]>
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 */ |
bd0388ae | 441 | __u8 *hash_location; /* temporary pointer, overloaded */ |
2100c8d2 YC |
442 | __u32 tsval, tsecr; /* need to include OPTION_TS */ |
443 | struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */ | |
eda7acdd | 444 | struct mptcp_out_options mptcp; |
33ad798c AL |
445 | }; |
446 | ||
eda7acdd PK |
447 | static void mptcp_options_write(__be32 *ptr, struct tcp_out_options *opts) |
448 | { | |
449 | #if IS_ENABLED(CONFIG_MPTCP) | |
450 | if (unlikely(OPTION_MPTCP & opts->options)) | |
451 | mptcp_write_options(ptr, &opts->mptcp); | |
452 | #endif | |
453 | } | |
454 | ||
67edfef7 AK |
455 | /* Write previously computed TCP options to the packet. |
456 | * | |
457 | * Beware: Something in the Internet is very sensitive to the ordering of | |
fd6149d3 IJ |
458 | * TCP options, we learned this through the hard way, so be careful here. |
459 | * Luckily we can at least blame others for their non-compliance but from | |
8e3bff96 | 460 | * inter-operability perspective it seems that we're somewhat stuck with |
fd6149d3 IJ |
461 | * the ordering which we have been using if we want to keep working with |
462 | * those broken things (not that it currently hurts anybody as there isn't | |
463 | * particular reason why the ordering would need to be changed). | |
464 | * | |
465 | * At least SACK_PERM as the first option is known to lead to a disaster | |
466 | * (but it may well be that other scenarios fail similarly). | |
467 | */ | |
33ad798c | 468 | static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp, |
bd0388ae WAS |
469 | struct tcp_out_options *opts) |
470 | { | |
2100c8d2 | 471 | u16 options = opts->options; /* mungable copy */ |
bd0388ae | 472 | |
bd0388ae | 473 | if (unlikely(OPTION_MD5 & options)) { |
1a2c6181 CP |
474 | *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | |
475 | (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG); | |
bd0388ae WAS |
476 | /* overload cookie hash location */ |
477 | opts->hash_location = (__u8 *)ptr; | |
33ad798c | 478 | ptr += 4; |
40efc6fa | 479 | } |
33ad798c | 480 | |
fd6149d3 IJ |
481 | if (unlikely(opts->mss)) { |
482 | *ptr++ = htonl((TCPOPT_MSS << 24) | | |
483 | (TCPOLEN_MSS << 16) | | |
484 | opts->mss); | |
485 | } | |
486 | ||
bd0388ae WAS |
487 | if (likely(OPTION_TS & options)) { |
488 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { | |
33ad798c AL |
489 | *ptr++ = htonl((TCPOPT_SACK_PERM << 24) | |
490 | (TCPOLEN_SACK_PERM << 16) | | |
491 | (TCPOPT_TIMESTAMP << 8) | | |
492 | TCPOLEN_TIMESTAMP); | |
bd0388ae | 493 | options &= ~OPTION_SACK_ADVERTISE; |
33ad798c AL |
494 | } else { |
495 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
496 | (TCPOPT_NOP << 16) | | |
497 | (TCPOPT_TIMESTAMP << 8) | | |
498 | TCPOLEN_TIMESTAMP); | |
499 | } | |
500 | *ptr++ = htonl(opts->tsval); | |
501 | *ptr++ = htonl(opts->tsecr); | |
502 | } | |
503 | ||
bd0388ae | 504 | if (unlikely(OPTION_SACK_ADVERTISE & options)) { |
33ad798c AL |
505 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
506 | (TCPOPT_NOP << 16) | | |
507 | (TCPOPT_SACK_PERM << 8) | | |
508 | TCPOLEN_SACK_PERM); | |
509 | } | |
510 | ||
bd0388ae | 511 | if (unlikely(OPTION_WSCALE & options)) { |
33ad798c AL |
512 | *ptr++ = htonl((TCPOPT_NOP << 24) | |
513 | (TCPOPT_WINDOW << 16) | | |
514 | (TCPOLEN_WINDOW << 8) | | |
515 | opts->ws); | |
516 | } | |
517 | ||
518 | if (unlikely(opts->num_sack_blocks)) { | |
519 | struct tcp_sack_block *sp = tp->rx_opt.dsack ? | |
520 | tp->duplicate_sack : tp->selective_acks; | |
40efc6fa SH |
521 | int this_sack; |
522 | ||
523 | *ptr++ = htonl((TCPOPT_NOP << 24) | | |
524 | (TCPOPT_NOP << 16) | | |
525 | (TCPOPT_SACK << 8) | | |
33ad798c | 526 | (TCPOLEN_SACK_BASE + (opts->num_sack_blocks * |
40efc6fa | 527 | TCPOLEN_SACK_PERBLOCK))); |
2de979bd | 528 | |
33ad798c AL |
529 | for (this_sack = 0; this_sack < opts->num_sack_blocks; |
530 | ++this_sack) { | |
40efc6fa SH |
531 | *ptr++ = htonl(sp[this_sack].start_seq); |
532 | *ptr++ = htonl(sp[this_sack].end_seq); | |
533 | } | |
2de979bd | 534 | |
5861f8e5 | 535 | tp->rx_opt.dsack = 0; |
40efc6fa | 536 | } |
2100c8d2 YC |
537 | |
538 | if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) { | |
539 | struct tcp_fastopen_cookie *foc = opts->fastopen_cookie; | |
7f9b838b DL |
540 | u8 *p = (u8 *)ptr; |
541 | u32 len; /* Fast Open option length */ | |
542 | ||
543 | if (foc->exp) { | |
544 | len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len; | |
545 | *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) | | |
546 | TCPOPT_FASTOPEN_MAGIC); | |
547 | p += TCPOLEN_EXP_FASTOPEN_BASE; | |
548 | } else { | |
549 | len = TCPOLEN_FASTOPEN_BASE + foc->len; | |
550 | *p++ = TCPOPT_FASTOPEN; | |
551 | *p++ = len; | |
552 | } | |
2100c8d2 | 553 | |
7f9b838b DL |
554 | memcpy(p, foc->val, foc->len); |
555 | if ((len & 3) == 2) { | |
556 | p[foc->len] = TCPOPT_NOP; | |
557 | p[foc->len + 1] = TCPOPT_NOP; | |
2100c8d2 | 558 | } |
7f9b838b | 559 | ptr += (len + 3) >> 2; |
2100c8d2 | 560 | } |
60e2a778 UB |
561 | |
562 | smc_options_write(ptr, &options); | |
eda7acdd PK |
563 | |
564 | mptcp_options_write(ptr, opts); | |
60e2a778 UB |
565 | } |
566 | ||
567 | static void smc_set_option(const struct tcp_sock *tp, | |
568 | struct tcp_out_options *opts, | |
569 | unsigned int *remaining) | |
570 | { | |
571 | #if IS_ENABLED(CONFIG_SMC) | |
572 | if (static_branch_unlikely(&tcp_have_smc)) { | |
573 | if (tp->syn_smc) { | |
574 | if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { | |
575 | opts->options |= OPTION_SMC; | |
576 | *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; | |
577 | } | |
578 | } | |
579 | } | |
580 | #endif | |
581 | } | |
582 | ||
583 | static void smc_set_option_cond(const struct tcp_sock *tp, | |
584 | const struct inet_request_sock *ireq, | |
585 | struct tcp_out_options *opts, | |
586 | unsigned int *remaining) | |
587 | { | |
588 | #if IS_ENABLED(CONFIG_SMC) | |
589 | if (static_branch_unlikely(&tcp_have_smc)) { | |
590 | if (tp->syn_smc && ireq->smc_ok) { | |
591 | if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) { | |
592 | opts->options |= OPTION_SMC; | |
593 | *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED; | |
594 | } | |
595 | } | |
596 | } | |
597 | #endif | |
33ad798c AL |
598 | } |
599 | ||
cec37a6e PK |
600 | static void mptcp_set_option_cond(const struct request_sock *req, |
601 | struct tcp_out_options *opts, | |
602 | unsigned int *remaining) | |
603 | { | |
604 | if (rsk_is_mptcp(req)) { | |
605 | unsigned int size; | |
606 | ||
607 | if (mptcp_synack_options(req, &size, &opts->mptcp)) { | |
608 | if (*remaining >= size) { | |
609 | opts->options |= OPTION_MPTCP; | |
610 | *remaining -= size; | |
611 | } | |
612 | } | |
613 | } | |
614 | } | |
615 | ||
67edfef7 AK |
616 | /* Compute TCP options for SYN packets. This is not the final |
617 | * network wire format yet. | |
618 | */ | |
95c96174 | 619 | static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb, |
33ad798c | 620 | struct tcp_out_options *opts, |
cf533ea5 ED |
621 | struct tcp_md5sig_key **md5) |
622 | { | |
33ad798c | 623 | struct tcp_sock *tp = tcp_sk(sk); |
95c96174 | 624 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
783237e8 | 625 | struct tcp_fastopen_request *fastopen = tp->fastopen_req; |
33ad798c | 626 | |
8c2320e8 | 627 | *md5 = NULL; |
cfb6eeb4 | 628 | #ifdef CONFIG_TCP_MD5SIG |
921f9a0f | 629 | if (static_branch_unlikely(&tcp_md5_needed) && |
6015c71e | 630 | rcu_access_pointer(tp->md5sig_info)) { |
8c2320e8 ED |
631 | *md5 = tp->af_specific->md5_lookup(sk, sk); |
632 | if (*md5) { | |
633 | opts->options |= OPTION_MD5; | |
634 | remaining -= TCPOLEN_MD5SIG_ALIGNED; | |
635 | } | |
cfb6eeb4 YH |
636 | } |
637 | #endif | |
33ad798c AL |
638 | |
639 | /* We always get an MSS option. The option bytes which will be seen in | |
640 | * normal data packets should timestamps be used, must be in the MSS | |
641 | * advertised. But we subtract them from tp->mss_cache so that | |
642 | * calculations in tcp_sendmsg are simpler etc. So account for this | |
643 | * fact here if necessary. If we don't do this correctly, as a | |
644 | * receiver we won't recognize data packets as being full sized when we | |
645 | * should, and thus we won't abide by the delayed ACK rules correctly. | |
646 | * SACKs don't matter, we never delay an ACK when we have any of those | |
647 | * going out. */ | |
648 | opts->mss = tcp_advertise_mss(sk); | |
bd0388ae | 649 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c | 650 | |
5d2ed052 | 651 | if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) { |
33ad798c | 652 | opts->options |= OPTION_TS; |
7faee5c0 | 653 | opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset; |
33ad798c | 654 | opts->tsecr = tp->rx_opt.ts_recent; |
bd0388ae | 655 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c | 656 | } |
9bb37ef0 | 657 | if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) { |
33ad798c | 658 | opts->ws = tp->rx_opt.rcv_wscale; |
89e95a61 | 659 | opts->options |= OPTION_WSCALE; |
bd0388ae | 660 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 661 | } |
f9301034 | 662 | if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) { |
33ad798c | 663 | opts->options |= OPTION_SACK_ADVERTISE; |
b32d1310 | 664 | if (unlikely(!(OPTION_TS & opts->options))) |
bd0388ae | 665 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c AL |
666 | } |
667 | ||
783237e8 | 668 | if (fastopen && fastopen->cookie.len >= 0) { |
2646c831 DL |
669 | u32 need = fastopen->cookie.len; |
670 | ||
671 | need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE : | |
672 | TCPOLEN_FASTOPEN_BASE; | |
783237e8 YC |
673 | need = (need + 3) & ~3U; /* Align to 32 bits */ |
674 | if (remaining >= need) { | |
675 | opts->options |= OPTION_FAST_OPEN_COOKIE; | |
676 | opts->fastopen_cookie = &fastopen->cookie; | |
677 | remaining -= need; | |
678 | tp->syn_fastopen = 1; | |
2646c831 | 679 | tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0; |
783237e8 YC |
680 | } |
681 | } | |
bd0388ae | 682 | |
60e2a778 UB |
683 | smc_set_option(tp, opts, &remaining); |
684 | ||
cec37a6e PK |
685 | if (sk_is_mptcp(sk)) { |
686 | unsigned int size; | |
687 | ||
cc7972ea | 688 | if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) { |
cec37a6e PK |
689 | opts->options |= OPTION_MPTCP; |
690 | remaining -= size; | |
691 | } | |
692 | } | |
693 | ||
bd0388ae | 694 | return MAX_TCP_OPTION_SPACE - remaining; |
40efc6fa SH |
695 | } |
696 | ||
67edfef7 | 697 | /* Set up TCP options for SYN-ACKs. */ |
60e2a778 UB |
698 | static unsigned int tcp_synack_options(const struct sock *sk, |
699 | struct request_sock *req, | |
37bfbdda ED |
700 | unsigned int mss, struct sk_buff *skb, |
701 | struct tcp_out_options *opts, | |
702 | const struct tcp_md5sig_key *md5, | |
e114e1e8 ED |
703 | struct tcp_fastopen_cookie *foc, |
704 | enum tcp_synack_type synack_type) | |
4957faad | 705 | { |
33ad798c | 706 | struct inet_request_sock *ireq = inet_rsk(req); |
95c96174 | 707 | unsigned int remaining = MAX_TCP_OPTION_SPACE; |
33ad798c | 708 | |
cfb6eeb4 | 709 | #ifdef CONFIG_TCP_MD5SIG |
80f03e27 | 710 | if (md5) { |
33ad798c | 711 | opts->options |= OPTION_MD5; |
4957faad WAS |
712 | remaining -= TCPOLEN_MD5SIG_ALIGNED; |
713 | ||
714 | /* We can't fit any SACK blocks in a packet with MD5 + TS | |
715 | * options. There was discussion about disabling SACK | |
716 | * rather than TS in order to fit in better with old, | |
717 | * buggy kernels, but that was deemed to be unnecessary. | |
718 | */ | |
e114e1e8 ED |
719 | if (synack_type != TCP_SYNACK_COOKIE) |
720 | ireq->tstamp_ok &= !ireq->sack_ok; | |
cfb6eeb4 YH |
721 | } |
722 | #endif | |
33ad798c | 723 | |
4957faad | 724 | /* We always send an MSS option. */ |
33ad798c | 725 | opts->mss = mss; |
4957faad | 726 | remaining -= TCPOLEN_MSS_ALIGNED; |
33ad798c AL |
727 | |
728 | if (likely(ireq->wscale_ok)) { | |
729 | opts->ws = ireq->rcv_wscale; | |
89e95a61 | 730 | opts->options |= OPTION_WSCALE; |
4957faad | 731 | remaining -= TCPOLEN_WSCALE_ALIGNED; |
33ad798c | 732 | } |
de213e5e | 733 | if (likely(ireq->tstamp_ok)) { |
33ad798c | 734 | opts->options |= OPTION_TS; |
95a22cae | 735 | opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off; |
33ad798c | 736 | opts->tsecr = req->ts_recent; |
4957faad | 737 | remaining -= TCPOLEN_TSTAMP_ALIGNED; |
33ad798c AL |
738 | } |
739 | if (likely(ireq->sack_ok)) { | |
740 | opts->options |= OPTION_SACK_ADVERTISE; | |
de213e5e | 741 | if (unlikely(!ireq->tstamp_ok)) |
4957faad | 742 | remaining -= TCPOLEN_SACKPERM_ALIGNED; |
33ad798c | 743 | } |
7f9b838b DL |
744 | if (foc != NULL && foc->len >= 0) { |
745 | u32 need = foc->len; | |
746 | ||
747 | need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE : | |
748 | TCPOLEN_FASTOPEN_BASE; | |
8336886f JC |
749 | need = (need + 3) & ~3U; /* Align to 32 bits */ |
750 | if (remaining >= need) { | |
751 | opts->options |= OPTION_FAST_OPEN_COOKIE; | |
752 | opts->fastopen_cookie = foc; | |
