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 | |
66f5fe62 | 64 | static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 65 | { |
9e412ba7 | 66 | struct tcp_sock *tp = tcp_sk(sk); |
66f5fe62 | 67 | unsigned int prior_packets = tp->packets_out; |
9e412ba7 | 68 | |
fe067e8a | 69 | tcp_advance_send_head(sk, skb); |
1da177e4 | 70 | tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; |
8512430e IJ |
71 | |
72 | /* Don't override Nagle indefinately with F-RTO */ | |
73 | if (tp->frto_counter == 2) | |
74 | tp->frto_counter = 3; | |
66f5fe62 IJ |
75 | |
76 | tp->packets_out += tcp_skb_pcount(skb); | |
77 | if (!prior_packets) | |
78 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, | |
79 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
1da177e4 LT |
80 | } |
81 | ||
82 | /* SND.NXT, if window was not shrunk. | |
83 | * If window has been shrunk, what should we make? It is not clear at all. | |
84 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( | |
85 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already | |
86 | * invalid. OK, let's make this for now: | |
87 | */ | |
9e412ba7 | 88 | static inline __u32 tcp_acceptable_seq(struct sock *sk) |
1da177e4 | 89 | { |
9e412ba7 IJ |
90 | struct tcp_sock *tp = tcp_sk(sk); |
91 | ||
1da177e4 LT |
92 | if (!before(tp->snd_una+tp->snd_wnd, tp->snd_nxt)) |
93 | return tp->snd_nxt; | |
94 | else | |
95 | return tp->snd_una+tp->snd_wnd; | |
96 | } | |
97 | ||
98 | /* Calculate mss to advertise in SYN segment. | |
99 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: | |
100 | * | |
101 | * 1. It is independent of path mtu. | |
102 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. | |
103 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of | |
104 | * attached devices, because some buggy hosts are confused by | |
105 | * large MSS. | |
106 | * 4. We do not make 3, we advertise MSS, calculated from first | |
107 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. | |
108 | * This may be overridden via information stored in routing table. | |
109 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, | |
110 | * probably even Jumbo". | |
111 | */ | |
112 | static __u16 tcp_advertise_mss(struct sock *sk) | |
113 | { | |
114 | struct tcp_sock *tp = tcp_sk(sk); | |
115 | struct dst_entry *dst = __sk_dst_get(sk); | |
116 | int mss = tp->advmss; | |
117 | ||
118 | if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) { | |
119 | mss = dst_metric(dst, RTAX_ADVMSS); | |
120 | tp->advmss = mss; | |
121 | } | |
122 | ||
123 | return (__u16)mss; | |
124 | } | |
125 | ||
126 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". | |
127 | * This is the first part of cwnd validation mechanism. */ | |
463c84b9 | 128 | static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst) |
1da177e4 | 129 | { |
463c84b9 | 130 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
131 | s32 delta = tcp_time_stamp - tp->lsndtime; |
132 | u32 restart_cwnd = tcp_init_cwnd(tp, dst); | |
133 | u32 cwnd = tp->snd_cwnd; | |
134 | ||
6687e988 | 135 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); |
1da177e4 | 136 | |
6687e988 | 137 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
1da177e4 LT |
138 | restart_cwnd = min(restart_cwnd, cwnd); |
139 | ||
463c84b9 | 140 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) |
1da177e4 LT |
141 | cwnd >>= 1; |
142 | tp->snd_cwnd = max(cwnd, restart_cwnd); | |
143 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
144 | tp->snd_cwnd_used = 0; | |
145 | } | |
146 | ||
40efc6fa SH |
147 | static void tcp_event_data_sent(struct tcp_sock *tp, |
148 | struct sk_buff *skb, struct sock *sk) | |
1da177e4 | 149 | { |
463c84b9 ACM |
150 | struct inet_connection_sock *icsk = inet_csk(sk); |
151 | const u32 now = tcp_time_stamp; | |
1da177e4 | 152 | |
35089bb2 DM |
153 | if (sysctl_tcp_slow_start_after_idle && |
154 | (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) | |
463c84b9 | 155 | tcp_cwnd_restart(sk, __sk_dst_get(sk)); |
1da177e4 LT |
156 | |
157 | tp->lsndtime = now; | |
158 | ||
159 | /* If it is a reply for ato after last received | |
160 | * packet, enter pingpong mode. | |
161 | */ | |
463c84b9 ACM |
162 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) |
163 | icsk->icsk_ack.pingpong = 1; | |
1da177e4 LT |
164 | } |
165 | ||
40efc6fa | 166 | static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) |
1da177e4 | 167 | { |
463c84b9 ACM |
168 | tcp_dec_quickack_mode(sk, pkts); |
169 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); | |
1da177e4 LT |
170 | } |
171 | ||
172 | /* Determine a window scaling and initial window to offer. | |
173 | * Based on the assumption that the given amount of space | |
174 | * will be offered. Store the results in the tp structure. | |
175 | * NOTE: for smooth operation initial space offering should | |
176 | * be a multiple of mss if possible. We assume here that mss >= 1. | |
177 | * This MUST be enforced by all callers. | |
178 | */ | |
179 | void tcp_select_initial_window(int __space, __u32 mss, | |
180 | __u32 *rcv_wnd, __u32 *window_clamp, | |
181 | int wscale_ok, __u8 *rcv_wscale) | |
182 | { | |
183 | unsigned int space = (__space < 0 ? 0 : __space); | |
184 | ||
185 | /* If no clamp set the clamp to the max possible scaled window */ | |
186 | if (*window_clamp == 0) | |
187 | (*window_clamp) = (65535 << 14); | |
188 | space = min(*window_clamp, space); | |
189 | ||
190 | /* Quantize space offering to a multiple of mss if possible. */ | |
