Commit | Line | Data |
---|---|---|
457c8996 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
1da177e4 LT |
2 | /* |
3 | * INET An implementation of the TCP/IP protocol suite for the LINUX | |
4 | * operating system. INET is implemented using the BSD Socket | |
5 | * interface as the means of communication with the user level. | |
6 | * | |
7 | * Implementation of the Transmission Control Protocol(TCP). | |
8 | * | |
02c30a84 | 9 | * Authors: Ross Biro |
1da177e4 LT |
10 | * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> |
11 | * Mark Evans, <evansmp@uhura.aston.ac.uk> | |
12 | * Corey Minyard <wf-rch!minyard@relay.EU.net> | |
13 | * Florian La Roche, <flla@stud.uni-sb.de> | |
14 | * Charles Hedrick, <hedrick@klinzhai.rutgers.edu> | |
15 | * Linus Torvalds, <torvalds@cs.helsinki.fi> | |
16 | * Alan Cox, <gw4pts@gw4pts.ampr.org> | |
17 | * Matthew Dillon, <dillon@apollo.west.oic.com> | |
18 | * Arnt Gulbrandsen, <agulbra@nvg.unit.no> | |
19 | * Jorge Cwik, <jorge@laser.satlink.net> | |
20 | */ | |
21 | ||
22 | /* | |
23 | * Changes: Pedro Roque : Retransmit queue handled by TCP. | |
24 | * : Fragmentation on mtu decrease | |
25 | * : Segment collapse on retransmit | |
26 | * : AF independence | |
27 | * | |
28 | * Linus Torvalds : send_delayed_ack | |
29 | * David S. Miller : Charge memory using the right skb | |
30 | * during syn/ack processing. | |
31 | * David S. Miller : Output engine completely rewritten. | |
32 | * Andrea Arcangeli: SYNACK carry ts_recent in tsecr. | |
33 | * Cacophonix Gaul : draft-minshall-nagle-01 | |
34 | * J Hadi Salim : ECN support | |
35 | * | |
36 | */ | |
37 | ||
91df42be JP |
38 | #define pr_fmt(fmt) "TCP: " fmt |
39 | ||
1da177e4 | 40 | #include <net/tcp.h> |
eda7acdd | 41 | #include <net/mptcp.h> |
f3d93817 | 42 | #include <net/proto_memory.h> |
1da177e4 LT |
43 | |
44 | #include <linux/compiler.h> | |
5a0e3ad6 | 45 | #include <linux/gfp.h> |
1da177e4 | 46 | #include <linux/module.h> |
60e2a778 | 47 | #include <linux/static_key.h> |
f6d827b1 | 48 | #include <linux/skbuff_ref.h> |
1da177e4 | 49 | |
e086101b | 50 | #include <trace/events/tcp.h> |
35089bb2 | 51 | |
9799ccb0 ED |
52 | /* Refresh clocks of a TCP socket, |
53 | * ensuring monotically increasing values. | |
54 | */ | |
55 | void tcp_mstamp_refresh(struct tcp_sock *tp) | |
56 | { | |
57 | u64 val = tcp_clock_ns(); | |
58 | ||
e6d14070 ED |
59 | tp->tcp_clock_cache = val; |
60 | tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC); | |
9799ccb0 ED |
61 | } |
62 | ||
46d3ceab ED |
63 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
64 | int push_one, gfp_t gfp); | |
519855c5 | 65 | |
67edfef7 | 66 | /* Account for new data that has been sent to the network. */ |
75c119af | 67 | static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 68 | { |
6ba8a3b1 | 69 | struct inet_connection_sock *icsk = inet_csk(sk); |
9e412ba7 | 70 | struct tcp_sock *tp = tcp_sk(sk); |
66f5fe62 | 71 | unsigned int prior_packets = tp->packets_out; |
9e412ba7 | 72 | |
e0d694d6 | 73 | WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq); |
8512430e | 74 | |
75c119af ED |
75 | __skb_unlink(skb, &sk->sk_write_queue); |
76 | tcp_rbtree_insert(&sk->tcp_rtx_queue, skb); | |
77 | ||
85369750 CZ |
78 | if (tp->highest_sack == NULL) |
79 | tp->highest_sack = skb; | |
80 | ||
66f5fe62 | 81 | tp->packets_out += tcp_skb_pcount(skb); |
bec41a11 | 82 | if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) |
750ea2ba | 83 | tcp_rearm_rto(sk); |
f19c29e3 | 84 | |
f7324acd DM |
85 | NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, |
86 | tcp_skb_pcount(skb)); | |
4bfe744f | 87 | tcp_check_space(sk); |
1da177e4 LT |
88 | } |
89 | ||
a4ecb15a CC |
90 | /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one |
91 | * window scaling factor due to loss of precision. | |
1da177e4 LT |
92 | * If window has been shrunk, what should we make? It is not clear at all. |
93 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( | |
94 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already | |
95 | * invalid. OK, let's make this for now: | |
96 | */ | |
cf533ea5 | 97 | static inline __u32 tcp_acceptable_seq(const struct sock *sk) |
1da177e4 | 98 | { |
cf533ea5 | 99 | const struct tcp_sock *tp = tcp_sk(sk); |
9e412ba7 | 100 | |
a4ecb15a CC |
101 | if (!before(tcp_wnd_end(tp), tp->snd_nxt) || |
102 | (tp->rx_opt.wscale_ok && | |
103 | ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale)))) | |
1da177e4 LT |
104 | return tp->snd_nxt; |
105 | else | |
90840def | 106 | return tcp_wnd_end(tp); |
1da177e4 LT |
107 | } |
108 | ||
109 | /* Calculate mss to advertise in SYN segment. | |
110 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: | |
111 | * | |
112 | * 1. It is independent of path mtu. | |
113 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. | |
114 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of | |
115 | * attached devices, because some buggy hosts are confused by | |
116 | * large MSS. | |
117 | * 4. We do not make 3, we advertise MSS, calculated from first | |
118 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. | |
119 | * This may be overridden via information stored in routing table. | |
120 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, | |
121 | * probably even Jumbo". | |
122 | */ | |
123 | static __u16 tcp_advertise_mss(struct sock *sk) | |
124 | { | |
125 | struct tcp_sock *tp = tcp_sk(sk); | |
cf533ea5 | 126 | const struct dst_entry *dst = __sk_dst_get(sk); |
1da177e4 LT |
127 | int mss = tp->advmss; |
128 | ||
0dbaee3b DM |
129 | if (dst) { |
130 | unsigned int metric = dst_metric_advmss(dst); | |
131 | ||
132 | if (metric < mss) { | |
133 | mss = metric; | |
134 | tp->advmss = mss; | |
135 | } | |
1da177e4 LT |
136 | } |
137 | ||
138 | return (__u16)mss; | |
139 | } | |
140 | ||
141 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". | |
6f021c62 ED |
142 | * This is the first part of cwnd validation mechanism. |
143 | */ | |
144 | void tcp_cwnd_restart(struct sock *sk, s32 delta) | |
1da177e4 | 145 | { |
463c84b9 | 146 | struct tcp_sock *tp = tcp_sk(sk); |
6f021c62 | 147 | u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); |
40570375 | 148 | u32 cwnd = tcp_snd_cwnd(tp); |
1da177e4 | 149 | |
6687e988 | 150 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); |
1da177e4 | 151 | |
6687e988 | 152 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
1da177e4 LT |
153 | restart_cwnd = min(restart_cwnd, cwnd); |
154 | ||
463c84b9 | 155 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) |
1da177e4 | 156 | cwnd >>= 1; |
40570375 | 157 | tcp_snd_cwnd_set(tp, max(cwnd, restart_cwnd)); |
c2203cf7 | 158 | tp->snd_cwnd_stamp = tcp_jiffies32; |
1da177e4 LT |
159 | tp->snd_cwnd_used = 0; |
160 | } | |
161 | ||
67edfef7 | 162 | /* Congestion state accounting after a packet has been sent. */ |
40efc6fa | 163 | static void tcp_event_data_sent(struct tcp_sock *tp, |
cf533ea5 | 164 | struct sock *sk) |
1da177e4 | 165 | { |
463c84b9 | 166 | struct inet_connection_sock *icsk = inet_csk(sk); |
d635fbe2 | 167 | const u32 now = tcp_jiffies32; |
1da177e4 | 168 | |
05c5a46d NC |
169 | if (tcp_packets_in_flight(tp) == 0) |
170 | tcp_ca_event(sk, CA_EVENT_TX_START); | |
171 | ||
4a41f453 | 172 | tp->lsndtime = now; |
4d8f24ee WW |
173 | |
174 | /* If it is a reply for ato after last received | |
562b1fdf | 175 | * packet, increase pingpong count. |
4d8f24ee WW |
176 | */ |
177 | if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) | |
562b1fdf | 178 | inet_csk_inc_pingpong_cnt(sk); |
1da177e4 LT |
179 | } |
180 | ||
67edfef7 | 181 | /* Account for an ACK we sent. */ |
059217c1 | 182 | static inline void tcp_event_ack_sent(struct sock *sk, u32 rcv_nxt) |
1da177e4 | 183 | { |
5d9f4262 ED |
184 | struct tcp_sock *tp = tcp_sk(sk); |
185 | ||
2b195850 | 186 | if (unlikely(tp->compressed_ack)) { |
200d95f4 | 187 | NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, |
2b195850 ED |
188 | tp->compressed_ack); |
189 | tp->compressed_ack = 0; | |
5d9f4262 ED |
190 | if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1) |
191 | __sock_put(sk); | |
192 | } | |
27cde44a YC |
193 | |
194 | if (unlikely(rcv_nxt != tp->rcv_nxt)) | |
195 | return; /* Special ACK sent by DCTCP to reflect ECN */ | |
059217c1 | 196 | tcp_dec_quickack_mode(sk); |
463c84b9 | 197 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); |
1da177e4 LT |
198 | } |
199 | ||
200 | /* Determine a window scaling and initial window to offer. | |
201 | * Based on the assumption that the given amount of space | |
202 | * will be offered. Store the results in the tp structure. | |
203 | * NOTE: for smooth operation initial space offering should | |
204 | * be a multiple of mss if possible. We assume here that mss >= 1. | |
205 | * This MUST be enforced by all callers. | |
206 | */ | |
ceef9ab6 | 207 | void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss, |
f410cbea | 208 | __u32 *rcv_wnd, __u32 *__window_clamp, |
31d12926 | 209 | int wscale_ok, __u8 *rcv_wscale, |
210 | __u32 init_rcv_wnd) | |
1da177e4 LT |
211 | { |
212 | unsigned int space = (__space < 0 ? 0 : __space); | |
f410cbea | 213 | u32 window_clamp = READ_ONCE(*__window_clamp); |
1da177e4 LT |
214 | |
215 | /* If no clamp set the clamp to the max possible scaled window */ | |
f410cbea ED |
216 | if (window_clamp == 0) |
217 | window_clamp = (U16_MAX << TCP_MAX_WSCALE); | |
218 | space = min(window_clamp, space); | |
1da177e4 LT |
219 | |
220 | /* Quantize space offering to a multiple of mss if possible. */ | |
221 | if (space > mss) | |
589c49cb | 222 | space = rounddown(space, mss); |
1da177e4 LT |
223 | |
224 | /* NOTE: offering an initial window larger than 32767 | |
15d99e02 RJ |
225 | * will break some buggy TCP stacks. If the admin tells us |
226 | * it is likely we could be speaking with such a buggy stack | |
227 | * we will truncate our initial window offering to 32K-1 | |
228 | * unless the remote has sent us a window scaling option, | |
229 | * which we interpret as a sign the remote TCP is not | |
230 | * misinterpreting the window field as a signed quantity. | |
1da177e4 | 231 | */ |
0f1e4d06 | 232 | if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)) |
15d99e02 RJ |
233 | (*rcv_wnd) = min(space, MAX_TCP_WINDOW); |
234 | else | |
a337531b YC |
235 | (*rcv_wnd) = min_t(u32, space, U16_MAX); |
236 | ||
237 | if (init_rcv_wnd) | |
238 | *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss); | |
15d99e02 | 239 | |
19bf6261 | 240 | *rcv_wscale = 0; |
1da177e4 | 241 | if (wscale_ok) { |
589c49cb | 242 | /* Set window scaling on max possible window */ |
02739545 | 243 | space = max_t(u32, space, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2])); |
1227c177 | 244 | space = max_t(u32, space, READ_ONCE(sysctl_rmem_max)); |
f410cbea | 245 | space = min_t(u32, space, window_clamp); |
19bf6261 ED |
246 | *rcv_wscale = clamp_t(int, ilog2(space) - 15, |
247 | 0, TCP_MAX_WSCALE); | |
1da177e4 | 248 | } |
1da177e4 | 249 | /* Set the clamp no higher than max representable value */ |
f410cbea ED |
250 | WRITE_ONCE(*__window_clamp, |
251 | min_t(__u32, U16_MAX << (*rcv_wscale), window_clamp)); | |
1da177e4 | 252 | } |
4bc2f18b | 253 | EXPORT_SYMBOL(tcp_select_initial_window); |
1da177e4 LT |
254 | |
255 | /* Chose a new window to advertise, update state in tcp_sock for the | |
256 | * socket, and return result with RFC1323 scaling applied. The return | |
257 | * value can be stuffed directly into th->window for an outgoing | |
258 | * frame. | |
259 | */ | |
40efc6fa | 260 | static u16 tcp_select_window(struct sock *sk) |
1da177e4 LT |
261 | { |
262 | struct tcp_sock *tp = tcp_sk(sk); | |
b650d953 | 263 | struct net *net = sock_net(sk); |
e2142825 MD |
264 | u32 old_win = tp->rcv_wnd; |
265 | u32 cur_win, new_win; | |
266 | ||
267 | /* Make the window 0 if we failed to queue the data because we | |
268 | * are out of memory. The window is temporary, so we don't store | |
269 | * it on the socket. | |
270 | */ | |
271 | if (unlikely(inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOMEM)) | |
272 | return 0; | |
1da177e4 | 273 | |
e2142825 MD |
274 | cur_win = tcp_receive_window(tp); |
275 | new_win = __tcp_select_window(sk); | |
2de979bd | 276 | if (new_win < cur_win) { |
1da177e4 LT |
277 | /* Danger Will Robinson! |
278 | * Don't update rcv_wup/rcv_wnd here or else | |
279 | * we will not be able to advertise a zero | |
280 | * window in time. --DaveM | |
281 | * | |
282 | * Relax Will Robinson. | |
283 | */ | |
b650d953 | 284 | if (!READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) || !tp->rx_opt.rcv_wscale) { |
285 | /* Never shrink the offered window */ | |
286 | if (new_win == 0) | |
287 | NET_INC_STATS(net, LINUX_MIB_TCPWANTZEROWINDOWADV); | |
288 | new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); | |
289 | } | |
1da177e4 | 290 | } |
b650d953 | 291 | |
1da177e4 LT |
292 | tp->rcv_wnd = new_win; |
293 | tp->rcv_wup = tp->rcv_nxt; | |
294 | ||
295 | /* Make sure we do not exceed the maximum possible | |
296 | * scaled window. | |
297 | */ | |
ceef9ab6 | 298 | if (!tp->rx_opt.rcv_wscale && |
b650d953 | 299 | READ_ONCE(net->ipv4.sysctl_tcp_workaround_signed_windows)) |
1da177e4 LT |
300 | new_win = min(new_win, MAX_TCP_WINDOW); |
301 | else | |
302 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); | |
303 | ||
304 | /* RFC1323 scaling applied */ | |
305 | new_win >>= tp->rx_opt.rcv_wscale; | |
306 | ||
31770e34 | 307 | /* If we advertise zero window, disable fast path. */ |
8e165e20 | 308 | if (new_win == 0) { |
31770e34 | 309 | tp->pred_flags = 0; |
8e165e20 | 310 | if (old_win) |
b650d953 | 311 | NET_INC_STATS(net, LINUX_MIB_TCPTOZEROWINDOWADV); |
8e165e20 | 312 | } else if (old_win == 0) { |
b650d953 | 313 | NET_INC_STATS(net, LINUX_MIB_TCPFROMZEROWINDOWADV); |
8e165e20 | 314 | } |
1da177e4 LT |
315 | |
316 | return new_win; | |
317 | } | |
318 | ||
67edfef7 | 319 | /* Packet ECN state for a SYN-ACK */ |
735d3831 | 320 | static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) |
bdf1ee5d | 321 | { |
30e502a3 DB |
322 | const struct tcp_sock *tp = tcp_sk(sk); |
323 | ||
4de075e0 | 324 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; |
056834d9 | 325 | if (!(tp->ecn_flags & TCP_ECN_OK)) |
4de075e0 | 326 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; |
91b5b21c LB |
327 | else if (tcp_ca_needs_ecn(sk) || |
328 | tcp_bpf_ca_needs_ecn(sk)) | |
30e502a3 | 329 | INET_ECN_xmit(sk); |
bdf1ee5d IJ |
330 | } |
331 | ||
67edfef7 | 332 | /* Packet ECN state for a SYN. */ |
735d3831 | 333 | static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) |
bdf1ee5d IJ |
334 | { |
335 | struct tcp_sock *tp = tcp_sk(sk); | |
91b5b21c | 336 | bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk); |
4785a667 | 337 | bool use_ecn = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn) == 1 || |
91b5b21c | 338 | tcp_ca_needs_ecn(sk) || bpf_needs_ecn; |
f7b3bec6 FW |
339 | |
340 | if (!use_ecn) { | |
341 | const struct dst_entry *dst = __sk_dst_get(sk); | |
342 | ||
343 | if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) | |
344 | use_ecn = true; | |
345 | } | |
bdf1ee5d IJ |
346 | |
347 | tp->ecn_flags = 0; | |
f7b3bec6 FW |
348 | |
349 | if (use_ecn) { | |
4de075e0 | 350 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; |
bdf1ee5d | 351 | tp->ecn_flags = TCP_ECN_OK; |
91b5b21c | 352 | if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn) |
30e502a3 | 353 | INET_ECN_xmit(sk); |
bdf1ee5d IJ |
354 | } |
355 | } | |
356 | ||
49213555 DB |
357 | static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) |
358 | { | |
12b8d9ca | 359 | if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)) |
49213555 DB |
360 | /* tp->ecn_flags are cleared at a later point in time when |
361 | * SYN ACK is ultimatively being received. | |
362 | */ | |
363 | TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); | |
364 | } | |
365 | ||
735d3831 | 366 | static void |
6ac705b1 | 367 | tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) |
bdf1ee5d | 368 | { |
6ac705b1 | 369 | if (inet_rsk(req)->ecn_ok) |
bdf1ee5d IJ |
370 | th->ece = 1; |
371 | } | |
372 | ||
67edfef7 AK |
373 | /* Set up ECN state for a packet on a ESTABLISHED socket that is about to |
374 | * be sent. | |
375 | */ | |
735d3831 | 376 | static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, |
ea1627c2 | 377 | struct tcphdr *th, int tcp_header_len) |
bdf1ee5d IJ |
378 | { |
379 | struct tcp_sock *tp = tcp_sk(sk); | |
380 | ||
381 | if (tp->ecn_flags & TCP_ECN_OK) { | |
382 | /* Not-retransmitted data segment: set ECT and inject CWR. */ | |
383 | if (skb->len != tcp_header_len && | |
384 | !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { | |
385 | INET_ECN_xmit(sk); | |
056834d9 | 386 | if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { |
bdf1ee5d | 387 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; |
ea1627c2 | 388 | th->cwr = 1; |
bdf1ee5d IJ |
389 | skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; |
390 | } | |
30e502a3 | 391 | } else if (!tcp_ca_needs_ecn(sk)) { |
bdf1ee5d IJ |
392 | /* ACK or retransmitted segment: clear ECT|CE */ |
393 | INET_ECN_dontxmit(sk); | |
394 | } | |
395 | if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) | |
ea1627c2 | 396 | th->ece = 1; |
bdf1ee5d IJ |
397 | } |
398 | } | |
399 | ||
e870a8ef IJ |
400 | /* Constructs common control bits of non-data skb. If SYN/FIN is present, |
401 | * auto increment end seqno. | |
402 | */ | |
403 | static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) | |
404 | { | |
2e8e18ef | 405 | skb->ip_summed = CHECKSUM_PARTIAL; |
e870a8ef | 406 | |
4de075e0 | 407 | TCP_SKB_CB(skb)->tcp_flags = flags; |
e870a8ef | 408 | |
cd7d8498 | 409 | tcp_skb_pcount_set(skb, 1); |
e870a8ef IJ |
410 | |
411 | TCP_SKB_CB(skb)->seq = seq; | |
a3433f35 | 412 | if (flags & (TCPHDR_SYN | TCPHDR_FIN)) |
e870a8ef IJ |
413 | seq++; |
414 | TCP_SKB_CB(skb)->end_seq = seq; | |
415 | } | |
416 | ||
a2a385d6 | 417 | static inline bool tcp_urg_mode(const struct tcp_sock *tp) |
33f5f57e IJ |
418 | { |
419 | return tp->snd_una != tp->snd_up; | |
420 | } | |
421 | ||
3b65abb8 LC |
422 | #define OPTION_SACK_ADVERTISE BIT(0) |
423 | #define OPTION_TS BIT(1) | |
424 | #define OPTION_MD5 BIT(2) | |
425 | #define OPTION_WSCALE BIT(3) | |
426 | #define OPTION_FAST_OPEN_COOKIE BIT(8) | |
427 | #define OPTION_SMC BIT(9) | |
428 | #define OPTION_MPTCP BIT(10) | |
1e03d32b | 429 | #define OPTION_AO BIT(11) |
60e2a778 UB |
430 | |
431 | static void smc_options_write(__be32 *ptr, u16 *options) | |
432 | { | |
433 | #if IS_ENABLED(CONFIG_SMC) | |
434 | if (static_branch_unlikely(&tcp_have_smc)) { | |
435 | if (unlikely(OPTION_SMC & *options)) { | |
436 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
437 | (TCPOPT_NOP << 16) | | |
438 | (TCPOPT_EXP << 8) | | |
439 | (TCPOLEN_EXP_SMC_BASE)); | |
440 | *ptr++ = htonl(TCPOPT_SMC_MAGIC); | |
441 | } | |
442 | } | |
443 | #endif | |
444 | } | |
33ad798c AL |
445 | |
446 | struct tcp_out_options { | |
2100c8d2 YC |
447 | u16 options; /* bit field of OPTION_* */ |
448 | u16 mss; /* 0 to disable */ | |
33ad798c AL |
449 | u8 ws; /* window scale, 0 to disable */ |
450 | u8 num_sack_blocks; /* number of SACK blocks to include */ | |
bd0388ae | 451 | u8 hash_size; /* bytes in hash_location */ |
331fca43 | 452 | u8 bpf_opt_len; /* length of BPF hdr option */ |
bd0388ae | 453 | __u8 *hash_location; /* temporary pointer, overloaded */ |
2100c8d2 YC |
454 | __u32 tsval, tsecr; /* need to include OPTION_TS */ |
455 | struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ | |
eda7acdd | 456 | struct mptcp_out_options mptcp; |
33ad798c AL |
457 | }; |
458 | ||
ea66758c PA |
459 | static void mptcp_options_write(struct tcphdr *th, __be32 *ptr, |
460 | struct tcp_sock *tp, | |
fa3fe2b1 | 461 | struct tcp_out_options *opts) |
eda7acdd PK |
462 | { |
463 | #if IS_ENABLED(CONFIG_MPTCP) | |
464 | if (unlikely(OPTION_MPTCP & opts->options)) | |
ea66758c | 465 | mptcp_write_options(th, ptr, tp, &opts->mptcp); |
eda7acdd PK |
466 | #endif |
467 | } | |
468 | ||
331fca43 | 469 | #ifdef CONFIG_CGROUP_BPF |
0813a841 MKL |
470 | static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb, |
471 | enum tcp_synack_type synack_type) | |
472 | { | |
473 | if (unlikely(!skb)) | |
474 | return BPF_WRITE_HDR_TCP_CURRENT_MSS; | |
475 | ||
476 | if (unlikely(synack_type == TCP_SYNACK_COOKIE)) | |
477 | return BPF_WRITE_HDR_TCP_SYNACK_COOKIE; | |
478 | ||
479 | return 0; | |
480 | } | |
481 | ||
331fca43 MKL |
482 | /* req, syn_skb and synack_type are used when writing synack */ |
483 | static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb, | |
484 | struct request_sock *req, | |
485 | struct sk_buff *syn_skb, | |
486 | enum tcp_synack_type synack_type, | |
487 | struct tcp_out_options *opts, | |
488 | unsigned int *remaining) | |
489 | { | |
0813a841 MKL |
490 | struct bpf_sock_ops_kern sock_ops; |
491 | int err; | |
492 | ||
331fca43 MKL |
493 | if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), |
494 | BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) || | |
495 | !*remaining) | |
496 | return; | |
497 | ||
0813a841 MKL |
498 | /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */ |
499 | ||
500 | /* init sock_ops */ | |
501 | memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); | |
502 | ||
503 | sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB; | |
504 | ||
505 | if (req) { | |
506 | /* The listen "sk" cannot be passed here because | |
507 | * it is not locked. It would not make too much | |
508 | * sense to do bpf_setsockopt(listen_sk) based | |
509 | * on individual connection request also. | |
510 | * | |
511 | * Thus, "req" is passed here and the cgroup-bpf-progs | |
512 | * of the listen "sk" will be run. | |
513 | * | |
514 | * "req" is also used here for fastopen even the "sk" here is | |
515 | * a fullsock "child" sk. It is to keep the behavior | |
516 | * consistent between fastopen and non-fastopen on | |
517 | * the bpf programming side. | |
518 | */ | |
519 | sock_ops.sk = (struct sock *)req; | |
520 | sock_ops.syn_skb = syn_skb; | |
521 | } else { | |
522 | sock_owned_by_me(sk); | |
523 | ||
524 | sock_ops.is_fullsock = 1; | |
525 | sock_ops.sk = sk; | |
526 | } | |
527 | ||
528 | sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type); | |
529 | sock_ops.remaining_opt_len = *remaining; | |
530 | /* tcp_current_mss() does not pass a skb */ | |
531 | if (skb) | |
532 | bpf_skops_init_skb(&sock_ops, skb, 0); | |
533 | ||
534 | err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk); | |
535 | ||
536 | if (err || sock_ops.remaining_opt_len == *remaining) | |
537 | return; | |
538 | ||
539 | opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len; | |
540 | /* round up to 4 bytes */ | |
541 | opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3; | |
542 | ||
543 | *remaining -= opts->bpf_opt_len; | |
331fca43 MKL |
544 | } |
545 | ||
546 | static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb, | |
547 | struct request_sock *req, | |
548 | struct sk_buff *syn_skb, | |
549 | enum tcp_synack_type synack_type, | |
550 | struct tcp_out_options *opts) | |
551 | { | |
0813a841 MKL |
552 | u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len; |
553 | struct bpf_sock_ops_kern sock_ops; | |
554 | int err; | |
555 | ||
556 | if (likely(!max_opt_len)) | |
331fca43 MKL |
557 | return; |
558 | ||
0813a841 MKL |
559 | memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); |
560 | ||
561 | sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB; | |
562 | ||
563 | if (req) { | |
564 | sock_ops.sk = (struct sock *)req; | |
565 | sock_ops.syn_skb = syn_skb; | |
566 | } else { | |
567 | sock_owned_by_me(sk); | |
568 | ||
569 | sock_ops.is_fullsock = 1; | |
570 | sock_ops.sk = sk; | |
571 | } | |
572 | ||
