]>
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 | 57 | int sysctl_tcp_mtu_probing __read_mostly = 0; |
97b1ce25 | 58 | int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS; |
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 | { | |
f1987257 ED |
388 | int cookie_size; |
389 | ||
390 | if (desired > 0) | |
bd0388ae WAS |
391 | /* previously specified */ |
392 | return desired; | |
f1987257 ED |
393 | |
394 | cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size); | |
395 | if (cookie_size <= 0) | |
bd0388ae WAS |
396 | /* no default specified */ |
397 | return 0; | |
f1987257 ED |
398 | |
399 | if (cookie_size <= TCP_COOKIE_MIN) | |
bd0388ae WAS |
400 | /* value too small, specify minimum */ |
401 | return TCP_COOKIE_MIN; | |
f1987257 ED |
402 | |
403 | if (cookie_size >= TCP_COOKIE_MAX) | |
bd0388ae WAS |
404 | /* value too large, specify maximum */ |
405 | return TCP_COOKIE_MAX; | |
f1987257 ED |
406 | |
407 | if (cookie_size & 1) | |
bd0388ae | 408 | /* 8-bit multiple, illegal, fix it */ |
f1987257 ED |
409 | cookie_size++; |
410 | ||
411 | return (u8)cookie_size; | |
bd0388ae WAS |
412 | } |
413 | ||
67edfef7 AK |
414 | /* Write previously computed TCP options to the packet. |
415 | * | |
416 | * Beware: Something in the Internet is very sensitive to the ordering of | |
fd6149d3 IJ |
417 | * TCP options, we learned this through the hard way, so be careful here. |
418 | * Luckily we can at least blame others for their non-compliance but from | |
419 | * inter-operatibility perspective it seems that we're somewhat stuck with | |
420 | * the ordering which we have been using if we want to keep working with | |
421 | * those broken things (not that it currently hurts anybody as there isn't | |
422 | * particular reason why the ordering would need to be changed). | |
423 | * | |
424 | * At least SACK_PERM as the first option is known to lead to a disaster | |
425 | * (but it may well be that other scenarios fail similarly). | |
426 | */ | |
33ad798c | 427 | static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, |
bd0388ae WAS |
428 | struct tcp_out_options *opts) |
429 | { | |
430 | u8 options = opts->options; /* mungable copy */ | |
431 | ||
432 | /* Having both authentication and cookies for security is redundant, | |
433 | * and there's certainly not enough room. Instead, the cookie-less | |
434 | * extension variant is proposed. | |
435 | * | |
436 | * Consider the pessimal case with authentication. The options | |
437 | * could look like: | |
438 | * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40 | |
439 | */ | |
440 | if (unlikely(OPTION_MD5 & options)) { | |
441 | if (unlikely(OPTION_COOKIE_EXTENSION & options)) { | |
442 | *ptr++ = htonl((TCPOPT_COOKIE << 24) | | |
443 | (TCPOLEN_COOKIE_BASE << 16) | | |
444 | (TCPOPT_MD5SIG << 8) | | |
445 | TCPOLEN_MD5SIG); | |
446 | } else { | |
447 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
448 | (TCPOPT_NOP << 16) | | |
449 | (TCPOPT_MD5SIG << 8) | | |
450 | TCPOLEN_MD5SIG); | |
451 | } | |
452 | options &= ~OPTION_COOKIE_EXTENSION; | |
453 | /* overload cookie hash location */ | |
454 | opts->hash_location = (__u8 *)ptr; | |
33ad798c | 455 | ptr += 4; |
40efc6fa | 456 | } |
33ad798c | 457 | |
fd6149d3 IJ |
458 | if (unlikely(opts->mss)) { |
459 | *ptr++ = htonl((TCPOPT_MSS << 24) | | |
460 | (TCPOLEN_MSS << 16) | | |
461 | opts->mss); | |
462 | } | |
463 | ||
bd0388ae WAS |
464 | if (likely(OPTION_TS & options)) { |
465 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { | |
33ad798c AL |
466 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
467 | (TCPOLEN_SACK_PERM << 16) | | |
468 | (TCPOPT_TIMESTAMP << 8) | | |
469 | TCPOLEN_TIMESTAMP); | |
bd0388ae | 470 | options &= ~OPTION_SACK_ADVERTISE; |
33ad798c AL |
471 | } else { |
472 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
473 | (TCPOPT_NOP << 16) | | |
474 | (TCPOPT_TIMESTAMP << 8) | | |
475 | TCPOLEN_TIMESTAMP); | |
476 | } | |
477 | *ptr++ = htonl(opts->tsval); | |
478 | *ptr++ = htonl(opts->tsecr); | |
479 | } | |
480 | ||
bd0388ae WAS |
481 | /* Specification requires after timestamp, so do it now. |
482 | * | |
483 | * Consider the pessimal case without authentication. The options | |
484 | * could look like: | |
485 | * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40 | |
486 | */ | |
487 | if (unlikely(OPTION_COOKIE_EXTENSION & options)) { | |
488 | __u8 *cookie_copy = opts->hash_location; | |
489 | u8 cookie_size = opts->hash_size; | |
490 | ||
491 | /* 8-bit multiple handled in tcp_cookie_size_check() above, | |
492 | * and elsewhere. | |
493 | */ | |
494 | if (0x2 & cookie_size) { | |
495 | __u8 *p = (__u8 *)ptr; | |
496 | ||
497 | /* 16-bit multiple */ | |
498 | *p++ = TCPOPT_COOKIE; | |
499 | *p++ = TCPOLEN_COOKIE_BASE + cookie_size; | |
500 | *p++ = *cookie_copy++; | |
501 | *p++ = *cookie_copy++; | |
502 | ptr++; | |
503 | cookie_size -= 2; | |
504 | } else { | |
505 | /* 32-bit multiple */ | |
506 | *ptr++ = htonl(((TCPOPT_NOP << 24) | | |
507 | (TCPOPT_NOP << 16) | | |
508 | (TCPOPT_COOKIE << 8) | | |
509 | TCPOLEN_COOKIE_BASE) + | |
510 | cookie_size); | |
511 | } | |
512 | ||
513 | if (cookie_size > 0) { | |
514 | memcpy(ptr, cookie_copy, cookie_size); | |
515 | ptr += (cookie_size / 4); | |
516 | } | |
517 | } | |
518 | ||
519 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { | |
33ad798c AL |
520 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
521 | (TCPOPT_NOP << 16) | | |
522 | (TCPOPT_SACK_PERM << 8) | | |
523 | TCPOLEN_SACK_PERM); | |
524 | } | |
525 | ||
bd0388ae | 526 | if (unlikely(OPTION_WSCALE & options)) { |
33ad798c AL |
527 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
528 | (TCPOPT_WINDOW << 16) | | |
529 | (TCPOLEN_WINDOW << 8) | | |
530 | opts->ws); | |
531 | } | |
532 | ||
533 | if (unlikely(opts->num_sack_blocks)) { | |
534 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? | |
535 | tp->duplicate_sack : tp->selective_acks; | |
40efc6fa SH |
536 | int this_sack; |
537 | ||
538 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
539 | (TCPOPT_NOP << 16) | | |
540 | (TCPOPT_SACK << 8) | | |
33ad798c | 541 | (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * |
40efc6fa | 542 | TCPOLEN_SACK_PERBLOCK))); |
2de979bd | 543 | |
33ad798c AL |
544 | for (this_sack = 0; this_sack < opts->num_sack_blocks; |
545 | ++this_sack) { | |
40efc6fa SH |
546 | *ptr++ = htonl(sp[this_sack].start_seq); |
547 | *ptr++ = htonl(sp[this_sack].end_seq); | |
548 | } | |
2de979bd | 549 | |
5861f8e5 | 550 | tp->rx_opt.dsack = 0; |
40efc6fa | 551 | } |
33ad798c AL |
552 | } |
553 | ||
67edfef7 AK |
554 | /* Compute TCP options for SYN packets. This is not the final |
555 | * network wire format yet. | |
556 | */ | |
33ad798c AL |
557 | static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb, |
558 | struct tcp_out_options *opts, | |
559 | struct tcp_md5sig_key **md5) { | |
560 | struct tcp_sock *tp = tcp_sk(sk); | |
bd0388ae | 561 | struct tcp_cookie_values *cvp = tp->cookie_values; |
bd0388ae WAS |
562 | unsigned remaining = MAX_TCP_OPTION_SPACE; |
563 | u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ? | |
564 | tcp_cookie_size_check(cvp->cookie_desired) : | |
565 | 0; | |
33ad798c | 566 | |
cfb6eeb4 | 567 | #ifdef CONFIG_TCP_MD5SIG |
33ad798c AL |
568 | *md5 = tp->af_specific->md5_lookup(sk, sk); |
569 | if (*md5) { | |
570 | opts->options |= OPTION_MD5; | |
bd0388ae | 571 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
cfb6eeb4 | 572 | } |
33ad798c AL |
573 | #else |
574 | *md5 = NULL; | |
cfb6eeb4 | 575 | #endif |
33ad798c AL |
576 | |
577 | /* We always get an MSS option. The option bytes which will be seen in | |
578 | * normal data packets should timestamps be used, must be in the MSS | |
579 | * advertised. But we subtract them from tp->mss_cache so that | |
580 | * calculations in tcp_sendmsg are simpler etc. So account for this | |
581 | * fact here if necessary. If we don't do this correctly, as a | |
582 | * receiver we won't recognize data packets as being full sized when we | |
583 | * should, and thus we won't abide by the delayed ACK rules correctly. | |
584 | * SACKs don't matter, we never delay an ACK when we have any of those | |
585 | * going out. */ | |
586 | opts->mss = tcp_advertise_mss(sk); | |
bd0388ae | 587 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c | 588 | |
bb5b7c11 | 589 | if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { |
33ad798c AL |
590 | opts->options |= OPTION_TS; |
591 | opts->tsval = TCP_SKB_CB(skb)->when; | |
592 | opts->tsecr = tp->rx_opt.ts_recent; | |
bd0388ae | 593 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c | 594 | } |
bb5b7c11 | 595 | if (likely(sysctl_tcp_window_scaling)) { |
33ad798c | 596 | opts->ws = tp->rx_opt.rcv_wscale; |
89e95a61 | 597 | opts->options |= OPTION_WSCALE; |
bd0388ae | 598 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 599 | } |
bb5b7c11 | 600 | if (likely(sysctl_tcp_sack)) { |
33ad798c | 601 | opts->options |= OPTION_SACK_ADVERTISE; |
b32d1310 | 602 | if (unlikely(!(OPTION_TS & opts->options))) |
bd0388ae | 603 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c AL |
604 | } |
605 | ||
bd0388ae WAS |
606 | /* Note that timestamps are required by the specification. |
607 | * | |
608 | * Odd numbers of bytes are prohibited by the specification, ensuring | |
609 | * that the cookie is 16-bit aligned, and the resulting cookie pair is | |
610 | * 32-bit aligned. | |
611 | */ | |
612 | if (*md5 == NULL && | |
613 | (OPTION_TS & opts->options) && | |
614 | cookie_size > 0) { | |
615 | int need = TCPOLEN_COOKIE_BASE + cookie_size; | |
616 | ||
617 | if (0x2 & need) { | |
618 | /* 32-bit multiple */ | |
619 | need += 2; /* NOPs */ | |
620 | ||
621 | if (need > remaining) { | |
622 | /* try shrinking cookie to fit */ | |
623 | cookie_size -= 2; | |
624 | need -= 4; | |
625 | } | |
626 | } | |
627 | while (need > remaining && TCP_COOKIE_MIN <= cookie_size) { | |
628 | cookie_size -= 4; | |
629 | need -= 4; | |
630 | } | |
631 | if (TCP_COOKIE_MIN <= cookie_size) { | |
632 | opts->options |= OPTION_COOKIE_EXTENSION; | |
633 | opts->hash_location = (__u8 *)&cvp->cookie_pair[0]; | |
634 | opts->hash_size = cookie_size; | |
635 | ||
636 | /* Remember for future incarnations. */ | |
637 | cvp->cookie_desired = cookie_size; | |
638 | ||
639 | if (cvp->cookie_desired != cvp->cookie_pair_size) { | |
640 | /* Currently use random bytes as a nonce, | |
641 | * assuming these are completely unpredictable | |
642 | * by hostile users of the same system. | |
643 | */ | |
644 | get_random_bytes(&cvp->cookie_pair[0], | |
645 | cookie_size); | |
646 | cvp->cookie_pair_size = cookie_size; | |
647 | } | |
648 | ||
649 | remaining -= need; | |
650 | } | |
651 | } | |
652 | return MAX_TCP_OPTION_SPACE - remaining; | |
40efc6fa SH |
653 | } |
654 | ||
67edfef7 | 655 | /* Set up TCP options for SYN-ACKs. */ |
33ad798c AL |
656 | static unsigned tcp_synack_options(struct sock *sk, |
657 | struct request_sock *req, | |
658 | unsigned mss, struct sk_buff *skb, | |
659 | struct tcp_out_options *opts, | |
4957faad WAS |
660 | struct tcp_md5sig_key **md5, |
661 | struct tcp_extend_values *xvp) | |
662 | { | |
33ad798c | 663 | struct inet_request_sock *ireq = inet_rsk(req); |
4957faad WAS |
664 | unsigned remaining = MAX_TCP_OPTION_SPACE; |
665 | u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ? | |
666 | xvp->cookie_plus : | |
667 | 0; | |
33ad798c | 668 | |
cfb6eeb4 | 669 | #ifdef CONFIG_TCP_MD5SIG |
33ad798c AL |
670 | *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); |
671 | if (*md5) { | |
672 | opts->options |= OPTION_MD5; | |
4957faad WAS |
673 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
674 | ||
675 | /* We can't fit any SACK blocks in a packet with MD5 + TS | |
676 | * options. There was discussion about disabling SACK | |
677 | * rather than TS in order to fit in better with old, | |
678 | * buggy kernels, but that was deemed to be unnecessary. | |
679 | */ | |
de213e5e | 680 | ireq->tstamp_ok &= !ireq->sack_ok; |
cfb6eeb4 | 681 | } |
33ad798c AL |
682 | #else |
683 | *md5 = NULL; | |
cfb6eeb4 | 684 | #endif |
33ad798c | 685 | |
4957faad | 686 | /* We always send an MSS option. */ |
33ad798c | 687 | opts->mss = mss; |
4957faad | 688 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c AL |
689 | |
690 | if (likely(ireq->wscale_ok)) { | |
691 | opts->ws = ireq->rcv_wscale; | |
89e95a61 | 692 | opts->options |= OPTION_WSCALE; |
4957faad | 693 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 694 | } |
de213e5e | 695 | if (likely(ireq->tstamp_ok)) { |
33ad798c AL |
696 | opts->options |= OPTION_TS; |
697 | opts->tsval = TCP_SKB_CB(skb)->when; | |
698 | opts->tsecr = req->ts_recent; | |
4957faad | 699 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c AL |
700 | } |
701 | if (likely(ireq->sack_ok)) { | |
