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 | * | |
02c30a84 | 8 | * Authors: Ross Biro |
1da177e4 LT |
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. | |
caa20d9a | 43 | * Andrey Savochkin: Fix RTT measurements in the presence of |
1da177e4 LT |
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. | |
e905a9ed | 51 | * Andi Kleen: Add tcp_measure_rcv_mss to make |
1da177e4 | 52 | * connections with MSS<min(MTU,ann. MSS) |
e905a9ed | 53 | * work without delayed acks. |
1da177e4 LT |
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 | |
1da177e4 LT |
62 | */ |
63 | ||
afd46503 JP |
64 | #define pr_fmt(fmt) "TCP: " fmt |
65 | ||
1da177e4 | 66 | #include <linux/mm.h> |
5a0e3ad6 | 67 | #include <linux/slab.h> |
1da177e4 LT |
68 | #include <linux/module.h> |
69 | #include <linux/sysctl.h> | |
a0bffffc | 70 | #include <linux/kernel.h> |
5ffc02a1 | 71 | #include <net/dst.h> |
1da177e4 LT |
72 | #include <net/tcp.h> |
73 | #include <net/inet_common.h> | |
74 | #include <linux/ipsec.h> | |
75 | #include <asm/unaligned.h> | |
1a2449a8 | 76 | #include <net/netdma.h> |
1da177e4 | 77 | |
ab32ea5d BH |
78 | int sysctl_tcp_timestamps __read_mostly = 1; |
79 | int sysctl_tcp_window_scaling __read_mostly = 1; | |
80 | int sysctl_tcp_sack __read_mostly = 1; | |
81 | int sysctl_tcp_fack __read_mostly = 1; | |
82 | int sysctl_tcp_reordering __read_mostly = TCP_FASTRETRANS_THRESH; | |
4bc2f18b | 83 | EXPORT_SYMBOL(sysctl_tcp_reordering); |
255cac91 | 84 | int sysctl_tcp_ecn __read_mostly = 2; |
4bc2f18b | 85 | EXPORT_SYMBOL(sysctl_tcp_ecn); |
ab32ea5d BH |
86 | int sysctl_tcp_dsack __read_mostly = 1; |
87 | int sysctl_tcp_app_win __read_mostly = 31; | |
88 | int sysctl_tcp_adv_win_scale __read_mostly = 2; | |
4bc2f18b | 89 | EXPORT_SYMBOL(sysctl_tcp_adv_win_scale); |
1da177e4 | 90 | |
ab32ea5d BH |
91 | int sysctl_tcp_stdurg __read_mostly; |
92 | int sysctl_tcp_rfc1337 __read_mostly; | |
93 | int sysctl_tcp_max_orphans __read_mostly = NR_FILE; | |
c96fd3d4 | 94 | int sysctl_tcp_frto __read_mostly = 2; |
3cfe3baa | 95 | int sysctl_tcp_frto_response __read_mostly; |
ab32ea5d | 96 | int sysctl_tcp_nometrics_save __read_mostly; |
1da177e4 | 97 | |
7e380175 AP |
98 | int sysctl_tcp_thin_dupack __read_mostly; |
99 | ||
ab32ea5d BH |
100 | int sysctl_tcp_moderate_rcvbuf __read_mostly = 1; |
101 | int sysctl_tcp_abc __read_mostly; | |
1da177e4 | 102 | |
1da177e4 LT |
103 | #define FLAG_DATA 0x01 /* Incoming frame contained data. */ |
104 | #define FLAG_WIN_UPDATE 0x02 /* Incoming ACK was a window update. */ | |
105 | #define FLAG_DATA_ACKED 0x04 /* This ACK acknowledged new data. */ | |
106 | #define FLAG_RETRANS_DATA_ACKED 0x08 /* "" "" some of which was retransmitted. */ | |
107 | #define FLAG_SYN_ACKED 0x10 /* This ACK acknowledged SYN. */ | |
108 | #define FLAG_DATA_SACKED 0x20 /* New SACK. */ | |
109 | #define FLAG_ECE 0x40 /* ECE in this ACK */ | |
1da177e4 | 110 | #define FLAG_SLOWPATH 0x100 /* Do not skip RFC checks for window update.*/ |
4dc2665e | 111 | #define FLAG_ONLY_ORIG_SACKED 0x200 /* SACKs only non-rexmit sent before RTO */ |
2e605294 | 112 | #define FLAG_SND_UNA_ADVANCED 0x400 /* Snd_una was changed (!= FLAG_DATA_ACKED) */ |
564262c1 | 113 | #define FLAG_DSACKING_ACK 0x800 /* SACK blocks contained D-SACK info */ |
009a2e3e | 114 | #define FLAG_NONHEAD_RETRANS_ACKED 0x1000 /* Non-head rexmitted data was ACKed */ |
cadbd031 | 115 | #define FLAG_SACK_RENEGING 0x2000 /* snd_una advanced to a sacked seq */ |
1da177e4 LT |
116 | |
117 | #define FLAG_ACKED (FLAG_DATA_ACKED|FLAG_SYN_ACKED) | |
118 | #define FLAG_NOT_DUP (FLAG_DATA|FLAG_WIN_UPDATE|FLAG_ACKED) | |
119 | #define FLAG_CA_ALERT (FLAG_DATA_SACKED|FLAG_ECE) | |
120 | #define FLAG_FORWARD_PROGRESS (FLAG_ACKED|FLAG_DATA_SACKED) | |
2e605294 | 121 | #define FLAG_ANY_PROGRESS (FLAG_FORWARD_PROGRESS|FLAG_SND_UNA_ADVANCED) |
1da177e4 | 122 | |
1da177e4 | 123 | #define TCP_REMNANT (TCP_FLAG_FIN|TCP_FLAG_URG|TCP_FLAG_SYN|TCP_FLAG_PSH) |
bdf1ee5d | 124 | #define TCP_HP_BITS (~(TCP_RESERVED_BITS|TCP_FLAG_PSH)) |
1da177e4 | 125 | |
e905a9ed | 126 | /* Adapt the MSS value used to make delayed ack decision to the |
1da177e4 | 127 | * real world. |
e905a9ed | 128 | */ |
056834d9 | 129 | static void tcp_measure_rcv_mss(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 130 | { |
463c84b9 | 131 | struct inet_connection_sock *icsk = inet_csk(sk); |
e905a9ed | 132 | const unsigned int lss = icsk->icsk_ack.last_seg_size; |
463c84b9 | 133 | unsigned int len; |
1da177e4 | 134 | |
e905a9ed | 135 | icsk->icsk_ack.last_seg_size = 0; |
1da177e4 LT |
136 | |
137 | /* skb->len may jitter because of SACKs, even if peer | |
138 | * sends good full-sized frames. | |
139 | */ | |
056834d9 | 140 | len = skb_shinfo(skb)->gso_size ? : skb->len; |
463c84b9 ACM |
141 | if (len >= icsk->icsk_ack.rcv_mss) { |
142 | icsk->icsk_ack.rcv_mss = len; | |
1da177e4 LT |
143 | } else { |
144 | /* Otherwise, we make more careful check taking into account, | |
145 | * that SACKs block is variable. | |
146 | * | |
147 | * "len" is invariant segment length, including TCP header. | |
148 | */ | |
9c70220b | 149 | len += skb->data - skb_transport_header(skb); |
bee7ca9e | 150 | if (len >= TCP_MSS_DEFAULT + sizeof(struct tcphdr) || |
1da177e4 LT |
151 | /* If PSH is not set, packet should be |
152 | * full sized, provided peer TCP is not badly broken. | |
153 | * This observation (if it is correct 8)) allows | |
154 | * to handle super-low mtu links fairly. | |
155 | */ | |
156 | (len >= TCP_MIN_MSS + sizeof(struct tcphdr) && | |
aa8223c7 | 157 | !(tcp_flag_word(tcp_hdr(skb)) & TCP_REMNANT))) { |
1da177e4 LT |
158 | /* Subtract also invariant (if peer is RFC compliant), |
159 | * tcp header plus fixed timestamp option length. | |
160 | * Resulting "len" is MSS free of SACK jitter. | |
161 | */ | |
463c84b9 ACM |
162 | len -= tcp_sk(sk)->tcp_header_len; |
163 | icsk->icsk_ack.last_seg_size = len; | |
1da177e4 | 164 | if (len == lss) { |
463c84b9 | 165 | icsk->icsk_ack.rcv_mss = len; |
1da177e4 LT |
166 | return; |
167 | } | |
168 | } | |
1ef9696c AK |
169 | if (icsk->icsk_ack.pending & ICSK_ACK_PUSHED) |
170 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED2; | |
463c84b9 | 171 | icsk->icsk_ack.pending |= ICSK_ACK_PUSHED; |
1da177e4 LT |
172 | } |
173 | } | |
174 | ||
463c84b9 | 175 | static void tcp_incr_quickack(struct sock *sk) |
1da177e4 | 176 | { |
463c84b9 ACM |
177 | struct inet_connection_sock *icsk = inet_csk(sk); |
178 | unsigned quickacks = tcp_sk(sk)->rcv_wnd / (2 * icsk->icsk_ack.rcv_mss); | |
1da177e4 | 179 | |
056834d9 IJ |
180 | if (quickacks == 0) |
181 | quickacks = 2; | |
463c84b9 ACM |
182 | if (quickacks > icsk->icsk_ack.quick) |
183 | icsk->icsk_ack.quick = min(quickacks, TCP_MAX_QUICKACKS); | |
1da177e4 LT |
184 | } |
185 | ||
1b9f4092 | 186 | static void tcp_enter_quickack_mode(struct sock *sk) |
1da177e4 | 187 | { |
463c84b9 ACM |
188 | struct inet_connection_sock *icsk = inet_csk(sk); |
189 | tcp_incr_quickack(sk); | |
190 | icsk->icsk_ack.pingpong = 0; | |
191 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 LT |
192 | } |
193 | ||
194 | /* Send ACKs quickly, if "quick" count is not exhausted | |
195 | * and the session is not interactive. | |
196 | */ | |
197 | ||
463c84b9 | 198 | static inline int tcp_in_quickack_mode(const struct sock *sk) |
1da177e4 | 199 | { |
463c84b9 ACM |
200 | const struct inet_connection_sock *icsk = inet_csk(sk); |
201 | return icsk->icsk_ack.quick && !icsk->icsk_ack.pingpong; | |
1da177e4 LT |
202 | } |
203 | ||
bdf1ee5d IJ |
204 | static inline void TCP_ECN_queue_cwr(struct tcp_sock *tp) |
205 | { | |
056834d9 | 206 | if (tp->ecn_flags & TCP_ECN_OK) |
bdf1ee5d IJ |
207 | tp->ecn_flags |= TCP_ECN_QUEUE_CWR; |
208 | } | |
209 | ||
cf533ea5 | 210 | static inline void TCP_ECN_accept_cwr(struct tcp_sock *tp, const struct sk_buff *skb) |
bdf1ee5d IJ |
211 | { |
212 | if (tcp_hdr(skb)->cwr) | |
213 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
214 | } | |
215 | ||
216 | static inline void TCP_ECN_withdraw_cwr(struct tcp_sock *tp) | |
217 | { | |
218 | tp->ecn_flags &= ~TCP_ECN_DEMAND_CWR; | |
219 | } | |
220 | ||
7a269ffa | 221 | static inline void TCP_ECN_check_ce(struct tcp_sock *tp, const struct sk_buff *skb) |
bdf1ee5d | 222 | { |
7a269ffa ED |
223 | if (!(tp->ecn_flags & TCP_ECN_OK)) |
224 | return; | |
225 | ||
b82d1bb4 | 226 | switch (TCP_SKB_CB(skb)->ip_dsfield & INET_ECN_MASK) { |
7a269ffa | 227 | case INET_ECN_NOT_ECT: |
bdf1ee5d | 228 | /* Funny extension: if ECT is not set on a segment, |
7a269ffa ED |
229 | * and we already seen ECT on a previous segment, |
230 | * it is probably a retransmit. | |
231 | */ | |
232 | if (tp->ecn_flags & TCP_ECN_SEEN) | |
bdf1ee5d | 233 | tcp_enter_quickack_mode((struct sock *)tp); |
7a269ffa ED |
234 | break; |
235 | case INET_ECN_CE: | |
236 | tp->ecn_flags |= TCP_ECN_DEMAND_CWR; | |
237 | /* fallinto */ | |
238 | default: | |
239 | tp->ecn_flags |= TCP_ECN_SEEN; | |
bdf1ee5d IJ |
240 | } |
241 | } | |
242 | ||
cf533ea5 | 243 | static inline void TCP_ECN_rcv_synack(struct tcp_sock *tp, const struct tcphdr *th) |
bdf1ee5d | 244 | { |
056834d9 | 245 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || th->cwr)) |
bdf1ee5d IJ |
246 | tp->ecn_flags &= ~TCP_ECN_OK; |
247 | } | |
248 | ||
cf533ea5 | 249 | static inline void TCP_ECN_rcv_syn(struct tcp_sock *tp, const struct tcphdr *th) |