573 | sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type); | |
574 | sock_ops.remaining_opt_len = max_opt_len; | |
575 | first_opt_off = tcp_hdrlen(skb) - max_opt_len; | |
576 | bpf_skops_init_skb(&sock_ops, skb, first_opt_off); | |
577 | ||
578 | err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk); | |
579 | ||
580 | if (err) | |
581 | nr_written = 0; | |
582 | else | |
583 | nr_written = max_opt_len - sock_ops.remaining_opt_len; | |
584 | ||
585 | if (nr_written < max_opt_len) | |
586 | memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP, | |
587 | max_opt_len - nr_written); | |
331fca43 MKL |
588 | } |
589 | #else | |
590 | static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb, | |
591 | struct request_sock *req, | |
592 | struct sk_buff *syn_skb, | |
593 | enum tcp_synack_type synack_type, | |
594 | struct tcp_out_options *opts, | |
595 | unsigned int *remaining) | |
596 | { | |
597 | } | |
598 | ||
599 | static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb, | |
600 | struct request_sock *req, | |
601 | struct sk_buff *syn_skb, | |
602 | enum tcp_synack_type synack_type, | |
603 | struct tcp_out_options *opts) | |
604 | { | |
605 | } | |
606 | #endif | |
607 | ||
7425627b NC |
608 | static __be32 *process_tcp_ao_options(struct tcp_sock *tp, |
609 | const struct tcp_request_sock *tcprsk, | |
610 | struct tcp_out_options *opts, | |
611 | struct tcp_key *key, __be32 *ptr) | |
612 | { | |
613 | #ifdef CONFIG_TCP_AO | |
614 | u8 maclen = tcp_ao_maclen(key->ao_key); | |
615 | ||
616 | if (tcprsk) { | |
617 | u8 aolen = maclen + sizeof(struct tcp_ao_hdr); | |
618 | ||
619 | *ptr++ = htonl((TCPOPT_AO << 24) | (aolen << 16) | | |
620 | (tcprsk->ao_keyid << 8) | | |
621 | (tcprsk->ao_rcv_next)); | |
622 | } else { | |
623 | struct tcp_ao_key *rnext_key; | |
624 | struct tcp_ao_info *ao_info; | |
625 | ||
626 | ao_info = rcu_dereference_check(tp->ao_info, | |
627 | lockdep_sock_is_held(&tp->inet_conn.icsk_inet.sk)); | |
628 | rnext_key = READ_ONCE(ao_info->rnext_key); | |
629 | if (WARN_ON_ONCE(!rnext_key)) | |
630 | return ptr; | |
631 | *ptr++ = htonl((TCPOPT_AO << 24) | | |
632 | (tcp_ao_len(key->ao_key) << 16) | | |
633 | (key->ao_key->sndid << 8) | | |
634 | (rnext_key->rcvid)); | |
635 | } | |
636 | opts->hash_location = (__u8 *)ptr; | |
637 | ptr += maclen / sizeof(*ptr); | |
638 | if (unlikely(maclen % sizeof(*ptr))) { | |
639 | memset(ptr, TCPOPT_NOP, sizeof(*ptr)); | |
640 | ptr++; | |
641 | } | |
642 | #endif | |
643 | return ptr; | |
644 | } | |
645 | ||
67edfef7 AK |
646 | /* Write previously computed TCP options to the packet. |
647 | * | |
648 | * Beware: Something in the Internet is very sensitive to the ordering of | |
fd6149d3 IJ |
649 | * TCP options, we learned this through the hard way, so be careful here. |
650 | * Luckily we can at least blame others for their non-compliance but from | |
8e3bff96 | 651 | * inter-operability perspective it seems that we're somewhat stuck with |
fd6149d3 IJ |
652 | * the ordering which we have been using if we want to keep working with |
653 | * those broken things (not that it currently hurts anybody as there isn't | |
654 | * particular reason why the ordering would need to be changed). | |
655 | * | |
656 | * At least SACK_PERM as the first option is known to lead to a disaster | |
657 | * (but it may well be that other scenarios fail similarly). | |
658 | */ | |
ea66758c | 659 | static void tcp_options_write(struct tcphdr *th, struct tcp_sock *tp, |
06b22ef2 | 660 | const struct tcp_request_sock *tcprsk, |
1e03d32b DS |
661 | struct tcp_out_options *opts, |
662 | struct tcp_key *key) | |
bd0388ae | 663 | { |
ea66758c | 664 | __be32 *ptr = (__be32 *)(th + 1); |
2100c8d2 | 665 | u16 options = opts->options; /* mungable copy */ |
bd0388ae | 666 | |
1e03d32b | 667 | if (tcp_key_is_md5(key)) { |
1a2c6181 CP |
668 | *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | |
669 | (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); | |
bd0388ae WAS |
670 | /* overload cookie hash location */ |
671 | opts->hash_location = (__u8 *)ptr; | |
33ad798c | 672 | ptr += 4; |
1e03d32b | 673 | } else if (tcp_key_is_ao(key)) { |
7425627b | 674 | ptr = process_tcp_ao_options(tp, tcprsk, opts, key, ptr); |
40efc6fa | 675 | } |
fd6149d3 IJ |
676 | if (unlikely(opts->mss)) { |
677 | *ptr++ = htonl((TCPOPT_MSS << 24) | | |
678 | (TCPOLEN_MSS << 16) | | |
679 | opts->mss); | |
680 | } | |
681 | ||
bd0388ae WAS |
682 | if (likely(OPTION_TS & options)) { |
683 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { | |
33ad798c AL |
684 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
685 | (TCPOLEN_SACK_PERM << 16) | | |
686 | (TCPOPT_TIMESTAMP << 8) | | |
687 | TCPOLEN_TIMESTAMP); | |
bd0388ae | 688 | options &= ~OPTION_SACK_ADVERTISE; |
33ad798c AL |
689 | } else { |
690 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
691 | (TCPOPT_NOP << 16) | | |
692 | (TCPOPT_TIMESTAMP << 8) | | |
693 | TCPOLEN_TIMESTAMP); | |
694 | } | |
695 | *ptr++ = htonl(opts->tsval); | |
696 | *ptr++ = htonl(opts->tsecr); | |
697 | } | |
698 | ||
bd0388ae | 699 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { |
33ad798c AL |
700 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
701 | (TCPOPT_NOP << 16) | | |
702 | (TCPOPT_SACK_PERM << 8) | | |
703 | TCPOLEN_SACK_PERM); | |
704 | } | |
705 | ||
bd0388ae | 706 | if (unlikely(OPTION_WSCALE & options)) { |
33ad798c AL |
707 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
708 | (TCPOPT_WINDOW << 16) | | |
709 | (TCPOLEN_WINDOW << 8) | | |
710 | opts->ws); | |
711 | } | |
712 | ||
713 | if (unlikely(opts->num_sack_blocks)) { | |
714 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? | |
715 | tp->duplicate_sack : tp->selective_acks; | |
40efc6fa SH |
716 | int this_sack; |
717 | ||
718 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
719 | (TCPOPT_NOP << 16) | | |
720 | (TCPOPT_SACK << 8) | | |
33ad798c | 721 | (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * |
40efc6fa | 722 | TCPOLEN_SACK_PERBLOCK))); |
2de979bd | 723 | |
33ad798c AL |
724 | for (this_sack = 0; this_sack < opts->num_sack_blocks; |
725 | ++this_sack) { | |
40efc6fa SH |
726 | *ptr++ = htonl(sp[this_sack].start_seq); |
727 | *ptr++ = htonl(sp[this_sack].end_seq); | |
728 | } | |
2de979bd | 729 | |
5861f8e5 | 730 | tp->rx_opt.dsack = 0; |
40efc6fa | 731 | } |
2100c8d2 YC |
732 | |
733 | if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { | |
734 | struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; | |
7f9b838b DL |
735 | u8 *p = (u8 *)ptr; |
736 | u32 len; /* Fast Open option length */ | |
737 | ||
738 | if (foc->exp) { | |
739 | len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; | |
740 | *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | | |
741 | TCPOPT_FASTOPEN_MAGIC); | |
742 | p += TCPOLEN_EXP_FASTOPEN_BASE; | |
743 | } else { | |
744 | len = TCPOLEN_FASTOPEN_BASE + foc->len; | |
745 | *p++ = TCPOPT_FASTOPEN; | |
746 | *p++ = len; | |
747 | } | |
2100c8d2 | 748 | |
7f9b838b DL |
749 | memcpy(p, foc->val, foc->len); |
750 | if ((len & 3) == 2) { | |
751 | p[foc->len] = TCPOPT_NOP; | |
752 | p[foc->len + 1] = TCPOPT_NOP; | |
2100c8d2 | 753 | } |
7f9b838b | 754 | ptr += (len + 3) >> 2; |
2100c8d2 | 755 | } |
60e2a778 UB |
756 | |
757 | smc_options_write(ptr, &options); | |
eda7acdd | 758 | |
ea66758c | 759 | mptcp_options_write(th, ptr, tp, opts); |
60e2a778 UB |
760 | } |
761 | ||
762 | static void smc_set_option(const struct tcp_sock *tp, | |
763 | struct tcp_out_options *opts, | |
764 | unsigned int *remaining) | |
765 | { | |
766 | #if IS_ENABLED(CONFIG_SMC) | |
767 | if (static_branch_unlikely(&tcp_have_smc)) { | |
768 | if (tp->syn_smc) { | |
769 | if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { | |
770 | opts->options |= OPTION_SMC; | |
771 | *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; | |
772 | } | |
773 | } | |
774 | } | |
775 | #endif | |
776 | } | |
777 | ||
778 | static void smc_set_option_cond(const struct tcp_sock *tp, | |
779 | const struct inet_request_sock *ireq, | |
780 | struct tcp_out_options *opts, | |
781 | unsigned int *remaining) | |
782 | { | |
783 | #if IS_ENABLED(CONFIG_SMC) | |
784 | if (static_branch_unlikely(&tcp_have_smc)) { | |
785 | if (tp->syn_smc && ireq->smc_ok) { | |
786 | if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { | |
787 | opts->options |= OPTION_SMC; | |
788 | *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; | |
789 | } | |
790 | } | |
791 | } | |
792 | #endif | |
33ad798c AL |
793 | } |
794 | ||
cec37a6e PK |
795 | static void mptcp_set_option_cond(const struct request_sock *req, |
796 | struct tcp_out_options *opts, | |
797 | unsigned int *remaining) | |
798 | { | |
799 | if (rsk_is_mptcp(req)) { | |
800 | unsigned int size; | |
801 | ||
802 | if (mptcp_synack_options(req, &size, &opts->mptcp)) { | |
803 | if (*remaining >= size) { | |
804 | opts->options |= OPTION_MPTCP; | |
805 | *remaining -= size; | |
806 | } | |
807 | } | |
808 | } | |
809 | } | |
810 | ||
67edfef7 AK |
811 | /* Compute TCP options for SYN packets. This is not the final |
812 | * network wire format yet. | |
813 | */ | |
95c96174 | 814 | static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, |
33ad798c | 815 | struct tcp_out_options *opts, |
1e03d32b | 816 | struct tcp_key *key) |
cf533ea5 | 817 | { |
33ad798c | 818 | struct tcp_sock *tp = tcp_sk(sk); |
95c96174 | 819 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
783237e8 | 820 | struct tcp_fastopen_request *fastopen = tp->fastopen_req; |
1e03d32b | 821 | bool timestamps; |
33ad798c | 822 | |
1e03d32b DS |
823 | /* Better than switch (key.type) as it has static branches */ |
824 | if (tcp_key_is_md5(key)) { | |
825 | timestamps = false; | |
826 | opts->options |= OPTION_MD5; | |
827 | remaining -= TCPOLEN_MD5SIG_ALIGNED; | |
828 | } else { | |
829 | timestamps = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps); | |
830 | if (tcp_key_is_ao(key)) { | |
831 | opts->options |= OPTION_AO; | |
da7dfaa6 | 832 | remaining -= tcp_ao_len_aligned(key->ao_key); |
8c2320e8 | 833 | } |
cfb6eeb4 | 834 | } |
33ad798c AL |
835 | |
836 | /* We always get an MSS option. The option bytes which will be seen in | |
837 | * normal data packets should timestamps be used, must be in the MSS | |
838 | * advertised. But we subtract them from tp->mss_cache so that | |
839 | * calculations in tcp_sendmsg are simpler etc. So account for this | |
840 | * fact here if necessary. If we don't do this correctly, as a | |
841 | * receiver we won't recognize data packets as being full sized when we | |
842 | * should, and thus we won't abide by the delayed ACK rules correctly. | |
843 | * SACKs don't matter, we never delay an ACK when we have any of those | |
844 | * going out. */ | |
845 | opts->mss = tcp_advertise_mss(sk); | |
bd0388ae | 846 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c | 847 | |
1e03d32b | 848 | if (likely(timestamps)) { |
33ad798c | 849 | opts->options |= OPTION_TS; |
614e8316 | 850 | opts->tsval = tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb) + tp->tsoffset; |
33ad798c | 851 | opts->tsecr = tp->rx_opt.ts_recent; |
bd0388ae | 852 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c | 853 | } |
3666f666 | 854 | if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling))) { |
33ad798c | 855 | opts->ws = tp->rx_opt.rcv_wscale; |
89e95a61 | 856 | opts->options |= OPTION_WSCALE; |
bd0388ae | 857 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 858 | } |
3666f666 | 859 | if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_sack))) { |
33ad798c | 860 | opts->options |= OPTION_SACK_ADVERTISE; |
b32d1310 | 861 | if (unlikely(!(OPTION_TS & opts->options))) |
bd0388ae | 862 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c AL |
863 | } |
864 | ||
783237e8 | 865 | if (fastopen && fastopen->cookie.len >= 0) { |
2646c831 DL |
866 | u32 need = fastopen->cookie.len; |
867 | ||
868 | need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : | |
869 | TCPOLEN_FASTOPEN_BASE; | |
783237e8 YC |
870 | need = (need + 3) & ~3U; /* Align to 32 bits */ |
871 | if (remaining >= need) { | |
872 | opts->options |= OPTION_FAST_OPEN_COOKIE; | |
873 | opts->fastopen_cookie = &fastopen->cookie; | |
874 | remaining -= need; | |
875 | tp->syn_fastopen = 1; | |
2646c831 | 876 | tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; |
783237e8 YC |
877 | } |
878 | } | |
bd0388ae | 879 | |
60e2a778 UB |
880 | smc_set_option(tp, opts, &remaining); |
881 | ||
cec37a6e PK |
882 | if (sk_is_mptcp(sk)) { |
883 | unsigned int size; | |
884 | ||
cc7972ea | 885 | if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) { |
cec37a6e PK |
886 | opts->options |= OPTION_MPTCP; |
887 | remaining -= size; | |
888 | } | |
889 | } | |
890 | ||
331fca43 MKL |
891 | bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining); |
892 | ||
bd0388ae | 893 | return MAX_TCP_OPTION_SPACE - remaining; |
40efc6fa SH |
894 | } |
895 | ||
67edfef7 | 896 | /* Set up TCP options for SYN-ACKs. */ |
60e2a778 UB |
897 | static unsigned int tcp_synack_options(const struct sock *sk, |
898 | struct request_sock *req, | |
37bfbdda ED |
899 | unsigned int mss, struct sk_buff *skb, |
900 | struct tcp_out_options *opts, | |
9427c6aa | 901 | const struct tcp_key *key, |
e114e1e8 | 902 | struct tcp_fastopen_cookie *foc, |
331fca43 MKL |
903 | enum tcp_synack_type synack_type, |
904 | struct sk_buff *syn_skb) | |
4957faad | 905 | { |
33ad798c | 906 | struct inet_request_sock *ireq = inet_rsk(req); |
95c96174 | 907 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
33ad798c | 908 | |
9427c6aa | 909 | if (tcp_key_is_md5(key)) { |
33ad798c | 910 | opts->options |= OPTION_MD5; |
4957faad WAS |
911 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
912 | ||
913 | /* We can't fit any SACK blocks in a packet with MD5 + TS | |
914 | * options. There was discussion about disabling SACK | |
915 | * rather than TS in order to fit in better with old, | |
916 | * buggy kernels, but that was deemed to be unnecessary. | |
917 | */ | |
e114e1e8 ED |
918 | if (synack_type != TCP_SYNACK_COOKIE) |
919 | ireq->tstamp_ok &= !ireq->sack_ok; | |
9427c6aa DS |
920 | } else if (tcp_key_is_ao(key)) { |
921 | opts->options |= OPTION_AO; | |
da7dfaa6 | 922 | remaining -= tcp_ao_len_aligned(key->ao_key); |
9427c6aa | 923 | ireq->tstamp_ok &= !ireq->sack_ok; |
cfb6eeb4 | 924 | } |
33ad798c | 925 | |
4957faad | 926 | /* We always send an MSS option. */ |
33ad798c | 927 | opts->mss = mss; |
4957faad | 928 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c AL |
929 | |
930 | if (likely(ireq->wscale_ok)) { | |
931 | opts->ws = ireq->rcv_wscale; | |
89e95a61 | 932 | opts->options |= OPTION_WSCALE; |
4957faad | 933 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 934 | } |
de213e5e | 935 | if (likely(ireq->tstamp_ok)) { |
33ad798c | 936 | opts->options |= OPTION_TS; |
614e8316 ED |
937 | opts->tsval = tcp_skb_timestamp_ts(tcp_rsk(req)->req_usec_ts, skb) + |
938 | tcp_rsk(req)->ts_off; | |
eba20811 | 939 | opts->tsecr = READ_ONCE(req->ts_recent); |
4957faad | 940 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c AL |
941 | } |
942 | if (likely(ireq->sack_ok)) { | |
943 | opts->options |= OPTION_SACK_ADVERTISE; | |
de213e5e | 944 | if (unlikely(!ireq->tstamp_ok)) |
4957faad | 945 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c | 946 | } |
7f9b838b DL |
947 | if (foc != NULL && foc->len >= 0) { |
948 | u32 need = foc->len; | |
949 | ||
950 | need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : | |
951 | TCPOLEN_FASTOPEN_BASE; | |
8336886f JC |
952 | need = (need + 3) & ~3U; /* Align to 32 bits */ |
953 | if (remaining >= need) { | |
954 | opts->options |= OPTION_FAST_OPEN_COOKIE; | |
955 | opts->fastopen_cookie = foc; | |
956 | remaining -= need; | |
957 | } | |
958 | } | |
1a2c6181 | 959 | |
cec37a6e PK |
960 | mptcp_set_option_cond(req, opts, &remaining); |
961 | ||
60e2a778 UB |
962 | smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining); |
963 | ||
331fca43 MKL |
964 | bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb, |
965 | synack_type, opts, &remaining); | |
966 | ||
4957faad | 967 | return MAX_TCP_OPTION_SPACE - remaining; |
33ad798c AL |
968 | } |
969 | ||
67edfef7 AK |
970 | /* Compute TCP options for ESTABLISHED sockets. This is not the |
971 | * final wire format yet. | |
972 | */ | |
95c96174 | 973 | static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, |
33ad798c | 974 | struct tcp_out_options *opts, |
1e03d32b | 975 | struct tcp_key *key) |
cf533ea5 | 976 | { |
33ad798c | 977 | struct tcp_sock *tp = tcp_sk(sk); |
95c96174 | 978 | unsigned int size = 0; |
cabeccbd | 979 | unsigned int eff_sacks; |
33ad798c | 980 | |
5843ef42 AK |
981 | opts->options = 0; |
982 | ||
1e03d32b DS |
983 | /* Better than switch (key.type) as it has static branches */ |
984 | if (tcp_key_is_md5(key)) { | |
985 | opts->options |= OPTION_MD5; | |
986 | size += TCPOLEN_MD5SIG_ALIGNED; | |
987 | } else if (tcp_key_is_ao(key)) { | |
988 | opts->options |= OPTION_AO; | |
da7dfaa6 | 989 | size += tcp_ao_len_aligned(key->ao_key); |
33ad798c | 990 | } |
33ad798c AL |
991 | |
992 | if (likely(tp->rx_opt.tstamp_ok)) { | |
993 | opts->options |= OPTION_TS; | |
614e8316 ED |
994 | opts->tsval = skb ? tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb) + |
995 | tp->tsoffset : 0; | |
33ad798c AL |
996 | opts->tsecr = tp->rx_opt.ts_recent; |
997 | size += TCPOLEN_TSTAMP_ALIGNED; | |
998 | } | |
999 | ||
cec37a6e PK |
1000 | /* MPTCP options have precedence over SACK for the limited TCP |
1001 | * option space because a MPTCP connection would be forced to | |
1002 | * fall back to regular TCP if a required multipath option is | |
1003 | * missing. SACK still gets a chance to use whatever space is | |
1004 | * left. | |
1005 | */ | |
1006 | if (sk_is_mptcp(sk)) { | |
1007 | unsigned int remaining = MAX_TCP_OPTION_SPACE - size; | |
1008 | unsigned int opt_size = 0; | |
1009 | ||
1010 | if (mptcp_established_options(sk, skb, &opt_size, remaining, | |
1011 | &opts->mptcp)) { | |
1012 | opts->options |= OPTION_MPTCP; | |
1013 | size += opt_size; | |
1014 | } | |
1015 | } | |
1016 | ||
cabeccbd IJ |
1017 | eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; |
1018 | if (unlikely(eff_sacks)) { | |
95c96174 | 1019 | const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; |
9cfcca23 MM |
1020 | if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED + |
1021 | TCPOLEN_SACK_PERBLOCK)) | |
1022 | return size; | |
1023 | ||
33ad798c | 1024 | opts->num_sack_blocks = |
95c96174 | 1025 | min_t(unsigned int, eff_sacks, |
33ad798c AL |
1026 | (remaining - TCPOLEN_SACK_BASE_ALIGNED) / |
1027 | TCPOLEN_SACK_PERBLOCK); | |
9cfcca23 MM |
1028 | |
1029 | size += TCPOLEN_SACK_BASE_ALIGNED + | |
1030 | opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; | |
33ad798c AL |
1031 | } |
1032 | ||
331fca43 MKL |
1033 | if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp, |
1034 | BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) { | |
1035 | unsigned int remaining = MAX_TCP_OPTION_SPACE - size; | |
1036 | ||
1037 | bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining); | |
1038 | ||
1039 | size = MAX_TCP_OPTION_SPACE - remaining; | |
1040 | } | |
1041 | ||
33ad798c | 1042 | return size; |
40efc6fa | 1043 | } |
1da177e4 | 1044 | |
46d3ceab ED |
1045 | |
1046 | /* TCP SMALL QUEUES (TSQ) | |
1047 | * | |
1048 | * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) | |
1049 | * to reduce RTT and bufferbloat. | |
1050 | * We do this using a special skb destructor (tcp_wfree). | |
1051 | * | |
1052 | * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb | |
1053 | * needs to be reallocated in a driver. | |
8e3bff96 | 1054 | * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc |
46d3ceab ED |
1055 | * |
1056 | * Since transmit from skb destructor is forbidden, we use a tasklet | |
1057 | * to process all sockets that eventually need to send more skbs. | |
1058 | * We use one tasklet per cpu, with its own queue of sockets. | |
1059 | */ | |
1060 | struct tsq_tasklet { | |
1061 | struct tasklet_struct tasklet; | |
1062 | struct list_head head; /* queue of tcp sockets */ | |
1063 | }; | |
1064 | static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); | |
1065 | ||
73a6bab5 | 1066 | static void tcp_tsq_write(struct sock *sk) |
6f458dfb ED |
1067 | { |
1068 | if ((1 << sk->sk_state) & | |
1069 | (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | | |
f9616c35 ED |
1070 | TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { |
1071 | struct tcp_sock *tp = tcp_sk(sk); | |
1072 | ||
1073 | if (tp->lost_out > tp->retrans_out && | |
40570375 | 1074 | tcp_snd_cwnd(tp) > tcp_packets_in_flight(tp)) { |
3a91d29f | 1075 | tcp_mstamp_refresh(tp); |
f9616c35 | 1076 | tcp_xmit_retransmit_queue(sk); |
3a91d29f | 1077 | } |
f9616c35 ED |
1078 | |
1079 | tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, | |
bf06200e | 1080 | 0, GFP_ATOMIC); |
f9616c35 | 1081 | } |
6f458dfb | 1082 | } |
73a6bab5 ED |
1083 | |
1084 | static void tcp_tsq_handler(struct sock *sk) | |
1085 | { | |
1086 | bh_lock_sock(sk); | |
1087 | if (!sock_owned_by_user(sk)) | |
1088 | tcp_tsq_write(sk); | |
1089 | else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) | |
1090 | sock_hold(sk); | |
1091 | bh_unlock_sock(sk); | |
1092 | } | |
46d3ceab | 1093 | /* |
8e3bff96 | 1094 | * One tasklet per cpu tries to send more skbs. |
46d3ceab | 1095 | * We run in tasklet context but need to disable irqs when |
8e3bff96 | 1096 | * transferring tsq->head because tcp_wfree() might |
46d3ceab ED |
1097 | * interrupt us (non NAPI drivers) |
1098 | */ | |
c6533ca8 | 1099 | static void tcp_tasklet_func(struct tasklet_struct *t) |
46d3ceab | 1100 | { |
c6533ca8 | 1101 | struct tsq_tasklet *tsq = from_tasklet(tsq, t, tasklet); |
46d3ceab ED |
1102 | LIST_HEAD(list); |
1103 | unsigned long flags; | |
1104 | struct list_head *q, *n; | |
1105 | struct tcp_sock *tp; | |
1106 | struct sock *sk; | |
1107 | ||
1108 | local_irq_save(flags); | |
1109 | list_splice_init(&tsq->head, &list); | |
1110 | local_irq_restore(flags); | |
1111 | ||
1112 | list_for_each_safe(q, n, &list) { | |
1113 | tp = list_entry(q, struct tcp_sock, tsq_node); | |
1114 | list_del(&tp->tsq_node); | |
1115 | ||
1116 | sk = (struct sock *)tp; | |
0a9648f1 | 1117 | smp_mb__before_atomic(); |
7aa5470c ED |
1118 | clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); |
1119 | ||
73a6bab5 | 1120 | tcp_tsq_handler(sk); |
46d3ceab ED |
1121 | sk_free(sk); |
1122 | } | |
1123 | } | |
1124 | ||
40fc3423 ED |
1125 | #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ |
1126 | TCPF_WRITE_TIMER_DEFERRED | \ | |
1127 | TCPF_DELACK_TIMER_DEFERRED | \ | |
133c4c0d ED |
1128 | TCPF_MTU_REDUCED_DEFERRED | \ |
1129 | TCPF_ACK_DEFERRED) | |
46d3ceab ED |
1130 | /** |
1131 | * tcp_release_cb - tcp release_sock() callback | |
1132 | * @sk: socket | |
1133 | * | |
1134 | * called from release_sock() to perform protocol dependent | |
1135 | * actions before socket release. | |
1136 | */ | |
1137 | void tcp_release_cb(struct sock *sk) | |
1138 | { | |
fac30731 ED |
1139 | unsigned long flags = smp_load_acquire(&sk->sk_tsq_flags); |
1140 | unsigned long nflags; | |
46d3ceab | 1141 | |
6f458dfb ED |
1142 | /* perform an atomic operation only if at least one flag is set */ |
1143 | do { | |
6f458dfb ED |
1144 | if (!(flags & TCP_DEFERRED_ALL)) |
1145 | return; | |
1146 | nflags = flags & ~TCP_DEFERRED_ALL; | |
fac30731 | 1147 | } while (!try_cmpxchg(&sk->sk_tsq_flags, &flags, nflags)); |
6f458dfb | 1148 | |
73a6bab5 ED |
1149 | if (flags & TCPF_TSQ_DEFERRED) { |
1150 | tcp_tsq_write(sk); | |
1151 | __sock_put(sk); | |
1152 | } | |
c3f9b018 | 1153 | |
40fc3423 | 1154 | if (flags & TCPF_WRITE_TIMER_DEFERRED) { |
6f458dfb | 1155 | tcp_write_timer_handler(sk); |
