Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * INET An implementation of the TCP/IP protocol suite for the LINUX | |
3 | * operating system. INET is implemented using the BSD Socket | |
4 | * interface as the means of communication with the user level. | |
5 | * | |
6 | * Implementation of the Transmission Control Protocol(TCP). | |
7 | * | |
8 | * Version: $Id: tcp_input.c,v 1.243 2002/02/01 22:01:04 davem Exp $ | |
9 | * | |
02c30a84 | 10 | * Authors: Ross Biro |
1da177e4 LT |
11 | * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> |
12 | * Mark Evans, <evansmp@uhura.aston.ac.uk> | |
13 | * Corey Minyard <wf-rch!minyard@relay.EU.net> | |
14 | * Florian La Roche, <flla@stud.uni-sb.de> | |
15 | * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> | |
16 | * Linus Torvalds, <torvalds@cs.helsinki.fi> | |
17 | * Alan Cox, <gw4pts@gw4pts.ampr.org> | |
18 | * Matthew Dillon, <dillon@apollo.west.oic.com> | |
19 | * Arnt Gulbrandsen, <agulbra@nvg.unit.no> | |
20 | * Jorge Cwik, <jorge@laser.satlink.net> | |
21 | */ | |
22 | ||
23 | /* | |
24 | * Changes: | |
25 | * Pedro Roque : Fast Retransmit/Recovery. | |
26 | * Two receive queues. | |
27 | * Retransmit queue handled by TCP. | |
28 | * Better retransmit timer handling. | |
29 | * New congestion avoidance. | |
30 | * Header prediction. | |
31 | * Variable renaming. | |
32 | * | |
33 | * Eric : Fast Retransmit. | |
34 | * Randy Scott : MSS option defines. | |
35 | * Eric Schenk : Fixes to slow start algorithm. | |
36 | * Eric Schenk : Yet another double ACK bug. | |
37 | * Eric Schenk : Delayed ACK bug fixes. | |
38 | * Eric Schenk : Floyd style fast retrans war avoidance. | |
39 | * David S. Miller : Don't allow zero congestion window. | |
40 | * Eric Schenk : Fix retransmitter so that it sends | |
41 | * next packet on ack of previous packet. | |
42 | * Andi Kleen : Moved open_request checking here | |
43 | * and process RSTs for open_requests. | |
44 | * Andi Kleen : Better prune_queue, and other fixes. | |
caa20d9a | 45 | * Andrey Savochkin: Fix RTT measurements in the presence of |
1da177e4 LT |
46 | * timestamps. |
47 | * Andrey Savochkin: Check sequence numbers correctly when | |
48 | * removing SACKs due to in sequence incoming | |
49 | * data segments. | |
50 | * Andi Kleen: Make sure we never ack data there is not | |
51 | * enough room for. Also make this condition | |
52 | * a fatal error if it might still happen. | |
e905a9ed | 53 | * Andi Kleen: Add tcp_measure_rcv_mss to make |
1da177e4 | 54 | * connections with MSS<min(MTU,ann. MSS) |
e905a9ed | 55 | * work without delayed acks. |
1da177e4 LT |
56 | * Andi Kleen: Process packets with PSH set in the |
57 | * fast path. | |
58 | * J Hadi Salim: ECN support | |
59 | * Andrei Gurtov, | |
60 | * Pasi Sarolahti, | |
61 | * Panu Kuhlberg: Experimental audit of TCP (re)transmission | |
62 | * engine. Lots of bugs are found. | |
63 | * Pasi Sarolahti: F-RTO for dealing with spurious RTOs | |
1da177e4 LT |
64 | */ |
65 | ||
1da177e4 LT |
66 | #include <linux/mm.h> |
67 | #include <linux/module.h> | |
68 | #include <linux/sysctl.h> | |
69 | #include <net/tcp.h> | |
70 | #include <net/inet_common.h> | |
71 | #include <linux/ipsec.h> | |
72 | #include <asm/unaligned.h> | |
1a2449a8 | 73 | #include <net/netdma.h> |
1da177e4 | 74 | |
ab32ea5d BH |
75 | int sysctl_tcp_timestamps __read_mostly = 1; |
76 | int sysctl_tcp_window_scaling __read_mostly = 1; | |
77 | int sysctl_tcp_sack __read_mostly = 1; | |
78 | int sysctl_tcp_fack __read_mostly = 1; | |
79 | int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH; | |
80 | int sysctl_tcp_ecn __read_mostly; | |
81 | int sysctl_tcp_dsack __read_mostly = 1; | |
82 | int sysctl_tcp_app_win __read_mostly = 31; | |
83 | int sysctl_tcp_adv_win_scale __read_mostly = 2; | |
1da177e4 | 84 | |
ab32ea5d BH |
85 | int sysctl_tcp_stdurg __read_mostly; |
86 | int sysctl_tcp_rfc1337 __read_mostly; | |
87 | int sysctl_tcp_max_orphans __read_mostly = NR_FILE; | |
c96fd3d4 | 88 | int sysctl_tcp_frto __read_mostly = 2; |
3cfe3baa | 89 | int sysctl_tcp_frto_response __read_mostly; |
ab32ea5d | 90 | int sysctl_tcp_nometrics_save __read_mostly; |
1da177e4 | 91 | |
ab32ea5d BH |
92 | int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; |
93 | int sysctl_tcp_abc __read_mostly; | |
1da177e4 | 94 | |
1da177e4 LT |
95 | #define FLAG_DATA 0x01 /* Incoming frame contained data. */ |
96 | #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ | |
97 | #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ | |
98 | #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ | |
99 | #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ | |
100 | #define FLAG_DATA_SACKED 0x20 /* New SACK. */ | |
101 | #define FLAG_ECE 0x40 /* ECE in this ACK */ | |
102 | #define FLAG_DATA_LOST 0x80 /* SACK detected data lossage. */ | |
103 | #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ | |
4dc2665e | 104 | #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */ |
2e605294 | 105 | #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ |
564262c1 | 106 | #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ |
009a2e3e | 107 | #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */ |
cadbd031 | 108 | #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ |
1da177e4 LT |
109 | |
110 | #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) | |
111 | #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) | |
112 | #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) | |
113 | #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) | |
2e605294 | 114 | #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED) |
1da177e4 | 115 | |
4dc2665e IJ |
116 | #define IsSackFrto() (sysctl_tcp_frto == 0x2) |
117 | ||
1da177e4 | 118 | #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) |
bdf1ee5d | 119 | #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) |
1da177e4 | 120 | |
e905a9ed | 121 | /* Adapt the MSS value used to make delayed ack decision to the |
1da177e4 | 122 | * real world. |
e905a9ed | 123 | */ |
056834d9 | 124 | static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 125 | { |
463c84b9 | 126 | struct inet_connection_sock *icsk = inet_csk(sk); |
e905a9ed | 127 | const unsigned int lss = icsk->icsk_ack.last_seg_size; |
463c84b9 | 128 | unsigned int len; |
1da177e4 | 129 | |
e905a9ed | 130 | icsk->icsk_ack.last_seg_size = 0; |
1da177e4 LT |
131 | |
132 | /* skb->len may jitter because of SACKs, even if peer | |
133 | * sends good full-sized frames. | |
134 | */ | |
056834d9 | 135 | len = skb_shinfo(skb)->gso_size ? : skb->len; |
463c84b9 ACM |
136 | if (len >= icsk->icsk_ack.rcv_mss) { |
137 | icsk->icsk_ack.rcv_mss = len; | |
1da177e4 LT |
138 | } else { |
139 | /* Otherwise, we make more careful check taking into account, | |
140 | * that SACKs block is variable. | |
141 | * | |
142 | * "len" is invariant segment length, including TCP header. | |
143 | */ | |
9c70220b | 144 | len += skb->data - skb_transport_header(skb); |
1da177e4 LT |
145 | if (len >= TCP_MIN_RCVMSS + sizeof(struct tcphdr) || |
146 | /* If PSH is not set, packet should be | |
147 | * full sized, provided peer TCP is not badly broken. | |
148 | * This observation (if it is correct 8)) allows | |
149 | * to handle super-low mtu links fairly. | |
150 | */ | |
151 | (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && | |
aa8223c7 | 152 | !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { |
1da177e4 LT |
153 | /* Subtract also invariant (if peer is RFC compliant), |
154 | * tcp header plus fixed timestamp option length. | |
155 | * Resulting "len" is MSS free of SACK jitter. | |
156 | */ | |
463c84b9 ACM |
157 | len -= tcp_sk(sk)->tcp_header_len; |
158 | icsk->icsk_ack.last_seg_size = len; | |
1da177e4 | 159 | if (len == lss) { |
