]>
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 | |
c39c4c6a WL |
53 | /* Default TSQ limit of four TSO segments */ |
54 | int sysctl_tcp_limit_output_bytes __read_mostly = 262144; | |
46d3ceab | 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 | |
35089bb2 | 62 | /* By default, RFC2861 behavior. */ |
ab32ea5d | 63 | int sysctl_tcp_slow_start_after_idle __read_mostly = 1; |
35089bb2 | 64 | |
46d3ceab ED |
65 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
66 | int push_one, gfp_t gfp); | |
519855c5 | 67 | |
67edfef7 | 68 | /* Account for new data that has been sent to the network. */ |
cf533ea5 | 69 | static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb) |
1da177e4 | 70 | { |
6ba8a3b1 | 71 | struct inet_connection_sock *icsk = inet_csk(sk); |
9e412ba7 | 72 | struct tcp_sock *tp = tcp_sk(sk); |
66f5fe62 | 73 | unsigned int prior_packets = tp->packets_out; |
9e412ba7 | 74 | |
fe067e8a | 75 | tcp_advance_send_head(sk, skb); |
1da177e4 | 76 | tp->snd_nxt = TCP_SKB_CB(skb)->end_seq; |
8512430e | 77 | |
66f5fe62 | 78 | tp->packets_out += tcp_skb_pcount(skb); |
6ba8a3b1 | 79 | if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS || |
6a5dc9e5 | 80 | icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) { |
750ea2ba | 81 | tcp_rearm_rto(sk); |
6a5dc9e5 | 82 | } |
f19c29e3 | 83 | |
f7324acd DM |
84 | NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, |
85 | tcp_skb_pcount(skb)); | |
1da177e4 LT |
86 | } |
87 | ||
88 | /* SND.NXT, if window was not shrunk. | |
89 | * If window has been shrunk, what should we make? It is not clear at all. | |
90 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( | |
91 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already | |
92 | * invalid. OK, let's make this for now: | |
93 | */ | |
cf533ea5 | 94 | static inline __u32 tcp_acceptable_seq(const struct sock *sk) |
1da177e4 | 95 | { |
cf533ea5 | 96 | const struct tcp_sock *tp = tcp_sk(sk); |
9e412ba7 | 97 | |
90840def | 98 | if (!before(tcp_wnd_end(tp), tp->snd_nxt)) |
1da177e4 LT |
99 | return tp->snd_nxt; |
100 | else | |
90840def | 101 | return tcp_wnd_end(tp); |
1da177e4 LT |
102 | } |
103 | ||
104 | /* Calculate mss to advertise in SYN segment. | |
105 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: | |
106 | * | |
107 | * 1. It is independent of path mtu. | |
108 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. | |
109 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of | |
110 | * attached devices, because some buggy hosts are confused by | |
111 | * large MSS. | |
112 | * 4. We do not make 3, we advertise MSS, calculated from first | |
113 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. | |
114 | * This may be overridden via information stored in routing table. | |
115 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, | |
116 | * probably even Jumbo". | |
117 | */ | |
118 | static __u16 tcp_advertise_mss(struct sock *sk) | |
119 | { | |
120 | struct tcp_sock *tp = tcp_sk(sk); | |
cf533ea5 | 121 | const struct dst_entry *dst = __sk_dst_get(sk); |
1da177e4 LT |
122 | int mss = tp->advmss; |
123 | ||
0dbaee3b DM |
124 | if (dst) { |
125 | unsigned int metric = dst_metric_advmss(dst); | |
126 | ||
127 | if (metric < mss) { | |
128 | mss = metric; | |
129 | tp->advmss = mss; | |
130 | } | |
1da177e4 LT |
131 | } |
132 | ||
133 | return (__u16)mss; | |
134 | } | |
135 | ||
136 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". | |
6f021c62 ED |
137 | * This is the first part of cwnd validation mechanism. |
138 | */ | |
139 | void tcp_cwnd_restart(struct sock *sk, s32 delta) | |
1da177e4 | 140 | { |
463c84b9 | 141 | struct tcp_sock *tp = tcp_sk(sk); |
6f021c62 | 142 | u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); |
1da177e4 LT |
143 | u32 cwnd = tp->snd_cwnd; |
144 | ||
6687e988 | 145 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); |
1da177e4 | 146 | |
6687e988 | 147 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
1da177e4 LT |
148 | restart_cwnd = min(restart_cwnd, cwnd); |
149 | ||
463c84b9 | 150 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) |
1da177e4 LT |
151 | cwnd >>= 1; |
152 | tp->snd_cwnd = max(cwnd, restart_cwnd); | |
153 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
154 | tp->snd_cwnd_used = 0; | |
155 | } | |
156 | ||
67edfef7 | 157 | /* Congestion state accounting after a packet has been sent. */ |
40efc6fa | 158 | static void tcp_event_data_sent(struct tcp_sock *tp, |
cf533ea5 | 159 | struct sock *sk) |
1da177e4 | 160 | { |
463c84b9 ACM |
161 | struct inet_connection_sock *icsk = inet_csk(sk); |
162 | const u32 now = tcp_time_stamp; | |
1da177e4 | 163 | |
05c5a46d NC |
164 | if (tcp_packets_in_flight(tp) == 0) |
165 | tcp_ca_event(sk, CA_EVENT_TX_START); | |
166 | ||
1da177e4 LT |
167 | tp->lsndtime = now; |
168 | ||
169 | /* If it is a reply for ato after last received | |
170 | * packet, enter pingpong mode. | |
171 | */ | |
2251ae46 JM |
172 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) |
173 | icsk->icsk_ack.pingpong = 1; | |
1da177e4 LT |
174 | } |
175 | ||
67edfef7 | 176 | /* Account for an ACK we sent. */ |
40efc6fa | 177 | static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) |
1da177e4 | 178 | { |
463c84b9 ACM |
179 | tcp_dec_quickack_mode(sk, pkts); |
180 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); | |
1da177e4 LT |
181 | } |
182 | ||
85f16525 YC |
183 | |
184 | u32 tcp_default_init_rwnd(u32 mss) | |
185 | { | |
186 | /* Initial receive window should be twice of TCP_INIT_CWND to | |
9ef71e0c | 187 | * enable proper sending of new unsent data during fast recovery |
85f16525 YC |
188 | * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a |
189 | * limit when mss is larger than 1460. | |
190 | */ | |
191 | u32 init_rwnd = TCP_INIT_CWND * 2; | |
192 | ||
193 | if (mss > 1460) | |
194 | init_rwnd = max((1460 * init_rwnd) / mss, 2U); | |
195 | return init_rwnd; | |
196 | } | |
197 | ||
1da177e4 LT |
198 | /* Determine a window scaling and initial window to offer. |
199 | * Based on the assumption that the given amount of space | |
200 | * will be offered. Store the results in the tp structure. | |
201 | * NOTE: for smooth operation initial space offering should | |
202 | * be a multiple of mss if possible. We assume here that mss >= 1. | |
203 | * This MUST be enforced by all callers. | |
204 | */ | |
205 | void tcp_select_initial_window(int __space, __u32 mss, | |
206 | __u32 *rcv_wnd, __u32 *window_clamp, | |
31d12926 | 207 | int wscale_ok, __u8 *rcv_wscale, |
208 | __u32 init_rcv_wnd) | |
1da177e4 LT |
209 | { |
210 | unsigned int space = (__space < 0 ? 0 : __space); | |
211 | ||
212 | /* If no clamp set the clamp to the max possible scaled window */ | |
213 | if (*window_clamp == 0) | |
214 | (*window_clamp) = (65535 << 14); | |
215 | space = min(*window_clamp, space); | |
216 | ||
217 | /* Quantize space offering to a multiple of mss if possible. */ | |
218 | if (space > mss) | |
219 | space = (space / mss) * mss; | |
220 | ||
221 | /* NOTE: offering an initial window larger than 32767 | |
15d99e02 RJ |
222 | * will break some buggy TCP stacks. If the admin tells us |
223 | * it is likely we could be speaking with such a buggy stack | |
224 | * we will truncate our initial window offering to 32K-1 | |
225 | * unless the remote has sent us a window scaling option, | |
226 | * which we interpret as a sign the remote TCP is not | |
227 | * misinterpreting the window field as a signed quantity. | |
1da177e4 | 228 | */ |
15d99e02 RJ |
229 | if (sysctl_tcp_workaround_signed_windows) |
230 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); | |
231 | else | |
232 | (*rcv_wnd) = space; | |
233 | ||
1da177e4 LT |
234 | (*rcv_wscale) = 0; |
235 | if (wscale_ok) { | |
236 | /* Set window scaling on max possible window | |
e905a9ed | 237 | * See RFC1323 for an explanation of the limit to 14 |
1da177e4 | 238 | */ |
f626300a SHY |
239 | space = max_t(u32, space, sysctl_tcp_rmem[2]); |
240 | space = max_t(u32, space, sysctl_rmem_max); | |
316c1592 | 241 | space = min_t(u32, space, *window_clamp); |
1da177e4 LT |
242 | while (space > 65535 && (*rcv_wscale) < 14) { |
243 | space >>= 1; | |
244 | (*rcv_wscale)++; | |
245 | } | |
246 | } | |
247 | ||
056834d9 | 248 | if (mss > (1 << *rcv_wscale)) { |
85f16525 YC |
249 | if (!init_rcv_wnd) /* Use default unless specified otherwise */ |
250 | init_rcv_wnd = tcp_default_init_rwnd(mss); | |
251 | *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); | |
1da177e4 LT |
252 | } |
253 | ||
254 | /* Set the clamp no higher than max representable value */ | |
255 | (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); | |
256 | } | |
4bc2f18b | 257 | EXPORT_SYMBOL(tcp_select_initial_window); |
1da177e4 LT |
258 | |
259 | /* Chose a new window to advertise, update state in tcp_sock for the | |
260 | * socket, and return result with RFC1323 scaling applied. The return | |
261 | * value can be stuffed directly into th->window for an outgoing | |
262 | * frame. | |
263 | */ | |
40efc6fa | 264 | static u16 tcp_select_window(struct sock *sk) |
1da177e4 LT |
265 | { |
266 | struct tcp_sock *tp = tcp_sk(sk); | |
8e165e20 | 267 | u32 old_win = tp->rcv_wnd; |
1da177e4 LT |
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 | */ | |
8e165e20 FW |
280 | if (new_win == 0) |
281 | NET_INC_STATS(sock_net(sk), | |
282 | LINUX_MIB_TCPWANTZEROWINDOWADV); | |
607bfbf2 | 283 | new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); |
1da177e4 LT |
284 | } |
285 | tp->rcv_wnd = new_win; | |
286 | tp->rcv_wup = tp->rcv_nxt; | |
287 | ||
288 | /* Make sure we do not exceed the maximum possible | |
289 | * scaled window. | |
290 | */ | |
15d99e02 | 291 | if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) |
1da177e4 LT |
292 | new_win = min(new_win, MAX_TCP_WINDOW); |
293 | else | |
294 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); | |
295 | ||
296 | /* RFC1323 scaling applied */ | |
297 | new_win >>= tp->rx_opt.rcv_wscale; | |
298 | ||
299 | /* If we advertise zero window, disable fast path. */ | |
8e165e20 | 300 | if (new_win == 0) { |
1da177e4 | 301 | tp->pred_flags = 0; |
8e165e20 FW |
302 | if (old_win) |
303 | NET_INC_STATS(sock_net(sk), | |
304 | LINUX_MIB_TCPTOZEROWINDOWADV); | |
305 | } else if (old_win == 0) { | |
306 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); | |
307 | } | |
1da177e4 LT |
308 | |
309 | return new_win; | |
310 | } | |
311 | ||
67edfef7 | 312 | /* Packet ECN state for a SYN-ACK */ |
735d3831 | 313 | static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) |
bdf1ee5d | 314 | { |
30e502a3 DB |
315 | const struct tcp_sock *tp = tcp_sk(sk); |
316 | ||
4de075e0 | 317 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; |
056834d9 | 318 | if (!(tp->ecn_flags & TCP_ECN_OK)) |
4de075e0 | 319 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; |
30e502a3 DB |
320 | else if (tcp_ca_needs_ecn(sk)) |
321 | INET_ECN_xmit(sk); | |
bdf1ee5d IJ |
322 | } |
323 | ||
67edfef7 | 324 | /* Packet ECN state for a SYN. */ |
735d3831 | 325 | static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) |
bdf1ee5d IJ |
326 | { |
327 | struct tcp_sock *tp = tcp_sk(sk); | |
f7b3bec6 FW |
328 | bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || |
329 | tcp_ca_needs_ecn(sk); | |
330 | ||
331 | if (!use_ecn) { | |
332 | const struct dst_entry *dst = __sk_dst_get(sk); | |
333 | ||
334 | if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) | |
335 | use_ecn = true; | |
336 | } | |
bdf1ee5d IJ |
337 | |
338 | tp->ecn_flags = 0; | |
f7b3bec6 FW |
339 | |
340 | if (use_ecn) { | |
4de075e0 | 341 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; |
bdf1ee5d | 342 | tp->ecn_flags = TCP_ECN_OK; |
30e502a3 DB |
343 | if (tcp_ca_needs_ecn(sk)) |
344 | INET_ECN_xmit(sk); | |
bdf1ee5d IJ |
345 | } |
346 | } | |
347 | ||
49213555 DB |
348 | static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) |
349 | { | |
350 | if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) | |
351 | /* tp->ecn_flags are cleared at a later point in time when | |
352 | * SYN ACK is ultimatively being received. | |
353 | */ | |
354 | TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); | |
355 | } | |
356 | ||
735d3831 | 357 | static void |
6ac705b1 | 358 | tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) |
bdf1ee5d | 359 | { |
6ac705b1 | 360 | if (inet_rsk(req)->ecn_ok) |
bdf1ee5d IJ |
361 | th->ece = 1; |
362 | } | |
363 | ||
67edfef7 AK |
364 | /* Set up ECN state for a packet on a ESTABLISHED socket that is about to |
365 | * be sent. | |
366 | */ | |
735d3831 | 367 | static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, |
ea1627c2 | 368 | struct tcphdr *th, int tcp_header_len) |
bdf1ee5d IJ |
369 | { |
370 | struct tcp_sock *tp = tcp_sk(sk); | |
371 | ||
372 | if (tp->ecn_flags & TCP_ECN_OK) { | |
373 | /* Not-retransmitted data segment: set ECT and inject CWR. */ | |
374 | if (skb->len != tcp_header_len && | |
375 | !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { | |
376 | INET_ECN_xmit(sk); | |
056834d9 | 377 | if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { |
bdf1ee5d | 378 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; |
ea1627c2 | 379 | th->cwr = 1; |
bdf1ee5d IJ |
380 | skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; |
381 | } | |
30e502a3 | 382 | } else if (!tcp_ca_needs_ecn(sk)) { |
bdf1ee5d IJ |
383 | /* ACK or retransmitted segment: clear ECT|CE */ |
384 | INET_ECN_dontxmit(sk); | |
385 | } | |
386 | if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) | |
ea1627c2 | 387 | th->ece = 1; |
bdf1ee5d IJ |
388 | } |
389 | } | |
390 | ||
e870a8ef IJ |
391 | /* Constructs common control bits of non-data skb. If SYN/FIN is present, |
392 | * auto increment end seqno. | |
393 | */ | |
394 | static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) | |
395 | { | |
2e8e18ef | 396 | skb->ip_summed = CHECKSUM_PARTIAL; |
e870a8ef IJ |
397 | skb->csum = 0; |
398 | ||
4de075e0 | 399 | TCP_SKB_CB(skb)->tcp_flags = flags; |
e870a8ef IJ |
400 | TCP_SKB_CB(skb)->sacked = 0; |
401 | ||
cd7d8498 | 402 | tcp_skb_pcount_set(skb, 1); |
e870a8ef IJ |
403 | |
404 | TCP_SKB_CB(skb)->seq = seq; | |
a3433f35 | 405 | if (flags & (TCPHDR_SYN | TCPHDR_FIN)) |
e870a8ef IJ |
406 | seq++; |
407 | TCP_SKB_CB(skb)->end_seq = seq; | |
408 | } | |
409 | ||
a2a385d6 | 410 | static inline bool tcp_urg_mode(const struct tcp_sock *tp) |
33f5f57e IJ |
411 | { |
412 | return tp->snd_una != tp->snd_up; | |
413 | } | |
414 | ||
33ad798c AL |
415 | #define OPTION_SACK_ADVERTISE (1 << 0) |
416 | #define OPTION_TS (1 << 1) | |
417 | #define OPTION_MD5 (1 << 2) | |
89e95a61 | 418 | #define OPTION_WSCALE (1 << 3) |
2100c8d2 | 419 | #define OPTION_FAST_OPEN_COOKIE (1 << 8) |
33ad798c AL |
420 | |
421 | struct tcp_out_options { | |
2100c8d2 YC |
422 | u16 options; /* bit field of OPTION_* */ |
423 | u16 mss; /* 0 to disable */ | |
33ad798c AL |
424 | u8 ws; /* window scale, 0 to disable */ |
425 | u8 num_sack_blocks; /* number of SACK blocks to include */ | |
bd0388ae | 426 | u8 hash_size; /* bytes in hash_location */ |
bd0388ae | 427 | __u8 *hash_location; /* temporary pointer, overloaded */ |
2100c8d2 YC |
428 | __u32 tsval, tsecr; /* need to include OPTION_TS */ |
429 | struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ | |
33ad798c AL |
430 | }; |
431 | ||
67edfef7 AK |
432 | /* Write previously computed TCP options to the packet. |
433 | * | |
434 | * Beware: Something in the Internet is very sensitive to the ordering of | |
fd6149d3 IJ |
435 | * TCP options, we learned this through the hard way, so be careful here. |
436 | * Luckily we can at least blame others for their non-compliance but from | |
8e3bff96 | 437 | * inter-operability perspective it seems that we're somewhat stuck with |
fd6149d3 IJ |
438 | * the ordering which we have been using if we want to keep working with |
439 | * those broken things (not that it currently hurts anybody as there isn't | |
440 | * particular reason why the ordering would need to be changed). | |
441 | * | |
442 | * At least SACK_PERM as the first option is known to lead to a disaster | |
443 | * (but it may well be that other scenarios fail similarly). | |
444 | */ | |
33ad798c | 445 | static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, |
bd0388ae WAS |
446 | struct tcp_out_options *opts) |
447 | { | |
2100c8d2 | 448 | u16 options = opts->options; /* mungable copy */ |
bd0388ae | 449 | |
bd0388ae | 450 | if (unlikely(OPTION_MD5 & options)) { |
1a2c6181 CP |
451 | *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | |
452 | (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); | |
bd0388ae WAS |
453 | /* overload cookie hash location */ |
454 | opts->hash_location = (__u8 *)ptr; | |
33ad798c | 455 | ptr += 4; |
40efc6fa | 456 | } |
33ad798c | 457 | |
fd6149d3 IJ |
458 | if (unlikely(opts->mss)) { |
459 | *ptr++ = htonl((TCPOPT_MSS << 24) | | |
460 | (TCPOLEN_MSS << 16) | | |
461 | opts->mss); | |
462 | } | |
463 | ||
bd0388ae WAS |
464 | if (likely(OPTION_TS & options)) { |
465 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { | |
33ad798c AL |
466 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
467 | (TCPOLEN_SACK_PERM << 16) | | |
468 | (TCPOPT_TIMESTAMP << 8) | | |
469 | TCPOLEN_TIMESTAMP); | |
bd0388ae | 470 | options &= ~OPTION_SACK_ADVERTISE; |
33ad798c AL |
471 | } else { |
472 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
473 | (TCPOPT_NOP << 16) | | |
474 | (TCPOPT_TIMESTAMP << 8) | | |
475 | TCPOLEN_TIMESTAMP); | |
476 | } | |
477 | *ptr++ = htonl(opts->tsval); | |
478 | *ptr++ = htonl(opts->tsecr); | |
479 | } | |
480 | ||
bd0388ae | 481 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { |
33ad798c AL |
482 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
483 | (TCPOPT_NOP << 16) | | |
484 | (TCPOPT_SACK_PERM << 8) | | |
485 | TCPOLEN_SACK_PERM); | |
486 | } | |
487 | ||
bd0388ae | 488 | if (unlikely(OPTION_WSCALE & options)) { |
33ad798c AL |
489 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
490 | (TCPOPT_WINDOW << 16) | | |
491 | (TCPOLEN_WINDOW << 8) | | |
492 | opts->ws); | |
493 | } | |
494 | ||
495 | if (unlikely(opts->num_sack_blocks)) { | |
496 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? | |
497 | tp->duplicate_sack : tp->selective_acks; | |
40efc6fa SH |
498 | int this_sack; |
499 | ||
500 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
501 | (TCPOPT_NOP << 16) | | |
502 | (TCPOPT_SACK << 8) | | |
33ad798c | 503 | (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * |
40efc6fa | 504 | TCPOLEN_SACK_PERBLOCK))); |
2de979bd | 505 | |
33ad798c AL |
506 | for (this_sack = 0; this_sack < opts->num_sack_blocks; |
507 | ++this_sack) { | |
40efc6fa SH |
