]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * INET An implementation of the TCP/IP protocol suite for the LINUX | |
3 | * operating system. INET is implemented using the BSD Socket | |
4 | * interface as the means of communication with the user level. | |
5 | * | |
6 | * Implementation of the Transmission Control Protocol(TCP). | |
7 | * | |
02c30a84 | 8 | * Authors: Ross Biro |
1da177e4 LT |
9 | * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> |
10 | * Mark Evans, <evansmp@uhura.aston.ac.uk> | |
11 | * Corey Minyard <wf-rch!minyard@relay.EU.net> | |
12 | * Florian La Roche, <flla@stud.uni-sb.de> | |
13 | * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> | |
14 | * Linus Torvalds, <torvalds@cs.helsinki.fi> | |
15 | * Alan Cox, <gw4pts@gw4pts.ampr.org> | |
16 | * Matthew Dillon, <dillon@apollo.west.oic.com> | |
17 | * Arnt Gulbrandsen, <agulbra@nvg.unit.no> | |
18 | * Jorge Cwik, <jorge@laser.satlink.net> | |
19 | */ | |
20 | ||
21 | /* | |
22 | * Changes: Pedro Roque : Retransmit queue handled by TCP. | |
23 | * : Fragmentation on mtu decrease | |
24 | * : Segment collapse on retransmit | |
25 | * : AF independence | |
26 | * | |
27 | * Linus Torvalds : send_delayed_ack | |
28 | * David S. Miller : Charge memory using the right skb | |
29 | * during syn/ack processing. | |
30 | * David S. Miller : Output engine completely rewritten. | |
31 | * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. | |
32 | * Cacophonix Gaul : draft-minshall-nagle-01 | |
33 | * J Hadi Salim : ECN support | |
34 | * | |
35 | */ | |
36 | ||
37 | #include <net/tcp.h> | |
38 | ||
39 | #include <linux/compiler.h> | |
5a0e3ad6 | 40 | #include <linux/gfp.h> |
1da177e4 | 41 | #include <linux/module.h> |
1da177e4 LT |
42 | |
43 | /* People can turn this off for buggy TCP's found in printers etc. */ | |
ab32ea5d | 44 | int sysctl_tcp_retrans_collapse __read_mostly = 1; |
1da177e4 | 45 | |
09cb105e | 46 | /* People can turn this on to work with those rare, broken TCPs that |
15d99e02 RJ |
47 | * interpret the window field as a signed quantity. |
48 | */ | |
ab32ea5d | 49 | int sysctl_tcp_workaround_signed_windows __read_mostly = 0; |
15d99e02 | 50 | |
1da177e4 LT |
51 | /* This limits the percentage of the congestion window which we |
52 | * will allow a single TSO frame to consume. Building TSO frames | |
53 | * which are too large can cause TCP streams to be bursty. | |
54 | */ | |
ab32ea5d | 55 | int sysctl_tcp_tso_win_divisor __read_mostly = 3; |
1da177e4 | 56 | |
ab32ea5d BH |
57 | int sysctl_tcp_mtu_probing __read_mostly = 0; |
58 | int sysctl_tcp_base_mss __read_mostly = 512; | |
5d424d5a | 59 | |
35089bb2 | 60 | /* By default, RFC2861 behavior. */ |
ab32ea5d | 61 | int sysctl_tcp_slow_start_after_idle __read_mostly = 1; |
35089bb2 | 62 | |
519855c5 | 63 | int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */ |
e6b09cca | 64 | EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size); |
519855c5 WAS |
65 | |
66 | ||
67edfef7 | 67 | /* Account for new data that has been sent to the network. */ |
66f5fe62 | 68 | static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 69 | { |
9e412ba7 | 70 | struct tcp_sock *tp = tcp_sk(sk); |
66f5fe62 | 71 | unsigned int prior_packets = tp->packets_out; |
9e412ba7 | 72 | |
fe067e8a | 73 | tcp_advance_send_head(sk, skb); |
1da177e4 | 74 | tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; |
8512430e IJ |
75 | |
76 | /* Don't override Nagle indefinately with F-RTO */ | |
77 | if (tp->frto_counter == 2) | |
78 | tp->frto_counter = 3; | |
66f5fe62 IJ |
79 | |
80 | tp->packets_out += tcp_skb_pcount(skb); | |
81 | if (!prior_packets) | |
82 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, | |
83 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
1da177e4 LT |
84 | } |
85 | ||
86 | /* SND.NXT, if window was not shrunk. | |
87 | * If window has been shrunk, what should we make? It is not clear at all. | |
88 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( | |
89 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already | |
90 | * invalid. OK, let's make this for now: | |
91 | */ | |
9e412ba7 | 92 | static inline __u32 tcp_acceptable_seq(struct sock *sk) |
1da177e4 | 93 | { |
9e412ba7 IJ |
94 | struct tcp_sock *tp = tcp_sk(sk); |
95 | ||
90840def | 96 | if (!before(tcp_wnd_end(tp), tp->snd_nxt)) |
1da177e4 LT |
97 | return tp->snd_nxt; |
98 | else | |
90840def | 99 | return tcp_wnd_end(tp); |
1da177e4 LT |
100 | } |
101 | ||
102 | /* Calculate mss to advertise in SYN segment. | |
103 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: | |
104 | * | |
105 | * 1. It is independent of path mtu. | |
106 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. | |
107 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of | |
108 | * attached devices, because some buggy hosts are confused by | |
109 | * large MSS. | |
110 | * 4. We do not make 3, we advertise MSS, calculated from first | |
111 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. | |
112 | * This may be overridden via information stored in routing table. | |
113 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, | |
114 | * probably even Jumbo". | |
115 | */ | |
116 | static __u16 tcp_advertise_mss(struct sock *sk) | |
117 | { | |
118 | struct tcp_sock *tp = tcp_sk(sk); | |
119 | struct dst_entry *dst = __sk_dst_get(sk); | |
120 | int mss = tp->advmss; | |
121 | ||
122 | if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) { | |
123 | mss = dst_metric(dst, RTAX_ADVMSS); | |
124 | tp->advmss = mss; | |
125 | } | |
126 | ||
127 | return (__u16)mss; | |
128 | } | |
129 | ||
130 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". | |
131 | * This is the first part of cwnd validation mechanism. */ | |
463c84b9 | 132 | static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst) |
1da177e4 | 133 | { |
463c84b9 | 134 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
135 | s32 delta = tcp_time_stamp - tp->lsndtime; |
136 | u32 restart_cwnd = tcp_init_cwnd(tp, dst); | |
137 | u32 cwnd = tp->snd_cwnd; | |
138 | ||
6687e988 | 139 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); |
1da177e4 | 140 | |
6687e988 | 141 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
1da177e4 LT |
142 | restart_cwnd = min(restart_cwnd, cwnd); |
143 | ||
463c84b9 | 144 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) |
1da177e4 LT |
145 | cwnd >>= 1; |
146 | tp->snd_cwnd = max(cwnd, restart_cwnd); | |
147 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
148 | tp->snd_cwnd_used = 0; | |
149 | } | |
150 | ||
67edfef7 | 151 | /* Congestion state accounting after a packet has been sent. */ |
40efc6fa SH |
152 | static void tcp_event_data_sent(struct tcp_sock *tp, |
153 | struct sk_buff *skb, struct sock *sk) | |
1da177e4 | 154 | { |
463c84b9 ACM |
155 | struct inet_connection_sock *icsk = inet_csk(sk); |
156 | const u32 now = tcp_time_stamp; | |
1da177e4 | 157 | |
35089bb2 DM |
158 | if (sysctl_tcp_slow_start_after_idle && |
159 | (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) | |
463c84b9 | 160 | tcp_cwnd_restart(sk, __sk_dst_get(sk)); |
1da177e4 LT |
161 | |
162 | tp->lsndtime = now; | |
163 | ||
164 | /* If it is a reply for ato after last received | |
165 | * packet, enter pingpong mode. | |
166 | */ | |
463c84b9 ACM |
167 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) |
168 | icsk->icsk_ack.pingpong = 1; | |
1da177e4 LT |
169 | } |
170 | ||
67edfef7 | 171 | /* Account for an ACK we sent. */ |
40efc6fa | 172 | static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) |
1da177e4 | 173 | { |
463c84b9 ACM |
174 | tcp_dec_quickack_mode(sk, pkts); |
175 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); | |
1da177e4 LT |
176 | } |
177 | ||
178 | /* Determine a window scaling and initial window to offer. | |
179 | * Based on the assumption that the given amount of space | |
180 | * will be offered. Store the results in the tp structure. | |
181 | * NOTE: for smooth operation initial space offering should | |
182 | * be a multiple of mss if possible. We assume here that mss >= 1. | |
183 | * This MUST be enforced by all callers. | |
184 | */ | |
185 | void tcp_select_initial_window(int __space, __u32 mss, | |
186 | __u32 *rcv_wnd, __u32 *window_clamp, | |
31d12926 | 187 | int wscale_ok, __u8 *rcv_wscale, |
188 | __u32 init_rcv_wnd) | |
1da177e4 LT |
189 | { |
190 | unsigned int space = (__space < 0 ? 0 : __space); | |
191 | ||
192 | /* If no clamp set the clamp to the max possible scaled window */ | |
193 | if (*window_clamp == 0) | |
194 | (*window_clamp) = (65535 << 14); | |
195 | space = min(*window_clamp, space); | |
196 | ||
197 | /* Quantize space offering to a multiple of mss if possible. */ | |
198 | if (space > mss) | |
199 | space = (space / mss) * mss; | |
200 | ||
201 | /* NOTE: offering an initial window larger than 32767 | |
15d99e02 RJ |
202 | * will break some buggy TCP stacks. If the admin tells us |
203 | * it is likely we could be speaking with such a buggy stack | |
204 | * we will truncate our initial window offering to 32K-1 | |
205 | * unless the remote has sent us a window scaling option, | |
206 | * which we interpret as a sign the remote TCP is not | |
207 | * misinterpreting the window field as a signed quantity. | |
1da177e4 | 208 | */ |
15d99e02 RJ |
209 | if (sysctl_tcp_workaround_signed_windows) |
210 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); | |
211 | else | |
212 | (*rcv_wnd) = space; | |
213 | ||
1da177e4 LT |
214 | (*rcv_wscale) = 0; |
215 | if (wscale_ok) { | |
216 | /* Set window scaling on max possible window | |
e905a9ed | 217 | * See RFC1323 for an explanation of the limit to 14 |
1da177e4 LT |
218 | */ |
219 | space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); | |
316c1592 | 220 | space = min_t(u32, space, *window_clamp); |
1da177e4 LT |
221 | while (space > 65535 && (*rcv_wscale) < 14) { |
222 | space >>= 1; | |
223 | (*rcv_wscale)++; | |
224 | } | |
225 | } | |
226 | ||
0705c6f0 | 227 | /* Set initial window to value enough for senders, following RFC5681. */ |
056834d9 | 228 | if (mss > (1 << *rcv_wscale)) { |
0705c6f0 GR |
229 | int init_cwnd = rfc3390_bytes_to_packets(mss); |
230 | ||
31d12926 | 231 | /* when initializing use the value from init_rcv_wnd |
232 | * rather than the default from above | |
233 | */ | |
b1afde60 ND |
234 | if (init_rcv_wnd) |
235 | *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); | |
236 | else | |
237 | *rcv_wnd = min(*rcv_wnd, init_cwnd * mss); | |
1da177e4 LT |
238 | } |
239 | ||
240 | /* Set the clamp no higher than max representable value */ | |
241 | (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); | |
242 | } | |
4bc2f18b | 243 | EXPORT_SYMBOL(tcp_select_initial_window); |
1da177e4 LT |
244 | |
245 | /* Chose a new window to advertise, update state in tcp_sock for the | |
246 | * socket, and return result with RFC1323 scaling applied. The return | |
247 | * value can be stuffed directly into th->window for an outgoing | |
248 | * frame. | |
249 | */ | |
40efc6fa | 250 | static u16 tcp_select_window(struct sock *sk) |
1da177e4 LT |
251 | { |
252 | struct tcp_sock *tp = tcp_sk(sk); | |
253 | u32 cur_win = tcp_receive_window(tp); | |
254 | u32 new_win = __tcp_select_window(sk); | |
255 | ||
256 | /* Never shrink the offered window */ | |
2de979bd | 257 | if (new_win < cur_win) { |
1da177e4 LT |
258 | /* Danger Will Robinson! |
259 | * Don't update rcv_wup/rcv_wnd here or else | |
260 | * we will not be able to advertise a zero | |
261 | * window in time. --DaveM | |
262 | * | |
263 | * Relax Will Robinson. | |
264 | */ | |
607bfbf2 | 265 | new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); |
1da177e4 LT |
266 | } |
267 | tp->rcv_wnd = new_win; | |
268 | tp->rcv_wup = tp->rcv_nxt; | |
269 | ||
270 | /* Make sure we do not exceed the maximum possible | |
271 | * scaled window. | |
272 | */ | |
15d99e02 | 273 | if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) |
1da177e4 LT |
274 | new_win = min(new_win, MAX_TCP_WINDOW); |
275 | else | |
276 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); | |
277 | ||
278 | /* RFC1323 scaling applied */ | |
279 | new_win >>= tp->rx_opt.rcv_wscale; | |
280 | ||
281 | /* If we advertise zero window, disable fast path. */ | |
282 | if (new_win == 0) | |
283 | tp->pred_flags = 0; | |
284 | ||
285 | return new_win; | |
286 | } | |
287 | ||
67edfef7 | 288 | /* Packet ECN state for a SYN-ACK */ |
056834d9 | 289 | static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb) |
bdf1ee5d | 290 | { |
a3433f35 | 291 | TCP_SKB_CB(skb)->flags &= ~TCPHDR_CWR; |
056834d9 | 292 | if (!(tp->ecn_flags & TCP_ECN_OK)) |
a3433f35 | 293 | TCP_SKB_CB(skb)->flags &= ~TCPHDR_ECE; |
bdf1ee5d IJ |
294 | } |
295 | ||
67edfef7 | 296 | /* Packet ECN state for a SYN. */ |
bdf1ee5d IJ |
297 | static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) |
298 | { | |
299 | struct tcp_sock *tp = tcp_sk(sk); | |
300 | ||
301 | tp->ecn_flags = 0; | |
255cac91 | 302 | if (sysctl_tcp_ecn == 1) { |
a3433f35 | 303 | TCP_SKB_CB(skb)->flags |= TCPHDR_ECE | TCPHDR_CWR; |
bdf1ee5d IJ |
304 | tp->ecn_flags = TCP_ECN_OK; |
305 | } | |
306 | } | |
307 | ||
308 | static __inline__ void | |
309 | TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th) | |
310 | { | |
311 | if (inet_rsk(req)->ecn_ok) | |
312 | th->ece = 1; | |
313 | } | |
314 | ||
67edfef7 AK |
315 | /* Set up ECN state for a packet on a ESTABLISHED socket that is about to |
316 | * be sent. | |
317 | */ | |
bdf1ee5d IJ |
318 | static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, |
319 | int tcp_header_len) | |
320 | { | |
321 | struct tcp_sock *tp = tcp_sk(sk); | |
322 | ||
323 | if (tp->ecn_flags & TCP_ECN_OK) { | |
324 | /* Not-retransmitted data segment: set ECT and inject CWR. */ | |
325 | if (skb->len != tcp_header_len && | |
326 | !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { | |
327 | INET_ECN_xmit(sk); | |
056834d9 | 328 | if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { |
bdf1ee5d IJ |
329 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; |
330 | tcp_hdr(skb)->cwr = 1; | |
331 | skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; | |
332 | } | |
333 | } else { | |
334 | /* ACK or retransmitted segment: clear ECT|CE */ | |
335 | INET_ECN_dontxmit(sk); | |
336 | } | |
337 | if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) | |
338 | tcp_hdr(skb)->ece = 1; | |
339 | } | |
340 | } | |
341 | ||
e870a8ef IJ |
342 | /* Constructs common control bits of non-data skb. If SYN/FIN is present, |
343 | * auto increment end seqno. | |
344 | */ | |
345 | static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) | |
346 | { | |
2e8e18ef | 347 | skb->ip_summed = CHECKSUM_PARTIAL; |
e870a8ef IJ |
348 | skb->csum = 0; |
349 | ||
350 | TCP_SKB_CB(skb)->flags = flags; | |