bdf1ee5d | 250 | { |
056834d9 | 251 | if ((tp->ecn_flags & TCP_ECN_OK) && (!th->ece || !th->cwr)) |
bdf1ee5d IJ |
252 | tp->ecn_flags &= ~TCP_ECN_OK; |
253 | } | |
254 | ||
cf533ea5 | 255 | static inline int TCP_ECN_rcv_ecn_echo(const struct tcp_sock *tp, const struct tcphdr *th) |
bdf1ee5d | 256 | { |
056834d9 | 257 | if (th->ece && !th->syn && (tp->ecn_flags & TCP_ECN_OK)) |
bdf1ee5d IJ |
258 | return 1; |
259 | return 0; | |
260 | } | |
261 | ||
1da177e4 LT |
262 | /* Buffer size and advertised window tuning. |
263 | * | |
264 | * 1. Tuning sk->sk_sndbuf, when connection enters established state. | |
265 | */ | |
266 | ||
267 | static void tcp_fixup_sndbuf(struct sock *sk) | |
268 | { | |
87fb4b7b | 269 | int sndmem = SKB_TRUESIZE(tcp_sk(sk)->rx_opt.mss_clamp + MAX_TCP_HEADER); |
1da177e4 | 270 | |
06a59ecb ED |
271 | sndmem *= TCP_INIT_CWND; |
272 | if (sk->sk_sndbuf < sndmem) | |
273 | sk->sk_sndbuf = min(sndmem, sysctl_tcp_wmem[2]); | |
1da177e4 LT |
274 | } |
275 | ||
276 | /* 2. Tuning advertised window (window_clamp, rcv_ssthresh) | |
277 | * | |
278 | * All tcp_full_space() is split to two parts: "network" buffer, allocated | |
279 | * forward and advertised in receiver window (tp->rcv_wnd) and | |
280 | * "application buffer", required to isolate scheduling/application | |
281 | * latencies from network. | |
282 | * window_clamp is maximal advertised window. It can be less than | |
283 | * tcp_full_space(), in this case tcp_full_space() - window_clamp | |
284 | * is reserved for "application" buffer. The less window_clamp is | |
285 | * the smoother our behaviour from viewpoint of network, but the lower | |
286 | * throughput and the higher sensitivity of the connection to losses. 8) | |
287 | * | |
288 | * rcv_ssthresh is more strict window_clamp used at "slow start" | |
289 | * phase to predict further behaviour of this connection. | |
290 | * It is used for two goals: | |
291 | * - to enforce header prediction at sender, even when application | |
292 | * requires some significant "application buffer". It is check #1. | |
293 | * - to prevent pruning of receive queue because of misprediction | |
294 | * of receiver window. Check #2. | |
295 | * | |
296 | * The scheme does not work when sender sends good segments opening | |
caa20d9a | 297 | * window and then starts to feed us spaghetti. But it should work |
1da177e4 LT |
298 | * in common situations. Otherwise, we have to rely on queue collapsing. |
299 | */ | |
300 | ||
301 | /* Slow part of check#2. */ | |
9e412ba7 | 302 | static int __tcp_grow_window(const struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 303 | { |
9e412ba7 | 304 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 305 | /* Optimize this! */ |
dfd4f0ae ED |
306 | int truesize = tcp_win_from_space(skb->truesize) >> 1; |
307 | int window = tcp_win_from_space(sysctl_tcp_rmem[2]) >> 1; | |
1da177e4 LT |
308 | |
309 | while (tp->rcv_ssthresh <= window) { | |
310 | if (truesize <= skb->len) | |
463c84b9 | 311 | return 2 * inet_csk(sk)->icsk_ack.rcv_mss; |
1da177e4 LT |
312 | |
313 | truesize >>= 1; | |
314 | window >>= 1; | |
315 | } | |
316 | return 0; | |
317 | } | |
318 | ||
cf533ea5 | 319 | static void tcp_grow_window(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 320 | { |
9e412ba7 IJ |
321 | struct tcp_sock *tp = tcp_sk(sk); |
322 | ||
1da177e4 LT |
323 | /* Check #1 */ |
324 | if (tp->rcv_ssthresh < tp->window_clamp && | |
325 | (int)tp->rcv_ssthresh < tcp_space(sk) && | |
180d8cd9 | 326 | !sk_under_memory_pressure(sk)) { |
1da177e4 LT |
327 | int incr; |
328 | ||
329 | /* Check #2. Increase window, if skb with such overhead | |
330 | * will fit to rcvbuf in future. | |
331 | */ | |
332 | if (tcp_win_from_space(skb->truesize) <= skb->len) | |
056834d9 | 333 | incr = 2 * tp->advmss; |
1da177e4 | 334 | else |
9e412ba7 | 335 | incr = __tcp_grow_window(sk, skb); |
1da177e4 LT |
336 | |
337 | if (incr) { | |
056834d9 IJ |
338 | tp->rcv_ssthresh = min(tp->rcv_ssthresh + incr, |
339 | tp->window_clamp); | |
463c84b9 | 340 | inet_csk(sk)->icsk_ack.quick |= 1; |
1da177e4 LT |
341 | } |
342 | } | |
343 | } | |
344 | ||
345 | /* 3. Tuning rcvbuf, when connection enters established state. */ | |
346 | ||
347 | static void tcp_fixup_rcvbuf(struct sock *sk) | |
348 | { | |
e9266a02 ED |
349 | u32 mss = tcp_sk(sk)->advmss; |
350 | u32 icwnd = TCP_DEFAULT_INIT_RCVWND; | |
351 | int rcvmem; | |
1da177e4 | 352 | |
e9266a02 ED |
353 | /* Limit to 10 segments if mss <= 1460, |
354 | * or 14600/mss segments, with a minimum of two segments. | |
1da177e4 | 355 | */ |
e9266a02 ED |
356 | if (mss > 1460) |
357 | icwnd = max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2); | |
358 | ||
359 | rcvmem = SKB_TRUESIZE(mss + MAX_TCP_HEADER); | |
360 | while (tcp_win_from_space(rcvmem) < mss) | |
1da177e4 | 361 | rcvmem += 128; |
e9266a02 ED |
362 | |
363 | rcvmem *= icwnd; | |
364 | ||
365 | if (sk->sk_rcvbuf < rcvmem) | |
366 | sk->sk_rcvbuf = min(rcvmem, sysctl_tcp_rmem[2]); | |
1da177e4 LT |
367 | } |
368 | ||
caa20d9a | 369 | /* 4. Try to fixup all. It is made immediately after connection enters |
1da177e4 LT |
370 | * established state. |
371 | */ | |
372 | static void tcp_init_buffer_space(struct sock *sk) | |
373 | { | |
374 | struct tcp_sock *tp = tcp_sk(sk); | |
375 | int maxwin; | |
376 | ||
377 | if (!(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) | |
378 | tcp_fixup_rcvbuf(sk); | |
379 | if (!(sk->sk_userlocks & SOCK_SNDBUF_LOCK)) | |
380 | tcp_fixup_sndbuf(sk); | |
381 | ||
382 | tp->rcvq_space.space = tp->rcv_wnd; | |
383 | ||
384 | maxwin = tcp_full_space(sk); | |
385 | ||
386 | if (tp->window_clamp >= maxwin) { | |
387 | tp->window_clamp = maxwin; | |
388 | ||
389 | if (sysctl_tcp_app_win && maxwin > 4 * tp->advmss) | |
390 | tp->window_clamp = max(maxwin - | |
391 | (maxwin >> sysctl_tcp_app_win), | |
392 | 4 * tp->advmss); | |
393 | } | |
394 | ||
395 | /* Force reservation of one segment. */ | |
396 | if (sysctl_tcp_app_win && | |
397 | tp->window_clamp > 2 * tp->advmss && | |
398 | tp->window_clamp + tp->advmss > maxwin) | |
399 | tp->window_clamp = max(2 * tp->advmss, maxwin - tp->advmss); | |
400 | ||
401 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, tp->window_clamp); | |
402 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
403 | } | |
404 | ||
1da177e4 | 405 | /* 5. Recalculate window clamp after socket hit its memory bounds. */ |
9e412ba7 | 406 | static void tcp_clamp_window(struct sock *sk) |
1da177e4 | 407 | { |
9e412ba7 | 408 | struct tcp_sock *tp = tcp_sk(sk); |
6687e988 | 409 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 410 | |
6687e988 | 411 | icsk->icsk_ack.quick = 0; |
1da177e4 | 412 | |
326f36e9 JH |
413 | if (sk->sk_rcvbuf < sysctl_tcp_rmem[2] && |
414 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK) && | |
180d8cd9 GC |
415 | !sk_under_memory_pressure(sk) && |
416 | sk_memory_allocated(sk) < sk_prot_mem_limits(sk, 0)) { | |
326f36e9 JH |
417 | sk->sk_rcvbuf = min(atomic_read(&sk->sk_rmem_alloc), |
418 | sysctl_tcp_rmem[2]); | |
1da177e4 | 419 | } |
326f36e9 | 420 | if (atomic_read(&sk->sk_rmem_alloc) > sk->sk_rcvbuf) |
056834d9 | 421 | tp->rcv_ssthresh = min(tp->window_clamp, 2U * tp->advmss); |
1da177e4 LT |
422 | } |
423 | ||
40efc6fa SH |
424 | /* Initialize RCV_MSS value. |
425 | * RCV_MSS is an our guess about MSS used by the peer. | |
426 | * We haven't any direct information about the MSS. | |
427 | * It's better to underestimate the RCV_MSS rather than overestimate. | |
428 | * Overestimations make us ACKing less frequently than needed. | |
429 | * Underestimations are more easy to detect and fix by tcp_measure_rcv_mss(). | |
430 | */ | |
431 | void tcp_initialize_rcv_mss(struct sock *sk) | |
432 | { | |
cf533ea5 | 433 | const struct tcp_sock *tp = tcp_sk(sk); |
40efc6fa SH |
434 | unsigned int hint = min_t(unsigned int, tp->advmss, tp->mss_cache); |
435 | ||
056834d9 | 436 | hint = min(hint, tp->rcv_wnd / 2); |
bee7ca9e | 437 | hint = min(hint, TCP_MSS_DEFAULT); |
40efc6fa SH |
438 | hint = max(hint, TCP_MIN_MSS); |
439 | ||
440 | inet_csk(sk)->icsk_ack.rcv_mss = hint; | |
441 | } | |
4bc2f18b | 442 | EXPORT_SYMBOL(tcp_initialize_rcv_mss); |
40efc6fa | 443 | |
1da177e4 LT |
444 | /* Receiver "autotuning" code. |
445 | * | |
446 | * The algorithm for RTT estimation w/o timestamps is based on | |
447 | * Dynamic Right-Sizing (DRS) by Wu Feng and Mike Fisk of LANL. | |
631dd1a8 | 448 | * <http://public.lanl.gov/radiant/pubs.html#DRS> |
1da177e4 LT |
449 | * |
450 | * More detail on this code can be found at | |
631dd1a8 | 451 | * <http://staff.psc.edu/jheffner/>, |
1da177e4 LT |
452 | * though this reference is out of date. A new paper |
453 | * is pending. | |
454 | */ | |
455 | static void tcp_rcv_rtt_update(struct tcp_sock *tp, u32 sample, int win_dep) | |
456 | { | |
457 | u32 new_sample = tp->rcv_rtt_est.rtt; | |
458 | long m = sample; | |
459 | ||
460 | if (m == 0) | |
461 | m = 1; | |
462 | ||
463 | if (new_sample != 0) { | |
464 | /* If we sample in larger samples in the non-timestamp | |
465 | * case, we could grossly overestimate the RTT especially | |
466 | * with chatty applications or bulk transfer apps which | |
467 | * are stalled on filesystem I/O. | |
468 | * | |
469 | * Also, since we are only going for a minimum in the | |
