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_output.c,v 1.146 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: Pedro Roque : Retransmit queue handled by TCP. | |
25 | * : Fragmentation on mtu decrease | |
26 | * : Segment collapse on retransmit | |
27 | * : AF independence | |
28 | * | |
29 | * Linus Torvalds : send_delayed_ack | |
30 | * David S. Miller : Charge memory using the right skb | |
31 | * during syn/ack processing. | |
32 | * David S. Miller : Output engine completely rewritten. | |
33 | * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. | |
34 | * Cacophonix Gaul : draft-minshall-nagle-01 | |
35 | * J Hadi Salim : ECN support | |
36 | * | |
37 | */ | |
38 | ||
39 | #include <net/tcp.h> | |
40 | ||
41 | #include <linux/compiler.h> | |
42 | #include <linux/module.h> | |
1da177e4 LT |
43 | |
44 | /* People can turn this off for buggy TCP's found in printers etc. */ | |
ab32ea5d | 45 | int sysctl_tcp_retrans_collapse __read_mostly = 1; |
1da177e4 | 46 | |
15d99e02 RJ |
47 | /* People can turn this on to work with those rare, broken TCPs that |
48 | * interpret the window field as a signed quantity. | |
49 | */ | |
ab32ea5d | 50 | int sysctl_tcp_workaround_signed_windows __read_mostly = 0; |
15d99e02 | 51 | |
1da177e4 LT |
52 | /* This limits the percentage of the congestion window which we |
53 | * will allow a single TSO frame to consume. Building TSO frames | |
54 | * which are too large can cause TCP streams to be bursty. | |
55 | */ | |
ab32ea5d | 56 | int sysctl_tcp_tso_win_divisor __read_mostly = 3; |
1da177e4 | 57 | |
ab32ea5d BH |
58 | int sysctl_tcp_mtu_probing __read_mostly = 0; |
59 | int sysctl_tcp_base_mss __read_mostly = 512; | |
5d424d5a | 60 | |
35089bb2 | 61 | /* By default, RFC2861 behavior. */ |
ab32ea5d | 62 | int sysctl_tcp_slow_start_after_idle __read_mostly = 1; |
35089bb2 | 63 | |
6ff03ac3 IJ |
64 | static inline void tcp_packets_out_inc(struct sock *sk, |
65 | const struct sk_buff *skb) | |
66 | { | |
67 | struct tcp_sock *tp = tcp_sk(sk); | |
68 | int orig = tp->packets_out; | |
69 | ||
70 | tp->packets_out += tcp_skb_pcount(skb); | |
71 | if (!orig) | |
72 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, | |
73 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
74 | } | |
75 | ||
9e412ba7 | 76 | static void update_send_head(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 77 | { |
9e412ba7 IJ |
78 | struct tcp_sock *tp = tcp_sk(sk); |
79 | ||
fe067e8a | 80 | tcp_advance_send_head(sk, skb); |
1da177e4 | 81 | tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; |
9e412ba7 | 82 | tcp_packets_out_inc(sk, skb); |
1da177e4 LT |
83 | } |
84 | ||
85 | /* SND.NXT, if window was not shrunk. | |
86 | * If window has been shrunk, what should we make? It is not clear at all. | |
87 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( | |
88 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already | |
89 | * invalid. OK, let's make this for now: | |
90 | */ | |
9e412ba7 | 91 | static inline __u32 tcp_acceptable_seq(struct sock *sk) |
1da177e4 | 92 | { |
9e412ba7 IJ |
93 | struct tcp_sock *tp = tcp_sk(sk); |
94 | ||
1da177e4 LT |
95 | if (!before(tp->snd_una+tp->snd_wnd, tp->snd_nxt)) |
96 | return tp->snd_nxt; | |
97 | else | |
98 | return tp->snd_una+tp->snd_wnd; | |
99 | } | |
100 | ||
101 | /* Calculate mss to advertise in SYN segment. | |
102 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: | |
103 | * | |
104 | * 1. It is independent of path mtu. | |
105 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. | |
106 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of | |
107 | * attached devices, because some buggy hosts are confused by | |
108 | * large MSS. | |
109 | * 4. We do not make 3, we advertise MSS, calculated from first | |
110 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. | |
111 | * This may be overridden via information stored in routing table. | |
112 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, | |
113 | * probably even Jumbo". | |
114 | */ | |
115 | static __u16 tcp_advertise_mss(struct sock *sk) | |
116 | { | |
117 | struct tcp_sock *tp = tcp_sk(sk); | |
118 | struct dst_entry *dst = __sk_dst_get(sk); | |
119 | int mss = tp->advmss; | |
120 | ||
121 | if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) { | |
122 | mss = dst_metric(dst, RTAX_ADVMSS); | |
123 | tp->advmss = mss; | |
124 | } | |
125 | ||
126 | return (__u16)mss; | |
127 | } | |
128 | ||
129 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". | |
130 | * This is the first part of cwnd validation mechanism. */ | |
463c84b9 | 131 | static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst) |
1da177e4 | 132 | { |
463c84b9 | 133 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
134 | s32 delta = tcp_time_stamp - tp->lsndtime; |
135 | u32 restart_cwnd = tcp_init_cwnd(tp, dst); | |
136 | u32 cwnd = tp->snd_cwnd; | |
137 | ||
6687e988 | 138 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); |
1da177e4 | 139 | |
6687e988 | 140 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
1da177e4 LT |
141 | restart_cwnd = min(restart_cwnd, cwnd); |
142 | ||
463c84b9 | 143 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) |
1da177e4 LT |
144 | cwnd >>= 1; |
145 | tp->snd_cwnd = max(cwnd, restart_cwnd); | |
146 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
147 | tp->snd_cwnd_used = 0; | |
148 | } | |
149 | ||
40efc6fa SH |
150 | static void tcp_event_data_sent(struct tcp_sock *tp, |
151 | struct sk_buff *skb, struct sock *sk) | |
1da177e4 | 152 | { |
463c84b9 ACM |
153 | struct inet_connection_sock *icsk = inet_csk(sk); |
154 | const u32 now = tcp_time_stamp; | |