702 | opts->options |= OPTION_SACK_ADVERTISE; | |
de213e5e | 703 | if (unlikely(!ireq->tstamp_ok)) |
4957faad | 704 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c AL |
705 | } |
706 | ||
4957faad WAS |
707 | /* Similar rationale to tcp_syn_options() applies here, too. |
708 | * If the <SYN> options fit, the same options should fit now! | |
709 | */ | |
710 | if (*md5 == NULL && | |
de213e5e | 711 | ireq->tstamp_ok && |
4957faad WAS |
712 | cookie_plus > TCPOLEN_COOKIE_BASE) { |
713 | int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */ | |
714 | ||
715 | if (0x2 & need) { | |
716 | /* 32-bit multiple */ | |
717 | need += 2; /* NOPs */ | |
718 | } | |
719 | if (need <= remaining) { | |
720 | opts->options |= OPTION_COOKIE_EXTENSION; | |
721 | opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE; | |
722 | remaining -= need; | |
723 | } else { | |
724 | /* There's no error return, so flag it. */ | |
725 | xvp->cookie_out_never = 1; /* true */ | |
726 | opts->hash_size = 0; | |
727 | } | |
728 | } | |
729 | return MAX_TCP_OPTION_SPACE - remaining; | |
33ad798c AL |
730 | } |
731 | ||
67edfef7 AK |
732 | /* Compute TCP options for ESTABLISHED sockets. This is not the |
733 | * final wire format yet. | |
734 | */ | |
33ad798c AL |
735 | static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb, |
736 | struct tcp_out_options *opts, | |
737 | struct tcp_md5sig_key **md5) { | |
738 | struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; | |
739 | struct tcp_sock *tp = tcp_sk(sk); | |
740 | unsigned size = 0; | |
cabeccbd | 741 | unsigned int eff_sacks; |
33ad798c AL |
742 | |
743 | #ifdef CONFIG_TCP_MD5SIG | |
744 | *md5 = tp->af_specific->md5_lookup(sk, sk); | |
745 | if (unlikely(*md5)) { | |
746 | opts->options |= OPTION_MD5; | |
747 | size += TCPOLEN_MD5SIG_ALIGNED; | |
748 | } | |
749 | #else | |
750 | *md5 = NULL; | |
751 | #endif | |
752 | ||
753 | if (likely(tp->rx_opt.tstamp_ok)) { | |
754 | opts->options |= OPTION_TS; | |
755 | opts->tsval = tcb ? tcb->when : 0; | |
756 | opts->tsecr = tp->rx_opt.ts_recent; | |
757 | size += TCPOLEN_TSTAMP_ALIGNED; | |
758 | } | |
759 | ||
cabeccbd IJ |
760 | eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; |
761 | if (unlikely(eff_sacks)) { | |
33ad798c AL |
762 | const unsigned remaining = MAX_TCP_OPTION_SPACE - size; |
763 | opts->num_sack_blocks = | |
cabeccbd | 764 | min_t(unsigned, eff_sacks, |
33ad798c AL |
765 | (remaining - TCPOLEN_SACK_BASE_ALIGNED) / |
766 | TCPOLEN_SACK_PERBLOCK); | |
767 | size += TCPOLEN_SACK_BASE_ALIGNED + | |
768 | opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; | |
769 | } | |
770 | ||
771 | return size; | |
40efc6fa | 772 | } |
1da177e4 LT |
773 | |
774 | /* This routine actually transmits TCP packets queued in by | |
775 | * tcp_do_sendmsg(). This is used by both the initial | |
776 | * transmission and possible later retransmissions. | |
777 | * All SKB's seen here are completely headerless. It is our | |
778 | * job to build the TCP header, and pass the packet down to | |
779 | * IP so it can do the same plus pass the packet off to the | |
780 | * device. | |
781 | * | |
782 | * We are working here with either a clone of the original | |
783 | * SKB, or a fresh unique copy made by the retransmit engine. | |
784 | */ | |
056834d9 IJ |
785 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, |
786 | gfp_t gfp_mask) | |
1da177e4 | 787 | { |
dfb4b9dc DM |
788 | const struct inet_connection_sock *icsk = inet_csk(sk); |
789 | struct inet_sock *inet; | |
790 | struct tcp_sock *tp; | |
791 | struct tcp_skb_cb *tcb; | |
33ad798c AL |
792 | struct tcp_out_options opts; |
793 | unsigned tcp_options_size, tcp_header_size; | |
cfb6eeb4 | 794 | struct tcp_md5sig_key *md5; |
dfb4b9dc | 795 | struct tcphdr *th; |
dfb4b9dc DM |
796 | int err; |
797 | ||
798 | BUG_ON(!skb || !tcp_skb_pcount(skb)); | |
799 | ||
800 | /* If congestion control is doing timestamping, we must | |
801 | * take such a timestamp before we potentially clone/copy. | |
802 | */ | |
164891aa | 803 | if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) |
dfb4b9dc DM |
804 | __net_timestamp(skb); |
805 | ||
806 | if (likely(clone_it)) { | |
807 | if (unlikely(skb_cloned(skb))) | |
808 | skb = pskb_copy(skb, gfp_mask); | |
809 | else | |
810 | skb = skb_clone(skb, gfp_mask); | |
811 | if (unlikely(!skb)) | |
812 | return -ENOBUFS; | |
813 | } | |
1da177e4 | 814 | |
dfb4b9dc DM |
815 | inet = inet_sk(sk); |
816 | tp = tcp_sk(sk); | |
817 | tcb = TCP_SKB_CB(skb); | |
33ad798c | 818 | memset(&opts, 0, sizeof(opts)); |
1da177e4 | 819 | |
a3433f35 | 820 | if (unlikely(tcb->flags & TCPHDR_SYN)) |
33ad798c AL |
821 | tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); |
822 | else | |
823 | tcp_options_size = tcp_established_options(sk, skb, &opts, | |
824 | &md5); | |
825 | tcp_header_size = tcp_options_size + sizeof(struct tcphdr); | |
e905a9ed | 826 | |
3853b584 | 827 | if (tcp_packets_in_flight(tp) == 0) { |
dfb4b9dc | 828 | tcp_ca_event(sk, CA_EVENT_TX_START); |
3853b584 TH |
829 | skb->ooo_okay = 1; |
830 | } else | |
831 | skb->ooo_okay = 0; | |
dfb4b9dc | 832 | |
aa8223c7 ACM |
833 | skb_push(skb, tcp_header_size); |
834 | skb_reset_transport_header(skb); | |
e89862f4 | 835 | skb_set_owner_w(skb, sk); |
dfb4b9dc DM |
836 | |
837 | /* Build TCP header and checksum it. */ | |
aa8223c7 | 838 | th = tcp_hdr(skb); |
c720c7e8 ED |
839 | th->source = inet->inet_sport; |
840 | th->dest = inet->inet_dport; | |
dfb4b9dc DM |
841 | th->seq = htonl(tcb->seq); |
842 | th->ack_seq = htonl(tp->rcv_nxt); | |
df7a3b07 | 843 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
dfb4b9dc DM |
844 | tcb->flags); |
845 | ||
a3433f35 | 846 | if (unlikely(tcb->flags & TCPHDR_SYN)) { |
dfb4b9dc DM |
847 | /* RFC1323: The window in SYN & SYN/ACK segments |
848 | * is never scaled. | |
849 | */ | |
600ff0c2 | 850 | th->window = htons(min(tp->rcv_wnd, 65535U)); |
dfb4b9dc DM |
851 | } else { |
852 | th->window = htons(tcp_select_window(sk)); | |
853 | } | |
854 | th->check = 0; | |
855 | th->urg_ptr = 0; | |
1da177e4 | 856 | |
33f5f57e | 857 | /* The urg_mode check is necessary during a below snd_una win probe */ |
7691367d HX |
858 | if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { |
859 | if (before(tp->snd_up, tcb->seq + 0x10000)) { | |
860 | th->urg_ptr = htons(tp->snd_up - tcb->seq); | |
861 | th->urg = 1; | |
862 | } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { | |
0eae88f3 | 863 | th->urg_ptr = htons(0xFFFF); |
7691367d HX |
864 | th->urg = 1; |
865 | } | |
dfb4b9dc | 866 | } |
1da177e4 | 867 | |
bd0388ae | 868 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
a3433f35 | 869 | if (likely((tcb->flags & TCPHDR_SYN) == 0)) |
9e412ba7 | 870 | TCP_ECN_send(sk, skb, tcp_header_size); |
1da177e4 | 871 | |
cfb6eeb4 YH |
872 | #ifdef CONFIG_TCP_MD5SIG |
873 | /* Calculate the MD5 hash, as we have all we need now */ | |
874 | if (md5) { | |
a465419b | 875 | sk_nocaps_add(sk, NETIF_F_GSO_MASK); |
bd0388ae | 876 | tp->af_specific->calc_md5_hash(opts.hash_location, |
49a72dfb | 877 | md5, sk, NULL, skb); |
cfb6eeb4 YH |
878 | } |
879 | #endif | |
880 | ||
bb296246 | 881 | icsk->icsk_af_ops->send_check(sk, skb); |
1da177e4 | 882 | |
a3433f35 | 883 | if (likely(tcb->flags & TCPHDR_ACK)) |
dfb4b9dc | 884 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); |
1da177e4 | 885 | |
dfb4b9dc DM |
886 | if (skb->len != tcp_header_size) |
887 | tcp_event_data_sent(tp, skb, sk); | |
1da177e4 | 888 | |
bd37a088 | 889 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
aa2ea058 TH |
890 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, |
891 | tcp_skb_pcount(skb)); | |
1da177e4 | 892 | |
4e15ed4d | 893 | err = icsk->icsk_af_ops->queue_xmit(skb); |
83de47cd | 894 | if (likely(err <= 0)) |
dfb4b9dc DM |
895 | return err; |
896 | ||
3cfe3baa | 897 | tcp_enter_cwr(sk, 1); |
dfb4b9dc | 898 | |
b9df3cb8 | 899 | return net_xmit_eval(err); |
1da177e4 LT |
900 | } |
901 | ||
67edfef7 | 902 | /* This routine just queues the buffer for sending. |
1da177e4 LT |
903 | * |
904 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, | |
905 | * otherwise socket can stall. | |
906 | */ | |
907 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) | |
908 | { | |
909 | struct tcp_sock *tp = tcp_sk(sk); | |
910 | ||
911 | /* Advance write_seq and place onto the write_queue. */ | |
912 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; | |
913 | skb_header_release(skb); | |
fe067e8a | 914 | tcp_add_write_queue_tail(sk, skb); |
3ab224be HA |
915 | sk->sk_wmem_queued += skb->truesize; |
916 | sk_mem_charge(sk, skb->truesize); | |
1da177e4 LT |
917 | } |
918 | ||
67edfef7 | 919 | /* Initialize TSO segments for a packet. */ |
056834d9 IJ |
920 | static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, |
921 | unsigned int mss_now) | |
f6302d1d | 922 | { |
8e5b9dda HX |
923 | if (skb->len <= mss_now || !sk_can_gso(sk) || |
924 | skb->ip_summed == CHECKSUM_NONE) { | |
f6302d1d DM |
925 | /* Avoid the costly divide in the normal |
926 | * non-TSO case. | |
927 | */ | |
7967168c HX |
928 | skb_shinfo(skb)->gso_segs = 1; |
929 | skb_shinfo(skb)->gso_size = 0; | |
930 | skb_shinfo(skb)->gso_type = 0; | |
f6302d1d | 931 | } else { |
356f89e1 | 932 | skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); |
7967168c | 933 | skb_shinfo(skb)->gso_size = mss_now; |
bcd76111 | 934 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
1da177e4 LT |
935 | } |
936 | } | |
937 | ||
91fed7a1 | 938 | /* When a modification to fackets out becomes necessary, we need to check |
68f8353b | 939 | * skb is counted to fackets_out or not. |
91fed7a1 | 940 | */ |
a47e5a98 | 941 | static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb, |
91fed7a1 IJ |
942 | int decr) |
943 | { | |
a47e5a98 IJ |
944 | struct tcp_sock *tp = tcp_sk(sk); |
945 | ||
dc86967b | 946 | if (!tp->sacked_out || tcp_is_reno(tp)) |
91fed7a1 IJ |
947 | return; |
948 | ||
6859d494 | 949 | if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) |
91fed7a1 | 950 | tp->fackets_out -= decr; |
91fed7a1 IJ |
951 | } |
952 | ||
797108d1 IJ |
953 | /* Pcount in the middle of the write queue got changed, we need to do various |
954 | * tweaks to fix counters | |
955 | */ | |
956 | static void tcp_adjust_pcount(struct sock *sk, struct sk_buff *skb, int decr) | |
957 | { | |
958 | struct tcp_sock *tp = tcp_sk(sk); | |
959 | ||
960 | tp->packets_out -= decr; | |
961 | ||
962 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
963 | tp->sacked_out -= decr; | |
964 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
965 | tp->retrans_out -= decr; | |
966 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) | |
967 | tp->lost_out -= decr; | |
968 | ||
969 | /* Reno case is special. Sigh... */ | |
970 | if (tcp_is_reno(tp) && decr > 0) | |
971 | tp->sacked_out -= min_t(u32, tp->sacked_out, decr); | |
972 | ||
973 | tcp_adjust_fackets_out(sk, skb, decr); | |
974 | ||
975 | if (tp->lost_skb_hint && | |
976 | before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && | |
52cf3cc8 | 977 | (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) |
797108d1 IJ |
978 | tp->lost_cnt_hint -= decr; |
979 | ||
980 | tcp_verify_left_out(tp); | |
981 | } | |
982 | ||
1da177e4 LT |
983 | /* Function to create two new TCP segments. Shrinks the given segment |
984 | * to the specified size and appends a new segment with the rest of the | |
e905a9ed | 985 | * packet to the list. This won't be called frequently, I hope. |
1da177e4 LT |
986 | * Remember, these are still headerless SKBs at this point. |
987 | */ | |
056834d9 IJ |
988 | int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, |
989 | unsigned int mss_now) | |
1da177e4 LT |
990 | { |
991 | struct tcp_sock *tp = tcp_sk(sk); | |
992 | struct sk_buff *buff; | |
6475be16 | 993 | int nsize, old_factor; |
b60b49ea | 994 | int nlen; |
9ce01461 | 995 | u8 flags; |
1da177e4 | 996 | |
b2cc99f0 | 997 | BUG_ON(len > skb->len); |
6a438bbe | 998 | |
1da177e4 LT |
999 | nsize = skb_headlen(skb) - len; |
1000 | if (nsize < 0) | |
1001 | nsize = 0; | |
1002 | ||
1003 | if (skb_cloned(skb) && | |
1004 | skb_is_nonlinear(skb) && | |
1005 | pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) | |
1006 | return -ENOMEM; | |
1007 | ||
1008 | /* Get a new skb... force flag on. */ | |
1009 | buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); | |
1010 | if (buff == NULL) | |
1011 | return -ENOMEM; /* We'll just try again later. */ | |
ef5cb973 | 1012 | |
3ab224be HA |
1013 | sk->sk_wmem_queued += buff->truesize; |
1014 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea HX |
1015 | nlen = skb->len - len - nsize; |
1016 | buff->truesize += nlen; | |
1017 | skb->truesize -= nlen; | |
1da177e4 LT |
1018 | |
1019 | /* Correct the sequence numbers. */ | |
