Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * INET An implementation of the TCP/IP protocol suite for the LINUX | |
3 | * operating system. INET is implemented using the BSD Socket | |
4 | * interface as the means of communication with the user level. | |
5 | * | |
6 | * Implementation of the Transmission Control Protocol(TCP). | |
7 | * | |
02c30a84 | 8 | * Authors: Ross Biro |
1da177e4 LT |
9 | * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> |
10 | * Mark Evans, <evansmp@uhura.aston.ac.uk> | |
11 | * Corey Minyard <wf-rch!minyard@relay.EU.net> | |
12 | * Florian La Roche, <flla@stud.uni-sb.de> | |
13 | * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> | |
14 | * Linus Torvalds, <torvalds@cs.helsinki.fi> | |
15 | * Alan Cox, <gw4pts@gw4pts.ampr.org> | |
16 | * Matthew Dillon, <dillon@apollo.west.oic.com> | |
17 | * Arnt Gulbrandsen, <agulbra@nvg.unit.no> | |
18 | * Jorge Cwik, <jorge@laser.satlink.net> | |
19 | */ | |
20 | ||
21 | /* | |
22 | * Changes: Pedro Roque : Retransmit queue handled by TCP. | |
23 | * : Fragmentation on mtu decrease | |
24 | * : Segment collapse on retransmit | |
25 | * : AF independence | |
26 | * | |
27 | * Linus Torvalds : send_delayed_ack | |
28 | * David S. Miller : Charge memory using the right skb | |
29 | * during syn/ack processing. | |
30 | * David S. Miller : Output engine completely rewritten. | |
31 | * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. | |
32 | * Cacophonix Gaul : draft-minshall-nagle-01 | |
33 | * J Hadi Salim : ECN support | |
34 | * | |
35 | */ | |
36 | ||
91df42be JP |
37 | #define pr_fmt(fmt) "TCP: " fmt |
38 | ||
1da177e4 LT |
39 | #include <net/tcp.h> |
40 | ||
41 | #include <linux/compiler.h> | |
5a0e3ad6 | 42 | #include <linux/gfp.h> |
1da177e4 | 43 | #include <linux/module.h> |
1da177e4 LT |
44 | |
45 | /* People can turn this off for buggy TCP's found in printers etc. */ | |
ab32ea5d | 46 | int sysctl_tcp_retrans_collapse __read_mostly = 1; |
1da177e4 | 47 | |
09cb105e | 48 | /* People can turn this on to work with those rare, broken TCPs that |
15d99e02 RJ |
49 | * interpret the window field as a signed quantity. |
50 | */ | |
ab32ea5d | 51 | int sysctl_tcp_workaround_signed_windows __read_mostly = 0; |
15d99e02 | 52 | |
46d3ceab ED |
53 | /* Default TSQ limit of two TSO segments */ |
54 | int sysctl_tcp_limit_output_bytes __read_mostly = 131072; | |
55 | ||
1da177e4 LT |
56 | /* This limits the percentage of the congestion window which we |
57 | * will allow a single TSO frame to consume. Building TSO frames | |
58 | * which are too large can cause TCP streams to be bursty. | |
59 | */ | |
ab32ea5d | 60 | int sysctl_tcp_tso_win_divisor __read_mostly = 3; |
1da177e4 | 61 | |
ab32ea5d | 62 | int sysctl_tcp_mtu_probing __read_mostly = 0; |
97b1ce25 | 63 | int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS; |
5d424d5a | 64 | |
35089bb2 | 65 | /* By default, RFC2861 behavior. */ |
ab32ea5d | 66 | int sysctl_tcp_slow_start_after_idle __read_mostly = 1; |
35089bb2 | 67 | |
46d3ceab ED |
68 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
69 | int push_one, gfp_t gfp); | |
519855c5 | 70 | |
67edfef7 | 71 | /* Account for new data that has been sent to the network. */ |
cf533ea5 | 72 | static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 73 | { |
6ba8a3b1 | 74 | struct inet_connection_sock *icsk = inet_csk(sk); |
9e412ba7 | 75 | struct tcp_sock *tp = tcp_sk(sk); |
66f5fe62 | 76 | unsigned int prior_packets = tp->packets_out; |
9e412ba7 | 77 | |
fe067e8a | 78 | tcp_advance_send_head(sk, skb); |
1da177e4 | 79 | tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; |
8512430e | 80 | |
66f5fe62 | 81 | tp->packets_out += tcp_skb_pcount(skb); |
6ba8a3b1 | 82 | if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || |
6a5dc9e5 | 83 | icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { |
750ea2ba | 84 | tcp_rearm_rto(sk); |
6a5dc9e5 | 85 | } |
1da177e4 LT |
86 | } |
87 | ||
88 | /* SND.NXT, if window was not shrunk. | |
89 | * If window has been shrunk, what should we make? It is not clear at all. | |
90 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( | |
91 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already | |
92 | * invalid. OK, let's make this for now: | |
93 | */ | |
cf533ea5 | 94 | static inline __u32 tcp_acceptable_seq(const struct sock *sk) |
1da177e4 | 95 | { |
cf533ea5 | 96 | const struct tcp_sock *tp = tcp_sk(sk); |
9e412ba7 | 97 | |
90840def | 98 | if (!before(tcp_wnd_end(tp), tp->snd_nxt)) |
1da177e4 LT |
99 | return tp->snd_nxt; |
100 | else | |
90840def | 101 | return tcp_wnd_end(tp); |
1da177e4 LT |
102 | } |
103 | ||
104 | /* Calculate mss to advertise in SYN segment. | |
105 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: | |
106 | * | |
107 | * 1. It is independent of path mtu. | |
108 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. | |
109 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of | |
110 | * attached devices, because some buggy hosts are confused by | |
111 | * large MSS. | |
112 | * 4. We do not make 3, we advertise MSS, calculated from first | |
113 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. | |
114 | * This may be overridden via information stored in routing table. | |
115 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, | |
116 | * probably even Jumbo". | |
117 | */ | |
118 | static __u16 tcp_advertise_mss(struct sock *sk) | |
119 | { | |
120 | struct tcp_sock *tp = tcp_sk(sk); | |
cf533ea5 | 121 | const struct dst_entry *dst = __sk_dst_get(sk); |
1da177e4 LT |
122 | int mss = tp->advmss; |
123 | ||
0dbaee3b DM |
124 | if (dst) { |
125 | unsigned int metric = dst_metric_advmss(dst); | |
126 | ||
127 | if (metric < mss) { | |
128 | mss = metric; | |
129 | tp->advmss = mss; | |
130 | } | |
1da177e4 LT |
131 | } |
132 | ||
133 | return (__u16)mss; | |
134 | } | |
135 | ||
136 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". | |
137 | * This is the first part of cwnd validation mechanism. */ | |
cf533ea5 | 138 | static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst) |
1da177e4 | 139 | { |
463c84b9 | 140 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
141 | s32 delta = tcp_time_stamp - tp->lsndtime; |
142 | u32 restart_cwnd = tcp_init_cwnd(tp, dst); | |
143 | u32 cwnd = tp->snd_cwnd; | |
144 | ||
6687e988 | 145 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); |
1da177e4 | 146 | |
6687e988 | 147 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
1da177e4 LT |
148 | restart_cwnd = min(restart_cwnd, cwnd); |
149 | ||
463c84b9 | 150 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) |
1da177e4 LT |
151 | cwnd >>= 1; |
152 | tp->snd_cwnd = max(cwnd, restart_cwnd); | |
153 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
154 | tp->snd_cwnd_used = 0; | |
155 | } | |
156 | ||
67edfef7 | 157 | /* Congestion state accounting after a packet has been sent. */ |
40efc6fa | 158 | static void tcp_event_data_sent(struct tcp_sock *tp, |
cf533ea5 | 159 | struct sock *sk) |
1da177e4 | 160 | { |
463c84b9 ACM |
161 | struct inet_connection_sock *icsk = inet_csk(sk); |
162 | const u32 now = tcp_time_stamp; | |
1da177e4 | 163 | |
35089bb2 DM |
164 | if (sysctl_tcp_slow_start_after_idle && |
165 | (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) | |
463c84b9 | 166 | tcp_cwnd_restart(sk, __sk_dst_get(sk)); |
1da177e4 LT |
167 | |
168 | tp->lsndtime = now; | |
169 | ||
170 | /* If it is a reply for ato after last received | |
171 | * packet, enter pingpong mode. | |
172 | */ | |
463c84b9 ACM |
173 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) |
174 | icsk->icsk_ack.pingpong = 1; | |
1da177e4 LT |
175 | } |
176 | ||
67edfef7 | 177 | /* Account for an ACK we sent. */ |
40efc6fa | 178 | static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) |
1da177e4 | 179 | { |
463c84b9 ACM |
180 | tcp_dec_quickack_mode(sk, pkts); |
181 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); | |
1da177e4 LT |
182 | } |
183 | ||
184 | /* Determine a window scaling and initial window to offer. | |
185 | * Based on the assumption that the given amount of space | |
186 | * will be offered. Store the results in the tp structure. | |
187 | * NOTE: for smooth operation initial space offering should | |
188 | * be a multiple of mss if possible. We assume here that mss >= 1. | |
189 | * This MUST be enforced by all callers. | |
190 | */ | |
191 | void tcp_select_initial_window(int __space, __u32 mss, | |
192 | __u32 *rcv_wnd, __u32 *window_clamp, | |
31d12926 | 193 | int wscale_ok, __u8 *rcv_wscale, |
194 | __u32 init_rcv_wnd) | |
1da177e4 LT |
195 | { |
196 | unsigned int space = (__space < 0 ? 0 : __space); | |
197 | ||
198 | /* If no clamp set the clamp to the max possible scaled window */ | |
199 | if (*window_clamp == 0) | |
200 | (*window_clamp) = (65535 << 14); | |
201 | space = min(*window_clamp, space); | |
202 | ||
203 | /* Quantize space offering to a multiple of mss if possible. */ | |
204 | if (space > mss) | |
205 | space = (space / mss) * mss; | |
206 | ||
207 | /* NOTE: offering an initial window larger than 32767 | |
15d99e02 RJ |
208 | * will break some buggy TCP stacks. If the admin tells us |
209 | * it is likely we could be speaking with such a buggy stack | |
210 | * we will truncate our initial window offering to 32K-1 | |
211 | * unless the remote has sent us a window scaling option, | |
212 | * which we interpret as a sign the remote TCP is not | |
213 | * misinterpreting the window field as a signed quantity. | |
1da177e4 | 214 | */ |
15d99e02 RJ |
215 | if (sysctl_tcp_workaround_signed_windows) |
216 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); | |
217 | else | |
218 | (*rcv_wnd) = space; | |
219 | ||
1da177e4 LT |
220 | (*rcv_wscale) = 0; |
221 | if (wscale_ok) { | |
222 | /* Set window scaling on max possible window | |
e905a9ed | 223 | * See RFC1323 for an explanation of the limit to 14 |
1da177e4 | 224 | */ |
aadf1c47 SHY |
225 | space = max_t(u32, space, sysctl_tcp_rmem[2]); |
226 | space = max_t(u32, space, sysctl_rmem_max); | |
316c1592 | 227 | space = min_t(u32, space, *window_clamp); |
1da177e4 LT |
228 | while (space > 65535 && (*rcv_wscale) < 14) { |
229 | space >>= 1; | |
230 | (*rcv_wscale)++; | |
231 | } | |
232 | } | |
233 | ||
356f0398 | 234 | /* Set initial window to a value enough for senders starting with |
6fa3eb70 | 235 | * initial congestion window of sysctl_tcp_default_init_rwnd. Place |
356f0398 ND |
236 | * a limit on the initial window when mss is larger than 1460. |
237 | */ | |
056834d9 | 238 | if (mss > (1 << *rcv_wscale)) { |
6fa3eb70 | 239 | int init_cwnd = sysctl_tcp_default_init_rwnd; |
356f0398 | 240 | if (mss > 1460) |
6fa3eb70 | 241 | init_cwnd = max_t(u32, (1460 * init_cwnd) / mss, 2); |
31d12926 | 242 | /* when initializing use the value from init_rcv_wnd |
243 | * rather than the default from above | |
244 | */ | |
b1afde60 ND |
245 | if (init_rcv_wnd) |
246 | *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); | |
247 | else | |
248 | *rcv_wnd = min(*rcv_wnd, init_cwnd * mss); | |
1da177e4 LT |
249 | } |
250 | ||
251 | /* Set the clamp no higher than max representable value */ | |
252 | (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); | |
253 | } | |
4bc2f18b | 254 | EXPORT_SYMBOL(tcp_select_initial_window); |
1da177e4 LT |
255 | |
256 | /* Chose a new window to advertise, update state in tcp_sock for the | |
257 | * socket, and return result with RFC1323 scaling applied. The return | |
258 | * value can be stuffed directly into th->window for an outgoing | |
259 | * frame. | |
260 | */ | |
40efc6fa | 261 | static u16 tcp_select_window(struct sock *sk) |
1da177e4 LT |
262 | { |
263 | struct tcp_sock *tp = tcp_sk(sk); | |
264 | u32 cur_win = tcp_receive_window(tp); | |
265 | u32 new_win = __tcp_select_window(sk); | |
266 | ||
267 | /* Never shrink the offered window */ | |
2de979bd | 268 | if (new_win < cur_win) { |
1da177e4 LT |
269 | /* Danger Will Robinson! |
270 | * Don't update rcv_wup/rcv_wnd here or else | |
271 | * we will not be able to advertise a zero | |
272 | * window in time. --DaveM | |
273 | * | |
274 | * Relax Will Robinson. | |
275 | */ | |
607bfbf2 | 276 | new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); |
1da177e4 LT |
277 | } |
278 | tp->rcv_wnd = new_win; | |
279 | tp->rcv_wup = tp->rcv_nxt; | |
280 | ||
281 | /* Make sure we do not exceed the maximum possible | |
282 | * scaled window. | |
283 | */ | |
15d99e02 | 284 | if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) |
1da177e4 LT |
285 | new_win = min(new_win, MAX_TCP_WINDOW); |
286 | else | |
287 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); | |
288 | ||
289 | /* RFC1323 scaling applied */ | |
290 | new_win >>= tp->rx_opt.rcv_wscale; | |
291 | ||
292 | /* If we advertise zero window, disable fast path. */ | |
293 | if (new_win == 0) | |
294 | tp->pred_flags = 0; | |
295 | ||
296 | return new_win; | |
297 | } | |
298 | ||
67edfef7 | 299 | /* Packet ECN state for a SYN-ACK */ |
cf533ea5 | 300 | static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb) |
bdf1ee5d | 301 | { |
4de075e0 | 302 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; |
056834d9 | 303 | if (!(tp->ecn_flags & TCP_ECN_OK)) |
4de075e0 | 304 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; |
bdf1ee5d IJ |
305 | } |
306 | ||
67edfef7 | 307 | /* Packet ECN state for a SYN. */ |
bdf1ee5d IJ |
308 | static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) |
309 | { | |
310 | struct tcp_sock *tp = tcp_sk(sk); | |
311 | ||
312 | tp->ecn_flags = 0; | |
5d134f1c | 313 | if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) { |
4de075e0 | 314 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; |
bdf1ee5d IJ |
315 | tp->ecn_flags = TCP_ECN_OK; |
316 | } | |
317 | } | |
318 | ||
319 | static __inline__ void | |
cf533ea5 | 320 | TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th) |
bdf1ee5d IJ |
321 | { |
322 | if (inet_rsk(req)->ecn_ok) | |
323 | th->ece = 1; | |
324 | } | |
325 | ||
67edfef7 AK |
326 | /* Set up ECN state for a packet on a ESTABLISHED socket that is about to |
327 | * be sent. | |
328 | */ | |
bdf1ee5d IJ |
329 | static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, |
330 | int tcp_header_len) | |
331 | { | |
332 | struct tcp_sock *tp = tcp_sk(sk); | |
333 | ||
334 | if (tp->ecn_flags & TCP_ECN_OK) { | |
335 | /* Not-retransmitted data segment: set ECT and inject CWR. */ | |
336 | if (skb->len != tcp_header_len && | |
337 | !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { | |
338 | INET_ECN_xmit(sk); | |
056834d9 | 339 | if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { |
bdf1ee5d IJ |
340 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; |
341 | tcp_hdr(skb)->cwr = 1; | |
342 | skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; | |
343 | } | |
344 | } else { | |
345 | /* ACK or retransmitted segment: clear ECT|CE */ | |
346 | INET_ECN_dontxmit(sk); | |
347 | } | |
348 | if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) | |
349 | tcp_hdr(skb)->ece = 1; | |
350 | } | |
351 | } | |
352 | ||
e870a8ef IJ |
353 | /* Constructs common control bits of non-data skb. If SYN/FIN is present, |
354 | * auto increment end seqno. | |
355 | */ | |
356 | static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) | |
357 | { | |
2e8e18ef | 358 | skb->ip_summed = CHECKSUM_PARTIAL; |
e870a8ef IJ |
359 | skb->csum = 0; |
360 | ||
4de075e0 | 361 | TCP_SKB_CB(skb)->tcp_flags = flags; |
e870a8ef IJ |
362 | TCP_SKB_CB(skb)->sacked = 0; |
363 | ||
364 | skb_shinfo(skb)->gso_segs = 1; | |
365 | skb_shinfo(skb)->gso_size = 0; | |
366 | skb_shinfo(skb)->gso_type = 0; | |
367 | ||
368 | TCP_SKB_CB(skb)->seq = seq; | |
a3433f35 | 369 | if (flags & (TCPHDR_SYN | TCPHDR_FIN)) |
e870a8ef IJ |
370 | seq++; |
371 | TCP_SKB_CB(skb)->end_seq = seq; | |
372 | } | |
373 | ||
a2a385d6 | 374 | static inline bool tcp_urg_mode(const struct tcp_sock *tp) |
33f5f57e IJ |
375 | { |
376 | return tp->snd_una != tp->snd_up; | |
377 | } | |
378 | ||
33ad798c AL |
379 | #define OPTION_SACK_ADVERTISE (1 << 0) |
380 | #define OPTION_TS (1 << 1) | |
381 | #define OPTION_MD5 (1 << 2) | |
89e95a61 | 382 | #define OPTION_WSCALE (1 << 3) |
2100c8d2 | 383 | #define OPTION_FAST_OPEN_COOKIE (1 << 8) |
33ad798c AL |
384 | |
385 | struct tcp_out_options { | |
2100c8d2 YC |
386 | u16 options; /* bit field of OPTION_* */ |
387 | u16 mss; /* 0 to disable */ | |
33ad798c AL |
388 | u8 ws; /* window scale, 0 to disable */ |
389 | u8 num_sack_blocks; /* number of SACK blocks to include */ | |
bd0388ae | 390 | u8 hash_size; /* bytes in hash_location */ |
bd0388ae | 391 | __u8 *hash_location; /* temporary pointer, overloaded */ |
2100c8d2 YC |
392 | __u32 tsval, tsecr; /* need to include OPTION_TS */ |
393 | struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ | |
33ad798c AL |
394 | }; |
395 | ||
67edfef7 AK |
396 | /* Write previously computed TCP options to the packet. |
397 | * | |
398 | * Beware: Something in the Internet is very sensitive to the ordering of | |
fd6149d3 IJ |
399 | * TCP options, we learned this through the hard way, so be careful here. |
400 | * Luckily we can at least blame others for their non-compliance but from | |