191 | if (space > mss) | |
192 | space = (space / mss) * mss; | |
193 | ||
194 | /* NOTE: offering an initial window larger than 32767 | |
15d99e02 RJ |
195 | * will break some buggy TCP stacks. If the admin tells us |
196 | * it is likely we could be speaking with such a buggy stack | |
197 | * we will truncate our initial window offering to 32K-1 | |
198 | * unless the remote has sent us a window scaling option, | |
199 | * which we interpret as a sign the remote TCP is not | |
200 | * misinterpreting the window field as a signed quantity. | |
1da177e4 | 201 | */ |
15d99e02 RJ |
202 | if (sysctl_tcp_workaround_signed_windows) |
203 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); | |
204 | else | |
205 | (*rcv_wnd) = space; | |
206 | ||
1da177e4 LT |
207 | (*rcv_wscale) = 0; |
208 | if (wscale_ok) { | |
209 | /* Set window scaling on max possible window | |
e905a9ed | 210 | * See RFC1323 for an explanation of the limit to 14 |
1da177e4 LT |
211 | */ |
212 | space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); | |
316c1592 | 213 | space = min_t(u32, space, *window_clamp); |
1da177e4 LT |
214 | while (space > 65535 && (*rcv_wscale) < 14) { |
215 | space >>= 1; | |
216 | (*rcv_wscale)++; | |
217 | } | |
218 | } | |
219 | ||
220 | /* Set initial window to value enough for senders, | |
6b251858 | 221 | * following RFC2414. Senders, not following this RFC, |
1da177e4 LT |
222 | * will be satisfied with 2. |
223 | */ | |
224 | if (mss > (1<<*rcv_wscale)) { | |
01ff367e DM |
225 | int init_cwnd = 4; |
226 | if (mss > 1460*3) | |
1da177e4 | 227 | init_cwnd = 2; |
01ff367e DM |
228 | else if (mss > 1460) |
229 | init_cwnd = 3; | |
1da177e4 LT |
230 | if (*rcv_wnd > init_cwnd*mss) |
231 | *rcv_wnd = init_cwnd*mss; | |
232 | } | |
233 | ||
234 | /* Set the clamp no higher than max representable value */ | |
235 | (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); | |
236 | } | |
237 | ||
238 | /* Chose a new window to advertise, update state in tcp_sock for the | |
239 | * socket, and return result with RFC1323 scaling applied. The return | |
240 | * value can be stuffed directly into th->window for an outgoing | |
241 | * frame. | |
242 | */ | |
40efc6fa | 243 | static u16 tcp_select_window(struct sock *sk) |
1da177e4 LT |
244 | { |
245 | struct tcp_sock *tp = tcp_sk(sk); | |
246 | u32 cur_win = tcp_receive_window(tp); | |
247 | u32 new_win = __tcp_select_window(sk); | |
248 | ||
249 | /* Never shrink the offered window */ | |
2de979bd | 250 | if (new_win < cur_win) { |
1da177e4 LT |
251 | /* Danger Will Robinson! |
252 | * Don't update rcv_wup/rcv_wnd here or else | |
253 | * we will not be able to advertise a zero | |
254 | * window in time. --DaveM | |
255 | * | |
256 | * Relax Will Robinson. | |
257 | */ | |
258 | new_win = cur_win; | |
259 | } | |
260 | tp->rcv_wnd = new_win; | |
261 | tp->rcv_wup = tp->rcv_nxt; | |
262 | ||
263 | /* Make sure we do not exceed the maximum possible | |
264 | * scaled window. | |
265 | */ | |
15d99e02 | 266 | if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) |
1da177e4 LT |
267 | new_win = min(new_win, MAX_TCP_WINDOW); |
268 | else | |
269 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); | |
270 | ||
271 | /* RFC1323 scaling applied */ | |
272 | new_win >>= tp->rx_opt.rcv_wscale; | |
273 | ||
274 | /* If we advertise zero window, disable fast path. */ | |
275 | if (new_win == 0) | |
276 | tp->pred_flags = 0; | |
277 | ||
278 | return new_win; | |
279 | } | |
280 | ||
bdf1ee5d IJ |
281 | static inline void TCP_ECN_send_synack(struct tcp_sock *tp, |
282 | struct sk_buff *skb) | |
283 | { | |
284 | TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR; | |
285 | if (!(tp->ecn_flags&TCP_ECN_OK)) | |
286 | TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE; | |
287 | } | |
288 | ||
289 | static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) | |
290 | { | |
291 | struct tcp_sock *tp = tcp_sk(sk); | |
292 | ||
293 | tp->ecn_flags = 0; | |
294 | if (sysctl_tcp_ecn) { | |
295 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE|TCPCB_FLAG_CWR; | |
296 | tp->ecn_flags = TCP_ECN_OK; | |
297 | } | |
298 | } | |
299 | ||
300 | static __inline__ void | |
301 | TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th) | |
302 | { | |
303 | if (inet_rsk(req)->ecn_ok) | |
304 | th->ece = 1; | |
305 | } | |
306 | ||
307 | static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, | |
308 | int tcp_header_len) | |
309 | { | |
310 | struct tcp_sock *tp = tcp_sk(sk); | |
311 | ||
312 | if (tp->ecn_flags & TCP_ECN_OK) { | |
313 | /* Not-retransmitted data segment: set ECT and inject CWR. */ | |
314 | if (skb->len != tcp_header_len && | |
315 | !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { | |
316 | INET_ECN_xmit(sk); | |
317 | if (tp->ecn_flags&TCP_ECN_QUEUE_CWR) { | |
318 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; | |
319 | tcp_hdr(skb)->cwr = 1; | |
320 | skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; | |
321 | } | |
322 | } else { | |
323 | /* ACK or retransmitted segment: clear ECT|CE */ | |
324 | INET_ECN_dontxmit(sk); | |
325 | } | |
326 | if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) | |
327 | tcp_hdr(skb)->ece = 1; | |
328 | } | |
329 | } | |
330 | ||
df7a3b07 | 331 | static void tcp_build_and_update_options(__be32 *ptr, struct tcp_sock *tp, |
cfb6eeb4 | 332 | __u32 tstamp, __u8 **md5_hash) |
40efc6fa SH |
333 | { |
334 | if (tp->rx_opt.tstamp_ok) { | |
