]>
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 | ||
91df42be JP |
37 | #define pr_fmt(fmt) "TCP: " fmt |
38 | ||
1da177e4 LT |
39 | #include <net/tcp.h> |
40 | ||
41 | #include <linux/compiler.h> | |
5a0e3ad6 | 42 | #include <linux/gfp.h> |
1da177e4 | 43 | #include <linux/module.h> |
1da177e4 LT |
44 | |
45 | /* People can turn this off for buggy TCP's found in printers etc. */ | |
ab32ea5d | 46 | int sysctl_tcp_retrans_collapse __read_mostly = 1; |
1da177e4 | 47 | |
09cb105e | 48 | /* People can turn this on to work with those rare, broken TCPs that |
15d99e02 RJ |
49 | * interpret the window field as a signed quantity. |
50 | */ | |
ab32ea5d | 51 | int sysctl_tcp_workaround_signed_windows __read_mostly = 0; |
15d99e02 | 52 | |
1da177e4 LT |
53 | /* This limits the percentage of the congestion window which we |
54 | * will allow a single TSO frame to consume. Building TSO frames | |
55 | * which are too large can cause TCP streams to be bursty. | |
56 | */ | |
ab32ea5d | 57 | int sysctl_tcp_tso_win_divisor __read_mostly = 3; |
1da177e4 | 58 | |
ab32ea5d | 59 | int sysctl_tcp_mtu_probing __read_mostly = 0; |
97b1ce25 | 60 | int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS; |
5d424d5a | 61 | |
35089bb2 | 62 | /* By default, RFC2861 behavior. */ |
ab32ea5d | 63 | int sysctl_tcp_slow_start_after_idle __read_mostly = 1; |
35089bb2 | 64 | |
519855c5 | 65 | int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */ |
e6b09cca | 66 | EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size); |
519855c5 WAS |
67 | |
68 | ||
67edfef7 | 69 | /* Account for new data that has been sent to the network. */ |
cf533ea5 | 70 | static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 71 | { |
9e412ba7 | 72 | struct tcp_sock *tp = tcp_sk(sk); |
66f5fe62 | 73 | unsigned int prior_packets = tp->packets_out; |
9e412ba7 | 74 | |
fe067e8a | 75 | tcp_advance_send_head(sk, skb); |
1da177e4 | 76 | tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; |
8512430e | 77 | |
25985edc | 78 | /* Don't override Nagle indefinitely with F-RTO */ |
8512430e IJ |
79 | if (tp->frto_counter == 2) |
80 | tp->frto_counter = 3; | |
66f5fe62 IJ |
81 | |
82 | tp->packets_out += tcp_skb_pcount(skb); | |
750ea2ba YC |
83 | if (!prior_packets || tp->early_retrans_delayed) |
84 | tcp_rearm_rto(sk); | |
1da177e4 LT |
85 | } |
86 | ||
87 | /* SND.NXT, if window was not shrunk. | |
88 | * If window has been shrunk, what should we make? It is not clear at all. | |
89 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( | |
90 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already | |
91 | * invalid. OK, let's make this for now: | |
92 | */ | |
cf533ea5 | 93 | static inline __u32 tcp_acceptable_seq(const struct sock *sk) |
1da177e4 | 94 | { |
cf533ea5 | 95 | const struct tcp_sock *tp = tcp_sk(sk); |
9e412ba7 | 96 | |
90840def | 97 | if (!before(tcp_wnd_end(tp), tp->snd_nxt)) |
1da177e4 LT |
98 | return tp->snd_nxt; |
99 | else | |
90840def | 100 | return tcp_wnd_end(tp); |
1da177e4 LT |
101 | } |
102 | ||
103 | /* Calculate mss to advertise in SYN segment. | |
104 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: | |
105 | * | |
106 | * 1. It is independent of path mtu. | |
107 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. | |
108 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of | |
109 | * attached devices, because some buggy hosts are confused by | |
110 | * large MSS. | |
111 | * 4. We do not make 3, we advertise MSS, calculated from first | |
112 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. | |
113 | * This may be overridden via information stored in routing table. | |
114 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, | |
115 | * probably even Jumbo". | |
116 | */ | |
117 | static __u16 tcp_advertise_mss(struct sock *sk) | |
118 | { | |
119 | struct tcp_sock *tp = tcp_sk(sk); | |
cf533ea5 | 120 | const struct dst_entry *dst = __sk_dst_get(sk); |
1da177e4 LT |
121 | int mss = tp->advmss; |
122 | ||
0dbaee3b DM |
123 | if (dst) { |
124 | unsigned int metric = dst_metric_advmss(dst); | |
125 | ||
126 | if (metric < mss) { | |
127 | mss = metric; | |
128 | tp->advmss = mss; | |
129 | } | |
1da177e4 LT |
130 | } |
131 | ||
132 | return (__u16)mss; | |
133 | } | |
134 | ||
135 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". | |
136 | * This is the first part of cwnd validation mechanism. */ | |
cf533ea5 | 137 | static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst) |
1da177e4 | 138 | { |
463c84b9 | 139 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
140 | s32 delta = tcp_time_stamp - tp->lsndtime; |
141 | u32 restart_cwnd = tcp_init_cwnd(tp, dst); | |
142 | u32 cwnd = tp->snd_cwnd; | |
143 | ||
6687e988 | 144 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); |
1da177e4 | 145 | |
6687e988 | 146 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
1da177e4 LT |
147 | restart_cwnd = min(restart_cwnd, cwnd); |
148 | ||
463c84b9 | 149 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) |
1da177e4 LT |
150 | cwnd >>= 1; |
151 | tp->snd_cwnd = max(cwnd, restart_cwnd); | |
152 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
153 | tp->snd_cwnd_used = 0; | |
154 | } | |
155 | ||
67edfef7 | 156 | /* Congestion state accounting after a packet has been sent. */ |
40efc6fa | 157 | static void tcp_event_data_sent(struct tcp_sock *tp, |
cf533ea5 | 158 | struct sock *sk) |
1da177e4 | 159 | { |
463c84b9 ACM |
160 | struct inet_connection_sock *icsk = inet_csk(sk); |
161 | const u32 now = tcp_time_stamp; | |
1da177e4 | 162 | |
35089bb2 DM |
163 | if (sysctl_tcp_slow_start_after_idle && |
164 | (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) | |
463c84b9 | 165 | tcp_cwnd_restart(sk, __sk_dst_get(sk)); |
1da177e4 LT |
166 | |
167 | tp->lsndtime = now; | |
168 | ||
169 | /* If it is a reply for ato after last received | |
170 | * packet, enter pingpong mode. | |
171 | */ | |
463c84b9 ACM |
172 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) |
173 | icsk->icsk_ack.pingpong = 1; | |
1da177e4 LT |
174 | } |
175 | ||
67edfef7 | 176 | /* Account for an ACK we sent. */ |
40efc6fa | 177 | static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) |
1da177e4 | 178 | { |
463c84b9 ACM |
179 | tcp_dec_quickack_mode(sk, pkts); |
180 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); | |
1da177e4 LT |
181 | } |
182 | ||
183 | /* Determine a window scaling and initial window to offer. | |
184 | * Based on the assumption that the given amount of space | |
185 | * will be offered. Store the results in the tp structure. | |
186 | * NOTE: for smooth operation initial space offering should | |
187 | * be a multiple of mss if possible. We assume here that mss >= 1. | |
188 | * This MUST be enforced by all callers. | |
189 | */ | |
190 | void tcp_select_initial_window(int __space, __u32 mss, | |
191 | __u32 *rcv_wnd, __u32 *window_clamp, | |
31d12926 | 192 | int wscale_ok, __u8 *rcv_wscale, |
193 | __u32 init_rcv_wnd) | |
1da177e4 LT |
194 | { |
195 | unsigned int space = (__space < 0 ? 0 : __space); | |
196 | ||
197 | /* If no clamp set the clamp to the max possible scaled window */ | |
198 | if (*window_clamp == 0) | |
199 | (*window_clamp) = (65535 << 14); | |
200 | space = min(*window_clamp, space); | |
201 | ||
202 | /* Quantize space offering to a multiple of mss if possible. */ | |
203 | if (space > mss) | |
204 | space = (space / mss) * mss; | |
205 | ||
206 | /* NOTE: offering an initial window larger than 32767 | |
15d99e02 RJ |
207 | * will break some buggy TCP stacks. If the admin tells us |
208 | * it is likely we could be speaking with such a buggy stack | |
209 | * we will truncate our initial window offering to 32K-1 | |
210 | * unless the remote has sent us a window scaling option, | |
211 | * which we interpret as a sign the remote TCP is not | |
212 | * misinterpreting the window field as a signed quantity. | |
1da177e4 | 213 | */ |
15d99e02 RJ |
214 | if (sysctl_tcp_workaround_signed_windows) |
215 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); | |
216 | else | |
217 | (*rcv_wnd) = space; | |
218 | ||
1da177e4 LT |
219 | (*rcv_wscale) = 0; |
220 | if (wscale_ok) { | |
221 | /* Set window scaling on max possible window | |
e905a9ed | 222 | * See RFC1323 for an explanation of the limit to 14 |
1da177e4 LT |
223 | */ |
224 | space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); | |
316c1592 | 225 | space = min_t(u32, space, *window_clamp); |
1da177e4 LT |
226 | while (space > 65535 && (*rcv_wscale) < 14) { |
227 | space >>= 1; | |
228 | (*rcv_wscale)++; | |
229 | } | |
230 | } | |
231 | ||
356f0398 ND |
232 | /* Set initial window to a value enough for senders starting with |
233 | * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place | |
234 | * a limit on the initial window when mss is larger than 1460. | |
235 | */ | |
056834d9 | 236 | if (mss > (1 << *rcv_wscale)) { |
356f0398 ND |
237 | int init_cwnd = TCP_DEFAULT_INIT_RCVWND; |
238 | if (mss > 1460) | |
239 | init_cwnd = | |
240 | max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2); | |
31d12926 | 241 | /* when initializing use the value from init_rcv_wnd |
242 | * rather than the default from above | |
243 | */ | |
b1afde60 ND |
244 | if (init_rcv_wnd) |
245 | *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); | |
246 | else | |
247 | *rcv_wnd = min(*rcv_wnd, init_cwnd * mss); | |
1da177e4 LT |
248 | } |
249 | ||
250 | /* Set the clamp no higher than max representable value */ | |
251 | (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); | |
252 | } | |
4bc2f18b | 253 | EXPORT_SYMBOL(tcp_select_initial_window); |
1da177e4 LT |
254 | |
255 | /* Chose a new window to advertise, update state in tcp_sock for the | |
256 | * socket, and return result with RFC1323 scaling applied. The return | |
257 | * value can be stuffed directly into th->window for an outgoing | |
258 | * frame. | |
259 | */ | |
40efc6fa | 260 | static u16 tcp_select_window(struct sock *sk) |
1da177e4 LT |
261 | { |
262 | struct tcp_sock *tp = tcp_sk(sk); | |
263 | u32 cur_win = tcp_receive_window(tp); | |
264 | u32 new_win = __tcp_select_window(sk); | |
265 | ||
266 | /* Never shrink the offered window */ | |
2de979bd | 267 | if (new_win < cur_win) { |
1da177e4 LT |
268 | /* Danger Will Robinson! |
269 | * Don't update rcv_wup/rcv_wnd here or else | |
270 | * we will not be able to advertise a zero | |
271 | * window in time. --DaveM | |
272 | * | |
273 | * Relax Will Robinson. | |
274 | */ | |
607bfbf2 | 275 | new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); |
1da177e4 LT |
276 | } |
277 | tp->rcv_wnd = new_win; | |
278 | tp->rcv_wup = tp->rcv_nxt; | |
279 | ||
280 | /* Make sure we do not exceed the maximum possible | |
281 | * scaled window. | |
282 | */ | |
15d99e02 | 283 | if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) |
1da177e4 LT |
284 | new_win = min(new_win, MAX_TCP_WINDOW); |
285 | else | |
286 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); | |
287 | ||
288 | /* RFC1323 scaling applied */ | |
289 | new_win >>= tp->rx_opt.rcv_wscale; | |
290 | ||
291 | /* If we advertise zero window, disable fast path. */ | |
292 | if (new_win == 0) | |
293 | tp->pred_flags = 0; | |
294 | ||
295 | return new_win; | |
296 | } | |
297 | ||
67edfef7 | 298 | /* Packet ECN state for a SYN-ACK */ |
cf533ea5 | 299 | static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb) |
bdf1ee5d | 300 | { |
4de075e0 | 301 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; |
056834d9 | 302 | if (!(tp->ecn_flags & TCP_ECN_OK)) |
4de075e0 | 303 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; |
bdf1ee5d IJ |
304 | } |
305 | ||
67edfef7 | 306 | /* Packet ECN state for a SYN. */ |
bdf1ee5d IJ |
307 | static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) |
308 | { | |
309 | struct tcp_sock *tp = tcp_sk(sk); | |
310 | ||
311 | tp->ecn_flags = 0; | |
255cac91 | 312 | if (sysctl_tcp_ecn == 1) { |
4de075e0 | 313 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; |
bdf1ee5d IJ |
314 | tp->ecn_flags = TCP_ECN_OK; |
315 | } | |
316 | } | |
317 | ||
318 | static __inline__ void | |
cf533ea5 | 319 | TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th) |
bdf1ee5d IJ |
320 | { |
321 | if (inet_rsk(req)->ecn_ok) | |
322 | th->ece = 1; | |
323 | } | |
324 | ||
67edfef7 AK |
325 | /* Set up ECN state for a packet on a ESTABLISHED socket that is about to |
326 | * be sent. | |
327 | */ | |
bdf1ee5d IJ |
328 | static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, |
329 | int tcp_header_len) | |
330 | { | |
331 | struct tcp_sock *tp = tcp_sk(sk); | |
332 | ||
333 | if (tp->ecn_flags & TCP_ECN_OK) { | |
334 | /* Not-retransmitted data segment: set ECT and inject CWR. */ | |
335 | if (skb->len != tcp_header_len && | |
336 | !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { | |
337 | INET_ECN_xmit(sk); | |
056834d9 | 338 | if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { |
bdf1ee5d IJ |
339 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; |
340 | tcp_hdr(skb)->cwr = 1; | |
341 | skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; | |
342 | } | |
343 | } else { | |
344 | /* ACK or retransmitted segment: clear ECT|CE */ | |
345 | INET_ECN_dontxmit(sk); | |
346 | } | |
347 | if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) | |
348 | tcp_hdr(skb)->ece = 1; | |
349 | } | |
350 | } | |
351 | ||
e870a8ef IJ |
352 | /* Constructs common control bits of non-data skb. If SYN/FIN is present, |
353 | * auto increment end seqno. | |
354 | */ | |
355 | static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) | |
356 | { | |
2e8e18ef | 357 | skb->ip_summed = CHECKSUM_PARTIAL; |
e870a8ef IJ |
358 | skb->csum = 0; |
359 | ||
4de075e0 | 360 | TCP_SKB_CB(skb)->tcp_flags = flags; |
e870a8ef IJ |
361 | TCP_SKB_CB(skb)->sacked = 0; |
362 | ||
363 | skb_shinfo(skb)->gso_segs = 1; | |
364 | skb_shinfo(skb)->gso_size = 0; | |
365 | skb_shinfo(skb)->gso_type = 0; | |
366 | ||
367 | TCP_SKB_CB(skb)->seq = seq; | |
a3433f35 | 368 | if (flags & (TCPHDR_SYN | TCPHDR_FIN)) |
e870a8ef IJ |
369 | seq++; |
370 | TCP_SKB_CB(skb)->end_seq = seq; | |
371 | } | |
372 | ||
a2a385d6 | 373 | static inline bool tcp_urg_mode(const struct tcp_sock *tp) |
33f5f57e IJ |
374 | { |
375 | return tp->snd_una != tp->snd_up; | |
376 | } | |
377 | ||
33ad798c AL |
378 | #define OPTION_SACK_ADVERTISE (1 << 0) |
379 | #define OPTION_TS (1 << 1) | |
380 | #define OPTION_MD5 (1 << 2) | |
89e95a61 | 381 | #define OPTION_WSCALE (1 << 3) |
bd0388ae | 382 | #define OPTION_COOKIE_EXTENSION (1 << 4) |
33ad798c AL |
383 | |
384 | struct tcp_out_options { | |
385 | u8 options; /* bit field of OPTION_* */ | |
386 | u8 ws; /* window scale, 0 to disable */ | |
387 | u8 num_sack_blocks; /* number of SACK blocks to include */ | |
bd0388ae | 388 | u8 hash_size; /* bytes in hash_location */ |
33ad798c AL |
389 | u16 mss; /* 0 to disable */ |
390 | __u32 tsval, tsecr; /* need to include OPTION_TS */ | |
bd0388ae | 391 | __u8 *hash_location; /* temporary pointer, overloaded */ |
33ad798c AL |
392 | }; |
393 | ||
bd0388ae WAS |
394 | /* The sysctl int routines are generic, so check consistency here. |
