]>
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); |
bec41a11 | 79 | if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) |
750ea2ba | 80 | tcp_rearm_rto(sk); |
f19c29e3 | 81 | |
f7324acd DM |
82 | NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, |
83 | tcp_skb_pcount(skb)); | |
1da177e4 LT |
84 | } |
85 | ||
86 | /* SND.NXT, if window was not shrunk. | |
87 | * If window has been shrunk, what should we make? It is not clear at all. | |
88 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( | |
89 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already | |
90 | * invalid. OK, let's make this for now: | |
91 | */ | |
cf533ea5 | 92 | static inline __u32 tcp_acceptable_seq(const struct sock *sk) |
1da177e4 | 93 | { |
cf533ea5 | 94 | const struct tcp_sock *tp = tcp_sk(sk); |
9e412ba7 | 95 | |
90840def | 96 | if (!before(tcp_wnd_end(tp), tp->snd_nxt)) |
1da177e4 LT |
97 | return tp->snd_nxt; |
98 | else | |
90840def | 99 | return tcp_wnd_end(tp); |
1da177e4 LT |
100 | } |
101 | ||
102 | /* Calculate mss to advertise in SYN segment. | |
103 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: | |
104 | * | |
105 | * 1. It is independent of path mtu. | |
106 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. | |
107 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of | |
108 | * attached devices, because some buggy hosts are confused by | |
109 | * large MSS. | |
110 | * 4. We do not make 3, we advertise MSS, calculated from first | |
111 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. | |
112 | * This may be overridden via information stored in routing table. | |
113 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, | |
114 | * probably even Jumbo". | |
115 | */ | |
116 | static __u16 tcp_advertise_mss(struct sock *sk) | |
117 | { | |
118 | struct tcp_sock *tp = tcp_sk(sk); | |
cf533ea5 | 119 | const struct dst_entry *dst = __sk_dst_get(sk); |
1da177e4 LT |
120 | int mss = tp->advmss; |
121 | ||
0dbaee3b DM |
122 | if (dst) { |
123 | unsigned int metric = dst_metric_advmss(dst); | |
124 | ||
125 | if (metric < mss) { | |
126 | mss = metric; | |
127 | tp->advmss = mss; | |
128 | } | |
1da177e4 LT |
129 | } |
130 | ||
131 | return (__u16)mss; | |
132 | } | |
133 | ||
134 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". | |
6f021c62 ED |
135 | * This is the first part of cwnd validation mechanism. |
136 | */ | |
137 | void tcp_cwnd_restart(struct sock *sk, s32 delta) | |
1da177e4 | 138 | { |
463c84b9 | 139 | struct tcp_sock *tp = tcp_sk(sk); |
6f021c62 | 140 | u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); |
1da177e4 LT |
141 | u32 cwnd = tp->snd_cwnd; |
142 | ||
6687e988 | 143 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); |
1da177e4 | 144 | |
6687e988 | 145 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
1da177e4 LT |
146 | restart_cwnd = min(restart_cwnd, cwnd); |
147 | ||
463c84b9 | 148 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) |
1da177e4 LT |
149 | cwnd >>= 1; |
150 | tp->snd_cwnd = max(cwnd, restart_cwnd); | |
151 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
152 | tp->snd_cwnd_used = 0; | |
153 | } | |
154 | ||
67edfef7 | 155 | /* Congestion state accounting after a packet has been sent. */ |
40efc6fa | 156 | static void tcp_event_data_sent(struct tcp_sock *tp, |
cf533ea5 | 157 | struct sock *sk) |
1da177e4 | 158 | { |
463c84b9 ACM |
159 | struct inet_connection_sock *icsk = inet_csk(sk); |
160 | const u32 now = tcp_time_stamp; | |
1da177e4 | 161 | |
05c5a46d NC |
162 | if (tcp_packets_in_flight(tp) == 0) |
163 | tcp_ca_event(sk, CA_EVENT_TX_START); | |
164 | ||
1da177e4 LT |
165 | tp->lsndtime = now; |
166 | ||
167 | /* If it is a reply for ato after last received | |
168 | * packet, enter pingpong mode. | |
169 | */ | |
2251ae46 JM |
170 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) |
171 | icsk->icsk_ack.pingpong = 1; | |
1da177e4 LT |
172 | } |
173 | ||
67edfef7 | 174 | /* Account for an ACK we sent. */ |
40efc6fa | 175 | static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts) |
1da177e4 | 176 | { |
463c84b9 ACM |
177 | tcp_dec_quickack_mode(sk, pkts); |
178 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); | |
1da177e4 LT |
179 | } |
180 | ||
85f16525 YC |
181 | |
182 | u32 tcp_default_init_rwnd(u32 mss) | |
183 | { | |
184 | /* Initial receive window should be twice of TCP_INIT_CWND to | |
9ef71e0c | 185 | * enable proper sending of new unsent data during fast recovery |
85f16525 YC |
186 | * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a |
187 | * limit when mss is larger than 1460. | |
188 | */ | |
189 | u32 init_rwnd = TCP_INIT_CWND * 2; | |
190 | ||
191 | if (mss > 1460) | |
192 | init_rwnd = max((1460 * init_rwnd) / mss, 2U); | |
193 | return init_rwnd; | |
194 | } | |
195 | ||
1da177e4 LT |
196 | /* Determine a window scaling and initial window to offer. |
197 | * Based on the assumption that the given amount of space | |
198 | * will be offered. Store the results in the tp structure. | |
199 | * NOTE: for smooth operation initial space offering should | |
200 | * be a multiple of mss if possible. We assume here that mss >= 1. | |
201 | * This MUST be enforced by all callers. | |
202 | */ | |
203 | void tcp_select_initial_window(int __space, __u32 mss, | |
204 | __u32 *rcv_wnd, __u32 *window_clamp, | |
31d12926 | 205 | int wscale_ok, __u8 *rcv_wscale, |
206 | __u32 init_rcv_wnd) | |
1da177e4 LT |
207 | { |
208 | unsigned int space = (__space < 0 ? 0 : __space); | |
209 | ||
210 | /* If no clamp set the clamp to the max possible scaled window */ | |
211 | if (*window_clamp == 0) | |
212 | (*window_clamp) = (65535 << 14); | |
213 | space = min(*window_clamp, space); | |
214 | ||
215 | /* Quantize space offering to a multiple of mss if possible. */ | |
216 | if (space > mss) | |
217 | space = (space / mss) * mss; | |
218 | ||
219 | /* NOTE: offering an initial window larger than 32767 | |
15d99e02 RJ |
220 | * will break some buggy TCP stacks. If the admin tells us |
221 | * it is likely we could be speaking with such a buggy stack | |
222 | * we will truncate our initial window offering to 32K-1 | |
223 | * unless the remote has sent us a window scaling option, | |
224 | * which we interpret as a sign the remote TCP is not | |
225 | * misinterpreting the window field as a signed quantity. | |
1da177e4 | 226 | */ |
15d99e02 RJ |
227 | if (sysctl_tcp_workaround_signed_windows) |
228 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); | |
229 | else | |
230 | (*rcv_wnd) = space; | |
231 | ||
1da177e4 LT |
232 | (*rcv_wscale) = 0; |
233 | if (wscale_ok) { | |
234 | /* Set window scaling on max possible window | |
e905a9ed | 235 | * See RFC1323 for an explanation of the limit to 14 |
1da177e4 | 236 | */ |
f626300a SHY |
237 | space = max_t(u32, space, sysctl_tcp_rmem[2]); |
238 | space = max_t(u32, space, sysctl_rmem_max); | |
316c1592 | 239 | space = min_t(u32, space, *window_clamp); |
1da177e4 LT |
240 | while (space > 65535 && (*rcv_wscale) < 14) { |
241 | space >>= 1; | |
242 | (*rcv_wscale)++; | |
243 | } | |
244 | } | |
245 | ||
056834d9 | 246 | if (mss > (1 << *rcv_wscale)) { |
85f16525 YC |
247 | if (!init_rcv_wnd) /* Use default unless specified otherwise */ |
248 | init_rcv_wnd = tcp_default_init_rwnd(mss); | |
249 | *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); | |
1da177e4 LT |
250 | } |
251 | ||
252 | /* Set the clamp no higher than max representable value */ | |
253 | (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp); | |
254 | } | |
4bc2f18b | 255 | EXPORT_SYMBOL(tcp_select_initial_window); |
1da177e4 LT |
256 | |
257 | /* Chose a new window to advertise, update state in tcp_sock for the | |
258 | * socket, and return result with RFC1323 scaling applied. The return | |
259 | * value can be stuffed directly into th->window for an outgoing | |
260 | * frame. | |
261 | */ | |
40efc6fa | 262 | static u16 tcp_select_window(struct sock *sk) |
1da177e4 LT |
263 | { |
264 | struct tcp_sock *tp = tcp_sk(sk); | |
8e165e20 | 265 | u32 old_win = tp->rcv_wnd; |
1da177e4 LT |
266 | u32 cur_win = tcp_receive_window(tp); |
267 | u32 new_win = __tcp_select_window(sk); | |
268 | ||
269 | /* Never shrink the offered window */ | |
2de979bd | 270 | if (new_win < cur_win) { |
1da177e4 LT |
271 | /* Danger Will Robinson! |
272 | * Don't update rcv_wup/rcv_wnd here or else | |
273 | * we will not be able to advertise a zero | |
274 | * window in time. --DaveM | |
275 | * | |
276 | * Relax Will Robinson. | |
277 | */ | |
8e165e20 FW |
278 | if (new_win == 0) |
279 | NET_INC_STATS(sock_net(sk), | |
280 | LINUX_MIB_TCPWANTZEROWINDOWADV); | |
607bfbf2 | 281 | new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); |
1da177e4 LT |
282 | } |
283 | tp->rcv_wnd = new_win; | |
284 | tp->rcv_wup = tp->rcv_nxt; | |
285 | ||
286 | /* Make sure we do not exceed the maximum possible | |
287 | * scaled window. | |
288 | */ | |
15d99e02 | 289 | if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows) |
1da177e4 LT |
290 | new_win = min(new_win, MAX_TCP_WINDOW); |
291 | else | |
292 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); | |
293 | ||
294 | /* RFC1323 scaling applied */ | |
295 | new_win >>= tp->rx_opt.rcv_wscale; | |
296 | ||
297 | /* If we advertise zero window, disable fast path. */ | |
8e165e20 | 298 | if (new_win == 0) { |
1da177e4 | 299 | tp->pred_flags = 0; |
8e165e20 FW |
300 | if (old_win) |
301 | NET_INC_STATS(sock_net(sk), | |
302 | LINUX_MIB_TCPTOZEROWINDOWADV); | |
303 | } else if (old_win == 0) { | |
304 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); | |
305 | } | |
1da177e4 LT |
306 | |
307 | return new_win; | |
308 | } | |
309 | ||
67edfef7 | 310 | /* Packet ECN state for a SYN-ACK */ |
735d3831 | 311 | static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) |
bdf1ee5d | 312 | { |
30e502a3 DB |
313 | const struct tcp_sock *tp = tcp_sk(sk); |
314 | ||
4de075e0 | 315 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; |
056834d9 | 316 | if (!(tp->ecn_flags & TCP_ECN_OK)) |
4de075e0 | 317 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; |
30e502a3 DB |
318 | else if (tcp_ca_needs_ecn(sk)) |
319 | INET_ECN_xmit(sk); | |
bdf1ee5d IJ |
320 | } |
321 | ||
67edfef7 | 322 | /* Packet ECN state for a SYN. */ |
735d3831 | 323 | static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) |
bdf1ee5d IJ |
324 | { |
325 | struct tcp_sock *tp = tcp_sk(sk); | |
f7b3bec6 FW |
326 | bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || |
327 | tcp_ca_needs_ecn(sk); | |
328 | ||
329 | if (!use_ecn) { | |
330 | const struct dst_entry *dst = __sk_dst_get(sk); | |
331 | ||
332 | if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) | |
333 | use_ecn = true; | |
334 | } | |
bdf1ee5d IJ |
335 | |
336 | tp->ecn_flags = 0; | |
f7b3bec6 FW |
337 | |
338 | if (use_ecn) { | |
4de075e0 | 339 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; |
bdf1ee5d | 340 | tp->ecn_flags = TCP_ECN_OK; |
30e502a3 DB |
341 | if (tcp_ca_needs_ecn(sk)) |
342 | INET_ECN_xmit(sk); | |
bdf1ee5d IJ |
343 | } |
344 | } | |
345 | ||
49213555 DB |
346 | static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) |
347 | { | |
348 | if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) | |
349 | /* tp->ecn_flags are cleared at a later point in time when | |
350 | * SYN ACK is ultimatively being received. | |
351 | */ | |
352 | TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); | |
353 | } | |
354 | ||
735d3831 | 355 | static void |
6ac705b1 | 356 | tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) |
bdf1ee5d | 357 | { |
6ac705b1 | 358 | if (inet_rsk(req)->ecn_ok) |
bdf1ee5d IJ |
359 | th->ece = 1; |
360 | } | |
361 | ||
67edfef7 AK |
362 | /* Set up ECN state for a packet on a ESTABLISHED socket that is about to |
363 | * be sent. | |
364 | */ | |
735d3831 | 365 | static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, |
ea1627c2 | 366 | struct tcphdr *th, int tcp_header_len) |
bdf1ee5d IJ |
367 | { |
368 | struct tcp_sock *tp = tcp_sk(sk); | |
369 | ||
370 | if (tp->ecn_flags & TCP_ECN_OK) { | |
371 | /* Not-retransmitted data segment: set ECT and inject CWR. */ | |
372 | if (skb->len != tcp_header_len && | |
373 | !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { | |
374 | INET_ECN_xmit(sk); | |
056834d9 | 375 | if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { |
bdf1ee5d | 376 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; |
ea1627c2 | 377 | th->cwr = 1; |
bdf1ee5d IJ |
378 | skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; |
379 | } | |
30e502a3 | 380 | } else if (!tcp_ca_needs_ecn(sk)) { |
bdf1ee5d IJ |
381 | /* ACK or retransmitted segment: clear ECT|CE */ |
382 | INET_ECN_dontxmit(sk); | |
383 | } | |
384 | if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) | |
ea1627c2 | 385 | th->ece = 1; |
bdf1ee5d IJ |
386 | } |
387 | } | |
388 | ||
e870a8ef IJ |
389 | /* Constructs common control bits of non-data skb. If SYN/FIN is present, |
390 | * auto increment end seqno. | |
391 | */ | |
392 | static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) | |
393 | { | |
2e8e18ef | 394 | skb->ip_summed = CHECKSUM_PARTIAL; |
e870a8ef IJ |
395 | skb->csum = 0; |
396 | ||
4de075e0 | 397 | TCP_SKB_CB(skb)->tcp_flags = flags; |
e870a8ef IJ |
398 | TCP_SKB_CB(skb)->sacked = 0; |
399 | ||
cd7d8498 | 400 | tcp_skb_pcount_set(skb, 1); |
e870a8ef IJ |
401 | |
402 | TCP_SKB_CB(skb)->seq = seq; | |
a3433f35 | 403 | if (flags & (TCPHDR_SYN | TCPHDR_FIN)) |
e870a8ef IJ |
404 | seq++; |
405 | TCP_SKB_CB(skb)->end_seq = seq; | |
406 | } | |
407 | ||
a2a385d6 | 408 | static inline bool tcp_urg_mode(const struct tcp_sock *tp) |
33f5f57e IJ |
409 | { |
410 | return tp->snd_una != tp->snd_up; | |
411 | } | |
412 | ||
33ad798c AL |
413 | #define OPTION_SACK_ADVERTISE (1 << 0) |
414 | #define OPTION_TS (1 << 1) | |
415 | #define OPTION_MD5 (1 << 2) | |
89e95a61 | 416 | #define OPTION_WSCALE (1 << 3) |
2100c8d2 | 417 | #define OPTION_FAST_OPEN_COOKIE (1 << 8) |
33ad798c AL |
418 | |
419 | struct tcp_out_options { | |
2100c8d2 YC |
420 | u16 options; /* bit field of OPTION_* */ |
421 | u16 mss; /* 0 to disable */ | |
33ad798c AL |
422 | u8 ws; /* window scale, 0 to disable */ |
423 | u8 num_sack_blocks; /* number of SACK blocks to include */ | |
bd0388ae | 424 | u8 hash_size; /* bytes in hash_location */ |
bd0388ae | 425 | __u8 *hash_location; /* temporary pointer, overloaded */ |
2100c8d2 YC |
426 | __u32 tsval, tsecr; /* need to include OPTION_TS */ |
427 | struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ | |
33ad798c AL |
428 | }; |
429 | ||
67edfef7 AK |
430 | /* Write previously computed TCP options to the packet. |
431 | * | |
432 | * Beware: Something in the Internet is very sensitive to the ordering of | |
fd6149d3 IJ |
433 | * TCP options, we learned this through the hard way, so be careful here. |
434 | * Luckily we can at least blame others for their non-compliance but from | |
8e3bff96 | 435 | * inter-operability perspective it seems that we're somewhat stuck with |
fd6149d3 IJ |
436 | * the ordering which we have been using if we want to keep working with |
437 | * those broken things (not that it currently hurts anybody as there isn't | |
438 | * particular reason why the ordering would need to be changed). | |
439 | * | |
440 | * At least SACK_PERM as the first option is known to lead to a disaster | |
441 | * (but it may well be that other scenarios fail similarly). | |
442 | */ | |
33ad798c | 443 | static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, |
bd0388ae WAS |
444 | struct tcp_out_options *opts) |
445 | { | |
2100c8d2 | 446 | u16 options = opts->options; /* mungable copy */ |
bd0388ae | 447 | |
bd0388ae | 448 | if (unlikely(OPTION_MD5 & options)) { |
1a2c6181 CP |
449 | *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | |
450 | (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); | |
bd0388ae WAS |
451 | /* overload cookie hash location */ |
452 | opts->hash_location = (__u8 *)ptr; | |
33ad798c | 453 | ptr += 4; |
40efc6fa | 454 | } |
33ad798c | 455 | |
fd6149d3 IJ |
456 | if (unlikely(opts->mss)) { |
457 | *ptr++ = htonl((TCPOPT_MSS << 24) | | |
458 | (TCPOLEN_MSS << 16) | | |
459 | opts->mss); | |
460 | } | |
461 | ||
bd0388ae WAS |
462 | if (likely(OPTION_TS & options)) { |
463 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { | |
33ad798c AL |
464 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
465 | (TCPOLEN_SACK_PERM << 16) | | |
466 | (TCPOPT_TIMESTAMP << 8) | | |
467 | TCPOLEN_TIMESTAMP); | |
bd0388ae | 468 | options &= ~OPTION_SACK_ADVERTISE; |
33ad798c AL |
469 | } else { |
470 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
471 | (TCPOPT_NOP << 16) | | |
472 | (TCPOPT_TIMESTAMP << 8) | | |
473 | TCPOLEN_TIMESTAMP); | |
474 | } | |
475 | *ptr++ = htonl(opts->tsval); | |
476 | *ptr++ = htonl(opts->tsecr); | |
477 | } | |
478 | ||
bd0388ae | 479 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { |
33ad798c AL |
480 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
481 | (TCPOPT_NOP << 16) | | |
482 | (TCPOPT_SACK_PERM << 8) | | |
483 | TCPOLEN_SACK_PERM); | |
484 | } | |
485 | ||
bd0388ae | 486 | if (unlikely(OPTION_WSCALE & options)) { |
33ad798c AL |
487 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
488 | (TCPOPT_WINDOW << 16) | | |
489 | (TCPOLEN_WINDOW << 8) | | |
490 | opts->ws); | |
491 | } | |
492 | ||
493 | if (unlikely(opts->num_sack_blocks)) { | |
494 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? | |
495 | tp->duplicate_sack : tp->selective_acks; | |
40efc6fa SH |
496 | int this_sack; |
497 | ||
498 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
499 | (TCPOPT_NOP << 16) | | |
500 | (TCPOPT_SACK << 8) | | |
33ad798c | 501 | (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * |
40efc6fa | 502 | TCPOLEN_SACK_PERBLOCK))); |
2de979bd | 503 | |
33ad798c AL |
504 | for (this_sack = 0; this_sack < opts->num_sack_blocks; |
505 | ++this_sack) { | |
40efc6fa SH |
506 | *ptr++ = htonl(sp[this_sack].start_seq); |
507 | *ptr++ = htonl(sp[this_sack].end_seq); | |
508 | } | |
2de979bd | 509 | |
5861f8e5 | 510 | tp->rx_opt.dsack = 0; |
40efc6fa | 511 | } |
2100c8d2 YC |
512 | |
513 | if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { | |
514 | struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; | |
7f9b838b DL |
515 | u8 *p = (u8 *)ptr; |
516 | u32 len; /* Fast Open option length */ | |
517 | ||
518 | if (foc->exp) { | |
519 | len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; | |
520 | *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | | |
521 | TCPOPT_FASTOPEN_MAGIC); | |
522 | p += TCPOLEN_EXP_FASTOPEN_BASE; | |
523 | } else { | |
524 | len = TCPOLEN_FASTOPEN_BASE + foc->len; | |
525 | *p++ = TCPOPT_FASTOPEN; | |
526 | *p++ = len; | |
527 | } | |
2100c8d2 | 528 | |
7f9b838b DL |
529 | memcpy(p, foc->val, foc->len); |
530 | if ((len & 3) == 2) { | |
531 | p[foc->len] = TCPOPT_NOP; | |
532 | p[foc->len + 1] = TCPOPT_NOP; | |
