be less 50msec, it is hallucination. | |
692 | * It cannot be less due to utterly erratic ACK generation made | |
693 | * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ | |
694 | * to do with delayed acks, because at cwnd>2 true delack timeout | |
695 | * is invisible. Actually, Linux-2.4 also generates erratic | |
696 | * ACKs in some circumstances. | |
697 | */ | |
698 | inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); | |
699 | ||
700 | /* 2. Fixups made earlier cannot be right. | |
701 | * If we do not estimate RTO correctly without them, | |
702 | * all the algo is pure shit and should be replaced | |
703 | * with correct one. It is exactly, which we pretend to do. | |
704 | */ | |
705 | ||
706 | /* NOTE: clamping at TCP_RTO_MIN is not required, current algo | |
707 | * guarantees that rto is higher. | |
708 | */ | |
709 | tcp_bound_rto(sk); | |
710 | } | |
711 | ||
712 | /* Save metrics learned by this TCP session. | |
713 | This function is called only, when TCP finishes successfully | |
714 | i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. | |
715 | */ | |
716 | void tcp_update_metrics(struct sock *sk) | |
717 | { | |
718 | struct tcp_sock *tp = tcp_sk(sk); | |
719 | struct dst_entry *dst = __sk_dst_get(sk); | |
720 | ||
721 | if (sysctl_tcp_nometrics_save) | |
722 | return; | |
723 | ||
724 | dst_confirm(dst); | |
725 | ||
726 | if (dst && (dst->flags & DST_HOST)) { | |
727 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
728 | int m; | |
729 | unsigned long rtt; | |
730 | ||
731 | if (icsk->icsk_backoff || !tp->srtt) { | |
732 | /* This session failed to estimate rtt. Why? | |
733 | * Probably, no packets returned in time. | |
734 | * Reset our results. | |
735 | */ | |
736 | if (!(dst_metric_locked(dst, RTAX_RTT))) | |
737 | dst->metrics[RTAX_RTT - 1] = 0; | |
738 | return; | |
739 | } | |
740 | ||
741 | rtt = dst_metric_rtt(dst, RTAX_RTT); | |
742 | m = rtt - tp->srtt; | |
743 | ||
744 | /* If newly calculated rtt larger than stored one, | |
745 | * store new one. Otherwise, use EWMA. Remember, | |
746 | * rtt overestimation is always better than underestimation. | |
747 | */ | |
748 | if (!(dst_metric_locked(dst, RTAX_RTT))) { | |
749 | if (m <= 0) | |
750 | set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt); | |
751 | else | |
752 | set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3)); | |
753 | } | |
754 | ||
755 | if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { | |
756 | unsigned long var; | |
757 | if (m < 0) | |
758 | m = -m; | |
759 | ||
760 | /* Scale deviation to rttvar fixed point */ | |
761 | m >>= 1; | |
762 | if (m < tp->mdev) | |
763 | m = tp->mdev; | |
764 | ||
765 | var = dst_metric_rtt(dst, RTAX_RTTVAR); | |
766 | if (m >= var) | |
767 | var = m; | |
768 | else | |
769 | var -= (var - m) >> 2; | |
770 | ||
771 | set_dst_metric_rtt(dst, RTAX_RTTVAR, var); | |
772 | } | |
773 | ||
774 | if (tcp_in_initial_slowstart(tp)) { | |
775 | /* Slow start still did not finish. */ | |
776 | if (dst_metric(dst, RTAX_SSTHRESH) && | |
777 | !dst_metric_locked(dst, RTAX_SSTHRESH) && | |
778 | (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) | |
779 | dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1; | |
780 | if (!dst_metric_locked(dst, RTAX_CWND) && | |
781 | tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) | |
782 | dst->metrics[RTAX_CWND - 1] = tp->snd_cwnd; | |
783 | } else if (tp->snd_cwnd > tp->snd_ssthresh && | |
784 | icsk->icsk_ca_state == TCP_CA_Open) { | |
785 | /* Cong. avoidance phase, cwnd is reliable. */ | |
786 | if (!dst_metric_locked(dst, RTAX_SSTHRESH)) | |
787 | dst->metrics[RTAX_SSTHRESH-1] = | |
788 | max(tp->snd_cwnd >> 1, tp->snd_ssthresh); | |
789 | if (!dst_metric_locked(dst, RTAX_CWND)) | |
790 | dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_cwnd) >> 1; | |
791 | } else { | |
792 | /* Else slow start did not finish, cwnd is non-sense, | |
793 | ssthresh may be also invalid. | |
794 | */ | |
795 | if (!dst_metric_locked(dst, RTAX_CWND)) | |
796 | dst->metrics[RTAX_CWND-1] = (dst_metric(dst, RTAX_CWND) + tp->snd_ssthresh) >> 1; | |
797 | if (dst_metric(dst, RTAX_SSTHRESH) && | |
798 | !dst_metric_locked(dst, RTAX_SSTHRESH) && | |
799 | tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH)) | |
800 | dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh; | |
801 | } | |
802 | ||
803 | if (!dst_metric_locked(dst, RTAX_REORDERING)) { | |
804 | if (dst_metric(dst, RTAX_REORDERING) < tp->reordering && | |
805 | tp->reordering != sysctl_tcp_reordering) | |
806 | dst->metrics[RTAX_REORDERING-1] = tp->reordering; | |
807 | } | |
808 | } | |
809 | } | |
810 | ||
811 | __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst) | |
812 | { | |
813 | __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); | |
814 | ||
815 | if (!cwnd) | |
816 | cwnd = rfc3390_bytes_to_packets(tp->mss_cache); | |
817 | return min_t(__u32, cwnd, tp->snd_cwnd_clamp); | |
818 | } | |
819 | ||
820 | /* Set slow start threshold and cwnd not falling to slow start */ | |
821 | void tcp_enter_cwr(struct sock *sk, const int set_ssthresh) | |
822 | { | |
823 | struct tcp_sock *tp = tcp_sk(sk); | |
824 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
825 | ||
826 | tp->prior_ssthresh = 0; | |
827 | tp->bytes_acked = 0; | |
828 | if (icsk->icsk_ca_state < TCP_CA_CWR) { | |
829 | tp->undo_marker = 0; | |
830 | if (set_ssthresh) | |
831 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
832 | tp->snd_cwnd = min(tp->snd_cwnd, | |
833 | tcp_packets_in_flight(tp) + 1U); | |
834 | tp->snd_cwnd_cnt = 0; | |
835 | tp->high_seq = tp->snd_nxt; | |
836 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
837 | TCP_ECN_queue_cwr(tp); | |
838 | ||
839 | tcp_set_ca_state(sk, TCP_CA_CWR); | |
840 | } | |
841 | } | |
842 | ||
843 | /* | |
844 | * Packet counting of FACK is based on in-order assumptions, therefore TCP | |
845 | * disables it when reordering is detected | |
846 | */ | |
847 | static void tcp_disable_fack(struct tcp_sock *tp) | |
848 | { | |
849 | /* RFC3517 uses different metric in lost marker => reset on change */ | |
850 | if (tcp_is_fack(tp)) | |
851 | tp->lost_skb_hint = NULL; | |
852 | tp->rx_opt.sack_ok &= ~2; | |
853 | } | |
854 | ||
855 | /* Take a notice that peer is sending D-SACKs */ | |
856 | static void tcp_dsack_seen(struct tcp_sock *tp) | |
857 | { | |
858 | tp->rx_opt.sack_ok |= 4; | |
859 | } | |
860 | ||
861 | /* Initialize metrics on socket. */ | |
862 | ||
863 | static void tcp_init_metrics(struct sock *sk) | |
864 | { | |
865 | struct tcp_sock *tp = tcp_sk(sk); | |
866 | struct dst_entry *dst = __sk_dst_get(sk); | |
867 | ||
868 | if (dst == NULL) | |
869 | goto reset; | |
870 | ||
871 | dst_confirm(dst); | |
872 | ||
873 | if (dst_metric_locked(dst, RTAX_CWND)) | |
874 | tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); | |
875 | if (dst_metric(dst, RTAX_SSTHRESH)) { | |
876 | tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); | |
877 | if (tp->snd_ssthresh > tp->snd_cwnd_clamp) | |
878 | tp->snd_ssthresh = tp->snd_cwnd_clamp; | |
879 | } | |
880 | if (dst_metric(dst, RTAX_REORDERING) && | |
881 | tp->reordering != dst_metric(dst, RTAX_REORDERING)) { | |
882 | tcp_disable_fack(tp); | |
883 | tp->reordering = dst_metric(dst, RTAX_REORDERING); | |
884 | } | |
885 | ||
886 | if (dst_metric(dst, RTAX_RTT) == 0) | |
887 | goto reset; | |
888 | ||
889 | if (!tp->srtt && dst_metric_rtt(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3)) | |
890 | goto reset; | |
891 | ||
892 | /* Initial rtt is determined from SYN,SYN-ACK. | |
893 | * The segment is small and rtt may appear much | |
894 | * less than real one. Use per-dst memory | |
895 | * to make it more realistic. | |
896 | * | |
897 | * A bit of theory. RTT is time passed after "normal" sized packet | |
898 | * is sent until it is ACKed. In normal circumstances sending small | |
899 | * packets force peer to delay ACKs and calculation is correct too. | |
900 | * The algorithm is adaptive and, provided we follow specs, it | |
901 | * NEVER underestimate RTT. BUT! If peer tries to make some clever | |
902 | * tricks sort of "quick acks" for time long enough to decrease RTT | |
903 | * to low value, and then abruptly stops to do it and starts to delay | |
904 | * ACKs, wait for troubles. | |
905 | */ | |
906 | if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) { | |
907 | tp->srtt = dst_metric_rtt(dst, RTAX_RTT); | |
908 | tp->rtt_seq = tp->snd_nxt; | |
909 | } | |
910 | if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) { | |
911 | tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR); | |
912 | tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); | |
913 | } | |
914 | tcp_set_rto(sk); | |
915 | if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp) | |
916 | goto reset; | |
917 | ||
918 | cwnd: | |
919 | tp->snd_cwnd = tcp_init_cwnd(tp, dst); | |
920 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
921 | return; | |
922 | ||
923 | reset: | |
924 | /* Play conservative. If timestamps are not | |
925 | * supported, TCP will fail to recalculate correct | |
926 | * rtt, if initial rto is too small. FORGET ALL AND RESET! | |
927 | */ | |
928 | if (!tp->rx_opt.saw_tstamp && tp->srtt) { | |
929 | tp->srtt = 0; | |
930 | tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT; | |
931 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; | |
932 | } | |
933 | goto cwnd; | |
934 | } | |
935 | ||
936 | static void tcp_update_reordering(struct sock *sk, const int metric, | |
937 | const int ts) | |
938 | { | |
939 | struct tcp_sock *tp = tcp_sk(sk); | |
940 | if (metric > tp->reordering) { | |
941 | int mib_idx; | |
942 | ||
943 | tp->reordering = min(TCP_MAX_REORDERING, metric); | |
944 | ||
945 | /* This exciting event is worth to be remembered. 8) */ | |
946 | if (ts) | |
947 | mib_idx = LINUX_MIB_TCPTSREORDER; | |
948 | else if (tcp_is_reno(tp)) | |
949 | mib_idx = LINUX_MIB_TCPRENOREORDER; | |
950 | else if (tcp_is_fack(tp)) | |
951 | mib_idx = LINUX_MIB_TCPFACKREORDER; | |
952 | else | |
953 | mib_idx = LINUX_MIB_TCPSACKREORDER; | |
954 | ||
955 | NET_INC_STATS_BH(sock_net(sk), mib_idx); | |
956 | #if FASTRETRANS_DEBUG > 1 | |
957 | printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", | |
958 | tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, | |
959 | tp->reordering, | |
960 | tp->fackets_out, | |
961 | tp->sacked_out, | |
962 | tp->undo_marker ? tp->undo_retrans : 0); | |
963 | #endif | |
964 | tcp_disable_fack(tp); | |
965 | } | |
966 | } | |
967 | ||
968 | /* This must be called before lost_out is incremented */ | |
969 | static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) | |
970 | { | |
971 | if ((tp->retransmit_skb_hint == NULL) || | |
972 | before(TCP_SKB_CB(skb)->seq, | |
973 | TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) | |
974 | tp->retransmit_skb_hint = skb; | |
975 | ||
976 | if (!tp->lost_out || | |
977 | after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high)) | |
978 | tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; | |
979 | } | |
980 | ||
981 | static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb) | |
982 | { | |
983 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { | |
984 | tcp_verify_retransmit_hint(tp, skb); | |
985 | ||
986 | tp->lost_out += tcp_skb_pcount(skb); | |
987 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
988 | } | |
989 | } | |
990 | ||
991 | static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, | |
992 | struct sk_buff *skb) | |
993 | { | |
994 | tcp_verify_retransmit_hint(tp, skb); | |
995 | ||
996 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { | |
997 | tp->lost_out += tcp_skb_pcount(skb); | |
998 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
999 | } | |
1000 | } | |
1001 | ||
1002 | /* This procedure tags the retransmission queue when SACKs arrive. | |
1003 | * | |
1004 | * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). | |
1005 | * Packets in queue with these bits set are counted in variables | |
1006 | * sacked_out, retrans_out and lost_out, correspondingly. | |
1007 | * | |
1008 | * Valid combinations are: | |
1009 | * Tag InFlight Description | |
1010 | * 0 1 - orig segment is in flight. | |
1011 | * S 0 - nothing flies, orig reached receiver. | |
1012 | * L 0 - nothing flies, orig lost by net. | |
1013 | * R 2 - both orig and retransmit are in flight. | |
1014 | * L|R 1 - orig is lost, retransmit is in flight. | |
1015 | * S|R 1 - orig reached receiver, retrans is still in flight. | |
1016 | * (L|S|R is logically valid, it could occur when L|R is sacked, | |
1017 | * but it is equivalent to plain S and code short-curcuits it to S. | |
1018 | * L|S is logically invalid, it would mean -1 packet in flight 8)) | |
1019 | * | |
1020 | * These 6 states form finite state machine, controlled by the following events: | |
1021 | * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) | |
1022 | * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) | |
1023 | * 3. Loss detection event of one of three flavors: | |
1024 | * A. Scoreboard estimator decided the packet is lost. | |
1025 | * A'. Reno "three dupacks" marks head of queue lost. | |
1026 | * A''. Its FACK modfication, head until snd.fack is lost. | |
1027 | * B. SACK arrives sacking data transmitted after never retransmitted | |
1028 | * hole was sent out. | |
1029 | * C. SACK arrives sacking SND.NXT at the moment, when the | |
1030 | * segment was retransmitted. | |
1031 | * 4. D-SACK added new rule: D-SACK changes any tag to S. | |
1032 | * | |
1033 | * It is pleasant to note, that state diagram turns out to be commutative, | |
1034 | * so that we are allowed not to be bothered by order of our actions, | |
1035 | * when multiple events arrive simultaneously. (see the function below). | |
1036 | * | |
1037 | * Reordering detection. | |
1038 | * -------------------- | |
1039 | * Reordering metric is maximal distance, which a packet can be displaced | |
1040 | * in packet stream. With SACKs we can estimate it: | |
1041 | * | |
1042 | * 1. SACK fills old hole and the corresponding segment was not | |
1043 | * ever retransmitted -> reordering. Alas, we cannot use it | |
1044 | * when segment was retransmitted. | |
1045 | * 2. The last flaw is solved with D-SACK. D-SACK arrives | |
1046 | * for retransmitted and already SACKed segment -> reordering.. | |
1047 | * Both of these heuristics are not used in Loss state, when we cannot | |
1048 | * account for retransmits accurately. | |
1049 | * | |
1050 | * SACK block validation. | |
1051 | * ---------------------- | |
1052 | * | |
1053 | * SACK block range validation checks that the received SACK block fits to | |
1054 | * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. | |
1055 | * Note that SND.UNA is not included to the range though being valid because | |
1056 | * it means that the receiver is rather inconsistent with itself reporting | |
1057 | * SACK reneging when it should advance SND.UNA. Such SACK block this is | |
1058 | * perfectly valid, however, in light of RFC2018 which explicitly states | |
1059 | * that "SACK block MUST reflect the newest segment. Even if the newest | |
1060 | * segment is going to be discarded ...", not that it looks very clever | |
1061 | * in case of head skb. Due to potentional receiver driven attacks, we | |
1062 | * choose to avoid immediate execution of a walk in write queue due to | |
1063 | * reneging and defer head skb's loss recovery to standard loss recovery | |
1064 | * procedure that will eventually trigger (nothing forbids us doing this). | |
1065 | * | |
1066 | * Implements also blockage to start_seq wrap-around. Problem lies in the | |
1067 | * fact that though start_seq (s) is before end_seq (i.e., not reversed), | |
1068 | * there's no guarantee that it will be before snd_nxt (n). The problem | |
1069 | * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt | |
1070 | * wrap (s_w): | |
1071 | * | |
1072 | * <- outs wnd -> <- wrapzone -> | |
1073 | * u e n u_w e_w s n_w | |
1074 | * | | | | | | | | |
1075 | * |<------------+------+----- TCP seqno space --------------+---------->| | |
1076 | * ...-- <2^31 ->| |<--------... | |
1077 | * ...---- >2^31 ------>| |<--------... | |
1078 | * | |
1079 | * Current code wouldn't be vulnerable but it's better still to discard such | |
1080 | * crazy SACK blocks. Doing this check for start_seq alone closes somewhat | |
1081 | * similar case (end_seq after snd_nxt wrap) as earlier reversed check in | |
1082 | * snd_nxt wrap -> snd_una region will then become "well defined", i.e., | |
1083 | * equal to the ideal case (infinite seqno space without wrap caused issues). | |
1084 | * | |
1085 | * With D-SACK the lower bound is extended to cover sequence space below | |
1086 | * SND.UNA down to undo_marker, which is the last point of interest. Yet | |
1087 | * again, D-SACK block must not to go across snd_una (for the same reason as | |
1088 | * for the normal SACK blocks, explained above). But there all simplicity | |
1089 | * ends, TCP might receive valid D-SACKs below that. As long as they reside | |
1090 | * fully below undo_marker they do not affect behavior in anyway and can | |
1091 | * therefore be safely ignored. In rare cases (which are more or less | |
1092 | * theoretical ones), the D-SACK will nicely cross that boundary due to skb | |
1093 | * fragmentation and packet reordering past skb's retransmission. To consider | |
1094 | * them correctly, the acceptable range must be extended even more though | |
1095 | * the exact amount is rather hard to quantify. However, tp->max_window can | |
1096 | * be used as an exaggerated estimate. | |
1097 | */ | |
1098 | static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack, | |
1099 | u32 start_seq, u32 end_seq) | |
1100 | { | |
1101 | /* Too far in future, or reversed (interpretation is ambiguous) */ | |
1102 | if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) | |
1103 | return 0; | |
1104 | ||
1105 | /* Nasty start_seq wrap-around check (see comments above) */ | |
1106 | if (!before(start_seq, tp->snd_nxt)) | |
1107 | return 0; | |
1108 | ||
1109 | /* In outstanding window? ...This is valid exit for D-SACKs too. | |
1110 | * start_seq == snd_una is non-sensical (see comments above) | |
1111 | */ | |
1112 | if (after(start_seq, tp->snd_una)) | |
1113 | return 1; | |
1114 | ||
1115 | if (!is_dsack || !tp->undo_marker) | |
1116 | return 0; | |
1117 | ||
1118 | /* ...Then it's D-SACK, and must reside below snd_una completely */ | |
1119 | if (!after(end_seq, tp->snd_una)) | |
1120 | return 0; | |
1121 | ||
1122 | if (!before(start_seq, tp->undo_marker)) | |
1123 | return 1; | |
1124 | ||
1125 | /* Too old */ | |
1126 | if (!after(end_seq, tp->undo_marker)) | |
1127 | return 0; | |
1128 | ||
1129 | /* Undo_marker boundary crossing (overestimates a lot). Known already: | |
1130 | * start_seq < undo_marker and end_seq >= undo_marker. | |
1131 | */ | |
1132 | return !before(start_seq, end_seq - tp->max_window); | |
1133 | } | |
1134 | ||
1135 | /* Check for lost retransmit. This superb idea is borrowed from "ratehalving". | |
1136 | * Event "C". Later note: FACK people cheated me again 8), we have to account | |
1137 | * for reordering! Ugly, but should help. | |
1138 | * | |
1139 | * Search retransmitted skbs from write_queue that were sent when snd_nxt was | |
1140 | * less than what is now known to be received by the other end (derived from | |
1141 | * highest SACK block). Also calculate the lowest snd_nxt among the remaining | |
1142 | * retransmitted skbs to avoid some costly processing per ACKs. | |
1143 | */ | |
1144 | static void tcp_mark_lost_retrans(struct sock *sk) | |
1145 | { | |
1146 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
1147 | struct tcp_sock *tp = tcp_sk(sk); | |
1148 | struct sk_buff *skb; | |
1149 | int cnt = 0; | |
1150 | u32 new_low_seq = tp->snd_nxt; | |
1151 | u32 received_upto = tcp_highest_sack_seq(tp); | |
1152 | ||
1153 | if (!tcp_is_fack(tp) || !tp->retrans_out || | |
1154 | !after(received_upto, tp->lost_retrans_low) || | |
1155 | icsk->icsk_ca_state != TCP_CA_Recovery) | |
1156 | return; | |
1157 | ||
1158 | tcp_for_write_queue(skb, sk) { | |
1159 | u32 ack_seq = TCP_SKB_CB(skb)->ack_seq; | |
1160 | ||
1161 | if (skb == tcp_send_head(sk)) | |
1162 | break; | |
1163 | if (cnt == tp->retrans_out) | |
1164 | break; | |
1165 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1166 | continue; | |
1167 | ||
1168 | if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) | |
1169 | continue; | |
1170 | ||
1171 | /* TODO: We would like to get rid of tcp_is_fack(tp) only | |
1172 | * constraint here (see above) but figuring out that at | |
1173 | * least tp->reordering SACK blocks reside between ack_seq | |
1174 | * and received_upto is not easy task to do cheaply with | |
1175 | * the available datastructures. | |
1176 | * | |
1177 | * Whether FACK should check here for tp->reordering segs | |
1178 | * in-between one could argue for either way (it would be | |
1179 | * rather simple to implement as we could count fack_count | |
1180 | * during the walk and do tp->fackets_out - fack_count). | |
1181 | */ | |
1182 | if (after(received_upto, ack_seq)) { | |
1183 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; | |
1184 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1185 | ||
1186 | tcp_skb_mark_lost_uncond_verify(tp, skb); | |
1187 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT); | |
1188 | } else { | |
1189 | if (before(ack_seq, new_low_seq)) | |
1190 | new_low_seq = ack_seq; | |
1191 | cnt += tcp_skb_pcount(skb); | |
1192 | } | |
1193 | } | |
1194 | ||
1195 | if (tp->retrans_out) | |
1196 | tp->lost_retrans_low = new_low_seq; | |
1197 | } | |
1198 | ||
1199 | static int tcp_check_dsack(struct sock *sk, struct sk_buff *ack_skb, | |
1200 | struct tcp_sack_block_wire *sp, int num_sacks, | |
1201 | u32 prior_snd_una) | |
1202 | { | |
1203 | struct tcp_sock *tp = tcp_sk(sk); | |
1204 | u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); | |
1205 | u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); | |
1206 | int dup_sack = 0; | |
1207 | ||
1208 | if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { | |
1209 | dup_sack = 1; | |
1210 | tcp_dsack_seen(tp); | |
1211 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV); | |
1212 | } else if (num_sacks > 1) { | |
1213 | u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); | |
1214 | u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); | |
1215 | ||
1216 | if (!after(end_seq_0, end_seq_1) && | |
1217 | !before(start_seq_0, start_seq_1)) { | |
1218 | dup_sack = 1; | |
1219 | tcp_dsack_seen(tp); | |
1220 | NET_INC_STATS_BH(sock_net(sk), | |
1221 | LINUX_MIB_TCPDSACKOFORECV); | |
1222 | } | |
1223 | } | |
1224 | ||
1225 | /* D-SACK for already forgotten data... Do dumb counting. */ | |
1226 | if (dup_sack && | |
1227 | !after(end_seq_0, prior_snd_una) && | |
1228 | after(end_seq_0, tp->undo_marker)) | |
1229 | tp->undo_retrans--; | |
1230 | ||
1231 | return dup_sack; | |
1232 | } | |
1233 | ||
1234 | struct tcp_sacktag_state { | |
1235 | int reord; | |
1236 | int fack_count; | |
1237 | int flag; | |
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 | * FIXME: this could be merged to shift decision code | |
1247 | */ | |
1248 | static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, | |
1249 | u32 start_seq, u32 end_seq) | |
1250 | { | |
1251 | int in_sack, err; | |
1252 | unsigned int pkt_len; | |
1253 | unsigned int mss; | |
1254 | ||
1255 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && | |
1256 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1257 | ||
1258 | if (tcp_skb_pcount(skb) > 1 && !in_sack && | |
1259 | after(TCP_SKB_CB(skb)->end_seq, start_seq)) { | |
1260 | mss = tcp_skb_mss(skb); | |
1261 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); | |
1262 | ||
1263 | if (!in_sack) { | |
1264 | pkt_len = start_seq - TCP_SKB_CB(skb)->seq; | |
1265 | if (pkt_len < mss) | |
1266 | pkt_len = mss; | |
1267 | } else { | |
1268 | pkt_len = end_seq - TCP_SKB_CB(skb)->seq; | |
1269 | if (pkt_len < mss) | |
1270 | return -EINVAL; | |
1271 | } | |
1272 | ||
1273 | /* Round if necessary so that SACKs cover only full MSSes | |
1274 | * and/or the remaining small portion (if present) | |
1275 | */ | |
1276 | if (pkt_len > mss) { | |
1277 | unsigned int new_len = (pkt_len / mss) * mss; | |