753 | remaining -= need; | |
754 | } | |
755 | } | |
1a2c6181 | 756 | |
cec37a6e PK |
757 | mptcp_set_option_cond(req, opts, &remaining); |
758 | ||
60e2a778 UB |
759 | smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining); |
760 | ||
4957faad | 761 | return MAX_TCP_OPTION_SPACE - remaining; |
33ad798c AL |
762 | } |
763 | ||
67edfef7 AK |
764 | /* Compute TCP options for ESTABLISHED sockets. This is not the |
765 | * final wire format yet. | |
766 | */ | |
95c96174 | 767 | static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb, |
33ad798c | 768 | struct tcp_out_options *opts, |
cf533ea5 ED |
769 | struct tcp_md5sig_key **md5) |
770 | { | |
33ad798c | 771 | struct tcp_sock *tp = tcp_sk(sk); |
95c96174 | 772 | unsigned int size = 0; |
cabeccbd | 773 | unsigned int eff_sacks; |
33ad798c | 774 | |
5843ef42 AK |
775 | opts->options = 0; |
776 | ||
8c2320e8 | 777 | *md5 = NULL; |
33ad798c | 778 | #ifdef CONFIG_TCP_MD5SIG |
921f9a0f | 779 | if (static_branch_unlikely(&tcp_md5_needed) && |
6015c71e | 780 | rcu_access_pointer(tp->md5sig_info)) { |
8c2320e8 ED |
781 | *md5 = tp->af_specific->md5_lookup(sk, sk); |
782 | if (*md5) { | |
783 | opts->options |= OPTION_MD5; | |
784 | size += TCPOLEN_MD5SIG_ALIGNED; | |
785 | } | |
33ad798c | 786 | } |
33ad798c AL |
787 | #endif |
788 | ||
789 | if (likely(tp->rx_opt.tstamp_ok)) { | |
790 | opts->options |= OPTION_TS; | |
7faee5c0 | 791 | opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0; |
33ad798c AL |
792 | opts->tsecr = tp->rx_opt.ts_recent; |
793 | size += TCPOLEN_TSTAMP_ALIGNED; | |
794 | } | |
795 | ||
cec37a6e PK |
796 | /* MPTCP options have precedence over SACK for the limited TCP |
797 | * option space because a MPTCP connection would be forced to | |
798 | * fall back to regular TCP if a required multipath option is | |
799 | * missing. SACK still gets a chance to use whatever space is | |
800 | * left. | |
801 | */ | |
802 | if (sk_is_mptcp(sk)) { | |
803 | unsigned int remaining = MAX_TCP_OPTION_SPACE - size; | |
804 | unsigned int opt_size = 0; | |
805 | ||
806 | if (mptcp_established_options(sk, skb, &opt_size, remaining, | |
807 | &opts->mptcp)) { | |
808 | opts->options |= OPTION_MPTCP; | |
809 | size += opt_size; | |
810 | } | |
811 | } | |
812 | ||
cabeccbd IJ |
813 | eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack; |
814 | if (unlikely(eff_sacks)) { | |
95c96174 | 815 | const unsigned int remaining = MAX_TCP_OPTION_SPACE - size; |
9cfcca23 MM |
816 | if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED + |
817 | TCPOLEN_SACK_PERBLOCK)) | |
818 | return size; | |
819 | ||
33ad798c | 820 | opts->num_sack_blocks = |
95c96174 | 821 | min_t(unsigned int, eff_sacks, |
33ad798c AL |
822 | (remaining - TCPOLEN_SACK_BASE_ALIGNED) / |
823 | TCPOLEN_SACK_PERBLOCK); | |
9cfcca23 MM |
824 | |
825 | size += TCPOLEN_SACK_BASE_ALIGNED + | |
826 | opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK; | |
33ad798c AL |
827 | } |
828 | ||
829 | return size; | |
40efc6fa | 830 | } |
1da177e4 | 831 | |
46d3ceab ED |
832 | |
833 | /* TCP SMALL QUEUES (TSQ) | |
834 | * | |
835 | * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev) | |
836 | * to reduce RTT and bufferbloat. | |
837 | * We do this using a special skb destructor (tcp_wfree). | |
838 | * | |
839 | * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb | |
840 | * needs to be reallocated in a driver. | |
8e3bff96 | 841 | * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc |
46d3ceab ED |
842 | * |
843 | * Since transmit from skb destructor is forbidden, we use a tasklet | |
844 | * to process all sockets that eventually need to send more skbs. | |
845 | * We use one tasklet per cpu, with its own queue of sockets. | |
846 | */ | |
847 | struct tsq_tasklet { | |
848 | struct tasklet_struct tasklet; | |
849 | struct list_head head; /* queue of tcp sockets */ | |
850 | }; | |
851 | static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet); | |
852 | ||
73a6bab5 | 853 | static void tcp_tsq_write(struct sock *sk) |
6f458dfb ED |
854 | { |
855 | if ((1 << sk->sk_state) & | |
856 | (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING | | |
f9616c35 ED |
857 | TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) { |
858 | struct tcp_sock *tp = tcp_sk(sk); | |
859 | ||
860 | if (tp->lost_out > tp->retrans_out && | |
3a91d29f KD |
861 | tp->snd_cwnd > tcp_packets_in_flight(tp)) { |
862 | tcp_mstamp_refresh(tp); | |
f9616c35 | 863 | tcp_xmit_retransmit_queue(sk); |
3a91d29f | 864 | } |
f9616c35 ED |
865 | |
866 | tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle, | |
bf06200e | 867 | 0, GFP_ATOMIC); |
f9616c35 | 868 | } |
6f458dfb | 869 | } |
73a6bab5 ED |
870 | |
871 | static void tcp_tsq_handler(struct sock *sk) | |
872 | { | |
873 | bh_lock_sock(sk); | |
874 | if (!sock_owned_by_user(sk)) | |
875 | tcp_tsq_write(sk); | |
876 | else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) | |
877 | sock_hold(sk); | |
878 | bh_unlock_sock(sk); | |
879 | } | |
46d3ceab | 880 | /* |
8e3bff96 | 881 | * One tasklet per cpu tries to send more skbs. |
46d3ceab | 882 | * We run in tasklet context but need to disable irqs when |
8e3bff96 | 883 | * transferring tsq->head because tcp_wfree() might |
46d3ceab ED |
884 | * interrupt us (non NAPI drivers) |
885 | */ | |
886 | static void tcp_tasklet_func(unsigned long data) | |
887 | { | |
888 | struct tsq_tasklet *tsq = (struct tsq_tasklet *)data; | |
889 | LIST_HEAD(list); | |
890 | unsigned long flags; | |
891 | struct list_head *q, *n; | |
892 | struct tcp_sock *tp; | |
893 | struct sock *sk; | |
894 | ||
895 | local_irq_save(flags); | |
896 | list_splice_init(&tsq->head, &list); | |
897 | local_irq_restore(flags); | |
898 | ||
899 | list_for_each_safe(q, n, &list) { | |
900 | tp = list_entry(q, struct tcp_sock, tsq_node); | |
901 | list_del(&tp->tsq_node); | |
902 | ||
903 | sk = (struct sock *)tp; | |
0a9648f1 | 904 | smp_mb__before_atomic(); |
7aa5470c ED |
905 | clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags); |
906 | ||
73a6bab5 | 907 | tcp_tsq_handler(sk); |
46d3ceab ED |
908 | sk_free(sk); |
909 | } | |
910 | } | |
911 | ||
40fc3423 ED |
912 | #define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \ |
913 | TCPF_WRITE_TIMER_DEFERRED | \ | |
914 | TCPF_DELACK_TIMER_DEFERRED | \ | |
915 | TCPF_MTU_REDUCED_DEFERRED) | |
46d3ceab ED |
916 | /** |
917 | * tcp_release_cb - tcp release_sock() callback | |
918 | * @sk: socket | |
919 | * | |
920 | * called from release_sock() to perform protocol dependent | |
921 | * actions before socket release. | |
922 | */ | |
923 | void tcp_release_cb(struct sock *sk) | |
924 | { | |
6f458dfb | 925 | unsigned long flags, nflags; |
46d3ceab | 926 | |
6f458dfb ED |
927 | /* perform an atomic operation only if at least one flag is set */ |
928 | do { | |
7aa5470c | 929 | flags = sk->sk_tsq_flags; |
6f458dfb ED |
930 | if (!(flags & TCP_DEFERRED_ALL)) |
931 | return; | |
932 | nflags = flags & ~TCP_DEFERRED_ALL; | |
7aa5470c | 933 | } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags); |
6f458dfb | 934 | |
73a6bab5 ED |
935 | if (flags & TCPF_TSQ_DEFERRED) { |
936 | tcp_tsq_write(sk); | |
937 | __sock_put(sk); | |
938 | } | |
c3f9b018 ED |
939 | /* Here begins the tricky part : |
940 | * We are called from release_sock() with : | |
941 | * 1) BH disabled | |
942 | * 2) sk_lock.slock spinlock held | |
943 | * 3) socket owned by us (sk->sk_lock.owned == 1) | |
944 | * | |
945 | * But following code is meant to be called from BH handlers, | |
946 | * so we should keep BH disabled, but early release socket ownership | |
947 | */ | |
948 | sock_release_ownership(sk); | |
949 | ||
40fc3423 | 950 | if (flags & TCPF_WRITE_TIMER_DEFERRED) { |
6f458dfb | 951 | tcp_write_timer_handler(sk); |
144d56e9 ED |
952 | __sock_put(sk); |
953 | } | |
40fc3423 | 954 | if (flags & TCPF_DELACK_TIMER_DEFERRED) { |
6f458dfb | 955 | tcp_delack_timer_handler(sk); |
144d56e9 ED |
956 | __sock_put(sk); |
957 | } | |
40fc3423 | 958 | if (flags & TCPF_MTU_REDUCED_DEFERRED) { |
4fab9071 | 959 | inet_csk(sk)->icsk_af_ops->mtu_reduced(sk); |
144d56e9 ED |
960 | __sock_put(sk); |
961 | } | |
46d3ceab ED |
962 | } |
963 | EXPORT_SYMBOL(tcp_release_cb); | |
964 | ||
965 | void __init tcp_tasklet_init(void) | |
966 | { | |
967 | int i; | |
968 | ||
969 | for_each_possible_cpu(i) { | |
970 | struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i); | |
971 | ||
972 | INIT_LIST_HEAD(&tsq->head); | |
973 | tasklet_init(&tsq->tasklet, | |
974 | tcp_tasklet_func, | |
975 | (unsigned long)tsq); | |
976 | } | |
977 | } | |
978 | ||
979 | /* | |
980 | * Write buffer destructor automatically called from kfree_skb. | |
8e3bff96 | 981 | * We can't xmit new skbs from this context, as we might already |
46d3ceab ED |
982 | * hold qdisc lock. |
983 | */ | |
d6a4a104 | 984 | void tcp_wfree(struct sk_buff *skb) |
46d3ceab ED |
985 | { |
986 | struct sock *sk = skb->sk; | |
987 | struct tcp_sock *tp = tcp_sk(sk); | |
408f0a6c | 988 | unsigned long flags, nval, oval; |
9b462d02 ED |
989 | |
990 | /* Keep one reference on sk_wmem_alloc. | |
991 | * Will be released by sk_free() from here or tcp_tasklet_func() | |
992 | */ | |
14afee4b | 993 | WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc)); |
9b462d02 ED |
994 | |
995 | /* If this softirq is serviced by ksoftirqd, we are likely under stress. | |
996 | * Wait until our queues (qdisc + devices) are drained. | |
997 | * This gives : | |
998 | * - less callbacks to tcp_write_xmit(), reducing stress (batches) | |
999 | * - chance for incoming ACK (processed by another cpu maybe) | |
1000 | * to migrate this flow (skb->ooo_okay will be eventually set) | |
1001 | */ | |
14afee4b | 1002 | if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current) |
9b462d02 | 1003 | goto out; |
46d3ceab | 1004 | |
7aa5470c | 1005 | for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) { |
46d3ceab | 1006 | struct tsq_tasklet *tsq; |
a9b204d1 | 1007 | bool empty; |
46d3ceab | 1008 | |
408f0a6c ED |
1009 | if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED)) |
1010 | goto out; | |
1011 | ||
73a6bab5 | 1012 | nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED; |
7aa5470c | 1013 | nval = cmpxchg(&sk->sk_tsq_flags, oval, nval); |
408f0a6c ED |
1014 | if (nval != oval) |
1015 | continue; | |
1016 | ||
46d3ceab ED |
1017 | /* queue this socket to tasklet queue */ |
1018 | local_irq_save(flags); | |
903ceff7 | 1019 | tsq = this_cpu_ptr(&tsq_tasklet); |
a9b204d1 | 1020 | empty = list_empty(&tsq->head); |
46d3ceab | 1021 | list_add(&tp->tsq_node, &tsq->head); |
a9b204d1 ED |
1022 | if (empty) |
1023 | tasklet_schedule(&tsq->tasklet); | |
46d3ceab | 1024 | local_irq_restore(flags); |
9b462d02 | 1025 | return; |
46d3ceab | 1026 | } |
9b462d02 ED |
1027 | out: |
1028 | sk_free(sk); | |
46d3ceab ED |
1029 | } |
1030 | ||
73a6bab5 ED |
1031 | /* Note: Called under soft irq. |
1032 | * We can call TCP stack right away, unless socket is owned by user. | |
218af599 ED |
1033 | */ |
1034 | enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer) | |
1035 | { | |
1036 | struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer); | |
1037 | struct sock *sk = (struct sock *)tp; | |
218af599 | 1038 | |
73a6bab5 ED |
1039 | tcp_tsq_handler(sk); |
1040 | sock_put(sk); | |
218af599 | 1041 | |
218af599 ED |
1042 | return HRTIMER_NORESTART; |
1043 | } | |
1044 | ||
a7a25630 ED |
1045 | static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb, |
1046 | u64 prior_wstamp) | |
e2080072 | 1047 | { |
ab408b6d ED |
1048 | struct tcp_sock *tp = tcp_sk(sk); |
1049 | ||
ab408b6d | 1050 | if (sk->sk_pacing_status != SK_PACING_NONE) { |
76a9ebe8 | 1051 | unsigned long rate = sk->sk_pacing_rate; |
ab408b6d ED |
1052 | |
1053 | /* Original sch_fq does not pace first 10 MSS | |
1054 | * Note that tp->data_segs_out overflows after 2^32 packets, | |
1055 | * this is a minor annoyance. | |
1056 | */ | |
76a9ebe8 | 1057 | if (rate != ~0UL && rate && tp->data_segs_out >= 10) { |
a7a25630 ED |
1058 | u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate); |
1059 | u64 credit = tp->tcp_wstamp_ns - prior_wstamp; | |
1060 | ||
1061 | /* take into account OS jitter */ | |
1062 | len_ns -= min_t(u64, len_ns / 2, credit); | |
1063 | tp->tcp_wstamp_ns += len_ns; | |
ab408b6d ED |
1064 | } |
1065 | } | |
e2080072 ED |
1066 | list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); |
1067 | } | |
1068 | ||
05e22e83 ED |
1069 | INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); |
1070 | INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl)); | |
dd2e0b86 | 1071 | INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb)); |
05e22e83 | 1072 | |
1da177e4 LT |
1073 | /* This routine actually transmits TCP packets queued in by |
1074 | * tcp_do_sendmsg(). This is used by both the initial | |
1075 | * transmission and possible later retransmissions. | |
1076 | * All SKB's seen here are completely headerless. It is our | |
1077 | * job to build the TCP header, and pass the packet down to | |
1078 | * IP so it can do the same plus pass the packet off to the | |