31f34269 | 470 | * non-timestamp case, we do not smooth things out |
caa20d9a | 471 | * else with timestamps disabled convergence takes too |
1da177e4 LT |
472 | * long. |
473 | */ | |
474 | if (!win_dep) { | |
475 | m -= (new_sample >> 3); | |
476 | new_sample += m; | |
477 | } else if (m < new_sample) | |
478 | new_sample = m << 3; | |
479 | } else { | |
caa20d9a | 480 | /* No previous measure. */ |
1da177e4 LT |
481 | new_sample = m << 3; |
482 | } | |
483 | ||
484 | if (tp->rcv_rtt_est.rtt != new_sample) | |
485 | tp->rcv_rtt_est.rtt = new_sample; | |
486 | } | |
487 | ||
488 | static inline void tcp_rcv_rtt_measure(struct tcp_sock *tp) | |
489 | { | |
490 | if (tp->rcv_rtt_est.time == 0) | |
491 | goto new_measure; | |
492 | if (before(tp->rcv_nxt, tp->rcv_rtt_est.seq)) | |
493 | return; | |
056834d9 | 494 | tcp_rcv_rtt_update(tp, jiffies - tp->rcv_rtt_est.time, 1); |
1da177e4 LT |
495 | |
496 | new_measure: | |
497 | tp->rcv_rtt_est.seq = tp->rcv_nxt + tp->rcv_wnd; | |
498 | tp->rcv_rtt_est.time = tcp_time_stamp; | |
499 | } | |
500 | ||
056834d9 IJ |
501 | static inline void tcp_rcv_rtt_measure_ts(struct sock *sk, |
502 | const struct sk_buff *skb) | |
1da177e4 | 503 | { |
463c84b9 | 504 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
505 | if (tp->rx_opt.rcv_tsecr && |
506 | (TCP_SKB_CB(skb)->end_seq - | |
463c84b9 | 507 | TCP_SKB_CB(skb)->seq >= inet_csk(sk)->icsk_ack.rcv_mss)) |
1da177e4 LT |
508 | tcp_rcv_rtt_update(tp, tcp_time_stamp - tp->rx_opt.rcv_tsecr, 0); |
509 | } | |
510 | ||
511 | /* | |
512 | * This function should be called every time data is copied to user space. | |
513 | * It calculates the appropriate TCP receive buffer space. | |
514 | */ | |
515 | void tcp_rcv_space_adjust(struct sock *sk) | |
516 | { | |
517 | struct tcp_sock *tp = tcp_sk(sk); | |
518 | int time; | |
519 | int space; | |
e905a9ed | 520 | |
1da177e4 LT |
521 | if (tp->rcvq_space.time == 0) |
522 | goto new_measure; | |
e905a9ed | 523 | |
1da177e4 | 524 | time = tcp_time_stamp - tp->rcvq_space.time; |
056834d9 | 525 | if (time < (tp->rcv_rtt_est.rtt >> 3) || tp->rcv_rtt_est.rtt == 0) |
1da177e4 | 526 | return; |
e905a9ed | 527 | |
1da177e4 LT |
528 | space = 2 * (tp->copied_seq - tp->rcvq_space.seq); |
529 | ||
530 | space = max(tp->rcvq_space.space, space); | |
531 | ||
532 | if (tp->rcvq_space.space != space) { | |
533 | int rcvmem; | |
534 | ||
535 | tp->rcvq_space.space = space; | |
536 | ||
6fcf9412 JH |
537 | if (sysctl_tcp_moderate_rcvbuf && |
538 | !(sk->sk_userlocks & SOCK_RCVBUF_LOCK)) { | |
1da177e4 LT |
539 | int new_clamp = space; |
540 | ||
541 | /* Receive space grows, normalize in order to | |
542 | * take into account packet headers and sk_buff | |
543 | * structure overhead. | |
544 | */ | |
545 | space /= tp->advmss; | |
546 | if (!space) | |
547 | space = 1; | |
87fb4b7b | 548 | rcvmem = SKB_TRUESIZE(tp->advmss + MAX_TCP_HEADER); |
1da177e4 LT |
549 | while (tcp_win_from_space(rcvmem) < tp->advmss) |
550 | rcvmem += 128; | |
551 | space *= rcvmem; | |
552 | space = min(space, sysctl_tcp_rmem[2]); | |
553 | if (space > sk->sk_rcvbuf) { | |
554 | sk->sk_rcvbuf = space; | |
555 | ||
556 | /* Make the window clamp follow along. */ | |
557 | tp->window_clamp = new_clamp; | |
558 | } | |
559 | } | |
560 | } | |
e905a9ed | 561 | |
1da177e4 LT |
562 | new_measure: |
563 | tp->rcvq_space.seq = tp->copied_seq; | |
564 | tp->rcvq_space.time = tcp_time_stamp; | |
565 | } | |
566 | ||
567 | /* There is something which you must keep in mind when you analyze the | |
568 | * behavior of the tp->ato delayed ack timeout interval. When a | |
569 | * connection starts up, we want to ack as quickly as possible. The | |
570 | * problem is that "good" TCP's do slow start at the beginning of data | |
571 | * transmission. The means that until we send the first few ACK's the | |
572 | * sender will sit on his end and only queue most of his data, because | |
573 | * he can only send snd_cwnd unacked packets at any given time. For | |
574 | * each ACK we send, he increments snd_cwnd and transmits more of his | |
575 | * queue. -DaveM | |
576 | */ | |
9e412ba7 | 577 | static void tcp_event_data_recv(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 578 | { |
9e412ba7 | 579 | struct tcp_sock *tp = tcp_sk(sk); |
463c84b9 | 580 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
581 | u32 now; |
582 | ||
463c84b9 | 583 | inet_csk_schedule_ack(sk); |
1da177e4 | 584 | |
463c84b9 | 585 | tcp_measure_rcv_mss(sk, skb); |
1da177e4 LT |
586 | |
587 | tcp_rcv_rtt_measure(tp); | |
e905a9ed | 588 | |
1da177e4 LT |
589 | now = tcp_time_stamp; |
590 | ||
463c84b9 | 591 | if (!icsk->icsk_ack.ato) { |
1da177e4 LT |
592 | /* The _first_ data packet received, initialize |
593 | * delayed ACK engine. | |
594 | */ | |
463c84b9 ACM |
595 | tcp_incr_quickack(sk); |
596 | icsk->icsk_ack.ato = TCP_ATO_MIN; | |
1da177e4 | 597 | } else { |
463c84b9 | 598 | int m = now - icsk->icsk_ack.lrcvtime; |
1da177e4 | 599 | |
056834d9 | 600 | if (m <= TCP_ATO_MIN / 2) { |
1da177e4 | 601 | /* The fastest case is the first. */ |
463c84b9 ACM |
602 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + TCP_ATO_MIN / 2; |
603 | } else if (m < icsk->icsk_ack.ato) { | |
604 | icsk->icsk_ack.ato = (icsk->icsk_ack.ato >> 1) + m; | |
605 | if (icsk->icsk_ack.ato > icsk->icsk_rto) | |
606 | icsk->icsk_ack.ato = icsk->icsk_rto; | |
607 | } else if (m > icsk->icsk_rto) { | |
caa20d9a | 608 | /* Too long gap. Apparently sender failed to |
1da177e4 LT |
609 | * restart window, so that we send ACKs quickly. |
610 | */ | |
463c84b9 | 611 | tcp_incr_quickack(sk); |
3ab224be | 612 | sk_mem_reclaim(sk); |
1da177e4 LT |
613 | } |
614 | } | |
463c84b9 | 615 | icsk->icsk_ack.lrcvtime = now; |
1da177e4 LT |
616 | |
617 | TCP_ECN_check_ce(tp, skb); | |
618 | ||
619 | if (skb->len >= 128) | |
9e412ba7 | 620 | tcp_grow_window(sk, skb); |
1da177e4 LT |
621 | } |
622 | ||
1da177e4 LT |
623 | /* Called to compute a smoothed rtt estimate. The data fed to this |
624 | * routine either comes from timestamps, or from segments that were | |
625 | * known _not_ to have been retransmitted [see Karn/Partridge | |
626 | * Proceedings SIGCOMM 87]. The algorithm is from the SIGCOMM 88 | |
627 | * piece by Van Jacobson. | |
628 | * NOTE: the next three routines used to be one big routine. | |
629 | * To save cycles in the RFC 1323 implementation it was better to break | |
630 | * it up into three procedures. -- erics | |
631 | */ | |
2d2abbab | 632 | static void tcp_rtt_estimator(struct sock *sk, const __u32 mrtt) |
1da177e4 | 633 | { |
6687e988 | 634 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
635 | long m = mrtt; /* RTT */ |
636 | ||
1da177e4 LT |
637 | /* The following amusing code comes from Jacobson's |
638 | * article in SIGCOMM '88. Note that rtt and mdev | |
639 | * are scaled versions of rtt and mean deviation. | |
e905a9ed | 640 | * This is designed to be as fast as possible |
1da177e4 LT |
641 | * m stands for "measurement". |
642 | * | |
643 | * On a 1990 paper the rto value is changed to: | |
644 | * RTO = rtt + 4 * mdev | |
645 | * | |
646 | * Funny. This algorithm seems to be very broken. | |
647 | * These formulae increase RTO, when it should be decreased, increase | |
31f34269 | 648 | * too slowly, when it should be increased quickly, decrease too quickly |
1da177e4 LT |
649 | * etc. I guess in BSD RTO takes ONE value, so that it is absolutely |
650 | * does not matter how to _calculate_ it. Seems, it was trap | |
651 | * that VJ failed to avoid. 8) | |
652 | */ | |
2de979bd | 653 | if (m == 0) |
1da177e4 LT |
654 | m = 1; |
655 | if (tp->srtt != 0) { | |
656 | m -= (tp->srtt >> 3); /* m is now error in rtt est */ | |
657 | tp->srtt += m; /* rtt = 7/8 rtt + 1/8 new */ | |
658 | if (m < 0) { | |
659 | m = -m; /* m is now abs(error) */ | |
660 | m -= (tp->mdev >> 2); /* similar update on mdev */ | |
661 | /* This is similar to one of Eifel findings. | |
662 | * Eifel blocks mdev updates when rtt decreases. | |
663 | * This solution is a bit different: we use finer gain | |
664 | * for mdev in this case (alpha*beta). | |
665 | * Like Eifel it also prevents growth of rto, | |
666 | * but also it limits too fast rto decreases, | |
667 | * happening in pure Eifel. | |
668 | */ | |
669 | if (m > 0) | |
670 | m >>= 3; | |
671 | } else { | |
672 | m -= (tp->mdev >> 2); /* similar update on mdev */ | |
673 | } | |
674 | tp->mdev += m; /* mdev = 3/4 mdev + 1/4 new */ | |
675 | if (tp->mdev > tp->mdev_max) { | |
676 | tp->mdev_max = tp->mdev; | |
677 | if (tp->mdev_max > tp->rttvar) | |
678 | tp->rttvar = tp->mdev_max; | |
679 | } | |
680 | if (after(tp->snd_una, tp->rtt_seq)) { | |
681 | if (tp->mdev_max < tp->rttvar) | |
056834d9 | 682 | tp->rttvar -= (tp->rttvar - tp->mdev_max) >> 2; |
1da177e4 | 683 | tp->rtt_seq = tp->snd_nxt; |
05bb1fad | 684 | tp->mdev_max = tcp_rto_min(sk); |
1da177e4 LT |
685 | } |
686 | } else { | |
687 | /* no previous measure. */ | |
056834d9 IJ |
688 | tp->srtt = m << 3; /* take the measured time to be rtt */ |
689 | tp->mdev = m << 1; /* make sure rto = 3*rtt */ | |
05bb1fad | 690 | tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); |
1da177e4 LT |
691 | tp->rtt_seq = tp->snd_nxt; |
692 | } | |
1da177e4 LT |
693 | } |
694 | ||
695 | /* Calculate rto without backoff. This is the second half of Van Jacobson's | |
