]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * INET An implementation of the TCP/IP protocol suite for the LINUX | |
3 | * operating system. INET is implemented using the BSD Socket | |
4 | * interface as the means of communication with the user level. | |
5 | * | |
6 | * Implementation of the Transmission Control Protocol(TCP). | |
7 | * | |
02c30a84 | 8 | * Authors: Ross Biro |
1da177e4 LT |
9 | * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> |
10 | * Mark Evans, <evansmp@uhura.aston.ac.uk> | |
11 | * Corey Minyard <wf-rch!minyard@relay.EU.net> | |
12 | * Florian La Roche, <flla@stud.uni-sb.de> | |
13 | * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> | |
14 | * Linus Torvalds, <torvalds@cs.helsinki.fi> | |
15 | * Alan Cox, <gw4pts@gw4pts.ampr.org> | |
16 | * Matthew Dillon, <dillon@apollo.west.oic.com> | |
17 | * Arnt Gulbrandsen, <agulbra@nvg.unit.no> | |
18 | * Jorge Cwik, <jorge@laser.satlink.net> | |
19 | */ | |
20 | ||
21 | /* | |
22 | * Changes: Pedro Roque : Retransmit queue handled by TCP. | |
23 | * : Fragmentation on mtu decrease | |
24 | * : Segment collapse on retransmit | |
25 | * : AF independence | |
26 | * | |
27 | * Linus Torvalds : send_delayed_ack | |
28 | * David S. Miller : Charge memory using the right skb | |
29 | * during syn/ack processing. | |
30 | * David S. Miller : Output engine completely rewritten. | |
31 | * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. | |
32 | * Cacophonix Gaul : draft-minshall-nagle-01 | |
33 | * J Hadi Salim : ECN support | |
34 | * | |
35 | */ | |
36 | ||
37 | #include <net/tcp.h> | |
38 | ||
39 | #include <linux/compiler.h> | |
5a0e3ad6 | 40 | #include <linux/gfp.h> |
1da177e4 | 41 | #include <linux/module.h> |
1da177e4 LT |
42 | |
43 | /* People can turn this off for buggy TCP's found in printers etc. */ | |
ab32ea5d | 44 | int sysctl_tcp_retrans_collapse __read_mostly = 1; |
1da177e4 | 45 | |
09cb105e | 46 | /* People can turn this on to work with those rare, broken TCPs that |
15d99e02 RJ |
47 | * interpret the window field as a signed quantity. |
48 | */ | |
ab32ea5d | 49 | int sysctl_tcp_workaround_signed_windows __read_mostly = 0; |
15d99e02 | 50 | |
1da177e4 LT |
51 | /* This limits the percentage of the congestion window which we |
52 | * will allow a single TSO frame to consume. Building TSO frames | |
53 | * which are too large can cause TCP streams to be bursty. | |
54 | */ | |
ab32ea5d | 55 | int sysctl_tcp_tso_win_divisor __read_mostly = 3; |
1da177e4 | 56 | |
ab32ea5d BH |
57 | int sysctl_tcp_mtu_probing __read_mostly = 0; |
58 | int sysctl_tcp_base_mss __read_mostly = 512; | |
5d424d5a | 59 | |
35089bb2 | 60 | /* By default, RFC2861 behavior. */ |
ab32ea5d | 61 | int sysctl_tcp_slow_start_after_idle __read_mostly = 1; |
35089bb2 | 62 | |
519855c5 | 63 | int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */ |
e6b09cca | 64 | EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size); |
519855c5 WAS |
65 | |
66 | ||
67edfef7 | 67 | /* Account for new data that has been sent to the network. */ |
66f5fe62 | 68 | static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 69 | { |
9e412ba7 | 70 | struct tcp_sock *tp = tcp_sk(sk); |
66f5fe62 | 71 | unsigned int prior_packets = tp->packets_out; |
9e412ba7 | 72 | |
fe067e8a | 73 | tcp_advance_send_head(sk, skb); |
1da177e4 | 74 | tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; |
8512430e IJ |
75 | |
76 | /* Don't override Nagle indefinately with F-RTO */ | |
77 | if (tp->frto_counter == 2) | |
78 | tp->frto_counter = 3; | |
66f5fe62 IJ |
79 | |
80 | tp->packets_out += tcp_skb_pcount(skb); | |
81 | if (!prior_packets) | |
82 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, | |
83 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
1da177e4 LT |
84 | } |
85 | ||
86 | /* SND.NXT, if window was not shrunk. | |
87 | * If window has been shrunk, what should we make? It is not clear at all. | |
88 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( | |
89 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already | |
90 | * invalid. OK, let's make this for now: | |
91 | */ | |
9e412ba7 | 92 | static inline __u32 tcp_acceptable_seq(struct sock *sk) |
1da177e4 | 93 | { |
9e412ba7 IJ |
94 | struct tcp_sock *tp = tcp_sk(sk); |
95 | ||
90840def | 96 | if (!before(tcp_wnd_end(tp), tp->snd_nxt)) |
1da177e4 LT |
97 | return tp->snd_nxt; |
98 | else | |
90840def | 99 | return tcp_wnd_end(tp); |
1da177e4 LT |
100 | } |
101 | ||
102 | /* Calculate mss to advertise in SYN segment. | |
103 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: | |
104 | * | |
105 | * 1. It is independent of path mtu. | |
106 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. | |
107 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of | |
108 | * attached devices, because some buggy hosts are confused by | |
109 | * large MSS. | |
110 | * 4. We do not make 3, we advertise MSS, calculated from first | |
111 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. | |
112 | * This may be overridden via information stored in routing table. | |
113 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, | |
114 | * probably even Jumbo". | |
115 | */ | |
116 | static __u16 tcp_advertise_mss(struct sock *sk) | |
117 | { | |
118 | struct tcp_sock *tp = tcp_sk(sk); | |
119 | struct dst_entry *dst = __sk_dst_get(sk); | |
120 | int mss = tp->advmss; | |
121 | ||
122 | if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) { | |
123 | mss = dst_metric(dst, RTAX_ADVMSS); | |
124 | tp->advmss = mss; | |
125 | } | |
126 | ||
127 | return (__u16)mss; | |
128 | } | |
129 | ||
130 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". | |
131 | * This is the first part of cwnd validation mechanism. */ | |
463c84b9 | 132 | static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst) |
1da177e4 | 133 | { |
463c84b9 | 134 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
135 | s32 delta = tcp_time_stamp - tp->lsndtime; |
136 | u32 restart_cwnd = tcp_init_cwnd(tp, dst); | |
137 | u32 cwnd = tp->snd_cwnd; | |
138 | ||
6687e988 | 139 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); |
1da177e4 | 140 | |
6687e988 | 141 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
1da177e4 LT |
142 | restart_cwnd = min(restart_cwnd, cwnd); |
143 | ||
463c84b9 | 144 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) |
1da177e4 LT |
145 | cwnd >>= 1; |
146 | tp->snd_cwnd = max(cwnd, restart_cwnd); | |
147 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
148 | tp->snd_cwnd_used = 0; | |
149 | } | |
150 | ||
67edfef7 | 151 | /* Congestion state accounting after a packet has been sent. */ |
40efc6fa SH |
152 | static void tcp_event_data_sent(struct tcp_sock *tp, |
153 | struct sk_buff *skb, struct sock *sk) | |
1da177e4 | 154 | { |
463c84b9 ACM |
155 | struct inet_connection_sock *icsk = inet_csk(sk); |
156 | const u32 now = tcp_time_stamp; | |
1da177e4 | 157 | |
35089bb2 DM |
158 | if (sysctl_tcp_slow_start_after_idle && |
159 | (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) | |
463c84b9 | 160 | tcp_cwnd_restart(sk, __sk_dst_get(sk)); |
1da177e4 LT |
161 | |
162 | tp->lsndtime = now; | |
163 | ||
164 | /* If it is a reply for ato after last received | |
165 | * packet, enter pingpong mode. | |
166 | */ | |
463c84b9 ACM |
167 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) |
168 | icsk->icsk_ack.pingpong = 1; | |
1da177e4 LT |
169 | } |
170 | ||
67edfef7 | 171 | /* Account for an ACK we sent. */ |
40efc6fa | 172 | static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) |
1da177e4 | 173 | { |
463c84b9 ACM |
174 | tcp_dec_quickack_mode(sk, pkts); |
175 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); | |
1da177e4 LT |
176 | } |
177 | ||
178 | /* Determine a window scaling and initial window to offer. | |
179 | * Based on the assumption that the given amount of space | |
180 | * will be offered. Store the results in the tp structure. | |
181 | * NOTE: for smooth operation initial space offering should | |
182 | * be a multiple of mss if possible. We assume here that mss >= 1. | |
183 | * This MUST be enforced by all callers. | |
184 | */ | |
185 | void tcp_select_initial_window(int __space, __u32 mss, | |
186 | __u32 *rcv_wnd, __u32 *window_clamp, | |
31d12926 | 187 | int wscale_ok, __u8 *rcv_wscale, |
188 | __u32 init_rcv_wnd) | |
1da177e4 LT |
189 | { |
190 | unsigned int space = (__space < 0 ? 0 : __space); | |
191 | ||
192 | /* If no clamp set the clamp to the max possible scaled window */ | |
193 | if (*window_clamp == 0) | |
194 | (*window_clamp) = (65535 << 14); | |
195 | space = min(*window_clamp, space); | |
196 | ||
197 | /* Quantize space offering to a multiple of mss if possible. */ | |
198 | if (space > mss) | |
199 | space = (space / mss) * mss; | |
200 | ||
201 | /* NOTE: offering an initial window larger than 32767 | |
15d99e02 RJ |
202 | * will break some buggy TCP stacks. If the admin tells us |
203 | * it is likely we could be speaking with such a buggy stack | |
204 | * we will truncate our initial window offering to 32K-1 | |
205 | * unless the remote has sent us a window scaling option, | |
206 | * which we interpret as a sign the remote TCP is not | |
207 | * misinterpreting the window field as a signed quantity. | |
1da177e4 | 208 | */ |
15d99e02 RJ |
209 | if (sysctl_tcp_workaround_signed_windows) |
210 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); | |
211 | else | |
212 | (*rcv_wnd) = space; | |
213 | ||
1da177e4 LT |
214 | (*rcv_wscale) = 0; |
215 | if (wscale_ok) { | |
216 | /* Set window scaling on max possible window | |
e905a9ed | 217 | * See RFC1323 for an explanation of the limit to 14 |
1da177e4 LT |
218 | */ |
219 | space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); | |
316c1592 | 220 | space = min_t(u32, space, *window_clamp); |
1da177e4 LT |
221 | while (space > 65535 && (*rcv_wscale) < 14) { |
222 | space >>= 1; | |
223 | (*rcv_wscale)++; | |
224 | } | |
225 | } | |
226 | ||
0705c6f0 | 227 | /* Set initial window to value enough for senders, following RFC5681. */ |
056834d9 | 228 | if (mss > (1 << *rcv_wscale)) { |
0705c6f0 GR |
229 | int init_cwnd = rfc3390_bytes_to_packets(mss); |
230 | ||
31d12926 | 231 | /* when initializing use the value from init_rcv_wnd |
232 | * rather than the default from above | |
233 | */ | |
234 | if (init_rcv_wnd && | |
235 | (*rcv_wnd > init_rcv_wnd * mss)) | |
236 | *rcv_wnd = init_rcv_wnd * mss; | |
237 | else if (*rcv_wnd > init_cwnd * mss) | |
056834d9 | 238 | *rcv_wnd = init_cwnd * mss; |
1da177e4 LT |
239 | } |
240 | ||
241 | /* Set the clamp no higher than max representable value */ | |
242 | (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); | |
243 | } | |
4bc2f18b | 244 | EXPORT_SYMBOL(tcp_select_initial_window); |
1da177e4 LT |
245 | |
246 | /* Chose a new window to advertise, update state in tcp_sock for the | |
247 | * socket, and return result with RFC1323 scaling applied. The return | |
248 | * value can be stuffed directly into th->window for an outgoing | |
249 | * frame. | |
250 | */ | |
40efc6fa | 251 | static u16 tcp_select_window(struct sock *sk) |
1da177e4 LT |
252 | { |
253 | struct tcp_sock *tp = tcp_sk(sk); | |
254 | u32 cur_win = tcp_receive_window(tp); | |
255 | u32 new_win = __tcp_select_window(sk); | |
256 | ||
257 | /* Never shrink the offered window */ | |
2de979bd | 258 | if (new_win < cur_win) { |
1da177e4 LT |
259 | /* Danger Will Robinson! |
260 | * Don't update rcv_wup/rcv_wnd here or else | |
261 | * we will not be able to advertise a zero | |
262 | * window in time. --DaveM | |
263 | * | |
264 | * Relax Will Robinson. | |
265 | */ | |
607bfbf2 | 266 | new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); |
1da177e4 LT |
267 | } |
268 | tp->rcv_wnd = new_win; | |
269 | tp->rcv_wup = tp->rcv_nxt; | |
270 | ||
271 | /* Make sure we do not exceed the maximum possible | |
272 | * scaled window. | |
273 | */ | |
15d99e02 | 274 | if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) |
1da177e4 LT |
275 | new_win = min(new_win, MAX_TCP_WINDOW); |
276 | else | |
277 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); | |
278 | ||
279 | /* RFC1323 scaling applied */ | |
280 | new_win >>= tp->rx_opt.rcv_wscale; | |
281 | ||
282 | /* If we advertise zero window, disable fast path. */ | |
283 | if (new_win == 0) | |
284 | tp->pred_flags = 0; | |
285 | ||
286 | return new_win; | |
287 | } | |
288 | ||
67edfef7 | 289 | /* Packet ECN state for a SYN-ACK */ |
056834d9 | 290 | static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb) |
bdf1ee5d | 291 | { |
a3433f35 | 292 | TCP_SKB_CB(skb)->flags &= ~TCPHDR_CWR; |
056834d9 | 293 | if (!(tp->ecn_flags & TCP_ECN_OK)) |
a3433f35 | 294 | TCP_SKB_CB(skb)->flags &= ~TCPHDR_ECE; |
bdf1ee5d IJ |
295 | } |
296 | ||
67edfef7 | 297 | /* Packet ECN state for a SYN. */ |
bdf1ee5d IJ |
298 | static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) |
299 | { | |
300 | struct tcp_sock *tp = tcp_sk(sk); | |
301 | ||
302 | tp->ecn_flags = 0; | |
255cac91 | 303 | if (sysctl_tcp_ecn == 1) { |
a3433f35 | 304 | TCP_SKB_CB(skb)->flags |= TCPHDR_ECE | TCPHDR_CWR; |
bdf1ee5d IJ |
305 | tp->ecn_flags = TCP_ECN_OK; |
306 | } | |
307 | } | |
308 | ||
309 | static __inline__ void | |
310 | TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th) | |
311 | { | |
312 | if (inet_rsk(req)->ecn_ok) | |
313 | th->ece = 1; | |
314 | } | |
315 | ||
67edfef7 AK |
316 | /* Set up ECN state for a packet on a ESTABLISHED socket that is about to |
317 | * be sent. | |
318 | */ | |
bdf1ee5d IJ |
319 | static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, |
320 | int tcp_header_len) | |
321 | { | |
322 | struct tcp_sock *tp = tcp_sk(sk); | |
323 | ||
324 | if (tp->ecn_flags & TCP_ECN_OK) { | |
325 | /* Not-retransmitted data segment: set ECT and inject CWR. */ | |
326 | if (skb->len != tcp_header_len && | |
327 | !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { | |
328 | INET_ECN_xmit(sk); | |
056834d9 | 329 | if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { |
bdf1ee5d IJ |
330 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; |
331 | tcp_hdr(skb)->cwr = 1; | |
332 | skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; | |
333 | } | |
334 | } else { | |
335 | /* ACK or retransmitted segment: clear ECT|CE */ | |
336 | INET_ECN_dontxmit(sk); | |
337 | } | |
338 | if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) | |
339 | tcp_hdr(skb)->ece = 1; | |
340 | } | |
341 | } | |
342 | ||
e870a8ef IJ |
343 | /* Constructs common control bits of non-data skb. If SYN/FIN is present, |
344 | * auto increment end seqno. | |
345 | */ | |
346 | static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) | |
347 | { | |
2e8e18ef | 348 | skb->ip_summed = CHECKSUM_PARTIAL; |
e870a8ef IJ |
349 | skb->csum = 0; |
350 | ||
351 | TCP_SKB_CB(skb)->flags = flags; | |