1278 | if (!in_sack && new_len < pkt_len) { | |
1279 | new_len += mss; | |
1280 | if (new_len > skb->len) | |
1281 | return 0; | |
1282 | } | |
1283 | pkt_len = new_len; | |
1284 | } | |
1285 | err = tcp_fragment(sk, skb, pkt_len, mss); | |
1286 | if (err < 0) | |
1287 | return err; | |
1288 | } | |
1289 | ||
1290 | return in_sack; | |
1291 | } | |
1292 | ||
1293 | static u8 tcp_sacktag_one(struct sk_buff *skb, struct sock *sk, | |
1294 | struct tcp_sacktag_state *state, | |
1295 | int dup_sack, int pcount) | |
1296 | { | |
1297 | struct tcp_sock *tp = tcp_sk(sk); | |
1298 | u8 sacked = TCP_SKB_CB(skb)->sacked; | |
1299 | int fack_count = state->fack_count; | |
1300 | ||
1301 | /* Account D-SACK for retransmitted packet. */ | |
1302 | if (dup_sack && (sacked & TCPCB_RETRANS)) { | |
1303 | if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker)) | |
1304 | tp->undo_retrans--; | |
1305 | if (sacked & TCPCB_SACKED_ACKED) | |
1306 | state->reord = min(fack_count, state->reord); | |
1307 | } | |
1308 | ||
1309 | /* Nothing to do; acked frame is about to be dropped (was ACKed). */ | |
1310 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1311 | return sacked; | |
1312 | ||
1313 | if (!(sacked & TCPCB_SACKED_ACKED)) { | |
1314 | if (sacked & TCPCB_SACKED_RETRANS) { | |
1315 | /* If the segment is not tagged as lost, | |
1316 | * we do not clear RETRANS, believing | |
1317 | * that retransmission is still in flight. | |
1318 | */ | |
1319 | if (sacked & TCPCB_LOST) { | |
1320 | sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); | |
1321 | tp->lost_out -= pcount; | |
1322 | tp->retrans_out -= pcount; | |
1323 | } | |
1324 | } else { | |
1325 | if (!(sacked & TCPCB_RETRANS)) { | |
1326 | /* New sack for not retransmitted frame, | |
1327 | * which was in hole. It is reordering. | |
1328 | */ | |
1329 | if (before(TCP_SKB_CB(skb)->seq, | |
1330 | tcp_highest_sack_seq(tp))) | |
1331 | state->reord = min(fack_count, | |
1332 | state->reord); | |
1333 | ||
1334 | /* SACK enhanced F-RTO (RFC4138; Appendix B) */ | |
1335 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) | |
1336 | state->flag |= FLAG_ONLY_ORIG_SACKED; | |
1337 | } | |
1338 | ||
1339 | if (sacked & TCPCB_LOST) { | |
1340 | sacked &= ~TCPCB_LOST; | |
1341 | tp->lost_out -= pcount; | |
1342 | } | |
1343 | } | |
1344 | ||
1345 | sacked |= TCPCB_SACKED_ACKED; | |
1346 | state->flag |= FLAG_DATA_SACKED; | |
1347 | tp->sacked_out += pcount; | |
1348 | ||
1349 | fack_count += pcount; | |
1350 | ||
1351 | /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ | |
1352 | if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) && | |
1353 | before(TCP_SKB_CB(skb)->seq, | |
1354 | TCP_SKB_CB(tp->lost_skb_hint)->seq)) | |
1355 | tp->lost_cnt_hint += pcount; | |
1356 | ||
1357 | if (fack_count > tp->fackets_out) | |
1358 | tp->fackets_out = fack_count; | |
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 && (sacked & TCPCB_SACKED_RETRANS)) { | |
1366 | sacked &= ~TCPCB_SACKED_RETRANS; | |
1367 | tp->retrans_out -= pcount; | |
1368 | } | |
1369 | ||
1370 | return sacked; | |
1371 | } | |
1372 | ||
1373 | static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb, | |
1374 | struct tcp_sacktag_state *state, | |
1375 | unsigned int pcount, int shifted, int mss, | |
1376 | int dup_sack) | |
1377 | { | |
1378 | struct tcp_sock *tp = tcp_sk(sk); | |
1379 | struct sk_buff *prev = tcp_write_queue_prev(sk, skb); | |
1380 | ||
1381 | BUG_ON(!pcount); | |
1382 | ||
1383 | /* Tweak before seqno plays */ | |
1384 | if (!tcp_is_fack(tp) && tcp_is_sack(tp) && tp->lost_skb_hint && | |
1385 | !before(TCP_SKB_CB(tp->lost_skb_hint)->seq, TCP_SKB_CB(skb)->seq)) | |
1386 | tp->lost_cnt_hint += pcount; | |
1387 | ||
1388 | TCP_SKB_CB(prev)->end_seq += shifted; | |
1389 | TCP_SKB_CB(skb)->seq += shifted; | |
1390 | ||
1391 | skb_shinfo(prev)->gso_segs += pcount; | |
1392 | BUG_ON(skb_shinfo(skb)->gso_segs < pcount); | |
1393 | skb_shinfo(skb)->gso_segs -= pcount; | |
1394 | ||
1395 | /* When we're adding to gso_segs == 1, gso_size will be zero, | |
1396 | * in theory this shouldn't be necessary but as long as DSACK | |
1397 | * code can come after this skb later on it's better to keep | |
1398 | * setting gso_size to something. | |
1399 | */ | |
1400 | if (!skb_shinfo(prev)->gso_size) { | |
1401 | skb_shinfo(prev)->gso_size = mss; | |
1402 | skb_shinfo(prev)->gso_type = sk->sk_gso_type; | |
1403 | } | |
1404 | ||
1405 | /* CHECKME: To clear or not to clear? Mimics normal skb currently */ | |
1406 | if (skb_shinfo(skb)->gso_segs <= 1) { | |
1407 | skb_shinfo(skb)->gso_size = 0; | |
1408 | skb_shinfo(skb)->gso_type = 0; | |
1409 | } | |
1410 | ||
1411 | /* We discard results */ | |
1412 | tcp_sacktag_one(skb, sk, state, dup_sack, pcount); | |
1413 | ||
1414 | /* Difference in this won't matter, both ACKed by the same cumul. ACK */ | |
1415 | TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); | |
1416 | ||
1417 | if (skb->len > 0) { | |
1418 | BUG_ON(!tcp_skb_pcount(skb)); | |
1419 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED); | |
1420 | return 0; | |
1421 | } | |
1422 | ||
1423 | /* Whole SKB was eaten :-) */ | |
1424 | ||
1425 | if (skb == tp->retransmit_skb_hint) | |
1426 | tp->retransmit_skb_hint = prev; | |
1427 | if (skb == tp->scoreboard_skb_hint) | |
1428 | tp->scoreboard_skb_hint = prev; | |
1429 | if (skb == tp->lost_skb_hint) { | |
1430 | tp->lost_skb_hint = prev; | |
1431 | tp->lost_cnt_hint -= tcp_skb_pcount(prev); | |
1432 | } | |
1433 | ||
1434 | TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(prev)->flags; | |
1435 | if (skb == tcp_highest_sack(sk)) | |
1436 | tcp_advance_highest_sack(sk, skb); | |
1437 | ||
1438 | tcp_unlink_write_queue(skb, sk); | |
1439 | sk_wmem_free_skb(sk, skb); | |
1440 | ||
1441 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED); | |
1442 | ||
1443 | return 1; | |
1444 | } | |
1445 | ||
1446 | /* I wish gso_size would have a bit more sane initialization than | |
1447 | * something-or-zero which complicates things | |
1448 | */ | |
1449 | static int tcp_skb_seglen(struct sk_buff *skb) | |
1450 | { | |
1451 | return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); | |
1452 | } | |
1453 | ||
1454 | /* Shifting pages past head area doesn't work */ | |
1455 | static int skb_can_shift(struct sk_buff *skb) | |
1456 | { | |
1457 | return !skb_headlen(skb) && skb_is_nonlinear(skb); | |
1458 | } | |
1459 | ||
1460 | /* Try collapsing SACK blocks spanning across multiple skbs to a single | |
1461 | * skb. | |
1462 | */ | |
1463 | static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, | |
1464 | struct tcp_sacktag_state *state, | |
1465 | u32 start_seq, u32 end_seq, | |
1466 | int dup_sack) | |
1467 | { | |
1468 | struct tcp_sock *tp = tcp_sk(sk); | |
1469 | struct sk_buff *prev; | |
1470 | int mss; | |
1471 | int pcount = 0; | |
1472 | int len; | |
1473 | int in_sack; | |
1474 | ||
1475 | if (!sk_can_gso(sk)) | |
1476 | goto fallback; | |
1477 | ||
1478 | /* Normally R but no L won't result in plain S */ | |
1479 | if (!dup_sack && | |
1480 | (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) | |
1481 | goto fallback; | |
1482 | if (!skb_can_shift(skb)) | |
1483 | goto fallback; | |
1484 | /* This frame is about to be dropped (was ACKed). */ | |
1485 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1486 | goto fallback; | |
1487 | ||
1488 | /* Can only happen with delayed DSACK + discard craziness */ | |
1489 | if (unlikely(skb == tcp_write_queue_head(sk))) | |
1490 | goto fallback; | |
1491 | prev = tcp_write_queue_prev(sk, skb); | |
1492 | ||
1493 | if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) | |
1494 | goto fallback; | |
1495 | ||
1496 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && | |
1497 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1498 | ||
1499 | if (in_sack) { | |
1500 | len = skb->len; | |
1501 | pcount = tcp_skb_pcount(skb); | |
1502 | mss = tcp_skb_seglen(skb); | |
1503 | ||
1504 | /* TODO: Fix DSACKs to not fragment already SACKed and we can | |
1505 | * drop this restriction as unnecessary | |
1506 | */ | |
1507 | if (mss != tcp_skb_seglen(prev)) | |
1508 | goto fallback; | |
1509 | } else { | |
1510 | if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) | |
1511 | goto noop; | |
1512 | /* CHECKME: This is non-MSS split case only?, this will | |
1513 | * cause skipped skbs due to advancing loop btw, original | |
1514 | * has that feature too | |
1515 | */ | |
1516 | if (tcp_skb_pcount(skb) <= 1) | |
1517 | goto noop; | |
1518 | ||
1519 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); | |
1520 | if (!in_sack) { | |
1521 | /* TODO: head merge to next could be attempted here | |
1522 | * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), | |
1523 | * though it might not be worth of the additional hassle | |
1524 | * | |
1525 | * ...we can probably just fallback to what was done | |
1526 | * previously. We could try merging non-SACKed ones | |
1527 | * as well but it probably isn't going to buy off | |
1528 | * because later SACKs might again split them, and | |
1529 | * it would make skb timestamp tracking considerably | |
1530 | * harder problem. | |
1531 | */ | |
1532 | goto fallback; | |
1533 | } | |
1534 | ||
1535 | len = end_seq - TCP_SKB_CB(skb)->seq; | |
1536 | BUG_ON(len < 0); | |
1537 | BUG_ON(len > skb->len); | |
1538 | ||
1539 | /* MSS boundaries should be honoured or else pcount will | |
1540 | * severely break even though it makes things bit trickier. | |
1541 | * Optimize common case to avoid most of the divides | |
1542 | */ | |
1543 | mss = tcp_skb_mss(skb); | |
1544 | ||
1545 | /* TODO: Fix DSACKs to not fragment already SACKed and we can | |
1546 | * drop this restriction as unnecessary | |
1547 | */ | |
1548 | if (mss != tcp_skb_seglen(prev)) | |
1549 | goto fallback; | |
1550 | ||
1551 | if (len == mss) { | |
1552 | pcount = 1; | |
1553 | } else if (len < mss) { | |
1554 | goto noop; | |
1555 | } else { | |
1556 | pcount = len / mss; | |
1557 | len = pcount * mss; | |
1558 | } | |
1559 | } | |
1560 | ||
1561 | if (!skb_shift(prev, skb, len)) | |
1562 | goto fallback; | |
1563 | if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack)) | |
1564 | goto out; | |
1565 | ||
1566 | /* Hole filled allows collapsing with the next as well, this is very | |
1567 | * useful when hole on every nth skb pattern happens | |
1568 | */ | |
1569 | if (prev == tcp_write_queue_tail(sk)) | |
1570 | goto out; | |
1571 | skb = tcp_write_queue_next(sk, prev); | |
1572 | ||
1573 | if (!skb_can_shift(skb) || | |
1574 | (skb == tcp_send_head(sk)) || | |
1575 | ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || | |
1576 | (mss != tcp_skb_seglen(skb))) | |
1577 | goto out; | |
1578 | ||
1579 | len = skb->len; | |
1580 | if (skb_shift(prev, skb, len)) { | |
1581 | pcount += tcp_skb_pcount(skb); | |
1582 | tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0); | |
1583 | } | |
1584 | ||
1585 | out: | |
1586 | state->fack_count += pcount; | |
1587 | return prev; | |
1588 | ||
1589 | noop: | |
1590 | return skb; | |
1591 | ||
1592 | fallback: | |
1593 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); | |
1594 | return NULL; | |
1595 | } | |
1596 | ||
1597 | static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, | |
1598 | struct tcp_sack_block *next_dup, | |
1599 | struct tcp_sacktag_state *state, | |
1600 | u32 start_seq, u32 end_seq, | |
1601 | int dup_sack_in) | |
1602 | { | |
1603 | struct tcp_sock *tp = tcp_sk(sk); | |
1604 | struct sk_buff *tmp; | |
1605 | ||
1606 | tcp_for_write_queue_from(skb, sk) { | |
1607 | int in_sack = 0; | |
1608 | int dup_sack = dup_sack_in; | |
1609 | ||
1610 | if (skb == tcp_send_head(sk)) | |
1611 | break; | |
1612 | ||
1613 | /* queue is in-order => we can short-circuit the walk early */ | |
1614 | if (!before(TCP_SKB_CB(skb)->seq, end_seq)) | |
1615 | break; | |
1616 | ||
1617 | if ((next_dup != NULL) && | |
1618 | before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { | |
1619 | in_sack = tcp_match_skb_to_sack(sk, skb, | |
1620 | next_dup->start_seq, | |
1621 | next_dup->end_seq); | |
1622 | if (in_sack > 0) | |
1623 | dup_sack = 1; | |
1624 | } | |
1625 | ||
1626 | /* skb reference here is a bit tricky to get right, since | |
1627 | * shifting can eat and free both this skb and the next, | |
1628 | * so not even _safe variant of the loop is enough. | |
1629 | */ | |
1630 | if (in_sack <= 0) { | |
1631 | tmp = tcp_shift_skb_data(sk, skb, state, | |
1632 | start_seq, end_seq, dup_sack); | |
1633 | if (tmp != NULL) { | |
1634 | if (tmp != skb) { | |
1635 | skb = tmp; | |
1636 | continue; | |
1637 | } | |
1638 | ||
1639 | in_sack = 0; | |
1640 | } else { | |
1641 | in_sack = tcp_match_skb_to_sack(sk, skb, | |
1642 | start_seq, | |
1643 | end_seq); | |
1644 | } | |
1645 | } | |
1646 | ||
1647 | if (unlikely(in_sack < 0)) | |
1648 | break; | |
1649 | ||
1650 | if (in_sack) { | |
1651 | TCP_SKB_CB(skb)->sacked = tcp_sacktag_one(skb, sk, | |
1652 | state, | |
1653 | dup_sack, | |
1654 | tcp_skb_pcount(skb)); | |
1655 | ||
1656 | if (!before(TCP_SKB_CB(skb)->seq, | |
1657 | tcp_highest_sack_seq(tp))) | |
1658 | tcp_advance_highest_sack(sk, skb); | |
1659 | } | |
1660 | ||
1661 | state->fack_count += tcp_skb_pcount(skb); | |
1662 | } | |
1663 | return skb; | |
1664 | } | |
1665 | ||
1666 | /* Avoid all extra work that is being done by sacktag while walking in | |
1667 | * a normal way | |
1668 | */ | |
1669 | static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, | |
1670 | struct tcp_sacktag_state *state, | |
1671 | u32 skip_to_seq) | |
1672 | { | |
1673 | tcp_for_write_queue_from(skb, sk) { | |
1674 | if (skb == tcp_send_head(sk)) | |
1675 | break; | |
1676 | ||
1677 | if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq)) | |
1678 | break; | |
1679 | ||
1680 | state->fack_count += tcp_skb_pcount(skb); | |
1681 | } | |
1682 | return skb; | |
1683 | } | |
1684 | ||
1685 | static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, | |
1686 | struct sock *sk, | |
1687 | struct tcp_sack_block *next_dup, | |
1688 | struct tcp_sacktag_state *state, | |
1689 | u32 skip_to_seq) | |
1690 | { | |
1691 | if (next_dup == NULL) | |
1692 | return skb; | |
1693 | ||
1694 | if (before(next_dup->start_seq, skip_to_seq)) { | |
1695 | skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq); | |
1696 | skb = tcp_sacktag_walk(skb, sk, NULL, state, | |
1697 | next_dup->start_seq, next_dup->end_seq, | |
1698 | 1); | |
1699 | } | |
1700 | ||
1701 | return skb; | |
1702 | } | |
1703 | ||
1704 | static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache) | |
1705 | { | |
1706 | return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); | |
1707 | } | |
1708 | ||
1709 | static int | |
1710 | tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, | |
1711 | u32 prior_snd_una) | |
1712 | { | |
1713 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
1714 | struct tcp_sock *tp = tcp_sk(sk); | |
1715 | unsigned char *ptr = (skb_transport_header(ack_skb) + | |
1716 | TCP_SKB_CB(ack_skb)->sacked); | |
1717 | struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); | |
1718 | struct tcp_sack_block sp[TCP_NUM_SACKS]; | |
1719 | struct tcp_sack_block *cache; | |
1720 | struct tcp_sacktag_state state; | |
1721 | struct sk_buff *skb; | |
1722 | int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); | |
1723 | int used_sacks; | |
1724 | int found_dup_sack = 0; | |
1725 | int i, j; | |
1726 | int first_sack_index; | |
1727 | ||
1728 | state.flag = 0; | |
1729 | state.reord = tp->packets_out; | |
1730 | ||
1731 | if (!tp->sacked_out) { | |
1732 | if (WARN_ON(tp->fackets_out)) | |
1733 | tp->fackets_out = 0; | |
1734 | tcp_highest_sack_reset(sk); | |
1735 | } | |
1736 | ||
1737 | found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, | |
1738 | num_sacks, prior_snd_una); | |
1739 | if (found_dup_sack) | |
1740 | state.flag |= FLAG_DSACKING_ACK; | |
1741 | ||
1742 | /* Eliminate too old ACKs, but take into | |
1743 | * account more or less fresh ones, they can | |
1744 | * contain valid SACK info. | |
1745 | */ | |
1746 | if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) | |
1747 | return 0; | |
1748 | ||
1749 | if (!tp->packets_out) | |
1750 | goto out; | |
1751 | ||
1752 | used_sacks = 0; | |
1753 | first_sack_index = 0; | |
1754 | for (i = 0; i < num_sacks; i++) { | |
1755 | int dup_sack = !i && found_dup_sack; | |
1756 | ||
1757 | sp[used_sacks].start_seq = get_unaligned_be32(&sp_wire[i].start_seq); | |
1758 | sp[used_sacks].end_seq = get_unaligned_be32(&sp_wire[i].end_seq); | |
1759 | ||
1760 | if (!tcp_is_sackblock_valid(tp, dup_sack, | |
1761 | sp[used_sacks].start_seq, | |
1762 | sp[used_sacks].end_seq)) { | |
1763 | int mib_idx; | |
1764 | ||
1765 | if (dup_sack) { | |
1766 | if (!tp->undo_marker) | |
1767 | mib_idx = LINUX_MIB_TCPDSACKIGNOREDNOUNDO; | |
1768 | else | |
1769 | mib_idx = LINUX_MIB_TCPDSACKIGNOREDOLD; | |
1770 | } else { | |
1771 | /* Don't count olds caused by ACK reordering */ | |
1772 | if ((TCP_SKB_CB(ack_skb)->ack_seq != tp->snd_una) && | |
1773 | !after(sp[used_sacks].end_seq, tp->snd_una)) | |
1774 | continue; | |
1775 | mib_idx = LINUX_MIB_TCPSACKDISCARD; | |
1776 | } | |
1777 | ||
1778 | NET_INC_STATS_BH(sock_net(sk), mib_idx); | |
1779 | if (i == 0) | |
1780 | first_sack_index = -1; | |
1781 | continue; | |
1782 | } | |
1783 | ||
1784 | /* Ignore very old stuff early */ | |
1785 | if (!after(sp[used_sacks].end_seq, prior_snd_una)) | |
1786 | continue; | |
1787 | ||
1788 | used_sacks++; | |
1789 | } | |
1790 | ||
1791 | /* order SACK blocks to allow in order walk of the retrans queue */ | |
1792 | for (i = used_sacks - 1; i > 0; i--) { | |
1793 | for (j = 0; j < i; j++) { | |
1794 | if (after(sp[j].start_seq, sp[j + 1].start_seq)) { | |
1795 | swap(sp[j], sp[j + 1]); | |
1796 | ||
1797 | /* Track where the first SACK block goes to */ | |
1798 | if (j == first_sack_index) | |
1799 | first_sack_index = j + 1; | |
1800 | } | |
1801 | } | |
1802 | } | |
1803 | ||
1804 | skb = tcp_write_queue_head(sk); | |
1805 | state.fack_count = 0; | |
1806 | i = 0; | |
1807 | ||
1808 | if (!tp->sacked_out) { | |
1809 | /* It's already past, so skip checking against it */ | |
1810 | cache = tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); | |
1811 | } else { | |
1812 | cache = tp->recv_sack_cache; | |
1813 | /* Skip empty blocks in at head of the cache */ | |
1814 | while (tcp_sack_cache_ok(tp, cache) && !cache->start_seq && | |
1815 | !cache->end_seq) | |
1816 | cache++; | |
1817 | } | |
1818 | ||
1819 | while (i < used_sacks) { | |
1820 | u32 start_seq = sp[i].start_seq; | |
1821 | u32 end_seq = sp[i].end_seq; | |
1822 | int dup_sack = (found_dup_sack && (i == first_sack_index)); | |
1823 | struct tcp_sack_block *next_dup = NULL; | |
1824 | ||
1825 | if (found_dup_sack && ((i + 1) == first_sack_index)) | |
1826 | next_dup = &sp[i + 1]; | |
1827 | ||
1828 | /* Event "B" in the comment above. */ | |
1829 | if (after(end_seq, tp->high_seq)) | |
1830 | state.flag |= FLAG_DATA_LOST; | |
1831 | ||
1832 | /* Skip too early cached blocks */ | |
1833 | while (tcp_sack_cache_ok(tp, cache) && | |
1834 | !before(start_seq, cache->end_seq)) | |
1835 | cache++; | |
1836 | ||
1837 | /* Can skip some work by looking recv_sack_cache? */ | |
1838 | if (tcp_sack_cache_ok(tp, cache) && !dup_sack && | |
1839 | after(end_seq, cache->start_seq)) { | |
1840 | ||
1841 | /* Head todo? */ | |
1842 | if (before(start_seq, cache->start_seq)) { | |
1843 | skb = tcp_sacktag_skip(skb, sk, &state, | |
1844 | start_seq); | |
1845 | skb = tcp_sacktag_walk(skb, sk, next_dup, | |
1846 | &state, | |
1847 | start_seq, | |
1848 | cache->start_seq, | |
1849 | dup_sack); | |
1850 | } | |
1851 | ||
1852 | /* Rest of the block already fully processed? */ | |
1853 | if (!after(end_seq, cache->end_seq)) | |
1854 | goto advance_sp; | |
1855 | ||
1856 | skb = tcp_maybe_skipping_dsack(skb, sk, next_dup, | |
1857 | &state, | |
1858 | cache->end_seq); | |
1859 | ||
1860 | /* ...tail remains todo... */ | |
1861 | if (tcp_highest_sack_seq(tp) == cache->end_seq) { | |
1862 | /* ...but better entrypoint exists! */ | |
1863 | skb = tcp_highest_sack(sk); | |
1864 | if (skb == NULL) | |
1865 | break; | |
1866 | state.fack_count = tp->fackets_out; | |
1867 | cache++; | |
1868 | goto walk; | |
1869 | } | |
1870 | ||
1871 | skb = tcp_sacktag_skip(skb, sk, &state, cache->end_seq); | |
1872 | /* Check overlap against next cached too (past this one already) */ | |
1873 | cache++; | |
1874 | continue; | |
1875 | } | |
1876 | ||
1877 | if (!before(start_seq, tcp_highest_sack_seq(tp))) { | |
1878 | skb = tcp_highest_sack(sk); | |
1879 | if (skb == NULL) | |
1880 | break; | |
1881 | state.fack_count = tp->fackets_out; | |
1882 | } | |
1883 | skb = tcp_sacktag_skip(skb, sk, &state, start_seq); | |
1884 | ||
1885 | walk: | |
1886 | skb = tcp_sacktag_walk(skb, sk, next_dup, &state, | |
1887 | start_seq, end_seq, dup_sack); | |
1888 | ||
1889 | advance_sp: | |
1890 | /* SACK enhanced FRTO (RFC4138, Appendix B): Clearing correct | |
1891 | * due to in-order walk | |
1892 | */ | |
1893 | if (after(end_seq, tp->frto_highmark)) | |
1894 | state.flag &= ~FLAG_ONLY_ORIG_SACKED; | |
1895 | ||
1896 | i++; | |
1897 | } | |
1898 | ||
1899 | /* Clear the head of the cache sack blocks so we can skip it next time */ | |
1900 | for (i = 0; i < ARRAY_SIZE(tp->recv_sack_cache) - used_sacks; i++) { | |
1901 | tp->recv_sack_cache[i].start_seq = 0; | |
1902 | tp->recv_sack_cache[i].end_seq = 0; | |
1903 | } | |
1904 | for (j = 0; j < used_sacks; j++) | |
1905 | tp->recv_sack_cache[i++] = sp[j]; | |
1906 | ||
1907 | tcp_mark_lost_retrans(sk); | |
1908 | ||
1909 | tcp_verify_left_out(tp); | |
1910 | ||
1911 | if ((state.reord < tp->fackets_out) && | |
1912 | ((icsk->icsk_ca_state != TCP_CA_Loss) || tp->undo_marker) && | |
1913 | (!tp->frto_highmark || after(tp->snd_una, tp->frto_highmark))) | |
1914 | tcp_update_reordering(sk, tp->fackets_out - state.reord, 0); | |
1915 | ||
1916 | out: | |
1917 | ||
1918 | #if FASTRETRANS_DEBUG > 0 | |
1919 | WARN_ON((int)tp->sacked_out < 0); | |
1920 | WARN_ON((int)tp->lost_out < 0); | |
1921 | WARN_ON((int)tp->retrans_out < 0); | |
1922 | WARN_ON((int)tcp_packets_in_flight(tp) < 0); | |
1923 | #endif | |
1924 | return state.flag; | |
1925 | } | |
1926 | ||
1927 | /* Limits sacked_out so that sum with lost_out isn't ever larger than | |
1928 | * packets_out. Returns zero if sacked_out adjustement wasn't necessary. | |
1929 | */ | |
1930 | static int tcp_limit_reno_sacked(struct tcp_sock *tp) | |
1931 | { | |
1932 | u32 holes; | |
1933 | ||
1934 | holes = max(tp->lost_out, 1U); | |
1935 | holes = min(holes, tp->packets_out); | |
1936 | ||
1937 | if ((tp->sacked_out + holes) > tp->packets_out) { | |
1938 | tp->sacked_out = tp->packets_out - holes; | |
1939 | return 1; | |
1940 | } | |
1941 | return 0; | |
1942 | } | |
1943 | ||
1944 | /* If we receive more dupacks than we expected counting segments | |
1945 | * in assumption of absent reordering, interpret this as reordering. | |
1946 | * The only another reason could be bug in receiver TCP. | |
1947 | */ | |
1948 | static void tcp_check_reno_reordering(struct sock *sk, const int addend) | |
1949 | { | |
1950 | struct tcp_sock *tp = tcp_sk(sk); | |
1951 | if (tcp_limit_reno_sacked(tp)) | |
1952 | tcp_update_reordering(sk, tp->packets_out + addend, 0); | |
1953 | } | |
1954 | ||
1955 | /* Emulate SACKs for SACKless connection: account for a new dupack. */ | |
1956 | ||
1957 | static void tcp_add_reno_sack(struct sock *sk) | |
1958 | { | |
1959 | struct tcp_sock *tp = tcp_sk(sk); | |
1960 | tp->sacked_out++; | |
1961 | tcp_check_reno_reordering(sk, 0); | |
1962 | tcp_verify_left_out(tp); | |
1963 | } | |
1964 | ||
1965 | /* Account for ACK, ACKing some data in Reno Recovery phase. */ | |
1966 | ||
1967 | static void tcp_remove_reno_sacks(struct sock *sk, int acked) | |
1968 | { | |
1969 | struct tcp_sock *tp = tcp_sk(sk); | |
1970 | ||
1971 | if (acked > 0) { | |
1972 | /* One ACK acked hole. The rest eat duplicate ACKs. */ | |
1973 | if (acked - 1 >= tp->sacked_out) | |
1974 | tp->sacked_out = 0; | |
1975 | else | |
1976 | tp->sacked_out -= acked - 1; | |
1977 | } | |
1978 | tcp_check_reno_reordering(sk, acked); | |
1979 | tcp_verify_left_out(tp); | |
1980 | } | |
1981 | ||
1982 | static inline void tcp_reset_reno_sack(struct tcp_sock *tp) | |
1983 | { | |
1984 | tp->sacked_out = 0; | |
