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 | |
46d3ceab ED |
53 | /* Default TSQ limit of two TSO segments */ |
54 | int sysctl_tcp_limit_output_bytes __read_mostly = 131072; | |
55 | ||
1da177e4 LT |
56 | /* This limits the percentage of the congestion window which we |
57 | * will allow a single TSO frame to consume. Building TSO frames | |
58 | * which are too large can cause TCP streams to be bursty. | |
59 | */ | |
ab32ea5d | 60 | int sysctl_tcp_tso_win_divisor __read_mostly = 3; |
1da177e4 | 61 | |
ab32ea5d | 62 | int sysctl_tcp_mtu_probing __read_mostly = 0; |
97b1ce25 | 63 | int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS; |
5d424d5a | 64 | |
35089bb2 | 65 | /* By default, RFC2861 behavior. */ |
ab32ea5d | 66 | int sysctl_tcp_slow_start_after_idle __read_mostly = 1; |
35089bb2 | 67 | |
46d3ceab ED |
68 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
69 | int push_one, gfp_t gfp); | |
519855c5 | 70 | |
67edfef7 | 71 | /* Account for new data that has been sent to the network. */ |
cf533ea5 | 72 | static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 73 | { |
6ba8a3b1 | 74 | struct inet_connection_sock *icsk = inet_csk(sk); |
9e412ba7 | 75 | struct tcp_sock *tp = tcp_sk(sk); |
66f5fe62 | 76 | unsigned int prior_packets = tp->packets_out; |
9e412ba7 | 77 | |
fe067e8a | 78 | tcp_advance_send_head(sk, skb); |
1da177e4 | 79 | tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; |
8512430e | 80 | |
66f5fe62 | 81 | tp->packets_out += tcp_skb_pcount(skb); |
6ba8a3b1 | 82 | if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || |
6a5dc9e5 | 83 | icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { |
750ea2ba | 84 | tcp_rearm_rto(sk); |
6a5dc9e5 | 85 | } |
1da177e4 LT |
86 | } |
87 | ||
88 | /* SND.NXT, if window was not shrunk. | |
89 | * If window has been shrunk, what should we make? It is not clear at all. | |
90 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( | |
91 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already | |
92 | * invalid. OK, let's make this for now: | |
93 | */ | |
cf533ea5 | 94 | static inline __u32 tcp_acceptable_seq(const struct sock *sk) |
1da177e4 | 95 | { |
cf533ea5 | 96 | const struct tcp_sock *tp = tcp_sk(sk); |
9e412ba7 | 97 | |
90840def | 98 | if (!before(tcp_wnd_end(tp), tp->snd_nxt)) |
1da177e4 LT |
99 | return tp->snd_nxt; |
100 | else | |
90840def | 101 | return tcp_wnd_end(tp); |
1da177e4 LT |
102 | } |
103 | ||
104 | /* Calculate mss to advertise in SYN segment. | |
105 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: | |
106 | * | |
107 | * 1. It is independent of path mtu. | |
108 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. | |
109 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of | |
110 | * attached devices, because some buggy hosts are confused by | |
111 | * large MSS. | |
112 | * 4. We do not make 3, we advertise MSS, calculated from first | |
113 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. | |
114 | * This may be overridden via information stored in routing table. | |
115 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, | |
116 | * probably even Jumbo". | |
117 | */ | |
118 | static __u16 tcp_advertise_mss(struct sock *sk) | |
119 | { | |
120 | struct tcp_sock *tp = tcp_sk(sk); | |
cf533ea5 | 121 | const struct dst_entry *dst = __sk_dst_get(sk); |
1da177e4 LT |
122 | int mss = tp->advmss; |
123 | ||
0dbaee3b DM |
124 | if (dst) { |
125 | unsigned int metric = dst_metric_advmss(dst); | |
126 | ||
127 | if (metric < mss) { | |
128 | mss = metric; | |
129 | tp->advmss = mss; | |
130 | } | |
1da177e4 LT |
131 | } |
132 | ||
133 | return (__u16)mss; | |
134 | } | |
135 | ||
136 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". | |
137 | * This is the first part of cwnd validation mechanism. */ | |
cf533ea5 | 138 | static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst) |
1da177e4 | 139 | { |
463c84b9 | 140 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
141 | s32 delta = tcp_time_stamp - tp->lsndtime; |
142 | u32 restart_cwnd = tcp_init_cwnd(tp, dst); | |
143 | u32 cwnd = tp->snd_cwnd; | |
144 | ||
6687e988 | 145 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); |
1da177e4 | 146 | |
6687e988 | 147 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
1da177e4 LT |
148 | restart_cwnd = min(restart_cwnd, cwnd); |
149 | ||
463c84b9 | 150 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) |
1da177e4 LT |
151 | cwnd >>= 1; |
152 | tp->snd_cwnd = max(cwnd, restart_cwnd); | |
153 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
154 | tp->snd_cwnd_used = 0; | |
155 | } | |
156 | ||
67edfef7 | 157 | /* Congestion state accounting after a packet has been sent. */ |
40efc6fa | 158 | static void tcp_event_data_sent(struct tcp_sock *tp, |
cf533ea5 | 159 | struct sock *sk) |
1da177e4 | 160 | { |
463c84b9 ACM |
161 | struct inet_connection_sock *icsk = inet_csk(sk); |
162 | const u32 now = tcp_time_stamp; | |
1da177e4 | 163 | |
35089bb2 DM |
164 | if (sysctl_tcp_slow_start_after_idle && |
165 | (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) | |
463c84b9 | 166 | tcp_cwnd_restart(sk, __sk_dst_get(sk)); |
1da177e4 LT |
167 | |
168 | tp->lsndtime = now; | |
169 | ||
170 | /* If it is a reply for ato after last received | |
171 | * packet, enter pingpong mode. | |
172 | */ | |
463c84b9 ACM |
173 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) |
174 | icsk->icsk_ack.pingpong = 1; | |
1da177e4 LT |
175 | } |
176 | ||
67edfef7 | 177 | /* Account for an ACK we sent. */ |
40efc6fa | 178 | static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) |
1da177e4 | 179 | { |
463c84b9 ACM |
180 | tcp_dec_quickack_mode(sk, pkts); |
181 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); | |
1da177e4 LT |
182 | } |
183 | ||
184 | /* Determine a window scaling and initial window to offer. | |
185 | * Based on the assumption that the given amount of space | |
186 | * will be offered. Store the results in the tp structure. | |
187 | * NOTE: for smooth operation initial space offering should | |
188 | * be a multiple of mss if possible. We assume here that mss >= 1. | |
189 | * This MUST be enforced by all callers. | |
190 | */ | |
191 | void tcp_select_initial_window(int __space, __u32 mss, | |
192 | __u32 *rcv_wnd, __u32 *window_clamp, | |
31d12926 | 193 | int wscale_ok, __u8 *rcv_wscale, |
194 | __u32 init_rcv_wnd) | |
1da177e4 LT |
195 | { |
196 | unsigned int space = (__space < 0 ? 0 : __space); | |
197 | ||
198 | /* If no clamp set the clamp to the max possible scaled window */ | |
199 | if (*window_clamp == 0) | |
200 | (*window_clamp) = (65535 << 14); | |
201 | space = min(*window_clamp, space); | |
202 | ||
203 | /* Quantize space offering to a multiple of mss if possible. */ | |
204 | if (space > mss) | |
205 | space = (space / mss) * mss; | |
206 | ||
207 | /* NOTE: offering an initial window larger than 32767 | |
15d99e02 RJ |
208 | * will break some buggy TCP stacks. If the admin tells us |
209 | * it is likely we could be speaking with such a buggy stack | |
210 | * we will truncate our initial window offering to 32K-1 | |
211 | * unless the remote has sent us a window scaling option, | |
212 | * which we interpret as a sign the remote TCP is not | |
213 | * misinterpreting the window field as a signed quantity. | |
1da177e4 | 214 | */ |
15d99e02 RJ |
215 | if (sysctl_tcp_workaround_signed_windows) |
216 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); | |
217 | else | |
218 | (*rcv_wnd) = space; | |
219 | ||
1da177e4 LT |
220 | (*rcv_wscale) = 0; |
221 | if (wscale_ok) { | |
222 | /* Set window scaling on max possible window | |
e905a9ed | 223 | * See RFC1323 for an explanation of the limit to 14 |
1da177e4 | 224 | */ |
aadf1c47 SHY |
225 | space = max_t(u32, space, sysctl_tcp_rmem[2]); |
226 | space = max_t(u32, space, sysctl_rmem_max); | |
316c1592 | 227 | space = min_t(u32, space, *window_clamp); |
1da177e4 LT |
228 | while (space > 65535 && (*rcv_wscale) < 14) { |
229 | space >>= 1; | |
230 | (*rcv_wscale)++; | |
231 | } | |
232 | } | |
233 | ||
356f0398 ND |
234 | /* Set initial window to a value enough for senders starting with |
235 | * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place | |
236 | * a limit on the initial window when mss is larger than 1460. | |
237 | */ | |
056834d9 | 238 | if (mss > (1 << *rcv_wscale)) { |
356f0398 ND |
239 | int init_cwnd = TCP_DEFAULT_INIT_RCVWND; |
240 | if (mss > 1460) | |
241 | init_cwnd = | |
242 | max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2); | |
31d12926 | 243 | /* when initializing use the value from init_rcv_wnd |
244 | * rather than the default from above | |
245 | */ | |
b1afde60 ND |
246 | if (init_rcv_wnd) |
247 | *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); | |
248 | else | |
249 | *rcv_wnd = min(*rcv_wnd, init_cwnd * mss); | |
1da177e4 LT |
250 | } |
251 | ||
252 | /* Set the clamp no higher than max representable value */ | |
253 | (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); | |
254 | } | |
4bc2f18b | 255 | EXPORT_SYMBOL(tcp_select_initial_window); |
1da177e4 LT |
256 | |
257 | /* Chose a new window to advertise, update state in tcp_sock for the | |
258 | * socket, and return result with RFC1323 scaling applied. The return | |
259 | * value can be stuffed directly into th->window for an outgoing | |
260 | * frame. | |
261 | */ | |
40efc6fa | 262 | static u16 tcp_select_window(struct sock *sk) |
1da177e4 LT |
263 | { |
264 | struct tcp_sock *tp = tcp_sk(sk); | |
265 | u32 cur_win = tcp_receive_window(tp); | |
266 | u32 new_win = __tcp_select_window(sk); | |
267 | ||
268 | /* Never shrink the offered window */ | |
2de979bd | 269 | if (new_win < cur_win) { |
1da177e4 LT |
270 | /* Danger Will Robinson! |
271 | * Don't update rcv_wup/rcv_wnd here or else | |
272 | * we will not be able to advertise a zero | |
273 | * window in time. --DaveM | |
274 | * | |
275 | * Relax Will Robinson. | |
276 | */ | |
607bfbf2 | 277 | new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); |
1da177e4 LT |
278 | } |
279 | tp->rcv_wnd = new_win; | |
280 | tp->rcv_wup = tp->rcv_nxt; | |
281 | ||
282 | /* Make sure we do not exceed the maximum possible | |
283 | * scaled window. | |
284 | */ | |
15d99e02 | 285 | if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) |
1da177e4 LT |
286 | new_win = min(new_win, MAX_TCP_WINDOW); |
287 | else | |
288 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); | |
289 | ||
290 | /* RFC1323 scaling applied */ | |
291 | new_win >>= tp->rx_opt.rcv_wscale; | |
292 | ||
293 | /* If we advertise zero window, disable fast path. */ | |
294 | if (new_win == 0) | |
295 | tp->pred_flags = 0; | |
296 | ||
297 | return new_win; | |
298 | } | |
299 | ||
67edfef7 | 300 | /* Packet ECN state for a SYN-ACK */ |
cf533ea5 | 301 | static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb) |
bdf1ee5d | 302 | { |
4de075e0 | 303 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; |
056834d9 | 304 | if (!(tp->ecn_flags & TCP_ECN_OK)) |
4de075e0 | 305 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; |
bdf1ee5d IJ |
306 | } |
307 | ||
67edfef7 | 308 | /* Packet ECN state for a SYN. */ |
bdf1ee5d IJ |
309 | static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) |
310 | { | |
311 | struct tcp_sock *tp = tcp_sk(sk); | |
312 | ||
313 | tp->ecn_flags = 0; | |
5d134f1c | 314 | if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) { |
4de075e0 | 315 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; |
bdf1ee5d IJ |
316 | tp->ecn_flags = TCP_ECN_OK; |
317 | } | |
318 | } | |
319 | ||
320 | static __inline__ void | |
cf533ea5 | 321 | TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th) |
bdf1ee5d IJ |
322 | { |
323 | if (inet_rsk(req)->ecn_ok) | |
324 | th->ece = 1; | |
325 | } | |
326 | ||
67edfef7 AK |
327 | /* Set up ECN state for a packet on a ESTABLISHED socket that is about to |
328 | * be sent. | |
329 | */ | |
bdf1ee5d IJ |
330 | static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, |
331 | int tcp_header_len) | |
332 | { | |
333 | struct tcp_sock *tp = tcp_sk(sk); | |
334 | ||
335 | if (tp->ecn_flags & TCP_ECN_OK) { | |
336 | /* Not-retransmitted data segment: set ECT and inject CWR. */ | |
337 | if (skb->len != tcp_header_len && | |
338 | !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { | |
339 | INET_ECN_xmit(sk); | |
056834d9 | 340 | if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { |
bdf1ee5d IJ |
341 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; |
342 | tcp_hdr(skb)->cwr = 1; | |
343 | skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; | |
344 | } | |
345 | } else { | |
346 | /* ACK or retransmitted segment: clear ECT|CE */ | |
347 | INET_ECN_dontxmit(sk); | |
348 | } | |
349 | if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) | |
350 | tcp_hdr(skb)->ece = 1; | |
351 | } | |
352 | } | |
353 | ||
e870a8ef IJ |
354 | /* Constructs common control bits of non-data skb. If SYN/FIN is present, |
355 | * auto increment end seqno. | |
356 | */ | |
357 | static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) | |
358 | { | |
2e8e18ef | 359 | skb->ip_summed = CHECKSUM_PARTIAL; |
e870a8ef IJ |
360 | skb->csum = 0; |
361 | ||
4de075e0 | 362 | TCP_SKB_CB(skb)->tcp_flags = flags; |
e870a8ef IJ |
363 | TCP_SKB_CB(skb)->sacked = 0; |
364 | ||
365 | skb_shinfo(skb)->gso_segs = 1; | |
366 | skb_shinfo(skb)->gso_size = 0; | |
367 | skb_shinfo(skb)->gso_type = 0; | |
368 | ||
369 | TCP_SKB_CB(skb)->seq = seq; | |
a3433f35 | 370 | if (flags & (TCPHDR_SYN | TCPHDR_FIN)) |
e870a8ef IJ |
371 | seq++; |
372 | TCP_SKB_CB(skb)->end_seq = seq; | |
373 | } | |
374 | ||
a2a385d6 | 375 | static inline bool tcp_urg_mode(const struct tcp_sock *tp) |
33f5f57e IJ |
376 | { |
377 | return tp->snd_una != tp->snd_up; | |
378 | } | |
379 | ||
33ad798c AL |
380 | #define OPTION_SACK_ADVERTISE (1 << 0) |
381 | #define OPTION_TS (1 << 1) | |
382 | #define OPTION_MD5 (1 << 2) | |
89e95a61 | 383 | #define OPTION_WSCALE (1 << 3) |
2100c8d2 | 384 | #define OPTION_FAST_OPEN_COOKIE (1 << 8) |
33ad798c AL |
385 | |
386 | struct tcp_out_options { | |
2100c8d2 YC |
387 | u16 options; /* bit field of OPTION_* */ |
388 | u16 mss; /* 0 to disable */ | |
33ad798c AL |
389 | u8 ws; /* window scale, 0 to disable */ |
390 | u8 num_sack_blocks; /* number of SACK blocks to include */ | |
bd0388ae | 391 | u8 hash_size; /* bytes in hash_location */ |
bd0388ae | 392 | __u8 *hash_location; /* temporary pointer, overloaded */ |
2100c8d2 YC |
393 | __u32 tsval, tsecr; /* need to include OPTION_TS */ |
394 | struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ | |
33ad798c AL |
395 | }; |
396 | ||
67edfef7 AK |
397 | /* Write previously computed TCP options to the packet. |
398 | * | |
399 | * Beware: Something in the Internet is very sensitive to the ordering of | |
fd6149d3 IJ |
400 | * TCP options, we learned this through the hard way, so be careful here. |
401 | * Luckily we can at least blame others for their non-compliance but from | |
402 | * inter-operatibility perspective it seems that we're somewhat stuck with | |
403 | * the ordering which we have been using if we want to keep working with | |
404 | * those broken things (not that it currently hurts anybody as there isn't | |
405 | * particular reason why the ordering would need to be changed). | |
406 | * | |
407 | * At least SACK_PERM as the first option is known to lead to a disaster | |
408 | * (but it may well be that other scenarios fail similarly). | |
409 | */ | |
33ad798c | 410 | static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, |
bd0388ae WAS |
411 | struct tcp_out_options *opts) |
412 | { | |
2100c8d2 | 413 | u16 options = opts->options; /* mungable copy */ |
bd0388ae | 414 | |
bd0388ae | 415 | if (unlikely(OPTION_MD5 & options)) { |
1a2c6181 CP |
416 | *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | |
417 | (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); | |
bd0388ae WAS |
418 | /* overload cookie hash location */ |
419 | opts->hash_location = (__u8 *)ptr; | |
33ad798c | 420 | ptr += 4; |
40efc6fa | 421 | } |
33ad798c | 422 | |
fd6149d3 IJ |
423 | if (unlikely(opts->mss)) { |
424 | *ptr++ = htonl((TCPOPT_MSS << 24) | | |
425 | (TCPOLEN_MSS << 16) | | |
426 | opts->mss); | |
427 | } | |
428 | ||
bd0388ae WAS |
429 | if (likely(OPTION_TS & options)) { |
430 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { | |
33ad798c AL |
431 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
432 | (TCPOLEN_SACK_PERM << 16) | | |
433 | (TCPOPT_TIMESTAMP << 8) | | |
434 | TCPOLEN_TIMESTAMP); | |
bd0388ae | 435 | options &= ~OPTION_SACK_ADVERTISE; |
33ad798c AL |
436 | } else { |
437 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
438 | (TCPOPT_NOP << 16) | | |
439 | (TCPOPT_TIMESTAMP << 8) | | |
440 | TCPOLEN_TIMESTAMP); | |
441 | } | |
442 | *ptr++ = htonl(opts->tsval); | |