352 | TCP_SKB_CB(skb)->sacked = 0; | |
353 | ||
354 | skb_shinfo(skb)->gso_segs = 1; | |
355 | skb_shinfo(skb)->gso_size = 0; | |
356 | skb_shinfo(skb)->gso_type = 0; | |
357 | ||
358 | TCP_SKB_CB(skb)->seq = seq; | |
a3433f35 | 359 | if (flags & (TCPHDR_SYN | TCPHDR_FIN)) |
e870a8ef IJ |
360 | seq++; |
361 | TCP_SKB_CB(skb)->end_seq = seq; | |
362 | } | |
363 | ||
33f5f57e IJ |
364 | static inline int tcp_urg_mode(const struct tcp_sock *tp) |
365 | { | |
366 | return tp->snd_una != tp->snd_up; | |
367 | } | |
368 | ||
33ad798c AL |
369 | #define OPTION_SACK_ADVERTISE (1 << 0) |
370 | #define OPTION_TS (1 << 1) | |
371 | #define OPTION_MD5 (1 << 2) | |
89e95a61 | 372 | #define OPTION_WSCALE (1 << 3) |
bd0388ae | 373 | #define OPTION_COOKIE_EXTENSION (1 << 4) |
33ad798c AL |
374 | |
375 | struct tcp_out_options { | |
376 | u8 options; /* bit field of OPTION_* */ | |
377 | u8 ws; /* window scale, 0 to disable */ | |
378 | u8 num_sack_blocks; /* number of SACK blocks to include */ | |
bd0388ae | 379 | u8 hash_size; /* bytes in hash_location */ |
33ad798c AL |
380 | u16 mss; /* 0 to disable */ |
381 | __u32 tsval, tsecr; /* need to include OPTION_TS */ | |
bd0388ae | 382 | __u8 *hash_location; /* temporary pointer, overloaded */ |
33ad798c AL |
383 | }; |
384 | ||
bd0388ae WAS |
385 | /* The sysctl int routines are generic, so check consistency here. |
386 | */ | |
387 | static u8 tcp_cookie_size_check(u8 desired) | |
388 | { | |
389 | if (desired > 0) { | |
390 | /* previously specified */ | |
391 | return desired; | |
392 | } | |
393 | if (sysctl_tcp_cookie_size <= 0) { | |
394 | /* no default specified */ | |
395 | return 0; | |
396 | } | |
397 | if (sysctl_tcp_cookie_size <= TCP_COOKIE_MIN) { | |
398 | /* value too small, specify minimum */ | |
399 | return TCP_COOKIE_MIN; | |
400 | } | |
401 | if (sysctl_tcp_cookie_size >= TCP_COOKIE_MAX) { | |
402 | /* value too large, specify maximum */ | |
403 | return TCP_COOKIE_MAX; | |
404 | } | |
405 | if (0x1 & sysctl_tcp_cookie_size) { | |
406 | /* 8-bit multiple, illegal, fix it */ | |
407 | return (u8)(sysctl_tcp_cookie_size + 0x1); | |
408 | } | |
409 | return (u8)sysctl_tcp_cookie_size; | |
410 | } | |
411 | ||
67edfef7 AK |
412 | /* Write previously computed TCP options to the packet. |
413 | * | |
414 | * Beware: Something in the Internet is very sensitive to the ordering of | |
fd6149d3 IJ |
415 | * TCP options, we learned this through the hard way, so be careful here. |
416 | * Luckily we can at least blame others for their non-compliance but from | |
417 | * inter-operatibility perspective it seems that we're somewhat stuck with | |
418 | * the ordering which we have been using if we want to keep working with | |
419 | * those broken things (not that it currently hurts anybody as there isn't | |
420 | * particular reason why the ordering would need to be changed). | |
421 | * | |
422 | * At least SACK_PERM as the first option is known to lead to a disaster | |
423 | * (but it may well be that other scenarios fail similarly). | |
424 | */ | |
33ad798c | 425 | static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, |
bd0388ae WAS |
426 | struct tcp_out_options *opts) |
427 | { | |
428 | u8 options = opts->options; /* mungable copy */ | |
429 | ||
430 | /* Having both authentication and cookies for security is redundant, | |
431 | * and there's certainly not enough room. Instead, the cookie-less | |
432 | * extension variant is proposed. | |
433 | * | |
434 | * Consider the pessimal case with authentication. The options | |
435 | * could look like: | |
436 | * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40 | |
437 | */ | |
438 | if (unlikely(OPTION_MD5 & options)) { | |
439 | if (unlikely(OPTION_COOKIE_EXTENSION & options)) { | |
440 | *ptr++ = htonl((TCPOPT_COOKIE << 24) | | |
441 | (TCPOLEN_COOKIE_BASE << 16) | | |
442 | (TCPOPT_MD5SIG << 8) | | |
443 | TCPOLEN_MD5SIG); | |
444 | } else { | |
445 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
446 | (TCPOPT_NOP << 16) | | |
447 | (TCPOPT_MD5SIG << 8) | | |
448 | TCPOLEN_MD5SIG); | |
449 | } | |
450 | options &= ~OPTION_COOKIE_EXTENSION; | |
451 | /* overload cookie hash location */ | |
452 | opts->hash_location = (__u8 *)ptr; | |
33ad798c | 453 | ptr += 4; |
40efc6fa | 454 | } |
33ad798c | 455 | |
fd6149d3 IJ |
456 | if (unlikely(opts->mss)) { |
457 | *ptr++ = htonl((TCPOPT_MSS << 24) | | |
458 | (TCPOLEN_MSS << 16) | | |
459 | opts->mss); | |
460 | } | |
461 | ||
bd0388ae WAS |
462 | if (likely(OPTION_TS & options)) { |
463 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { | |
33ad798c AL |
464 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
465 | (TCPOLEN_SACK_PERM << 16) | | |
466 | (TCPOPT_TIMESTAMP << 8) | | |
467 | TCPOLEN_TIMESTAMP); | |
bd0388ae | 468 | options &= ~OPTION_SACK_ADVERTISE; |
33ad798c AL |
469 | } else { |
470 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
471 | (TCPOPT_NOP << 16) | | |
472 | (TCPOPT_TIMESTAMP << 8) | | |
473 | TCPOLEN_TIMESTAMP); | |
474 | } | |
475 | *ptr++ = htonl(opts->tsval); | |
476 | *ptr++ = htonl(opts->tsecr); | |
477 | } | |
478 | ||
bd0388ae WAS |
479 | /* Specification requires after timestamp, so do it now. |
480 | * | |
481 | * Consider the pessimal case without authentication. The options | |
482 | * could look like: | |
483 | * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40 | |
484 | */ | |
485 | if (unlikely(OPTION_COOKIE_EXTENSION & options)) { | |
486 | __u8 *cookie_copy = opts->hash_location; | |
487 | u8 cookie_size = opts->hash_size; | |
488 | ||
489 | /* 8-bit multiple handled in tcp_cookie_size_check() above, | |
490 | * and elsewhere. | |
491 | */ | |
492 | if (0x2 & cookie_size) { | |
493 | __u8 *p = (__u8 *)ptr; | |
494 | ||
495 | /* 16-bit multiple */ | |
496 | *p++ = TCPOPT_COOKIE; | |
497 | *p++ = TCPOLEN_COOKIE_BASE + cookie_size; | |
498 | *p++ = *cookie_copy++; | |
499 | *p++ = *cookie_copy++; | |
500 | ptr++; | |
501 | cookie_size -= 2; | |
502 | } else { | |
503 | /* 32-bit multiple */ | |
504 | *ptr++ = htonl(((TCPOPT_NOP << 24) | | |
505 | (TCPOPT_NOP << 16) | | |
506 | (TCPOPT_COOKIE << 8) | | |
507 | TCPOLEN_COOKIE_BASE) + | |
508 | cookie_size); | |
509 | } | |
510 | ||
511 | if (cookie_size > 0) { | |
512 | memcpy(ptr, cookie_copy, cookie_size); | |
513 | ptr += (cookie_size / 4); | |
514 | } | |
515 | } | |
516 | ||
517 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { | |
33ad798c AL |
518 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
519 | (TCPOPT_NOP << 16) | | |
520 | (TCPOPT_SACK_PERM << 8) | | |
521 | TCPOLEN_SACK_PERM); | |
522 | } | |
523 | ||
bd0388ae | 524 | if (unlikely(OPTION_WSCALE & options)) { |
33ad798c AL |
525 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
526 | (TCPOPT_WINDOW << 16) | | |
527 | (TCPOLEN_WINDOW << 8) | | |
528 | opts->ws); | |
529 | } | |
530 | ||
531 | if (unlikely(opts->num_sack_blocks)) { | |
532 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? | |
533 | tp->duplicate_sack : tp->selective_acks; | |
40efc6fa SH |
534 | int this_sack; |
535 | ||
536 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
537 | (TCPOPT_NOP << 16) | | |
538 | (TCPOPT_SACK << 8) | | |
33ad798c | 539 | (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * |
40efc6fa | 540 | TCPOLEN_SACK_PERBLOCK))); |
2de979bd | 541 | |
33ad798c AL |
542 | for (this_sack = 0; this_sack < opts->num_sack_blocks; |
543 | ++this_sack) { | |
40efc6fa SH |
544 | *ptr++ = htonl(sp[this_sack].start_seq); |
545 | *ptr++ = htonl(sp[this_sack].end_seq); | |
546 | } | |
2de979bd | 547 | |
5861f8e5 | 548 | tp->rx_opt.dsack = 0; |
40efc6fa | 549 | } |
33ad798c AL |
550 | } |
551 | ||
67edfef7 AK |
552 | /* Compute TCP options for SYN packets. This is not the final |
553 | * network wire format yet. | |
554 | */ | |
33ad798c AL |
555 | static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb, |
556 | struct tcp_out_options *opts, | |
557 | struct tcp_md5sig_key **md5) { | |
558 | struct tcp_sock *tp = tcp_sk(sk); | |
bd0388ae | 559 | struct tcp_cookie_values *cvp = tp->cookie_values; |
bd0388ae WAS |
560 | unsigned remaining = MAX_TCP_OPTION_SPACE; |
561 | u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ? | |
562 | tcp_cookie_size_check(cvp->cookie_desired) : | |
563 | 0; | |
33ad798c | 564 | |
cfb6eeb4 | 565 | #ifdef CONFIG_TCP_MD5SIG |
33ad798c AL |
566 | *md5 = tp->af_specific->md5_lookup(sk, sk); |
567 | if (*md5) { | |
568 | opts->options |= OPTION_MD5; | |
bd0388ae | 569 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
cfb6eeb4 | 570 | } |
33ad798c AL |
571 | #else |
572 | *md5 = NULL; | |
cfb6eeb4 | 573 | #endif |
33ad798c AL |
574 | |
575 | /* We always get an MSS option. The option bytes which will be seen in | |
576 | * normal data packets should timestamps be used, must be in the MSS | |
577 | * advertised. But we subtract them from tp->mss_cache so that | |
578 | * calculations in tcp_sendmsg are simpler etc. So account for this | |
579 | * fact here if necessary. If we don't do this correctly, as a | |
580 | * receiver we won't recognize data packets as being full sized when we | |
581 | * should, and thus we won't abide by the delayed ACK rules correctly. | |
582 | * SACKs don't matter, we never delay an ACK when we have any of those | |
583 | * going out. */ | |
584 | opts->mss = tcp_advertise_mss(sk); | |
bd0388ae | 585 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c | 586 | |
bb5b7c11 | 587 | if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { |
33ad798c AL |
588 | opts->options |= OPTION_TS; |
589 | opts->tsval = TCP_SKB_CB(skb)->when; | |
590 | opts->tsecr = tp->rx_opt.ts_recent; | |
bd0388ae | 591 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c | 592 | } |
bb5b7c11 | 593 | if (likely(sysctl_tcp_window_scaling)) { |
33ad798c | 594 | opts->ws = tp->rx_opt.rcv_wscale; |
89e95a61 | 595 | opts->options |= OPTION_WSCALE; |
bd0388ae | 596 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 597 | } |
bb5b7c11 | 598 | if (likely(sysctl_tcp_sack)) { |
33ad798c | 599 | opts->options |= OPTION_SACK_ADVERTISE; |
b32d1310 | 600 | if (unlikely(!(OPTION_TS & opts->options))) |
bd0388ae | 601 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c AL |
602 | } |
603 | ||
bd0388ae WAS |
604 | /* Note that timestamps are required by the specification. |
605 | * | |
606 | * Odd numbers of bytes are prohibited by the specification, ensuring | |
607 | * that the cookie is 16-bit aligned, and the resulting cookie pair is | |
608 | * 32-bit aligned. | |
609 | */ | |
610 | if (*md5 == NULL && | |
611 | (OPTION_TS & opts->options) && | |
612 | cookie_size > 0) { | |
613 | int need = TCPOLEN_COOKIE_BASE + cookie_size; | |
614 | ||
615 | if (0x2 & need) { | |
616 | /* 32-bit multiple */ | |
617 | need += 2; /* NOPs */ | |
618 | ||
619 | if (need > remaining) { | |
620 | /* try shrinking cookie to fit */ | |
621 | cookie_size -= 2; | |
622 | need -= 4; | |
623 | } | |
624 | } | |
625 | while (need > remaining && TCP_COOKIE_MIN <= cookie_size) { | |
626 | cookie_size -= 4; | |
627 | need -= 4; | |
628 | } | |
629 | if (TCP_COOKIE_MIN <= cookie_size) { | |
630 | opts->options |= OPTION_COOKIE_EXTENSION; | |
631 | opts->hash_location = (__u8 *)&cvp->cookie_pair[0]; | |
632 | opts->hash_size = cookie_size; | |
633 | ||
634 | /* Remember for future incarnations. */ | |
635 | cvp->cookie_desired = cookie_size; | |
636 | ||
637 | if (cvp->cookie_desired != cvp->cookie_pair_size) { | |
638 | /* Currently use random bytes as a nonce, | |
639 | * assuming these are completely unpredictable | |
640 | * by hostile users of the same system. | |
641 | */ | |
642 | get_random_bytes(&cvp->cookie_pair[0], | |
643 | cookie_size); | |
644 | cvp->cookie_pair_size = cookie_size; | |
645 | } | |
646 | ||
647 | remaining -= need; | |
648 | } | |
649 | } | |
650 | return MAX_TCP_OPTION_SPACE - remaining; | |
40efc6fa SH |
651 | } |
652 | ||
67edfef7 | 653 | /* Set up TCP options for SYN-ACKs. */ |
33ad798c AL |
654 | static unsigned tcp_synack_options(struct sock *sk, |
655 | struct request_sock *req, | |
656 | unsigned mss, struct sk_buff *skb, | |
657 | struct tcp_out_options *opts, | |
4957faad WAS |
658 | struct tcp_md5sig_key **md5, |
659 | struct tcp_extend_values *xvp) | |
660 | { | |
33ad798c | 661 | struct inet_request_sock *ireq = inet_rsk(req); |
4957faad WAS |
662 | unsigned remaining = MAX_TCP_OPTION_SPACE; |
663 | u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ? | |
664 | xvp->cookie_plus : | |
665 | 0; | |
33ad798c | 666 | |
cfb6eeb4 | 667 | #ifdef CONFIG_TCP_MD5SIG |
33ad798c AL |
668 | *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); |
669 | if (*md5) { | |
670 | opts->options |= OPTION_MD5; | |
4957faad WAS |
671 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
672 | ||
673 | /* We can't fit any SACK blocks in a packet with MD5 + TS | |
674 | * options. There was discussion about disabling SACK | |
675 | * rather than TS in order to fit in better with old, | |
676 | * buggy kernels, but that was deemed to be unnecessary. | |
677 | */ | |
de213e5e | 678 | ireq->tstamp_ok &= !ireq->sack_ok; |
cfb6eeb4 | 679 | } |
33ad798c AL |
680 | #else |
681 | *md5 = NULL; | |
cfb6eeb4 | 682 | #endif |
33ad798c | 683 | |
4957faad | 684 | /* We always send an MSS option. */ |
33ad798c | 685 | opts->mss = mss; |
4957faad | 686 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c AL |
687 | |
688 | if (likely(ireq->wscale_ok)) { | |
689 | opts->ws = ireq->rcv_wscale; | |
89e95a61 | 690 | opts->options |= OPTION_WSCALE; |
4957faad | 691 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 692 | } |
de213e5e | 693 | if (likely(ireq->tstamp_ok)) { |
33ad798c AL |
694 | opts->options |= OPTION_TS; |
695 | opts->tsval = TCP_SKB_CB(skb)->when; | |
696 | opts->tsecr = req->ts_recent; | |
4957faad | 697 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c AL |
698 | } |
699 | if (likely(ireq->sack_ok)) { | |
700 | opts->options |= OPTION_SACK_ADVERTISE; | |
de213e5e | 701 | if (unlikely(!ireq->tstamp_ok)) |
4957faad | 702 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c AL |
703 | } |
704 | ||
4957faad WAS |
705 | /* Similar rationale to tcp_syn_options() applies here, too. |
706 | * If the <SYN> options fit, the same options should fit now! | |
707 | */ | |
708 | if (*md5 == NULL && | |
de213e5e | 709 | ireq->tstamp_ok && |
4957faad WAS |
710 | cookie_plus > TCPOLEN_COOKIE_BASE) { |
711 | int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */ | |
712 | ||
713 | if (0x2 & need) { | |
714 | /* 32-bit multiple */ | |
715 | need += 2; /* NOPs */ | |
716 | } | |
717 | if (need <= remaining) { | |
718 | opts->options |= OPTION_COOKIE_EXTENSION; | |
719 | opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE; | |
720 | remaining -= need; | |
721 | } else { | |
722 | /* There's no error return, so flag it. */ | |