401 | * inter-operatibility perspective it seems that we're somewhat stuck with | |
402 | * the ordering which we have been using if we want to keep working with | |
403 | * those broken things (not that it currently hurts anybody as there isn't | |
404 | * particular reason why the ordering would need to be changed). | |
405 | * | |
406 | * At least SACK_PERM as the first option is known to lead to a disaster | |
407 | * (but it may well be that other scenarios fail similarly). | |
408 | */ | |
33ad798c | 409 | static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, |
bd0388ae WAS |
410 | struct tcp_out_options *opts) |
411 | { | |
2100c8d2 | 412 | u16 options = opts->options; /* mungable copy */ |
bd0388ae | 413 | |
bd0388ae | 414 | if (unlikely(OPTION_MD5 & options)) { |
1a2c6181 CP |
415 | *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | |
416 | (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); | |
bd0388ae WAS |
417 | /* overload cookie hash location */ |
418 | opts->hash_location = (__u8 *)ptr; | |
33ad798c | 419 | ptr += 4; |
40efc6fa | 420 | } |
33ad798c | 421 | |
fd6149d3 IJ |
422 | if (unlikely(opts->mss)) { |
423 | *ptr++ = htonl((TCPOPT_MSS << 24) | | |
424 | (TCPOLEN_MSS << 16) | | |
425 | opts->mss); | |
426 | } | |
427 | ||
bd0388ae WAS |
428 | if (likely(OPTION_TS & options)) { |
429 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { | |
33ad798c AL |
430 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
431 | (TCPOLEN_SACK_PERM << 16) | | |
432 | (TCPOPT_TIMESTAMP << 8) | | |
433 | TCPOLEN_TIMESTAMP); | |
bd0388ae | 434 | options &= ~OPTION_SACK_ADVERTISE; |
33ad798c AL |
435 | } else { |
436 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
437 | (TCPOPT_NOP << 16) | | |
438 | (TCPOPT_TIMESTAMP << 8) | | |
439 | TCPOLEN_TIMESTAMP); | |
440 | } | |
441 | *ptr++ = htonl(opts->tsval); | |
442 | *ptr++ = htonl(opts->tsecr); | |
443 | } | |
444 | ||
bd0388ae | 445 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { |
33ad798c AL |
446 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
447 | (TCPOPT_NOP << 16) | | |
448 | (TCPOPT_SACK_PERM << 8) | | |
449 | TCPOLEN_SACK_PERM); | |
450 | } | |
451 | ||
bd0388ae | 452 | if (unlikely(OPTION_WSCALE & options)) { |
33ad798c AL |
453 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
454 | (TCPOPT_WINDOW << 16) | | |
455 | (TCPOLEN_WINDOW << 8) | | |
456 | opts->ws); | |
457 | } | |
458 | ||
459 | if (unlikely(opts->num_sack_blocks)) { | |
460 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? | |
461 | tp->duplicate_sack : tp->selective_acks; | |
40efc6fa SH |
462 | int this_sack; |
463 | ||
464 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
465 | (TCPOPT_NOP << 16) | | |
466 | (TCPOPT_SACK << 8) | | |
33ad798c | 467 | (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * |
40efc6fa | 468 | TCPOLEN_SACK_PERBLOCK))); |
2de979bd | 469 | |
33ad798c AL |
470 | for (this_sack = 0; this_sack < opts->num_sack_blocks; |
471 | ++this_sack) { | |
40efc6fa SH |
472 | *ptr++ = htonl(sp[this_sack].start_seq); |
473 | *ptr++ = htonl(sp[this_sack].end_seq); | |
474 | } | |
2de979bd | 475 | |
5861f8e5 | 476 | tp->rx_opt.dsack = 0; |
40efc6fa | 477 | } |
2100c8d2 YC |
478 | |
479 | if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { | |
480 | struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; | |
481 | ||
482 | *ptr++ = htonl((TCPOPT_EXP << 24) | | |
483 | ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) | | |
484 | TCPOPT_FASTOPEN_MAGIC); | |
485 | ||
486 | memcpy(ptr, foc->val, foc->len); | |
487 | if ((foc->len & 3) == 2) { | |
488 | u8 *align = ((u8 *)ptr) + foc->len; | |
489 | align[0] = align[1] = TCPOPT_NOP; | |
490 | } | |
491 | ptr += (foc->len + 3) >> 2; | |
492 | } | |
33ad798c AL |
493 | } |
494 | ||
67edfef7 AK |
495 | /* Compute TCP options for SYN packets. This is not the final |
496 | * network wire format yet. | |
497 | */ | |
95c96174 | 498 | static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, |
33ad798c | 499 | struct tcp_out_options *opts, |
cf533ea5 ED |
500 | struct tcp_md5sig_key **md5) |
501 | { | |
33ad798c | 502 | struct tcp_sock *tp = tcp_sk(sk); |
95c96174 | 503 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
783237e8 | 504 | struct tcp_fastopen_request *fastopen = tp->fastopen_req; |
33ad798c | 505 | |
cfb6eeb4 | 506 | #ifdef CONFIG_TCP_MD5SIG |
33ad798c AL |
507 | *md5 = tp->af_specific->md5_lookup(sk, sk); |
508 | if (*md5) { | |
509 | opts->options |= OPTION_MD5; | |
bd0388ae | 510 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
cfb6eeb4 | 511 | } |
33ad798c AL |
512 | #else |
513 | *md5 = NULL; | |
cfb6eeb4 | 514 | #endif |
33ad798c AL |
515 | |
516 | /* We always get an MSS option. The option bytes which will be seen in | |
517 | * normal data packets should timestamps be used, must be in the MSS | |
518 | * advertised. But we subtract them from tp->mss_cache so that | |
519 | * calculations in tcp_sendmsg are simpler etc. So account for this | |
520 | * fact here if necessary. If we don't do this correctly, as a | |
521 | * receiver we won't recognize data packets as being full sized when we | |
522 | * should, and thus we won't abide by the delayed ACK rules correctly. | |
523 | * SACKs don't matter, we never delay an ACK when we have any of those | |
524 | * going out. */ | |
525 | opts->mss = tcp_advertise_mss(sk); | |
bd0388ae | 526 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c | 527 | |
bb5b7c11 | 528 | if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { |
33ad798c | 529 | opts->options |= OPTION_TS; |
ee684b6f | 530 | opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset; |
33ad798c | 531 | opts->tsecr = tp->rx_opt.ts_recent; |
bd0388ae | 532 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c | 533 | } |
bb5b7c11 | 534 | if (likely(sysctl_tcp_window_scaling)) { |
33ad798c | 535 | opts->ws = tp->rx_opt.rcv_wscale; |
89e95a61 | 536 | opts->options |= OPTION_WSCALE; |
bd0388ae | 537 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 538 | } |
bb5b7c11 | 539 | if (likely(sysctl_tcp_sack)) { |
33ad798c | 540 | opts->options |= OPTION_SACK_ADVERTISE; |
b32d1310 | 541 | if (unlikely(!(OPTION_TS & opts->options))) |
bd0388ae | 542 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c AL |
543 | } |
544 | ||
783237e8 YC |
545 | if (fastopen && fastopen->cookie.len >= 0) { |
546 | u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len; | |
547 | need = (need + 3) & ~3U; /* Align to 32 bits */ | |
548 | if (remaining >= need) { | |
549 | opts->options |= OPTION_FAST_OPEN_COOKIE; | |
550 | opts->fastopen_cookie = &fastopen->cookie; | |
551 | remaining -= need; | |
552 | tp->syn_fastopen = 1; | |
553 | } | |
554 | } | |
bd0388ae | 555 | |
bd0388ae | 556 | return MAX_TCP_OPTION_SPACE - remaining; |
40efc6fa SH |
557 | } |
558 | ||
67edfef7 | 559 | /* Set up TCP options for SYN-ACKs. */ |
95c96174 | 560 | static unsigned int tcp_synack_options(struct sock *sk, |
33ad798c | 561 | struct request_sock *req, |
95c96174 | 562 | unsigned int mss, struct sk_buff *skb, |
33ad798c | 563 | struct tcp_out_options *opts, |
4957faad | 564 | struct tcp_md5sig_key **md5, |
8336886f | 565 | struct tcp_fastopen_cookie *foc) |
4957faad | 566 | { |
33ad798c | 567 | struct inet_request_sock *ireq = inet_rsk(req); |
95c96174 | 568 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
33ad798c | 569 | |
cfb6eeb4 | 570 | #ifdef CONFIG_TCP_MD5SIG |
33ad798c AL |
571 | *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); |
572 | if (*md5) { | |
573 | opts->options |= OPTION_MD5; | |
4957faad WAS |
574 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
575 | ||
576 | /* We can't fit any SACK blocks in a packet with MD5 + TS | |
577 | * options. There was discussion about disabling SACK | |
578 | * rather than TS in order to fit in better with old, | |
579 | * buggy kernels, but that was deemed to be unnecessary. | |
580 | */ | |
de213e5e | 581 | ireq->tstamp_ok &= !ireq->sack_ok; |
cfb6eeb4 | 582 | } |
33ad798c AL |
583 | #else |
584 | *md5 = NULL; | |
cfb6eeb4 | 585 | #endif |
33ad798c | 586 | |
4957faad | 587 | /* We always send an MSS option. */ |
33ad798c | 588 | opts->mss = mss; |
4957faad | 589 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c AL |
590 | |
591 | if (likely(ireq->wscale_ok)) { | |
592 | opts->ws = ireq->rcv_wscale; | |
89e95a61 | 593 | opts->options |= OPTION_WSCALE; |
4957faad | 594 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 595 | } |
de213e5e | 596 | if (likely(ireq->tstamp_ok)) { |
33ad798c AL |
597 | opts->options |= OPTION_TS; |
598 | opts->tsval = TCP_SKB_CB(skb)->when; | |
599 | opts->tsecr = req->ts_recent; | |
4957faad | 600 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c AL |
601 | } |
602 | if (likely(ireq->sack_ok)) { | |
603 | opts->options |= OPTION_SACK_ADVERTISE; | |
de213e5e | 604 | if (unlikely(!ireq->tstamp_ok)) |
4957faad | 605 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c | 606 | } |
8336886f JC |
607 | if (foc != NULL) { |
608 | u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; | |
609 | need = (need + 3) & ~3U; /* Align to 32 bits */ | |
610 | if (remaining >= need) { | |
611 | opts->options |= OPTION_FAST_OPEN_COOKIE; | |
612 | opts->fastopen_cookie = foc; | |
613 | remaining -= need; | |
614 | } | |
615 | } | |
1a2c6181 | 616 | |
4957faad | 617 | return MAX_TCP_OPTION_SPACE - remaining; |
33ad798c AL |
618 | } |
619 | ||
67edfef7 AK |
620 | /* Compute TCP options for ESTABLISHED sockets. This is not the |
621 | * final wire format yet. | |
622 | */ | |
95c96174 | 623 | static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, |
33ad798c | 624 | struct tcp_out_options *opts, |
cf533ea5 ED |
625 | struct tcp_md5sig_key **md5) |
626 | { | |
33ad798c AL |
627 | struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; |
628 | struct tcp_sock *tp = tcp_sk(sk); | |
95c96174 | 629 | unsigned int size = 0; |
cabeccbd | 630 | unsigned int eff_sacks; |
33ad798c AL |
631 | |
632 | #ifdef CONFIG_TCP_MD5SIG | |
633 | *md5 = tp->af_specific->md5_lookup(sk, sk); | |
634 | if (unlikely(*md5)) { | |
635 | opts->options |= OPTION_MD5; | |
636 | size += TCPOLEN_MD5SIG_ALIGNED; | |
637 | } | |
638 | #else | |
639 | *md5 = NULL; | |
640 | #endif | |
641 | ||
642 | if (likely(tp->rx_opt.tstamp_ok)) { | |
643 | opts->options |= OPTION_TS; | |
ee684b6f | 644 | opts->tsval = tcb ? tcb->when + tp->tsoffset : 0; |
33ad798c AL |
645 | opts->tsecr = tp->rx_opt.ts_recent; |
646 | size += TCPOLEN_TSTAMP_ALIGNED; | |
647 | } | |
648 | ||
cabeccbd IJ |
649 | eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; |
650 | if (unlikely(eff_sacks)) { | |
95c96174 | 651 | const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; |
33ad798c | 652 | opts->num_sack_blocks = |
95c96174 | 653 | min_t(unsigned int, eff_sacks, |
33ad798c AL |
654 | (remaining - TCPOLEN_SACK_BASE_ALIGNED) / |
655 | TCPOLEN_SACK_PERBLOCK); | |
656 | size += TCPOLEN_SACK_BASE_ALIGNED + | |
657 | opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; | |
658 | } | |
659 | ||
660 | return size; | |
40efc6fa | 661 | } |
1da177e4 | 662 | |
46d3ceab ED |
663 | |
664 | /* TCP SMALL QUEUES (TSQ) | |
665 | * | |
666 | * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) | |
667 | * to reduce RTT and bufferbloat. | |
668 | * We do this using a special skb destructor (tcp_wfree). | |
669 | * | |
670 | * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb | |
671 | * needs to be reallocated in a driver. | |
672 | * The invariant being skb->truesize substracted from sk->sk_wmem_alloc | |
673 | * | |
674 | * Since transmit from skb destructor is forbidden, we use a tasklet | |
675 | * to process all sockets that eventually need to send more skbs. | |
676 | * We use one tasklet per cpu, with its own queue of sockets. | |
677 | */ | |
678 | struct tsq_tasklet { | |
679 | struct tasklet_struct tasklet; | |
680 | struct list_head head; /* queue of tcp sockets */ | |
681 | }; | |
682 | static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); | |
683 | ||
6f458dfb ED |
684 | static void tcp_tsq_handler(struct sock *sk) |
685 | { | |
686 | if ((1 << sk->sk_state) & | |
687 | (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | | |
688 | TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) | |
94ee16ae JO |
689 | tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle, |
690 | 0, GFP_ATOMIC); | |
6f458dfb | 691 | } |
46d3ceab ED |
692 | /* |
693 | * One tasklest per cpu tries to send more skbs. | |
694 | * We run in tasklet context but need to disable irqs when | |
695 | * transfering tsq->head because tcp_wfree() might | |
696 | * interrupt us (non NAPI drivers) | |
697 | */ | |
698 | static void tcp_tasklet_func(unsigned long data) | |
699 | { | |
700 | struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; | |
701 | LIST_HEAD(list); | |
702 | unsigned long flags; | |
703 | struct list_head *q, *n; | |
704 | struct tcp_sock *tp; | |
705 | struct sock *sk; | |
706 | ||
707 | local_irq_save(flags); | |
708 | list_splice_init(&tsq->head, &list); | |
709 | local_irq_restore(flags); | |
710 | ||
711 | list_for_each_safe(q, n, &list) { | |
712 | tp = list_entry(q, struct tcp_sock, tsq_node); | |
713 | list_del(&tp->tsq_node); | |
714 | ||
715 | sk = (struct sock *)tp; | |
716 | bh_lock_sock(sk); | |
717 | ||
718 | if (!sock_owned_by_user(sk)) { | |
6f458dfb | 719 | tcp_tsq_handler(sk); |
46d3ceab ED |
720 | } else { |
721 | /* defer the work to tcp_release_cb() */ | |
6f458dfb | 722 | set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); |
46d3ceab ED |
723 | } |
724 | bh_unlock_sock(sk); | |
725 | ||
726 | clear_bit(TSQ_QUEUED, &tp->tsq_flags); | |
727 | sk_free(sk); | |
728 | } | |
729 | } | |
730 | ||
6f458dfb ED |
731 | #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ |
732 | (1UL << TCP_WRITE_TIMER_DEFERRED) | \ | |
563d34d0 ED |
733 | (1UL << TCP_DELACK_TIMER_DEFERRED) | \ |
734 | (1UL << TCP_MTU_REDUCED_DEFERRED)) | |
46d3ceab ED |
735 | /** |
736 | * tcp_release_cb - tcp release_sock() callback | |
737 | * @sk: socket | |
738 | * | |
739 | * called from release_sock() to perform protocol dependent | |
740 | * actions before socket release. | |
741 | */ | |
742 | void tcp_release_cb(struct sock *sk) | |
743 | { | |
744 | struct tcp_sock *tp = tcp_sk(sk); | |
6f458dfb | 745 | unsigned long flags, nflags; |
46d3ceab | 746 | |
6f458dfb ED |
747 | /* perform an atomic operation only if at least one flag is set */ |
748 | do { | |
749 | flags = tp->tsq_flags; | |
750 | if (!(flags & TCP_DEFERRED_ALL)) | |
751 | return; | |
752 | nflags = flags & ~TCP_DEFERRED_ALL; | |
753 | } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); | |
754 | ||
755 | if (flags & (1UL << TCP_TSQ_DEFERRED)) | |
756 | tcp_tsq_handler(sk); | |
757 | ||
cbbb5a25 ED |
758 | /* Here begins the tricky part : |
759 | * We are called from release_sock() with : | |
760 | * 1) BH disabled | |
761 | * 2) sk_lock.slock spinlock held | |
762 | * 3) socket owned by us (sk->sk_lock.owned == 1) | |
763 | * | |
764 | * But following code is meant to be called from BH handlers, | |
765 | * so we should keep BH disabled, but early release socket ownership | |
766 | */ | |
767 | sock_release_ownership(sk); | |
768 | ||
144d56e9 | 769 | if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { |
6f458dfb | 770 | tcp_write_timer_handler(sk); |
144d56e9 ED |
771 | __sock_put(sk); |
772 | } | |
773 | if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { | |
6f458dfb | 774 | tcp_delack_timer_handler(sk); |
144d56e9 ED |
775 | __sock_put(sk); |
776 | } | |
777 | if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { | |
5f80f4d8 | 778 | inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); |
144d56e9 ED |
779 | __sock_put(sk); |
780 | } | |
46d3ceab ED |
781 | } |
782 | EXPORT_SYMBOL(tcp_release_cb); | |
783 | ||
784 | void __init tcp_tasklet_init(void) | |
785 | { | |
786 | int i; | |
787 | ||
788 | for_each_possible_cpu(i) { | |
789 | struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); | |
790 | ||
791 | INIT_LIST_HEAD(&tsq->head); | |
792 | tasklet_init(&tsq->tasklet, | |
793 | tcp_tasklet_func, | |
794 | (unsigned long)tsq); | |
795 | } | |
796 | } | |
797 | ||
798 | /* | |
799 | * Write buffer destructor automatically called from kfree_skb. | |
800 | * We cant xmit new skbs from this context, as we might already | |
801 | * hold qdisc lock. | |
802 | */ | |
d6a4a104 | 803 | void tcp_wfree(struct sk_buff *skb) |
46d3ceab ED |
804 | { |
805 | struct sock *sk = skb->sk; | |
806 | struct tcp_sock *tp = tcp_sk(sk); | |
807 | ||
808 | if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && | |
809 | !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { | |
810 | unsigned long flags; | |
811 | struct tsq_tasklet *tsq; | |
812 | ||
813 | /* Keep a ref on socket. | |
814 | * This last ref will be released in tcp_tasklet_func() | |
815 | */ | |
816 | atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc); | |
817 | ||
818 | /* queue this socket to tasklet queue */ | |
819 | local_irq_save(flags); | |
820 | tsq = &__get_cpu_var(tsq_tasklet); | |
821 | list_add(&tp->tsq_node, &tsq->head); | |
822 | tasklet_schedule(&tsq->tasklet); | |