1020 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1021 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1022 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1023 | ||
1024 | /* PSH and FIN should only be set in the second packet. */ | |
1025 | flags = TCP_SKB_CB(skb)->flags; | |
a3433f35 | 1026 | TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); |
1da177e4 | 1027 | TCP_SKB_CB(buff)->flags = flags; |
e14c3caf | 1028 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
1da177e4 | 1029 | |
84fa7933 | 1030 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { |
1da177e4 | 1031 | /* Copy and checksum data tail into the new buffer. */ |
056834d9 IJ |
1032 | buff->csum = csum_partial_copy_nocheck(skb->data + len, |
1033 | skb_put(buff, nsize), | |
1da177e4 LT |
1034 | nsize, 0); |
1035 | ||
1036 | skb_trim(skb, len); | |
1037 | ||
1038 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); | |
1039 | } else { | |
84fa7933 | 1040 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
1041 | skb_split(skb, buff, len); |
1042 | } | |
1043 | ||
1044 | buff->ip_summed = skb->ip_summed; | |
1045 | ||
1046 | /* Looks stupid, but our code really uses when of | |
1047 | * skbs, which it never sent before. --ANK | |
1048 | */ | |
1049 | TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; | |
a61bbcf2 | 1050 | buff->tstamp = skb->tstamp; |
1da177e4 | 1051 | |
6475be16 DM |
1052 | old_factor = tcp_skb_pcount(skb); |
1053 | ||
1da177e4 | 1054 | /* Fix up tso_factor for both original and new SKB. */ |
846998ae DM |
1055 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
1056 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
1da177e4 | 1057 | |
6475be16 DM |
1058 | /* If this packet has been sent out already, we must |
1059 | * adjust the various packet counters. | |
1060 | */ | |
cf0b450c | 1061 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
6475be16 DM |
1062 | int diff = old_factor - tcp_skb_pcount(skb) - |
1063 | tcp_skb_pcount(buff); | |
1da177e4 | 1064 | |
797108d1 IJ |
1065 | if (diff) |
1066 | tcp_adjust_pcount(sk, skb, diff); | |
1da177e4 LT |
1067 | } |
1068 | ||
1069 | /* Link BUFF into the send queue. */ | |
f44b5271 | 1070 | skb_header_release(buff); |
fe067e8a | 1071 | tcp_insert_write_queue_after(skb, buff, sk); |
1da177e4 LT |
1072 | |
1073 | return 0; | |
1074 | } | |
1075 | ||
1076 | /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c | |
1077 | * eventually). The difference is that pulled data not copied, but | |
1078 | * immediately discarded. | |
1079 | */ | |
f2911969 | 1080 | static void __pskb_trim_head(struct sk_buff *skb, int len) |
1da177e4 LT |
1081 | { |
1082 | int i, k, eat; | |
1083 | ||
1084 | eat = len; | |
1085 | k = 0; | |
056834d9 | 1086 | for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { |
1da177e4 LT |
1087 | if (skb_shinfo(skb)->frags[i].size <= eat) { |
1088 | put_page(skb_shinfo(skb)->frags[i].page); | |
1089 | eat -= skb_shinfo(skb)->frags[i].size; | |
1090 | } else { | |
1091 | skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; | |
1092 | if (eat) { | |
1093 | skb_shinfo(skb)->frags[k].page_offset += eat; | |
1094 | skb_shinfo(skb)->frags[k].size -= eat; | |
1095 | eat = 0; | |
1096 | } | |
1097 | k++; | |
1098 | } | |
1099 | } | |
1100 | skb_shinfo(skb)->nr_frags = k; | |
1101 | ||
27a884dc | 1102 | skb_reset_tail_pointer(skb); |
1da177e4 LT |
1103 | skb->data_len -= len; |
1104 | skb->len = skb->data_len; | |
1da177e4 LT |
1105 | } |
1106 | ||
67edfef7 | 1107 | /* Remove acked data from a packet in the transmit queue. */ |
1da177e4 LT |
1108 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) |
1109 | { | |
056834d9 | 1110 | if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) |
1da177e4 LT |
1111 | return -ENOMEM; |
1112 | ||
f2911969 HX |
1113 | /* If len == headlen, we avoid __skb_pull to preserve alignment. */ |
1114 | if (unlikely(len < skb_headlen(skb))) | |
1da177e4 | 1115 | __skb_pull(skb, len); |
f2911969 HX |
1116 | else |
1117 | __pskb_trim_head(skb, len - skb_headlen(skb)); | |
1da177e4 LT |
1118 | |
1119 | TCP_SKB_CB(skb)->seq += len; | |
84fa7933 | 1120 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
1121 | |
1122 | skb->truesize -= len; | |
1123 | sk->sk_wmem_queued -= len; | |
3ab224be | 1124 | sk_mem_uncharge(sk, len); |
1da177e4 LT |
1125 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); |
1126 | ||
1127 | /* Any change of skb->len requires recalculation of tso | |
1128 | * factor and mss. | |
1129 | */ | |
1130 | if (tcp_skb_pcount(skb) > 1) | |
0c54b85f | 1131 | tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk)); |
1da177e4 LT |
1132 | |
1133 | return 0; | |
1134 | } | |
1135 | ||
67edfef7 | 1136 | /* Calculate MSS. Not accounting for SACKs here. */ |
5d424d5a JH |
1137 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) |
1138 | { | |
1139 | struct tcp_sock *tp = tcp_sk(sk); | |
1140 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1141 | int mss_now; | |
1142 | ||
1143 | /* Calculate base mss without TCP options: | |
1144 | It is MMS_S - sizeof(tcphdr) of rfc1122 | |
1145 | */ | |
1146 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); | |
1147 | ||
1148 | /* Clamp it (mss_clamp does not include tcp options) */ | |
1149 | if (mss_now > tp->rx_opt.mss_clamp) | |
1150 | mss_now = tp->rx_opt.mss_clamp; | |
1151 | ||
1152 | /* Now subtract optional transport overhead */ | |
1153 | mss_now -= icsk->icsk_ext_hdr_len; | |
1154 | ||
1155 | /* Then reserve room for full set of TCP options and 8 bytes of data */ | |
1156 | if (mss_now < 48) | |
1157 | mss_now = 48; | |
1158 | ||
1159 | /* Now subtract TCP options size, not including SACKs */ | |
1160 | mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); | |
1161 | ||
1162 | return mss_now; | |
1163 | } | |
1164 | ||
1165 | /* Inverse of above */ | |
1166 | int tcp_mss_to_mtu(struct sock *sk, int mss) | |
1167 | { | |
1168 | struct tcp_sock *tp = tcp_sk(sk); | |
1169 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1170 | int mtu; | |
1171 | ||
1172 | mtu = mss + | |
1173 | tp->tcp_header_len + | |
1174 | icsk->icsk_ext_hdr_len + | |
1175 | icsk->icsk_af_ops->net_header_len; | |
1176 | ||
1177 | return mtu; | |
1178 | } | |
1179 | ||
67edfef7 | 1180 | /* MTU probing init per socket */ |
5d424d5a JH |
1181 | void tcp_mtup_init(struct sock *sk) |
1182 | { | |
1183 | struct tcp_sock *tp = tcp_sk(sk); | |
1184 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1185 | ||
1186 | icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; | |
1187 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + | |
e905a9ed | 1188 | icsk->icsk_af_ops->net_header_len; |
5d424d5a JH |
1189 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); |
1190 | icsk->icsk_mtup.probe_size = 0; | |
1191 | } | |
4bc2f18b | 1192 | EXPORT_SYMBOL(tcp_mtup_init); |
5d424d5a | 1193 | |
1da177e4 LT |
1194 | /* This function synchronize snd mss to current pmtu/exthdr set. |
1195 | ||
1196 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts | |
1197 | for TCP options, but includes only bare TCP header. | |
1198 | ||
1199 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. | |
caa20d9a | 1200 | It is minimum of user_mss and mss received with SYN. |
1da177e4 LT |
1201 | It also does not include TCP options. |
1202 | ||
d83d8461 | 1203 | inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. |
1da177e4 LT |
1204 | |
1205 | tp->mss_cache is current effective sending mss, including | |
1206 | all tcp options except for SACKs. It is evaluated, | |
1207 | taking into account current pmtu, but never exceeds | |
1208 | tp->rx_opt.mss_clamp. | |
1209 | ||
1210 | NOTE1. rfc1122 clearly states that advertised MSS | |
1211 | DOES NOT include either tcp or ip options. | |
1212 | ||
d83d8461 ACM |
1213 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
1214 | are READ ONLY outside this function. --ANK (980731) | |
1da177e4 | 1215 | */ |
1da177e4 LT |
1216 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) |
1217 | { | |
1218 | struct tcp_sock *tp = tcp_sk(sk); | |
d83d8461 | 1219 | struct inet_connection_sock *icsk = inet_csk(sk); |
5d424d5a | 1220 | int mss_now; |
1da177e4 | 1221 | |
5d424d5a JH |
1222 | if (icsk->icsk_mtup.search_high > pmtu) |
1223 | icsk->icsk_mtup.search_high = pmtu; | |
1da177e4 | 1224 | |
5d424d5a | 1225 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
409d22b4 | 1226 | mss_now = tcp_bound_to_half_wnd(tp, mss_now); |
1da177e4 LT |
1227 | |
1228 | /* And store cached results */ | |
d83d8461 | 1229 | icsk->icsk_pmtu_cookie = pmtu; |
5d424d5a JH |
1230 | if (icsk->icsk_mtup.enabled) |
1231 | mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); | |
c1b4a7e6 | 1232 | tp->mss_cache = mss_now; |
1da177e4 LT |
1233 | |
1234 | return mss_now; | |
1235 | } | |
4bc2f18b | 1236 | EXPORT_SYMBOL(tcp_sync_mss); |
1da177e4 LT |
1237 | |
1238 | /* Compute the current effective MSS, taking SACKs and IP options, | |
1239 | * and even PMTU discovery events into account. | |
1da177e4 | 1240 | */ |
0c54b85f | 1241 | unsigned int tcp_current_mss(struct sock *sk) |
1da177e4 LT |
1242 | { |
1243 | struct tcp_sock *tp = tcp_sk(sk); | |
1244 | struct dst_entry *dst = __sk_dst_get(sk); | |
c1b4a7e6 | 1245 | u32 mss_now; |
33ad798c AL |
1246 | unsigned header_len; |
1247 | struct tcp_out_options opts; | |
1248 | struct tcp_md5sig_key *md5; | |
c1b4a7e6 DM |
1249 | |
1250 | mss_now = tp->mss_cache; | |
1251 | ||
1da177e4 LT |
1252 | if (dst) { |
1253 | u32 mtu = dst_mtu(dst); | |
d83d8461 | 1254 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
1da177e4 LT |
1255 | mss_now = tcp_sync_mss(sk, mtu); |
1256 | } | |
1257 | ||
33ad798c AL |
1258 | header_len = tcp_established_options(sk, NULL, &opts, &md5) + |
1259 | sizeof(struct tcphdr); | |
1260 | /* The mss_cache is sized based on tp->tcp_header_len, which assumes | |
1261 | * some common options. If this is an odd packet (because we have SACK | |
1262 | * blocks etc) then our calculated header_len will be different, and | |
1263 | * we have to adjust mss_now correspondingly */ | |
1264 | if (header_len != tp->tcp_header_len) { | |
1265 | int delta = (int) header_len - tp->tcp_header_len; | |
1266 | mss_now -= delta; | |
1267 | } | |
cfb6eeb4 | 1268 | |
1da177e4 LT |
1269 | return mss_now; |
1270 | } | |
1271 | ||
a762a980 | 1272 | /* Congestion window validation. (RFC2861) */ |
9e412ba7 | 1273 | static void tcp_cwnd_validate(struct sock *sk) |
a762a980 | 1274 | { |
9e412ba7 | 1275 | struct tcp_sock *tp = tcp_sk(sk); |
a762a980 | 1276 | |
d436d686 | 1277 | if (tp->packets_out >= tp->snd_cwnd) { |
a762a980 DM |
1278 | /* Network is feed fully. */ |
1279 | tp->snd_cwnd_used = 0; | |
1280 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1281 | } else { | |
1282 | /* Network starves. */ | |
1283 | if (tp->packets_out > tp->snd_cwnd_used) | |
1284 | tp->snd_cwnd_used = tp->packets_out; | |
1285 | ||
15d33c07 DM |
1286 | if (sysctl_tcp_slow_start_after_idle && |
1287 | (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) | |
a762a980 DM |
1288 | tcp_cwnd_application_limited(sk); |
1289 | } | |
1290 | } | |
1291 | ||
0e3a4803 IJ |
1292 | /* Returns the portion of skb which can be sent right away without |
1293 | * introducing MSS oddities to segment boundaries. In rare cases where | |
1294 | * mss_now != mss_cache, we will request caller to create a small skb | |
1295 | * per input skb which could be mostly avoided here (if desired). | |
5ea3a748 IJ |
1296 | * |
1297 | * We explicitly want to create a request for splitting write queue tail | |
1298 | * to a small skb for Nagle purposes while avoiding unnecessary modulos, | |
1299 | * thus all the complexity (cwnd_len is always MSS multiple which we | |
1300 | * return whenever allowed by the other factors). Basically we need the | |
1301 | * modulo only when the receiver window alone is the limiting factor or | |
1302 | * when we would be allowed to send the split-due-to-Nagle skb fully. | |
0e3a4803 IJ |
1303 | */ |
1304 | static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb, | |
056834d9 | 1305 | unsigned int mss_now, unsigned int cwnd) |
c1b4a7e6 | 1306 | { |
0e3a4803 IJ |
1307 | struct tcp_sock *tp = tcp_sk(sk); |
1308 | u32 needed, window, cwnd_len; | |
c1b4a7e6 | 1309 | |
90840def | 1310 | window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
c1b4a7e6 | 1311 | cwnd_len = mss_now * cwnd; |
0e3a4803 IJ |
1312 | |
1313 | if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk))) | |
1314 | return cwnd_len; | |
1315 | ||
5ea3a748 IJ |
1316 | needed = min(skb->len, window); |
1317 | ||
17515408 | 1318 | if (cwnd_len <= needed) |
0e3a4803 IJ |
1319 | return cwnd_len; |
1320 | ||
0e3a4803 | 1321 | return needed - needed % mss_now; |
c1b4a7e6 DM |
1322 | } |
1323 | ||
1324 | /* Can at least one segment of SKB be sent right now, according to the | |
1325 | * congestion window rules? If so, return how many segments are allowed. | |
1326 | */ | |