1079 | * device. | |
1080 | * | |
1081 | * We are working here with either a clone of the original | |
1082 | * SKB, or a fresh unique copy made by the retransmit engine. | |
1083 | */ | |
2987babb YC |
1084 | static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, |
1085 | int clone_it, gfp_t gfp_mask, u32 rcv_nxt) | |
1da177e4 | 1086 | { |
dfb4b9dc DM |
1087 | const struct inet_connection_sock *icsk = inet_csk(sk); |
1088 | struct inet_sock *inet; | |
1089 | struct tcp_sock *tp; | |
1090 | struct tcp_skb_cb *tcb; | |
33ad798c | 1091 | struct tcp_out_options opts; |
95c96174 | 1092 | unsigned int tcp_options_size, tcp_header_size; |
8c72c65b | 1093 | struct sk_buff *oskb = NULL; |
cfb6eeb4 | 1094 | struct tcp_md5sig_key *md5; |
dfb4b9dc | 1095 | struct tcphdr *th; |
a7a25630 | 1096 | u64 prior_wstamp; |
dfb4b9dc DM |
1097 | int err; |
1098 | ||
1099 | BUG_ON(!skb || !tcp_skb_pcount(skb)); | |
6f094b9e | 1100 | tp = tcp_sk(sk); |
7f12422c YC |
1101 | prior_wstamp = tp->tcp_wstamp_ns; |
1102 | tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache); | |
1103 | skb->skb_mstamp_ns = tp->tcp_wstamp_ns; | |
ccdbb6e9 | 1104 | if (clone_it) { |
6f094b9e LB |
1105 | TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq |
1106 | - tp->snd_una; | |
8c72c65b | 1107 | oskb = skb; |
e2080072 ED |
1108 | |
1109 | tcp_skb_tsorted_save(oskb) { | |
1110 | if (unlikely(skb_cloned(oskb))) | |
1111 | skb = pskb_copy(oskb, gfp_mask); | |
1112 | else | |
1113 | skb = skb_clone(oskb, gfp_mask); | |
1114 | } tcp_skb_tsorted_restore(oskb); | |
1115 | ||
dfb4b9dc DM |
1116 | if (unlikely(!skb)) |
1117 | return -ENOBUFS; | |
b738a185 ED |
1118 | /* retransmit skbs might have a non zero value in skb->dev |
1119 | * because skb->dev is aliased with skb->rbnode.rb_left | |
1120 | */ | |
1121 | skb->dev = NULL; | |
dfb4b9dc | 1122 | } |
5f6188a8 | 1123 | |
dfb4b9dc | 1124 | inet = inet_sk(sk); |
dfb4b9dc | 1125 | tcb = TCP_SKB_CB(skb); |
33ad798c | 1126 | memset(&opts, 0, sizeof(opts)); |
1da177e4 | 1127 | |
051ba674 | 1128 | if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) { |
33ad798c | 1129 | tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5); |
051ba674 | 1130 | } else { |
33ad798c AL |
1131 | tcp_options_size = tcp_established_options(sk, skb, &opts, |
1132 | &md5); | |
051ba674 ED |
1133 | /* Force a PSH flag on all (GSO) packets to expedite GRO flush |
1134 | * at receiver : This slightly improve GRO performance. | |
1135 | * Note that we do not force the PSH flag for non GSO packets, | |
1136 | * because they might be sent under high congestion events, | |
1137 | * and in this case it is better to delay the delivery of 1-MSS | |
1138 | * packets and thus the corresponding ACK packet that would | |
1139 | * release the following packet. | |
1140 | */ | |
1141 | if (tcp_skb_pcount(skb) > 1) | |
1142 | tcb->tcp_flags |= TCPHDR_PSH; | |
1143 | } | |
33ad798c | 1144 | tcp_header_size = tcp_options_size + sizeof(struct tcphdr); |
e905a9ed | 1145 | |
547669d4 | 1146 | /* if no packet is in qdisc/device queue, then allow XPS to select |
b2532eb9 | 1147 | * another queue. We can be called from tcp_tsq_handler() |
73a6bab5 | 1148 | * which holds one reference to sk. |
b2532eb9 ED |
1149 | * |
1150 | * TODO: Ideally, in-flight pure ACK packets should not matter here. | |
1151 | * One way to get this would be to set skb->truesize = 2 on them. | |
547669d4 | 1152 | */ |
b2532eb9 | 1153 | skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1); |
dfb4b9dc | 1154 | |
38ab52e8 ED |
1155 | /* If we had to use memory reserve to allocate this skb, |
1156 | * this might cause drops if packet is looped back : | |
1157 | * Other socket might not have SOCK_MEMALLOC. | |
1158 | * Packets not looped back do not care about pfmemalloc. | |
1159 | */ | |
1160 | skb->pfmemalloc = 0; | |
1161 | ||
aa8223c7 ACM |
1162 | skb_push(skb, tcp_header_size); |
1163 | skb_reset_transport_header(skb); | |
46d3ceab ED |
1164 | |
1165 | skb_orphan(skb); | |
1166 | skb->sk = sk; | |
1d2077ac | 1167 | skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree; |
b73c3d0e | 1168 | skb_set_hash_from_sk(skb, sk); |
14afee4b | 1169 | refcount_add(skb->truesize, &sk->sk_wmem_alloc); |
dfb4b9dc | 1170 | |
c3a2e837 JA |
1171 | skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm); |
1172 | ||
dfb4b9dc | 1173 | /* Build TCP header and checksum it. */ |
ea1627c2 | 1174 | th = (struct tcphdr *)skb->data; |
c720c7e8 ED |
1175 | th->source = inet->inet_sport; |
1176 | th->dest = inet->inet_dport; | |
dfb4b9dc | 1177 | th->seq = htonl(tcb->seq); |
2987babb | 1178 | th->ack_seq = htonl(rcv_nxt); |
df7a3b07 | 1179 | *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) | |
4de075e0 | 1180 | tcb->tcp_flags); |
dfb4b9dc | 1181 | |
dfb4b9dc DM |
1182 | th->check = 0; |
1183 | th->urg_ptr = 0; | |
1da177e4 | 1184 | |
33f5f57e | 1185 | /* The urg_mode check is necessary during a below snd_una win probe */ |
7691367d HX |
1186 | if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) { |
1187 | if (before(tp->snd_up, tcb->seq + 0x10000)) { | |
1188 | th->urg_ptr = htons(tp->snd_up - tcb->seq); | |
1189 | th->urg = 1; | |
1190 | } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) { | |
0eae88f3 | 1191 | th->urg_ptr = htons(0xFFFF); |
7691367d HX |
1192 | th->urg = 1; |
1193 | } | |
dfb4b9dc | 1194 | } |
1da177e4 | 1195 | |
bd0388ae | 1196 | tcp_options_write((__be32 *)(th + 1), tp, &opts); |
51466a75 | 1197 | skb_shinfo(skb)->gso_type = sk->sk_gso_type; |
ea1627c2 ED |
1198 | if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) { |
1199 | th->window = htons(tcp_select_window(sk)); | |
1200 | tcp_ecn_send(sk, skb, th, tcp_header_size); | |
1201 | } else { | |
1202 | /* RFC1323: The window in SYN & SYN/ACK segments | |
1203 | * is never scaled. | |
1204 | */ | |
1205 | th->window = htons(min(tp->rcv_wnd, 65535U)); | |
1206 | } | |
cfb6eeb4 YH |
1207 | #ifdef CONFIG_TCP_MD5SIG |
1208 | /* Calculate the MD5 hash, as we have all we need now */ | |
1209 | if (md5) { | |
a465419b | 1210 | sk_nocaps_add(sk, NETIF_F_GSO_MASK); |
bd0388ae | 1211 | tp->af_specific->calc_md5_hash(opts.hash_location, |
39f8e58e | 1212 | md5, sk, skb); |
cfb6eeb4 YH |
1213 | } |
1214 | #endif | |
1215 | ||
dd2e0b86 ED |
1216 | INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check, |
1217 | tcp_v6_send_check, tcp_v4_send_check, | |
1218 | sk, skb); | |
1da177e4 | 1219 | |
4de075e0 | 1220 | if (likely(tcb->tcp_flags & TCPHDR_ACK)) |
27cde44a | 1221 | tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt); |
1da177e4 | 1222 | |
a44d6eac | 1223 | if (skb->len != tcp_header_size) { |
cf533ea5 | 1224 | tcp_event_data_sent(tp, sk); |
a44d6eac | 1225 | tp->data_segs_out += tcp_skb_pcount(skb); |
ba113c3a | 1226 | tp->bytes_sent += skb->len - tcp_header_size; |
a44d6eac | 1227 | } |
1da177e4 | 1228 | |
bd37a088 | 1229 | if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq) |
aa2ea058 TH |
1230 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, |
1231 | tcp_skb_pcount(skb)); | |
1da177e4 | 1232 | |
2efd055c | 1233 | tp->segs_out += tcp_skb_pcount(skb); |
f69ad292 | 1234 | /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */ |
cd7d8498 | 1235 | skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb); |
f69ad292 | 1236 | skb_shinfo(skb)->gso_size = tcp_skb_mss(skb); |
cd7d8498 | 1237 | |
d3edd06e | 1238 | /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */ |
971f10ec ED |
1239 | |
1240 | /* Cleanup our debris for IP stacks */ | |
1241 | memset(skb->cb, 0, max(sizeof(struct inet_skb_parm), | |
1242 | sizeof(struct inet6_skb_parm))); | |
1243 | ||
a842fe14 ED |
1244 | tcp_add_tx_delay(skb, tp); |
1245 | ||
05e22e83 ED |
1246 | err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit, |
1247 | inet6_csk_xmit, ip_queue_xmit, | |
1248 | sk, skb, &inet->cork.fl); | |
7faee5c0 | 1249 | |
8c72c65b ED |
1250 | if (unlikely(err > 0)) { |
1251 | tcp_enter_cwr(sk); | |
1252 | err = net_xmit_eval(err); | |
1253 | } | |
fc225799 | 1254 | if (!err && oskb) { |
a7a25630 | 1255 | tcp_update_skb_after_send(sk, oskb, prior_wstamp); |
fc225799 ED |
1256 | tcp_rate_skb_sent(sk, oskb); |
1257 | } | |
8c72c65b | 1258 | return err; |
1da177e4 LT |
1259 | } |
1260 | ||
2987babb YC |
1261 | static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it, |
1262 | gfp_t gfp_mask) | |
1263 | { | |
1264 | return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask, | |
1265 | tcp_sk(sk)->rcv_nxt); | |
1266 | } | |
1267 | ||
67edfef7 | 1268 | /* This routine just queues the buffer for sending. |
1da177e4 LT |
1269 | * |
1270 | * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames, | |
1271 | * otherwise socket can stall. | |
1272 | */ | |
1273 | static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb) | |
1274 | { | |
1275 | struct tcp_sock *tp = tcp_sk(sk); | |
1276 | ||
1277 | /* Advance write_seq and place onto the write_queue. */ | |
0f317464 | 1278 | WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq); |
f4a775d1 | 1279 | __skb_header_release(skb); |
fe067e8a | 1280 | tcp_add_write_queue_tail(sk, skb); |
ab4e846a | 1281 | sk_wmem_queued_add(sk, skb->truesize); |
3ab224be | 1282 | sk_mem_charge(sk, skb->truesize); |
1da177e4 LT |
1283 | } |
1284 | ||
67edfef7 | 1285 | /* Initialize TSO segments for a packet. */ |
5bbb432c | 1286 | static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now) |
f6302d1d | 1287 | { |
4a64fd6c | 1288 | if (skb->len <= mss_now) { |
f6302d1d DM |
1289 | /* Avoid the costly divide in the normal |
1290 | * non-TSO case. | |
1291 | */ | |
cd7d8498 | 1292 | tcp_skb_pcount_set(skb, 1); |
f69ad292 | 1293 | TCP_SKB_CB(skb)->tcp_gso_size = 0; |
f6302d1d | 1294 | } else { |
cd7d8498 | 1295 | tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now)); |
f69ad292 | 1296 | TCP_SKB_CB(skb)->tcp_gso_size = mss_now; |
1da177e4 LT |
1297 | } |
1298 | } | |
1299 | ||
797108d1 IJ |
1300 | /* Pcount in the middle of the write queue got changed, we need to do various |
1301 | * tweaks to fix counters | |
1302 | */ | |
cf533ea5 | 1303 | static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr) |
797108d1 IJ |
1304 | { |
1305 | struct tcp_sock *tp = tcp_sk(sk); | |
1306 | ||
1307 | tp->packets_out -= decr; | |
1308 | ||
1309 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) | |
1310 | tp->sacked_out -= decr; | |
1311 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) | |
1312 | tp->retrans_out -= decr; | |
1313 | if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST) | |
1314 | tp->lost_out -= decr; | |
1315 | ||
1316 | /* Reno case is special. Sigh... */ | |
1317 | if (tcp_is_reno(tp) && decr > 0) | |
1318 | tp->sacked_out -= min_t(u32, tp->sacked_out, decr); | |
1319 | ||
797108d1 IJ |
1320 | if (tp->lost_skb_hint && |
1321 | before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) && | |
713bafea | 1322 | (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)) |
797108d1 IJ |
1323 | tp->lost_cnt_hint -= decr; |
1324 | ||
1325 | tcp_verify_left_out(tp); | |
1326 | } | |
1327 | ||
0a2cf20c SHY |
1328 | static bool tcp_has_tx_tstamp(const struct sk_buff *skb) |
1329 | { | |
1330 | return TCP_SKB_CB(skb)->txstamp_ack || | |
1331 | (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP); | |
1332 | } | |
1333 | ||
490cc7d0 WB |
1334 | static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2) |
1335 | { | |
1336 | struct skb_shared_info *shinfo = skb_shinfo(skb); | |
1337 | ||
0a2cf20c | 1338 | if (unlikely(tcp_has_tx_tstamp(skb)) && |
490cc7d0 WB |
1339 | !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) { |
1340 | struct skb_shared_info *shinfo2 = skb_shinfo(skb2); | |
1341 | u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP; | |
1342 | ||
1343 | shinfo->tx_flags &= ~tsflags; | |
1344 | shinfo2->tx_flags |= tsflags; | |
1345 | swap(shinfo->tskey, shinfo2->tskey); | |
b51e13fa MKL |
1346 | TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack; |
1347 | TCP_SKB_CB(skb)->txstamp_ack = 0; | |
490cc7d0 WB |
1348 | } |
1349 | } | |
1350 | ||
a166140e MKL |
1351 | static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2) |
1352 | { | |
1353 | TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor; | |
1354 | TCP_SKB_CB(skb)->eor = 0; | |
1355 | } | |
1356 | ||
75c119af ED |
1357 | /* Insert buff after skb on the write or rtx queue of sk. */ |
1358 | static void tcp_insert_write_queue_after(struct sk_buff *skb, | |
1359 | struct sk_buff *buff, | |
1360 | struct sock *sk, | |
1361 | enum tcp_queue tcp_queue) | |
1362 | { | |
1363 | if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE) | |
1364 | __skb_queue_after(&sk->sk_write_queue, skb, buff); | |
1365 | else | |
1366 | tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); | |
1367 | } | |
1368 | ||
1da177e4 LT |
1369 | /* Function to create two new TCP segments. Shrinks the given segment |
1370 | * to the specified size and appends a new segment with the rest of the | |
e905a9ed | 1371 | * packet to the list. This won't be called frequently, I hope. |
1da177e4 LT |