823 | local_irq_restore(flags); | |
824 | } else { | |
825 | sock_wfree(skb); | |
826 | } | |
827 | } | |
828 | ||
1da177e4 LT |
829 | /* This routine actually transmits TCP packets queued in by |
830 | * tcp_do_sendmsg(). This is used by both the initial | |
831 | * transmission and possible later retransmissions. | |
832 | * All SKB's seen here are completely headerless. It is our | |
833 | * job to build the TCP header, and pass the packet down to | |
834 | * IP so it can do the same plus pass the packet off to the | |
835 | * device. | |
836 | * | |
837 | * We are working here with either a clone of the original | |
838 | * SKB, or a fresh unique copy made by the retransmit engine. | |
839 | */ | |
056834d9 IJ |
840 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, |
841 | gfp_t gfp_mask) | |
1da177e4 | 842 | { |
dfb4b9dc DM |
843 | const struct inet_connection_sock *icsk = inet_csk(sk); |
844 | struct inet_sock *inet; | |
845 | struct tcp_sock *tp; | |
846 | struct tcp_skb_cb *tcb; | |
33ad798c | 847 | struct tcp_out_options opts; |
95c96174 | 848 | unsigned int tcp_options_size, tcp_header_size; |
cfb6eeb4 | 849 | struct tcp_md5sig_key *md5; |
dfb4b9dc | 850 | struct tcphdr *th; |
dfb4b9dc DM |
851 | int err; |
852 | ||
853 | BUG_ON(!skb || !tcp_skb_pcount(skb)); | |
854 | ||
855 | /* If congestion control is doing timestamping, we must | |
856 | * take such a timestamp before we potentially clone/copy. | |
857 | */ | |
164891aa | 858 | if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) |
dfb4b9dc DM |
859 | __net_timestamp(skb); |
860 | ||
861 | if (likely(clone_it)) { | |
0e280af0 ED |
862 | const struct sk_buff *fclone = skb + 1; |
863 | ||
864 | if (unlikely(skb->fclone == SKB_FCLONE_ORIG && | |
865 | fclone->fclone == SKB_FCLONE_CLONE)) | |
866 | NET_INC_STATS_BH(sock_net(sk), | |
867 | LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); | |
868 | ||
dfb4b9dc DM |
869 | if (unlikely(skb_cloned(skb))) |
870 | skb = pskb_copy(skb, gfp_mask); | |
871 | else | |
872 | skb = skb_clone(skb, gfp_mask); | |
873 | if (unlikely(!skb)) | |
874 | return -ENOBUFS; | |
875 | } | |
1da177e4 | 876 | |
dfb4b9dc DM |
877 | inet = inet_sk(sk); |
878 | tp = tcp_sk(sk); | |
879 | tcb = TCP_SKB_CB(skb); | |
33ad798c | 880 | memset(&opts, 0, sizeof(opts)); |
1da177e4 | 881 | |
4de075e0 | 882 | if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) |
33ad798c AL |
883 | tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); |
884 | else | |
885 | tcp_options_size = tcp_established_options(sk, skb, &opts, | |
886 | &md5); | |
887 | tcp_header_size = tcp_options_size + sizeof(struct tcphdr); | |
e905a9ed | 888 | |
547669d4 | 889 | if (tcp_packets_in_flight(tp) == 0) |
dfb4b9dc | 890 | tcp_ca_event(sk, CA_EVENT_TX_START); |
547669d4 ED |
891 | |
892 | /* if no packet is in qdisc/device queue, then allow XPS to select | |
893 | * another queue. | |
894 | */ | |
895 | skb->ooo_okay = sk_wmem_alloc_get(sk) == 0; | |
dfb4b9dc | 896 | |
aa8223c7 ACM |
897 | skb_push(skb, tcp_header_size); |
898 | skb_reset_transport_header(skb); | |
46d3ceab ED |
899 | |
900 | skb_orphan(skb); | |
901 | skb->sk = sk; | |
0ae5f47e | 902 | skb->destructor = tcp_wfree; |
46d3ceab | 903 | atomic_add(skb->truesize, &sk->sk_wmem_alloc); |
dfb4b9dc DM |
904 | |
905 | /* Build TCP header and checksum it. */ | |
aa8223c7 | 906 | th = tcp_hdr(skb); |
c720c7e8 ED |
907 | th->source = inet->inet_sport; |
908 | th->dest = inet->inet_dport; | |
dfb4b9dc DM |
909 | th->seq = htonl(tcb->seq); |
910 | th->ack_seq = htonl(tp->rcv_nxt); | |
df7a3b07 | 911 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
4de075e0 | 912 | tcb->tcp_flags); |
dfb4b9dc | 913 | |
4de075e0 | 914 | if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { |
dfb4b9dc DM |
915 | /* RFC1323: The window in SYN & SYN/ACK segments |
916 | * is never scaled. | |
917 | */ | |
600ff0c2 | 918 | th->window = htons(min(tp->rcv_wnd, 65535U)); |
dfb4b9dc DM |
919 | } else { |
920 | th->window = htons(tcp_select_window(sk)); | |
921 | } | |
922 | th->check = 0; | |
923 | th->urg_ptr = 0; | |
1da177e4 | 924 | |
33f5f57e | 925 | /* The urg_mode check is necessary during a below snd_una win probe */ |
7691367d HX |
926 | if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { |
927 | if (before(tp->snd_up, tcb->seq + 0x10000)) { | |
928 | th->urg_ptr = htons(tp->snd_up - tcb->seq); | |
929 | th->urg = 1; | |
930 | } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { | |
0eae88f3 | 931 | th->urg_ptr = htons(0xFFFF); |
7691367d HX |
932 | th->urg = 1; |
933 | } | |
dfb4b9dc | 934 | } |
1da177e4 | 935 | |
bd0388ae | 936 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
4de075e0 | 937 | if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) |
9e412ba7 | 938 | TCP_ECN_send(sk, skb, tcp_header_size); |
1da177e4 | 939 | |
cfb6eeb4 YH |
940 | #ifdef CONFIG_TCP_MD5SIG |
941 | /* Calculate the MD5 hash, as we have all we need now */ | |
942 | if (md5) { | |
a465419b | 943 | sk_nocaps_add(sk, NETIF_F_GSO_MASK); |
bd0388ae | 944 | tp->af_specific->calc_md5_hash(opts.hash_location, |
49a72dfb | 945 | md5, sk, NULL, skb); |
cfb6eeb4 YH |
946 | } |
947 | #endif | |
948 | ||
bb296246 | 949 | icsk->icsk_af_ops->send_check(sk, skb); |
1da177e4 | 950 | |
4de075e0 | 951 | if (likely(tcb->tcp_flags & TCPHDR_ACK)) |
dfb4b9dc | 952 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); |
1da177e4 | 953 | |
dfb4b9dc | 954 | if (skb->len != tcp_header_size) |
cf533ea5 | 955 | tcp_event_data_sent(tp, sk); |
1da177e4 | 956 | |
bd37a088 | 957 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
aa2ea058 TH |
958 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, |
959 | tcp_skb_pcount(skb)); | |
1da177e4 | 960 | |
d9d8da80 | 961 | err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl); |
83de47cd | 962 | if (likely(err <= 0)) |
dfb4b9dc DM |
963 | return err; |
964 | ||
3cfe3baa | 965 | tcp_enter_cwr(sk, 1); |
dfb4b9dc | 966 | |
b9df3cb8 | 967 | return net_xmit_eval(err); |
1da177e4 LT |
968 | } |
969 | ||
67edfef7 | 970 | /* This routine just queues the buffer for sending. |
1da177e4 LT |
971 | * |
972 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, | |
973 | * otherwise socket can stall. | |
974 | */ | |
975 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) | |
976 | { | |
977 | struct tcp_sock *tp = tcp_sk(sk); | |
978 | ||
979 | /* Advance write_seq and place onto the write_queue. */ | |
980 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; | |
981 | skb_header_release(skb); | |
fe067e8a | 982 | tcp_add_write_queue_tail(sk, skb); |
3ab224be HA |
983 | sk->sk_wmem_queued += skb->truesize; |
984 | sk_mem_charge(sk, skb->truesize); | |
1da177e4 LT |
985 | } |
986 | ||
67edfef7 | 987 | /* Initialize TSO segments for a packet. */ |
cf533ea5 | 988 | static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, |
056834d9 | 989 | unsigned int mss_now) |
f6302d1d | 990 | { |
b81908e1 ED |
991 | /* Make sure we own this skb before messing gso_size/gso_segs */ |
992 | WARN_ON_ONCE(skb_cloned(skb)); | |
993 | ||
8e5b9dda HX |
994 | if (skb->len <= mss_now || !sk_can_gso(sk) || |
995 | skb->ip_summed == CHECKSUM_NONE) { | |
f6302d1d DM |
996 | /* Avoid the costly divide in the normal |
997 | * non-TSO case. | |
998 | */ | |
7967168c HX |
999 | skb_shinfo(skb)->gso_segs = 1; |
1000 | skb_shinfo(skb)->gso_size = 0; | |
c9af6db4 | 1001 | skb_shinfo(skb)->gso_type = 0; |
f6302d1d | 1002 | } else { |
356f89e1 | 1003 | skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); |
7967168c | 1004 | skb_shinfo(skb)->gso_size = mss_now; |
c9af6db4 | 1005 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
1da177e4 LT |
1006 | } |
1007 | } | |
1008 | ||
91fed7a1 | 1009 | /* When a modification to fackets out becomes necessary, we need to check |
68f8353b | 1010 | * skb is counted to fackets_out or not. |
91fed7a1 | 1011 | */ |
cf533ea5 | 1012 | static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, |
91fed7a1 IJ |
1013 | int decr) |
1014 | { | |
a47e5a98 IJ |
1015 | struct tcp_sock *tp = tcp_sk(sk); |
1016 | ||
dc86967b | 1017 | if (!tp->sacked_out || tcp_is_reno(tp)) |
91fed7a1 IJ |
1018 | return; |
1019 | ||
6859d494 | 1020 | if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) |
91fed7a1 | 1021 | tp->fackets_out -= decr; |
91fed7a1 IJ |
1022 | } |
1023 | ||
797108d1 IJ |
1024 | /* Pcount in the middle of the write queue got changed, we need to do various |
1025 | * tweaks to fix counters | |
1026 | */ | |
cf533ea5 | 1027 | static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) |
797108d1 IJ |
1028 | { |
1029 | struct tcp_sock *tp = tcp_sk(sk); | |
1030 | ||
1031 | tp->packets_out -= decr; | |
1032 | ||
1033 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
1034 | tp->sacked_out -= decr; | |
1035 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
1036 | tp->retrans_out -= decr; | |
1037 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) | |
1038 | tp->lost_out -= decr; | |
1039 | ||
1040 | /* Reno case is special. Sigh... */ | |
1041 | if (tcp_is_reno(tp) && decr > 0) | |
1042 | tp->sacked_out -= min_t(u32, tp->sacked_out, decr); | |
1043 | ||
1044 | tcp_adjust_fackets_out(sk, skb, decr); | |
1045 | ||
1046 | if (tp->lost_skb_hint && | |
1047 | before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && | |
52cf3cc8 | 1048 | (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) |
797108d1 IJ |
1049 | tp->lost_cnt_hint -= decr; |
1050 | ||
1051 | tcp_verify_left_out(tp); | |
1052 | } | |
1053 | ||
1da177e4 LT |
1054 | /* Function to create two new TCP segments. Shrinks the given segment |
1055 | * to the specified size and appends a new segment with the rest of the | |
e905a9ed | 1056 | * packet to the list. This won't be called frequently, I hope. |
1da177e4 LT |
1057 | * Remember, these are still headerless SKBs at this point. |
1058 | */ | |
056834d9 IJ |
1059 | int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, |
1060 | unsigned int mss_now) | |
1da177e4 LT |
1061 | { |
1062 | struct tcp_sock *tp = tcp_sk(sk); | |
1063 | struct sk_buff *buff; | |
6475be16 | 1064 | int nsize, old_factor; |
b60b49ea | 1065 | int nlen; |
9ce01461 | 1066 | u8 flags; |
1da177e4 | 1067 | |
2fceec13 IJ |
1068 | if (WARN_ON(len > skb->len)) |
1069 | return -EINVAL; | |
6a438bbe | 1070 | |
1da177e4 LT |
1071 | nsize = skb_headlen(skb) - len; |
1072 | if (nsize < 0) | |
1073 | nsize = 0; | |
1074 | ||
b81908e1 | 1075 | if (skb_unclone(skb, GFP_ATOMIC)) |
1da177e4 LT |
1076 | return -ENOMEM; |
1077 | ||
1078 | /* Get a new skb... force flag on. */ | |
1079 | buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); | |
1080 | if (buff == NULL) | |
1081 | return -ENOMEM; /* We'll just try again later. */ | |
ef5cb973 | 1082 | |
3ab224be HA |
1083 | sk->sk_wmem_queued += buff->truesize; |
1084 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea HX |
1085 | nlen = skb->len - len - nsize; |
1086 | buff->truesize += nlen; | |
1087 | skb->truesize -= nlen; | |
1da177e4 LT |
1088 | |
1089 | /* Correct the sequence numbers. */ | |
1090 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1091 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1092 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1093 | ||
1094 | /* PSH and FIN should only be set in the second packet. */ | |
4de075e0 ED |
1095 | flags = TCP_SKB_CB(skb)->tcp_flags; |
1096 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
1097 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
e14c3caf | 1098 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
1da177e4 | 1099 | |
84fa7933 | 1100 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { |
1da177e4 | 1101 | /* Copy and checksum data tail into the new buffer. */ |
056834d9 IJ |
1102 | buff->csum = csum_partial_copy_nocheck(skb->data + len, |
1103 | skb_put(buff, nsize), | |
1da177e4 LT |
1104 | nsize, 0); |
1105 | ||
1106 | skb_trim(skb, len); | |
1107 | ||
1108 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); | |
1109 | } else { | |
84fa7933 | 1110 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
1111 | skb_split(skb, buff, len); |
1112 | } | |
1113 | ||
1114 | buff->ip_summed = skb->ip_summed; | |
1115 | ||
1116 | /* Looks stupid, but our code really uses when of | |
1117 | * skbs, which it never sent before. --ANK | |
1118 | */ | |
1119 | TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; | |
a61bbcf2 | 1120 | buff->tstamp = skb->tstamp; |
1da177e4 | 1121 | |
6475be16 DM |
1122 | old_factor = tcp_skb_pcount(skb); |
1123 | ||
1da177e4 | 1124 | /* Fix up tso_factor for both original and new SKB. */ |
846998ae DM |
1125 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
1126 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
1da177e4 | 1127 | |
6475be16 DM |
1128 | /* If this packet has been sent out already, we must |
1129 | * adjust the various packet counters. | |
1130 | */ | |
cf0b450c | 1131 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
6475be16 DM |
1132 | int diff = old_factor - tcp_skb_pcount(skb) - |
1133 | tcp_skb_pcount(buff); | |
1da177e4 | 1134 | |
797108d1 IJ |
1135 | if (diff) |
1136 | tcp_adjust_pcount(sk, skb, diff); | |
1da177e4 LT |
1137 | } |
1138 | ||
1139 | /* Link BUFF into the send queue. */ | |
f44b5271 | 1140 | skb_header_release(buff); |
fe067e8a | 1141 | tcp_insert_write_queue_after(skb, buff, sk); |
1da177e4 LT |
1142 | |
1143 | return 0; | |
1144 | } | |
1145 | ||
1146 | /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c | |
1147 | * eventually). The difference is that pulled data not copied, but | |
1148 | * immediately discarded. | |
1149 | */ | |
f2911969 | 1150 | static void __pskb_trim_head(struct sk_buff *skb, int len) |
1da177e4 LT |
1151 | { |
1152 | int i, k, eat; | |
1153 | ||
4fa48bf3 ED |
1154 | eat = min_t(int, len, skb_headlen(skb)); |
1155 | if (eat) { | |
1156 | __skb_pull(skb, eat); | |
1157 | len -= eat; | |
1158 | if (!len) | |
1159 | return; | |
1160 | } | |
1da177e4 LT |
1161 | eat = len; |
1162 | k = 0; | |
056834d9 | 1163 | for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { |
9e903e08 ED |
1164 | int size = skb_frag_size(&skb_shinfo(skb)->frags[i]); |
1165 | ||
1166 | if (size <= eat) { | |
aff65da0 | 1167 | skb_frag_unref(skb, i); |
9e903e08 | 1168 | eat -= size; |
1da177e4 LT |
1169 | } else { |
1170 | skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; | |
1171 | if (eat) { | |
1172 | skb_shinfo(skb)->frags[k].page_offset += eat; | |
9e903e08 | 1173 | skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat); |
1da177e4 LT |
1174 | eat = 0; |
1175 | } | |
1176 | k++; | |
1177 | } | |
1178 | } | |
1179 | skb_shinfo(skb)->nr_frags = k; | |
1180 | ||
27a884dc | 1181 | skb_reset_tail_pointer(skb); |
1da177e4 LT |
1182 | skb->data_len -= len; |
1183 | skb->len = skb->data_len; | |
1da177e4 LT |
1184 | } |
1185 | ||
67edfef7 | 1186 | /* Remove acked data from a packet in the transmit queue. */ |
1da177e4 LT |
1187 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) |
1188 | { | |
14bbd6a5 | 1189 | if (skb_unclone(skb, GFP_ATOMIC)) |
1da177e4 LT |
1190 | return -ENOMEM; |
1191 | ||
4fa48bf3 | 1192 | __pskb_trim_head(skb, len); |
1da177e4 LT |
1193 | |
1194 | TCP_SKB_CB(skb)->seq += len; | |
84fa7933 | 1195 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
1196 | |
1197 | skb->truesize -= len; | |
1198 | sk->sk_wmem_queued -= len; | |
3ab224be | 1199 | sk_mem_uncharge(sk, len); |
1da177e4 LT |
1200 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); |
1201 | ||
5b35e1e6 | 1202 | /* Any change of skb->len requires recalculation of tso factor. */ |
1da177e4 | 1203 | if (tcp_skb_pcount(skb) > 1) |
5b35e1e6 | 1204 | tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); |
1da177e4 LT |
1205 | |
1206 | return 0; | |
1207 | } | |
1208 | ||
1b63edd6 YC |
1209 | /* Calculate MSS not accounting any TCP options. */ |
1210 | static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
5d424d5a | 1211 | { |
cf533ea5 ED |
1212 | const struct tcp_sock *tp = tcp_sk(sk); |
1213 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
5d424d5a JH |
1214 | int mss_now; |
1215 | ||
1216 | /* Calculate base mss without TCP options: | |
1217 | It is MMS_S - sizeof(tcphdr) of rfc1122 | |
1218 | */ | |
1219 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); | |
1220 | ||
67469601 ED |
1221 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
1222 | if (icsk->icsk_af_ops->net_frag_header_len) { | |
1223 | const struct dst_entry *dst = __sk_dst_get(sk); | |
1224 | ||
1225 | if (dst && dst_allfrag(dst)) | |
1226 | mss_now -= icsk->icsk_af_ops->net_frag_header_len; | |
1227 | } | |
1228 | ||
5d424d5a JH |
1229 | /* Clamp it (mss_clamp does not include tcp options) */ |
1230 | if (mss_now > tp->rx_opt.mss_clamp) | |
1231 | mss_now = tp->rx_opt.mss_clamp; | |
1232 | ||
1233 | /* Now subtract optional transport overhead */ | |
1234 | mss_now -= icsk->icsk_ext_hdr_len; | |
1235 | ||
1236 | /* Then reserve room for full set of TCP options and 8 bytes of data */ | |
1237 | if (mss_now < 48) | |
1238 | mss_now = 48; | |
5d424d5a JH |
1239 | return mss_now; |
1240 | } | |
1241 | ||
1b63edd6 YC |
1242 | /* Calculate MSS. Not accounting for SACKs here. */ |