1da177e4 | 155 | |
35089bb2 DM |
156 | if (sysctl_tcp_slow_start_after_idle && |
157 | (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) | |
463c84b9 | 158 | tcp_cwnd_restart(sk, __sk_dst_get(sk)); |
1da177e4 LT |
159 | |
160 | tp->lsndtime = now; | |
161 | ||
162 | /* If it is a reply for ato after last received | |
163 | * packet, enter pingpong mode. | |
164 | */ | |
463c84b9 ACM |
165 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) |
166 | icsk->icsk_ack.pingpong = 1; | |
1da177e4 LT |
167 | } |
168 | ||
40efc6fa | 169 | static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) |
1da177e4 | 170 | { |
463c84b9 ACM |
171 | tcp_dec_quickack_mode(sk, pkts); |
172 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); | |
1da177e4 LT |
173 | } |
174 | ||
175 | /* Determine a window scaling and initial window to offer. | |
176 | * Based on the assumption that the given amount of space | |
177 | * will be offered. Store the results in the tp structure. | |
178 | * NOTE: for smooth operation initial space offering should | |
179 | * be a multiple of mss if possible. We assume here that mss >= 1. | |
180 | * This MUST be enforced by all callers. | |
181 | */ | |
182 | void tcp_select_initial_window(int __space, __u32 mss, | |
183 | __u32 *rcv_wnd, __u32 *window_clamp, | |
184 | int wscale_ok, __u8 *rcv_wscale) | |
185 | { | |
186 | unsigned int space = (__space < 0 ? 0 : __space); | |
187 | ||
188 | /* If no clamp set the clamp to the max possible scaled window */ | |
189 | if (*window_clamp == 0) | |
190 | (*window_clamp) = (65535 << 14); | |
191 | space = min(*window_clamp, space); | |
192 | ||
193 | /* Quantize space offering to a multiple of mss if possible. */ | |
194 | if (space > mss) | |
195 | space = (space / mss) * mss; | |
196 | ||
197 | /* NOTE: offering an initial window larger than 32767 | |
15d99e02 RJ |
198 | * will break some buggy TCP stacks. If the admin tells us |
199 | * it is likely we could be speaking with such a buggy stack | |
200 | * we will truncate our initial window offering to 32K-1 | |
201 | * unless the remote has sent us a window scaling option, | |
202 | * which we interpret as a sign the remote TCP is not | |
203 | * misinterpreting the window field as a signed quantity. | |
1da177e4 | 204 | */ |
15d99e02 RJ |
205 | if (sysctl_tcp_workaround_signed_windows) |
206 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); | |
207 | else | |
208 | (*rcv_wnd) = space; | |
209 | ||
1da177e4 LT |
210 | (*rcv_wscale) = 0; |
211 | if (wscale_ok) { | |
212 | /* Set window scaling on max possible window | |
e905a9ed | 213 | * See RFC1323 for an explanation of the limit to 14 |
1da177e4 LT |
214 | */ |
215 | space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); | |
316c1592 | 216 | space = min_t(u32, space, *window_clamp); |
1da177e4 LT |
217 | while (space > 65535 && (*rcv_wscale) < 14) { |
218 | space >>= 1; | |
219 | (*rcv_wscale)++; | |
220 | } | |
221 | } | |
222 | ||
223 | /* Set initial window to value enough for senders, | |
6b251858 | 224 | * following RFC2414. Senders, not following this RFC, |
1da177e4 LT |
225 | * will be satisfied with 2. |
226 | */ | |
227 | if (mss > (1<<*rcv_wscale)) { | |
01ff367e DM |
228 | int init_cwnd = 4; |
229 | if (mss > 1460*3) | |
1da177e4 | 230 | init_cwnd = 2; |
01ff367e DM |
231 | else if (mss > 1460) |
232 | init_cwnd = 3; | |
1da177e4 LT |
233 | if (*rcv_wnd > init_cwnd*mss) |
234 | *rcv_wnd = init_cwnd*mss; | |
235 | } | |
236 | ||
237 | /* Set the clamp no higher than max representable value */ | |
238 | (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); | |
239 | } | |
240 | ||
241 | /* Chose a new window to advertise, update state in tcp_sock for the | |
242 | * socket, and return result with RFC1323 scaling applied. The return | |
243 | * value can be stuffed directly into th->window for an outgoing | |
244 | * frame. | |
245 | */ | |
40efc6fa | 246 | static u16 tcp_select_window(struct sock *sk) |
1da177e4 LT |
247 | { |
248 | struct tcp_sock *tp = tcp_sk(sk); | |
249 | u32 cur_win = tcp_receive_window(tp); | |
250 | u32 new_win = __tcp_select_window(sk); | |
251 | ||
252 | /* Never shrink the offered window */ | |
2de979bd | 253 | if (new_win < cur_win) { |
1da177e4 LT |
254 | /* Danger Will Robinson! |
255 | * Don't update rcv_wup/rcv_wnd here or else | |
256 | * we will not be able to advertise a zero | |
257 | * window in time. --DaveM | |
258 | * | |
259 | * Relax Will Robinson. | |
260 | */ | |
261 | new_win = cur_win; | |
262 | } | |
263 | tp->rcv_wnd = new_win; | |
264 | tp->rcv_wup = tp->rcv_nxt; | |
265 | ||
266 | /* Make sure we do not exceed the maximum possible | |
267 | * scaled window. | |
268 | */ | |
15d99e02 | 269 | if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) |
1da177e4 LT |
270 | new_win = min(new_win, MAX_TCP_WINDOW); |
271 | else | |
272 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); | |
273 | ||
274 | /* RFC1323 scaling applied */ | |
275 | new_win >>= tp->rx_opt.rcv_wscale; | |
276 | ||
277 | /* If we advertise zero window, disable fast path. */ | |
278 | if (new_win == 0) | |
279 | tp->pred_flags = 0; | |
280 | ||
281 | return new_win; | |
282 | } | |
283 | ||
bdf1ee5d IJ |
284 | static inline void TCP_ECN_send_synack(struct tcp_sock *tp, |
285 | struct sk_buff *skb) | |
286 | { | |
287 | TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR; | |
288 | if (!(tp->ecn_flags&TCP_ECN_OK)) | |
289 | TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE; | |
290 | } | |
291 | ||
292 | static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) | |
293 | { | |
294 | struct tcp_sock *tp = tcp_sk(sk); | |
295 | ||