056834d9 IJ |
1327 | static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, |
1328 | struct sk_buff *skb) | |
c1b4a7e6 DM |
1329 | { |
1330 | u32 in_flight, cwnd; | |
1331 | ||
1332 | /* Don't be strict about the congestion window for the final FIN. */ | |
a3433f35 | 1333 | if ((TCP_SKB_CB(skb)->flags & TCPHDR_FIN) && tcp_skb_pcount(skb) == 1) |
c1b4a7e6 DM |
1334 | return 1; |
1335 | ||
1336 | in_flight = tcp_packets_in_flight(tp); | |
1337 | cwnd = tp->snd_cwnd; | |
1338 | if (in_flight < cwnd) | |
1339 | return (cwnd - in_flight); | |
1340 | ||
1341 | return 0; | |
1342 | } | |
1343 | ||
67edfef7 AK |
1344 | /* Intialize TSO state of a skb. |
1345 | * This must be invoked the first time we consider transmitting | |
c1b4a7e6 DM |
1346 | * SKB onto the wire. |
1347 | */ | |
056834d9 IJ |
1348 | static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, |
1349 | unsigned int mss_now) | |
c1b4a7e6 DM |
1350 | { |
1351 | int tso_segs = tcp_skb_pcount(skb); | |
1352 | ||
f8269a49 | 1353 | if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { |
846998ae | 1354 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
c1b4a7e6 DM |
1355 | tso_segs = tcp_skb_pcount(skb); |
1356 | } | |
1357 | return tso_segs; | |
1358 | } | |
1359 | ||
67edfef7 | 1360 | /* Minshall's variant of the Nagle send check. */ |
c1b4a7e6 DM |
1361 | static inline int tcp_minshall_check(const struct tcp_sock *tp) |
1362 | { | |
09cb105e | 1363 | return after(tp->snd_sml, tp->snd_una) && |
c1b4a7e6 DM |
1364 | !after(tp->snd_sml, tp->snd_nxt); |
1365 | } | |
1366 | ||
1367 | /* Return 0, if packet can be sent now without violation Nagle's rules: | |
1368 | * 1. It is full sized. | |
1369 | * 2. Or it contains FIN. (already checked by caller) | |
1370 | * 3. Or TCP_NODELAY was set. | |
1371 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. | |
1372 | * With Minshall's modification: all sent small packets are ACKed. | |
1373 | */ | |
c1b4a7e6 | 1374 | static inline int tcp_nagle_check(const struct tcp_sock *tp, |
e905a9ed | 1375 | const struct sk_buff *skb, |
c1b4a7e6 DM |
1376 | unsigned mss_now, int nonagle) |
1377 | { | |
a02cec21 | 1378 | return skb->len < mss_now && |
056834d9 | 1379 | ((nonagle & TCP_NAGLE_CORK) || |
a02cec21 | 1380 | (!nonagle && tp->packets_out && tcp_minshall_check(tp))); |
c1b4a7e6 DM |
1381 | } |
1382 | ||
1383 | /* Return non-zero if the Nagle test allows this packet to be | |
1384 | * sent now. | |
1385 | */ | |
1386 | static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, | |
1387 | unsigned int cur_mss, int nonagle) | |
1388 | { | |
1389 | /* Nagle rule does not apply to frames, which sit in the middle of the | |
1390 | * write_queue (they have no chances to get new data). | |
1391 | * | |
1392 | * This is implemented in the callers, where they modify the 'nonagle' | |
1393 | * argument based upon the location of SKB in the send queue. | |
1394 | */ | |
1395 | if (nonagle & TCP_NAGLE_PUSH) | |
1396 | return 1; | |
1397 | ||
d551e454 IJ |
1398 | /* Don't use the nagle rule for urgent data (or for the final FIN). |
1399 | * Nagle can be ignored during F-RTO too (see RFC4138). | |
1400 | */ | |
33f5f57e | 1401 | if (tcp_urg_mode(tp) || (tp->frto_counter == 2) || |
a3433f35 | 1402 | (TCP_SKB_CB(skb)->flags & TCPHDR_FIN)) |
c1b4a7e6 DM |
1403 | return 1; |
1404 | ||
1405 | if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) | |
1406 | return 1; | |
1407 | ||
1408 | return 0; | |
1409 | } | |
1410 | ||
1411 | /* Does at least the first segment of SKB fit into the send window? */ | |
056834d9 IJ |
1412 | static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, |
1413 | unsigned int cur_mss) | |
c1b4a7e6 DM |
1414 | { |
1415 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
1416 | ||
1417 | if (skb->len > cur_mss) | |
1418 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; | |
1419 | ||
90840def | 1420 | return !after(end_seq, tcp_wnd_end(tp)); |
c1b4a7e6 DM |
1421 | } |
1422 | ||
fe067e8a | 1423 | /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) |
c1b4a7e6 DM |
1424 | * should be put on the wire right now. If so, it returns the number of |
1425 | * packets allowed by the congestion window. | |
1426 | */ | |
1427 | static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, | |
1428 | unsigned int cur_mss, int nonagle) | |
1429 | { | |
1430 | struct tcp_sock *tp = tcp_sk(sk); | |
1431 | unsigned int cwnd_quota; | |
1432 | ||
846998ae | 1433 | tcp_init_tso_segs(sk, skb, cur_mss); |
c1b4a7e6 DM |
1434 | |
1435 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) | |
1436 | return 0; | |
1437 | ||
1438 | cwnd_quota = tcp_cwnd_test(tp, skb); | |
056834d9 | 1439 | if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) |
c1b4a7e6 DM |
1440 | cwnd_quota = 0; |
1441 | ||
1442 | return cwnd_quota; | |
1443 | } | |
1444 | ||
67edfef7 | 1445 | /* Test if sending is allowed right now. */ |
9e412ba7 | 1446 | int tcp_may_send_now(struct sock *sk) |
c1b4a7e6 | 1447 | { |
9e412ba7 | 1448 | struct tcp_sock *tp = tcp_sk(sk); |
fe067e8a | 1449 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 | 1450 | |
a02cec21 | 1451 | return skb && |
0c54b85f | 1452 | tcp_snd_test(sk, skb, tcp_current_mss(sk), |
c1b4a7e6 | 1453 | (tcp_skb_is_last(sk, skb) ? |
a02cec21 | 1454 | tp->nonagle : TCP_NAGLE_PUSH)); |
c1b4a7e6 DM |
1455 | } |
1456 | ||
1457 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet | |
1458 | * which is put after SKB on the list. It is very much like | |
1459 | * tcp_fragment() except that it may make several kinds of assumptions | |
1460 | * in order to speed up the splitting operation. In particular, we | |
1461 | * know that all the data is in scatter-gather pages, and that the | |
1462 | * packet has never been sent out before (and thus is not cloned). | |
1463 | */ | |
056834d9 | 1464 | static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, |
c4ead4c5 | 1465 | unsigned int mss_now, gfp_t gfp) |
c1b4a7e6 DM |
1466 | { |
1467 | struct sk_buff *buff; | |
1468 | int nlen = skb->len - len; | |
9ce01461 | 1469 | u8 flags; |
c1b4a7e6 DM |
1470 | |
1471 | /* All of a TSO frame must be composed of paged data. */ | |
c8ac3774 HX |
1472 | if (skb->len != skb->data_len) |
1473 | return tcp_fragment(sk, skb, len, mss_now); | |
c1b4a7e6 | 1474 | |
c4ead4c5 | 1475 | buff = sk_stream_alloc_skb(sk, 0, gfp); |
c1b4a7e6 DM |
1476 | if (unlikely(buff == NULL)) |
1477 | return -ENOMEM; | |
1478 | ||
3ab224be HA |
1479 | sk->sk_wmem_queued += buff->truesize; |
1480 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea | 1481 | buff->truesize += nlen; |
c1b4a7e6 DM |
1482 | skb->truesize -= nlen; |
1483 | ||
1484 | /* Correct the sequence numbers. */ | |
1485 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1486 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1487 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1488 | ||
1489 | /* PSH and FIN should only be set in the second packet. */ | |
1490 | flags = TCP_SKB_CB(skb)->flags; | |
a3433f35 | 1491 | TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); |
c1b4a7e6 DM |
1492 | TCP_SKB_CB(buff)->flags = flags; |
1493 | ||
1494 | /* This packet was never sent out yet, so no SACK bits. */ | |
1495 | TCP_SKB_CB(buff)->sacked = 0; | |
1496 | ||
84fa7933 | 1497 | buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; |
c1b4a7e6 DM |
1498 | skb_split(skb, buff, len); |
1499 | ||
1500 | /* Fix up tso_factor for both original and new SKB. */ | |
846998ae DM |
1501 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
1502 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
c1b4a7e6 DM |
1503 | |
1504 | /* Link BUFF into the send queue. */ | |
1505 | skb_header_release(buff); | |
fe067e8a | 1506 | tcp_insert_write_queue_after(skb, buff, sk); |
c1b4a7e6 DM |
1507 | |
1508 | return 0; | |
1509 | } | |
1510 | ||
1511 | /* Try to defer sending, if possible, in order to minimize the amount | |
1512 | * of TSO splitting we do. View it as a kind of TSO Nagle test. | |
1513 | * | |
1514 | * This algorithm is from John Heffner. | |
1515 | */ | |
9e412ba7 | 1516 | static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) |
c1b4a7e6 | 1517 | { |
9e412ba7 | 1518 | struct tcp_sock *tp = tcp_sk(sk); |
6687e988 | 1519 | const struct inet_connection_sock *icsk = inet_csk(sk); |
c1b4a7e6 | 1520 | u32 send_win, cong_win, limit, in_flight; |
ad9f4f50 | 1521 | int win_divisor; |
c1b4a7e6 | 1522 | |
a3433f35 | 1523 | if (TCP_SKB_CB(skb)->flags & TCPHDR_FIN) |
ae8064ac | 1524 | goto send_now; |
c1b4a7e6 | 1525 | |
6687e988 | 1526 | if (icsk->icsk_ca_state != TCP_CA_Open) |
ae8064ac JH |
1527 | goto send_now; |
1528 | ||
1529 | /* Defer for less than two clock ticks. */ | |
bd515c3e | 1530 | if (tp->tso_deferred && |
a2acde07 | 1531 | (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) |
ae8064ac | 1532 | goto send_now; |
908a75c1 | 1533 | |
c1b4a7e6 DM |
1534 | in_flight = tcp_packets_in_flight(tp); |
1535 | ||
056834d9 | 1536 | BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); |
c1b4a7e6 | 1537 | |
90840def | 1538 | send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
c1b4a7e6 DM |
1539 | |
1540 | /* From in_flight test above, we know that cwnd > in_flight. */ | |
1541 | cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; | |
1542 | ||
1543 | limit = min(send_win, cong_win); | |
1544 | ||
ba244fe9 | 1545 | /* If a full-sized TSO skb can be sent, do it. */ |
82cc1a7a | 1546 | if (limit >= sk->sk_gso_max_size) |
ae8064ac | 1547 | goto send_now; |
ba244fe9 | 1548 | |
62ad2761 IJ |
1549 | /* Middle in queue won't get any more data, full sendable already? */ |
1550 | if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) | |
1551 | goto send_now; | |
1552 | ||
ad9f4f50 ED |
1553 | win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); |
1554 | if (win_divisor) { | |
c1b4a7e6 DM |
1555 | u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); |
1556 | ||
1557 | /* If at least some fraction of a window is available, | |
1558 | * just use it. | |
1559 | */ | |
ad9f4f50 | 1560 | chunk /= win_divisor; |
c1b4a7e6 | 1561 | if (limit >= chunk) |
ae8064ac | 1562 | goto send_now; |
c1b4a7e6 DM |
1563 | } else { |
1564 | /* Different approach, try not to defer past a single | |
1565 | * ACK. Receiver should ACK every other full sized | |
1566 | * frame, so if we have space for more than 3 frames | |
1567 | * then send now. | |
1568 | */ | |
1569 | if (limit > tcp_max_burst(tp) * tp->mss_cache) | |
ae8064ac | 1570 | goto send_now; |
c1b4a7e6 DM |
1571 | } |
1572 | ||
1573 | /* Ok, it looks like it is advisable to defer. */ | |
056834d9 | 1574 | tp->tso_deferred = 1 | (jiffies << 1); |
ae8064ac | 1575 | |
c1b4a7e6 | 1576 | return 1; |
ae8064ac JH |
1577 | |
1578 | send_now: | |
1579 | tp->tso_deferred = 0; | |
1580 | return 0; | |
c1b4a7e6 DM |
1581 | } |
1582 | ||
5d424d5a | 1583 | /* Create a new MTU probe if we are ready. |
67edfef7 AK |
1584 | * MTU probe is regularly attempting to increase the path MTU by |
1585 | * deliberately sending larger packets. This discovers routing | |
1586 | * changes resulting in larger path MTUs. | |
1587 | * | |
5d424d5a JH |
1588 | * Returns 0 if we should wait to probe (no cwnd available), |
1589 | * 1 if a probe was sent, | |
056834d9 IJ |
1590 | * -1 otherwise |
1591 | */ | |
5d424d5a JH |
1592 | static int tcp_mtu_probe(struct sock *sk) |
1593 | { | |
1594 | struct tcp_sock *tp = tcp_sk(sk); | |
1595 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1596 | struct sk_buff *skb, *nskb, *next; | |
1597 | int len; | |
1598 | int probe_size; | |
91cc17c0 | 1599 | int size_needed; |
5d424d5a JH |
1600 | int copy; |
1601 | int mss_now; | |
1602 | ||
1603 | /* Not currently probing/verifying, | |
1604 | * not in recovery, | |
1605 | * have enough cwnd, and | |
1606 | * not SACKing (the variable headers throw things off) */ | |
1607 | if (!icsk->icsk_mtup.enabled || | |
1608 | icsk->icsk_mtup.probe_size || | |
1609 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || | |
1610 | tp->snd_cwnd < 11 || | |
cabeccbd | 1611 | tp->rx_opt.num_sacks || tp->rx_opt.dsack) |
5d424d5a JH |
1612 | return -1; |
1613 | ||
1614 | /* Very simple search strategy: just double the MSS. */ | |
0c54b85f | 1615 | mss_now = tcp_current_mss(sk); |
056834d9 | 1616 | probe_size = 2 * tp->mss_cache; |
91cc17c0 | 1617 | size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; |
5d424d5a JH |
1618 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { |
1619 | /* TODO: set timer for probe_converge_event */ | |
1620 | return -1; | |
1621 | } | |
1622 | ||
1623 | /* Have enough data in the send queue to probe? */ | |
7f9c33e5 | 1624 | if (tp->write_seq - tp->snd_nxt < size_needed) |
5d424d5a JH |
1625 | return -1; |
1626 | ||
91cc17c0 IJ |
1627 | if (tp->snd_wnd < size_needed) |
1628 | return -1; | |
90840def | 1629 | if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) |