144d56e9 ED |
1156 | __sock_put(sk); |
1157 | } | |
40fc3423 | 1158 | if (flags & TCPF_DELACK_TIMER_DEFERRED) { |
6f458dfb | 1159 | tcp_delack_timer_handler(sk); |
144d56e9 ED |
1160 | __sock_put(sk); |
1161 | } | |
40fc3423 | 1162 | if (flags & TCPF_MTU_REDUCED_DEFERRED) { |
4fab9071 | 1163 | inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); |
144d56e9 ED |
1164 | __sock_put(sk); |
1165 | } | |
133c4c0d ED |
1166 | if ((flags & TCPF_ACK_DEFERRED) && inet_csk_ack_scheduled(sk)) |
1167 | tcp_send_ack(sk); | |
46d3ceab ED |
1168 | } |
1169 | EXPORT_SYMBOL(tcp_release_cb); | |
1170 | ||
1171 | void __init tcp_tasklet_init(void) | |
1172 | { | |
1173 | int i; | |
1174 | ||
1175 | for_each_possible_cpu(i) { | |
1176 | struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); | |
1177 | ||
1178 | INIT_LIST_HEAD(&tsq->head); | |
c6533ca8 | 1179 | tasklet_setup(&tsq->tasklet, tcp_tasklet_func); |
46d3ceab ED |
1180 | } |
1181 | } | |
1182 | ||
1183 | /* | |
1184 | * Write buffer destructor automatically called from kfree_skb. | |
8e3bff96 | 1185 | * We can't xmit new skbs from this context, as we might already |
46d3ceab ED |
1186 | * hold qdisc lock. |
1187 | */ | |
d6a4a104 | 1188 | void tcp_wfree(struct sk_buff *skb) |
46d3ceab ED |
1189 | { |
1190 | struct sock *sk = skb->sk; | |
1191 | struct tcp_sock *tp = tcp_sk(sk); | |
408f0a6c | 1192 | unsigned long flags, nval, oval; |
b548b17a ED |
1193 | struct tsq_tasklet *tsq; |
1194 | bool empty; | |
9b462d02 ED |
1195 | |
1196 | /* Keep one reference on sk_wmem_alloc. | |
1197 | * Will be released by sk_free() from here or tcp_tasklet_func() | |
1198 | */ | |
14afee4b | 1199 | WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); |
9b462d02 ED |
1200 | |
1201 | /* If this softirq is serviced by ksoftirqd, we are likely under stress. | |
1202 | * Wait until our queues (qdisc + devices) are drained. | |
1203 | * This gives : | |
1204 | * - less callbacks to tcp_write_xmit(), reducing stress (batches) | |
1205 | * - chance for incoming ACK (processed by another cpu maybe) | |
1206 | * to migrate this flow (skb->ooo_okay will be eventually set) | |
1207 | */ | |
14afee4b | 1208 | if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) |
9b462d02 | 1209 | goto out; |
46d3ceab | 1210 | |
b548b17a ED |
1211 | oval = smp_load_acquire(&sk->sk_tsq_flags); |
1212 | do { | |
408f0a6c ED |
1213 | if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) |
1214 | goto out; | |
1215 | ||
73a6bab5 | 1216 | nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED; |
b548b17a ED |
1217 | } while (!try_cmpxchg(&sk->sk_tsq_flags, &oval, nval)); |
1218 | ||
1219 | /* queue this socket to tasklet queue */ | |
1220 | local_irq_save(flags); | |
1221 | tsq = this_cpu_ptr(&tsq_tasklet); | |
1222 | empty = list_empty(&tsq->head); | |
1223 | list_add(&tp->tsq_node, &tsq->head); | |
1224 | if (empty) | |
1225 | tasklet_schedule(&tsq->tasklet); | |
1226 | local_irq_restore(flags); | |
1227 | return; | |
9b462d02 ED |
1228 | out: |
1229 | sk_free(sk); | |
46d3ceab ED |
1230 | } |
1231 | ||
73a6bab5 ED |
1232 | /* Note: Called under soft irq. |
1233 | * We can call TCP stack right away, unless socket is owned by user. | |
218af599 ED |
1234 | */ |
1235 | enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) | |
1236 | { | |
1237 | struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); | |
1238 | struct sock *sk = (struct sock *)tp; | |
218af599 | 1239 | |
73a6bab5 ED |
1240 | tcp_tsq_handler(sk); |
1241 | sock_put(sk); | |
218af599 | 1242 | |
218af599 ED |
1243 | return HRTIMER_NORESTART; |
1244 | } | |
1245 | ||
a7a25630 ED |
1246 | static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb, |
1247 | u64 prior_wstamp) | |
e2080072 | 1248 | { |
ab408b6d ED |
1249 | struct tcp_sock *tp = tcp_sk(sk); |
1250 | ||
ab408b6d | 1251 | if (sk->sk_pacing_status != SK_PACING_NONE) { |
28b24f90 | 1252 | unsigned long rate = READ_ONCE(sk->sk_pacing_rate); |
ab408b6d ED |
1253 | |
1254 | /* Original sch_fq does not pace first 10 MSS | |
1255 | * Note that tp->data_segs_out overflows after 2^32 packets, | |
1256 | * this is a minor annoyance. | |
1257 | */ | |
76a9ebe8 | 1258 | if (rate != ~0UL && rate && tp->data_segs_out >= 10) { |
a7a25630 ED |
1259 | u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate); |
1260 | u64 credit = tp->tcp_wstamp_ns - prior_wstamp; | |
1261 | ||
1262 | /* take into account OS jitter */ | |
1263 | len_ns -= min_t(u64, len_ns / 2, credit); | |
1264 | tp->tcp_wstamp_ns += len_ns; | |
ab408b6d ED |
1265 | } |
1266 | } | |
e2080072 ED |
1267 | list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); |
1268 | } | |
1269 | ||
05e22e83 ED |
1270 | INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); |
1271 | INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); | |
dd2e0b86 | 1272 | INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb)); |
05e22e83 | 1273 | |
1da177e4 LT |
1274 | /* This routine actually transmits TCP packets queued in by |
1275 | * tcp_do_sendmsg(). This is used by both the initial | |
1276 | * transmission and possible later retransmissions. | |
1277 | * All SKB's seen here are completely headerless. It is our | |
1278 | * job to build the TCP header, and pass the packet down to | |
1279 | * IP so it can do the same plus pass the packet off to the | |
1280 | * device. | |
1281 | * | |
1282 | * We are working here with either a clone of the original | |
1283 | * SKB, or a fresh unique copy made by the retransmit engine. | |
1284 | */ | |
2987babb YC |
1285 | static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, |
1286 | int clone_it, gfp_t gfp_mask, u32 rcv_nxt) | |
1da177e4 | 1287 | { |
dfb4b9dc DM |
1288 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1289 | struct inet_sock *inet; | |
1290 | struct tcp_sock *tp; | |
1291 | struct tcp_skb_cb *tcb; | |
33ad798c | 1292 | struct tcp_out_options opts; |
95c96174 | 1293 | unsigned int tcp_options_size, tcp_header_size; |
8c72c65b | 1294 | struct sk_buff *oskb = NULL; |
1e03d32b | 1295 | struct tcp_key key; |
dfb4b9dc | 1296 | struct tcphdr *th; |
a7a25630 | 1297 | u64 prior_wstamp; |
dfb4b9dc DM |
1298 | int err; |
1299 | ||
1300 | BUG_ON(!skb || !tcp_skb_pcount(skb)); | |
6f094b9e | 1301 | tp = tcp_sk(sk); |
7f12422c YC |
1302 | prior_wstamp = tp->tcp_wstamp_ns; |
1303 | tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache); | |
a1ac9c8a | 1304 | skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true); |
ccdbb6e9 | 1305 | if (clone_it) { |
8c72c65b | 1306 | oskb = skb; |
e2080072 ED |
1307 | |
1308 | tcp_skb_tsorted_save(oskb) { | |
1309 | if (unlikely(skb_cloned(oskb))) | |
1310 | skb = pskb_copy(oskb, gfp_mask); | |
1311 | else | |
1312 | skb = skb_clone(oskb, gfp_mask); | |
1313 | } tcp_skb_tsorted_restore(oskb); | |
1314 | ||
dfb4b9dc DM |
1315 | if (unlikely(!skb)) |
1316 | return -ENOBUFS; | |
b738a185 ED |
1317 | /* retransmit skbs might have a non zero value in skb->dev |
1318 | * because skb->dev is aliased with skb->rbnode.rb_left | |
1319 | */ | |
1320 | skb->dev = NULL; | |
dfb4b9dc | 1321 | } |
5f6188a8 | 1322 | |
dfb4b9dc | 1323 | inet = inet_sk(sk); |
dfb4b9dc | 1324 | tcb = TCP_SKB_CB(skb); |
33ad798c | 1325 | memset(&opts, 0, sizeof(opts)); |
1da177e4 | 1326 | |
1e03d32b | 1327 | tcp_get_current_key(sk, &key); |
051ba674 | 1328 | if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { |
1e03d32b | 1329 | tcp_options_size = tcp_syn_options(sk, skb, &opts, &key); |
051ba674 | 1330 | } else { |
1e03d32b | 1331 | tcp_options_size = tcp_established_options(sk, skb, &opts, &key); |
051ba674 ED |
1332 | /* Force a PSH flag on all (GSO) packets to expedite GRO flush |
1333 | * at receiver : This slightly improve GRO performance. | |
1334 | * Note that we do not force the PSH flag for non GSO packets, | |
1335 | * because they might be sent under high congestion events, | |
1336 | * and in this case it is better to delay the delivery of 1-MSS | |
1337 | * packets and thus the corresponding ACK packet that would | |
1338 | * release the following packet. | |
1339 | */ | |
1340 | if (tcp_skb_pcount(skb) > 1) | |
1341 | tcb->tcp_flags |= TCPHDR_PSH; | |
1342 | } | |
33ad798c | 1343 | tcp_header_size = tcp_options_size + sizeof(struct tcphdr); |
e905a9ed | 1344 | |
726e9e8b ED |
1345 | /* We set skb->ooo_okay to one if this packet can select |
1346 | * a different TX queue than prior packets of this flow, | |
1347 | * to avoid self inflicted reorders. | |
1348 | * The 'other' queue decision is based on current cpu number | |
1349 | * if XPS is enabled, or sk->sk_txhash otherwise. | |
1350 | * We can switch to another (and better) queue if: | |
1351 | * 1) No packet with payload is in qdisc/device queues. | |
1352 | * Delays in TX completion can defeat the test | |
1353 | * even if packets were already sent. | |
1354 | * 2) Or rtx queue is empty. | |
1355 | * This mitigates above case if ACK packets for | |
1356 | * all prior packets were already processed. | |
547669d4 | 1357 | */ |
726e9e8b ED |
1358 | skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1) || |
1359 | tcp_rtx_queue_empty(sk); | |
dfb4b9dc | 1360 | |
38ab52e8 ED |
1361 | /* If we had to use memory reserve to allocate this skb, |
1362 | * this might cause drops if packet is looped back : | |
1363 | * Other socket might not have SOCK_MEMALLOC. | |
1364 | * Packets not looped back do not care about pfmemalloc. | |
1365 | */ | |
1366 | skb->pfmemalloc = 0; | |
1367 | ||
aa8223c7 ACM |
1368 | skb_push(skb, tcp_header_size); |
1369 | skb_reset_transport_header(skb); | |
46d3ceab ED |
1370 | |
1371 | skb_orphan(skb); | |
1372 | skb->sk = sk; | |
1d2077ac | 1373 | skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; |
14afee4b | 1374 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); |
dfb4b9dc | 1375 | |
eb44ad4e | 1376 | skb_set_dst_pending_confirm(skb, READ_ONCE(sk->sk_dst_pending_confirm)); |
c3a2e837 | 1377 | |
dfb4b9dc | 1378 | /* Build TCP header and checksum it. */ |
ea1627c2 | 1379 | th = (struct tcphdr *)skb->data; |
c720c7e8 ED |
1380 | th->source = inet->inet_sport; |
1381 | th->dest = inet->inet_dport; | |
dfb4b9dc | 1382 | th->seq = htonl(tcb->seq); |
2987babb | 1383 | th->ack_seq = htonl(rcv_nxt); |
df7a3b07 | 1384 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
4de075e0 | 1385 | tcb->tcp_flags); |
dfb4b9dc | 1386 | |
dfb4b9dc DM |
1387 | th->check = 0; |
1388 | th->urg_ptr = 0; | |
1da177e4 | 1389 | |
33f5f57e | 1390 | /* The urg_mode check is necessary during a below snd_una win probe */ |
7691367d HX |
1391 | if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { |
1392 | if (before(tp->snd_up, tcb->seq + 0x10000)) { | |
1393 | th->urg_ptr = htons(tp->snd_up - tcb->seq); | |
1394 | th->urg = 1; | |
1395 | } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { | |
0eae88f3 | 1396 | th->urg_ptr = htons(0xFFFF); |
7691367d HX |
1397 | th->urg = 1; |
1398 | } | |
dfb4b9dc | 1399 | } |
1da177e4 | 1400 | |
51466a75 | 1401 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
ea1627c2 ED |
1402 | if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { |
1403 | th->window = htons(tcp_select_window(sk)); | |
1404 | tcp_ecn_send(sk, skb, th, tcp_header_size); | |
1405 | } else { | |
1406 | /* RFC1323: The window in SYN & SYN/ACK segments | |
1407 | * is never scaled. | |
1408 | */ | |
1409 | th->window = htons(min(tp->rcv_wnd, 65535U)); | |
1410 | } | |
fa3fe2b1 | 1411 | |
06b22ef2 | 1412 | tcp_options_write(th, tp, NULL, &opts, &key); |
fa3fe2b1 | 1413 | |
1e03d32b | 1414 | if (tcp_key_is_md5(&key)) { |
cfb6eeb4 | 1415 | #ifdef CONFIG_TCP_MD5SIG |
1e03d32b | 1416 | /* Calculate the MD5 hash, as we have all we need now */ |
aba54656 | 1417 | sk_gso_disable(sk); |
bd0388ae | 1418 | tp->af_specific->calc_md5_hash(opts.hash_location, |
1e03d32b | 1419 | key.md5_key, sk, skb); |
cfb6eeb4 | 1420 | #endif |
1e03d32b DS |
1421 | } else if (tcp_key_is_ao(&key)) { |
1422 | int err; | |
1423 | ||
1424 | err = tcp_ao_transmit_skb(sk, skb, key.ao_key, th, | |
1425 | opts.hash_location); | |
1426 | if (err) { | |
1427 | kfree_skb_reason(skb, SKB_DROP_REASON_NOT_SPECIFIED); | |
1428 | return -ENOMEM; | |
1429 | } | |
1430 | } | |
cfb6eeb4 | 1431 | |
331fca43 MKL |
1432 | /* BPF prog is the last one writing header option */ |
1433 | bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts); | |
1434 | ||
dd2e0b86 ED |
1435 | INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check, |
1436 | tcp_v6_send_check, tcp_v4_send_check, | |
1437 | sk, skb); | |
1da177e4 | 1438 | |
4de075e0 | 1439 | if (likely(tcb->tcp_flags & TCPHDR_ACK)) |
059217c1 | 1440 | tcp_event_ack_sent(sk, rcv_nxt); |
1da177e4 | 1441 | |
a44d6eac | 1442 | if (skb->len != tcp_header_size) { |
cf533ea5 | 1443 | tcp_event_data_sent(tp, sk); |
a44d6eac | 1444 | tp->data_segs_out += tcp_skb_pcount(skb); |
ba113c3a | 1445 | tp->bytes_sent += skb->len - tcp_header_size; |
a44d6eac | 1446 | } |
1da177e4 | 1447 | |
bd37a088 | 1448 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
aa2ea058 TH |
1449 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, |
1450 | tcp_skb_pcount(skb)); | |
1da177e4 | 1451 | |
2efd055c | 1452 | tp->segs_out += tcp_skb_pcount(skb); |
0ae5b43d | 1453 | skb_set_hash_from_sk(skb, sk); |
f69ad292 | 1454 | /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ |
cd7d8498 | 1455 | skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); |
f69ad292 | 1456 | skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); |
cd7d8498 | 1457 | |
d3edd06e | 1458 | /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */ |
971f10ec ED |
1459 | |
1460 | /* Cleanup our debris for IP stacks */ | |
1461 | memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), | |
1462 | sizeof(struct inet6_skb_parm))); | |
1463 | ||
a842fe14 ED |
1464 | tcp_add_tx_delay(skb, tp); |
1465 | ||
05e22e83 ED |
1466 | err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit, |
1467 | inet6_csk_xmit, ip_queue_xmit, | |
1468 | sk, skb, &inet->cork.fl); | |
7faee5c0 | 1469 | |
8c72c65b ED |
1470 | if (unlikely(err > 0)) { |
1471 | tcp_enter_cwr(sk); | |
1472 | err = net_xmit_eval(err); | |
1473 | } | |
fc225799 | 1474 | if (!err && oskb) { |
a7a25630 | 1475 | tcp_update_skb_after_send(sk, oskb, prior_wstamp); |
fc225799 ED |
1476 | tcp_rate_skb_sent(sk, oskb); |
1477 | } | |
8c72c65b | 1478 | return err; |
1da177e4 LT |
1479 | } |
1480 | ||
2987babb YC |
1481 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, |
1482 | gfp_t gfp_mask) | |
1483 | { | |
1484 | return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask, | |
1485 | tcp_sk(sk)->rcv_nxt); | |
1486 | } | |
1487 | ||
67edfef7 | 1488 | /* This routine just queues the buffer for sending. |
1da177e4 LT |
1489 | * |
1490 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, | |
1491 | * otherwise socket can stall. | |
1492 | */ | |
1493 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) | |
1494 | { | |
1495 | struct tcp_sock *tp = tcp_sk(sk); | |
1496 | ||
1497 | /* Advance write_seq and place onto the write_queue. */ | |
0f317464 | 1498 | WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq); |
f4a775d1 | 1499 | __skb_header_release(skb); |
fe067e8a | 1500 | tcp_add_write_queue_tail(sk, skb); |
ab4e846a | 1501 | sk_wmem_queued_add(sk, skb->truesize); |
3ab224be | 1502 | sk_mem_charge(sk, skb->truesize); |
1da177e4 LT |
1503 | } |
1504 | ||
67edfef7 | 1505 | /* Initialize TSO segments for a packet. */ |
d5b38a71 | 1506 | static int tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) |
f6302d1d | 1507 | { |
d5b38a71 ED |
1508 | int tso_segs; |
1509 | ||
4a64fd6c | 1510 | if (skb->len <= mss_now) { |
f6302d1d DM |
1511 | /* Avoid the costly divide in the normal |
1512 | * non-TSO case. | |
1513 | */ | |
f69ad292 | 1514 | TCP_SKB_CB(skb)->tcp_gso_size = 0; |
d5b38a71 ED |
1515 | tcp_skb_pcount_set(skb, 1); |
1516 | return 1; | |
1da177e4 | 1517 | } |
d5b38a71 ED |
1518 | TCP_SKB_CB(skb)->tcp_gso_size = mss_now; |
1519 | tso_segs = DIV_ROUND_UP(skb->len, mss_now); | |
1520 | tcp_skb_pcount_set(skb, tso_segs); | |
1521 | return tso_segs; | |
1da177e4 LT |
1522 | } |
1523 | ||
797108d1 IJ |
1524 | /* Pcount in the middle of the write queue got changed, we need to do various |
1525 | * tweaks to fix counters | |
1526 | */ | |
cf533ea5 | 1527 | static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) |
797108d1 IJ |
1528 | { |
1529 | struct tcp_sock *tp = tcp_sk(sk); | |
1530 | ||
1531 | tp->packets_out -= decr; | |
1532 | ||
1533 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
1534 | tp->sacked_out -= decr; | |
1535 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
1536 | tp->retrans_out -= decr; | |
1537 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) | |
1538 | tp->lost_out -= decr; | |
1539 | ||
1540 | /* Reno case is special. Sigh... */ | |
1541 | if (tcp_is_reno(tp) && decr > 0) | |
1542 | tp->sacked_out -= min_t(u32, tp->sacked_out, decr); | |
1543 | ||
797108d1 IJ |
1544 | if (tp->lost_skb_hint && |
1545 | before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && | |
713bafea | 1546 | (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) |
797108d1 IJ |
1547 | tp->lost_cnt_hint -= decr; |
1548 | ||
1549 | tcp_verify_left_out(tp); | |
1550 | } | |
1551 | ||
0a2cf20c SHY |
1552 | static bool tcp_has_tx_tstamp(const struct sk_buff *skb) |
1553 | { | |
1554 | return TCP_SKB_CB(skb)->txstamp_ack || | |
1555 | (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); | |
1556 | } | |
1557 | ||
490cc7d0 WB |
1558 | static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) |
1559 | { | |
1560 | struct skb_shared_info *shinfo = skb_shinfo(skb); | |
1561 | ||
0a2cf20c | 1562 | if (unlikely(tcp_has_tx_tstamp(skb)) && |
490cc7d0 WB |
1563 | !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { |
1564 | struct skb_shared_info *shinfo2 = skb_shinfo(skb2); | |
1565 | u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; | |
1566 | ||
1567 | shinfo->tx_flags &= ~tsflags; | |
1568 | shinfo2->tx_flags |= tsflags; | |
1569 | swap(shinfo->tskey, shinfo2->tskey); | |
b51e13fa MKL |
1570 | TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; |
1571 | TCP_SKB_CB(skb)->txstamp_ack = 0; | |
490cc7d0 WB |
1572 | } |
1573 | } | |
1574 | ||
a166140e MKL |
1575 | static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) |
1576 | { | |
1577 | TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; | |
1578 | TCP_SKB_CB(skb)->eor = 0; | |
1579 | } | |
1580 | ||
75c119af ED |
1581 | /* Insert buff after skb on the write or rtx queue of sk. */ |
1582 | static void tcp_insert_write_queue_after(struct sk_buff *skb, | |
1583 | struct sk_buff *buff, | |
1584 | struct sock *sk, | |
1585 | enum tcp_queue tcp_queue) | |
1586 | { | |
1587 | if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) | |
1588 | __skb_queue_after(&sk->sk_write_queue, skb, buff); | |
1589 | else | |
1590 | tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); | |
1591 | } | |
1592 | ||
1da177e4 LT |
1593 | /* Function to create two new TCP segments. Shrinks the given segment |
1594 | * to the specified size and appends a new segment with the rest of the | |
e905a9ed | 1595 | * packet to the list. This won't be called frequently, I hope. |
1da177e4 LT |
1596 | * Remember, these are still headerless SKBs at this point. |
1597 | */ | |
75c119af ED |
1598 | int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, |
1599 | struct sk_buff *skb, u32 len, | |
6cc55e09 | 1600 | unsigned int mss_now, gfp_t gfp) |
1da177e4 LT |
1601 | { |
1602 | struct tcp_sock *tp = tcp_sk(sk); | |
1603 | struct sk_buff *buff; | |
b4a24397 | 1604 | int old_factor; |
b617158d | 1605 | long limit; |
b60b49ea | 1606 | int nlen; |
9ce01461 | 1607 | u8 flags; |
1da177e4 | 1608 | |
2fceec13 IJ |
1609 | if (WARN_ON(len > skb->len)) |
1610 | return -EINVAL; | |
6a438bbe | 1611 | |
b4a24397 | 1612 | DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb)); |
1da177e4 | 1613 | |
b617158d ED |
1614 | /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb. |
1615 | * We need some allowance to not penalize applications setting small | |
1616 | * SO_SNDBUF values. | |
1617 | * Also allow first and last skb in retransmit queue to be split. | |
1618 | */ | |
7c4e983c | 1619 | limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_LEGACY_MAX_SIZE); |
b617158d ED |
1620 | if (unlikely((sk->sk_wmem_queued >> 1) > limit && |
1621 | tcp_queue != TCP_FRAG_IN_WRITE_QUEUE && | |
1622 | skb != tcp_rtx_queue_head(sk) && | |
1623 | skb != tcp_rtx_queue_tail(sk))) { | |
f070ef2a ED |
1624 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG); |
1625 | return -ENOMEM; | |
1626 | } | |
1627 | ||
c4777efa | 1628 | if (skb_unclone_keeptruesize(skb, gfp)) |
1da177e4 LT |
1629 | return -ENOMEM; |
1630 | ||
1631 | /* Get a new skb... force flag on. */ | |
5882efff | 1632 | buff = tcp_stream_alloc_skb(sk, gfp, true); |
51456b29 | 1633 | if (!buff) |
1da177e4 | 1634 | return -ENOMEM; /* We'll just try again later. */ |
41477662 | 1635 | skb_copy_decrypted(buff, skb); |
5a369ca6 | 1636 | mptcp_skb_ext_copy(buff, skb); |
ef5cb973 | 1637 | |
ab4e846a | 1638 | sk_wmem_queued_add(sk, buff->truesize); |
3ab224be | 1639 | sk_mem_charge(sk, buff->truesize); |
b4a24397 | 1640 | nlen = skb->len - len; |
b60b49ea HX |
1641 | buff->truesize += nlen; |
1642 | skb->truesize -= nlen; | |
1da177e4 LT |
1643 | |
1644 | /* Correct the sequence numbers. */ | |
1645 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1646 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1647 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1648 | ||
1649 | /* PSH and FIN should only be set in the second packet. */ | |
4de075e0 ED |
1650 | flags = TCP_SKB_CB(skb)->tcp_flags; |
1651 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
1652 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
e14c3caf | 1653 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
a166140e | 1654 | tcp_skb_fragment_eor(skb, buff); |
1da177e4 | 1655 | |
98be9b12 | 1656 | skb_split(skb, buff, len); |
1da177e4 | 1657 | |
a1ac9c8a | 1658 | skb_set_delivery_time(buff, skb->tstamp, true); |
490cc7d0 | 1659 | tcp_fragment_tstamp(skb, buff); |
1da177e4 | 1660 | |
6475be16 DM |
1661 | old_factor = tcp_skb_pcount(skb); |
1662 | ||
1da177e4 | 1663 | /* Fix up tso_factor for both original and new SKB. */ |
5bbb432c ED |
1664 | tcp_set_skb_tso_segs(skb, mss_now); |
1665 | tcp_set_skb_tso_segs(buff, mss_now); | |
1da177e4 | 1666 | |
b9f64820 YC |
1667 | /* Update delivered info for the new segment */ |
1668 | TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; | |
1669 | ||
6475be16 DM |
1670 | /* If this packet has been sent out already, we must |
1671 | * adjust the various packet counters. | |
1672 | */ | |
cf0b450c | 1673 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
6475be16 DM |
1674 | int diff = old_factor - tcp_skb_pcount(skb) - |
1675 | tcp_skb_pcount(buff); | |
1da177e4 | 1676 | |
797108d1 IJ |
1677 | if (diff) |
1678 | tcp_adjust_pcount(sk, skb, diff); | |
1da177e4 LT |
1679 | } |
1680 | ||
1681 | /* Link BUFF into the send queue. */ | |
f4a775d1 | 1682 | __skb_header_release(buff); |
75c119af | 1683 | tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); |
f67971e6 ED |
1684 | if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) |
1685 | list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); | |
1da177e4 LT |
1686 | |
1687 | return 0; | |
1688 | } | |
1689 | ||
f4d01666 ED |
1690 | /* This is similar to __pskb_pull_tail(). The difference is that pulled |
1691 | * data is not copied, but immediately discarded. | |
1da177e4 | 1692 | */ |
7162fb24 | 1693 | static int __pskb_trim_head(struct sk_buff *skb, int len) |
1da177e4 | 1694 | { |
7b7fc97a | 1695 | struct skb_shared_info *shinfo; |
1da177e4 LT |
1696 | int i, k, eat; |
1697 | ||