508 | *ptr++ = htonl(sp[this_sack].start_seq); |
509 | *ptr++ = htonl(sp[this_sack].end_seq); | |
510 | } | |
2de979bd | 511 | |
5861f8e5 | 512 | tp->rx_opt.dsack = 0; |
40efc6fa | 513 | } |
2100c8d2 YC |
514 | |
515 | if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { | |
516 | struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; | |
7f9b838b DL |
517 | u8 *p = (u8 *)ptr; |
518 | u32 len; /* Fast Open option length */ | |
519 | ||
520 | if (foc->exp) { | |
521 | len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; | |
522 | *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | | |
523 | TCPOPT_FASTOPEN_MAGIC); | |
524 | p += TCPOLEN_EXP_FASTOPEN_BASE; | |
525 | } else { | |
526 | len = TCPOLEN_FASTOPEN_BASE + foc->len; | |
527 | *p++ = TCPOPT_FASTOPEN; | |
528 | *p++ = len; | |
529 | } | |
2100c8d2 | 530 | |
7f9b838b DL |
531 | memcpy(p, foc->val, foc->len); |
532 | if ((len & 3) == 2) { | |
533 | p[foc->len] = TCPOPT_NOP; | |
534 | p[foc->len + 1] = TCPOPT_NOP; | |
2100c8d2 | 535 | } |
7f9b838b | 536 | ptr += (len + 3) >> 2; |
2100c8d2 | 537 | } |
33ad798c AL |
538 | } |
539 | ||
67edfef7 AK |
540 | /* Compute TCP options for SYN packets. This is not the final |
541 | * network wire format yet. | |
542 | */ | |
95c96174 | 543 | static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, |
33ad798c | 544 | struct tcp_out_options *opts, |
cf533ea5 ED |
545 | struct tcp_md5sig_key **md5) |
546 | { | |
33ad798c | 547 | struct tcp_sock *tp = tcp_sk(sk); |
95c96174 | 548 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
783237e8 | 549 | struct tcp_fastopen_request *fastopen = tp->fastopen_req; |
33ad798c | 550 | |
cfb6eeb4 | 551 | #ifdef CONFIG_TCP_MD5SIG |
33ad798c AL |
552 | *md5 = tp->af_specific->md5_lookup(sk, sk); |
553 | if (*md5) { | |
554 | opts->options |= OPTION_MD5; | |
bd0388ae | 555 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
cfb6eeb4 | 556 | } |
33ad798c AL |
557 | #else |
558 | *md5 = NULL; | |
cfb6eeb4 | 559 | #endif |
33ad798c AL |
560 | |
561 | /* We always get an MSS option. The option bytes which will be seen in | |
562 | * normal data packets should timestamps be used, must be in the MSS | |
563 | * advertised. But we subtract them from tp->mss_cache so that | |
564 | * calculations in tcp_sendmsg are simpler etc. So account for this | |
565 | * fact here if necessary. If we don't do this correctly, as a | |
566 | * receiver we won't recognize data packets as being full sized when we | |
567 | * should, and thus we won't abide by the delayed ACK rules correctly. | |
568 | * SACKs don't matter, we never delay an ACK when we have any of those | |
569 | * going out. */ | |
570 | opts->mss = tcp_advertise_mss(sk); | |
bd0388ae | 571 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c | 572 | |
51456b29 | 573 | if (likely(sysctl_tcp_timestamps && !*md5)) { |
33ad798c | 574 | opts->options |= OPTION_TS; |
7faee5c0 | 575 | opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; |
33ad798c | 576 | opts->tsecr = tp->rx_opt.ts_recent; |
bd0388ae | 577 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c | 578 | } |
bb5b7c11 | 579 | if (likely(sysctl_tcp_window_scaling)) { |
33ad798c | 580 | opts->ws = tp->rx_opt.rcv_wscale; |
89e95a61 | 581 | opts->options |= OPTION_WSCALE; |
bd0388ae | 582 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 583 | } |
bb5b7c11 | 584 | if (likely(sysctl_tcp_sack)) { |
33ad798c | 585 | opts->options |= OPTION_SACK_ADVERTISE; |
b32d1310 | 586 | if (unlikely(!(OPTION_TS & opts->options))) |
bd0388ae | 587 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c AL |
588 | } |
589 | ||
783237e8 | 590 | if (fastopen && fastopen->cookie.len >= 0) { |
2646c831 DL |
591 | u32 need = fastopen->cookie.len; |
592 | ||
593 | need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : | |
594 | TCPOLEN_FASTOPEN_BASE; | |
783237e8 YC |
595 | need = (need + 3) & ~3U; /* Align to 32 bits */ |
596 | if (remaining >= need) { | |
597 | opts->options |= OPTION_FAST_OPEN_COOKIE; | |
598 | opts->fastopen_cookie = &fastopen->cookie; | |
599 | remaining -= need; | |
600 | tp->syn_fastopen = 1; | |
2646c831 | 601 | tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; |
783237e8 YC |
602 | } |
603 | } | |
bd0388ae | 604 | |
bd0388ae | 605 | return MAX_TCP_OPTION_SPACE - remaining; |
40efc6fa SH |
606 | } |
607 | ||
67edfef7 | 608 | /* Set up TCP options for SYN-ACKs. */ |
37bfbdda ED |
609 | static unsigned int tcp_synack_options(struct request_sock *req, |
610 | unsigned int mss, struct sk_buff *skb, | |
611 | struct tcp_out_options *opts, | |
612 | const struct tcp_md5sig_key *md5, | |
613 | struct tcp_fastopen_cookie *foc) | |
4957faad | 614 | { |
33ad798c | 615 | struct inet_request_sock *ireq = inet_rsk(req); |
95c96174 | 616 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
33ad798c | 617 | |
cfb6eeb4 | 618 | #ifdef CONFIG_TCP_MD5SIG |
80f03e27 | 619 | if (md5) { |
33ad798c | 620 | opts->options |= OPTION_MD5; |
4957faad WAS |
621 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
622 | ||
623 | /* We can't fit any SACK blocks in a packet with MD5 + TS | |
624 | * options. There was discussion about disabling SACK | |
625 | * rather than TS in order to fit in better with old, | |
626 | * buggy kernels, but that was deemed to be unnecessary. | |
627 | */ | |
de213e5e | 628 | ireq->tstamp_ok &= !ireq->sack_ok; |
cfb6eeb4 YH |
629 | } |
630 | #endif | |
33ad798c | 631 | |
4957faad | 632 | /* We always send an MSS option. */ |
33ad798c | 633 | opts->mss = mss; |
4957faad | 634 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c AL |
635 | |
636 | if (likely(ireq->wscale_ok)) { | |
637 | opts->ws = ireq->rcv_wscale; | |
89e95a61 | 638 | opts->options |= OPTION_WSCALE; |
4957faad | 639 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 640 | } |
de213e5e | 641 | if (likely(ireq->tstamp_ok)) { |
33ad798c | 642 | opts->options |= OPTION_TS; |
7faee5c0 | 643 | opts->tsval = tcp_skb_timestamp(skb); |
33ad798c | 644 | opts->tsecr = req->ts_recent; |
4957faad | 645 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c AL |
646 | } |
647 | if (likely(ireq->sack_ok)) { | |
648 | opts->options |= OPTION_SACK_ADVERTISE; | |
de213e5e | 649 | if (unlikely(!ireq->tstamp_ok)) |
4957faad | 650 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c | 651 | } |
7f9b838b DL |
652 | if (foc != NULL && foc->len >= 0) { |
653 | u32 need = foc->len; | |
654 | ||
655 | need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : | |
656 | TCPOLEN_FASTOPEN_BASE; | |
8336886f JC |
657 | need = (need + 3) & ~3U; /* Align to 32 bits */ |
658 | if (remaining >= need) { | |
659 | opts->options |= OPTION_FAST_OPEN_COOKIE; | |
660 | opts->fastopen_cookie = foc; | |
661 | remaining -= need; | |
662 | } | |
663 | } | |
1a2c6181 | 664 | |
4957faad | 665 | return MAX_TCP_OPTION_SPACE - remaining; |
33ad798c AL |
666 | } |
667 | ||
67edfef7 AK |
668 | /* Compute TCP options for ESTABLISHED sockets. This is not the |
669 | * final wire format yet. | |
670 | */ | |
95c96174 | 671 | static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, |
33ad798c | 672 | struct tcp_out_options *opts, |
cf533ea5 ED |
673 | struct tcp_md5sig_key **md5) |
674 | { | |
33ad798c | 675 | struct tcp_sock *tp = tcp_sk(sk); |
95c96174 | 676 | unsigned int size = 0; |
cabeccbd | 677 | unsigned int eff_sacks; |
33ad798c | 678 | |
5843ef42 AK |
679 | opts->options = 0; |
680 | ||
33ad798c AL |
681 | #ifdef CONFIG_TCP_MD5SIG |
682 | *md5 = tp->af_specific->md5_lookup(sk, sk); | |
683 | if (unlikely(*md5)) { | |
684 | opts->options |= OPTION_MD5; | |
685 | size += TCPOLEN_MD5SIG_ALIGNED; | |
686 | } | |
687 | #else | |
688 | *md5 = NULL; | |
689 | #endif | |
690 | ||
691 | if (likely(tp->rx_opt.tstamp_ok)) { | |
692 | opts->options |= OPTION_TS; | |
7faee5c0 | 693 | opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; |
33ad798c AL |
694 | opts->tsecr = tp->rx_opt.ts_recent; |
695 | size += TCPOLEN_TSTAMP_ALIGNED; | |
696 | } | |
697 | ||
cabeccbd IJ |
698 | eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; |
699 | if (unlikely(eff_sacks)) { | |
95c96174 | 700 | const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; |
33ad798c | 701 | opts->num_sack_blocks = |
95c96174 | 702 | min_t(unsigned int, eff_sacks, |
33ad798c AL |
703 | (remaining - TCPOLEN_SACK_BASE_ALIGNED) / |
704 | TCPOLEN_SACK_PERBLOCK); | |
705 | size += TCPOLEN_SACK_BASE_ALIGNED + | |
706 | opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; | |
707 | } | |
708 | ||
709 | return size; | |
40efc6fa | 710 | } |
1da177e4 | 711 | |
46d3ceab ED |
712 | |
713 | /* TCP SMALL QUEUES (TSQ) | |
714 | * | |
715 | * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) | |
716 | * to reduce RTT and bufferbloat. | |
717 | * We do this using a special skb destructor (tcp_wfree). | |
718 | * | |
719 | * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb | |
720 | * needs to be reallocated in a driver. | |
8e3bff96 | 721 | * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc |
46d3ceab ED |
722 | * |
723 | * Since transmit from skb destructor is forbidden, we use a tasklet | |
724 | * to process all sockets that eventually need to send more skbs. | |
725 | * We use one tasklet per cpu, with its own queue of sockets. | |
726 | */ | |
727 | struct tsq_tasklet { | |
728 | struct tasklet_struct tasklet; | |
729 | struct list_head head; /* queue of tcp sockets */ | |
730 | }; | |
731 | static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); | |
732 | ||
6f458dfb ED |
733 | static void tcp_tsq_handler(struct sock *sk) |
734 | { | |
735 | if ((1 << sk->sk_state) & | |
736 | (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | | |
737 | TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) | |
bf06200e JO |
738 | tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle, |
739 | 0, GFP_ATOMIC); | |
6f458dfb | 740 | } |
46d3ceab | 741 | /* |
8e3bff96 | 742 | * One tasklet per cpu tries to send more skbs. |
46d3ceab | 743 | * We run in tasklet context but need to disable irqs when |
8e3bff96 | 744 | * transferring tsq->head because tcp_wfree() might |
46d3ceab ED |
745 | * interrupt us (non NAPI drivers) |
746 | */ | |
747 | static void tcp_tasklet_func(unsigned long data) | |
748 | { | |
749 | struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; | |
750 | LIST_HEAD(list); | |
751 | unsigned long flags; | |
752 | struct list_head *q, *n; | |
753 | struct tcp_sock *tp; | |
754 | struct sock *sk; | |
755 | ||
756 | local_irq_save(flags); | |
757 | list_splice_init(&tsq->head, &list); | |
758 | local_irq_restore(flags); | |
759 | ||
760 | list_for_each_safe(q, n, &list) { | |
761 | tp = list_entry(q, struct tcp_sock, tsq_node); | |
762 | list_del(&tp->tsq_node); | |
763 | ||
764 | sk = (struct sock *)tp; | |
765 | bh_lock_sock(sk); | |
766 | ||
767 | if (!sock_owned_by_user(sk)) { | |
6f458dfb | 768 | tcp_tsq_handler(sk); |
46d3ceab ED |
769 | } else { |
770 | /* defer the work to tcp_release_cb() */ | |
6f458dfb | 771 | set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags); |
46d3ceab ED |
772 | } |
773 | bh_unlock_sock(sk); | |
774 | ||
775 | clear_bit(TSQ_QUEUED, &tp->tsq_flags); | |
776 | sk_free(sk); | |
777 | } | |
778 | } | |
779 | ||
6f458dfb ED |
780 | #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \ |
781 | (1UL << TCP_WRITE_TIMER_DEFERRED) | \ | |
563d34d0 ED |
782 | (1UL << TCP_DELACK_TIMER_DEFERRED) | \ |
783 | (1UL << TCP_MTU_REDUCED_DEFERRED)) | |
46d3ceab ED |
784 | /** |
785 | * tcp_release_cb - tcp release_sock() callback | |
786 | * @sk: socket | |
787 | * | |
788 | * called from release_sock() to perform protocol dependent | |
789 | * actions before socket release. | |
790 | */ | |
791 | void tcp_release_cb(struct sock *sk) | |
792 | { | |
793 | struct tcp_sock *tp = tcp_sk(sk); | |
6f458dfb | 794 | unsigned long flags, nflags; |
46d3ceab | 795 | |
6f458dfb ED |
796 | /* perform an atomic operation only if at least one flag is set */ |
797 | do { | |
798 | flags = tp->tsq_flags; | |
799 | if (!(flags & TCP_DEFERRED_ALL)) | |
800 | return; | |
801 | nflags = flags & ~TCP_DEFERRED_ALL; | |
802 | } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags); | |
803 | ||
804 | if (flags & (1UL << TCP_TSQ_DEFERRED)) | |
805 | tcp_tsq_handler(sk); | |
806 | ||
c3f9b018 ED |
807 | /* Here begins the tricky part : |
808 | * We are called from release_sock() with : | |
809 | * 1) BH disabled | |
810 | * 2) sk_lock.slock spinlock held | |
811 | * 3) socket owned by us (sk->sk_lock.owned == 1) | |
812 | * | |
813 | * But following code is meant to be called from BH handlers, | |
814 | * so we should keep BH disabled, but early release socket ownership | |
815 | */ | |
816 | sock_release_ownership(sk); | |
817 | ||
144d56e9 | 818 | if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) { |
6f458dfb | 819 | tcp_write_timer_handler(sk); |
144d56e9 ED |
820 | __sock_put(sk); |
821 | } | |
822 | if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) { | |
6f458dfb | 823 | tcp_delack_timer_handler(sk); |
144d56e9 ED |
824 | __sock_put(sk); |
825 | } | |
826 | if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) { | |
4fab9071 | 827 | inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); |
144d56e9 ED |
828 | __sock_put(sk); |
829 | } | |
46d3ceab ED |
830 | } |
831 | EXPORT_SYMBOL(tcp_release_cb); | |
832 | ||
833 | void __init tcp_tasklet_init(void) | |
834 | { | |
835 | int i; | |
836 | ||
837 | for_each_possible_cpu(i) { | |
838 | struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); | |
839 | ||
840 | INIT_LIST_HEAD(&tsq->head); | |
841 | tasklet_init(&tsq->tasklet, | |
842 | tcp_tasklet_func, | |
843 | (unsigned long)tsq); | |
844 | } | |
845 | } | |
846 | ||
847 | /* | |
848 | * Write buffer destructor automatically called from kfree_skb. | |
8e3bff96 | 849 | * We can't xmit new skbs from this context, as we might already |
46d3ceab ED |
850 | * hold qdisc lock. |
851 | */ | |
d6a4a104 | 852 | void tcp_wfree(struct sk_buff *skb) |
46d3ceab ED |
853 | { |
854 | struct sock *sk = skb->sk; | |
855 | struct tcp_sock *tp = tcp_sk(sk); | |
9b462d02 ED |
856 | int wmem; |
857 | ||
858 | /* Keep one reference on sk_wmem_alloc. | |
859 | * Will be released by sk_free() from here or tcp_tasklet_func() | |
860 | */ | |
861 | wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc); | |
862 | ||
863 | /* If this softirq is serviced by ksoftirqd, we are likely under stress. | |
864 | * Wait until our queues (qdisc + devices) are drained. | |
865 | * This gives : | |
866 | * - less callbacks to tcp_write_xmit(), reducing stress (batches) | |
867 | * - chance for incoming ACK (processed by another cpu maybe) | |
868 | * to migrate this flow (skb->ooo_okay will be eventually set) | |
869 | */ | |
870 | if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) | |
871 | goto out; | |
46d3ceab ED |
872 | |
873 | if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) && | |
874 | !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) { | |
875 | unsigned long flags; | |
876 | struct tsq_tasklet *tsq; | |
877 | ||
46d3ceab ED |
878 | /* queue this socket to tasklet queue */ |
879 | local_irq_save(flags); | |
903ceff7 | 880 | tsq = this_cpu_ptr(&tsq_tasklet); |
46d3ceab ED |
881 | list_add(&tp->tsq_node, &tsq->head); |
882 | tasklet_schedule(&tsq->tasklet); | |
883 | local_irq_restore(flags); | |
9b462d02 | 884 | return; |
46d3ceab | 885 | } |
9b462d02 ED |
886 | out: |
887 | sk_free(sk); | |
46d3ceab ED |
888 | } |
889 | ||
1da177e4 LT |
890 | /* This routine actually transmits TCP packets queued in by |
891 | * tcp_do_sendmsg(). This is used by both the initial | |
892 | * transmission and possible later retransmissions. | |
893 | * All SKB's seen here are completely headerless. It is our | |
894 | * job to build the TCP header, and pass the packet down to | |
895 | * IP so it can do the same plus pass the packet off to the | |
896 | * device. | |
897 | * | |
898 | * We are working here with either a clone of the original | |
899 | * SKB, or a fresh unique copy made by the retransmit engine. | |
900 | */ | |
056834d9 IJ |
901 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, |
902 | gfp_t gfp_mask) | |
1da177e4 | 903 | { |
dfb4b9dc DM |
904 | const struct inet_connection_sock *icsk = inet_csk(sk); |
905 | struct inet_sock *inet; | |
906 | struct tcp_sock *tp; | |
907 | struct tcp_skb_cb *tcb; | |
33ad798c | 908 | struct tcp_out_options opts; |
95c96174 | 909 | unsigned int tcp_options_size, tcp_header_size; |
cfb6eeb4 | 910 | struct tcp_md5sig_key *md5; |
dfb4b9dc | 911 | struct tcphdr *th; |
dfb4b9dc DM |
912 | int err; |
913 | ||
914 | BUG_ON(!skb || !tcp_skb_pcount(skb)); | |
6f094b9e | 915 | tp = tcp_sk(sk); |
dfb4b9dc | 916 | |
ccdbb6e9 | 917 | if (clone_it) { |
740b0f18 | 918 | skb_mstamp_get(&skb->skb_mstamp); |
6f094b9e LB |
919 | TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq |
920 | - tp->snd_una; | |
b9f64820 | 921 | tcp_rate_skb_sent(sk, skb); |
ccdbb6e9 | 922 | |
dfb4b9dc DM |
923 | if (unlikely(skb_cloned(skb))) |
924 | skb = pskb_copy(skb, gfp_mask); | |
925 | else | |
926 | skb = skb_clone(skb, gfp_mask); | |
927 | if (unlikely(!skb)) | |
928 | return -ENOBUFS; | |
929 | } | |
1da177e4 | 930 | |
dfb4b9dc | 931 | inet = inet_sk(sk); |
dfb4b9dc | 932 | tcb = TCP_SKB_CB(skb); |
33ad798c | 933 | memset(&opts, 0, sizeof(opts)); |
1da177e4 | 934 | |
4de075e0 | 935 | if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) |
33ad798c AL |
936 | tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); |
937 | else | |
938 | tcp_options_size = tcp_established_options(sk, skb, &opts, | |
939 | &md5); | |
940 | tcp_header_size = tcp_options_size + sizeof(struct tcphdr); | |
e905a9ed | 941 | |
547669d4 | 942 | /* if no packet is in qdisc/device queue, then allow XPS to select |
b2532eb9 ED |
943 | * another queue. We can be called from tcp_tsq_handler() |
944 | * which holds one reference to sk_wmem_alloc. | |
945 | * | |
946 | * TODO: Ideally, in-flight pure ACK packets should not matter here. | |
947 | * One way to get this would be to set skb->truesize = 2 on them. | |
547669d4 | 948 | */ |
b2532eb9 | 949 | skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); |
dfb4b9dc | 950 | |
aa8223c7 ACM |
951 | skb_push(skb, tcp_header_size); |
952 | skb_reset_transport_header(skb); | |
46d3ceab ED |
953 | |
954 | skb_orphan(skb); | |
955 | skb->sk = sk; | |
1d2077ac | 956 | skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; |
b73c3d0e | 957 | skb_set_hash_from_sk(skb, sk); |
46d3ceab | 958 | atomic_add(skb->truesize, &sk->sk_wmem_alloc); |
dfb4b9dc DM |
959 | |
960 | /* Build TCP header and checksum it. */ | |
ea1627c2 | 961 | th = (struct tcphdr *)skb->data; |
c720c7e8 ED |
962 | th->source = inet->inet_sport; |