696 | * routine referred to above. | |
697 | */ | |
463c84b9 | 698 | static inline void tcp_set_rto(struct sock *sk) |
1da177e4 | 699 | { |
463c84b9 | 700 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
701 | /* Old crap is replaced with new one. 8) |
702 | * | |
703 | * More seriously: | |
704 | * 1. If rtt variance happened to be less 50msec, it is hallucination. | |
705 | * It cannot be less due to utterly erratic ACK generation made | |
706 | * at least by solaris and freebsd. "Erratic ACKs" has _nothing_ | |
707 | * to do with delayed acks, because at cwnd>2 true delack timeout | |
708 | * is invisible. Actually, Linux-2.4 also generates erratic | |
caa20d9a | 709 | * ACKs in some circumstances. |
1da177e4 | 710 | */ |
f1ecd5d9 | 711 | inet_csk(sk)->icsk_rto = __tcp_set_rto(tp); |
1da177e4 LT |
712 | |
713 | /* 2. Fixups made earlier cannot be right. | |
714 | * If we do not estimate RTO correctly without them, | |
715 | * all the algo is pure shit and should be replaced | |
caa20d9a | 716 | * with correct one. It is exactly, which we pretend to do. |
1da177e4 | 717 | */ |
1da177e4 | 718 | |
ee6aac59 IJ |
719 | /* NOTE: clamping at TCP_RTO_MIN is not required, current algo |
720 | * guarantees that rto is higher. | |
721 | */ | |
f1ecd5d9 | 722 | tcp_bound_rto(sk); |
1da177e4 LT |
723 | } |
724 | ||
725 | /* Save metrics learned by this TCP session. | |
726 | This function is called only, when TCP finishes successfully | |
727 | i.e. when it enters TIME-WAIT or goes from LAST-ACK to CLOSE. | |
728 | */ | |
729 | void tcp_update_metrics(struct sock *sk) | |
730 | { | |
731 | struct tcp_sock *tp = tcp_sk(sk); | |
732 | struct dst_entry *dst = __sk_dst_get(sk); | |
733 | ||
734 | if (sysctl_tcp_nometrics_save) | |
735 | return; | |
736 | ||
737 | dst_confirm(dst); | |
738 | ||
056834d9 | 739 | if (dst && (dst->flags & DST_HOST)) { |
6687e988 | 740 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 741 | int m; |
c1e20f7c | 742 | unsigned long rtt; |
1da177e4 | 743 | |
6687e988 | 744 | if (icsk->icsk_backoff || !tp->srtt) { |
1da177e4 LT |
745 | /* This session failed to estimate rtt. Why? |
746 | * Probably, no packets returned in time. | |
747 | * Reset our results. | |
748 | */ | |
749 | if (!(dst_metric_locked(dst, RTAX_RTT))) | |
defb3519 | 750 | dst_metric_set(dst, RTAX_RTT, 0); |
1da177e4 LT |
751 | return; |
752 | } | |
753 | ||
c1e20f7c SH |
754 | rtt = dst_metric_rtt(dst, RTAX_RTT); |
755 | m = rtt - tp->srtt; | |
1da177e4 LT |
756 | |
757 | /* If newly calculated rtt larger than stored one, | |
758 | * store new one. Otherwise, use EWMA. Remember, | |
759 | * rtt overestimation is always better than underestimation. | |
760 | */ | |
761 | if (!(dst_metric_locked(dst, RTAX_RTT))) { | |
762 | if (m <= 0) | |
c1e20f7c | 763 | set_dst_metric_rtt(dst, RTAX_RTT, tp->srtt); |
1da177e4 | 764 | else |
c1e20f7c | 765 | set_dst_metric_rtt(dst, RTAX_RTT, rtt - (m >> 3)); |
1da177e4 LT |
766 | } |
767 | ||
768 | if (!(dst_metric_locked(dst, RTAX_RTTVAR))) { | |
c1e20f7c | 769 | unsigned long var; |
1da177e4 LT |
770 | if (m < 0) |
771 | m = -m; | |
772 | ||
773 | /* Scale deviation to rttvar fixed point */ | |
774 | m >>= 1; | |
775 | if (m < tp->mdev) | |
776 | m = tp->mdev; | |
777 | ||
c1e20f7c SH |
778 | var = dst_metric_rtt(dst, RTAX_RTTVAR); |
779 | if (m >= var) | |
780 | var = m; | |
1da177e4 | 781 | else |
c1e20f7c SH |
782 | var -= (var - m) >> 2; |
783 | ||
784 | set_dst_metric_rtt(dst, RTAX_RTTVAR, var); | |
1da177e4 LT |
785 | } |
786 | ||
0b6a05c1 | 787 | if (tcp_in_initial_slowstart(tp)) { |
1da177e4 LT |
788 | /* Slow start still did not finish. */ |
789 | if (dst_metric(dst, RTAX_SSTHRESH) && | |
790 | !dst_metric_locked(dst, RTAX_SSTHRESH) && | |
791 | (tp->snd_cwnd >> 1) > dst_metric(dst, RTAX_SSTHRESH)) | |
defb3519 | 792 | dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_cwnd >> 1); |
1da177e4 LT |
793 | if (!dst_metric_locked(dst, RTAX_CWND) && |
794 | tp->snd_cwnd > dst_metric(dst, RTAX_CWND)) | |
defb3519 | 795 | dst_metric_set(dst, RTAX_CWND, tp->snd_cwnd); |
1da177e4 | 796 | } else if (tp->snd_cwnd > tp->snd_ssthresh && |
6687e988 | 797 | icsk->icsk_ca_state == TCP_CA_Open) { |
1da177e4 LT |
798 | /* Cong. avoidance phase, cwnd is reliable. */ |
799 | if (!dst_metric_locked(dst, RTAX_SSTHRESH)) | |
defb3519 DM |
800 | dst_metric_set(dst, RTAX_SSTHRESH, |
801 | max(tp->snd_cwnd >> 1, tp->snd_ssthresh)); | |
1da177e4 | 802 | if (!dst_metric_locked(dst, RTAX_CWND)) |
defb3519 DM |
803 | dst_metric_set(dst, RTAX_CWND, |
804 | (dst_metric(dst, RTAX_CWND) + | |
805 | tp->snd_cwnd) >> 1); | |
1da177e4 LT |
806 | } else { |
807 | /* Else slow start did not finish, cwnd is non-sense, | |
808 | ssthresh may be also invalid. | |
809 | */ | |
810 | if (!dst_metric_locked(dst, RTAX_CWND)) | |
defb3519 DM |
811 | dst_metric_set(dst, RTAX_CWND, |
812 | (dst_metric(dst, RTAX_CWND) + | |
813 | tp->snd_ssthresh) >> 1); | |
5ffc02a1 | 814 | if (dst_metric(dst, RTAX_SSTHRESH) && |
1da177e4 | 815 | !dst_metric_locked(dst, RTAX_SSTHRESH) && |
5ffc02a1 | 816 | tp->snd_ssthresh > dst_metric(dst, RTAX_SSTHRESH)) |
defb3519 | 817 | dst_metric_set(dst, RTAX_SSTHRESH, tp->snd_ssthresh); |
1da177e4 LT |
818 | } |
819 | ||
820 | if (!dst_metric_locked(dst, RTAX_REORDERING)) { | |
5ffc02a1 | 821 | if (dst_metric(dst, RTAX_REORDERING) < tp->reordering && |
1da177e4 | 822 | tp->reordering != sysctl_tcp_reordering) |
defb3519 | 823 | dst_metric_set(dst, RTAX_REORDERING, tp->reordering); |
1da177e4 LT |
824 | } |
825 | } | |
826 | } | |
827 | ||
cf533ea5 | 828 | __u32 tcp_init_cwnd(const struct tcp_sock *tp, const struct dst_entry *dst) |
1da177e4 LT |
829 | { |
830 | __u32 cwnd = (dst ? dst_metric(dst, RTAX_INITCWND) : 0); | |
831 | ||
22b71c8f | 832 | if (!cwnd) |
442b9635 | 833 | cwnd = TCP_INIT_CWND; |
1da177e4 LT |
834 | return min_t(__u32, cwnd, tp->snd_cwnd_clamp); |
835 | } | |
836 | ||
40efc6fa | 837 | /* Set slow start threshold and cwnd not falling to slow start */ |
3cfe3baa | 838 | void tcp_enter_cwr(struct sock *sk, const int set_ssthresh) |
40efc6fa SH |
839 | { |
840 | struct tcp_sock *tp = tcp_sk(sk); | |
3cfe3baa | 841 | const struct inet_connection_sock *icsk = inet_csk(sk); |
40efc6fa SH |
842 | |
843 | tp->prior_ssthresh = 0; | |
844 | tp->bytes_acked = 0; | |
e01f9d77 | 845 | if (icsk->icsk_ca_state < TCP_CA_CWR) { |
40efc6fa | 846 | tp->undo_marker = 0; |
3cfe3baa IJ |
847 | if (set_ssthresh) |
848 | tp->snd_ssthresh = icsk->icsk_ca_ops->ssthresh(sk); | |
40efc6fa SH |
849 | tp->snd_cwnd = min(tp->snd_cwnd, |
850 | tcp_packets_in_flight(tp) + 1U); | |
851 | tp->snd_cwnd_cnt = 0; | |
852 | tp->high_seq = tp->snd_nxt; | |
853 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
854 | TCP_ECN_queue_cwr(tp); | |
855 | ||
856 | tcp_set_ca_state(sk, TCP_CA_CWR); | |
857 | } | |
858 | } | |
859 | ||
e60402d0 IJ |
860 | /* |
861 | * Packet counting of FACK is based on in-order assumptions, therefore TCP | |
862 | * disables it when reordering is detected | |
863 | */ | |
864 | static void tcp_disable_fack(struct tcp_sock *tp) | |
865 | { | |
85cc391c IJ |
866 | /* RFC3517 uses different metric in lost marker => reset on change */ |
867 | if (tcp_is_fack(tp)) | |
868 | tp->lost_skb_hint = NULL; | |
ab56222a | 869 | tp->rx_opt.sack_ok &= ~TCP_FACK_ENABLED; |
e60402d0 IJ |
870 | } |
871 | ||
564262c1 | 872 | /* Take a notice that peer is sending D-SACKs */ |
e60402d0 IJ |
873 | static void tcp_dsack_seen(struct tcp_sock *tp) |
874 | { | |
ab56222a | 875 | tp->rx_opt.sack_ok |= TCP_DSACK_SEEN; |
e60402d0 IJ |
876 | } |
877 | ||
1da177e4 LT |
878 | /* Initialize metrics on socket. */ |
879 | ||
880 | static void tcp_init_metrics(struct sock *sk) | |
881 | { | |
882 | struct tcp_sock *tp = tcp_sk(sk); | |
883 | struct dst_entry *dst = __sk_dst_get(sk); | |
884 | ||
885 | if (dst == NULL) | |
886 | goto reset; | |
887 | ||
888 | dst_confirm(dst); | |
889 | ||
890 | if (dst_metric_locked(dst, RTAX_CWND)) | |
891 | tp->snd_cwnd_clamp = dst_metric(dst, RTAX_CWND); | |
892 | if (dst_metric(dst, RTAX_SSTHRESH)) { | |
893 | tp->snd_ssthresh = dst_metric(dst, RTAX_SSTHRESH); | |
894 | if (tp->snd_ssthresh > tp->snd_cwnd_clamp) | |
895 | tp->snd_ssthresh = tp->snd_cwnd_clamp; | |
9ad7c049 JC |
896 | } else { |
897 | /* ssthresh may have been reduced unnecessarily during. | |
898 | * 3WHS. Restore it back to its initial default. | |
899 | */ | |
900 | tp->snd_ssthresh = TCP_INFINITE_SSTHRESH; | |
1da177e4 LT |
901 | } |
902 | if (dst_metric(dst, RTAX_REORDERING) && | |
903 | tp->reordering != dst_metric(dst, RTAX_REORDERING)) { | |
e60402d0 | 904 | tcp_disable_fack(tp); |
1da177e4 LT |
905 | tp->reordering = dst_metric(dst, RTAX_REORDERING); |
906 | } | |
907 | ||
9ad7c049 | 908 | if (dst_metric(dst, RTAX_RTT) == 0 || tp->srtt == 0) |
1da177e4 LT |
909 | goto reset; |
910 | ||
911 | /* Initial rtt is determined from SYN,SYN-ACK. | |
912 | * The segment is small and rtt may appear much | |
913 | * less than real one. Use per-dst memory | |
914 | * to make it more realistic. | |
915 | * | |
916 | * A bit of theory. RTT is time passed after "normal" sized packet | |
caa20d9a | 917 | * is sent until it is ACKed. In normal circumstances sending small |
1da177e4 LT |
918 | * packets force peer to delay ACKs and calculation is correct too. |
919 | * The algorithm is adaptive and, provided we follow specs, it | |