351 | TCP_SKB_CB(skb)->sacked = 0; | |
352 | ||
353 | skb_shinfo(skb)->gso_segs = 1; | |
354 | skb_shinfo(skb)->gso_size = 0; | |
355 | skb_shinfo(skb)->gso_type = 0; | |
356 | ||
357 | TCP_SKB_CB(skb)->seq = seq; | |
a3433f35 | 358 | if (flags & (TCPHDR_SYN | TCPHDR_FIN)) |
e870a8ef IJ |
359 | seq++; |
360 | TCP_SKB_CB(skb)->end_seq = seq; | |
361 | } | |
362 | ||
33f5f57e IJ |
363 | static inline int tcp_urg_mode(const struct tcp_sock *tp) |
364 | { | |
365 | return tp->snd_una != tp->snd_up; | |
366 | } | |
367 | ||
33ad798c AL |
368 | #define OPTION_SACK_ADVERTISE (1 << 0) |
369 | #define OPTION_TS (1 << 1) | |
370 | #define OPTION_MD5 (1 << 2) | |
89e95a61 | 371 | #define OPTION_WSCALE (1 << 3) |
bd0388ae | 372 | #define OPTION_COOKIE_EXTENSION (1 << 4) |
33ad798c AL |
373 | |
374 | struct tcp_out_options { | |
375 | u8 options; /* bit field of OPTION_* */ | |
376 | u8 ws; /* window scale, 0 to disable */ | |
377 | u8 num_sack_blocks; /* number of SACK blocks to include */ | |
bd0388ae | 378 | u8 hash_size; /* bytes in hash_location */ |
33ad798c AL |
379 | u16 mss; /* 0 to disable */ |
380 | __u32 tsval, tsecr; /* need to include OPTION_TS */ | |
bd0388ae | 381 | __u8 *hash_location; /* temporary pointer, overloaded */ |
33ad798c AL |
382 | }; |
383 | ||
bd0388ae WAS |
384 | /* The sysctl int routines are generic, so check consistency here. |
385 | */ | |
386 | static u8 tcp_cookie_size_check(u8 desired) | |
387 | { | |
388 | if (desired > 0) { | |
389 | /* previously specified */ | |
390 | return desired; | |
391 | } | |
392 | if (sysctl_tcp_cookie_size <= 0) { | |
393 | /* no default specified */ | |
394 | return 0; | |
395 | } | |
396 | if (sysctl_tcp_cookie_size <= TCP_COOKIE_MIN) { | |
397 | /* value too small, specify minimum */ | |
398 | return TCP_COOKIE_MIN; | |
399 | } | |
400 | if (sysctl_tcp_cookie_size >= TCP_COOKIE_MAX) { | |
401 | /* value too large, specify maximum */ | |
402 | return TCP_COOKIE_MAX; | |
403 | } | |
404 | if (0x1 & sysctl_tcp_cookie_size) { | |
405 | /* 8-bit multiple, illegal, fix it */ | |
406 | return (u8)(sysctl_tcp_cookie_size + 0x1); | |
407 | } | |
408 | return (u8)sysctl_tcp_cookie_size; | |
409 | } | |
410 | ||
67edfef7 AK |
411 | /* Write previously computed TCP options to the packet. |
412 | * | |
413 | * Beware: Something in the Internet is very sensitive to the ordering of | |
fd6149d3 IJ |
414 | * TCP options, we learned this through the hard way, so be careful here. |
415 | * Luckily we can at least blame others for their non-compliance but from | |
416 | * inter-operatibility perspective it seems that we're somewhat stuck with | |
417 | * the ordering which we have been using if we want to keep working with | |
418 | * those broken things (not that it currently hurts anybody as there isn't | |
419 | * particular reason why the ordering would need to be changed). | |
420 | * | |
421 | * At least SACK_PERM as the first option is known to lead to a disaster | |
422 | * (but it may well be that other scenarios fail similarly). | |
423 | */ | |
33ad798c | 424 | static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, |
bd0388ae WAS |
425 | struct tcp_out_options *opts) |
426 | { | |
427 | u8 options = opts->options; /* mungable copy */ | |
428 | ||
429 | /* Having both authentication and cookies for security is redundant, | |
430 | * and there's certainly not enough room. Instead, the cookie-less | |
431 | * extension variant is proposed. | |
432 | * | |
433 | * Consider the pessimal case with authentication. The options | |
434 | * could look like: | |
435 | * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40 | |
436 | */ | |
437 | if (unlikely(OPTION_MD5 & options)) { | |
438 | if (unlikely(OPTION_COOKIE_EXTENSION & options)) { | |
439 | *ptr++ = htonl((TCPOPT_COOKIE << 24) | | |
440 | (TCPOLEN_COOKIE_BASE << 16) | | |
441 | (TCPOPT_MD5SIG << 8) | | |
442 | TCPOLEN_MD5SIG); | |
443 | } else { | |
444 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
445 | (TCPOPT_NOP << 16) | | |
446 | (TCPOPT_MD5SIG << 8) | | |
447 | TCPOLEN_MD5SIG); | |
448 | } | |
449 | options &= ~OPTION_COOKIE_EXTENSION; | |
450 | /* overload cookie hash location */ | |
451 | opts->hash_location = (__u8 *)ptr; | |
33ad798c | 452 | ptr += 4; |
40efc6fa | 453 | } |
33ad798c | 454 | |
fd6149d3 IJ |
455 | if (unlikely(opts->mss)) { |
456 | *ptr++ = htonl((TCPOPT_MSS << 24) | | |
457 | (TCPOLEN_MSS << 16) | | |
458 | opts->mss); | |
459 | } | |
460 | ||
bd0388ae WAS |
461 | if (likely(OPTION_TS & options)) { |
462 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { | |
33ad798c AL |
463 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
464 | (TCPOLEN_SACK_PERM << 16) | | |
465 | (TCPOPT_TIMESTAMP << 8) | | |
466 | TCPOLEN_TIMESTAMP); | |
bd0388ae | 467 | options &= ~OPTION_SACK_ADVERTISE; |
33ad798c AL |
468 | } else { |
469 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
470 | (TCPOPT_NOP << 16) | | |
471 | (TCPOPT_TIMESTAMP << 8) | | |
472 | TCPOLEN_TIMESTAMP); | |
473 | } | |
474 | *ptr++ = htonl(opts->tsval); | |
475 | *ptr++ = htonl(opts->tsecr); | |
476 | } | |
477 | ||
bd0388ae WAS |
478 | /* Specification requires after timestamp, so do it now. |
479 | * | |
480 | * Consider the pessimal case without authentication. The options | |
481 | * could look like: | |
482 | * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40 | |
483 | */ | |
484 | if (unlikely(OPTION_COOKIE_EXTENSION & options)) { | |
485 | __u8 *cookie_copy = opts->hash_location; | |
486 | u8 cookie_size = opts->hash_size; | |
487 | ||
488 | /* 8-bit multiple handled in tcp_cookie_size_check() above, | |
489 | * and elsewhere. | |
490 | */ | |
491 | if (0x2 & cookie_size) { | |
492 | __u8 *p = (__u8 *)ptr; | |
493 | ||
494 | /* 16-bit multiple */ | |
495 | *p++ = TCPOPT_COOKIE; | |
496 | *p++ = TCPOLEN_COOKIE_BASE + cookie_size; | |
497 | *p++ = *cookie_copy++; | |
498 | *p++ = *cookie_copy++; | |
499 | ptr++; | |
500 | cookie_size -= 2; | |
501 | } else { | |
502 | /* 32-bit multiple */ | |
503 | *ptr++ = htonl(((TCPOPT_NOP << 24) | | |
504 | (TCPOPT_NOP << 16) | | |
505 | (TCPOPT_COOKIE << 8) | | |
506 | TCPOLEN_COOKIE_BASE) + | |
507 | cookie_size); | |
508 | } | |
509 | ||
510 | if (cookie_size > 0) { | |
511 | memcpy(ptr, cookie_copy, cookie_size); | |
512 | ptr += (cookie_size / 4); | |
513 | } | |
514 | } | |
515 | ||
516 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { | |
33ad798c AL |
517 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
518 | (TCPOPT_NOP << 16) | | |
519 | (TCPOPT_SACK_PERM << 8) | | |
520 | TCPOLEN_SACK_PERM); | |
521 | } | |
522 | ||
bd0388ae | 523 | if (unlikely(OPTION_WSCALE & options)) { |
33ad798c AL |
524 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
525 | (TCPOPT_WINDOW << 16) | | |
526 | (TCPOLEN_WINDOW << 8) | | |
527 | opts->ws); | |
528 | } | |
529 | ||
530 | if (unlikely(opts->num_sack_blocks)) { | |
531 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? | |
532 | tp->duplicate_sack : tp->selective_acks; | |
40efc6fa SH |
533 | int this_sack; |
534 | ||
535 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
536 | (TCPOPT_NOP << 16) | | |
537 | (TCPOPT_SACK << 8) | | |
33ad798c | 538 | (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * |
40efc6fa | 539 | TCPOLEN_SACK_PERBLOCK))); |
2de979bd | 540 | |
33ad798c AL |
541 | for (this_sack = 0; this_sack < opts->num_sack_blocks; |
542 | ++this_sack) { | |
40efc6fa SH |
543 | *ptr++ = htonl(sp[this_sack].start_seq); |
544 | *ptr++ = htonl(sp[this_sack].end_seq); | |
545 | } | |
2de979bd | 546 | |
5861f8e5 | 547 | tp->rx_opt.dsack = 0; |
40efc6fa | 548 | } |
33ad798c AL |
549 | } |
550 | ||
67edfef7 AK |
551 | /* Compute TCP options for SYN packets. This is not the final |
552 | * network wire format yet. | |
553 | */ | |
33ad798c AL |
554 | static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb, |
555 | struct tcp_out_options *opts, | |
556 | struct tcp_md5sig_key **md5) { | |
557 | struct tcp_sock *tp = tcp_sk(sk); | |
bd0388ae | 558 | struct tcp_cookie_values *cvp = tp->cookie_values; |
bd0388ae WAS |
559 | unsigned remaining = MAX_TCP_OPTION_SPACE; |
560 | u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ? | |
561 | tcp_cookie_size_check(cvp->cookie_desired) : | |
562 | 0; | |
33ad798c | 563 | |
cfb6eeb4 | 564 | #ifdef CONFIG_TCP_MD5SIG |
33ad798c AL |
565 | *md5 = tp->af_specific->md5_lookup(sk, sk); |
566 | if (*md5) { | |
567 | opts->options |= OPTION_MD5; | |
bd0388ae | 568 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
cfb6eeb4 | 569 | } |
33ad798c AL |
570 | #else |
571 | *md5 = NULL; | |
cfb6eeb4 | 572 | #endif |
33ad798c AL |
573 | |
574 | /* We always get an MSS option. The option bytes which will be seen in | |
575 | * normal data packets should timestamps be used, must be in the MSS | |
576 | * advertised. But we subtract them from tp->mss_cache so that | |
577 | * calculations in tcp_sendmsg are simpler etc. So account for this | |
578 | * fact here if necessary. If we don't do this correctly, as a | |
579 | * receiver we won't recognize data packets as being full sized when we | |
580 | * should, and thus we won't abide by the delayed ACK rules correctly. | |
581 | * SACKs don't matter, we never delay an ACK when we have any of those | |
582 | * going out. */ | |
583 | opts->mss = tcp_advertise_mss(sk); | |
bd0388ae | 584 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c | 585 | |
bb5b7c11 | 586 | if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { |
33ad798c AL |
587 | opts->options |= OPTION_TS; |
588 | opts->tsval = TCP_SKB_CB(skb)->when; | |
589 | opts->tsecr = tp->rx_opt.ts_recent; | |
bd0388ae | 590 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c | 591 | } |
bb5b7c11 | 592 | if (likely(sysctl_tcp_window_scaling)) { |
33ad798c | 593 | opts->ws = tp->rx_opt.rcv_wscale; |
89e95a61 | 594 | opts->options |= OPTION_WSCALE; |
bd0388ae | 595 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 596 | } |
bb5b7c11 | 597 | if (likely(sysctl_tcp_sack)) { |
33ad798c | 598 | opts->options |= OPTION_SACK_ADVERTISE; |
b32d1310 | 599 | if (unlikely(!(OPTION_TS & opts->options))) |
bd0388ae | 600 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c AL |
601 | } |
602 | ||
bd0388ae WAS |
603 | /* Note that timestamps are required by the specification. |
604 | * | |
605 | * Odd numbers of bytes are prohibited by the specification, ensuring | |
606 | * that the cookie is 16-bit aligned, and the resulting cookie pair is | |
607 | * 32-bit aligned. | |
608 | */ | |
609 | if (*md5 == NULL && | |
610 | (OPTION_TS & opts->options) && | |
611 | cookie_size > 0) { | |
612 | int need = TCPOLEN_COOKIE_BASE + cookie_size; | |
613 | ||
614 | if (0x2 & need) { | |
615 | /* 32-bit multiple */ | |
616 | need += 2; /* NOPs */ | |
617 | ||
618 | if (need > remaining) { | |
619 | /* try shrinking cookie to fit */ | |
620 | cookie_size -= 2; | |
621 | need -= 4; | |
622 | } | |
623 | } | |
624 | while (need > remaining && TCP_COOKIE_MIN <= cookie_size) { | |
625 | cookie_size -= 4; | |
626 | need -= 4; | |
627 | } | |
628 | if (TCP_COOKIE_MIN <= cookie_size) { | |
629 | opts->options |= OPTION_COOKIE_EXTENSION; | |
630 | opts->hash_location = (__u8 *)&cvp->cookie_pair[0]; | |
631 | opts->hash_size = cookie_size; | |
632 | ||
633 | /* Remember for future incarnations. */ | |
634 | cvp->cookie_desired = cookie_size; | |
635 | ||
636 | if (cvp->cookie_desired != cvp->cookie_pair_size) { | |
637 | /* Currently use random bytes as a nonce, | |
638 | * assuming these are completely unpredictable | |
639 | * by hostile users of the same system. | |
640 | */ | |
641 | get_random_bytes(&cvp->cookie_pair[0], | |
642 | cookie_size); | |
643 | cvp->cookie_pair_size = cookie_size; | |
644 | } | |
645 | ||
646 | remaining -= need; | |
647 | } | |
648 | } | |
649 | return MAX_TCP_OPTION_SPACE - remaining; | |
40efc6fa SH |
650 | } |
651 | ||
67edfef7 | 652 | /* Set up TCP options for SYN-ACKs. */ |
33ad798c AL |
653 | static unsigned tcp_synack_options(struct sock *sk, |
654 | struct request_sock *req, | |
655 | unsigned mss, struct sk_buff *skb, | |
656 | struct tcp_out_options *opts, | |
4957faad WAS |
657 | struct tcp_md5sig_key **md5, |
658 | struct tcp_extend_values *xvp) | |
659 | { | |
33ad798c | 660 | struct inet_request_sock *ireq = inet_rsk(req); |
4957faad WAS |
661 | unsigned remaining = MAX_TCP_OPTION_SPACE; |
662 | u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ? | |
663 | xvp->cookie_plus : | |
664 | 0; | |
33ad798c | 665 | |
cfb6eeb4 | 666 | #ifdef CONFIG_TCP_MD5SIG |
33ad798c AL |
667 | *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); |
668 | if (*md5) { | |
669 | opts->options |= OPTION_MD5; | |
4957faad WAS |
670 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
671 | ||
672 | /* We can't fit any SACK blocks in a packet with MD5 + TS | |
673 | * options. There was discussion about disabling SACK | |
674 | * rather than TS in order to fit in better with old, | |
675 | * buggy kernels, but that was deemed to be unnecessary. | |
676 | */ | |
de213e5e | 677 | ireq->tstamp_ok &= !ireq->sack_ok; |
cfb6eeb4 | 678 | } |
33ad798c AL |
679 | #else |
680 | *md5 = NULL; | |
cfb6eeb4 | 681 | #endif |
33ad798c | 682 | |
4957faad | 683 | /* We always send an MSS option. */ |
33ad798c | 684 | opts->mss = mss; |
4957faad | 685 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c AL |
686 | |
687 | if (likely(ireq->wscale_ok)) { | |
688 | opts->ws = ireq->rcv_wscale; | |
89e95a61 | 689 | opts->options |= OPTION_WSCALE; |
4957faad | 690 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 691 | } |
de213e5e | 692 | if (likely(ireq->tstamp_ok)) { |
33ad798c AL |
693 | opts->options |= OPTION_TS; |
694 | opts->tsval = TCP_SKB_CB(skb)->when; | |
695 | opts->tsecr = req->ts_recent; | |
4957faad | 696 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c AL |
697 | } |
698 | if (likely(ireq->sack_ok)) { | |
699 | opts->options |= OPTION_SACK_ADVERTISE; | |
de213e5e | 700 | if (unlikely(!ireq->tstamp_ok)) |
4957faad | 701 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c AL |
702 | } |
703 | ||
4957faad WAS |
704 | /* Similar rationale to tcp_syn_options() applies here, too. |
705 | * If the <SYN> options fit, the same options should fit now! | |
706 | */ | |
707 | if (*md5 == NULL && | |
de213e5e | 708 | ireq->tstamp_ok && |
4957faad WAS |
709 | cookie_plus > TCPOLEN_COOKIE_BASE) { |
710 | int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */ | |
711 | ||
712 | if (0x2 & need) { | |
713 | /* 32-bit multiple */ | |
714 | need += 2; /* NOPs */ | |
715 | } | |
716 | if (need <= remaining) { | |
717 | opts->options |= OPTION_COOKIE_EXTENSION; | |
718 | opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE; | |
719 | remaining -= need; | |
720 | } else { | |
721 | /* There's no error return, so flag it. */ | |