296 | tp->ecn_flags = 0; | |
297 | if (sysctl_tcp_ecn) { | |
298 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE|TCPCB_FLAG_CWR; | |
299 | tp->ecn_flags = TCP_ECN_OK; | |
300 | } | |
301 | } | |
302 | ||
303 | static __inline__ void | |
304 | TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th) | |
305 | { | |
306 | if (inet_rsk(req)->ecn_ok) | |
307 | th->ece = 1; | |
308 | } | |
309 | ||
310 | static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, | |
311 | int tcp_header_len) | |
312 | { | |
313 | struct tcp_sock *tp = tcp_sk(sk); | |
314 | ||
315 | if (tp->ecn_flags & TCP_ECN_OK) { | |
316 | /* Not-retransmitted data segment: set ECT and inject CWR. */ | |
317 | if (skb->len != tcp_header_len && | |
318 | !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { | |
319 | INET_ECN_xmit(sk); | |
320 | if (tp->ecn_flags&TCP_ECN_QUEUE_CWR) { | |
321 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; | |
322 | tcp_hdr(skb)->cwr = 1; | |
323 | skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; | |
324 | } | |
325 | } else { | |
326 | /* ACK or retransmitted segment: clear ECT|CE */ | |
327 | INET_ECN_dontxmit(sk); | |
328 | } | |
329 | if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) | |
330 | tcp_hdr(skb)->ece = 1; | |
331 | } | |
332 | } | |
333 | ||
df7a3b07 | 334 | static void tcp_build_and_update_options(__be32 *ptr, struct tcp_sock *tp, |
cfb6eeb4 | 335 | __u32 tstamp, __u8 **md5_hash) |
40efc6fa SH |
336 | { |
337 | if (tp->rx_opt.tstamp_ok) { | |
496c98df YH |
338 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
339 | (TCPOPT_NOP << 16) | | |
340 | (TCPOPT_TIMESTAMP << 8) | | |
341 | TCPOLEN_TIMESTAMP); | |
40efc6fa SH |
342 | *ptr++ = htonl(tstamp); |
343 | *ptr++ = htonl(tp->rx_opt.ts_recent); | |
344 | } | |
345 | if (tp->rx_opt.eff_sacks) { | |
346 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks; | |
347 | int this_sack; | |
348 | ||
349 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
350 | (TCPOPT_NOP << 16) | | |
351 | (TCPOPT_SACK << 8) | | |
352 | (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks * | |
353 | TCPOLEN_SACK_PERBLOCK))); | |
2de979bd SH |
354 | |
355 | for (this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) { | |
40efc6fa SH |
356 | *ptr++ = htonl(sp[this_sack].start_seq); |
357 | *ptr++ = htonl(sp[this_sack].end_seq); | |
358 | } | |
2de979bd | 359 | |
40efc6fa SH |
360 | if (tp->rx_opt.dsack) { |
361 | tp->rx_opt.dsack = 0; | |
362 | tp->rx_opt.eff_sacks--; | |
363 | } | |
364 | } | |
cfb6eeb4 YH |
365 | #ifdef CONFIG_TCP_MD5SIG |
366 | if (md5_hash) { | |
367 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
368 | (TCPOPT_NOP << 16) | | |
369 | (TCPOPT_MD5SIG << 8) | | |
370 | TCPOLEN_MD5SIG); | |
371 | *md5_hash = (__u8 *)ptr; | |
372 | } | |
373 | #endif | |
40efc6fa SH |
374 | } |
375 | ||
376 | /* Construct a tcp options header for a SYN or SYN_ACK packet. | |
377 | * If this is every changed make sure to change the definition of | |
378 | * MAX_SYN_SIZE to match the new maximum number of options that you | |
379 | * can generate. | |
cfb6eeb4 YH |
380 | * |
381 | * Note - that with the RFC2385 TCP option, we make room for the | |
382 | * 16 byte MD5 hash. This will be filled in later, so the pointer for the | |
383 | * location to be filled is passed back up. | |
40efc6fa | 384 | */ |
df7a3b07 | 385 | static void tcp_syn_build_options(__be32 *ptr, int mss, int ts, int sack, |
40efc6fa | 386 | int offer_wscale, int wscale, __u32 tstamp, |
cfb6eeb4 | 387 | __u32 ts_recent, __u8 **md5_hash) |
40efc6fa SH |
388 | { |
389 | /* We always get an MSS option. | |
390 | * The option bytes which will be seen in normal data | |
391 | * packets should timestamps be used, must be in the MSS | |
392 | * advertised. But we subtract them from tp->mss_cache so | |
393 | * that calculations in tcp_sendmsg are simpler etc. | |
394 | * So account for this fact here if necessary. If we | |
395 | * don't do this correctly, as a receiver we won't | |
396 | * recognize data packets as being full sized when we | |
397 | * should, and thus we won't abide by the delayed ACK | |
398 | * rules correctly. | |
399 | * SACKs don't matter, we never delay an ACK when we | |
400 | * have any of those going out. | |
401 | */ | |
402 | *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss); | |
403 | if (ts) { | |
2de979bd | 404 | if (sack) |
496c98df YH |
405 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
406 | (TCPOLEN_SACK_PERM << 16) | | |
407 | (TCPOPT_TIMESTAMP << 8) | | |
408 | TCPOLEN_TIMESTAMP); | |
40efc6fa | 409 | else |
496c98df YH |
410 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
411 | (TCPOPT_NOP << 16) | | |
412 | (TCPOPT_TIMESTAMP << 8) | | |
413 | TCPOLEN_TIMESTAMP); | |
40efc6fa SH |
414 | *ptr++ = htonl(tstamp); /* TSVAL */ |
415 | *ptr++ = htonl(ts_recent); /* TSECR */ | |
2de979bd | 416 | } else if (sack) |
496c98df YH |
417 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
418 | (TCPOPT_NOP << 16) | | |
419 | (TCPOPT_SACK_PERM << 8) | | |
420 | TCPOLEN_SACK_PERM); | |
40efc6fa | 421 | if (offer_wscale) |
496c98df YH |
422 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
423 | (TCPOPT_WINDOW << 16) | | |
424 | (TCPOLEN_WINDOW << 8) | | |
425 | (wscale)); | |
cfb6eeb4 YH |
426 | #ifdef CONFIG_TCP_MD5SIG |
427 | /* | |
428 | * If MD5 is enabled, then we set the option, and include the size | |
429 | * (always 18). The actual MD5 hash is added just before the | |
430 | * packet is sent. | |
431 | */ | |
432 | if (md5_hash) { | |
433 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
434 | (TCPOPT_NOP << 16) | | |
435 | (TCPOPT_MD5SIG << 8) | | |