395 | */ | |
396 | static u8 tcp_cookie_size_check(u8 desired) | |
397 | { | |
f1987257 ED |
398 | int cookie_size; |
399 | ||
400 | if (desired > 0) | |
bd0388ae WAS |
401 | /* previously specified */ |
402 | return desired; | |
f1987257 ED |
403 | |
404 | cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size); | |
405 | if (cookie_size <= 0) | |
bd0388ae WAS |
406 | /* no default specified */ |
407 | return 0; | |
f1987257 ED |
408 | |
409 | if (cookie_size <= TCP_COOKIE_MIN) | |
bd0388ae WAS |
410 | /* value too small, specify minimum */ |
411 | return TCP_COOKIE_MIN; | |
f1987257 ED |
412 | |
413 | if (cookie_size >= TCP_COOKIE_MAX) | |
bd0388ae WAS |
414 | /* value too large, specify maximum */ |
415 | return TCP_COOKIE_MAX; | |
f1987257 ED |
416 | |
417 | if (cookie_size & 1) | |
bd0388ae | 418 | /* 8-bit multiple, illegal, fix it */ |
f1987257 ED |
419 | cookie_size++; |
420 | ||
421 | return (u8)cookie_size; | |
bd0388ae WAS |
422 | } |
423 | ||
67edfef7 AK |
424 | /* Write previously computed TCP options to the packet. |
425 | * | |
426 | * Beware: Something in the Internet is very sensitive to the ordering of | |
fd6149d3 IJ |
427 | * TCP options, we learned this through the hard way, so be careful here. |
428 | * Luckily we can at least blame others for their non-compliance but from | |
429 | * inter-operatibility perspective it seems that we're somewhat stuck with | |
430 | * the ordering which we have been using if we want to keep working with | |
431 | * those broken things (not that it currently hurts anybody as there isn't | |
432 | * particular reason why the ordering would need to be changed). | |
433 | * | |
434 | * At least SACK_PERM as the first option is known to lead to a disaster | |
435 | * (but it may well be that other scenarios fail similarly). | |
436 | */ | |
33ad798c | 437 | static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, |
bd0388ae WAS |
438 | struct tcp_out_options *opts) |
439 | { | |
440 | u8 options = opts->options; /* mungable copy */ | |
441 | ||
442 | /* Having both authentication and cookies for security is redundant, | |
443 | * and there's certainly not enough room. Instead, the cookie-less | |
444 | * extension variant is proposed. | |
445 | * | |
446 | * Consider the pessimal case with authentication. The options | |
447 | * could look like: | |
448 | * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40 | |
449 | */ | |
450 | if (unlikely(OPTION_MD5 & options)) { | |
451 | if (unlikely(OPTION_COOKIE_EXTENSION & options)) { | |
452 | *ptr++ = htonl((TCPOPT_COOKIE << 24) | | |
453 | (TCPOLEN_COOKIE_BASE << 16) | | |
454 | (TCPOPT_MD5SIG << 8) | | |
455 | TCPOLEN_MD5SIG); | |
456 | } else { | |
457 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
458 | (TCPOPT_NOP << 16) | | |
459 | (TCPOPT_MD5SIG << 8) | | |
460 | TCPOLEN_MD5SIG); | |
461 | } | |
462 | options &= ~OPTION_COOKIE_EXTENSION; | |
463 | /* overload cookie hash location */ | |
464 | opts->hash_location = (__u8 *)ptr; | |
33ad798c | 465 | ptr += 4; |
40efc6fa | 466 | } |
33ad798c | 467 | |
fd6149d3 IJ |
468 | if (unlikely(opts->mss)) { |
469 | *ptr++ = htonl((TCPOPT_MSS << 24) | | |
470 | (TCPOLEN_MSS << 16) | | |
471 | opts->mss); | |
472 | } | |
473 | ||
bd0388ae WAS |
474 | if (likely(OPTION_TS & options)) { |
475 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { | |
33ad798c AL |
476 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
477 | (TCPOLEN_SACK_PERM << 16) | | |
478 | (TCPOPT_TIMESTAMP << 8) | | |
479 | TCPOLEN_TIMESTAMP); | |
bd0388ae | 480 | options &= ~OPTION_SACK_ADVERTISE; |
33ad798c AL |
481 | } else { |
482 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
483 | (TCPOPT_NOP << 16) | | |
484 | (TCPOPT_TIMESTAMP << 8) | | |
485 | TCPOLEN_TIMESTAMP); | |
486 | } | |
487 | *ptr++ = htonl(opts->tsval); | |
488 | *ptr++ = htonl(opts->tsecr); | |
489 | } | |
490 | ||
bd0388ae WAS |
491 | /* Specification requires after timestamp, so do it now. |
492 | * | |
493 | * Consider the pessimal case without authentication. The options | |
494 | * could look like: | |
495 | * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40 | |
496 | */ | |
497 | if (unlikely(OPTION_COOKIE_EXTENSION & options)) { | |
498 | __u8 *cookie_copy = opts->hash_location; | |
499 | u8 cookie_size = opts->hash_size; | |
500 | ||
501 | /* 8-bit multiple handled in tcp_cookie_size_check() above, | |
502 | * and elsewhere. | |
503 | */ | |
504 | if (0x2 & cookie_size) { | |
505 | __u8 *p = (__u8 *)ptr; | |
506 | ||
507 | /* 16-bit multiple */ | |
508 | *p++ = TCPOPT_COOKIE; | |
509 | *p++ = TCPOLEN_COOKIE_BASE + cookie_size; | |
510 | *p++ = *cookie_copy++; | |
511 | *p++ = *cookie_copy++; | |
512 | ptr++; | |
513 | cookie_size -= 2; | |
514 | } else { | |
515 | /* 32-bit multiple */ | |
516 | *ptr++ = htonl(((TCPOPT_NOP << 24) | | |
517 | (TCPOPT_NOP << 16) | | |
518 | (TCPOPT_COOKIE << 8) | | |
519 | TCPOLEN_COOKIE_BASE) + | |
520 | cookie_size); | |
521 | } | |
522 | ||
523 | if (cookie_size > 0) { | |
524 | memcpy(ptr, cookie_copy, cookie_size); | |
525 | ptr += (cookie_size / 4); | |
526 | } | |
527 | } | |
528 | ||
529 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { | |
33ad798c AL |
530 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
531 | (TCPOPT_NOP << 16) | | |
532 | (TCPOPT_SACK_PERM << 8) | | |
533 | TCPOLEN_SACK_PERM); | |
534 | } | |
535 | ||
bd0388ae | 536 | if (unlikely(OPTION_WSCALE & options)) { |
33ad798c AL |
537 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
538 | (TCPOPT_WINDOW << 16) | | |
539 | (TCPOLEN_WINDOW << 8) | | |
540 | opts->ws); | |
541 | } | |
542 | ||
543 | if (unlikely(opts->num_sack_blocks)) { | |
544 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? | |
545 | tp->duplicate_sack : tp->selective_acks; | |
40efc6fa SH |
546 | int this_sack; |
547 | ||
548 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
549 | (TCPOPT_NOP << 16) | | |
550 | (TCPOPT_SACK << 8) | | |
33ad798c | 551 | (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * |
40efc6fa | 552 | TCPOLEN_SACK_PERBLOCK))); |
2de979bd | 553 | |
33ad798c AL |
554 | for (this_sack = 0; this_sack < opts->num_sack_blocks; |
555 | ++this_sack) { | |
40efc6fa SH |
556 | *ptr++ = htonl(sp[this_sack].start_seq); |
557 | *ptr++ = htonl(sp[this_sack].end_seq); | |
558 | } | |
2de979bd | 559 | |
5861f8e5 | 560 | tp->rx_opt.dsack = 0; |
40efc6fa | 561 | } |
33ad798c AL |
562 | } |
563 | ||
67edfef7 AK |
564 | /* Compute TCP options for SYN packets. This is not the final |
565 | * network wire format yet. | |
566 | */ | |
95c96174 | 567 | static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, |
33ad798c | 568 | struct tcp_out_options *opts, |
cf533ea5 ED |
569 | struct tcp_md5sig_key **md5) |
570 | { | |
33ad798c | 571 | struct tcp_sock *tp = tcp_sk(sk); |
bd0388ae | 572 | struct tcp_cookie_values *cvp = tp->cookie_values; |
95c96174 | 573 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
bd0388ae WAS |
574 | u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ? |
575 | tcp_cookie_size_check(cvp->cookie_desired) : | |
576 | 0; | |
33ad798c | 577 | |
cfb6eeb4 | 578 | #ifdef CONFIG_TCP_MD5SIG |
33ad798c AL |
579 | *md5 = tp->af_specific->md5_lookup(sk, sk); |
580 | if (*md5) { | |
581 | opts->options |= OPTION_MD5; | |
bd0388ae | 582 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
cfb6eeb4 | 583 | } |
33ad798c AL |
584 | #else |
585 | *md5 = NULL; | |
cfb6eeb4 | 586 | #endif |
33ad798c AL |
587 | |
588 | /* We always get an MSS option. The option bytes which will be seen in | |
589 | * normal data packets should timestamps be used, must be in the MSS | |
590 | * advertised. But we subtract them from tp->mss_cache so that | |
591 | * calculations in tcp_sendmsg are simpler etc. So account for this | |
592 | * fact here if necessary. If we don't do this correctly, as a | |
593 | * receiver we won't recognize data packets as being full sized when we | |
594 | * should, and thus we won't abide by the delayed ACK rules correctly. | |
595 | * SACKs don't matter, we never delay an ACK when we have any of those | |
596 | * going out. */ | |
597 | opts->mss = tcp_advertise_mss(sk); | |
bd0388ae | 598 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c | 599 | |
bb5b7c11 | 600 | if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { |
33ad798c AL |
601 | opts->options |= OPTION_TS; |
602 | opts->tsval = TCP_SKB_CB(skb)->when; | |
603 | opts->tsecr = tp->rx_opt.ts_recent; | |
bd0388ae | 604 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c | 605 | } |
bb5b7c11 | 606 | if (likely(sysctl_tcp_window_scaling)) { |
33ad798c | 607 | opts->ws = tp->rx_opt.rcv_wscale; |
89e95a61 | 608 | opts->options |= OPTION_WSCALE; |
bd0388ae | 609 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 610 | } |
bb5b7c11 | 611 | if (likely(sysctl_tcp_sack)) { |
33ad798c | 612 | opts->options |= OPTION_SACK_ADVERTISE; |
b32d1310 | 613 | if (unlikely(!(OPTION_TS & opts->options))) |
bd0388ae | 614 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c AL |
615 | } |
616 | ||
bd0388ae WAS |
617 | /* Note that timestamps are required by the specification. |
618 | * | |
619 | * Odd numbers of bytes are prohibited by the specification, ensuring | |
620 | * that the cookie is 16-bit aligned, and the resulting cookie pair is | |
621 | * 32-bit aligned. | |
622 | */ | |
623 | if (*md5 == NULL && | |
624 | (OPTION_TS & opts->options) && | |
625 | cookie_size > 0) { | |
626 | int need = TCPOLEN_COOKIE_BASE + cookie_size; | |
627 | ||
628 | if (0x2 & need) { | |
629 | /* 32-bit multiple */ | |
630 | need += 2; /* NOPs */ | |
631 | ||
632 | if (need > remaining) { | |
633 | /* try shrinking cookie to fit */ | |
634 | cookie_size -= 2; | |
635 | need -= 4; | |
636 | } | |
637 | } | |
638 | while (need > remaining && TCP_COOKIE_MIN <= cookie_size) { | |
639 | cookie_size -= 4; | |
640 | need -= 4; | |
641 | } | |
642 | if (TCP_COOKIE_MIN <= cookie_size) { | |
643 | opts->options |= OPTION_COOKIE_EXTENSION; | |
644 | opts->hash_location = (__u8 *)&cvp->cookie_pair[0]; | |
645 | opts->hash_size = cookie_size; | |
646 | ||
647 | /* Remember for future incarnations. */ | |
648 | cvp->cookie_desired = cookie_size; | |
649 | ||
650 | if (cvp->cookie_desired != cvp->cookie_pair_size) { | |
651 | /* Currently use random bytes as a nonce, | |
652 | * assuming these are completely unpredictable | |
653 | * by hostile users of the same system. | |
654 | */ | |
655 | get_random_bytes(&cvp->cookie_pair[0], | |
656 | cookie_size); | |
657 | cvp->cookie_pair_size = cookie_size; | |
658 | } | |
659 | ||
660 | remaining -= need; | |
661 | } | |
662 | } | |
663 | return MAX_TCP_OPTION_SPACE - remaining; | |
40efc6fa SH |
664 | } |
665 | ||
67edfef7 | 666 | /* Set up TCP options for SYN-ACKs. */ |
95c96174 | 667 | static unsigned int tcp_synack_options(struct sock *sk, |
33ad798c | 668 | struct request_sock *req, |
95c96174 | 669 | unsigned int mss, struct sk_buff *skb, |
33ad798c | 670 | struct tcp_out_options *opts, |
4957faad WAS |
671 | struct tcp_md5sig_key **md5, |
672 | struct tcp_extend_values *xvp) | |
673 | { | |
33ad798c | 674 | struct inet_request_sock *ireq = inet_rsk(req); |
95c96174 | 675 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
4957faad WAS |
676 | u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ? |
677 | xvp->cookie_plus : | |
678 | 0; | |
33ad798c | 679 | |
cfb6eeb4 | 680 | #ifdef CONFIG_TCP_MD5SIG |
33ad798c AL |
681 | *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); |
682 | if (*md5) { | |
683 | opts->options |= OPTION_MD5; | |
4957faad WAS |
684 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
685 | ||
686 | /* We can't fit any SACK blocks in a packet with MD5 + TS | |
687 | * options. There was discussion about disabling SACK | |
688 | * rather than TS in order to fit in better with old, | |
689 | * buggy kernels, but that was deemed to be unnecessary. | |
690 | */ | |
de213e5e | 691 | ireq->tstamp_ok &= !ireq->sack_ok; |
cfb6eeb4 | 692 | } |
33ad798c AL |
693 | #else |
694 | *md5 = NULL; | |
cfb6eeb4 | 695 | #endif |
33ad798c | 696 | |
4957faad | 697 | /* We always send an MSS option. */ |
33ad798c | 698 | opts->mss = mss; |
4957faad | 699 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c AL |
700 | |
701 | if (likely(ireq->wscale_ok)) { | |
702 | opts->ws = ireq->rcv_wscale; | |
89e95a61 | 703 | opts->options |= OPTION_WSCALE; |
4957faad | 704 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 705 | } |
de213e5e | 706 | if (likely(ireq->tstamp_ok)) { |
33ad798c AL |
707 | opts->options |= OPTION_TS; |
708 | opts->tsval = TCP_SKB_CB(skb)->when; | |
709 | opts->tsecr = req->ts_recent; | |
4957faad | 710 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c AL |
711 | } |
712 | if (likely(ireq->sack_ok)) { | |
713 | opts->options |= OPTION_SACK_ADVERTISE; | |
de213e5e | 714 | if (unlikely(!ireq->tstamp_ok)) |
4957faad | 715 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c AL |
716 | } |
717 | ||
4957faad WAS |
718 | /* Similar rationale to tcp_syn_options() applies here, too. |
719 | * If the <SYN> options fit, the same options should fit now! | |
720 | */ | |
721 | if (*md5 == NULL && | |
de213e5e | 722 | ireq->tstamp_ok && |
4957faad WAS |
723 | cookie_plus > TCPOLEN_COOKIE_BASE) { |
724 | int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */ | |
725 | ||
726 | if (0x2 & need) { | |
727 | /* 32-bit multiple */ | |
728 | need += 2; /* NOPs */ | |
729 | } | |
730 | if (need <= remaining) { | |
731 | opts->options |= OPTION_COOKIE_EXTENSION; | |
732 | opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE; | |
733 | remaining -= need; | |
734 | } else { | |
735 | /* There's no error return, so flag it. */ | |
736 | xvp->cookie_out_never = 1; /* true */ | |
737 | opts->hash_size = 0; | |
738 | } | |
739 | } | |
740 | return MAX_TCP_OPTION_SPACE - remaining; | |
33ad798c AL |
741 | } |
742 | ||
67edfef7 AK |
743 | /* Compute TCP options for ESTABLISHED sockets. This is not the |
744 | * final wire format yet. | |
745 | */ | |
95c96174 | 746 | static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, |
33ad798c | 747 | struct tcp_out_options *opts, |
cf533ea5 ED |
748 | struct tcp_md5sig_key **md5) |
749 | { | |
33ad798c AL |
750 | struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; |
751 | struct tcp_sock *tp = tcp_sk(sk); | |
95c96174 | 752 | unsigned int size = 0; |
cabeccbd | 753 | unsigned int eff_sacks; |
33ad798c AL |
754 | |
755 | #ifdef CONFIG_TCP_MD5SIG | |
756 | *md5 = tp->af_specific->md5_lookup(sk, sk); | |
757 | if (unlikely(*md5)) { | |
758 | opts->options |= OPTION_MD5; | |
759 | size += TCPOLEN_MD5SIG_ALIGNED; | |
760 | } | |
761 | #else | |
762 | *md5 = NULL; | |
763 | #endif | |
764 | ||
765 | if (likely(tp->rx_opt.tstamp_ok)) { | |
766 | opts->options |= OPTION_TS; | |
767 | opts->tsval = tcb ? tcb->when : 0; | |
768 | opts->tsecr = tp->rx_opt.ts_recent; | |
769 | size += TCPOLEN_TSTAMP_ALIGNED; | |
770 | } | |
771 | ||
cabeccbd IJ |
772 | eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; |
773 | if (unlikely(eff_sacks)) { | |
95c96174 | 774 | const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; |