723 | xvp->cookie_out_never = 1; /* true */ | |
724 | opts->hash_size = 0; | |
725 | } | |
726 | } | |
727 | return MAX_TCP_OPTION_SPACE - remaining; | |
33ad798c AL |
728 | } |
729 | ||
67edfef7 AK |
730 | /* Compute TCP options for ESTABLISHED sockets. This is not the |
731 | * final wire format yet. | |
732 | */ | |
33ad798c AL |
733 | static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb, |
734 | struct tcp_out_options *opts, | |
735 | struct tcp_md5sig_key **md5) { | |
736 | struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; | |
737 | struct tcp_sock *tp = tcp_sk(sk); | |
738 | unsigned size = 0; | |
cabeccbd | 739 | unsigned int eff_sacks; |
33ad798c AL |
740 | |
741 | #ifdef CONFIG_TCP_MD5SIG | |
742 | *md5 = tp->af_specific->md5_lookup(sk, sk); | |
743 | if (unlikely(*md5)) { | |
744 | opts->options |= OPTION_MD5; | |
745 | size += TCPOLEN_MD5SIG_ALIGNED; | |
746 | } | |
747 | #else | |
748 | *md5 = NULL; | |
749 | #endif | |
750 | ||
751 | if (likely(tp->rx_opt.tstamp_ok)) { | |
752 | opts->options |= OPTION_TS; | |
753 | opts->tsval = tcb ? tcb->when : 0; | |
754 | opts->tsecr = tp->rx_opt.ts_recent; | |
755 | size += TCPOLEN_TSTAMP_ALIGNED; | |
756 | } | |
757 | ||
cabeccbd IJ |
758 | eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; |
759 | if (unlikely(eff_sacks)) { | |
33ad798c AL |
760 | const unsigned remaining = MAX_TCP_OPTION_SPACE - size; |
761 | opts->num_sack_blocks = | |
cabeccbd | 762 | min_t(unsigned, eff_sacks, |
33ad798c AL |
763 | (remaining - TCPOLEN_SACK_BASE_ALIGNED) / |
764 | TCPOLEN_SACK_PERBLOCK); | |
765 | size += TCPOLEN_SACK_BASE_ALIGNED + | |
766 | opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; | |
767 | } | |
768 | ||
769 | return size; | |
40efc6fa | 770 | } |
1da177e4 LT |
771 | |
772 | /* This routine actually transmits TCP packets queued in by | |
773 | * tcp_do_sendmsg(). This is used by both the initial | |
774 | * transmission and possible later retransmissions. | |
775 | * All SKB's seen here are completely headerless. It is our | |
776 | * job to build the TCP header, and pass the packet down to | |
777 | * IP so it can do the same plus pass the packet off to the | |
778 | * device. | |
779 | * | |
780 | * We are working here with either a clone of the original | |
781 | * SKB, or a fresh unique copy made by the retransmit engine. | |
782 | */ | |
056834d9 IJ |
783 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, |
784 | gfp_t gfp_mask) | |
1da177e4 | 785 | { |
dfb4b9dc DM |
786 | const struct inet_connection_sock *icsk = inet_csk(sk); |
787 | struct inet_sock *inet; | |
788 | struct tcp_sock *tp; | |
789 | struct tcp_skb_cb *tcb; | |
33ad798c AL |
790 | struct tcp_out_options opts; |
791 | unsigned tcp_options_size, tcp_header_size; | |
cfb6eeb4 | 792 | struct tcp_md5sig_key *md5; |
dfb4b9dc | 793 | struct tcphdr *th; |
dfb4b9dc DM |
794 | int err; |
795 | ||
796 | BUG_ON(!skb || !tcp_skb_pcount(skb)); | |
797 | ||
798 | /* If congestion control is doing timestamping, we must | |
799 | * take such a timestamp before we potentially clone/copy. | |
800 | */ | |
164891aa | 801 | if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) |
dfb4b9dc DM |
802 | __net_timestamp(skb); |
803 | ||
804 | if (likely(clone_it)) { | |
805 | if (unlikely(skb_cloned(skb))) | |
806 | skb = pskb_copy(skb, gfp_mask); | |
807 | else | |
808 | skb = skb_clone(skb, gfp_mask); | |
809 | if (unlikely(!skb)) | |
810 | return -ENOBUFS; | |
811 | } | |
1da177e4 | 812 | |
dfb4b9dc DM |
813 | inet = inet_sk(sk); |
814 | tp = tcp_sk(sk); | |
815 | tcb = TCP_SKB_CB(skb); | |
33ad798c | 816 | memset(&opts, 0, sizeof(opts)); |
1da177e4 | 817 | |
a3433f35 | 818 | if (unlikely(tcb->flags & TCPHDR_SYN)) |
33ad798c AL |
819 | tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); |
820 | else | |
821 | tcp_options_size = tcp_established_options(sk, skb, &opts, | |
822 | &md5); | |
823 | tcp_header_size = tcp_options_size + sizeof(struct tcphdr); | |
e905a9ed | 824 | |
dfb4b9dc DM |
825 | if (tcp_packets_in_flight(tp) == 0) |
826 | tcp_ca_event(sk, CA_EVENT_TX_START); | |
827 | ||
aa8223c7 ACM |
828 | skb_push(skb, tcp_header_size); |
829 | skb_reset_transport_header(skb); | |
e89862f4 | 830 | skb_set_owner_w(skb, sk); |
dfb4b9dc DM |
831 | |
832 | /* Build TCP header and checksum it. */ | |
aa8223c7 | 833 | th = tcp_hdr(skb); |
c720c7e8 ED |
834 | th->source = inet->inet_sport; |
835 | th->dest = inet->inet_dport; | |
dfb4b9dc DM |
836 | th->seq = htonl(tcb->seq); |
837 | th->ack_seq = htonl(tp->rcv_nxt); | |
df7a3b07 | 838 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
dfb4b9dc DM |
839 | tcb->flags); |
840 | ||
a3433f35 | 841 | if (unlikely(tcb->flags & TCPHDR_SYN)) { |
dfb4b9dc DM |
842 | /* RFC1323: The window in SYN & SYN/ACK segments |
843 | * is never scaled. | |
844 | */ | |
600ff0c2 | 845 | th->window = htons(min(tp->rcv_wnd, 65535U)); |
dfb4b9dc DM |
846 | } else { |
847 | th->window = htons(tcp_select_window(sk)); | |
848 | } | |
849 | th->check = 0; | |
850 | th->urg_ptr = 0; | |
1da177e4 | 851 | |
33f5f57e | 852 | /* The urg_mode check is necessary during a below snd_una win probe */ |
7691367d HX |
853 | if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { |
854 | if (before(tp->snd_up, tcb->seq + 0x10000)) { | |
855 | th->urg_ptr = htons(tp->snd_up - tcb->seq); | |
856 | th->urg = 1; | |
857 | } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { | |
0eae88f3 | 858 | th->urg_ptr = htons(0xFFFF); |
7691367d HX |
859 | th->urg = 1; |
860 | } | |
dfb4b9dc | 861 | } |
1da177e4 | 862 | |
bd0388ae | 863 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
a3433f35 | 864 | if (likely((tcb->flags & TCPHDR_SYN) == 0)) |
9e412ba7 | 865 | TCP_ECN_send(sk, skb, tcp_header_size); |
1da177e4 | 866 | |
cfb6eeb4 YH |
867 | #ifdef CONFIG_TCP_MD5SIG |
868 | /* Calculate the MD5 hash, as we have all we need now */ | |
869 | if (md5) { | |
a465419b | 870 | sk_nocaps_add(sk, NETIF_F_GSO_MASK); |
bd0388ae | 871 | tp->af_specific->calc_md5_hash(opts.hash_location, |
49a72dfb | 872 | md5, sk, NULL, skb); |
cfb6eeb4 YH |
873 | } |
874 | #endif | |
875 | ||
bb296246 | 876 | icsk->icsk_af_ops->send_check(sk, skb); |
1da177e4 | 877 | |
a3433f35 | 878 | if (likely(tcb->flags & TCPHDR_ACK)) |
dfb4b9dc | 879 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); |
1da177e4 | 880 | |
dfb4b9dc DM |
881 | if (skb->len != tcp_header_size) |
882 | tcp_event_data_sent(tp, skb, sk); | |
1da177e4 | 883 | |
bd37a088 | 884 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
aa2ea058 TH |
885 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, |
886 | tcp_skb_pcount(skb)); | |
1da177e4 | 887 | |
4e15ed4d | 888 | err = icsk->icsk_af_ops->queue_xmit(skb); |
83de47cd | 889 | if (likely(err <= 0)) |
dfb4b9dc DM |
890 | return err; |
891 | ||
3cfe3baa | 892 | tcp_enter_cwr(sk, 1); |
dfb4b9dc | 893 | |
b9df3cb8 | 894 | return net_xmit_eval(err); |
1da177e4 LT |
895 | } |
896 | ||
67edfef7 | 897 | /* This routine just queues the buffer for sending. |
1da177e4 LT |
898 | * |
899 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, | |
900 | * otherwise socket can stall. | |
901 | */ | |
902 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) | |
903 | { | |
904 | struct tcp_sock *tp = tcp_sk(sk); | |
905 | ||
906 | /* Advance write_seq and place onto the write_queue. */ | |
907 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; | |
908 | skb_header_release(skb); | |
fe067e8a | 909 | tcp_add_write_queue_tail(sk, skb); |
3ab224be HA |
910 | sk->sk_wmem_queued += skb->truesize; |
911 | sk_mem_charge(sk, skb->truesize); | |
1da177e4 LT |
912 | } |
913 | ||
67edfef7 | 914 | /* Initialize TSO segments for a packet. */ |
056834d9 IJ |
915 | static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, |
916 | unsigned int mss_now) | |
f6302d1d | 917 | { |
8e5b9dda HX |
918 | if (skb->len <= mss_now || !sk_can_gso(sk) || |
919 | skb->ip_summed == CHECKSUM_NONE) { | |
f6302d1d DM |
920 | /* Avoid the costly divide in the normal |
921 | * non-TSO case. | |
922 | */ | |
7967168c HX |
923 | skb_shinfo(skb)->gso_segs = 1; |
924 | skb_shinfo(skb)->gso_size = 0; | |
925 | skb_shinfo(skb)->gso_type = 0; | |
f6302d1d | 926 | } else { |
356f89e1 | 927 | skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); |
7967168c | 928 | skb_shinfo(skb)->gso_size = mss_now; |
bcd76111 | 929 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
1da177e4 LT |
930 | } |
931 | } | |
932 | ||
91fed7a1 | 933 | /* When a modification to fackets out becomes necessary, we need to check |
68f8353b | 934 | * skb is counted to fackets_out or not. |
91fed7a1 | 935 | */ |
a47e5a98 | 936 | static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb, |
91fed7a1 IJ |
937 | int decr) |
938 | { | |
a47e5a98 IJ |
939 | struct tcp_sock *tp = tcp_sk(sk); |
940 | ||
dc86967b | 941 | if (!tp->sacked_out || tcp_is_reno(tp)) |
91fed7a1 IJ |
942 | return; |
943 | ||
6859d494 | 944 | if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) |
91fed7a1 | 945 | tp->fackets_out -= decr; |
91fed7a1 IJ |
946 | } |
947 | ||
797108d1 IJ |
948 | /* Pcount in the middle of the write queue got changed, we need to do various |
949 | * tweaks to fix counters | |
950 | */ | |
951 | static void tcp_adjust_pcount(struct sock *sk, struct sk_buff *skb, int decr) | |
952 | { | |
953 | struct tcp_sock *tp = tcp_sk(sk); | |
954 | ||
955 | tp->packets_out -= decr; | |
956 | ||
957 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
958 | tp->sacked_out -= decr; | |
959 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
960 | tp->retrans_out -= decr; | |
961 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) | |
962 | tp->lost_out -= decr; | |
963 | ||
964 | /* Reno case is special. Sigh... */ | |
965 | if (tcp_is_reno(tp) && decr > 0) | |
966 | tp->sacked_out -= min_t(u32, tp->sacked_out, decr); | |
967 | ||
968 | tcp_adjust_fackets_out(sk, skb, decr); | |
969 | ||
970 | if (tp->lost_skb_hint && | |
971 | before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && | |
52cf3cc8 | 972 | (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) |
797108d1 IJ |
973 | tp->lost_cnt_hint -= decr; |
974 | ||
975 | tcp_verify_left_out(tp); | |
976 | } | |
977 | ||
1da177e4 LT |
978 | /* Function to create two new TCP segments. Shrinks the given segment |
979 | * to the specified size and appends a new segment with the rest of the | |
e905a9ed | 980 | * packet to the list. This won't be called frequently, I hope. |
1da177e4 LT |
981 | * Remember, these are still headerless SKBs at this point. |
982 | */ | |
056834d9 IJ |
983 | int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, |
984 | unsigned int mss_now) | |
1da177e4 LT |
985 | { |
986 | struct tcp_sock *tp = tcp_sk(sk); | |
987 | struct sk_buff *buff; | |
6475be16 | 988 | int nsize, old_factor; |
b60b49ea | 989 | int nlen; |
9ce01461 | 990 | u8 flags; |
1da177e4 | 991 | |
b2cc99f0 | 992 | BUG_ON(len > skb->len); |
6a438bbe | 993 | |
1da177e4 LT |
994 | nsize = skb_headlen(skb) - len; |
995 | if (nsize < 0) | |
996 | nsize = 0; | |
997 | ||
998 | if (skb_cloned(skb) && | |
999 | skb_is_nonlinear(skb) && | |
1000 | pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) | |
1001 | return -ENOMEM; | |
1002 | ||
1003 | /* Get a new skb... force flag on. */ | |
1004 | buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); | |
1005 | if (buff == NULL) | |
1006 | return -ENOMEM; /* We'll just try again later. */ | |
ef5cb973 | 1007 | |
3ab224be HA |
1008 | sk->sk_wmem_queued += buff->truesize; |
1009 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea HX |
1010 | nlen = skb->len - len - nsize; |
1011 | buff->truesize += nlen; | |
1012 | skb->truesize -= nlen; | |
1da177e4 LT |
1013 | |
1014 | /* Correct the sequence numbers. */ | |
1015 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1016 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1017 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1018 | ||
1019 | /* PSH and FIN should only be set in the second packet. */ | |
1020 | flags = TCP_SKB_CB(skb)->flags; | |
a3433f35 | 1021 | TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); |
1da177e4 | 1022 | TCP_SKB_CB(buff)->flags = flags; |
e14c3caf | 1023 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
1da177e4 | 1024 | |
84fa7933 | 1025 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { |
1da177e4 | 1026 | /* Copy and checksum data tail into the new buffer. */ |
056834d9 IJ |
1027 | buff->csum = csum_partial_copy_nocheck(skb->data + len, |
1028 | skb_put(buff, nsize), | |
1da177e4 LT |
1029 | nsize, 0); |
1030 | ||
1031 | skb_trim(skb, len); | |
1032 | ||
1033 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); | |
1034 | } else { | |
84fa7933 | 1035 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
1036 | skb_split(skb, buff, len); |
1037 | } | |
1038 | ||
1039 | buff->ip_summed = skb->ip_summed; | |
1040 | ||
1041 | /* Looks stupid, but our code really uses when of | |
1042 | * skbs, which it never sent before. --ANK | |
1043 | */ | |
1044 | TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; | |
a61bbcf2 | 1045 | buff->tstamp = skb->tstamp; |
1da177e4 | 1046 | |
6475be16 DM |
1047 | old_factor = tcp_skb_pcount(skb); |
1048 | ||
1da177e4 | 1049 | /* Fix up tso_factor for both original and new SKB. */ |
846998ae DM |
1050 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
1051 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
1da177e4 | 1052 | |
6475be16 DM |
1053 | /* If this packet has been sent out already, we must |
1054 | * adjust the various packet counters. | |
1055 | */ | |
cf0b450c | 1056 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
6475be16 DM |
1057 | int diff = old_factor - tcp_skb_pcount(skb) - |
1058 | tcp_skb_pcount(buff); | |
1da177e4 | 1059 | |
797108d1 IJ |
1060 | if (diff) |
1061 | tcp_adjust_pcount(sk, skb, diff); | |
1da177e4 LT |
1062 | } |
1063 | ||
1064 | /* Link BUFF into the send queue. */ | |
f44b5271 | 1065 | skb_header_release(buff); |
fe067e8a | 1066 | tcp_insert_write_queue_after(skb, buff, sk); |
1da177e4 LT |
1067 | |
1068 | return 0; | |
1069 | } | |
1070 | ||
1071 | /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c | |
1072 | * eventually). The difference is that pulled data not copied, but | |
1073 | * immediately discarded. | |
1074 | */ | |
f2911969 | 1075 | static void __pskb_trim_head(struct sk_buff *skb, int len) |
1da177e4 LT |
1076 | { |
1077 | int i, k, eat; | |
1078 | ||
1079 | eat = len; | |
1080 | k = 0; | |
056834d9 | 1081 | for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { |
1da177e4 LT |
1082 | if (skb_shinfo(skb)->frags[i].size <= eat) { |
1083 | put_page(skb_shinfo(skb)->frags[i].page); | |
1084 | eat -= skb_shinfo(skb)->frags[i].size; | |
1085 | } else { | |
1086 | skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; | |