436 | TCPOLEN_MD5SIG); | |
437 | *md5_hash = (__u8 *) ptr; | |
438 | } | |
439 | #endif | |
40efc6fa | 440 | } |
1da177e4 LT |
441 | |
442 | /* This routine actually transmits TCP packets queued in by | |
443 | * tcp_do_sendmsg(). This is used by both the initial | |
444 | * transmission and possible later retransmissions. | |
445 | * All SKB's seen here are completely headerless. It is our | |
446 | * job to build the TCP header, and pass the packet down to | |
447 | * IP so it can do the same plus pass the packet off to the | |
448 | * device. | |
449 | * | |
450 | * We are working here with either a clone of the original | |
451 | * SKB, or a fresh unique copy made by the retransmit engine. | |
452 | */ | |
dfb4b9dc | 453 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, gfp_t gfp_mask) |
1da177e4 | 454 | { |
dfb4b9dc DM |
455 | const struct inet_connection_sock *icsk = inet_csk(sk); |
456 | struct inet_sock *inet; | |
457 | struct tcp_sock *tp; | |
458 | struct tcp_skb_cb *tcb; | |
459 | int tcp_header_size; | |
cfb6eeb4 YH |
460 | #ifdef CONFIG_TCP_MD5SIG |
461 | struct tcp_md5sig_key *md5; | |
462 | __u8 *md5_hash_location; | |
463 | #endif | |
dfb4b9dc DM |
464 | struct tcphdr *th; |
465 | int sysctl_flags; | |
466 | int err; | |
467 | ||
468 | BUG_ON(!skb || !tcp_skb_pcount(skb)); | |
469 | ||
470 | /* If congestion control is doing timestamping, we must | |
471 | * take such a timestamp before we potentially clone/copy. | |
472 | */ | |
164891aa | 473 | if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) |
dfb4b9dc DM |
474 | __net_timestamp(skb); |
475 | ||
476 | if (likely(clone_it)) { | |
477 | if (unlikely(skb_cloned(skb))) | |
478 | skb = pskb_copy(skb, gfp_mask); | |
479 | else | |
480 | skb = skb_clone(skb, gfp_mask); | |
481 | if (unlikely(!skb)) | |
482 | return -ENOBUFS; | |
483 | } | |
1da177e4 | 484 | |
dfb4b9dc DM |
485 | inet = inet_sk(sk); |
486 | tp = tcp_sk(sk); | |
487 | tcb = TCP_SKB_CB(skb); | |
488 | tcp_header_size = tp->tcp_header_len; | |
1da177e4 LT |
489 | |
490 | #define SYSCTL_FLAG_TSTAMPS 0x1 | |
491 | #define SYSCTL_FLAG_WSCALE 0x2 | |
492 | #define SYSCTL_FLAG_SACK 0x4 | |
493 | ||
dfb4b9dc DM |
494 | sysctl_flags = 0; |
495 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { | |
496 | tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS; | |
2de979bd | 497 | if (sysctl_tcp_timestamps) { |
dfb4b9dc DM |
498 | tcp_header_size += TCPOLEN_TSTAMP_ALIGNED; |
499 | sysctl_flags |= SYSCTL_FLAG_TSTAMPS; | |
1da177e4 | 500 | } |
dfb4b9dc DM |
501 | if (sysctl_tcp_window_scaling) { |
502 | tcp_header_size += TCPOLEN_WSCALE_ALIGNED; | |
503 | sysctl_flags |= SYSCTL_FLAG_WSCALE; | |
1da177e4 | 504 | } |
dfb4b9dc DM |
505 | if (sysctl_tcp_sack) { |
506 | sysctl_flags |= SYSCTL_FLAG_SACK; | |
507 | if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS)) | |
508 | tcp_header_size += TCPOLEN_SACKPERM_ALIGNED; | |
1da177e4 | 509 | } |
dfb4b9dc DM |
510 | } else if (unlikely(tp->rx_opt.eff_sacks)) { |
511 | /* A SACK is 2 pad bytes, a 2 byte header, plus | |
512 | * 2 32-bit sequence numbers for each SACK block. | |
513 | */ | |
514 | tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED + | |
515 | (tp->rx_opt.eff_sacks * | |
516 | TCPOLEN_SACK_PERBLOCK)); | |
517 | } | |
e905a9ed | 518 | |
dfb4b9dc DM |
519 | if (tcp_packets_in_flight(tp) == 0) |
520 | tcp_ca_event(sk, CA_EVENT_TX_START); | |
521 | ||
cfb6eeb4 YH |
522 | #ifdef CONFIG_TCP_MD5SIG |
523 | /* | |
524 | * Are we doing MD5 on this segment? If so - make | |
525 | * room for it. | |
526 | */ | |
527 | md5 = tp->af_specific->md5_lookup(sk, sk); | |
528 | if (md5) | |
529 | tcp_header_size += TCPOLEN_MD5SIG_ALIGNED; | |
530 | #endif | |
531 | ||
aa8223c7 ACM |
532 | skb_push(skb, tcp_header_size); |
533 | skb_reset_transport_header(skb); | |
e89862f4 | 534 | skb_set_owner_w(skb, sk); |
dfb4b9dc DM |
535 | |
536 | /* Build TCP header and checksum it. */ | |
aa8223c7 | 537 | th = tcp_hdr(skb); |
dfb4b9dc DM |
538 | th->source = inet->sport; |
539 | th->dest = inet->dport; | |
540 | th->seq = htonl(tcb->seq); | |
541 | th->ack_seq = htonl(tp->rcv_nxt); | |
df7a3b07 | 542 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
dfb4b9dc DM |
543 | tcb->flags); |
544 | ||
545 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { | |
546 | /* RFC1323: The window in SYN & SYN/ACK segments | |
547 | * is never scaled. | |
548 | */ | |
600ff0c2 | 549 | th->window = htons(min(tp->rcv_wnd, 65535U)); |
dfb4b9dc DM |
550 | } else { |
551 | th->window = htons(tcp_select_window(sk)); | |
552 | } | |
553 | th->check = 0; | |
554 | th->urg_ptr = 0; | |
1da177e4 | 555 | |
dfb4b9dc DM |
556 | if (unlikely(tp->urg_mode && |
557 | between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF))) { | |
558 | th->urg_ptr = htons(tp->snd_up-tcb->seq); | |
559 | th->urg = 1; | |
560 | } | |
1da177e4 | 561 | |
dfb4b9dc | 562 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { |
df7a3b07 | 563 | tcp_syn_build_options((__be32 *)(th + 1), |
dfb4b9dc DM |
564 | tcp_advertise_mss(sk), |
565 | (sysctl_flags & SYSCTL_FLAG_TSTAMPS), | |
566 | (sysctl_flags & SYSCTL_FLAG_SACK), | |
567 | (sysctl_flags & SYSCTL_FLAG_WSCALE), | |
568 | tp->rx_opt.rcv_wscale, | |
569 | tcb->when, | |
cfb6eeb4 YH |
570 | tp->rx_opt.ts_recent, |
571 | ||
572 | #ifdef CONFIG_TCP_MD5SIG | |
573 | md5 ? &md5_hash_location : | |
574 | #endif | |
575 | NULL); | |
dfb4b9dc | 576 | } else { |
df7a3b07 | 577 | tcp_build_and_update_options((__be32 *)(th + 1), |
cfb6eeb4 YH |
578 | tp, tcb->when, |
579 | #ifdef CONFIG_TCP_MD5SIG | |
580 | md5 ? &md5_hash_location : | |
581 | #endif | |
582 | NULL); | |