91cc17c0 | 1630 | return 0; |
5d424d5a | 1631 | |
d67c58e9 IJ |
1632 | /* Do we need to wait to drain cwnd? With none in flight, don't stall */ |
1633 | if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { | |
1634 | if (!tcp_packets_in_flight(tp)) | |
5d424d5a JH |
1635 | return -1; |
1636 | else | |
1637 | return 0; | |
1638 | } | |
1639 | ||
1640 | /* We're allowed to probe. Build it now. */ | |
1641 | if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) | |
1642 | return -1; | |
3ab224be HA |
1643 | sk->sk_wmem_queued += nskb->truesize; |
1644 | sk_mem_charge(sk, nskb->truesize); | |
5d424d5a | 1645 | |
fe067e8a | 1646 | skb = tcp_send_head(sk); |
5d424d5a JH |
1647 | |
1648 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; | |
1649 | TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; | |
a3433f35 | 1650 | TCP_SKB_CB(nskb)->flags = TCPHDR_ACK; |
5d424d5a JH |
1651 | TCP_SKB_CB(nskb)->sacked = 0; |
1652 | nskb->csum = 0; | |
84fa7933 | 1653 | nskb->ip_summed = skb->ip_summed; |
5d424d5a | 1654 | |
50c4817e IJ |
1655 | tcp_insert_write_queue_before(nskb, skb, sk); |
1656 | ||
5d424d5a | 1657 | len = 0; |
234b6860 | 1658 | tcp_for_write_queue_from_safe(skb, next, sk) { |
5d424d5a JH |
1659 | copy = min_t(int, skb->len, probe_size - len); |
1660 | if (nskb->ip_summed) | |
1661 | skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); | |
1662 | else | |
1663 | nskb->csum = skb_copy_and_csum_bits(skb, 0, | |
056834d9 IJ |
1664 | skb_put(nskb, copy), |
1665 | copy, nskb->csum); | |
5d424d5a JH |
1666 | |
1667 | if (skb->len <= copy) { | |
1668 | /* We've eaten all the data from this skb. | |
1669 | * Throw it away. */ | |
1670 | TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags; | |
fe067e8a | 1671 | tcp_unlink_write_queue(skb, sk); |
3ab224be | 1672 | sk_wmem_free_skb(sk, skb); |
5d424d5a JH |
1673 | } else { |
1674 | TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags & | |
a3433f35 | 1675 | ~(TCPHDR_FIN|TCPHDR_PSH); |
5d424d5a JH |
1676 | if (!skb_shinfo(skb)->nr_frags) { |
1677 | skb_pull(skb, copy); | |
84fa7933 | 1678 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
056834d9 IJ |
1679 | skb->csum = csum_partial(skb->data, |
1680 | skb->len, 0); | |
5d424d5a JH |
1681 | } else { |
1682 | __pskb_trim_head(skb, copy); | |
1683 | tcp_set_skb_tso_segs(sk, skb, mss_now); | |
1684 | } | |
1685 | TCP_SKB_CB(skb)->seq += copy; | |
1686 | } | |
1687 | ||
1688 | len += copy; | |
234b6860 IJ |
1689 | |
1690 | if (len >= probe_size) | |
1691 | break; | |
5d424d5a JH |
1692 | } |
1693 | tcp_init_tso_segs(sk, nskb, nskb->len); | |
1694 | ||
1695 | /* We're ready to send. If this fails, the probe will | |
1696 | * be resegmented into mss-sized pieces by tcp_write_xmit(). */ | |
1697 | TCP_SKB_CB(nskb)->when = tcp_time_stamp; | |
1698 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { | |
1699 | /* Decrement cwnd here because we are sending | |
056834d9 | 1700 | * effectively two packets. */ |
5d424d5a | 1701 | tp->snd_cwnd--; |
66f5fe62 | 1702 | tcp_event_new_data_sent(sk, nskb); |
5d424d5a JH |
1703 | |
1704 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); | |
0e7b1368 JH |
1705 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; |
1706 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; | |
5d424d5a JH |
1707 | |
1708 | return 1; | |
1709 | } | |
1710 | ||
1711 | return -1; | |
1712 | } | |
1713 | ||
1da177e4 LT |
1714 | /* This routine writes packets to the network. It advances the |
1715 | * send_head. This happens as incoming acks open up the remote | |
1716 | * window for us. | |
1717 | * | |
f8269a49 IJ |
1718 | * LARGESEND note: !tcp_urg_mode is overkill, only frames between |
1719 | * snd_up-64k-mss .. snd_up cannot be large. However, taking into | |
1720 | * account rare use of URG, this is not a big flaw. | |
1721 | * | |
1da177e4 LT |
1722 | * Returns 1, if no segments are in flight and we have queued segments, but |
1723 | * cannot send anything now because of SWS or another problem. | |
1724 | */ | |
d5dd9175 IJ |
1725 | static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
1726 | int push_one, gfp_t gfp) | |
1da177e4 LT |
1727 | { |
1728 | struct tcp_sock *tp = tcp_sk(sk); | |
92df7b51 | 1729 | struct sk_buff *skb; |
c1b4a7e6 DM |
1730 | unsigned int tso_segs, sent_pkts; |
1731 | int cwnd_quota; | |
5d424d5a | 1732 | int result; |
1da177e4 | 1733 | |
92df7b51 | 1734 | sent_pkts = 0; |
5d424d5a | 1735 | |
d5dd9175 IJ |
1736 | if (!push_one) { |
1737 | /* Do MTU probing. */ | |
1738 | result = tcp_mtu_probe(sk); | |
1739 | if (!result) { | |
1740 | return 0; | |
1741 | } else if (result > 0) { | |
1742 | sent_pkts = 1; | |
1743 | } | |
5d424d5a JH |
1744 | } |
1745 | ||
fe067e8a | 1746 | while ((skb = tcp_send_head(sk))) { |
c8ac3774 HX |
1747 | unsigned int limit; |
1748 | ||
b68e9f85 | 1749 | tso_segs = tcp_init_tso_segs(sk, skb, mss_now); |
c1b4a7e6 | 1750 | BUG_ON(!tso_segs); |
aa93466b | 1751 | |
b68e9f85 HX |
1752 | cwnd_quota = tcp_cwnd_test(tp, skb); |
1753 | if (!cwnd_quota) | |
1754 | break; | |
1755 | ||
1756 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) | |
1757 | break; | |
1758 | ||
c1b4a7e6 DM |
1759 | if (tso_segs == 1) { |
1760 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, | |
1761 | (tcp_skb_is_last(sk, skb) ? | |
1762 | nonagle : TCP_NAGLE_PUSH)))) | |
1763 | break; | |
1764 | } else { | |
d5dd9175 | 1765 | if (!push_one && tcp_tso_should_defer(sk, skb)) |
c1b4a7e6 DM |
1766 | break; |
1767 | } | |
aa93466b | 1768 | |
c8ac3774 | 1769 | limit = mss_now; |
f8269a49 | 1770 | if (tso_segs > 1 && !tcp_urg_mode(tp)) |
0e3a4803 IJ |
1771 | limit = tcp_mss_split_point(sk, skb, mss_now, |
1772 | cwnd_quota); | |
1da177e4 | 1773 | |
c8ac3774 | 1774 | if (skb->len > limit && |
c4ead4c5 | 1775 | unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) |
c8ac3774 HX |
1776 | break; |
1777 | ||
92df7b51 | 1778 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
c1b4a7e6 | 1779 | |
d5dd9175 | 1780 | if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) |
92df7b51 | 1781 | break; |
1da177e4 | 1782 | |
92df7b51 DM |
1783 | /* Advance the send_head. This one is sent out. |
1784 | * This call will increment packets_out. | |
1785 | */ | |
66f5fe62 | 1786 | tcp_event_new_data_sent(sk, skb); |
1da177e4 | 1787 | |
92df7b51 | 1788 | tcp_minshall_update(tp, mss_now, skb); |
aa93466b | 1789 | sent_pkts++; |
d5dd9175 IJ |
1790 | |
1791 | if (push_one) | |
1792 | break; | |
92df7b51 | 1793 | } |
1da177e4 | 1794 | |
aa93466b | 1795 | if (likely(sent_pkts)) { |
9e412ba7 | 1796 | tcp_cwnd_validate(sk); |
92df7b51 | 1797 | return 0; |
1da177e4 | 1798 | } |
fe067e8a | 1799 | return !tp->packets_out && tcp_send_head(sk); |
1da177e4 LT |
1800 | } |
1801 | ||
a762a980 DM |
1802 | /* Push out any pending frames which were held back due to |
1803 | * TCP_CORK or attempt at coalescing tiny packets. | |
1804 | * The socket must be locked by the caller. | |
1805 | */ | |
9e412ba7 IJ |
1806 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, |
1807 | int nonagle) | |
a762a980 | 1808 | { |
726e07a8 IJ |
1809 | /* If we are closed, the bytes will have to remain here. |
1810 | * In time closedown will finish, we empty the write queue and | |
1811 | * all will be happy. | |
1812 | */ | |
1813 | if (unlikely(sk->sk_state == TCP_CLOSE)) | |
1814 | return; | |
1815 | ||
d5dd9175 | 1816 | if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC)) |
726e07a8 | 1817 | tcp_check_probe_timer(sk); |
a762a980 DM |
1818 | } |
1819 | ||
c1b4a7e6 DM |
1820 | /* Send _single_ skb sitting at the send head. This function requires |
1821 | * true push pending frames to setup probe timer etc. | |
1822 | */ | |
1823 | void tcp_push_one(struct sock *sk, unsigned int mss_now) | |
1824 | { | |
fe067e8a | 1825 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 DM |
1826 | |
1827 | BUG_ON(!skb || skb->len < mss_now); | |
1828 | ||
d5dd9175 | 1829 | tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); |
c1b4a7e6 DM |
1830 | } |
1831 | ||
1da177e4 LT |
1832 | /* This function returns the amount that we can raise the |
1833 | * usable window based on the following constraints | |
e905a9ed | 1834 | * |
1da177e4 LT |
1835 | * 1. The window can never be shrunk once it is offered (RFC 793) |
1836 | * 2. We limit memory per socket | |
1837 | * | |
1838 | * RFC 1122: | |
1839 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep | |
1840 | * RECV.NEXT + RCV.WIN fixed until: | |
1841 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" | |
1842 | * | |
1843 | * i.e. don't raise the right edge of the window until you can raise | |
1844 | * it at least MSS bytes. | |
1845 | * | |
1846 | * Unfortunately, the recommended algorithm breaks header prediction, | |
1847 | * since header prediction assumes th->window stays fixed. | |
1848 | * | |
1849 | * Strictly speaking, keeping th->window fixed violates the receiver | |
1850 | * side SWS prevention criteria. The problem is that under this rule | |
1851 | * a stream of single byte packets will cause the right side of the | |
1852 | * window to always advance by a single byte. | |
e905a9ed | 1853 | * |
1da177e4 LT |
1854 | * Of course, if the sender implements sender side SWS prevention |
1855 | * then this will not be a problem. | |
e905a9ed | 1856 | * |
1da177e4 | 1857 | * BSD seems to make the following compromise: |
e905a9ed | 1858 | * |
1da177e4 LT |
1859 | * If the free space is less than the 1/4 of the maximum |
1860 | * space available and the free space is less than 1/2 mss, | |
1861 | * then set the window to 0. | |
1862 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] | |
1863 | * Otherwise, just prevent the window from shrinking | |
1864 | * and from being larger than the largest representable value. | |
1865 | * | |
1866 | * This prevents incremental opening of the window in the regime | |
1867 | * where TCP is limited by the speed of the reader side taking | |
1868 | * data out of the TCP receive queue. It does nothing about | |
1869 | * those cases where the window is constrained on the sender side | |
1870 | * because the pipeline is full. | |
1871 | * | |
1872 | * BSD also seems to "accidentally" limit itself to windows that are a | |
1873 | * multiple of MSS, at least until the free space gets quite small. | |
1874 | * This would appear to be a side effect of the mbuf implementation. | |
1875 | * Combining these two algorithms results in the observed behavior | |
1876 | * of having a fixed window size at almost all times. | |
1877 | * | |
1878 | * Below we obtain similar behavior by forcing the offered window to | |
1879 | * a multiple of the mss when it is feasible to do so. | |
1880 | * | |
1881 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. | |
1882 | * Regular options like TIMESTAMP are taken into account. | |
1883 | */ | |
1884 | u32 __tcp_select_window(struct sock *sk) | |
1885 | { | |
463c84b9 | 1886 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 1887 | struct tcp_sock *tp = tcp_sk(sk); |
caa20d9a | 1888 | /* MSS for the peer's data. Previous versions used mss_clamp |
1da177e4 LT |
1889 | * here. I don't know if the value based on our guesses |
1890 | * of peer's MSS is better for the performance. It's more correct | |
1891 | * but may be worse for the performance because of rcv_mss | |
1892 | * fluctuations. --SAW 1998/11/1 | |
1893 | */ | |
463c84b9 | 1894 | int mss = icsk->icsk_ack.rcv_mss; |
1da177e4 LT |
1895 | int free_space = tcp_space(sk); |
1896 | int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); | |
1897 | int window; | |
1898 | ||
1899 | if (mss > full_space) | |
e905a9ed | 1900 | mss = full_space; |
1da177e4 | 1901 | |
b92edbe0 | 1902 | if (free_space < (full_space >> 1)) { |
463c84b9 | 1903 | icsk->icsk_ack.quick = 0; |
1da177e4 LT |
1904 | |
1905 | if (tcp_memory_pressure) | |
056834d9 IJ |
1906 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, |
1907 | 4U * tp->advmss); | |
1da177e4 LT |
1908 | |
1909 | if (free_space < mss) | |
1910 | return 0; | |
1911 | } | |
1912 | ||
1913 | if (free_space > tp->rcv_ssthresh) | |
1914 | free_space = tp->rcv_ssthresh; | |
1915 | ||
1916 | /* Don't do rounding if we are using window scaling, since the | |
1917 | * scaled window will not line up with the MSS boundary anyway. | |
1918 | */ | |
1919 | window = tp->rcv_wnd; | |
1920 | if (tp->rx_opt.rcv_wscale) { | |
1921 | window = free_space; | |
1922 | ||
1923 | /* Advertise enough space so that it won't get scaled away. | |
1924 | * Import case: prevent zero window announcement if | |
1925 | * 1<<rcv_wscale > mss. | |
1926 | */ | |
1927 | if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) | |
1928 | window = (((window >> tp->rx_opt.rcv_wscale) + 1) | |
1929 | << tp->rx_opt.rcv_wscale); | |
1930 | } else { | |
1931 | /* Get the largest window that is a nice multiple of mss. | |
