]>
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> |
1da177e4 LT |
42 | |
43 | #include <linux/compiler.h> | |
5a0e3ad6 | 44 | #include <linux/gfp.h> |
1da177e4 | 45 | #include <linux/module.h> |
60e2a778 | 46 | #include <linux/static_key.h> |
1da177e4 | 47 | |
e086101b | 48 | #include <trace/events/tcp.h> |
35089bb2 | 49 | |
9799ccb0 ED |
50 | /* Refresh clocks of a TCP socket, |
51 | * ensuring monotically increasing values. | |
52 | */ | |
53 | void tcp_mstamp_refresh(struct tcp_sock *tp) | |
54 | { | |
55 | u64 val = tcp_clock_ns(); | |
56 | ||
e6d14070 ED |
57 | tp->tcp_clock_cache = val; |
58 | tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC); | |
9799ccb0 ED |
59 | } |
60 | ||
46d3ceab ED |
61 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
62 | int push_one, gfp_t gfp); | |
519855c5 | 63 | |
67edfef7 | 64 | /* Account for new data that has been sent to the network. */ |
75c119af | 65 | static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb) |
1da177e4 | 66 | { |
6ba8a3b1 | 67 | struct inet_connection_sock *icsk = inet_csk(sk); |
9e412ba7 | 68 | struct tcp_sock *tp = tcp_sk(sk); |
66f5fe62 | 69 | unsigned int prior_packets = tp->packets_out; |
9e412ba7 | 70 | |
e0d694d6 | 71 | WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq); |
8512430e | 72 | |
75c119af ED |
73 | __skb_unlink(skb, &sk->sk_write_queue); |
74 | tcp_rbtree_insert(&sk->tcp_rtx_queue, skb); | |
75 | ||
85369750 CZ |
76 | if (tp->highest_sack == NULL) |
77 | tp->highest_sack = skb; | |
78 | ||
66f5fe62 | 79 | tp->packets_out += tcp_skb_pcount(skb); |
bec41a11 | 80 | if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) |
750ea2ba | 81 | tcp_rearm_rto(sk); |
f19c29e3 | 82 | |
f7324acd DM |
83 | NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT, |
84 | tcp_skb_pcount(skb)); | |
1da177e4 LT |
85 | } |
86 | ||
a4ecb15a CC |
87 | /* SND.NXT, if window was not shrunk or the amount of shrunk was less than one |
88 | * window scaling factor due to loss of precision. | |
1da177e4 LT |
89 | * If window has been shrunk, what should we make? It is not clear at all. |
90 | * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-( | |
91 | * Anything in between SND.UNA...SND.UNA+SND.WND also can be already | |
92 | * invalid. OK, let's make this for now: | |
93 | */ | |
cf533ea5 | 94 | static inline __u32 tcp_acceptable_seq(const struct sock *sk) |
1da177e4 | 95 | { |
cf533ea5 | 96 | const struct tcp_sock *tp = tcp_sk(sk); |
9e412ba7 | 97 | |
a4ecb15a CC |
98 | if (!before(tcp_wnd_end(tp), tp->snd_nxt) || |
99 | (tp->rx_opt.wscale_ok && | |
100 | ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale)))) | |
1da177e4 LT |
101 | return tp->snd_nxt; |
102 | else | |
90840def | 103 | return tcp_wnd_end(tp); |
1da177e4 LT |
104 | } |
105 | ||
106 | /* Calculate mss to advertise in SYN segment. | |
107 | * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that: | |
108 | * | |
109 | * 1. It is independent of path mtu. | |
110 | * 2. Ideally, it is maximal possible segment size i.e. 65535-40. | |
111 | * 3. For IPv4 it is reasonable to calculate it from maximal MTU of | |
112 | * attached devices, because some buggy hosts are confused by | |
113 | * large MSS. | |
114 | * 4. We do not make 3, we advertise MSS, calculated from first | |
115 | * hop device mtu, but allow to raise it to ip_rt_min_advmss. | |
116 | * This may be overridden via information stored in routing table. | |
117 | * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible, | |
118 | * probably even Jumbo". | |
119 | */ | |
120 | static __u16 tcp_advertise_mss(struct sock *sk) | |
121 | { | |
122 | struct tcp_sock *tp = tcp_sk(sk); | |
cf533ea5 | 123 | const struct dst_entry *dst = __sk_dst_get(sk); |
1da177e4 LT |
124 | int mss = tp->advmss; |
125 | ||
0dbaee3b DM |
126 | if (dst) { |
127 | unsigned int metric = dst_metric_advmss(dst); | |
128 | ||
129 | if (metric < mss) { | |
130 | mss = metric; | |
131 | tp->advmss = mss; | |
132 | } | |
1da177e4 LT |
133 | } |
134 | ||
135 | return (__u16)mss; | |
136 | } | |
137 | ||
138 | /* RFC2861. Reset CWND after idle period longer RTO to "restart window". | |
6f021c62 ED |
139 | * This is the first part of cwnd validation mechanism. |
140 | */ | |
141 | void tcp_cwnd_restart(struct sock *sk, s32 delta) | |
1da177e4 | 142 | { |
463c84b9 | 143 | struct tcp_sock *tp = tcp_sk(sk); |
6f021c62 | 144 | u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk)); |
1da177e4 LT |
145 | u32 cwnd = tp->snd_cwnd; |
146 | ||
6687e988 | 147 | tcp_ca_event(sk, CA_EVENT_CWND_RESTART); |
1da177e4 | 148 | |
6687e988 | 149 | tp->snd_ssthresh = tcp_current_ssthresh(sk); |
1da177e4 LT |
150 | restart_cwnd = min(restart_cwnd, cwnd); |
151 | ||
463c84b9 | 152 | while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd) |
1da177e4 LT |
153 | cwnd >>= 1; |
154 | tp->snd_cwnd = max(cwnd, restart_cwnd); | |
c2203cf7 | 155 | tp->snd_cwnd_stamp = tcp_jiffies32; |
1da177e4 LT |
156 | tp->snd_cwnd_used = 0; |
157 | } | |
158 | ||
67edfef7 | 159 | /* Congestion state accounting after a packet has been sent. */ |
40efc6fa | 160 | static void tcp_event_data_sent(struct tcp_sock *tp, |
cf533ea5 | 161 | struct sock *sk) |
1da177e4 | 162 | { |
463c84b9 | 163 | struct inet_connection_sock *icsk = inet_csk(sk); |
d635fbe2 | 164 | const u32 now = tcp_jiffies32; |
1da177e4 | 165 | |
05c5a46d NC |
166 | if (tcp_packets_in_flight(tp) == 0) |
167 | tcp_ca_event(sk, CA_EVENT_TX_START); | |
168 | ||
4a41f453 WW |
169 | /* If this is the first data packet sent in response to the |
170 | * previous received data, | |
171 | * and it is a reply for ato after last received packet, | |
172 | * increase pingpong count. | |
1da177e4 | 173 | */ |
4a41f453 WW |
174 | if (before(tp->lsndtime, icsk->icsk_ack.lrcvtime) && |
175 | (u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato) | |
176 | inet_csk_inc_pingpong_cnt(sk); | |
177 | ||
178 | tp->lsndtime = now; | |
1da177e4 LT |
179 | } |
180 | ||
67edfef7 | 181 | /* Account for an ACK we sent. */ |
27cde44a YC |
182 | static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts, |
183 | u32 rcv_nxt) | |
1da177e4 | 184 | { |
5d9f4262 ED |
185 | struct tcp_sock *tp = tcp_sk(sk); |
186 | ||
2b195850 | 187 | if (unlikely(tp->compressed_ack)) { |
200d95f4 | 188 | NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED, |
2b195850 ED |
189 | tp->compressed_ack); |
190 | tp->compressed_ack = 0; | |
5d9f4262 ED |
191 | if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1) |
192 | __sock_put(sk); | |
193 | } | |
27cde44a YC |
194 | |
195 | if (unlikely(rcv_nxt != tp->rcv_nxt)) | |
196 | return; /* Special ACK sent by DCTCP to reflect ECN */ | |
463c84b9 ACM |
197 | tcp_dec_quickack_mode(sk, pkts); |
198 | inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK); | |
1da177e4 LT |
199 | } |
200 | ||
201 | /* Determine a window scaling and initial window to offer. | |
202 | * Based on the assumption that the given amount of space | |
203 | * will be offered. Store the results in the tp structure. | |
204 | * NOTE: for smooth operation initial space offering should | |
205 | * be a multiple of mss if possible. We assume here that mss >= 1. | |
206 | * This MUST be enforced by all callers. | |
207 | */ | |
ceef9ab6 | 208 | void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss, |
1da177e4 | 209 | __u32 *rcv_wnd, __u32 *window_clamp, |
31d12926 | 210 | int wscale_ok, __u8 *rcv_wscale, |
211 | __u32 init_rcv_wnd) | |
1da177e4 LT |
212 | { |
213 | unsigned int space = (__space < 0 ? 0 : __space); | |
214 | ||
215 | /* If no clamp set the clamp to the max possible scaled window */ | |
216 | if (*window_clamp == 0) | |
589c49cb | 217 | (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE); |
1da177e4 LT |
218 | space = min(*window_clamp, space); |
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 | */ |
ceef9ab6 | 232 | if (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 */ |
356d1833 | 243 | space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]); |
f626300a | 244 | space = max_t(u32, space, sysctl_rmem_max); |
316c1592 | 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 */ |
589c49cb | 250 | (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp); |
1da177e4 | 251 | } |
4bc2f18b | 252 | EXPORT_SYMBOL(tcp_select_initial_window); |
1da177e4 LT |
253 | |
254 | /* Chose a new window to advertise, update state in tcp_sock for the | |
255 | * socket, and return result with RFC1323 scaling applied. The return | |
256 | * value can be stuffed directly into th->window for an outgoing | |
257 | * frame. | |
258 | */ | |
40efc6fa | 259 | static u16 tcp_select_window(struct sock *sk) |
1da177e4 LT |
260 | { |
261 | struct tcp_sock *tp = tcp_sk(sk); | |
8e165e20 | 262 | u32 old_win = tp->rcv_wnd; |
1da177e4 LT |
263 | u32 cur_win = tcp_receive_window(tp); |
264 | u32 new_win = __tcp_select_window(sk); | |
265 | ||
266 | /* Never shrink the offered window */ | |
2de979bd | 267 | if (new_win < cur_win) { |
1da177e4 LT |
268 | /* Danger Will Robinson! |
269 | * Don't update rcv_wup/rcv_wnd here or else | |
270 | * we will not be able to advertise a zero | |
271 | * window in time. --DaveM | |
272 | * | |
273 | * Relax Will Robinson. | |
274 | */ | |
8e165e20 FW |
275 | if (new_win == 0) |
276 | NET_INC_STATS(sock_net(sk), | |
277 | LINUX_MIB_TCPWANTZEROWINDOWADV); | |
607bfbf2 | 278 | new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale); |
1da177e4 LT |
279 | } |
280 | tp->rcv_wnd = new_win; | |
281 | tp->rcv_wup = tp->rcv_nxt; | |
282 | ||
283 | /* Make sure we do not exceed the maximum possible | |
284 | * scaled window. | |
285 | */ | |
ceef9ab6 ED |
286 | if (!tp->rx_opt.rcv_wscale && |
287 | sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows) | |
1da177e4 LT |
288 | new_win = min(new_win, MAX_TCP_WINDOW); |
289 | else | |
290 | new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale)); | |
291 | ||
292 | /* RFC1323 scaling applied */ | |
293 | new_win >>= tp->rx_opt.rcv_wscale; | |
294 | ||
31770e34 | 295 | /* If we advertise zero window, disable fast path. */ |
8e165e20 | 296 | if (new_win == 0) { |
31770e34 | 297 | tp->pred_flags = 0; |
8e165e20 FW |
298 | if (old_win) |
299 | NET_INC_STATS(sock_net(sk), | |
300 | LINUX_MIB_TCPTOZEROWINDOWADV); | |
301 | } else if (old_win == 0) { | |
302 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV); | |
303 | } | |
1da177e4 LT |
304 | |
305 | return new_win; | |
306 | } | |
307 | ||
67edfef7 | 308 | /* Packet ECN state for a SYN-ACK */ |
735d3831 | 309 | static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb) |
bdf1ee5d | 310 | { |
30e502a3 DB |
311 | const struct tcp_sock *tp = tcp_sk(sk); |
312 | ||
4de075e0 | 313 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR; |
056834d9 | 314 | if (!(tp->ecn_flags & TCP_ECN_OK)) |
4de075e0 | 315 | TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE; |
91b5b21c LB |
316 | else if (tcp_ca_needs_ecn(sk) || |
317 | tcp_bpf_ca_needs_ecn(sk)) | |
30e502a3 | 318 | INET_ECN_xmit(sk); |
bdf1ee5d IJ |
319 | } |
320 | ||
67edfef7 | 321 | /* Packet ECN state for a SYN. */ |
735d3831 | 322 | static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb) |
bdf1ee5d IJ |
323 | { |
324 | struct tcp_sock *tp = tcp_sk(sk); | |
91b5b21c | 325 | bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk); |
f7b3bec6 | 326 | bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 || |
91b5b21c | 327 | tcp_ca_needs_ecn(sk) || bpf_needs_ecn; |
f7b3bec6 FW |
328 | |
329 | if (!use_ecn) { | |
330 | const struct dst_entry *dst = __sk_dst_get(sk); | |
331 | ||
332 | if (dst && dst_feature(dst, RTAX_FEATURE_ECN)) | |
333 | use_ecn = true; | |
334 | } | |
bdf1ee5d IJ |
335 | |
336 | tp->ecn_flags = 0; | |
f7b3bec6 FW |
337 | |
338 | if (use_ecn) { | |
4de075e0 | 339 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR; |
bdf1ee5d | 340 | tp->ecn_flags = TCP_ECN_OK; |
91b5b21c | 341 | if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn) |
30e502a3 | 342 | INET_ECN_xmit(sk); |
bdf1ee5d IJ |
343 | } |
344 | } | |
345 | ||
49213555 DB |
346 | static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb) |
347 | { | |
348 | if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback) | |
349 | /* tp->ecn_flags are cleared at a later point in time when | |
350 | * SYN ACK is ultimatively being received. | |
351 | */ | |
352 | TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR); | |
353 | } | |
354 | ||
735d3831 | 355 | static void |
6ac705b1 | 356 | tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th) |
bdf1ee5d | 357 | { |
6ac705b1 | 358 | if (inet_rsk(req)->ecn_ok) |
bdf1ee5d IJ |
359 | th->ece = 1; |
360 | } | |
361 | ||
67edfef7 AK |
362 | /* Set up ECN state for a packet on a ESTABLISHED socket that is about to |
363 | * be sent. | |
364 | */ | |
735d3831 | 365 | static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb, |
ea1627c2 | 366 | struct tcphdr *th, int tcp_header_len) |
bdf1ee5d IJ |
367 | { |
368 | struct tcp_sock *tp = tcp_sk(sk); | |
369 | ||
370 | if (tp->ecn_flags & TCP_ECN_OK) { | |
371 | /* Not-retransmitted data segment: set ECT and inject CWR. */ | |
372 | if (skb->len != tcp_header_len && | |
373 | !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) { | |
374 | INET_ECN_xmit(sk); | |
056834d9 | 375 | if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) { |
bdf1ee5d | 376 | tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR; |
ea1627c2 | 377 | th->cwr = 1; |
bdf1ee5d IJ |
378 | skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN; |
379 | } | |
30e502a3 | 380 | } else if (!tcp_ca_needs_ecn(sk)) { |
bdf1ee5d IJ |
381 | /* ACK or retransmitted segment: clear ECT|CE */ |
382 | INET_ECN_dontxmit(sk); | |
383 | } | |
384 | if (tp->ecn_flags & TCP_ECN_DEMAND_CWR) | |
ea1627c2 | 385 | th->ece = 1; |
bdf1ee5d IJ |
386 | } |
387 | } | |
388 | ||
e870a8ef IJ |
389 | /* Constructs common control bits of non-data skb. If SYN/FIN is present, |
390 | * auto increment end seqno. | |
391 | */ | |
392 | static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags) | |
393 | { | |
2e8e18ef | 394 | skb->ip_summed = CHECKSUM_PARTIAL; |
e870a8ef | 395 | |
4de075e0 | 396 | TCP_SKB_CB(skb)->tcp_flags = flags; |
e870a8ef IJ |
397 | TCP_SKB_CB(skb)->sacked = 0; |
398 | ||
cd7d8498 | 399 | tcp_skb_pcount_set(skb, 1); |
e870a8ef IJ |
400 | |
401 | TCP_SKB_CB(skb)->seq = seq; | |
a3433f35 | 402 | if (flags & (TCPHDR_SYN | TCPHDR_FIN)) |
e870a8ef IJ |
403 | seq++; |
404 | TCP_SKB_CB(skb)->end_seq = seq; | |
405 | } | |
406 | ||
a2a385d6 | 407 | static inline bool tcp_urg_mode(const struct tcp_sock *tp) |
33f5f57e IJ |
408 | { |
409 | return tp->snd_una != tp->snd_up; | |
410 | } | |
411 | ||
33ad798c AL |
412 | #define OPTION_SACK_ADVERTISE (1 << 0) |
413 | #define OPTION_TS (1 << 1) | |
414 | #define OPTION_MD5 (1 << 2) | |
89e95a61 | 415 | #define OPTION_WSCALE (1 << 3) |
2100c8d2 | 416 | #define OPTION_FAST_OPEN_COOKIE (1 << 8) |
60e2a778 | 417 | #define OPTION_SMC (1 << 9) |
eda7acdd | 418 | #define OPTION_MPTCP (1 << 10) |
60e2a778 UB |
419 | |
420 | static void smc_options_write(__be32 *ptr, u16 *options) | |
421 | { | |
422 | #if IS_ENABLED(CONFIG_SMC) | |
423 | if (static_branch_unlikely(&tcp_have_smc)) { | |
424 | if (unlikely(OPTION_SMC & *options)) { | |
425 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
426 | (TCPOPT_NOP << 16) | | |
427 | (TCPOPT_EXP << 8) | | |
428 | (TCPOLEN_EXP_SMC_BASE)); | |
429 | *ptr++ = htonl(TCPOPT_SMC_MAGIC); | |
430 | } | |
431 | } | |
432 | #endif | |
433 | } | |
33ad798c AL |
434 | |
435 | struct tcp_out_options { | |
2100c8d2 YC |
436 | u16 options; /* bit field of OPTION_* */ |
437 | u16 mss; /* 0 to disable */ | |
33ad798c AL |
438 | u8 ws; /* window scale, 0 to disable */ |
439 | u8 num_sack_blocks; /* number of SACK blocks to include */ | |
bd0388ae | 440 | u8 hash_size; /* bytes in hash_location */ |
331fca43 | 441 | u8 bpf_opt_len; /* length of BPF hdr option */ |
bd0388ae | 442 | __u8 *hash_location; /* temporary pointer, overloaded */ |
2100c8d2 YC |
443 | __u32 tsval, tsecr; /* need to include OPTION_TS */ |
444 | struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ | |
eda7acdd | 445 | struct mptcp_out_options mptcp; |
33ad798c AL |
446 | }; |
447 | ||
fa3fe2b1 FW |
448 | static void mptcp_options_write(__be32 *ptr, const struct tcp_sock *tp, |
449 | struct tcp_out_options *opts) | |
eda7acdd PK |
450 | { |
451 | #if IS_ENABLED(CONFIG_MPTCP) | |
452 | if (unlikely(OPTION_MPTCP & opts->options)) | |
fa3fe2b1 | 453 | mptcp_write_options(ptr, tp, &opts->mptcp); |
eda7acdd PK |
454 | #endif |
455 | } | |
456 | ||
331fca43 | 457 | #ifdef CONFIG_CGROUP_BPF |
0813a841 MKL |
458 | static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb, |
459 | enum tcp_synack_type synack_type) | |
460 | { | |
461 | if (unlikely(!skb)) | |
462 | return BPF_WRITE_HDR_TCP_CURRENT_MSS; | |
463 | ||
464 | if (unlikely(synack_type == TCP_SYNACK_COOKIE)) | |
465 | return BPF_WRITE_HDR_TCP_SYNACK_COOKIE; | |
466 | ||
467 | return 0; | |
468 | } | |
469 | ||
331fca43 MKL |
470 | /* req, syn_skb and synack_type are used when writing synack */ |
471 | static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb, | |
472 | struct request_sock *req, | |
473 | struct sk_buff *syn_skb, | |
474 | enum tcp_synack_type synack_type, | |
475 | struct tcp_out_options *opts, | |
476 | unsigned int *remaining) | |
477 | { | |
0813a841 MKL |
478 | struct bpf_sock_ops_kern sock_ops; |
479 | int err; | |
480 | ||
331fca43 MKL |
481 | if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk), |
482 | BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) || | |
483 | !*remaining) | |
484 | return; | |
485 | ||
0813a841 MKL |
486 | /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */ |
487 | ||
488 | /* init sock_ops */ | |
489 | memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); | |
490 | ||
491 | sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB; | |
492 | ||
493 | if (req) { | |
494 | /* The listen "sk" cannot be passed here because | |
495 | * it is not locked. It would not make too much | |
496 | * sense to do bpf_setsockopt(listen_sk) based | |
497 | * on individual connection request also. | |
498 | * | |
499 | * Thus, "req" is passed here and the cgroup-bpf-progs | |
500 | * of the listen "sk" will be run. | |
501 | * | |
502 | * "req" is also used here for fastopen even the "sk" here is | |
503 | * a fullsock "child" sk. It is to keep the behavior | |
504 | * consistent between fastopen and non-fastopen on | |
505 | * the bpf programming side. | |
506 | */ | |
507 | sock_ops.sk = (struct sock *)req; | |
508 | sock_ops.syn_skb = syn_skb; | |
509 | } else { | |
510 | sock_owned_by_me(sk); | |
511 | ||
512 | sock_ops.is_fullsock = 1; | |
513 | sock_ops.sk = sk; | |
514 | } | |
515 | ||
516 | sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type); | |
517 | sock_ops.remaining_opt_len = *remaining; | |
518 | /* tcp_current_mss() does not pass a skb */ | |
519 | if (skb) | |
520 | bpf_skops_init_skb(&sock_ops, skb, 0); | |
521 | ||
522 | err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk); | |
523 | ||
524 | if (err || sock_ops.remaining_opt_len == *remaining) | |
525 | return; | |
526 | ||
527 | opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len; | |
528 | /* round up to 4 bytes */ | |
529 | opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3; | |
530 | ||
531 | *remaining -= opts->bpf_opt_len; | |
331fca43 MKL |
532 | } |
533 | ||
534 | static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb, | |
535 | struct request_sock *req, | |
536 | struct sk_buff *syn_skb, | |
537 | enum tcp_synack_type synack_type, | |
538 | struct tcp_out_options *opts) | |
539 | { | |
0813a841 MKL |
540 | u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len; |
541 | struct bpf_sock_ops_kern sock_ops; | |
542 | int err; | |
543 | ||
544 | if (likely(!max_opt_len)) | |
331fca43 MKL |
545 | return; |
546 | ||
0813a841 MKL |
547 | memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp)); |
548 | ||
549 | sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB; | |
550 | ||
551 | if (req) { | |
552 | sock_ops.sk = (struct sock *)req; | |
553 | sock_ops.syn_skb = syn_skb; | |
554 | } else { | |
555 | sock_owned_by_me(sk); | |
556 | ||
557 | sock_ops.is_fullsock = 1; | |
558 | sock_ops.sk = sk; | |
559 | } | |
560 | ||
561 | sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type); | |
562 | sock_ops.remaining_opt_len = max_opt_len; | |
563 | first_opt_off = tcp_hdrlen(skb) - max_opt_len; | |
564 | bpf_skops_init_skb(&sock_ops, skb, first_opt_off); | |
565 | ||
566 | err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk); | |
567 | ||
568 | if (err) | |
569 | nr_written = 0; | |
570 | else | |
571 | nr_written = max_opt_len - sock_ops.remaining_opt_len; | |
572 | ||
573 | if (nr_written < max_opt_len) | |
574 | memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP, | |
575 | max_opt_len - nr_written); | |
331fca43 MKL |
576 | } |
577 | #else | |
578 | static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb, | |
579 | struct request_sock *req, | |
580 | struct sk_buff *syn_skb, | |
581 | enum tcp_synack_type synack_type, | |
582 | struct tcp_out_options *opts, | |
583 | unsigned int *remaining) | |
584 | { | |
585 | } | |
586 | ||
587 | static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb, | |
588 | struct request_sock *req, | |
589 | struct sk_buff *syn_skb, | |
590 | enum tcp_synack_type synack_type, | |
591 | struct tcp_out_options *opts) | |
592 | { | |
593 | } | |
594 | #endif | |
595 | ||
67edfef7 AK |
596 | /* Write previously computed TCP options to the packet. |
597 | * | |
598 | * Beware: Something in the Internet is very sensitive to the ordering of | |