443 | *ptr++ = htonl(opts->tsecr); | |
444 | } | |
445 | ||
bd0388ae | 446 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { |
33ad798c AL |
447 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
448 | (TCPOPT_NOP << 16) | | |
449 | (TCPOPT_SACK_PERM << 8) | | |
450 | TCPOLEN_SACK_PERM); | |
451 | } | |
452 | ||
bd0388ae | 453 | if (unlikely(OPTION_WSCALE & options)) { |
33ad798c AL |
454 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
455 | (TCPOPT_WINDOW << 16) | | |
456 | (TCPOLEN_WINDOW << 8) | | |
457 | opts->ws); | |
458 | } | |
459 | ||
460 | if (unlikely(opts->num_sack_blocks)) { | |
461 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? | |
462 | tp->duplicate_sack : tp->selective_acks; | |
40efc6fa SH |
463 | int this_sack; |
464 | ||
465 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
466 | (TCPOPT_NOP << 16) | | |
467 | (TCPOPT_SACK << 8) | | |
33ad798c | 468 | (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * |
40efc6fa | 469 | TCPOLEN_SACK_PERBLOCK))); |
2de979bd | 470 | |
33ad798c AL |
471 | for (this_sack = 0; this_sack < opts->num_sack_blocks; |
472 | ++this_sack) { | |
40efc6fa SH |
473 | *ptr++ = htonl(sp[this_sack].start_seq); |
474 | *ptr++ = htonl(sp[this_sack].end_seq); | |
475 | } | |
2de979bd | 476 | |
5861f8e5 | 477 | tp->rx_opt.dsack = 0; |
40efc6fa | 478 | } |
2100c8d2 YC |
479 | |
480 | if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { | |
481 | struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; | |
482 | ||
483 | *ptr++ = htonl((TCPOPT_EXP << 24) | | |
484 | ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) | | |
485 | TCPOPT_FASTOPEN_MAGIC); | |
486 | ||
487 | memcpy(ptr, foc->val, foc->len); | |
488 | if ((foc->len & 3) == 2) { | |
489 | u8 *align = ((u8 *)ptr) + foc->len; | |
490 | align[0] = align[1] = TCPOPT_NOP; | |
491 | } | |
492 | ptr += (foc->len + 3) >> 2; | |
493 | } | |
33ad798c AL |
494 | } |
495 | ||
67edfef7 AK |
496 | /* Compute TCP options for SYN packets. This is not the final |
497 | * network wire format yet. | |
498 | */ | |
95c96174 | 499 | static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, |
33ad798c | 500 | struct tcp_out_options *opts, |
cf533ea5 ED |
501 | struct tcp_md5sig_key **md5) |
502 | { | |
33ad798c | 503 | struct tcp_sock *tp = tcp_sk(sk); |
95c96174 | 504 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
783237e8 | 505 | struct tcp_fastopen_request *fastopen = tp->fastopen_req; |
33ad798c | 506 | |
cfb6eeb4 | 507 | #ifdef CONFIG_TCP_MD5SIG |
33ad798c AL |
508 | *md5 = tp->af_specific->md5_lookup(sk, sk); |
509 | if (*md5) { | |
510 | opts->options |= OPTION_MD5; | |
bd0388ae | 511 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
cfb6eeb4 | 512 | } |
33ad798c AL |
513 | #else |
514 | *md5 = NULL; | |
cfb6eeb4 | 515 | #endif |
33ad798c AL |
516 | |
517 | /* We always get an MSS option. The option bytes which will be seen in | |
518 | * normal data packets should timestamps be used, must be in the MSS | |
519 | * advertised. But we subtract them from tp->mss_cache so that | |
520 | * calculations in tcp_sendmsg are simpler etc. So account for this | |
521 | * fact here if necessary. If we don't do this correctly, as a | |
522 | * receiver we won't recognize data packets as being full sized when we | |
523 | * should, and thus we won't abide by the delayed ACK rules correctly. | |
524 | * SACKs don't matter, we never delay an ACK when we have any of those | |
525 | * going out. */ | |
526 | opts->mss = tcp_advertise_mss(sk); | |
bd0388ae | 527 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c | 528 | |
bb5b7c11 | 529 | if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { |
33ad798c | 530 | opts->options |= OPTION_TS; |
ee684b6f | 531 | opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset; |
33ad798c | 532 | opts->tsecr = tp->rx_opt.ts_recent; |
bd0388ae | 533 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c | 534 | } |
bb5b7c11 | 535 | if (likely(sysctl_tcp_window_scaling)) { |
33ad798c | 536 | opts->ws = tp->rx_opt.rcv_wscale; |
89e95a61 | 537 | opts->options |= OPTION_WSCALE; |
bd0388ae | 538 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 539 | } |
bb5b7c11 | 540 | if (likely(sysctl_tcp_sack)) { |
33ad798c | 541 | opts->options |= OPTION_SACK_ADVERTISE; |
b32d1310 | 542 | if (unlikely(!(OPTION_TS & opts->options))) |
bd0388ae | 543 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c AL |
544 | } |
545 | ||
783237e8 YC |
546 | if (fastopen && fastopen->cookie.len >= 0) { |
547 | u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len; | |
548 | need = (need + 3) & ~3U; /* Align to 32 bits */ | |
549 | if (remaining >= need) { | |
550 | opts->options |= OPTION_FAST_OPEN_COOKIE; | |
551 | opts->fastopen_cookie = &fastopen->cookie; | |
552 | remaining -= need; | |
553 | tp->syn_fastopen = 1; | |
554 | } | |
555 | } | |
bd0388ae | 556 | |
bd0388ae | 557 | return MAX_TCP_OPTION_SPACE - remaining; |
40efc6fa SH |
558 | } |
559 | ||
67edfef7 | 560 | /* Set up TCP options for SYN-ACKs. */ |
95c96174 | 561 | static unsigned int tcp_synack_options(struct sock *sk, |
33ad798c | 562 | struct request_sock *req, |
95c96174 | 563 | unsigned int mss, struct sk_buff *skb, |
33ad798c | 564 | struct tcp_out_options *opts, |
4957faad | 565 | struct tcp_md5sig_key **md5, |
8336886f | 566 | struct tcp_fastopen_cookie *foc) |
4957faad | 567 | { |
33ad798c | 568 | struct inet_request_sock *ireq = inet_rsk(req); |
95c96174 | 569 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
33ad798c | 570 | |
cfb6eeb4 | 571 | #ifdef CONFIG_TCP_MD5SIG |
33ad798c AL |
572 | *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); |
573 | if (*md5) { | |
574 | opts->options |= OPTION_MD5; | |
4957faad WAS |
575 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
576 | ||
577 | /* We can't fit any SACK blocks in a packet with MD5 + TS | |
578 | * options. There was discussion about disabling SACK | |
579 | * rather than TS in order to fit in better with old, | |
580 | * buggy kernels, but that was deemed to be unnecessary. | |
581 | */ | |
de213e5e | 582 | ireq->tstamp_ok &= !ireq->sack_ok; |
cfb6eeb4 | 583 | } |
33ad798c AL |
584 | #else |
585 | *md5 = NULL; | |
cfb6eeb4 | 586 | #endif |
33ad798c | 587 | |
4957faad | 588 | /* We always send an MSS option. */ |
33ad798c | 589 | opts->mss = mss; |
4957faad | 590 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c AL |
591 | |
592 | if (likely(ireq->wscale_ok)) { | |
593 | opts->ws = ireq->rcv_wscale; | |
89e95a61 | 594 | opts->options |= OPTION_WSCALE; |
4957faad | 595 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 596 | } |
de213e5e | 597 | if (likely(ireq->tstamp_ok)) { |
33ad798c AL |
598 | opts->options |= OPTION_TS; |
599 | opts->tsval = TCP_SKB_CB(skb)->when; | |
600 | opts->tsecr = req->ts_recent; | |
4957faad | 601 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c AL |
602 | } |
603 | if (likely(ireq->sack_ok)) { | |
604 | opts->options |= OPTION_SACK_ADVERTISE; | |
de213e5e | 605 | if (unlikely(!ireq->tstamp_ok)) |
4957faad | 606 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c | 607 | } |
8336886f JC |
608 | if (foc != NULL) { |
609 | u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; | |
610 | need = (need + 3) & ~3U; /* Align to 32 bits */ | |
611 | if (remaining >= need) { | |
612 | opts->options |= OPTION_FAST_OPEN_COOKIE; | |
613 | opts->fastopen_cookie = foc; | |
614 | remaining -= need; | |
615 | } | |
616 | } | |
1a2c6181 | 617 | |
4957faad | 618 | return MAX_TCP_OPTION_SPACE - remaining; |
33ad798c AL |
619 | } |
620 | ||
67edfef7 AK |
621 | /* Compute TCP options for ESTABLISHED sockets. This is not the |
622 | * final wire format yet. | |
623 | */ | |
95c96174 | 624 | static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, |
33ad798c | 625 | struct tcp_out_options *opts, |
cf533ea5 ED |
626 | struct tcp_md5sig_key **md5) |
627 | { | |
33ad798c AL |
628 | struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; |
629 | struct tcp_sock *tp = tcp_sk(sk); | |
95c96174 | 630 | unsigned int size = 0; |
cabeccbd | 631 | unsigned int eff_sacks; |
33ad798c AL |
632 | |
633 | #ifdef CONFIG_TCP_MD5SIG | |
634 | *md5 = tp->af_specific->md5_lookup(sk, sk); | |
635 | if (unlikely(*md5)) { | |
636 | opts->options |= OPTION_MD5; | |
637 | size += TCPOLEN_MD5SIG_ALIGNED; | |
638 | } | |
639 | #else | |
640 | *md5 = NULL; | |
641 | #endif | |
642 | ||
643 | if (likely(tp->rx_opt.tstamp_ok)) { | |
644 | opts->options |= OPTION_TS; | |
ee684b6f | 645 | opts->tsval = tcb ? tcb->when + tp->tsoffset : 0; |
33ad798c AL |
646 | opts->tsecr = tp->rx_opt.ts_recent; |
647 | size += TCPOLEN_TSTAMP_ALIGNED; | |
648 | } | |
649 | ||
cabeccbd IJ |
650 | eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; |
651 | if (unlikely(eff_sacks)) { | |
95c96174 | 652 | const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; |
33ad798c | 653 | opts->num_sack_blocks = |
95c96174 | 654 | min_t(unsigned int, eff_sacks, |
33ad798c AL |
655 | (remaining - TCPOLEN_SACK_BASE_ALIGNED) / |
656 | TCPOLEN_SACK_PERBLOCK); | |
657 | size += TCPOLEN_SACK_BASE_ALIGNED + | |
658 | opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; | |
659 | } | |
660 | ||
661 | return size; | |
40efc6fa | 662 | } |
1da177e4 | 663 | |
46d3ceab ED |
664 | |
665 | /* TCP SMALL QUEUES (TSQ) | |
666 | * | |
667 | * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) | |
668 | * to reduce RTT and bufferbloat. | |
669 | * We do this using a special skb destructor (tcp_wfree). | |
670 | * | |
671 | * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb | |
672 | * needs to be reallocated in a driver. | |
673 | * The invariant being skb->truesize substracted from sk->sk_wmem_alloc | |
674 | * | |
675 | * Since transmit from skb destructor is forbidden, we use a tasklet | |
676 | * to process all sockets that eventually need to send more skbs. | |
677 | * We use one tasklet per cpu, with its own queue of sockets. | |
678 | */ | |
679 | struct tsq_tasklet { | |
680 | struct tasklet_struct tasklet; | |
681 | struct list_head head; /* queue of tcp sockets */ | |
682 | }; | |
683 | static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); | |
684 | ||
6f458dfb ED |
685 | static void tcp_tsq_handler(struct sock *sk) |
686 | { | |
687 | if ((1 << sk->sk_state) & | |
688 | (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | | |
689 | TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) | |
94ee16ae JO |
690 | tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle, |
691 | 0, GFP_ATOMIC); | |
6f458dfb | 692 | } |
46d3ceab ED |
693 | /* |
694 | * One tasklest per cpu tries to send more skbs. | |
695 | * We run in tasklet context but need to disable irqs when | |
696 | * transfering tsq->head because tcp_wfree() might | |
697 | * interrupt us (non NAPI drivers) | |
698 | */ | |
699 | static void tcp_tasklet_func(unsigned long data) | |
700 | { | |
701 | struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; | |
702 | LIST_HEAD(list); | |
703 | unsigned long flags; | |
704 | struct list_head *q, *n; | |
705 | struct tcp_sock *tp; | |
706 | struct sock *sk; | |
707 | ||
708 | local_irq_save(flags); | |
709 | list_splice_init(&tsq->head, &list); | |
710 | local_irq_restore(flags); | |
711 | ||
712 | list_for_each_safe(q, n, &list) { | |
713 | tp = list_entry(q, struct tcp_sock, tsq_node); | |
714 | list_del(&tp->tsq_node); | |
715 | ||
716 | sk = (struct sock *)tp; | |
717 | bh_lock_sock(sk); | |
718 | ||
719 | if (!sock_owned_by_user(sk)) { | |
6f458dfb | 720 | tcp_tsq_handler(sk); |
46d3ceab ED |
721 | } else { |
722 | /* defer the work to tcp_release_cb() */ | |
6f458dfb | 723 | set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); |
46d3ceab ED |
724 | } |
725 | bh_unlock_sock(sk); | |
726 | ||
727 | clear_bit(TSQ_QUEUED, &tp->tsq_flags); | |
728 | sk_free(sk); | |
729 | } | |
730 | } | |
731 | ||
6f458dfb ED |
732 | #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ |
733 | (1UL << TCP_WRITE_TIMER_DEFERRED) | \ | |
563d34d0 ED |
734 | (1UL << TCP_DELACK_TIMER_DEFERRED) | \ |
735 | (1UL << TCP_MTU_REDUCED_DEFERRED)) | |
46d3ceab ED |
736 | /** |
737 | * tcp_release_cb - tcp release_sock() callback | |
738 | * @sk: socket | |
739 | * | |
740 | * called from release_sock() to perform protocol dependent | |
741 | * actions before socket release. | |
742 | */ | |
743 | void tcp_release_cb(struct sock *sk) | |
744 | { | |
745 | struct tcp_sock *tp = tcp_sk(sk); | |
6f458dfb | 746 | unsigned long flags, nflags; |
46d3ceab | 747 | |
6f458dfb ED |
748 | /* perform an atomic operation only if at least one flag is set */ |
749 | do { | |
750 | flags = tp->tsq_flags; | |
751 | if (!(flags & TCP_DEFERRED_ALL)) | |
752 | return; | |
753 | nflags = flags & ~TCP_DEFERRED_ALL; | |
754 | } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); | |
755 | ||
756 | if (flags & (1UL << TCP_TSQ_DEFERRED)) | |
757 | tcp_tsq_handler(sk); | |
758 | ||
cbbb5a25 ED |
759 | /* Here begins the tricky part : |
760 | * We are called from release_sock() with : | |
761 | * 1) BH disabled | |
762 | * 2) sk_lock.slock spinlock held | |
763 | * 3) socket owned by us (sk->sk_lock.owned == 1) | |
764 | * | |
765 | * But following code is meant to be called from BH handlers, | |
766 | * so we should keep BH disabled, but early release socket ownership | |
767 | */ | |
768 | sock_release_ownership(sk); | |
769 | ||
144d56e9 | 770 | if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { |
6f458dfb | 771 | tcp_write_timer_handler(sk); |
144d56e9 ED |
772 | __sock_put(sk); |
773 | } | |
774 | if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { | |
6f458dfb | 775 | tcp_delack_timer_handler(sk); |
144d56e9 ED |
776 | __sock_put(sk); |
777 | } | |
778 | if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { | |
5f80f4d8 | 779 | inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); |
144d56e9 ED |
780 | __sock_put(sk); |
781 | } | |
46d3ceab ED |
782 | } |
783 | EXPORT_SYMBOL(tcp_release_cb); | |
784 | ||
785 | void __init tcp_tasklet_init(void) | |
786 | { | |
787 | int i; | |
788 | ||
789 | for_each_possible_cpu(i) { | |
790 | struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); | |
791 | ||
792 | INIT_LIST_HEAD(&tsq->head); | |
793 | tasklet_init(&tsq->tasklet, | |
794 | tcp_tasklet_func, | |
795 | (unsigned long)tsq); | |
796 | } | |
797 | } | |
798 | ||
799 | /* | |
800 | * Write buffer destructor automatically called from kfree_skb. | |
801 | * We cant xmit new skbs from this context, as we might already | |
802 | * hold qdisc lock. | |
803 | */ | |
d6a4a104 | 804 | void tcp_wfree(struct sk_buff *skb) |
46d3ceab ED |
805 | { |
806 | struct sock *sk = skb->sk; | |
807 | struct tcp_sock *tp = tcp_sk(sk); | |
808 | ||
809 | if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && | |
810 | !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { | |
811 | unsigned long flags; | |
812 | struct tsq_tasklet *tsq; | |
813 | ||
814 | /* Keep a ref on socket. | |
815 | * This last ref will be released in tcp_tasklet_func() | |
816 | */ | |
817 | atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc); | |
818 | ||
819 | /* queue this socket to tasklet queue */ | |
820 | local_irq_save(flags); | |
821 | tsq = &__get_cpu_var(tsq_tasklet); | |
822 | list_add(&tp->tsq_node, &tsq->head); | |
823 | tasklet_schedule(&tsq->tasklet); | |
824 | local_irq_restore(flags); | |
825 | } else { | |
826 | sock_wfree(skb); | |
827 | } | |
828 | } | |
829 | ||
1da177e4 LT |
830 | /* This routine actually transmits TCP packets queued in by |
831 | * tcp_do_sendmsg(). This is used by both the initial | |
832 | * transmission and possible later retransmissions. | |
833 | * All SKB's seen here are completely headerless. It is our | |
834 | * job to build the TCP header, and pass the packet down to | |
835 | * IP so it can do the same plus pass the packet off to the | |
836 | * device. | |
837 | * | |
838 | * We are working here with either a clone of the original | |
839 | * SKB, or a fresh unique copy made by the retransmit engine. | |
840 | */ | |
056834d9 IJ |
841 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, |
842 | gfp_t gfp_mask) | |
1da177e4 | 843 | { |
dfb4b9dc DM |
844 | const struct inet_connection_sock *icsk = inet_csk(sk); |
845 | struct inet_sock *inet; | |
846 | struct tcp_sock *tp; | |
847 | struct tcp_skb_cb *tcb; | |
33ad798c | 848 | struct tcp_out_options opts; |
95c96174 | 849 | unsigned int tcp_options_size, tcp_header_size; |
cfb6eeb4 | 850 | struct tcp_md5sig_key *md5; |
dfb4b9dc | 851 | struct tcphdr *th; |
dfb4b9dc DM |
852 | int err; |
853 | ||
854 | BUG_ON(!skb || !tcp_skb_pcount(skb)); | |
855 | ||
856 | /* If congestion control is doing timestamping, we must | |
857 | * take such a timestamp before we potentially clone/copy. | |
858 | */ | |
