]>
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 | |
519855c5 | 68 | int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */ |
e6b09cca | 69 | EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size); |
519855c5 | 70 | |
46d3ceab ED |
71 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
72 | int push_one, gfp_t gfp); | |
519855c5 | 73 | |
67edfef7 | 74 | /* Account for new data that has been sent to the network. */ |
cf533ea5 | 75 | static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 76 | { |
9e412ba7 | 77 | struct tcp_sock *tp = tcp_sk(sk); |
66f5fe62 | 78 | unsigned int prior_packets = tp->packets_out; |
9e412ba7 | 79 | |
fe067e8a | 80 | tcp_advance_send_head(sk, skb); |
1da177e4 | 81 | tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; |
8512430e | 82 | |
25985edc | 83 | /* Don't override Nagle indefinitely with F-RTO */ |
8512430e IJ |
84 | if (tp->frto_counter == 2) |
85 | tp->frto_counter = 3; | |
66f5fe62 IJ |
86 | |
87 | tp->packets_out += tcp_skb_pcount(skb); | |
750ea2ba YC |
88 | if (!prior_packets || tp->early_retrans_delayed) |
89 | tcp_rearm_rto(sk); | |
1da177e4 LT |
90 | } |
91 | ||
92 | /* SND.NXT, if window was not shrunk. | |
93 | * If window has been shrunk, what should we make? It is not clear at all. | |
94 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( | |
95 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already | |
96 | * invalid. OK, let's make this for now: | |
97 | */ | |
cf533ea5 | 98 | static inline __u32 tcp_acceptable_seq(const struct sock *sk) |
1da177e4 | 99 | { |
cf533ea5 | 100 | const struct tcp_sock *tp = tcp_sk(sk); |
9e412ba7 | 101 | |
90840def | 102 | if (!before(tcp_wnd_end(tp), tp->snd_nxt)) |
1da177e4 LT |
103 | return tp->snd_nxt; |
104 | else | |
90840def | 105 | return tcp_wnd_end(tp); |
1da177e4 LT |
106 | } |
107 | ||
108 | /* Calculate mss to advertise in SYN segment. | |
109 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: | |
110 | * | |
111 | * 1. It is independent of path mtu. | |
112 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. | |
113 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of | |
114 | * attached devices, because some buggy hosts are confused by | |
115 | * large MSS. | |
116 | * 4. We do not make 3, we advertise MSS, calculated from first | |
117 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. | |
118 | * This may be overridden via information stored in routing table. | |
119 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, | |
120 | * probably even Jumbo". | |
121 | */ | |
122 | static __u16 tcp_advertise_mss(struct sock *sk) | |
123 | { | |
124 | struct tcp_sock *tp = tcp_sk(sk); | |
cf533ea5 | 125 | const struct dst_entry *dst = __sk_dst_get(sk); |
1da177e4 LT |
126 | int mss = tp->advmss; |
127 | ||
0dbaee3b DM |
128 | if (dst) { |
129 | unsigned int metric = dst_metric_advmss(dst); | |
130 | ||
131 | if (metric < mss) { | |
132 | mss = metric; | |
133 | tp->advmss = mss; | |
134 | } | |
1da177e4 LT |
135 | } |
136 | ||
137 | return (__u16)mss; | |
138 | } | |
139 | ||
140 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". | |
141 | * This is the first part of cwnd validation mechanism. */ | |
cf533ea5 | 142 | static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst) |
1da177e4 | 143 | { |
463c84b9 | 144 | struct tcp_sock *tp = tcp_sk(sk); |
1da177e4 LT |
145 | s32 delta = tcp_time_stamp - tp->lsndtime; |
146 | u32 restart_cwnd = tcp_init_cwnd(tp, dst); | |
147 | u32 cwnd = tp->snd_cwnd; | |
148 | ||
6687e988 | 149 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); |
1da177e4 | 150 | |
6687e988 | 151 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
1da177e4 LT |
152 | restart_cwnd = min(restart_cwnd, cwnd); |
153 | ||
463c84b9 | 154 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) |
1da177e4 LT |
155 | cwnd >>= 1; |
156 | tp->snd_cwnd = max(cwnd, restart_cwnd); | |
157 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
158 | tp->snd_cwnd_used = 0; | |
159 | } | |
160 | ||
67edfef7 | 161 | /* Congestion state accounting after a packet has been sent. */ |
40efc6fa | 162 | static void tcp_event_data_sent(struct tcp_sock *tp, |
cf533ea5 | 163 | struct sock *sk) |
1da177e4 | 164 | { |
463c84b9 ACM |
165 | struct inet_connection_sock *icsk = inet_csk(sk); |
166 | const u32 now = tcp_time_stamp; | |
1da177e4 | 167 | |
35089bb2 DM |
168 | if (sysctl_tcp_slow_start_after_idle && |
169 | (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto)) | |
463c84b9 | 170 | tcp_cwnd_restart(sk, __sk_dst_get(sk)); |
1da177e4 LT |
171 | |
172 | tp->lsndtime = now; | |
173 | ||
174 | /* If it is a reply for ato after last received | |
175 | * packet, enter pingpong mode. | |
176 | */ | |
463c84b9 ACM |
177 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) |
178 | icsk->icsk_ack.pingpong = 1; | |
1da177e4 LT |
179 | } |
180 | ||
67edfef7 | 181 | /* Account for an ACK we sent. */ |
40efc6fa | 182 | static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) |
1da177e4 | 183 | { |
463c84b9 ACM |
184 | tcp_dec_quickack_mode(sk, pkts); |
185 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); | |
1da177e4 LT |
186 | } |
187 | ||
188 | /* Determine a window scaling and initial window to offer. | |
189 | * Based on the assumption that the given amount of space | |
190 | * will be offered. Store the results in the tp structure. | |
191 | * NOTE: for smooth operation initial space offering should | |
192 | * be a multiple of mss if possible. We assume here that mss >= 1. | |
193 | * This MUST be enforced by all callers. | |
194 | */ | |
195 | void tcp_select_initial_window(int __space, __u32 mss, | |
196 | __u32 *rcv_wnd, __u32 *window_clamp, | |
31d12926 | 197 | int wscale_ok, __u8 *rcv_wscale, |
198 | __u32 init_rcv_wnd) | |
1da177e4 LT |
199 | { |
200 | unsigned int space = (__space < 0 ? 0 : __space); | |
201 | ||
202 | /* If no clamp set the clamp to the max possible scaled window */ | |
203 | if (*window_clamp == 0) | |
204 | (*window_clamp) = (65535 << 14); | |
205 | space = min(*window_clamp, space); | |
206 | ||
207 | /* Quantize space offering to a multiple of mss if possible. */ | |
208 | if (space > mss) | |
209 | space = (space / mss) * mss; | |
210 | ||
211 | /* NOTE: offering an initial window larger than 32767 | |
15d99e02 RJ |
212 | * will break some buggy TCP stacks. If the admin tells us |
213 | * it is likely we could be speaking with such a buggy stack | |
214 | * we will truncate our initial window offering to 32K-1 | |
215 | * unless the remote has sent us a window scaling option, | |
216 | * which we interpret as a sign the remote TCP is not | |
217 | * misinterpreting the window field as a signed quantity. | |
1da177e4 | 218 | */ |
15d99e02 RJ |
219 | if (sysctl_tcp_workaround_signed_windows) |
220 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); | |
221 | else | |
222 | (*rcv_wnd) = space; | |
223 | ||
1da177e4 LT |
224 | (*rcv_wscale) = 0; |
225 | if (wscale_ok) { | |
226 | /* Set window scaling on max possible window | |
e905a9ed | 227 | * See RFC1323 for an explanation of the limit to 14 |
1da177e4 LT |
228 | */ |
229 | space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max); | |
316c1592 | 230 | space = min_t(u32, space, *window_clamp); |
1da177e4 LT |
231 | while (space > 65535 && (*rcv_wscale) < 14) { |
232 | space >>= 1; | |
233 | (*rcv_wscale)++; | |
234 | } | |
235 | } | |
236 | ||
356f0398 ND |
237 | /* Set initial window to a value enough for senders starting with |
238 | * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place | |
239 | * a limit on the initial window when mss is larger than 1460. | |
240 | */ | |
056834d9 | 241 | if (mss > (1 << *rcv_wscale)) { |
356f0398 ND |
242 | int init_cwnd = TCP_DEFAULT_INIT_RCVWND; |
243 | if (mss > 1460) | |
244 | init_cwnd = | |
245 | max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2); | |
31d12926 | 246 | /* when initializing use the value from init_rcv_wnd |
247 | * rather than the default from above | |
248 | */ | |
b1afde60 ND |
249 | if (init_rcv_wnd) |
250 | *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); | |
251 | else | |
252 | *rcv_wnd = min(*rcv_wnd, init_cwnd * mss); | |
1da177e4 LT |
253 | } |
254 | ||
255 | /* Set the clamp no higher than max representable value */ | |
256 | (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); | |
257 | } | |
4bc2f18b | 258 | EXPORT_SYMBOL(tcp_select_initial_window); |
1da177e4 LT |
259 | |
260 | /* Chose a new window to advertise, update state in tcp_sock for the | |
261 | * socket, and return result with RFC1323 scaling applied. The return | |
262 | * value can be stuffed directly into th->window for an outgoing | |
263 | * frame. | |
264 | */ | |
40efc6fa | 265 | static u16 tcp_select_window(struct sock *sk) |
1da177e4 LT |
266 | { |
267 | struct tcp_sock *tp = tcp_sk(sk); | |
268 | u32 cur_win = tcp_receive_window(tp); | |
269 | u32 new_win = __tcp_select_window(sk); | |
270 | ||
271 | /* Never shrink the offered window */ | |
2de979bd | 272 | if (new_win < cur_win) { |
1da177e4 LT |
273 | /* Danger Will Robinson! |
274 | * Don't update rcv_wup/rcv_wnd here or else | |
275 | * we will not be able to advertise a zero | |
276 | * window in time. --DaveM | |
277 | * | |
278 | * Relax Will Robinson. | |
279 | */ | |
607bfbf2 | 280 | new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); |
1da177e4 LT |
281 | } |
282 | tp->rcv_wnd = new_win; | |
283 | tp->rcv_wup = tp->rcv_nxt; | |
284 | ||
285 | /* Make sure we do not exceed the maximum possible | |
286 | * scaled window. | |
287 | */ | |
15d99e02 | 288 | if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) |
1da177e4 LT |
289 | new_win = min(new_win, MAX_TCP_WINDOW); |
290 | else | |
291 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); | |
292 | ||
293 | /* RFC1323 scaling applied */ | |
294 | new_win >>= tp->rx_opt.rcv_wscale; | |
295 | ||
296 | /* If we advertise zero window, disable fast path. */ | |
297 | if (new_win == 0) | |
298 | tp->pred_flags = 0; | |
299 | ||
300 | return new_win; | |
301 | } | |
302 | ||
67edfef7 | 303 | /* Packet ECN state for a SYN-ACK */ |
cf533ea5 | 304 | static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb) |
bdf1ee5d | 305 | { |
4de075e0 | 306 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; |
056834d9 | 307 | if (!(tp->ecn_flags & TCP_ECN_OK)) |
4de075e0 | 308 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; |
bdf1ee5d IJ |
309 | } |
310 | ||
67edfef7 | 311 | /* Packet ECN state for a SYN. */ |
bdf1ee5d IJ |
312 | static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb) |
313 | { | |
314 | struct tcp_sock *tp = tcp_sk(sk); | |
315 | ||
316 | tp->ecn_flags = 0; | |
255cac91 | 317 | if (sysctl_tcp_ecn == 1) { |
4de075e0 | 318 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; |
bdf1ee5d IJ |
319 | tp->ecn_flags = TCP_ECN_OK; |
320 | } | |
321 | } | |
322 | ||
323 | static __inline__ void | |
cf533ea5 | 324 | TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th) |
bdf1ee5d IJ |
325 | { |
326 | if (inet_rsk(req)->ecn_ok) | |
327 | th->ece = 1; | |
328 | } | |
329 | ||
67edfef7 AK |
330 | /* Set up ECN state for a packet on a ESTABLISHED socket that is about to |
331 | * be sent. | |
332 | */ | |
bdf1ee5d IJ |
333 | static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb, |
334 | int tcp_header_len) | |
335 | { | |
336 | struct tcp_sock *tp = tcp_sk(sk); | |
337 | ||
338 | if (tp->ecn_flags & TCP_ECN_OK) { | |
339 | /* Not-retransmitted data segment: set ECT and inject CWR. */ | |
340 | if (skb->len != tcp_header_len && | |
341 | !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { | |
342 | INET_ECN_xmit(sk); | |
056834d9 | 343 | if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { |
bdf1ee5d IJ |
344 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; |
345 | tcp_hdr(skb)->cwr = 1; | |
346 | skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; | |
347 | } | |
348 | } else { | |
349 | /* ACK or retransmitted segment: clear ECT|CE */ | |
350 | INET_ECN_dontxmit(sk); | |
351 | } | |
352 | if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) | |
353 | tcp_hdr(skb)->ece = 1; | |
354 | } | |
355 | } | |
356 | ||
e870a8ef IJ |
357 | /* Constructs common control bits of non-data skb. If SYN/FIN is present, |
358 | * auto increment end seqno. | |
359 | */ | |
360 | static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) | |
361 | { | |
2e8e18ef | 362 | skb->ip_summed = CHECKSUM_PARTIAL; |
e870a8ef IJ |
363 | skb->csum = 0; |
364 | ||
4de075e0 | 365 | TCP_SKB_CB(skb)->tcp_flags = flags; |
e870a8ef IJ |
366 | TCP_SKB_CB(skb)->sacked = 0; |
367 | ||
368 | skb_shinfo(skb)->gso_segs = 1; | |
369 | skb_shinfo(skb)->gso_size = 0; | |
370 | skb_shinfo(skb)->gso_type = 0; | |
371 | ||
372 | TCP_SKB_CB(skb)->seq = seq; | |
a3433f35 | 373 | if (flags & (TCPHDR_SYN | TCPHDR_FIN)) |
e870a8ef IJ |
374 | seq++; |
375 | TCP_SKB_CB(skb)->end_seq = seq; | |
376 | } | |
377 | ||
a2a385d6 | 378 | static inline bool tcp_urg_mode(const struct tcp_sock *tp) |
33f5f57e IJ |
379 | { |
380 | return tp->snd_una != tp->snd_up; | |
381 | } | |
382 | ||
33ad798c AL |
383 | #define OPTION_SACK_ADVERTISE (1 << 0) |
384 | #define OPTION_TS (1 << 1) | |
385 | #define OPTION_MD5 (1 << 2) | |
89e95a61 | 386 | #define OPTION_WSCALE (1 << 3) |
bd0388ae | 387 | #define OPTION_COOKIE_EXTENSION (1 << 4) |
2100c8d2 | 388 | #define OPTION_FAST_OPEN_COOKIE (1 << 8) |
33ad798c AL |
389 | |
390 | struct tcp_out_options { | |
2100c8d2 YC |
391 | u16 options; /* bit field of OPTION_* */ |
392 | u16 mss; /* 0 to disable */ | |
33ad798c AL |
393 | u8 ws; /* window scale, 0 to disable */ |
394 | u8 num_sack_blocks; /* number of SACK blocks to include */ | |
bd0388ae | 395 | u8 hash_size; /* bytes in hash_location */ |
bd0388ae | 396 | __u8 *hash_location; /* temporary pointer, overloaded */ |
2100c8d2 YC |
397 | __u32 tsval, tsecr; /* need to include OPTION_TS */ |
398 | struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ | |
33ad798c AL |
399 | }; |
400 | ||
bd0388ae WAS |
401 | /* The sysctl int routines are generic, so check consistency here. |
402 | */ | |
403 | static u8 tcp_cookie_size_check(u8 desired) | |
404 | { | |
f1987257 ED |
405 | int cookie_size; |
406 | ||
407 | if (desired > 0) | |
bd0388ae WAS |
408 | /* previously specified */ |
409 | return desired; | |
f1987257 ED |
410 | |
411 | cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size); | |
412 | if (cookie_size <= 0) | |
bd0388ae WAS |
413 | /* no default specified */ |
414 | return 0; | |
f1987257 ED |
415 | |
416 | if (cookie_size <= TCP_COOKIE_MIN) | |
bd0388ae WAS |
417 | /* value too small, specify minimum */ |
418 | return TCP_COOKIE_MIN; | |
f1987257 ED |
419 | |
420 | if (cookie_size >= TCP_COOKIE_MAX) | |
bd0388ae WAS |
421 | /* value too large, specify maximum */ |
422 | return TCP_COOKIE_MAX; | |
f1987257 ED |
423 | |
424 | if (cookie_size & 1) | |
bd0388ae | 425 | /* 8-bit multiple, illegal, fix it */ |
f1987257 ED |
426 | cookie_size++; |
427 | ||
428 | return (u8)cookie_size; | |
bd0388ae WAS |
429 | } |
430 | ||
67edfef7 AK |
431 | /* Write previously computed TCP options to the packet. |
432 | * | |
433 | * Beware: Something in the Internet is very sensitive to the ordering of | |
fd6149d3 IJ |
434 | * TCP options, we learned this through the hard way, so be careful here. |
435 | * Luckily we can at least blame others for their non-compliance but from | |
436 | * inter-operatibility perspective it seems that we're somewhat stuck with | |
437 | * the ordering which we have been using if we want to keep working with | |
438 | * those broken things (not that it currently hurts anybody as there isn't | |
439 | * particular reason why the ordering would need to be changed). | |
440 | * | |
441 | * At least SACK_PERM as the first option is known to lead to a disaster | |
442 | * (but it may well be that other scenarios fail similarly). | |
443 | */ | |
33ad798c | 444 | static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, |
bd0388ae WAS |
445 | struct tcp_out_options *opts) |
446 | { | |
2100c8d2 | 447 | u16 options = opts->options; /* mungable copy */ |
bd0388ae WAS |
448 | |
449 | /* Having both authentication and cookies for security is redundant, | |
450 | * and there's certainly not enough room. Instead, the cookie-less | |
451 | * extension variant is proposed. | |
452 | * | |
453 | * Consider the pessimal case with authentication. The options | |
454 | * could look like: | |
455 | * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40 | |
456 | */ | |