496c98df YH |
335 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
336 | (TCPOPT_NOP << 16) | | |
337 | (TCPOPT_TIMESTAMP << 8) | | |
338 | TCPOLEN_TIMESTAMP); | |
40efc6fa SH |
339 | *ptr++ = htonl(tstamp); |
340 | *ptr++ = htonl(tp->rx_opt.ts_recent); | |
341 | } | |
342 | if (tp->rx_opt.eff_sacks) { | |
343 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks; | |
344 | int this_sack; | |
345 | ||
346 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
347 | (TCPOPT_NOP << 16) | | |
348 | (TCPOPT_SACK << 8) | | |
349 | (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks * | |
350 | TCPOLEN_SACK_PERBLOCK))); | |
2de979bd SH |
351 | |
352 | for (this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) { | |
40efc6fa SH |
353 | *ptr++ = htonl(sp[this_sack].start_seq); |
354 | *ptr++ = htonl(sp[this_sack].end_seq); | |
355 | } | |
2de979bd | 356 | |
40efc6fa SH |
357 | if (tp->rx_opt.dsack) { |
358 | tp->rx_opt.dsack = 0; | |
359 | tp->rx_opt.eff_sacks--; | |
360 | } | |
361 | } | |
cfb6eeb4 YH |
362 | #ifdef CONFIG_TCP_MD5SIG |
363 | if (md5_hash) { | |
364 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
365 | (TCPOPT_NOP << 16) | | |
366 | (TCPOPT_MD5SIG << 8) | | |
367 | TCPOLEN_MD5SIG); | |
368 | *md5_hash = (__u8 *)ptr; | |
369 | } | |
370 | #endif | |
40efc6fa SH |
371 | } |
372 | ||
373 | /* Construct a tcp options header for a SYN or SYN_ACK packet. | |
374 | * If this is every changed make sure to change the definition of | |
375 | * MAX_SYN_SIZE to match the new maximum number of options that you | |
376 | * can generate. | |
cfb6eeb4 YH |
377 | * |
378 | * Note - that with the RFC2385 TCP option, we make room for the | |
379 | * 16 byte MD5 hash. This will be filled in later, so the pointer for the | |
380 | * location to be filled is passed back up. | |
40efc6fa | 381 | */ |
df7a3b07 | 382 | static void tcp_syn_build_options(__be32 *ptr, int mss, int ts, int sack, |
40efc6fa | 383 | int offer_wscale, int wscale, __u32 tstamp, |
cfb6eeb4 | 384 | __u32 ts_recent, __u8 **md5_hash) |
40efc6fa SH |
385 | { |
386 | /* We always get an MSS option. | |
387 | * The option bytes which will be seen in normal data | |
388 | * packets should timestamps be used, must be in the MSS | |
389 | * advertised. But we subtract them from tp->mss_cache so | |
390 | * that calculations in tcp_sendmsg are simpler etc. | |
391 | * So account for this fact here if necessary. If we | |
392 | * don't do this correctly, as a receiver we won't | |
393 | * recognize data packets as being full sized when we | |
394 | * should, and thus we won't abide by the delayed ACK | |
395 | * rules correctly. | |
396 | * SACKs don't matter, we never delay an ACK when we | |
397 | * have any of those going out. | |
398 | */ | |
399 | *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss); | |
400 | if (ts) { | |
2de979bd | 401 | if (sack) |
496c98df YH |
402 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
403 | (TCPOLEN_SACK_PERM << 16) | | |
404 | (TCPOPT_TIMESTAMP << 8) | | |
405 | TCPOLEN_TIMESTAMP); | |
40efc6fa | 406 | else |
496c98df YH |
407 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
408 | (TCPOPT_NOP << 16) | | |
409 | (TCPOPT_TIMESTAMP << 8) | | |
410 | TCPOLEN_TIMESTAMP); | |
40efc6fa SH |
411 | *ptr++ = htonl(tstamp); /* TSVAL */ |
412 | *ptr++ = htonl(ts_recent); /* TSECR */ | |
2de979bd | 413 | } else if (sack) |
496c98df YH |
414 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
415 | (TCPOPT_NOP << 16) | | |
416 | (TCPOPT_SACK_PERM << 8) | | |
417 | TCPOLEN_SACK_PERM); | |
40efc6fa | 418 | if (offer_wscale) |
496c98df YH |
419 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
420 | (TCPOPT_WINDOW << 16) | | |
421 | (TCPOLEN_WINDOW << 8) | | |
422 | (wscale)); | |
cfb6eeb4 YH |
423 | #ifdef CONFIG_TCP_MD5SIG |
424 | /* | |
425 | * If MD5 is enabled, then we set the option, and include the size | |
426 | * (always 18). The actual MD5 hash is added just before the | |
427 | * packet is sent. | |
428 | */ | |
429 | if (md5_hash) { | |
430 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
431 | (TCPOPT_NOP << 16) | | |
432 | (TCPOPT_MD5SIG << 8) | | |
433 | TCPOLEN_MD5SIG); | |
434 | *md5_hash = (__u8 *) ptr; | |
435 | } | |
436 | #endif | |
40efc6fa | 437 | } |
1da177e4 LT |
438 | |
439 | /* This routine actually transmits TCP packets queued in by | |
440 | * tcp_do_sendmsg(). This is used by both the initial | |
441 | * transmission and possible later retransmissions. | |
442 | * All SKB's seen here are completely headerless. It is our | |
443 | * job to build the TCP header, and pass the packet down to | |
444 | * IP so it can do the same plus pass the packet off to the | |
445 | * device. | |
446 | * | |
447 | * We are working here with either a clone of the original | |
448 | * SKB, or a fresh unique copy made by the retransmit engine. | |
449 | */ | |
dfb4b9dc | 450 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, gfp_t gfp_mask) |
1da177e4 | 451 | { |
dfb4b9dc DM |
452 | const struct inet_connection_sock *icsk = inet_csk(sk); |
453 | struct inet_sock *inet; | |
454 | struct tcp_sock *tp; | |
455 | struct tcp_skb_cb *tcb; | |
456 | int tcp_header_size; | |
cfb6eeb4 YH |
457 | #ifdef CONFIG_TCP_MD5SIG |
458 | struct tcp_md5sig_key *md5; | |
459 | __u8 *md5_hash_location; | |
460 | #endif | |
dfb4b9dc DM |
461 | struct tcphdr *th; |
462 | int sysctl_flags; | |
463 | int err; | |
464 | ||
465 | BUG_ON(!skb || !tcp_skb_pcount(skb)); | |
466 | ||