164891aa | 859 | if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) |
dfb4b9dc DM |
860 | __net_timestamp(skb); |
861 | ||
862 | if (likely(clone_it)) { | |
0e280af0 ED |
863 | const struct sk_buff *fclone = skb + 1; |
864 | ||
865 | if (unlikely(skb->fclone == SKB_FCLONE_ORIG && | |
866 | fclone->fclone == SKB_FCLONE_CLONE)) | |
867 | NET_INC_STATS_BH(sock_net(sk), | |
868 | LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); | |
869 | ||
dfb4b9dc DM |
870 | if (unlikely(skb_cloned(skb))) |
871 | skb = pskb_copy(skb, gfp_mask); | |
872 | else | |
873 | skb = skb_clone(skb, gfp_mask); | |
874 | if (unlikely(!skb)) | |
875 | return -ENOBUFS; | |
876 | } | |
1da177e4 | 877 | |
dfb4b9dc DM |
878 | inet = inet_sk(sk); |
879 | tp = tcp_sk(sk); | |
880 | tcb = TCP_SKB_CB(skb); | |
33ad798c | 881 | memset(&opts, 0, sizeof(opts)); |
1da177e4 | 882 | |
4de075e0 | 883 | if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) |
33ad798c AL |
884 | tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); |
885 | else | |
886 | tcp_options_size = tcp_established_options(sk, skb, &opts, | |
887 | &md5); | |
888 | tcp_header_size = tcp_options_size + sizeof(struct tcphdr); | |
e905a9ed | 889 | |
547669d4 | 890 | if (tcp_packets_in_flight(tp) == 0) |
dfb4b9dc | 891 | tcp_ca_event(sk, CA_EVENT_TX_START); |
547669d4 ED |
892 | |
893 | /* if no packet is in qdisc/device queue, then allow XPS to select | |
894 | * another queue. | |
895 | */ | |
896 | skb->ooo_okay = sk_wmem_alloc_get(sk) == 0; | |
dfb4b9dc | 897 | |
aa8223c7 ACM |
898 | skb_push(skb, tcp_header_size); |
899 | skb_reset_transport_header(skb); | |
46d3ceab ED |
900 | |
901 | skb_orphan(skb); | |
902 | skb->sk = sk; | |
0ae5f47e | 903 | skb->destructor = tcp_wfree; |
46d3ceab | 904 | atomic_add(skb->truesize, &sk->sk_wmem_alloc); |
dfb4b9dc DM |
905 | |
906 | /* Build TCP header and checksum it. */ | |
aa8223c7 | 907 | th = tcp_hdr(skb); |
c720c7e8 ED |
908 | th->source = inet->inet_sport; |
909 | th->dest = inet->inet_dport; | |
dfb4b9dc DM |
910 | th->seq = htonl(tcb->seq); |
911 | th->ack_seq = htonl(tp->rcv_nxt); | |
df7a3b07 | 912 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
4de075e0 | 913 | tcb->tcp_flags); |
dfb4b9dc | 914 | |
4de075e0 | 915 | if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { |
dfb4b9dc DM |
916 | /* RFC1323: The window in SYN & SYN/ACK segments |
917 | * is never scaled. | |
918 | */ | |
600ff0c2 | 919 | th->window = htons(min(tp->rcv_wnd, 65535U)); |
dfb4b9dc DM |
920 | } else { |
921 | th->window = htons(tcp_select_window(sk)); | |
922 | } | |
923 | th->check = 0; | |
924 | th->urg_ptr = 0; | |
1da177e4 | 925 | |
33f5f57e | 926 | /* The urg_mode check is necessary during a below snd_una win probe */ |
7691367d HX |
927 | if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { |
928 | if (before(tp->snd_up, tcb->seq + 0x10000)) { | |
929 | th->urg_ptr = htons(tp->snd_up - tcb->seq); | |
930 | th->urg = 1; | |
931 | } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { | |
0eae88f3 | 932 | th->urg_ptr = htons(0xFFFF); |
7691367d HX |
933 | th->urg = 1; |
934 | } | |
dfb4b9dc | 935 | } |
1da177e4 | 936 | |
bd0388ae | 937 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
4de075e0 | 938 | if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) |
9e412ba7 | 939 | TCP_ECN_send(sk, skb, tcp_header_size); |
1da177e4 | 940 | |
cfb6eeb4 YH |
941 | #ifdef CONFIG_TCP_MD5SIG |
942 | /* Calculate the MD5 hash, as we have all we need now */ | |
943 | if (md5) { | |
a465419b | 944 | sk_nocaps_add(sk, NETIF_F_GSO_MASK); |
bd0388ae | 945 | tp->af_specific->calc_md5_hash(opts.hash_location, |
49a72dfb | 946 | md5, sk, NULL, skb); |
cfb6eeb4 YH |
947 | } |
948 | #endif | |
949 | ||
bb296246 | 950 | icsk->icsk_af_ops->send_check(sk, skb); |
1da177e4 | 951 | |
4de075e0 | 952 | if (likely(tcb->tcp_flags & TCPHDR_ACK)) |
dfb4b9dc | 953 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); |
1da177e4 | 954 | |
dfb4b9dc | 955 | if (skb->len != tcp_header_size) |
cf533ea5 | 956 | tcp_event_data_sent(tp, sk); |
1da177e4 | 957 | |
bd37a088 | 958 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
aa2ea058 TH |
959 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, |
960 | tcp_skb_pcount(skb)); | |
1da177e4 | 961 | |
d9d8da80 | 962 | err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl); |
83de47cd | 963 | if (likely(err <= 0)) |
dfb4b9dc DM |
964 | return err; |
965 | ||
3cfe3baa | 966 | tcp_enter_cwr(sk, 1); |
dfb4b9dc | 967 | |
b9df3cb8 | 968 | return net_xmit_eval(err); |
1da177e4 LT |
969 | } |
970 | ||
67edfef7 | 971 | /* This routine just queues the buffer for sending. |
1da177e4 LT |
972 | * |
973 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, | |
974 | * otherwise socket can stall. | |
975 | */ | |
976 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) | |
977 | { | |
978 | struct tcp_sock *tp = tcp_sk(sk); | |
979 | ||
980 | /* Advance write_seq and place onto the write_queue. */ | |
981 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; | |
982 | skb_header_release(skb); | |
fe067e8a | 983 | tcp_add_write_queue_tail(sk, skb); |
3ab224be HA |
984 | sk->sk_wmem_queued += skb->truesize; |
985 | sk_mem_charge(sk, skb->truesize); | |
1da177e4 LT |
986 | } |
987 | ||
67edfef7 | 988 | /* Initialize TSO segments for a packet. */ |
cf533ea5 | 989 | static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, |
056834d9 | 990 | unsigned int mss_now) |
f6302d1d | 991 | { |
b81908e1 ED |
992 | /* Make sure we own this skb before messing gso_size/gso_segs */ |
993 | WARN_ON_ONCE(skb_cloned(skb)); | |
994 | ||
8e5b9dda HX |
995 | if (skb->len <= mss_now || !sk_can_gso(sk) || |
996 | skb->ip_summed == CHECKSUM_NONE) { | |
f6302d1d DM |
997 | /* Avoid the costly divide in the normal |
998 | * non-TSO case. | |
999 | */ | |
7967168c HX |
1000 | skb_shinfo(skb)->gso_segs = 1; |
1001 | skb_shinfo(skb)->gso_size = 0; | |
c9af6db4 | 1002 | skb_shinfo(skb)->gso_type = 0; |
f6302d1d | 1003 | } else { |
356f89e1 | 1004 | skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); |
7967168c | 1005 | skb_shinfo(skb)->gso_size = mss_now; |
c9af6db4 | 1006 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
1da177e4 LT |
1007 | } |
1008 | } | |
1009 | ||
91fed7a1 | 1010 | /* When a modification to fackets out becomes necessary, we need to check |
68f8353b | 1011 | * skb is counted to fackets_out or not. |
91fed7a1 | 1012 | */ |
cf533ea5 | 1013 | static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, |
91fed7a1 IJ |
1014 | int decr) |
1015 | { | |
a47e5a98 IJ |
1016 | struct tcp_sock *tp = tcp_sk(sk); |
1017 | ||
dc86967b | 1018 | if (!tp->sacked_out || tcp_is_reno(tp)) |
91fed7a1 IJ |
1019 | return; |
1020 | ||
6859d494 | 1021 | if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) |
91fed7a1 | 1022 | tp->fackets_out -= decr; |
91fed7a1 IJ |
1023 | } |
1024 | ||
797108d1 IJ |
1025 | /* Pcount in the middle of the write queue got changed, we need to do various |
1026 | * tweaks to fix counters | |
1027 | */ | |
cf533ea5 | 1028 | static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) |
797108d1 IJ |
1029 | { |
1030 | struct tcp_sock *tp = tcp_sk(sk); | |
1031 | ||
1032 | tp->packets_out -= decr; | |
1033 | ||
1034 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
1035 | tp->sacked_out -= decr; | |
1036 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
1037 | tp->retrans_out -= decr; | |
1038 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) | |
1039 | tp->lost_out -= decr; | |
1040 | ||
1041 | /* Reno case is special. Sigh... */ | |
1042 | if (tcp_is_reno(tp) && decr > 0) | |
1043 | tp->sacked_out -= min_t(u32, tp->sacked_out, decr); | |
1044 | ||
1045 | tcp_adjust_fackets_out(sk, skb, decr); | |
1046 | ||
1047 | if (tp->lost_skb_hint && | |
1048 | before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && | |
52cf3cc8 | 1049 | (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) |
797108d1 IJ |
1050 | tp->lost_cnt_hint -= decr; |
1051 | ||
1052 | tcp_verify_left_out(tp); | |
1053 | } | |
1054 | ||
1da177e4 LT |
1055 | /* Function to create two new TCP segments. Shrinks the given segment |
1056 | * to the specified size and appends a new segment with the rest of the | |
e905a9ed | 1057 | * packet to the list. This won't be called frequently, I hope. |
1da177e4 LT |
1058 | * Remember, these are still headerless SKBs at this point. |
1059 | */ | |
056834d9 IJ |
1060 | int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, |
1061 | unsigned int mss_now) | |
1da177e4 LT |
1062 | { |
1063 | struct tcp_sock *tp = tcp_sk(sk); | |
1064 | struct sk_buff *buff; | |
6475be16 | 1065 | int nsize, old_factor; |
b60b49ea | 1066 | int nlen; |
9ce01461 | 1067 | u8 flags; |
1da177e4 | 1068 | |
2fceec13 IJ |
1069 | if (WARN_ON(len > skb->len)) |
1070 | return -EINVAL; | |
6a438bbe | 1071 | |
1da177e4 LT |
1072 | nsize = skb_headlen(skb) - len; |
1073 | if (nsize < 0) | |
1074 | nsize = 0; | |
1075 | ||
b81908e1 | 1076 | if (skb_unclone(skb, GFP_ATOMIC)) |
1da177e4 LT |
1077 | return -ENOMEM; |
1078 | ||
1079 | /* Get a new skb... force flag on. */ | |
1080 | buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); | |
1081 | if (buff == NULL) | |
1082 | return -ENOMEM; /* We'll just try again later. */ | |
ef5cb973 | 1083 | |
3ab224be HA |
1084 | sk->sk_wmem_queued += buff->truesize; |
1085 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea HX |
1086 | nlen = skb->len - len - nsize; |
1087 | buff->truesize += nlen; | |
1088 | skb->truesize -= nlen; | |
1da177e4 LT |
1089 | |
1090 | /* Correct the sequence numbers. */ | |
1091 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1092 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1093 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1094 | ||
1095 | /* PSH and FIN should only be set in the second packet. */ | |
4de075e0 ED |
1096 | flags = TCP_SKB_CB(skb)->tcp_flags; |
1097 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
1098 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
e14c3caf | 1099 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
1da177e4 | 1100 | |
84fa7933 | 1101 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { |
1da177e4 | 1102 | /* Copy and checksum data tail into the new buffer. */ |
056834d9 IJ |
1103 | buff->csum = csum_partial_copy_nocheck(skb->data + len, |
1104 | skb_put(buff, nsize), | |
1da177e4 LT |
1105 | nsize, 0); |
1106 | ||
1107 | skb_trim(skb, len); | |
1108 | ||
1109 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); | |
1110 | } else { | |
84fa7933 | 1111 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
1112 | skb_split(skb, buff, len); |
1113 | } | |
1114 | ||
1115 | buff->ip_summed = skb->ip_summed; | |
1116 | ||
1117 | /* Looks stupid, but our code really uses when of | |
1118 | * skbs, which it never sent before. --ANK | |
1119 | */ | |
1120 | TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; | |
a61bbcf2 | 1121 | buff->tstamp = skb->tstamp; |
1da177e4 | 1122 | |
6475be16 DM |
1123 | old_factor = tcp_skb_pcount(skb); |
1124 | ||
1da177e4 | 1125 | /* Fix up tso_factor for both original and new SKB. */ |
846998ae DM |
1126 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
1127 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
1da177e4 | 1128 | |
6475be16 DM |
1129 | /* If this packet has been sent out already, we must |
1130 | * adjust the various packet counters. | |
1131 | */ | |
cf0b450c | 1132 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
6475be16 DM |
1133 | int diff = old_factor - tcp_skb_pcount(skb) - |
1134 | tcp_skb_pcount(buff); | |
1da177e4 | 1135 | |
797108d1 IJ |
1136 | if (diff) |
1137 | tcp_adjust_pcount(sk, skb, diff); | |
1da177e4 LT |
1138 | } |
1139 | ||
1140 | /* Link BUFF into the send queue. */ | |
f44b5271 | 1141 | skb_header_release(buff); |
fe067e8a | 1142 | tcp_insert_write_queue_after(skb, buff, sk); |
1da177e4 LT |
1143 | |
1144 | return 0; | |
1145 | } | |
1146 | ||
1147 | /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c | |
1148 | * eventually). The difference is that pulled data not copied, but | |
1149 | * immediately discarded. | |
1150 | */ | |
f2911969 | 1151 | static void __pskb_trim_head(struct sk_buff *skb, int len) |
1da177e4 LT |
1152 | { |
1153 | int i, k, eat; | |
1154 | ||
4fa48bf3 ED |
1155 | eat = min_t(int, len, skb_headlen(skb)); |
1156 | if (eat) { | |
1157 | __skb_pull(skb, eat); | |
1158 | len -= eat; | |
1159 | if (!len) | |
1160 | return; | |
1161 | } | |
1da177e4 LT |
1162 | eat = len; |
1163 | k = 0; | |
056834d9 | 1164 | for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { |
9e903e08 ED |
1165 | int size = skb_frag_size(&skb_shinfo(skb)->frags[i]); |
1166 | ||
1167 | if (size <= eat) { | |
aff65da0 | 1168 | skb_frag_unref(skb, i); |
9e903e08 | 1169 | eat -= size; |
1da177e4 LT |
1170 | } else { |
1171 | skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; | |
1172 | if (eat) { | |
1173 | skb_shinfo(skb)->frags[k].page_offset += eat; | |
9e903e08 | 1174 | skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat); |
1da177e4 LT |
1175 | eat = 0; |
1176 | } | |
1177 | k++; | |
1178 | } | |
1179 | } | |
1180 | skb_shinfo(skb)->nr_frags = k; | |
1181 | ||
27a884dc | 1182 | skb_reset_tail_pointer(skb); |
1da177e4 LT |
1183 | skb->data_len -= len; |
1184 | skb->len = skb->data_len; | |
1da177e4 LT |
1185 | } |
1186 | ||
67edfef7 | 1187 | /* Remove acked data from a packet in the transmit queue. */ |
1da177e4 LT |
1188 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) |
1189 | { | |
14bbd6a5 | 1190 | if (skb_unclone(skb, GFP_ATOMIC)) |
1da177e4 LT |
1191 | return -ENOMEM; |
1192 | ||
4fa48bf3 | 1193 | __pskb_trim_head(skb, len); |
1da177e4 LT |
1194 | |
1195 | TCP_SKB_CB(skb)->seq += len; | |
84fa7933 | 1196 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
1197 | |
1198 | skb->truesize -= len; | |
1199 | sk->sk_wmem_queued -= len; | |
3ab224be | 1200 | sk_mem_uncharge(sk, len); |
1da177e4 LT |
1201 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); |
1202 | ||
5b35e1e6 | 1203 | /* Any change of skb->len requires recalculation of tso factor. */ |
1da177e4 | 1204 | if (tcp_skb_pcount(skb) > 1) |
5b35e1e6 | 1205 | tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); |
1da177e4 LT |
1206 | |
1207 | return 0; | |
1208 | } | |
1209 | ||
1b63edd6 YC |
1210 | /* Calculate MSS not accounting any TCP options. */ |
1211 | static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
5d424d5a | 1212 | { |
cf533ea5 ED |
1213 | const struct tcp_sock *tp = tcp_sk(sk); |
1214 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
5d424d5a JH |
1215 | int mss_now; |
1216 | ||
1217 | /* Calculate base mss without TCP options: | |
1218 | It is MMS_S - sizeof(tcphdr) of rfc1122 | |
1219 | */ | |
1220 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); | |
1221 | ||
67469601 ED |
1222 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
1223 | if (icsk->icsk_af_ops->net_frag_header_len) { | |
1224 | const struct dst_entry *dst = __sk_dst_get(sk); | |
1225 | ||
1226 | if (dst && dst_allfrag(dst)) | |
1227 | mss_now -= icsk->icsk_af_ops->net_frag_header_len; | |
1228 | } | |
1229 | ||
5d424d5a JH |
1230 | /* Clamp it (mss_clamp does not include tcp options) */ |
1231 | if (mss_now > tp->rx_opt.mss_clamp) | |
1232 | mss_now = tp->rx_opt.mss_clamp; | |
1233 | ||
1234 | /* Now subtract optional transport overhead */ | |
1235 | mss_now -= icsk->icsk_ext_hdr_len; | |
1236 | ||
1237 | /* Then reserve room for full set of TCP options and 8 bytes of data */ | |
1238 | if (mss_now < 48) | |
1239 | mss_now = 48; | |
5d424d5a JH |
1240 | return mss_now; |
1241 | } | |
1242 | ||
1b63edd6 YC |
1243 | /* Calculate MSS. Not accounting for SACKs here. */ |
1244 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
1245 | { | |
1246 | /* Subtract TCP options size, not including SACKs */ | |
1247 | return __tcp_mtu_to_mss(sk, pmtu) - | |
1248 | (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); | |
1249 | } | |
1250 | ||
5d424d5a | 1251 | /* Inverse of above */ |
67469601 | 1252 | int tcp_mss_to_mtu(struct sock *sk, int mss) |
5d424d5a | 1253 | { |
cf533ea5 ED |
1254 | const struct tcp_sock *tp = tcp_sk(sk); |
1255 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
5d424d5a JH |
1256 | int mtu; |
1257 | ||
1258 | mtu = mss + | |
1259 | tp->tcp_header_len + | |
1260 | icsk->icsk_ext_hdr_len + | |
1261 | icsk->icsk_af_ops->net_header_len; | |
1262 | ||
67469601 ED |
1263 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
1264 | if (icsk->icsk_af_ops->net_frag_header_len) { | |
1265 | const struct dst_entry *dst = __sk_dst_get(sk); | |
1266 | ||
1267 | if (dst && dst_allfrag(dst)) | |
1268 | mtu += icsk->icsk_af_ops->net_frag_header_len; | |
1269 | } | |
5d424d5a JH |
1270 | return mtu; |
1271 | } | |
1272 | ||
67edfef7 | 1273 | /* MTU probing init per socket */ |