fd6149d3 IJ |
599 | * TCP options, we learned this through the hard way, so be careful here. |
600 | * Luckily we can at least blame others for their non-compliance but from | |
8e3bff96 | 601 | * inter-operability perspective it seems that we're somewhat stuck with |
fd6149d3 IJ |
602 | * the ordering which we have been using if we want to keep working with |
603 | * those broken things (not that it currently hurts anybody as there isn't | |
604 | * particular reason why the ordering would need to be changed). | |
605 | * | |
606 | * At least SACK_PERM as the first option is known to lead to a disaster | |
607 | * (but it may well be that other scenarios fail similarly). | |
608 | */ | |
33ad798c | 609 | static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, |
bd0388ae WAS |
610 | struct tcp_out_options *opts) |
611 | { | |
2100c8d2 | 612 | u16 options = opts->options; /* mungable copy */ |
bd0388ae | 613 | |
bd0388ae | 614 | if (unlikely(OPTION_MD5 & options)) { |
1a2c6181 CP |
615 | *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | |
616 | (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); | |
bd0388ae WAS |
617 | /* overload cookie hash location */ |
618 | opts->hash_location = (__u8 *)ptr; | |
33ad798c | 619 | ptr += 4; |
40efc6fa | 620 | } |
33ad798c | 621 | |
fd6149d3 IJ |
622 | if (unlikely(opts->mss)) { |
623 | *ptr++ = htonl((TCPOPT_MSS << 24) | | |
624 | (TCPOLEN_MSS << 16) | | |
625 | opts->mss); | |
626 | } | |
627 | ||
bd0388ae WAS |
628 | if (likely(OPTION_TS & options)) { |
629 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { | |
33ad798c AL |
630 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
631 | (TCPOLEN_SACK_PERM << 16) | | |
632 | (TCPOPT_TIMESTAMP << 8) | | |
633 | TCPOLEN_TIMESTAMP); | |
bd0388ae | 634 | options &= ~OPTION_SACK_ADVERTISE; |
33ad798c AL |
635 | } else { |
636 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
637 | (TCPOPT_NOP << 16) | | |
638 | (TCPOPT_TIMESTAMP << 8) | | |
639 | TCPOLEN_TIMESTAMP); | |
640 | } | |
641 | *ptr++ = htonl(opts->tsval); | |
642 | *ptr++ = htonl(opts->tsecr); | |
643 | } | |
644 | ||
bd0388ae | 645 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { |
33ad798c AL |
646 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
647 | (TCPOPT_NOP << 16) | | |
648 | (TCPOPT_SACK_PERM << 8) | | |
649 | TCPOLEN_SACK_PERM); | |
650 | } | |
651 | ||
bd0388ae | 652 | if (unlikely(OPTION_WSCALE & options)) { |
33ad798c AL |
653 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
654 | (TCPOPT_WINDOW << 16) | | |
655 | (TCPOLEN_WINDOW << 8) | | |
656 | opts->ws); | |
657 | } | |
658 | ||
659 | if (unlikely(opts->num_sack_blocks)) { | |
660 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? | |
661 | tp->duplicate_sack : tp->selective_acks; | |
40efc6fa SH |
662 | int this_sack; |
663 | ||
664 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
665 | (TCPOPT_NOP << 16) | | |
666 | (TCPOPT_SACK << 8) | | |
33ad798c | 667 | (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * |
40efc6fa | 668 | TCPOLEN_SACK_PERBLOCK))); |
2de979bd | 669 | |
33ad798c AL |
670 | for (this_sack = 0; this_sack < opts->num_sack_blocks; |
671 | ++this_sack) { | |
40efc6fa SH |
672 | *ptr++ = htonl(sp[this_sack].start_seq); |
673 | *ptr++ = htonl(sp[this_sack].end_seq); | |
674 | } | |
2de979bd | 675 | |
5861f8e5 | 676 | tp->rx_opt.dsack = 0; |
40efc6fa | 677 | } |
2100c8d2 YC |
678 | |
679 | if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { | |
680 | struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; | |
7f9b838b DL |
681 | u8 *p = (u8 *)ptr; |
682 | u32 len; /* Fast Open option length */ | |
683 | ||
684 | if (foc->exp) { | |
685 | len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; | |
686 | *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | | |
687 | TCPOPT_FASTOPEN_MAGIC); | |
688 | p += TCPOLEN_EXP_FASTOPEN_BASE; | |
689 | } else { | |
690 | len = TCPOLEN_FASTOPEN_BASE + foc->len; | |
691 | *p++ = TCPOPT_FASTOPEN; | |
692 | *p++ = len; | |
693 | } | |
2100c8d2 | 694 | |
7f9b838b DL |
695 | memcpy(p, foc->val, foc->len); |
696 | if ((len & 3) == 2) { | |
697 | p[foc->len] = TCPOPT_NOP; | |
698 | p[foc->len + 1] = TCPOPT_NOP; | |
2100c8d2 | 699 | } |
7f9b838b | 700 | ptr += (len + 3) >> 2; |
2100c8d2 | 701 | } |
60e2a778 UB |
702 | |
703 | smc_options_write(ptr, &options); | |
eda7acdd | 704 | |
fa3fe2b1 | 705 | mptcp_options_write(ptr, tp, opts); |
60e2a778 UB |
706 | } |
707 | ||
708 | static void smc_set_option(const struct tcp_sock *tp, | |
709 | struct tcp_out_options *opts, | |
710 | unsigned int *remaining) | |
711 | { | |
712 | #if IS_ENABLED(CONFIG_SMC) | |
713 | if (static_branch_unlikely(&tcp_have_smc)) { | |
714 | if (tp->syn_smc) { | |
715 | if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { | |
716 | opts->options |= OPTION_SMC; | |
717 | *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; | |
718 | } | |
719 | } | |
720 | } | |
721 | #endif | |
722 | } | |
723 | ||
724 | static void smc_set_option_cond(const struct tcp_sock *tp, | |
725 | const struct inet_request_sock *ireq, | |
726 | struct tcp_out_options *opts, | |
727 | unsigned int *remaining) | |
728 | { | |
729 | #if IS_ENABLED(CONFIG_SMC) | |
730 | if (static_branch_unlikely(&tcp_have_smc)) { | |
731 | if (tp->syn_smc && ireq->smc_ok) { | |
732 | if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { | |
733 | opts->options |= OPTION_SMC; | |
734 | *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; | |
735 | } | |
736 | } | |
737 | } | |
738 | #endif | |
33ad798c AL |
739 | } |
740 | ||
cec37a6e PK |
741 | static void mptcp_set_option_cond(const struct request_sock *req, |
742 | struct tcp_out_options *opts, | |
743 | unsigned int *remaining) | |
744 | { | |
745 | if (rsk_is_mptcp(req)) { | |
746 | unsigned int size; | |
747 | ||
748 | if (mptcp_synack_options(req, &size, &opts->mptcp)) { | |
749 | if (*remaining >= size) { | |
750 | opts->options |= OPTION_MPTCP; | |
751 | *remaining -= size; | |
752 | } | |
753 | } | |
754 | } | |
755 | } | |
756 | ||
67edfef7 AK |
757 | /* Compute TCP options for SYN packets. This is not the final |
758 | * network wire format yet. | |
759 | */ | |
95c96174 | 760 | static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, |
33ad798c | 761 | struct tcp_out_options *opts, |
cf533ea5 ED |
762 | struct tcp_md5sig_key **md5) |
763 | { | |
33ad798c | 764 | struct tcp_sock *tp = tcp_sk(sk); |
95c96174 | 765 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
783237e8 | 766 | struct tcp_fastopen_request *fastopen = tp->fastopen_req; |
33ad798c | 767 | |
8c2320e8 | 768 | *md5 = NULL; |
cfb6eeb4 | 769 | #ifdef CONFIG_TCP_MD5SIG |
921f9a0f | 770 | if (static_branch_unlikely(&tcp_md5_needed) && |
6015c71e | 771 | rcu_access_pointer(tp->md5sig_info)) { |
8c2320e8 ED |
772 | *md5 = tp->af_specific->md5_lookup(sk, sk); |
773 | if (*md5) { | |
774 | opts->options |= OPTION_MD5; | |
775 | remaining -= TCPOLEN_MD5SIG_ALIGNED; | |
776 | } | |
cfb6eeb4 YH |
777 | } |
778 | #endif | |
33ad798c AL |
779 | |
780 | /* We always get an MSS option. The option bytes which will be seen in | |
781 | * normal data packets should timestamps be used, must be in the MSS | |
782 | * advertised. But we subtract them from tp->mss_cache so that | |
783 | * calculations in tcp_sendmsg are simpler etc. So account for this | |
784 | * fact here if necessary. If we don't do this correctly, as a | |
785 | * receiver we won't recognize data packets as being full sized when we | |
786 | * should, and thus we won't abide by the delayed ACK rules correctly. | |
787 | * SACKs don't matter, we never delay an ACK when we have any of those | |
788 | * going out. */ | |
789 | opts->mss = tcp_advertise_mss(sk); | |
bd0388ae | 790 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c | 791 | |
5d2ed052 | 792 | if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) { |
33ad798c | 793 | opts->options |= OPTION_TS; |
7faee5c0 | 794 | opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; |
33ad798c | 795 | opts->tsecr = tp->rx_opt.ts_recent; |
bd0388ae | 796 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c | 797 | } |
9bb37ef0 | 798 | if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) { |
33ad798c | 799 | opts->ws = tp->rx_opt.rcv_wscale; |
89e95a61 | 800 | opts->options |= OPTION_WSCALE; |
bd0388ae | 801 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 802 | } |
f9301034 | 803 | if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) { |
33ad798c | 804 | opts->options |= OPTION_SACK_ADVERTISE; |
b32d1310 | 805 | if (unlikely(!(OPTION_TS & opts->options))) |
bd0388ae | 806 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c AL |
807 | } |
808 | ||
783237e8 | 809 | if (fastopen && fastopen->cookie.len >= 0) { |
2646c831 DL |
810 | u32 need = fastopen->cookie.len; |
811 | ||
812 | need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : | |
813 | TCPOLEN_FASTOPEN_BASE; | |
783237e8 YC |
814 | need = (need + 3) & ~3U; /* Align to 32 bits */ |
815 | if (remaining >= need) { | |
816 | opts->options |= OPTION_FAST_OPEN_COOKIE; | |
817 | opts->fastopen_cookie = &fastopen->cookie; | |
818 | remaining -= need; | |
819 | tp->syn_fastopen = 1; | |
2646c831 | 820 | tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; |
783237e8 YC |
821 | } |
822 | } | |
bd0388ae | 823 | |
60e2a778 UB |
824 | smc_set_option(tp, opts, &remaining); |
825 | ||
cec37a6e PK |
826 | if (sk_is_mptcp(sk)) { |
827 | unsigned int size; | |
828 | ||
cc7972ea | 829 | if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) { |
cec37a6e PK |
830 | opts->options |= OPTION_MPTCP; |
831 | remaining -= size; | |
832 | } | |
833 | } | |
834 | ||
331fca43 MKL |
835 | bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining); |
836 | ||
bd0388ae | 837 | return MAX_TCP_OPTION_SPACE - remaining; |
40efc6fa SH |
838 | } |
839 | ||
67edfef7 | 840 | /* Set up TCP options for SYN-ACKs. */ |
60e2a778 UB |
841 | static unsigned int tcp_synack_options(const struct sock *sk, |
842 | struct request_sock *req, | |
37bfbdda ED |
843 | unsigned int mss, struct sk_buff *skb, |
844 | struct tcp_out_options *opts, | |
845 | const struct tcp_md5sig_key *md5, | |
e114e1e8 | 846 | struct tcp_fastopen_cookie *foc, |
331fca43 MKL |
847 | enum tcp_synack_type synack_type, |
848 | struct sk_buff *syn_skb) | |
4957faad | 849 | { |
33ad798c | 850 | struct inet_request_sock *ireq = inet_rsk(req); |
95c96174 | 851 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
33ad798c | 852 | |
cfb6eeb4 | 853 | #ifdef CONFIG_TCP_MD5SIG |
80f03e27 | 854 | if (md5) { |
33ad798c | 855 | opts->options |= OPTION_MD5; |
4957faad WAS |
856 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
857 | ||
858 | /* We can't fit any SACK blocks in a packet with MD5 + TS | |
859 | * options. There was discussion about disabling SACK | |
860 | * rather than TS in order to fit in better with old, | |
861 | * buggy kernels, but that was deemed to be unnecessary. | |
862 | */ | |
e114e1e8 ED |
863 | if (synack_type != TCP_SYNACK_COOKIE) |
864 | ireq->tstamp_ok &= !ireq->sack_ok; | |
cfb6eeb4 YH |
865 | } |
866 | #endif | |
33ad798c | 867 | |
4957faad | 868 | /* We always send an MSS option. */ |
33ad798c | 869 | opts->mss = mss; |
4957faad | 870 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c AL |
871 | |
872 | if (likely(ireq->wscale_ok)) { | |
873 | opts->ws = ireq->rcv_wscale; | |
89e95a61 | 874 | opts->options |= OPTION_WSCALE; |
4957faad | 875 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 876 | } |
de213e5e | 877 | if (likely(ireq->tstamp_ok)) { |
33ad798c | 878 | opts->options |= OPTION_TS; |
95a22cae | 879 | opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; |
33ad798c | 880 | opts->tsecr = req->ts_recent; |
4957faad | 881 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c AL |
882 | } |
883 | if (likely(ireq->sack_ok)) { | |
884 | opts->options |= OPTION_SACK_ADVERTISE; | |
de213e5e | 885 | if (unlikely(!ireq->tstamp_ok)) |
4957faad | 886 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c | 887 | } |
7f9b838b DL |
888 | if (foc != NULL && foc->len >= 0) { |
889 | u32 need = foc->len; | |
890 | ||
891 | need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : | |
892 | TCPOLEN_FASTOPEN_BASE; | |
8336886f JC |
893 | need = (need + 3) & ~3U; /* Align to 32 bits */ |
894 | if (remaining >= need) { | |
895 | opts->options |= OPTION_FAST_OPEN_COOKIE; | |
896 | opts->fastopen_cookie = foc; | |
897 | remaining -= need; | |
898 | } | |
899 | } | |
1a2c6181 | 900 | |
cec37a6e PK |
901 | mptcp_set_option_cond(req, opts, &remaining); |
902 | ||
60e2a778 UB |
903 | smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining); |
904 | ||
331fca43 MKL |
905 | bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb, |
906 | synack_type, opts, &remaining); | |
907 | ||
4957faad | 908 | return MAX_TCP_OPTION_SPACE - remaining; |
33ad798c AL |
909 | } |
910 | ||
67edfef7 AK |
911 | /* Compute TCP options for ESTABLISHED sockets. This is not the |
912 | * final wire format yet. | |
913 | */ | |
95c96174 | 914 | static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, |
33ad798c | 915 | struct tcp_out_options *opts, |
cf533ea5 ED |
916 | struct tcp_md5sig_key **md5) |
917 | { | |
33ad798c | 918 | struct tcp_sock *tp = tcp_sk(sk); |
95c96174 | 919 | unsigned int size = 0; |
cabeccbd | 920 | unsigned int eff_sacks; |
33ad798c | 921 | |
5843ef42 AK |
922 | opts->options = 0; |
923 | ||
8c2320e8 | 924 | *md5 = NULL; |
33ad798c | 925 | #ifdef CONFIG_TCP_MD5SIG |
921f9a0f | 926 | if (static_branch_unlikely(&tcp_md5_needed) && |
6015c71e | 927 | rcu_access_pointer(tp->md5sig_info)) { |
8c2320e8 ED |
928 | *md5 = tp->af_specific->md5_lookup(sk, sk); |
929 | if (*md5) { | |
930 | opts->options |= OPTION_MD5; | |
931 | size += TCPOLEN_MD5SIG_ALIGNED; | |
932 | } | |
33ad798c | 933 | } |
33ad798c AL |
934 | #endif |
935 | ||
936 | if (likely(tp->rx_opt.tstamp_ok)) { | |
937 | opts->options |= OPTION_TS; | |
7faee5c0 | 938 | opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; |
33ad798c AL |
939 | opts->tsecr = tp->rx_opt.ts_recent; |
940 | size += TCPOLEN_TSTAMP_ALIGNED; | |
941 | } | |
942 | ||
cec37a6e PK |
943 | /* MPTCP options have precedence over SACK for the limited TCP |
944 | * option space because a MPTCP connection would be forced to | |
945 | * fall back to regular TCP if a required multipath option is | |
946 | * missing. SACK still gets a chance to use whatever space is | |
947 | * left. | |
948 | */ | |
949 | if (sk_is_mptcp(sk)) { | |
950 | unsigned int remaining = MAX_TCP_OPTION_SPACE - size; | |
951 | unsigned int opt_size = 0; | |
952 | ||
953 | if (mptcp_established_options(sk, skb, &opt_size, remaining, | |
954 | &opts->mptcp)) { | |
955 | opts->options |= OPTION_MPTCP; | |
956 | size += opt_size; | |
957 | } | |
958 | } | |
959 | ||
cabeccbd IJ |
960 | eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; |
961 | if (unlikely(eff_sacks)) { | |
95c96174 | 962 | const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; |
9cfcca23 MM |
963 | if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED + |
964 | TCPOLEN_SACK_PERBLOCK)) | |
965 | return size; | |
966 | ||
33ad798c | 967 | opts->num_sack_blocks = |
95c96174 | 968 | min_t(unsigned int, eff_sacks, |
33ad798c AL |
969 | (remaining - TCPOLEN_SACK_BASE_ALIGNED) / |
970 | TCPOLEN_SACK_PERBLOCK); | |
9cfcca23 MM |
971 | |
972 | size += TCPOLEN_SACK_BASE_ALIGNED + | |
973 | opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; | |
33ad798c AL |
974 | } |
975 | ||
331fca43 MKL |
976 | if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp, |
977 | BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) { | |
978 | unsigned int remaining = MAX_TCP_OPTION_SPACE - size; | |
979 | ||
980 | bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining); | |
981 | ||
982 | size = MAX_TCP_OPTION_SPACE - remaining; | |
983 | } | |
984 | ||
33ad798c | 985 | return size; |
40efc6fa | 986 | } |
1da177e4 | 987 | |
46d3ceab ED |
988 | |
989 | /* TCP SMALL QUEUES (TSQ) | |
990 | * | |
991 | * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) | |
992 | * to reduce RTT and bufferbloat. | |
993 | * We do this using a special skb destructor (tcp_wfree). | |
994 | * | |
995 | * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb | |
996 | * needs to be reallocated in a driver. | |
8e3bff96 | 997 | * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc |
46d3ceab ED |
998 | * |
999 | * Since transmit from skb destructor is forbidden, we use a tasklet | |
1000 | * to process all sockets that eventually need to send more skbs. | |
1001 | * We use one tasklet per cpu, with its own queue of sockets. | |
1002 | */ | |
1003 | struct tsq_tasklet { | |
1004 | struct tasklet_struct tasklet; | |
1005 | struct list_head head; /* queue of tcp sockets */ | |
1006 | }; | |
1007 | static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); | |
1008 | ||
73a6bab5 | 1009 | static void tcp_tsq_write(struct sock *sk) |
6f458dfb ED |
1010 | { |
1011 | if ((1 << sk->sk_state) & | |
1012 | (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | | |
f9616c35 ED |
1013 | TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { |
1014 | struct tcp_sock *tp = tcp_sk(sk); | |
1015 | ||
1016 | if (tp->lost_out > tp->retrans_out && | |
3a91d29f KD |
1017 | tp->snd_cwnd > tcp_packets_in_flight(tp)) { |
1018 | tcp_mstamp_refresh(tp); | |
f9616c35 | 1019 | tcp_xmit_retransmit_queue(sk); |
3a91d29f | 1020 | } |
f9616c35 ED |
1021 | |
1022 | tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, | |
bf06200e | 1023 | 0, GFP_ATOMIC); |
f9616c35 | 1024 | } |
6f458dfb | 1025 | } |
73a6bab5 ED |
1026 | |
1027 | static void tcp_tsq_handler(struct sock *sk) | |
1028 | { | |
1029 | bh_lock_sock(sk); | |
1030 | if (!sock_owned_by_user(sk)) | |
1031 | tcp_tsq_write(sk); | |
1032 | else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) | |
1033 | sock_hold(sk); | |
1034 | bh_unlock_sock(sk); | |
1035 | } | |
46d3ceab | 1036 | /* |
8e3bff96 | 1037 | * One tasklet per cpu tries to send more skbs. |
46d3ceab | 1038 | * We run in tasklet context but need to disable irqs when |
8e3bff96 | 1039 | * transferring tsq->head because tcp_wfree() might |
46d3ceab ED |
1040 | * interrupt us (non NAPI drivers) |
1041 | */ | |
c6533ca8 | 1042 | static void tcp_tasklet_func(struct tasklet_struct *t) |
46d3ceab | 1043 | { |
c6533ca8 | 1044 | struct tsq_tasklet *tsq = from_tasklet(tsq, t, tasklet); |
46d3ceab ED |
1045 | LIST_HEAD(list); |
1046 | unsigned long flags; | |
1047 | struct list_head *q, *n; | |
1048 | struct tcp_sock *tp; | |
1049 | struct sock *sk; | |
1050 | ||
1051 | local_irq_save(flags); | |
1052 | list_splice_init(&tsq->head, &list); | |
1053 | local_irq_restore(flags); | |
1054 | ||
1055 | list_for_each_safe(q, n, &list) { | |
1056 | tp = list_entry(q, struct tcp_sock, tsq_node); | |
1057 | list_del(&tp->tsq_node); | |
1058 | ||
1059 | sk = (struct sock *)tp; | |
0a9648f1 | 1060 | smp_mb__before_atomic(); |
7aa5470c ED |
1061 | clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); |
1062 | ||
73a6bab5 | 1063 | tcp_tsq_handler(sk); |
46d3ceab ED |
1064 | sk_free(sk); |
1065 | } | |
1066 | } | |
1067 | ||
40fc3423 ED |
1068 | #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ |
1069 | TCPF_WRITE_TIMER_DEFERRED | \ | |
1070 | TCPF_DELACK_TIMER_DEFERRED | \ | |
1071 | TCPF_MTU_REDUCED_DEFERRED) | |
46d3ceab ED |
1072 | /** |
1073 | * tcp_release_cb - tcp release_sock() callback | |
1074 | * @sk: socket | |
1075 | * | |
1076 | * called from release_sock() to perform protocol dependent | |
1077 | * actions before socket release. | |
1078 | */ | |
1079 | void tcp_release_cb(struct sock *sk) | |
1080 | { | |
6f458dfb | 1081 | unsigned long flags, nflags; |
46d3ceab | 1082 | |
6f458dfb ED |
1083 | /* perform an atomic operation only if at least one flag is set */ |
1084 | do { | |
7aa5470c | 1085 | flags = sk->sk_tsq_flags; |
6f458dfb ED |
1086 | if (!(flags & TCP_DEFERRED_ALL)) |
1087 | return; | |
1088 | nflags = flags & ~TCP_DEFERRED_ALL; | |
7aa5470c | 1089 | } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); |
6f458dfb | 1090 | |
73a6bab5 ED |
1091 | if (flags & TCPF_TSQ_DEFERRED) { |
1092 | tcp_tsq_write(sk); | |
1093 | __sock_put(sk); | |
1094 | } | |
c3f9b018 ED |
1095 | /* Here begins the tricky part : |
1096 | * We are called from release_sock() with : | |
1097 | * 1) BH disabled | |
1098 | * 2) sk_lock.slock spinlock held | |
1099 | * 3) socket owned by us (sk->sk_lock.owned == 1) | |
1100 | * | |
1101 | * But following code is meant to be called from BH handlers, | |
1102 | * so we should keep BH disabled, but early release socket ownership | |
1103 | */ | |
1104 | sock_release_ownership(sk); | |
1105 | ||
40fc3423 | 1106 | if (flags & TCPF_WRITE_TIMER_DEFERRED) { |
6f458dfb | 1107 | tcp_write_timer_handler(sk); |
144d56e9 ED |
1108 | __sock_put(sk); |
1109 | } | |
40fc3423 | 1110 | if (flags & TCPF_DELACK_TIMER_DEFERRED) { |
6f458dfb | 1111 | tcp_delack_timer_handler(sk); |
144d56e9 ED |
1112 | __sock_put(sk); |
1113 | } | |
40fc3423 | 1114 | if (flags & TCPF_MTU_REDUCED_DEFERRED) { |
4fab9071 | 1115 | inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); |
144d56e9 ED |
1116 | __sock_put(sk); |
1117 | } | |
46d3ceab ED |
1118 | } |
1119 | EXPORT_SYMBOL(tcp_release_cb); | |
1120 | ||
1121 | void __init tcp_tasklet_init(void) | |
1122 | { | |
1123 | int i; | |
1124 | ||
1125 | for_each_possible_cpu(i) { | |
1126 | struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); | |
1127 | ||
1128 | INIT_LIST_HEAD(&tsq->head); | |
c6533ca8 | 1129 | tasklet_setup(&tsq->tasklet, tcp_tasklet_func); |
46d3ceab ED |
1130 | } |
1131 | } | |
1132 | ||
1133 | /* | |