1087 | if (eat) { | |
1088 | skb_shinfo(skb)->frags[k].page_offset += eat; | |
1089 | skb_shinfo(skb)->frags[k].size -= eat; | |
1090 | eat = 0; | |
1091 | } | |
1092 | k++; | |
1093 | } | |
1094 | } | |
1095 | skb_shinfo(skb)->nr_frags = k; | |
1096 | ||
27a884dc | 1097 | skb_reset_tail_pointer(skb); |
1da177e4 LT |
1098 | skb->data_len -= len; |
1099 | skb->len = skb->data_len; | |
1da177e4 LT |
1100 | } |
1101 | ||
67edfef7 | 1102 | /* Remove acked data from a packet in the transmit queue. */ |
1da177e4 LT |
1103 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) |
1104 | { | |
056834d9 | 1105 | if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) |
1da177e4 LT |
1106 | return -ENOMEM; |
1107 | ||
f2911969 HX |
1108 | /* If len == headlen, we avoid __skb_pull to preserve alignment. */ |
1109 | if (unlikely(len < skb_headlen(skb))) | |
1da177e4 | 1110 | __skb_pull(skb, len); |
f2911969 HX |
1111 | else |
1112 | __pskb_trim_head(skb, len - skb_headlen(skb)); | |
1da177e4 LT |
1113 | |
1114 | TCP_SKB_CB(skb)->seq += len; | |
84fa7933 | 1115 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
1116 | |
1117 | skb->truesize -= len; | |
1118 | sk->sk_wmem_queued -= len; | |
3ab224be | 1119 | sk_mem_uncharge(sk, len); |
1da177e4 LT |
1120 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); |
1121 | ||
1122 | /* Any change of skb->len requires recalculation of tso | |
1123 | * factor and mss. | |
1124 | */ | |
1125 | if (tcp_skb_pcount(skb) > 1) | |
0c54b85f | 1126 | tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk)); |
1da177e4 LT |
1127 | |
1128 | return 0; | |
1129 | } | |
1130 | ||
67edfef7 | 1131 | /* Calculate MSS. Not accounting for SACKs here. */ |
5d424d5a JH |
1132 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) |
1133 | { | |
1134 | struct tcp_sock *tp = tcp_sk(sk); | |
1135 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1136 | int mss_now; | |
1137 | ||
1138 | /* Calculate base mss without TCP options: | |
1139 | It is MMS_S - sizeof(tcphdr) of rfc1122 | |
1140 | */ | |
1141 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); | |
1142 | ||
1143 | /* Clamp it (mss_clamp does not include tcp options) */ | |
1144 | if (mss_now > tp->rx_opt.mss_clamp) | |
1145 | mss_now = tp->rx_opt.mss_clamp; | |
1146 | ||
1147 | /* Now subtract optional transport overhead */ | |
1148 | mss_now -= icsk->icsk_ext_hdr_len; | |
1149 | ||
1150 | /* Then reserve room for full set of TCP options and 8 bytes of data */ | |
1151 | if (mss_now < 48) | |
1152 | mss_now = 48; | |
1153 | ||
1154 | /* Now subtract TCP options size, not including SACKs */ | |
1155 | mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); | |
1156 | ||
1157 | return mss_now; | |
1158 | } | |
1159 | ||
1160 | /* Inverse of above */ | |
1161 | int tcp_mss_to_mtu(struct sock *sk, int mss) | |
1162 | { | |
1163 | struct tcp_sock *tp = tcp_sk(sk); | |
1164 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1165 | int mtu; | |
1166 | ||
1167 | mtu = mss + | |
1168 | tp->tcp_header_len + | |
1169 | icsk->icsk_ext_hdr_len + | |
1170 | icsk->icsk_af_ops->net_header_len; | |
1171 | ||
1172 | return mtu; | |
1173 | } | |
1174 | ||
67edfef7 | 1175 | /* MTU probing init per socket */ |
5d424d5a JH |
1176 | void tcp_mtup_init(struct sock *sk) |
1177 | { | |
1178 | struct tcp_sock *tp = tcp_sk(sk); | |
1179 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1180 | ||
1181 | icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; | |
1182 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + | |
e905a9ed | 1183 | icsk->icsk_af_ops->net_header_len; |
5d424d5a JH |
1184 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); |
1185 | icsk->icsk_mtup.probe_size = 0; | |
1186 | } | |
4bc2f18b | 1187 | EXPORT_SYMBOL(tcp_mtup_init); |
5d424d5a | 1188 | |
1da177e4 LT |
1189 | /* This function synchronize snd mss to current pmtu/exthdr set. |
1190 | ||
1191 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts | |
1192 | for TCP options, but includes only bare TCP header. | |
1193 | ||
1194 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. | |
caa20d9a | 1195 | It is minimum of user_mss and mss received with SYN. |
1da177e4 LT |
1196 | It also does not include TCP options. |
1197 | ||
d83d8461 | 1198 | inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. |
1da177e4 LT |
1199 | |
1200 | tp->mss_cache is current effective sending mss, including | |
1201 | all tcp options except for SACKs. It is evaluated, | |
1202 | taking into account current pmtu, but never exceeds | |
1203 | tp->rx_opt.mss_clamp. | |
1204 | ||
1205 | NOTE1. rfc1122 clearly states that advertised MSS | |
1206 | DOES NOT include either tcp or ip options. | |
1207 | ||
d83d8461 ACM |
1208 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
1209 | are READ ONLY outside this function. --ANK (980731) | |
1da177e4 | 1210 | */ |
1da177e4 LT |
1211 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) |
1212 | { | |
1213 | struct tcp_sock *tp = tcp_sk(sk); | |
d83d8461 | 1214 | struct inet_connection_sock *icsk = inet_csk(sk); |
5d424d5a | 1215 | int mss_now; |
1da177e4 | 1216 | |
5d424d5a JH |
1217 | if (icsk->icsk_mtup.search_high > pmtu) |
1218 | icsk->icsk_mtup.search_high = pmtu; | |
1da177e4 | 1219 | |
5d424d5a | 1220 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
409d22b4 | 1221 | mss_now = tcp_bound_to_half_wnd(tp, mss_now); |
1da177e4 LT |
1222 | |
1223 | /* And store cached results */ | |
d83d8461 | 1224 | icsk->icsk_pmtu_cookie = pmtu; |
5d424d5a JH |
1225 | if (icsk->icsk_mtup.enabled) |
1226 | mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); | |
c1b4a7e6 | 1227 | tp->mss_cache = mss_now; |
1da177e4 LT |
1228 | |
1229 | return mss_now; | |
1230 | } | |
4bc2f18b | 1231 | EXPORT_SYMBOL(tcp_sync_mss); |
1da177e4 LT |
1232 | |
1233 | /* Compute the current effective MSS, taking SACKs and IP options, | |
1234 | * and even PMTU discovery events into account. | |
1da177e4 | 1235 | */ |
0c54b85f | 1236 | unsigned int tcp_current_mss(struct sock *sk) |
1da177e4 LT |
1237 | { |
1238 | struct tcp_sock *tp = tcp_sk(sk); | |
1239 | struct dst_entry *dst = __sk_dst_get(sk); | |
c1b4a7e6 | 1240 | u32 mss_now; |
33ad798c AL |
1241 | unsigned header_len; |
1242 | struct tcp_out_options opts; | |
1243 | struct tcp_md5sig_key *md5; | |
c1b4a7e6 DM |
1244 | |
1245 | mss_now = tp->mss_cache; | |
1246 | ||
1da177e4 LT |
1247 | if (dst) { |
1248 | u32 mtu = dst_mtu(dst); | |
d83d8461 | 1249 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
1da177e4 LT |
1250 | mss_now = tcp_sync_mss(sk, mtu); |
1251 | } | |
1252 | ||
33ad798c AL |
1253 | header_len = tcp_established_options(sk, NULL, &opts, &md5) + |
1254 | sizeof(struct tcphdr); | |
1255 | /* The mss_cache is sized based on tp->tcp_header_len, which assumes | |
1256 | * some common options. If this is an odd packet (because we have SACK | |
1257 | * blocks etc) then our calculated header_len will be different, and | |
1258 | * we have to adjust mss_now correspondingly */ | |
1259 | if (header_len != tp->tcp_header_len) { | |
1260 | int delta = (int) header_len - tp->tcp_header_len; | |
1261 | mss_now -= delta; | |
1262 | } | |
cfb6eeb4 | 1263 | |
1da177e4 LT |
1264 | return mss_now; |
1265 | } | |
1266 | ||
a762a980 | 1267 | /* Congestion window validation. (RFC2861) */ |
9e412ba7 | 1268 | static void tcp_cwnd_validate(struct sock *sk) |
a762a980 | 1269 | { |
9e412ba7 | 1270 | struct tcp_sock *tp = tcp_sk(sk); |
a762a980 | 1271 | |
d436d686 | 1272 | if (tp->packets_out >= tp->snd_cwnd) { |
a762a980 DM |
1273 | /* Network is feed fully. */ |
1274 | tp->snd_cwnd_used = 0; | |
1275 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1276 | } else { | |
1277 | /* Network starves. */ | |
1278 | if (tp->packets_out > tp->snd_cwnd_used) | |
1279 | tp->snd_cwnd_used = tp->packets_out; | |
1280 | ||
15d33c07 DM |
1281 | if (sysctl_tcp_slow_start_after_idle && |
1282 | (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) | |
a762a980 DM |
1283 | tcp_cwnd_application_limited(sk); |
1284 | } | |
1285 | } | |
1286 | ||
0e3a4803 IJ |
1287 | /* Returns the portion of skb which can be sent right away without |
1288 | * introducing MSS oddities to segment boundaries. In rare cases where | |
1289 | * mss_now != mss_cache, we will request caller to create a small skb | |
1290 | * per input skb which could be mostly avoided here (if desired). | |
5ea3a748 IJ |
1291 | * |
1292 | * We explicitly want to create a request for splitting write queue tail | |
1293 | * to a small skb for Nagle purposes while avoiding unnecessary modulos, | |
1294 | * thus all the complexity (cwnd_len is always MSS multiple which we | |
1295 | * return whenever allowed by the other factors). Basically we need the | |
1296 | * modulo only when the receiver window alone is the limiting factor or | |
1297 | * when we would be allowed to send the split-due-to-Nagle skb fully. | |
0e3a4803 IJ |
1298 | */ |
1299 | static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb, | |
056834d9 | 1300 | unsigned int mss_now, unsigned int cwnd) |
c1b4a7e6 | 1301 | { |
0e3a4803 IJ |
1302 | struct tcp_sock *tp = tcp_sk(sk); |
1303 | u32 needed, window, cwnd_len; | |
c1b4a7e6 | 1304 | |
90840def | 1305 | window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
c1b4a7e6 | 1306 | cwnd_len = mss_now * cwnd; |
0e3a4803 IJ |
1307 | |
1308 | if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk))) | |
1309 | return cwnd_len; | |
1310 | ||
5ea3a748 IJ |
1311 | needed = min(skb->len, window); |
1312 | ||
17515408 | 1313 | if (cwnd_len <= needed) |
0e3a4803 IJ |
1314 | return cwnd_len; |
1315 | ||
0e3a4803 | 1316 | return needed - needed % mss_now; |
c1b4a7e6 DM |
1317 | } |
1318 | ||
1319 | /* Can at least one segment of SKB be sent right now, according to the | |
1320 | * congestion window rules? If so, return how many segments are allowed. | |
1321 | */ | |
056834d9 IJ |
1322 | static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, |
1323 | struct sk_buff *skb) | |
c1b4a7e6 DM |
1324 | { |
1325 | u32 in_flight, cwnd; | |
1326 | ||
1327 | /* Don't be strict about the congestion window for the final FIN. */ | |
a3433f35 | 1328 | if ((TCP_SKB_CB(skb)->flags & TCPHDR_FIN) && tcp_skb_pcount(skb) == 1) |
c1b4a7e6 DM |
1329 | return 1; |
1330 | ||
1331 | in_flight = tcp_packets_in_flight(tp); | |
1332 | cwnd = tp->snd_cwnd; | |
1333 | if (in_flight < cwnd) | |
1334 | return (cwnd - in_flight); | |
1335 | ||
1336 | return 0; | |
1337 | } | |
1338 | ||
67edfef7 AK |
1339 | /* Intialize TSO state of a skb. |
1340 | * This must be invoked the first time we consider transmitting | |
c1b4a7e6 DM |
1341 | * SKB onto the wire. |
1342 | */ | |
056834d9 IJ |
1343 | static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, |
1344 | unsigned int mss_now) | |
c1b4a7e6 DM |
1345 | { |
1346 | int tso_segs = tcp_skb_pcount(skb); | |
1347 | ||
f8269a49 | 1348 | if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { |
846998ae | 1349 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
c1b4a7e6 DM |
1350 | tso_segs = tcp_skb_pcount(skb); |
1351 | } | |
1352 | return tso_segs; | |
1353 | } | |
1354 | ||
67edfef7 | 1355 | /* Minshall's variant of the Nagle send check. */ |
c1b4a7e6 DM |
1356 | static inline int tcp_minshall_check(const struct tcp_sock *tp) |
1357 | { | |
09cb105e | 1358 | return after(tp->snd_sml, tp->snd_una) && |
c1b4a7e6 DM |
1359 | !after(tp->snd_sml, tp->snd_nxt); |
1360 | } | |
1361 | ||
1362 | /* Return 0, if packet can be sent now without violation Nagle's rules: | |
1363 | * 1. It is full sized. | |
1364 | * 2. Or it contains FIN. (already checked by caller) | |
1365 | * 3. Or TCP_NODELAY was set. | |
1366 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. | |
1367 | * With Minshall's modification: all sent small packets are ACKed. | |
1368 | */ | |
c1b4a7e6 | 1369 | static inline int tcp_nagle_check(const struct tcp_sock *tp, |
e905a9ed | 1370 | const struct sk_buff *skb, |
c1b4a7e6 DM |
1371 | unsigned mss_now, int nonagle) |
1372 | { | |
a02cec21 | 1373 | return skb->len < mss_now && |
056834d9 | 1374 | ((nonagle & TCP_NAGLE_CORK) || |
a02cec21 | 1375 | (!nonagle && tp->packets_out && tcp_minshall_check(tp))); |
c1b4a7e6 DM |
1376 | } |
1377 | ||
1378 | /* Return non-zero if the Nagle test allows this packet to be | |
1379 | * sent now. | |
1380 | */ | |
1381 | static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, | |
1382 | unsigned int cur_mss, int nonagle) | |
1383 | { | |
1384 | /* Nagle rule does not apply to frames, which sit in the middle of the | |
1385 | * write_queue (they have no chances to get new data). | |
1386 | * | |
1387 | * This is implemented in the callers, where they modify the 'nonagle' | |
1388 | * argument based upon the location of SKB in the send queue. | |
1389 | */ | |
1390 | if (nonagle & TCP_NAGLE_PUSH) | |
1391 | return 1; | |
1392 | ||
d551e454 IJ |
1393 | /* Don't use the nagle rule for urgent data (or for the final FIN). |
1394 | * Nagle can be ignored during F-RTO too (see RFC4138). | |
1395 | */ | |
33f5f57e | 1396 | if (tcp_urg_mode(tp) || (tp->frto_counter == 2) || |
a3433f35 | 1397 | (TCP_SKB_CB(skb)->flags & TCPHDR_FIN)) |
c1b4a7e6 DM |
1398 | return 1; |
1399 | ||
1400 | if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) | |
1401 | return 1; | |
1402 | ||
1403 | return 0; | |
1404 | } | |
1405 | ||
1406 | /* Does at least the first segment of SKB fit into the send window? */ | |
056834d9 IJ |
1407 | static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, |
1408 | unsigned int cur_mss) | |
c1b4a7e6 DM |
1409 | { |
1410 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
1411 | ||
1412 | if (skb->len > cur_mss) | |
1413 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; | |
1414 | ||
90840def | 1415 | return !after(end_seq, tcp_wnd_end(tp)); |
c1b4a7e6 DM |
1416 | } |
1417 | ||
fe067e8a | 1418 | /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) |
c1b4a7e6 DM |
1419 | * should be put on the wire right now. If so, it returns the number of |
1420 | * packets allowed by the congestion window. | |
1421 | */ | |
1422 | static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, | |
1423 | unsigned int cur_mss, int nonagle) | |
1424 | { | |
1425 | struct tcp_sock *tp = tcp_sk(sk); | |
1426 | unsigned int cwnd_quota; | |
1427 | ||
846998ae | 1428 | tcp_init_tso_segs(sk, skb, cur_mss); |
c1b4a7e6 DM |
1429 | |
1430 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) | |
1431 | return 0; | |
1432 | ||
1433 | cwnd_quota = tcp_cwnd_test(tp, skb); | |
056834d9 | 1434 | if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) |
c1b4a7e6 DM |
1435 | cwnd_quota = 0; |
1436 | ||
1437 | return cwnd_quota; | |
1438 | } | |
1439 | ||
67edfef7 | 1440 | /* Test if sending is allowed right now. */ |