5d424d5a JH |
1274 | void tcp_mtup_init(struct sock *sk) |
1275 | { | |
1276 | struct tcp_sock *tp = tcp_sk(sk); | |
1277 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1278 | ||
1279 | icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; | |
1280 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + | |
e905a9ed | 1281 | icsk->icsk_af_ops->net_header_len; |
5d424d5a JH |
1282 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); |
1283 | icsk->icsk_mtup.probe_size = 0; | |
1284 | } | |
4bc2f18b | 1285 | EXPORT_SYMBOL(tcp_mtup_init); |
5d424d5a | 1286 | |
1da177e4 LT |
1287 | /* This function synchronize snd mss to current pmtu/exthdr set. |
1288 | ||
1289 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts | |
1290 | for TCP options, but includes only bare TCP header. | |
1291 | ||
1292 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. | |
caa20d9a | 1293 | It is minimum of user_mss and mss received with SYN. |
1da177e4 LT |
1294 | It also does not include TCP options. |
1295 | ||
d83d8461 | 1296 | inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. |
1da177e4 LT |
1297 | |
1298 | tp->mss_cache is current effective sending mss, including | |
1299 | all tcp options except for SACKs. It is evaluated, | |
1300 | taking into account current pmtu, but never exceeds | |
1301 | tp->rx_opt.mss_clamp. | |
1302 | ||
1303 | NOTE1. rfc1122 clearly states that advertised MSS | |
1304 | DOES NOT include either tcp or ip options. | |
1305 | ||
d83d8461 ACM |
1306 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
1307 | are READ ONLY outside this function. --ANK (980731) | |
1da177e4 | 1308 | */ |
1da177e4 LT |
1309 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) |
1310 | { | |
1311 | struct tcp_sock *tp = tcp_sk(sk); | |
d83d8461 | 1312 | struct inet_connection_sock *icsk = inet_csk(sk); |
5d424d5a | 1313 | int mss_now; |
1da177e4 | 1314 | |
5d424d5a JH |
1315 | if (icsk->icsk_mtup.search_high > pmtu) |
1316 | icsk->icsk_mtup.search_high = pmtu; | |
1da177e4 | 1317 | |
5d424d5a | 1318 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
409d22b4 | 1319 | mss_now = tcp_bound_to_half_wnd(tp, mss_now); |
1da177e4 LT |
1320 | |
1321 | /* And store cached results */ | |
d83d8461 | 1322 | icsk->icsk_pmtu_cookie = pmtu; |
5d424d5a JH |
1323 | if (icsk->icsk_mtup.enabled) |
1324 | mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); | |
c1b4a7e6 | 1325 | tp->mss_cache = mss_now; |
1da177e4 LT |
1326 | |
1327 | return mss_now; | |
1328 | } | |
4bc2f18b | 1329 | EXPORT_SYMBOL(tcp_sync_mss); |
1da177e4 LT |
1330 | |
1331 | /* Compute the current effective MSS, taking SACKs and IP options, | |
1332 | * and even PMTU discovery events into account. | |
1da177e4 | 1333 | */ |
0c54b85f | 1334 | unsigned int tcp_current_mss(struct sock *sk) |
1da177e4 | 1335 | { |
cf533ea5 ED |
1336 | const struct tcp_sock *tp = tcp_sk(sk); |
1337 | const struct dst_entry *dst = __sk_dst_get(sk); | |
c1b4a7e6 | 1338 | u32 mss_now; |
95c96174 | 1339 | unsigned int header_len; |
33ad798c AL |
1340 | struct tcp_out_options opts; |
1341 | struct tcp_md5sig_key *md5; | |
c1b4a7e6 DM |
1342 | |
1343 | mss_now = tp->mss_cache; | |
1344 | ||
1da177e4 LT |
1345 | if (dst) { |
1346 | u32 mtu = dst_mtu(dst); | |
d83d8461 | 1347 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
1da177e4 LT |
1348 | mss_now = tcp_sync_mss(sk, mtu); |
1349 | } | |
1350 | ||
33ad798c AL |
1351 | header_len = tcp_established_options(sk, NULL, &opts, &md5) + |
1352 | sizeof(struct tcphdr); | |
1353 | /* The mss_cache is sized based on tp->tcp_header_len, which assumes | |
1354 | * some common options. If this is an odd packet (because we have SACK | |
1355 | * blocks etc) then our calculated header_len will be different, and | |
1356 | * we have to adjust mss_now correspondingly */ | |
1357 | if (header_len != tp->tcp_header_len) { | |
1358 | int delta = (int) header_len - tp->tcp_header_len; | |
1359 | mss_now -= delta; | |
1360 | } | |
cfb6eeb4 | 1361 | |
1da177e4 LT |
1362 | return mss_now; |
1363 | } | |
1364 | ||
a762a980 | 1365 | /* Congestion window validation. (RFC2861) */ |
9e412ba7 | 1366 | static void tcp_cwnd_validate(struct sock *sk) |
a762a980 | 1367 | { |
9e412ba7 | 1368 | struct tcp_sock *tp = tcp_sk(sk); |
a762a980 | 1369 | |
d436d686 | 1370 | if (tp->packets_out >= tp->snd_cwnd) { |
a762a980 DM |
1371 | /* Network is feed fully. */ |
1372 | tp->snd_cwnd_used = 0; | |
1373 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1374 | } else { | |
1375 | /* Network starves. */ | |
1376 | if (tp->packets_out > tp->snd_cwnd_used) | |
1377 | tp->snd_cwnd_used = tp->packets_out; | |
1378 | ||
15d33c07 DM |
1379 | if (sysctl_tcp_slow_start_after_idle && |
1380 | (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) | |
a762a980 DM |
1381 | tcp_cwnd_application_limited(sk); |
1382 | } | |
1383 | } | |
1384 | ||
0e3a4803 IJ |
1385 | /* Returns the portion of skb which can be sent right away without |
1386 | * introducing MSS oddities to segment boundaries. In rare cases where | |
1387 | * mss_now != mss_cache, we will request caller to create a small skb | |
1388 | * per input skb which could be mostly avoided here (if desired). | |
5ea3a748 IJ |
1389 | * |
1390 | * We explicitly want to create a request for splitting write queue tail | |
1391 | * to a small skb for Nagle purposes while avoiding unnecessary modulos, | |
1392 | * thus all the complexity (cwnd_len is always MSS multiple which we | |
1393 | * return whenever allowed by the other factors). Basically we need the | |
1394 | * modulo only when the receiver window alone is the limiting factor or | |
1395 | * when we would be allowed to send the split-due-to-Nagle skb fully. | |
0e3a4803 | 1396 | */ |
cf533ea5 | 1397 | static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb, |
1485348d | 1398 | unsigned int mss_now, unsigned int max_segs) |
c1b4a7e6 | 1399 | { |
cf533ea5 | 1400 | const struct tcp_sock *tp = tcp_sk(sk); |
1485348d | 1401 | u32 needed, window, max_len; |
c1b4a7e6 | 1402 | |
90840def | 1403 | window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
1485348d | 1404 | max_len = mss_now * max_segs; |
0e3a4803 | 1405 | |
1485348d BH |
1406 | if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) |
1407 | return max_len; | |
0e3a4803 | 1408 | |
5ea3a748 IJ |
1409 | needed = min(skb->len, window); |
1410 | ||
1485348d BH |
1411 | if (max_len <= needed) |
1412 | return max_len; | |
0e3a4803 | 1413 | |
0e3a4803 | 1414 | return needed - needed % mss_now; |
c1b4a7e6 DM |
1415 | } |
1416 | ||
1417 | /* Can at least one segment of SKB be sent right now, according to the | |
1418 | * congestion window rules? If so, return how many segments are allowed. | |
1419 | */ | |
cf533ea5 ED |
1420 | static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, |
1421 | const struct sk_buff *skb) | |
c1b4a7e6 DM |
1422 | { |
1423 | u32 in_flight, cwnd; | |
1424 | ||
1425 | /* Don't be strict about the congestion window for the final FIN. */ | |
4de075e0 ED |
1426 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && |
1427 | tcp_skb_pcount(skb) == 1) | |
c1b4a7e6 DM |
1428 | return 1; |
1429 | ||
1430 | in_flight = tcp_packets_in_flight(tp); | |
1431 | cwnd = tp->snd_cwnd; | |
1432 | if (in_flight < cwnd) | |
1433 | return (cwnd - in_flight); | |
1434 | ||
1435 | return 0; | |
1436 | } | |
1437 | ||
b595076a | 1438 | /* Initialize TSO state of a skb. |
67edfef7 | 1439 | * This must be invoked the first time we consider transmitting |
c1b4a7e6 DM |
1440 | * SKB onto the wire. |
1441 | */ | |
cf533ea5 | 1442 | static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, |
056834d9 | 1443 | unsigned int mss_now) |
c1b4a7e6 DM |
1444 | { |
1445 | int tso_segs = tcp_skb_pcount(skb); | |
1446 | ||
f8269a49 | 1447 | if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { |
846998ae | 1448 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
c1b4a7e6 DM |
1449 | tso_segs = tcp_skb_pcount(skb); |
1450 | } | |
1451 | return tso_segs; | |
1452 | } | |
1453 | ||
67edfef7 | 1454 | /* Minshall's variant of the Nagle send check. */ |
a2a385d6 | 1455 | static inline bool tcp_minshall_check(const struct tcp_sock *tp) |
c1b4a7e6 | 1456 | { |
09cb105e | 1457 | return after(tp->snd_sml, tp->snd_una) && |
c1b4a7e6 DM |
1458 | !after(tp->snd_sml, tp->snd_nxt); |
1459 | } | |
1460 | ||
a2a385d6 | 1461 | /* Return false, if packet can be sent now without violation Nagle's rules: |
c1b4a7e6 DM |
1462 | * 1. It is full sized. |
1463 | * 2. Or it contains FIN. (already checked by caller) | |
6d67e9be | 1464 | * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. |
c1b4a7e6 DM |
1465 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. |
1466 | * With Minshall's modification: all sent small packets are ACKed. | |
1467 | */ | |
a2a385d6 | 1468 | static inline bool tcp_nagle_check(const struct tcp_sock *tp, |
e905a9ed | 1469 | const struct sk_buff *skb, |
95c96174 | 1470 | unsigned int mss_now, int nonagle) |
c1b4a7e6 | 1471 | { |
a02cec21 | 1472 | return skb->len < mss_now && |
056834d9 | 1473 | ((nonagle & TCP_NAGLE_CORK) || |
a02cec21 | 1474 | (!nonagle && tp->packets_out && tcp_minshall_check(tp))); |
c1b4a7e6 DM |
1475 | } |
1476 | ||
a2a385d6 | 1477 | /* Return true if the Nagle test allows this packet to be |
c1b4a7e6 DM |
1478 | * sent now. |
1479 | */ | |
a2a385d6 ED |
1480 | static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, |
1481 | unsigned int cur_mss, int nonagle) | |
c1b4a7e6 DM |
1482 | { |
1483 | /* Nagle rule does not apply to frames, which sit in the middle of the | |
1484 | * write_queue (they have no chances to get new data). | |
1485 | * | |
1486 | * This is implemented in the callers, where they modify the 'nonagle' | |
1487 | * argument based upon the location of SKB in the send queue. | |
1488 | */ | |
1489 | if (nonagle & TCP_NAGLE_PUSH) | |
a2a385d6 | 1490 | return true; |
c1b4a7e6 | 1491 | |
9b44190d YC |
1492 | /* Don't use the nagle rule for urgent data (or for the final FIN). */ |
1493 | if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) | |
a2a385d6 | 1494 | return true; |
c1b4a7e6 DM |
1495 | |
1496 | if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) | |
a2a385d6 | 1497 | return true; |
c1b4a7e6 | 1498 | |
a2a385d6 | 1499 | return false; |
c1b4a7e6 DM |
1500 | } |
1501 | ||
1502 | /* Does at least the first segment of SKB fit into the send window? */ | |
a2a385d6 ED |
1503 | static bool tcp_snd_wnd_test(const struct tcp_sock *tp, |
1504 | const struct sk_buff *skb, | |
1505 | unsigned int cur_mss) | |
c1b4a7e6 DM |
1506 | { |
1507 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
1508 | ||
1509 | if (skb->len > cur_mss) | |
1510 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; | |
1511 | ||
90840def | 1512 | return !after(end_seq, tcp_wnd_end(tp)); |
c1b4a7e6 DM |
1513 | } |
1514 | ||
fe067e8a | 1515 | /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) |
c1b4a7e6 DM |
1516 | * should be put on the wire right now. If so, it returns the number of |
1517 | * packets allowed by the congestion window. | |
1518 | */ | |
cf533ea5 | 1519 | static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, |
c1b4a7e6 DM |
1520 | unsigned int cur_mss, int nonagle) |
1521 | { | |
cf533ea5 | 1522 | const struct tcp_sock *tp = tcp_sk(sk); |
c1b4a7e6 DM |
1523 | unsigned int cwnd_quota; |
1524 | ||
846998ae | 1525 | tcp_init_tso_segs(sk, skb, cur_mss); |
c1b4a7e6 DM |
1526 | |
1527 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) | |
1528 | return 0; | |
1529 | ||
1530 | cwnd_quota = tcp_cwnd_test(tp, skb); | |
056834d9 | 1531 | if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) |
c1b4a7e6 DM |
1532 | cwnd_quota = 0; |
1533 | ||
1534 | return cwnd_quota; | |
1535 | } | |
1536 | ||
67edfef7 | 1537 | /* Test if sending is allowed right now. */ |
a2a385d6 | 1538 | bool tcp_may_send_now(struct sock *sk) |
c1b4a7e6 | 1539 | { |
cf533ea5 | 1540 | const struct tcp_sock *tp = tcp_sk(sk); |
fe067e8a | 1541 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 | 1542 | |
a02cec21 | 1543 | return skb && |
0c54b85f | 1544 | tcp_snd_test(sk, skb, tcp_current_mss(sk), |
c1b4a7e6 | 1545 | (tcp_skb_is_last(sk, skb) ? |
a02cec21 | 1546 | tp->nonagle : TCP_NAGLE_PUSH)); |
c1b4a7e6 DM |
1547 | } |
1548 | ||
1549 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet | |
1550 | * which is put after SKB on the list. It is very much like | |
1551 | * tcp_fragment() except that it may make several kinds of assumptions | |
1552 | * in order to speed up the splitting operation. In particular, we | |
1553 | * know that all the data is in scatter-gather pages, and that the | |
1554 | * packet has never been sent out before (and thus is not cloned). | |
1555 | */ | |
056834d9 | 1556 | static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, |
c4ead4c5 | 1557 | unsigned int mss_now, gfp_t gfp) |
c1b4a7e6 DM |
1558 | { |
1559 | struct sk_buff *buff; | |
1560 | int nlen = skb->len - len; | |
9ce01461 | 1561 | u8 flags; |
c1b4a7e6 DM |
1562 | |
1563 | /* All of a TSO frame must be composed of paged data. */ | |
c8ac3774 HX |
1564 | if (skb->len != skb->data_len) |
1565 | return tcp_fragment(sk, skb, len, mss_now); | |
c1b4a7e6 | 1566 | |
c4ead4c5 | 1567 | buff = sk_stream_alloc_skb(sk, 0, gfp); |
c1b4a7e6 DM |
1568 | if (unlikely(buff == NULL)) |
1569 | return -ENOMEM; | |
1570 | ||
3ab224be HA |
1571 | sk->sk_wmem_queued += buff->truesize; |
1572 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea | 1573 | buff->truesize += nlen; |
c1b4a7e6 DM |
1574 | skb->truesize -= nlen; |
1575 | ||
1576 | /* Correct the sequence numbers. */ | |
1577 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1578 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1579 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1580 | ||
1581 | /* PSH and FIN should only be set in the second packet. */ | |
4de075e0 ED |
1582 | flags = TCP_SKB_CB(skb)->tcp_flags; |
1583 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
1584 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
c1b4a7e6 DM |
1585 | |
1586 | /* This packet was never sent out yet, so no SACK bits. */ | |
1587 | TCP_SKB_CB(buff)->sacked = 0; | |
1588 | ||
84fa7933 | 1589 | buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; |
c1b4a7e6 DM |
1590 | skb_split(skb, buff, len); |
1591 | ||
1592 | /* Fix up tso_factor for both original and new SKB. */ | |
846998ae DM |
1593 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
1594 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
c1b4a7e6 DM |
1595 | |
1596 | /* Link BUFF into the send queue. */ | |
1597 | skb_header_release(buff); | |
fe067e8a | 1598 | tcp_insert_write_queue_after(skb, buff, sk); |
c1b4a7e6 DM |
1599 | |
1600 | return 0; | |
1601 | } | |
1602 | ||
1603 | /* Try to defer sending, if possible, in order to minimize the amount | |
1604 | * of TSO splitting we do. View it as a kind of TSO Nagle test. | |
1605 | * | |
1606 | * This algorithm is from John Heffner. | |
1607 | */ | |
a2a385d6 | 1608 | static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) |
c1b4a7e6 | 1609 | { |
9e412ba7 | 1610 | struct tcp_sock *tp = tcp_sk(sk); |
6687e988 | 1611 | const struct inet_connection_sock *icsk = inet_csk(sk); |
c1b4a7e6 | 1612 | u32 send_win, cong_win, limit, in_flight; |
ad9f4f50 | 1613 | int win_divisor; |
c1b4a7e6 | 1614 | |
4de075e0 | 1615 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) |
ae8064ac | 1616 | goto send_now; |
c1b4a7e6 | 1617 | |
6687e988 | 1618 | if (icsk->icsk_ca_state != TCP_CA_Open) |
ae8064ac JH |
1619 | goto send_now; |
1620 | ||
1621 | /* Defer for less than two clock ticks. */ | |
bd515c3e | 1622 | if (tp->tso_deferred && |
a2acde07 | 1623 | (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) |
ae8064ac | 1624 | goto send_now; |
908a75c1 | 1625 | |
c1b4a7e6 DM |
1626 | in_flight = tcp_packets_in_flight(tp); |
1627 | ||
056834d9 | 1628 | BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); |
c1b4a7e6 | 1629 | |
90840def | 1630 | send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
c1b4a7e6 DM |
1631 | |
1632 | /* From in_flight test above, we know that cwnd > in_flight. */ | |
1633 | cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; | |
1634 | ||
1635 | limit = min(send_win, cong_win); | |
1636 | ||
ba244fe9 | 1637 | /* If a full-sized TSO skb can be sent, do it. */ |
1485348d | 1638 | if (limit >= min_t(unsigned int, sk->sk_gso_max_size, |
5e25ba50 | 1639 | tp->xmit_size_goal_segs * tp->mss_cache)) |
ae8064ac | 1640 | goto send_now; |
ba244fe9 | 1641 | |
62ad2761 IJ |
1642 | /* Middle in queue won't get any more data, full sendable already? */ |
1643 | if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) | |
1644 | goto send_now; | |
1645 | ||
ad9f4f50 ED |
1646 | win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); |