1134 | * Write buffer destructor automatically called from kfree_skb. | |
8e3bff96 | 1135 | * We can't xmit new skbs from this context, as we might already |
46d3ceab ED |
1136 | * hold qdisc lock. |
1137 | */ | |
d6a4a104 | 1138 | void tcp_wfree(struct sk_buff *skb) |
46d3ceab ED |
1139 | { |
1140 | struct sock *sk = skb->sk; | |
1141 | struct tcp_sock *tp = tcp_sk(sk); | |
408f0a6c | 1142 | unsigned long flags, nval, oval; |
9b462d02 ED |
1143 | |
1144 | /* Keep one reference on sk_wmem_alloc. | |
1145 | * Will be released by sk_free() from here or tcp_tasklet_func() | |
1146 | */ | |
14afee4b | 1147 | WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); |
9b462d02 ED |
1148 | |
1149 | /* If this softirq is serviced by ksoftirqd, we are likely under stress. | |
1150 | * Wait until our queues (qdisc + devices) are drained. | |
1151 | * This gives : | |
1152 | * - less callbacks to tcp_write_xmit(), reducing stress (batches) | |
1153 | * - chance for incoming ACK (processed by another cpu maybe) | |
1154 | * to migrate this flow (skb->ooo_okay will be eventually set) | |
1155 | */ | |
14afee4b | 1156 | if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) |
9b462d02 | 1157 | goto out; |
46d3ceab | 1158 | |
7aa5470c | 1159 | for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { |
46d3ceab | 1160 | struct tsq_tasklet *tsq; |
a9b204d1 | 1161 | bool empty; |
46d3ceab | 1162 | |
408f0a6c ED |
1163 | if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) |
1164 | goto out; | |
1165 | ||
73a6bab5 | 1166 | nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED; |
7aa5470c | 1167 | nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); |
408f0a6c ED |
1168 | if (nval != oval) |
1169 | continue; | |
1170 | ||
46d3ceab ED |
1171 | /* queue this socket to tasklet queue */ |
1172 | local_irq_save(flags); | |
903ceff7 | 1173 | tsq = this_cpu_ptr(&tsq_tasklet); |
a9b204d1 | 1174 | empty = list_empty(&tsq->head); |
46d3ceab | 1175 | list_add(&tp->tsq_node, &tsq->head); |
a9b204d1 ED |
1176 | if (empty) |
1177 | tasklet_schedule(&tsq->tasklet); | |
46d3ceab | 1178 | local_irq_restore(flags); |
9b462d02 | 1179 | return; |
46d3ceab | 1180 | } |
9b462d02 ED |
1181 | out: |
1182 | sk_free(sk); | |
46d3ceab ED |
1183 | } |
1184 | ||
73a6bab5 ED |
1185 | /* Note: Called under soft irq. |
1186 | * We can call TCP stack right away, unless socket is owned by user. | |
218af599 ED |
1187 | */ |
1188 | enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) | |
1189 | { | |
1190 | struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); | |
1191 | struct sock *sk = (struct sock *)tp; | |
218af599 | 1192 | |
73a6bab5 ED |
1193 | tcp_tsq_handler(sk); |
1194 | sock_put(sk); | |
218af599 | 1195 | |
218af599 ED |
1196 | return HRTIMER_NORESTART; |
1197 | } | |
1198 | ||
a7a25630 ED |
1199 | static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb, |
1200 | u64 prior_wstamp) | |
e2080072 | 1201 | { |
ab408b6d ED |
1202 | struct tcp_sock *tp = tcp_sk(sk); |
1203 | ||
ab408b6d | 1204 | if (sk->sk_pacing_status != SK_PACING_NONE) { |
76a9ebe8 | 1205 | unsigned long rate = sk->sk_pacing_rate; |
ab408b6d ED |
1206 | |
1207 | /* Original sch_fq does not pace first 10 MSS | |
1208 | * Note that tp->data_segs_out overflows after 2^32 packets, | |
1209 | * this is a minor annoyance. | |
1210 | */ | |
76a9ebe8 | 1211 | if (rate != ~0UL && rate && tp->data_segs_out >= 10) { |
a7a25630 ED |
1212 | u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate); |
1213 | u64 credit = tp->tcp_wstamp_ns - prior_wstamp; | |
1214 | ||
1215 | /* take into account OS jitter */ | |
1216 | len_ns -= min_t(u64, len_ns / 2, credit); | |
1217 | tp->tcp_wstamp_ns += len_ns; | |
ab408b6d ED |
1218 | } |
1219 | } | |
e2080072 ED |
1220 | list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); |
1221 | } | |
1222 | ||
05e22e83 ED |
1223 | INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); |
1224 | INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); | |
dd2e0b86 | 1225 | INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb)); |
05e22e83 | 1226 | |
1da177e4 LT |
1227 | /* This routine actually transmits TCP packets queued in by |
1228 | * tcp_do_sendmsg(). This is used by both the initial | |
1229 | * transmission and possible later retransmissions. | |
1230 | * All SKB's seen here are completely headerless. It is our | |
1231 | * job to build the TCP header, and pass the packet down to | |
1232 | * IP so it can do the same plus pass the packet off to the | |
1233 | * device. | |
1234 | * | |
1235 | * We are working here with either a clone of the original | |
1236 | * SKB, or a fresh unique copy made by the retransmit engine. | |
1237 | */ | |
2987babb YC |
1238 | static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, |
1239 | int clone_it, gfp_t gfp_mask, u32 rcv_nxt) | |
1da177e4 | 1240 | { |
dfb4b9dc DM |
1241 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1242 | struct inet_sock *inet; | |
1243 | struct tcp_sock *tp; | |
1244 | struct tcp_skb_cb *tcb; | |
33ad798c | 1245 | struct tcp_out_options opts; |
95c96174 | 1246 | unsigned int tcp_options_size, tcp_header_size; |
8c72c65b | 1247 | struct sk_buff *oskb = NULL; |
cfb6eeb4 | 1248 | struct tcp_md5sig_key *md5; |
dfb4b9dc | 1249 | struct tcphdr *th; |
a7a25630 | 1250 | u64 prior_wstamp; |
dfb4b9dc DM |
1251 | int err; |
1252 | ||
1253 | BUG_ON(!skb || !tcp_skb_pcount(skb)); | |
6f094b9e | 1254 | tp = tcp_sk(sk); |
7f12422c YC |
1255 | prior_wstamp = tp->tcp_wstamp_ns; |
1256 | tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache); | |
1257 | skb->skb_mstamp_ns = tp->tcp_wstamp_ns; | |
ccdbb6e9 | 1258 | if (clone_it) { |
6f094b9e LB |
1259 | TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq |
1260 | - tp->snd_una; | |
8c72c65b | 1261 | oskb = skb; |
e2080072 ED |
1262 | |
1263 | tcp_skb_tsorted_save(oskb) { | |
1264 | if (unlikely(skb_cloned(oskb))) | |
1265 | skb = pskb_copy(oskb, gfp_mask); | |
1266 | else | |
1267 | skb = skb_clone(oskb, gfp_mask); | |
1268 | } tcp_skb_tsorted_restore(oskb); | |
1269 | ||
dfb4b9dc DM |
1270 | if (unlikely(!skb)) |
1271 | return -ENOBUFS; | |
b738a185 ED |
1272 | /* retransmit skbs might have a non zero value in skb->dev |
1273 | * because skb->dev is aliased with skb->rbnode.rb_left | |
1274 | */ | |
1275 | skb->dev = NULL; | |
dfb4b9dc | 1276 | } |
5f6188a8 | 1277 | |
dfb4b9dc | 1278 | inet = inet_sk(sk); |
dfb4b9dc | 1279 | tcb = TCP_SKB_CB(skb); |
33ad798c | 1280 | memset(&opts, 0, sizeof(opts)); |
1da177e4 | 1281 | |
051ba674 | 1282 | if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { |
33ad798c | 1283 | tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); |
051ba674 | 1284 | } else { |
33ad798c AL |
1285 | tcp_options_size = tcp_established_options(sk, skb, &opts, |
1286 | &md5); | |
051ba674 ED |
1287 | /* Force a PSH flag on all (GSO) packets to expedite GRO flush |
1288 | * at receiver : This slightly improve GRO performance. | |
1289 | * Note that we do not force the PSH flag for non GSO packets, | |
1290 | * because they might be sent under high congestion events, | |
1291 | * and in this case it is better to delay the delivery of 1-MSS | |
1292 | * packets and thus the corresponding ACK packet that would | |
1293 | * release the following packet. | |
1294 | */ | |
1295 | if (tcp_skb_pcount(skb) > 1) | |
1296 | tcb->tcp_flags |= TCPHDR_PSH; | |
1297 | } | |
33ad798c | 1298 | tcp_header_size = tcp_options_size + sizeof(struct tcphdr); |
e905a9ed | 1299 | |
547669d4 | 1300 | /* if no packet is in qdisc/device queue, then allow XPS to select |
b2532eb9 | 1301 | * another queue. We can be called from tcp_tsq_handler() |
73a6bab5 | 1302 | * which holds one reference to sk. |
b2532eb9 ED |
1303 | * |
1304 | * TODO: Ideally, in-flight pure ACK packets should not matter here. | |
1305 | * One way to get this would be to set skb->truesize = 2 on them. | |
547669d4 | 1306 | */ |
b2532eb9 | 1307 | skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); |
dfb4b9dc | 1308 | |
38ab52e8 ED |
1309 | /* If we had to use memory reserve to allocate this skb, |
1310 | * this might cause drops if packet is looped back : | |
1311 | * Other socket might not have SOCK_MEMALLOC. | |
1312 | * Packets not looped back do not care about pfmemalloc. | |
1313 | */ | |
1314 | skb->pfmemalloc = 0; | |
1315 | ||
aa8223c7 ACM |
1316 | skb_push(skb, tcp_header_size); |
1317 | skb_reset_transport_header(skb); | |
46d3ceab ED |
1318 | |
1319 | skb_orphan(skb); | |
1320 | skb->sk = sk; | |
1d2077ac | 1321 | skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; |
14afee4b | 1322 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); |
dfb4b9dc | 1323 | |
c3a2e837 JA |
1324 | skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); |
1325 | ||
dfb4b9dc | 1326 | /* Build TCP header and checksum it. */ |
ea1627c2 | 1327 | th = (struct tcphdr *)skb->data; |
c720c7e8 ED |
1328 | th->source = inet->inet_sport; |
1329 | th->dest = inet->inet_dport; | |
dfb4b9dc | 1330 | th->seq = htonl(tcb->seq); |
2987babb | 1331 | th->ack_seq = htonl(rcv_nxt); |
df7a3b07 | 1332 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
4de075e0 | 1333 | tcb->tcp_flags); |
dfb4b9dc | 1334 | |
dfb4b9dc DM |
1335 | th->check = 0; |
1336 | th->urg_ptr = 0; | |
1da177e4 | 1337 | |
33f5f57e | 1338 | /* The urg_mode check is necessary during a below snd_una win probe */ |
7691367d HX |
1339 | if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { |
1340 | if (before(tp->snd_up, tcb->seq + 0x10000)) { | |
1341 | th->urg_ptr = htons(tp->snd_up - tcb->seq); | |
1342 | th->urg = 1; | |
1343 | } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { | |
0eae88f3 | 1344 | th->urg_ptr = htons(0xFFFF); |
7691367d HX |
1345 | th->urg = 1; |
1346 | } | |
dfb4b9dc | 1347 | } |
1da177e4 | 1348 | |
51466a75 | 1349 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
ea1627c2 ED |
1350 | if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { |
1351 | th->window = htons(tcp_select_window(sk)); | |
1352 | tcp_ecn_send(sk, skb, th, tcp_header_size); | |
1353 | } else { | |
1354 | /* RFC1323: The window in SYN & SYN/ACK segments | |
1355 | * is never scaled. | |
1356 | */ | |
1357 | th->window = htons(min(tp->rcv_wnd, 65535U)); | |
1358 | } | |
fa3fe2b1 FW |
1359 | |
1360 | tcp_options_write((__be32 *)(th + 1), tp, &opts); | |
1361 | ||
cfb6eeb4 YH |
1362 | #ifdef CONFIG_TCP_MD5SIG |
1363 | /* Calculate the MD5 hash, as we have all we need now */ | |
1364 | if (md5) { | |
a465419b | 1365 | sk_nocaps_add(sk, NETIF_F_GSO_MASK); |
bd0388ae | 1366 | tp->af_specific->calc_md5_hash(opts.hash_location, |
39f8e58e | 1367 | md5, sk, skb); |
cfb6eeb4 YH |
1368 | } |
1369 | #endif | |
1370 | ||
331fca43 MKL |
1371 | /* BPF prog is the last one writing header option */ |
1372 | bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts); | |
1373 | ||
dd2e0b86 ED |
1374 | INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check, |
1375 | tcp_v6_send_check, tcp_v4_send_check, | |
1376 | sk, skb); | |
1da177e4 | 1377 | |
4de075e0 | 1378 | if (likely(tcb->tcp_flags & TCPHDR_ACK)) |
27cde44a | 1379 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt); |
1da177e4 | 1380 | |
a44d6eac | 1381 | if (skb->len != tcp_header_size) { |
cf533ea5 | 1382 | tcp_event_data_sent(tp, sk); |
a44d6eac | 1383 | tp->data_segs_out += tcp_skb_pcount(skb); |
ba113c3a | 1384 | tp->bytes_sent += skb->len - tcp_header_size; |
a44d6eac | 1385 | } |
1da177e4 | 1386 | |
bd37a088 | 1387 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
aa2ea058 TH |
1388 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, |
1389 | tcp_skb_pcount(skb)); | |
1da177e4 | 1390 | |
2efd055c | 1391 | tp->segs_out += tcp_skb_pcount(skb); |
0ae5b43d | 1392 | skb_set_hash_from_sk(skb, sk); |
f69ad292 | 1393 | /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ |
cd7d8498 | 1394 | skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); |
f69ad292 | 1395 | skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); |
cd7d8498 | 1396 | |
d3edd06e | 1397 | /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */ |
971f10ec ED |
1398 | |
1399 | /* Cleanup our debris for IP stacks */ | |
1400 | memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), | |
1401 | sizeof(struct inet6_skb_parm))); | |
1402 | ||
a842fe14 ED |
1403 | tcp_add_tx_delay(skb, tp); |
1404 | ||
05e22e83 ED |
1405 | err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit, |
1406 | inet6_csk_xmit, ip_queue_xmit, | |
1407 | sk, skb, &inet->cork.fl); | |
7faee5c0 | 1408 | |
8c72c65b ED |
1409 | if (unlikely(err > 0)) { |
1410 | tcp_enter_cwr(sk); | |
1411 | err = net_xmit_eval(err); | |
1412 | } | |
fc225799 | 1413 | if (!err && oskb) { |
a7a25630 | 1414 | tcp_update_skb_after_send(sk, oskb, prior_wstamp); |
fc225799 ED |
1415 | tcp_rate_skb_sent(sk, oskb); |
1416 | } | |
8c72c65b | 1417 | return err; |
1da177e4 LT |
1418 | } |
1419 | ||
2987babb YC |
1420 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, |
1421 | gfp_t gfp_mask) | |
1422 | { | |
1423 | return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask, | |
1424 | tcp_sk(sk)->rcv_nxt); | |
1425 | } | |
1426 | ||
67edfef7 | 1427 | /* This routine just queues the buffer for sending. |
1da177e4 LT |
1428 | * |
1429 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, | |
1430 | * otherwise socket can stall. | |
1431 | */ | |
1432 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) | |
1433 | { | |
1434 | struct tcp_sock *tp = tcp_sk(sk); | |
1435 | ||
1436 | /* Advance write_seq and place onto the write_queue. */ | |
0f317464 | 1437 | WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq); |
f4a775d1 | 1438 | __skb_header_release(skb); |
fe067e8a | 1439 | tcp_add_write_queue_tail(sk, skb); |
ab4e846a | 1440 | sk_wmem_queued_add(sk, skb->truesize); |
3ab224be | 1441 | sk_mem_charge(sk, skb->truesize); |
1da177e4 LT |
1442 | } |
1443 | ||
67edfef7 | 1444 | /* Initialize TSO segments for a packet. */ |
5bbb432c | 1445 | static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) |
f6302d1d | 1446 | { |
4a64fd6c | 1447 | if (skb->len <= mss_now) { |
f6302d1d DM |
1448 | /* Avoid the costly divide in the normal |
1449 | * non-TSO case. | |
1450 | */ | |
cd7d8498 | 1451 | tcp_skb_pcount_set(skb, 1); |
f69ad292 | 1452 | TCP_SKB_CB(skb)->tcp_gso_size = 0; |
f6302d1d | 1453 | } else { |
cd7d8498 | 1454 | tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); |
f69ad292 | 1455 | TCP_SKB_CB(skb)->tcp_gso_size = mss_now; |
1da177e4 LT |
1456 | } |
1457 | } | |
1458 | ||
797108d1 IJ |
1459 | /* Pcount in the middle of the write queue got changed, we need to do various |
1460 | * tweaks to fix counters | |
1461 | */ | |
cf533ea5 | 1462 | static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) |
797108d1 IJ |
1463 | { |
1464 | struct tcp_sock *tp = tcp_sk(sk); | |
1465 | ||
1466 | tp->packets_out -= decr; | |
1467 | ||
1468 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
1469 | tp->sacked_out -= decr; | |
1470 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
1471 | tp->retrans_out -= decr; | |
1472 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) | |
1473 | tp->lost_out -= decr; | |
1474 | ||
1475 | /* Reno case is special. Sigh... */ | |
1476 | if (tcp_is_reno(tp) && decr > 0) | |
1477 | tp->sacked_out -= min_t(u32, tp->sacked_out, decr); | |
1478 | ||
797108d1 IJ |
1479 | if (tp->lost_skb_hint && |
1480 | before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && | |
713bafea | 1481 | (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) |
797108d1 IJ |
1482 | tp->lost_cnt_hint -= decr; |
1483 | ||
1484 | tcp_verify_left_out(tp); | |
1485 | } | |
1486 | ||
0a2cf20c SHY |
1487 | static bool tcp_has_tx_tstamp(const struct sk_buff *skb) |
1488 | { | |
1489 | return TCP_SKB_CB(skb)->txstamp_ack || | |
1490 | (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); | |
1491 | } | |
1492 | ||
490cc7d0 WB |
1493 | static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) |
1494 | { | |
1495 | struct skb_shared_info *shinfo = skb_shinfo(skb); | |
1496 | ||
0a2cf20c | 1497 | if (unlikely(tcp_has_tx_tstamp(skb)) && |
490cc7d0 WB |
1498 | !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { |
1499 | struct skb_shared_info *shinfo2 = skb_shinfo(skb2); | |
1500 | u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; | |
1501 | ||
1502 | shinfo->tx_flags &= ~tsflags; | |
1503 | shinfo2->tx_flags |= tsflags; | |
1504 | swap(shinfo->tskey, shinfo2->tskey); | |
b51e13fa MKL |
1505 | TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; |
1506 | TCP_SKB_CB(skb)->txstamp_ack = 0; | |
490cc7d0 WB |
1507 | } |
1508 | } | |
1509 | ||
a166140e MKL |
1510 | static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) |
1511 | { | |
1512 | TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; | |
1513 | TCP_SKB_CB(skb)->eor = 0; | |
1514 | } | |
1515 | ||
75c119af ED |
1516 | /* Insert buff after skb on the write or rtx queue of sk. */ |
1517 | static void tcp_insert_write_queue_after(struct sk_buff *skb, | |
1518 | struct sk_buff *buff, | |
1519 | struct sock *sk, | |
1520 | enum tcp_queue tcp_queue) | |
1521 | { | |
1522 | if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) | |
1523 | __skb_queue_after(&sk->sk_write_queue, skb, buff); | |
1524 | else | |
1525 | tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); | |
1526 | } | |
1527 | ||
1da177e4 LT |
1528 | /* Function to create two new TCP segments. Shrinks the given segment |
1529 | * to the specified size and appends a new segment with the rest of the | |
e905a9ed | 1530 | * packet to the list. This won't be called frequently, I hope. |
1da177e4 LT |
1531 | * Remember, these are still headerless SKBs at this point. |
1532 | */ | |
75c119af ED |
1533 | int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, |
1534 | struct sk_buff *skb, u32 len, | |
6cc55e09 | 1535 | unsigned int mss_now, gfp_t gfp) |
1da177e4 LT |
1536 | { |
1537 | struct tcp_sock *tp = tcp_sk(sk); | |
1538 | struct sk_buff *buff; | |
6475be16 | 1539 | int nsize, old_factor; |
b617158d | 1540 | long limit; |
b60b49ea | 1541 | int nlen; |
9ce01461 | 1542 | u8 flags; |
1da177e4 | 1543 | |
2fceec13 IJ |
1544 | if (WARN_ON(len > skb->len)) |
1545 | return -EINVAL; | |
6a438bbe | 1546 | |
1da177e4 LT |
1547 | nsize = skb_headlen(skb) - len; |
1548 | if (nsize < 0) | |
1549 | nsize = 0; | |
1550 | ||
b617158d ED |
1551 | /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb. |
1552 | * We need some allowance to not penalize applications setting small | |
1553 | * SO_SNDBUF values. | |
1554 | * Also allow first and last skb in retransmit queue to be split. | |
1555 | */ | |
1556 | limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE); | |
1557 | if (unlikely((sk->sk_wmem_queued >> 1) > limit && | |
1558 | tcp_queue != TCP_FRAG_IN_WRITE_QUEUE && | |
1559 | skb != tcp_rtx_queue_head(sk) && | |
1560 | skb != tcp_rtx_queue_tail(sk))) { | |
f070ef2a ED |
1561 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG); |
1562 | return -ENOMEM; | |
1563 | } | |
1564 | ||
6cc55e09 | 1565 | if (skb_unclone(skb, gfp)) |
1da177e4 LT |
1566 | return -ENOMEM; |
1567 | ||
1568 | /* Get a new skb... force flag on. */ | |
eb934478 | 1569 | buff = sk_stream_alloc_skb(sk, nsize, gfp, true); |
51456b29 | 1570 | if (!buff) |
1da177e4 | 1571 | return -ENOMEM; /* We'll just try again later. */ |
41477662 | 1572 | skb_copy_decrypted(buff, skb); |
5a369ca6 | 1573 | mptcp_skb_ext_copy(buff, skb); |
ef5cb973 | 1574 | |
ab4e846a | 1575 | sk_wmem_queued_add(sk, buff->truesize); |
3ab224be | 1576 | sk_mem_charge(sk, buff->truesize); |
b60b49ea HX |
1577 | nlen = skb->len - len - nsize; |
1578 | buff->truesize += nlen; | |
1579 | skb->truesize -= nlen; | |
1da177e4 LT |
1580 | |
1581 | /* Correct the sequence numbers. */ | |
1582 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1583 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1584 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1585 | ||
1586 | /* PSH and FIN should only be set in the second packet. */ | |
4de075e0 ED |
1587 | flags = TCP_SKB_CB(skb)->tcp_flags; |
1588 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
1589 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
e14c3caf | 1590 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
a166140e | 1591 | tcp_skb_fragment_eor(skb, buff); |
1da177e4 | 1592 | |
98be9b12 | 1593 | skb_split(skb, buff, len); |
1da177e4 | 1594 | |
98be9b12 | 1595 | buff->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 | 1596 | |
a61bbcf2 | 1597 | buff->tstamp = skb->tstamp; |
490cc7d0 | 1598 | tcp_fragment_tstamp(skb, buff); |
1da177e4 | 1599 | |
6475be16 DM |
1600 | old_factor = tcp_skb_pcount(skb); |
1601 | ||
1da177e4 | 1602 | /* Fix up tso_factor for both original and new SKB. */ |
5bbb432c ED |
1603 | tcp_set_skb_tso_segs(skb, mss_now); |
1604 | tcp_set_skb_tso_segs(buff, mss_now); | |
1da177e4 | 1605 | |
b9f64820 YC |
1606 | /* Update delivered info for the new segment */ |
1607 | TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; | |
1608 | ||
6475be16 DM |
1609 | /* If this packet has been sent out already, we must |
1610 | * adjust the various packet counters. | |
1611 | */ | |
cf0b450c | 1612 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
6475be16 DM |
1613 | int diff = old_factor - tcp_skb_pcount(skb) - |
1614 | tcp_skb_pcount(buff); | |
1da177e4 | 1615 | |
797108d1 IJ |
1616 | if (diff) |
1617 | tcp_adjust_pcount(sk, skb, diff); | |
1da177e4 LT |
1618 | } |
1619 | ||
1620 | /* Link BUFF into the send queue. */ | |
f4a775d1 | 1621 | __skb_header_release(buff); |
75c119af | 1622 | tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); |
f67971e6 ED |
1623 | if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) |