33ad798c | 775 | opts->num_sack_blocks = |
95c96174 | 776 | min_t(unsigned int, eff_sacks, |
33ad798c AL |
777 | (remaining - TCPOLEN_SACK_BASE_ALIGNED) / |
778 | TCPOLEN_SACK_PERBLOCK); | |
779 | size += TCPOLEN_SACK_BASE_ALIGNED + | |
780 | opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; | |
781 | } | |
782 | ||
783 | return size; | |
40efc6fa | 784 | } |
1da177e4 LT |
785 | |
786 | /* This routine actually transmits TCP packets queued in by | |
787 | * tcp_do_sendmsg(). This is used by both the initial | |
788 | * transmission and possible later retransmissions. | |
789 | * All SKB's seen here are completely headerless. It is our | |
790 | * job to build the TCP header, and pass the packet down to | |
791 | * IP so it can do the same plus pass the packet off to the | |
792 | * device. | |
793 | * | |
794 | * We are working here with either a clone of the original | |
795 | * SKB, or a fresh unique copy made by the retransmit engine. | |
796 | */ | |
056834d9 IJ |
797 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, |
798 | gfp_t gfp_mask) | |
1da177e4 | 799 | { |
dfb4b9dc DM |
800 | const struct inet_connection_sock *icsk = inet_csk(sk); |
801 | struct inet_sock *inet; | |
802 | struct tcp_sock *tp; | |
803 | struct tcp_skb_cb *tcb; | |
33ad798c | 804 | struct tcp_out_options opts; |
95c96174 | 805 | unsigned int tcp_options_size, tcp_header_size; |
cfb6eeb4 | 806 | struct tcp_md5sig_key *md5; |
dfb4b9dc | 807 | struct tcphdr *th; |
dfb4b9dc DM |
808 | int err; |
809 | ||
810 | BUG_ON(!skb || !tcp_skb_pcount(skb)); | |
811 | ||
812 | /* If congestion control is doing timestamping, we must | |
813 | * take such a timestamp before we potentially clone/copy. | |
814 | */ | |
164891aa | 815 | if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) |
dfb4b9dc DM |
816 | __net_timestamp(skb); |
817 | ||
818 | if (likely(clone_it)) { | |
819 | if (unlikely(skb_cloned(skb))) | |
820 | skb = pskb_copy(skb, gfp_mask); | |
821 | else | |
822 | skb = skb_clone(skb, gfp_mask); | |
823 | if (unlikely(!skb)) | |
824 | return -ENOBUFS; | |
825 | } | |
1da177e4 | 826 | |
dfb4b9dc DM |
827 | inet = inet_sk(sk); |
828 | tp = tcp_sk(sk); | |
829 | tcb = TCP_SKB_CB(skb); | |
33ad798c | 830 | memset(&opts, 0, sizeof(opts)); |
1da177e4 | 831 | |
4de075e0 | 832 | if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) |
33ad798c AL |
833 | tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); |
834 | else | |
835 | tcp_options_size = tcp_established_options(sk, skb, &opts, | |
836 | &md5); | |
837 | tcp_header_size = tcp_options_size + sizeof(struct tcphdr); | |
e905a9ed | 838 | |
3853b584 | 839 | if (tcp_packets_in_flight(tp) == 0) { |
dfb4b9dc | 840 | tcp_ca_event(sk, CA_EVENT_TX_START); |
3853b584 TH |
841 | skb->ooo_okay = 1; |
842 | } else | |
843 | skb->ooo_okay = 0; | |
dfb4b9dc | 844 | |
aa8223c7 ACM |
845 | skb_push(skb, tcp_header_size); |
846 | skb_reset_transport_header(skb); | |
e89862f4 | 847 | skb_set_owner_w(skb, sk); |
dfb4b9dc DM |
848 | |
849 | /* Build TCP header and checksum it. */ | |
aa8223c7 | 850 | th = tcp_hdr(skb); |
c720c7e8 ED |
851 | th->source = inet->inet_sport; |
852 | th->dest = inet->inet_dport; | |
dfb4b9dc DM |
853 | th->seq = htonl(tcb->seq); |
854 | th->ack_seq = htonl(tp->rcv_nxt); | |
df7a3b07 | 855 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
4de075e0 | 856 | tcb->tcp_flags); |
dfb4b9dc | 857 | |
4de075e0 | 858 | if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { |
dfb4b9dc DM |
859 | /* RFC1323: The window in SYN & SYN/ACK segments |
860 | * is never scaled. | |
861 | */ | |
600ff0c2 | 862 | th->window = htons(min(tp->rcv_wnd, 65535U)); |
dfb4b9dc DM |
863 | } else { |
864 | th->window = htons(tcp_select_window(sk)); | |
865 | } | |
866 | th->check = 0; | |
867 | th->urg_ptr = 0; | |
1da177e4 | 868 | |
33f5f57e | 869 | /* The urg_mode check is necessary during a below snd_una win probe */ |
7691367d HX |
870 | if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { |
871 | if (before(tp->snd_up, tcb->seq + 0x10000)) { | |
872 | th->urg_ptr = htons(tp->snd_up - tcb->seq); | |
873 | th->urg = 1; | |
874 | } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { | |
0eae88f3 | 875 | th->urg_ptr = htons(0xFFFF); |
7691367d HX |
876 | th->urg = 1; |
877 | } | |
dfb4b9dc | 878 | } |
1da177e4 | 879 | |
bd0388ae | 880 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
4de075e0 | 881 | if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) |
9e412ba7 | 882 | TCP_ECN_send(sk, skb, tcp_header_size); |
1da177e4 | 883 | |
cfb6eeb4 YH |
884 | #ifdef CONFIG_TCP_MD5SIG |
885 | /* Calculate the MD5 hash, as we have all we need now */ | |
886 | if (md5) { | |
a465419b | 887 | sk_nocaps_add(sk, NETIF_F_GSO_MASK); |
bd0388ae | 888 | tp->af_specific->calc_md5_hash(opts.hash_location, |
49a72dfb | 889 | md5, sk, NULL, skb); |
cfb6eeb4 YH |
890 | } |
891 | #endif | |
892 | ||
bb296246 | 893 | icsk->icsk_af_ops->send_check(sk, skb); |
1da177e4 | 894 | |
4de075e0 | 895 | if (likely(tcb->tcp_flags & TCPHDR_ACK)) |
dfb4b9dc | 896 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); |
1da177e4 | 897 | |
dfb4b9dc | 898 | if (skb->len != tcp_header_size) |
cf533ea5 | 899 | tcp_event_data_sent(tp, sk); |
1da177e4 | 900 | |
bd37a088 | 901 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
aa2ea058 TH |
902 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, |
903 | tcp_skb_pcount(skb)); | |
1da177e4 | 904 | |
d9d8da80 | 905 | err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl); |
83de47cd | 906 | if (likely(err <= 0)) |
dfb4b9dc DM |
907 | return err; |
908 | ||
3cfe3baa | 909 | tcp_enter_cwr(sk, 1); |
dfb4b9dc | 910 | |
b9df3cb8 | 911 | return net_xmit_eval(err); |
1da177e4 LT |
912 | } |
913 | ||
67edfef7 | 914 | /* This routine just queues the buffer for sending. |
1da177e4 LT |
915 | * |
916 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, | |
917 | * otherwise socket can stall. | |
918 | */ | |
919 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) | |
920 | { | |
921 | struct tcp_sock *tp = tcp_sk(sk); | |
922 | ||
923 | /* Advance write_seq and place onto the write_queue. */ | |
924 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; | |
925 | skb_header_release(skb); | |
fe067e8a | 926 | tcp_add_write_queue_tail(sk, skb); |
3ab224be HA |
927 | sk->sk_wmem_queued += skb->truesize; |
928 | sk_mem_charge(sk, skb->truesize); | |
1da177e4 LT |
929 | } |
930 | ||
67edfef7 | 931 | /* Initialize TSO segments for a packet. */ |
cf533ea5 | 932 | static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, |
056834d9 | 933 | unsigned int mss_now) |
f6302d1d | 934 | { |
8e5b9dda HX |
935 | if (skb->len <= mss_now || !sk_can_gso(sk) || |
936 | skb->ip_summed == CHECKSUM_NONE) { | |
f6302d1d DM |
937 | /* Avoid the costly divide in the normal |
938 | * non-TSO case. | |
939 | */ | |
7967168c HX |
940 | skb_shinfo(skb)->gso_segs = 1; |
941 | skb_shinfo(skb)->gso_size = 0; | |
942 | skb_shinfo(skb)->gso_type = 0; | |
f6302d1d | 943 | } else { |
356f89e1 | 944 | skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); |
7967168c | 945 | skb_shinfo(skb)->gso_size = mss_now; |
bcd76111 | 946 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
1da177e4 LT |
947 | } |
948 | } | |
949 | ||
91fed7a1 | 950 | /* When a modification to fackets out becomes necessary, we need to check |
68f8353b | 951 | * skb is counted to fackets_out or not. |
91fed7a1 | 952 | */ |
cf533ea5 | 953 | static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, |
91fed7a1 IJ |
954 | int decr) |
955 | { | |
a47e5a98 IJ |
956 | struct tcp_sock *tp = tcp_sk(sk); |
957 | ||
dc86967b | 958 | if (!tp->sacked_out || tcp_is_reno(tp)) |
91fed7a1 IJ |
959 | return; |
960 | ||
6859d494 | 961 | if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) |
91fed7a1 | 962 | tp->fackets_out -= decr; |
91fed7a1 IJ |
963 | } |
964 | ||
797108d1 IJ |
965 | /* Pcount in the middle of the write queue got changed, we need to do various |
966 | * tweaks to fix counters | |
967 | */ | |
cf533ea5 | 968 | static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) |
797108d1 IJ |
969 | { |
970 | struct tcp_sock *tp = tcp_sk(sk); | |
971 | ||
972 | tp->packets_out -= decr; | |
973 | ||
974 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
975 | tp->sacked_out -= decr; | |
976 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
977 | tp->retrans_out -= decr; | |
978 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) | |
979 | tp->lost_out -= decr; | |
980 | ||
981 | /* Reno case is special. Sigh... */ | |
982 | if (tcp_is_reno(tp) && decr > 0) | |
983 | tp->sacked_out -= min_t(u32, tp->sacked_out, decr); | |
984 | ||
985 | tcp_adjust_fackets_out(sk, skb, decr); | |
986 | ||
987 | if (tp->lost_skb_hint && | |
988 | before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && | |
52cf3cc8 | 989 | (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) |
797108d1 IJ |
990 | tp->lost_cnt_hint -= decr; |
991 | ||
992 | tcp_verify_left_out(tp); | |
993 | } | |
994 | ||
1da177e4 LT |
995 | /* Function to create two new TCP segments. Shrinks the given segment |
996 | * to the specified size and appends a new segment with the rest of the | |
e905a9ed | 997 | * packet to the list. This won't be called frequently, I hope. |
1da177e4 LT |
998 | * Remember, these are still headerless SKBs at this point. |
999 | */ | |
056834d9 IJ |
1000 | int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, |
1001 | unsigned int mss_now) | |
1da177e4 LT |
1002 | { |
1003 | struct tcp_sock *tp = tcp_sk(sk); | |
1004 | struct sk_buff *buff; | |
6475be16 | 1005 | int nsize, old_factor; |
b60b49ea | 1006 | int nlen; |
9ce01461 | 1007 | u8 flags; |
1da177e4 | 1008 | |
2fceec13 IJ |
1009 | if (WARN_ON(len > skb->len)) |
1010 | return -EINVAL; | |
6a438bbe | 1011 | |
1da177e4 LT |
1012 | nsize = skb_headlen(skb) - len; |
1013 | if (nsize < 0) | |
1014 | nsize = 0; | |
1015 | ||
1016 | if (skb_cloned(skb) && | |
1017 | skb_is_nonlinear(skb) && | |
1018 | pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) | |
1019 | return -ENOMEM; | |
1020 | ||
1021 | /* Get a new skb... force flag on. */ | |
1022 | buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); | |
1023 | if (buff == NULL) | |
1024 | return -ENOMEM; /* We'll just try again later. */ | |
ef5cb973 | 1025 | |
3ab224be HA |
1026 | sk->sk_wmem_queued += buff->truesize; |
1027 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea HX |
1028 | nlen = skb->len - len - nsize; |
1029 | buff->truesize += nlen; | |
1030 | skb->truesize -= nlen; | |
1da177e4 LT |
1031 | |
1032 | /* Correct the sequence numbers. */ | |
1033 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1034 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1035 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1036 | ||
1037 | /* PSH and FIN should only be set in the second packet. */ | |
4de075e0 ED |
1038 | flags = TCP_SKB_CB(skb)->tcp_flags; |
1039 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
1040 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
e14c3caf | 1041 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
1da177e4 | 1042 | |
84fa7933 | 1043 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { |
1da177e4 | 1044 | /* Copy and checksum data tail into the new buffer. */ |
056834d9 IJ |
1045 | buff->csum = csum_partial_copy_nocheck(skb->data + len, |
1046 | skb_put(buff, nsize), | |
1da177e4 LT |
1047 | nsize, 0); |
1048 | ||
1049 | skb_trim(skb, len); | |
1050 | ||
1051 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); | |
1052 | } else { | |
84fa7933 | 1053 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
1054 | skb_split(skb, buff, len); |
1055 | } | |
1056 | ||
1057 | buff->ip_summed = skb->ip_summed; | |
1058 | ||
1059 | /* Looks stupid, but our code really uses when of | |
1060 | * skbs, which it never sent before. --ANK | |
1061 | */ | |
1062 | TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; | |
a61bbcf2 | 1063 | buff->tstamp = skb->tstamp; |
1da177e4 | 1064 | |
6475be16 DM |
1065 | old_factor = tcp_skb_pcount(skb); |
1066 | ||
1da177e4 | 1067 | /* Fix up tso_factor for both original and new SKB. */ |
846998ae DM |
1068 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
1069 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
1da177e4 | 1070 | |
6475be16 DM |
1071 | /* If this packet has been sent out already, we must |
1072 | * adjust the various packet counters. | |
1073 | */ | |
cf0b450c | 1074 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
6475be16 DM |
1075 | int diff = old_factor - tcp_skb_pcount(skb) - |
1076 | tcp_skb_pcount(buff); | |
1da177e4 | 1077 | |
797108d1 IJ |
1078 | if (diff) |
1079 | tcp_adjust_pcount(sk, skb, diff); | |
1da177e4 LT |
1080 | } |
1081 | ||
1082 | /* Link BUFF into the send queue. */ | |
f44b5271 | 1083 | skb_header_release(buff); |
fe067e8a | 1084 | tcp_insert_write_queue_after(skb, buff, sk); |
1da177e4 LT |
1085 | |
1086 | return 0; | |
1087 | } | |
1088 | ||
1089 | /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c | |
1090 | * eventually). The difference is that pulled data not copied, but | |
1091 | * immediately discarded. | |
1092 | */ | |
f2911969 | 1093 | static void __pskb_trim_head(struct sk_buff *skb, int len) |
1da177e4 LT |
1094 | { |
1095 | int i, k, eat; | |
1096 | ||
4fa48bf3 ED |
1097 | eat = min_t(int, len, skb_headlen(skb)); |
1098 | if (eat) { | |
1099 | __skb_pull(skb, eat); | |
22b4a4f2 | 1100 | skb->avail_size -= eat; |
4fa48bf3 ED |
1101 | len -= eat; |
1102 | if (!len) | |
1103 | return; | |
1104 | } | |
1da177e4 LT |
1105 | eat = len; |
1106 | k = 0; | |
056834d9 | 1107 | for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { |
9e903e08 ED |
1108 | int size = skb_frag_size(&skb_shinfo(skb)->frags[i]); |
1109 | ||
1110 | if (size <= eat) { | |
aff65da0 | 1111 | skb_frag_unref(skb, i); |
9e903e08 | 1112 | eat -= size; |
1da177e4 LT |
1113 | } else { |
1114 | skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; | |
1115 | if (eat) { | |
1116 | skb_shinfo(skb)->frags[k].page_offset += eat; | |
9e903e08 | 1117 | skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat); |
1da177e4 LT |
1118 | eat = 0; |
1119 | } | |
1120 | k++; | |
1121 | } | |
1122 | } | |
1123 | skb_shinfo(skb)->nr_frags = k; | |
1124 | ||
27a884dc | 1125 | skb_reset_tail_pointer(skb); |
1da177e4 LT |
1126 | skb->data_len -= len; |
1127 | skb->len = skb->data_len; | |
1da177e4 LT |
1128 | } |
1129 | ||
67edfef7 | 1130 | /* Remove acked data from a packet in the transmit queue. */ |
1da177e4 LT |
1131 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) |
1132 | { | |
056834d9 | 1133 | if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) |
1da177e4 LT |
1134 | return -ENOMEM; |
1135 | ||
4fa48bf3 | 1136 | __pskb_trim_head(skb, len); |
1da177e4 LT |
1137 | |
1138 | TCP_SKB_CB(skb)->seq += len; | |
84fa7933 | 1139 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
1140 | |
1141 | skb->truesize -= len; | |
1142 | sk->sk_wmem_queued -= len; | |
3ab224be | 1143 | sk_mem_uncharge(sk, len); |
1da177e4 LT |
1144 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); |
1145 | ||