9e412ba7 | 1441 | int tcp_may_send_now(struct sock *sk) |
c1b4a7e6 | 1442 | { |
9e412ba7 | 1443 | struct tcp_sock *tp = tcp_sk(sk); |
fe067e8a | 1444 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 | 1445 | |
a02cec21 | 1446 | return skb && |
0c54b85f | 1447 | tcp_snd_test(sk, skb, tcp_current_mss(sk), |
c1b4a7e6 | 1448 | (tcp_skb_is_last(sk, skb) ? |
a02cec21 | 1449 | tp->nonagle : TCP_NAGLE_PUSH)); |
c1b4a7e6 DM |
1450 | } |
1451 | ||
1452 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet | |
1453 | * which is put after SKB on the list. It is very much like | |
1454 | * tcp_fragment() except that it may make several kinds of assumptions | |
1455 | * in order to speed up the splitting operation. In particular, we | |
1456 | * know that all the data is in scatter-gather pages, and that the | |
1457 | * packet has never been sent out before (and thus is not cloned). | |
1458 | */ | |
056834d9 | 1459 | static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, |
c4ead4c5 | 1460 | unsigned int mss_now, gfp_t gfp) |
c1b4a7e6 DM |
1461 | { |
1462 | struct sk_buff *buff; | |
1463 | int nlen = skb->len - len; | |
9ce01461 | 1464 | u8 flags; |
c1b4a7e6 DM |
1465 | |
1466 | /* All of a TSO frame must be composed of paged data. */ | |
c8ac3774 HX |
1467 | if (skb->len != skb->data_len) |
1468 | return tcp_fragment(sk, skb, len, mss_now); | |
c1b4a7e6 | 1469 | |
c4ead4c5 | 1470 | buff = sk_stream_alloc_skb(sk, 0, gfp); |
c1b4a7e6 DM |
1471 | if (unlikely(buff == NULL)) |
1472 | return -ENOMEM; | |
1473 | ||
3ab224be HA |
1474 | sk->sk_wmem_queued += buff->truesize; |
1475 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea | 1476 | buff->truesize += nlen; |
c1b4a7e6 DM |
1477 | skb->truesize -= nlen; |
1478 | ||
1479 | /* Correct the sequence numbers. */ | |
1480 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1481 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1482 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1483 | ||
1484 | /* PSH and FIN should only be set in the second packet. */ | |
1485 | flags = TCP_SKB_CB(skb)->flags; | |
a3433f35 | 1486 | TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); |
c1b4a7e6 DM |
1487 | TCP_SKB_CB(buff)->flags = flags; |
1488 | ||
1489 | /* This packet was never sent out yet, so no SACK bits. */ | |
1490 | TCP_SKB_CB(buff)->sacked = 0; | |
1491 | ||
84fa7933 | 1492 | buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; |
c1b4a7e6 DM |
1493 | skb_split(skb, buff, len); |
1494 | ||
1495 | /* Fix up tso_factor for both original and new SKB. */ | |
846998ae DM |
1496 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
1497 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
c1b4a7e6 DM |
1498 | |
1499 | /* Link BUFF into the send queue. */ | |
1500 | skb_header_release(buff); | |
fe067e8a | 1501 | tcp_insert_write_queue_after(skb, buff, sk); |
c1b4a7e6 DM |
1502 | |
1503 | return 0; | |
1504 | } | |
1505 | ||
1506 | /* Try to defer sending, if possible, in order to minimize the amount | |
1507 | * of TSO splitting we do. View it as a kind of TSO Nagle test. | |
1508 | * | |
1509 | * This algorithm is from John Heffner. | |
1510 | */ | |
9e412ba7 | 1511 | static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) |
c1b4a7e6 | 1512 | { |
9e412ba7 | 1513 | struct tcp_sock *tp = tcp_sk(sk); |
6687e988 | 1514 | const struct inet_connection_sock *icsk = inet_csk(sk); |
c1b4a7e6 DM |
1515 | u32 send_win, cong_win, limit, in_flight; |
1516 | ||
a3433f35 | 1517 | if (TCP_SKB_CB(skb)->flags & TCPHDR_FIN) |
ae8064ac | 1518 | goto send_now; |
c1b4a7e6 | 1519 | |
6687e988 | 1520 | if (icsk->icsk_ca_state != TCP_CA_Open) |
ae8064ac JH |
1521 | goto send_now; |
1522 | ||
1523 | /* Defer for less than two clock ticks. */ | |
bd515c3e | 1524 | if (tp->tso_deferred && |
a2acde07 | 1525 | (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) |
ae8064ac | 1526 | goto send_now; |
908a75c1 | 1527 | |
c1b4a7e6 DM |
1528 | in_flight = tcp_packets_in_flight(tp); |
1529 | ||
056834d9 | 1530 | BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); |
c1b4a7e6 | 1531 | |
90840def | 1532 | send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
c1b4a7e6 DM |
1533 | |
1534 | /* From in_flight test above, we know that cwnd > in_flight. */ | |
1535 | cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; | |
1536 | ||
1537 | limit = min(send_win, cong_win); | |
1538 | ||
ba244fe9 | 1539 | /* If a full-sized TSO skb can be sent, do it. */ |
82cc1a7a | 1540 | if (limit >= sk->sk_gso_max_size) |
ae8064ac | 1541 | goto send_now; |
ba244fe9 | 1542 | |
62ad2761 IJ |
1543 | /* Middle in queue won't get any more data, full sendable already? */ |
1544 | if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) | |
1545 | goto send_now; | |
1546 | ||
c1b4a7e6 DM |
1547 | if (sysctl_tcp_tso_win_divisor) { |
1548 | u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); | |
1549 | ||
1550 | /* If at least some fraction of a window is available, | |
1551 | * just use it. | |
1552 | */ | |
1553 | chunk /= sysctl_tcp_tso_win_divisor; | |
1554 | if (limit >= chunk) | |
ae8064ac | 1555 | goto send_now; |
c1b4a7e6 DM |
1556 | } else { |
1557 | /* Different approach, try not to defer past a single | |
1558 | * ACK. Receiver should ACK every other full sized | |
1559 | * frame, so if we have space for more than 3 frames | |
1560 | * then send now. | |
1561 | */ | |
1562 | if (limit > tcp_max_burst(tp) * tp->mss_cache) | |
ae8064ac | 1563 | goto send_now; |
c1b4a7e6 DM |
1564 | } |
1565 | ||
1566 | /* Ok, it looks like it is advisable to defer. */ | |
056834d9 | 1567 | tp->tso_deferred = 1 | (jiffies << 1); |
ae8064ac | 1568 | |
c1b4a7e6 | 1569 | return 1; |
ae8064ac JH |
1570 | |
1571 | send_now: | |
1572 | tp->tso_deferred = 0; | |
1573 | return 0; | |
c1b4a7e6 DM |
1574 | } |
1575 | ||
5d424d5a | 1576 | /* Create a new MTU probe if we are ready. |
67edfef7 AK |
1577 | * MTU probe is regularly attempting to increase the path MTU by |
1578 | * deliberately sending larger packets. This discovers routing | |
1579 | * changes resulting in larger path MTUs. | |
1580 | * | |
5d424d5a JH |
1581 | * Returns 0 if we should wait to probe (no cwnd available), |
1582 | * 1 if a probe was sent, | |
056834d9 IJ |
1583 | * -1 otherwise |
1584 | */ | |
5d424d5a JH |
1585 | static int tcp_mtu_probe(struct sock *sk) |
1586 | { | |
1587 | struct tcp_sock *tp = tcp_sk(sk); | |
1588 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1589 | struct sk_buff *skb, *nskb, *next; | |
1590 | int len; | |
1591 | int probe_size; | |
91cc17c0 | 1592 | int size_needed; |
5d424d5a JH |
1593 | int copy; |
1594 | int mss_now; | |
1595 | ||
1596 | /* Not currently probing/verifying, | |
1597 | * not in recovery, | |
1598 | * have enough cwnd, and | |
1599 | * not SACKing (the variable headers throw things off) */ | |
1600 | if (!icsk->icsk_mtup.enabled || | |
1601 | icsk->icsk_mtup.probe_size || | |
1602 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || | |
1603 | tp->snd_cwnd < 11 || | |
cabeccbd | 1604 | tp->rx_opt.num_sacks || tp->rx_opt.dsack) |
5d424d5a JH |
1605 | return -1; |
1606 | ||
1607 | /* Very simple search strategy: just double the MSS. */ | |
0c54b85f | 1608 | mss_now = tcp_current_mss(sk); |
056834d9 | 1609 | probe_size = 2 * tp->mss_cache; |
91cc17c0 | 1610 | size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; |
5d424d5a JH |
1611 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { |
1612 | /* TODO: set timer for probe_converge_event */ | |
1613 | return -1; | |
1614 | } | |
1615 | ||
1616 | /* Have enough data in the send queue to probe? */ | |
7f9c33e5 | 1617 | if (tp->write_seq - tp->snd_nxt < size_needed) |
5d424d5a JH |
1618 | return -1; |
1619 | ||
91cc17c0 IJ |
1620 | if (tp->snd_wnd < size_needed) |
1621 | return -1; | |
90840def | 1622 | if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) |
91cc17c0 | 1623 | return 0; |
5d424d5a | 1624 | |
d67c58e9 IJ |
1625 | /* Do we need to wait to drain cwnd? With none in flight, don't stall */ |
1626 | if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { | |
1627 | if (!tcp_packets_in_flight(tp)) | |
5d424d5a JH |
1628 | return -1; |
1629 | else | |
1630 | return 0; | |
1631 | } | |
1632 | ||
1633 | /* We're allowed to probe. Build it now. */ | |
1634 | if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) | |
1635 | return -1; | |
3ab224be HA |
1636 | sk->sk_wmem_queued += nskb->truesize; |
1637 | sk_mem_charge(sk, nskb->truesize); | |
5d424d5a | 1638 | |
fe067e8a | 1639 | skb = tcp_send_head(sk); |
5d424d5a JH |
1640 | |
1641 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; | |
1642 | TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; | |
a3433f35 | 1643 | TCP_SKB_CB(nskb)->flags = TCPHDR_ACK; |
5d424d5a JH |
1644 | TCP_SKB_CB(nskb)->sacked = 0; |
1645 | nskb->csum = 0; | |
84fa7933 | 1646 | nskb->ip_summed = skb->ip_summed; |
5d424d5a | 1647 | |
50c4817e IJ |
1648 | tcp_insert_write_queue_before(nskb, skb, sk); |
1649 | ||
5d424d5a | 1650 | len = 0; |
234b6860 | 1651 | tcp_for_write_queue_from_safe(skb, next, sk) { |
5d424d5a JH |
1652 | copy = min_t(int, skb->len, probe_size - len); |
1653 | if (nskb->ip_summed) | |
1654 | skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); | |
1655 | else | |
1656 | nskb->csum = skb_copy_and_csum_bits(skb, 0, | |
056834d9 IJ |
1657 | skb_put(nskb, copy), |
1658 | copy, nskb->csum); | |
5d424d5a JH |
1659 | |
1660 | if (skb->len <= copy) { | |
1661 | /* We've eaten all the data from this skb. | |
1662 | * Throw it away. */ | |
1663 | TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags; | |
fe067e8a | 1664 | tcp_unlink_write_queue(skb, sk); |
3ab224be | 1665 | sk_wmem_free_skb(sk, skb); |
5d424d5a JH |
1666 | } else { |
1667 | TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags & | |
a3433f35 | 1668 | ~(TCPHDR_FIN|TCPHDR_PSH); |
5d424d5a JH |
1669 | if (!skb_shinfo(skb)->nr_frags) { |
1670 | skb_pull(skb, copy); | |
84fa7933 | 1671 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
056834d9 IJ |
1672 | skb->csum = csum_partial(skb->data, |
1673 | skb->len, 0); | |
5d424d5a JH |
1674 | } else { |
1675 | __pskb_trim_head(skb, copy); | |
1676 | tcp_set_skb_tso_segs(sk, skb, mss_now); | |
1677 | } | |
1678 | TCP_SKB_CB(skb)->seq += copy; | |
1679 | } | |
1680 | ||
1681 | len += copy; | |
234b6860 IJ |
1682 | |
1683 | if (len >= probe_size) | |
1684 | break; | |
5d424d5a JH |
1685 | } |
1686 | tcp_init_tso_segs(sk, nskb, nskb->len); | |
1687 | ||
1688 | /* We're ready to send. If this fails, the probe will | |
1689 | * be resegmented into mss-sized pieces by tcp_write_xmit(). */ | |
1690 | TCP_SKB_CB(nskb)->when = tcp_time_stamp; | |
1691 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { | |
1692 | /* Decrement cwnd here because we are sending | |
056834d9 | 1693 | * effectively two packets. */ |
5d424d5a | 1694 | tp->snd_cwnd--; |
66f5fe62 | 1695 | tcp_event_new_data_sent(sk, nskb); |
5d424d5a JH |
1696 | |
1697 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); | |
0e7b1368 JH |
1698 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; |
1699 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; | |
5d424d5a JH |
1700 | |
1701 | return 1; | |
1702 | } | |
1703 | ||
1704 | return -1; | |
1705 | } | |
1706 | ||
1da177e4 LT |
1707 | /* This routine writes packets to the network. It advances the |
1708 | * send_head. This happens as incoming acks open up the remote | |
1709 | * window for us. | |
1710 | * | |
f8269a49 IJ |
1711 | * LARGESEND note: !tcp_urg_mode is overkill, only frames between |
1712 | * snd_up-64k-mss .. snd_up cannot be large. However, taking into | |
1713 | * account rare use of URG, this is not a big flaw. | |
1714 | * | |
1da177e4 LT |
1715 | * Returns 1, if no segments are in flight and we have queued segments, but |
1716 | * cannot send anything now because of SWS or another problem. | |
1717 | */ | |
d5dd9175 IJ |
1718 | static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
1719 | int push_one, gfp_t gfp) | |
1da177e4 LT |
1720 | { |
1721 | struct tcp_sock *tp = tcp_sk(sk); | |
92df7b51 | 1722 | struct sk_buff *skb; |
c1b4a7e6 DM |
1723 | unsigned int tso_segs, sent_pkts; |
1724 | int cwnd_quota; | |
5d424d5a | 1725 | int result; |
1da177e4 | 1726 | |
92df7b51 | 1727 | sent_pkts = 0; |
5d424d5a | 1728 | |
d5dd9175 IJ |
1729 | if (!push_one) { |
1730 | /* Do MTU probing. */ | |
1731 | result = tcp_mtu_probe(sk); | |
1732 | if (!result) { | |
1733 | return 0; | |
1734 | } else if (result > 0) { | |
1735 | sent_pkts = 1; | |
1736 | } | |
5d424d5a JH |
1737 | } |
1738 | ||
fe067e8a | 1739 | while ((skb = tcp_send_head(sk))) { |
c8ac3774 HX |
1740 | unsigned int limit; |
1741 | ||
b68e9f85 | 1742 | tso_segs = tcp_init_tso_segs(sk, skb, mss_now); |
c1b4a7e6 | 1743 | BUG_ON(!tso_segs); |
aa93466b | 1744 | |
b68e9f85 HX |
1745 | cwnd_quota = tcp_cwnd_test(tp, skb); |
1746 | if (!cwnd_quota) | |
1747 | break; | |
1748 | ||
1749 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) | |
1750 | break; | |
1751 | ||
c1b4a7e6 DM |
1752 | if (tso_segs == 1) { |
1753 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, | |
1754 | (tcp_skb_is_last(sk, skb) ? | |
1755 | nonagle : TCP_NAGLE_PUSH)))) | |
1756 | break; | |
1757 | } else { | |
d5dd9175 | 1758 | if (!push_one && tcp_tso_should_defer(sk, skb)) |
c1b4a7e6 DM |
1759 | break; |
1760 | } | |
aa93466b | 1761 | |
c8ac3774 | 1762 | limit = mss_now; |
f8269a49 | 1763 | if (tso_segs > 1 && !tcp_urg_mode(tp)) |
0e3a4803 IJ |
1764 | limit = tcp_mss_split_point(sk, skb, mss_now, |
1765 | cwnd_quota); | |
1da177e4 | 1766 | |
c8ac3774 | 1767 | if (skb->len > limit && |
c4ead4c5 | 1768 | unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) |
c8ac3774 HX |
1769 | break; |
1770 | ||
92df7b51 | 1771 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
c1b4a7e6 | 1772 | |
d5dd9175 | 1773 | if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) |
92df7b51 | 1774 | break; |
1da177e4 | 1775 | |
92df7b51 DM |
1776 | /* Advance the send_head. This one is sent out. |
1777 | * This call will increment packets_out. | |
1778 | */ | |
66f5fe62 | 1779 | tcp_event_new_data_sent(sk, skb); |
1da177e4 | 1780 | |
92df7b51 | 1781 | tcp_minshall_update(tp, mss_now, skb); |
aa93466b | 1782 | sent_pkts++; |
d5dd9175 IJ |
1783 | |
1784 | if (push_one) | |
1785 | break; | |
92df7b51 | 1786 | } |
1da177e4 | 1787 | |
aa93466b | 1788 | if (likely(sent_pkts)) { |
9e412ba7 | 1789 | tcp_cwnd_validate(sk); |
92df7b51 | 1790 | return 0; |
1da177e4 | 1791 | } |
fe067e8a | 1792 | return !tp->packets_out && tcp_send_head(sk); |
1da177e4 LT |
1793 | } |
1794 | ||
a762a980 DM |
1795 | /* Push out any pending frames which were held back due to |
1796 | * TCP_CORK or attempt at coalescing tiny packets. | |
1797 | * The socket must be locked by the caller. | |
1798 | */ | |
9e412ba7 IJ |