457 | if (unlikely(OPTION_MD5 & options)) { | |
458 | if (unlikely(OPTION_COOKIE_EXTENSION & options)) { | |
459 | *ptr++ = htonl((TCPOPT_COOKIE << 24) | | |
460 | (TCPOLEN_COOKIE_BASE << 16) | | |
461 | (TCPOPT_MD5SIG << 8) | | |
462 | TCPOLEN_MD5SIG); | |
463 | } else { | |
464 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
465 | (TCPOPT_NOP << 16) | | |
466 | (TCPOPT_MD5SIG << 8) | | |
467 | TCPOLEN_MD5SIG); | |
468 | } | |
469 | options &= ~OPTION_COOKIE_EXTENSION; | |
470 | /* overload cookie hash location */ | |
471 | opts->hash_location = (__u8 *)ptr; | |
33ad798c | 472 | ptr += 4; |
40efc6fa | 473 | } |
33ad798c | 474 | |
fd6149d3 IJ |
475 | if (unlikely(opts->mss)) { |
476 | *ptr++ = htonl((TCPOPT_MSS << 24) | | |
477 | (TCPOLEN_MSS << 16) | | |
478 | opts->mss); | |
479 | } | |
480 | ||
bd0388ae WAS |
481 | if (likely(OPTION_TS & options)) { |
482 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { | |
33ad798c AL |
483 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
484 | (TCPOLEN_SACK_PERM << 16) | | |
485 | (TCPOPT_TIMESTAMP << 8) | | |
486 | TCPOLEN_TIMESTAMP); | |
bd0388ae | 487 | options &= ~OPTION_SACK_ADVERTISE; |
33ad798c AL |
488 | } else { |
489 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
490 | (TCPOPT_NOP << 16) | | |
491 | (TCPOPT_TIMESTAMP << 8) | | |
492 | TCPOLEN_TIMESTAMP); | |
493 | } | |
494 | *ptr++ = htonl(opts->tsval); | |
495 | *ptr++ = htonl(opts->tsecr); | |
496 | } | |
497 | ||
bd0388ae WAS |
498 | /* Specification requires after timestamp, so do it now. |
499 | * | |
500 | * Consider the pessimal case without authentication. The options | |
501 | * could look like: | |
502 | * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40 | |
503 | */ | |
504 | if (unlikely(OPTION_COOKIE_EXTENSION & options)) { | |
505 | __u8 *cookie_copy = opts->hash_location; | |
506 | u8 cookie_size = opts->hash_size; | |
507 | ||
508 | /* 8-bit multiple handled in tcp_cookie_size_check() above, | |
509 | * and elsewhere. | |
510 | */ | |
511 | if (0x2 & cookie_size) { | |
512 | __u8 *p = (__u8 *)ptr; | |
513 | ||
514 | /* 16-bit multiple */ | |
515 | *p++ = TCPOPT_COOKIE; | |
516 | *p++ = TCPOLEN_COOKIE_BASE + cookie_size; | |
517 | *p++ = *cookie_copy++; | |
518 | *p++ = *cookie_copy++; | |
519 | ptr++; | |
520 | cookie_size -= 2; | |
521 | } else { | |
522 | /* 32-bit multiple */ | |
523 | *ptr++ = htonl(((TCPOPT_NOP << 24) | | |
524 | (TCPOPT_NOP << 16) | | |
525 | (TCPOPT_COOKIE << 8) | | |
526 | TCPOLEN_COOKIE_BASE) + | |
527 | cookie_size); | |
528 | } | |
529 | ||
530 | if (cookie_size > 0) { | |
531 | memcpy(ptr, cookie_copy, cookie_size); | |
532 | ptr += (cookie_size / 4); | |
533 | } | |
534 | } | |
535 | ||
536 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { | |
33ad798c AL |
537 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
538 | (TCPOPT_NOP << 16) | | |
539 | (TCPOPT_SACK_PERM << 8) | | |
540 | TCPOLEN_SACK_PERM); | |
541 | } | |
542 | ||
bd0388ae | 543 | if (unlikely(OPTION_WSCALE & options)) { |
33ad798c AL |
544 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
545 | (TCPOPT_WINDOW << 16) | | |
546 | (TCPOLEN_WINDOW << 8) | | |
547 | opts->ws); | |
548 | } | |
549 | ||
550 | if (unlikely(opts->num_sack_blocks)) { | |
551 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? | |
552 | tp->duplicate_sack : tp->selective_acks; | |
40efc6fa SH |
553 | int this_sack; |
554 | ||
555 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
556 | (TCPOPT_NOP << 16) | | |
557 | (TCPOPT_SACK << 8) | | |
33ad798c | 558 | (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * |
40efc6fa | 559 | TCPOLEN_SACK_PERBLOCK))); |
2de979bd | 560 | |
33ad798c AL |
561 | for (this_sack = 0; this_sack < opts->num_sack_blocks; |
562 | ++this_sack) { | |
40efc6fa SH |
563 | *ptr++ = htonl(sp[this_sack].start_seq); |
564 | *ptr++ = htonl(sp[this_sack].end_seq); | |
565 | } | |
2de979bd | 566 | |
5861f8e5 | 567 | tp->rx_opt.dsack = 0; |
40efc6fa | 568 | } |
2100c8d2 YC |
569 | |
570 | if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { | |
571 | struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; | |
572 | ||
573 | *ptr++ = htonl((TCPOPT_EXP << 24) | | |
574 | ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) | | |
575 | TCPOPT_FASTOPEN_MAGIC); | |
576 | ||
577 | memcpy(ptr, foc->val, foc->len); | |
578 | if ((foc->len & 3) == 2) { | |
579 | u8 *align = ((u8 *)ptr) + foc->len; | |
580 | align[0] = align[1] = TCPOPT_NOP; | |
581 | } | |
582 | ptr += (foc->len + 3) >> 2; | |
583 | } | |
33ad798c AL |
584 | } |
585 | ||
67edfef7 AK |
586 | /* Compute TCP options for SYN packets. This is not the final |
587 | * network wire format yet. | |
588 | */ | |
95c96174 | 589 | static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, |
33ad798c | 590 | struct tcp_out_options *opts, |
cf533ea5 ED |
591 | struct tcp_md5sig_key **md5) |
592 | { | |
33ad798c | 593 | struct tcp_sock *tp = tcp_sk(sk); |
bd0388ae | 594 | struct tcp_cookie_values *cvp = tp->cookie_values; |
95c96174 | 595 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
bd0388ae WAS |
596 | u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ? |
597 | tcp_cookie_size_check(cvp->cookie_desired) : | |
598 | 0; | |
783237e8 | 599 | struct tcp_fastopen_request *fastopen = tp->fastopen_req; |
33ad798c | 600 | |
cfb6eeb4 | 601 | #ifdef CONFIG_TCP_MD5SIG |
33ad798c AL |
602 | *md5 = tp->af_specific->md5_lookup(sk, sk); |
603 | if (*md5) { | |
604 | opts->options |= OPTION_MD5; | |
bd0388ae | 605 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
cfb6eeb4 | 606 | } |
33ad798c AL |
607 | #else |
608 | *md5 = NULL; | |
cfb6eeb4 | 609 | #endif |
33ad798c AL |
610 | |
611 | /* We always get an MSS option. The option bytes which will be seen in | |
612 | * normal data packets should timestamps be used, must be in the MSS | |
613 | * advertised. But we subtract them from tp->mss_cache so that | |
614 | * calculations in tcp_sendmsg are simpler etc. So account for this | |
615 | * fact here if necessary. If we don't do this correctly, as a | |
616 | * receiver we won't recognize data packets as being full sized when we | |
617 | * should, and thus we won't abide by the delayed ACK rules correctly. | |
618 | * SACKs don't matter, we never delay an ACK when we have any of those | |
619 | * going out. */ | |
620 | opts->mss = tcp_advertise_mss(sk); | |
bd0388ae | 621 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c | 622 | |
bb5b7c11 | 623 | if (likely(sysctl_tcp_timestamps && *md5 == NULL)) { |
33ad798c AL |
624 | opts->options |= OPTION_TS; |
625 | opts->tsval = TCP_SKB_CB(skb)->when; | |
626 | opts->tsecr = tp->rx_opt.ts_recent; | |
bd0388ae | 627 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c | 628 | } |
bb5b7c11 | 629 | if (likely(sysctl_tcp_window_scaling)) { |
33ad798c | 630 | opts->ws = tp->rx_opt.rcv_wscale; |
89e95a61 | 631 | opts->options |= OPTION_WSCALE; |
bd0388ae | 632 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 633 | } |
bb5b7c11 | 634 | if (likely(sysctl_tcp_sack)) { |
33ad798c | 635 | opts->options |= OPTION_SACK_ADVERTISE; |
b32d1310 | 636 | if (unlikely(!(OPTION_TS & opts->options))) |
bd0388ae | 637 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c AL |
638 | } |
639 | ||
783237e8 YC |
640 | if (fastopen && fastopen->cookie.len >= 0) { |
641 | u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len; | |
642 | need = (need + 3) & ~3U; /* Align to 32 bits */ | |
643 | if (remaining >= need) { | |
644 | opts->options |= OPTION_FAST_OPEN_COOKIE; | |
645 | opts->fastopen_cookie = &fastopen->cookie; | |
646 | remaining -= need; | |
647 | tp->syn_fastopen = 1; | |
648 | } | |
649 | } | |
bd0388ae WAS |
650 | /* Note that timestamps are required by the specification. |
651 | * | |
652 | * Odd numbers of bytes are prohibited by the specification, ensuring | |
653 | * that the cookie is 16-bit aligned, and the resulting cookie pair is | |
654 | * 32-bit aligned. | |
655 | */ | |
656 | if (*md5 == NULL && | |
657 | (OPTION_TS & opts->options) && | |
658 | cookie_size > 0) { | |
659 | int need = TCPOLEN_COOKIE_BASE + cookie_size; | |
660 | ||
661 | if (0x2 & need) { | |
662 | /* 32-bit multiple */ | |
663 | need += 2; /* NOPs */ | |
664 | ||
665 | if (need > remaining) { | |
666 | /* try shrinking cookie to fit */ | |
667 | cookie_size -= 2; | |
668 | need -= 4; | |
669 | } | |
670 | } | |
671 | while (need > remaining && TCP_COOKIE_MIN <= cookie_size) { | |
672 | cookie_size -= 4; | |
673 | need -= 4; | |
674 | } | |
675 | if (TCP_COOKIE_MIN <= cookie_size) { | |
676 | opts->options |= OPTION_COOKIE_EXTENSION; | |
677 | opts->hash_location = (__u8 *)&cvp->cookie_pair[0]; | |
678 | opts->hash_size = cookie_size; | |
679 | ||
680 | /* Remember for future incarnations. */ | |
681 | cvp->cookie_desired = cookie_size; | |
682 | ||
683 | if (cvp->cookie_desired != cvp->cookie_pair_size) { | |
684 | /* Currently use random bytes as a nonce, | |
685 | * assuming these are completely unpredictable | |
686 | * by hostile users of the same system. | |
687 | */ | |
688 | get_random_bytes(&cvp->cookie_pair[0], | |
689 | cookie_size); | |
690 | cvp->cookie_pair_size = cookie_size; | |
691 | } | |
692 | ||
693 | remaining -= need; | |
694 | } | |
695 | } | |
696 | return MAX_TCP_OPTION_SPACE - remaining; | |
40efc6fa SH |
697 | } |
698 | ||
67edfef7 | 699 | /* Set up TCP options for SYN-ACKs. */ |
95c96174 | 700 | static unsigned int tcp_synack_options(struct sock *sk, |
33ad798c | 701 | struct request_sock *req, |
95c96174 | 702 | unsigned int mss, struct sk_buff *skb, |
33ad798c | 703 | struct tcp_out_options *opts, |
4957faad WAS |
704 | struct tcp_md5sig_key **md5, |
705 | struct tcp_extend_values *xvp) | |
706 | { | |
33ad798c | 707 | struct inet_request_sock *ireq = inet_rsk(req); |
95c96174 | 708 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
4957faad WAS |
709 | u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ? |
710 | xvp->cookie_plus : | |
711 | 0; | |
33ad798c | 712 | |
cfb6eeb4 | 713 | #ifdef CONFIG_TCP_MD5SIG |
33ad798c AL |
714 | *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req); |
715 | if (*md5) { | |
716 | opts->options |= OPTION_MD5; | |
4957faad WAS |
717 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
718 | ||
719 | /* We can't fit any SACK blocks in a packet with MD5 + TS | |
720 | * options. There was discussion about disabling SACK | |
721 | * rather than TS in order to fit in better with old, | |
722 | * buggy kernels, but that was deemed to be unnecessary. | |
723 | */ | |
de213e5e | 724 | ireq->tstamp_ok &= !ireq->sack_ok; |
cfb6eeb4 | 725 | } |
33ad798c AL |
726 | #else |
727 | *md5 = NULL; | |
cfb6eeb4 | 728 | #endif |
33ad798c | 729 | |
4957faad | 730 | /* We always send an MSS option. */ |
33ad798c | 731 | opts->mss = mss; |
4957faad | 732 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c AL |
733 | |
734 | if (likely(ireq->wscale_ok)) { | |
735 | opts->ws = ireq->rcv_wscale; | |
89e95a61 | 736 | opts->options |= OPTION_WSCALE; |
4957faad | 737 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 738 | } |
de213e5e | 739 | if (likely(ireq->tstamp_ok)) { |
33ad798c AL |
740 | opts->options |= OPTION_TS; |
741 | opts->tsval = TCP_SKB_CB(skb)->when; | |
742 | opts->tsecr = req->ts_recent; | |
4957faad | 743 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c AL |
744 | } |
745 | if (likely(ireq->sack_ok)) { | |
746 | opts->options |= OPTION_SACK_ADVERTISE; | |
de213e5e | 747 | if (unlikely(!ireq->tstamp_ok)) |
4957faad | 748 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c AL |
749 | } |
750 | ||
4957faad WAS |
751 | /* Similar rationale to tcp_syn_options() applies here, too. |
752 | * If the <SYN> options fit, the same options should fit now! | |
753 | */ | |
754 | if (*md5 == NULL && | |
de213e5e | 755 | ireq->tstamp_ok && |
4957faad WAS |
756 | cookie_plus > TCPOLEN_COOKIE_BASE) { |
757 | int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */ | |
758 | ||
759 | if (0x2 & need) { | |
760 | /* 32-bit multiple */ | |
761 | need += 2; /* NOPs */ | |
762 | } | |
763 | if (need <= remaining) { | |
764 | opts->options |= OPTION_COOKIE_EXTENSION; | |
765 | opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE; | |
766 | remaining -= need; | |
767 | } else { | |
768 | /* There's no error return, so flag it. */ | |
769 | xvp->cookie_out_never = 1; /* true */ | |
770 | opts->hash_size = 0; | |
771 | } | |
772 | } | |
773 | return MAX_TCP_OPTION_SPACE - remaining; | |
33ad798c AL |
774 | } |
775 | ||
67edfef7 AK |
776 | /* Compute TCP options for ESTABLISHED sockets. This is not the |
777 | * final wire format yet. | |
778 | */ | |
95c96174 | 779 | static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, |
33ad798c | 780 | struct tcp_out_options *opts, |
cf533ea5 ED |
781 | struct tcp_md5sig_key **md5) |
782 | { | |
33ad798c AL |
783 | struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL; |
784 | struct tcp_sock *tp = tcp_sk(sk); | |
95c96174 | 785 | unsigned int size = 0; |
cabeccbd | 786 | unsigned int eff_sacks; |
33ad798c AL |
787 | |
788 | #ifdef CONFIG_TCP_MD5SIG | |
789 | *md5 = tp->af_specific->md5_lookup(sk, sk); | |
790 | if (unlikely(*md5)) { | |
791 | opts->options |= OPTION_MD5; | |
792 | size += TCPOLEN_MD5SIG_ALIGNED; | |
793 | } | |
794 | #else | |
795 | *md5 = NULL; | |
796 | #endif | |
797 | ||
798 | if (likely(tp->rx_opt.tstamp_ok)) { | |
799 | opts->options |= OPTION_TS; | |
800 | opts->tsval = tcb ? tcb->when : 0; | |
801 | opts->tsecr = tp->rx_opt.ts_recent; | |
802 | size += TCPOLEN_TSTAMP_ALIGNED; | |
803 | } | |
804 | ||
cabeccbd IJ |
805 | eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; |
806 | if (unlikely(eff_sacks)) { | |
95c96174 | 807 | const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; |
33ad798c | 808 | opts->num_sack_blocks = |
95c96174 | 809 | min_t(unsigned int, eff_sacks, |
33ad798c AL |
810 | (remaining - TCPOLEN_SACK_BASE_ALIGNED) / |
811 | TCPOLEN_SACK_PERBLOCK); | |
812 | size += TCPOLEN_SACK_BASE_ALIGNED + | |
813 | opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; | |
814 | } | |
815 | ||
816 | return size; | |
40efc6fa | 817 | } |
1da177e4 | 818 | |
46d3ceab ED |
819 | |
820 | /* TCP SMALL QUEUES (TSQ) | |
821 | * | |
822 | * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) | |
823 | * to reduce RTT and bufferbloat. | |
824 | * We do this using a special skb destructor (tcp_wfree). | |
825 | * | |
826 | * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb | |
827 | * needs to be reallocated in a driver. | |
828 | * The invariant being skb->truesize substracted from sk->sk_wmem_alloc | |
829 | * | |
830 | * Since transmit from skb destructor is forbidden, we use a tasklet | |
831 | * to process all sockets that eventually need to send more skbs. | |
832 | * We use one tasklet per cpu, with its own queue of sockets. | |
833 | */ | |
834 | struct tsq_tasklet { | |
835 | struct tasklet_struct tasklet; | |
836 | struct list_head head; /* queue of tcp sockets */ | |
837 | }; | |
838 | static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); | |
839 | ||
6f458dfb ED |
840 | static void tcp_tsq_handler(struct sock *sk) |
841 | { | |
842 | if ((1 << sk->sk_state) & | |
843 | (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | | |
844 | TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) | |
845 | tcp_write_xmit(sk, tcp_current_mss(sk), 0, 0, GFP_ATOMIC); | |
846 | } | |
46d3ceab ED |
847 | /* |
848 | * One tasklest per cpu tries to send more skbs. | |
849 | * We run in tasklet context but need to disable irqs when | |
850 | * transfering tsq->head because tcp_wfree() might | |
851 | * interrupt us (non NAPI drivers) | |
852 | */ | |
853 | static void tcp_tasklet_func(unsigned long data) | |
854 | { | |
855 | struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; | |
856 | LIST_HEAD(list); | |
857 | unsigned long flags; | |
858 | struct list_head *q, *n; | |
859 | struct tcp_sock *tp; | |
860 | struct sock *sk; | |
861 | ||
862 | local_irq_save(flags); | |
863 | list_splice_init(&tsq->head, &list); | |
864 | local_irq_restore(flags); | |
865 | ||
866 | list_for_each_safe(q, n, &list) { | |
867 | tp = list_entry(q, struct tcp_sock, tsq_node); | |
868 | list_del(&tp->tsq_node); | |
869 | ||
870 | sk = (struct sock *)tp; | |
871 | bh_lock_sock(sk); | |
872 | ||
873 | if (!sock_owned_by_user(sk)) { | |
6f458dfb | 874 | tcp_tsq_handler(sk); |
46d3ceab ED |
875 | } else { |
876 | /* defer the work to tcp_release_cb() */ | |
6f458dfb | 877 | set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); |