1932 | * Window clamp already applied above. | |
1933 | * If our current window offering is within 1 mss of the | |
1934 | * free space we just keep it. This prevents the divide | |
1935 | * and multiply from happening most of the time. | |
1936 | * We also don't do any window rounding when the free space | |
1937 | * is too small. | |
1938 | */ | |
1939 | if (window <= free_space - mss || window > free_space) | |
056834d9 | 1940 | window = (free_space / mss) * mss; |
84565070 | 1941 | else if (mss == full_space && |
b92edbe0 | 1942 | free_space > window + (full_space >> 1)) |
84565070 | 1943 | window = free_space; |
1da177e4 LT |
1944 | } |
1945 | ||
1946 | return window; | |
1947 | } | |
1948 | ||
4a17fc3a IJ |
1949 | /* Collapses two adjacent SKB's during retransmission. */ |
1950 | static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) | |
1da177e4 LT |
1951 | { |
1952 | struct tcp_sock *tp = tcp_sk(sk); | |
fe067e8a | 1953 | struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); |
058dc334 | 1954 | int skb_size, next_skb_size; |
1da177e4 | 1955 | |
058dc334 IJ |
1956 | skb_size = skb->len; |
1957 | next_skb_size = next_skb->len; | |
1da177e4 | 1958 | |
058dc334 | 1959 | BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); |
a6963a6b | 1960 | |
058dc334 | 1961 | tcp_highest_sack_combine(sk, next_skb, skb); |
1da177e4 | 1962 | |
058dc334 | 1963 | tcp_unlink_write_queue(next_skb, sk); |
1da177e4 | 1964 | |
058dc334 IJ |
1965 | skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), |
1966 | next_skb_size); | |
1da177e4 | 1967 | |
058dc334 IJ |
1968 | if (next_skb->ip_summed == CHECKSUM_PARTIAL) |
1969 | skb->ip_summed = CHECKSUM_PARTIAL; | |
1da177e4 | 1970 | |
058dc334 IJ |
1971 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
1972 | skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); | |
1da177e4 | 1973 | |
058dc334 IJ |
1974 | /* Update sequence range on original skb. */ |
1975 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; | |
1da177e4 | 1976 | |
e6c7d085 IJ |
1977 | /* Merge over control information. This moves PSH/FIN etc. over */ |
1978 | TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(next_skb)->flags; | |
058dc334 IJ |
1979 | |
1980 | /* All done, get rid of second SKB and account for it so | |
1981 | * packet counting does not break. | |
1982 | */ | |
1983 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; | |
058dc334 IJ |
1984 | |
1985 | /* changed transmit queue under us so clear hints */ | |
ef9da47c IJ |
1986 | tcp_clear_retrans_hints_partial(tp); |
1987 | if (next_skb == tp->retransmit_skb_hint) | |
1988 | tp->retransmit_skb_hint = skb; | |
058dc334 | 1989 | |
797108d1 IJ |
1990 | tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); |
1991 | ||
058dc334 | 1992 | sk_wmem_free_skb(sk, next_skb); |
1da177e4 LT |
1993 | } |
1994 | ||
67edfef7 | 1995 | /* Check if coalescing SKBs is legal. */ |
4a17fc3a IJ |
1996 | static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb) |
1997 | { | |
1998 | if (tcp_skb_pcount(skb) > 1) | |
1999 | return 0; | |
2000 | /* TODO: SACK collapsing could be used to remove this condition */ | |
2001 | if (skb_shinfo(skb)->nr_frags != 0) | |
2002 | return 0; | |
2003 | if (skb_cloned(skb)) | |
2004 | return 0; | |
2005 | if (skb == tcp_send_head(sk)) | |
2006 | return 0; | |
2007 | /* Some heurestics for collapsing over SACK'd could be invented */ | |
2008 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
2009 | return 0; | |
2010 | ||
2011 | return 1; | |
2012 | } | |
2013 | ||
67edfef7 AK |
2014 | /* Collapse packets in the retransmit queue to make to create |
2015 | * less packets on the wire. This is only done on retransmission. | |
2016 | */ | |
4a17fc3a IJ |
2017 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, |
2018 | int space) | |
2019 | { | |
2020 | struct tcp_sock *tp = tcp_sk(sk); | |
2021 | struct sk_buff *skb = to, *tmp; | |
2022 | int first = 1; | |
2023 | ||
2024 | if (!sysctl_tcp_retrans_collapse) | |
2025 | return; | |
a3433f35 | 2026 | if (TCP_SKB_CB(skb)->flags & TCPHDR_SYN) |
4a17fc3a IJ |
2027 | return; |
2028 | ||
2029 | tcp_for_write_queue_from_safe(skb, tmp, sk) { | |
2030 | if (!tcp_can_collapse(sk, skb)) | |
2031 | break; | |
2032 | ||
2033 | space -= skb->len; | |
2034 | ||
2035 | if (first) { | |
2036 | first = 0; | |
2037 | continue; | |
2038 | } | |
2039 | ||
2040 | if (space < 0) | |
2041 | break; | |
2042 | /* Punt if not enough space exists in the first SKB for | |
2043 | * the data in the second | |
2044 | */ | |
2045 | if (skb->len > skb_tailroom(to)) | |
2046 | break; | |
2047 | ||
2048 | if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) | |
2049 | break; | |
2050 | ||
2051 | tcp_collapse_retrans(sk, to); | |
2052 | } | |
2053 | } | |
2054 | ||
1da177e4 LT |
2055 | /* This retransmits one SKB. Policy decisions and retransmit queue |
2056 | * state updates are done by the caller. Returns non-zero if an | |
2057 | * error occurred which prevented the send. | |
2058 | */ | |
2059 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) | |
2060 | { | |
2061 | struct tcp_sock *tp = tcp_sk(sk); | |
5d424d5a | 2062 | struct inet_connection_sock *icsk = inet_csk(sk); |
7d227cd2 | 2063 | unsigned int cur_mss; |
1da177e4 LT |
2064 | int err; |
2065 | ||
5d424d5a JH |
2066 | /* Inconslusive MTU probe */ |
2067 | if (icsk->icsk_mtup.probe_size) { | |
2068 | icsk->icsk_mtup.probe_size = 0; | |
2069 | } | |
2070 | ||
1da177e4 | 2071 | /* Do not sent more than we queued. 1/4 is reserved for possible |
caa20d9a | 2072 | * copying overhead: fragmentation, tunneling, mangling etc. |
1da177e4 LT |
2073 | */ |
2074 | if (atomic_read(&sk->sk_wmem_alloc) > | |
2075 | min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) | |
2076 | return -EAGAIN; | |
2077 | ||
2078 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { | |
2079 | if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
2080 | BUG(); | |
1da177e4 LT |
2081 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
2082 | return -ENOMEM; | |
2083 | } | |
2084 | ||
7d227cd2 SS |
2085 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) |
2086 | return -EHOSTUNREACH; /* Routing failure or similar. */ | |
2087 | ||
0c54b85f | 2088 | cur_mss = tcp_current_mss(sk); |
7d227cd2 | 2089 | |
1da177e4 LT |
2090 | /* If receiver has shrunk his window, and skb is out of |
2091 | * new window, do not retransmit it. The exception is the | |
2092 | * case, when window is shrunk to zero. In this case | |
2093 | * our retransmit serves as a zero window probe. | |
2094 | */ | |
9d4fb27d JP |
2095 | if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && |
2096 | TCP_SKB_CB(skb)->seq != tp->snd_una) | |
1da177e4 LT |
2097 | return -EAGAIN; |
2098 | ||
2099 | if (skb->len > cur_mss) { | |
846998ae | 2100 | if (tcp_fragment(sk, skb, cur_mss, cur_mss)) |
1da177e4 | 2101 | return -ENOMEM; /* We'll try again later. */ |
02276f3c | 2102 | } else { |
9eb9362e IJ |
2103 | int oldpcount = tcp_skb_pcount(skb); |
2104 | ||
2105 | if (unlikely(oldpcount > 1)) { | |
2106 | tcp_init_tso_segs(sk, skb, cur_mss); | |
2107 | tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); | |
2108 | } | |
1da177e4 LT |
2109 | } |
2110 | ||
4a17fc3a | 2111 | tcp_retrans_try_collapse(sk, skb, cur_mss); |
1da177e4 | 2112 | |
1da177e4 LT |
2113 | /* Some Solaris stacks overoptimize and ignore the FIN on a |
2114 | * retransmit when old data is attached. So strip it off | |
2115 | * since it is cheap to do so and saves bytes on the network. | |
2116 | */ | |
2de979bd | 2117 | if (skb->len > 0 && |
a3433f35 | 2118 | (TCP_SKB_CB(skb)->flags & TCPHDR_FIN) && |
2de979bd | 2119 | tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { |
1da177e4 | 2120 | if (!pskb_trim(skb, 0)) { |
e870a8ef IJ |
2121 | /* Reuse, even though it does some unnecessary work */ |
2122 | tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, | |
2123 | TCP_SKB_CB(skb)->flags); | |
1da177e4 | 2124 | skb->ip_summed = CHECKSUM_NONE; |
1da177e4 LT |
2125 | } |
2126 | } | |
2127 | ||
2128 | /* Make a copy, if the first transmission SKB clone we made | |
2129 | * is still in somebody's hands, else make a clone. | |
2130 | */ | |
2131 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
1da177e4 | 2132 | |
dfb4b9dc | 2133 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
1da177e4 LT |
2134 | |
2135 | if (err == 0) { | |
2136 | /* Update global TCP statistics. */ | |
81cc8a75 | 2137 | TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); |
1da177e4 LT |
2138 | |
2139 | tp->total_retrans++; | |
2140 | ||
2141 | #if FASTRETRANS_DEBUG > 0 | |
056834d9 | 2142 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
1da177e4 LT |
2143 | if (net_ratelimit()) |
2144 | printk(KERN_DEBUG "retrans_out leaked.\n"); | |
2145 | } | |
2146 | #endif | |
b08d6cb2 IJ |
2147 | if (!tp->retrans_out) |
2148 | tp->lost_retrans_low = tp->snd_nxt; | |
1da177e4 LT |
2149 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; |
2150 | tp->retrans_out += tcp_skb_pcount(skb); | |
2151 | ||
2152 | /* Save stamp of the first retransmit. */ | |
2153 | if (!tp->retrans_stamp) | |
2154 | tp->retrans_stamp = TCP_SKB_CB(skb)->when; | |
2155 | ||
2156 | tp->undo_retrans++; | |
2157 | ||
2158 | /* snd_nxt is stored to detect loss of retransmitted segment, | |
2159 | * see tcp_input.c tcp_sacktag_write_queue(). | |
2160 | */ | |
2161 | TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; | |
2162 | } | |
2163 | return err; | |
2164 | } | |
2165 | ||
67edfef7 AK |
2166 | /* Check if we forward retransmits are possible in the current |
2167 | * window/congestion state. | |
2168 | */ | |
b5afe7bc IJ |
2169 | static int tcp_can_forward_retransmit(struct sock *sk) |
2170 | { | |
2171 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
2172 | struct tcp_sock *tp = tcp_sk(sk); | |
2173 | ||
2174 | /* Forward retransmissions are possible only during Recovery. */ | |
2175 | if (icsk->icsk_ca_state != TCP_CA_Recovery) | |
2176 | return 0; | |
2177 | ||
2178 | /* No forward retransmissions in Reno are possible. */ | |
2179 | if (tcp_is_reno(tp)) | |
2180 | return 0; | |
2181 | ||
2182 | /* Yeah, we have to make difficult choice between forward transmission | |
2183 | * and retransmission... Both ways have their merits... | |
2184 | * | |
2185 | * For now we do not retransmit anything, while we have some new | |
2186 | * segments to send. In the other cases, follow rule 3 for | |
2187 | * NextSeg() specified in RFC3517. | |
2188 | */ | |
2189 | ||
2190 | if (tcp_may_send_now(sk)) | |
2191 | return 0; | |
2192 | ||
2193 | return 1; | |
2194 | } | |
2195 | ||
1da177e4 LT |
2196 | /* This gets called after a retransmit timeout, and the initially |
2197 | * retransmitted data is acknowledged. It tries to continue | |
2198 | * resending the rest of the retransmit queue, until either | |
2199 | * we've sent it all or the congestion window limit is reached. | |
2200 | * If doing SACK, the first ACK which comes back for a timeout | |
2201 | * based retransmit packet might feed us FACK information again. | |
2202 | * If so, we use it to avoid unnecessarily retransmissions. | |
2203 | */ | |
2204 | void tcp_xmit_retransmit_queue(struct sock *sk) | |
2205 | { | |
6687e988 | 2206 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2207 | struct tcp_sock *tp = tcp_sk(sk); |
2208 | struct sk_buff *skb; | |
0e1c54c2 | 2209 | struct sk_buff *hole = NULL; |
618d9f25 | 2210 | u32 last_lost; |
61eb55f4 | 2211 | int mib_idx; |
0e1c54c2 | 2212 | int fwd_rexmitting = 0; |
6a438bbe | 2213 | |
45e77d31 IJ |
2214 | if (!tp->packets_out) |
2215 | return; | |
2216 | ||
08ebd172 IJ |
2217 | if (!tp->lost_out) |
2218 | tp->retransmit_high = tp->snd_una; | |
2219 | ||
618d9f25 | 2220 | if (tp->retransmit_skb_hint) { |
6a438bbe | 2221 | skb = tp->retransmit_skb_hint; |
618d9f25 IJ |
2222 | last_lost = TCP_SKB_CB(skb)->end_seq; |
2223 | if (after(last_lost, tp->retransmit_high)) | |
2224 | last_lost = tp->retransmit_high; | |
2225 | } else { | |
fe067e8a | 2226 | skb = tcp_write_queue_head(sk); |
618d9f25 IJ |
2227 | last_lost = tp->snd_una; |
2228 | } | |
1da177e4 | 2229 | |
08ebd172 IJ |
2230 | tcp_for_write_queue_from(skb, sk) { |
2231 | __u8 sacked = TCP_SKB_CB(skb)->sacked; | |
1da177e4 | 2232 | |
08ebd172 IJ |
2233 | if (skb == tcp_send_head(sk)) |
2234 | break; | |
2235 | /* we could do better than to assign each time */ | |
0e1c54c2 IJ |
2236 | if (hole == NULL) |
2237 | tp->retransmit_skb_hint = skb; | |
08ebd172 IJ |
2238 | |
2239 | /* Assume this retransmit will generate | |
2240 | * only one packet for congestion window | |
2241 | * calculation purposes. This works because | |
2242 | * tcp_retransmit_skb() will chop up the | |
2243 | * packet to be MSS sized and all the | |
2244 | * packet counting works out. | |
2245 | */ | |
2246 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) | |