1799 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, |
1800 | int nonagle) | |
a762a980 | 1801 | { |
726e07a8 IJ |
1802 | /* If we are closed, the bytes will have to remain here. |
1803 | * In time closedown will finish, we empty the write queue and | |
1804 | * all will be happy. | |
1805 | */ | |
1806 | if (unlikely(sk->sk_state == TCP_CLOSE)) | |
1807 | return; | |
1808 | ||
d5dd9175 | 1809 | if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC)) |
726e07a8 | 1810 | tcp_check_probe_timer(sk); |
a762a980 DM |
1811 | } |
1812 | ||
c1b4a7e6 DM |
1813 | /* Send _single_ skb sitting at the send head. This function requires |
1814 | * true push pending frames to setup probe timer etc. | |
1815 | */ | |
1816 | void tcp_push_one(struct sock *sk, unsigned int mss_now) | |
1817 | { | |
fe067e8a | 1818 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 DM |
1819 | |
1820 | BUG_ON(!skb || skb->len < mss_now); | |
1821 | ||
d5dd9175 | 1822 | tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); |
c1b4a7e6 DM |
1823 | } |
1824 | ||
1da177e4 LT |
1825 | /* This function returns the amount that we can raise the |
1826 | * usable window based on the following constraints | |
e905a9ed | 1827 | * |
1da177e4 LT |
1828 | * 1. The window can never be shrunk once it is offered (RFC 793) |
1829 | * 2. We limit memory per socket | |
1830 | * | |
1831 | * RFC 1122: | |
1832 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep | |
1833 | * RECV.NEXT + RCV.WIN fixed until: | |
1834 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" | |
1835 | * | |
1836 | * i.e. don't raise the right edge of the window until you can raise | |
1837 | * it at least MSS bytes. | |
1838 | * | |
1839 | * Unfortunately, the recommended algorithm breaks header prediction, | |
1840 | * since header prediction assumes th->window stays fixed. | |
1841 | * | |
1842 | * Strictly speaking, keeping th->window fixed violates the receiver | |
1843 | * side SWS prevention criteria. The problem is that under this rule | |
1844 | * a stream of single byte packets will cause the right side of the | |
1845 | * window to always advance by a single byte. | |
e905a9ed | 1846 | * |
1da177e4 LT |
1847 | * Of course, if the sender implements sender side SWS prevention |
1848 | * then this will not be a problem. | |
e905a9ed | 1849 | * |
1da177e4 | 1850 | * BSD seems to make the following compromise: |
e905a9ed | 1851 | * |
1da177e4 LT |
1852 | * If the free space is less than the 1/4 of the maximum |
1853 | * space available and the free space is less than 1/2 mss, | |
1854 | * then set the window to 0. | |
1855 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] | |
1856 | * Otherwise, just prevent the window from shrinking | |
1857 | * and from being larger than the largest representable value. | |
1858 | * | |
1859 | * This prevents incremental opening of the window in the regime | |
1860 | * where TCP is limited by the speed of the reader side taking | |
1861 | * data out of the TCP receive queue. It does nothing about | |
1862 | * those cases where the window is constrained on the sender side | |
1863 | * because the pipeline is full. | |
1864 | * | |
1865 | * BSD also seems to "accidentally" limit itself to windows that are a | |
1866 | * multiple of MSS, at least until the free space gets quite small. | |
1867 | * This would appear to be a side effect of the mbuf implementation. | |
1868 | * Combining these two algorithms results in the observed behavior | |
1869 | * of having a fixed window size at almost all times. | |
1870 | * | |
1871 | * Below we obtain similar behavior by forcing the offered window to | |
1872 | * a multiple of the mss when it is feasible to do so. | |
1873 | * | |
1874 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. | |
1875 | * Regular options like TIMESTAMP are taken into account. | |
1876 | */ | |
1877 | u32 __tcp_select_window(struct sock *sk) | |
1878 | { | |
463c84b9 | 1879 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 1880 | struct tcp_sock *tp = tcp_sk(sk); |
caa20d9a | 1881 | /* MSS for the peer's data. Previous versions used mss_clamp |
1da177e4 LT |
1882 | * here. I don't know if the value based on our guesses |
1883 | * of peer's MSS is better for the performance. It's more correct | |
1884 | * but may be worse for the performance because of rcv_mss | |
1885 | * fluctuations. --SAW 1998/11/1 | |
1886 | */ | |
463c84b9 | 1887 | int mss = icsk->icsk_ack.rcv_mss; |
1da177e4 LT |
1888 | int free_space = tcp_space(sk); |
1889 | int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); | |
1890 | int window; | |
1891 | ||
1892 | if (mss > full_space) | |
e905a9ed | 1893 | mss = full_space; |
1da177e4 | 1894 | |
b92edbe0 | 1895 | if (free_space < (full_space >> 1)) { |
463c84b9 | 1896 | icsk->icsk_ack.quick = 0; |
1da177e4 LT |
1897 | |
1898 | if (tcp_memory_pressure) | |
056834d9 IJ |
1899 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, |
1900 | 4U * tp->advmss); | |
1da177e4 LT |
1901 | |
1902 | if (free_space < mss) | |
1903 | return 0; | |
1904 | } | |
1905 | ||
1906 | if (free_space > tp->rcv_ssthresh) | |
1907 | free_space = tp->rcv_ssthresh; | |
1908 | ||
1909 | /* Don't do rounding if we are using window scaling, since the | |
1910 | * scaled window will not line up with the MSS boundary anyway. | |
1911 | */ | |
1912 | window = tp->rcv_wnd; | |
1913 | if (tp->rx_opt.rcv_wscale) { | |
1914 | window = free_space; | |
1915 | ||
1916 | /* Advertise enough space so that it won't get scaled away. | |
1917 | * Import case: prevent zero window announcement if | |
1918 | * 1<<rcv_wscale > mss. | |
1919 | */ | |
1920 | if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) | |
1921 | window = (((window >> tp->rx_opt.rcv_wscale) + 1) | |
1922 | << tp->rx_opt.rcv_wscale); | |
1923 | } else { | |
1924 | /* Get the largest window that is a nice multiple of mss. | |
1925 | * Window clamp already applied above. | |
1926 | * If our current window offering is within 1 mss of the | |
1927 | * free space we just keep it. This prevents the divide | |
1928 | * and multiply from happening most of the time. | |
1929 | * We also don't do any window rounding when the free space | |
1930 | * is too small. | |
1931 | */ | |
1932 | if (window <= free_space - mss || window > free_space) | |
056834d9 | 1933 | window = (free_space / mss) * mss; |
84565070 | 1934 | else if (mss == full_space && |
b92edbe0 | 1935 | free_space > window + (full_space >> 1)) |
84565070 | 1936 | window = free_space; |
1da177e4 LT |
1937 | } |
1938 | ||
1939 | return window; | |
1940 | } | |
1941 | ||
4a17fc3a IJ |
1942 | /* Collapses two adjacent SKB's during retransmission. */ |
1943 | static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) | |
1da177e4 LT |
1944 | { |
1945 | struct tcp_sock *tp = tcp_sk(sk); | |
fe067e8a | 1946 | struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); |
058dc334 | 1947 | int skb_size, next_skb_size; |
1da177e4 | 1948 | |
058dc334 IJ |
1949 | skb_size = skb->len; |
1950 | next_skb_size = next_skb->len; | |
1da177e4 | 1951 | |
058dc334 | 1952 | BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); |
a6963a6b | 1953 | |
058dc334 | 1954 | tcp_highest_sack_combine(sk, next_skb, skb); |
1da177e4 | 1955 | |
058dc334 | 1956 | tcp_unlink_write_queue(next_skb, sk); |
1da177e4 | 1957 | |
058dc334 IJ |
1958 | skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), |
1959 | next_skb_size); | |
1da177e4 | 1960 | |
058dc334 IJ |
1961 | if (next_skb->ip_summed == CHECKSUM_PARTIAL) |
1962 | skb->ip_summed = CHECKSUM_PARTIAL; | |
1da177e4 | 1963 | |
058dc334 IJ |
1964 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
1965 | skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); | |
1da177e4 | 1966 | |
058dc334 IJ |
1967 | /* Update sequence range on original skb. */ |
1968 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; | |
1da177e4 | 1969 | |
e6c7d085 IJ |
1970 | /* Merge over control information. This moves PSH/FIN etc. over */ |
1971 | TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(next_skb)->flags; | |
058dc334 IJ |
1972 | |
1973 | /* All done, get rid of second SKB and account for it so | |
1974 | * packet counting does not break. | |
1975 | */ | |
1976 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; | |
058dc334 IJ |
1977 | |
1978 | /* changed transmit queue under us so clear hints */ | |
ef9da47c IJ |
1979 | tcp_clear_retrans_hints_partial(tp); |
1980 | if (next_skb == tp->retransmit_skb_hint) | |
1981 | tp->retransmit_skb_hint = skb; | |
058dc334 | 1982 | |
797108d1 IJ |
1983 | tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); |
1984 | ||
058dc334 | 1985 | sk_wmem_free_skb(sk, next_skb); |
1da177e4 LT |
1986 | } |
1987 | ||
67edfef7 | 1988 | /* Check if coalescing SKBs is legal. */ |
4a17fc3a IJ |
1989 | static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb) |
1990 | { | |
1991 | if (tcp_skb_pcount(skb) > 1) | |
1992 | return 0; | |
1993 | /* TODO: SACK collapsing could be used to remove this condition */ | |
1994 | if (skb_shinfo(skb)->nr_frags != 0) | |
1995 | return 0; | |
1996 | if (skb_cloned(skb)) | |
1997 | return 0; | |
1998 | if (skb == tcp_send_head(sk)) | |
1999 | return 0; | |
2000 | /* Some heurestics for collapsing over SACK'd could be invented */ | |
2001 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
2002 | return 0; | |
2003 | ||
2004 | return 1; | |
2005 | } | |
2006 | ||
67edfef7 AK |
2007 | /* Collapse packets in the retransmit queue to make to create |
2008 | * less packets on the wire. This is only done on retransmission. | |
2009 | */ | |
4a17fc3a IJ |
2010 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, |
2011 | int space) | |
2012 | { | |
2013 | struct tcp_sock *tp = tcp_sk(sk); | |
2014 | struct sk_buff *skb = to, *tmp; | |
2015 | int first = 1; | |
2016 | ||
2017 | if (!sysctl_tcp_retrans_collapse) | |
2018 | return; | |
a3433f35 | 2019 | if (TCP_SKB_CB(skb)->flags & TCPHDR_SYN) |
4a17fc3a IJ |
2020 | return; |
2021 | ||
2022 | tcp_for_write_queue_from_safe(skb, tmp, sk) { | |
2023 | if (!tcp_can_collapse(sk, skb)) | |
2024 | break; | |
2025 | ||
2026 | space -= skb->len; | |
2027 | ||
2028 | if (first) { | |
2029 | first = 0; | |
2030 | continue; | |
2031 | } | |
2032 | ||
2033 | if (space < 0) | |
2034 | break; | |
2035 | /* Punt if not enough space exists in the first SKB for | |
2036 | * the data in the second | |
2037 | */ | |
2038 | if (skb->len > skb_tailroom(to)) | |
2039 | break; | |
2040 | ||
2041 | if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) | |
2042 | break; | |
2043 | ||
2044 | tcp_collapse_retrans(sk, to); | |
2045 | } | |
2046 | } | |
2047 | ||
1da177e4 LT |
2048 | /* This retransmits one SKB. Policy decisions and retransmit queue |
2049 | * state updates are done by the caller. Returns non-zero if an | |
2050 | * error occurred which prevented the send. | |
2051 | */ | |
2052 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) | |
2053 | { | |
2054 | struct tcp_sock *tp = tcp_sk(sk); | |
5d424d5a | 2055 | struct inet_connection_sock *icsk = inet_csk(sk); |
7d227cd2 | 2056 | unsigned int cur_mss; |
1da177e4 LT |
2057 | int err; |
2058 | ||
5d424d5a JH |
2059 | /* Inconslusive MTU probe */ |
2060 | if (icsk->icsk_mtup.probe_size) { | |
2061 | icsk->icsk_mtup.probe_size = 0; | |
2062 | } | |
2063 | ||
1da177e4 | 2064 | /* Do not sent more than we queued. 1/4 is reserved for possible |
caa20d9a | 2065 | * copying overhead: fragmentation, tunneling, mangling etc. |
1da177e4 LT |
2066 | */ |
2067 | if (atomic_read(&sk->sk_wmem_alloc) > | |
2068 | min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) | |
2069 | return -EAGAIN; | |
2070 | ||
2071 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { | |
2072 | if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
2073 | BUG(); | |
1da177e4 LT |
2074 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
2075 | return -ENOMEM; | |
2076 | } | |
2077 | ||
7d227cd2 SS |
2078 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) |
2079 | return -EHOSTUNREACH; /* Routing failure or similar. */ | |
2080 | ||
0c54b85f | 2081 | cur_mss = tcp_current_mss(sk); |
7d227cd2 | 2082 | |
1da177e4 LT |
2083 | /* If receiver has shrunk his window, and skb is out of |
2084 | * new window, do not retransmit it. The exception is the | |
2085 | * case, when window is shrunk to zero. In this case | |
2086 | * our retransmit serves as a zero window probe. | |
2087 | */ | |
9d4fb27d JP |
2088 | if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && |
2089 | TCP_SKB_CB(skb)->seq != tp->snd_una) | |
1da177e4 LT |
2090 | return -EAGAIN; |
2091 | ||
2092 | if (skb->len > cur_mss) { | |
846998ae | 2093 | if (tcp_fragment(sk, skb, cur_mss, cur_mss)) |
1da177e4 | 2094 | return -ENOMEM; /* We'll try again later. */ |
02276f3c | 2095 | } else { |
9eb9362e IJ |
2096 | int oldpcount = tcp_skb_pcount(skb); |
2097 | ||
2098 | if (unlikely(oldpcount > 1)) { | |
2099 | tcp_init_tso_segs(sk, skb, cur_mss); | |
2100 | tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); | |
2101 | } | |
1da177e4 LT |
2102 | } |
2103 | ||
4a17fc3a | 2104 | tcp_retrans_try_collapse(sk, skb, cur_mss); |
1da177e4 | 2105 | |
1da177e4 LT |
2106 | /* Some Solaris stacks overoptimize and ignore the FIN on a |
2107 | * retransmit when old data is attached. So strip it off | |
2108 | * since it is cheap to do so and saves bytes on the network. | |
2109 | */ | |
2de979bd | 2110 | if (skb->len > 0 && |
a3433f35 | 2111 | (TCP_SKB_CB(skb)->flags & TCPHDR_FIN) && |
2de979bd | 2112 | tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { |
1da177e4 | 2113 | if (!pskb_trim(skb, 0)) { |
e870a8ef IJ |
2114 | /* Reuse, even though it does some unnecessary work */ |
2115 | tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, | |
2116 | TCP_SKB_CB(skb)->flags); | |
1da177e4 | 2117 | skb->ip_summed = CHECKSUM_NONE; |
1da177e4 LT |
2118 | } |
2119 | } | |
2120 | ||
2121 | /* Make a copy, if the first transmission SKB clone we made | |
2122 | * is still in somebody's hands, else make a clone. | |
2123 | */ | |
2124 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
1da177e4 | 2125 | |
dfb4b9dc | 2126 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
1da177e4 LT |
2127 | |
2128 | if (err == 0) { | |
2129 | /* Update global TCP statistics. */ | |
81cc8a75 | 2130 | TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); |
1da177e4 LT |
2131 | |
2132 | tp->total_retrans++; | |
2133 | ||
2134 | #if FASTRETRANS_DEBUG > 0 | |
056834d9 | 2135 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
1da177e4 LT |
2136 | if (net_ratelimit()) |
2137 | printk(KERN_DEBUG "retrans_out leaked.\n"); | |
2138 | } | |
2139 | #endif | |
b08d6cb2 IJ |
2140 | if (!tp->retrans_out) |
2141 | tp->lost_retrans_low = tp->snd_nxt; | |
1da177e4 LT |
2142 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; |