1372 | * Remember, these are still headerless SKBs at this point. |
1373 | */ | |
75c119af ED |
1374 | int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue, |
1375 | struct sk_buff *skb, u32 len, | |
6cc55e09 | 1376 | unsigned int mss_now, gfp_t gfp) |
1da177e4 LT |
1377 | { |
1378 | struct tcp_sock *tp = tcp_sk(sk); | |
1379 | struct sk_buff *buff; | |
6475be16 | 1380 | int nsize, old_factor; |
b617158d | 1381 | long limit; |
b60b49ea | 1382 | int nlen; |
9ce01461 | 1383 | u8 flags; |
1da177e4 | 1384 | |
2fceec13 IJ |
1385 | if (WARN_ON(len > skb->len)) |
1386 | return -EINVAL; | |
6a438bbe | 1387 | |
1da177e4 LT |
1388 | nsize = skb_headlen(skb) - len; |
1389 | if (nsize < 0) | |
1390 | nsize = 0; | |
1391 | ||
b617158d ED |
1392 | /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb. |
1393 | * We need some allowance to not penalize applications setting small | |
1394 | * SO_SNDBUF values. | |
1395 | * Also allow first and last skb in retransmit queue to be split. | |
1396 | */ | |
1397 | limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE); | |
1398 | if (unlikely((sk->sk_wmem_queued >> 1) > limit && | |
1399 | tcp_queue != TCP_FRAG_IN_WRITE_QUEUE && | |
1400 | skb != tcp_rtx_queue_head(sk) && | |
1401 | skb != tcp_rtx_queue_tail(sk))) { | |
f070ef2a ED |
1402 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG); |
1403 | return -ENOMEM; | |
1404 | } | |
1405 | ||
6cc55e09 | 1406 | if (skb_unclone(skb, gfp)) |
1da177e4 LT |
1407 | return -ENOMEM; |
1408 | ||
1409 | /* Get a new skb... force flag on. */ | |
eb934478 | 1410 | buff = sk_stream_alloc_skb(sk, nsize, gfp, true); |
51456b29 | 1411 | if (!buff) |
1da177e4 | 1412 | return -ENOMEM; /* We'll just try again later. */ |
41477662 | 1413 | skb_copy_decrypted(buff, skb); |
ef5cb973 | 1414 | |
ab4e846a | 1415 | sk_wmem_queued_add(sk, buff->truesize); |
3ab224be | 1416 | sk_mem_charge(sk, buff->truesize); |
b60b49ea HX |
1417 | nlen = skb->len - len - nsize; |
1418 | buff->truesize += nlen; | |
1419 | skb->truesize -= nlen; | |
1da177e4 LT |
1420 | |
1421 | /* Correct the sequence numbers. */ | |
1422 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1423 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1424 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1425 | ||
1426 | /* PSH and FIN should only be set in the second packet. */ | |
4de075e0 ED |
1427 | flags = TCP_SKB_CB(skb)->tcp_flags; |
1428 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
1429 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
e14c3caf | 1430 | TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked; |
a166140e | 1431 | tcp_skb_fragment_eor(skb, buff); |
1da177e4 | 1432 | |
98be9b12 | 1433 | skb_split(skb, buff, len); |
1da177e4 | 1434 | |
98be9b12 | 1435 | buff->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 | 1436 | |
a61bbcf2 | 1437 | buff->tstamp = skb->tstamp; |
490cc7d0 | 1438 | tcp_fragment_tstamp(skb, buff); |
1da177e4 | 1439 | |
6475be16 DM |
1440 | old_factor = tcp_skb_pcount(skb); |
1441 | ||
1da177e4 | 1442 | /* Fix up tso_factor for both original and new SKB. */ |
5bbb432c ED |
1443 | tcp_set_skb_tso_segs(skb, mss_now); |
1444 | tcp_set_skb_tso_segs(buff, mss_now); | |
1da177e4 | 1445 | |
b9f64820 YC |
1446 | /* Update delivered info for the new segment */ |
1447 | TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx; | |
1448 | ||
6475be16 DM |
1449 | /* If this packet has been sent out already, we must |
1450 | * adjust the various packet counters. | |
1451 | */ | |
cf0b450c | 1452 | if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) { |
6475be16 DM |
1453 | int diff = old_factor - tcp_skb_pcount(skb) - |
1454 | tcp_skb_pcount(buff); | |
1da177e4 | 1455 | |
797108d1 IJ |
1456 | if (diff) |
1457 | tcp_adjust_pcount(sk, skb, diff); | |
1da177e4 LT |
1458 | } |
1459 | ||
1460 | /* Link BUFF into the send queue. */ | |
f4a775d1 | 1461 | __skb_header_release(buff); |
75c119af | 1462 | tcp_insert_write_queue_after(skb, buff, sk, tcp_queue); |
f67971e6 ED |
1463 | if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE) |
1464 | list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor); | |
1da177e4 LT |
1465 | |
1466 | return 0; | |
1467 | } | |
1468 | ||
f4d01666 ED |
1469 | /* This is similar to __pskb_pull_tail(). The difference is that pulled |
1470 | * data is not copied, but immediately discarded. | |
1da177e4 | 1471 | */ |
7162fb24 | 1472 | static int __pskb_trim_head(struct sk_buff *skb, int len) |
1da177e4 | 1473 | { |
7b7fc97a | 1474 | struct skb_shared_info *shinfo; |
1da177e4 LT |
1475 | int i, k, eat; |
1476 | ||
4fa48bf3 ED |
1477 | eat = min_t(int, len, skb_headlen(skb)); |
1478 | if (eat) { | |
1479 | __skb_pull(skb, eat); | |
1480 | len -= eat; | |
1481 | if (!len) | |
7162fb24 | 1482 | return 0; |
4fa48bf3 | 1483 | } |
1da177e4 LT |
1484 | eat = len; |
1485 | k = 0; | |
7b7fc97a ED |
1486 | shinfo = skb_shinfo(skb); |
1487 | for (i = 0; i < shinfo->nr_frags; i++) { | |
1488 | int size = skb_frag_size(&shinfo->frags[i]); | |
9e903e08 ED |
1489 | |
1490 | if (size <= eat) { | |
aff65da0 | 1491 | skb_frag_unref(skb, i); |
9e903e08 | 1492 | eat -= size; |
1da177e4 | 1493 | } else { |
7b7fc97a | 1494 | shinfo->frags[k] = shinfo->frags[i]; |
1da177e4 | 1495 | if (eat) { |
b54c9d5b | 1496 | skb_frag_off_add(&shinfo->frags[k], eat); |
7b7fc97a | 1497 | skb_frag_size_sub(&shinfo->frags[k], eat); |
1da177e4 LT |
1498 | eat = 0; |
1499 | } | |
1500 | k++; | |
1501 | } | |
1502 | } | |
7b7fc97a | 1503 | shinfo->nr_frags = k; |
1da177e4 | 1504 | |
1da177e4 LT |
1505 | skb->data_len -= len; |
1506 | skb->len = skb->data_len; | |
7162fb24 | 1507 | return len; |
1da177e4 LT |
1508 | } |
1509 | ||
67edfef7 | 1510 | /* Remove acked data from a packet in the transmit queue. */ |
1da177e4 LT |
1511 | int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len) |
1512 | { | |
7162fb24 ED |
1513 | u32 delta_truesize; |
1514 | ||
14bbd6a5 | 1515 | if (skb_unclone(skb, GFP_ATOMIC)) |
1da177e4 LT |
1516 | return -ENOMEM; |
1517 | ||
7162fb24 | 1518 | delta_truesize = __pskb_trim_head(skb, len); |
1da177e4 LT |
1519 | |
1520 | TCP_SKB_CB(skb)->seq += len; | |
84fa7933 | 1521 | skb->ip_summed = CHECKSUM_PARTIAL; |
1da177e4 | 1522 | |
7162fb24 ED |
1523 | if (delta_truesize) { |
1524 | skb->truesize -= delta_truesize; | |
ab4e846a | 1525 | sk_wmem_queued_add(sk, -delta_truesize); |
7162fb24 ED |
1526 | sk_mem_uncharge(sk, delta_truesize); |
1527 | sock_set_flag(sk, SOCK_QUEUE_SHRUNK); | |
1528 | } | |
1da177e4 | 1529 | |
5b35e1e6 | 1530 | /* Any change of skb->len requires recalculation of tso factor. */ |
1da177e4 | 1531 | if (tcp_skb_pcount(skb) > 1) |
5bbb432c | 1532 | tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb)); |
1da177e4 LT |
1533 | |
1534 | return 0; | |
1535 | } | |
1536 | ||
1b63edd6 YC |
1537 | /* Calculate MSS not accounting any TCP options. */ |
1538 | static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
5d424d5a | 1539 | { |
cf533ea5 ED |
1540 | const struct tcp_sock *tp = tcp_sk(sk); |
1541 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
5d424d5a JH |
1542 | int mss_now; |
1543 | ||
1544 | /* Calculate base mss without TCP options: | |
1545 | It is MMS_S - sizeof(tcphdr) of rfc1122 | |
1546 | */ | |
1547 | mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr); | |
1548 | ||
67469601 ED |
1549 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
1550 | if (icsk->icsk_af_ops->net_frag_header_len) { | |
1551 | const struct dst_entry *dst = __sk_dst_get(sk); | |
1552 | ||
1553 | if (dst && dst_allfrag(dst)) | |
1554 | mss_now -= icsk->icsk_af_ops->net_frag_header_len; | |
1555 | } | |
1556 | ||
5d424d5a JH |
1557 | /* Clamp it (mss_clamp does not include tcp options) */ |
1558 | if (mss_now > tp->rx_opt.mss_clamp) | |
1559 | mss_now = tp->rx_opt.mss_clamp; | |
1560 | ||
1561 | /* Now subtract optional transport overhead */ | |
1562 | mss_now -= icsk->icsk_ext_hdr_len; | |
1563 | ||
1564 | /* Then reserve room for full set of TCP options and 8 bytes of data */ | |
5f3e2bf0 | 1565 | mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss); |
5d424d5a JH |
1566 | return mss_now; |
1567 | } | |
1568 | ||
1b63edd6 YC |
1569 | /* Calculate MSS. Not accounting for SACKs here. */ |
1570 | int tcp_mtu_to_mss(struct sock *sk, int pmtu) | |
1571 | { | |
1572 | /* Subtract TCP options size, not including SACKs */ | |
1573 | return __tcp_mtu_to_mss(sk, pmtu) - | |
1574 | (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr)); | |
1575 | } | |
1576 | ||
5d424d5a | 1577 | /* Inverse of above */ |
67469601 | 1578 | int tcp_mss_to_mtu(struct sock *sk, int mss) |
5d424d5a | 1579 | { |
cf533ea5 ED |
1580 | const struct tcp_sock *tp = tcp_sk(sk); |
1581 | const struct inet_connection_sock *icsk = inet_csk(sk); | |
5d424d5a JH |
1582 | int mtu; |
1583 | ||
1584 | mtu = mss + | |
1585 | tp->tcp_header_len + | |
1586 | icsk->icsk_ext_hdr_len + | |
1587 | icsk->icsk_af_ops->net_header_len; | |
1588 | ||
67469601 ED |
1589 | /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */ |
1590 | if (icsk->icsk_af_ops->net_frag_header_len) { | |
1591 | const struct dst_entry *dst = __sk_dst_get(sk); | |
1592 | ||
1593 | if (dst && dst_allfrag(dst)) | |
1594 | mtu += icsk->icsk_af_ops->net_frag_header_len; | |
1595 | } | |
5d424d5a JH |
1596 | return mtu; |
1597 | } | |
556c6b46 | 1598 | EXPORT_SYMBOL(tcp_mss_to_mtu); |
5d424d5a | 1599 | |
67edfef7 | 1600 | /* MTU probing init per socket */ |
5d424d5a JH |
1601 | void tcp_mtup_init(struct sock *sk) |
1602 | { | |
1603 | struct tcp_sock *tp = tcp_sk(sk); | |
1604 | struct inet_connection_sock *icsk = inet_csk(sk); | |
b0f9ca53 | 1605 | struct net *net = sock_net(sk); |
5d424d5a | 1606 | |
b0f9ca53 | 1607 | icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1; |
5d424d5a | 1608 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) + |
e905a9ed | 1609 | icsk->icsk_af_ops->net_header_len; |
b0f9ca53 | 1610 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss); |
5d424d5a | 1611 | icsk->icsk_mtup.probe_size = 0; |
05cbc0db | 1612 | if (icsk->icsk_mtup.enabled) |
c74df29a | 1613 | icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; |
5d424d5a | 1614 | } |
4bc2f18b | 1615 | EXPORT_SYMBOL(tcp_mtup_init); |
5d424d5a | 1616 | |
1da177e4 LT |
1617 | /* This function synchronize snd mss to current pmtu/exthdr set. |
1618 | ||
1619 | tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts | |
1620 | for TCP options, but includes only bare TCP header. | |
1621 | ||
1622 | tp->rx_opt.mss_clamp is mss negotiated at connection setup. | |
caa20d9a | 1623 | It is minimum of user_mss and mss received with SYN. |
1da177e4 LT |
1624 | It also does not include TCP options. |
1625 | ||
d83d8461 | 1626 | inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function. |
1da177e4 LT |
1627 | |
1628 | tp->mss_cache is current effective sending mss, including | |
1629 | all tcp options except for SACKs. It is evaluated, | |
1630 | taking into account current pmtu, but never exceeds | |
1631 | tp->rx_opt.mss_clamp. | |
1632 | ||
1633 | NOTE1. rfc1122 clearly states that advertised MSS | |
1634 | DOES NOT include either tcp or ip options. | |
1635 | ||
d83d8461 ACM |
1636 | NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache |
1637 | are READ ONLY outside this function. --ANK (980731) | |
1da177e4 | 1638 | */ |
1da177e4 LT |
1639 | unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu) |
1640 | { | |
1641 | struct tcp_sock *tp = tcp_sk(sk); | |
d83d8461 | 1642 | struct inet_connection_sock *icsk = inet_csk(sk); |
5d424d5a | 1643 | int mss_now; |
1da177e4 | 1644 | |
5d424d5a JH |
1645 | if (icsk->icsk_mtup.search_high > pmtu) |
1646 | icsk->icsk_mtup.search_high = pmtu; | |
1da177e4 | 1647 | |
5d424d5a | 1648 | mss_now = tcp_mtu_to_mss(sk, pmtu); |
409d22b4 | 1649 | mss_now = tcp_bound_to_half_wnd(tp, mss_now); |
1da177e4 LT |
1650 | |
1651 | /* And store cached results */ | |
d83d8461 | 1652 | icsk->icsk_pmtu_cookie = pmtu; |
5d424d5a JH |
1653 | if (icsk->icsk_mtup.enabled) |
1654 | mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low)); | |
c1b4a7e6 | 1655 | tp->mss_cache = mss_now; |
1da177e4 LT |
1656 | |
1657 | return mss_now; | |
1658 | } | |
4bc2f18b | 1659 | EXPORT_SYMBOL(tcp_sync_mss); |
1da177e4 LT |
1660 | |
1661 | /* Compute the current effective MSS, taking SACKs and IP options, | |
1662 | * and even PMTU discovery events into account. | |
1da177e4 | 1663 | */ |
0c54b85f | 1664 | unsigned int tcp_current_mss(struct sock *sk) |
1da177e4 | 1665 | { |
cf533ea5 ED |
1666 | const struct tcp_sock *tp = tcp_sk(sk); |
1667 | const struct dst_entry *dst = __sk_dst_get(sk); | |
c1b4a7e6 | 1668 | u32 mss_now; |
95c96174 | 1669 | unsigned int header_len; |
33ad798c AL |
1670 | struct tcp_out_options opts; |
1671 | struct tcp_md5sig_key *md5; | |
c1b4a7e6 DM |
1672 | |
1673 | mss_now = tp->mss_cache; | |
1674 | ||
1da177e4 LT |
1675 | if (dst) { |
1676 | u32 mtu = dst_mtu(dst); | |
d83d8461 | 1677 | if (mtu != inet_csk(sk)->icsk_pmtu_cookie) |
1da177e4 LT |
1678 | mss_now = tcp_sync_mss(sk, mtu); |