46d3ceab ED |
878 | } |
879 | bh_unlock_sock(sk); | |
880 | ||
881 | clear_bit(TSQ_QUEUED, &tp->tsq_flags); | |
882 | sk_free(sk); | |
883 | } | |
884 | } | |
885 | ||
6f458dfb ED |
886 | #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ |
887 | (1UL << TCP_WRITE_TIMER_DEFERRED) | \ | |
563d34d0 ED |
888 | (1UL << TCP_DELACK_TIMER_DEFERRED) | \ |
889 | (1UL << TCP_MTU_REDUCED_DEFERRED)) | |
46d3ceab ED |
890 | /** |
891 | * tcp_release_cb - tcp release_sock() callback | |
892 | * @sk: socket | |
893 | * | |
894 | * called from release_sock() to perform protocol dependent | |
895 | * actions before socket release. | |
896 | */ | |
897 | void tcp_release_cb(struct sock *sk) | |
898 | { | |
899 | struct tcp_sock *tp = tcp_sk(sk); | |
6f458dfb | 900 | unsigned long flags, nflags; |
46d3ceab | 901 | |
6f458dfb ED |
902 | /* perform an atomic operation only if at least one flag is set */ |
903 | do { | |
904 | flags = tp->tsq_flags; | |
905 | if (!(flags & TCP_DEFERRED_ALL)) | |
906 | return; | |
907 | nflags = flags & ~TCP_DEFERRED_ALL; | |
908 | } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); | |
909 | ||
910 | if (flags & (1UL << TCP_TSQ_DEFERRED)) | |
911 | tcp_tsq_handler(sk); | |
912 | ||
913 | if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) | |
914 | tcp_write_timer_handler(sk); | |
915 | ||
916 | if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) | |
917 | tcp_delack_timer_handler(sk); | |
563d34d0 ED |
918 | |
919 | if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) | |
920 | sk->sk_prot->mtu_reduced(sk); | |
46d3ceab ED |
921 | } |
922 | EXPORT_SYMBOL(tcp_release_cb); | |
923 | ||
924 | void __init tcp_tasklet_init(void) | |
925 | { | |
926 | int i; | |
927 | ||
928 | for_each_possible_cpu(i) { | |
929 | struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); | |
930 | ||
931 | INIT_LIST_HEAD(&tsq->head); | |
932 | tasklet_init(&tsq->tasklet, | |
933 | tcp_tasklet_func, | |
934 | (unsigned long)tsq); | |
935 | } | |
936 | } | |
937 | ||
938 | /* | |
939 | * Write buffer destructor automatically called from kfree_skb. | |
940 | * We cant xmit new skbs from this context, as we might already | |
941 | * hold qdisc lock. | |
942 | */ | |
8e7dfbc8 | 943 | static void tcp_wfree(struct sk_buff *skb) |
46d3ceab ED |
944 | { |
945 | struct sock *sk = skb->sk; | |
946 | struct tcp_sock *tp = tcp_sk(sk); | |
947 | ||
948 | if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && | |
949 | !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { | |
950 | unsigned long flags; | |
951 | struct tsq_tasklet *tsq; | |
952 | ||
953 | /* Keep a ref on socket. | |
954 | * This last ref will be released in tcp_tasklet_func() | |
955 | */ | |
956 | atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc); | |
957 | ||
958 | /* queue this socket to tasklet queue */ | |
959 | local_irq_save(flags); | |
960 | tsq = &__get_cpu_var(tsq_tasklet); | |
961 | list_add(&tp->tsq_node, &tsq->head); | |
962 | tasklet_schedule(&tsq->tasklet); | |
963 | local_irq_restore(flags); | |
964 | } else { | |
965 | sock_wfree(skb); | |
966 | } | |
967 | } | |
968 | ||
1da177e4 LT |
969 | /* This routine actually transmits TCP packets queued in by |
970 | * tcp_do_sendmsg(). This is used by both the initial | |
971 | * transmission and possible later retransmissions. | |
972 | * All SKB's seen here are completely headerless. It is our | |
973 | * job to build the TCP header, and pass the packet down to | |
974 | * IP so it can do the same plus pass the packet off to the | |
975 | * device. | |
976 | * | |
977 | * We are working here with either a clone of the original | |
978 | * SKB, or a fresh unique copy made by the retransmit engine. | |
979 | */ | |
056834d9 IJ |
980 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, |
981 | gfp_t gfp_mask) | |
1da177e4 | 982 | { |
dfb4b9dc DM |
983 | const struct inet_connection_sock *icsk = inet_csk(sk); |
984 | struct inet_sock *inet; | |
985 | struct tcp_sock *tp; | |
986 | struct tcp_skb_cb *tcb; | |
33ad798c | 987 | struct tcp_out_options opts; |
95c96174 | 988 | unsigned int tcp_options_size, tcp_header_size; |
cfb6eeb4 | 989 | struct tcp_md5sig_key *md5; |
dfb4b9dc | 990 | struct tcphdr *th; |
dfb4b9dc DM |
991 | int err; |
992 | ||
993 | BUG_ON(!skb || !tcp_skb_pcount(skb)); | |
994 | ||
995 | /* If congestion control is doing timestamping, we must | |
996 | * take such a timestamp before we potentially clone/copy. | |
997 | */ | |
164891aa | 998 | if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP) |
dfb4b9dc DM |
999 | __net_timestamp(skb); |
1000 | ||
1001 | if (likely(clone_it)) { | |
1002 | if (unlikely(skb_cloned(skb))) | |
1003 | skb = pskb_copy(skb, gfp_mask); | |
1004 | else | |
1005 | skb = skb_clone(skb, gfp_mask); | |
1006 | if (unlikely(!skb)) | |
1007 | return -ENOBUFS; | |
1008 | } | |
1da177e4 | 1009 | |
dfb4b9dc DM |
1010 | inet = inet_sk(sk); |
1011 | tp = tcp_sk(sk); | |
1012 | tcb = TCP_SKB_CB(skb); | |
33ad798c | 1013 | memset(&opts, 0, sizeof(opts)); |
1da177e4 | 1014 | |
4de075e0 | 1015 | if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) |
33ad798c AL |
1016 | tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); |
1017 | else | |
1018 | tcp_options_size = tcp_established_options(sk, skb, &opts, | |
1019 | &md5); | |
1020 | tcp_header_size = tcp_options_size + sizeof(struct tcphdr); | |
e905a9ed | 1021 | |
3853b584 | 1022 | if (tcp_packets_in_flight(tp) == 0) { |
dfb4b9dc | 1023 | tcp_ca_event(sk, CA_EVENT_TX_START); |
3853b584 TH |
1024 | skb->ooo_okay = 1; |
1025 | } else | |
1026 | skb->ooo_okay = 0; | |
dfb4b9dc | 1027 | |
aa8223c7 ACM |
1028 | skb_push(skb, tcp_header_size); |
1029 | skb_reset_transport_header(skb); | |
46d3ceab ED |
1030 | |
1031 | skb_orphan(skb); | |
1032 | skb->sk = sk; | |
1033 | skb->destructor = (sysctl_tcp_limit_output_bytes > 0) ? | |
1034 | tcp_wfree : sock_wfree; | |
1035 | atomic_add(skb->truesize, &sk->sk_wmem_alloc); | |
dfb4b9dc DM |
1036 | |
1037 | /* Build TCP header and checksum it. */ | |
aa8223c7 | 1038 | th = tcp_hdr(skb); |
c720c7e8 ED |
1039 | th->source = inet->inet_sport; |
1040 | th->dest = inet->inet_dport; | |
dfb4b9dc DM |
1041 | th->seq = htonl(tcb->seq); |
1042 | th->ack_seq = htonl(tp->rcv_nxt); | |
df7a3b07 | 1043 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
4de075e0 | 1044 | tcb->tcp_flags); |
dfb4b9dc | 1045 | |
4de075e0 | 1046 | if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { |
dfb4b9dc DM |
1047 | /* RFC1323: The window in SYN & SYN/ACK segments |
1048 | * is never scaled. | |
1049 | */ | |
600ff0c2 | 1050 | th->window = htons(min(tp->rcv_wnd, 65535U)); |
dfb4b9dc DM |
1051 | } else { |
1052 | th->window = htons(tcp_select_window(sk)); | |
1053 | } | |
1054 | th->check = 0; | |
1055 | th->urg_ptr = 0; | |
1da177e4 | 1056 | |
33f5f57e | 1057 | /* The urg_mode check is necessary during a below snd_una win probe */ |
7691367d HX |
1058 | if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { |
1059 | if (before(tp->snd_up, tcb->seq + 0x10000)) { | |
1060 | th->urg_ptr = htons(tp->snd_up - tcb->seq); | |
1061 | th->urg = 1; | |
1062 | } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { | |
0eae88f3 | 1063 | th->urg_ptr = htons(0xFFFF); |
7691367d HX |
1064 | th->urg = 1; |
1065 | } | |
dfb4b9dc | 1066 | } |
1da177e4 | 1067 | |
bd0388ae | 1068 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
4de075e0 | 1069 | if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0)) |
9e412ba7 | 1070 | TCP_ECN_send(sk, skb, tcp_header_size); |
1da177e4 | 1071 | |
cfb6eeb4 YH |
1072 | #ifdef CONFIG_TCP_MD5SIG |
1073 | /* Calculate the MD5 hash, as we have all we need now */ | |
1074 | if (md5) { | |
a465419b | 1075 | sk_nocaps_add(sk, NETIF_F_GSO_MASK); |
bd0388ae | 1076 | tp->af_specific->calc_md5_hash(opts.hash_location, |
49a72dfb | 1077 | md5, sk, NULL, skb); |
cfb6eeb4 YH |
1078 | } |
1079 | #endif | |
1080 | ||
bb296246 | 1081 | icsk->icsk_af_ops->send_check(sk, skb); |
1da177e4 | 1082 | |
4de075e0 | 1083 | if (likely(tcb->tcp_flags & TCPHDR_ACK)) |
dfb4b9dc | 1084 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); |
1da177e4 | 1085 | |
dfb4b9dc | 1086 | if (skb->len != tcp_header_size) |
cf533ea5 | 1087 | tcp_event_data_sent(tp, sk); |
1da177e4 | 1088 | |
bd37a088 | 1089 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
aa2ea058 TH |
1090 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, |
1091 | tcp_skb_pcount(skb)); | |
1da177e4 | 1092 | |
d9d8da80 | 1093 | err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl); |
83de47cd | 1094 | if (likely(err <= 0)) |
dfb4b9dc DM |
1095 | return err; |
1096 | ||
3cfe3baa | 1097 | tcp_enter_cwr(sk, 1); |
dfb4b9dc | 1098 | |
b9df3cb8 | 1099 | return net_xmit_eval(err); |
1da177e4 LT |
1100 | } |
1101 | ||
67edfef7 | 1102 | /* This routine just queues the buffer for sending. |
1da177e4 LT |
1103 | * |
1104 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, | |
1105 | * otherwise socket can stall. | |
1106 | */ | |
1107 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) | |
1108 | { | |
1109 | struct tcp_sock *tp = tcp_sk(sk); | |
1110 | ||
1111 | /* Advance write_seq and place onto the write_queue. */ | |
1112 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; | |
1113 | skb_header_release(skb); | |
fe067e8a | 1114 | tcp_add_write_queue_tail(sk, skb); |
3ab224be HA |
1115 | sk->sk_wmem_queued += skb->truesize; |
1116 | sk_mem_charge(sk, skb->truesize); | |
1da177e4 LT |
1117 | } |
1118 | ||
67edfef7 | 1119 | /* Initialize TSO segments for a packet. */ |
cf533ea5 | 1120 | static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb, |
056834d9 | 1121 | unsigned int mss_now) |
f6302d1d | 1122 | { |
8e5b9dda HX |
1123 | if (skb->len <= mss_now || !sk_can_gso(sk) || |
1124 | skb->ip_summed == CHECKSUM_NONE) { | |
f6302d1d DM |
1125 | /* Avoid the costly divide in the normal |
1126 | * non-TSO case. | |
1127 | */ | |
7967168c HX |
1128 | skb_shinfo(skb)->gso_segs = 1; |
1129 | skb_shinfo(skb)->gso_size = 0; | |
1130 | skb_shinfo(skb)->gso_type = 0; | |
f6302d1d | 1131 | } else { |
356f89e1 | 1132 | skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now); |
7967168c | 1133 | skb_shinfo(skb)->gso_size = mss_now; |
bcd76111 | 1134 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
1da177e4 LT |
1135 | } |
1136 | } | |
1137 | ||
91fed7a1 | 1138 | /* When a modification to fackets out becomes necessary, we need to check |
68f8353b | 1139 | * skb is counted to fackets_out or not. |
91fed7a1 | 1140 | */ |
cf533ea5 | 1141 | static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, |
91fed7a1 IJ |
1142 | int decr) |
1143 | { | |
a47e5a98 IJ |
1144 | struct tcp_sock *tp = tcp_sk(sk); |
1145 | ||
dc86967b | 1146 | if (!tp->sacked_out || tcp_is_reno(tp)) |
91fed7a1 IJ |
1147 | return; |
1148 | ||
6859d494 | 1149 | if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) |
91fed7a1 | 1150 | tp->fackets_out -= decr; |
91fed7a1 IJ |
1151 | } |
1152 | ||
797108d1 IJ |
1153 | /* Pcount in the middle of the write queue got changed, we need to do various |
1154 | * tweaks to fix counters | |
1155 | */ | |
cf533ea5 | 1156 | static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) |
797108d1 IJ |
1157 | { |
1158 | struct tcp_sock *tp = tcp_sk(sk); | |
1159 | ||
1160 | tp->packets_out -= decr; | |
1161 | ||
1162 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
1163 | tp->sacked_out -= decr; | |
1164 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
1165 | tp->retrans_out -= decr; | |
1166 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) | |
1167 | tp->lost_out -= decr; | |
1168 | ||
1169 | /* Reno case is special. Sigh... */ | |
1170 | if (tcp_is_reno(tp) && decr > 0) | |
1171 | tp->sacked_out -= min_t(u32, tp->sacked_out, decr); | |
1172 | ||
1173 | tcp_adjust_fackets_out(sk, skb, decr); | |
1174 | ||
1175 | if (tp->lost_skb_hint && | |
1176 | before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && | |
52cf3cc8 | 1177 | (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) |
797108d1 IJ |
1178 | tp->lost_cnt_hint -= decr; |
1179 | ||
1180 | tcp_verify_left_out(tp); | |
1181 | } | |
1182 | ||
1da177e4 LT |
1183 | /* Function to create two new TCP segments. Shrinks the given segment |
1184 | * to the specified size and appends a new segment with the rest of the | |
e905a9ed | 1185 | * packet to the list. This won't be called frequently, I hope. |
1da177e4 LT |
1186 | * Remember, these are still headerless SKBs at this point. |
1187 | */ | |
056834d9 IJ |
1188 | int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, |
1189 | unsigned int mss_now) | |
1da177e4 LT |
1190 | { |
1191 | struct tcp_sock *tp = tcp_sk(sk); | |
1192 | struct sk_buff *buff; | |
6475be16 | 1193 | int nsize, old_factor; |
b60b49ea | 1194 | int nlen; |
9ce01461 | 1195 | u8 flags; |
1da177e4 | 1196 | |
2fceec13 IJ |
1197 | if (WARN_ON(len > skb->len)) |
1198 | return -EINVAL; | |
6a438bbe | 1199 | |
1da177e4 LT |
1200 | nsize = skb_headlen(skb) - len; |
1201 | if (nsize < 0) | |
1202 | nsize = 0; | |
1203 | ||
1204 | if (skb_cloned(skb) && | |
1205 | skb_is_nonlinear(skb) && | |
1206 | pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) | |
1207 | return -ENOMEM; | |
1208 | ||
1209 | /* Get a new skb... force flag on. */ | |
1210 | buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC); | |
1211 | if (buff == NULL) | |
1212 | return -ENOMEM; /* We'll just try again later. */ | |
ef5cb973 | 1213 | |
3ab224be HA |
1214 | sk->sk_wmem_queued += buff->truesize; |
1215 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea HX |
1216 | nlen = skb->len - len - nsize; |
1217 | buff->truesize += nlen; | |
1218 | skb->truesize -= nlen; | |
1da177e4 LT |
1219 | |
1220 | /* Correct the sequence numbers. */ | |
1221 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1222 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1223 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1224 | ||
1225 | /* PSH and FIN should only be set in the second packet. */ | |
4de075e0 ED |
1226 | flags = TCP_SKB_CB(skb)->tcp_flags; |
1227 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
1228 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
e14c3caf | 1229 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
1da177e4 | 1230 | |
84fa7933 | 1231 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { |
1da177e4 | 1232 | /* Copy and checksum data tail into the new buffer. */ |
056834d9 IJ |
1233 | buff->csum = csum_partial_copy_nocheck(skb->data + len, |
1234 | skb_put(buff, nsize), | |
1da177e4 LT |
1235 | nsize, 0); |
1236 | ||
1237 | skb_trim(skb, len); | |
1238 | ||
1239 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); | |
1240 | } else { | |
84fa7933 | 1241 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
1242 | skb_split(skb, buff, len); |
1243 | } | |
1244 | ||
1245 | buff->ip_summed = skb->ip_summed; | |
1246 | ||
1247 | /* Looks stupid, but our code really uses when of | |
1248 | * skbs, which it never sent before. --ANK | |
1249 | */ | |
1250 | TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when; | |
a61bbcf2 | 1251 | buff->tstamp = skb->tstamp; |
1da177e4 | 1252 | |
6475be16 DM |
1253 | old_factor = tcp_skb_pcount(skb); |
1254 | ||
1da177e4 | 1255 | /* Fix up tso_factor for both original and new SKB. */ |
846998ae DM |
1256 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
1257 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
1da177e4 | 1258 | |
6475be16 DM |
1259 | /* If this packet has been sent out already, we must |
1260 | * adjust the various packet counters. | |
1261 | */ | |
cf0b450c | 1262 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
6475be16 DM |
1263 | int diff = old_factor - tcp_skb_pcount(skb) - |
1264 | tcp_skb_pcount(buff); | |
1da177e4 | 1265 | |
797108d1 IJ |
1266 | if (diff) |
1267 | tcp_adjust_pcount(sk, skb, diff); | |
1da177e4 LT |
1268 | } |
1269 | ||
1270 | /* Link BUFF into the send queue. */ | |
f44b5271 | 1271 | skb_header_release(buff); |
fe067e8a | 1272 | tcp_insert_write_queue_after(skb, buff, sk); |
1da177e4 LT |
1273 | |
1274 | return 0; | |
1275 | } | |
1276 | ||
1277 | /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c | |
1278 | * eventually). The difference is that pulled data not copied, but | |
1279 | * immediately discarded. | |
1280 | */ | |
f2911969 | 1281 | static void __pskb_trim_head(struct sk_buff *skb, int len) |
1da177e4 LT |
1282 | { |
1283 | int i, k, eat; | |
1284 | ||
4fa48bf3 ED |
1285 | eat = min_t(int, len, skb_headlen(skb)); |
1286 | if (eat) { | |