1985 | } | |
1986 | ||
1987 | static int tcp_is_sackfrto(const struct tcp_sock *tp) | |
1988 | { | |
1989 | return (sysctl_tcp_frto == 0x2) && !tcp_is_reno(tp); | |
1990 | } | |
1991 | ||
1992 | /* F-RTO can only be used if TCP has never retransmitted anything other than | |
1993 | * head (SACK enhanced variant from Appendix B of RFC4138 is more robust here) | |
1994 | */ | |
1995 | int tcp_use_frto(struct sock *sk) | |
1996 | { | |
1997 | const struct tcp_sock *tp = tcp_sk(sk); | |
1998 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
1999 | struct sk_buff *skb; | |
2000 | ||
2001 | if (!sysctl_tcp_frto) | |
2002 | return 0; | |
2003 | ||
2004 | /* MTU probe and F-RTO won't really play nicely along currently */ | |
2005 | if (icsk->icsk_mtup.probe_size) | |
2006 | return 0; | |
2007 | ||
2008 | if (tcp_is_sackfrto(tp)) | |
2009 | return 1; | |
2010 | ||
2011 | /* Avoid expensive walking of rexmit queue if possible */ | |
2012 | if (tp->retrans_out > 1) | |
2013 | return 0; | |
2014 | ||
2015 | skb = tcp_write_queue_head(sk); | |
2016 | if (tcp_skb_is_last(sk, skb)) | |
2017 | return 1; | |
2018 | skb = tcp_write_queue_next(sk, skb); /* Skips head */ | |
2019 | tcp_for_write_queue_from(skb, sk) { | |
2020 | if (skb == tcp_send_head(sk)) | |
2021 | break; | |
2022 | if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) | |
2023 | return 0; | |
2024 | /* Short-circuit when first non-SACKed skb has been checked */ | |
2025 | if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) | |
2026 | break; | |
2027 | } | |
2028 | return 1; | |
2029 | } | |
2030 | ||
2031 | /* RTO occurred, but do not yet enter Loss state. Instead, defer RTO | |
2032 | * recovery a bit and use heuristics in tcp_process_frto() to detect if | |
2033 | * the RTO was spurious. Only clear SACKED_RETRANS of the head here to | |
2034 | * keep retrans_out counting accurate (with SACK F-RTO, other than head | |
2035 | * may still have that bit set); TCPCB_LOST and remaining SACKED_RETRANS | |
2036 | * bits are handled if the Loss state is really to be entered (in | |
2037 | * tcp_enter_frto_loss). | |
2038 | * | |
2039 | * Do like tcp_enter_loss() would; when RTO expires the second time it | |
2040 | * does: | |
2041 | * "Reduce ssthresh if it has not yet been made inside this window." | |
2042 | */ | |
2043 | void tcp_enter_frto(struct sock *sk) | |
2044 | { | |
2045 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
2046 | struct tcp_sock *tp = tcp_sk(sk); | |
2047 | struct sk_buff *skb; | |
2048 | ||
2049 | if ((!tp->frto_counter && icsk->icsk_ca_state <= TCP_CA_Disorder) || | |
2050 | tp->snd_una == tp->high_seq || | |
2051 | ((icsk->icsk_ca_state == TCP_CA_Loss || tp->frto_counter) && | |
2052 | !icsk->icsk_retransmits)) { | |
2053 | tp->prior_ssthresh = tcp_current_ssthresh(sk); | |
2054 | /* Our state is too optimistic in ssthresh() call because cwnd | |
2055 | * is not reduced until tcp_enter_frto_loss() when previous F-RTO | |
2056 | * recovery has not yet completed. Pattern would be this: RTO, | |
2057 | * Cumulative ACK, RTO (2xRTO for the same segment does not end | |
2058 | * up here twice). | |
2059 | * RFC4138 should be more specific on what to do, even though | |
2060 | * RTO is quite unlikely to occur after the first Cumulative ACK | |
2061 | * due to back-off and complexity of triggering events ... | |
2062 | */ | |
2063 | if (tp->frto_counter) { | |
2064 | u32 stored_cwnd; | |
2065 | stored_cwnd = tp->snd_cwnd; | |
2066 | tp->snd_cwnd = 2; | |
2067 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
2068 | tp->snd_cwnd = stored_cwnd; | |
2069 | } else { | |
2070 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
2071 | } | |
2072 | /* ... in theory, cong.control module could do "any tricks" in | |
2073 | * ssthresh(), which means that ca_state, lost bits and lost_out | |
2074 | * counter would have to be faked before the call occurs. We | |
2075 | * consider that too expensive, unlikely and hacky, so modules | |
2076 | * using these in ssthresh() must deal these incompatibility | |
2077 | * issues if they receives CA_EVENT_FRTO and frto_counter != 0 | |
2078 | */ | |
2079 | tcp_ca_event(sk, CA_EVENT_FRTO); | |
2080 | } | |
2081 | ||
2082 | tp->undo_marker = tp->snd_una; | |
2083 | tp->undo_retrans = 0; | |
2084 | ||
2085 | skb = tcp_write_queue_head(sk); | |
2086 | if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) | |
2087 | tp->undo_marker = 0; | |
2088 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { | |
2089 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; | |
2090 | tp->retrans_out -= tcp_skb_pcount(skb); | |
2091 | } | |
2092 | tcp_verify_left_out(tp); | |
2093 | ||
2094 | /* Too bad if TCP was application limited */ | |
2095 | tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1); | |
2096 | ||
2097 | /* Earlier loss recovery underway (see RFC4138; Appendix B). | |
2098 | * The last condition is necessary at least in tp->frto_counter case. | |
2099 | */ | |
2100 | if (tcp_is_sackfrto(tp) && (tp->frto_counter || | |
2101 | ((1 << icsk->icsk_ca_state) & (TCPF_CA_Recovery|TCPF_CA_Loss))) && | |
2102 | after(tp->high_seq, tp->snd_una)) { | |
2103 | tp->frto_highmark = tp->high_seq; | |
2104 | } else { | |
2105 | tp->frto_highmark = tp->snd_nxt; | |
2106 | } | |
2107 | tcp_set_ca_state(sk, TCP_CA_Disorder); | |
2108 | tp->high_seq = tp->snd_nxt; | |
2109 | tp->frto_counter = 1; | |
2110 | } | |
2111 | ||
2112 | /* Enter Loss state after F-RTO was applied. Dupack arrived after RTO, | |
2113 | * which indicates that we should follow the traditional RTO recovery, | |
2114 | * i.e. mark everything lost and do go-back-N retransmission. | |
2115 | */ | |
2116 | static void tcp_enter_frto_loss(struct sock *sk, int allowed_segments, int flag) | |
2117 | { | |
2118 | struct tcp_sock *tp = tcp_sk(sk); | |
2119 | struct sk_buff *skb; | |
2120 | ||
2121 | tp->lost_out = 0; | |
2122 | tp->retrans_out = 0; | |
2123 | if (tcp_is_reno(tp)) | |
2124 | tcp_reset_reno_sack(tp); | |
2125 | ||
2126 | tcp_for_write_queue(skb, sk) { | |
2127 | if (skb == tcp_send_head(sk)) | |
2128 | break; | |
2129 | ||
2130 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; | |
2131 | /* | |
2132 | * Count the retransmission made on RTO correctly (only when | |
2133 | * waiting for the first ACK and did not get it)... | |
2134 | */ | |
2135 | if ((tp->frto_counter == 1) && !(flag & FLAG_DATA_ACKED)) { | |
2136 | /* For some reason this R-bit might get cleared? */ | |
2137 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
2138 | tp->retrans_out += tcp_skb_pcount(skb); | |
2139 | /* ...enter this if branch just for the first segment */ | |
2140 | flag |= FLAG_DATA_ACKED; | |
2141 | } else { | |
2142 | if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) | |
2143 | tp->undo_marker = 0; | |
2144 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; | |
2145 | } | |
2146 | ||
2147 | /* Marking forward transmissions that were made after RTO lost | |
2148 | * can cause unnecessary retransmissions in some scenarios, | |
2149 | * SACK blocks will mitigate that in some but not in all cases. | |
2150 | * We used to not mark them but it was causing break-ups with | |
2151 | * receivers that do only in-order receival. | |
2152 | * | |
2153 | * TODO: we could detect presence of such receiver and select | |
2154 | * different behavior per flow. | |
2155 | */ | |
2156 | if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { | |
2157 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
2158 | tp->lost_out += tcp_skb_pcount(skb); | |
2159 | tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; | |
2160 | } | |
2161 | } | |
2162 | tcp_verify_left_out(tp); | |
2163 | ||
2164 | tp->snd_cwnd = tcp_packets_in_flight(tp) + allowed_segments; | |
2165 | tp->snd_cwnd_cnt = 0; | |
2166 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
2167 | tp->frto_counter = 0; | |
2168 | tp->bytes_acked = 0; | |
2169 | ||
2170 | tp->reordering = min_t(unsigned int, tp->reordering, | |
2171 | sysctl_tcp_reordering); | |
2172 | tcp_set_ca_state(sk, TCP_CA_Loss); | |
2173 | tp->high_seq = tp->snd_nxt; | |
2174 | TCP_ECN_queue_cwr(tp); | |
2175 | ||
2176 | tcp_clear_all_retrans_hints(tp); | |
2177 | } | |
2178 | ||
2179 | static void tcp_clear_retrans_partial(struct tcp_sock *tp) | |
2180 | { | |
2181 | tp->retrans_out = 0; | |
2182 | tp->lost_out = 0; | |
2183 | ||
2184 | tp->undo_marker = 0; | |
2185 | tp->undo_retrans = 0; | |
2186 | } | |
2187 | ||
2188 | void tcp_clear_retrans(struct tcp_sock *tp) | |
2189 | { | |
2190 | tcp_clear_retrans_partial(tp); | |
2191 | ||
2192 | tp->fackets_out = 0; | |
2193 | tp->sacked_out = 0; | |
2194 | } | |
2195 | ||
2196 | /* Enter Loss state. If "how" is not zero, forget all SACK information | |
2197 | * and reset tags completely, otherwise preserve SACKs. If receiver | |
2198 | * dropped its ofo queue, we will know this due to reneging detection. | |
2199 | */ | |
2200 | void tcp_enter_loss(struct sock *sk, int how) | |
2201 | { | |
2202 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
2203 | struct tcp_sock *tp = tcp_sk(sk); | |
2204 | struct sk_buff *skb; | |
2205 | ||
2206 | /* Reduce ssthresh if it has not yet been made inside this window. */ | |
2207 | if (icsk->icsk_ca_state <= TCP_CA_Disorder || tp->snd_una == tp->high_seq || | |
2208 | (icsk->icsk_ca_state == TCP_CA_Loss && !icsk->icsk_retransmits)) { | |
2209 | tp->prior_ssthresh = tcp_current_ssthresh(sk); | |
2210 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
2211 | tcp_ca_event(sk, CA_EVENT_LOSS); | |
2212 | } | |
2213 | tp->snd_cwnd = 1; | |
2214 | tp->snd_cwnd_cnt = 0; | |
2215 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
2216 | ||
2217 | tp->bytes_acked = 0; | |
2218 | tcp_clear_retrans_partial(tp); | |
2219 | ||
2220 | if (tcp_is_reno(tp)) | |
2221 | tcp_reset_reno_sack(tp); | |
2222 | ||
2223 | if (!how) { | |
2224 | /* Push undo marker, if it was plain RTO and nothing | |
2225 | * was retransmitted. */ | |
2226 | tp->undo_marker = tp->snd_una; | |
2227 | } else { | |
2228 | tp->sacked_out = 0; | |
2229 | tp->fackets_out = 0; | |
2230 | } | |
2231 | tcp_clear_all_retrans_hints(tp); | |
2232 | ||
2233 | tcp_for_write_queue(skb, sk) { | |
2234 | if (skb == tcp_send_head(sk)) | |
2235 | break; | |
2236 | ||
2237 | if (TCP_SKB_CB(skb)->sacked & TCPCB_RETRANS) | |
2238 | tp->undo_marker = 0; | |
2239 | TCP_SKB_CB(skb)->sacked &= (~TCPCB_TAGBITS)|TCPCB_SACKED_ACKED; | |
2240 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED) || how) { | |
2241 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_ACKED; | |
2242 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
2243 | tp->lost_out += tcp_skb_pcount(skb); | |
2244 | tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; | |
2245 | } | |
2246 | } | |
2247 | tcp_verify_left_out(tp); | |
2248 | ||
2249 | tp->reordering = min_t(unsigned int, tp->reordering, | |
2250 | sysctl_tcp_reordering); | |
2251 | tcp_set_ca_state(sk, TCP_CA_Loss); | |
2252 | tp->high_seq = tp->snd_nxt; | |
2253 | TCP_ECN_queue_cwr(tp); | |
2254 | /* Abort F-RTO algorithm if one is in progress */ | |
2255 | tp->frto_counter = 0; | |
2256 | } | |
2257 | ||
2258 | /* If ACK arrived pointing to a remembered SACK, it means that our | |
2259 | * remembered SACKs do not reflect real state of receiver i.e. | |
2260 | * receiver _host_ is heavily congested (or buggy). | |
2261 | * | |
2262 | * Do processing similar to RTO timeout. | |
2263 | */ | |
2264 | static int tcp_check_sack_reneging(struct sock *sk, int flag) | |
2265 | { | |
2266 | if (flag & FLAG_SACK_RENEGING) { | |
2267 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2268 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSACKRENEGING); | |
2269 | ||
2270 | tcp_enter_loss(sk, 1); | |
2271 | icsk->icsk_retransmits++; | |
2272 | tcp_retransmit_skb(sk, tcp_write_queue_head(sk)); | |
2273 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, | |
2274 | icsk->icsk_rto, TCP_RTO_MAX); | |
2275 | return 1; | |
2276 | } | |
2277 | return 0; | |
2278 | } | |
2279 | ||
2280 | static inline int tcp_fackets_out(struct tcp_sock *tp) | |
2281 | { | |
2282 | return tcp_is_reno(tp) ? tp->sacked_out + 1 : tp->fackets_out; | |
2283 | } | |
2284 | ||
2285 | /* Heurestics to calculate number of duplicate ACKs. There's no dupACKs | |
2286 | * counter when SACK is enabled (without SACK, sacked_out is used for | |
2287 | * that purpose). | |
2288 | * | |
2289 | * Instead, with FACK TCP uses fackets_out that includes both SACKed | |
2290 | * segments up to the highest received SACK block so far and holes in | |
2291 | * between them. | |
2292 | * | |
2293 | * With reordering, holes may still be in flight, so RFC3517 recovery | |
2294 | * uses pure sacked_out (total number of SACKed segments) even though | |
2295 | * it violates the RFC that uses duplicate ACKs, often these are equal | |
2296 | * but when e.g. out-of-window ACKs or packet duplication occurs, | |
2297 | * they differ. Since neither occurs due to loss, TCP should really | |
2298 | * ignore them. | |
2299 | */ | |
2300 | static inline int tcp_dupack_heuristics(struct tcp_sock *tp) | |
2301 | { | |
2302 | return tcp_is_fack(tp) ? tp->fackets_out : tp->sacked_out + 1; | |
2303 | } | |
2304 | ||
2305 | static inline int tcp_skb_timedout(struct sock *sk, struct sk_buff *skb) | |
2306 | { | |
2307 | return tcp_time_stamp - TCP_SKB_CB(skb)->when > inet_csk(sk)->icsk_rto; | |
2308 | } | |
2309 | ||
2310 | static inline int tcp_head_timedout(struct sock *sk) | |
2311 | { | |
2312 | struct tcp_sock *tp = tcp_sk(sk); | |
2313 | ||
2314 | return tp->packets_out && | |
2315 | tcp_skb_timedout(sk, tcp_write_queue_head(sk)); | |
2316 | } | |
2317 | ||
2318 | /* Linux NewReno/SACK/FACK/ECN state machine. | |
2319 | * -------------------------------------- | |
2320 | * | |
2321 | * "Open" Normal state, no dubious events, fast path. | |
2322 | * "Disorder" In all the respects it is "Open", | |
2323 | * but requires a bit more attention. It is entered when | |
2324 | * we see some SACKs or dupacks. It is split of "Open" | |
2325 | * mainly to move some processing from fast path to slow one. | |
2326 | * "CWR" CWND was reduced due to some Congestion Notification event. | |
2327 | * It can be ECN, ICMP source quench, local device congestion. | |
2328 | * "Recovery" CWND was reduced, we are fast-retransmitting. | |
2329 | * "Loss" CWND was reduced due to RTO timeout or SACK reneging. | |
2330 | * | |
2331 | * tcp_fastretrans_alert() is entered: | |
2332 | * - each incoming ACK, if state is not "Open" | |
2333 | * - when arrived ACK is unusual, namely: | |
2334 | * * SACK | |
2335 | * * Duplicate ACK. | |
2336 | * * ECN ECE. | |
2337 | * | |
2338 | * Counting packets in flight is pretty simple. | |
2339 | * | |
2340 | * in_flight = packets_out - left_out + retrans_out | |
2341 | * | |
2342 | * packets_out is SND.NXT-SND.UNA counted in packets. | |
2343 | * | |
2344 | * retrans_out is number of retransmitted segments. | |
2345 | * | |
2346 | * left_out is number of segments left network, but not ACKed yet. | |
2347 | * | |
2348 | * left_out = sacked_out + lost_out | |
2349 | * | |
2350 | * sacked_out: Packets, which arrived to receiver out of order | |
2351 | * and hence not ACKed. With SACKs this number is simply | |
2352 | * amount of SACKed data. Even without SACKs | |
2353 | * it is easy to give pretty reliable estimate of this number, | |
2354 | * counting duplicate ACKs. | |
2355 | * | |
2356 | * lost_out: Packets lost by network. TCP has no explicit | |
2357 | * "loss notification" feedback from network (for now). | |
2358 | * It means that this number can be only _guessed_. | |
2359 | * Actually, it is the heuristics to predict lossage that | |
2360 | * distinguishes different algorithms. | |
2361 | * | |
2362 | * F.e. after RTO, when all the queue is considered as lost, | |
2363 | * lost_out = packets_out and in_flight = retrans_out. | |
2364 | * | |
2365 | * Essentially, we have now two algorithms counting | |
2366 | * lost packets. | |
2367 | * | |
2368 | * FACK: It is the simplest heuristics. As soon as we decided | |
2369 | * that something is lost, we decide that _all_ not SACKed | |
2370 | * packets until the most forward SACK are lost. I.e. | |
2371 | * lost_out = fackets_out - sacked_out and left_out = fackets_out. | |
2372 | * It is absolutely correct estimate, if network does not reorder | |
2373 | * packets. And it loses any connection to reality when reordering | |
2374 | * takes place. We use FACK by default until reordering | |
2375 | * is suspected on the path to this destination. | |
2376 | * | |
2377 | * NewReno: when Recovery is entered, we assume that one segment | |
2378 | * is lost (classic Reno). While we are in Recovery and | |
2379 | * a partial ACK arrives, we assume that one more packet | |
2380 | * is lost (NewReno). This heuristics are the same in NewReno | |
2381 | * and SACK. | |
2382 | * | |
2383 | * Imagine, that's all! Forget about all this shamanism about CWND inflation | |
2384 | * deflation etc. CWND is real congestion window, never inflated, changes | |
2385 | * only according to classic VJ rules. | |
2386 | * | |
2387 | * Really tricky (and requiring careful tuning) part of algorithm | |
2388 | * is hidden in functions tcp_time_to_recover() and tcp_xmit_retransmit_queue(). | |
2389 | * The first determines the moment _when_ we should reduce CWND and, | |
2390 | * hence, slow down forward transmission. In fact, it determines the moment | |
2391 | * when we decide that hole is caused by loss, rather than by a reorder. | |
2392 | * | |
2393 | * tcp_xmit_retransmit_queue() decides, _what_ we should retransmit to fill | |
2394 | * holes, caused by lost packets. | |
2395 | * | |
2396 | * And the most logically complicated part of algorithm is undo | |
2397 | * heuristics. We detect false retransmits due to both too early | |
2398 | * fast retransmit (reordering) and underestimated RTO, analyzing | |
2399 | * timestamps and D-SACKs. When we detect that some segments were | |
2400 | * retransmitted by mistake and CWND reduction was wrong, we undo | |
2401 | * window reduction and abort recovery phase. This logic is hidden | |
2402 | * inside several functions named tcp_try_undo_<something>. | |
2403 | */ | |
2404 | ||
2405 | /* This function decides, when we should leave Disordered state | |
2406 | * and enter Recovery phase, reducing congestion window. | |
2407 | * | |
2408 | * Main question: may we further continue forward transmission | |
2409 | * with the same cwnd? | |
2410 | */ | |
2411 | static int tcp_time_to_recover(struct sock *sk) | |
2412 | { | |
2413 | struct tcp_sock *tp = tcp_sk(sk); | |
2414 | __u32 packets_out; | |
2415 | ||
2416 | /* Do not perform any recovery during F-RTO algorithm */ | |
2417 | if (tp->frto_counter) | |
2418 | return 0; | |
2419 | ||
2420 | /* Trick#1: The loss is proven. */ | |
2421 | if (tp->lost_out) | |
2422 | return 1; | |
2423 | ||
2424 | /* Not-A-Trick#2 : Classic rule... */ | |
2425 | if (tcp_dupack_heuristics(tp) > tp->reordering) | |
2426 | return 1; | |
2427 | ||
2428 | /* Trick#3 : when we use RFC2988 timer restart, fast | |
2429 | * retransmit can be triggered by timeout of queue head. | |
2430 | */ | |
2431 | if (tcp_is_fack(tp) && tcp_head_timedout(sk)) | |
2432 | return 1; | |
2433 | ||
2434 | /* Trick#4: It is still not OK... But will it be useful to delay | |
2435 | * recovery more? | |
2436 | */ | |
2437 | packets_out = tp->packets_out; | |
2438 | if (packets_out <= tp->reordering && | |
2439 | tp->sacked_out >= max_t(__u32, packets_out/2, sysctl_tcp_reordering) && | |
2440 | !tcp_may_send_now(sk)) { | |
2441 | /* We have nothing to send. This connection is limited | |
2442 | * either by receiver window or by application. | |
2443 | */ | |
2444 | return 1; | |
2445 | } | |
2446 | ||
2447 | /* If a thin stream is detected, retransmit after first | |
2448 | * received dupack. Employ only if SACK is supported in order | |
2449 | * to avoid possible corner-case series of spurious retransmissions | |
2450 | * Use only if there are no unsent data. | |
2451 | */ | |
2452 | if ((tp->thin_dupack || sysctl_tcp_thin_dupack) && | |
2453 | tcp_stream_is_thin(tp) && tcp_dupack_heuristics(tp) > 1 && | |
2454 | tcp_is_sack(tp) && !tcp_send_head(sk)) | |
2455 | return 1; | |
2456 | ||
2457 | return 0; | |
2458 | } | |
2459 | ||
2460 | /* New heuristics: it is possible only after we switched to restart timer | |
2461 | * each time when something is ACKed. Hence, we can detect timed out packets | |
2462 | * during fast retransmit without falling to slow start. | |
2463 | * | |
2464 | * Usefulness of this as is very questionable, since we should know which of | |
2465 | * the segments is the next to timeout which is relatively expensive to find | |
2466 | * in general case unless we add some data structure just for that. The | |
2467 | * current approach certainly won't find the right one too often and when it | |
2468 | * finally does find _something_ it usually marks large part of the window | |
2469 | * right away (because a retransmission with a larger timestamp blocks the | |
2470 | * loop from advancing). -ij | |
2471 | */ | |
2472 | static void tcp_timeout_skbs(struct sock *sk) | |
2473 | { | |
2474 | struct tcp_sock *tp = tcp_sk(sk); | |
2475 | struct sk_buff *skb; | |
2476 | ||
2477 | if (!tcp_is_fack(tp) || !tcp_head_timedout(sk)) | |
2478 | return; | |
2479 | ||
2480 | skb = tp->scoreboard_skb_hint; | |
2481 | if (tp->scoreboard_skb_hint == NULL) | |
2482 | skb = tcp_write_queue_head(sk); | |
2483 | ||
2484 | tcp_for_write_queue_from(skb, sk) { | |
2485 | if (skb == tcp_send_head(sk)) | |
2486 | break; | |
2487 | if (!tcp_skb_timedout(sk, skb)) | |
2488 | break; | |
2489 | ||
2490 | tcp_skb_mark_lost(tp, skb); | |
2491 | } | |
2492 | ||
2493 | tp->scoreboard_skb_hint = skb; | |
2494 | ||
2495 | tcp_verify_left_out(tp); | |
2496 | } | |
2497 | ||
2498 | /* Mark head of queue up as lost. With RFC3517 SACK, the packets is | |
2499 | * is against sacked "cnt", otherwise it's against facked "cnt" | |
2500 | */ | |
2501 | static void tcp_mark_head_lost(struct sock *sk, int packets, int mark_head) | |
2502 | { | |
2503 | struct tcp_sock *tp = tcp_sk(sk); | |
2504 | struct sk_buff *skb; | |
2505 | int cnt, oldcnt; | |
2506 | int err; | |
2507 | unsigned int mss; | |
2508 | ||
2509 | WARN_ON(packets > tp->packets_out); | |
2510 | if (tp->lost_skb_hint) { | |
2511 | skb = tp->lost_skb_hint; | |
2512 | cnt = tp->lost_cnt_hint; | |
2513 | /* Head already handled? */ | |
2514 | if (mark_head && skb != tcp_write_queue_head(sk)) | |
2515 | return; | |
2516 | } else { | |
2517 | skb = tcp_write_queue_head(sk); | |
2518 | cnt = 0; | |
2519 | } | |
2520 | ||
2521 | tcp_for_write_queue_from(skb, sk) { | |
2522 | if (skb == tcp_send_head(sk)) | |
2523 | break; | |
2524 | /* TODO: do this better */ | |
2525 | /* this is not the most efficient way to do this... */ | |
2526 | tp->lost_skb_hint = skb; | |
2527 | tp->lost_cnt_hint = cnt; | |
2528 | ||
2529 | if (after(TCP_SKB_CB(skb)->end_seq, tp->high_seq)) | |
2530 | break; | |
2531 | ||
2532 | oldcnt = cnt; | |
2533 | if (tcp_is_fack(tp) || tcp_is_reno(tp) || | |
2534 | (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) | |
2535 | cnt += tcp_skb_pcount(skb); | |
2536 | ||
2537 | if (cnt > packets) { | |
2538 | if ((tcp_is_sack(tp) && !tcp_is_fack(tp)) || | |
2539 | (oldcnt >= packets)) | |
2540 | break; | |
2541 | ||
2542 | mss = skb_shinfo(skb)->gso_size; | |
2543 | err = tcp_fragment(sk, skb, (packets - oldcnt) * mss, mss); | |
2544 | if (err < 0) | |
2545 | break; | |
2546 | cnt = packets; | |
2547 | } | |
2548 | ||
2549 | tcp_skb_mark_lost(tp, skb); | |
2550 | ||
2551 | if (mark_head) | |
2552 | break; | |
2553 | } | |
2554 | tcp_verify_left_out(tp); | |
2555 | } | |
2556 | ||
2557 | /* Account newly detected lost packet(s) */ | |
2558 | ||
2559 | static void tcp_update_scoreboard(struct sock *sk, int fast_rexmit) | |
2560 | { | |
2561 | struct tcp_sock *tp = tcp_sk(sk); | |
2562 | ||
2563 | if (tcp_is_reno(tp)) { | |
2564 | tcp_mark_head_lost(sk, 1, 1); | |
2565 | } else if (tcp_is_fack(tp)) { | |
2566 | int lost = tp->fackets_out - tp->reordering; | |
2567 | if (lost <= 0) | |
2568 | lost = 1; | |
2569 | tcp_mark_head_lost(sk, lost, 0); | |
2570 | } else { | |
2571 | int sacked_upto = tp->sacked_out - tp->reordering; | |
2572 | if (sacked_upto >= 0) | |
2573 | tcp_mark_head_lost(sk, sacked_upto, 0); | |