920 | * NEVER underestimate RTT. BUT! If peer tries to make some clever | |
921 | * tricks sort of "quick acks" for time long enough to decrease RTT | |
922 | * to low value, and then abruptly stops to do it and starts to delay | |
923 | * ACKs, wait for troubles. | |
924 | */ | |
c1e20f7c SH |
925 | if (dst_metric_rtt(dst, RTAX_RTT) > tp->srtt) { |
926 | tp->srtt = dst_metric_rtt(dst, RTAX_RTT); | |
1da177e4 LT |
927 | tp->rtt_seq = tp->snd_nxt; |
928 | } | |
c1e20f7c SH |
929 | if (dst_metric_rtt(dst, RTAX_RTTVAR) > tp->mdev) { |
930 | tp->mdev = dst_metric_rtt(dst, RTAX_RTTVAR); | |
488faa2a | 931 | tp->mdev_max = tp->rttvar = max(tp->mdev, tcp_rto_min(sk)); |
1da177e4 | 932 | } |
463c84b9 | 933 | tcp_set_rto(sk); |
1da177e4 | 934 | reset: |
9ad7c049 JC |
935 | if (tp->srtt == 0) { |
936 | /* RFC2988bis: We've failed to get a valid RTT sample from | |
937 | * 3WHS. This is most likely due to retransmission, | |
938 | * including spurious one. Reset the RTO back to 3secs | |
939 | * from the more aggressive 1sec to avoid more spurious | |
940 | * retransmission. | |
d9f4fbaf | 941 | */ |
9ad7c049 JC |
942 | tp->mdev = tp->mdev_max = tp->rttvar = TCP_TIMEOUT_FALLBACK; |
943 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_FALLBACK; | |
1da177e4 | 944 | } |
9ad7c049 JC |
945 | /* Cut cwnd down to 1 per RFC5681 if SYN or SYN-ACK has been |
946 | * retransmitted. In light of RFC2988bis' more aggressive 1sec | |
947 | * initRTO, we only reset cwnd when more than 1 SYN/SYN-ACK | |
948 | * retransmission has occurred. | |
949 | */ | |
950 | if (tp->total_retrans > 1) | |
951 | tp->snd_cwnd = 1; | |
952 | else | |
953 | tp->snd_cwnd = tcp_init_cwnd(tp, dst); | |
d9f4fbaf | 954 | tp->snd_cwnd_stamp = tcp_time_stamp; |
1da177e4 LT |
955 | } |
956 | ||
6687e988 ACM |
957 | static void tcp_update_reordering(struct sock *sk, const int metric, |
958 | const int ts) | |
1da177e4 | 959 | { |
6687e988 | 960 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 | 961 | if (metric > tp->reordering) { |
40b215e5 PE |
962 | int mib_idx; |
963 | ||
1da177e4 LT |
964 | tp->reordering = min(TCP_MAX_REORDERING, metric); |
965 | ||
966 | /* This exciting event is worth to be remembered. 8) */ | |
967 | if (ts) | |
40b215e5 | 968 | mib_idx = LINUX_MIB_TCPTSREORDER; |
e60402d0 | 969 | else if (tcp_is_reno(tp)) |
40b215e5 | 970 | mib_idx = LINUX_MIB_TCPRENOREORDER; |
e60402d0 | 971 | else if (tcp_is_fack(tp)) |
40b215e5 | 972 | mib_idx = LINUX_MIB_TCPFACKREORDER; |
1da177e4 | 973 | else |
40b215e5 PE |
974 | mib_idx = LINUX_MIB_TCPSACKREORDER; |
975 | ||
de0744af | 976 | NET_INC_STATS_BH(sock_net(sk), mib_idx); |
1da177e4 LT |
977 | #if FASTRETRANS_DEBUG > 1 |
978 | printk(KERN_DEBUG "Disorder%d %d %u f%u s%u rr%d\n", | |
6687e988 | 979 | tp->rx_opt.sack_ok, inet_csk(sk)->icsk_ca_state, |
1da177e4 LT |
980 | tp->reordering, |
981 | tp->fackets_out, | |
982 | tp->sacked_out, | |
983 | tp->undo_marker ? tp->undo_retrans : 0); | |
984 | #endif | |
e60402d0 | 985 | tcp_disable_fack(tp); |
1da177e4 LT |
986 | } |
987 | } | |
988 | ||
006f582c | 989 | /* This must be called before lost_out is incremented */ |
c8c213f2 IJ |
990 | static void tcp_verify_retransmit_hint(struct tcp_sock *tp, struct sk_buff *skb) |
991 | { | |
006f582c | 992 | if ((tp->retransmit_skb_hint == NULL) || |
c8c213f2 IJ |
993 | before(TCP_SKB_CB(skb)->seq, |
994 | TCP_SKB_CB(tp->retransmit_skb_hint)->seq)) | |
006f582c IJ |
995 | tp->retransmit_skb_hint = skb; |
996 | ||
997 | if (!tp->lost_out || | |
998 | after(TCP_SKB_CB(skb)->end_seq, tp->retransmit_high)) | |
999 | tp->retransmit_high = TCP_SKB_CB(skb)->end_seq; | |
c8c213f2 IJ |
1000 | } |
1001 | ||
41ea36e3 IJ |
1002 | static void tcp_skb_mark_lost(struct tcp_sock *tp, struct sk_buff *skb) |
1003 | { | |
1004 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { | |
1005 | tcp_verify_retransmit_hint(tp, skb); | |
1006 | ||
1007 | tp->lost_out += tcp_skb_pcount(skb); | |
1008 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1009 | } | |
1010 | } | |
1011 | ||
e1aa680f IJ |
1012 | static void tcp_skb_mark_lost_uncond_verify(struct tcp_sock *tp, |
1013 | struct sk_buff *skb) | |
006f582c IJ |
1014 | { |
1015 | tcp_verify_retransmit_hint(tp, skb); | |
1016 | ||
1017 | if (!(TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_ACKED))) { | |
1018 | tp->lost_out += tcp_skb_pcount(skb); | |
1019 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1020 | } | |
1021 | } | |
1022 | ||
1da177e4 LT |
1023 | /* This procedure tags the retransmission queue when SACKs arrive. |
1024 | * | |
1025 | * We have three tag bits: SACKED(S), RETRANS(R) and LOST(L). | |
1026 | * Packets in queue with these bits set are counted in variables | |
1027 | * sacked_out, retrans_out and lost_out, correspondingly. | |
1028 | * | |
1029 | * Valid combinations are: | |
1030 | * Tag InFlight Description | |
1031 | * 0 1 - orig segment is in flight. | |
1032 | * S 0 - nothing flies, orig reached receiver. | |
1033 | * L 0 - nothing flies, orig lost by net. | |
1034 | * R 2 - both orig and retransmit are in flight. | |
1035 | * L|R 1 - orig is lost, retransmit is in flight. | |
1036 | * S|R 1 - orig reached receiver, retrans is still in flight. | |
1037 | * (L|S|R is logically valid, it could occur when L|R is sacked, | |
1038 | * but it is equivalent to plain S and code short-curcuits it to S. | |
1039 | * L|S is logically invalid, it would mean -1 packet in flight 8)) | |
1040 | * | |
1041 | * These 6 states form finite state machine, controlled by the following events: | |
1042 | * 1. New ACK (+SACK) arrives. (tcp_sacktag_write_queue()) | |
1043 | * 2. Retransmission. (tcp_retransmit_skb(), tcp_xmit_retransmit_queue()) | |
974c1236 | 1044 | * 3. Loss detection event of two flavors: |
1da177e4 LT |
1045 | * A. Scoreboard estimator decided the packet is lost. |
1046 | * A'. Reno "three dupacks" marks head of queue lost. | |
974c1236 YC |
1047 | * A''. Its FACK modification, head until snd.fack is lost. |
1048 | * B. SACK arrives sacking SND.NXT at the moment, when the | |
1da177e4 LT |
1049 | * segment was retransmitted. |
1050 | * 4. D-SACK added new rule: D-SACK changes any tag to S. | |
1051 | * | |
1052 | * It is pleasant to note, that state diagram turns out to be commutative, | |
1053 | * so that we are allowed not to be bothered by order of our actions, | |
1054 | * when multiple events arrive simultaneously. (see the function below). | |
1055 | * | |
1056 | * Reordering detection. | |
1057 | * -------------------- | |
1058 | * Reordering metric is maximal distance, which a packet can be displaced | |
1059 | * in packet stream. With SACKs we can estimate it: | |
1060 | * | |
1061 | * 1. SACK fills old hole and the corresponding segment was not | |
1062 | * ever retransmitted -> reordering. Alas, we cannot use it | |
1063 | * when segment was retransmitted. | |
1064 | * 2. The last flaw is solved with D-SACK. D-SACK arrives | |
1065 | * for retransmitted and already SACKed segment -> reordering.. | |
1066 | * Both of these heuristics are not used in Loss state, when we cannot | |
1067 | * account for retransmits accurately. | |
5b3c9882 IJ |
1068 | * |
1069 | * SACK block validation. | |
1070 | * ---------------------- | |
1071 | * | |
1072 | * SACK block range validation checks that the received SACK block fits to | |
1073 | * the expected sequence limits, i.e., it is between SND.UNA and SND.NXT. | |
1074 | * Note that SND.UNA is not included to the range though being valid because | |
0e835331 IJ |
1075 | * it means that the receiver is rather inconsistent with itself reporting |
1076 | * SACK reneging when it should advance SND.UNA. Such SACK block this is | |
1077 | * perfectly valid, however, in light of RFC2018 which explicitly states | |
1078 | * that "SACK block MUST reflect the newest segment. Even if the newest | |
1079 | * segment is going to be discarded ...", not that it looks very clever | |
1080 | * in case of head skb. Due to potentional receiver driven attacks, we | |
1081 | * choose to avoid immediate execution of a walk in write queue due to | |
1082 | * reneging and defer head skb's loss recovery to standard loss recovery | |
1083 | * procedure that will eventually trigger (nothing forbids us doing this). | |
5b3c9882 IJ |
1084 | * |
1085 | * Implements also blockage to start_seq wrap-around. Problem lies in the | |
1086 | * fact that though start_seq (s) is before end_seq (i.e., not reversed), | |
1087 | * there's no guarantee that it will be before snd_nxt (n). The problem | |
1088 | * happens when start_seq resides between end_seq wrap (e_w) and snd_nxt | |
1089 | * wrap (s_w): | |
1090 | * | |
1091 | * <- outs wnd -> <- wrapzone -> | |
1092 | * u e n u_w e_w s n_w | |
1093 | * | | | | | | | | |
1094 | * |<------------+------+----- TCP seqno space --------------+---------->| | |
1095 | * ...-- <2^31 ->| |<--------... | |
1096 | * ...---- >2^31 ------>| |<--------... | |
1097 | * | |
1098 | * Current code wouldn't be vulnerable but it's better still to discard such | |
1099 | * crazy SACK blocks. Doing this check for start_seq alone closes somewhat | |
1100 | * similar case (end_seq after snd_nxt wrap) as earlier reversed check in | |
1101 | * snd_nxt wrap -> snd_una region will then become "well defined", i.e., | |