722 | xvp->cookie_out_never = 1; /* true */ | |
723 | opts->hash_size = 0; | |
724 | } | |
725 | } | |
726 | return MAX_TCP_OPTION_SPACE - remaining; | |
33ad798c AL |
727 | } |
728 | ||
67edfef7 AK |
729 | /* Compute TCP options for ESTABLISHED sockets. This is not the |
730 | * final wire format yet. | |
731 | */ | |
33ad798c AL |
732 | static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb, |
733 | struct tcp_out_options *opts, | |
734 | struct tcp_md5sig_key **md5) { | |
735 | struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; | |
736 | struct tcp_sock *tp = tcp_sk(sk); | |
737 | unsigned size = 0; | |
cabeccbd | 738 | unsigned int eff_sacks; |
33ad798c AL |
739 | |
740 | #ifdef CONFIG_TCP_MD5SIG | |
741 | *md5 = tp->af_specific->md5_lookup(sk, sk); | |
742 | if (unlikely(*md5)) { | |
743 | opts->options |= OPTION_MD5; | |
744 | size += TCPOLEN_MD5SIG_ALIGNED; | |
745 | } | |
746 | #else | |
747 | *md5 = NULL; | |
748 | #endif | |
749 | ||
750 | if (likely(tp->rx_opt.tstamp_ok)) { | |
751 | opts->options |= OPTION_TS; | |
752 | opts->tsval = tcb ? tcb->when : 0; | |
753 | opts->tsecr = tp->rx_opt.ts_recent; | |
754 | size += TCPOLEN_TSTAMP_ALIGNED; | |
755 | } | |
756 | ||
cabeccbd IJ |
757 | eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; |
758 | if (unlikely(eff_sacks)) { | |
33ad798c AL |
759 | const unsigned remaining = MAX_TCP_OPTION_SPACE - size; |
760 | opts->num_sack_blocks = | |
cabeccbd | 761 | min_t(unsigned, eff_sacks, |
33ad798c AL |
762 | (remaining - TCPOLEN_SACK_BASE_ALIGNED) / |
763 | TCPOLEN_SACK_PERBLOCK); | |
764 | size += TCPOLEN_SACK_BASE_ALIGNED + | |
765 | opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; | |
766 | } | |
767 | ||
768 | return size; | |
40efc6fa | 769 | } |
1da177e4 LT |
770 | |
771 | /* This routine actually transmits TCP packets queued in by | |
772 | * tcp_do_sendmsg(). This is used by both the initial | |
773 | * transmission and possible later retransmissions. | |
774 | * All SKB's seen here are completely headerless. It is our | |
775 | * job to build the TCP header, and pass the packet down to | |
776 | * IP so it can do the same plus pass the packet off to the | |
777 | * device. | |
778 | * | |
779 | * We are working here with either a clone of the original | |
780 | * SKB, or a fresh unique copy made by the retransmit engine. | |
781 | */ | |
056834d9 IJ |
782 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, |
783 | gfp_t gfp_mask) | |
1da177e4 | 784 | { |
dfb4b9dc DM |
785 | const struct inet_connection_sock *icsk = inet_csk(sk); |
786 | struct inet_sock *inet; | |
787 | struct tcp_sock *tp; | |
788 | struct tcp_skb_cb *tcb; | |
33ad798c AL |
789 | struct tcp_out_options opts; |
790 | unsigned tcp_options_size, tcp_header_size; | |
cfb6eeb4 | 791 | struct tcp_md5sig_key *md5; |
dfb4b9dc | 792 | struct tcphdr *th; |
dfb4b9dc DM |
793 | int err; |
794 | ||
795 | BUG_ON(!skb || !tcp_skb_pcount(skb)); | |
796 | ||
797 | /* If congestion control is doing timestamping, we must | |
798 | * take such a timestamp before we potentially clone/copy. | |
799 | */ | |
164891aa | 800 | if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) |
dfb4b9dc DM |
801 | __net_timestamp(skb); |
802 | ||
803 | if (likely(clone_it)) { | |
804 | if (unlikely(skb_cloned(skb))) | |
805 | skb = pskb_copy(skb, gfp_mask); | |
806 | else | |
807 | skb = skb_clone(skb, gfp_mask); | |
808 | if (unlikely(!skb)) | |
809 | return -ENOBUFS; | |
810 | } | |
1da177e4 | 811 | |
dfb4b9dc DM |
812 | inet = inet_sk(sk); |
813 | tp = tcp_sk(sk); | |
814 | tcb = TCP_SKB_CB(skb); | |
33ad798c | 815 | memset(&opts, 0, sizeof(opts)); |
1da177e4 | 816 | |
a3433f35 | 817 | if (unlikely(tcb->flags & TCPHDR_SYN)) |
33ad798c AL |
818 | tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); |
819 | else | |
820 | tcp_options_size = tcp_established_options(sk, skb, &opts, | |
821 | &md5); | |
822 | tcp_header_size = tcp_options_size + sizeof(struct tcphdr); | |
e905a9ed | 823 | |
dfb4b9dc DM |
824 | if (tcp_packets_in_flight(tp) == 0) |
825 | tcp_ca_event(sk, CA_EVENT_TX_START); | |
826 | ||
aa8223c7 ACM |
827 | skb_push(skb, tcp_header_size); |
828 | skb_reset_transport_header(skb); | |
e89862f4 | 829 | skb_set_owner_w(skb, sk); |
dfb4b9dc DM |
830 | |
831 | /* Build TCP header and checksum it. */ | |
aa8223c7 | 832 | th = tcp_hdr(skb); |
c720c7e8 ED |
833 | th->source = inet->inet_sport; |
834 | th->dest = inet->inet_dport; | |
dfb4b9dc DM |
835 | th->seq = htonl(tcb->seq); |
836 | th->ack_seq = htonl(tp->rcv_nxt); | |
df7a3b07 | 837 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
dfb4b9dc DM |
838 | tcb->flags); |
839 | ||
a3433f35 | 840 | if (unlikely(tcb->flags & TCPHDR_SYN)) { |
dfb4b9dc DM |
841 | /* RFC1323: The window in SYN & SYN/ACK segments |
842 | * is never scaled. | |
843 | */ | |
600ff0c2 | 844 | th->window = htons(min(tp->rcv_wnd, 65535U)); |
dfb4b9dc DM |
845 | } else { |
846 | th->window = htons(tcp_select_window(sk)); | |
847 | } | |
848 | th->check = 0; | |
849 | th->urg_ptr = 0; | |
1da177e4 | 850 | |
33f5f57e | 851 | /* The urg_mode check is necessary during a below snd_una win probe */ |
7691367d HX |
852 | if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { |
853 | if (before(tp->snd_up, tcb->seq + 0x10000)) { | |
854 | th->urg_ptr = htons(tp->snd_up - tcb->seq); | |
855 | th->urg = 1; | |
856 | } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { | |
0eae88f3 | 857 | th->urg_ptr = htons(0xFFFF); |
7691367d HX |
858 | th->urg = 1; |
859 | } | |
dfb4b9dc | 860 | } |
1da177e4 | 861 | |
bd0388ae | 862 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
a3433f35 | 863 | if (likely((tcb->flags & TCPHDR_SYN) == 0)) |
9e412ba7 | 864 | TCP_ECN_send(sk, skb, tcp_header_size); |
1da177e4 | 865 | |
cfb6eeb4 YH |
866 | #ifdef CONFIG_TCP_MD5SIG |
867 | /* Calculate the MD5 hash, as we have all we need now */ | |
868 | if (md5) { | |
a465419b | 869 | sk_nocaps_add(sk, NETIF_F_GSO_MASK); |
bd0388ae | 870 | tp->af_specific->calc_md5_hash(opts.hash_location, |
49a72dfb | 871 | md5, sk, NULL, skb); |
cfb6eeb4 YH |
872 | } |
873 | #endif | |
874 | ||
bb296246 | 875 | icsk->icsk_af_ops->send_check(sk, skb); |
1da177e4 | 876 | |
a3433f35 | 877 | if (likely(tcb->flags & TCPHDR_ACK)) |
dfb4b9dc | 878 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); |
1da177e4 | 879 | |
dfb4b9dc DM |
880 | if (skb->len != tcp_header_size) |
881 | tcp_event_data_sent(tp, skb, sk); | |
1da177e4 | 882 | |
bd37a088 | 883 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
aa2ea058 TH |
884 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, |
885 | tcp_skb_pcount(skb)); | |
1da177e4 | 886 | |
4e15ed4d | 887 | err = icsk->icsk_af_ops->queue_xmit(skb); |
83de47cd | 888 | if (likely(err <= 0)) |
dfb4b9dc DM |
889 | return err; |
890 | ||
3cfe3baa | 891 | tcp_enter_cwr(sk, 1); |
dfb4b9dc | 892 | |
b9df3cb8 | 893 | return net_xmit_eval(err); |
1da177e4 LT |
894 | } |
895 | ||
67edfef7 | 896 | /* This routine just queues the buffer for sending. |
1da177e4 LT |
897 | * |
898 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, | |
899 | * otherwise socket can stall. | |
900 | */ | |
901 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) | |
902 | { | |
903 | struct tcp_sock *tp = tcp_sk(sk); | |
904 | ||
905 | /* Advance write_seq and place onto the write_queue. */ | |
906 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; | |
907 | skb_header_release(skb); | |
fe067e8a | 908 | tcp_add_write_queue_tail(sk, skb); |
3ab224be HA |
909 | sk->sk_wmem_queued += skb->truesize; |
910 | sk_mem_charge(sk, skb->truesize); | |
1da177e4 LT |
911 | } |
912 | ||
67edfef7 | 913 | /* Initialize TSO segments for a packet. */ |
056834d9 IJ |
914 | static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, |
915 | unsigned int mss_now) | |
f6302d1d | 916 | { |
8e5b9dda HX |
917 | if (skb->len <= mss_now || !sk_can_gso(sk) || |
918 | skb->ip_summed == CHECKSUM_NONE) { | |
f6302d1d DM |
919 | /* Avoid the costly divide in the normal |
920 | * non-TSO case. | |
921 | */ | |
7967168c HX |
922 | skb_shinfo(skb)->gso_segs = 1; |
923 | skb_shinfo(skb)->gso_size = 0; | |
924 | skb_shinfo(skb)->gso_type = 0; | |
f6302d1d | 925 | } else { |
356f89e1 | 926 | skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); |
7967168c | 927 | skb_shinfo(skb)->gso_size = mss_now; |
bcd76111 | 928 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
1da177e4 LT |
929 | } |
930 | } | |
931 | ||
91fed7a1 | 932 | /* When a modification to fackets out becomes necessary, we need to check |
68f8353b | 933 | * skb is counted to fackets_out or not. |
91fed7a1 | 934 | */ |
a47e5a98 | 935 | static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb, |
91fed7a1 IJ |
936 | int decr) |
937 | { | |
a47e5a98 IJ |
938 | struct tcp_sock *tp = tcp_sk(sk); |
939 | ||
dc86967b | 940 | if (!tp->sacked_out || tcp_is_reno(tp)) |
91fed7a1 IJ |
941 | return; |
942 | ||
6859d494 | 943 | if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) |
91fed7a1 | 944 | tp->fackets_out -= decr; |
91fed7a1 IJ |
945 | } |
946 | ||
797108d1 IJ |
947 | /* Pcount in the middle of the write queue got changed, we need to do various |
948 | * tweaks to fix counters | |
949 | */ | |
950 | static void tcp_adjust_pcount(struct sock *sk, struct sk_buff *skb, int decr) | |
951 | { | |
952 | struct tcp_sock *tp = tcp_sk(sk); | |
953 | ||
954 | tp->packets_out -= decr; | |
955 | ||
956 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
957 | tp->sacked_out -= decr; | |
958 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
959 | tp->retrans_out -= decr; | |
960 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) | |
961 | tp->lost_out -= decr; | |
962 | ||
963 | /* Reno case is special. Sigh... */ | |
964 | if (tcp_is_reno(tp) && decr > 0) | |
965 | tp->sacked_out -= min_t(u32, tp->sacked_out, decr); | |
966 | ||
967 | tcp_adjust_fackets_out(sk, skb, decr); | |
968 | ||
969 | if (tp->lost_skb_hint && | |
970 | before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && | |
52cf3cc8 | 971 | (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) |
797108d1 IJ |
972 | tp->lost_cnt_hint -= decr; |
973 | ||
974 | tcp_verify_left_out(tp); | |
975 | } | |
976 | ||
1da177e4 LT |
977 | /* Function to create two new TCP segments. Shrinks the given segment |
978 | * to the specified size and appends a new segment with the rest of the | |
e905a9ed | 979 | * packet to the list. This won't be called frequently, I hope. |
1da177e4 LT |
980 | * Remember, these are still headerless SKBs at this point. |
981 | */ | |
056834d9 IJ |
982 | int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, |
983 | unsigned int mss_now) | |
1da177e4 LT |
984 | { |
985 | struct tcp_sock *tp = tcp_sk(sk); | |
986 | struct sk_buff *buff; | |
6475be16 | 987 | int nsize, old_factor; |
b60b49ea | 988 | int nlen; |
9ce01461 | 989 | u8 flags; |
1da177e4 | 990 | |
b2cc99f0 | 991 | BUG_ON(len > skb->len); |
6a438bbe | 992 | |
1da177e4 LT |
993 | nsize = skb_headlen(skb) - len; |
994 | if (nsize < 0) | |
995 | nsize = 0; | |
996 | ||
997 | if (skb_cloned(skb) && | |
998 | skb_is_nonlinear(skb) && | |
999 | pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) | |
1000 | return -ENOMEM; | |
1001 | ||
1002 | /* Get a new skb... force flag on. */ | |
1003 | buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); | |
1004 | if (buff == NULL) | |
1005 | return -ENOMEM; /* We'll just try again later. */ | |
ef5cb973 | 1006 | |
3ab224be HA |
1007 | sk->sk_wmem_queued += buff->truesize; |
1008 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea HX |
1009 | nlen = skb->len - len - nsize; |
1010 | buff->truesize += nlen; | |
1011 | skb->truesize -= nlen; | |
1da177e4 LT |
1012 | |
1013 | /* Correct the sequence numbers. */ | |
1014 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1015 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1016 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1017 | ||
1018 | /* PSH and FIN should only be set in the second packet. */ | |
1019 | flags = TCP_SKB_CB(skb)->flags; | |
a3433f35 | 1020 | TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); |
1da177e4 | 1021 | TCP_SKB_CB(buff)->flags = flags; |
e14c3caf | 1022 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
1da177e4 | 1023 | |
84fa7933 | 1024 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { |
1da177e4 | 1025 | /* Copy and checksum data tail into the new buffer. */ |
056834d9 IJ |
1026 | buff->csum = csum_partial_copy_nocheck(skb->data + len, |
1027 | skb_put(buff, nsize), | |
1da177e4 LT |
1028 | nsize, 0); |
1029 | ||
1030 | skb_trim(skb, len); | |
1031 | ||
1032 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); | |
1033 | } else { | |
84fa7933 | 1034 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
1035 | skb_split(skb, buff, len); |
1036 | } | |
1037 | ||
1038 | buff->ip_summed = skb->ip_summed; | |
1039 | ||
1040 | /* Looks stupid, but our code really uses when of | |
1041 | * skbs, which it never sent before. --ANK | |
1042 | */ | |
1043 | TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; | |
a61bbcf2 | 1044 | buff->tstamp = skb->tstamp; |
1da177e4 | 1045 | |
6475be16 DM |
1046 | old_factor = tcp_skb_pcount(skb); |
1047 | ||
1da177e4 | 1048 | /* Fix up tso_factor for both original and new SKB. */ |
846998ae DM |
1049 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
1050 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
1da177e4 | 1051 | |
6475be16 DM |
1052 | /* If this packet has been sent out already, we must |
1053 | * adjust the various packet counters. | |
1054 | */ | |
cf0b450c | 1055 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
6475be16 DM |
1056 | int diff = old_factor - tcp_skb_pcount(skb) - |
1057 | tcp_skb_pcount(buff); | |
1da177e4 | 1058 | |
797108d1 IJ |
1059 | if (diff) |
1060 | tcp_adjust_pcount(sk, skb, diff); | |
1da177e4 LT |
1061 | } |
1062 | ||
1063 | /* Link BUFF into the send queue. */ | |
f44b5271 | 1064 | skb_header_release(buff); |
fe067e8a | 1065 | tcp_insert_write_queue_after(skb, buff, sk); |
1da177e4 LT |
1066 | |
1067 | return 0; | |
1068 | } | |
1069 | ||
1070 | /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c | |
1071 | * eventually). The difference is that pulled data not copied, but | |
1072 | * immediately discarded. | |
1073 | */ | |
f2911969 | 1074 | static void __pskb_trim_head(struct sk_buff *skb, int len) |
1da177e4 LT |
1075 | { |
1076 | int i, k, eat; | |
1077 | ||
1078 | eat = len; | |
1079 | k = 0; | |
056834d9 | 1080 | for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { |
1da177e4 LT |
1081 | if (skb_shinfo(skb)->frags[i].size <= eat) { |
1082 | put_page(skb_shinfo(skb)->frags[i].page); | |
1083 | eat -= skb_shinfo(skb)->frags[i].size; | |
1084 | } else { | |