1647 | if (win_divisor) { | |
c1b4a7e6 DM |
1648 | u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); |
1649 | ||
1650 | /* If at least some fraction of a window is available, | |
1651 | * just use it. | |
1652 | */ | |
ad9f4f50 | 1653 | chunk /= win_divisor; |
c1b4a7e6 | 1654 | if (limit >= chunk) |
ae8064ac | 1655 | goto send_now; |
c1b4a7e6 DM |
1656 | } else { |
1657 | /* Different approach, try not to defer past a single | |
1658 | * ACK. Receiver should ACK every other full sized | |
1659 | * frame, so if we have space for more than 3 frames | |
1660 | * then send now. | |
1661 | */ | |
6b5a5c0d | 1662 | if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) |
ae8064ac | 1663 | goto send_now; |
c1b4a7e6 DM |
1664 | } |
1665 | ||
f4541d60 ED |
1666 | /* Ok, it looks like it is advisable to defer. |
1667 | * Do not rearm the timer if already set to not break TCP ACK clocking. | |
1668 | */ | |
1669 | if (!tp->tso_deferred) | |
1670 | tp->tso_deferred = 1 | (jiffies << 1); | |
ae8064ac | 1671 | |
a2a385d6 | 1672 | return true; |
ae8064ac JH |
1673 | |
1674 | send_now: | |
1675 | tp->tso_deferred = 0; | |
a2a385d6 | 1676 | return false; |
c1b4a7e6 DM |
1677 | } |
1678 | ||
5d424d5a | 1679 | /* Create a new MTU probe if we are ready. |
67edfef7 AK |
1680 | * MTU probe is regularly attempting to increase the path MTU by |
1681 | * deliberately sending larger packets. This discovers routing | |
1682 | * changes resulting in larger path MTUs. | |
1683 | * | |
5d424d5a JH |
1684 | * Returns 0 if we should wait to probe (no cwnd available), |
1685 | * 1 if a probe was sent, | |
056834d9 IJ |
1686 | * -1 otherwise |
1687 | */ | |
5d424d5a JH |
1688 | static int tcp_mtu_probe(struct sock *sk) |
1689 | { | |
1690 | struct tcp_sock *tp = tcp_sk(sk); | |
1691 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1692 | struct sk_buff *skb, *nskb, *next; | |
1693 | int len; | |
1694 | int probe_size; | |
91cc17c0 | 1695 | int size_needed; |
5d424d5a JH |
1696 | int copy; |
1697 | int mss_now; | |
1698 | ||
1699 | /* Not currently probing/verifying, | |
1700 | * not in recovery, | |
1701 | * have enough cwnd, and | |
1702 | * not SACKing (the variable headers throw things off) */ | |
1703 | if (!icsk->icsk_mtup.enabled || | |
1704 | icsk->icsk_mtup.probe_size || | |
1705 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || | |
1706 | tp->snd_cwnd < 11 || | |
cabeccbd | 1707 | tp->rx_opt.num_sacks || tp->rx_opt.dsack) |
5d424d5a JH |
1708 | return -1; |
1709 | ||
1710 | /* Very simple search strategy: just double the MSS. */ | |
0c54b85f | 1711 | mss_now = tcp_current_mss(sk); |
056834d9 | 1712 | probe_size = 2 * tp->mss_cache; |
91cc17c0 | 1713 | size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; |
5d424d5a JH |
1714 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { |
1715 | /* TODO: set timer for probe_converge_event */ | |
1716 | return -1; | |
1717 | } | |
1718 | ||
1719 | /* Have enough data in the send queue to probe? */ | |
7f9c33e5 | 1720 | if (tp->write_seq - tp->snd_nxt < size_needed) |
5d424d5a JH |
1721 | return -1; |
1722 | ||
91cc17c0 IJ |
1723 | if (tp->snd_wnd < size_needed) |
1724 | return -1; | |
90840def | 1725 | if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) |
91cc17c0 | 1726 | return 0; |
5d424d5a | 1727 | |
d67c58e9 IJ |
1728 | /* Do we need to wait to drain cwnd? With none in flight, don't stall */ |
1729 | if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { | |
1730 | if (!tcp_packets_in_flight(tp)) | |
5d424d5a JH |
1731 | return -1; |
1732 | else | |
1733 | return 0; | |
1734 | } | |
1735 | ||
1736 | /* We're allowed to probe. Build it now. */ | |
1737 | if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) | |
1738 | return -1; | |
3ab224be HA |
1739 | sk->sk_wmem_queued += nskb->truesize; |
1740 | sk_mem_charge(sk, nskb->truesize); | |
5d424d5a | 1741 | |
fe067e8a | 1742 | skb = tcp_send_head(sk); |
5d424d5a JH |
1743 | |
1744 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; | |
1745 | TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; | |
4de075e0 | 1746 | TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; |
5d424d5a JH |
1747 | TCP_SKB_CB(nskb)->sacked = 0; |
1748 | nskb->csum = 0; | |
84fa7933 | 1749 | nskb->ip_summed = skb->ip_summed; |
5d424d5a | 1750 | |
50c4817e IJ |
1751 | tcp_insert_write_queue_before(nskb, skb, sk); |
1752 | ||
5d424d5a | 1753 | len = 0; |
234b6860 | 1754 | tcp_for_write_queue_from_safe(skb, next, sk) { |
5d424d5a | 1755 | copy = min_t(int, skb->len, probe_size - len); |
d318f82f | 1756 | if (nskb->ip_summed) { |
5d424d5a | 1757 | skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); |
d318f82f DCS |
1758 | } else { |
1759 | __wsum csum = skb_copy_and_csum_bits(skb, 0, | |
1760 | skb_put(nskb, copy), | |
1761 | copy, 0); | |
1762 | nskb->csum = csum_block_add(nskb->csum, csum, len); | |
1763 | } | |
5d424d5a JH |
1764 | |
1765 | if (skb->len <= copy) { | |
1766 | /* We've eaten all the data from this skb. | |
1767 | * Throw it away. */ | |
4de075e0 | 1768 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; |
fe067e8a | 1769 | tcp_unlink_write_queue(skb, sk); |
3ab224be | 1770 | sk_wmem_free_skb(sk, skb); |
5d424d5a | 1771 | } else { |
4de075e0 | 1772 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & |
a3433f35 | 1773 | ~(TCPHDR_FIN|TCPHDR_PSH); |
5d424d5a JH |
1774 | if (!skb_shinfo(skb)->nr_frags) { |
1775 | skb_pull(skb, copy); | |
84fa7933 | 1776 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
056834d9 IJ |
1777 | skb->csum = csum_partial(skb->data, |
1778 | skb->len, 0); | |
5d424d5a JH |
1779 | } else { |
1780 | __pskb_trim_head(skb, copy); | |
1781 | tcp_set_skb_tso_segs(sk, skb, mss_now); | |
1782 | } | |
1783 | TCP_SKB_CB(skb)->seq += copy; | |
1784 | } | |
1785 | ||
1786 | len += copy; | |
234b6860 IJ |
1787 | |
1788 | if (len >= probe_size) | |
1789 | break; | |
5d424d5a JH |
1790 | } |
1791 | tcp_init_tso_segs(sk, nskb, nskb->len); | |
1792 | ||
1793 | /* We're ready to send. If this fails, the probe will | |
1794 | * be resegmented into mss-sized pieces by tcp_write_xmit(). */ | |
1795 | TCP_SKB_CB(nskb)->when = tcp_time_stamp; | |
1796 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { | |
1797 | /* Decrement cwnd here because we are sending | |
056834d9 | 1798 | * effectively two packets. */ |
5d424d5a | 1799 | tp->snd_cwnd--; |
66f5fe62 | 1800 | tcp_event_new_data_sent(sk, nskb); |
5d424d5a JH |
1801 | |
1802 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); | |
0e7b1368 JH |
1803 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; |
1804 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; | |
5d424d5a JH |
1805 | |
1806 | return 1; | |
1807 | } | |
1808 | ||
1809 | return -1; | |
1810 | } | |
1811 | ||
1da177e4 LT |
1812 | /* This routine writes packets to the network. It advances the |
1813 | * send_head. This happens as incoming acks open up the remote | |
1814 | * window for us. | |
1815 | * | |
f8269a49 IJ |
1816 | * LARGESEND note: !tcp_urg_mode is overkill, only frames between |
1817 | * snd_up-64k-mss .. snd_up cannot be large. However, taking into | |
1818 | * account rare use of URG, this is not a big flaw. | |
1819 | * | |
6ba8a3b1 ND |
1820 | * Send at most one packet when push_one > 0. Temporarily ignore |
1821 | * cwnd limit to force at most one packet out when push_one == 2. | |
1822 | ||
a2a385d6 ED |
1823 | * Returns true, if no segments are in flight and we have queued segments, |
1824 | * but cannot send anything now because of SWS or another problem. | |
1da177e4 | 1825 | */ |
a2a385d6 ED |
1826 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
1827 | int push_one, gfp_t gfp) | |
1da177e4 LT |
1828 | { |
1829 | struct tcp_sock *tp = tcp_sk(sk); | |
92df7b51 | 1830 | struct sk_buff *skb; |
c1b4a7e6 DM |
1831 | unsigned int tso_segs, sent_pkts; |
1832 | int cwnd_quota; | |
5d424d5a | 1833 | int result; |
1da177e4 | 1834 | |
92df7b51 | 1835 | sent_pkts = 0; |
5d424d5a | 1836 | |
d5dd9175 IJ |
1837 | if (!push_one) { |
1838 | /* Do MTU probing. */ | |
1839 | result = tcp_mtu_probe(sk); | |
1840 | if (!result) { | |
a2a385d6 | 1841 | return false; |
d5dd9175 IJ |
1842 | } else if (result > 0) { |
1843 | sent_pkts = 1; | |
1844 | } | |
5d424d5a JH |
1845 | } |
1846 | ||
fe067e8a | 1847 | while ((skb = tcp_send_head(sk))) { |
c8ac3774 HX |
1848 | unsigned int limit; |
1849 | ||
b68e9f85 | 1850 | tso_segs = tcp_init_tso_segs(sk, skb, mss_now); |
c1b4a7e6 | 1851 | BUG_ON(!tso_segs); |
aa93466b | 1852 | |
ec342325 AV |
1853 | if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) |
1854 | goto repair; /* Skip network transmission */ | |
1855 | ||
b68e9f85 | 1856 | cwnd_quota = tcp_cwnd_test(tp, skb); |
6ba8a3b1 ND |
1857 | if (!cwnd_quota) { |
1858 | if (push_one == 2) | |
1859 | /* Force out a loss probe pkt. */ | |
1860 | cwnd_quota = 1; | |
1861 | else | |
1862 | break; | |
1863 | } | |
b68e9f85 HX |
1864 | |
1865 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) | |
1866 | break; | |
1867 | ||
afd67b68 | 1868 | if (tso_segs == 1 || !sk->sk_gso_max_segs) { |
c1b4a7e6 DM |
1869 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, |
1870 | (tcp_skb_is_last(sk, skb) ? | |
1871 | nonagle : TCP_NAGLE_PUSH)))) | |
1872 | break; | |
1873 | } else { | |
d5dd9175 | 1874 | if (!push_one && tcp_tso_should_defer(sk, skb)) |
c1b4a7e6 DM |
1875 | break; |
1876 | } | |
aa93466b | 1877 | |
0ae5f47e ED |
1878 | /* TCP Small Queues : |
1879 | * Control number of packets in qdisc/devices to two packets / or ~1 ms. | |
1880 | * This allows for : | |
1881 | * - better RTT estimation and ACK scheduling | |
1882 | * - faster recovery | |
1883 | * - high rates | |
6ef30bda ED |
1884 | * Alas, some drivers / subsystems require a fair amount |
1885 | * of queued bytes to ensure line rate. | |
1886 | * One example is wifi aggregation (802.11 AMPDU) | |
46d3ceab | 1887 | */ |
6ef30bda ED |
1888 | limit = max_t(unsigned int, sysctl_tcp_limit_output_bytes, |
1889 | sk->sk_pacing_rate >> 10); | |
0ae5f47e ED |
1890 | |
1891 | if (atomic_read(&sk->sk_wmem_alloc) > limit) { | |
46d3ceab | 1892 | set_bit(TSQ_THROTTLED, &tp->tsq_flags); |
94ee16ae JO |
1893 | /* It is possible TX completion already happened |
1894 | * before we set TSQ_THROTTLED, so we must | |
1895 | * test again the condition. | |
1896 | * We abuse smp_mb__after_clear_bit() because | |
1897 | * there is no smp_mb__after_set_bit() yet | |
1898 | */ | |
1899 | smp_mb__after_clear_bit(); | |
1900 | if (atomic_read(&sk->sk_wmem_alloc) > limit) | |
1901 | break; | |
46d3ceab | 1902 | } |
0ae5f47e | 1903 | |
c8ac3774 | 1904 | limit = mss_now; |
afd67b68 | 1905 | if (tso_segs > 1 && sk->sk_gso_max_segs && !tcp_urg_mode(tp)) |
0e3a4803 | 1906 | limit = tcp_mss_split_point(sk, skb, mss_now, |
1485348d BH |
1907 | min_t(unsigned int, |
1908 | cwnd_quota, | |
1909 | sk->sk_gso_max_segs)); | |
1da177e4 | 1910 | |
c8ac3774 | 1911 | if (skb->len > limit && |
c4ead4c5 | 1912 | unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) |
c8ac3774 HX |
1913 | break; |
1914 | ||
92df7b51 | 1915 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
c1b4a7e6 | 1916 | |
d5dd9175 | 1917 | if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) |
92df7b51 | 1918 | break; |
1da177e4 | 1919 | |
ec342325 | 1920 | repair: |
92df7b51 DM |
1921 | /* Advance the send_head. This one is sent out. |
1922 | * This call will increment packets_out. | |
1923 | */ | |
66f5fe62 | 1924 | tcp_event_new_data_sent(sk, skb); |
1da177e4 | 1925 | |
92df7b51 | 1926 | tcp_minshall_update(tp, mss_now, skb); |
a262f0cd | 1927 | sent_pkts += tcp_skb_pcount(skb); |
d5dd9175 IJ |
1928 | |
1929 | if (push_one) | |
1930 | break; | |
92df7b51 | 1931 | } |
1da177e4 | 1932 | |
aa93466b | 1933 | if (likely(sent_pkts)) { |
684bad11 YC |
1934 | if (tcp_in_cwnd_reduction(sk)) |
1935 | tp->prr_out += sent_pkts; | |
6ba8a3b1 ND |
1936 | |
1937 | /* Send one loss probe per tail loss episode. */ | |
1938 | if (push_one != 2) | |
1939 | tcp_schedule_loss_probe(sk); | |
9e412ba7 | 1940 | tcp_cwnd_validate(sk); |
a2a385d6 | 1941 | return false; |
1da177e4 | 1942 | } |
6ba8a3b1 ND |
1943 | return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk)); |
1944 | } | |
1945 | ||
1946 | bool tcp_schedule_loss_probe(struct sock *sk) | |
1947 | { | |
6ba8a3b1 | 1948 | struct tcp_sock *tp = tcp_sk(sk); |
6ba8a3b1 | 1949 | u32 rtt = tp->srtt >> 3; |
aa923a7b | 1950 | u32 timeout, rto_delta; |
6ba8a3b1 | 1951 | |
6ba8a3b1 ND |
1952 | /* Don't do any loss probe on a Fast Open connection before 3WHS |
1953 | * finishes. | |
1954 | */ | |
1955 | if (sk->sk_state == TCP_SYN_RECV) | |
1956 | return false; | |
1957 | ||
6ba8a3b1 ND |
1958 | /* Schedule a loss probe in 2*RTT for SACK capable connections |
1959 | * in Open state, that are either limited by cwnd or application. | |
1960 | */ | |
1961 | if (sysctl_tcp_early_retrans < 3 || !rtt || !tp->packets_out || | |
1962 | !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) | |
1963 | return false; | |
1964 | ||
1965 | if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && | |
1966 | tcp_send_head(sk)) | |
1967 | return false; | |
1968 | ||
1969 | /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account | |
1970 | * for delayed ack when there's one outstanding packet. | |
1971 | */ | |
1972 | timeout = rtt << 1; | |
1973 | if (tp->packets_out == 1) | |
1974 | timeout = max_t(u32, timeout, | |
1975 | (rtt + (rtt >> 1) + TCP_DELACK_MAX)); | |
1976 | timeout = max_t(u32, timeout, msecs_to_jiffies(10)); | |
1977 | ||
aa923a7b NC |
1978 | /* If the RTO formula yields an earlier time, then use that time. */ |
1979 | rto_delta = tcp_rto_delta(sk); /* How far in future is RTO? */ | |
1980 | if (rto_delta > 0) | |
1981 | timeout = min_t(u32, timeout, rto_delta); | |
6ba8a3b1 ND |
1982 | |
1983 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, | |
1984 | TCP_RTO_MAX); | |
1985 | return true; | |
1986 | } | |
1987 | ||
1988 | /* When probe timeout (PTO) fires, send a new segment if one exists, else | |
1989 | * retransmit the last segment. | |
1990 | */ | |
1991 | void tcp_send_loss_probe(struct sock *sk) | |
1992 | { | |
9b717a8d | 1993 | struct tcp_sock *tp = tcp_sk(sk); |
6ba8a3b1 ND |
1994 | struct sk_buff *skb; |
1995 | int pcount; | |
1996 | int mss = tcp_current_mss(sk); | |
1997 | int err = -1; | |
1998 | ||
1999 | if (tcp_send_head(sk) != NULL) { | |
2000 | err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); | |
2001 | goto rearm_timer; | |
2002 | } | |
2003 | ||
9b717a8d ND |
2004 | /* At most one outstanding TLP retransmission. */ |
2005 | if (tp->tlp_high_seq) | |
2006 | goto rearm_timer; | |
2007 | ||
6ba8a3b1 ND |
2008 | /* Retransmit last segment. */ |
2009 | skb = tcp_write_queue_tail(sk); | |
2010 | if (WARN_ON(!skb)) | |
2011 | goto rearm_timer; | |
2012 | ||
2013 | pcount = tcp_skb_pcount(skb); | |
2014 | if (WARN_ON(!pcount)) | |
2015 | goto rearm_timer; | |
2016 | ||
2017 | if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { | |
2018 | if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss))) | |
2019 | goto rearm_timer; | |
2020 | skb = tcp_write_queue_tail(sk); | |
2021 | } | |
2022 | ||
2023 | if (WARN_ON(!skb || !tcp_skb_pcount(skb))) | |
2024 | goto rearm_timer; | |
2025 | ||
dc5a1700 | 2026 | err = __tcp_retransmit_skb(sk, skb); |
6ba8a3b1 | 2027 | |
9b717a8d ND |
2028 | /* Record snd_nxt for loss detection. */ |
2029 | if (likely(!err)) | |
2030 | tp->tlp_high_seq = tp->snd_nxt; | |
2031 | ||
6ba8a3b1 ND |
2032 | rearm_timer: |
2033 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, | |
2034 | inet_csk(sk)->icsk_rto, | |
2035 | TCP_RTO_MAX); | |
2036 | ||
2037 | if (likely(!err)) | |
2038 | NET_INC_STATS_BH(sock_net(sk), | |
2039 | LINUX_MIB_TCPLOSSPROBES); | |
2040 | return; | |
1da177e4 LT |
2041 | } |
2042 | ||
a762a980 DM |
2043 | /* Push out any pending frames which were held back due to |
2044 | * TCP_CORK or attempt at coalescing tiny packets. | |
2045 | * The socket must be locked by the caller. | |
2046 | */ | |
9e412ba7 IJ |
2047 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, |
2048 | int nonagle) | |
a762a980 | 2049 | { |
726e07a8 IJ |
2050 | /* If we are closed, the bytes will have to remain here. |
2051 | * In time closedown will finish, we empty the write queue and | |
2052 | * all will be happy. | |
2053 | */ | |
2054 | if (unlikely(sk->sk_state == TCP_CLOSE)) | |
2055 | return; | |
2056 | ||
99a1dec7 MG |
2057 | if (tcp_write_xmit(sk, cur_mss, nonagle, 0, |
2058 | sk_gfp_atomic(sk, GFP_ATOMIC))) | |
726e07a8 | 2059 | tcp_check_probe_timer(sk); |
a762a980 DM |
2060 | } |
2061 | ||
c1b4a7e6 DM |
2062 | /* Send _single_ skb sitting at the send head. This function requires |
2063 | * true push pending frames to setup probe timer etc. | |
2064 | */ | |
2065 | void tcp_push_one(struct sock *sk, unsigned int mss_now) | |