1624 | list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); | |
1da177e4 LT |
1625 | |
1626 | return 0; | |
1627 | } | |
1628 | ||
f4d01666 ED |
1629 | /* This is similar to __pskb_pull_tail(). The difference is that pulled |
1630 | * data is not copied, but immediately discarded. | |
1da177e4 | 1631 | */ |
7162fb24 | 1632 | static int __pskb_trim_head(struct sk_buff *skb, int len) |
1da177e4 | 1633 | { |
7b7fc97a | 1634 | struct skb_shared_info *shinfo; |
1da177e4 LT |
1635 | int i, k, eat; |
1636 | ||
4fa48bf3 ED |
1637 | eat = min_t(int, len, skb_headlen(skb)); |
1638 | if (eat) { | |
1639 | __skb_pull(skb, eat); | |
1640 | len -= eat; | |
1641 | if (!len) | |
7162fb24 | 1642 | return 0; |
4fa48bf3 | 1643 | } |
1da177e4 LT |
1644 | eat = len; |
1645 | k = 0; | |
7b7fc97a ED |
1646 | shinfo = skb_shinfo(skb); |
1647 | for (i = 0; i < shinfo->nr_frags; i++) { | |
1648 | int size = skb_frag_size(&shinfo->frags[i]); | |
9e903e08 ED |
1649 | |
1650 | if (size <= eat) { | |
aff65da0 | 1651 | skb_frag_unref(skb, i); |
9e903e08 | 1652 | eat -= size; |
1da177e4 | 1653 | } else { |
7b7fc97a | 1654 | shinfo->frags[k] = shinfo->frags[i]; |
1da177e4 | 1655 | if (eat) { |
b54c9d5b | 1656 | skb_frag_off_add(&shinfo->frags[k], eat); |
7b7fc97a | 1657 | skb_frag_size_sub(&shinfo->frags[k], eat); |
1da177e4 LT |
1658 | eat = 0; |
1659 | } | |
1660 | k++; | |
1661 | } | |
1662 | } | |
7b7fc97a | 1663 | shinfo->nr_frags = k; |
1da177e4 | 1664 | |
1da177e4 LT |
1665 | skb->data_len -= len; |
1666 | skb->len = skb->data_len; | |
7162fb24 | 1667 | return len; |
1da177e4 LT |
1668 | } |
1669 | ||
67edfef7 | 1670 | /* Remove acked data from a packet in the transmit queue. */ |
1da177e4 LT |
1671 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) |
1672 | { | |
7162fb24 ED |
1673 | u32 delta_truesize; |
1674 | ||
14bbd6a5 | 1675 | if (skb_unclone(skb, GFP_ATOMIC)) |
1da177e4 LT |
1676 | return -ENOMEM; |
1677 | ||
7162fb24 | 1678 | delta_truesize = __pskb_trim_head(skb, len); |
1da177e4 LT |
1679 | |
1680 | TCP_SKB_CB(skb)->seq += len; | |
84fa7933 | 1681 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 | 1682 | |
7162fb24 ED |
1683 | if (delta_truesize) { |
1684 | skb->truesize -= delta_truesize; | |
ab4e846a | 1685 | sk_wmem_queued_add(sk, -delta_truesize); |
7162fb24 | 1686 | sk_mem_uncharge(sk, delta_truesize); |
7162fb24 | 1687 | } |
1da177e4 | 1688 | |
5b35e1e6 | 1689 | /* Any change of skb->len requires recalculation of tso factor. */ |
1da177e4 | 1690 | if (tcp_skb_pcount(skb) > 1) |
5bbb432c | 1691 | tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); |
1da177e4 LT |
1692 | |
1693 | return 0; | |
1694 | } | |
1695 | ||
1b63edd6 YC |
1696 | /* Calculate MSS not accounting any TCP options. */ |
1697 | static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
5d424d5a | 1698 | { |
cf533ea5 ED |
1699 | const struct tcp_sock *tp = tcp_sk(sk); |
1700 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
5d424d5a JH |
1701 | int mss_now; |
1702 | ||
1703 | /* Calculate base mss without TCP options: | |
1704 | It is MMS_S - sizeof(tcphdr) of rfc1122 | |
1705 | */ | |
1706 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); | |
1707 | ||
67469601 ED |
1708 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
1709 | if (icsk->icsk_af_ops->net_frag_header_len) { | |
1710 | const struct dst_entry *dst = __sk_dst_get(sk); | |
1711 | ||
1712 | if (dst && dst_allfrag(dst)) | |
1713 | mss_now -= icsk->icsk_af_ops->net_frag_header_len; | |
1714 | } | |
1715 | ||
5d424d5a JH |
1716 | /* Clamp it (mss_clamp does not include tcp options) */ |
1717 | if (mss_now > tp->rx_opt.mss_clamp) | |
1718 | mss_now = tp->rx_opt.mss_clamp; | |
1719 | ||
1720 | /* Now subtract optional transport overhead */ | |
1721 | mss_now -= icsk->icsk_ext_hdr_len; | |
1722 | ||
1723 | /* Then reserve room for full set of TCP options and 8 bytes of data */ | |
5f3e2bf0 | 1724 | mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss); |
5d424d5a JH |
1725 | return mss_now; |
1726 | } | |
1727 | ||
1b63edd6 YC |
1728 | /* Calculate MSS. Not accounting for SACKs here. */ |
1729 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
1730 | { | |
1731 | /* Subtract TCP options size, not including SACKs */ | |
1732 | return __tcp_mtu_to_mss(sk, pmtu) - | |
1733 | (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); | |
1734 | } | |
c7bb4b89 | 1735 | EXPORT_SYMBOL(tcp_mtu_to_mss); |
1b63edd6 | 1736 | |
5d424d5a | 1737 | /* Inverse of above */ |
67469601 | 1738 | int tcp_mss_to_mtu(struct sock *sk, int mss) |
5d424d5a | 1739 | { |
cf533ea5 ED |
1740 | const struct tcp_sock *tp = tcp_sk(sk); |
1741 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
5d424d5a JH |
1742 | int mtu; |
1743 | ||
1744 | mtu = mss + | |
1745 | tp->tcp_header_len + | |
1746 | icsk->icsk_ext_hdr_len + | |
1747 | icsk->icsk_af_ops->net_header_len; | |
1748 | ||
67469601 ED |
1749 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
1750 | if (icsk->icsk_af_ops->net_frag_header_len) { | |
1751 | const struct dst_entry *dst = __sk_dst_get(sk); | |
1752 | ||
1753 | if (dst && dst_allfrag(dst)) | |
1754 | mtu += icsk->icsk_af_ops->net_frag_header_len; | |
1755 | } | |
5d424d5a JH |
1756 | return mtu; |
1757 | } | |
556c6b46 | 1758 | EXPORT_SYMBOL(tcp_mss_to_mtu); |
5d424d5a | 1759 | |
67edfef7 | 1760 | /* MTU probing init per socket */ |
5d424d5a JH |
1761 | void tcp_mtup_init(struct sock *sk) |
1762 | { | |
1763 | struct tcp_sock *tp = tcp_sk(sk); | |
1764 | struct inet_connection_sock *icsk = inet_csk(sk); | |
b0f9ca53 | 1765 | struct net *net = sock_net(sk); |
5d424d5a | 1766 | |
b0f9ca53 | 1767 | icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; |
5d424d5a | 1768 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + |
e905a9ed | 1769 | icsk->icsk_af_ops->net_header_len; |
b0f9ca53 | 1770 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); |
5d424d5a | 1771 | icsk->icsk_mtup.probe_size = 0; |
05cbc0db | 1772 | if (icsk->icsk_mtup.enabled) |
c74df29a | 1773 | icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; |
5d424d5a | 1774 | } |
4bc2f18b | 1775 | EXPORT_SYMBOL(tcp_mtup_init); |
5d424d5a | 1776 | |
1da177e4 LT |
1777 | /* This function synchronize snd mss to current pmtu/exthdr set. |
1778 | ||
1779 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts | |
1780 | for TCP options, but includes only bare TCP header. | |
1781 | ||
1782 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. | |
caa20d9a | 1783 | It is minimum of user_mss and mss received with SYN. |
1da177e4 LT |
1784 | It also does not include TCP options. |
1785 | ||
d83d8461 | 1786 | inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. |
1da177e4 LT |
1787 | |
1788 | tp->mss_cache is current effective sending mss, including | |
1789 | all tcp options except for SACKs. It is evaluated, | |
1790 | taking into account current pmtu, but never exceeds | |
1791 | tp->rx_opt.mss_clamp. | |
1792 | ||
1793 | NOTE1. rfc1122 clearly states that advertised MSS | |
1794 | DOES NOT include either tcp or ip options. | |
1795 | ||
d83d8461 ACM |
1796 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
1797 | are READ ONLY outside this function. --ANK (980731) | |
1da177e4 | 1798 | */ |
1da177e4 LT |
1799 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) |
1800 | { | |
1801 | struct tcp_sock *tp = tcp_sk(sk); | |
d83d8461 | 1802 | struct inet_connection_sock *icsk = inet_csk(sk); |
5d424d5a | 1803 | int mss_now; |
1da177e4 | 1804 | |
5d424d5a JH |
1805 | if (icsk->icsk_mtup.search_high > pmtu) |
1806 | icsk->icsk_mtup.search_high = pmtu; | |
1da177e4 | 1807 | |
5d424d5a | 1808 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
409d22b4 | 1809 | mss_now = tcp_bound_to_half_wnd(tp, mss_now); |
1da177e4 LT |
1810 | |
1811 | /* And store cached results */ | |
d83d8461 | 1812 | icsk->icsk_pmtu_cookie = pmtu; |
5d424d5a JH |
1813 | if (icsk->icsk_mtup.enabled) |
1814 | mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); | |
c1b4a7e6 | 1815 | tp->mss_cache = mss_now; |
1da177e4 LT |
1816 | |
1817 | return mss_now; | |
1818 | } | |
4bc2f18b | 1819 | EXPORT_SYMBOL(tcp_sync_mss); |
1da177e4 LT |
1820 | |
1821 | /* Compute the current effective MSS, taking SACKs and IP options, | |
1822 | * and even PMTU discovery events into account. | |
1da177e4 | 1823 | */ |
0c54b85f | 1824 | unsigned int tcp_current_mss(struct sock *sk) |
1da177e4 | 1825 | { |
cf533ea5 ED |
1826 | const struct tcp_sock *tp = tcp_sk(sk); |
1827 | const struct dst_entry *dst = __sk_dst_get(sk); | |
c1b4a7e6 | 1828 | u32 mss_now; |
95c96174 | 1829 | unsigned int header_len; |
33ad798c AL |
1830 | struct tcp_out_options opts; |
1831 | struct tcp_md5sig_key *md5; | |
c1b4a7e6 DM |
1832 | |
1833 | mss_now = tp->mss_cache; | |
1834 | ||
1da177e4 LT |
1835 | if (dst) { |
1836 | u32 mtu = dst_mtu(dst); | |
d83d8461 | 1837 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
1da177e4 LT |
1838 | mss_now = tcp_sync_mss(sk, mtu); |
1839 | } | |
1840 | ||
33ad798c AL |
1841 | header_len = tcp_established_options(sk, NULL, &opts, &md5) + |
1842 | sizeof(struct tcphdr); | |
1843 | /* The mss_cache is sized based on tp->tcp_header_len, which assumes | |
1844 | * some common options. If this is an odd packet (because we have SACK | |
1845 | * blocks etc) then our calculated header_len will be different, and | |
1846 | * we have to adjust mss_now correspondingly */ | |
1847 | if (header_len != tp->tcp_header_len) { | |
1848 | int delta = (int) header_len - tp->tcp_header_len; | |
1849 | mss_now -= delta; | |
1850 | } | |
cfb6eeb4 | 1851 | |
1da177e4 LT |
1852 | return mss_now; |
1853 | } | |
1854 | ||
86fd14ad WP |
1855 | /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. |
1856 | * As additional protections, we do not touch cwnd in retransmission phases, | |
1857 | * and if application hit its sndbuf limit recently. | |
1858 | */ | |
1859 | static void tcp_cwnd_application_limited(struct sock *sk) | |
1860 | { | |
1861 | struct tcp_sock *tp = tcp_sk(sk); | |
1862 | ||
1863 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && | |
1864 | sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { | |
1865 | /* Limited by application or receiver window. */ | |
1866 | u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); | |
1867 | u32 win_used = max(tp->snd_cwnd_used, init_win); | |
1868 | if (win_used < tp->snd_cwnd) { | |
1869 | tp->snd_ssthresh = tcp_current_ssthresh(sk); | |
1870 | tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; | |
1871 | } | |
1872 | tp->snd_cwnd_used = 0; | |
1873 | } | |
c2203cf7 | 1874 | tp->snd_cwnd_stamp = tcp_jiffies32; |
86fd14ad WP |
1875 | } |
1876 | ||
ca8a2263 | 1877 | static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) |
a762a980 | 1878 | { |
1b1fc3fd | 1879 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; |
9e412ba7 | 1880 | struct tcp_sock *tp = tcp_sk(sk); |
a762a980 | 1881 | |
ca8a2263 NC |
1882 | /* Track the maximum number of outstanding packets in each |
1883 | * window, and remember whether we were cwnd-limited then. | |
1884 | */ | |
1885 | if (!before(tp->snd_una, tp->max_packets_seq) || | |
299bcb55 NC |
1886 | tp->packets_out > tp->max_packets_out || |
1887 | is_cwnd_limited) { | |
ca8a2263 NC |
1888 | tp->max_packets_out = tp->packets_out; |
1889 | tp->max_packets_seq = tp->snd_nxt; | |
1890 | tp->is_cwnd_limited = is_cwnd_limited; | |
1891 | } | |
e114a710 | 1892 | |
24901551 | 1893 | if (tcp_is_cwnd_limited(sk)) { |
a762a980 DM |
1894 | /* Network is feed fully. */ |
1895 | tp->snd_cwnd_used = 0; | |
c2203cf7 | 1896 | tp->snd_cwnd_stamp = tcp_jiffies32; |
a762a980 DM |
1897 | } else { |
1898 | /* Network starves. */ | |
1899 | if (tp->packets_out > tp->snd_cwnd_used) | |
1900 | tp->snd_cwnd_used = tp->packets_out; | |
1901 | ||
b510f0d2 | 1902 | if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle && |
c2203cf7 | 1903 | (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && |
1b1fc3fd | 1904 | !ca_ops->cong_control) |
a762a980 | 1905 | tcp_cwnd_application_limited(sk); |
b0f71bd3 FY |
1906 | |
1907 | /* The following conditions together indicate the starvation | |
1908 | * is caused by insufficient sender buffer: | |
1909 | * 1) just sent some data (see tcp_write_xmit) | |
1910 | * 2) not cwnd limited (this else condition) | |
75c119af | 1911 | * 3) no more data to send (tcp_write_queue_empty()) |
b0f71bd3 FY |
1912 | * 4) application is hitting buffer limit (SOCK_NOSPACE) |
1913 | */ | |
75c119af | 1914 | if (tcp_write_queue_empty(sk) && sk->sk_socket && |
b0f71bd3 FY |
1915 | test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && |
1916 | (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) | |
1917 | tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); | |
a762a980 DM |
1918 | } |
1919 | } | |
1920 | ||
d4589926 ED |
1921 | /* Minshall's variant of the Nagle send check. */ |
1922 | static bool tcp_minshall_check(const struct tcp_sock *tp) | |
1923 | { | |
1924 | return after(tp->snd_sml, tp->snd_una) && | |
1925 | !after(tp->snd_sml, tp->snd_nxt); | |
1926 | } | |
1927 | ||
1928 | /* Update snd_sml if this skb is under mss | |
1929 | * Note that a TSO packet might end with a sub-mss segment | |
1930 | * The test is really : | |
1931 | * if ((skb->len % mss) != 0) | |
1932 | * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; | |
1933 | * But we can avoid doing the divide again given we already have | |
1934 | * skb_pcount = skb->len / mss_now | |
0e3a4803 | 1935 | */ |
d4589926 ED |
1936 | static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, |
1937 | const struct sk_buff *skb) | |
1938 | { | |
1939 | if (skb->len < tcp_skb_pcount(skb) * mss_now) | |
1940 | tp->snd_sml = TCP_SKB_CB(skb)->end_seq; | |
1941 | } | |
1942 | ||
1943 | /* Return false, if packet can be sent now without violation Nagle's rules: | |
1944 | * 1. It is full sized. (provided by caller in %partial bool) | |
1945 | * 2. Or it contains FIN. (already checked by caller) | |
1946 | * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. | |
1947 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. | |
1948 | * With Minshall's modification: all sent small packets are ACKed. | |
1949 | */ | |
1950 | static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, | |
cc93fc51 | 1951 | int nonagle) |
d4589926 ED |
1952 | { |
1953 | return partial && | |
1954 | ((nonagle & TCP_NAGLE_CORK) || | |
1955 | (!nonagle && tp->packets_out && tcp_minshall_check(tp))); | |
1956 | } | |
605ad7f1 ED |
1957 | |
1958 | /* Return how many segs we'd like on a TSO packet, | |
1959 | * to send one TSO packet per ms | |
1960 | */ | |
dcb8c9b4 ED |
1961 | static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, |
1962 | int min_tso_segs) | |
605ad7f1 ED |
1963 | { |
1964 | u32 bytes, segs; | |
1965 | ||
76a9ebe8 | 1966 | bytes = min_t(unsigned long, |
7c68fa2b | 1967 | sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift), |
76a9ebe8 | 1968 | sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); |
605ad7f1 ED |
1969 | |
1970 | /* Goal is to send at least one packet per ms, | |
1971 | * not one big TSO packet every 100 ms. | |
1972 | * This preserves ACK clocking and is consistent | |
1973 | * with tcp_tso_should_defer() heuristic. | |
1974 | */ | |
1b3878ca | 1975 | segs = max_t(u32, bytes / mss_now, min_tso_segs); |
605ad7f1 | 1976 | |
350c9f48 | 1977 | return segs; |
605ad7f1 ED |
1978 | } |
1979 | ||
ed6e7268 NC |
1980 | /* Return the number of segments we want in the skb we are transmitting. |
1981 | * See if congestion control module wants to decide; otherwise, autosize. | |
1982 | */ | |
1983 | static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) | |
1984 | { | |
1985 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; | |
dcb8c9b4 | 1986 | u32 min_tso, tso_segs; |
ed6e7268 | 1987 | |
dcb8c9b4 ED |
1988 | min_tso = ca_ops->min_tso_segs ? |
1989 | ca_ops->min_tso_segs(sk) : | |
1990 | sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs; | |
1991 | ||
1992 | tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); | |
350c9f48 | 1993 | return min_t(u32, tso_segs, sk->sk_gso_max_segs); |
ed6e7268 NC |
1994 | } |
1995 | ||
d4589926 ED |
1996 | /* Returns the portion of skb which can be sent right away */ |
1997 | static unsigned int tcp_mss_split_point(const struct sock *sk, | |
1998 | const struct sk_buff *skb, | |
1999 | unsigned int mss_now, | |
2000 | unsigned int max_segs, | |
2001 | int nonagle) | |
c1b4a7e6 | 2002 | { |
cf533ea5 | 2003 | const struct tcp_sock *tp = tcp_sk(sk); |
d4589926 | 2004 | u32 partial, needed, window, max_len; |
c1b4a7e6 | 2005 | |
90840def | 2006 | window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
1485348d | 2007 | max_len = mss_now * max_segs; |
0e3a4803 | 2008 | |
1485348d BH |
2009 | if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) |
2010 | return max_len; | |
0e3a4803 | 2011 | |
5ea3a748 IJ |
2012 | needed = min(skb->len, window); |
2013 | ||
1485348d BH |
2014 | if (max_len <= needed) |
2015 | return max_len; | |
0e3a4803 | 2016 | |
d4589926 ED |
2017 | partial = needed % mss_now; |
2018 | /* If last segment is not a full MSS, check if Nagle rules allow us | |
2019 | * to include this last segment in this skb. | |
2020 | * Otherwise, we'll split the skb at last MSS boundary | |
2021 | */ | |
cc93fc51 | 2022 | if (tcp_nagle_check(partial != 0, tp, nonagle)) |
d4589926 ED |
2023 | return needed - partial; |
2024 | ||
2025 | return needed; | |
c1b4a7e6 DM |
2026 | } |
2027 | ||
2028 | /* Can at least one segment of SKB be sent right now, according to the | |
2029 | * congestion window rules? If so, return how many segments are allowed. | |
2030 | */ | |
cf533ea5 ED |
2031 | static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, |
2032 | const struct sk_buff *skb) | |
c1b4a7e6 | 2033 | { |
d649a7a8 | 2034 | u32 in_flight, cwnd, halfcwnd; |
c1b4a7e6 DM |
2035 | |
2036 | /* Don't be strict about the congestion window for the final FIN. */ | |
4de075e0 ED |
2037 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && |
2038 | tcp_skb_pcount(skb) == 1) | |
c1b4a7e6 DM |
2039 | return 1; |
2040 | ||
2041 | in_flight = tcp_packets_in_flight(tp); | |
2042 | cwnd = tp->snd_cwnd; | |
d649a7a8 ED |
2043 | if (in_flight >= cwnd) |
2044 | return 0; | |
c1b4a7e6 | 2045 | |
d649a7a8 ED |
2046 | /* For better scheduling, ensure we have at least |
2047 | * 2 GSO packets in flight. | |
2048 | */ | |
2049 | halfcwnd = max(cwnd >> 1, 1U); | |
2050 | return min(halfcwnd, cwnd - in_flight); | |
c1b4a7e6 DM |
2051 | } |
2052 | ||
b595076a | 2053 | /* Initialize TSO state of a skb. |
67edfef7 | 2054 | * This must be invoked the first time we consider transmitting |
c1b4a7e6 DM |
2055 | * SKB onto the wire. |
2056 | */ | |
5bbb432c | 2057 | static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) |
c1b4a7e6 DM |
2058 | { |
2059 | int tso_segs = tcp_skb_pcount(skb); | |
2060 | ||
f8269a49 | 2061 | if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { |
5bbb432c | 2062 | tcp_set_skb_tso_segs(skb, mss_now); |
c1b4a7e6 DM |
2063 | tso_segs = tcp_skb_pcount(skb); |
2064 | } | |
2065 | return tso_segs; | |
2066 | } | |
2067 | ||
c1b4a7e6 | 2068 | |
a2a385d6 | 2069 | /* Return true if the Nagle test allows this packet to be |
c1b4a7e6 DM |
2070 | * sent now. |
2071 | */ | |
a2a385d6 ED |
2072 | static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, |
2073 | unsigned int cur_mss, int nonagle) | |
c1b4a7e6 DM |
2074 | { |
2075 | /* Nagle rule does not apply to frames, which sit in the middle of the | |
2076 | * write_queue (they have no chances to get new data). | |
2077 | * | |
2078 | * This is implemented in the callers, where they modify the 'nonagle' | |
2079 | * argument based upon the location of SKB in the send queue. | |
2080 | */ | |
2081 | if (nonagle & TCP_NAGLE_PUSH) | |
a2a385d6 | 2082 | return true; |
c1b4a7e6 | 2083 | |
9b44190d YC |
2084 | /* Don't use the nagle rule for urgent data (or for the final FIN). */ |
2085 | if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) | |
a2a385d6 | 2086 | return true; |
c1b4a7e6 | 2087 | |
cc93fc51 | 2088 | if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) |
a2a385d6 | 2089 | return true; |
c1b4a7e6 | 2090 | |
a2a385d6 | 2091 | return false; |
c1b4a7e6 DM |
2092 | } |
2093 | ||
2094 | /* Does at least the first segment of SKB fit into the send window? */ | |
a2a385d6 ED |
2095 | static bool tcp_snd_wnd_test(const struct tcp_sock *tp, |
2096 | const struct sk_buff *skb, | |
2097 | unsigned int cur_mss) | |
c1b4a7e6 DM |
2098 | { |
2099 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
2100 | ||
2101 | if (skb->len > cur_mss) | |
2102 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; | |
2103 | ||
90840def | 2104 | return !after(end_seq, tcp_wnd_end(tp)); |
c1b4a7e6 DM |
2105 | } |
2106 | ||
c1b4a7e6 DM |
2107 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet |
2108 | * which is put after SKB on the list. It is very much like | |
2109 | * tcp_fragment() except that it may make several kinds of assumptions | |
2110 | * in order to speed up the splitting operation. In particular, we | |
2111 | * know that all the data is in scatter-gather pages, and that the | |
2112 | * packet has never been sent out before (and thus is not cloned). | |
2113 | */ | |
56483341 | 2114 | static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, |
c4ead4c5 | 2115 | unsigned int mss_now, gfp_t gfp) |
c1b4a7e6 | 2116 | { |
c1b4a7e6 | 2117 | int nlen = skb->len - len; |
56483341 | 2118 | struct sk_buff *buff; |
9ce01461 | 2119 | u8 flags; |
c1b4a7e6 DM |
2120 | |
2121 | /* All of a TSO frame must be composed of paged data. */ | |
c8ac3774 | 2122 | if (skb->len != skb->data_len) |
56483341 ED |