9e412ba7 | 583 | TCP_ECN_send(sk, skb, tcp_header_size); |
dfb4b9dc | 584 | } |
1da177e4 | 585 | |
cfb6eeb4 YH |
586 | #ifdef CONFIG_TCP_MD5SIG |
587 | /* Calculate the MD5 hash, as we have all we need now */ | |
588 | if (md5) { | |
589 | tp->af_specific->calc_md5_hash(md5_hash_location, | |
590 | md5, | |
591 | sk, NULL, NULL, | |
aa8223c7 | 592 | tcp_hdr(skb), |
cfb6eeb4 YH |
593 | sk->sk_protocol, |
594 | skb->len); | |
595 | } | |
596 | #endif | |
597 | ||
8292a17a | 598 | icsk->icsk_af_ops->send_check(sk, skb->len, skb); |
1da177e4 | 599 | |
dfb4b9dc DM |
600 | if (likely(tcb->flags & TCPCB_FLAG_ACK)) |
601 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); | |
1da177e4 | 602 | |
dfb4b9dc DM |
603 | if (skb->len != tcp_header_size) |
604 | tcp_event_data_sent(tp, skb, sk); | |
1da177e4 | 605 | |
bd37a088 WY |
606 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
607 | TCP_INC_STATS(TCP_MIB_OUTSEGS); | |
1da177e4 | 608 | |
e89862f4 | 609 | err = icsk->icsk_af_ops->queue_xmit(skb, 0); |
83de47cd | 610 | if (likely(err <= 0)) |
dfb4b9dc DM |
611 | return err; |
612 | ||
3cfe3baa | 613 | tcp_enter_cwr(sk, 1); |
dfb4b9dc | 614 | |
b9df3cb8 | 615 | return net_xmit_eval(err); |
1da177e4 | 616 | |
1da177e4 LT |
617 | #undef SYSCTL_FLAG_TSTAMPS |
618 | #undef SYSCTL_FLAG_WSCALE | |
619 | #undef SYSCTL_FLAG_SACK | |
620 | } | |
621 | ||
622 | ||
e905a9ed | 623 | /* This routine just queue's the buffer |
1da177e4 LT |
624 | * |
625 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, | |
626 | * otherwise socket can stall. | |
627 | */ | |
628 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) | |
629 | { | |
630 | struct tcp_sock *tp = tcp_sk(sk); | |
631 | ||
632 | /* Advance write_seq and place onto the write_queue. */ | |
633 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; | |
634 | skb_header_release(skb); | |
fe067e8a | 635 | tcp_add_write_queue_tail(sk, skb); |
1da177e4 | 636 | sk_charge_skb(sk, skb); |
1da177e4 LT |
637 | } |
638 | ||
846998ae | 639 | static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now) |
f6302d1d | 640 | { |
bcd76111 | 641 | if (skb->len <= mss_now || !sk_can_gso(sk)) { |
f6302d1d DM |
642 | /* Avoid the costly divide in the normal |
643 | * non-TSO case. | |
644 | */ | |
7967168c HX |
645 | skb_shinfo(skb)->gso_segs = 1; |
646 | skb_shinfo(skb)->gso_size = 0; | |
647 | skb_shinfo(skb)->gso_type = 0; | |
f6302d1d | 648 | } else { |
356f89e1 | 649 | skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); |
7967168c | 650 | skb_shinfo(skb)->gso_size = mss_now; |
bcd76111 | 651 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
1da177e4 LT |
652 | } |
653 | } | |
654 | ||
91fed7a1 IJ |
655 | /* When a modification to fackets out becomes necessary, we need to check |
656 | * skb is counted to fackets_out or not. Another important thing is to | |
657 | * tweak SACK fastpath hint too as it would overwrite all changes unless | |
658 | * hint is also changed. | |
659 | */ | |
660 | static void tcp_adjust_fackets_out(struct tcp_sock *tp, struct sk_buff *skb, | |
661 | int decr) | |
662 | { | |
dc86967b | 663 | if (!tp->sacked_out || tcp_is_reno(tp)) |
91fed7a1 IJ |
664 | return; |
665 | ||
666 | if (!before(tp->highest_sack, TCP_SKB_CB(skb)->seq)) | |
667 | tp->fackets_out -= decr; | |
668 | ||
669 | /* cnt_hint is "off-by-one" compared with fackets_out (see sacktag) */ | |
670 | if (tp->fastpath_skb_hint != NULL && | |
671 | after(TCP_SKB_CB(tp->fastpath_skb_hint)->seq, TCP_SKB_CB(skb)->seq)) | |
672 | tp->fastpath_cnt_hint -= decr; | |
673 | } | |
674 | ||
1da177e4 LT |
675 | /* Function to create two new TCP segments. Shrinks the given segment |
676 | * to the specified size and appends a new segment with the rest of the | |
e905a9ed | 677 | * packet to the list. This won't be called frequently, I hope. |
1da177e4 LT |
678 | * Remember, these are still headerless SKBs at this point. |
679 | */ | |
6475be16 | 680 | int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now) |
1da177e4 LT |
681 | { |
682 | struct tcp_sock *tp = tcp_sk(sk); | |
683 | struct sk_buff *buff; | |
6475be16 | 684 | int nsize, old_factor; |
b60b49ea | 685 | int nlen; |
1da177e4 LT |
686 | u16 flags; |
687 | ||
b2cc99f0 | 688 | BUG_ON(len > skb->len); |
6a438bbe | 689 | |
b7689205 | 690 | tcp_clear_retrans_hints_partial(tp); |
1da177e4 LT |
691 | nsize = skb_headlen(skb) - len; |
692 | if (nsize < 0) | |
693 | nsize = 0; | |
694 | ||
695 | if (skb_cloned(skb) && | |
696 | skb_is_nonlinear(skb) && | |
697 | pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) | |
698 | return -ENOMEM; | |
699 | ||
700 | /* Get a new skb... force flag on. */ | |
701 | buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); | |
702 | if (buff == NULL) | |
703 | return -ENOMEM; /* We'll just try again later. */ | |
ef5cb973 | 704 | |
b60b49ea HX |
705 | sk_charge_skb(sk, buff); |
706 | nlen = skb->len - len - nsize; | |
707 | buff->truesize += nlen; | |
708 | skb->truesize -= nlen; | |
1da177e4 LT |
709 | |
710 | /* Correct the sequence numbers. */ | |
711 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
712 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
713 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
714 | ||
dc86967b IJ |
715 | if (tcp_is_sack(tp) && tp->sacked_out && |
716 | (TCP_SKB_CB(skb)->seq == tp->highest_sack)) | |
0dde7b54 IJ |
717 | tp->highest_sack = TCP_SKB_CB(buff)->seq; |
718 | ||
1da177e4 LT |
719 | /* PSH and FIN should only be set in the second packet. */ |
720 | flags = TCP_SKB_CB(skb)->flags; | |
721 | TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); | |
722 | TCP_SKB_CB(buff)->flags = flags; | |