1243 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
1244 | { | |
1245 | /* Subtract TCP options size, not including SACKs */ | |
1246 | return __tcp_mtu_to_mss(sk, pmtu) - | |
1247 | (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); | |
1248 | } | |
1249 | ||
5d424d5a | 1250 | /* Inverse of above */ |
67469601 | 1251 | int tcp_mss_to_mtu(struct sock *sk, int mss) |
5d424d5a | 1252 | { |
cf533ea5 ED |
1253 | const struct tcp_sock *tp = tcp_sk(sk); |
1254 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
5d424d5a JH |
1255 | int mtu; |
1256 | ||
1257 | mtu = mss + | |
1258 | tp->tcp_header_len + | |
1259 | icsk->icsk_ext_hdr_len + | |
1260 | icsk->icsk_af_ops->net_header_len; | |
1261 | ||
67469601 ED |
1262 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
1263 | if (icsk->icsk_af_ops->net_frag_header_len) { | |
1264 | const struct dst_entry *dst = __sk_dst_get(sk); | |
1265 | ||
1266 | if (dst && dst_allfrag(dst)) | |
1267 | mtu += icsk->icsk_af_ops->net_frag_header_len; | |
1268 | } | |
5d424d5a JH |
1269 | return mtu; |
1270 | } | |
1271 | ||
67edfef7 | 1272 | /* MTU probing init per socket */ |
5d424d5a JH |
1273 | void tcp_mtup_init(struct sock *sk) |
1274 | { | |
1275 | struct tcp_sock *tp = tcp_sk(sk); | |
1276 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1277 | ||
1278 | icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; | |
1279 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + | |
e905a9ed | 1280 | icsk->icsk_af_ops->net_header_len; |
5d424d5a JH |
1281 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); |
1282 | icsk->icsk_mtup.probe_size = 0; | |
1283 | } | |
4bc2f18b | 1284 | EXPORT_SYMBOL(tcp_mtup_init); |
5d424d5a | 1285 | |
1da177e4 LT |
1286 | /* This function synchronize snd mss to current pmtu/exthdr set. |
1287 | ||
1288 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts | |
1289 | for TCP options, but includes only bare TCP header. | |
1290 | ||
1291 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. | |
caa20d9a | 1292 | It is minimum of user_mss and mss received with SYN. |
1da177e4 LT |
1293 | It also does not include TCP options. |
1294 | ||
d83d8461 | 1295 | inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. |
1da177e4 LT |
1296 | |
1297 | tp->mss_cache is current effective sending mss, including | |
1298 | all tcp options except for SACKs. It is evaluated, | |
1299 | taking into account current pmtu, but never exceeds | |
1300 | tp->rx_opt.mss_clamp. | |
1301 | ||
1302 | NOTE1. rfc1122 clearly states that advertised MSS | |
1303 | DOES NOT include either tcp or ip options. | |
1304 | ||
d83d8461 ACM |
1305 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
1306 | are READ ONLY outside this function. --ANK (980731) | |
1da177e4 | 1307 | */ |
1da177e4 LT |
1308 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) |
1309 | { | |
1310 | struct tcp_sock *tp = tcp_sk(sk); | |
d83d8461 | 1311 | struct inet_connection_sock *icsk = inet_csk(sk); |
5d424d5a | 1312 | int mss_now; |
1da177e4 | 1313 | |
5d424d5a JH |
1314 | if (icsk->icsk_mtup.search_high > pmtu) |
1315 | icsk->icsk_mtup.search_high = pmtu; | |
1da177e4 | 1316 | |
5d424d5a | 1317 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
409d22b4 | 1318 | mss_now = tcp_bound_to_half_wnd(tp, mss_now); |
1da177e4 LT |
1319 | |
1320 | /* And store cached results */ | |
d83d8461 | 1321 | icsk->icsk_pmtu_cookie = pmtu; |
5d424d5a JH |
1322 | if (icsk->icsk_mtup.enabled) |
1323 | mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); | |
c1b4a7e6 | 1324 | tp->mss_cache = mss_now; |
1da177e4 LT |
1325 | |
1326 | return mss_now; | |
1327 | } | |
4bc2f18b | 1328 | EXPORT_SYMBOL(tcp_sync_mss); |
1da177e4 LT |
1329 | |
1330 | /* Compute the current effective MSS, taking SACKs and IP options, | |
1331 | * and even PMTU discovery events into account. | |
1da177e4 | 1332 | */ |
0c54b85f | 1333 | unsigned int tcp_current_mss(struct sock *sk) |
1da177e4 | 1334 | { |
cf533ea5 ED |
1335 | const struct tcp_sock *tp = tcp_sk(sk); |
1336 | const struct dst_entry *dst = __sk_dst_get(sk); | |
c1b4a7e6 | 1337 | u32 mss_now; |
95c96174 | 1338 | unsigned int header_len; |
33ad798c AL |
1339 | struct tcp_out_options opts; |
1340 | struct tcp_md5sig_key *md5; | |
c1b4a7e6 DM |
1341 | |
1342 | mss_now = tp->mss_cache; | |
1343 | ||
1da177e4 LT |
1344 | if (dst) { |
1345 | u32 mtu = dst_mtu(dst); | |
d83d8461 | 1346 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
1da177e4 LT |
1347 | mss_now = tcp_sync_mss(sk, mtu); |
1348 | } | |
1349 | ||
33ad798c AL |
1350 | header_len = tcp_established_options(sk, NULL, &opts, &md5) + |
1351 | sizeof(struct tcphdr); | |
1352 | /* The mss_cache is sized based on tp->tcp_header_len, which assumes | |
1353 | * some common options. If this is an odd packet (because we have SACK | |
1354 | * blocks etc) then our calculated header_len will be different, and | |
1355 | * we have to adjust mss_now correspondingly */ | |
1356 | if (header_len != tp->tcp_header_len) { | |
1357 | int delta = (int) header_len - tp->tcp_header_len; | |
1358 | mss_now -= delta; | |
1359 | } | |
cfb6eeb4 | 1360 | |
1da177e4 LT |
1361 | return mss_now; |
1362 | } | |
1363 | ||
a762a980 | 1364 | /* Congestion window validation. (RFC2861) */ |
9e412ba7 | 1365 | static void tcp_cwnd_validate(struct sock *sk) |
a762a980 | 1366 | { |
9e412ba7 | 1367 | struct tcp_sock *tp = tcp_sk(sk); |
a762a980 | 1368 | |
d436d686 | 1369 | if (tp->packets_out >= tp->snd_cwnd) { |
a762a980 DM |
1370 | /* Network is feed fully. */ |
1371 | tp->snd_cwnd_used = 0; | |
1372 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1373 | } else { | |
1374 | /* Network starves. */ | |
1375 | if (tp->packets_out > tp->snd_cwnd_used) | |
1376 | tp->snd_cwnd_used = tp->packets_out; | |
1377 | ||
15d33c07 DM |
1378 | if (sysctl_tcp_slow_start_after_idle && |
1379 | (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) | |
a762a980 DM |
1380 | tcp_cwnd_application_limited(sk); |
1381 | } | |
1382 | } | |
1383 | ||
0e3a4803 IJ |
1384 | /* Returns the portion of skb which can be sent right away without |
1385 | * introducing MSS oddities to segment boundaries. In rare cases where | |
1386 | * mss_now != mss_cache, we will request caller to create a small skb | |
1387 | * per input skb which could be mostly avoided here (if desired). | |
5ea3a748 IJ |
1388 | * |
1389 | * We explicitly want to create a request for splitting write queue tail | |
1390 | * to a small skb for Nagle purposes while avoiding unnecessary modulos, | |
1391 | * thus all the complexity (cwnd_len is always MSS multiple which we | |
1392 | * return whenever allowed by the other factors). Basically we need the | |
1393 | * modulo only when the receiver window alone is the limiting factor or | |
1394 | * when we would be allowed to send the split-due-to-Nagle skb fully. | |
0e3a4803 | 1395 | */ |
cf533ea5 | 1396 | static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb, |
1485348d | 1397 | unsigned int mss_now, unsigned int max_segs) |
c1b4a7e6 | 1398 | { |
cf533ea5 | 1399 | const struct tcp_sock *tp = tcp_sk(sk); |
1485348d | 1400 | u32 needed, window, max_len; |
c1b4a7e6 | 1401 | |
90840def | 1402 | window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
1485348d | 1403 | max_len = mss_now * max_segs; |
0e3a4803 | 1404 | |
1485348d BH |
1405 | if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) |
1406 | return max_len; | |
0e3a4803 | 1407 | |
5ea3a748 IJ |
1408 | needed = min(skb->len, window); |
1409 | ||
1485348d BH |
1410 | if (max_len <= needed) |
1411 | return max_len; | |
0e3a4803 | 1412 | |
0e3a4803 | 1413 | return needed - needed % mss_now; |
c1b4a7e6 DM |
1414 | } |
1415 | ||
1416 | /* Can at least one segment of SKB be sent right now, according to the | |
1417 | * congestion window rules? If so, return how many segments are allowed. | |
1418 | */ | |
cf533ea5 ED |
1419 | static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, |
1420 | const struct sk_buff *skb) | |
c1b4a7e6 DM |
1421 | { |
1422 | u32 in_flight, cwnd; | |
1423 | ||
1424 | /* Don't be strict about the congestion window for the final FIN. */ | |
4de075e0 ED |
1425 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && |
1426 | tcp_skb_pcount(skb) == 1) | |
c1b4a7e6 DM |
1427 | return 1; |
1428 | ||
1429 | in_flight = tcp_packets_in_flight(tp); | |
1430 | cwnd = tp->snd_cwnd; | |
1431 | if (in_flight < cwnd) | |
1432 | return (cwnd - in_flight); | |
1433 | ||
1434 | return 0; | |
1435 | } | |
1436 | ||
b595076a | 1437 | /* Initialize TSO state of a skb. |
67edfef7 | 1438 | * This must be invoked the first time we consider transmitting |
c1b4a7e6 DM |
1439 | * SKB onto the wire. |
1440 | */ | |
cf533ea5 | 1441 | static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, |
056834d9 | 1442 | unsigned int mss_now) |
c1b4a7e6 DM |
1443 | { |
1444 | int tso_segs = tcp_skb_pcount(skb); | |
1445 | ||
f8269a49 | 1446 | if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { |
846998ae | 1447 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
c1b4a7e6 DM |
1448 | tso_segs = tcp_skb_pcount(skb); |
1449 | } | |
1450 | return tso_segs; | |
1451 | } | |
1452 | ||
67edfef7 | 1453 | /* Minshall's variant of the Nagle send check. */ |
a2a385d6 | 1454 | static inline bool tcp_minshall_check(const struct tcp_sock *tp) |
c1b4a7e6 | 1455 | { |
09cb105e | 1456 | return after(tp->snd_sml, tp->snd_una) && |
c1b4a7e6 DM |
1457 | !after(tp->snd_sml, tp->snd_nxt); |
1458 | } | |
1459 | ||
a2a385d6 | 1460 | /* Return false, if packet can be sent now without violation Nagle's rules: |
c1b4a7e6 DM |
1461 | * 1. It is full sized. |
1462 | * 2. Or it contains FIN. (already checked by caller) | |
6d67e9be | 1463 | * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. |
c1b4a7e6 DM |
1464 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. |
1465 | * With Minshall's modification: all sent small packets are ACKed. | |
1466 | */ | |
a2a385d6 | 1467 | static inline bool tcp_nagle_check(const struct tcp_sock *tp, |
e905a9ed | 1468 | const struct sk_buff *skb, |
95c96174 | 1469 | unsigned int mss_now, int nonagle) |
c1b4a7e6 | 1470 | { |
a02cec21 | 1471 | return skb->len < mss_now && |
056834d9 | 1472 | ((nonagle & TCP_NAGLE_CORK) || |
a02cec21 | 1473 | (!nonagle && tp->packets_out && tcp_minshall_check(tp))); |
c1b4a7e6 DM |
1474 | } |
1475 | ||
a2a385d6 | 1476 | /* Return true if the Nagle test allows this packet to be |
c1b4a7e6 DM |
1477 | * sent now. |
1478 | */ | |
a2a385d6 ED |
1479 | static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, |
1480 | unsigned int cur_mss, int nonagle) | |
c1b4a7e6 DM |
1481 | { |
1482 | /* Nagle rule does not apply to frames, which sit in the middle of the | |
1483 | * write_queue (they have no chances to get new data). | |
1484 | * | |
1485 | * This is implemented in the callers, where they modify the 'nonagle' | |
1486 | * argument based upon the location of SKB in the send queue. | |
1487 | */ | |
1488 | if (nonagle & TCP_NAGLE_PUSH) | |
a2a385d6 | 1489 | return true; |
c1b4a7e6 | 1490 | |
9b44190d YC |
1491 | /* Don't use the nagle rule for urgent data (or for the final FIN). */ |
1492 | if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) | |
a2a385d6 | 1493 | return true; |
c1b4a7e6 DM |
1494 | |
1495 | if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) | |
a2a385d6 | 1496 | return true; |
c1b4a7e6 | 1497 | |
a2a385d6 | 1498 | return false; |
c1b4a7e6 DM |
1499 | } |
1500 | ||
1501 | /* Does at least the first segment of SKB fit into the send window? */ | |
a2a385d6 ED |
1502 | static bool tcp_snd_wnd_test(const struct tcp_sock *tp, |
1503 | const struct sk_buff *skb, | |
1504 | unsigned int cur_mss) | |
c1b4a7e6 DM |
1505 | { |
1506 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
1507 | ||
1508 | if (skb->len > cur_mss) | |
1509 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; | |
1510 | ||
90840def | 1511 | return !after(end_seq, tcp_wnd_end(tp)); |
c1b4a7e6 DM |
1512 | } |
1513 | ||
fe067e8a | 1514 | /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) |
c1b4a7e6 DM |
1515 | * should be put on the wire right now. If so, it returns the number of |
1516 | * packets allowed by the congestion window. | |
1517 | */ | |
cf533ea5 | 1518 | static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, |
c1b4a7e6 DM |
1519 | unsigned int cur_mss, int nonagle) |
1520 | { | |
cf533ea5 | 1521 | const struct tcp_sock *tp = tcp_sk(sk); |
c1b4a7e6 DM |
1522 | unsigned int cwnd_quota; |
1523 | ||
846998ae | 1524 | tcp_init_tso_segs(sk, skb, cur_mss); |
c1b4a7e6 DM |
1525 | |
1526 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) | |
1527 | return 0; | |
1528 | ||
1529 | cwnd_quota = tcp_cwnd_test(tp, skb); | |
056834d9 | 1530 | if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) |
c1b4a7e6 DM |
1531 | cwnd_quota = 0; |
1532 | ||
1533 | return cwnd_quota; | |
1534 | } | |
1535 | ||
67edfef7 | 1536 | /* Test if sending is allowed right now. */ |
a2a385d6 | 1537 | bool tcp_may_send_now(struct sock *sk) |
c1b4a7e6 | 1538 | { |
cf533ea5 | 1539 | const struct tcp_sock *tp = tcp_sk(sk); |
fe067e8a | 1540 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 | 1541 | |
a02cec21 | 1542 | return skb && |
0c54b85f | 1543 | tcp_snd_test(sk, skb, tcp_current_mss(sk), |
c1b4a7e6 | 1544 | (tcp_skb_is_last(sk, skb) ? |
a02cec21 | 1545 | tp->nonagle : TCP_NAGLE_PUSH)); |
c1b4a7e6 DM |
1546 | } |
1547 | ||
1548 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet | |
1549 | * which is put after SKB on the list. It is very much like | |
1550 | * tcp_fragment() except that it may make several kinds of assumptions | |
1551 | * in order to speed up the splitting operation. In particular, we | |
1552 | * know that all the data is in scatter-gather pages, and that the | |
1553 | * packet has never been sent out before (and thus is not cloned). | |
1554 | */ | |
056834d9 | 1555 | static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, |
c4ead4c5 | 1556 | unsigned int mss_now, gfp_t gfp) |
c1b4a7e6 DM |
1557 | { |
1558 | struct sk_buff *buff; | |
1559 | int nlen = skb->len - len; | |
9ce01461 | 1560 | u8 flags; |
c1b4a7e6 DM |
1561 | |
1562 | /* All of a TSO frame must be composed of paged data. */ | |
c8ac3774 HX |
1563 | if (skb->len != skb->data_len) |
1564 | return tcp_fragment(sk, skb, len, mss_now); | |
c1b4a7e6 | 1565 | |
c4ead4c5 | 1566 | buff = sk_stream_alloc_skb(sk, 0, gfp); |
c1b4a7e6 DM |
1567 | if (unlikely(buff == NULL)) |
1568 | return -ENOMEM; | |
1569 | ||
3ab224be HA |
1570 | sk->sk_wmem_queued += buff->truesize; |
1571 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea | 1572 | buff->truesize += nlen; |
c1b4a7e6 DM |
1573 | skb->truesize -= nlen; |
1574 | ||
1575 | /* Correct the sequence numbers. */ | |
1576 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1577 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1578 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1579 | ||
1580 | /* PSH and FIN should only be set in the second packet. */ | |
4de075e0 ED |
1581 | flags = TCP_SKB_CB(skb)->tcp_flags; |
1582 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
1583 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
c1b4a7e6 DM |
1584 | |
1585 | /* This packet was never sent out yet, so no SACK bits. */ | |
1586 | TCP_SKB_CB(buff)->sacked = 0; | |
1587 | ||
84fa7933 | 1588 | buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; |
c1b4a7e6 DM |
1589 | skb_split(skb, buff, len); |
1590 | ||
1591 | /* Fix up tso_factor for both original and new SKB. */ | |
846998ae DM |
1592 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
1593 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
c1b4a7e6 DM |
1594 | |
1595 | /* Link BUFF into the send queue. */ | |
1596 | skb_header_release(buff); | |
fe067e8a | 1597 | tcp_insert_write_queue_after(skb, buff, sk); |
c1b4a7e6 DM |
1598 | |
1599 | return 0; | |
1600 | } | |
1601 | ||
1602 | /* Try to defer sending, if possible, in order to minimize the amount | |
1603 | * of TSO splitting we do. View it as a kind of TSO Nagle test. | |
1604 | * | |
1605 | * This algorithm is from John Heffner. | |
1606 | */ | |
a2a385d6 | 1607 | static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) |
c1b4a7e6 | 1608 | { |
9e412ba7 | 1609 | struct tcp_sock *tp = tcp_sk(sk); |
6687e988 | 1610 | const struct inet_connection_sock *icsk = inet_csk(sk); |
c1b4a7e6 | 1611 | u32 send_win, cong_win, limit, in_flight; |
ad9f4f50 | 1612 | int win_divisor; |
c1b4a7e6 | 1613 | |
4de075e0 | 1614 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) |
ae8064ac | 1615 | goto send_now; |
c1b4a7e6 | 1616 | |
6687e988 | 1617 | if (icsk->icsk_ca_state != TCP_CA_Open) |
ae8064ac JH |
1618 | goto send_now; |
1619 | ||
1620 | /* Defer for less than two clock ticks. */ | |
bd515c3e | 1621 | if (tp->tso_deferred && |
a2acde07 | 1622 | (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) |
ae8064ac | 1623 | goto send_now; |
908a75c1 | 1624 | |
c1b4a7e6 DM |
1625 | in_flight = tcp_packets_in_flight(tp); |
1626 | ||
056834d9 | 1627 | BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); |
c1b4a7e6 | 1628 | |
90840def | 1629 | send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
c1b4a7e6 DM |
1630 | |