2247 | return; | |
1da177e4 | 2248 | |
0e1c54c2 IJ |
2249 | if (fwd_rexmitting) { |
2250 | begin_fwd: | |
2251 | if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) | |
2252 | break; | |
2253 | mib_idx = LINUX_MIB_TCPFORWARDRETRANS; | |
6a438bbe | 2254 | |
0e1c54c2 | 2255 | } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { |
618d9f25 | 2256 | tp->retransmit_high = last_lost; |
0e1c54c2 IJ |
2257 | if (!tcp_can_forward_retransmit(sk)) |
2258 | break; | |
2259 | /* Backtrack if necessary to non-L'ed skb */ | |
2260 | if (hole != NULL) { | |
2261 | skb = hole; | |
2262 | hole = NULL; | |
2263 | } | |
2264 | fwd_rexmitting = 1; | |
2265 | goto begin_fwd; | |
1da177e4 | 2266 | |
0e1c54c2 | 2267 | } else if (!(sacked & TCPCB_LOST)) { |
ac11ba75 | 2268 | if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) |
0e1c54c2 IJ |
2269 | hole = skb; |
2270 | continue; | |
1da177e4 | 2271 | |
0e1c54c2 | 2272 | } else { |
618d9f25 | 2273 | last_lost = TCP_SKB_CB(skb)->end_seq; |
0e1c54c2 IJ |
2274 | if (icsk->icsk_ca_state != TCP_CA_Loss) |
2275 | mib_idx = LINUX_MIB_TCPFASTRETRANS; | |
2276 | else | |
2277 | mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; | |
2278 | } | |
1da177e4 | 2279 | |
0e1c54c2 | 2280 | if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) |
1da177e4 LT |
2281 | continue; |
2282 | ||
f0ceb0ed | 2283 | if (tcp_retransmit_skb(sk, skb)) |
0e1c54c2 | 2284 | return; |
0e1c54c2 | 2285 | NET_INC_STATS_BH(sock_net(sk), mib_idx); |
1da177e4 | 2286 | |
fe067e8a | 2287 | if (skb == tcp_write_queue_head(sk)) |
3f421baa ACM |
2288 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
2289 | inet_csk(sk)->icsk_rto, | |
2290 | TCP_RTO_MAX); | |
1da177e4 LT |
2291 | } |
2292 | } | |
2293 | ||
1da177e4 LT |
2294 | /* Send a fin. The caller locks the socket for us. This cannot be |
2295 | * allowed to fail queueing a FIN frame under any circumstances. | |
2296 | */ | |
2297 | void tcp_send_fin(struct sock *sk) | |
2298 | { | |
e905a9ed | 2299 | struct tcp_sock *tp = tcp_sk(sk); |
fe067e8a | 2300 | struct sk_buff *skb = tcp_write_queue_tail(sk); |
1da177e4 | 2301 | int mss_now; |
e905a9ed | 2302 | |
1da177e4 LT |
2303 | /* Optimization, tack on the FIN if we have a queue of |
2304 | * unsent frames. But be careful about outgoing SACKS | |
2305 | * and IP options. | |
2306 | */ | |
0c54b85f | 2307 | mss_now = tcp_current_mss(sk); |
1da177e4 | 2308 | |
fe067e8a | 2309 | if (tcp_send_head(sk) != NULL) { |
a3433f35 | 2310 | TCP_SKB_CB(skb)->flags |= TCPHDR_FIN; |
1da177e4 LT |
2311 | TCP_SKB_CB(skb)->end_seq++; |
2312 | tp->write_seq++; | |
2313 | } else { | |
2314 | /* Socket is locked, keep trying until memory is available. */ | |
2315 | for (;;) { | |
aa133076 WF |
2316 | skb = alloc_skb_fclone(MAX_TCP_HEADER, |
2317 | sk->sk_allocation); | |
1da177e4 LT |
2318 | if (skb) |
2319 | break; | |
2320 | yield(); | |
2321 | } | |
2322 | ||
2323 | /* Reserve space for headers and prepare control bits. */ | |
2324 | skb_reserve(skb, MAX_TCP_HEADER); | |
1da177e4 | 2325 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ |
e870a8ef | 2326 | tcp_init_nondata_skb(skb, tp->write_seq, |
a3433f35 | 2327 | TCPHDR_ACK | TCPHDR_FIN); |
1da177e4 LT |
2328 | tcp_queue_skb(sk, skb); |
2329 | } | |
9e412ba7 | 2330 | __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); |
1da177e4 LT |
2331 | } |
2332 | ||
2333 | /* We get here when a process closes a file descriptor (either due to | |
2334 | * an explicit close() or as a byproduct of exit()'ing) and there | |
2335 | * was unread data in the receive queue. This behavior is recommended | |
65bb723c | 2336 | * by RFC 2525, section 2.17. -DaveM |
1da177e4 | 2337 | */ |
dd0fc66f | 2338 | void tcp_send_active_reset(struct sock *sk, gfp_t priority) |
1da177e4 | 2339 | { |
1da177e4 LT |
2340 | struct sk_buff *skb; |
2341 | ||
2342 | /* NOTE: No TCP options attached and we never retransmit this. */ | |
2343 | skb = alloc_skb(MAX_TCP_HEADER, priority); | |
2344 | if (!skb) { | |
4e673444 | 2345 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
1da177e4 LT |
2346 | return; |
2347 | } | |
2348 | ||
2349 | /* Reserve space for headers and prepare control bits. */ | |
2350 | skb_reserve(skb, MAX_TCP_HEADER); | |
e870a8ef | 2351 | tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), |
a3433f35 | 2352 | TCPHDR_ACK | TCPHDR_RST); |
1da177e4 | 2353 | /* Send it off. */ |
1da177e4 | 2354 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
dfb4b9dc | 2355 | if (tcp_transmit_skb(sk, skb, 0, priority)) |
4e673444 | 2356 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
26af65cb | 2357 | |
81cc8a75 | 2358 | TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); |
1da177e4 LT |
2359 | } |
2360 | ||
67edfef7 AK |
2361 | /* Send a crossed SYN-ACK during socket establishment. |
2362 | * WARNING: This routine must only be called when we have already sent | |
1da177e4 LT |
2363 | * a SYN packet that crossed the incoming SYN that caused this routine |
2364 | * to get called. If this assumption fails then the initial rcv_wnd | |
2365 | * and rcv_wscale values will not be correct. | |
2366 | */ | |
2367 | int tcp_send_synack(struct sock *sk) | |
2368 | { | |
056834d9 | 2369 | struct sk_buff *skb; |
1da177e4 | 2370 | |
fe067e8a | 2371 | skb = tcp_write_queue_head(sk); |
a3433f35 | 2372 | if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPHDR_SYN)) { |
1da177e4 LT |
2373 | printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); |
2374 | return -EFAULT; | |
2375 | } | |
a3433f35 | 2376 | if (!(TCP_SKB_CB(skb)->flags & TCPHDR_ACK)) { |
1da177e4 LT |
2377 | if (skb_cloned(skb)) { |
2378 | struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); | |
2379 | if (nskb == NULL) | |
2380 | return -ENOMEM; | |
fe067e8a | 2381 | tcp_unlink_write_queue(skb, sk); |
1da177e4 | 2382 | skb_header_release(nskb); |
fe067e8a | 2383 | __tcp_add_write_queue_head(sk, nskb); |
3ab224be HA |
2384 | sk_wmem_free_skb(sk, skb); |
2385 | sk->sk_wmem_queued += nskb->truesize; | |
2386 | sk_mem_charge(sk, nskb->truesize); | |
1da177e4 LT |
2387 | skb = nskb; |
2388 | } | |
2389 | ||
a3433f35 | 2390 | TCP_SKB_CB(skb)->flags |= TCPHDR_ACK; |
1da177e4 LT |
2391 | TCP_ECN_send_synack(tcp_sk(sk), skb); |
2392 | } | |
2393 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
dfb4b9dc | 2394 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
1da177e4 LT |
2395 | } |
2396 | ||
67edfef7 | 2397 | /* Prepare a SYN-ACK. */ |
056834d9 | 2398 | struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, |
e6b4d113 WAS |
2399 | struct request_sock *req, |
2400 | struct request_values *rvp) | |
1da177e4 | 2401 | { |
bd0388ae | 2402 | struct tcp_out_options opts; |
4957faad | 2403 | struct tcp_extend_values *xvp = tcp_xv(rvp); |
2e6599cb | 2404 | struct inet_request_sock *ireq = inet_rsk(req); |
1da177e4 | 2405 | struct tcp_sock *tp = tcp_sk(sk); |
28b2774a | 2406 | const struct tcp_cookie_values *cvp = tp->cookie_values; |
1da177e4 | 2407 | struct tcphdr *th; |
1da177e4 | 2408 | struct sk_buff *skb; |
cfb6eeb4 | 2409 | struct tcp_md5sig_key *md5; |
bd0388ae | 2410 | int tcp_header_size; |
f5fff5dc | 2411 | int mss; |
28b2774a | 2412 | int s_data_desired = 0; |
1da177e4 | 2413 | |
28b2774a ED |
2414 | if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired) |
2415 | s_data_desired = cvp->s_data_desired; | |
2416 | skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15 + s_data_desired, 1, GFP_ATOMIC); | |
1da177e4 LT |
2417 | if (skb == NULL) |
2418 | return NULL; | |
2419 | ||
2420 | /* Reserve space for headers. */ | |
2421 | skb_reserve(skb, MAX_TCP_HEADER); | |
2422 | ||
adf30907 | 2423 | skb_dst_set(skb, dst_clone(dst)); |
1da177e4 | 2424 | |
f5fff5dc TQ |
2425 | mss = dst_metric(dst, RTAX_ADVMSS); |
2426 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) | |
2427 | mss = tp->rx_opt.user_mss; | |
2428 | ||
33ad798c AL |
2429 | if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ |
2430 | __u8 rcv_wscale; | |
2431 | /* Set this up on the first call only */ | |
2432 | req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); | |
e88c64f0 HPP |
2433 | |
2434 | /* limit the window selection if the user enforce a smaller rx buffer */ | |
2435 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && | |
2436 | (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) | |
2437 | req->window_clamp = tcp_full_space(sk); | |
2438 | ||
33ad798c AL |
2439 | /* tcp_full_space because it is guaranteed to be the first packet */ |
2440 | tcp_select_initial_window(tcp_full_space(sk), | |
f5fff5dc | 2441 | mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), |
33ad798c AL |
2442 | &req->rcv_wnd, |
2443 | &req->window_clamp, | |
2444 | ireq->wscale_ok, | |
31d12926 | 2445 | &rcv_wscale, |
2446 | dst_metric(dst, RTAX_INITRWND)); | |
33ad798c AL |
2447 | ireq->rcv_wscale = rcv_wscale; |
2448 | } | |
2449 | ||
2450 | memset(&opts, 0, sizeof(opts)); | |
8b5f12d0 FW |
2451 | #ifdef CONFIG_SYN_COOKIES |
2452 | if (unlikely(req->cookie_ts)) | |
2453 | TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); | |
2454 | else | |
2455 | #endif | |
33ad798c | 2456 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
f5fff5dc | 2457 | tcp_header_size = tcp_synack_options(sk, req, mss, |
4957faad WAS |
2458 | skb, &opts, &md5, xvp) |
2459 | + sizeof(*th); | |
cfb6eeb4 | 2460 | |
aa8223c7 ACM |
2461 | skb_push(skb, tcp_header_size); |
2462 | skb_reset_transport_header(skb); | |
1da177e4 | 2463 | |
aa8223c7 | 2464 | th = tcp_hdr(skb); |
1da177e4 LT |
2465 | memset(th, 0, sizeof(struct tcphdr)); |
2466 | th->syn = 1; | |
2467 | th->ack = 1; | |
1da177e4 | 2468 | TCP_ECN_make_synack(req, th); |
a3116ac5 | 2469 | th->source = ireq->loc_port; |
2e6599cb | 2470 | th->dest = ireq->rmt_port; |
e870a8ef IJ |
2471 | /* Setting of flags are superfluous here for callers (and ECE is |
2472 | * not even correctly set) | |
2473 | */ | |
2474 | tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, | |
a3433f35 | 2475 | TCPHDR_SYN | TCPHDR_ACK); |
4957faad WAS |
2476 | |
2477 | if (OPTION_COOKIE_EXTENSION & opts.options) { | |
28b2774a ED |
2478 | if (s_data_desired) { |
2479 | u8 *buf = skb_put(skb, s_data_desired); | |
4957faad WAS |
2480 | |
2481 | /* copy data directly from the listening socket. */ | |
28b2774a ED |
2482 | memcpy(buf, cvp->s_data_payload, s_data_desired); |
2483 | TCP_SKB_CB(skb)->end_seq += s_data_desired; | |
4957faad WAS |
2484 | } |
2485 | ||
2486 | if (opts.hash_size > 0) { | |
2487 | __u32 workspace[SHA_WORKSPACE_WORDS]; | |
2488 | u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS]; | |
2489 | u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1]; | |
2490 | ||
2491 | /* Secret recipe depends on the Timestamp, (future) | |
2492 | * Sequence and Acknowledgment Numbers, Initiator | |
2493 | * Cookie, and others handled by IP variant caller. | |
2494 | */ | |
2495 | *tail-- ^= opts.tsval; | |
2496 | *tail-- ^= tcp_rsk(req)->rcv_isn + 1; | |
2497 | *tail-- ^= TCP_SKB_CB(skb)->seq + 1; | |
2498 | ||
2499 | /* recommended */ | |
0eae88f3 | 2500 | *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source); |
f9a2e69e | 2501 | *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */ |
4957faad WAS |
2502 | |
2503 | sha_transform((__u32 *)&xvp->cookie_bakery[0], | |
2504 | (char *)mess, | |
2505 | &workspace[0]); | |
2506 | opts.hash_location = | |
2507 | (__u8 *)&xvp->cookie_bakery[0]; | |
2508 | } | |
2509 | } | |
2510 | ||
1da177e4 | 2511 | th->seq = htonl(TCP_SKB_CB(skb)->seq); |
2e6599cb | 2512 | th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); |
1da177e4 LT |
2513 | |
2514 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ | |
600ff0c2 | 2515 | th->window = htons(min(req->rcv_wnd, 65535U)); |
bd0388ae | 2516 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
1da177e4 | 2517 | th->doff = (tcp_header_size >> 2); |
aa2ea058 | 2518 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb)); |
cfb6eeb4 YH |
2519 | |
2520 | #ifdef CONFIG_TCP_MD5SIG | |
2521 | /* Okay, we have all we need - do the md5 hash if needed */ | |
2522 | if (md5) { | |
bd0388ae | 2523 | tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, |
49a72dfb | 2524 | md5, NULL, req, skb); |
cfb6eeb4 YH |
2525 | } |
2526 | #endif | |
2527 | ||
1da177e4 LT |
2528 | return skb; |
2529 | } | |
4bc2f18b | 2530 | EXPORT_SYMBOL(tcp_make_synack); |
1da177e4 | 2531 | |
67edfef7 | 2532 | /* Do all connect socket setups that can be done AF independent. */ |
40efc6fa | 2533 | static void tcp_connect_init(struct sock *sk) |
1da177e4 LT |
2534 | { |
2535 | struct dst_entry *dst = __sk_dst_get(sk); | |
2536 | struct tcp_sock *tp = tcp_sk(sk); | |
2537 | __u8 rcv_wscale; | |
2538 | ||
2539 | /* We'll fix this up when we get a response from the other end. | |
2540 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. | |
2541 | */ | |