463c84b9 | 160 | icsk->icsk_ack.rcv_mss = len; |
1da177e4 LT |
161 | return; |
162 | } | |
163 | } | |
1ef9696c AK |
164 | if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) |
165 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; | |
463c84b9 | 166 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; |
1da177e4 LT |
167 | } |
168 | } | |
169 | ||
463c84b9 | 170 | static void tcp_incr_quickack(struct sock *sk) |
1da177e4 | 171 | { |
463c84b9 ACM |
172 | struct inet_connection_sock *icsk = inet_csk(sk); |
173 | unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); | |
1da177e4 | 174 | |
056834d9 IJ |
175 | if (quickacks == 0) |
176 | quickacks = 2; | |
463c84b9 ACM |
177 | if (quickacks > icsk->icsk_ack.quick) |
178 | icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); | |
1da177e4 LT |
179 | } |
180 | ||
463c84b9 | 181 | void tcp_enter_quickack_mode(struct sock *sk) |
1da177e4 | 182 | { |
463c84b9 ACM |
183 | struct inet_connection_sock *icsk = inet_csk(sk); |
184 | tcp_incr_quickack(sk); | |
185 | icsk->icsk_ack.pingpong = 0; | |
186 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 LT |
187 | } |
188 | ||
189 | /* Send ACKs quickly, if "quick" count is not exhausted | |
190 | * and the session is not interactive. | |
191 | */ | |
192 | ||
463c84b9 | 193 | static inline int tcp_in_quickack_mode(const struct sock *sk) |
1da177e4 | 194 | { |
463c84b9 ACM |
195 | const struct inet_connection_sock *icsk = inet_csk(sk); |
196 | return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; | |
1da177e4 LT |
197 | } |
198 | ||
bdf1ee5d IJ |
199 | static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp) |
200 | { | |
056834d9 | 201 | if (tp->ecn_flags & TCP_ECN_OK) |
bdf1ee5d IJ |
202 | tp->ecn_flags |= TCP_ECN_QUEUE_CWR; |
203 | } | |
204 | ||
205 | static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, struct sk_buff *skb) | |
206 | { | |
207 | if (tcp_hdr(skb)->cwr) | |
208 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
209 | } | |
210 | ||
211 | static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp) | |
212 | { | |
213 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
214 | } | |
215 | ||
216 | static inline void TCP_ECN_check_ce(struct tcp_sock *tp, struct sk_buff *skb) | |
217 | { | |
056834d9 | 218 | if (tp->ecn_flags & TCP_ECN_OK) { |
bdf1ee5d IJ |
219 | if (INET_ECN_is_ce(TCP_SKB_CB(skb)->flags)) |
220 | tp->ecn_flags |= TCP_ECN_DEMAND_CWR; | |
221 | /* Funny extension: if ECT is not set on a segment, | |
222 | * it is surely retransmit. It is not in ECN RFC, | |
223 | * but Linux follows this rule. */ | |
224 | else if (INET_ECN_is_not_ect((TCP_SKB_CB(skb)->flags))) | |
225 | tcp_enter_quickack_mode((struct sock *)tp); | |
226 | } | |
227 | } | |
228 | ||
229 | static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, struct tcphdr *th) | |
230 | { | |
056834d9 | 231 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) |
bdf1ee5d IJ |
232 | tp->ecn_flags &= ~TCP_ECN_OK; |
233 | } | |
234 | ||
235 | static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, struct tcphdr *th) | |
236 | { | |
056834d9 | 237 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) |
bdf1ee5d IJ |
238 | tp->ecn_flags &= ~TCP_ECN_OK; |
239 | } | |
240 | ||
241 | static inline int TCP_ECN_rcv_ecn_echo(struct tcp_sock *tp, struct tcphdr *th) | |
242 | { | |
056834d9 | 243 | if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) |
bdf1ee5d IJ |
244 | return 1; |
245 | return 0; | |
246 | } | |
247 | ||
1da177e4 LT |
248 | /* Buffer size and advertised window tuning. |
249 | * | |
250 | * 1. Tuning sk->sk_sndbuf, when connection enters established state. | |
251 | */ | |
252 | ||
253 | static void tcp_fixup_sndbuf(struct sock *sk) | |
254 | { | |
255 | int sndmem = tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER + 16 + | |
256 | sizeof(struct sk_buff); | |
257 | ||
258 | if (sk->sk_sndbuf < 3 * sndmem) | |
259 | sk->sk_sndbuf = min(3 * sndmem, sysctl_tcp_wmem[2]); | |
260 | } | |
261 | ||
262 | /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) | |
263 | * | |
264 | * All tcp_full_space() is split to two parts: "network" buffer, allocated | |
265 | * forward and advertised in receiver window (tp->rcv_wnd) and | |
266 | * "application buffer", required to isolate scheduling/application | |
267 | * latencies from network. | |
268 | * window_clamp is maximal advertised window. It can be less than | |
269 | * tcp_full_space(), in this case tcp_full_space() - window_clamp | |
270 | * is reserved for "application" buffer. The less window_clamp is | |
271 | * the smoother our behaviour from viewpoint of network, but the lower | |
272 | * throughput and the higher sensitivity of the connection to losses. 8) | |
273 | * | |
274 | * rcv_ssthresh is more strict window_clamp used at "slow start" | |
275 | * phase to predict further behaviour of this connection. | |
276 | * It is used for two goals: | |
277 | * - to enforce header prediction at sender, even when application | |
278 | * requires some significant "application buffer". It is check #1. | |
279 | * - to prevent pruning of receive queue because of misprediction | |
280 | * of receiver window. Check #2. | |
281 | * | |
282 | * The scheme does not work when sender sends good segments opening | |
caa20d9a | 283 | * window and then starts to feed us spaghetti. But it should work |
1da177e4 LT |
284 | * in common situations. Otherwise, we have to rely on queue collapsing. |
285 | */ | |
286 | ||
287 | /* Slow part of check#2. */ | |
9e412ba7 | 288 | static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 289 | { |
9e412ba7 | 290 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 291 | /* Optimize this! */ |
dfd4f0ae ED |
292 | int truesize = tcp_win_from_space(skb->truesize) >> 1; |
293 | int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1; | |
1da177e4 LT |
294 | |
295 | while (tp->rcv_ssthresh <= window) { | |
296 | if (truesize <= skb->len) | |
463c84b9 | 297 | return 2 * inet_csk(sk)->icsk_ack.rcv_mss; |
1da177e4 LT |
298 | |
299 | truesize >>= 1; | |
300 | window >>= 1; | |
301 | } | |
302 | return 0; | |
303 | } | |
304 | ||
056834d9 | 305 | static void tcp_grow_window(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 306 | { |
9e412ba7 IJ |
307 | struct tcp_sock *tp = tcp_sk(sk); |
308 | ||
1da177e4 LT |
309 | /* Check #1 */ |
310 | if (tp->rcv_ssthresh < tp->window_clamp && | |
311 | (int)tp->rcv_ssthresh < tcp_space(sk) && | |
312 | !tcp_memory_pressure) { | |
313 | int incr; | |
314 | ||
315 | /* Check #2. Increase window, if skb with such overhead | |
316 | * will fit to rcvbuf in future. | |
317 | */ | |
318 | if (tcp_win_from_space(skb->truesize) <= skb->len) | |
056834d9 | 319 | incr = 2 * tp->advmss; |
1da177e4 | 320 | else |
9e412ba7 | 321 | incr = __tcp_grow_window(sk, skb); |
1da177e4 LT |
322 | |
323 | if (incr) { | |
056834d9 IJ |
324 | tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, |
325 | tp->window_clamp); | |
463c84b9 | 326 | inet_csk(sk)->icsk_ack.quick |= 1; |
1da177e4 LT |
327 | } |
328 | } | |
329 | } | |
330 | ||
331 | /* 3. Tuning rcvbuf, when connection enters established state. */ | |
332 | ||
333 | static void tcp_fixup_rcvbuf(struct sock *sk) | |
334 | { | |
335 | struct tcp_sock *tp = tcp_sk(sk); | |
336 | int rcvmem = tp->advmss + MAX_TCP_HEADER + 16 + sizeof(struct sk_buff); | |
337 | ||
338 | /* Try to select rcvbuf so that 4 mss-sized segments | |
caa20d9a | 339 | * will fit to window and corresponding skbs will fit to our rcvbuf. |
1da177e4 LT |
340 | * (was 3; 4 is minimum to allow fast retransmit to work.) |
341 | */ | |
342 | while (tcp_win_from_space(rcvmem) < tp->advmss) | |
343 | rcvmem += 128; | |
344 | if (sk->sk_rcvbuf < 4 * rcvmem) | |