2100c8d2 | 533 | } |
7f9b838b | 534 | ptr += (len + 3) >> 2; |
2100c8d2 | 535 | } |
33ad798c AL |
536 | } |
537 | ||
67edfef7 AK |
538 | /* Compute TCP options for SYN packets. This is not the final |
539 | * network wire format yet. | |
540 | */ | |
95c96174 | 541 | static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, |
33ad798c | 542 | struct tcp_out_options *opts, |
cf533ea5 ED |
543 | struct tcp_md5sig_key **md5) |
544 | { | |
33ad798c | 545 | struct tcp_sock *tp = tcp_sk(sk); |
95c96174 | 546 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
783237e8 | 547 | struct tcp_fastopen_request *fastopen = tp->fastopen_req; |
33ad798c | 548 | |
cfb6eeb4 | 549 | #ifdef CONFIG_TCP_MD5SIG |
33ad798c AL |
550 | *md5 = tp->af_specific->md5_lookup(sk, sk); |
551 | if (*md5) { | |
552 | opts->options |= OPTION_MD5; | |
bd0388ae | 553 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
cfb6eeb4 | 554 | } |
33ad798c AL |
555 | #else |
556 | *md5 = NULL; | |
cfb6eeb4 | 557 | #endif |
33ad798c AL |
558 | |
559 | /* We always get an MSS option. The option bytes which will be seen in | |
560 | * normal data packets should timestamps be used, must be in the MSS | |
561 | * advertised. But we subtract them from tp->mss_cache so that | |
562 | * calculations in tcp_sendmsg are simpler etc. So account for this | |
563 | * fact here if necessary. If we don't do this correctly, as a | |
564 | * receiver we won't recognize data packets as being full sized when we | |
565 | * should, and thus we won't abide by the delayed ACK rules correctly. | |
566 | * SACKs don't matter, we never delay an ACK when we have any of those | |
567 | * going out. */ | |
568 | opts->mss = tcp_advertise_mss(sk); | |
bd0388ae | 569 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c | 570 | |
51456b29 | 571 | if (likely(sysctl_tcp_timestamps && !*md5)) { |
33ad798c | 572 | opts->options |= OPTION_TS; |
7faee5c0 | 573 | opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; |
33ad798c | 574 | opts->tsecr = tp->rx_opt.ts_recent; |
bd0388ae | 575 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c | 576 | } |
bb5b7c11 | 577 | if (likely(sysctl_tcp_window_scaling)) { |
33ad798c | 578 | opts->ws = tp->rx_opt.rcv_wscale; |
89e95a61 | 579 | opts->options |= OPTION_WSCALE; |
bd0388ae | 580 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 581 | } |
bb5b7c11 | 582 | if (likely(sysctl_tcp_sack)) { |
33ad798c | 583 | opts->options |= OPTION_SACK_ADVERTISE; |
b32d1310 | 584 | if (unlikely(!(OPTION_TS & opts->options))) |
bd0388ae | 585 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c AL |
586 | } |
587 | ||
783237e8 | 588 | if (fastopen && fastopen->cookie.len >= 0) { |
2646c831 DL |
589 | u32 need = fastopen->cookie.len; |
590 | ||
591 | need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : | |
592 | TCPOLEN_FASTOPEN_BASE; | |
783237e8 YC |
593 | need = (need + 3) & ~3U; /* Align to 32 bits */ |
594 | if (remaining >= need) { | |
595 | opts->options |= OPTION_FAST_OPEN_COOKIE; | |
596 | opts->fastopen_cookie = &fastopen->cookie; | |
597 | remaining -= need; | |
598 | tp->syn_fastopen = 1; | |
2646c831 | 599 | tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; |
783237e8 YC |
600 | } |
601 | } | |
bd0388ae | 602 | |
bd0388ae | 603 | return MAX_TCP_OPTION_SPACE - remaining; |
40efc6fa SH |
604 | } |
605 | ||
67edfef7 | 606 | /* Set up TCP options for SYN-ACKs. */ |
37bfbdda ED |
607 | static unsigned int tcp_synack_options(struct request_sock *req, |
608 | unsigned int mss, struct sk_buff *skb, | |
609 | struct tcp_out_options *opts, | |
610 | const struct tcp_md5sig_key *md5, | |
611 | struct tcp_fastopen_cookie *foc) | |
4957faad | 612 | { |
33ad798c | 613 | struct inet_request_sock *ireq = inet_rsk(req); |
95c96174 | 614 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
33ad798c | 615 | |
cfb6eeb4 | 616 | #ifdef CONFIG_TCP_MD5SIG |
80f03e27 | 617 | if (md5) { |
33ad798c | 618 | opts->options |= OPTION_MD5; |
4957faad WAS |
619 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
620 | ||
621 | /* We can't fit any SACK blocks in a packet with MD5 + TS | |
622 | * options. There was discussion about disabling SACK | |
623 | * rather than TS in order to fit in better with old, | |
624 | * buggy kernels, but that was deemed to be unnecessary. | |
625 | */ | |
de213e5e | 626 | ireq->tstamp_ok &= !ireq->sack_ok; |
cfb6eeb4 YH |
627 | } |
628 | #endif | |
33ad798c | 629 | |
4957faad | 630 | /* We always send an MSS option. */ |
33ad798c | 631 | opts->mss = mss; |
4957faad | 632 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c AL |
633 | |
634 | if (likely(ireq->wscale_ok)) { | |
635 | opts->ws = ireq->rcv_wscale; | |
89e95a61 | 636 | opts->options |= OPTION_WSCALE; |
4957faad | 637 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 638 | } |
de213e5e | 639 | if (likely(ireq->tstamp_ok)) { |
33ad798c | 640 | opts->options |= OPTION_TS; |
95a22cae | 641 | opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; |
33ad798c | 642 | opts->tsecr = req->ts_recent; |
4957faad | 643 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c AL |
644 | } |
645 | if (likely(ireq->sack_ok)) { | |
646 | opts->options |= OPTION_SACK_ADVERTISE; | |
de213e5e | 647 | if (unlikely(!ireq->tstamp_ok)) |
4957faad | 648 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c | 649 | } |
7f9b838b DL |
650 | if (foc != NULL && foc->len >= 0) { |
651 | u32 need = foc->len; | |
652 | ||
653 | need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : | |
654 | TCPOLEN_FASTOPEN_BASE; | |
8336886f JC |
655 | need = (need + 3) & ~3U; /* Align to 32 bits */ |
656 | if (remaining >= need) { | |
657 | opts->options |= OPTION_FAST_OPEN_COOKIE; | |
658 | opts->fastopen_cookie = foc; | |
659 | remaining -= need; | |
660 | } | |
661 | } | |
1a2c6181 | 662 | |
4957faad | 663 | return MAX_TCP_OPTION_SPACE - remaining; |
33ad798c AL |
664 | } |
665 | ||
67edfef7 AK |
666 | /* Compute TCP options for ESTABLISHED sockets. This is not the |
667 | * final wire format yet. | |
668 | */ | |
95c96174 | 669 | static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, |
33ad798c | 670 | struct tcp_out_options *opts, |
cf533ea5 ED |
671 | struct tcp_md5sig_key **md5) |
672 | { | |
33ad798c | 673 | struct tcp_sock *tp = tcp_sk(sk); |
95c96174 | 674 | unsigned int size = 0; |
cabeccbd | 675 | unsigned int eff_sacks; |
33ad798c | 676 | |
5843ef42 AK |
677 | opts->options = 0; |
678 | ||
33ad798c AL |
679 | #ifdef CONFIG_TCP_MD5SIG |
680 | *md5 = tp->af_specific->md5_lookup(sk, sk); | |
681 | if (unlikely(*md5)) { | |
682 | opts->options |= OPTION_MD5; | |
683 | size += TCPOLEN_MD5SIG_ALIGNED; | |
684 | } | |
685 | #else | |
686 | *md5 = NULL; | |
687 | #endif | |
688 | ||
689 | if (likely(tp->rx_opt.tstamp_ok)) { | |
690 | opts->options |= OPTION_TS; | |
7faee5c0 | 691 | opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; |
33ad798c AL |
692 | opts->tsecr = tp->rx_opt.ts_recent; |
693 | size += TCPOLEN_TSTAMP_ALIGNED; | |
694 | } | |
695 | ||
cabeccbd IJ |
696 | eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; |
697 | if (unlikely(eff_sacks)) { | |
95c96174 | 698 | const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; |
33ad798c | 699 | opts->num_sack_blocks = |
95c96174 | 700 | min_t(unsigned int, eff_sacks, |
33ad798c AL |
701 | (remaining - TCPOLEN_SACK_BASE_ALIGNED) / |
702 | TCPOLEN_SACK_PERBLOCK); | |
703 | size += TCPOLEN_SACK_BASE_ALIGNED + | |
704 | opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; | |
705 | } | |
706 | ||
707 | return size; | |
40efc6fa | 708 | } |
1da177e4 | 709 | |
46d3ceab ED |
710 | |
711 | /* TCP SMALL QUEUES (TSQ) | |
712 | * | |
713 | * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) | |
714 | * to reduce RTT and bufferbloat. | |
715 | * We do this using a special skb destructor (tcp_wfree). | |
716 | * | |
717 | * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb | |
718 | * needs to be reallocated in a driver. | |
8e3bff96 | 719 | * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc |
46d3ceab ED |
720 | * |
721 | * Since transmit from skb destructor is forbidden, we use a tasklet | |
722 | * to process all sockets that eventually need to send more skbs. | |
723 | * We use one tasklet per cpu, with its own queue of sockets. | |
724 | */ | |
725 | struct tsq_tasklet { | |
726 | struct tasklet_struct tasklet; | |
727 | struct list_head head; /* queue of tcp sockets */ | |
728 | }; | |
729 | static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); | |
730 | ||
6f458dfb ED |
731 | static void tcp_tsq_handler(struct sock *sk) |
732 | { | |
733 | if ((1 << sk->sk_state) & | |
734 | (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | | |
f9616c35 ED |
735 | TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { |
736 | struct tcp_sock *tp = tcp_sk(sk); | |
737 | ||
738 | if (tp->lost_out > tp->retrans_out && | |
739 | tp->snd_cwnd > tcp_packets_in_flight(tp)) | |
740 | tcp_xmit_retransmit_queue(sk); | |
741 | ||
742 | tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, | |
bf06200e | 743 | 0, GFP_ATOMIC); |
f9616c35 | 744 | } |
6f458dfb | 745 | } |
46d3ceab | 746 | /* |
8e3bff96 | 747 | * One tasklet per cpu tries to send more skbs. |
46d3ceab | 748 | * We run in tasklet context but need to disable irqs when |
8e3bff96 | 749 | * transferring tsq->head because tcp_wfree() might |
46d3ceab ED |
750 | * interrupt us (non NAPI drivers) |
751 | */ | |
752 | static void tcp_tasklet_func(unsigned long data) | |
753 | { | |
754 | struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; | |
755 | LIST_HEAD(list); | |
756 | unsigned long flags; | |
757 | struct list_head *q, *n; | |
758 | struct tcp_sock *tp; | |
759 | struct sock *sk; | |
760 | ||
761 | local_irq_save(flags); | |
762 | list_splice_init(&tsq->head, &list); | |
763 | local_irq_restore(flags); | |
764 | ||
765 | list_for_each_safe(q, n, &list) { | |
766 | tp = list_entry(q, struct tcp_sock, tsq_node); | |
767 | list_del(&tp->tsq_node); | |
768 | ||
769 | sk = (struct sock *)tp; | |
0a9648f1 | 770 | smp_mb__before_atomic(); |
7aa5470c ED |
771 | clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); |
772 | ||
b223feb9 | 773 | if (!sk->sk_lock.owned && |
7aa5470c | 774 | test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) { |
b223feb9 ED |
775 | bh_lock_sock(sk); |
776 | if (!sock_owned_by_user(sk)) { | |
7aa5470c | 777 | clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); |
b223feb9 ED |
778 | tcp_tsq_handler(sk); |
779 | } | |
780 | bh_unlock_sock(sk); | |
46d3ceab | 781 | } |
46d3ceab | 782 | |
46d3ceab ED |
783 | sk_free(sk); |
784 | } | |
785 | } | |
786 | ||
40fc3423 ED |
787 | #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ |
788 | TCPF_WRITE_TIMER_DEFERRED | \ | |
789 | TCPF_DELACK_TIMER_DEFERRED | \ | |
790 | TCPF_MTU_REDUCED_DEFERRED) | |
46d3ceab ED |
791 | /** |
792 | * tcp_release_cb - tcp release_sock() callback | |
793 | * @sk: socket | |
794 | * | |
795 | * called from release_sock() to perform protocol dependent | |
796 | * actions before socket release. | |
797 | */ | |
798 | void tcp_release_cb(struct sock *sk) | |
799 | { | |
6f458dfb | 800 | unsigned long flags, nflags; |
46d3ceab | 801 | |
6f458dfb ED |
802 | /* perform an atomic operation only if at least one flag is set */ |
803 | do { | |
7aa5470c | 804 | flags = sk->sk_tsq_flags; |
6f458dfb ED |
805 | if (!(flags & TCP_DEFERRED_ALL)) |
806 | return; | |
807 | nflags = flags & ~TCP_DEFERRED_ALL; | |
7aa5470c | 808 | } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); |
6f458dfb | 809 | |
40fc3423 | 810 | if (flags & TCPF_TSQ_DEFERRED) |
6f458dfb ED |
811 | tcp_tsq_handler(sk); |
812 | ||
c3f9b018 ED |
813 | /* Here begins the tricky part : |
814 | * We are called from release_sock() with : | |
815 | * 1) BH disabled | |
816 | * 2) sk_lock.slock spinlock held | |
817 | * 3) socket owned by us (sk->sk_lock.owned == 1) | |
818 | * | |
819 | * But following code is meant to be called from BH handlers, | |
820 | * so we should keep BH disabled, but early release socket ownership | |
821 | */ | |
822 | sock_release_ownership(sk); | |
823 | ||
40fc3423 | 824 | if (flags & TCPF_WRITE_TIMER_DEFERRED) { |
6f458dfb | 825 | tcp_write_timer_handler(sk); |
144d56e9 ED |
826 | __sock_put(sk); |
827 | } | |
40fc3423 | 828 | if (flags & TCPF_DELACK_TIMER_DEFERRED) { |
6f458dfb | 829 | tcp_delack_timer_handler(sk); |
144d56e9 ED |
830 | __sock_put(sk); |
831 | } | |
40fc3423 | 832 | if (flags & TCPF_MTU_REDUCED_DEFERRED) { |
4fab9071 | 833 | inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); |
144d56e9 ED |
834 | __sock_put(sk); |
835 | } | |
46d3ceab ED |
836 | } |
837 | EXPORT_SYMBOL(tcp_release_cb); | |
838 | ||
839 | void __init tcp_tasklet_init(void) | |
840 | { | |
841 | int i; | |
842 | ||
843 | for_each_possible_cpu(i) { | |
844 | struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); | |
845 | ||
846 | INIT_LIST_HEAD(&tsq->head); | |
847 | tasklet_init(&tsq->tasklet, | |
848 | tcp_tasklet_func, | |
849 | (unsigned long)tsq); | |
850 | } | |
851 | } | |
852 | ||
853 | /* | |
854 | * Write buffer destructor automatically called from kfree_skb. | |
8e3bff96 | 855 | * We can't xmit new skbs from this context, as we might already |
46d3ceab ED |
856 | * hold qdisc lock. |
857 | */ | |
d6a4a104 | 858 | void tcp_wfree(struct sk_buff *skb) |
46d3ceab ED |
859 | { |
860 | struct sock *sk = skb->sk; | |
861 | struct tcp_sock *tp = tcp_sk(sk); | |
408f0a6c | 862 | unsigned long flags, nval, oval; |
9b462d02 ED |
863 | int wmem; |
864 | ||
865 | /* Keep one reference on sk_wmem_alloc. | |
866 | * Will be released by sk_free() from here or tcp_tasklet_func() | |
867 | */ | |
868 | wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc); | |
869 | ||
870 | /* If this softirq is serviced by ksoftirqd, we are likely under stress. | |
871 | * Wait until our queues (qdisc + devices) are drained. | |
872 | * This gives : | |
873 | * - less callbacks to tcp_write_xmit(), reducing stress (batches) | |
874 | * - chance for incoming ACK (processed by another cpu maybe) | |
875 | * to migrate this flow (skb->ooo_okay will be eventually set) | |
876 | */ | |
877 | if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) | |
878 | goto out; | |
46d3ceab | 879 | |
7aa5470c | 880 | for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { |
46d3ceab | 881 | struct tsq_tasklet *tsq; |
a9b204d1 | 882 | bool empty; |
46d3ceab | 883 | |
408f0a6c ED |
884 | if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) |
885 | goto out; | |
886 | ||
b223feb9 | 887 | nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED; |
7aa5470c | 888 | nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); |
408f0a6c ED |
889 | if (nval != oval) |
890 | continue; | |
891 | ||
46d3ceab ED |
892 | /* queue this socket to tasklet queue */ |
893 | local_irq_save(flags); | |
903ceff7 | 894 | tsq = this_cpu_ptr(&tsq_tasklet); |
a9b204d1 | 895 | empty = list_empty(&tsq->head); |
46d3ceab | 896 | list_add(&tp->tsq_node, &tsq->head); |
a9b204d1 ED |
897 | if (empty) |
898 | tasklet_schedule(&tsq->tasklet); | |
46d3ceab | 899 | local_irq_restore(flags); |
9b462d02 | 900 | return; |
46d3ceab | 901 | } |
9b462d02 ED |
902 | out: |
903 | sk_free(sk); | |
46d3ceab ED |
904 | } |
905 | ||
1da177e4 LT |
906 | /* This routine actually transmits TCP packets queued in by |
907 | * tcp_do_sendmsg(). This is used by both the initial | |
908 | * transmission and possible later retransmissions. | |
909 | * All SKB's seen here are completely headerless. It is our | |
910 | * job to build the TCP header, and pass the packet down to | |
911 | * IP so it can do the same plus pass the packet off to the | |
912 | * device. | |
913 | * | |
914 | * We are working here with either a clone of the original | |
915 | * SKB, or a fresh unique copy made by the retransmit engine. | |
916 | */ | |
056834d9 IJ |
917 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, |
918 | gfp_t gfp_mask) | |
1da177e4 | 919 | { |
dfb4b9dc DM |
920 | const struct inet_connection_sock *icsk = inet_csk(sk); |
921 | struct inet_sock *inet; | |
922 | struct tcp_sock *tp; | |
923 | struct tcp_skb_cb *tcb; | |
33ad798c | 924 | struct tcp_out_options opts; |
95c96174 | 925 | unsigned int tcp_options_size, tcp_header_size; |
cfb6eeb4 | 926 | struct tcp_md5sig_key *md5; |
dfb4b9dc | 927 | struct tcphdr *th; |
dfb4b9dc DM |
928 | int err; |
929 | ||
930 | BUG_ON(!skb || !tcp_skb_pcount(skb)); | |
6f094b9e | 931 | tp = tcp_sk(sk); |
dfb4b9dc | 932 | |
ccdbb6e9 | 933 | if (clone_it) { |
740b0f18 | 934 | skb_mstamp_get(&skb->skb_mstamp); |
6f094b9e LB |
935 | TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq |
936 | - tp->snd_una; | |
b9f64820 | 937 | tcp_rate_skb_sent(sk, skb); |
ccdbb6e9 | 938 | |
dfb4b9dc DM |
939 | if (unlikely(skb_cloned(skb))) |
940 | skb = pskb_copy(skb, gfp_mask); | |
941 | else | |
942 | skb = skb_clone(skb, gfp_mask); | |
943 | if (unlikely(!skb)) | |
944 | return -ENOBUFS; | |
945 | } | |
1da177e4 | 946 | |
dfb4b9dc | 947 | inet = inet_sk(sk); |
dfb4b9dc | 948 | tcb = TCP_SKB_CB(skb); |
33ad798c | 949 | memset(&opts, 0, sizeof(opts)); |
1da177e4 | 950 | |
4de075e0 | 951 | if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) |
33ad798c AL |
952 | tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); |
953 | else | |
954 | tcp_options_size = tcp_established_options(sk, skb, &opts, | |
955 | &md5); | |
956 | tcp_header_size = tcp_options_size + sizeof(struct tcphdr); | |
e905a9ed | 957 | |
547669d4 | 958 | /* if no packet is in qdisc/device queue, then allow XPS to select |
b2532eb9 ED |
959 | * another queue. We can be called from tcp_tsq_handler() |
960 | * which holds one reference to sk_wmem_alloc. | |
961 | * | |
962 | * TODO: Ideally, in-flight pure ACK packets should not matter here. | |
963 | * One way to get this would be to set skb->truesize = 2 on them. | |
547669d4 | 964 | */ |
b2532eb9 | 965 | skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); |
dfb4b9dc | 966 | |
aa8223c7 ACM |
967 | skb_push(skb, tcp_header_size); |
968 | skb_reset_transport_header(skb); | |
46d3ceab ED |
969 | |
970 | skb_orphan(skb); | |
971 | skb->sk = sk; | |
1d2077ac | 972 | skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; |
b73c3d0e | 973 | skb_set_hash_from_sk(skb, sk); |
46d3ceab | 974 | atomic_add(skb->truesize, &sk->sk_wmem_alloc); |
dfb4b9dc DM |
975 | |
976 | /* Build TCP header and checksum it. */ | |
ea1627c2 | 977 | th = (struct tcphdr *)skb->data; |
c720c7e8 ED |
978 | th->source = inet->inet_sport; |
979 | th->dest = inet->inet_dport; | |
dfb4b9dc DM |
980 | th->seq = htonl(tcb->seq); |
981 | th->ack_seq = htonl(tp->rcv_nxt); | |
df7a3b07 | 982 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
4de075e0 | 983 | tcb->tcp_flags); |
dfb4b9dc | 984 | |
dfb4b9dc DM |
985 | th->check = 0; |
986 | th->urg_ptr = 0; | |
1da177e4 | 987 | |
33f5f57e | 988 | /* The urg_mode check is necessary during a below snd_una win probe */ |