1287 | __skb_pull(skb, eat); | |
22b4a4f2 | 1288 | skb->avail_size -= eat; |
4fa48bf3 ED |
1289 | len -= eat; |
1290 | if (!len) | |
1291 | return; | |
1292 | } | |
1da177e4 LT |
1293 | eat = len; |
1294 | k = 0; | |
056834d9 | 1295 | for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { |
9e903e08 ED |
1296 | int size = skb_frag_size(&skb_shinfo(skb)->frags[i]); |
1297 | ||
1298 | if (size <= eat) { | |
aff65da0 | 1299 | skb_frag_unref(skb, i); |
9e903e08 | 1300 | eat -= size; |
1da177e4 LT |
1301 | } else { |
1302 | skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i]; | |
1303 | if (eat) { | |
1304 | skb_shinfo(skb)->frags[k].page_offset += eat; | |
9e903e08 | 1305 | skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat); |
1da177e4 LT |
1306 | eat = 0; |
1307 | } | |
1308 | k++; | |
1309 | } | |
1310 | } | |
1311 | skb_shinfo(skb)->nr_frags = k; | |
1312 | ||
27a884dc | 1313 | skb_reset_tail_pointer(skb); |
1da177e4 LT |
1314 | skb->data_len -= len; |
1315 | skb->len = skb->data_len; | |
1da177e4 LT |
1316 | } |
1317 | ||
67edfef7 | 1318 | /* Remove acked data from a packet in the transmit queue. */ |
1da177e4 LT |
1319 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) |
1320 | { | |
056834d9 | 1321 | if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC)) |
1da177e4 LT |
1322 | return -ENOMEM; |
1323 | ||
4fa48bf3 | 1324 | __pskb_trim_head(skb, len); |
1da177e4 LT |
1325 | |
1326 | TCP_SKB_CB(skb)->seq += len; | |
84fa7933 | 1327 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
1328 | |
1329 | skb->truesize -= len; | |
1330 | sk->sk_wmem_queued -= len; | |
3ab224be | 1331 | sk_mem_uncharge(sk, len); |
1da177e4 LT |
1332 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); |
1333 | ||
5b35e1e6 | 1334 | /* Any change of skb->len requires recalculation of tso factor. */ |
1da177e4 | 1335 | if (tcp_skb_pcount(skb) > 1) |
5b35e1e6 | 1336 | tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb)); |
1da177e4 LT |
1337 | |
1338 | return 0; | |
1339 | } | |
1340 | ||
67edfef7 | 1341 | /* Calculate MSS. Not accounting for SACKs here. */ |
67469601 | 1342 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) |
5d424d5a | 1343 | { |
cf533ea5 ED |
1344 | const struct tcp_sock *tp = tcp_sk(sk); |
1345 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
5d424d5a JH |
1346 | int mss_now; |
1347 | ||
1348 | /* Calculate base mss without TCP options: | |
1349 | It is MMS_S - sizeof(tcphdr) of rfc1122 | |
1350 | */ | |
1351 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); | |
1352 | ||
67469601 ED |
1353 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
1354 | if (icsk->icsk_af_ops->net_frag_header_len) { | |
1355 | const struct dst_entry *dst = __sk_dst_get(sk); | |
1356 | ||
1357 | if (dst && dst_allfrag(dst)) | |
1358 | mss_now -= icsk->icsk_af_ops->net_frag_header_len; | |
1359 | } | |
1360 | ||
5d424d5a JH |
1361 | /* Clamp it (mss_clamp does not include tcp options) */ |
1362 | if (mss_now > tp->rx_opt.mss_clamp) | |
1363 | mss_now = tp->rx_opt.mss_clamp; | |
1364 | ||
1365 | /* Now subtract optional transport overhead */ | |
1366 | mss_now -= icsk->icsk_ext_hdr_len; | |
1367 | ||
1368 | /* Then reserve room for full set of TCP options and 8 bytes of data */ | |
1369 | if (mss_now < 48) | |
1370 | mss_now = 48; | |
1371 | ||
1372 | /* Now subtract TCP options size, not including SACKs */ | |
1373 | mss_now -= tp->tcp_header_len - sizeof(struct tcphdr); | |
1374 | ||
1375 | return mss_now; | |
1376 | } | |
1377 | ||
1378 | /* Inverse of above */ | |
67469601 | 1379 | int tcp_mss_to_mtu(struct sock *sk, int mss) |
5d424d5a | 1380 | { |
cf533ea5 ED |
1381 | const struct tcp_sock *tp = tcp_sk(sk); |
1382 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
5d424d5a JH |
1383 | int mtu; |
1384 | ||
1385 | mtu = mss + | |
1386 | tp->tcp_header_len + | |
1387 | icsk->icsk_ext_hdr_len + | |
1388 | icsk->icsk_af_ops->net_header_len; | |
1389 | ||
67469601 ED |
1390 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
1391 | if (icsk->icsk_af_ops->net_frag_header_len) { | |
1392 | const struct dst_entry *dst = __sk_dst_get(sk); | |
1393 | ||
1394 | if (dst && dst_allfrag(dst)) | |
1395 | mtu += icsk->icsk_af_ops->net_frag_header_len; | |
1396 | } | |
5d424d5a JH |
1397 | return mtu; |
1398 | } | |
1399 | ||
67edfef7 | 1400 | /* MTU probing init per socket */ |
5d424d5a JH |
1401 | void tcp_mtup_init(struct sock *sk) |
1402 | { | |
1403 | struct tcp_sock *tp = tcp_sk(sk); | |
1404 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1405 | ||
1406 | icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1; | |
1407 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + | |
e905a9ed | 1408 | icsk->icsk_af_ops->net_header_len; |
5d424d5a JH |
1409 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss); |
1410 | icsk->icsk_mtup.probe_size = 0; | |
1411 | } | |
4bc2f18b | 1412 | EXPORT_SYMBOL(tcp_mtup_init); |
5d424d5a | 1413 | |
1da177e4 LT |
1414 | /* This function synchronize snd mss to current pmtu/exthdr set. |
1415 | ||
1416 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts | |
1417 | for TCP options, but includes only bare TCP header. | |
1418 | ||
1419 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. | |
caa20d9a | 1420 | It is minimum of user_mss and mss received with SYN. |
1da177e4 LT |
1421 | It also does not include TCP options. |
1422 | ||
d83d8461 | 1423 | inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. |
1da177e4 LT |
1424 | |
1425 | tp->mss_cache is current effective sending mss, including | |
1426 | all tcp options except for SACKs. It is evaluated, | |
1427 | taking into account current pmtu, but never exceeds | |
1428 | tp->rx_opt.mss_clamp. | |
1429 | ||
1430 | NOTE1. rfc1122 clearly states that advertised MSS | |
1431 | DOES NOT include either tcp or ip options. | |
1432 | ||
d83d8461 ACM |
1433 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
1434 | are READ ONLY outside this function. --ANK (980731) | |
1da177e4 | 1435 | */ |
1da177e4 LT |
1436 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) |
1437 | { | |
1438 | struct tcp_sock *tp = tcp_sk(sk); | |
d83d8461 | 1439 | struct inet_connection_sock *icsk = inet_csk(sk); |
5d424d5a | 1440 | int mss_now; |
1da177e4 | 1441 | |
5d424d5a JH |
1442 | if (icsk->icsk_mtup.search_high > pmtu) |
1443 | icsk->icsk_mtup.search_high = pmtu; | |
1da177e4 | 1444 | |
5d424d5a | 1445 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
409d22b4 | 1446 | mss_now = tcp_bound_to_half_wnd(tp, mss_now); |
1da177e4 LT |
1447 | |
1448 | /* And store cached results */ | |
d83d8461 | 1449 | icsk->icsk_pmtu_cookie = pmtu; |
5d424d5a JH |
1450 | if (icsk->icsk_mtup.enabled) |
1451 | mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); | |
c1b4a7e6 | 1452 | tp->mss_cache = mss_now; |
1da177e4 LT |
1453 | |
1454 | return mss_now; | |
1455 | } | |
4bc2f18b | 1456 | EXPORT_SYMBOL(tcp_sync_mss); |
1da177e4 LT |
1457 | |
1458 | /* Compute the current effective MSS, taking SACKs and IP options, | |
1459 | * and even PMTU discovery events into account. | |
1da177e4 | 1460 | */ |
0c54b85f | 1461 | unsigned int tcp_current_mss(struct sock *sk) |
1da177e4 | 1462 | { |
cf533ea5 ED |
1463 | const struct tcp_sock *tp = tcp_sk(sk); |
1464 | const struct dst_entry *dst = __sk_dst_get(sk); | |
c1b4a7e6 | 1465 | u32 mss_now; |
95c96174 | 1466 | unsigned int header_len; |
33ad798c AL |
1467 | struct tcp_out_options opts; |
1468 | struct tcp_md5sig_key *md5; | |
c1b4a7e6 DM |
1469 | |
1470 | mss_now = tp->mss_cache; | |
1471 | ||
1da177e4 LT |
1472 | if (dst) { |
1473 | u32 mtu = dst_mtu(dst); | |
d83d8461 | 1474 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
1da177e4 LT |
1475 | mss_now = tcp_sync_mss(sk, mtu); |
1476 | } | |
1477 | ||
33ad798c AL |
1478 | header_len = tcp_established_options(sk, NULL, &opts, &md5) + |
1479 | sizeof(struct tcphdr); | |
1480 | /* The mss_cache is sized based on tp->tcp_header_len, which assumes | |
1481 | * some common options. If this is an odd packet (because we have SACK | |
1482 | * blocks etc) then our calculated header_len will be different, and | |
1483 | * we have to adjust mss_now correspondingly */ | |
1484 | if (header_len != tp->tcp_header_len) { | |
1485 | int delta = (int) header_len - tp->tcp_header_len; | |
1486 | mss_now -= delta; | |
1487 | } | |
cfb6eeb4 | 1488 | |
1da177e4 LT |
1489 | return mss_now; |
1490 | } | |
1491 | ||
a762a980 | 1492 | /* Congestion window validation. (RFC2861) */ |
9e412ba7 | 1493 | static void tcp_cwnd_validate(struct sock *sk) |
a762a980 | 1494 | { |
9e412ba7 | 1495 | struct tcp_sock *tp = tcp_sk(sk); |
a762a980 | 1496 | |
d436d686 | 1497 | if (tp->packets_out >= tp->snd_cwnd) { |
a762a980 DM |
1498 | /* Network is feed fully. */ |
1499 | tp->snd_cwnd_used = 0; | |
1500 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1501 | } else { | |
1502 | /* Network starves. */ | |
1503 | if (tp->packets_out > tp->snd_cwnd_used) | |
1504 | tp->snd_cwnd_used = tp->packets_out; | |
1505 | ||
15d33c07 DM |
1506 | if (sysctl_tcp_slow_start_after_idle && |
1507 | (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) | |
a762a980 DM |
1508 | tcp_cwnd_application_limited(sk); |
1509 | } | |
1510 | } | |
1511 | ||
0e3a4803 IJ |
1512 | /* Returns the portion of skb which can be sent right away without |
1513 | * introducing MSS oddities to segment boundaries. In rare cases where | |
1514 | * mss_now != mss_cache, we will request caller to create a small skb | |
1515 | * per input skb which could be mostly avoided here (if desired). | |
5ea3a748 IJ |
1516 | * |
1517 | * We explicitly want to create a request for splitting write queue tail | |
1518 | * to a small skb for Nagle purposes while avoiding unnecessary modulos, | |
1519 | * thus all the complexity (cwnd_len is always MSS multiple which we | |
1520 | * return whenever allowed by the other factors). Basically we need the | |
1521 | * modulo only when the receiver window alone is the limiting factor or | |
1522 | * when we would be allowed to send the split-due-to-Nagle skb fully. | |
0e3a4803 | 1523 | */ |
cf533ea5 | 1524 | static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb, |
1485348d | 1525 | unsigned int mss_now, unsigned int max_segs) |
c1b4a7e6 | 1526 | { |
cf533ea5 | 1527 | const struct tcp_sock *tp = tcp_sk(sk); |
1485348d | 1528 | u32 needed, window, max_len; |
c1b4a7e6 | 1529 | |
90840def | 1530 | window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
1485348d | 1531 | max_len = mss_now * max_segs; |
0e3a4803 | 1532 | |
1485348d BH |
1533 | if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) |
1534 | return max_len; | |
0e3a4803 | 1535 | |
5ea3a748 IJ |
1536 | needed = min(skb->len, window); |
1537 | ||
1485348d BH |
1538 | if (max_len <= needed) |
1539 | return max_len; | |
0e3a4803 | 1540 | |
0e3a4803 | 1541 | return needed - needed % mss_now; |
c1b4a7e6 DM |
1542 | } |
1543 | ||
1544 | /* Can at least one segment of SKB be sent right now, according to the | |
1545 | * congestion window rules? If so, return how many segments are allowed. | |
1546 | */ | |
cf533ea5 ED |
1547 | static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, |
1548 | const struct sk_buff *skb) | |
c1b4a7e6 DM |
1549 | { |
1550 | u32 in_flight, cwnd; | |
1551 | ||
1552 | /* Don't be strict about the congestion window for the final FIN. */ | |
4de075e0 ED |
1553 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && |
1554 | tcp_skb_pcount(skb) == 1) | |
c1b4a7e6 DM |
1555 | return 1; |
1556 | ||
1557 | in_flight = tcp_packets_in_flight(tp); | |
1558 | cwnd = tp->snd_cwnd; | |
1559 | if (in_flight < cwnd) | |
1560 | return (cwnd - in_flight); | |
1561 | ||
1562 | return 0; | |
1563 | } | |
1564 | ||
b595076a | 1565 | /* Initialize TSO state of a skb. |
67edfef7 | 1566 | * This must be invoked the first time we consider transmitting |
c1b4a7e6 DM |
1567 | * SKB onto the wire. |
1568 | */ | |
cf533ea5 | 1569 | static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb, |
056834d9 | 1570 | unsigned int mss_now) |
c1b4a7e6 DM |
1571 | { |
1572 | int tso_segs = tcp_skb_pcount(skb); | |
1573 | ||
f8269a49 | 1574 | if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { |
846998ae | 1575 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
c1b4a7e6 DM |
1576 | tso_segs = tcp_skb_pcount(skb); |
1577 | } | |
1578 | return tso_segs; | |
1579 | } | |
1580 | ||
67edfef7 | 1581 | /* Minshall's variant of the Nagle send check. */ |
a2a385d6 | 1582 | static inline bool tcp_minshall_check(const struct tcp_sock *tp) |
c1b4a7e6 | 1583 | { |
09cb105e | 1584 | return after(tp->snd_sml, tp->snd_una) && |
c1b4a7e6 DM |
1585 | !after(tp->snd_sml, tp->snd_nxt); |
1586 | } | |
1587 | ||
a2a385d6 | 1588 | /* Return false, if packet can be sent now without violation Nagle's rules: |
c1b4a7e6 DM |
1589 | * 1. It is full sized. |
1590 | * 2. Or it contains FIN. (already checked by caller) | |
6d67e9be | 1591 | * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. |
c1b4a7e6 DM |
1592 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. |
1593 | * With Minshall's modification: all sent small packets are ACKed. | |
1594 | */ | |
a2a385d6 | 1595 | static inline bool tcp_nagle_check(const struct tcp_sock *tp, |
e905a9ed | 1596 | const struct sk_buff *skb, |
95c96174 | 1597 | unsigned int mss_now, int nonagle) |
c1b4a7e6 | 1598 | { |
a02cec21 | 1599 | return skb->len < mss_now && |
056834d9 | 1600 | ((nonagle & TCP_NAGLE_CORK) || |
a02cec21 | 1601 | (!nonagle && tp->packets_out && tcp_minshall_check(tp))); |
c1b4a7e6 DM |
1602 | } |
1603 | ||
a2a385d6 | 1604 | /* Return true if the Nagle test allows this packet to be |
c1b4a7e6 DM |
1605 | * sent now. |
1606 | */ | |
a2a385d6 ED |
1607 | static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, |
1608 | unsigned int cur_mss, int nonagle) | |
c1b4a7e6 DM |
1609 | { |
1610 | /* Nagle rule does not apply to frames, which sit in the middle of the | |
1611 | * write_queue (they have no chances to get new data). | |
1612 | * | |
1613 | * This is implemented in the callers, where they modify the 'nonagle' | |
1614 | * argument based upon the location of SKB in the send queue. | |
1615 | */ | |
1616 | if (nonagle & TCP_NAGLE_PUSH) | |
a2a385d6 | 1617 | return true; |
c1b4a7e6 | 1618 | |
d551e454 IJ |
1619 | /* Don't use the nagle rule for urgent data (or for the final FIN). |
1620 | * Nagle can be ignored during F-RTO too (see RFC4138). | |
1621 | */ | |
33f5f57e | 1622 | if (tcp_urg_mode(tp) || (tp->frto_counter == 2) || |
4de075e0 | 1623 | (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) |
a2a385d6 | 1624 | return true; |
c1b4a7e6 DM |
1625 | |
1626 | if (!tcp_nagle_check(tp, skb, cur_mss, nonagle)) | |
a2a385d6 | 1627 | return true; |
c1b4a7e6 | 1628 | |
a2a385d6 | 1629 | return false; |
c1b4a7e6 DM |
1630 | } |
1631 | ||
1632 | /* Does at least the first segment of SKB fit into the send window? */ | |
a2a385d6 ED |
1633 | static bool tcp_snd_wnd_test(const struct tcp_sock *tp, |
1634 | const struct sk_buff *skb, | |
1635 | unsigned int cur_mss) | |
c1b4a7e6 DM |
1636 | { |
1637 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
1638 | ||
1639 | if (skb->len > cur_mss) | |
1640 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; | |
1641 | ||
90840def | 1642 | return !after(end_seq, tcp_wnd_end(tp)); |
c1b4a7e6 DM |
1643 | } |
1644 | ||
fe067e8a | 1645 | /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) |
c1b4a7e6 DM |
1646 | * should be put on the wire right now. If so, it returns the number of |
1647 | * packets allowed by the congestion window. | |
1648 | */ | |
cf533ea5 | 1649 | static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, |
c1b4a7e6 DM |
1650 | unsigned int cur_mss, int nonagle) |
1651 | { | |
cf533ea5 | 1652 | const struct tcp_sock *tp = tcp_sk(sk); |
c1b4a7e6 DM |
1653 | unsigned int cwnd_quota; |
1654 | ||
846998ae | 1655 | tcp_init_tso_segs(sk, skb, cur_mss); |
c1b4a7e6 DM |
1656 | |
1657 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) | |
1658 | return 0; | |
1659 | ||
1660 | cwnd_quota = tcp_cwnd_test(tp, skb); | |
056834d9 | 1661 | if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) |
c1b4a7e6 DM |
1662 | cwnd_quota = 0; |
1663 | ||
1664 | return cwnd_quota; | |
1665 | } | |
1666 | ||
67edfef7 | 1667 | /* Test if sending is allowed right now. */ |
a2a385d6 | 1668 | bool tcp_may_send_now(struct sock *sk) |
c1b4a7e6 | 1669 | { |
cf533ea5 | 1670 | const struct tcp_sock *tp = tcp_sk(sk); |
fe067e8a | 1671 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 | 1672 | |