345 | sk->sk_rcvbuf = min(4 * rcvmem, sysctl_tcp_rmem[2]); | |
346 | } | |
347 | ||
caa20d9a | 348 | /* 4. Try to fixup all. It is made immediately after connection enters |
1da177e4 LT |
349 | * established state. |
350 | */ | |
351 | static void tcp_init_buffer_space(struct sock *sk) | |
352 | { | |
353 | struct tcp_sock *tp = tcp_sk(sk); | |
354 | int maxwin; | |
355 | ||
356 | if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) | |
357 | tcp_fixup_rcvbuf(sk); | |
358 | if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) | |
359 | tcp_fixup_sndbuf(sk); | |
360 | ||
361 | tp->rcvq_space.space = tp->rcv_wnd; | |
362 | ||
363 | maxwin = tcp_full_space(sk); | |
364 | ||
365 | if (tp->window_clamp >= maxwin) { | |
366 | tp->window_clamp = maxwin; | |
367 | ||
368 | if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) | |
369 | tp->window_clamp = max(maxwin - | |
370 | (maxwin >> sysctl_tcp_app_win), | |
371 | 4 * tp->advmss); | |
372 | } | |
373 | ||
374 | /* Force reservation of one segment. */ | |
375 | if (sysctl_tcp_app_win && | |
376 | tp->window_clamp > 2 * tp->advmss && | |
377 | tp->window_clamp + tp->advmss > maxwin) | |
378 | tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); | |
379 | ||
380 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); | |
381 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
382 | } | |
383 | ||
1da177e4 | 384 | /* 5. Recalculate window clamp after socket hit its memory bounds. */ |
9e412ba7 | 385 | static void tcp_clamp_window(struct sock *sk) |
1da177e4 | 386 | { |
9e412ba7 | 387 | struct tcp_sock *tp = tcp_sk(sk); |
6687e988 | 388 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 389 | |
6687e988 | 390 | icsk->icsk_ack.quick = 0; |
1da177e4 | 391 | |
326f36e9 JH |
392 | if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && |
393 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && | |
394 | !tcp_memory_pressure && | |
395 | atomic_read(&tcp_memory_allocated) < sysctl_tcp_mem[0]) { | |
396 | sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), | |
397 | sysctl_tcp_rmem[2]); | |
1da177e4 | 398 | } |
326f36e9 | 399 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) |
056834d9 | 400 | tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); |
1da177e4 LT |
401 | } |
402 | ||
40efc6fa SH |
403 | /* Initialize RCV_MSS value. |
404 | * RCV_MSS is an our guess about MSS used by the peer. | |
405 | * We haven't any direct information about the MSS. | |
406 | * It's better to underestimate the RCV_MSS rather than overestimate. | |
407 | * Overestimations make us ACKing less frequently than needed. | |
408 | * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). | |
409 | */ | |
410 | void tcp_initialize_rcv_mss(struct sock *sk) | |
411 | { | |
412 | struct tcp_sock *tp = tcp_sk(sk); | |
413 | unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); | |
414 | ||
056834d9 | 415 | hint = min(hint, tp->rcv_wnd / 2); |
40efc6fa SH |
416 | hint = min(hint, TCP_MIN_RCVMSS); |
417 | hint = max(hint, TCP_MIN_MSS); | |
418 | ||
419 | inet_csk(sk)->icsk_ack.rcv_mss = hint; | |
420 | } | |
421 | ||
1da177e4 LT |
422 | /* Receiver "autotuning" code. |
423 | * | |
424 | * The algorithm for RTT estimation w/o timestamps is based on | |
425 | * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. | |
426 | * <http://www.lanl.gov/radiant/website/pubs/drs/lacsi2001.ps> | |
427 | * | |
428 | * More detail on this code can be found at | |
429 | * <http://www.psc.edu/~jheffner/senior_thesis.ps>, | |
430 | * though this reference is out of date. A new paper | |
431 | * is pending. | |
432 | */ | |
433 | static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) | |
434 | { | |
435 | u32 new_sample = tp->rcv_rtt_est.rtt; | |
436 | long m = sample; | |
437 | ||
438 | if (m == 0) | |
439 | m = 1; | |
440 | ||
441 | if (new_sample != 0) { | |
442 | /* If we sample in larger samples in the non-timestamp | |
443 | * case, we could grossly overestimate the RTT especially | |
444 | * with chatty applications or bulk transfer apps which | |
445 | * are stalled on filesystem I/O. | |
446 | * | |
447 | * Also, since we are only going for a minimum in the | |
31f34269 | 448 | * non-timestamp case, we do not smooth things out |
caa20d9a | 449 | * else with timestamps disabled convergence takes too |
1da177e4 LT |
450 | * long. |
451 | */ | |
452 | if (!win_dep) { | |
453 | m -= (new_sample >> 3); | |
454 | new_sample += m; | |
455 | } else if (m < new_sample) | |
456 | new_sample = m << 3; | |
457 | } else { | |
caa20d9a | 458 | /* No previous measure. */ |
1da177e4 LT |
459 | new_sample = m << 3; |
460 | } | |
461 | ||
462 | if (tp->rcv_rtt_est.rtt != new_sample) | |
463 | tp->rcv_rtt_est.rtt = new_sample; | |
464 | } | |
465 | ||
466 | static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) | |
467 | { | |
468 | if (tp->rcv_rtt_est.time == 0) | |
469 | goto new_measure; | |
470 | if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) | |
471 | return; | |
056834d9 | 472 | tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1); |
1da177e4 LT |
473 | |
474 | new_measure: | |
475 | tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; | |
476 | tp->rcv_rtt_est.time = tcp_time_stamp; | |
477 | } | |
478 | ||
056834d9 IJ |
479 | static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, |
480 | const struct sk_buff *skb) | |
1da177e4 | 481 | { |
463c84b9 | 482 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
483 | if (tp->rx_opt.rcv_tsecr && |
484 | (TCP_SKB_CB(skb)->end_seq - | |
463c84b9 | 485 | TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) |
1da177e4 LT |
486 | tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); |
487 | } | |
488 | ||
489 | /* | |
490 | * This function should be called every time data is copied to user space. | |
491 | * It calculates the appropriate TCP receive buffer space. | |
492 | */ | |
493 | void tcp_rcv_space_adjust(struct sock *sk) | |
494 | { | |
495 | struct tcp_sock *tp = tcp_sk(sk); | |
496 | int time; | |
497 | int space; | |
e905a9ed | 498 | |
1da177e4 LT |
499 | if (tp->rcvq_space.time == 0) |
500 | goto new_measure; | |
e905a9ed | 501 | |
1da177e4 | 502 | time = tcp_time_stamp - tp->rcvq_space.time; |
056834d9 | 503 | if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0) |
1da177e4 | 504 | return; |
e905a9ed | 505 | |
1da177e4 LT |
506 | space = 2 * (tp->copied_seq - tp->rcvq_space.seq); |
507 | ||
508 | space = max(tp->rcvq_space.space, space); | |
509 | ||
510 | if (tp->rcvq_space.space != space) { | |
511 | int rcvmem; | |
512 | ||
513 | tp->rcvq_space.space = space; | |
514 | ||
6fcf9412 JH |
515 | if (sysctl_tcp_moderate_rcvbuf && |
516 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { | |
1da177e4 LT |
517 | int new_clamp = space; |
518 | ||
519 | /* Receive space grows, normalize in order to | |
520 | * take into account packet headers and sk_buff | |
521 | * structure overhead. | |
522 | */ | |
523 | space /= tp->advmss; | |
524 | if (!space) | |
525 | space = 1; | |
526 | rcvmem = (tp->advmss + MAX_TCP_HEADER + | |
527 | 16 + sizeof(struct sk_buff)); | |
528 | while (tcp_win_from_space(rcvmem) < tp->advmss) | |
529 | rcvmem += 128; | |
530 | space *= rcvmem; | |
531 | space = min(space, sysctl_tcp_rmem[2]); | |
532 | if (space > sk->sk_rcvbuf) { | |
533 | sk->sk_rcvbuf = space; | |
534 | ||
535 | /* Make the window clamp follow along. */ | |
536 | tp->window_clamp = new_clamp; | |
537 | } | |
538 | } | |
539 | } | |
e905a9ed | 540 | |
1da177e4 LT |
541 | new_measure: |
542 | tp->rcvq_space.seq = tp->copied_seq; | |
543 | tp->rcvq_space.time = tcp_time_stamp; | |
544 | } | |
545 | ||
546 | /* There is something which you must keep in mind when you analyze the | |
