Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * INET An implementation of the TCP/IP protocol suite for the LINUX | |
3 | * operating system. INET is implemented using the BSD Socket | |
4 | * interface as the means of communication with the user level. | |
5 | * | |
6 | * Implementation of the Transmission Control Protocol(TCP). | |
7 | * | |
02c30a84 | 8 | * Authors: Ross Biro |
1da177e4 LT |
9 | * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> |
10 | * Mark Evans, <evansmp@uhura.aston.ac.uk> | |
11 | * Corey Minyard <wf-rch!minyard@relay.EU.net> | |
12 | * Florian La Roche, <flla@stud.uni-sb.de> | |
13 | * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> | |
14 | * Linus Torvalds, <torvalds@cs.helsinki.fi> | |
15 | * Alan Cox, <gw4pts@gw4pts.ampr.org> | |
16 | * Matthew Dillon, <dillon@apollo.west.oic.com> | |
17 | * Arnt Gulbrandsen, <agulbra@nvg.unit.no> | |
18 | * Jorge Cwik, <jorge@laser.satlink.net> | |
19 | */ | |
20 | ||
21 | /* | |
22 | * Changes: Pedro Roque : Retransmit queue handled by TCP. | |
23 | * : Fragmentation on mtu decrease | |
24 | * : Segment collapse on retransmit | |
25 | * : AF independence | |
26 | * | |
27 | * Linus Torvalds : send_delayed_ack | |
28 | * David S. Miller : Charge memory using the right skb | |
29 | * during syn/ack processing. | |
30 | * David S. Miller : Output engine completely rewritten. | |
31 | * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. | |
32 | * Cacophonix Gaul : draft-minshall-nagle-01 | |
33 | * J Hadi Salim : ECN support | |
34 | * | |
35 | */ | |
36 | ||
37 | #include <net/tcp.h> | |
38 | ||
39 | #include <linux/compiler.h> | |
40 | #include <linux/module.h> | |
1da177e4 LT |
41 | |
42 | /* People can turn this off for buggy TCP's found in printers etc. */ | |
ab32ea5d | 43 | int sysctl_tcp_retrans_collapse __read_mostly = 1; |
1da177e4 | 44 | |
15d99e02 RJ |
45 | /* People can turn this on to work with those rare, broken TCPs that |
46 | * interpret the window field as a signed quantity. | |
47 | */ | |
ab32ea5d | 48 | int sysctl_tcp_workaround_signed_windows __read_mostly = 0; |
15d99e02 | 49 | |
1da177e4 LT |
50 | /* This limits the percentage of the congestion window which we |
51 | * will allow a single TSO frame to consume. Building TSO frames | |
52 | * which are too large can cause TCP streams to be bursty. | |
53 | */ | |
ab32ea5d | 54 | int sysctl_tcp_tso_win_divisor __read_mostly = 3; |
1da177e4 | 55 | |
ab32ea5d BH |
56 | int sysctl_tcp_mtu_probing __read_mostly = 0; |
57 | int sysctl_tcp_base_mss __read_mostly = 512; | |
5d424d5a | 58 | |
35089bb2 | 59 | /* By default, RFC2861 behavior. */ |
ab32ea5d | 60 | int sysctl_tcp_slow_start_after_idle __read_mostly = 1; |
35089bb2 | 61 | |
66f5fe62 | 62 | static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 63 | { |
9e412ba7 | 64 | struct tcp_sock *tp = tcp_sk(sk); |
66f5fe62 | 65 | unsigned int prior_packets = tp->packets_out; |
9e412ba7 | 66 | |
fe067e8a | 67 | tcp_advance_send_head(sk, skb); |
1da177e4 | 68 | tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; |
8512430e IJ |
69 | |
70 | /* Don't override Nagle indefinately with F-RTO */ | |
71 | if (tp->frto_counter == 2) | |
72 | tp->frto_counter = 3; | |
66f5fe62 IJ |
73 | |
74 | tp->packets_out += tcp_skb_pcount(skb); | |
75 | if (!prior_packets) | |
76 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, | |
77 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
1da177e4 LT |
78 | } |
79 | ||
80 | /* SND.NXT, if window was not shrunk. | |
81 | * If window has been shrunk, what should we make? It is not clear at all. | |
82 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( | |
83 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already | |
84 | * invalid. OK, let's make this for now: | |
85 | */ | |
9e412ba7 | 86 | static inline __u32 tcp_acceptable_seq(struct sock *sk) |
1da177e4 | 87 | { |
9e412ba7 IJ |
88 | struct tcp_sock *tp = tcp_sk(sk); |
89 | ||
90840def | 90 | if (!before(tcp_wnd_end(tp), tp->snd_nxt)) |
1da177e4 LT |
91 | return tp->snd_nxt; |
92 | else | |
90840def | 93 | return tcp_wnd_end(tp); |
1da177e4 LT |
94 | } |
95 | ||
96 | /* Calculate mss to advertise in SYN segment. | |
97 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: | |
98 | * | |
99 | * 1. It is independent of path mtu. | |
100 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. | |
101 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of | |
102 | * attached devices, because some buggy hosts are confused by | |
103 | * large MSS. | |
104 | * 4. We do not make 3, we advertise MSS, calculated from first | |
105 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. | |
106 | * This may be overridden via information stored in routing table. | |
107 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, | |
108 | * probably even Jumbo". | |
109 | */ | |
110 | static __u16 tcp_advertise_mss(struct sock *sk) | |
111 | { | |
112 | struct tcp_sock *tp = tcp_sk(sk); | |
113 | struct dst_entry *dst = __sk_dst_get(sk); | |
114 | int mss = tp->advmss; | |
115 | ||
116 | if (dst && dst_metric(dst, RTAX_ADVMSS) < mss) { | |
117 | mss = dst_metric(dst, RTAX_ADVMSS); | |
118 | tp->advmss = mss; | |
119 | } | |
120 | ||
121 | return (__u16)mss; | |
122 | } | |
123 | ||
124 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". | |
125 | * This is the first part of cwnd validation mechanism. */ | |
463c84b9 | 126 | static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst) |
1da177e4 | 127 | { |
463c84b9 | 128 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
129 | s32 delta = tcp_time_stamp - tp->lsndtime; |
130 | u32 restart_cwnd = tcp_init_cwnd(tp, dst); | |
131 | u32 cwnd = tp->snd_cwnd; | |
132 | ||
6687e988 | 133 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); |
1da177e4 | 134 | |
6687e988 | 135 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
1da177e4 LT |
136 | restart_cwnd = min(restart_cwnd, cwnd); |
137 | ||
463c84b9 | 138 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) |
1da177e4 LT |
139 | cwnd >>= 1; |
140 | tp->snd_cwnd = max(cwnd, restart_cwnd); | |
141 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
142 | tp->snd_cwnd_used = 0; | |
143 | } | |
144 | ||
40efc6fa SH |
145 | static void tcp_event_data_sent(struct tcp_sock *tp, |