2574 | else if (fast_rexmit) | |
2575 | tcp_mark_head_lost(sk, 1, 1); | |
2576 | } | |
2577 | ||
2578 | tcp_timeout_skbs(sk); | |
2579 | } | |
2580 | ||
2581 | /* CWND moderation, preventing bursts due to too big ACKs | |
2582 | * in dubious situations. | |
2583 | */ | |
2584 | static inline void tcp_moderate_cwnd(struct tcp_sock *tp) | |
2585 | { | |
2586 | tp->snd_cwnd = min(tp->snd_cwnd, | |
2587 | tcp_packets_in_flight(tp) + tcp_max_burst(tp)); | |
2588 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
2589 | } | |
2590 | ||
2591 | /* Lower bound on congestion window is slow start threshold | |
2592 | * unless congestion avoidance choice decides to overide it. | |
2593 | */ | |
2594 | static inline u32 tcp_cwnd_min(const struct sock *sk) | |
2595 | { | |
2596 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; | |
2597 | ||
2598 | return ca_ops->min_cwnd ? ca_ops->min_cwnd(sk) : tcp_sk(sk)->snd_ssthresh; | |
2599 | } | |
2600 | ||
2601 | /* Decrease cwnd each second ack. */ | |
2602 | static void tcp_cwnd_down(struct sock *sk, int flag) | |
2603 | { | |
2604 | struct tcp_sock *tp = tcp_sk(sk); | |
2605 | int decr = tp->snd_cwnd_cnt + 1; | |
2606 | ||
2607 | if ((flag & (FLAG_ANY_PROGRESS | FLAG_DSACKING_ACK)) || | |
2608 | (tcp_is_reno(tp) && !(flag & FLAG_NOT_DUP))) { | |
2609 | tp->snd_cwnd_cnt = decr & 1; | |
2610 | decr >>= 1; | |
2611 | ||
2612 | if (decr && tp->snd_cwnd > tcp_cwnd_min(sk)) | |
2613 | tp->snd_cwnd -= decr; | |
2614 | ||
2615 | tp->snd_cwnd = min(tp->snd_cwnd, tcp_packets_in_flight(tp) + 1); | |
2616 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
2617 | } | |
2618 | } | |
2619 | ||
2620 | /* Nothing was retransmitted or returned timestamp is less | |
2621 | * than timestamp of the first retransmission. | |
2622 | */ | |
2623 | static inline int tcp_packet_delayed(struct tcp_sock *tp) | |
2624 | { | |
2625 | return !tp->retrans_stamp || | |
2626 | (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && | |
2627 | before(tp->rx_opt.rcv_tsecr, tp->retrans_stamp)); | |
2628 | } | |
2629 | ||
2630 | /* Undo procedures. */ | |
2631 | ||
2632 | #if FASTRETRANS_DEBUG > 1 | |
2633 | static void DBGUNDO(struct sock *sk, const char *msg) | |
2634 | { | |
2635 | struct tcp_sock *tp = tcp_sk(sk); | |
2636 | struct inet_sock *inet = inet_sk(sk); | |
2637 | ||
2638 | if (sk->sk_family == AF_INET) { | |
2639 | printk(KERN_DEBUG "Undo %s %pI4/%u c%u l%u ss%u/%u p%u\n", | |
2640 | msg, | |
2641 | &inet->inet_daddr, ntohs(inet->inet_dport), | |
2642 | tp->snd_cwnd, tcp_left_out(tp), | |
2643 | tp->snd_ssthresh, tp->prior_ssthresh, | |
2644 | tp->packets_out); | |
2645 | } | |
2646 | #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) | |
2647 | else if (sk->sk_family == AF_INET6) { | |
2648 | struct ipv6_pinfo *np = inet6_sk(sk); | |
2649 | printk(KERN_DEBUG "Undo %s %pI6/%u c%u l%u ss%u/%u p%u\n", | |
2650 | msg, | |
2651 | &np->daddr, ntohs(inet->inet_dport), | |
2652 | tp->snd_cwnd, tcp_left_out(tp), | |
2653 | tp->snd_ssthresh, tp->prior_ssthresh, | |
2654 | tp->packets_out); | |
2655 | } | |
2656 | #endif | |
2657 | } | |
2658 | #else | |
2659 | #define DBGUNDO(x...) do { } while (0) | |
2660 | #endif | |
2661 | ||
2662 | static void tcp_undo_cwr(struct sock *sk, const int undo) | |
2663 | { | |
2664 | struct tcp_sock *tp = tcp_sk(sk); | |
2665 | ||
2666 | if (tp->prior_ssthresh) { | |
2667 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
2668 | ||
2669 | if (icsk->icsk_ca_ops->undo_cwnd) | |
2670 | tp->snd_cwnd = icsk->icsk_ca_ops->undo_cwnd(sk); | |
2671 | else | |
2672 | tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh << 1); | |
2673 | ||
2674 | if (undo && tp->prior_ssthresh > tp->snd_ssthresh) { | |
2675 | tp->snd_ssthresh = tp->prior_ssthresh; | |
2676 | TCP_ECN_withdraw_cwr(tp); | |
2677 | } | |
2678 | } else { | |
2679 | tp->snd_cwnd = max(tp->snd_cwnd, tp->snd_ssthresh); | |
2680 | } | |
2681 | tcp_moderate_cwnd(tp); | |
2682 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
2683 | } | |
2684 | ||
2685 | static inline int tcp_may_undo(struct tcp_sock *tp) | |
2686 | { | |
2687 | return tp->undo_marker && (!tp->undo_retrans || tcp_packet_delayed(tp)); | |
2688 | } | |
2689 | ||
2690 | /* People celebrate: "We love our President!" */ | |
2691 | static int tcp_try_undo_recovery(struct sock *sk) | |
2692 | { | |
2693 | struct tcp_sock *tp = tcp_sk(sk); | |
2694 | ||
2695 | if (tcp_may_undo(tp)) { | |
2696 | int mib_idx; | |
2697 | ||
2698 | /* Happy end! We did not retransmit anything | |
2699 | * or our original transmission succeeded. | |
2700 | */ | |
2701 | DBGUNDO(sk, inet_csk(sk)->icsk_ca_state == TCP_CA_Loss ? "loss" : "retrans"); | |
2702 | tcp_undo_cwr(sk, 1); | |
2703 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Loss) | |
2704 | mib_idx = LINUX_MIB_TCPLOSSUNDO; | |
2705 | else | |
2706 | mib_idx = LINUX_MIB_TCPFULLUNDO; | |
2707 | ||
2708 | NET_INC_STATS_BH(sock_net(sk), mib_idx); | |
2709 | tp->undo_marker = 0; | |
2710 | } | |
2711 | if (tp->snd_una == tp->high_seq && tcp_is_reno(tp)) { | |
2712 | /* Hold old state until something *above* high_seq | |
2713 | * is ACKed. For Reno it is MUST to prevent false | |
2714 | * fast retransmits (RFC2582). SACK TCP is safe. */ | |
2715 | tcp_moderate_cwnd(tp); | |
2716 | return 1; | |
2717 | } | |
2718 | tcp_set_ca_state(sk, TCP_CA_Open); | |
2719 | return 0; | |
2720 | } | |
2721 | ||
2722 | /* Try to undo cwnd reduction, because D-SACKs acked all retransmitted data */ | |
2723 | static void tcp_try_undo_dsack(struct sock *sk) | |
2724 | { | |
2725 | struct tcp_sock *tp = tcp_sk(sk); | |
2726 | ||
2727 | if (tp->undo_marker && !tp->undo_retrans) { | |
2728 | DBGUNDO(sk, "D-SACK"); | |
2729 | tcp_undo_cwr(sk, 1); | |
2730 | tp->undo_marker = 0; | |
2731 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKUNDO); | |
2732 | } | |
2733 | } | |
2734 | ||
2735 | /* We can clear retrans_stamp when there are no retransmissions in the | |
2736 | * window. It would seem that it is trivially available for us in | |
2737 | * tp->retrans_out, however, that kind of assumptions doesn't consider | |
2738 | * what will happen if errors occur when sending retransmission for the | |
2739 | * second time. ...It could the that such segment has only | |
2740 | * TCPCB_EVER_RETRANS set at the present time. It seems that checking | |
2741 | * the head skb is enough except for some reneging corner cases that | |
2742 | * are not worth the effort. | |
2743 | * | |
2744 | * Main reason for all this complexity is the fact that connection dying | |
2745 | * time now depends on the validity of the retrans_stamp, in particular, | |
2746 | * that successive retransmissions of a segment must not advance | |
2747 | * retrans_stamp under any conditions. | |
2748 | */ | |
2749 | static int tcp_any_retrans_done(struct sock *sk) | |
2750 | { | |
2751 | struct tcp_sock *tp = tcp_sk(sk); | |
2752 | struct sk_buff *skb; | |
2753 | ||
2754 | if (tp->retrans_out) | |
2755 | return 1; | |
2756 | ||
2757 | skb = tcp_write_queue_head(sk); | |
2758 | if (unlikely(skb && TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS)) | |
2759 | return 1; | |
2760 | ||
2761 | return 0; | |
2762 | } | |
2763 | ||
2764 | /* Undo during fast recovery after partial ACK. */ | |
2765 | ||
2766 | static int tcp_try_undo_partial(struct sock *sk, int acked) | |
2767 | { | |
2768 | struct tcp_sock *tp = tcp_sk(sk); | |
2769 | /* Partial ACK arrived. Force Hoe's retransmit. */ | |
2770 | int failed = tcp_is_reno(tp) || (tcp_fackets_out(tp) > tp->reordering); | |
2771 | ||
2772 | if (tcp_may_undo(tp)) { | |
2773 | /* Plain luck! Hole if filled with delayed | |
2774 | * packet, rather than with a retransmit. | |
2775 | */ | |
2776 | if (!tcp_any_retrans_done(sk)) | |
2777 | tp->retrans_stamp = 0; | |
2778 | ||
2779 | tcp_update_reordering(sk, tcp_fackets_out(tp) + acked, 1); | |
2780 | ||
2781 | DBGUNDO(sk, "Hoe"); | |
2782 | tcp_undo_cwr(sk, 0); | |
2783 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPARTIALUNDO); | |
2784 | ||
2785 | /* So... Do not make Hoe's retransmit yet. | |
2786 | * If the first packet was delayed, the rest | |
2787 | * ones are most probably delayed as well. | |
2788 | */ | |
2789 | failed = 0; | |
2790 | } | |
2791 | return failed; | |
2792 | } | |
2793 | ||
2794 | /* Undo during loss recovery after partial ACK. */ | |
2795 | static int tcp_try_undo_loss(struct sock *sk) | |
2796 | { | |
2797 | struct tcp_sock *tp = tcp_sk(sk); | |
2798 | ||
2799 | if (tcp_may_undo(tp)) { | |
2800 | struct sk_buff *skb; | |
2801 | tcp_for_write_queue(skb, sk) { | |
2802 | if (skb == tcp_send_head(sk)) | |
2803 | break; | |
2804 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; | |
2805 | } | |
2806 | ||
2807 | tcp_clear_all_retrans_hints(tp); | |
2808 | ||
2809 | DBGUNDO(sk, "partial loss"); | |
2810 | tp->lost_out = 0; | |
2811 | tcp_undo_cwr(sk, 1); | |
2812 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSUNDO); | |
2813 | inet_csk(sk)->icsk_retransmits = 0; | |
2814 | tp->undo_marker = 0; | |
2815 | if (tcp_is_sack(tp)) | |
2816 | tcp_set_ca_state(sk, TCP_CA_Open); | |
2817 | return 1; | |
2818 | } | |
2819 | return 0; | |
2820 | } | |
2821 | ||
2822 | static inline void tcp_complete_cwr(struct sock *sk) | |
2823 | { | |
2824 | struct tcp_sock *tp = tcp_sk(sk); | |
2825 | tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); | |
2826 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
2827 | tcp_ca_event(sk, CA_EVENT_COMPLETE_CWR); | |
2828 | } | |
2829 | ||
2830 | static void tcp_try_keep_open(struct sock *sk) | |
2831 | { | |
2832 | struct tcp_sock *tp = tcp_sk(sk); | |
2833 | int state = TCP_CA_Open; | |
2834 | ||
2835 | if (tcp_left_out(tp) || tcp_any_retrans_done(sk) || tp->undo_marker) | |
2836 | state = TCP_CA_Disorder; | |
2837 | ||
2838 | if (inet_csk(sk)->icsk_ca_state != state) { | |
2839 | tcp_set_ca_state(sk, state); | |
2840 | tp->high_seq = tp->snd_nxt; | |
2841 | } | |
2842 | } | |
2843 | ||
2844 | static void tcp_try_to_open(struct sock *sk, int flag) | |
2845 | { | |
2846 | struct tcp_sock *tp = tcp_sk(sk); | |
2847 | ||
2848 | tcp_verify_left_out(tp); | |
2849 | ||
2850 | if (!tp->frto_counter && !tcp_any_retrans_done(sk)) | |
2851 | tp->retrans_stamp = 0; | |
2852 | ||
2853 | if (flag & FLAG_ECE) | |
2854 | tcp_enter_cwr(sk, 1); | |
2855 | ||
2856 | if (inet_csk(sk)->icsk_ca_state != TCP_CA_CWR) { | |
2857 | tcp_try_keep_open(sk); | |
2858 | tcp_moderate_cwnd(tp); | |
2859 | } else { | |
2860 | tcp_cwnd_down(sk, flag); | |
2861 | } | |
2862 | } | |
2863 | ||
2864 | static void tcp_mtup_probe_failed(struct sock *sk) | |
2865 | { | |
2866 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2867 | ||
2868 | icsk->icsk_mtup.search_high = icsk->icsk_mtup.probe_size - 1; | |
2869 | icsk->icsk_mtup.probe_size = 0; | |
2870 | } | |
2871 | ||
2872 | static void tcp_mtup_probe_success(struct sock *sk) | |
2873 | { | |
2874 | struct tcp_sock *tp = tcp_sk(sk); | |
2875 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2876 | ||
2877 | /* FIXME: breaks with very large cwnd */ | |
2878 | tp->prior_ssthresh = tcp_current_ssthresh(sk); | |
2879 | tp->snd_cwnd = tp->snd_cwnd * | |
2880 | tcp_mss_to_mtu(sk, tp->mss_cache) / | |
2881 | icsk->icsk_mtup.probe_size; | |
2882 | tp->snd_cwnd_cnt = 0; | |
2883 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
2884 | tp->snd_ssthresh = tcp_current_ssthresh(sk); | |
2885 | ||
2886 | icsk->icsk_mtup.search_low = icsk->icsk_mtup.probe_size; | |
2887 | icsk->icsk_mtup.probe_size = 0; | |
2888 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); | |
2889 | } | |
2890 | ||
2891 | /* Do a simple retransmit without using the backoff mechanisms in | |
2892 | * tcp_timer. This is used for path mtu discovery. | |
2893 | * The socket is already locked here. | |
2894 | */ | |
2895 | void tcp_simple_retransmit(struct sock *sk) | |
2896 | { | |
2897 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
2898 | struct tcp_sock *tp = tcp_sk(sk); | |
2899 | struct sk_buff *skb; | |
2900 | unsigned int mss = tcp_current_mss(sk); | |
2901 | u32 prior_lost = tp->lost_out; | |
2902 | ||
2903 | tcp_for_write_queue(skb, sk) { | |
2904 | if (skb == tcp_send_head(sk)) | |
2905 | break; | |
2906 | if (tcp_skb_seglen(skb) > mss && | |
2907 | !(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) { | |
2908 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { | |
2909 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; | |
2910 | tp->retrans_out -= tcp_skb_pcount(skb); | |
2911 | } | |
2912 | tcp_skb_mark_lost_uncond_verify(tp, skb); | |
2913 | } | |
2914 | } | |
2915 | ||
2916 | tcp_clear_retrans_hints_partial(tp); | |
2917 | ||
2918 | if (prior_lost == tp->lost_out) | |
2919 | return; | |
2920 | ||
2921 | if (tcp_is_reno(tp)) | |
2922 | tcp_limit_reno_sacked(tp); | |
2923 | ||
2924 | tcp_verify_left_out(tp); | |
2925 | ||
2926 | /* Don't muck with the congestion window here. | |
2927 | * Reason is that we do not increase amount of _data_ | |
2928 | * in network, but units changed and effective | |
2929 | * cwnd/ssthresh really reduced now. | |
2930 | */ | |
2931 | if (icsk->icsk_ca_state != TCP_CA_Loss) { | |
2932 | tp->high_seq = tp->snd_nxt; | |
2933 | tp->snd_ssthresh = tcp_current_ssthresh(sk); | |
2934 | tp->prior_ssthresh = 0; | |
2935 | tp->undo_marker = 0; | |
2936 | tcp_set_ca_state(sk, TCP_CA_Loss); | |
2937 | } | |
2938 | tcp_xmit_retransmit_queue(sk); | |
2939 | } | |
2940 | EXPORT_SYMBOL(tcp_simple_retransmit); | |
2941 | ||
2942 | /* Process an event, which can update packets-in-flight not trivially. | |
2943 | * Main goal of this function is to calculate new estimate for left_out, | |
2944 | * taking into account both packets sitting in receiver's buffer and | |
2945 | * packets lost by network. | |
2946 | * | |
2947 | * Besides that it does CWND reduction, when packet loss is detected | |
2948 | * and changes state of machine. | |
2949 | * | |
2950 | * It does _not_ decide what to send, it is made in function | |
2951 | * tcp_xmit_retransmit_queue(). | |
2952 | */ | |
2953 | static void tcp_fastretrans_alert(struct sock *sk, int pkts_acked, int flag) | |
2954 | { | |
2955 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2956 | struct tcp_sock *tp = tcp_sk(sk); | |
2957 | int is_dupack = !(flag & (FLAG_SND_UNA_ADVANCED | FLAG_NOT_DUP)); | |
2958 | int do_lost = is_dupack || ((flag & FLAG_DATA_SACKED) && | |
2959 | (tcp_fackets_out(tp) > tp->reordering)); | |
2960 | int fast_rexmit = 0, mib_idx; | |
2961 | ||
2962 | if (WARN_ON(!tp->packets_out && tp->sacked_out)) | |
2963 | tp->sacked_out = 0; | |
2964 | if (WARN_ON(!tp->sacked_out && tp->fackets_out)) | |
2965 | tp->fackets_out = 0; | |
2966 | ||
2967 | /* Now state machine starts. | |
2968 | * A. ECE, hence prohibit cwnd undoing, the reduction is required. */ | |
2969 | if (flag & FLAG_ECE) | |
2970 | tp->prior_ssthresh = 0; | |
2971 | ||
2972 | /* B. In all the states check for reneging SACKs. */ | |
2973 | if (tcp_check_sack_reneging(sk, flag)) | |
2974 | return; | |
2975 | ||
2976 | /* C. Process data loss notification, provided it is valid. */ | |
2977 | if (tcp_is_fack(tp) && (flag & FLAG_DATA_LOST) && | |
2978 | before(tp->snd_una, tp->high_seq) && | |
2979 | icsk->icsk_ca_state != TCP_CA_Open && | |
2980 | tp->fackets_out > tp->reordering) { | |
2981 | tcp_mark_head_lost(sk, tp->fackets_out - tp->reordering, 0); | |
2982 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSS); | |
2983 | } | |
2984 | ||
2985 | /* D. Check consistency of the current state. */ | |
2986 | tcp_verify_left_out(tp); | |
2987 | ||
2988 | /* E. Check state exit conditions. State can be terminated | |
2989 | * when high_seq is ACKed. */ | |
2990 | if (icsk->icsk_ca_state == TCP_CA_Open) { | |
2991 | WARN_ON(tp->retrans_out != 0); | |
2992 | tp->retrans_stamp = 0; | |
2993 | } else if (!before(tp->snd_una, tp->high_seq)) { | |
2994 | switch (icsk->icsk_ca_state) { | |
2995 | case TCP_CA_Loss: | |
2996 | icsk->icsk_retransmits = 0; | |
2997 | if (tcp_try_undo_recovery(sk)) | |
2998 | return; | |
2999 | break; | |
3000 | ||
3001 | case TCP_CA_CWR: | |
3002 | /* CWR is to be held something *above* high_seq | |
3003 | * is ACKed for CWR bit to reach receiver. */ | |
3004 | if (tp->snd_una != tp->high_seq) { | |
3005 | tcp_complete_cwr(sk); | |
3006 | tcp_set_ca_state(sk, TCP_CA_Open); | |
3007 | } | |
3008 | break; | |
3009 | ||
3010 | case TCP_CA_Disorder: | |
3011 | tcp_try_undo_dsack(sk); | |
3012 | if (!tp->undo_marker || | |
3013 | /* For SACK case do not Open to allow to undo | |
3014 | * catching for all duplicate ACKs. */ | |
3015 | tcp_is_reno(tp) || tp->snd_una != tp->high_seq) { | |
3016 | tp->undo_marker = 0; | |
3017 | tcp_set_ca_state(sk, TCP_CA_Open); | |
3018 | } | |
3019 | break; | |
3020 | ||
3021 | case TCP_CA_Recovery: | |
3022 | if (tcp_is_reno(tp)) | |
3023 | tcp_reset_reno_sack(tp); | |
3024 | if (tcp_try_undo_recovery(sk)) | |
3025 | return; | |
3026 | tcp_complete_cwr(sk); | |
3027 | break; | |
3028 | } | |
3029 | } | |
3030 | ||
3031 | /* F. Process state. */ | |
3032 | switch (icsk->icsk_ca_state) { | |
3033 | case TCP_CA_Recovery: | |
3034 | if (!(flag & FLAG_SND_UNA_ADVANCED)) { | |
3035 | if (tcp_is_reno(tp) && is_dupack) | |
3036 | tcp_add_reno_sack(sk); | |
3037 | } else | |
3038 | do_lost = tcp_try_undo_partial(sk, pkts_acked); | |
3039 | break; | |
3040 | case TCP_CA_Loss: | |
3041 | if (flag & FLAG_DATA_ACKED) | |
3042 | icsk->icsk_retransmits = 0; | |
3043 | if (tcp_is_reno(tp) && flag & FLAG_SND_UNA_ADVANCED) | |
3044 | tcp_reset_reno_sack(tp); | |
3045 | if (!tcp_try_undo_loss(sk)) { | |
3046 | tcp_moderate_cwnd(tp); | |
3047 | tcp_xmit_retransmit_queue(sk); | |
3048 | return; | |
3049 | } | |
3050 | if (icsk->icsk_ca_state != TCP_CA_Open) | |
3051 | return; | |
3052 | /* Loss is undone; fall through to processing in Open state. */ | |
3053 | default: | |
3054 | if (tcp_is_reno(tp)) { | |
3055 | if (flag & FLAG_SND_UNA_ADVANCED) | |
3056 | tcp_reset_reno_sack(tp); | |
3057 | if (is_dupack) | |
3058 | tcp_add_reno_sack(sk); | |
3059 | } | |
3060 | ||
3061 | if (icsk->icsk_ca_state == TCP_CA_Disorder) | |
3062 | tcp_try_undo_dsack(sk); | |
3063 | ||
3064 | if (!tcp_time_to_recover(sk)) { | |
3065 | tcp_try_to_open(sk, flag); | |
3066 | return; | |
3067 | } | |
3068 | ||
3069 | /* MTU probe failure: don't reduce cwnd */ | |
3070 | if (icsk->icsk_ca_state < TCP_CA_CWR && | |
3071 | icsk->icsk_mtup.probe_size && | |
3072 | tp->snd_una == tp->mtu_probe.probe_seq_start) { | |
3073 | tcp_mtup_probe_failed(sk); | |
3074 | /* Restores the reduction we did in tcp_mtup_probe() */ | |
3075 | tp->snd_cwnd++; | |
3076 | tcp_simple_retransmit(sk); | |
3077 | return; | |
3078 | } | |
3079 | ||
3080 | /* Otherwise enter Recovery state */ | |
3081 | ||
3082 | if (tcp_is_reno(tp)) | |
3083 | mib_idx = LINUX_MIB_TCPRENORECOVERY; | |
3084 | else | |
3085 | mib_idx = LINUX_MIB_TCPSACKRECOVERY; | |
3086 | ||
3087 | NET_INC_STATS_BH(sock_net(sk), mib_idx); | |
3088 | ||
3089 | tp->high_seq = tp->snd_nxt; | |
3090 | tp->prior_ssthresh = 0; | |
3091 | tp->undo_marker = tp->snd_una; | |
3092 | tp->undo_retrans = tp->retrans_out; | |
3093 | ||
3094 | if (icsk->icsk_ca_state < TCP_CA_CWR) { | |
3095 | if (!(flag & FLAG_ECE)) | |
3096 | tp->prior_ssthresh = tcp_current_ssthresh(sk); | |
3097 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
3098 | TCP_ECN_queue_cwr(tp); | |
3099 | } | |
3100 | ||
3101 | tp->bytes_acked = 0; | |
3102 | tp->snd_cwnd_cnt = 0; | |
3103 | tcp_set_ca_state(sk, TCP_CA_Recovery); | |
3104 | fast_rexmit = 1; | |
3105 | } | |
3106 | ||
3107 | if (do_lost || (tcp_is_fack(tp) && tcp_head_timedout(sk))) | |
3108 | tcp_update_scoreboard(sk, fast_rexmit); | |
3109 | tcp_cwnd_down(sk, flag); | |
3110 | tcp_xmit_retransmit_queue(sk); | |
3111 | } | |
3112 | ||
3113 | static void tcp_valid_rtt_meas(struct sock *sk, u32 seq_rtt) | |
3114 | { | |
3115 | tcp_rtt_estimator(sk, seq_rtt); | |
3116 | tcp_set_rto(sk); | |
3117 | inet_csk(sk)->icsk_backoff = 0; | |
3118 | } | |
3119 | ||
3120 | /* Read draft-ietf-tcplw-high-performance before mucking | |
3121 | * with this code. (Supersedes RFC1323) | |
3122 | */ | |
3123 | static void tcp_ack_saw_tstamp(struct sock *sk, int flag) | |
3124 | { | |
3125 | /* RTTM Rule: A TSecr value received in a segment is used to | |
3126 | * update the averaged RTT measurement only if the segment | |
3127 | * acknowledges some new data, i.e., only if it advances the | |
3128 | * left edge of the send window. | |
3129 | * | |
3130 | * See draft-ietf-tcplw-high-performance-00, section 3.3. | |
3131 | * 1998/04/10 Andrey V. Savochkin <saw@msu.ru> | |
3132 | * | |
3133 | * Changed: reset backoff as soon as we see the first valid sample. | |
3134 | * If we do not, we get strongly overestimated rto. With timestamps | |
3135 | * samples are accepted even from very old segments: f.e., when rtt=1 | |
3136 | * increases to 8, we retransmit 5 times and after 8 seconds delayed | |
3137 | * answer arrives rto becomes 120 seconds! If at least one of segments | |
3138 | * in window is lost... Voila. --ANK (010210) | |
3139 | */ | |
3140 | struct tcp_sock *tp = tcp_sk(sk); | |
3141 | ||
3142 | tcp_valid_rtt_meas(sk, tcp_time_stamp - tp->rx_opt.rcv_tsecr); | |
3143 | } | |
3144 | ||
3145 | static void tcp_ack_no_tstamp(struct sock *sk, u32 seq_rtt, int flag) | |
3146 | { | |
3147 | /* We don't have a timestamp. Can only use | |
3148 | * packets that are not retransmitted to determine | |
3149 | * rtt estimates. Also, we must not reset the | |
3150 | * backoff for rto until we get a non-retransmitted | |
3151 | * packet. This allows us to deal with a situation | |
3152 | * where the network delay has increased suddenly. | |
3153 | * I.e. Karn's algorithm. (SIGCOMM '87, p5.) | |
3154 | */ | |
3155 | ||
3156 | if (flag & FLAG_RETRANS_DATA_ACKED) | |
3157 | return; | |
3158 | ||
3159 | tcp_valid_rtt_meas(sk, seq_rtt); | |
3160 | } | |
3161 | ||
3162 | static inline void tcp_ack_update_rtt(struct sock *sk, const int flag, | |
3163 | const s32 seq_rtt) | |
3164 | { | |
3165 | const struct tcp_sock *tp = tcp_sk(sk); | |
3166 | /* Note that peer MAY send zero echo. In this case it is ignored. (rfc1323) */ | |
3167 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr) | |
3168 | tcp_ack_saw_tstamp(sk, flag); | |
3169 | else if (seq_rtt >= 0) | |
3170 | tcp_ack_no_tstamp(sk, seq_rtt, flag); | |
3171 | } | |
3172 | ||
3173 | static void tcp_cong_avoid(struct sock *sk, u32 ack, u32 in_flight) | |
3174 | { | |
3175 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
3176 | icsk->icsk_ca_ops->cong_avoid(sk, ack, in_flight); | |
3177 | tcp_sk(sk)->snd_cwnd_stamp = tcp_time_stamp; | |
3178 | } | |
3179 | ||
3180 | /* Restart timer after forward progress on connection. | |
3181 | * RFC2988 recommends to restart timer to now+rto. | |
3182 | */ | |
3183 | static void tcp_rearm_rto(struct sock *sk) | |
3184 | { | |
3185 | struct tcp_sock *tp = tcp_sk(sk); | |
3186 | ||
3187 | if (!tp->packets_out) { | |
3188 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_RETRANS); | |
3189 | } else { | |
3190 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, | |
3191 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
3192 | } | |
3193 | } | |
3194 | ||
3195 | /* If we get here, the whole TSO packet has not been acked. */ | |
3196 | static u32 tcp_tso_acked(struct sock *sk, struct sk_buff *skb) | |
3197 | { | |
3198 | struct tcp_sock *tp = tcp_sk(sk); | |
3199 | u32 packets_acked; | |
3200 | ||
3201 | BUG_ON(!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)); | |
3202 | ||
3203 | packets_acked = tcp_skb_pcount(skb); | |
3204 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) | |
3205 | return 0; | |
3206 | packets_acked -= tcp_skb_pcount(skb); | |
3207 | ||
3208 | if (packets_acked) { | |
3209 | BUG_ON(tcp_skb_pcount(skb) == 0); | |
3210 | BUG_ON(!before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)); | |
3211 | } | |
3212 | ||
3213 | return packets_acked; | |
3214 | } | |
3215 | ||
3216 | /* Remove acknowledged frames from the retransmission queue. If our packet | |
3217 | * is before the ack sequence we can discard it as it's confirmed to have | |
3218 | * arrived at the other end. | |
3219 | */ | |
3220 | static int tcp_clean_rtx_queue(struct sock *sk, int prior_fackets, | |
3221 | u32 prior_snd_una) | |
3222 | { | |
3223 | struct tcp_sock *tp = tcp_sk(sk); | |