547 | * behavior of the tp->ato delayed ack timeout interval. When a | |
548 | * connection starts up, we want to ack as quickly as possible. The | |
549 | * problem is that "good" TCP's do slow start at the beginning of data | |
550 | * transmission. The means that until we send the first few ACK's the | |
551 | * sender will sit on his end and only queue most of his data, because | |
552 | * he can only send snd_cwnd unacked packets at any given time. For | |
553 | * each ACK we send, he increments snd_cwnd and transmits more of his | |
554 | * queue. -DaveM | |
555 | */ | |
9e412ba7 | 556 | static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 557 | { |
9e412ba7 | 558 | struct tcp_sock *tp = tcp_sk(sk); |
463c84b9 | 559 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
560 | u32 now; |
561 | ||
463c84b9 | 562 | inet_csk_schedule_ack(sk); |
1da177e4 | 563 | |
463c84b9 | 564 | tcp_measure_rcv_mss(sk, skb); |
1da177e4 LT |
565 | |
566 | tcp_rcv_rtt_measure(tp); | |
e905a9ed | 567 | |
1da177e4 LT |
568 | now = tcp_time_stamp; |
569 | ||
463c84b9 | 570 | if (!icsk->icsk_ack.ato) { |
1da177e4 LT |
571 | /* The _first_ data packet received, initialize |
572 | * delayed ACK engine. | |
573 | */ | |
463c84b9 ACM |
574 | tcp_incr_quickack(sk); |
575 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 | 576 | } else { |
463c84b9 | 577 | int m = now - icsk->icsk_ack.lrcvtime; |
1da177e4 | 578 | |
056834d9 | 579 | if (m <= TCP_ATO_MIN / 2) { |
1da177e4 | 580 | /* The fastest case is the first. */ |
463c84b9 ACM |
581 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; |
582 | } else if (m < icsk->icsk_ack.ato) { | |
583 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; | |
584 | if (icsk->icsk_ack.ato > icsk->icsk_rto) | |
585 | icsk->icsk_ack.ato = icsk->icsk_rto; | |
586 | } else if (m > icsk->icsk_rto) { | |
caa20d9a | 587 | /* Too long gap. Apparently sender failed to |
1da177e4 LT |
588 | * restart window, so that we send ACKs quickly. |
589 | */ | |
463c84b9 | 590 | tcp_incr_quickack(sk); |
3ab224be | 591 | sk_mem_reclaim(sk); |
1da177e4 LT |
592 | } |
593 | } | |
463c84b9 | 594 | icsk->icsk_ack.lrcvtime = now; |
1da177e4 LT |
595 | |
596 | TCP_ECN_check_ce(tp, skb); | |
597 | ||
598 | if (skb->len >= 128) | |
9e412ba7 | 599 | tcp_grow_window(sk, skb); |
1da177e4 LT |
600 | } |
601 | ||
05bb1fad DM |
602 | static u32 tcp_rto_min(struct sock *sk) |
603 | { | |
604 | struct dst_entry *dst = __sk_dst_get(sk); | |
605 | u32 rto_min = TCP_RTO_MIN; | |
606 | ||
5c127c58 | 607 | if (dst && dst_metric_locked(dst, RTAX_RTO_MIN)) |
056834d9 | 608 | rto_min = dst->metrics[RTAX_RTO_MIN - 1]; |
05bb1fad DM |
609 | return rto_min; |
610 | } | |
611 | ||
1da177e4 LT |
612 | /* Called to compute a smoothed rtt estimate. The data fed to this |
613 | * routine either comes from timestamps, or from segments that were | |
614 | * known _not_ to have been retransmitted [see Karn/Partridge | |
615 | * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 | |
616 | * piece by Van Jacobson. | |
617 | * NOTE: the next three routines used to be one big routine. | |
618 | * To save cycles in the RFC 1323 implementation it was better to break | |
619 | * it up into three procedures. -- erics | |
620 | */ | |
2d2abbab | 621 | static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) |
1da177e4 | 622 | { |
6687e988 | 623 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
624 | long m = mrtt; /* RTT */ |
625 | ||
1da177e4 LT |
626 | /* The following amusing code comes from Jacobson's |
627 | * article in SIGCOMM '88. Note that rtt and mdev | |
628 | * are scaled versions of rtt and mean deviation. | |
e905a9ed | 629 | * This is designed to be as fast as possible |
1da177e4 LT |
630 | * m stands for "measurement". |
631 | * | |
632 | * On a 1990 paper the rto value is changed to: | |
633 | * RTO = rtt + 4 * mdev | |
634 | * | |
635 | * Funny. This algorithm seems to be very broken. | |
636 | * These formulae increase RTO, when it should be decreased, increase | |
31f34269 | 637 | * too slowly, when it should be increased quickly, decrease too quickly |
1da177e4 LT |
638 | * etc. I guess in BSD RTO takes ONE value, so that it is absolutely |
639 | * does not matter how to _calculate_ it. Seems, it was trap | |
640 | * that VJ failed to avoid. 8) | |
641 | */ | |
2de979bd | 642 | if (m == 0) |
1da177e4 LT |
643 | m = 1; |
644 | if (tp->srtt != 0) { | |
645 | m -= (tp->srtt >> 3); /* m is now error in rtt est */ | |
646 | tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ | |
647 | if (m < 0) { | |
648 | m = -m; /* m is now abs(error) */ | |
649 | m -= (tp->mdev >> 2); /* similar update on mdev */ | |
650 | /* This is similar to one of Eifel findings. | |
651 | * Eifel blocks mdev updates when rtt decreases. | |
652 | * This solution is a bit different: we use finer gain | |
653 | * for mdev in this case (alpha*beta). | |
654 | * Like Eifel it also prevents growth of rto, | |
655 | * but also it limits too fast rto decreases, | |
656 | * happening in pure Eifel. | |
657 | */ | |
658 | if (m > 0) | |
659 | m >>= 3; | |
660 | } else { | |
661 | m -= (tp->mdev >> 2); /* similar update on mdev */ | |
662 | } | |
663 | tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ | |
664 | if (tp->mdev > tp->mdev_max) { | |
665 | tp->mdev_max = tp->mdev; | |
666 | if (tp->mdev_max > tp->rttvar) | |
667 | tp->rttvar = tp->mdev_max; | |
668 | } | |
669 | if (after(tp->snd_una, tp->rtt_seq)) { | |
670 | if (tp->mdev_max < tp->rttvar) | |
056834d9 | 671 | tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2; |
1da177e4 | 672 | tp->rtt_seq = tp->snd_nxt; |
05bb1fad | 673 | tp->mdev_max = tcp_rto_min(sk); |
1da177e4 LT |
674 | } |
675 | } else { | |
676 | /* no previous measure. */ | |
056834d9 IJ |
677 | tp->srtt = m << 3; /* take the measured time to be rtt */ |
678 | tp->mdev = m << 1; /* make sure rto = 3*rtt */ | |
05bb1fad | 679 | tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); |
1da177e4 LT |
680 | tp->rtt_seq = tp->snd_nxt; |
681 | } | |
1da177e4 LT |
682 | } |
683 | ||
684 | /* Calculate rto without backoff. This is the second half of Van Jacobson's | |
685 | * routine referred to above. | |
686 | */ | |
463c84b9 | 687 | static inline void tcp_set_rto(struct sock *sk) |
1da177e4 | 688 | { |
463c84b9 | 689 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
690 | /* Old crap is replaced with new one. 8) |
691 | * | |
692 | * More seriously: | |
693 | * 1. If rtt variance happened to be less 50msec, it is hallucination. | |
694 | * It cannot be less due to utterly erratic ACK generation made | |
695 | * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ | |
696 | * to do with delayed acks, because at cwnd>2 true delack timeout | |
697 | * is invisible. Actually, Linux-2.4 also generates erratic | |
caa20d9a | 698 | * ACKs in some circumstances. |
1da177e4 | 699 | */ |
463c84b9 | 700 | inet_csk(sk)->icsk_rto = (tp->srtt >> 3) + tp->rttvar; |
1da177e4 LT |
701 | |
702 | /* 2. Fixups made earlier cannot be right. | |
703 | * If we do not estimate RTO correctly without them, | |
704 | * all the algo is pure shit and should be replaced | |
caa20d9a | 705 | * with correct one. It is exactly, which we pretend to do. |
1da177e4 LT |
706 | */ |
707 | } | |
708 | ||
709 | /* NOTE: clamping at TCP_RTO_MIN is not required, current algo | |
710 | * guarantees that rto is higher. | |
711 | */ | |
463c84b9 | 712 | static inline void tcp_bound_rto(struct sock *sk) |
1da177e4 | 713 | { |
463c84b9 ACM |
714 | if (inet_csk(sk)->icsk_rto > TCP_RTO_MAX) |
715 | inet_csk(sk)->icsk_rto = TCP_RTO_MAX; | |
1da177e4 LT |
716 | } |
717 | ||
718 | /* Save metrics learned by this TCP session. | |
719 | This function is called only, when TCP finishes successfully | |
