]>
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 | |
9e412ba7 | 64 | static void update_send_head(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 65 | { |
9e412ba7 IJ |
66 | struct tcp_sock *tp = tcp_sk(sk); |
67 | ||
fe067e8a | 68 | tcp_advance_send_head(sk, skb); |
1da177e4 | 69 | tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; |
9e412ba7 | 70 | tcp_packets_out_inc(sk, skb); |
1da177e4 LT |
71 | } |
72 | ||
73 | /* SND.NXT, if window was not shrunk. | |
74 | * If window has been shrunk, what should we make? It is not clear at all. | |
75 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( | |
76 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already | |
77 | * invalid. OK, let's make this for now: | |
78 | */ | |
9e412ba7 | 79 | static inline __u32 tcp_acceptable_seq(struct sock *sk) |
1da177e4 | 80 | { |
9e412ba7 IJ |
81 | struct tcp_sock *tp = tcp_sk(sk); |
82 | ||
1da177e4 LT |
83 | if (!before(tp->snd_una+tp->snd_wnd, tp->snd_nxt)) |
84 | return tp->snd_nxt; | |
85 | else | |
86 | return tp->snd_una+tp->snd_wnd; | |
87 | } | |
88 | ||
89 | /* Calculate mss to advertise in SYN segment. | |
90 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: | |
91 | * | |
92 | * 1. It is independent of path mtu. | |
93 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. | |
94 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of | |
95 | * attached devices, because some buggy hosts are confused by | |
96 | * large MSS. | |
97 | * 4. We do not make 3, we advertise MSS, calculated from first | |
98 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. | |
99 | * This may be overridden via information stored in routing table. | |
100 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, | |
101 | * probably even Jumbo". | |
102 | */ | |
103 | static __u16 tcp_advertise_mss(struct sock *sk) | |
104 | { | |
105 | struct tcp_sock *tp = tcp_sk(sk); | |
106 | struct dst_entry *dst = __sk_dst_get(sk); | |
107 | int mss = tp->advmss; | |
108 | ||
109 | if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) { | |
110 | mss = dst_metric(dst, RTAX_ADVMSS); | |
111 | tp->advmss = mss; | |
112 | } | |
113 | ||
114 | return (__u16)mss; | |
115 | } | |
116 | ||
117 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". | |
118 | * This is the first part of cwnd validation mechanism. */ | |
463c84b9 | 119 | static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst) |
1da177e4 | 120 | { |
463c84b9 | 121 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
122 | s32 delta = tcp_time_stamp - tp->lsndtime; |
123 | u32 restart_cwnd = tcp_init_cwnd(tp, dst); | |
124 | u32 cwnd = tp->snd_cwnd; | |
125 | ||
6687e988 | 126 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); |
1da177e4 | 127 | |
6687e988 | 128 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
1da177e4 LT |
129 | restart_cwnd = min(restart_cwnd, cwnd); |
130 | ||
463c84b9 | 131 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) |
1da177e4 LT |
132 | cwnd >>= 1; |
133 | tp->snd_cwnd = max(cwnd, restart_cwnd); | |
134 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
135 | tp->snd_cwnd_used = 0; | |
136 | } | |
137 | ||
40efc6fa SH |
138 | static void tcp_event_data_sent(struct tcp_sock *tp, |
139 | struct sk_buff *skb, struct sock *sk) | |
1da177e4 | 140 | { |
463c84b9 ACM |
141 | struct inet_connection_sock *icsk = inet_csk(sk); |
142 | const u32 now = tcp_time_stamp; | |
1da177e4 | 143 | |
35089bb2 DM |
144 | if (sysctl_tcp_slow_start_after_idle && |
145 | (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) | |
463c84b9 | 146 | tcp_cwnd_restart(sk, __sk_dst_get(sk)); |
1da177e4 LT |
147 | |
148 | tp->lsndtime = now; | |
149 | ||
150 | /* If it is a reply for ato after last received | |
151 | * packet, enter pingpong mode. | |
152 | */ | |
463c84b9 ACM |
153 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) |
154 | icsk->icsk_ack.pingpong = 1; | |
1da177e4 LT |
155 | } |
156 | ||
40efc6fa | 157 | static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) |
1da177e4 | 158 | { |
463c84b9 ACM |
159 | tcp_dec_quickack_mode(sk, pkts); |
160 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); | |
1da177e4 LT |
161 | } |
162 | ||
163 | /* Determine a window scaling and initial window to offer. | |
164 | * Based on the assumption that the given amount of space | |
165 | * will be offered. Store the results in the tp structure. | |
166 | * NOTE: for smooth operation initial space offering should | |
167 | * be a multiple of mss if possible. We assume here that mss >= 1. | |
168 | * This MUST be enforced by all callers. | |
169 | */ | |
170 | void tcp_select_initial_window(int __space, __u32 mss, | |
171 | __u32 *rcv_wnd, __u32 *window_clamp, | |
172 | int wscale_ok, __u8 *rcv_wscale) | |
173 | { | |
174 | unsigned int space = (__space < 0 ? 0 : __space); | |
175 | ||
176 | /* If no clamp set the clamp to the max possible scaled window */ | |
177 | if (*window_clamp == 0) | |
178 | (*window_clamp) = (65535 << 14); | |
179 | space = min(*window_clamp, space); | |
180 | ||
181 | /* Quantize space offering to a multiple of mss if possible. */ | |
182 | if (space > mss) | |
183 | space = (space / mss) * mss; | |
184 | ||
185 | /* NOTE: offering an initial window larger than 32767 | |
15d99e02 RJ |
186 | * will break some buggy TCP stacks. If the admin tells us |
187 | * it is likely we could be speaking with such a buggy stack | |
188 | * we will truncate our initial window offering to 32K-1 | |
189 | * unless the remote has sent us a window scaling option, | |
190 | * which we interpret as a sign the remote TCP is not | |
191 | * misinterpreting the window field as a signed quantity. | |
1da177e4 | 192 | */ |
15d99e02 RJ |
193 | if (sysctl_tcp_workaround_signed_windows) |
194 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); | |
195 | else | |
196 | (*rcv_wnd) = space; | |
197 | ||
1da177e4 LT |
198 | (*rcv_wscale) = 0; |
199 | if (wscale_ok) { | |
200 | /* Set window scaling on max possible window | |
e905a9ed | 201 | * See RFC1323 for an explanation of the limit to 14 |
1da177e4 LT |
202 | */ |
203 | space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); | |
316c1592 | 204 | space = min_t(u32, space, *window_clamp); |
1da177e4 LT |
205 | while (space > 65535 && (*rcv_wscale) < 14) { |
206 | space >>= 1; | |
207 | (*rcv_wscale)++; | |
208 | } | |
209 | } | |
210 | ||
211 | /* Set initial window to value enough for senders, | |
6b251858 | 212 | * following RFC2414. Senders, not following this RFC, |
1da177e4 LT |
213 | * will be satisfied with 2. |
214 | */ | |
215 | if (mss > (1<<*rcv_wscale)) { | |
01ff367e DM |
216 | int init_cwnd = 4; |
217 | if (mss > 1460*3) | |
1da177e4 | 218 | init_cwnd = 2; |
01ff367e DM |
219 | else if (mss > 1460) |
220 | init_cwnd = 3; | |
1da177e4 LT |
221 | if (*rcv_wnd > init_cwnd*mss) |
222 | *rcv_wnd = init_cwnd*mss; | |
223 | } | |
224 | ||
225 | /* Set the clamp no higher than max representable value */ | |
226 | (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); | |
227 | } | |
228 | ||
229 | /* Chose a new window to advertise, update state in tcp_sock for the | |
230 | * socket, and return result with RFC1323 scaling applied. The return | |
231 | * value can be stuffed directly into th->window for an outgoing | |
232 | * frame. | |
233 | */ | |
40efc6fa | 234 | static u16 tcp_select_window(struct sock *sk) |
1da177e4 LT |
235 | { |
236 | struct tcp_sock *tp = tcp_sk(sk); | |
237 | u32 cur_win = tcp_receive_window(tp); | |
238 | u32 new_win = __tcp_select_window(sk); | |
239 | ||
240 | /* Never shrink the offered window */ | |
2de979bd | 241 | if (new_win < cur_win) { |
1da177e4 LT |
242 | /* Danger Will Robinson! |
243 | * Don't update rcv_wup/rcv_wnd here or else | |
244 | * we will not be able to advertise a zero | |
245 | * window in time. --DaveM | |
246 | * | |
247 | * Relax Will Robinson. | |
248 | */ | |
249 | new_win = cur_win; | |
250 | } | |
251 | tp->rcv_wnd = new_win; | |
252 | tp->rcv_wup = tp->rcv_nxt; | |
253 | ||
254 | /* Make sure we do not exceed the maximum possible | |
255 | * scaled window. | |
256 | */ | |
15d99e02 | 257 | if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) |
1da177e4 LT |
258 | new_win = min(new_win, MAX_TCP_WINDOW); |
259 | else | |
260 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); | |
261 | ||
262 | /* RFC1323 scaling applied */ | |
263 | new_win >>= tp->rx_opt.rcv_wscale; | |
264 | ||
265 | /* If we advertise zero window, disable fast path. */ | |
266 | if (new_win == 0) | |
267 | tp->pred_flags = 0; | |
268 | ||
269 | return new_win; | |
270 | } | |
271 | ||
df7a3b07 | 272 | static void tcp_build_and_update_options(__be32 *ptr, struct tcp_sock *tp, |
cfb6eeb4 | 273 | __u32 tstamp, __u8 **md5_hash) |
40efc6fa SH |
274 | { |
275 | if (tp->rx_opt.tstamp_ok) { | |
496c98df YH |
276 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
277 | (TCPOPT_NOP << 16) | | |
278 | (TCPOPT_TIMESTAMP << 8) | | |
279 | TCPOLEN_TIMESTAMP); | |
40efc6fa SH |
280 | *ptr++ = htonl(tstamp); |
281 | *ptr++ = htonl(tp->rx_opt.ts_recent); | |
282 | } | |
283 | if (tp->rx_opt.eff_sacks) { | |
284 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? tp->duplicate_sack : tp->selective_acks; | |
285 | int this_sack; | |
286 | ||
287 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
288 | (TCPOPT_NOP << 16) | | |
289 | (TCPOPT_SACK << 8) | | |
290 | (TCPOLEN_SACK_BASE + (tp->rx_opt.eff_sacks * | |
291 | TCPOLEN_SACK_PERBLOCK))); | |
2de979bd SH |
292 | |
293 | for (this_sack = 0; this_sack < tp->rx_opt.eff_sacks; this_sack++) { | |
40efc6fa SH |
294 | *ptr++ = htonl(sp[this_sack].start_seq); |
295 | *ptr++ = htonl(sp[this_sack].end_seq); | |
296 | } | |
2de979bd | 297 | |
40efc6fa SH |
298 | if (tp->rx_opt.dsack) { |
299 | tp->rx_opt.dsack = 0; | |
300 | tp->rx_opt.eff_sacks--; | |
301 | } | |
302 | } | |
cfb6eeb4 YH |
303 | #ifdef CONFIG_TCP_MD5SIG |
304 | if (md5_hash) { | |
305 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
306 | (TCPOPT_NOP << 16) | | |
307 | (TCPOPT_MD5SIG << 8) | | |
308 | TCPOLEN_MD5SIG); | |
309 | *md5_hash = (__u8 *)ptr; | |
310 | } | |
311 | #endif | |
40efc6fa SH |
312 | } |
313 | ||
314 | /* Construct a tcp options header for a SYN or SYN_ACK packet. | |
315 | * If this is every changed make sure to change the definition of | |
316 | * MAX_SYN_SIZE to match the new maximum number of options that you | |
317 | * can generate. | |
cfb6eeb4 YH |
318 | * |
319 | * Note - that with the RFC2385 TCP option, we make room for the | |
320 | * 16 byte MD5 hash. This will be filled in later, so the pointer for the | |
321 | * location to be filled is passed back up. | |
40efc6fa | 322 | */ |
df7a3b07 | 323 | static void tcp_syn_build_options(__be32 *ptr, int mss, int ts, int sack, |
40efc6fa | 324 | int offer_wscale, int wscale, __u32 tstamp, |
cfb6eeb4 | 325 | __u32 ts_recent, __u8 **md5_hash) |
40efc6fa SH |
326 | { |
327 | /* We always get an MSS option. | |
328 | * The option bytes which will be seen in normal data | |
329 | * packets should timestamps be used, must be in the MSS | |
330 | * advertised. But we subtract them from tp->mss_cache so | |
331 | * that calculations in tcp_sendmsg are simpler etc. | |
332 | * So account for this fact here if necessary. If we | |
333 | * don't do this correctly, as a receiver we won't | |
334 | * recognize data packets as being full sized when we | |
335 | * should, and thus we won't abide by the delayed ACK | |
336 | * rules correctly. | |
337 | * SACKs don't matter, we never delay an ACK when we | |
338 | * have any of those going out. | |
339 | */ | |
340 | *ptr++ = htonl((TCPOPT_MSS << 24) | (TCPOLEN_MSS << 16) | mss); | |
341 | if (ts) { | |
2de979bd | 342 | if (sack) |
496c98df YH |
343 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
344 | (TCPOLEN_SACK_PERM << 16) | | |
345 | (TCPOPT_TIMESTAMP << 8) | | |
346 | TCPOLEN_TIMESTAMP); | |
40efc6fa | 347 | else |
496c98df YH |
348 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
349 | (TCPOPT_NOP << 16) | | |
350 | (TCPOPT_TIMESTAMP << 8) | | |
351 | TCPOLEN_TIMESTAMP); | |
40efc6fa SH |
352 | *ptr++ = htonl(tstamp); /* TSVAL */ |
353 | *ptr++ = htonl(ts_recent); /* TSECR */ | |
2de979bd | 354 | } else if (sack) |
496c98df YH |
355 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
356 | (TCPOPT_NOP << 16) | | |
357 | (TCPOPT_SACK_PERM << 8) | | |
358 | TCPOLEN_SACK_PERM); | |
40efc6fa | 359 | if (offer_wscale) |
496c98df YH |
360 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
361 | (TCPOPT_WINDOW << 16) | | |