1679 | } | |
1680 | ||
33ad798c AL |
1681 | header_len = tcp_established_options(sk, NULL, &opts, &md5) + |
1682 | sizeof(struct tcphdr); | |
1683 | /* The mss_cache is sized based on tp->tcp_header_len, which assumes | |
1684 | * some common options. If this is an odd packet (because we have SACK | |
1685 | * blocks etc) then our calculated header_len will be different, and | |
1686 | * we have to adjust mss_now correspondingly */ | |
1687 | if (header_len != tp->tcp_header_len) { | |
1688 | int delta = (int) header_len - tp->tcp_header_len; | |
1689 | mss_now -= delta; | |
1690 | } | |
cfb6eeb4 | 1691 | |
1da177e4 LT |
1692 | return mss_now; |
1693 | } | |
1694 | ||
86fd14ad WP |
1695 | /* RFC2861, slow part. Adjust cwnd, after it was not full during one rto. |
1696 | * As additional protections, we do not touch cwnd in retransmission phases, | |
1697 | * and if application hit its sndbuf limit recently. | |
1698 | */ | |
1699 | static void tcp_cwnd_application_limited(struct sock *sk) | |
1700 | { | |
1701 | struct tcp_sock *tp = tcp_sk(sk); | |
1702 | ||
1703 | if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open && | |
1704 | sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) { | |
1705 | /* Limited by application or receiver window. */ | |
1706 | u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk)); | |
1707 | u32 win_used = max(tp->snd_cwnd_used, init_win); | |
1708 | if (win_used < tp->snd_cwnd) { | |
1709 | tp->snd_ssthresh = tcp_current_ssthresh(sk); | |
1710 | tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1; | |
1711 | } | |
1712 | tp->snd_cwnd_used = 0; | |
1713 | } | |
c2203cf7 | 1714 | tp->snd_cwnd_stamp = tcp_jiffies32; |
86fd14ad WP |
1715 | } |
1716 | ||
ca8a2263 | 1717 | static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited) |
a762a980 | 1718 | { |
1b1fc3fd | 1719 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; |
9e412ba7 | 1720 | struct tcp_sock *tp = tcp_sk(sk); |
a762a980 | 1721 | |
ca8a2263 NC |
1722 | /* Track the maximum number of outstanding packets in each |
1723 | * window, and remember whether we were cwnd-limited then. | |
1724 | */ | |
1725 | if (!before(tp->snd_una, tp->max_packets_seq) || | |
1726 | tp->packets_out > tp->max_packets_out) { | |
1727 | tp->max_packets_out = tp->packets_out; | |
1728 | tp->max_packets_seq = tp->snd_nxt; | |
1729 | tp->is_cwnd_limited = is_cwnd_limited; | |
1730 | } | |
e114a710 | 1731 | |
24901551 | 1732 | if (tcp_is_cwnd_limited(sk)) { |
a762a980 DM |
1733 | /* Network is feed fully. */ |
1734 | tp->snd_cwnd_used = 0; | |
c2203cf7 | 1735 | tp->snd_cwnd_stamp = tcp_jiffies32; |
a762a980 DM |
1736 | } else { |
1737 | /* Network starves. */ | |
1738 | if (tp->packets_out > tp->snd_cwnd_used) | |
1739 | tp->snd_cwnd_used = tp->packets_out; | |
1740 | ||
b510f0d2 | 1741 | if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle && |
c2203cf7 | 1742 | (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto && |
1b1fc3fd | 1743 | !ca_ops->cong_control) |
a762a980 | 1744 | tcp_cwnd_application_limited(sk); |
b0f71bd3 FY |
1745 | |
1746 | /* The following conditions together indicate the starvation | |
1747 | * is caused by insufficient sender buffer: | |
1748 | * 1) just sent some data (see tcp_write_xmit) | |
1749 | * 2) not cwnd limited (this else condition) | |
75c119af | 1750 | * 3) no more data to send (tcp_write_queue_empty()) |
b0f71bd3 FY |
1751 | * 4) application is hitting buffer limit (SOCK_NOSPACE) |
1752 | */ | |
75c119af | 1753 | if (tcp_write_queue_empty(sk) && sk->sk_socket && |
b0f71bd3 FY |
1754 | test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) && |
1755 | (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)) | |
1756 | tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED); | |
a762a980 DM |
1757 | } |
1758 | } | |
1759 | ||
d4589926 ED |
1760 | /* Minshall's variant of the Nagle send check. */ |
1761 | static bool tcp_minshall_check(const struct tcp_sock *tp) | |
1762 | { | |
1763 | return after(tp->snd_sml, tp->snd_una) && | |
1764 | !after(tp->snd_sml, tp->snd_nxt); | |
1765 | } | |
1766 | ||
1767 | /* Update snd_sml if this skb is under mss | |
1768 | * Note that a TSO packet might end with a sub-mss segment | |
1769 | * The test is really : | |
1770 | * if ((skb->len % mss) != 0) | |
1771 | * tp->snd_sml = TCP_SKB_CB(skb)->end_seq; | |
1772 | * But we can avoid doing the divide again given we already have | |
1773 | * skb_pcount = skb->len / mss_now | |
0e3a4803 | 1774 | */ |
d4589926 ED |
1775 | static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now, |
1776 | const struct sk_buff *skb) | |
1777 | { | |
1778 | if (skb->len < tcp_skb_pcount(skb) * mss_now) | |
1779 | tp->snd_sml = TCP_SKB_CB(skb)->end_seq; | |
1780 | } | |
1781 | ||
1782 | /* Return false, if packet can be sent now without violation Nagle's rules: | |
1783 | * 1. It is full sized. (provided by caller in %partial bool) | |
1784 | * 2. Or it contains FIN. (already checked by caller) | |
1785 | * 3. Or TCP_CORK is not set, and TCP_NODELAY is set. | |
1786 | * 4. Or TCP_CORK is not set, and all sent packets are ACKed. | |
1787 | * With Minshall's modification: all sent small packets are ACKed. | |
1788 | */ | |
1789 | static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp, | |
cc93fc51 | 1790 | int nonagle) |
d4589926 ED |
1791 | { |
1792 | return partial && | |
1793 | ((nonagle & TCP_NAGLE_CORK) || | |
1794 | (!nonagle && tp->packets_out && tcp_minshall_check(tp))); | |
1795 | } | |
605ad7f1 ED |
1796 | |
1797 | /* Return how many segs we'd like on a TSO packet, | |
1798 | * to send one TSO packet per ms | |
1799 | */ | |
dcb8c9b4 ED |
1800 | static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now, |
1801 | int min_tso_segs) | |
605ad7f1 ED |
1802 | { |
1803 | u32 bytes, segs; | |
1804 | ||
76a9ebe8 | 1805 | bytes = min_t(unsigned long, |
7c68fa2b | 1806 | sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift), |
76a9ebe8 | 1807 | sk->sk_gso_max_size - 1 - MAX_TCP_HEADER); |
605ad7f1 ED |
1808 | |
1809 | /* Goal is to send at least one packet per ms, | |
1810 | * not one big TSO packet every 100 ms. | |
1811 | * This preserves ACK clocking and is consistent | |
1812 | * with tcp_tso_should_defer() heuristic. | |
1813 | */ | |
1b3878ca | 1814 | segs = max_t(u32, bytes / mss_now, min_tso_segs); |
605ad7f1 | 1815 | |
350c9f48 | 1816 | return segs; |
605ad7f1 ED |
1817 | } |
1818 | ||
ed6e7268 NC |
1819 | /* Return the number of segments we want in the skb we are transmitting. |
1820 | * See if congestion control module wants to decide; otherwise, autosize. | |
1821 | */ | |
1822 | static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now) | |
1823 | { | |
1824 | const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops; | |
dcb8c9b4 | 1825 | u32 min_tso, tso_segs; |
ed6e7268 | 1826 | |
dcb8c9b4 ED |
1827 | min_tso = ca_ops->min_tso_segs ? |
1828 | ca_ops->min_tso_segs(sk) : | |
1829 | sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs; | |
1830 | ||
1831 | tso_segs = tcp_tso_autosize(sk, mss_now, min_tso); | |
350c9f48 | 1832 | return min_t(u32, tso_segs, sk->sk_gso_max_segs); |
ed6e7268 NC |
1833 | } |
1834 | ||
d4589926 ED |
1835 | /* Returns the portion of skb which can be sent right away */ |
1836 | static unsigned int tcp_mss_split_point(const struct sock *sk, | |
1837 | const struct sk_buff *skb, | |
1838 | unsigned int mss_now, | |
1839 | unsigned int max_segs, | |
1840 | int nonagle) | |
c1b4a7e6 | 1841 | { |
cf533ea5 | 1842 | const struct tcp_sock *tp = tcp_sk(sk); |
d4589926 | 1843 | u32 partial, needed, window, max_len; |
c1b4a7e6 | 1844 | |
90840def | 1845 | window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
1485348d | 1846 | max_len = mss_now * max_segs; |
0e3a4803 | 1847 | |
1485348d BH |
1848 | if (likely(max_len <= window && skb != tcp_write_queue_tail(sk))) |
1849 | return max_len; | |
0e3a4803 | 1850 | |
5ea3a748 IJ |
1851 | needed = min(skb->len, window); |
1852 | ||
1485348d BH |
1853 | if (max_len <= needed) |
1854 | return max_len; | |
0e3a4803 | 1855 | |
d4589926 ED |
1856 | partial = needed % mss_now; |
1857 | /* If last segment is not a full MSS, check if Nagle rules allow us | |
1858 | * to include this last segment in this skb. | |
1859 | * Otherwise, we'll split the skb at last MSS boundary | |
1860 | */ | |
cc93fc51 | 1861 | if (tcp_nagle_check(partial != 0, tp, nonagle)) |
d4589926 ED |
1862 | return needed - partial; |
1863 | ||
1864 | return needed; | |
c1b4a7e6 DM |
1865 | } |
1866 | ||
1867 | /* Can at least one segment of SKB be sent right now, according to the | |
1868 | * congestion window rules? If so, return how many segments are allowed. | |
1869 | */ | |
cf533ea5 ED |
1870 | static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp, |
1871 | const struct sk_buff *skb) | |
c1b4a7e6 | 1872 | { |
d649a7a8 | 1873 | u32 in_flight, cwnd, halfcwnd; |
c1b4a7e6 DM |
1874 | |
1875 | /* Don't be strict about the congestion window for the final FIN. */ | |
4de075e0 ED |
1876 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) && |
1877 | tcp_skb_pcount(skb) == 1) | |
c1b4a7e6 DM |
1878 | return 1; |
1879 | ||
1880 | in_flight = tcp_packets_in_flight(tp); | |
1881 | cwnd = tp->snd_cwnd; | |
d649a7a8 ED |
1882 | if (in_flight >= cwnd) |
1883 | return 0; | |
c1b4a7e6 | 1884 | |
d649a7a8 ED |
1885 | /* For better scheduling, ensure we have at least |
1886 | * 2 GSO packets in flight. | |
1887 | */ | |
1888 | halfcwnd = max(cwnd >> 1, 1U); | |
1889 | return min(halfcwnd, cwnd - in_flight); | |
c1b4a7e6 DM |
1890 | } |
1891 | ||
b595076a | 1892 | /* Initialize TSO state of a skb. |
67edfef7 | 1893 | * This must be invoked the first time we consider transmitting |
c1b4a7e6 DM |
1894 | * SKB onto the wire. |
1895 | */ | |
5bbb432c | 1896 | static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now) |
c1b4a7e6 DM |
1897 | { |
1898 | int tso_segs = tcp_skb_pcount(skb); | |
1899 | ||
f8269a49 | 1900 | if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) { |
5bbb432c | 1901 | tcp_set_skb_tso_segs(skb, mss_now); |
c1b4a7e6 DM |
1902 | tso_segs = tcp_skb_pcount(skb); |
1903 | } | |
1904 | return tso_segs; | |
1905 | } | |
1906 | ||
c1b4a7e6 | 1907 | |
a2a385d6 | 1908 | /* Return true if the Nagle test allows this packet to be |
c1b4a7e6 DM |
1909 | * sent now. |
1910 | */ | |
a2a385d6 ED |
1911 | static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb, |
1912 | unsigned int cur_mss, int nonagle) | |
c1b4a7e6 DM |
1913 | { |
1914 | /* Nagle rule does not apply to frames, which sit in the middle of the | |
1915 | * write_queue (they have no chances to get new data). | |
1916 | * | |
1917 | * This is implemented in the callers, where they modify the 'nonagle' | |
1918 | * argument based upon the location of SKB in the send queue. | |
1919 | */ | |
1920 | if (nonagle & TCP_NAGLE_PUSH) | |
a2a385d6 | 1921 | return true; |
c1b4a7e6 | 1922 | |
9b44190d YC |
1923 | /* Don't use the nagle rule for urgent data (or for the final FIN). */ |
1924 | if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)) | |
a2a385d6 | 1925 | return true; |
c1b4a7e6 | 1926 | |
cc93fc51 | 1927 | if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle)) |
a2a385d6 | 1928 | return true; |
c1b4a7e6 | 1929 | |
a2a385d6 | 1930 | return false; |
c1b4a7e6 DM |
1931 | } |
1932 | ||
1933 | /* Does at least the first segment of SKB fit into the send window? */ | |
a2a385d6 ED |
1934 | static bool tcp_snd_wnd_test(const struct tcp_sock *tp, |
1935 | const struct sk_buff *skb, | |
1936 | unsigned int cur_mss) | |
c1b4a7e6 DM |
1937 | { |
1938 | u32 end_seq = TCP_SKB_CB(skb)->end_seq; | |
1939 | ||
1940 | if (skb->len > cur_mss) | |
1941 | end_seq = TCP_SKB_CB(skb)->seq + cur_mss; | |
1942 | ||
90840def | 1943 | return !after(end_seq, tcp_wnd_end(tp)); |
c1b4a7e6 DM |
1944 | } |
1945 | ||
c1b4a7e6 DM |
1946 | /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet |
1947 | * which is put after SKB on the list. It is very much like | |
1948 | * tcp_fragment() except that it may make several kinds of assumptions | |
1949 | * in order to speed up the splitting operation. In particular, we | |
1950 | * know that all the data is in scatter-gather pages, and that the | |
1951 | * packet has never been sent out before (and thus is not cloned). | |
1952 | */ | |
56483341 | 1953 | static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len, |
c4ead4c5 | 1954 | unsigned int mss_now, gfp_t gfp) |
c1b4a7e6 | 1955 | { |
c1b4a7e6 | 1956 | int nlen = skb->len - len; |
56483341 | 1957 | struct sk_buff *buff; |
9ce01461 | 1958 | u8 flags; |
c1b4a7e6 DM |
1959 | |
1960 | /* All of a TSO frame must be composed of paged data. */ | |
c8ac3774 | 1961 | if (skb->len != skb->data_len) |
56483341 ED |
1962 | return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, |
1963 | skb, len, mss_now, gfp); | |
c1b4a7e6 | 1964 | |
eb934478 | 1965 | buff = sk_stream_alloc_skb(sk, 0, gfp, true); |
51456b29 | 1966 | if (unlikely(!buff)) |
c1b4a7e6 | 1967 | return -ENOMEM; |
41477662 | 1968 | skb_copy_decrypted(buff, skb); |