1085 | skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; | |
1086 | if (eat) { | |
1087 | skb_shinfo(skb)->frags[k].page_offset += eat; | |
1088 | skb_shinfo(skb)->frags[k].size -= eat; | |
1089 | eat = 0; | |
1090 | } | |
1091 | k++; | |
1092 | } | |
1093 | } | |
1094 | skb_shinfo(skb)->nr_frags = k; | |
1095 | ||
27a884dc | 1096 | skb_reset_tail_pointer(skb); |
1da177e4 LT |
1097 | skb->data_len -= len; |
1098 | skb->len = skb->data_len; | |
1da177e4 LT |
1099 | } |
1100 | ||
67edfef7 | 1101 | /* Remove acked data from a packet in the transmit queue. */ |
1da177e4 LT |
1102 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) |
1103 | { | |
056834d9 | 1104 | if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) |
1da177e4 LT |
1105 | return -ENOMEM; |
1106 | ||
f2911969 HX |
1107 | /* If len == headlen, we avoid __skb_pull to preserve alignment. */ |
1108 | if (unlikely(len < skb_headlen(skb))) | |
1da177e4 | 1109 | __skb_pull(skb, len); |
f2911969 HX |
1110 | else |
1111 | __pskb_trim_head(skb, len - skb_headlen(skb)); | |
1da177e4 LT |
1112 | |
1113 | TCP_SKB_CB(skb)->seq += len; | |
84fa7933 | 1114 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
1115 | |
1116 | skb->truesize -= len; | |
1117 | sk->sk_wmem_queued -= len; | |
3ab224be | 1118 | sk_mem_uncharge(sk, len); |
1da177e4 LT |
1119 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); |
1120 | ||
1121 | /* Any change of skb->len requires recalculation of tso | |
1122 | * factor and mss. | |
1123 | */ | |
1124 | if (tcp_skb_pcount(skb) > 1) | |
0c54b85f | 1125 | tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk)); |
1da177e4 LT |
1126 | |
1127 | return 0; | |
1128 | } | |
1129 | ||
67edfef7 | 1130 | /* Calculate MSS. Not accounting for SACKs here. */ |
5d424d5a JH |
1131 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) |
1132 | { | |
1133 | struct tcp_sock *tp = tcp_sk(sk); | |
1134 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1135 | int mss_now; | |
1136 | ||
1137 | /* Calculate base mss without TCP options: | |
1138 | It is MMS_S - sizeof(tcphdr) of rfc1122 | |
1139 | */ | |
1140 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); | |
1141 | ||
1142 | /* Clamp it (mss_clamp does not include tcp options) */ | |
1143 | if (mss_now > tp->rx_opt.mss_clamp) | |
1144 | mss_now = tp->rx_opt.mss_clamp; | |
1145 | ||
1146 | /* Now subtract optional transport overhead */ | |
1147 | mss_now -= icsk->icsk_ext_hdr_len; | |
1148 | ||
1149 | /* Then reserve room for full set of TCP options and 8 bytes of data */ | |
1150 | if (mss_now < 48) | |
1151 | mss_now = 48; | |
1152 | ||
1153 | /* Now subtract TCP options size, not including SACKs */ | |
1154 | mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); | |
1155 | ||
1156 | return mss_now; | |
1157 | } | |
1158 | ||
1159 | /* Inverse of above */ | |
1160 | int tcp_mss_to_mtu(struct sock *sk, int mss) | |
1161 | { | |
1162 | struct tcp_sock *tp = tcp_sk(sk); | |
1163 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1164 | int mtu; | |
1165 | ||
1166 | mtu = mss + | |
1167 | tp->tcp_header_len + | |
1168 | icsk->icsk_ext_hdr_len + | |
1169 | icsk->icsk_af_ops->net_header_len; | |
1170 | ||
1171 | return mtu; | |
1172 | } | |
1173 | ||
67edfef7 | 1174 | /* MTU probing init per socket */ |
5d424d5a JH |
1175 | void tcp_mtup_init(struct sock *sk) |
1176 | { | |
1177 | struct tcp_sock *tp = tcp_sk(sk); | |
1178 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1179 | ||
1180 | icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; | |
1181 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + | |
e905a9ed | 1182 | icsk->icsk_af_ops->net_header_len; |
5d424d5a JH |
1183 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); |
1184 | icsk->icsk_mtup.probe_size = 0; | |
1185 | } | |
4bc2f18b | 1186 | EXPORT_SYMBOL(tcp_mtup_init); |
5d424d5a | 1187 | |
1da177e4 LT |
1188 | /* This function synchronize snd mss to current pmtu/exthdr set. |
1189 | ||
1190 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts | |
1191 | for TCP options, but includes only bare TCP header. | |
1192 | ||
1193 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. | |
caa20d9a | 1194 | It is minimum of user_mss and mss received with SYN. |
1da177e4 LT |
1195 | It also does not include TCP options. |
1196 | ||
d83d8461 | 1197 | inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. |
1da177e4 LT |
1198 | |
1199 | tp->mss_cache is current effective sending mss, including | |
1200 | all tcp options except for SACKs. It is evaluated, | |
1201 | taking into account current pmtu, but never exceeds | |
1202 | tp->rx_opt.mss_clamp. | |
1203 | ||
1204 | NOTE1. rfc1122 clearly states that advertised MSS | |
1205 | DOES NOT include either tcp or ip options. | |
1206 | ||
d83d8461 ACM |
1207 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
1208 | are READ ONLY outside this function. --ANK (980731) | |
1da177e4 | 1209 | */ |
1da177e4 LT |
1210 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) |
1211 | { | |
1212 | struct tcp_sock *tp = tcp_sk(sk); | |
d83d8461 | 1213 | struct inet_connection_sock *icsk = inet_csk(sk); |
5d424d5a | 1214 | int mss_now; |
1da177e4 | 1215 | |
5d424d5a JH |
1216 | if (icsk->icsk_mtup.search_high > pmtu) |
1217 | icsk->icsk_mtup.search_high = pmtu; | |
1da177e4 | 1218 | |
5d424d5a | 1219 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
409d22b4 | 1220 | mss_now = tcp_bound_to_half_wnd(tp, mss_now); |
1da177e4 LT |
1221 | |
1222 | /* And store cached results */ | |
d83d8461 | 1223 | icsk->icsk_pmtu_cookie = pmtu; |
5d424d5a JH |
1224 | if (icsk->icsk_mtup.enabled) |
1225 | mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); | |
c1b4a7e6 | 1226 | tp->mss_cache = mss_now; |
1da177e4 LT |
1227 | |
1228 | return mss_now; | |
1229 | } | |
4bc2f18b | 1230 | EXPORT_SYMBOL(tcp_sync_mss); |
1da177e4 LT |
1231 | |
1232 | /* Compute the current effective MSS, taking SACKs and IP options, | |
1233 | * and even PMTU discovery events into account. | |
1da177e4 | 1234 | */ |
0c54b85f | 1235 | unsigned int tcp_current_mss(struct sock *sk) |
1da177e4 LT |
1236 | { |
1237 | struct tcp_sock *tp = tcp_sk(sk); | |
1238 | struct dst_entry *dst = __sk_dst_get(sk); | |
c1b4a7e6 | 1239 | u32 mss_now; |
33ad798c AL |
1240 | unsigned header_len; |
1241 | struct tcp_out_options opts; | |
1242 | struct tcp_md5sig_key *md5; | |
c1b4a7e6 DM |
1243 | |
1244 | mss_now = tp->mss_cache; | |
1245 | ||
1da177e4 LT |
1246 | if (dst) { |
1247 | u32 mtu = dst_mtu(dst); | |
d83d8461 | 1248 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
1da177e4 LT |
1249 | mss_now = tcp_sync_mss(sk, mtu); |
1250 | } | |
1251 | ||
33ad798c AL |
1252 | header_len = tcp_established_options(sk, NULL, &opts, &md5) + |
1253 | sizeof(struct tcphdr); | |
1254 | /* The mss_cache is sized based on tp->tcp_header_len, which assumes | |
1255 | * some common options. If this is an odd packet (because we have SACK | |
1256 | * blocks etc) then our calculated header_len will be different, and | |
1257 | * we have to adjust mss_now correspondingly */ | |
1258 | if (header_len != tp->tcp_header_len) { | |
1259 | int delta = (int) header_len - tp->tcp_header_len; | |
1260 | mss_now -= delta; | |
1261 | } | |
cfb6eeb4 | 1262 | |
1da177e4 LT |
1263 | return mss_now; |
1264 | } | |
1265 | ||
a762a980 | 1266 | /* Congestion window validation. (RFC2861) */ |
9e412ba7 | 1267 | static void tcp_cwnd_validate(struct sock *sk) |
a762a980 | 1268 | { |
9e412ba7 | 1269 | struct tcp_sock *tp = tcp_sk(sk); |
a762a980 | 1270 | |
d436d686 | 1271 | if (tp->packets_out >= tp->snd_cwnd) { |
a762a980 DM |
1272 | /* Network is feed fully. */ |
1273 | tp->snd_cwnd_used = 0; | |
1274 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1275 | } else { | |
1276 | /* Network starves. */ | |
1277 | if (tp->packets_out > tp->snd_cwnd_used) | |
1278 | tp->snd_cwnd_used = tp->packets_out; | |
1279 | ||
15d33c07 DM |
1280 | if (sysctl_tcp_slow_start_after_idle && |
1281 | (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) | |
a762a980 DM |
1282 | tcp_cwnd_application_limited(sk); |
1283 | } | |
1284 | } | |
1285 | ||
0e3a4803 IJ |
1286 | /* Returns the portion of skb which can be sent right away without |
1287 | * introducing MSS oddities to segment boundaries. In rare cases where | |
1288 | * mss_now != mss_cache, we will request caller to create a small skb | |
1289 | * per input skb which could be mostly avoided here (if desired). | |
5ea3a748 IJ |
1290 | * |
1291 | * We explicitly want to create a request for splitting write queue tail | |
1292 | * to a small skb for Nagle purposes while avoiding unnecessary modulos, | |
1293 | * thus all the complexity (cwnd_len is always MSS multiple which we | |
1294 | * return whenever allowed by the other factors). Basically we need the | |
1295 | * modulo only when the receiver window alone is the limiting factor or | |
1296 | * when we would be allowed to send the split-due-to-Nagle skb fully. | |
0e3a4803 IJ |
1297 | */ |
1298 | static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb, | |
056834d9 | 1299 | unsigned int mss_now, unsigned int cwnd) |
c1b4a7e6 | 1300 | { |
0e3a4803 IJ |
1301 | struct tcp_sock *tp = tcp_sk(sk); |
1302 | u32 needed, window, cwnd_len; | |
c1b4a7e6 | 1303 | |
90840def | 1304 | window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
c1b4a7e6 | 1305 | cwnd_len = mss_now * cwnd; |
0e3a4803 IJ |
1306 | |
1307 | if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk))) | |
1308 | return cwnd_len; | |
1309 | ||
5ea3a748 IJ |
1310 | needed = min(skb->len, window); |
1311 | ||
17515408 | 1312 | if (cwnd_len <= needed) |
0e3a4803 IJ |
1313 | return cwnd_len; |
1314 | ||
0e3a4803 | 1315 | return needed - needed % mss_now; |
c1b4a7e6 DM |
1316 | } |
1317 | ||
1318 | /* Can at least one segment of SKB be sent right now, according to the | |
1319 | * congestion window rules? If so, return how many segments are allowed. | |
1320 | */ | |
056834d9 IJ |
1321 | static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, |
1322 | struct sk_buff *skb) | |
c1b4a7e6 DM |
1323 | { |
1324 | u32 in_flight, cwnd; | |
1325 | ||
1326 | /* Don't be strict about the congestion window for the final FIN. */ | |
a3433f35 | 1327 | if ((TCP_SKB_CB(skb)->flags & TCPHDR_FIN) && tcp_skb_pcount(skb) == 1) |
c1b4a7e6 DM |
1328 | return 1; |
1329 | ||
1330 | in_flight = tcp_packets_in_flight(tp); | |
1331 | cwnd = tp->snd_cwnd; | |
1332 | if (in_flight < cwnd) | |
1333 | return (cwnd - in_flight); | |
1334 | ||
1335 | return 0; | |
1336 | } | |
1337 | ||
67edfef7 AK |
1338 | /* Intialize TSO state of a skb. |
1339 | * This must be invoked the first time we consider transmitting | |
c1b4a7e6 DM |
1340 | * SKB onto the wire. |
1341 | */ | |
056834d9 IJ |
1342 | static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, |
1343 | unsigned int mss_now) | |
c1b4a7e6 DM |
1344 | { |
1345 | int tso_segs = tcp_skb_pcount(skb); | |
1346 | ||
f8269a49 | 1347 | if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { |
846998ae | 1348 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
c1b4a7e6 DM |
1349 | tso_segs = tcp_skb_pcount(skb); |
1350 | } | |
1351 | return tso_segs; | |
1352 | } | |
1353 | ||
67edfef7 | 1354 | /* Minshall's variant of the Nagle send check. */ |
c1b4a7e6 DM |
1355 | static inline int tcp_minshall_check(const struct tcp_sock *tp) |
1356 | { | |
09cb105e | 1357 | return after(tp->snd_sml, tp->snd_una) && |
c1b4a7e6 DM |
1358 | !after(tp->snd_sml, tp->snd_nxt); |
1359 | } | |
1360 | ||
1361 | /* Return 0, if packet can be sent now without violation Nagle's rules: | |
1362 | * 1. It is full sized. | |
1363 | * 2. Or it contains FIN. (already checked by caller) | |
1364 | * 3. Or TCP_NODELAY was set. | |
1365 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. | |
1366 | * With Minshall's modification: all sent small packets are ACKed. | |
1367 | */ | |
c1b4a7e6 | 1368 | static inline int tcp_nagle_check(const struct tcp_sock *tp, |
e905a9ed | 1369 | const struct sk_buff *skb, |
c1b4a7e6 DM |
1370 | unsigned mss_now, int nonagle) |
1371 | { | |
a02cec21 | 1372 | return skb->len < mss_now && |
056834d9 | 1373 | ((nonagle & TCP_NAGLE_CORK) || |
a02cec21 | 1374 | (!nonagle && tp->packets_out && tcp_minshall_check(tp))); |
c1b4a7e6 DM |
1375 | } |
1376 | ||
1377 | /* Return non-zero if the Nagle test allows this packet to be | |
1378 | * sent now. | |
1379 | */ | |
1380 | static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, | |
1381 | unsigned int cur_mss, int nonagle) | |
1382 | { | |
1383 | /* Nagle rule does not apply to frames, which sit in the middle of the | |
1384 | * write_queue (they have no chances to get new data). | |
1385 | * | |
1386 | * This is implemented in the callers, where they modify the 'nonagle' | |
1387 | * argument based upon the location of SKB in the send queue. | |
1388 | */ | |
1389 | if (nonagle & TCP_NAGLE_PUSH) | |
1390 | return 1; | |
1391 | ||
d551e454 IJ |
1392 | /* Don't use the nagle rule for urgent data (or for the final FIN). |
1393 | * Nagle can be ignored during F-RTO too (see RFC4138). | |
1394 | */ | |
33f5f57e | 1395 | if (tcp_urg_mode(tp) || (tp->frto_counter == 2) || |
a3433f35 | 1396 | (TCP_SKB_CB(skb)->flags & TCPHDR_FIN)) |
c1b4a7e6 DM |
1397 | return 1; |
1398 | ||
1399 | if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) | |
1400 | return 1; | |
1401 | ||
1402 | return 0; | |
1403 | } | |
1404 | ||
1405 | /* Does at least the first segment of SKB fit into the send window? */ | |
056834d9 IJ |
1406 | static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, |
1407 | unsigned int cur_mss) | |
c1b4a7e6 DM |
1408 | { |
1409 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
1410 | ||
1411 | if (skb->len > cur_mss) | |
1412 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; | |
1413 | ||
90840def | 1414 | return !after(end_seq, tcp_wnd_end(tp)); |
c1b4a7e6 DM |
1415 | } |
1416 | ||
fe067e8a | 1417 | /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) |
c1b4a7e6 DM |
1418 | * should be put on the wire right now. If so, it returns the number of |
1419 | * packets allowed by the congestion window. | |
1420 | */ | |
1421 | static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, | |
1422 | unsigned int cur_mss, int nonagle) | |
1423 | { | |
1424 | struct tcp_sock *tp = tcp_sk(sk); | |
1425 | unsigned int cwnd_quota; | |
1426 | ||
846998ae | 1427 | tcp_init_tso_segs(sk, skb, cur_mss); |
c1b4a7e6 DM |
1428 | |
1429 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) | |
1430 | return 0; | |
1431 | ||
1432 | cwnd_quota = tcp_cwnd_test(tp, skb); | |
056834d9 | 1433 | if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) |
c1b4a7e6 DM |
1434 | cwnd_quota = 0; |
1435 | ||
1436 | return cwnd_quota; | |
1437 | } | |
1438 | ||
67edfef7 | 1439 | /* Test if sending is allowed right now. */ |