1102 | * equal to the ideal case (infinite seqno space without wrap caused issues). | |
1103 | * | |
1104 | * With D-SACK the lower bound is extended to cover sequence space below | |
1105 | * SND.UNA down to undo_marker, which is the last point of interest. Yet | |
564262c1 | 1106 | * again, D-SACK block must not to go across snd_una (for the same reason as |
5b3c9882 IJ |
1107 | * for the normal SACK blocks, explained above). But there all simplicity |
1108 | * ends, TCP might receive valid D-SACKs below that. As long as they reside | |
1109 | * fully below undo_marker they do not affect behavior in anyway and can | |
1110 | * therefore be safely ignored. In rare cases (which are more or less | |
1111 | * theoretical ones), the D-SACK will nicely cross that boundary due to skb | |
1112 | * fragmentation and packet reordering past skb's retransmission. To consider | |
1113 | * them correctly, the acceptable range must be extended even more though | |
1114 | * the exact amount is rather hard to quantify. However, tp->max_window can | |
1115 | * be used as an exaggerated estimate. | |
1da177e4 | 1116 | */ |
5b3c9882 IJ |
1117 | static int tcp_is_sackblock_valid(struct tcp_sock *tp, int is_dsack, |
1118 | u32 start_seq, u32 end_seq) | |
1119 | { | |
1120 | /* Too far in future, or reversed (interpretation is ambiguous) */ | |
1121 | if (after(end_seq, tp->snd_nxt) || !before(start_seq, end_seq)) | |
1122 | return 0; | |
1123 | ||
1124 | /* Nasty start_seq wrap-around check (see comments above) */ | |
1125 | if (!before(start_seq, tp->snd_nxt)) | |
1126 | return 0; | |
1127 | ||
564262c1 | 1128 | /* In outstanding window? ...This is valid exit for D-SACKs too. |
5b3c9882 IJ |
1129 | * start_seq == snd_una is non-sensical (see comments above) |
1130 | */ | |
1131 | if (after(start_seq, tp->snd_una)) | |
1132 | return 1; | |
1133 | ||
1134 | if (!is_dsack || !tp->undo_marker) | |
1135 | return 0; | |
1136 | ||
1137 | /* ...Then it's D-SACK, and must reside below snd_una completely */ | |
f779b2d6 | 1138 | if (after(end_seq, tp->snd_una)) |
5b3c9882 IJ |
1139 | return 0; |
1140 | ||
1141 | if (!before(start_seq, tp->undo_marker)) | |
1142 | return 1; | |
1143 | ||
1144 | /* Too old */ | |
1145 | if (!after(end_seq, tp->undo_marker)) | |
1146 | return 0; | |
1147 | ||
1148 | /* Undo_marker boundary crossing (overestimates a lot). Known already: | |
1149 | * start_seq < undo_marker and end_seq >= undo_marker. | |
1150 | */ | |
1151 | return !before(start_seq, end_seq - tp->max_window); | |
1152 | } | |
1153 | ||
1c1e87ed | 1154 | /* Check for lost retransmit. This superb idea is borrowed from "ratehalving". |
974c1236 | 1155 | * Event "B". Later note: FACK people cheated me again 8), we have to account |
1c1e87ed | 1156 | * for reordering! Ugly, but should help. |
f785a8e2 IJ |
1157 | * |
1158 | * Search retransmitted skbs from write_queue that were sent when snd_nxt was | |
1159 | * less than what is now known to be received by the other end (derived from | |
9f58f3b7 IJ |
1160 | * highest SACK block). Also calculate the lowest snd_nxt among the remaining |
1161 | * retransmitted skbs to avoid some costly processing per ACKs. | |
1c1e87ed | 1162 | */ |
407ef1de | 1163 | static void tcp_mark_lost_retrans(struct sock *sk) |
1c1e87ed | 1164 | { |
9f58f3b7 | 1165 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1c1e87ed IJ |
1166 | struct tcp_sock *tp = tcp_sk(sk); |
1167 | struct sk_buff *skb; | |
f785a8e2 | 1168 | int cnt = 0; |
df2e014b | 1169 | u32 new_low_seq = tp->snd_nxt; |
6859d494 | 1170 | u32 received_upto = tcp_highest_sack_seq(tp); |
9f58f3b7 IJ |
1171 | |
1172 | if (!tcp_is_fack(tp) || !tp->retrans_out || | |
1173 | !after(received_upto, tp->lost_retrans_low) || | |
1174 | icsk->icsk_ca_state != TCP_CA_Recovery) | |
407ef1de | 1175 | return; |
1c1e87ed IJ |
1176 | |
1177 | tcp_for_write_queue(skb, sk) { | |
1178 | u32 ack_seq = TCP_SKB_CB(skb)->ack_seq; | |
1179 | ||
1180 | if (skb == tcp_send_head(sk)) | |
1181 | break; | |
f785a8e2 | 1182 | if (cnt == tp->retrans_out) |
1c1e87ed IJ |
1183 | break; |
1184 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1185 | continue; | |
1186 | ||
f785a8e2 IJ |
1187 | if (!(TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)) |
1188 | continue; | |
1189 | ||
d0af4160 IJ |
1190 | /* TODO: We would like to get rid of tcp_is_fack(tp) only |
1191 | * constraint here (see above) but figuring out that at | |
1192 | * least tp->reordering SACK blocks reside between ack_seq | |
1193 | * and received_upto is not easy task to do cheaply with | |
1194 | * the available datastructures. | |
1195 | * | |
1196 | * Whether FACK should check here for tp->reordering segs | |
1197 | * in-between one could argue for either way (it would be | |
1198 | * rather simple to implement as we could count fack_count | |
1199 | * during the walk and do tp->fackets_out - fack_count). | |
1200 | */ | |
1201 | if (after(received_upto, ack_seq)) { | |
1c1e87ed IJ |
1202 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; |
1203 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1204 | ||
006f582c | 1205 | tcp_skb_mark_lost_uncond_verify(tp, skb); |
de0744af | 1206 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSTRETRANSMIT); |
f785a8e2 | 1207 | } else { |
df2e014b | 1208 | if (before(ack_seq, new_low_seq)) |
b08d6cb2 | 1209 | new_low_seq = ack_seq; |
f785a8e2 | 1210 | cnt += tcp_skb_pcount(skb); |
1c1e87ed IJ |
1211 | } |
1212 | } | |
b08d6cb2 IJ |
1213 | |
1214 | if (tp->retrans_out) | |
1215 | tp->lost_retrans_low = new_low_seq; | |
1c1e87ed | 1216 | } |
5b3c9882 | 1217 | |
cf533ea5 | 1218 | static int tcp_check_dsack(struct sock *sk, const struct sk_buff *ack_skb, |
d06e021d DM |
1219 | struct tcp_sack_block_wire *sp, int num_sacks, |
1220 | u32 prior_snd_una) | |
1221 | { | |
1ed83465 | 1222 | struct tcp_sock *tp = tcp_sk(sk); |
d3e2ce3b HH |
1223 | u32 start_seq_0 = get_unaligned_be32(&sp[0].start_seq); |
1224 | u32 end_seq_0 = get_unaligned_be32(&sp[0].end_seq); | |
d06e021d DM |
1225 | int dup_sack = 0; |
1226 | ||
1227 | if (before(start_seq_0, TCP_SKB_CB(ack_skb)->ack_seq)) { | |
1228 | dup_sack = 1; | |
e60402d0 | 1229 | tcp_dsack_seen(tp); |
de0744af | 1230 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPDSACKRECV); |
d06e021d | 1231 | } else if (num_sacks > 1) { |
d3e2ce3b HH |
1232 | u32 end_seq_1 = get_unaligned_be32(&sp[1].end_seq); |
1233 | u32 start_seq_1 = get_unaligned_be32(&sp[1].start_seq); | |
d06e021d DM |
1234 | |
1235 | if (!after(end_seq_0, end_seq_1) && | |
1236 | !before(start_seq_0, start_seq_1)) { | |
1237 | dup_sack = 1; | |
e60402d0 | 1238 | tcp_dsack_seen(tp); |
de0744af PE |
1239 | NET_INC_STATS_BH(sock_net(sk), |
1240 | LINUX_MIB_TCPDSACKOFORECV); | |
d06e021d DM |
1241 | } |
1242 | } | |
1243 | ||
1244 | /* D-SACK for already forgotten data... Do dumb counting. */ | |
c24f691b | 1245 | if (dup_sack && tp->undo_marker && tp->undo_retrans && |
d06e021d DM |
1246 | !after(end_seq_0, prior_snd_una) && |
1247 | after(end_seq_0, tp->undo_marker)) | |
1248 | tp->undo_retrans--; | |
1249 | ||
1250 | return dup_sack; | |
1251 | } | |
1252 | ||
a1197f5a IJ |
1253 | struct tcp_sacktag_state { |
1254 | int reord; | |
1255 | int fack_count; | |
1256 | int flag; | |
1257 | }; | |
1258 | ||
d1935942 IJ |
1259 | /* Check if skb is fully within the SACK block. In presence of GSO skbs, |
1260 | * the incoming SACK may not exactly match but we can find smaller MSS | |
1261 | * aligned portion of it that matches. Therefore we might need to fragment | |
1262 | * which may fail and creates some hassle (caller must handle error case | |
1263 | * returns). | |
832d11c5 IJ |
1264 | * |
1265 | * FIXME: this could be merged to shift decision code | |
d1935942 | 1266 | */ |
0f79efdc AB |
1267 | static int tcp_match_skb_to_sack(struct sock *sk, struct sk_buff *skb, |
1268 | u32 start_seq, u32 end_seq) | |
d1935942 IJ |
1269 | { |
1270 | int in_sack, err; | |
1271 | unsigned int pkt_len; | |
adb92db8 | 1272 | unsigned int mss; |
d1935942 IJ |
1273 | |
1274 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && | |
1275 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1276 | ||
1277 | if (tcp_skb_pcount(skb) > 1 && !in_sack && | |
1278 | after(TCP_SKB_CB(skb)->end_seq, start_seq)) { | |
adb92db8 | 1279 | mss = tcp_skb_mss(skb); |
d1935942 IJ |
1280 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); |
1281 | ||
adb92db8 | 1282 | if (!in_sack) { |
d1935942 | 1283 | pkt_len = start_seq - TCP_SKB_CB(skb)->seq; |
adb92db8 IJ |
1284 | if (pkt_len < mss) |
1285 | pkt_len = mss; | |
1286 | } else { | |
d1935942 | 1287 | pkt_len = end_seq - TCP_SKB_CB(skb)->seq; |
adb92db8 IJ |
1288 | if (pkt_len < mss) |
1289 | return -EINVAL; | |
1290 | } | |
1291 | ||
1292 | /* Round if necessary so that SACKs cover only full MSSes | |
1293 | * and/or the remaining small portion (if present) | |
1294 | */ | |
1295 | if (pkt_len > mss) { | |
1296 | unsigned int new_len = (pkt_len / mss) * mss; | |
1297 | if (!in_sack && new_len < pkt_len) { | |
1298 | new_len += mss; | |
1299 | if (new_len > skb->len) | |
1300 | return 0; | |
1301 | } | |
1302 | pkt_len = new_len; | |
1303 | } | |
1304 | err = tcp_fragment(sk, skb, pkt_len, mss); | |
d1935942 IJ |
1305 | if (err < 0) |
1306 | return err; | |
1307 | } | |
1308 | ||
1309 | return in_sack; | |
1310 | } | |
1311 | ||
cc9a672e NC |
