]>
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 | |
c9bee3b7 ED |
68 | unsigned int sysctl_tcp_notsent_lowat __read_mostly = UINT_MAX; |
69 | EXPORT_SYMBOL(sysctl_tcp_notsent_lowat); | |
70 | ||
46d3ceab ED |
71 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
72 | int push_one, gfp_t gfp); | |
519855c5 | 73 | |
67edfef7 | 74 | /* Account for new data that has been sent to the network. */ |
cf533ea5 | 75 | static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 76 | { |
6ba8a3b1 | 77 | struct inet_connection_sock *icsk = inet_csk(sk); |
9e412ba7 | 78 | struct tcp_sock *tp = tcp_sk(sk); |
66f5fe62 | 79 | unsigned int prior_packets = tp->packets_out; |
9e412ba7 | 80 | |
fe067e8a | 81 | tcp_advance_send_head(sk, skb); |
1da177e4 | 82 | tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; |
8512430e | 83 | |
66f5fe62 | 84 | tp->packets_out += tcp_skb_pcount(skb); |
6ba8a3b1 | 85 | if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || |
6a5dc9e5 | 86 | icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { |
750ea2ba | 87 | tcp_rearm_rto(sk); |
6a5dc9e5 | 88 | } |
f19c29e3 | 89 | |
f7324acd DM |
90 | NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, |
91 | tcp_skb_pcount(skb)); | |
1da177e4 LT |
92 | } |
93 | ||
94 | /* SND.NXT, if window was not shrunk. | |
95 | * If window has been shrunk, what should we make? It is not clear at all. | |
96 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( | |
97 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already | |
98 | * invalid. OK, let's make this for now: | |
99 | */ | |
cf533ea5 | 100 | static inline __u32 tcp_acceptable_seq(const struct sock *sk) |
1da177e4 | 101 | { |
cf533ea5 | 102 | const struct tcp_sock *tp = tcp_sk(sk); |
9e412ba7 | 103 | |
90840def | 104 | if (!before(tcp_wnd_end(tp), tp->snd_nxt)) |
1da177e4 LT |
105 | return tp->snd_nxt; |
106 | else | |
90840def | 107 | return tcp_wnd_end(tp); |
1da177e4 LT |
108 | } |
109 | ||
110 | /* Calculate mss to advertise in SYN segment. | |
111 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: | |
112 | * | |
113 | * 1. It is independent of path mtu. | |
114 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. | |
115 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of | |
116 | * attached devices, because some buggy hosts are confused by | |
117 | * large MSS. | |
118 | * 4. We do not make 3, we advertise MSS, calculated from first | |
119 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. | |
120 | * This may be overridden via information stored in routing table. | |
121 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, | |
122 | * probably even Jumbo". | |
123 | */ | |
124 | static __u16 tcp_advertise_mss(struct sock *sk) | |
125 | { | |
126 | struct tcp_sock *tp = tcp_sk(sk); | |
cf533ea5 | 127 | const struct dst_entry *dst = __sk_dst_get(sk); |
1da177e4 LT |
128 | int mss = tp->advmss; |
129 | ||
0dbaee3b DM |
130 | if (dst) { |
131 | unsigned int metric = dst_metric_advmss(dst); | |
132 | ||
133 | if (metric < mss) { | |
134 | mss = metric; | |
135 | tp->advmss = mss; | |
136 | } | |
1da177e4 LT |
137 | } |
138 | ||
139 | return (__u16)mss; | |
140 | } | |
141 | ||
142 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". | |
143 | * This is the first part of cwnd validation mechanism. */ | |
cf533ea5 | 144 | static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst) |
1da177e4 | 145 | { |
463c84b9 | 146 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
147 | s32 delta = tcp_time_stamp - tp->lsndtime; |
148 | u32 restart_cwnd = tcp_init_cwnd(tp, dst); | |
149 | u32 cwnd = tp->snd_cwnd; | |
150 | ||
6687e988 | 151 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); |
1da177e4 | 152 | |
6687e988 | 153 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
1da177e4 LT |
154 | restart_cwnd = min(restart_cwnd, cwnd); |
155 | ||
463c84b9 | 156 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) |
1da177e4 LT |
157 | cwnd >>= 1; |
158 | tp->snd_cwnd = max(cwnd, restart_cwnd); | |
159 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
160 | tp->snd_cwnd_used = 0; | |
161 | } | |
162 | ||
67edfef7 | 163 | /* Congestion state accounting after a packet has been sent. */ |
40efc6fa | 164 | static void tcp_event_data_sent(struct tcp_sock *tp, |
cf533ea5 | 165 | struct sock *sk) |
1da177e4 | 166 | { |
463c84b9 ACM |
167 | struct inet_connection_sock *icsk = inet_csk(sk); |
168 | const u32 now = tcp_time_stamp; | |
bcefe17c | 169 | const struct dst_entry *dst = __sk_dst_get(sk); |
1da177e4 | 170 | |
35089bb2 DM |
171 | if (sysctl_tcp_slow_start_after_idle && |
172 | (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) | |
463c84b9 | 173 | tcp_cwnd_restart(sk, __sk_dst_get(sk)); |
1da177e4 LT |
174 | |
175 | tp->lsndtime = now; | |
176 | ||
177 | /* If it is a reply for ato after last received | |
178 | * packet, enter pingpong mode. | |
179 | */ | |
bcefe17c CW |
180 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato && |
181 | (!dst || !dst_metric(dst, RTAX_QUICKACK))) | |
182 | icsk->icsk_ack.pingpong = 1; | |
1da177e4 LT |
183 | } |
184 | ||
67edfef7 | 185 | /* Account for an ACK we sent. */ |
40efc6fa | 186 | static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) |
1da177e4 | 187 | { |
463c84b9 ACM |
188 | tcp_dec_quickack_mode(sk, pkts); |
189 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); | |
1da177e4 LT |
190 | } |
191 | ||
85f16525 YC |
192 | |
193 | u32 tcp_default_init_rwnd(u32 mss) | |
194 | { | |
195 | /* Initial receive window should be twice of TCP_INIT_CWND to | |
9ef71e0c | 196 | * enable proper sending of new unsent data during fast recovery |
85f16525 YC |
197 | * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a |
198 | * limit when mss is larger than 1460. | |
199 | */ | |
200 | u32 init_rwnd = TCP_INIT_CWND * 2; | |
201 | ||
202 | if (mss > 1460) | |
203 | init_rwnd = max((1460 * init_rwnd) / mss, 2U); | |
204 | return init_rwnd; | |
205 | } | |
206 | ||
1da177e4 LT |
207 | /* Determine a window scaling and initial window to offer. |
208 | * Based on the assumption that the given amount of space | |
209 | * will be offered. Store the results in the tp structure. | |
210 | * NOTE: for smooth operation initial space offering should | |
211 | * be a multiple of mss if possible. We assume here that mss >= 1. | |
212 | * This MUST be enforced by all callers. | |
213 | */ | |
214 | void tcp_select_initial_window(int __space, __u32 mss, | |
215 | __u32 *rcv_wnd, __u32 *window_clamp, | |
31d12926 | 216 | int wscale_ok, __u8 *rcv_wscale, |
217 | __u32 init_rcv_wnd) | |
1da177e4 LT |
218 | { |
219 | unsigned int space = (__space < 0 ? 0 : __space); | |
220 | ||
221 | /* If no clamp set the clamp to the max possible scaled window */ | |
222 | if (*window_clamp == 0) | |
223 | (*window_clamp) = (65535 << 14); | |
224 | space = min(*window_clamp, space); | |
225 | ||
226 | /* Quantize space offering to a multiple of mss if possible. */ | |
227 | if (space > mss) | |
228 | space = (space / mss) * mss; | |
229 | ||
230 | /* NOTE: offering an initial window larger than 32767 | |
15d99e02 RJ |
231 | * will break some buggy TCP stacks. If the admin tells us |
232 | * it is likely we could be speaking with such a buggy stack | |
233 | * we will truncate our initial window offering to 32K-1 | |
234 | * unless the remote has sent us a window scaling option, | |
235 | * which we interpret as a sign the remote TCP is not | |
236 | * misinterpreting the window field as a signed quantity. | |
1da177e4 | 237 | */ |
15d99e02 RJ |
238 | if (sysctl_tcp_workaround_signed_windows) |
239 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); | |
240 | else | |
241 | (*rcv_wnd) = space; | |
242 | ||
1da177e4 LT |
243 | (*rcv_wscale) = 0; |
244 | if (wscale_ok) { | |
245 | /* Set window scaling on max possible window | |
e905a9ed | 246 | * See RFC1323 for an explanation of the limit to 14 |
1da177e4 LT |
247 | */ |
248 | space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); | |
316c1592 | 249 | space = min_t(u32, space, *window_clamp); |
1da177e4 LT |
250 | while (space > 65535 && (*rcv_wscale) < 14) { |
251 | space >>= 1; | |
252 | (*rcv_wscale)++; | |
253 | } | |
254 | } | |
255 | ||
056834d9 | 256 | if (mss > (1 << *rcv_wscale)) { |
85f16525 YC |
257 | if (!init_rcv_wnd) /* Use default unless specified otherwise */ |
258 | init_rcv_wnd = tcp_default_init_rwnd(mss); | |
259 | *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); | |
1da177e4 LT |
260 | } |
261 | ||
262 | /* Set the clamp no higher than max representable value */ | |
263 | (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); | |
264 | } | |
4bc2f18b | 265 | EXPORT_SYMBOL(tcp_select_initial_window); |
1da177e4 LT |
266 | |
267 | /* Chose a new window to advertise, update state in tcp_sock for the | |
268 | * socket, and return result with RFC1323 scaling applied. The return | |
269 | * value can be stuffed directly into th->window for an outgoing | |
270 | * frame. | |
271 | */ | |
40efc6fa | 272 | static u16 tcp_select_window(struct sock *sk) |
1da177e4 LT |
273 | { |
274 | struct tcp_sock *tp = tcp_sk(sk); | |
8e165e20 | 275 | u32 old_win = tp->rcv_wnd; |
1da177e4 LT |
276 | u32 cur_win = tcp_receive_window(tp); |
277 | u32 new_win = __tcp_select_window(sk); | |
278 | ||
279 | /* Never shrink the offered window */ | |
2de979bd | 280 | if (new_win < cur_win) { |
1da177e4 LT |
281 | /* Danger Will Robinson! |
282 | * Don't update rcv_wup/rcv_wnd here or else | |
283 | * we will not be able to advertise a zero | |
284 | * window in time. --DaveM | |
285 | * | |
286 | * Relax Will Robinson. | |
287 | */ | |
8e165e20 FW |
288 | if (new_win == 0) |
289 | NET_INC_STATS(sock_net(sk), | |
290 | LINUX_MIB_TCPWANTZEROWINDOWADV); | |
607bfbf2 | 291 | new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); |
1da177e4 LT |
292 | } |
293 | tp->rcv_wnd = new_win; | |
294 | tp->rcv_wup = tp->rcv_nxt; | |
295 | ||
296 | /* Make sure we do not exceed the maximum possible | |
297 | * scaled window. | |
298 | */ | |
15d99e02 | 299 | if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) |
1da177e4 LT |
300 | new_win = min(new_win, MAX_TCP_WINDOW); |
301 | else | |
302 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); | |
303 | ||
304 | /* RFC1323 scaling applied */ | |
305 | new_win >>= tp->rx_opt.rcv_wscale; | |
306 | ||
307 | /* If we advertise zero window, disable fast path. */ | |
8e165e20 | 308 | if (new_win == 0) { |
1da177e4 | 309 | tp->pred_flags = 0; |
8e165e20 FW |
310 | if (old_win) |
311 | NET_INC_STATS(sock_net(sk), | |
312 | LINUX_MIB_TCPTOZEROWINDOWADV); | |
313 | } else if (old_win == 0) { | |
314 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); | |
315 | } | |
1da177e4 LT |
316 | |
317 | return new_win; | |
318 | } | |
319 | ||
67edfef7 | 320 | /* Packet ECN state for a SYN-ACK */ |
cf533ea5 | 321 | static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb) |
bdf1ee5d | 322 | { |
4de075e0 | 323 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; |
056834d9 | 324 | if (!(tp->ecn_flags & TCP_ECN_OK)) |
4de075e0 | 325 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; |
bdf1ee5d IJ |
326 | } |
327 | ||
67edfef7 | 328 | /* Packet ECN state for a SYN. */ |
bdf1ee5d IJ |
329 | static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) |
330 | { | |
331 | struct tcp_sock *tp = tcp_sk(sk); | |
332 | ||
333 | tp->ecn_flags = 0; | |
5d134f1c | 334 | if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) { |
4de075e0 | 335 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; |
bdf1ee5d IJ |
336 | tp->ecn_flags = TCP_ECN_OK; |
337 | } | |
338 | } | |
339 | ||
340 | static __inline__ void | |
cf533ea5 | 341 | TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th) |
bdf1ee5d IJ |
342 | { |
343 | if (inet_rsk(req)->ecn_ok) | |
344 | th->ece = 1; | |
345 | } | |
346 | ||
67edfef7 AK |
347 | /* Set up ECN state for a packet on a ESTABLISHED socket that is about to |
348 | * be sent. | |
349 | */ | |
bdf1ee5d IJ |
350 | static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, |
351 | int tcp_header_len) | |
352 | { | |
353 | struct tcp_sock *tp = tcp_sk(sk); | |
354 | ||
355 | if (tp->ecn_flags & TCP_ECN_OK) { | |
356 | /* Not-retransmitted data segment: set ECT and inject CWR. */ | |
357 | if (skb->len != tcp_header_len && | |
358 | !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { | |
359 | INET_ECN_xmit(sk); | |
056834d9 | 360 | if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { |
bdf1ee5d IJ |
361 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; |
362 | tcp_hdr(skb)->cwr = 1; | |
363 | skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; | |
364 | } | |
365 | } else { | |
366 | /* ACK or retransmitted segment: clear ECT|CE */ | |
367 | INET_ECN_dontxmit(sk); | |
368 | } | |
369 | if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) | |
370 | tcp_hdr(skb)->ece = 1; | |
371 | } | |
372 | } | |
373 | ||
e870a8ef IJ |
374 | /* Constructs common control bits of non-data skb. If SYN/FIN is present, |
375 | * auto increment end seqno. | |
376 | */ | |
377 | static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) | |
378 | { | |
7b7fc97a ED |
379 | struct skb_shared_info *shinfo = skb_shinfo(skb); |
380 | ||
2e8e18ef | 381 | skb->ip_summed = CHECKSUM_PARTIAL; |
e870a8ef IJ |
382 | skb->csum = 0; |
383 | ||
4de075e0 | 384 | TCP_SKB_CB(skb)->tcp_flags = flags; |
e870a8ef IJ |
385 | TCP_SKB_CB(skb)->sacked = 0; |
386 | ||
7b7fc97a ED |
387 | shinfo->gso_segs = 1; |
388 | shinfo->gso_size = 0; | |
389 | shinfo->gso_type = 0; | |
e870a8ef IJ |
390 | |
391 | TCP_SKB_CB(skb)->seq = seq; | |
a3433f35 | 392 | if (flags & (TCPHDR_SYN | TCPHDR_FIN)) |
e870a8ef IJ |
393 | seq++; |
394 | TCP_SKB_CB(skb)->end_seq = seq; | |
395 | } | |
396 | ||
a2a385d6 | 397 | static inline bool tcp_urg_mode(const struct tcp_sock *tp) |
33f5f57e IJ |
398 | { |
399 | return tp->snd_una != tp->snd_up; | |
400 | } | |
401 | ||
33ad798c AL |
402 | #define OPTION_SACK_ADVERTISE (1 << 0) |
403 | #define OPTION_TS (1 << 1) | |
404 | #define OPTION_MD5 (1 << 2) | |
89e95a61 | 405 | #define OPTION_WSCALE (1 << 3) |
2100c8d2 | 406 | #define OPTION_FAST_OPEN_COOKIE (1 << 8) |
33ad798c AL |
407 | |
408 | struct tcp_out_options { | |
2100c8d2 YC |
409 | u16 options; /* bit field of OPTION_* */ |
410 | u16 mss; /* 0 to disable */ | |
33ad798c AL |
411 | u8 ws; /* window scale, 0 to disable */ |
412 | u8 num_sack_blocks; /* number of SACK blocks to include */ | |
bd0388ae | 413 | u8 hash_size; /* bytes in hash_location */ |
bd0388ae | 414 | __u8 *hash_location; /* temporary pointer, overloaded */ |
2100c8d2 YC |
415 | __u32 tsval, tsecr; /* need to include OPTION_TS */ |
416 | struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ | |
33ad798c AL |
417 | }; |
418 | ||
67edfef7 AK |
419 | /* Write previously computed TCP options to the packet. |
420 | * | |
421 | * Beware: Something in the Internet is very sensitive to the ordering of | |
fd6149d3 IJ |
422 | * TCP options, we learned this through the hard way, so be careful here. |
423 | * Luckily we can at least blame others for their non-compliance but from | |
8e3bff96 | 424 | * inter-operability perspective it seems that we're somewhat stuck with |
fd6149d3 IJ |
425 | * the ordering which we have been using if we want to keep working with |
426 | * those broken things (not that it currently hurts anybody as there isn't | |
427 | * particular reason why the ordering would need to be changed). | |
428 | * | |
429 | * At least SACK_PERM as the first option is known to lead to a disaster | |
430 | * (but it may well be that other scenarios fail similarly). | |
431 | */ | |
33ad798c | 432 | static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, |
bd0388ae WAS |
433 | struct tcp_out_options *opts) |
434 | { | |
2100c8d2 | 435 | u16 options = opts->options; /* mungable copy */ |
bd0388ae | 436 | |
bd0388ae | 437 | if (unlikely(OPTION_MD5 & options)) { |
1a2c6181 CP |
438 | *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | |
439 | (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); | |
bd0388ae WAS |
440 | /* overload cookie hash location */ |
441 | opts->hash_location = (__u8 *)ptr; | |
33ad798c | 442 | ptr += 4; |
40efc6fa | 443 | } |
33ad798c | 444 | |
fd6149d3 IJ |
445 | if (unlikely(opts->mss)) { |
446 | *ptr++ = htonl((TCPOPT_MSS << 24) | | |
447 | (TCPOLEN_MSS << 16) | | |
448 | opts->mss); | |
449 | } | |
450 | ||
bd0388ae WAS |
451 | if (likely(OPTION_TS & options)) { |
452 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { | |
33ad798c AL |
453 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
454 | (TCPOLEN_SACK_PERM << 16) | | |
455 | (TCPOPT_TIMESTAMP << 8) | | |
456 | TCPOLEN_TIMESTAMP); | |
bd0388ae | 457 | options &= ~OPTION_SACK_ADVERTISE; |
33ad798c AL |
458 | } else { |
459 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
460 | (TCPOPT_NOP << 16) | | |
461 | (TCPOPT_TIMESTAMP << 8) | | |
462 | TCPOLEN_TIMESTAMP); | |
463 | } | |
464 | *ptr++ = htonl(opts->tsval); | |
465 | *ptr++ = htonl(opts->tsecr); | |
466 | } | |
467 | ||
bd0388ae | 468 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { |
33ad798c AL |
469 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
470 | (TCPOPT_NOP << 16) | | |
471 | (TCPOPT_SACK_PERM << 8) | | |
472 | TCPOLEN_SACK_PERM); | |
473 | } | |
474 | ||
bd0388ae | 475 | if (unlikely(OPTION_WSCALE & options)) { |
33ad798c AL |
476 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
477 | (TCPOPT_WINDOW << 16) | | |
478 | (TCPOLEN_WINDOW << 8) | | |
479 | opts->ws); | |
480 | } | |
481 | ||
482 | if (unlikely(opts->num_sack_blocks)) { | |
483 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? | |
484 | tp->duplicate_sack : tp->selective_acks; | |
40efc6fa SH |
485 | int this_sack; |
486 | ||
487 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
488 | (TCPOPT_NOP << 16) | | |
489 | (TCPOPT_SACK << 8) | | |
33ad798c | 490 | (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * |
40efc6fa | 491 | TCPOLEN_SACK_PERBLOCK))); |
2de979bd | 492 | |
33ad798c AL |
493 | for (this_sack = 0; this_sack < opts->num_sack_blocks; |
494 | ++this_sack) { | |
40efc6fa SH |
495 | *ptr++ = htonl(sp[this_sack].start_seq); |
496 | *ptr++ = htonl(sp[this_sack].end_seq); | |
497 | } | |
2de979bd | 498 | |
5861f8e5 | 499 | tp->rx_opt.dsack = 0; |
40efc6fa | 500 | } |
2100c8d2 YC |
501 | |
502 | if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { | |
503 | struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; | |
504 | ||
505 | *ptr++ = htonl((TCPOPT_EXP << 24) | | |
506 | ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) | | |
507 | TCPOPT_FASTOPEN_MAGIC); | |
508 | ||
509 | memcpy(ptr, foc->val, foc->len); | |
510 | if ((foc->len & 3) == 2) { | |
511 | u8 *align = ((u8 *)ptr) + foc->len; | |
512 | align[0] = align[1] = TCPOPT_NOP; | |
513 | } | |
514 | ptr += (foc->len + 3) >> 2; | |
515 | } | |
33ad798c AL |
516 | } |
517 | ||
67edfef7 AK |
518 | /* Compute TCP options for SYN packets. This is not the final |
519 | * network wire format yet. | |
520 | */ | |
95c96174 | 521 | static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, |
33ad798c | 522 | struct tcp_out_options *opts, |
cf533ea5 ED |
523 | struct tcp_md5sig_key **md5) |
524 | { | |
33ad798c | 525 | struct tcp_sock *tp = tcp_sk(sk); |
95c96174 | 526 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
783237e8 | 527 | struct tcp_fastopen_request *fastopen = tp->fastopen_req; |
33ad798c | 528 | |
cfb6eeb4 | 529 | #ifdef CONFIG_TCP_MD5SIG |
33ad798c AL |
530 | *md5 = tp->af_specific->md5_lookup(sk, sk); |
531 | if (*md5) { | |
532 | opts->options |= OPTION_MD5; | |
bd0388ae | 533 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
cfb6eeb4 | 534 | } |
33ad798c AL |
535 | #else |
536 | *md5 = NULL; | |
cfb6eeb4 | 537 | #endif |
33ad798c AL |
538 | |
539 | /* We always get an MSS option. The option bytes which will be seen in | |
540 | * normal data packets should timestamps be used, must be in the MSS | |
541 | * advertised. But we subtract them from tp->mss_cache so that | |
542 | * calculations in tcp_sendmsg are simpler etc. So account for this | |
543 | * fact here if necessary. If we don't do this correctly, as a | |
544 | * receiver we won't recognize data packets as being full sized when we | |
545 | * should, and thus we won't abide by the delayed ACK rules correctly. | |
546 | * SACKs don't matter, we never delay an ACK when we have any of those | |
547 | * going out. */ | |
548 | opts->mss = tcp_advertise_mss(sk); | |
bd0388ae | 549 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c | 550 | |
bb5b7c11 | 551 | if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { |
33ad798c | 552 | opts->options |= OPTION_TS; |
ee684b6f | 553 | opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset; |
33ad798c | 554 | opts->tsecr = tp->rx_opt.ts_recent; |
bd0388ae | 555 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c | 556 | } |
bb5b7c11 | 557 | if (likely(sysctl_tcp_window_scaling)) { |
33ad798c | 558 | opts->ws = tp->rx_opt.rcv_wscale; |
89e95a61 | 559 | opts->options |= OPTION_WSCALE; |
bd0388ae | 560 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 561 | } |
bb5b7c11 | 562 | if (likely(sysctl_tcp_sack)) { |
33ad798c | 563 | opts->options |= OPTION_SACK_ADVERTISE; |
b32d1310 | 564 | if (unlikely(!(OPTION_TS & opts->options))) |
bd0388ae | 565 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c AL |
566 | } |
567 | ||
783237e8 YC |
568 | if (fastopen && fastopen->cookie.len >= 0) { |
569 | u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len; | |
570 | need = (need + 3) & ~3U; /* Align to 32 bits */ | |
571 | if (remaining >= need) { | |
572 | opts->options |= OPTION_FAST_OPEN_COOKIE; | |
573 | opts->fastopen_cookie = &fastopen->cookie; | |
574 | remaining -= need; | |
575 | tp->syn_fastopen = 1; | |
576 | } | |
577 | } | |
bd0388ae | 578 | |
bd0388ae | 579 | return MAX_TCP_OPTION_SPACE - remaining; |
40efc6fa SH |
580 | } |
581 | ||
67edfef7 | 582 | /* Set up TCP options for SYN-ACKs. */ |
95c96174 | 583 | static unsigned int tcp_synack_options(struct sock *sk, |
33ad798c | 584 | struct request_sock *req, |
95c96174 | 585 | unsigned int mss, struct sk_buff *skb, |
33ad798c | 586 | struct tcp_out_options *opts, |
4957faad | 587 | struct tcp_md5sig_key **md5, |
8336886f | 588 | struct tcp_fastopen_cookie *foc) |
4957faad | 589 | { |
33ad798c | 590 | struct inet_request_sock *ireq = inet_rsk(req); |
95c96174 | 591 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
33ad798c | 592 | |
cfb6eeb4 | 593 | #ifdef CONFIG_TCP_MD5SIG |
33ad798c AL |
594 | *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); |
595 | if (*md5) { | |
596 | opts->options |= OPTION_MD5; | |
4957faad WAS |
597 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
598 | ||
599 | /* We can't fit any SACK blocks in a packet with MD5 + TS | |
600 | * options. There was discussion about disabling SACK | |
601 | * rather than TS in order to fit in better with old, | |
602 | * buggy kernels, but that was deemed to be unnecessary. | |
603 | */ | |
de213e5e | 604 | ireq->tstamp_ok &= !ireq->sack_ok; |
cfb6eeb4 | 605 | } |
33ad798c AL |
606 | #else |
607 | *md5 = NULL; | |
cfb6eeb4 | 608 | #endif |
33ad798c | 609 | |
4957faad | 610 | /* We always send an MSS option. */ |
33ad798c | 611 | opts->mss = mss; |
4957faad | 612 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c AL |
613 | |
614 | if (likely(ireq->wscale_ok)) { | |
615 | opts->ws = ireq->rcv_wscale; | |
89e95a61 | 616 | opts->options |= OPTION_WSCALE; |
4957faad | 617 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 618 | } |
de213e5e | 619 | if (likely(ireq->tstamp_ok)) { |
33ad798c AL |
620 | opts->options |= OPTION_TS; |
621 | opts->tsval = TCP_SKB_CB(skb)->when; | |
622 | opts->tsecr = req->ts_recent; | |
4957faad | 623 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c AL |
624 | } |
625 | if (likely(ireq->sack_ok)) { | |
626 | opts->options |= OPTION_SACK_ADVERTISE; | |
de213e5e | 627 | if (unlikely(!ireq->tstamp_ok)) |
4957faad | 628 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c | 629 | } |
8336886f JC |
630 | if (foc != NULL) { |
631 | u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; | |
632 | need = (need + 3) & ~3U; /* Align to 32 bits */ | |
633 | if (remaining >= need) { | |
634 | opts->options |= OPTION_FAST_OPEN_COOKIE; | |
635 | opts->fastopen_cookie = foc; | |
636 | remaining -= need; | |
637 | } | |
638 | } | |
1a2c6181 | 639 | |
4957faad | 640 | return MAX_TCP_OPTION_SPACE - remaining; |
33ad798c AL |
641 | } |
642 | ||
67edfef7 AK |
643 | /* Compute TCP options for ESTABLISHED sockets. This is not the |
644 | * final wire format yet. | |
645 | */ | |
95c96174 | 646 | static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, |
33ad798c | 647 | struct tcp_out_options *opts, |
cf533ea5 ED |
648 | struct tcp_md5sig_key **md5) |
649 | { | |
33ad798c AL |
650 | struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; |
651 | struct tcp_sock *tp = tcp_sk(sk); | |
95c96174 | 652 | unsigned int size = 0; |
cabeccbd | 653 | unsigned int eff_sacks; |
33ad798c | 654 | |
5843ef42 AK |
655 | opts->options = 0; |
656 | ||
33ad798c AL |
657 | #ifdef CONFIG_TCP_MD5SIG |
658 | *md5 = tp->af_specific->md5_lookup(sk, sk); | |
659 | if (unlikely(*md5)) { | |
660 | opts->options |= OPTION_MD5; | |
661 | size += TCPOLEN_MD5SIG_ALIGNED; | |
662 | } | |
663 | #else | |
664 | *md5 = NULL; | |
665 | #endif | |
666 | ||
667 | if (likely(tp->rx_opt.tstamp_ok)) { | |
668 | opts->options |= OPTION_TS; | |
ee684b6f | 669 | opts->tsval = tcb ? tcb->when + tp->tsoffset : 0; |
33ad798c AL |
670 | opts->tsecr = tp->rx_opt.ts_recent; |
671 | size += TCPOLEN_TSTAMP_ALIGNED; | |
672 | } | |
673 | ||
cabeccbd IJ |
674 | eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; |
675 | if (unlikely(eff_sacks)) { | |
95c96174 | 676 | const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; |
33ad798c | 677 | opts->num_sack_blocks = |
95c96174 | 678 | min_t(unsigned int, eff_sacks, |
33ad798c AL |
679 | (remaining - TCPOLEN_SACK_BASE_ALIGNED) / |
680 | TCPOLEN_SACK_PERBLOCK); | |
681 | size += TCPOLEN_SACK_BASE_ALIGNED + | |
682 | opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; | |
683 | } | |
684 | ||
685 | return size; | |
40efc6fa | 686 | } |
1da177e4 | 687 | |
46d3ceab ED |
688 | |
689 | /* TCP SMALL QUEUES (TSQ) | |
690 | * | |
691 | * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) | |
692 | * to reduce RTT and bufferbloat. | |
693 | * We do this using a special skb destructor (tcp_wfree). | |
694 | * | |
695 | * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb | |
696 | * needs to be reallocated in a driver. | |
8e3bff96 | 697 | * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc |
46d3ceab ED |
698 | * |
699 | * Since transmit from skb destructor is forbidden, we use a tasklet | |
700 | * to process all sockets that eventually need to send more skbs. | |
701 | * We use one tasklet per cpu, with its own queue of sockets. | |
702 | */ | |
703 | struct tsq_tasklet { | |
704 | struct tasklet_struct tasklet; | |
705 | struct list_head head; /* queue of tcp sockets */ | |
706 | }; | |
707 | static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); | |
708 | ||
6f458dfb ED |
709 | static void tcp_tsq_handler(struct sock *sk) |
710 | { | |
711 | if ((1 << sk->sk_state) & | |
712 | (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | | |
713 | TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) | |
bf06200e JO |
714 | tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle, |
715 | 0, GFP_ATOMIC); | |
6f458dfb | 716 | } |
46d3ceab | 717 | /* |
8e3bff96 | 718 | * One tasklet per cpu tries to send more skbs. |
46d3ceab | 719 | * We run in tasklet context but need to disable irqs when |
8e3bff96 | 720 | * transferring tsq->head because tcp_wfree() might |
46d3ceab ED |
721 | * interrupt us (non NAPI drivers) |
722 | */ | |
723 | static void tcp_tasklet_func(unsigned long data) | |
724 | { | |
725 | struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; | |
726 | LIST_HEAD(list); | |
727 | unsigned long flags; | |
728 | struct list_head *q, *n; | |
729 | struct tcp_sock *tp; | |
730 | struct sock *sk; | |
731 | ||
732 | local_irq_save(flags); | |
733 | list_splice_init(&tsq->head, &list); | |
734 | local_irq_restore(flags); | |
735 | ||
736 | list_for_each_safe(q, n, &list) { | |
737 | tp = list_entry(q, struct tcp_sock, tsq_node); | |
738 | list_del(&tp->tsq_node); | |
739 | ||
740 | sk = (struct sock *)tp; | |
741 | bh_lock_sock(sk); | |
742 | ||
743 | if (!sock_owned_by_user(sk)) { | |
6f458dfb | 744 | tcp_tsq_handler(sk); |
46d3ceab ED |
745 | } else { |
746 | /* defer the work to tcp_release_cb() */ | |
6f458dfb | 747 | set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); |
46d3ceab ED |
748 | } |
749 | bh_unlock_sock(sk); | |
750 | ||
751 | clear_bit(TSQ_QUEUED, &tp->tsq_flags); | |
752 | sk_free(sk); | |
753 | } | |
754 | } | |
755 | ||
6f458dfb ED |
756 | #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ |
757 | (1UL << TCP_WRITE_TIMER_DEFERRED) | \ | |
563d34d0 ED |
758 | (1UL << TCP_DELACK_TIMER_DEFERRED) | \ |
759 | (1UL << TCP_MTU_REDUCED_DEFERRED)) | |
46d3ceab ED |
760 | /** |
761 | * tcp_release_cb - tcp release_sock() callback | |
762 | * @sk: socket | |
763 | * | |
764 | * called from release_sock() to perform protocol dependent | |
765 | * actions before socket release. | |
766 | */ | |
767 | void tcp_release_cb(struct sock *sk) | |
768 | { | |
769 | struct tcp_sock *tp = tcp_sk(sk); | |
6f458dfb | 770 | unsigned long flags, nflags; |
46d3ceab | 771 | |
6f458dfb ED |
772 | /* perform an atomic operation only if at least one flag is set */ |
773 | do { | |
774 | flags = tp->tsq_flags; | |
775 | if (!(flags & TCP_DEFERRED_ALL)) | |
776 | return; | |
777 | nflags = flags & ~TCP_DEFERRED_ALL; | |
778 | } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); | |
779 | ||
780 | if (flags & (1UL << TCP_TSQ_DEFERRED)) | |
781 | tcp_tsq_handler(sk); | |
782 | ||
c3f9b018 ED |
783 | /* Here begins the tricky part : |
784 | * We are called from release_sock() with : | |
785 | * 1) BH disabled | |
786 | * 2) sk_lock.slock spinlock held | |
787 | * 3) socket owned by us (sk->sk_lock.owned == 1) | |
788 | * | |
789 | * But following code is meant to be called from BH handlers, | |
790 | * so we should keep BH disabled, but early release socket ownership | |
791 | */ | |
792 | sock_release_ownership(sk); | |
793 | ||
144d56e9 | 794 | if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { |
6f458dfb | 795 | tcp_write_timer_handler(sk); |
144d56e9 ED |
796 | __sock_put(sk); |
797 | } | |
798 | if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { | |
6f458dfb | 799 | tcp_delack_timer_handler(sk); |
144d56e9 ED |
800 | __sock_put(sk); |
801 | } | |
802 | if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { | |
563d34d0 | 803 | sk->sk_prot->mtu_reduced(sk); |
144d56e9 ED |
804 | __sock_put(sk); |
805 | } | |
46d3ceab ED |
806 | } |
807 | EXPORT_SYMBOL(tcp_release_cb); | |
808 | ||
809 | void __init tcp_tasklet_init(void) | |
810 | { | |
811 | int i; | |
812 | ||
813 | for_each_possible_cpu(i) { | |
814 | struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); | |
815 | ||
816 | INIT_LIST_HEAD(&tsq->head); | |
817 | tasklet_init(&tsq->tasklet, | |
818 | tcp_tasklet_func, | |
819 | (unsigned long)tsq); | |
820 | } | |
821 | } | |
822 | ||
823 | /* | |
824 | * Write buffer destructor automatically called from kfree_skb. | |
8e3bff96 | 825 | * We can't xmit new skbs from this context, as we might already |
46d3ceab ED |
826 | * hold qdisc lock. |
827 | */ | |
d6a4a104 | 828 | void tcp_wfree(struct sk_buff *skb) |
46d3ceab ED |
829 | { |
830 | struct sock *sk = skb->sk; | |
831 | struct tcp_sock *tp = tcp_sk(sk); | |
832 | ||
833 | if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && | |
834 | !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { | |
835 | unsigned long flags; | |
836 | struct tsq_tasklet *tsq; | |
837 | ||
838 | /* Keep a ref on socket. | |
839 | * This last ref will be released in tcp_tasklet_func() | |
840 | */ | |
841 | atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc); | |
842 | ||
843 | /* queue this socket to tasklet queue */ | |
844 | local_irq_save(flags); | |
845 | tsq = &__get_cpu_var(tsq_tasklet); | |
846 | list_add(&tp->tsq_node, &tsq->head); | |
847 | tasklet_schedule(&tsq->tasklet); | |
848 | local_irq_restore(flags); | |
849 | } else { | |
850 | sock_wfree(skb); | |
851 | } | |
852 | } | |
853 | ||
1da177e4 LT |
854 | /* This routine actually transmits TCP packets queued in by |
855 | * tcp_do_sendmsg(). This is used by both the initial | |
856 | * transmission and possible later retransmissions. | |
857 | * All SKB's seen here are completely headerless. It is our | |
858 | * job to build the TCP header, and pass the packet down to | |
859 | * IP so it can do the same plus pass the packet off to the | |
860 | * device. | |
861 | * | |
862 | * We are working here with either a clone of the original | |
863 | * SKB, or a fresh unique copy made by the retransmit engine. | |
864 | */ | |
056834d9 IJ |
865 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, |
866 | gfp_t gfp_mask) | |
1da177e4 | 867 | { |
dfb4b9dc DM |
868 | const struct inet_connection_sock *icsk = inet_csk(sk); |
869 | struct inet_sock *inet; | |
870 | struct tcp_sock *tp; | |
871 | struct tcp_skb_cb *tcb; | |
33ad798c | 872 | struct tcp_out_options opts; |
95c96174 | 873 | unsigned int tcp_options_size, tcp_header_size; |
cfb6eeb4 | 874 | struct tcp_md5sig_key *md5; |
dfb4b9dc | 875 | struct tcphdr *th; |
dfb4b9dc DM |
876 | int err; |
877 | ||
878 | BUG_ON(!skb || !tcp_skb_pcount(skb)); | |
879 | ||
ccdbb6e9 | 880 | if (clone_it) { |
740b0f18 | 881 | skb_mstamp_get(&skb->skb_mstamp); |
ccdbb6e9 | 882 | |
dfb4b9dc DM |
883 | if (unlikely(skb_cloned(skb))) |
884 | skb = pskb_copy(skb, gfp_mask); | |
885 | else | |
886 | skb = skb_clone(skb, gfp_mask); | |
887 | if (unlikely(!skb)) | |
888 | return -ENOBUFS; | |
21962692 ED |
889 | /* Our usage of tstamp should remain private */ |
890 | skb->tstamp.tv64 = 0; | |
dfb4b9dc | 891 | } |
1da177e4 | 892 | |
dfb4b9dc DM |
893 | inet = inet_sk(sk); |