c1b4a7e6 | 1969 | |
ab4e846a | 1970 | sk_wmem_queued_add(sk, buff->truesize); |
3ab224be | 1971 | sk_mem_charge(sk, buff->truesize); |
b60b49ea | 1972 | buff->truesize += nlen; |
c1b4a7e6 DM |
1973 | skb->truesize -= nlen; |
1974 | ||
1975 | /* Correct the sequence numbers. */ | |
1976 | TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len; | |
1977 | TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq; | |
1978 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq; | |
1979 | ||
1980 | /* PSH and FIN should only be set in the second packet. */ | |
4de075e0 ED |
1981 | flags = TCP_SKB_CB(skb)->tcp_flags; |
1982 | TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH); | |
1983 | TCP_SKB_CB(buff)->tcp_flags = flags; | |
c1b4a7e6 DM |
1984 | |
1985 | /* This packet was never sent out yet, so no SACK bits. */ | |
1986 | TCP_SKB_CB(buff)->sacked = 0; | |
1987 | ||
a166140e MKL |
1988 | tcp_skb_fragment_eor(skb, buff); |
1989 | ||
98be9b12 | 1990 | buff->ip_summed = CHECKSUM_PARTIAL; |
c1b4a7e6 | 1991 | skb_split(skb, buff, len); |
490cc7d0 | 1992 | tcp_fragment_tstamp(skb, buff); |
c1b4a7e6 DM |
1993 | |
1994 | /* Fix up tso_factor for both original and new SKB. */ | |
5bbb432c ED |
1995 | tcp_set_skb_tso_segs(skb, mss_now); |
1996 | tcp_set_skb_tso_segs(buff, mss_now); | |
c1b4a7e6 DM |
1997 | |
1998 | /* Link BUFF into the send queue. */ | |
f4a775d1 | 1999 | __skb_header_release(buff); |
56483341 | 2000 | tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE); |
c1b4a7e6 DM |
2001 | |
2002 | return 0; | |
2003 | } | |
2004 | ||
2005 | /* Try to defer sending, if possible, in order to minimize the amount | |
2006 | * of TSO splitting we do. View it as a kind of TSO Nagle test. | |
2007 | * | |
2008 | * This algorithm is from John Heffner. | |
2009 | */ | |
ca8a2263 | 2010 | static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb, |
f9bfe4e6 ED |
2011 | bool *is_cwnd_limited, |
2012 | bool *is_rwnd_limited, | |
2013 | u32 max_segs) | |
c1b4a7e6 | 2014 | { |
6687e988 | 2015 | const struct inet_connection_sock *icsk = inet_csk(sk); |
f1c6ea38 | 2016 | u32 send_win, cong_win, limit, in_flight; |
50c8339e | 2017 | struct tcp_sock *tp = tcp_sk(sk); |
50c8339e | 2018 | struct sk_buff *head; |
ad9f4f50 | 2019 | int win_divisor; |
f1c6ea38 | 2020 | s64 delta; |
c1b4a7e6 | 2021 | |
99d7662a | 2022 | if (icsk->icsk_ca_state >= TCP_CA_Recovery) |
ae8064ac JH |
2023 | goto send_now; |
2024 | ||
5f852eb5 | 2025 | /* Avoid bursty behavior by allowing defer |
a682850a ED |
2026 | * only if the last write was recent (1 ms). |
2027 | * Note that tp->tcp_wstamp_ns can be in the future if we have | |
2028 | * packets waiting in a qdisc or device for EDT delivery. | |
5f852eb5 | 2029 | */ |
a682850a ED |
2030 | delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC; |
2031 | if (delta > 0) | |
ae8064ac | 2032 | goto send_now; |
908a75c1 | 2033 | |
c1b4a7e6 DM |
2034 | in_flight = tcp_packets_in_flight(tp); |
2035 | ||
c8c9aeb5 SB |
2036 | BUG_ON(tcp_skb_pcount(skb) <= 1); |
2037 | BUG_ON(tp->snd_cwnd <= in_flight); | |
c1b4a7e6 | 2038 | |
90840def | 2039 | send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
c1b4a7e6 DM |
2040 | |
2041 | /* From in_flight test above, we know that cwnd > in_flight. */ | |
2042 | cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache; | |
2043 | ||
2044 | limit = min(send_win, cong_win); | |
2045 | ||
ba244fe9 | 2046 | /* If a full-sized TSO skb can be sent, do it. */ |
605ad7f1 | 2047 | if (limit >= max_segs * tp->mss_cache) |
ae8064ac | 2048 | goto send_now; |
ba244fe9 | 2049 | |
62ad2761 IJ |
2050 | /* Middle in queue won't get any more data, full sendable already? */ |
2051 | if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len)) | |
2052 | goto send_now; | |
2053 | ||
5bbcc0f5 | 2054 | win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor); |
ad9f4f50 | 2055 | if (win_divisor) { |
c1b4a7e6 DM |
2056 | u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache); |
2057 | ||
2058 | /* If at least some fraction of a window is available, | |
2059 | * just use it. | |
2060 | */ | |
ad9f4f50 | 2061 | chunk /= win_divisor; |
c1b4a7e6 | 2062 | if (limit >= chunk) |
ae8064ac | 2063 | goto send_now; |
c1b4a7e6 DM |
2064 | } else { |
2065 | /* Different approach, try not to defer past a single | |
2066 | * ACK. Receiver should ACK every other full sized | |
2067 | * frame, so if we have space for more than 3 frames | |
2068 | * then send now. | |
2069 | */ | |
6b5a5c0d | 2070 | if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache) |
ae8064ac | 2071 | goto send_now; |
c1b4a7e6 DM |
2072 | } |
2073 | ||
75c119af ED |
2074 | /* TODO : use tsorted_sent_queue ? */ |
2075 | head = tcp_rtx_queue_head(sk); | |
2076 | if (!head) | |
2077 | goto send_now; | |
f1c6ea38 | 2078 | delta = tp->tcp_clock_cache - head->tstamp; |
50c8339e | 2079 | /* If next ACK is likely to come too late (half srtt), do not defer */ |
f1c6ea38 | 2080 | if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0) |
50c8339e ED |
2081 | goto send_now; |
2082 | ||
f9bfe4e6 ED |
2083 | /* Ok, it looks like it is advisable to defer. |
2084 | * Three cases are tracked : | |
2085 | * 1) We are cwnd-limited | |
2086 | * 2) We are rwnd-limited | |
2087 | * 3) We are application limited. | |
2088 | */ | |
2089 | if (cong_win < send_win) { | |
2090 | if (cong_win <= skb->len) { | |
2091 | *is_cwnd_limited = true; | |
2092 | return true; | |
2093 | } | |
2094 | } else { | |
2095 | if (send_win <= skb->len) { | |
2096 | *is_rwnd_limited = true; | |
2097 | return true; | |
2098 | } | |
2099 | } | |
ae8064ac | 2100 | |
f9bfe4e6 | 2101 | /* If this packet won't get more data, do not wait. */ |
d8ed257f ED |
2102 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) || |
2103 | TCP_SKB_CB(skb)->eor) | |
f9bfe4e6 | 2104 | goto send_now; |
ca8a2263 | 2105 | |
a2a385d6 | 2106 | return true; |
ae8064ac JH |
2107 | |
2108 | send_now: | |
a2a385d6 | 2109 | return false; |
c1b4a7e6 DM |
2110 | } |
2111 | ||
05cbc0db FD |
2112 | static inline void tcp_mtu_check_reprobe(struct sock *sk) |
2113 | { | |
2114 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2115 | struct tcp_sock *tp = tcp_sk(sk); | |
2116 | struct net *net = sock_net(sk); | |
2117 | u32 interval; | |
2118 | s32 delta; | |
2119 | ||
2120 | interval = net->ipv4.sysctl_tcp_probe_interval; | |
c74df29a | 2121 | delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp; |
05cbc0db FD |
2122 | if (unlikely(delta >= interval * HZ)) { |
2123 | int mss = tcp_current_mss(sk); | |
2124 | ||
2125 | /* Update current search range */ | |
2126 | icsk->icsk_mtup.probe_size = 0; | |
2127 | icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + | |
2128 | sizeof(struct tcphdr) + | |
2129 | icsk->icsk_af_ops->net_header_len; | |
2130 | icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss); | |
2131 | ||
2132 | /* Update probe time stamp */ | |
c74df29a | 2133 | icsk->icsk_mtup.probe_timestamp = tcp_jiffies32; |
05cbc0db FD |
2134 | } |
2135 | } | |
2136 | ||
808cf9e3 IL |
2137 | static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len) |
2138 | { | |
2139 | struct sk_buff *skb, *next; | |
2140 | ||
2141 | skb = tcp_send_head(sk); | |
2142 | tcp_for_write_queue_from_safe(skb, next, sk) { | |
2143 | if (len <= skb->len) | |
2144 | break; | |
2145 | ||
888a5c53 | 2146 | if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb)) |
808cf9e3 IL |
2147 | return false; |
2148 | ||
2149 | len -= skb->len; | |
2150 | } | |
2151 | ||
2152 | return true; | |
2153 | } | |
2154 | ||
5d424d5a | 2155 | /* Create a new MTU probe if we are ready. |
67edfef7 AK |
2156 | * MTU probe is regularly attempting to increase the path MTU by |
2157 | * deliberately sending larger packets. This discovers routing | |
2158 | * changes resulting in larger path MTUs. | |
2159 | * | |
5d424d5a JH |
2160 | * Returns 0 if we should wait to probe (no cwnd available), |
2161 | * 1 if a probe was sent, | |
056834d9 IJ |
2162 | * -1 otherwise |
2163 | */ | |
5d424d5a JH |
2164 | static int tcp_mtu_probe(struct sock *sk) |
2165 | { | |
5d424d5a | 2166 | struct inet_connection_sock *icsk = inet_csk(sk); |
12a59abc | 2167 | struct tcp_sock *tp = tcp_sk(sk); |
5d424d5a | 2168 | struct sk_buff *skb, *nskb, *next; |
6b58e0a5 | 2169 | struct net *net = sock_net(sk); |
5d424d5a | 2170 | int probe_size; |
91cc17c0 | 2171 | int size_needed; |
12a59abc | 2172 | int copy, len; |
5d424d5a | 2173 | int mss_now; |
6b58e0a5 | 2174 | int interval; |
5d424d5a JH |
2175 | |
2176 | /* Not currently probing/verifying, | |
2177 | * not in recovery, | |
2178 | * have enough cwnd, and | |
12a59abc ED |
2179 | * not SACKing (the variable headers throw things off) |
2180 | */ | |
2181 | if (likely(!icsk->icsk_mtup.enabled || | |
2182 | icsk->icsk_mtup.probe_size || | |
2183 | inet_csk(sk)->icsk_ca_state != TCP_CA_Open || | |
2184 | tp->snd_cwnd < 11 || | |
2185 | tp->rx_opt.num_sacks || tp->rx_opt.dsack)) | |
5d424d5a JH |
2186 | return -1; |
2187 | ||
6b58e0a5 FD |
2188 | /* Use binary search for probe_size between tcp_mss_base, |
2189 | * and current mss_clamp. if (search_high - search_low) | |
2190 | * smaller than a threshold, backoff from probing. | |
2191 | */ | |
0c54b85f | 2192 | mss_now = tcp_current_mss(sk); |
6b58e0a5 FD |
2193 | probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high + |
2194 | icsk->icsk_mtup.search_low) >> 1); | |
91cc17c0 | 2195 | size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache; |
6b58e0a5 | 2196 | interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low; |
05cbc0db FD |
2197 | /* When misfortune happens, we are reprobing actively, |
2198 | * and then reprobe timer has expired. We stick with current | |
2199 | * probing process by not resetting search range to its orignal. | |
2200 | */ | |
6b58e0a5 | 2201 | if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) || |
05cbc0db FD |
2202 | interval < net->ipv4.sysctl_tcp_probe_threshold) { |
2203 | /* Check whether enough time has elaplased for | |
2204 | * another round of probing. | |
2205 | */ | |
2206 | tcp_mtu_check_reprobe(sk); | |
5d424d5a JH |
2207 | return -1; |
2208 | } | |
2209 | ||
2210 | /* Have enough data in the send queue to probe? */ | |
7f9c33e5 | 2211 | if (tp->write_seq - tp->snd_nxt < size_needed) |
5d424d5a JH |
2212 | return -1; |
2213 | ||
91cc17c0 IJ |
2214 | if (tp->snd_wnd < size_needed) |
2215 | return -1; | |
90840def | 2216 | if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp))) |
91cc17c0 | 2217 | return 0; |
5d424d5a | 2218 | |
d67c58e9 IJ |
2219 | /* Do we need to wait to drain cwnd? With none in flight, don't stall */ |
2220 | if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) { | |
2221 | if (!tcp_packets_in_flight(tp)) | |
5d424d5a JH |
2222 | return -1; |
2223 | else | |
2224 | return 0; | |
2225 | } | |
2226 | ||
808cf9e3 IL |
2227 | if (!tcp_can_coalesce_send_queue_head(sk, probe_size)) |
2228 | return -1; | |
2229 | ||
5d424d5a | 2230 | /* We're allowed to probe. Build it now. */ |
eb934478 | 2231 | nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false); |
51456b29 | 2232 | if (!nskb) |
5d424d5a | 2233 | return -1; |
ab4e846a | 2234 | sk_wmem_queued_add(sk, nskb->truesize); |
3ab224be | 2235 | sk_mem_charge(sk, nskb->truesize); |
5d424d5a | 2236 | |
fe067e8a | 2237 | skb = tcp_send_head(sk); |
41477662 | 2238 | skb_copy_decrypted(nskb, skb); |
5d424d5a JH |
2239 | |
2240 | TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq; | |
2241 | TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size; | |
4de075e0 | 2242 | TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK; |
5d424d5a JH |
2243 | TCP_SKB_CB(nskb)->sacked = 0; |
2244 | nskb->csum = 0; | |
98be9b12 | 2245 | nskb->ip_summed = CHECKSUM_PARTIAL; |
5d424d5a | 2246 | |
50c4817e | 2247 | tcp_insert_write_queue_before(nskb, skb, sk); |
2b7cda9c | 2248 | tcp_highest_sack_replace(sk, skb, nskb); |
50c4817e | 2249 | |
5d424d5a | 2250 | len = 0; |
234b6860 | 2251 | tcp_for_write_queue_from_safe(skb, next, sk) { |
5d424d5a | 2252 | copy = min_t(int, skb->len, probe_size - len); |
98be9b12 | 2253 | skb_copy_bits(skb, 0, skb_put(nskb, copy), copy); |
5d424d5a JH |
2254 | |
2255 | if (skb->len <= copy) { | |
2256 | /* We've eaten all the data from this skb. | |
2257 | * Throw it away. */ | |
4de075e0 | 2258 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags; |
808cf9e3 IL |
2259 | /* If this is the last SKB we copy and eor is set |
2260 | * we need to propagate it to the new skb. | |
2261 | */ | |
2262 | TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor; | |
888a5c53 | 2263 | tcp_skb_collapse_tstamp(nskb, skb); |
fe067e8a | 2264 | tcp_unlink_write_queue(skb, sk); |
3ab224be | 2265 | sk_wmem_free_skb(sk, skb); |
5d424d5a | 2266 | } else { |
4de075e0 | 2267 | TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags & |
a3433f35 | 2268 | ~(TCPHDR_FIN|TCPHDR_PSH); |
5d424d5a JH |
2269 | if (!skb_shinfo(skb)->nr_frags) { |
2270 | skb_pull(skb, copy); | |
5d424d5a JH |
2271 | } else { |
2272 | __pskb_trim_head(skb, copy); | |