963 | th->dest = inet->inet_dport; | |
dfb4b9dc DM |
964 | th->seq = htonl(tcb->seq); |
965 | th->ack_seq = htonl(tp->rcv_nxt); | |
df7a3b07 | 966 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
4de075e0 | 967 | tcb->tcp_flags); |
dfb4b9dc | 968 | |
dfb4b9dc DM |
969 | th->check = 0; |
970 | th->urg_ptr = 0; | |
1da177e4 | 971 | |
33f5f57e | 972 | /* The urg_mode check is necessary during a below snd_una win probe */ |
7691367d HX |
973 | if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { |
974 | if (before(tp->snd_up, tcb->seq + 0x10000)) { | |
975 | th->urg_ptr = htons(tp->snd_up - tcb->seq); | |
976 | th->urg = 1; | |
977 | } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { | |
0eae88f3 | 978 | th->urg_ptr = htons(0xFFFF); |
7691367d HX |
979 | th->urg = 1; |
980 | } | |
dfb4b9dc | 981 | } |
1da177e4 | 982 | |
bd0388ae | 983 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
51466a75 | 984 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
ea1627c2 ED |
985 | if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { |
986 | th->window = htons(tcp_select_window(sk)); | |
987 | tcp_ecn_send(sk, skb, th, tcp_header_size); | |
988 | } else { | |
989 | /* RFC1323: The window in SYN & SYN/ACK segments | |
990 | * is never scaled. | |
991 | */ | |
992 | th->window = htons(min(tp->rcv_wnd, 65535U)); | |
993 | } | |
cfb6eeb4 YH |
994 | #ifdef CONFIG_TCP_MD5SIG |
995 | /* Calculate the MD5 hash, as we have all we need now */ | |
996 | if (md5) { | |
a465419b | 997 | sk_nocaps_add(sk, NETIF_F_GSO_MASK); |
bd0388ae | 998 | tp->af_specific->calc_md5_hash(opts.hash_location, |
39f8e58e | 999 | md5, sk, skb); |
cfb6eeb4 YH |
1000 | } |
1001 | #endif | |
1002 | ||
bb296246 | 1003 | icsk->icsk_af_ops->send_check(sk, skb); |
1da177e4 | 1004 | |
4de075e0 | 1005 | if (likely(tcb->tcp_flags & TCPHDR_ACK)) |
dfb4b9dc | 1006 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); |
1da177e4 | 1007 | |
a44d6eac | 1008 | if (skb->len != tcp_header_size) { |
cf533ea5 | 1009 | tcp_event_data_sent(tp, sk); |
a44d6eac MKL |
1010 | tp->data_segs_out += tcp_skb_pcount(skb); |
1011 | } | |
1da177e4 | 1012 | |
bd37a088 | 1013 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
aa2ea058 TH |
1014 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, |
1015 | tcp_skb_pcount(skb)); | |
1da177e4 | 1016 | |
2efd055c | 1017 | tp->segs_out += tcp_skb_pcount(skb); |
f69ad292 | 1018 | /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ |
cd7d8498 | 1019 | skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); |
f69ad292 | 1020 | skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); |
cd7d8498 | 1021 | |
7faee5c0 ED |
1022 | /* Our usage of tstamp should remain private */ |
1023 | skb->tstamp.tv64 = 0; | |
971f10ec ED |
1024 | |
1025 | /* Cleanup our debris for IP stacks */ | |
1026 | memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), | |
1027 | sizeof(struct inet6_skb_parm))); | |
1028 | ||
b0270e91 | 1029 | err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); |
7faee5c0 | 1030 | |
83de47cd | 1031 | if (likely(err <= 0)) |
dfb4b9dc DM |
1032 | return err; |
1033 | ||
5ee2c941 | 1034 | tcp_enter_cwr(sk); |
dfb4b9dc | 1035 | |
b9df3cb8 | 1036 | return net_xmit_eval(err); |
1da177e4 LT |
1037 | } |
1038 | ||
67edfef7 | 1039 | /* This routine just queues the buffer for sending. |
1da177e4 LT |
1040 | * |
1041 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, | |
1042 | * otherwise socket can stall. | |
1043 | */ | |
1044 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) | |
1045 | { | |
1046 | struct tcp_sock *tp = tcp_sk(sk); | |
1047 | ||
1048 | /* Advance write_seq and place onto the write_queue. */ | |
1049 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; | |
f4a775d1 | 1050 | __skb_header_release(skb); |
fe067e8a | 1051 | tcp_add_write_queue_tail(sk, skb); |
3ab224be HA |
1052 | sk->sk_wmem_queued += skb->truesize; |
1053 | sk_mem_charge(sk, skb->truesize); | |
1da177e4 LT |
1054 | } |
1055 | ||
67edfef7 | 1056 | /* Initialize TSO segments for a packet. */ |
5bbb432c | 1057 | static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) |
f6302d1d | 1058 | { |
8f26fb1c | 1059 | if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { |
f6302d1d DM |
1060 | /* Avoid the costly divide in the normal |
1061 | * non-TSO case. | |
1062 | */ | |
cd7d8498 | 1063 | tcp_skb_pcount_set(skb, 1); |
f69ad292 | 1064 | TCP_SKB_CB(skb)->tcp_gso_size = 0; |
f6302d1d | 1065 | } else { |
cd7d8498 | 1066 | tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); |
f69ad292 | 1067 | TCP_SKB_CB(skb)->tcp_gso_size = mss_now; |
1da177e4 LT |
1068 | } |
1069 | } | |
1070 | ||
91fed7a1 | 1071 | /* When a modification to fackets out becomes necessary, we need to check |
68f8353b | 1072 | * skb is counted to fackets_out or not. |
91fed7a1 | 1073 | */ |
cf533ea5 | 1074 | static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, |
91fed7a1 IJ |
1075 | int decr) |
1076 | { | |
a47e5a98 IJ |
1077 | struct tcp_sock *tp = tcp_sk(sk); |
1078 | ||
dc86967b | 1079 | if (!tp->sacked_out || tcp_is_reno(tp)) |
91fed7a1 IJ |
1080 | return; |
1081 | ||
6859d494 | 1082 | if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) |
91fed7a1 | 1083 | tp->fackets_out -= decr; |
91fed7a1 IJ |
1084 | } |
1085 | ||
797108d1 IJ |
1086 | /* Pcount in the middle of the write queue got changed, we need to do various |
1087 | * tweaks to fix counters | |
1088 | */ | |
cf533ea5 | 1089 | static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) |
797108d1 IJ |
1090 | { |
1091 | struct tcp_sock *tp = tcp_sk(sk); | |
1092 | ||
1093 | tp->packets_out -= decr; | |
1094 | ||
1095 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
1096 | tp->sacked_out -= decr; | |
1097 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
1098 | tp->retrans_out -= decr; | |
1099 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) | |
1100 | tp->lost_out -= decr; | |
1101 | ||
1102 | /* Reno case is special. Sigh... */ | |
1103 | if (tcp_is_reno(tp) && decr > 0) | |
1104 | tp->sacked_out -= min_t(u32, tp->sacked_out, decr); | |
1105 | ||
1106 | tcp_adjust_fackets_out(sk, skb, decr); | |
1107 | ||
1108 | if (tp->lost_skb_hint && | |
1109 | before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && | |
52cf3cc8 | 1110 | (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) |
797108d1 IJ |
1111 | tp->lost_cnt_hint -= decr; |
1112 | ||
1113 | tcp_verify_left_out(tp); | |
1114 | } | |
1115 | ||
0a2cf20c SHY |
1116 | static bool tcp_has_tx_tstamp(const struct sk_buff *skb) |
1117 | { | |
1118 | return TCP_SKB_CB(skb)->txstamp_ack || | |
1119 | (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); | |
1120 | } | |
1121 | ||
490cc7d0 WB |
1122 | static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) |
1123 | { | |
1124 | struct skb_shared_info *shinfo = skb_shinfo(skb); | |
1125 | ||
0a2cf20c | 1126 | if (unlikely(tcp_has_tx_tstamp(skb)) && |
490cc7d0 WB |
1127 | !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { |
1128 | struct skb_shared_info *shinfo2 = skb_shinfo(skb2); | |
1129 | u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; | |
1130 | ||
1131 | shinfo->tx_flags &= ~tsflags; | |
1132 | shinfo2->tx_flags |= tsflags; | |
1133 | swap(shinfo->tskey, shinfo2->tskey); | |
b51e13fa MKL |
1134 | TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; |
1135 | TCP_SKB_CB(skb)->txstamp_ack = 0; | |
490cc7d0 WB |
1136 | } |
1137 | } | |
1138 | ||
a166140e MKL |
1139 | static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) |
1140 | { | |
1141 | TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; | |
1142 | TCP_SKB_CB(skb)->eor = 0; | |
1143 | } | |
1144 | ||
1da177e4 LT |
1145 | /* Function to create two new TCP segments. Shrinks the given segment |
1146 | * to the specified size and appends a new segment with the rest of the | |
e905a9ed | 1147 | * packet to the list. This won't be called frequently, I hope. |
1da177e4 LT |
1148 | * Remember, these are still headerless SKBs at this point. |
1149 | */ | |
056834d9 | 1150 | int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, |
6cc55e09 | 1151 | unsigned int mss_now, gfp_t gfp) |
1da177e4 LT |
1152 | { |
1153 | struct tcp_sock *tp = tcp_sk(sk); | |
1154 | struct sk_buff *buff; | |
6475be16 | 1155 | int nsize, old_factor; |
b60b49ea | 1156 | int nlen; |
9ce01461 | 1157 | u8 flags; |
1da177e4 | 1158 | |
2fceec13 IJ |
1159 | if (WARN_ON(len > skb->len)) |
1160 | return -EINVAL; | |
6a438bbe | 1161 | |
1da177e4 LT |
1162 | nsize = skb_headlen(skb) - len; |
1163 | if (nsize < 0) | |
1164 | nsize = 0; | |
1165 | ||
6cc55e09 | 1166 | if (skb_unclone(skb, gfp)) |
1da177e4 LT |
1167 | return -ENOMEM; |
1168 | ||
1169 | /* Get a new skb... force flag on. */ | |
eb934478 | 1170 | buff = sk_stream_alloc_skb(sk, nsize, gfp, true); |
51456b29 | 1171 | if (!buff) |
1da177e4 | 1172 | return -ENOMEM; /* We'll just try again later. */ |
ef5cb973 | 1173 | |
3ab224be HA |
1174 | sk->sk_wmem_queued += buff->truesize; |
1175 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea HX |
1176 | nlen = skb->len - len - nsize; |
1177 | buff->truesize += nlen; | |
1178 | skb->truesize -= nlen; | |
1da177e4 LT |
1179 | |
1180 | /* Correct the sequence numbers. */ | |
1181 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1182 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1183 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1184 | ||
1185 | /* PSH and FIN should only be set in the second packet. */ | |
4de075e0 ED |
1186 | flags = TCP_SKB_CB(skb)->tcp_flags; |
1187 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
1188 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
e14c3caf | 1189 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
a166140e | 1190 | tcp_skb_fragment_eor(skb, buff); |
1da177e4 | 1191 | |
84fa7933 | 1192 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { |
1da177e4 | 1193 | /* Copy and checksum data tail into the new buffer. */ |
056834d9 IJ |
1194 | buff->csum = csum_partial_copy_nocheck(skb->data + len, |
1195 | skb_put(buff, nsize), | |
1da177e4 LT |
1196 | nsize, 0); |
1197 | ||
1198 | skb_trim(skb, len); | |
1199 | ||
1200 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); | |
1201 | } else { | |
84fa7933 | 1202 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
1203 | skb_split(skb, buff, len); |
1204 | } | |
1205 | ||
1206 | buff->ip_summed = skb->ip_summed; | |
1207 | ||
a61bbcf2 | 1208 | buff->tstamp = skb->tstamp; |
490cc7d0 | 1209 | tcp_fragment_tstamp(skb, buff); |
1da177e4 | 1210 | |
6475be16 DM |
1211 | old_factor = tcp_skb_pcount(skb); |
1212 | ||
1da177e4 | 1213 | /* Fix up tso_factor for both original and new SKB. */ |
5bbb432c ED |
1214 | tcp_set_skb_tso_segs(skb, mss_now); |
1215 | tcp_set_skb_tso_segs(buff, mss_now); | |
1da177e4 | 1216 | |
b9f64820 YC |
1217 | /* Update delivered info for the new segment */ |
1218 | TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; | |
1219 | ||
6475be16 DM |
1220 | /* If this packet has been sent out already, we must |
1221 | * adjust the various packet counters. | |
1222 | */ | |
cf0b450c | 1223 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
6475be16 DM |
1224 | int diff = old_factor - tcp_skb_pcount(skb) - |
1225 | tcp_skb_pcount(buff); | |
1da177e4 | 1226 | |
797108d1 IJ |
1227 | if (diff) |
1228 | tcp_adjust_pcount(sk, skb, diff); | |
1da177e4 LT |
1229 | } |
1230 | ||
1231 | /* Link BUFF into the send queue. */ | |
f4a775d1 | 1232 | __skb_header_release(buff); |
fe067e8a | 1233 | tcp_insert_write_queue_after(skb, buff, sk); |
1da177e4 LT |
1234 | |
1235 | return 0; | |
1236 | } | |
1237 | ||
1238 | /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c | |
1239 | * eventually). The difference is that pulled data not copied, but | |
1240 | * immediately discarded. | |
1241 | */ | |
f2911969 | 1242 | static void __pskb_trim_head(struct sk_buff *skb, int len) |
1da177e4 | 1243 | { |
7b7fc97a | 1244 | struct skb_shared_info *shinfo; |
1da177e4 LT |
1245 | int i, k, eat; |
1246 | ||
4fa48bf3 ED |
1247 | eat = min_t(int, len, skb_headlen(skb)); |
1248 | if (eat) { | |
1249 | __skb_pull(skb, eat); | |
1250 | len -= eat; | |
1251 | if (!len) | |
1252 | return; | |
1253 | } | |
1da177e4 LT |
1254 | eat = len; |
1255 | k = 0; | |
7b7fc97a ED |
1256 | shinfo = skb_shinfo(skb); |
1257 | for (i = 0; i < shinfo->nr_frags; i++) { | |
1258 | int size = skb_frag_size(&shinfo->frags[i]); | |
9e903e08 ED |
1259 | |
1260 | if (size <= eat) { | |
aff65da0 | 1261 | skb_frag_unref(skb, i); |
9e903e08 | 1262 | eat -= size; |
1da177e4 | 1263 | } else { |
7b7fc97a | 1264 | shinfo->frags[k] = shinfo->frags[i]; |
1da177e4 | 1265 | if (eat) { |
7b7fc97a ED |
1266 | shinfo->frags[k].page_offset += eat; |
1267 | skb_frag_size_sub(&shinfo->frags[k], eat); | |
1da177e4 LT |
1268 | eat = 0; |
1269 | } | |
1270 | k++; | |
1271 | } | |
1272 | } | |
7b7fc97a | 1273 | shinfo->nr_frags = k; |
1da177e4 | 1274 | |
27a884dc | 1275 | skb_reset_tail_pointer(skb); |
1da177e4 LT |
1276 | skb->data_len -= len; |
1277 | skb->len = skb->data_len; | |
1da177e4 LT |
1278 | } |
1279 | ||
67edfef7 | 1280 | /* Remove acked data from a packet in the transmit queue. */ |
1da177e4 LT |
1281 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) |
1282 | { | |
14bbd6a5 | 1283 | if (skb_unclone(skb, GFP_ATOMIC)) |
1da177e4 LT |
1284 | return -ENOMEM; |
1285 | ||
4fa48bf3 | 1286 | __pskb_trim_head(skb, len); |
1da177e4 LT |
1287 | |
1288 | TCP_SKB_CB(skb)->seq += len; | |
84fa7933 | 1289 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
1290 | |
1291 | skb->truesize -= len; | |
1292 | sk->sk_wmem_queued -= len; | |
3ab224be | 1293 | sk_mem_uncharge(sk, len); |
1da177e4 LT |
1294 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); |
1295 | ||
5b35e1e6 | 1296 | /* Any change of skb->len requires recalculation of tso factor. */ |
1da177e4 | 1297 | if (tcp_skb_pcount(skb) > 1) |
5bbb432c | 1298 | tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); |
1da177e4 LT |
1299 | |
1300 | return 0; | |
1301 | } | |
1302 | ||
1b63edd6 YC |
1303 | /* Calculate MSS not accounting any TCP options. */ |
1304 | static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
5d424d5a | 1305 | { |
cf533ea5 ED |
1306 | const struct tcp_sock *tp = tcp_sk(sk); |
1307 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
5d424d5a JH |
1308 | int mss_now; |
1309 | ||
1310 | /* Calculate base mss without TCP options: | |
1311 | It is MMS_S - sizeof(tcphdr) of rfc1122 | |
1312 | */ | |
1313 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); | |
1314 | ||
67469601 ED |
1315 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
1316 | if (icsk->icsk_af_ops->net_frag_header_len) { | |
1317 | const struct dst_entry *dst = __sk_dst_get(sk); | |
1318 | ||
1319 | if (dst && dst_allfrag(dst)) | |
1320 | mss_now -= icsk->icsk_af_ops->net_frag_header_len; | |
1321 | } | |
1322 | ||
5d424d5a JH |
1323 | /* Clamp it (mss_clamp does not include tcp options) */ |
1324 | if (mss_now > tp->rx_opt.mss_clamp) | |
1325 | mss_now = tp->rx_opt.mss_clamp; | |
1326 | ||
1327 | /* Now subtract optional transport overhead */ | |
1328 | mss_now -= icsk->icsk_ext_hdr_len; | |
1329 | ||
1330 | /* Then reserve room for full set of TCP options and 8 bytes of data */ | |
1331 | if (mss_now < 48) | |
1332 | mss_now = 48; | |
5d424d5a JH |
1333 | return mss_now; |
1334 | } | |
1335 | ||
1b63edd6 YC |
1336 | /* Calculate MSS. Not accounting for SACKs here. */ |
1337 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
1338 | { | |
1339 | /* Subtract TCP options size, not including SACKs */ | |
1340 | return __tcp_mtu_to_mss(sk, pmtu) - | |
1341 | (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); | |
1342 | } | |
1343 | ||
5d424d5a | 1344 | /* Inverse of above */ |
67469601 | 1345 | int tcp_mss_to_mtu(struct sock *sk, int mss) |
5d424d5a | 1346 | { |
cf533ea5 ED |
1347 | const struct tcp_sock *tp = tcp_sk(sk); |
1348 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
5d424d5a JH |
1349 | int mtu; |
1350 | ||
1351 | mtu = mss + | |
1352 | tp->tcp_header_len + | |
1353 | icsk->icsk_ext_hdr_len + | |
1354 | icsk->icsk_af_ops->net_header_len; | |
1355 | ||
67469601 ED |
1356 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
1357 | if (icsk->icsk_af_ops->net_frag_header_len) { | |
1358 | const struct dst_entry *dst = __sk_dst_get(sk); | |
1359 | ||
1360 | if (dst && dst_allfrag(dst)) | |
1361 | mtu += icsk->icsk_af_ops->net_frag_header_len; | |
1362 | } | |
5d424d5a JH |
1363 | return mtu; |
1364 | } | |
1365 | ||
67edfef7 | 1366 | /* MTU probing init per socket */ |
5d424d5a JH |
1367 | void tcp_mtup_init(struct sock *sk) |
1368 | { | |
1369 | struct tcp_sock *tp = tcp_sk(sk); | |
1370 | struct inet_connection_sock *icsk = inet_csk(sk); | |
b0f9ca53 | 1371 | struct net *net = sock_net(sk); |
5d424d5a | 1372 | |
b0f9ca53 | 1373 | icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; |
5d424d5a | 1374 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + |
e905a9ed | 1375 | icsk->icsk_af_ops->net_header_len; |
b0f9ca53 | 1376 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); |
5d424d5a | 1377 | icsk->icsk_mtup.probe_size = 0; |
05cbc0db FD |
1378 | if (icsk->icsk_mtup.enabled) |
1379 | icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; | |
5d424d5a | 1380 | } |
4bc2f18b | 1381 | EXPORT_SYMBOL(tcp_mtup_init); |
5d424d5a | 1382 | |
1da177e4 LT |
1383 | /* This function synchronize snd mss to current pmtu/exthdr set. |
1384 | ||
1385 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts | |
1386 | for TCP options, but includes only bare TCP header. | |
1387 | ||
1388 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. | |
caa20d9a | 1389 | It is minimum of user_mss and mss received with SYN. |
1da177e4 LT |
1390 | It also does not include TCP options. |
1391 | ||
d83d8461 | 1392 | inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. |
1da177e4 LT |
1393 | |
1394 | tp->mss_cache is current effective sending mss, including | |
1395 | all tcp options except for SACKs. It is evaluated, | |
1396 | taking into account current pmtu, but never exceeds | |
1397 | tp->rx_opt.mss_clamp. | |
1398 | ||
1399 | NOTE1. rfc1122 clearly states that advertised MSS | |
1400 | DOES NOT include either tcp or ip options. | |
1401 | ||
d83d8461 ACM |
1402 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
1403 | are READ ONLY outside this function. --ANK (980731) | |
1da177e4 | 1404 | */ |
1da177e4 LT |
1405 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) |
1406 | { | |
1407 | struct tcp_sock *tp = tcp_sk(sk); | |
d83d8461 | 1408 | struct inet_connection_sock *icsk = inet_csk(sk); |
5d424d5a | 1409 | int mss_now; |
1da177e4 | 1410 | |
5d424d5a JH |