2066 | { | |
fe067e8a | 2067 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 DM |
2068 | |
2069 | BUG_ON(!skb || skb->len < mss_now); | |
2070 | ||
d5dd9175 | 2071 | tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); |
c1b4a7e6 DM |
2072 | } |
2073 | ||
1da177e4 LT |
2074 | /* This function returns the amount that we can raise the |
2075 | * usable window based on the following constraints | |
e905a9ed | 2076 | * |
1da177e4 LT |
2077 | * 1. The window can never be shrunk once it is offered (RFC 793) |
2078 | * 2. We limit memory per socket | |
2079 | * | |
2080 | * RFC 1122: | |
2081 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep | |
2082 | * RECV.NEXT + RCV.WIN fixed until: | |
2083 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" | |
2084 | * | |
2085 | * i.e. don't raise the right edge of the window until you can raise | |
2086 | * it at least MSS bytes. | |
2087 | * | |
2088 | * Unfortunately, the recommended algorithm breaks header prediction, | |
2089 | * since header prediction assumes th->window stays fixed. | |
2090 | * | |
2091 | * Strictly speaking, keeping th->window fixed violates the receiver | |
2092 | * side SWS prevention criteria. The problem is that under this rule | |
2093 | * a stream of single byte packets will cause the right side of the | |
2094 | * window to always advance by a single byte. | |
e905a9ed | 2095 | * |
1da177e4 LT |
2096 | * Of course, if the sender implements sender side SWS prevention |
2097 | * then this will not be a problem. | |
e905a9ed | 2098 | * |
1da177e4 | 2099 | * BSD seems to make the following compromise: |
e905a9ed | 2100 | * |
1da177e4 LT |
2101 | * If the free space is less than the 1/4 of the maximum |
2102 | * space available and the free space is less than 1/2 mss, | |
2103 | * then set the window to 0. | |
2104 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] | |
2105 | * Otherwise, just prevent the window from shrinking | |
2106 | * and from being larger than the largest representable value. | |
2107 | * | |
2108 | * This prevents incremental opening of the window in the regime | |
2109 | * where TCP is limited by the speed of the reader side taking | |
2110 | * data out of the TCP receive queue. It does nothing about | |
2111 | * those cases where the window is constrained on the sender side | |
2112 | * because the pipeline is full. | |
2113 | * | |
2114 | * BSD also seems to "accidentally" limit itself to windows that are a | |
2115 | * multiple of MSS, at least until the free space gets quite small. | |
2116 | * This would appear to be a side effect of the mbuf implementation. | |
2117 | * Combining these two algorithms results in the observed behavior | |
2118 | * of having a fixed window size at almost all times. | |
2119 | * | |
2120 | * Below we obtain similar behavior by forcing the offered window to | |
2121 | * a multiple of the mss when it is feasible to do so. | |
2122 | * | |
2123 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. | |
2124 | * Regular options like TIMESTAMP are taken into account. | |
2125 | */ | |
2126 | u32 __tcp_select_window(struct sock *sk) | |
2127 | { | |
463c84b9 | 2128 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 2129 | struct tcp_sock *tp = tcp_sk(sk); |
caa20d9a | 2130 | /* MSS for the peer's data. Previous versions used mss_clamp |
1da177e4 LT |
2131 | * here. I don't know if the value based on our guesses |
2132 | * of peer's MSS is better for the performance. It's more correct | |
2133 | * but may be worse for the performance because of rcv_mss | |
2134 | * fluctuations. --SAW 1998/11/1 | |
2135 | */ | |
463c84b9 | 2136 | int mss = icsk->icsk_ack.rcv_mss; |
1da177e4 LT |
2137 | int free_space = tcp_space(sk); |
2138 | int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); | |
2139 | int window; | |
2140 | ||
5c235ec4 | 2141 | if (unlikely(mss > full_space)) { |
e905a9ed | 2142 | mss = full_space; |
5c235ec4 ED |
2143 | if (mss <= 0) |
2144 | return 0; | |
2145 | } | |
b92edbe0 | 2146 | if (free_space < (full_space >> 1)) { |
463c84b9 | 2147 | icsk->icsk_ack.quick = 0; |
1da177e4 | 2148 | |
180d8cd9 | 2149 | if (sk_under_memory_pressure(sk)) |
056834d9 IJ |
2150 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, |
2151 | 4U * tp->advmss); | |
1da177e4 LT |
2152 | |
2153 | if (free_space < mss) | |
2154 | return 0; | |
2155 | } | |
2156 | ||
2157 | if (free_space > tp->rcv_ssthresh) | |
2158 | free_space = tp->rcv_ssthresh; | |
2159 | ||
2160 | /* Don't do rounding if we are using window scaling, since the | |
2161 | * scaled window will not line up with the MSS boundary anyway. | |
2162 | */ | |
2163 | window = tp->rcv_wnd; | |
2164 | if (tp->rx_opt.rcv_wscale) { | |
2165 | window = free_space; | |
2166 | ||
2167 | /* Advertise enough space so that it won't get scaled away. | |
2168 | * Import case: prevent zero window announcement if | |
2169 | * 1<<rcv_wscale > mss. | |
2170 | */ | |
2171 | if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) | |
2172 | window = (((window >> tp->rx_opt.rcv_wscale) + 1) | |
2173 | << tp->rx_opt.rcv_wscale); | |
2174 | } else { | |
2175 | /* Get the largest window that is a nice multiple of mss. | |
2176 | * Window clamp already applied above. | |
2177 | * If our current window offering is within 1 mss of the | |
2178 | * free space we just keep it. This prevents the divide | |
2179 | * and multiply from happening most of the time. | |
2180 | * We also don't do any window rounding when the free space | |
2181 | * is too small. | |
2182 | */ | |
2183 | if (window <= free_space - mss || window > free_space) | |
056834d9 | 2184 | window = (free_space / mss) * mss; |
84565070 | 2185 | else if (mss == full_space && |
b92edbe0 | 2186 | free_space > window + (full_space >> 1)) |
84565070 | 2187 | window = free_space; |
1da177e4 LT |
2188 | } |
2189 | ||
2190 | return window; | |
2191 | } | |
2192 | ||
4a17fc3a IJ |
2193 | /* Collapses two adjacent SKB's during retransmission. */ |
2194 | static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) | |
1da177e4 LT |
2195 | { |
2196 | struct tcp_sock *tp = tcp_sk(sk); | |
fe067e8a | 2197 | struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); |
058dc334 | 2198 | int skb_size, next_skb_size; |
1da177e4 | 2199 | |
058dc334 IJ |
2200 | skb_size = skb->len; |
2201 | next_skb_size = next_skb->len; | |
1da177e4 | 2202 | |
058dc334 | 2203 | BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); |
a6963a6b | 2204 | |
058dc334 | 2205 | tcp_highest_sack_combine(sk, next_skb, skb); |
1da177e4 | 2206 | |
058dc334 | 2207 | tcp_unlink_write_queue(next_skb, sk); |
1da177e4 | 2208 | |
058dc334 IJ |
2209 | skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), |
2210 | next_skb_size); | |
1da177e4 | 2211 | |
058dc334 IJ |
2212 | if (next_skb->ip_summed == CHECKSUM_PARTIAL) |
2213 | skb->ip_summed = CHECKSUM_PARTIAL; | |
1da177e4 | 2214 | |
058dc334 IJ |
2215 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
2216 | skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); | |
1da177e4 | 2217 | |
058dc334 IJ |
2218 | /* Update sequence range on original skb. */ |
2219 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; | |
1da177e4 | 2220 | |
e6c7d085 | 2221 | /* Merge over control information. This moves PSH/FIN etc. over */ |
4de075e0 | 2222 | TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; |
058dc334 IJ |
2223 | |
2224 | /* All done, get rid of second SKB and account for it so | |
2225 | * packet counting does not break. | |
2226 | */ | |
2227 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; | |
058dc334 IJ |
2228 | |
2229 | /* changed transmit queue under us so clear hints */ | |
ef9da47c IJ |
2230 | tcp_clear_retrans_hints_partial(tp); |
2231 | if (next_skb == tp->retransmit_skb_hint) | |
2232 | tp->retransmit_skb_hint = skb; | |
058dc334 | 2233 | |
797108d1 IJ |
2234 | tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); |
2235 | ||
058dc334 | 2236 | sk_wmem_free_skb(sk, next_skb); |
1da177e4 LT |
2237 | } |
2238 | ||
67edfef7 | 2239 | /* Check if coalescing SKBs is legal. */ |
a2a385d6 | 2240 | static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) |
4a17fc3a IJ |
2241 | { |
2242 | if (tcp_skb_pcount(skb) > 1) | |
a2a385d6 | 2243 | return false; |
4a17fc3a IJ |
2244 | /* TODO: SACK collapsing could be used to remove this condition */ |
2245 | if (skb_shinfo(skb)->nr_frags != 0) | |
a2a385d6 | 2246 | return false; |
4a17fc3a | 2247 | if (skb_cloned(skb)) |
a2a385d6 | 2248 | return false; |
4a17fc3a | 2249 | if (skb == tcp_send_head(sk)) |
a2a385d6 | 2250 | return false; |
4a17fc3a IJ |
2251 | /* Some heurestics for collapsing over SACK'd could be invented */ |
2252 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
a2a385d6 | 2253 | return false; |
4a17fc3a | 2254 | |
a2a385d6 | 2255 | return true; |
4a17fc3a IJ |
2256 | } |
2257 | ||
67edfef7 AK |
2258 | /* Collapse packets in the retransmit queue to make to create |
2259 | * less packets on the wire. This is only done on retransmission. | |
2260 | */ | |
4a17fc3a IJ |
2261 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, |
2262 | int space) | |
2263 | { | |
2264 | struct tcp_sock *tp = tcp_sk(sk); | |
2265 | struct sk_buff *skb = to, *tmp; | |
a2a385d6 | 2266 | bool first = true; |
4a17fc3a IJ |
2267 | |
2268 | if (!sysctl_tcp_retrans_collapse) | |
2269 | return; | |
4de075e0 | 2270 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) |
4a17fc3a IJ |
2271 | return; |
2272 | ||
2273 | tcp_for_write_queue_from_safe(skb, tmp, sk) { | |
2274 | if (!tcp_can_collapse(sk, skb)) | |
2275 | break; | |
2276 | ||
2277 | space -= skb->len; | |
2278 | ||
2279 | if (first) { | |
a2a385d6 | 2280 | first = false; |
4a17fc3a IJ |
2281 | continue; |
2282 | } | |
2283 | ||
2284 | if (space < 0) | |
2285 | break; | |
2286 | /* Punt if not enough space exists in the first SKB for | |
2287 | * the data in the second | |
2288 | */ | |
a21d4572 | 2289 | if (skb->len > skb_availroom(to)) |
4a17fc3a IJ |
2290 | break; |
2291 | ||
2292 | if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) | |
2293 | break; | |
2294 | ||
2295 | tcp_collapse_retrans(sk, to); | |
2296 | } | |
2297 | } | |
2298 | ||
1da177e4 LT |
2299 | /* This retransmits one SKB. Policy decisions and retransmit queue |
2300 | * state updates are done by the caller. Returns non-zero if an | |
2301 | * error occurred which prevented the send. | |
2302 | */ | |
93b174ad | 2303 | int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) |
1da177e4 LT |
2304 | { |
2305 | struct tcp_sock *tp = tcp_sk(sk); | |
5d424d5a | 2306 | struct inet_connection_sock *icsk = inet_csk(sk); |
7d227cd2 | 2307 | unsigned int cur_mss; |
1da177e4 | 2308 | |
5d424d5a JH |
2309 | /* Inconslusive MTU probe */ |
2310 | if (icsk->icsk_mtup.probe_size) { | |
2311 | icsk->icsk_mtup.probe_size = 0; | |
2312 | } | |
2313 | ||
1da177e4 | 2314 | /* Do not sent more than we queued. 1/4 is reserved for possible |
caa20d9a | 2315 | * copying overhead: fragmentation, tunneling, mangling etc. |
1da177e4 LT |
2316 | */ |
2317 | if (atomic_read(&sk->sk_wmem_alloc) > | |
93522d31 ED |
2318 | min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), |
2319 | sk->sk_sndbuf)) | |
1da177e4 LT |
2320 | return -EAGAIN; |
2321 | ||
2322 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { | |
2323 | if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
2324 | BUG(); | |
1da177e4 LT |
2325 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
2326 | return -ENOMEM; | |
2327 | } | |
2328 | ||
7d227cd2 SS |
2329 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) |
2330 | return -EHOSTUNREACH; /* Routing failure or similar. */ | |
2331 | ||
0c54b85f | 2332 | cur_mss = tcp_current_mss(sk); |
7d227cd2 | 2333 | |
1da177e4 LT |
2334 | /* If receiver has shrunk his window, and skb is out of |
2335 | * new window, do not retransmit it. The exception is the | |
2336 | * case, when window is shrunk to zero. In this case | |
2337 | * our retransmit serves as a zero window probe. | |
2338 | */ | |
9d4fb27d JP |
2339 | if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && |
2340 | TCP_SKB_CB(skb)->seq != tp->snd_una) | |
1da177e4 LT |
2341 | return -EAGAIN; |
2342 | ||
2343 | if (skb->len > cur_mss) { | |
846998ae | 2344 | if (tcp_fragment(sk, skb, cur_mss, cur_mss)) |
1da177e4 | 2345 | return -ENOMEM; /* We'll try again later. */ |
02276f3c | 2346 | } else { |
9eb9362e IJ |
2347 | int oldpcount = tcp_skb_pcount(skb); |
2348 | ||
2349 | if (unlikely(oldpcount > 1)) { | |
b81908e1 ED |
2350 | if (skb_unclone(skb, GFP_ATOMIC)) |
2351 | return -ENOMEM; | |
9eb9362e IJ |
2352 | tcp_init_tso_segs(sk, skb, cur_mss); |
2353 | tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); | |
2354 | } | |
1da177e4 LT |
2355 | } |
2356 | ||
4a17fc3a | 2357 | tcp_retrans_try_collapse(sk, skb, cur_mss); |
1da177e4 | 2358 | |
1da177e4 LT |
2359 | /* Some Solaris stacks overoptimize and ignore the FIN on a |
2360 | * retransmit when old data is attached. So strip it off | |
2361 | * since it is cheap to do so and saves bytes on the network. | |
2362 | */ | |
2de979bd | 2363 | if (skb->len > 0 && |
4de075e0 | 2364 | (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && |
2de979bd | 2365 | tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { |
1da177e4 | 2366 | if (!pskb_trim(skb, 0)) { |
e870a8ef IJ |
2367 | /* Reuse, even though it does some unnecessary work */ |
2368 | tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, | |
4de075e0 | 2369 | TCP_SKB_CB(skb)->tcp_flags); |
1da177e4 | 2370 | skb->ip_summed = CHECKSUM_NONE; |
1da177e4 LT |
2371 | } |
2372 | } | |
2373 | ||
2374 | /* Make a copy, if the first transmission SKB clone we made | |
2375 | * is still in somebody's hands, else make a clone. | |
2376 | */ | |
2377 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
1da177e4 | 2378 | |
50bceae9 TG |
2379 | /* make sure skb->data is aligned on arches that require it |
2380 | * and check if ack-trimming & collapsing extended the headroom | |
2381 | * beyond what csum_start can cover. | |
2382 | */ | |
2383 | if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || | |
2384 | skb_headroom(skb) >= 0xFFFF)) { | |
117632e6 ED |
2385 | struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, |
2386 | GFP_ATOMIC); | |
93b174ad YC |
2387 | return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : |
2388 | -ENOBUFS; | |
117632e6 | 2389 | } else { |
93b174ad | 2390 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
117632e6 | 2391 | } |
93b174ad YC |
2392 | } |
2393 | ||
2394 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) | |
2395 | { | |
2396 | struct tcp_sock *tp = tcp_sk(sk); | |
2397 | int err = __tcp_retransmit_skb(sk, skb); | |
1da177e4 LT |
2398 | |
2399 | if (err == 0) { | |
2400 | /* Update global TCP statistics. */ | |
81cc8a75 | 2401 | TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); |
1da177e4 LT |
2402 | |
2403 | tp->total_retrans++; | |
2404 | ||
2405 | #if FASTRETRANS_DEBUG > 0 | |
056834d9 | 2406 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
e87cc472 | 2407 | net_dbg_ratelimited("retrans_out leaked\n"); |
1da177e4 LT |
2408 | } |
2409 | #endif | |
b08d6cb2 IJ |
2410 | if (!tp->retrans_out) |
2411 | tp->lost_retrans_low = tp->snd_nxt; | |
1da177e4 LT |
2412 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; |
2413 | tp->retrans_out += tcp_skb_pcount(skb); | |
2414 | ||
2415 | /* Save stamp of the first retransmit. */ | |
2416 | if (!tp->retrans_stamp) | |
2417 | tp->retrans_stamp = TCP_SKB_CB(skb)->when; | |
2418 | ||
1da177e4 LT |
2419 | /* snd_nxt is stored to detect loss of retransmitted segment, |
2420 | * see tcp_input.c tcp_sacktag_write_queue(). | |
2421 | */ | |
2422 | TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; | |
2423 | } | |
76fd3c89 YC |
2424 | |
2425 | if (tp->undo_retrans < 0) | |
2426 | tp->undo_retrans = 0; | |
2427 | tp->undo_retrans += tcp_skb_pcount(skb); | |
1da177e4 LT |
2428 | return err; |
2429 | } | |
2430 | ||
67edfef7 AK |
2431 | /* Check if we forward retransmits are possible in the current |
2432 | * window/congestion state. | |
2433 | */ | |
a2a385d6 | 2434 | static bool tcp_can_forward_retransmit(struct sock *sk) |
b5afe7bc IJ |
2435 | { |
2436 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
cf533ea5 | 2437 | const struct tcp_sock *tp = tcp_sk(sk); |
b5afe7bc IJ |
2438 | |
2439 | /* Forward retransmissions are possible only during Recovery. */ | |
2440 | if (icsk->icsk_ca_state != TCP_CA_Recovery) | |
a2a385d6 | 2441 | return false; |
b5afe7bc IJ |
2442 | |
2443 | /* No forward retransmissions in Reno are possible. */ | |
2444 | if (tcp_is_reno(tp)) | |
a2a385d6 | 2445 | return false; |
b5afe7bc IJ |
2446 | |
2447 | /* Yeah, we have to make difficult choice between forward transmission | |
2448 | * and retransmission... Both ways have their merits... | |