e14c3caf | 723 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
1da177e4 LT |
724 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL; |
725 | ||
84fa7933 | 726 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { |
1da177e4 LT |
727 | /* Copy and checksum data tail into the new buffer. */ |
728 | buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize), | |
729 | nsize, 0); | |
730 | ||
731 | skb_trim(skb, len); | |
732 | ||
733 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); | |
734 | } else { | |
84fa7933 | 735 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
736 | skb_split(skb, buff, len); |
737 | } | |
738 | ||
739 | buff->ip_summed = skb->ip_summed; | |
740 | ||
741 | /* Looks stupid, but our code really uses when of | |
742 | * skbs, which it never sent before. --ANK | |
743 | */ | |
744 | TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; | |
a61bbcf2 | 745 | buff->tstamp = skb->tstamp; |
1da177e4 | 746 | |
6475be16 DM |
747 | old_factor = tcp_skb_pcount(skb); |
748 | ||
1da177e4 | 749 | /* Fix up tso_factor for both original and new SKB. */ |
846998ae DM |
750 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
751 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
1da177e4 | 752 | |
6475be16 DM |
753 | /* If this packet has been sent out already, we must |
754 | * adjust the various packet counters. | |
755 | */ | |
cf0b450c | 756 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
6475be16 DM |
757 | int diff = old_factor - tcp_skb_pcount(skb) - |
758 | tcp_skb_pcount(buff); | |
1da177e4 | 759 | |
6475be16 | 760 | tp->packets_out -= diff; |
e14c3caf HX |
761 | |
762 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
763 | tp->sacked_out -= diff; | |
764 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
765 | tp->retrans_out -= diff; | |
766 | ||
b5860bba | 767 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) |
6475be16 | 768 | tp->lost_out -= diff; |
83ca28be | 769 | |
91fed7a1 IJ |
770 | /* Adjust Reno SACK estimate. */ |
771 | if (tcp_is_reno(tp) && diff > 0) { | |
772 | tcp_dec_pcount_approx_int(&tp->sacked_out, diff); | |
773 | tcp_verify_left_out(tp); | |
6475be16 | 774 | } |
91fed7a1 | 775 | tcp_adjust_fackets_out(tp, skb, diff); |
1da177e4 LT |
776 | } |
777 | ||
778 | /* Link BUFF into the send queue. */ | |
f44b5271 | 779 | skb_header_release(buff); |
fe067e8a | 780 | tcp_insert_write_queue_after(skb, buff, sk); |
1da177e4 LT |
781 | |
782 | return 0; | |
783 | } | |
784 | ||
785 | /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c | |
786 | * eventually). The difference is that pulled data not copied, but | |
787 | * immediately discarded. | |
788 | */ | |
f2911969 | 789 | static void __pskb_trim_head(struct sk_buff *skb, int len) |
1da177e4 LT |
790 | { |
791 | int i, k, eat; | |
792 | ||
793 | eat = len; | |
794 | k = 0; | |
795 | for (i=0; i<skb_shinfo(skb)->nr_frags; i++) { | |
796 | if (skb_shinfo(skb)->frags[i].size <= eat) { | |
797 | put_page(skb_shinfo(skb)->frags[i].page); | |
798 | eat -= skb_shinfo(skb)->frags[i].size; | |
799 | } else { | |
800 | skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; | |
801 | if (eat) { | |
802 | skb_shinfo(skb)->frags[k].page_offset += eat; | |
803 | skb_shinfo(skb)->frags[k].size -= eat; | |
804 | eat = 0; | |
805 | } | |
806 | k++; | |
807 | } | |
808 | } | |
809 | skb_shinfo(skb)->nr_frags = k; | |
810 | ||
27a884dc | 811 | skb_reset_tail_pointer(skb); |
1da177e4 LT |
812 | skb->data_len -= len; |
813 | skb->len = skb->data_len; | |
1da177e4 LT |
814 | } |
815 | ||
816 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) | |
817 | { | |
818 | if (skb_cloned(skb) && | |
819 | pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) | |
820 | return -ENOMEM; | |
821 | ||
f2911969 HX |
822 | /* If len == headlen, we avoid __skb_pull to preserve alignment. */ |
823 | if (unlikely(len < skb_headlen(skb))) | |
1da177e4 | 824 | __skb_pull(skb, len); |
f2911969 HX |
825 | else |
826 | __pskb_trim_head(skb, len - skb_headlen(skb)); | |
1da177e4 LT |
827 | |
828 | TCP_SKB_CB(skb)->seq += len; | |
84fa7933 | 829 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
830 | |
831 | skb->truesize -= len; | |
832 | sk->sk_wmem_queued -= len; | |
833 | sk->sk_forward_alloc += len; | |
834 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); | |
835 | ||
836 | /* Any change of skb->len requires recalculation of tso | |
837 | * factor and mss. | |
838 | */ | |
839 | if (tcp_skb_pcount(skb) > 1) | |
846998ae | 840 | tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1)); |
1da177e4 LT |
841 | |
842 | return 0; | |
843 | } | |
844 | ||
5d424d5a JH |
845 | /* Not accounting for SACKs here. */ |
846 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
847 | { | |
848 | struct tcp_sock *tp = tcp_sk(sk); | |
849 | struct inet_connection_sock *icsk = inet_csk(sk); | |
850 | int mss_now; | |
851 | ||
852 | /* Calculate base mss without TCP options: | |
853 | It is MMS_S - sizeof(tcphdr) of rfc1122 | |
854 | */ | |
855 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); | |
856 | ||
857 | /* Clamp it (mss_clamp does not include tcp options) */ | |
858 | if (mss_now > tp->rx_opt.mss_clamp) | |
859 | mss_now = tp->rx_opt.mss_clamp; | |
860 | ||
861 | /* Now subtract optional transport overhead */ | |
862 | mss_now -= icsk->icsk_ext_hdr_len; | |
863 | ||
864 | /* Then reserve room for full set of TCP options and 8 bytes of data */ | |
865 | if (mss_now < 48) | |
866 | mss_now = 48; | |
867 | ||
868 | /* Now subtract TCP options size, not including SACKs */ | |
869 | mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); | |
870 | ||
871 | return mss_now; | |