894 | tp = tcp_sk(sk); | |
895 | tcb = TCP_SKB_CB(skb); | |
33ad798c | 896 | memset(&opts, 0, sizeof(opts)); |
1da177e4 | 897 | |
4de075e0 | 898 | if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) |
33ad798c AL |
899 | tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); |
900 | else | |
901 | tcp_options_size = tcp_established_options(sk, skb, &opts, | |
902 | &md5); | |
903 | tcp_header_size = tcp_options_size + sizeof(struct tcphdr); | |
e905a9ed | 904 | |
547669d4 | 905 | if (tcp_packets_in_flight(tp) == 0) |
dfb4b9dc | 906 | tcp_ca_event(sk, CA_EVENT_TX_START); |
547669d4 ED |
907 | |
908 | /* if no packet is in qdisc/device queue, then allow XPS to select | |
909 | * another queue. | |
910 | */ | |
911 | skb->ooo_okay = sk_wmem_alloc_get(sk) == 0; | |
dfb4b9dc | 912 | |
aa8223c7 ACM |
913 | skb_push(skb, tcp_header_size); |
914 | skb_reset_transport_header(skb); | |
46d3ceab ED |
915 | |
916 | skb_orphan(skb); | |
917 | skb->sk = sk; | |
c9eeec26 | 918 | skb->destructor = tcp_wfree; |
46d3ceab | 919 | atomic_add(skb->truesize, &sk->sk_wmem_alloc); |
dfb4b9dc DM |
920 | |
921 | /* Build TCP header and checksum it. */ | |
aa8223c7 | 922 | th = tcp_hdr(skb); |
c720c7e8 ED |
923 | th->source = inet->inet_sport; |
924 | th->dest = inet->inet_dport; | |
dfb4b9dc DM |
925 | th->seq = htonl(tcb->seq); |
926 | th->ack_seq = htonl(tp->rcv_nxt); | |
df7a3b07 | 927 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
4de075e0 | 928 | tcb->tcp_flags); |
dfb4b9dc | 929 | |
4de075e0 | 930 | if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { |
dfb4b9dc DM |
931 | /* RFC1323: The window in SYN & SYN/ACK segments |
932 | * is never scaled. | |
933 | */ | |
600ff0c2 | 934 | th->window = htons(min(tp->rcv_wnd, 65535U)); |
dfb4b9dc DM |
935 | } else { |
936 | th->window = htons(tcp_select_window(sk)); | |
937 | } | |
938 | th->check = 0; | |
939 | th->urg_ptr = 0; | |
1da177e4 | 940 | |
33f5f57e | 941 | /* The urg_mode check is necessary during a below snd_una win probe */ |
7691367d HX |
942 | if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { |
943 | if (before(tp->snd_up, tcb->seq + 0x10000)) { | |
944 | th->urg_ptr = htons(tp->snd_up - tcb->seq); | |
945 | th->urg = 1; | |
946 | } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { | |
0eae88f3 | 947 | th->urg_ptr = htons(0xFFFF); |
7691367d HX |
948 | th->urg = 1; |
949 | } | |
dfb4b9dc | 950 | } |
1da177e4 | 951 | |
bd0388ae | 952 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
4de075e0 | 953 | if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) |
9e412ba7 | 954 | TCP_ECN_send(sk, skb, tcp_header_size); |
1da177e4 | 955 | |
cfb6eeb4 YH |
956 | #ifdef CONFIG_TCP_MD5SIG |
957 | /* Calculate the MD5 hash, as we have all we need now */ | |
958 | if (md5) { | |
a465419b | 959 | sk_nocaps_add(sk, NETIF_F_GSO_MASK); |
bd0388ae | 960 | tp->af_specific->calc_md5_hash(opts.hash_location, |
49a72dfb | 961 | md5, sk, NULL, skb); |
cfb6eeb4 YH |
962 | } |
963 | #endif | |
964 | ||
bb296246 | 965 | icsk->icsk_af_ops->send_check(sk, skb); |
1da177e4 | 966 | |
4de075e0 | 967 | if (likely(tcb->tcp_flags & TCPHDR_ACK)) |
dfb4b9dc | 968 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); |
1da177e4 | 969 | |
dfb4b9dc | 970 | if (skb->len != tcp_header_size) |
cf533ea5 | 971 | tcp_event_data_sent(tp, sk); |
1da177e4 | 972 | |
bd37a088 | 973 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
aa2ea058 TH |
974 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, |
975 | tcp_skb_pcount(skb)); | |
1da177e4 | 976 | |
b0270e91 | 977 | err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); |
83de47cd | 978 | if (likely(err <= 0)) |
dfb4b9dc DM |
979 | return err; |
980 | ||
3cfe3baa | 981 | tcp_enter_cwr(sk, 1); |
dfb4b9dc | 982 | |
b9df3cb8 | 983 | return net_xmit_eval(err); |
1da177e4 LT |
984 | } |
985 | ||
67edfef7 | 986 | /* This routine just queues the buffer for sending. |
1da177e4 LT |
987 | * |
988 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, | |
989 | * otherwise socket can stall. | |
990 | */ | |
991 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) | |
992 | { | |
993 | struct tcp_sock *tp = tcp_sk(sk); | |
994 | ||
995 | /* Advance write_seq and place onto the write_queue. */ | |
996 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; | |
997 | skb_header_release(skb); | |
fe067e8a | 998 | tcp_add_write_queue_tail(sk, skb); |
3ab224be HA |
999 | sk->sk_wmem_queued += skb->truesize; |
1000 | sk_mem_charge(sk, skb->truesize); | |
1da177e4 LT |
1001 | } |
1002 | ||
67edfef7 | 1003 | /* Initialize TSO segments for a packet. */ |
cf533ea5 | 1004 | static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, |
056834d9 | 1005 | unsigned int mss_now) |
f6302d1d | 1006 | { |
7b7fc97a ED |
1007 | struct skb_shared_info *shinfo = skb_shinfo(skb); |
1008 | ||
c52e2421 ED |
1009 | /* Make sure we own this skb before messing gso_size/gso_segs */ |
1010 | WARN_ON_ONCE(skb_cloned(skb)); | |
1011 | ||
8f26fb1c | 1012 | if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { |
f6302d1d DM |
1013 | /* Avoid the costly divide in the normal |
1014 | * non-TSO case. | |
1015 | */ | |
7b7fc97a ED |
1016 | shinfo->gso_segs = 1; |
1017 | shinfo->gso_size = 0; | |
1018 | shinfo->gso_type = 0; | |
f6302d1d | 1019 | } else { |
7b7fc97a ED |
1020 | shinfo->gso_segs = DIV_ROUND_UP(skb->len, mss_now); |
1021 | shinfo->gso_size = mss_now; | |
1022 | shinfo->gso_type = sk->sk_gso_type; | |
1da177e4 LT |
1023 | } |
1024 | } | |
1025 | ||
91fed7a1 | 1026 | /* When a modification to fackets out becomes necessary, we need to check |
68f8353b | 1027 | * skb is counted to fackets_out or not. |
91fed7a1 | 1028 | */ |
cf533ea5 | 1029 | static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, |
91fed7a1 IJ |
1030 | int decr) |
1031 | { | |
a47e5a98 IJ |
1032 | struct tcp_sock *tp = tcp_sk(sk); |
1033 | ||
dc86967b | 1034 | if (!tp->sacked_out || tcp_is_reno(tp)) |
91fed7a1 IJ |
1035 | return; |
1036 | ||
6859d494 | 1037 | if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) |
91fed7a1 | 1038 | tp->fackets_out -= decr; |
91fed7a1 IJ |
1039 | } |
1040 | ||
797108d1 IJ |
1041 | /* Pcount in the middle of the write queue got changed, we need to do various |
1042 | * tweaks to fix counters | |
1043 | */ | |
cf533ea5 | 1044 | static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) |
797108d1 IJ |
1045 | { |
1046 | struct tcp_sock *tp = tcp_sk(sk); | |
1047 | ||
1048 | tp->packets_out -= decr; | |
1049 | ||
1050 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
1051 | tp->sacked_out -= decr; | |
1052 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
1053 | tp->retrans_out -= decr; | |
1054 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) | |
1055 | tp->lost_out -= decr; | |
1056 | ||
1057 | /* Reno case is special. Sigh... */ | |
1058 | if (tcp_is_reno(tp) && decr > 0) | |
1059 | tp->sacked_out -= min_t(u32, tp->sacked_out, decr); | |
1060 | ||
1061 | tcp_adjust_fackets_out(sk, skb, decr); | |
1062 | ||
1063 | if (tp->lost_skb_hint && | |
1064 | before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && | |
52cf3cc8 | 1065 | (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) |
797108d1 IJ |
1066 | tp->lost_cnt_hint -= decr; |
1067 | ||
1068 | tcp_verify_left_out(tp); | |
1069 | } | |
1070 | ||
1da177e4 LT |
1071 | /* Function to create two new TCP segments. Shrinks the given segment |
1072 | * to the specified size and appends a new segment with the rest of the | |
e905a9ed | 1073 | * packet to the list. This won't be called frequently, I hope. |
1da177e4 LT |
1074 | * Remember, these are still headerless SKBs at this point. |
1075 | */ | |
056834d9 IJ |
1076 | int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, |
1077 | unsigned int mss_now) | |
1da177e4 LT |
1078 | { |
1079 | struct tcp_sock *tp = tcp_sk(sk); | |
1080 | struct sk_buff *buff; | |
6475be16 | 1081 | int nsize, old_factor; |
b60b49ea | 1082 | int nlen; |
9ce01461 | 1083 | u8 flags; |
1da177e4 | 1084 | |
2fceec13 IJ |
1085 | if (WARN_ON(len > skb->len)) |
1086 | return -EINVAL; | |
6a438bbe | 1087 | |
1da177e4 LT |
1088 | nsize = skb_headlen(skb) - len; |
1089 | if (nsize < 0) | |
1090 | nsize = 0; | |
1091 | ||
c52e2421 | 1092 | if (skb_unclone(skb, GFP_ATOMIC)) |
1da177e4 LT |
1093 | return -ENOMEM; |
1094 | ||
1095 | /* Get a new skb... force flag on. */ | |
1096 | buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); | |
1097 | if (buff == NULL) | |
1098 | return -ENOMEM; /* We'll just try again later. */ | |
ef5cb973 | 1099 | |
3ab224be HA |
1100 | sk->sk_wmem_queued += buff->truesize; |
1101 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea HX |
1102 | nlen = skb->len - len - nsize; |
1103 | buff->truesize += nlen; | |
1104 | skb->truesize -= nlen; | |
1da177e4 LT |
1105 | |
1106 | /* Correct the sequence numbers. */ | |
1107 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1108 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1109 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1110 | ||
1111 | /* PSH and FIN should only be set in the second packet. */ | |
4de075e0 ED |
1112 | flags = TCP_SKB_CB(skb)->tcp_flags; |
1113 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
1114 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
e14c3caf | 1115 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
1da177e4 | 1116 | |
84fa7933 | 1117 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { |
1da177e4 | 1118 | /* Copy and checksum data tail into the new buffer. */ |
056834d9 IJ |
1119 | buff->csum = csum_partial_copy_nocheck(skb->data + len, |
1120 | skb_put(buff, nsize), | |
1da177e4 LT |
1121 | nsize, 0); |
1122 | ||
1123 | skb_trim(skb, len); | |
1124 | ||
1125 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); | |
1126 | } else { | |
84fa7933 | 1127 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
1128 | skb_split(skb, buff, len); |
1129 | } | |
1130 | ||
1131 | buff->ip_summed = skb->ip_summed; | |
1132 | ||
1133 | /* Looks stupid, but our code really uses when of | |
1134 | * skbs, which it never sent before. --ANK | |
1135 | */ | |
1136 | TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; | |
a61bbcf2 | 1137 | buff->tstamp = skb->tstamp; |
1da177e4 | 1138 | |
6475be16 DM |
1139 | old_factor = tcp_skb_pcount(skb); |
1140 | ||
1da177e4 | 1141 | /* Fix up tso_factor for both original and new SKB. */ |
846998ae DM |
1142 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
1143 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
1da177e4 | 1144 | |
6475be16 DM |
1145 | /* If this packet has been sent out already, we must |
1146 | * adjust the various packet counters. | |
1147 | */ | |
cf0b450c | 1148 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
6475be16 DM |
1149 | int diff = old_factor - tcp_skb_pcount(skb) - |
1150 | tcp_skb_pcount(buff); | |
1da177e4 | 1151 | |
797108d1 IJ |
1152 | if (diff) |
1153 | tcp_adjust_pcount(sk, skb, diff); | |
1da177e4 LT |
1154 | } |
1155 | ||
1156 | /* Link BUFF into the send queue. */ | |
f44b5271 | 1157 | skb_header_release(buff); |
fe067e8a | 1158 | tcp_insert_write_queue_after(skb, buff, sk); |
1da177e4 LT |
1159 | |
1160 | return 0; | |
1161 | } | |
1162 | ||
1163 | /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c | |
1164 | * eventually). The difference is that pulled data not copied, but | |
1165 | * immediately discarded. | |
1166 | */ | |
f2911969 | 1167 | static void __pskb_trim_head(struct sk_buff *skb, int len) |
1da177e4 | 1168 | { |
7b7fc97a | 1169 | struct skb_shared_info *shinfo; |
1da177e4 LT |
1170 | int i, k, eat; |
1171 | ||
4fa48bf3 ED |
1172 | eat = min_t(int, len, skb_headlen(skb)); |
1173 | if (eat) { | |
1174 | __skb_pull(skb, eat); | |
1175 | len -= eat; | |
1176 | if (!len) | |
1177 | return; | |
1178 | } | |
1da177e4 LT |
1179 | eat = len; |
1180 | k = 0; | |
7b7fc97a ED |
1181 | shinfo = skb_shinfo(skb); |
1182 | for (i = 0; i < shinfo->nr_frags; i++) { | |
1183 | int size = skb_frag_size(&shinfo->frags[i]); | |
9e903e08 ED |
1184 | |
1185 | if (size <= eat) { | |
aff65da0 | 1186 | skb_frag_unref(skb, i); |
9e903e08 | 1187 | eat -= size; |
1da177e4 | 1188 | } else { |
7b7fc97a | 1189 | shinfo->frags[k] = shinfo->frags[i]; |
1da177e4 | 1190 | if (eat) { |
7b7fc97a ED |
1191 | shinfo->frags[k].page_offset += eat; |
1192 | skb_frag_size_sub(&shinfo->frags[k], eat); | |
1da177e4 LT |
1193 | eat = 0; |
1194 | } | |
1195 | k++; | |
1196 | } | |
1197 | } | |
7b7fc97a | 1198 | shinfo->nr_frags = k; |
1da177e4 | 1199 | |
27a884dc | 1200 | skb_reset_tail_pointer(skb); |
1da177e4 LT |
1201 | skb->data_len -= len; |
1202 | skb->len = skb->data_len; | |
1da177e4 LT |
1203 | } |
1204 | ||
67edfef7 | 1205 | /* Remove acked data from a packet in the transmit queue. */ |
1da177e4 LT |
1206 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) |
1207 | { | |
14bbd6a5 | 1208 | if (skb_unclone(skb, GFP_ATOMIC)) |
1da177e4 LT |
1209 | return -ENOMEM; |
1210 | ||
4fa48bf3 | 1211 | __pskb_trim_head(skb, len); |
1da177e4 LT |
1212 | |
1213 | TCP_SKB_CB(skb)->seq += len; | |
84fa7933 | 1214 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
1215 | |
1216 | skb->truesize -= len; | |
1217 | sk->sk_wmem_queued -= len; | |
3ab224be | 1218 | sk_mem_uncharge(sk, len); |
1da177e4 LT |
1219 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); |
1220 | ||
5b35e1e6 | 1221 | /* Any change of skb->len requires recalculation of tso factor. */ |
1da177e4 | 1222 | if (tcp_skb_pcount(skb) > 1) |
5b35e1e6 | 1223 | tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); |
1da177e4 LT |
1224 | |
1225 | return 0; | |
1226 | } | |
1227 | ||
1b63edd6 YC |
1228 | /* Calculate MSS not accounting any TCP options. */ |
1229 | static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
5d424d5a | 1230 | { |
cf533ea5 ED |
1231 | const struct tcp_sock *tp = tcp_sk(sk); |
1232 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
5d424d5a JH |
1233 | int mss_now; |
1234 | ||
1235 | /* Calculate base mss without TCP options: | |
1236 | It is MMS_S - sizeof(tcphdr) of rfc1122 | |
1237 | */ | |
1238 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); | |
1239 | ||
67469601 ED |
1240 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
1241 | if (icsk->icsk_af_ops->net_frag_header_len) { | |
1242 | const struct dst_entry *dst = __sk_dst_get(sk); | |
1243 | ||
1244 | if (dst && dst_allfrag(dst)) | |
1245 | mss_now -= icsk->icsk_af_ops->net_frag_header_len; | |
1246 | } | |
1247 | ||
5d424d5a JH |
1248 | /* Clamp it (mss_clamp does not include tcp options) */ |
1249 | if (mss_now > tp->rx_opt.mss_clamp) | |
1250 | mss_now = tp->rx_opt.mss_clamp; | |
1251 | ||
1252 | /* Now subtract optional transport overhead */ | |
1253 | mss_now -= icsk->icsk_ext_hdr_len; | |
1254 | ||
1255 | /* Then reserve room for full set of TCP options and 8 bytes of data */ | |
1256 | if (mss_now < 48) | |
1257 | mss_now = 48; | |
5d424d5a JH |
1258 | return mss_now; |
1259 | } | |
1260 | ||
1b63edd6 YC |
1261 | /* Calculate MSS. Not accounting for SACKs here. */ |
1262 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
1263 | { | |
1264 | /* Subtract TCP options size, not including SACKs */ | |
1265 | return __tcp_mtu_to_mss(sk, pmtu) - | |
1266 | (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); | |
1267 | } | |
1268 | ||
5d424d5a | 1269 | /* Inverse of above */ |
67469601 | 1270 | int tcp_mss_to_mtu(struct sock *sk, int mss) |
5d424d5a | 1271 | { |
cf533ea5 ED |
1272 | const struct tcp_sock *tp = tcp_sk(sk); |
1273 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
5d424d5a JH |
1274 | int mtu; |
1275 | ||
1276 | mtu = mss + | |
1277 | tp->tcp_header_len + | |
1278 | icsk->icsk_ext_hdr_len + | |
1279 | icsk->icsk_af_ops->net_header_len; | |
1280 | ||
67469601 ED |
1281 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
1282 | if (icsk->icsk_af_ops->net_frag_header_len) { | |
1283 | const struct dst_entry *dst = __sk_dst_get(sk); | |
1284 | ||
1285 | if (dst && dst_allfrag(dst)) | |
1286 | mtu += icsk->icsk_af_ops->net_frag_header_len; | |
1287 | } | |
5d424d5a JH |
1288 | return mtu; |
1289 | } | |
1290 | ||
67edfef7 | 1291 | /* MTU probing init per socket */ |
5d424d5a JH |
1292 | void tcp_mtup_init(struct sock *sk) |
1293 | { | |
1294 | struct tcp_sock *tp = tcp_sk(sk); | |
1295 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1296 | ||
1297 | icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; | |
1298 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + | |
e905a9ed | 1299 | icsk->icsk_af_ops->net_header_len; |
5d424d5a JH |
1300 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); |
1301 | icsk->icsk_mtup.probe_size = 0; | |
1302 | } | |
4bc2f18b | 1303 | EXPORT_SYMBOL(tcp_mtup_init); |
5d424d5a | 1304 | |
1da177e4 LT |
1305 | /* This function synchronize snd mss to current pmtu/exthdr set. |
1306 | ||
1307 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts | |
1308 | for TCP options, but includes only bare TCP header. | |
1309 | ||
1310 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. | |
caa20d9a | 1311 | It is minimum of user_mss and mss received with SYN. |
1da177e4 LT |