b4a24397 | 1698 | DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb)); |
1da177e4 LT |
1699 | eat = len; |
1700 | k = 0; | |
7b7fc97a ED |
1701 | shinfo = skb_shinfo(skb); |
1702 | for (i = 0; i < shinfo->nr_frags; i++) { | |
1703 | int size = skb_frag_size(&shinfo->frags[i]); | |
9e903e08 ED |
1704 | |
1705 | if (size <= eat) { | |
aff65da0 | 1706 | skb_frag_unref(skb, i); |
9e903e08 | 1707 | eat -= size; |
1da177e4 | 1708 | } else { |
7b7fc97a | 1709 | shinfo->frags[k] = shinfo->frags[i]; |
1da177e4 | 1710 | if (eat) { |
b54c9d5b | 1711 | skb_frag_off_add(&shinfo->frags[k], eat); |
7b7fc97a | 1712 | skb_frag_size_sub(&shinfo->frags[k], eat); |
1da177e4 LT |
1713 | eat = 0; |
1714 | } | |
1715 | k++; | |
1716 | } | |
1717 | } | |
7b7fc97a | 1718 | shinfo->nr_frags = k; |
1da177e4 | 1719 | |
1da177e4 LT |
1720 | skb->data_len -= len; |
1721 | skb->len = skb->data_len; | |
7162fb24 | 1722 | return len; |
1da177e4 LT |
1723 | } |
1724 | ||
67edfef7 | 1725 | /* Remove acked data from a packet in the transmit queue. */ |
1da177e4 LT |
1726 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) |
1727 | { | |
7162fb24 ED |
1728 | u32 delta_truesize; |
1729 | ||
c4777efa | 1730 | if (skb_unclone_keeptruesize(skb, GFP_ATOMIC)) |
1da177e4 LT |
1731 | return -ENOMEM; |
1732 | ||
7162fb24 | 1733 | delta_truesize = __pskb_trim_head(skb, len); |
1da177e4 LT |
1734 | |
1735 | TCP_SKB_CB(skb)->seq += len; | |
1da177e4 | 1736 | |
b4a24397 ED |
1737 | skb->truesize -= delta_truesize; |
1738 | sk_wmem_queued_add(sk, -delta_truesize); | |
1739 | if (!skb_zcopy_pure(skb)) | |
1740 | sk_mem_uncharge(sk, delta_truesize); | |
1da177e4 | 1741 | |
5b35e1e6 | 1742 | /* Any change of skb->len requires recalculation of tso factor. */ |
1da177e4 | 1743 | if (tcp_skb_pcount(skb) > 1) |
5bbb432c | 1744 | tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); |
1da177e4 LT |
1745 | |
1746 | return 0; | |
1747 | } | |
1748 | ||
1b63edd6 YC |
1749 | /* Calculate MSS not accounting any TCP options. */ |
1750 | static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
5d424d5a | 1751 | { |
cf533ea5 ED |
1752 | const struct tcp_sock *tp = tcp_sk(sk); |
1753 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
5d424d5a JH |
1754 | int mss_now; |
1755 | ||
1756 | /* Calculate base mss without TCP options: | |
1757 | It is MMS_S - sizeof(tcphdr) of rfc1122 | |
1758 | */ | |
1759 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); | |
1760 | ||
1761 | /* Clamp it (mss_clamp does not include tcp options) */ | |
1762 | if (mss_now > tp->rx_opt.mss_clamp) | |
1763 | mss_now = tp->rx_opt.mss_clamp; | |
1764 | ||
1765 | /* Now subtract optional transport overhead */ | |
1766 | mss_now -= icsk->icsk_ext_hdr_len; | |
1767 | ||
1768 | /* Then reserve room for full set of TCP options and 8 bytes of data */ | |
78eb166c KI |
1769 | mss_now = max(mss_now, |
1770 | READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss)); | |
5d424d5a JH |
1771 | return mss_now; |
1772 | } | |
1773 | ||
1b63edd6 YC |
1774 | /* Calculate MSS. Not accounting for SACKs here. */ |
1775 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
1776 | { | |
1777 | /* Subtract TCP options size, not including SACKs */ | |
1778 | return __tcp_mtu_to_mss(sk, pmtu) - | |
1779 | (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); | |
1780 | } | |
c7bb4b89 | 1781 | EXPORT_SYMBOL(tcp_mtu_to_mss); |
1b63edd6 | 1782 | |
5d424d5a | 1783 | /* Inverse of above */ |
67469601 | 1784 | int tcp_mss_to_mtu(struct sock *sk, int mss) |
5d424d5a | 1785 | { |
cf533ea5 ED |
1786 | const struct tcp_sock *tp = tcp_sk(sk); |
1787 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
5d424d5a | 1788 | |
e57a3447 | 1789 | return mss + |
5d424d5a JH |
1790 | tp->tcp_header_len + |
1791 | icsk->icsk_ext_hdr_len + | |
1792 | icsk->icsk_af_ops->net_header_len; | |
5d424d5a | 1793 | } |
556c6b46 | 1794 | EXPORT_SYMBOL(tcp_mss_to_mtu); |
5d424d5a | 1795 | |
67edfef7 | 1796 | /* MTU probing init per socket */ |
5d424d5a JH |
1797 | void tcp_mtup_init(struct sock *sk) |
1798 | { | |
1799 | struct tcp_sock *tp = tcp_sk(sk); | |
1800 | struct inet_connection_sock *icsk = inet_csk(sk); | |
b0f9ca53 | 1801 | struct net *net = sock_net(sk); |
5d424d5a | 1802 | |
f47d00e0 | 1803 | icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1; |
5d424d5a | 1804 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + |
e905a9ed | 1805 | icsk->icsk_af_ops->net_header_len; |
88d78bc0 | 1806 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss)); |
5d424d5a | 1807 | icsk->icsk_mtup.probe_size = 0; |
05cbc0db | 1808 | if (icsk->icsk_mtup.enabled) |
c74df29a | 1809 | icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; |
5d424d5a | 1810 | } |
4bc2f18b | 1811 | EXPORT_SYMBOL(tcp_mtup_init); |
5d424d5a | 1812 | |
1da177e4 LT |
1813 | /* This function synchronize snd mss to current pmtu/exthdr set. |
1814 | ||
1815 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts | |
1816 | for TCP options, but includes only bare TCP header. | |
1817 | ||
1818 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. | |
caa20d9a | 1819 | It is minimum of user_mss and mss received with SYN. |
1da177e4 LT |
1820 | It also does not include TCP options. |
1821 | ||
d83d8461 | 1822 | inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. |
1da177e4 LT |
1823 | |
1824 | tp->mss_cache is current effective sending mss, including | |
1825 | all tcp options except for SACKs. It is evaluated, | |
1826 | taking into account current pmtu, but never exceeds | |
1827 | tp->rx_opt.mss_clamp. | |
1828 | ||
1829 | NOTE1. rfc1122 clearly states that advertised MSS | |
1830 | DOES NOT include either tcp or ip options. | |
1831 | ||
d83d8461 ACM |
1832 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
1833 | are READ ONLY outside this function. --ANK (980731) | |
1da177e4 | 1834 | */ |
1da177e4 LT |
1835 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) |
1836 | { | |
1837 | struct tcp_sock *tp = tcp_sk(sk); | |
d83d8461 | 1838 | struct inet_connection_sock *icsk = inet_csk(sk); |
5d424d5a | 1839 | int mss_now; |
1da177e4 | 1840 | |
5d424d5a JH |
1841 | if (icsk->icsk_mtup.search_high > pmtu) |
1842 | icsk->icsk_mtup.search_high = pmtu; | |
1da177e4 | 1843 | |
5d424d5a | 1844 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
409d22b4 | 1845 | mss_now = tcp_bound_to_half_wnd(tp, mss_now); |
1da177e4 LT |
1846 | |
1847 | /* And store cached results */ | |
d83d8461 | 1848 | icsk->icsk_pmtu_cookie = pmtu; |
5d424d5a JH |
1849 | if (icsk->icsk_mtup.enabled) |
1850 | mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); | |
c1b4a7e6 | 1851 | tp->mss_cache = mss_now; |
1da177e4 LT |
1852 | |
1853 | return mss_now; | |
1854 | } | |
4bc2f18b | 1855 | EXPORT_SYMBOL(tcp_sync_mss); |
1da177e4 LT |
1856 | |
1857 | /* Compute the current effective MSS, taking SACKs and IP options, | |
1858 | * and even PMTU discovery events into account. | |
1da177e4 | 1859 | */ |
0c54b85f | 1860 | unsigned int tcp_current_mss(struct sock *sk) |
1da177e4 | 1861 | { |
cf533ea5 ED |
1862 | const struct tcp_sock *tp = tcp_sk(sk); |
1863 | const struct dst_entry *dst = __sk_dst_get(sk); | |
c1b4a7e6 | 1864 | u32 mss_now; |
95c96174 | 1865 | unsigned int header_len; |
33ad798c | 1866 | struct tcp_out_options opts; |
1e03d32b | 1867 | struct tcp_key key; |
c1b4a7e6 DM |
1868 | |
1869 | mss_now = tp->mss_cache; | |
1870 | ||
1da177e4 LT |
1871 | if (dst) { |
1872 | u32 mtu = dst_mtu(dst); | |
d83d8461 | 1873 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
1da177e4 LT |
1874 | mss_now = tcp_sync_mss(sk, mtu); |
1875 | } | |
1e03d32b DS |
1876 | tcp_get_current_key(sk, &key); |
1877 | header_len = tcp_established_options(sk, NULL, &opts, &key) + | |
33ad798c AL |
1878 | sizeof(struct tcphdr); |
1879 | /* The mss_cache is sized based on tp->tcp_header_len, which assumes | |
1880 | * some common options. If this is an odd packet (because we have SACK | |
1881 | * blocks etc) then our calculated header_len will be different, and | |
1882 | * we have to adjust mss_now correspondingly */ | |
1883 | if (header_len != tp->tcp_header_len) { | |
1884 | int delta = (int) header_len - tp->tcp_header_len; | |
1885 | mss_now -= delta; | |
1886 | } | |
cfb6eeb4 | 1887 | |
1da177e4 LT |
1888 | return mss_now; |
1889 | } | |
1890 | ||
86fd14ad WP |
1891 | /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. |
1892 | * As additional protections, we do not touch cwnd in retransmission phases, | |
1893 | * and if application hit its sndbuf limit recently. | |
1894 | */ | |
1895 | static void tcp_cwnd_application_limited(struct sock *sk) | |
1896 | { | |
1897 | struct tcp_sock *tp = tcp_sk(sk); | |
1898 | ||
1899 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && | |
1900 | sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { | |
1901 | /* Limited by application or receiver window. */ | |
1902 | u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); | |
1903 | u32 win_used = max(tp->snd_cwnd_used, init_win); | |
40570375 | 1904 | if (win_used < tcp_snd_cwnd(tp)) { |
86fd14ad | 1905 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
40570375 | 1906 | tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1); |
86fd14ad WP |
1907 | } |
1908 | tp->snd_cwnd_used = 0; | |
1909 | } | |
c2203cf7 | 1910 | tp->snd_cwnd_stamp = tcp_jiffies32; |
86fd14ad WP |
1911 | } |
1912 | ||
ca8a2263 | 1913 | static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) |
a762a980 | 1914 | { |
1b1fc3fd | 1915 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; |
9e412ba7 | 1916 | struct tcp_sock *tp = tcp_sk(sk); |
a762a980 | 1917 | |
f4ce91ce NC |
1918 | /* Track the strongest available signal of the degree to which the cwnd |
1919 | * is fully utilized. If cwnd-limited then remember that fact for the | |
1920 | * current window. If not cwnd-limited then track the maximum number of | |
1921 | * outstanding packets in the current window. (If cwnd-limited then we | |
1922 | * chose to not update tp->max_packets_out to avoid an extra else | |
1923 | * clause with no functional impact.) | |
ca8a2263 | 1924 | */ |
f4ce91ce NC |
1925 | if (!before(tp->snd_una, tp->cwnd_usage_seq) || |
1926 | is_cwnd_limited || | |
1927 | (!tp->is_cwnd_limited && | |
1928 | tp->packets_out > tp->max_packets_out)) { | |
ca8a2263 | 1929 | tp->is_cwnd_limited = is_cwnd_limited; |
f4ce91ce NC |
1930 | tp->max_packets_out = tp->packets_out; |
1931 | tp->cwnd_usage_seq = tp->snd_nxt; | |
ca8a2263 | 1932 | } |
e114a710 | 1933 | |
24901551 | 1934 | if (tcp_is_cwnd_limited(sk)) { |
a762a980 DM |
1935 | /* Network is feed fully. */ |
1936 | tp->snd_cwnd_used = 0; | |
c2203cf7 | 1937 | tp->snd_cwnd_stamp = tcp_jiffies32; |
a762a980 DM |
1938 | } else { |
1939 | /* Network starves. */ | |
1940 | if (tp->packets_out > tp->snd_cwnd_used) | |
1941 | tp->snd_cwnd_used = tp->packets_out; | |
1942 | ||
4845b571 | 1943 | if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) && |
c2203cf7 | 1944 | (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && |
1b1fc3fd | 1945 | !ca_ops->cong_control) |
a762a980 | 1946 | tcp_cwnd_application_limited(sk); |
b0f71bd3 FY |
1947 | |
1948 | /* The following conditions together indicate the starvation | |
1949 | * is caused by insufficient sender buffer: | |
1950 | * 1) just sent some data (see tcp_write_xmit) | |
1951 | * 2) not cwnd limited (this else condition) | |
75c119af | 1952 | * 3) no more data to send (tcp_write_queue_empty()) |
b0f71bd3 FY |
1953 | * 4) application is hitting buffer limit (SOCK_NOSPACE) |
1954 | */ | |
75c119af | 1955 | if (tcp_write_queue_empty(sk) && sk->sk_socket && |
b0f71bd3 FY |
1956 | test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && |
1957 | (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) | |
1958 | tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); | |
a762a980 DM |
1959 | } |
1960 | } | |
1961 | ||
d4589926 ED |
1962 | /* Minshall's variant of the Nagle send check. */ |
1963 | static bool tcp_minshall_check(const struct tcp_sock *tp) | |
1964 | { | |
1965 | return after(tp->snd_sml, tp->snd_una) && | |
1966 | !after(tp->snd_sml, tp->snd_nxt); | |
1967 | } | |
1968 | ||
1969 | /* Update snd_sml if this skb is under mss | |
1970 | * Note that a TSO packet might end with a sub-mss segment | |
1971 | * The test is really : | |
1972 | * if ((skb->len % mss) != 0) | |
1973 | * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; | |
1974 | * But we can avoid doing the divide again given we already have | |
1975 | * skb_pcount = skb->len / mss_now | |
0e3a4803 | 1976 | */ |
d4589926 ED |
1977 | static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, |
1978 | const struct sk_buff *skb) | |
1979 | { | |
1980 | if (skb->len < tcp_skb_pcount(skb) * mss_now) | |
1981 | tp->snd_sml = TCP_SKB_CB(skb)->end_seq; | |
1982 | } | |
1983 | ||
1984 | /* Return false, if packet can be sent now without violation Nagle's rules: | |
1985 | * 1. It is full sized. (provided by caller in %partial bool) | |
1986 | * 2. Or it contains FIN. (already checked by caller) | |
1987 | * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. | |
1988 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. | |
1989 | * With Minshall's modification: all sent small packets are ACKed. | |
1990 | */ | |
1991 | static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, | |
cc93fc51 | 1992 | int nonagle) |
d4589926 ED |
1993 | { |
1994 | return partial && | |
1995 | ((nonagle & TCP_NAGLE_CORK) || | |
1996 | (!nonagle && tp->packets_out && tcp_minshall_check(tp))); | |
1997 | } | |
605ad7f1 ED |
1998 | |
1999 | /* Return how many segs we'd like on a TSO packet, | |
65466904 ED |
2000 | * depending on current pacing rate, and how close the peer is. |
2001 | * | |
2002 | * Rationale is: | |
2003 | * - For close peers, we rather send bigger packets to reduce | |
2004 | * cpu costs, because occasional losses will be repaired fast. | |
2005 | * - For long distance/rtt flows, we would like to get ACK clocking | |
2006 | * with 1 ACK per ms. | |
2007 | * | |
2008 | * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting | |
2009 | * in bigger TSO bursts. We we cut the RTT-based allowance in half | |
2010 | * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance | |
2011 | * is below 1500 bytes after 6 * ~500 usec = 3ms. | |
605ad7f1 | 2012 | */ |
dcb8c9b4 ED |
2013 | static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, |
2014 | int min_tso_segs) | |
605ad7f1 | 2015 | { |
65466904 ED |
2016 | unsigned long bytes; |
2017 | u32 r; | |
605ad7f1 | 2018 | |
28b24f90 | 2019 | bytes = READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift); |
605ad7f1 | 2020 | |
2455e61b | 2021 | r = tcp_min_rtt(tcp_sk(sk)) >> READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_rtt_log); |
65466904 ED |
2022 | if (r < BITS_PER_TYPE(sk->sk_gso_max_size)) |
2023 | bytes += sk->sk_gso_max_size >> r; | |
2024 | ||
2025 | bytes = min_t(unsigned long, bytes, sk->sk_gso_max_size); | |
605ad7f1 | 2026 | |
65466904 | 2027 | return max_t(u32, bytes / mss_now, min_tso_segs); |
605ad7f1 ED |
2028 | } |
2029 | ||
ed6e7268 NC |
2030 | /* Return the number of segments we want in the skb we are transmitting. |
2031 | * See if congestion control module wants to decide; otherwise, autosize. | |
2032 | */ | |
2033 | static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) | |
2034 | { | |
2035 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; | |
dcb8c9b4 | 2036 | u32 min_tso, tso_segs; |
ed6e7268 | 2037 | |
dcb8c9b4 ED |
2038 | min_tso = ca_ops->min_tso_segs ? |
2039 | ca_ops->min_tso_segs(sk) : | |
e0bb4ab9 | 2040 | READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs); |
dcb8c9b4 ED |
2041 | |
2042 | tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); | |
350c9f48 | 2043 | return min_t(u32, tso_segs, sk->sk_gso_max_segs); |
ed6e7268 NC |
2044 | } |
2045 | ||
d4589926 ED |
2046 | /* Returns the portion of skb which can be sent right away */ |
2047 | static unsigned int tcp_mss_split_point(const struct sock *sk, | |
2048 | const struct sk_buff *skb, | |
2049 | unsigned int mss_now, | |
2050 | unsigned int max_segs, | |
2051 | int nonagle) | |
c1b4a7e6 | 2052 | { |
cf533ea5 | 2053 | const struct tcp_sock *tp = tcp_sk(sk); |
d4589926 | 2054 | u32 partial, needed, window, max_len; |
c1b4a7e6 | 2055 | |
90840def | 2056 | window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
1485348d | 2057 | max_len = mss_now * max_segs; |
0e3a4803 | 2058 | |
1485348d BH |
2059 | if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) |
2060 | return max_len; | |
0e3a4803 | 2061 | |
5ea3a748 IJ |
2062 | needed = min(skb->len, window); |
2063 | ||
1485348d BH |
2064 | if (max_len <= needed) |
2065 | return max_len; | |
0e3a4803 | 2066 | |
d4589926 ED |
2067 | partial = needed % mss_now; |
2068 | /* If last segment is not a full MSS, check if Nagle rules allow us | |
2069 | * to include this last segment in this skb. | |
2070 | * Otherwise, we'll split the skb at last MSS boundary | |
2071 | */ | |
cc93fc51 | 2072 | if (tcp_nagle_check(partial != 0, tp, nonagle)) |
d4589926 ED |
2073 | return needed - partial; |
2074 | ||
2075 | return needed; | |
c1b4a7e6 DM |
2076 | } |
2077 | ||
2078 | /* Can at least one segment of SKB be sent right now, according to the | |
2079 | * congestion window rules? If so, return how many segments are allowed. | |
2080 | */ | |
22555032 | 2081 | static u32 tcp_cwnd_test(const struct tcp_sock *tp) |
c1b4a7e6 | 2082 | { |
d649a7a8 | 2083 | u32 in_flight, cwnd, halfcwnd; |
c1b4a7e6 | 2084 | |
c1b4a7e6 | 2085 | in_flight = tcp_packets_in_flight(tp); |
40570375 | 2086 | cwnd = tcp_snd_cwnd(tp); |
d649a7a8 ED |
2087 | if (in_flight >= cwnd) |
2088 | return 0; | |
c1b4a7e6 | 2089 | |
d649a7a8 ED |
2090 | /* For better scheduling, ensure we have at least |
2091 | * 2 GSO packets in flight. | |
2092 | */ | |
2093 | halfcwnd = max(cwnd >> 1, 1U); | |
2094 | return min(halfcwnd, cwnd - in_flight); | |
c1b4a7e6 DM |
2095 | } |
2096 | ||
b595076a | 2097 | /* Initialize TSO state of a skb. |
67edfef7 | 2098 | * This must be invoked the first time we consider transmitting |
c1b4a7e6 DM |
2099 | * SKB onto the wire. |
2100 | */ | |
5bbb432c | 2101 | static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) |
c1b4a7e6 DM |
2102 | { |
2103 | int tso_segs = tcp_skb_pcount(skb); | |
2104 | ||
d5b38a71 ED |
2105 | if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) |
2106 | return tcp_set_skb_tso_segs(skb, mss_now); | |
2107 | ||
c1b4a7e6 DM |
2108 | return tso_segs; |
2109 | } | |
2110 | ||
c1b4a7e6 | 2111 | |
a2a385d6 | 2112 | /* Return true if the Nagle test allows this packet to be |
c1b4a7e6 DM |
2113 | * sent now. |
2114 | */ | |
a2a385d6 ED |
2115 | static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, |
2116 | unsigned int cur_mss, int nonagle) | |
c1b4a7e6 DM |
2117 | { |
2118 | /* Nagle rule does not apply to frames, which sit in the middle of the | |
2119 | * write_queue (they have no chances to get new data). | |
2120 | * | |
2121 | * This is implemented in the callers, where they modify the 'nonagle' | |
2122 | * argument based upon the location of SKB in the send queue. | |
2123 | */ | |
2124 | if (nonagle & TCP_NAGLE_PUSH) | |
a2a385d6 | 2125 | return true; |
c1b4a7e6 | 2126 | |
9b44190d YC |
2127 | /* Don't use the nagle rule for urgent data (or for the final FIN). */ |
2128 | if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) | |
a2a385d6 | 2129 | return true; |
c1b4a7e6 | 2130 | |
cc93fc51 | 2131 | if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) |
a2a385d6 | 2132 | return true; |
c1b4a7e6 | 2133 | |
a2a385d6 | 2134 | return false; |
c1b4a7e6 DM |
2135 | } |
2136 | ||
2137 | /* Does at least the first segment of SKB fit into the send window? */ | |
a2a385d6 ED |
2138 | static bool tcp_snd_wnd_test(const struct tcp_sock *tp, |
2139 | const struct sk_buff *skb, | |
2140 | unsigned int cur_mss) | |
c1b4a7e6 DM |
2141 | { |
2142 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
2143 | ||
2144 | if (skb->len > cur_mss) | |
2145 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; | |
2146 | ||
90840def | 2147 | return !after(end_seq, tcp_wnd_end(tp)); |
c1b4a7e6 DM |
2148 | } |
2149 | ||
c1b4a7e6 DM |
2150 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet |
2151 | * which is put after SKB on the list. It is very much like | |
2152 | * tcp_fragment() except that it may make several kinds of assumptions | |
2153 | * in order to speed up the splitting operation. In particular, we | |
2154 | * know that all the data is in scatter-gather pages, and that the | |
2155 | * packet has never been sent out before (and thus is not cloned). | |
2156 | */ | |
56483341 | 2157 | static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, |
c4ead4c5 | 2158 | unsigned int mss_now, gfp_t gfp) |
c1b4a7e6 | 2159 | { |
c1b4a7e6 | 2160 | int nlen = skb->len - len; |
56483341 | 2161 | struct sk_buff *buff; |
9ce01461 | 2162 | u8 flags; |
c1b4a7e6 DM |
2163 | |
2164 | /* All of a TSO frame must be composed of paged data. */ | |
b4a24397 | 2165 | DEBUG_NET_WARN_ON_ONCE(skb->len != skb->data_len); |
c1b4a7e6 | 2166 | |
5882efff | 2167 | buff = tcp_stream_alloc_skb(sk, gfp, true); |
51456b29 | 2168 | if (unlikely(!buff)) |
c1b4a7e6 | 2169 | return -ENOMEM; |
41477662 | 2170 | skb_copy_decrypted(buff, skb); |
5a369ca6 | 2171 | mptcp_skb_ext_copy(buff, skb); |
c1b4a7e6 | 2172 | |
ab4e846a | 2173 | sk_wmem_queued_add(sk, buff->truesize); |
3ab224be | 2174 | sk_mem_charge(sk, buff->truesize); |
b60b49ea | 2175 | buff->truesize += nlen; |
c1b4a7e6 DM |
2176 | skb->truesize -= nlen; |
2177 | ||
2178 | /* Correct the sequence numbers. */ | |
2179 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
2180 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
2181 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
2182 | ||
2183 | /* PSH and FIN should only be set in the second packet. */ | |
4de075e0 ED |
2184 | flags = TCP_SKB_CB(skb)->tcp_flags; |
2185 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
2186 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
c1b4a7e6 | 2187 | |
a166140e MKL |
2188 | tcp_skb_fragment_eor(skb, buff); |
2189 | ||
c1b4a7e6 | 2190 | skb_split(skb, buff, len); |
490cc7d0 | 2191 | tcp_fragment_tstamp(skb, buff); |
c1b4a7e6 DM |
2192 | |
2193 | /* Fix up tso_factor for both original and new SKB. */ | |
5bbb432c ED |
2194 | tcp_set_skb_tso_segs(skb, mss_now); |
2195 | tcp_set_skb_tso_segs(buff, mss_now); | |
c1b4a7e6 DM |
2196 | |
2197 | /* Link BUFF into the send queue. */ | |
f4a775d1 | 2198 | __skb_header_release(buff); |
56483341 | 2199 | tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE); |
c1b4a7e6 DM |
2200 | |
2201 | return 0; | |
2202 | } | |
2203 | ||
2204 | /* Try to defer sending, if possible, in order to minimize the amount | |
2205 | * of TSO splitting we do. View it as a kind of TSO Nagle test. | |
2206 | * | |
2207 | * This algorithm is from John Heffner. | |
2208 | */ | |
ca8a2263 | 2209 | static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, |
f9bfe4e6 ED |
2210 | bool *is_cwnd_limited, |
2211 | bool *is_rwnd_limited, | |
2212 | u32 max_segs) | |
c1b4a7e6 | 2213 | { |
6687e988 | 2214 | const struct inet_connection_sock *icsk = inet_csk(sk); |
f1c6ea38 | 2215 | u32 send_win, cong_win, limit, in_flight; |
50c8339e | 2216 | struct tcp_sock *tp = tcp_sk(sk); |
50c8339e | 2217 | struct sk_buff *head; |
ad9f4f50 | 2218 | int win_divisor; |
f1c6ea38 | 2219 | s64 delta; |
c1b4a7e6 | 2220 | |
99d7662a | 2221 | if (icsk->icsk_ca_state >= TCP_CA_Recovery) |
ae8064ac JH |
2222 | goto send_now; |
2223 | ||
5f852eb5 | 2224 | /* Avoid bursty behavior by allowing defer |
a682850a ED |
2225 | * only if the last write was recent (1 ms). |
2226 | * Note that tp->tcp_wstamp_ns can be in the future if we have | |