146 | struct sk_buff *skb, struct sock *sk) | |
1da177e4 | 147 | { |
463c84b9 ACM |
148 | struct inet_connection_sock *icsk = inet_csk(sk); |
149 | const u32 now = tcp_time_stamp; | |
1da177e4 | 150 | |
35089bb2 DM |
151 | if (sysctl_tcp_slow_start_after_idle && |
152 | (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) | |
463c84b9 | 153 | tcp_cwnd_restart(sk, __sk_dst_get(sk)); |
1da177e4 LT |
154 | |
155 | tp->lsndtime = now; | |
156 | ||
157 | /* If it is a reply for ato after last received | |
158 | * packet, enter pingpong mode. | |
159 | */ | |
463c84b9 ACM |
160 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) |
161 | icsk->icsk_ack.pingpong = 1; | |
1da177e4 LT |
162 | } |
163 | ||
40efc6fa | 164 | static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) |
1da177e4 | 165 | { |
463c84b9 ACM |
166 | tcp_dec_quickack_mode(sk, pkts); |
167 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); | |
1da177e4 LT |
168 | } |
169 | ||
170 | /* Determine a window scaling and initial window to offer. | |
171 | * Based on the assumption that the given amount of space | |
172 | * will be offered. Store the results in the tp structure. | |
173 | * NOTE: for smooth operation initial space offering should | |
174 | * be a multiple of mss if possible. We assume here that mss >= 1. | |
175 | * This MUST be enforced by all callers. | |
176 | */ | |
177 | void tcp_select_initial_window(int __space, __u32 mss, | |
178 | __u32 *rcv_wnd, __u32 *window_clamp, | |
179 | int wscale_ok, __u8 *rcv_wscale) | |
180 | { | |
181 | unsigned int space = (__space < 0 ? 0 : __space); | |
182 | ||
183 | /* If no clamp set the clamp to the max possible scaled window */ | |
184 | if (*window_clamp == 0) | |
185 | (*window_clamp) = (65535 << 14); | |
186 | space = min(*window_clamp, space); | |
187 | ||
188 | /* Quantize space offering to a multiple of mss if possible. */ | |
189 | if (space > mss) | |
190 | space = (space / mss) * mss; | |
191 | ||
192 | /* NOTE: offering an initial window larger than 32767 | |
15d99e02 RJ |
193 | * will break some buggy TCP stacks. If the admin tells us |
194 | * it is likely we could be speaking with such a buggy stack | |
195 | * we will truncate our initial window offering to 32K-1 | |
196 | * unless the remote has sent us a window scaling option, | |
197 | * which we interpret as a sign the remote TCP is not | |
198 | * misinterpreting the window field as a signed quantity. | |
1da177e4 | 199 | */ |
15d99e02 RJ |
200 | if (sysctl_tcp_workaround_signed_windows) |
201 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); | |
202 | else | |
203 | (*rcv_wnd) = space; | |
204 | ||
1da177e4 LT |
205 | (*rcv_wscale) = 0; |
206 | if (wscale_ok) { | |
207 | /* Set window scaling on max possible window | |
e905a9ed | 208 | * See RFC1323 for an explanation of the limit to 14 |
1da177e4 LT |
209 | */ |
210 | space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); | |
316c1592 | 211 | space = min_t(u32, space, *window_clamp); |
1da177e4 LT |
212 | while (space > 65535 && (*rcv_wscale) < 14) { |
213 | space >>= 1; | |
214 | (*rcv_wscale)++; | |
215 | } | |
216 | } | |
217 | ||
218 | /* Set initial window to value enough for senders, | |
6b251858 | 219 | * following RFC2414. Senders, not following this RFC, |
1da177e4 LT |
220 | * will be satisfied with 2. |
221 | */ | |
056834d9 | 222 | if (mss > (1 << *rcv_wscale)) { |
01ff367e | 223 | int init_cwnd = 4; |
056834d9 | 224 | if (mss > 1460 * 3) |
1da177e4 | 225 | init_cwnd = 2; |
01ff367e DM |
226 | else if (mss > 1460) |
227 | init_cwnd = 3; | |
056834d9 IJ |
228 | if (*rcv_wnd > init_cwnd * mss) |
229 | *rcv_wnd = init_cwnd * mss; | |
1da177e4 LT |
230 | } |
231 | ||
232 | /* Set the clamp no higher than max representable value */ | |
233 | (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); | |
234 | } | |
235 | ||
236 | /* Chose a new window to advertise, update state in tcp_sock for the | |
237 | * socket, and return result with RFC1323 scaling applied. The return | |
238 | * value can be stuffed directly into th->window for an outgoing | |
239 | * frame. | |
240 | */ | |
40efc6fa | 241 | static u16 tcp_select_window(struct sock *sk) |
1da177e4 LT |
242 | { |
243 | struct tcp_sock *tp = tcp_sk(sk); | |
244 | u32 cur_win = tcp_receive_window(tp); | |
245 | u32 new_win = __tcp_select_window(sk); | |
246 | ||
247 | /* Never shrink the offered window */ | |
2de979bd | 248 | if (new_win < cur_win) { |
1da177e4 LT |
249 | /* Danger Will Robinson! |
250 | * Don't update rcv_wup/rcv_wnd here or else | |
251 | * we will not be able to advertise a zero | |
252 | * window in time. --DaveM | |
253 | * | |
254 | * Relax Will Robinson. | |
255 | */ | |
607bfbf2 | 256 | new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); |
1da177e4 LT |
257 | } |
258 | tp->rcv_wnd = new_win; | |
259 | tp->rcv_wup = tp->rcv_nxt; | |
260 | ||
261 | /* Make sure we do not exceed the maximum possible | |
262 | * scaled window. | |
263 | */ | |
15d99e02 | 264 | if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) |
1da177e4 LT |
265 | new_win = min(new_win, MAX_TCP_WINDOW); |
266 | else | |
267 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); | |
268 | ||
269 | /* RFC1323 scaling applied */ | |
270 | new_win >>= tp->rx_opt.rcv_wscale; | |
271 | ||
272 | /* If we advertise zero window, disable fast path. */ | |
273 | if (new_win == 0) | |
274 | tp->pred_flags = 0; | |
275 | ||
276 | return new_win; | |
277 | } | |
278 | ||
056834d9 | 279 | static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb) |
bdf1ee5d IJ |
280 | { |
281 | TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_CWR; | |
056834d9 | 282 | if (!(tp->ecn_flags & TCP_ECN_OK)) |
bdf1ee5d IJ |
283 | TCP_SKB_CB(skb)->flags &= ~TCPCB_FLAG_ECE; |
284 | } | |
285 | ||
286 | static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) | |
287 | { | |
288 | struct tcp_sock *tp = tcp_sk(sk); | |
289 | ||
290 | tp->ecn_flags = 0; | |
291 | if (sysctl_tcp_ecn) { | |
056834d9 | 292 | TCP_SKB_CB(skb)->flags |= TCPCB_FLAG_ECE | TCPCB_FLAG_CWR; |
bdf1ee5d IJ |
293 | tp->ecn_flags = TCP_ECN_OK; |
294 | } | |
295 | } | |
296 | ||
297 | static __inline__ void | |
298 | TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th) | |
299 | { | |