7691367d HX |
989 | if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { |
990 | if (before(tp->snd_up, tcb->seq + 0x10000)) { | |
991 | th->urg_ptr = htons(tp->snd_up - tcb->seq); | |
992 | th->urg = 1; | |
993 | } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { | |
0eae88f3 | 994 | th->urg_ptr = htons(0xFFFF); |
7691367d HX |
995 | th->urg = 1; |
996 | } | |
dfb4b9dc | 997 | } |
1da177e4 | 998 | |
bd0388ae | 999 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
51466a75 | 1000 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
ea1627c2 ED |
1001 | if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { |
1002 | th->window = htons(tcp_select_window(sk)); | |
1003 | tcp_ecn_send(sk, skb, th, tcp_header_size); | |
1004 | } else { | |
1005 | /* RFC1323: The window in SYN & SYN/ACK segments | |
1006 | * is never scaled. | |
1007 | */ | |
1008 | th->window = htons(min(tp->rcv_wnd, 65535U)); | |
1009 | } | |
cfb6eeb4 YH |
1010 | #ifdef CONFIG_TCP_MD5SIG |
1011 | /* Calculate the MD5 hash, as we have all we need now */ | |
1012 | if (md5) { | |
a465419b | 1013 | sk_nocaps_add(sk, NETIF_F_GSO_MASK); |
bd0388ae | 1014 | tp->af_specific->calc_md5_hash(opts.hash_location, |
39f8e58e | 1015 | md5, sk, skb); |
cfb6eeb4 YH |
1016 | } |
1017 | #endif | |
1018 | ||
bb296246 | 1019 | icsk->icsk_af_ops->send_check(sk, skb); |
1da177e4 | 1020 | |
4de075e0 | 1021 | if (likely(tcb->tcp_flags & TCPHDR_ACK)) |
dfb4b9dc | 1022 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb)); |
1da177e4 | 1023 | |
a44d6eac | 1024 | if (skb->len != tcp_header_size) { |
cf533ea5 | 1025 | tcp_event_data_sent(tp, sk); |
a44d6eac MKL |
1026 | tp->data_segs_out += tcp_skb_pcount(skb); |
1027 | } | |
1da177e4 | 1028 | |
bd37a088 | 1029 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
aa2ea058 TH |
1030 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, |
1031 | tcp_skb_pcount(skb)); | |
1da177e4 | 1032 | |
2efd055c | 1033 | tp->segs_out += tcp_skb_pcount(skb); |
f69ad292 | 1034 | /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ |
cd7d8498 | 1035 | skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); |
f69ad292 | 1036 | skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); |
cd7d8498 | 1037 | |
7faee5c0 | 1038 | /* Our usage of tstamp should remain private */ |
2456e855 | 1039 | skb->tstamp = 0; |
971f10ec ED |
1040 | |
1041 | /* Cleanup our debris for IP stacks */ | |
1042 | memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), | |
1043 | sizeof(struct inet6_skb_parm))); | |
1044 | ||
b0270e91 | 1045 | err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl); |
7faee5c0 | 1046 | |
83de47cd | 1047 | if (likely(err <= 0)) |
dfb4b9dc DM |
1048 | return err; |
1049 | ||
5ee2c941 | 1050 | tcp_enter_cwr(sk); |
dfb4b9dc | 1051 | |
b9df3cb8 | 1052 | return net_xmit_eval(err); |
1da177e4 LT |
1053 | } |
1054 | ||
67edfef7 | 1055 | /* This routine just queues the buffer for sending. |
1da177e4 LT |
1056 | * |
1057 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, | |
1058 | * otherwise socket can stall. | |
1059 | */ | |
1060 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) | |
1061 | { | |
1062 | struct tcp_sock *tp = tcp_sk(sk); | |
1063 | ||
1064 | /* Advance write_seq and place onto the write_queue. */ | |
1065 | tp->write_seq = TCP_SKB_CB(skb)->end_seq; | |
f4a775d1 | 1066 | __skb_header_release(skb); |
fe067e8a | 1067 | tcp_add_write_queue_tail(sk, skb); |
3ab224be HA |
1068 | sk->sk_wmem_queued += skb->truesize; |
1069 | sk_mem_charge(sk, skb->truesize); | |
1da177e4 LT |
1070 | } |
1071 | ||
67edfef7 | 1072 | /* Initialize TSO segments for a packet. */ |
5bbb432c | 1073 | static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) |
f6302d1d | 1074 | { |
8f26fb1c | 1075 | if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) { |
f6302d1d DM |
1076 | /* Avoid the costly divide in the normal |
1077 | * non-TSO case. | |
1078 | */ | |
cd7d8498 | 1079 | tcp_skb_pcount_set(skb, 1); |
f69ad292 | 1080 | TCP_SKB_CB(skb)->tcp_gso_size = 0; |
f6302d1d | 1081 | } else { |
cd7d8498 | 1082 | tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); |
f69ad292 | 1083 | TCP_SKB_CB(skb)->tcp_gso_size = mss_now; |
1da177e4 LT |
1084 | } |
1085 | } | |
1086 | ||
91fed7a1 | 1087 | /* When a modification to fackets out becomes necessary, we need to check |
68f8353b | 1088 | * skb is counted to fackets_out or not. |
91fed7a1 | 1089 | */ |
cf533ea5 | 1090 | static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb, |
91fed7a1 IJ |
1091 | int decr) |
1092 | { | |
a47e5a98 IJ |
1093 | struct tcp_sock *tp = tcp_sk(sk); |
1094 | ||
dc86967b | 1095 | if (!tp->sacked_out || tcp_is_reno(tp)) |
91fed7a1 IJ |
1096 | return; |
1097 | ||
6859d494 | 1098 | if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq)) |
91fed7a1 | 1099 | tp->fackets_out -= decr; |
91fed7a1 IJ |
1100 | } |
1101 | ||
797108d1 IJ |
1102 | /* Pcount in the middle of the write queue got changed, we need to do various |
1103 | * tweaks to fix counters | |
1104 | */ | |
cf533ea5 | 1105 | static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) |
797108d1 IJ |
1106 | { |
1107 | struct tcp_sock *tp = tcp_sk(sk); | |
1108 | ||
1109 | tp->packets_out -= decr; | |
1110 | ||
1111 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
1112 | tp->sacked_out -= decr; | |
1113 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
1114 | tp->retrans_out -= decr; | |
1115 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) | |
1116 | tp->lost_out -= decr; | |
1117 | ||
1118 | /* Reno case is special. Sigh... */ | |
1119 | if (tcp_is_reno(tp) && decr > 0) | |
1120 | tp->sacked_out -= min_t(u32, tp->sacked_out, decr); | |
1121 | ||
1122 | tcp_adjust_fackets_out(sk, skb, decr); | |
1123 | ||
1124 | if (tp->lost_skb_hint && | |
1125 | before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && | |
52cf3cc8 | 1126 | (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))) |
797108d1 IJ |
1127 | tp->lost_cnt_hint -= decr; |
1128 | ||
1129 | tcp_verify_left_out(tp); | |
1130 | } | |
1131 | ||
0a2cf20c SHY |
1132 | static bool tcp_has_tx_tstamp(const struct sk_buff *skb) |
1133 | { | |
1134 | return TCP_SKB_CB(skb)->txstamp_ack || | |
1135 | (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); | |
1136 | } | |
1137 | ||
490cc7d0 WB |
1138 | static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) |
1139 | { | |
1140 | struct skb_shared_info *shinfo = skb_shinfo(skb); | |
1141 | ||
0a2cf20c | 1142 | if (unlikely(tcp_has_tx_tstamp(skb)) && |
490cc7d0 WB |
1143 | !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { |
1144 | struct skb_shared_info *shinfo2 = skb_shinfo(skb2); | |
1145 | u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; | |
1146 | ||
1147 | shinfo->tx_flags &= ~tsflags; | |
1148 | shinfo2->tx_flags |= tsflags; | |
1149 | swap(shinfo->tskey, shinfo2->tskey); | |
b51e13fa MKL |
1150 | TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; |
1151 | TCP_SKB_CB(skb)->txstamp_ack = 0; | |
490cc7d0 WB |
1152 | } |
1153 | } | |
1154 | ||
a166140e MKL |
1155 | static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) |
1156 | { | |
1157 | TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; | |
1158 | TCP_SKB_CB(skb)->eor = 0; | |
1159 | } | |
1160 | ||
1da177e4 LT |
1161 | /* Function to create two new TCP segments. Shrinks the given segment |
1162 | * to the specified size and appends a new segment with the rest of the | |
e905a9ed | 1163 | * packet to the list. This won't be called frequently, I hope. |
1da177e4 LT |
1164 | * Remember, these are still headerless SKBs at this point. |
1165 | */ | |
056834d9 | 1166 | int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len, |
6cc55e09 | 1167 | unsigned int mss_now, gfp_t gfp) |
1da177e4 LT |
1168 | { |
1169 | struct tcp_sock *tp = tcp_sk(sk); | |
1170 | struct sk_buff *buff; | |
6475be16 | 1171 | int nsize, old_factor; |
b60b49ea | 1172 | int nlen; |
9ce01461 | 1173 | u8 flags; |
1da177e4 | 1174 | |
2fceec13 IJ |
1175 | if (WARN_ON(len > skb->len)) |
1176 | return -EINVAL; | |
6a438bbe | 1177 | |
1da177e4 LT |
1178 | nsize = skb_headlen(skb) - len; |
1179 | if (nsize < 0) | |
1180 | nsize = 0; | |
1181 | ||
6cc55e09 | 1182 | if (skb_unclone(skb, gfp)) |
1da177e4 LT |
1183 | return -ENOMEM; |
1184 | ||
1185 | /* Get a new skb... force flag on. */ | |
eb934478 | 1186 | buff = sk_stream_alloc_skb(sk, nsize, gfp, true); |
51456b29 | 1187 | if (!buff) |
1da177e4 | 1188 | return -ENOMEM; /* We'll just try again later. */ |
ef5cb973 | 1189 | |
3ab224be HA |
1190 | sk->sk_wmem_queued += buff->truesize; |
1191 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea HX |
1192 | nlen = skb->len - len - nsize; |
1193 | buff->truesize += nlen; | |
1194 | skb->truesize -= nlen; | |
1da177e4 LT |
1195 | |
1196 | /* Correct the sequence numbers. */ | |
1197 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1198 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1199 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1200 | ||
1201 | /* PSH and FIN should only be set in the second packet. */ | |
4de075e0 ED |
1202 | flags = TCP_SKB_CB(skb)->tcp_flags; |
1203 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
1204 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
e14c3caf | 1205 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
a166140e | 1206 | tcp_skb_fragment_eor(skb, buff); |
1da177e4 | 1207 | |
84fa7933 | 1208 | if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) { |
1da177e4 | 1209 | /* Copy and checksum data tail into the new buffer. */ |
056834d9 IJ |
1210 | buff->csum = csum_partial_copy_nocheck(skb->data + len, |
1211 | skb_put(buff, nsize), | |
1da177e4 LT |
1212 | nsize, 0); |
1213 | ||
1214 | skb_trim(skb, len); | |
1215 | ||
1216 | skb->csum = csum_block_sub(skb->csum, buff->csum, len); | |
1217 | } else { | |
84fa7933 | 1218 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
1219 | skb_split(skb, buff, len); |
1220 | } | |
1221 | ||
1222 | buff->ip_summed = skb->ip_summed; | |
1223 | ||
a61bbcf2 | 1224 | buff->tstamp = skb->tstamp; |
490cc7d0 | 1225 | tcp_fragment_tstamp(skb, buff); |
1da177e4 | 1226 | |
6475be16 DM |
1227 | old_factor = tcp_skb_pcount(skb); |
1228 | ||
1da177e4 | 1229 | /* Fix up tso_factor for both original and new SKB. */ |
5bbb432c ED |
1230 | tcp_set_skb_tso_segs(skb, mss_now); |
1231 | tcp_set_skb_tso_segs(buff, mss_now); | |
1da177e4 | 1232 | |
b9f64820 YC |
1233 | /* Update delivered info for the new segment */ |
1234 | TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; | |
1235 | ||
6475be16 DM |
1236 | /* If this packet has been sent out already, we must |
1237 | * adjust the various packet counters. | |
1238 | */ | |
cf0b450c | 1239 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
6475be16 DM |
1240 | int diff = old_factor - tcp_skb_pcount(skb) - |
1241 | tcp_skb_pcount(buff); | |
1da177e4 | 1242 | |
797108d1 IJ |
1243 | if (diff) |
1244 | tcp_adjust_pcount(sk, skb, diff); | |
1da177e4 LT |
1245 | } |
1246 | ||
1247 | /* Link BUFF into the send queue. */ | |
f4a775d1 | 1248 | __skb_header_release(buff); |
fe067e8a | 1249 | tcp_insert_write_queue_after(skb, buff, sk); |
1da177e4 LT |
1250 | |
1251 | return 0; | |
1252 | } | |
1253 | ||
1254 | /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c | |
1255 | * eventually). The difference is that pulled data not copied, but | |
1256 | * immediately discarded. | |
1257 | */ | |
f2911969 | 1258 | static void __pskb_trim_head(struct sk_buff *skb, int len) |
1da177e4 | 1259 | { |
7b7fc97a | 1260 | struct skb_shared_info *shinfo; |
1da177e4 LT |
1261 | int i, k, eat; |
1262 | ||
4fa48bf3 ED |
1263 | eat = min_t(int, len, skb_headlen(skb)); |
1264 | if (eat) { | |
1265 | __skb_pull(skb, eat); | |
1266 | len -= eat; | |
1267 | if (!len) | |
1268 | return; | |
1269 | } | |
1da177e4 LT |
1270 | eat = len; |
1271 | k = 0; | |
7b7fc97a ED |
1272 | shinfo = skb_shinfo(skb); |
1273 | for (i = 0; i < shinfo->nr_frags; i++) { | |
1274 | int size = skb_frag_size(&shinfo->frags[i]); | |
9e903e08 ED |
1275 | |
1276 | if (size <= eat) { | |
aff65da0 | 1277 | skb_frag_unref(skb, i); |
9e903e08 | 1278 | eat -= size; |
1da177e4 | 1279 | } else { |
7b7fc97a | 1280 | shinfo->frags[k] = shinfo->frags[i]; |
1da177e4 | 1281 | if (eat) { |
7b7fc97a ED |
1282 | shinfo->frags[k].page_offset += eat; |
1283 | skb_frag_size_sub(&shinfo->frags[k], eat); | |
1da177e4 LT |
1284 | eat = 0; |
1285 | } | |
1286 | k++; | |
1287 | } | |
1288 | } | |
7b7fc97a | 1289 | shinfo->nr_frags = k; |
1da177e4 | 1290 | |
27a884dc | 1291 | skb_reset_tail_pointer(skb); |
1da177e4 LT |
1292 | skb->data_len -= len; |
1293 | skb->len = skb->data_len; | |
1da177e4 LT |
1294 | } |
1295 | ||
67edfef7 | 1296 | /* Remove acked data from a packet in the transmit queue. */ |
1da177e4 LT |
1297 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) |
1298 | { | |
14bbd6a5 | 1299 | if (skb_unclone(skb, GFP_ATOMIC)) |
1da177e4 LT |
1300 | return -ENOMEM; |
1301 | ||
4fa48bf3 | 1302 | __pskb_trim_head(skb, len); |
1da177e4 LT |
1303 | |
1304 | TCP_SKB_CB(skb)->seq += len; | |
84fa7933 | 1305 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 LT |
1306 | |
1307 | skb->truesize -= len; | |
1308 | sk->sk_wmem_queued -= len; | |
3ab224be | 1309 | sk_mem_uncharge(sk, len); |
1da177e4 LT |
1310 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); |
1311 | ||
5b35e1e6 | 1312 | /* Any change of skb->len requires recalculation of tso factor. */ |
1da177e4 | 1313 | if (tcp_skb_pcount(skb) > 1) |
5bbb432c | 1314 | tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); |
1da177e4 LT |
1315 | |
1316 | return 0; | |
1317 | } | |
1318 | ||
1b63edd6 YC |
1319 | /* Calculate MSS not accounting any TCP options. */ |
1320 | static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
5d424d5a | 1321 | { |
cf533ea5 ED |
1322 | const struct tcp_sock *tp = tcp_sk(sk); |
1323 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
5d424d5a JH |
1324 | int mss_now; |
1325 | ||
1326 | /* Calculate base mss without TCP options: | |
1327 | It is MMS_S - sizeof(tcphdr) of rfc1122 | |
1328 | */ | |
1329 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); | |
1330 | ||
67469601 ED |
1331 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
1332 | if (icsk->icsk_af_ops->net_frag_header_len) { | |
1333 | const struct dst_entry *dst = __sk_dst_get(sk); | |
1334 | ||
1335 | if (dst && dst_allfrag(dst)) | |
1336 | mss_now -= icsk->icsk_af_ops->net_frag_header_len; | |
1337 | } | |
1338 | ||
5d424d5a JH |
1339 | /* Clamp it (mss_clamp does not include tcp options) */ |
1340 | if (mss_now > tp->rx_opt.mss_clamp) | |
1341 | mss_now = tp->rx_opt.mss_clamp; | |
1342 | ||
1343 | /* Now subtract optional transport overhead */ | |
1344 | mss_now -= icsk->icsk_ext_hdr_len; | |
1345 | ||
1346 | /* Then reserve room for full set of TCP options and 8 bytes of data */ | |
1347 | if (mss_now < 48) | |
1348 | mss_now = 48; | |
5d424d5a JH |
1349 | return mss_now; |
1350 | } | |
1351 | ||
1b63edd6 YC |
1352 | /* Calculate MSS. Not accounting for SACKs here. */ |
1353 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
1354 | { | |
1355 | /* Subtract TCP options size, not including SACKs */ | |
1356 | return __tcp_mtu_to_mss(sk, pmtu) - | |
1357 | (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); | |
1358 | } | |
1359 | ||
5d424d5a | 1360 | /* Inverse of above */ |
67469601 | 1361 | int tcp_mss_to_mtu(struct sock *sk, int mss) |
5d424d5a | 1362 | { |
cf533ea5 ED |
1363 | const struct tcp_sock *tp = tcp_sk(sk); |
1364 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
5d424d5a JH |
1365 | int mtu; |
1366 | ||
1367 | mtu = mss + | |
1368 | tp->tcp_header_len + | |
1369 | icsk->icsk_ext_hdr_len + | |
1370 | icsk->icsk_af_ops->net_header_len; | |
1371 | ||
67469601 ED |
1372 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
1373 | if (icsk->icsk_af_ops->net_frag_header_len) { | |
1374 | const struct dst_entry *dst = __sk_dst_get(sk); | |
1375 | ||
1376 | if (dst && dst_allfrag(dst)) | |
1377 | mtu += icsk->icsk_af_ops->net_frag_header_len; | |
1378 | } | |
5d424d5a JH |
1379 | return mtu; |
1380 | } | |
556c6b46 | 1381 | EXPORT_SYMBOL(tcp_mss_to_mtu); |
5d424d5a | 1382 | |
67edfef7 | 1383 | /* MTU probing init per socket */ |
5d424d5a JH |
1384 | void tcp_mtup_init(struct sock *sk) |
1385 | { | |
1386 | struct tcp_sock *tp = tcp_sk(sk); | |
1387 | struct inet_connection_sock *icsk = inet_csk(sk); | |
b0f9ca53 | 1388 | struct net *net = sock_net(sk); |
5d424d5a | 1389 | |
b0f9ca53 | 1390 | icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; |
5d424d5a | 1391 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + |
e905a9ed | 1392 | icsk->icsk_af_ops->net_header_len; |
b0f9ca53 | 1393 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); |
5d424d5a | 1394 | icsk->icsk_mtup.probe_size = 0; |
05cbc0db FD |
1395 | if (icsk->icsk_mtup.enabled) |
1396 | icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; | |
5d424d5a | 1397 | } |
4bc2f18b | 1398 | EXPORT_SYMBOL(tcp_mtup_init); |
5d424d5a | 1399 | |
1da177e4 LT |
1400 | /* This function synchronize snd mss to current pmtu/exthdr set. |
1401 | ||
1402 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts | |
1403 | for TCP options, but includes only bare TCP header. | |
1404 | ||
1405 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. | |
caa20d9a | 1406 | It is minimum of user_mss and mss received with SYN. |
1da177e4 LT |
1407 | It also does not include TCP options. |
1408 | ||
d83d8461 | 1409 | inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. |
1da177e4 LT |
1410 | |
1411 | tp->mss_cache is current effective sending mss, including | |
1412 | all tcp options except for SACKs. It is evaluated, | |
1413 | taking into account current pmtu, but never exceeds | |
1414 | tp->rx_opt.mss_clamp. | |
1415 | ||
1416 | NOTE1. rfc1122 clearly states that advertised MSS | |
1417 | DOES NOT include either tcp or ip options. | |
1418 | ||
d83d8461 ACM |
1419 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
1420 | are READ ONLY outside this function. --ANK (980731) | |
1da177e4 | 1421 | */ |
1da177e4 LT |
1422 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) |
1423 | { | |
1424 | struct tcp_sock *tp = tcp_sk(sk); | |
d83d8461 | 1425 | struct inet_connection_sock *icsk = inet_csk(sk); |
5d424d5a | 1426 | int mss_now; |
1da177e4 | 1427 | |
5d424d5a JH |
1428 | if (icsk->icsk_mtup.search_high > pmtu) |
1429 | icsk->icsk_mtup.search_high = pmtu; | |
1da177e4 | 1430 | |
5d424d5a | 1431 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
409d22b4 | 1432 | mss_now = tcp_bound_to_half_wnd(tp, mss_now); |
1da177e4 LT |
1433 | |
1434 | /* And store cached results */ | |
d83d8461 | 1435 | icsk->icsk_pmtu_cookie = pmtu; |
5d424d5a JH |
1436 | if (icsk->icsk_mtup.enabled) |
1437 | mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); | |
c1b4a7e6 | 1438 | tp->mss_cache = mss_now; |
1da177e4 LT |
1439 | |