2227 | * packets waiting in a qdisc or device for EDT delivery. | |
5f852eb5 | 2228 | */ |
a682850a ED |
2229 | delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC; |
2230 | if (delta > 0) | |
ae8064ac | 2231 | goto send_now; |
908a75c1 | 2232 | |
c1b4a7e6 DM |
2233 | in_flight = tcp_packets_in_flight(tp); |
2234 | ||
c8c9aeb5 | 2235 | BUG_ON(tcp_skb_pcount(skb) <= 1); |
40570375 | 2236 | BUG_ON(tcp_snd_cwnd(tp) <= in_flight); |
c1b4a7e6 | 2237 | |
90840def | 2238 | send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
c1b4a7e6 DM |
2239 | |
2240 | /* From in_flight test above, we know that cwnd > in_flight. */ | |
40570375 | 2241 | cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache; |
c1b4a7e6 DM |
2242 | |
2243 | limit = min(send_win, cong_win); | |
2244 | ||
ba244fe9 | 2245 | /* If a full-sized TSO skb can be sent, do it. */ |
605ad7f1 | 2246 | if (limit >= max_segs * tp->mss_cache) |
ae8064ac | 2247 | goto send_now; |
ba244fe9 | 2248 | |
62ad2761 IJ |
2249 | /* Middle in queue won't get any more data, full sendable already? */ |
2250 | if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) | |
2251 | goto send_now; | |
2252 | ||
5bbcc0f5 | 2253 | win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); |
ad9f4f50 | 2254 | if (win_divisor) { |
40570375 | 2255 | u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache); |
c1b4a7e6 DM |
2256 | |
2257 | /* If at least some fraction of a window is available, | |
2258 | * just use it. | |
2259 | */ | |
ad9f4f50 | 2260 | chunk /= win_divisor; |
c1b4a7e6 | 2261 | if (limit >= chunk) |
ae8064ac | 2262 | goto send_now; |
c1b4a7e6 DM |
2263 | } else { |
2264 | /* Different approach, try not to defer past a single | |
2265 | * ACK. Receiver should ACK every other full sized | |
2266 | * frame, so if we have space for more than 3 frames | |
2267 | * then send now. | |
2268 | */ | |
6b5a5c0d | 2269 | if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) |
ae8064ac | 2270 | goto send_now; |
c1b4a7e6 DM |
2271 | } |
2272 | ||
75c119af ED |
2273 | /* TODO : use tsorted_sent_queue ? */ |
2274 | head = tcp_rtx_queue_head(sk); | |
2275 | if (!head) | |
2276 | goto send_now; | |
f1c6ea38 | 2277 | delta = tp->tcp_clock_cache - head->tstamp; |
50c8339e | 2278 | /* If next ACK is likely to come too late (half srtt), do not defer */ |
f1c6ea38 | 2279 | if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0) |
50c8339e ED |
2280 | goto send_now; |
2281 | ||
f9bfe4e6 ED |
2282 | /* Ok, it looks like it is advisable to defer. |
2283 | * Three cases are tracked : | |
2284 | * 1) We are cwnd-limited | |
2285 | * 2) We are rwnd-limited | |
2286 | * 3) We are application limited. | |
2287 | */ | |
2288 | if (cong_win < send_win) { | |
2289 | if (cong_win <= skb->len) { | |
2290 | *is_cwnd_limited = true; | |
2291 | return true; | |
2292 | } | |
2293 | } else { | |
2294 | if (send_win <= skb->len) { | |
2295 | *is_rwnd_limited = true; | |
2296 | return true; | |
2297 | } | |
2298 | } | |
ae8064ac | 2299 | |
f9bfe4e6 | 2300 | /* If this packet won't get more data, do not wait. */ |
d8ed257f ED |
2301 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) || |
2302 | TCP_SKB_CB(skb)->eor) | |
f9bfe4e6 | 2303 | goto send_now; |
ca8a2263 | 2304 | |
a2a385d6 | 2305 | return true; |
ae8064ac JH |
2306 | |
2307 | send_now: | |
a2a385d6 | 2308 | return false; |
c1b4a7e6 DM |
2309 | } |
2310 | ||
05cbc0db FD |
2311 | static inline void tcp_mtu_check_reprobe(struct sock *sk) |
2312 | { | |
2313 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2314 | struct tcp_sock *tp = tcp_sk(sk); | |
2315 | struct net *net = sock_net(sk); | |
2316 | u32 interval; | |
2317 | s32 delta; | |
2318 | ||
2a85388f | 2319 | interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval); |
c74df29a | 2320 | delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; |
05cbc0db FD |
2321 | if (unlikely(delta >= interval * HZ)) { |
2322 | int mss = tcp_current_mss(sk); | |
2323 | ||
2324 | /* Update current search range */ | |
2325 | icsk->icsk_mtup.probe_size = 0; | |
2326 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + | |
2327 | sizeof(struct tcphdr) + | |
2328 | icsk->icsk_af_ops->net_header_len; | |
2329 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); | |
2330 | ||
2331 | /* Update probe time stamp */ | |
c74df29a | 2332 | icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; |
05cbc0db FD |
2333 | } |
2334 | } | |
2335 | ||
808cf9e3 IL |
2336 | static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) |
2337 | { | |
2338 | struct sk_buff *skb, *next; | |
2339 | ||
2340 | skb = tcp_send_head(sk); | |
2341 | tcp_for_write_queue_from_safe(skb, next, sk) { | |
2342 | if (len <= skb->len) | |
2343 | break; | |
2344 | ||
9b65b17d TA |
2345 | if (unlikely(TCP_SKB_CB(skb)->eor) || |
2346 | tcp_has_tx_tstamp(skb) || | |
2347 | !skb_pure_zcopy_same(skb, next)) | |
808cf9e3 IL |
2348 | return false; |
2349 | ||
2350 | len -= skb->len; | |
2351 | } | |
2352 | ||
2353 | return true; | |
2354 | } | |
2355 | ||
73601329 ED |
2356 | static int tcp_clone_payload(struct sock *sk, struct sk_buff *to, |
2357 | int probe_size) | |
2358 | { | |
2359 | skb_frag_t *lastfrag = NULL, *fragto = skb_shinfo(to)->frags; | |
2360 | int i, todo, len = 0, nr_frags = 0; | |
2361 | const struct sk_buff *skb; | |
2362 | ||
2363 | if (!sk_wmem_schedule(sk, to->truesize + probe_size)) | |
2364 | return -ENOMEM; | |
2365 | ||
2366 | skb_queue_walk(&sk->sk_write_queue, skb) { | |
2367 | const skb_frag_t *fragfrom = skb_shinfo(skb)->frags; | |
2368 | ||
2369 | if (skb_headlen(skb)) | |
2370 | return -EINVAL; | |
2371 | ||
2372 | for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, fragfrom++) { | |
2373 | if (len >= probe_size) | |
2374 | goto commit; | |
2375 | todo = min_t(int, skb_frag_size(fragfrom), | |
2376 | probe_size - len); | |
2377 | len += todo; | |
2378 | if (lastfrag && | |
2379 | skb_frag_page(fragfrom) == skb_frag_page(lastfrag) && | |
2380 | skb_frag_off(fragfrom) == skb_frag_off(lastfrag) + | |
2381 | skb_frag_size(lastfrag)) { | |
2382 | skb_frag_size_add(lastfrag, todo); | |
2383 | continue; | |
2384 | } | |
2385 | if (unlikely(nr_frags == MAX_SKB_FRAGS)) | |
2386 | return -E2BIG; | |
2387 | skb_frag_page_copy(fragto, fragfrom); | |
2388 | skb_frag_off_copy(fragto, fragfrom); | |
2389 | skb_frag_size_set(fragto, todo); | |
2390 | nr_frags++; | |
2391 | lastfrag = fragto++; | |
2392 | } | |
2393 | } | |
2394 | commit: | |
2395 | WARN_ON_ONCE(len != probe_size); | |
2396 | for (i = 0; i < nr_frags; i++) | |
2397 | skb_frag_ref(to, i); | |
2398 | ||
2399 | skb_shinfo(to)->nr_frags = nr_frags; | |
2400 | to->truesize += probe_size; | |
2401 | to->len += probe_size; | |
2402 | to->data_len += probe_size; | |
2403 | __skb_header_release(to); | |
2404 | return 0; | |
2405 | } | |
2406 | ||
1bede0a1 ED |
2407 | /* tcp_mtu_probe() and tcp_grow_skb() can both eat an skb (src) if |
2408 | * all its payload was moved to another one (dst). | |
2409 | * Make sure to transfer tcp_flags, eor, and tstamp. | |
2410 | */ | |
2411 | static void tcp_eat_one_skb(struct sock *sk, | |
2412 | struct sk_buff *dst, | |
2413 | struct sk_buff *src) | |
2414 | { | |
2415 | TCP_SKB_CB(dst)->tcp_flags |= TCP_SKB_CB(src)->tcp_flags; | |
2416 | TCP_SKB_CB(dst)->eor = TCP_SKB_CB(src)->eor; | |
2417 | tcp_skb_collapse_tstamp(dst, src); | |
2418 | tcp_unlink_write_queue(src, sk); | |
2419 | tcp_wmem_free_skb(sk, src); | |
2420 | } | |
2421 | ||
5d424d5a | 2422 | /* Create a new MTU probe if we are ready. |
67edfef7 AK |
2423 | * MTU probe is regularly attempting to increase the path MTU by |
2424 | * deliberately sending larger packets. This discovers routing | |
2425 | * changes resulting in larger path MTUs. | |
2426 | * | |
5d424d5a JH |
2427 | * Returns 0 if we should wait to probe (no cwnd available), |
2428 | * 1 if a probe was sent, | |
056834d9 IJ |
2429 | * -1 otherwise |
2430 | */ | |
5d424d5a JH |
2431 | static int tcp_mtu_probe(struct sock *sk) |
2432 | { | |
5d424d5a | 2433 | struct inet_connection_sock *icsk = inet_csk(sk); |
12a59abc | 2434 | struct tcp_sock *tp = tcp_sk(sk); |
5d424d5a | 2435 | struct sk_buff *skb, *nskb, *next; |
6b58e0a5 | 2436 | struct net *net = sock_net(sk); |
5d424d5a | 2437 | int probe_size; |
91cc17c0 | 2438 | int size_needed; |
12a59abc | 2439 | int copy, len; |
5d424d5a | 2440 | int mss_now; |
6b58e0a5 | 2441 | int interval; |
5d424d5a JH |
2442 | |
2443 | /* Not currently probing/verifying, | |
2444 | * not in recovery, | |
2445 | * have enough cwnd, and | |
12a59abc ED |
2446 | * not SACKing (the variable headers throw things off) |
2447 | */ | |
2448 | if (likely(!icsk->icsk_mtup.enabled || | |
2449 | icsk->icsk_mtup.probe_size || | |
2450 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || | |
40570375 | 2451 | tcp_snd_cwnd(tp) < 11 || |
12a59abc | 2452 | tp->rx_opt.num_sacks || tp->rx_opt.dsack)) |
5d424d5a JH |
2453 | return -1; |
2454 | ||
6b58e0a5 FD |
2455 | /* Use binary search for probe_size between tcp_mss_base, |
2456 | * and current mss_clamp. if (search_high - search_low) | |
2457 | * smaller than a threshold, backoff from probing. | |
2458 | */ | |
0c54b85f | 2459 | mss_now = tcp_current_mss(sk); |
6b58e0a5 FD |
2460 | probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + |
2461 | icsk->icsk_mtup.search_low) >> 1); | |
91cc17c0 | 2462 | size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; |
6b58e0a5 | 2463 | interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; |
05cbc0db FD |
2464 | /* When misfortune happens, we are reprobing actively, |
2465 | * and then reprobe timer has expired. We stick with current | |
2466 | * probing process by not resetting search range to its orignal. | |
2467 | */ | |
6b58e0a5 | 2468 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || |
92c0aa41 | 2469 | interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) { |
05cbc0db FD |
2470 | /* Check whether enough time has elaplased for |
2471 | * another round of probing. | |
2472 | */ | |
2473 | tcp_mtu_check_reprobe(sk); | |
5d424d5a JH |
2474 | return -1; |
2475 | } | |
2476 | ||
2477 | /* Have enough data in the send queue to probe? */ | |
7f9c33e5 | 2478 | if (tp->write_seq - tp->snd_nxt < size_needed) |
5d424d5a JH |
2479 | return -1; |
2480 | ||
91cc17c0 IJ |
2481 | if (tp->snd_wnd < size_needed) |
2482 | return -1; | |
90840def | 2483 | if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) |
91cc17c0 | 2484 | return 0; |
5d424d5a | 2485 | |
d67c58e9 | 2486 | /* Do we need to wait to drain cwnd? With none in flight, don't stall */ |
40570375 | 2487 | if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) { |
d67c58e9 | 2488 | if (!tcp_packets_in_flight(tp)) |
5d424d5a JH |
2489 | return -1; |
2490 | else | |
2491 | return 0; | |
2492 | } | |
2493 | ||
808cf9e3 IL |
2494 | if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) |
2495 | return -1; | |
2496 | ||
5d424d5a | 2497 | /* We're allowed to probe. Build it now. */ |
5882efff | 2498 | nskb = tcp_stream_alloc_skb(sk, GFP_ATOMIC, false); |
51456b29 | 2499 | if (!nskb) |
5d424d5a | 2500 | return -1; |
73601329 ED |
2501 | |
2502 | /* build the payload, and be prepared to abort if this fails. */ | |
2503 | if (tcp_clone_payload(sk, nskb, probe_size)) { | |
71c299c7 | 2504 | tcp_skb_tsorted_anchor_cleanup(nskb); |
73601329 ED |
2505 | consume_skb(nskb); |
2506 | return -1; | |
2507 | } | |
ab4e846a | 2508 | sk_wmem_queued_add(sk, nskb->truesize); |
3ab224be | 2509 | sk_mem_charge(sk, nskb->truesize); |
5d424d5a | 2510 | |
fe067e8a | 2511 | skb = tcp_send_head(sk); |
41477662 | 2512 | skb_copy_decrypted(nskb, skb); |
5a369ca6 | 2513 | mptcp_skb_ext_copy(nskb, skb); |
5d424d5a JH |
2514 | |
2515 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; | |
2516 | TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; | |
4de075e0 | 2517 | TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; |
5d424d5a | 2518 | |
50c4817e | 2519 | tcp_insert_write_queue_before(nskb, skb, sk); |
2b7cda9c | 2520 | tcp_highest_sack_replace(sk, skb, nskb); |
50c4817e | 2521 | |
5d424d5a | 2522 | len = 0; |
234b6860 | 2523 | tcp_for_write_queue_from_safe(skb, next, sk) { |
5d424d5a | 2524 | copy = min_t(int, skb->len, probe_size - len); |
5d424d5a JH |
2525 | |
2526 | if (skb->len <= copy) { | |
1bede0a1 | 2527 | tcp_eat_one_skb(sk, nskb, skb); |
5d424d5a | 2528 | } else { |
4de075e0 | 2529 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & |
a3433f35 | 2530 | ~(TCPHDR_FIN|TCPHDR_PSH); |
b4a24397 ED |
2531 | __pskb_trim_head(skb, copy); |
2532 | tcp_set_skb_tso_segs(skb, mss_now); | |
5d424d5a JH |
2533 | TCP_SKB_CB(skb)->seq += copy; |
2534 | } | |
2535 | ||
2536 | len += copy; | |
234b6860 IJ |
2537 | |
2538 | if (len >= probe_size) | |
2539 | break; | |
5d424d5a | 2540 | } |
5bbb432c | 2541 | tcp_init_tso_segs(nskb, nskb->len); |
5d424d5a JH |
2542 | |
2543 | /* We're ready to send. If this fails, the probe will | |
7faee5c0 ED |
2544 | * be resegmented into mss-sized pieces by tcp_write_xmit(). |
2545 | */ | |
5d424d5a JH |
2546 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { |
2547 | /* Decrement cwnd here because we are sending | |
056834d9 | 2548 | * effectively two packets. */ |
40570375 | 2549 | tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1); |
66f5fe62 | 2550 | tcp_event_new_data_sent(sk, nskb); |
5d424d5a JH |
2551 | |
2552 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); | |
0e7b1368 JH |
2553 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; |
2554 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; | |
5d424d5a JH |
2555 | |
2556 | return 1; | |
2557 | } | |
2558 | ||
2559 | return -1; | |
2560 | } | |
2561 | ||
864e5c09 | 2562 | static bool tcp_pacing_check(struct sock *sk) |
218af599 | 2563 | { |
864e5c09 ED |
2564 | struct tcp_sock *tp = tcp_sk(sk); |
2565 | ||
2566 | if (!tcp_needs_internal_pacing(sk)) | |
2567 | return false; | |
2568 | ||
2569 | if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache) | |
2570 | return false; | |
2571 | ||
2572 | if (!hrtimer_is_queued(&tp->pacing_timer)) { | |
2573 | hrtimer_start(&tp->pacing_timer, | |
2574 | ns_to_ktime(tp->tcp_wstamp_ns), | |
2575 | HRTIMER_MODE_ABS_PINNED_SOFT); | |
2576 | sock_hold(sk); | |
2577 | } | |
2578 | return true; | |
218af599 ED |
2579 | } |
2580 | ||
f921a4a5 ED |
2581 | static bool tcp_rtx_queue_empty_or_single_skb(const struct sock *sk) |
2582 | { | |
2583 | const struct rb_node *node = sk->tcp_rtx_queue.rb_node; | |
2584 | ||
2585 | /* No skb in the rtx queue. */ | |
2586 | if (!node) | |
2587 | return true; | |
2588 | ||
2589 | /* Only one skb in rtx queue. */ | |
2590 | return !node->rb_left && !node->rb_right; | |
2591 | } | |
2592 | ||
f9616c35 ED |
2593 | /* TCP Small Queues : |
2594 | * Control number of packets in qdisc/devices to two packets / or ~1 ms. | |
2595 | * (These limits are doubled for retransmits) | |
2596 | * This allows for : | |
2597 | * - better RTT estimation and ACK scheduling | |
2598 | * - faster recovery | |
2599 | * - high rates | |
2600 | * Alas, some drivers / subsystems require a fair amount | |
2601 | * of queued bytes to ensure line rate. | |
2602 | * One example is wifi aggregation (802.11 AMPDU) | |
2603 | */ | |
2604 | static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, | |
2605 | unsigned int factor) | |
2606 | { | |
76a9ebe8 | 2607 | unsigned long limit; |
f9616c35 | 2608 | |
76a9ebe8 ED |
2609 | limit = max_t(unsigned long, |
2610 | 2 * skb->truesize, | |
28b24f90 | 2611 | READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift)); |
c73e5807 ED |
2612 | if (sk->sk_pacing_status == SK_PACING_NONE) |
2613 | limit = min_t(unsigned long, limit, | |
9fb90193 | 2614 | READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes)); |
f9616c35 ED |
2615 | limit <<= factor; |
2616 | ||
a842fe14 ED |
2617 | if (static_branch_unlikely(&tcp_tx_delay_enabled) && |
2618 | tcp_sk(sk)->tcp_tx_delay) { | |
28b24f90 ED |
2619 | u64 extra_bytes = (u64)READ_ONCE(sk->sk_pacing_rate) * |
2620 | tcp_sk(sk)->tcp_tx_delay; | |
a842fe14 ED |
2621 | |
2622 | /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we | |
2623 | * approximate our needs assuming an ~100% skb->truesize overhead. | |
2624 | * USEC_PER_SEC is approximated by 2^20. | |
2625 | * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift. | |
2626 | */ | |
2627 | extra_bytes >>= (20 - 1); | |
2628 | limit += extra_bytes; | |
2629 | } | |
14afee4b | 2630 | if (refcount_read(&sk->sk_wmem_alloc) > limit) { |
f921a4a5 | 2631 | /* Always send skb if rtx queue is empty or has one skb. |
75eefc6c ED |
2632 | * No need to wait for TX completion to call us back, |
2633 | * after softirq/tasklet schedule. | |
2634 | * This helps when TX completions are delayed too much. | |
2635 | */ | |
f921a4a5 | 2636 | if (tcp_rtx_queue_empty_or_single_skb(sk)) |
75eefc6c ED |
2637 | return false; |
2638 | ||
7aa5470c | 2639 | set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); |
f9616c35 ED |
2640 | /* It is possible TX completion already happened |
2641 | * before we set TSQ_THROTTLED, so we must | |
2642 | * test again the condition. | |
2643 | */ | |
2644 | smp_mb__after_atomic(); | |
ce8299b6 | 2645 | if (refcount_read(&sk->sk_wmem_alloc) > limit) |
f9616c35 ED |
2646 | return true; |
2647 | } | |
2648 | return false; | |
2649 | } | |
2650 | ||
05b055e8 FY |
2651 | static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) |
2652 | { | |
628174cc | 2653 | const u32 now = tcp_jiffies32; |
efe967cd | 2654 | enum tcp_chrono old = tp->chrono_type; |
05b055e8 | 2655 | |
efe967cd AB |
2656 | if (old > TCP_CHRONO_UNSPEC) |
2657 | tp->chrono_stat[old - 1] += now - tp->chrono_start; | |
05b055e8 FY |
2658 | tp->chrono_start = now; |
2659 | tp->chrono_type = new; | |
2660 | } | |
2661 | ||
2662 | void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) | |
2663 | { | |
2664 | struct tcp_sock *tp = tcp_sk(sk); | |
2665 | ||
2666 | /* If there are multiple conditions worthy of tracking in a | |
0f87230d FY |
2667 | * chronograph then the highest priority enum takes precedence |
2668 | * over the other conditions. So that if something "more interesting" | |
05b055e8 FY |
2669 | * starts happening, stop the previous chrono and start a new one. |
2670 | */ | |
2671 | if (type > tp->chrono_type) | |
2672 | tcp_chrono_set(tp, type); | |
2673 | } | |
2674 | ||
2675 | void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) | |
2676 | { | |
2677 | struct tcp_sock *tp = tcp_sk(sk); | |
2678 | ||
0f87230d FY |
2679 | |
2680 | /* There are multiple conditions worthy of tracking in a | |
2681 | * chronograph, so that the highest priority enum takes | |
2682 | * precedence over the other conditions (see tcp_chrono_start). | |
2683 | * If a condition stops, we only stop chrono tracking if | |
2684 | * it's the "most interesting" or current chrono we are | |
2685 | * tracking and starts busy chrono if we have pending data. | |
2686 | */ | |
75c119af | 2687 | if (tcp_rtx_and_write_queues_empty(sk)) |
0f87230d FY |
2688 | tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); |
2689 | else if (type == tp->chrono_type) | |
2690 | tcp_chrono_set(tp, TCP_CHRONO_BUSY); | |
05b055e8 FY |
2691 | } |
2692 | ||
8ee602c6 ED |
2693 | /* First skb in the write queue is smaller than ideal packet size. |
2694 | * Check if we can move payload from the second skb in the queue. | |
2695 | */ | |
2696 | static void tcp_grow_skb(struct sock *sk, struct sk_buff *skb, int amount) | |
2697 | { | |
2698 | struct sk_buff *next_skb = skb->next; | |
2699 | unsigned int nlen; | |
2700 | ||
2701 | if (tcp_skb_is_last(sk, skb)) | |
2702 | return; | |
2703 | ||
2704 | if (!tcp_skb_can_collapse(skb, next_skb)) | |
2705 | return; | |
2706 | ||
2707 | nlen = min_t(u32, amount, next_skb->len); | |
2708 | if (!nlen || !skb_shift(skb, next_skb, nlen)) | |
2709 | return; | |
2710 | ||
2711 | TCP_SKB_CB(skb)->end_seq += nlen; | |
2712 | TCP_SKB_CB(next_skb)->seq += nlen; | |
2713 | ||
2714 | if (!next_skb->len) { | |
1bede0a1 | 2715 | /* In case FIN is set, we need to update end_seq */ |
8ee602c6 | 2716 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; |
1bede0a1 ED |
2717 | |
2718 | tcp_eat_one_skb(sk, skb, next_skb); | |
8ee602c6 ED |
2719 | } |
2720 | } | |
2721 | ||
1da177e4 LT |
2722 | /* This routine writes packets to the network. It advances the |
2723 | * send_head. This happens as incoming acks open up the remote | |
2724 | * window for us. | |
2725 | * | |
f8269a49 IJ |
2726 | * LARGESEND note: !tcp_urg_mode is overkill, only frames between |
2727 | * snd_up-64k-mss .. snd_up cannot be large. However, taking into | |
2728 | * account rare use of URG, this is not a big flaw. | |
2729 | * | |
6ba8a3b1 ND |
2730 | * Send at most one packet when push_one > 0. Temporarily ignore |
2731 | * cwnd limit to force at most one packet out when push_one == 2. | |
2732 | ||
a2a385d6 ED |
2733 | * Returns true, if no segments are in flight and we have queued segments, |
2734 | * but cannot send anything now because of SWS or another problem. | |
1da177e4 | 2735 | */ |
a2a385d6 ED |
2736 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
2737 | int push_one, gfp_t gfp) | |
1da177e4 LT |
2738 | { |
2739 | struct tcp_sock *tp = tcp_sk(sk); | |
92df7b51 | 2740 | struct sk_buff *skb; |
c1b4a7e6 | 2741 | unsigned int tso_segs, sent_pkts; |
22555032 | 2742 | u32 cwnd_quota, max_segs; |
5d424d5a | 2743 | int result; |
5615f886 | 2744 | bool is_cwnd_limited = false, is_rwnd_limited = false; |
1da177e4 | 2745 | |
92df7b51 | 2746 | sent_pkts = 0; |
5d424d5a | 2747 | |
ee1836ae | 2748 | tcp_mstamp_refresh(tp); |
d5dd9175 IJ |
2749 | if (!push_one) { |
2750 | /* Do MTU probing. */ | |
2751 | result = tcp_mtu_probe(sk); | |
2752 | if (!result) { | |
a2a385d6 | 2753 | return false; |
d5dd9175 IJ |
2754 | } else if (result > 0) { |
2755 | sent_pkts = 1; | |
2756 | } | |
5d424d5a JH |
2757 | } |
2758 | ||
ed6e7268 | 2759 | max_segs = tcp_tso_segs(sk, mss_now); |
fe067e8a | 2760 | while ((skb = tcp_send_head(sk))) { |
c8ac3774 | 2761 | unsigned int limit; |
8ee602c6 | 2762 | int missing_bytes; |
c8ac3774 | 2763 | |
79861919 ED |
2764 | if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { |
2765 | /* "skb_mstamp_ns" is used as a start point for the retransmit timer */ | |
a1ac9c8a MKL |
2766 | tp->tcp_wstamp_ns = tp->tcp_clock_cache; |
2767 | skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true); | |
79861919 | 2768 | list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); |
bf50b606 | 2769 | tcp_init_tso_segs(skb, mss_now); |
79861919 ED |
2770 | goto repair; /* Skip network transmission */ |
2771 | } | |
2772 | ||
218af599 ED |
2773 | if (tcp_pacing_check(sk)) |
2774 | break; | |
2775 | ||
22555032 | 2776 | cwnd_quota = tcp_cwnd_test(tp); |
6ba8a3b1 ND |
2777 | if (!cwnd_quota) { |
2778 | if (push_one == 2) | |
2779 | /* Force out a loss probe pkt. */ | |
2780 | cwnd_quota = 1; | |
2781 | else | |
2782 | break; | |
2783 | } | |
8ee602c6 ED |
2784 | cwnd_quota = min(cwnd_quota, max_segs); |
2785 | missing_bytes = cwnd_quota * mss_now - skb->len; | |
2786 | if (missing_bytes > 0) | |
2787 | tcp_grow_skb(sk, skb, missing_bytes); | |
b68e9f85 | 2788 | |
d5b38a71 ED |
2789 | tso_segs = tcp_set_skb_tso_segs(skb, mss_now); |
2790 | ||
5615f886 FY |
2791 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { |
2792 | is_rwnd_limited = true; | |
b68e9f85 | 2793 | break; |
5615f886 | 2794 | } |
b68e9f85 | 2795 | |
d6a4e26a | 2796 | if (tso_segs == 1) { |
c1b4a7e6 DM |
2797 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, |
2798 | (tcp_skb_is_last(sk, skb) ? | |
2799 | nonagle : TCP_NAGLE_PUSH)))) | |
2800 | break; | |
2801 | } else { | |
ca8a2263 | 2802 | if (!push_one && |
605ad7f1 | 2803 | tcp_tso_should_defer(sk, skb, &is_cwnd_limited, |
f9bfe4e6 | 2804 | &is_rwnd_limited, max_segs)) |