300 | if (inet_rsk(req)->ecn_ok) | |
301 | th->ece = 1; | |
302 | } | |
303 | ||
304 | static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, | |
305 | int tcp_header_len) | |
306 | { | |
307 | struct tcp_sock *tp = tcp_sk(sk); | |
308 | ||
309 | if (tp->ecn_flags & TCP_ECN_OK) { | |
310 | /* Not-retransmitted data segment: set ECT and inject CWR. */ | |
311 | if (skb->len != tcp_header_len && | |
312 | !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { | |
313 | INET_ECN_xmit(sk); | |
056834d9 | 314 | if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { |
bdf1ee5d IJ |
315 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; |
316 | tcp_hdr(skb)->cwr = 1; | |
317 | skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; | |
318 | } | |
319 | } else { | |
320 | /* ACK or retransmitted segment: clear ECT|CE */ | |
321 | INET_ECN_dontxmit(sk); | |
322 | } | |
323 | if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) | |
324 | tcp_hdr(skb)->ece = 1; | |
325 | } | |
326 | } | |
327 | ||
e870a8ef IJ |
328 | /* Constructs common control bits of non-data skb. If SYN/FIN is present, |
329 | * auto increment end seqno. | |
330 | */ | |
331 | static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) | |
332 | { | |
333 | skb->csum = 0; | |
334 | ||
335 | TCP_SKB_CB(skb)->flags = flags; | |
336 | TCP_SKB_CB(skb)->sacked = 0; | |
337 | ||
338 | skb_shinfo(skb)->gso_segs = 1; | |
339 | skb_shinfo(skb)->gso_size = 0; | |
340 | skb_shinfo(skb)->gso_type = 0; | |
341 | ||
342 | TCP_SKB_CB(skb)->seq = seq; | |
343 | if (flags & (TCPCB_FLAG_SYN | TCPCB_FLAG_FIN)) | |
344 | seq++; | |
345 | TCP_SKB_CB(skb)->end_seq = seq; | |
346 | } | |
347 | ||
33ad798c AL |
348 | #define OPTION_SACK_ADVERTISE (1 << 0) |
349 | #define OPTION_TS (1 << 1) | |
350 | #define OPTION_MD5 (1 << 2) | |
351 | ||
352 | struct tcp_out_options { | |
353 | u8 options; /* bit field of OPTION_* */ | |
354 | u8 ws; /* window scale, 0 to disable */ | |
355 | u8 num_sack_blocks; /* number of SACK blocks to include */ | |
356 | u16 mss; /* 0 to disable */ | |
357 | __u32 tsval, tsecr; /* need to include OPTION_TS */ | |
358 | }; | |
359 | ||
360 | static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, | |
361 | const struct tcp_out_options *opts, | |
362 | __u8 **md5_hash) { | |
363 | if (unlikely(OPTION_MD5 & opts->options)) { | |
496c98df YH |
364 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
365 | (TCPOPT_NOP << 16) | | |
33ad798c AL |
366 | (TCPOPT_MD5SIG << 8) | |
367 | TCPOLEN_MD5SIG); | |
368 | *md5_hash = (__u8 *)ptr; | |
369 | ptr += 4; | |
370 | } else { | |
371 | *md5_hash = NULL; | |
40efc6fa | 372 | } |
33ad798c AL |
373 | |
374 | if (likely(OPTION_TS & opts->options)) { | |
375 | if (unlikely(OPTION_SACK_ADVERTISE & opts->options)) { | |
376 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | | |
377 | (TCPOLEN_SACK_PERM << 16) | | |
378 | (TCPOPT_TIMESTAMP << 8) | | |
379 | TCPOLEN_TIMESTAMP); | |
380 | } else { | |
381 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
382 | (TCPOPT_NOP << 16) | | |
383 | (TCPOPT_TIMESTAMP << 8) | | |
384 | TCPOLEN_TIMESTAMP); | |
385 | } | |
386 | *ptr++ = htonl(opts->tsval); | |
387 | *ptr++ = htonl(opts->tsecr); | |
388 | } | |
389 | ||
390 | if (unlikely(opts->mss)) { | |
391 | *ptr++ = htonl((TCPOPT_MSS << 24) | | |
392 | (TCPOLEN_MSS << 16) | | |
393 | opts->mss); | |
394 | } | |
395 | ||
396 | if (unlikely(OPTION_SACK_ADVERTISE & opts->options && | |
397 | !(OPTION_TS & opts->options))) { | |
398 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
399 | (TCPOPT_NOP << 16) | | |
400 | (TCPOPT_SACK_PERM << 8) | | |
401 | TCPOLEN_SACK_PERM); | |
402 | } | |
403 | ||
404 | if (unlikely(opts->ws)) { | |
405 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
406 | (TCPOPT_WINDOW << 16) | | |
407 | (TCPOLEN_WINDOW << 8) | | |
408 | opts->ws); | |
409 | } | |
410 | ||
411 | if (unlikely(opts->num_sack_blocks)) { | |
412 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? | |
413 | tp->duplicate_sack : tp->selective_acks; | |
40efc6fa SH |
414 | int this_sack; |
415 | ||
416 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
417 | (TCPOPT_NOP << 16) | | |
418 | (TCPOPT_SACK << 8) | | |
33ad798c | 419 | (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * |
40efc6fa | 420 | TCPOLEN_SACK_PERBLOCK))); |
2de979bd | 421 | |
33ad798c AL |
422 | for (this_sack = 0; this_sack < opts->num_sack_blocks; |
423 | ++this_sack) { | |
40efc6fa SH |
424 | *ptr++ = htonl(sp[this_sack].start_seq); |
425 | *ptr++ = htonl(sp[this_sack].end_seq); | |
426 | } | |
2de979bd | 427 | |
40efc6fa SH |
428 | if (tp->rx_opt.dsack) { |
429 | tp->rx_opt.dsack = 0; | |
430 | tp->rx_opt.eff_sacks--; | |
431 | } | |
432 | } | |
33ad798c AL |
433 | } |
434 | ||
435 | static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb, | |
436 | struct tcp_out_options *opts, | |
437 | struct tcp_md5sig_key **md5) { | |
438 | struct tcp_sock *tp = tcp_sk(sk); | |
439 | unsigned size = 0; | |
440 | ||
cfb6eeb4 | 441 | #ifdef CONFIG_TCP_MD5SIG |
33ad798c AL |
442 | *md5 = tp->af_specific->md5_lookup(sk, sk); |
443 | if (*md5) { | |
444 | opts->options |= OPTION_MD5; | |
445 | size += TCPOLEN_MD5SIG_ALIGNED; | |
cfb6eeb4 | 446 | } |
33ad798c AL |
447 | #else |
448 | *md5 = NULL; | |
cfb6eeb4 | 449 | #endif |
33ad798c AL |
450 | |
451 | /* We always get an MSS option. The option bytes which will be seen in | |
452 | * normal data packets should timestamps be used, must be in the MSS | |
453 | * advertised. But we subtract them from tp->mss_cache so that | |
454 | * calculations in tcp_sendmsg are simpler etc. So account for this | |
455 | * fact here if necessary. If we don't do this correctly, as a | |
456 | * receiver we won't recognize data packets as being full sized when we | |
457 | * should, and thus we won't abide by the delayed ACK rules correctly. | |
458 | * SACKs don't matter, we never delay an ACK when we have any of those | |
459 | * going out. */ | |
460 | opts->mss = tcp_advertise_mss(sk); | |
461 | size += TCPOLEN_MSS_ALIGNED; | |
462 | ||
463 | if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { | |
464 | opts->options |= OPTION_TS; | |
465 | opts->tsval = TCP_SKB_CB(skb)->when; | |
466 | opts->tsecr = tp->rx_opt.ts_recent; | |