1440 | return mss_now; | |
1441 | } | |
4bc2f18b | 1442 | EXPORT_SYMBOL(tcp_sync_mss); |
1da177e4 LT |
1443 | |
1444 | /* Compute the current effective MSS, taking SACKs and IP options, | |
1445 | * and even PMTU discovery events into account. | |
1da177e4 | 1446 | */ |
0c54b85f | 1447 | unsigned int tcp_current_mss(struct sock *sk) |
1da177e4 | 1448 | { |
cf533ea5 ED |
1449 | const struct tcp_sock *tp = tcp_sk(sk); |
1450 | const struct dst_entry *dst = __sk_dst_get(sk); | |
c1b4a7e6 | 1451 | u32 mss_now; |
95c96174 | 1452 | unsigned int header_len; |
33ad798c AL |
1453 | struct tcp_out_options opts; |
1454 | struct tcp_md5sig_key *md5; | |
c1b4a7e6 DM |
1455 | |
1456 | mss_now = tp->mss_cache; | |
1457 | ||
1da177e4 LT |
1458 | if (dst) { |
1459 | u32 mtu = dst_mtu(dst); | |
d83d8461 | 1460 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
1da177e4 LT |
1461 | mss_now = tcp_sync_mss(sk, mtu); |
1462 | } | |
1463 | ||
33ad798c AL |
1464 | header_len = tcp_established_options(sk, NULL, &opts, &md5) + |
1465 | sizeof(struct tcphdr); | |
1466 | /* The mss_cache is sized based on tp->tcp_header_len, which assumes | |
1467 | * some common options. If this is an odd packet (because we have SACK | |
1468 | * blocks etc) then our calculated header_len will be different, and | |
1469 | * we have to adjust mss_now correspondingly */ | |
1470 | if (header_len != tp->tcp_header_len) { | |
1471 | int delta = (int) header_len - tp->tcp_header_len; | |
1472 | mss_now -= delta; | |
1473 | } | |
cfb6eeb4 | 1474 | |
1da177e4 LT |
1475 | return mss_now; |
1476 | } | |
1477 | ||
86fd14ad WP |
1478 | /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. |
1479 | * As additional protections, we do not touch cwnd in retransmission phases, | |
1480 | * and if application hit its sndbuf limit recently. | |
1481 | */ | |
1482 | static void tcp_cwnd_application_limited(struct sock *sk) | |
1483 | { | |
1484 | struct tcp_sock *tp = tcp_sk(sk); | |
1485 | ||
1486 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && | |
1487 | sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { | |
1488 | /* Limited by application or receiver window. */ | |
1489 | u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); | |
1490 | u32 win_used = max(tp->snd_cwnd_used, init_win); | |
1491 | if (win_used < tp->snd_cwnd) { | |
1492 | tp->snd_ssthresh = tcp_current_ssthresh(sk); | |
1493 | tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; | |
1494 | } | |
1495 | tp->snd_cwnd_used = 0; | |
1496 | } | |
1497 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1498 | } | |
1499 | ||
ca8a2263 | 1500 | static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) |
a762a980 | 1501 | { |
9e412ba7 | 1502 | struct tcp_sock *tp = tcp_sk(sk); |
a762a980 | 1503 | |
ca8a2263 NC |
1504 | /* Track the maximum number of outstanding packets in each |
1505 | * window, and remember whether we were cwnd-limited then. | |
1506 | */ | |
1507 | if (!before(tp->snd_una, tp->max_packets_seq) || | |
1508 | tp->packets_out > tp->max_packets_out) { | |
1509 | tp->max_packets_out = tp->packets_out; | |
1510 | tp->max_packets_seq = tp->snd_nxt; | |
1511 | tp->is_cwnd_limited = is_cwnd_limited; | |
1512 | } | |
e114a710 | 1513 | |
24901551 | 1514 | if (tcp_is_cwnd_limited(sk)) { |
a762a980 DM |
1515 | /* Network is feed fully. */ |
1516 | tp->snd_cwnd_used = 0; | |
1517 | tp->snd_cwnd_stamp = tcp_time_stamp; | |
1518 | } else { | |
1519 | /* Network starves. */ | |
1520 | if (tp->packets_out > tp->snd_cwnd_used) | |
1521 | tp->snd_cwnd_used = tp->packets_out; | |
1522 | ||
15d33c07 DM |
1523 | if (sysctl_tcp_slow_start_after_idle && |
1524 | (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto) | |
a762a980 | 1525 | tcp_cwnd_application_limited(sk); |
b0f71bd3 FY |
1526 | |
1527 | /* The following conditions together indicate the starvation | |
1528 | * is caused by insufficient sender buffer: | |
1529 | * 1) just sent some data (see tcp_write_xmit) | |
1530 | * 2) not cwnd limited (this else condition) | |
1531 | * 3) no more data to send (null tcp_send_head ) | |
1532 | * 4) application is hitting buffer limit (SOCK_NOSPACE) | |
1533 | */ | |
1534 | if (!tcp_send_head(sk) && sk->sk_socket && | |
1535 | test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && | |
1536 | (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) | |
1537 | tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); | |
a762a980 DM |
1538 | } |
1539 | } | |
1540 | ||
d4589926 ED |
1541 | /* Minshall's variant of the Nagle send check. */ |
1542 | static bool tcp_minshall_check(const struct tcp_sock *tp) | |
1543 | { | |
1544 | return after(tp->snd_sml, tp->snd_una) && | |
1545 | !after(tp->snd_sml, tp->snd_nxt); | |
1546 | } | |
1547 | ||
1548 | /* Update snd_sml if this skb is under mss | |
1549 | * Note that a TSO packet might end with a sub-mss segment | |
1550 | * The test is really : | |
1551 | * if ((skb->len % mss) != 0) | |
1552 | * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; | |
1553 | * But we can avoid doing the divide again given we already have | |
1554 | * skb_pcount = skb->len / mss_now | |
0e3a4803 | 1555 | */ |
d4589926 ED |
1556 | static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, |
1557 | const struct sk_buff *skb) | |
1558 | { | |
1559 | if (skb->len < tcp_skb_pcount(skb) * mss_now) | |
1560 | tp->snd_sml = TCP_SKB_CB(skb)->end_seq; | |
1561 | } | |
1562 | ||
1563 | /* Return false, if packet can be sent now without violation Nagle's rules: | |
1564 | * 1. It is full sized. (provided by caller in %partial bool) | |
1565 | * 2. Or it contains FIN. (already checked by caller) | |
1566 | * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. | |
1567 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. | |
1568 | * With Minshall's modification: all sent small packets are ACKed. | |
1569 | */ | |
1570 | static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, | |
cc93fc51 | 1571 | int nonagle) |
d4589926 ED |
1572 | { |
1573 | return partial && | |
1574 | ((nonagle & TCP_NAGLE_CORK) || | |
1575 | (!nonagle && tp->packets_out && tcp_minshall_check(tp))); | |
1576 | } | |
605ad7f1 ED |
1577 | |
1578 | /* Return how many segs we'd like on a TSO packet, | |
1579 | * to send one TSO packet per ms | |
1580 | */ | |
1b3878ca NC |
1581 | u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, |
1582 | int min_tso_segs) | |
605ad7f1 ED |
1583 | { |
1584 | u32 bytes, segs; | |
1585 | ||
1586 | bytes = min(sk->sk_pacing_rate >> 10, | |
1587 | sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); | |
1588 | ||
1589 | /* Goal is to send at least one packet per ms, | |
1590 | * not one big TSO packet every 100 ms. | |
1591 | * This preserves ACK clocking and is consistent | |
1592 | * with tcp_tso_should_defer() heuristic. | |
1593 | */ | |
1b3878ca | 1594 | segs = max_t(u32, bytes / mss_now, min_tso_segs); |
605ad7f1 ED |
1595 | |
1596 | return min_t(u32, segs, sk->sk_gso_max_segs); | |
1597 | } | |
1b3878ca | 1598 | EXPORT_SYMBOL(tcp_tso_autosize); |
605ad7f1 | 1599 | |
ed6e7268 NC |
1600 | /* Return the number of segments we want in the skb we are transmitting. |
1601 | * See if congestion control module wants to decide; otherwise, autosize. | |
1602 | */ | |
1603 | static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) | |
1604 | { | |
1605 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; | |
1606 | u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0; | |
1607 | ||
1b3878ca NC |
1608 | return tso_segs ? : |
1609 | tcp_tso_autosize(sk, mss_now, sysctl_tcp_min_tso_segs); | |
ed6e7268 NC |
1610 | } |
1611 | ||
d4589926 ED |
1612 | /* Returns the portion of skb which can be sent right away */ |
1613 | static unsigned int tcp_mss_split_point(const struct sock *sk, | |
1614 | const struct sk_buff *skb, | |
1615 | unsigned int mss_now, | |
1616 | unsigned int max_segs, | |
1617 | int nonagle) | |
c1b4a7e6 | 1618 | { |
cf533ea5 | 1619 | const struct tcp_sock *tp = tcp_sk(sk); |
d4589926 | 1620 | u32 partial, needed, window, max_len; |
c1b4a7e6 | 1621 | |
90840def | 1622 | window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
1485348d | 1623 | max_len = mss_now * max_segs; |
0e3a4803 | 1624 | |
1485348d BH |
1625 | if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) |
1626 | return max_len; | |
0e3a4803 | 1627 | |
5ea3a748 IJ |
1628 | needed = min(skb->len, window); |
1629 | ||
1485348d BH |
1630 | if (max_len <= needed) |
1631 | return max_len; | |
0e3a4803 | 1632 | |
d4589926 ED |
1633 | partial = needed % mss_now; |
1634 | /* If last segment is not a full MSS, check if Nagle rules allow us | |
1635 | * to include this last segment in this skb. | |
1636 | * Otherwise, we'll split the skb at last MSS boundary | |
1637 | */ | |
cc93fc51 | 1638 | if (tcp_nagle_check(partial != 0, tp, nonagle)) |
d4589926 ED |
1639 | return needed - partial; |
1640 | ||
1641 | return needed; | |
c1b4a7e6 DM |
1642 | } |
1643 | ||
1644 | /* Can at least one segment of SKB be sent right now, according to the | |
1645 | * congestion window rules? If so, return how many segments are allowed. | |
1646 | */ | |
cf533ea5 ED |
1647 | static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, |
1648 | const struct sk_buff *skb) | |
c1b4a7e6 | 1649 | { |
d649a7a8 | 1650 | u32 in_flight, cwnd, halfcwnd; |
c1b4a7e6 DM |
1651 | |
1652 | /* Don't be strict about the congestion window for the final FIN. */ | |
4de075e0 ED |
1653 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && |
1654 | tcp_skb_pcount(skb) == 1) | |
c1b4a7e6 DM |
1655 | return 1; |
1656 | ||
1657 | in_flight = tcp_packets_in_flight(tp); | |
1658 | cwnd = tp->snd_cwnd; | |
d649a7a8 ED |
1659 | if (in_flight >= cwnd) |
1660 | return 0; | |
c1b4a7e6 | 1661 | |
d649a7a8 ED |
1662 | /* For better scheduling, ensure we have at least |
1663 | * 2 GSO packets in flight. | |
1664 | */ | |
1665 | halfcwnd = max(cwnd >> 1, 1U); | |
1666 | return min(halfcwnd, cwnd - in_flight); | |
c1b4a7e6 DM |
1667 | } |
1668 | ||
b595076a | 1669 | /* Initialize TSO state of a skb. |
67edfef7 | 1670 | * This must be invoked the first time we consider transmitting |
c1b4a7e6 DM |
1671 | * SKB onto the wire. |
1672 | */ | |
5bbb432c | 1673 | static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) |
c1b4a7e6 DM |
1674 | { |
1675 | int tso_segs = tcp_skb_pcount(skb); | |
1676 | ||
f8269a49 | 1677 | if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { |
5bbb432c | 1678 | tcp_set_skb_tso_segs(skb, mss_now); |
c1b4a7e6 DM |
1679 | tso_segs = tcp_skb_pcount(skb); |
1680 | } | |
1681 | return tso_segs; | |
1682 | } | |
1683 | ||
c1b4a7e6 | 1684 | |
a2a385d6 | 1685 | /* Return true if the Nagle test allows this packet to be |
c1b4a7e6 DM |
1686 | * sent now. |
1687 | */ | |
a2a385d6 ED |
1688 | static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, |
1689 | unsigned int cur_mss, int nonagle) | |
c1b4a7e6 DM |
1690 | { |
1691 | /* Nagle rule does not apply to frames, which sit in the middle of the | |
1692 | * write_queue (they have no chances to get new data). | |
1693 | * | |
1694 | * This is implemented in the callers, where they modify the 'nonagle' | |
1695 | * argument based upon the location of SKB in the send queue. | |
1696 | */ | |
1697 | if (nonagle & TCP_NAGLE_PUSH) | |
a2a385d6 | 1698 | return true; |
c1b4a7e6 | 1699 | |
9b44190d YC |
1700 | /* Don't use the nagle rule for urgent data (or for the final FIN). */ |
1701 | if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) | |
a2a385d6 | 1702 | return true; |
c1b4a7e6 | 1703 | |
cc93fc51 | 1704 | if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) |
a2a385d6 | 1705 | return true; |
c1b4a7e6 | 1706 | |
a2a385d6 | 1707 | return false; |
c1b4a7e6 DM |
1708 | } |
1709 | ||
1710 | /* Does at least the first segment of SKB fit into the send window? */ | |
a2a385d6 ED |
1711 | static bool tcp_snd_wnd_test(const struct tcp_sock *tp, |
1712 | const struct sk_buff *skb, | |
1713 | unsigned int cur_mss) | |
c1b4a7e6 DM |
1714 | { |
1715 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
1716 | ||
1717 | if (skb->len > cur_mss) | |
1718 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; | |
1719 | ||
90840def | 1720 | return !after(end_seq, tcp_wnd_end(tp)); |
c1b4a7e6 DM |
1721 | } |
1722 | ||
fe067e8a | 1723 | /* This checks if the data bearing packet SKB (usually tcp_send_head(sk)) |
c1b4a7e6 DM |
1724 | * should be put on the wire right now. If so, it returns the number of |
1725 | * packets allowed by the congestion window. | |
1726 | */ | |
cf533ea5 | 1727 | static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb, |
c1b4a7e6 DM |
1728 | unsigned int cur_mss, int nonagle) |
1729 | { | |
cf533ea5 | 1730 | const struct tcp_sock *tp = tcp_sk(sk); |
c1b4a7e6 DM |
1731 | unsigned int cwnd_quota; |
1732 | ||
5bbb432c | 1733 | tcp_init_tso_segs(skb, cur_mss); |
c1b4a7e6 DM |
1734 | |
1735 | if (!tcp_nagle_test(tp, skb, cur_mss, nonagle)) | |
1736 | return 0; | |
1737 | ||
1738 | cwnd_quota = tcp_cwnd_test(tp, skb); | |
056834d9 | 1739 | if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss)) |
c1b4a7e6 DM |
1740 | cwnd_quota = 0; |
1741 | ||
1742 | return cwnd_quota; | |
1743 | } | |
1744 | ||
67edfef7 | 1745 | /* Test if sending is allowed right now. */ |
a2a385d6 | 1746 | bool tcp_may_send_now(struct sock *sk) |
c1b4a7e6 | 1747 | { |
cf533ea5 | 1748 | const struct tcp_sock *tp = tcp_sk(sk); |
fe067e8a | 1749 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 | 1750 | |
a02cec21 | 1751 | return skb && |
0c54b85f | 1752 | tcp_snd_test(sk, skb, tcp_current_mss(sk), |
c1b4a7e6 | 1753 | (tcp_skb_is_last(sk, skb) ? |
a02cec21 | 1754 | tp->nonagle : TCP_NAGLE_PUSH)); |
c1b4a7e6 DM |
1755 | } |
1756 | ||
1757 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet | |
1758 | * which is put after SKB on the list. It is very much like | |
1759 | * tcp_fragment() except that it may make several kinds of assumptions | |
1760 | * in order to speed up the splitting operation. In particular, we | |
1761 | * know that all the data is in scatter-gather pages, and that the | |
1762 | * packet has never been sent out before (and thus is not cloned). | |
1763 | */ | |
056834d9 | 1764 | static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, |
c4ead4c5 | 1765 | unsigned int mss_now, gfp_t gfp) |
c1b4a7e6 DM |
1766 | { |
1767 | struct sk_buff *buff; | |
1768 | int nlen = skb->len - len; | |
9ce01461 | 1769 | u8 flags; |
c1b4a7e6 DM |
1770 | |
1771 | /* All of a TSO frame must be composed of paged data. */ | |
c8ac3774 | 1772 | if (skb->len != skb->data_len) |
6cc55e09 | 1773 | return tcp_fragment(sk, skb, len, mss_now, gfp); |
c1b4a7e6 | 1774 | |
eb934478 | 1775 | buff = sk_stream_alloc_skb(sk, 0, gfp, true); |
51456b29 | 1776 | if (unlikely(!buff)) |
c1b4a7e6 DM |
1777 | return -ENOMEM; |
1778 | ||
3ab224be HA |
1779 | sk->sk_wmem_queued += buff->truesize; |
1780 | sk_mem_charge(sk, buff->truesize); | |
b60b49ea | 1781 | buff->truesize += nlen; |
c1b4a7e6 DM |
1782 | skb->truesize -= nlen; |
1783 | ||
1784 | /* Correct the sequence numbers. */ | |
1785 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1786 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1787 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1788 | ||
1789 | /* PSH and FIN should only be set in the second packet. */ | |
4de075e0 ED |
1790 | flags = TCP_SKB_CB(skb)->tcp_flags; |
1791 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
1792 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
c1b4a7e6 DM |
1793 | |
1794 | /* This packet was never sent out yet, so no SACK bits. */ | |
1795 | TCP_SKB_CB(buff)->sacked = 0; | |
1796 | ||
a166140e MKL |
1797 | tcp_skb_fragment_eor(skb, buff); |
1798 | ||
84fa7933 | 1799 | buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL; |
c1b4a7e6 | 1800 | skb_split(skb, buff, len); |
490cc7d0 | 1801 | tcp_fragment_tstamp(skb, buff); |
c1b4a7e6 DM |
1802 | |
1803 | /* Fix up tso_factor for both original and new SKB. */ | |
5bbb432c ED |
1804 | tcp_set_skb_tso_segs(skb, mss_now); |
1805 | tcp_set_skb_tso_segs(buff, mss_now); | |
c1b4a7e6 DM |
1806 | |
1807 | /* Link BUFF into the send queue. */ | |
f4a775d1 | 1808 | __skb_header_release(buff); |
fe067e8a | 1809 | tcp_insert_write_queue_after(skb, buff, sk); |
c1b4a7e6 DM |
1810 | |
1811 | return 0; | |
1812 | } | |
1813 | ||
1814 | /* Try to defer sending, if possible, in order to minimize the amount | |
1815 | * of TSO splitting we do. View it as a kind of TSO Nagle test. | |
1816 | * | |
1817 | * This algorithm is from John Heffner. | |
1818 | */ | |
ca8a2263 | 1819 | static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, |
605ad7f1 | 1820 | bool *is_cwnd_limited, u32 max_segs) |
c1b4a7e6 | 1821 | { |
6687e988 | 1822 | const struct inet_connection_sock *icsk = inet_csk(sk); |
50c8339e ED |
1823 | u32 age, send_win, cong_win, limit, in_flight; |
1824 | struct tcp_sock *tp = tcp_sk(sk); | |
1825 | struct skb_mstamp now; | |
1826 | struct sk_buff *head; | |
ad9f4f50 | 1827 | int win_divisor; |
c1b4a7e6 | 1828 | |
4de075e0 | 1829 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) |
ae8064ac | 1830 | goto send_now; |
c1b4a7e6 | 1831 | |
99d7662a | 1832 | if (icsk->icsk_ca_state >= TCP_CA_Recovery) |
ae8064ac JH |
1833 | goto send_now; |
1834 | ||
5f852eb5 ED |
1835 | /* Avoid bursty behavior by allowing defer |
1836 | * only if the last write was recent. | |
1837 | */ | |
1838 | if ((s32)(tcp_time_stamp - tp->lsndtime) > 0) | |
ae8064ac | 1839 | goto send_now; |
908a75c1 | 1840 | |
c1b4a7e6 DM |
1841 | in_flight = tcp_packets_in_flight(tp); |
1842 | ||
056834d9 | 1843 | BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight)); |
c1b4a7e6 | 1844 | |
90840def | 1845 | send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
c1b4a7e6 DM |
1846 | |
1847 | /* From in_flight test above, we know that cwnd > in_flight. */ | |
1848 | cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; | |
1849 | ||
1850 | limit = min(send_win, cong_win); | |
1851 | ||
ba244fe9 | 1852 | /* If a full-sized TSO skb can be sent, do it. */ |
605ad7f1 | 1853 | if (limit >= max_segs * tp->mss_cache) |
ae8064ac | 1854 | goto send_now; |
ba244fe9 | 1855 | |
62ad2761 IJ |
1856 | /* Middle in queue won't get any more data, full sendable already? */ |
1857 | if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) | |
1858 | goto send_now; | |
1859 | ||
ad9f4f50 ED |
1860 | win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor); |
1861 | if (win_divisor) { | |
c1b4a7e6 DM |
1862 | u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); |
1863 | ||
1864 | /* If at least some fraction of a window is available, | |
1865 | * just use it. | |
1866 | */ | |
ad9f4f50 | 1867 | chunk /= win_divisor; |
c1b4a7e6 | 1868 | if (limit >= chunk) |
ae8064ac | 1869 | goto send_now; |
c1b4a7e6 DM |
1870 | } else { |
1871 | /* Different approach, try not to defer past a single | |
1872 | * ACK. Receiver should ACK every other full sized | |