3224 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
3225 | struct sk_buff *skb; | |
3226 | u32 now = tcp_time_stamp; | |
3227 | int fully_acked = 1; | |
3228 | int flag = 0; | |
3229 | u32 pkts_acked = 0; | |
3230 | u32 reord = tp->packets_out; | |
3231 | u32 prior_sacked = tp->sacked_out; | |
3232 | s32 seq_rtt = -1; | |
3233 | s32 ca_seq_rtt = -1; | |
3234 | ktime_t last_ackt = net_invalid_timestamp(); | |
3235 | ||
3236 | while ((skb = tcp_write_queue_head(sk)) && skb != tcp_send_head(sk)) { | |
3237 | struct tcp_skb_cb *scb = TCP_SKB_CB(skb); | |
3238 | u32 acked_pcount; | |
3239 | u8 sacked = scb->sacked; | |
3240 | ||
3241 | /* Determine how many packets and what bytes were acked, tso and else */ | |
3242 | if (after(scb->end_seq, tp->snd_una)) { | |
3243 | if (tcp_skb_pcount(skb) == 1 || | |
3244 | !after(tp->snd_una, scb->seq)) | |
3245 | break; | |
3246 | ||
3247 | acked_pcount = tcp_tso_acked(sk, skb); | |
3248 | if (!acked_pcount) | |
3249 | break; | |
3250 | ||
3251 | fully_acked = 0; | |
3252 | } else { | |
3253 | acked_pcount = tcp_skb_pcount(skb); | |
3254 | } | |
3255 | ||
3256 | if (sacked & TCPCB_RETRANS) { | |
3257 | if (sacked & TCPCB_SACKED_RETRANS) | |
3258 | tp->retrans_out -= acked_pcount; | |
3259 | flag |= FLAG_RETRANS_DATA_ACKED; | |
3260 | ca_seq_rtt = -1; | |
3261 | seq_rtt = -1; | |
3262 | if ((flag & FLAG_DATA_ACKED) || (acked_pcount > 1)) | |
3263 | flag |= FLAG_NONHEAD_RETRANS_ACKED; | |
3264 | } else { | |
3265 | ca_seq_rtt = now - scb->when; | |
3266 | last_ackt = skb->tstamp; | |
3267 | if (seq_rtt < 0) { | |
3268 | seq_rtt = ca_seq_rtt; | |
3269 | } | |
3270 | if (!(sacked & TCPCB_SACKED_ACKED)) | |
3271 | reord = min(pkts_acked, reord); | |
3272 | } | |
3273 | ||
3274 | if (sacked & TCPCB_SACKED_ACKED) | |
3275 | tp->sacked_out -= acked_pcount; | |
3276 | if (sacked & TCPCB_LOST) | |
3277 | tp->lost_out -= acked_pcount; | |
3278 | ||
3279 | tp->packets_out -= acked_pcount; | |
3280 | pkts_acked += acked_pcount; | |
3281 | ||
3282 | /* Initial outgoing SYN's get put onto the write_queue | |
3283 | * just like anything else we transmit. It is not | |
3284 | * true data, and if we misinform our callers that | |
3285 | * this ACK acks real data, we will erroneously exit | |
3286 | * connection startup slow start one packet too | |
3287 | * quickly. This is severely frowned upon behavior. | |
3288 | */ | |
3289 | if (!(scb->flags & TCPHDR_SYN)) { | |
3290 | flag |= FLAG_DATA_ACKED; | |
3291 | } else { | |
3292 | flag |= FLAG_SYN_ACKED; | |
3293 | tp->retrans_stamp = 0; | |
3294 | } | |
3295 | ||
3296 | if (!fully_acked) | |
3297 | break; | |
3298 | ||
3299 | tcp_unlink_write_queue(skb, sk); | |
3300 | sk_wmem_free_skb(sk, skb); | |
3301 | tp->scoreboard_skb_hint = NULL; | |
3302 | if (skb == tp->retransmit_skb_hint) | |
3303 | tp->retransmit_skb_hint = NULL; | |
3304 | if (skb == tp->lost_skb_hint) | |
3305 | tp->lost_skb_hint = NULL; | |
3306 | } | |
3307 | ||
3308 | if (likely(between(tp->snd_up, prior_snd_una, tp->snd_una))) | |
3309 | tp->snd_up = tp->snd_una; | |
3310 | ||
3311 | if (skb && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) | |
3312 | flag |= FLAG_SACK_RENEGING; | |
3313 | ||
3314 | if (flag & FLAG_ACKED) { | |
3315 | const struct tcp_congestion_ops *ca_ops | |
3316 | = inet_csk(sk)->icsk_ca_ops; | |
3317 | ||
3318 | if (unlikely(icsk->icsk_mtup.probe_size && | |
3319 | !after(tp->mtu_probe.probe_seq_end, tp->snd_una))) { | |
3320 | tcp_mtup_probe_success(sk); | |
3321 | } | |
3322 | ||
3323 | tcp_ack_update_rtt(sk, flag, seq_rtt); | |
3324 | tcp_rearm_rto(sk); | |
3325 | ||
3326 | if (tcp_is_reno(tp)) { | |
3327 | tcp_remove_reno_sacks(sk, pkts_acked); | |
3328 | } else { | |
3329 | int delta; | |
3330 | ||
3331 | /* Non-retransmitted hole got filled? That's reordering */ | |
3332 | if (reord < prior_fackets) | |
3333 | tcp_update_reordering(sk, tp->fackets_out - reord, 0); | |
3334 | ||
3335 | delta = tcp_is_fack(tp) ? pkts_acked : | |
3336 | prior_sacked - tp->sacked_out; | |
3337 | tp->lost_cnt_hint -= min(tp->lost_cnt_hint, delta); | |
3338 | } | |
3339 | ||
3340 | tp->fackets_out -= min(pkts_acked, tp->fackets_out); | |
3341 | ||
3342 | if (ca_ops->pkts_acked) { | |
3343 | s32 rtt_us = -1; | |
3344 | ||
3345 | /* Is the ACK triggering packet unambiguous? */ | |
3346 | if (!(flag & FLAG_RETRANS_DATA_ACKED)) { | |
3347 | /* High resolution needed and available? */ | |
3348 | if (ca_ops->flags & TCP_CONG_RTT_STAMP && | |
3349 | !ktime_equal(last_ackt, | |
3350 | net_invalid_timestamp())) | |
3351 | rtt_us = ktime_us_delta(ktime_get_real(), | |
3352 | last_ackt); | |
3353 | else if (ca_seq_rtt > 0) | |
3354 | rtt_us = jiffies_to_usecs(ca_seq_rtt); | |
3355 | } | |
3356 | ||
3357 | ca_ops->pkts_acked(sk, pkts_acked, rtt_us); | |
3358 | } | |
3359 | } | |
3360 | ||
3361 | #if FASTRETRANS_DEBUG > 0 | |
3362 | WARN_ON((int)tp->sacked_out < 0); | |
3363 | WARN_ON((int)tp->lost_out < 0); | |
3364 | WARN_ON((int)tp->retrans_out < 0); | |
3365 | if (!tp->packets_out && tcp_is_sack(tp)) { | |
3366 | icsk = inet_csk(sk); | |
3367 | if (tp->lost_out) { | |
3368 | printk(KERN_DEBUG "Leak l=%u %d\n", | |
3369 | tp->lost_out, icsk->icsk_ca_state); | |
3370 | tp->lost_out = 0; | |
3371 | } | |
3372 | if (tp->sacked_out) { | |
3373 | printk(KERN_DEBUG "Leak s=%u %d\n", | |
3374 | tp->sacked_out, icsk->icsk_ca_state); | |
3375 | tp->sacked_out = 0; | |
3376 | } | |
3377 | if (tp->retrans_out) { | |
3378 | printk(KERN_DEBUG "Leak r=%u %d\n", | |
3379 | tp->retrans_out, icsk->icsk_ca_state); | |
3380 | tp->retrans_out = 0; | |
3381 | } | |
3382 | } | |
3383 | #endif | |
3384 | return flag; | |
3385 | } | |
3386 | ||
3387 | static void tcp_ack_probe(struct sock *sk) | |
3388 | { | |
3389 | const struct tcp_sock *tp = tcp_sk(sk); | |
3390 | struct inet_connection_sock *icsk = inet_csk(sk); | |
3391 | ||
3392 | /* Was it a usable window open? */ | |
3393 | ||
3394 | if (!after(TCP_SKB_CB(tcp_send_head(sk))->end_seq, tcp_wnd_end(tp))) { | |
3395 | icsk->icsk_backoff = 0; | |
3396 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_PROBE0); | |
3397 | /* Socket must be waked up by subsequent tcp_data_snd_check(). | |
3398 | * This function is not for random using! | |
3399 | */ | |
3400 | } else { | |
3401 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, | |
3402 | min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), | |
3403 | TCP_RTO_MAX); | |
3404 | } | |
3405 | } | |
3406 | ||
3407 | static inline int tcp_ack_is_dubious(const struct sock *sk, const int flag) | |
3408 | { | |
3409 | return !(flag & FLAG_NOT_DUP) || (flag & FLAG_CA_ALERT) || | |
3410 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open; | |
3411 | } | |
3412 | ||
3413 | static inline int tcp_may_raise_cwnd(const struct sock *sk, const int flag) | |
3414 | { | |
3415 | const struct tcp_sock *tp = tcp_sk(sk); | |
3416 | return (!(flag & FLAG_ECE) || tp->snd_cwnd < tp->snd_ssthresh) && | |
3417 | !((1 << inet_csk(sk)->icsk_ca_state) & (TCPF_CA_Recovery | TCPF_CA_CWR)); | |
3418 | } | |
3419 | ||
3420 | /* Check that window update is acceptable. | |
3421 | * The function assumes that snd_una<=ack<=snd_next. | |
3422 | */ | |
3423 | static inline int tcp_may_update_window(const struct tcp_sock *tp, | |
3424 | const u32 ack, const u32 ack_seq, | |
3425 | const u32 nwin) | |
3426 | { | |
3427 | return after(ack, tp->snd_una) || | |
3428 | after(ack_seq, tp->snd_wl1) || | |
3429 | (ack_seq == tp->snd_wl1 && nwin > tp->snd_wnd); | |
3430 | } | |
3431 | ||
3432 | /* Update our send window. | |
3433 | * | |
3434 | * Window update algorithm, described in RFC793/RFC1122 (used in linux-2.2 | |
3435 | * and in FreeBSD. NetBSD's one is even worse.) is wrong. | |
3436 | */ | |
3437 | static int tcp_ack_update_window(struct sock *sk, struct sk_buff *skb, u32 ack, | |
3438 | u32 ack_seq) | |
3439 | { | |
3440 | struct tcp_sock *tp = tcp_sk(sk); | |
3441 | int flag = 0; | |
3442 | u32 nwin = ntohs(tcp_hdr(skb)->window); | |
3443 | ||
3444 | if (likely(!tcp_hdr(skb)->syn)) | |
3445 | nwin <<= tp->rx_opt.snd_wscale; | |
3446 | ||
3447 | if (tcp_may_update_window(tp, ack, ack_seq, nwin)) { | |
3448 | flag |= FLAG_WIN_UPDATE; | |
3449 | tcp_update_wl(tp, ack_seq); | |
3450 | ||
3451 | if (tp->snd_wnd != nwin) { | |
3452 | tp->snd_wnd = nwin; | |
3453 | ||
3454 | /* Note, it is the only place, where | |
3455 | * fast path is recovered for sending TCP. | |
3456 | */ | |
3457 | tp->pred_flags = 0; | |
3458 | tcp_fast_path_check(sk); | |
3459 | ||
3460 | if (nwin > tp->max_window) { | |
3461 | tp->max_window = nwin; | |
3462 | tcp_sync_mss(sk, inet_csk(sk)->icsk_pmtu_cookie); | |
3463 | } | |
3464 | } | |
3465 | } | |
3466 | ||
3467 | tp->snd_una = ack; | |
3468 | ||
3469 | return flag; | |
3470 | } | |
3471 | ||
3472 | /* A very conservative spurious RTO response algorithm: reduce cwnd and | |
3473 | * continue in congestion avoidance. | |
3474 | */ | |
3475 | static void tcp_conservative_spur_to_response(struct tcp_sock *tp) | |
3476 | { | |
3477 | tp->snd_cwnd = min(tp->snd_cwnd, tp->snd_ssthresh); | |
3478 | tp->snd_cwnd_cnt = 0; | |
3479 | tp->bytes_acked = 0; | |
3480 | TCP_ECN_queue_cwr(tp); | |
3481 | tcp_moderate_cwnd(tp); | |
3482 | } | |
3483 | ||
3484 | /* A conservative spurious RTO response algorithm: reduce cwnd using | |
3485 | * rate halving and continue in congestion avoidance. | |
3486 | */ | |
3487 | static void tcp_ratehalving_spur_to_response(struct sock *sk) | |
3488 | { | |
3489 | tcp_enter_cwr(sk, 0); | |
3490 | } | |
3491 | ||
3492 | static void tcp_undo_spur_to_response(struct sock *sk, int flag) | |
3493 | { | |
3494 | if (flag & FLAG_ECE) | |
3495 | tcp_ratehalving_spur_to_response(sk); | |
3496 | else | |
3497 | tcp_undo_cwr(sk, 1); | |
3498 | } | |
3499 | ||
3500 | /* F-RTO spurious RTO detection algorithm (RFC4138) | |
3501 | * | |
3502 | * F-RTO affects during two new ACKs following RTO (well, almost, see inline | |
3503 | * comments). State (ACK number) is kept in frto_counter. When ACK advances | |
3504 | * window (but not to or beyond highest sequence sent before RTO): | |
3505 | * On First ACK, send two new segments out. | |
3506 | * On Second ACK, RTO was likely spurious. Do spurious response (response | |
3507 | * algorithm is not part of the F-RTO detection algorithm | |
3508 | * given in RFC4138 but can be selected separately). | |
3509 | * Otherwise (basically on duplicate ACK), RTO was (likely) caused by a loss | |
3510 | * and TCP falls back to conventional RTO recovery. F-RTO allows overriding | |
3511 | * of Nagle, this is done using frto_counter states 2 and 3, when a new data | |
3512 | * segment of any size sent during F-RTO, state 2 is upgraded to 3. | |
3513 | * | |
3514 | * Rationale: if the RTO was spurious, new ACKs should arrive from the | |
3515 | * original window even after we transmit two new data segments. | |
3516 | * | |
3517 | * SACK version: | |
3518 | * on first step, wait until first cumulative ACK arrives, then move to | |
3519 | * the second step. In second step, the next ACK decides. | |
3520 | * | |
3521 | * F-RTO is implemented (mainly) in four functions: | |
3522 | * - tcp_use_frto() is used to determine if TCP is can use F-RTO | |
3523 | * - tcp_enter_frto() prepares TCP state on RTO if F-RTO is used, it is | |
3524 | * called when tcp_use_frto() showed green light | |
3525 | * - tcp_process_frto() handles incoming ACKs during F-RTO algorithm | |
3526 | * - tcp_enter_frto_loss() is called if there is not enough evidence | |
3527 | * to prove that the RTO is indeed spurious. It transfers the control | |
3528 | * from F-RTO to the conventional RTO recovery | |
3529 | */ | |
3530 | static int tcp_process_frto(struct sock *sk, int flag) | |
3531 | { | |
3532 | struct tcp_sock *tp = tcp_sk(sk); | |
3533 | ||
3534 | tcp_verify_left_out(tp); | |
3535 | ||
3536 | /* Duplicate the behavior from Loss state (fastretrans_alert) */ | |
3537 | if (flag & FLAG_DATA_ACKED) | |
3538 | inet_csk(sk)->icsk_retransmits = 0; | |
3539 | ||
3540 | if ((flag & FLAG_NONHEAD_RETRANS_ACKED) || | |
3541 | ((tp->frto_counter >= 2) && (flag & FLAG_RETRANS_DATA_ACKED))) | |
3542 | tp->undo_marker = 0; | |
3543 | ||
3544 | if (!before(tp->snd_una, tp->frto_highmark)) { | |
3545 | tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 2 : 3), flag); | |
3546 | return 1; | |
3547 | } | |
3548 | ||
3549 | if (!tcp_is_sackfrto(tp)) { | |
3550 | /* RFC4138 shortcoming in step 2; should also have case c): | |
3551 | * ACK isn't duplicate nor advances window, e.g., opposite dir | |
3552 | * data, winupdate | |
3553 | */ | |
3554 | if (!(flag & FLAG_ANY_PROGRESS) && (flag & FLAG_NOT_DUP)) | |
3555 | return 1; | |
3556 | ||
3557 | if (!(flag & FLAG_DATA_ACKED)) { | |
3558 | tcp_enter_frto_loss(sk, (tp->frto_counter == 1 ? 0 : 3), | |
3559 | flag); | |
3560 | return 1; | |
3561 | } | |
3562 | } else { | |
3563 | if (!(flag & FLAG_DATA_ACKED) && (tp->frto_counter == 1)) { | |
3564 | /* Prevent sending of new data. */ | |
3565 | tp->snd_cwnd = min(tp->snd_cwnd, | |
3566 | tcp_packets_in_flight(tp)); | |
3567 | return 1; | |
3568 | } | |
3569 | ||
3570 | if ((tp->frto_counter >= 2) && | |
3571 | (!(flag & FLAG_FORWARD_PROGRESS) || | |
3572 | ((flag & FLAG_DATA_SACKED) && | |
3573 | !(flag & FLAG_ONLY_ORIG_SACKED)))) { | |
3574 | /* RFC4138 shortcoming (see comment above) */ | |
3575 | if (!(flag & FLAG_FORWARD_PROGRESS) && | |
3576 | (flag & FLAG_NOT_DUP)) | |
3577 | return 1; | |
3578 | ||
3579 | tcp_enter_frto_loss(sk, 3, flag); | |
3580 | return 1; | |
3581 | } | |
3582 | } | |
3583 | ||
3584 | if (tp->frto_counter == 1) { | |
3585 | /* tcp_may_send_now needs to see updated state */ | |
3586 | tp->snd_cwnd = tcp_packets_in_flight(tp) + 2; | |
3587 | tp->frto_counter = 2; | |
3588 | ||
3589 | if (!tcp_may_send_now(sk)) | |
3590 | tcp_enter_frto_loss(sk, 2, flag); | |
3591 | ||
3592 | return 1; | |
3593 | } else { | |
3594 | switch (sysctl_tcp_frto_response) { | |
3595 | case 2: | |
3596 | tcp_undo_spur_to_response(sk, flag); | |
3597 | break; | |
3598 | case 1: | |
3599 | tcp_conservative_spur_to_response(tp); | |
3600 | break; | |
3601 | default: | |
3602 | tcp_ratehalving_spur_to_response(sk); | |
3603 | break; | |
3604 | } | |
3605 | tp->frto_counter = 0; | |
3606 | tp->undo_marker = 0; | |
3607 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSPURIOUSRTOS); | |
3608 | } | |
3609 | return 0; | |
3610 | } | |
3611 | ||
3612 | /* This routine deals with incoming acks, but not outgoing ones. */ | |
3613 | static int tcp_ack(struct sock *sk, struct sk_buff *skb, int flag) | |
3614 | { | |
3615 | struct inet_connection_sock *icsk = inet_csk(sk); | |
3616 | struct tcp_sock *tp = tcp_sk(sk); | |
3617 | u32 prior_snd_una = tp->snd_una; | |
3618 | u32 ack_seq = TCP_SKB_CB(skb)->seq; | |
3619 | u32 ack = TCP_SKB_CB(skb)->ack_seq; | |
3620 | u32 prior_in_flight; | |
3621 | u32 prior_fackets; | |
3622 | int prior_packets; | |
3623 | int frto_cwnd = 0; | |
3624 | ||
3625 | /* If the ack is older than previous acks | |
3626 | * then we can probably ignore it. | |
3627 | */ | |
3628 | if (before(ack, prior_snd_una)) | |
3629 | goto old_ack; | |
3630 | ||
3631 | /* If the ack includes data we haven't sent yet, discard | |
3632 | * this segment (RFC793 Section 3.9). | |
3633 | */ | |
3634 | if (after(ack, tp->snd_nxt)) | |
3635 | goto invalid_ack; | |
3636 | ||
3637 | if (after(ack, prior_snd_una)) | |
3638 | flag |= FLAG_SND_UNA_ADVANCED; | |
3639 | ||
3640 | if (sysctl_tcp_abc) { | |
3641 | if (icsk->icsk_ca_state < TCP_CA_CWR) | |
3642 | tp->bytes_acked += ack - prior_snd_una; | |
3643 | else if (icsk->icsk_ca_state == TCP_CA_Loss) | |
3644 | /* we assume just one segment left network */ | |
3645 | tp->bytes_acked += min(ack - prior_snd_una, | |
3646 | tp->mss_cache); | |
3647 | } | |
3648 | ||
3649 | prior_fackets = tp->fackets_out; | |
3650 | prior_in_flight = tcp_packets_in_flight(tp); | |
3651 | ||
3652 | if (!(flag & FLAG_SLOWPATH) && after(ack, prior_snd_una)) { | |
3653 | /* Window is constant, pure forward advance. | |
3654 | * No more checks are required. | |
3655 | * Note, we use the fact that SND.UNA>=SND.WL2. | |
3656 | */ | |
3657 | tcp_update_wl(tp, ack_seq); | |
3658 | tp->snd_una = ack; | |
3659 | flag |= FLAG_WIN_UPDATE; | |
3660 | ||
3661 | tcp_ca_event(sk, CA_EVENT_FAST_ACK); | |
3662 | ||
3663 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPACKS); | |
3664 | } else { | |
3665 | if (ack_seq != TCP_SKB_CB(skb)->end_seq) | |
3666 | flag |= FLAG_DATA; | |
3667 | else | |
3668 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPPUREACKS); | |
3669 | ||
3670 | flag |= tcp_ack_update_window(sk, skb, ack, ack_seq); | |
3671 | ||
3672 | if (TCP_SKB_CB(skb)->sacked) | |
3673 | flag |= tcp_sacktag_write_queue(sk, skb, prior_snd_una); | |
3674 | ||
3675 | if (TCP_ECN_rcv_ecn_echo(tp, tcp_hdr(skb))) | |
3676 | flag |= FLAG_ECE; | |
3677 | ||
3678 | tcp_ca_event(sk, CA_EVENT_SLOW_ACK); | |
3679 | } | |
3680 | ||
3681 | /* We passed data and got it acked, remove any soft error | |
3682 | * log. Something worked... | |
3683 | */ | |
3684 | sk->sk_err_soft = 0; | |
3685 | icsk->icsk_probes_out = 0; | |
3686 | tp->rcv_tstamp = tcp_time_stamp; | |
3687 | prior_packets = tp->packets_out; | |
3688 | if (!prior_packets) | |
3689 | goto no_queue; | |
3690 | ||
3691 | /* See if we can take anything off of the retransmit queue. */ | |
3692 | flag |= tcp_clean_rtx_queue(sk, prior_fackets, prior_snd_una); | |
3693 | ||
3694 | if (tp->frto_counter) | |
3695 | frto_cwnd = tcp_process_frto(sk, flag); | |
3696 | /* Guarantee sacktag reordering detection against wrap-arounds */ | |
3697 | if (before(tp->frto_highmark, tp->snd_una)) | |
3698 | tp->frto_highmark = 0; | |
3699 | ||
3700 | if (tcp_ack_is_dubious(sk, flag)) { | |
3701 | /* Advance CWND, if state allows this. */ | |
3702 | if ((flag & FLAG_DATA_ACKED) && !frto_cwnd && | |
3703 | tcp_may_raise_cwnd(sk, flag)) | |
3704 | tcp_cong_avoid(sk, ack, prior_in_flight); | |
3705 | tcp_fastretrans_alert(sk, prior_packets - tp->packets_out, | |
3706 | flag); | |
3707 | } else { | |
3708 | if ((flag & FLAG_DATA_ACKED) && !frto_cwnd) | |
3709 | tcp_cong_avoid(sk, ack, prior_in_flight); | |
3710 | } | |
3711 | ||
3712 | if ((flag & FLAG_FORWARD_PROGRESS) || !(flag & FLAG_NOT_DUP)) | |
3713 | dst_confirm(__sk_dst_get(sk)); | |
3714 | ||
3715 | return 1; | |
3716 | ||
3717 | no_queue: | |
3718 | /* If this ack opens up a zero window, clear backoff. It was | |
3719 | * being used to time the probes, and is probably far higher than | |
3720 | * it needs to be for normal retransmission. | |
3721 | */ | |
3722 | if (tcp_send_head(sk)) | |
3723 | tcp_ack_probe(sk); | |
3724 | return 1; | |
3725 | ||
3726 | invalid_ack: | |
3727 | SOCK_DEBUG(sk, "Ack %u after %u:%u\n", ack, tp->snd_una, tp->snd_nxt); | |
3728 | return -1; | |
3729 | ||
3730 | old_ack: | |
3731 | if (TCP_SKB_CB(skb)->sacked) { | |
3732 | tcp_sacktag_write_queue(sk, skb, prior_snd_una); | |
3733 | if (icsk->icsk_ca_state == TCP_CA_Open) | |
3734 | tcp_try_keep_open(sk); | |
3735 | } | |
3736 | ||
3737 | SOCK_DEBUG(sk, "Ack %u before %u:%u\n", ack, tp->snd_una, tp->snd_nxt); | |
3738 | return 0; | |
3739 | } | |
3740 | ||
3741 | /* Look for tcp options. Normally only called on SYN and SYNACK packets. | |
3742 | * But, this can also be called on packets in the established flow when | |
3743 | * the fast version below fails. | |
3744 | */ | |
3745 | void tcp_parse_options(struct sk_buff *skb, struct tcp_options_received *opt_rx, | |
3746 | u8 **hvpp, int estab) | |
3747 | { | |
3748 | unsigned char *ptr; | |
3749 | struct tcphdr *th = tcp_hdr(skb); | |
3750 | int length = (th->doff * 4) - sizeof(struct tcphdr); | |
3751 | ||
3752 | ptr = (unsigned char *)(th + 1); | |
3753 | opt_rx->saw_tstamp = 0; | |
3754 | ||
3755 | while (length > 0) { | |
3756 | int opcode = *ptr++; | |
3757 | int opsize; | |
3758 | ||
3759 | switch (opcode) { | |
3760 | case TCPOPT_EOL: | |
3761 | return; | |
3762 | case TCPOPT_NOP: /* Ref: RFC 793 section 3.1 */ | |
3763 | length--; | |
3764 | continue; | |
3765 | default: | |
3766 | opsize = *ptr++; | |
3767 | if (opsize < 2) /* "silly options" */ | |
3768 | return; | |
3769 | if (opsize > length) | |
3770 | return; /* don't parse partial options */ | |
3771 | switch (opcode) { | |
3772 | case TCPOPT_MSS: | |
3773 | if (opsize == TCPOLEN_MSS && th->syn && !estab) { | |
3774 | u16 in_mss = get_unaligned_be16(ptr); | |
3775 | if (in_mss) { | |
3776 | if (opt_rx->user_mss && | |
3777 | opt_rx->user_mss < in_mss) | |
3778 | in_mss = opt_rx->user_mss; | |
3779 | opt_rx->mss_clamp = in_mss; | |
3780 | } | |
3781 | } | |
3782 | break; | |
3783 | case TCPOPT_WINDOW: | |
3784 | if (opsize == TCPOLEN_WINDOW && th->syn && | |
3785 | !estab && sysctl_tcp_window_scaling) { | |
3786 | __u8 snd_wscale = *(__u8 *)ptr; | |
3787 | opt_rx->wscale_ok = 1; | |
3788 | if (snd_wscale > 14) { | |
3789 | if (net_ratelimit()) | |
3790 | printk(KERN_INFO "tcp_parse_options: Illegal window " | |
3791 | "scaling value %d >14 received.\n", | |
3792 | snd_wscale); | |
3793 | snd_wscale = 14; | |
3794 | } | |
3795 | opt_rx->snd_wscale = snd_wscale; | |
3796 | } | |
3797 | break; | |
3798 | case TCPOPT_TIMESTAMP: | |
3799 | if ((opsize == TCPOLEN_TIMESTAMP) && | |
3800 | ((estab && opt_rx->tstamp_ok) || | |
3801 | (!estab && sysctl_tcp_timestamps))) { | |
3802 | opt_rx->saw_tstamp = 1; | |
3803 | opt_rx->rcv_tsval = get_unaligned_be32(ptr); | |
3804 | opt_rx->rcv_tsecr = get_unaligned_be32(ptr + 4); | |
3805 | } | |
3806 | break; | |
3807 | case TCPOPT_SACK_PERM: | |
3808 | if (opsize == TCPOLEN_SACK_PERM && th->syn && | |
3809 | !estab && sysctl_tcp_sack) { | |
3810 | opt_rx->sack_ok = 1; | |
3811 | tcp_sack_reset(opt_rx); | |
3812 | } | |
3813 | break; | |
3814 | ||
3815 | case TCPOPT_SACK: | |
3816 | if ((opsize >= (TCPOLEN_SACK_BASE + TCPOLEN_SACK_PERBLOCK)) && | |
3817 | !((opsize - TCPOLEN_SACK_BASE) % TCPOLEN_SACK_PERBLOCK) && | |
3818 | opt_rx->sack_ok) { | |
3819 | TCP_SKB_CB(skb)->sacked = (ptr - 2) - (unsigned char *)th; | |
3820 | } | |
3821 | break; | |
3822 | #ifdef CONFIG_TCP_MD5SIG | |
3823 | case TCPOPT_MD5SIG: | |
3824 | /* | |
3825 | * The MD5 Hash has already been | |
3826 | * checked (see tcp_v{4,6}_do_rcv()). | |
3827 | */ | |
3828 | break; | |
3829 | #endif | |
3830 | case TCPOPT_COOKIE: | |
3831 | /* This option is variable length. | |
3832 | */ | |
3833 | switch (opsize) { | |
3834 | case TCPOLEN_COOKIE_BASE: | |
3835 | /* not yet implemented */ | |
3836 | break; | |
3837 | case TCPOLEN_COOKIE_PAIR: | |
3838 | /* not yet implemented */ | |
3839 | break; | |
3840 | case TCPOLEN_COOKIE_MIN+0: | |
3841 | case TCPOLEN_COOKIE_MIN+2: | |
3842 | case TCPOLEN_COOKIE_MIN+4: | |
3843 | case TCPOLEN_COOKIE_MIN+6: | |
3844 | case TCPOLEN_COOKIE_MAX: | |
3845 | /* 16-bit multiple */ | |
3846 | opt_rx->cookie_plus = opsize; | |
3847 | *hvpp = ptr; | |
3848 | break; | |
3849 | default: | |
3850 | /* ignore option */ | |
3851 | break; | |
3852 | } | |
3853 | break; | |
3854 | } | |
3855 | ||
3856 | ptr += opsize-2; | |
3857 | length -= opsize; | |
3858 | } | |
3859 | } | |
3860 | } | |
3861 | EXPORT_SYMBOL(tcp_parse_options); | |
3862 | ||
3863 | static int tcp_parse_aligned_timestamp(struct tcp_sock *tp, struct tcphdr *th) | |
3864 | { | |
3865 | __be32 *ptr = (__be32 *)(th + 1); | |
3866 | ||
3867 | if (*ptr == htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | |
3868 | | (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP)) { | |
3869 | tp->rx_opt.saw_tstamp = 1; | |
3870 | ++ptr; | |
3871 | tp->rx_opt.rcv_tsval = ntohl(*ptr); | |
3872 | ++ptr; | |
3873 | tp->rx_opt.rcv_tsecr = ntohl(*ptr); | |
3874 | return 1; | |
3875 | } | |
3876 | return 0; | |
3877 | } | |
3878 | ||
3879 | /* Fast parse options. This hopes to only see timestamps. | |
3880 | * If it is wrong it falls back on tcp_parse_options(). | |
3881 | */ | |
3882 | static int tcp_fast_parse_options(struct sk_buff *skb, struct tcphdr *th, | |
3883 | struct tcp_sock *tp, u8 **hvpp) | |
3884 | { | |
3885 | /* In the spirit of fast parsing, compare doff directly to constant | |
3886 | * values. Because equality is used, short doff can be ignored here. | |
3887 | */ | |
3888 | if (th->doff == (sizeof(*th) / 4)) { | |
3889 | tp->rx_opt.saw_tstamp = 0; | |
3890 | return 0; | |
3891 | } else if (tp->rx_opt.tstamp_ok && | |
3892 | th->doff == ((sizeof(*th) + TCPOLEN_TSTAMP_ALIGNED) / 4)) { | |
3893 | if (tcp_parse_aligned_timestamp(tp, th)) | |