5bbb432c | 2273 | tcp_set_skb_tso_segs(skb, mss_now); |
5d424d5a JH |
2274 | } |
2275 | TCP_SKB_CB(skb)->seq += copy; | |
2276 | } | |
2277 | ||
2278 | len += copy; | |
234b6860 IJ |
2279 | |
2280 | if (len >= probe_size) | |
2281 | break; | |
5d424d5a | 2282 | } |
5bbb432c | 2283 | tcp_init_tso_segs(nskb, nskb->len); |
5d424d5a JH |
2284 | |
2285 | /* We're ready to send. If this fails, the probe will | |
7faee5c0 ED |
2286 | * be resegmented into mss-sized pieces by tcp_write_xmit(). |
2287 | */ | |
5d424d5a JH |
2288 | if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) { |
2289 | /* Decrement cwnd here because we are sending | |
056834d9 | 2290 | * effectively two packets. */ |
5d424d5a | 2291 | tp->snd_cwnd--; |
66f5fe62 | 2292 | tcp_event_new_data_sent(sk, nskb); |
5d424d5a JH |
2293 | |
2294 | icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len); | |
0e7b1368 JH |
2295 | tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq; |
2296 | tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq; | |
5d424d5a JH |
2297 | |
2298 | return 1; | |
2299 | } | |
2300 | ||
2301 | return -1; | |
2302 | } | |
2303 | ||
864e5c09 | 2304 | static bool tcp_pacing_check(struct sock *sk) |
218af599 | 2305 | { |
864e5c09 ED |
2306 | struct tcp_sock *tp = tcp_sk(sk); |
2307 | ||
2308 | if (!tcp_needs_internal_pacing(sk)) | |
2309 | return false; | |
2310 | ||
2311 | if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache) | |
2312 | return false; | |
2313 | ||
2314 | if (!hrtimer_is_queued(&tp->pacing_timer)) { | |
2315 | hrtimer_start(&tp->pacing_timer, | |
2316 | ns_to_ktime(tp->tcp_wstamp_ns), | |
2317 | HRTIMER_MODE_ABS_PINNED_SOFT); | |
2318 | sock_hold(sk); | |
2319 | } | |
2320 | return true; | |
218af599 ED |
2321 | } |
2322 | ||
f9616c35 ED |
2323 | /* TCP Small Queues : |
2324 | * Control number of packets in qdisc/devices to two packets / or ~1 ms. | |
2325 | * (These limits are doubled for retransmits) | |
2326 | * This allows for : | |
2327 | * - better RTT estimation and ACK scheduling | |
2328 | * - faster recovery | |
2329 | * - high rates | |
2330 | * Alas, some drivers / subsystems require a fair amount | |
2331 | * of queued bytes to ensure line rate. | |
2332 | * One example is wifi aggregation (802.11 AMPDU) | |
2333 | */ | |
2334 | static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb, | |
2335 | unsigned int factor) | |
2336 | { | |
76a9ebe8 | 2337 | unsigned long limit; |
f9616c35 | 2338 | |
76a9ebe8 ED |
2339 | limit = max_t(unsigned long, |
2340 | 2 * skb->truesize, | |
7c68fa2b | 2341 | sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift)); |
c73e5807 ED |
2342 | if (sk->sk_pacing_status == SK_PACING_NONE) |
2343 | limit = min_t(unsigned long, limit, | |
2344 | sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes); | |
f9616c35 ED |
2345 | limit <<= factor; |
2346 | ||
a842fe14 ED |
2347 | if (static_branch_unlikely(&tcp_tx_delay_enabled) && |
2348 | tcp_sk(sk)->tcp_tx_delay) { | |
2349 | u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay; | |
2350 | ||
2351 | /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we | |
2352 | * approximate our needs assuming an ~100% skb->truesize overhead. | |
2353 | * USEC_PER_SEC is approximated by 2^20. | |
2354 | * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift. | |
2355 | */ | |
2356 | extra_bytes >>= (20 - 1); | |
2357 | limit += extra_bytes; | |
2358 | } | |
14afee4b | 2359 | if (refcount_read(&sk->sk_wmem_alloc) > limit) { |
75c119af | 2360 | /* Always send skb if rtx queue is empty. |
75eefc6c ED |
2361 | * No need to wait for TX completion to call us back, |
2362 | * after softirq/tasklet schedule. | |
2363 | * This helps when TX completions are delayed too much. | |
2364 | */ | |
75c119af | 2365 | if (tcp_rtx_queue_empty(sk)) |
75eefc6c ED |
2366 | return false; |
2367 | ||
7aa5470c | 2368 | set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags); |
f9616c35 ED |
2369 | /* It is possible TX completion already happened |
2370 | * before we set TSQ_THROTTLED, so we must | |
2371 | * test again the condition. | |
2372 | */ | |
2373 | smp_mb__after_atomic(); | |
14afee4b | 2374 | if (refcount_read(&sk->sk_wmem_alloc) > limit) |
f9616c35 ED |
2375 | return true; |
2376 | } | |
2377 | return false; | |
2378 | } | |
2379 | ||
05b055e8 FY |
2380 | static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new) |
2381 | { | |
628174cc | 2382 | const u32 now = tcp_jiffies32; |
efe967cd | 2383 | enum tcp_chrono old = tp->chrono_type; |
05b055e8 | 2384 | |
efe967cd AB |
2385 | if (old > TCP_CHRONO_UNSPEC) |
2386 | tp->chrono_stat[old - 1] += now - tp->chrono_start; | |
05b055e8 FY |
2387 | tp->chrono_start = now; |
2388 | tp->chrono_type = new; | |
2389 | } | |
2390 | ||
2391 | void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type) | |
2392 | { | |
2393 | struct tcp_sock *tp = tcp_sk(sk); | |
2394 | ||
2395 | /* If there are multiple conditions worthy of tracking in a | |
0f87230d FY |
2396 | * chronograph then the highest priority enum takes precedence |
2397 | * over the other conditions. So that if something "more interesting" | |
05b055e8 FY |
2398 | * starts happening, stop the previous chrono and start a new one. |
2399 | */ | |
2400 | if (type > tp->chrono_type) | |
2401 | tcp_chrono_set(tp, type); | |
2402 | } | |
2403 | ||
2404 | void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type) | |
2405 | { | |
2406 | struct tcp_sock *tp = tcp_sk(sk); | |
2407 | ||
0f87230d FY |
2408 | |
2409 | /* There are multiple conditions worthy of tracking in a | |
2410 | * chronograph, so that the highest priority enum takes | |
2411 | * precedence over the other conditions (see tcp_chrono_start). | |
2412 | * If a condition stops, we only stop chrono tracking if | |
2413 | * it's the "most interesting" or current chrono we are | |
2414 | * tracking and starts busy chrono if we have pending data. | |
2415 | */ | |
75c119af | 2416 | if (tcp_rtx_and_write_queues_empty(sk)) |
0f87230d FY |
2417 | tcp_chrono_set(tp, TCP_CHRONO_UNSPEC); |
2418 | else if (type == tp->chrono_type) | |
2419 | tcp_chrono_set(tp, TCP_CHRONO_BUSY); | |
05b055e8 FY |
2420 | } |
2421 | ||
1da177e4 LT |
2422 | /* This routine writes packets to the network. It advances the |
2423 | * send_head. This happens as incoming acks open up the remote | |
2424 | * window for us. | |
2425 | * | |
f8269a49 IJ |
2426 | * LARGESEND note: !tcp_urg_mode is overkill, only frames between |
2427 | * snd_up-64k-mss .. snd_up cannot be large. However, taking into | |
2428 | * account rare use of URG, this is not a big flaw. | |
2429 | * | |
6ba8a3b1 ND |
2430 | * Send at most one packet when push_one > 0. Temporarily ignore |
2431 | * cwnd limit to force at most one packet out when push_one == 2. | |
2432 | ||
a2a385d6 ED |
2433 | * Returns true, if no segments are in flight and we have queued segments, |
2434 | * but cannot send anything now because of SWS or another problem. | |
1da177e4 | 2435 | */ |
a2a385d6 ED |
2436 | static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle, |
2437 | int push_one, gfp_t gfp) | |
1da177e4 LT |
2438 | { |
2439 | struct tcp_sock *tp = tcp_sk(sk); | |
92df7b51 | 2440 | struct sk_buff *skb; |
c1b4a7e6 DM |
2441 | unsigned int tso_segs, sent_pkts; |
2442 | int cwnd_quota; | |
5d424d5a | 2443 | int result; |
5615f886 | 2444 | bool is_cwnd_limited = false, is_rwnd_limited = false; |
605ad7f1 | 2445 | u32 max_segs; |
1da177e4 | 2446 | |
92df7b51 | 2447 | sent_pkts = 0; |
5d424d5a | 2448 | |
ee1836ae | 2449 | tcp_mstamp_refresh(tp); |
d5dd9175 IJ |
2450 | if (!push_one) { |
2451 | /* Do MTU probing. */ | |
2452 | result = tcp_mtu_probe(sk); | |
2453 | if (!result) { | |
a2a385d6 | 2454 | return false; |
d5dd9175 IJ |
2455 | } else if (result > 0) { |
2456 | sent_pkts = 1; | |
2457 | } | |
5d424d5a JH |
2458 | } |
2459 | ||
ed6e7268 | 2460 | max_segs = tcp_tso_segs(sk, mss_now); |
fe067e8a | 2461 | while ((skb = tcp_send_head(sk))) { |
c8ac3774 HX |
2462 | unsigned int limit; |
2463 | ||
79861919 ED |
2464 | if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) { |
2465 | /* "skb_mstamp_ns" is used as a start point for the retransmit timer */ | |
2466 | skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache; | |
2467 | list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue); | |
bf50b606 | 2468 | tcp_init_tso_segs(skb, mss_now); |
79861919 ED |
2469 | goto repair; /* Skip network transmission */ |
2470 | } | |
2471 | ||
218af599 ED |
2472 | if (tcp_pacing_check(sk)) |
2473 | break; | |
2474 | ||
5bbb432c | 2475 | tso_segs = tcp_init_tso_segs(skb, mss_now); |
c1b4a7e6 | 2476 | BUG_ON(!tso_segs); |
aa93466b | 2477 | |
b68e9f85 | 2478 | cwnd_quota = tcp_cwnd_test(tp, skb); |
6ba8a3b1 ND |
2479 | if (!cwnd_quota) { |
2480 | if (push_one == 2) | |
2481 | /* Force out a loss probe pkt. */ | |
2482 | cwnd_quota = 1; | |
2483 | else | |
2484 | break; | |
2485 | } | |
b68e9f85 | 2486 | |
5615f886 FY |
2487 | if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) { |
2488 | is_rwnd_limited = true; | |
b68e9f85 | 2489 | break; |
5615f886 | 2490 | } |
b68e9f85 | 2491 | |
d6a4e26a | 2492 | if (tso_segs == 1) { |
c1b4a7e6 DM |
2493 | if (unlikely(!tcp_nagle_test(tp, skb, mss_now, |
2494 | (tcp_skb_is_last(sk, skb) ? | |
2495 | nonagle : TCP_NAGLE_PUSH)))) | |
2496 | break; | |
2497 | } else { | |
ca8a2263 | 2498 | if (!push_one && |
605ad7f1 | 2499 | tcp_tso_should_defer(sk, skb, &is_cwnd_limited, |
f9bfe4e6 | 2500 | &is_rwnd_limited, max_segs)) |
c1b4a7e6 DM |
2501 | break; |
2502 | } | |
aa93466b | 2503 | |
605ad7f1 | 2504 | limit = mss_now; |
d6a4e26a | 2505 | if (tso_segs > 1 && !tcp_urg_mode(tp)) |
605ad7f1 ED |
2506 | limit = tcp_mss_split_point(sk, skb, mss_now, |
2507 | min_t(unsigned int, | |
2508 | cwnd_quota, | |
2509 | max_segs), | |
2510 | nonagle); | |
2511 | ||
2512 | if (skb->len > limit && | |
56483341 | 2513 | unlikely(tso_fragment(sk, skb, limit, mss_now, gfp))) |
605ad7f1 ED |
2514 | break; |
2515 | ||
f9616c35 ED |
2516 | if (tcp_small_queue_check(sk, skb, 0)) |
2517 | break; | |
c9eeec26 | 2518 | |
1f85e626 ED |
2519 | /* Argh, we hit an empty skb(), presumably a thread |
2520 | * is sleeping in sendmsg()/sk_stream_wait_memory(). | |
2521 | * We do not want to send a pure-ack packet and have | |
2522 | * a strange looking rtx queue with empty packet(s). | |
2523 | */ | |
2524 | if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq) | |
2525 | break; | |
2526 | ||
d5dd9175 | 2527 | if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp))) |
92df7b51 | 2528 | break; |
1da177e4 | 2529 | |
ec342325 | 2530 | repair: |
92df7b51 DM |
2531 | /* Advance the send_head. This one is sent out. |
2532 | * This call will increment packets_out. | |
2533 | */ | |
66f5fe62 | 2534 | tcp_event_new_data_sent(sk, skb); |
1da177e4 | 2535 | |
92df7b51 | 2536 | tcp_minshall_update(tp, mss_now, skb); |
a262f0cd | 2537 | sent_pkts += tcp_skb_pcount(skb); |
d5dd9175 IJ |
2538 | |
2539 | if (push_one) | |
2540 | break; | |
92df7b51 | 2541 | } |
1da177e4 | 2542 | |
5615f886 FY |
2543 | if (is_rwnd_limited) |
2544 | tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED); | |
2545 | else | |
2546 | tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED); | |
2547 | ||
aa93466b | 2548 | if (likely(sent_pkts)) { |
684bad11 YC |
2549 | if (tcp_in_cwnd_reduction(sk)) |
2550 | tp->prr_out += sent_pkts; | |
6ba8a3b1 ND |
2551 | |
2552 | /* Send one loss probe per tail loss episode. */ | |
2553 | if (push_one != 2) | |
ed66dfaf | 2554 | tcp_schedule_loss_probe(sk, false); |
d2e1339f | 2555 | is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd); |
ca8a2263 | 2556 | tcp_cwnd_validate(sk, is_cwnd_limited); |
a2a385d6 | 2557 | return false; |
1da177e4 | 2558 | } |
75c119af | 2559 | return !tp->packets_out && !tcp_write_queue_empty(sk); |
6ba8a3b1 ND |
2560 | } |
2561 | ||
ed66dfaf | 2562 | bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto) |
6ba8a3b1 ND |
2563 | { |
2564 | struct inet_connection_sock *icsk = inet_csk(sk); | |
2565 | struct tcp_sock *tp = tcp_sk(sk); | |
a2815817 | 2566 | u32 timeout, rto_delta_us; |
2ae21cf5 | 2567 | int early_retrans; |
6ba8a3b1 | 2568 | |
6ba8a3b1 ND |
2569 | /* Don't do any loss probe on a Fast Open connection before 3WHS |
2570 | * finishes. | |
2571 | */ | |
d983ea6f | 2572 | if (rcu_access_pointer(tp->fastopen_rsk)) |
6ba8a3b1 ND |
2573 | return false; |
2574 | ||
2ae21cf5 | 2575 | early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans; |
6ba8a3b1 | 2576 | /* Schedule a loss probe in 2*RTT for SACK capable connections |
b4f70c3d | 2577 | * not in loss recovery, that are either limited by cwnd or application. |
6ba8a3b1 | 2578 | */ |
2ae21cf5 | 2579 | if ((early_retrans != 3 && early_retrans != 4) || |
bec41a11 | 2580 | !tp->packets_out || !tcp_is_sack(tp) || |
b4f70c3d NC |
2581 | (icsk->icsk_ca_state != TCP_CA_Open && |
2582 | icsk->icsk_ca_state != TCP_CA_CWR)) | |
6ba8a3b1 ND |
2583 | return false; |
2584 | ||
bb4d991a | 2585 | /* Probe timeout is 2*rtt. Add minimum RTO to account |
f9b99582 YC |