467 | size += TCPOLEN_TSTAMP_ALIGNED; | |
468 | } | |
469 | if (likely(sysctl_tcp_window_scaling)) { | |
470 | opts->ws = tp->rx_opt.rcv_wscale; | |
471 | size += TCPOLEN_WSCALE_ALIGNED; | |
472 | } | |
473 | if (likely(sysctl_tcp_sack)) { | |
474 | opts->options |= OPTION_SACK_ADVERTISE; | |
b32d1310 | 475 | if (unlikely(!(OPTION_TS & opts->options))) |
33ad798c AL |
476 | size += TCPOLEN_SACKPERM_ALIGNED; |
477 | } | |
478 | ||
479 | return size; | |
40efc6fa SH |
480 | } |
481 | ||
33ad798c AL |
482 | static unsigned tcp_synack_options(struct sock *sk, |
483 | struct request_sock *req, | |
484 | unsigned mss, struct sk_buff *skb, | |
485 | struct tcp_out_options *opts, | |
486 | struct tcp_md5sig_key **md5) { | |
487 | unsigned size = 0; | |
488 | struct inet_request_sock *ireq = inet_rsk(req); | |
489 | char doing_ts; | |
490 | ||
cfb6eeb4 | 491 | #ifdef CONFIG_TCP_MD5SIG |
33ad798c AL |
492 | *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); |
493 | if (*md5) { | |
494 | opts->options |= OPTION_MD5; | |
495 | size += TCPOLEN_MD5SIG_ALIGNED; | |
cfb6eeb4 | 496 | } |
33ad798c AL |
497 | #else |
498 | *md5 = NULL; | |
cfb6eeb4 | 499 | #endif |
33ad798c AL |
500 | |
501 | /* we can't fit any SACK blocks in a packet with MD5 + TS | |
502 | options. There was discussion about disabling SACK rather than TS in | |
503 | order to fit in better with old, buggy kernels, but that was deemed | |
504 | to be unnecessary. */ | |
505 | doing_ts = ireq->tstamp_ok && !(*md5 && ireq->sack_ok); | |
506 | ||
507 | opts->mss = mss; | |
508 | size += TCPOLEN_MSS_ALIGNED; | |
509 | ||
510 | if (likely(ireq->wscale_ok)) { | |
511 | opts->ws = ireq->rcv_wscale; | |
512 | size += TCPOLEN_WSCALE_ALIGNED; | |
513 | } | |
514 | if (likely(doing_ts)) { | |
515 | opts->options |= OPTION_TS; | |
516 | opts->tsval = TCP_SKB_CB(skb)->when; | |
517 | opts->tsecr = req->ts_recent; | |
518 | size += TCPOLEN_TSTAMP_ALIGNED; | |
519 | } | |
520 | if (likely(ireq->sack_ok)) { | |
521 | opts->options |= OPTION_SACK_ADVERTISE; | |
522 | if (unlikely(!doing_ts)) | |
523 | size += TCPOLEN_SACKPERM_ALIGNED; | |
524 | } | |
525 | ||
526 | return size; | |
527 | } | |
528 | ||
529 | static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb, | |
530 | struct tcp_out_options *opts, | |
531 | struct tcp_md5sig_key **md5) { | |
532 | struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; | |
533 | struct tcp_sock *tp = tcp_sk(sk); | |
534 | unsigned size = 0; | |
535 | ||
536 | #ifdef CONFIG_TCP_MD5SIG | |
537 | *md5 = tp->af_specific->md5_lookup(sk, sk); | |
538 | if (unlikely(*md5)) { | |
539 | opts->options |= OPTION_MD5; | |
540 | size += TCPOLEN_MD5SIG_ALIGNED; | |
541 | } | |
542 | #else | |
543 | *md5 = NULL; | |
544 | #endif | |
545 | ||
546 | if (likely(tp->rx_opt.tstamp_ok)) { | |
547 | opts->options |= OPTION_TS; | |
548 | opts->tsval = tcb ? tcb->when : 0; | |
549 | opts->tsecr = tp->rx_opt.ts_recent; | |
550 | size += TCPOLEN_TSTAMP_ALIGNED; | |
551 | } | |
552 | ||
553 | if (unlikely(tp->rx_opt.eff_sacks)) { | |
554 | const unsigned remaining = MAX_TCP_OPTION_SPACE - size; | |
555 | opts->num_sack_blocks = | |
556 | min_t(unsigned, tp->rx_opt.eff_sacks, | |
557 | (remaining - TCPOLEN_SACK_BASE_ALIGNED) / | |
558 | TCPOLEN_SACK_PERBLOCK); | |
559 | size += TCPOLEN_SACK_BASE_ALIGNED + | |
560 | opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; | |
561 | } | |
562 | ||
563 | return size; | |
40efc6fa | 564 | } |
1da177e4 LT |
565 | |
566 | /* This routine actually transmits TCP packets queued in by | |
567 | * tcp_do_sendmsg(). This is used by both the initial | |
568 | * transmission and possible later retransmissions. | |
569 | * All SKB's seen here are completely headerless. It is our | |
570 | * job to build the TCP header, and pass the packet down to | |
571 | * IP so it can do the same plus pass the packet off to the | |
572 | * device. | |
573 | * | |
574 | * We are working here with either a clone of the original | |
575 | * SKB, or a fresh unique copy made by the retransmit engine. | |
576 | */ | |
056834d9 IJ |
577 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, |
578 | gfp_t gfp_mask) | |
1da177e4 | 579 | { |
dfb4b9dc DM |
580 | const struct inet_connection_sock *icsk = inet_csk(sk); |
581 | struct inet_sock *inet; | |
582 | struct tcp_sock *tp; | |
583 | struct tcp_skb_cb *tcb; | |
33ad798c AL |
584 | struct tcp_out_options opts; |
585 | unsigned tcp_options_size, tcp_header_size; | |
cfb6eeb4 YH |
586 | struct tcp_md5sig_key *md5; |
587 | __u8 *md5_hash_location; | |
dfb4b9dc | 588 | struct tcphdr *th; |
dfb4b9dc DM |
589 | int err; |
590 | ||
591 | BUG_ON(!skb || !tcp_skb_pcount(skb)); | |
592 | ||
593 | /* If congestion control is doing timestamping, we must | |
594 | * take such a timestamp before we potentially clone/copy. | |
595 | */ | |
164891aa | 596 | if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) |
dfb4b9dc DM |
597 | __net_timestamp(skb); |
598 | ||
599 | if (likely(clone_it)) { | |
600 | if (unlikely(skb_cloned(skb))) | |
601 | skb = pskb_copy(skb, gfp_mask); | |
602 | else | |
603 | skb = skb_clone(skb, gfp_mask); | |
604 | if (unlikely(!skb)) | |
605 | return -ENOBUFS; | |
606 | } | |
1da177e4 | 607 | |
dfb4b9dc DM |
608 | inet = inet_sk(sk); |
609 | tp = tcp_sk(sk); | |
610 | tcb = TCP_SKB_CB(skb); | |
33ad798c | 611 | memset(&opts, 0, sizeof(opts)); |
1da177e4 | 612 | |
33ad798c AL |
613 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) |
614 | tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); | |
615 | else | |
616 | tcp_options_size = tcp_established_options(sk, skb, &opts, | |
617 | &md5); | |
618 | tcp_header_size = tcp_options_size + sizeof(struct tcphdr); | |
e905a9ed | 619 | |
dfb4b9dc DM |
620 | if (tcp_packets_in_flight(tp) == 0) |
621 | tcp_ca_event(sk, CA_EVENT_TX_START); | |
622 | ||
aa8223c7 ACM |
623 | skb_push(skb, tcp_header_size); |
624 | skb_reset_transport_header(skb); | |
e89862f4 | 625 | skb_set_owner_w(skb, sk); |
dfb4b9dc DM |
626 | |
627 | /* Build TCP header and checksum it. */ | |
aa8223c7 | 628 | th = tcp_hdr(skb); |
dfb4b9dc DM |
629 | th->source = inet->sport; |
630 | th->dest = inet->dport; | |
631 | th->seq = htonl(tcb->seq); | |
632 | th->ack_seq = htonl(tp->rcv_nxt); | |
df7a3b07 | 633 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