1873 | * frame, so if we have space for more than 3 frames | |
1874 | * then send now. | |
1875 | */ | |
6b5a5c0d | 1876 | if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) |
ae8064ac | 1877 | goto send_now; |
c1b4a7e6 DM |
1878 | } |
1879 | ||
50c8339e ED |
1880 | head = tcp_write_queue_head(sk); |
1881 | skb_mstamp_get(&now); | |
1882 | age = skb_mstamp_us_delta(&now, &head->skb_mstamp); | |
1883 | /* If next ACK is likely to come too late (half srtt), do not defer */ | |
1884 | if (age < (tp->srtt_us >> 4)) | |
1885 | goto send_now; | |
1886 | ||
5f852eb5 | 1887 | /* Ok, it looks like it is advisable to defer. */ |
ae8064ac | 1888 | |
d2e1339f | 1889 | if (cong_win < send_win && cong_win <= skb->len) |
ca8a2263 NC |
1890 | *is_cwnd_limited = true; |
1891 | ||
a2a385d6 | 1892 | return true; |
ae8064ac JH |
1893 | |
1894 | send_now: | |
a2a385d6 | 1895 | return false; |
c1b4a7e6 DM |
1896 | } |
1897 | ||
05cbc0db FD |
1898 | static inline void tcp_mtu_check_reprobe(struct sock *sk) |
1899 | { | |
1900 | struct inet_connection_sock *icsk = inet_csk(sk); | |
1901 | struct tcp_sock *tp = tcp_sk(sk); | |
1902 | struct net *net = sock_net(sk); | |
1903 | u32 interval; | |
1904 | s32 delta; | |
1905 | ||
1906 | interval = net->ipv4.sysctl_tcp_probe_interval; | |
1907 | delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp; | |
1908 | if (unlikely(delta >= interval * HZ)) { | |
1909 | int mss = tcp_current_mss(sk); | |
1910 | ||
1911 | /* Update current search range */ | |
1912 | icsk->icsk_mtup.probe_size = 0; | |
1913 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + | |
1914 | sizeof(struct tcphdr) + | |
1915 | icsk->icsk_af_ops->net_header_len; | |
1916 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); | |
1917 | ||
1918 | /* Update probe time stamp */ | |
1919 | icsk->icsk_mtup.probe_timestamp = tcp_time_stamp; | |
1920 | } | |
1921 | } | |
1922 | ||
5d424d5a | 1923 | /* Create a new MTU probe if we are ready. |
67edfef7 AK |
1924 | * MTU probe is regularly attempting to increase the path MTU by |
1925 | * deliberately sending larger packets. This discovers routing | |
1926 | * changes resulting in larger path MTUs. | |
1927 | * | |
5d424d5a JH |
1928 | * Returns 0 if we should wait to probe (no cwnd available), |
1929 | * 1 if a probe was sent, | |
056834d9 IJ |
1930 | * -1 otherwise |
1931 | */ | |
5d424d5a JH |
1932 | static int tcp_mtu_probe(struct sock *sk) |
1933 | { | |
5d424d5a | 1934 | struct inet_connection_sock *icsk = inet_csk(sk); |
12a59abc | 1935 | struct tcp_sock *tp = tcp_sk(sk); |
5d424d5a | 1936 | struct sk_buff *skb, *nskb, *next; |
6b58e0a5 | 1937 | struct net *net = sock_net(sk); |
5d424d5a | 1938 | int probe_size; |
91cc17c0 | 1939 | int size_needed; |
12a59abc | 1940 | int copy, len; |
5d424d5a | 1941 | int mss_now; |
6b58e0a5 | 1942 | int interval; |
5d424d5a JH |
1943 | |
1944 | /* Not currently probing/verifying, | |
1945 | * not in recovery, | |
1946 | * have enough cwnd, and | |
12a59abc ED |
1947 | * not SACKing (the variable headers throw things off) |
1948 | */ | |
1949 | if (likely(!icsk->icsk_mtup.enabled || | |
1950 | icsk->icsk_mtup.probe_size || | |
1951 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || | |
1952 | tp->snd_cwnd < 11 || | |
1953 | tp->rx_opt.num_sacks || tp->rx_opt.dsack)) | |
5d424d5a JH |
1954 | return -1; |
1955 | ||
6b58e0a5 FD |
1956 | /* Use binary search for probe_size between tcp_mss_base, |
1957 | * and current mss_clamp. if (search_high - search_low) | |
1958 | * smaller than a threshold, backoff from probing. | |
1959 | */ | |
0c54b85f | 1960 | mss_now = tcp_current_mss(sk); |
6b58e0a5 FD |
1961 | probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + |
1962 | icsk->icsk_mtup.search_low) >> 1); | |
91cc17c0 | 1963 | size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; |
6b58e0a5 | 1964 | interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; |
05cbc0db FD |
1965 | /* When misfortune happens, we are reprobing actively, |
1966 | * and then reprobe timer has expired. We stick with current | |
1967 | * probing process by not resetting search range to its orignal. | |
1968 | */ | |
6b58e0a5 | 1969 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || |
05cbc0db FD |
1970 | interval < net->ipv4.sysctl_tcp_probe_threshold) { |
1971 | /* Check whether enough time has elaplased for | |
1972 | * another round of probing. | |
1973 | */ | |
1974 | tcp_mtu_check_reprobe(sk); | |
5d424d5a JH |
1975 | return -1; |
1976 | } | |
1977 | ||
1978 | /* Have enough data in the send queue to probe? */ | |
7f9c33e5 | 1979 | if (tp->write_seq - tp->snd_nxt < size_needed) |
5d424d5a JH |
1980 | return -1; |
1981 | ||
91cc17c0 IJ |
1982 | if (tp->snd_wnd < size_needed) |
1983 | return -1; | |
90840def | 1984 | if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) |
91cc17c0 | 1985 | return 0; |
5d424d5a | 1986 | |
d67c58e9 IJ |
1987 | /* Do we need to wait to drain cwnd? With none in flight, don't stall */ |
1988 | if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { | |
1989 | if (!tcp_packets_in_flight(tp)) | |
5d424d5a JH |
1990 | return -1; |
1991 | else | |
1992 | return 0; | |
1993 | } | |
1994 | ||
1995 | /* We're allowed to probe. Build it now. */ | |
eb934478 | 1996 | nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); |
51456b29 | 1997 | if (!nskb) |
5d424d5a | 1998 | return -1; |
3ab224be HA |
1999 | sk->sk_wmem_queued += nskb->truesize; |
2000 | sk_mem_charge(sk, nskb->truesize); | |
5d424d5a | 2001 | |
fe067e8a | 2002 | skb = tcp_send_head(sk); |
5d424d5a JH |
2003 | |
2004 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; | |
2005 | TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; | |
4de075e0 | 2006 | TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; |
5d424d5a JH |
2007 | TCP_SKB_CB(nskb)->sacked = 0; |
2008 | nskb->csum = 0; | |
84fa7933 | 2009 | nskb->ip_summed = skb->ip_summed; |
5d424d5a | 2010 | |
50c4817e IJ |
2011 | tcp_insert_write_queue_before(nskb, skb, sk); |
2012 | ||
5d424d5a | 2013 | len = 0; |
234b6860 | 2014 | tcp_for_write_queue_from_safe(skb, next, sk) { |
5d424d5a | 2015 | copy = min_t(int, skb->len, probe_size - len); |
2fe664f1 | 2016 | if (nskb->ip_summed) { |
5d424d5a | 2017 | skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); |
2fe664f1 DCS |
2018 | } else { |
2019 | __wsum csum = skb_copy_and_csum_bits(skb, 0, | |
2020 | skb_put(nskb, copy), | |
2021 | copy, 0); | |
2022 | nskb->csum = csum_block_add(nskb->csum, csum, len); | |
2023 | } | |
5d424d5a JH |
2024 | |
2025 | if (skb->len <= copy) { | |
2026 | /* We've eaten all the data from this skb. | |
2027 | * Throw it away. */ | |
4de075e0 | 2028 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; |
fe067e8a | 2029 | tcp_unlink_write_queue(skb, sk); |
3ab224be | 2030 | sk_wmem_free_skb(sk, skb); |
5d424d5a | 2031 | } else { |
4de075e0 | 2032 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & |
a3433f35 | 2033 | ~(TCPHDR_FIN|TCPHDR_PSH); |
5d424d5a JH |
2034 | if (!skb_shinfo(skb)->nr_frags) { |
2035 | skb_pull(skb, copy); | |
84fa7933 | 2036 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
056834d9 IJ |
2037 | skb->csum = csum_partial(skb->data, |
2038 | skb->len, 0); | |
5d424d5a JH |
2039 | } else { |
2040 | __pskb_trim_head(skb, copy); | |
5bbb432c | 2041 | tcp_set_skb_tso_segs(skb, mss_now); |
5d424d5a JH |
2042 | } |
2043 | TCP_SKB_CB(skb)->seq += copy; | |
2044 | } | |
2045 | ||
2046 | len += copy; | |
234b6860 IJ |
2047 | |
2048 | if (len >= probe_size) | |
2049 | break; | |
5d424d5a | 2050 | } |
5bbb432c | 2051 | tcp_init_tso_segs(nskb, nskb->len); |
5d424d5a JH |
2052 | |
2053 | /* We're ready to send. If this fails, the probe will | |
7faee5c0 ED |
2054 | * be resegmented into mss-sized pieces by tcp_write_xmit(). |
2055 | */ | |
5d424d5a JH |
2056 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { |
2057 | /* Decrement cwnd here because we are sending | |
056834d9 | 2058 | * effectively two packets. */ |
5d424d5a | 2059 | tp->snd_cwnd--; |
66f5fe62 | 2060 | tcp_event_new_data_sent(sk, nskb); |
5d424d5a JH |
2061 | |
2062 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); | |
0e7b1368 JH |
2063 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; |
2064 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; | |
5d424d5a JH |
2065 | |
2066 | return 1; | |
2067 | } | |
2068 | ||
2069 | return -1; | |
2070 | } | |
2071 | ||
f9616c35 ED |
2072 | /* TCP Small Queues : |
2073 | * Control number of packets in qdisc/devices to two packets / or ~1 ms. | |
2074 | * (These limits are doubled for retransmits) | |
2075 | * This allows for : | |
2076 | * - better RTT estimation and ACK scheduling | |
2077 | * - faster recovery | |
2078 | * - high rates | |
2079 | * Alas, some drivers / subsystems require a fair amount | |
2080 | * of queued bytes to ensure line rate. | |
2081 | * One example is wifi aggregation (802.11 AMPDU) | |
2082 | */ | |
2083 | static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, | |
2084 | unsigned int factor) | |
2085 | { | |
2086 | unsigned int limit; | |
2087 | ||
2088 | limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10); | |
2089 | limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes); | |
2090 | limit <<= factor; | |
2091 | ||
2092 | if (atomic_read(&sk->sk_wmem_alloc) > limit) { | |
75eefc6c ED |
2093 | /* Always send the 1st or 2nd skb in write queue. |
2094 | * No need to wait for TX completion to call us back, | |
2095 | * after softirq/tasklet schedule. | |
2096 | * This helps when TX completions are delayed too much. | |
2097 | */ | |
2098 | if (skb == sk->sk_write_queue.next || | |
2099 | skb->prev == sk->sk_write_queue.next) | |
2100 | return false; | |
2101 | ||
7aa5470c | 2102 | set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); |
f9616c35 ED |
2103 | /* It is possible TX completion already happened |
2104 | * before we set TSQ_THROTTLED, so we must | |
2105 | * test again the condition. | |
2106 | */ | |
2107 | smp_mb__after_atomic(); | |
2108 | if (atomic_read(&sk->sk_wmem_alloc) > limit) | |
2109 | return true; | |
2110 | } | |
2111 | return false; | |
2112 | } | |
2113 | ||
05b055e8 FY |
2114 | static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) |
2115 | { | |
2116 | const u32 now = tcp_time_stamp; | |
2117 | ||
2118 | if (tp->chrono_type > TCP_CHRONO_UNSPEC) | |
2119 | tp->chrono_stat[tp->chrono_type - 1] += now - tp->chrono_start; | |
2120 | tp->chrono_start = now; | |
2121 | tp->chrono_type = new; | |
2122 | } | |
2123 | ||
2124 | void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) | |
2125 | { | |
2126 | struct tcp_sock *tp = tcp_sk(sk); | |
2127 | ||
2128 | /* If there are multiple conditions worthy of tracking in a | |
0f87230d FY |
2129 | * chronograph then the highest priority enum takes precedence |
2130 | * over the other conditions. So that if something "more interesting" | |
05b055e8 FY |
2131 | * starts happening, stop the previous chrono and start a new one. |
2132 | */ | |
2133 | if (type > tp->chrono_type) | |
2134 | tcp_chrono_set(tp, type); | |
2135 | } | |
2136 | ||
2137 | void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) | |
2138 | { | |
2139 | struct tcp_sock *tp = tcp_sk(sk); | |
2140 | ||
0f87230d FY |
2141 | |
2142 | /* There are multiple conditions worthy of tracking in a | |
2143 | * chronograph, so that the highest priority enum takes | |
2144 | * precedence over the other conditions (see tcp_chrono_start). | |
2145 | * If a condition stops, we only stop chrono tracking if | |
2146 | * it's the "most interesting" or current chrono we are | |
2147 | * tracking and starts busy chrono if we have pending data. | |
2148 | */ | |
2149 | if (tcp_write_queue_empty(sk)) | |
2150 | tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); | |
2151 | else if (type == tp->chrono_type) | |
2152 | tcp_chrono_set(tp, TCP_CHRONO_BUSY); | |
05b055e8 FY |
2153 | } |
2154 | ||
1da177e4 LT |
2155 | /* This routine writes packets to the network. It advances the |
2156 | * send_head. This happens as incoming acks open up the remote | |
2157 | * window for us. | |
2158 | * | |
f8269a49 IJ |
2159 | * LARGESEND note: !tcp_urg_mode is overkill, only frames between |
2160 | * snd_up-64k-mss .. snd_up cannot be large. However, taking into | |
2161 | * account rare use of URG, this is not a big flaw. | |
2162 | * | |
6ba8a3b1 ND |
2163 | * Send at most one packet when push_one > 0. Temporarily ignore |
2164 | * cwnd limit to force at most one packet out when push_one == 2. | |
2165 | ||
a2a385d6 ED |
2166 | * Returns true, if no segments are in flight and we have queued segments, |
2167 | * but cannot send anything now because of SWS or another problem. | |
1da177e4 | 2168 | */ |
a2a385d6 ED |
2169 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
2170 | int push_one, gfp_t gfp) | |
1da177e4 LT |
2171 | { |
2172 | struct tcp_sock *tp = tcp_sk(sk); | |
92df7b51 | 2173 | struct sk_buff *skb; |
c1b4a7e6 DM |
2174 | unsigned int tso_segs, sent_pkts; |
2175 | int cwnd_quota; | |
5d424d5a | 2176 | int result; |
5615f886 | 2177 | bool is_cwnd_limited = false, is_rwnd_limited = false; |
605ad7f1 | 2178 | u32 max_segs; |
1da177e4 | 2179 | |
92df7b51 | 2180 | sent_pkts = 0; |
5d424d5a | 2181 | |
d5dd9175 IJ |
2182 | if (!push_one) { |
2183 | /* Do MTU probing. */ | |
2184 | result = tcp_mtu_probe(sk); | |
2185 | if (!result) { | |
a2a385d6 | 2186 | return false; |
d5dd9175 IJ |
2187 | } else if (result > 0) { |
2188 | sent_pkts = 1; | |
2189 | } | |
5d424d5a JH |
2190 | } |
2191 | ||
ed6e7268 | 2192 | max_segs = tcp_tso_segs(sk, mss_now); |
fe067e8a | 2193 | while ((skb = tcp_send_head(sk))) { |
c8ac3774 HX |
2194 | unsigned int limit; |
2195 | ||
5bbb432c | 2196 | tso_segs = tcp_init_tso_segs(skb, mss_now); |
c1b4a7e6 | 2197 | BUG_ON(!tso_segs); |
aa93466b | 2198 | |
9d186cac | 2199 | if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { |
7faee5c0 ED |
2200 | /* "skb_mstamp" is used as a start point for the retransmit timer */ |
2201 | skb_mstamp_get(&skb->skb_mstamp); | |
ec342325 | 2202 | goto repair; /* Skip network transmission */ |
9d186cac | 2203 | } |
ec342325 | 2204 | |
b68e9f85 | 2205 | cwnd_quota = tcp_cwnd_test(tp, skb); |
6ba8a3b1 ND |
2206 | if (!cwnd_quota) { |
2207 | if (push_one == 2) | |
2208 | /* Force out a loss probe pkt. */ | |
2209 | cwnd_quota = 1; | |
2210 | else | |
2211 | break; | |
2212 | } | |
b68e9f85 | 2213 | |
5615f886 FY |
2214 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { |
2215 | is_rwnd_limited = true; | |
b68e9f85 | 2216 | break; |
5615f886 | 2217 | } |
b68e9f85 | 2218 | |
d6a4e26a | 2219 | if (tso_segs == 1) { |
c1b4a7e6 DM |
2220 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, |
2221 | (tcp_skb_is_last(sk, skb) ? | |
2222 | nonagle : TCP_NAGLE_PUSH)))) | |
2223 | break; | |
2224 | } else { | |
ca8a2263 | 2225 | if (!push_one && |
605ad7f1 ED |
2226 | tcp_tso_should_defer(sk, skb, &is_cwnd_limited, |
2227 | max_segs)) | |
c1b4a7e6 DM |
2228 | break; |
2229 | } | |
aa93466b | 2230 | |
605ad7f1 | 2231 | limit = mss_now; |
d6a4e26a | 2232 | if (tso_segs > 1 && !tcp_urg_mode(tp)) |
605ad7f1 ED |
2233 | limit = tcp_mss_split_point(sk, skb, mss_now, |
2234 | min_t(unsigned int, | |
2235 | cwnd_quota, | |
2236 | max_segs), | |
2237 | nonagle); | |
2238 | ||
2239 | if (skb->len > limit && | |
2240 | unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) | |
2241 | break; | |
2242 | ||
7aa5470c ED |
2243 | if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) |
2244 | clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags); | |
f9616c35 ED |
2245 | if (tcp_small_queue_check(sk, skb, 0)) |
2246 | break; | |
c9eeec26 | 2247 | |
d5dd9175 | 2248 | if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) |
92df7b51 | 2249 | break; |
1da177e4 | 2250 | |
ec342325 | 2251 | repair: |
92df7b51 DM |
2252 | /* Advance the send_head. This one is sent out. |
2253 | * This call will increment packets_out. | |
2254 | */ | |
66f5fe62 | 2255 | tcp_event_new_data_sent(sk, skb); |
1da177e4 | 2256 | |
92df7b51 | 2257 | tcp_minshall_update(tp, mss_now, skb); |
a262f0cd | 2258 | sent_pkts += tcp_skb_pcount(skb); |
d5dd9175 IJ |
2259 | |
2260 | if (push_one) | |
2261 | break; | |
92df7b51 | 2262 | } |
1da177e4 | 2263 | |
5615f886 FY |
2264 | if (is_rwnd_limited) |
2265 | tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); | |
2266 | else | |
2267 | tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); | |
2268 | ||
aa93466b | 2269 | if (likely(sent_pkts)) { |
684bad11 YC |
2270 | if (tcp_in_cwnd_reduction(sk)) |
2271 | tp->prr_out += sent_pkts; | |
6ba8a3b1 ND |
2272 | |
2273 | /* Send one loss probe per tail loss episode. */ | |
2274 | if (push_one != 2) | |
2275 | tcp_schedule_loss_probe(sk); | |
d2e1339f | 2276 | is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); |
ca8a2263 | 2277 | tcp_cwnd_validate(sk, is_cwnd_limited); |
a2a385d6 | 2278 | return false; |
1da177e4 | 2279 | } |
b340b264 | 2280 | return !tp->packets_out && tcp_send_head(sk); |
6ba8a3b1 ND |
2281 | } |
2282 | ||
2283 | bool tcp_schedule_loss_probe(struct sock *sk) | |
2284 | { | |
2285 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2286 | struct tcp_sock *tp = tcp_sk(sk); | |
2287 | u32 timeout, tlp_time_stamp, rto_time_stamp; | |
740b0f18 | 2288 | u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3); |
6ba8a3b1 | 2289 | |
6ba8a3b1 ND |
2290 | /* No consecutive loss probes. */ |
2291 | if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) { | |
2292 | tcp_rearm_rto(sk); | |
2293 | return false; | |
2294 | } | |
2295 | /* Don't do any loss probe on a Fast Open connection before 3WHS | |
2296 | * finishes. | |
2297 | */ | |
f9b99582 | 2298 | if (tp->fastopen_rsk) |
6ba8a3b1 ND |
2299 | return false; |
2300 | ||
2301 | /* TLP is only scheduled when next timer event is RTO. */ | |
2302 | if (icsk->icsk_pending != ICSK_TIME_RETRANS) | |
2303 | return false; | |
2304 | ||
2305 | /* Schedule a loss probe in 2*RTT for SACK capable connections | |
2306 | * in Open state, that are either limited by cwnd or application. | |
2307 | */ | |
bec41a11 YC |
2308 | if ((sysctl_tcp_early_retrans != 3 && sysctl_tcp_early_retrans != 4) || |
2309 | !tp->packets_out || !tcp_is_sack(tp) || | |
2310 | icsk->icsk_ca_state != TCP_CA_Open) | |
6ba8a3b1 ND |
2311 | return false; |
2312 | ||
2313 | if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) && | |
2314 | tcp_send_head(sk)) | |
2315 | return false; | |
2316 | ||
2317 | /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account | |
f9b99582 YC |
2318 | * for delayed ack when there's one outstanding packet. If no RTT |
2319 | * sample is available then probe after TCP_TIMEOUT_INIT. | |
6ba8a3b1 | 2320 | */ |
f9b99582 | 2321 | timeout = rtt << 1 ? : TCP_TIMEOUT_INIT; |
6ba8a3b1 ND |
2322 | if (tp->packets_out == 1) |
2323 | timeout = max_t(u32, timeout, | |
2324 | (rtt + (rtt >> 1) + TCP_DELACK_MAX)); | |
2325 | timeout = max_t(u32, timeout, msecs_to_jiffies(10)); | |