a02cec21 | 1673 | return skb && |
0c54b85f | 1674 | tcp_snd_test(sk, skb, tcp_current_mss(sk), |
c1b4a7e6 | 1675 | (tcp_skb_is_last(sk, skb) ? |
a02cec21 | 1676 | tp->nonagle : TCP_NAGLE_PUSH)); |
c1b4a7e6 DM |
1677 | } |
1678 | ||
1679 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet | |
1680 | * which is put after SKB on the list. It is very much like | |
1681 | * tcp_fragment() except that it may make several kinds of assumptions | |
1682 | * in order to speed up the splitting operation. In particular, we | |
1683 | * know that all the data is in scatter-gather pages, and that the | |
1684 | * packet has never been sent out before (and thus is not cloned). | |
1685 | */ | |
056834d9 | 1686 | static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, |
c4ead4c5 | 1687 | unsigned int mss_now, gfp_t gfp) |
c1b4a7e6 DM |
1688 | { |
1689 | struct sk_buff *buff; | |
1690 | int nlen = skb->len - len; | |
9ce01461 | 1691 | u8 flags; |
c1b4a7e6 DM |
1692 | |
1693 | /* All of a TSO frame must be composed of paged data. */ | |
c8ac3774 HX |
1694 | if (skb->len != skb->data_len) |
1695 | return tcp_fragment(sk, skb, len, mss_now); | |
c1b4a7e6 | 1696 | |
c4ead4c5 | 1697 | buff = sk_stream_alloc_skb(sk, 0, gfp); |
c1b4a7e6 DM |
1698 | if (unlikely(buff == NULL)) |
1699 | return -ENOMEM; | |
1700 | ||
3ab224be HA |
1701 | sk->sk_wmem_queued += buff->truesize; |
1702 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea | 1703 | buff->truesize += nlen; |
c1b4a7e6 DM |
1704 | skb->truesize -= nlen; |
1705 | ||
1706 | /* Correct the sequence numbers. */ | |
1707 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1708 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1709 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1710 | ||
1711 | /* PSH and FIN should only be set in the second packet. */ | |
4de075e0 ED |
1712 | flags = TCP_SKB_CB(skb)->tcp_flags; |
1713 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
1714 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
c1b4a7e6 DM |
1715 | |
1716 | /* This packet was never sent out yet, so no SACK bits. */ | |
1717 | TCP_SKB_CB(buff)->sacked = 0; | |
1718 | ||
84fa7933 | 1719 | buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; |
c1b4a7e6 DM |
1720 | skb_split(skb, buff, len); |
1721 | ||
1722 | /* Fix up tso_factor for both original and new SKB. */ | |
846998ae DM |
1723 | tcp_set_skb_tso_segs(sk, skb, mss_now); |
1724 | tcp_set_skb_tso_segs(sk, buff, mss_now); | |
c1b4a7e6 DM |
1725 | |
1726 | /* Link BUFF into the send queue. */ | |
1727 | skb_header_release(buff); | |
fe067e8a | 1728 | tcp_insert_write_queue_after(skb, buff, sk); |
c1b4a7e6 DM |
1729 | |
1730 | return 0; | |
1731 | } | |
1732 | ||
1733 | /* Try to defer sending, if possible, in order to minimize the amount | |
1734 | * of TSO splitting we do. View it as a kind of TSO Nagle test. | |
1735 | * | |
1736 | * This algorithm is from John Heffner. | |
1737 | */ | |
a2a385d6 | 1738 | static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb) |
c1b4a7e6 | 1739 | { |
9e412ba7 | 1740 | struct tcp_sock *tp = tcp_sk(sk); |
6687e988 | 1741 | const struct inet_connection_sock *icsk = inet_csk(sk); |
c1b4a7e6 | 1742 | u32 send_win, cong_win, limit, in_flight; |
ad9f4f50 | 1743 | int win_divisor; |
c1b4a7e6 | 1744 | |
4de075e0 | 1745 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) |
ae8064ac | 1746 | goto send_now; |
c1b4a7e6 | 1747 | |
6687e988 | 1748 | if (icsk->icsk_ca_state != TCP_CA_Open) |
ae8064ac JH |
1749 | goto send_now; |
1750 | ||
1751 | /* Defer for less than two clock ticks. */ | |
bd515c3e | 1752 | if (tp->tso_deferred && |
a2acde07 | 1753 | (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1) |
ae8064ac | 1754 | goto send_now; |
908a75c1 | 1755 | |
c1b4a7e6 DM |
1756 | in_flight = tcp_packets_in_flight(tp); |
1757 | ||
056834d9 | 1758 | BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); |
c1b4a7e6 | 1759 | |
90840def | 1760 | send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
c1b4a7e6 DM |
1761 | |
1762 | /* From in_flight test above, we know that cwnd > in_flight. */ | |
1763 | cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; | |
1764 | ||
1765 | limit = min(send_win, cong_win); | |
1766 | ||
ba244fe9 | 1767 | /* If a full-sized TSO skb can be sent, do it. */ |
1485348d BH |
1768 | if (limit >= min_t(unsigned int, sk->sk_gso_max_size, |
1769 | sk->sk_gso_max_segs * tp->mss_cache)) | |
ae8064ac | 1770 | goto send_now; |
ba244fe9 | 1771 | |
62ad2761 IJ |
1772 | /* Middle in queue won't get any more data, full sendable already? */ |
1773 | if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) | |
1774 | goto send_now; | |
1775 | ||
ad9f4f50 ED |
1776 | win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); |
1777 | if (win_divisor) { | |
c1b4a7e6 DM |
1778 | u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); |
1779 | ||
1780 | /* If at least some fraction of a window is available, | |
1781 | * just use it. | |
1782 | */ | |
ad9f4f50 | 1783 | chunk /= win_divisor; |
c1b4a7e6 | 1784 | if (limit >= chunk) |
ae8064ac | 1785 | goto send_now; |
c1b4a7e6 DM |
1786 | } else { |
1787 | /* Different approach, try not to defer past a single | |
1788 | * ACK. Receiver should ACK every other full sized | |
1789 | * frame, so if we have space for more than 3 frames | |
1790 | * then send now. | |
1791 | */ | |
6b5a5c0d | 1792 | if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) |
ae8064ac | 1793 | goto send_now; |
c1b4a7e6 DM |
1794 | } |
1795 | ||
1796 | /* Ok, it looks like it is advisable to defer. */ | |
056834d9 | 1797 | tp->tso_deferred = 1 | (jiffies << 1); |
ae8064ac | 1798 | |
a2a385d6 | 1799 | return true; |
ae8064ac JH |
1800 | |
1801 | send_now: | |
1802 | tp->tso_deferred = 0; | |
a2a385d6 | 1803 | return false; |
c1b4a7e6 DM |
1804 | } |
1805 | ||
5d424d5a | 1806 | /* Create a new MTU probe if we are ready. |
67edfef7 AK |
1807 | * MTU probe is regularly attempting to increase the path MTU by |
1808 | * deliberately sending larger packets. This discovers routing | |
1809 | * changes resulting in larger path MTUs. | |
1810 | * | |
5d424d5a JH |
1811 | * Returns 0 if we should wait to probe (no cwnd available), |
1812 | * 1 if a probe was sent, | |
056834d9 IJ |
1813 | * -1 otherwise |
1814 | */ | |
5d424d5a JH |
1815 | static int tcp_mtu_probe(struct sock *sk) |
1816 | { | |
1817 | struct tcp_sock *tp = tcp_sk(sk); | |
1818 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1819 | struct sk_buff *skb, *nskb, *next; | |
1820 | int len; | |
1821 | int probe_size; | |
91cc17c0 | 1822 | int size_needed; |
5d424d5a JH |
1823 | int copy; |
1824 | int mss_now; | |
1825 | ||
1826 | /* Not currently probing/verifying, | |
1827 | * not in recovery, | |
1828 | * have enough cwnd, and | |
1829 | * not SACKing (the variable headers throw things off) */ | |
1830 | if (!icsk->icsk_mtup.enabled || | |
1831 | icsk->icsk_mtup.probe_size || | |
1832 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || | |
1833 | tp->snd_cwnd < 11 || | |
cabeccbd | 1834 | tp->rx_opt.num_sacks || tp->rx_opt.dsack) |
5d424d5a JH |
1835 | return -1; |
1836 | ||
1837 | /* Very simple search strategy: just double the MSS. */ | |
0c54b85f | 1838 | mss_now = tcp_current_mss(sk); |
056834d9 | 1839 | probe_size = 2 * tp->mss_cache; |
91cc17c0 | 1840 | size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; |
5d424d5a JH |
1841 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) { |
1842 | /* TODO: set timer for probe_converge_event */ | |
1843 | return -1; | |
1844 | } | |
1845 | ||
1846 | /* Have enough data in the send queue to probe? */ | |
7f9c33e5 | 1847 | if (tp->write_seq - tp->snd_nxt < size_needed) |
5d424d5a JH |
1848 | return -1; |
1849 | ||
91cc17c0 IJ |
1850 | if (tp->snd_wnd < size_needed) |
1851 | return -1; | |
90840def | 1852 | if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) |
91cc17c0 | 1853 | return 0; |
5d424d5a | 1854 | |
d67c58e9 IJ |
1855 | /* Do we need to wait to drain cwnd? With none in flight, don't stall */ |
1856 | if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { | |
1857 | if (!tcp_packets_in_flight(tp)) | |
5d424d5a JH |
1858 | return -1; |
1859 | else | |
1860 | return 0; | |
1861 | } | |
1862 | ||
1863 | /* We're allowed to probe. Build it now. */ | |
1864 | if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL) | |
1865 | return -1; | |
3ab224be HA |
1866 | sk->sk_wmem_queued += nskb->truesize; |
1867 | sk_mem_charge(sk, nskb->truesize); | |
5d424d5a | 1868 | |
fe067e8a | 1869 | skb = tcp_send_head(sk); |
5d424d5a JH |
1870 | |
1871 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; | |
1872 | TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; | |
4de075e0 | 1873 | TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; |
5d424d5a JH |
1874 | TCP_SKB_CB(nskb)->sacked = 0; |
1875 | nskb->csum = 0; | |
84fa7933 | 1876 | nskb->ip_summed = skb->ip_summed; |
5d424d5a | 1877 | |
50c4817e IJ |
1878 | tcp_insert_write_queue_before(nskb, skb, sk); |
1879 | ||
5d424d5a | 1880 | len = 0; |
234b6860 | 1881 | tcp_for_write_queue_from_safe(skb, next, sk) { |
5d424d5a JH |
1882 | copy = min_t(int, skb->len, probe_size - len); |
1883 | if (nskb->ip_summed) | |
1884 | skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); | |
1885 | else | |
1886 | nskb->csum = skb_copy_and_csum_bits(skb, 0, | |
056834d9 IJ |
1887 | skb_put(nskb, copy), |
1888 | copy, nskb->csum); | |
5d424d5a JH |
1889 | |
1890 | if (skb->len <= copy) { | |
1891 | /* We've eaten all the data from this skb. | |
1892 | * Throw it away. */ | |
4de075e0 | 1893 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; |
fe067e8a | 1894 | tcp_unlink_write_queue(skb, sk); |
3ab224be | 1895 | sk_wmem_free_skb(sk, skb); |
5d424d5a | 1896 | } else { |
4de075e0 | 1897 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & |
a3433f35 | 1898 | ~(TCPHDR_FIN|TCPHDR_PSH); |
5d424d5a JH |
1899 | if (!skb_shinfo(skb)->nr_frags) { |
1900 | skb_pull(skb, copy); | |
84fa7933 | 1901 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
056834d9 IJ |
1902 | skb->csum = csum_partial(skb->data, |
1903 | skb->len, 0); | |
5d424d5a JH |
1904 | } else { |
1905 | __pskb_trim_head(skb, copy); | |
1906 | tcp_set_skb_tso_segs(sk, skb, mss_now); | |
1907 | } | |
1908 | TCP_SKB_CB(skb)->seq += copy; | |
1909 | } | |
1910 | ||
1911 | len += copy; | |
234b6860 IJ |
1912 | |
1913 | if (len >= probe_size) | |
1914 | break; | |
5d424d5a JH |
1915 | } |
1916 | tcp_init_tso_segs(sk, nskb, nskb->len); | |
1917 | ||
1918 | /* We're ready to send. If this fails, the probe will | |
1919 | * be resegmented into mss-sized pieces by tcp_write_xmit(). */ | |
1920 | TCP_SKB_CB(nskb)->when = tcp_time_stamp; | |
1921 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { | |
1922 | /* Decrement cwnd here because we are sending | |
056834d9 | 1923 | * effectively two packets. */ |
5d424d5a | 1924 | tp->snd_cwnd--; |
66f5fe62 | 1925 | tcp_event_new_data_sent(sk, nskb); |
5d424d5a JH |
1926 | |
1927 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); | |
0e7b1368 JH |
1928 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; |
1929 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; | |
5d424d5a JH |
1930 | |
1931 | return 1; | |
1932 | } | |
1933 | ||
1934 | return -1; | |
1935 | } | |
1936 | ||
1da177e4 LT |
1937 | /* This routine writes packets to the network. It advances the |
1938 | * send_head. This happens as incoming acks open up the remote | |
1939 | * window for us. | |
1940 | * | |
f8269a49 IJ |
1941 | * LARGESEND note: !tcp_urg_mode is overkill, only frames between |
1942 | * snd_up-64k-mss .. snd_up cannot be large. However, taking into | |
1943 | * account rare use of URG, this is not a big flaw. | |
1944 | * | |
a2a385d6 ED |
1945 | * Returns true, if no segments are in flight and we have queued segments, |
1946 | * but cannot send anything now because of SWS or another problem. | |
1da177e4 | 1947 | */ |
a2a385d6 ED |
1948 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
1949 | int push_one, gfp_t gfp) | |
1da177e4 LT |
1950 | { |
1951 | struct tcp_sock *tp = tcp_sk(sk); | |
92df7b51 | 1952 | struct sk_buff *skb; |
c1b4a7e6 DM |
1953 | unsigned int tso_segs, sent_pkts; |
1954 | int cwnd_quota; | |
5d424d5a | 1955 | int result; |
1da177e4 | 1956 | |
92df7b51 | 1957 | sent_pkts = 0; |
5d424d5a | 1958 | |
d5dd9175 IJ |
1959 | if (!push_one) { |
1960 | /* Do MTU probing. */ | |
1961 | result = tcp_mtu_probe(sk); | |
1962 | if (!result) { | |
a2a385d6 | 1963 | return false; |
d5dd9175 IJ |
1964 | } else if (result > 0) { |
1965 | sent_pkts = 1; | |
1966 | } | |
5d424d5a JH |
1967 | } |
1968 | ||
fe067e8a | 1969 | while ((skb = tcp_send_head(sk))) { |
c8ac3774 HX |
1970 | unsigned int limit; |
1971 | ||
46d3ceab | 1972 | |
b68e9f85 | 1973 | tso_segs = tcp_init_tso_segs(sk, skb, mss_now); |
c1b4a7e6 | 1974 | BUG_ON(!tso_segs); |
aa93466b | 1975 | |
b68e9f85 HX |
1976 | cwnd_quota = tcp_cwnd_test(tp, skb); |
1977 | if (!cwnd_quota) | |
1978 | break; | |
1979 | ||
1980 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) | |
1981 | break; | |
1982 | ||
c1b4a7e6 DM |
1983 | if (tso_segs == 1) { |
1984 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, | |
1985 | (tcp_skb_is_last(sk, skb) ? | |
1986 | nonagle : TCP_NAGLE_PUSH)))) | |
1987 | break; | |
1988 | } else { | |
d5dd9175 | 1989 | if (!push_one && tcp_tso_should_defer(sk, skb)) |
c1b4a7e6 DM |
1990 | break; |
1991 | } | |
aa93466b | 1992 | |
46d3ceab ED |
1993 | /* TSQ : sk_wmem_alloc accounts skb truesize, |
1994 | * including skb overhead. But thats OK. | |
1995 | */ | |
1996 | if (atomic_read(&sk->sk_wmem_alloc) >= sysctl_tcp_limit_output_bytes) { | |
1997 | set_bit(TSQ_THROTTLED, &tp->tsq_flags); | |
1998 | break; | |
1999 | } | |
c8ac3774 | 2000 | limit = mss_now; |
f8269a49 | 2001 | if (tso_segs > 1 && !tcp_urg_mode(tp)) |
0e3a4803 | 2002 | limit = tcp_mss_split_point(sk, skb, mss_now, |
1485348d BH |
2003 | min_t(unsigned int, |
2004 | cwnd_quota, | |
2005 | sk->sk_gso_max_segs)); | |
1da177e4 | 2006 | |
c8ac3774 | 2007 | if (skb->len > limit && |
c4ead4c5 | 2008 | unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) |
c8ac3774 HX |
2009 | break; |
2010 | ||
92df7b51 | 2011 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
c1b4a7e6 | 2012 | |
d5dd9175 | 2013 | if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) |
92df7b51 | 2014 | break; |
1da177e4 | 2015 | |
92df7b51 DM |
2016 | /* Advance the send_head. This one is sent out. |
2017 | * This call will increment packets_out. | |
2018 | */ | |
66f5fe62 | 2019 | tcp_event_new_data_sent(sk, skb); |
1da177e4 | 2020 | |
92df7b51 | 2021 | tcp_minshall_update(tp, mss_now, skb); |
a262f0cd | 2022 | sent_pkts += tcp_skb_pcount(skb); |
d5dd9175 IJ |
2023 | |
2024 | if (push_one) | |
2025 | break; | |
92df7b51 | 2026 | } |
a262f0cd ND |
2027 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery) |
2028 | tp->prr_out += sent_pkts; | |
1da177e4 | 2029 | |
aa93466b | 2030 | if (likely(sent_pkts)) { |
9e412ba7 | 2031 | tcp_cwnd_validate(sk); |
a2a385d6 | 2032 | return false; |
1da177e4 | 2033 | } |
fe067e8a | 2034 | return !tp->packets_out && tcp_send_head(sk); |
1da177e4 LT |
2035 | } |
2036 | ||
a762a980 DM |
2037 | /* Push out any pending frames which were held back due to |
2038 | * TCP_CORK or attempt at coalescing tiny packets. | |
2039 | * The socket must be locked by the caller. | |
2040 | */ | |
9e412ba7 IJ |
2041 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, |
2042 | int nonagle) | |
a762a980 | 2043 | { |
726e07a8 IJ |
2044 | /* If we are closed, the bytes will have to remain here. |
2045 | * In time closedown will finish, we empty the write queue and | |
2046 | * all will be happy. | |
2047 | */ | |
2048 | if (unlikely(sk->sk_state == TCP_CLOSE)) | |
2049 | return; | |
2050 | ||
99a1dec7 MG |
2051 | if (tcp_write_xmit(sk, cur_mss, nonagle, 0, |
2052 | sk_gfp_atomic(sk, GFP_ATOMIC))) | |
726e07a8 | 2053 | tcp_check_probe_timer(sk); |
a762a980 DM |
2054 | } |
2055 | ||
c1b4a7e6 DM |
2056 | /* Send _single_ skb sitting at the send head. This function requires |
2057 | * true push pending frames to setup probe timer etc. | |
2058 | */ | |
2059 | void tcp_push_one(struct sock *sk, unsigned int mss_now) | |
2060 | { | |
fe067e8a | 2061 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 DM |
2062 | |
2063 | BUG_ON(!skb || skb->len < mss_now); | |
2064 | ||
d5dd9175 | 2065 | tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); |
c1b4a7e6 DM |
2066 | } |
2067 | ||
1da177e4 LT |