1631 | /* From in_flight test above, we know that cwnd > in_flight. */ | |
1632 | cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; | |
1633 | ||
1634 | limit = min(send_win, cong_win); | |
1635 | ||
ba244fe9 | 1636 | /* If a full-sized TSO skb can be sent, do it. */ |
1485348d | 1637 | if (limit >= min_t(unsigned int, sk->sk_gso_max_size, |
5e25ba50 | 1638 | tp->xmit_size_goal_segs * tp->mss_cache)) |
ae8064ac | 1639 | goto send_now; |
ba244fe9 | 1640 | |
62ad2761 IJ |
1641 | /* Middle in queue won't get any more data, full sendable already? */ |
1642 | if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) | |
1643 | goto send_now; | |
1644 | ||
ad9f4f50 ED |
1645 | win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); |
1646 | if (win_divisor) { | |
c1b4a7e6 DM |
1647 | u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); |
1648 | ||
1649 | /* If at least some fraction of a window is available, | |
1650 | * just use it. | |
1651 | */ | |
ad9f4f50 | 1652 | chunk /= win_divisor; |
c1b4a7e6 | 1653 | if (limit >= chunk) |
ae8064ac | 1654 | goto send_now; |
c1b4a7e6 DM |
1655 | } else { |
1656 | /* Different approach, try not to defer past a single | |
1657 | * ACK. Receiver should ACK every other full sized | |
1658 | * frame, so if we have space for more than 3 frames | |
1659 | * then send now. | |
1660 | */ | |
6b5a5c0d | 1661 | if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) |
ae8064ac | 1662 | goto send_now; |
c1b4a7e6 DM |
1663 | } |
1664 | ||
f4541d60 ED |
1665 | /* Ok, it looks like it is advisable to defer. |
1666 | * Do not rearm the timer if already set to not break TCP ACK clocking. | |
1667 | */ | |
1668 | if (!tp->tso_deferred) | |
1669 | tp->tso_deferred = 1 | (jiffies << 1); | |
ae8064ac | 1670 | |
a2a385d6 | 1671 | return true; |
ae8064ac JH |
1672 | |
1673 | send_now: | |
1674 | tp->tso_deferred = 0; | |
a2a385d6 | 1675 | return false; |
c1b4a7e6 DM |
1676 | } |
1677 | ||
5d424d5a | 1678 | /* Create a new MTU probe if we are ready. |
67edfef7 AK |
1679 | * MTU probe is regularly attempting to increase the path MTU by |
1680 | * deliberately sending larger packets. This discovers routing | |
1681 | * changes resulting in larger path MTUs. | |
1682 | * | |
5d424d5a JH |
1683 | * Returns 0 if we should wait to probe (no cwnd available), |
1684 | * 1 if a probe was sent, | |
056834d9 IJ |
1685 | * -1 otherwise |
1686 | */ | |
5d424d5a JH |
1687 | static int tcp_mtu_probe(struct sock *sk) |
1688 | { | |
1689 | struct tcp_sock *tp = tcp_sk(sk); | |
1690 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1691 | struct sk_buff *skb, *nskb, *next; | |
1692 | int len; | |
1693 | int probe_size; | |
91cc17c0 | 1694 | int size_needed; |
5d424d5a JH |
1695 | int copy; |
1696 | int mss_now; | |
1697 | ||
1698 | /* Not currently probing/verifying, | |
1699 | * not in recovery, | |
1700 | * have enough cwnd, and | |
1701 | * not SACKing (the variable headers throw things off) */ | |
1702 | if (!icsk->icsk_mtup.enabled || | |
1703 | icsk->icsk_mtup.probe_size || | |
1704 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || | |
1705 | tp->snd_cwnd < 11 || | |
cabeccbd | 1706 | tp->rx_opt.num_sacks || tp->rx_opt.dsack) |
5d424d5a JH |
1707 | return -1; |
1708 | ||
1709 | /* Very simple search strategy: just double the MSS. */ | |
0c54b85f | 1710 | mss_now = tcp_current_mss(sk); |
056834d9 | 1711 | probe_size = 2 * tp->mss_cache; |
91cc17c0 | 1712 | size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; |
5d424d5a JH |
1713 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { |
1714 | /* TODO: set timer for probe_converge_event */ | |
1715 | return -1; | |
1716 | } | |
1717 | ||
1718 | /* Have enough data in the send queue to probe? */ | |
7f9c33e5 | 1719 | if (tp->write_seq - tp->snd_nxt < size_needed) |
5d424d5a JH |
1720 | return -1; |
1721 | ||
91cc17c0 IJ |
1722 | if (tp->snd_wnd < size_needed) |
1723 | return -1; | |
90840def | 1724 | if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) |
91cc17c0 | 1725 | return 0; |
5d424d5a | 1726 | |
d67c58e9 IJ |
1727 | /* Do we need to wait to drain cwnd? With none in flight, don't stall */ |
1728 | if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { | |
1729 | if (!tcp_packets_in_flight(tp)) | |
5d424d5a JH |
1730 | return -1; |
1731 | else | |
1732 | return 0; | |
1733 | } | |
1734 | ||
1735 | /* We're allowed to probe. Build it now. */ | |
1736 | if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) | |
1737 | return -1; | |
3ab224be HA |
1738 | sk->sk_wmem_queued += nskb->truesize; |
1739 | sk_mem_charge(sk, nskb->truesize); | |
5d424d5a | 1740 | |
fe067e8a | 1741 | skb = tcp_send_head(sk); |
5d424d5a JH |
1742 | |
1743 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; | |
1744 | TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; | |
4de075e0 | 1745 | TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; |
5d424d5a JH |
1746 | TCP_SKB_CB(nskb)->sacked = 0; |
1747 | nskb->csum = 0; | |
84fa7933 | 1748 | nskb->ip_summed = skb->ip_summed; |
5d424d5a | 1749 | |
50c4817e IJ |
1750 | tcp_insert_write_queue_before(nskb, skb, sk); |
1751 | ||
5d424d5a | 1752 | len = 0; |
234b6860 | 1753 | tcp_for_write_queue_from_safe(skb, next, sk) { |
5d424d5a | 1754 | copy = min_t(int, skb->len, probe_size - len); |
d318f82f | 1755 | if (nskb->ip_summed) { |
5d424d5a | 1756 | skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); |
d318f82f DCS |
1757 | } else { |
1758 | __wsum csum = skb_copy_and_csum_bits(skb, 0, | |
1759 | skb_put(nskb, copy), | |
1760 | copy, 0); | |
1761 | nskb->csum = csum_block_add(nskb->csum, csum, len); | |
1762 | } | |
5d424d5a JH |
1763 | |
1764 | if (skb->len <= copy) { | |
1765 | /* We've eaten all the data from this skb. | |
1766 | * Throw it away. */ | |
4de075e0 | 1767 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; |
fe067e8a | 1768 | tcp_unlink_write_queue(skb, sk); |
3ab224be | 1769 | sk_wmem_free_skb(sk, skb); |
5d424d5a | 1770 | } else { |
4de075e0 | 1771 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & |
a3433f35 | 1772 | ~(TCPHDR_FIN|TCPHDR_PSH); |
5d424d5a JH |
1773 | if (!skb_shinfo(skb)->nr_frags) { |
1774 | skb_pull(skb, copy); | |
84fa7933 | 1775 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
056834d9 IJ |
1776 | skb->csum = csum_partial(skb->data, |
1777 | skb->len, 0); | |
5d424d5a JH |
1778 | } else { |
1779 | __pskb_trim_head(skb, copy); | |
1780 | tcp_set_skb_tso_segs(sk, skb, mss_now); | |
1781 | } | |
1782 | TCP_SKB_CB(skb)->seq += copy; | |
1783 | } | |
1784 | ||
1785 | len += copy; | |
234b6860 IJ |
1786 | |
1787 | if (len >= probe_size) | |
1788 | break; | |
5d424d5a JH |
1789 | } |
1790 | tcp_init_tso_segs(sk, nskb, nskb->len); | |
1791 | ||
1792 | /* We're ready to send. If this fails, the probe will | |
1793 | * be resegmented into mss-sized pieces by tcp_write_xmit(). */ | |
1794 | TCP_SKB_CB(nskb)->when = tcp_time_stamp; | |
1795 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { | |
1796 | /* Decrement cwnd here because we are sending | |
056834d9 | 1797 | * effectively two packets. */ |
5d424d5a | 1798 | tp->snd_cwnd--; |
66f5fe62 | 1799 | tcp_event_new_data_sent(sk, nskb); |
5d424d5a JH |
1800 | |
1801 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); | |
0e7b1368 JH |
1802 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; |
1803 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; | |
5d424d5a JH |
1804 | |
1805 | return 1; | |
1806 | } | |
1807 | ||
1808 | return -1; | |
1809 | } | |
1810 | ||
1da177e4 LT |
1811 | /* This routine writes packets to the network. It advances the |
1812 | * send_head. This happens as incoming acks open up the remote | |
1813 | * window for us. | |
1814 | * | |
f8269a49 IJ |
1815 | * LARGESEND note: !tcp_urg_mode is overkill, only frames between |
1816 | * snd_up-64k-mss .. snd_up cannot be large. However, taking into | |
1817 | * account rare use of URG, this is not a big flaw. | |
1818 | * | |
6ba8a3b1 ND |
1819 | * Send at most one packet when push_one > 0. Temporarily ignore |
1820 | * cwnd limit to force at most one packet out when push_one == 2. | |
1821 | ||
a2a385d6 ED |
1822 | * Returns true, if no segments are in flight and we have queued segments, |
1823 | * but cannot send anything now because of SWS or another problem. | |
1da177e4 | 1824 | */ |
a2a385d6 ED |
1825 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
1826 | int push_one, gfp_t gfp) | |
1da177e4 LT |
1827 | { |
1828 | struct tcp_sock *tp = tcp_sk(sk); | |
92df7b51 | 1829 | struct sk_buff *skb; |
c1b4a7e6 DM |
1830 | unsigned int tso_segs, sent_pkts; |
1831 | int cwnd_quota; | |
5d424d5a | 1832 | int result; |
1da177e4 | 1833 | |
92df7b51 | 1834 | sent_pkts = 0; |
5d424d5a | 1835 | |
d5dd9175 IJ |
1836 | if (!push_one) { |
1837 | /* Do MTU probing. */ | |
1838 | result = tcp_mtu_probe(sk); | |
1839 | if (!result) { | |
a2a385d6 | 1840 | return false; |
d5dd9175 IJ |
1841 | } else if (result > 0) { |
1842 | sent_pkts = 1; | |
1843 | } | |
5d424d5a JH |
1844 | } |
1845 | ||
fe067e8a | 1846 | while ((skb = tcp_send_head(sk))) { |
c8ac3774 HX |
1847 | unsigned int limit; |
1848 | ||
b68e9f85 | 1849 | tso_segs = tcp_init_tso_segs(sk, skb, mss_now); |
c1b4a7e6 | 1850 | BUG_ON(!tso_segs); |
aa93466b | 1851 | |
ec342325 AV |
1852 | if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) |
1853 | goto repair; /* Skip network transmission */ | |
1854 | ||
b68e9f85 | 1855 | cwnd_quota = tcp_cwnd_test(tp, skb); |
6ba8a3b1 ND |
1856 | if (!cwnd_quota) { |
1857 | if (push_one == 2) | |
1858 | /* Force out a loss probe pkt. */ | |
1859 | cwnd_quota = 1; | |
1860 | else | |
1861 | break; | |
1862 | } | |
b68e9f85 HX |
1863 | |
1864 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) | |
1865 | break; | |
1866 | ||
afd67b68 | 1867 | if (tso_segs == 1 || !sk->sk_gso_max_segs) { |
c1b4a7e6 DM |
1868 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, |
1869 | (tcp_skb_is_last(sk, skb) ? | |
1870 | nonagle : TCP_NAGLE_PUSH)))) | |
1871 | break; | |
1872 | } else { | |
d5dd9175 | 1873 | if (!push_one && tcp_tso_should_defer(sk, skb)) |
c1b4a7e6 DM |
1874 | break; |
1875 | } | |
aa93466b | 1876 | |
0ae5f47e ED |
1877 | /* TCP Small Queues : |
1878 | * Control number of packets in qdisc/devices to two packets / or ~1 ms. | |
1879 | * This allows for : | |
1880 | * - better RTT estimation and ACK scheduling | |
1881 | * - faster recovery | |
1882 | * - high rates | |
6ef30bda ED |
1883 | * Alas, some drivers / subsystems require a fair amount |
1884 | * of queued bytes to ensure line rate. | |
1885 | * One example is wifi aggregation (802.11 AMPDU) | |
46d3ceab | 1886 | */ |
6ef30bda ED |
1887 | limit = max_t(unsigned int, sysctl_tcp_limit_output_bytes, |
1888 | sk->sk_pacing_rate >> 10); | |
0ae5f47e ED |
1889 | |
1890 | if (atomic_read(&sk->sk_wmem_alloc) > limit) { | |
46d3ceab | 1891 | set_bit(TSQ_THROTTLED, &tp->tsq_flags); |
94ee16ae JO |
1892 | /* It is possible TX completion already happened |
1893 | * before we set TSQ_THROTTLED, so we must | |
1894 | * test again the condition. | |
1895 | * We abuse smp_mb__after_clear_bit() because | |
1896 | * there is no smp_mb__after_set_bit() yet | |
1897 | */ | |
1898 | smp_mb__after_clear_bit(); | |
1899 | if (atomic_read(&sk->sk_wmem_alloc) > limit) | |
1900 | break; | |
46d3ceab | 1901 | } |
0ae5f47e | 1902 | |
c8ac3774 | 1903 | limit = mss_now; |
afd67b68 | 1904 | if (tso_segs > 1 && sk->sk_gso_max_segs && !tcp_urg_mode(tp)) |
0e3a4803 | 1905 | limit = tcp_mss_split_point(sk, skb, mss_now, |
1485348d BH |
1906 | min_t(unsigned int, |
1907 | cwnd_quota, | |
1908 | sk->sk_gso_max_segs)); | |
1da177e4 | 1909 | |
c8ac3774 | 1910 | if (skb->len > limit && |
c4ead4c5 | 1911 | unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) |
c8ac3774 HX |
1912 | break; |
1913 | ||
92df7b51 | 1914 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
c1b4a7e6 | 1915 | |
d5dd9175 | 1916 | if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) |
92df7b51 | 1917 | break; |
1da177e4 | 1918 | |
ec342325 | 1919 | repair: |
92df7b51 DM |
1920 | /* Advance the send_head. This one is sent out. |
1921 | * This call will increment packets_out. | |
1922 | */ | |
66f5fe62 | 1923 | tcp_event_new_data_sent(sk, skb); |
1da177e4 | 1924 | |
92df7b51 | 1925 | tcp_minshall_update(tp, mss_now, skb); |
a262f0cd | 1926 | sent_pkts += tcp_skb_pcount(skb); |
d5dd9175 IJ |
1927 | |
1928 | if (push_one) | |
1929 | break; | |
92df7b51 | 1930 | } |
1da177e4 | 1931 | |
aa93466b | 1932 | if (likely(sent_pkts)) { |
684bad11 YC |
1933 | if (tcp_in_cwnd_reduction(sk)) |
1934 | tp->prr_out += sent_pkts; | |
6ba8a3b1 ND |
1935 | |
1936 | /* Send one loss probe per tail loss episode. */ | |
1937 | if (push_one != 2) | |
1938 | tcp_schedule_loss_probe(sk); | |
9e412ba7 | 1939 | tcp_cwnd_validate(sk); |
a2a385d6 | 1940 | return false; |
1da177e4 | 1941 | } |
6ba8a3b1 ND |
1942 | return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk)); |
1943 | } | |
1944 | ||
1945 | bool tcp_schedule_loss_probe(struct sock *sk) | |
1946 | { | |
1947 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1948 | struct tcp_sock *tp = tcp_sk(sk); | |
1949 | u32 timeout, tlp_time_stamp, rto_time_stamp; | |
1950 | u32 rtt = tp->srtt >> 3; | |
1951 | ||
1952 | if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) | |
1953 | return false; | |
1954 | /* No consecutive loss probes. */ | |
1955 | if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { | |
1956 | tcp_rearm_rto(sk); | |
1957 | return false; | |
1958 | } | |
1959 | /* Don't do any loss probe on a Fast Open connection before 3WHS | |
1960 | * finishes. | |
1961 | */ | |
1962 | if (sk->sk_state == TCP_SYN_RECV) | |
1963 | return false; | |
1964 | ||
1965 | /* TLP is only scheduled when next timer event is RTO. */ | |
1966 | if (icsk->icsk_pending != ICSK_TIME_RETRANS) | |
1967 | return false; | |
1968 | ||
1969 | /* Schedule a loss probe in 2*RTT for SACK capable connections | |
1970 | * in Open state, that are either limited by cwnd or application. | |
1971 | */ | |
1972 | if (sysctl_tcp_early_retrans < 3 || !rtt || !tp->packets_out || | |
1973 | !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) | |
1974 | return false; | |
1975 | ||
1976 | if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && | |
1977 | tcp_send_head(sk)) | |
1978 | return false; | |
1979 | ||
1980 | /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account | |
1981 | * for delayed ack when there's one outstanding packet. | |
1982 | */ | |
1983 | timeout = rtt << 1; | |
1984 | if (tp->packets_out == 1) | |
1985 | timeout = max_t(u32, timeout, | |
1986 | (rtt + (rtt >> 1) + TCP_DELACK_MAX)); | |
1987 | timeout = max_t(u32, timeout, msecs_to_jiffies(10)); | |
1988 | ||
1989 | /* If RTO is shorter, just schedule TLP in its place. */ | |
1990 | tlp_time_stamp = tcp_time_stamp + timeout; | |
1991 | rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; | |
1992 | if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { | |
1993 | s32 delta = rto_time_stamp - tcp_time_stamp; | |
1994 | if (delta > 0) | |
1995 | timeout = delta; | |
1996 | } | |
1997 | ||
1998 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, | |
6fa3eb70 | 1999 | sysctl_tcp_rto_max); |
6ba8a3b1 ND |
2000 | return true; |
2001 | } | |
2002 | ||
2003 | /* When probe timeout (PTO) fires, send a new segment if one exists, else | |
2004 | * retransmit the last segment. | |
2005 | */ | |
2006 | void tcp_send_loss_probe(struct sock *sk) | |
2007 | { | |
9b717a8d | 2008 | struct tcp_sock *tp = tcp_sk(sk); |
6ba8a3b1 ND |
2009 | struct sk_buff *skb; |
2010 | int pcount; | |
2011 | int mss = tcp_current_mss(sk); | |
2012 | int err = -1; | |
2013 | ||
2014 | if (tcp_send_head(sk) != NULL) { | |
2015 | err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); | |
2016 | goto rearm_timer; | |
2017 | } | |
2018 | ||
9b717a8d ND |
2019 | /* At most one outstanding TLP retransmission. */ |
2020 | if (tp->tlp_high_seq) | |
2021 | goto rearm_timer; | |
2022 | ||
6ba8a3b1 ND |
2023 | /* Retransmit last segment. */ |
2024 | skb = tcp_write_queue_tail(sk); | |
2025 | if (WARN_ON(!skb)) | |
2026 | goto rearm_timer; | |
2027 | ||
2028 | pcount = tcp_skb_pcount(skb); | |
2029 | if (WARN_ON(!pcount)) | |
2030 | goto rearm_timer; | |
2031 | ||
2032 | if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { | |
2033 | if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss))) | |
2034 | goto rearm_timer; | |
2035 | skb = tcp_write_queue_tail(sk); | |
2036 | } | |
2037 | ||
2038 | if (WARN_ON(!skb || !tcp_skb_pcount(skb))) | |
2039 | goto rearm_timer; | |
2040 | ||
dc5a1700 | 2041 | err = __tcp_retransmit_skb(sk, skb); |
6ba8a3b1 | 2042 | |
9b717a8d ND |
2043 | /* Record snd_nxt for loss detection. */ |
2044 | if (likely(!err)) | |
2045 | tp->tlp_high_seq = tp->snd_nxt; | |
2046 | ||
6ba8a3b1 ND |
2047 | rearm_timer: |
2048 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, | |
2049 | inet_csk(sk)->icsk_rto, | |
6fa3eb70 | 2050 | sysctl_tcp_rto_max); |
6ba8a3b1 ND |
2051 | |
2052 | if (likely(!err)) | |
2053 | NET_INC_STATS_BH(sock_net(sk), | |
2054 | LINUX_MIB_TCPLOSSPROBES); | |
2055 | return; | |
1da177e4 LT |
2056 | } |
2057 | ||
a762a980 DM |