1312 | It also does not include TCP options. |
1313 | ||
d83d8461 | 1314 | inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. |
1da177e4 LT |
1315 | |
1316 | tp->mss_cache is current effective sending mss, including | |
1317 | all tcp options except for SACKs. It is evaluated, | |
1318 | taking into account current pmtu, but never exceeds | |
1319 | tp->rx_opt.mss_clamp. | |
1320 | ||
1321 | NOTE1. rfc1122 clearly states that advertised MSS | |
1322 | DOES NOT include either tcp or ip options. | |
1323 | ||
d83d8461 ACM |
1324 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
1325 | are READ ONLY outside this function. --ANK (980731) | |
1da177e4 | 1326 | */ |
1da177e4 LT |
1327 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) |
1328 | { | |
1329 | struct tcp_sock *tp = tcp_sk(sk); | |
d83d8461 | 1330 | struct inet_connection_sock *icsk = inet_csk(sk); |
5d424d5a | 1331 | int mss_now; |
1da177e4 | 1332 | |
5d424d5a JH |
1333 | if (icsk->icsk_mtup.search_high > pmtu) |
1334 | icsk->icsk_mtup.search_high = pmtu; | |
1da177e4 | 1335 | |
5d424d5a | 1336 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
409d22b4 | 1337 | mss_now = tcp_bound_to_half_wnd(tp, mss_now); |
1da177e4 LT |
1338 | |
1339 | /* And store cached results */ | |
d83d8461 | 1340 | icsk->icsk_pmtu_cookie = pmtu; |
5d424d5a JH |
1341 | if (icsk->icsk_mtup.enabled) |
1342 | mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); | |
c1b4a7e6 | 1343 | tp->mss_cache = mss_now; |
1da177e4 LT |
1344 | |
1345 | return mss_now; | |
1346 | } | |
4bc2f18b | 1347 | EXPORT_SYMBOL(tcp_sync_mss); |
1da177e4 LT |
1348 | |
1349 | /* Compute the current effective MSS, taking SACKs and IP options, | |
1350 | * and even PMTU discovery events into account. | |
1da177e4 | 1351 | */ |
0c54b85f | 1352 | unsigned int tcp_current_mss(struct sock *sk) |
1da177e4 | 1353 | { |
cf533ea5 ED |
1354 | const struct tcp_sock *tp = tcp_sk(sk); |
1355 | const struct dst_entry *dst = __sk_dst_get(sk); | |
c1b4a7e6 | 1356 | u32 mss_now; |
95c96174 | 1357 | unsigned int header_len; |
33ad798c AL |
1358 | struct tcp_out_options opts; |
1359 | struct tcp_md5sig_key *md5; | |
c1b4a7e6 DM |
1360 | |
1361 | mss_now = tp->mss_cache; | |
1362 | ||
1da177e4 LT |
1363 | if (dst) { |
1364 | u32 mtu = dst_mtu(dst); | |
d83d8461 | 1365 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
1da177e4 LT |
1366 | mss_now = tcp_sync_mss(sk, mtu); |
1367 | } | |
1368 | ||
33ad798c AL |
1369 | header_len = tcp_established_options(sk, NULL, &opts, &md5) + |
1370 | sizeof(struct tcphdr); | |
1371 | /* The mss_cache is sized based on tp->tcp_header_len, which assumes | |
1372 | * some common options. If this is an odd packet (because we have SACK | |
1373 | * blocks etc) then our calculated header_len will be different, and | |
1374 | * we have to adjust mss_now correspondingly */ | |
1375 | if (header_len != tp->tcp_header_len) { | |
1376 | int delta = (int) header_len - tp->tcp_header_len; | |
1377 | mss_now -= delta; | |
1378 | } | |
cfb6eeb4 | 1379 | |
1da177e4 LT |
1380 | return mss_now; |
1381 | } | |
1382 | ||
86fd14ad WP |
1383 | /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. |
1384 | * As additional protections, we do not touch cwnd in retransmission phases, | |
1385 | * and if application hit its sndbuf limit recently. | |
1386 | */ | |
1387 | static void tcp_cwnd_application_limited(struct sock *sk) | |
1388 | { | |
1389 | struct tcp_sock *tp = tcp_sk(sk); | |
1390 | ||
1391 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && | |
1392 | sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { | |
1393 | /* Limited by application or receiver window. */ | |
1394 | u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); | |
1395 | u32 win_used = max(tp->snd_cwnd_used, init_win); | |
1396 | if (win_used < tp->snd_cwnd) { | |
1397 | tp->snd_ssthresh = tcp_current_ssthresh(sk); | |
1398 | tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; | |
1399 | } | |
1400 | tp->snd_cwnd_used = 0; | |
1401 | } | |
1402 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1403 | } | |
1404 | ||
e114a710 | 1405 | static void tcp_cwnd_validate(struct sock *sk, u32 unsent_segs) |
a762a980 | 1406 | { |
9e412ba7 | 1407 | struct tcp_sock *tp = tcp_sk(sk); |
a762a980 | 1408 | |
e114a710 ED |
1409 | tp->lsnd_pending = tp->packets_out + unsent_segs; |
1410 | ||
1411 | if (tcp_is_cwnd_limited(sk, 0)) { | |
a762a980 DM |
1412 | /* Network is feed fully. */ |
1413 | tp->snd_cwnd_used = 0; | |
1414 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1415 | } else { | |
1416 | /* Network starves. */ | |
1417 | if (tp->packets_out > tp->snd_cwnd_used) | |
1418 | tp->snd_cwnd_used = tp->packets_out; | |
1419 | ||
15d33c07 DM |
1420 | if (sysctl_tcp_slow_start_after_idle && |
1421 | (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) | |
a762a980 DM |
1422 | tcp_cwnd_application_limited(sk); |
1423 | } | |
1424 | } | |
1425 | ||
d4589926 ED |
1426 | /* Minshall's variant of the Nagle send check. */ |
1427 | static bool tcp_minshall_check(const struct tcp_sock *tp) | |
1428 | { | |
1429 | return after(tp->snd_sml, tp->snd_una) && | |
1430 | !after(tp->snd_sml, tp->snd_nxt); | |
1431 | } | |
1432 | ||
1433 | /* Update snd_sml if this skb is under mss | |
1434 | * Note that a TSO packet might end with a sub-mss segment | |
1435 | * The test is really : | |
1436 | * if ((skb->len % mss) != 0) | |
1437 | * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; | |
1438 | * But we can avoid doing the divide again given we already have | |
1439 | * skb_pcount = skb->len / mss_now | |
0e3a4803 | 1440 | */ |
d4589926 ED |
1441 | static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, |
1442 | const struct sk_buff *skb) | |
1443 | { | |
1444 | if (skb->len < tcp_skb_pcount(skb) * mss_now) | |
1445 | tp->snd_sml = TCP_SKB_CB(skb)->end_seq; | |
1446 | } | |
1447 | ||
1448 | /* Return false, if packet can be sent now without violation Nagle's rules: | |
1449 | * 1. It is full sized. (provided by caller in %partial bool) | |
1450 | * 2. Or it contains FIN. (already checked by caller) | |
1451 | * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. | |
1452 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. | |
1453 | * With Minshall's modification: all sent small packets are ACKed. | |
1454 | */ | |
1455 | static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, | |
cc93fc51 | 1456 | int nonagle) |
d4589926 ED |
1457 | { |
1458 | return partial && | |
1459 | ((nonagle & TCP_NAGLE_CORK) || | |
1460 | (!nonagle && tp->packets_out && tcp_minshall_check(tp))); | |
1461 | } | |
1462 | /* Returns the portion of skb which can be sent right away */ | |
1463 | static unsigned int tcp_mss_split_point(const struct sock *sk, | |
1464 | const struct sk_buff *skb, | |
1465 | unsigned int mss_now, | |
1466 | unsigned int max_segs, | |
1467 | int nonagle) | |
c1b4a7e6 | 1468 | { |
cf533ea5 | 1469 | const struct tcp_sock *tp = tcp_sk(sk); |
d4589926 | 1470 | u32 partial, needed, window, max_len; |
c1b4a7e6 | 1471 | |
90840def | 1472 | window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
1485348d | 1473 | max_len = mss_now * max_segs; |
0e3a4803 | 1474 | |
1485348d BH |
1475 | if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) |
1476 | return max_len; | |
0e3a4803 | 1477 | |
5ea3a748 IJ |
1478 | needed = min(skb->len, window); |
1479 | ||
1485348d BH |
1480 | if (max_len <= needed) |
1481 | return max_len; | |
0e3a4803 | 1482 | |
d4589926 ED |
1483 | partial = needed % mss_now; |
1484 | /* If last segment is not a full MSS, check if Nagle rules allow us | |
1485 | * to include this last segment in this skb. | |
1486 | * Otherwise, we'll split the skb at last MSS boundary | |
1487 | */ | |
cc93fc51 | 1488 | if (tcp_nagle_check(partial != 0, tp, nonagle)) |
d4589926 ED |
1489 | return needed - partial; |
1490 | ||
1491 | return needed; | |
c1b4a7e6 DM |
1492 | } |
1493 | ||
1494 | /* Can at least one segment of SKB be sent right now, according to the | |
1495 | * congestion window rules? If so, return how many segments are allowed. | |
1496 | */ | |
cf533ea5 ED |
1497 | static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, |
1498 | const struct sk_buff *skb) | |
c1b4a7e6 DM |
1499 | { |
1500 | u32 in_flight, cwnd; | |
1501 | ||
1502 | /* Don't be strict about the congestion window for the final FIN. */ | |
4de075e0 ED |
1503 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && |
1504 | tcp_skb_pcount(skb) == 1) | |
c1b4a7e6 DM |
1505 | return 1; |
1506 | ||
1507 | in_flight = tcp_packets_in_flight(tp); | |
1508 | cwnd = tp->snd_cwnd; | |
1509 | if (in_flight < cwnd) | |
1510 | return (cwnd - in_flight); | |
1511 | ||
1512 | return 0; | |
1513 | } | |
1514 | ||
b595076a | 1515 | /* Initialize TSO state of a skb. |
67edfef7 | 1516 | * This must be invoked the first time we consider transmitting |
c1b4a7e6 DM |
1517 | * SKB onto the wire. |
1518 | */ | |
cf533ea5 | 1519 | static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, |
056834d9 | 1520 | unsigned int mss_now) |
c1b4a7e6 DM |
1521 | { |
1522 | int tso_segs = tcp_skb_pcount(skb); | |
1523 | ||
f8269a49 | 1524 | if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { |
846998ae | 1525 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
c1b4a7e6 DM |
1526 | tso_segs = tcp_skb_pcount(skb); |
1527 | } | |
1528 | return tso_segs; | |
1529 | } | |
1530 | ||
c1b4a7e6 | 1531 | |
a2a385d6 | 1532 | /* Return true if the Nagle test allows this packet to be |
c1b4a7e6 DM |
1533 | * sent now. |
1534 | */ | |
a2a385d6 ED |
1535 | static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, |
1536 | unsigned int cur_mss, int nonagle) | |
c1b4a7e6 DM |
1537 | { |
1538 | /* Nagle rule does not apply to frames, which sit in the middle of the | |
1539 | * write_queue (they have no chances to get new data). | |
1540 | * | |
1541 | * This is implemented in the callers, where they modify the 'nonagle' | |
1542 | * argument based upon the location of SKB in the send queue. | |
1543 | */ | |
1544 | if (nonagle & TCP_NAGLE_PUSH) | |
a2a385d6 | 1545 | return true; |
c1b4a7e6 | 1546 | |
9b44190d YC |
1547 | /* Don't use the nagle rule for urgent data (or for the final FIN). */ |
1548 | if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) | |
a2a385d6 | 1549 | return true; |
c1b4a7e6 | 1550 | |
cc93fc51 | 1551 | if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) |
a2a385d6 | 1552 | return true; |
c1b4a7e6 | 1553 | |
a2a385d6 | 1554 | return false; |
c1b4a7e6 DM |
1555 | } |
1556 | ||
1557 | /* Does at least the first segment of SKB fit into the send window? */ | |
a2a385d6 ED |
1558 | static bool tcp_snd_wnd_test(const struct tcp_sock *tp, |
1559 | const struct sk_buff *skb, | |
1560 | unsigned int cur_mss) | |
c1b4a7e6 DM |
1561 | { |
1562 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
1563 | ||
1564 | if (skb->len > cur_mss) | |
1565 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; | |
1566 | ||
90840def | 1567 | return !after(end_seq, tcp_wnd_end(tp)); |
c1b4a7e6 DM |
1568 | } |
1569 | ||
fe067e8a | 1570 | /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) |
c1b4a7e6 DM |
1571 | * should be put on the wire right now. If so, it returns the number of |
1572 | * packets allowed by the congestion window. | |
1573 | */ | |
cf533ea5 | 1574 | static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, |
c1b4a7e6 DM |
1575 | unsigned int cur_mss, int nonagle) |
1576 | { | |
cf533ea5 | 1577 | const struct tcp_sock *tp = tcp_sk(sk); |
c1b4a7e6 DM |
1578 | unsigned int cwnd_quota; |
1579 | ||
846998ae | 1580 | tcp_init_tso_segs(sk, skb, cur_mss); |
c1b4a7e6 DM |
1581 | |
1582 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) | |
1583 | return 0; | |
1584 | ||
1585 | cwnd_quota = tcp_cwnd_test(tp, skb); | |
056834d9 | 1586 | if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) |
c1b4a7e6 DM |
1587 | cwnd_quota = 0; |
1588 | ||
1589 | return cwnd_quota; | |
1590 | } | |
1591 | ||
67edfef7 | 1592 | /* Test if sending is allowed right now. */ |
a2a385d6 | 1593 | bool tcp_may_send_now(struct sock *sk) |
c1b4a7e6 | 1594 | { |
cf533ea5 | 1595 | const struct tcp_sock *tp = tcp_sk(sk); |
fe067e8a | 1596 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 | 1597 | |
a02cec21 | 1598 | return skb && |
0c54b85f | 1599 | tcp_snd_test(sk, skb, tcp_current_mss(sk), |
c1b4a7e6 | 1600 | (tcp_skb_is_last(sk, skb) ? |
a02cec21 | 1601 | tp->nonagle : TCP_NAGLE_PUSH)); |
c1b4a7e6 DM |
1602 | } |
1603 | ||
1604 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet | |
1605 | * which is put after SKB on the list. It is very much like | |
1606 | * tcp_fragment() except that it may make several kinds of assumptions | |
1607 | * in order to speed up the splitting operation. In particular, we | |
1608 | * know that all the data is in scatter-gather pages, and that the | |
1609 | * packet has never been sent out before (and thus is not cloned). | |
1610 | */ | |
056834d9 | 1611 | static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, |
c4ead4c5 | 1612 | unsigned int mss_now, gfp_t gfp) |
c1b4a7e6 DM |
1613 | { |
1614 | struct sk_buff *buff; | |
1615 | int nlen = skb->len - len; | |
9ce01461 | 1616 | u8 flags; |
c1b4a7e6 DM |
1617 | |
1618 | /* All of a TSO frame must be composed of paged data. */ | |
c8ac3774 HX |
1619 | if (skb->len != skb->data_len) |
1620 | return tcp_fragment(sk, skb, len, mss_now); | |
c1b4a7e6 | 1621 | |
c4ead4c5 | 1622 | buff = sk_stream_alloc_skb(sk, 0, gfp); |
c1b4a7e6 DM |
1623 | if (unlikely(buff == NULL)) |
1624 | return -ENOMEM; | |
1625 | ||
3ab224be HA |
1626 | sk->sk_wmem_queued += buff->truesize; |
1627 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea | 1628 | buff->truesize += nlen; |
c1b4a7e6 DM |
1629 | skb->truesize -= nlen; |
1630 | ||
1631 | /* Correct the sequence numbers. */ | |
1632 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1633 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1634 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1635 | ||
1636 | /* PSH and FIN should only be set in the second packet. */ | |
4de075e0 ED |
1637 | flags = TCP_SKB_CB(skb)->tcp_flags; |
1638 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
1639 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
c1b4a7e6 DM |
1640 | |
1641 | /* This packet was never sent out yet, so no SACK bits. */ | |
1642 | TCP_SKB_CB(buff)->sacked = 0; | |
1643 | ||
84fa7933 | 1644 | buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; |
c1b4a7e6 DM |
1645 | skb_split(skb, buff, len); |
1646 | ||
1647 | /* Fix up tso_factor for both original and new SKB. */ | |
846998ae DM |
1648 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
1649 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
c1b4a7e6 DM |
1650 | |
1651 | /* Link BUFF into the send queue. */ | |
1652 | skb_header_release(buff); | |
fe067e8a | 1653 | tcp_insert_write_queue_after(skb, buff, sk); |
c1b4a7e6 DM |
1654 | |
1655 | return 0; | |
1656 | } | |
1657 | ||
1658 | /* Try to defer sending, if possible, in order to minimize the amount | |
1659 | * of TSO splitting we do. View it as a kind of TSO Nagle test. | |
1660 | * | |
1661 | * This algorithm is from John Heffner. | |
1662 | */ | |
a2a385d6 | 1663 | static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) |
c1b4a7e6 | 1664 | { |
9e412ba7 | 1665 | struct tcp_sock *tp = tcp_sk(sk); |
6687e988 | 1666 | const struct inet_connection_sock *icsk = inet_csk(sk); |
c1b4a7e6 | 1667 | u32 send_win, cong_win, limit, in_flight; |
ad9f4f50 | 1668 | int win_divisor; |
c1b4a7e6 | 1669 | |
4de075e0 | 1670 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) |
ae8064ac | 1671 | goto send_now; |
c1b4a7e6 | 1672 | |
6687e988 | 1673 | if (icsk->icsk_ca_state != TCP_CA_Open) |
ae8064ac JH |
1674 | goto send_now; |
1675 | ||
1676 | /* Defer for less than two clock ticks. */ | |
bd515c3e | 1677 | if (tp->tso_deferred && |
a2acde07 | 1678 | (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) |
ae8064ac | 1679 | goto send_now; |
908a75c1 | 1680 | |
c1b4a7e6 DM |
1681 | in_flight = tcp_packets_in_flight(tp); |
1682 | ||
056834d9 | 1683 | BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); |
c1b4a7e6 | 1684 | |
90840def | 1685 | send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
c1b4a7e6 DM |
1686 | |
1687 | /* From in_flight test above, we know that cwnd > in_flight. */ | |
1688 | cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; | |
1689 | ||
1690 | limit = min(send_win, cong_win); | |
1691 | ||
ba244fe9 | 1692 | /* If a full-sized TSO skb can be sent, do it. */ |
1485348d | 1693 | if (limit >= min_t(unsigned int, sk->sk_gso_max_size, |
95bd09eb | 1694 | tp->xmit_size_goal_segs * tp->mss_cache)) |
ae8064ac | 1695 | goto send_now; |
ba244fe9 | 1696 | |
62ad2761 IJ |
1697 | /* Middle in queue won't get any more data, full sendable already? */ |
1698 | if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) | |
1699 | goto send_now; | |
1700 | ||
ad9f4f50 ED |
1701 | win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); |
1702 | if (win_divisor) { | |
c1b4a7e6 DM |
1703 | u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); |