9e412ba7 | 1440 | int tcp_may_send_now(struct sock *sk) |
c1b4a7e6 | 1441 | { |
9e412ba7 | 1442 | struct tcp_sock *tp = tcp_sk(sk); |
fe067e8a | 1443 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 | 1444 | |
a02cec21 | 1445 | return skb && |
0c54b85f | 1446 | tcp_snd_test(sk, skb, tcp_current_mss(sk), |
c1b4a7e6 | 1447 | (tcp_skb_is_last(sk, skb) ? |
a02cec21 | 1448 | tp->nonagle : TCP_NAGLE_PUSH)); |
c1b4a7e6 DM |
1449 | } |
1450 | ||
1451 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet | |
1452 | * which is put after SKB on the list. It is very much like | |
1453 | * tcp_fragment() except that it may make several kinds of assumptions | |
1454 | * in order to speed up the splitting operation. In particular, we | |
1455 | * know that all the data is in scatter-gather pages, and that the | |
1456 | * packet has never been sent out before (and thus is not cloned). | |
1457 | */ | |
056834d9 | 1458 | static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, |
c4ead4c5 | 1459 | unsigned int mss_now, gfp_t gfp) |
c1b4a7e6 DM |
1460 | { |
1461 | struct sk_buff *buff; | |
1462 | int nlen = skb->len - len; | |
9ce01461 | 1463 | u8 flags; |
c1b4a7e6 DM |
1464 | |
1465 | /* All of a TSO frame must be composed of paged data. */ | |
c8ac3774 HX |
1466 | if (skb->len != skb->data_len) |
1467 | return tcp_fragment(sk, skb, len, mss_now); | |
c1b4a7e6 | 1468 | |
c4ead4c5 | 1469 | buff = sk_stream_alloc_skb(sk, 0, gfp); |
c1b4a7e6 DM |
1470 | if (unlikely(buff == NULL)) |
1471 | return -ENOMEM; | |
1472 | ||
3ab224be HA |
1473 | sk->sk_wmem_queued += buff->truesize; |
1474 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea | 1475 | buff->truesize += nlen; |
c1b4a7e6 DM |
1476 | skb->truesize -= nlen; |
1477 | ||
1478 | /* Correct the sequence numbers. */ | |
1479 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1480 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1481 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1482 | ||
1483 | /* PSH and FIN should only be set in the second packet. */ | |
1484 | flags = TCP_SKB_CB(skb)->flags; | |
a3433f35 | 1485 | TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); |
c1b4a7e6 DM |
1486 | TCP_SKB_CB(buff)->flags = flags; |
1487 | ||
1488 | /* This packet was never sent out yet, so no SACK bits. */ | |
1489 | TCP_SKB_CB(buff)->sacked = 0; | |
1490 | ||
84fa7933 | 1491 | buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; |
c1b4a7e6 DM |
1492 | skb_split(skb, buff, len); |
1493 | ||
1494 | /* Fix up tso_factor for both original and new SKB. */ | |
846998ae DM |
1495 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
1496 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
c1b4a7e6 DM |
1497 | |
1498 | /* Link BUFF into the send queue. */ | |
1499 | skb_header_release(buff); | |
fe067e8a | 1500 | tcp_insert_write_queue_after(skb, buff, sk); |
c1b4a7e6 DM |
1501 | |
1502 | return 0; | |
1503 | } | |
1504 | ||
1505 | /* Try to defer sending, if possible, in order to minimize the amount | |
1506 | * of TSO splitting we do. View it as a kind of TSO Nagle test. | |
1507 | * | |
1508 | * This algorithm is from John Heffner. | |
1509 | */ | |
9e412ba7 | 1510 | static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) |
c1b4a7e6 | 1511 | { |
9e412ba7 | 1512 | struct tcp_sock *tp = tcp_sk(sk); |
6687e988 | 1513 | const struct inet_connection_sock *icsk = inet_csk(sk); |
c1b4a7e6 DM |
1514 | u32 send_win, cong_win, limit, in_flight; |
1515 | ||
a3433f35 | 1516 | if (TCP_SKB_CB(skb)->flags & TCPHDR_FIN) |
ae8064ac | 1517 | goto send_now; |
c1b4a7e6 | 1518 | |
6687e988 | 1519 | if (icsk->icsk_ca_state != TCP_CA_Open) |
ae8064ac JH |
1520 | goto send_now; |
1521 | ||
1522 | /* Defer for less than two clock ticks. */ | |
bd515c3e | 1523 | if (tp->tso_deferred && |
a2acde07 | 1524 | (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) |
ae8064ac | 1525 | goto send_now; |
908a75c1 | 1526 | |
c1b4a7e6 DM |
1527 | in_flight = tcp_packets_in_flight(tp); |
1528 | ||
056834d9 | 1529 | BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); |
c1b4a7e6 | 1530 | |
90840def | 1531 | send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
c1b4a7e6 DM |
1532 | |
1533 | /* From in_flight test above, we know that cwnd > in_flight. */ | |
1534 | cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; | |
1535 | ||
1536 | limit = min(send_win, cong_win); | |
1537 | ||
ba244fe9 | 1538 | /* If a full-sized TSO skb can be sent, do it. */ |
82cc1a7a | 1539 | if (limit >= sk->sk_gso_max_size) |
ae8064ac | 1540 | goto send_now; |
ba244fe9 | 1541 | |
62ad2761 IJ |
1542 | /* Middle in queue won't get any more data, full sendable already? */ |
1543 | if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) | |
1544 | goto send_now; | |
1545 | ||
c1b4a7e6 DM |
1546 | if (sysctl_tcp_tso_win_divisor) { |
1547 | u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); | |
1548 | ||
1549 | /* If at least some fraction of a window is available, | |
1550 | * just use it. | |
1551 | */ | |
1552 | chunk /= sysctl_tcp_tso_win_divisor; | |
1553 | if (limit >= chunk) | |
ae8064ac | 1554 | goto send_now; |
c1b4a7e6 DM |
1555 | } else { |
1556 | /* Different approach, try not to defer past a single | |
1557 | * ACK. Receiver should ACK every other full sized | |
1558 | * frame, so if we have space for more than 3 frames | |
1559 | * then send now. | |
1560 | */ | |
1561 | if (limit > tcp_max_burst(tp) * tp->mss_cache) | |
ae8064ac | 1562 | goto send_now; |
c1b4a7e6 DM |
1563 | } |
1564 | ||
1565 | /* Ok, it looks like it is advisable to defer. */ | |
056834d9 | 1566 | tp->tso_deferred = 1 | (jiffies << 1); |
ae8064ac | 1567 | |
c1b4a7e6 | 1568 | return 1; |
ae8064ac JH |
1569 | |
1570 | send_now: | |
1571 | tp->tso_deferred = 0; | |
1572 | return 0; | |
c1b4a7e6 DM |
1573 | } |
1574 | ||
5d424d5a | 1575 | /* Create a new MTU probe if we are ready. |
67edfef7 AK |
1576 | * MTU probe is regularly attempting to increase the path MTU by |
1577 | * deliberately sending larger packets. This discovers routing | |
1578 | * changes resulting in larger path MTUs. | |
1579 | * | |
5d424d5a JH |
1580 | * Returns 0 if we should wait to probe (no cwnd available), |
1581 | * 1 if a probe was sent, | |
056834d9 IJ |
1582 | * -1 otherwise |
1583 | */ | |
5d424d5a JH |
1584 | static int tcp_mtu_probe(struct sock *sk) |
1585 | { | |
1586 | struct tcp_sock *tp = tcp_sk(sk); | |
1587 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1588 | struct sk_buff *skb, *nskb, *next; | |
1589 | int len; | |
1590 | int probe_size; | |
91cc17c0 | 1591 | int size_needed; |
5d424d5a JH |
1592 | int copy; |
1593 | int mss_now; | |
1594 | ||
1595 | /* Not currently probing/verifying, | |
1596 | * not in recovery, | |
1597 | * have enough cwnd, and | |
1598 | * not SACKing (the variable headers throw things off) */ | |
1599 | if (!icsk->icsk_mtup.enabled || | |
1600 | icsk->icsk_mtup.probe_size || | |
1601 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || | |
1602 | tp->snd_cwnd < 11 || | |
cabeccbd | 1603 | tp->rx_opt.num_sacks || tp->rx_opt.dsack) |
5d424d5a JH |
1604 | return -1; |
1605 | ||
1606 | /* Very simple search strategy: just double the MSS. */ | |
0c54b85f | 1607 | mss_now = tcp_current_mss(sk); |
056834d9 | 1608 | probe_size = 2 * tp->mss_cache; |
91cc17c0 | 1609 | size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; |
5d424d5a JH |
1610 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { |
1611 | /* TODO: set timer for probe_converge_event */ | |
1612 | return -1; | |
1613 | } | |
1614 | ||
1615 | /* Have enough data in the send queue to probe? */ | |
7f9c33e5 | 1616 | if (tp->write_seq - tp->snd_nxt < size_needed) |
5d424d5a JH |
1617 | return -1; |
1618 | ||
91cc17c0 IJ |
1619 | if (tp->snd_wnd < size_needed) |
1620 | return -1; | |
90840def | 1621 | if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) |
91cc17c0 | 1622 | return 0; |
5d424d5a | 1623 | |
d67c58e9 IJ |
1624 | /* Do we need to wait to drain cwnd? With none in flight, don't stall */ |
1625 | if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { | |
1626 | if (!tcp_packets_in_flight(tp)) | |
5d424d5a JH |
1627 | return -1; |
1628 | else | |
1629 | return 0; | |
1630 | } | |
1631 | ||
1632 | /* We're allowed to probe. Build it now. */ | |
1633 | if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) | |
1634 | return -1; | |
3ab224be HA |
1635 | sk->sk_wmem_queued += nskb->truesize; |
1636 | sk_mem_charge(sk, nskb->truesize); | |
5d424d5a | 1637 | |
fe067e8a | 1638 | skb = tcp_send_head(sk); |
5d424d5a JH |
1639 | |
1640 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; | |
1641 | TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; | |
a3433f35 | 1642 | TCP_SKB_CB(nskb)->flags = TCPHDR_ACK; |
5d424d5a JH |
1643 | TCP_SKB_CB(nskb)->sacked = 0; |
1644 | nskb->csum = 0; | |
84fa7933 | 1645 | nskb->ip_summed = skb->ip_summed; |
5d424d5a | 1646 | |
50c4817e IJ |
1647 | tcp_insert_write_queue_before(nskb, skb, sk); |
1648 | ||
5d424d5a | 1649 | len = 0; |
234b6860 | 1650 | tcp_for_write_queue_from_safe(skb, next, sk) { |
5d424d5a JH |
1651 | copy = min_t(int, skb->len, probe_size - len); |
1652 | if (nskb->ip_summed) | |
1653 | skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); | |
1654 | else | |
1655 | nskb->csum = skb_copy_and_csum_bits(skb, 0, | |
056834d9 IJ |
1656 | skb_put(nskb, copy), |
1657 | copy, nskb->csum); | |
5d424d5a JH |
1658 | |
1659 | if (skb->len <= copy) { | |
1660 | /* We've eaten all the data from this skb. | |
1661 | * Throw it away. */ | |
1662 | TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags; | |
fe067e8a | 1663 | tcp_unlink_write_queue(skb, sk); |
3ab224be | 1664 | sk_wmem_free_skb(sk, skb); |
5d424d5a JH |
1665 | } else { |
1666 | TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags & | |
a3433f35 | 1667 | ~(TCPHDR_FIN|TCPHDR_PSH); |
5d424d5a JH |
1668 | if (!skb_shinfo(skb)->nr_frags) { |
1669 | skb_pull(skb, copy); | |
84fa7933 | 1670 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
056834d9 IJ |
1671 | skb->csum = csum_partial(skb->data, |
1672 | skb->len, 0); | |
5d424d5a JH |
1673 | } else { |
1674 | __pskb_trim_head(skb, copy); | |
1675 | tcp_set_skb_tso_segs(sk, skb, mss_now); | |
1676 | } | |
1677 | TCP_SKB_CB(skb)->seq += copy; | |
1678 | } | |
1679 | ||
1680 | len += copy; | |
234b6860 IJ |
1681 | |
1682 | if (len >= probe_size) | |
1683 | break; | |
5d424d5a JH |
1684 | } |
1685 | tcp_init_tso_segs(sk, nskb, nskb->len); | |
1686 | ||
1687 | /* We're ready to send. If this fails, the probe will | |
1688 | * be resegmented into mss-sized pieces by tcp_write_xmit(). */ | |
1689 | TCP_SKB_CB(nskb)->when = tcp_time_stamp; | |
1690 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { | |
1691 | /* Decrement cwnd here because we are sending | |
056834d9 | 1692 | * effectively two packets. */ |
5d424d5a | 1693 | tp->snd_cwnd--; |
66f5fe62 | 1694 | tcp_event_new_data_sent(sk, nskb); |
5d424d5a JH |
1695 | |
1696 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); | |
0e7b1368 JH |
1697 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; |
1698 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; | |
5d424d5a JH |
1699 | |
1700 | return 1; | |
1701 | } | |
1702 | ||
1703 | return -1; | |
1704 | } | |
1705 | ||
1da177e4 LT |
1706 | /* This routine writes packets to the network. It advances the |
1707 | * send_head. This happens as incoming acks open up the remote | |
1708 | * window for us. | |
1709 | * | |
f8269a49 IJ |
1710 | * LARGESEND note: !tcp_urg_mode is overkill, only frames between |
1711 | * snd_up-64k-mss .. snd_up cannot be large. However, taking into | |
1712 | * account rare use of URG, this is not a big flaw. | |
1713 | * | |
1da177e4 LT |
1714 | * Returns 1, if no segments are in flight and we have queued segments, but |
1715 | * cannot send anything now because of SWS or another problem. | |
1716 | */ | |
d5dd9175 IJ |
1717 | static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
1718 | int push_one, gfp_t gfp) | |
1da177e4 LT |
1719 | { |
1720 | struct tcp_sock *tp = tcp_sk(sk); | |
92df7b51 | 1721 | struct sk_buff *skb; |
c1b4a7e6 DM |
1722 | unsigned int tso_segs, sent_pkts; |
1723 | int cwnd_quota; | |
5d424d5a | 1724 | int result; |
1da177e4 | 1725 | |
92df7b51 | 1726 | sent_pkts = 0; |
5d424d5a | 1727 | |
d5dd9175 IJ |
1728 | if (!push_one) { |
1729 | /* Do MTU probing. */ | |
1730 | result = tcp_mtu_probe(sk); | |
1731 | if (!result) { | |
1732 | return 0; | |
1733 | } else if (result > 0) { | |
1734 | sent_pkts = 1; | |
1735 | } | |
5d424d5a JH |
1736 | } |
1737 | ||
fe067e8a | 1738 | while ((skb = tcp_send_head(sk))) { |
c8ac3774 HX |
1739 | unsigned int limit; |
1740 | ||
b68e9f85 | 1741 | tso_segs = tcp_init_tso_segs(sk, skb, mss_now); |
c1b4a7e6 | 1742 | BUG_ON(!tso_segs); |
aa93466b | 1743 | |
b68e9f85 HX |
1744 | cwnd_quota = tcp_cwnd_test(tp, skb); |
1745 | if (!cwnd_quota) | |
1746 | break; | |
1747 | ||
1748 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) | |
1749 | break; | |
1750 | ||
c1b4a7e6 DM |
1751 | if (tso_segs == 1) { |
1752 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, | |
1753 | (tcp_skb_is_last(sk, skb) ? | |
1754 | nonagle : TCP_NAGLE_PUSH)))) | |
1755 | break; | |
1756 | } else { | |
d5dd9175 | 1757 | if (!push_one && tcp_tso_should_defer(sk, skb)) |
c1b4a7e6 DM |
1758 | break; |
1759 | } | |
aa93466b | 1760 | |
c8ac3774 | 1761 | limit = mss_now; |
f8269a49 | 1762 | if (tso_segs > 1 && !tcp_urg_mode(tp)) |
0e3a4803 IJ |
1763 | limit = tcp_mss_split_point(sk, skb, mss_now, |
1764 | cwnd_quota); | |
1da177e4 | 1765 | |
c8ac3774 | 1766 | if (skb->len > limit && |
c4ead4c5 | 1767 | unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) |
c8ac3774 HX |
1768 | break; |
1769 | ||
92df7b51 | 1770 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
c1b4a7e6 | 1771 | |
d5dd9175 | 1772 | if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) |
92df7b51 | 1773 | break; |
1da177e4 | 1774 | |
92df7b51 DM |
1775 | /* Advance the send_head. This one is sent out. |
1776 | * This call will increment packets_out. | |
1777 | */ | |
66f5fe62 | 1778 | tcp_event_new_data_sent(sk, skb); |
1da177e4 | 1779 | |
92df7b51 | 1780 | tcp_minshall_update(tp, mss_now, skb); |
aa93466b | 1781 | sent_pkts++; |
d5dd9175 IJ |
1782 | |
1783 | if (push_one) | |
1784 | break; | |
92df7b51 | 1785 | } |
1da177e4 | 1786 | |
aa93466b | 1787 | if (likely(sent_pkts)) { |
9e412ba7 | 1788 | tcp_cwnd_validate(sk); |
92df7b51 | 1789 | return 0; |
1da177e4 | 1790 | } |
fe067e8a | 1791 | return !tp->packets_out && tcp_send_head(sk); |
1da177e4 LT |
1792 | } |
1793 | ||
a762a980 DM |
1794 | /* Push out any pending frames which were held back due to |
1795 | * TCP_CORK or attempt at coalescing tiny packets. | |
1796 | * The socket must be locked by the caller. | |