2068 | /* This function returns the amount that we can raise the |
2069 | * usable window based on the following constraints | |
e905a9ed | 2070 | * |
1da177e4 LT |
2071 | * 1. The window can never be shrunk once it is offered (RFC 793) |
2072 | * 2. We limit memory per socket | |
2073 | * | |
2074 | * RFC 1122: | |
2075 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep | |
2076 | * RECV.NEXT + RCV.WIN fixed until: | |
2077 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" | |
2078 | * | |
2079 | * i.e. don't raise the right edge of the window until you can raise | |
2080 | * it at least MSS bytes. | |
2081 | * | |
2082 | * Unfortunately, the recommended algorithm breaks header prediction, | |
2083 | * since header prediction assumes th->window stays fixed. | |
2084 | * | |
2085 | * Strictly speaking, keeping th->window fixed violates the receiver | |
2086 | * side SWS prevention criteria. The problem is that under this rule | |
2087 | * a stream of single byte packets will cause the right side of the | |
2088 | * window to always advance by a single byte. | |
e905a9ed | 2089 | * |
1da177e4 LT |
2090 | * Of course, if the sender implements sender side SWS prevention |
2091 | * then this will not be a problem. | |
e905a9ed | 2092 | * |
1da177e4 | 2093 | * BSD seems to make the following compromise: |
e905a9ed | 2094 | * |
1da177e4 LT |
2095 | * If the free space is less than the 1/4 of the maximum |
2096 | * space available and the free space is less than 1/2 mss, | |
2097 | * then set the window to 0. | |
2098 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] | |
2099 | * Otherwise, just prevent the window from shrinking | |
2100 | * and from being larger than the largest representable value. | |
2101 | * | |
2102 | * This prevents incremental opening of the window in the regime | |
2103 | * where TCP is limited by the speed of the reader side taking | |
2104 | * data out of the TCP receive queue. It does nothing about | |
2105 | * those cases where the window is constrained on the sender side | |
2106 | * because the pipeline is full. | |
2107 | * | |
2108 | * BSD also seems to "accidentally" limit itself to windows that are a | |
2109 | * multiple of MSS, at least until the free space gets quite small. | |
2110 | * This would appear to be a side effect of the mbuf implementation. | |
2111 | * Combining these two algorithms results in the observed behavior | |
2112 | * of having a fixed window size at almost all times. | |
2113 | * | |
2114 | * Below we obtain similar behavior by forcing the offered window to | |
2115 | * a multiple of the mss when it is feasible to do so. | |
2116 | * | |
2117 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. | |
2118 | * Regular options like TIMESTAMP are taken into account. | |
2119 | */ | |
2120 | u32 __tcp_select_window(struct sock *sk) | |
2121 | { | |
463c84b9 | 2122 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 2123 | struct tcp_sock *tp = tcp_sk(sk); |
caa20d9a | 2124 | /* MSS for the peer's data. Previous versions used mss_clamp |
1da177e4 LT |
2125 | * here. I don't know if the value based on our guesses |
2126 | * of peer's MSS is better for the performance. It's more correct | |
2127 | * but may be worse for the performance because of rcv_mss | |
2128 | * fluctuations. --SAW 1998/11/1 | |
2129 | */ | |
463c84b9 | 2130 | int mss = icsk->icsk_ack.rcv_mss; |
1da177e4 LT |
2131 | int free_space = tcp_space(sk); |
2132 | int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk)); | |
2133 | int window; | |
2134 | ||
2135 | if (mss > full_space) | |
e905a9ed | 2136 | mss = full_space; |
1da177e4 | 2137 | |
b92edbe0 | 2138 | if (free_space < (full_space >> 1)) { |
463c84b9 | 2139 | icsk->icsk_ack.quick = 0; |
1da177e4 | 2140 | |
180d8cd9 | 2141 | if (sk_under_memory_pressure(sk)) |
056834d9 IJ |
2142 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, |
2143 | 4U * tp->advmss); | |
1da177e4 LT |
2144 | |
2145 | if (free_space < mss) | |
2146 | return 0; | |
2147 | } | |
2148 | ||
2149 | if (free_space > tp->rcv_ssthresh) | |
2150 | free_space = tp->rcv_ssthresh; | |
2151 | ||
2152 | /* Don't do rounding if we are using window scaling, since the | |
2153 | * scaled window will not line up with the MSS boundary anyway. | |
2154 | */ | |
2155 | window = tp->rcv_wnd; | |
2156 | if (tp->rx_opt.rcv_wscale) { | |
2157 | window = free_space; | |
2158 | ||
2159 | /* Advertise enough space so that it won't get scaled away. | |
2160 | * Import case: prevent zero window announcement if | |
2161 | * 1<<rcv_wscale > mss. | |
2162 | */ | |
2163 | if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) | |
2164 | window = (((window >> tp->rx_opt.rcv_wscale) + 1) | |
2165 | << tp->rx_opt.rcv_wscale); | |
2166 | } else { | |
2167 | /* Get the largest window that is a nice multiple of mss. | |
2168 | * Window clamp already applied above. | |
2169 | * If our current window offering is within 1 mss of the | |
2170 | * free space we just keep it. This prevents the divide | |
2171 | * and multiply from happening most of the time. | |
2172 | * We also don't do any window rounding when the free space | |
2173 | * is too small. | |
2174 | */ | |
2175 | if (window <= free_space - mss || window > free_space) | |
056834d9 | 2176 | window = (free_space / mss) * mss; |
84565070 | 2177 | else if (mss == full_space && |
b92edbe0 | 2178 | free_space > window + (full_space >> 1)) |
84565070 | 2179 | window = free_space; |
1da177e4 LT |
2180 | } |
2181 | ||
2182 | return window; | |
2183 | } | |
2184 | ||
4a17fc3a IJ |
2185 | /* Collapses two adjacent SKB's during retransmission. */ |
2186 | static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) | |
1da177e4 LT |
2187 | { |
2188 | struct tcp_sock *tp = tcp_sk(sk); | |
fe067e8a | 2189 | struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); |
058dc334 | 2190 | int skb_size, next_skb_size; |
1da177e4 | 2191 | |
058dc334 IJ |
2192 | skb_size = skb->len; |
2193 | next_skb_size = next_skb->len; | |
1da177e4 | 2194 | |
058dc334 | 2195 | BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); |
a6963a6b | 2196 | |
058dc334 | 2197 | tcp_highest_sack_combine(sk, next_skb, skb); |
1da177e4 | 2198 | |
058dc334 | 2199 | tcp_unlink_write_queue(next_skb, sk); |
1da177e4 | 2200 | |
058dc334 IJ |
2201 | skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), |
2202 | next_skb_size); | |
1da177e4 | 2203 | |
058dc334 IJ |
2204 | if (next_skb->ip_summed == CHECKSUM_PARTIAL) |
2205 | skb->ip_summed = CHECKSUM_PARTIAL; | |
1da177e4 | 2206 | |
058dc334 IJ |
2207 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
2208 | skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); | |
1da177e4 | 2209 | |
058dc334 IJ |
2210 | /* Update sequence range on original skb. */ |
2211 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; | |
1da177e4 | 2212 | |
e6c7d085 | 2213 | /* Merge over control information. This moves PSH/FIN etc. over */ |
4de075e0 | 2214 | TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; |
058dc334 IJ |
2215 | |
2216 | /* All done, get rid of second SKB and account for it so | |
2217 | * packet counting does not break. | |
2218 | */ | |
2219 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; | |
058dc334 IJ |
2220 | |
2221 | /* changed transmit queue under us so clear hints */ | |
ef9da47c IJ |
2222 | tcp_clear_retrans_hints_partial(tp); |
2223 | if (next_skb == tp->retransmit_skb_hint) | |
2224 | tp->retransmit_skb_hint = skb; | |
058dc334 | 2225 | |
797108d1 IJ |
2226 | tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); |
2227 | ||
058dc334 | 2228 | sk_wmem_free_skb(sk, next_skb); |
1da177e4 LT |
2229 | } |
2230 | ||
67edfef7 | 2231 | /* Check if coalescing SKBs is legal. */ |
a2a385d6 | 2232 | static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) |
4a17fc3a IJ |
2233 | { |
2234 | if (tcp_skb_pcount(skb) > 1) | |
a2a385d6 | 2235 | return false; |
4a17fc3a IJ |
2236 | /* TODO: SACK collapsing could be used to remove this condition */ |
2237 | if (skb_shinfo(skb)->nr_frags != 0) | |
a2a385d6 | 2238 | return false; |
4a17fc3a | 2239 | if (skb_cloned(skb)) |
a2a385d6 | 2240 | return false; |
4a17fc3a | 2241 | if (skb == tcp_send_head(sk)) |
a2a385d6 | 2242 | return false; |
4a17fc3a IJ |
2243 | /* Some heurestics for collapsing over SACK'd could be invented */ |
2244 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
a2a385d6 | 2245 | return false; |
4a17fc3a | 2246 | |
a2a385d6 | 2247 | return true; |
4a17fc3a IJ |
2248 | } |
2249 | ||
67edfef7 AK |
2250 | /* Collapse packets in the retransmit queue to make to create |
2251 | * less packets on the wire. This is only done on retransmission. | |
2252 | */ | |
4a17fc3a IJ |
2253 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, |
2254 | int space) | |
2255 | { | |
2256 | struct tcp_sock *tp = tcp_sk(sk); | |
2257 | struct sk_buff *skb = to, *tmp; | |
a2a385d6 | 2258 | bool first = true; |
4a17fc3a IJ |
2259 | |
2260 | if (!sysctl_tcp_retrans_collapse) | |
2261 | return; | |
4de075e0 | 2262 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) |
4a17fc3a IJ |
2263 | return; |
2264 | ||
2265 | tcp_for_write_queue_from_safe(skb, tmp, sk) { | |
2266 | if (!tcp_can_collapse(sk, skb)) | |
2267 | break; | |
2268 | ||
2269 | space -= skb->len; | |
2270 | ||
2271 | if (first) { | |
a2a385d6 | 2272 | first = false; |
4a17fc3a IJ |
2273 | continue; |
2274 | } | |
2275 | ||
2276 | if (space < 0) | |
2277 | break; | |
2278 | /* Punt if not enough space exists in the first SKB for | |
2279 | * the data in the second | |
2280 | */ | |
a21d4572 | 2281 | if (skb->len > skb_availroom(to)) |
4a17fc3a IJ |
2282 | break; |
2283 | ||
2284 | if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) | |
2285 | break; | |
2286 | ||
2287 | tcp_collapse_retrans(sk, to); | |
2288 | } | |
2289 | } | |
2290 | ||
1da177e4 LT |
2291 | /* This retransmits one SKB. Policy decisions and retransmit queue |
2292 | * state updates are done by the caller. Returns non-zero if an | |
2293 | * error occurred which prevented the send. | |
2294 | */ | |
2295 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb) | |
2296 | { | |
2297 | struct tcp_sock *tp = tcp_sk(sk); | |
5d424d5a | 2298 | struct inet_connection_sock *icsk = inet_csk(sk); |
7d227cd2 | 2299 | unsigned int cur_mss; |
1da177e4 LT |
2300 | int err; |
2301 | ||
5d424d5a JH |
2302 | /* Inconslusive MTU probe */ |
2303 | if (icsk->icsk_mtup.probe_size) { | |
2304 | icsk->icsk_mtup.probe_size = 0; | |
2305 | } | |
2306 | ||
1da177e4 | 2307 | /* Do not sent more than we queued. 1/4 is reserved for possible |
caa20d9a | 2308 | * copying overhead: fragmentation, tunneling, mangling etc. |
1da177e4 LT |
2309 | */ |
2310 | if (atomic_read(&sk->sk_wmem_alloc) > | |
2311 | min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) | |
2312 | return -EAGAIN; | |
2313 | ||
2314 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { | |
2315 | if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
2316 | BUG(); | |
1da177e4 LT |
2317 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
2318 | return -ENOMEM; | |
2319 | } | |
2320 | ||
7d227cd2 SS |
2321 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) |
2322 | return -EHOSTUNREACH; /* Routing failure or similar. */ | |
2323 | ||
0c54b85f | 2324 | cur_mss = tcp_current_mss(sk); |
7d227cd2 | 2325 | |
1da177e4 LT |
2326 | /* If receiver has shrunk his window, and skb is out of |
2327 | * new window, do not retransmit it. The exception is the | |
2328 | * case, when window is shrunk to zero. In this case | |
2329 | * our retransmit serves as a zero window probe. | |
2330 | */ | |
9d4fb27d JP |
2331 | if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && |
2332 | TCP_SKB_CB(skb)->seq != tp->snd_una) | |
1da177e4 LT |
2333 | return -EAGAIN; |
2334 | ||
2335 | if (skb->len > cur_mss) { | |
846998ae | 2336 | if (tcp_fragment(sk, skb, cur_mss, cur_mss)) |
1da177e4 | 2337 | return -ENOMEM; /* We'll try again later. */ |
02276f3c | 2338 | } else { |
9eb9362e IJ |
2339 | int oldpcount = tcp_skb_pcount(skb); |
2340 | ||
2341 | if (unlikely(oldpcount > 1)) { | |
2342 | tcp_init_tso_segs(sk, skb, cur_mss); | |
2343 | tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb)); | |
2344 | } | |
1da177e4 LT |
2345 | } |
2346 | ||
4a17fc3a | 2347 | tcp_retrans_try_collapse(sk, skb, cur_mss); |
1da177e4 | 2348 | |
1da177e4 LT |
2349 | /* Some Solaris stacks overoptimize and ignore the FIN on a |
2350 | * retransmit when old data is attached. So strip it off | |
2351 | * since it is cheap to do so and saves bytes on the network. | |
2352 | */ | |
2de979bd | 2353 | if (skb->len > 0 && |
4de075e0 | 2354 | (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && |
2de979bd | 2355 | tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) { |
1da177e4 | 2356 | if (!pskb_trim(skb, 0)) { |
e870a8ef IJ |
2357 | /* Reuse, even though it does some unnecessary work */ |
2358 | tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1, | |
4de075e0 | 2359 | TCP_SKB_CB(skb)->tcp_flags); |
1da177e4 | 2360 | skb->ip_summed = CHECKSUM_NONE; |
1da177e4 LT |
2361 | } |
2362 | } | |
2363 | ||
2364 | /* Make a copy, if the first transmission SKB clone we made | |
2365 | * is still in somebody's hands, else make a clone. | |
2366 | */ | |
2367 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
1da177e4 | 2368 | |
117632e6 ED |
2369 | /* make sure skb->data is aligned on arches that require it */ |
2370 | if (unlikely(NET_IP_ALIGN && ((unsigned long)skb->data & 3))) { | |
2371 | struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER, | |
2372 | GFP_ATOMIC); | |
2373 | err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : | |
2374 | -ENOBUFS; | |
2375 | } else { | |
2376 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); | |
2377 | } | |
1da177e4 LT |
2378 | |
2379 | if (err == 0) { | |
2380 | /* Update global TCP statistics. */ | |
81cc8a75 | 2381 | TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); |
1da177e4 LT |
2382 | |
2383 | tp->total_retrans++; | |
2384 | ||
2385 | #if FASTRETRANS_DEBUG > 0 | |
056834d9 | 2386 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
e87cc472 | 2387 | net_dbg_ratelimited("retrans_out leaked\n"); |
1da177e4 LT |
2388 | } |
2389 | #endif | |
b08d6cb2 IJ |
2390 | if (!tp->retrans_out) |
2391 | tp->lost_retrans_low = tp->snd_nxt; | |
1da177e4 LT |
2392 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; |
2393 | tp->retrans_out += tcp_skb_pcount(skb); | |
2394 | ||
2395 | /* Save stamp of the first retransmit. */ | |
2396 | if (!tp->retrans_stamp) | |
2397 | tp->retrans_stamp = TCP_SKB_CB(skb)->when; | |
2398 | ||
c24f691b | 2399 | tp->undo_retrans += tcp_skb_pcount(skb); |
1da177e4 LT |
2400 | |
2401 | /* snd_nxt is stored to detect loss of retransmitted segment, | |
2402 | * see tcp_input.c tcp_sacktag_write_queue(). | |
2403 | */ | |
2404 | TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt; | |
2405 | } | |
2406 | return err; | |
2407 | } | |
2408 | ||
67edfef7 AK |
2409 | /* Check if we forward retransmits are possible in the current |
2410 | * window/congestion state. | |
2411 | */ | |
a2a385d6 | 2412 | static bool tcp_can_forward_retransmit(struct sock *sk) |
b5afe7bc IJ |
2413 | { |
2414 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
cf533ea5 | 2415 | const struct tcp_sock *tp = tcp_sk(sk); |
b5afe7bc IJ |
2416 | |
2417 | /* Forward retransmissions are possible only during Recovery. */ | |
2418 | if (icsk->icsk_ca_state != TCP_CA_Recovery) | |
a2a385d6 | 2419 | return false; |
b5afe7bc IJ |
2420 | |
2421 | /* No forward retransmissions in Reno are possible. */ | |
2422 | if (tcp_is_reno(tp)) | |
a2a385d6 | 2423 | return false; |
b5afe7bc IJ |
2424 | |
2425 | /* Yeah, we have to make difficult choice between forward transmission | |
2426 | * and retransmission... Both ways have their merits... | |
2427 | * | |
2428 | * For now we do not retransmit anything, while we have some new | |
2429 | * segments to send. In the other cases, follow rule 3 for | |
2430 | * NextSeg() specified in RFC3517. | |
2431 | */ | |
2432 | ||
2433 | if (tcp_may_send_now(sk)) | |
a2a385d6 | 2434 | return false; |
b5afe7bc | 2435 | |
a2a385d6 | 2436 | return true; |
b5afe7bc IJ |
2437 | } |
2438 | ||
1da177e4 LT |
2439 | /* This gets called after a retransmit timeout, and the initially |
2440 | * retransmitted data is acknowledged. It tries to continue | |
2441 | * resending the rest of the retransmit queue, until either | |
2442 | * we've sent it all or the congestion window limit is reached. | |
2443 | * If doing SACK, the first ACK which comes back for a timeout | |