dfb4b9dc DM |
634 | tcb->flags); |
635 | ||
636 | if (unlikely(tcb->flags & TCPCB_FLAG_SYN)) { | |
637 | /* RFC1323: The window in SYN & SYN/ACK segments | |
638 | * is never scaled. | |
639 | */ | |
600ff0c2 | 640 | th->window = htons(min(tp->rcv_wnd, 65535U)); |
dfb4b9dc DM |
641 | } else { |
642 | th->window = htons(tcp_select_window(sk)); | |
643 | } | |
644 | th->check = 0; | |
645 | th->urg_ptr = 0; | |
1da177e4 | 646 | |
dfb4b9dc | 647 | if (unlikely(tp->urg_mode && |
056834d9 IJ |
648 | between(tp->snd_up, tcb->seq + 1, tcb->seq + 0xFFFF))) { |
649 | th->urg_ptr = htons(tp->snd_up - tcb->seq); | |
dfb4b9dc DM |
650 | th->urg = 1; |
651 | } | |
1da177e4 | 652 | |
33ad798c AL |
653 | tcp_options_write((__be32 *)(th + 1), tp, &opts, &md5_hash_location); |
654 | if (likely((tcb->flags & TCPCB_FLAG_SYN) == 0)) | |
9e412ba7 | 655 | TCP_ECN_send(sk, skb, tcp_header_size); |
1da177e4 | 656 | |
cfb6eeb4 YH |
657 | #ifdef CONFIG_TCP_MD5SIG |
658 | /* Calculate the MD5 hash, as we have all we need now */ | |
659 | if (md5) { | |
33ad798c | 660 | sk->sk_route_caps &= ~NETIF_F_GSO_MASK; |
cfb6eeb4 | 661 | tp->af_specific->calc_md5_hash(md5_hash_location, |
49a72dfb | 662 | md5, sk, NULL, skb); |
cfb6eeb4 YH |
663 | } |
664 | #endif | |
665 | ||
8292a17a | 666 | icsk->icsk_af_ops->send_check(sk, skb->len, skb); |
1da177e4 | 667 | |
dfb4b9dc DM |
668 | if (likely(tcb->flags & TCPCB_FLAG_ACK)) |
669 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); | |
1da177e4 | 670 | |
dfb4b9dc DM |
671 | if (skb->len != tcp_header_size) |
672 | tcp_event_data_sent(tp, skb, sk); | |
1da177e4 | 673 | |
bd37a088 | 674 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
81cc8a75 | 675 | TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); |
1da177e4 | 676 | |
e89862f4 | 677 | err = icsk->icsk_af_ops->queue_xmit(skb, 0); |
83de47cd | 678 | if (likely(err <= 0)) |
dfb4b9dc DM |
679 | return err; |
680 | ||
3cfe3baa | 681 | tcp_enter_cwr(sk, 1); |
dfb4b9dc | 682 | |
b9df3cb8 | 683 | return net_xmit_eval(err); |
1da177e4 LT |
684 | } |
685 | ||
e905a9ed | 686 | /* This routine just queue's the buffer |
1da177e4 LT |
687 | * |
688 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, | |
689 | * otherwise socket can stall. | |
690 | */ | |
691 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) | |
692 | { | |
693 | struct tcp_sock *tp = tcp_sk(sk); | |
694 | ||
695 | /* Advance write_seq and place onto the write_queue. */ | |
696 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; | |
697 | skb_header_release(skb); | |
fe067e8a | 698 | tcp_add_write_queue_tail(sk, skb); |
3ab224be HA |
699 | sk->sk_wmem_queued += skb->truesize; |
700 | sk_mem_charge(sk, skb->truesize); | |
1da177e4 LT |
701 | } |
702 | ||
056834d9 IJ |
703 | static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb, |
704 | unsigned int mss_now) | |
f6302d1d | 705 | { |
bcd76111 | 706 | if (skb->len <= mss_now || !sk_can_gso(sk)) { |
f6302d1d DM |
707 | /* Avoid the costly divide in the normal |
708 | * non-TSO case. | |
709 | */ | |
7967168c HX |
710 | skb_shinfo(skb)->gso_segs = 1; |
711 | skb_shinfo(skb)->gso_size = 0; | |
712 | skb_shinfo(skb)->gso_type = 0; | |
f6302d1d | 713 | } else { |
356f89e1 | 714 | skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); |
7967168c | 715 | skb_shinfo(skb)->gso_size = mss_now; |
bcd76111 | 716 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
1da177e4 LT |
717 | } |
718 | } | |
719 | ||
91fed7a1 | 720 | /* When a modification to fackets out becomes necessary, we need to check |
68f8353b | 721 | * skb is counted to fackets_out or not. |
91fed7a1 | 722 | */ |
a47e5a98 | 723 | static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb, |
91fed7a1 IJ |
724 | int decr) |
725 | { | |
a47e5a98 IJ |
726 | struct tcp_sock *tp = tcp_sk(sk); |
727 | ||
dc86967b | 728 | if (!tp->sacked_out || tcp_is_reno(tp)) |
91fed7a1 IJ |
729 | return; |
730 | ||
6859d494 | 731 | if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) |
91fed7a1 | 732 | tp->fackets_out -= decr; |
91fed7a1 IJ |
733 | } |
734 | ||
1da177e4 LT |
735 | /* Function to create two new TCP segments. Shrinks the given segment |
736 | * to the specified size and appends a new segment with the rest of the | |
e905a9ed | 737 | * packet to the list. This won't be called frequently, I hope. |
1da177e4 LT |
738 | * Remember, these are still headerless SKBs at this point. |
739 | */ | |
056834d9 IJ |
740 | int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, |
741 | unsigned int mss_now) | |
1da177e4 LT |
742 | { |
743 | struct tcp_sock *tp = tcp_sk(sk); | |
744 | struct sk_buff *buff; | |
6475be16 | 745 | int nsize, old_factor; |
b60b49ea | 746 | int nlen; |
1da177e4 LT |
747 | u16 flags; |
748 | ||
b2cc99f0 | 749 | BUG_ON(len > skb->len); |
6a438bbe | 750 | |
b7689205 | 751 | tcp_clear_retrans_hints_partial(tp); |
1da177e4 LT |
752 | nsize = skb_headlen(skb) - len; |
753 | if (nsize < 0) | |
754 | nsize = 0; | |
755 | ||
756 | if (skb_cloned(skb) && | |
757 | skb_is_nonlinear(skb) && | |
758 | pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) | |
759 | return -ENOMEM; | |
760 | ||
761 | /* Get a new skb... force flag on. */ | |
762 | buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); | |
763 | if (buff == NULL) | |
764 | return -ENOMEM; /* We'll just try again later. */ | |
ef5cb973 | 765 | |
3ab224be HA |
766 | sk->sk_wmem_queued += buff->truesize; |
767 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea HX |
768 | nlen = skb->len - len - nsize; |
769 | buff->truesize += nlen; | |
770 | skb->truesize -= nlen; | |
1da177e4 LT |
771 | |
772 | /* Correct the sequence numbers. */ | |
773 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
774 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
775 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
776 | ||
777 | /* PSH and FIN should only be set in the second packet. */ | |
778 | flags = TCP_SKB_CB(skb)->flags; | |
056834d9 | 779 | TCP_SKB_CB(skb)->flags = flags & ~(TCPCB_FLAG_FIN | TCPCB_FLAG_PSH); |
1da177e4 | 780 | TCP_SKB_CB(buff)->flags = flags; |
e14c3caf | 781 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
1da177e4 | 782 | |
84fa7933 | 783 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { |
1da177e4 | 784 | /* Copy and checksum data tail into the new buffer. */ |
056834d9 IJ |
785 | buff->csum = csum_partial_copy_nocheck(skb->data + len, |
786 | skb_put(buff, nsize), | |
1da177e4 LT |
787 | nsize, 0); |
788 | ||
789 | skb_trim(skb, len); | |
790 | ||
791 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); | |
792 | } else { | |
84fa7933 | 793 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
794 | skb_split(skb, buff, len); |
795 | } | |
796 | ||
797 | buff->ip_summed = skb->ip_summed; | |
798 | ||
799 | /* Looks stupid, but our code really uses when of | |
800 | * skbs, which it never sent before. --ANK | |
801 | */ | |
802 | TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; | |
a61bbcf2 | 803 | buff->tstamp = skb->tstamp; |
1da177e4 | 804 | |
6475be16 DM |
805 | old_factor = tcp_skb_pcount(skb); |
806 | ||
1da177e4 | 807 | /* Fix up tso_factor for both original and new SKB. */ |
846998ae DM |
808 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
809 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
1da177e4 | 810 | |
6475be16 DM |
811 | /* If this packet has been sent out already, we must |
812 | * adjust the various packet counters. | |
813 | */ | |
cf0b450c | 814 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
6475be16 DM |
815 | int diff = old_factor - tcp_skb_pcount(skb) - |
816 | tcp_skb_pcount(buff); | |
1da177e4 | 817 | |
6475be16 | 818 | tp->packets_out -= diff; |
e14c3caf HX |
819 | |
820 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
821 | tp->sacked_out -= diff; | |
822 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
823 | tp->retrans_out -= diff; | |
824 | ||
b5860bba | 825 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) |
6475be16 | 826 | tp->lost_out -= diff; |
83ca28be | 827 | |
91fed7a1 IJ |
828 | /* Adjust Reno SACK estimate. */ |
829 | if (tcp_is_reno(tp) && diff > 0) { | |
830 | tcp_dec_pcount_approx_int(&tp->sacked_out, diff); | |
831 | tcp_verify_left_out(tp); | |
6475be16 | 832 | } |
a47e5a98 | 833 | tcp_adjust_fackets_out(sk, skb, diff); |
1da177e4 LT |
834 | } |
835 | ||
836 | /* Link BUFF into the send queue. */ | |
f44b5271 | 837 | skb_header_release(buff); |
fe067e8a | 838 | tcp_insert_write_queue_after(skb, buff, sk); |
1da177e4 LT |
839 | |
840 | return 0; | |
841 | } | |
842 | ||
843 | /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c | |
844 | * eventually). The difference is that pulled data not copied, but | |
845 | * immediately discarded. | |
846 | */ | |
f2911969 | 847 | static void __pskb_trim_head(struct sk_buff *skb, int len) |
1da177e4 LT |
848 | { |
849 | int i, k, eat; | |
850 | ||
851 | eat = len; | |
852 | k = 0; | |
056834d9 | 853 | for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { |
1da177e4 LT |
854 | if (skb_shinfo(skb)->frags[i].size <= eat) { |
855 | put_page(skb_shinfo(skb)->frags[i].page); | |
856 | eat -= skb_shinfo(skb)->frags[i].size; | |
857 | } else { | |
858 | skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; | |
859 | if (eat) { | |
860 | skb_shinfo(skb)->frags[k].page_offset += eat; | |
861 | skb_shinfo(skb)->frags[k].size -= eat; | |
862 | eat = 0; | |
863 | } | |
864 | k++; | |
865 | } | |
866 | } | |
867 | skb_shinfo(skb)->nr_frags = k; | |
868 | ||
27a884dc | 869 | skb_reset_tail_pointer(skb); |
1da177e4 LT |
870 | skb->data_len -= len; |
871 | skb->len = skb->data_len; | |
1da177e4 LT |
872 | } |
873 | ||
874 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) | |
875 | { | |
056834d9 | 876 | if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) |
1da177e4 LT |
877 | return -ENOMEM; |
878 | ||
f2911969 HXP |
879 | /* If len == headlen, we avoid __skb_pull to preserve alignment. */ |
880 | if (unlikely(len < skb_headlen(skb))) | |
1da177e4 | 881 | __skb_pull(skb, len); |
f2911969 HXP |
882 | else |
883 | __pskb_trim_head(skb, len - skb_headlen(skb)); | |
1da177e4 LT |
884 | |
885 | TCP_SKB_CB(skb)->seq += len; | |
84fa7933 | 886 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
887 | |
888 | skb->truesize -= len; | |
889 | sk->sk_wmem_queued -= len; | |
3ab224be | 890 | sk_mem_uncharge(sk, len); |
1da177e4 LT |
891 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); |
892 | ||
893 | /* Any change of skb->len requires recalculation of tso | |
894 | * factor and mss. | |
895 | */ | |
896 | if (tcp_skb_pcount(skb) > 1) | |
846998ae | 897 | tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk, 1)); |
1da177e4 LT |
898 | |
899 | return 0; | |
900 | } | |
901 | ||
5d424d5a JH |
902 | /* Not accounting for SACKs here. */ |
903 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
904 | { | |
905 | struct tcp_sock *tp = tcp_sk(sk); | |
906 | struct inet_connection_sock *icsk = inet_csk(sk); | |
907 | int mss_now; | |
908 | ||
909 | /* Calculate base mss without TCP options: | |
910 | It is MMS_S - sizeof(tcphdr) of rfc1122 | |
911 | */ | |
912 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); | |
913 | ||
914 | /* Clamp it (mss_clamp does not include tcp options) */ | |
915 | if (mss_now > tp->rx_opt.mss_clamp) | |
916 | mss_now = tp->rx_opt.mss_clamp; | |
917 | ||
918 | /* Now subtract optional transport overhead */ | |
919 | mss_now -= icsk->icsk_ext_hdr_len; | |
920 | ||
921 | /* Then reserve room for full set of TCP options and 8 bytes of data */ | |
922 | if (mss_now < 48) | |
923 | mss_now = 48; | |
924 | ||
925 | /* Now subtract TCP options size, not including SACKs */ | |
926 | mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); | |
927 | ||
928 | return mss_now; | |
929 | } | |
930 | ||
931 | /* Inverse of above */ | |
932 | int tcp_mss_to_mtu(struct sock *sk, int mss) | |
933 | { | |
934 | struct tcp_sock *tp = tcp_sk(sk); | |
935 | struct inet_connection_sock *icsk = inet_csk(sk); | |
936 | int mtu; | |
937 | ||
938 | mtu = mss + | |
939 | tp->tcp_header_len + | |
940 | icsk->icsk_ext_hdr_len + | |
941 | icsk->icsk_af_ops->net_header_len; | |
942 | ||
943 | return mtu; | |
944 | } | |
945 | ||
946 | void tcp_mtup_init(struct sock *sk) | |
947 | { | |
948 | struct tcp_sock *tp = tcp_sk(sk); | |
949 | struct inet_connection_sock *icsk = inet_csk(sk); | |
950 | ||
951 | icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; | |
952 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + | |
e905a9ed | 953 | icsk->icsk_af_ops->net_header_len; |
5d424d5a JH |
954 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); |
955 | icsk->icsk_mtup.probe_size = 0; | |
956 | } | |
957 | ||
409d22b4 IJ |
958 | /* Bound MSS / TSO packet size with the half of the window */ |
959 | static int tcp_bound_to_half_wnd(struct tcp_sock *tp, int pktsize) | |
960 | { | |
961 | if (tp->max_window && pktsize > (tp->max_window >> 1)) | |
962 | return max(tp->max_window >> 1, 68U - tp->tcp_header_len); | |
963 | else | |
964 | return pktsize; | |
965 | } | |
966 | ||
1da177e4 LT |
967 | /* This function synchronize snd mss to current pmtu/exthdr set. |
968 | ||
969 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts | |
970 | for TCP options, but includes only bare TCP header. | |
971 | ||
972 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. | |
caa20d9a | 973 | It is minimum of user_mss and mss received with SYN. |
1da177e4 LT |
974 | It also does not include TCP options. |
975 | ||
d83d8461 | 976 | inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. |
1da177e4 LT |
977 | |
978 | tp->mss_cache is current effective sending mss, including | |
979 | all tcp options except for SACKs. It is evaluated, | |
980 | taking into account current pmtu, but never exceeds | |
981 | tp->rx_opt.mss_clamp. | |
982 | ||
983 | NOTE1. rfc1122 clearly states that advertised MSS | |
984 | DOES NOT include either tcp or ip options. | |
985 | ||
d83d8461 ACM |
986 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
987 | are READ ONLY outside this function. --ANK (980731) | |
1da177e4 | 988 | */ |
1da177e4 LT |
989 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) |
990 | { | |
991 | struct tcp_sock *tp = tcp_sk(sk); | |
d83d8461 | 992 | struct inet_connection_sock *icsk = inet_csk(sk); |
5d424d5a | 993 | int mss_now; |
1da177e4 | 994 | |
5d424d5a JH |
995 | if (icsk->icsk_mtup.search_high > pmtu) |
996 | icsk->icsk_mtup.search_high = pmtu; | |
1da177e4 | 997 | |
5d424d5a | 998 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
409d22b4 | 999 | mss_now = tcp_bound_to_half_wnd(tp, mss_now); |
1da177e4 LT |
1000 | |
1001 | /* And store cached results */ | |
d83d8461 | 1002 | icsk->icsk_pmtu_cookie = pmtu; |
5d424d5a JH |
1003 | if (icsk->icsk_mtup.enabled) |
1004 | mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); | |
c1b4a7e6 | 1005 | tp->mss_cache = mss_now; |
1da177e4 LT |
1006 | |
1007 | return mss_now; | |
1008 | } | |
1009 | ||
1010 | /* Compute the current effective MSS, taking SACKs and IP options, | |
1011 | * and even PMTU discovery events into account. | |
1012 | * | |
1013 | * LARGESEND note: !urg_mode is overkill, only frames up to snd_up | |
1014 | * cannot be large. However, taking into account rare use of URG, this | |
1015 | * is not a big flaw. | |
1016 | */ | |
c1b4a7e6 | 1017 | unsigned int tcp_current_mss(struct sock *sk, int large_allowed) |
1da177e4 LT |
1018 | { |
1019 | struct tcp_sock *tp = tcp_sk(sk); | |
1020 | struct dst_entry *dst = __sk_dst_get(sk); | |
c1b4a7e6 DM |
1021 | u32 mss_now; |
1022 | u16 xmit_size_goal; | |
1023 | int doing_tso = 0; | |
33ad798c AL |
1024 | unsigned header_len; |
1025 | struct tcp_out_options opts; | |
1026 | struct tcp_md5sig_key *md5; | |
c1b4a7e6 DM |
1027 | |
1028 | mss_now = tp->mss_cache; | |
1029 | ||
bcd76111 | 1030 | if (large_allowed && sk_can_gso(sk) && !tp->urg_mode) |
c1b4a7e6 | 1031 | doing_tso = 1; |
1da177e4 | 1032 | |
1da177e4 LT |
1033 | if (dst) { |
1034 | u32 mtu = dst_mtu(dst); | |
d83d8461 | 1035 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
1da177e4 LT |
1036 | mss_now = tcp_sync_mss(sk, mtu); |
1037 | } | |
1038 | ||
33ad798c AL |
1039 | header_len = tcp_established_options(sk, NULL, &opts, &md5) + |
1040 | sizeof(struct tcphdr); | |
1041 | /* The mss_cache is sized based on tp->tcp_header_len, which assumes | |
1042 | * some common options. If this is an odd packet (because we have SACK | |
1043 | * blocks etc) then our calculated header_len will be different, and | |
1044 | * we have to adjust mss_now correspondingly */ | |
1045 | if (header_len != tp->tcp_header_len) { | |
1046 | int delta = (int) header_len - tp->tcp_header_len; | |
1047 | mss_now -= delta; | |
1048 | } | |
cfb6eeb4 | 1049 | |
c1b4a7e6 | 1050 | xmit_size_goal = mss_now; |
1da177e4 | 1051 | |
c1b4a7e6 | 1052 | if (doing_tso) { |
82cc1a7a | 1053 | xmit_size_goal = ((sk->sk_gso_max_size - 1) - |
8292a17a | 1054 | inet_csk(sk)->icsk_af_ops->net_header_len - |
d83d8461 ACM |
1055 | inet_csk(sk)->icsk_ext_hdr_len - |
1056 | tp->tcp_header_len); | |
1da177e4 | 1057 | |
409d22b4 | 1058 | xmit_size_goal = tcp_bound_to_half_wnd(tp, xmit_size_goal); |
c1b4a7e6 | 1059 | xmit_size_goal -= (xmit_size_goal % mss_now); |
1da177e4 | 1060 | } |
c1b4a7e6 | 1061 | tp->xmit_size_goal = xmit_size_goal; |
1da177e4 | 1062 | |
1da177e4 LT |
1063 | return mss_now; |
1064 | } | |
1065 | ||
a762a980 | 1066 | /* Congestion window validation. (RFC2861) */ |
9e412ba7 | 1067 | static void tcp_cwnd_validate(struct sock *sk) |
a762a980 | 1068 | { |
9e412ba7 | 1069 | struct tcp_sock *tp = tcp_sk(sk); |
a762a980 | 1070 | |
d436d686 | 1071 | if (tp->packets_out >= tp->snd_cwnd) { |
a762a980 DM |
1072 | /* Network is feed fully. */ |
1073 | tp->snd_cwnd_used = 0; | |
1074 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1075 | } else { | |
1076 | /* Network starves. */ | |
1077 | if (tp->packets_out > tp->snd_cwnd_used) | |
1078 | tp->snd_cwnd_used = tp->packets_out; | |
1079 | ||
15d33c07 DM |
1080 | if (sysctl_tcp_slow_start_after_idle && |
1081 | (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) | |
a762a980 DM |
1082 | tcp_cwnd_application_limited(sk); |
1083 | } | |
1084 | } | |
1085 | ||
0e3a4803 IJ |
1086 | /* Returns the portion of skb which can be sent right away without |
1087 | * introducing MSS oddities to segment boundaries. In rare cases where | |
1088 | * mss_now != mss_cache, we will request caller to create a small skb | |
1089 | * per input skb which could be mostly avoided here (if desired). | |
5ea3a748 IJ |
1090 | * |
1091 | * We explicitly want to create a request for splitting write queue tail | |
1092 | * to a small skb for Nagle purposes while avoiding unnecessary modulos, | |