362 | (TCPOLEN_WINDOW << 8) | | |
363 | (wscale)); | |
cfb6eeb4 YH |
364 | #ifdef CONFIG_TCP_MD5SIG |
365 | /* | |
366 | * If MD5 is enabled, then we set the option, and include the size | |
367 | * (always 18). The actual MD5 hash is added just before the | |
368 | * packet is sent. | |
369 | */ | |
370 | if (md5_hash) { | |
371 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
372 | (TCPOPT_NOP << 16) | | |
373 | (TCPOPT_MD5SIG << 8) | | |
374 | TCPOLEN_MD5SIG); | |
375 | *md5_hash = (__u8 *) ptr; | |
376 | } | |
377 | #endif | |
40efc6fa | 378 | } |
1da177e4 LT |
379 | |
380 | /* This routine actually transmits TCP packets queued in by | |
381 | * tcp_do_sendmsg(). This is used by both the initial | |
382 | * transmission and possible later retransmissions. | |
383 | * All SKB's seen here are completely headerless. It is our | |
384 | * job to build the TCP header, and pass the packet down to | |
385 | * IP so it can do the same plus pass the packet off to the | |
386 | * device. | |
387 | * | |
388 | * We are working here with either a clone of the original | |
389 | * SKB, or a fresh unique copy made by the retransmit engine. | |
390 | */ | |
dfb4b9dc | 391 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, gfp_t gfp_mask) |
1da177e4 | 392 | { |
dfb4b9dc DM |
393 | const struct inet_connection_sock *icsk = inet_csk(sk); |
394 | struct inet_sock *inet; | |
395 | struct tcp_sock *tp; | |
396 | struct tcp_skb_cb *tcb; | |
397 | int tcp_header_size; | |
cfb6eeb4 YH |
398 | #ifdef CONFIG_TCP_MD5SIG |
399 | struct tcp_md5sig_key *md5; | |
400 | __u8 *md5_hash_location; | |
401 | #endif | |
dfb4b9dc DM |
402 | struct tcphdr *th; |
403 | int sysctl_flags; | |
404 | int err; | |
405 | ||
406 | BUG_ON(!skb || !tcp_skb_pcount(skb)); | |
407 | ||
408 | /* If congestion control is doing timestamping, we must | |
409 | * take such a timestamp before we potentially clone/copy. | |
410 | */ | |
164891aa | 411 | if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) |
dfb4b9dc DM |
412 | __net_timestamp(skb); |
413 | ||
414 | if (likely(clone_it)) { | |
415 | if (unlikely(skb_cloned(skb))) | |
416 | skb = pskb_copy(skb, gfp_mask); | |
417 | else | |
418 | skb = skb_clone(skb, gfp_mask); | |
419 | if (unlikely(!skb)) | |
420 | return -ENOBUFS; | |
421 | } | |
1da177e4 | 422 | |
dfb4b9dc DM |
423 | inet = inet_sk(sk); |
424 | tp = tcp_sk(sk); | |
425 | tcb = TCP_SKB_CB(skb); | |
426 | tcp_header_size = tp->tcp_header_len; | |
1da177e4 LT |
427 | |
428 | #define SYSCTL_FLAG_TSTAMPS 0x1 | |
429 | #define SYSCTL_FLAG_WSCALE 0x2 | |
430 | #define SYSCTL_FLAG_SACK 0x4 | |
431 | ||
dfb4b9dc DM |
432 | sysctl_flags = 0; |
433 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { | |
434 | tcp_header_size = sizeof(struct tcphdr) + TCPOLEN_MSS; | |
2de979bd | 435 | if (sysctl_tcp_timestamps) { |
dfb4b9dc DM |
436 | tcp_header_size += TCPOLEN_TSTAMP_ALIGNED; |
437 | sysctl_flags |= SYSCTL_FLAG_TSTAMPS; | |
1da177e4 | 438 | } |
dfb4b9dc DM |
439 | if (sysctl_tcp_window_scaling) { |
440 | tcp_header_size += TCPOLEN_WSCALE_ALIGNED; | |
441 | sysctl_flags |= SYSCTL_FLAG_WSCALE; | |
1da177e4 | 442 | } |
dfb4b9dc DM |
443 | if (sysctl_tcp_sack) { |
444 | sysctl_flags |= SYSCTL_FLAG_SACK; | |
445 | if (!(sysctl_flags & SYSCTL_FLAG_TSTAMPS)) | |
446 | tcp_header_size += TCPOLEN_SACKPERM_ALIGNED; | |
1da177e4 | 447 | } |
dfb4b9dc DM |
448 | } else if (unlikely(tp->rx_opt.eff_sacks)) { |
449 | /* A SACK is 2 pad bytes, a 2 byte header, plus | |
450 | * 2 32-bit sequence numbers for each SACK block. | |
451 | */ | |
452 | tcp_header_size += (TCPOLEN_SACK_BASE_ALIGNED + | |
453 | (tp->rx_opt.eff_sacks * | |
454 | TCPOLEN_SACK_PERBLOCK)); | |
455 | } | |
e905a9ed | 456 | |
dfb4b9dc DM |
457 | if (tcp_packets_in_flight(tp) == 0) |
458 | tcp_ca_event(sk, CA_EVENT_TX_START); | |
459 | ||
cfb6eeb4 YH |
460 | #ifdef CONFIG_TCP_MD5SIG |
461 | /* | |
462 | * Are we doing MD5 on this segment? If so - make | |
463 | * room for it. | |
464 | */ | |
465 | md5 = tp->af_specific->md5_lookup(sk, sk); | |
466 | if (md5) | |
467 | tcp_header_size += TCPOLEN_MD5SIG_ALIGNED; | |
468 | #endif | |
469 | ||
aa8223c7 ACM |
470 | skb_push(skb, tcp_header_size); |
471 | skb_reset_transport_header(skb); | |
e89862f4 | 472 | skb_set_owner_w(skb, sk); |
dfb4b9dc DM |
473 | |
474 | /* Build TCP header and checksum it. */ | |
aa8223c7 | 475 | th = tcp_hdr(skb); |
dfb4b9dc DM |
476 | th->source = inet->sport; |
477 | th->dest = inet->dport; | |
478 | th->seq = htonl(tcb->seq); | |
479 | th->ack_seq = htonl(tp->rcv_nxt); | |
df7a3b07 | 480 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
dfb4b9dc DM |
481 | tcb->flags); |
482 | ||
483 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { | |
484 | /* RFC1323: The window in SYN & SYN/ACK segments | |
485 | * is never scaled. | |
486 | */ | |
600ff0c2 | 487 | th->window = htons(min(tp->rcv_wnd, 65535U)); |
dfb4b9dc DM |
488 | } else { |
489 | th->window = htons(tcp_select_window(sk)); | |
490 | } | |
491 | th->check = 0; | |
492 | th->urg_ptr = 0; | |
1da177e4 | 493 | |
dfb4b9dc DM |
494 | if (unlikely(tp->urg_mode && |
495 | between(tp->snd_up, tcb->seq+1, tcb->seq+0xFFFF))) { | |
496 | th->urg_ptr = htons(tp->snd_up-tcb->seq); | |
497 | th->urg = 1; | |
498 | } | |
1da177e4 | 499 | |
dfb4b9dc | 500 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { |
df7a3b07 | 501 | tcp_syn_build_options((__be32 *)(th + 1), |
dfb4b9dc DM |
502 | tcp_advertise_mss(sk), |
503 | (sysctl_flags & SYSCTL_FLAG_TSTAMPS), | |
504 | (sysctl_flags & SYSCTL_FLAG_SACK), | |
505 | (sysctl_flags & SYSCTL_FLAG_WSCALE), | |
506 | tp->rx_opt.rcv_wscale, | |
507 | tcb->when, | |
cfb6eeb4 YH |
508 | tp->rx_opt.ts_recent, |
509 | ||
510 | #ifdef CONFIG_TCP_MD5SIG | |
511 | md5 ? &md5_hash_location : | |
512 | #endif | |
513 | NULL); | |
dfb4b9dc | 514 | } else { |
df7a3b07 | 515 | tcp_build_and_update_options((__be32 *)(th + 1), |
cfb6eeb4 YH |
516 | tp, tcb->when, |
517 | #ifdef CONFIG_TCP_MD5SIG | |
518 | md5 ? &md5_hash_location : | |
519 | #endif | |
520 | NULL); | |
9e412ba7 | 521 | TCP_ECN_send(sk, skb, tcp_header_size); |
dfb4b9dc | 522 | } |
1da177e4 | 523 | |
cfb6eeb4 YH |
524 | #ifdef CONFIG_TCP_MD5SIG |
525 | /* Calculate the MD5 hash, as we have all we need now */ | |
526 | if (md5) { | |
527 | tp->af_specific->calc_md5_hash(md5_hash_location, | |
528 | md5, | |
529 | sk, NULL, NULL, | |
aa8223c7 | 530 | tcp_hdr(skb), |
cfb6eeb4 YH |
531 | sk->sk_protocol, |
532 | skb->len); | |
533 | } | |
534 | #endif | |
535 | ||
8292a17a | 536 | icsk->icsk_af_ops->send_check(sk, skb->len, skb); |
1da177e4 | 537 | |
dfb4b9dc DM |
538 | if (likely(tcb->flags & TCPCB_FLAG_ACK)) |
539 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); | |
1da177e4 | 540 | |
dfb4b9dc DM |
541 | if (skb->len != tcp_header_size) |
542 | tcp_event_data_sent(tp, skb, sk); | |
1da177e4 | 543 | |
bd37a088 WY |
544 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
545 | TCP_INC_STATS(TCP_MIB_OUTSEGS); | |
1da177e4 | 546 | |
e89862f4 | 547 | err = icsk->icsk_af_ops->queue_xmit(skb, 0); |
83de47cd | 548 | if (likely(err <= 0)) |
dfb4b9dc DM |
549 | return err; |
550 | ||
3cfe3baa | 551 | tcp_enter_cwr(sk, 1); |
dfb4b9dc | 552 | |
b9df3cb8 | 553 | return net_xmit_eval(err); |
1da177e4 | 554 | |
1da177e4 LT |
555 | #undef SYSCTL_FLAG_TSTAMPS |
556 | #undef SYSCTL_FLAG_WSCALE | |
557 | #undef SYSCTL_FLAG_SACK | |
558 | } | |
559 | ||
560 | ||
e905a9ed | 561 | /* This routine just queue's the buffer |
1da177e4 LT |
562 | * |
563 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, | |
564 | * otherwise socket can stall. | |
565 | */ | |
566 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) | |
567 | { | |
568 | struct tcp_sock *tp = tcp_sk(sk); | |
569 | ||
570 | /* Advance write_seq and place onto the write_queue. */ | |
571 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; | |
572 | skb_header_release(skb); | |
fe067e8a | 573 | tcp_add_write_queue_tail(sk, skb); |
1da177e4 | 574 | sk_charge_skb(sk, skb); |
1da177e4 LT |
575 | } |
576 | ||
846998ae | 577 | static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now) |
f6302d1d | 578 | { |
bcd76111 | 579 | if (skb->len <= mss_now || !sk_can_gso(sk)) { |
f6302d1d DM |
580 | /* Avoid the costly divide in the normal |
581 | * non-TSO case. | |
582 | */ | |
7967168c HX |
583 | skb_shinfo(skb)->gso_segs = 1; |
584 | skb_shinfo(skb)->gso_size = 0; | |
585 | skb_shinfo(skb)->gso_type = 0; | |
f6302d1d DM |
586 | } else { |
587 | unsigned int factor; | |
588 | ||
846998ae DM |
589 | factor = skb->len + (mss_now - 1); |
590 | factor /= mss_now; | |
7967168c HX |
591 | skb_shinfo(skb)->gso_segs = factor; |
592 | skb_shinfo(skb)->gso_size = mss_now; | |
bcd76111 | 593 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
1da177e4 LT |
594 | } |
595 | } | |
596 | ||
1da177e4 LT |
597 | /* Function to create two new TCP segments. Shrinks the given segment |
598 | * to the specified size and appends a new segment with the rest of the | |
e905a9ed | 599 | * packet to the list. This won't be called frequently, I hope. |
1da177e4 LT |
600 | * Remember, these are still headerless SKBs at this point. |
601 | */ | |
6475be16 | 602 | int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, unsigned int mss_now) |
1da177e4 LT |
603 | { |
604 | struct tcp_sock *tp = tcp_sk(sk); | |
605 | struct sk_buff *buff; | |
6475be16 | 606 | int nsize, old_factor; |
b60b49ea | 607 | int nlen; |
1da177e4 LT |
608 | u16 flags; |
609 | ||
b2cc99f0 | 610 | BUG_ON(len > skb->len); |
6a438bbe | 611 | |
e905a9ed | 612 | clear_all_retrans_hints(tp); |
1da177e4 LT |
613 | nsize = skb_headlen(skb) - len; |
614 | if (nsize < 0) | |
615 | nsize = 0; | |
616 | ||
617 | if (skb_cloned(skb) && | |
618 | skb_is_nonlinear(skb) && | |
619 | pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) | |
620 | return -ENOMEM; | |
621 | ||
622 | /* Get a new skb... force flag on. */ | |
623 | buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); | |
624 | if (buff == NULL) | |
625 | return -ENOMEM; /* We'll just try again later. */ | |
ef5cb973 | 626 | |
b60b49ea HX |
627 | sk_charge_skb(sk, buff); |
628 | nlen = skb->len - len - nsize; | |
629 | buff->truesize += nlen; | |
630 | skb->truesize -= nlen; | |
1da177e4 LT |
631 | |
632 | /* Correct the sequence numbers. */ | |
633 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
634 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
635 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
636 | ||
637 | /* PSH and FIN should only be set in the second packet. */ | |
638 | flags = TCP_SKB_CB(skb)->flags; | |
639 | TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); | |
640 | TCP_SKB_CB(buff)->flags = flags; | |
e14c3caf | 641 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
1da177e4 LT |
642 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_AT_TAIL; |
643 | ||
84fa7933 | 644 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { |
1da177e4 LT |
645 | /* Copy and checksum data tail into the new buffer. */ |
646 | buff->csum = csum_partial_copy_nocheck(skb->data + len, skb_put(buff, nsize), | |
647 | nsize, 0); | |
648 | ||
649 | skb_trim(skb, len); | |
650 | ||
651 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); | |
652 | } else { | |
84fa7933 | 653 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
654 | skb_split(skb, buff, len); |
655 | } | |
656 | ||
657 | buff->ip_summed = skb->ip_summed; | |
658 | ||
659 | /* Looks stupid, but our code really uses when of | |
660 | * skbs, which it never sent before. --ANK | |
661 | */ | |
662 | TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; | |
a61bbcf2 | 663 | buff->tstamp = skb->tstamp; |
1da177e4 | 664 | |
6475be16 DM |
665 | old_factor = tcp_skb_pcount(skb); |
666 | ||
1da177e4 | 667 | /* Fix up tso_factor for both original and new SKB. */ |
846998ae DM |
668 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
669 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
1da177e4 | 670 | |
6475be16 DM |
671 | /* If this packet has been sent out already, we must |
672 | * adjust the various packet counters. | |
673 | */ | |
cf0b450c | 674 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
6475be16 DM |
675 | int diff = old_factor - tcp_skb_pcount(skb) - |
676 | tcp_skb_pcount(buff); | |
1da177e4 | 677 | |
6475be16 | 678 | tp->packets_out -= diff; |
e14c3caf HX |
679 | |
680 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
681 | tp->sacked_out -= diff; | |
682 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