1411 | if (icsk->icsk_mtup.search_high > pmtu) |
1412 | icsk->icsk_mtup.search_high = pmtu; | |
1da177e4 | 1413 | |
5d424d5a | 1414 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
409d22b4 | 1415 | mss_now = tcp_bound_to_half_wnd(tp, mss_now); |
1da177e4 LT |
1416 | |
1417 | /* And store cached results */ | |
d83d8461 | 1418 | icsk->icsk_pmtu_cookie = pmtu; |
5d424d5a JH |
1419 | if (icsk->icsk_mtup.enabled) |
1420 | mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); | |
c1b4a7e6 | 1421 | tp->mss_cache = mss_now; |
1da177e4 LT |
1422 | |
1423 | return mss_now; | |
1424 | } | |
4bc2f18b | 1425 | EXPORT_SYMBOL(tcp_sync_mss); |
1da177e4 LT |
1426 | |
1427 | /* Compute the current effective MSS, taking SACKs and IP options, | |
1428 | * and even PMTU discovery events into account. | |
1da177e4 | 1429 | */ |
0c54b85f | 1430 | unsigned int tcp_current_mss(struct sock *sk) |
1da177e4 | 1431 | { |
cf533ea5 ED |
1432 | const struct tcp_sock *tp = tcp_sk(sk); |
1433 | const struct dst_entry *dst = __sk_dst_get(sk); | |
c1b4a7e6 | 1434 | u32 mss_now; |
95c96174 | 1435 | unsigned int header_len; |
33ad798c AL |
1436 | struct tcp_out_options opts; |
1437 | struct tcp_md5sig_key *md5; | |
c1b4a7e6 DM |
1438 | |
1439 | mss_now = tp->mss_cache; | |
1440 | ||
1da177e4 LT |
1441 | if (dst) { |
1442 | u32 mtu = dst_mtu(dst); | |
d83d8461 | 1443 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
1da177e4 LT |
1444 | mss_now = tcp_sync_mss(sk, mtu); |
1445 | } | |
1446 | ||
33ad798c AL |
1447 | header_len = tcp_established_options(sk, NULL, &opts, &md5) + |
1448 | sizeof(struct tcphdr); | |
1449 | /* The mss_cache is sized based on tp->tcp_header_len, which assumes | |
1450 | * some common options. If this is an odd packet (because we have SACK | |
1451 | * blocks etc) then our calculated header_len will be different, and | |
1452 | * we have to adjust mss_now correspondingly */ | |
1453 | if (header_len != tp->tcp_header_len) { | |
1454 | int delta = (int) header_len - tp->tcp_header_len; | |
1455 | mss_now -= delta; | |
1456 | } | |
cfb6eeb4 | 1457 | |
1da177e4 LT |
1458 | return mss_now; |
1459 | } | |
1460 | ||
86fd14ad WP |
1461 | /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. |
1462 | * As additional protections, we do not touch cwnd in retransmission phases, | |
1463 | * and if application hit its sndbuf limit recently. | |
1464 | */ | |
1465 | static void tcp_cwnd_application_limited(struct sock *sk) | |
1466 | { | |
1467 | struct tcp_sock *tp = tcp_sk(sk); | |
1468 | ||
1469 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && | |
1470 | sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { | |
1471 | /* Limited by application or receiver window. */ | |
1472 | u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); | |
1473 | u32 win_used = max(tp->snd_cwnd_used, init_win); | |
1474 | if (win_used < tp->snd_cwnd) { | |
1475 | tp->snd_ssthresh = tcp_current_ssthresh(sk); | |
1476 | tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; | |
1477 | } | |
1478 | tp->snd_cwnd_used = 0; | |
1479 | } | |
1480 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1481 | } | |
1482 | ||
ca8a2263 | 1483 | static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) |
a762a980 | 1484 | { |
9e412ba7 | 1485 | struct tcp_sock *tp = tcp_sk(sk); |
a762a980 | 1486 | |
ca8a2263 NC |
1487 | /* Track the maximum number of outstanding packets in each |
1488 | * window, and remember whether we were cwnd-limited then. | |
1489 | */ | |
1490 | if (!before(tp->snd_una, tp->max_packets_seq) || | |
1491 | tp->packets_out > tp->max_packets_out) { | |
1492 | tp->max_packets_out = tp->packets_out; | |
1493 | tp->max_packets_seq = tp->snd_nxt; | |
1494 | tp->is_cwnd_limited = is_cwnd_limited; | |
1495 | } | |
e114a710 | 1496 | |
24901551 | 1497 | if (tcp_is_cwnd_limited(sk)) { |
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 | ||
d4589926 ED |
1512 | /* Minshall's variant of the Nagle send check. */ |
1513 | static bool tcp_minshall_check(const struct tcp_sock *tp) | |
1514 | { | |
1515 | return after(tp->snd_sml, tp->snd_una) && | |
1516 | !after(tp->snd_sml, tp->snd_nxt); | |
1517 | } | |
1518 | ||
1519 | /* Update snd_sml if this skb is under mss | |
1520 | * Note that a TSO packet might end with a sub-mss segment | |
1521 | * The test is really : | |
1522 | * if ((skb->len % mss) != 0) | |
1523 | * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; | |
1524 | * But we can avoid doing the divide again given we already have | |
1525 | * skb_pcount = skb->len / mss_now | |
0e3a4803 | 1526 | */ |
d4589926 ED |
1527 | static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, |
1528 | const struct sk_buff *skb) | |
1529 | { | |
1530 | if (skb->len < tcp_skb_pcount(skb) * mss_now) | |
1531 | tp->snd_sml = TCP_SKB_CB(skb)->end_seq; | |
1532 | } | |
1533 | ||
1534 | /* Return false, if packet can be sent now without violation Nagle's rules: | |
1535 | * 1. It is full sized. (provided by caller in %partial bool) | |
1536 | * 2. Or it contains FIN. (already checked by caller) | |
1537 | * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. | |
1538 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. | |
1539 | * With Minshall's modification: all sent small packets are ACKed. | |
1540 | */ | |
1541 | static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, | |
cc93fc51 | 1542 | int nonagle) |
d4589926 ED |
1543 | { |
1544 | return partial && | |
1545 | ((nonagle & TCP_NAGLE_CORK) || | |
1546 | (!nonagle && tp->packets_out && tcp_minshall_check(tp))); | |
1547 | } | |
605ad7f1 ED |
1548 | |
1549 | /* Return how many segs we'd like on a TSO packet, | |
1550 | * to send one TSO packet per ms | |
1551 | */ | |
1b3878ca NC |
1552 | u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, |
1553 | int min_tso_segs) | |
605ad7f1 ED |
1554 | { |
1555 | u32 bytes, segs; | |
1556 | ||
1557 | bytes = min(sk->sk_pacing_rate >> 10, | |
1558 | sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); | |
1559 | ||
1560 | /* Goal is to send at least one packet per ms, | |
1561 | * not one big TSO packet every 100 ms. | |
1562 | * This preserves ACK clocking and is consistent | |
1563 | * with tcp_tso_should_defer() heuristic. | |
1564 | */ | |
1b3878ca | 1565 | segs = max_t(u32, bytes / mss_now, min_tso_segs); |
605ad7f1 ED |
1566 | |
1567 | return min_t(u32, segs, sk->sk_gso_max_segs); | |
1568 | } | |
1b3878ca | 1569 | EXPORT_SYMBOL(tcp_tso_autosize); |
605ad7f1 | 1570 | |
ed6e7268 NC |
1571 | /* Return the number of segments we want in the skb we are transmitting. |
1572 | * See if congestion control module wants to decide; otherwise, autosize. | |
1573 | */ | |
1574 | static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) | |
1575 | { | |
1576 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; | |
1577 | u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0; | |
1578 | ||
1b3878ca NC |
1579 | return tso_segs ? : |
1580 | tcp_tso_autosize(sk, mss_now, sysctl_tcp_min_tso_segs); | |
ed6e7268 NC |
1581 | } |
1582 | ||
d4589926 ED |
1583 | /* Returns the portion of skb which can be sent right away */ |
1584 | static unsigned int tcp_mss_split_point(const struct sock *sk, | |
1585 | const struct sk_buff *skb, | |
1586 | unsigned int mss_now, | |
1587 | unsigned int max_segs, | |
1588 | int nonagle) | |
c1b4a7e6 | 1589 | { |
cf533ea5 | 1590 | const struct tcp_sock *tp = tcp_sk(sk); |
d4589926 | 1591 | u32 partial, needed, window, max_len; |
c1b4a7e6 | 1592 | |
90840def | 1593 | window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
1485348d | 1594 | max_len = mss_now * max_segs; |
0e3a4803 | 1595 | |
1485348d BH |
1596 | if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) |
1597 | return max_len; | |
0e3a4803 | 1598 | |
5ea3a748 IJ |
1599 | needed = min(skb->len, window); |
1600 | ||
1485348d BH |
1601 | if (max_len <= needed) |
1602 | return max_len; | |
0e3a4803 | 1603 | |
d4589926 ED |
1604 | partial = needed % mss_now; |
1605 | /* If last segment is not a full MSS, check if Nagle rules allow us | |
1606 | * to include this last segment in this skb. | |
1607 | * Otherwise, we'll split the skb at last MSS boundary | |
1608 | */ | |
cc93fc51 | 1609 | if (tcp_nagle_check(partial != 0, tp, nonagle)) |
d4589926 ED |
1610 | return needed - partial; |
1611 | ||
1612 | return needed; | |
c1b4a7e6 DM |
1613 | } |
1614 | ||
1615 | /* Can at least one segment of SKB be sent right now, according to the | |
1616 | * congestion window rules? If so, return how many segments are allowed. | |
1617 | */ | |
cf533ea5 ED |
1618 | static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, |
1619 | const struct sk_buff *skb) | |
c1b4a7e6 | 1620 | { |
d649a7a8 | 1621 | u32 in_flight, cwnd, halfcwnd; |
c1b4a7e6 DM |
1622 | |
1623 | /* Don't be strict about the congestion window for the final FIN. */ | |
4de075e0 ED |
1624 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && |
1625 | tcp_skb_pcount(skb) == 1) | |
c1b4a7e6 DM |
1626 | return 1; |
1627 | ||
1628 | in_flight = tcp_packets_in_flight(tp); | |
1629 | cwnd = tp->snd_cwnd; | |
d649a7a8 ED |
1630 | if (in_flight >= cwnd) |
1631 | return 0; | |
c1b4a7e6 | 1632 | |
d649a7a8 ED |
1633 | /* For better scheduling, ensure we have at least |
1634 | * 2 GSO packets in flight. | |
1635 | */ | |
1636 | halfcwnd = max(cwnd >> 1, 1U); | |
1637 | return min(halfcwnd, cwnd - in_flight); | |
c1b4a7e6 DM |
1638 | } |
1639 | ||
b595076a | 1640 | /* Initialize TSO state of a skb. |
67edfef7 | 1641 | * This must be invoked the first time we consider transmitting |
c1b4a7e6 DM |
1642 | * SKB onto the wire. |
1643 | */ | |
5bbb432c | 1644 | static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) |
c1b4a7e6 DM |
1645 | { |
1646 | int tso_segs = tcp_skb_pcount(skb); | |
1647 | ||
f8269a49 | 1648 | if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { |
5bbb432c | 1649 | tcp_set_skb_tso_segs(skb, mss_now); |
c1b4a7e6 DM |
1650 | tso_segs = tcp_skb_pcount(skb); |
1651 | } | |
1652 | return tso_segs; | |
1653 | } | |
1654 | ||
c1b4a7e6 | 1655 | |
a2a385d6 | 1656 | /* Return true if the Nagle test allows this packet to be |
c1b4a7e6 DM |
1657 | * sent now. |
1658 | */ | |
a2a385d6 ED |
1659 | static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, |
1660 | unsigned int cur_mss, int nonagle) | |
c1b4a7e6 DM |
1661 | { |
1662 | /* Nagle rule does not apply to frames, which sit in the middle of the | |
1663 | * write_queue (they have no chances to get new data). | |
1664 | * | |
1665 | * This is implemented in the callers, where they modify the 'nonagle' | |
1666 | * argument based upon the location of SKB in the send queue. | |
1667 | */ | |
1668 | if (nonagle & TCP_NAGLE_PUSH) | |
a2a385d6 | 1669 | return true; |
c1b4a7e6 | 1670 | |
9b44190d YC |
1671 | /* Don't use the nagle rule for urgent data (or for the final FIN). */ |
1672 | if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) | |
a2a385d6 | 1673 | return true; |
c1b4a7e6 | 1674 | |
cc93fc51 | 1675 | if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) |
a2a385d6 | 1676 | return true; |
c1b4a7e6 | 1677 | |
a2a385d6 | 1678 | return false; |
c1b4a7e6 DM |
1679 | } |
1680 | ||
1681 | /* Does at least the first segment of SKB fit into the send window? */ | |
a2a385d6 ED |
1682 | static bool tcp_snd_wnd_test(const struct tcp_sock *tp, |
1683 | const struct sk_buff *skb, | |
1684 | unsigned int cur_mss) | |
c1b4a7e6 DM |
1685 | { |
1686 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
1687 | ||
1688 | if (skb->len > cur_mss) | |
1689 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; | |
1690 | ||
90840def | 1691 | return !after(end_seq, tcp_wnd_end(tp)); |
c1b4a7e6 DM |
1692 | } |
1693 | ||
fe067e8a | 1694 | /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) |
c1b4a7e6 DM |
1695 | * should be put on the wire right now. If so, it returns the number of |
1696 | * packets allowed by the congestion window. | |
1697 | */ | |
cf533ea5 | 1698 | static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, |
c1b4a7e6 DM |
1699 | unsigned int cur_mss, int nonagle) |
1700 | { | |
cf533ea5 | 1701 | const struct tcp_sock *tp = tcp_sk(sk); |
c1b4a7e6 DM |
1702 | unsigned int cwnd_quota; |
1703 | ||
5bbb432c | 1704 | tcp_init_tso_segs(skb, cur_mss); |
c1b4a7e6 DM |
1705 | |
1706 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) | |
1707 | return 0; | |
1708 | ||
1709 | cwnd_quota = tcp_cwnd_test(tp, skb); | |
056834d9 | 1710 | if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) |
c1b4a7e6 DM |
1711 | cwnd_quota = 0; |
1712 | ||
1713 | return cwnd_quota; | |
1714 | } | |
1715 | ||
67edfef7 | 1716 | /* Test if sending is allowed right now. */ |
a2a385d6 | 1717 | bool tcp_may_send_now(struct sock *sk) |
c1b4a7e6 | 1718 | { |
cf533ea5 | 1719 | const struct tcp_sock *tp = tcp_sk(sk); |
fe067e8a | 1720 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 | 1721 | |
a02cec21 | 1722 | return skb && |
0c54b85f | 1723 | tcp_snd_test(sk, skb, tcp_current_mss(sk), |
c1b4a7e6 | 1724 | (tcp_skb_is_last(sk, skb) ? |
a02cec21 | 1725 | tp->nonagle : TCP_NAGLE_PUSH)); |
c1b4a7e6 DM |
1726 | } |
1727 | ||
1728 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet | |
1729 | * which is put after SKB on the list. It is very much like | |
1730 | * tcp_fragment() except that it may make several kinds of assumptions | |
1731 | * in order to speed up the splitting operation. In particular, we | |
1732 | * know that all the data is in scatter-gather pages, and that the | |
1733 | * packet has never been sent out before (and thus is not cloned). | |
1734 | */ | |
056834d9 | 1735 | static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, |
c4ead4c5 | 1736 | unsigned int mss_now, gfp_t gfp) |
c1b4a7e6 DM |
1737 | { |
1738 | struct sk_buff *buff; | |
1739 | int nlen = skb->len - len; | |
9ce01461 | 1740 | u8 flags; |
c1b4a7e6 DM |
1741 | |
1742 | /* All of a TSO frame must be composed of paged data. */ | |
c8ac3774 | 1743 | if (skb->len != skb->data_len) |
6cc55e09 | 1744 | return tcp_fragment(sk, skb, len, mss_now, gfp); |
c1b4a7e6 | 1745 | |
eb934478 | 1746 | buff = sk_stream_alloc_skb(sk, 0, gfp, true); |
51456b29 | 1747 | if (unlikely(!buff)) |
c1b4a7e6 DM |
1748 | return -ENOMEM; |
1749 | ||
3ab224be HA |
1750 | sk->sk_wmem_queued += buff->truesize; |
1751 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea | 1752 | buff->truesize += nlen; |
c1b4a7e6 DM |
1753 | skb->truesize -= nlen; |
1754 | ||
1755 | /* Correct the sequence numbers. */ | |
1756 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1757 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1758 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1759 | ||
1760 | /* PSH and FIN should only be set in the second packet. */ | |
4de075e0 ED |
1761 | flags = TCP_SKB_CB(skb)->tcp_flags; |
1762 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
1763 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
c1b4a7e6 DM |
1764 | |
1765 | /* This packet was never sent out yet, so no SACK bits. */ | |
1766 | TCP_SKB_CB(buff)->sacked = 0; | |
1767 | ||
a166140e MKL |
1768 | tcp_skb_fragment_eor(skb, buff); |
1769 | ||
84fa7933 | 1770 | buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; |
c1b4a7e6 | 1771 | skb_split(skb, buff, len); |
490cc7d0 | 1772 | tcp_fragment_tstamp(skb, buff); |
c1b4a7e6 DM |
1773 | |
1774 | /* Fix up tso_factor for both original and new SKB. */ | |
5bbb432c ED |
1775 | tcp_set_skb_tso_segs(skb, mss_now); |
1776 | tcp_set_skb_tso_segs(buff, mss_now); | |
c1b4a7e6 DM |
1777 | |
1778 | /* Link BUFF into the send queue. */ | |
f4a775d1 | 1779 | __skb_header_release(buff); |
fe067e8a | 1780 | tcp_insert_write_queue_after(skb, buff, sk); |
c1b4a7e6 DM |
1781 | |
1782 | return 0; | |
1783 | } | |
1784 | ||
1785 | /* Try to defer sending, if possible, in order to minimize the amount | |
1786 | * of TSO splitting we do. View it as a kind of TSO Nagle test. | |
1787 | * | |
1788 | * This algorithm is from John Heffner. | |
1789 | */ | |
ca8a2263 | 1790 | static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, |
605ad7f1 | 1791 | bool *is_cwnd_limited, u32 max_segs) |
c1b4a7e6 | 1792 | { |
6687e988 | 1793 | const struct inet_connection_sock *icsk = inet_csk(sk); |
50c8339e ED |
1794 | u32 age, send_win, cong_win, limit, in_flight; |
1795 | struct tcp_sock *tp = tcp_sk(sk); | |
1796 | struct skb_mstamp now; | |
1797 | struct sk_buff *head; | |
ad9f4f50 | 1798 | int win_divisor; |
c1b4a7e6 | 1799 | |
4de075e0 | 1800 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) |
ae8064ac | 1801 | goto send_now; |
c1b4a7e6 | 1802 | |
99d7662a | 1803 | if (icsk->icsk_ca_state >= TCP_CA_Recovery) |
ae8064ac JH |
1804 | goto send_now; |
1805 | ||
5f852eb5 ED |
1806 | /* Avoid bursty behavior by allowing defer |
1807 | * only if the last write was recent. | |
1808 | */ | |
1809 | if ((s32)(tcp_time_stamp - tp->lsndtime) > 0) | |
ae8064ac | 1810 | goto send_now; |
908a75c1 | 1811 | |
c1b4a7e6 DM |
1812 | in_flight = tcp_packets_in_flight(tp); |
1813 | ||
056834d9 | 1814 | BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); |
c1b4a7e6 | 1815 | |
90840def | 1816 | send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
c1b4a7e6 DM |
1817 | |
1818 | /* From in_flight test above, we know that cwnd > in_flight. */ | |
1819 | cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; | |
1820 | ||
1821 | limit = min(send_win, cong_win); | |
1822 | ||
ba244fe9 | 1823 | /* If a full-sized TSO skb can be sent, do it. */ |
605ad7f1 | 1824 | if (limit >= max_segs * tp->mss_cache) |
ae8064ac | 1825 | goto send_now; |
ba244fe9 | 1826 | |
62ad2761 IJ |
1827 | /* Middle in queue won't get any more data, full sendable already? */ |
1828 | if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) | |
1829 | goto send_now; | |
1830 | ||
ad9f4f50 ED |
1831 | win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); |
1832 | if (win_divisor) { | |
c1b4a7e6 DM |
1833 | u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); |
1834 | ||
1835 | /* If at least some fraction of a window is available, | |
1836 | * just use it. | |
1837 | */ | |
ad9f4f50 | 1838 | chunk /= win_divisor; |
c1b4a7e6 | 1839 | if (limit >= chunk) |
ae8064ac | 1840 | goto send_now; |
c1b4a7e6 DM |
1841 | } else { |
1842 | /* Different approach, try not to defer past a single | |
1843 | * ACK. Receiver should ACK every other full sized | |
1844 | * frame, so if we have space for more than 3 frames | |
1845 | * then send now. | |
1846 | */ | |
6b5a5c0d | 1847 | if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) |
ae8064ac | 1848 | goto send_now; |
c1b4a7e6 DM |
1849 | } |
1850 | ||
50c8339e ED |
1851 | head = tcp_write_queue_head(sk); |
1852 | skb_mstamp_get(&now); | |
1853 | age = skb_mstamp_us_delta(&now, &head->skb_mstamp); | |