467 | /* If congestion control is doing timestamping, we must | |
468 | * take such a timestamp before we potentially clone/copy. | |
469 | */ | |
164891aa | 470 | if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) |
dfb4b9dc DM |
471 | __net_timestamp(skb); |
472 | ||
473 | if (likely(clone_it)) { | |
474 | if (unlikely(skb_cloned(skb))) | |
475 | skb = pskb_copy(skb, gfp_mask); | |
476 | else | |
477 | skb = skb_clone(skb, gfp_mask); | |
478 | if (unlikely(!skb)) | |
479 | return -ENOBUFS; | |
480 | } | |
1da177e4 | 481 | |
dfb4b9dc DM |
482 | inet = inet_sk(sk); |
483 | tp = tcp_sk(sk); | |
484 | tcb = TCP_SKB_CB(skb); | |
485 | tcp_header_size = tp->tcp_header_len; | |
1da177e4 LT |
486 | |
487 | #define SYSCTL_FLAG_TSTAMPS 0x1 | |
488 | #define SYSCTL_FLAG_WSCALE 0x2 | |
489 | #define SYSCTL_FLAG_SACK 0x4 | |
490 | ||
dfb4b9dc DM |
491 | sysctl_flags = 0; |
492 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { | |
493 | tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS; | |
2de979bd | 494 | if (sysctl_tcp_timestamps) { |
dfb4b9dc DM |
495 | tcp_header_size += TCPOLEN_TSTAMP_ALIGNED; |
496 | sysctl_flags |= SYSCTL_FLAG_TSTAMPS; | |
1da177e4 | 497 | } |
dfb4b9dc DM |
498 | if (sysctl_tcp_window_scaling) { |
499 | tcp_header_size += TCPOLEN_WSCALE_ALIGNED; | |
500 | sysctl_flags |= SYSCTL_FLAG_WSCALE; | |
1da177e4 | 501 | } |
dfb4b9dc DM |
502 | if (sysctl_tcp_sack) { |
503 | sysctl_flags |= SYSCTL_FLAG_SACK; | |
504 | if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS)) | |
505 | tcp_header_size += TCPOLEN_SACKPERM_ALIGNED; | |
1da177e4 | 506 | } |
dfb4b9dc DM |
507 | } else if (unlikely(tp->rx_opt.eff_sacks)) { |
508 | /* A SACK is 2 pad bytes, a 2 byte header, plus | |
509 | * 2 32-bit sequence numbers for each SACK block. | |
510 | */ | |
511 | tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED + | |
512 | (tp->rx_opt.eff_sacks * | |
513 | TCPOLEN_SACK_PERBLOCK)); | |
514 | } | |
e905a9ed | 515 | |
dfb4b9dc DM |
516 | if (tcp_packets_in_flight(tp) == 0) |
517 | tcp_ca_event(sk, CA_EVENT_TX_START); | |
518 | ||
cfb6eeb4 YH |
519 | #ifdef CONFIG_TCP_MD5SIG |
520 | /* | |
521 | * Are we doing MD5 on this segment? If so - make | |
522 | * room for it. | |
523 | */ | |
524 | md5 = tp->af_specific->md5_lookup(sk, sk); | |
525 | if (md5) | |
526 | tcp_header_size += TCPOLEN_MD5SIG_ALIGNED; | |
527 | #endif | |
528 | ||
aa8223c7 ACM |
529 | skb_push(skb, tcp_header_size); |
530 | skb_reset_transport_header(skb); | |
e89862f4 | 531 | skb_set_owner_w(skb, sk); |
dfb4b9dc DM |
532 | |
533 | /* Build TCP header and checksum it. */ | |
aa8223c7 | 534 | th = tcp_hdr(skb); |
dfb4b9dc DM |
535 | th->source = inet->sport; |
536 | th->dest = inet->dport; | |
537 | th->seq = htonl(tcb->seq); | |
538 | th->ack_seq = htonl(tp->rcv_nxt); | |
df7a3b07 | 539 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
dfb4b9dc DM |
540 | tcb->flags); |
541 | ||
542 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { | |
543 | /* RFC1323: The window in SYN & SYN/ACK segments | |
544 | * is never scaled. | |
545 | */ | |
600ff0c2 | 546 | th->window = htons(min(tp->rcv_wnd, 65535U)); |
dfb4b9dc DM |
547 | } else { |
548 | th->window = htons(tcp_select_window(sk)); | |
549 | } | |
550 | th->check = 0; | |
551 | th->urg_ptr = 0; | |
1da177e4 | 552 | |
dfb4b9dc DM |
553 | if (unlikely(tp->urg_mode && |
554 | between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF))) { | |
555 | th->urg_ptr = htons(tp->snd_up-tcb->seq); | |
556 | th->urg = 1; | |
557 | } | |
1da177e4 | 558 | |
dfb4b9dc | 559 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { |
df7a3b07 | 560 | tcp_syn_build_options((__be32 *)(th + 1), |
dfb4b9dc DM |
561 | tcp_advertise_mss(sk), |
562 | (sysctl_flags & SYSCTL_FLAG_TSTAMPS), | |
563 | (sysctl_flags & SYSCTL_FLAG_SACK), | |
564 | (sysctl_flags & SYSCTL_FLAG_WSCALE), | |
565 | tp->rx_opt.rcv_wscale, | |
566 | tcb->when, | |
cfb6eeb4 YH |
567 | tp->rx_opt.ts_recent, |
568 | ||
569 | #ifdef CONFIG_TCP_MD5SIG | |
570 | md5 ? &md5_hash_location : | |
571 | #endif | |
572 | NULL); | |
dfb4b9dc | 573 | } else { |
df7a3b07 | 574 | tcp_build_and_update_options((__be32 *)(th + 1), |
cfb6eeb4 YH |
575 | tp, tcb->when, |
576 | #ifdef CONFIG_TCP_MD5SIG | |
577 | md5 ? &md5_hash_location : | |
578 | #endif | |
579 | NULL); | |
9e412ba7 | 580 | TCP_ECN_send(sk, skb, tcp_header_size); |
dfb4b9dc | 581 | } |
1da177e4 | 582 | |
cfb6eeb4 YH |
583 | #ifdef CONFIG_TCP_MD5SIG |
584 | /* Calculate the MD5 hash, as we have all we need now */ | |
585 | if (md5) { | |
586 | tp->af_specific->calc_md5_hash(md5_hash_location, | |
587 | md5, | |
588 | sk, NULL, NULL, | |
aa8223c7 | 589 | tcp_hdr(skb), |
cfb6eeb4 YH |
590 | sk->sk_protocol, |
591 | skb->len); | |
592 | } | |
593 | #endif | |
594 | ||
8292a17a | 595 | icsk->icsk_af_ops->send_check(sk, skb->len, skb); |
1da177e4 | 596 | |
dfb4b9dc DM |
597 | if (likely(tcb->flags & TCPCB_FLAG_ACK)) |
598 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); | |
1da177e4 | 599 | |
dfb4b9dc DM |
600 | if (skb->len != tcp_header_size) |
601 | tcp_event_data_sent(tp, skb, sk); | |
1da177e4 | 602 | |
bd37a088 WY |
603 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
604 | TCP_INC_STATS(TCP_MIB_OUTSEGS); | |
1da177e4 | 605 | |
e89862f4 | 606 | err = icsk->icsk_af_ops->queue_xmit(skb, 0); |
83de47cd | 607 | if (likely(err <= 0)) |
dfb4b9dc DM |
608 | return err; |
609 | ||
3cfe3baa | 610 | tcp_enter_cwr(sk, 1); |
dfb4b9dc | 611 | |