683 | tp->retrans_out -= diff; | |
684 | ||
6475be16 DM |
685 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) { |
686 | tp->lost_out -= diff; | |
687 | tp->left_out -= diff; | |
688 | } | |
83ca28be | 689 | |
6475be16 | 690 | if (diff > 0) { |
83ca28be HX |
691 | /* Adjust Reno SACK estimate. */ |
692 | if (!tp->rx_opt.sack_ok) { | |
693 | tp->sacked_out -= diff; | |
694 | if ((int)tp->sacked_out < 0) | |
695 | tp->sacked_out = 0; | |
696 | tcp_sync_left_out(tp); | |
697 | } | |
698 | ||
6475be16 DM |
699 | tp->fackets_out -= diff; |
700 | if ((int)tp->fackets_out < 0) | |
701 | tp->fackets_out = 0; | |
d0410051 IJ |
702 | /* SACK fastpath might overwrite it unless dealt with */ |
703 | if (tp->fastpath_skb_hint != NULL && | |
704 | after(TCP_SKB_CB(tp->fastpath_skb_hint)->seq, | |
705 | TCP_SKB_CB(skb)->seq)) { | |
706 | tp->fastpath_cnt_hint -= diff; | |
707 | if ((int)tp->fastpath_cnt_hint < 0) | |
708 | tp->fastpath_cnt_hint = 0; | |
709 | } | |
6475be16 | 710 | } |
1da177e4 LT |
711 | } |
712 | ||
713 | /* Link BUFF into the send queue. */ | |
f44b5271 | 714 | skb_header_release(buff); |
fe067e8a | 715 | tcp_insert_write_queue_after(skb, buff, sk); |
1da177e4 LT |
716 | |
717 | return 0; | |
718 | } | |
719 | ||
720 | /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c | |
721 | * eventually). The difference is that pulled data not copied, but | |
722 | * immediately discarded. | |
723 | */ | |
f2911969 | 724 | static void __pskb_trim_head(struct sk_buff *skb, int len) |
1da177e4 LT |
725 | { |
726 | int i, k, eat; | |
727 | ||
728 | eat = len; | |
729 | k = 0; | |
730 | for (i=0; i<skb_shinfo(skb)->nr_frags; i++) { | |
731 | if (skb_shinfo(skb)->frags[i].size <= eat) { | |
732 | put_page(skb_shinfo(skb)->frags[i].page); | |
733 | eat -= skb_shinfo(skb)->frags[i].size; | |
734 | } else { | |
735 | skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; | |
736 | if (eat) { | |
737 | skb_shinfo(skb)->frags[k].page_offset += eat; | |
738 | skb_shinfo(skb)->frags[k].size -= eat; | |
739 | eat = 0; | |
740 | } | |
741 | k++; | |
742 | } | |
743 | } | |
744 | skb_shinfo(skb)->nr_frags = k; | |
745 | ||
27a884dc | 746 | skb_reset_tail_pointer(skb); |
1da177e4 LT |
747 | skb->data_len -= len; |
748 | skb->len = skb->data_len; | |
1da177e4 LT |
749 | } |
750 | ||
751 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) | |
752 | { | |
753 | if (skb_cloned(skb) && | |
754 | pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) | |
755 | return -ENOMEM; | |
756 | ||
f2911969 HX |
757 | /* If len == headlen, we avoid __skb_pull to preserve alignment. */ |
758 | if (unlikely(len < skb_headlen(skb))) | |
1da177e4 | 759 | __skb_pull(skb, len); |
f2911969 HX |
760 | else |
761 | __pskb_trim_head(skb, len - skb_headlen(skb)); | |
1da177e4 LT |
762 | |
763 | TCP_SKB_CB(skb)->seq += len; | |
84fa7933 | 764 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
765 | |
766 | skb->truesize -= len; | |
767 | sk->sk_wmem_queued -= len; | |
768 | sk->sk_forward_alloc += len; | |
769 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); | |
770 | ||
771 | /* Any change of skb->len requires recalculation of tso | |
772 | * factor and mss. | |
773 | */ | |
774 | if (tcp_skb_pcount(skb) > 1) | |
846998ae | 775 | tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1)); |
1da177e4 LT |
776 | |
777 | return 0; | |
778 | } | |
779 | ||
5d424d5a JH |
780 | /* Not accounting for SACKs here. */ |
781 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
782 | { | |
783 | struct tcp_sock *tp = tcp_sk(sk); | |
784 | struct inet_connection_sock *icsk = inet_csk(sk); | |
785 | int mss_now; | |
786 | ||
787 | /* Calculate base mss without TCP options: | |
788 | It is MMS_S - sizeof(tcphdr) of rfc1122 | |
789 | */ | |
790 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); | |
791 | ||
792 | /* Clamp it (mss_clamp does not include tcp options) */ | |
793 | if (mss_now > tp->rx_opt.mss_clamp) | |
794 | mss_now = tp->rx_opt.mss_clamp; | |
795 | ||
796 | /* Now subtract optional transport overhead */ | |
797 | mss_now -= icsk->icsk_ext_hdr_len; | |
798 | ||
799 | /* Then reserve room for full set of TCP options and 8 bytes of data */ | |
800 | if (mss_now < 48) | |
801 | mss_now = 48; | |
802 | ||
803 | /* Now subtract TCP options size, not including SACKs */ | |
804 | mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); | |
805 | ||
806 | return mss_now; | |
807 | } | |
808 | ||
809 | /* Inverse of above */ | |
810 | int tcp_mss_to_mtu(struct sock *sk, int mss) | |
811 | { | |
812 | struct tcp_sock *tp = tcp_sk(sk); | |
813 | struct inet_connection_sock *icsk = inet_csk(sk); | |
814 | int mtu; | |
815 | ||
816 | mtu = mss + | |
817 | tp->tcp_header_len + | |
818 | icsk->icsk_ext_hdr_len + | |
819 | icsk->icsk_af_ops->net_header_len; | |
820 | ||
821 | return mtu; | |
822 | } | |
823 | ||
824 | void tcp_mtup_init(struct sock *sk) | |
825 | { | |
826 | struct tcp_sock *tp = tcp_sk(sk); | |
827 | struct inet_connection_sock *icsk = inet_csk(sk); | |
828 | ||
829 | icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; | |
830 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + | |
e905a9ed | 831 | icsk->icsk_af_ops->net_header_len; |
5d424d5a JH |
832 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); |
833 | icsk->icsk_mtup.probe_size = 0; | |
834 | } | |
835 | ||
1da177e4 LT |
836 | /* This function synchronize snd mss to current pmtu/exthdr set. |
837 | ||
838 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts | |
839 | for TCP options, but includes only bare TCP header. | |
840 | ||
841 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. | |
caa20d9a | 842 | It is minimum of user_mss and mss received with SYN. |
1da177e4 LT |
843 | It also does not include TCP options. |
844 | ||
d83d8461 | 845 | inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. |
1da177e4 LT |
846 | |
847 | tp->mss_cache is current effective sending mss, including | |
848 | all tcp options except for SACKs. It is evaluated, | |
849 | taking into account current pmtu, but never exceeds | |
850 | tp->rx_opt.mss_clamp. | |
851 | ||
852 | NOTE1. rfc1122 clearly states that advertised MSS | |
853 | DOES NOT include either tcp or ip options. | |
854 | ||
d83d8461 ACM |
855 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
856 | are READ ONLY outside this function. --ANK (980731) | |
1da177e4 LT |
857 | */ |
858 | ||
859 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) | |
860 | { | |
861 | struct tcp_sock *tp = tcp_sk(sk); | |
d83d8461 | 862 | struct inet_connection_sock *icsk = inet_csk(sk); |
5d424d5a | 863 | int mss_now; |
1da177e4 | 864 | |
5d424d5a JH |
865 | if (icsk->icsk_mtup.search_high > pmtu) |
866 | icsk->icsk_mtup.search_high = pmtu; | |
1da177e4 | 867 | |
5d424d5a | 868 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
1da177e4 LT |
869 | |
870 | /* Bound mss with half of window */ | |
871 | if (tp->max_window && mss_now > (tp->max_window>>1)) | |
872 | mss_now = max((tp->max_window>>1), 68U - tp->tcp_header_len); | |
873 | ||
874 | /* And store cached results */ | |
d83d8461 | 875 | icsk->icsk_pmtu_cookie = pmtu; |
5d424d5a JH |
876 | if (icsk->icsk_mtup.enabled) |
877 | mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); | |
c1b4a7e6 | 878 | tp->mss_cache = mss_now; |
1da177e4 LT |
879 | |
880 | return mss_now; | |
881 | } | |
882 | ||
883 | /* Compute the current effective MSS, taking SACKs and IP options, | |
884 | * and even PMTU discovery events into account. | |
885 | * | |
886 | * LARGESEND note: !urg_mode is overkill, only frames up to snd_up | |
887 | * cannot be large. However, taking into account rare use of URG, this | |
888 | * is not a big flaw. | |
889 | */ | |
c1b4a7e6 | 890 | unsigned int tcp_current_mss(struct sock *sk, int large_allowed) |
1da177e4 LT |
891 | { |
892 | struct tcp_sock *tp = tcp_sk(sk); | |
893 | struct dst_entry *dst = __sk_dst_get(sk); | |
c1b4a7e6 DM |
894 | u32 mss_now; |
895 | u16 xmit_size_goal; | |
896 | int doing_tso = 0; | |
897 | ||
898 | mss_now = tp->mss_cache; | |
899 | ||
bcd76111 | 900 | if (large_allowed && sk_can_gso(sk) && !tp->urg_mode) |
c1b4a7e6 | 901 | doing_tso = 1; |
1da177e4 | 902 | |
1da177e4 LT |
903 | if (dst) { |
904 | u32 mtu = dst_mtu(dst); | |
d83d8461 | 905 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
1da177e4 LT |
906 | mss_now = tcp_sync_mss(sk, mtu); |
907 | } | |
908 | ||
c1b4a7e6 DM |
909 | if (tp->rx_opt.eff_sacks) |
910 | mss_now -= (TCPOLEN_SACK_BASE_ALIGNED + | |
911 | (tp->rx_opt.eff_sacks * TCPOLEN_SACK_PERBLOCK)); | |
1da177e4 | 912 | |
cfb6eeb4 YH |
913 | #ifdef CONFIG_TCP_MD5SIG |
914 | if (tp->af_specific->md5_lookup(sk, sk)) | |
915 | mss_now -= TCPOLEN_MD5SIG_ALIGNED; | |
916 | #endif | |
917 | ||
c1b4a7e6 | 918 | xmit_size_goal = mss_now; |
1da177e4 | 919 | |
c1b4a7e6 | 920 | if (doing_tso) { |
8292a17a ACM |
921 | xmit_size_goal = (65535 - |
922 | inet_csk(sk)->icsk_af_ops->net_header_len - | |
d83d8461 ACM |
923 | inet_csk(sk)->icsk_ext_hdr_len - |
924 | tp->tcp_header_len); | |
1da177e4 | 925 | |
c1b4a7e6 DM |
926 | if (tp->max_window && |
927 | (xmit_size_goal > (tp->max_window >> 1))) | |
928 | xmit_size_goal = max((tp->max_window >> 1), | |
929 | 68U - tp->tcp_header_len); | |
1da177e4 | 930 | |
c1b4a7e6 | 931 | xmit_size_goal -= (xmit_size_goal % mss_now); |
1da177e4 | 932 | } |
c1b4a7e6 | 933 | tp->xmit_size_goal = xmit_size_goal; |
1da177e4 | 934 | |
1da177e4 LT |
935 | return mss_now; |
936 | } | |
937 | ||
a762a980 DM |
938 | /* Congestion window validation. (RFC2861) */ |
939 | ||
9e412ba7 | 940 | static void tcp_cwnd_validate(struct sock *sk) |
a762a980 | 941 | { |
9e412ba7 | 942 | struct tcp_sock *tp = tcp_sk(sk); |
a762a980 DM |
943 | __u32 packets_out = tp->packets_out; |
944 | ||
945 | if (packets_out >= tp->snd_cwnd) { | |
946 | /* Network is feed fully. */ | |
947 | tp->snd_cwnd_used = 0; | |
948 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
949 | } else { | |
950 | /* Network starves. */ | |
951 | if (tp->packets_out > tp->snd_cwnd_used) | |
952 | tp->snd_cwnd_used = tp->packets_out; | |
953 | ||
15d33c07 DM |
954 | if (sysctl_tcp_slow_start_after_idle && |
955 | (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) | |
a762a980 DM |
956 | tcp_cwnd_application_limited(sk); |
957 | } | |
958 | } | |
959 | ||
c1b4a7e6 DM |
960 | static unsigned int tcp_window_allows(struct tcp_sock *tp, struct sk_buff *skb, unsigned int mss_now, unsigned int cwnd) |
961 | { | |
962 | u32 window, cwnd_len; | |
963 | ||
964 | window = (tp->snd_una + tp->snd_wnd - TCP_SKB_CB(skb)->seq); | |
965 | cwnd_len = mss_now * cwnd; | |
966 | return min(window, cwnd_len); | |
967 | } | |
968 | ||
969 | /* Can at least one segment of SKB be sent right now, according to the | |
970 | * congestion window rules? If so, return how many segments are allowed. | |
971 | */ | |
972 | static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, struct sk_buff *skb) | |
973 | { | |
974 | u32 in_flight, cwnd; | |
975 | ||
976 | /* Don't be strict about the congestion window for the final FIN. */ | |
104439a8 JH |
977 | if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && |
978 | tcp_skb_pcount(skb) == 1) | |
c1b4a7e6 DM |
979 | return 1; |
980 | ||
981 | in_flight = tcp_packets_in_flight(tp); | |
982 | cwnd = tp->snd_cwnd; | |
983 | if (in_flight < cwnd) | |
984 | return (cwnd - in_flight); | |
985 | ||
986 | return 0; | |
987 | } | |
988 | ||
989 | /* This must be invoked the first time we consider transmitting | |
990 | * SKB onto the wire. | |
991 | */ | |
40efc6fa | 992 | static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, unsigned int mss_now) |
c1b4a7e6 DM |
993 | { |
994 | int tso_segs = tcp_skb_pcount(skb); | |
995 | ||
846998ae DM |
996 | if (!tso_segs || |
997 | (tso_segs > 1 && | |
7967168c | 998 | tcp_skb_mss(skb) != mss_now)) { |
846998ae | 999 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
c1b4a7e6 DM |
1000 | tso_segs = tcp_skb_pcount(skb); |
1001 | } | |
1002 | return tso_segs; | |
1003 | } | |
1004 | ||
1005 | static inline int tcp_minshall_check(const struct tcp_sock *tp) | |
1006 | { | |
1007 | return after(tp->snd_sml,tp->snd_una) && | |
1008 | !after(tp->snd_sml, tp->snd_nxt); | |
1009 | } | |
1010 | ||
1011 | /* Return 0, if packet can be sent now without violation Nagle's rules: | |
1012 | * 1. It is full sized. | |
1013 | * 2. Or it contains FIN. (already checked by caller) | |
1014 | * 3. Or TCP_NODELAY was set. | |
1015 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. | |