2143 | tp->retrans_out += tcp_skb_pcount(skb); | |
2144 | ||
2145 | /* Save stamp of the first retransmit. */ | |
2146 | if (!tp->retrans_stamp) | |
2147 | tp->retrans_stamp = TCP_SKB_CB(skb)->when; | |
2148 | ||
2149 | tp->undo_retrans++; | |
2150 | ||
2151 | /* snd_nxt is stored to detect loss of retransmitted segment, | |
2152 | * see tcp_input.c tcp_sacktag_write_queue(). | |
2153 | */ | |
2154 | TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; | |
2155 | } | |
2156 | return err; | |
2157 | } | |
2158 | ||
67edfef7 AK |
2159 | /* Check if we forward retransmits are possible in the current |
2160 | * window/congestion state. | |
2161 | */ | |
b5afe7bc IJ |
2162 | static int tcp_can_forward_retransmit(struct sock *sk) |
2163 | { | |
2164 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
2165 | struct tcp_sock *tp = tcp_sk(sk); | |
2166 | ||
2167 | /* Forward retransmissions are possible only during Recovery. */ | |
2168 | if (icsk->icsk_ca_state != TCP_CA_Recovery) | |
2169 | return 0; | |
2170 | ||
2171 | /* No forward retransmissions in Reno are possible. */ | |
2172 | if (tcp_is_reno(tp)) | |
2173 | return 0; | |
2174 | ||
2175 | /* Yeah, we have to make difficult choice between forward transmission | |
2176 | * and retransmission... Both ways have their merits... | |
2177 | * | |
2178 | * For now we do not retransmit anything, while we have some new | |
2179 | * segments to send. In the other cases, follow rule 3 for | |
2180 | * NextSeg() specified in RFC3517. | |
2181 | */ | |
2182 | ||
2183 | if (tcp_may_send_now(sk)) | |
2184 | return 0; | |
2185 | ||
2186 | return 1; | |
2187 | } | |
2188 | ||
1da177e4 LT |
2189 | /* This gets called after a retransmit timeout, and the initially |
2190 | * retransmitted data is acknowledged. It tries to continue | |
2191 | * resending the rest of the retransmit queue, until either | |
2192 | * we've sent it all or the congestion window limit is reached. | |
2193 | * If doing SACK, the first ACK which comes back for a timeout | |
2194 | * based retransmit packet might feed us FACK information again. | |
2195 | * If so, we use it to avoid unnecessarily retransmissions. | |
2196 | */ | |
2197 | void tcp_xmit_retransmit_queue(struct sock *sk) | |
2198 | { | |
6687e988 | 2199 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2200 | struct tcp_sock *tp = tcp_sk(sk); |
2201 | struct sk_buff *skb; | |
0e1c54c2 | 2202 | struct sk_buff *hole = NULL; |
618d9f25 | 2203 | u32 last_lost; |
61eb55f4 | 2204 | int mib_idx; |
0e1c54c2 | 2205 | int fwd_rexmitting = 0; |
6a438bbe | 2206 | |
45e77d31 IJ |
2207 | if (!tp->packets_out) |
2208 | return; | |
2209 | ||
08ebd172 IJ |
2210 | if (!tp->lost_out) |
2211 | tp->retransmit_high = tp->snd_una; | |
2212 | ||
618d9f25 | 2213 | if (tp->retransmit_skb_hint) { |
6a438bbe | 2214 | skb = tp->retransmit_skb_hint; |
618d9f25 IJ |
2215 | last_lost = TCP_SKB_CB(skb)->end_seq; |
2216 | if (after(last_lost, tp->retransmit_high)) | |
2217 | last_lost = tp->retransmit_high; | |
2218 | } else { | |
fe067e8a | 2219 | skb = tcp_write_queue_head(sk); |
618d9f25 IJ |
2220 | last_lost = tp->snd_una; |
2221 | } | |
1da177e4 | 2222 | |
08ebd172 IJ |
2223 | tcp_for_write_queue_from(skb, sk) { |
2224 | __u8 sacked = TCP_SKB_CB(skb)->sacked; | |
1da177e4 | 2225 | |
08ebd172 IJ |
2226 | if (skb == tcp_send_head(sk)) |
2227 | break; | |
2228 | /* we could do better than to assign each time */ | |
0e1c54c2 IJ |
2229 | if (hole == NULL) |
2230 | tp->retransmit_skb_hint = skb; | |
08ebd172 IJ |
2231 | |
2232 | /* Assume this retransmit will generate | |
2233 | * only one packet for congestion window | |
2234 | * calculation purposes. This works because | |
2235 | * tcp_retransmit_skb() will chop up the | |
2236 | * packet to be MSS sized and all the | |
2237 | * packet counting works out. | |
2238 | */ | |
2239 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) | |
2240 | return; | |
1da177e4 | 2241 | |
0e1c54c2 IJ |
2242 | if (fwd_rexmitting) { |
2243 | begin_fwd: | |
2244 | if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) | |
2245 | break; | |
2246 | mib_idx = LINUX_MIB_TCPFORWARDRETRANS; | |
6a438bbe | 2247 | |
0e1c54c2 | 2248 | } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { |
618d9f25 | 2249 | tp->retransmit_high = last_lost; |
0e1c54c2 IJ |
2250 | if (!tcp_can_forward_retransmit(sk)) |
2251 | break; | |
2252 | /* Backtrack if necessary to non-L'ed skb */ | |
2253 | if (hole != NULL) { | |
2254 | skb = hole; | |
2255 | hole = NULL; | |
2256 | } | |
2257 | fwd_rexmitting = 1; | |
2258 | goto begin_fwd; | |
1da177e4 | 2259 | |
0e1c54c2 | 2260 | } else if (!(sacked & TCPCB_LOST)) { |
ac11ba75 | 2261 | if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) |
0e1c54c2 IJ |
2262 | hole = skb; |
2263 | continue; | |
1da177e4 | 2264 | |
0e1c54c2 | 2265 | } else { |
618d9f25 | 2266 | last_lost = TCP_SKB_CB(skb)->end_seq; |
0e1c54c2 IJ |
2267 | if (icsk->icsk_ca_state != TCP_CA_Loss) |
2268 | mib_idx = LINUX_MIB_TCPFASTRETRANS; | |
2269 | else | |
2270 | mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; | |
2271 | } | |
1da177e4 | 2272 | |
0e1c54c2 | 2273 | if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) |
1da177e4 LT |
2274 | continue; |
2275 | ||
f0ceb0ed | 2276 | if (tcp_retransmit_skb(sk, skb)) |
0e1c54c2 | 2277 | return; |
0e1c54c2 | 2278 | NET_INC_STATS_BH(sock_net(sk), mib_idx); |
1da177e4 | 2279 | |
fe067e8a | 2280 | if (skb == tcp_write_queue_head(sk)) |
3f421baa ACM |
2281 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
2282 | inet_csk(sk)->icsk_rto, | |
2283 | TCP_RTO_MAX); | |
1da177e4 LT |
2284 | } |
2285 | } | |
2286 | ||
1da177e4 LT |
2287 | /* Send a fin. The caller locks the socket for us. This cannot be |
2288 | * allowed to fail queueing a FIN frame under any circumstances. | |
2289 | */ | |
2290 | void tcp_send_fin(struct sock *sk) | |
2291 | { | |
e905a9ed | 2292 | struct tcp_sock *tp = tcp_sk(sk); |
fe067e8a | 2293 | struct sk_buff *skb = tcp_write_queue_tail(sk); |
1da177e4 | 2294 | int mss_now; |
e905a9ed | 2295 | |
1da177e4 LT |
2296 | /* Optimization, tack on the FIN if we have a queue of |
2297 | * unsent frames. But be careful about outgoing SACKS | |
2298 | * and IP options. | |
2299 | */ | |
0c54b85f | 2300 | mss_now = tcp_current_mss(sk); |
1da177e4 | 2301 | |
fe067e8a | 2302 | if (tcp_send_head(sk) != NULL) { |
a3433f35 | 2303 | TCP_SKB_CB(skb)->flags |= TCPHDR_FIN; |
1da177e4 LT |
2304 | TCP_SKB_CB(skb)->end_seq++; |
2305 | tp->write_seq++; | |
2306 | } else { | |
2307 | /* Socket is locked, keep trying until memory is available. */ | |
2308 | for (;;) { | |
aa133076 WF |
2309 | skb = alloc_skb_fclone(MAX_TCP_HEADER, |
2310 | sk->sk_allocation); | |
1da177e4 LT |
2311 | if (skb) |
2312 | break; | |
2313 | yield(); | |
2314 | } | |
2315 | ||
2316 | /* Reserve space for headers and prepare control bits. */ | |
2317 | skb_reserve(skb, MAX_TCP_HEADER); | |
1da177e4 | 2318 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ |
e870a8ef | 2319 | tcp_init_nondata_skb(skb, tp->write_seq, |
a3433f35 | 2320 | TCPHDR_ACK | TCPHDR_FIN); |
1da177e4 LT |
2321 | tcp_queue_skb(sk, skb); |
2322 | } | |
9e412ba7 | 2323 | __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); |
1da177e4 LT |
2324 | } |
2325 | ||
2326 | /* We get here when a process closes a file descriptor (either due to | |
2327 | * an explicit close() or as a byproduct of exit()'ing) and there | |
2328 | * was unread data in the receive queue. This behavior is recommended | |
65bb723c | 2329 | * by RFC 2525, section 2.17. -DaveM |
1da177e4 | 2330 | */ |
dd0fc66f | 2331 | void tcp_send_active_reset(struct sock *sk, gfp_t priority) |
1da177e4 | 2332 | { |
1da177e4 LT |
2333 | struct sk_buff *skb; |
2334 | ||
2335 | /* NOTE: No TCP options attached and we never retransmit this. */ | |
2336 | skb = alloc_skb(MAX_TCP_HEADER, priority); | |
2337 | if (!skb) { | |
4e673444 | 2338 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
1da177e4 LT |
2339 | return; |
2340 | } | |
2341 | ||
2342 | /* Reserve space for headers and prepare control bits. */ | |
2343 | skb_reserve(skb, MAX_TCP_HEADER); | |
e870a8ef | 2344 | tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), |
a3433f35 | 2345 | TCPHDR_ACK | TCPHDR_RST); |
1da177e4 | 2346 | /* Send it off. */ |
1da177e4 | 2347 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
dfb4b9dc | 2348 | if (tcp_transmit_skb(sk, skb, 0, priority)) |
4e673444 | 2349 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
26af65cb | 2350 | |
81cc8a75 | 2351 | TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); |
1da177e4 LT |
2352 | } |
2353 | ||
67edfef7 AK |
2354 | /* Send a crossed SYN-ACK during socket establishment. |
2355 | * WARNING: This routine must only be called when we have already sent | |
1da177e4 LT |
2356 | * a SYN packet that crossed the incoming SYN that caused this routine |
2357 | * to get called. If this assumption fails then the initial rcv_wnd | |
2358 | * and rcv_wscale values will not be correct. | |
2359 | */ | |
2360 | int tcp_send_synack(struct sock *sk) | |
2361 | { | |
056834d9 | 2362 | struct sk_buff *skb; |
1da177e4 | 2363 | |
fe067e8a | 2364 | skb = tcp_write_queue_head(sk); |
a3433f35 | 2365 | if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPHDR_SYN)) { |
1da177e4 LT |
2366 | printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n"); |
2367 | return -EFAULT; | |
2368 | } | |
a3433f35 | 2369 | if (!(TCP_SKB_CB(skb)->flags & TCPHDR_ACK)) { |
1da177e4 LT |
2370 | if (skb_cloned(skb)) { |
2371 | struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); | |
2372 | if (nskb == NULL) | |
2373 | return -ENOMEM; | |
fe067e8a | 2374 | tcp_unlink_write_queue(skb, sk); |
1da177e4 | 2375 | skb_header_release(nskb); |
fe067e8a | 2376 | __tcp_add_write_queue_head(sk, nskb); |
3ab224be HA |
2377 | sk_wmem_free_skb(sk, skb); |
2378 | sk->sk_wmem_queued += nskb->truesize; | |
2379 | sk_mem_charge(sk, nskb->truesize); | |
1da177e4 LT |
2380 | skb = nskb; |
2381 | } | |
2382 | ||
a3433f35 | 2383 | TCP_SKB_CB(skb)->flags |= TCPHDR_ACK; |
1da177e4 LT |
2384 | TCP_ECN_send_synack(tcp_sk(sk), skb); |
2385 | } | |
2386 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
dfb4b9dc | 2387 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
1da177e4 LT |
2388 | } |
2389 | ||
67edfef7 | 2390 | /* Prepare a SYN-ACK. */ |
056834d9 | 2391 | struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, |
e6b4d113 WAS |
2392 | struct request_sock *req, |
2393 | struct request_values *rvp) | |
1da177e4 | 2394 | { |
bd0388ae | 2395 | struct tcp_out_options opts; |
4957faad | 2396 | struct tcp_extend_values *xvp = tcp_xv(rvp); |
2e6599cb | 2397 | struct inet_request_sock *ireq = inet_rsk(req); |
1da177e4 | 2398 | struct tcp_sock *tp = tcp_sk(sk); |
28b2774a | 2399 | const struct tcp_cookie_values *cvp = tp->cookie_values; |
1da177e4 | 2400 | struct tcphdr *th; |
1da177e4 | 2401 | struct sk_buff *skb; |
cfb6eeb4 | 2402 | struct tcp_md5sig_key *md5; |
bd0388ae | 2403 | int tcp_header_size; |
f5fff5dc | 2404 | int mss; |
28b2774a | 2405 | int s_data_desired = 0; |
1da177e4 | 2406 | |
28b2774a ED |
2407 | if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired) |
2408 | s_data_desired = cvp->s_data_desired; | |
2409 | skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15 + s_data_desired, 1, GFP_ATOMIC); | |
1da177e4 LT |
2410 | if (skb == NULL) |
2411 | return NULL; | |
2412 | ||
2413 | /* Reserve space for headers. */ | |
2414 | skb_reserve(skb, MAX_TCP_HEADER); | |
2415 | ||
adf30907 | 2416 | skb_dst_set(skb, dst_clone(dst)); |
1da177e4 | 2417 | |
f5fff5dc TQ |
2418 | mss = dst_metric(dst, RTAX_ADVMSS); |
2419 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) | |
2420 | mss = tp->rx_opt.user_mss; | |
2421 | ||
33ad798c AL |
2422 | if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ |
2423 | __u8 rcv_wscale; | |
2424 | /* Set this up on the first call only */ | |
2425 | req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); | |
e88c64f0 HPP |
2426 | |
2427 | /* limit the window selection if the user enforce a smaller rx buffer */ | |
2428 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && | |
2429 | (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) | |
2430 | req->window_clamp = tcp_full_space(sk); | |
2431 | ||
33ad798c AL |
2432 | /* tcp_full_space because it is guaranteed to be the first packet */ |
2433 | tcp_select_initial_window(tcp_full_space(sk), | |
f5fff5dc | 2434 | mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), |
33ad798c AL |
2435 | &req->rcv_wnd, |
2436 | &req->window_clamp, | |
2437 | ireq->wscale_ok, | |
31d12926 | 2438 | &rcv_wscale, |
2439 | dst_metric(dst, RTAX_INITRWND)); | |
33ad798c AL |
2440 | ireq->rcv_wscale = rcv_wscale; |
2441 | } | |
2442 | ||
2443 | memset(&opts, 0, sizeof(opts)); | |
8b5f12d0 FW |
2444 | #ifdef CONFIG_SYN_COOKIES |
2445 | if (unlikely(req->cookie_ts)) | |
2446 | TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); | |
2447 | else | |
2448 | #endif | |
33ad798c | 2449 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
f5fff5dc | 2450 | tcp_header_size = tcp_synack_options(sk, req, mss, |
4957faad WAS |
2451 | skb, &opts, &md5, xvp) |
2452 | + sizeof(*th); | |
cfb6eeb4 | 2453 | |
aa8223c7 ACM |
2454 | skb_push(skb, tcp_header_size); |
2455 | skb_reset_transport_header(skb); | |
1da177e4 | 2456 | |
aa8223c7 | 2457 | th = tcp_hdr(skb); |
1da177e4 LT |
2458 | memset(th, 0, sizeof(struct tcphdr)); |
2459 | th->syn = 1; | |
2460 | th->ack = 1; | |
1da177e4 | 2461 | TCP_ECN_make_synack(req, th); |
a3116ac5 | 2462 | th->source = ireq->loc_port; |
2e6599cb | 2463 | th->dest = ireq->rmt_port; |
e870a8ef IJ |
2464 | /* Setting of flags are superfluous here for callers (and ECE is |
2465 | * not even correctly set) | |
2466 | */ | |
2467 | tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, | |
a3433f35 | 2468 | TCPHDR_SYN | TCPHDR_ACK); |
4957faad WAS |
2469 | |
2470 | if (OPTION_COOKIE_EXTENSION & opts.options) { | |
28b2774a ED |
2471 | if (s_data_desired) { |
2472 | u8 *buf = skb_put(skb, s_data_desired); | |
4957faad WAS |
2473 | |
2474 | /* copy data directly from the listening socket. */ | |
28b2774a ED |
2475 | memcpy(buf, cvp->s_data_payload, s_data_desired); |
2476 | TCP_SKB_CB(skb)->end_seq += s_data_desired; | |
4957faad WAS |
2477 | } |
2478 | ||
2479 | if (opts.hash_size > 0) { | |
2480 | __u32 workspace[SHA_WORKSPACE_WORDS]; | |
2481 | u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS]; | |
2482 | u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1]; | |
2483 | ||
2484 | /* Secret recipe depends on the Timestamp, (future) | |
2485 | * Sequence and Acknowledgment Numbers, Initiator | |
2486 | * Cookie, and others handled by IP variant caller. | |
2487 | */ | |
2488 | *tail-- ^= opts.tsval; | |