1704 | ||
1705 | /* If at least some fraction of a window is available, | |
1706 | * just use it. | |
1707 | */ | |
ad9f4f50 | 1708 | chunk /= win_divisor; |
c1b4a7e6 | 1709 | if (limit >= chunk) |
ae8064ac | 1710 | goto send_now; |
c1b4a7e6 DM |
1711 | } else { |
1712 | /* Different approach, try not to defer past a single | |
1713 | * ACK. Receiver should ACK every other full sized | |
1714 | * frame, so if we have space for more than 3 frames | |
1715 | * then send now. | |
1716 | */ | |
6b5a5c0d | 1717 | if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) |
ae8064ac | 1718 | goto send_now; |
c1b4a7e6 DM |
1719 | } |
1720 | ||
f4541d60 ED |
1721 | /* Ok, it looks like it is advisable to defer. |
1722 | * Do not rearm the timer if already set to not break TCP ACK clocking. | |
1723 | */ | |
1724 | if (!tp->tso_deferred) | |
1725 | tp->tso_deferred = 1 | (jiffies << 1); | |
ae8064ac | 1726 | |
a2a385d6 | 1727 | return true; |
ae8064ac JH |
1728 | |
1729 | send_now: | |
1730 | tp->tso_deferred = 0; | |
a2a385d6 | 1731 | return false; |
c1b4a7e6 DM |
1732 | } |
1733 | ||
5d424d5a | 1734 | /* Create a new MTU probe if we are ready. |
67edfef7 AK |
1735 | * MTU probe is regularly attempting to increase the path MTU by |
1736 | * deliberately sending larger packets. This discovers routing | |
1737 | * changes resulting in larger path MTUs. | |
1738 | * | |
5d424d5a JH |
1739 | * Returns 0 if we should wait to probe (no cwnd available), |
1740 | * 1 if a probe was sent, | |
056834d9 IJ |
1741 | * -1 otherwise |
1742 | */ | |
5d424d5a JH |
1743 | static int tcp_mtu_probe(struct sock *sk) |
1744 | { | |
1745 | struct tcp_sock *tp = tcp_sk(sk); | |
1746 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1747 | struct sk_buff *skb, *nskb, *next; | |
1748 | int len; | |
1749 | int probe_size; | |
91cc17c0 | 1750 | int size_needed; |
5d424d5a JH |
1751 | int copy; |
1752 | int mss_now; | |
1753 | ||
1754 | /* Not currently probing/verifying, | |
1755 | * not in recovery, | |
1756 | * have enough cwnd, and | |
1757 | * not SACKing (the variable headers throw things off) */ | |
1758 | if (!icsk->icsk_mtup.enabled || | |
1759 | icsk->icsk_mtup.probe_size || | |
1760 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || | |
1761 | tp->snd_cwnd < 11 || | |
cabeccbd | 1762 | tp->rx_opt.num_sacks || tp->rx_opt.dsack) |
5d424d5a JH |
1763 | return -1; |
1764 | ||
1765 | /* Very simple search strategy: just double the MSS. */ | |
0c54b85f | 1766 | mss_now = tcp_current_mss(sk); |
056834d9 | 1767 | probe_size = 2 * tp->mss_cache; |
91cc17c0 | 1768 | size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; |
5d424d5a JH |
1769 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { |
1770 | /* TODO: set timer for probe_converge_event */ | |
1771 | return -1; | |
1772 | } | |
1773 | ||
1774 | /* Have enough data in the send queue to probe? */ | |
7f9c33e5 | 1775 | if (tp->write_seq - tp->snd_nxt < size_needed) |
5d424d5a JH |
1776 | return -1; |
1777 | ||
91cc17c0 IJ |
1778 | if (tp->snd_wnd < size_needed) |
1779 | return -1; | |
90840def | 1780 | if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) |
91cc17c0 | 1781 | return 0; |
5d424d5a | 1782 | |
d67c58e9 IJ |
1783 | /* Do we need to wait to drain cwnd? With none in flight, don't stall */ |
1784 | if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { | |
1785 | if (!tcp_packets_in_flight(tp)) | |
5d424d5a JH |
1786 | return -1; |
1787 | else | |
1788 | return 0; | |
1789 | } | |
1790 | ||
1791 | /* We're allowed to probe. Build it now. */ | |
1792 | if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) | |
1793 | return -1; | |
3ab224be HA |
1794 | sk->sk_wmem_queued += nskb->truesize; |
1795 | sk_mem_charge(sk, nskb->truesize); | |
5d424d5a | 1796 | |
fe067e8a | 1797 | skb = tcp_send_head(sk); |
5d424d5a JH |
1798 | |
1799 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; | |
1800 | TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; | |
4de075e0 | 1801 | TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; |
5d424d5a JH |
1802 | TCP_SKB_CB(nskb)->sacked = 0; |
1803 | nskb->csum = 0; | |
84fa7933 | 1804 | nskb->ip_summed = skb->ip_summed; |
5d424d5a | 1805 | |
50c4817e IJ |
1806 | tcp_insert_write_queue_before(nskb, skb, sk); |
1807 | ||
5d424d5a | 1808 | len = 0; |
234b6860 | 1809 | tcp_for_write_queue_from_safe(skb, next, sk) { |
5d424d5a JH |
1810 | copy = min_t(int, skb->len, probe_size - len); |
1811 | if (nskb->ip_summed) | |
1812 | skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); | |
1813 | else | |
1814 | nskb->csum = skb_copy_and_csum_bits(skb, 0, | |
056834d9 IJ |
1815 | skb_put(nskb, copy), |
1816 | copy, nskb->csum); | |
5d424d5a JH |
1817 | |
1818 | if (skb->len <= copy) { | |
1819 | /* We've eaten all the data from this skb. | |
1820 | * Throw it away. */ | |
4de075e0 | 1821 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; |
fe067e8a | 1822 | tcp_unlink_write_queue(skb, sk); |
3ab224be | 1823 | sk_wmem_free_skb(sk, skb); |
5d424d5a | 1824 | } else { |
4de075e0 | 1825 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & |
a3433f35 | 1826 | ~(TCPHDR_FIN|TCPHDR_PSH); |
5d424d5a JH |
1827 | if (!skb_shinfo(skb)->nr_frags) { |
1828 | skb_pull(skb, copy); | |
84fa7933 | 1829 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
056834d9 IJ |
1830 | skb->csum = csum_partial(skb->data, |
1831 | skb->len, 0); | |
5d424d5a JH |
1832 | } else { |
1833 | __pskb_trim_head(skb, copy); | |
1834 | tcp_set_skb_tso_segs(sk, skb, mss_now); | |
1835 | } | |
1836 | TCP_SKB_CB(skb)->seq += copy; | |
1837 | } | |
1838 | ||
1839 | len += copy; | |
234b6860 IJ |
1840 | |
1841 | if (len >= probe_size) | |
1842 | break; | |
5d424d5a JH |
1843 | } |
1844 | tcp_init_tso_segs(sk, nskb, nskb->len); | |
1845 | ||
1846 | /* We're ready to send. If this fails, the probe will | |
1847 | * be resegmented into mss-sized pieces by tcp_write_xmit(). */ | |
1848 | TCP_SKB_CB(nskb)->when = tcp_time_stamp; | |
1849 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { | |
1850 | /* Decrement cwnd here because we are sending | |
056834d9 | 1851 | * effectively two packets. */ |
5d424d5a | 1852 | tp->snd_cwnd--; |
66f5fe62 | 1853 | tcp_event_new_data_sent(sk, nskb); |
5d424d5a JH |
1854 | |
1855 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); | |
0e7b1368 JH |
1856 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; |
1857 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; | |
5d424d5a JH |
1858 | |
1859 | return 1; | |
1860 | } | |
1861 | ||
1862 | return -1; | |
1863 | } | |
1864 | ||
1da177e4 LT |
1865 | /* This routine writes packets to the network. It advances the |
1866 | * send_head. This happens as incoming acks open up the remote | |
1867 | * window for us. | |
1868 | * | |
f8269a49 IJ |
1869 | * LARGESEND note: !tcp_urg_mode is overkill, only frames between |
1870 | * snd_up-64k-mss .. snd_up cannot be large. However, taking into | |
1871 | * account rare use of URG, this is not a big flaw. | |
1872 | * | |
6ba8a3b1 ND |
1873 | * Send at most one packet when push_one > 0. Temporarily ignore |
1874 | * cwnd limit to force at most one packet out when push_one == 2. | |
1875 | ||
a2a385d6 ED |
1876 | * Returns true, if no segments are in flight and we have queued segments, |
1877 | * but cannot send anything now because of SWS or another problem. | |
1da177e4 | 1878 | */ |
a2a385d6 ED |
1879 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
1880 | int push_one, gfp_t gfp) | |
1da177e4 LT |
1881 | { |
1882 | struct tcp_sock *tp = tcp_sk(sk); | |
92df7b51 | 1883 | struct sk_buff *skb; |
e114a710 | 1884 | unsigned int tso_segs, sent_pkts, unsent_segs = 0; |
c1b4a7e6 | 1885 | int cwnd_quota; |
5d424d5a | 1886 | int result; |
1da177e4 | 1887 | |
92df7b51 | 1888 | sent_pkts = 0; |
5d424d5a | 1889 | |
d5dd9175 IJ |
1890 | if (!push_one) { |
1891 | /* Do MTU probing. */ | |
1892 | result = tcp_mtu_probe(sk); | |
1893 | if (!result) { | |
a2a385d6 | 1894 | return false; |
d5dd9175 IJ |
1895 | } else if (result > 0) { |
1896 | sent_pkts = 1; | |
1897 | } | |
5d424d5a JH |
1898 | } |
1899 | ||
fe067e8a | 1900 | while ((skb = tcp_send_head(sk))) { |
c8ac3774 HX |
1901 | unsigned int limit; |
1902 | ||
b68e9f85 | 1903 | tso_segs = tcp_init_tso_segs(sk, skb, mss_now); |
c1b4a7e6 | 1904 | BUG_ON(!tso_segs); |
aa93466b | 1905 | |
ec342325 AV |
1906 | if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) |
1907 | goto repair; /* Skip network transmission */ | |
1908 | ||
b68e9f85 | 1909 | cwnd_quota = tcp_cwnd_test(tp, skb); |
6ba8a3b1 ND |
1910 | if (!cwnd_quota) { |
1911 | if (push_one == 2) | |
1912 | /* Force out a loss probe pkt. */ | |
1913 | cwnd_quota = 1; | |
1914 | else | |
1915 | break; | |
1916 | } | |
b68e9f85 HX |
1917 | |
1918 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) | |
1919 | break; | |
1920 | ||
c1b4a7e6 DM |
1921 | if (tso_segs == 1) { |
1922 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, | |
1923 | (tcp_skb_is_last(sk, skb) ? | |
1924 | nonagle : TCP_NAGLE_PUSH)))) | |
1925 | break; | |
1926 | } else { | |
d5dd9175 | 1927 | if (!push_one && tcp_tso_should_defer(sk, skb)) |
e114a710 | 1928 | goto compute_unsent_segs; |
c1b4a7e6 | 1929 | } |
aa93466b | 1930 | |
c9eeec26 ED |
1931 | /* TCP Small Queues : |
1932 | * Control number of packets in qdisc/devices to two packets / or ~1 ms. | |
1933 | * This allows for : | |
1934 | * - better RTT estimation and ACK scheduling | |
1935 | * - faster recovery | |
1936 | * - high rates | |
98e09386 ED |
1937 | * Alas, some drivers / subsystems require a fair amount |
1938 | * of queued bytes to ensure line rate. | |
1939 | * One example is wifi aggregation (802.11 AMPDU) | |
46d3ceab | 1940 | */ |
98e09386 ED |
1941 | limit = max_t(unsigned int, sysctl_tcp_limit_output_bytes, |
1942 | sk->sk_pacing_rate >> 10); | |
c9eeec26 ED |
1943 | |
1944 | if (atomic_read(&sk->sk_wmem_alloc) > limit) { | |
46d3ceab | 1945 | set_bit(TSQ_THROTTLED, &tp->tsq_flags); |
bf06200e JO |
1946 | /* It is possible TX completion already happened |
1947 | * before we set TSQ_THROTTLED, so we must | |
1948 | * test again the condition. | |
1949 | * We abuse smp_mb__after_clear_bit() because | |
1950 | * there is no smp_mb__after_set_bit() yet | |
1951 | */ | |
1952 | smp_mb__after_clear_bit(); | |
e114a710 ED |
1953 | if (atomic_read(&sk->sk_wmem_alloc) > limit) { |
1954 | u32 unsent_bytes; | |
1955 | ||
1956 | compute_unsent_segs: | |
1957 | unsent_bytes = tp->write_seq - tp->snd_nxt; | |
1958 | unsent_segs = DIV_ROUND_UP(unsent_bytes, mss_now); | |
bf06200e | 1959 | break; |
e114a710 | 1960 | } |
46d3ceab | 1961 | } |
c9eeec26 | 1962 | |
c8ac3774 | 1963 | limit = mss_now; |
f8269a49 | 1964 | if (tso_segs > 1 && !tcp_urg_mode(tp)) |
0e3a4803 | 1965 | limit = tcp_mss_split_point(sk, skb, mss_now, |
1485348d BH |
1966 | min_t(unsigned int, |
1967 | cwnd_quota, | |
d4589926 ED |
1968 | sk->sk_gso_max_segs), |
1969 | nonagle); | |
1da177e4 | 1970 | |
c8ac3774 | 1971 | if (skb->len > limit && |
c4ead4c5 | 1972 | unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) |
c8ac3774 HX |
1973 | break; |
1974 | ||
92df7b51 | 1975 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
c1b4a7e6 | 1976 | |
d5dd9175 | 1977 | if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) |
92df7b51 | 1978 | break; |
1da177e4 | 1979 | |
ec342325 | 1980 | repair: |
92df7b51 DM |
1981 | /* Advance the send_head. This one is sent out. |
1982 | * This call will increment packets_out. | |
1983 | */ | |
66f5fe62 | 1984 | tcp_event_new_data_sent(sk, skb); |
1da177e4 | 1985 | |
92df7b51 | 1986 | tcp_minshall_update(tp, mss_now, skb); |
a262f0cd | 1987 | sent_pkts += tcp_skb_pcount(skb); |
d5dd9175 IJ |
1988 | |
1989 | if (push_one) | |
1990 | break; | |
92df7b51 | 1991 | } |
1da177e4 | 1992 | |
aa93466b | 1993 | if (likely(sent_pkts)) { |
684bad11 YC |
1994 | if (tcp_in_cwnd_reduction(sk)) |
1995 | tp->prr_out += sent_pkts; | |
6ba8a3b1 ND |
1996 | |
1997 | /* Send one loss probe per tail loss episode. */ | |
1998 | if (push_one != 2) | |
1999 | tcp_schedule_loss_probe(sk); | |
e114a710 | 2000 | tcp_cwnd_validate(sk, unsent_segs); |
a2a385d6 | 2001 | return false; |
1da177e4 | 2002 | } |
6ba8a3b1 ND |
2003 | return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk)); |
2004 | } | |
2005 | ||
2006 | bool tcp_schedule_loss_probe(struct sock *sk) | |
2007 | { | |
2008 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2009 | struct tcp_sock *tp = tcp_sk(sk); | |
2010 | u32 timeout, tlp_time_stamp, rto_time_stamp; | |
740b0f18 | 2011 | u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3); |
6ba8a3b1 ND |
2012 | |
2013 | if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) | |
2014 | return false; | |
2015 | /* No consecutive loss probes. */ | |
2016 | if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { | |
2017 | tcp_rearm_rto(sk); | |
2018 | return false; | |
2019 | } | |
2020 | /* Don't do any loss probe on a Fast Open connection before 3WHS | |
2021 | * finishes. | |
2022 | */ | |
2023 | if (sk->sk_state == TCP_SYN_RECV) | |
2024 | return false; | |
2025 | ||
2026 | /* TLP is only scheduled when next timer event is RTO. */ | |
2027 | if (icsk->icsk_pending != ICSK_TIME_RETRANS) | |
2028 | return false; | |
2029 | ||
2030 | /* Schedule a loss probe in 2*RTT for SACK capable connections | |
2031 | * in Open state, that are either limited by cwnd or application. | |
2032 | */ | |
740b0f18 | 2033 | if (sysctl_tcp_early_retrans < 3 || !tp->srtt_us || !tp->packets_out || |
6ba8a3b1 ND |
2034 | !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) |
2035 | return false; | |
2036 | ||
2037 | if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && | |
2038 | tcp_send_head(sk)) | |
2039 | return false; | |
2040 | ||
2041 | /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account | |
2042 | * for delayed ack when there's one outstanding packet. | |
2043 | */ | |
2044 | timeout = rtt << 1; | |
2045 | if (tp->packets_out == 1) | |
2046 | timeout = max_t(u32, timeout, | |
2047 | (rtt + (rtt >> 1) + TCP_DELACK_MAX)); | |
2048 | timeout = max_t(u32, timeout, msecs_to_jiffies(10)); | |
2049 | ||
2050 | /* If RTO is shorter, just schedule TLP in its place. */ | |
2051 | tlp_time_stamp = tcp_time_stamp + timeout; | |
2052 | rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; | |
2053 | if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { | |
2054 | s32 delta = rto_time_stamp - tcp_time_stamp; | |
2055 | if (delta > 0) | |
2056 | timeout = delta; | |
2057 | } | |
2058 | ||
2059 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, | |
2060 | TCP_RTO_MAX); | |
2061 | return true; | |
2062 | } | |
2063 | ||
1f3279ae ED |
2064 | /* Thanks to skb fast clones, we can detect if a prior transmit of |
2065 | * a packet is still in a qdisc or driver queue. | |
2066 | * In this case, there is very little point doing a retransmit ! | |
2067 | * Note: This is called from BH context only. | |
2068 | */ | |
2069 | static bool skb_still_in_host_queue(const struct sock *sk, | |
2070 | const struct sk_buff *skb) | |
2071 | { | |
2072 | const struct sk_buff *fclone = skb + 1; | |
2073 | ||
2074 | if (unlikely(skb->fclone == SKB_FCLONE_ORIG && | |
2075 | fclone->fclone == SKB_FCLONE_CLONE)) { | |
2076 | NET_INC_STATS_BH(sock_net(sk), | |
2077 | LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); | |
2078 | return true; | |
2079 | } | |
2080 | return false; | |
2081 | } | |
2082 | ||
6ba8a3b1 ND |
2083 | /* When probe timeout (PTO) fires, send a new segment if one exists, else |
2084 | * retransmit the last segment. | |
2085 | */ | |
2086 | void tcp_send_loss_probe(struct sock *sk) | |
2087 | { | |
9b717a8d | 2088 | struct tcp_sock *tp = tcp_sk(sk); |
6ba8a3b1 ND |
2089 | struct sk_buff *skb; |
2090 | int pcount; | |
2091 | int mss = tcp_current_mss(sk); | |
2092 | int err = -1; | |
2093 | ||
2094 | if (tcp_send_head(sk) != NULL) { | |
2095 | err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); | |
2096 | goto rearm_timer; | |
2097 | } | |
2098 | ||
9b717a8d ND |
2099 | /* At most one outstanding TLP retransmission. */ |
2100 | if (tp->tlp_high_seq) | |
2101 | goto rearm_timer; | |
2102 | ||
6ba8a3b1 ND |
2103 | /* Retransmit last segment. */ |
2104 | skb = tcp_write_queue_tail(sk); | |
2105 | if (WARN_ON(!skb)) | |
2106 | goto rearm_timer; | |
2107 | ||
1f3279ae ED |
2108 | if (skb_still_in_host_queue(sk, skb)) |
2109 | goto rearm_timer; | |
2110 | ||
6ba8a3b1 ND |
2111 | pcount = tcp_skb_pcount(skb); |
2112 | if (WARN_ON(!pcount)) | |
2113 | goto rearm_timer; | |
2114 | ||
2115 | if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { | |
2116 | if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss))) | |
2117 | goto rearm_timer; | |
2118 | skb = tcp_write_queue_tail(sk); | |
2119 | } | |
2120 | ||
2121 | if (WARN_ON(!skb || !tcp_skb_pcount(skb))) | |
2122 | goto rearm_timer; | |
2123 | ||
2124 | /* Probe with zero data doesn't trigger fast recovery. */ | |
2125 | if (skb->len > 0) | |
2126 | err = __tcp_retransmit_skb(sk, skb); | |
2127 | ||
9b717a8d ND |
2128 | /* Record snd_nxt for loss detection. */ |