2326 | ||
2327 | /* If RTO is shorter, just schedule TLP in its place. */ | |
2328 | tlp_time_stamp = tcp_time_stamp + timeout; | |
2329 | rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout; | |
2330 | if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) { | |
2331 | s32 delta = rto_time_stamp - tcp_time_stamp; | |
2332 | if (delta > 0) | |
2333 | timeout = delta; | |
2334 | } | |
2335 | ||
2336 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, | |
2337 | TCP_RTO_MAX); | |
2338 | return true; | |
2339 | } | |
2340 | ||
1f3279ae ED |
2341 | /* Thanks to skb fast clones, we can detect if a prior transmit of |
2342 | * a packet is still in a qdisc or driver queue. | |
2343 | * In this case, there is very little point doing a retransmit ! | |
1f3279ae ED |
2344 | */ |
2345 | static bool skb_still_in_host_queue(const struct sock *sk, | |
2346 | const struct sk_buff *skb) | |
2347 | { | |
39bb5e62 | 2348 | if (unlikely(skb_fclone_busy(sk, skb))) { |
c10d9310 ED |
2349 | NET_INC_STATS(sock_net(sk), |
2350 | LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); | |
1f3279ae ED |
2351 | return true; |
2352 | } | |
2353 | return false; | |
2354 | } | |
2355 | ||
b340b264 | 2356 | /* When probe timeout (PTO) fires, try send a new segment if possible, else |
6ba8a3b1 ND |
2357 | * retransmit the last segment. |
2358 | */ | |
2359 | void tcp_send_loss_probe(struct sock *sk) | |
2360 | { | |
9b717a8d | 2361 | struct tcp_sock *tp = tcp_sk(sk); |
6ba8a3b1 ND |
2362 | struct sk_buff *skb; |
2363 | int pcount; | |
2364 | int mss = tcp_current_mss(sk); | |
6ba8a3b1 | 2365 | |
b340b264 YC |
2366 | skb = tcp_send_head(sk); |
2367 | if (skb) { | |
2368 | if (tcp_snd_wnd_test(tp, skb, mss)) { | |
2369 | pcount = tp->packets_out; | |
2370 | tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); | |
2371 | if (tp->packets_out > pcount) | |
2372 | goto probe_sent; | |
2373 | goto rearm_timer; | |
2374 | } | |
2375 | skb = tcp_write_queue_prev(sk, skb); | |
2376 | } else { | |
2377 | skb = tcp_write_queue_tail(sk); | |
6ba8a3b1 ND |
2378 | } |
2379 | ||
9b717a8d ND |
2380 | /* At most one outstanding TLP retransmission. */ |
2381 | if (tp->tlp_high_seq) | |
2382 | goto rearm_timer; | |
2383 | ||
6ba8a3b1 | 2384 | /* Retransmit last segment. */ |
6ba8a3b1 ND |
2385 | if (WARN_ON(!skb)) |
2386 | goto rearm_timer; | |
2387 | ||
1f3279ae ED |
2388 | if (skb_still_in_host_queue(sk, skb)) |
2389 | goto rearm_timer; | |
2390 | ||
6ba8a3b1 ND |
2391 | pcount = tcp_skb_pcount(skb); |
2392 | if (WARN_ON(!pcount)) | |
2393 | goto rearm_timer; | |
2394 | ||
2395 | if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { | |
6cc55e09 OP |
2396 | if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss, |
2397 | GFP_ATOMIC))) | |
6ba8a3b1 | 2398 | goto rearm_timer; |
b340b264 | 2399 | skb = tcp_write_queue_next(sk, skb); |
6ba8a3b1 ND |
2400 | } |
2401 | ||
2402 | if (WARN_ON(!skb || !tcp_skb_pcount(skb))) | |
2403 | goto rearm_timer; | |
2404 | ||
10d3be56 | 2405 | if (__tcp_retransmit_skb(sk, skb, 1)) |
b340b264 | 2406 | goto rearm_timer; |
6ba8a3b1 | 2407 | |
9b717a8d | 2408 | /* Record snd_nxt for loss detection. */ |
b340b264 | 2409 | tp->tlp_high_seq = tp->snd_nxt; |
9b717a8d | 2410 | |
b340b264 | 2411 | probe_sent: |
c10d9310 | 2412 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); |
b340b264 YC |
2413 | /* Reset s.t. tcp_rearm_rto will restart timer from now */ |
2414 | inet_csk(sk)->icsk_pending = 0; | |
6ba8a3b1 | 2415 | rearm_timer: |
fcd16c0a | 2416 | tcp_rearm_rto(sk); |
1da177e4 LT |
2417 | } |
2418 | ||
a762a980 DM |
2419 | /* Push out any pending frames which were held back due to |
2420 | * TCP_CORK or attempt at coalescing tiny packets. | |
2421 | * The socket must be locked by the caller. | |
2422 | */ | |
9e412ba7 IJ |
2423 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, |
2424 | int nonagle) | |
a762a980 | 2425 | { |
726e07a8 IJ |
2426 | /* If we are closed, the bytes will have to remain here. |
2427 | * In time closedown will finish, we empty the write queue and | |
2428 | * all will be happy. | |
2429 | */ | |
2430 | if (unlikely(sk->sk_state == TCP_CLOSE)) | |
2431 | return; | |
2432 | ||
99a1dec7 | 2433 | if (tcp_write_xmit(sk, cur_mss, nonagle, 0, |
7450aaf6 | 2434 | sk_gfp_mask(sk, GFP_ATOMIC))) |
726e07a8 | 2435 | tcp_check_probe_timer(sk); |
a762a980 DM |
2436 | } |
2437 | ||
c1b4a7e6 DM |
2438 | /* Send _single_ skb sitting at the send head. This function requires |
2439 | * true push pending frames to setup probe timer etc. | |
2440 | */ | |
2441 | void tcp_push_one(struct sock *sk, unsigned int mss_now) | |
2442 | { | |
fe067e8a | 2443 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 DM |
2444 | |
2445 | BUG_ON(!skb || skb->len < mss_now); | |
2446 | ||
d5dd9175 | 2447 | tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); |
c1b4a7e6 DM |
2448 | } |
2449 | ||
1da177e4 LT |
2450 | /* This function returns the amount that we can raise the |
2451 | * usable window based on the following constraints | |
e905a9ed | 2452 | * |
1da177e4 LT |
2453 | * 1. The window can never be shrunk once it is offered (RFC 793) |
2454 | * 2. We limit memory per socket | |
2455 | * | |
2456 | * RFC 1122: | |
2457 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep | |
2458 | * RECV.NEXT + RCV.WIN fixed until: | |
2459 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" | |
2460 | * | |
2461 | * i.e. don't raise the right edge of the window until you can raise | |
2462 | * it at least MSS bytes. | |
2463 | * | |
2464 | * Unfortunately, the recommended algorithm breaks header prediction, | |
2465 | * since header prediction assumes th->window stays fixed. | |
2466 | * | |
2467 | * Strictly speaking, keeping th->window fixed violates the receiver | |
2468 | * side SWS prevention criteria. The problem is that under this rule | |
2469 | * a stream of single byte packets will cause the right side of the | |
2470 | * window to always advance by a single byte. | |
e905a9ed | 2471 | * |
1da177e4 LT |
2472 | * Of course, if the sender implements sender side SWS prevention |
2473 | * then this will not be a problem. | |
e905a9ed | 2474 | * |
1da177e4 | 2475 | * BSD seems to make the following compromise: |
e905a9ed | 2476 | * |
1da177e4 LT |
2477 | * If the free space is less than the 1/4 of the maximum |
2478 | * space available and the free space is less than 1/2 mss, | |
2479 | * then set the window to 0. | |
2480 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] | |
2481 | * Otherwise, just prevent the window from shrinking | |
2482 | * and from being larger than the largest representable value. | |
2483 | * | |
2484 | * This prevents incremental opening of the window in the regime | |
2485 | * where TCP is limited by the speed of the reader side taking | |
2486 | * data out of the TCP receive queue. It does nothing about | |
2487 | * those cases where the window is constrained on the sender side | |
2488 | * because the pipeline is full. | |
2489 | * | |
2490 | * BSD also seems to "accidentally" limit itself to windows that are a | |
2491 | * multiple of MSS, at least until the free space gets quite small. | |
2492 | * This would appear to be a side effect of the mbuf implementation. | |
2493 | * Combining these two algorithms results in the observed behavior | |
2494 | * of having a fixed window size at almost all times. | |
2495 | * | |
2496 | * Below we obtain similar behavior by forcing the offered window to | |
2497 | * a multiple of the mss when it is feasible to do so. | |
2498 | * | |
2499 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. | |
2500 | * Regular options like TIMESTAMP are taken into account. | |
2501 | */ | |
2502 | u32 __tcp_select_window(struct sock *sk) | |
2503 | { | |
463c84b9 | 2504 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 2505 | struct tcp_sock *tp = tcp_sk(sk); |
caa20d9a | 2506 | /* MSS for the peer's data. Previous versions used mss_clamp |
1da177e4 LT |
2507 | * here. I don't know if the value based on our guesses |
2508 | * of peer's MSS is better for the performance. It's more correct | |
2509 | * but may be worse for the performance because of rcv_mss | |
2510 | * fluctuations. --SAW 1998/11/1 | |
2511 | */ | |
463c84b9 | 2512 | int mss = icsk->icsk_ack.rcv_mss; |
1da177e4 | 2513 | int free_space = tcp_space(sk); |
86c1a045 FW |
2514 | int allowed_space = tcp_full_space(sk); |
2515 | int full_space = min_t(int, tp->window_clamp, allowed_space); | |
1da177e4 LT |
2516 | int window; |
2517 | ||
06425c30 | 2518 | if (unlikely(mss > full_space)) { |
e905a9ed | 2519 | mss = full_space; |
06425c30 ED |
2520 | if (mss <= 0) |
2521 | return 0; | |
2522 | } | |
b92edbe0 | 2523 | if (free_space < (full_space >> 1)) { |
463c84b9 | 2524 | icsk->icsk_ack.quick = 0; |
1da177e4 | 2525 | |
b8da51eb | 2526 | if (tcp_under_memory_pressure(sk)) |
056834d9 IJ |
2527 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, |
2528 | 4U * tp->advmss); | |
1da177e4 | 2529 | |
86c1a045 FW |
2530 | /* free_space might become our new window, make sure we don't |
2531 | * increase it due to wscale. | |
2532 | */ | |
2533 | free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); | |
2534 | ||
2535 | /* if free space is less than mss estimate, or is below 1/16th | |
2536 | * of the maximum allowed, try to move to zero-window, else | |
2537 | * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and | |
2538 | * new incoming data is dropped due to memory limits. | |
2539 | * With large window, mss test triggers way too late in order | |
2540 | * to announce zero window in time before rmem limit kicks in. | |
2541 | */ | |
2542 | if (free_space < (allowed_space >> 4) || free_space < mss) | |
1da177e4 LT |
2543 | return 0; |
2544 | } | |
2545 | ||
2546 | if (free_space > tp->rcv_ssthresh) | |
2547 | free_space = tp->rcv_ssthresh; | |
2548 | ||
2549 | /* Don't do rounding if we are using window scaling, since the | |
2550 | * scaled window will not line up with the MSS boundary anyway. | |
2551 | */ | |
2552 | window = tp->rcv_wnd; | |
2553 | if (tp->rx_opt.rcv_wscale) { | |
2554 | window = free_space; | |
2555 | ||
2556 | /* Advertise enough space so that it won't get scaled away. | |
2557 | * Import case: prevent zero window announcement if | |
2558 | * 1<<rcv_wscale > mss. | |
2559 | */ | |
2560 | if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window) | |
2561 | window = (((window >> tp->rx_opt.rcv_wscale) + 1) | |
2562 | << tp->rx_opt.rcv_wscale); | |
2563 | } else { | |
2564 | /* Get the largest window that is a nice multiple of mss. | |
2565 | * Window clamp already applied above. | |
2566 | * If our current window offering is within 1 mss of the | |
2567 | * free space we just keep it. This prevents the divide | |
2568 | * and multiply from happening most of the time. | |
2569 | * We also don't do any window rounding when the free space | |
2570 | * is too small. | |
2571 | */ | |
2572 | if (window <= free_space - mss || window > free_space) | |
056834d9 | 2573 | window = (free_space / mss) * mss; |
84565070 | 2574 | else if (mss == full_space && |
b92edbe0 | 2575 | free_space > window + (full_space >> 1)) |
84565070 | 2576 | window = free_space; |
1da177e4 LT |
2577 | } |
2578 | ||
2579 | return window; | |
2580 | } | |
2581 | ||
cfea5a68 MKL |
2582 | void tcp_skb_collapse_tstamp(struct sk_buff *skb, |
2583 | const struct sk_buff *next_skb) | |
082ac2d5 | 2584 | { |
0a2cf20c SHY |
2585 | if (unlikely(tcp_has_tx_tstamp(next_skb))) { |
2586 | const struct skb_shared_info *next_shinfo = | |
2587 | skb_shinfo(next_skb); | |
082ac2d5 MKL |
2588 | struct skb_shared_info *shinfo = skb_shinfo(skb); |
2589 | ||
0a2cf20c | 2590 | shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; |
082ac2d5 | 2591 | shinfo->tskey = next_shinfo->tskey; |
2de8023e MKL |
2592 | TCP_SKB_CB(skb)->txstamp_ack |= |
2593 | TCP_SKB_CB(next_skb)->txstamp_ack; | |
082ac2d5 MKL |
2594 | } |
2595 | } | |
2596 | ||
4a17fc3a | 2597 | /* Collapses two adjacent SKB's during retransmission. */ |
f8071cde | 2598 | static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) |
1da177e4 LT |
2599 | { |
2600 | struct tcp_sock *tp = tcp_sk(sk); | |
fe067e8a | 2601 | struct sk_buff *next_skb = tcp_write_queue_next(sk, skb); |
058dc334 | 2602 | int skb_size, next_skb_size; |
1da177e4 | 2603 | |
058dc334 IJ |
2604 | skb_size = skb->len; |
2605 | next_skb_size = next_skb->len; | |
1da177e4 | 2606 | |
058dc334 | 2607 | BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); |
a6963a6b | 2608 | |
f8071cde ED |
2609 | if (next_skb_size) { |
2610 | if (next_skb_size <= skb_availroom(skb)) | |
2611 | skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), | |
2612 | next_skb_size); | |
2613 | else if (!skb_shift(skb, next_skb, next_skb_size)) | |
2614 | return false; | |
2615 | } | |
058dc334 | 2616 | tcp_highest_sack_combine(sk, next_skb, skb); |
1da177e4 | 2617 | |
058dc334 | 2618 | tcp_unlink_write_queue(next_skb, sk); |
1da177e4 | 2619 | |
058dc334 IJ |
2620 | if (next_skb->ip_summed == CHECKSUM_PARTIAL) |
2621 | skb->ip_summed = CHECKSUM_PARTIAL; | |
1da177e4 | 2622 | |
058dc334 IJ |
2623 | if (skb->ip_summed != CHECKSUM_PARTIAL) |
2624 | skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size); | |
1da177e4 | 2625 | |
058dc334 IJ |
2626 | /* Update sequence range on original skb. */ |
2627 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; | |
1da177e4 | 2628 | |
e6c7d085 | 2629 | /* Merge over control information. This moves PSH/FIN etc. over */ |
4de075e0 | 2630 | TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; |
058dc334 IJ |
2631 | |
2632 | /* All done, get rid of second SKB and account for it so | |
2633 | * packet counting does not break. | |
2634 | */ | |
2635 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; | |
a643b5d4 | 2636 | TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; |
058dc334 IJ |
2637 | |
2638 | /* changed transmit queue under us so clear hints */ | |
ef9da47c IJ |
2639 | tcp_clear_retrans_hints_partial(tp); |
2640 | if (next_skb == tp->retransmit_skb_hint) | |
2641 | tp->retransmit_skb_hint = skb; | |
058dc334 | 2642 | |
797108d1 IJ |
2643 | tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); |
2644 | ||
082ac2d5 MKL |
2645 | tcp_skb_collapse_tstamp(skb, next_skb); |
2646 | ||
058dc334 | 2647 | sk_wmem_free_skb(sk, next_skb); |
f8071cde | 2648 | return true; |
1da177e4 LT |
2649 | } |
2650 | ||
67edfef7 | 2651 | /* Check if coalescing SKBs is legal. */ |
a2a385d6 | 2652 | static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) |
4a17fc3a IJ |
2653 | { |
2654 | if (tcp_skb_pcount(skb) > 1) | |
a2a385d6 | 2655 | return false; |
4a17fc3a | 2656 | if (skb_cloned(skb)) |
a2a385d6 | 2657 | return false; |
4a17fc3a | 2658 | if (skb == tcp_send_head(sk)) |
a2a385d6 | 2659 | return false; |
2331ccc5 | 2660 | /* Some heuristics for collapsing over SACK'd could be invented */ |
4a17fc3a | 2661 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) |
a2a385d6 | 2662 | return false; |
4a17fc3a | 2663 | |
a2a385d6 | 2664 | return true; |
4a17fc3a IJ |
2665 | } |
2666 | ||
67edfef7 AK |
2667 | /* Collapse packets in the retransmit queue to make to create |
2668 | * less packets on the wire. This is only done on retransmission. | |
2669 | */ | |
4a17fc3a IJ |
2670 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, |
2671 | int space) | |
2672 | { | |
2673 | struct tcp_sock *tp = tcp_sk(sk); | |
2674 | struct sk_buff *skb = to, *tmp; | |
a2a385d6 | 2675 | bool first = true; |
4a17fc3a IJ |
2676 | |
2677 | if (!sysctl_tcp_retrans_collapse) | |
2678 | return; | |
4de075e0 | 2679 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) |
4a17fc3a IJ |
2680 | return; |
2681 | ||
2682 | tcp_for_write_queue_from_safe(skb, tmp, sk) { | |
2683 | if (!tcp_can_collapse(sk, skb)) | |
2684 | break; | |
2685 | ||
a643b5d4 MKL |
2686 | if (!tcp_skb_can_collapse_to(to)) |
2687 | break; | |
2688 | ||
4a17fc3a IJ |
2689 | space -= skb->len; |
2690 | ||
2691 | if (first) { | |
a2a385d6 | 2692 | first = false; |
4a17fc3a IJ |
2693 | continue; |
2694 | } | |
2695 | ||
2696 | if (space < 0) | |
2697 | break; | |
4a17fc3a IJ |
2698 | |
2699 | if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) | |
2700 | break; | |
2701 | ||
f8071cde ED |
2702 | if (!tcp_collapse_retrans(sk, to)) |
2703 | break; | |
4a17fc3a IJ |
2704 | } |
2705 | } | |
2706 | ||
1da177e4 LT |
2707 | /* This retransmits one SKB. Policy decisions and retransmit queue |
2708 | * state updates are done by the caller. Returns non-zero if an | |
2709 | * error occurred which prevented the send. | |
2710 | */ | |
10d3be56 | 2711 | int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) |
1da177e4 | 2712 | { |
5d424d5a | 2713 | struct inet_connection_sock *icsk = inet_csk(sk); |
10d3be56 | 2714 | struct tcp_sock *tp = tcp_sk(sk); |
7d227cd2 | 2715 | unsigned int cur_mss; |
10d3be56 ED |
2716 | int diff, len, err; |
2717 | ||
1da177e4 | 2718 | |
10d3be56 ED |
2719 | /* Inconclusive MTU probe */ |
2720 | if (icsk->icsk_mtup.probe_size) | |
5d424d5a | 2721 | icsk->icsk_mtup.probe_size = 0; |
5d424d5a | 2722 | |
1da177e4 | 2723 | /* Do not sent more than we queued. 1/4 is reserved for possible |
caa20d9a | 2724 | * copying overhead: fragmentation, tunneling, mangling etc. |
1da177e4 LT |
2725 | */ |
2726 | if (atomic_read(&sk->sk_wmem_alloc) > | |
ffb4d6c8 ED |
2727 | min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), |
2728 | sk->sk_sndbuf)) | |
1da177e4 LT |
2729 | return -EAGAIN; |
2730 | ||
1f3279ae ED |
2731 | if (skb_still_in_host_queue(sk, skb)) |
2732 | return -EBUSY; | |
2733 | ||
1da177e4 LT |
2734 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { |
2735 | if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una)) | |
2736 | BUG(); | |
1da177e4 LT |
2737 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
2738 | return -ENOMEM; | |
2739 | } | |
2740 | ||
7d227cd2 SS |
2741 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) |
2742 | return -EHOSTUNREACH; /* Routing failure or similar. */ | |
2743 | ||
0c54b85f | 2744 | cur_mss = tcp_current_mss(sk); |
7d227cd2 | 2745 | |
1da177e4 LT |
2746 | /* If receiver has shrunk his window, and skb is out of |
2747 | * new window, do not retransmit it. The exception is the | |
2748 | * case, when window is shrunk to zero. In this case | |
2749 | * our retransmit serves as a zero window probe. | |
2750 | */ | |
9d4fb27d JP |
2751 | if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && |
2752 | TCP_SKB_CB(skb)->seq != tp->snd_una) | |
1da177e4 LT |
2753 | return -EAGAIN; |
2754 | ||
10d3be56 ED |
2755 | len = cur_mss * segs; |
2756 | if (skb->len > len) { | |
2757 | if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC)) | |
1da177e4 | 2758 | return -ENOMEM; /* We'll try again later. */ |
02276f3c | 2759 | } else { |
10d3be56 ED |
2760 | if (skb_unclone(skb, GFP_ATOMIC)) |
2761 | return -ENOMEM; | |
9eb9362e | 2762 | |