1016 | * With Minshall's modification: all sent small packets are ACKed. | |
1017 | */ | |
1018 | ||
1019 | static inline int tcp_nagle_check(const struct tcp_sock *tp, | |
e905a9ed | 1020 | const struct sk_buff *skb, |
c1b4a7e6 DM |
1021 | unsigned mss_now, int nonagle) |
1022 | { | |
1023 | return (skb->len < mss_now && | |
1024 | ((nonagle&TCP_NAGLE_CORK) || | |
1025 | (!nonagle && | |
1026 | tp->packets_out && | |
1027 | tcp_minshall_check(tp)))); | |
1028 | } | |
1029 | ||
1030 | /* Return non-zero if the Nagle test allows this packet to be | |
1031 | * sent now. | |
1032 | */ | |
1033 | static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, | |
1034 | unsigned int cur_mss, int nonagle) | |
1035 | { | |
1036 | /* Nagle rule does not apply to frames, which sit in the middle of the | |
1037 | * write_queue (they have no chances to get new data). | |
1038 | * | |
1039 | * This is implemented in the callers, where they modify the 'nonagle' | |
1040 | * argument based upon the location of SKB in the send queue. | |
1041 | */ | |
1042 | if (nonagle & TCP_NAGLE_PUSH) | |
1043 | return 1; | |
1044 | ||
d551e454 IJ |
1045 | /* Don't use the nagle rule for urgent data (or for the final FIN). |
1046 | * Nagle can be ignored during F-RTO too (see RFC4138). | |
1047 | */ | |
1048 | if (tp->urg_mode || (tp->frto_counter == 2) || | |
c1b4a7e6 DM |
1049 | (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)) |
1050 | return 1; | |
1051 | ||
1052 | if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) | |
1053 | return 1; | |
1054 | ||
1055 | return 0; | |
1056 | } | |
1057 | ||
1058 | /* Does at least the first segment of SKB fit into the send window? */ | |
1059 | static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, unsigned int cur_mss) | |
1060 | { | |
1061 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
1062 | ||
1063 | if (skb->len > cur_mss) | |
1064 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; | |
1065 | ||
1066 | return !after(end_seq, tp->snd_una + tp->snd_wnd); | |
1067 | } | |
1068 | ||
fe067e8a | 1069 | /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) |
c1b4a7e6 DM |
1070 | * should be put on the wire right now. If so, it returns the number of |
1071 | * packets allowed by the congestion window. | |
1072 | */ | |
1073 | static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, | |
1074 | unsigned int cur_mss, int nonagle) | |
1075 | { | |
1076 | struct tcp_sock *tp = tcp_sk(sk); | |
1077 | unsigned int cwnd_quota; | |
1078 | ||
846998ae | 1079 | tcp_init_tso_segs(sk, skb, cur_mss); |
c1b4a7e6 DM |
1080 | |
1081 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) | |
1082 | return 0; | |
1083 | ||
1084 | cwnd_quota = tcp_cwnd_test(tp, skb); | |
1085 | if (cwnd_quota && | |
1086 | !tcp_snd_wnd_test(tp, skb, cur_mss)) | |
1087 | cwnd_quota = 0; | |
1088 | ||
1089 | return cwnd_quota; | |
1090 | } | |
1091 | ||
9e412ba7 | 1092 | int tcp_may_send_now(struct sock *sk) |
c1b4a7e6 | 1093 | { |
9e412ba7 | 1094 | struct tcp_sock *tp = tcp_sk(sk); |
fe067e8a | 1095 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 DM |
1096 | |
1097 | return (skb && | |
1098 | tcp_snd_test(sk, skb, tcp_current_mss(sk, 1), | |
1099 | (tcp_skb_is_last(sk, skb) ? | |
1100 | TCP_NAGLE_PUSH : | |
1101 | tp->nonagle))); | |
1102 | } | |
1103 | ||
1104 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet | |
1105 | * which is put after SKB on the list. It is very much like | |
1106 | * tcp_fragment() except that it may make several kinds of assumptions | |
1107 | * in order to speed up the splitting operation. In particular, we | |
1108 | * know that all the data is in scatter-gather pages, and that the | |
1109 | * packet has never been sent out before (and thus is not cloned). | |
1110 | */ | |
846998ae | 1111 | static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, unsigned int mss_now) |
c1b4a7e6 DM |
1112 | { |
1113 | struct sk_buff *buff; | |
1114 | int nlen = skb->len - len; | |
1115 | u16 flags; | |
1116 | ||
1117 | /* All of a TSO frame must be composed of paged data. */ | |
c8ac3774 HX |
1118 | if (skb->len != skb->data_len) |
1119 | return tcp_fragment(sk, skb, len, mss_now); | |
c1b4a7e6 DM |
1120 | |
1121 | buff = sk_stream_alloc_pskb(sk, 0, 0, GFP_ATOMIC); | |
1122 | if (unlikely(buff == NULL)) | |
1123 | return -ENOMEM; | |
1124 | ||
b60b49ea HX |
1125 | sk_charge_skb(sk, buff); |
1126 | buff->truesize += nlen; | |
c1b4a7e6 DM |
1127 | skb->truesize -= nlen; |
1128 | ||
1129 | /* Correct the sequence numbers. */ | |
1130 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1131 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1132 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1133 | ||
1134 | /* PSH and FIN should only be set in the second packet. */ | |
1135 | flags = TCP_SKB_CB(skb)->flags; | |
1136 | TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); | |
1137 | TCP_SKB_CB(buff)->flags = flags; | |
1138 | ||
1139 | /* This packet was never sent out yet, so no SACK bits. */ | |
1140 | TCP_SKB_CB(buff)->sacked = 0; | |
1141 | ||
84fa7933 | 1142 | buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; |
c1b4a7e6 DM |
1143 | skb_split(skb, buff, len); |
1144 | ||
1145 | /* Fix up tso_factor for both original and new SKB. */ | |
846998ae DM |
1146 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
1147 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
c1b4a7e6 DM |
1148 | |
1149 | /* Link BUFF into the send queue. */ | |
1150 | skb_header_release(buff); | |
fe067e8a | 1151 | tcp_insert_write_queue_after(skb, buff, sk); |
c1b4a7e6 DM |
1152 | |
1153 | return 0; | |
1154 | } | |
1155 | ||
1156 | /* Try to defer sending, if possible, in order to minimize the amount | |
1157 | * of TSO splitting we do. View it as a kind of TSO Nagle test. | |
1158 | * | |
1159 | * This algorithm is from John Heffner. | |
1160 | */ | |
9e412ba7 | 1161 | static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) |
c1b4a7e6 | 1162 | { |
9e412ba7 | 1163 | struct tcp_sock *tp = tcp_sk(sk); |
6687e988 | 1164 | const struct inet_connection_sock *icsk = inet_csk(sk); |
c1b4a7e6 DM |
1165 | u32 send_win, cong_win, limit, in_flight; |
1166 | ||
1167 | if (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) | |
ae8064ac | 1168 | goto send_now; |
c1b4a7e6 | 1169 | |
6687e988 | 1170 | if (icsk->icsk_ca_state != TCP_CA_Open) |
ae8064ac JH |
1171 | goto send_now; |
1172 | ||
1173 | /* Defer for less than two clock ticks. */ | |
1174 | if (!tp->tso_deferred && ((jiffies<<1)>>1) - (tp->tso_deferred>>1) > 1) | |
1175 | goto send_now; | |
908a75c1 | 1176 | |
c1b4a7e6 DM |
1177 | in_flight = tcp_packets_in_flight(tp); |
1178 | ||
1179 | BUG_ON(tcp_skb_pcount(skb) <= 1 || | |
1180 | (tp->snd_cwnd <= in_flight)); | |
1181 | ||
1182 | send_win = (tp->snd_una + tp->snd_wnd) - TCP_SKB_CB(skb)->seq; | |
1183 | ||
1184 | /* From in_flight test above, we know that cwnd > in_flight. */ | |
1185 | cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; | |
1186 | ||
1187 | limit = min(send_win, cong_win); | |
1188 | ||
ba244fe9 DM |
1189 | /* If a full-sized TSO skb can be sent, do it. */ |
1190 | if (limit >= 65536) | |
ae8064ac | 1191 | goto send_now; |
ba244fe9 | 1192 | |
c1b4a7e6 DM |
1193 | if (sysctl_tcp_tso_win_divisor) { |
1194 | u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); | |
1195 | ||
1196 | /* If at least some fraction of a window is available, | |
1197 | * just use it. | |
1198 | */ | |
1199 | chunk /= sysctl_tcp_tso_win_divisor; | |
1200 | if (limit >= chunk) | |
ae8064ac | 1201 | goto send_now; |
c1b4a7e6 DM |
1202 | } else { |
1203 | /* Different approach, try not to defer past a single | |
1204 | * ACK. Receiver should ACK every other full sized | |
1205 | * frame, so if we have space for more than 3 frames | |
1206 | * then send now. | |
1207 | */ | |
1208 | if (limit > tcp_max_burst(tp) * tp->mss_cache) | |
ae8064ac | 1209 | goto send_now; |
c1b4a7e6 DM |
1210 | } |
1211 | ||
1212 | /* Ok, it looks like it is advisable to defer. */ | |
ae8064ac JH |
1213 | tp->tso_deferred = 1 | (jiffies<<1); |
1214 | ||
c1b4a7e6 | 1215 | return 1; |
ae8064ac JH |
1216 | |
1217 | send_now: | |
1218 | tp->tso_deferred = 0; | |
1219 | return 0; | |
c1b4a7e6 DM |
1220 | } |
1221 | ||
5d424d5a JH |
1222 | /* Create a new MTU probe if we are ready. |
1223 | * Returns 0 if we should wait to probe (no cwnd available), | |
1224 | * 1 if a probe was sent, | |
1225 | * -1 otherwise */ | |
1226 | static int tcp_mtu_probe(struct sock *sk) | |
1227 | { | |
1228 | struct tcp_sock *tp = tcp_sk(sk); | |
1229 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1230 | struct sk_buff *skb, *nskb, *next; | |
1231 | int len; | |
1232 | int probe_size; | |
1233 | unsigned int pif; | |
1234 | int copy; | |
1235 | int mss_now; | |
1236 | ||
1237 | /* Not currently probing/verifying, | |
1238 | * not in recovery, | |
1239 | * have enough cwnd, and | |
1240 | * not SACKing (the variable headers throw things off) */ | |
1241 | if (!icsk->icsk_mtup.enabled || | |
1242 | icsk->icsk_mtup.probe_size || | |
1243 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || | |
1244 | tp->snd_cwnd < 11 || | |
1245 | tp->rx_opt.eff_sacks) | |
1246 | return -1; | |
1247 | ||
1248 | /* Very simple search strategy: just double the MSS. */ | |
1249 | mss_now = tcp_current_mss(sk, 0); | |
1250 | probe_size = 2*tp->mss_cache; | |
1251 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { | |
1252 | /* TODO: set timer for probe_converge_event */ | |
1253 | return -1; | |
1254 | } | |
1255 | ||
1256 | /* Have enough data in the send queue to probe? */ | |
1257 | len = 0; | |
fe067e8a | 1258 | if ((skb = tcp_send_head(sk)) == NULL) |
5d424d5a JH |
1259 | return -1; |
1260 | while ((len += skb->len) < probe_size && !tcp_skb_is_last(sk, skb)) | |
fe067e8a | 1261 | skb = tcp_write_queue_next(sk, skb); |
5d424d5a JH |
1262 | if (len < probe_size) |
1263 | return -1; | |
1264 | ||
1265 | /* Receive window check. */ | |
1266 | if (after(TCP_SKB_CB(skb)->seq + probe_size, tp->snd_una + tp->snd_wnd)) { | |
1267 | if (tp->snd_wnd < probe_size) | |
1268 | return -1; | |
1269 | else | |
1270 | return 0; | |
1271 | } | |
1272 | ||
1273 | /* Do we need to wait to drain cwnd? */ | |
1274 | pif = tcp_packets_in_flight(tp); | |
1275 | if (pif + 2 > tp->snd_cwnd) { | |
1276 | /* With no packets in flight, don't stall. */ | |
1277 | if (pif == 0) | |
1278 | return -1; | |
1279 | else | |
1280 | return 0; | |
1281 | } | |
1282 | ||
1283 | /* We're allowed to probe. Build it now. */ | |
1284 | if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) | |
1285 | return -1; | |
1286 | sk_charge_skb(sk, nskb); | |
1287 | ||
fe067e8a DM |
1288 | skb = tcp_send_head(sk); |
1289 | tcp_insert_write_queue_before(nskb, skb, sk); | |
1290 | tcp_advance_send_head(sk, skb); | |
5d424d5a JH |
1291 | |
1292 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; | |
1293 | TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; | |
1294 | TCP_SKB_CB(nskb)->flags = TCPCB_FLAG_ACK; | |
1295 | TCP_SKB_CB(nskb)->sacked = 0; | |
1296 | nskb->csum = 0; | |
84fa7933 | 1297 | nskb->ip_summed = skb->ip_summed; |
5d424d5a JH |
1298 | |
1299 | len = 0; | |
1300 | while (len < probe_size) { | |
fe067e8a | 1301 | next = tcp_write_queue_next(sk, skb); |
5d424d5a JH |
1302 | |
1303 | copy = min_t(int, skb->len, probe_size - len); | |
1304 | if (nskb->ip_summed) | |
1305 | skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); | |
1306 | else | |
1307 | nskb->csum = skb_copy_and_csum_bits(skb, 0, | |
e905a9ed | 1308 | skb_put(nskb, copy), copy, nskb->csum); |
5d424d5a JH |
1309 | |
1310 | if (skb->len <= copy) { | |
1311 | /* We've eaten all the data from this skb. | |
1312 | * Throw it away. */ | |
1313 | TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags; | |
fe067e8a | 1314 | tcp_unlink_write_queue(skb, sk); |
5d424d5a JH |
1315 | sk_stream_free_skb(sk, skb); |
1316 | } else { | |
1317 | TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags & | |
e905a9ed | 1318 | ~(TCPCB_FLAG_FIN|TCPCB_FLAG_PSH); |
5d424d5a JH |
1319 | if (!skb_shinfo(skb)->nr_frags) { |
1320 | skb_pull(skb, copy); | |
84fa7933 | 1321 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
5d424d5a JH |
1322 | skb->csum = csum_partial(skb->data, skb->len, 0); |
1323 | } else { | |
1324 | __pskb_trim_head(skb, copy); | |
1325 | tcp_set_skb_tso_segs(sk, skb, mss_now); | |
1326 | } | |
1327 | TCP_SKB_CB(skb)->seq += copy; | |
1328 | } | |
1329 | ||
1330 | len += copy; | |
1331 | skb = next; | |
1332 | } | |