2129 | if (likely(!err)) | |
2130 | tp->tlp_high_seq = tp->snd_nxt; | |
2131 | ||
6ba8a3b1 ND |
2132 | rearm_timer: |
2133 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, | |
2134 | inet_csk(sk)->icsk_rto, | |
2135 | TCP_RTO_MAX); | |
2136 | ||
2137 | if (likely(!err)) | |
2138 | NET_INC_STATS_BH(sock_net(sk), | |
2139 | LINUX_MIB_TCPLOSSPROBES); | |
1da177e4 LT |
2140 | } |
2141 | ||
a762a980 DM |
2142 | /* Push out any pending frames which were held back due to |
2143 | * TCP_CORK or attempt at coalescing tiny packets. | |
2144 | * The socket must be locked by the caller. | |
2145 | */ | |
9e412ba7 IJ |
2146 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, |
2147 | int nonagle) | |
a762a980 | 2148 | { |
726e07a8 IJ |
2149 | /* If we are closed, the bytes will have to remain here. |
2150 | * In time closedown will finish, we empty the write queue and | |
2151 | * all will be happy. | |
2152 | */ | |
2153 | if (unlikely(sk->sk_state == TCP_CLOSE)) | |
2154 | return; | |
2155 | ||
99a1dec7 MG |
2156 | if (tcp_write_xmit(sk, cur_mss, nonagle, 0, |
2157 | sk_gfp_atomic(sk, GFP_ATOMIC))) | |
726e07a8 | 2158 | tcp_check_probe_timer(sk); |
a762a980 DM |
2159 | } |
2160 | ||
c1b4a7e6 DM |
2161 | /* Send _single_ skb sitting at the send head. This function requires |
2162 | * true push pending frames to setup probe timer etc. | |
2163 | */ | |
2164 | void tcp_push_one(struct sock *sk, unsigned int mss_now) | |
2165 | { | |
fe067e8a | 2166 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 DM |
2167 | |
2168 | BUG_ON(!skb || skb->len < mss_now); | |
2169 | ||
d5dd9175 | 2170 | tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); |
c1b4a7e6 DM |
2171 | } |
2172 | ||
1da177e4 LT |
2173 | /* This function returns the amount that we can raise the |
2174 | * usable window based on the following constraints | |
e905a9ed | 2175 | * |
1da177e4 LT |
2176 | * 1. The window can never be shrunk once it is offered (RFC 793) |
2177 | * 2. We limit memory per socket | |
2178 | * | |
2179 | * RFC 1122: | |
2180 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep | |
2181 | * RECV.NEXT + RCV.WIN fixed until: | |
2182 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" | |
2183 | * | |
2184 | * i.e. don't raise the right edge of the window until you can raise | |
2185 | * it at least MSS bytes. | |
2186 | * | |
2187 | * Unfortunately, the recommended algorithm breaks header prediction, | |
2188 | * since header prediction assumes th->window stays fixed. | |
2189 | * | |
2190 | * Strictly speaking, keeping th->window fixed violates the receiver | |
2191 | * side SWS prevention criteria. The problem is that under this rule | |
2192 | * a stream of single byte packets will cause the right side of the | |
2193 | * window to always advance by a single byte. | |
e905a9ed | 2194 | * |
1da177e4 LT |
2195 | * Of course, if the sender implements sender side SWS prevention |
2196 | * then this will not be a problem. | |
e905a9ed | 2197 | * |
1da177e4 | 2198 | * BSD seems to make the following compromise: |
e905a9ed | 2199 | * |
1da177e4 LT |
2200 | * If the free space is less than the 1/4 of the maximum |
2201 | * space available and the free space is less than 1/2 mss, | |
2202 | * then set the window to 0. | |
2203 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] | |
2204 | * Otherwise, just prevent the window from shrinking | |
2205 | * and from being larger than the largest representable value. | |
2206 | * | |
2207 | * This prevents incremental opening of the window in the regime | |
2208 | * where TCP is limited by the speed of the reader side taking | |
2209 | * data out of the TCP receive queue. It does nothing about | |
2210 | * those cases where the window is constrained on the sender side | |
2211 | * because the pipeline is full. | |
2212 | * | |
2213 | * BSD also seems to "accidentally" limit itself to windows that are a | |
2214 | * multiple of MSS, at least until the free space gets quite small. | |
2215 | * This would appear to be a side effect of the mbuf implementation. | |
2216 | * Combining these two algorithms results in the observed behavior | |
2217 | * of having a fixed window size at almost all times. | |
2218 | * | |
2219 | * Below we obtain similar behavior by forcing the offered window to | |
2220 | * a multiple of the mss when it is feasible to do so. | |
2221 | * | |
2222 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. | |
2223 | * Regular options like TIMESTAMP are taken into account. | |
2224 | */ | |
2225 | u32 __tcp_select_window(struct sock *sk) | |
2226 | { | |
463c84b9 | 2227 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 2228 | struct tcp_sock *tp = tcp_sk(sk); |
caa20d9a | 2229 | /* MSS for the peer's data. Previous versions used mss_clamp |
1da177e4 LT |
2230 | * here. I don't know if the value based on our guesses |
2231 | * of peer's MSS is better for the performance. It's more correct | |
2232 | * but may be worse for the performance because of rcv_mss | |
2233 | * fluctuations. --SAW 1998/11/1 | |
2234 | */ | |
463c84b9 | 2235 | int mss = icsk->icsk_ack.rcv_mss; |
1da177e4 | 2236 | int free_space = tcp_space(sk); |
86c1a045 FW |
2237 | int allowed_space = tcp_full_space(sk); |
2238 | int full_space = min_t(int, tp->window_clamp, allowed_space); | |
1da177e4 LT |
2239 | int window; |
2240 | ||
2241 | if (mss > full_space) | |
e905a9ed | 2242 | mss = full_space; |
1da177e4 | 2243 | |
b92edbe0 | 2244 | if (free_space < (full_space >> 1)) { |
463c84b9 | 2245 | icsk->icsk_ack.quick = 0; |
1da177e4 | 2246 | |
180d8cd9 | 2247 | if (sk_under_memory_pressure(sk)) |
056834d9 IJ |
2248 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, |
2249 | 4U * tp->advmss); | |
1da177e4 | 2250 | |
86c1a045 FW |
2251 | /* free_space might become our new window, make sure we don't |
2252 | * increase it due to wscale. | |
2253 | */ | |
2254 | free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); | |
2255 | ||
2256 | /* if free space is less than mss estimate, or is below 1/16th | |
2257 | * of the maximum allowed, try to move to zero-window, else | |
2258 | * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and | |
2259 | * new incoming data is dropped due to memory limits. | |
2260 | * With large window, mss test triggers way too late in order | |
2261 | * to announce zero window in time before rmem limit kicks in. | |
2262 | */ | |
2263 | if (free_space < (allowed_space >> 4) || free_space < mss) | |
1da177e4 LT |
2264 | return 0; |
2265 | } | |
2266 | ||
2267 | if (free_space > tp->rcv_ssthresh) | |
2268 | free_space = tp->rcv_ssthresh; | |
2269 | ||
2270 | /* Don't do rounding if we are using window scaling, since the | |
2271 | * scaled window will not line up with the MSS boundary anyway. | |
2272 | */ | |
2273 | window = tp->rcv_wnd; | |
2274 | if (tp->rx_opt.rcv_wscale) { | |
2275 | window = free_space; | |
2276 | ||
2277 | /* Advertise enough space so that it won't get scaled away. | |
2278 | * Import case: prevent zero window announcement if | |
2279 | * 1<<rcv_wscale > mss. | |
2280 | */ | |
2281 | if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) | |
2282 | window = (((window >> tp->rx_opt.rcv_wscale) + 1) | |
2283 | << tp->rx_opt.rcv_wscale); | |
2284 | } else { | |
2285 | /* Get the largest window that is a nice multiple of mss. | |
2286 | * Window clamp already applied above. | |
2287 | * If our current window offering is within 1 mss of the | |
2288 | * free space we just keep it. This prevents the divide | |
2289 | * and multiply from happening most of the time. | |
2290 | * We also don't do any window rounding when the free space | |
2291 | * is too small. | |
2292 | */ | |
2293 | if (window <= free_space - mss || window > free_space) | |
056834d9 | 2294 | window = (free_space / mss) * mss; |
84565070 | 2295 | else if (mss == full_space && |
b92edbe0 | 2296 | free_space > window + (full_space >> 1)) |
84565070 | 2297 | window = free_space; |
1da177e4 LT |
2298 | } |
2299 | ||
2300 | return window; | |
2301 | } | |
2302 | ||
4a17fc3a IJ |
2303 | /* Collapses two adjacent SKB's during retransmission. */ |
2304 | static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) | |
1da177e4 LT |
2305 | { |
2306 | struct tcp_sock *tp = tcp_sk(sk); | |
fe067e8a | 2307 | struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); |
058dc334 | 2308 | int skb_size, next_skb_size; |
1da177e4 | 2309 | |
058dc334 IJ |
2310 | skb_size = skb->len; |
2311 | next_skb_size = next_skb->len; | |
1da177e4 | 2312 | |
058dc334 | 2313 | BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); |
a6963a6b | 2314 | |
058dc334 | 2315 | tcp_highest_sack_combine(sk, next_skb, skb); |
1da177e4 | 2316 | |
058dc334 | 2317 | tcp_unlink_write_queue(next_skb, sk); |
1da177e4 | 2318 | |
058dc334 IJ |
2319 | skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), |
2320 | next_skb_size); | |
1da177e4 | 2321 | |
058dc334 IJ |
2322 | if (next_skb->ip_summed == CHECKSUM_PARTIAL) |
2323 | skb->ip_summed = CHECKSUM_PARTIAL; | |
1da177e4 | 2324 | |
058dc334 IJ |
2325 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
2326 | skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); | |
1da177e4 | 2327 | |
058dc334 IJ |
2328 | /* Update sequence range on original skb. */ |
2329 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; | |
1da177e4 | 2330 | |
e6c7d085 | 2331 | /* Merge over control information. This moves PSH/FIN etc. over */ |
4de075e0 | 2332 | TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; |
058dc334 IJ |
2333 | |
2334 | /* All done, get rid of second SKB and account for it so | |
2335 | * packet counting does not break. | |
2336 | */ | |
2337 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; | |
058dc334 IJ |
2338 | |
2339 | /* changed transmit queue under us so clear hints */ | |
ef9da47c IJ |
2340 | tcp_clear_retrans_hints_partial(tp); |
2341 | if (next_skb == tp->retransmit_skb_hint) | |
2342 | tp->retransmit_skb_hint = skb; | |
058dc334 | 2343 | |
797108d1 IJ |
2344 | tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); |
2345 | ||
058dc334 | 2346 | sk_wmem_free_skb(sk, next_skb); |
1da177e4 LT |
2347 | } |
2348 | ||
67edfef7 | 2349 | /* Check if coalescing SKBs is legal. */ |
a2a385d6 | 2350 | static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) |
4a17fc3a IJ |
2351 | { |
2352 | if (tcp_skb_pcount(skb) > 1) | |
a2a385d6 | 2353 | return false; |
4a17fc3a IJ |
2354 | /* TODO: SACK collapsing could be used to remove this condition */ |
2355 | if (skb_shinfo(skb)->nr_frags != 0) | |
a2a385d6 | 2356 | return false; |
4a17fc3a | 2357 | if (skb_cloned(skb)) |
a2a385d6 | 2358 | return false; |
4a17fc3a | 2359 | if (skb == tcp_send_head(sk)) |
a2a385d6 | 2360 | return false; |
4a17fc3a IJ |
2361 | /* Some heurestics for collapsing over SACK'd could be invented */ |
2362 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
a2a385d6 | 2363 | return false; |
4a17fc3a | 2364 | |
a2a385d6 | 2365 | return true; |
4a17fc3a IJ |
2366 | } |
2367 | ||
67edfef7 AK |
2368 | /* Collapse packets in the retransmit queue to make to create |
2369 | * less packets on the wire. This is only done on retransmission. | |
2370 | */ | |
4a17fc3a IJ |
2371 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, |
2372 | int space) | |
2373 | { | |
2374 | struct tcp_sock *tp = tcp_sk(sk); | |
2375 | struct sk_buff *skb = to, *tmp; | |
a2a385d6 | 2376 | bool first = true; |
4a17fc3a IJ |
2377 | |
2378 | if (!sysctl_tcp_retrans_collapse) | |
2379 | return; | |
4de075e0 | 2380 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) |
4a17fc3a IJ |
2381 | return; |
2382 | ||
2383 | tcp_for_write_queue_from_safe(skb, tmp, sk) { | |
2384 | if (!tcp_can_collapse(sk, skb)) | |
2385 | break; | |
2386 | ||
2387 | space -= skb->len; | |
2388 | ||
2389 | if (first) { | |
a2a385d6 | 2390 | first = false; |
4a17fc3a IJ |
2391 | continue; |
2392 | } | |
2393 | ||
2394 | if (space < 0) | |
2395 | break; | |
2396 | /* Punt if not enough space exists in the first SKB for | |
2397 | * the data in the second | |
2398 | */ | |
a21d4572 | 2399 | if (skb->len > skb_availroom(to)) |
4a17fc3a IJ |
2400 | break; |
2401 | ||
2402 | if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) | |
2403 | break; | |
2404 | ||
2405 | tcp_collapse_retrans(sk, to); | |
2406 | } | |
2407 | } | |
2408 | ||
1da177e4 LT |
2409 | /* This retransmits one SKB. Policy decisions and retransmit queue |
2410 | * state updates are done by the caller. Returns non-zero if an | |
2411 | * error occurred which prevented the send. | |
2412 | */ | |
93b174ad | 2413 | int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) |
1da177e4 LT |
2414 | { |
2415 | struct tcp_sock *tp = tcp_sk(sk); | |
5d424d5a | 2416 | struct inet_connection_sock *icsk = inet_csk(sk); |
7d227cd2 | 2417 | unsigned int cur_mss; |
c84a5711 | 2418 | int err; |
1da177e4 | 2419 | |
5d424d5a JH |
2420 | /* Inconslusive MTU probe */ |
2421 | if (icsk->icsk_mtup.probe_size) { | |
2422 | icsk->icsk_mtup.probe_size = 0; | |
2423 | } | |
2424 | ||
1da177e4 | 2425 | /* Do not sent more than we queued. 1/4 is reserved for possible |
caa20d9a | 2426 | * copying overhead: fragmentation, tunneling, mangling etc. |
1da177e4 LT |
2427 | */ |
2428 | if (atomic_read(&sk->sk_wmem_alloc) > | |
2429 | min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) | |
2430 | return -EAGAIN; | |
2431 | ||
1f3279ae ED |
2432 | if (skb_still_in_host_queue(sk, skb)) |
2433 | return -EBUSY; | |
2434 | ||
1da177e4 LT |
2435 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { |
2436 | if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
2437 | BUG(); | |
1da177e4 LT |
2438 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
2439 | return -ENOMEM; | |
2440 | } | |
2441 | ||
7d227cd2 SS |
2442 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) |
2443 | return -EHOSTUNREACH; /* Routing failure or similar. */ | |
2444 | ||
0c54b85f | 2445 | cur_mss = tcp_current_mss(sk); |
7d227cd2 | 2446 | |
1da177e4 LT |
2447 | /* If receiver has shrunk his window, and skb is out of |
2448 | * new window, do not retransmit it. The exception is the | |
2449 | * case, when window is shrunk to zero. In this case | |
2450 | * our retransmit serves as a zero window probe. | |
2451 | */ | |
9d4fb27d JP |
2452 | if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && |
2453 | TCP_SKB_CB(skb)->seq != tp->snd_una) | |
1da177e4 LT |
2454 | return -EAGAIN; |
2455 | ||
2456 | if (skb->len > cur_mss) { | |
846998ae | 2457 | if (tcp_fragment(sk, skb, cur_mss, cur_mss)) |
1da177e4 | 2458 | return -ENOMEM; /* We'll try again later. */ |
02276f3c | 2459 | } else { |
9eb9362e IJ |
2460 | int oldpcount = tcp_skb_pcount(skb); |
2461 | ||
2462 | if (unlikely(oldpcount > 1)) { | |
c52e2421 ED |
2463 | if (skb_unclone(skb, GFP_ATOMIC)) |
2464 | return -ENOMEM; | |
9eb9362e IJ |
2465 | tcp_init_tso_segs(sk, skb, cur_mss); |
2466 | tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); | |
2467 | } | |
1da177e4 LT |
2468 | } |
2469 | ||
4a17fc3a | 2470 | tcp_retrans_try_collapse(sk, skb, cur_mss); |
1da177e4 | 2471 | |
1da177e4 LT |
2472 | /* Make a copy, if the first transmission SKB clone we made |
2473 | * is still in somebody's hands, else make a clone. | |
2474 | */ | |
2475 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
1da177e4 | 2476 | |
50bceae9 TG |
2477 | /* make sure skb->data is aligned on arches that require it |
2478 | * and check if ack-trimming & collapsing extended the headroom | |
2479 | * beyond what csum_start can cover. | |
2480 | */ | |
2481 | if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || | |
2482 | skb_headroom(skb) >= 0xFFFF)) { | |
117632e6 ED |
2483 | struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, |
2484 | GFP_ATOMIC); | |
c84a5711 YC |
2485 | err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : |
2486 | -ENOBUFS; | |
117632e6 | 2487 | } else { |
c84a5711 | 2488 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
117632e6 | 2489 | } |
c84a5711 YC |
2490 | |
2491 | if (likely(!err)) | |
2492 | TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; | |
2493 | return err; | |
93b174ad YC |
2494 | } |
2495 | ||
2496 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) | |
2497 | { | |
2498 | struct tcp_sock *tp = tcp_sk(sk); | |
2499 | int err = __tcp_retransmit_skb(sk, skb); | |
1da177e4 LT |
2500 | |
2501 | if (err == 0) { | |
2502 | /* Update global TCP statistics. */ | |
81cc8a75 | 2503 | TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); |
f19c29e3 YC |
2504 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) |
2505 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); | |
1da177e4 LT |
2506 | tp->total_retrans++; |
2507 | ||
2508 | #if FASTRETRANS_DEBUG > 0 | |
056834d9 | 2509 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
e87cc472 | 2510 | net_dbg_ratelimited("retrans_out leaked\n"); |
1da177e4 LT |
2511 | } |
2512 | #endif | |
b08d6cb2 IJ |
2513 | if (!tp->retrans_out) |
2514 | tp->lost_retrans_low = tp->snd_nxt; | |
1da177e4 LT |
2515 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; |
2516 | tp->retrans_out += tcp_skb_pcount(skb); | |
2517 | ||