2058 | /* Push out any pending frames which were held back due to |
2059 | * TCP_CORK or attempt at coalescing tiny packets. | |
2060 | * The socket must be locked by the caller. | |
2061 | */ | |
9e412ba7 IJ |
2062 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, |
2063 | int nonagle) | |
a762a980 | 2064 | { |
726e07a8 IJ |
2065 | /* If we are closed, the bytes will have to remain here. |
2066 | * In time closedown will finish, we empty the write queue and | |
2067 | * all will be happy. | |
2068 | */ | |
2069 | if (unlikely(sk->sk_state == TCP_CLOSE)) | |
2070 | return; | |
2071 | ||
99a1dec7 MG |
2072 | if (tcp_write_xmit(sk, cur_mss, nonagle, 0, |
2073 | sk_gfp_atomic(sk, GFP_ATOMIC))) | |
726e07a8 | 2074 | tcp_check_probe_timer(sk); |
a762a980 DM |
2075 | } |
2076 | ||
c1b4a7e6 DM |
2077 | /* Send _single_ skb sitting at the send head. This function requires |
2078 | * true push pending frames to setup probe timer etc. | |
2079 | */ | |
2080 | void tcp_push_one(struct sock *sk, unsigned int mss_now) | |
2081 | { | |
fe067e8a | 2082 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 DM |
2083 | |
2084 | BUG_ON(!skb || skb->len < mss_now); | |
2085 | ||
d5dd9175 | 2086 | tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); |
c1b4a7e6 DM |
2087 | } |
2088 | ||
1da177e4 LT |
2089 | /* This function returns the amount that we can raise the |
2090 | * usable window based on the following constraints | |
e905a9ed | 2091 | * |
1da177e4 LT |
2092 | * 1. The window can never be shrunk once it is offered (RFC 793) |
2093 | * 2. We limit memory per socket | |
2094 | * | |
2095 | * RFC 1122: | |
2096 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep | |
2097 | * RECV.NEXT + RCV.WIN fixed until: | |
2098 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" | |
2099 | * | |
2100 | * i.e. don't raise the right edge of the window until you can raise | |
2101 | * it at least MSS bytes. | |
2102 | * | |
2103 | * Unfortunately, the recommended algorithm breaks header prediction, | |
2104 | * since header prediction assumes th->window stays fixed. | |
2105 | * | |
2106 | * Strictly speaking, keeping th->window fixed violates the receiver | |
2107 | * side SWS prevention criteria. The problem is that under this rule | |
2108 | * a stream of single byte packets will cause the right side of the | |
2109 | * window to always advance by a single byte. | |
e905a9ed | 2110 | * |
1da177e4 LT |
2111 | * Of course, if the sender implements sender side SWS prevention |
2112 | * then this will not be a problem. | |
e905a9ed | 2113 | * |
1da177e4 | 2114 | * BSD seems to make the following compromise: |
e905a9ed | 2115 | * |
1da177e4 LT |
2116 | * If the free space is less than the 1/4 of the maximum |
2117 | * space available and the free space is less than 1/2 mss, | |
2118 | * then set the window to 0. | |
2119 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] | |
2120 | * Otherwise, just prevent the window from shrinking | |
2121 | * and from being larger than the largest representable value. | |
2122 | * | |
2123 | * This prevents incremental opening of the window in the regime | |
2124 | * where TCP is limited by the speed of the reader side taking | |
2125 | * data out of the TCP receive queue. It does nothing about | |
2126 | * those cases where the window is constrained on the sender side | |
2127 | * because the pipeline is full. | |
2128 | * | |
2129 | * BSD also seems to "accidentally" limit itself to windows that are a | |
2130 | * multiple of MSS, at least until the free space gets quite small. | |
2131 | * This would appear to be a side effect of the mbuf implementation. | |
2132 | * Combining these two algorithms results in the observed behavior | |
2133 | * of having a fixed window size at almost all times. | |
2134 | * | |
2135 | * Below we obtain similar behavior by forcing the offered window to | |
2136 | * a multiple of the mss when it is feasible to do so. | |
2137 | * | |
2138 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. | |
2139 | * Regular options like TIMESTAMP are taken into account. | |
2140 | */ | |
2141 | u32 __tcp_select_window(struct sock *sk) | |
2142 | { | |
463c84b9 | 2143 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 2144 | struct tcp_sock *tp = tcp_sk(sk); |
caa20d9a | 2145 | /* MSS for the peer's data. Previous versions used mss_clamp |
1da177e4 LT |
2146 | * here. I don't know if the value based on our guesses |
2147 | * of peer's MSS is better for the performance. It's more correct | |
2148 | * but may be worse for the performance because of rcv_mss | |
2149 | * fluctuations. --SAW 1998/11/1 | |
2150 | */ | |
463c84b9 | 2151 | int mss = icsk->icsk_ack.rcv_mss; |
1da177e4 LT |
2152 | int free_space = tcp_space(sk); |
2153 | int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); | |
2154 | int window; | |
2155 | ||
5c235ec4 | 2156 | if (unlikely(mss > full_space)) { |
e905a9ed | 2157 | mss = full_space; |
5c235ec4 ED |
2158 | if (mss <= 0) |
2159 | return 0; | |
2160 | } | |
b92edbe0 | 2161 | if (free_space < (full_space >> 1)) { |
463c84b9 | 2162 | icsk->icsk_ack.quick = 0; |
1da177e4 | 2163 | |
180d8cd9 | 2164 | if (sk_under_memory_pressure(sk)) |
056834d9 IJ |
2165 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, |
2166 | 4U * tp->advmss); | |
1da177e4 LT |
2167 | |
2168 | if (free_space < mss) | |
2169 | return 0; | |
2170 | } | |
2171 | ||
2172 | if (free_space > tp->rcv_ssthresh) | |
2173 | free_space = tp->rcv_ssthresh; | |
2174 | ||
2175 | /* Don't do rounding if we are using window scaling, since the | |
2176 | * scaled window will not line up with the MSS boundary anyway. | |
2177 | */ | |
2178 | window = tp->rcv_wnd; | |
2179 | if (tp->rx_opt.rcv_wscale) { | |
2180 | window = free_space; | |
2181 | ||
2182 | /* Advertise enough space so that it won't get scaled away. | |
2183 | * Import case: prevent zero window announcement if | |
2184 | * 1<<rcv_wscale > mss. | |
2185 | */ | |
2186 | if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) | |
2187 | window = (((window >> tp->rx_opt.rcv_wscale) + 1) | |
2188 | << tp->rx_opt.rcv_wscale); | |
2189 | } else { | |
2190 | /* Get the largest window that is a nice multiple of mss. | |
2191 | * Window clamp already applied above. | |
2192 | * If our current window offering is within 1 mss of the | |
2193 | * free space we just keep it. This prevents the divide | |
2194 | * and multiply from happening most of the time. | |
2195 | * We also don't do any window rounding when the free space | |
2196 | * is too small. | |
2197 | */ | |
2198 | if (window <= free_space - mss || window > free_space) | |
056834d9 | 2199 | window = (free_space / mss) * mss; |
84565070 | 2200 | else if (mss == full_space && |
b92edbe0 | 2201 | free_space > window + (full_space >> 1)) |
84565070 | 2202 | window = free_space; |
1da177e4 LT |
2203 | } |
2204 | ||
2205 | return window; | |
2206 | } | |
2207 | ||
4a17fc3a IJ |
2208 | /* Collapses two adjacent SKB's during retransmission. */ |
2209 | static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) | |
1da177e4 LT |
2210 | { |
2211 | struct tcp_sock *tp = tcp_sk(sk); | |
fe067e8a | 2212 | struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); |
058dc334 | 2213 | int skb_size, next_skb_size; |
1da177e4 | 2214 | |
058dc334 IJ |
2215 | skb_size = skb->len; |
2216 | next_skb_size = next_skb->len; | |
1da177e4 | 2217 | |
058dc334 | 2218 | BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); |
a6963a6b | 2219 | |
058dc334 | 2220 | tcp_highest_sack_combine(sk, next_skb, skb); |
1da177e4 | 2221 | |
058dc334 | 2222 | tcp_unlink_write_queue(next_skb, sk); |
1da177e4 | 2223 | |
058dc334 IJ |
2224 | skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), |
2225 | next_skb_size); | |
1da177e4 | 2226 | |
058dc334 IJ |
2227 | if (next_skb->ip_summed == CHECKSUM_PARTIAL) |
2228 | skb->ip_summed = CHECKSUM_PARTIAL; | |
1da177e4 | 2229 | |
058dc334 IJ |
2230 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
2231 | skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); | |
1da177e4 | 2232 | |
058dc334 IJ |
2233 | /* Update sequence range on original skb. */ |
2234 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; | |
1da177e4 | 2235 | |
e6c7d085 | 2236 | /* Merge over control information. This moves PSH/FIN etc. over */ |
4de075e0 | 2237 | TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; |
058dc334 IJ |
2238 | |
2239 | /* All done, get rid of second SKB and account for it so | |
2240 | * packet counting does not break. | |
2241 | */ | |
2242 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; | |
058dc334 IJ |
2243 | |
2244 | /* changed transmit queue under us so clear hints */ | |
ef9da47c IJ |
2245 | tcp_clear_retrans_hints_partial(tp); |
2246 | if (next_skb == tp->retransmit_skb_hint) | |
2247 | tp->retransmit_skb_hint = skb; | |
058dc334 | 2248 | |
797108d1 IJ |
2249 | tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); |
2250 | ||
058dc334 | 2251 | sk_wmem_free_skb(sk, next_skb); |
1da177e4 LT |
2252 | } |
2253 | ||
67edfef7 | 2254 | /* Check if coalescing SKBs is legal. */ |
a2a385d6 | 2255 | static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) |
4a17fc3a IJ |
2256 | { |
2257 | if (tcp_skb_pcount(skb) > 1) | |
a2a385d6 | 2258 | return false; |
4a17fc3a IJ |
2259 | /* TODO: SACK collapsing could be used to remove this condition */ |
2260 | if (skb_shinfo(skb)->nr_frags != 0) | |
a2a385d6 | 2261 | return false; |
4a17fc3a | 2262 | if (skb_cloned(skb)) |
a2a385d6 | 2263 | return false; |
4a17fc3a | 2264 | if (skb == tcp_send_head(sk)) |
a2a385d6 | 2265 | return false; |
4a17fc3a IJ |
2266 | /* Some heurestics for collapsing over SACK'd could be invented */ |
2267 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
a2a385d6 | 2268 | return false; |
4a17fc3a | 2269 | |
a2a385d6 | 2270 | return true; |
4a17fc3a IJ |
2271 | } |
2272 | ||
67edfef7 AK |
2273 | /* Collapse packets in the retransmit queue to make to create |
2274 | * less packets on the wire. This is only done on retransmission. | |
2275 | */ | |
4a17fc3a IJ |
2276 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, |
2277 | int space) | |
2278 | { | |
2279 | struct tcp_sock *tp = tcp_sk(sk); | |
2280 | struct sk_buff *skb = to, *tmp; | |
a2a385d6 | 2281 | bool first = true; |
4a17fc3a IJ |
2282 | |
2283 | if (!sysctl_tcp_retrans_collapse) | |
2284 | return; | |
4de075e0 | 2285 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) |
4a17fc3a IJ |
2286 | return; |
2287 | ||
2288 | tcp_for_write_queue_from_safe(skb, tmp, sk) { | |
2289 | if (!tcp_can_collapse(sk, skb)) | |
2290 | break; | |
2291 | ||
2292 | space -= skb->len; | |
2293 | ||
2294 | if (first) { | |
a2a385d6 | 2295 | first = false; |
4a17fc3a IJ |
2296 | continue; |
2297 | } | |
2298 | ||
2299 | if (space < 0) | |
2300 | break; | |
2301 | /* Punt if not enough space exists in the first SKB for | |
2302 | * the data in the second | |
2303 | */ | |
a21d4572 | 2304 | if (skb->len > skb_availroom(to)) |
4a17fc3a IJ |
2305 | break; |
2306 | ||
2307 | if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) | |
2308 | break; | |
2309 | ||
2310 | tcp_collapse_retrans(sk, to); | |
2311 | } | |
2312 | } | |
2313 | ||
1da177e4 LT |
2314 | /* This retransmits one SKB. Policy decisions and retransmit queue |
2315 | * state updates are done by the caller. Returns non-zero if an | |
2316 | * error occurred which prevented the send. | |
2317 | */ | |
93b174ad | 2318 | int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) |
1da177e4 LT |
2319 | { |
2320 | struct tcp_sock *tp = tcp_sk(sk); | |
5d424d5a | 2321 | struct inet_connection_sock *icsk = inet_csk(sk); |
7d227cd2 | 2322 | unsigned int cur_mss; |
1da177e4 | 2323 | |
5d424d5a JH |
2324 | /* Inconslusive MTU probe */ |
2325 | if (icsk->icsk_mtup.probe_size) { | |
2326 | icsk->icsk_mtup.probe_size = 0; | |
2327 | } | |
2328 | ||
1da177e4 | 2329 | /* Do not sent more than we queued. 1/4 is reserved for possible |
caa20d9a | 2330 | * copying overhead: fragmentation, tunneling, mangling etc. |
1da177e4 LT |
2331 | */ |
2332 | if (atomic_read(&sk->sk_wmem_alloc) > | |
93522d31 ED |
2333 | min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), |
2334 | sk->sk_sndbuf)) | |
1da177e4 LT |
2335 | return -EAGAIN; |
2336 | ||
2337 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { | |
2338 | if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
2339 | BUG(); | |
1da177e4 LT |
2340 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
2341 | return -ENOMEM; | |
2342 | } | |
2343 | ||
7d227cd2 SS |
2344 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) |
2345 | return -EHOSTUNREACH; /* Routing failure or similar. */ | |
2346 | ||
0c54b85f | 2347 | cur_mss = tcp_current_mss(sk); |
7d227cd2 | 2348 | |
1da177e4 LT |
2349 | /* If receiver has shrunk his window, and skb is out of |
2350 | * new window, do not retransmit it. The exception is the | |
2351 | * case, when window is shrunk to zero. In this case | |
2352 | * our retransmit serves as a zero window probe. | |
2353 | */ | |
9d4fb27d JP |
2354 | if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && |
2355 | TCP_SKB_CB(skb)->seq != tp->snd_una) | |
1da177e4 LT |
2356 | return -EAGAIN; |
2357 | ||
2358 | if (skb->len > cur_mss) { | |
846998ae | 2359 | if (tcp_fragment(sk, skb, cur_mss, cur_mss)) |
1da177e4 | 2360 | return -ENOMEM; /* We'll try again later. */ |
02276f3c | 2361 | } else { |
9eb9362e IJ |
2362 | int oldpcount = tcp_skb_pcount(skb); |
2363 | ||
2364 | if (unlikely(oldpcount > 1)) { | |
b81908e1 ED |
2365 | if (skb_unclone(skb, GFP_ATOMIC)) |
2366 | return -ENOMEM; | |
9eb9362e IJ |
2367 | tcp_init_tso_segs(sk, skb, cur_mss); |
2368 | tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); | |
2369 | } | |
1da177e4 LT |
2370 | } |
2371 | ||
4a17fc3a | 2372 | tcp_retrans_try_collapse(sk, skb, cur_mss); |
1da177e4 | 2373 | |
1da177e4 LT |
2374 | /* Some Solaris stacks overoptimize and ignore the FIN on a |
2375 | * retransmit when old data is attached. So strip it off | |
2376 | * since it is cheap to do so and saves bytes on the network. | |
2377 | */ | |
2de979bd | 2378 | if (skb->len > 0 && |
4de075e0 | 2379 | (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && |
2de979bd | 2380 | tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { |
1da177e4 | 2381 | if (!pskb_trim(skb, 0)) { |
e870a8ef IJ |
2382 | /* Reuse, even though it does some unnecessary work */ |
2383 | tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, | |
4de075e0 | 2384 | TCP_SKB_CB(skb)->tcp_flags); |
1da177e4 | 2385 | skb->ip_summed = CHECKSUM_NONE; |
1da177e4 LT |
2386 | } |
2387 | } | |
2388 | ||
2389 | /* Make a copy, if the first transmission SKB clone we made | |
2390 | * is still in somebody's hands, else make a clone. | |
2391 | */ | |
2392 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
1da177e4 | 2393 | |
50bceae9 TG |
2394 | /* make sure skb->data is aligned on arches that require it |
2395 | * and check if ack-trimming & collapsing extended the headroom | |
2396 | * beyond what csum_start can cover. | |
2397 | */ | |
2398 | if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || | |
2399 | skb_headroom(skb) >= 0xFFFF)) { | |
117632e6 ED |
2400 | struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, |
2401 | GFP_ATOMIC); | |
93b174ad YC |
2402 | return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : |
2403 | -ENOBUFS; | |
117632e6 | 2404 | } else { |
93b174ad | 2405 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
117632e6 | 2406 | } |
93b174ad YC |
2407 | } |
2408 | ||
2409 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) | |
2410 | { | |
2411 | struct tcp_sock *tp = tcp_sk(sk); | |
2412 | int err = __tcp_retransmit_skb(sk, skb); | |
1da177e4 LT |
2413 | |
2414 | if (err == 0) { | |
2415 | /* Update global TCP statistics. */ | |
81cc8a75 | 2416 | TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); |
1da177e4 LT |
2417 | |
2418 | tp->total_retrans++; | |
2419 | ||
2420 | #if FASTRETRANS_DEBUG > 0 | |
056834d9 | 2421 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
e87cc472 | 2422 | net_dbg_ratelimited("retrans_out leaked\n"); |
1da177e4 LT |
2423 | } |
2424 | #endif | |
b08d6cb2 IJ |
2425 | if (!tp->retrans_out) |
2426 | tp->lost_retrans_low = tp->snd_nxt; | |
1da177e4 LT |
2427 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; |
2428 | tp->retrans_out += tcp_skb_pcount(skb); | |
2429 | ||
2430 | /* Save stamp of the first retransmit. */ | |
2431 | if (!tp->retrans_stamp) | |
2432 | tp->retrans_stamp = TCP_SKB_CB(skb)->when; | |
2433 | ||
1da177e4 LT |
2434 | /* snd_nxt is stored to detect loss of retransmitted segment, |
2435 | * see tcp_input.c tcp_sacktag_write_queue(). | |
2436 | */ | |
2437 | TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; | |
2438 | } | |
76fd3c89 YC |
2439 | |
2440 | if (tp->undo_retrans < 0) | |
2441 | tp->undo_retrans = 0; | |
2442 | tp->undo_retrans += tcp_skb_pcount(skb); | |
1da177e4 LT |
2443 | return err; |
2444 | } | |
2445 | ||
67edfef7 AK |
2446 | /* Check if we forward retransmits are possible in the current |
2447 | * window/congestion state. | |
2448 | */ | |
a2a385d6 | 2449 | static bool tcp_can_forward_retransmit(struct sock *sk) |