5b35e1e6 | 1146 | /* Any change of skb->len requires recalculation of tso factor. */ |
1da177e4 | 1147 | if (tcp_skb_pcount(skb) > 1) |
5b35e1e6 | 1148 | tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); |
1da177e4 LT |
1149 | |
1150 | return 0; | |
1151 | } | |
1152 | ||
67edfef7 | 1153 | /* Calculate MSS. Not accounting for SACKs here. */ |
67469601 | 1154 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) |
5d424d5a | 1155 | { |
cf533ea5 ED |
1156 | const struct tcp_sock *tp = tcp_sk(sk); |
1157 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
5d424d5a JH |
1158 | int mss_now; |
1159 | ||
1160 | /* Calculate base mss without TCP options: | |
1161 | It is MMS_S - sizeof(tcphdr) of rfc1122 | |
1162 | */ | |
1163 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); | |
1164 | ||
67469601 ED |
1165 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
1166 | if (icsk->icsk_af_ops->net_frag_header_len) { | |
1167 | const struct dst_entry *dst = __sk_dst_get(sk); | |
1168 | ||
1169 | if (dst && dst_allfrag(dst)) | |
1170 | mss_now -= icsk->icsk_af_ops->net_frag_header_len; | |
1171 | } | |
1172 | ||
5d424d5a JH |
1173 | /* Clamp it (mss_clamp does not include tcp options) */ |
1174 | if (mss_now > tp->rx_opt.mss_clamp) | |
1175 | mss_now = tp->rx_opt.mss_clamp; | |
1176 | ||
1177 | /* Now subtract optional transport overhead */ | |
1178 | mss_now -= icsk->icsk_ext_hdr_len; | |
1179 | ||
1180 | /* Then reserve room for full set of TCP options and 8 bytes of data */ | |
1181 | if (mss_now < 48) | |
1182 | mss_now = 48; | |
1183 | ||
1184 | /* Now subtract TCP options size, not including SACKs */ | |
1185 | mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); | |
1186 | ||
1187 | return mss_now; | |
1188 | } | |
1189 | ||
1190 | /* Inverse of above */ | |
67469601 | 1191 | int tcp_mss_to_mtu(struct sock *sk, int mss) |
5d424d5a | 1192 | { |
cf533ea5 ED |
1193 | const struct tcp_sock *tp = tcp_sk(sk); |
1194 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
5d424d5a JH |
1195 | int mtu; |
1196 | ||
1197 | mtu = mss + | |
1198 | tp->tcp_header_len + | |
1199 | icsk->icsk_ext_hdr_len + | |
1200 | icsk->icsk_af_ops->net_header_len; | |
1201 | ||
67469601 ED |
1202 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
1203 | if (icsk->icsk_af_ops->net_frag_header_len) { | |
1204 | const struct dst_entry *dst = __sk_dst_get(sk); | |
1205 | ||
1206 | if (dst && dst_allfrag(dst)) | |
1207 | mtu += icsk->icsk_af_ops->net_frag_header_len; | |
1208 | } | |
5d424d5a JH |
1209 | return mtu; |
1210 | } | |
1211 | ||
67edfef7 | 1212 | /* MTU probing init per socket */ |
5d424d5a JH |
1213 | void tcp_mtup_init(struct sock *sk) |
1214 | { | |
1215 | struct tcp_sock *tp = tcp_sk(sk); | |
1216 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1217 | ||
1218 | icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; | |
1219 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + | |
e905a9ed | 1220 | icsk->icsk_af_ops->net_header_len; |
5d424d5a JH |
1221 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); |
1222 | icsk->icsk_mtup.probe_size = 0; | |
1223 | } | |
4bc2f18b | 1224 | EXPORT_SYMBOL(tcp_mtup_init); |
5d424d5a | 1225 | |
1da177e4 LT |
1226 | /* This function synchronize snd mss to current pmtu/exthdr set. |
1227 | ||
1228 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts | |
1229 | for TCP options, but includes only bare TCP header. | |
1230 | ||
1231 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. | |
caa20d9a | 1232 | It is minimum of user_mss and mss received with SYN. |
1da177e4 LT |
1233 | It also does not include TCP options. |
1234 | ||
d83d8461 | 1235 | inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. |
1da177e4 LT |
1236 | |
1237 | tp->mss_cache is current effective sending mss, including | |
1238 | all tcp options except for SACKs. It is evaluated, | |
1239 | taking into account current pmtu, but never exceeds | |
1240 | tp->rx_opt.mss_clamp. | |
1241 | ||
1242 | NOTE1. rfc1122 clearly states that advertised MSS | |
1243 | DOES NOT include either tcp or ip options. | |
1244 | ||
d83d8461 ACM |
1245 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
1246 | are READ ONLY outside this function. --ANK (980731) | |
1da177e4 | 1247 | */ |
1da177e4 LT |
1248 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) |
1249 | { | |
1250 | struct tcp_sock *tp = tcp_sk(sk); | |
d83d8461 | 1251 | struct inet_connection_sock *icsk = inet_csk(sk); |
5d424d5a | 1252 | int mss_now; |
1da177e4 | 1253 | |
5d424d5a JH |
1254 | if (icsk->icsk_mtup.search_high > pmtu) |
1255 | icsk->icsk_mtup.search_high = pmtu; | |
1da177e4 | 1256 | |
5d424d5a | 1257 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
409d22b4 | 1258 | mss_now = tcp_bound_to_half_wnd(tp, mss_now); |
1da177e4 LT |
1259 | |
1260 | /* And store cached results */ | |
d83d8461 | 1261 | icsk->icsk_pmtu_cookie = pmtu; |
5d424d5a JH |
1262 | if (icsk->icsk_mtup.enabled) |
1263 | mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); | |
c1b4a7e6 | 1264 | tp->mss_cache = mss_now; |
1da177e4 LT |
1265 | |
1266 | return mss_now; | |
1267 | } | |
4bc2f18b | 1268 | EXPORT_SYMBOL(tcp_sync_mss); |
1da177e4 LT |
1269 | |
1270 | /* Compute the current effective MSS, taking SACKs and IP options, | |
1271 | * and even PMTU discovery events into account. | |
1da177e4 | 1272 | */ |
0c54b85f | 1273 | unsigned int tcp_current_mss(struct sock *sk) |
1da177e4 | 1274 | { |
cf533ea5 ED |
1275 | const struct tcp_sock *tp = tcp_sk(sk); |
1276 | const struct dst_entry *dst = __sk_dst_get(sk); | |
c1b4a7e6 | 1277 | u32 mss_now; |
95c96174 | 1278 | unsigned int header_len; |
33ad798c AL |
1279 | struct tcp_out_options opts; |
1280 | struct tcp_md5sig_key *md5; | |
c1b4a7e6 DM |
1281 | |
1282 | mss_now = tp->mss_cache; | |
1283 | ||
1da177e4 LT |
1284 | if (dst) { |
1285 | u32 mtu = dst_mtu(dst); | |
d83d8461 | 1286 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
1da177e4 LT |
1287 | mss_now = tcp_sync_mss(sk, mtu); |
1288 | } | |
1289 | ||
33ad798c AL |
1290 | header_len = tcp_established_options(sk, NULL, &opts, &md5) + |
1291 | sizeof(struct tcphdr); | |
1292 | /* The mss_cache is sized based on tp->tcp_header_len, which assumes | |
1293 | * some common options. If this is an odd packet (because we have SACK | |
1294 | * blocks etc) then our calculated header_len will be different, and | |
1295 | * we have to adjust mss_now correspondingly */ | |
1296 | if (header_len != tp->tcp_header_len) { | |
1297 | int delta = (int) header_len - tp->tcp_header_len; | |
1298 | mss_now -= delta; | |
1299 | } | |
cfb6eeb4 | 1300 | |
1da177e4 LT |
1301 | return mss_now; |
1302 | } | |
1303 | ||
a762a980 | 1304 | /* Congestion window validation. (RFC2861) */ |
9e412ba7 | 1305 | static void tcp_cwnd_validate(struct sock *sk) |
a762a980 | 1306 | { |
9e412ba7 | 1307 | struct tcp_sock *tp = tcp_sk(sk); |
a762a980 | 1308 | |
d436d686 | 1309 | if (tp->packets_out >= tp->snd_cwnd) { |
a762a980 DM |
1310 | /* Network is feed fully. */ |
1311 | tp->snd_cwnd_used = 0; | |
1312 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1313 | } else { | |
1314 | /* Network starves. */ | |
1315 | if (tp->packets_out > tp->snd_cwnd_used) | |
1316 | tp->snd_cwnd_used = tp->packets_out; | |
1317 | ||
15d33c07 DM |
1318 | if (sysctl_tcp_slow_start_after_idle && |
1319 | (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) | |
a762a980 DM |
1320 | tcp_cwnd_application_limited(sk); |
1321 | } | |
1322 | } | |
1323 | ||
0e3a4803 IJ |
1324 | /* Returns the portion of skb which can be sent right away without |
1325 | * introducing MSS oddities to segment boundaries. In rare cases where | |
1326 | * mss_now != mss_cache, we will request caller to create a small skb | |
1327 | * per input skb which could be mostly avoided here (if desired). | |
5ea3a748 IJ |
1328 | * |
1329 | * We explicitly want to create a request for splitting write queue tail | |
1330 | * to a small skb for Nagle purposes while avoiding unnecessary modulos, | |
1331 | * thus all the complexity (cwnd_len is always MSS multiple which we | |
1332 | * return whenever allowed by the other factors). Basically we need the | |
1333 | * modulo only when the receiver window alone is the limiting factor or | |
1334 | * when we would be allowed to send the split-due-to-Nagle skb fully. | |
0e3a4803 | 1335 | */ |
cf533ea5 | 1336 | static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb, |
056834d9 | 1337 | unsigned int mss_now, unsigned int cwnd) |
c1b4a7e6 | 1338 | { |
cf533ea5 | 1339 | const struct tcp_sock *tp = tcp_sk(sk); |
0e3a4803 | 1340 | u32 needed, window, cwnd_len; |
c1b4a7e6 | 1341 | |
90840def | 1342 | window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
c1b4a7e6 | 1343 | cwnd_len = mss_now * cwnd; |
0e3a4803 IJ |
1344 | |
1345 | if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk))) | |
1346 | return cwnd_len; | |
1347 | ||
5ea3a748 IJ |
1348 | needed = min(skb->len, window); |
1349 | ||
17515408 | 1350 | if (cwnd_len <= needed) |
0e3a4803 IJ |
1351 | return cwnd_len; |
1352 | ||
0e3a4803 | 1353 | return needed - needed % mss_now; |
c1b4a7e6 DM |
1354 | } |
1355 | ||
1356 | /* Can at least one segment of SKB be sent right now, according to the | |
1357 | * congestion window rules? If so, return how many segments are allowed. | |
1358 | */ | |
cf533ea5 ED |
1359 | static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, |
1360 | const struct sk_buff *skb) | |
c1b4a7e6 DM |
1361 | { |
1362 | u32 in_flight, cwnd; | |
1363 | ||
1364 | /* Don't be strict about the congestion window for the final FIN. */ | |
4de075e0 ED |
1365 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && |
1366 | tcp_skb_pcount(skb) == 1) | |
c1b4a7e6 DM |
1367 | return 1; |
1368 | ||
1369 | in_flight = tcp_packets_in_flight(tp); | |
1370 | cwnd = tp->snd_cwnd; | |
1371 | if (in_flight < cwnd) | |
1372 | return (cwnd - in_flight); | |
1373 | ||
1374 | return 0; | |
1375 | } | |
1376 | ||
b595076a | 1377 | /* Initialize TSO state of a skb. |
67edfef7 | 1378 | * This must be invoked the first time we consider transmitting |
c1b4a7e6 DM |
1379 | * SKB onto the wire. |
1380 | */ | |
cf533ea5 | 1381 | static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, |
056834d9 | 1382 | unsigned int mss_now) |
c1b4a7e6 DM |
1383 | { |
1384 | int tso_segs = tcp_skb_pcount(skb); | |
1385 | ||
f8269a49 | 1386 | if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { |
846998ae | 1387 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
c1b4a7e6 DM |
1388 | tso_segs = tcp_skb_pcount(skb); |
1389 | } | |
1390 | return tso_segs; | |
1391 | } | |
1392 | ||
67edfef7 | 1393 | /* Minshall's variant of the Nagle send check. */ |
a2a385d6 | 1394 | static inline bool tcp_minshall_check(const struct tcp_sock *tp) |
c1b4a7e6 | 1395 | { |
09cb105e | 1396 | return after(tp->snd_sml, tp->snd_una) && |
c1b4a7e6 DM |
1397 | !after(tp->snd_sml, tp->snd_nxt); |
1398 | } | |
1399 | ||
a2a385d6 | 1400 | /* Return false, if packet can be sent now without violation Nagle's rules: |
c1b4a7e6 DM |
1401 | * 1. It is full sized. |
1402 | * 2. Or it contains FIN. (already checked by caller) | |
6d67e9be | 1403 | * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. |
c1b4a7e6 DM |
1404 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. |
1405 | * With Minshall's modification: all sent small packets are ACKed. | |
1406 | */ | |
a2a385d6 | 1407 | static inline bool tcp_nagle_check(const struct tcp_sock *tp, |
e905a9ed | 1408 | const struct sk_buff *skb, |
95c96174 | 1409 | unsigned int mss_now, int nonagle) |
c1b4a7e6 | 1410 | { |
a02cec21 | 1411 | return skb->len < mss_now && |
056834d9 | 1412 | ((nonagle & TCP_NAGLE_CORK) || |
a02cec21 | 1413 | (!nonagle && tp->packets_out && tcp_minshall_check(tp))); |
c1b4a7e6 DM |
1414 | } |
1415 | ||
a2a385d6 | 1416 | /* Return true if the Nagle test allows this packet to be |
c1b4a7e6 DM |
1417 | * sent now. |
1418 | */ | |
a2a385d6 ED |
1419 | static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, |
1420 | unsigned int cur_mss, int nonagle) | |
c1b4a7e6 DM |
1421 | { |
1422 | /* Nagle rule does not apply to frames, which sit in the middle of the | |
1423 | * write_queue (they have no chances to get new data). | |
1424 | * | |
1425 | * This is implemented in the callers, where they modify the 'nonagle' | |
1426 | * argument based upon the location of SKB in the send queue. | |
1427 | */ | |
1428 | if (nonagle & TCP_NAGLE_PUSH) | |
a2a385d6 | 1429 | return true; |
c1b4a7e6 | 1430 | |
d551e454 IJ |
1431 | /* Don't use the nagle rule for urgent data (or for the final FIN). |
1432 | * Nagle can be ignored during F-RTO too (see RFC4138). | |
1433 | */ | |
33f5f57e | 1434 | if (tcp_urg_mode(tp) || (tp->frto_counter == 2) || |
4de075e0 | 1435 | (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) |
a2a385d6 | 1436 | return true; |
c1b4a7e6 DM |
1437 | |
1438 | if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) | |
a2a385d6 | 1439 | return true; |
c1b4a7e6 | 1440 | |
a2a385d6 | 1441 | return false; |
c1b4a7e6 DM |
1442 | } |
1443 | ||
1444 | /* Does at least the first segment of SKB fit into the send window? */ | |
a2a385d6 ED |
1445 | static bool tcp_snd_wnd_test(const struct tcp_sock *tp, |
1446 | const struct sk_buff *skb, | |
1447 | unsigned int cur_mss) | |
c1b4a7e6 DM |
1448 | { |
1449 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
1450 | ||
1451 | if (skb->len > cur_mss) | |
1452 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; | |
1453 | ||
90840def | 1454 | return !after(end_seq, tcp_wnd_end(tp)); |
c1b4a7e6 DM |
1455 | } |
1456 | ||
fe067e8a | 1457 | /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) |
c1b4a7e6 DM |
1458 | * should be put on the wire right now. If so, it returns the number of |
1459 | * packets allowed by the congestion window. | |
1460 | */ | |
cf533ea5 | 1461 | static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, |
c1b4a7e6 DM |
1462 | unsigned int cur_mss, int nonagle) |
1463 | { | |
cf533ea5 | 1464 | const struct tcp_sock *tp = tcp_sk(sk); |
c1b4a7e6 DM |
1465 | unsigned int cwnd_quota; |
1466 | ||
846998ae | 1467 | tcp_init_tso_segs(sk, skb, cur_mss); |
c1b4a7e6 DM |
1468 | |
1469 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) | |
1470 | return 0; | |
1471 | ||
1472 | cwnd_quota = tcp_cwnd_test(tp, skb); | |
056834d9 | 1473 | if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) |
c1b4a7e6 DM |
1474 | cwnd_quota = 0; |
1475 | ||
1476 | return cwnd_quota; | |
1477 | } | |
1478 | ||
67edfef7 | 1479 | /* Test if sending is allowed right now. */ |
a2a385d6 | 1480 | bool tcp_may_send_now(struct sock *sk) |
c1b4a7e6 | 1481 | { |
cf533ea5 | 1482 | const struct tcp_sock *tp = tcp_sk(sk); |
fe067e8a | 1483 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 | 1484 | |
a02cec21 | 1485 | return skb && |
0c54b85f | 1486 | tcp_snd_test(sk, skb, tcp_current_mss(sk), |
c1b4a7e6 | 1487 | (tcp_skb_is_last(sk, skb) ? |
a02cec21 | 1488 | tp->nonagle : TCP_NAGLE_PUSH)); |
c1b4a7e6 DM |
1489 | } |
1490 | ||
1491 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet | |
1492 | * which is put after SKB on the list. It is very much like | |