b9df3cb8 | 612 | return net_xmit_eval(err); |
1da177e4 | 613 | |
1da177e4 LT |
614 | #undef SYSCTL_FLAG_TSTAMPS |
615 | #undef SYSCTL_FLAG_WSCALE | |
616 | #undef SYSCTL_FLAG_SACK | |
617 | } | |
618 | ||
619 | ||
e905a9ed | 620 | /* This routine just queue's the buffer |
1da177e4 LT |
621 | * |
622 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, | |
623 | * otherwise socket can stall. | |
624 | */ | |
625 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) | |
626 | { | |
627 | struct tcp_sock *tp = tcp_sk(sk); | |
628 | ||
629 | /* Advance write_seq and place onto the write_queue. */ | |
630 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; | |
631 | skb_header_release(skb); | |
fe067e8a | 632 | tcp_add_write_queue_tail(sk, skb); |
3ab224be HA |
633 | sk->sk_wmem_queued += skb->truesize; |
634 | sk_mem_charge(sk, skb->truesize); | |
1da177e4 LT |
635 | } |
636 | ||
846998ae | 637 | static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now) |
f6302d1d | 638 | { |
bcd76111 | 639 | if (skb->len <= mss_now || !sk_can_gso(sk)) { |
f6302d1d DM |
640 | /* Avoid the costly divide in the normal |
641 | * non-TSO case. | |
642 | */ | |
7967168c HX |
643 | skb_shinfo(skb)->gso_segs = 1; |
644 | skb_shinfo(skb)->gso_size = 0; | |
645 | skb_shinfo(skb)->gso_type = 0; | |
f6302d1d | 646 | } else { |
356f89e1 | 647 | skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); |
7967168c | 648 | skb_shinfo(skb)->gso_size = mss_now; |
bcd76111 | 649 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
1da177e4 LT |
650 | } |
651 | } | |
652 | ||
91fed7a1 | 653 | /* When a modification to fackets out becomes necessary, we need to check |
68f8353b | 654 | * skb is counted to fackets_out or not. |
91fed7a1 | 655 | */ |
a47e5a98 | 656 | static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb, |
91fed7a1 IJ |
657 | int decr) |
658 | { | |
a47e5a98 IJ |
659 | struct tcp_sock *tp = tcp_sk(sk); |
660 | ||
dc86967b | 661 | if (!tp->sacked_out || tcp_is_reno(tp)) |
91fed7a1 IJ |
662 | return; |
663 | ||
6859d494 | 664 | if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) |
91fed7a1 | 665 | tp->fackets_out -= decr; |
91fed7a1 IJ |
666 | } |
667 | ||
1da177e4 LT |
668 | /* Function to create two new TCP segments. Shrinks the given segment |
669 | * to the specified size and appends a new segment with the rest of the | |
e905a9ed | 670 | * packet to the list. This won't be called frequently, I hope. |
1da177e4 LT |
671 | * Remember, these are still headerless SKBs at this point. |
672 | */ | |
6475be16 | 673 | int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now) |
1da177e4 LT |
674 | { |
675 | struct tcp_sock *tp = tcp_sk(sk); | |
676 | struct sk_buff *buff; | |
6475be16 | 677 | int nsize, old_factor; |
b60b49ea | 678 | int nlen; |
1da177e4 LT |
679 | u16 flags; |
680 | ||
b2cc99f0 | 681 | BUG_ON(len > skb->len); |
6a438bbe | 682 | |
b7689205 | 683 | tcp_clear_retrans_hints_partial(tp); |
1da177e4 LT |
684 | nsize = skb_headlen(skb) - len; |
685 | if (nsize < 0) | |
686 | nsize = 0; | |
687 | ||
688 | if (skb_cloned(skb) && | |
689 | skb_is_nonlinear(skb) && | |
690 | pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) | |
691 | return -ENOMEM; | |
692 | ||
693 | /* Get a new skb... force flag on. */ | |
694 | buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); | |
695 | if (buff == NULL) | |
696 | return -ENOMEM; /* We'll just try again later. */ | |
ef5cb973 | 697 | |
3ab224be HA |
698 | sk->sk_wmem_queued += buff->truesize; |
699 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea HX |
700 | nlen = skb->len - len - nsize; |
701 | buff->truesize += nlen; | |
702 | skb->truesize -= nlen; | |
1da177e4 LT |
703 | |
704 | /* Correct the sequence numbers. */ | |
705 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
706 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
707 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
708 | ||
709 | /* PSH and FIN should only be set in the second packet. */ | |
710 | flags = TCP_SKB_CB(skb)->flags; | |
711 | TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); | |
712 | TCP_SKB_CB(buff)->flags = flags; | |
e14c3caf | 713 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
1da177e4 LT |
714 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL; |
715 | ||
84fa7933 | 716 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { |
1da177e4 LT |
717 | /* Copy and checksum data tail into the new buffer. */ |
718 | buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize), | |
719 | nsize, 0); | |
720 | ||
721 | skb_trim(skb, len); | |
722 | ||
723 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); | |
724 | } else { | |
84fa7933 | 725 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
726 | skb_split(skb, buff, len); |
727 | } | |
728 | ||
729 | buff->ip_summed = skb->ip_summed; | |
730 | ||
731 | /* Looks stupid, but our code really uses when of | |
732 | * skbs, which it never sent before. --ANK | |
733 | */ | |
734 | TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; | |
a61bbcf2 | 735 | buff->tstamp = skb->tstamp; |
1da177e4 | 736 | |
6475be16 DM |
737 | old_factor = tcp_skb_pcount(skb); |
738 | ||
1da177e4 | 739 | /* Fix up tso_factor for both original and new SKB. */ |
846998ae DM |
740 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
741 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
1da177e4 | 742 | |
6475be16 DM |
743 | /* If this packet has been sent out already, we must |