2123 | return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, |
2124 | skb, len, mss_now, gfp); | |
c1b4a7e6 | 2125 | |
eb934478 | 2126 | buff = sk_stream_alloc_skb(sk, 0, gfp, true); |
51456b29 | 2127 | if (unlikely(!buff)) |
c1b4a7e6 | 2128 | return -ENOMEM; |
41477662 | 2129 | skb_copy_decrypted(buff, skb); |
5a369ca6 | 2130 | mptcp_skb_ext_copy(buff, skb); |
c1b4a7e6 | 2131 | |
ab4e846a | 2132 | sk_wmem_queued_add(sk, buff->truesize); |
3ab224be | 2133 | sk_mem_charge(sk, buff->truesize); |
b60b49ea | 2134 | buff->truesize += nlen; |
c1b4a7e6 DM |
2135 | skb->truesize -= nlen; |
2136 | ||
2137 | /* Correct the sequence numbers. */ | |
2138 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
2139 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
2140 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
2141 | ||
2142 | /* PSH and FIN should only be set in the second packet. */ | |
4de075e0 ED |
2143 | flags = TCP_SKB_CB(skb)->tcp_flags; |
2144 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
2145 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
c1b4a7e6 DM |
2146 | |
2147 | /* This packet was never sent out yet, so no SACK bits. */ | |
2148 | TCP_SKB_CB(buff)->sacked = 0; | |
2149 | ||
a166140e MKL |
2150 | tcp_skb_fragment_eor(skb, buff); |
2151 | ||
98be9b12 | 2152 | buff->ip_summed = CHECKSUM_PARTIAL; |
c1b4a7e6 | 2153 | skb_split(skb, buff, len); |
490cc7d0 | 2154 | tcp_fragment_tstamp(skb, buff); |
c1b4a7e6 DM |
2155 | |
2156 | /* Fix up tso_factor for both original and new SKB. */ | |
5bbb432c ED |
2157 | tcp_set_skb_tso_segs(skb, mss_now); |
2158 | tcp_set_skb_tso_segs(buff, mss_now); | |
c1b4a7e6 DM |
2159 | |
2160 | /* Link BUFF into the send queue. */ | |
f4a775d1 | 2161 | __skb_header_release(buff); |
56483341 | 2162 | tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE); |
c1b4a7e6 DM |
2163 | |
2164 | return 0; | |
2165 | } | |
2166 | ||
2167 | /* Try to defer sending, if possible, in order to minimize the amount | |
2168 | * of TSO splitting we do. View it as a kind of TSO Nagle test. | |
2169 | * | |
2170 | * This algorithm is from John Heffner. | |
2171 | */ | |
ca8a2263 | 2172 | static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, |
f9bfe4e6 ED |
2173 | bool *is_cwnd_limited, |
2174 | bool *is_rwnd_limited, | |
2175 | u32 max_segs) | |
c1b4a7e6 | 2176 | { |
6687e988 | 2177 | const struct inet_connection_sock *icsk = inet_csk(sk); |
f1c6ea38 | 2178 | u32 send_win, cong_win, limit, in_flight; |
50c8339e | 2179 | struct tcp_sock *tp = tcp_sk(sk); |
50c8339e | 2180 | struct sk_buff *head; |
ad9f4f50 | 2181 | int win_divisor; |
f1c6ea38 | 2182 | s64 delta; |
c1b4a7e6 | 2183 | |
99d7662a | 2184 | if (icsk->icsk_ca_state >= TCP_CA_Recovery) |
ae8064ac JH |
2185 | goto send_now; |
2186 | ||
5f852eb5 | 2187 | /* Avoid bursty behavior by allowing defer |
a682850a ED |
2188 | * only if the last write was recent (1 ms). |
2189 | * Note that tp->tcp_wstamp_ns can be in the future if we have | |
2190 | * packets waiting in a qdisc or device for EDT delivery. | |
5f852eb5 | 2191 | */ |
a682850a ED |
2192 | delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC; |
2193 | if (delta > 0) | |
ae8064ac | 2194 | goto send_now; |
908a75c1 | 2195 | |
c1b4a7e6 DM |
2196 | in_flight = tcp_packets_in_flight(tp); |
2197 | ||
c8c9aeb5 SB |
2198 | BUG_ON(tcp_skb_pcount(skb) <= 1); |
2199 | BUG_ON(tp->snd_cwnd <= in_flight); | |
c1b4a7e6 | 2200 | |
90840def | 2201 | send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
c1b4a7e6 DM |
2202 | |
2203 | /* From in_flight test above, we know that cwnd > in_flight. */ | |
2204 | cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; | |
2205 | ||
2206 | limit = min(send_win, cong_win); | |
2207 | ||
ba244fe9 | 2208 | /* If a full-sized TSO skb can be sent, do it. */ |
605ad7f1 | 2209 | if (limit >= max_segs * tp->mss_cache) |
ae8064ac | 2210 | goto send_now; |
ba244fe9 | 2211 | |
62ad2761 IJ |
2212 | /* Middle in queue won't get any more data, full sendable already? */ |
2213 | if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) | |
2214 | goto send_now; | |
2215 | ||
5bbcc0f5 | 2216 | win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); |
ad9f4f50 | 2217 | if (win_divisor) { |
c1b4a7e6 DM |
2218 | u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); |
2219 | ||
2220 | /* If at least some fraction of a window is available, | |
2221 | * just use it. | |
2222 | */ | |
ad9f4f50 | 2223 | chunk /= win_divisor; |
c1b4a7e6 | 2224 | if (limit >= chunk) |
ae8064ac | 2225 | goto send_now; |
c1b4a7e6 DM |
2226 | } else { |
2227 | /* Different approach, try not to defer past a single | |
2228 | * ACK. Receiver should ACK every other full sized | |
2229 | * frame, so if we have space for more than 3 frames | |
2230 | * then send now. | |
2231 | */ | |
6b5a5c0d | 2232 | if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) |
ae8064ac | 2233 | goto send_now; |
c1b4a7e6 DM |
2234 | } |
2235 | ||
75c119af ED |
2236 | /* TODO : use tsorted_sent_queue ? */ |
2237 | head = tcp_rtx_queue_head(sk); | |
2238 | if (!head) | |
2239 | goto send_now; | |
f1c6ea38 | 2240 | delta = tp->tcp_clock_cache - head->tstamp; |
50c8339e | 2241 | /* If next ACK is likely to come too late (half srtt), do not defer */ |
f1c6ea38 | 2242 | if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0) |
50c8339e ED |
2243 | goto send_now; |
2244 | ||
f9bfe4e6 ED |
2245 | /* Ok, it looks like it is advisable to defer. |
2246 | * Three cases are tracked : | |
2247 | * 1) We are cwnd-limited | |
2248 | * 2) We are rwnd-limited | |
2249 | * 3) We are application limited. | |
2250 | */ | |
2251 | if (cong_win < send_win) { | |
2252 | if (cong_win <= skb->len) { | |
2253 | *is_cwnd_limited = true; | |
2254 | return true; | |
2255 | } | |
2256 | } else { | |
2257 | if (send_win <= skb->len) { | |
2258 | *is_rwnd_limited = true; | |
2259 | return true; | |
2260 | } | |
2261 | } | |
ae8064ac | 2262 | |
f9bfe4e6 | 2263 | /* If this packet won't get more data, do not wait. */ |
d8ed257f ED |
2264 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) || |
2265 | TCP_SKB_CB(skb)->eor) | |
f9bfe4e6 | 2266 | goto send_now; |
ca8a2263 | 2267 | |
a2a385d6 | 2268 | return true; |
ae8064ac JH |
2269 | |
2270 | send_now: | |
a2a385d6 | 2271 | return false; |
c1b4a7e6 DM |
2272 | } |
2273 | ||
05cbc0db FD |
2274 | static inline void tcp_mtu_check_reprobe(struct sock *sk) |
2275 | { | |
2276 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2277 | struct tcp_sock *tp = tcp_sk(sk); | |
2278 | struct net *net = sock_net(sk); | |
2279 | u32 interval; | |
2280 | s32 delta; | |
2281 | ||
2282 | interval = net->ipv4.sysctl_tcp_probe_interval; | |
c74df29a | 2283 | delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; |
05cbc0db FD |
2284 | if (unlikely(delta >= interval * HZ)) { |
2285 | int mss = tcp_current_mss(sk); | |
2286 | ||
2287 | /* Update current search range */ | |
2288 | icsk->icsk_mtup.probe_size = 0; | |
2289 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + | |
2290 | sizeof(struct tcphdr) + | |
2291 | icsk->icsk_af_ops->net_header_len; | |
2292 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); | |
2293 | ||
2294 | /* Update probe time stamp */ | |
c74df29a | 2295 | icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; |
05cbc0db FD |
2296 | } |
2297 | } | |
2298 | ||
808cf9e3 IL |
2299 | static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) |
2300 | { | |
2301 | struct sk_buff *skb, *next; | |
2302 | ||
2303 | skb = tcp_send_head(sk); | |
2304 | tcp_for_write_queue_from_safe(skb, next, sk) { | |
2305 | if (len <= skb->len) | |
2306 | break; | |
2307 | ||
888a5c53 | 2308 | if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb)) |
808cf9e3 IL |
2309 | return false; |
2310 | ||
2311 | len -= skb->len; | |
2312 | } | |
2313 | ||
2314 | return true; | |
2315 | } | |
2316 | ||
5d424d5a | 2317 | /* Create a new MTU probe if we are ready. |
67edfef7 AK |
2318 | * MTU probe is regularly attempting to increase the path MTU by |
2319 | * deliberately sending larger packets. This discovers routing | |
2320 | * changes resulting in larger path MTUs. | |
2321 | * | |
5d424d5a JH |
2322 | * Returns 0 if we should wait to probe (no cwnd available), |
2323 | * 1 if a probe was sent, | |
056834d9 IJ |
2324 | * -1 otherwise |
2325 | */ | |
5d424d5a JH |
2326 | static int tcp_mtu_probe(struct sock *sk) |
2327 | { | |
5d424d5a | 2328 | struct inet_connection_sock *icsk = inet_csk(sk); |
12a59abc | 2329 | struct tcp_sock *tp = tcp_sk(sk); |
5d424d5a | 2330 | struct sk_buff *skb, *nskb, *next; |
6b58e0a5 | 2331 | struct net *net = sock_net(sk); |
5d424d5a | 2332 | int probe_size; |
91cc17c0 | 2333 | int size_needed; |
12a59abc | 2334 | int copy, len; |
5d424d5a | 2335 | int mss_now; |
6b58e0a5 | 2336 | int interval; |
5d424d5a JH |
2337 | |
2338 | /* Not currently probing/verifying, | |
2339 | * not in recovery, | |
2340 | * have enough cwnd, and | |
12a59abc ED |
2341 | * not SACKing (the variable headers throw things off) |
2342 | */ | |
2343 | if (likely(!icsk->icsk_mtup.enabled || | |
2344 | icsk->icsk_mtup.probe_size || | |
2345 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || | |
2346 | tp->snd_cwnd < 11 || | |
2347 | tp->rx_opt.num_sacks || tp->rx_opt.dsack)) | |
5d424d5a JH |
2348 | return -1; |
2349 | ||
6b58e0a5 FD |
2350 | /* Use binary search for probe_size between tcp_mss_base, |
2351 | * and current mss_clamp. if (search_high - search_low) | |
2352 | * smaller than a threshold, backoff from probing. | |
2353 | */ | |
0c54b85f | 2354 | mss_now = tcp_current_mss(sk); |
6b58e0a5 FD |
2355 | probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + |
2356 | icsk->icsk_mtup.search_low) >> 1); | |
91cc17c0 | 2357 | size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; |
6b58e0a5 | 2358 | interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; |
05cbc0db FD |
2359 | /* When misfortune happens, we are reprobing actively, |
2360 | * and then reprobe timer has expired. We stick with current | |
2361 | * probing process by not resetting search range to its orignal. | |
2362 | */ | |
6b58e0a5 | 2363 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || |
05cbc0db FD |
2364 | interval < net->ipv4.sysctl_tcp_probe_threshold) { |
2365 | /* Check whether enough time has elaplased for | |
2366 | * another round of probing. | |
2367 | */ | |
2368 | tcp_mtu_check_reprobe(sk); | |
5d424d5a JH |
2369 | return -1; |
2370 | } | |
2371 | ||
2372 | /* Have enough data in the send queue to probe? */ | |
7f9c33e5 | 2373 | if (tp->write_seq - tp->snd_nxt < size_needed) |
5d424d5a JH |
2374 | return -1; |
2375 | ||
91cc17c0 IJ |
2376 | if (tp->snd_wnd < size_needed) |
2377 | return -1; | |
90840def | 2378 | if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) |
91cc17c0 | 2379 | return 0; |
5d424d5a | 2380 | |
d67c58e9 IJ |
2381 | /* Do we need to wait to drain cwnd? With none in flight, don't stall */ |
2382 | if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { | |
2383 | if (!tcp_packets_in_flight(tp)) | |
5d424d5a JH |
2384 | return -1; |
2385 | else | |
2386 | return 0; | |
2387 | } | |
2388 | ||
808cf9e3 IL |
2389 | if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) |
2390 | return -1; | |
2391 | ||
5d424d5a | 2392 | /* We're allowed to probe. Build it now. */ |
eb934478 | 2393 | nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); |
51456b29 | 2394 | if (!nskb) |
5d424d5a | 2395 | return -1; |
ab4e846a | 2396 | sk_wmem_queued_add(sk, nskb->truesize); |
3ab224be | 2397 | sk_mem_charge(sk, nskb->truesize); |
5d424d5a | 2398 | |
fe067e8a | 2399 | skb = tcp_send_head(sk); |
41477662 | 2400 | skb_copy_decrypted(nskb, skb); |
5a369ca6 | 2401 | mptcp_skb_ext_copy(nskb, skb); |
5d424d5a JH |
2402 | |
2403 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; | |
2404 | TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; | |
4de075e0 | 2405 | TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; |
5d424d5a JH |
2406 | TCP_SKB_CB(nskb)->sacked = 0; |
2407 | nskb->csum = 0; | |
98be9b12 | 2408 | nskb->ip_summed = CHECKSUM_PARTIAL; |
5d424d5a | 2409 | |
50c4817e | 2410 | tcp_insert_write_queue_before(nskb, skb, sk); |
2b7cda9c | 2411 | tcp_highest_sack_replace(sk, skb, nskb); |
50c4817e | 2412 | |
5d424d5a | 2413 | len = 0; |
234b6860 | 2414 | tcp_for_write_queue_from_safe(skb, next, sk) { |
5d424d5a | 2415 | copy = min_t(int, skb->len, probe_size - len); |
98be9b12 | 2416 | skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); |
5d424d5a JH |
2417 | |
2418 | if (skb->len <= copy) { | |
2419 | /* We've eaten all the data from this skb. | |
2420 | * Throw it away. */ | |
4de075e0 | 2421 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; |
808cf9e3 IL |
2422 | /* If this is the last SKB we copy and eor is set |
2423 | * we need to propagate it to the new skb. | |
2424 | */ | |
2425 | TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; | |
888a5c53 | 2426 | tcp_skb_collapse_tstamp(nskb, skb); |
fe067e8a | 2427 | tcp_unlink_write_queue(skb, sk); |
3ab224be | 2428 | sk_wmem_free_skb(sk, skb); |
5d424d5a | 2429 | } else { |
4de075e0 | 2430 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & |
a3433f35 | 2431 | ~(TCPHDR_FIN|TCPHDR_PSH); |
5d424d5a JH |
2432 | if (!skb_shinfo(skb)->nr_frags) { |
2433 | skb_pull(skb, copy); | |
5d424d5a JH |
2434 | } else { |
2435 | __pskb_trim_head(skb, copy); | |
5bbb432c | 2436 | tcp_set_skb_tso_segs(skb, mss_now); |
5d424d5a JH |
2437 | } |
2438 | TCP_SKB_CB(skb)->seq += copy; | |
2439 | } | |
2440 | ||
2441 | len += copy; | |
234b6860 IJ |
2442 | |
2443 | if (len >= probe_size) | |
2444 | break; | |
5d424d5a | 2445 | } |
5bbb432c | 2446 | tcp_init_tso_segs(nskb, nskb->len); |
5d424d5a JH |
2447 | |
2448 | /* We're ready to send. If this fails, the probe will | |
7faee5c0 ED |
2449 | * be resegmented into mss-sized pieces by tcp_write_xmit(). |
2450 | */ | |
5d424d5a JH |
2451 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { |
2452 | /* Decrement cwnd here because we are sending | |
056834d9 | 2453 | * effectively two packets. */ |
5d424d5a | 2454 | tp->snd_cwnd--; |
66f5fe62 | 2455 | tcp_event_new_data_sent(sk, nskb); |
5d424d5a JH |
2456 | |
2457 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); | |
0e7b1368 JH |
2458 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; |
2459 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; | |
5d424d5a JH |
2460 | |
2461 | return 1; | |
2462 | } | |
2463 | ||
2464 | return -1; | |
2465 | } | |
2466 | ||
864e5c09 | 2467 | static bool tcp_pacing_check(struct sock *sk) |
218af599 | 2468 | { |
864e5c09 ED |
2469 | struct tcp_sock *tp = tcp_sk(sk); |
2470 | ||
2471 | if (!tcp_needs_internal_pacing(sk)) | |
2472 | return false; | |
2473 | ||
2474 | if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache) | |
2475 | return false; | |
2476 | ||
2477 | if (!hrtimer_is_queued(&tp->pacing_timer)) { | |
2478 | hrtimer_start(&tp->pacing_timer, | |
2479 | ns_to_ktime(tp->tcp_wstamp_ns), | |
2480 | HRTIMER_MODE_ABS_PINNED_SOFT); | |
2481 | sock_hold(sk); | |
2482 | } | |
2483 | return true; | |
218af599 ED |
2484 | } |
2485 | ||
f9616c35 ED |
2486 | /* TCP Small Queues : |
2487 | * Control number of packets in qdisc/devices to two packets / or ~1 ms. | |
2488 | * (These limits are doubled for retransmits) | |
2489 | * This allows for : | |
2490 | * - better RTT estimation and ACK scheduling | |
2491 | * - faster recovery | |
2492 | * - high rates | |
2493 | * Alas, some drivers / subsystems require a fair amount | |
2494 | * of queued bytes to ensure line rate. | |
2495 | * One example is wifi aggregation (802.11 AMPDU) | |
2496 | */ | |
2497 | static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, | |
2498 | unsigned int factor) | |
2499 | { | |
76a9ebe8 | 2500 | unsigned long limit; |
f9616c35 | 2501 | |
76a9ebe8 ED |
2502 | limit = max_t(unsigned long, |
2503 | 2 * skb->truesize, | |
7c68fa2b | 2504 | sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift)); |
c73e5807 ED |
2505 | if (sk->sk_pacing_status == SK_PACING_NONE) |
2506 | limit = min_t(unsigned long, limit, | |
2507 | sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes); | |
f9616c35 ED |
2508 | limit <<= factor; |
2509 | ||
a842fe14 ED |
2510 | if (static_branch_unlikely(&tcp_tx_delay_enabled) && |
2511 | tcp_sk(sk)->tcp_tx_delay) { | |
2512 | u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay; | |
2513 | ||
2514 | /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we | |
2515 | * approximate our needs assuming an ~100% skb->truesize overhead. | |
2516 | * USEC_PER_SEC is approximated by 2^20. | |
2517 | * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift. | |
2518 | */ | |
2519 | extra_bytes >>= (20 - 1); | |
2520 | limit += extra_bytes; | |
2521 | } | |
14afee4b | 2522 | if (refcount_read(&sk->sk_wmem_alloc) > limit) { |
75c119af | 2523 | /* Always send skb if rtx queue is empty. |
75eefc6c ED |
2524 | * No need to wait for TX completion to call us back, |
2525 | * after softirq/tasklet schedule. | |
2526 | * This helps when TX completions are delayed too much. | |
2527 | */ | |
75c119af | 2528 | if (tcp_rtx_queue_empty(sk)) |
75eefc6c ED |
2529 | return false; |
2530 | ||
7aa5470c | 2531 | set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); |
f9616c35 ED |
2532 | /* It is possible TX completion already happened |
2533 | * before we set TSQ_THROTTLED, so we must | |
2534 | * test again the condition. | |
2535 | */ | |
2536 | smp_mb__after_atomic(); | |
14afee4b | 2537 | if (refcount_read(&sk->sk_wmem_alloc) > limit) |
f9616c35 ED |
2538 | return true; |
2539 | } | |
2540 | return false; | |
2541 | } | |
2542 | ||
05b055e8 FY |
2543 | static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) |
2544 | { | |
628174cc | 2545 | const u32 now = tcp_jiffies32; |
efe967cd | 2546 | enum tcp_chrono old = tp->chrono_type; |
05b055e8 | 2547 | |
efe967cd AB |
2548 | if (old > TCP_CHRONO_UNSPEC) |
2549 | tp->chrono_stat[old - 1] += now - tp->chrono_start; | |
05b055e8 FY |
2550 | tp->chrono_start = now; |
2551 | tp->chrono_type = new; | |
2552 | } | |
2553 | ||
2554 | void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) | |
2555 | { | |
2556 | struct tcp_sock *tp = tcp_sk(sk); | |
2557 | ||
2558 | /* If there are multiple conditions worthy of tracking in a | |
0f87230d FY |
2559 | * chronograph then the highest priority enum takes precedence |
2560 | * over the other conditions. So that if something "more interesting" | |
05b055e8 FY |
2561 | * starts happening, stop the previous chrono and start a new one. |
2562 | */ | |
2563 | if (type > tp->chrono_type) | |
2564 | tcp_chrono_set(tp, type); | |
2565 | } | |
2566 | ||
2567 | void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) | |
2568 | { | |
2569 | struct tcp_sock *tp = tcp_sk(sk); | |
2570 | ||
0f87230d FY |
2571 | |
2572 | /* There are multiple conditions worthy of tracking in a | |
2573 | * chronograph, so that the highest priority enum takes | |
2574 | * precedence over the other conditions (see tcp_chrono_start). | |
2575 | * If a condition stops, we only stop chrono tracking if | |
2576 | * it's the "most interesting" or current chrono we are | |
2577 | * tracking and starts busy chrono if we have pending data. | |
2578 | */ | |
75c119af | 2579 | if (tcp_rtx_and_write_queues_empty(sk)) |
0f87230d FY |
2580 | tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); |
2581 | else if (type == tp->chrono_type) | |
2582 | tcp_chrono_set(tp, TCP_CHRONO_BUSY); | |
05b055e8 FY |
2583 | } |
2584 | ||
1da177e4 LT |
2585 | /* This routine writes packets to the network. It advances the |
2586 | * send_head. This happens as incoming acks open up the remote | |
2587 | * window for us. | |
2588 | * | |
f8269a49 IJ |
2589 | * LARGESEND note: !tcp_urg_mode is overkill, only frames between |
2590 | * snd_up-64k-mss .. snd_up cannot be large. However, taking into | |
2591 | * account rare use of URG, this is not a big flaw. | |
2592 | * | |
6ba8a3b1 ND |
2593 | * Send at most one packet when push_one > 0. Temporarily ignore |
2594 | * cwnd limit to force at most one packet out when push_one == 2. | |
2595 | ||
a2a385d6 ED |
2596 | * Returns true, if no segments are in flight and we have queued segments, |
2597 | * but cannot send anything now because of SWS or another problem. | |
1da177e4 | 2598 | */ |
a2a385d6 ED |
2599 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
2600 | int push_one, gfp_t gfp) | |
1da177e4 LT |
2601 | { |
2602 | struct tcp_sock *tp = tcp_sk(sk); | |
92df7b51 | 2603 | struct sk_buff *skb; |
c1b4a7e6 DM |
2604 | unsigned int tso_segs, sent_pkts; |
2605 | int cwnd_quota; | |
5d424d5a | 2606 | int result; |
5615f886 | 2607 | bool is_cwnd_limited = false, is_rwnd_limited = false; |
605ad7f1 | 2608 | u32 max_segs; |
1da177e4 | 2609 | |
92df7b51 | 2610 | sent_pkts = 0; |
5d424d5a | 2611 | |
ee1836ae | 2612 | tcp_mstamp_refresh(tp); |
d5dd9175 IJ |
2613 | if (!push_one) { |
2614 | /* Do MTU probing. */ | |
2615 | result = tcp_mtu_probe(sk); | |
2616 | if (!result) { | |
a2a385d6 | 2617 | return false; |
d5dd9175 IJ |
2618 | } else if (result > 0) { |
2619 | sent_pkts = 1; | |
2620 | } | |
5d424d5a JH |
2621 | } |
2622 | ||
ed6e7268 | 2623 | max_segs = tcp_tso_segs(sk, mss_now); |
fe067e8a | 2624 | while ((skb = tcp_send_head(sk))) { |
c8ac3774 HX |
2625 | unsigned int limit; |
2626 | ||
79861919 ED |
2627 | if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { |
2628 | /* "skb_mstamp_ns" is used as a start point for the retransmit timer */ | |
2629 | skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache; | |
2630 | list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); | |
bf50b606 | 2631 | tcp_init_tso_segs(skb, mss_now); |