c1b4a7e6 DM |
2805 | break; |
2806 | } | |
aa93466b | 2807 | |
605ad7f1 | 2808 | limit = mss_now; |
d6a4e26a | 2809 | if (tso_segs > 1 && !tcp_urg_mode(tp)) |
605ad7f1 | 2810 | limit = tcp_mss_split_point(sk, skb, mss_now, |
8ee602c6 | 2811 | cwnd_quota, |
605ad7f1 ED |
2812 | nonagle); |
2813 | ||
2814 | if (skb->len > limit && | |
56483341 | 2815 | unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) |
605ad7f1 ED |
2816 | break; |
2817 | ||
f9616c35 ED |
2818 | if (tcp_small_queue_check(sk, skb, 0)) |
2819 | break; | |
c9eeec26 | 2820 | |
1f85e626 ED |
2821 | /* Argh, we hit an empty skb(), presumably a thread |
2822 | * is sleeping in sendmsg()/sk_stream_wait_memory(). | |
2823 | * We do not want to send a pure-ack packet and have | |
2824 | * a strange looking rtx queue with empty packet(s). | |
2825 | */ | |
2826 | if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) | |
2827 | break; | |
2828 | ||
d5dd9175 | 2829 | if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) |
92df7b51 | 2830 | break; |
1da177e4 | 2831 | |
ec342325 | 2832 | repair: |
92df7b51 DM |
2833 | /* Advance the send_head. This one is sent out. |
2834 | * This call will increment packets_out. | |
2835 | */ | |
66f5fe62 | 2836 | tcp_event_new_data_sent(sk, skb); |
1da177e4 | 2837 | |
92df7b51 | 2838 | tcp_minshall_update(tp, mss_now, skb); |
a262f0cd | 2839 | sent_pkts += tcp_skb_pcount(skb); |
d5dd9175 IJ |
2840 | |
2841 | if (push_one) | |
2842 | break; | |
92df7b51 | 2843 | } |
1da177e4 | 2844 | |
5615f886 FY |
2845 | if (is_rwnd_limited) |
2846 | tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); | |
2847 | else | |
2848 | tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); | |
2849 | ||
40570375 | 2850 | is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp)); |
299bcb55 NC |
2851 | if (likely(sent_pkts || is_cwnd_limited)) |
2852 | tcp_cwnd_validate(sk, is_cwnd_limited); | |
2853 | ||
aa93466b | 2854 | if (likely(sent_pkts)) { |
684bad11 YC |
2855 | if (tcp_in_cwnd_reduction(sk)) |
2856 | tp->prr_out += sent_pkts; | |
6ba8a3b1 ND |
2857 | |
2858 | /* Send one loss probe per tail loss episode. */ | |
2859 | if (push_one != 2) | |
ed66dfaf | 2860 | tcp_schedule_loss_probe(sk, false); |
a2a385d6 | 2861 | return false; |
1da177e4 | 2862 | } |
75c119af | 2863 | return !tp->packets_out && !tcp_write_queue_empty(sk); |
6ba8a3b1 ND |
2864 | } |
2865 | ||
ed66dfaf | 2866 | bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) |
6ba8a3b1 ND |
2867 | { |
2868 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2869 | struct tcp_sock *tp = tcp_sk(sk); | |
1c2709cf | 2870 | u32 timeout, timeout_us, rto_delta_us; |
2ae21cf5 | 2871 | int early_retrans; |
6ba8a3b1 | 2872 | |
6ba8a3b1 ND |
2873 | /* Don't do any loss probe on a Fast Open connection before 3WHS |
2874 | * finishes. | |
2875 | */ | |
d983ea6f | 2876 | if (rcu_access_pointer(tp->fastopen_rsk)) |
6ba8a3b1 ND |
2877 | return false; |
2878 | ||
52e65865 | 2879 | early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans); |
6ba8a3b1 | 2880 | /* Schedule a loss probe in 2*RTT for SACK capable connections |
b4f70c3d | 2881 | * not in loss recovery, that are either limited by cwnd or application. |
6ba8a3b1 | 2882 | */ |
2ae21cf5 | 2883 | if ((early_retrans != 3 && early_retrans != 4) || |
bec41a11 | 2884 | !tp->packets_out || !tcp_is_sack(tp) || |
b4f70c3d NC |
2885 | (icsk->icsk_ca_state != TCP_CA_Open && |
2886 | icsk->icsk_ca_state != TCP_CA_CWR)) | |
6ba8a3b1 ND |
2887 | return false; |
2888 | ||
bb4d991a | 2889 | /* Probe timeout is 2*rtt. Add minimum RTO to account |
f9b99582 YC |
2890 | * for delayed ack when there's one outstanding packet. If no RTT |
2891 | * sample is available then probe after TCP_TIMEOUT_INIT. | |
6ba8a3b1 | 2892 | */ |
bb4d991a | 2893 | if (tp->srtt_us) { |
1c2709cf | 2894 | timeout_us = tp->srtt_us >> 2; |
bb4d991a | 2895 | if (tp->packets_out == 1) |
1c2709cf | 2896 | timeout_us += tcp_rto_min_us(sk); |
bb4d991a | 2897 | else |
1c2709cf NC |
2898 | timeout_us += TCP_TIMEOUT_MIN_US; |
2899 | timeout = usecs_to_jiffies(timeout_us); | |
bb4d991a YC |
2900 | } else { |
2901 | timeout = TCP_TIMEOUT_INIT; | |
2902 | } | |
6ba8a3b1 | 2903 | |
a2815817 | 2904 | /* If the RTO formula yields an earlier time, then use that time. */ |
ed66dfaf NC |
2905 | rto_delta_us = advancing_rto ? |
2906 | jiffies_to_usecs(inet_csk(sk)->icsk_rto) : | |
2907 | tcp_rto_delta_us(sk); /* How far in future is RTO? */ | |
a2815817 NC |
2908 | if (rto_delta_us > 0) |
2909 | timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); | |
6ba8a3b1 | 2910 | |
8dc242ad | 2911 | tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX); |
6ba8a3b1 ND |
2912 | return true; |
2913 | } | |
2914 | ||
1f3279ae ED |
2915 | /* Thanks to skb fast clones, we can detect if a prior transmit of |
2916 | * a packet is still in a qdisc or driver queue. | |
2917 | * In this case, there is very little point doing a retransmit ! | |
1f3279ae | 2918 | */ |
f4dae54e | 2919 | static bool skb_still_in_host_queue(struct sock *sk, |
1f3279ae ED |
2920 | const struct sk_buff *skb) |
2921 | { | |
39bb5e62 | 2922 | if (unlikely(skb_fclone_busy(sk, skb))) { |
f4dae54e ED |
2923 | set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); |
2924 | smp_mb__after_atomic(); | |
2925 | if (skb_fclone_busy(sk, skb)) { | |
2926 | NET_INC_STATS(sock_net(sk), | |
2927 | LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); | |
2928 | return true; | |
2929 | } | |
1f3279ae ED |
2930 | } |
2931 | return false; | |
2932 | } | |
2933 | ||
b340b264 | 2934 | /* When probe timeout (PTO) fires, try send a new segment if possible, else |
6ba8a3b1 ND |
2935 | * retransmit the last segment. |
2936 | */ | |
2937 | void tcp_send_loss_probe(struct sock *sk) | |
2938 | { | |
9b717a8d | 2939 | struct tcp_sock *tp = tcp_sk(sk); |
6ba8a3b1 ND |
2940 | struct sk_buff *skb; |
2941 | int pcount; | |
2942 | int mss = tcp_current_mss(sk); | |
6ba8a3b1 | 2943 | |
76be93fc YC |
2944 | /* At most one outstanding TLP */ |
2945 | if (tp->tlp_high_seq) | |
2946 | goto rearm_timer; | |
2947 | ||
2948 | tp->tlp_retrans = 0; | |
b340b264 | 2949 | skb = tcp_send_head(sk); |
75c119af ED |
2950 | if (skb && tcp_snd_wnd_test(tp, skb, mss)) { |
2951 | pcount = tp->packets_out; | |
2952 | tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); | |
2953 | if (tp->packets_out > pcount) | |
2954 | goto probe_sent; | |
2955 | goto rearm_timer; | |
6ba8a3b1 | 2956 | } |
75c119af | 2957 | skb = skb_rb_last(&sk->tcp_rtx_queue); |
b2b7af86 YC |
2958 | if (unlikely(!skb)) { |
2959 | WARN_ONCE(tp->packets_out, | |
2960 | "invalid inflight: %u state %u cwnd %u mss %d\n", | |
40570375 | 2961 | tp->packets_out, sk->sk_state, tcp_snd_cwnd(tp), mss); |
b2b7af86 YC |
2962 | inet_csk(sk)->icsk_pending = 0; |
2963 | return; | |
2964 | } | |
6ba8a3b1 | 2965 | |
1f3279ae ED |
2966 | if (skb_still_in_host_queue(sk, skb)) |
2967 | goto rearm_timer; | |
2968 | ||
6ba8a3b1 ND |
2969 | pcount = tcp_skb_pcount(skb); |
2970 | if (WARN_ON(!pcount)) | |
2971 | goto rearm_timer; | |
2972 | ||
2973 | if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { | |
75c119af ED |
2974 | if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, |
2975 | (pcount - 1) * mss, mss, | |
6cc55e09 | 2976 | GFP_ATOMIC))) |
6ba8a3b1 | 2977 | goto rearm_timer; |
75c119af | 2978 | skb = skb_rb_next(skb); |
6ba8a3b1 ND |
2979 | } |
2980 | ||
2981 | if (WARN_ON(!skb || !tcp_skb_pcount(skb))) | |
2982 | goto rearm_timer; | |
2983 | ||
10d3be56 | 2984 | if (__tcp_retransmit_skb(sk, skb, 1)) |
b340b264 | 2985 | goto rearm_timer; |
6ba8a3b1 | 2986 | |
76be93fc YC |
2987 | tp->tlp_retrans = 1; |
2988 | ||
2989 | probe_sent: | |
9b717a8d | 2990 | /* Record snd_nxt for loss detection. */ |
b340b264 | 2991 | tp->tlp_high_seq = tp->snd_nxt; |
9b717a8d | 2992 | |
c10d9310 | 2993 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); |
b340b264 YC |
2994 | /* Reset s.t. tcp_rearm_rto will restart timer from now */ |
2995 | inet_csk(sk)->icsk_pending = 0; | |
6ba8a3b1 | 2996 | rearm_timer: |
fcd16c0a | 2997 | tcp_rearm_rto(sk); |
1da177e4 LT |
2998 | } |
2999 | ||
a762a980 DM |
3000 | /* Push out any pending frames which were held back due to |
3001 | * TCP_CORK or attempt at coalescing tiny packets. | |
3002 | * The socket must be locked by the caller. | |
3003 | */ | |
9e412ba7 IJ |
3004 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, |
3005 | int nonagle) | |
a762a980 | 3006 | { |
726e07a8 IJ |
3007 | /* If we are closed, the bytes will have to remain here. |
3008 | * In time closedown will finish, we empty the write queue and | |
3009 | * all will be happy. | |
3010 | */ | |
3011 | if (unlikely(sk->sk_state == TCP_CLOSE)) | |
3012 | return; | |
3013 | ||
99a1dec7 | 3014 | if (tcp_write_xmit(sk, cur_mss, nonagle, 0, |
7450aaf6 | 3015 | sk_gfp_mask(sk, GFP_ATOMIC))) |
726e07a8 | 3016 | tcp_check_probe_timer(sk); |
a762a980 DM |
3017 | } |
3018 | ||
c1b4a7e6 DM |
3019 | /* Send _single_ skb sitting at the send head. This function requires |
3020 | * true push pending frames to setup probe timer etc. | |
3021 | */ | |
3022 | void tcp_push_one(struct sock *sk, unsigned int mss_now) | |
3023 | { | |
fe067e8a | 3024 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 DM |
3025 | |
3026 | BUG_ON(!skb || skb->len < mss_now); | |
3027 | ||
d5dd9175 | 3028 | tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); |
c1b4a7e6 DM |
3029 | } |
3030 | ||
1da177e4 LT |
3031 | /* This function returns the amount that we can raise the |
3032 | * usable window based on the following constraints | |
e905a9ed | 3033 | * |
1da177e4 LT |
3034 | * 1. The window can never be shrunk once it is offered (RFC 793) |
3035 | * 2. We limit memory per socket | |
3036 | * | |
3037 | * RFC 1122: | |
3038 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep | |
3039 | * RECV.NEXT + RCV.WIN fixed until: | |
3040 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" | |
3041 | * | |
3042 | * i.e. don't raise the right edge of the window until you can raise | |
3043 | * it at least MSS bytes. | |
3044 | * | |
3045 | * Unfortunately, the recommended algorithm breaks header prediction, | |
3046 | * since header prediction assumes th->window stays fixed. | |
3047 | * | |
3048 | * Strictly speaking, keeping th->window fixed violates the receiver | |
3049 | * side SWS prevention criteria. The problem is that under this rule | |
3050 | * a stream of single byte packets will cause the right side of the | |
3051 | * window to always advance by a single byte. | |
e905a9ed | 3052 | * |
1da177e4 LT |
3053 | * Of course, if the sender implements sender side SWS prevention |
3054 | * then this will not be a problem. | |
e905a9ed | 3055 | * |
1da177e4 | 3056 | * BSD seems to make the following compromise: |
e905a9ed | 3057 | * |
1da177e4 LT |
3058 | * If the free space is less than the 1/4 of the maximum |
3059 | * space available and the free space is less than 1/2 mss, | |
3060 | * then set the window to 0. | |
3061 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] | |
3062 | * Otherwise, just prevent the window from shrinking | |
3063 | * and from being larger than the largest representable value. | |
3064 | * | |
3065 | * This prevents incremental opening of the window in the regime | |
3066 | * where TCP is limited by the speed of the reader side taking | |
3067 | * data out of the TCP receive queue. It does nothing about | |
3068 | * those cases where the window is constrained on the sender side | |
3069 | * because the pipeline is full. | |
3070 | * | |
3071 | * BSD also seems to "accidentally" limit itself to windows that are a | |
3072 | * multiple of MSS, at least until the free space gets quite small. | |
3073 | * This would appear to be a side effect of the mbuf implementation. | |
3074 | * Combining these two algorithms results in the observed behavior | |
3075 | * of having a fixed window size at almost all times. | |
3076 | * | |
3077 | * Below we obtain similar behavior by forcing the offered window to | |
3078 | * a multiple of the mss when it is feasible to do so. | |
3079 | * | |
3080 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. | |
3081 | * Regular options like TIMESTAMP are taken into account. | |
3082 | */ | |
3083 | u32 __tcp_select_window(struct sock *sk) | |
3084 | { | |
463c84b9 | 3085 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 3086 | struct tcp_sock *tp = tcp_sk(sk); |
b650d953 | 3087 | struct net *net = sock_net(sk); |
caa20d9a | 3088 | /* MSS for the peer's data. Previous versions used mss_clamp |
1da177e4 LT |
3089 | * here. I don't know if the value based on our guesses |
3090 | * of peer's MSS is better for the performance. It's more correct | |
3091 | * but may be worse for the performance because of rcv_mss | |
3092 | * fluctuations. --SAW 1998/11/1 | |
3093 | */ | |
463c84b9 | 3094 | int mss = icsk->icsk_ack.rcv_mss; |
1da177e4 | 3095 | int free_space = tcp_space(sk); |
86c1a045 | 3096 | int allowed_space = tcp_full_space(sk); |
071c8ed6 FW |
3097 | int full_space, window; |
3098 | ||
3099 | if (sk_is_mptcp(sk)) | |
3100 | mptcp_space(sk, &free_space, &allowed_space); | |
3101 | ||
3102 | full_space = min_t(int, tp->window_clamp, allowed_space); | |
1da177e4 | 3103 | |
06425c30 | 3104 | if (unlikely(mss > full_space)) { |
e905a9ed | 3105 | mss = full_space; |
06425c30 ED |
3106 | if (mss <= 0) |
3107 | return 0; | |
3108 | } | |
b650d953 | 3109 | |
3110 | /* Only allow window shrink if the sysctl is enabled and we have | |
3111 | * a non-zero scaling factor in effect. | |
3112 | */ | |
3113 | if (READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) && tp->rx_opt.rcv_wscale) | |
3114 | goto shrink_window_allowed; | |
3115 | ||
3116 | /* do not allow window to shrink */ | |
3117 | ||
b92edbe0 | 3118 | if (free_space < (full_space >> 1)) { |
463c84b9 | 3119 | icsk->icsk_ack.quick = 0; |
1da177e4 | 3120 | |
b8da51eb | 3121 | if (tcp_under_memory_pressure(sk)) |
053f3684 | 3122 | tcp_adjust_rcv_ssthresh(sk); |
1da177e4 | 3123 | |
86c1a045 FW |
3124 | /* free_space might become our new window, make sure we don't |
3125 | * increase it due to wscale. | |
3126 | */ | |
3127 | free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); | |
3128 | ||
3129 | /* if free space is less than mss estimate, or is below 1/16th | |
3130 | * of the maximum allowed, try to move to zero-window, else | |
3131 | * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and | |
3132 | * new incoming data is dropped due to memory limits. | |
3133 | * With large window, mss test triggers way too late in order | |
3134 | * to announce zero window in time before rmem limit kicks in. | |
3135 | */ | |
3136 | if (free_space < (allowed_space >> 4) || free_space < mss) | |
1da177e4 LT |
3137 | return 0; |
3138 | } | |
3139 | ||
3140 | if (free_space > tp->rcv_ssthresh) | |
3141 | free_space = tp->rcv_ssthresh; | |
3142 | ||
3143 | /* Don't do rounding if we are using window scaling, since the | |
3144 | * scaled window will not line up with the MSS boundary anyway. | |
3145 | */ | |
1da177e4 LT |
3146 | if (tp->rx_opt.rcv_wscale) { |
3147 | window = free_space; | |
3148 | ||
3149 | /* Advertise enough space so that it won't get scaled away. | |
3150 | * Import case: prevent zero window announcement if | |
3151 | * 1<<rcv_wscale > mss. | |
3152 | */ | |
1935299d | 3153 | window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); |
1da177e4 | 3154 | } else { |
1935299d | 3155 | window = tp->rcv_wnd; |
1da177e4 LT |
3156 | /* Get the largest window that is a nice multiple of mss. |
3157 | * Window clamp already applied above. | |
3158 | * If our current window offering is within 1 mss of the | |
3159 | * free space we just keep it. This prevents the divide | |
3160 | * and multiply from happening most of the time. | |
3161 | * We also don't do any window rounding when the free space | |
3162 | * is too small. | |
3163 | */ | |
3164 | if (window <= free_space - mss || window > free_space) | |
1935299d | 3165 | window = rounddown(free_space, mss); |
84565070 | 3166 | else if (mss == full_space && |
b92edbe0 | 3167 | free_space > window + (full_space >> 1)) |
84565070 | 3168 | window = free_space; |
1da177e4 LT |
3169 | } |
3170 | ||
3171 | return window; | |
b650d953 | 3172 | |
3173 | shrink_window_allowed: | |
3174 | /* new window should always be an exact multiple of scaling factor */ | |
3175 | free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); | |
3176 | ||
3177 | if (free_space < (full_space >> 1)) { | |
3178 | icsk->icsk_ack.quick = 0; | |
3179 | ||
3180 | if (tcp_under_memory_pressure(sk)) | |
3181 | tcp_adjust_rcv_ssthresh(sk); | |
3182 | ||
3183 | /* if free space is too low, return a zero window */ | |
3184 | if (free_space < (allowed_space >> 4) || free_space < mss || | |
3185 | free_space < (1 << tp->rx_opt.rcv_wscale)) | |
3186 | return 0; | |
3187 | } | |
3188 | ||
3189 | if (free_space > tp->rcv_ssthresh) { | |
3190 | free_space = tp->rcv_ssthresh; | |
3191 | /* new window should always be an exact multiple of scaling factor | |
3192 | * | |
3193 | * For this case, we ALIGN "up" (increase free_space) because | |
3194 | * we know free_space is not zero here, it has been reduced from | |
3195 | * the memory-based limit, and rcv_ssthresh is not a hard limit | |
3196 | * (unlike sk_rcvbuf). | |
3197 | */ | |
3198 | free_space = ALIGN(free_space, (1 << tp->rx_opt.rcv_wscale)); | |
3199 | } | |
3200 | ||
3201 | return free_space; | |
1da177e4 LT |
3202 | } |
3203 | ||
cfea5a68 MKL |
3204 | void tcp_skb_collapse_tstamp(struct sk_buff *skb, |
3205 | const struct sk_buff *next_skb) | |
082ac2d5 | 3206 | { |
0a2cf20c SHY |
3207 | if (unlikely(tcp_has_tx_tstamp(next_skb))) { |
3208 | const struct skb_shared_info *next_shinfo = | |
3209 | skb_shinfo(next_skb); | |
082ac2d5 MKL |
3210 | struct skb_shared_info *shinfo = skb_shinfo(skb); |
3211 | ||
0a2cf20c | 3212 | shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; |
082ac2d5 | 3213 | shinfo->tskey = next_shinfo->tskey; |
2de8023e MKL |
3214 | TCP_SKB_CB(skb)->txstamp_ack |= |
3215 | TCP_SKB_CB(next_skb)->txstamp_ack; | |
082ac2d5 MKL |
3216 | } |
3217 | } | |
3218 | ||
4a17fc3a | 3219 | /* Collapses two adjacent SKB's during retransmission. */ |
f8071cde | 3220 | static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) |
1da177e4 LT |
3221 | { |
3222 | struct tcp_sock *tp = tcp_sk(sk); | |
75c119af | 3223 | struct sk_buff *next_skb = skb_rb_next(skb); |
13dde04f | 3224 | int next_skb_size; |
1da177e4 | 3225 | |
058dc334 | 3226 | next_skb_size = next_skb->len; |
1da177e4 | 3227 | |
058dc334 | 3228 | BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); |
a6963a6b | 3229 | |
bd446314 ED |
3230 | if (next_skb_size && !tcp_skb_shift(skb, next_skb, 1, next_skb_size)) |
3231 | return false; | |
3232 | ||
2b7cda9c | 3233 | tcp_highest_sack_replace(sk, next_skb, skb); |
1da177e4 | 3234 | |
058dc334 IJ |
3235 | /* Update sequence range on original skb. */ |
3236 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; | |
1da177e4 | 3237 | |
e6c7d085 | 3238 | /* Merge over control information. This moves PSH/FIN etc. over */ |
4de075e0 | 3239 | TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; |
058dc334 IJ |
3240 | |
3241 | /* All done, get rid of second SKB and account for it so | |
3242 | * packet counting does not break. | |
3243 | */ | |
3244 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; | |
a643b5d4 | 3245 | TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; |
058dc334 IJ |
3246 | |
3247 | /* changed transmit queue under us so clear hints */ | |
ef9da47c IJ |
3248 | tcp_clear_retrans_hints_partial(tp); |
3249 | if (next_skb == tp->retransmit_skb_hint) | |
3250 | tp->retransmit_skb_hint = skb; | |
058dc334 | 3251 | |
797108d1 IJ |
3252 | tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); |
3253 | ||
082ac2d5 MKL |
3254 | tcp_skb_collapse_tstamp(skb, next_skb); |
3255 | ||
75c119af | 3256 | tcp_rtx_queue_unlink_and_free(next_skb, sk); |
f8071cde | 3257 | return true; |
1da177e4 LT |
3258 | } |
3259 | ||
67edfef7 | 3260 | /* Check if coalescing SKBs is legal. */ |
a2a385d6 | 3261 | static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) |
4a17fc3a IJ |
3262 | { |
3263 | if (tcp_skb_pcount(skb) > 1) | |
a2a385d6 | 3264 | return false; |
4a17fc3a | 3265 | if (skb_cloned(skb)) |
a2a385d6 | 3266 | return false; |
2331ccc5 | 3267 | /* Some heuristics for collapsing over SACK'd could be invented */ |
4a17fc3a | 3268 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) |
a2a385d6 | 3269 | return false; |
4a17fc3a | 3270 | |
a2a385d6 | 3271 | return true; |
4a17fc3a IJ |
3272 | } |
3273 | ||
67edfef7 AK |
3274 | /* Collapse packets in the retransmit queue to make to create |
3275 | * less packets on the wire. This is only done on retransmission. | |
3276 | */ | |
4a17fc3a IJ |
3277 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, |
3278 | int space) | |
3279 | { | |
3280 | struct tcp_sock *tp = tcp_sk(sk); | |
3281 | struct sk_buff *skb = to, *tmp; | |
a2a385d6 | 3282 | bool first = true; |
4a17fc3a | 3283 | |
1a63cb91 | 3284 | if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)) |
4a17fc3a | 3285 | return; |
4de075e0 | 3286 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) |
4a17fc3a IJ |
3287 | return; |
3288 | ||
75c119af | 3289 | skb_rbtree_walk_from_safe(skb, tmp) { |
4a17fc3a IJ |
3290 | if (!tcp_can_collapse(sk, skb)) |
3291 | break; | |
3292 | ||
85712484 | 3293 | if (!tcp_skb_can_collapse(to, skb)) |
a643b5d4 MKL |
3294 | break; |
3295 | ||
4a17fc3a IJ |
3296 | space -= skb->len; |
3297 | ||
3298 | if (first) { | |
a2a385d6 | 3299 | first = false; |
4a17fc3a IJ |
3300 | continue; |
3301 | } | |
3302 | ||
3303 | if (space < 0) | |
3304 | break; | |
4a17fc3a IJ |
3305 | |
3306 | if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) | |
3307 | break; | |
3308 | ||
f8071cde ED |
3309 | if (!tcp_collapse_retrans(sk, to)) |
3310 | break; | |
4a17fc3a IJ |
3311 | } |
3312 | } | |
3313 | ||
1da177e4 LT |
3314 | /* This retransmits one SKB. Policy decisions and retransmit queue |
3315 | * state updates are done by the caller. Returns non-zero if an | |
3316 | * error occurred which prevented the send. | |
3317 | */ | |
10d3be56 | 3318 | int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) |
1da177e4 | 3319 | { |
5d424d5a | 3320 | struct inet_connection_sock *icsk = inet_csk(sk); |
10d3be56 | 3321 | struct tcp_sock *tp = tcp_sk(sk); |
7d227cd2 | 3322 | unsigned int cur_mss; |
10d3be56 | 3323 | int diff, len, err; |
536a6c8e | 3324 | int avail_wnd; |
1da177e4 | 3325 | |
10d3be56 ED |
3326 | /* Inconclusive MTU probe */ |
3327 | if (icsk->icsk_mtup.probe_size) | |
5d424d5a | 3328 | icsk->icsk_mtup.probe_size = 0; |
1da177e4 | 3329 | |
1f3279ae ED |
3330 | if (skb_still_in_host_queue(sk, skb)) |
3331 | return -EBUSY; | |
3332 | ||
f99cd562 | 3333 | start: |
1da177e4 | 3334 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { |
f99cd562 DC |
3335 | if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { |
3336 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN; | |
3337 | TCP_SKB_CB(skb)->seq++; | |
3338 | goto start; | |
3339 | } | |
7f582b24 ED |
3340 | if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) { |
3341 | WARN_ON_ONCE(1); | |
3342 | return -EINVAL; | |
3343 | } | |
1da177e4 LT |
3344 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
3345 | return -ENOMEM; | |
3346 | } | |
3347 | ||
7d227cd2 SS |
3348 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) |
3349 | return -EHOSTUNREACH; /* Routing failure or similar. */ | |
3350 | ||
0c54b85f | 3351 | cur_mss = tcp_current_mss(sk); |
536a6c8e | 3352 | avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
7d227cd2 | 3353 | |
1da177e4 LT |
3354 | /* If receiver has shrunk his window, and skb is out of |