720 | i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. | |
721 | */ | |
722 | void tcp_update_metrics(struct sock *sk) | |
723 | { | |
724 | struct tcp_sock *tp = tcp_sk(sk); | |
725 | struct dst_entry *dst = __sk_dst_get(sk); | |
726 | ||
727 | if (sysctl_tcp_nometrics_save) | |
728 | return; | |
729 | ||
730 | dst_confirm(dst); | |
731 | ||
056834d9 | 732 | if (dst && (dst->flags & DST_HOST)) { |
6687e988 | 733 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
734 | int m; |
735 | ||
6687e988 | 736 | if (icsk->icsk_backoff || !tp->srtt) { |
1da177e4 LT |
737 | /* This session failed to estimate rtt. Why? |
738 | * Probably, no packets returned in time. | |
739 | * Reset our results. | |
740 | */ | |
741 | if (!(dst_metric_locked(dst, RTAX_RTT))) | |
056834d9 | 742 | dst->metrics[RTAX_RTT - 1] = 0; |
1da177e4 LT |
743 | return; |
744 | } | |
745 | ||
746 | m = dst_metric(dst, RTAX_RTT) - tp->srtt; | |
747 | ||
748 | /* If newly calculated rtt larger than stored one, | |
749 | * store new one. Otherwise, use EWMA. Remember, | |
750 | * rtt overestimation is always better than underestimation. | |
751 | */ | |
752 | if (!(dst_metric_locked(dst, RTAX_RTT))) { | |
753 | if (m <= 0) | |
056834d9 | 754 | dst->metrics[RTAX_RTT - 1] = tp->srtt; |
1da177e4 | 755 | else |
056834d9 | 756 | dst->metrics[RTAX_RTT - 1] -= (m >> 3); |
1da177e4 LT |
757 | } |
758 | ||
759 | if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { | |
760 | if (m < 0) | |
761 | m = -m; | |
762 | ||
763 | /* Scale deviation to rttvar fixed point */ | |
764 | m >>= 1; | |
765 | if (m < tp->mdev) | |
766 | m = tp->mdev; | |
767 | ||
768 | if (m >= dst_metric(dst, RTAX_RTTVAR)) | |
056834d9 | 769 | dst->metrics[RTAX_RTTVAR - 1] = m; |
1da177e4 LT |
770 | else |
771 | dst->metrics[RTAX_RTTVAR-1] -= | |
772 | (dst->metrics[RTAX_RTTVAR-1] - m)>>2; | |
773 | } | |
774 | ||
775 | if (tp->snd_ssthresh >= 0xFFFF) { | |
776 | /* Slow start still did not finish. */ | |
777 | if (dst_metric(dst, RTAX_SSTHRESH) && | |
778 | !dst_metric_locked(dst, RTAX_SSTHRESH) && | |
779 | (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) | |
780 | dst->metrics[RTAX_SSTHRESH-1] = tp->snd_cwnd >> 1; | |
781 | if (!dst_metric_locked(dst, RTAX_CWND) && | |
782 | tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) | |
056834d9 | 783 | dst->metrics[RTAX_CWND - 1] = tp->snd_cwnd; |
1da177e4 | 784 | } else if (tp->snd_cwnd > tp->snd_ssthresh && |
6687e988 | 785 | icsk->icsk_ca_state == TCP_CA_Open) { |
1da177e4 LT |
786 | /* Cong. avoidance phase, cwnd is reliable. */ |
787 | if (!dst_metric_locked(dst, RTAX_SSTHRESH)) | |
788 | dst->metrics[RTAX_SSTHRESH-1] = | |
789 | max(tp->snd_cwnd >> 1, tp->snd_ssthresh); | |
790 | if (!dst_metric_locked(dst, RTAX_CWND)) | |
791 | dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_cwnd) >> 1; | |
792 | } else { | |
793 | /* Else slow start did not finish, cwnd is non-sense, | |
794 | ssthresh may be also invalid. | |
795 | */ | |
796 | if (!dst_metric_locked(dst, RTAX_CWND)) | |
797 | dst->metrics[RTAX_CWND-1] = (dst->metrics[RTAX_CWND-1] + tp->snd_ssthresh) >> 1; | |
798 | if (dst->metrics[RTAX_SSTHRESH-1] && | |
799 | !dst_metric_locked(dst, RTAX_SSTHRESH) && | |
800 | tp->snd_ssthresh > dst->metrics[RTAX_SSTHRESH-1]) | |
801 | dst->metrics[RTAX_SSTHRESH-1] = tp->snd_ssthresh; | |
802 | } | |
803 | ||
804 | if (!dst_metric_locked(dst, RTAX_REORDERING)) { | |
805 | if (dst->metrics[RTAX_REORDERING-1] < tp->reordering && | |
806 | tp->reordering != sysctl_tcp_reordering) | |
807 | dst->metrics[RTAX_REORDERING-1] = tp->reordering; | |
808 | } | |
809 | } | |
810 | } | |
811 | ||
26722873 DM |
812 | /* Numbers are taken from RFC3390. |
813 | * | |
814 | * John Heffner states: | |
815 | * | |
816 | * The RFC specifies a window of no more than 4380 bytes | |
817 | * unless 2*MSS > 4380. Reading the pseudocode in the RFC | |
818 | * is a bit misleading because they use a clamp at 4380 bytes | |
819 | * rather than use a multiplier in the relevant range. | |
820 | */ | |
1da177e4 LT |
821 | __u32 tcp_init_cwnd(struct tcp_sock *tp, struct dst_entry *dst) |
822 | { | |
823 | __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); | |
824 | ||
825 | if (!cwnd) { | |
c1b4a7e6 | 826 | if (tp->mss_cache > 1460) |
1da177e4 LT |
827 | cwnd = 2; |
828 | else | |
c1b4a7e6 | 829 | cwnd = (tp->mss_cache > 1095) ? 3 : 4; |
1da177e4 LT |
830 | } |
831 | return min_t(__u32, cwnd, tp->snd_cwnd_clamp); | |
832 | } | |
833 | ||
40efc6fa | 834 | /* Set slow start threshold and cwnd not falling to slow start */ |
3cfe3baa | 835 | void tcp_enter_cwr(struct sock *sk, const int set_ssthresh) |
40efc6fa SH |
836 | { |
837 | struct tcp_sock *tp = tcp_sk(sk); | |
3cfe3baa | 838 | const struct inet_connection_sock *icsk = inet_csk(sk); |
40efc6fa SH |
839 | |
840 | tp->prior_ssthresh = 0; | |
841 | tp->bytes_acked = 0; | |
e01f9d77 | 842 | if (icsk->icsk_ca_state < TCP_CA_CWR) { |
40efc6fa | 843 | tp->undo_marker = 0; |
3cfe3baa IJ |
844 | if (set_ssthresh) |
845 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
40efc6fa SH |
846 | tp->snd_cwnd = min(tp->snd_cwnd, |
847 | tcp_packets_in_flight(tp) + 1U); | |
848 | tp->snd_cwnd_cnt = 0; | |
849 | tp->high_seq = tp->snd_nxt; | |
850 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
851 | TCP_ECN_queue_cwr(tp); | |
852 | ||
853 | tcp_set_ca_state(sk, TCP_CA_CWR); | |
854 | } | |
855 | } | |
856 | ||
e60402d0 IJ |
857 | /* |
858 | * Packet counting of FACK is based on in-order assumptions, therefore TCP | |
859 | * disables it when reordering is detected | |
860 | */ | |
861 | static void tcp_disable_fack(struct tcp_sock *tp) | |
862 | { | |
85cc391c IJ |
863 | /* RFC3517 uses different metric in lost marker => reset on change */ |
864 | if (tcp_is_fack(tp)) | |
865 | tp->lost_skb_hint = NULL; | |
e60402d0 IJ |
866 | tp->rx_opt.sack_ok &= ~2; |
867 | } | |
868 | ||
564262c1 | 869 | /* Take a notice that peer is sending D-SACKs */ |
e60402d0 IJ |
870 | static void tcp_dsack_seen(struct tcp_sock *tp) |
871 | { | |
872 | tp->rx_opt.sack_ok |= 4; | |
873 | } | |
874 | ||
1da177e4 LT |
875 | /* Initialize metrics on socket. */ |
876 | ||
877 | static void tcp_init_metrics(struct sock *sk) | |
878 | { | |
879 | struct tcp_sock *tp = tcp_sk(sk); | |
880 | struct dst_entry *dst = __sk_dst_get(sk); | |
881 | ||
882 | if (dst == NULL) | |
883 | goto reset; | |
884 | ||
885 | dst_confirm(dst); | |
886 | ||
887 | if (dst_metric_locked(dst, RTAX_CWND)) | |
888 | tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); | |
889 | if (dst_metric(dst, RTAX_SSTHRESH)) { | |
890 | tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); | |
891 | if (tp->snd_ssthresh > tp->snd_cwnd_clamp) | |
892 | tp->snd_ssthresh = tp->snd_cwnd_clamp; | |
893 | } | |
894 | if (dst_metric(dst, RTAX_REORDERING) && | |
895 | tp->reordering != dst_metric(dst, RTAX_REORDERING)) { | |
e60402d0 | 896 | tcp_disable_fack(tp); |
1da177e4 LT |
897 | tp->reordering = dst_metric(dst, RTAX_REORDERING); |
898 | } | |
899 | ||
900 | if (dst_metric(dst, RTAX_RTT) == 0) | |
901 | goto reset; | |
902 | ||
903 | if (!tp->srtt && dst_metric(dst, RTAX_RTT) < (TCP_TIMEOUT_INIT << 3)) | |
904 | goto reset; | |
905 | ||
906 | /* Initial rtt is determined from SYN,SYN-ACK. | |
907 | * The segment is small and rtt may appear much | |
908 | * less than real one. Use per-dst memory | |
909 | * to make it more realistic. | |
910 | * | |
911 | * A bit of theory. RTT is time passed after "normal" sized packet | |
caa20d9a | 912 | * is sent until it is ACKed. In normal circumstances sending small |
1da177e4 LT |
913 | * packets force peer to delay ACKs and calculation is correct too. |
914 | * The algorithm is adaptive and, provided we follow specs, it | |