79861919 ED |
2632 | goto repair; /* Skip network transmission */ |
2633 | } | |
2634 | ||
218af599 ED |
2635 | if (tcp_pacing_check(sk)) |
2636 | break; | |
2637 | ||
5bbb432c | 2638 | tso_segs = tcp_init_tso_segs(skb, mss_now); |
c1b4a7e6 | 2639 | BUG_ON(!tso_segs); |
aa93466b | 2640 | |
b68e9f85 | 2641 | cwnd_quota = tcp_cwnd_test(tp, skb); |
6ba8a3b1 ND |
2642 | if (!cwnd_quota) { |
2643 | if (push_one == 2) | |
2644 | /* Force out a loss probe pkt. */ | |
2645 | cwnd_quota = 1; | |
2646 | else | |
2647 | break; | |
2648 | } | |
b68e9f85 | 2649 | |
5615f886 FY |
2650 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { |
2651 | is_rwnd_limited = true; | |
b68e9f85 | 2652 | break; |
5615f886 | 2653 | } |
b68e9f85 | 2654 | |
d6a4e26a | 2655 | if (tso_segs == 1) { |
c1b4a7e6 DM |
2656 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, |
2657 | (tcp_skb_is_last(sk, skb) ? | |
2658 | nonagle : TCP_NAGLE_PUSH)))) | |
2659 | break; | |
2660 | } else { | |
ca8a2263 | 2661 | if (!push_one && |
605ad7f1 | 2662 | tcp_tso_should_defer(sk, skb, &is_cwnd_limited, |
f9bfe4e6 | 2663 | &is_rwnd_limited, max_segs)) |
c1b4a7e6 DM |
2664 | break; |
2665 | } | |
aa93466b | 2666 | |
605ad7f1 | 2667 | limit = mss_now; |
d6a4e26a | 2668 | if (tso_segs > 1 && !tcp_urg_mode(tp)) |
605ad7f1 ED |
2669 | limit = tcp_mss_split_point(sk, skb, mss_now, |
2670 | min_t(unsigned int, | |
2671 | cwnd_quota, | |
2672 | max_segs), | |
2673 | nonagle); | |
2674 | ||
2675 | if (skb->len > limit && | |
56483341 | 2676 | unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) |
605ad7f1 ED |
2677 | break; |
2678 | ||
f9616c35 ED |
2679 | if (tcp_small_queue_check(sk, skb, 0)) |
2680 | break; | |
c9eeec26 | 2681 | |
1f85e626 ED |
2682 | /* Argh, we hit an empty skb(), presumably a thread |
2683 | * is sleeping in sendmsg()/sk_stream_wait_memory(). | |
2684 | * We do not want to send a pure-ack packet and have | |
2685 | * a strange looking rtx queue with empty packet(s). | |
2686 | */ | |
2687 | if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) | |
2688 | break; | |
2689 | ||
d5dd9175 | 2690 | if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) |
92df7b51 | 2691 | break; |
1da177e4 | 2692 | |
ec342325 | 2693 | repair: |
92df7b51 DM |
2694 | /* Advance the send_head. This one is sent out. |
2695 | * This call will increment packets_out. | |
2696 | */ | |
66f5fe62 | 2697 | tcp_event_new_data_sent(sk, skb); |
1da177e4 | 2698 | |
92df7b51 | 2699 | tcp_minshall_update(tp, mss_now, skb); |
a262f0cd | 2700 | sent_pkts += tcp_skb_pcount(skb); |
d5dd9175 IJ |
2701 | |
2702 | if (push_one) | |
2703 | break; | |
92df7b51 | 2704 | } |
1da177e4 | 2705 | |
5615f886 FY |
2706 | if (is_rwnd_limited) |
2707 | tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); | |
2708 | else | |
2709 | tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); | |
2710 | ||
299bcb55 NC |
2711 | is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); |
2712 | if (likely(sent_pkts || is_cwnd_limited)) | |
2713 | tcp_cwnd_validate(sk, is_cwnd_limited); | |
2714 | ||
aa93466b | 2715 | if (likely(sent_pkts)) { |
684bad11 YC |
2716 | if (tcp_in_cwnd_reduction(sk)) |
2717 | tp->prr_out += sent_pkts; | |
6ba8a3b1 ND |
2718 | |
2719 | /* Send one loss probe per tail loss episode. */ | |
2720 | if (push_one != 2) | |
ed66dfaf | 2721 | tcp_schedule_loss_probe(sk, false); |
a2a385d6 | 2722 | return false; |
1da177e4 | 2723 | } |
75c119af | 2724 | return !tp->packets_out && !tcp_write_queue_empty(sk); |
6ba8a3b1 ND |
2725 | } |
2726 | ||
ed66dfaf | 2727 | bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) |
6ba8a3b1 ND |
2728 | { |
2729 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2730 | struct tcp_sock *tp = tcp_sk(sk); | |
a2815817 | 2731 | u32 timeout, rto_delta_us; |
2ae21cf5 | 2732 | int early_retrans; |
6ba8a3b1 | 2733 | |
6ba8a3b1 ND |
2734 | /* Don't do any loss probe on a Fast Open connection before 3WHS |
2735 | * finishes. | |
2736 | */ | |
d983ea6f | 2737 | if (rcu_access_pointer(tp->fastopen_rsk)) |
6ba8a3b1 ND |
2738 | return false; |
2739 | ||
2ae21cf5 | 2740 | early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans; |
6ba8a3b1 | 2741 | /* Schedule a loss probe in 2*RTT for SACK capable connections |
b4f70c3d | 2742 | * not in loss recovery, that are either limited by cwnd or application. |
6ba8a3b1 | 2743 | */ |
2ae21cf5 | 2744 | if ((early_retrans != 3 && early_retrans != 4) || |
bec41a11 | 2745 | !tp->packets_out || !tcp_is_sack(tp) || |
b4f70c3d NC |
2746 | (icsk->icsk_ca_state != TCP_CA_Open && |
2747 | icsk->icsk_ca_state != TCP_CA_CWR)) | |
6ba8a3b1 ND |
2748 | return false; |
2749 | ||
bb4d991a | 2750 | /* Probe timeout is 2*rtt. Add minimum RTO to account |
f9b99582 YC |
2751 | * for delayed ack when there's one outstanding packet. If no RTT |
2752 | * sample is available then probe after TCP_TIMEOUT_INIT. | |
6ba8a3b1 | 2753 | */ |
bb4d991a YC |
2754 | if (tp->srtt_us) { |
2755 | timeout = usecs_to_jiffies(tp->srtt_us >> 2); | |
2756 | if (tp->packets_out == 1) | |
2757 | timeout += TCP_RTO_MIN; | |
2758 | else | |
2759 | timeout += TCP_TIMEOUT_MIN; | |
2760 | } else { | |
2761 | timeout = TCP_TIMEOUT_INIT; | |
2762 | } | |
6ba8a3b1 | 2763 | |
a2815817 | 2764 | /* If the RTO formula yields an earlier time, then use that time. */ |
ed66dfaf NC |
2765 | rto_delta_us = advancing_rto ? |
2766 | jiffies_to_usecs(inet_csk(sk)->icsk_rto) : | |
2767 | tcp_rto_delta_us(sk); /* How far in future is RTO? */ | |
a2815817 NC |
2768 | if (rto_delta_us > 0) |
2769 | timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); | |
6ba8a3b1 | 2770 | |
8dc242ad | 2771 | tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX); |
6ba8a3b1 ND |
2772 | return true; |
2773 | } | |
2774 | ||
1f3279ae ED |
2775 | /* Thanks to skb fast clones, we can detect if a prior transmit of |
2776 | * a packet is still in a qdisc or driver queue. | |
2777 | * In this case, there is very little point doing a retransmit ! | |
1f3279ae | 2778 | */ |
f4dae54e | 2779 | static bool skb_still_in_host_queue(struct sock *sk, |
1f3279ae ED |
2780 | const struct sk_buff *skb) |
2781 | { | |
39bb5e62 | 2782 | if (unlikely(skb_fclone_busy(sk, skb))) { |
f4dae54e ED |
2783 | set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); |
2784 | smp_mb__after_atomic(); | |
2785 | if (skb_fclone_busy(sk, skb)) { | |
2786 | NET_INC_STATS(sock_net(sk), | |
2787 | LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); | |
2788 | return true; | |
2789 | } | |
1f3279ae ED |
2790 | } |
2791 | return false; | |
2792 | } | |
2793 | ||
b340b264 | 2794 | /* When probe timeout (PTO) fires, try send a new segment if possible, else |
6ba8a3b1 ND |
2795 | * retransmit the last segment. |
2796 | */ | |
2797 | void tcp_send_loss_probe(struct sock *sk) | |
2798 | { | |
9b717a8d | 2799 | struct tcp_sock *tp = tcp_sk(sk); |
6ba8a3b1 ND |
2800 | struct sk_buff *skb; |
2801 | int pcount; | |
2802 | int mss = tcp_current_mss(sk); | |
6ba8a3b1 | 2803 | |
76be93fc YC |
2804 | /* At most one outstanding TLP */ |
2805 | if (tp->tlp_high_seq) | |
2806 | goto rearm_timer; | |
2807 | ||
2808 | tp->tlp_retrans = 0; | |
b340b264 | 2809 | skb = tcp_send_head(sk); |
75c119af ED |
2810 | if (skb && tcp_snd_wnd_test(tp, skb, mss)) { |
2811 | pcount = tp->packets_out; | |
2812 | tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); | |
2813 | if (tp->packets_out > pcount) | |
2814 | goto probe_sent; | |
2815 | goto rearm_timer; | |
6ba8a3b1 | 2816 | } |
75c119af | 2817 | skb = skb_rb_last(&sk->tcp_rtx_queue); |
b2b7af86 YC |
2818 | if (unlikely(!skb)) { |
2819 | WARN_ONCE(tp->packets_out, | |
2820 | "invalid inflight: %u state %u cwnd %u mss %d\n", | |
2821 | tp->packets_out, sk->sk_state, tp->snd_cwnd, mss); | |
2822 | inet_csk(sk)->icsk_pending = 0; | |
2823 | return; | |
2824 | } | |
6ba8a3b1 | 2825 | |
1f3279ae ED |
2826 | if (skb_still_in_host_queue(sk, skb)) |
2827 | goto rearm_timer; | |
2828 | ||
6ba8a3b1 ND |
2829 | pcount = tcp_skb_pcount(skb); |
2830 | if (WARN_ON(!pcount)) | |
2831 | goto rearm_timer; | |
2832 | ||
2833 | if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { | |
75c119af ED |
2834 | if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, |
2835 | (pcount - 1) * mss, mss, | |
6cc55e09 | 2836 | GFP_ATOMIC))) |
6ba8a3b1 | 2837 | goto rearm_timer; |
75c119af | 2838 | skb = skb_rb_next(skb); |
6ba8a3b1 ND |
2839 | } |
2840 | ||
2841 | if (WARN_ON(!skb || !tcp_skb_pcount(skb))) | |
2842 | goto rearm_timer; | |
2843 | ||
10d3be56 | 2844 | if (__tcp_retransmit_skb(sk, skb, 1)) |
b340b264 | 2845 | goto rearm_timer; |
6ba8a3b1 | 2846 | |
76be93fc YC |
2847 | tp->tlp_retrans = 1; |
2848 | ||
2849 | probe_sent: | |
9b717a8d | 2850 | /* Record snd_nxt for loss detection. */ |
b340b264 | 2851 | tp->tlp_high_seq = tp->snd_nxt; |
9b717a8d | 2852 | |
c10d9310 | 2853 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); |
b340b264 YC |
2854 | /* Reset s.t. tcp_rearm_rto will restart timer from now */ |
2855 | inet_csk(sk)->icsk_pending = 0; | |
6ba8a3b1 | 2856 | rearm_timer: |
fcd16c0a | 2857 | tcp_rearm_rto(sk); |
1da177e4 LT |
2858 | } |
2859 | ||
a762a980 DM |
2860 | /* Push out any pending frames which were held back due to |
2861 | * TCP_CORK or attempt at coalescing tiny packets. | |
2862 | * The socket must be locked by the caller. | |
2863 | */ | |
9e412ba7 IJ |
2864 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, |
2865 | int nonagle) | |
a762a980 | 2866 | { |
726e07a8 IJ |
2867 | /* If we are closed, the bytes will have to remain here. |
2868 | * In time closedown will finish, we empty the write queue and | |
2869 | * all will be happy. | |
2870 | */ | |
2871 | if (unlikely(sk->sk_state == TCP_CLOSE)) | |
2872 | return; | |
2873 | ||
99a1dec7 | 2874 | if (tcp_write_xmit(sk, cur_mss, nonagle, 0, |
7450aaf6 | 2875 | sk_gfp_mask(sk, GFP_ATOMIC))) |
726e07a8 | 2876 | tcp_check_probe_timer(sk); |
a762a980 DM |
2877 | } |
2878 | ||
c1b4a7e6 DM |
2879 | /* Send _single_ skb sitting at the send head. This function requires |
2880 | * true push pending frames to setup probe timer etc. | |
2881 | */ | |
2882 | void tcp_push_one(struct sock *sk, unsigned int mss_now) | |
2883 | { | |
fe067e8a | 2884 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 DM |
2885 | |
2886 | BUG_ON(!skb || skb->len < mss_now); | |
2887 | ||
d5dd9175 | 2888 | tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); |
c1b4a7e6 DM |
2889 | } |
2890 | ||
1da177e4 LT |
2891 | /* This function returns the amount that we can raise the |
2892 | * usable window based on the following constraints | |
e905a9ed | 2893 | * |
1da177e4 LT |
2894 | * 1. The window can never be shrunk once it is offered (RFC 793) |
2895 | * 2. We limit memory per socket | |
2896 | * | |
2897 | * RFC 1122: | |
2898 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep | |
2899 | * RECV.NEXT + RCV.WIN fixed until: | |
2900 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" | |
2901 | * | |
2902 | * i.e. don't raise the right edge of the window until you can raise | |
2903 | * it at least MSS bytes. | |
2904 | * | |
2905 | * Unfortunately, the recommended algorithm breaks header prediction, | |
2906 | * since header prediction assumes th->window stays fixed. | |
2907 | * | |
2908 | * Strictly speaking, keeping th->window fixed violates the receiver | |
2909 | * side SWS prevention criteria. The problem is that under this rule | |
2910 | * a stream of single byte packets will cause the right side of the | |
2911 | * window to always advance by a single byte. | |
e905a9ed | 2912 | * |
1da177e4 LT |
2913 | * Of course, if the sender implements sender side SWS prevention |
2914 | * then this will not be a problem. | |
e905a9ed | 2915 | * |
1da177e4 | 2916 | * BSD seems to make the following compromise: |
e905a9ed | 2917 | * |
1da177e4 LT |
2918 | * If the free space is less than the 1/4 of the maximum |
2919 | * space available and the free space is less than 1/2 mss, | |
2920 | * then set the window to 0. | |
2921 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] | |
2922 | * Otherwise, just prevent the window from shrinking | |
2923 | * and from being larger than the largest representable value. | |
2924 | * | |
2925 | * This prevents incremental opening of the window in the regime | |
2926 | * where TCP is limited by the speed of the reader side taking | |
2927 | * data out of the TCP receive queue. It does nothing about | |
2928 | * those cases where the window is constrained on the sender side | |
2929 | * because the pipeline is full. | |
2930 | * | |
2931 | * BSD also seems to "accidentally" limit itself to windows that are a | |
2932 | * multiple of MSS, at least until the free space gets quite small. | |
2933 | * This would appear to be a side effect of the mbuf implementation. | |
2934 | * Combining these two algorithms results in the observed behavior | |
2935 | * of having a fixed window size at almost all times. | |
2936 | * | |
2937 | * Below we obtain similar behavior by forcing the offered window to | |
2938 | * a multiple of the mss when it is feasible to do so. | |
2939 | * | |
2940 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. | |
2941 | * Regular options like TIMESTAMP are taken into account. | |
2942 | */ | |
2943 | u32 __tcp_select_window(struct sock *sk) | |
2944 | { | |
463c84b9 | 2945 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 2946 | struct tcp_sock *tp = tcp_sk(sk); |
caa20d9a | 2947 | /* MSS for the peer's data. Previous versions used mss_clamp |
1da177e4 LT |
2948 | * here. I don't know if the value based on our guesses |
2949 | * of peer's MSS is better for the performance. It's more correct | |
2950 | * but may be worse for the performance because of rcv_mss | |
2951 | * fluctuations. --SAW 1998/11/1 | |
2952 | */ | |
463c84b9 | 2953 | int mss = icsk->icsk_ack.rcv_mss; |
1da177e4 | 2954 | int free_space = tcp_space(sk); |
86c1a045 | 2955 | int allowed_space = tcp_full_space(sk); |
071c8ed6 FW |
2956 | int full_space, window; |
2957 | ||
2958 | if (sk_is_mptcp(sk)) | |
2959 | mptcp_space(sk, &free_space, &allowed_space); | |
2960 | ||
2961 | full_space = min_t(int, tp->window_clamp, allowed_space); | |
1da177e4 | 2962 | |
06425c30 | 2963 | if (unlikely(mss > full_space)) { |
e905a9ed | 2964 | mss = full_space; |
06425c30 ED |
2965 | if (mss <= 0) |
2966 | return 0; | |
2967 | } | |
b92edbe0 | 2968 | if (free_space < (full_space >> 1)) { |
463c84b9 | 2969 | icsk->icsk_ack.quick = 0; |
1da177e4 | 2970 | |
b8da51eb | 2971 | if (tcp_under_memory_pressure(sk)) |
056834d9 IJ |
2972 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, |
2973 | 4U * tp->advmss); | |
1da177e4 | 2974 | |
86c1a045 FW |
2975 | /* free_space might become our new window, make sure we don't |
2976 | * increase it due to wscale. | |
2977 | */ | |
2978 | free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); | |
2979 | ||
2980 | /* if free space is less than mss estimate, or is below 1/16th | |
2981 | * of the maximum allowed, try to move to zero-window, else | |
2982 | * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and | |
2983 | * new incoming data is dropped due to memory limits. | |
2984 | * With large window, mss test triggers way too late in order | |
2985 | * to announce zero window in time before rmem limit kicks in. | |
2986 | */ | |
2987 | if (free_space < (allowed_space >> 4) || free_space < mss) | |
1da177e4 LT |
2988 | return 0; |
2989 | } | |
2990 | ||
2991 | if (free_space > tp->rcv_ssthresh) | |
2992 | free_space = tp->rcv_ssthresh; | |
2993 | ||
2994 | /* Don't do rounding if we are using window scaling, since the | |
2995 | * scaled window will not line up with the MSS boundary anyway. | |
2996 | */ | |
1da177e4 LT |
2997 | if (tp->rx_opt.rcv_wscale) { |
2998 | window = free_space; | |
2999 | ||
3000 | /* Advertise enough space so that it won't get scaled away. | |
3001 | * Import case: prevent zero window announcement if | |
3002 | * 1<<rcv_wscale > mss. | |
3003 | */ | |
1935299d | 3004 | window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); |
1da177e4 | 3005 | } else { |
1935299d | 3006 | window = tp->rcv_wnd; |
1da177e4 LT |
3007 | /* Get the largest window that is a nice multiple of mss. |
3008 | * Window clamp already applied above. | |
3009 | * If our current window offering is within 1 mss of the | |
3010 | * free space we just keep it. This prevents the divide | |
3011 | * and multiply from happening most of the time. | |
3012 | * We also don't do any window rounding when the free space | |
3013 | * is too small. | |
3014 | */ | |
3015 | if (window <= free_space - mss || window > free_space) | |
1935299d | 3016 | window = rounddown(free_space, mss); |
84565070 | 3017 | else if (mss == full_space && |
b92edbe0 | 3018 | free_space > window + (full_space >> 1)) |
84565070 | 3019 | window = free_space; |
1da177e4 LT |
3020 | } |
3021 | ||
3022 | return window; | |
3023 | } | |
3024 | ||
cfea5a68 MKL |
3025 | void tcp_skb_collapse_tstamp(struct sk_buff *skb, |
3026 | const struct sk_buff *next_skb) | |
082ac2d5 | 3027 | { |
0a2cf20c SHY |
3028 | if (unlikely(tcp_has_tx_tstamp(next_skb))) { |
3029 | const struct skb_shared_info *next_shinfo = | |
3030 | skb_shinfo(next_skb); | |
082ac2d5 MKL |
3031 | struct skb_shared_info *shinfo = skb_shinfo(skb); |
3032 | ||
0a2cf20c | 3033 | shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; |
082ac2d5 | 3034 | shinfo->tskey = next_shinfo->tskey; |
2de8023e MKL |
3035 | TCP_SKB_CB(skb)->txstamp_ack |= |
3036 | TCP_SKB_CB(next_skb)->txstamp_ack; | |
082ac2d5 MKL |
3037 | } |
3038 | } | |
3039 | ||
4a17fc3a | 3040 | /* Collapses two adjacent SKB's during retransmission. */ |
f8071cde | 3041 | static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) |
1da177e4 LT |
3042 | { |
3043 | struct tcp_sock *tp = tcp_sk(sk); | |
75c119af | 3044 | struct sk_buff *next_skb = skb_rb_next(skb); |
13dde04f | 3045 | int next_skb_size; |
1da177e4 | 3046 | |
058dc334 | 3047 | next_skb_size = next_skb->len; |
1da177e4 | 3048 | |
058dc334 | 3049 | BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); |
a6963a6b | 3050 | |
f8071cde ED |
3051 | if (next_skb_size) { |
3052 | if (next_skb_size <= skb_availroom(skb)) | |
3053 | skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), | |
3054 | next_skb_size); | |
3b4929f6 | 3055 | else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size)) |
f8071cde ED |
3056 | return false; |
3057 | } | |
2b7cda9c | 3058 | tcp_highest_sack_replace(sk, next_skb, skb); |
1da177e4 | 3059 | |
058dc334 IJ |
3060 | /* Update sequence range on original skb. */ |
3061 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; | |
1da177e4 | 3062 | |
e6c7d085 | 3063 | /* Merge over control information. This moves PSH/FIN etc. over */ |
4de075e0 | 3064 | TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; |
058dc334 IJ |
3065 | |
3066 | /* All done, get rid of second SKB and account for it so | |
3067 | * packet counting does not break. | |
3068 | */ | |
3069 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; | |
a643b5d4 | 3070 | TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; |
058dc334 IJ |
3071 | |
3072 | /* changed transmit queue under us so clear hints */ | |
ef9da47c IJ |
3073 | tcp_clear_retrans_hints_partial(tp); |
3074 | if (next_skb == tp->retransmit_skb_hint) | |
3075 | tp->retransmit_skb_hint = skb; | |
058dc334 | 3076 | |
797108d1 IJ |
3077 | tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); |
3078 | ||
082ac2d5 MKL |
3079 | tcp_skb_collapse_tstamp(skb, next_skb); |
3080 | ||
75c119af | 3081 | tcp_rtx_queue_unlink_and_free(next_skb, sk); |
f8071cde | 3082 | return true; |
1da177e4 LT |
3083 | } |
3084 | ||
67edfef7 | 3085 | /* Check if coalescing SKBs is legal. */ |
a2a385d6 | 3086 | static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) |
4a17fc3a IJ |
3087 | { |
3088 | if (tcp_skb_pcount(skb) > 1) | |
a2a385d6 | 3089 | return false; |
4a17fc3a | 3090 | if (skb_cloned(skb)) |
a2a385d6 | 3091 | return false; |
2331ccc5 | 3092 | /* Some heuristics for collapsing over SACK'd could be invented */ |
4a17fc3a | 3093 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) |
a2a385d6 | 3094 | return false; |
4a17fc3a | 3095 | |
a2a385d6 | 3096 | return true; |
4a17fc3a IJ |
3097 | } |
3098 | ||
67edfef7 AK |
3099 | /* Collapse packets in the retransmit queue to make to create |
3100 | * less packets on the wire. This is only done on retransmission. | |
3101 | */ | |
4a17fc3a IJ |
3102 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, |
3103 | int space) | |
3104 | { | |
3105 | struct tcp_sock *tp = tcp_sk(sk); | |
3106 | struct sk_buff *skb = to, *tmp; | |
a2a385d6 | 3107 | bool first = true; |
4a17fc3a | 3108 | |
e0a1e5b5 | 3109 | if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse) |
4a17fc3a | 3110 | return; |
4de075e0 | 3111 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) |
4a17fc3a IJ |
3112 | return; |
3113 | ||
75c119af | 3114 | skb_rbtree_walk_from_safe(skb, tmp) { |
4a17fc3a IJ |
3115 | if (!tcp_can_collapse(sk, skb)) |
3116 | break; | |
3117 | ||
85712484 | 3118 | if (!tcp_skb_can_collapse(to, skb)) |
a643b5d4 MKL |
3119 | break; |
3120 | ||
4a17fc3a IJ |
3121 | space -= skb->len; |
3122 | ||
3123 | if (first) { | |
a2a385d6 | 3124 | first = false; |
4a17fc3a IJ |
3125 | continue; |
3126 | } | |
3127 | ||
3128 | if (space < 0) | |
3129 | break; | |
4a17fc3a IJ |
3130 | |
3131 | if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) | |
3132 | break; | |
3133 | ||
f8071cde ED |
3134 | if (!tcp_collapse_retrans(sk, to)) |
3135 | break; | |
4a17fc3a IJ |
3136 | } |
3137 | } | |
3138 | ||
1da177e4 LT |
3139 | /* This retransmits one SKB. Policy decisions and retransmit queue |
3140 | * state updates are done by the caller. Returns non-zero if an | |