872 | } | |
873 | ||
874 | /* Inverse of above */ | |
875 | int tcp_mss_to_mtu(struct sock *sk, int mss) | |
876 | { | |
877 | struct tcp_sock *tp = tcp_sk(sk); | |
878 | struct inet_connection_sock *icsk = inet_csk(sk); | |
879 | int mtu; | |
880 | ||
881 | mtu = mss + | |
882 | tp->tcp_header_len + | |
883 | icsk->icsk_ext_hdr_len + | |
884 | icsk->icsk_af_ops->net_header_len; | |
885 | ||
886 | return mtu; | |
887 | } | |
888 | ||
889 | void tcp_mtup_init(struct sock *sk) | |
890 | { | |
891 | struct tcp_sock *tp = tcp_sk(sk); | |
892 | struct inet_connection_sock *icsk = inet_csk(sk); | |
893 | ||
894 | icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; | |
895 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + | |
e905a9ed | 896 | icsk->icsk_af_ops->net_header_len; |
5d424d5a JH |
897 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); |
898 | icsk->icsk_mtup.probe_size = 0; | |
899 | } | |
900 | ||
1da177e4 LT |
901 | /* This function synchronize snd mss to current pmtu/exthdr set. |
902 | ||
903 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts | |
904 | for TCP options, but includes only bare TCP header. | |
905 | ||
906 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. | |
caa20d9a | 907 | It is minimum of user_mss and mss received with SYN. |
1da177e4 LT |
908 | It also does not include TCP options. |
909 | ||
d83d8461 | 910 | inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. |
1da177e4 LT |
911 | |
912 | tp->mss_cache is current effective sending mss, including | |
913 | all tcp options except for SACKs. It is evaluated, | |
914 | taking into account current pmtu, but never exceeds | |
915 | tp->rx_opt.mss_clamp. | |
916 | ||
917 | NOTE1. rfc1122 clearly states that advertised MSS | |
918 | DOES NOT include either tcp or ip options. | |
919 | ||
d83d8461 ACM |
920 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
921 | are READ ONLY outside this function. --ANK (980731) | |
1da177e4 LT |
922 | */ |
923 | ||
924 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) | |
925 | { | |
926 | struct tcp_sock *tp = tcp_sk(sk); | |
d83d8461 | 927 | struct inet_connection_sock *icsk = inet_csk(sk); |
5d424d5a | 928 | int mss_now; |
1da177e4 | 929 | |
5d424d5a JH |
930 | if (icsk->icsk_mtup.search_high > pmtu) |
931 | icsk->icsk_mtup.search_high = pmtu; | |
1da177e4 | 932 | |
5d424d5a | 933 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
1da177e4 LT |
934 | |
935 | /* Bound mss with half of window */ | |
936 | if (tp->max_window && mss_now > (tp->max_window>>1)) | |
937 | mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len); | |
938 | ||
939 | /* And store cached results */ | |
d83d8461 | 940 | icsk->icsk_pmtu_cookie = pmtu; |
5d424d5a JH |
941 | if (icsk->icsk_mtup.enabled) |
942 | mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); | |
c1b4a7e6 | 943 | tp->mss_cache = mss_now; |
1da177e4 LT |
944 | |
945 | return mss_now; | |
946 | } | |
947 | ||
948 | /* Compute the current effective MSS, taking SACKs and IP options, | |
949 | * and even PMTU discovery events into account. | |
950 | * | |
951 | * LARGESEND note: !urg_mode is overkill, only frames up to snd_up | |
952 | * cannot be large. However, taking into account rare use of URG, this | |
953 | * is not a big flaw. | |
954 | */ | |
c1b4a7e6 | 955 | unsigned int tcp_current_mss(struct sock *sk, int large_allowed) |
1da177e4 LT |
956 | { |
957 | struct tcp_sock *tp = tcp_sk(sk); | |
958 | struct dst_entry *dst = __sk_dst_get(sk); | |
c1b4a7e6 DM |
959 | u32 mss_now; |
960 | u16 xmit_size_goal; | |
961 | int doing_tso = 0; | |
962 | ||
963 | mss_now = tp->mss_cache; | |
964 | ||
bcd76111 | 965 | if (large_allowed && sk_can_gso(sk) && !tp->urg_mode) |
c1b4a7e6 | 966 | doing_tso = 1; |
1da177e4 | 967 | |
1da177e4 LT |
968 | if (dst) { |
969 | u32 mtu = dst_mtu(dst); | |
d83d8461 | 970 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
1da177e4 LT |
971 | mss_now = tcp_sync_mss(sk, mtu); |
972 | } | |
973 | ||
c1b4a7e6 DM |
974 | if (tp->rx_opt.eff_sacks) |
975 | mss_now -= (TCPOLEN_SACK_BASE_ALIGNED + | |
976 | (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK)); | |
1da177e4 | 977 | |
cfb6eeb4 YH |
978 | #ifdef CONFIG_TCP_MD5SIG |
979 | if (tp->af_specific->md5_lookup(sk, sk)) | |
980 | mss_now -= TCPOLEN_MD5SIG_ALIGNED; | |
981 | #endif | |
982 | ||
c1b4a7e6 | 983 | xmit_size_goal = mss_now; |
1da177e4 | 984 | |
c1b4a7e6 | 985 | if (doing_tso) { |
8292a17a ACM |
986 | xmit_size_goal = (65535 - |
987 | inet_csk(sk)->icsk_af_ops->net_header_len - | |
d83d8461 ACM |
988 | inet_csk(sk)->icsk_ext_hdr_len - |
989 | tp->tcp_header_len); | |
1da177e4 | 990 | |
c1b4a7e6 DM |
991 | if (tp->max_window && |
992 | (xmit_size_goal > (tp->max_window >> 1))) | |
993 | xmit_size_goal = max((tp->max_window >> 1), | |
994 | 68U - tp->tcp_header_len); | |
1da177e4 | 995 | |
c1b4a7e6 | 996 | xmit_size_goal -= (xmit_size_goal % mss_now); |
1da177e4 | 997 | } |
c1b4a7e6 | 998 | tp->xmit_size_goal = xmit_size_goal; |
1da177e4 | 999 | |
1da177e4 LT |
1000 | return mss_now; |
1001 | } | |
1002 | ||
a762a980 DM |
1003 | /* Congestion window validation. (RFC2861) */ |
1004 | ||
9e412ba7 | 1005 | static void tcp_cwnd_validate(struct sock *sk) |
a762a980 | 1006 | { |
9e412ba7 | 1007 | struct tcp_sock *tp = tcp_sk(sk); |
a762a980 DM |
1008 | __u32 packets_out = tp->packets_out; |
1009 | ||
1010 | if (packets_out >= tp->snd_cwnd) { | |
1011 | /* Network is feed fully. */ | |
1012 | tp->snd_cwnd_used = 0; | |
1013 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1014 | } else { | |
1015 | /* Network starves. */ | |
1016 | if (tp->packets_out > tp->snd_cwnd_used) | |