b5afe7bc IJ |
2450 | { |
2451 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
cf533ea5 | 2452 | const struct tcp_sock *tp = tcp_sk(sk); |
b5afe7bc IJ |
2453 | |
2454 | /* Forward retransmissions are possible only during Recovery. */ | |
2455 | if (icsk->icsk_ca_state != TCP_CA_Recovery) | |
a2a385d6 | 2456 | return false; |
b5afe7bc IJ |
2457 | |
2458 | /* No forward retransmissions in Reno are possible. */ | |
2459 | if (tcp_is_reno(tp)) | |
a2a385d6 | 2460 | return false; |
b5afe7bc IJ |
2461 | |
2462 | /* Yeah, we have to make difficult choice between forward transmission | |
2463 | * and retransmission... Both ways have their merits... | |
2464 | * | |
2465 | * For now we do not retransmit anything, while we have some new | |
2466 | * segments to send. In the other cases, follow rule 3 for | |
2467 | * NextSeg() specified in RFC3517. | |
2468 | */ | |
2469 | ||
2470 | if (tcp_may_send_now(sk)) | |
a2a385d6 | 2471 | return false; |
b5afe7bc | 2472 | |
a2a385d6 | 2473 | return true; |
b5afe7bc IJ |
2474 | } |
2475 | ||
1da177e4 LT |
2476 | /* This gets called after a retransmit timeout, and the initially |
2477 | * retransmitted data is acknowledged. It tries to continue | |
2478 | * resending the rest of the retransmit queue, until either | |
2479 | * we've sent it all or the congestion window limit is reached. | |
2480 | * If doing SACK, the first ACK which comes back for a timeout | |
2481 | * based retransmit packet might feed us FACK information again. | |
2482 | * If so, we use it to avoid unnecessarily retransmissions. | |
2483 | */ | |
2484 | void tcp_xmit_retransmit_queue(struct sock *sk) | |
2485 | { | |
6687e988 | 2486 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2487 | struct tcp_sock *tp = tcp_sk(sk); |
2488 | struct sk_buff *skb; | |
0e1c54c2 | 2489 | struct sk_buff *hole = NULL; |
618d9f25 | 2490 | u32 last_lost; |
61eb55f4 | 2491 | int mib_idx; |
0e1c54c2 | 2492 | int fwd_rexmitting = 0; |
6a438bbe | 2493 | |
45e77d31 IJ |
2494 | if (!tp->packets_out) |
2495 | return; | |
2496 | ||
08ebd172 IJ |
2497 | if (!tp->lost_out) |
2498 | tp->retransmit_high = tp->snd_una; | |
2499 | ||
618d9f25 | 2500 | if (tp->retransmit_skb_hint) { |
6a438bbe | 2501 | skb = tp->retransmit_skb_hint; |
618d9f25 IJ |
2502 | last_lost = TCP_SKB_CB(skb)->end_seq; |
2503 | if (after(last_lost, tp->retransmit_high)) | |
2504 | last_lost = tp->retransmit_high; | |
2505 | } else { | |
fe067e8a | 2506 | skb = tcp_write_queue_head(sk); |
618d9f25 IJ |
2507 | last_lost = tp->snd_una; |
2508 | } | |
1da177e4 | 2509 | |
08ebd172 IJ |
2510 | tcp_for_write_queue_from(skb, sk) { |
2511 | __u8 sacked = TCP_SKB_CB(skb)->sacked; | |
1da177e4 | 2512 | |
08ebd172 IJ |
2513 | if (skb == tcp_send_head(sk)) |
2514 | break; | |
2515 | /* we could do better than to assign each time */ | |
0e1c54c2 IJ |
2516 | if (hole == NULL) |
2517 | tp->retransmit_skb_hint = skb; | |
08ebd172 IJ |
2518 | |
2519 | /* Assume this retransmit will generate | |
2520 | * only one packet for congestion window | |
2521 | * calculation purposes. This works because | |
2522 | * tcp_retransmit_skb() will chop up the | |
2523 | * packet to be MSS sized and all the | |
2524 | * packet counting works out. | |
2525 | */ | |
2526 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) | |
2527 | return; | |
1da177e4 | 2528 | |
0e1c54c2 IJ |
2529 | if (fwd_rexmitting) { |
2530 | begin_fwd: | |
2531 | if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) | |
2532 | break; | |
2533 | mib_idx = LINUX_MIB_TCPFORWARDRETRANS; | |
6a438bbe | 2534 | |
0e1c54c2 | 2535 | } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { |
618d9f25 | 2536 | tp->retransmit_high = last_lost; |
0e1c54c2 IJ |
2537 | if (!tcp_can_forward_retransmit(sk)) |
2538 | break; | |
2539 | /* Backtrack if necessary to non-L'ed skb */ | |
2540 | if (hole != NULL) { | |
2541 | skb = hole; | |
2542 | hole = NULL; | |
2543 | } | |
2544 | fwd_rexmitting = 1; | |
2545 | goto begin_fwd; | |
1da177e4 | 2546 | |
0e1c54c2 | 2547 | } else if (!(sacked & TCPCB_LOST)) { |
ac11ba75 | 2548 | if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) |
0e1c54c2 IJ |
2549 | hole = skb; |
2550 | continue; | |
1da177e4 | 2551 | |
0e1c54c2 | 2552 | } else { |
618d9f25 | 2553 | last_lost = TCP_SKB_CB(skb)->end_seq; |
0e1c54c2 IJ |
2554 | if (icsk->icsk_ca_state != TCP_CA_Loss) |
2555 | mib_idx = LINUX_MIB_TCPFASTRETRANS; | |
2556 | else | |
2557 | mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; | |
2558 | } | |
1da177e4 | 2559 | |
0e1c54c2 | 2560 | if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) |
1da177e4 LT |
2561 | continue; |
2562 | ||
09e9b813 ED |
2563 | if (tcp_retransmit_skb(sk, skb)) { |
2564 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); | |
0e1c54c2 | 2565 | return; |
09e9b813 | 2566 | } |
0e1c54c2 | 2567 | NET_INC_STATS_BH(sock_net(sk), mib_idx); |
1da177e4 | 2568 | |
684bad11 | 2569 | if (tcp_in_cwnd_reduction(sk)) |
a262f0cd ND |
2570 | tp->prr_out += tcp_skb_pcount(skb); |
2571 | ||
fe067e8a | 2572 | if (skb == tcp_write_queue_head(sk)) |
3f421baa ACM |
2573 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
2574 | inet_csk(sk)->icsk_rto, | |
6fa3eb70 | 2575 | sysctl_tcp_rto_max); |
1da177e4 LT |
2576 | } |
2577 | } | |
2578 | ||
aac9fda3 ED |
2579 | /* We allow to exceed memory limits for FIN packets to expedite |
2580 | * connection tear down and (memory) recovery. | |
bc9f0ea1 ED |
2581 | * Otherwise tcp_send_fin() could be tempted to either delay FIN |
2582 | * or even be forced to close flow without any FIN. | |
aac9fda3 ED |
2583 | */ |
2584 | static void sk_forced_wmem_schedule(struct sock *sk, int size) | |
2585 | { | |
2586 | int amt, status; | |
2587 | ||
2588 | if (size <= sk->sk_forward_alloc) | |
2589 | return; | |
2590 | amt = sk_mem_pages(size); | |
2591 | sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; | |
2592 | sk_memory_allocated_add(sk, amt, &status); | |
2593 | } | |
2594 | ||
bc9f0ea1 ED |
2595 | /* Send a FIN. The caller locks the socket for us. |
2596 | * We should try to send a FIN packet really hard, but eventually give up. | |
1da177e4 LT |
2597 | */ |
2598 | void tcp_send_fin(struct sock *sk) | |
2599 | { | |
bc9f0ea1 | 2600 | struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); |
e905a9ed | 2601 | struct tcp_sock *tp = tcp_sk(sk); |
e905a9ed | 2602 | |
bc9f0ea1 ED |
2603 | /* Optimization, tack on the FIN if we have one skb in write queue and |
2604 | * this skb was not yet sent, or we are under memory pressure. | |
2605 | * Note: in the latter case, FIN packet will be sent after a timeout, | |
2606 | * as TCP stack thinks it has already been transmitted. | |
1da177e4 | 2607 | */ |
bc9f0ea1 ED |
2608 | if (tskb && (tcp_send_head(sk) || sk_under_memory_pressure(sk))) { |
2609 | coalesce: | |
2610 | TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; | |
2611 | TCP_SKB_CB(tskb)->end_seq++; | |
1da177e4 | 2612 | tp->write_seq++; |
bc9f0ea1 ED |
2613 | if (!tcp_send_head(sk)) { |
2614 | /* This means tskb was already sent. | |
2615 | * Pretend we included the FIN on previous transmit. | |
2616 | * We need to set tp->snd_nxt to the value it would have | |
2617 | * if FIN had been sent. This is because retransmit path | |
2618 | * does not change tp->snd_nxt. | |
2619 | */ | |
2620 | tp->snd_nxt++; | |
2621 | return; | |
2622 | } | |
1da177e4 | 2623 | } else { |
bc9f0ea1 ED |
2624 | skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); |
2625 | if (unlikely(!skb)) { | |
2626 | if (tskb) | |
2627 | goto coalesce; | |
2628 | return; | |
1da177e4 | 2629 | } |
aac9fda3 ED |
2630 | skb_reserve(skb, MAX_TCP_HEADER); |
2631 | sk_forced_wmem_schedule(sk, skb->truesize); | |
1da177e4 | 2632 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ |
e870a8ef | 2633 | tcp_init_nondata_skb(skb, tp->write_seq, |
a3433f35 | 2634 | TCPHDR_ACK | TCPHDR_FIN); |
1da177e4 LT |
2635 | tcp_queue_skb(sk, skb); |
2636 | } | |
bc9f0ea1 | 2637 | __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); |
1da177e4 LT |
2638 | } |
2639 | ||
2640 | /* We get here when a process closes a file descriptor (either due to | |
2641 | * an explicit close() or as a byproduct of exit()'ing) and there | |
2642 | * was unread data in the receive queue. This behavior is recommended | |
65bb723c | 2643 | * by RFC 2525, section 2.17. -DaveM |
1da177e4 | 2644 | */ |
dd0fc66f | 2645 | void tcp_send_active_reset(struct sock *sk, gfp_t priority) |
1da177e4 | 2646 | { |
1da177e4 LT |
2647 | struct sk_buff *skb; |
2648 | ||
2649 | /* NOTE: No TCP options attached and we never retransmit this. */ | |
2650 | skb = alloc_skb(MAX_TCP_HEADER, priority); | |
2651 | if (!skb) { | |
4e673444 | 2652 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
1da177e4 LT |
2653 | return; |
2654 | } | |
2655 | ||
2656 | /* Reserve space for headers and prepare control bits. */ | |
2657 | skb_reserve(skb, MAX_TCP_HEADER); | |
e870a8ef | 2658 | tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), |
a3433f35 | 2659 | TCPHDR_ACK | TCPHDR_RST); |
1da177e4 | 2660 | /* Send it off. */ |
1da177e4 | 2661 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
dfb4b9dc | 2662 | if (tcp_transmit_skb(sk, skb, 0, priority)) |
4e673444 | 2663 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
26af65cb | 2664 | |
81cc8a75 | 2665 | TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); |
1da177e4 LT |
2666 | } |
2667 | ||
67edfef7 AK |
2668 | /* Send a crossed SYN-ACK during socket establishment. |
2669 | * WARNING: This routine must only be called when we have already sent | |
1da177e4 LT |
2670 | * a SYN packet that crossed the incoming SYN that caused this routine |
2671 | * to get called. If this assumption fails then the initial rcv_wnd | |
2672 | * and rcv_wscale values will not be correct. | |
2673 | */ | |
2674 | int tcp_send_synack(struct sock *sk) | |
2675 | { | |
056834d9 | 2676 | struct sk_buff *skb; |
1da177e4 | 2677 | |
fe067e8a | 2678 | skb = tcp_write_queue_head(sk); |
4de075e0 | 2679 | if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { |
91df42be | 2680 | pr_debug("%s: wrong queue state\n", __func__); |
1da177e4 LT |
2681 | return -EFAULT; |
2682 | } | |
4de075e0 | 2683 | if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { |
1da177e4 LT |
2684 | if (skb_cloned(skb)) { |
2685 | struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); | |
2686 | if (nskb == NULL) | |
2687 | return -ENOMEM; | |
fe067e8a | 2688 | tcp_unlink_write_queue(skb, sk); |
1da177e4 | 2689 | skb_header_release(nskb); |
fe067e8a | 2690 | __tcp_add_write_queue_head(sk, nskb); |
3ab224be HA |
2691 | sk_wmem_free_skb(sk, skb); |
2692 | sk->sk_wmem_queued += nskb->truesize; | |
2693 | sk_mem_charge(sk, nskb->truesize); | |
1da177e4 LT |
2694 | skb = nskb; |
2695 | } | |
2696 | ||
4de075e0 | 2697 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; |
1da177e4 LT |
2698 | TCP_ECN_send_synack(tcp_sk(sk), skb); |
2699 | } | |
2700 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
dfb4b9dc | 2701 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
1da177e4 LT |
2702 | } |
2703 | ||
4aea39c1 ED |
2704 | /** |
2705 | * tcp_make_synack - Prepare a SYN-ACK. | |
2706 | * sk: listener socket | |
2707 | * dst: dst entry attached to the SYNACK | |
2708 | * req: request_sock pointer | |
4aea39c1 ED |
2709 | * |
2710 | * Allocate one skb and build a SYNACK packet. | |
2711 | * @dst is consumed : Caller should not use it again. | |
2712 | */ | |
056834d9 | 2713 | struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, |
e6b4d113 | 2714 | struct request_sock *req, |
8336886f | 2715 | struct tcp_fastopen_cookie *foc) |
1da177e4 | 2716 | { |
bd0388ae | 2717 | struct tcp_out_options opts; |
2e6599cb | 2718 | struct inet_request_sock *ireq = inet_rsk(req); |
1da177e4 LT |
2719 | struct tcp_sock *tp = tcp_sk(sk); |
2720 | struct tcphdr *th; | |
1da177e4 | 2721 | struct sk_buff *skb; |
cfb6eeb4 | 2722 | struct tcp_md5sig_key *md5; |
bd0388ae | 2723 | int tcp_header_size; |
f5fff5dc | 2724 | int mss; |
1da177e4 | 2725 | |
b70a23ab | 2726 | skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15, 1, GFP_ATOMIC); |
4aea39c1 ED |
2727 | if (unlikely(!skb)) { |
2728 | dst_release(dst); | |
1da177e4 | 2729 | return NULL; |
4aea39c1 | 2730 | } |
1da177e4 LT |
2731 | /* Reserve space for headers. */ |
2732 | skb_reserve(skb, MAX_TCP_HEADER); | |
2733 | ||
4aea39c1 | 2734 | skb_dst_set(skb, dst); |
ca10b9e9 | 2735 | security_skb_owned_by(skb, sk); |
1da177e4 | 2736 | |
0dbaee3b | 2737 | mss = dst_metric_advmss(dst); |
f5fff5dc TQ |
2738 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) |
2739 | mss = tp->rx_opt.user_mss; | |
2740 | ||
33ad798c AL |
2741 | if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ |
2742 | __u8 rcv_wscale; | |
2743 | /* Set this up on the first call only */ | |
2744 | req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); | |
e88c64f0 HPP |
2745 | |
2746 | /* limit the window selection if the user enforce a smaller rx buffer */ | |
2747 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && | |
2748 | (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) | |
2749 | req->window_clamp = tcp_full_space(sk); | |
2750 | ||
33ad798c AL |
2751 | /* tcp_full_space because it is guaranteed to be the first packet */ |
2752 | tcp_select_initial_window(tcp_full_space(sk), | |
f5fff5dc | 2753 | mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), |
33ad798c AL |
2754 | &req->rcv_wnd, |
2755 | &req->window_clamp, | |
2756 | ireq->wscale_ok, | |
31d12926 | 2757 | &rcv_wscale, |
2758 | dst_metric(dst, RTAX_INITRWND)); | |
33ad798c AL |
2759 | ireq->rcv_wscale = rcv_wscale; |
2760 | } | |
2761 | ||
2762 | memset(&opts, 0, sizeof(opts)); | |
8b5f12d0 FW |
2763 | #ifdef CONFIG_SYN_COOKIES |
2764 | if (unlikely(req->cookie_ts)) | |
2765 | TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); | |
2766 | else | |
2767 | #endif | |
33ad798c | 2768 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
1a2c6181 CP |
2769 | tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5, |
2770 | foc) + sizeof(*th); | |
cfb6eeb4 | 2771 | |
aa8223c7 ACM |
2772 | skb_push(skb, tcp_header_size); |
2773 | skb_reset_transport_header(skb); | |
1da177e4 | 2774 | |
aa8223c7 | 2775 | th = tcp_hdr(skb); |
1da177e4 LT |
2776 | memset(th, 0, sizeof(struct tcphdr)); |
2777 | th->syn = 1; | |
2778 | th->ack = 1; | |
1da177e4 | 2779 | TCP_ECN_make_synack(req, th); |
a3116ac5 | 2780 | th->source = ireq->loc_port; |
2e6599cb | 2781 | th->dest = ireq->rmt_port; |
e870a8ef IJ |
2782 | /* Setting of flags are superfluous here for callers (and ECE is |
2783 | * not even correctly set) | |
2784 | */ | |
2785 | tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, | |
a3433f35 | 2786 | TCPHDR_SYN | TCPHDR_ACK); |
4957faad | 2787 | |
1da177e4 | 2788 | th->seq = htonl(TCP_SKB_CB(skb)->seq); |
8336886f JC |
2789 | /* XXX data is queued and acked as is. No buffer/window check */ |
2790 | th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); | |
1da177e4 LT |
2791 | |
2792 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ | |
600ff0c2 | 2793 | th->window = htons(min(req->rcv_wnd, 65535U)); |
bd0388ae | 2794 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
1da177e4 | 2795 | th->doff = (tcp_header_size >> 2); |
aa2ea058 | 2796 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb)); |
cfb6eeb4 YH |
2797 | |
2798 | #ifdef CONFIG_TCP_MD5SIG | |
2799 | /* Okay, we have all we need - do the md5 hash if needed */ | |
2800 | if (md5) { | |
bd0388ae | 2801 | tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, |
49a72dfb | 2802 | md5, NULL, req, skb); |
cfb6eeb4 YH |
2803 | } |
2804 | #endif | |
2805 | ||
ef15025a ED |
2806 | /* Do not fool tcpdump (if any), clean our debris */ |
2807 | skb->tstamp.tv64 = 0; | |
1da177e4 LT |
2808 | return skb; |
2809 | } | |
4bc2f18b | 2810 | EXPORT_SYMBOL(tcp_make_synack); |
1da177e4 | 2811 | |
67edfef7 | 2812 | /* Do all connect socket setups that can be done AF independent. */ |
370816ae | 2813 | void tcp_connect_init(struct sock *sk) |
1da177e4 | 2814 | { |
cf533ea5 | 2815 | const struct dst_entry *dst = __sk_dst_get(sk); |
1da177e4 LT |
2816 | struct tcp_sock *tp = tcp_sk(sk); |
2817 | __u8 rcv_wscale; | |
2818 | ||
2819 | /* We'll fix this up when we get a response from the other end. | |
2820 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. | |
2821 | */ | |
2822 | tp->tcp_header_len = sizeof(struct tcphdr) + | |
bb5b7c11 | 2823 | (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); |
1da177e4 | 2824 | |
cfb6eeb4 YH |
2825 | #ifdef CONFIG_TCP_MD5SIG |
2826 | if (tp->af_specific->md5_lookup(sk, sk) != NULL) | |
2827 | tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; | |
2828 | #endif | |
2829 | ||
1da177e4 LT |
2830 | /* If user gave his TCP_MAXSEG, record it to clamp */ |
2831 | if (tp->rx_opt.user_mss) | |
2832 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
2833 | tp->max_window = 0; | |
5d424d5a | 2834 | tcp_mtup_init(sk); |
1da177e4 LT |
2835 | tcp_sync_mss(sk, dst_mtu(dst)); |
2836 | ||
2837 | if (!tp->window_clamp) | |
2838 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); | |
0dbaee3b | 2839 | tp->advmss = dst_metric_advmss(dst); |
f5fff5dc TQ |