1797 | */ | |
9e412ba7 IJ |
1798 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, |
1799 | int nonagle) | |
a762a980 | 1800 | { |
726e07a8 IJ |
1801 | /* If we are closed, the bytes will have to remain here. |
1802 | * In time closedown will finish, we empty the write queue and | |
1803 | * all will be happy. | |
1804 | */ | |
1805 | if (unlikely(sk->sk_state == TCP_CLOSE)) | |
1806 | return; | |
1807 | ||
d5dd9175 | 1808 | if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC)) |
726e07a8 | 1809 | tcp_check_probe_timer(sk); |
a762a980 DM |
1810 | } |
1811 | ||
c1b4a7e6 DM |
1812 | /* Send _single_ skb sitting at the send head. This function requires |
1813 | * true push pending frames to setup probe timer etc. | |
1814 | */ | |
1815 | void tcp_push_one(struct sock *sk, unsigned int mss_now) | |
1816 | { | |
fe067e8a | 1817 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 DM |
1818 | |
1819 | BUG_ON(!skb || skb->len < mss_now); | |
1820 | ||
d5dd9175 | 1821 | tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); |
c1b4a7e6 DM |
1822 | } |
1823 | ||
1da177e4 LT |
1824 | /* This function returns the amount that we can raise the |
1825 | * usable window based on the following constraints | |
e905a9ed | 1826 | * |
1da177e4 LT |
1827 | * 1. The window can never be shrunk once it is offered (RFC 793) |
1828 | * 2. We limit memory per socket | |
1829 | * | |
1830 | * RFC 1122: | |
1831 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep | |
1832 | * RECV.NEXT + RCV.WIN fixed until: | |
1833 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" | |
1834 | * | |
1835 | * i.e. don't raise the right edge of the window until you can raise | |
1836 | * it at least MSS bytes. | |
1837 | * | |
1838 | * Unfortunately, the recommended algorithm breaks header prediction, | |
1839 | * since header prediction assumes th->window stays fixed. | |
1840 | * | |
1841 | * Strictly speaking, keeping th->window fixed violates the receiver | |
1842 | * side SWS prevention criteria. The problem is that under this rule | |
1843 | * a stream of single byte packets will cause the right side of the | |
1844 | * window to always advance by a single byte. | |
e905a9ed | 1845 | * |
1da177e4 LT |
1846 | * Of course, if the sender implements sender side SWS prevention |
1847 | * then this will not be a problem. | |
e905a9ed | 1848 | * |
1da177e4 | 1849 | * BSD seems to make the following compromise: |
e905a9ed | 1850 | * |
1da177e4 LT |
1851 | * If the free space is less than the 1/4 of the maximum |
1852 | * space available and the free space is less than 1/2 mss, | |
1853 | * then set the window to 0. | |
1854 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] | |
1855 | * Otherwise, just prevent the window from shrinking | |
1856 | * and from being larger than the largest representable value. | |
1857 | * | |
1858 | * This prevents incremental opening of the window in the regime | |
1859 | * where TCP is limited by the speed of the reader side taking | |
1860 | * data out of the TCP receive queue. It does nothing about | |
1861 | * those cases where the window is constrained on the sender side | |
1862 | * because the pipeline is full. | |
1863 | * | |
1864 | * BSD also seems to "accidentally" limit itself to windows that are a | |
1865 | * multiple of MSS, at least until the free space gets quite small. | |
1866 | * This would appear to be a side effect of the mbuf implementation. | |
1867 | * Combining these two algorithms results in the observed behavior | |
1868 | * of having a fixed window size at almost all times. | |
1869 | * | |
1870 | * Below we obtain similar behavior by forcing the offered window to | |
1871 | * a multiple of the mss when it is feasible to do so. | |
1872 | * | |
1873 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. | |
1874 | * Regular options like TIMESTAMP are taken into account. | |
1875 | */ | |
1876 | u32 __tcp_select_window(struct sock *sk) | |
1877 | { | |
463c84b9 | 1878 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 1879 | struct tcp_sock *tp = tcp_sk(sk); |
caa20d9a | 1880 | /* MSS for the peer's data. Previous versions used mss_clamp |
1da177e4 LT |
1881 | * here. I don't know if the value based on our guesses |
1882 | * of peer's MSS is better for the performance. It's more correct | |
1883 | * but may be worse for the performance because of rcv_mss | |
1884 | * fluctuations. --SAW 1998/11/1 | |
1885 | */ | |
463c84b9 | 1886 | int mss = icsk->icsk_ack.rcv_mss; |
1da177e4 LT |
1887 | int free_space = tcp_space(sk); |
1888 | int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); | |
1889 | int window; | |
1890 | ||
1891 | if (mss > full_space) | |
e905a9ed | 1892 | mss = full_space; |
1da177e4 | 1893 | |
b92edbe0 | 1894 | if (free_space < (full_space >> 1)) { |
463c84b9 | 1895 | icsk->icsk_ack.quick = 0; |
1da177e4 LT |
1896 | |
1897 | if (tcp_memory_pressure) | |
056834d9 IJ |
1898 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, |
1899 | 4U * tp->advmss); | |
1da177e4 LT |
1900 | |
1901 | if (free_space < mss) | |
1902 | return 0; | |
1903 | } | |
1904 | ||
1905 | if (free_space > tp->rcv_ssthresh) | |
1906 | free_space = tp->rcv_ssthresh; | |
1907 | ||
1908 | /* Don't do rounding if we are using window scaling, since the | |
1909 | * scaled window will not line up with the MSS boundary anyway. | |
1910 | */ | |
1911 | window = tp->rcv_wnd; | |
1912 | if (tp->rx_opt.rcv_wscale) { | |
1913 | window = free_space; | |
1914 | ||
1915 | /* Advertise enough space so that it won't get scaled away. | |
1916 | * Import case: prevent zero window announcement if | |
1917 | * 1<<rcv_wscale > mss. | |
1918 | */ | |
1919 | if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) | |
1920 | window = (((window >> tp->rx_opt.rcv_wscale) + 1) | |
1921 | << tp->rx_opt.rcv_wscale); | |
1922 | } else { | |
1923 | /* Get the largest window that is a nice multiple of mss. | |
1924 | * Window clamp already applied above. | |
1925 | * If our current window offering is within 1 mss of the | |
1926 | * free space we just keep it. This prevents the divide | |
1927 | * and multiply from happening most of the time. | |
1928 | * We also don't do any window rounding when the free space | |
1929 | * is too small. | |
1930 | */ | |
1931 | if (window <= free_space - mss || window > free_space) | |
056834d9 | 1932 | window = (free_space / mss) * mss; |
84565070 | 1933 | else if (mss == full_space && |
b92edbe0 | 1934 | free_space > window + (full_space >> 1)) |
84565070 | 1935 | window = free_space; |
1da177e4 LT |
1936 | } |
1937 | ||
1938 | return window; | |
1939 | } | |
1940 | ||
4a17fc3a IJ |
1941 | /* Collapses two adjacent SKB's during retransmission. */ |
1942 | static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) | |
1da177e4 LT |
1943 | { |
1944 | struct tcp_sock *tp = tcp_sk(sk); | |
fe067e8a | 1945 | struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); |
058dc334 | 1946 | int skb_size, next_skb_size; |
1da177e4 | 1947 | |
058dc334 IJ |
1948 | skb_size = skb->len; |
1949 | next_skb_size = next_skb->len; | |
1da177e4 | 1950 | |
058dc334 | 1951 | BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); |
a6963a6b | 1952 | |
058dc334 | 1953 | tcp_highest_sack_combine(sk, next_skb, skb); |
1da177e4 | 1954 | |
058dc334 | 1955 | tcp_unlink_write_queue(next_skb, sk); |
1da177e4 | 1956 | |
058dc334 IJ |
1957 | skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), |
1958 | next_skb_size); | |
1da177e4 | 1959 | |
058dc334 IJ |
1960 | if (next_skb->ip_summed == CHECKSUM_PARTIAL) |
1961 | skb->ip_summed = CHECKSUM_PARTIAL; | |
1da177e4 | 1962 | |
058dc334 IJ |
1963 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
1964 | skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); | |
1da177e4 | 1965 | |
058dc334 IJ |
1966 | /* Update sequence range on original skb. */ |
1967 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; | |
1da177e4 | 1968 | |
e6c7d085 IJ |
1969 | /* Merge over control information. This moves PSH/FIN etc. over */ |
1970 | TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(next_skb)->flags; | |
058dc334 IJ |
1971 | |
1972 | /* All done, get rid of second SKB and account for it so | |
1973 | * packet counting does not break. | |
1974 | */ | |
1975 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; | |
058dc334 IJ |
1976 | |
1977 | /* changed transmit queue under us so clear hints */ | |
ef9da47c IJ |
1978 | tcp_clear_retrans_hints_partial(tp); |
1979 | if (next_skb == tp->retransmit_skb_hint) | |
1980 | tp->retransmit_skb_hint = skb; | |
058dc334 | 1981 | |
797108d1 IJ |
1982 | tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); |
1983 | ||
058dc334 | 1984 | sk_wmem_free_skb(sk, next_skb); |
1da177e4 LT |
1985 | } |
1986 | ||
67edfef7 | 1987 | /* Check if coalescing SKBs is legal. */ |
4a17fc3a IJ |
1988 | static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb) |
1989 | { | |
1990 | if (tcp_skb_pcount(skb) > 1) | |
1991 | return 0; | |
1992 | /* TODO: SACK collapsing could be used to remove this condition */ | |
1993 | if (skb_shinfo(skb)->nr_frags != 0) | |
1994 | return 0; | |
1995 | if (skb_cloned(skb)) | |
1996 | return 0; | |
1997 | if (skb == tcp_send_head(sk)) | |
1998 | return 0; | |
1999 | /* Some heurestics for collapsing over SACK'd could be invented */ | |
2000 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
2001 | return 0; | |
2002 | ||
2003 | return 1; | |
2004 | } | |
2005 | ||
67edfef7 AK |
2006 | /* Collapse packets in the retransmit queue to make to create |
2007 | * less packets on the wire. This is only done on retransmission. | |
2008 | */ | |
4a17fc3a IJ |
2009 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, |
2010 | int space) | |
2011 | { | |
2012 | struct tcp_sock *tp = tcp_sk(sk); | |
2013 | struct sk_buff *skb = to, *tmp; | |
2014 | int first = 1; | |
2015 | ||
2016 | if (!sysctl_tcp_retrans_collapse) | |
2017 | return; | |
a3433f35 | 2018 | if (TCP_SKB_CB(skb)->flags & TCPHDR_SYN) |
4a17fc3a IJ |
2019 | return; |
2020 | ||
2021 | tcp_for_write_queue_from_safe(skb, tmp, sk) { | |
2022 | if (!tcp_can_collapse(sk, skb)) | |
2023 | break; | |
2024 | ||
2025 | space -= skb->len; | |
2026 | ||
2027 | if (first) { | |
2028 | first = 0; | |
2029 | continue; | |
2030 | } | |
2031 | ||
2032 | if (space < 0) | |
2033 | break; | |
2034 | /* Punt if not enough space exists in the first SKB for | |
2035 | * the data in the second | |
2036 | */ | |
2037 | if (skb->len > skb_tailroom(to)) | |
2038 | break; | |
2039 | ||
2040 | if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) | |
2041 | break; | |
2042 | ||
2043 | tcp_collapse_retrans(sk, to); | |
2044 | } | |
2045 | } | |
2046 | ||
1da177e4 LT |
2047 | /* This retransmits one SKB. Policy decisions and retransmit queue |
2048 | * state updates are done by the caller. Returns non-zero if an | |
2049 | * error occurred which prevented the send. | |
2050 | */ | |
2051 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) | |
2052 | { | |
2053 | struct tcp_sock *tp = tcp_sk(sk); | |
5d424d5a | 2054 | struct inet_connection_sock *icsk = inet_csk(sk); |
7d227cd2 | 2055 | unsigned int cur_mss; |
1da177e4 LT |
2056 | int err; |
2057 | ||
5d424d5a JH |
2058 | /* Inconslusive MTU probe */ |
2059 | if (icsk->icsk_mtup.probe_size) { | |
2060 | icsk->icsk_mtup.probe_size = 0; | |
2061 | } | |
2062 | ||
1da177e4 | 2063 | /* Do not sent more than we queued. 1/4 is reserved for possible |
caa20d9a | 2064 | * copying overhead: fragmentation, tunneling, mangling etc. |
1da177e4 LT |
2065 | */ |
2066 | if (atomic_read(&sk->sk_wmem_alloc) > | |
2067 | min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) | |
2068 | return -EAGAIN; | |
2069 | ||
2070 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { | |
2071 | if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
2072 | BUG(); | |
1da177e4 LT |
2073 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
2074 | return -ENOMEM; | |
2075 | } | |
2076 | ||
7d227cd2 SS |
2077 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) |
2078 | return -EHOSTUNREACH; /* Routing failure or similar. */ | |
2079 | ||
0c54b85f | 2080 | cur_mss = tcp_current_mss(sk); |
7d227cd2 | 2081 | |
1da177e4 LT |
2082 | /* If receiver has shrunk his window, and skb is out of |
2083 | * new window, do not retransmit it. The exception is the | |
2084 | * case, when window is shrunk to zero. In this case | |
2085 | * our retransmit serves as a zero window probe. | |
2086 | */ | |
9d4fb27d JP |
2087 | if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && |
2088 | TCP_SKB_CB(skb)->seq != tp->snd_una) | |
1da177e4 LT |
2089 | return -EAGAIN; |
2090 | ||
2091 | if (skb->len > cur_mss) { | |
846998ae | 2092 | if (tcp_fragment(sk, skb, cur_mss, cur_mss)) |
1da177e4 | 2093 | return -ENOMEM; /* We'll try again later. */ |
02276f3c | 2094 | } else { |
9eb9362e IJ |
2095 | int oldpcount = tcp_skb_pcount(skb); |
2096 | ||
2097 | if (unlikely(oldpcount > 1)) { | |
2098 | tcp_init_tso_segs(sk, skb, cur_mss); | |
2099 | tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); | |
2100 | } | |
1da177e4 LT |
2101 | } |
2102 | ||
4a17fc3a | 2103 | tcp_retrans_try_collapse(sk, skb, cur_mss); |
1da177e4 | 2104 | |
1da177e4 LT |
2105 | /* Some Solaris stacks overoptimize and ignore the FIN on a |
2106 | * retransmit when old data is attached. So strip it off | |
2107 | * since it is cheap to do so and saves bytes on the network. | |
2108 | */ | |
2de979bd | 2109 | if (skb->len > 0 && |
a3433f35 | 2110 | (TCP_SKB_CB(skb)->flags & TCPHDR_FIN) && |
2de979bd | 2111 | tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { |
1da177e4 | 2112 | if (!pskb_trim(skb, 0)) { |
e870a8ef IJ |
2113 | /* Reuse, even though it does some unnecessary work */ |
2114 | tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, | |
2115 | TCP_SKB_CB(skb)->flags); | |
1da177e4 | 2116 | skb->ip_summed = CHECKSUM_NONE; |
1da177e4 LT |
2117 | } |
2118 | } | |
2119 | ||
2120 | /* Make a copy, if the first transmission SKB clone we made | |
2121 | * is still in somebody's hands, else make a clone. | |
2122 | */ | |
2123 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
1da177e4 | 2124 | |
dfb4b9dc | 2125 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
1da177e4 LT |
2126 | |
2127 | if (err == 0) { | |
2128 | /* Update global TCP statistics. */ | |
81cc8a75 | 2129 | TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); |
1da177e4 LT |
2130 | |
2131 | tp->total_retrans++; | |
2132 | ||
2133 | #if FASTRETRANS_DEBUG > 0 | |
056834d9 | 2134 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
1da177e4 LT |
2135 | if (net_ratelimit()) |
2136 | printk(KERN_DEBUG "retrans_out leaked.\n"); | |
2137 | } | |
2138 | #endif | |
b08d6cb2 IJ |
2139 | if (!tp->retrans_out) |