10d3be56 ED |
2763 | diff = tcp_skb_pcount(skb); |
2764 | tcp_set_skb_tso_segs(skb, cur_mss); | |
2765 | diff -= tcp_skb_pcount(skb); | |
2766 | if (diff) | |
2767 | tcp_adjust_pcount(sk, skb, diff); | |
2768 | if (skb->len < cur_mss) | |
2769 | tcp_retrans_try_collapse(sk, skb, cur_mss); | |
1da177e4 LT |
2770 | } |
2771 | ||
49213555 DB |
2772 | /* RFC3168, section 6.1.1.1. ECN fallback */ |
2773 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) | |
2774 | tcp_ecn_clear_syn(sk, skb); | |
2775 | ||
678550c6 YC |
2776 | /* Update global and local TCP statistics. */ |
2777 | segs = tcp_skb_pcount(skb); | |
2778 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); | |
2779 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) | |
2780 | __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); | |
2781 | tp->total_retrans += segs; | |
2782 | ||
50bceae9 TG |
2783 | /* make sure skb->data is aligned on arches that require it |
2784 | * and check if ack-trimming & collapsing extended the headroom | |
2785 | * beyond what csum_start can cover. | |
2786 | */ | |
2787 | if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || | |
2788 | skb_headroom(skb) >= 0xFFFF)) { | |
10a81980 ED |
2789 | struct sk_buff *nskb; |
2790 | ||
2791 | skb_mstamp_get(&skb->skb_mstamp); | |
2792 | nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); | |
c84a5711 YC |
2793 | err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) : |
2794 | -ENOBUFS; | |
117632e6 | 2795 | } else { |
c84a5711 | 2796 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
117632e6 | 2797 | } |
c84a5711 | 2798 | |
fc9f3501 | 2799 | if (likely(!err)) { |
c84a5711 | 2800 | TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; |
678550c6 YC |
2801 | } else if (err != -EBUSY) { |
2802 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL); | |
fc9f3501 | 2803 | } |
c84a5711 | 2804 | return err; |
93b174ad YC |
2805 | } |
2806 | ||
10d3be56 | 2807 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) |
93b174ad YC |
2808 | { |
2809 | struct tcp_sock *tp = tcp_sk(sk); | |
10d3be56 | 2810 | int err = __tcp_retransmit_skb(sk, skb, segs); |
1da177e4 LT |
2811 | |
2812 | if (err == 0) { | |
1da177e4 | 2813 | #if FASTRETRANS_DEBUG > 0 |
056834d9 | 2814 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
e87cc472 | 2815 | net_dbg_ratelimited("retrans_out leaked\n"); |
1da177e4 LT |
2816 | } |
2817 | #endif | |
2818 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; | |
2819 | tp->retrans_out += tcp_skb_pcount(skb); | |
2820 | ||
2821 | /* Save stamp of the first retransmit. */ | |
2822 | if (!tp->retrans_stamp) | |
7faee5c0 | 2823 | tp->retrans_stamp = tcp_skb_timestamp(skb); |
1da177e4 | 2824 | |
1da177e4 | 2825 | } |
6e08d5e3 YC |
2826 | |
2827 | if (tp->undo_retrans < 0) | |
2828 | tp->undo_retrans = 0; | |
2829 | tp->undo_retrans += tcp_skb_pcount(skb); | |
1da177e4 LT |
2830 | return err; |
2831 | } | |
2832 | ||
2833 | /* This gets called after a retransmit timeout, and the initially | |
2834 | * retransmitted data is acknowledged. It tries to continue | |
2835 | * resending the rest of the retransmit queue, until either | |
2836 | * we've sent it all or the congestion window limit is reached. | |
2837 | * If doing SACK, the first ACK which comes back for a timeout | |
2838 | * based retransmit packet might feed us FACK information again. | |
2839 | * If so, we use it to avoid unnecessarily retransmissions. | |
2840 | */ | |
2841 | void tcp_xmit_retransmit_queue(struct sock *sk) | |
2842 | { | |
6687e988 | 2843 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 LT |
2844 | struct tcp_sock *tp = tcp_sk(sk); |
2845 | struct sk_buff *skb; | |
0e1c54c2 | 2846 | struct sk_buff *hole = NULL; |
840a3cbe | 2847 | u32 max_segs; |
61eb55f4 | 2848 | int mib_idx; |
6a438bbe | 2849 | |
45e77d31 IJ |
2850 | if (!tp->packets_out) |
2851 | return; | |
2852 | ||
618d9f25 | 2853 | if (tp->retransmit_skb_hint) { |
6a438bbe | 2854 | skb = tp->retransmit_skb_hint; |
618d9f25 | 2855 | } else { |
fe067e8a | 2856 | skb = tcp_write_queue_head(sk); |
618d9f25 | 2857 | } |
1da177e4 | 2858 | |
ed6e7268 | 2859 | max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); |
08ebd172 | 2860 | tcp_for_write_queue_from(skb, sk) { |
dca0aaf8 | 2861 | __u8 sacked; |
10d3be56 | 2862 | int segs; |
1da177e4 | 2863 | |
08ebd172 IJ |
2864 | if (skb == tcp_send_head(sk)) |
2865 | break; | |
2866 | /* we could do better than to assign each time */ | |
51456b29 | 2867 | if (!hole) |
0e1c54c2 | 2868 | tp->retransmit_skb_hint = skb; |
08ebd172 | 2869 | |
10d3be56 ED |
2870 | segs = tp->snd_cwnd - tcp_packets_in_flight(tp); |
2871 | if (segs <= 0) | |
08ebd172 | 2872 | return; |
dca0aaf8 | 2873 | sacked = TCP_SKB_CB(skb)->sacked; |
a3d2e9f8 ED |
2874 | /* In case tcp_shift_skb_data() have aggregated large skbs, |
2875 | * we need to make sure not sending too bigs TSO packets | |
2876 | */ | |
2877 | segs = min_t(int, segs, max_segs); | |
1da177e4 | 2878 | |
840a3cbe YC |
2879 | if (tp->retrans_out >= tp->lost_out) { |
2880 | break; | |
0e1c54c2 | 2881 | } else if (!(sacked & TCPCB_LOST)) { |
51456b29 | 2882 | if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) |
0e1c54c2 IJ |
2883 | hole = skb; |
2884 | continue; | |
1da177e4 | 2885 | |
0e1c54c2 IJ |
2886 | } else { |
2887 | if (icsk->icsk_ca_state != TCP_CA_Loss) | |
2888 | mib_idx = LINUX_MIB_TCPFASTRETRANS; | |
2889 | else | |
2890 | mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; | |
2891 | } | |
1da177e4 | 2892 | |
0e1c54c2 | 2893 | if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) |
1da177e4 LT |
2894 | continue; |
2895 | ||
f9616c35 ED |
2896 | if (tcp_small_queue_check(sk, skb, 1)) |
2897 | return; | |
2898 | ||
10d3be56 | 2899 | if (tcp_retransmit_skb(sk, skb, segs)) |
0e1c54c2 | 2900 | return; |
24ab6bec | 2901 | |
de1d6578 | 2902 | NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); |
1da177e4 | 2903 | |
684bad11 | 2904 | if (tcp_in_cwnd_reduction(sk)) |
a262f0cd ND |
2905 | tp->prr_out += tcp_skb_pcount(skb); |
2906 | ||
57dde7f7 YC |
2907 | if (skb == tcp_write_queue_head(sk) && |
2908 | icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) | |
3f421baa ACM |
2909 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
2910 | inet_csk(sk)->icsk_rto, | |
2911 | TCP_RTO_MAX); | |
1da177e4 LT |
2912 | } |
2913 | } | |
2914 | ||
d83769a5 ED |
2915 | /* We allow to exceed memory limits for FIN packets to expedite |
2916 | * connection tear down and (memory) recovery. | |
845704a5 ED |
2917 | * Otherwise tcp_send_fin() could be tempted to either delay FIN |
2918 | * or even be forced to close flow without any FIN. | |
a6c5ea4c ED |
2919 | * In general, we want to allow one skb per socket to avoid hangs |
2920 | * with edge trigger epoll() | |
d83769a5 | 2921 | */ |
a6c5ea4c | 2922 | void sk_forced_mem_schedule(struct sock *sk, int size) |
d83769a5 | 2923 | { |
e805605c | 2924 | int amt; |
d83769a5 ED |
2925 | |
2926 | if (size <= sk->sk_forward_alloc) | |
2927 | return; | |
2928 | amt = sk_mem_pages(size); | |
2929 | sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; | |
e805605c JW |
2930 | sk_memory_allocated_add(sk, amt); |
2931 | ||
baac50bb JW |
2932 | if (mem_cgroup_sockets_enabled && sk->sk_memcg) |
2933 | mem_cgroup_charge_skmem(sk->sk_memcg, amt); | |
d83769a5 ED |
2934 | } |
2935 | ||
845704a5 ED |
2936 | /* Send a FIN. The caller locks the socket for us. |
2937 | * We should try to send a FIN packet really hard, but eventually give up. | |
1da177e4 LT |
2938 | */ |
2939 | void tcp_send_fin(struct sock *sk) | |
2940 | { | |
845704a5 | 2941 | struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk); |
e905a9ed | 2942 | struct tcp_sock *tp = tcp_sk(sk); |
e905a9ed | 2943 | |
845704a5 ED |
2944 | /* Optimization, tack on the FIN if we have one skb in write queue and |
2945 | * this skb was not yet sent, or we are under memory pressure. | |
2946 | * Note: in the latter case, FIN packet will be sent after a timeout, | |
2947 | * as TCP stack thinks it has already been transmitted. | |
1da177e4 | 2948 | */ |
b8da51eb | 2949 | if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) { |
845704a5 ED |
2950 | coalesce: |
2951 | TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; | |
2952 | TCP_SKB_CB(tskb)->end_seq++; | |
1da177e4 | 2953 | tp->write_seq++; |
845704a5 ED |
2954 | if (!tcp_send_head(sk)) { |
2955 | /* This means tskb was already sent. | |
2956 | * Pretend we included the FIN on previous transmit. | |
2957 | * We need to set tp->snd_nxt to the value it would have | |
2958 | * if FIN had been sent. This is because retransmit path | |
2959 | * does not change tp->snd_nxt. | |
2960 | */ | |
2961 | tp->snd_nxt++; | |
2962 | return; | |
2963 | } | |
1da177e4 | 2964 | } else { |
845704a5 ED |
2965 | skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); |
2966 | if (unlikely(!skb)) { | |
2967 | if (tskb) | |
2968 | goto coalesce; | |
2969 | return; | |
1da177e4 | 2970 | } |
d83769a5 | 2971 | skb_reserve(skb, MAX_TCP_HEADER); |
a6c5ea4c | 2972 | sk_forced_mem_schedule(sk, skb->truesize); |
1da177e4 | 2973 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ |
e870a8ef | 2974 | tcp_init_nondata_skb(skb, tp->write_seq, |
a3433f35 | 2975 | TCPHDR_ACK | TCPHDR_FIN); |
1da177e4 LT |
2976 | tcp_queue_skb(sk, skb); |
2977 | } | |
845704a5 | 2978 | __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); |
1da177e4 LT |
2979 | } |
2980 | ||
2981 | /* We get here when a process closes a file descriptor (either due to | |
2982 | * an explicit close() or as a byproduct of exit()'ing) and there | |
2983 | * was unread data in the receive queue. This behavior is recommended | |
65bb723c | 2984 | * by RFC 2525, section 2.17. -DaveM |
1da177e4 | 2985 | */ |
dd0fc66f | 2986 | void tcp_send_active_reset(struct sock *sk, gfp_t priority) |
1da177e4 | 2987 | { |
1da177e4 LT |
2988 | struct sk_buff *skb; |
2989 | ||
2990 | /* NOTE: No TCP options attached and we never retransmit this. */ | |
2991 | skb = alloc_skb(MAX_TCP_HEADER, priority); | |
2992 | if (!skb) { | |
4e673444 | 2993 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
1da177e4 LT |
2994 | return; |
2995 | } | |
2996 | ||
2997 | /* Reserve space for headers and prepare control bits. */ | |
2998 | skb_reserve(skb, MAX_TCP_HEADER); | |
e870a8ef | 2999 | tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), |
a3433f35 | 3000 | TCPHDR_ACK | TCPHDR_RST); |
675ee231 | 3001 | skb_mstamp_get(&skb->skb_mstamp); |
1da177e4 | 3002 | /* Send it off. */ |
dfb4b9dc | 3003 | if (tcp_transmit_skb(sk, skb, 0, priority)) |
4e673444 | 3004 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
26af65cb | 3005 | |
81cc8a75 | 3006 | TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); |
1da177e4 LT |
3007 | } |
3008 | ||
67edfef7 AK |
3009 | /* Send a crossed SYN-ACK during socket establishment. |
3010 | * WARNING: This routine must only be called when we have already sent | |
1da177e4 LT |
3011 | * a SYN packet that crossed the incoming SYN that caused this routine |
3012 | * to get called. If this assumption fails then the initial rcv_wnd | |
3013 | * and rcv_wscale values will not be correct. | |
3014 | */ | |
3015 | int tcp_send_synack(struct sock *sk) | |
3016 | { | |
056834d9 | 3017 | struct sk_buff *skb; |
1da177e4 | 3018 | |
fe067e8a | 3019 | skb = tcp_write_queue_head(sk); |
51456b29 | 3020 | if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { |
91df42be | 3021 | pr_debug("%s: wrong queue state\n", __func__); |
1da177e4 LT |
3022 | return -EFAULT; |
3023 | } | |
4de075e0 | 3024 | if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { |
1da177e4 LT |
3025 | if (skb_cloned(skb)) { |
3026 | struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC); | |
51456b29 | 3027 | if (!nskb) |
1da177e4 | 3028 | return -ENOMEM; |
fe067e8a | 3029 | tcp_unlink_write_queue(skb, sk); |
f4a775d1 | 3030 | __skb_header_release(nskb); |
fe067e8a | 3031 | __tcp_add_write_queue_head(sk, nskb); |
3ab224be HA |
3032 | sk_wmem_free_skb(sk, skb); |
3033 | sk->sk_wmem_queued += nskb->truesize; | |
3034 | sk_mem_charge(sk, nskb->truesize); | |
1da177e4 LT |
3035 | skb = nskb; |
3036 | } | |
3037 | ||
4de075e0 | 3038 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; |
735d3831 | 3039 | tcp_ecn_send_synack(sk, skb); |
1da177e4 | 3040 | } |
dfb4b9dc | 3041 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
1da177e4 LT |
3042 | } |
3043 | ||
4aea39c1 ED |
3044 | /** |
3045 | * tcp_make_synack - Prepare a SYN-ACK. | |
3046 | * sk: listener socket | |
3047 | * dst: dst entry attached to the SYNACK | |
3048 | * req: request_sock pointer | |
4aea39c1 ED |
3049 | * |
3050 | * Allocate one skb and build a SYNACK packet. | |
3051 | * @dst is consumed : Caller should not use it again. | |
3052 | */ | |
5d062de7 | 3053 | struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, |
e6b4d113 | 3054 | struct request_sock *req, |
ca6fb065 | 3055 | struct tcp_fastopen_cookie *foc, |
b3d05147 | 3056 | enum tcp_synack_type synack_type) |
1da177e4 | 3057 | { |
2e6599cb | 3058 | struct inet_request_sock *ireq = inet_rsk(req); |
5d062de7 | 3059 | const struct tcp_sock *tp = tcp_sk(sk); |
80f03e27 | 3060 | struct tcp_md5sig_key *md5 = NULL; |
5d062de7 ED |
3061 | struct tcp_out_options opts; |
3062 | struct sk_buff *skb; | |
bd0388ae | 3063 | int tcp_header_size; |
5d062de7 ED |
3064 | struct tcphdr *th; |
3065 | u16 user_mss; | |
f5fff5dc | 3066 | int mss; |
1da177e4 | 3067 | |
ca6fb065 | 3068 | skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); |
4aea39c1 ED |
3069 | if (unlikely(!skb)) { |
3070 | dst_release(dst); | |
1da177e4 | 3071 | return NULL; |
4aea39c1 | 3072 | } |
1da177e4 LT |
3073 | /* Reserve space for headers. */ |
3074 | skb_reserve(skb, MAX_TCP_HEADER); | |
3075 | ||
b3d05147 ED |
3076 | switch (synack_type) { |
3077 | case TCP_SYNACK_NORMAL: | |
9e17f8a4 | 3078 | skb_set_owner_w(skb, req_to_sk(req)); |
b3d05147 ED |
3079 | break; |
3080 | case TCP_SYNACK_COOKIE: | |
3081 | /* Under synflood, we do not attach skb to a socket, | |
3082 | * to avoid false sharing. | |
3083 | */ | |
3084 | break; | |
3085 | case TCP_SYNACK_FASTOPEN: | |
ca6fb065 ED |
3086 | /* sk is a const pointer, because we want to express multiple |
3087 | * cpu might call us concurrently. | |
3088 | * sk->sk_wmem_alloc in an atomic, we can promote to rw. | |
3089 | */ | |
3090 | skb_set_owner_w(skb, (struct sock *)sk); | |
b3d05147 | 3091 | break; |
ca6fb065 | 3092 | } |
4aea39c1 | 3093 | skb_dst_set(skb, dst); |
1da177e4 | 3094 | |
0dbaee3b | 3095 | mss = dst_metric_advmss(dst); |
5d062de7 ED |
3096 | user_mss = READ_ONCE(tp->rx_opt.user_mss); |
3097 | if (user_mss && user_mss < mss) | |
3098 | mss = user_mss; | |
f5fff5dc | 3099 | |
33ad798c | 3100 | memset(&opts, 0, sizeof(opts)); |
8b5f12d0 FW |
3101 | #ifdef CONFIG_SYN_COOKIES |
3102 | if (unlikely(req->cookie_ts)) | |
7faee5c0 | 3103 | skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req); |
8b5f12d0 FW |
3104 | else |
3105 | #endif | |
7faee5c0 | 3106 | skb_mstamp_get(&skb->skb_mstamp); |
80f03e27 ED |
3107 | |
3108 | #ifdef CONFIG_TCP_MD5SIG | |
3109 | rcu_read_lock(); | |
fd3a154a | 3110 | md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); |
80f03e27 | 3111 | #endif |
58d607d3 | 3112 | skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); |
37bfbdda ED |
3113 | tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) + |
3114 | sizeof(*th); | |
cfb6eeb4 | 3115 | |
aa8223c7 ACM |
3116 | skb_push(skb, tcp_header_size); |
3117 | skb_reset_transport_header(skb); | |
1da177e4 | 3118 | |
ea1627c2 | 3119 | th = (struct tcphdr *)skb->data; |
1da177e4 LT |
3120 | memset(th, 0, sizeof(struct tcphdr)); |
3121 | th->syn = 1; | |
3122 | th->ack = 1; | |
6ac705b1 | 3123 | tcp_ecn_make_synack(req, th); |
b44084c2 | 3124 | th->source = htons(ireq->ir_num); |
634fb979 | 3125 | th->dest = ireq->ir_rmt_port; |
e870a8ef IJ |
3126 | /* Setting of flags are superfluous here for callers (and ECE is |
3127 | * not even correctly set) | |
3128 | */ | |
3129 | tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn, | |
a3433f35 | 3130 | TCPHDR_SYN | TCPHDR_ACK); |
4957faad | 3131 | |
1da177e4 | 3132 | th->seq = htonl(TCP_SKB_CB(skb)->seq); |
8336886f JC |
3133 | /* XXX data is queued and acked as is. No buffer/window check */ |
3134 | th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); | |
1da177e4 LT |
3135 | |
3136 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ | |
ed53d0ab | 3137 | th->window = htons(min(req->rsk_rcv_wnd, 65535U)); |
5d062de7 | 3138 | tcp_options_write((__be32 *)(th + 1), NULL, &opts); |
1da177e4 | 3139 | th->doff = (tcp_header_size >> 2); |
90bbcc60 | 3140 | __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); |
cfb6eeb4 YH |
3141 | |
3142 | #ifdef CONFIG_TCP_MD5SIG | |
3143 | /* Okay, we have all we need - do the md5 hash if needed */ | |
80f03e27 | 3144 | if (md5) |
bd0388ae | 3145 | tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, |
39f8e58e | 3146 | md5, req_to_sk(req), skb); |
80f03e27 | 3147 | rcu_read_unlock(); |
cfb6eeb4 YH |
3148 | #endif |
3149 | ||
b50edd78 | 3150 | /* Do not fool tcpdump (if any), clean our debris */ |
2456e855 | 3151 | skb->tstamp = 0; |
1da177e4 LT |
3152 | return skb; |
3153 | } | |
4bc2f18b | 3154 | EXPORT_SYMBOL(tcp_make_synack); |
1da177e4 | 3155 | |
81164413 DB |
3156 | static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) |
3157 | { | |
3158 | struct inet_connection_sock *icsk = inet_csk(sk); | |
3159 | const struct tcp_congestion_ops *ca; | |
3160 | u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); | |
3161 | ||
3162 | if (ca_key == TCP_CA_UNSPEC) | |
3163 | return; | |
3164 | ||
3165 | rcu_read_lock(); | |
3166 | ca = tcp_ca_find_key(ca_key); | |
3167 | if (likely(ca && try_module_get(ca->owner))) { | |
3168 | module_put(icsk->icsk_ca_ops->owner); | |
3169 | icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); | |
3170 | icsk->icsk_ca_ops = ca; | |
3171 | } | |
3172 | rcu_read_unlock(); | |
3173 | } | |
3174 | ||
67edfef7 | 3175 | /* Do all connect socket setups that can be done AF independent. */ |
f7e56a76 | 3176 | static void tcp_connect_init(struct sock *sk) |
1da177e4 | 3177 | { |
cf533ea5 | 3178 | const struct dst_entry *dst = __sk_dst_get(sk); |
1da177e4 LT |
3179 | struct tcp_sock *tp = tcp_sk(sk); |
3180 | __u8 rcv_wscale; | |
3181 | ||
3182 | /* We'll fix this up when we get a response from the other end. | |
3183 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. | |
3184 | */ | |
3185 | tp->tcp_header_len = sizeof(struct tcphdr) + | |
bb5b7c11 | 3186 | (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0); |
1da177e4 | 3187 | |
cfb6eeb4 | 3188 | #ifdef CONFIG_TCP_MD5SIG |
00db4124 | 3189 | if (tp->af_specific->md5_lookup(sk, sk)) |
cfb6eeb4 YH |
3190 | tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; |
3191 | #endif | |
3192 | ||
1da177e4 LT |
3193 | /* If user gave his TCP_MAXSEG, record it to clamp */ |