2449 | * | |
2450 | * For now we do not retransmit anything, while we have some new | |
2451 | * segments to send. In the other cases, follow rule 3 for | |
2452 | * NextSeg() specified in RFC3517. | |
2453 | */ | |
2454 | ||
2455 | if (tcp_may_send_now(sk)) | |
a2a385d6 | 2456 | return false; |
b5afe7bc | 2457 | |
a2a385d6 | 2458 | return true; |
b5afe7bc IJ |
2459 | } |
2460 | ||
1da177e4 LT |
2461 | /* This gets called after a retransmit timeout, and the initially |
2462 | * retransmitted data is acknowledged. It tries to continue | |
2463 | * resending the rest of the retransmit queue, until either | |
2464 | * we've sent it all or the congestion window limit is reached. | |
2465 | * If doing SACK, the first ACK which comes back for a timeout | |
2466 | * based retransmit packet might feed us FACK information again. | |
2467 | * If so, we use it to avoid unnecessarily retransmissions. | |
2468 | */ | |
2469 | void tcp_xmit_retransmit_queue(struct sock *sk) | |
2470 | { | |
6687e988 | 2471 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2472 | struct tcp_sock *tp = tcp_sk(sk); |
2473 | struct sk_buff *skb; | |
0e1c54c2 | 2474 | struct sk_buff *hole = NULL; |
618d9f25 | 2475 | u32 last_lost; |
61eb55f4 | 2476 | int mib_idx; |
0e1c54c2 | 2477 | int fwd_rexmitting = 0; |
6a438bbe | 2478 | |
45e77d31 IJ |
2479 | if (!tp->packets_out) |
2480 | return; | |
2481 | ||
08ebd172 IJ |
2482 | if (!tp->lost_out) |
2483 | tp->retransmit_high = tp->snd_una; | |
2484 | ||
618d9f25 | 2485 | if (tp->retransmit_skb_hint) { |
6a438bbe | 2486 | skb = tp->retransmit_skb_hint; |
618d9f25 IJ |
2487 | last_lost = TCP_SKB_CB(skb)->end_seq; |
2488 | if (after(last_lost, tp->retransmit_high)) | |
2489 | last_lost = tp->retransmit_high; | |
2490 | } else { | |
fe067e8a | 2491 | skb = tcp_write_queue_head(sk); |
618d9f25 IJ |
2492 | last_lost = tp->snd_una; |
2493 | } | |
1da177e4 | 2494 | |
08ebd172 IJ |
2495 | tcp_for_write_queue_from(skb, sk) { |
2496 | __u8 sacked = TCP_SKB_CB(skb)->sacked; | |
1da177e4 | 2497 | |
08ebd172 IJ |
2498 | if (skb == tcp_send_head(sk)) |
2499 | break; | |
2500 | /* we could do better than to assign each time */ | |
0e1c54c2 IJ |
2501 | if (hole == NULL) |
2502 | tp->retransmit_skb_hint = skb; | |
08ebd172 IJ |
2503 | |
2504 | /* Assume this retransmit will generate | |
2505 | * only one packet for congestion window | |
2506 | * calculation purposes. This works because | |
2507 | * tcp_retransmit_skb() will chop up the | |
2508 | * packet to be MSS sized and all the | |
2509 | * packet counting works out. | |
2510 | */ | |
2511 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) | |
2512 | return; | |
1da177e4 | 2513 | |
0e1c54c2 IJ |
2514 | if (fwd_rexmitting) { |
2515 | begin_fwd: | |
2516 | if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) | |
2517 | break; | |
2518 | mib_idx = LINUX_MIB_TCPFORWARDRETRANS; | |
6a438bbe | 2519 | |
0e1c54c2 | 2520 | } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { |
618d9f25 | 2521 | tp->retransmit_high = last_lost; |
0e1c54c2 IJ |
2522 | if (!tcp_can_forward_retransmit(sk)) |
2523 | break; | |
2524 | /* Backtrack if necessary to non-L'ed skb */ | |
2525 | if (hole != NULL) { | |
2526 | skb = hole; | |
2527 | hole = NULL; | |
2528 | } | |
2529 | fwd_rexmitting = 1; | |
2530 | goto begin_fwd; | |
1da177e4 | 2531 | |
0e1c54c2 | 2532 | } else if (!(sacked & TCPCB_LOST)) { |
ac11ba75 | 2533 | if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) |
0e1c54c2 IJ |
2534 | hole = skb; |
2535 | continue; | |
1da177e4 | 2536 | |
0e1c54c2 | 2537 | } else { |
618d9f25 | 2538 | last_lost = TCP_SKB_CB(skb)->end_seq; |
0e1c54c2 IJ |
2539 | if (icsk->icsk_ca_state != TCP_CA_Loss) |
2540 | mib_idx = LINUX_MIB_TCPFASTRETRANS; | |
2541 | else | |
2542 | mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; | |
2543 | } | |
1da177e4 | 2544 | |
0e1c54c2 | 2545 | if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) |
1da177e4 LT |
2546 | continue; |
2547 | ||
09e9b813 ED |
2548 | if (tcp_retransmit_skb(sk, skb)) { |
2549 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); | |
0e1c54c2 | 2550 | return; |
09e9b813 | 2551 | } |
0e1c54c2 | 2552 | NET_INC_STATS_BH(sock_net(sk), mib_idx); |
1da177e4 | 2553 | |
684bad11 | 2554 | if (tcp_in_cwnd_reduction(sk)) |
a262f0cd ND |
2555 | tp->prr_out += tcp_skb_pcount(skb); |
2556 | ||
fe067e8a | 2557 | if (skb == tcp_write_queue_head(sk)) |
3f421baa ACM |
2558 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
2559 | inet_csk(sk)->icsk_rto, | |
2560 | TCP_RTO_MAX); | |
1da177e4 LT |
2561 | } |
2562 | } | |
2563 | ||
aac9fda3 ED |
2564 | /* We allow to exceed memory limits for FIN packets to expedite |
2565 | * connection tear down and (memory) recovery. | |
bc9f0ea1 ED |
2566 | * Otherwise tcp_send_fin() could be tempted to either delay FIN |
2567 | * or even be forced to close flow without any FIN. | |
aac9fda3 ED |
2568 | */ |
2569 | static void sk_forced_wmem_schedule(struct sock *sk, int size) | |
2570 | { | |
2571 | int amt, status; | |
2572 | ||
2573 | if (size <= sk->sk_forward_alloc) | |
2574 | return; | |
2575 | amt = sk_mem_pages(size); | |
2576 | sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; | |
2577 | sk_memory_allocated_add(sk, amt, &status); | |
2578 | } | |
2579 | ||
bc9f0ea1 ED |
2580 | /* Send a FIN. The caller locks the socket for us. |
2581 | * We should try to send a FIN packet really hard, but eventually give up. | |
1da177e4 LT |
2582 | */ |
2583 | void tcp_send_fin(struct sock *sk) | |
2584 | { | |
bc9f0ea1 | 2585 | struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); |
e905a9ed | 2586 | struct tcp_sock *tp = tcp_sk(sk); |
e905a9ed | 2587 | |
bc9f0ea1 ED |
2588 | /* Optimization, tack on the FIN if we have one skb in write queue and |
2589 | * this skb was not yet sent, or we are under memory pressure. | |
2590 | * Note: in the latter case, FIN packet will be sent after a timeout, | |
2591 | * as TCP stack thinks it has already been transmitted. | |
1da177e4 | 2592 | */ |
bc9f0ea1 ED |
2593 | if (tskb && (tcp_send_head(sk) || sk_under_memory_pressure(sk))) { |
2594 | coalesce: | |
2595 | TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; | |
2596 | TCP_SKB_CB(tskb)->end_seq++; | |
1da177e4 | 2597 | tp->write_seq++; |
bc9f0ea1 ED |
2598 | if (!tcp_send_head(sk)) { |
2599 | /* This means tskb was already sent. | |
2600 | * Pretend we included the FIN on previous transmit. | |
2601 | * We need to set tp->snd_nxt to the value it would have | |
2602 | * if FIN had been sent. This is because retransmit path | |
2603 | * does not change tp->snd_nxt. | |
2604 | */ | |
2605 | tp->snd_nxt++; | |
2606 | return; | |
2607 | } | |
1da177e4 | 2608 | } else { |
bc9f0ea1 ED |
2609 | skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); |
2610 | if (unlikely(!skb)) { | |
2611 | if (tskb) | |
2612 | goto coalesce; | |
2613 | return; | |
1da177e4 | 2614 | } |
aac9fda3 ED |
2615 | skb_reserve(skb, MAX_TCP_HEADER); |
2616 | sk_forced_wmem_schedule(sk, skb->truesize); | |
1da177e4 | 2617 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ |
e870a8ef | 2618 | tcp_init_nondata_skb(skb, tp->write_seq, |
a3433f35 | 2619 | TCPHDR_ACK | TCPHDR_FIN); |
1da177e4 LT |
2620 | tcp_queue_skb(sk, skb); |
2621 | } | |
bc9f0ea1 | 2622 | __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); |
1da177e4 LT |
2623 | } |
2624 | ||
2625 | /* We get here when a process closes a file descriptor (either due to | |
2626 | * an explicit close() or as a byproduct of exit()'ing) and there | |
2627 | * was unread data in the receive queue. This behavior is recommended | |
65bb723c | 2628 | * by RFC 2525, section 2.17. -DaveM |
1da177e4 | 2629 | */ |
dd0fc66f | 2630 | void tcp_send_active_reset(struct sock *sk, gfp_t priority) |
1da177e4 | 2631 | { |
1da177e4 LT |
2632 | struct sk_buff *skb; |
2633 | ||
2634 | /* NOTE: No TCP options attached and we never retransmit this. */ | |
2635 | skb = alloc_skb(MAX_TCP_HEADER, priority); | |
2636 | if (!skb) { | |
4e673444 | 2637 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
1da177e4 LT |
2638 | return; |
2639 | } | |
2640 | ||
2641 | /* Reserve space for headers and prepare control bits. */ | |
2642 | skb_reserve(skb, MAX_TCP_HEADER); | |
e870a8ef | 2643 | tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), |
a3433f35 | 2644 | TCPHDR_ACK | TCPHDR_RST); |
1da177e4 | 2645 | /* Send it off. */ |
1da177e4 | 2646 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
dfb4b9dc | 2647 | if (tcp_transmit_skb(sk, skb, 0, priority)) |
4e673444 | 2648 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
26af65cb | 2649 | |
81cc8a75 | 2650 | TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); |
1da177e4 LT |
2651 | } |
2652 | ||
67edfef7 AK |
2653 | /* Send a crossed SYN-ACK during socket establishment. |
2654 | * WARNING: This routine must only be called when we have already sent | |
1da177e4 LT |
2655 | * a SYN packet that crossed the incoming SYN that caused this routine |
2656 | * to get called. If this assumption fails then the initial rcv_wnd | |
2657 | * and rcv_wscale values will not be correct. | |
2658 | */ | |
2659 | int tcp_send_synack(struct sock *sk) | |
2660 | { | |
056834d9 | 2661 | struct sk_buff *skb; |
1da177e4 | 2662 | |
fe067e8a | 2663 | skb = tcp_write_queue_head(sk); |
4de075e0 | 2664 | if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { |
91df42be | 2665 | pr_debug("%s: wrong queue state\n", __func__); |
1da177e4 LT |
2666 | return -EFAULT; |
2667 | } | |
4de075e0 | 2668 | if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { |
1da177e4 LT |
2669 | if (skb_cloned(skb)) { |
2670 | struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); | |
2671 | if (nskb == NULL) | |
2672 | return -ENOMEM; | |
fe067e8a | 2673 | tcp_unlink_write_queue(skb, sk); |
1da177e4 | 2674 | skb_header_release(nskb); |
fe067e8a | 2675 | __tcp_add_write_queue_head(sk, nskb); |
3ab224be HA |
2676 | sk_wmem_free_skb(sk, skb); |
2677 | sk->sk_wmem_queued += nskb->truesize; | |
2678 | sk_mem_charge(sk, nskb->truesize); | |
1da177e4 LT |
2679 | skb = nskb; |
2680 | } | |
2681 | ||
4de075e0 | 2682 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; |
1da177e4 LT |
2683 | TCP_ECN_send_synack(tcp_sk(sk), skb); |
2684 | } | |
2685 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
dfb4b9dc | 2686 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
1da177e4 LT |
2687 | } |
2688 | ||
4aea39c1 ED |
2689 | /** |
2690 | * tcp_make_synack - Prepare a SYN-ACK. | |
2691 | * sk: listener socket | |
2692 | * dst: dst entry attached to the SYNACK | |
2693 | * req: request_sock pointer | |
4aea39c1 ED |
2694 | * |
2695 | * Allocate one skb and build a SYNACK packet. | |
2696 | * @dst is consumed : Caller should not use it again. | |
2697 | */ | |
056834d9 | 2698 | struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, |
e6b4d113 | 2699 | struct request_sock *req, |
8336886f | 2700 | struct tcp_fastopen_cookie *foc) |
1da177e4 | 2701 | { |
bd0388ae | 2702 | struct tcp_out_options opts; |
2e6599cb | 2703 | struct inet_request_sock *ireq = inet_rsk(req); |
1da177e4 LT |
2704 | struct tcp_sock *tp = tcp_sk(sk); |
2705 | struct tcphdr *th; | |
1da177e4 | 2706 | struct sk_buff *skb; |
cfb6eeb4 | 2707 | struct tcp_md5sig_key *md5; |
bd0388ae | 2708 | int tcp_header_size; |
f5fff5dc | 2709 | int mss; |
1da177e4 | 2710 | |
b70a23ab | 2711 | skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC); |
4aea39c1 ED |
2712 | if (unlikely(!skb)) { |
2713 | dst_release(dst); | |
1da177e4 | 2714 | return NULL; |
4aea39c1 | 2715 | } |
1da177e4 LT |
2716 | /* Reserve space for headers. */ |
2717 | skb_reserve(skb, MAX_TCP_HEADER); | |
2718 | ||
4aea39c1 | 2719 | skb_dst_set(skb, dst); |
ca10b9e9 | 2720 | security_skb_owned_by(skb, sk); |
1da177e4 | 2721 | |
0dbaee3b | 2722 | mss = dst_metric_advmss(dst); |
f5fff5dc TQ |
2723 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) |
2724 | mss = tp->rx_opt.user_mss; | |
2725 | ||
33ad798c AL |
2726 | if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ |
2727 | __u8 rcv_wscale; | |
2728 | /* Set this up on the first call only */ | |
2729 | req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); | |
e88c64f0 HPP |
2730 | |
2731 | /* limit the window selection if the user enforce a smaller rx buffer */ | |
2732 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && | |
2733 | (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) | |
2734 | req->window_clamp = tcp_full_space(sk); | |
2735 | ||
33ad798c AL |
2736 | /* tcp_full_space because it is guaranteed to be the first packet */ |
2737 | tcp_select_initial_window(tcp_full_space(sk), | |
f5fff5dc | 2738 | mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), |
33ad798c AL |
2739 | &req->rcv_wnd, |
2740 | &req->window_clamp, | |
2741 | ireq->wscale_ok, | |
31d12926 | 2742 | &rcv_wscale, |
2743 | dst_metric(dst, RTAX_INITRWND)); | |
33ad798c AL |
2744 | ireq->rcv_wscale = rcv_wscale; |
2745 | } | |
2746 | ||
2747 | memset(&opts, 0, sizeof(opts)); | |
8b5f12d0 FW |
2748 | #ifdef CONFIG_SYN_COOKIES |
2749 | if (unlikely(req->cookie_ts)) | |
2750 | TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); | |
2751 | else | |
2752 | #endif | |
33ad798c | 2753 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
1a2c6181 CP |
2754 | tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5, |
2755 | foc) + sizeof(*th); | |
cfb6eeb4 | 2756 | |
aa8223c7 ACM |
2757 | skb_push(skb, tcp_header_size); |
2758 | skb_reset_transport_header(skb); | |
1da177e4 | 2759 | |
aa8223c7 | 2760 | th = tcp_hdr(skb); |
1da177e4 LT |
2761 | memset(th, 0, sizeof(struct tcphdr)); |
2762 | th->syn = 1; | |
2763 | th->ack = 1; | |
1da177e4 | 2764 | TCP_ECN_make_synack(req, th); |
a3116ac5 | 2765 | th->source = ireq->loc_port; |
2e6599cb | 2766 | th->dest = ireq->rmt_port; |
e870a8ef IJ |
2767 | /* Setting of flags are superfluous here for callers (and ECE is |
2768 | * not even correctly set) | |
2769 | */ | |
2770 | tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, | |
a3433f35 | 2771 | TCPHDR_SYN | TCPHDR_ACK); |
4957faad | 2772 | |
1da177e4 | 2773 | th->seq = htonl(TCP_SKB_CB(skb)->seq); |
8336886f JC |
2774 | /* XXX data is queued and acked as is. No buffer/window check */ |
2775 | th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); | |
1da177e4 LT |
2776 | |
2777 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ | |
600ff0c2 | 2778 | th->window = htons(min(req->rcv_wnd, 65535U)); |
bd0388ae | 2779 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
1da177e4 | 2780 | th->doff = (tcp_header_size >> 2); |
aa2ea058 | 2781 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb)); |
cfb6eeb4 YH |
2782 | |
2783 | #ifdef CONFIG_TCP_MD5SIG | |
2784 | /* Okay, we have all we need - do the md5 hash if needed */ | |
2785 | if (md5) { | |
bd0388ae | 2786 | tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, |
49a72dfb | 2787 | md5, NULL, req, skb); |
cfb6eeb4 YH |
2788 | } |
2789 | #endif | |
2790 | ||
ef15025a ED |
2791 | /* Do not fool tcpdump (if any), clean our debris */ |
2792 | skb->tstamp.tv64 = 0; | |
1da177e4 LT |
2793 | return skb; |
2794 | } | |
4bc2f18b | 2795 | EXPORT_SYMBOL(tcp_make_synack); |
1da177e4 | 2796 | |
67edfef7 | 2797 | /* Do all connect socket setups that can be done AF independent. */ |
370816ae | 2798 | void tcp_connect_init(struct sock *sk) |
1da177e4 | 2799 | { |
cf533ea5 | 2800 | const struct dst_entry *dst = __sk_dst_get(sk); |
1da177e4 LT |
2801 | struct tcp_sock *tp = tcp_sk(sk); |
2802 | __u8 rcv_wscale; | |
2803 | ||
2804 | /* We'll fix this up when we get a response from the other end. | |
2805 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. | |
2806 | */ | |
2807 | tp->tcp_header_len = sizeof(struct tcphdr) + | |
bb5b7c11 | 2808 | (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); |
1da177e4 | 2809 | |
cfb6eeb4 YH |
2810 | #ifdef CONFIG_TCP_MD5SIG |
2811 | if (tp->af_specific->md5_lookup(sk, sk) != NULL) | |
2812 | tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; | |
2813 | #endif | |
2814 | ||
1da177e4 LT |
2815 | /* If user gave his TCP_MAXSEG, record it to clamp */ |
2816 | if (tp->rx_opt.user_mss) | |
2817 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
2818 | tp->max_window = 0; | |
5d424d5a | 2819 | tcp_mtup_init(sk); |
1da177e4 LT |
2820 | tcp_sync_mss(sk, dst_mtu(dst)); |
2821 | ||
2822 | if (!tp->window_clamp) | |
2823 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); | |
0dbaee3b | 2824 | tp->advmss = dst_metric_advmss(dst); |
f5fff5dc TQ |
2825 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) |
2826 | tp->advmss = tp->rx_opt.user_mss; | |
2827 | ||
1da177e4 | 2828 | tcp_initialize_rcv_mss(sk); |
1da177e4 | 2829 | |
e88c64f0 HPP |
2830 | /* limit the window selection if the user enforce a smaller rx buffer */ |