915 | * NEVER underestimate RTT. BUT! If peer tries to make some clever | |
916 | * tricks sort of "quick acks" for time long enough to decrease RTT | |
917 | * to low value, and then abruptly stops to do it and starts to delay | |
918 | * ACKs, wait for troubles. | |
919 | */ | |
920 | if (dst_metric(dst, RTAX_RTT) > tp->srtt) { | |
921 | tp->srtt = dst_metric(dst, RTAX_RTT); | |
922 | tp->rtt_seq = tp->snd_nxt; | |
923 | } | |
924 | if (dst_metric(dst, RTAX_RTTVAR) > tp->mdev) { | |
925 | tp->mdev = dst_metric(dst, RTAX_RTTVAR); | |
488faa2a | 926 | tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); |
1da177e4 | 927 | } |
463c84b9 ACM |
928 | tcp_set_rto(sk); |
929 | tcp_bound_rto(sk); | |
930 | if (inet_csk(sk)->icsk_rto < TCP_TIMEOUT_INIT && !tp->rx_opt.saw_tstamp) | |
1da177e4 LT |
931 | goto reset; |
932 | tp->snd_cwnd = tcp_init_cwnd(tp, dst); | |
933 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
934 | return; | |
935 | ||
936 | reset: | |
937 | /* Play conservative. If timestamps are not | |
938 | * supported, TCP will fail to recalculate correct | |
939 | * rtt, if initial rto is too small. FORGET ALL AND RESET! | |
940 | */ | |
941 | if (!tp->rx_opt.saw_tstamp && tp->srtt) { | |
942 | tp->srtt = 0; | |
943 | tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_INIT; | |
463c84b9 | 944 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
1da177e4 LT |
945 | } |
946 | } | |
947 | ||
6687e988 ACM |
948 | static void tcp_update_reordering(struct sock *sk, const int metric, |
949 | const int ts) | |
1da177e4 | 950 | { |
6687e988 | 951 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
952 | if (metric > tp->reordering) { |
953 | tp->reordering = min(TCP_MAX_REORDERING, metric); | |
954 | ||
955 | /* This exciting event is worth to be remembered. 8) */ | |
956 | if (ts) | |
957 | NET_INC_STATS_BH(LINUX_MIB_TCPTSREORDER); | |
e60402d0 | 958 | else if (tcp_is_reno(tp)) |
1da177e4 | 959 | NET_INC_STATS_BH(LINUX_MIB_TCPRENOREORDER); |
e60402d0 | 960 | else if (tcp_is_fack(tp)) |
1da177e4 LT |
961 | NET_INC_STATS_BH(LINUX_MIB_TCPFACKREORDER); |
962 | else | |
963 | NET_INC_STATS_BH(LINUX_MIB_TCPSACKREORDER); | |
964 | #if FASTRETRANS_DEBUG > 1 | |
965 | printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", | |
6687e988 | 966 | tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, |
1da177e4 LT |
967 | tp->reordering, |
968 | tp->fackets_out, | |
969 | tp->sacked_out, | |
970 | tp->undo_marker ? tp->undo_retrans : 0); | |
971 | #endif | |
e60402d0 | 972 | tcp_disable_fack(tp); |
1da177e4 LT |
973 | } |
974 | } | |
975 | ||
976 | /* This procedure tags the retransmission queue when SACKs arrive. | |
977 | * | |
978 | * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). | |
979 | * Packets in queue with these bits set are counted in variables | |
980 | * sacked_out, retrans_out and lost_out, correspondingly. | |
981 | * | |
982 | * Valid combinations are: | |
983 | * Tag InFlight Description | |
984 | * 0 1 - orig segment is in flight. | |
985 | * S 0 - nothing flies, orig reached receiver. | |
986 | * L 0 - nothing flies, orig lost by net. | |
987 | * R 2 - both orig and retransmit are in flight. | |
988 | * L|R 1 - orig is lost, retransmit is in flight. | |
989 | * S|R 1 - orig reached receiver, retrans is still in flight. | |
990 | * (L|S|R is logically valid, it could occur when L|R is sacked, | |
991 | * but it is equivalent to plain S and code short-curcuits it to S. | |
992 | * L|S is logically invalid, it would mean -1 packet in flight 8)) | |
993 | * | |
994 | * These 6 states form finite state machine, controlled by the following events: | |
995 | * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) | |
996 | * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) | |
997 | * 3. Loss detection event of one of three flavors: | |
998 | * A. Scoreboard estimator decided the packet is lost. | |
999 | * A'. Reno "three dupacks" marks head of queue lost. | |
1000 | * A''. Its FACK modfication, head until snd.fack is lost. | |
1001 | * B. SACK arrives sacking data transmitted after never retransmitted | |
1002 | * hole was sent out. | |
1003 | * C. SACK arrives sacking SND.NXT at the moment, when the | |
1004 | * segment was retransmitted. | |
1005 | * 4. D-SACK added new rule: D-SACK changes any tag to S. | |
1006 | * | |
1007 | * It is pleasant to note, that state diagram turns out to be commutative, | |
1008 | * so that we are allowed not to be bothered by order of our actions, | |
1009 | * when multiple events arrive simultaneously. (see the function below). | |
1010 | * | |
1011 | * Reordering detection. | |
1012 | * -------------------- | |
1013 | * Reordering metric is maximal distance, which a packet can be displaced | |
1014 | * in packet stream. With SACKs we can estimate it: | |
1015 | * | |
1016 | * 1. SACK fills old hole and the corresponding segment was not | |
1017 | * ever retransmitted -> reordering. Alas, we cannot use it | |
1018 | * when segment was retransmitted. | |
1019 | * 2. The last flaw is solved with D-SACK. D-SACK arrives | |
1020 | * for retransmitted and already SACKed segment -> reordering.. | |
1021 | * Both of these heuristics are not used in Loss state, when we cannot | |
1022 | * account for retransmits accurately. | |
5b3c9882 IJ |
1023 | * |
1024 | * SACK block validation. | |
1025 | * ---------------------- | |
1026 | * | |
1027 | * SACK block range validation checks that the received SACK block fits to | |
1028 | * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. | |
1029 | * Note that SND.UNA is not included to the range though being valid because | |
0e835331 IJ |
1030 | * it means that the receiver is rather inconsistent with itself reporting |
1031 | * SACK reneging when it should advance SND.UNA. Such SACK block this is | |
1032 | * perfectly valid, however, in light of RFC2018 which explicitly states | |
1033 | * that "SACK block MUST reflect the newest segment. Even if the newest | |
1034 | * segment is going to be discarded ...", not that it looks very clever | |
1035 | * in case of head skb. Due to potentional receiver driven attacks, we | |
1036 | * choose to avoid immediate execution of a walk in write queue due to | |
1037 | * reneging and defer head skb's loss recovery to standard loss recovery | |
1038 | * procedure that will eventually trigger (nothing forbids us doing this). | |
5b3c9882 IJ |
1039 | * |
1040 | * Implements also blockage to start_seq wrap-around. Problem lies in the | |
1041 | * fact that though start_seq (s) is before end_seq (i.e., not reversed), | |
1042 | * there's no guarantee that it will be before snd_nxt (n). The problem | |
1043 | * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt | |
1044 | * wrap (s_w): | |
1045 | * | |
1046 | * <- outs wnd -> <- wrapzone -> | |
1047 | * u e n u_w e_w s n_w | |
1048 | * | | | | | | | | |
1049 | * |<------------+------+----- TCP seqno space --------------+---------->| | |
1050 | * ...-- <2^31 ->| |<--------... | |
1051 | * ...---- >2^31 ------>| |<--------... | |
1052 | * | |
1053 | * Current code wouldn't be vulnerable but it's better still to discard such | |
1054 | * crazy SACK blocks. Doing this check for start_seq alone closes somewhat | |
1055 | * similar case (end_seq after snd_nxt wrap) as earlier reversed check in | |
1056 | * snd_nxt wrap -> snd_una region will then become "well defined", i.e., | |
1057 | * equal to the ideal case (infinite seqno space without wrap caused issues). | |
1058 | * | |
1059 | * With D-SACK the lower bound is extended to cover sequence space below | |
1060 | * SND.UNA down to undo_marker, which is the last point of interest. Yet | |
564262c1 | 1061 | * again, D-SACK block must not to go across snd_una (for the same reason as |
5b3c9882 IJ |