1854 | /* If next ACK is likely to come too late (half srtt), do not defer */ | |
1855 | if (age < (tp->srtt_us >> 4)) | |
1856 | goto send_now; | |
1857 | ||
5f852eb5 | 1858 | /* Ok, it looks like it is advisable to defer. */ |
ae8064ac | 1859 | |
d2e1339f | 1860 | if (cong_win < send_win && cong_win <= skb->len) |
ca8a2263 NC |
1861 | *is_cwnd_limited = true; |
1862 | ||
a2a385d6 | 1863 | return true; |
ae8064ac JH |
1864 | |
1865 | send_now: | |
a2a385d6 | 1866 | return false; |
c1b4a7e6 DM |
1867 | } |
1868 | ||
05cbc0db FD |
1869 | static inline void tcp_mtu_check_reprobe(struct sock *sk) |
1870 | { | |
1871 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1872 | struct tcp_sock *tp = tcp_sk(sk); | |
1873 | struct net *net = sock_net(sk); | |
1874 | u32 interval; | |
1875 | s32 delta; | |
1876 | ||
1877 | interval = net->ipv4.sysctl_tcp_probe_interval; | |
1878 | delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp; | |
1879 | if (unlikely(delta >= interval * HZ)) { | |
1880 | int mss = tcp_current_mss(sk); | |
1881 | ||
1882 | /* Update current search range */ | |
1883 | icsk->icsk_mtup.probe_size = 0; | |
1884 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + | |
1885 | sizeof(struct tcphdr) + | |
1886 | icsk->icsk_af_ops->net_header_len; | |
1887 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); | |
1888 | ||
1889 | /* Update probe time stamp */ | |
1890 | icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; | |
1891 | } | |
1892 | } | |
1893 | ||
5d424d5a | 1894 | /* Create a new MTU probe if we are ready. |
67edfef7 AK |
1895 | * MTU probe is regularly attempting to increase the path MTU by |
1896 | * deliberately sending larger packets. This discovers routing | |
1897 | * changes resulting in larger path MTUs. | |
1898 | * | |
5d424d5a JH |
1899 | * Returns 0 if we should wait to probe (no cwnd available), |
1900 | * 1 if a probe was sent, | |
056834d9 IJ |
1901 | * -1 otherwise |
1902 | */ | |
5d424d5a JH |
1903 | static int tcp_mtu_probe(struct sock *sk) |
1904 | { | |
1905 | struct tcp_sock *tp = tcp_sk(sk); | |
1906 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1907 | struct sk_buff *skb, *nskb, *next; | |
6b58e0a5 | 1908 | struct net *net = sock_net(sk); |
5d424d5a JH |
1909 | int len; |
1910 | int probe_size; | |
91cc17c0 | 1911 | int size_needed; |
5d424d5a JH |
1912 | int copy; |
1913 | int mss_now; | |
6b58e0a5 | 1914 | int interval; |
5d424d5a JH |
1915 | |
1916 | /* Not currently probing/verifying, | |
1917 | * not in recovery, | |
1918 | * have enough cwnd, and | |
1919 | * not SACKing (the variable headers throw things off) */ | |
1920 | if (!icsk->icsk_mtup.enabled || | |
1921 | icsk->icsk_mtup.probe_size || | |
1922 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || | |
1923 | tp->snd_cwnd < 11 || | |
cabeccbd | 1924 | tp->rx_opt.num_sacks || tp->rx_opt.dsack) |
5d424d5a JH |
1925 | return -1; |
1926 | ||
6b58e0a5 FD |
1927 | /* Use binary search for probe_size between tcp_mss_base, |
1928 | * and current mss_clamp. if (search_high - search_low) | |
1929 | * smaller than a threshold, backoff from probing. | |
1930 | */ | |
0c54b85f | 1931 | mss_now = tcp_current_mss(sk); |
6b58e0a5 FD |
1932 | probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + |
1933 | icsk->icsk_mtup.search_low) >> 1); | |
91cc17c0 | 1934 | size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; |
6b58e0a5 | 1935 | interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; |
05cbc0db FD |
1936 | /* When misfortune happens, we are reprobing actively, |
1937 | * and then reprobe timer has expired. We stick with current | |
1938 | * probing process by not resetting search range to its orignal. | |
1939 | */ | |
6b58e0a5 | 1940 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || |
05cbc0db FD |
1941 | interval < net->ipv4.sysctl_tcp_probe_threshold) { |
1942 | /* Check whether enough time has elaplased for | |
1943 | * another round of probing. | |
1944 | */ | |
1945 | tcp_mtu_check_reprobe(sk); | |
5d424d5a JH |
1946 | return -1; |
1947 | } | |
1948 | ||
1949 | /* Have enough data in the send queue to probe? */ | |
7f9c33e5 | 1950 | if (tp->write_seq - tp->snd_nxt < size_needed) |
5d424d5a JH |
1951 | return -1; |
1952 | ||
91cc17c0 IJ |
1953 | if (tp->snd_wnd < size_needed) |
1954 | return -1; | |
90840def | 1955 | if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) |
91cc17c0 | 1956 | return 0; |
5d424d5a | 1957 | |
d67c58e9 IJ |
1958 | /* Do we need to wait to drain cwnd? With none in flight, don't stall */ |
1959 | if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { | |
1960 | if (!tcp_packets_in_flight(tp)) | |
5d424d5a JH |
1961 | return -1; |
1962 | else | |
1963 | return 0; | |
1964 | } | |
1965 | ||
1966 | /* We're allowed to probe. Build it now. */ | |
eb934478 | 1967 | nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); |
51456b29 | 1968 | if (!nskb) |
5d424d5a | 1969 | return -1; |
3ab224be HA |
1970 | sk->sk_wmem_queued += nskb->truesize; |
1971 | sk_mem_charge(sk, nskb->truesize); | |
5d424d5a | 1972 | |
fe067e8a | 1973 | skb = tcp_send_head(sk); |
5d424d5a JH |
1974 | |
1975 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; | |
1976 | TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; | |
4de075e0 | 1977 | TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; |
5d424d5a JH |
1978 | TCP_SKB_CB(nskb)->sacked = 0; |
1979 | nskb->csum = 0; | |
84fa7933 | 1980 | nskb->ip_summed = skb->ip_summed; |
5d424d5a | 1981 | |
50c4817e IJ |
1982 | tcp_insert_write_queue_before(nskb, skb, sk); |
1983 | ||
5d424d5a | 1984 | len = 0; |
234b6860 | 1985 | tcp_for_write_queue_from_safe(skb, next, sk) { |
5d424d5a JH |
1986 | copy = min_t(int, skb->len, probe_size - len); |
1987 | if (nskb->ip_summed) | |
1988 | skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); | |
1989 | else | |
1990 | nskb->csum = skb_copy_and_csum_bits(skb, 0, | |
056834d9 IJ |
1991 | skb_put(nskb, copy), |
1992 | copy, nskb->csum); | |
5d424d5a JH |
1993 | |
1994 | if (skb->len <= copy) { | |
1995 | /* We've eaten all the data from this skb. | |
1996 | * Throw it away. */ | |
4de075e0 | 1997 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; |
fe067e8a | 1998 | tcp_unlink_write_queue(skb, sk); |
3ab224be | 1999 | sk_wmem_free_skb(sk, skb); |
5d424d5a | 2000 | } else { |
4de075e0 | 2001 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & |
a3433f35 | 2002 | ~(TCPHDR_FIN|TCPHDR_PSH); |
5d424d5a JH |
2003 | if (!skb_shinfo(skb)->nr_frags) { |
2004 | skb_pull(skb, copy); | |
84fa7933 | 2005 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
056834d9 IJ |
2006 | skb->csum = csum_partial(skb->data, |
2007 | skb->len, 0); | |
5d424d5a JH |
2008 | } else { |
2009 | __pskb_trim_head(skb, copy); | |
5bbb432c | 2010 | tcp_set_skb_tso_segs(skb, mss_now); |
5d424d5a JH |
2011 | } |
2012 | TCP_SKB_CB(skb)->seq += copy; | |
2013 | } | |
2014 | ||
2015 | len += copy; | |
234b6860 IJ |
2016 | |
2017 | if (len >= probe_size) | |
2018 | break; | |
5d424d5a | 2019 | } |
5bbb432c | 2020 | tcp_init_tso_segs(nskb, nskb->len); |
5d424d5a JH |
2021 | |
2022 | /* We're ready to send. If this fails, the probe will | |
7faee5c0 ED |
2023 | * be resegmented into mss-sized pieces by tcp_write_xmit(). |
2024 | */ | |
5d424d5a JH |
2025 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { |
2026 | /* Decrement cwnd here because we are sending | |
056834d9 | 2027 | * effectively two packets. */ |
5d424d5a | 2028 | tp->snd_cwnd--; |
66f5fe62 | 2029 | tcp_event_new_data_sent(sk, nskb); |
5d424d5a JH |
2030 | |
2031 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); | |
0e7b1368 JH |
2032 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; |
2033 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; | |
5d424d5a JH |
2034 | |
2035 | return 1; | |
2036 | } | |
2037 | ||
2038 | return -1; | |
2039 | } | |
2040 | ||
1da177e4 LT |
2041 | /* This routine writes packets to the network. It advances the |
2042 | * send_head. This happens as incoming acks open up the remote | |
2043 | * window for us. | |
2044 | * | |
f8269a49 IJ |
2045 | * LARGESEND note: !tcp_urg_mode is overkill, only frames between |
2046 | * snd_up-64k-mss .. snd_up cannot be large. However, taking into | |
2047 | * account rare use of URG, this is not a big flaw. | |
2048 | * | |
6ba8a3b1 ND |
2049 | * Send at most one packet when push_one > 0. Temporarily ignore |
2050 | * cwnd limit to force at most one packet out when push_one == 2. | |
2051 | ||
a2a385d6 ED |
2052 | * Returns true, if no segments are in flight and we have queued segments, |
2053 | * but cannot send anything now because of SWS or another problem. | |
1da177e4 | 2054 | */ |
a2a385d6 ED |
2055 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
2056 | int push_one, gfp_t gfp) | |
1da177e4 LT |
2057 | { |
2058 | struct tcp_sock *tp = tcp_sk(sk); | |
92df7b51 | 2059 | struct sk_buff *skb; |
c1b4a7e6 DM |
2060 | unsigned int tso_segs, sent_pkts; |
2061 | int cwnd_quota; | |
5d424d5a | 2062 | int result; |
ca8a2263 | 2063 | bool is_cwnd_limited = false; |
605ad7f1 | 2064 | u32 max_segs; |
1da177e4 | 2065 | |
92df7b51 | 2066 | sent_pkts = 0; |
5d424d5a | 2067 | |
d5dd9175 IJ |
2068 | if (!push_one) { |
2069 | /* Do MTU probing. */ | |
2070 | result = tcp_mtu_probe(sk); | |
2071 | if (!result) { | |
a2a385d6 | 2072 | return false; |
d5dd9175 IJ |
2073 | } else if (result > 0) { |
2074 | sent_pkts = 1; | |
2075 | } | |
5d424d5a JH |
2076 | } |
2077 | ||
ed6e7268 | 2078 | max_segs = tcp_tso_segs(sk, mss_now); |
fe067e8a | 2079 | while ((skb = tcp_send_head(sk))) { |
c8ac3774 HX |
2080 | unsigned int limit; |
2081 | ||
5bbb432c | 2082 | tso_segs = tcp_init_tso_segs(skb, mss_now); |
c1b4a7e6 | 2083 | BUG_ON(!tso_segs); |
aa93466b | 2084 | |
9d186cac | 2085 | if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { |
7faee5c0 ED |
2086 | /* "skb_mstamp" is used as a start point for the retransmit timer */ |
2087 | skb_mstamp_get(&skb->skb_mstamp); | |
ec342325 | 2088 | goto repair; /* Skip network transmission */ |
9d186cac | 2089 | } |
ec342325 | 2090 | |
b68e9f85 | 2091 | cwnd_quota = tcp_cwnd_test(tp, skb); |
6ba8a3b1 ND |
2092 | if (!cwnd_quota) { |
2093 | if (push_one == 2) | |
2094 | /* Force out a loss probe pkt. */ | |
2095 | cwnd_quota = 1; | |
2096 | else | |
2097 | break; | |
2098 | } | |
b68e9f85 HX |
2099 | |
2100 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) | |
2101 | break; | |
2102 | ||
d6a4e26a | 2103 | if (tso_segs == 1) { |
c1b4a7e6 DM |
2104 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, |
2105 | (tcp_skb_is_last(sk, skb) ? | |
2106 | nonagle : TCP_NAGLE_PUSH)))) | |
2107 | break; | |
2108 | } else { | |
ca8a2263 | 2109 | if (!push_one && |
605ad7f1 ED |
2110 | tcp_tso_should_defer(sk, skb, &is_cwnd_limited, |
2111 | max_segs)) | |
c1b4a7e6 DM |
2112 | break; |
2113 | } | |
aa93466b | 2114 | |
605ad7f1 | 2115 | limit = mss_now; |
d6a4e26a | 2116 | if (tso_segs > 1 && !tcp_urg_mode(tp)) |
605ad7f1 ED |
2117 | limit = tcp_mss_split_point(sk, skb, mss_now, |
2118 | min_t(unsigned int, | |
2119 | cwnd_quota, | |
2120 | max_segs), | |
2121 | nonagle); | |
2122 | ||
2123 | if (skb->len > limit && | |
2124 | unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) | |
2125 | break; | |
2126 | ||
c9eeec26 ED |
2127 | /* TCP Small Queues : |
2128 | * Control number of packets in qdisc/devices to two packets / or ~1 ms. | |
2129 | * This allows for : | |
2130 | * - better RTT estimation and ACK scheduling | |
2131 | * - faster recovery | |
2132 | * - high rates | |
98e09386 ED |
2133 | * Alas, some drivers / subsystems require a fair amount |
2134 | * of queued bytes to ensure line rate. | |
2135 | * One example is wifi aggregation (802.11 AMPDU) | |
46d3ceab | 2136 | */ |
605ad7f1 ED |
2137 | limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10); |
2138 | limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes); | |
c9eeec26 ED |
2139 | |
2140 | if (atomic_read(&sk->sk_wmem_alloc) > limit) { | |
46d3ceab | 2141 | set_bit(TSQ_THROTTLED, &tp->tsq_flags); |
bf06200e JO |
2142 | /* It is possible TX completion already happened |
2143 | * before we set TSQ_THROTTLED, so we must | |
2144 | * test again the condition. | |
bf06200e | 2145 | */ |
4e857c58 | 2146 | smp_mb__after_atomic(); |
bf06200e JO |
2147 | if (atomic_read(&sk->sk_wmem_alloc) > limit) |
2148 | break; | |
46d3ceab | 2149 | } |
c9eeec26 | 2150 | |
d5dd9175 | 2151 | if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) |
92df7b51 | 2152 | break; |
1da177e4 | 2153 | |
ec342325 | 2154 | repair: |
92df7b51 DM |
2155 | /* Advance the send_head. This one is sent out. |
2156 | * This call will increment packets_out. | |
2157 | */ | |
66f5fe62 | 2158 | tcp_event_new_data_sent(sk, skb); |
1da177e4 | 2159 | |
92df7b51 | 2160 | tcp_minshall_update(tp, mss_now, skb); |
a262f0cd | 2161 | sent_pkts += tcp_skb_pcount(skb); |
d5dd9175 IJ |
2162 | |
2163 | if (push_one) | |
2164 | break; | |
92df7b51 | 2165 | } |
1da177e4 | 2166 | |
aa93466b | 2167 | if (likely(sent_pkts)) { |
684bad11 YC |
2168 | if (tcp_in_cwnd_reduction(sk)) |
2169 | tp->prr_out += sent_pkts; | |
6ba8a3b1 ND |
2170 | |
2171 | /* Send one loss probe per tail loss episode. */ | |
2172 | if (push_one != 2) | |
2173 | tcp_schedule_loss_probe(sk); | |
d2e1339f | 2174 | is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); |
ca8a2263 | 2175 | tcp_cwnd_validate(sk, is_cwnd_limited); |
a2a385d6 | 2176 | return false; |
1da177e4 | 2177 | } |
b340b264 | 2178 | return !tp->packets_out && tcp_send_head(sk); |
6ba8a3b1 ND |
2179 | } |
2180 | ||
2181 | bool tcp_schedule_loss_probe(struct sock *sk) | |
2182 | { | |
2183 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2184 | struct tcp_sock *tp = tcp_sk(sk); | |
2185 | u32 timeout, tlp_time_stamp, rto_time_stamp; | |
740b0f18 | 2186 | u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3); |
6ba8a3b1 ND |
2187 | |
2188 | if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS)) | |
2189 | return false; | |
2190 | /* No consecutive loss probes. */ | |
2191 | if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { | |
2192 | tcp_rearm_rto(sk); | |
2193 | return false; | |
2194 | } | |
2195 | /* Don't do any loss probe on a Fast Open connection before 3WHS | |
2196 | * finishes. | |
2197 | */ | |
f9b99582 | 2198 | if (tp->fastopen_rsk) |
6ba8a3b1 ND |
2199 | return false; |
2200 | ||
2201 | /* TLP is only scheduled when next timer event is RTO. */ | |
2202 | if (icsk->icsk_pending != ICSK_TIME_RETRANS) | |
2203 | return false; | |
2204 | ||
2205 | /* Schedule a loss probe in 2*RTT for SACK capable connections | |
2206 | * in Open state, that are either limited by cwnd or application. | |
2207 | */ | |
f9b99582 | 2208 | if (sysctl_tcp_early_retrans < 3 || !tp->packets_out || |
6ba8a3b1 ND |
2209 | !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open) |
2210 | return false; | |
2211 | ||
2212 | if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && | |
2213 | tcp_send_head(sk)) | |
2214 | return false; | |
2215 | ||
2216 | /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account | |
f9b99582 YC |
2217 | * for delayed ack when there's one outstanding packet. If no RTT |
2218 | * sample is available then probe after TCP_TIMEOUT_INIT. | |
6ba8a3b1 | 2219 | */ |
f9b99582 | 2220 | timeout = rtt << 1 ? : TCP_TIMEOUT_INIT; |
6ba8a3b1 ND |
2221 | if (tp->packets_out == 1) |
2222 | timeout = max_t(u32, timeout, | |
2223 | (rtt + (rtt >> 1) + TCP_DELACK_MAX)); | |
2224 | timeout = max_t(u32, timeout, msecs_to_jiffies(10)); | |
2225 | ||
2226 | /* If RTO is shorter, just schedule TLP in its place. */ | |
2227 | tlp_time_stamp = tcp_time_stamp + timeout; | |
2228 | rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; | |
2229 | if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { | |
2230 | s32 delta = rto_time_stamp - tcp_time_stamp; | |
2231 | if (delta > 0) | |
2232 | timeout = delta; | |
2233 | } | |
2234 | ||
2235 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, | |
2236 | TCP_RTO_MAX); | |
2237 | return true; | |
2238 | } | |
2239 | ||
1f3279ae ED |
2240 | /* Thanks to skb fast clones, we can detect if a prior transmit of |
2241 | * a packet is still in a qdisc or driver queue. | |
2242 | * In this case, there is very little point doing a retransmit ! | |
1f3279ae ED |
2243 | */ |
2244 | static bool skb_still_in_host_queue(const struct sock *sk, | |
2245 | const struct sk_buff *skb) | |
2246 | { | |
39bb5e62 | 2247 | if (unlikely(skb_fclone_busy(sk, skb))) { |
c10d9310 ED |
2248 | NET_INC_STATS(sock_net(sk), |
2249 | LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); | |
1f3279ae ED |
2250 | return true; |
2251 | } | |
2252 | return false; | |
2253 | } | |
2254 | ||
b340b264 | 2255 | /* When probe timeout (PTO) fires, try send a new segment if possible, else |
6ba8a3b1 ND |
2256 | * retransmit the last segment. |
2257 | */ | |
2258 | void tcp_send_loss_probe(struct sock *sk) | |
2259 | { | |
9b717a8d | 2260 | struct tcp_sock *tp = tcp_sk(sk); |
6ba8a3b1 ND |
2261 | struct sk_buff *skb; |
2262 | int pcount; | |
2263 | int mss = tcp_current_mss(sk); | |
6ba8a3b1 | 2264 | |
b340b264 YC |
2265 | skb = tcp_send_head(sk); |
2266 | if (skb) { | |
2267 | if (tcp_snd_wnd_test(tp, skb, mss)) { | |
2268 | pcount = tp->packets_out; | |
2269 | tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); | |
2270 | if (tp->packets_out > pcount) | |
2271 | goto probe_sent; | |
2272 | goto rearm_timer; | |
2273 | } | |
2274 | skb = tcp_write_queue_prev(sk, skb); | |
2275 | } else { | |
2276 | skb = tcp_write_queue_tail(sk); | |
6ba8a3b1 ND |
2277 | } |
2278 | ||
9b717a8d ND |
2279 | /* At most one outstanding TLP retransmission. */ |
2280 | if (tp->tlp_high_seq) | |
2281 | goto rearm_timer; | |
2282 | ||
6ba8a3b1 | 2283 | /* Retransmit last segment. */ |
6ba8a3b1 ND |
2284 | if (WARN_ON(!skb)) |
2285 | goto rearm_timer; | |
2286 | ||
1f3279ae ED |
2287 | if (skb_still_in_host_queue(sk, skb)) |
2288 | goto rearm_timer; | |
2289 | ||
6ba8a3b1 ND |
2290 | pcount = tcp_skb_pcount(skb); |
2291 | if (WARN_ON(!pcount)) | |
2292 | goto rearm_timer; | |
2293 | ||
2294 | if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { | |
6cc55e09 OP |
2295 | if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss, |
2296 | GFP_ATOMIC))) | |
6ba8a3b1 | 2297 | goto rearm_timer; |
b340b264 | 2298 | skb = tcp_write_queue_next(sk, skb); |
6ba8a3b1 ND |
2299 | } |
2300 | ||
2301 | if (WARN_ON(!skb || !tcp_skb_pcount(skb))) | |
2302 | goto rearm_timer; | |
2303 | ||
10d3be56 | 2304 | if (__tcp_retransmit_skb(sk, skb, 1)) |
b340b264 | 2305 | goto rearm_timer; |
6ba8a3b1 | 2306 | |
9b717a8d | 2307 | /* Record snd_nxt for loss detection. */ |