3141 | * error occurred which prevented the send. | |
3142 | */ | |
10d3be56 | 3143 | int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) |
1da177e4 | 3144 | { |
5d424d5a | 3145 | struct inet_connection_sock *icsk = inet_csk(sk); |
10d3be56 | 3146 | struct tcp_sock *tp = tcp_sk(sk); |
7d227cd2 | 3147 | unsigned int cur_mss; |
10d3be56 ED |
3148 | int diff, len, err; |
3149 | ||
1da177e4 | 3150 | |
10d3be56 ED |
3151 | /* Inconclusive MTU probe */ |
3152 | if (icsk->icsk_mtup.probe_size) | |
5d424d5a | 3153 | icsk->icsk_mtup.probe_size = 0; |
1da177e4 | 3154 | |
1f3279ae ED |
3155 | if (skb_still_in_host_queue(sk, skb)) |
3156 | return -EBUSY; | |
3157 | ||
1da177e4 | 3158 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { |
7f582b24 ED |
3159 | if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) { |
3160 | WARN_ON_ONCE(1); | |
3161 | return -EINVAL; | |
3162 | } | |
1da177e4 LT |
3163 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
3164 | return -ENOMEM; | |
3165 | } | |
3166 | ||
7d227cd2 SS |
3167 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) |
3168 | return -EHOSTUNREACH; /* Routing failure or similar. */ | |
3169 | ||
0c54b85f | 3170 | cur_mss = tcp_current_mss(sk); |
7d227cd2 | 3171 | |
1da177e4 LT |
3172 | /* If receiver has shrunk his window, and skb is out of |
3173 | * new window, do not retransmit it. The exception is the | |
3174 | * case, when window is shrunk to zero. In this case | |
3175 | * our retransmit serves as a zero window probe. | |
3176 | */ | |
9d4fb27d JP |
3177 | if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && |
3178 | TCP_SKB_CB(skb)->seq != tp->snd_una) | |
1da177e4 LT |
3179 | return -EAGAIN; |
3180 | ||
10d3be56 ED |
3181 | len = cur_mss * segs; |
3182 | if (skb->len > len) { | |
75c119af ED |
3183 | if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, |
3184 | cur_mss, GFP_ATOMIC)) | |
1da177e4 | 3185 | return -ENOMEM; /* We'll try again later. */ |
02276f3c | 3186 | } else { |
10d3be56 ED |
3187 | if (skb_unclone(skb, GFP_ATOMIC)) |
3188 | return -ENOMEM; | |
9eb9362e | 3189 | |
10d3be56 ED |
3190 | diff = tcp_skb_pcount(skb); |
3191 | tcp_set_skb_tso_segs(skb, cur_mss); | |
3192 | diff -= tcp_skb_pcount(skb); | |
3193 | if (diff) | |
3194 | tcp_adjust_pcount(sk, skb, diff); | |
3195 | if (skb->len < cur_mss) | |
3196 | tcp_retrans_try_collapse(sk, skb, cur_mss); | |
1da177e4 LT |
3197 | } |
3198 | ||
49213555 DB |
3199 | /* RFC3168, section 6.1.1.1. ECN fallback */ |
3200 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) | |
3201 | tcp_ecn_clear_syn(sk, skb); | |
3202 | ||
678550c6 YC |
3203 | /* Update global and local TCP statistics. */ |
3204 | segs = tcp_skb_pcount(skb); | |
3205 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); | |
3206 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) | |
3207 | __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); | |
3208 | tp->total_retrans += segs; | |
fb31c9b9 | 3209 | tp->bytes_retrans += skb->len; |
678550c6 | 3210 | |
50bceae9 TG |
3211 | /* make sure skb->data is aligned on arches that require it |
3212 | * and check if ack-trimming & collapsing extended the headroom | |
3213 | * beyond what csum_start can cover. | |
3214 | */ | |
3215 | if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || | |
3216 | skb_headroom(skb) >= 0xFFFF)) { | |
10a81980 ED |
3217 | struct sk_buff *nskb; |
3218 | ||
e2080072 ED |
3219 | tcp_skb_tsorted_save(skb) { |
3220 | nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); | |
07f8e4d0 FW |
3221 | if (nskb) { |
3222 | nskb->dev = NULL; | |
3223 | err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC); | |
3224 | } else { | |
3225 | err = -ENOBUFS; | |
3226 | } | |
e2080072 ED |
3227 | } tcp_skb_tsorted_restore(skb); |
3228 | ||
5889e2c0 | 3229 | if (!err) { |
a7a25630 | 3230 | tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns); |
5889e2c0 YS |
3231 | tcp_rate_skb_sent(sk, skb); |
3232 | } | |
117632e6 | 3233 | } else { |
c84a5711 | 3234 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
117632e6 | 3235 | } |
c84a5711 | 3236 | |
7f12422c YC |
3237 | /* To avoid taking spuriously low RTT samples based on a timestamp |
3238 | * for a transmit that never happened, always mark EVER_RETRANS | |
3239 | */ | |
3240 | TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; | |
3241 | ||
a31ad29e LB |
3242 | if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) |
3243 | tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, | |
3244 | TCP_SKB_CB(skb)->seq, segs, err); | |
3245 | ||
fc9f3501 | 3246 | if (likely(!err)) { |
e086101b | 3247 | trace_tcp_retransmit_skb(sk, skb); |
678550c6 | 3248 | } else if (err != -EBUSY) { |
ec641b39 | 3249 | NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs); |
fc9f3501 | 3250 | } |
c84a5711 | 3251 | return err; |
93b174ad YC |
3252 | } |
3253 | ||
10d3be56 | 3254 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) |
93b174ad YC |
3255 | { |
3256 | struct tcp_sock *tp = tcp_sk(sk); | |
10d3be56 | 3257 | int err = __tcp_retransmit_skb(sk, skb, segs); |
1da177e4 LT |
3258 | |
3259 | if (err == 0) { | |
1da177e4 | 3260 | #if FASTRETRANS_DEBUG > 0 |
056834d9 | 3261 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
e87cc472 | 3262 | net_dbg_ratelimited("retrans_out leaked\n"); |
1da177e4 LT |
3263 | } |
3264 | #endif | |
3265 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; | |
3266 | tp->retrans_out += tcp_skb_pcount(skb); | |
1da177e4 | 3267 | } |
6e08d5e3 | 3268 | |
7ae18975 YC |
3269 | /* Save stamp of the first (attempted) retransmit. */ |
3270 | if (!tp->retrans_stamp) | |
3271 | tp->retrans_stamp = tcp_skb_timestamp(skb); | |
3272 | ||
6e08d5e3 YC |
3273 | if (tp->undo_retrans < 0) |
3274 | tp->undo_retrans = 0; | |
3275 | tp->undo_retrans += tcp_skb_pcount(skb); | |
1da177e4 LT |
3276 | return err; |
3277 | } | |
3278 | ||
3279 | /* This gets called after a retransmit timeout, and the initially | |
3280 | * retransmitted data is acknowledged. It tries to continue | |
3281 | * resending the rest of the retransmit queue, until either | |
3282 | * we've sent it all or the congestion window limit is reached. | |
1da177e4 LT |
3283 | */ |
3284 | void tcp_xmit_retransmit_queue(struct sock *sk) | |
3285 | { | |
6687e988 | 3286 | const struct inet_connection_sock *icsk = inet_csk(sk); |
b9f1f1ce | 3287 | struct sk_buff *skb, *rtx_head, *hole = NULL; |
1da177e4 | 3288 | struct tcp_sock *tp = tcp_sk(sk); |
916e6d1a | 3289 | bool rearm_timer = false; |
840a3cbe | 3290 | u32 max_segs; |
61eb55f4 | 3291 | int mib_idx; |
6a438bbe | 3292 | |
45e77d31 IJ |
3293 | if (!tp->packets_out) |
3294 | return; | |
3295 | ||
b9f1f1ce ED |
3296 | rtx_head = tcp_rtx_queue_head(sk); |
3297 | skb = tp->retransmit_skb_hint ?: rtx_head; | |
ed6e7268 | 3298 | max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); |
75c119af | 3299 | skb_rbtree_walk_from(skb) { |
dca0aaf8 | 3300 | __u8 sacked; |
10d3be56 | 3301 | int segs; |
1da177e4 | 3302 | |
218af599 ED |
3303 | if (tcp_pacing_check(sk)) |
3304 | break; | |
3305 | ||
08ebd172 | 3306 | /* we could do better than to assign each time */ |
51456b29 | 3307 | if (!hole) |
0e1c54c2 | 3308 | tp->retransmit_skb_hint = skb; |
08ebd172 | 3309 | |
10d3be56 ED |
3310 | segs = tp->snd_cwnd - tcp_packets_in_flight(tp); |
3311 | if (segs <= 0) | |
916e6d1a | 3312 | break; |
dca0aaf8 | 3313 | sacked = TCP_SKB_CB(skb)->sacked; |
a3d2e9f8 ED |
3314 | /* In case tcp_shift_skb_data() have aggregated large skbs, |
3315 | * we need to make sure not sending too bigs TSO packets | |
3316 | */ | |
3317 | segs = min_t(int, segs, max_segs); | |
1da177e4 | 3318 | |
840a3cbe YC |
3319 | if (tp->retrans_out >= tp->lost_out) { |
3320 | break; | |
0e1c54c2 | 3321 | } else if (!(sacked & TCPCB_LOST)) { |
51456b29 | 3322 | if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) |
0e1c54c2 IJ |
3323 | hole = skb; |
3324 | continue; | |
1da177e4 | 3325 | |
0e1c54c2 IJ |
3326 | } else { |
3327 | if (icsk->icsk_ca_state != TCP_CA_Loss) | |
3328 | mib_idx = LINUX_MIB_TCPFASTRETRANS; | |
3329 | else | |
3330 | mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; | |
3331 | } | |
1da177e4 | 3332 | |
0e1c54c2 | 3333 | if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) |
1da177e4 LT |
3334 | continue; |
3335 | ||
f9616c35 | 3336 | if (tcp_small_queue_check(sk, skb, 1)) |
916e6d1a | 3337 | break; |
f9616c35 | 3338 | |
10d3be56 | 3339 | if (tcp_retransmit_skb(sk, skb, segs)) |
916e6d1a | 3340 | break; |
24ab6bec | 3341 | |
de1d6578 | 3342 | NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); |
1da177e4 | 3343 | |
684bad11 | 3344 | if (tcp_in_cwnd_reduction(sk)) |
a262f0cd ND |
3345 | tp->prr_out += tcp_skb_pcount(skb); |
3346 | ||
75c119af | 3347 | if (skb == rtx_head && |
57dde7f7 | 3348 | icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) |
916e6d1a ED |
3349 | rearm_timer = true; |
3350 | ||
1da177e4 | 3351 | } |
916e6d1a ED |
3352 | if (rearm_timer) |
3353 | tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, | |
3354 | inet_csk(sk)->icsk_rto, | |
3355 | TCP_RTO_MAX); | |
1da177e4 LT |
3356 | } |
3357 | ||
d83769a5 ED |
3358 | /* We allow to exceed memory limits for FIN packets to expedite |
3359 | * connection tear down and (memory) recovery. | |
845704a5 ED |
3360 | * Otherwise tcp_send_fin() could be tempted to either delay FIN |
3361 | * or even be forced to close flow without any FIN. | |
a6c5ea4c ED |
3362 | * In general, we want to allow one skb per socket to avoid hangs |
3363 | * with edge trigger epoll() | |
d83769a5 | 3364 | */ |
a6c5ea4c | 3365 | void sk_forced_mem_schedule(struct sock *sk, int size) |
d83769a5 | 3366 | { |
e805605c | 3367 | int amt; |
d83769a5 ED |
3368 | |
3369 | if (size <= sk->sk_forward_alloc) | |
3370 | return; | |
3371 | amt = sk_mem_pages(size); | |
3372 | sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; | |
e805605c JW |
3373 | sk_memory_allocated_add(sk, amt); |
3374 | ||
baac50bb | 3375 | if (mem_cgroup_sockets_enabled && sk->sk_memcg) |
4b1327be WW |
3376 | mem_cgroup_charge_skmem(sk->sk_memcg, amt, |
3377 | gfp_memcg_charge() | __GFP_NOFAIL); | |
d83769a5 ED |
3378 | } |
3379 | ||
845704a5 ED |
3380 | /* Send a FIN. The caller locks the socket for us. |
3381 | * We should try to send a FIN packet really hard, but eventually give up. | |
1da177e4 LT |
3382 | */ |
3383 | void tcp_send_fin(struct sock *sk) | |
3384 | { | |
ee2aabd3 | 3385 | struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk); |
e905a9ed | 3386 | struct tcp_sock *tp = tcp_sk(sk); |
e905a9ed | 3387 | |
845704a5 ED |
3388 | /* Optimization, tack on the FIN if we have one skb in write queue and |
3389 | * this skb was not yet sent, or we are under memory pressure. | |
3390 | * Note: in the latter case, FIN packet will be sent after a timeout, | |
3391 | * as TCP stack thinks it has already been transmitted. | |
1da177e4 | 3392 | */ |
ee2aabd3 | 3393 | tskb = tail; |
75c119af ED |
3394 | if (!tskb && tcp_under_memory_pressure(sk)) |
3395 | tskb = skb_rb_last(&sk->tcp_rtx_queue); | |
3396 | ||
3397 | if (tskb) { | |
845704a5 ED |
3398 | TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; |
3399 | TCP_SKB_CB(tskb)->end_seq++; | |
1da177e4 | 3400 | tp->write_seq++; |
ee2aabd3 | 3401 | if (!tail) { |
845704a5 ED |
3402 | /* This means tskb was already sent. |
3403 | * Pretend we included the FIN on previous transmit. | |
3404 | * We need to set tp->snd_nxt to the value it would have | |
3405 | * if FIN had been sent. This is because retransmit path | |
3406 | * does not change tp->snd_nxt. | |
3407 | */ | |
e0d694d6 | 3408 | WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1); |
845704a5 ED |
3409 | return; |
3410 | } | |
1da177e4 | 3411 | } else { |
845704a5 | 3412 | skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); |
d1edc085 | 3413 | if (unlikely(!skb)) |
845704a5 | 3414 | return; |
d1edc085 | 3415 | |
e2080072 | 3416 | INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); |
d83769a5 | 3417 | skb_reserve(skb, MAX_TCP_HEADER); |
a6c5ea4c | 3418 | sk_forced_mem_schedule(sk, skb->truesize); |
1da177e4 | 3419 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ |
e870a8ef | 3420 | tcp_init_nondata_skb(skb, tp->write_seq, |
a3433f35 | 3421 | TCPHDR_ACK | TCPHDR_FIN); |
1da177e4 LT |
3422 | tcp_queue_skb(sk, skb); |
3423 | } | |
845704a5 | 3424 | __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); |
1da177e4 LT |
3425 | } |
3426 | ||
3427 | /* We get here when a process closes a file descriptor (either due to | |
3428 | * an explicit close() or as a byproduct of exit()'ing) and there | |
3429 | * was unread data in the receive queue. This behavior is recommended | |
65bb723c | 3430 | * by RFC 2525, section 2.17. -DaveM |
1da177e4 | 3431 | */ |
dd0fc66f | 3432 | void tcp_send_active_reset(struct sock *sk, gfp_t priority) |
1da177e4 | 3433 | { |
1da177e4 LT |
3434 | struct sk_buff *skb; |
3435 | ||
7cc2b043 GF |
3436 | TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); |
3437 | ||
1da177e4 LT |
3438 | /* NOTE: No TCP options attached and we never retransmit this. */ |
3439 | skb = alloc_skb(MAX_TCP_HEADER, priority); | |
3440 | if (!skb) { | |
4e673444 | 3441 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
1da177e4 LT |
3442 | return; |
3443 | } | |
3444 | ||
3445 | /* Reserve space for headers and prepare control bits. */ | |
3446 | skb_reserve(skb, MAX_TCP_HEADER); | |
e870a8ef | 3447 | tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), |
a3433f35 | 3448 | TCPHDR_ACK | TCPHDR_RST); |
9a568de4 | 3449 | tcp_mstamp_refresh(tcp_sk(sk)); |
1da177e4 | 3450 | /* Send it off. */ |
dfb4b9dc | 3451 | if (tcp_transmit_skb(sk, skb, 0, priority)) |
4e673444 | 3452 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
c24b14c4 SL |
3453 | |
3454 | /* skb of trace_tcp_send_reset() keeps the skb that caused RST, | |
3455 | * skb here is different to the troublesome skb, so use NULL | |
3456 | */ | |
3457 | trace_tcp_send_reset(sk, NULL); | |
1da177e4 LT |
3458 | } |
3459 | ||
67edfef7 AK |
3460 | /* Send a crossed SYN-ACK during socket establishment. |
3461 | * WARNING: This routine must only be called when we have already sent | |
1da177e4 LT |
3462 | * a SYN packet that crossed the incoming SYN that caused this routine |
3463 | * to get called. If this assumption fails then the initial rcv_wnd | |
3464 | * and rcv_wscale values will not be correct. | |
3465 | */ | |
3466 | int tcp_send_synack(struct sock *sk) | |
3467 | { | |
056834d9 | 3468 | struct sk_buff *skb; |
1da177e4 | 3469 | |
75c119af | 3470 | skb = tcp_rtx_queue_head(sk); |
51456b29 | 3471 | if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { |
75c119af | 3472 | pr_err("%s: wrong queue state\n", __func__); |
1da177e4 LT |
3473 | return -EFAULT; |
3474 | } | |
4de075e0 | 3475 | if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { |
1da177e4 | 3476 | if (skb_cloned(skb)) { |
e2080072 ED |
3477 | struct sk_buff *nskb; |
3478 | ||
3479 | tcp_skb_tsorted_save(skb) { | |
3480 | nskb = skb_copy(skb, GFP_ATOMIC); | |
3481 | } tcp_skb_tsorted_restore(skb); | |
51456b29 | 3482 | if (!nskb) |
1da177e4 | 3483 | return -ENOMEM; |
e2080072 | 3484 | INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); |
2bec445f | 3485 | tcp_highest_sack_replace(sk, skb, nskb); |
75c119af | 3486 | tcp_rtx_queue_unlink_and_free(skb, sk); |
f4a775d1 | 3487 | __skb_header_release(nskb); |
75c119af | 3488 | tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); |
ab4e846a | 3489 | sk_wmem_queued_add(sk, nskb->truesize); |
3ab224be | 3490 | sk_mem_charge(sk, nskb->truesize); |
1da177e4 LT |
3491 | skb = nskb; |
3492 | } | |
3493 | ||
4de075e0 | 3494 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; |
735d3831 | 3495 | tcp_ecn_send_synack(sk, skb); |
1da177e4 | 3496 | } |
dfb4b9dc | 3497 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
1da177e4 LT |
3498 | } |
3499 | ||
4aea39c1 | 3500 | /** |
331fca43 MKL |
3501 | * tcp_make_synack - Allocate one skb and build a SYNACK packet. |
3502 | * @sk: listener socket | |
3503 | * @dst: dst entry attached to the SYNACK. It is consumed and caller | |
3504 | * should not use it again. | |
3505 | * @req: request_sock pointer | |
3506 | * @foc: cookie for tcp fast open | |
3507 | * @synack_type: Type of synack to prepare | |
3508 | * @syn_skb: SYN packet just received. It could be NULL for rtx case. | |
4aea39c1 | 3509 | */ |
5d062de7 | 3510 | struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, |
e6b4d113 | 3511 | struct request_sock *req, |
ca6fb065 | 3512 | struct tcp_fastopen_cookie *foc, |
331fca43 MKL |
3513 | enum tcp_synack_type synack_type, |
3514 | struct sk_buff *syn_skb) | |
1da177e4 | 3515 | { |
2e6599cb | 3516 | struct inet_request_sock *ireq = inet_rsk(req); |
5d062de7 | 3517 | const struct tcp_sock *tp = tcp_sk(sk); |
80f03e27 | 3518 | struct tcp_md5sig_key *md5 = NULL; |
5d062de7 ED |
3519 | struct tcp_out_options opts; |
3520 | struct sk_buff *skb; | |
bd0388ae | 3521 | int tcp_header_size; |
5d062de7 | 3522 | struct tcphdr *th; |
f5fff5dc | 3523 | int mss; |
a842fe14 | 3524 | u64 now; |
1da177e4 | 3525 | |
ca6fb065 | 3526 | skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); |
4aea39c1 ED |
3527 | if (unlikely(!skb)) { |
3528 | dst_release(dst); | |
1da177e4 | 3529 | return NULL; |
4aea39c1 | 3530 | } |
1da177e4 LT |
3531 | /* Reserve space for headers. */ |
3532 | skb_reserve(skb, MAX_TCP_HEADER); | |
3533 | ||
b3d05147 ED |
3534 | switch (synack_type) { |
3535 | case TCP_SYNACK_NORMAL: | |
9e17f8a4 | 3536 | skb_set_owner_w(skb, req_to_sk(req)); |
b3d05147 ED |
3537 | break; |
3538 | case TCP_SYNACK_COOKIE: | |
3539 | /* Under synflood, we do not attach skb to a socket, | |
3540 | * to avoid false sharing. | |
3541 | */ | |
3542 | break; | |
3543 | case TCP_SYNACK_FASTOPEN: | |
ca6fb065 ED |
3544 | /* sk is a const pointer, because we want to express multiple |
3545 | * cpu might call us concurrently. | |
3546 | * sk->sk_wmem_alloc in an atomic, we can promote to rw. | |
3547 | */ | |
3548 | skb_set_owner_w(skb, (struct sock *)sk); | |
b3d05147 | 3549 | break; |
ca6fb065 | 3550 | } |
4aea39c1 | 3551 | skb_dst_set(skb, dst); |
1da177e4 | 3552 | |
3541f9e8 | 3553 | mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); |
f5fff5dc | 3554 | |
33ad798c | 3555 | memset(&opts, 0, sizeof(opts)); |
a842fe14 | 3556 | now = tcp_clock_ns(); |
8b5f12d0 | 3557 | #ifdef CONFIG_SYN_COOKIES |
f8ace8d9 | 3558 | if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok)) |
200ecef6 | 3559 | skb->skb_mstamp_ns = cookie_init_timestamp(req, now); |
8b5f12d0 FW |
3560 | else |
3561 | #endif | |
9e450c1e | 3562 | { |
a842fe14 | 3563 | skb->skb_mstamp_ns = now; |
9e450c1e YC |
3564 | if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */ |
3565 | tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb); | |
3566 | } | |
80f03e27 ED |
3567 | |
3568 | #ifdef CONFIG_TCP_MD5SIG | |
3569 | rcu_read_lock(); | |
fd3a154a | 3570 | md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); |
80f03e27 | 3571 | #endif |
58d607d3 | 3572 | skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); |
331fca43 MKL |
3573 | /* bpf program will be interested in the tcp_flags */ |
3574 | TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK; | |
60e2a778 | 3575 | tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, |
331fca43 MKL |
3576 | foc, synack_type, |
3577 | syn_skb) + sizeof(*th); | |
cfb6eeb4 | 3578 | |
aa8223c7 ACM |
3579 | skb_push(skb, tcp_header_size); |
3580 | skb_reset_transport_header(skb); | |
1da177e4 | 3581 | |
ea1627c2 | 3582 | th = (struct tcphdr *)skb->data; |
1da177e4 LT |
3583 | memset(th, 0, sizeof(struct tcphdr)); |
3584 | th->syn = 1; | |
3585 | th->ack = 1; | |
6ac705b1 | 3586 | tcp_ecn_make_synack(req, th); |
b44084c2 | 3587 | th->source = htons(ireq->ir_num); |
634fb979 | 3588 | th->dest = ireq->ir_rmt_port; |
e05a90ec | 3589 | skb->mark = ireq->ir_mark; |
3b117750 ED |
3590 | skb->ip_summed = CHECKSUM_PARTIAL; |
3591 | th->seq = htonl(tcp_rsk(req)->snt_isn); | |
8336886f JC |
3592 | /* XXX data is queued and acked as is. No buffer/window check */ |
3593 | th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); | |
1da177e4 LT |
3594 | |
3595 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ | |
ed53d0ab | 3596 | th->window = htons(min(req->rsk_rcv_wnd, 65535U)); |
5d062de7 | 3597 | tcp_options_write((__be32 *)(th + 1), NULL, &opts); |
1da177e4 | 3598 | th->doff = (tcp_header_size >> 2); |
90bbcc60 | 3599 | __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); |
cfb6eeb4 YH |
3600 | |
3601 | #ifdef CONFIG_TCP_MD5SIG | |
3602 | /* Okay, we have all we need - do the md5 hash if needed */ | |
80f03e27 | 3603 | if (md5) |
bd0388ae | 3604 | tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, |
39f8e58e | 3605 | md5, req_to_sk(req), skb); |
80f03e27 | 3606 | rcu_read_unlock(); |
cfb6eeb4 YH |
3607 | #endif |
3608 | ||
331fca43 MKL |
3609 | bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb, |
3610 | synack_type, &opts); | |
3611 | ||
a842fe14 ED |
3612 | skb->skb_mstamp_ns = now; |
3613 | tcp_add_tx_delay(skb, tp); | |
3614 | ||
1da177e4 LT |
3615 | return skb; |
3616 | } | |
4bc2f18b | 3617 | EXPORT_SYMBOL(tcp_make_synack); |
1da177e4 | 3618 | |
81164413 DB |
3619 | static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) |
3620 | { | |
3621 | struct inet_connection_sock *icsk = inet_csk(sk); | |
3622 | const struct tcp_congestion_ops *ca; | |
3623 | u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); | |
3624 | ||
3625 | if (ca_key == TCP_CA_UNSPEC) | |
3626 | return; | |
3627 | ||
3628 | rcu_read_lock(); | |
3629 | ca = tcp_ca_find_key(ca_key); | |
0baf26b0 MKL |
3630 | if (likely(ca && bpf_try_module_get(ca, ca->owner))) { |