3894 | return 1; | |
3895 | } | |
3896 | tcp_parse_options(skb, &tp->rx_opt, hvpp, 1); | |
3897 | return 1; | |
3898 | } | |
3899 | ||
3900 | #ifdef CONFIG_TCP_MD5SIG | |
3901 | /* | |
3902 | * Parse MD5 Signature option | |
3903 | */ | |
3904 | u8 *tcp_parse_md5sig_option(struct tcphdr *th) | |
3905 | { | |
3906 | int length = (th->doff << 2) - sizeof (*th); | |
3907 | u8 *ptr = (u8*)(th + 1); | |
3908 | ||
3909 | /* If the TCP option is too short, we can short cut */ | |
3910 | if (length < TCPOLEN_MD5SIG) | |
3911 | return NULL; | |
3912 | ||
3913 | while (length > 0) { | |
3914 | int opcode = *ptr++; | |
3915 | int opsize; | |
3916 | ||
3917 | switch(opcode) { | |
3918 | case TCPOPT_EOL: | |
3919 | return NULL; | |
3920 | case TCPOPT_NOP: | |
3921 | length--; | |
3922 | continue; | |
3923 | default: | |
3924 | opsize = *ptr++; | |
3925 | if (opsize < 2 || opsize > length) | |
3926 | return NULL; | |
3927 | if (opcode == TCPOPT_MD5SIG) | |
3928 | return opsize == TCPOLEN_MD5SIG ? ptr : NULL; | |
3929 | } | |
3930 | ptr += opsize - 2; | |
3931 | length -= opsize; | |
3932 | } | |
3933 | return NULL; | |
3934 | } | |
3935 | EXPORT_SYMBOL(tcp_parse_md5sig_option); | |
3936 | #endif | |
3937 | ||
3938 | static inline void tcp_store_ts_recent(struct tcp_sock *tp) | |
3939 | { | |
3940 | tp->rx_opt.ts_recent = tp->rx_opt.rcv_tsval; | |
3941 | tp->rx_opt.ts_recent_stamp = get_seconds(); | |
3942 | } | |
3943 | ||
3944 | static inline void tcp_replace_ts_recent(struct tcp_sock *tp, u32 seq) | |
3945 | { | |
3946 | if (tp->rx_opt.saw_tstamp && !after(seq, tp->rcv_wup)) { | |
3947 | /* PAWS bug workaround wrt. ACK frames, the PAWS discard | |
3948 | * extra check below makes sure this can only happen | |
3949 | * for pure ACK frames. -DaveM | |
3950 | * | |
3951 | * Not only, also it occurs for expired timestamps. | |
3952 | */ | |
3953 | ||
3954 | if (tcp_paws_check(&tp->rx_opt, 0)) | |
3955 | tcp_store_ts_recent(tp); | |
3956 | } | |
3957 | } | |
3958 | ||
3959 | /* Sorry, PAWS as specified is broken wrt. pure-ACKs -DaveM | |
3960 | * | |
3961 | * It is not fatal. If this ACK does _not_ change critical state (seqs, window) | |
3962 | * it can pass through stack. So, the following predicate verifies that | |
3963 | * this segment is not used for anything but congestion avoidance or | |
3964 | * fast retransmit. Moreover, we even are able to eliminate most of such | |
3965 | * second order effects, if we apply some small "replay" window (~RTO) | |
3966 | * to timestamp space. | |
3967 | * | |
3968 | * All these measures still do not guarantee that we reject wrapped ACKs | |
3969 | * on networks with high bandwidth, when sequence space is recycled fastly, | |
3970 | * but it guarantees that such events will be very rare and do not affect | |
3971 | * connection seriously. This doesn't look nice, but alas, PAWS is really | |
3972 | * buggy extension. | |
3973 | * | |
3974 | * [ Later note. Even worse! It is buggy for segments _with_ data. RFC | |
3975 | * states that events when retransmit arrives after original data are rare. | |
3976 | * It is a blatant lie. VJ forgot about fast retransmit! 8)8) It is | |
3977 | * the biggest problem on large power networks even with minor reordering. | |
3978 | * OK, let's give it small replay window. If peer clock is even 1hz, it is safe | |
3979 | * up to bandwidth of 18Gigabit/sec. 8) ] | |
3980 | */ | |
3981 | ||
3982 | static int tcp_disordered_ack(const struct sock *sk, const struct sk_buff *skb) | |
3983 | { | |
3984 | struct tcp_sock *tp = tcp_sk(sk); | |
3985 | struct tcphdr *th = tcp_hdr(skb); | |
3986 | u32 seq = TCP_SKB_CB(skb)->seq; | |
3987 | u32 ack = TCP_SKB_CB(skb)->ack_seq; | |
3988 | ||
3989 | return (/* 1. Pure ACK with correct sequence number. */ | |
3990 | (th->ack && seq == TCP_SKB_CB(skb)->end_seq && seq == tp->rcv_nxt) && | |
3991 | ||
3992 | /* 2. ... and duplicate ACK. */ | |
3993 | ack == tp->snd_una && | |
3994 | ||
3995 | /* 3. ... and does not update window. */ | |
3996 | !tcp_may_update_window(tp, ack, seq, ntohs(th->window) << tp->rx_opt.snd_wscale) && | |
3997 | ||
3998 | /* 4. ... and sits in replay window. */ | |
3999 | (s32)(tp->rx_opt.ts_recent - tp->rx_opt.rcv_tsval) <= (inet_csk(sk)->icsk_rto * 1024) / HZ); | |
4000 | } | |
4001 | ||
4002 | static inline int tcp_paws_discard(const struct sock *sk, | |
4003 | const struct sk_buff *skb) | |
4004 | { | |
4005 | const struct tcp_sock *tp = tcp_sk(sk); | |
4006 | ||
4007 | return !tcp_paws_check(&tp->rx_opt, TCP_PAWS_WINDOW) && | |
4008 | !tcp_disordered_ack(sk, skb); | |
4009 | } | |
4010 | ||
4011 | /* Check segment sequence number for validity. | |
4012 | * | |
4013 | * Segment controls are considered valid, if the segment | |
4014 | * fits to the window after truncation to the window. Acceptability | |
4015 | * of data (and SYN, FIN, of course) is checked separately. | |
4016 | * See tcp_data_queue(), for example. | |
4017 | * | |
4018 | * Also, controls (RST is main one) are accepted using RCV.WUP instead | |
4019 | * of RCV.NXT. Peer still did not advance his SND.UNA when we | |
4020 | * delayed ACK, so that hisSND.UNA<=ourRCV.WUP. | |
4021 | * (borrowed from freebsd) | |
4022 | */ | |
4023 | ||
4024 | static inline int tcp_sequence(struct tcp_sock *tp, u32 seq, u32 end_seq) | |
4025 | { | |
4026 | return !before(end_seq, tp->rcv_wup) && | |
4027 | !after(seq, tp->rcv_nxt + tcp_receive_window(tp)); | |
4028 | } | |
4029 | ||
4030 | /* When we get a reset we do this. */ | |
4031 | static void tcp_reset(struct sock *sk) | |
4032 | { | |
4033 | /* We want the right error as BSD sees it (and indeed as we do). */ | |
4034 | switch (sk->sk_state) { | |
4035 | case TCP_SYN_SENT: | |
4036 | sk->sk_err = ECONNREFUSED; | |
4037 | break; | |
4038 | case TCP_CLOSE_WAIT: | |
4039 | sk->sk_err = EPIPE; | |
4040 | break; | |
4041 | case TCP_CLOSE: | |
4042 | return; | |
4043 | default: | |
4044 | sk->sk_err = ECONNRESET; | |
4045 | } | |
4046 | /* This barrier is coupled with smp_rmb() in tcp_poll() */ | |
4047 | smp_wmb(); | |
4048 | ||
4049 | if (!sock_flag(sk, SOCK_DEAD)) | |
4050 | sk->sk_error_report(sk); | |
4051 | ||
4052 | tcp_done(sk); | |
4053 | } | |
4054 | ||
4055 | /* | |
4056 | * Process the FIN bit. This now behaves as it is supposed to work | |
4057 | * and the FIN takes effect when it is validly part of sequence | |
4058 | * space. Not before when we get holes. | |
4059 | * | |
4060 | * If we are ESTABLISHED, a received fin moves us to CLOSE-WAIT | |
4061 | * (and thence onto LAST-ACK and finally, CLOSE, we never enter | |
4062 | * TIME-WAIT) | |
4063 | * | |
4064 | * If we are in FINWAIT-1, a received FIN indicates simultaneous | |
4065 | * close and we go into CLOSING (and later onto TIME-WAIT) | |
4066 | * | |
4067 | * If we are in FINWAIT-2, a received FIN moves us to TIME-WAIT. | |
4068 | */ | |
4069 | static void tcp_fin(struct sk_buff *skb, struct sock *sk, struct tcphdr *th) | |
4070 | { | |
4071 | struct tcp_sock *tp = tcp_sk(sk); | |
4072 | ||
4073 | inet_csk_schedule_ack(sk); | |
4074 | ||
4075 | sk->sk_shutdown |= RCV_SHUTDOWN; | |
4076 | sock_set_flag(sk, SOCK_DONE); | |
4077 | ||
4078 | switch (sk->sk_state) { | |
4079 | case TCP_SYN_RECV: | |
4080 | case TCP_ESTABLISHED: | |
4081 | /* Move to CLOSE_WAIT */ | |
4082 | tcp_set_state(sk, TCP_CLOSE_WAIT); | |
4083 | inet_csk(sk)->icsk_ack.pingpong = 1; | |
4084 | break; | |
4085 | ||
4086 | case TCP_CLOSE_WAIT: | |
4087 | case TCP_CLOSING: | |
4088 | /* Received a retransmission of the FIN, do | |
4089 | * nothing. | |
4090 | */ | |
4091 | break; | |
4092 | case TCP_LAST_ACK: | |
4093 | /* RFC793: Remain in the LAST-ACK state. */ | |
4094 | break; | |
4095 | ||
4096 | case TCP_FIN_WAIT1: | |
4097 | /* This case occurs when a simultaneous close | |
4098 | * happens, we must ack the received FIN and | |
4099 | * enter the CLOSING state. | |
4100 | */ | |
4101 | tcp_send_ack(sk); | |
4102 | tcp_set_state(sk, TCP_CLOSING); | |
4103 | break; | |
4104 | case TCP_FIN_WAIT2: | |
4105 | /* Received a FIN -- send ACK and enter TIME_WAIT. */ | |
4106 | tcp_send_ack(sk); | |
4107 | tcp_time_wait(sk, TCP_TIME_WAIT, 0); | |
4108 | break; | |
4109 | default: | |
4110 | /* Only TCP_LISTEN and TCP_CLOSE are left, in these | |
4111 | * cases we should never reach this piece of code. | |
4112 | */ | |
4113 | printk(KERN_ERR "%s: Impossible, sk->sk_state=%d\n", | |
4114 | __func__, sk->sk_state); | |
4115 | break; | |
4116 | } | |
4117 | ||
4118 | /* It _is_ possible, that we have something out-of-order _after_ FIN. | |
4119 | * Probably, we should reset in this case. For now drop them. | |
4120 | */ | |
4121 | __skb_queue_purge(&tp->out_of_order_queue); | |
4122 | if (tcp_is_sack(tp)) | |
4123 | tcp_sack_reset(&tp->rx_opt); | |
4124 | sk_mem_reclaim(sk); | |
4125 | ||
4126 | if (!sock_flag(sk, SOCK_DEAD)) { | |
4127 | sk->sk_state_change(sk); | |
4128 | ||
4129 | /* Do not send POLL_HUP for half duplex close. */ | |
4130 | if (sk->sk_shutdown == SHUTDOWN_MASK || | |
4131 | sk->sk_state == TCP_CLOSE) | |
4132 | sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP); | |
4133 | else | |
4134 | sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN); | |
4135 | } | |
4136 | } | |
4137 | ||
4138 | static inline int tcp_sack_extend(struct tcp_sack_block *sp, u32 seq, | |
4139 | u32 end_seq) | |
4140 | { | |
4141 | if (!after(seq, sp->end_seq) && !after(sp->start_seq, end_seq)) { | |
4142 | if (before(seq, sp->start_seq)) | |
4143 | sp->start_seq = seq; | |
4144 | if (after(end_seq, sp->end_seq)) | |
4145 | sp->end_seq = end_seq; | |
4146 | return 1; | |
4147 | } | |
4148 | return 0; | |
4149 | } | |
4150 | ||
4151 | static void tcp_dsack_set(struct sock *sk, u32 seq, u32 end_seq) | |
4152 | { | |
4153 | struct tcp_sock *tp = tcp_sk(sk); | |
4154 | ||
4155 | if (tcp_is_sack(tp) && sysctl_tcp_dsack) { | |
4156 | int mib_idx; | |
4157 | ||
4158 | if (before(seq, tp->rcv_nxt)) | |
4159 | mib_idx = LINUX_MIB_TCPDSACKOLDSENT; | |
4160 | else | |
4161 | mib_idx = LINUX_MIB_TCPDSACKOFOSENT; | |
4162 | ||
4163 | NET_INC_STATS_BH(sock_net(sk), mib_idx); | |
4164 | ||
4165 | tp->rx_opt.dsack = 1; | |
4166 | tp->duplicate_sack[0].start_seq = seq; | |
4167 | tp->duplicate_sack[0].end_seq = end_seq; | |
4168 | } | |
4169 | } | |
4170 | ||
4171 | static void tcp_dsack_extend(struct sock *sk, u32 seq, u32 end_seq) | |
4172 | { | |
4173 | struct tcp_sock *tp = tcp_sk(sk); | |
4174 | ||
4175 | if (!tp->rx_opt.dsack) | |
4176 | tcp_dsack_set(sk, seq, end_seq); | |
4177 | else | |
4178 | tcp_sack_extend(tp->duplicate_sack, seq, end_seq); | |
4179 | } | |
4180 | ||
4181 | static void tcp_send_dupack(struct sock *sk, struct sk_buff *skb) | |
4182 | { | |
4183 | struct tcp_sock *tp = tcp_sk(sk); | |
4184 | ||
4185 | if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && | |
4186 | before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
4187 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); | |
4188 | tcp_enter_quickack_mode(sk); | |
4189 | ||
4190 | if (tcp_is_sack(tp) && sysctl_tcp_dsack) { | |
4191 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
4192 | ||
4193 | if (after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) | |
4194 | end_seq = tp->rcv_nxt; | |
4195 | tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, end_seq); | |
4196 | } | |
4197 | } | |
4198 | ||
4199 | tcp_send_ack(sk); | |
4200 | } | |
4201 | ||
4202 | /* These routines update the SACK block as out-of-order packets arrive or | |
4203 | * in-order packets close up the sequence space. | |
4204 | */ | |
4205 | static void tcp_sack_maybe_coalesce(struct tcp_sock *tp) | |
4206 | { | |
4207 | int this_sack; | |
4208 | struct tcp_sack_block *sp = &tp->selective_acks[0]; | |
4209 | struct tcp_sack_block *swalk = sp + 1; | |
4210 | ||
4211 | /* See if the recent change to the first SACK eats into | |
4212 | * or hits the sequence space of other SACK blocks, if so coalesce. | |
4213 | */ | |
4214 | for (this_sack = 1; this_sack < tp->rx_opt.num_sacks;) { | |
4215 | if (tcp_sack_extend(sp, swalk->start_seq, swalk->end_seq)) { | |
4216 | int i; | |
4217 | ||
4218 | /* Zap SWALK, by moving every further SACK up by one slot. | |
4219 | * Decrease num_sacks. | |
4220 | */ | |
4221 | tp->rx_opt.num_sacks--; | |
4222 | for (i = this_sack; i < tp->rx_opt.num_sacks; i++) | |
4223 | sp[i] = sp[i + 1]; | |
4224 | continue; | |
4225 | } | |
4226 | this_sack++, swalk++; | |
4227 | } | |
4228 | } | |
4229 | ||
4230 | static void tcp_sack_new_ofo_skb(struct sock *sk, u32 seq, u32 end_seq) | |
4231 | { | |
4232 | struct tcp_sock *tp = tcp_sk(sk); | |
4233 | struct tcp_sack_block *sp = &tp->selective_acks[0]; | |
4234 | int cur_sacks = tp->rx_opt.num_sacks; | |
4235 | int this_sack; | |
4236 | ||
4237 | if (!cur_sacks) | |
4238 | goto new_sack; | |
4239 | ||
4240 | for (this_sack = 0; this_sack < cur_sacks; this_sack++, sp++) { | |
4241 | if (tcp_sack_extend(sp, seq, end_seq)) { | |
4242 | /* Rotate this_sack to the first one. */ | |
4243 | for (; this_sack > 0; this_sack--, sp--) | |
4244 | swap(*sp, *(sp - 1)); | |
4245 | if (cur_sacks > 1) | |
4246 | tcp_sack_maybe_coalesce(tp); | |
4247 | return; | |
4248 | } | |
4249 | } | |
4250 | ||
4251 | /* Could not find an adjacent existing SACK, build a new one, | |
4252 | * put it at the front, and shift everyone else down. We | |
4253 | * always know there is at least one SACK present already here. | |
4254 | * | |
4255 | * If the sack array is full, forget about the last one. | |
4256 | */ | |
4257 | if (this_sack >= TCP_NUM_SACKS) { | |
4258 | this_sack--; | |
4259 | tp->rx_opt.num_sacks--; | |
4260 | sp--; | |
4261 | } | |
4262 | for (; this_sack > 0; this_sack--, sp--) | |
4263 | *sp = *(sp - 1); | |
4264 | ||
4265 | new_sack: | |
4266 | /* Build the new head SACK, and we're done. */ | |
4267 | sp->start_seq = seq; | |
4268 | sp->end_seq = end_seq; | |
4269 | tp->rx_opt.num_sacks++; | |
4270 | } | |
4271 | ||
4272 | /* RCV.NXT advances, some SACKs should be eaten. */ | |
4273 | ||
4274 | static void tcp_sack_remove(struct tcp_sock *tp) | |
4275 | { | |
4276 | struct tcp_sack_block *sp = &tp->selective_acks[0]; | |
4277 | int num_sacks = tp->rx_opt.num_sacks; | |
4278 | int this_sack; | |
4279 | ||
4280 | /* Empty ofo queue, hence, all the SACKs are eaten. Clear. */ | |
4281 | if (skb_queue_empty(&tp->out_of_order_queue)) { | |
4282 | tp->rx_opt.num_sacks = 0; | |
4283 | return; | |
4284 | } | |
4285 | ||
4286 | for (this_sack = 0; this_sack < num_sacks;) { | |
4287 | /* Check if the start of the sack is covered by RCV.NXT. */ | |
4288 | if (!before(tp->rcv_nxt, sp->start_seq)) { | |
4289 | int i; | |
4290 | ||
4291 | /* RCV.NXT must cover all the block! */ | |
4292 | WARN_ON(before(tp->rcv_nxt, sp->end_seq)); | |
4293 | ||
4294 | /* Zap this SACK, by moving forward any other SACKS. */ | |
4295 | for (i=this_sack+1; i < num_sacks; i++) | |
4296 | tp->selective_acks[i-1] = tp->selective_acks[i]; | |
4297 | num_sacks--; | |
4298 | continue; | |
4299 | } | |
4300 | this_sack++; | |
4301 | sp++; | |
4302 | } | |
4303 | tp->rx_opt.num_sacks = num_sacks; | |
4304 | } | |
4305 | ||
4306 | /* This one checks to see if we can put data from the | |
4307 | * out_of_order queue into the receive_queue. | |
4308 | */ | |
4309 | static void tcp_ofo_queue(struct sock *sk) | |
4310 | { | |
4311 | struct tcp_sock *tp = tcp_sk(sk); | |
4312 | __u32 dsack_high = tp->rcv_nxt; | |
4313 | struct sk_buff *skb; | |
4314 | ||
4315 | while ((skb = skb_peek(&tp->out_of_order_queue)) != NULL) { | |
4316 | if (after(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) | |
4317 | break; | |
4318 | ||
4319 | if (before(TCP_SKB_CB(skb)->seq, dsack_high)) { | |
4320 | __u32 dsack = dsack_high; | |
4321 | if (before(TCP_SKB_CB(skb)->end_seq, dsack_high)) | |
4322 | dsack_high = TCP_SKB_CB(skb)->end_seq; | |
4323 | tcp_dsack_extend(sk, TCP_SKB_CB(skb)->seq, dsack); | |
4324 | } | |
4325 | ||
4326 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { | |
4327 | SOCK_DEBUG(sk, "ofo packet was already received\n"); | |
4328 | __skb_unlink(skb, &tp->out_of_order_queue); | |
4329 | __kfree_skb(skb); | |
4330 | continue; | |
4331 | } | |
4332 | SOCK_DEBUG(sk, "ofo requeuing : rcv_next %X seq %X - %X\n", | |
4333 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, | |
4334 | TCP_SKB_CB(skb)->end_seq); | |
4335 | ||
4336 | __skb_unlink(skb, &tp->out_of_order_queue); | |
4337 | __skb_queue_tail(&sk->sk_receive_queue, skb); | |
4338 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
4339 | if (tcp_hdr(skb)->fin) | |
4340 | tcp_fin(skb, sk, tcp_hdr(skb)); | |
4341 | } | |
4342 | } | |
4343 | ||
4344 | static int tcp_prune_ofo_queue(struct sock *sk); | |
4345 | static int tcp_prune_queue(struct sock *sk); | |
4346 | ||
4347 | static inline int tcp_try_rmem_schedule(struct sock *sk, unsigned int size) | |
4348 | { | |
4349 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf || | |
4350 | !sk_rmem_schedule(sk, size)) { | |
4351 | ||
4352 | if (tcp_prune_queue(sk) < 0) | |
4353 | return -1; | |
4354 | ||
4355 | if (!sk_rmem_schedule(sk, size)) { | |
4356 | if (!tcp_prune_ofo_queue(sk)) | |
4357 | return -1; | |
4358 | ||
4359 | if (!sk_rmem_schedule(sk, size)) | |
4360 | return -1; | |
4361 | } | |
4362 | } | |
4363 | return 0; | |
4364 | } | |
4365 | ||
4366 | static void tcp_data_queue(struct sock *sk, struct sk_buff *skb) | |
4367 | { | |
4368 | struct tcphdr *th = tcp_hdr(skb); | |
4369 | struct tcp_sock *tp = tcp_sk(sk); | |
4370 | int eaten = -1; | |
4371 | ||
4372 | if (TCP_SKB_CB(skb)->seq == TCP_SKB_CB(skb)->end_seq) | |
4373 | goto drop; | |
4374 | ||
4375 | skb_dst_drop(skb); | |
4376 | __skb_pull(skb, th->doff * 4); | |
4377 | ||
4378 | TCP_ECN_accept_cwr(tp, skb); | |
4379 | ||
4380 | tp->rx_opt.dsack = 0; | |
4381 | ||
4382 | /* Queue data for delivery to the user. | |
4383 | * Packets in sequence go to the receive queue. | |
4384 | * Out of sequence packets to the out_of_order_queue. | |
4385 | */ | |
4386 | if (TCP_SKB_CB(skb)->seq == tp->rcv_nxt) { | |
4387 | if (tcp_receive_window(tp) == 0) | |
4388 | goto out_of_window; | |
4389 | ||
4390 | /* Ok. In sequence. In window. */ | |
4391 | if (tp->ucopy.task == current && | |
4392 | tp->copied_seq == tp->rcv_nxt && tp->ucopy.len && | |
4393 | sock_owned_by_user(sk) && !tp->urg_data) { | |
4394 | int chunk = min_t(unsigned int, skb->len, | |
4395 | tp->ucopy.len); | |
4396 | ||
4397 | __set_current_state(TASK_RUNNING); | |
4398 | ||
4399 | local_bh_enable(); | |
4400 | if (!skb_copy_datagram_iovec(skb, 0, tp->ucopy.iov, chunk)) { | |
4401 | tp->ucopy.len -= chunk; | |
4402 | tp->copied_seq += chunk; | |
4403 | eaten = (chunk == skb->len && !th->fin); | |
4404 | tcp_rcv_space_adjust(sk); | |
4405 | } | |
4406 | local_bh_disable(); | |
4407 | } | |
4408 | ||
4409 | if (eaten <= 0) { | |
4410 | queue_and_out: | |
4411 | if (eaten < 0 && | |
4412 | tcp_try_rmem_schedule(sk, skb->truesize)) | |
4413 | goto drop; | |
4414 | ||
4415 | skb_set_owner_r(skb, sk); | |
4416 | __skb_queue_tail(&sk->sk_receive_queue, skb); | |
4417 | } | |
4418 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
4419 | if (skb->len) | |
4420 | tcp_event_data_recv(sk, skb); | |
4421 | if (th->fin) | |
4422 | tcp_fin(skb, sk, th); | |
4423 | ||
4424 | if (!skb_queue_empty(&tp->out_of_order_queue)) { | |
4425 | tcp_ofo_queue(sk); | |
4426 | ||
4427 | /* RFC2581. 4.2. SHOULD send immediate ACK, when | |
4428 | * gap in queue is filled. | |
4429 | */ | |
4430 | if (skb_queue_empty(&tp->out_of_order_queue)) | |
4431 | inet_csk(sk)->icsk_ack.pingpong = 0; | |
4432 | } | |
4433 | ||
4434 | if (tp->rx_opt.num_sacks) | |
4435 | tcp_sack_remove(tp); | |
4436 | ||
4437 | tcp_fast_path_check(sk); | |
4438 | ||
4439 | if (eaten > 0) | |
4440 | __kfree_skb(skb); | |
4441 | else if (!sock_flag(sk, SOCK_DEAD)) | |
4442 | sk->sk_data_ready(sk, 0); | |
4443 | return; | |
4444 | } | |
4445 | ||
4446 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->rcv_nxt)) { | |
4447 | /* A retransmit, 2nd most common case. Force an immediate ack. */ | |
4448 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_DELAYEDACKLOST); | |
4449 | tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); | |
4450 | ||
4451 | out_of_window: | |
4452 | tcp_enter_quickack_mode(sk); | |
4453 | inet_csk_schedule_ack(sk); | |
4454 | drop: | |
4455 | __kfree_skb(skb); | |
4456 | return; | |
4457 | } | |
4458 | ||
4459 | /* Out of window. F.e. zero window probe. */ | |
4460 | if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt + tcp_receive_window(tp))) | |
4461 | goto out_of_window; | |
4462 | ||
4463 | tcp_enter_quickack_mode(sk); | |
4464 | ||
4465 | if (before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
4466 | /* Partial packet, seq < rcv_next < end_seq */ | |
4467 | SOCK_DEBUG(sk, "partial packet: rcv_next %X seq %X - %X\n", | |
4468 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, | |
4469 | TCP_SKB_CB(skb)->end_seq); | |
4470 | ||
4471 | tcp_dsack_set(sk, TCP_SKB_CB(skb)->seq, tp->rcv_nxt); | |
4472 | ||
4473 | /* If window is closed, drop tail of packet. But after | |
4474 | * remembering D-SACK for its head made in previous line. | |
4475 | */ | |
4476 | if (!tcp_receive_window(tp)) | |
4477 | goto out_of_window; | |
4478 | goto queue_and_out; | |
4479 | } | |
4480 | ||
4481 | TCP_ECN_check_ce(tp, skb); | |
4482 | ||
4483 | if (tcp_try_rmem_schedule(sk, skb->truesize)) | |
4484 | goto drop; | |
4485 | ||
4486 | /* Disable header prediction. */ | |
4487 | tp->pred_flags = 0; | |
4488 | inet_csk_schedule_ack(sk); | |
4489 | ||
4490 | SOCK_DEBUG(sk, "out of order segment: rcv_next %X seq %X - %X\n", | |
4491 | tp->rcv_nxt, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq); | |
4492 | ||
4493 | skb_set_owner_r(skb, sk); | |
4494 | ||
4495 | if (!skb_peek(&tp->out_of_order_queue)) { | |
4496 | /* Initial out of order segment, build 1 SACK. */ | |
4497 | if (tcp_is_sack(tp)) { | |
4498 | tp->rx_opt.num_sacks = 1; | |
4499 | tp->selective_acks[0].start_seq = TCP_SKB_CB(skb)->seq; | |
4500 | tp->selective_acks[0].end_seq = | |
4501 | TCP_SKB_CB(skb)->end_seq; | |
4502 | } | |
4503 | __skb_queue_head(&tp->out_of_order_queue, skb); | |
4504 | } else { | |
4505 | struct sk_buff *skb1 = skb_peek_tail(&tp->out_of_order_queue); | |
4506 | u32 seq = TCP_SKB_CB(skb)->seq; | |
4507 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
4508 | ||
4509 | if (seq == TCP_SKB_CB(skb1)->end_seq) { | |
4510 | __skb_queue_after(&tp->out_of_order_queue, skb1, skb); | |
4511 | ||
4512 | if (!tp->rx_opt.num_sacks || | |
4513 | tp->selective_acks[0].end_seq != seq) | |
4514 | goto add_sack; | |
4515 | ||
4516 | /* Common case: data arrive in order after hole. */ | |
4517 | tp->selective_acks[0].end_seq = end_seq; | |
4518 | return; | |
4519 | } | |
4520 | ||
4521 | /* Find place to insert this segment. */ | |
4522 | while (1) { | |
4523 | if (!after(TCP_SKB_CB(skb1)->seq, seq)) | |
4524 | break; | |
4525 | if (skb_queue_is_first(&tp->out_of_order_queue, skb1)) { | |
4526 | skb1 = NULL; | |
4527 | break; | |
4528 | } | |
4529 | skb1 = skb_queue_prev(&tp->out_of_order_queue, skb1); | |
4530 | } | |
4531 | ||
4532 | /* Do skb overlap to previous one? */ | |
4533 | if (skb1 && before(seq, TCP_SKB_CB(skb1)->end_seq)) { | |
4534 | if (!after(end_seq, TCP_SKB_CB(skb1)->end_seq)) { | |
4535 | /* All the bits are present. Drop. */ | |
4536 | __kfree_skb(skb); | |
4537 | tcp_dsack_set(sk, seq, end_seq); | |
4538 | goto add_sack; | |
4539 | } | |
4540 | if (after(seq, TCP_SKB_CB(skb1)->seq)) { | |
4541 | /* Partial overlap. */ | |
4542 | tcp_dsack_set(sk, seq, | |
4543 | TCP_SKB_CB(skb1)->end_seq); | |
4544 | } else { | |
4545 | if (skb_queue_is_first(&tp->out_of_order_queue, | |
4546 | skb1)) | |
4547 | skb1 = NULL; | |
4548 | else | |
4549 | skb1 = skb_queue_prev( | |
4550 | &tp->out_of_order_queue, | |
4551 | skb1); | |
4552 | } | |
4553 | } | |
4554 | if (!skb1) | |
4555 | __skb_queue_head(&tp->out_of_order_queue, skb); | |
4556 | else | |
4557 | __skb_queue_after(&tp->out_of_order_queue, skb1, skb); | |
4558 | ||
4559 | /* And clean segments covered by new one as whole. */ | |
4560 | while (!skb_queue_is_last(&tp->out_of_order_queue, skb)) { | |
4561 | skb1 = skb_queue_next(&tp->out_of_order_queue, skb); | |
4562 | ||
4563 | if (!after(end_seq, TCP_SKB_CB(skb1)->seq)) | |
4564 | break; | |
4565 | if (before(end_seq, TCP_SKB_CB(skb1)->end_seq)) { | |
4566 | tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, | |