1312 | /* Mark the given newly-SACKed range as such, adjusting counters and hints. */ |
1313 | static u8 tcp_sacktag_one(struct sock *sk, | |
1314 | struct tcp_sacktag_state *state, u8 sacked, | |
1315 | u32 start_seq, u32 end_seq, | |
a1197f5a | 1316 | int dup_sack, int pcount) |
9e10c47c | 1317 | { |
6859d494 | 1318 | struct tcp_sock *tp = tcp_sk(sk); |
a1197f5a | 1319 | int fack_count = state->fack_count; |
9e10c47c IJ |
1320 | |
1321 | /* Account D-SACK for retransmitted packet. */ | |
1322 | if (dup_sack && (sacked & TCPCB_RETRANS)) { | |
c24f691b | 1323 | if (tp->undo_marker && tp->undo_retrans && |
cc9a672e | 1324 | after(end_seq, tp->undo_marker)) |
9e10c47c | 1325 | tp->undo_retrans--; |
ede9f3b1 | 1326 | if (sacked & TCPCB_SACKED_ACKED) |
a1197f5a | 1327 | state->reord = min(fack_count, state->reord); |
9e10c47c IJ |
1328 | } |
1329 | ||
1330 | /* Nothing to do; acked frame is about to be dropped (was ACKed). */ | |
cc9a672e | 1331 | if (!after(end_seq, tp->snd_una)) |
a1197f5a | 1332 | return sacked; |
9e10c47c IJ |
1333 | |
1334 | if (!(sacked & TCPCB_SACKED_ACKED)) { | |
1335 | if (sacked & TCPCB_SACKED_RETRANS) { | |
1336 | /* If the segment is not tagged as lost, | |
1337 | * we do not clear RETRANS, believing | |
1338 | * that retransmission is still in flight. | |
1339 | */ | |
1340 | if (sacked & TCPCB_LOST) { | |
a1197f5a | 1341 | sacked &= ~(TCPCB_LOST|TCPCB_SACKED_RETRANS); |
f58b22fd IJ |
1342 | tp->lost_out -= pcount; |
1343 | tp->retrans_out -= pcount; | |
9e10c47c IJ |
1344 | } |
1345 | } else { | |
1346 | if (!(sacked & TCPCB_RETRANS)) { | |
1347 | /* New sack for not retransmitted frame, | |
1348 | * which was in hole. It is reordering. | |
1349 | */ | |
cc9a672e | 1350 | if (before(start_seq, |
9e10c47c | 1351 | tcp_highest_sack_seq(tp))) |
a1197f5a IJ |
1352 | state->reord = min(fack_count, |
1353 | state->reord); | |
9e10c47c IJ |
1354 | |
1355 | /* SACK enhanced F-RTO (RFC4138; Appendix B) */ | |
cc9a672e | 1356 | if (!after(end_seq, tp->frto_highmark)) |
a1197f5a | 1357 | state->flag |= FLAG_ONLY_ORIG_SACKED; |
9e10c47c IJ |
1358 | } |
1359 | ||
1360 | if (sacked & TCPCB_LOST) { | |
a1197f5a | 1361 | sacked &= ~TCPCB_LOST; |
f58b22fd | 1362 | tp->lost_out -= pcount; |
9e10c47c IJ |
1363 | } |
1364 | } | |
1365 | ||
a1197f5a IJ |
1366 | sacked |= TCPCB_SACKED_ACKED; |
1367 | state->flag |= FLAG_DATA_SACKED; | |
f58b22fd | 1368 | tp->sacked_out += pcount; |
9e10c47c | 1369 | |
f58b22fd | 1370 | fack_count += pcount; |
9e10c47c IJ |
1371 | |
1372 | /* Lost marker hint past SACKed? Tweak RFC3517 cnt */ | |
1373 | if (!tcp_is_fack(tp) && (tp->lost_skb_hint != NULL) && | |
cc9a672e | 1374 | before(start_seq, TCP_SKB_CB(tp->lost_skb_hint)->seq)) |
f58b22fd | 1375 | tp->lost_cnt_hint += pcount; |
9e10c47c IJ |
1376 | |
1377 | if (fack_count > tp->fackets_out) | |
1378 | tp->fackets_out = fack_count; | |
9e10c47c IJ |
1379 | } |
1380 | ||
1381 | /* D-SACK. We can detect redundant retransmission in S|R and plain R | |
1382 | * frames and clear it. undo_retrans is decreased above, L|R frames | |
1383 | * are accounted above as well. | |
1384 | */ | |
a1197f5a IJ |
1385 | if (dup_sack && (sacked & TCPCB_SACKED_RETRANS)) { |
1386 | sacked &= ~TCPCB_SACKED_RETRANS; | |
f58b22fd | 1387 | tp->retrans_out -= pcount; |
9e10c47c IJ |
1388 | } |
1389 | ||
a1197f5a | 1390 | return sacked; |
9e10c47c IJ |
1391 | } |
1392 | ||
daef52ba NC |
1393 | /* Shift newly-SACKed bytes from this skb to the immediately previous |
1394 | * already-SACKed sk_buff. Mark the newly-SACKed bytes as such. | |
1395 | */ | |
50133161 | 1396 | static int tcp_shifted_skb(struct sock *sk, struct sk_buff *skb, |
a1197f5a | 1397 | struct tcp_sacktag_state *state, |
9ec06ff5 IJ |
1398 | unsigned int pcount, int shifted, int mss, |
1399 | int dup_sack) | |
832d11c5 IJ |
1400 | { |
1401 | struct tcp_sock *tp = tcp_sk(sk); | |
50133161 | 1402 | struct sk_buff *prev = tcp_write_queue_prev(sk, skb); |
daef52ba NC |
1403 | u32 start_seq = TCP_SKB_CB(skb)->seq; /* start of newly-SACKed */ |
1404 | u32 end_seq = start_seq + shifted; /* end of newly-SACKed */ | |
832d11c5 IJ |
1405 | |
1406 | BUG_ON(!pcount); | |
1407 | ||
4c90d3b3 NC |
1408 | /* Adjust counters and hints for the newly sacked sequence |
1409 | * range but discard the return value since prev is already | |
1410 | * marked. We must tag the range first because the seq | |
1411 | * advancement below implicitly advances | |
1412 | * tcp_highest_sack_seq() when skb is highest_sack. | |
1413 | */ | |
1414 | tcp_sacktag_one(sk, state, TCP_SKB_CB(skb)->sacked, | |
1415 | start_seq, end_seq, dup_sack, pcount); | |
1416 | ||
1417 | if (skb == tp->lost_skb_hint) | |
0af2a0d0 NC |
1418 | tp->lost_cnt_hint += pcount; |
1419 | ||
832d11c5 IJ |
1420 | TCP_SKB_CB(prev)->end_seq += shifted; |
1421 | TCP_SKB_CB(skb)->seq += shifted; | |
1422 | ||
1423 | skb_shinfo(prev)->gso_segs += pcount; | |
1424 | BUG_ON(skb_shinfo(skb)->gso_segs < pcount); | |
1425 | skb_shinfo(skb)->gso_segs -= pcount; | |
1426 | ||
1427 | /* When we're adding to gso_segs == 1, gso_size will be zero, | |
1428 | * in theory this shouldn't be necessary but as long as DSACK | |
1429 | * code can come after this skb later on it's better to keep | |
1430 | * setting gso_size to something. | |
1431 | */ | |
1432 | if (!skb_shinfo(prev)->gso_size) { | |
1433 | skb_shinfo(prev)->gso_size = mss; | |
1434 | skb_shinfo(prev)->gso_type = sk->sk_gso_type; | |
1435 | } | |
1436 | ||
1437 | /* CHECKME: To clear or not to clear? Mimics normal skb currently */ | |
1438 | if (skb_shinfo(skb)->gso_segs <= 1) { | |
1439 | skb_shinfo(skb)->gso_size = 0; | |
1440 | skb_shinfo(skb)->gso_type = 0; | |
1441 | } | |
1442 | ||
832d11c5 IJ |
1443 | /* Difference in this won't matter, both ACKed by the same cumul. ACK */ |
1444 | TCP_SKB_CB(prev)->sacked |= (TCP_SKB_CB(skb)->sacked & TCPCB_EVER_RETRANS); | |
1445 | ||
832d11c5 IJ |
1446 | if (skb->len > 0) { |
1447 | BUG_ON(!tcp_skb_pcount(skb)); | |
111cc8b9 | 1448 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTED); |
832d11c5 IJ |
1449 | return 0; |
1450 | } | |
1451 | ||
1452 | /* Whole SKB was eaten :-) */ | |
1453 | ||
92ee76b6 IJ |
1454 | if (skb == tp->retransmit_skb_hint) |
1455 | tp->retransmit_skb_hint = prev; | |
1456 | if (skb == tp->scoreboard_skb_hint) | |
1457 | tp->scoreboard_skb_hint = prev; | |
1458 | if (skb == tp->lost_skb_hint) { | |
1459 | tp->lost_skb_hint = prev; | |
1460 | tp->lost_cnt_hint -= tcp_skb_pcount(prev); | |
1461 | } | |
1462 | ||
4de075e0 | 1463 | TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(prev)->tcp_flags; |
832d11c5 IJ |
1464 | if (skb == tcp_highest_sack(sk)) |
1465 | tcp_advance_highest_sack(sk, skb); | |
1466 | ||
1467 | tcp_unlink_write_queue(skb, sk); | |
1468 | sk_wmem_free_skb(sk, skb); | |
1469 | ||
111cc8b9 IJ |
1470 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKMERGED); |
1471 | ||
832d11c5 IJ |
1472 | return 1; |
1473 | } | |
1474 | ||
1475 | /* I wish gso_size would have a bit more sane initialization than | |
1476 | * something-or-zero which complicates things | |
1477 | */ | |
cf533ea5 | 1478 | static int tcp_skb_seglen(const struct sk_buff *skb) |
832d11c5 | 1479 | { |
775ffabf | 1480 | return tcp_skb_pcount(skb) == 1 ? skb->len : tcp_skb_mss(skb); |
832d11c5 IJ |
1481 | } |
1482 | ||
1483 | /* Shifting pages past head area doesn't work */ | |
cf533ea5 | 1484 | static int skb_can_shift(const struct sk_buff *skb) |
832d11c5 IJ |
1485 | { |
1486 | return !skb_headlen(skb) && skb_is_nonlinear(skb); | |
1487 | } | |
1488 | ||
1489 | /* Try collapsing SACK blocks spanning across multiple skbs to a single | |
1490 | * skb. | |
1491 | */ | |
1492 | static struct sk_buff *tcp_shift_skb_data(struct sock *sk, struct sk_buff *skb, | |
a1197f5a | 1493 | struct tcp_sacktag_state *state, |
832d11c5 | 1494 | u32 start_seq, u32 end_seq, |
a1197f5a | 1495 | int dup_sack) |
832d11c5 IJ |
1496 | { |
1497 | struct tcp_sock *tp = tcp_sk(sk); | |
1498 | struct sk_buff *prev; | |
1499 | int mss; | |
1500 | int pcount = 0; | |
1501 | int len; | |
1502 | int in_sack; | |
1503 | ||
1504 | if (!sk_can_gso(sk)) | |
1505 | goto fallback; | |
1506 | ||
1507 | /* Normally R but no L won't result in plain S */ | |
1508 | if (!dup_sack && | |
9969ca5f | 1509 | (TCP_SKB_CB(skb)->sacked & (TCPCB_LOST|TCPCB_SACKED_RETRANS)) == TCPCB_SACKED_RETRANS) |
832d11c5 IJ |
1510 | goto fallback; |
1511 | if (!skb_can_shift(skb)) | |
1512 | goto fallback; | |
1513 | /* This frame is about to be dropped (was ACKed). */ | |
1514 | if (!after(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1515 | goto fallback; | |
1516 | ||
1517 | /* Can only happen with delayed DSACK + discard craziness */ | |
1518 | if (unlikely(skb == tcp_write_queue_head(sk))) | |
1519 | goto fallback; | |
1520 | prev = tcp_write_queue_prev(sk, skb); | |
1521 | ||
1522 | if ((TCP_SKB_CB(prev)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) | |
1523 | goto fallback; | |
1524 | ||
1525 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq) && | |
1526 | !before(end_seq, TCP_SKB_CB(skb)->end_seq); | |
1527 | ||
1528 | if (in_sack) { | |
1529 | len = skb->len; | |
1530 | pcount = tcp_skb_pcount(skb); | |
775ffabf | 1531 | mss = tcp_skb_seglen(skb); |
832d11c5 IJ |
1532 | |
1533 | /* TODO: Fix DSACKs to not fragment already SACKed and we can | |
1534 | * drop this restriction as unnecessary | |