1017 | tp->snd_cwnd_used = tp->packets_out; | |
1018 | ||
15d33c07 DM |
1019 | if (sysctl_tcp_slow_start_after_idle && |
1020 | (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) | |
a762a980 DM |
1021 | tcp_cwnd_application_limited(sk); |
1022 | } | |
1023 | } | |
1024 | ||
c1b4a7e6 DM |
1025 | static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd) |
1026 | { | |
1027 | u32 window, cwnd_len; | |
1028 | ||
1029 | window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq); | |
1030 | cwnd_len = mss_now * cwnd; | |
1031 | return min(window, cwnd_len); | |
1032 | } | |
1033 | ||
1034 | /* Can at least one segment of SKB be sent right now, according to the | |
1035 | * congestion window rules? If so, return how many segments are allowed. | |
1036 | */ | |
1037 | static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb) | |
1038 | { | |
1039 | u32 in_flight, cwnd; | |
1040 | ||
1041 | /* Don't be strict about the congestion window for the final FIN. */ | |
104439a8 JH |
1042 | if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && |
1043 | tcp_skb_pcount(skb) == 1) | |
c1b4a7e6 DM |
1044 | return 1; |
1045 | ||
1046 | in_flight = tcp_packets_in_flight(tp); | |
1047 | cwnd = tp->snd_cwnd; | |
1048 | if (in_flight < cwnd) | |
1049 | return (cwnd - in_flight); | |
1050 | ||
1051 | return 0; | |
1052 | } | |
1053 | ||
1054 | /* This must be invoked the first time we consider transmitting | |
1055 | * SKB onto the wire. | |
1056 | */ | |
40efc6fa | 1057 | static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now) |
c1b4a7e6 DM |
1058 | { |
1059 | int tso_segs = tcp_skb_pcount(skb); | |
1060 | ||
846998ae DM |
1061 | if (!tso_segs || |
1062 | (tso_segs > 1 && | |
7967168c | 1063 | tcp_skb_mss(skb) != mss_now)) { |
846998ae | 1064 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
c1b4a7e6 DM |
1065 | tso_segs = tcp_skb_pcount(skb); |
1066 | } | |
1067 | return tso_segs; | |
1068 | } | |
1069 | ||
1070 | static inline int tcp_minshall_check(const struct tcp_sock *tp) | |
1071 | { | |
1072 | return after(tp->snd_sml,tp->snd_una) && | |
1073 | !after(tp->snd_sml, tp->snd_nxt); | |
1074 | } | |
1075 | ||
1076 | /* Return 0, if packet can be sent now without violation Nagle's rules: | |
1077 | * 1. It is full sized. | |
1078 | * 2. Or it contains FIN. (already checked by caller) | |
1079 | * 3. Or TCP_NODELAY was set. | |
1080 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. | |
1081 | * With Minshall's modification: all sent small packets are ACKed. | |
1082 | */ | |
1083 | ||
1084 | static inline int tcp_nagle_check(const struct tcp_sock *tp, | |
e905a9ed | 1085 | const struct sk_buff *skb, |
c1b4a7e6 DM |
1086 | unsigned mss_now, int nonagle) |
1087 | { | |
1088 | return (skb->len < mss_now && | |
1089 | ((nonagle&TCP_NAGLE_CORK) || | |
1090 | (!nonagle && | |
1091 | tp->packets_out && | |
1092 | tcp_minshall_check(tp)))); | |
1093 | } | |
1094 | ||
1095 | /* Return non-zero if the Nagle test allows this packet to be | |
1096 | * sent now. | |
1097 | */ | |
1098 | static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, | |
1099 | unsigned int cur_mss, int nonagle) | |
1100 | { | |
1101 | /* Nagle rule does not apply to frames, which sit in the middle of the | |
1102 | * write_queue (they have no chances to get new data). | |
1103 | * | |
1104 | * This is implemented in the callers, where they modify the 'nonagle' | |
1105 | * argument based upon the location of SKB in the send queue. | |
1106 | */ | |
1107 | if (nonagle & TCP_NAGLE_PUSH) | |
1108 | return 1; | |
1109 | ||
d551e454 IJ |
1110 | /* Don't use the nagle rule for urgent data (or for the final FIN). |
1111 | * Nagle can be ignored during F-RTO too (see RFC4138). | |
1112 | */ | |
1113 | if (tp->urg_mode || (tp->frto_counter == 2) || | |
c1b4a7e6 DM |
1114 | (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)) |
1115 | return 1; | |
1116 | ||
1117 | if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) | |
1118 | return 1; | |
1119 | ||
1120 | return 0; | |
1121 | } | |
1122 | ||
1123 | /* Does at least the first segment of SKB fit into the send window? */ | |
1124 | static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss) | |
1125 | { | |
1126 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
1127 | ||
1128 | if (skb->len > cur_mss) | |
1129 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; | |
1130 | ||
1131 | return !after(end_seq, tp->snd_una + tp->snd_wnd); | |
1132 | } | |
1133 | ||
fe067e8a | 1134 | /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) |
c1b4a7e6 DM |
1135 | * should be put on the wire right now. If so, it returns the number of |
1136 | * packets allowed by the congestion window. | |
1137 | */ | |
1138 | static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, | |
1139 | unsigned int cur_mss, int nonagle) | |
1140 | { | |
1141 | struct tcp_sock *tp = tcp_sk(sk); | |
1142 | unsigned int cwnd_quota; | |
1143 | ||
846998ae | 1144 | tcp_init_tso_segs(sk, skb, cur_mss); |
c1b4a7e6 DM |
1145 | |
1146 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) | |
1147 | return 0; | |
1148 | ||
1149 | cwnd_quota = tcp_cwnd_test(tp, skb); | |
1150 | if (cwnd_quota && | |
1151 | !tcp_snd_wnd_test(tp, skb, cur_mss)) | |
1152 | cwnd_quota = 0; | |
1153 | ||
1154 | return cwnd_quota; | |
1155 | } | |
1156 | ||
9e412ba7 | 1157 | int tcp_may_send_now(struct sock *sk) |
c1b4a7e6 | 1158 | { |
9e412ba7 | 1159 | struct tcp_sock *tp = tcp_sk(sk); |
fe067e8a | 1160 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 DM |
1161 | |
1162 | return (skb && | |
1163 | tcp_snd_test(sk, skb, tcp_current_mss(sk, 1), | |
1164 | (tcp_skb_is_last(sk, skb) ? | |