744 | * adjust the various packet counters. | |
745 | */ | |
cf0b450c | 746 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
6475be16 DM |
747 | int diff = old_factor - tcp_skb_pcount(skb) - |
748 | tcp_skb_pcount(buff); | |
1da177e4 | 749 | |
6475be16 | 750 | tp->packets_out -= diff; |
e14c3caf HX |
751 | |
752 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
753 | tp->sacked_out -= diff; | |
754 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
755 | tp->retrans_out -= diff; | |
756 | ||
b5860bba | 757 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) |
6475be16 | 758 | tp->lost_out -= diff; |
83ca28be | 759 | |
91fed7a1 IJ |
760 | /* Adjust Reno SACK estimate. */ |
761 | if (tcp_is_reno(tp) && diff > 0) { | |
762 | tcp_dec_pcount_approx_int(&tp->sacked_out, diff); | |
763 | tcp_verify_left_out(tp); | |
6475be16 | 764 | } |
a47e5a98 | 765 | tcp_adjust_fackets_out(sk, skb, diff); |
1da177e4 LT |
766 | } |
767 | ||
768 | /* Link BUFF into the send queue. */ | |
f44b5271 | 769 | skb_header_release(buff); |
fe067e8a | 770 | tcp_insert_write_queue_after(skb, buff, sk); |
1da177e4 LT |
771 | |
772 | return 0; | |
773 | } | |
774 | ||
775 | /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c | |
776 | * eventually). The difference is that pulled data not copied, but | |
777 | * immediately discarded. | |
778 | */ | |
f2911969 | 779 | static void __pskb_trim_head(struct sk_buff *skb, int len) |
1da177e4 LT |
780 | { |
781 | int i, k, eat; | |
782 | ||
783 | eat = len; | |
784 | k = 0; | |
785 | for (i=0; i<skb_shinfo(skb)->nr_frags; i++) { | |
786 | if (skb_shinfo(skb)->frags[i].size <= eat) { | |
787 | put_page(skb_shinfo(skb)->frags[i].page); | |
788 | eat -= skb_shinfo(skb)->frags[i].size; | |
789 | } else { | |
790 | skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; | |
791 | if (eat) { | |
792 | skb_shinfo(skb)->frags[k].page_offset += eat; | |
793 | skb_shinfo(skb)->frags[k].size -= eat; | |
794 | eat = 0; | |
795 | } | |
796 | k++; | |
797 | } | |
798 | } | |
799 | skb_shinfo(skb)->nr_frags = k; | |
800 | ||
27a884dc | 801 | skb_reset_tail_pointer(skb); |
1da177e4 LT |
802 | skb->data_len -= len; |
803 | skb->len = skb->data_len; | |
1da177e4 LT |
804 | } |
805 | ||
806 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) | |
807 | { | |
808 | if (skb_cloned(skb) && | |
809 | pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) | |
810 | return -ENOMEM; | |
811 | ||
f2911969 HXP |
812 | /* If len == headlen, we avoid __skb_pull to preserve alignment. */ |
813 | if (unlikely(len < skb_headlen(skb))) | |
1da177e4 | 814 | __skb_pull(skb, len); |
f2911969 HXP |
815 | else |
816 | __pskb_trim_head(skb, len - skb_headlen(skb)); | |
1da177e4 LT |
817 | |
818 | TCP_SKB_CB(skb)->seq += len; | |
84fa7933 | 819 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
820 | |
821 | skb->truesize -= len; | |
822 | sk->sk_wmem_queued -= len; | |
3ab224be | 823 | sk_mem_uncharge(sk, len); |
1da177e4 LT |
824 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); |
825 | ||
826 | /* Any change of skb->len requires recalculation of tso | |
827 | * factor and mss. | |
828 | */ | |
829 | if (tcp_skb_pcount(skb) > 1) | |
846998ae | 830 | tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1)); |
1da177e4 LT |
831 | |
832 | return 0; | |
833 | } | |
834 | ||
5d424d5a JH |
835 | /* Not accounting for SACKs here. */ |
836 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
837 | { | |
838 | struct tcp_sock *tp = tcp_sk(sk); | |
839 | struct inet_connection_sock *icsk = inet_csk(sk); | |
840 | int mss_now; | |
841 | ||
842 | /* Calculate base mss without TCP options: | |
843 | It is MMS_S - sizeof(tcphdr) of rfc1122 | |
844 | */ | |
845 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); | |
846 | ||
847 | /* Clamp it (mss_clamp does not include tcp options) */ | |
848 | if (mss_now > tp->rx_opt.mss_clamp) | |
849 | mss_now = tp->rx_opt.mss_clamp; | |
850 | ||
851 | /* Now subtract optional transport overhead */ | |
852 | mss_now -= icsk->icsk_ext_hdr_len; | |
853 | ||
854 | /* Then reserve room for full set of TCP options and 8 bytes of data */ | |
855 | if (mss_now < 48) | |
856 | mss_now = 48; | |
857 | ||
858 | /* Now subtract TCP options size, not including SACKs */ | |
859 | mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); | |
860 | ||
861 | return mss_now; | |
862 | } | |
863 | ||
864 | /* Inverse of above */ | |
865 | int tcp_mss_to_mtu(struct sock *sk, int mss) | |
866 | { | |
867 | struct tcp_sock *tp = tcp_sk(sk); | |
868 | struct inet_connection_sock *icsk = inet_csk(sk); | |
869 | int mtu; | |
870 | ||
871 | mtu = mss + | |
872 | tp->tcp_header_len + | |
873 | icsk->icsk_ext_hdr_len + | |
874 | icsk->icsk_af_ops->net_header_len; | |
875 | ||
876 | return mtu; | |
877 | } | |
878 | ||
879 | void tcp_mtup_init(struct sock *sk) | |
880 | { | |
881 | struct tcp_sock *tp = tcp_sk(sk); | |
882 | struct inet_connection_sock *icsk = inet_csk(sk); | |
883 | ||
884 | icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; | |
885 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + | |
e905a9ed | 886 | icsk->icsk_af_ops->net_header_len; |
5d424d5a JH |
887 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); |
888 | icsk->icsk_mtup.probe_size = 0; | |
889 | } | |
890 | ||
1da177e4 LT |
891 | /* This function synchronize snd mss to current pmtu/exthdr set. |
892 | ||
893 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts | |
894 | for TCP options, but includes only bare TCP header. | |