2518 | /* Save stamp of the first retransmit. */ | |
2519 | if (!tp->retrans_stamp) | |
2520 | tp->retrans_stamp = TCP_SKB_CB(skb)->when; | |
2521 | ||
c24f691b | 2522 | tp->undo_retrans += tcp_skb_pcount(skb); |
1da177e4 LT |
2523 | |
2524 | /* snd_nxt is stored to detect loss of retransmitted segment, | |
2525 | * see tcp_input.c tcp_sacktag_write_queue(). | |
2526 | */ | |
2527 | TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; | |
1f3279ae | 2528 | } else if (err != -EBUSY) { |
24ab6bec | 2529 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); |
1da177e4 LT |
2530 | } |
2531 | return err; | |
2532 | } | |
2533 | ||
67edfef7 AK |
2534 | /* Check if we forward retransmits are possible in the current |
2535 | * window/congestion state. | |
2536 | */ | |
a2a385d6 | 2537 | static bool tcp_can_forward_retransmit(struct sock *sk) |
b5afe7bc IJ |
2538 | { |
2539 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
cf533ea5 | 2540 | const struct tcp_sock *tp = tcp_sk(sk); |
b5afe7bc IJ |
2541 | |
2542 | /* Forward retransmissions are possible only during Recovery. */ | |
2543 | if (icsk->icsk_ca_state != TCP_CA_Recovery) | |
a2a385d6 | 2544 | return false; |
b5afe7bc IJ |
2545 | |
2546 | /* No forward retransmissions in Reno are possible. */ | |
2547 | if (tcp_is_reno(tp)) | |
a2a385d6 | 2548 | return false; |
b5afe7bc IJ |
2549 | |
2550 | /* Yeah, we have to make difficult choice between forward transmission | |
2551 | * and retransmission... Both ways have their merits... | |
2552 | * | |
2553 | * For now we do not retransmit anything, while we have some new | |
2554 | * segments to send. In the other cases, follow rule 3 for | |
2555 | * NextSeg() specified in RFC3517. | |
2556 | */ | |
2557 | ||
2558 | if (tcp_may_send_now(sk)) | |
a2a385d6 | 2559 | return false; |
b5afe7bc | 2560 | |
a2a385d6 | 2561 | return true; |
b5afe7bc IJ |
2562 | } |
2563 | ||
1da177e4 LT |
2564 | /* This gets called after a retransmit timeout, and the initially |
2565 | * retransmitted data is acknowledged. It tries to continue | |
2566 | * resending the rest of the retransmit queue, until either | |
2567 | * we've sent it all or the congestion window limit is reached. | |
2568 | * If doing SACK, the first ACK which comes back for a timeout | |
2569 | * based retransmit packet might feed us FACK information again. | |
2570 | * If so, we use it to avoid unnecessarily retransmissions. | |
2571 | */ | |
2572 | void tcp_xmit_retransmit_queue(struct sock *sk) | |
2573 | { | |
6687e988 | 2574 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2575 | struct tcp_sock *tp = tcp_sk(sk); |
2576 | struct sk_buff *skb; | |
0e1c54c2 | 2577 | struct sk_buff *hole = NULL; |
618d9f25 | 2578 | u32 last_lost; |
61eb55f4 | 2579 | int mib_idx; |
0e1c54c2 | 2580 | int fwd_rexmitting = 0; |
6a438bbe | 2581 | |
45e77d31 IJ |
2582 | if (!tp->packets_out) |
2583 | return; | |
2584 | ||
08ebd172 IJ |
2585 | if (!tp->lost_out) |
2586 | tp->retransmit_high = tp->snd_una; | |
2587 | ||
618d9f25 | 2588 | if (tp->retransmit_skb_hint) { |
6a438bbe | 2589 | skb = tp->retransmit_skb_hint; |
618d9f25 IJ |
2590 | last_lost = TCP_SKB_CB(skb)->end_seq; |
2591 | if (after(last_lost, tp->retransmit_high)) | |
2592 | last_lost = tp->retransmit_high; | |
2593 | } else { | |
fe067e8a | 2594 | skb = tcp_write_queue_head(sk); |
618d9f25 IJ |
2595 | last_lost = tp->snd_una; |
2596 | } | |
1da177e4 | 2597 | |
08ebd172 IJ |
2598 | tcp_for_write_queue_from(skb, sk) { |
2599 | __u8 sacked = TCP_SKB_CB(skb)->sacked; | |
1da177e4 | 2600 | |
08ebd172 IJ |
2601 | if (skb == tcp_send_head(sk)) |
2602 | break; | |
2603 | /* we could do better than to assign each time */ | |
0e1c54c2 IJ |
2604 | if (hole == NULL) |
2605 | tp->retransmit_skb_hint = skb; | |
08ebd172 IJ |
2606 | |
2607 | /* Assume this retransmit will generate | |
2608 | * only one packet for congestion window | |
2609 | * calculation purposes. This works because | |
2610 | * tcp_retransmit_skb() will chop up the | |
2611 | * packet to be MSS sized and all the | |
2612 | * packet counting works out. | |
2613 | */ | |
2614 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) | |
2615 | return; | |
1da177e4 | 2616 | |
0e1c54c2 IJ |
2617 | if (fwd_rexmitting) { |
2618 | begin_fwd: | |
2619 | if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) | |
2620 | break; | |
2621 | mib_idx = LINUX_MIB_TCPFORWARDRETRANS; | |
6a438bbe | 2622 | |
0e1c54c2 | 2623 | } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { |
618d9f25 | 2624 | tp->retransmit_high = last_lost; |
0e1c54c2 IJ |
2625 | if (!tcp_can_forward_retransmit(sk)) |
2626 | break; | |
2627 | /* Backtrack if necessary to non-L'ed skb */ | |
2628 | if (hole != NULL) { | |
2629 | skb = hole; | |
2630 | hole = NULL; | |
2631 | } | |
2632 | fwd_rexmitting = 1; | |
2633 | goto begin_fwd; | |
1da177e4 | 2634 | |
0e1c54c2 | 2635 | } else if (!(sacked & TCPCB_LOST)) { |
ac11ba75 | 2636 | if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) |
0e1c54c2 IJ |
2637 | hole = skb; |
2638 | continue; | |
1da177e4 | 2639 | |
0e1c54c2 | 2640 | } else { |
618d9f25 | 2641 | last_lost = TCP_SKB_CB(skb)->end_seq; |
0e1c54c2 IJ |
2642 | if (icsk->icsk_ca_state != TCP_CA_Loss) |
2643 | mib_idx = LINUX_MIB_TCPFASTRETRANS; | |
2644 | else | |
2645 | mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; | |
2646 | } | |
1da177e4 | 2647 | |
0e1c54c2 | 2648 | if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) |
1da177e4 LT |
2649 | continue; |
2650 | ||
24ab6bec | 2651 | if (tcp_retransmit_skb(sk, skb)) |
0e1c54c2 | 2652 | return; |
24ab6bec | 2653 | |
0e1c54c2 | 2654 | NET_INC_STATS_BH(sock_net(sk), mib_idx); |
1da177e4 | 2655 | |
684bad11 | 2656 | if (tcp_in_cwnd_reduction(sk)) |
a262f0cd ND |
2657 | tp->prr_out += tcp_skb_pcount(skb); |
2658 | ||
fe067e8a | 2659 | if (skb == tcp_write_queue_head(sk)) |
3f421baa ACM |
2660 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
2661 | inet_csk(sk)->icsk_rto, | |
2662 | TCP_RTO_MAX); | |
1da177e4 LT |
2663 | } |
2664 | } | |
2665 | ||
1da177e4 LT |
2666 | /* Send a fin. The caller locks the socket for us. This cannot be |
2667 | * allowed to fail queueing a FIN frame under any circumstances. | |
2668 | */ | |
2669 | void tcp_send_fin(struct sock *sk) | |
2670 | { | |
e905a9ed | 2671 | struct tcp_sock *tp = tcp_sk(sk); |
fe067e8a | 2672 | struct sk_buff *skb = tcp_write_queue_tail(sk); |
1da177e4 | 2673 | int mss_now; |
e905a9ed | 2674 | |
1da177e4 LT |
2675 | /* Optimization, tack on the FIN if we have a queue of |
2676 | * unsent frames. But be careful about outgoing SACKS | |
2677 | * and IP options. | |
2678 | */ | |
0c54b85f | 2679 | mss_now = tcp_current_mss(sk); |
1da177e4 | 2680 | |
fe067e8a | 2681 | if (tcp_send_head(sk) != NULL) { |
4de075e0 | 2682 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN; |
1da177e4 LT |
2683 | TCP_SKB_CB(skb)->end_seq++; |
2684 | tp->write_seq++; | |
2685 | } else { | |
2686 | /* Socket is locked, keep trying until memory is available. */ | |
2687 | for (;;) { | |
aa133076 WF |
2688 | skb = alloc_skb_fclone(MAX_TCP_HEADER, |
2689 | sk->sk_allocation); | |
1da177e4 LT |
2690 | if (skb) |
2691 | break; | |
2692 | yield(); | |
2693 | } | |
2694 | ||
2695 | /* Reserve space for headers and prepare control bits. */ | |
2696 | skb_reserve(skb, MAX_TCP_HEADER); | |
1da177e4 | 2697 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ |
e870a8ef | 2698 | tcp_init_nondata_skb(skb, tp->write_seq, |
a3433f35 | 2699 | TCPHDR_ACK | TCPHDR_FIN); |
1da177e4 LT |
2700 | tcp_queue_skb(sk, skb); |
2701 | } | |
9e412ba7 | 2702 | __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); |
1da177e4 LT |
2703 | } |
2704 | ||
2705 | /* We get here when a process closes a file descriptor (either due to | |
2706 | * an explicit close() or as a byproduct of exit()'ing) and there | |
2707 | * was unread data in the receive queue. This behavior is recommended | |
65bb723c | 2708 | * by RFC 2525, section 2.17. -DaveM |
1da177e4 | 2709 | */ |
dd0fc66f | 2710 | void tcp_send_active_reset(struct sock *sk, gfp_t priority) |
1da177e4 | 2711 | { |
1da177e4 LT |
2712 | struct sk_buff *skb; |
2713 | ||
2714 | /* NOTE: No TCP options attached and we never retransmit this. */ | |
2715 | skb = alloc_skb(MAX_TCP_HEADER, priority); | |
2716 | if (!skb) { | |
4e673444 | 2717 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
1da177e4 LT |
2718 | return; |
2719 | } | |
2720 | ||
2721 | /* Reserve space for headers and prepare control bits. */ | |
2722 | skb_reserve(skb, MAX_TCP_HEADER); | |
e870a8ef | 2723 | tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), |
a3433f35 | 2724 | TCPHDR_ACK | TCPHDR_RST); |
1da177e4 | 2725 | /* Send it off. */ |
1da177e4 | 2726 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
dfb4b9dc | 2727 | if (tcp_transmit_skb(sk, skb, 0, priority)) |
4e673444 | 2728 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
26af65cb | 2729 | |
81cc8a75 | 2730 | TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); |
1da177e4 LT |
2731 | } |
2732 | ||
67edfef7 AK |
2733 | /* Send a crossed SYN-ACK during socket establishment. |
2734 | * WARNING: This routine must only be called when we have already sent | |
1da177e4 LT |
2735 | * a SYN packet that crossed the incoming SYN that caused this routine |
2736 | * to get called. If this assumption fails then the initial rcv_wnd | |
2737 | * and rcv_wscale values will not be correct. | |
2738 | */ | |
2739 | int tcp_send_synack(struct sock *sk) | |
2740 | { | |
056834d9 | 2741 | struct sk_buff *skb; |
1da177e4 | 2742 | |
fe067e8a | 2743 | skb = tcp_write_queue_head(sk); |
4de075e0 | 2744 | if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { |
91df42be | 2745 | pr_debug("%s: wrong queue state\n", __func__); |
1da177e4 LT |
2746 | return -EFAULT; |
2747 | } | |
4de075e0 | 2748 | if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { |
1da177e4 LT |
2749 | if (skb_cloned(skb)) { |
2750 | struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); | |
2751 | if (nskb == NULL) | |
2752 | return -ENOMEM; | |
fe067e8a | 2753 | tcp_unlink_write_queue(skb, sk); |
1da177e4 | 2754 | skb_header_release(nskb); |
fe067e8a | 2755 | __tcp_add_write_queue_head(sk, nskb); |
3ab224be HA |
2756 | sk_wmem_free_skb(sk, skb); |
2757 | sk->sk_wmem_queued += nskb->truesize; | |
2758 | sk_mem_charge(sk, nskb->truesize); | |
1da177e4 LT |
2759 | skb = nskb; |
2760 | } | |
2761 | ||
4de075e0 | 2762 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; |
1da177e4 LT |
2763 | TCP_ECN_send_synack(tcp_sk(sk), skb); |
2764 | } | |
2765 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
dfb4b9dc | 2766 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
1da177e4 LT |
2767 | } |
2768 | ||
4aea39c1 ED |
2769 | /** |
2770 | * tcp_make_synack - Prepare a SYN-ACK. | |
2771 | * sk: listener socket | |
2772 | * dst: dst entry attached to the SYNACK | |
2773 | * req: request_sock pointer | |
4aea39c1 ED |
2774 | * |
2775 | * Allocate one skb and build a SYNACK packet. | |
2776 | * @dst is consumed : Caller should not use it again. | |
2777 | */ | |
056834d9 | 2778 | struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, |
e6b4d113 | 2779 | struct request_sock *req, |
8336886f | 2780 | struct tcp_fastopen_cookie *foc) |
1da177e4 | 2781 | { |
bd0388ae | 2782 | struct tcp_out_options opts; |
2e6599cb | 2783 | struct inet_request_sock *ireq = inet_rsk(req); |
1da177e4 LT |
2784 | struct tcp_sock *tp = tcp_sk(sk); |
2785 | struct tcphdr *th; | |
1da177e4 | 2786 | struct sk_buff *skb; |
cfb6eeb4 | 2787 | struct tcp_md5sig_key *md5; |
bd0388ae | 2788 | int tcp_header_size; |
f5fff5dc | 2789 | int mss; |
1da177e4 | 2790 | |
a0b8486c | 2791 | skb = sock_wmalloc(sk, MAX_TCP_HEADER, 1, GFP_ATOMIC); |
4aea39c1 ED |
2792 | if (unlikely(!skb)) { |
2793 | dst_release(dst); | |
1da177e4 | 2794 | return NULL; |
4aea39c1 | 2795 | } |
1da177e4 LT |
2796 | /* Reserve space for headers. */ |
2797 | skb_reserve(skb, MAX_TCP_HEADER); | |
2798 | ||
4aea39c1 | 2799 | skb_dst_set(skb, dst); |
ca10b9e9 | 2800 | security_skb_owned_by(skb, sk); |
1da177e4 | 2801 | |
0dbaee3b | 2802 | mss = dst_metric_advmss(dst); |
f5fff5dc TQ |
2803 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) |
2804 | mss = tp->rx_opt.user_mss; | |
2805 | ||
33ad798c AL |
2806 | if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ |
2807 | __u8 rcv_wscale; | |
2808 | /* Set this up on the first call only */ | |
2809 | req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); | |
e88c64f0 HPP |
2810 | |
2811 | /* limit the window selection if the user enforce a smaller rx buffer */ | |
2812 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && | |
2813 | (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) | |
2814 | req->window_clamp = tcp_full_space(sk); | |
2815 | ||
33ad798c AL |
2816 | /* tcp_full_space because it is guaranteed to be the first packet */ |
2817 | tcp_select_initial_window(tcp_full_space(sk), | |
f5fff5dc | 2818 | mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), |
33ad798c AL |
2819 | &req->rcv_wnd, |
2820 | &req->window_clamp, | |
2821 | ireq->wscale_ok, | |
31d12926 | 2822 | &rcv_wscale, |
2823 | dst_metric(dst, RTAX_INITRWND)); | |
33ad798c AL |
2824 | ireq->rcv_wscale = rcv_wscale; |
2825 | } | |
2826 | ||
2827 | memset(&opts, 0, sizeof(opts)); | |
8b5f12d0 FW |
2828 | #ifdef CONFIG_SYN_COOKIES |
2829 | if (unlikely(req->cookie_ts)) | |
2830 | TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); | |
2831 | else | |
2832 | #endif | |
33ad798c | 2833 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
1a2c6181 CP |
2834 | tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5, |
2835 | foc) + sizeof(*th); | |
cfb6eeb4 | 2836 | |
aa8223c7 ACM |
2837 | skb_push(skb, tcp_header_size); |
2838 | skb_reset_transport_header(skb); | |
1da177e4 | 2839 | |
aa8223c7 | 2840 | th = tcp_hdr(skb); |
1da177e4 LT |
2841 | memset(th, 0, sizeof(struct tcphdr)); |
2842 | th->syn = 1; | |
2843 | th->ack = 1; | |
1da177e4 | 2844 | TCP_ECN_make_synack(req, th); |
b44084c2 | 2845 | th->source = htons(ireq->ir_num); |
634fb979 | 2846 | th->dest = ireq->ir_rmt_port; |
e870a8ef IJ |
2847 | /* Setting of flags are superfluous here for callers (and ECE is |
2848 | * not even correctly set) | |
2849 | */ | |
2850 | tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, | |
a3433f35 | 2851 | TCPHDR_SYN | TCPHDR_ACK); |
4957faad | 2852 | |
1da177e4 | 2853 | th->seq = htonl(TCP_SKB_CB(skb)->seq); |
8336886f JC |
2854 | /* XXX data is queued and acked as is. No buffer/window check */ |
2855 | th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); | |
1da177e4 LT |
2856 | |
2857 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ | |
600ff0c2 | 2858 | th->window = htons(min(req->rcv_wnd, 65535U)); |
bd0388ae | 2859 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
1da177e4 | 2860 | th->doff = (tcp_header_size >> 2); |
a0b8486c | 2861 | TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS); |
cfb6eeb4 YH |
2862 | |
2863 | #ifdef CONFIG_TCP_MD5SIG | |
2864 | /* Okay, we have all we need - do the md5 hash if needed */ | |
2865 | if (md5) { | |
bd0388ae | 2866 | tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, |
49a72dfb | 2867 | md5, NULL, req, skb); |
cfb6eeb4 YH |
2868 | } |
2869 | #endif | |
2870 | ||
1da177e4 LT |
2871 | return skb; |
2872 | } | |
4bc2f18b | 2873 | EXPORT_SYMBOL(tcp_make_synack); |
1da177e4 | 2874 | |
67edfef7 | 2875 | /* Do all connect socket setups that can be done AF independent. */ |
f7e56a76 | 2876 | static void tcp_connect_init(struct sock *sk) |
1da177e4 | 2877 | { |
cf533ea5 | 2878 | const struct dst_entry *dst = __sk_dst_get(sk); |
1da177e4 LT |
2879 | struct tcp_sock *tp = tcp_sk(sk); |
2880 | __u8 rcv_wscale; | |
2881 | ||
2882 | /* We'll fix this up when we get a response from the other end. | |
2883 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. | |
2884 | */ | |
2885 | tp->tcp_header_len = sizeof(struct tcphdr) + | |
bb5b7c11 | 2886 | (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); |
1da177e4 | 2887 | |
cfb6eeb4 YH |
2888 | #ifdef CONFIG_TCP_MD5SIG |
2889 | if (tp->af_specific->md5_lookup(sk, sk) != NULL) | |
2890 | tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; | |
2891 | #endif | |
2892 | ||
1da177e4 LT |
2893 | /* If user gave his TCP_MAXSEG, record it to clamp */ |
2894 | if (tp->rx_opt.user_mss) | |
2895 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
2896 | tp->max_window = 0; | |
5d424d5a | 2897 | tcp_mtup_init(sk); |
1da177e4 LT |
2898 | tcp_sync_mss(sk, dst_mtu(dst)); |
2899 | ||
2900 | if (!tp->window_clamp) | |
2901 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); | |
0dbaee3b | 2902 | tp->advmss = dst_metric_advmss(dst); |
f5fff5dc TQ |
2903 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) |
2904 | tp->advmss = tp->rx_opt.user_mss; | |
2905 | ||
1da177e4 | 2906 | tcp_initialize_rcv_mss(sk); |
1da177e4 | 2907 | |
e88c64f0 HPP |
2908 | /* limit the window selection if the user enforce a smaller rx buffer */ |
2909 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && | |
2910 | (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) | |