3355 | * new window, do not retransmit it. The exception is the | |
3356 | * case, when window is shrunk to zero. In this case | |
536a6c8e | 3357 | * our retransmit of one segment serves as a zero window probe. |
1da177e4 | 3358 | */ |
536a6c8e YL |
3359 | if (avail_wnd <= 0) { |
3360 | if (TCP_SKB_CB(skb)->seq != tp->snd_una) | |
3361 | return -EAGAIN; | |
3362 | avail_wnd = cur_mss; | |
3363 | } | |
1da177e4 | 3364 | |
10d3be56 | 3365 | len = cur_mss * segs; |
536a6c8e YL |
3366 | if (len > avail_wnd) { |
3367 | len = rounddown(avail_wnd, cur_mss); | |
3368 | if (!len) | |
3369 | len = avail_wnd; | |
3370 | } | |
10d3be56 | 3371 | if (skb->len > len) { |
75c119af ED |
3372 | if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, |
3373 | cur_mss, GFP_ATOMIC)) | |
1da177e4 | 3374 | return -ENOMEM; /* We'll try again later. */ |
02276f3c | 3375 | } else { |
c4777efa | 3376 | if (skb_unclone_keeptruesize(skb, GFP_ATOMIC)) |
10d3be56 | 3377 | return -ENOMEM; |
9eb9362e | 3378 | |
10d3be56 ED |
3379 | diff = tcp_skb_pcount(skb); |
3380 | tcp_set_skb_tso_segs(skb, cur_mss); | |
3381 | diff -= tcp_skb_pcount(skb); | |
3382 | if (diff) | |
3383 | tcp_adjust_pcount(sk, skb, diff); | |
536a6c8e YL |
3384 | avail_wnd = min_t(int, avail_wnd, cur_mss); |
3385 | if (skb->len < avail_wnd) | |
3386 | tcp_retrans_try_collapse(sk, skb, avail_wnd); | |
1da177e4 LT |
3387 | } |
3388 | ||
49213555 DB |
3389 | /* RFC3168, section 6.1.1.1. ECN fallback */ |
3390 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) | |
3391 | tcp_ecn_clear_syn(sk, skb); | |
3392 | ||
678550c6 YC |
3393 | /* Update global and local TCP statistics. */ |
3394 | segs = tcp_skb_pcount(skb); | |
3395 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); | |
3396 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) | |
3397 | __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); | |
3398 | tp->total_retrans += segs; | |
fb31c9b9 | 3399 | tp->bytes_retrans += skb->len; |
678550c6 | 3400 | |
50bceae9 TG |
3401 | /* make sure skb->data is aligned on arches that require it |
3402 | * and check if ack-trimming & collapsing extended the headroom | |
3403 | * beyond what csum_start can cover. | |
3404 | */ | |
3405 | if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || | |
3406 | skb_headroom(skb) >= 0xFFFF)) { | |
10a81980 ED |
3407 | struct sk_buff *nskb; |
3408 | ||
e2080072 ED |
3409 | tcp_skb_tsorted_save(skb) { |
3410 | nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); | |
07f8e4d0 FW |
3411 | if (nskb) { |
3412 | nskb->dev = NULL; | |
3413 | err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC); | |
3414 | } else { | |
3415 | err = -ENOBUFS; | |
3416 | } | |
e2080072 ED |
3417 | } tcp_skb_tsorted_restore(skb); |
3418 | ||
5889e2c0 | 3419 | if (!err) { |
a7a25630 | 3420 | tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns); |
5889e2c0 YS |
3421 | tcp_rate_skb_sent(sk, skb); |
3422 | } | |
117632e6 | 3423 | } else { |
c84a5711 | 3424 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
117632e6 | 3425 | } |
c84a5711 | 3426 | |
a31ad29e LB |
3427 | if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) |
3428 | tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, | |
3429 | TCP_SKB_CB(skb)->seq, segs, err); | |
3430 | ||
fc9f3501 | 3431 | if (likely(!err)) { |
e086101b | 3432 | trace_tcp_retransmit_skb(sk, skb); |
678550c6 | 3433 | } else if (err != -EBUSY) { |
ec641b39 | 3434 | NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs); |
fc9f3501 | 3435 | } |
2bf90a57 PL |
3436 | |
3437 | /* To avoid taking spuriously low RTT samples based on a timestamp | |
3438 | * for a transmit that never happened, always mark EVER_RETRANS | |
3439 | */ | |
3440 | TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; | |
3441 | ||
c84a5711 | 3442 | return err; |
93b174ad YC |
3443 | } |
3444 | ||
10d3be56 | 3445 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) |
93b174ad YC |
3446 | { |
3447 | struct tcp_sock *tp = tcp_sk(sk); | |
10d3be56 | 3448 | int err = __tcp_retransmit_skb(sk, skb, segs); |
1da177e4 LT |
3449 | |
3450 | if (err == 0) { | |
1da177e4 | 3451 | #if FASTRETRANS_DEBUG > 0 |
056834d9 | 3452 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
e87cc472 | 3453 | net_dbg_ratelimited("retrans_out leaked\n"); |
1da177e4 LT |
3454 | } |
3455 | #endif | |
3456 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; | |
3457 | tp->retrans_out += tcp_skb_pcount(skb); | |
1da177e4 | 3458 | } |
6e08d5e3 | 3459 | |
7ae18975 YC |
3460 | /* Save stamp of the first (attempted) retransmit. */ |
3461 | if (!tp->retrans_stamp) | |
614e8316 | 3462 | tp->retrans_stamp = tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb); |
7ae18975 | 3463 | |
6e08d5e3 YC |
3464 | if (tp->undo_retrans < 0) |
3465 | tp->undo_retrans = 0; | |
3466 | tp->undo_retrans += tcp_skb_pcount(skb); | |
1da177e4 LT |
3467 | return err; |
3468 | } | |
3469 | ||
3470 | /* This gets called after a retransmit timeout, and the initially | |
3471 | * retransmitted data is acknowledged. It tries to continue | |
3472 | * resending the rest of the retransmit queue, until either | |
3473 | * we've sent it all or the congestion window limit is reached. | |
1da177e4 LT |
3474 | */ |
3475 | void tcp_xmit_retransmit_queue(struct sock *sk) | |
3476 | { | |
6687e988 | 3477 | const struct inet_connection_sock *icsk = inet_csk(sk); |
b9f1f1ce | 3478 | struct sk_buff *skb, *rtx_head, *hole = NULL; |
1da177e4 | 3479 | struct tcp_sock *tp = tcp_sk(sk); |
916e6d1a | 3480 | bool rearm_timer = false; |
840a3cbe | 3481 | u32 max_segs; |
61eb55f4 | 3482 | int mib_idx; |
6a438bbe | 3483 | |
45e77d31 IJ |
3484 | if (!tp->packets_out) |
3485 | return; | |
3486 | ||
b9f1f1ce ED |
3487 | rtx_head = tcp_rtx_queue_head(sk); |
3488 | skb = tp->retransmit_skb_hint ?: rtx_head; | |
ed6e7268 | 3489 | max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); |
75c119af | 3490 | skb_rbtree_walk_from(skb) { |
dca0aaf8 | 3491 | __u8 sacked; |
10d3be56 | 3492 | int segs; |
1da177e4 | 3493 | |
218af599 ED |
3494 | if (tcp_pacing_check(sk)) |
3495 | break; | |
3496 | ||
08ebd172 | 3497 | /* we could do better than to assign each time */ |
51456b29 | 3498 | if (!hole) |
0e1c54c2 | 3499 | tp->retransmit_skb_hint = skb; |
08ebd172 | 3500 | |
40570375 | 3501 | segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp); |
10d3be56 | 3502 | if (segs <= 0) |
916e6d1a | 3503 | break; |
dca0aaf8 | 3504 | sacked = TCP_SKB_CB(skb)->sacked; |
a3d2e9f8 ED |
3505 | /* In case tcp_shift_skb_data() have aggregated large skbs, |
3506 | * we need to make sure not sending too bigs TSO packets | |
3507 | */ | |
3508 | segs = min_t(int, segs, max_segs); | |
1da177e4 | 3509 | |
840a3cbe YC |
3510 | if (tp->retrans_out >= tp->lost_out) { |
3511 | break; | |
0e1c54c2 | 3512 | } else if (!(sacked & TCPCB_LOST)) { |
51456b29 | 3513 | if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) |
0e1c54c2 IJ |
3514 | hole = skb; |
3515 | continue; | |
1da177e4 | 3516 | |
0e1c54c2 IJ |
3517 | } else { |
3518 | if (icsk->icsk_ca_state != TCP_CA_Loss) | |
3519 | mib_idx = LINUX_MIB_TCPFASTRETRANS; | |
3520 | else | |
3521 | mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; | |
3522 | } | |
1da177e4 | 3523 | |
0e1c54c2 | 3524 | if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) |
1da177e4 LT |
3525 | continue; |
3526 | ||
f9616c35 | 3527 | if (tcp_small_queue_check(sk, skb, 1)) |
916e6d1a | 3528 | break; |
f9616c35 | 3529 | |
10d3be56 | 3530 | if (tcp_retransmit_skb(sk, skb, segs)) |
916e6d1a | 3531 | break; |
24ab6bec | 3532 | |
de1d6578 | 3533 | NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); |
1da177e4 | 3534 | |
684bad11 | 3535 | if (tcp_in_cwnd_reduction(sk)) |
a262f0cd ND |
3536 | tp->prr_out += tcp_skb_pcount(skb); |
3537 | ||
75c119af | 3538 | if (skb == rtx_head && |
57dde7f7 | 3539 | icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) |
916e6d1a ED |
3540 | rearm_timer = true; |
3541 | ||
1da177e4 | 3542 | } |
916e6d1a ED |
3543 | if (rearm_timer) |
3544 | tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, | |
3545 | inet_csk(sk)->icsk_rto, | |
3546 | TCP_RTO_MAX); | |
1da177e4 LT |
3547 | } |
3548 | ||
d83769a5 ED |
3549 | /* We allow to exceed memory limits for FIN packets to expedite |
3550 | * connection tear down and (memory) recovery. | |
845704a5 ED |
3551 | * Otherwise tcp_send_fin() could be tempted to either delay FIN |
3552 | * or even be forced to close flow without any FIN. | |
a6c5ea4c ED |
3553 | * In general, we want to allow one skb per socket to avoid hangs |
3554 | * with edge trigger epoll() | |
d83769a5 | 3555 | */ |
a6c5ea4c | 3556 | void sk_forced_mem_schedule(struct sock *sk, int size) |
d83769a5 | 3557 | { |
c4ee1185 | 3558 | int delta, amt; |
d83769a5 | 3559 | |
c4ee1185 ED |
3560 | delta = size - sk->sk_forward_alloc; |
3561 | if (delta <= 0) | |
d83769a5 | 3562 | return; |
c4ee1185 | 3563 | amt = sk_mem_pages(delta); |
5e6300e7 | 3564 | sk_forward_alloc_add(sk, amt << PAGE_SHIFT); |
e805605c JW |
3565 | sk_memory_allocated_add(sk, amt); |
3566 | ||
baac50bb | 3567 | if (mem_cgroup_sockets_enabled && sk->sk_memcg) |
4b1327be WW |
3568 | mem_cgroup_charge_skmem(sk->sk_memcg, amt, |
3569 | gfp_memcg_charge() | __GFP_NOFAIL); | |
d83769a5 ED |
3570 | } |
3571 | ||
845704a5 ED |
3572 | /* Send a FIN. The caller locks the socket for us. |
3573 | * We should try to send a FIN packet really hard, but eventually give up. | |
1da177e4 LT |
3574 | */ |
3575 | void tcp_send_fin(struct sock *sk) | |
3576 | { | |
ee2aabd3 | 3577 | struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk); |
e905a9ed | 3578 | struct tcp_sock *tp = tcp_sk(sk); |
e905a9ed | 3579 | |
845704a5 ED |
3580 | /* Optimization, tack on the FIN if we have one skb in write queue and |
3581 | * this skb was not yet sent, or we are under memory pressure. | |
3582 | * Note: in the latter case, FIN packet will be sent after a timeout, | |
3583 | * as TCP stack thinks it has already been transmitted. | |
1da177e4 | 3584 | */ |
ee2aabd3 | 3585 | tskb = tail; |
75c119af ED |
3586 | if (!tskb && tcp_under_memory_pressure(sk)) |
3587 | tskb = skb_rb_last(&sk->tcp_rtx_queue); | |
3588 | ||
3589 | if (tskb) { | |
845704a5 ED |
3590 | TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; |
3591 | TCP_SKB_CB(tskb)->end_seq++; | |
1da177e4 | 3592 | tp->write_seq++; |
ee2aabd3 | 3593 | if (!tail) { |
845704a5 ED |
3594 | /* This means tskb was already sent. |
3595 | * Pretend we included the FIN on previous transmit. | |
3596 | * We need to set tp->snd_nxt to the value it would have | |
3597 | * if FIN had been sent. This is because retransmit path | |
3598 | * does not change tp->snd_nxt. | |
3599 | */ | |
e0d694d6 | 3600 | WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1); |
845704a5 ED |
3601 | return; |
3602 | } | |
1da177e4 | 3603 | } else { |
94062790 ED |
3604 | skb = alloc_skb_fclone(MAX_TCP_HEADER, |
3605 | sk_gfp_mask(sk, GFP_ATOMIC | | |
3606 | __GFP_NOWARN)); | |
d1edc085 | 3607 | if (unlikely(!skb)) |
845704a5 | 3608 | return; |
d1edc085 | 3609 | |
e2080072 | 3610 | INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); |
d83769a5 | 3611 | skb_reserve(skb, MAX_TCP_HEADER); |
a6c5ea4c | 3612 | sk_forced_mem_schedule(sk, skb->truesize); |
1da177e4 | 3613 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ |
e870a8ef | 3614 | tcp_init_nondata_skb(skb, tp->write_seq, |
a3433f35 | 3615 | TCPHDR_ACK | TCPHDR_FIN); |
1da177e4 LT |
3616 | tcp_queue_skb(sk, skb); |
3617 | } | |
845704a5 | 3618 | __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); |
1da177e4 LT |
3619 | } |
3620 | ||
3621 | /* We get here when a process closes a file descriptor (either due to | |
3622 | * an explicit close() or as a byproduct of exit()'ing) and there | |
3623 | * was unread data in the receive queue. This behavior is recommended | |
65bb723c | 3624 | * by RFC 2525, section 2.17. -DaveM |
1da177e4 | 3625 | */ |
5691276b JX |
3626 | void tcp_send_active_reset(struct sock *sk, gfp_t priority, |
3627 | enum sk_rst_reason reason) | |
1da177e4 | 3628 | { |
1da177e4 LT |
3629 | struct sk_buff *skb; |
3630 | ||
7cc2b043 GF |
3631 | TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); |
3632 | ||
1da177e4 LT |
3633 | /* NOTE: No TCP options attached and we never retransmit this. */ |
3634 | skb = alloc_skb(MAX_TCP_HEADER, priority); | |
3635 | if (!skb) { | |
4e673444 | 3636 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
1da177e4 LT |
3637 | return; |
3638 | } | |
3639 | ||
3640 | /* Reserve space for headers and prepare control bits. */ | |
3641 | skb_reserve(skb, MAX_TCP_HEADER); | |
e870a8ef | 3642 | tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), |
a3433f35 | 3643 | TCPHDR_ACK | TCPHDR_RST); |
9a568de4 | 3644 | tcp_mstamp_refresh(tcp_sk(sk)); |
1da177e4 | 3645 | /* Send it off. */ |
dfb4b9dc | 3646 | if (tcp_transmit_skb(sk, skb, 0, priority)) |
4e673444 | 3647 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
c24b14c4 SL |
3648 | |
3649 | /* skb of trace_tcp_send_reset() keeps the skb that caused RST, | |
3650 | * skb here is different to the troublesome skb, so use NULL | |
3651 | */ | |
b533fb9c | 3652 | trace_tcp_send_reset(sk, NULL, SK_RST_REASON_NOT_SPECIFIED); |
1da177e4 LT |
3653 | } |
3654 | ||
67edfef7 AK |
3655 | /* Send a crossed SYN-ACK during socket establishment. |
3656 | * WARNING: This routine must only be called when we have already sent | |
1da177e4 LT |
3657 | * a SYN packet that crossed the incoming SYN that caused this routine |
3658 | * to get called. If this assumption fails then the initial rcv_wnd | |
3659 | * and rcv_wscale values will not be correct. | |
3660 | */ | |
3661 | int tcp_send_synack(struct sock *sk) | |
3662 | { | |
056834d9 | 3663 | struct sk_buff *skb; |
1da177e4 | 3664 | |
75c119af | 3665 | skb = tcp_rtx_queue_head(sk); |
51456b29 | 3666 | if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { |
75c119af | 3667 | pr_err("%s: wrong queue state\n", __func__); |
1da177e4 LT |
3668 | return -EFAULT; |
3669 | } | |
4de075e0 | 3670 | if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { |
1da177e4 | 3671 | if (skb_cloned(skb)) { |
e2080072 ED |
3672 | struct sk_buff *nskb; |
3673 | ||
3674 | tcp_skb_tsorted_save(skb) { | |
3675 | nskb = skb_copy(skb, GFP_ATOMIC); | |
3676 | } tcp_skb_tsorted_restore(skb); | |
51456b29 | 3677 | if (!nskb) |
1da177e4 | 3678 | return -ENOMEM; |
e2080072 | 3679 | INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); |
2bec445f | 3680 | tcp_highest_sack_replace(sk, skb, nskb); |
75c119af | 3681 | tcp_rtx_queue_unlink_and_free(skb, sk); |
f4a775d1 | 3682 | __skb_header_release(nskb); |
75c119af | 3683 | tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); |
ab4e846a | 3684 | sk_wmem_queued_add(sk, nskb->truesize); |
3ab224be | 3685 | sk_mem_charge(sk, nskb->truesize); |
1da177e4 LT |
3686 | skb = nskb; |
3687 | } | |
3688 | ||
4de075e0 | 3689 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; |
735d3831 | 3690 | tcp_ecn_send_synack(sk, skb); |
1da177e4 | 3691 | } |
dfb4b9dc | 3692 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
1da177e4 LT |
3693 | } |
3694 | ||
4aea39c1 | 3695 | /** |
331fca43 MKL |
3696 | * tcp_make_synack - Allocate one skb and build a SYNACK packet. |
3697 | * @sk: listener socket | |
3698 | * @dst: dst entry attached to the SYNACK. It is consumed and caller | |
3699 | * should not use it again. | |
3700 | * @req: request_sock pointer | |
3701 | * @foc: cookie for tcp fast open | |
3702 | * @synack_type: Type of synack to prepare | |
3703 | * @syn_skb: SYN packet just received. It could be NULL for rtx case. | |
4aea39c1 | 3704 | */ |
5d062de7 | 3705 | struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, |
e6b4d113 | 3706 | struct request_sock *req, |
ca6fb065 | 3707 | struct tcp_fastopen_cookie *foc, |
331fca43 MKL |
3708 | enum tcp_synack_type synack_type, |
3709 | struct sk_buff *syn_skb) | |
1da177e4 | 3710 | { |
2e6599cb | 3711 | struct inet_request_sock *ireq = inet_rsk(req); |
5d062de7 | 3712 | const struct tcp_sock *tp = tcp_sk(sk); |
5d062de7 | 3713 | struct tcp_out_options opts; |
1e03d32b | 3714 | struct tcp_key key = {}; |
5d062de7 | 3715 | struct sk_buff *skb; |
bd0388ae | 3716 | int tcp_header_size; |
5d062de7 | 3717 | struct tcphdr *th; |
f5fff5dc | 3718 | int mss; |
a842fe14 | 3719 | u64 now; |
1da177e4 | 3720 | |
ca6fb065 | 3721 | skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); |
4aea39c1 ED |
3722 | if (unlikely(!skb)) { |
3723 | dst_release(dst); | |
1da177e4 | 3724 | return NULL; |
4aea39c1 | 3725 | } |
1da177e4 LT |
3726 | /* Reserve space for headers. */ |
3727 | skb_reserve(skb, MAX_TCP_HEADER); | |
3728 | ||
b3d05147 ED |
3729 | switch (synack_type) { |
3730 | case TCP_SYNACK_NORMAL: | |
9e17f8a4 | 3731 | skb_set_owner_w(skb, req_to_sk(req)); |
b3d05147 ED |
3732 | break; |
3733 | case TCP_SYNACK_COOKIE: | |
3734 | /* Under synflood, we do not attach skb to a socket, | |
3735 | * to avoid false sharing. | |
3736 | */ | |
3737 | break; | |
3738 | case TCP_SYNACK_FASTOPEN: | |
ca6fb065 ED |
3739 | /* sk is a const pointer, because we want to express multiple |
3740 | * cpu might call us concurrently. | |
3741 | * sk->sk_wmem_alloc in an atomic, we can promote to rw. | |
3742 | */ | |
3743 | skb_set_owner_w(skb, (struct sock *)sk); | |
b3d05147 | 3744 | break; |
ca6fb065 | 3745 | } |
4aea39c1 | 3746 | skb_dst_set(skb, dst); |
1da177e4 | 3747 | |
3541f9e8 | 3748 | mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); |
f5fff5dc | 3749 | |
33ad798c | 3750 | memset(&opts, 0, sizeof(opts)); |
a842fe14 | 3751 | now = tcp_clock_ns(); |
8b5f12d0 | 3752 | #ifdef CONFIG_SYN_COOKIES |
f8ace8d9 | 3753 | if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok)) |
a1ac9c8a MKL |
3754 | skb_set_delivery_time(skb, cookie_init_timestamp(req, now), |
3755 | true); | |
8b5f12d0 FW |
3756 | else |
3757 | #endif | |
9e450c1e | 3758 | { |
a1ac9c8a | 3759 | skb_set_delivery_time(skb, now, true); |
9e450c1e YC |
3760 | if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */ |
3761 | tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb); | |
3762 | } | |
80f03e27 | 3763 | |
9427c6aa | 3764 | #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO) |
80f03e27 | 3765 | rcu_read_lock(); |
80f03e27 | 3766 | #endif |
9427c6aa DS |
3767 | if (tcp_rsk_used_ao(req)) { |
3768 | #ifdef CONFIG_TCP_AO | |
3769 | struct tcp_ao_key *ao_key = NULL; | |
9427c6aa DS |
3770 | u8 keyid = tcp_rsk(req)->ao_keyid; |
3771 | ||
3772 | ao_key = tcp_sk(sk)->af_specific->ao_lookup(sk, req_to_sk(req), | |
3773 | keyid, -1); | |
3774 | /* If there is no matching key - avoid sending anything, | |
3775 | * especially usigned segments. It could try harder and lookup | |
3776 | * for another peer-matching key, but the peer has requested | |
3777 | * ao_keyid (RFC5925 RNextKeyID), so let's keep it simple here. | |
3778 | */ | |
9396c4ee | 3779 | if (unlikely(!ao_key)) { |
9427c6aa DS |
3780 | rcu_read_unlock(); |
3781 | kfree_skb(skb); | |
9396c4ee DS |
3782 | net_warn_ratelimited("TCP-AO: the keyid %u from SYN packet is not present - not sending SYNACK\n", |
3783 | keyid); | |
9427c6aa DS |
3784 | return NULL; |
3785 | } | |
3786 | key.ao_key = ao_key; | |
3787 | key.type = TCP_KEY_AO; | |
3788 | #endif | |
3789 | } else { | |
3790 | #ifdef CONFIG_TCP_MD5SIG | |
3791 | key.md5_key = tcp_rsk(req)->af_specific->req_md5_lookup(sk, | |
3792 | req_to_sk(req)); | |
3793 | if (key.md5_key) | |
3794 | key.type = TCP_KEY_MD5; | |
3795 | #endif | |
3796 | } | |
5e526552 | 3797 | skb_set_hash(skb, READ_ONCE(tcp_rsk(req)->txhash), PKT_HASH_TYPE_L4); |
331fca43 MKL |
3798 | /* bpf program will be interested in the tcp_flags */ |
3799 | TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK; | |
9427c6aa DS |
3800 | tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, |
3801 | &key, foc, synack_type, syn_skb) | |
3802 | + sizeof(*th); | |
cfb6eeb4 | 3803 | |
aa8223c7 ACM |
3804 | skb_push(skb, tcp_header_size); |
3805 | skb_reset_transport_header(skb); | |
1da177e4 | 3806 | |
ea1627c2 | 3807 | th = (struct tcphdr *)skb->data; |
1da177e4 LT |
3808 | memset(th, 0, sizeof(struct tcphdr)); |
3809 | th->syn = 1; | |
3810 | th->ack = 1; | |
6ac705b1 | 3811 | tcp_ecn_make_synack(req, th); |
b44084c2 | 3812 | th->source = htons(ireq->ir_num); |
634fb979 | 3813 | th->dest = ireq->ir_rmt_port; |
e05a90ec | 3814 | skb->mark = ireq->ir_mark; |
3b117750 ED |
3815 | skb->ip_summed = CHECKSUM_PARTIAL; |
3816 | th->seq = htonl(tcp_rsk(req)->snt_isn); | |
8336886f JC |
3817 | /* XXX data is queued and acked as is. No buffer/window check */ |
3818 | th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); | |
1da177e4 LT |
3819 | |
3820 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ | |
ed53d0ab | 3821 | th->window = htons(min(req->rsk_rcv_wnd, 65535U)); |
9427c6aa | 3822 | tcp_options_write(th, NULL, tcp_rsk(req), &opts, &key); |
1da177e4 | 3823 | th->doff = (tcp_header_size >> 2); |
bced3f7d | 3824 | TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); |
cfb6eeb4 | 3825 | |
cfb6eeb4 | 3826 | /* Okay, we have all we need - do the md5 hash if needed */ |
9427c6aa DS |
3827 | if (tcp_key_is_md5(&key)) { |
3828 | #ifdef CONFIG_TCP_MD5SIG | |
bd0388ae | 3829 | tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, |
9427c6aa DS |
3830 | key.md5_key, req_to_sk(req), skb); |
3831 | #endif | |
3832 | } else if (tcp_key_is_ao(&key)) { | |
3833 | #ifdef CONFIG_TCP_AO | |
3834 | tcp_rsk(req)->af_specific->ao_synack_hash(opts.hash_location, | |
3835 | key.ao_key, req, skb, | |
3836 | opts.hash_location - (u8 *)th, 0); | |
3837 | #endif | |
3838 | } | |
3839 | #if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO) | |
80f03e27 | 3840 | rcu_read_unlock(); |
cfb6eeb4 YH |
3841 | #endif |
3842 | ||
331fca43 MKL |
3843 | bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb, |
3844 | synack_type, &opts); | |
3845 | ||
a1ac9c8a | 3846 | skb_set_delivery_time(skb, now, true); |
a842fe14 ED |
3847 | tcp_add_tx_delay(skb, tp); |
3848 | ||
1da177e4 LT |
3849 | return skb; |
3850 | } | |
4bc2f18b | 3851 | EXPORT_SYMBOL(tcp_make_synack); |
1da177e4 | 3852 | |
81164413 DB |
3853 | static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) |
3854 | { | |
3855 | struct inet_connection_sock *icsk = inet_csk(sk); | |
3856 | const struct tcp_congestion_ops *ca; | |
3857 | u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); | |
3858 | ||
3859 | if (ca_key == TCP_CA_UNSPEC) | |
3860 | return; | |
3861 | ||
3862 | rcu_read_lock(); | |
3863 | ca = tcp_ca_find_key(ca_key); | |
0baf26b0 MKL |
3864 | if (likely(ca && bpf_try_module_get(ca, ca->owner))) { |
3865 | bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner); | |
81164413 DB |
3866 | icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); |
3867 | icsk->icsk_ca_ops = ca; | |
3868 | } | |
3869 | rcu_read_unlock(); | |
3870 | } | |
3871 | ||
67edfef7 | 3872 | /* Do all connect socket setups that can be done AF independent. */ |
f7e56a76 | 3873 | static void tcp_connect_init(struct sock *sk) |
1da177e4 | 3874 | { |
cf533ea5 | 3875 | const struct dst_entry *dst = __sk_dst_get(sk); |
1da177e4 LT |
3876 | struct tcp_sock *tp = tcp_sk(sk); |
3877 | __u8 rcv_wscale; | |
13d3b1eb | 3878 | u32 rcv_wnd; |
1da177e4 LT |
3879 | |
3880 | /* We'll fix this up when we get a response from the other end. | |
3881 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. | |
3882 | */ | |
5d2ed052 | 3883 | tp->tcp_header_len = sizeof(struct tcphdr); |
3666f666 | 3884 | if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps)) |
5d2ed052 | 3885 | tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; |
1da177e4 | 3886 | |
7c2ffaf2 DS |
3887 | tcp_ao_connect_init(sk); |
3888 | ||
1da177e4 LT |
3889 | /* If user gave his TCP_MAXSEG, record it to clamp */ |
3890 | if (tp->rx_opt.user_mss) | |