3631 | bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner); | |
81164413 DB |
3632 | icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); |
3633 | icsk->icsk_ca_ops = ca; | |
3634 | } | |
3635 | rcu_read_unlock(); | |
3636 | } | |
3637 | ||
67edfef7 | 3638 | /* Do all connect socket setups that can be done AF independent. */ |
f7e56a76 | 3639 | static void tcp_connect_init(struct sock *sk) |
1da177e4 | 3640 | { |
cf533ea5 | 3641 | const struct dst_entry *dst = __sk_dst_get(sk); |
1da177e4 LT |
3642 | struct tcp_sock *tp = tcp_sk(sk); |
3643 | __u8 rcv_wscale; | |
13d3b1eb | 3644 | u32 rcv_wnd; |
1da177e4 LT |
3645 | |
3646 | /* We'll fix this up when we get a response from the other end. | |
3647 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. | |
3648 | */ | |
5d2ed052 ED |
3649 | tp->tcp_header_len = sizeof(struct tcphdr); |
3650 | if (sock_net(sk)->ipv4.sysctl_tcp_timestamps) | |
3651 | tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; | |
1da177e4 | 3652 | |
cfb6eeb4 | 3653 | #ifdef CONFIG_TCP_MD5SIG |
00db4124 | 3654 | if (tp->af_specific->md5_lookup(sk, sk)) |
cfb6eeb4 YH |
3655 | tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; |
3656 | #endif | |
3657 | ||
1da177e4 LT |
3658 | /* If user gave his TCP_MAXSEG, record it to clamp */ |
3659 | if (tp->rx_opt.user_mss) | |
3660 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
3661 | tp->max_window = 0; | |
5d424d5a | 3662 | tcp_mtup_init(sk); |
1da177e4 LT |
3663 | tcp_sync_mss(sk, dst_mtu(dst)); |
3664 | ||
81164413 DB |
3665 | tcp_ca_dst_init(sk, dst); |
3666 | ||
1da177e4 LT |
3667 | if (!tp->window_clamp) |
3668 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); | |
3541f9e8 | 3669 | tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); |
f5fff5dc | 3670 | |
1da177e4 | 3671 | tcp_initialize_rcv_mss(sk); |
1da177e4 | 3672 | |
e88c64f0 HPP |
3673 | /* limit the window selection if the user enforce a smaller rx buffer */ |
3674 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && | |
3675 | (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) | |
3676 | tp->window_clamp = tcp_full_space(sk); | |
3677 | ||
13d3b1eb LB |
3678 | rcv_wnd = tcp_rwnd_init_bpf(sk); |
3679 | if (rcv_wnd == 0) | |
3680 | rcv_wnd = dst_metric(dst, RTAX_INITRWND); | |
3681 | ||
ceef9ab6 | 3682 | tcp_select_initial_window(sk, tcp_full_space(sk), |
1da177e4 LT |
3683 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), |
3684 | &tp->rcv_wnd, | |
3685 | &tp->window_clamp, | |
9bb37ef0 | 3686 | sock_net(sk)->ipv4.sysctl_tcp_window_scaling, |
31d12926 | 3687 | &rcv_wscale, |
13d3b1eb | 3688 | rcv_wnd); |
1da177e4 LT |
3689 | |
3690 | tp->rx_opt.rcv_wscale = rcv_wscale; | |
3691 | tp->rcv_ssthresh = tp->rcv_wnd; | |
3692 | ||
3693 | sk->sk_err = 0; | |
3694 | sock_reset_flag(sk, SOCK_DONE); | |
3695 | tp->snd_wnd = 0; | |
ee7537b6 | 3696 | tcp_init_wl(tp, 0); |
7f582b24 | 3697 | tcp_write_queue_purge(sk); |
1da177e4 LT |
3698 | tp->snd_una = tp->write_seq; |
3699 | tp->snd_sml = tp->write_seq; | |
33f5f57e | 3700 | tp->snd_up = tp->write_seq; |
e0d694d6 | 3701 | WRITE_ONCE(tp->snd_nxt, tp->write_seq); |
ee995283 PE |
3702 | |
3703 | if (likely(!tp->repair)) | |
3704 | tp->rcv_nxt = 0; | |
c7781a6e | 3705 | else |
70eabf0e | 3706 | tp->rcv_tstamp = tcp_jiffies32; |
ee995283 | 3707 | tp->rcv_wup = tp->rcv_nxt; |
7db48e98 | 3708 | WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); |
1da177e4 | 3709 | |
8550f328 | 3710 | inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); |
463c84b9 | 3711 | inet_csk(sk)->icsk_retransmits = 0; |
1da177e4 LT |
3712 | tcp_clear_retrans(tp); |
3713 | } | |
3714 | ||
783237e8 YC |
3715 | static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) |
3716 | { | |
3717 | struct tcp_sock *tp = tcp_sk(sk); | |
3718 | struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); | |
3719 | ||
3720 | tcb->end_seq += skb->len; | |
f4a775d1 | 3721 | __skb_header_release(skb); |
ab4e846a | 3722 | sk_wmem_queued_add(sk, skb->truesize); |
783237e8 | 3723 | sk_mem_charge(sk, skb->truesize); |
0f317464 | 3724 | WRITE_ONCE(tp->write_seq, tcb->end_seq); |
783237e8 YC |
3725 | tp->packets_out += tcp_skb_pcount(skb); |
3726 | } | |
3727 | ||
3728 | /* Build and send a SYN with data and (cached) Fast Open cookie. However, | |
3729 | * queue a data-only packet after the regular SYN, such that regular SYNs | |
3730 | * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges | |
3731 | * only the SYN sequence, the data are retransmitted in the first ACK. | |
3732 | * If cookie is not cached or other error occurs, falls back to send a | |
3733 | * regular SYN with Fast Open cookie request option. | |
3734 | */ | |
3735 | static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) | |
3736 | { | |
3737 | struct tcp_sock *tp = tcp_sk(sk); | |
3738 | struct tcp_fastopen_request *fo = tp->fastopen_req; | |
065263f4 | 3739 | int space, err = 0; |
355a901e | 3740 | struct sk_buff *syn_data; |
aab48743 | 3741 | |
67da22d2 | 3742 | tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ |
065263f4 | 3743 | if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) |
783237e8 YC |
3744 | goto fallback; |
3745 | ||
3746 | /* MSS for SYN-data is based on cached MSS and bounded by PMTU and | |
3747 | * user-MSS. Reserve maximum option space for middleboxes that add | |
3748 | * private TCP options. The cost is reduced data space in SYN :( | |
3749 | */ | |
3541f9e8 ED |
3750 | tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); |
3751 | ||
1b63edd6 | 3752 | space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - |
783237e8 YC |
3753 | MAX_TCP_OPTION_SPACE; |
3754 | ||
f5ddcbbb ED |
3755 | space = min_t(size_t, space, fo->size); |
3756 | ||
3757 | /* limit to order-0 allocations */ | |
3758 | space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); | |
3759 | ||
eb934478 | 3760 | syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); |
355a901e | 3761 | if (!syn_data) |
783237e8 | 3762 | goto fallback; |
355a901e ED |
3763 | syn_data->ip_summed = CHECKSUM_PARTIAL; |
3764 | memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); | |
07e100f9 ED |
3765 | if (space) { |
3766 | int copied = copy_from_iter(skb_put(syn_data, space), space, | |
3767 | &fo->data->msg_iter); | |
3768 | if (unlikely(!copied)) { | |
ba233b34 | 3769 | tcp_skb_tsorted_anchor_cleanup(syn_data); |
07e100f9 ED |
3770 | kfree_skb(syn_data); |
3771 | goto fallback; | |
3772 | } | |
3773 | if (copied != space) { | |
3774 | skb_trim(syn_data, copied); | |
3775 | space = copied; | |
3776 | } | |
f859a448 | 3777 | skb_zcopy_set(syn_data, fo->uarg, NULL); |
57be5bda | 3778 | } |
355a901e ED |
3779 | /* No more data pending in inet_wait_for_connect() */ |
3780 | if (space == fo->size) | |
3781 | fo->data = NULL; | |
3782 | fo->copied = space; | |
783237e8 | 3783 | |
355a901e | 3784 | tcp_connect_queue_skb(sk, syn_data); |
0f87230d FY |
3785 | if (syn_data->len) |
3786 | tcp_chrono_start(sk, TCP_CHRONO_BUSY); | |
783237e8 | 3787 | |
355a901e | 3788 | err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); |
783237e8 | 3789 | |
d3edd06e | 3790 | syn->skb_mstamp_ns = syn_data->skb_mstamp_ns; |
431a9124 | 3791 | |
355a901e ED |
3792 | /* Now full SYN+DATA was cloned and sent (or not), |
3793 | * remove the SYN from the original skb (syn_data) | |
3794 | * we keep in write queue in case of a retransmit, as we | |
3795 | * also have the SYN packet (with no data) in the same queue. | |
3796 | */ | |
3797 | TCP_SKB_CB(syn_data)->seq++; | |
3798 | TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; | |
3799 | if (!err) { | |
67da22d2 | 3800 | tp->syn_data = (fo->copied > 0); |
75c119af | 3801 | tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); |
f19c29e3 | 3802 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); |
783237e8 YC |
3803 | goto done; |
3804 | } | |
783237e8 | 3805 | |
75c119af ED |
3806 | /* data was not sent, put it in write_queue */ |
3807 | __skb_queue_tail(&sk->sk_write_queue, syn_data); | |
b5b7db8d ED |
3808 | tp->packets_out -= tcp_skb_pcount(syn_data); |
3809 | ||
783237e8 YC |
3810 | fallback: |
3811 | /* Send a regular SYN with Fast Open cookie request option */ | |
3812 | if (fo->cookie.len > 0) | |
3813 | fo->cookie.len = 0; | |
3814 | err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); | |
3815 | if (err) | |
3816 | tp->syn_fastopen = 0; | |
783237e8 YC |
3817 | done: |
3818 | fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ | |
3819 | return err; | |
3820 | } | |
3821 | ||
67edfef7 | 3822 | /* Build a SYN and send it off. */ |
1da177e4 LT |
3823 | int tcp_connect(struct sock *sk) |
3824 | { | |
3825 | struct tcp_sock *tp = tcp_sk(sk); | |
3826 | struct sk_buff *buff; | |
ee586811 | 3827 | int err; |
1da177e4 | 3828 | |
de525be2 | 3829 | tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); |
8ba60924 ED |
3830 | |
3831 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) | |
3832 | return -EHOSTUNREACH; /* Routing failure or similar. */ | |
3833 | ||
1da177e4 LT |
3834 | tcp_connect_init(sk); |
3835 | ||
2b916477 AV |
3836 | if (unlikely(tp->repair)) { |
3837 | tcp_finish_connect(sk, NULL); | |
3838 | return 0; | |
3839 | } | |
3840 | ||
eb934478 | 3841 | buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); |
355a901e | 3842 | if (unlikely(!buff)) |
1da177e4 LT |
3843 | return -ENOBUFS; |
3844 | ||
a3433f35 | 3845 | tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); |
9a568de4 ED |
3846 | tcp_mstamp_refresh(tp); |
3847 | tp->retrans_stamp = tcp_time_stamp(tp); | |
783237e8 | 3848 | tcp_connect_queue_skb(sk, buff); |
735d3831 | 3849 | tcp_ecn_send_syn(sk, buff); |
75c119af | 3850 | tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); |
1da177e4 | 3851 | |
783237e8 YC |
3852 | /* Send off SYN; include data in Fast Open. */ |
3853 | err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : | |
3854 | tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); | |
ee586811 EP |
3855 | if (err == -ECONNREFUSED) |
3856 | return err; | |
bd37a088 WY |
3857 | |
3858 | /* We change tp->snd_nxt after the tcp_transmit_skb() call | |
3859 | * in order to make this packet get counted in tcpOutSegs. | |
3860 | */ | |
e0d694d6 | 3861 | WRITE_ONCE(tp->snd_nxt, tp->write_seq); |
bd37a088 | 3862 | tp->pushed_seq = tp->write_seq; |
b5b7db8d ED |
3863 | buff = tcp_send_head(sk); |
3864 | if (unlikely(buff)) { | |
e0d694d6 | 3865 | WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq); |
b5b7db8d ED |
3866 | tp->pushed_seq = TCP_SKB_CB(buff)->seq; |
3867 | } | |
81cc8a75 | 3868 | TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); |
1da177e4 LT |
3869 | |
3870 | /* Timer for repeating the SYN until an answer. */ | |
3f421baa ACM |
3871 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
3872 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
1da177e4 LT |
3873 | return 0; |
3874 | } | |
4bc2f18b | 3875 | EXPORT_SYMBOL(tcp_connect); |
1da177e4 LT |
3876 | |
3877 | /* Send out a delayed ack, the caller does the policy checking | |
3878 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() | |
3879 | * for details. | |
3880 | */ | |
3881 | void tcp_send_delayed_ack(struct sock *sk) | |
3882 | { | |
463c84b9 ACM |
3883 | struct inet_connection_sock *icsk = inet_csk(sk); |
3884 | int ato = icsk->icsk_ack.ato; | |
1da177e4 LT |
3885 | unsigned long timeout; |
3886 | ||
3887 | if (ato > TCP_DELACK_MIN) { | |
463c84b9 | 3888 | const struct tcp_sock *tp = tcp_sk(sk); |
056834d9 | 3889 | int max_ato = HZ / 2; |
1da177e4 | 3890 | |
31954cd8 | 3891 | if (inet_csk_in_pingpong_mode(sk) || |
056834d9 | 3892 | (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) |
1da177e4 LT |
3893 | max_ato = TCP_DELACK_MAX; |
3894 | ||
3895 | /* Slow path, intersegment interval is "high". */ | |
3896 | ||
3897 | /* If some rtt estimate is known, use it to bound delayed ack. | |
463c84b9 | 3898 | * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements |
1da177e4 LT |
3899 | * directly. |
3900 | */ | |
740b0f18 ED |
3901 | if (tp->srtt_us) { |
3902 | int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), | |
3903 | TCP_DELACK_MIN); | |
1da177e4 LT |
3904 | |
3905 | if (rtt < max_ato) | |
3906 | max_ato = rtt; | |
3907 | } | |
3908 | ||
3909 | ato = min(ato, max_ato); | |
3910 | } | |
3911 | ||
2b8ee4f0 MKL |
3912 | ato = min_t(u32, ato, inet_csk(sk)->icsk_delack_max); |
3913 | ||
1da177e4 LT |
3914 | /* Stay within the limit we were given */ |
3915 | timeout = jiffies + ato; | |
3916 | ||
3917 | /* Use new timeout only if there wasn't a older one earlier. */ | |
463c84b9 | 3918 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { |
b6b6d653 ED |
3919 | /* If delack timer is about to expire, send ACK now. */ |
3920 | if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { | |
1da177e4 LT |
3921 | tcp_send_ack(sk); |
3922 | return; | |
3923 | } | |
3924 | ||
463c84b9 ACM |
3925 | if (!time_before(timeout, icsk->icsk_ack.timeout)) |
3926 | timeout = icsk->icsk_ack.timeout; | |
1da177e4 | 3927 | } |
463c84b9 ACM |
3928 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; |
3929 | icsk->icsk_ack.timeout = timeout; | |
3930 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); | |
1da177e4 LT |
3931 | } |
3932 | ||
3933 | /* This routine sends an ack and also updates the window. */ | |
2987babb | 3934 | void __tcp_send_ack(struct sock *sk, u32 rcv_nxt) |
1da177e4 | 3935 | { |
058dc334 | 3936 | struct sk_buff *buff; |
1da177e4 | 3937 | |
058dc334 IJ |
3938 | /* If we have been reset, we may not send again. */ |
3939 | if (sk->sk_state == TCP_CLOSE) | |
3940 | return; | |
1da177e4 | 3941 | |
058dc334 IJ |
3942 | /* We are not putting this on the write queue, so |
3943 | * tcp_transmit_skb() will set the ownership to this | |
3944 | * sock. | |
3945 | */ | |
7450aaf6 ED |
3946 | buff = alloc_skb(MAX_TCP_HEADER, |
3947 | sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); | |
3948 | if (unlikely(!buff)) { | |
a37c2134 ED |
3949 | struct inet_connection_sock *icsk = inet_csk(sk); |
3950 | unsigned long delay; | |
3951 | ||
3952 | delay = TCP_DELACK_MAX << icsk->icsk_ack.retry; | |
3953 | if (delay < TCP_RTO_MAX) | |
3954 | icsk->icsk_ack.retry++; | |
058dc334 | 3955 | inet_csk_schedule_ack(sk); |
a37c2134 ED |
3956 | icsk->icsk_ack.ato = TCP_ATO_MIN; |
3957 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX); | |
058dc334 | 3958 | return; |
1da177e4 | 3959 | } |
058dc334 IJ |
3960 | |
3961 | /* Reserve space for headers and prepare control bits. */ | |
3962 | skb_reserve(buff, MAX_TCP_HEADER); | |
a3433f35 | 3963 | tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); |
058dc334 | 3964 | |
98781965 ED |
3965 | /* We do not want pure acks influencing TCP Small Queues or fq/pacing |
3966 | * too much. | |
3967 | * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 | |
98781965 ED |
3968 | */ |
3969 | skb_set_tcp_pure_ack(buff); | |
3970 | ||
058dc334 | 3971 | /* Send it off, this clears delayed acks for us. */ |
2987babb YC |
3972 | __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt); |
3973 | } | |
27cde44a | 3974 | EXPORT_SYMBOL_GPL(__tcp_send_ack); |
2987babb YC |
3975 | |
3976 | void tcp_send_ack(struct sock *sk) | |
3977 | { | |
3978 | __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt); | |
1da177e4 LT |
3979 | } |
3980 | ||
3981 | /* This routine sends a packet with an out of date sequence | |
3982 | * number. It assumes the other end will try to ack it. | |
3983 | * | |
3984 | * Question: what should we make while urgent mode? | |
3985 | * 4.4BSD forces sending single byte of data. We cannot send | |
3986 | * out of window data, because we have SND.NXT==SND.MAX... | |
3987 | * | |
3988 | * Current solution: to send TWO zero-length segments in urgent mode: | |
3989 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is | |
3990 | * out-of-date with SND.UNA-1 to probe window. | |
3991 | */ | |
e520af48 | 3992 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) |
1da177e4 LT |
3993 | { |
3994 | struct tcp_sock *tp = tcp_sk(sk); | |
3995 | struct sk_buff *skb; | |
3996 | ||
3997 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ | |
7450aaf6 ED |
3998 | skb = alloc_skb(MAX_TCP_HEADER, |
3999 | sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); | |
51456b29 | 4000 | if (!skb) |
1da177e4 LT |
4001 | return -1; |
4002 | ||
4003 | /* Reserve space for headers and set control bits. */ | |
4004 | skb_reserve(skb, MAX_TCP_HEADER); | |
1da177e4 LT |
4005 | /* Use a previous sequence. This should cause the other |
4006 | * end to send an ack. Don't queue or clone SKB, just | |
4007 | * send it. | |
4008 | */ | |
a3433f35 | 4009 | tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); |
e2e8009f | 4010 | NET_INC_STATS(sock_net(sk), mib); |
7450aaf6 | 4011 | return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); |
1da177e4 LT |
4012 | } |
4013 | ||
385e2070 | 4014 | /* Called from setsockopt( ... TCP_REPAIR ) */ |
ee995283 PE |
4015 | void tcp_send_window_probe(struct sock *sk) |
4016 | { | |
4017 | if (sk->sk_state == TCP_ESTABLISHED) { | |
4018 | tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; | |
9a568de4 | 4019 | tcp_mstamp_refresh(tcp_sk(sk)); |
e520af48 | 4020 | tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); |
ee995283 PE |
4021 | } |
4022 | } | |
4023 | ||
67edfef7 | 4024 | /* Initiate keepalive or window probe from timer. */ |
e520af48 | 4025 | int tcp_write_wakeup(struct sock *sk, int mib) |
1da177e4 | 4026 | { |
058dc334 IJ |
4027 | struct tcp_sock *tp = tcp_sk(sk); |
4028 | struct sk_buff *skb; | |
1da177e4 | 4029 | |
058dc334 IJ |
4030 | if (sk->sk_state == TCP_CLOSE) |
4031 | return -1; | |
4032 | ||
00db4124 IM |
4033 | skb = tcp_send_head(sk); |
4034 | if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { | |
058dc334 | 4035 | int err; |
0c54b85f | 4036 | unsigned int mss = tcp_current_mss(sk); |
058dc334 IJ |
4037 | unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
4038 | ||
4039 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) | |
4040 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; | |
4041 | ||
4042 | /* We are probing the opening of a window | |
4043 | * but the window size is != 0 | |
4044 | * must have been a result SWS avoidance ( sender ) | |
4045 | */ | |
4046 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || | |
4047 | skb->len > mss) { | |
4048 | seg_size = min(seg_size, mss); | |
4de075e0 | 4049 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
75c119af ED |
4050 | if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, |
4051 | skb, seg_size, mss, GFP_ATOMIC)) | |
058dc334 IJ |
4052 | return -1; |
4053 | } else if (!tcp_skb_pcount(skb)) | |
5bbb432c | 4054 | tcp_set_skb_tso_segs(skb, mss); |
058dc334 | 4055 | |
4de075e0 | 4056 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
058dc334 IJ |
4057 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
4058 | if (!err) | |
4059 | tcp_event_new_data_sent(sk, skb); | |
4060 | return err; | |
4061 | } else { | |
33f5f57e | 4062 | if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) |
e520af48 ED |
4063 | tcp_xmit_probe_skb(sk, 1, mib); |
4064 | return tcp_xmit_probe_skb(sk, 0, mib); | |
1da177e4 | 4065 | } |
1da177e4 LT |
4066 | } |
4067 | ||
4068 | /* A window probe timeout has occurred. If window is not closed send | |
4069 | * a partial packet else a zero probe. | |
4070 | */ | |
4071 | void tcp_send_probe0(struct sock *sk) | |
4072 | { | |
463c84b9 | 4073 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 4074 | struct tcp_sock *tp = tcp_sk(sk); |
c6214a97 | 4075 | struct net *net = sock_net(sk); |
c1d5674f | 4076 | unsigned long timeout; |
1da177e4 LT |
4077 | int err; |
4078 | ||
e520af48 | 4079 | err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); |
1da177e4 | 4080 | |
75c119af | 4081 | if (tp->packets_out || tcp_write_queue_empty(sk)) { |
1da177e4 | 4082 | /* Cancel probe timer, if it is not required. */ |
6687e988 | 4083 | icsk->icsk_probes_out = 0; |
463c84b9 | 4084 | icsk->icsk_backoff = 0; |
9d9b1ee0 | 4085 | icsk->icsk_probes_tstamp = 0; |
1da177e4 LT |
4086 | return; |
4087 | } | |
4088 | ||
c1d5674f | 4089 | icsk->icsk_probes_out++; |
1da177e4 | 4090 | if (err <= 0) { |
c6214a97 | 4091 | if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) |
463c84b9 | 4092 | icsk->icsk_backoff++; |
c1d5674f | 4093 | timeout = tcp_probe0_when(sk, TCP_RTO_MAX); |
1da177e4 LT |
4094 | } else { |
4095 | /* If packet was not sent due to local congestion, | |
c1d5674f | 4096 | * Let senders fight for local resources conservatively. |
1da177e4 | 4097 | */ |
c1d5674f YC |
4098 | timeout = TCP_RESOURCE_PROBE_INTERVAL; |
4099 | } | |
344db93a EC |
4100 | |
4101 | timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout); | |
8dc242ad | 4102 | tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX); |
1da177e4 | 4103 | } |
5db92c99 | 4104 | |
ea3bea3a | 4105 | int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) |
5db92c99 OP |
4106 | { |
4107 | const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; | |
4108 | struct flowi fl; | |
4109 | int res; | |
4110 | ||
58d607d3 | 4111 | tcp_rsk(req)->txhash = net_tx_rndhash(); |
331fca43 MKL |
4112 | res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL, |
4113 | NULL); | |
5db92c99 | 4114 | if (!res) { |
90bbcc60 | 4115 | __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); |
02a1d6e7 | 4116 | __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); |
7e32b443 YC |
4117 | if (unlikely(tcp_passive_fastopen(sk))) |
4118 | tcp_sk(sk)->total_retrans++; | |
cf34ce3d | 4119 | trace_tcp_retransmit_synack(sk, req); |
5db92c99 OP |
4120 | } |
4121 | return res; | |
4122 | } | |
4123 | EXPORT_SYMBOL(tcp_rtx_synack); |