1493 | * tcp_fragment() except that it may make several kinds of assumptions | |
1494 | * in order to speed up the splitting operation. In particular, we | |
1495 | * know that all the data is in scatter-gather pages, and that the | |
1496 | * packet has never been sent out before (and thus is not cloned). | |
1497 | */ | |
056834d9 | 1498 | static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, |
c4ead4c5 | 1499 | unsigned int mss_now, gfp_t gfp) |
c1b4a7e6 DM |
1500 | { |
1501 | struct sk_buff *buff; | |
1502 | int nlen = skb->len - len; | |
9ce01461 | 1503 | u8 flags; |
c1b4a7e6 DM |
1504 | |
1505 | /* All of a TSO frame must be composed of paged data. */ | |
c8ac3774 HX |
1506 | if (skb->len != skb->data_len) |
1507 | return tcp_fragment(sk, skb, len, mss_now); | |
c1b4a7e6 | 1508 | |
c4ead4c5 | 1509 | buff = sk_stream_alloc_skb(sk, 0, gfp); |
c1b4a7e6 DM |
1510 | if (unlikely(buff == NULL)) |
1511 | return -ENOMEM; | |
1512 | ||
3ab224be HA |
1513 | sk->sk_wmem_queued += buff->truesize; |
1514 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea | 1515 | buff->truesize += nlen; |
c1b4a7e6 DM |
1516 | skb->truesize -= nlen; |
1517 | ||
1518 | /* Correct the sequence numbers. */ | |
1519 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1520 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1521 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1522 | ||
1523 | /* PSH and FIN should only be set in the second packet. */ | |
4de075e0 ED |
1524 | flags = TCP_SKB_CB(skb)->tcp_flags; |
1525 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
1526 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
c1b4a7e6 DM |
1527 | |
1528 | /* This packet was never sent out yet, so no SACK bits. */ | |
1529 | TCP_SKB_CB(buff)->sacked = 0; | |
1530 | ||
84fa7933 | 1531 | buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; |
c1b4a7e6 DM |
1532 | skb_split(skb, buff, len); |
1533 | ||
1534 | /* Fix up tso_factor for both original and new SKB. */ | |
846998ae DM |
1535 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
1536 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
c1b4a7e6 DM |
1537 | |
1538 | /* Link BUFF into the send queue. */ | |
1539 | skb_header_release(buff); | |
fe067e8a | 1540 | tcp_insert_write_queue_after(skb, buff, sk); |
c1b4a7e6 DM |
1541 | |
1542 | return 0; | |
1543 | } | |
1544 | ||
1545 | /* Try to defer sending, if possible, in order to minimize the amount | |
1546 | * of TSO splitting we do. View it as a kind of TSO Nagle test. | |
1547 | * | |
1548 | * This algorithm is from John Heffner. | |
1549 | */ | |
a2a385d6 | 1550 | static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) |
c1b4a7e6 | 1551 | { |
9e412ba7 | 1552 | struct tcp_sock *tp = tcp_sk(sk); |
6687e988 | 1553 | const struct inet_connection_sock *icsk = inet_csk(sk); |
c1b4a7e6 | 1554 | u32 send_win, cong_win, limit, in_flight; |
ad9f4f50 | 1555 | int win_divisor; |
c1b4a7e6 | 1556 | |
4de075e0 | 1557 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) |
ae8064ac | 1558 | goto send_now; |
c1b4a7e6 | 1559 | |
6687e988 | 1560 | if (icsk->icsk_ca_state != TCP_CA_Open) |
ae8064ac JH |
1561 | goto send_now; |
1562 | ||
1563 | /* Defer for less than two clock ticks. */ | |
bd515c3e | 1564 | if (tp->tso_deferred && |
a2acde07 | 1565 | (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) |
ae8064ac | 1566 | goto send_now; |
908a75c1 | 1567 | |
c1b4a7e6 DM |
1568 | in_flight = tcp_packets_in_flight(tp); |
1569 | ||
056834d9 | 1570 | BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); |
c1b4a7e6 | 1571 | |
90840def | 1572 | send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
c1b4a7e6 DM |
1573 | |
1574 | /* From in_flight test above, we know that cwnd > in_flight. */ | |
1575 | cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; | |
1576 | ||
1577 | limit = min(send_win, cong_win); | |
1578 | ||
ba244fe9 | 1579 | /* If a full-sized TSO skb can be sent, do it. */ |
82cc1a7a | 1580 | if (limit >= sk->sk_gso_max_size) |
ae8064ac | 1581 | goto send_now; |
ba244fe9 | 1582 | |
62ad2761 IJ |
1583 | /* Middle in queue won't get any more data, full sendable already? */ |
1584 | if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) | |
1585 | goto send_now; | |
1586 | ||
ad9f4f50 ED |
1587 | win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); |
1588 | if (win_divisor) { | |
c1b4a7e6 DM |
1589 | u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); |
1590 | ||
1591 | /* If at least some fraction of a window is available, | |
1592 | * just use it. | |
1593 | */ | |
ad9f4f50 | 1594 | chunk /= win_divisor; |
c1b4a7e6 | 1595 | if (limit >= chunk) |
ae8064ac | 1596 | goto send_now; |
c1b4a7e6 DM |
1597 | } else { |
1598 | /* Different approach, try not to defer past a single | |
1599 | * ACK. Receiver should ACK every other full sized | |
1600 | * frame, so if we have space for more than 3 frames | |
1601 | * then send now. | |
1602 | */ | |
6b5a5c0d | 1603 | if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) |
ae8064ac | 1604 | goto send_now; |
c1b4a7e6 DM |
1605 | } |
1606 | ||
1607 | /* Ok, it looks like it is advisable to defer. */ | |
056834d9 | 1608 | tp->tso_deferred = 1 | (jiffies << 1); |
ae8064ac | 1609 | |
a2a385d6 | 1610 | return true; |
ae8064ac JH |
1611 | |
1612 | send_now: | |
1613 | tp->tso_deferred = 0; | |
a2a385d6 | 1614 | return false; |
c1b4a7e6 DM |
1615 | } |
1616 | ||
5d424d5a | 1617 | /* Create a new MTU probe if we are ready. |
67edfef7 AK |
1618 | * MTU probe is regularly attempting to increase the path MTU by |
1619 | * deliberately sending larger packets. This discovers routing | |
1620 | * changes resulting in larger path MTUs. | |
1621 | * | |
5d424d5a JH |
1622 | * Returns 0 if we should wait to probe (no cwnd available), |
1623 | * 1 if a probe was sent, | |
056834d9 IJ |
1624 | * -1 otherwise |
1625 | */ | |
5d424d5a JH |
1626 | static int tcp_mtu_probe(struct sock *sk) |
1627 | { | |
1628 | struct tcp_sock *tp = tcp_sk(sk); | |
1629 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1630 | struct sk_buff *skb, *nskb, *next; | |
1631 | int len; | |
1632 | int probe_size; | |
91cc17c0 | 1633 | int size_needed; |
5d424d5a JH |
1634 | int copy; |
1635 | int mss_now; | |
1636 | ||
1637 | /* Not currently probing/verifying, | |
1638 | * not in recovery, | |
1639 | * have enough cwnd, and | |
1640 | * not SACKing (the variable headers throw things off) */ | |
1641 | if (!icsk->icsk_mtup.enabled || | |
1642 | icsk->icsk_mtup.probe_size || | |
1643 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || | |
1644 | tp->snd_cwnd < 11 || | |
cabeccbd | 1645 | tp->rx_opt.num_sacks || tp->rx_opt.dsack) |
5d424d5a JH |
1646 | return -1; |
1647 | ||
1648 | /* Very simple search strategy: just double the MSS. */ | |
0c54b85f | 1649 | mss_now = tcp_current_mss(sk); |
056834d9 | 1650 | probe_size = 2 * tp->mss_cache; |
91cc17c0 | 1651 | size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; |
5d424d5a JH |
1652 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { |
1653 | /* TODO: set timer for probe_converge_event */ | |
1654 | return -1; | |
1655 | } | |
1656 | ||
1657 | /* Have enough data in the send queue to probe? */ | |
7f9c33e5 | 1658 | if (tp->write_seq - tp->snd_nxt < size_needed) |
5d424d5a JH |
1659 | return -1; |
1660 | ||
91cc17c0 IJ |
1661 | if (tp->snd_wnd < size_needed) |
1662 | return -1; | |
90840def | 1663 | if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) |
91cc17c0 | 1664 | return 0; |
5d424d5a | 1665 | |
d67c58e9 IJ |
1666 | /* Do we need to wait to drain cwnd? With none in flight, don't stall */ |
1667 | if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { | |
1668 | if (!tcp_packets_in_flight(tp)) | |
5d424d5a JH |
1669 | return -1; |
1670 | else | |
1671 | return 0; | |
1672 | } | |
1673 | ||
1674 | /* We're allowed to probe. Build it now. */ | |
1675 | if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) | |
1676 | return -1; | |
3ab224be HA |
1677 | sk->sk_wmem_queued += nskb->truesize; |
1678 | sk_mem_charge(sk, nskb->truesize); | |
5d424d5a | 1679 | |
fe067e8a | 1680 | skb = tcp_send_head(sk); |
5d424d5a JH |
1681 | |
1682 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; | |
1683 | TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; | |
4de075e0 | 1684 | TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; |
5d424d5a JH |
1685 | TCP_SKB_CB(nskb)->sacked = 0; |
1686 | nskb->csum = 0; | |
84fa7933 | 1687 | nskb->ip_summed = skb->ip_summed; |
5d424d5a | 1688 | |
50c4817e IJ |
1689 | tcp_insert_write_queue_before(nskb, skb, sk); |
1690 | ||
5d424d5a | 1691 | len = 0; |
234b6860 | 1692 | tcp_for_write_queue_from_safe(skb, next, sk) { |
5d424d5a JH |
1693 | copy = min_t(int, skb->len, probe_size - len); |
1694 | if (nskb->ip_summed) | |
1695 | skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); | |
1696 | else | |
1697 | nskb->csum = skb_copy_and_csum_bits(skb, 0, | |
056834d9 IJ |
1698 | skb_put(nskb, copy), |
1699 | copy, nskb->csum); | |
5d424d5a JH |
1700 | |
1701 | if (skb->len <= copy) { | |
1702 | /* We've eaten all the data from this skb. | |
1703 | * Throw it away. */ | |
4de075e0 | 1704 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; |
fe067e8a | 1705 | tcp_unlink_write_queue(skb, sk); |
3ab224be | 1706 | sk_wmem_free_skb(sk, skb); |
5d424d5a | 1707 | } else { |
4de075e0 | 1708 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & |
a3433f35 | 1709 | ~(TCPHDR_FIN|TCPHDR_PSH); |
5d424d5a JH |
1710 | if (!skb_shinfo(skb)->nr_frags) { |
1711 | skb_pull(skb, copy); | |
84fa7933 | 1712 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
056834d9 IJ |
1713 | skb->csum = csum_partial(skb->data, |
1714 | skb->len, 0); | |
5d424d5a JH |
1715 | } else { |
1716 | __pskb_trim_head(skb, copy); | |
1717 | tcp_set_skb_tso_segs(sk, skb, mss_now); | |
1718 | } | |
1719 | TCP_SKB_CB(skb)->seq += copy; | |
1720 | } | |
1721 | ||
1722 | len += copy; | |
234b6860 IJ |
1723 | |
1724 | if (len >= probe_size) | |
1725 | break; | |
5d424d5a JH |
1726 | } |
1727 | tcp_init_tso_segs(sk, nskb, nskb->len); | |
1728 | ||
1729 | /* We're ready to send. If this fails, the probe will | |
1730 | * be resegmented into mss-sized pieces by tcp_write_xmit(). */ | |
1731 | TCP_SKB_CB(nskb)->when = tcp_time_stamp; | |
1732 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { | |
1733 | /* Decrement cwnd here because we are sending | |
056834d9 | 1734 | * effectively two packets. */ |
5d424d5a | 1735 | tp->snd_cwnd--; |
66f5fe62 | 1736 | tcp_event_new_data_sent(sk, nskb); |
5d424d5a JH |
1737 | |
1738 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); | |
0e7b1368 JH |
1739 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; |
1740 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; | |
5d424d5a JH |
1741 | |
1742 | return 1; | |
1743 | } | |
1744 | ||
1745 | return -1; | |
1746 | } | |
1747 | ||
1da177e4 LT |
1748 | /* This routine writes packets to the network. It advances the |
1749 | * send_head. This happens as incoming acks open up the remote | |
1750 | * window for us. | |
1751 | * | |
f8269a49 IJ |
1752 | * LARGESEND note: !tcp_urg_mode is overkill, only frames between |
1753 | * snd_up-64k-mss .. snd_up cannot be large. However, taking into | |
1754 | * account rare use of URG, this is not a big flaw. | |
1755 | * | |
a2a385d6 ED |
1756 | * Returns true, if no segments are in flight and we have queued segments, |
1757 | * but cannot send anything now because of SWS or another problem. | |
1da177e4 | 1758 | */ |
a2a385d6 ED |
1759 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
1760 | int push_one, gfp_t gfp) | |
1da177e4 LT |
1761 | { |
1762 | struct tcp_sock *tp = tcp_sk(sk); | |
92df7b51 | 1763 | struct sk_buff *skb; |
c1b4a7e6 DM |
1764 | unsigned int tso_segs, sent_pkts; |
1765 | int cwnd_quota; | |
5d424d5a | 1766 | int result; |
1da177e4 | 1767 | |
92df7b51 | 1768 | sent_pkts = 0; |
5d424d5a | 1769 | |
d5dd9175 IJ |
1770 | if (!push_one) { |
1771 | /* Do MTU probing. */ | |
1772 | result = tcp_mtu_probe(sk); | |
1773 | if (!result) { | |
a2a385d6 | 1774 | return false; |
d5dd9175 IJ |
1775 | } else if (result > 0) { |
1776 | sent_pkts = 1; | |
1777 | } | |
5d424d5a JH |
1778 | } |
1779 | ||
fe067e8a | 1780 | while ((skb = tcp_send_head(sk))) { |
c8ac3774 HX |
1781 | unsigned int limit; |
1782 | ||
b68e9f85 | 1783 | tso_segs = tcp_init_tso_segs(sk, skb, mss_now); |
c1b4a7e6 | 1784 | BUG_ON(!tso_segs); |
aa93466b | 1785 | |
b68e9f85 HX |
1786 | cwnd_quota = tcp_cwnd_test(tp, skb); |
1787 | if (!cwnd_quota) | |
1788 | break; | |
1789 | ||
1790 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) | |
1791 | break; | |
1792 | ||
c1b4a7e6 DM |
1793 | if (tso_segs == 1) { |
1794 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, | |
1795 | (tcp_skb_is_last(sk, skb) ? | |
1796 | nonagle : TCP_NAGLE_PUSH)))) | |
1797 | break; | |
1798 | } else { | |
d5dd9175 | 1799 | if (!push_one && tcp_tso_should_defer(sk, skb)) |
c1b4a7e6 DM |
1800 | break; |
1801 | } | |
aa93466b | 1802 | |
c8ac3774 | 1803 | limit = mss_now; |
f8269a49 | 1804 | if (tso_segs > 1 && !tcp_urg_mode(tp)) |
0e3a4803 IJ |
1805 | limit = tcp_mss_split_point(sk, skb, mss_now, |
1806 | cwnd_quota); | |
1da177e4 | 1807 | |
c8ac3774 | 1808 | if (skb->len > limit && |
c4ead4c5 | 1809 | unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) |
c8ac3774 HX |
1810 | break; |
1811 | ||
92df7b51 | 1812 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
c1b4a7e6 | 1813 | |
d5dd9175 | 1814 | if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) |
92df7b51 | 1815 | break; |
1da177e4 | 1816 | |
92df7b51 DM |
1817 | /* Advance the send_head. This one is sent out. |
1818 | * This call will increment packets_out. | |
1819 | */ | |
66f5fe62 | 1820 | tcp_event_new_data_sent(sk, skb); |
1da177e4 | 1821 | |
92df7b51 | 1822 | tcp_minshall_update(tp, mss_now, skb); |
a262f0cd | 1823 | sent_pkts += tcp_skb_pcount(skb); |
d5dd9175 IJ |
1824 | |
1825 | if (push_one) | |
1826 | break; | |
92df7b51 | 1827 | } |
a262f0cd ND |
1828 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery) |
1829 | tp->prr_out += sent_pkts; | |
1da177e4 | 1830 | |
aa93466b | 1831 | if (likely(sent_pkts)) { |
9e412ba7 | 1832 | tcp_cwnd_validate(sk); |
a2a385d6 | 1833 | return false; |
1da177e4 | 1834 | } |
fe067e8a | 1835 | return !tp->packets_out && tcp_send_head(sk); |
1da177e4 LT |
1836 | } |
1837 | ||
a762a980 DM |
1838 | /* Push out any pending frames which were held back due to |
1839 | * TCP_CORK or attempt at coalescing tiny packets. | |
1840 | * The socket must be locked by the caller. | |
1841 | */ | |
9e412ba7 IJ |
1842 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, |
1843 | int nonagle) | |
a762a980 | 1844 | { |
726e07a8 IJ |
1845 | /* If we are closed, the bytes will have to remain here. |
1846 | * In time closedown will finish, we empty the write queue and | |
1847 | * all will be happy. | |
1848 | */ | |
1849 | if (unlikely(sk->sk_state == TCP_CLOSE)) | |
1850 | return; | |
1851 | ||
d5dd9175 | 1852 | if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC)) |
726e07a8 | 1853 | tcp_check_probe_timer(sk); |
a762a980 DM |