895 | ||
896 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. | |
caa20d9a | 897 | It is minimum of user_mss and mss received with SYN. |
1da177e4 LT |
898 | It also does not include TCP options. |
899 | ||
d83d8461 | 900 | inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. |
1da177e4 LT |
901 | |
902 | tp->mss_cache is current effective sending mss, including | |
903 | all tcp options except for SACKs. It is evaluated, | |
904 | taking into account current pmtu, but never exceeds | |
905 | tp->rx_opt.mss_clamp. | |
906 | ||
907 | NOTE1. rfc1122 clearly states that advertised MSS | |
908 | DOES NOT include either tcp or ip options. | |
909 | ||
d83d8461 ACM |
910 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
911 | are READ ONLY outside this function. --ANK (980731) | |
1da177e4 LT |
912 | */ |
913 | ||
914 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) | |
915 | { | |
916 | struct tcp_sock *tp = tcp_sk(sk); | |
d83d8461 | 917 | struct inet_connection_sock *icsk = inet_csk(sk); |
5d424d5a | 918 | int mss_now; |
1da177e4 | 919 | |
5d424d5a JH |
920 | if (icsk->icsk_mtup.search_high > pmtu) |
921 | icsk->icsk_mtup.search_high = pmtu; | |
1da177e4 | 922 | |
5d424d5a | 923 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
1da177e4 LT |
924 | |
925 | /* Bound mss with half of window */ | |
926 | if (tp->max_window && mss_now > (tp->max_window>>1)) | |
927 | mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len); | |
928 | ||
929 | /* And store cached results */ | |
d83d8461 | 930 | icsk->icsk_pmtu_cookie = pmtu; |
5d424d5a JH |
931 | if (icsk->icsk_mtup.enabled) |
932 | mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); | |
c1b4a7e6 | 933 | tp->mss_cache = mss_now; |
1da177e4 LT |
934 | |
935 | return mss_now; | |
936 | } | |
937 | ||
938 | /* Compute the current effective MSS, taking SACKs and IP options, | |
939 | * and even PMTU discovery events into account. | |
940 | * | |
941 | * LARGESEND note: !urg_mode is overkill, only frames up to snd_up | |
942 | * cannot be large. However, taking into account rare use of URG, this | |
943 | * is not a big flaw. | |
944 | */ | |
c1b4a7e6 | 945 | unsigned int tcp_current_mss(struct sock *sk, int large_allowed) |
1da177e4 LT |
946 | { |
947 | struct tcp_sock *tp = tcp_sk(sk); | |
948 | struct dst_entry *dst = __sk_dst_get(sk); | |
c1b4a7e6 DM |
949 | u32 mss_now; |
950 | u16 xmit_size_goal; | |
951 | int doing_tso = 0; | |
952 | ||
953 | mss_now = tp->mss_cache; | |
954 | ||
bcd76111 | 955 | if (large_allowed && sk_can_gso(sk) && !tp->urg_mode) |
c1b4a7e6 | 956 | doing_tso = 1; |
1da177e4 | 957 | |
1da177e4 LT |
958 | if (dst) { |
959 | u32 mtu = dst_mtu(dst); | |
d83d8461 | 960 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
1da177e4 LT |
961 | mss_now = tcp_sync_mss(sk, mtu); |
962 | } | |
963 | ||
c1b4a7e6 DM |
964 | if (tp->rx_opt.eff_sacks) |
965 | mss_now -= (TCPOLEN_SACK_BASE_ALIGNED + | |
966 | (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK)); | |
1da177e4 | 967 | |
cfb6eeb4 YH |
968 | #ifdef CONFIG_TCP_MD5SIG |
969 | if (tp->af_specific->md5_lookup(sk, sk)) | |
970 | mss_now -= TCPOLEN_MD5SIG_ALIGNED; | |
971 | #endif | |
972 | ||
c1b4a7e6 | 973 | xmit_size_goal = mss_now; |
1da177e4 | 974 | |
c1b4a7e6 | 975 | if (doing_tso) { |
8292a17a ACM |
976 | xmit_size_goal = (65535 - |
977 | inet_csk(sk)->icsk_af_ops->net_header_len - | |
d83d8461 ACM |
978 | inet_csk(sk)->icsk_ext_hdr_len - |
979 | tp->tcp_header_len); | |
1da177e4 | 980 | |
c1b4a7e6 DM |
981 | if (tp->max_window && |
982 | (xmit_size_goal > (tp->max_window >> 1))) | |
983 | xmit_size_goal = max((tp->max_window >> 1), | |
984 | 68U - tp->tcp_header_len); | |
1da177e4 | 985 | |
c1b4a7e6 | 986 | xmit_size_goal -= (xmit_size_goal % mss_now); |
1da177e4 | 987 | } |
c1b4a7e6 | 988 | tp->xmit_size_goal = xmit_size_goal; |
1da177e4 | 989 | |
1da177e4 LT |
990 | return mss_now; |
991 | } | |
992 | ||
a762a980 DM |
993 | /* Congestion window validation. (RFC2861) */ |
994 | ||
9e412ba7 | 995 | static void tcp_cwnd_validate(struct sock *sk) |
a762a980 | 996 | { |
9e412ba7 | 997 | struct tcp_sock *tp = tcp_sk(sk); |
a762a980 DM |
998 | __u32 packets_out = tp->packets_out; |
999 | ||
1000 | if (packets_out >= tp->snd_cwnd) { | |
1001 | /* Network is feed fully. */ | |
1002 | tp->snd_cwnd_used = 0; | |
1003 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1004 | } else { | |
1005 | /* Network starves. */ | |
1006 | if (tp->packets_out > tp->snd_cwnd_used) | |
1007 | tp->snd_cwnd_used = tp->packets_out; | |
1008 | ||
15d33c07 DM |
1009 | if (sysctl_tcp_slow_start_after_idle && |
1010 | (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) | |
a762a980 DM |
1011 | tcp_cwnd_application_limited(sk); |
1012 | } | |
1013 | } | |
1014 | ||
0e3a4803 IJ |
1015 | /* Returns the portion of skb which can be sent right away without |
1016 | * introducing MSS oddities to segment boundaries. In rare cases where | |
1017 | * mss_now != mss_cache, we will request caller to create a small skb | |
1018 | * per input skb which could be mostly avoided here (if desired). | |
1019 | */ | |
1020 | static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb, | |
1021 | unsigned int mss_now, | |
1022 | unsigned int cwnd) | |
c1b4a7e6 | 1023 | { |
0e3a4803 IJ |
1024 | struct tcp_sock *tp = tcp_sk(sk); |
1025 | u32 needed, window, cwnd_len; | |
c1b4a7e6 DM |
1026 | |
1027 | window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq); | |