b340b264 | 2308 | tp->tlp_high_seq = tp->snd_nxt; |
9b717a8d | 2309 | |
b340b264 | 2310 | probe_sent: |
c10d9310 | 2311 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); |
b340b264 YC |
2312 | /* Reset s.t. tcp_rearm_rto will restart timer from now */ |
2313 | inet_csk(sk)->icsk_pending = 0; | |
6ba8a3b1 | 2314 | rearm_timer: |
fcd16c0a | 2315 | tcp_rearm_rto(sk); |
1da177e4 LT |
2316 | } |
2317 | ||
a762a980 DM |
2318 | /* Push out any pending frames which were held back due to |
2319 | * TCP_CORK or attempt at coalescing tiny packets. | |
2320 | * The socket must be locked by the caller. | |
2321 | */ | |
9e412ba7 IJ |
2322 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, |
2323 | int nonagle) | |
a762a980 | 2324 | { |
726e07a8 IJ |
2325 | /* If we are closed, the bytes will have to remain here. |
2326 | * In time closedown will finish, we empty the write queue and | |
2327 | * all will be happy. | |
2328 | */ | |
2329 | if (unlikely(sk->sk_state == TCP_CLOSE)) | |
2330 | return; | |
2331 | ||
99a1dec7 | 2332 | if (tcp_write_xmit(sk, cur_mss, nonagle, 0, |
7450aaf6 | 2333 | sk_gfp_mask(sk, GFP_ATOMIC))) |
726e07a8 | 2334 | tcp_check_probe_timer(sk); |
a762a980 DM |
2335 | } |
2336 | ||
c1b4a7e6 DM |
2337 | /* Send _single_ skb sitting at the send head. This function requires |
2338 | * true push pending frames to setup probe timer etc. | |
2339 | */ | |
2340 | void tcp_push_one(struct sock *sk, unsigned int mss_now) | |
2341 | { | |
fe067e8a | 2342 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 DM |
2343 | |
2344 | BUG_ON(!skb || skb->len < mss_now); | |
2345 | ||
d5dd9175 | 2346 | tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); |
c1b4a7e6 DM |
2347 | } |
2348 | ||
1da177e4 LT |
2349 | /* This function returns the amount that we can raise the |
2350 | * usable window based on the following constraints | |
e905a9ed | 2351 | * |
1da177e4 LT |
2352 | * 1. The window can never be shrunk once it is offered (RFC 793) |
2353 | * 2. We limit memory per socket | |
2354 | * | |
2355 | * RFC 1122: | |
2356 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep | |
2357 | * RECV.NEXT + RCV.WIN fixed until: | |
2358 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" | |
2359 | * | |
2360 | * i.e. don't raise the right edge of the window until you can raise | |
2361 | * it at least MSS bytes. | |
2362 | * | |
2363 | * Unfortunately, the recommended algorithm breaks header prediction, | |
2364 | * since header prediction assumes th->window stays fixed. | |
2365 | * | |
2366 | * Strictly speaking, keeping th->window fixed violates the receiver | |
2367 | * side SWS prevention criteria. The problem is that under this rule | |
2368 | * a stream of single byte packets will cause the right side of the | |
2369 | * window to always advance by a single byte. | |
e905a9ed | 2370 | * |
1da177e4 LT |
2371 | * Of course, if the sender implements sender side SWS prevention |
2372 | * then this will not be a problem. | |
e905a9ed | 2373 | * |
1da177e4 | 2374 | * BSD seems to make the following compromise: |
e905a9ed | 2375 | * |
1da177e4 LT |
2376 | * If the free space is less than the 1/4 of the maximum |
2377 | * space available and the free space is less than 1/2 mss, | |
2378 | * then set the window to 0. | |
2379 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] | |
2380 | * Otherwise, just prevent the window from shrinking | |
2381 | * and from being larger than the largest representable value. | |
2382 | * | |
2383 | * This prevents incremental opening of the window in the regime | |
2384 | * where TCP is limited by the speed of the reader side taking | |
2385 | * data out of the TCP receive queue. It does nothing about | |
2386 | * those cases where the window is constrained on the sender side | |
2387 | * because the pipeline is full. | |
2388 | * | |
2389 | * BSD also seems to "accidentally" limit itself to windows that are a | |
2390 | * multiple of MSS, at least until the free space gets quite small. | |
2391 | * This would appear to be a side effect of the mbuf implementation. | |
2392 | * Combining these two algorithms results in the observed behavior | |
2393 | * of having a fixed window size at almost all times. | |
2394 | * | |
2395 | * Below we obtain similar behavior by forcing the offered window to | |
2396 | * a multiple of the mss when it is feasible to do so. | |
2397 | * | |
2398 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. | |
2399 | * Regular options like TIMESTAMP are taken into account. | |
2400 | */ | |
2401 | u32 __tcp_select_window(struct sock *sk) | |
2402 | { | |
463c84b9 | 2403 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 2404 | struct tcp_sock *tp = tcp_sk(sk); |
caa20d9a | 2405 | /* MSS for the peer's data. Previous versions used mss_clamp |
1da177e4 LT |
2406 | * here. I don't know if the value based on our guesses |
2407 | * of peer's MSS is better for the performance. It's more correct | |
2408 | * but may be worse for the performance because of rcv_mss | |
2409 | * fluctuations. --SAW 1998/11/1 | |
2410 | */ | |
463c84b9 | 2411 | int mss = icsk->icsk_ack.rcv_mss; |
1da177e4 | 2412 | int free_space = tcp_space(sk); |
86c1a045 FW |
2413 | int allowed_space = tcp_full_space(sk); |
2414 | int full_space = min_t(int, tp->window_clamp, allowed_space); | |
1da177e4 LT |
2415 | int window; |
2416 | ||
2417 | if (mss > full_space) | |
e905a9ed | 2418 | mss = full_space; |
1da177e4 | 2419 | |
b92edbe0 | 2420 | if (free_space < (full_space >> 1)) { |
463c84b9 | 2421 | icsk->icsk_ack.quick = 0; |
1da177e4 | 2422 | |
b8da51eb | 2423 | if (tcp_under_memory_pressure(sk)) |
056834d9 IJ |
2424 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, |
2425 | 4U * tp->advmss); | |
1da177e4 | 2426 | |
86c1a045 FW |
2427 | /* free_space might become our new window, make sure we don't |
2428 | * increase it due to wscale. | |
2429 | */ | |
2430 | free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); | |
2431 | ||
2432 | /* if free space is less than mss estimate, or is below 1/16th | |
2433 | * of the maximum allowed, try to move to zero-window, else | |
2434 | * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and | |
2435 | * new incoming data is dropped due to memory limits. | |
2436 | * With large window, mss test triggers way too late in order | |
2437 | * to announce zero window in time before rmem limit kicks in. | |
2438 | */ | |
2439 | if (free_space < (allowed_space >> 4) || free_space < mss) | |
1da177e4 LT |
2440 | return 0; |
2441 | } | |
2442 | ||
2443 | if (free_space > tp->rcv_ssthresh) | |
2444 | free_space = tp->rcv_ssthresh; | |
2445 | ||
2446 | /* Don't do rounding if we are using window scaling, since the | |
2447 | * scaled window will not line up with the MSS boundary anyway. | |
2448 | */ | |
2449 | window = tp->rcv_wnd; | |
2450 | if (tp->rx_opt.rcv_wscale) { | |
2451 | window = free_space; | |
2452 | ||
2453 | /* Advertise enough space so that it won't get scaled away. | |
2454 | * Import case: prevent zero window announcement if | |
2455 | * 1<<rcv_wscale > mss. | |
2456 | */ | |
2457 | if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) | |
2458 | window = (((window >> tp->rx_opt.rcv_wscale) + 1) | |
2459 | << tp->rx_opt.rcv_wscale); | |
2460 | } else { | |
2461 | /* Get the largest window that is a nice multiple of mss. | |
2462 | * Window clamp already applied above. | |
2463 | * If our current window offering is within 1 mss of the | |
2464 | * free space we just keep it. This prevents the divide | |
2465 | * and multiply from happening most of the time. | |
2466 | * We also don't do any window rounding when the free space | |
2467 | * is too small. | |
2468 | */ | |
2469 | if (window <= free_space - mss || window > free_space) | |
056834d9 | 2470 | window = (free_space / mss) * mss; |
84565070 | 2471 | else if (mss == full_space && |
b92edbe0 | 2472 | free_space > window + (full_space >> 1)) |
84565070 | 2473 | window = free_space; |
1da177e4 LT |
2474 | } |
2475 | ||
2476 | return window; | |
2477 | } | |
2478 | ||
cfea5a68 MKL |
2479 | void tcp_skb_collapse_tstamp(struct sk_buff *skb, |
2480 | const struct sk_buff *next_skb) | |
082ac2d5 | 2481 | { |
0a2cf20c SHY |
2482 | if (unlikely(tcp_has_tx_tstamp(next_skb))) { |
2483 | const struct skb_shared_info *next_shinfo = | |
2484 | skb_shinfo(next_skb); | |
082ac2d5 MKL |
2485 | struct skb_shared_info *shinfo = skb_shinfo(skb); |
2486 | ||
0a2cf20c | 2487 | shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; |
082ac2d5 | 2488 | shinfo->tskey = next_shinfo->tskey; |
2de8023e MKL |
2489 | TCP_SKB_CB(skb)->txstamp_ack |= |
2490 | TCP_SKB_CB(next_skb)->txstamp_ack; | |
082ac2d5 MKL |
2491 | } |
2492 | } | |
2493 | ||
4a17fc3a IJ |
2494 | /* Collapses two adjacent SKB's during retransmission. */ |
2495 | static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) | |
1da177e4 LT |
2496 | { |
2497 | struct tcp_sock *tp = tcp_sk(sk); | |
fe067e8a | 2498 | struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); |
058dc334 | 2499 | int skb_size, next_skb_size; |
1da177e4 | 2500 | |
058dc334 IJ |
2501 | skb_size = skb->len; |
2502 | next_skb_size = next_skb->len; | |
1da177e4 | 2503 | |
058dc334 | 2504 | BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); |
a6963a6b | 2505 | |
058dc334 | 2506 | tcp_highest_sack_combine(sk, next_skb, skb); |
1da177e4 | 2507 | |
058dc334 | 2508 | tcp_unlink_write_queue(next_skb, sk); |
1da177e4 | 2509 | |
058dc334 IJ |
2510 | skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size), |
2511 | next_skb_size); | |
1da177e4 | 2512 | |
058dc334 IJ |
2513 | if (next_skb->ip_summed == CHECKSUM_PARTIAL) |
2514 | skb->ip_summed = CHECKSUM_PARTIAL; | |
1da177e4 | 2515 | |
058dc334 IJ |
2516 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
2517 | skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); | |
1da177e4 | 2518 | |
058dc334 IJ |
2519 | /* Update sequence range on original skb. */ |
2520 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; | |
1da177e4 | 2521 | |
e6c7d085 | 2522 | /* Merge over control information. This moves PSH/FIN etc. over */ |
4de075e0 | 2523 | TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; |
058dc334 IJ |
2524 | |
2525 | /* All done, get rid of second SKB and account for it so | |
2526 | * packet counting does not break. | |
2527 | */ | |
2528 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; | |
a643b5d4 | 2529 | TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; |
058dc334 IJ |
2530 | |
2531 | /* changed transmit queue under us so clear hints */ | |
ef9da47c IJ |
2532 | tcp_clear_retrans_hints_partial(tp); |
2533 | if (next_skb == tp->retransmit_skb_hint) | |
2534 | tp->retransmit_skb_hint = skb; | |
058dc334 | 2535 | |
797108d1 IJ |
2536 | tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); |
2537 | ||
082ac2d5 MKL |
2538 | tcp_skb_collapse_tstamp(skb, next_skb); |
2539 | ||
058dc334 | 2540 | sk_wmem_free_skb(sk, next_skb); |
1da177e4 LT |
2541 | } |
2542 | ||
67edfef7 | 2543 | /* Check if coalescing SKBs is legal. */ |
a2a385d6 | 2544 | static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) |
4a17fc3a IJ |
2545 | { |
2546 | if (tcp_skb_pcount(skb) > 1) | |
a2a385d6 | 2547 | return false; |
4a17fc3a IJ |
2548 | /* TODO: SACK collapsing could be used to remove this condition */ |
2549 | if (skb_shinfo(skb)->nr_frags != 0) | |
a2a385d6 | 2550 | return false; |
4a17fc3a | 2551 | if (skb_cloned(skb)) |
a2a385d6 | 2552 | return false; |
4a17fc3a | 2553 | if (skb == tcp_send_head(sk)) |
a2a385d6 | 2554 | return false; |
4a17fc3a IJ |
2555 | /* Some heurestics for collapsing over SACK'd could be invented */ |
2556 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
a2a385d6 | 2557 | return false; |
4a17fc3a | 2558 | |
a2a385d6 | 2559 | return true; |
4a17fc3a IJ |
2560 | } |
2561 | ||
67edfef7 AK |
2562 | /* Collapse packets in the retransmit queue to make to create |
2563 | * less packets on the wire. This is only done on retransmission. | |
2564 | */ | |
4a17fc3a IJ |
2565 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, |
2566 | int space) | |
2567 | { | |
2568 | struct tcp_sock *tp = tcp_sk(sk); | |
2569 | struct sk_buff *skb = to, *tmp; | |
a2a385d6 | 2570 | bool first = true; |
4a17fc3a IJ |
2571 | |
2572 | if (!sysctl_tcp_retrans_collapse) | |
2573 | return; | |
4de075e0 | 2574 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) |
4a17fc3a IJ |
2575 | return; |
2576 | ||
2577 | tcp_for_write_queue_from_safe(skb, tmp, sk) { | |
2578 | if (!tcp_can_collapse(sk, skb)) | |
2579 | break; | |
2580 | ||
a643b5d4 MKL |
2581 | if (!tcp_skb_can_collapse_to(to)) |
2582 | break; | |
2583 | ||
4a17fc3a IJ |
2584 | space -= skb->len; |
2585 | ||
2586 | if (first) { | |
a2a385d6 | 2587 | first = false; |
4a17fc3a IJ |
2588 | continue; |
2589 | } | |
2590 | ||
2591 | if (space < 0) | |
2592 | break; | |
2593 | /* Punt if not enough space exists in the first SKB for | |
2594 | * the data in the second | |
2595 | */ | |
a21d4572 | 2596 | if (skb->len > skb_availroom(to)) |
4a17fc3a IJ |
2597 | break; |
2598 | ||
2599 | if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) | |
2600 | break; | |
2601 | ||
2602 | tcp_collapse_retrans(sk, to); | |
2603 | } | |
2604 | } | |
2605 | ||
1da177e4 LT |
2606 | /* This retransmits one SKB. Policy decisions and retransmit queue |
2607 | * state updates are done by the caller. Returns non-zero if an | |
2608 | * error occurred which prevented the send. | |
2609 | */ | |
10d3be56 | 2610 | int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) |
1da177e4 | 2611 | { |
5d424d5a | 2612 | struct inet_connection_sock *icsk = inet_csk(sk); |
10d3be56 | 2613 | struct tcp_sock *tp = tcp_sk(sk); |
7d227cd2 | 2614 | unsigned int cur_mss; |
10d3be56 ED |
2615 | int diff, len, err; |
2616 | ||
1da177e4 | 2617 | |
10d3be56 ED |
2618 | /* Inconclusive MTU probe */ |
2619 | if (icsk->icsk_mtup.probe_size) | |
5d424d5a | 2620 | icsk->icsk_mtup.probe_size = 0; |
5d424d5a | 2621 | |
1da177e4 | 2622 | /* Do not sent more than we queued. 1/4 is reserved for possible |
caa20d9a | 2623 | * copying overhead: fragmentation, tunneling, mangling etc. |
1da177e4 LT |
2624 | */ |
2625 | if (atomic_read(&sk->sk_wmem_alloc) > | |
2626 | min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf)) | |
2627 | return -EAGAIN; | |
2628 | ||
1f3279ae ED |
2629 | if (skb_still_in_host_queue(sk, skb)) |
2630 | return -EBUSY; | |
2631 | ||
1da177e4 LT |
2632 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { |
2633 | if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
2634 | BUG(); | |
1da177e4 LT |
2635 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
2636 | return -ENOMEM; | |
2637 | } | |
2638 | ||
7d227cd2 SS |
2639 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) |
2640 | return -EHOSTUNREACH; /* Routing failure or similar. */ | |
2641 | ||
0c54b85f | 2642 | cur_mss = tcp_current_mss(sk); |
7d227cd2 | 2643 | |
1da177e4 LT |
2644 | /* If receiver has shrunk his window, and skb is out of |
2645 | * new window, do not retransmit it. The exception is the | |
2646 | * case, when window is shrunk to zero. In this case | |
2647 | * our retransmit serves as a zero window probe. | |
2648 | */ | |
9d4fb27d JP |
2649 | if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && |
2650 | TCP_SKB_CB(skb)->seq != tp->snd_una) | |
1da177e4 LT |
2651 | return -EAGAIN; |
2652 | ||
10d3be56 ED |
2653 | len = cur_mss * segs; |
2654 | if (skb->len > len) { | |
2655 | if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC)) | |
1da177e4 | 2656 | return -ENOMEM; /* We'll try again later. */ |
02276f3c | 2657 | } else { |
10d3be56 ED |
2658 | if (skb_unclone(skb, GFP_ATOMIC)) |
2659 | return -ENOMEM; | |
9eb9362e | 2660 | |
10d3be56 ED |
2661 | diff = tcp_skb_pcount(skb); |
2662 | tcp_set_skb_tso_segs(skb, cur_mss); | |
2663 | diff -= tcp_skb_pcount(skb); | |
2664 | if (diff) | |
2665 | tcp_adjust_pcount(sk, skb, diff); | |
2666 | if (skb->len < cur_mss) | |
2667 | tcp_retrans_try_collapse(sk, skb, cur_mss); | |
1da177e4 LT |
2668 | } |
2669 | ||
49213555 DB |
2670 | /* RFC3168, section 6.1.1.1. ECN fallback */ |
2671 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) | |
2672 | tcp_ecn_clear_syn(sk, skb); | |
2673 | ||
50bceae9 TG |
2674 | /* make sure skb->data is aligned on arches that require it |
2675 | * and check if ack-trimming & collapsing extended the headroom | |
2676 | * beyond what csum_start can cover. | |
2677 | */ | |
2678 | if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || | |
2679 | skb_headroom(skb) >= 0xFFFF)) { | |
10a81980 ED |
2680 | struct sk_buff *nskb; |
2681 | ||
2682 | skb_mstamp_get(&skb->skb_mstamp); | |
2683 | nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); | |
c84a5711 YC |
2684 | err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : |
2685 | -ENOBUFS; | |
117632e6 | 2686 | } else { |
c84a5711 | 2687 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
117632e6 | 2688 | } |
c84a5711 | 2689 | |
fc9f3501 | 2690 | if (likely(!err)) { |
10d3be56 ED |
2691 | segs = tcp_skb_pcount(skb); |
2692 | ||
c84a5711 | 2693 | TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; |
fc9f3501 | 2694 | /* Update global TCP statistics. */ |
10d3be56 | 2695 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); |
fc9f3501 | 2696 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) |
02a1d6e7 | 2697 | __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); |
10d3be56 | 2698 | tp->total_retrans += segs; |
fc9f3501 | 2699 | } |
c84a5711 | 2700 | return err; |
93b174ad YC |
2701 | } |
2702 | ||
10d3be56 | 2703 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) |
93b174ad YC |
2704 | { |
2705 | struct tcp_sock *tp = tcp_sk(sk); | |
10d3be56 | 2706 | int err = __tcp_retransmit_skb(sk, skb, segs); |
1da177e4 LT |
2707 | |
2708 | if (err == 0) { | |
1da177e4 | 2709 | #if FASTRETRANS_DEBUG > 0 |
056834d9 | 2710 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
e87cc472 | 2711 | net_dbg_ratelimited("retrans_out leaked\n"); |
1da177e4 LT |
2712 | } |
2713 | #endif | |
2714 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; | |
2715 | tp->retrans_out += tcp_skb_pcount(skb); | |
2716 | ||
2717 | /* Save stamp of the first retransmit. */ | |
2718 | if (!tp->retrans_stamp) | |
7faee5c0 | 2719 | tp->retrans_stamp = tcp_skb_timestamp(skb); |
1da177e4 | 2720 | |
1f3279ae | 2721 | } else if (err != -EBUSY) { |
c10d9310 | 2722 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); |
1da177e4 | 2723 | } |
6e08d5e3 YC |
2724 | |
2725 | if (tp->undo_retrans < 0) | |
2726 | tp->undo_retrans = 0; | |
2727 | tp->undo_retrans += tcp_skb_pcount(skb); | |
1da177e4 LT |
2728 | return err; |
2729 | } | |
2730 | ||
67edfef7 AK |