4567 | end_seq); | |
4568 | break; | |
4569 | } | |
4570 | __skb_unlink(skb1, &tp->out_of_order_queue); | |
4571 | tcp_dsack_extend(sk, TCP_SKB_CB(skb1)->seq, | |
4572 | TCP_SKB_CB(skb1)->end_seq); | |
4573 | __kfree_skb(skb1); | |
4574 | } | |
4575 | ||
4576 | add_sack: | |
4577 | if (tcp_is_sack(tp)) | |
4578 | tcp_sack_new_ofo_skb(sk, seq, end_seq); | |
4579 | } | |
4580 | } | |
4581 | ||
4582 | static struct sk_buff *tcp_collapse_one(struct sock *sk, struct sk_buff *skb, | |
4583 | struct sk_buff_head *list) | |
4584 | { | |
4585 | struct sk_buff *next = NULL; | |
4586 | ||
4587 | if (!skb_queue_is_last(list, skb)) | |
4588 | next = skb_queue_next(list, skb); | |
4589 | ||
4590 | __skb_unlink(skb, list); | |
4591 | __kfree_skb(skb); | |
4592 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRCVCOLLAPSED); | |
4593 | ||
4594 | return next; | |
4595 | } | |
4596 | ||
4597 | /* Collapse contiguous sequence of skbs head..tail with | |
4598 | * sequence numbers start..end. | |
4599 | * | |
4600 | * If tail is NULL, this means until the end of the list. | |
4601 | * | |
4602 | * Segments with FIN/SYN are not collapsed (only because this | |
4603 | * simplifies code) | |
4604 | */ | |
4605 | static void | |
4606 | tcp_collapse(struct sock *sk, struct sk_buff_head *list, | |
4607 | struct sk_buff *head, struct sk_buff *tail, | |
4608 | u32 start, u32 end) | |
4609 | { | |
4610 | struct sk_buff *skb, *n; | |
4611 | bool end_of_skbs; | |
4612 | ||
4613 | /* First, check that queue is collapsible and find | |
4614 | * the point where collapsing can be useful. */ | |
4615 | skb = head; | |
4616 | restart: | |
4617 | end_of_skbs = true; | |
4618 | skb_queue_walk_from_safe(list, skb, n) { | |
4619 | if (skb == tail) | |
4620 | break; | |
4621 | /* No new bits? It is possible on ofo queue. */ | |
4622 | if (!before(start, TCP_SKB_CB(skb)->end_seq)) { | |
4623 | skb = tcp_collapse_one(sk, skb, list); | |
4624 | if (!skb) | |
4625 | break; | |
4626 | goto restart; | |
4627 | } | |
4628 | ||
4629 | /* The first skb to collapse is: | |
4630 | * - not SYN/FIN and | |
4631 | * - bloated or contains data before "start" or | |
4632 | * overlaps to the next one. | |
4633 | */ | |
4634 | if (!tcp_hdr(skb)->syn && !tcp_hdr(skb)->fin && | |
4635 | (tcp_win_from_space(skb->truesize) > skb->len || | |
4636 | before(TCP_SKB_CB(skb)->seq, start))) { | |
4637 | end_of_skbs = false; | |
4638 | break; | |
4639 | } | |
4640 | ||
4641 | if (!skb_queue_is_last(list, skb)) { | |
4642 | struct sk_buff *next = skb_queue_next(list, skb); | |
4643 | if (next != tail && | |
4644 | TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(next)->seq) { | |
4645 | end_of_skbs = false; | |
4646 | break; | |
4647 | } | |
4648 | } | |
4649 | ||
4650 | /* Decided to skip this, advance start seq. */ | |
4651 | start = TCP_SKB_CB(skb)->end_seq; | |
4652 | } | |
4653 | if (end_of_skbs || tcp_hdr(skb)->syn || tcp_hdr(skb)->fin) | |
4654 | return; | |
4655 | ||
4656 | while (before(start, end)) { | |
4657 | struct sk_buff *nskb; | |
4658 | unsigned int header = skb_headroom(skb); | |
4659 | int copy = SKB_MAX_ORDER(header, 0); | |
4660 | ||
4661 | /* Too big header? This can happen with IPv6. */ | |
4662 | if (copy < 0) | |
4663 | return; | |
4664 | if (end - start < copy) | |
4665 | copy = end - start; | |
4666 | nskb = alloc_skb(copy + header, GFP_ATOMIC); | |
4667 | if (!nskb) | |
4668 | return; | |
4669 | ||
4670 | skb_set_mac_header(nskb, skb_mac_header(skb) - skb->head); | |
4671 | skb_set_network_header(nskb, (skb_network_header(skb) - | |
4672 | skb->head)); | |
4673 | skb_set_transport_header(nskb, (skb_transport_header(skb) - | |
4674 | skb->head)); | |
4675 | skb_reserve(nskb, header); | |
4676 | memcpy(nskb->head, skb->head, header); | |
4677 | memcpy(nskb->cb, skb->cb, sizeof(skb->cb)); | |
4678 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(nskb)->end_seq = start; | |
4679 | __skb_queue_before(list, skb, nskb); | |
4680 | skb_set_owner_r(nskb, sk); | |
4681 | ||
4682 | /* Copy data, releasing collapsed skbs. */ | |
4683 | while (copy > 0) { | |
4684 | int offset = start - TCP_SKB_CB(skb)->seq; | |
4685 | int size = TCP_SKB_CB(skb)->end_seq - start; | |
4686 | ||
4687 | BUG_ON(offset < 0); | |
4688 | if (size > 0) { | |
4689 | size = min(copy, size); | |
4690 | if (skb_copy_bits(skb, offset, skb_put(nskb, size), size)) | |
4691 | BUG(); | |
4692 | TCP_SKB_CB(nskb)->end_seq += size; | |
4693 | copy -= size; | |
4694 | start += size; | |
4695 | } | |
4696 | if (!before(start, TCP_SKB_CB(skb)->end_seq)) { | |
4697 | skb = tcp_collapse_one(sk, skb, list); | |
4698 | if (!skb || | |
4699 | skb == tail || | |
4700 | tcp_hdr(skb)->syn || | |
4701 | tcp_hdr(skb)->fin) | |
4702 | return; | |
4703 | } | |
4704 | } | |
4705 | } | |
4706 | } | |
4707 | ||
4708 | /* Collapse ofo queue. Algorithm: select contiguous sequence of skbs | |
4709 | * and tcp_collapse() them until all the queue is collapsed. | |
4710 | */ | |
4711 | static void tcp_collapse_ofo_queue(struct sock *sk) | |
4712 | { | |
4713 | struct tcp_sock *tp = tcp_sk(sk); | |
4714 | struct sk_buff *skb = skb_peek(&tp->out_of_order_queue); | |
4715 | struct sk_buff *head; | |
4716 | u32 start, end; | |
4717 | ||
4718 | if (skb == NULL) | |
4719 | return; | |
4720 | ||
4721 | start = TCP_SKB_CB(skb)->seq; | |
4722 | end = TCP_SKB_CB(skb)->end_seq; | |
4723 | head = skb; | |
4724 | ||
4725 | for (;;) { | |
4726 | struct sk_buff *next = NULL; | |
4727 | ||
4728 | if (!skb_queue_is_last(&tp->out_of_order_queue, skb)) | |
4729 | next = skb_queue_next(&tp->out_of_order_queue, skb); | |
4730 | skb = next; | |
4731 | ||
4732 | /* Segment is terminated when we see gap or when | |
4733 | * we are at the end of all the queue. */ | |
4734 | if (!skb || | |
4735 | after(TCP_SKB_CB(skb)->seq, end) || | |
4736 | before(TCP_SKB_CB(skb)->end_seq, start)) { | |
4737 | tcp_collapse(sk, &tp->out_of_order_queue, | |
4738 | head, skb, start, end); | |
4739 | head = skb; | |
4740 | if (!skb) | |
4741 | break; | |
4742 | /* Start new segment */ | |
4743 | start = TCP_SKB_CB(skb)->seq; | |
4744 | end = TCP_SKB_CB(skb)->end_seq; | |
4745 | } else { | |
4746 | if (before(TCP_SKB_CB(skb)->seq, start)) | |
4747 | start = TCP_SKB_CB(skb)->seq; | |
4748 | if (after(TCP_SKB_CB(skb)->end_seq, end)) | |
4749 | end = TCP_SKB_CB(skb)->end_seq; | |
4750 | } | |
4751 | } | |
4752 | } | |
4753 | ||
4754 | /* | |
4755 | * Purge the out-of-order queue. | |
4756 | * Return true if queue was pruned. | |
4757 | */ | |
4758 | static int tcp_prune_ofo_queue(struct sock *sk) | |
4759 | { | |
4760 | struct tcp_sock *tp = tcp_sk(sk); | |
4761 | int res = 0; | |
4762 | ||
4763 | if (!skb_queue_empty(&tp->out_of_order_queue)) { | |
4764 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_OFOPRUNED); | |
4765 | __skb_queue_purge(&tp->out_of_order_queue); | |
4766 | ||
4767 | /* Reset SACK state. A conforming SACK implementation will | |
4768 | * do the same at a timeout based retransmit. When a connection | |
4769 | * is in a sad state like this, we care only about integrity | |
4770 | * of the connection not performance. | |
4771 | */ | |
4772 | if (tp->rx_opt.sack_ok) | |
4773 | tcp_sack_reset(&tp->rx_opt); | |
4774 | sk_mem_reclaim(sk); | |
4775 | res = 1; | |
4776 | } | |
4777 | return res; | |
4778 | } | |
4779 | ||
4780 | /* Reduce allocated memory if we can, trying to get | |
4781 | * the socket within its memory limits again. | |
4782 | * | |
4783 | * Return less than zero if we should start dropping frames | |
4784 | * until the socket owning process reads some of the data | |
4785 | * to stabilize the situation. | |
4786 | */ | |
4787 | static int tcp_prune_queue(struct sock *sk) | |
4788 | { | |
4789 | struct tcp_sock *tp = tcp_sk(sk); | |
4790 | ||
4791 | SOCK_DEBUG(sk, "prune_queue: c=%x\n", tp->copied_seq); | |
4792 | ||
4793 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PRUNECALLED); | |
4794 | ||
4795 | if (atomic_read(&sk->sk_rmem_alloc) >= sk->sk_rcvbuf) | |
4796 | tcp_clamp_window(sk); | |
4797 | else if (tcp_memory_pressure) | |
4798 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U * tp->advmss); | |
4799 | ||
4800 | tcp_collapse_ofo_queue(sk); | |
4801 | if (!skb_queue_empty(&sk->sk_receive_queue)) | |
4802 | tcp_collapse(sk, &sk->sk_receive_queue, | |
4803 | skb_peek(&sk->sk_receive_queue), | |
4804 | NULL, | |
4805 | tp->copied_seq, tp->rcv_nxt); | |
4806 | sk_mem_reclaim(sk); | |
4807 | ||
4808 | if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) | |
4809 | return 0; | |
4810 | ||
4811 | /* Collapsing did not help, destructive actions follow. | |
4812 | * This must not ever occur. */ | |
4813 | ||
4814 | tcp_prune_ofo_queue(sk); | |
4815 | ||
4816 | if (atomic_read(&sk->sk_rmem_alloc) <= sk->sk_rcvbuf) | |
4817 | return 0; | |
4818 | ||
4819 | /* If we are really being abused, tell the caller to silently | |
4820 | * drop receive data on the floor. It will get retransmitted | |
4821 | * and hopefully then we'll have sufficient space. | |
4822 | */ | |
4823 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_RCVPRUNED); | |
4824 | ||
4825 | /* Massive buffer overcommit. */ | |
4826 | tp->pred_flags = 0; | |
4827 | return -1; | |
4828 | } | |
4829 | ||
4830 | /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. | |
4831 | * As additional protections, we do not touch cwnd in retransmission phases, | |
4832 | * and if application hit its sndbuf limit recently. | |
4833 | */ | |
4834 | void tcp_cwnd_application_limited(struct sock *sk) | |
4835 | { | |
4836 | struct tcp_sock *tp = tcp_sk(sk); | |
4837 | ||
4838 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && | |
4839 | sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { | |
4840 | /* Limited by application or receiver window. */ | |
4841 | u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); | |
4842 | u32 win_used = max(tp->snd_cwnd_used, init_win); | |
4843 | if (win_used < tp->snd_cwnd) { | |
4844 | tp->snd_ssthresh = tcp_current_ssthresh(sk); | |
4845 | tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; | |
4846 | } | |
4847 | tp->snd_cwnd_used = 0; | |
4848 | } | |
4849 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
4850 | } | |
4851 | ||
4852 | static int tcp_should_expand_sndbuf(struct sock *sk) | |
4853 | { | |
4854 | struct tcp_sock *tp = tcp_sk(sk); | |
4855 | ||
4856 | /* If the user specified a specific send buffer setting, do | |
4857 | * not modify it. | |
4858 | */ | |
4859 | if (sk->sk_userlocks & SOCK_SNDBUF_LOCK) | |
4860 | return 0; | |
4861 | ||
4862 | /* If we are under global TCP memory pressure, do not expand. */ | |
4863 | if (tcp_memory_pressure) | |
4864 | return 0; | |
4865 | ||
4866 | /* If we are under soft global TCP memory pressure, do not expand. */ | |
4867 | if (atomic_long_read(&tcp_memory_allocated) >= sysctl_tcp_mem[0]) | |
4868 | return 0; | |
4869 | ||
4870 | /* If we filled the congestion window, do not expand. */ | |
4871 | if (tp->packets_out >= tp->snd_cwnd) | |
4872 | return 0; | |
4873 | ||
4874 | return 1; | |
4875 | } | |
4876 | ||
4877 | /* When incoming ACK allowed to free some skb from write_queue, | |
4878 | * we remember this event in flag SOCK_QUEUE_SHRUNK and wake up socket | |
4879 | * on the exit from tcp input handler. | |
4880 | * | |
4881 | * PROBLEM: sndbuf expansion does not work well with largesend. | |
4882 | */ | |
4883 | static void tcp_new_space(struct sock *sk) | |
4884 | { | |
4885 | struct tcp_sock *tp = tcp_sk(sk); | |
4886 | ||
4887 | if (tcp_should_expand_sndbuf(sk)) { | |
4888 | int sndmem = max_t(u32, tp->rx_opt.mss_clamp, tp->mss_cache) + | |
4889 | MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); | |
4890 | int demanded = max_t(unsigned int, tp->snd_cwnd, | |
4891 | tp->reordering + 1); | |
4892 | sndmem *= 2 * demanded; | |
4893 | if (sndmem > sk->sk_sndbuf) | |
4894 | sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); | |
4895 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
4896 | } | |
4897 | ||
4898 | sk->sk_write_space(sk); | |
4899 | } | |
4900 | ||
4901 | static void tcp_check_space(struct sock *sk) | |
4902 | { | |
4903 | if (sock_flag(sk, SOCK_QUEUE_SHRUNK)) { | |
4904 | sock_reset_flag(sk, SOCK_QUEUE_SHRUNK); | |
4905 | if (sk->sk_socket && | |
4906 | test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) | |
4907 | tcp_new_space(sk); | |
4908 | } | |
4909 | } | |
4910 | ||
4911 | static inline void tcp_data_snd_check(struct sock *sk) | |
4912 | { | |
4913 | tcp_push_pending_frames(sk); | |
4914 | tcp_check_space(sk); | |
4915 | } | |
4916 | ||
4917 | /* | |
4918 | * Check if sending an ack is needed. | |
4919 | */ | |
4920 | static void __tcp_ack_snd_check(struct sock *sk, int ofo_possible) | |
4921 | { | |
4922 | struct tcp_sock *tp = tcp_sk(sk); | |
4923 | ||
4924 | /* More than one full frame received... */ | |
4925 | if (((tp->rcv_nxt - tp->rcv_wup) > inet_csk(sk)->icsk_ack.rcv_mss && | |
4926 | /* ... and right edge of window advances far enough. | |
4927 | * (tcp_recvmsg() will send ACK otherwise). Or... | |
4928 | */ | |
4929 | __tcp_select_window(sk) >= tp->rcv_wnd) || | |
4930 | /* We ACK each frame or... */ | |
4931 | tcp_in_quickack_mode(sk) || | |
4932 | /* We have out of order data. */ | |
4933 | (ofo_possible && skb_peek(&tp->out_of_order_queue))) { | |
4934 | /* Then ack it now */ | |
4935 | tcp_send_ack(sk); | |
4936 | } else { | |
4937 | /* Else, send delayed ack. */ | |
4938 | tcp_send_delayed_ack(sk); | |
4939 | } | |
4940 | } | |
4941 | ||
4942 | static inline void tcp_ack_snd_check(struct sock *sk) | |
4943 | { | |
4944 | if (!inet_csk_ack_scheduled(sk)) { | |
4945 | /* We sent a data segment already. */ | |
4946 | return; | |
4947 | } | |
4948 | __tcp_ack_snd_check(sk, 1); | |
4949 | } | |
4950 | ||
4951 | /* | |
4952 | * This routine is only called when we have urgent data | |
4953 | * signaled. Its the 'slow' part of tcp_urg. It could be | |
4954 | * moved inline now as tcp_urg is only called from one | |
4955 | * place. We handle URGent data wrong. We have to - as | |
4956 | * BSD still doesn't use the correction from RFC961. | |
4957 | * For 1003.1g we should support a new option TCP_STDURG to permit | |
4958 | * either form (or just set the sysctl tcp_stdurg). | |
4959 | */ | |
4960 | ||
4961 | static void tcp_check_urg(struct sock *sk, struct tcphdr *th) | |
4962 | { | |
4963 | struct tcp_sock *tp = tcp_sk(sk); | |
4964 | u32 ptr = ntohs(th->urg_ptr); | |
4965 | ||
4966 | if (ptr && !sysctl_tcp_stdurg) | |
4967 | ptr--; | |
4968 | ptr += ntohl(th->seq); | |
4969 | ||
4970 | /* Ignore urgent data that we've already seen and read. */ | |
4971 | if (after(tp->copied_seq, ptr)) | |
4972 | return; | |
4973 | ||
4974 | /* Do not replay urg ptr. | |
4975 | * | |
4976 | * NOTE: interesting situation not covered by specs. | |
4977 | * Misbehaving sender may send urg ptr, pointing to segment, | |
4978 | * which we already have in ofo queue. We are not able to fetch | |
4979 | * such data and will stay in TCP_URG_NOTYET until will be eaten | |
4980 | * by recvmsg(). Seems, we are not obliged to handle such wicked | |
4981 | * situations. But it is worth to think about possibility of some | |
4982 | * DoSes using some hypothetical application level deadlock. | |
4983 | */ | |
4984 | if (before(ptr, tp->rcv_nxt)) | |
4985 | return; | |
4986 | ||
4987 | /* Do we already have a newer (or duplicate) urgent pointer? */ | |
4988 | if (tp->urg_data && !after(ptr, tp->urg_seq)) | |
4989 | return; | |
4990 | ||
4991 | /* Tell the world about our new urgent pointer. */ | |
4992 | sk_send_sigurg(sk); | |
4993 | ||
4994 | /* We may be adding urgent data when the last byte read was | |
4995 | * urgent. To do this requires some care. We cannot just ignore | |
4996 | * tp->copied_seq since we would read the last urgent byte again | |
4997 | * as data, nor can we alter copied_seq until this data arrives | |
4998 | * or we break the semantics of SIOCATMARK (and thus sockatmark()) | |
4999 | * | |
5000 | * NOTE. Double Dutch. Rendering to plain English: author of comment | |
5001 | * above did something sort of send("A", MSG_OOB); send("B", MSG_OOB); | |
5002 | * and expect that both A and B disappear from stream. This is _wrong_. | |
5003 | * Though this happens in BSD with high probability, this is occasional. | |
5004 | * Any application relying on this is buggy. Note also, that fix "works" | |
5005 | * only in this artificial test. Insert some normal data between A and B and we will | |
5006 | * decline of BSD again. Verdict: it is better to remove to trap | |
5007 | * buggy users. | |
5008 | */ | |
5009 | if (tp->urg_seq == tp->copied_seq && tp->urg_data && | |
5010 | !sock_flag(sk, SOCK_URGINLINE) && tp->copied_seq != tp->rcv_nxt) { | |
5011 | struct sk_buff *skb = skb_peek(&sk->sk_receive_queue); | |
5012 | tp->copied_seq++; | |
5013 | if (skb && !before(tp->copied_seq, TCP_SKB_CB(skb)->end_seq)) { | |
5014 | __skb_unlink(skb, &sk->sk_receive_queue); | |
5015 | __kfree_skb(skb); | |
5016 | } | |
5017 | } | |
5018 | ||
5019 | tp->urg_data = TCP_URG_NOTYET; | |
5020 | tp->urg_seq = ptr; | |
5021 | ||
5022 | /* Disable header prediction. */ | |
5023 | tp->pred_flags = 0; | |
5024 | } | |
5025 | ||
5026 | /* This is the 'fast' part of urgent handling. */ | |
5027 | static void tcp_urg(struct sock *sk, struct sk_buff *skb, struct tcphdr *th) | |
5028 | { | |
5029 | struct tcp_sock *tp = tcp_sk(sk); | |
5030 | ||
5031 | /* Check if we get a new urgent pointer - normally not. */ | |
5032 | if (th->urg) | |
5033 | tcp_check_urg(sk, th); | |
5034 | ||
5035 | /* Do we wait for any urgent data? - normally not... */ | |
5036 | if (tp->urg_data == TCP_URG_NOTYET) { | |
5037 | u32 ptr = tp->urg_seq - ntohl(th->seq) + (th->doff * 4) - | |
5038 | th->syn; | |
5039 | ||
5040 | /* Is the urgent pointer pointing into this packet? */ | |
5041 | if (ptr < skb->len) { | |
5042 | u8 tmp; | |
5043 | if (skb_copy_bits(skb, ptr, &tmp, 1)) | |
5044 | BUG(); | |
5045 | tp->urg_data = TCP_URG_VALID | tmp; | |
5046 | if (!sock_flag(sk, SOCK_DEAD)) | |
5047 | sk->sk_data_ready(sk, 0); | |
5048 | } | |
5049 | } | |
5050 | } | |
5051 | ||
5052 | static int tcp_copy_to_iovec(struct sock *sk, struct sk_buff *skb, int hlen) | |
5053 | { | |
5054 | struct tcp_sock *tp = tcp_sk(sk); | |
5055 | int chunk = skb->len - hlen; | |
5056 | int err; | |
5057 | ||
5058 | local_bh_enable(); | |
5059 | if (skb_csum_unnecessary(skb)) | |
5060 | err = skb_copy_datagram_iovec(skb, hlen, tp->ucopy.iov, chunk); | |
5061 | else | |
5062 | err = skb_copy_and_csum_datagram_iovec(skb, hlen, | |
5063 | tp->ucopy.iov); | |
5064 | ||
5065 | if (!err) { | |
5066 | tp->ucopy.len -= chunk; | |
5067 | tp->copied_seq += chunk; | |
5068 | tcp_rcv_space_adjust(sk); | |
5069 | } | |
5070 | ||
5071 | local_bh_disable(); | |
5072 | return err; | |
5073 | } | |
5074 | ||
5075 | static __sum16 __tcp_checksum_complete_user(struct sock *sk, | |
5076 | struct sk_buff *skb) | |
5077 | { | |
5078 | __sum16 result; | |
5079 | ||
5080 | if (sock_owned_by_user(sk)) { | |
5081 | local_bh_enable(); | |
5082 | result = __tcp_checksum_complete(skb); | |
5083 | local_bh_disable(); | |
5084 | } else { | |
5085 | result = __tcp_checksum_complete(skb); | |
5086 | } | |
5087 | return result; | |
5088 | } | |
5089 | ||
5090 | static inline int tcp_checksum_complete_user(struct sock *sk, | |
5091 | struct sk_buff *skb) | |
5092 | { | |
5093 | return !skb_csum_unnecessary(skb) && | |
5094 | __tcp_checksum_complete_user(sk, skb); | |
5095 | } | |
5096 | ||
5097 | #ifdef CONFIG_NET_DMA | |
5098 | static int tcp_dma_try_early_copy(struct sock *sk, struct sk_buff *skb, | |
5099 | int hlen) | |
5100 | { | |
5101 | struct tcp_sock *tp = tcp_sk(sk); | |
5102 | int chunk = skb->len - hlen; | |
5103 | int dma_cookie; | |
5104 | int copied_early = 0; | |
5105 | ||
5106 | if (tp->ucopy.wakeup) | |
5107 | return 0; | |
5108 | ||
5109 | if (!tp->ucopy.dma_chan && tp->ucopy.pinned_list) | |
5110 | tp->ucopy.dma_chan = dma_find_channel(DMA_MEMCPY); | |
5111 | ||
5112 | if (tp->ucopy.dma_chan && skb_csum_unnecessary(skb)) { | |
5113 | ||
5114 | dma_cookie = dma_skb_copy_datagram_iovec(tp->ucopy.dma_chan, | |
5115 | skb, hlen, | |
5116 | tp->ucopy.iov, chunk, | |
5117 | tp->ucopy.pinned_list); | |
5118 | ||
5119 | if (dma_cookie < 0) | |
5120 | goto out; | |
5121 | ||
5122 | tp->ucopy.dma_cookie = dma_cookie; | |
5123 | copied_early = 1; | |
5124 | ||
5125 | tp->ucopy.len -= chunk; | |
5126 | tp->copied_seq += chunk; | |
5127 | tcp_rcv_space_adjust(sk); | |
5128 | ||
5129 | if ((tp->ucopy.len == 0) || | |
5130 | (tcp_flag_word(tcp_hdr(skb)) & TCP_FLAG_PSH) || | |
5131 | (atomic_read(&sk->sk_rmem_alloc) > (sk->sk_rcvbuf >> 1))) { | |
5132 | tp->ucopy.wakeup = 1; | |
5133 | sk->sk_data_ready(sk, 0); | |
5134 | } | |
5135 | } else if (chunk > 0) { | |
5136 | tp->ucopy.wakeup = 1; | |
5137 | sk->sk_data_ready(sk, 0); | |
5138 | } | |
5139 | out: | |
5140 | return copied_early; | |
5141 | } | |
5142 | #endif /* CONFIG_NET_DMA */ | |
5143 | ||
5144 | /* Does PAWS and seqno based validation of an incoming segment, flags will | |
5145 | * play significant role here. | |
5146 | */ | |
5147 | static int tcp_validate_incoming(struct sock *sk, struct sk_buff *skb, | |
5148 | struct tcphdr *th, int syn_inerr) | |
5149 | { | |
5150 | u8 *hash_location; | |
5151 | struct tcp_sock *tp = tcp_sk(sk); | |
5152 | ||
5153 | /* RFC1323: H1. Apply PAWS check first. */ | |
5154 | if (tcp_fast_parse_options(skb, th, tp, &hash_location) && | |
5155 | tp->rx_opt.saw_tstamp && | |
5156 | tcp_paws_discard(sk, skb)) { | |
5157 | if (!th->rst) { | |
5158 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSESTABREJECTED); | |
5159 | tcp_send_dupack(sk, skb); | |
5160 | goto discard; | |
5161 | } | |
5162 | /* Reset is accepted even if it did not pass PAWS. */ | |
5163 | } | |
5164 | ||
5165 | /* Step 1: check sequence number */ | |
5166 | if (!tcp_sequence(tp, TCP_SKB_CB(skb)->seq, TCP_SKB_CB(skb)->end_seq)) { | |
5167 | /* RFC793, page 37: "In all states except SYN-SENT, all reset | |
5168 | * (RST) segments are validated by checking their SEQ-fields." | |
5169 | * And page 69: "If an incoming segment is not acceptable, | |
5170 | * an acknowledgment should be sent in reply (unless the RST | |
5171 | * bit is set, if so drop the segment and return)". | |
5172 | */ | |
5173 | if (!th->rst) | |
5174 | tcp_send_dupack(sk, skb); | |
5175 | goto discard; | |
5176 | } | |
5177 | ||
5178 | /* Step 2: check RST bit */ | |
5179 | if (th->rst) { | |
5180 | tcp_reset(sk); | |
5181 | goto discard; | |
5182 | } | |
5183 | ||
5184 | /* ts_recent update must be made after we are sure that the packet | |
5185 | * is in window. | |
5186 | */ | |
5187 | tcp_replace_ts_recent(tp, TCP_SKB_CB(skb)->seq); | |
5188 | ||
5189 | /* step 3: check security and precedence [ignored] */ | |
5190 | ||
5191 | /* step 4: Check for a SYN in window. */ | |
5192 | if (th->syn && !before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) { | |
5193 | if (syn_inerr) | |
5194 | TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); | |
5195 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONSYN); | |
5196 | tcp_reset(sk); | |
5197 | return -1; | |
5198 | } | |
5199 | ||
5200 | return 1; | |
5201 | ||
5202 | discard: | |
5203 | __kfree_skb(skb); | |
5204 | return 0; | |
5205 | } | |
5206 | ||
5207 | /* | |
5208 | * TCP receive function for the ESTABLISHED state. | |
5209 | * | |
5210 | * It is split into a fast path and a slow path. The fast path is | |
5211 | * disabled when: | |
5212 | * - A zero window was announced from us - zero window probing | |
5213 | * is only handled properly in the slow path. | |
5214 | * - Out of order segments arrived. | |
5215 | * - Urgent data is expected. | |
5216 | * - There is no buffer space left | |
5217 | * - Unexpected TCP flags/window values/header lengths are received | |
5218 | * (detected by checking the TCP header against pred_flags) | |
5219 | * - Data is sent in both directions. Fast path only supports pure senders | |
5220 | * or pure receivers (this means either the sequence number or the ack | |
5221 | * value must stay constant) | |
5222 | * - Unexpected TCP option. | |
5223 | * | |
5224 | * When these conditions are not satisfied it drops into a standard | |
5225 | * receive procedure patterned after RFC793 to handle all cases. | |
5226 | * The first three cases are guaranteed by proper pred_flags setting, | |