1062 | * for the normal SACK blocks, explained above). But there all simplicity |
1063 | * ends, TCP might receive valid D-SACKs below that. As long as they reside | |
1064 | * fully below undo_marker they do not affect behavior in anyway and can | |
1065 | * therefore be safely ignored. In rare cases (which are more or less | |
1066 | * theoretical ones), the D-SACK will nicely cross that boundary due to skb | |
1067 | * fragmentation and packet reordering past skb's retransmission. To consider | |
1068 | * them correctly, the acceptable range must be extended even more though | |
1069 | * the exact amount is rather hard to quantify. However, tp->max_window can | |
1070 | * be used as an exaggerated estimate. | |
1da177e4 | 1071 | */ |
5b3c9882 IJ |
1072 | static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack, |
1073 | u32 start_seq, u32 end_seq) | |
1074 | { | |
1075 | /* Too far in future, or reversed (interpretation is ambiguous) */ | |
1076 | if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) | |
1077 | return 0; | |
1078 | ||
1079 | /* Nasty start_seq wrap-around check (see comments above) */ | |
1080 | if (!before(start_seq, tp->snd_nxt)) | |
1081 | return 0; | |
1082 | ||
564262c1 | 1083 | /* In outstanding window? ...This is valid exit for D-SACKs too. |
5b3c9882 IJ |
1084 | * start_seq == snd_una is non-sensical (see comments above) |
1085 | */ | |
1086 | if (after(start_seq, tp->snd_una)) | |
1087 | return 1; | |
1088 | ||
1089 | if (!is_dsack || !tp->undo_marker) | |
1090 | return 0; | |
1091 | ||
1092 | /* ...Then it's D-SACK, and must reside below snd_una completely */ | |
1093 | if (!after(end_seq, tp->snd_una)) | |
1094 | return 0; | |
1095 | ||
1096 | if (!before(start_seq, tp->undo_marker)) | |
1097 | return 1; | |
1098 | ||
1099 | /* Too old */ | |
1100 | if (!after(end_seq, tp->undo_marker)) | |
1101 | return 0; | |
1102 | ||
1103 | /* Undo_marker boundary crossing (overestimates a lot). Known already: | |
1104 | * start_seq < undo_marker and end_seq >= undo_marker. | |
1105 | */ | |
1106 | return !before(start_seq, end_seq - tp->max_window); | |
1107 | } | |
1108 | ||
1c1e87ed IJ |
1109 | /* Check for lost retransmit. This superb idea is borrowed from "ratehalving". |
1110 | * Event "C". Later note: FACK people cheated me again 8), we have to account | |
1111 | * for reordering! Ugly, but should help. | |
f785a8e2 IJ |
1112 | * |
1113 | * Search retransmitted skbs from write_queue that were sent when snd_nxt was | |
1114 | * less than what is now known to be received by the other end (derived from | |
9f58f3b7 IJ |
1115 | * highest SACK block). Also calculate the lowest snd_nxt among the remaining |
1116 | * retransmitted skbs to avoid some costly processing per ACKs. | |
1c1e87ed | 1117 | */ |
407ef1de | 1118 | static void tcp_mark_lost_retrans(struct sock *sk) |
1c1e87ed | 1119 | { |
9f58f3b7 | 1120 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1c1e87ed IJ |
1121 | struct tcp_sock *tp = tcp_sk(sk); |
1122 | struct sk_buff *skb; | |
f785a8e2 | 1123 | int cnt = 0; |
df2e014b | 1124 | u32 new_low_seq = tp->snd_nxt; |
6859d494 | 1125 | u32 received_upto = tcp_highest_sack_seq(tp); |
9f58f3b7 IJ |
1126 | |
1127 | if (!tcp_is_fack(tp) || !tp->retrans_out || | |
1128 | !after(received_upto, tp->lost_retrans_low) || | |
1129 | icsk->icsk_ca_state != TCP_CA_Recovery) | |
407ef1de | 1130 | return; |
1c1e87ed IJ |
1131 | |
1132 | tcp_for_write_queue(skb, sk) { | |
1133 | u32 ack_seq = TCP_SKB_CB(skb)->ack_seq; | |
1134 | ||
1135 | if (skb == tcp_send_head(sk)) | |
1136 | break; | |
f785a8e2 | 1137 | if (cnt == tp->retrans_out) |
1c1e87ed IJ |
1138 | break; |
1139 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1140 | continue; | |
1141 | ||
f785a8e2 IJ |
1142 | if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) |
1143 | continue; | |
1144 | ||
1145 | if (after(received_upto, ack_seq) && | |
1c1e87ed | 1146 | (tcp_is_fack(tp) || |
f785a8e2 | 1147 | !before(received_upto, |
1c1e87ed IJ |
1148 | ack_seq + tp->reordering * tp->mss_cache))) { |
1149 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; | |
1150 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1151 | ||
1152 | /* clear lost hint */ | |
1153 | tp->retransmit_skb_hint = NULL; | |
1154 | ||
1155 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { | |
1156 | tp->lost_out += tcp_skb_pcount(skb); | |
1157 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1c1e87ed | 1158 | } |
bce392f3 | 1159 | NET_INC_STATS_BH(LINUX_MIB_TCPLOSTRETRANSMIT); |
f785a8e2 | 1160 | } else { |
df2e014b | 1161 | if (before(ack_seq, new_low_seq)) |
b08d6cb2 | 1162 | new_low_seq = ack_seq; |
f785a8e2 | 1163 | cnt += tcp_skb_pcount(skb); |
1c1e87ed IJ |
1164 | } |
1165 | } | |
b08d6cb2 IJ |
1166 | |
1167 | if (tp->retrans_out) | |
1168 | tp->lost_retrans_low = new_low_seq; | |
1c1e87ed | 1169 | } |
5b3c9882 | 1170 | |
d06e021d DM |
1171 | static int tcp_check_dsack(struct tcp_sock *tp, struct sk_buff *ack_skb, |
1172 | struct tcp_sack_block_wire *sp, int num_sacks, | |
1173 | u32 prior_snd_una) | |
1174 | { | |
1175 | u32 start_seq_0 = ntohl(get_unaligned(&sp[0].start_seq)); | |
1176 | u32 end_seq_0 = ntohl(get_unaligned(&sp[0].end_seq)); | |
1177 | int dup_sack = 0; | |
1178 | ||
1179 | if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { | |
1180 | dup_sack = 1; | |
e60402d0 | 1181 | tcp_dsack_seen(tp); |
d06e021d DM |
1182 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKRECV); |
1183 | } else if (num_sacks > 1) { | |
1184 | u32 end_seq_1 = ntohl(get_unaligned(&sp[1].end_seq)); | |
1185 | u32 start_seq_1 = ntohl(get_unaligned(&sp[1].start_seq)); | |
1186 | ||
1187 | if (!after(end_seq_0, end_seq_1) && | |
1188 | !before(start_seq_0, start_seq_1)) { | |
1189 | dup_sack = 1; | |
e60402d0 | 1190 | tcp_dsack_seen(tp); |
d06e021d DM |
1191 | NET_INC_STATS_BH(LINUX_MIB_TCPDSACKOFORECV); |
1192 | } | |
1193 | } | |
1194 | ||
1195 | /* D-SACK for already forgotten data... Do dumb counting. */ | |
1196 | if (dup_sack && | |
1197 | !after(end_seq_0, prior_snd_una) && | |
1198 | after(end_seq_0, tp->undo_marker)) | |
1199 | tp->undo_retrans--; | |
1200 | ||
1201 | return dup_sack; | |
1202 | } | |
1203 | ||
d1935942 IJ |
1204 | /* Check if skb is fully within the SACK block. In presence of GSO skbs, |
1205 | * the incoming SACK may not exactly match but we can find smaller MSS | |
1206 | * aligned portion of it that matches. Therefore we might need to fragment | |
1207 | * which may fail and creates some hassle (caller must handle error case | |
1208 | * returns). | |
1209 | */ | |
0f79efdc AB |
1210 | static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, |
1211 | u32 start_seq, u32 end_seq) | |
d1935942 IJ |
1212 | { |
1213 | int in_sack, err; | |
1214 | unsigned int pkt_len; | |
1215 | ||
1216 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && | |
1217 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1218 | ||
1219 | if (tcp_skb_pcount(skb) > 1 && !in_sack && | |
1220 | after(TCP_SKB_CB(skb)->end_seq, start_seq)) { | |
1221 | ||
1222 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); | |
1223 | ||
1224 | if (!in_sack) | |
1225 | pkt_len = start_seq - TCP_SKB_CB(skb)->seq; | |
1226 | else | |
1227 | pkt_len = end_seq - TCP_SKB_CB(skb)->seq; | |
1228 | err = tcp_fragment(sk, skb, pkt_len, skb_shinfo(skb)->gso_size); | |
1229 | if (err < 0) | |
1230 | return err; | |
1231 | } | |
1232 | ||
1233 | return in_sack; | |
1234 | } | |
1235 | ||
6859d494 | 1236 | static int tcp_sacktag_one(struct sk_buff *skb, struct sock *sk, |
9e10c47c IJ |
1237 | int *reord, int dup_sack, int fack_count) |
1238 | { | |
6859d494 | 1239 | struct tcp_sock *tp = tcp_sk(sk); |
9e10c47c IJ |
1240 | u8 sacked = TCP_SKB_CB(skb)->sacked; |
1241 | int flag = 0; | |