1333 | tcp_init_tso_segs(sk, nskb, nskb->len); | |
1334 | ||
1335 | /* We're ready to send. If this fails, the probe will | |
1336 | * be resegmented into mss-sized pieces by tcp_write_xmit(). */ | |
1337 | TCP_SKB_CB(nskb)->when = tcp_time_stamp; | |
1338 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { | |
1339 | /* Decrement cwnd here because we are sending | |
1340 | * effectively two packets. */ | |
1341 | tp->snd_cwnd--; | |
9e412ba7 | 1342 | update_send_head(sk, nskb); |
5d424d5a JH |
1343 | |
1344 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); | |
0e7b1368 JH |
1345 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; |
1346 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; | |
5d424d5a JH |
1347 | |
1348 | return 1; | |
1349 | } | |
1350 | ||
1351 | return -1; | |
1352 | } | |
1353 | ||
1354 | ||
1da177e4 LT |
1355 | /* This routine writes packets to the network. It advances the |
1356 | * send_head. This happens as incoming acks open up the remote | |
1357 | * window for us. | |
1358 | * | |
1359 | * Returns 1, if no segments are in flight and we have queued segments, but | |
1360 | * cannot send anything now because of SWS or another problem. | |
1361 | */ | |
a2e2a59c | 1362 | static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle) |
1da177e4 LT |
1363 | { |
1364 | struct tcp_sock *tp = tcp_sk(sk); | |
92df7b51 | 1365 | struct sk_buff *skb; |
c1b4a7e6 DM |
1366 | unsigned int tso_segs, sent_pkts; |
1367 | int cwnd_quota; | |
5d424d5a | 1368 | int result; |
1da177e4 LT |
1369 | |
1370 | /* If we are closed, the bytes will have to remain here. | |
1371 | * In time closedown will finish, we empty the write queue and all | |
1372 | * will be happy. | |
1373 | */ | |
92df7b51 DM |
1374 | if (unlikely(sk->sk_state == TCP_CLOSE)) |
1375 | return 0; | |
1da177e4 | 1376 | |
92df7b51 | 1377 | sent_pkts = 0; |
5d424d5a JH |
1378 | |
1379 | /* Do MTU probing. */ | |
1380 | if ((result = tcp_mtu_probe(sk)) == 0) { | |
1381 | return 0; | |
1382 | } else if (result > 0) { | |
1383 | sent_pkts = 1; | |
1384 | } | |
1385 | ||
fe067e8a | 1386 | while ((skb = tcp_send_head(sk))) { |
c8ac3774 HX |
1387 | unsigned int limit; |
1388 | ||
b68e9f85 | 1389 | tso_segs = tcp_init_tso_segs(sk, skb, mss_now); |
c1b4a7e6 | 1390 | BUG_ON(!tso_segs); |
aa93466b | 1391 | |
b68e9f85 HX |
1392 | cwnd_quota = tcp_cwnd_test(tp, skb); |
1393 | if (!cwnd_quota) | |
1394 | break; | |
1395 | ||
1396 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) | |
1397 | break; | |
1398 | ||
c1b4a7e6 DM |
1399 | if (tso_segs == 1) { |
1400 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, | |
1401 | (tcp_skb_is_last(sk, skb) ? | |
1402 | nonagle : TCP_NAGLE_PUSH)))) | |
1403 | break; | |
1404 | } else { | |
9e412ba7 | 1405 | if (tcp_tso_should_defer(sk, skb)) |
c1b4a7e6 DM |
1406 | break; |
1407 | } | |
aa93466b | 1408 | |
c8ac3774 | 1409 | limit = mss_now; |
c1b4a7e6 | 1410 | if (tso_segs > 1) { |
c8ac3774 HX |
1411 | limit = tcp_window_allows(tp, skb, |
1412 | mss_now, cwnd_quota); | |
c1b4a7e6 DM |
1413 | |
1414 | if (skb->len < limit) { | |
1415 | unsigned int trim = skb->len % mss_now; | |
aa93466b | 1416 | |
c1b4a7e6 DM |
1417 | if (trim) |
1418 | limit = skb->len - trim; | |
1419 | } | |
92df7b51 | 1420 | } |
1da177e4 | 1421 | |
c8ac3774 HX |
1422 | if (skb->len > limit && |
1423 | unlikely(tso_fragment(sk, skb, limit, mss_now))) | |
1424 | break; | |
1425 | ||
92df7b51 | 1426 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
c1b4a7e6 | 1427 | |
dfb4b9dc | 1428 | if (unlikely(tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC))) |
92df7b51 | 1429 | break; |
1da177e4 | 1430 | |
92df7b51 DM |
1431 | /* Advance the send_head. This one is sent out. |
1432 | * This call will increment packets_out. | |
1433 | */ | |
9e412ba7 | 1434 | update_send_head(sk, skb); |
1da177e4 | 1435 | |
92df7b51 | 1436 | tcp_minshall_update(tp, mss_now, skb); |
aa93466b | 1437 | sent_pkts++; |
92df7b51 | 1438 | } |
1da177e4 | 1439 | |
aa93466b | 1440 | if (likely(sent_pkts)) { |
9e412ba7 | 1441 | tcp_cwnd_validate(sk); |
92df7b51 | 1442 | return 0; |
1da177e4 | 1443 | } |
fe067e8a | 1444 | return !tp->packets_out && tcp_send_head(sk); |
1da177e4 LT |
1445 | } |
1446 | ||
a762a980 DM |
1447 | /* Push out any pending frames which were held back due to |
1448 | * TCP_CORK or attempt at coalescing tiny packets. | |
1449 | * The socket must be locked by the caller. | |
1450 | */ | |
9e412ba7 IJ |
1451 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, |
1452 | int nonagle) | |
a762a980 | 1453 | { |
fe067e8a | 1454 | struct sk_buff *skb = tcp_send_head(sk); |
a762a980 DM |
1455 | |
1456 | if (skb) { | |
55c97f3e | 1457 | if (tcp_write_xmit(sk, cur_mss, nonagle)) |
9e412ba7 | 1458 | tcp_check_probe_timer(sk); |
a762a980 DM |
1459 | } |
1460 | } | |
1461 | ||
c1b4a7e6 DM |
1462 | /* Send _single_ skb sitting at the send head. This function requires |
1463 | * true push pending frames to setup probe timer etc. | |
1464 | */ | |
1465 | void tcp_push_one(struct sock *sk, unsigned int mss_now) | |
1466 | { | |
1467 | struct tcp_sock *tp = tcp_sk(sk); | |
fe067e8a | 1468 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 DM |
1469 | unsigned int tso_segs, cwnd_quota; |
1470 | ||
1471 | BUG_ON(!skb || skb->len < mss_now); | |
1472 | ||
846998ae | 1473 | tso_segs = tcp_init_tso_segs(sk, skb, mss_now); |
c1b4a7e6 DM |
1474 | cwnd_quota = tcp_snd_test(sk, skb, mss_now, TCP_NAGLE_PUSH); |
1475 | ||
1476 | if (likely(cwnd_quota)) { | |
c8ac3774 HX |
1477 | unsigned int limit; |
1478 | ||
c1b4a7e6 DM |
1479 | BUG_ON(!tso_segs); |
1480 | ||
c8ac3774 | 1481 | limit = mss_now; |
c1b4a7e6 | 1482 | if (tso_segs > 1) { |
c8ac3774 HX |
1483 | limit = tcp_window_allows(tp, skb, |
1484 | mss_now, cwnd_quota); | |
c1b4a7e6 DM |
1485 | |
1486 | if (skb->len < limit) { | |
1487 | unsigned int trim = skb->len % mss_now; | |
1488 | ||
1489 | if (trim) | |
1490 | limit = skb->len - trim; | |
1491 | } | |
c1b4a7e6 DM |
1492 | } |
1493 | ||
c8ac3774 HX |
1494 | if (skb->len > limit && |
1495 | unlikely(tso_fragment(sk, skb, limit, mss_now))) | |
1496 | return; | |
1497 | ||
c1b4a7e6 DM |
1498 | /* Send it out now. */ |
1499 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
1500 | ||
dfb4b9dc | 1501 | if (likely(!tcp_transmit_skb(sk, skb, 1, sk->sk_allocation))) { |
9e412ba7 IJ |
1502 | update_send_head(sk, skb); |
1503 | tcp_cwnd_validate(sk); | |
c1b4a7e6 DM |
1504 | return; |
1505 | } | |
1506 | } | |
1507 | } | |
1508 | ||
1da177e4 LT |
1509 | /* This function returns the amount that we can raise the |
1510 | * usable window based on the following constraints | |
e905a9ed | 1511 | * |
1da177e4 LT |
1512 | * 1. The window can never be shrunk once it is offered (RFC 793) |
1513 | * 2. We limit memory per socket | |
1514 | * | |
1515 | * RFC 1122: | |
1516 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep | |
1517 | * RECV.NEXT + RCV.WIN fixed until: | |
1518 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" | |
1519 | * | |
1520 | * i.e. don't raise the right edge of the window until you can raise | |
1521 | * it at least MSS bytes. | |
1522 | * | |
1523 | * Unfortunately, the recommended algorithm breaks header prediction, | |
1524 | * since header prediction assumes th->window stays fixed. | |
1525 | * | |
1526 | * Strictly speaking, keeping th->window fixed violates the receiver | |
1527 | * side SWS prevention criteria. The problem is that under this rule | |
1528 | * a stream of single byte packets will cause the right side of the | |
1529 | * window to always advance by a single byte. | |
e905a9ed | 1530 | * |
1da177e4 LT |
1531 | * Of course, if the sender implements sender side SWS prevention |
1532 | * then this will not be a problem. | |
e905a9ed | 1533 | * |
1da177e4 | 1534 | * BSD seems to make the following compromise: |
e905a9ed | 1535 | * |
1da177e4 LT |
1536 | * If the free space is less than the 1/4 of the maximum |
1537 | * space available and the free space is less than 1/2 mss, | |
1538 | * then set the window to 0. | |
1539 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] | |
1540 | * Otherwise, just prevent the window from shrinking | |
1541 | * and from being larger than the largest representable value. | |
1542 | * | |
1543 | * This prevents incremental opening of the window in the regime | |
1544 | * where TCP is limited by the speed of the reader side taking | |
1545 | * data out of the TCP receive queue. It does nothing about | |
1546 | * those cases where the window is constrained on the sender side | |
1547 | * because the pipeline is full. | |
1548 | * | |
1549 | * BSD also seems to "accidentally" limit itself to windows that are a | |
1550 | * multiple of MSS, at least until the free space gets quite small. | |
1551 | * This would appear to be a side effect of the mbuf implementation. | |
1552 | * Combining these two algorithms results in the observed behavior | |
1553 | * of having a fixed window size at almost all times. | |
1554 | * | |
1555 | * Below we obtain similar behavior by forcing the offered window to | |
1556 | * a multiple of the mss when it is feasible to do so. | |
1557 | * | |
1558 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. | |
1559 | * Regular options like TIMESTAMP are taken into account. | |
1560 | */ | |
1561 | u32 __tcp_select_window(struct sock *sk) | |
1562 | { | |
463c84b9 | 1563 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 1564 | struct tcp_sock *tp = tcp_sk(sk); |
caa20d9a | 1565 | /* MSS for the peer's data. Previous versions used mss_clamp |
1da177e4 LT |
1566 | * here. I don't know if the value based on our guesses |
1567 | * of peer's MSS is better for the performance. It's more correct | |
1568 | * but may be worse for the performance because of rcv_mss | |
1569 | * fluctuations. --SAW 1998/11/1 | |
1570 | */ | |
463c84b9 | 1571 | int mss = icsk->icsk_ack.rcv_mss; |
1da177e4 LT |
1572 | int free_space = tcp_space(sk); |
1573 | int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); | |
1574 | int window; | |
1575 | ||
1576 | if (mss > full_space) | |
e905a9ed | 1577 | mss = full_space; |
1da177e4 LT |
1578 | |
1579 | if (free_space < full_space/2) { | |
463c84b9 | 1580 | icsk->icsk_ack.quick = 0; |
1da177e4 LT |
1581 | |
1582 | if (tcp_memory_pressure) | |
1583 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, 4U*tp->advmss); | |
1584 | ||
1585 | if (free_space < mss) | |
1586 | return 0; | |
1587 | } | |
1588 | ||
1589 | if (free_space > tp->rcv_ssthresh) | |
1590 | free_space = tp->rcv_ssthresh; | |
1591 | ||
1592 | /* Don't do rounding if we are using window scaling, since the | |
1593 | * scaled window will not line up with the MSS boundary anyway. | |
1594 | */ | |
1595 | window = tp->rcv_wnd; | |
1596 | if (tp->rx_opt.rcv_wscale) { | |
1597 | window = free_space; | |
1598 | ||
1599 | /* Advertise enough space so that it won't get scaled away. | |
1600 | * Import case: prevent zero window announcement if | |
1601 | * 1<<rcv_wscale > mss. | |
1602 | */ | |
1603 | if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) | |
1604 | window = (((window >> tp->rx_opt.rcv_wscale) + 1) | |
1605 | << tp->rx_opt.rcv_wscale); | |
1606 | } else { | |
1607 | /* Get the largest window that is a nice multiple of mss. | |
1608 | * Window clamp already applied above. | |
1609 | * If our current window offering is within 1 mss of the | |
1610 | * free space we just keep it. This prevents the divide | |
1611 | * and multiply from happening most of the time. | |
1612 | * We also don't do any window rounding when the free space | |
1613 | * is too small. | |
1614 | */ | |
1615 | if (window <= free_space - mss || window > free_space) | |
1616 | window = (free_space/mss)*mss; | |
84565070 | 1617 | else if (mss == full_space && |
9c681b43 | 1618 | free_space > window + full_space/2) |
84565070 | 1619 | window = free_space; |
1da177e4 LT |
1620 | } |
1621 | ||
1622 | return window; | |
1623 | } | |
1624 | ||
1625 | /* Attempt to collapse two adjacent SKB's during retransmission. */ | |
1626 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *skb, int mss_now) | |
1627 | { | |
1628 | struct tcp_sock *tp = tcp_sk(sk); | |
fe067e8a | 1629 | struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); |
1da177e4 LT |
1630 | |
1631 | /* The first test we must make is that neither of these two | |
1632 | * SKB's are still referenced by someone else. | |
1633 | */ | |
1634 | if (!skb_cloned(skb) && !skb_cloned(next_skb)) { | |
1635 | int skb_size = skb->len, next_skb_size = next_skb->len; | |