5227 | * the rest is checked inline. Fast processing is turned on in | |
5228 | * tcp_data_queue when everything is OK. | |
5229 | */ | |
5230 | int tcp_rcv_established(struct sock *sk, struct sk_buff *skb, | |
5231 | struct tcphdr *th, unsigned len) | |
5232 | { | |
5233 | struct tcp_sock *tp = tcp_sk(sk); | |
5234 | int res; | |
5235 | ||
5236 | /* | |
5237 | * Header prediction. | |
5238 | * The code loosely follows the one in the famous | |
5239 | * "30 instruction TCP receive" Van Jacobson mail. | |
5240 | * | |
5241 | * Van's trick is to deposit buffers into socket queue | |
5242 | * on a device interrupt, to call tcp_recv function | |
5243 | * on the receive process context and checksum and copy | |
5244 | * the buffer to user space. smart... | |
5245 | * | |
5246 | * Our current scheme is not silly either but we take the | |
5247 | * extra cost of the net_bh soft interrupt processing... | |
5248 | * We do checksum and copy also but from device to kernel. | |
5249 | */ | |
5250 | ||
5251 | tp->rx_opt.saw_tstamp = 0; | |
5252 | ||
5253 | /* pred_flags is 0xS?10 << 16 + snd_wnd | |
5254 | * if header_prediction is to be made | |
5255 | * 'S' will always be tp->tcp_header_len >> 2 | |
5256 | * '?' will be 0 for the fast path, otherwise pred_flags is 0 to | |
5257 | * turn it off (when there are holes in the receive | |
5258 | * space for instance) | |
5259 | * PSH flag is ignored. | |
5260 | */ | |
5261 | ||
5262 | if ((tcp_flag_word(th) & TCP_HP_BITS) == tp->pred_flags && | |
5263 | TCP_SKB_CB(skb)->seq == tp->rcv_nxt && | |
5264 | !after(TCP_SKB_CB(skb)->ack_seq, tp->snd_nxt)) { | |
5265 | int tcp_header_len = tp->tcp_header_len; | |
5266 | ||
5267 | /* Timestamp header prediction: tcp_header_len | |
5268 | * is automatically equal to th->doff*4 due to pred_flags | |
5269 | * match. | |
5270 | */ | |
5271 | ||
5272 | /* Check timestamp */ | |
5273 | if (tcp_header_len == sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) { | |
5274 | /* No? Slow path! */ | |
5275 | if (!tcp_parse_aligned_timestamp(tp, th)) | |
5276 | goto slow_path; | |
5277 | ||
5278 | /* If PAWS failed, check it more carefully in slow path */ | |
5279 | if ((s32)(tp->rx_opt.rcv_tsval - tp->rx_opt.ts_recent) < 0) | |
5280 | goto slow_path; | |
5281 | ||
5282 | /* DO NOT update ts_recent here, if checksum fails | |
5283 | * and timestamp was corrupted part, it will result | |
5284 | * in a hung connection since we will drop all | |
5285 | * future packets due to the PAWS test. | |
5286 | */ | |
5287 | } | |
5288 | ||
5289 | if (len <= tcp_header_len) { | |
5290 | /* Bulk data transfer: sender */ | |
5291 | if (len == tcp_header_len) { | |
5292 | /* Predicted packet is in window by definition. | |
5293 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. | |
5294 | * Hence, check seq<=rcv_wup reduces to: | |
5295 | */ | |
5296 | if (tcp_header_len == | |
5297 | (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && | |
5298 | tp->rcv_nxt == tp->rcv_wup) | |
5299 | tcp_store_ts_recent(tp); | |
5300 | ||
5301 | /* We know that such packets are checksummed | |
5302 | * on entry. | |
5303 | */ | |
5304 | tcp_ack(sk, skb, 0); | |
5305 | __kfree_skb(skb); | |
5306 | tcp_data_snd_check(sk); | |
5307 | return 0; | |
5308 | } else { /* Header too small */ | |
5309 | TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); | |
5310 | goto discard; | |
5311 | } | |
5312 | } else { | |
5313 | int eaten = 0; | |
5314 | int copied_early = 0; | |
5315 | ||
5316 | if (tp->copied_seq == tp->rcv_nxt && | |
5317 | len - tcp_header_len <= tp->ucopy.len) { | |
5318 | #ifdef CONFIG_NET_DMA | |
5319 | if (tcp_dma_try_early_copy(sk, skb, tcp_header_len)) { | |
5320 | copied_early = 1; | |
5321 | eaten = 1; | |
5322 | } | |
5323 | #endif | |
5324 | if (tp->ucopy.task == current && | |
5325 | sock_owned_by_user(sk) && !copied_early) { | |
5326 | __set_current_state(TASK_RUNNING); | |
5327 | ||
5328 | if (!tcp_copy_to_iovec(sk, skb, tcp_header_len)) | |
5329 | eaten = 1; | |
5330 | } | |
5331 | if (eaten) { | |
5332 | /* Predicted packet is in window by definition. | |
5333 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. | |
5334 | * Hence, check seq<=rcv_wup reduces to: | |
5335 | */ | |
5336 | if (tcp_header_len == | |
5337 | (sizeof(struct tcphdr) + | |
5338 | TCPOLEN_TSTAMP_ALIGNED) && | |
5339 | tp->rcv_nxt == tp->rcv_wup) | |
5340 | tcp_store_ts_recent(tp); | |
5341 | ||
5342 | tcp_rcv_rtt_measure_ts(sk, skb); | |
5343 | ||
5344 | __skb_pull(skb, tcp_header_len); | |
5345 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
5346 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITSTOUSER); | |
5347 | } | |
5348 | if (copied_early) | |
5349 | tcp_cleanup_rbuf(sk, skb->len); | |
5350 | } | |
5351 | if (!eaten) { | |
5352 | if (tcp_checksum_complete_user(sk, skb)) | |
5353 | goto csum_error; | |
5354 | ||
5355 | /* Predicted packet is in window by definition. | |
5356 | * seq == rcv_nxt and rcv_wup <= rcv_nxt. | |
5357 | * Hence, check seq<=rcv_wup reduces to: | |
5358 | */ | |
5359 | if (tcp_header_len == | |
5360 | (sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED) && | |
5361 | tp->rcv_nxt == tp->rcv_wup) | |
5362 | tcp_store_ts_recent(tp); | |
5363 | ||
5364 | tcp_rcv_rtt_measure_ts(sk, skb); | |
5365 | ||
5366 | if ((int)skb->truesize > sk->sk_forward_alloc) | |
5367 | goto step5; | |
5368 | ||
5369 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPHPHITS); | |
5370 | ||
5371 | /* Bulk data transfer: receiver */ | |
5372 | __skb_pull(skb, tcp_header_len); | |
5373 | __skb_queue_tail(&sk->sk_receive_queue, skb); | |
5374 | skb_set_owner_r(skb, sk); | |
5375 | tp->rcv_nxt = TCP_SKB_CB(skb)->end_seq; | |
5376 | } | |
5377 | ||
5378 | tcp_event_data_recv(sk, skb); | |
5379 | ||
5380 | if (TCP_SKB_CB(skb)->ack_seq != tp->snd_una) { | |
5381 | /* Well, only one small jumplet in fast path... */ | |
5382 | tcp_ack(sk, skb, FLAG_DATA); | |
5383 | tcp_data_snd_check(sk); | |
5384 | if (!inet_csk_ack_scheduled(sk)) | |
5385 | goto no_ack; | |
5386 | } | |
5387 | ||
5388 | if (!copied_early || tp->rcv_nxt != tp->rcv_wup) | |
5389 | __tcp_ack_snd_check(sk, 0); | |
5390 | no_ack: | |
5391 | #ifdef CONFIG_NET_DMA | |
5392 | if (copied_early) | |
5393 | __skb_queue_tail(&sk->sk_async_wait_queue, skb); | |
5394 | else | |
5395 | #endif | |
5396 | if (eaten) | |
5397 | __kfree_skb(skb); | |
5398 | else | |
5399 | sk->sk_data_ready(sk, 0); | |
5400 | return 0; | |
5401 | } | |
5402 | } | |
5403 | ||
5404 | slow_path: | |
5405 | if (len < (th->doff << 2) || tcp_checksum_complete_user(sk, skb)) | |
5406 | goto csum_error; | |
5407 | ||
5408 | /* | |
5409 | * Standard slow path. | |
5410 | */ | |
5411 | ||
5412 | res = tcp_validate_incoming(sk, skb, th, 1); | |
5413 | if (res <= 0) | |
5414 | return -res; | |
5415 | ||
5416 | step5: | |
5417 | if (th->ack && tcp_ack(sk, skb, FLAG_SLOWPATH) < 0) | |
5418 | goto discard; | |
5419 | ||
5420 | tcp_rcv_rtt_measure_ts(sk, skb); | |
5421 | ||
5422 | /* Process urgent data. */ | |
5423 | tcp_urg(sk, skb, th); | |
5424 | ||
5425 | /* step 7: process the segment text */ | |
5426 | tcp_data_queue(sk, skb); | |
5427 | ||
5428 | tcp_data_snd_check(sk); | |
5429 | tcp_ack_snd_check(sk); | |
5430 | return 0; | |
5431 | ||
5432 | csum_error: | |
5433 | TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_INERRS); | |
5434 | ||
5435 | discard: | |
5436 | __kfree_skb(skb); | |
5437 | return 0; | |
5438 | } | |
5439 | EXPORT_SYMBOL(tcp_rcv_established); | |
5440 | ||
5441 | static int tcp_rcv_synsent_state_process(struct sock *sk, struct sk_buff *skb, | |
5442 | struct tcphdr *th, unsigned len) | |
5443 | { | |
5444 | u8 *hash_location; | |
5445 | struct inet_connection_sock *icsk = inet_csk(sk); | |
5446 | struct tcp_sock *tp = tcp_sk(sk); | |
5447 | struct tcp_cookie_values *cvp = tp->cookie_values; | |
5448 | int saved_clamp = tp->rx_opt.mss_clamp; | |
5449 | ||
5450 | tcp_parse_options(skb, &tp->rx_opt, &hash_location, 0); | |
5451 | ||
5452 | if (th->ack) { | |
5453 | /* rfc793: | |
5454 | * "If the state is SYN-SENT then | |
5455 | * first check the ACK bit | |
5456 | * If the ACK bit is set | |
5457 | * If SEG.ACK =< ISS, or SEG.ACK > SND.NXT, send | |
5458 | * a reset (unless the RST bit is set, if so drop | |
5459 | * the segment and return)" | |
5460 | * | |
5461 | * We do not send data with SYN, so that RFC-correct | |
5462 | * test reduces to: | |
5463 | */ | |
5464 | if (TCP_SKB_CB(skb)->ack_seq != tp->snd_nxt) | |
5465 | goto reset_and_undo; | |
5466 | ||
5467 | if (tp->rx_opt.saw_tstamp && tp->rx_opt.rcv_tsecr && | |
5468 | !between(tp->rx_opt.rcv_tsecr, tp->retrans_stamp, | |
5469 | tcp_time_stamp)) { | |
5470 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_PAWSACTIVEREJECTED); | |
5471 | goto reset_and_undo; | |
5472 | } | |
5473 | ||
5474 | /* Now ACK is acceptable. | |
5475 | * | |
5476 | * "If the RST bit is set | |
5477 | * If the ACK was acceptable then signal the user "error: | |
5478 | * connection reset", drop the segment, enter CLOSED state, | |
5479 | * delete TCB, and return." | |
5480 | */ | |
5481 | ||
5482 | if (th->rst) { | |
5483 | tcp_reset(sk); | |
5484 | goto discard; | |
5485 | } | |
5486 | ||
5487 | /* rfc793: | |
5488 | * "fifth, if neither of the SYN or RST bits is set then | |
5489 | * drop the segment and return." | |
5490 | * | |
5491 | * See note below! | |
5492 | * --ANK(990513) | |
5493 | */ | |
5494 | if (!th->syn) | |
5495 | goto discard_and_undo; | |
5496 | ||
5497 | /* rfc793: | |
5498 | * "If the SYN bit is on ... | |
5499 | * are acceptable then ... | |
5500 | * (our SYN has been ACKed), change the connection | |
5501 | * state to ESTABLISHED..." | |
5502 | */ | |
5503 | ||
5504 | TCP_ECN_rcv_synack(tp, th); | |
5505 | ||
5506 | tp->snd_wl1 = TCP_SKB_CB(skb)->seq; | |
5507 | tcp_ack(sk, skb, FLAG_SLOWPATH); | |
5508 | ||
5509 | /* Ok.. it's good. Set up sequence numbers and | |
5510 | * move to established. | |
5511 | */ | |
5512 | tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; | |
5513 | tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; | |
5514 | ||
5515 | /* RFC1323: The window in SYN & SYN/ACK segments is | |
5516 | * never scaled. | |
5517 | */ | |
5518 | tp->snd_wnd = ntohs(th->window); | |
5519 | tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); | |
5520 | ||
5521 | if (!tp->rx_opt.wscale_ok) { | |
5522 | tp->rx_opt.snd_wscale = tp->rx_opt.rcv_wscale = 0; | |
5523 | tp->window_clamp = min(tp->window_clamp, 65535U); | |
5524 | } | |
5525 | ||
5526 | if (tp->rx_opt.saw_tstamp) { | |
5527 | tp->rx_opt.tstamp_ok = 1; | |
5528 | tp->tcp_header_len = | |
5529 | sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; | |
5530 | tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; | |
5531 | tcp_store_ts_recent(tp); | |
5532 | } else { | |
5533 | tp->tcp_header_len = sizeof(struct tcphdr); | |
5534 | } | |
5535 | ||
5536 | if (tcp_is_sack(tp) && sysctl_tcp_fack) | |
5537 | tcp_enable_fack(tp); | |
5538 | ||
5539 | tcp_mtup_init(sk); | |
5540 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); | |
5541 | tcp_initialize_rcv_mss(sk); | |
5542 | ||
5543 | /* Remember, tcp_poll() does not lock socket! | |
5544 | * Change state from SYN-SENT only after copied_seq | |
5545 | * is initialized. */ | |
5546 | tp->copied_seq = tp->rcv_nxt; | |
5547 | ||
5548 | if (cvp != NULL && | |
5549 | cvp->cookie_pair_size > 0 && | |
5550 | tp->rx_opt.cookie_plus > 0) { | |
5551 | int cookie_size = tp->rx_opt.cookie_plus | |
5552 | - TCPOLEN_COOKIE_BASE; | |
5553 | int cookie_pair_size = cookie_size | |
5554 | + cvp->cookie_desired; | |
5555 | ||
5556 | /* A cookie extension option was sent and returned. | |
5557 | * Note that each incoming SYNACK replaces the | |
5558 | * Responder cookie. The initial exchange is most | |
5559 | * fragile, as protection against spoofing relies | |
5560 | * entirely upon the sequence and timestamp (above). | |
5561 | * This replacement strategy allows the correct pair to | |
5562 | * pass through, while any others will be filtered via | |
5563 | * Responder verification later. | |
5564 | */ | |
5565 | if (sizeof(cvp->cookie_pair) >= cookie_pair_size) { | |
5566 | memcpy(&cvp->cookie_pair[cvp->cookie_desired], | |
5567 | hash_location, cookie_size); | |
5568 | cvp->cookie_pair_size = cookie_pair_size; | |
5569 | } | |
5570 | } | |
5571 | ||
5572 | smp_mb(); | |
5573 | tcp_set_state(sk, TCP_ESTABLISHED); | |
5574 | ||
5575 | security_inet_conn_established(sk, skb); | |
5576 | ||
5577 | /* Make sure socket is routed, for correct metrics. */ | |
5578 | icsk->icsk_af_ops->rebuild_header(sk); | |
5579 | ||
5580 | tcp_init_metrics(sk); | |
5581 | ||
5582 | tcp_init_congestion_control(sk); | |
5583 | ||
5584 | /* Prevent spurious tcp_cwnd_restart() on first data | |
5585 | * packet. | |
5586 | */ | |
5587 | tp->lsndtime = tcp_time_stamp; | |
5588 | ||
5589 | tcp_init_buffer_space(sk); | |
5590 | ||
5591 | if (sock_flag(sk, SOCK_KEEPOPEN)) | |
5592 | inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tp)); | |
5593 | ||
5594 | if (!tp->rx_opt.snd_wscale) | |
5595 | __tcp_fast_path_on(tp, tp->snd_wnd); | |
5596 | else | |
5597 | tp->pred_flags = 0; | |
5598 | ||
5599 | if (!sock_flag(sk, SOCK_DEAD)) { | |
5600 | sk->sk_state_change(sk); | |
5601 | sk_wake_async(sk, SOCK_WAKE_IO, POLL_OUT); | |
5602 | } | |
5603 | ||
5604 | if (sk->sk_write_pending || | |
5605 | icsk->icsk_accept_queue.rskq_defer_accept || | |
5606 | icsk->icsk_ack.pingpong) { | |
5607 | /* Save one ACK. Data will be ready after | |
5608 | * several ticks, if write_pending is set. | |
5609 | * | |
5610 | * It may be deleted, but with this feature tcpdumps | |
5611 | * look so _wonderfully_ clever, that I was not able | |
5612 | * to stand against the temptation 8) --ANK | |
5613 | */ | |
5614 | inet_csk_schedule_ack(sk); | |
5615 | icsk->icsk_ack.lrcvtime = tcp_time_stamp; | |
5616 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
5617 | tcp_incr_quickack(sk); | |
5618 | tcp_enter_quickack_mode(sk); | |
5619 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, | |
5620 | TCP_DELACK_MAX, TCP_RTO_MAX); | |
5621 | ||
5622 | discard: | |
5623 | __kfree_skb(skb); | |
5624 | return 0; | |
5625 | } else { | |
5626 | tcp_send_ack(sk); | |
5627 | } | |
5628 | return -1; | |
5629 | } | |
5630 | ||
5631 | /* No ACK in the segment */ | |
5632 | ||
5633 | if (th->rst) { | |
5634 | /* rfc793: | |
5635 | * "If the RST bit is set | |
5636 | * | |
5637 | * Otherwise (no ACK) drop the segment and return." | |
5638 | */ | |
5639 | ||
5640 | goto discard_and_undo; | |
5641 | } | |
5642 | ||
5643 | /* PAWS check. */ | |
5644 | if (tp->rx_opt.ts_recent_stamp && tp->rx_opt.saw_tstamp && | |
5645 | tcp_paws_reject(&tp->rx_opt, 0)) | |
5646 | goto discard_and_undo; | |
5647 | ||
5648 | if (th->syn) { | |
5649 | /* We see SYN without ACK. It is attempt of | |
5650 | * simultaneous connect with crossed SYNs. | |
5651 | * Particularly, it can be connect to self. | |
5652 | */ | |
5653 | tcp_set_state(sk, TCP_SYN_RECV); | |
5654 | ||
5655 | if (tp->rx_opt.saw_tstamp) { | |
5656 | tp->rx_opt.tstamp_ok = 1; | |
5657 | tcp_store_ts_recent(tp); | |
5658 | tp->tcp_header_len = | |
5659 | sizeof(struct tcphdr) + TCPOLEN_TSTAMP_ALIGNED; | |
5660 | } else { | |
5661 | tp->tcp_header_len = sizeof(struct tcphdr); | |
5662 | } | |
5663 | ||
5664 | tp->rcv_nxt = TCP_SKB_CB(skb)->seq + 1; | |
5665 | tp->rcv_wup = TCP_SKB_CB(skb)->seq + 1; | |
5666 | ||
5667 | /* RFC1323: The window in SYN & SYN/ACK segments is | |
5668 | * never scaled. | |
5669 | */ | |
5670 | tp->snd_wnd = ntohs(th->window); | |
5671 | tp->snd_wl1 = TCP_SKB_CB(skb)->seq; | |
5672 | tp->max_window = tp->snd_wnd; | |
5673 | ||
5674 | TCP_ECN_rcv_syn(tp, th); | |
5675 | ||
5676 | tcp_mtup_init(sk); | |
5677 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); | |
5678 | tcp_initialize_rcv_mss(sk); | |
5679 | ||
5680 | tcp_send_synack(sk); | |
5681 | #if 0 | |
5682 | /* Note, we could accept data and URG from this segment. | |
5683 | * There are no obstacles to make this. | |
5684 | * | |
5685 | * However, if we ignore data in ACKless segments sometimes, | |
5686 | * we have no reasons to accept it sometimes. | |
5687 | * Also, seems the code doing it in step6 of tcp_rcv_state_process | |
5688 | * is not flawless. So, discard packet for sanity. | |
5689 | * Uncomment this return to process the data. | |
5690 | */ | |
5691 | return -1; | |
5692 | #else | |
5693 | goto discard; | |
5694 | #endif | |
5695 | } | |
5696 | /* "fifth, if neither of the SYN or RST bits is set then | |
5697 | * drop the segment and return." | |
5698 | */ | |
5699 | ||
5700 | discard_and_undo: | |
5701 | tcp_clear_options(&tp->rx_opt); | |
5702 | tp->rx_opt.mss_clamp = saved_clamp; | |
5703 | goto discard; | |
5704 | ||
5705 | reset_and_undo: | |
5706 | tcp_clear_options(&tp->rx_opt); | |
5707 | tp->rx_opt.mss_clamp = saved_clamp; | |
5708 | return 1; | |
5709 | } | |
5710 | ||
5711 | /* | |
5712 | * This function implements the receiving procedure of RFC 793 for | |
5713 | * all states except ESTABLISHED and TIME_WAIT. | |
5714 | * It's called from both tcp_v4_rcv and tcp_v6_rcv and should be | |
5715 | * address independent. | |
5716 | */ | |
5717 | ||
5718 | int tcp_rcv_state_process(struct sock *sk, struct sk_buff *skb, | |
5719 | struct tcphdr *th, unsigned len) | |
5720 | { | |
5721 | struct tcp_sock *tp = tcp_sk(sk); | |
5722 | struct inet_connection_sock *icsk = inet_csk(sk); | |
5723 | int queued = 0; | |
5724 | int res; | |
5725 | ||
5726 | tp->rx_opt.saw_tstamp = 0; | |
5727 | ||
5728 | switch (sk->sk_state) { | |
5729 | case TCP_CLOSE: | |
5730 | goto discard; | |
5731 | ||
5732 | case TCP_LISTEN: | |
5733 | if (th->ack) | |
5734 | return 1; | |
5735 | ||
5736 | if (th->rst) | |
5737 | goto discard; | |
5738 | ||
5739 | if (th->syn) { | |
5740 | if (icsk->icsk_af_ops->conn_request(sk, skb) < 0) | |
5741 | return 1; | |
5742 | ||
5743 | /* Now we have several options: In theory there is | |
5744 | * nothing else in the frame. KA9Q has an option to | |
5745 | * send data with the syn, BSD accepts data with the | |
5746 | * syn up to the [to be] advertised window and | |
5747 | * Solaris 2.1 gives you a protocol error. For now | |
5748 | * we just ignore it, that fits the spec precisely | |
5749 | * and avoids incompatibilities. It would be nice in | |
5750 | * future to drop through and process the data. | |
5751 | * | |
5752 | * Now that TTCP is starting to be used we ought to | |
5753 | * queue this data. | |
5754 | * But, this leaves one open to an easy denial of | |
5755 | * service attack, and SYN cookies can't defend | |
5756 | * against this problem. So, we drop the data | |
5757 | * in the interest of security over speed unless | |
5758 | * it's still in use. | |
5759 | */ | |
5760 | kfree_skb(skb); | |
5761 | return 0; | |
5762 | } | |
5763 | goto discard; | |
5764 | ||
5765 | case TCP_SYN_SENT: | |
5766 | queued = tcp_rcv_synsent_state_process(sk, skb, th, len); | |
5767 | if (queued >= 0) | |
5768 | return queued; | |
5769 | ||
5770 | /* Do step6 onward by hand. */ | |
5771 | tcp_urg(sk, skb, th); | |
5772 | __kfree_skb(skb); | |
5773 | tcp_data_snd_check(sk); | |
5774 | return 0; | |
5775 | } | |
5776 | ||
5777 | res = tcp_validate_incoming(sk, skb, th, 0); | |
5778 | if (res <= 0) | |
5779 | return -res; | |
5780 | ||
5781 | /* step 5: check the ACK field */ | |
5782 | if (th->ack) { | |
5783 | int acceptable = tcp_ack(sk, skb, FLAG_SLOWPATH) > 0; | |
5784 | ||
5785 | switch (sk->sk_state) { | |
5786 | case TCP_SYN_RECV: | |
5787 | if (acceptable) { | |
5788 | tp->copied_seq = tp->rcv_nxt; | |
5789 | smp_mb(); | |
5790 | tcp_set_state(sk, TCP_ESTABLISHED); | |
5791 | sk->sk_state_change(sk); | |
5792 | ||
5793 | /* Note, that this wakeup is only for marginal | |
5794 | * crossed SYN case. Passively open sockets | |
5795 | * are not waked up, because sk->sk_sleep == | |
5796 | * NULL and sk->sk_socket == NULL. | |
5797 | */ | |
5798 | if (sk->sk_socket) | |
5799 | sk_wake_async(sk, | |
5800 | SOCK_WAKE_IO, POLL_OUT); | |
5801 | ||
5802 | tp->snd_una = TCP_SKB_CB(skb)->ack_seq; | |
5803 | tp->snd_wnd = ntohs(th->window) << | |
5804 | tp->rx_opt.snd_wscale; | |
5805 | tcp_init_wl(tp, TCP_SKB_CB(skb)->seq); | |
5806 | ||
5807 | /* tcp_ack considers this ACK as duplicate | |
5808 | * and does not calculate rtt. | |
5809 | * Force it here. | |
5810 | */ | |
5811 | tcp_ack_update_rtt(sk, 0, 0); | |
5812 | ||
5813 | if (tp->rx_opt.tstamp_ok) | |
5814 | tp->advmss -= TCPOLEN_TSTAMP_ALIGNED; | |
5815 | ||
5816 | /* Make sure socket is routed, for | |
5817 | * correct metrics. | |
5818 | */ | |
5819 | icsk->icsk_af_ops->rebuild_header(sk); | |
5820 | ||
5821 | tcp_init_metrics(sk); | |
5822 | ||
5823 | tcp_init_congestion_control(sk); | |
5824 | ||
5825 | /* Prevent spurious tcp_cwnd_restart() on | |
5826 | * first data packet. | |
5827 | */ | |
5828 | tp->lsndtime = tcp_time_stamp; | |
5829 | ||
5830 | tcp_mtup_init(sk); | |
5831 | tcp_initialize_rcv_mss(sk); | |
5832 | tcp_init_buffer_space(sk); | |
5833 | tcp_fast_path_on(tp); | |
5834 | } else { | |
5835 | return 1; | |
5836 | } | |
5837 | break; | |
5838 | ||
5839 | case TCP_FIN_WAIT1: | |
5840 | if (tp->snd_una == tp->write_seq) { | |
5841 | tcp_set_state(sk, TCP_FIN_WAIT2); | |
5842 | sk->sk_shutdown |= SEND_SHUTDOWN; | |
5843 | dst_confirm(__sk_dst_get(sk)); | |
5844 | ||
5845 | if (!sock_flag(sk, SOCK_DEAD)) | |
5846 | /* Wake up lingering close() */ | |
5847 | sk->sk_state_change(sk); | |
5848 | else { | |
5849 | int tmo; | |
5850 | ||
5851 | if (tp->linger2 < 0 || | |
5852 | (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && | |
5853 | after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt))) { | |
5854 | tcp_done(sk); | |
5855 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA); | |
5856 | return 1; | |
5857 | } | |
5858 | ||
5859 | tmo = tcp_fin_time(sk); | |
5860 | if (tmo > TCP_TIMEWAIT_LEN) { | |
5861 | inet_csk_reset_keepalive_timer(sk, tmo - TCP_TIMEWAIT_LEN); | |
5862 | } else if (th->fin || sock_owned_by_user(sk)) { | |
5863 | /* Bad case. We could lose such FIN otherwise. | |
5864 | * It is not a big problem, but it looks confusing | |
5865 | * and not so rare event. We still can lose it now, | |
5866 | * if it spins in bh_lock_sock(), but it is really | |
5867 | * marginal case. | |
5868 | */ | |
5869 | inet_csk_reset_keepalive_timer(sk, tmo); | |
5870 | } else { | |
5871 | tcp_time_wait(sk, TCP_FIN_WAIT2, tmo); | |
5872 | goto discard; | |
5873 | } | |
5874 | } | |
5875 | } | |
5876 | break; | |
5877 | ||
5878 | case TCP_CLOSING: | |
5879 | if (tp->snd_una == tp->write_seq) { | |
5880 | tcp_time_wait(sk, TCP_TIME_WAIT, 0); | |
5881 | goto discard; | |
5882 | } | |
5883 | break; | |
5884 | ||
5885 | case TCP_LAST_ACK: | |
5886 | if (tp->snd_una == tp->write_seq) { | |
5887 | tcp_update_metrics(sk); | |
5888 | tcp_done(sk); | |
5889 | goto discard; | |
5890 | } | |
5891 | break; | |
5892 | } | |
5893 | } else | |
5894 | goto discard; | |
5895 | ||
5896 | /* step 6: check the URG bit */ | |
5897 | tcp_urg(sk, skb, th); | |
5898 | ||
5899 | /* step 7: process the segment text */ | |
5900 | switch (sk->sk_state) { | |
5901 | case TCP_CLOSE_WAIT: | |
5902 | case TCP_CLOSING: | |
5903 | case TCP_LAST_ACK: | |
5904 | if (!before(TCP_SKB_CB(skb)->seq, tp->rcv_nxt)) | |
5905 | break; | |
5906 | case TCP_FIN_WAIT1: | |
5907 | case TCP_FIN_WAIT2: | |
5908 | /* RFC 793 says to queue data in these states, | |
5909 | * RFC 1122 says we MUST send a reset. | |
5910 | * BSD 4.4 also does reset. | |
5911 | */ | |
5912 | if (sk->sk_shutdown & RCV_SHUTDOWN) { | |
5913 | if (TCP_SKB_CB(skb)->end_seq != TCP_SKB_CB(skb)->seq && | |
5914 | after(TCP_SKB_CB(skb)->end_seq - th->fin, tp->rcv_nxt)) { | |
5915 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPABORTONDATA); | |
5916 | tcp_reset(sk); | |
5917 | return 1; | |
5918 | } | |
5919 | } | |
5920 | /* Fall through */ | |
5921 | case TCP_ESTABLISHED: | |
5922 | tcp_data_queue(sk, skb); | |
5923 | queued = 1; | |
5924 | break; | |
5925 | } | |
5926 | ||
5927 | /* tcp_data could move socket to TIME-WAIT */ | |
5928 | if (sk->sk_state != TCP_CLOSE) { | |
5929 | tcp_data_snd_check(sk); | |
5930 | tcp_ack_snd_check(sk); | |
5931 | } | |
5932 | ||
5933 | if (!queued) { | |
5934 | discard: | |
5935 | __kfree_skb(skb); | |
5936 | } | |
5937 | return 0; | |
5938 | } | |
5939 | EXPORT_SYMBOL(tcp_rcv_state_process); |