2140 | tp->lost_retrans_low = tp->snd_nxt; | |
1da177e4 LT |
2141 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; |
2142 | tp->retrans_out += tcp_skb_pcount(skb); | |
2143 | ||
2144 | /* Save stamp of the first retransmit. */ | |
2145 | if (!tp->retrans_stamp) | |
2146 | tp->retrans_stamp = TCP_SKB_CB(skb)->when; | |
2147 | ||
2148 | tp->undo_retrans++; | |
2149 | ||
2150 | /* snd_nxt is stored to detect loss of retransmitted segment, | |
2151 | * see tcp_input.c tcp_sacktag_write_queue(). | |
2152 | */ | |
2153 | TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; | |
2154 | } | |
2155 | return err; | |
2156 | } | |
2157 | ||
67edfef7 AK |
2158 | /* Check if we forward retransmits are possible in the current |
2159 | * window/congestion state. | |
2160 | */ | |
b5afe7bc IJ |
2161 | static int tcp_can_forward_retransmit(struct sock *sk) |
2162 | { | |
2163 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
2164 | struct tcp_sock *tp = tcp_sk(sk); | |
2165 | ||
2166 | /* Forward retransmissions are possible only during Recovery. */ | |
2167 | if (icsk->icsk_ca_state != TCP_CA_Recovery) | |
2168 | return 0; | |
2169 | ||
2170 | /* No forward retransmissions in Reno are possible. */ | |
2171 | if (tcp_is_reno(tp)) | |
2172 | return 0; | |
2173 | ||
2174 | /* Yeah, we have to make difficult choice between forward transmission | |
2175 | * and retransmission... Both ways have their merits... | |
2176 | * | |
2177 | * For now we do not retransmit anything, while we have some new | |
2178 | * segments to send. In the other cases, follow rule 3 for | |
2179 | * NextSeg() specified in RFC3517. | |
2180 | */ | |
2181 | ||
2182 | if (tcp_may_send_now(sk)) | |
2183 | return 0; | |
2184 | ||
2185 | return 1; | |
2186 | } | |
2187 | ||
1da177e4 LT |
2188 | /* This gets called after a retransmit timeout, and the initially |
2189 | * retransmitted data is acknowledged. It tries to continue | |
2190 | * resending the rest of the retransmit queue, until either | |
2191 | * we've sent it all or the congestion window limit is reached. | |
2192 | * If doing SACK, the first ACK which comes back for a timeout | |
2193 | * based retransmit packet might feed us FACK information again. | |
2194 | * If so, we use it to avoid unnecessarily retransmissions. | |
2195 | */ | |
2196 | void tcp_xmit_retransmit_queue(struct sock *sk) | |
2197 | { | |
6687e988 | 2198 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2199 | struct tcp_sock *tp = tcp_sk(sk); |
2200 | struct sk_buff *skb; | |
0e1c54c2 | 2201 | struct sk_buff *hole = NULL; |
618d9f25 | 2202 | u32 last_lost; |
61eb55f4 | 2203 | int mib_idx; |
0e1c54c2 | 2204 | int fwd_rexmitting = 0; |
6a438bbe | 2205 | |
45e77d31 IJ |
2206 | if (!tp->packets_out) |
2207 | return; | |
2208 | ||
08ebd172 IJ |
2209 | if (!tp->lost_out) |
2210 | tp->retransmit_high = tp->snd_una; | |
2211 | ||
618d9f25 | 2212 | if (tp->retransmit_skb_hint) { |
6a438bbe | 2213 | skb = tp->retransmit_skb_hint; |
618d9f25 IJ |
2214 | last_lost = TCP_SKB_CB(skb)->end_seq; |
2215 | if (after(last_lost, tp->retransmit_high)) | |
2216 | last_lost = tp->retransmit_high; | |
2217 | } else { | |
fe067e8a | 2218 | skb = tcp_write_queue_head(sk); |
618d9f25 IJ |
2219 | last_lost = tp->snd_una; |
2220 | } | |
1da177e4 | 2221 | |
08ebd172 IJ |
2222 | tcp_for_write_queue_from(skb, sk) { |
2223 | __u8 sacked = TCP_SKB_CB(skb)->sacked; | |
1da177e4 | 2224 | |
08ebd172 IJ |
2225 | if (skb == tcp_send_head(sk)) |
2226 | break; | |
2227 | /* we could do better than to assign each time */ | |
0e1c54c2 IJ |
2228 | if (hole == NULL) |
2229 | tp->retransmit_skb_hint = skb; | |
08ebd172 IJ |
2230 | |
2231 | /* Assume this retransmit will generate | |
2232 | * only one packet for congestion window | |
2233 | * calculation purposes. This works because | |
2234 | * tcp_retransmit_skb() will chop up the | |
2235 | * packet to be MSS sized and all the | |
2236 | * packet counting works out. | |
2237 | */ | |
2238 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) | |
2239 | return; | |
1da177e4 | 2240 | |
0e1c54c2 IJ |
2241 | if (fwd_rexmitting) { |
2242 | begin_fwd: | |
2243 | if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) | |
2244 | break; | |
2245 | mib_idx = LINUX_MIB_TCPFORWARDRETRANS; | |
6a438bbe | 2246 | |
0e1c54c2 | 2247 | } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { |
618d9f25 | 2248 | tp->retransmit_high = last_lost; |
0e1c54c2 IJ |
2249 | if (!tcp_can_forward_retransmit(sk)) |
2250 | break; | |
2251 | /* Backtrack if necessary to non-L'ed skb */ | |
2252 | if (hole != NULL) { | |
2253 | skb = hole; | |
2254 | hole = NULL; | |
2255 | } | |
2256 | fwd_rexmitting = 1; | |
2257 | goto begin_fwd; | |
1da177e4 | 2258 | |
0e1c54c2 | 2259 | } else if (!(sacked & TCPCB_LOST)) { |
ac11ba75 | 2260 | if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) |
0e1c54c2 IJ |
2261 | hole = skb; |
2262 | continue; | |
1da177e4 | 2263 | |
0e1c54c2 | 2264 | } else { |
618d9f25 | 2265 | last_lost = TCP_SKB_CB(skb)->end_seq; |
0e1c54c2 IJ |
2266 | if (icsk->icsk_ca_state != TCP_CA_Loss) |
2267 | mib_idx = LINUX_MIB_TCPFASTRETRANS; | |
2268 | else | |
2269 | mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; | |
2270 | } | |
1da177e4 | 2271 | |
0e1c54c2 | 2272 | if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) |
1da177e4 LT |
2273 | continue; |
2274 | ||
f0ceb0ed | 2275 | if (tcp_retransmit_skb(sk, skb)) |
0e1c54c2 | 2276 | return; |
0e1c54c2 | 2277 | NET_INC_STATS_BH(sock_net(sk), mib_idx); |
1da177e4 | 2278 | |
fe067e8a | 2279 | if (skb == tcp_write_queue_head(sk)) |
3f421baa ACM |
2280 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
2281 | inet_csk(sk)->icsk_rto, | |
2282 | TCP_RTO_MAX); | |
1da177e4 LT |
2283 | } |
2284 | } | |
2285 | ||
1da177e4 LT |
2286 | /* Send a fin. The caller locks the socket for us. This cannot be |
2287 | * allowed to fail queueing a FIN frame under any circumstances. | |
2288 | */ | |
2289 | void tcp_send_fin(struct sock *sk) | |
2290 | { | |
e905a9ed | 2291 | struct tcp_sock *tp = tcp_sk(sk); |
fe067e8a | 2292 | struct sk_buff *skb = tcp_write_queue_tail(sk); |
1da177e4 | 2293 | int mss_now; |
e905a9ed | 2294 | |
1da177e4 LT |
2295 | /* Optimization, tack on the FIN if we have a queue of |
2296 | * unsent frames. But be careful about outgoing SACKS | |
2297 | * and IP options. | |
2298 | */ | |
0c54b85f | 2299 | mss_now = tcp_current_mss(sk); |
1da177e4 | 2300 | |
fe067e8a | 2301 | if (tcp_send_head(sk) != NULL) { |
a3433f35 | 2302 | TCP_SKB_CB(skb)->flags |= TCPHDR_FIN; |
1da177e4 LT |
2303 | TCP_SKB_CB(skb)->end_seq++; |
2304 | tp->write_seq++; | |
2305 | } else { | |
2306 | /* Socket is locked, keep trying until memory is available. */ | |
2307 | for (;;) { | |
aa133076 WF |
2308 | skb = alloc_skb_fclone(MAX_TCP_HEADER, |
2309 | sk->sk_allocation); | |
1da177e4 LT |
2310 | if (skb) |
2311 | break; | |
2312 | yield(); | |
2313 | } | |
2314 | ||
2315 | /* Reserve space for headers and prepare control bits. */ | |
2316 | skb_reserve(skb, MAX_TCP_HEADER); | |
1da177e4 | 2317 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ |
e870a8ef | 2318 | tcp_init_nondata_skb(skb, tp->write_seq, |
a3433f35 | 2319 | TCPHDR_ACK | TCPHDR_FIN); |
1da177e4 LT |
2320 | tcp_queue_skb(sk, skb); |
2321 | } | |
9e412ba7 | 2322 | __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); |
1da177e4 LT |
2323 | } |
2324 | ||
2325 | /* We get here when a process closes a file descriptor (either due to | |
2326 | * an explicit close() or as a byproduct of exit()'ing) and there | |
2327 | * was unread data in the receive queue. This behavior is recommended | |
65bb723c | 2328 | * by RFC 2525, section 2.17. -DaveM |
1da177e4 | 2329 | */ |
dd0fc66f | 2330 | void tcp_send_active_reset(struct sock *sk, gfp_t priority) |
1da177e4 | 2331 | { |
1da177e4 LT |
2332 | struct sk_buff *skb; |
2333 | ||
2334 | /* NOTE: No TCP options attached and we never retransmit this. */ | |
2335 | skb = alloc_skb(MAX_TCP_HEADER, priority); | |
2336 | if (!skb) { | |
4e673444 | 2337 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
1da177e4 LT |
2338 | return; |
2339 | } | |
2340 | ||
2341 | /* Reserve space for headers and prepare control bits. */ | |
2342 | skb_reserve(skb, MAX_TCP_HEADER); | |
e870a8ef | 2343 | tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), |
a3433f35 | 2344 | TCPHDR_ACK | TCPHDR_RST); |
1da177e4 | 2345 | /* Send it off. */ |
1da177e4 | 2346 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
dfb4b9dc | 2347 | if (tcp_transmit_skb(sk, skb, 0, priority)) |
4e673444 | 2348 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
26af65cb | 2349 | |
81cc8a75 | 2350 | TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); |
1da177e4 LT |
2351 | } |
2352 | ||
67edfef7 AK |
2353 | /* Send a crossed SYN-ACK during socket establishment. |
2354 | * WARNING: This routine must only be called when we have already sent | |
1da177e4 LT |
2355 | * a SYN packet that crossed the incoming SYN that caused this routine |
2356 | * to get called. If this assumption fails then the initial rcv_wnd | |
2357 | * and rcv_wscale values will not be correct. | |
2358 | */ | |
2359 | int tcp_send_synack(struct sock *sk) | |
2360 | { | |
056834d9 | 2361 | struct sk_buff *skb; |
1da177e4 | 2362 | |
fe067e8a | 2363 | skb = tcp_write_queue_head(sk); |
a3433f35 | 2364 | if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPHDR_SYN)) { |
1da177e4 LT |
2365 | printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); |
2366 | return -EFAULT; | |
2367 | } | |
a3433f35 | 2368 | if (!(TCP_SKB_CB(skb)->flags & TCPHDR_ACK)) { |
1da177e4 LT |
2369 | if (skb_cloned(skb)) { |
2370 | struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); | |
2371 | if (nskb == NULL) | |
2372 | return -ENOMEM; | |
fe067e8a | 2373 | tcp_unlink_write_queue(skb, sk); |
1da177e4 | 2374 | skb_header_release(nskb); |
fe067e8a | 2375 | __tcp_add_write_queue_head(sk, nskb); |
3ab224be HA |
2376 | sk_wmem_free_skb(sk, skb); |
2377 | sk->sk_wmem_queued += nskb->truesize; | |
2378 | sk_mem_charge(sk, nskb->truesize); | |
1da177e4 LT |
2379 | skb = nskb; |
2380 | } | |
2381 | ||
a3433f35 | 2382 | TCP_SKB_CB(skb)->flags |= TCPHDR_ACK; |
1da177e4 LT |
2383 | TCP_ECN_send_synack(tcp_sk(sk), skb); |
2384 | } | |
2385 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
dfb4b9dc | 2386 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
1da177e4 LT |
2387 | } |
2388 | ||
67edfef7 | 2389 | /* Prepare a SYN-ACK. */ |
056834d9 | 2390 | struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, |
e6b4d113 WAS |
2391 | struct request_sock *req, |
2392 | struct request_values *rvp) | |
1da177e4 | 2393 | { |
bd0388ae | 2394 | struct tcp_out_options opts; |
4957faad | 2395 | struct tcp_extend_values *xvp = tcp_xv(rvp); |
2e6599cb | 2396 | struct inet_request_sock *ireq = inet_rsk(req); |
1da177e4 | 2397 | struct tcp_sock *tp = tcp_sk(sk); |
28b2774a | 2398 | const struct tcp_cookie_values *cvp = tp->cookie_values; |
1da177e4 | 2399 | struct tcphdr *th; |
1da177e4 | 2400 | struct sk_buff *skb; |
cfb6eeb4 | 2401 | struct tcp_md5sig_key *md5; |
bd0388ae | 2402 | int tcp_header_size; |
f5fff5dc | 2403 | int mss; |
28b2774a | 2404 | int s_data_desired = 0; |
1da177e4 | 2405 | |
28b2774a ED |
2406 | if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired) |
2407 | s_data_desired = cvp->s_data_desired; | |
2408 | skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15 + s_data_desired, 1, GFP_ATOMIC); | |
1da177e4 LT |
2409 | if (skb == NULL) |
2410 | return NULL; | |
2411 | ||
2412 | /* Reserve space for headers. */ | |
2413 | skb_reserve(skb, MAX_TCP_HEADER); | |
2414 | ||
adf30907 | 2415 | skb_dst_set(skb, dst_clone(dst)); |
1da177e4 | 2416 | |
f5fff5dc TQ |
2417 | mss = dst_metric(dst, RTAX_ADVMSS); |
2418 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) | |
2419 | mss = tp->rx_opt.user_mss; | |
2420 | ||
33ad798c AL |
2421 | if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ |
2422 | __u8 rcv_wscale; | |
2423 | /* Set this up on the first call only */ | |
2424 | req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); | |
e88c64f0 HPP |
2425 | |
2426 | /* limit the window selection if the user enforce a smaller rx buffer */ | |
2427 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && | |
2428 | (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) | |
2429 | req->window_clamp = tcp_full_space(sk); | |
2430 | ||
33ad798c AL |
2431 | /* tcp_full_space because it is guaranteed to be the first packet */ |
2432 | tcp_select_initial_window(tcp_full_space(sk), | |
f5fff5dc | 2433 | mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), |
33ad798c AL |
2434 | &req->rcv_wnd, |
2435 | &req->window_clamp, | |
2436 | ireq->wscale_ok, | |
31d12926 | 2437 | &rcv_wscale, |
2438 | dst_metric(dst, RTAX_INITRWND)); | |
33ad798c AL |
2439 | ireq->rcv_wscale = rcv_wscale; |
2440 | } | |
2441 | ||
2442 | memset(&opts, 0, sizeof(opts)); | |
8b5f12d0 FW |
2443 | #ifdef CONFIG_SYN_COOKIES |
2444 | if (unlikely(req->cookie_ts)) | |
2445 | TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); | |
2446 | else | |
2447 | #endif | |
33ad798c | 2448 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
f5fff5dc | 2449 | tcp_header_size = tcp_synack_options(sk, req, mss, |
4957faad WAS |
2450 | skb, &opts, &md5, xvp) |
2451 | + sizeof(*th); | |
cfb6eeb4 | 2452 | |
aa8223c7 ACM |
2453 | skb_push(skb, tcp_header_size); |
2454 | skb_reset_transport_header(skb); | |
1da177e4 | 2455 | |
aa8223c7 | 2456 | th = tcp_hdr(skb); |
1da177e4 LT |
2457 | memset(th, 0, sizeof(struct tcphdr)); |
2458 | th->syn = 1; | |
2459 | th->ack = 1; | |
1da177e4 | 2460 | TCP_ECN_make_synack(req, th); |
a3116ac5 | 2461 | th->source = ireq->loc_port; |
2e6599cb | 2462 | th->dest = ireq->rmt_port; |
e870a8ef IJ |
2463 | /* Setting of flags are superfluous here for callers (and ECE is |
2464 | * not even correctly set) | |
2465 | */ | |
2466 | tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, | |
a3433f35 | 2467 | TCPHDR_SYN | TCPHDR_ACK); |
4957faad WAS |
2468 | |
2469 | if (OPTION_COOKIE_EXTENSION & opts.options) { | |
28b2774a ED |
2470 | if (s_data_desired) { |
2471 | u8 *buf = skb_put(skb, s_data_desired); | |
4957faad WAS |
2472 | |
2473 | /* copy data directly from the listening socket. */ | |
28b2774a ED |
2474 | memcpy(buf, cvp->s_data_payload, s_data_desired); |
2475 | TCP_SKB_CB(skb)->end_seq += s_data_desired; | |
4957faad WAS |
2476 | } |
2477 | ||
2478 | if (opts.hash_size > 0) { | |
2479 | __u32 workspace[SHA_WORKSPACE_WORDS]; | |
2480 | u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS]; | |
2481 | u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1]; | |
2482 | ||
2483 | /* Secret recipe depends on the Timestamp, (future) | |
2484 | * Sequence and Acknowledgment Numbers, Initiator | |