2542 | tp->tcp_header_len = sizeof(struct tcphdr) + | |
bb5b7c11 | 2543 | (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); |
1da177e4 | 2544 | |
cfb6eeb4 YH |
2545 | #ifdef CONFIG_TCP_MD5SIG |
2546 | if (tp->af_specific->md5_lookup(sk, sk) != NULL) | |
2547 | tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; | |
2548 | #endif | |
2549 | ||
1da177e4 LT |
2550 | /* If user gave his TCP_MAXSEG, record it to clamp */ |
2551 | if (tp->rx_opt.user_mss) | |
2552 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
2553 | tp->max_window = 0; | |
5d424d5a | 2554 | tcp_mtup_init(sk); |
1da177e4 LT |
2555 | tcp_sync_mss(sk, dst_mtu(dst)); |
2556 | ||
2557 | if (!tp->window_clamp) | |
2558 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); | |
2559 | tp->advmss = dst_metric(dst, RTAX_ADVMSS); | |
f5fff5dc TQ |
2560 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) |
2561 | tp->advmss = tp->rx_opt.user_mss; | |
2562 | ||
1da177e4 | 2563 | tcp_initialize_rcv_mss(sk); |
1da177e4 | 2564 | |
e88c64f0 HPP |
2565 | /* limit the window selection if the user enforce a smaller rx buffer */ |
2566 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && | |
2567 | (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) | |
2568 | tp->window_clamp = tcp_full_space(sk); | |
2569 | ||
1da177e4 LT |
2570 | tcp_select_initial_window(tcp_full_space(sk), |
2571 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), | |
2572 | &tp->rcv_wnd, | |
2573 | &tp->window_clamp, | |
bb5b7c11 | 2574 | sysctl_tcp_window_scaling, |
31d12926 | 2575 | &rcv_wscale, |
2576 | dst_metric(dst, RTAX_INITRWND)); | |
1da177e4 LT |
2577 | |
2578 | tp->rx_opt.rcv_wscale = rcv_wscale; | |
2579 | tp->rcv_ssthresh = tp->rcv_wnd; | |
2580 | ||
2581 | sk->sk_err = 0; | |
2582 | sock_reset_flag(sk, SOCK_DONE); | |
2583 | tp->snd_wnd = 0; | |
ee7537b6 | 2584 | tcp_init_wl(tp, 0); |
1da177e4 LT |
2585 | tp->snd_una = tp->write_seq; |
2586 | tp->snd_sml = tp->write_seq; | |
33f5f57e | 2587 | tp->snd_up = tp->write_seq; |
1da177e4 LT |
2588 | tp->rcv_nxt = 0; |
2589 | tp->rcv_wup = 0; | |
2590 | tp->copied_seq = 0; | |
2591 | ||
463c84b9 ACM |
2592 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
2593 | inet_csk(sk)->icsk_retransmits = 0; | |
1da177e4 LT |
2594 | tcp_clear_retrans(tp); |
2595 | } | |
2596 | ||
67edfef7 | 2597 | /* Build a SYN and send it off. */ |
1da177e4 LT |
2598 | int tcp_connect(struct sock *sk) |
2599 | { | |
2600 | struct tcp_sock *tp = tcp_sk(sk); | |
2601 | struct sk_buff *buff; | |
ee586811 | 2602 | int err; |
1da177e4 LT |
2603 | |
2604 | tcp_connect_init(sk); | |
2605 | ||
d179cd12 | 2606 | buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); |
1da177e4 LT |
2607 | if (unlikely(buff == NULL)) |
2608 | return -ENOBUFS; | |
2609 | ||
2610 | /* Reserve space for headers. */ | |
2611 | skb_reserve(buff, MAX_TCP_HEADER); | |
2612 | ||
bd37a088 | 2613 | tp->snd_nxt = tp->write_seq; |
a3433f35 | 2614 | tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); |
e870a8ef | 2615 | TCP_ECN_send_syn(sk, buff); |
1da177e4 LT |
2616 | |
2617 | /* Send it off. */ | |
2618 | TCP_SKB_CB(buff)->when = tcp_time_stamp; | |
2619 | tp->retrans_stamp = TCP_SKB_CB(buff)->when; | |
2620 | skb_header_release(buff); | |
fe067e8a | 2621 | __tcp_add_write_queue_tail(sk, buff); |
3ab224be HA |
2622 | sk->sk_wmem_queued += buff->truesize; |
2623 | sk_mem_charge(sk, buff->truesize); | |
1da177e4 | 2624 | tp->packets_out += tcp_skb_pcount(buff); |
ee586811 EP |
2625 | err = tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); |
2626 | if (err == -ECONNREFUSED) | |
2627 | return err; | |
bd37a088 WY |
2628 | |
2629 | /* We change tp->snd_nxt after the tcp_transmit_skb() call | |
2630 | * in order to make this packet get counted in tcpOutSegs. | |
2631 | */ | |
2632 | tp->snd_nxt = tp->write_seq; | |
2633 | tp->pushed_seq = tp->write_seq; | |
81cc8a75 | 2634 | TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); |
1da177e4 LT |
2635 | |
2636 | /* Timer for repeating the SYN until an answer. */ | |
3f421baa ACM |
2637 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
2638 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
1da177e4 LT |
2639 | return 0; |
2640 | } | |
4bc2f18b | 2641 | EXPORT_SYMBOL(tcp_connect); |
1da177e4 LT |
2642 | |
2643 | /* Send out a delayed ack, the caller does the policy checking | |
2644 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() | |
2645 | * for details. | |
2646 | */ | |
2647 | void tcp_send_delayed_ack(struct sock *sk) | |
2648 | { | |
463c84b9 ACM |
2649 | struct inet_connection_sock *icsk = inet_csk(sk); |
2650 | int ato = icsk->icsk_ack.ato; | |
1da177e4 LT |
2651 | unsigned long timeout; |
2652 | ||
2653 | if (ato > TCP_DELACK_MIN) { | |
463c84b9 | 2654 | const struct tcp_sock *tp = tcp_sk(sk); |
056834d9 | 2655 | int max_ato = HZ / 2; |
1da177e4 | 2656 | |
056834d9 IJ |
2657 | if (icsk->icsk_ack.pingpong || |
2658 | (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) | |
1da177e4 LT |
2659 | max_ato = TCP_DELACK_MAX; |
2660 | ||
2661 | /* Slow path, intersegment interval is "high". */ | |
2662 | ||
2663 | /* If some rtt estimate is known, use it to bound delayed ack. | |
463c84b9 | 2664 | * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements |
1da177e4 LT |
2665 | * directly. |
2666 | */ | |
2667 | if (tp->srtt) { | |
056834d9 | 2668 | int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); |
1da177e4 LT |
2669 | |
2670 | if (rtt < max_ato) | |
2671 | max_ato = rtt; | |
2672 | } | |
2673 | ||
2674 | ato = min(ato, max_ato); | |
2675 | } | |
2676 | ||
2677 | /* Stay within the limit we were given */ | |
2678 | timeout = jiffies + ato; | |
2679 | ||
2680 | /* Use new timeout only if there wasn't a older one earlier. */ | |
463c84b9 | 2681 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { |
1da177e4 LT |
2682 | /* If delack timer was blocked or is about to expire, |
2683 | * send ACK now. | |
2684 | */ | |
463c84b9 ACM |
2685 | if (icsk->icsk_ack.blocked || |
2686 | time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { | |
1da177e4 LT |
2687 | tcp_send_ack(sk); |
2688 | return; | |
2689 | } | |
2690 | ||
463c84b9 ACM |
2691 | if (!time_before(timeout, icsk->icsk_ack.timeout)) |
2692 | timeout = icsk->icsk_ack.timeout; | |
1da177e4 | 2693 | } |
463c84b9 ACM |
2694 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; |
2695 | icsk->icsk_ack.timeout = timeout; | |
2696 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); | |
1da177e4 LT |
2697 | } |
2698 | ||
2699 | /* This routine sends an ack and also updates the window. */ | |
2700 | void tcp_send_ack(struct sock *sk) | |
2701 | { | |
058dc334 | 2702 | struct sk_buff *buff; |
1da177e4 | 2703 | |
058dc334 IJ |
2704 | /* If we have been reset, we may not send again. */ |
2705 | if (sk->sk_state == TCP_CLOSE) | |
2706 | return; | |
1da177e4 | 2707 | |
058dc334 IJ |
2708 | /* We are not putting this on the write queue, so |
2709 | * tcp_transmit_skb() will set the ownership to this | |
2710 | * sock. | |
2711 | */ | |
2712 | buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); | |
2713 | if (buff == NULL) { | |
2714 | inet_csk_schedule_ack(sk); | |
2715 | inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; | |
2716 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, | |
2717 | TCP_DELACK_MAX, TCP_RTO_MAX); | |
2718 | return; | |
1da177e4 | 2719 | } |
058dc334 IJ |
2720 | |
2721 | /* Reserve space for headers and prepare control bits. */ | |
2722 | skb_reserve(buff, MAX_TCP_HEADER); | |
a3433f35 | 2723 | tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); |
058dc334 IJ |
2724 | |
2725 | /* Send it off, this clears delayed acks for us. */ | |
058dc334 IJ |
2726 | TCP_SKB_CB(buff)->when = tcp_time_stamp; |
2727 | tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); | |
1da177e4 LT |
2728 | } |
2729 | ||
2730 | /* This routine sends a packet with an out of date sequence | |
2731 | * number. It assumes the other end will try to ack it. | |
2732 | * | |
2733 | * Question: what should we make while urgent mode? | |
2734 | * 4.4BSD forces sending single byte of data. We cannot send | |
2735 | * out of window data, because we have SND.NXT==SND.MAX... | |
2736 | * | |
2737 | * Current solution: to send TWO zero-length segments in urgent mode: | |
2738 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is | |
2739 | * out-of-date with SND.UNA-1 to probe window. | |
2740 | */ | |
2741 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent) | |
2742 | { | |
2743 | struct tcp_sock *tp = tcp_sk(sk); | |
2744 | struct sk_buff *skb; | |
2745 | ||
2746 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ | |
2747 | skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); | |
e905a9ed | 2748 | if (skb == NULL) |
1da177e4 LT |
2749 | return -1; |
2750 | ||
2751 | /* Reserve space for headers and set control bits. */ | |
2752 | skb_reserve(skb, MAX_TCP_HEADER); | |
1da177e4 LT |
2753 | /* Use a previous sequence. This should cause the other |
2754 | * end to send an ack. Don't queue or clone SKB, just | |
2755 | * send it. | |
2756 | */ | |
a3433f35 | 2757 | tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); |
1da177e4 | 2758 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
dfb4b9dc | 2759 | return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); |
1da177e4 LT |
2760 | } |
2761 | ||
67edfef7 | 2762 | /* Initiate keepalive or window probe from timer. */ |
1da177e4 LT |
2763 | int tcp_write_wakeup(struct sock *sk) |
2764 | { | |
058dc334 IJ |
2765 | struct tcp_sock *tp = tcp_sk(sk); |
2766 | struct sk_buff *skb; | |
1da177e4 | 2767 | |
058dc334 IJ |
2768 | if (sk->sk_state == TCP_CLOSE) |
2769 | return -1; | |
2770 | ||
2771 | if ((skb = tcp_send_head(sk)) != NULL && | |
2772 | before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { | |
2773 | int err; | |
0c54b85f | 2774 | unsigned int mss = tcp_current_mss(sk); |
058dc334 IJ |
2775 | unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
2776 | ||
2777 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) | |
2778 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; | |
2779 | ||
2780 | /* We are probing the opening of a window | |
2781 | * but the window size is != 0 | |
2782 | * must have been a result SWS avoidance ( sender ) | |
2783 | */ | |
2784 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || | |
2785 | skb->len > mss) { | |
2786 | seg_size = min(seg_size, mss); | |
a3433f35 | 2787 | TCP_SKB_CB(skb)->flags |= TCPHDR_PSH; |
058dc334 IJ |
2788 | if (tcp_fragment(sk, skb, seg_size, mss)) |
2789 | return -1; | |
2790 | } else if (!tcp_skb_pcount(skb)) | |
2791 | tcp_set_skb_tso_segs(sk, skb, mss); | |
2792 | ||
a3433f35 | 2793 | TCP_SKB_CB(skb)->flags |= TCPHDR_PSH; |
058dc334 IJ |
2794 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
2795 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); | |
2796 | if (!err) | |
2797 | tcp_event_new_data_sent(sk, skb); | |
2798 | return err; | |
2799 | } else { | |
33f5f57e | 2800 | if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) |
058dc334 IJ |
2801 | tcp_xmit_probe_skb(sk, 1); |
2802 | return tcp_xmit_probe_skb(sk, 0); | |
1da177e4 | 2803 | } |
1da177e4 LT |
2804 | } |
2805 | ||
2806 | /* A window probe timeout has occurred. If window is not closed send | |
2807 | * a partial packet else a zero probe. | |
2808 | */ | |
2809 | void tcp_send_probe0(struct sock *sk) | |
2810 | { | |
463c84b9 | 2811 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2812 | struct tcp_sock *tp = tcp_sk(sk); |
2813 | int err; | |
2814 | ||
2815 | err = tcp_write_wakeup(sk); | |
2816 | ||
fe067e8a | 2817 | if (tp->packets_out || !tcp_send_head(sk)) { |
1da177e4 | 2818 | /* Cancel probe timer, if it is not required. */ |
6687e988 | 2819 | icsk->icsk_probes_out = 0; |
463c84b9 | 2820 | icsk->icsk_backoff = 0; |
1da177e4 LT |
2821 | return; |
2822 | } | |
2823 | ||
2824 | if (err <= 0) { | |
463c84b9 ACM |
2825 | if (icsk->icsk_backoff < sysctl_tcp_retries2) |
2826 | icsk->icsk_backoff++; | |
6687e988 | 2827 | icsk->icsk_probes_out++; |
e905a9ed | 2828 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
3f421baa ACM |
2829 | min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), |
2830 | TCP_RTO_MAX); | |
1da177e4 LT |
2831 | } else { |
2832 | /* If packet was not sent due to local congestion, | |
6687e988 | 2833 | * do not backoff and do not remember icsk_probes_out. |
1da177e4 LT |
2834 | * Let local senders to fight for local resources. |
2835 | * | |
2836 | * Use accumulated backoff yet. | |
2837 | */ | |
6687e988 ACM |
2838 | if (!icsk->icsk_probes_out) |
2839 | icsk->icsk_probes_out = 1; | |
e905a9ed | 2840 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
463c84b9 | 2841 | min(icsk->icsk_rto << icsk->icsk_backoff, |
3f421baa ACM |
2842 | TCP_RESOURCE_PROBE_INTERVAL), |
2843 | TCP_RTO_MAX); | |
1da177e4 LT |
2844 | } |
2845 | } |