1242 | ||
1243 | /* Account D-SACK for retransmitted packet. */ | |
1244 | if (dup_sack && (sacked & TCPCB_RETRANS)) { | |
1245 | if (after(TCP_SKB_CB(skb)->end_seq, tp->undo_marker)) | |
1246 | tp->undo_retrans--; | |
ede9f3b1 | 1247 | if (sacked & TCPCB_SACKED_ACKED) |
9e10c47c IJ |
1248 | *reord = min(fack_count, *reord); |
1249 | } | |
1250 | ||
1251 | /* Nothing to do; acked frame is about to be dropped (was ACKed). */ | |
1252 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1253 | return flag; | |
1254 | ||
1255 | if (!(sacked & TCPCB_SACKED_ACKED)) { | |
1256 | if (sacked & TCPCB_SACKED_RETRANS) { | |
1257 | /* If the segment is not tagged as lost, | |
1258 | * we do not clear RETRANS, believing | |
1259 | * that retransmission is still in flight. | |
1260 | */ | |
1261 | if (sacked & TCPCB_LOST) { | |
1262 | TCP_SKB_CB(skb)->sacked &= | |
1263 | ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); | |
1264 | tp->lost_out -= tcp_skb_pcount(skb); | |
1265 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1266 | ||
1267 | /* clear lost hint */ | |
1268 | tp->retransmit_skb_hint = NULL; | |
1269 | } | |
1270 | } else { | |
1271 | if (!(sacked & TCPCB_RETRANS)) { | |
1272 | /* New sack for not retransmitted frame, | |
1273 | * which was in hole. It is reordering. | |
1274 | */ | |
1275 | if (before(TCP_SKB_CB(skb)->seq, | |
1276 | tcp_highest_sack_seq(tp))) | |
1277 | *reord = min(fack_count, *reord); | |
1278 | ||
1279 | /* SACK enhanced F-RTO (RFC4138; Appendix B) */ | |
1280 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->frto_highmark)) | |
1281 | flag |= FLAG_ONLY_ORIG_SACKED; | |
1282 | } | |
1283 | ||
1284 | if (sacked & TCPCB_LOST) { | |
1285 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_LOST; | |
1286 | tp->lost_out -= tcp_skb_pcount(skb); | |
1287 | ||
1288 | /* clear lost hint */ | |
1289 | tp->retransmit_skb_hint = NULL; | |
1290 | } | |
1291 | } | |
1292 | ||
1293 | TCP_SKB_CB(skb)->sacked |= TCPCB_SACKED_ACKED; | |
1294 | flag |= FLAG_DATA_SACKED; | |
1295 | tp->sacked_out += tcp_skb_pcount(skb); | |
1296 | ||
1297 | fack_count += tcp_skb_pcount(skb); | |
1298 | ||
1299 | /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ | |
1300 | if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) && | |
1301 | before(TCP_SKB_CB(skb)->seq, | |
1302 | TCP_SKB_CB(tp->lost_skb_hint)->seq)) | |
1303 | tp->lost_cnt_hint += tcp_skb_pcount(skb); | |
1304 | ||
1305 | if (fack_count > tp->fackets_out) | |
1306 | tp->fackets_out = fack_count; | |
1307 | ||
6859d494 IJ |
1308 | if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) |
1309 | tcp_advance_highest_sack(sk, skb); | |
9e10c47c IJ |
1310 | } |
1311 | ||
1312 | /* D-SACK. We can detect redundant retransmission in S|R and plain R | |
1313 | * frames and clear it. undo_retrans is decreased above, L|R frames | |
1314 | * are accounted above as well. | |
1315 | */ | |
1316 | if (dup_sack && (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) { | |
1317 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; | |
1318 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1319 | tp->retransmit_skb_hint = NULL; | |
1320 | } | |
1321 | ||
1322 | return flag; | |
1323 | } | |
1324 | ||
68f8353b IJ |
1325 | static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, |
1326 | struct tcp_sack_block *next_dup, | |
1327 | u32 start_seq, u32 end_seq, | |
1328 | int dup_sack_in, int *fack_count, | |
1329 | int *reord, int *flag) | |
1330 | { | |
68f8353b IJ |
1331 | tcp_for_write_queue_from(skb, sk) { |
1332 | int in_sack = 0; | |
1333 | int dup_sack = dup_sack_in; | |
1334 | ||
1335 | if (skb == tcp_send_head(sk)) | |
1336 | break; | |
1337 | ||
1338 | /* queue is in-order => we can short-circuit the walk early */ | |
1339 | if (!before(TCP_SKB_CB(skb)->seq, end_seq)) | |
1340 | break; | |
1341 | ||
1342 | if ((next_dup != NULL) && | |
1343 | before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { | |
1344 | in_sack = tcp_match_skb_to_sack(sk, skb, | |
1345 | next_dup->start_seq, | |
1346 | next_dup->end_seq); | |
1347 | if (in_sack > 0) | |
1348 | dup_sack = 1; | |
1349 | } | |
1350 | ||
1351 | if (in_sack <= 0) | |
056834d9 IJ |
1352 | in_sack = tcp_match_skb_to_sack(sk, skb, start_seq, |
1353 | end_seq); | |
68f8353b IJ |
1354 | if (unlikely(in_sack < 0)) |
1355 | break; | |
1356 | ||
1357 | if (in_sack) | |
056834d9 IJ |
1358 | *flag |= tcp_sacktag_one(skb, sk, reord, dup_sack, |
1359 | *fack_count); | |
68f8353b IJ |
1360 | |
1361 | *fack_count += tcp_skb_pcount(skb); | |
1362 | } | |
1363 | return skb; | |
1364 | } | |
1365 | ||
1366 | /* Avoid all extra work that is being done by sacktag while walking in | |
1367 | * a normal way | |
1368 | */ | |
1369 | static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, | |
d152a7d8 | 1370 | u32 skip_to_seq, int *fack_count) |
68f8353b IJ |
1371 | { |
1372 | tcp_for_write_queue_from(skb, sk) { | |
1373 | if (skb == tcp_send_head(sk)) | |
1374 | break; | |
1375 | ||
ea4f76ae | 1376 | if (!before(TCP_SKB_CB(skb)->end_seq, skip_to_seq)) |
68f8353b | 1377 | break; |
d152a7d8 IJ |
1378 | |
1379 | *fack_count += tcp_skb_pcount(skb); | |
68f8353b IJ |
1380 | } |
1381 | return skb; | |
1382 | } | |
1383 | ||
1384 | static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, | |
1385 | struct sock *sk, | |
1386 | struct tcp_sack_block *next_dup, | |
1387 | u32 skip_to_seq, | |
1388 | int *fack_count, int *reord, | |
1389 | int *flag) | |
1390 | { | |
1391 | if (next_dup == NULL) | |
1392 | return skb; | |
1393 | ||
1394 | if (before(next_dup->start_seq, skip_to_seq)) { | |
d152a7d8 | 1395 | skb = tcp_sacktag_skip(skb, sk, next_dup->start_seq, fack_count); |
68f8353b IJ |
1396 | tcp_sacktag_walk(skb, sk, NULL, |
1397 | next_dup->start_seq, next_dup->end_seq, | |
1398 | 1, fack_count, reord, flag); | |
1399 | } | |
1400 | ||
1401 | return skb; | |
1402 | } | |
1403 | ||
1404 | static int tcp_sack_cache_ok(struct tcp_sock *tp, struct tcp_sack_block *cache) | |
1405 | { | |
1406 | return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); | |
1407 | } | |
1408 | ||
1da177e4 | 1409 | static int |
056834d9 IJ |
1410 | tcp_sacktag_write_queue(struct sock *sk, struct sk_buff *ack_skb, |
1411 | u32 prior_snd_una) | |
1da177e4 | 1412 | { |
6687e988 | 1413 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 1414 | struct tcp_sock *tp = tcp_sk(sk); |
9c70220b ACM |
1415 | unsigned char *ptr = (skb_transport_header(ack_skb) + |
1416 | TCP_SKB_CB(ack_skb)->sacked); | |
fd6dad61 IJ |
1417 | struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); |
1418 | struct tcp_sack_block sp[4]; | |
68f8353b IJ |
1419 | struct tcp_sack_block *cache; |
1420 | struct sk_buff *skb; | |
056834d9 | 1421 | int num_sacks = (ptr[1] - TCPOLEN_SACK_BASE) >> 3; |
fd6dad61 | 1422 | int used_sacks; |
1da177e4 | 1423 | int reord = tp->packets_out; |
1da177e4 | 1424 | int flag = 0; |
7769f406 | 1425 | int found_dup_sack = 0; |
68f8353b IJ |
1426 | int fack_count; |
1427 | int i, j; | |
fda03fbb | 1428 | int first_sack_index; |
1da177e4 | 1429 | |
d738cd8f | 1430 | if (!tp->sacked_out) { |
de83c058 IJ |
1431 | if (WARN_ON(tp->fackets_out)) |
1432 | tp->fackets_out = 0; | |
6859d494 | 1433 | tcp_highest_sack_reset(sk); |
d738cd8f | 1434 | } |
1da177e4 | 1435 | |
fd6dad61 | 1436 | found_dup_sack = tcp_check_dsack(tp, ack_skb, sp_wire, |
d06e021d DM |
1437 | num_sacks, prior_snd_una); |
1438 | if (found_dup_sack) | |
49ff4bb4 | 1439 | flag |= FLAG_DSACKING_ACK; |
6f74651a BE |
1440 | |
1441 | /* Eliminate too old ACKs, but take into | |
1442 | * account more or less fresh ones, they can | |
1443 | * contain valid SACK info. | |
1444 | */ | |
1445 | if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) | |
1446 | return 0; | |
1447 |