2731 | /* Check if we forward retransmits are possible in the current |
2732 | * window/congestion state. | |
2733 | */ | |
a2a385d6 | 2734 | static bool tcp_can_forward_retransmit(struct sock *sk) |
b5afe7bc IJ |
2735 | { |
2736 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
cf533ea5 | 2737 | const struct tcp_sock *tp = tcp_sk(sk); |
b5afe7bc IJ |
2738 | |
2739 | /* Forward retransmissions are possible only during Recovery. */ | |
2740 | if (icsk->icsk_ca_state != TCP_CA_Recovery) | |
a2a385d6 | 2741 | return false; |
b5afe7bc IJ |
2742 | |
2743 | /* No forward retransmissions in Reno are possible. */ | |
2744 | if (tcp_is_reno(tp)) | |
a2a385d6 | 2745 | return false; |
b5afe7bc IJ |
2746 | |
2747 | /* Yeah, we have to make difficult choice between forward transmission | |
2748 | * and retransmission... Both ways have their merits... | |
2749 | * | |
2750 | * For now we do not retransmit anything, while we have some new | |
2751 | * segments to send. In the other cases, follow rule 3 for | |
2752 | * NextSeg() specified in RFC3517. | |
2753 | */ | |
2754 | ||
2755 | if (tcp_may_send_now(sk)) | |
a2a385d6 | 2756 | return false; |
b5afe7bc | 2757 | |
a2a385d6 | 2758 | return true; |
b5afe7bc IJ |
2759 | } |
2760 | ||
1da177e4 LT |
2761 | /* This gets called after a retransmit timeout, and the initially |
2762 | * retransmitted data is acknowledged. It tries to continue | |
2763 | * resending the rest of the retransmit queue, until either | |
2764 | * we've sent it all or the congestion window limit is reached. | |
2765 | * If doing SACK, the first ACK which comes back for a timeout | |
2766 | * based retransmit packet might feed us FACK information again. | |
2767 | * If so, we use it to avoid unnecessarily retransmissions. | |
2768 | */ | |
2769 | void tcp_xmit_retransmit_queue(struct sock *sk) | |
2770 | { | |
6687e988 | 2771 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2772 | struct tcp_sock *tp = tcp_sk(sk); |
2773 | struct sk_buff *skb; | |
0e1c54c2 | 2774 | struct sk_buff *hole = NULL; |
a3d2e9f8 | 2775 | u32 max_segs, last_lost; |
61eb55f4 | 2776 | int mib_idx; |
0e1c54c2 | 2777 | int fwd_rexmitting = 0; |
6a438bbe | 2778 | |
45e77d31 IJ |
2779 | if (!tp->packets_out) |
2780 | return; | |
2781 | ||
08ebd172 IJ |
2782 | if (!tp->lost_out) |
2783 | tp->retransmit_high = tp->snd_una; | |
2784 | ||
618d9f25 | 2785 | if (tp->retransmit_skb_hint) { |
6a438bbe | 2786 | skb = tp->retransmit_skb_hint; |
618d9f25 IJ |
2787 | last_lost = TCP_SKB_CB(skb)->end_seq; |
2788 | if (after(last_lost, tp->retransmit_high)) | |
2789 | last_lost = tp->retransmit_high; | |
2790 | } else { | |
fe067e8a | 2791 | skb = tcp_write_queue_head(sk); |
618d9f25 IJ |
2792 | last_lost = tp->snd_una; |
2793 | } | |
1da177e4 | 2794 | |
ed6e7268 | 2795 | max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); |
08ebd172 | 2796 | tcp_for_write_queue_from(skb, sk) { |
dca0aaf8 | 2797 | __u8 sacked; |
10d3be56 | 2798 | int segs; |
1da177e4 | 2799 | |
08ebd172 IJ |
2800 | if (skb == tcp_send_head(sk)) |
2801 | break; | |
2802 | /* we could do better than to assign each time */ | |
51456b29 | 2803 | if (!hole) |
0e1c54c2 | 2804 | tp->retransmit_skb_hint = skb; |
08ebd172 | 2805 | |
10d3be56 ED |
2806 | segs = tp->snd_cwnd - tcp_packets_in_flight(tp); |
2807 | if (segs <= 0) | |
08ebd172 | 2808 | return; |
dca0aaf8 | 2809 | sacked = TCP_SKB_CB(skb)->sacked; |
a3d2e9f8 ED |
2810 | /* In case tcp_shift_skb_data() have aggregated large skbs, |
2811 | * we need to make sure not sending too bigs TSO packets | |
2812 | */ | |
2813 | segs = min_t(int, segs, max_segs); | |
1da177e4 | 2814 | |
0e1c54c2 IJ |
2815 | if (fwd_rexmitting) { |
2816 | begin_fwd: | |
2817 | if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp))) | |
2818 | break; | |
2819 | mib_idx = LINUX_MIB_TCPFORWARDRETRANS; | |
6a438bbe | 2820 | |
0e1c54c2 | 2821 | } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) { |
618d9f25 | 2822 | tp->retransmit_high = last_lost; |
0e1c54c2 IJ |
2823 | if (!tcp_can_forward_retransmit(sk)) |
2824 | break; | |
2825 | /* Backtrack if necessary to non-L'ed skb */ | |
00db4124 | 2826 | if (hole) { |
0e1c54c2 IJ |
2827 | skb = hole; |
2828 | hole = NULL; | |
2829 | } | |
2830 | fwd_rexmitting = 1; | |
2831 | goto begin_fwd; | |
1da177e4 | 2832 | |
0e1c54c2 | 2833 | } else if (!(sacked & TCPCB_LOST)) { |
51456b29 | 2834 | if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) |
0e1c54c2 IJ |
2835 | hole = skb; |
2836 | continue; | |
1da177e4 | 2837 | |
0e1c54c2 | 2838 | } else { |
618d9f25 | 2839 | last_lost = TCP_SKB_CB(skb)->end_seq; |
0e1c54c2 IJ |
2840 | if (icsk->icsk_ca_state != TCP_CA_Loss) |
2841 | mib_idx = LINUX_MIB_TCPFASTRETRANS; | |
2842 | else | |
2843 | mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; | |
2844 | } | |
1da177e4 | 2845 | |
0e1c54c2 | 2846 | if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) |
1da177e4 LT |
2847 | continue; |
2848 | ||
10d3be56 | 2849 | if (tcp_retransmit_skb(sk, skb, segs)) |
0e1c54c2 | 2850 | return; |
24ab6bec | 2851 | |
c10d9310 | 2852 | NET_INC_STATS(sock_net(sk), mib_idx); |
1da177e4 | 2853 | |
684bad11 | 2854 | if (tcp_in_cwnd_reduction(sk)) |
a262f0cd ND |
2855 | tp->prr_out += tcp_skb_pcount(skb); |
2856 | ||
fe067e8a | 2857 | if (skb == tcp_write_queue_head(sk)) |
3f421baa ACM |
2858 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
2859 | inet_csk(sk)->icsk_rto, | |
2860 | TCP_RTO_MAX); | |
1da177e4 LT |
2861 | } |
2862 | } | |
2863 | ||
d83769a5 ED |
2864 | /* We allow to exceed memory limits for FIN packets to expedite |
2865 | * connection tear down and (memory) recovery. | |
845704a5 ED |
2866 | * Otherwise tcp_send_fin() could be tempted to either delay FIN |
2867 | * or even be forced to close flow without any FIN. | |
a6c5ea4c ED |
2868 | * In general, we want to allow one skb per socket to avoid hangs |
2869 | * with edge trigger epoll() | |
d83769a5 | 2870 | */ |
a6c5ea4c | 2871 | void sk_forced_mem_schedule(struct sock *sk, int size) |
d83769a5 | 2872 | { |
e805605c | 2873 | int amt; |
d83769a5 ED |
2874 | |
2875 | if (size <= sk->sk_forward_alloc) | |
2876 | return; | |
2877 | amt = sk_mem_pages(size); | |
2878 | sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; | |
e805605c JW |
2879 | sk_memory_allocated_add(sk, amt); |
2880 | ||
baac50bb JW |
2881 | if (mem_cgroup_sockets_enabled && sk->sk_memcg) |
2882 | mem_cgroup_charge_skmem(sk->sk_memcg, amt); | |
d83769a5 ED |
2883 | } |
2884 | ||
845704a5 ED |
2885 | /* Send a FIN. The caller locks the socket for us. |
2886 | * We should try to send a FIN packet really hard, but eventually give up. | |
1da177e4 LT |
2887 | */ |
2888 | void tcp_send_fin(struct sock *sk) | |
2889 | { | |
845704a5 | 2890 | struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); |
e905a9ed | 2891 | struct tcp_sock *tp = tcp_sk(sk); |
e905a9ed | 2892 | |
845704a5 ED |
2893 | /* Optimization, tack on the FIN if we have one skb in write queue and |
2894 | * this skb was not yet sent, or we are under memory pressure. | |
2895 | * Note: in the latter case, FIN packet will be sent after a timeout, | |
2896 | * as TCP stack thinks it has already been transmitted. | |
1da177e4 | 2897 | */ |
b8da51eb | 2898 | if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) { |
845704a5 ED |
2899 | coalesce: |
2900 | TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; | |
2901 | TCP_SKB_CB(tskb)->end_seq++; | |
1da177e4 | 2902 | tp->write_seq++; |
845704a5 ED |
2903 | if (!tcp_send_head(sk)) { |
2904 | /* This means tskb was already sent. | |
2905 | * Pretend we included the FIN on previous transmit. | |
2906 | * We need to set tp->snd_nxt to the value it would have | |
2907 | * if FIN had been sent. This is because retransmit path | |
2908 | * does not change tp->snd_nxt. | |
2909 | */ | |
2910 | tp->snd_nxt++; | |
2911 | return; | |
2912 | } | |
1da177e4 | 2913 | } else { |
845704a5 ED |
2914 | skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); |
2915 | if (unlikely(!skb)) { | |
2916 | if (tskb) | |
2917 | goto coalesce; | |
2918 | return; | |
1da177e4 | 2919 | } |
d83769a5 | 2920 | skb_reserve(skb, MAX_TCP_HEADER); |
a6c5ea4c | 2921 | sk_forced_mem_schedule(sk, skb->truesize); |
1da177e4 | 2922 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ |
e870a8ef | 2923 | tcp_init_nondata_skb(skb, tp->write_seq, |
a3433f35 | 2924 | TCPHDR_ACK | TCPHDR_FIN); |
1da177e4 LT |
2925 | tcp_queue_skb(sk, skb); |
2926 | } | |
845704a5 | 2927 | __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); |
1da177e4 LT |
2928 | } |
2929 | ||
2930 | /* We get here when a process closes a file descriptor (either due to | |
2931 | * an explicit close() or as a byproduct of exit()'ing) and there | |
2932 | * was unread data in the receive queue. This behavior is recommended | |
65bb723c | 2933 | * by RFC 2525, section 2.17. -DaveM |
1da177e4 | 2934 | */ |
dd0fc66f | 2935 | void tcp_send_active_reset(struct sock *sk, gfp_t priority) |
1da177e4 | 2936 | { |
1da177e4 LT |
2937 | struct sk_buff *skb; |
2938 | ||
2939 | /* NOTE: No TCP options attached and we never retransmit this. */ | |
2940 | skb = alloc_skb(MAX_TCP_HEADER, priority); | |
2941 | if (!skb) { | |
4e673444 | 2942 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
1da177e4 LT |
2943 | return; |
2944 | } | |
2945 | ||
2946 | /* Reserve space for headers and prepare control bits. */ | |
2947 | skb_reserve(skb, MAX_TCP_HEADER); | |
e870a8ef | 2948 | tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), |
a3433f35 | 2949 | TCPHDR_ACK | TCPHDR_RST); |
675ee231 | 2950 | skb_mstamp_get(&skb->skb_mstamp); |
1da177e4 | 2951 | /* Send it off. */ |
dfb4b9dc | 2952 | if (tcp_transmit_skb(sk, skb, 0, priority)) |
4e673444 | 2953 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
26af65cb | 2954 | |
81cc8a75 | 2955 | TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); |
1da177e4 LT |
2956 | } |
2957 | ||
67edfef7 AK |
2958 | /* Send a crossed SYN-ACK during socket establishment. |
2959 | * WARNING: This routine must only be called when we have already sent | |
1da177e4 LT |
2960 | * a SYN packet that crossed the incoming SYN that caused this routine |
2961 | * to get called. If this assumption fails then the initial rcv_wnd | |
2962 | * and rcv_wscale values will not be correct. | |
2963 | */ | |
2964 | int tcp_send_synack(struct sock *sk) | |
2965 | { | |
056834d9 | 2966 | struct sk_buff *skb; |
1da177e4 | 2967 | |
fe067e8a | 2968 | skb = tcp_write_queue_head(sk); |
51456b29 | 2969 | if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { |
91df42be | 2970 | pr_debug("%s: wrong queue state\n", __func__); |
1da177e4 LT |
2971 | return -EFAULT; |
2972 | } | |
4de075e0 | 2973 | if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { |
1da177e4 LT |
2974 | if (skb_cloned(skb)) { |
2975 | struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); | |
51456b29 | 2976 | if (!nskb) |
1da177e4 | 2977 | return -ENOMEM; |
fe067e8a | 2978 | tcp_unlink_write_queue(skb, sk); |
f4a775d1 | 2979 | __skb_header_release(nskb); |
fe067e8a | 2980 | __tcp_add_write_queue_head(sk, nskb); |
3ab224be HA |
2981 | sk_wmem_free_skb(sk, skb); |
2982 | sk->sk_wmem_queued += nskb->truesize; | |
2983 | sk_mem_charge(sk, nskb->truesize); | |
1da177e4 LT |
2984 | skb = nskb; |
2985 | } | |
2986 | ||
4de075e0 | 2987 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; |
735d3831 | 2988 | tcp_ecn_send_synack(sk, skb); |
1da177e4 | 2989 | } |
dfb4b9dc | 2990 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
1da177e4 LT |
2991 | } |
2992 | ||
4aea39c1 ED |
2993 | /** |
2994 | * tcp_make_synack - Prepare a SYN-ACK. | |
2995 | * sk: listener socket | |
2996 | * dst: dst entry attached to the SYNACK | |
2997 | * req: request_sock pointer | |
4aea39c1 ED |
2998 | * |
2999 | * Allocate one skb and build a SYNACK packet. | |
3000 | * @dst is consumed : Caller should not use it again. | |
3001 | */ | |
5d062de7 | 3002 | struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, |
e6b4d113 | 3003 | struct request_sock *req, |
ca6fb065 | 3004 | struct tcp_fastopen_cookie *foc, |
b3d05147 | 3005 | enum tcp_synack_type synack_type) |
1da177e4 | 3006 | { |
2e6599cb | 3007 | struct inet_request_sock *ireq = inet_rsk(req); |
5d062de7 | 3008 | const struct tcp_sock *tp = tcp_sk(sk); |
80f03e27 | 3009 | struct tcp_md5sig_key *md5 = NULL; |
5d062de7 ED |
3010 | struct tcp_out_options opts; |
3011 | struct sk_buff *skb; | |
bd0388ae | 3012 | int tcp_header_size; |
5d062de7 ED |
3013 | struct tcphdr *th; |
3014 | u16 user_mss; | |
f5fff5dc | 3015 | int mss; |
1da177e4 | 3016 | |
ca6fb065 | 3017 | skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); |
4aea39c1 ED |
3018 | if (unlikely(!skb)) { |
3019 | dst_release(dst); | |
1da177e4 | 3020 | return NULL; |
4aea39c1 | 3021 | } |
1da177e4 LT |
3022 | /* Reserve space for headers. */ |
3023 | skb_reserve(skb, MAX_TCP_HEADER); | |
3024 | ||
b3d05147 ED |
3025 | switch (synack_type) { |
3026 | case TCP_SYNACK_NORMAL: | |
9e17f8a4 | 3027 | skb_set_owner_w(skb, req_to_sk(req)); |
b3d05147 ED |
3028 | break; |
3029 | case TCP_SYNACK_COOKIE: | |
3030 | /* Under synflood, we do not attach skb to a socket, | |
3031 | * to avoid false sharing. | |
3032 | */ | |
3033 | break; | |
3034 | case TCP_SYNACK_FASTOPEN: | |
ca6fb065 ED |
3035 | /* sk is a const pointer, because we want to express multiple |
3036 | * cpu might call us concurrently. | |
3037 | * sk->sk_wmem_alloc in an atomic, we can promote to rw. | |
3038 | */ | |
3039 | skb_set_owner_w(skb, (struct sock *)sk); | |
b3d05147 | 3040 | break; |
ca6fb065 | 3041 | } |
4aea39c1 | 3042 | skb_dst_set(skb, dst); |
1da177e4 | 3043 | |
0dbaee3b | 3044 | mss = dst_metric_advmss(dst); |
5d062de7 ED |
3045 | user_mss = READ_ONCE(tp->rx_opt.user_mss); |
3046 | if (user_mss && user_mss < mss) | |
3047 | mss = user_mss; | |
f5fff5dc | 3048 | |
33ad798c | 3049 | memset(&opts, 0, sizeof(opts)); |
8b5f12d0 FW |
3050 | #ifdef CONFIG_SYN_COOKIES |
3051 | if (unlikely(req->cookie_ts)) | |
7faee5c0 | 3052 | skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req); |
8b5f12d0 FW |
3053 | else |
3054 | #endif | |
7faee5c0 | 3055 | skb_mstamp_get(&skb->skb_mstamp); |
80f03e27 ED |
3056 | |
3057 | #ifdef CONFIG_TCP_MD5SIG | |
3058 | rcu_read_lock(); | |
fd3a154a | 3059 | md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); |
80f03e27 | 3060 | #endif |
58d607d3 | 3061 | skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); |
37bfbdda ED |
3062 | tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) + |
3063 | sizeof(*th); | |
cfb6eeb4 | 3064 | |
aa8223c7 ACM |
3065 | skb_push(skb, tcp_header_size); |
3066 | skb_reset_transport_header(skb); | |
1da177e4 | 3067 | |
ea1627c2 | 3068 | th = (struct tcphdr *)skb->data; |
1da177e4 LT |
3069 | memset(th, 0, sizeof(struct tcphdr)); |
3070 | th->syn = 1; | |
3071 | th->ack = 1; | |
6ac705b1 | 3072 | tcp_ecn_make_synack(req, th); |
b44084c2 | 3073 | th->source = htons(ireq->ir_num); |
634fb979 | 3074 | th->dest = ireq->ir_rmt_port; |
e870a8ef IJ |
3075 | /* Setting of flags are superfluous here for callers (and ECE is |
3076 | * not even correctly set) | |
3077 | */ | |
3078 | tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, | |
a3433f35 | 3079 | TCPHDR_SYN | TCPHDR_ACK); |
4957faad | 3080 | |
1da177e4 | 3081 | th->seq = htonl(TCP_SKB_CB(skb)->seq); |
8336886f JC |
3082 | /* XXX data is queued and acked as is. No buffer/window check */ |
3083 | th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); | |
1da177e4 LT |
3084 | |
3085 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ | |
ed53d0ab | 3086 | th->window = htons(min(req->rsk_rcv_wnd, 65535U)); |
5d062de7 | 3087 | tcp_options_write((__be32 *)(th + 1), NULL, &opts); |
1da177e4 | 3088 | th->doff = (tcp_header_size >> 2); |
90bbcc60 | 3089 | __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); |
cfb6eeb4 YH |
3090 | |
3091 | #ifdef CONFIG_TCP_MD5SIG | |
3092 | /* Okay, we have all we need - do the md5 hash if needed */ | |
80f03e27 | 3093 | if (md5) |
bd0388ae | 3094 | tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, |
39f8e58e | 3095 | md5, req_to_sk(req), skb); |
80f03e27 | 3096 | rcu_read_unlock(); |
cfb6eeb4 YH |
3097 | #endif |
3098 | ||
b50edd78 ED |
3099 | /* Do not fool tcpdump (if any), clean our debris */ |
3100 | skb->tstamp.tv64 = 0; | |
1da177e4 LT |
3101 | return skb; |
3102 | } | |
4bc2f18b | 3103 | EXPORT_SYMBOL(tcp_make_synack); |
1da177e4 | 3104 | |
81164413 DB |
3105 | static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) |
3106 | { | |
3107 | struct inet_connection_sock *icsk = inet_csk(sk); | |
3108 | const struct tcp_congestion_ops *ca; | |
3109 | u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); | |
3110 | ||
3111 | if (ca_key == TCP_CA_UNSPEC) | |
3112 | return; | |
3113 | ||
3114 | rcu_read_lock(); | |
3115 | ca = tcp_ca_find_key(ca_key); | |
3116 | if (likely(ca && try_module_get(ca->owner))) { | |
3117 | module_put(icsk->icsk_ca_ops->owner); | |
3118 | icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); | |
3119 | icsk->icsk_ca_ops = ca; | |
3120 | } | |
3121 | rcu_read_unlock(); | |
3122 | } | |
3123 | ||
67edfef7 | 3124 | /* Do all connect socket setups that can be done AF independent. */ |
f7e56a76 | 3125 | static void tcp_connect_init(struct sock *sk) |
1da177e4 | 3126 | { |
cf533ea5 | 3127 | const struct dst_entry *dst = __sk_dst_get(sk); |
1da177e4 LT |
3128 | struct tcp_sock *tp = tcp_sk(sk); |
3129 | __u8 rcv_wscale; | |
3130 | ||
3131 | /* We'll fix this up when we get a response from the other end. | |
3132 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. | |
3133 | */ | |
3134 | tp->tcp_header_len = sizeof(struct tcphdr) + | |
bb5b7c11 | 3135 | (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); |
1da177e4 | 3136 | |
cfb6eeb4 | 3137 | #ifdef CONFIG_TCP_MD5SIG |
00db4124 | 3138 | if (tp->af_specific->md5_lookup(sk, sk)) |
cfb6eeb4 YH |
3139 | tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; |
3140 | #endif | |
3141 | ||
1da177e4 LT |
3142 | /* If user gave his TCP_MAXSEG, record it to clamp */ |
3143 | if (tp->rx_opt.user_mss) | |
3144 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
3145 | tp->max_window = 0; | |
5d424d5a | 3146 | tcp_mtup_init(sk); |
1da177e4 LT |
3147 | tcp_sync_mss(sk, dst_mtu(dst)); |
3148 | ||
81164413 DB |
3149 | tcp_ca_dst_init(sk, dst); |
3150 | ||
1da177e4 LT |
3151 | if (!tp->window_clamp) |
3152 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); | |
0dbaee3b | 3153 | tp->advmss = dst_metric_advmss(dst); |
f5fff5dc TQ |
3154 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) |
3155 | tp->advmss = tp->rx_opt.user_mss; | |
3156 | ||
1da177e4 | 3157 | tcp_initialize_rcv_mss(sk); |
1da177e4 | 3158 | |
e88c64f0 HPP |