1854 | } |
1855 | ||
c1b4a7e6 DM |
1856 | /* Send _single_ skb sitting at the send head. This function requires |
1857 | * true push pending frames to setup probe timer etc. | |
1858 | */ | |
1859 | void tcp_push_one(struct sock *sk, unsigned int mss_now) | |
1860 | { | |
fe067e8a | 1861 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 DM |
1862 | |
1863 | BUG_ON(!skb || skb->len < mss_now); | |
1864 | ||
d5dd9175 | 1865 | tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); |
c1b4a7e6 DM |
1866 | } |
1867 | ||
1da177e4 LT |
1868 | /* This function returns the amount that we can raise the |
1869 | * usable window based on the following constraints | |
e905a9ed | 1870 | * |
1da177e4 LT |
1871 | * 1. The window can never be shrunk once it is offered (RFC 793) |
1872 | * 2. We limit memory per socket | |
1873 | * | |
1874 | * RFC 1122: | |
1875 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep | |
1876 | * RECV.NEXT + RCV.WIN fixed until: | |
1877 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" | |
1878 | * | |
1879 | * i.e. don't raise the right edge of the window until you can raise | |
1880 | * it at least MSS bytes. | |
1881 | * | |
1882 | * Unfortunately, the recommended algorithm breaks header prediction, | |
1883 | * since header prediction assumes th->window stays fixed. | |
1884 | * | |
1885 | * Strictly speaking, keeping th->window fixed violates the receiver | |
1886 | * side SWS prevention criteria. The problem is that under this rule | |
1887 | * a stream of single byte packets will cause the right side of the | |
1888 | * window to always advance by a single byte. | |
e905a9ed | 1889 | * |
1da177e4 LT |
1890 | * Of course, if the sender implements sender side SWS prevention |
1891 | * then this will not be a problem. | |
e905a9ed | 1892 | * |
1da177e4 | 1893 | * BSD seems to make the following compromise: |
e905a9ed | 1894 | * |
1da177e4 LT |
1895 | * If the free space is less than the 1/4 of the maximum |
1896 | * space available and the free space is less than 1/2 mss, | |
1897 | * then set the window to 0. | |
1898 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] | |
1899 | * Otherwise, just prevent the window from shrinking | |
1900 | * and from being larger than the largest representable value. | |
1901 | * | |
1902 | * This prevents incremental opening of the window in the regime | |
1903 | * where TCP is limited by the speed of the reader side taking | |
1904 | * data out of the TCP receive queue. It does nothing about | |
1905 | * those cases where the window is constrained on the sender side | |
1906 | * because the pipeline is full. | |
1907 | * | |
1908 | * BSD also seems to "accidentally" limit itself to windows that are a | |
1909 | * multiple of MSS, at least until the free space gets quite small. | |
1910 | * This would appear to be a side effect of the mbuf implementation. | |
1911 | * Combining these two algorithms results in the observed behavior | |
1912 | * of having a fixed window size at almost all times. | |
1913 | * | |
1914 | * Below we obtain similar behavior by forcing the offered window to | |
1915 | * a multiple of the mss when it is feasible to do so. | |
1916 | * | |
1917 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. | |
1918 | * Regular options like TIMESTAMP are taken into account. | |
1919 | */ | |
1920 | u32 __tcp_select_window(struct sock *sk) | |
1921 | { | |
463c84b9 | 1922 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 1923 | struct tcp_sock *tp = tcp_sk(sk); |
caa20d9a | 1924 | /* MSS for the peer's data. Previous versions used mss_clamp |
1da177e4 LT |
1925 | * here. I don't know if the value based on our guesses |
1926 | * of peer's MSS is better for the performance. It's more correct | |
1927 | * but may be worse for the performance because of rcv_mss | |
1928 | * fluctuations. --SAW 1998/11/1 | |
1929 | */ | |
463c84b9 | 1930 | int mss = icsk->icsk_ack.rcv_mss; |
1da177e4 LT |
1931 | int free_space = tcp_space(sk); |
1932 | int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); | |
1933 | int window; | |
1934 | ||
1935 | if (mss > full_space) | |
e905a9ed | 1936 | mss = full_space; |
1da177e4 | 1937 | |
b92edbe0 | 1938 | if (free_space < (full_space >> 1)) { |
463c84b9 | 1939 | icsk->icsk_ack.quick = 0; |
1da177e4 | 1940 | |
180d8cd9 | 1941 | if (sk_under_memory_pressure(sk)) |
056834d9 IJ |
1942 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, |
1943 | 4U * tp->advmss); | |
1da177e4 LT |
1944 | |
1945 | if (free_space < mss) | |
1946 | return 0; | |
1947 | } | |
1948 | ||
1949 | if (free_space > tp->rcv_ssthresh) | |
1950 | free_space = tp->rcv_ssthresh; | |
1951 | ||
1952 | /* Don't do rounding if we are using window scaling, since the | |
1953 | * scaled window will not line up with the MSS boundary anyway. | |
1954 | */ | |
1955 | window = tp->rcv_wnd; | |
1956 | if (tp->rx_opt.rcv_wscale) { | |
1957 | window = free_space; | |
1958 | ||
1959 | /* Advertise enough space so that it won't get scaled away. | |
1960 | * Import case: prevent zero window announcement if | |
1961 | * 1<<rcv_wscale > mss. | |
1962 | */ | |
1963 | if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) | |
1964 | window = (((window >> tp->rx_opt.rcv_wscale) + 1) | |
1965 | << tp->rx_opt.rcv_wscale); | |
1966 | } else { | |
1967 | /* Get the largest window that is a nice multiple of mss. | |
1968 | * Window clamp already applied above. | |
1969 | * If our current window offering is within 1 mss of the | |
1970 | * free space we just keep it. This prevents the divide | |
1971 | * and multiply from happening most of the time. | |
1972 | * We also don't do any window rounding when the free space | |
1973 | * is too small. | |
1974 | */ | |
1975 | if (window <= free_space - mss || window > free_space) | |
056834d9 | 1976 | window = (free_space / mss) * mss; |
84565070 | 1977 | else if (mss == full_space && |
b92edbe0 | 1978 | free_space > window + (full_space >> 1)) |
84565070 | 1979 | window = free_space; |
1da177e4 LT |
1980 | } |
1981 | ||
1982 | return window; | |
1983 | } | |
1984 | ||
4a17fc3a IJ |
1985 | /* Collapses two adjacent SKB's during retransmission. */ |
1986 | static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) | |
1da177e4 LT |
1987 | { |
1988 | struct tcp_sock *tp = tcp_sk(sk); | |
fe067e8a | 1989 | struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); |
058dc334 | 1990 | int skb_size, next_skb_size; |
1da177e4 | 1991 | |
058dc334 IJ |
1992 | skb_size = skb->len; |
1993 | next_skb_size = next_skb->len; | |
1da177e4 | 1994 | |
058dc334 | 1995 | BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); |
a6963a6b | 1996 | |
058dc334 | 1997 | tcp_highest_sack_combine(sk, next_skb, skb); |
1da177e4 | 1998 | |
058dc334 | 1999 | tcp_unlink_write_queue(next_skb, sk); |
1da177e4 | 2000 | |
058dc334 IJ |
2001 | skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), |
2002 | next_skb_size); | |
1da177e4 | 2003 | |
058dc334 IJ |
2004 | if (next_skb->ip_summed == CHECKSUM_PARTIAL) |
2005 | skb->ip_summed = CHECKSUM_PARTIAL; | |
1da177e4 | 2006 | |
058dc334 IJ |
2007 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
2008 | skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); | |
1da177e4 | 2009 | |
058dc334 IJ |
2010 | /* Update sequence range on original skb. */ |
2011 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; | |
1da177e4 | 2012 | |
e6c7d085 | 2013 | /* Merge over control information. This moves PSH/FIN etc. over */ |
4de075e0 | 2014 | TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; |
058dc334 IJ |
2015 | |
2016 | /* All done, get rid of second SKB and account for it so | |
2017 | * packet counting does not break. | |
2018 | */ | |
2019 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; | |
058dc334 IJ |
2020 | |
2021 | /* changed transmit queue under us so clear hints */ | |
ef9da47c IJ |
2022 | tcp_clear_retrans_hints_partial(tp); |
2023 | if (next_skb == tp->retransmit_skb_hint) | |
2024 | tp->retransmit_skb_hint = skb; | |
058dc334 | 2025 | |
797108d1 IJ |
2026 | tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); |
2027 | ||
058dc334 | 2028 | sk_wmem_free_skb(sk, next_skb); |
1da177e4 LT |
2029 | } |
2030 | ||
67edfef7 | 2031 | /* Check if coalescing SKBs is legal. */ |
a2a385d6 | 2032 | static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) |
4a17fc3a IJ |
2033 | { |
2034 | if (tcp_skb_pcount(skb) > 1) | |
a2a385d6 | 2035 | return false; |
4a17fc3a IJ |
2036 | /* TODO: SACK collapsing could be used to remove this condition */ |
2037 | if (skb_shinfo(skb)->nr_frags != 0) | |
a2a385d6 | 2038 | return false; |
4a17fc3a | 2039 | if (skb_cloned(skb)) |
a2a385d6 | 2040 | return false; |
4a17fc3a | 2041 | if (skb == tcp_send_head(sk)) |
a2a385d6 | 2042 | return false; |
4a17fc3a IJ |
2043 | /* Some heurestics for collapsing over SACK'd could be invented */ |
2044 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
a2a385d6 | 2045 | return false; |
4a17fc3a | 2046 | |
a2a385d6 | 2047 | return true; |
4a17fc3a IJ |
2048 | } |
2049 | ||
67edfef7 AK |
2050 | /* Collapse packets in the retransmit queue to make to create |
2051 | * less packets on the wire. This is only done on retransmission. | |
2052 | */ | |
4a17fc3a IJ |
2053 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, |
2054 | int space) | |
2055 | { | |
2056 | struct tcp_sock *tp = tcp_sk(sk); | |
2057 | struct sk_buff *skb = to, *tmp; | |
a2a385d6 | 2058 | bool first = true; |
4a17fc3a IJ |
2059 | |
2060 | if (!sysctl_tcp_retrans_collapse) | |
2061 | return; | |
4de075e0 | 2062 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) |
4a17fc3a IJ |
2063 | return; |
2064 | ||
2065 | tcp_for_write_queue_from_safe(skb, tmp, sk) { | |
2066 | if (!tcp_can_collapse(sk, skb)) | |
2067 | break; | |
2068 | ||
2069 | space -= skb->len; | |
2070 | ||
2071 | if (first) { | |
a2a385d6 | 2072 | first = false; |
4a17fc3a IJ |
2073 | continue; |
2074 | } | |
2075 | ||
2076 | if (space < 0) | |
2077 | break; | |
2078 | /* Punt if not enough space exists in the first SKB for | |
2079 | * the data in the second | |
2080 | */ | |
a21d4572 | 2081 | if (skb->len > skb_availroom(to)) |
4a17fc3a IJ |
2082 | break; |
2083 | ||
2084 | if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) | |
2085 | break; | |
2086 | ||
2087 | tcp_collapse_retrans(sk, to); | |
2088 | } | |
2089 | } | |
2090 | ||
1da177e4 LT |
2091 | /* This retransmits one SKB. Policy decisions and retransmit queue |
2092 | * state updates are done by the caller. Returns non-zero if an | |
2093 | * error occurred which prevented the send. | |
2094 | */ | |
2095 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) | |
2096 | { | |
2097 | struct tcp_sock *tp = tcp_sk(sk); | |
5d424d5a | 2098 | struct inet_connection_sock *icsk = inet_csk(sk); |
7d227cd2 | 2099 | unsigned int cur_mss; |
1da177e4 LT |
2100 | int err; |
2101 | ||
5d424d5a JH |
2102 | /* Inconslusive MTU probe */ |
2103 | if (icsk->icsk_mtup.probe_size) { | |
2104 | icsk->icsk_mtup.probe_size = 0; | |
2105 | } | |
2106 | ||
1da177e4 | 2107 | /* Do not sent more than we queued. 1/4 is reserved for possible |
caa20d9a | 2108 | * copying overhead: fragmentation, tunneling, mangling etc. |
1da177e4 LT |
2109 | */ |
2110 | if (atomic_read(&sk->sk_wmem_alloc) > | |
2111 | min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) | |
2112 | return -EAGAIN; | |
2113 | ||
2114 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { | |
2115 | if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
2116 | BUG(); | |
1da177e4 LT |
2117 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
2118 | return -ENOMEM; | |
2119 | } | |
2120 | ||
7d227cd2 SS |
2121 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) |
2122 | return -EHOSTUNREACH; /* Routing failure or similar. */ | |
2123 | ||
0c54b85f | 2124 | cur_mss = tcp_current_mss(sk); |
7d227cd2 | 2125 | |
1da177e4 LT |
2126 | /* If receiver has shrunk his window, and skb is out of |
2127 | * new window, do not retransmit it. The exception is the | |
2128 | * case, when window is shrunk to zero. In this case | |
2129 | * our retransmit serves as a zero window probe. | |
2130 | */ | |
9d4fb27d JP |
2131 | if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && |
2132 | TCP_SKB_CB(skb)->seq != tp->snd_una) | |
1da177e4 LT |
2133 | return -EAGAIN; |
2134 | ||
2135 | if (skb->len > cur_mss) { | |
846998ae | 2136 | if (tcp_fragment(sk, skb, cur_mss, cur_mss)) |
1da177e4 | 2137 | return -ENOMEM; /* We'll try again later. */ |
02276f3c | 2138 | } else { |
9eb9362e IJ |
2139 | int oldpcount = tcp_skb_pcount(skb); |
2140 | ||
2141 | if (unlikely(oldpcount > 1)) { | |
2142 | tcp_init_tso_segs(sk, skb, cur_mss); | |
2143 | tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); | |
2144 | } | |
1da177e4 LT |
2145 | } |
2146 | ||
4a17fc3a | 2147 | tcp_retrans_try_collapse(sk, skb, cur_mss); |
1da177e4 | 2148 | |
1da177e4 LT |
2149 | /* Some Solaris stacks overoptimize and ignore the FIN on a |
2150 | * retransmit when old data is attached. So strip it off | |
2151 | * since it is cheap to do so and saves bytes on the network. | |
2152 | */ | |
2de979bd | 2153 | if (skb->len > 0 && |
4de075e0 | 2154 | (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && |
2de979bd | 2155 | tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { |
1da177e4 | 2156 | if (!pskb_trim(skb, 0)) { |
e870a8ef IJ |
2157 | /* Reuse, even though it does some unnecessary work */ |
2158 | tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, | |
4de075e0 | 2159 | TCP_SKB_CB(skb)->tcp_flags); |
1da177e4 | 2160 | skb->ip_summed = CHECKSUM_NONE; |
1da177e4 LT |
2161 | } |
2162 | } | |
2163 | ||
2164 | /* Make a copy, if the first transmission SKB clone we made | |
2165 | * is still in somebody's hands, else make a clone. | |
2166 | */ | |
2167 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
1da177e4 | 2168 | |
117632e6 ED |
2169 | /* make sure skb->data is aligned on arches that require it */ |
2170 | if (unlikely(NET_IP_ALIGN && ((unsigned long)skb->data & 3))) { | |
2171 | struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, | |
2172 | GFP_ATOMIC); | |
2173 | err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : | |
2174 | -ENOBUFS; | |
2175 | } else { | |
2176 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); | |
2177 | } | |
1da177e4 LT |
2178 | |
2179 | if (err == 0) { | |
2180 | /* Update global TCP statistics. */ | |
81cc8a75 | 2181 | TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); |
1da177e4 LT |
2182 | |
2183 | tp->total_retrans++; | |
2184 | ||
2185 | #if FASTRETRANS_DEBUG > 0 | |
056834d9 | 2186 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
e87cc472 | 2187 | net_dbg_ratelimited("retrans_out leaked\n"); |
1da177e4 LT |
2188 | } |
2189 | #endif | |
b08d6cb2 IJ |
2190 | if (!tp->retrans_out) |
2191 | tp->lost_retrans_low = tp->snd_nxt; | |
1da177e4 LT |
2192 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; |
2193 | tp->retrans_out += tcp_skb_pcount(skb); | |
2194 | ||
2195 | /* Save stamp of the first retransmit. */ | |
2196 | if (!tp->retrans_stamp) | |
2197 | tp->retrans_stamp = TCP_SKB_CB(skb)->when; | |
2198 | ||