3194 | if (tp->rx_opt.user_mss) | |
3195 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
3196 | tp->max_window = 0; | |
5d424d5a | 3197 | tcp_mtup_init(sk); |
1da177e4 LT |
3198 | tcp_sync_mss(sk, dst_mtu(dst)); |
3199 | ||
81164413 DB |
3200 | tcp_ca_dst_init(sk, dst); |
3201 | ||
1da177e4 LT |
3202 | if (!tp->window_clamp) |
3203 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); | |
0dbaee3b | 3204 | tp->advmss = dst_metric_advmss(dst); |
f5fff5dc TQ |
3205 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss) |
3206 | tp->advmss = tp->rx_opt.user_mss; | |
3207 | ||
1da177e4 | 3208 | tcp_initialize_rcv_mss(sk); |
1da177e4 | 3209 | |
e88c64f0 HPP |
3210 | /* limit the window selection if the user enforce a smaller rx buffer */ |
3211 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && | |
3212 | (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) | |
3213 | tp->window_clamp = tcp_full_space(sk); | |
3214 | ||
1da177e4 LT |
3215 | tcp_select_initial_window(tcp_full_space(sk), |
3216 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), | |
3217 | &tp->rcv_wnd, | |
3218 | &tp->window_clamp, | |
bb5b7c11 | 3219 | sysctl_tcp_window_scaling, |
31d12926 | 3220 | &rcv_wscale, |
3221 | dst_metric(dst, RTAX_INITRWND)); | |
1da177e4 LT |
3222 | |
3223 | tp->rx_opt.rcv_wscale = rcv_wscale; | |
3224 | tp->rcv_ssthresh = tp->rcv_wnd; | |
3225 | ||
3226 | sk->sk_err = 0; | |
3227 | sock_reset_flag(sk, SOCK_DONE); | |
3228 | tp->snd_wnd = 0; | |
ee7537b6 | 3229 | tcp_init_wl(tp, 0); |
1da177e4 LT |
3230 | tp->snd_una = tp->write_seq; |
3231 | tp->snd_sml = tp->write_seq; | |
33f5f57e | 3232 | tp->snd_up = tp->write_seq; |
370816ae | 3233 | tp->snd_nxt = tp->write_seq; |
ee995283 PE |
3234 | |
3235 | if (likely(!tp->repair)) | |
3236 | tp->rcv_nxt = 0; | |
c7781a6e AV |
3237 | else |
3238 | tp->rcv_tstamp = tcp_time_stamp; | |
ee995283 PE |
3239 | tp->rcv_wup = tp->rcv_nxt; |
3240 | tp->copied_seq = tp->rcv_nxt; | |
1da177e4 | 3241 | |
463c84b9 ACM |
3242 | inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT; |
3243 | inet_csk(sk)->icsk_retransmits = 0; | |
1da177e4 LT |
3244 | tcp_clear_retrans(tp); |
3245 | } | |
3246 | ||
783237e8 YC |
3247 | static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) |
3248 | { | |
3249 | struct tcp_sock *tp = tcp_sk(sk); | |
3250 | struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); | |
3251 | ||
3252 | tcb->end_seq += skb->len; | |
f4a775d1 | 3253 | __skb_header_release(skb); |
783237e8 YC |
3254 | __tcp_add_write_queue_tail(sk, skb); |
3255 | sk->sk_wmem_queued += skb->truesize; | |
3256 | sk_mem_charge(sk, skb->truesize); | |
3257 | tp->write_seq = tcb->end_seq; | |
3258 | tp->packets_out += tcp_skb_pcount(skb); | |
3259 | } | |
3260 | ||
3261 | /* Build and send a SYN with data and (cached) Fast Open cookie. However, | |
3262 | * queue a data-only packet after the regular SYN, such that regular SYNs | |
3263 | * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges | |
3264 | * only the SYN sequence, the data are retransmitted in the first ACK. | |
3265 | * If cookie is not cached or other error occurs, falls back to send a | |
3266 | * regular SYN with Fast Open cookie request option. | |
3267 | */ | |
3268 | static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) | |
3269 | { | |
3270 | struct tcp_sock *tp = tcp_sk(sk); | |
3271 | struct tcp_fastopen_request *fo = tp->fastopen_req; | |
065263f4 | 3272 | int space, err = 0; |
355a901e | 3273 | struct sk_buff *syn_data; |
aab48743 | 3274 | |
67da22d2 | 3275 | tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ |
065263f4 | 3276 | if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) |
783237e8 YC |
3277 | goto fallback; |
3278 | ||
3279 | /* MSS for SYN-data is based on cached MSS and bounded by PMTU and | |
3280 | * user-MSS. Reserve maximum option space for middleboxes that add | |
3281 | * private TCP options. The cost is reduced data space in SYN :( | |
3282 | */ | |
3283 | if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp) | |
3284 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
1b63edd6 | 3285 | space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - |
783237e8 YC |
3286 | MAX_TCP_OPTION_SPACE; |
3287 | ||
f5ddcbbb ED |
3288 | space = min_t(size_t, space, fo->size); |
3289 | ||
3290 | /* limit to order-0 allocations */ | |
3291 | space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); | |
3292 | ||
eb934478 | 3293 | syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); |
355a901e | 3294 | if (!syn_data) |
783237e8 | 3295 | goto fallback; |
355a901e ED |
3296 | syn_data->ip_summed = CHECKSUM_PARTIAL; |
3297 | memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); | |
07e100f9 ED |
3298 | if (space) { |
3299 | int copied = copy_from_iter(skb_put(syn_data, space), space, | |
3300 | &fo->data->msg_iter); | |
3301 | if (unlikely(!copied)) { | |
3302 | kfree_skb(syn_data); | |
3303 | goto fallback; | |
3304 | } | |
3305 | if (copied != space) { | |
3306 | skb_trim(syn_data, copied); | |
3307 | space = copied; | |
3308 | } | |
57be5bda | 3309 | } |
355a901e ED |
3310 | /* No more data pending in inet_wait_for_connect() */ |
3311 | if (space == fo->size) | |
3312 | fo->data = NULL; | |
3313 | fo->copied = space; | |
783237e8 | 3314 | |
355a901e | 3315 | tcp_connect_queue_skb(sk, syn_data); |
0f87230d FY |
3316 | if (syn_data->len) |
3317 | tcp_chrono_start(sk, TCP_CHRONO_BUSY); | |
783237e8 | 3318 | |
355a901e | 3319 | err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); |
783237e8 | 3320 | |
355a901e | 3321 | syn->skb_mstamp = syn_data->skb_mstamp; |
431a9124 | 3322 | |
355a901e ED |
3323 | /* Now full SYN+DATA was cloned and sent (or not), |
3324 | * remove the SYN from the original skb (syn_data) | |
3325 | * we keep in write queue in case of a retransmit, as we | |
3326 | * also have the SYN packet (with no data) in the same queue. | |
3327 | */ | |
3328 | TCP_SKB_CB(syn_data)->seq++; | |
3329 | TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; | |
3330 | if (!err) { | |
67da22d2 | 3331 | tp->syn_data = (fo->copied > 0); |
f19c29e3 | 3332 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); |
783237e8 YC |
3333 | goto done; |
3334 | } | |
783237e8 YC |
3335 | |
3336 | fallback: | |
3337 | /* Send a regular SYN with Fast Open cookie request option */ | |
3338 | if (fo->cookie.len > 0) | |
3339 | fo->cookie.len = 0; | |
3340 | err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); | |
3341 | if (err) | |
3342 | tp->syn_fastopen = 0; | |
783237e8 YC |
3343 | done: |
3344 | fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ | |
3345 | return err; | |
3346 | } | |
3347 | ||
67edfef7 | 3348 | /* Build a SYN and send it off. */ |
1da177e4 LT |
3349 | int tcp_connect(struct sock *sk) |
3350 | { | |
3351 | struct tcp_sock *tp = tcp_sk(sk); | |
3352 | struct sk_buff *buff; | |
ee586811 | 3353 | int err; |
1da177e4 LT |
3354 | |
3355 | tcp_connect_init(sk); | |
3356 | ||
2b916477 AV |
3357 | if (unlikely(tp->repair)) { |
3358 | tcp_finish_connect(sk, NULL); | |
3359 | return 0; | |
3360 | } | |
3361 | ||
eb934478 | 3362 | buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); |
355a901e | 3363 | if (unlikely(!buff)) |
1da177e4 LT |
3364 | return -ENOBUFS; |
3365 | ||
a3433f35 | 3366 | tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); |
7faee5c0 | 3367 | tp->retrans_stamp = tcp_time_stamp; |
783237e8 | 3368 | tcp_connect_queue_skb(sk, buff); |
735d3831 | 3369 | tcp_ecn_send_syn(sk, buff); |
1da177e4 | 3370 | |
783237e8 YC |
3371 | /* Send off SYN; include data in Fast Open. */ |
3372 | err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : | |
3373 | tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); | |
ee586811 EP |
3374 | if (err == -ECONNREFUSED) |
3375 | return err; | |
bd37a088 WY |
3376 | |
3377 | /* We change tp->snd_nxt after the tcp_transmit_skb() call | |
3378 | * in order to make this packet get counted in tcpOutSegs. | |
3379 | */ | |
3380 | tp->snd_nxt = tp->write_seq; | |
3381 | tp->pushed_seq = tp->write_seq; | |
81cc8a75 | 3382 | TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); |
1da177e4 LT |
3383 | |
3384 | /* Timer for repeating the SYN until an answer. */ | |
3f421baa ACM |
3385 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
3386 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
1da177e4 LT |
3387 | return 0; |
3388 | } | |
4bc2f18b | 3389 | EXPORT_SYMBOL(tcp_connect); |
1da177e4 LT |
3390 | |
3391 | /* Send out a delayed ack, the caller does the policy checking | |
3392 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() | |
3393 | * for details. | |
3394 | */ | |
3395 | void tcp_send_delayed_ack(struct sock *sk) | |
3396 | { | |
463c84b9 ACM |
3397 | struct inet_connection_sock *icsk = inet_csk(sk); |
3398 | int ato = icsk->icsk_ack.ato; | |
1da177e4 LT |
3399 | unsigned long timeout; |
3400 | ||
9890092e FW |
3401 | tcp_ca_event(sk, CA_EVENT_DELAYED_ACK); |
3402 | ||
1da177e4 | 3403 | if (ato > TCP_DELACK_MIN) { |
463c84b9 | 3404 | const struct tcp_sock *tp = tcp_sk(sk); |
056834d9 | 3405 | int max_ato = HZ / 2; |
1da177e4 | 3406 | |
056834d9 IJ |
3407 | if (icsk->icsk_ack.pingpong || |
3408 | (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) | |
1da177e4 LT |
3409 | max_ato = TCP_DELACK_MAX; |
3410 | ||
3411 | /* Slow path, intersegment interval is "high". */ | |
3412 | ||
3413 | /* If some rtt estimate is known, use it to bound delayed ack. | |
463c84b9 | 3414 | * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements |
1da177e4 LT |
3415 | * directly. |
3416 | */ | |
740b0f18 ED |
3417 | if (tp->srtt_us) { |
3418 | int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), | |
3419 | TCP_DELACK_MIN); | |
1da177e4 LT |
3420 | |
3421 | if (rtt < max_ato) | |
3422 | max_ato = rtt; | |
3423 | } | |
3424 | ||
3425 | ato = min(ato, max_ato); | |
3426 | } | |
3427 | ||
3428 | /* Stay within the limit we were given */ | |
3429 | timeout = jiffies + ato; | |
3430 | ||
3431 | /* Use new timeout only if there wasn't a older one earlier. */ | |
463c84b9 | 3432 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { |
1da177e4 LT |
3433 | /* If delack timer was blocked or is about to expire, |
3434 | * send ACK now. | |
3435 | */ | |
463c84b9 ACM |
3436 | if (icsk->icsk_ack.blocked || |
3437 | time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { | |
1da177e4 LT |
3438 | tcp_send_ack(sk); |
3439 | return; | |
3440 | } | |
3441 | ||
463c84b9 ACM |
3442 | if (!time_before(timeout, icsk->icsk_ack.timeout)) |
3443 | timeout = icsk->icsk_ack.timeout; | |
1da177e4 | 3444 | } |
463c84b9 ACM |
3445 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; |
3446 | icsk->icsk_ack.timeout = timeout; | |
3447 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); | |
1da177e4 LT |
3448 | } |
3449 | ||
3450 | /* This routine sends an ack and also updates the window. */ | |
3451 | void tcp_send_ack(struct sock *sk) | |
3452 | { | |
058dc334 | 3453 | struct sk_buff *buff; |
1da177e4 | 3454 | |
058dc334 IJ |
3455 | /* If we have been reset, we may not send again. */ |
3456 | if (sk->sk_state == TCP_CLOSE) | |
3457 | return; | |
1da177e4 | 3458 | |
9890092e FW |
3459 | tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK); |
3460 | ||
058dc334 IJ |
3461 | /* We are not putting this on the write queue, so |
3462 | * tcp_transmit_skb() will set the ownership to this | |
3463 | * sock. | |
3464 | */ | |
7450aaf6 ED |
3465 | buff = alloc_skb(MAX_TCP_HEADER, |
3466 | sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); | |
3467 | if (unlikely(!buff)) { | |
058dc334 IJ |
3468 | inet_csk_schedule_ack(sk); |
3469 | inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; | |
3470 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, | |
3471 | TCP_DELACK_MAX, TCP_RTO_MAX); | |
3472 | return; | |
1da177e4 | 3473 | } |
058dc334 IJ |
3474 | |
3475 | /* Reserve space for headers and prepare control bits. */ | |
3476 | skb_reserve(buff, MAX_TCP_HEADER); | |
a3433f35 | 3477 | tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); |
058dc334 | 3478 | |
98781965 ED |
3479 | /* We do not want pure acks influencing TCP Small Queues or fq/pacing |
3480 | * too much. | |
3481 | * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 | |
98781965 ED |
3482 | */ |
3483 | skb_set_tcp_pure_ack(buff); | |
3484 | ||
058dc334 | 3485 | /* Send it off, this clears delayed acks for us. */ |
7faee5c0 | 3486 | skb_mstamp_get(&buff->skb_mstamp); |
7450aaf6 | 3487 | tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0); |
1da177e4 | 3488 | } |
e3118e83 | 3489 | EXPORT_SYMBOL_GPL(tcp_send_ack); |
1da177e4 LT |
3490 | |
3491 | /* This routine sends a packet with an out of date sequence | |
3492 | * number. It assumes the other end will try to ack it. | |
3493 | * | |
3494 | * Question: what should we make while urgent mode? | |
3495 | * 4.4BSD forces sending single byte of data. We cannot send | |
3496 | * out of window data, because we have SND.NXT==SND.MAX... | |
3497 | * | |
3498 | * Current solution: to send TWO zero-length segments in urgent mode: | |
3499 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is | |
3500 | * out-of-date with SND.UNA-1 to probe window. | |
3501 | */ | |
e520af48 | 3502 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) |
1da177e4 LT |
3503 | { |
3504 | struct tcp_sock *tp = tcp_sk(sk); | |
3505 | struct sk_buff *skb; | |
3506 | ||
3507 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ | |
7450aaf6 ED |
3508 | skb = alloc_skb(MAX_TCP_HEADER, |
3509 | sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); | |
51456b29 | 3510 | if (!skb) |
1da177e4 LT |
3511 | return -1; |
3512 | ||
3513 | /* Reserve space for headers and set control bits. */ | |
3514 | skb_reserve(skb, MAX_TCP_HEADER); | |
1da177e4 LT |
3515 | /* Use a previous sequence. This should cause the other |
3516 | * end to send an ack. Don't queue or clone SKB, just | |
3517 | * send it. | |
3518 | */ | |
a3433f35 | 3519 | tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); |
7faee5c0 | 3520 | skb_mstamp_get(&skb->skb_mstamp); |
e2e8009f | 3521 | NET_INC_STATS(sock_net(sk), mib); |
7450aaf6 | 3522 | return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); |
1da177e4 LT |
3523 | } |
3524 | ||
ee995283 PE |
3525 | void tcp_send_window_probe(struct sock *sk) |
3526 | { | |
3527 | if (sk->sk_state == TCP_ESTABLISHED) { | |
3528 | tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; | |
e520af48 | 3529 | tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); |
ee995283 PE |
3530 | } |
3531 | } | |
3532 | ||
67edfef7 | 3533 | /* Initiate keepalive or window probe from timer. */ |
e520af48 | 3534 | int tcp_write_wakeup(struct sock *sk, int mib) |
1da177e4 | 3535 | { |
058dc334 IJ |
3536 | struct tcp_sock *tp = tcp_sk(sk); |
3537 | struct sk_buff *skb; | |
1da177e4 | 3538 | |
058dc334 IJ |
3539 | if (sk->sk_state == TCP_CLOSE) |
3540 | return -1; | |
3541 | ||
00db4124 IM |
3542 | skb = tcp_send_head(sk); |
3543 | if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { | |
058dc334 | 3544 | int err; |
0c54b85f | 3545 | unsigned int mss = tcp_current_mss(sk); |
058dc334 IJ |
3546 | unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
3547 | ||
3548 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) | |
3549 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; | |
3550 | ||
3551 | /* We are probing the opening of a window | |
3552 | * but the window size is != 0 | |
3553 | * must have been a result SWS avoidance ( sender ) | |
3554 | */ | |
3555 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || | |
3556 | skb->len > mss) { | |
3557 | seg_size = min(seg_size, mss); | |
4de075e0 | 3558 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
6cc55e09 | 3559 | if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC)) |
058dc334 IJ |
3560 | return -1; |
3561 | } else if (!tcp_skb_pcount(skb)) | |
5bbb432c | 3562 | tcp_set_skb_tso_segs(skb, mss); |
058dc334 | 3563 | |
4de075e0 | 3564 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
058dc334 IJ |
3565 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
3566 | if (!err) | |
3567 | tcp_event_new_data_sent(sk, skb); | |
3568 | return err; | |
3569 | } else { | |
33f5f57e | 3570 | if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) |
e520af48 ED |
3571 | tcp_xmit_probe_skb(sk, 1, mib); |
3572 | return tcp_xmit_probe_skb(sk, 0, mib); | |
1da177e4 | 3573 | } |
1da177e4 LT |
3574 | } |
3575 | ||
3576 | /* A window probe timeout has occurred. If window is not closed send | |
3577 | * a partial packet else a zero probe. | |
3578 | */ | |
3579 | void tcp_send_probe0(struct sock *sk) | |
3580 | { | |
463c84b9 | 3581 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 3582 | struct tcp_sock *tp = tcp_sk(sk); |
c6214a97 | 3583 | struct net *net = sock_net(sk); |
fcdd1cf4 | 3584 | unsigned long probe_max; |
1da177e4 LT |
3585 | int err; |
3586 | ||
e520af48 | 3587 | err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); |
1da177e4 | 3588 | |
fe067e8a | 3589 | if (tp->packets_out || !tcp_send_head(sk)) { |
1da177e4 | 3590 | /* Cancel probe timer, if it is not required. */ |
6687e988 | 3591 | icsk->icsk_probes_out = 0; |
463c84b9 | 3592 | icsk->icsk_backoff = 0; |
1da177e4 LT |
3593 | return; |
3594 | } | |
3595 | ||
3596 | if (err <= 0) { | |
c6214a97 | 3597 | if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) |
463c84b9 | 3598 | icsk->icsk_backoff++; |
6687e988 | 3599 | icsk->icsk_probes_out++; |
fcdd1cf4 | 3600 | probe_max = TCP_RTO_MAX; |
1da177e4 LT |
3601 | } else { |
3602 | /* If packet was not sent due to local congestion, | |
6687e988 | 3603 | * do not backoff and do not remember icsk_probes_out. |
1da177e4 LT |
3604 | * Let local senders to fight for local resources. |
3605 | * | |
3606 | * Use accumulated backoff yet. | |
3607 | */ | |
6687e988 ACM |
3608 | if (!icsk->icsk_probes_out) |
3609 | icsk->icsk_probes_out = 1; | |
fcdd1cf4 | 3610 | probe_max = TCP_RESOURCE_PROBE_INTERVAL; |
1da177e4 | 3611 | } |
fcdd1cf4 | 3612 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0, |
21c8fe99 | 3613 | tcp_probe0_when(sk, probe_max), |
fcdd1cf4 | 3614 | TCP_RTO_MAX); |
1da177e4 | 3615 | } |
5db92c99 | 3616 | |
ea3bea3a | 3617 | int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) |
5db92c99 OP |
3618 | { |
3619 | const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; | |
3620 | struct flowi fl; | |
3621 | int res; | |
3622 | ||
58d607d3 | 3623 | tcp_rsk(req)->txhash = net_tx_rndhash(); |
b3d05147 | 3624 | res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); |
5db92c99 | 3625 | if (!res) { |
90bbcc60 | 3626 | __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); |
02a1d6e7 | 3627 | __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); |
7e32b443 YC |
3628 | if (unlikely(tcp_passive_fastopen(sk))) |
3629 | tcp_sk(sk)->total_retrans++; | |
5db92c99 OP |
3630 | } |
3631 | return res; | |
3632 | } | |
3633 | EXPORT_SYMBOL(tcp_rtx_synack); |