2444 | * based retransmit packet might feed us FACK information again. | |
2445 | * If so, we use it to avoid unnecessarily retransmissions. | |
2446 | */ | |
2447 | void tcp_xmit_retransmit_queue(struct sock *sk) | |
2448 | { | |
6687e988 | 2449 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2450 | struct tcp_sock *tp = tcp_sk(sk); |
2451 | struct sk_buff *skb; | |
0e1c54c2 | 2452 | struct sk_buff *hole = NULL; |
618d9f25 | 2453 | u32 last_lost; |
61eb55f4 | 2454 | int mib_idx; |
0e1c54c2 | 2455 | int fwd_rexmitting = 0; |
6a438bbe | 2456 | |
45e77d31 IJ |
2457 | if (!tp->packets_out) |
2458 | return; | |
2459 | ||
08ebd172 IJ |
2460 | if (!tp->lost_out) |
2461 | tp->retransmit_high = tp->snd_una; | |
2462 | ||
618d9f25 | 2463 | if (tp->retransmit_skb_hint) { |
6a438bbe | 2464 | skb = tp->retransmit_skb_hint; |
618d9f25 IJ |
2465 | last_lost = TCP_SKB_CB(skb)->end_seq; |
2466 | if (after(last_lost, tp->retransmit_high)) | |
2467 | last_lost = tp->retransmit_high; | |
2468 | } else { | |
fe067e8a | 2469 | skb = tcp_write_queue_head(sk); |
618d9f25 IJ |
2470 | last_lost = tp->snd_una; |
2471 | } | |
1da177e4 | 2472 | |
08ebd172 IJ |
2473 | tcp_for_write_queue_from(skb, sk) { |
2474 | __u8 sacked = TCP_SKB_CB(skb)->sacked; | |
1da177e4 | 2475 | |
08ebd172 IJ |
2476 | if (skb == tcp_send_head(sk)) |
2477 | break; | |
2478 | /* we could do better than to assign each time */ | |
0e1c54c2 IJ |
2479 | if (hole == NULL) |
2480 | tp->retransmit_skb_hint = skb; | |
08ebd172 IJ |
2481 | |
2482 | /* Assume this retransmit will generate | |
2483 | * only one packet for congestion window | |
2484 | * calculation purposes. This works because | |
2485 | * tcp_retransmit_skb() will chop up the | |
2486 | * packet to be MSS sized and all the | |
2487 | * packet counting works out. | |
2488 | */ | |
2489 | if (tcp_packets_in_flight(tp) >= tp->snd_cwnd) | |
2490 | return; | |
1da177e4 | 2491 | |
0e1c54c2 IJ |
2492 | if (fwd_rexmitting) { |
2493 | begin_fwd: | |
2494 | if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) | |
2495 | break; | |
2496 | mib_idx = LINUX_MIB_TCPFORWARDRETRANS; | |
6a438bbe | 2497 | |
0e1c54c2 | 2498 | } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { |
618d9f25 | 2499 | tp->retransmit_high = last_lost; |
0e1c54c2 IJ |
2500 | if (!tcp_can_forward_retransmit(sk)) |
2501 | break; | |
2502 | /* Backtrack if necessary to non-L'ed skb */ | |
2503 | if (hole != NULL) { | |
2504 | skb = hole; | |
2505 | hole = NULL; | |
2506 | } | |
2507 | fwd_rexmitting = 1; | |
2508 | goto begin_fwd; | |
1da177e4 | 2509 | |
0e1c54c2 | 2510 | } else if (!(sacked & TCPCB_LOST)) { |
ac11ba75 | 2511 | if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) |
0e1c54c2 IJ |
2512 | hole = skb; |
2513 | continue; | |
1da177e4 | 2514 | |
0e1c54c2 | 2515 | } else { |
618d9f25 | 2516 | last_lost = TCP_SKB_CB(skb)->end_seq; |
0e1c54c2 IJ |
2517 | if (icsk->icsk_ca_state != TCP_CA_Loss) |
2518 | mib_idx = LINUX_MIB_TCPFASTRETRANS; | |
2519 | else | |
2520 | mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; | |
2521 | } | |
1da177e4 | 2522 | |
0e1c54c2 | 2523 | if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) |
1da177e4 LT |
2524 | continue; |
2525 | ||
09e9b813 ED |
2526 | if (tcp_retransmit_skb(sk, skb)) { |
2527 | NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); | |
0e1c54c2 | 2528 | return; |
09e9b813 | 2529 | } |
0e1c54c2 | 2530 | NET_INC_STATS_BH(sock_net(sk), mib_idx); |
1da177e4 | 2531 | |
a262f0cd ND |
2532 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery) |
2533 | tp->prr_out += tcp_skb_pcount(skb); | |
2534 | ||
fe067e8a | 2535 | if (skb == tcp_write_queue_head(sk)) |
3f421baa ACM |
2536 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
2537 | inet_csk(sk)->icsk_rto, | |
2538 | TCP_RTO_MAX); | |
1da177e4 LT |
2539 | } |
2540 | } | |
2541 | ||
1da177e4 LT |
2542 | /* Send a fin. The caller locks the socket for us. This cannot be |
2543 | * allowed to fail queueing a FIN frame under any circumstances. | |
2544 | */ | |
2545 | void tcp_send_fin(struct sock *sk) | |
2546 | { | |
e905a9ed | 2547 | struct tcp_sock *tp = tcp_sk(sk); |
fe067e8a | 2548 | struct sk_buff *skb = tcp_write_queue_tail(sk); |
1da177e4 | 2549 | int mss_now; |
e905a9ed | 2550 | |
1da177e4 LT |
2551 | /* Optimization, tack on the FIN if we have a queue of |
2552 | * unsent frames. But be careful about outgoing SACKS | |
2553 | * and IP options. | |
2554 | */ | |
0c54b85f | 2555 | mss_now = tcp_current_mss(sk); |
1da177e4 | 2556 | |
fe067e8a | 2557 | if (tcp_send_head(sk) != NULL) { |
4de075e0 | 2558 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN; |
1da177e4 LT |
2559 | TCP_SKB_CB(skb)->end_seq++; |
2560 | tp->write_seq++; | |
2561 | } else { | |
2562 | /* Socket is locked, keep trying until memory is available. */ | |
2563 | for (;;) { | |
aa133076 WF |
2564 | skb = alloc_skb_fclone(MAX_TCP_HEADER, |
2565 | sk->sk_allocation); | |
1da177e4 LT |
2566 | if (skb) |
2567 | break; | |
2568 | yield(); | |
2569 | } | |
2570 | ||
2571 | /* Reserve space for headers and prepare control bits. */ | |
2572 | skb_reserve(skb, MAX_TCP_HEADER); | |
1da177e4 | 2573 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ |
e870a8ef | 2574 | tcp_init_nondata_skb(skb, tp->write_seq, |
a3433f35 | 2575 | TCPHDR_ACK | TCPHDR_FIN); |
1da177e4 LT |
2576 | tcp_queue_skb(sk, skb); |
2577 | } | |
9e412ba7 | 2578 | __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF); |
1da177e4 LT |
2579 | } |
2580 | ||
2581 | /* We get here when a process closes a file descriptor (either due to | |
2582 | * an explicit close() or as a byproduct of exit()'ing) and there | |
2583 | * was unread data in the receive queue. This behavior is recommended | |
65bb723c | 2584 | * by RFC 2525, section 2.17. -DaveM |
1da177e4 | 2585 | */ |
dd0fc66f | 2586 | void tcp_send_active_reset(struct sock *sk, gfp_t priority) |
1da177e4 | 2587 | { |
1da177e4 LT |
2588 | struct sk_buff *skb; |
2589 | ||
2590 | /* NOTE: No TCP options attached and we never retransmit this. */ | |
2591 | skb = alloc_skb(MAX_TCP_HEADER, priority); | |
2592 | if (!skb) { | |
4e673444 | 2593 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
1da177e4 LT |
2594 | return; |
2595 | } | |
2596 | ||
2597 | /* Reserve space for headers and prepare control bits. */ | |
2598 | skb_reserve(skb, MAX_TCP_HEADER); | |
e870a8ef | 2599 | tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), |
a3433f35 | 2600 | TCPHDR_ACK | TCPHDR_RST); |
1da177e4 | 2601 | /* Send it off. */ |
1da177e4 | 2602 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
dfb4b9dc | 2603 | if (tcp_transmit_skb(sk, skb, 0, priority)) |
4e673444 | 2604 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
26af65cb | 2605 | |
81cc8a75 | 2606 | TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); |
1da177e4 LT |
2607 | } |
2608 | ||
67edfef7 AK |
2609 | /* Send a crossed SYN-ACK during socket establishment. |
2610 | * WARNING: This routine must only be called when we have already sent | |
1da177e4 LT |
2611 | * a SYN packet that crossed the incoming SYN that caused this routine |
2612 | * to get called. If this assumption fails then the initial rcv_wnd | |
2613 | * and rcv_wscale values will not be correct. | |
2614 | */ | |
2615 | int tcp_send_synack(struct sock *sk) | |
2616 | { | |
056834d9 | 2617 | struct sk_buff *skb; |
1da177e4 | 2618 | |
fe067e8a | 2619 | skb = tcp_write_queue_head(sk); |
4de075e0 | 2620 | if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { |
91df42be | 2621 | pr_debug("%s: wrong queue state\n", __func__); |
1da177e4 LT |
2622 | return -EFAULT; |
2623 | } | |
4de075e0 | 2624 | if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { |
1da177e4 LT |
2625 | if (skb_cloned(skb)) { |
2626 | struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); | |
2627 | if (nskb == NULL) | |
2628 | return -ENOMEM; | |
fe067e8a | 2629 | tcp_unlink_write_queue(skb, sk); |
1da177e4 | 2630 | skb_header_release(nskb); |
fe067e8a | 2631 | __tcp_add_write_queue_head(sk, nskb); |
3ab224be HA |
2632 | sk_wmem_free_skb(sk, skb); |
2633 | sk->sk_wmem_queued += nskb->truesize; | |
2634 | sk_mem_charge(sk, nskb->truesize); | |
1da177e4 LT |
2635 | skb = nskb; |
2636 | } | |
2637 | ||
4de075e0 | 2638 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; |
1da177e4 LT |
2639 | TCP_ECN_send_synack(tcp_sk(sk), skb); |
2640 | } | |
2641 | TCP_SKB_CB(skb)->when = tcp_time_stamp; | |
dfb4b9dc | 2642 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
1da177e4 LT |
2643 | } |
2644 | ||
4aea39c1 ED |
2645 | /** |
2646 | * tcp_make_synack - Prepare a SYN-ACK. | |
2647 | * sk: listener socket | |
2648 | * dst: dst entry attached to the SYNACK | |
2649 | * req: request_sock pointer | |
2650 | * rvp: request_values pointer | |
2651 | * | |
2652 | * Allocate one skb and build a SYNACK packet. | |
2653 | * @dst is consumed : Caller should not use it again. | |
2654 | */ | |
056834d9 | 2655 | struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst, |
e6b4d113 WAS |
2656 | struct request_sock *req, |
2657 | struct request_values *rvp) | |
1da177e4 | 2658 | { |
bd0388ae | 2659 | struct tcp_out_options opts; |
4957faad | 2660 | struct tcp_extend_values *xvp = tcp_xv(rvp); |
2e6599cb | 2661 | struct inet_request_sock *ireq = inet_rsk(req); |
1da177e4 | 2662 | struct tcp_sock *tp = tcp_sk(sk); |
28b2774a | 2663 | const struct tcp_cookie_values *cvp = tp->cookie_values; |
1da177e4 | 2664 | struct tcphdr *th; |
1da177e4 | 2665 | struct sk_buff *skb; |
cfb6eeb4 | 2666 | struct tcp_md5sig_key *md5; |
bd0388ae | 2667 | int tcp_header_size; |
f5fff5dc | 2668 | int mss; |
28b2774a | 2669 | int s_data_desired = 0; |
1da177e4 | 2670 | |
28b2774a ED |
2671 | if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired) |
2672 | s_data_desired = cvp->s_data_desired; | |
99a1dec7 MG |
2673 | skb = alloc_skb(MAX_TCP_HEADER + 15 + s_data_desired, |
2674 | sk_gfp_atomic(sk, GFP_ATOMIC)); | |
4aea39c1 ED |
2675 | if (unlikely(!skb)) { |
2676 | dst_release(dst); | |
1da177e4 | 2677 | return NULL; |
4aea39c1 | 2678 | } |
1da177e4 LT |
2679 | /* Reserve space for headers. */ |
2680 | skb_reserve(skb, MAX_TCP_HEADER); | |
2681 | ||
4aea39c1 | 2682 | skb_dst_set(skb, dst); |
1da177e4 | 2683 | |
0dbaee3b | 2684 | mss = dst_metric_advmss(dst); |
f5fff5dc TQ |
2685 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss) |
2686 | mss = tp->rx_opt.user_mss; | |
2687 | ||
33ad798c AL |
2688 | if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */ |
2689 | __u8 rcv_wscale; | |
2690 | /* Set this up on the first call only */ | |
2691 | req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW); | |
e88c64f0 HPP |
2692 | |
2693 | /* limit the window selection if the user enforce a smaller rx buffer */ | |
2694 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && | |
2695 | (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0)) | |
2696 | req->window_clamp = tcp_full_space(sk); | |
2697 | ||
33ad798c AL |
2698 | /* tcp_full_space because it is guaranteed to be the first packet */ |
2699 | tcp_select_initial_window(tcp_full_space(sk), | |
f5fff5dc | 2700 | mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0), |
33ad798c AL |
2701 | &req->rcv_wnd, |
2702 | &req->window_clamp, | |
2703 | ireq->wscale_ok, | |
31d12926 | 2704 | &rcv_wscale, |
2705 | dst_metric(dst, RTAX_INITRWND)); | |
33ad798c AL |
2706 | ireq->rcv_wscale = rcv_wscale; |
2707 | } | |
2708 | ||
2709 | memset(&opts, 0, sizeof(opts)); | |
8b5f12d0 FW |
2710 | #ifdef CONFIG_SYN_COOKIES |
2711 | if (unlikely(req->cookie_ts)) | |
2712 | TCP_SKB_CB(skb)->when = cookie_init_timestamp(req); | |
2713 | else | |
2714 | #endif | |
33ad798c | 2715 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
f5fff5dc | 2716 | tcp_header_size = tcp_synack_options(sk, req, mss, |
4957faad WAS |
2717 | skb, &opts, &md5, xvp) |
2718 | + sizeof(*th); | |
cfb6eeb4 | 2719 | |
aa8223c7 ACM |
2720 | skb_push(skb, tcp_header_size); |
2721 | skb_reset_transport_header(skb); | |
1da177e4 | 2722 | |
aa8223c7 | 2723 | th = tcp_hdr(skb); |
1da177e4 LT |
2724 | memset(th, 0, sizeof(struct tcphdr)); |
2725 | th->syn = 1; | |
2726 | th->ack = 1; | |
1da177e4 | 2727 | TCP_ECN_make_synack(req, th); |
a3116ac5 | 2728 | th->source = ireq->loc_port; |
2e6599cb | 2729 | th->dest = ireq->rmt_port; |
e870a8ef IJ |
2730 | /* Setting of flags are superfluous here for callers (and ECE is |
2731 | * not even correctly set) | |
2732 | */ | |
2733 | tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, | |
a3433f35 | 2734 | TCPHDR_SYN | TCPHDR_ACK); |
4957faad WAS |
2735 | |
2736 | if (OPTION_COOKIE_EXTENSION & opts.options) { | |
28b2774a ED |
2737 | if (s_data_desired) { |
2738 | u8 *buf = skb_put(skb, s_data_desired); | |
4957faad WAS |
2739 | |
2740 | /* copy data directly from the listening socket. */ | |
28b2774a ED |
2741 | memcpy(buf, cvp->s_data_payload, s_data_desired); |
2742 | TCP_SKB_CB(skb)->end_seq += s_data_desired; | |
4957faad WAS |
2743 | } |
2744 | ||
2745 | if (opts.hash_size > 0) { | |
2746 | __u32 workspace[SHA_WORKSPACE_WORDS]; | |
2747 | u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS]; | |
2748 | u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1]; | |
2749 | ||
2750 | /* Secret recipe depends on the Timestamp, (future) | |
2751 | * Sequence and Acknowledgment Numbers, Initiator | |
2752 | * Cookie, and others handled by IP variant caller. | |
2753 | */ | |
2754 | *tail-- ^= opts.tsval; | |
2755 | *tail-- ^= tcp_rsk(req)->rcv_isn + 1; | |
2756 | *tail-- ^= TCP_SKB_CB(skb)->seq + 1; | |
2757 | ||
2758 | /* recommended */ | |
0eae88f3 | 2759 | *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source); |
f9a2e69e | 2760 | *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */ |
4957faad WAS |
2761 | |
2762 | sha_transform((__u32 *)&xvp->cookie_bakery[0], | |
2763 | (char *)mess, | |
2764 | &workspace[0]); | |
2765 | opts.hash_location = | |
2766 | (__u8 *)&xvp->cookie_bakery[0]; | |
2767 | } | |
2768 | } | |
2769 | ||
1da177e4 | 2770 | th->seq = htonl(TCP_SKB_CB(skb)->seq); |
2e6599cb | 2771 | th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1); |
1da177e4 LT |
2772 | |
2773 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ | |
600ff0c2 | 2774 | th->window = htons(min(req->rcv_wnd, 65535U)); |
bd0388ae | 2775 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
1da177e4 | 2776 | th->doff = (tcp_header_size >> 2); |
aa2ea058 | 2777 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb)); |
cfb6eeb4 YH |
2778 | |
2779 | #ifdef CONFIG_TCP_MD5SIG | |
2780 | /* Okay, we have all we need - do the md5 hash if needed */ | |
2781 | if (md5) { | |
bd0388ae | 2782 | tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, |
49a72dfb | 2783 | md5, NULL, req, skb); |
cfb6eeb4 YH |
2784 | } |
2785 | #endif | |
2786 | ||
1da177e4 LT |
2787 | return skb; |
2788 | } | |
4bc2f18b | 2789 | EXPORT_SYMBOL(tcp_make_synack); |
1da177e4 | 2790 | |
67edfef7 | 2791 | /* Do all connect socket setups that can be done AF independent. */ |
370816ae | 2792 | void tcp_connect_init(struct sock *sk) |
1da177e4 | 2793 | { |
cf533ea5 | 2794 | const struct dst_entry *dst = __sk_dst_get(sk); |
1da177e4 LT |
2795 | struct tcp_sock *tp = tcp_sk(sk); |
2796 | __u8 rcv_wscale; | |
2797 | ||
2798 | /* We'll fix this up when we get a response from the other end. | |
2799 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. | |
2800 | */ | |
2801 | tp->tcp_header_len = sizeof(struct tcphdr) + | |
bb5b7c11 | 2802 | (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); |
1da177e4 | 2803 | |
cfb6eeb4 YH |
2804 | #ifdef CONFIG_TCP_MD5SIG |
2805 | if (tp->af_specific->md5_lookup(sk, sk) != NULL) | |
2806 | tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; | |
2807 | #endif | |
2808 | ||
1da177e4 LT |
2809 | /* If user gave his TCP_MAXSEG, record it to clamp */ |
2810 | if (tp->rx_opt.user_mss) | |
2811 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
2812 | tp->max_window = 0; | |
5d424d5a | 2813 | tcp_mtup_init(sk); |
1da177e4 LT |
2814 | tcp_sync_mss(sk, dst_mtu(dst)); |
2815 | ||
2816 | if (!tp->window_clamp) | |
2817 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); | |
0dbaee3b | 2818 | tp->advmss = dst_metric_advmss(dst); |
f5fff5dc TQ |
2819 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) |
2820 | tp->advmss = tp->rx_opt.user_mss; | |
2821 | ||
1da177e4 | 2822 | tcp_initialize_rcv_mss(sk); |
1da177e4 | 2823 | |
e88c64f0 HPP |
2824 | /* limit the window selection if the user enforce a smaller rx buffer */ |
2825 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && | |
2826 | (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) | |
2827 | tp->window_clamp = tcp_full_space(sk); | |
2828 | ||
1da177e4 LT |
2829 | tcp_select_initial_window(tcp_full_space(sk), |
2830 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), | |
2831 | &tp->rcv_wnd, | |
2832 | &tp->window_clamp, | |
bb5b7c11 | 2833 | sysctl_tcp_window_scaling, |
31d12926 | 2834 | &rcv_wscale, |
2835 | dst_metric(dst, RTAX_INITRWND)); | |
1da177e4 LT |
2836 | |
2837 | tp->rx_opt.rcv_wscale = rcv_wscale; | |
2838 | tp->rcv_ssthresh = tp->rcv_wnd; | |
2839 | ||
2840 | sk->sk_err = 0; | |
2841 | sock_reset_flag(sk, SOCK_DONE); | |
2842 | tp->snd_wnd = 0; | |
ee7537b6 | 2843 | tcp_init_wl(tp, 0); |
1da177e4 LT |
2844 | tp->snd_una = tp->write_seq; |
2845 | tp->snd_sml = tp->write_seq; | |
33f5f57e | 2846 | tp->snd_up = tp->write_seq; |
370816ae | 2847 | tp->snd_nxt = tp->write_seq; |
ee995283 PE |
2848 | |
2849 | if (likely(!tp->repair)) | |
2850 | tp->rcv_nxt = 0; | |
2851 | tp->rcv_wup = tp->rcv_nxt; | |
2852 | tp->copied_seq = tp->rcv_nxt; | |
1da177e4 | 2853 | |
463c84b9 ACM |
2854 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
2855 | inet_csk(sk)->icsk_retransmits = 0; | |
1da177e4 LT |
2856 | tcp_clear_retrans(tp); |
2857 | } | |
2858 | ||
783237e8 YC |
2859 | static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) |
2860 | { | |
2861 | struct tcp_sock *tp = tcp_sk(sk); | |
2862 | struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); | |
2863 | ||
2864 | tcb->end_seq += skb->len; | |
2865 | skb_header_release(skb); | |
2866 | __tcp_add_write_queue_tail(sk, skb); | |
2867 | sk->sk_wmem_queued += skb->truesize; | |
2868 | sk_mem_charge(sk, skb->truesize); | |
2869 | tp->write_seq = tcb->end_seq; | |
2870 | tp->packets_out += tcp_skb_pcount(skb); | |
2871 | } | |
2872 | ||
2873 | /* Build and send a SYN with data and (cached) Fast Open cookie. However, | |
2874 | * queue a data-only packet after the regular SYN, such that regular SYNs | |
2875 | * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges | |
2876 | * only the SYN sequence, the data are retransmitted in the first ACK. | |
2877 | * If cookie is not cached or other error occurs, falls back to send a | |
2878 | * regular SYN with Fast Open cookie request option. | |
2879 | */ | |
2880 | static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) | |
2881 | { | |
2882 | struct tcp_sock *tp = tcp_sk(sk); | |
2883 | struct tcp_fastopen_request *fo = tp->fastopen_req; | |
aab48743 | 2884 | int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen; |
783237e8 | 2885 | struct sk_buff *syn_data = NULL, *data; |
aab48743 YC |
2886 | unsigned long last_syn_loss = 0; |
2887 | ||
67da22d2 | 2888 | tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ |
aab48743 YC |
2889 | tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, |
2890 | &syn_loss, &last_syn_loss); | |
2891 | /* Recurring FO SYN losses: revert to regular handshake temporarily */ | |
2892 | if (syn_loss > 1 && | |
2893 | time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { | |
2894 | fo->cookie.len = -1; | |
2895 | goto fallback; | |
2896 | } | |
783237e8 | 2897 | |
67da22d2 YC |
2898 | if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) |
2899 | fo->cookie.len = -1; | |
2900 | else if (fo->cookie.len <= 0) | |
783237e8 YC |
2901 | goto fallback; |
2902 | ||
2903 | /* MSS for SYN-data is based on cached MSS and bounded by PMTU and | |
2904 | * user-MSS. Reserve maximum option space for middleboxes that add | |
2905 | * private TCP options. The cost is reduced data space in SYN :( | |
2906 | */ | |
2907 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) | |
2908 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
2909 | space = tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - | |
2910 | MAX_TCP_OPTION_SPACE; | |
2911 | ||
2912 | syn_data = skb_copy_expand(syn, skb_headroom(syn), space, | |
2913 | sk->sk_allocation); | |
2914 | if (syn_data == NULL) | |
2915 | goto fallback; | |
2916 | ||
2917 | for (i = 0; i < iovlen && syn_data->len < space; ++i) { | |
2918 | struct iovec *iov = &fo->data->msg_iov[i]; | |
2919 | unsigned char __user *from = iov->iov_base; | |
2920 | int len = iov->iov_len; | |
2921 | ||
2922 | if (syn_data->len + len > space) | |
2923 | len = space - syn_data->len; | |
2924 | else if (i + 1 == iovlen) | |
2925 | /* No more data pending in inet_wait_for_connect() */ | |
2926 | fo->data = NULL; | |
2927 | ||
2928 | if (skb_add_data(syn_data, from, len)) | |
2929 | goto fallback; | |
2930 | } | |
2931 | ||
2932 | /* Queue a data-only packet after the regular SYN for retransmission */ | |
2933 | data = pskb_copy(syn_data, sk->sk_allocation); | |
2934 | if (data == NULL) | |
2935 | goto fallback; | |
2936 | TCP_SKB_CB(data)->seq++; | |
2937 | TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN; | |
2938 | TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH); | |
2939 | tcp_connect_queue_skb(sk, data); | |
2940 | fo->copied = data->len; | |
2941 | ||
2942 | if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) { | |
67da22d2 | 2943 | tp->syn_data = (fo->copied > 0); |
783237e8 YC |
2944 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE); |
2945 | goto done; | |
2946 | } | |
2947 | syn_data = NULL; | |
2948 | ||
2949 | fallback: | |
2950 | /* Send a regular SYN with Fast Open cookie request option */ | |
2951 | if (fo->cookie.len > 0) | |
2952 | fo->cookie.len = 0; | |
2953 | err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); | |
2954 | if (err) | |
2955 | tp->syn_fastopen = 0; | |
2956 | kfree_skb(syn_data); | |
2957 | done: | |
2958 | fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ | |
2959 | return err; | |
2960 | } | |
2961 | ||
67edfef7 | 2962 | /* Build a SYN and send it off. */ |
1da177e4 LT |
2963 | int tcp_connect(struct sock *sk) |
2964 | { | |
2965 | struct tcp_sock *tp = tcp_sk(sk); | |
2966 | struct sk_buff *buff; | |
ee586811 | 2967 | int err; |
1da177e4 LT |
2968 | |
2969 | tcp_connect_init(sk); | |
2970 | ||
d179cd12 | 2971 | buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation); |
1da177e4 LT |
2972 | if (unlikely(buff == NULL)) |
2973 | return -ENOBUFS; | |
2974 | ||
2975 | /* Reserve space for headers. */ | |
2976 | skb_reserve(buff, MAX_TCP_HEADER); | |
2977 | ||
a3433f35 | 2978 | tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); |
783237e8 YC |
2979 | tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp; |
2980 | tcp_connect_queue_skb(sk, buff); | |
e870a8ef | 2981 | TCP_ECN_send_syn(sk, buff); |
1da177e4 | 2982 | |
783237e8 YC |
2983 | /* Send off SYN; include data in Fast Open. */ |
2984 | err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : | |
2985 | tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); | |
ee586811 EP |
2986 | if (err == -ECONNREFUSED) |
2987 | return err; | |
bd37a088 WY |
2988 | |
2989 | /* We change tp->snd_nxt after the tcp_transmit_skb() call | |
2990 | * in order to make this packet get counted in tcpOutSegs. | |
2991 | */ | |
2992 | tp->snd_nxt = tp->write_seq; | |
2993 | tp->pushed_seq = tp->write_seq; | |
81cc8a75 | 2994 | TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); |
1da177e4 LT |
2995 | |
2996 | /* Timer for repeating the SYN until an answer. */ | |
3f421baa ACM |
2997 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
2998 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
1da177e4 LT |
2999 | return 0; |
3000 | } | |
4bc2f18b | 3001 | EXPORT_SYMBOL(tcp_connect); |
1da177e4 LT |
3002 | |
3003 | /* Send out a delayed ack, the caller does the policy checking | |
3004 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() | |
3005 | * for details. | |
3006 | */ | |
3007 | void tcp_send_delayed_ack(struct sock *sk) | |
3008 | { | |
463c84b9 ACM |
3009 | struct inet_connection_sock *icsk = inet_csk(sk); |
3010 | int ato = icsk->icsk_ack.ato; | |
1da177e4 LT |
3011 | unsigned long timeout; |
3012 | ||
3013 | if (ato > TCP_DELACK_MIN) { | |
463c84b9 | 3014 | const struct tcp_sock *tp = tcp_sk(sk); |
056834d9 | 3015 | int max_ato = HZ / 2; |
1da177e4 | 3016 | |
056834d9 IJ |
3017 | if (icsk->icsk_ack.pingpong || |
3018 | (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) | |
1da177e4 LT |
3019 | max_ato = TCP_DELACK_MAX; |
3020 | ||
3021 | /* Slow path, intersegment interval is "high". */ | |
3022 | ||
3023 | /* If some rtt estimate is known, use it to bound delayed ack. | |
463c84b9 | 3024 | * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements |
1da177e4 LT |
3025 | * directly. |
3026 | */ | |
3027 | if (tp->srtt) { | |
056834d9 | 3028 | int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN); |
1da177e4 LT |
3029 | |
3030 | if (rtt < max_ato) | |
3031 | max_ato = rtt; | |
3032 | } | |
3033 | ||
3034 | ato = min(ato, max_ato); | |
3035 | } | |
3036 | ||
3037 | /* Stay within the limit we were given */ | |
3038 | timeout = jiffies + ato; | |
3039 | ||
3040 | /* Use new timeout only if there wasn't a older one earlier. */ | |
463c84b9 | 3041 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { |
1da177e4 LT |
3042 | /* If delack timer was blocked or is about to expire, |
3043 | * send ACK now. | |
3044 | */ | |
463c84b9 ACM |
3045 | if (icsk->icsk_ack.blocked || |
3046 | time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { | |
1da177e4 LT |
3047 | tcp_send_ack(sk); |
3048 | return; | |
3049 | } | |
3050 | ||
463c84b9 ACM |
3051 | if (!time_before(timeout, icsk->icsk_ack.timeout)) |
3052 | timeout = icsk->icsk_ack.timeout; | |
1da177e4 | 3053 | } |
463c84b9 ACM |
3054 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; |
3055 | icsk->icsk_ack.timeout = timeout; | |
3056 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); | |
1da177e4 LT |
3057 | } |
3058 | ||
3059 | /* This routine sends an ack and also updates the window. */ | |
3060 | void tcp_send_ack(struct sock *sk) | |
3061 | { | |
058dc334 | 3062 | struct sk_buff *buff; |
1da177e4 | 3063 | |
058dc334 IJ |
3064 | /* If we have been reset, we may not send again. */ |
3065 | if (sk->sk_state == TCP_CLOSE) | |
3066 | return; | |
1da177e4 | 3067 | |
058dc334 IJ |
3068 | /* We are not putting this on the write queue, so |
3069 | * tcp_transmit_skb() will set the ownership to this | |
3070 | * sock. | |
3071 | */ | |
99a1dec7 | 3072 | buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); |
058dc334 IJ |
3073 | if (buff == NULL) { |
3074 | inet_csk_schedule_ack(sk); | |
3075 | inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; | |
3076 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, | |
3077 | TCP_DELACK_MAX, TCP_RTO_MAX); | |
3078 | return; | |
1da177e4 | 3079 | } |
058dc334 IJ |
3080 | |
3081 | /* Reserve space for headers and prepare control bits. */ | |
3082 | skb_reserve(buff, MAX_TCP_HEADER); | |
a3433f35 | 3083 | tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); |
058dc334 IJ |
3084 | |
3085 | /* Send it off, this clears delayed acks for us. */ | |
058dc334 | 3086 | TCP_SKB_CB(buff)->when = tcp_time_stamp; |
99a1dec7 | 3087 | tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC)); |
1da177e4 LT |
3088 | } |
3089 | ||
3090 | /* This routine sends a packet with an out of date sequence | |
3091 | * number. It assumes the other end will try to ack it. | |
3092 | * | |
3093 | * Question: what should we make while urgent mode? | |
3094 | * 4.4BSD forces sending single byte of data. We cannot send | |
3095 | * out of window data, because we have SND.NXT==SND.MAX... | |
3096 | * | |
3097 | * Current solution: to send TWO zero-length segments in urgent mode: | |
3098 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is | |
3099 | * out-of-date with SND.UNA-1 to probe window. | |
3100 | */ | |
3101 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent) | |
3102 | { | |
3103 | struct tcp_sock *tp = tcp_sk(sk); | |
3104 | struct sk_buff *skb; | |
3105 | ||
3106 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ | |
99a1dec7 | 3107 | skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC)); |
e905a9ed | 3108 | if (skb == NULL) |
1da177e4 LT |
3109 | return -1; |
3110 | ||
3111 | /* Reserve space for headers and set control bits. */ | |
3112 | skb_reserve(skb, MAX_TCP_HEADER); | |
1da177e4 LT |
3113 | /* Use a previous sequence. This should cause the other |
3114 | * end to send an ack. Don't queue or clone SKB, just | |
3115 | * send it. | |
3116 | */ | |
a3433f35 | 3117 | tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); |
1da177e4 | 3118 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
dfb4b9dc | 3119 | return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC); |
1da177e4 LT |
3120 | } |
3121 | ||
ee995283 PE |
3122 | void tcp_send_window_probe(struct sock *sk) |
3123 | { | |
3124 | if (sk->sk_state == TCP_ESTABLISHED) { | |
3125 | tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; | |
c0e88ff0 | 3126 | tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq; |
ee995283 PE |
3127 | tcp_xmit_probe_skb(sk, 0); |
3128 | } | |
3129 | } | |
3130 | ||
67edfef7 | 3131 | /* Initiate keepalive or window probe from timer. */ |
1da177e4 LT |
3132 | int tcp_write_wakeup(struct sock *sk) |
3133 | { | |
058dc334 IJ |
3134 | struct tcp_sock *tp = tcp_sk(sk); |
3135 | struct sk_buff *skb; | |
1da177e4 | 3136 | |
058dc334 IJ |
3137 | if (sk->sk_state == TCP_CLOSE) |
3138 | return -1; | |
3139 | ||
3140 | if ((skb = tcp_send_head(sk)) != NULL && | |
3141 | before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { | |
3142 | int err; | |
0c54b85f | 3143 | unsigned int mss = tcp_current_mss(sk); |
058dc334 IJ |
3144 | unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
3145 | ||
3146 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) | |
3147 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; | |
3148 | ||
3149 | /* We are probing the opening of a window | |
3150 | * but the window size is != 0 | |
3151 | * must have been a result SWS avoidance ( sender ) | |
3152 | */ | |
3153 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || | |
3154 | skb->len > mss) { | |
3155 | seg_size = min(seg_size, mss); | |
4de075e0 | 3156 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
058dc334 IJ |
3157 | if (tcp_fragment(sk, skb, seg_size, mss)) |
3158 | return -1; | |
3159 | } else if (!tcp_skb_pcount(skb)) | |
3160 | tcp_set_skb_tso_segs(sk, skb, mss); | |
3161 | ||
4de075e0 | 3162 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
058dc334 IJ |
3163 | TCP_SKB_CB(skb)->when = tcp_time_stamp; |
3164 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); | |
3165 | if (!err) | |
3166 | tcp_event_new_data_sent(sk, skb); | |
3167 | return err; | |
3168 | } else { | |
33f5f57e | 3169 | if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) |
058dc334 IJ |
3170 | tcp_xmit_probe_skb(sk, 1); |
3171 | return tcp_xmit_probe_skb(sk, 0); | |
1da177e4 | 3172 | } |
1da177e4 LT |
3173 | } |
3174 | ||
3175 | /* A window probe timeout has occurred. If window is not closed send | |
3176 | * a partial packet else a zero probe. | |
3177 | */ | |
3178 | void tcp_send_probe0(struct sock *sk) | |
3179 | { | |
463c84b9 | 3180 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
3181 | struct tcp_sock *tp = tcp_sk(sk); |
3182 | int err; | |
3183 | ||
3184 | err = tcp_write_wakeup(sk); | |
3185 | ||
fe067e8a | 3186 | if (tp->packets_out || !tcp_send_head(sk)) { |
1da177e4 | 3187 | /* Cancel probe timer, if it is not required. */ |
6687e988 | 3188 | icsk->icsk_probes_out = 0; |
463c84b9 | 3189 | icsk->icsk_backoff = 0; |
1da177e4 LT |
3190 | return; |
3191 | } | |
3192 | ||
3193 | if (err <= 0) { | |
463c84b9 ACM |
3194 | if (icsk->icsk_backoff < sysctl_tcp_retries2) |
3195 | icsk->icsk_backoff++; | |
6687e988 | 3196 | icsk->icsk_probes_out++; |
e905a9ed | 3197 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
3f421baa ACM |
3198 | min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX), |
3199 | TCP_RTO_MAX); | |
1da177e4 LT |
3200 | } else { |
3201 | /* If packet was not sent due to local congestion, | |
6687e988 | 3202 | * do not backoff and do not remember icsk_probes_out. |
1da177e4 LT |
3203 | * Let local senders to fight for local resources. |
3204 | * | |
3205 | * Use accumulated backoff yet. | |
3206 | */ | |
6687e988 ACM |
3207 | if (!icsk->icsk_probes_out) |
3208 | icsk->icsk_probes_out = 1; | |
e905a9ed | 3209 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
463c84b9 | 3210 | min(icsk->icsk_rto << icsk->icsk_backoff, |
3f421baa ACM |
3211 | TCP_RESOURCE_PROBE_INTERVAL), |
3212 | TCP_RTO_MAX); | |
1da177e4 LT |
3213 | } |
3214 | } |