1636 | u16 flags = TCP_SKB_CB(skb)->flags; | |
1637 | ||
1638 | /* Also punt if next skb has been SACK'd. */ | |
2de979bd | 1639 | if (TCP_SKB_CB(next_skb)->sacked & TCPCB_SACKED_ACKED) |
1da177e4 LT |
1640 | return; |
1641 | ||
1642 | /* Next skb is out of window. */ | |
1643 | if (after(TCP_SKB_CB(next_skb)->end_seq, tp->snd_una+tp->snd_wnd)) | |
1644 | return; | |
1645 | ||
1646 | /* Punt if not enough space exists in the first SKB for | |
1647 | * the data in the second, or the total combined payload | |
1648 | * would exceed the MSS. | |
1649 | */ | |
1650 | if ((next_skb_size > skb_tailroom(skb)) || | |
1651 | ((skb_size + next_skb_size) > mss_now)) | |
1652 | return; | |
1653 | ||
1654 | BUG_ON(tcp_skb_pcount(skb) != 1 || | |
1655 | tcp_skb_pcount(next_skb) != 1); | |
1656 | ||
6a438bbe SH |
1657 | /* changing transmit queue under us so clear hints */ |
1658 | clear_all_retrans_hints(tp); | |
1659 | ||
1660 | /* Ok. We will be able to collapse the packet. */ | |
fe067e8a | 1661 | tcp_unlink_write_queue(next_skb, sk); |
1da177e4 | 1662 | |
1a4e2d09 ACM |
1663 | skb_copy_from_linear_data(next_skb, |
1664 | skb_put(skb, next_skb_size), | |
1665 | next_skb_size); | |
1da177e4 | 1666 | |
52d570aa JP |
1667 | if (next_skb->ip_summed == CHECKSUM_PARTIAL) |
1668 | skb->ip_summed = CHECKSUM_PARTIAL; | |
1da177e4 | 1669 | |
84fa7933 | 1670 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
1da177e4 LT |
1671 | skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); |
1672 | ||
1673 | /* Update sequence range on original skb. */ | |
1674 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; | |
1675 | ||
1676 | /* Merge over control information. */ | |
1677 | flags |= TCP_SKB_CB(next_skb)->flags; /* This moves PSH/FIN etc. over */ | |
1678 | TCP_SKB_CB(skb)->flags = flags; | |
1679 | ||
1680 | /* All done, get rid of second SKB and account for it so | |
1681 | * packet counting does not break. | |
1682 | */ | |
1683 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked&(TCPCB_EVER_RETRANS|TCPCB_AT_TAIL); | |
1684 | if (TCP_SKB_CB(next_skb)->sacked&TCPCB_SACKED_RETRANS) | |
1685 | tp->retrans_out -= tcp_skb_pcount(next_skb); | |
1686 | if (TCP_SKB_CB(next_skb)->sacked&TCPCB_LOST) { | |
1687 | tp->lost_out -= tcp_skb_pcount(next_skb); | |
1688 | tp->left_out -= tcp_skb_pcount(next_skb); | |
1689 | } | |
1690 | /* Reno case is special. Sigh... */ | |
1691 | if (!tp->rx_opt.sack_ok && tp->sacked_out) { | |
1692 | tcp_dec_pcount_approx(&tp->sacked_out, next_skb); | |
1693 | tp->left_out -= tcp_skb_pcount(next_skb); | |
1694 | } | |
1695 | ||
1696 | /* Not quite right: it can be > snd.fack, but | |
1697 | * it is better to underestimate fackets. | |
1698 | */ | |
1699 | tcp_dec_pcount_approx(&tp->fackets_out, next_skb); | |
1700 | tcp_packets_out_dec(tp, next_skb); | |
1701 | sk_stream_free_skb(sk, next_skb); | |
1702 | } | |
1703 | } | |
1704 | ||
1705 | /* Do a simple retransmit without using the backoff mechanisms in | |
e905a9ed | 1706 | * tcp_timer. This is used for path mtu discovery. |
1da177e4 | 1707 | * The socket is already locked here. |
e905a9ed | 1708 | */ |
1da177e4 LT |
1709 | void tcp_simple_retransmit(struct sock *sk) |
1710 | { | |
6687e988 | 1711 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
1712 | struct tcp_sock *tp = tcp_sk(sk); |
1713 | struct sk_buff *skb; | |
1714 | unsigned int mss = tcp_current_mss(sk, 0); | |
1715 | int lost = 0; | |
1716 | ||
fe067e8a DM |
1717 | tcp_for_write_queue(skb, sk) { |
1718 | if (skb == tcp_send_head(sk)) | |
1719 | break; | |
e905a9ed | 1720 | if (skb->len > mss && |
1da177e4 LT |
1721 | !(TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_ACKED)) { |
1722 | if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) { | |
1723 | TCP_SKB_CB(skb)->sacked &= ~TCPCB_SACKED_RETRANS; | |
1724 | tp->retrans_out -= tcp_skb_pcount(skb); | |
1725 | } | |
1726 | if (!(TCP_SKB_CB(skb)->sacked&TCPCB_LOST)) { | |
1727 | TCP_SKB_CB(skb)->sacked |= TCPCB_LOST; | |
1728 | tp->lost_out += tcp_skb_pcount(skb); | |
1729 | lost = 1; | |
1730 | } | |
1731 | } | |
1732 | } | |
1733 | ||
6a438bbe SH |
1734 | clear_all_retrans_hints(tp); |
1735 | ||
1da177e4 LT |
1736 | if (!lost) |
1737 | return; | |
1738 | ||
1739 | tcp_sync_left_out(tp); | |
1740 | ||
e905a9ed | 1741 | /* Don't muck with the congestion window here. |
1da177e4 LT |
1742 | * Reason is that we do not increase amount of _data_ |
1743 | * in network, but units changed and effective | |
1744 | * cwnd/ssthresh really reduced now. | |
1745 | */ | |
6687e988 | 1746 | if (icsk->icsk_ca_state != TCP_CA_Loss) { |
1da177e4 | 1747 | tp->high_seq = tp->snd_nxt; |
6687e988 | 1748 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
1da177e4 LT |
1749 | tp->prior_ssthresh = 0; |
1750 | tp->undo_marker = 0; | |
6687e988 | 1751 | tcp_set_ca_state(sk, TCP_CA_Loss); |
1da177e4 LT |
1752 | } |
1753 | tcp_xmit_retransmit_queue(sk); | |
1754 | } | |
1755 | ||
1756 | /* This retransmits one SKB. Policy decisions and retransmit queue | |
1757 | * state updates are done by the caller. Returns non-zero if an | |
1758 | * error occurred which prevented the send. | |
1759 | */ | |
1760 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) | |
1761 | { | |
1762 | struct tcp_sock *tp = tcp_sk(sk); | |
5d424d5a | 1763 | struct inet_connection_sock *icsk = inet_csk(sk); |
e905a9ed | 1764 | unsigned int cur_mss = tcp_current_mss(sk, 0); |
1da177e4 LT |
1765 | int err; |
1766 | ||
5d424d5a JH |
1767 | /* Inconslusive MTU probe */ |
1768 | if (icsk->icsk_mtup.probe_size) { | |
1769 | icsk->icsk_mtup.probe_size = 0; | |
1770 | } | |
1771 | ||
1da177e4 | 1772 | /* Do not sent more than we queued. 1/4 is reserved for possible |
caa20d9a | 1773 | * copying overhead: fragmentation, tunneling, mangling etc. |
1da177e4 LT |
1774 | */ |
1775 | if (atomic_read(&sk->sk_wmem_alloc) > | |
1776 | min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) | |
1777 | return -EAGAIN; | |
1778 | ||
1779 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { | |
1780 | if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
1781 | BUG(); | |
1da177e4 LT |
1782 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
1783 | return -ENOMEM; | |
1784 | } | |
1785 | ||
1786 | /* If receiver has shrunk his window, and skb is out of | |
1787 | * new window, do not retransmit it. The exception is the | |
1788 | * case, when window is shrunk to zero. In this case | |
1789 | * our retransmit serves as a zero window probe. | |
1790 | */ | |
1791 | if (!before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd) | |
1792 | && TCP_SKB_CB(skb)->seq != tp->snd_una) | |
1793 | return -EAGAIN; | |
1794 | ||
1795 | if (skb->len > cur_mss) { | |
846998ae | 1796 | if (tcp_fragment(sk, skb, cur_mss, cur_mss)) |
1da177e4 | 1797 | return -ENOMEM; /* We'll try again later. */ |
1da177e4 LT |
1798 | } |
1799 | ||
1800 | /* Collapse two adjacent packets if worthwhile and we can. */ | |
2de979bd SH |
1801 | if (!(TCP_SKB_CB(skb)->flags & TCPCB_FLAG_SYN) && |
1802 | (skb->len < (cur_mss >> 1)) && | |
1803 | (tcp_write_queue_next(sk, skb) != tcp_send_head(sk)) && | |
1804 | (!tcp_skb_is_last(sk, skb)) && | |
1805 | (skb_shinfo(skb)->nr_frags == 0 && skb_shinfo(tcp_write_queue_next(sk, skb))->nr_frags == 0) && | |
1806 | (tcp_skb_pcount(skb) == 1 && tcp_skb_pcount(tcp_write_queue_next(sk, skb)) == 1) && | |
1807 | (sysctl_tcp_retrans_collapse != 0)) | |
1da177e4 LT |
1808 | tcp_retrans_try_collapse(sk, skb, cur_mss); |
1809 | ||
8292a17a | 1810 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) |
1da177e4 LT |
1811 | return -EHOSTUNREACH; /* Routing failure or similar. */ |
1812 | ||
1813 | /* Some Solaris stacks overoptimize and ignore the FIN on a | |
1814 | * retransmit when old data is attached. So strip it off | |
1815 | * since it is cheap to do so and saves bytes on the network. | |
1816 | */ | |
2de979bd SH |
1817 | if (skb->len > 0 && |
1818 | (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && | |
1819 | tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { | |
1da177e4 LT |
1820 | if (!pskb_trim(skb, 0)) { |
1821 | TCP_SKB_CB(skb)->seq = TCP_SKB_CB(skb)->end_seq - 1; | |
7967168c HX |
1822 | skb_shinfo(skb)->gso_segs = 1; |
1823 | skb_shinfo(skb)->gso_size = 0; | |
1824 | skb_shinfo(skb)->gso_type = 0; | |
1da177e4 LT |
1825 | skb->ip_summed = CHECKSUM_NONE; |
1826 | skb->csum = 0; | |
1827 | } | |
1828 | } | |
1829 | ||
1830 | /* Make a copy, if the first transmission SKB clone we made | |
1831 | * is still in somebody's hands, else make a clone. | |
1832 | */ | |
1833 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
1da177e4 | 1834 | |
dfb4b9dc | 1835 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
1da177e4 LT |
1836 | |
1837 | if (err == 0) { | |
1838 | /* Update global TCP statistics. */ | |
1839 | TCP_INC_STATS(TCP_MIB_RETRANSSEGS); | |
1840 | ||
1841 | tp->total_retrans++; | |
1842 | ||
1843 | #if FASTRETRANS_DEBUG > 0 | |
1844 | if (TCP_SKB_CB(skb)->sacked&TCPCB_SACKED_RETRANS) { | |
1845 | if (net_ratelimit()) | |
1846 | printk(KERN_DEBUG "retrans_out leaked.\n"); | |
1847 | } | |
1848 | #endif | |
1849 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; | |
1850 | tp->retrans_out += tcp_skb_pcount(skb); | |
1851 | ||
1852 | /* Save stamp of the first retransmit. */ | |
1853 | if (!tp->retrans_stamp) | |
1854 | tp->retrans_stamp = TCP_SKB_CB(skb)->when; | |
1855 | ||
1856 | tp->undo_retrans++; | |
1857 | ||
1858 | /* snd_nxt is stored to detect loss of retransmitted segment, | |
1859 | * see tcp_input.c tcp_sacktag_write_queue(). | |
1860 | */ | |
1861 | TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; | |
1862 | } | |
1863 | return err; | |
1864 | } | |
1865 | ||
1866 | /* This gets called after a retransmit timeout, and the initially | |
1867 | * retransmitted data is acknowledged. It tries to continue | |
1868 | * resending the rest of the retransmit queue, until either | |
1869 | * we've sent it all or the congestion window limit is reached. | |
1870 | * If doing SACK, the first ACK which comes back for a timeout | |
1871 | * based retransmit packet might feed us FACK information again. | |
1872 | * If so, we use it to avoid unnecessarily retransmissions. | |
1873 | */ | |
1874 | void tcp_xmit_retransmit_queue(struct sock *sk) | |
1875 | { | |
6687e988 | 1876 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
1877 | struct tcp_sock *tp = tcp_sk(sk); |
1878 | struct sk_buff *skb; | |
6a438bbe SH |
1879 | int packet_cnt; |
1880 | ||
1881 | if (tp->retransmit_skb_hint) { | |
1882 | skb = tp->retransmit_skb_hint; | |
1883 | packet_cnt = tp->retransmit_cnt_hint; | |
1884 | }else{ | |
fe067e8a | 1885 | skb = tcp_write_queue_head(sk); |
6a438bbe SH |
1886 | packet_cnt = 0; |
1887 | } | |
1da177e4 LT |
1888 | |
1889 | /* First pass: retransmit lost packets. */ | |
6a438bbe | 1890 | if (tp->lost_out) { |
fe067e8a | 1891 | tcp_for_write_queue_from(skb, sk) { |
1da177e4 LT |
1892 | __u8 sacked = TCP_SKB_CB(skb)->sacked; |
1893 | ||
fe067e8a DM |
1894 | if (skb == tcp_send_head(sk)) |
1895 | break; | |
6a438bbe SH |
1896 | /* we could do better than to assign each time */ |
1897 | tp->retransmit_skb_hint = skb; | |
1898 | tp->retransmit_cnt_hint = packet_cnt; | |
1899 | ||
1da177e4 LT |
1900 | /* Assume this retransmit will generate |
1901 | * only one packet for congestion window | |
1902 | * calculation purposes. This works because | |
1903 | * tcp_retransmit_skb() will chop up the | |
1904 | * packet to be MSS sized and all the | |
1905 | * packet counting works out. | |
1906 | */ | |
1907 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) | |
1908 | return; | |
1909 | ||
6a438bbe | 1910 | if (sacked & TCPCB_LOST) { |
1da177e4 | 1911 | if (!(sacked&(TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))) { |
6a438bbe SH |
1912 | if (tcp_retransmit_skb(sk, skb)) { |
1913 | tp->retransmit_skb_hint = NULL; | |
1da177e4 | 1914 | return; |
6a438bbe | 1915 | } |
6687e988 | 1916 | if (icsk->icsk_ca_state != TCP_CA_Loss) |
1da177e4 LT |
1917 | NET_INC_STATS_BH(LINUX_MIB_TCPFASTRETRANS); |
1918 | else | |
1919 | NET_INC_STATS_BH(LINUX_MIB_TCPSLOWSTARTRETRANS); | |
1920 | ||
fe067e8a | 1921 | if (skb == tcp_write_queue_head(sk)) |
463c84b9 | 1922 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
3f421baa ACM |
1923 | inet_csk(sk)->icsk_rto, |
1924 | TCP_RTO_MAX); | |
1da177e4 LT |
1925 | } |
1926 | ||
6a438bbe SH |
1927 | packet_cnt += tcp_skb_pcount(skb); |
1928 | if (packet_cnt >= tp->lost_out) | |
1da177e4 LT |
1929 | break; |
1930 | } | |
1931 | } | |
1932 | } | |