2485 | * Cookie, and others handled by IP variant caller. | |
2486 | */ | |
2487 | *tail-- ^= opts.tsval; | |
2488 | *tail-- ^= tcp_rsk(req)->rcv_isn + 1; | |
2489 | *tail-- ^= TCP_SKB_CB(skb)->seq + 1; | |
2490 | ||
2491 | /* recommended */ | |
0eae88f3 | 2492 | *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source); |
f9a2e69e | 2493 | *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */ |
4957faad WAS |
2494 | |
2495 | sha_transform((__u32 *)&xvp->cookie_bakery[0], | |
2496 | (char *)mess, | |
2497 | &workspace[0]); | |
2498 | opts.hash_location = | |
2499 | (__u8 *)&xvp->cookie_bakery[0]; | |
2500 | } | |
2501 | } | |
2502 | ||
1da177e4 | 2503 | th->seq = htonl(TCP_SKB_CB(skb)->seq); |
2e6599cb | 2504 | th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); |
1da177e4 LT |
2505 | |
2506 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ | |
600ff0c2 | 2507 | th->window = htons(min(req->rcv_wnd, 65535U)); |
bd0388ae | 2508 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
1da177e4 | 2509 | th->doff = (tcp_header_size >> 2); |
aa2ea058 | 2510 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb)); |
cfb6eeb4 YH |
2511 | |
2512 | #ifdef CONFIG_TCP_MD5SIG | |
2513 | /* Okay, we have all we need - do the md5 hash if needed */ | |
2514 | if (md5) { | |
bd0388ae | 2515 | tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, |
49a72dfb | 2516 | md5, NULL, req, skb); |
cfb6eeb4 YH |
2517 | } |
2518 | #endif | |
2519 | ||
1da177e4 LT |
2520 | return skb; |
2521 | } | |
4bc2f18b | 2522 | EXPORT_SYMBOL(tcp_make_synack); |
1da177e4 | 2523 | |
67edfef7 | 2524 | /* Do all connect socket setups that can be done AF independent. */ |
40efc6fa | 2525 | static void tcp_connect_init(struct sock *sk) |
1da177e4 LT |
2526 | { |
2527 | struct dst_entry *dst = __sk_dst_get(sk); | |
2528 | struct tcp_sock *tp = tcp_sk(sk); | |
2529 | __u8 rcv_wscale; | |
2530 | ||
2531 | /* We'll fix this up when we get a response from the other end. | |
2532 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. | |
2533 | */ | |
2534 | tp->tcp_header_len = sizeof(struct tcphdr) + | |
bb5b7c11 | 2535 | (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); |
1da177e4 | 2536 | |
cfb6eeb4 YH |
2537 | #ifdef CONFIG_TCP_MD5SIG |
2538 | if (tp->af_specific->md5_lookup(sk, sk) != NULL) | |
2539 | tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; | |
2540 | #endif | |
2541 | ||
1da177e4 LT |
2542 | /* If user gave his TCP_MAXSEG, record it to clamp */ |
2543 | if (tp->rx_opt.user_mss) | |
2544 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
2545 | tp->max_window = 0; | |
5d424d5a | 2546 | tcp_mtup_init(sk); |
1da177e4 LT |
2547 | tcp_sync_mss(sk, dst_mtu(dst)); |
2548 | ||
2549 | if (!tp->window_clamp) | |
2550 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); | |
2551 | tp->advmss = dst_metric(dst, RTAX_ADVMSS); | |
f5fff5dc TQ |
2552 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) |
2553 | tp->advmss = tp->rx_opt.user_mss; | |
2554 | ||
1da177e4 | 2555 | tcp_initialize_rcv_mss(sk); |
1da177e4 | 2556 | |
e88c64f0 HPP |
2557 | /* limit the window selection if the user enforce a smaller rx buffer */ |
2558 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && | |
2559 | (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) | |
2560 | tp->window_clamp = tcp_full_space(sk); | |
2561 | ||
1da177e4 LT |
2562 | tcp_select_initial_window(tcp_full_space(sk), |
2563 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), | |
2564 | &tp->rcv_wnd, | |
2565 | &tp->window_clamp, | |
bb5b7c11 | 2566 | sysctl_tcp_window_scaling, |
31d12926 | 2567 | &rcv_wscale, |
2568 | dst_metric(dst, RTAX_INITRWND)); | |
1da177e4 LT |
2569 | |
2570 | tp->rx_opt.rcv_wscale = rcv_wscale; | |
2571 | tp->rcv_ssthresh = tp->rcv_wnd; | |
2572 | ||
2573 | sk->sk_err = 0; | |
2574 | sock_reset_flag(sk, SOCK_DONE); | |
2575 | tp->snd_wnd = 0; | |
ee7537b6 | 2576 | tcp_init_wl(tp, 0); |
1da177e4 LT |
2577 | tp->snd_una = tp->write_seq; |
2578 | tp->snd_sml = tp->write_seq; | |
33f5f57e | 2579 | tp->snd_up = tp->write_seq; |
1da177e4 LT |
2580 | tp->rcv_nxt = 0; |
2581 | tp->rcv_wup = 0; | |
2582 | tp->copied_seq = 0; | |
2583 | ||
463c84b9 ACM |
2584 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
2585 | inet_csk(sk)->icsk_retransmits = 0; | |
1da177e4 LT |
2586 | tcp_clear_retrans(tp); |
2587 | } | |
2588 | ||
67edfef7 | 2589 | /* Build a SYN and send it off. */ |
1da177e4 LT |
2590 | int tcp_connect(struct sock *sk) |
2591 | { | |
2592 | struct tcp_sock *tp = tcp_sk(sk); | |
2593 | struct sk_buff *buff; | |
2594 | ||
2595 | tcp_connect_init(sk); | |
2596 | ||
d179cd12 | 2597 | buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); |
1da177e4 LT |
2598 | if (unlikely(buff == NULL)) |
2599 | return -ENOBUFS; | |
2600 | ||
2601 | /* Reserve space for headers. */ | |
2602 | skb_reserve(buff, MAX_TCP_HEADER); | |
2603 | ||
bd37a088 | 2604 | tp->snd_nxt = tp->write_seq; |
a3433f35 | 2605 | tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); |
e870a8ef | 2606 | TCP_ECN_send_syn(sk, buff); |
1da177e4 LT |
2607 | |
2608 | /* Send it off. */ | |
2609 | TCP_SKB_CB(buff)->when = tcp_time_stamp; | |
2610 | tp->retrans_stamp = TCP_SKB_CB(buff)->when; | |
2611 | skb_header_release(buff); | |
fe067e8a | 2612 | __tcp_add_write_queue_tail(sk, buff); |
3ab224be HA |
2613 | sk->sk_wmem_queued += buff->truesize; |
2614 | sk_mem_charge(sk, buff->truesize); | |
1da177e4 | 2615 | tp->packets_out += tcp_skb_pcount(buff); |
aa133076 | 2616 | tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); |
bd37a088 WY |
2617 | |
2618 | /* We change tp->snd_nxt after the tcp_transmit_skb() call | |
2619 | * in order to make this packet get counted in tcpOutSegs. | |
2620 | */ | |
2621 | tp->snd_nxt = tp->write_seq; | |
2622 | tp->pushed_seq = tp->write_seq; | |
81cc8a75 | 2623 | TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); |
1da177e4 LT |
2624 | |
2625 | /* Timer for repeating the SYN until an answer. */ | |
3f421baa ACM |
2626 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
2627 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
1da177e4 LT |
2628 | return 0; |
2629 | } | |
4bc2f18b | 2630 | EXPORT_SYMBOL(tcp_connect); |
1da177e4 LT |
2631 | |
2632 | /* Send out a delayed ack, the caller does the policy checking | |
2633 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() | |
2634 | * for details. | |
2635 | */ | |
2636 | void tcp_send_delayed_ack(struct sock *sk) | |
2637 | { | |
463c84b9 ACM |
2638 | struct inet_connection_sock *icsk = inet_csk(sk); |
2639 | int ato = icsk->icsk_ack.ato; | |
1da177e4 LT |
2640 | unsigned long timeout; |
2641 | ||
2642 | if (ato > TCP_DELACK_MIN) { | |
463c84b9 | 2643 | const struct tcp_sock *tp = tcp_sk(sk); |
056834d9 | 2644 | int max_ato = HZ / 2; |
1da177e4 | 2645 | |
056834d9 IJ |
2646 | if (icsk->icsk_ack.pingpong || |
2647 | (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) | |
1da177e4 LT |
2648 | max_ato = TCP_DELACK_MAX; |
2649 | ||
2650 | /* Slow path, intersegment interval is "high". */ | |
2651 | ||
2652 | /* If some rtt estimate is known, use it to bound delayed ack. | |
463c84b9 | 2653 | * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements |
1da177e4 LT |
2654 | * directly. |
2655 | */ | |
2656 | if (tp->srtt) { | |
056834d9 | 2657 | int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); |
1da177e4 LT |
2658 | |
2659 | if (rtt < max_ato) | |
2660 | max_ato = rtt; | |
2661 | } | |
2662 | ||
2663 | ato = min(ato, max_ato); | |
2664 | } | |
2665 | ||
2666 | /* Stay within the limit we were given */ | |
2667 | timeout = jiffies + ato; | |
2668 | ||
2669 | /* Use new timeout only if there wasn't a older one earlier. */ | |
463c84b9 | 2670 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { |
1da177e4 LT |
2671 | /* If delack timer was blocked or is about to expire, |
2672 | * send ACK now. | |
2673 | */ | |
463c84b9 ACM |
2674 | if (icsk->icsk_ack.blocked || |
2675 | time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { | |
1da177e4 LT |
2676 | tcp_send_ack(sk); |
2677 | return; | |
2678 | } | |
2679 | ||
463c84b9 ACM |
2680 | if (!time_before(timeout, icsk->icsk_ack.timeout)) |
2681 | timeout = icsk->icsk_ack.timeout; | |
1da177e4 | 2682 | } |
463c84b9 ACM |
2683 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; |
2684 | icsk->icsk_ack.timeout = timeout; | |
2685 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); | |
1da177e4 LT |
2686 | } |
2687 | ||
2688 | /* This routine sends an ack and also updates the window. */ | |
2689 | void tcp_send_ack(struct sock *sk) | |
2690 | { | |
058dc334 | 2691 | struct sk_buff *buff; |
1da177e4 | 2692 | |
058dc334 IJ |
2693 | /* If we have been reset, we may not send again. */ |
2694 | if (sk->sk_state == TCP_CLOSE) | |
2695 | return; | |
1da177e4 | 2696 | |
058dc334 IJ |
2697 | /* We are not putting this on the write queue, so |
2698 | * tcp_transmit_skb() will set the ownership to this | |
2699 | * sock. | |
2700 | */ | |
2701 | buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); | |
2702 | if (buff == NULL) { | |
2703 | inet_csk_schedule_ack(sk); | |
2704 | inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; | |
2705 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, | |
2706 | TCP_DELACK_MAX, TCP_RTO_MAX); | |
2707 | return; | |
1da177e4 | 2708 | } |
058dc334 IJ |
2709 | |
2710 | /* Reserve space for headers and prepare control bits. */ | |
2711 | skb_reserve(buff, MAX_TCP_HEADER); | |
a3433f35 | 2712 | tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); |
058dc334 IJ |
2713 | |
2714 | /* Send it off, this clears delayed acks for us. */ | |
058dc334 IJ |
2715 | TCP_SKB_CB(buff)->when = tcp_time_stamp; |
2716 | tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); | |
1da177e4 LT |
2717 | } |
2718 | ||
2719 | /* This routine sends a packet with an out of date sequence | |
2720 | * number. It assumes the other end will try to ack it. | |
2721 | * | |
2722 | * Question: what should we make while urgent mode? | |
2723 | * 4.4BSD forces sending single byte of data. We cannot send | |
2724 | * out of window data, because we have SND.NXT==SND.MAX... | |
2725 | * | |
2726 | * Current solution: to send TWO zero-length segments in urgent mode: | |
2727 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is | |
2728 | * out-of-date with SND.UNA-1 to probe window. | |
2729 | */ | |
2730 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent) | |
2731 | { | |
2732 | struct tcp_sock *tp = tcp_sk(sk); | |
2733 | struct sk_buff *skb; | |
2734 | ||
2735 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ | |
2736 | skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); | |
e905a9ed | 2737 | if (skb == NULL) |
1da177e4 LT |
2738 | return -1; |
2739 | ||
2740 | /* Reserve space for headers and set control bits. */ | |
2741 | skb_reserve(skb, MAX_TCP_HEADER); | |
1da177e4 LT |
2742 | /* Use a previous sequence. This should cause the other |
2743 | * end to send an ack. Don't queue or clone SKB, just | |
2744 | * send it. | |
2745 | */ | |
a3433f35 | 2746 | tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); |
1da177e4 | 2747 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
dfb4b9dc | 2748 | return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); |
1da177e4 LT |
2749 | } |
2750 | ||
67edfef7 | 2751 | /* Initiate keepalive or window probe from timer. */ |
1da177e4 LT |
2752 | int tcp_write_wakeup(struct sock *sk) |
2753 | { | |
058dc334 IJ |
2754 | struct tcp_sock *tp = tcp_sk(sk); |
2755 | struct sk_buff *skb; | |
1da177e4 | 2756 | |
058dc334 IJ |
2757 | if (sk->sk_state == TCP_CLOSE) |
2758 | return -1; | |
2759 | ||
2760 | if ((skb = tcp_send_head(sk)) != NULL && | |
2761 | before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { | |
2762 | int err; | |
0c54b85f | 2763 | unsigned int mss = tcp_current_mss(sk); |
058dc334 IJ |
2764 | unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
2765 | ||
2766 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) | |
2767 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; | |
2768 | ||
2769 | /* We are probing the opening of a window | |
2770 | * but the window size is != 0 | |
2771 | * must have been a result SWS avoidance ( sender ) | |
2772 | */ | |
2773 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || | |
2774 | skb->len > mss) { | |
2775 | seg_size = min(seg_size, mss); | |
a3433f35 | 2776 | TCP_SKB_CB(skb)->flags |= TCPHDR_PSH; |
058dc334 IJ |
2777 | if (tcp_fragment(sk, skb, seg_size, mss)) |
2778 | return -1; | |
2779 | } else if (!tcp_skb_pcount(skb)) | |
2780 | tcp_set_skb_tso_segs(sk, skb, mss); | |
2781 | ||
a3433f35 | 2782 | TCP_SKB_CB(skb)->flags |= TCPHDR_PSH; |
058dc334 IJ |
2783 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
2784 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); | |
2785 | if (!err) | |
2786 | tcp_event_new_data_sent(sk, skb); | |
2787 | return err; | |
2788 | } else { | |
33f5f57e | 2789 | if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) |
058dc334 IJ |
2790 | tcp_xmit_probe_skb(sk, 1); |
2791 | return tcp_xmit_probe_skb(sk, 0); | |
1da177e4 | 2792 | } |
1da177e4 LT |
2793 | } |
2794 | ||
2795 | /* A window probe timeout has occurred. If window is not closed send | |
2796 | * a partial packet else a zero probe. | |
2797 | */ | |
2798 | void tcp_send_probe0(struct sock *sk) | |
2799 | { | |
463c84b9 | 2800 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2801 | struct tcp_sock *tp = tcp_sk(sk); |
2802 | int err; | |
2803 | ||
2804 | err = tcp_write_wakeup(sk); | |
2805 | ||
fe067e8a | 2806 | if (tp->packets_out || !tcp_send_head(sk)) { |
1da177e4 | 2807 | /* Cancel probe timer, if it is not required. */ |
6687e988 | 2808 | icsk->icsk_probes_out = 0; |
463c84b9 | 2809 | icsk->icsk_backoff = 0; |
1da177e4 LT |
2810 | return; |
2811 | } | |
2812 | ||
2813 | if (err <= 0) { | |
463c84b9 ACM |
2814 | if (icsk->icsk_backoff < sysctl_tcp_retries2) |
2815 | icsk->icsk_backoff++; | |
6687e988 | 2816 | icsk->icsk_probes_out++; |
e905a9ed | 2817 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
3f421baa ACM |
2818 | min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), |
2819 | TCP_RTO_MAX); | |
1da177e4 LT |
2820 | } else { |
2821 | /* If packet was not sent due to local congestion, | |
6687e988 | 2822 | * do not backoff and do not remember icsk_probes_out. |
1da177e4 LT |
2823 | * Let local senders to fight for local resources. |
2824 | * | |
2825 | * Use accumulated backoff yet. | |
2826 | */ | |
6687e988 ACM |
2827 | if (!icsk->icsk_probes_out) |
2828 | icsk->icsk_probes_out = 1; | |
e905a9ed | 2829 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
463c84b9 | 2830 | min(icsk->icsk_rto << icsk->icsk_backoff, |
3f421baa ACM |
2831 | TCP_RESOURCE_PROBE_INTERVAL), |
2832 | TCP_RTO_MAX); | |
1da177e4 LT |
2833 | } |
2834 | } |