3159 | /* limit the window selection if the user enforce a smaller rx buffer */ |
3160 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && | |
3161 | (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) | |
3162 | tp->window_clamp = tcp_full_space(sk); | |
3163 | ||
1da177e4 LT |
3164 | tcp_select_initial_window(tcp_full_space(sk), |
3165 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), | |
3166 | &tp->rcv_wnd, | |
3167 | &tp->window_clamp, | |
bb5b7c11 | 3168 | sysctl_tcp_window_scaling, |
31d12926 | 3169 | &rcv_wscale, |
3170 | dst_metric(dst, RTAX_INITRWND)); | |
1da177e4 LT |
3171 | |
3172 | tp->rx_opt.rcv_wscale = rcv_wscale; | |
3173 | tp->rcv_ssthresh = tp->rcv_wnd; | |
3174 | ||
3175 | sk->sk_err = 0; | |
3176 | sock_reset_flag(sk, SOCK_DONE); | |
3177 | tp->snd_wnd = 0; | |
ee7537b6 | 3178 | tcp_init_wl(tp, 0); |
1da177e4 LT |
3179 | tp->snd_una = tp->write_seq; |
3180 | tp->snd_sml = tp->write_seq; | |
33f5f57e | 3181 | tp->snd_up = tp->write_seq; |
370816ae | 3182 | tp->snd_nxt = tp->write_seq; |
ee995283 PE |
3183 | |
3184 | if (likely(!tp->repair)) | |
3185 | tp->rcv_nxt = 0; | |
c7781a6e AV |
3186 | else |
3187 | tp->rcv_tstamp = tcp_time_stamp; | |
ee995283 PE |
3188 | tp->rcv_wup = tp->rcv_nxt; |
3189 | tp->copied_seq = tp->rcv_nxt; | |
1da177e4 | 3190 | |
463c84b9 ACM |
3191 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
3192 | inet_csk(sk)->icsk_retransmits = 0; | |
1da177e4 LT |
3193 | tcp_clear_retrans(tp); |
3194 | } | |
3195 | ||
783237e8 YC |
3196 | static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) |
3197 | { | |
3198 | struct tcp_sock *tp = tcp_sk(sk); | |
3199 | struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); | |
3200 | ||
3201 | tcb->end_seq += skb->len; | |
f4a775d1 | 3202 | __skb_header_release(skb); |
783237e8 YC |
3203 | __tcp_add_write_queue_tail(sk, skb); |
3204 | sk->sk_wmem_queued += skb->truesize; | |
3205 | sk_mem_charge(sk, skb->truesize); | |
3206 | tp->write_seq = tcb->end_seq; | |
3207 | tp->packets_out += tcp_skb_pcount(skb); | |
3208 | } | |
3209 | ||
3210 | /* Build and send a SYN with data and (cached) Fast Open cookie. However, | |
3211 | * queue a data-only packet after the regular SYN, such that regular SYNs | |
3212 | * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges | |
3213 | * only the SYN sequence, the data are retransmitted in the first ACK. | |
3214 | * If cookie is not cached or other error occurs, falls back to send a | |
3215 | * regular SYN with Fast Open cookie request option. | |
3216 | */ | |
3217 | static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) | |
3218 | { | |
3219 | struct tcp_sock *tp = tcp_sk(sk); | |
3220 | struct tcp_fastopen_request *fo = tp->fastopen_req; | |
07e100f9 | 3221 | int syn_loss = 0, space, err = 0; |
aab48743 | 3222 | unsigned long last_syn_loss = 0; |
355a901e | 3223 | struct sk_buff *syn_data; |
aab48743 | 3224 | |
67da22d2 | 3225 | tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ |
aab48743 YC |
3226 | tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie, |
3227 | &syn_loss, &last_syn_loss); | |
3228 | /* Recurring FO SYN losses: revert to regular handshake temporarily */ | |
3229 | if (syn_loss > 1 && | |
3230 | time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) { | |
3231 | fo->cookie.len = -1; | |
3232 | goto fallback; | |
3233 | } | |
783237e8 | 3234 | |
67da22d2 YC |
3235 | if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE) |
3236 | fo->cookie.len = -1; | |
3237 | else if (fo->cookie.len <= 0) | |
783237e8 YC |
3238 | goto fallback; |
3239 | ||
3240 | /* MSS for SYN-data is based on cached MSS and bounded by PMTU and | |
3241 | * user-MSS. Reserve maximum option space for middleboxes that add | |
3242 | * private TCP options. The cost is reduced data space in SYN :( | |
3243 | */ | |
3244 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) | |
3245 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
1b63edd6 | 3246 | space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - |
783237e8 YC |
3247 | MAX_TCP_OPTION_SPACE; |
3248 | ||
f5ddcbbb ED |
3249 | space = min_t(size_t, space, fo->size); |
3250 | ||
3251 | /* limit to order-0 allocations */ | |
3252 | space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); | |
3253 | ||
eb934478 | 3254 | syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); |
355a901e | 3255 | if (!syn_data) |
783237e8 | 3256 | goto fallback; |
355a901e ED |
3257 | syn_data->ip_summed = CHECKSUM_PARTIAL; |
3258 | memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); | |
07e100f9 ED |
3259 | if (space) { |
3260 | int copied = copy_from_iter(skb_put(syn_data, space), space, | |
3261 | &fo->data->msg_iter); | |
3262 | if (unlikely(!copied)) { | |
3263 | kfree_skb(syn_data); | |
3264 | goto fallback; | |
3265 | } | |
3266 | if (copied != space) { | |
3267 | skb_trim(syn_data, copied); | |
3268 | space = copied; | |
3269 | } | |
57be5bda | 3270 | } |
355a901e ED |
3271 | /* No more data pending in inet_wait_for_connect() */ |
3272 | if (space == fo->size) | |
3273 | fo->data = NULL; | |
3274 | fo->copied = space; | |
783237e8 | 3275 | |
355a901e | 3276 | tcp_connect_queue_skb(sk, syn_data); |
783237e8 | 3277 | |
355a901e | 3278 | err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); |
783237e8 | 3279 | |
355a901e | 3280 | syn->skb_mstamp = syn_data->skb_mstamp; |
431a9124 | 3281 | |
355a901e ED |
3282 | /* Now full SYN+DATA was cloned and sent (or not), |
3283 | * remove the SYN from the original skb (syn_data) | |
3284 | * we keep in write queue in case of a retransmit, as we | |
3285 | * also have the SYN packet (with no data) in the same queue. | |
3286 | */ | |
3287 | TCP_SKB_CB(syn_data)->seq++; | |
3288 | TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; | |
3289 | if (!err) { | |
67da22d2 | 3290 | tp->syn_data = (fo->copied > 0); |
f19c29e3 | 3291 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); |
783237e8 YC |
3292 | goto done; |
3293 | } | |
783237e8 YC |
3294 | |
3295 | fallback: | |
3296 | /* Send a regular SYN with Fast Open cookie request option */ | |
3297 | if (fo->cookie.len > 0) | |
3298 | fo->cookie.len = 0; | |
3299 | err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); | |
3300 | if (err) | |
3301 | tp->syn_fastopen = 0; | |
783237e8 YC |
3302 | done: |
3303 | fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ | |
3304 | return err; | |
3305 | } | |
3306 | ||
67edfef7 | 3307 | /* Build a SYN and send it off. */ |
1da177e4 LT |
3308 | int tcp_connect(struct sock *sk) |
3309 | { | |
3310 | struct tcp_sock *tp = tcp_sk(sk); | |
3311 | struct sk_buff *buff; | |
ee586811 | 3312 | int err; |
1da177e4 LT |
3313 | |
3314 | tcp_connect_init(sk); | |
3315 | ||
2b916477 AV |
3316 | if (unlikely(tp->repair)) { |
3317 | tcp_finish_connect(sk, NULL); | |
3318 | return 0; | |
3319 | } | |
3320 | ||
eb934478 | 3321 | buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); |
355a901e | 3322 | if (unlikely(!buff)) |
1da177e4 LT |
3323 | return -ENOBUFS; |
3324 | ||
a3433f35 | 3325 | tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); |
7faee5c0 | 3326 | tp->retrans_stamp = tcp_time_stamp; |
783237e8 | 3327 | tcp_connect_queue_skb(sk, buff); |
735d3831 | 3328 | tcp_ecn_send_syn(sk, buff); |
1da177e4 | 3329 | |
783237e8 YC |
3330 | /* Send off SYN; include data in Fast Open. */ |
3331 | err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : | |
3332 | tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); | |
ee586811 EP |
3333 | if (err == -ECONNREFUSED) |
3334 | return err; | |
bd37a088 WY |
3335 | |
3336 | /* We change tp->snd_nxt after the tcp_transmit_skb() call | |
3337 | * in order to make this packet get counted in tcpOutSegs. | |
3338 | */ | |
3339 | tp->snd_nxt = tp->write_seq; | |
3340 | tp->pushed_seq = tp->write_seq; | |
81cc8a75 | 3341 | TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); |
1da177e4 LT |
3342 | |
3343 | /* Timer for repeating the SYN until an answer. */ | |
3f421baa ACM |
3344 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
3345 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
1da177e4 LT |
3346 | return 0; |
3347 | } | |
4bc2f18b | 3348 | EXPORT_SYMBOL(tcp_connect); |
1da177e4 LT |
3349 | |
3350 | /* Send out a delayed ack, the caller does the policy checking | |
3351 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() | |
3352 | * for details. | |
3353 | */ | |
3354 | void tcp_send_delayed_ack(struct sock *sk) | |
3355 | { | |
463c84b9 ACM |
3356 | struct inet_connection_sock *icsk = inet_csk(sk); |
3357 | int ato = icsk->icsk_ack.ato; | |
1da177e4 LT |
3358 | unsigned long timeout; |
3359 | ||
9890092e FW |
3360 | tcp_ca_event(sk, CA_EVENT_DELAYED_ACK); |
3361 | ||
1da177e4 | 3362 | if (ato > TCP_DELACK_MIN) { |
463c84b9 | 3363 | const struct tcp_sock *tp = tcp_sk(sk); |
056834d9 | 3364 | int max_ato = HZ / 2; |
1da177e4 | 3365 | |
056834d9 IJ |
3366 | if (icsk->icsk_ack.pingpong || |
3367 | (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) | |
1da177e4 LT |
3368 | max_ato = TCP_DELACK_MAX; |
3369 | ||
3370 | /* Slow path, intersegment interval is "high". */ | |
3371 | ||
3372 | /* If some rtt estimate is known, use it to bound delayed ack. | |
463c84b9 | 3373 | * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements |
1da177e4 LT |
3374 | * directly. |
3375 | */ | |
740b0f18 ED |
3376 | if (tp->srtt_us) { |
3377 | int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), | |
3378 | TCP_DELACK_MIN); | |
1da177e4 LT |
3379 | |
3380 | if (rtt < max_ato) | |
3381 | max_ato = rtt; | |
3382 | } | |
3383 | ||
3384 | ato = min(ato, max_ato); | |
3385 | } | |
3386 | ||
3387 | /* Stay within the limit we were given */ | |
3388 | timeout = jiffies + ato; | |
3389 | ||
3390 | /* Use new timeout only if there wasn't a older one earlier. */ | |
463c84b9 | 3391 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { |
1da177e4 LT |
3392 | /* If delack timer was blocked or is about to expire, |
3393 | * send ACK now. | |
3394 | */ | |
463c84b9 ACM |
3395 | if (icsk->icsk_ack.blocked || |
3396 | time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { | |
1da177e4 LT |
3397 | tcp_send_ack(sk); |
3398 | return; | |
3399 | } | |
3400 | ||
463c84b9 ACM |
3401 | if (!time_before(timeout, icsk->icsk_ack.timeout)) |
3402 | timeout = icsk->icsk_ack.timeout; | |
1da177e4 | 3403 | } |
463c84b9 ACM |
3404 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; |
3405 | icsk->icsk_ack.timeout = timeout; | |
3406 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); | |
1da177e4 LT |
3407 | } |
3408 | ||
3409 | /* This routine sends an ack and also updates the window. */ | |
3410 | void tcp_send_ack(struct sock *sk) | |
3411 | { | |
058dc334 | 3412 | struct sk_buff *buff; |
1da177e4 | 3413 | |
058dc334 IJ |
3414 | /* If we have been reset, we may not send again. */ |
3415 | if (sk->sk_state == TCP_CLOSE) | |
3416 | return; | |
1da177e4 | 3417 | |
9890092e FW |
3418 | tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK); |
3419 | ||
058dc334 IJ |
3420 | /* We are not putting this on the write queue, so |
3421 | * tcp_transmit_skb() will set the ownership to this | |
3422 | * sock. | |
3423 | */ | |
7450aaf6 ED |
3424 | buff = alloc_skb(MAX_TCP_HEADER, |
3425 | sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); | |
3426 | if (unlikely(!buff)) { | |
058dc334 IJ |
3427 | inet_csk_schedule_ack(sk); |
3428 | inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; | |
3429 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, | |
3430 | TCP_DELACK_MAX, TCP_RTO_MAX); | |
3431 | return; | |
1da177e4 | 3432 | } |
058dc334 IJ |
3433 | |
3434 | /* Reserve space for headers and prepare control bits. */ | |
3435 | skb_reserve(buff, MAX_TCP_HEADER); | |
a3433f35 | 3436 | tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); |
058dc334 | 3437 | |
98781965 ED |
3438 | /* We do not want pure acks influencing TCP Small Queues or fq/pacing |
3439 | * too much. | |
3440 | * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 | |
3441 | * We also avoid tcp_wfree() overhead (cache line miss accessing | |
3442 | * tp->tsq_flags) by using regular sock_wfree() | |
3443 | */ | |
3444 | skb_set_tcp_pure_ack(buff); | |
3445 | ||
058dc334 | 3446 | /* Send it off, this clears delayed acks for us. */ |
7faee5c0 | 3447 | skb_mstamp_get(&buff->skb_mstamp); |
7450aaf6 | 3448 | tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0); |
1da177e4 | 3449 | } |
e3118e83 | 3450 | EXPORT_SYMBOL_GPL(tcp_send_ack); |
1da177e4 LT |
3451 | |
3452 | /* This routine sends a packet with an out of date sequence | |
3453 | * number. It assumes the other end will try to ack it. | |
3454 | * | |
3455 | * Question: what should we make while urgent mode? | |
3456 | * 4.4BSD forces sending single byte of data. We cannot send | |
3457 | * out of window data, because we have SND.NXT==SND.MAX... | |
3458 | * | |
3459 | * Current solution: to send TWO zero-length segments in urgent mode: | |
3460 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is | |
3461 | * out-of-date with SND.UNA-1 to probe window. | |
3462 | */ | |
e520af48 | 3463 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) |
1da177e4 LT |
3464 | { |
3465 | struct tcp_sock *tp = tcp_sk(sk); | |
3466 | struct sk_buff *skb; | |
3467 | ||
3468 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ | |
7450aaf6 ED |
3469 | skb = alloc_skb(MAX_TCP_HEADER, |
3470 | sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); | |
51456b29 | 3471 | if (!skb) |
1da177e4 LT |
3472 | return -1; |
3473 | ||
3474 | /* Reserve space for headers and set control bits. */ | |
3475 | skb_reserve(skb, MAX_TCP_HEADER); | |
1da177e4 LT |
3476 | /* Use a previous sequence. This should cause the other |
3477 | * end to send an ack. Don't queue or clone SKB, just | |
3478 | * send it. | |
3479 | */ | |
a3433f35 | 3480 | tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); |
7faee5c0 | 3481 | skb_mstamp_get(&skb->skb_mstamp); |
e2e8009f | 3482 | NET_INC_STATS(sock_net(sk), mib); |
7450aaf6 | 3483 | return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); |
1da177e4 LT |
3484 | } |
3485 | ||
ee995283 PE |
3486 | void tcp_send_window_probe(struct sock *sk) |
3487 | { | |
3488 | if (sk->sk_state == TCP_ESTABLISHED) { | |
3489 | tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; | |
e520af48 | 3490 | tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); |
ee995283 PE |
3491 | } |
3492 | } | |
3493 | ||
67edfef7 | 3494 | /* Initiate keepalive or window probe from timer. */ |
e520af48 | 3495 | int tcp_write_wakeup(struct sock *sk, int mib) |
1da177e4 | 3496 | { |
058dc334 IJ |
3497 | struct tcp_sock *tp = tcp_sk(sk); |
3498 | struct sk_buff *skb; | |
1da177e4 | 3499 | |
058dc334 IJ |
3500 | if (sk->sk_state == TCP_CLOSE) |
3501 | return -1; | |
3502 | ||
00db4124 IM |
3503 | skb = tcp_send_head(sk); |
3504 | if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { | |
058dc334 | 3505 | int err; |
0c54b85f | 3506 | unsigned int mss = tcp_current_mss(sk); |
058dc334 IJ |
3507 | unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
3508 | ||
3509 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) | |
3510 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; | |
3511 | ||
3512 | /* We are probing the opening of a window | |
3513 | * but the window size is != 0 | |
3514 | * must have been a result SWS avoidance ( sender ) | |
3515 | */ | |
3516 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || | |
3517 | skb->len > mss) { | |
3518 | seg_size = min(seg_size, mss); | |
4de075e0 | 3519 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
6cc55e09 | 3520 | if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC)) |
058dc334 IJ |
3521 | return -1; |
3522 | } else if (!tcp_skb_pcount(skb)) | |
5bbb432c | 3523 | tcp_set_skb_tso_segs(skb, mss); |
058dc334 | 3524 | |
4de075e0 | 3525 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
058dc334 IJ |
3526 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
3527 | if (!err) | |
3528 | tcp_event_new_data_sent(sk, skb); | |
3529 | return err; | |
3530 | } else { | |
33f5f57e | 3531 | if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) |
e520af48 ED |
3532 | tcp_xmit_probe_skb(sk, 1, mib); |
3533 | return tcp_xmit_probe_skb(sk, 0, mib); | |
1da177e4 | 3534 | } |
1da177e4 LT |
3535 | } |
3536 | ||
3537 | /* A window probe timeout has occurred. If window is not closed send | |
3538 | * a partial packet else a zero probe. | |
3539 | */ | |
3540 | void tcp_send_probe0(struct sock *sk) | |
3541 | { | |
463c84b9 | 3542 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 3543 | struct tcp_sock *tp = tcp_sk(sk); |
c6214a97 | 3544 | struct net *net = sock_net(sk); |
fcdd1cf4 | 3545 | unsigned long probe_max; |
1da177e4 LT |
3546 | int err; |
3547 | ||
e520af48 | 3548 | err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); |
1da177e4 | 3549 | |
fe067e8a | 3550 | if (tp->packets_out || !tcp_send_head(sk)) { |
1da177e4 | 3551 | /* Cancel probe timer, if it is not required. */ |
6687e988 | 3552 | icsk->icsk_probes_out = 0; |
463c84b9 | 3553 | icsk->icsk_backoff = 0; |
1da177e4 LT |
3554 | return; |
3555 | } | |
3556 | ||
3557 | if (err <= 0) { | |
c6214a97 | 3558 | if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) |
463c84b9 | 3559 | icsk->icsk_backoff++; |
6687e988 | 3560 | icsk->icsk_probes_out++; |
fcdd1cf4 | 3561 | probe_max = TCP_RTO_MAX; |
1da177e4 LT |
3562 | } else { |
3563 | /* If packet was not sent due to local congestion, | |
6687e988 | 3564 | * do not backoff and do not remember icsk_probes_out. |
1da177e4 LT |
3565 | * Let local senders to fight for local resources. |
3566 | * | |
3567 | * Use accumulated backoff yet. | |
3568 | */ | |
6687e988 ACM |
3569 | if (!icsk->icsk_probes_out) |
3570 | icsk->icsk_probes_out = 1; | |
fcdd1cf4 | 3571 | probe_max = TCP_RESOURCE_PROBE_INTERVAL; |
1da177e4 | 3572 | } |
fcdd1cf4 | 3573 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
21c8fe99 | 3574 | tcp_probe0_when(sk, probe_max), |
fcdd1cf4 | 3575 | TCP_RTO_MAX); |
1da177e4 | 3576 | } |
5db92c99 | 3577 | |
ea3bea3a | 3578 | int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) |
5db92c99 OP |
3579 | { |
3580 | const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; | |
3581 | struct flowi fl; | |
3582 | int res; | |
3583 | ||
58d607d3 | 3584 | tcp_rsk(req)->txhash = net_tx_rndhash(); |
b3d05147 | 3585 | res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); |
5db92c99 | 3586 | if (!res) { |
90bbcc60 | 3587 | __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); |
02a1d6e7 | 3588 | __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); |
5db92c99 OP |
3589 | } |
3590 | return res; | |
3591 | } | |
3592 | EXPORT_SYMBOL(tcp_rtx_synack); |