1933 | ||
1934 | /* OK, demanded retransmission is finished. */ | |
1935 | ||
1936 | /* Forward retransmissions are possible only during Recovery. */ | |
6687e988 | 1937 | if (icsk->icsk_ca_state != TCP_CA_Recovery) |
1da177e4 LT |
1938 | return; |
1939 | ||
1940 | /* No forward retransmissions in Reno are possible. */ | |
1941 | if (!tp->rx_opt.sack_ok) | |
1942 | return; | |
1943 | ||
1944 | /* Yeah, we have to make difficult choice between forward transmission | |
1945 | * and retransmission... Both ways have their merits... | |
1946 | * | |
1947 | * For now we do not retransmit anything, while we have some new | |
539d243f IJ |
1948 | * segments to send. In the other cases, follow rule 3 for |
1949 | * NextSeg() specified in RFC3517. | |
1da177e4 LT |
1950 | */ |
1951 | ||
9e412ba7 | 1952 | if (tcp_may_send_now(sk)) |
1da177e4 LT |
1953 | return; |
1954 | ||
539d243f IJ |
1955 | /* If nothing is SACKed, highest_sack in the loop won't be valid */ |
1956 | if (!tp->sacked_out) | |
1957 | return; | |
1958 | ||
1959 | if (tp->forward_skb_hint) | |
6a438bbe | 1960 | skb = tp->forward_skb_hint; |
539d243f | 1961 | else |
fe067e8a | 1962 | skb = tcp_write_queue_head(sk); |
6a438bbe | 1963 | |
fe067e8a DM |
1964 | tcp_for_write_queue_from(skb, sk) { |
1965 | if (skb == tcp_send_head(sk)) | |
1966 | break; | |
6a438bbe | 1967 | tp->forward_skb_hint = skb; |
1da177e4 | 1968 | |
539d243f | 1969 | if (after(TCP_SKB_CB(skb)->seq, tp->highest_sack)) |
1da177e4 LT |
1970 | break; |
1971 | ||
1972 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) | |
1973 | break; | |
1974 | ||
1975 | if (TCP_SKB_CB(skb)->sacked & TCPCB_TAGBITS) | |
1976 | continue; | |
1977 | ||
1978 | /* Ok, retransmit it. */ | |
6a438bbe SH |
1979 | if (tcp_retransmit_skb(sk, skb)) { |
1980 | tp->forward_skb_hint = NULL; | |
1da177e4 | 1981 | break; |
6a438bbe | 1982 | } |
1da177e4 | 1983 | |
fe067e8a | 1984 | if (skb == tcp_write_queue_head(sk)) |
3f421baa ACM |
1985 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
1986 | inet_csk(sk)->icsk_rto, | |
1987 | TCP_RTO_MAX); | |
1da177e4 LT |
1988 | |
1989 | NET_INC_STATS_BH(LINUX_MIB_TCPFORWARDRETRANS); | |
1990 | } | |
1991 | } | |
1992 | ||
1993 | ||
1994 | /* Send a fin. The caller locks the socket for us. This cannot be | |
1995 | * allowed to fail queueing a FIN frame under any circumstances. | |
1996 | */ | |
1997 | void tcp_send_fin(struct sock *sk) | |
1998 | { | |
e905a9ed | 1999 | struct tcp_sock *tp = tcp_sk(sk); |
fe067e8a | 2000 | struct sk_buff *skb = tcp_write_queue_tail(sk); |
1da177e4 | 2001 | int mss_now; |
e905a9ed | 2002 | |
1da177e4 LT |
2003 | /* Optimization, tack on the FIN if we have a queue of |
2004 | * unsent frames. But be careful about outgoing SACKS | |
2005 | * and IP options. | |
2006 | */ | |
2007 | mss_now = tcp_current_mss(sk, 1); | |
2008 | ||
fe067e8a | 2009 | if (tcp_send_head(sk) != NULL) { |
1da177e4 LT |
2010 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_FIN; |
2011 | TCP_SKB_CB(skb)->end_seq++; | |
2012 | tp->write_seq++; | |
2013 | } else { | |
2014 | /* Socket is locked, keep trying until memory is available. */ | |
2015 | for (;;) { | |
d179cd12 | 2016 | skb = alloc_skb_fclone(MAX_TCP_HEADER, GFP_KERNEL); |
1da177e4 LT |
2017 | if (skb) |
2018 | break; | |
2019 | yield(); | |
2020 | } | |
2021 | ||
2022 | /* Reserve space for headers and prepare control bits. */ | |
2023 | skb_reserve(skb, MAX_TCP_HEADER); | |
2024 | skb->csum = 0; | |
2025 | TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_FIN); | |
2026 | TCP_SKB_CB(skb)->sacked = 0; | |
7967168c HX |
2027 | skb_shinfo(skb)->gso_segs = 1; |
2028 | skb_shinfo(skb)->gso_size = 0; | |
2029 | skb_shinfo(skb)->gso_type = 0; | |
1da177e4 LT |
2030 | |
2031 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ | |
2032 | TCP_SKB_CB(skb)->seq = tp->write_seq; | |
2033 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1; | |
2034 | tcp_queue_skb(sk, skb); | |
2035 | } | |
9e412ba7 | 2036 | __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); |
1da177e4 LT |
2037 | } |
2038 | ||
2039 | /* We get here when a process closes a file descriptor (either due to | |
2040 | * an explicit close() or as a byproduct of exit()'ing) and there | |
2041 | * was unread data in the receive queue. This behavior is recommended | |
65bb723c | 2042 | * by RFC 2525, section 2.17. -DaveM |
1da177e4 | 2043 | */ |
dd0fc66f | 2044 | void tcp_send_active_reset(struct sock *sk, gfp_t priority) |
1da177e4 | 2045 | { |
1da177e4 LT |
2046 | struct sk_buff *skb; |
2047 | ||
2048 | /* NOTE: No TCP options attached and we never retransmit this. */ | |
2049 | skb = alloc_skb(MAX_TCP_HEADER, priority); | |
2050 | if (!skb) { | |
2051 | NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); | |
2052 | return; | |
2053 | } | |
2054 | ||
2055 | /* Reserve space for headers and prepare control bits. */ | |
2056 | skb_reserve(skb, MAX_TCP_HEADER); | |
2057 | skb->csum = 0; | |
2058 | TCP_SKB_CB(skb)->flags = (TCPCB_FLAG_ACK | TCPCB_FLAG_RST); | |
2059 | TCP_SKB_CB(skb)->sacked = 0; | |
7967168c HX |
2060 | skb_shinfo(skb)->gso_segs = 1; |
2061 | skb_shinfo(skb)->gso_size = 0; | |
2062 | skb_shinfo(skb)->gso_type = 0; | |
1da177e4 LT |
2063 | |
2064 | /* Send it off. */ | |
9e412ba7 | 2065 | TCP_SKB_CB(skb)->seq = tcp_acceptable_seq(sk); |
1da177e4 LT |
2066 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq; |
2067 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
dfb4b9dc | 2068 | if (tcp_transmit_skb(sk, skb, 0, priority)) |
1da177e4 LT |
2069 | NET_INC_STATS(LINUX_MIB_TCPABORTFAILED); |
2070 | } | |
2071 | ||
2072 | /* WARNING: This routine must only be called when we have already sent | |
2073 | * a SYN packet that crossed the incoming SYN that caused this routine | |
2074 | * to get called. If this assumption fails then the initial rcv_wnd | |
2075 | * and rcv_wscale values will not be correct. | |
2076 | */ | |
2077 | int tcp_send_synack(struct sock *sk) | |
2078 | { | |
2079 | struct sk_buff* skb; | |
2080 | ||
fe067e8a | 2081 | skb = tcp_write_queue_head(sk); |
1da177e4 LT |
2082 | if (skb == NULL || !(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_SYN)) { |
2083 | printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); | |
2084 | return -EFAULT; | |
2085 | } | |
2086 | if (!(TCP_SKB_CB(skb)->flags&TCPCB_FLAG_ACK)) { | |
2087 | if (skb_cloned(skb)) { | |
2088 | struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); | |
2089 | if (nskb == NULL) | |
2090 | return -ENOMEM; | |
fe067e8a | 2091 | tcp_unlink_write_queue(skb, sk); |
1da177e4 | 2092 | skb_header_release(nskb); |
fe067e8a | 2093 | __tcp_add_write_queue_head(sk, nskb); |
1da177e4 LT |
2094 | sk_stream_free_skb(sk, skb); |
2095 | sk_charge_skb(sk, nskb); | |
2096 | skb = nskb; | |
2097 | } | |
2098 | ||
2099 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ACK; | |
2100 | TCP_ECN_send_synack(tcp_sk(sk), skb); | |
2101 | } | |
2102 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
dfb4b9dc | 2103 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
1da177e4 LT |
2104 | } |
2105 | ||
2106 | /* | |
2107 | * Prepare a SYN-ACK. | |
2108 | */ | |
2109 | struct sk_buff * tcp_make_synack(struct sock *sk, struct dst_entry *dst, | |
60236fdd | 2110 | struct request_sock *req) |
1da177e4 | 2111 | { |
2e6599cb | 2112 | struct inet_request_sock *ireq = inet_rsk(req); |
1da177e4 LT |
2113 | struct tcp_sock *tp = tcp_sk(sk); |
2114 | struct tcphdr *th; | |
2115 | int tcp_header_size; | |
2116 | struct sk_buff *skb; | |
cfb6eeb4 YH |
2117 | #ifdef CONFIG_TCP_MD5SIG |
2118 | struct tcp_md5sig_key *md5; | |
2119 | __u8 *md5_hash_location; | |
2120 | #endif | |
1da177e4 LT |
2121 | |
2122 | skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC); | |
2123 | if (skb == NULL) | |
2124 | return NULL; | |
2125 | ||
2126 | /* Reserve space for headers. */ | |
2127 | skb_reserve(skb, MAX_TCP_HEADER); | |
2128 | ||
2129 | skb->dst = dst_clone(dst); | |
2130 | ||
2131 | tcp_header_size = (sizeof(struct tcphdr) + TCPOLEN_MSS + | |
2e6599cb ACM |
2132 | (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0) + |
2133 | (ireq->wscale_ok ? TCPOLEN_WSCALE_ALIGNED : 0) + | |
1da177e4 | 2134 | /* SACK_PERM is in the place of NOP NOP of TS */ |
2e6599cb | 2135 | ((ireq->sack_ok && !ireq->tstamp_ok) ? TCPOLEN_SACKPERM_ALIGNED : 0)); |
cfb6eeb4 YH |
2136 | |
2137 | #ifdef CONFIG_TCP_MD5SIG | |
2138 | /* Are we doing MD5 on this segment? If so - make room for it */ | |
2139 | md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); | |
2140 | if (md5) | |
2141 | tcp_header_size += TCPOLEN_MD5SIG_ALIGNED; | |
2142 | #endif | |
aa8223c7 ACM |
2143 | skb_push(skb, tcp_header_size); |
2144 | skb_reset_transport_header(skb); | |
1da177e4 | 2145 | |
aa8223c7 | 2146 | th = tcp_hdr(skb); |
1da177e4 LT |
2147 | memset(th, 0, sizeof(struct tcphdr)); |
2148 | th->syn = 1; | |
2149 | th->ack = 1; | |
1da177e4 LT |
2150 | TCP_ECN_make_synack(req, th); |
2151 | th->source = inet_sk(sk)->sport; | |
2e6599cb ACM |
2152 | th->dest = ireq->rmt_port; |
2153 | TCP_SKB_CB(skb)->seq = tcp_rsk(req)->snt_isn; | |
1da177e4 LT |
2154 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq + 1; |
2155 | TCP_SKB_CB(skb)->sacked = 0; | |
7967168c HX |
2156 | skb_shinfo(skb)->gso_segs = 1; |
2157 | skb_shinfo(skb)->gso_size = 0; | |
2158 | skb_shinfo(skb)->gso_type = 0; | |
1da177e4 | 2159 | th->seq = htonl(TCP_SKB_CB(skb)->seq); |
2e6599cb | 2160 | th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); |
1da177e4 | 2161 | if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ |
e905a9ed | 2162 | __u8 rcv_wscale; |
1da177e4 LT |
2163 | /* Set this up on the first call only */ |
2164 | req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); | |
2165 | /* tcp_full_space because it is guaranteed to be the first packet */ | |
e905a9ed | 2166 | tcp_select_initial_window(tcp_full_space(sk), |
2e6599cb | 2167 | dst_metric(dst, RTAX_ADVMSS) - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), |
1da177e4 LT |
2168 | &req->rcv_wnd, |
2169 | &req->window_clamp, | |
2e6599cb | 2170 | ireq->wscale_ok, |
1da177e4 | 2171 | &rcv_wscale); |
e905a9ed | 2172 | ireq->rcv_wscale = rcv_wscale; |
1da177e4 LT |
2173 | } |
2174 | ||
2175 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ | |
600ff0c2 | 2176 | th->window = htons(min(req->rcv_wnd, 65535U)); |
1da177e4 LT |
2177 | |
2178 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
df7a3b07 | 2179 | tcp_syn_build_options((__be32 *)(th + 1), dst_metric(dst, RTAX_ADVMSS), ireq->tstamp_ok, |
2e6599cb | 2180 | ireq->sack_ok, ireq->wscale_ok, ireq->rcv_wscale, |
1da177e4 | 2181 | TCP_SKB_CB(skb)->when, |
cfb6eeb4 YH |
2182 | req->ts_recent, |
2183 | ( | |
2184 | #ifdef CONFIG_TCP_MD5SIG | |
2185 | md5 ? &md5_hash_location : | |
2186 | #endif | |
2187 | NULL) | |
2188 | ); | |
1da177e4 LT |
2189 | |
2190 | skb->csum = 0; | |
2191 | th->doff = (tcp_header_size >> 2); | |
2192 | TCP_INC_STATS(TCP_MIB_OUTSEGS); | |
cfb6eeb4 YH |
2193 | |
2194 | #ifdef CONFIG_TCP_MD5SIG | |
2195 | /* Okay, we have all we need - do the md5 hash if needed */ | |
2196 | if (md5) { | |
2197 | tp->af_specific->calc_md5_hash(md5_hash_location, | |
2198 | md5, | |
2199 | NULL, dst, req, | |
aa8223c7 | 2200 | tcp_hdr(skb), sk->sk_protocol, |
cfb6eeb4 YH |
2201 | skb->len); |
2202 | } | |
2203 | #endif | |
2204 | ||
1da177e4 LT |
2205 | return skb; |
2206 | } | |
2207 | ||
e905a9ed | 2208 | /* |
1da177e4 | 2209 | * Do all connect socket setups that can be done AF independent. |
e905a9ed | 2210 | */ |
40efc6fa | 2211 | static void tcp_connect_init(struct sock *sk) |
1da177e4 LT |
2212 | { |
2213 | struct dst_entry *dst = __sk_dst_get(sk); | |
2214 | struct tcp_sock *tp = tcp_sk(sk); | |
2215 | __u8 rcv_wscale; | |
2216 | ||
2217 | /* We'll fix this up when we get a response from the other end. | |
2218 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. | |
2219 | */ | |
2220 | tp->tcp_header_len = sizeof(struct tcphdr) + | |
2221 | (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); | |
2222 | ||
cfb6eeb4 YH |
2223 | #ifdef CONFIG_TCP_MD5SIG |
2224 | if (tp->af_specific->md5_lookup(sk, sk) != NULL) | |
2225 | tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; | |
2226 | #endif | |
2227 | ||
1da177e4 LT |
2228 | /* If user gave his TCP_MAXSEG, record it to clamp */ |
2229 | if (tp->rx_opt.user_mss) | |
2230 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
2231 | tp->max_window = 0; | |
5d424d5a | 2232 | tcp_mtup_init(sk); |
1da177e4 LT |
2233 | tcp_sync_mss(sk, dst_mtu(dst)); |