2911 | tp->window_clamp = tcp_full_space(sk); | |
2912 | ||
1da177e4 LT |
2913 | tcp_select_initial_window(tcp_full_space(sk), |
2914 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), | |
2915 | &tp->rcv_wnd, | |
2916 | &tp->window_clamp, | |
bb5b7c11 | 2917 | sysctl_tcp_window_scaling, |
31d12926 | 2918 | &rcv_wscale, |
2919 | dst_metric(dst, RTAX_INITRWND)); | |
1da177e4 LT |
2920 | |
2921 | tp->rx_opt.rcv_wscale = rcv_wscale; | |
2922 | tp->rcv_ssthresh = tp->rcv_wnd; | |
2923 | ||
2924 | sk->sk_err = 0; | |
2925 | sock_reset_flag(sk, SOCK_DONE); | |
2926 | tp->snd_wnd = 0; | |
ee7537b6 | 2927 | tcp_init_wl(tp, 0); |
1da177e4 LT |
2928 | tp->snd_una = tp->write_seq; |
2929 | tp->snd_sml = tp->write_seq; | |
33f5f57e | 2930 | tp->snd_up = tp->write_seq; |
370816ae | 2931 | tp->snd_nxt = tp->write_seq; |
ee995283 PE |
2932 | |
2933 | if (likely(!tp->repair)) | |
2934 | tp->rcv_nxt = 0; | |
c7781a6e AV |
2935 | else |
2936 | tp->rcv_tstamp = tcp_time_stamp; | |
ee995283 PE |
2937 | tp->rcv_wup = tp->rcv_nxt; |
2938 | tp->copied_seq = tp->rcv_nxt; | |
1da177e4 | 2939 | |
463c84b9 ACM |
2940 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
2941 | inet_csk(sk)->icsk_retransmits = 0; | |
1da177e4 LT |
2942 | tcp_clear_retrans(tp); |
2943 | } | |
2944 | ||
783237e8 YC |
2945 | static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) |
2946 | { | |
2947 | struct tcp_sock *tp = tcp_sk(sk); | |
2948 | struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); | |
2949 | ||
2950 | tcb->end_seq += skb->len; | |
2951 | skb_header_release(skb); | |
2952 | __tcp_add_write_queue_tail(sk, skb); | |
2953 | sk->sk_wmem_queued += skb->truesize; | |
2954 | sk_mem_charge(sk, skb->truesize); | |
2955 | tp->write_seq = tcb->end_seq; | |
2956 | tp->packets_out += tcp_skb_pcount(skb); | |
2957 | } | |
2958 | ||
2959 | /* Build and send a SYN with data and (cached) Fast Open cookie. However, | |
2960 | * queue a data-only packet after the regular SYN, such that regular SYNs | |
2961 | * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges | |
2962 | * only the SYN sequence, the data are retransmitted in the first ACK. | |
2963 | * If cookie is not cached or other error occurs, falls back to send a | |
2964 | * regular SYN with Fast Open cookie request option. | |
2965 | */ | |
2966 | static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) | |
2967 | { | |
2968 | struct tcp_sock *tp = tcp_sk(sk); | |
2969 | struct tcp_fastopen_request *fo = tp->fastopen_req; | |
aab48743 | 2970 | int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen; |
783237e8 | 2971 | struct sk_buff *syn_data = NULL, *data; |
aab48743 YC |
2972 | unsigned long last_syn_loss = 0; |
2973 | ||
67da22d2 | 2974 | tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ |
aab48743 YC |
2975 | tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, |
2976 | &syn_loss, &last_syn_loss); | |
2977 | /* Recurring FO SYN losses: revert to regular handshake temporarily */ | |
2978 | if (syn_loss > 1 && | |
2979 | time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { | |
2980 | fo->cookie.len = -1; | |
2981 | goto fallback; | |
2982 | } | |
783237e8 | 2983 | |
67da22d2 YC |
2984 | if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) |
2985 | fo->cookie.len = -1; | |
2986 | else if (fo->cookie.len <= 0) | |
783237e8 YC |
2987 | goto fallback; |
2988 | ||
2989 | /* MSS for SYN-data is based on cached MSS and bounded by PMTU and | |
2990 | * user-MSS. Reserve maximum option space for middleboxes that add | |
2991 | * private TCP options. The cost is reduced data space in SYN :( | |
2992 | */ | |
2993 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) | |
2994 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
1b63edd6 | 2995 | space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - |
783237e8 YC |
2996 | MAX_TCP_OPTION_SPACE; |
2997 | ||
f5ddcbbb ED |
2998 | space = min_t(size_t, space, fo->size); |
2999 | ||
3000 | /* limit to order-0 allocations */ | |
3001 | space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); | |
3002 | ||
3003 | syn_data = skb_copy_expand(syn, MAX_TCP_HEADER, space, | |
783237e8 YC |
3004 | sk->sk_allocation); |
3005 | if (syn_data == NULL) | |
3006 | goto fallback; | |
3007 | ||
3008 | for (i = 0; i < iovlen && syn_data->len < space; ++i) { | |
3009 | struct iovec *iov = &fo->data->msg_iov[i]; | |
3010 | unsigned char __user *from = iov->iov_base; | |
3011 | int len = iov->iov_len; | |
3012 | ||
3013 | if (syn_data->len + len > space) | |
3014 | len = space - syn_data->len; | |
3015 | else if (i + 1 == iovlen) | |
3016 | /* No more data pending in inet_wait_for_connect() */ | |
3017 | fo->data = NULL; | |
3018 | ||
3019 | if (skb_add_data(syn_data, from, len)) | |
3020 | goto fallback; | |
3021 | } | |
3022 | ||
3023 | /* Queue a data-only packet after the regular SYN for retransmission */ | |
3024 | data = pskb_copy(syn_data, sk->sk_allocation); | |
3025 | if (data == NULL) | |
3026 | goto fallback; | |
3027 | TCP_SKB_CB(data)->seq++; | |
3028 | TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN; | |
3029 | TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH); | |
3030 | tcp_connect_queue_skb(sk, data); | |
3031 | fo->copied = data->len; | |
3032 | ||
431a9124 ED |
3033 | /* syn_data is about to be sent, we need to take current time stamps |
3034 | * for the packets that are in write queue : SYN packet and DATA | |
3035 | */ | |
3036 | skb_mstamp_get(&syn->skb_mstamp); | |
3037 | data->skb_mstamp = syn->skb_mstamp; | |
3038 | ||
783237e8 | 3039 | if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) { |
67da22d2 | 3040 | tp->syn_data = (fo->copied > 0); |
f19c29e3 | 3041 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); |
783237e8 YC |
3042 | goto done; |
3043 | } | |
3044 | syn_data = NULL; | |
3045 | ||
3046 | fallback: | |
3047 | /* Send a regular SYN with Fast Open cookie request option */ | |
3048 | if (fo->cookie.len > 0) | |
3049 | fo->cookie.len = 0; | |
3050 | err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); | |
3051 | if (err) | |
3052 | tp->syn_fastopen = 0; | |
3053 | kfree_skb(syn_data); | |
3054 | done: | |
3055 | fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ | |
3056 | return err; | |
3057 | } | |
3058 | ||
67edfef7 | 3059 | /* Build a SYN and send it off. */ |
1da177e4 LT |
3060 | int tcp_connect(struct sock *sk) |
3061 | { | |
3062 | struct tcp_sock *tp = tcp_sk(sk); | |
3063 | struct sk_buff *buff; | |
ee586811 | 3064 | int err; |
1da177e4 LT |
3065 | |
3066 | tcp_connect_init(sk); | |
3067 | ||
2b916477 AV |
3068 | if (unlikely(tp->repair)) { |
3069 | tcp_finish_connect(sk, NULL); | |
3070 | return 0; | |
3071 | } | |
3072 | ||
d179cd12 | 3073 | buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); |
1da177e4 LT |
3074 | if (unlikely(buff == NULL)) |
3075 | return -ENOBUFS; | |
3076 | ||
3077 | /* Reserve space for headers. */ | |
3078 | skb_reserve(buff, MAX_TCP_HEADER); | |
3079 | ||
a3433f35 | 3080 | tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); |
783237e8 YC |
3081 | tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp; |
3082 | tcp_connect_queue_skb(sk, buff); | |
e870a8ef | 3083 | TCP_ECN_send_syn(sk, buff); |
1da177e4 | 3084 | |
783237e8 YC |
3085 | /* Send off SYN; include data in Fast Open. */ |
3086 | err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : | |
3087 | tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); | |
ee586811 EP |
3088 | if (err == -ECONNREFUSED) |
3089 | return err; | |
bd37a088 WY |
3090 | |
3091 | /* We change tp->snd_nxt after the tcp_transmit_skb() call | |
3092 | * in order to make this packet get counted in tcpOutSegs. | |
3093 | */ | |
3094 | tp->snd_nxt = tp->write_seq; | |
3095 | tp->pushed_seq = tp->write_seq; | |
81cc8a75 | 3096 | TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); |
1da177e4 LT |
3097 | |
3098 | /* Timer for repeating the SYN until an answer. */ | |
3f421baa ACM |
3099 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
3100 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
1da177e4 LT |
3101 | return 0; |
3102 | } | |
4bc2f18b | 3103 | EXPORT_SYMBOL(tcp_connect); |
1da177e4 LT |
3104 | |
3105 | /* Send out a delayed ack, the caller does the policy checking | |
3106 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() | |
3107 | * for details. | |
3108 | */ | |
3109 | void tcp_send_delayed_ack(struct sock *sk) | |
3110 | { | |
463c84b9 ACM |
3111 | struct inet_connection_sock *icsk = inet_csk(sk); |
3112 | int ato = icsk->icsk_ack.ato; | |
1da177e4 LT |
3113 | unsigned long timeout; |
3114 | ||
3115 | if (ato > TCP_DELACK_MIN) { | |
463c84b9 | 3116 | const struct tcp_sock *tp = tcp_sk(sk); |
056834d9 | 3117 | int max_ato = HZ / 2; |
1da177e4 | 3118 | |
056834d9 IJ |
3119 | if (icsk->icsk_ack.pingpong || |
3120 | (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) | |
1da177e4 LT |
3121 | max_ato = TCP_DELACK_MAX; |
3122 | ||
3123 | /* Slow path, intersegment interval is "high". */ | |
3124 | ||
3125 | /* If some rtt estimate is known, use it to bound delayed ack. | |
463c84b9 | 3126 | * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements |
1da177e4 LT |
3127 | * directly. |
3128 | */ | |
740b0f18 ED |
3129 | if (tp->srtt_us) { |
3130 | int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), | |
3131 | TCP_DELACK_MIN); | |
1da177e4 LT |
3132 | |
3133 | if (rtt < max_ato) | |
3134 | max_ato = rtt; | |
3135 | } | |
3136 | ||
3137 | ato = min(ato, max_ato); | |
3138 | } | |
3139 | ||
3140 | /* Stay within the limit we were given */ | |
3141 | timeout = jiffies + ato; | |
3142 | ||
3143 | /* Use new timeout only if there wasn't a older one earlier. */ | |
463c84b9 | 3144 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { |
1da177e4 LT |
3145 | /* If delack timer was blocked or is about to expire, |
3146 | * send ACK now. | |
3147 | */ | |
463c84b9 ACM |
3148 | if (icsk->icsk_ack.blocked || |
3149 | time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { | |
1da177e4 LT |
3150 | tcp_send_ack(sk); |
3151 | return; | |
3152 | } | |
3153 | ||
463c84b9 ACM |
3154 | if (!time_before(timeout, icsk->icsk_ack.timeout)) |
3155 | timeout = icsk->icsk_ack.timeout; | |
1da177e4 | 3156 | } |
463c84b9 ACM |
3157 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; |
3158 | icsk->icsk_ack.timeout = timeout; | |
3159 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); | |
1da177e4 LT |
3160 | } |
3161 | ||
3162 | /* This routine sends an ack and also updates the window. */ | |
3163 | void tcp_send_ack(struct sock *sk) | |
3164 | { | |
058dc334 | 3165 | struct sk_buff *buff; |
1da177e4 | 3166 | |
058dc334 IJ |
3167 | /* If we have been reset, we may not send again. */ |
3168 | if (sk->sk_state == TCP_CLOSE) | |
3169 | return; | |
1da177e4 | 3170 | |
058dc334 IJ |
3171 | /* We are not putting this on the write queue, so |
3172 | * tcp_transmit_skb() will set the ownership to this | |
3173 | * sock. | |
3174 | */ | |
99a1dec7 | 3175 | buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); |
058dc334 IJ |
3176 | if (buff == NULL) { |
3177 | inet_csk_schedule_ack(sk); | |
3178 | inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; | |
3179 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, | |
3180 | TCP_DELACK_MAX, TCP_RTO_MAX); | |
3181 | return; | |
1da177e4 | 3182 | } |
058dc334 IJ |
3183 | |
3184 | /* Reserve space for headers and prepare control bits. */ | |
3185 | skb_reserve(buff, MAX_TCP_HEADER); | |
a3433f35 | 3186 | tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); |
058dc334 IJ |
3187 | |
3188 | /* Send it off, this clears delayed acks for us. */ | |
058dc334 | 3189 | TCP_SKB_CB(buff)->when = tcp_time_stamp; |
99a1dec7 | 3190 | tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC)); |
1da177e4 LT |
3191 | } |
3192 | ||
3193 | /* This routine sends a packet with an out of date sequence | |
3194 | * number. It assumes the other end will try to ack it. | |
3195 | * | |
3196 | * Question: what should we make while urgent mode? | |
3197 | * 4.4BSD forces sending single byte of data. We cannot send | |
3198 | * out of window data, because we have SND.NXT==SND.MAX... | |
3199 | * | |
3200 | * Current solution: to send TWO zero-length segments in urgent mode: | |
3201 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is | |
3202 | * out-of-date with SND.UNA-1 to probe window. | |
3203 | */ | |
3204 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent) | |
3205 | { | |
3206 | struct tcp_sock *tp = tcp_sk(sk); | |
3207 | struct sk_buff *skb; | |
3208 | ||
3209 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ | |
99a1dec7 | 3210 | skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); |
e905a9ed | 3211 | if (skb == NULL) |
1da177e4 LT |
3212 | return -1; |
3213 | ||
3214 | /* Reserve space for headers and set control bits. */ | |
3215 | skb_reserve(skb, MAX_TCP_HEADER); | |
1da177e4 LT |
3216 | /* Use a previous sequence. This should cause the other |
3217 | * end to send an ack. Don't queue or clone SKB, just | |
3218 | * send it. | |
3219 | */ | |
a3433f35 | 3220 | tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); |
1da177e4 | 3221 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
dfb4b9dc | 3222 | return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); |
1da177e4 LT |
3223 | } |
3224 | ||
ee995283 PE |
3225 | void tcp_send_window_probe(struct sock *sk) |
3226 | { | |
3227 | if (sk->sk_state == TCP_ESTABLISHED) { | |
3228 | tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; | |
3229 | tcp_xmit_probe_skb(sk, 0); | |
3230 | } | |
3231 | } | |
3232 | ||
67edfef7 | 3233 | /* Initiate keepalive or window probe from timer. */ |
1da177e4 LT |
3234 | int tcp_write_wakeup(struct sock *sk) |
3235 | { | |
058dc334 IJ |
3236 | struct tcp_sock *tp = tcp_sk(sk); |
3237 | struct sk_buff *skb; | |
1da177e4 | 3238 | |
058dc334 IJ |
3239 | if (sk->sk_state == TCP_CLOSE) |
3240 | return -1; | |
3241 | ||
3242 | if ((skb = tcp_send_head(sk)) != NULL && | |
3243 | before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { | |
3244 | int err; | |
0c54b85f | 3245 | unsigned int mss = tcp_current_mss(sk); |
058dc334 IJ |
3246 | unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
3247 | ||
3248 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) | |
3249 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; | |
3250 | ||
3251 | /* We are probing the opening of a window | |
3252 | * but the window size is != 0 | |
3253 | * must have been a result SWS avoidance ( sender ) | |
3254 | */ | |
3255 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || | |
3256 | skb->len > mss) { | |
3257 | seg_size = min(seg_size, mss); | |
4de075e0 | 3258 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
058dc334 IJ |
3259 | if (tcp_fragment(sk, skb, seg_size, mss)) |
3260 | return -1; | |
3261 | } else if (!tcp_skb_pcount(skb)) | |
3262 | tcp_set_skb_tso_segs(sk, skb, mss); | |
3263 | ||
4de075e0 | 3264 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
058dc334 IJ |
3265 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
3266 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); | |
3267 | if (!err) | |
3268 | tcp_event_new_data_sent(sk, skb); | |
3269 | return err; | |
3270 | } else { | |
33f5f57e | 3271 | if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) |
058dc334 IJ |
3272 | tcp_xmit_probe_skb(sk, 1); |
3273 | return tcp_xmit_probe_skb(sk, 0); | |
1da177e4 | 3274 | } |
1da177e4 LT |
3275 | } |
3276 | ||
3277 | /* A window probe timeout has occurred. If window is not closed send | |
3278 | * a partial packet else a zero probe. | |
3279 | */ | |
3280 | void tcp_send_probe0(struct sock *sk) | |
3281 | { | |
463c84b9 | 3282 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
3283 | struct tcp_sock *tp = tcp_sk(sk); |
3284 | int err; | |
3285 | ||
3286 | err = tcp_write_wakeup(sk); | |
3287 | ||
fe067e8a | 3288 | if (tp->packets_out || !tcp_send_head(sk)) { |
1da177e4 | 3289 | /* Cancel probe timer, if it is not required. */ |
6687e988 | 3290 | icsk->icsk_probes_out = 0; |
463c84b9 | 3291 | icsk->icsk_backoff = 0; |
1da177e4 LT |
3292 | return; |
3293 | } | |
3294 | ||
3295 | if (err <= 0) { | |
463c84b9 ACM |
3296 | if (icsk->icsk_backoff < sysctl_tcp_retries2) |
3297 | icsk->icsk_backoff++; | |
6687e988 | 3298 | icsk->icsk_probes_out++; |
e905a9ed | 3299 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
3f421baa ACM |
3300 | min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), |
3301 | TCP_RTO_MAX); | |
1da177e4 LT |
3302 | } else { |
3303 | /* If packet was not sent due to local congestion, | |
6687e988 | 3304 | * do not backoff and do not remember icsk_probes_out. |
1da177e4 LT |
3305 | * Let local senders to fight for local resources. |
3306 | * | |
3307 | * Use accumulated backoff yet. | |
3308 | */ | |
6687e988 ACM |
3309 | if (!icsk->icsk_probes_out) |
3310 | icsk->icsk_probes_out = 1; | |
e905a9ed | 3311 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
463c84b9 | 3312 | min(icsk->icsk_rto << icsk->icsk_backoff, |
3f421baa ACM |
3313 | TCP_RESOURCE_PROBE_INTERVAL), |
3314 | TCP_RTO_MAX); | |
1da177e4 LT |
3315 | } |
3316 | } |