c24f691b | 2199 | tp->undo_retrans += tcp_skb_pcount(skb); |
1da177e4 LT |
2200 | |
2201 | /* snd_nxt is stored to detect loss of retransmitted segment, | |
2202 | * see tcp_input.c tcp_sacktag_write_queue(). | |
2203 | */ | |
2204 | TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; | |
2205 | } | |
2206 | return err; | |
2207 | } | |
2208 | ||
67edfef7 AK |
2209 | /* Check if we forward retransmits are possible in the current |
2210 | * window/congestion state. | |
2211 | */ | |
a2a385d6 | 2212 | static bool tcp_can_forward_retransmit(struct sock *sk) |
b5afe7bc IJ |
2213 | { |
2214 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
cf533ea5 | 2215 | const struct tcp_sock *tp = tcp_sk(sk); |
b5afe7bc IJ |
2216 | |
2217 | /* Forward retransmissions are possible only during Recovery. */ | |
2218 | if (icsk->icsk_ca_state != TCP_CA_Recovery) | |
a2a385d6 | 2219 | return false; |
b5afe7bc IJ |
2220 | |
2221 | /* No forward retransmissions in Reno are possible. */ | |
2222 | if (tcp_is_reno(tp)) | |
a2a385d6 | 2223 | return false; |
b5afe7bc IJ |
2224 | |
2225 | /* Yeah, we have to make difficult choice between forward transmission | |
2226 | * and retransmission... Both ways have their merits... | |
2227 | * | |
2228 | * For now we do not retransmit anything, while we have some new | |
2229 | * segments to send. In the other cases, follow rule 3 for | |
2230 | * NextSeg() specified in RFC3517. | |
2231 | */ | |
2232 | ||
2233 | if (tcp_may_send_now(sk)) | |
a2a385d6 | 2234 | return false; |
b5afe7bc | 2235 | |
a2a385d6 | 2236 | return true; |
b5afe7bc IJ |
2237 | } |
2238 | ||
1da177e4 LT |
2239 | /* This gets called after a retransmit timeout, and the initially |
2240 | * retransmitted data is acknowledged. It tries to continue | |
2241 | * resending the rest of the retransmit queue, until either | |
2242 | * we've sent it all or the congestion window limit is reached. | |
2243 | * If doing SACK, the first ACK which comes back for a timeout | |
2244 | * based retransmit packet might feed us FACK information again. | |
2245 | * If so, we use it to avoid unnecessarily retransmissions. | |
2246 | */ | |
2247 | void tcp_xmit_retransmit_queue(struct sock *sk) | |
2248 | { | |
6687e988 | 2249 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2250 | struct tcp_sock *tp = tcp_sk(sk); |
2251 | struct sk_buff *skb; | |
0e1c54c2 | 2252 | struct sk_buff *hole = NULL; |
618d9f25 | 2253 | u32 last_lost; |
61eb55f4 | 2254 | int mib_idx; |
0e1c54c2 | 2255 | int fwd_rexmitting = 0; |
6a438bbe | 2256 | |
45e77d31 IJ |
2257 | if (!tp->packets_out) |
2258 | return; | |
2259 | ||
08ebd172 IJ |
2260 | if (!tp->lost_out) |
2261 | tp->retransmit_high = tp->snd_una; | |
2262 | ||
618d9f25 | 2263 | if (tp->retransmit_skb_hint) { |
6a438bbe | 2264 | skb = tp->retransmit_skb_hint; |
618d9f25 IJ |
2265 | last_lost = TCP_SKB_CB(skb)->end_seq; |
2266 | if (after(last_lost, tp->retransmit_high)) | |
2267 | last_lost = tp->retransmit_high; | |
2268 | } else { | |
fe067e8a | 2269 | skb = tcp_write_queue_head(sk); |
618d9f25 IJ |
2270 | last_lost = tp->snd_una; |
2271 | } | |
1da177e4 | 2272 | |
08ebd172 IJ |
2273 | tcp_for_write_queue_from(skb, sk) { |
2274 | __u8 sacked = TCP_SKB_CB(skb)->sacked; | |
1da177e4 | 2275 | |
08ebd172 IJ |
2276 | if (skb == tcp_send_head(sk)) |
2277 | break; | |
2278 | /* we could do better than to assign each time */ | |
0e1c54c2 IJ |
2279 | if (hole == NULL) |
2280 | tp->retransmit_skb_hint = skb; | |
08ebd172 IJ |
2281 | |
2282 | /* Assume this retransmit will generate | |
2283 | * only one packet for congestion window | |
2284 | * calculation purposes. This works because | |
2285 | * tcp_retransmit_skb() will chop up the | |
2286 | * packet to be MSS sized and all the | |
2287 | * packet counting works out. | |
2288 | */ | |
2289 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) | |
2290 | return; | |
1da177e4 | 2291 | |
0e1c54c2 IJ |
2292 | if (fwd_rexmitting) { |
2293 | begin_fwd: | |
2294 | if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) | |
2295 | break; | |
2296 | mib_idx = LINUX_MIB_TCPFORWARDRETRANS; | |
6a438bbe | 2297 | |
0e1c54c2 | 2298 | } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { |
618d9f25 | 2299 | tp->retransmit_high = last_lost; |
0e1c54c2 IJ |
2300 | if (!tcp_can_forward_retransmit(sk)) |
2301 | break; | |
2302 | /* Backtrack if necessary to non-L'ed skb */ | |
2303 | if (hole != NULL) { | |
2304 | skb = hole; | |
2305 | hole = NULL; | |
2306 | } | |
2307 | fwd_rexmitting = 1; | |
2308 | goto begin_fwd; | |
1da177e4 | 2309 | |
0e1c54c2 | 2310 | } else if (!(sacked & TCPCB_LOST)) { |
ac11ba75 | 2311 | if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) |
0e1c54c2 IJ |
2312 | hole = skb; |
2313 | continue; | |
1da177e4 | 2314 | |
0e1c54c2 | 2315 | } else { |
618d9f25 | 2316 | last_lost = TCP_SKB_CB(skb)->end_seq; |
0e1c54c2 IJ |
2317 | if (icsk->icsk_ca_state != TCP_CA_Loss) |
2318 | mib_idx = LINUX_MIB_TCPFASTRETRANS; | |
2319 | else | |
2320 | mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; | |
2321 | } | |
1da177e4 | 2322 | |
0e1c54c2 | 2323 | if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) |
1da177e4 LT |
2324 | continue; |
2325 | ||
09e9b813 ED |
2326 | if (tcp_retransmit_skb(sk, skb)) { |
2327 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); | |
0e1c54c2 | 2328 | return; |
09e9b813 | 2329 | } |
0e1c54c2 | 2330 | NET_INC_STATS_BH(sock_net(sk), mib_idx); |
1da177e4 | 2331 | |
a262f0cd ND |
2332 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery) |
2333 | tp->prr_out += tcp_skb_pcount(skb); | |
2334 | ||
fe067e8a | 2335 | if (skb == tcp_write_queue_head(sk)) |
3f421baa ACM |
2336 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
2337 | inet_csk(sk)->icsk_rto, | |
2338 | TCP_RTO_MAX); | |
1da177e4 LT |
2339 | } |
2340 | } | |
2341 | ||
1da177e4 LT |
2342 | /* Send a fin. The caller locks the socket for us. This cannot be |
2343 | * allowed to fail queueing a FIN frame under any circumstances. | |
2344 | */ | |
2345 | void tcp_send_fin(struct sock *sk) | |
2346 | { | |
e905a9ed | 2347 | struct tcp_sock *tp = tcp_sk(sk); |
fe067e8a | 2348 | struct sk_buff *skb = tcp_write_queue_tail(sk); |
1da177e4 | 2349 | int mss_now; |
e905a9ed | 2350 | |
1da177e4 LT |
2351 | /* Optimization, tack on the FIN if we have a queue of |
2352 | * unsent frames. But be careful about outgoing SACKS | |
2353 | * and IP options. | |
2354 | */ | |
0c54b85f | 2355 | mss_now = tcp_current_mss(sk); |
1da177e4 | 2356 | |
fe067e8a | 2357 | if (tcp_send_head(sk) != NULL) { |
4de075e0 | 2358 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN; |
1da177e4 LT |
2359 | TCP_SKB_CB(skb)->end_seq++; |
2360 | tp->write_seq++; | |
2361 | } else { | |
2362 | /* Socket is locked, keep trying until memory is available. */ | |
2363 | for (;;) { | |
aa133076 WF |
2364 | skb = alloc_skb_fclone(MAX_TCP_HEADER, |
2365 | sk->sk_allocation); | |
1da177e4 LT |
2366 | if (skb) |
2367 | break; | |
2368 | yield(); | |
2369 | } | |
2370 | ||
2371 | /* Reserve space for headers and prepare control bits. */ | |
2372 | skb_reserve(skb, MAX_TCP_HEADER); | |
1da177e4 | 2373 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ |
e870a8ef | 2374 | tcp_init_nondata_skb(skb, tp->write_seq, |
a3433f35 | 2375 | TCPHDR_ACK | TCPHDR_FIN); |
1da177e4 LT |
2376 | tcp_queue_skb(sk, skb); |
2377 | } | |
9e412ba7 | 2378 | __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); |
1da177e4 LT |
2379 | } |
2380 | ||
2381 | /* We get here when a process closes a file descriptor (either due to | |
2382 | * an explicit close() or as a byproduct of exit()'ing) and there | |
2383 | * was unread data in the receive queue. This behavior is recommended | |
65bb723c | 2384 | * by RFC 2525, section 2.17. -DaveM |
1da177e4 | 2385 | */ |
dd0fc66f | 2386 | void tcp_send_active_reset(struct sock *sk, gfp_t priority) |
1da177e4 | 2387 | { |
1da177e4 LT |
2388 | struct sk_buff *skb; |
2389 | ||
2390 | /* NOTE: No TCP options attached and we never retransmit this. */ | |
2391 | skb = alloc_skb(MAX_TCP_HEADER, priority); | |
2392 | if (!skb) { | |
4e673444 | 2393 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
1da177e4 LT |
2394 | return; |
2395 | } | |
2396 | ||
2397 | /* Reserve space for headers and prepare control bits. */ | |
2398 | skb_reserve(skb, MAX_TCP_HEADER); | |
e870a8ef | 2399 | tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), |
a3433f35 | 2400 | TCPHDR_ACK | TCPHDR_RST); |
1da177e4 | 2401 | /* Send it off. */ |
1da177e4 | 2402 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
dfb4b9dc | 2403 | if (tcp_transmit_skb(sk, skb, 0, priority)) |
4e673444 | 2404 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
26af65cb | 2405 | |
81cc8a75 | 2406 | TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); |
1da177e4 LT |
2407 | } |
2408 | ||
67edfef7 AK |
2409 | /* Send a crossed SYN-ACK during socket establishment. |
2410 | * WARNING: This routine must only be called when we have already sent | |
1da177e4 LT |
2411 | * a SYN packet that crossed the incoming SYN that caused this routine |
2412 | * to get called. If this assumption fails then the initial rcv_wnd | |
2413 | * and rcv_wscale values will not be correct. | |
2414 | */ | |
2415 | int tcp_send_synack(struct sock *sk) | |
2416 | { | |
056834d9 | 2417 | struct sk_buff *skb; |
1da177e4 | 2418 | |
fe067e8a | 2419 | skb = tcp_write_queue_head(sk); |
4de075e0 | 2420 | if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { |
91df42be | 2421 | pr_debug("%s: wrong queue state\n", __func__); |
1da177e4 LT |
2422 | return -EFAULT; |
2423 | } | |
4de075e0 | 2424 | if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { |
1da177e4 LT |
2425 | if (skb_cloned(skb)) { |
2426 | struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); | |
2427 | if (nskb == NULL) | |
2428 | return -ENOMEM; | |
fe067e8a | 2429 | tcp_unlink_write_queue(skb, sk); |
1da177e4 | 2430 | skb_header_release(nskb); |
fe067e8a | 2431 | __tcp_add_write_queue_head(sk, nskb); |
3ab224be HA |
2432 | sk_wmem_free_skb(sk, skb); |
2433 | sk->sk_wmem_queued += nskb->truesize; | |
2434 | sk_mem_charge(sk, nskb->truesize); | |
1da177e4 LT |
2435 | skb = nskb; |
2436 | } | |
2437 | ||
4de075e0 | 2438 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; |
1da177e4 LT |
2439 | TCP_ECN_send_synack(tcp_sk(sk), skb); |
2440 | } | |
2441 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
dfb4b9dc | 2442 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
1da177e4 LT |
2443 | } |
2444 | ||
67edfef7 | 2445 | /* Prepare a SYN-ACK. */ |
056834d9 | 2446 | struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, |
e6b4d113 WAS |
2447 | struct request_sock *req, |
2448 | struct request_values *rvp) | |
1da177e4 | 2449 | { |
bd0388ae | 2450 | struct tcp_out_options opts; |
4957faad | 2451 | struct tcp_extend_values *xvp = tcp_xv(rvp); |
2e6599cb | 2452 | struct inet_request_sock *ireq = inet_rsk(req); |
1da177e4 | 2453 | struct tcp_sock *tp = tcp_sk(sk); |
28b2774a | 2454 | const struct tcp_cookie_values *cvp = tp->cookie_values; |
1da177e4 | 2455 | struct tcphdr *th; |
1da177e4 | 2456 | struct sk_buff *skb; |
cfb6eeb4 | 2457 | struct tcp_md5sig_key *md5; |
bd0388ae | 2458 | int tcp_header_size; |
f5fff5dc | 2459 | int mss; |
28b2774a | 2460 | int s_data_desired = 0; |
1da177e4 | 2461 | |
28b2774a ED |
2462 | if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired) |
2463 | s_data_desired = cvp->s_data_desired; | |
2464 | skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15 + s_data_desired, 1, GFP_ATOMIC); | |
1da177e4 LT |
2465 | if (skb == NULL) |
2466 | return NULL; | |
2467 | ||
2468 | /* Reserve space for headers. */ | |
2469 | skb_reserve(skb, MAX_TCP_HEADER); | |
2470 | ||
adf30907 | 2471 | skb_dst_set(skb, dst_clone(dst)); |
1da177e4 | 2472 | |
0dbaee3b | 2473 | mss = dst_metric_advmss(dst); |
f5fff5dc TQ |
2474 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) |
2475 | mss = tp->rx_opt.user_mss; | |
2476 | ||
33ad798c AL |
2477 | if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ |
2478 | __u8 rcv_wscale; | |
2479 | /* Set this up on the first call only */ | |
2480 | req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); | |
e88c64f0 HPP |
2481 | |
2482 | /* limit the window selection if the user enforce a smaller rx buffer */ | |
2483 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && | |
2484 | (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) | |
2485 | req->window_clamp = tcp_full_space(sk); | |
2486 | ||
33ad798c AL |
2487 | /* tcp_full_space because it is guaranteed to be the first packet */ |
2488 | tcp_select_initial_window(tcp_full_space(sk), | |
f5fff5dc | 2489 | mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), |
33ad798c AL |
2490 | &req->rcv_wnd, |
2491 | &req->window_clamp, | |
2492 | ireq->wscale_ok, | |
31d12926 | 2493 | &rcv_wscale, |
2494 | dst_metric(dst, RTAX_INITRWND)); | |
33ad798c AL |
2495 | ireq->rcv_wscale = rcv_wscale; |
2496 | } | |
2497 | ||
2498 | memset(&opts, 0, sizeof(opts)); | |
8b5f12d0 FW |
2499 | #ifdef CONFIG_SYN_COOKIES |
2500 | if (unlikely(req->cookie_ts)) | |
2501 | TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); | |
2502 | else | |
2503 | #endif | |
33ad798c | 2504 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
f5fff5dc | 2505 | tcp_header_size = tcp_synack_options(sk, req, mss, |
4957faad WAS |
2506 | skb, &opts, &md5, xvp) |
2507 | + sizeof(*th); | |
cfb6eeb4 | 2508 | |
aa8223c7 ACM |
2509 | skb_push(skb, tcp_header_size); |
2510 | skb_reset_transport_header(skb); | |
1da177e4 | 2511 | |
aa8223c7 | 2512 | th = tcp_hdr(skb); |
1da177e4 LT |
2513 | memset(th, 0, sizeof(struct tcphdr)); |
2514 | th->syn = 1; | |
2515 | th->ack = 1; | |
1da177e4 | 2516 | TCP_ECN_make_synack(req, th); |
a3116ac5 | 2517 | th->source = ireq->loc_port; |
2e6599cb | 2518 | th->dest = ireq->rmt_port; |
e870a8ef IJ |
2519 | /* Setting of flags are superfluous here for callers (and ECE is |
2520 | * not even correctly set) | |
2521 | */ | |
2522 | tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, | |
a3433f35 | 2523 | TCPHDR_SYN | TCPHDR_ACK); |
4957faad WAS |
2524 | |
2525 | if (OPTION_COOKIE_EXTENSION & opts.options) { | |
28b2774a ED |
2526 | if (s_data_desired) { |
2527 | u8 *buf = skb_put(skb, s_data_desired); | |
4957faad WAS |
2528 | |
2529 | /* copy data directly from the listening socket. */ | |
28b2774a ED |
2530 | memcpy(buf, cvp->s_data_payload, s_data_desired); |
2531 | TCP_SKB_CB(skb)->end_seq += s_data_desired; | |
4957faad WAS |
2532 | } |
2533 | ||
2534 | if (opts.hash_size > 0) { | |
2535 | __u32 workspace[SHA_WORKSPACE_WORDS]; | |
2536 | u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS]; | |
2537 | u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1]; | |
2538 | ||
2539 | /* Secret recipe depends on the Timestamp, (future) | |
2540 | * Sequence and Acknowledgment Numbers, Initiator | |
2541 | * Cookie, and others handled by IP variant caller. | |
2542 | */ | |
2543 | *tail-- ^= opts.tsval; | |
2544 | *tail-- ^= tcp_rsk(req)->rcv_isn + 1; | |
2545 | *tail-- ^= TCP_SKB_CB(skb)->seq + 1; | |
2546 | ||
2547 | /* recommended */ | |