1093 | * thus all the complexity (cwnd_len is always MSS multiple which we | |
1094 | * return whenever allowed by the other factors). Basically we need the | |
1095 | * modulo only when the receiver window alone is the limiting factor or | |
1096 | * when we would be allowed to send the split-due-to-Nagle skb fully. | |
0e3a4803 IJ |
1097 | */ |
1098 | static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb, | |
056834d9 | 1099 | unsigned int mss_now, unsigned int cwnd) |
c1b4a7e6 | 1100 | { |
0e3a4803 IJ |
1101 | struct tcp_sock *tp = tcp_sk(sk); |
1102 | u32 needed, window, cwnd_len; | |
c1b4a7e6 | 1103 | |
90840def | 1104 | window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
c1b4a7e6 | 1105 | cwnd_len = mss_now * cwnd; |
0e3a4803 IJ |
1106 | |
1107 | if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk))) | |
1108 | return cwnd_len; | |
1109 | ||
5ea3a748 IJ |
1110 | needed = min(skb->len, window); |
1111 | ||
17515408 | 1112 | if (cwnd_len <= needed) |
0e3a4803 IJ |
1113 | return cwnd_len; |
1114 | ||
0e3a4803 | 1115 | return needed - needed % mss_now; |
c1b4a7e6 DM |
1116 | } |
1117 | ||
1118 | /* Can at least one segment of SKB be sent right now, according to the | |
1119 | * congestion window rules? If so, return how many segments are allowed. | |
1120 | */ | |
056834d9 IJ |
1121 | static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp, |
1122 | struct sk_buff *skb) | |
c1b4a7e6 DM |
1123 | { |
1124 | u32 in_flight, cwnd; | |
1125 | ||
1126 | /* Don't be strict about the congestion window for the final FIN. */ | |
104439a8 JH |
1127 | if ((TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN) && |
1128 | tcp_skb_pcount(skb) == 1) | |
c1b4a7e6 DM |
1129 | return 1; |
1130 | ||
1131 | in_flight = tcp_packets_in_flight(tp); | |
1132 | cwnd = tp->snd_cwnd; | |
1133 | if (in_flight < cwnd) | |
1134 | return (cwnd - in_flight); | |
1135 | ||
1136 | return 0; | |
1137 | } | |
1138 | ||
1139 | /* This must be invoked the first time we consider transmitting | |
1140 | * SKB onto the wire. | |
1141 | */ | |
056834d9 IJ |
1142 | static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb, |
1143 | unsigned int mss_now) | |
c1b4a7e6 DM |
1144 | { |
1145 | int tso_segs = tcp_skb_pcount(skb); | |
1146 | ||
056834d9 | 1147 | if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { |
846998ae | 1148 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
c1b4a7e6 DM |
1149 | tso_segs = tcp_skb_pcount(skb); |
1150 | } | |
1151 | return tso_segs; | |
1152 | } | |
1153 | ||
1154 | static inline int tcp_minshall_check(const struct tcp_sock *tp) | |
1155 | { | |
1156 | return after(tp->snd_sml,tp->snd_una) && | |
1157 | !after(tp->snd_sml, tp->snd_nxt); | |
1158 | } | |
1159 | ||
1160 | /* Return 0, if packet can be sent now without violation Nagle's rules: | |
1161 | * 1. It is full sized. | |
1162 | * 2. Or it contains FIN. (already checked by caller) | |
1163 | * 3. Or TCP_NODELAY was set. | |
1164 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. | |
1165 | * With Minshall's modification: all sent small packets are ACKed. | |
1166 | */ | |
c1b4a7e6 | 1167 | static inline int tcp_nagle_check(const struct tcp_sock *tp, |
e905a9ed | 1168 | const struct sk_buff *skb, |
c1b4a7e6 DM |
1169 | unsigned mss_now, int nonagle) |
1170 | { | |
1171 | return (skb->len < mss_now && | |
056834d9 IJ |
1172 | ((nonagle & TCP_NAGLE_CORK) || |
1173 | (!nonagle && tp->packets_out && tcp_minshall_check(tp)))); | |
c1b4a7e6 DM |
1174 | } |
1175 | ||
1176 | /* Return non-zero if the Nagle test allows this packet to be | |
1177 | * sent now. | |
1178 | */ | |
1179 | static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb, | |
1180 | unsigned int cur_mss, int nonagle) | |
1181 | { | |
1182 | /* Nagle rule does not apply to frames, which sit in the middle of the | |
1183 | * write_queue (they have no chances to get new data). | |
1184 | * | |
1185 | * This is implemented in the callers, where they modify the 'nonagle' | |
1186 | * argument based upon the location of SKB in the send queue. | |
1187 | */ | |
1188 | if (nonagle & TCP_NAGLE_PUSH) | |
1189 | return 1; | |
1190 | ||
d551e454 IJ |
1191 | /* Don't use the nagle rule for urgent data (or for the final FIN). |
1192 | * Nagle can be ignored during F-RTO too (see RFC4138). | |
1193 | */ | |
1194 | if (tp->urg_mode || (tp->frto_counter == 2) || | |
c1b4a7e6 DM |
1195 | (TCP_SKB_CB(skb)->flags & TCPCB_FLAG_FIN)) |
1196 | return 1; | |
1197 | ||
1198 | if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) | |
1199 | return 1; | |
1200 | ||
1201 | return 0; | |
1202 | } | |
1203 | ||
1204 | /* Does at least the first segment of SKB fit into the send window? */ | |
056834d9 IJ |
1205 | static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb, |
1206 | unsigned int cur_mss) | |
c1b4a7e6 DM |
1207 | { |
1208 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
1209 | ||
1210 | if (skb->len > cur_mss) | |
1211 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; | |
1212 | ||
90840def | 1213 | return !after(end_seq, tcp_wnd_end(tp)); |
c1b4a7e6 DM |
1214 | } |
1215 | ||
fe067e8a | 1216 | /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) |
c1b4a7e6 DM |
1217 | * should be put on the wire right now. If so, it returns the number of |
1218 | * packets allowed by the congestion window. | |
1219 | */ | |
1220 | static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb, | |
1221 | unsigned int cur_mss, int nonagle) | |
1222 | { | |
1223 | struct tcp_sock *tp = tcp_sk(sk); | |
1224 | unsigned int cwnd_quota; | |
1225 | ||
846998ae | 1226 | tcp_init_tso_segs(sk, skb, cur_mss); |
c1b4a7e6 DM |
1227 | |
1228 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) | |
1229 | return 0; | |
1230 | ||
1231 | cwnd_quota = tcp_cwnd_test(tp, skb); | |
056834d9 | 1232 | if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) |
c1b4a7e6 DM |
1233 | cwnd_quota = 0; |
1234 | ||
1235 | return cwnd_quota; | |
1236 | } | |
1237 | ||
9e412ba7 | 1238 | int tcp_may_send_now(struct sock *sk) |
c1b4a7e6 | 1239 | { |
9e412ba7 | 1240 | struct tcp_sock *tp = tcp_sk(sk); |
fe067e8a | 1241 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 DM |
1242 | |
1243 | return (skb && | |
1244 | tcp_snd_test(sk, skb, tcp_current_mss(sk, 1), | |
1245 | (tcp_skb_is_last(sk, skb) ? |