2489 | *tail-- ^= tcp_rsk(req)->rcv_isn + 1; | |
2490 | *tail-- ^= TCP_SKB_CB(skb)->seq + 1; | |
2491 | ||
2492 | /* recommended */ | |
0eae88f3 | 2493 | *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source); |
f9a2e69e | 2494 | *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */ |
4957faad WAS |
2495 | |
2496 | sha_transform((__u32 *)&xvp->cookie_bakery[0], | |
2497 | (char *)mess, | |
2498 | &workspace[0]); | |
2499 | opts.hash_location = | |
2500 | (__u8 *)&xvp->cookie_bakery[0]; | |
2501 | } | |
2502 | } | |
2503 | ||
1da177e4 | 2504 | th->seq = htonl(TCP_SKB_CB(skb)->seq); |
2e6599cb | 2505 | th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); |
1da177e4 LT |
2506 | |
2507 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ | |
600ff0c2 | 2508 | th->window = htons(min(req->rcv_wnd, 65535U)); |
bd0388ae | 2509 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
1da177e4 | 2510 | th->doff = (tcp_header_size >> 2); |
aa2ea058 | 2511 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb)); |
cfb6eeb4 YH |
2512 | |
2513 | #ifdef CONFIG_TCP_MD5SIG | |
2514 | /* Okay, we have all we need - do the md5 hash if needed */ | |
2515 | if (md5) { | |
bd0388ae | 2516 | tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, |
49a72dfb | 2517 | md5, NULL, req, skb); |
cfb6eeb4 YH |
2518 | } |
2519 | #endif | |
2520 | ||
1da177e4 LT |
2521 | return skb; |
2522 | } | |
4bc2f18b | 2523 | EXPORT_SYMBOL(tcp_make_synack); |
1da177e4 | 2524 | |
67edfef7 | 2525 | /* Do all connect socket setups that can be done AF independent. */ |
40efc6fa | 2526 | static void tcp_connect_init(struct sock *sk) |
1da177e4 LT |
2527 | { |
2528 | struct dst_entry *dst = __sk_dst_get(sk); | |
2529 | struct tcp_sock *tp = tcp_sk(sk); | |
2530 | __u8 rcv_wscale; | |
2531 | ||
2532 | /* We'll fix this up when we get a response from the other end. | |
2533 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. | |
2534 | */ | |
2535 | tp->tcp_header_len = sizeof(struct tcphdr) + | |
bb5b7c11 | 2536 | (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); |
1da177e4 | 2537 | |
cfb6eeb4 YH |
2538 | #ifdef CONFIG_TCP_MD5SIG |
2539 | if (tp->af_specific->md5_lookup(sk, sk) != NULL) | |
2540 | tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; | |
2541 | #endif | |
2542 | ||
1da177e4 LT |
2543 | /* If user gave his TCP_MAXSEG, record it to clamp */ |
2544 | if (tp->rx_opt.user_mss) | |
2545 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
2546 | tp->max_window = 0; | |
5d424d5a | 2547 | tcp_mtup_init(sk); |
1da177e4 LT |
2548 | tcp_sync_mss(sk, dst_mtu(dst)); |
2549 | ||
2550 | if (!tp->window_clamp) | |
2551 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); | |
2552 | tp->advmss = dst_metric(dst, RTAX_ADVMSS); | |
f5fff5dc TQ |
2553 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) |
2554 | tp->advmss = tp->rx_opt.user_mss; | |
2555 | ||
1da177e4 | 2556 | tcp_initialize_rcv_mss(sk); |
1da177e4 | 2557 | |
e88c64f0 HPP |
2558 | /* limit the window selection if the user enforce a smaller rx buffer */ |
2559 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && | |
2560 | (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) | |
2561 | tp->window_clamp = tcp_full_space(sk); | |
2562 | ||
1da177e4 LT |
2563 | tcp_select_initial_window(tcp_full_space(sk), |
2564 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), | |
2565 | &tp->rcv_wnd, | |
2566 | &tp->window_clamp, | |
bb5b7c11 | 2567 | sysctl_tcp_window_scaling, |
31d12926 | 2568 | &rcv_wscale, |
2569 | dst_metric(dst, RTAX_INITRWND)); | |
1da177e4 LT |
2570 | |
2571 | tp->rx_opt.rcv_wscale = rcv_wscale; | |
2572 | tp->rcv_ssthresh = tp->rcv_wnd; | |
2573 | ||
2574 | sk->sk_err = 0; | |
2575 | sock_reset_flag(sk, SOCK_DONE); | |
2576 | tp->snd_wnd = 0; | |
ee7537b6 | 2577 | tcp_init_wl(tp, 0); |
1da177e4 LT |
2578 | tp->snd_una = tp->write_seq; |
2579 | tp->snd_sml = tp->write_seq; | |
33f5f57e | 2580 | tp->snd_up = tp->write_seq; |
1da177e4 LT |
2581 | tp->rcv_nxt = 0; |
2582 | tp->rcv_wup = 0; | |
2583 | tp->copied_seq = 0; | |
2584 | ||
463c84b9 ACM |
2585 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
2586 | inet_csk(sk)->icsk_retransmits = 0; | |
1da177e4 LT |
2587 | tcp_clear_retrans(tp); |
2588 | } | |
2589 | ||
67edfef7 | 2590 | /* Build a SYN and send it off. */ |
1da177e4 LT |
2591 | int tcp_connect(struct sock *sk) |
2592 | { | |
2593 | struct tcp_sock *tp = tcp_sk(sk); | |
2594 | struct sk_buff *buff; | |
ee586811 | 2595 | int err; |
1da177e4 LT |
2596 | |
2597 | tcp_connect_init(sk); | |
2598 | ||
d179cd12 | 2599 | buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); |
1da177e4 LT |
2600 | if (unlikely(buff == NULL)) |
2601 | return -ENOBUFS; | |
2602 | ||
2603 | /* Reserve space for headers. */ | |
2604 | skb_reserve(buff, MAX_TCP_HEADER); | |
2605 | ||
bd37a088 | 2606 | tp->snd_nxt = tp->write_seq; |
a3433f35 | 2607 | tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); |
e870a8ef | 2608 | TCP_ECN_send_syn(sk, buff); |
1da177e4 LT |
2609 | |
2610 | /* Send it off. */ | |
2611 | TCP_SKB_CB(buff)->when = tcp_time_stamp; | |
2612 | tp->retrans_stamp = TCP_SKB_CB(buff)->when; | |
2613 | skb_header_release(buff); | |
fe067e8a | 2614 | __tcp_add_write_queue_tail(sk, buff); |
3ab224be HA |
2615 | sk->sk_wmem_queued += buff->truesize; |
2616 | sk_mem_charge(sk, buff->truesize); | |
1da177e4 | 2617 | tp->packets_out += tcp_skb_pcount(buff); |
ee586811 EP |
2618 | err = tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); |
2619 | if (err == -ECONNREFUSED) | |
2620 | return err; | |
bd37a088 WY |
2621 | |
2622 | /* We change tp->snd_nxt after the tcp_transmit_skb() call | |
2623 | * in order to make this packet get counted in tcpOutSegs. | |
2624 | */ | |
2625 | tp->snd_nxt = tp->write_seq; | |
2626 | tp->pushed_seq = tp->write_seq; | |
81cc8a75 | 2627 | TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); |
1da177e4 LT |
2628 | |
2629 | /* Timer for repeating the SYN until an answer. */ | |
3f421baa ACM |
2630 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
2631 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
1da177e4 LT |
2632 | return 0; |
2633 | } | |
4bc2f18b | 2634 | EXPORT_SYMBOL(tcp_connect); |
1da177e4 LT |
2635 | |
2636 | /* Send out a delayed ack, the caller does the policy checking | |
2637 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() | |
2638 | * for details. | |
2639 | */ | |
2640 | void tcp_send_delayed_ack(struct sock *sk) | |
2641 | { | |
463c84b9 ACM |
2642 | struct inet_connection_sock *icsk = inet_csk(sk); |
2643 | int ato = icsk->icsk_ack.ato; | |
1da177e4 LT |
2644 | unsigned long timeout; |
2645 | ||
2646 | if (ato > TCP_DELACK_MIN) { | |
463c84b9 | 2647 | const struct tcp_sock *tp = tcp_sk(sk); |
056834d9 | 2648 | int max_ato = HZ / 2; |
1da177e4 | 2649 | |
056834d9 IJ |
2650 | if (icsk->icsk_ack.pingpong || |
2651 | (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) | |
1da177e4 LT |
2652 | max_ato = TCP_DELACK_MAX; |
2653 | ||
2654 | /* Slow path, intersegment interval is "high". */ | |
2655 | ||
2656 | /* If some rtt estimate is known, use it to bound delayed ack. | |
463c84b9 | 2657 | * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements |
1da177e4 LT |
2658 | * directly. |
2659 | */ | |
2660 | if (tp->srtt) { | |
056834d9 | 2661 | int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); |
1da177e4 LT |
2662 | |
2663 | if (rtt < max_ato) | |
2664 | max_ato = rtt; | |
2665 | } | |
2666 | ||
2667 | ato = min(ato, max_ato); | |
2668 | } | |
2669 | ||
2670 | /* Stay within the limit we were given */ | |
2671 | timeout = jiffies + ato; | |
2672 | ||
2673 | /* Use new timeout only if there wasn't a older one earlier. */ | |
463c84b9 | 2674 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { |
1da177e4 LT |
2675 | /* If delack timer was blocked or is about to expire, |
2676 | * send ACK now. | |
2677 | */ | |
463c84b9 ACM |
2678 | if (icsk->icsk_ack.blocked || |
2679 | time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { | |
1da177e4 LT |
2680 | tcp_send_ack(sk); |
2681 | return; | |
2682 | } | |
2683 | ||
463c84b9 ACM |
2684 | if (!time_before(timeout, icsk->icsk_ack.timeout)) |
2685 | timeout = icsk->icsk_ack.timeout; | |
1da177e4 | 2686 | } |
463c84b9 ACM |
2687 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; |
2688 | icsk->icsk_ack.timeout = timeout; | |
2689 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); | |
1da177e4 LT |
2690 | } |
2691 | ||
2692 | /* This routine sends an ack and also updates the window. */ | |
2693 | void tcp_send_ack(struct sock *sk) | |
2694 | { | |
058dc334 | 2695 | struct sk_buff *buff; |
1da177e4 | 2696 | |
058dc334 IJ |
2697 | /* If we have been reset, we may not send again. */ |
2698 | if (sk->sk_state == TCP_CLOSE) | |
2699 | return; | |
1da177e4 | 2700 | |
058dc334 IJ |
2701 | /* We are not putting this on the write queue, so |
2702 | * tcp_transmit_skb() will set the ownership to this | |
2703 | * sock. | |
2704 | */ | |
2705 | buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); | |
2706 | if (buff == NULL) { | |
2707 | inet_csk_schedule_ack(sk); | |
2708 | inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; | |
2709 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, | |
2710 | TCP_DELACK_MAX, TCP_RTO_MAX); | |
2711 | return; | |
1da177e4 | 2712 | } |
058dc334 IJ |
2713 | |
2714 | /* Reserve space for headers and prepare control bits. */ | |
2715 | skb_reserve(buff, MAX_TCP_HEADER); | |
a3433f35 | 2716 | tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); |
058dc334 IJ |
2717 | |
2718 | /* Send it off, this clears delayed acks for us. */ | |
058dc334 IJ |
2719 | TCP_SKB_CB(buff)->when = tcp_time_stamp; |
2720 | tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); | |
1da177e4 LT |
2721 | } |
2722 | ||
2723 | /* This routine sends a packet with an out of date sequence | |
2724 | * number. It assumes the other end will try to ack it. | |
2725 | * | |
2726 | * Question: what should we make while urgent mode? | |
2727 | * 4.4BSD forces sending single byte of data. We cannot send | |
2728 | * out of window data, because we have SND.NXT==SND.MAX... | |
2729 | * | |
2730 | * Current solution: to send TWO zero-length segments in urgent mode: | |
2731 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is | |
2732 | * out-of-date with SND.UNA-1 to probe window. | |
2733 | */ | |
2734 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent) | |
2735 | { | |
2736 | struct tcp_sock *tp = tcp_sk(sk); | |
2737 | struct sk_buff *skb; | |
2738 | ||
2739 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ | |
2740 | skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); | |
e905a9ed | 2741 | if (skb == NULL) |
1da177e4 LT |
2742 | return -1; |
2743 | ||
2744 | /* Reserve space for headers and set control bits. */ | |
2745 | skb_reserve(skb, MAX_TCP_HEADER); | |
1da177e4 LT |
2746 | /* Use a previous sequence. This should cause the other |
2747 | * end to send an ack. Don't queue or clone SKB, just | |
2748 | * send it. | |
2749 | */ | |
a3433f35 | 2750 | tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); |
1da177e4 | 2751 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
dfb4b9dc | 2752 | return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); |
1da177e4 LT |
2753 | } |
2754 | ||
67edfef7 | 2755 | /* Initiate keepalive or window probe from timer. */ |
1da177e4 LT |
2756 | int tcp_write_wakeup(struct sock *sk) |
2757 | { | |
058dc334 IJ |
2758 | struct tcp_sock *tp = tcp_sk(sk); |
2759 | struct sk_buff *skb; | |
1da177e4 | 2760 | |
058dc334 IJ |
2761 | if (sk->sk_state == TCP_CLOSE) |
2762 | return -1; | |
2763 | ||
2764 | if ((skb = tcp_send_head(sk)) != NULL && | |
2765 | before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { | |
2766 | int err; | |
0c54b85f | 2767 | unsigned int mss = tcp_current_mss(sk); |
058dc334 IJ |
2768 | unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
2769 | ||
2770 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) | |
2771 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; | |
2772 | ||
2773 | /* We are probing the opening of a window | |
2774 | * but the window size is != 0 | |
2775 | * must have been a result SWS avoidance ( sender ) | |
2776 | */ | |
2777 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || | |
2778 | skb->len > mss) { | |
2779 | seg_size = min(seg_size, mss); | |
a3433f35 | 2780 | TCP_SKB_CB(skb)->flags |= TCPHDR_PSH; |
058dc334 IJ |
2781 | if (tcp_fragment(sk, skb, seg_size, mss)) |
2782 | return -1; | |
2783 | } else if (!tcp_skb_pcount(skb)) | |
2784 | tcp_set_skb_tso_segs(sk, skb, mss); | |
2785 | ||
a3433f35 | 2786 | TCP_SKB_CB(skb)->flags |= TCPHDR_PSH; |
058dc334 IJ |
2787 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
2788 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); | |
2789 | if (!err) | |
2790 | tcp_event_new_data_sent(sk, skb); | |
2791 | return err; | |
2792 | } else { | |
33f5f57e | 2793 | if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) |
058dc334 IJ |
2794 | tcp_xmit_probe_skb(sk, 1); |
2795 | return tcp_xmit_probe_skb(sk, 0); | |
1da177e4 | 2796 | } |
1da177e4 LT |
2797 | } |
2798 | ||
2799 | /* A window probe timeout has occurred. If window is not closed send | |
2800 | * a partial packet else a zero probe. | |
2801 | */ | |
2802 | void tcp_send_probe0(struct sock *sk) | |
2803 | { | |
463c84b9 | 2804 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2805 | struct tcp_sock *tp = tcp_sk(sk); |
2806 | int err; | |
2807 | ||
2808 | err = tcp_write_wakeup(sk); | |
2809 | ||
fe067e8a | 2810 | if (tp->packets_out || !tcp_send_head(sk)) { |
1da177e4 | 2811 | /* Cancel probe timer, if it is not required. */ |
6687e988 | 2812 | icsk->icsk_probes_out = 0; |
463c84b9 | 2813 | icsk->icsk_backoff = 0; |
1da177e4 LT |
2814 | return; |
2815 | } | |
2816 | ||
2817 | if (err <= 0) { | |
463c84b9 ACM |
2818 | if (icsk->icsk_backoff < sysctl_tcp_retries2) |
2819 | icsk->icsk_backoff++; | |
6687e988 | 2820 | icsk->icsk_probes_out++; |
e905a9ed | 2821 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
3f421baa ACM |
2822 | min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), |
2823 | TCP_RTO_MAX); | |
1da177e4 LT |
2824 | } else { |
2825 | /* If packet was not sent due to local congestion, | |
6687e988 | 2826 | * do not backoff and do not remember icsk_probes_out. |
1da177e4 LT |
2827 | * Let local senders to fight for local resources. |
2828 | * | |
2829 | * Use accumulated backoff yet. | |
2830 | */ | |
6687e988 ACM |
2831 | if (!icsk->icsk_probes_out) |
2832 | icsk->icsk_probes_out = 1; | |
e905a9ed | 2833 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
463c84b9 | 2834 | min(icsk->icsk_rto << icsk->icsk_backoff, |
3f421baa ACM |
2835 | TCP_RESOURCE_PROBE_INTERVAL), |
2836 | TCP_RTO_MAX); | |
1da177e4 LT |
2837 | } |
2838 | } |