0eae88f3 | 2548 | *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source); |
f9a2e69e | 2549 | *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */ |
4957faad WAS |
2550 | |
2551 | sha_transform((__u32 *)&xvp->cookie_bakery[0], | |
2552 | (char *)mess, | |
2553 | &workspace[0]); | |
2554 | opts.hash_location = | |
2555 | (__u8 *)&xvp->cookie_bakery[0]; | |
2556 | } | |
2557 | } | |
2558 | ||
1da177e4 | 2559 | th->seq = htonl(TCP_SKB_CB(skb)->seq); |
2e6599cb | 2560 | th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); |
1da177e4 LT |
2561 | |
2562 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ | |
600ff0c2 | 2563 | th->window = htons(min(req->rcv_wnd, 65535U)); |
bd0388ae | 2564 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
1da177e4 | 2565 | th->doff = (tcp_header_size >> 2); |
aa2ea058 | 2566 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb)); |
cfb6eeb4 YH |
2567 | |
2568 | #ifdef CONFIG_TCP_MD5SIG | |
2569 | /* Okay, we have all we need - do the md5 hash if needed */ | |
2570 | if (md5) { | |
bd0388ae | 2571 | tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, |
49a72dfb | 2572 | md5, NULL, req, skb); |
cfb6eeb4 YH |
2573 | } |
2574 | #endif | |
2575 | ||
1da177e4 LT |
2576 | return skb; |
2577 | } | |
4bc2f18b | 2578 | EXPORT_SYMBOL(tcp_make_synack); |
1da177e4 | 2579 | |
67edfef7 | 2580 | /* Do all connect socket setups that can be done AF independent. */ |
370816ae | 2581 | void tcp_connect_init(struct sock *sk) |
1da177e4 | 2582 | { |
cf533ea5 | 2583 | const struct dst_entry *dst = __sk_dst_get(sk); |
1da177e4 LT |
2584 | struct tcp_sock *tp = tcp_sk(sk); |
2585 | __u8 rcv_wscale; | |
2586 | ||
2587 | /* We'll fix this up when we get a response from the other end. | |
2588 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. | |
2589 | */ | |
2590 | tp->tcp_header_len = sizeof(struct tcphdr) + | |
bb5b7c11 | 2591 | (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); |
1da177e4 | 2592 | |
cfb6eeb4 YH |
2593 | #ifdef CONFIG_TCP_MD5SIG |
2594 | if (tp->af_specific->md5_lookup(sk, sk) != NULL) | |
2595 | tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; | |
2596 | #endif | |
2597 | ||
1da177e4 LT |
2598 | /* If user gave his TCP_MAXSEG, record it to clamp */ |
2599 | if (tp->rx_opt.user_mss) | |
2600 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
2601 | tp->max_window = 0; | |
5d424d5a | 2602 | tcp_mtup_init(sk); |
1da177e4 LT |
2603 | tcp_sync_mss(sk, dst_mtu(dst)); |
2604 | ||
2605 | if (!tp->window_clamp) | |
2606 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); | |
0dbaee3b | 2607 | tp->advmss = dst_metric_advmss(dst); |
f5fff5dc TQ |
2608 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) |
2609 | tp->advmss = tp->rx_opt.user_mss; | |
2610 | ||
1da177e4 | 2611 | tcp_initialize_rcv_mss(sk); |
1da177e4 | 2612 | |
e88c64f0 HPP |
2613 | /* limit the window selection if the user enforce a smaller rx buffer */ |
2614 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && | |
2615 | (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) | |
2616 | tp->window_clamp = tcp_full_space(sk); | |
2617 | ||
1da177e4 LT |
2618 | tcp_select_initial_window(tcp_full_space(sk), |
2619 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), | |
2620 | &tp->rcv_wnd, | |
2621 | &tp->window_clamp, | |
bb5b7c11 | 2622 | sysctl_tcp_window_scaling, |
31d12926 | 2623 | &rcv_wscale, |
2624 | dst_metric(dst, RTAX_INITRWND)); | |
1da177e4 LT |
2625 | |
2626 | tp->rx_opt.rcv_wscale = rcv_wscale; | |
2627 | tp->rcv_ssthresh = tp->rcv_wnd; | |
2628 | ||
2629 | sk->sk_err = 0; | |
2630 | sock_reset_flag(sk, SOCK_DONE); | |
2631 | tp->snd_wnd = 0; | |
ee7537b6 | 2632 | tcp_init_wl(tp, 0); |
1da177e4 LT |
2633 | tp->snd_una = tp->write_seq; |
2634 | tp->snd_sml = tp->write_seq; | |
33f5f57e | 2635 | tp->snd_up = tp->write_seq; |
370816ae | 2636 | tp->snd_nxt = tp->write_seq; |
ee995283 PE |
2637 | |
2638 | if (likely(!tp->repair)) | |
2639 | tp->rcv_nxt = 0; | |
2640 | tp->rcv_wup = tp->rcv_nxt; | |
2641 | tp->copied_seq = tp->rcv_nxt; | |
1da177e4 | 2642 | |
463c84b9 ACM |
2643 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
2644 | inet_csk(sk)->icsk_retransmits = 0; | |
1da177e4 LT |
2645 | tcp_clear_retrans(tp); |
2646 | } | |
2647 | ||
67edfef7 | 2648 | /* Build a SYN and send it off. */ |
1da177e4 LT |
2649 | int tcp_connect(struct sock *sk) |
2650 | { | |
2651 | struct tcp_sock *tp = tcp_sk(sk); | |
2652 | struct sk_buff *buff; | |
ee586811 | 2653 | int err; |
1da177e4 LT |
2654 | |
2655 | tcp_connect_init(sk); | |
2656 | ||
d179cd12 | 2657 | buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); |
1da177e4 LT |
2658 | if (unlikely(buff == NULL)) |
2659 | return -ENOBUFS; | |
2660 | ||
2661 | /* Reserve space for headers. */ | |
2662 | skb_reserve(buff, MAX_TCP_HEADER); | |
2663 | ||
a3433f35 | 2664 | tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); |
e870a8ef | 2665 | TCP_ECN_send_syn(sk, buff); |
1da177e4 LT |
2666 | |
2667 | /* Send it off. */ | |
2668 | TCP_SKB_CB(buff)->when = tcp_time_stamp; | |
2669 | tp->retrans_stamp = TCP_SKB_CB(buff)->when; | |
2670 | skb_header_release(buff); | |
fe067e8a | 2671 | __tcp_add_write_queue_tail(sk, buff); |
3ab224be HA |
2672 | sk->sk_wmem_queued += buff->truesize; |
2673 | sk_mem_charge(sk, buff->truesize); | |
1da177e4 | 2674 | tp->packets_out += tcp_skb_pcount(buff); |
ee586811 EP |
2675 | err = tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); |
2676 | if (err == -ECONNREFUSED) | |
2677 | return err; | |
bd37a088 WY |
2678 | |
2679 | /* We change tp->snd_nxt after the tcp_transmit_skb() call | |
2680 | * in order to make this packet get counted in tcpOutSegs. | |
2681 | */ | |
2682 | tp->snd_nxt = tp->write_seq; | |
2683 | tp->pushed_seq = tp->write_seq; | |
81cc8a75 | 2684 | TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); |
1da177e4 LT |
2685 | |
2686 | /* Timer for repeating the SYN until an answer. */ | |
3f421baa ACM |
2687 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
2688 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
1da177e4 LT |
2689 | return 0; |
2690 | } | |
4bc2f18b | 2691 | EXPORT_SYMBOL(tcp_connect); |
1da177e4 LT |
2692 | |
2693 | /* Send out a delayed ack, the caller does the policy checking | |
2694 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() | |
2695 | * for details. | |
2696 | */ | |
2697 | void tcp_send_delayed_ack(struct sock *sk) | |
2698 | { | |
463c84b9 ACM |
2699 | struct inet_connection_sock *icsk = inet_csk(sk); |
2700 | int ato = icsk->icsk_ack.ato; | |
1da177e4 LT |
2701 | unsigned long timeout; |
2702 | ||
2703 | if (ato > TCP_DELACK_MIN) { | |
463c84b9 | 2704 | const struct tcp_sock *tp = tcp_sk(sk); |
056834d9 | 2705 | int max_ato = HZ / 2; |
1da177e4 | 2706 | |
056834d9 IJ |
2707 | if (icsk->icsk_ack.pingpong || |
2708 | (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) | |
1da177e4 LT |
2709 | max_ato = TCP_DELACK_MAX; |
2710 | ||
2711 | /* Slow path, intersegment interval is "high". */ | |
2712 | ||
2713 | /* If some rtt estimate is known, use it to bound delayed ack. | |
463c84b9 | 2714 | * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements |
1da177e4 LT |
2715 | * directly. |
2716 | */ | |
2717 | if (tp->srtt) { | |
056834d9 | 2718 | int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); |
1da177e4 LT |
2719 | |
2720 | if (rtt < max_ato) | |
2721 | max_ato = rtt; | |
2722 | } | |
2723 | ||
2724 | ato = min(ato, max_ato); | |
2725 | } | |
2726 | ||
2727 | /* Stay within the limit we were given */ | |
2728 | timeout = jiffies + ato; | |
2729 | ||
2730 | /* Use new timeout only if there wasn't a older one earlier. */ | |
463c84b9 | 2731 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { |
1da177e4 LT |
2732 | /* If delack timer was blocked or is about to expire, |
2733 | * send ACK now. | |
2734 | */ | |
463c84b9 ACM |
2735 | if (icsk->icsk_ack.blocked || |
2736 | time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { | |
1da177e4 LT |
2737 | tcp_send_ack(sk); |
2738 | return; | |
2739 | } | |
2740 | ||
463c84b9 ACM |
2741 | if (!time_before(timeout, icsk->icsk_ack.timeout)) |
2742 | timeout = icsk->icsk_ack.timeout; | |
1da177e4 | 2743 | } |
463c84b9 ACM |
2744 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; |
2745 | icsk->icsk_ack.timeout = timeout; | |
2746 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); | |
1da177e4 LT |
2747 | } |
2748 | ||
2749 | /* This routine sends an ack and also updates the window. */ | |
2750 | void tcp_send_ack(struct sock *sk) | |
2751 | { | |
058dc334 | 2752 | struct sk_buff *buff; |
1da177e4 | 2753 | |
058dc334 IJ |
2754 | /* If we have been reset, we may not send again. */ |
2755 | if (sk->sk_state == TCP_CLOSE) | |
2756 | return; | |
1da177e4 | 2757 | |
058dc334 IJ |
2758 | /* We are not putting this on the write queue, so |
2759 | * tcp_transmit_skb() will set the ownership to this | |
2760 | * sock. | |
2761 | */ | |
2762 | buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); | |
2763 | if (buff == NULL) { | |
2764 | inet_csk_schedule_ack(sk); | |
2765 | inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; | |
2766 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, | |
2767 | TCP_DELACK_MAX, TCP_RTO_MAX); | |
2768 | return; | |
1da177e4 | 2769 | } |
058dc334 IJ |
2770 | |
2771 | /* Reserve space for headers and prepare control bits. */ | |
2772 | skb_reserve(buff, MAX_TCP_HEADER); | |
a3433f35 | 2773 | tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); |
058dc334 IJ |
2774 | |
2775 | /* Send it off, this clears delayed acks for us. */ | |
058dc334 IJ |
2776 | TCP_SKB_CB(buff)->when = tcp_time_stamp; |
2777 | tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC); | |
1da177e4 LT |
2778 | } |
2779 | ||
2780 | /* This routine sends a packet with an out of date sequence | |
2781 | * number. It assumes the other end will try to ack it. | |
2782 | * | |
2783 | * Question: what should we make while urgent mode? | |
2784 | * 4.4BSD forces sending single byte of data. We cannot send | |
2785 | * out of window data, because we have SND.NXT==SND.MAX... | |
2786 | * | |
2787 | * Current solution: to send TWO zero-length segments in urgent mode: | |
2788 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is | |
2789 | * out-of-date with SND.UNA-1 to probe window. | |
2790 | */ | |
2791 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent) | |
2792 | { | |
2793 | struct tcp_sock *tp = tcp_sk(sk); | |
2794 | struct sk_buff *skb; | |
2795 | ||
2796 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ | |
2797 | skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); | |
e905a9ed | 2798 | if (skb == NULL) |
1da177e4 LT |
2799 | return -1; |
2800 | ||
2801 | /* Reserve space for headers and set control bits. */ | |
2802 | skb_reserve(skb, MAX_TCP_HEADER); | |
1da177e4 LT |
2803 | /* Use a previous sequence. This should cause the other |
2804 | * end to send an ack. Don't queue or clone SKB, just | |
2805 | * send it. | |
2806 | */ | |
a3433f35 | 2807 | tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); |
1da177e4 | 2808 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
dfb4b9dc | 2809 | return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); |
1da177e4 LT |
2810 | } |
2811 | ||
ee995283 PE |
2812 | void tcp_send_window_probe(struct sock *sk) |
2813 | { | |
2814 | if (sk->sk_state == TCP_ESTABLISHED) { | |
2815 | tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; | |
c0e88ff0 | 2816 | tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq; |
ee995283 PE |
2817 | tcp_xmit_probe_skb(sk, 0); |
2818 | } | |
2819 | } | |
2820 | ||
67edfef7 | 2821 | /* Initiate keepalive or window probe from timer. */ |
1da177e4 LT |
2822 | int tcp_write_wakeup(struct sock *sk) |
2823 | { | |
058dc334 IJ |
2824 | struct tcp_sock *tp = tcp_sk(sk); |
2825 | struct sk_buff *skb; | |
1da177e4 | 2826 | |
058dc334 IJ |
2827 | if (sk->sk_state == TCP_CLOSE) |
2828 | return -1; | |
2829 | ||
2830 | if ((skb = tcp_send_head(sk)) != NULL && | |
2831 | before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { | |
2832 | int err; | |
0c54b85f | 2833 | unsigned int mss = tcp_current_mss(sk); |
058dc334 IJ |
2834 | unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
2835 | ||
2836 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) | |
2837 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; | |
2838 | ||
2839 | /* We are probing the opening of a window | |
2840 | * but the window size is != 0 | |
2841 | * must have been a result SWS avoidance ( sender ) | |
2842 | */ | |
2843 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || | |
2844 | skb->len > mss) { | |
2845 | seg_size = min(seg_size, mss); | |
4de075e0 | 2846 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
058dc334 IJ |
2847 | if (tcp_fragment(sk, skb, seg_size, mss)) |
2848 | return -1; | |
2849 | } else if (!tcp_skb_pcount(skb)) | |
2850 | tcp_set_skb_tso_segs(sk, skb, mss); | |
2851 | ||
4de075e0 | 2852 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
058dc334 IJ |
2853 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
2854 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); | |
2855 | if (!err) | |
2856 | tcp_event_new_data_sent(sk, skb); | |
2857 | return err; | |
2858 | } else { | |
33f5f57e | 2859 | if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) |
058dc334 IJ |
2860 | tcp_xmit_probe_skb(sk, 1); |
2861 | return tcp_xmit_probe_skb(sk, 0); | |
1da177e4 | 2862 | } |
1da177e4 LT |
2863 | } |
2864 | ||
2865 | /* A window probe timeout has occurred. If window is not closed send | |
2866 | * a partial packet else a zero probe. | |
2867 | */ | |
2868 | void tcp_send_probe0(struct sock *sk) | |
2869 | { | |
463c84b9 | 2870 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2871 | struct tcp_sock *tp = tcp_sk(sk); |
2872 | int err; | |
2873 | ||
2874 | err = tcp_write_wakeup(sk); | |
2875 | ||
fe067e8a | 2876 | if (tp->packets_out || !tcp_send_head(sk)) { |
1da177e4 | 2877 | /* Cancel probe timer, if it is not required. */ |
6687e988 | 2878 | icsk->icsk_probes_out = 0; |
463c84b9 | 2879 | icsk->icsk_backoff = 0; |
1da177e4 LT |
2880 | return; |
2881 | } | |
2882 | ||
2883 | if (err <= 0) { | |
463c84b9 ACM |
2884 | if (icsk->icsk_backoff < sysctl_tcp_retries2) |
2885 | icsk->icsk_backoff++; | |
6687e988 | 2886 | icsk->icsk_probes_out++; |
e905a9ed | 2887 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
3f421baa ACM |
2888 | min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), |
2889 | TCP_RTO_MAX); | |
1da177e4 LT |
2890 | } else { |
2891 | /* If packet was not sent due to local congestion, | |
6687e988 | 2892 | * do not backoff and do not remember icsk_probes_out. |
1da177e4 LT |
2893 | * Let local senders to fight for local resources. |
2894 | * | |
2895 | * Use accumulated backoff yet. | |
2896 | */ | |
6687e988 ACM |
2897 | if (!icsk->icsk_probes_out) |
2898 | icsk->icsk_probes_out = 1; | |
e905a9ed | 2899 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
463c84b9 | 2900 | min(icsk->icsk_rto << icsk->icsk_backoff, |
3f421baa ACM |
2901 | TCP_RESOURCE_PROBE_INTERVAL), |
2902 | TCP_RTO_MAX); | |
1da177e4 LT |
2903 | } |
2904 | } |