1028 | cwnd_len = mss_now * cwnd; | |
0e3a4803 IJ |
1029 | |
1030 | if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk))) | |
1031 | return cwnd_len; | |
1032 | ||
1033 | if (skb == tcp_write_queue_tail(sk) && cwnd_len <= skb->len) | |
1034 | return cwnd_len; | |
1035 | ||
1036 | needed = min(skb->len, window); | |
1037 | return needed - needed % mss_now; | |
c1b4a7e6 DM |
1038 | } |
1039 | ||
1040 | /* Can at least one segment of SKB be sent right now, according to the | |
1041 | * congestion window rules? If so, return how many segments are allowed. | |
1042 | */ | |
1043 | static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb) | |
1044 | { | |
1045 | u32 in_flight, cwnd; | |
1046 | ||
1047 | /* Don't be strict about the congestion window for the final FIN. */ | |
104439a8 JH |
1048 | if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && |
1049 | tcp_skb_pcount(skb) == 1) | |
c1b4a7e6 DM |
1050 | return 1; |
1051 | ||
1052 | in_flight = tcp_packets_in_flight(tp); | |
1053 | cwnd = tp->snd_cwnd; | |
1054 | if (in_flight < cwnd) | |
1055 | return (cwnd - in_flight); | |
1056 | ||
1057 | return 0; | |
1058 | } | |
1059 | ||
1060 | /* This must be invoked the first time we consider transmitting | |
1061 | * SKB onto the wire. | |
1062 | */ | |
40efc6fa | 1063 | static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now) |
c1b4a7e6 DM |
1064 | { |
1065 | int tso_segs = tcp_skb_pcount(skb); | |
1066 | ||
846998ae DM |
1067 | if (!tso_segs || |
1068 | (tso_segs > 1 && | |
7967168c | 1069 | tcp_skb_mss(skb) != mss_now)) { |
846998ae | 1070 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
c1b4a7e6 DM |
1071 | tso_segs = tcp_skb_pcount(skb); |
1072 | } | |
1073 | return tso_segs; | |
1074 | } | |
1075 | ||
1076 | static inline int tcp_minshall_check(const struct tcp_sock *tp) | |
1077 | { | |
1078 | return after(tp->snd_sml,tp->snd_una) && | |
1079 | !after(tp->snd_sml, tp->snd_nxt); | |
1080 | } | |
1081 | ||
1082 | /* Return 0, if packet can be sent now without violation Nagle's rules: | |
1083 | * 1. It is full sized. | |
1084 | * 2. Or it contains FIN. (already checked by caller) | |
1085 | * 3. Or TCP_NODELAY was set. | |
1086 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. | |
1087 | * With Minshall's modification: all sent small packets are ACKed. | |
1088 | */ | |
1089 | ||
1090 | static inline int tcp_nagle_check(const struct tcp_sock *tp, | |
e905a9ed | 1091 | const struct sk_buff *skb, |
c1b4a7e6 DM |
1092 | unsigned mss_now, int nonagle) |
1093 | { | |
1094 | return (skb->len < mss_now && | |
1095 | ((nonagle&TCP_NAGLE_CORK) || | |
1096 | (!nonagle && | |
1097 | tp->packets_out && | |
1098 | tcp_minshall_check(tp)))); | |
1099 | } | |
1100 | ||
1101 | /* Return non-zero if the Nagle test allows this packet to be | |
1102 | * sent now. | |
1103 | */ | |
1104 | static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, | |
1105 | unsigned int cur_mss, int nonagle) | |
1106 | { | |
1107 | /* Nagle rule does not apply to frames, which sit in the middle of the | |
1108 | * write_queue (they have no chances to get new data). | |
1109 | * | |
1110 | * This is implemented in the callers, where they modify the 'nonagle' | |
1111 | * argument based upon the location of SKB in the send queue. | |
1112 | */ | |
1113 | if (nonagle & TCP_NAGLE_PUSH) | |
1114 | return 1; | |
1115 | ||
d551e454 IJ |
1116 | /* Don't use the nagle rule for urgent data (or for the final FIN). |
1117 | * Nagle can be ignored during F-RTO too (see RFC4138). | |
1118 | */ | |
1119 | if (tp->urg_mode || (tp->frto_counter == 2) || | |
c1b4a7e6 DM |
1120 | (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)) |
1121 | return 1; | |
1122 | ||
1123 | if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) | |
1124 | return 1; | |
1125 | ||
1126 | return 0; | |
1127 | } | |
1128 | ||
1129 | /* Does at least the first segment of SKB fit into the send window? */ | |
1130 | static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss) | |
1131 | { | |
1132 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
1133 | ||
1134 | if (skb->len > cur_mss) | |
1135 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; | |
1136 | ||
1137 | return !after(end_seq, tp->snd_una + tp->snd_wnd); | |
1138 | } | |
1139 | ||
fe067e8a | 1140 | /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) |
c1b4a7e6 DM |
1141 | * should be put on the wire right now. If so, it returns the number of |
1142 | * packets allowed by the congestion window. | |
1143 | */ | |
1144 | static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, | |
1145 | unsigned int cur_mss, int nonagle) | |
1146 | { | |
1147 | struct tcp_sock *tp = tcp_sk(sk); | |
1148 | unsigned int cwnd_quota; | |
1149 | ||
846998ae | 1150 | tcp_init_tso_segs(sk, skb, cur_mss); |
c1b4a7e6 DM |
1151 | |
1152 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) | |
1153 | return 0; | |
1154 | ||
1155 | cwnd_quota = tcp_cwnd_test(tp, skb); | |
1156 | if (cwnd_quota && | |
1157 | !tcp_snd_wnd_test(tp, skb, cur_mss)) | |
1158 | cwnd_quota = 0; | |
1159 | ||
1160 | return cwnd_quota; | |
1161 | } | |
1162 | ||
9e412ba7 | 1163 | int tcp_may_send_now(struct sock *sk) |
c1b4a7e6 | 1164 | { |
9e412ba7 | 1165 | struct tcp_sock *tp = tcp_sk(sk); |
fe067e8a | 1166 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 DM |
1167 | |
1168 | return (skb && | |
1169 | tcp_snd_test(sk, skb, tcp_current_mss(sk, 1), | |
1170 | (tcp_skb_is_last(sk, skb) ? | |