1535 | */ | |
775ffabf | 1536 | if (mss != tcp_skb_seglen(prev)) |
832d11c5 IJ |
1537 | goto fallback; |
1538 | } else { | |
1539 | if (!after(TCP_SKB_CB(skb)->end_seq, start_seq)) | |
1540 | goto noop; | |
1541 | /* CHECKME: This is non-MSS split case only?, this will | |
1542 | * cause skipped skbs due to advancing loop btw, original | |
1543 | * has that feature too | |
1544 | */ | |
1545 | if (tcp_skb_pcount(skb) <= 1) | |
1546 | goto noop; | |
1547 | ||
1548 | in_sack = !after(start_seq, TCP_SKB_CB(skb)->seq); | |
1549 | if (!in_sack) { | |
1550 | /* TODO: head merge to next could be attempted here | |
1551 | * if (!after(TCP_SKB_CB(skb)->end_seq, end_seq)), | |
1552 | * though it might not be worth of the additional hassle | |
1553 | * | |
1554 | * ...we can probably just fallback to what was done | |
1555 | * previously. We could try merging non-SACKed ones | |
1556 | * as well but it probably isn't going to buy off | |
1557 | * because later SACKs might again split them, and | |
1558 | * it would make skb timestamp tracking considerably | |
1559 | * harder problem. | |
1560 | */ | |
1561 | goto fallback; | |
1562 | } | |
1563 | ||
1564 | len = end_seq - TCP_SKB_CB(skb)->seq; | |
1565 | BUG_ON(len < 0); | |
1566 | BUG_ON(len > skb->len); | |
1567 | ||
1568 | /* MSS boundaries should be honoured or else pcount will | |
1569 | * severely break even though it makes things bit trickier. | |
1570 | * Optimize common case to avoid most of the divides | |
1571 | */ | |
1572 | mss = tcp_skb_mss(skb); | |
1573 | ||
1574 | /* TODO: Fix DSACKs to not fragment already SACKed and we can | |
1575 | * drop this restriction as unnecessary | |
1576 | */ | |
775ffabf | 1577 | if (mss != tcp_skb_seglen(prev)) |
832d11c5 IJ |
1578 | goto fallback; |
1579 | ||
1580 | if (len == mss) { | |
1581 | pcount = 1; | |
1582 | } else if (len < mss) { | |
1583 | goto noop; | |
1584 | } else { | |
1585 | pcount = len / mss; | |
1586 | len = pcount * mss; | |
1587 | } | |
1588 | } | |
1589 | ||
4648dc97 NC |
1590 | /* tcp_sacktag_one() won't SACK-tag ranges below snd_una */ |
1591 | if (!after(TCP_SKB_CB(skb)->seq + len, tp->snd_una)) | |
1592 | goto fallback; | |
1593 | ||
832d11c5 IJ |
1594 | if (!skb_shift(prev, skb, len)) |
1595 | goto fallback; | |
9ec06ff5 | 1596 | if (!tcp_shifted_skb(sk, skb, state, pcount, len, mss, dup_sack)) |
832d11c5 IJ |
1597 | goto out; |
1598 | ||
1599 | /* Hole filled allows collapsing with the next as well, this is very | |
1600 | * useful when hole on every nth skb pattern happens | |
1601 | */ | |
1602 | if (prev == tcp_write_queue_tail(sk)) | |
1603 | goto out; | |
1604 | skb = tcp_write_queue_next(sk, prev); | |
1605 | ||
f0bc52f3 IJ |
1606 | if (!skb_can_shift(skb) || |
1607 | (skb == tcp_send_head(sk)) || | |
1608 | ((TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) != TCPCB_SACKED_ACKED) || | |
775ffabf | 1609 | (mss != tcp_skb_seglen(skb))) |
832d11c5 IJ |
1610 | goto out; |
1611 | ||
1612 | len = skb->len; | |
1613 | if (skb_shift(prev, skb, len)) { | |
1614 | pcount += tcp_skb_pcount(skb); | |
9ec06ff5 | 1615 | tcp_shifted_skb(sk, skb, state, tcp_skb_pcount(skb), len, mss, 0); |
832d11c5 IJ |
1616 | } |
1617 | ||
1618 | out: | |
a1197f5a | 1619 | state->fack_count += pcount; |
832d11c5 IJ |
1620 | return prev; |
1621 | ||
1622 | noop: | |
1623 | return skb; | |
1624 | ||
1625 | fallback: | |
111cc8b9 | 1626 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_SACKSHIFTFALLBACK); |
832d11c5 IJ |
1627 | return NULL; |
1628 | } | |
1629 | ||
68f8353b IJ |
1630 | static struct sk_buff *tcp_sacktag_walk(struct sk_buff *skb, struct sock *sk, |
1631 | struct tcp_sack_block *next_dup, | |
a1197f5a | 1632 | struct tcp_sacktag_state *state, |
68f8353b | 1633 | u32 start_seq, u32 end_seq, |
a1197f5a | 1634 | int dup_sack_in) |
68f8353b | 1635 | { |
832d11c5 IJ |
1636 | struct tcp_sock *tp = tcp_sk(sk); |
1637 | struct sk_buff *tmp; | |
1638 | ||
68f8353b IJ |
1639 | tcp_for_write_queue_from(skb, sk) { |
1640 | int in_sack = 0; | |
1641 | int dup_sack = dup_sack_in; | |
1642 | ||
1643 | if (skb == tcp_send_head(sk)) | |
1644 | break; | |
1645 | ||
1646 | /* queue is in-order => we can short-circuit the walk early */ | |
1647 | if (!before(TCP_SKB_CB(skb)->seq, end_seq)) | |
1648 | break; | |
1649 | ||
1650 | if ((next_dup != NULL) && | |
1651 | before(TCP_SKB_CB(skb)->seq, next_dup->end_seq)) { | |
1652 | in_sack = tcp_match_skb_to_sack(sk, skb, | |
1653 | next_dup->start_seq, | |
1654 | next_dup->end_seq); | |
1655 | if (in_sack > 0) | |
1656 | dup_sack = 1; | |
1657 | } | |
1658 | ||
832d11c5 IJ |
1659 | /* skb reference here is a bit tricky to get right, since |
1660 | * shifting can eat and free both this skb and the next, | |
1661 | * so not even _safe variant of the loop is enough. | |
1662 | */ | |
1663 | if (in_sack <= 0) { | |
a1197f5a IJ |
1664 | tmp = tcp_shift_skb_data(sk, skb, state, |
1665 | start_seq, end_seq, dup_sack); | |
832d11c5 IJ |
1666 | if (tmp != NULL) { |
1667 | if (tmp != skb) { | |
1668 | skb = tmp; | |
1669 | continue; | |
1670 | } | |
1671 | ||
1672 | in_sack = 0; | |
1673 | } else { | |
1674 | in_sack = tcp_match_skb_to_sack(sk, skb, | |
1675 | start_seq, | |
1676 | end_seq); | |
1677 | } | |
1678 | } | |
1679 | ||
68f8353b IJ |
1680 | if (unlikely(in_sack < 0)) |
1681 | break; | |
1682 | ||
832d11c5 | 1683 | if (in_sack) { |
cc9a672e NC |
1684 | TCP_SKB_CB(skb)->sacked = |
1685 | tcp_sacktag_one(sk, | |
1686 | state, | |
1687 | TCP_SKB_CB(skb)->sacked, | |
1688 | TCP_SKB_CB(skb)->seq, | |
1689 | TCP_SKB_CB(skb)->end_seq, | |
1690 | dup_sack, | |
1691 | tcp_skb_pcount(skb)); | |
68f8353b | 1692 | |
832d11c5 IJ |
1693 | if (!before(TCP_SKB_CB(skb)->seq, |
1694 | tcp_highest_sack_seq(tp))) | |
1695 | tcp_advance_highest_sack(sk, skb); | |
1696 | } | |
1697 | ||
a1197f5a | 1698 | state->fack_count += tcp_skb_pcount(skb); |
68f8353b IJ |
1699 | } |
1700 | return skb; | |
1701 | } | |
1702 | ||
1703 | /* Avoid all extra work that is being done by sacktag while walking in | |
1704 | * a normal way | |
1705 | */ | |
1706 | static struct sk_buff *tcp_sacktag_skip(struct sk_buff *skb, struct sock *sk, | |
a1197f5a IJ |
1707 | struct tcp_sacktag_state *state, |
1708 | u32 skip_to_seq) | |
68f8353b IJ |
1709 | { |
1710 | tcp_for_write_queue_from(skb, sk) { | |
1711 | if (skb == tcp_send_head(sk)) | |
1712 | break; | |
1713 | ||
e8bae275 | 1714 | if (after(TCP_SKB_CB(skb)->end_seq, skip_to_seq)) |
68f8353b | 1715 | break; |
d152a7d8 | 1716 | |
a1197f5a | 1717 | state->fack_count += tcp_skb_pcount(skb); |
68f8353b IJ |
1718 | } |
1719 | return skb; | |
1720 | } | |
1721 | ||
1722 | static struct sk_buff *tcp_maybe_skipping_dsack(struct sk_buff *skb, | |
1723 | struct sock *sk, | |
1724 | struct tcp_sack_block *next_dup, | |
a1197f5a IJ |
1725 | struct tcp_sacktag_state *state, |
1726 | u32 skip_to_seq) | |
68f8353b IJ |
1727 | { |
1728 | if (next_dup == NULL) | |
1729 | return skb; | |
1730 | ||
1731 | if (before(next_dup->start_seq, skip_to_seq)) { | |
a1197f5a IJ |
1732 | skb = tcp_sacktag_skip(skb, sk, state, next_dup->start_seq); |
1733 | skb = tcp_sacktag_walk(skb, sk, NULL, state, | |
1734 | next_dup->start_seq, next_dup->end_seq, | |
1735 | 1); | |
68f8353b IJ |
1736 | } |
1737 | ||
1738 | return skb; | |
1739 | } | |
1740 | ||
cf533ea5 | 1741 | static int tcp_sack_cache_ok(const struct tcp_sock *tp, const struct tcp_sack_block *cache) |
68f8353b IJ |
1742 | { |
1743 | return cache < tp->recv_sack_cache + ARRAY_SIZE(tp->recv_sack_cache); | |
1744 | } | |
1745 | ||
1da177e4 | 1746 | static int |
cf533ea5 | 1747 | tcp_sacktag_write_queue(struct sock *sk, const struct sk_buff *ack_skb, |
056834d9 | 1748 | u32 prior_snd_una) |
1da177e4 | 1749 | { |
6687e988 | 1750 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 1751 | struct tcp_sock *tp = tcp_sk(sk); |
cf533ea5 ED |
1752 | const unsigned char *ptr = (skb_transport_header(ack_skb) + |
1753 | TCP_SKB_CB(ack_skb)->sacked); | |
fd6dad61 | 1754 | struct tcp_sack_block_wire *sp_wire = (struct tcp_sack_block_wire *)(ptr+2); |
4389dded | 1755 | struct tcp_sack_block sp[TCP_NUM_SACKS]; |
68f8353b | 1756 | struct tcp_sack_block *cache; |
a1197f5a | 1757 | struct tcp_sacktag_state state; |
68f8353b | 1758 | struct sk_buff *skb; |
4389dded | 1759 | int num_sacks = min(TCP_NUM_SACKS, (ptr[1] - TCPOLEN_SACK_BASE) >> 3); |
fd6dad61 | 1760 | int used_sacks; |
7769f406 | 1761 | int found_dup_sack = 0; |
68f8353b | 1762 | int i, j; |
fda03fbb | 1763 | int first_sack_index; |
1da177e4 | 1764 | |
a1197f5a IJ |
1765 | state.flag = 0; |
1766 | state.reord = tp->packets_out; | |
1767 | ||
d738cd8f | 1768 | if (!tp->sacked_out) { |
de83c058 IJ |
1769 | if (WARN_ON(tp->fackets_out)) |
1770 | tp->fackets_out = 0; | |
6859d494 | 1771 | tcp_highest_sack_reset(sk); |
d738cd8f | 1772 | } |
1da177e4 | 1773 | |
1ed83465 | 1774 | found_dup_sack = tcp_check_dsack(sk, ack_skb, sp_wire, |
d06e021d DM |
1775 | num_sacks, prior_snd_una); |
1776 | if (found_dup_sack) | |
a1197f5a | 1777 | state.flag |= FLAG_DSACKING_ACK; |
6f74651a BE |
1778 | |
1779 | /* Eliminate too old ACKs, but take into | |
1780 | * account more or less fresh ones, they can | |
1781 | * contain valid SACK info. | |
1782 | */ | |
1783 | if (before(TCP_SKB_CB(ack_skb)->ack_seq, prior_snd_una - tp->max_window)) | |
1784 | return 0; | |
1785 | ||