2840 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) |
2841 | tp->advmss = tp->rx_opt.user_mss; | |
2842 | ||
1da177e4 | 2843 | tcp_initialize_rcv_mss(sk); |
1da177e4 | 2844 | |
e88c64f0 HPP |
2845 | /* limit the window selection if the user enforce a smaller rx buffer */ |
2846 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && | |
2847 | (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) | |
2848 | tp->window_clamp = tcp_full_space(sk); | |
2849 | ||
1da177e4 LT |
2850 | tcp_select_initial_window(tcp_full_space(sk), |
2851 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), | |
2852 | &tp->rcv_wnd, | |
2853 | &tp->window_clamp, | |
bb5b7c11 | 2854 | sysctl_tcp_window_scaling, |
31d12926 | 2855 | &rcv_wscale, |
2856 | dst_metric(dst, RTAX_INITRWND)); | |
1da177e4 LT |
2857 | |
2858 | tp->rx_opt.rcv_wscale = rcv_wscale; | |
2859 | tp->rcv_ssthresh = tp->rcv_wnd; | |
2860 | ||
2861 | sk->sk_err = 0; | |
2862 | sock_reset_flag(sk, SOCK_DONE); | |
2863 | tp->snd_wnd = 0; | |
ee7537b6 | 2864 | tcp_init_wl(tp, 0); |
1da177e4 LT |
2865 | tp->snd_una = tp->write_seq; |
2866 | tp->snd_sml = tp->write_seq; | |
33f5f57e | 2867 | tp->snd_up = tp->write_seq; |
370816ae | 2868 | tp->snd_nxt = tp->write_seq; |
ee995283 PE |
2869 | |
2870 | if (likely(!tp->repair)) | |
2871 | tp->rcv_nxt = 0; | |
6f198dca AV |
2872 | else |
2873 | tp->rcv_tstamp = tcp_time_stamp; | |
ee995283 PE |
2874 | tp->rcv_wup = tp->rcv_nxt; |
2875 | tp->copied_seq = tp->rcv_nxt; | |
1da177e4 | 2876 | |
463c84b9 ACM |
2877 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
2878 | inet_csk(sk)->icsk_retransmits = 0; | |
1da177e4 LT |
2879 | tcp_clear_retrans(tp); |
2880 | } | |
2881 | ||
783237e8 YC |
2882 | static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) |
2883 | { | |
2884 | struct tcp_sock *tp = tcp_sk(sk); | |
2885 | struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); | |
2886 | ||
2887 | tcb->end_seq += skb->len; | |
2888 | skb_header_release(skb); | |
2889 | __tcp_add_write_queue_tail(sk, skb); | |
2890 | sk->sk_wmem_queued += skb->truesize; | |
2891 | sk_mem_charge(sk, skb->truesize); | |
2892 | tp->write_seq = tcb->end_seq; | |
2893 | tp->packets_out += tcp_skb_pcount(skb); | |
2894 | } | |
2895 | ||
2896 | /* Build and send a SYN with data and (cached) Fast Open cookie. However, | |
2897 | * queue a data-only packet after the regular SYN, such that regular SYNs | |
2898 | * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges | |
2899 | * only the SYN sequence, the data are retransmitted in the first ACK. | |
2900 | * If cookie is not cached or other error occurs, falls back to send a | |
2901 | * regular SYN with Fast Open cookie request option. | |
2902 | */ | |
2903 | static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) | |
2904 | { | |
2905 | struct tcp_sock *tp = tcp_sk(sk); | |
2906 | struct tcp_fastopen_request *fo = tp->fastopen_req; | |
e64a8519 | 2907 | int syn_loss = 0, space, err = 0; |
aab48743 | 2908 | unsigned long last_syn_loss = 0; |
e64a8519 | 2909 | struct sk_buff *syn_data; |
aab48743 | 2910 | |
67da22d2 | 2911 | tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ |
aab48743 YC |
2912 | tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, |
2913 | &syn_loss, &last_syn_loss); | |
2914 | /* Recurring FO SYN losses: revert to regular handshake temporarily */ | |
2915 | if (syn_loss > 1 && | |
2916 | time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { | |
2917 | fo->cookie.len = -1; | |
2918 | goto fallback; | |
2919 | } | |
783237e8 | 2920 | |
67da22d2 YC |
2921 | if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) |
2922 | fo->cookie.len = -1; | |
2923 | else if (fo->cookie.len <= 0) | |
783237e8 YC |
2924 | goto fallback; |
2925 | ||
2926 | /* MSS for SYN-data is based on cached MSS and bounded by PMTU and | |
2927 | * user-MSS. Reserve maximum option space for middleboxes that add | |
2928 | * private TCP options. The cost is reduced data space in SYN :( | |
2929 | */ | |
2930 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) | |
2931 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
1b63edd6 | 2932 | space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - |
783237e8 YC |
2933 | MAX_TCP_OPTION_SPACE; |
2934 | ||
fe42b170 ED |
2935 | space = min_t(size_t, space, fo->size); |
2936 | ||
2937 | /* limit to order-0 allocations */ | |
2938 | space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); | |
2939 | ||
e64a8519 ED |
2940 | syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation); |
2941 | if (!syn_data) | |
783237e8 | 2942 | goto fallback; |
e64a8519 ED |
2943 | syn_data->ip_summed = CHECKSUM_PARTIAL; |
2944 | memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); | |
75a518b0 | 2945 | skb_shinfo(syn_data)->gso_segs = 1; |
e64a8519 ED |
2946 | if (unlikely(memcpy_fromiovecend(skb_put(syn_data, space), |
2947 | fo->data->msg_iov, 0, space))) { | |
2948 | kfree_skb(syn_data); | |
2949 | goto fallback; | |
2950 | } | |
783237e8 | 2951 | |
e64a8519 ED |
2952 | /* No more data pending in inet_wait_for_connect() */ |
2953 | if (space == fo->size) | |
2954 | fo->data = NULL; | |
2955 | fo->copied = space; | |
783237e8 | 2956 | |
e64a8519 | 2957 | tcp_connect_queue_skb(sk, syn_data); |
783237e8 | 2958 | |
e64a8519 | 2959 | err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); |
783237e8 | 2960 | |
e64a8519 ED |
2961 | /* Now full SYN+DATA was cloned and sent (or not), |
2962 | * remove the SYN from the original skb (syn_data) | |
2963 | * we keep in write queue in case of a retransmit, as we | |
2964 | * also have the SYN packet (with no data) in the same queue. | |
2965 | */ | |
2966 | TCP_SKB_CB(syn_data)->seq++; | |
2967 | TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; | |
2968 | if (!err) { | |
67da22d2 | 2969 | tp->syn_data = (fo->copied > 0); |
783237e8 YC |
2970 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); |
2971 | goto done; | |
2972 | } | |
783237e8 YC |
2973 | |
2974 | fallback: | |
2975 | /* Send a regular SYN with Fast Open cookie request option */ | |
2976 | if (fo->cookie.len > 0) | |
2977 | fo->cookie.len = 0; | |
2978 | err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); | |
2979 | if (err) | |
2980 | tp->syn_fastopen = 0; | |
783237e8 YC |
2981 | done: |
2982 | fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ | |
2983 | return err; | |
2984 | } | |
2985 | ||
67edfef7 | 2986 | /* Build a SYN and send it off. */ |
1da177e4 LT |
2987 | int tcp_connect(struct sock *sk) |
2988 | { | |
2989 | struct tcp_sock *tp = tcp_sk(sk); | |
2990 | struct sk_buff *buff; | |
ee586811 | 2991 | int err; |
1da177e4 LT |
2992 | |
2993 | tcp_connect_init(sk); | |
2994 | ||
2b916477 AV |
2995 | if (unlikely(tp->repair)) { |
2996 | tcp_finish_connect(sk, NULL); | |
2997 | return 0; | |
2998 | } | |
2999 | ||
e64a8519 ED |
3000 | buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation); |
3001 | if (unlikely(!buff)) | |
1da177e4 LT |
3002 | return -ENOBUFS; |
3003 | ||
a3433f35 | 3004 | tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); |
783237e8 YC |
3005 | tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp; |
3006 | tcp_connect_queue_skb(sk, buff); | |
e870a8ef | 3007 | TCP_ECN_send_syn(sk, buff); |
1da177e4 | 3008 | |
783237e8 YC |
3009 | /* Send off SYN; include data in Fast Open. */ |
3010 | err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : | |
3011 | tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); | |
ee586811 EP |
3012 | if (err == -ECONNREFUSED) |
3013 | return err; | |
bd37a088 WY |
3014 | |
3015 | /* We change tp->snd_nxt after the tcp_transmit_skb() call | |
3016 | * in order to make this packet get counted in tcpOutSegs. | |
3017 | */ | |
3018 | tp->snd_nxt = tp->write_seq; | |
3019 | tp->pushed_seq = tp->write_seq; | |
81cc8a75 | 3020 | TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); |
1da177e4 LT |
3021 | |
3022 | /* Timer for repeating the SYN until an answer. */ | |
3f421baa | 3023 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
6fa3eb70 | 3024 | inet_csk(sk)->icsk_rto, sysctl_tcp_rto_max); |
1da177e4 LT |
3025 | return 0; |
3026 | } | |
4bc2f18b | 3027 | EXPORT_SYMBOL(tcp_connect); |
1da177e4 LT |
3028 | |
3029 | /* Send out a delayed ack, the caller does the policy checking | |
3030 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() | |
3031 | * for details. | |
3032 | */ | |
3033 | void tcp_send_delayed_ack(struct sock *sk) | |
3034 | { | |
463c84b9 ACM |
3035 | struct inet_connection_sock *icsk = inet_csk(sk); |
3036 | int ato = icsk->icsk_ack.ato; | |
1da177e4 LT |
3037 | unsigned long timeout; |
3038 | ||
3039 | if (ato > TCP_DELACK_MIN) { | |
463c84b9 | 3040 | const struct tcp_sock *tp = tcp_sk(sk); |
056834d9 | 3041 | int max_ato = HZ / 2; |
1da177e4 | 3042 | |
056834d9 IJ |
3043 | if (icsk->icsk_ack.pingpong || |
3044 | (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) | |
1da177e4 LT |
3045 | max_ato = TCP_DELACK_MAX; |
3046 | ||
3047 | /* Slow path, intersegment interval is "high". */ | |
3048 | ||
3049 | /* If some rtt estimate is known, use it to bound delayed ack. | |
463c84b9 | 3050 | * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements |
1da177e4 LT |
3051 | * directly. |
3052 | */ | |
3053 | if (tp->srtt) { | |
056834d9 | 3054 | int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); |
1da177e4 LT |
3055 | |
3056 | if (rtt < max_ato) | |
3057 | max_ato = rtt; | |
3058 | } | |
3059 | ||
3060 | ato = min(ato, max_ato); | |
3061 | } | |
3062 | ||
3063 | /* Stay within the limit we were given */ | |
3064 | timeout = jiffies + ato; | |
3065 | ||
3066 | /* Use new timeout only if there wasn't a older one earlier. */ | |
463c84b9 | 3067 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { |
1da177e4 LT |
3068 | /* If delack timer was blocked or is about to expire, |
3069 | * send ACK now. | |
3070 | */ | |
463c84b9 ACM |
3071 | if (icsk->icsk_ack.blocked || |
3072 | time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { | |
1da177e4 LT |
3073 | tcp_send_ack(sk); |
3074 | return; | |
3075 | } | |
3076 | ||
463c84b9 ACM |
3077 | if (!time_before(timeout, icsk->icsk_ack.timeout)) |
3078 | timeout = icsk->icsk_ack.timeout; | |
1da177e4 | 3079 | } |
463c84b9 ACM |
3080 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; |
3081 | icsk->icsk_ack.timeout = timeout; | |
3082 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); | |
1da177e4 LT |
3083 | } |
3084 | ||
3085 | /* This routine sends an ack and also updates the window. */ | |
3086 | void tcp_send_ack(struct sock *sk) | |
3087 | { | |
058dc334 | 3088 | struct sk_buff *buff; |
1da177e4 | 3089 | |
058dc334 IJ |
3090 | /* If we have been reset, we may not send again. */ |
3091 | if (sk->sk_state == TCP_CLOSE) | |
3092 | return; | |
1da177e4 | 3093 | |
058dc334 IJ |
3094 | /* We are not putting this on the write queue, so |
3095 | * tcp_transmit_skb() will set the ownership to this | |
3096 | * sock. | |
3097 | */ | |
99a1dec7 | 3098 | buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); |
058dc334 IJ |
3099 | if (buff == NULL) { |
3100 | inet_csk_schedule_ack(sk); | |
3101 | inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; | |
3102 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, | |
6fa3eb70 | 3103 | TCP_DELACK_MAX, sysctl_tcp_rto_max); |
058dc334 | 3104 | return; |
1da177e4 | 3105 | } |
058dc334 IJ |
3106 | |
3107 | /* Reserve space for headers and prepare control bits. */ | |
3108 | skb_reserve(buff, MAX_TCP_HEADER); | |
a3433f35 | 3109 | tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); |
058dc334 IJ |
3110 | |
3111 | /* Send it off, this clears delayed acks for us. */ | |
058dc334 | 3112 | TCP_SKB_CB(buff)->when = tcp_time_stamp; |
99a1dec7 | 3113 | tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC)); |
1da177e4 LT |
3114 | } |
3115 | ||
3116 | /* This routine sends a packet with an out of date sequence | |
3117 | * number. It assumes the other end will try to ack it. | |
3118 | * | |
3119 | * Question: what should we make while urgent mode? | |
3120 | * 4.4BSD forces sending single byte of data. We cannot send | |
3121 | * out of window data, because we have SND.NXT==SND.MAX... | |
3122 | * | |
3123 | * Current solution: to send TWO zero-length segments in urgent mode: | |
3124 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is | |
3125 | * out-of-date with SND.UNA-1 to probe window. | |
3126 | */ | |
3127 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent) | |
3128 | { | |
3129 | struct tcp_sock *tp = tcp_sk(sk); | |
3130 | struct sk_buff *skb; | |
3131 | ||
3132 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ | |
99a1dec7 | 3133 | skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); |
e905a9ed | 3134 | if (skb == NULL) |
1da177e4 LT |
3135 | return -1; |
3136 | ||
3137 | /* Reserve space for headers and set control bits. */ | |
3138 | skb_reserve(skb, MAX_TCP_HEADER); | |
1da177e4 LT |
3139 | /* Use a previous sequence. This should cause the other |
3140 | * end to send an ack. Don't queue or clone SKB, just | |
3141 | * send it. | |
3142 | */ | |
a3433f35 | 3143 | tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); |
1da177e4 | 3144 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
dfb4b9dc | 3145 | return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); |
1da177e4 LT |
3146 | } |
3147 | ||
ee995283 PE |
3148 | void tcp_send_window_probe(struct sock *sk) |
3149 | { | |
3150 | if (sk->sk_state == TCP_ESTABLISHED) { | |
3151 | tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; | |
3152 | tcp_xmit_probe_skb(sk, 0); | |
3153 | } | |
3154 | } | |
3155 | ||
67edfef7 | 3156 | /* Initiate keepalive or window probe from timer. */ |
1da177e4 LT |
3157 | int tcp_write_wakeup(struct sock *sk) |
3158 | { | |
058dc334 IJ |
3159 | struct tcp_sock *tp = tcp_sk(sk); |
3160 | struct sk_buff *skb; | |
1da177e4 | 3161 | |
058dc334 IJ |
3162 | if (sk->sk_state == TCP_CLOSE) |
3163 | return -1; | |
3164 | ||
3165 | if ((skb = tcp_send_head(sk)) != NULL && | |
3166 | before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { | |
3167 | int err; | |
0c54b85f | 3168 | unsigned int mss = tcp_current_mss(sk); |
058dc334 IJ |
3169 | unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
3170 | ||
3171 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) | |
3172 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; | |
3173 | ||
3174 | /* We are probing the opening of a window | |
3175 | * but the window size is != 0 | |
3176 | * must have been a result SWS avoidance ( sender ) | |
3177 | */ | |
3178 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || | |
3179 | skb->len > mss) { | |
3180 | seg_size = min(seg_size, mss); | |
4de075e0 | 3181 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
058dc334 IJ |
3182 | if (tcp_fragment(sk, skb, seg_size, mss)) |
3183 | return -1; | |
3184 | } else if (!tcp_skb_pcount(skb)) | |
3185 | tcp_set_skb_tso_segs(sk, skb, mss); | |
3186 | ||
4de075e0 | 3187 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
058dc334 IJ |
3188 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
3189 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); | |
3190 | if (!err) | |
3191 | tcp_event_new_data_sent(sk, skb); | |
3192 | return err; | |
3193 | } else { | |
33f5f57e | 3194 | if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) |
058dc334 IJ |
3195 | tcp_xmit_probe_skb(sk, 1); |
3196 | return tcp_xmit_probe_skb(sk, 0); | |
1da177e4 | 3197 | } |
1da177e4 LT |
3198 | } |
3199 | ||
3200 | /* A window probe timeout has occurred. If window is not closed send | |
3201 | * a partial packet else a zero probe. | |
3202 | */ | |
3203 | void tcp_send_probe0(struct sock *sk) | |
3204 | { | |
463c84b9 | 3205 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
3206 | struct tcp_sock *tp = tcp_sk(sk); |
3207 | int err; | |
3208 | ||
3209 | err = tcp_write_wakeup(sk); | |
3210 | ||
fe067e8a | 3211 | if (tp->packets_out || !tcp_send_head(sk)) { |
1da177e4 | 3212 | /* Cancel probe timer, if it is not required. */ |
6687e988 | 3213 | icsk->icsk_probes_out = 0; |
463c84b9 | 3214 | icsk->icsk_backoff = 0; |
1da177e4 LT |
3215 | return; |
3216 | } | |
3217 | ||
3218 | if (err <= 0) { | |
463c84b9 ACM |
3219 | if (icsk->icsk_backoff < sysctl_tcp_retries2) |
3220 | icsk->icsk_backoff++; | |
6687e988 | 3221 | icsk->icsk_probes_out++; |
e905a9ed | 3222 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
6fa3eb70 S |
3223 | min_t(unsigned int, icsk->icsk_rto << icsk->icsk_backoff, sysctl_tcp_rto_max), |
3224 | sysctl_tcp_rto_max); | |
1da177e4 LT |
3225 | } else { |
3226 | /* If packet was not sent due to local congestion, | |
6687e988 | 3227 | * do not backoff and do not remember icsk_probes_out. |
1da177e4 LT |
3228 | * Let local senders to fight for local resources. |
3229 | * | |
3230 | * Use accumulated backoff yet. | |
3231 | */ | |
6687e988 ACM |
3232 | if (!icsk->icsk_probes_out) |
3233 | icsk->icsk_probes_out = 1; | |
e905a9ed | 3234 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
463c84b9 | 3235 | min(icsk->icsk_rto << icsk->icsk_backoff, |
3f421baa | 3236 | TCP_RESOURCE_PROBE_INTERVAL), |
6fa3eb70 | 3237 | sysctl_tcp_rto_max); |
1da177e4 LT |
3238 | } |
3239 | } |