3891 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
3892 | tp->max_window = 0; | |
5d424d5a | 3893 | tcp_mtup_init(sk); |
1da177e4 LT |
3894 | tcp_sync_mss(sk, dst_mtu(dst)); |
3895 | ||
81164413 DB |
3896 | tcp_ca_dst_init(sk, dst); |
3897 | ||
1da177e4 | 3898 | if (!tp->window_clamp) |
f410cbea | 3899 | WRITE_ONCE(tp->window_clamp, dst_metric(dst, RTAX_WINDOW)); |
3541f9e8 | 3900 | tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); |
f5fff5dc | 3901 | |
1da177e4 | 3902 | tcp_initialize_rcv_mss(sk); |
1da177e4 | 3903 | |
e88c64f0 HPP |
3904 | /* limit the window selection if the user enforce a smaller rx buffer */ |
3905 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && | |
3906 | (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) | |
f410cbea | 3907 | WRITE_ONCE(tp->window_clamp, tcp_full_space(sk)); |
e88c64f0 | 3908 | |
13d3b1eb LB |
3909 | rcv_wnd = tcp_rwnd_init_bpf(sk); |
3910 | if (rcv_wnd == 0) | |
3911 | rcv_wnd = dst_metric(dst, RTAX_INITRWND); | |
3912 | ||
ceef9ab6 | 3913 | tcp_select_initial_window(sk, tcp_full_space(sk), |
1da177e4 LT |
3914 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), |
3915 | &tp->rcv_wnd, | |
3916 | &tp->window_clamp, | |
3666f666 | 3917 | READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling), |
31d12926 | 3918 | &rcv_wscale, |
13d3b1eb | 3919 | rcv_wnd); |
1da177e4 LT |
3920 | |
3921 | tp->rx_opt.rcv_wscale = rcv_wscale; | |
3922 | tp->rcv_ssthresh = tp->rcv_wnd; | |
3923 | ||
e13ec3da | 3924 | WRITE_ONCE(sk->sk_err, 0); |
1da177e4 LT |
3925 | sock_reset_flag(sk, SOCK_DONE); |
3926 | tp->snd_wnd = 0; | |
ee7537b6 | 3927 | tcp_init_wl(tp, 0); |
7f582b24 | 3928 | tcp_write_queue_purge(sk); |
1da177e4 LT |
3929 | tp->snd_una = tp->write_seq; |
3930 | tp->snd_sml = tp->write_seq; | |
33f5f57e | 3931 | tp->snd_up = tp->write_seq; |
e0d694d6 | 3932 | WRITE_ONCE(tp->snd_nxt, tp->write_seq); |
ee995283 PE |
3933 | |
3934 | if (likely(!tp->repair)) | |
3935 | tp->rcv_nxt = 0; | |
c7781a6e | 3936 | else |
70eabf0e | 3937 | tp->rcv_tstamp = tcp_jiffies32; |
ee995283 | 3938 | tp->rcv_wup = tp->rcv_nxt; |
7db48e98 | 3939 | WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); |
1da177e4 | 3940 | |
8550f328 | 3941 | inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); |
463c84b9 | 3942 | inet_csk(sk)->icsk_retransmits = 0; |
1da177e4 LT |
3943 | tcp_clear_retrans(tp); |
3944 | } | |
3945 | ||
783237e8 YC |
3946 | static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) |
3947 | { | |
3948 | struct tcp_sock *tp = tcp_sk(sk); | |
3949 | struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); | |
3950 | ||
3951 | tcb->end_seq += skb->len; | |
f4a775d1 | 3952 | __skb_header_release(skb); |
ab4e846a | 3953 | sk_wmem_queued_add(sk, skb->truesize); |
783237e8 | 3954 | sk_mem_charge(sk, skb->truesize); |
0f317464 | 3955 | WRITE_ONCE(tp->write_seq, tcb->end_seq); |
783237e8 YC |
3956 | tp->packets_out += tcp_skb_pcount(skb); |
3957 | } | |
3958 | ||
3959 | /* Build and send a SYN with data and (cached) Fast Open cookie. However, | |
3960 | * queue a data-only packet after the regular SYN, such that regular SYNs | |
3961 | * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges | |
3962 | * only the SYN sequence, the data are retransmitted in the first ACK. | |
3963 | * If cookie is not cached or other error occurs, falls back to send a | |
3964 | * regular SYN with Fast Open cookie request option. | |
3965 | */ | |
3966 | static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) | |
3967 | { | |
ed0c99dc | 3968 | struct inet_connection_sock *icsk = inet_csk(sk); |
783237e8 YC |
3969 | struct tcp_sock *tp = tcp_sk(sk); |
3970 | struct tcp_fastopen_request *fo = tp->fastopen_req; | |
fbf93406 | 3971 | struct page_frag *pfrag = sk_page_frag(sk); |
355a901e | 3972 | struct sk_buff *syn_data; |
fbf93406 | 3973 | int space, err = 0; |
aab48743 | 3974 | |
67da22d2 | 3975 | tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ |
065263f4 | 3976 | if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) |
783237e8 YC |
3977 | goto fallback; |
3978 | ||
3979 | /* MSS for SYN-data is based on cached MSS and bounded by PMTU and | |
3980 | * user-MSS. Reserve maximum option space for middleboxes that add | |
3981 | * private TCP options. The cost is reduced data space in SYN :( | |
3982 | */ | |
3541f9e8 | 3983 | tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); |
ed0c99dc JK |
3984 | /* Sync mss_cache after updating the mss_clamp */ |
3985 | tcp_sync_mss(sk, icsk->icsk_pmtu_cookie); | |
3541f9e8 | 3986 | |
ed0c99dc | 3987 | space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) - |
783237e8 YC |
3988 | MAX_TCP_OPTION_SPACE; |
3989 | ||
f5ddcbbb ED |
3990 | space = min_t(size_t, space, fo->size); |
3991 | ||
fbf93406 ED |
3992 | if (space && |
3993 | !skb_page_frag_refill(min_t(size_t, space, PAGE_SIZE), | |
3994 | pfrag, sk->sk_allocation)) | |
3995 | goto fallback; | |
5882efff | 3996 | syn_data = tcp_stream_alloc_skb(sk, sk->sk_allocation, false); |
355a901e | 3997 | if (!syn_data) |
783237e8 | 3998 | goto fallback; |
355a901e | 3999 | memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); |
07e100f9 | 4000 | if (space) { |
fbf93406 ED |
4001 | space = min_t(size_t, space, pfrag->size - pfrag->offset); |
4002 | space = tcp_wmem_schedule(sk, space); | |
4003 | } | |
4004 | if (space) { | |
4005 | space = copy_page_from_iter(pfrag->page, pfrag->offset, | |
4006 | space, &fo->data->msg_iter); | |
4007 | if (unlikely(!space)) { | |
ba233b34 | 4008 | tcp_skb_tsorted_anchor_cleanup(syn_data); |
07e100f9 ED |
4009 | kfree_skb(syn_data); |
4010 | goto fallback; | |
4011 | } | |
fbf93406 ED |
4012 | skb_fill_page_desc(syn_data, 0, pfrag->page, |
4013 | pfrag->offset, space); | |
4014 | page_ref_inc(pfrag->page); | |
4015 | pfrag->offset += space; | |
4016 | skb_len_add(syn_data, space); | |
f859a448 | 4017 | skb_zcopy_set(syn_data, fo->uarg, NULL); |
57be5bda | 4018 | } |
355a901e ED |
4019 | /* No more data pending in inet_wait_for_connect() */ |
4020 | if (space == fo->size) | |
4021 | fo->data = NULL; | |
4022 | fo->copied = space; | |
783237e8 | 4023 | |
355a901e | 4024 | tcp_connect_queue_skb(sk, syn_data); |
0f87230d FY |
4025 | if (syn_data->len) |
4026 | tcp_chrono_start(sk, TCP_CHRONO_BUSY); | |
783237e8 | 4027 | |
355a901e | 4028 | err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); |
783237e8 | 4029 | |
a1ac9c8a | 4030 | skb_set_delivery_time(syn, syn_data->skb_mstamp_ns, true); |
431a9124 | 4031 | |
355a901e ED |
4032 | /* Now full SYN+DATA was cloned and sent (or not), |
4033 | * remove the SYN from the original skb (syn_data) | |
4034 | * we keep in write queue in case of a retransmit, as we | |
4035 | * also have the SYN packet (with no data) in the same queue. | |
4036 | */ | |
4037 | TCP_SKB_CB(syn_data)->seq++; | |
4038 | TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; | |
4039 | if (!err) { | |
67da22d2 | 4040 | tp->syn_data = (fo->copied > 0); |
75c119af | 4041 | tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); |
f19c29e3 | 4042 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); |
783237e8 YC |
4043 | goto done; |
4044 | } | |
783237e8 | 4045 | |
75c119af ED |
4046 | /* data was not sent, put it in write_queue */ |
4047 | __skb_queue_tail(&sk->sk_write_queue, syn_data); | |
b5b7db8d ED |
4048 | tp->packets_out -= tcp_skb_pcount(syn_data); |
4049 | ||
783237e8 YC |
4050 | fallback: |
4051 | /* Send a regular SYN with Fast Open cookie request option */ | |
4052 | if (fo->cookie.len > 0) | |
4053 | fo->cookie.len = 0; | |
4054 | err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); | |
4055 | if (err) | |
4056 | tp->syn_fastopen = 0; | |
783237e8 YC |
4057 | done: |
4058 | fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ | |
4059 | return err; | |
4060 | } | |
4061 | ||
67edfef7 | 4062 | /* Build a SYN and send it off. */ |
1da177e4 LT |
4063 | int tcp_connect(struct sock *sk) |
4064 | { | |
4065 | struct tcp_sock *tp = tcp_sk(sk); | |
4066 | struct sk_buff *buff; | |
ee586811 | 4067 | int err; |
1da177e4 | 4068 | |
de525be2 | 4069 | tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); |
8ba60924 | 4070 | |
0aadc739 DS |
4071 | #if defined(CONFIG_TCP_MD5SIG) && defined(CONFIG_TCP_AO) |
4072 | /* Has to be checked late, after setting daddr/saddr/ops. | |
4073 | * Return error if the peer has both a md5 and a tcp-ao key | |
4074 | * configured as this is ambiguous. | |
4075 | */ | |
4076 | if (unlikely(rcu_dereference_protected(tp->md5sig_info, | |
4077 | lockdep_sock_is_held(sk)))) { | |
4078 | bool needs_ao = !!tp->af_specific->ao_lookup(sk, sk, -1, -1); | |
4079 | bool needs_md5 = !!tp->af_specific->md5_lookup(sk, sk); | |
4080 | struct tcp_ao_info *ao_info; | |
4081 | ||
4082 | ao_info = rcu_dereference_check(tp->ao_info, | |
4083 | lockdep_sock_is_held(sk)); | |
4084 | if (ao_info) { | |
4085 | /* This is an extra check: tcp_ao_required() in | |
4086 | * tcp_v{4,6}_parse_md5_keys() should prevent adding | |
4087 | * md5 keys on ao_required socket. | |
4088 | */ | |
4089 | needs_ao |= ao_info->ao_required; | |
4090 | WARN_ON_ONCE(ao_info->ao_required && needs_md5); | |
4091 | } | |
4092 | if (needs_md5 && needs_ao) | |
4093 | return -EKEYREJECTED; | |
4094 | ||
4095 | /* If we have a matching md5 key and no matching tcp-ao key | |
4096 | * then free up ao_info if allocated. | |
4097 | */ | |
4098 | if (needs_md5) { | |
decde258 | 4099 | tcp_ao_destroy_sock(sk, false); |
0aadc739 DS |
4100 | } else if (needs_ao) { |
4101 | tcp_clear_md5_list(sk); | |
4102 | kfree(rcu_replace_pointer(tp->md5sig_info, NULL, | |
4103 | lockdep_sock_is_held(sk))); | |
4104 | } | |
4105 | } | |
4106 | #endif | |
4107 | #ifdef CONFIG_TCP_AO | |
4108 | if (unlikely(rcu_dereference_protected(tp->ao_info, | |
4109 | lockdep_sock_is_held(sk)))) { | |
4110 | /* Don't allow connecting if ao is configured but no | |
4111 | * matching key is found. | |
4112 | */ | |
4113 | if (!tp->af_specific->ao_lookup(sk, sk, -1, -1)) | |
4114 | return -EKEYREJECTED; | |
4115 | } | |
4116 | #endif | |
4117 | ||
8ba60924 ED |
4118 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) |
4119 | return -EHOSTUNREACH; /* Routing failure or similar. */ | |
4120 | ||
1da177e4 LT |
4121 | tcp_connect_init(sk); |
4122 | ||
2b916477 AV |
4123 | if (unlikely(tp->repair)) { |
4124 | tcp_finish_connect(sk, NULL); | |
4125 | return 0; | |
4126 | } | |
4127 | ||
5882efff | 4128 | buff = tcp_stream_alloc_skb(sk, sk->sk_allocation, true); |
355a901e | 4129 | if (unlikely(!buff)) |
1da177e4 LT |
4130 | return -ENOBUFS; |
4131 | ||
a3433f35 | 4132 | tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); |
9a568de4 | 4133 | tcp_mstamp_refresh(tp); |
9d0c00f5 | 4134 | tp->retrans_stamp = tcp_time_stamp_ts(tp); |
783237e8 | 4135 | tcp_connect_queue_skb(sk, buff); |
735d3831 | 4136 | tcp_ecn_send_syn(sk, buff); |
75c119af | 4137 | tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); |
1da177e4 | 4138 | |
783237e8 YC |
4139 | /* Send off SYN; include data in Fast Open. */ |
4140 | err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : | |
4141 | tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); | |
ee586811 EP |
4142 | if (err == -ECONNREFUSED) |
4143 | return err; | |
bd37a088 WY |
4144 | |
4145 | /* We change tp->snd_nxt after the tcp_transmit_skb() call | |
4146 | * in order to make this packet get counted in tcpOutSegs. | |
4147 | */ | |
e0d694d6 | 4148 | WRITE_ONCE(tp->snd_nxt, tp->write_seq); |
bd37a088 | 4149 | tp->pushed_seq = tp->write_seq; |
b5b7db8d ED |
4150 | buff = tcp_send_head(sk); |
4151 | if (unlikely(buff)) { | |
e0d694d6 | 4152 | WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq); |
b5b7db8d ED |
4153 | tp->pushed_seq = TCP_SKB_CB(buff)->seq; |
4154 | } | |
81cc8a75 | 4155 | TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); |
1da177e4 LT |
4156 | |
4157 | /* Timer for repeating the SYN until an answer. */ | |
3f421baa ACM |
4158 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
4159 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
1da177e4 LT |
4160 | return 0; |
4161 | } | |
4bc2f18b | 4162 | EXPORT_SYMBOL(tcp_connect); |
1da177e4 | 4163 | |
bbf80d71 ED |
4164 | u32 tcp_delack_max(const struct sock *sk) |
4165 | { | |
4166 | const struct dst_entry *dst = __sk_dst_get(sk); | |
4167 | u32 delack_max = inet_csk(sk)->icsk_delack_max; | |
4168 | ||
4169 | if (dst && dst_metric_locked(dst, RTAX_RTO_MIN)) { | |
4170 | u32 rto_min = dst_metric_rtt(dst, RTAX_RTO_MIN); | |
4171 | u32 delack_from_rto_min = max_t(int, 1, rto_min - 1); | |
4172 | ||
4173 | delack_max = min_t(u32, delack_max, delack_from_rto_min); | |
4174 | } | |
4175 | return delack_max; | |
4176 | } | |
4177 | ||
1da177e4 LT |
4178 | /* Send out a delayed ack, the caller does the policy checking |
4179 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() | |
4180 | * for details. | |
4181 | */ | |
4182 | void tcp_send_delayed_ack(struct sock *sk) | |
4183 | { | |
463c84b9 ACM |
4184 | struct inet_connection_sock *icsk = inet_csk(sk); |
4185 | int ato = icsk->icsk_ack.ato; | |
1da177e4 LT |
4186 | unsigned long timeout; |
4187 | ||
4188 | if (ato > TCP_DELACK_MIN) { | |
463c84b9 | 4189 | const struct tcp_sock *tp = tcp_sk(sk); |
056834d9 | 4190 | int max_ato = HZ / 2; |
1da177e4 | 4191 | |
31954cd8 | 4192 | if (inet_csk_in_pingpong_mode(sk) || |
056834d9 | 4193 | (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) |
1da177e4 LT |
4194 | max_ato = TCP_DELACK_MAX; |
4195 | ||
4196 | /* Slow path, intersegment interval is "high". */ | |
4197 | ||
4198 | /* If some rtt estimate is known, use it to bound delayed ack. | |
463c84b9 | 4199 | * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements |
1da177e4 LT |
4200 | * directly. |
4201 | */ | |
740b0f18 ED |
4202 | if (tp->srtt_us) { |
4203 | int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), | |
4204 | TCP_DELACK_MIN); | |
1da177e4 LT |
4205 | |
4206 | if (rtt < max_ato) | |
4207 | max_ato = rtt; | |
4208 | } | |
4209 | ||
4210 | ato = min(ato, max_ato); | |
4211 | } | |
4212 | ||
bbf80d71 | 4213 | ato = min_t(u32, ato, tcp_delack_max(sk)); |
2b8ee4f0 | 4214 | |
1da177e4 LT |
4215 | /* Stay within the limit we were given */ |
4216 | timeout = jiffies + ato; | |
4217 | ||
4218 | /* Use new timeout only if there wasn't a older one earlier. */ | |
463c84b9 | 4219 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { |
b6b6d653 ED |
4220 | /* If delack timer is about to expire, send ACK now. */ |
4221 | if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { | |
1da177e4 LT |
4222 | tcp_send_ack(sk); |
4223 | return; | |
4224 | } | |
4225 | ||
463c84b9 ACM |
4226 | if (!time_before(timeout, icsk->icsk_ack.timeout)) |
4227 | timeout = icsk->icsk_ack.timeout; | |
1da177e4 | 4228 | } |
463c84b9 ACM |
4229 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; |
4230 | icsk->icsk_ack.timeout = timeout; | |
4231 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); | |
1da177e4 LT |
4232 | } |
4233 | ||
4234 | /* This routine sends an ack and also updates the window. */ | |
2987babb | 4235 | void __tcp_send_ack(struct sock *sk, u32 rcv_nxt) |
1da177e4 | 4236 | { |
058dc334 | 4237 | struct sk_buff *buff; |
1da177e4 | 4238 | |
058dc334 IJ |
4239 | /* If we have been reset, we may not send again. */ |
4240 | if (sk->sk_state == TCP_CLOSE) | |
4241 | return; | |
1da177e4 | 4242 | |
058dc334 IJ |
4243 | /* We are not putting this on the write queue, so |
4244 | * tcp_transmit_skb() will set the ownership to this | |
4245 | * sock. | |
4246 | */ | |
7450aaf6 ED |
4247 | buff = alloc_skb(MAX_TCP_HEADER, |
4248 | sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); | |
4249 | if (unlikely(!buff)) { | |
a37c2134 ED |
4250 | struct inet_connection_sock *icsk = inet_csk(sk); |
4251 | unsigned long delay; | |
4252 | ||
4253 | delay = TCP_DELACK_MAX << icsk->icsk_ack.retry; | |
4254 | if (delay < TCP_RTO_MAX) | |
4255 | icsk->icsk_ack.retry++; | |
058dc334 | 4256 | inet_csk_schedule_ack(sk); |
a37c2134 ED |
4257 | icsk->icsk_ack.ato = TCP_ATO_MIN; |
4258 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX); | |
058dc334 | 4259 | return; |
1da177e4 | 4260 | } |
058dc334 IJ |
4261 | |
4262 | /* Reserve space for headers and prepare control bits. */ | |
4263 | skb_reserve(buff, MAX_TCP_HEADER); | |
a3433f35 | 4264 | tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); |
058dc334 | 4265 | |
98781965 ED |
4266 | /* We do not want pure acks influencing TCP Small Queues or fq/pacing |
4267 | * too much. | |
4268 | * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 | |
98781965 ED |
4269 | */ |
4270 | skb_set_tcp_pure_ack(buff); | |
4271 | ||
058dc334 | 4272 | /* Send it off, this clears delayed acks for us. */ |
2987babb YC |
4273 | __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt); |
4274 | } | |
27cde44a | 4275 | EXPORT_SYMBOL_GPL(__tcp_send_ack); |
2987babb YC |
4276 | |
4277 | void tcp_send_ack(struct sock *sk) | |
4278 | { | |
4279 | __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt); | |
1da177e4 LT |
4280 | } |
4281 | ||
4282 | /* This routine sends a packet with an out of date sequence | |
4283 | * number. It assumes the other end will try to ack it. | |
4284 | * | |
4285 | * Question: what should we make while urgent mode? | |
4286 | * 4.4BSD forces sending single byte of data. We cannot send | |
4287 | * out of window data, because we have SND.NXT==SND.MAX... | |
4288 | * | |
4289 | * Current solution: to send TWO zero-length segments in urgent mode: | |
4290 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is | |
4291 | * out-of-date with SND.UNA-1 to probe window. | |
4292 | */ | |
e520af48 | 4293 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) |
1da177e4 LT |
4294 | { |
4295 | struct tcp_sock *tp = tcp_sk(sk); | |
4296 | struct sk_buff *skb; | |
4297 | ||
4298 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ | |
7450aaf6 ED |
4299 | skb = alloc_skb(MAX_TCP_HEADER, |
4300 | sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); | |
51456b29 | 4301 | if (!skb) |
1da177e4 LT |
4302 | return -1; |
4303 | ||
4304 | /* Reserve space for headers and set control bits. */ | |
4305 | skb_reserve(skb, MAX_TCP_HEADER); | |
1da177e4 LT |
4306 | /* Use a previous sequence. This should cause the other |
4307 | * end to send an ack. Don't queue or clone SKB, just | |
4308 | * send it. | |
4309 | */ | |
a3433f35 | 4310 | tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); |
e2e8009f | 4311 | NET_INC_STATS(sock_net(sk), mib); |
7450aaf6 | 4312 | return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); |
1da177e4 LT |
4313 | } |
4314 | ||
385e2070 | 4315 | /* Called from setsockopt( ... TCP_REPAIR ) */ |
ee995283 PE |
4316 | void tcp_send_window_probe(struct sock *sk) |
4317 | { | |
4318 | if (sk->sk_state == TCP_ESTABLISHED) { | |
4319 | tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; | |
9a568de4 | 4320 | tcp_mstamp_refresh(tcp_sk(sk)); |
e520af48 | 4321 | tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); |
ee995283 PE |
4322 | } |
4323 | } | |
4324 | ||
67edfef7 | 4325 | /* Initiate keepalive or window probe from timer. */ |
e520af48 | 4326 | int tcp_write_wakeup(struct sock *sk, int mib) |
1da177e4 | 4327 | { |
058dc334 IJ |
4328 | struct tcp_sock *tp = tcp_sk(sk); |
4329 | struct sk_buff *skb; | |
1da177e4 | 4330 | |
058dc334 IJ |
4331 | if (sk->sk_state == TCP_CLOSE) |
4332 | return -1; | |
4333 | ||
00db4124 IM |
4334 | skb = tcp_send_head(sk); |
4335 | if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { | |
058dc334 | 4336 | int err; |
0c54b85f | 4337 | unsigned int mss = tcp_current_mss(sk); |
058dc334 IJ |
4338 | unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
4339 | ||
4340 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) | |
4341 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; | |
4342 | ||
4343 | /* We are probing the opening of a window | |
4344 | * but the window size is != 0 | |
4345 | * must have been a result SWS avoidance ( sender ) | |
4346 | */ | |
4347 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || | |
4348 | skb->len > mss) { | |
4349 | seg_size = min(seg_size, mss); | |
4de075e0 | 4350 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
75c119af ED |
4351 | if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, |
4352 | skb, seg_size, mss, GFP_ATOMIC)) | |
058dc334 IJ |
4353 | return -1; |
4354 | } else if (!tcp_skb_pcount(skb)) | |
5bbb432c | 4355 | tcp_set_skb_tso_segs(skb, mss); |
058dc334 | 4356 | |
4de075e0 | 4357 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
058dc334 IJ |
4358 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
4359 | if (!err) | |
4360 | tcp_event_new_data_sent(sk, skb); | |
4361 | return err; | |
4362 | } else { | |
33f5f57e | 4363 | if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) |
e520af48 ED |
4364 | tcp_xmit_probe_skb(sk, 1, mib); |
4365 | return tcp_xmit_probe_skb(sk, 0, mib); | |
1da177e4 | 4366 | } |
1da177e4 LT |
4367 | } |
4368 | ||
4369 | /* A window probe timeout has occurred. If window is not closed send | |
4370 | * a partial packet else a zero probe. | |
4371 | */ | |
4372 | void tcp_send_probe0(struct sock *sk) | |
4373 | { | |
463c84b9 | 4374 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 4375 | struct tcp_sock *tp = tcp_sk(sk); |
c6214a97 | 4376 | struct net *net = sock_net(sk); |
c1d5674f | 4377 | unsigned long timeout; |
1da177e4 LT |
4378 | int err; |
4379 | ||
e520af48 | 4380 | err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); |
1da177e4 | 4381 | |
75c119af | 4382 | if (tp->packets_out || tcp_write_queue_empty(sk)) { |
1da177e4 | 4383 | /* Cancel probe timer, if it is not required. */ |
6687e988 | 4384 | icsk->icsk_probes_out = 0; |
463c84b9 | 4385 | icsk->icsk_backoff = 0; |
9d9b1ee0 | 4386 | icsk->icsk_probes_tstamp = 0; |
1da177e4 LT |
4387 | return; |
4388 | } | |
4389 | ||
c1d5674f | 4390 | icsk->icsk_probes_out++; |
1da177e4 | 4391 | if (err <= 0) { |
39e24435 | 4392 | if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2)) |
463c84b9 | 4393 | icsk->icsk_backoff++; |
c1d5674f | 4394 | timeout = tcp_probe0_when(sk, TCP_RTO_MAX); |
1da177e4 LT |
4395 | } else { |
4396 | /* If packet was not sent due to local congestion, | |
c1d5674f | 4397 | * Let senders fight for local resources conservatively. |
1da177e4 | 4398 | */ |
c1d5674f YC |
4399 | timeout = TCP_RESOURCE_PROBE_INTERVAL; |
4400 | } | |
344db93a EC |
4401 | |
4402 | timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout); | |
8dc242ad | 4403 | tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX); |
1da177e4 | 4404 | } |
5db92c99 | 4405 | |
ea3bea3a | 4406 | int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) |
5db92c99 OP |
4407 | { |
4408 | const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; | |
4409 | struct flowi fl; | |
4410 | int res; | |
4411 | ||
cb6cd2ce AK |
4412 | /* Paired with WRITE_ONCE() in sock_setsockopt() */ |
4413 | if (READ_ONCE(sk->sk_txrehash) == SOCK_TXREHASH_ENABLED) | |
5e526552 | 4414 | WRITE_ONCE(tcp_rsk(req)->txhash, net_tx_rndhash()); |
331fca43 MKL |
4415 | res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL, |
4416 | NULL); | |
5db92c99 | 4417 | if (!res) { |
0a375c82 ED |
4418 | TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); |
4419 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); | |
e9d9da91 ED |
4420 | if (unlikely(tcp_passive_fastopen(sk))) { |
4421 | /* sk has const attribute because listeners are lockless. | |
4422 | * However in this case, we are dealing with a passive fastopen | |
4423 | * socket thus we can change total_retrans value. | |
4424 | */ | |
4425 | tcp_sk_rw(sk)->total_retrans++; | |
4426 | } | |
cf34ce3d | 4427 | trace_tcp_retransmit_synack(sk, req); |
5db92c99 OP |
4428 | } |
4429 | return res; | |
4430 | } | |
4431 | EXPORT_SYMBOL(tcp_rtx_synack); |