2234 | ||
2235 | if (!tp->window_clamp) | |
2236 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); | |
2237 | tp->advmss = dst_metric(dst, RTAX_ADVMSS); | |
2238 | tcp_initialize_rcv_mss(sk); | |
1da177e4 LT |
2239 | |
2240 | tcp_select_initial_window(tcp_full_space(sk), | |
2241 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), | |
2242 | &tp->rcv_wnd, | |
2243 | &tp->window_clamp, | |
2244 | sysctl_tcp_window_scaling, | |
2245 | &rcv_wscale); | |
2246 | ||
2247 | tp->rx_opt.rcv_wscale = rcv_wscale; | |
2248 | tp->rcv_ssthresh = tp->rcv_wnd; | |
2249 | ||
2250 | sk->sk_err = 0; | |
2251 | sock_reset_flag(sk, SOCK_DONE); | |
2252 | tp->snd_wnd = 0; | |
2253 | tcp_init_wl(tp, tp->write_seq, 0); | |
2254 | tp->snd_una = tp->write_seq; | |
2255 | tp->snd_sml = tp->write_seq; | |
2256 | tp->rcv_nxt = 0; | |
2257 | tp->rcv_wup = 0; | |
2258 | tp->copied_seq = 0; | |
2259 | ||
463c84b9 ACM |
2260 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
2261 | inet_csk(sk)->icsk_retransmits = 0; | |
1da177e4 LT |
2262 | tcp_clear_retrans(tp); |
2263 | } | |
2264 | ||
2265 | /* | |
2266 | * Build a SYN and send it off. | |
e905a9ed | 2267 | */ |
1da177e4 LT |
2268 | int tcp_connect(struct sock *sk) |
2269 | { | |
2270 | struct tcp_sock *tp = tcp_sk(sk); | |
2271 | struct sk_buff *buff; | |
2272 | ||
2273 | tcp_connect_init(sk); | |
2274 | ||
d179cd12 | 2275 | buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); |
1da177e4 LT |
2276 | if (unlikely(buff == NULL)) |
2277 | return -ENOBUFS; | |
2278 | ||
2279 | /* Reserve space for headers. */ | |
2280 | skb_reserve(buff, MAX_TCP_HEADER); | |
2281 | ||
2282 | TCP_SKB_CB(buff)->flags = TCPCB_FLAG_SYN; | |
9e412ba7 | 2283 | TCP_ECN_send_syn(sk, buff); |
1da177e4 | 2284 | TCP_SKB_CB(buff)->sacked = 0; |
7967168c HX |
2285 | skb_shinfo(buff)->gso_segs = 1; |
2286 | skb_shinfo(buff)->gso_size = 0; | |
2287 | skb_shinfo(buff)->gso_type = 0; | |
1da177e4 | 2288 | buff->csum = 0; |
bd37a088 | 2289 | tp->snd_nxt = tp->write_seq; |
1da177e4 LT |
2290 | TCP_SKB_CB(buff)->seq = tp->write_seq++; |
2291 | TCP_SKB_CB(buff)->end_seq = tp->write_seq; | |
1da177e4 LT |
2292 | |
2293 | /* Send it off. */ | |
2294 | TCP_SKB_CB(buff)->when = tcp_time_stamp; | |
2295 | tp->retrans_stamp = TCP_SKB_CB(buff)->when; | |
2296 | skb_header_release(buff); | |
fe067e8a | 2297 | __tcp_add_write_queue_tail(sk, buff); |
1da177e4 LT |
2298 | sk_charge_skb(sk, buff); |
2299 | tp->packets_out += tcp_skb_pcount(buff); | |
dfb4b9dc | 2300 | tcp_transmit_skb(sk, buff, 1, GFP_KERNEL); |
bd37a088 WY |
2301 | |
2302 | /* We change tp->snd_nxt after the tcp_transmit_skb() call | |
2303 | * in order to make this packet get counted in tcpOutSegs. | |
2304 | */ | |
2305 | tp->snd_nxt = tp->write_seq; | |
2306 | tp->pushed_seq = tp->write_seq; | |
1da177e4 LT |
2307 | TCP_INC_STATS(TCP_MIB_ACTIVEOPENS); |
2308 | ||
2309 | /* Timer for repeating the SYN until an answer. */ | |
3f421baa ACM |
2310 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
2311 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
1da177e4 LT |
2312 | return 0; |
2313 | } | |
2314 | ||
2315 | /* Send out a delayed ack, the caller does the policy checking | |
2316 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() | |
2317 | * for details. | |
2318 | */ | |
2319 | void tcp_send_delayed_ack(struct sock *sk) | |
2320 | { | |
463c84b9 ACM |
2321 | struct inet_connection_sock *icsk = inet_csk(sk); |
2322 | int ato = icsk->icsk_ack.ato; | |
1da177e4 LT |
2323 | unsigned long timeout; |
2324 | ||
2325 | if (ato > TCP_DELACK_MIN) { | |
463c84b9 | 2326 | const struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
2327 | int max_ato = HZ/2; |
2328 | ||
463c84b9 | 2329 | if (icsk->icsk_ack.pingpong || (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) |
1da177e4 LT |
2330 | max_ato = TCP_DELACK_MAX; |
2331 | ||
2332 | /* Slow path, intersegment interval is "high". */ | |
2333 | ||
2334 | /* If some rtt estimate is known, use it to bound delayed ack. | |
463c84b9 | 2335 | * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements |
1da177e4 LT |
2336 | * directly. |
2337 | */ | |
2338 | if (tp->srtt) { | |
2339 | int rtt = max(tp->srtt>>3, TCP_DELACK_MIN); | |
2340 | ||
2341 | if (rtt < max_ato) | |
2342 | max_ato = rtt; | |
2343 | } | |
2344 | ||
2345 | ato = min(ato, max_ato); | |
2346 | } | |
2347 | ||
2348 | /* Stay within the limit we were given */ | |
2349 | timeout = jiffies + ato; | |
2350 | ||
2351 | /* Use new timeout only if there wasn't a older one earlier. */ | |
463c84b9 | 2352 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { |
1da177e4 LT |
2353 | /* If delack timer was blocked or is about to expire, |
2354 | * send ACK now. | |
2355 | */ | |
463c84b9 ACM |
2356 | if (icsk->icsk_ack.blocked || |
2357 | time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { | |
1da177e4 LT |
2358 | tcp_send_ack(sk); |
2359 | return; | |
2360 | } | |
2361 | ||
463c84b9 ACM |
2362 | if (!time_before(timeout, icsk->icsk_ack.timeout)) |
2363 | timeout = icsk->icsk_ack.timeout; | |
1da177e4 | 2364 | } |
463c84b9 ACM |
2365 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; |
2366 | icsk->icsk_ack.timeout = timeout; | |
2367 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); | |
1da177e4 LT |
2368 | } |
2369 | ||
2370 | /* This routine sends an ack and also updates the window. */ | |
2371 | void tcp_send_ack(struct sock *sk) | |
2372 | { | |
2373 | /* If we have been reset, we may not send again. */ | |
2374 | if (sk->sk_state != TCP_CLOSE) { | |
1da177e4 LT |
2375 | struct sk_buff *buff; |
2376 | ||
2377 | /* We are not putting this on the write queue, so | |
2378 | * tcp_transmit_skb() will set the ownership to this | |
2379 | * sock. | |
2380 | */ | |
2381 | buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); | |
2382 | if (buff == NULL) { | |
463c84b9 ACM |
2383 | inet_csk_schedule_ack(sk); |
2384 | inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; | |
3f421baa ACM |
2385 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, |
2386 | TCP_DELACK_MAX, TCP_RTO_MAX); | |
1da177e4 LT |
2387 | return; |
2388 | } | |
2389 | ||
2390 | /* Reserve space for headers and prepare control bits. */ | |
2391 | skb_reserve(buff, MAX_TCP_HEADER); | |
2392 | buff->csum = 0; | |
2393 | TCP_SKB_CB(buff)->flags = TCPCB_FLAG_ACK; | |
2394 | TCP_SKB_CB(buff)->sacked = 0; | |
7967168c HX |
2395 | skb_shinfo(buff)->gso_segs = 1; |
2396 | skb_shinfo(buff)->gso_size = 0; | |
2397 | skb_shinfo(buff)->gso_type = 0; | |
1da177e4 LT |
2398 | |
2399 | /* Send it off, this clears delayed acks for us. */ | |
9e412ba7 | 2400 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(buff)->end_seq = tcp_acceptable_seq(sk); |
1da177e4 | 2401 | TCP_SKB_CB(buff)->when = tcp_time_stamp; |
dfb4b9dc | 2402 | tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); |
1da177e4 LT |
2403 | } |
2404 | } | |
2405 | ||
2406 | /* This routine sends a packet with an out of date sequence | |
2407 | * number. It assumes the other end will try to ack it. | |
2408 | * | |
2409 | * Question: what should we make while urgent mode? | |
2410 | * 4.4BSD forces sending single byte of data. We cannot send | |
2411 | * out of window data, because we have SND.NXT==SND.MAX... | |
2412 | * | |
2413 | * Current solution: to send TWO zero-length segments in urgent mode: | |
2414 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is | |
2415 | * out-of-date with SND.UNA-1 to probe window. | |
2416 | */ | |
2417 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent) | |
2418 | { | |
2419 | struct tcp_sock *tp = tcp_sk(sk); | |
2420 | struct sk_buff *skb; | |
2421 | ||
2422 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ | |
2423 | skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); | |
e905a9ed | 2424 | if (skb == NULL) |
1da177e4 LT |
2425 | return -1; |
2426 | ||
2427 | /* Reserve space for headers and set control bits. */ | |
2428 | skb_reserve(skb, MAX_TCP_HEADER); | |
2429 | skb->csum = 0; | |
2430 | TCP_SKB_CB(skb)->flags = TCPCB_FLAG_ACK; | |
2431 | TCP_SKB_CB(skb)->sacked = urgent; | |
7967168c HX |
2432 | skb_shinfo(skb)->gso_segs = 1; |
2433 | skb_shinfo(skb)->gso_size = 0; | |
2434 | skb_shinfo(skb)->gso_type = 0; | |
1da177e4 LT |
2435 | |
2436 | /* Use a previous sequence. This should cause the other | |
2437 | * end to send an ack. Don't queue or clone SKB, just | |
2438 | * send it. | |
2439 | */ | |
2440 | TCP_SKB_CB(skb)->seq = urgent ? tp->snd_una : tp->snd_una - 1; | |
2441 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(skb)->seq; | |
2442 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
dfb4b9dc | 2443 | return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); |
1da177e4 LT |
2444 | } |
2445 | ||
2446 | int tcp_write_wakeup(struct sock *sk) | |
2447 | { | |
2448 | if (sk->sk_state != TCP_CLOSE) { | |
2449 | struct tcp_sock *tp = tcp_sk(sk); | |
2450 | struct sk_buff *skb; | |
2451 | ||
fe067e8a | 2452 | if ((skb = tcp_send_head(sk)) != NULL && |
1da177e4 LT |
2453 | before(TCP_SKB_CB(skb)->seq, tp->snd_una+tp->snd_wnd)) { |
2454 | int err; | |
2455 | unsigned int mss = tcp_current_mss(sk, 0); | |
2456 | unsigned int seg_size = tp->snd_una+tp->snd_wnd-TCP_SKB_CB(skb)->seq; | |
2457 | ||
2458 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) | |
2459 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; | |
2460 | ||
2461 | /* We are probing the opening of a window | |
2462 | * but the window size is != 0 | |
2463 | * must have been a result SWS avoidance ( sender ) | |
2464 | */ | |
2465 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || | |
2466 | skb->len > mss) { | |
2467 | seg_size = min(seg_size, mss); | |
2468 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; | |
846998ae | 2469 | if (tcp_fragment(sk, skb, seg_size, mss)) |
1da177e4 | 2470 | return -1; |
1da177e4 | 2471 | } else if (!tcp_skb_pcount(skb)) |
846998ae | 2472 | tcp_set_skb_tso_segs(sk, skb, mss); |
1da177e4 LT |
2473 | |
2474 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_PSH; | |
2475 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
dfb4b9dc | 2476 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
1da177e4 | 2477 | if (!err) { |
9e412ba7 | 2478 | update_send_head(sk, skb); |
1da177e4 LT |
2479 | } |
2480 | return err; | |
2481 | } else { | |
2482 | if (tp->urg_mode && | |
2483 | between(tp->snd_up, tp->snd_una+1, tp->snd_una+0xFFFF)) | |
2484 | tcp_xmit_probe_skb(sk, TCPCB_URG); | |
2485 | return tcp_xmit_probe_skb(sk, 0); | |
2486 | } | |
2487 | } | |
2488 | return -1; | |
2489 | } | |
2490 | ||
2491 | /* A window probe timeout has occurred. If window is not closed send | |
2492 | * a partial packet else a zero probe. | |
2493 | */ | |
2494 | void tcp_send_probe0(struct sock *sk) | |
2495 | { | |
463c84b9 | 2496 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2497 | struct tcp_sock *tp = tcp_sk(sk); |
2498 | int err; | |
2499 | ||
2500 | err = tcp_write_wakeup(sk); | |
2501 | ||
fe067e8a | 2502 | if (tp->packets_out || !tcp_send_head(sk)) { |
1da177e4 | 2503 | /* Cancel probe timer, if it is not required. */ |
6687e988 | 2504 | icsk->icsk_probes_out = 0; |
463c84b9 | 2505 | icsk->icsk_backoff = 0; |
1da177e4 LT |
2506 | return; |
2507 | } | |
2508 | ||
2509 | if (err <= 0) { | |
463c84b9 ACM |
2510 | if (icsk->icsk_backoff < sysctl_tcp_retries2) |
2511 | icsk->icsk_backoff++; | |
6687e988 | 2512 | icsk->icsk_probes_out++; |
e905a9ed | 2513 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
3f421baa ACM |
2514 | min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), |
2515 | TCP_RTO_MAX); | |
1da177e4 LT |
2516 | } else { |
2517 | /* If packet was not sent due to local congestion, | |
6687e988 | 2518 | * do not backoff and do not remember icsk_probes_out. |
1da177e4 LT |
2519 | * Let local senders to fight for local resources. |
2520 | * | |
2521 | * Use accumulated backoff yet. | |
2522 | */ | |
6687e988 ACM |
2523 | if (!icsk->icsk_probes_out) |
2524 | icsk->icsk_probes_out = 1; | |
e905a9ed | 2525 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
463c84b9 | 2526 | min(icsk->icsk_rto << icsk->icsk_backoff, |
3f421baa ACM |
2527 | TCP_RESOURCE_PROBE_INTERVAL), |
2528 | TCP_RTO_MAX); | |
1da177e4 LT |
2529 | } |
2530 | } | |
2531 | ||
2532 | EXPORT_SYMBOL(tcp_connect); | |
2533 | EXPORT_SYMBOL(tcp_make_synack); | |
2534 | EXPORT_SYMBOL(tcp_simple_retransmit); | |
2535 | EXPORT_SYMBOL(tcp_sync_mss); | |
f4805ede | 2536 | EXPORT_SYMBOL(sysctl_tcp_tso_win_divisor); |
5d424d5a | 2537 | EXPORT_SYMBOL(tcp_mtup_init); |