2831 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && | |
2832 | (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) | |
2833 | tp->window_clamp = tcp_full_space(sk); | |
2834 | ||
1da177e4 LT |
2835 | tcp_select_initial_window(tcp_full_space(sk), |
2836 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), | |
2837 | &tp->rcv_wnd, | |
2838 | &tp->window_clamp, | |
bb5b7c11 | 2839 | sysctl_tcp_window_scaling, |
31d12926 | 2840 | &rcv_wscale, |
2841 | dst_metric(dst, RTAX_INITRWND)); | |
1da177e4 LT |
2842 | |
2843 | tp->rx_opt.rcv_wscale = rcv_wscale; | |
2844 | tp->rcv_ssthresh = tp->rcv_wnd; | |
2845 | ||
2846 | sk->sk_err = 0; | |
2847 | sock_reset_flag(sk, SOCK_DONE); | |
2848 | tp->snd_wnd = 0; | |
ee7537b6 | 2849 | tcp_init_wl(tp, 0); |
1da177e4 LT |
2850 | tp->snd_una = tp->write_seq; |
2851 | tp->snd_sml = tp->write_seq; | |
33f5f57e | 2852 | tp->snd_up = tp->write_seq; |
370816ae | 2853 | tp->snd_nxt = tp->write_seq; |
ee995283 PE |
2854 | |
2855 | if (likely(!tp->repair)) | |
2856 | tp->rcv_nxt = 0; | |
6f198dca AV |
2857 | else |
2858 | tp->rcv_tstamp = tcp_time_stamp; | |
ee995283 PE |
2859 | tp->rcv_wup = tp->rcv_nxt; |
2860 | tp->copied_seq = tp->rcv_nxt; | |
1da177e4 | 2861 | |
463c84b9 ACM |
2862 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
2863 | inet_csk(sk)->icsk_retransmits = 0; | |
1da177e4 LT |
2864 | tcp_clear_retrans(tp); |
2865 | } | |
2866 | ||
783237e8 YC |
2867 | static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) |
2868 | { | |
2869 | struct tcp_sock *tp = tcp_sk(sk); | |
2870 | struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); | |
2871 | ||
2872 | tcb->end_seq += skb->len; | |
2873 | skb_header_release(skb); | |
2874 | __tcp_add_write_queue_tail(sk, skb); | |
2875 | sk->sk_wmem_queued += skb->truesize; | |
2876 | sk_mem_charge(sk, skb->truesize); | |
2877 | tp->write_seq = tcb->end_seq; | |
2878 | tp->packets_out += tcp_skb_pcount(skb); | |
2879 | } | |
2880 | ||
2881 | /* Build and send a SYN with data and (cached) Fast Open cookie. However, | |
2882 | * queue a data-only packet after the regular SYN, such that regular SYNs | |
2883 | * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges | |
2884 | * only the SYN sequence, the data are retransmitted in the first ACK. | |
2885 | * If cookie is not cached or other error occurs, falls back to send a | |
2886 | * regular SYN with Fast Open cookie request option. | |
2887 | */ | |
2888 | static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) | |
2889 | { | |
2890 | struct tcp_sock *tp = tcp_sk(sk); | |
2891 | struct tcp_fastopen_request *fo = tp->fastopen_req; | |
e64a8519 | 2892 | int syn_loss = 0, space, err = 0; |
aab48743 | 2893 | unsigned long last_syn_loss = 0; |
e64a8519 | 2894 | struct sk_buff *syn_data; |
aab48743 | 2895 | |
67da22d2 | 2896 | tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ |
aab48743 YC |
2897 | tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, |
2898 | &syn_loss, &last_syn_loss); | |
2899 | /* Recurring FO SYN losses: revert to regular handshake temporarily */ | |
2900 | if (syn_loss > 1 && | |
2901 | time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { | |
2902 | fo->cookie.len = -1; | |
2903 | goto fallback; | |
2904 | } | |
783237e8 | 2905 | |
67da22d2 YC |
2906 | if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) |
2907 | fo->cookie.len = -1; | |
2908 | else if (fo->cookie.len <= 0) | |
783237e8 YC |
2909 | goto fallback; |
2910 | ||
2911 | /* MSS for SYN-data is based on cached MSS and bounded by PMTU and | |
2912 | * user-MSS. Reserve maximum option space for middleboxes that add | |
2913 | * private TCP options. The cost is reduced data space in SYN :( | |
2914 | */ | |
2915 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) | |
2916 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
1b63edd6 | 2917 | space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - |
783237e8 YC |
2918 | MAX_TCP_OPTION_SPACE; |
2919 | ||
fe42b170 ED |
2920 | space = min_t(size_t, space, fo->size); |
2921 | ||
2922 | /* limit to order-0 allocations */ | |
2923 | space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); | |
2924 | ||
e64a8519 ED |
2925 | syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation); |
2926 | if (!syn_data) | |
783237e8 | 2927 | goto fallback; |
e64a8519 ED |
2928 | syn_data->ip_summed = CHECKSUM_PARTIAL; |
2929 | memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); | |
75a518b0 | 2930 | skb_shinfo(syn_data)->gso_segs = 1; |
e64a8519 ED |
2931 | if (unlikely(memcpy_fromiovecend(skb_put(syn_data, space), |
2932 | fo->data->msg_iov, 0, space))) { | |
2933 | kfree_skb(syn_data); | |
2934 | goto fallback; | |
2935 | } | |
783237e8 | 2936 | |
e64a8519 ED |
2937 | /* No more data pending in inet_wait_for_connect() */ |
2938 | if (space == fo->size) | |
2939 | fo->data = NULL; | |
2940 | fo->copied = space; | |
783237e8 | 2941 | |
e64a8519 | 2942 | tcp_connect_queue_skb(sk, syn_data); |
783237e8 | 2943 | |
e64a8519 | 2944 | err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); |
783237e8 | 2945 | |
e64a8519 ED |
2946 | /* Now full SYN+DATA was cloned and sent (or not), |
2947 | * remove the SYN from the original skb (syn_data) | |
2948 | * we keep in write queue in case of a retransmit, as we | |
2949 | * also have the SYN packet (with no data) in the same queue. | |
2950 | */ | |
2951 | TCP_SKB_CB(syn_data)->seq++; | |
2952 | TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; | |
2953 | if (!err) { | |
67da22d2 | 2954 | tp->syn_data = (fo->copied > 0); |
783237e8 YC |
2955 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); |
2956 | goto done; | |
2957 | } | |
783237e8 YC |
2958 | |
2959 | fallback: | |
2960 | /* Send a regular SYN with Fast Open cookie request option */ | |
2961 | if (fo->cookie.len > 0) | |
2962 | fo->cookie.len = 0; | |
2963 | err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); | |
2964 | if (err) | |
2965 | tp->syn_fastopen = 0; | |
783237e8 YC |
2966 | done: |
2967 | fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ | |
2968 | return err; | |
2969 | } | |
2970 | ||
67edfef7 | 2971 | /* Build a SYN and send it off. */ |
1da177e4 LT |
2972 | int tcp_connect(struct sock *sk) |
2973 | { | |
2974 | struct tcp_sock *tp = tcp_sk(sk); | |
2975 | struct sk_buff *buff; | |
ee586811 | 2976 | int err; |
1da177e4 LT |
2977 | |
2978 | tcp_connect_init(sk); | |
2979 | ||
2b916477 AV |
2980 | if (unlikely(tp->repair)) { |
2981 | tcp_finish_connect(sk, NULL); | |
2982 | return 0; | |
2983 | } | |
2984 | ||
e64a8519 ED |
2985 | buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation); |
2986 | if (unlikely(!buff)) | |
1da177e4 LT |
2987 | return -ENOBUFS; |
2988 | ||
a3433f35 | 2989 | tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); |
783237e8 YC |
2990 | tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp; |
2991 | tcp_connect_queue_skb(sk, buff); | |
e870a8ef | 2992 | TCP_ECN_send_syn(sk, buff); |
1da177e4 | 2993 | |
783237e8 YC |
2994 | /* Send off SYN; include data in Fast Open. */ |
2995 | err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : | |
2996 | tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); | |
ee586811 EP |
2997 | if (err == -ECONNREFUSED) |
2998 | return err; | |
bd37a088 WY |
2999 | |
3000 | /* We change tp->snd_nxt after the tcp_transmit_skb() call | |
3001 | * in order to make this packet get counted in tcpOutSegs. | |
3002 | */ | |
3003 | tp->snd_nxt = tp->write_seq; | |
3004 | tp->pushed_seq = tp->write_seq; | |
81cc8a75 | 3005 | TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); |
1da177e4 LT |
3006 | |
3007 | /* Timer for repeating the SYN until an answer. */ | |
3f421baa ACM |
3008 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
3009 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
1da177e4 LT |
3010 | return 0; |
3011 | } | |
4bc2f18b | 3012 | EXPORT_SYMBOL(tcp_connect); |
1da177e4 LT |
3013 | |
3014 | /* Send out a delayed ack, the caller does the policy checking | |
3015 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() | |
3016 | * for details. | |
3017 | */ | |
3018 | void tcp_send_delayed_ack(struct sock *sk) | |
3019 | { | |
463c84b9 ACM |
3020 | struct inet_connection_sock *icsk = inet_csk(sk); |
3021 | int ato = icsk->icsk_ack.ato; | |
1da177e4 LT |
3022 | unsigned long timeout; |
3023 | ||
3024 | if (ato > TCP_DELACK_MIN) { | |
463c84b9 | 3025 | const struct tcp_sock *tp = tcp_sk(sk); |
056834d9 | 3026 | int max_ato = HZ / 2; |
1da177e4 | 3027 | |
056834d9 IJ |
3028 | if (icsk->icsk_ack.pingpong || |
3029 | (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) | |
1da177e4 LT |
3030 | max_ato = TCP_DELACK_MAX; |
3031 | ||
3032 | /* Slow path, intersegment interval is "high". */ | |
3033 | ||
3034 | /* If some rtt estimate is known, use it to bound delayed ack. | |
463c84b9 | 3035 | * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements |
1da177e4 LT |
3036 | * directly. |
3037 | */ | |
3038 | if (tp->srtt) { | |
056834d9 | 3039 | int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); |
1da177e4 LT |
3040 | |
3041 | if (rtt < max_ato) | |
3042 | max_ato = rtt; | |
3043 | } | |
3044 | ||
3045 | ato = min(ato, max_ato); | |
3046 | } | |
3047 | ||
3048 | /* Stay within the limit we were given */ | |
3049 | timeout = jiffies + ato; | |
3050 | ||
3051 | /* Use new timeout only if there wasn't a older one earlier. */ | |
463c84b9 | 3052 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { |
1da177e4 LT |
3053 | /* If delack timer was blocked or is about to expire, |
3054 | * send ACK now. | |
3055 | */ | |
463c84b9 ACM |
3056 | if (icsk->icsk_ack.blocked || |
3057 | time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { | |
1da177e4 LT |
3058 | tcp_send_ack(sk); |
3059 | return; | |
3060 | } | |
3061 | ||
463c84b9 ACM |
3062 | if (!time_before(timeout, icsk->icsk_ack.timeout)) |
3063 | timeout = icsk->icsk_ack.timeout; | |
1da177e4 | 3064 | } |
463c84b9 ACM |
3065 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; |
3066 | icsk->icsk_ack.timeout = timeout; | |
3067 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); | |
1da177e4 LT |
3068 | } |
3069 | ||
3070 | /* This routine sends an ack and also updates the window. */ | |
3071 | void tcp_send_ack(struct sock *sk) | |
3072 | { | |
058dc334 | 3073 | struct sk_buff *buff; |
1da177e4 | 3074 | |
058dc334 IJ |
3075 | /* If we have been reset, we may not send again. */ |
3076 | if (sk->sk_state == TCP_CLOSE) | |
3077 | return; | |
1da177e4 | 3078 | |
058dc334 IJ |
3079 | /* We are not putting this on the write queue, so |
3080 | * tcp_transmit_skb() will set the ownership to this | |
3081 | * sock. | |
3082 | */ | |
99a1dec7 | 3083 | buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); |
058dc334 IJ |
3084 | if (buff == NULL) { |
3085 | inet_csk_schedule_ack(sk); | |
3086 | inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; | |
3087 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, | |
3088 | TCP_DELACK_MAX, TCP_RTO_MAX); | |
3089 | return; | |
1da177e4 | 3090 | } |
058dc334 IJ |
3091 | |
3092 | /* Reserve space for headers and prepare control bits. */ | |
3093 | skb_reserve(buff, MAX_TCP_HEADER); | |
a3433f35 | 3094 | tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); |
058dc334 IJ |
3095 | |
3096 | /* Send it off, this clears delayed acks for us. */ | |
058dc334 | 3097 | TCP_SKB_CB(buff)->when = tcp_time_stamp; |
99a1dec7 | 3098 | tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC)); |
1da177e4 LT |
3099 | } |
3100 | ||
3101 | /* This routine sends a packet with an out of date sequence | |
3102 | * number. It assumes the other end will try to ack it. | |
3103 | * | |
3104 | * Question: what should we make while urgent mode? | |
3105 | * 4.4BSD forces sending single byte of data. We cannot send | |
3106 | * out of window data, because we have SND.NXT==SND.MAX... | |
3107 | * | |
3108 | * Current solution: to send TWO zero-length segments in urgent mode: | |
3109 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is | |
3110 | * out-of-date with SND.UNA-1 to probe window. | |
3111 | */ | |
3112 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent) | |
3113 | { | |
3114 | struct tcp_sock *tp = tcp_sk(sk); | |
3115 | struct sk_buff *skb; | |
3116 | ||
3117 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ | |
99a1dec7 | 3118 | skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); |
e905a9ed | 3119 | if (skb == NULL) |
1da177e4 LT |
3120 | return -1; |
3121 | ||
3122 | /* Reserve space for headers and set control bits. */ | |
3123 | skb_reserve(skb, MAX_TCP_HEADER); | |
1da177e4 LT |
3124 | /* Use a previous sequence. This should cause the other |
3125 | * end to send an ack. Don't queue or clone SKB, just | |
3126 | * send it. | |
3127 | */ | |
a3433f35 | 3128 | tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); |
1da177e4 | 3129 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
dfb4b9dc | 3130 | return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); |
1da177e4 LT |
3131 | } |
3132 | ||
ee995283 PE |
3133 | void tcp_send_window_probe(struct sock *sk) |
3134 | { | |
3135 | if (sk->sk_state == TCP_ESTABLISHED) { | |
3136 | tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; | |
3137 | tcp_xmit_probe_skb(sk, 0); | |
3138 | } | |
3139 | } | |
3140 | ||
67edfef7 | 3141 | /* Initiate keepalive or window probe from timer. */ |
1da177e4 LT |
3142 | int tcp_write_wakeup(struct sock *sk) |
3143 | { | |
058dc334 IJ |
3144 | struct tcp_sock *tp = tcp_sk(sk); |
3145 | struct sk_buff *skb; | |
1da177e4 | 3146 | |
058dc334 IJ |
3147 | if (sk->sk_state == TCP_CLOSE) |
3148 | return -1; | |
3149 | ||
3150 | if ((skb = tcp_send_head(sk)) != NULL && | |
3151 | before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { | |
3152 | int err; | |
0c54b85f | 3153 | unsigned int mss = tcp_current_mss(sk); |
058dc334 IJ |
3154 | unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
3155 | ||
3156 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) | |
3157 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; | |
3158 | ||
3159 | /* We are probing the opening of a window | |
3160 | * but the window size is != 0 | |
3161 | * must have been a result SWS avoidance ( sender ) | |
3162 | */ | |
3163 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || | |
3164 | skb->len > mss) { | |
3165 | seg_size = min(seg_size, mss); | |
4de075e0 | 3166 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
058dc334 IJ |
3167 | if (tcp_fragment(sk, skb, seg_size, mss)) |
3168 | return -1; | |
3169 | } else if (!tcp_skb_pcount(skb)) | |
3170 | tcp_set_skb_tso_segs(sk, skb, mss); | |
3171 | ||
4de075e0 | 3172 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
058dc334 IJ |
3173 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
3174 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); | |
3175 | if (!err) | |
3176 | tcp_event_new_data_sent(sk, skb); | |
3177 | return err; | |
3178 | } else { | |
33f5f57e | 3179 | if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) |
058dc334 IJ |
3180 | tcp_xmit_probe_skb(sk, 1); |
3181 | return tcp_xmit_probe_skb(sk, 0); | |
1da177e4 | 3182 | } |
1da177e4 LT |
3183 | } |
3184 | ||
3185 | /* A window probe timeout has occurred. If window is not closed send | |
3186 | * a partial packet else a zero probe. | |
3187 | */ | |
3188 | void tcp_send_probe0(struct sock *sk) | |
3189 | { | |
463c84b9 | 3190 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
3191 | struct tcp_sock *tp = tcp_sk(sk); |
3192 | int err; | |
3193 | ||
3194 | err = tcp_write_wakeup(sk); | |
3195 | ||
fe067e8a | 3196 | if (tp->packets_out || !tcp_send_head(sk)) { |
1da177e4 | 3197 | /* Cancel probe timer, if it is not required. */ |
6687e988 | 3198 | icsk->icsk_probes_out = 0; |
463c84b9 | 3199 | icsk->icsk_backoff = 0; |
1da177e4 LT |
3200 | return; |
3201 | } | |
3202 | ||
3203 | if (err <= 0) { | |
463c84b9 ACM |
3204 | if (icsk->icsk_backoff < sysctl_tcp_retries2) |
3205 | icsk->icsk_backoff++; | |
6687e988 | 3206 | icsk->icsk_probes_out++; |
e905a9ed | 3207 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
3f421baa ACM |
3208 | min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), |
3209 | TCP_RTO_MAX); | |
1da177e4 LT |
3210 | } else { |
3211 | /* If packet was not sent due to local congestion, | |
6687e988 | 3212 | * do not backoff and do not remember icsk_probes_out. |
1da177e4 LT |
3213 | * Let local senders to fight for local resources. |
3214 | * | |
3215 | * Use accumulated backoff yet. | |
3216 | */ | |
6687e988 ACM |
3217 | if (!icsk->icsk_probes_out) |
3218 | icsk->icsk_probes_out = 1; | |
e905a9ed | 3219 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
463c84b9 | 3220 | min(icsk->icsk_rto << icsk->icsk_backoff, |
3f421baa ACM |
3221 | TCP_RESOURCE_PROBE_INTERVAL), |
3222 | TCP_RTO_MAX); | |
1da177e4 LT |
3223 | } |
3224 | } |