2586 | * for delayed ack when there's one outstanding packet. If no RTT |
2587 | * sample is available then probe after TCP_TIMEOUT_INIT. | |
6ba8a3b1 | 2588 | */ |
bb4d991a YC |
2589 | if (tp->srtt_us) { |
2590 | timeout = usecs_to_jiffies(tp->srtt_us >> 2); | |
2591 | if (tp->packets_out == 1) | |
2592 | timeout += TCP_RTO_MIN; | |
2593 | else | |
2594 | timeout += TCP_TIMEOUT_MIN; | |
2595 | } else { | |
2596 | timeout = TCP_TIMEOUT_INIT; | |
2597 | } | |
6ba8a3b1 | 2598 | |
a2815817 | 2599 | /* If the RTO formula yields an earlier time, then use that time. */ |
ed66dfaf NC |
2600 | rto_delta_us = advancing_rto ? |
2601 | jiffies_to_usecs(inet_csk(sk)->icsk_rto) : | |
2602 | tcp_rto_delta_us(sk); /* How far in future is RTO? */ | |
a2815817 NC |
2603 | if (rto_delta_us > 0) |
2604 | timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us)); | |
6ba8a3b1 | 2605 | |
8dc242ad | 2606 | tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX); |
6ba8a3b1 ND |
2607 | return true; |
2608 | } | |
2609 | ||
1f3279ae ED |
2610 | /* Thanks to skb fast clones, we can detect if a prior transmit of |
2611 | * a packet is still in a qdisc or driver queue. | |
2612 | * In this case, there is very little point doing a retransmit ! | |
1f3279ae ED |
2613 | */ |
2614 | static bool skb_still_in_host_queue(const struct sock *sk, | |
2615 | const struct sk_buff *skb) | |
2616 | { | |
39bb5e62 | 2617 | if (unlikely(skb_fclone_busy(sk, skb))) { |
c10d9310 ED |
2618 | NET_INC_STATS(sock_net(sk), |
2619 | LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES); | |
1f3279ae ED |
2620 | return true; |
2621 | } | |
2622 | return false; | |
2623 | } | |
2624 | ||
b340b264 | 2625 | /* When probe timeout (PTO) fires, try send a new segment if possible, else |
6ba8a3b1 ND |
2626 | * retransmit the last segment. |
2627 | */ | |
2628 | void tcp_send_loss_probe(struct sock *sk) | |
2629 | { | |
9b717a8d | 2630 | struct tcp_sock *tp = tcp_sk(sk); |
6ba8a3b1 ND |
2631 | struct sk_buff *skb; |
2632 | int pcount; | |
2633 | int mss = tcp_current_mss(sk); | |
6ba8a3b1 | 2634 | |
76be93fc YC |
2635 | /* At most one outstanding TLP */ |
2636 | if (tp->tlp_high_seq) | |
2637 | goto rearm_timer; | |
2638 | ||
2639 | tp->tlp_retrans = 0; | |
b340b264 | 2640 | skb = tcp_send_head(sk); |
75c119af ED |
2641 | if (skb && tcp_snd_wnd_test(tp, skb, mss)) { |
2642 | pcount = tp->packets_out; | |
2643 | tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC); | |
2644 | if (tp->packets_out > pcount) | |
2645 | goto probe_sent; | |
2646 | goto rearm_timer; | |
6ba8a3b1 | 2647 | } |
75c119af | 2648 | skb = skb_rb_last(&sk->tcp_rtx_queue); |
b2b7af86 YC |
2649 | if (unlikely(!skb)) { |
2650 | WARN_ONCE(tp->packets_out, | |
2651 | "invalid inflight: %u state %u cwnd %u mss %d\n", | |
2652 | tp->packets_out, sk->sk_state, tp->snd_cwnd, mss); | |
2653 | inet_csk(sk)->icsk_pending = 0; | |
2654 | return; | |
2655 | } | |
6ba8a3b1 | 2656 | |
1f3279ae ED |
2657 | if (skb_still_in_host_queue(sk, skb)) |
2658 | goto rearm_timer; | |
2659 | ||
6ba8a3b1 ND |
2660 | pcount = tcp_skb_pcount(skb); |
2661 | if (WARN_ON(!pcount)) | |
2662 | goto rearm_timer; | |
2663 | ||
2664 | if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) { | |
75c119af ED |
2665 | if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, |
2666 | (pcount - 1) * mss, mss, | |
6cc55e09 | 2667 | GFP_ATOMIC))) |
6ba8a3b1 | 2668 | goto rearm_timer; |
75c119af | 2669 | skb = skb_rb_next(skb); |
6ba8a3b1 ND |
2670 | } |
2671 | ||
2672 | if (WARN_ON(!skb || !tcp_skb_pcount(skb))) | |
2673 | goto rearm_timer; | |
2674 | ||
10d3be56 | 2675 | if (__tcp_retransmit_skb(sk, skb, 1)) |
b340b264 | 2676 | goto rearm_timer; |
6ba8a3b1 | 2677 | |
76be93fc YC |
2678 | tp->tlp_retrans = 1; |
2679 | ||
2680 | probe_sent: | |
9b717a8d | 2681 | /* Record snd_nxt for loss detection. */ |
b340b264 | 2682 | tp->tlp_high_seq = tp->snd_nxt; |
9b717a8d | 2683 | |
c10d9310 | 2684 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES); |
b340b264 YC |
2685 | /* Reset s.t. tcp_rearm_rto will restart timer from now */ |
2686 | inet_csk(sk)->icsk_pending = 0; | |
6ba8a3b1 | 2687 | rearm_timer: |
fcd16c0a | 2688 | tcp_rearm_rto(sk); |
1da177e4 LT |
2689 | } |
2690 | ||
a762a980 DM |
2691 | /* Push out any pending frames which were held back due to |
2692 | * TCP_CORK or attempt at coalescing tiny packets. | |
2693 | * The socket must be locked by the caller. | |
2694 | */ | |
9e412ba7 IJ |
2695 | void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss, |
2696 | int nonagle) | |
a762a980 | 2697 | { |
726e07a8 IJ |
2698 | /* If we are closed, the bytes will have to remain here. |
2699 | * In time closedown will finish, we empty the write queue and | |
2700 | * all will be happy. | |
2701 | */ | |
2702 | if (unlikely(sk->sk_state == TCP_CLOSE)) | |
2703 | return; | |
2704 | ||
99a1dec7 | 2705 | if (tcp_write_xmit(sk, cur_mss, nonagle, 0, |
7450aaf6 | 2706 | sk_gfp_mask(sk, GFP_ATOMIC))) |
726e07a8 | 2707 | tcp_check_probe_timer(sk); |
a762a980 DM |
2708 | } |
2709 | ||
c1b4a7e6 DM |
2710 | /* Send _single_ skb sitting at the send head. This function requires |
2711 | * true push pending frames to setup probe timer etc. | |
2712 | */ | |
2713 | void tcp_push_one(struct sock *sk, unsigned int mss_now) | |
2714 | { | |
fe067e8a | 2715 | struct sk_buff *skb = tcp_send_head(sk); |
c1b4a7e6 DM |
2716 | |
2717 | BUG_ON(!skb || skb->len < mss_now); | |
2718 | ||
d5dd9175 | 2719 | tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation); |
c1b4a7e6 DM |
2720 | } |
2721 | ||
1da177e4 LT |
2722 | /* This function returns the amount that we can raise the |
2723 | * usable window based on the following constraints | |
e905a9ed | 2724 | * |
1da177e4 LT |
2725 | * 1. The window can never be shrunk once it is offered (RFC 793) |
2726 | * 2. We limit memory per socket | |
2727 | * | |
2728 | * RFC 1122: | |
2729 | * "the suggested [SWS] avoidance algorithm for the receiver is to keep | |
2730 | * RECV.NEXT + RCV.WIN fixed until: | |
2731 | * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)" | |
2732 | * | |
2733 | * i.e. don't raise the right edge of the window until you can raise | |
2734 | * it at least MSS bytes. | |
2735 | * | |
2736 | * Unfortunately, the recommended algorithm breaks header prediction, | |
2737 | * since header prediction assumes th->window stays fixed. | |
2738 | * | |
2739 | * Strictly speaking, keeping th->window fixed violates the receiver | |
2740 | * side SWS prevention criteria. The problem is that under this rule | |
2741 | * a stream of single byte packets will cause the right side of the | |
2742 | * window to always advance by a single byte. | |
e905a9ed | 2743 | * |
1da177e4 LT |
2744 | * Of course, if the sender implements sender side SWS prevention |
2745 | * then this will not be a problem. | |
e905a9ed | 2746 | * |
1da177e4 | 2747 | * BSD seems to make the following compromise: |
e905a9ed | 2748 | * |
1da177e4 LT |
2749 | * If the free space is less than the 1/4 of the maximum |
2750 | * space available and the free space is less than 1/2 mss, | |
2751 | * then set the window to 0. | |
2752 | * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ] | |
2753 | * Otherwise, just prevent the window from shrinking | |
2754 | * and from being larger than the largest representable value. | |
2755 | * | |
2756 | * This prevents incremental opening of the window in the regime | |
2757 | * where TCP is limited by the speed of the reader side taking | |
2758 | * data out of the TCP receive queue. It does nothing about | |
2759 | * those cases where the window is constrained on the sender side | |
2760 | * because the pipeline is full. | |
2761 | * | |
2762 | * BSD also seems to "accidentally" limit itself to windows that are a | |
2763 | * multiple of MSS, at least until the free space gets quite small. | |
2764 | * This would appear to be a side effect of the mbuf implementation. | |
2765 | * Combining these two algorithms results in the observed behavior | |
2766 | * of having a fixed window size at almost all times. | |
2767 | * | |
2768 | * Below we obtain similar behavior by forcing the offered window to | |
2769 | * a multiple of the mss when it is feasible to do so. | |
2770 | * | |
2771 | * Note, we don't "adjust" for TIMESTAMP or SACK option bytes. | |
2772 | * Regular options like TIMESTAMP are taken into account. | |
2773 | */ | |
2774 | u32 __tcp_select_window(struct sock *sk) | |
2775 | { | |
463c84b9 | 2776 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 2777 | struct tcp_sock *tp = tcp_sk(sk); |
caa20d9a | 2778 | /* MSS for the peer's data. Previous versions used mss_clamp |
1da177e4 LT |
2779 | * here. I don't know if the value based on our guesses |
2780 | * of peer's MSS is better for the performance. It's more correct | |
2781 | * but may be worse for the performance because of rcv_mss | |
2782 | * fluctuations. --SAW 1998/11/1 | |
2783 | */ | |
463c84b9 | 2784 | int mss = icsk->icsk_ack.rcv_mss; |
1da177e4 | 2785 | int free_space = tcp_space(sk); |
86c1a045 | 2786 | int allowed_space = tcp_full_space(sk); |
071c8ed6 FW |
2787 | int full_space, window; |
2788 | ||
2789 | if (sk_is_mptcp(sk)) | |
2790 | mptcp_space(sk, &free_space, &allowed_space); | |
2791 | ||
2792 | full_space = min_t(int, tp->window_clamp, allowed_space); | |
1da177e4 | 2793 | |
06425c30 | 2794 | if (unlikely(mss > full_space)) { |
e905a9ed | 2795 | mss = full_space; |
06425c30 ED |
2796 | if (mss <= 0) |
2797 | return 0; | |
2798 | } | |
b92edbe0 | 2799 | if (free_space < (full_space >> 1)) { |
463c84b9 | 2800 | icsk->icsk_ack.quick = 0; |
1da177e4 | 2801 | |
b8da51eb | 2802 | if (tcp_under_memory_pressure(sk)) |
056834d9 IJ |
2803 | tp->rcv_ssthresh = min(tp->rcv_ssthresh, |
2804 | 4U * tp->advmss); | |
1da177e4 | 2805 | |
86c1a045 FW |
2806 | /* free_space might become our new window, make sure we don't |
2807 | * increase it due to wscale. | |
2808 | */ | |
2809 | free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale); | |
2810 | ||
2811 | /* if free space is less than mss estimate, or is below 1/16th | |
2812 | * of the maximum allowed, try to move to zero-window, else | |
2813 | * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and | |
2814 | * new incoming data is dropped due to memory limits. | |
2815 | * With large window, mss test triggers way too late in order | |
2816 | * to announce zero window in time before rmem limit kicks in. | |
2817 | */ | |
2818 | if (free_space < (allowed_space >> 4) || free_space < mss) | |
1da177e4 LT |
2819 | return 0; |
2820 | } | |
2821 | ||
2822 | if (free_space > tp->rcv_ssthresh) | |
2823 | free_space = tp->rcv_ssthresh; | |
2824 | ||
2825 | /* Don't do rounding if we are using window scaling, since the | |
2826 | * scaled window will not line up with the MSS boundary anyway. | |
2827 | */ | |
1da177e4 LT |
2828 | if (tp->rx_opt.rcv_wscale) { |
2829 | window = free_space; | |
2830 | ||
2831 | /* Advertise enough space so that it won't get scaled away. | |
2832 | * Import case: prevent zero window announcement if | |
2833 | * 1<<rcv_wscale > mss. | |
2834 | */ | |
1935299d | 2835 | window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale)); |
1da177e4 | 2836 | } else { |
1935299d | 2837 | window = tp->rcv_wnd; |
1da177e4 LT |
2838 | /* Get the largest window that is a nice multiple of mss. |
2839 | * Window clamp already applied above. | |
2840 | * If our current window offering is within 1 mss of the | |
2841 | * free space we just keep it. This prevents the divide | |
2842 | * and multiply from happening most of the time. | |
2843 | * We also don't do any window rounding when the free space | |
2844 | * is too small. | |
2845 | */ | |
2846 | if (window <= free_space - mss || window > free_space) | |
1935299d | 2847 | window = rounddown(free_space, mss); |
84565070 | 2848 | else if (mss == full_space && |
b92edbe0 | 2849 | free_space > window + (full_space >> 1)) |
84565070 | 2850 | window = free_space; |
1da177e4 LT |
2851 | } |
2852 | ||
2853 | return window; | |
2854 | } | |
2855 | ||
cfea5a68 MKL |
2856 | void tcp_skb_collapse_tstamp(struct sk_buff *skb, |
2857 | const struct sk_buff *next_skb) | |
082ac2d5 | 2858 | { |
0a2cf20c SHY |
2859 | if (unlikely(tcp_has_tx_tstamp(next_skb))) { |
2860 | const struct skb_shared_info *next_shinfo = | |
2861 | skb_shinfo(next_skb); | |
082ac2d5 MKL |
2862 | struct skb_shared_info *shinfo = skb_shinfo(skb); |
2863 | ||
0a2cf20c | 2864 | shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP; |
082ac2d5 | 2865 | shinfo->tskey = next_shinfo->tskey; |
2de8023e MKL |
2866 | TCP_SKB_CB(skb)->txstamp_ack |= |
2867 | TCP_SKB_CB(next_skb)->txstamp_ack; | |
082ac2d5 MKL |
2868 | } |
2869 | } | |
2870 | ||
4a17fc3a | 2871 | /* Collapses two adjacent SKB's during retransmission. */ |
f8071cde | 2872 | static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb) |
1da177e4 LT |
2873 | { |
2874 | struct tcp_sock *tp = tcp_sk(sk); | |
75c119af | 2875 | struct sk_buff *next_skb = skb_rb_next(skb); |
13dde04f | 2876 | int next_skb_size; |
1da177e4 | 2877 | |
058dc334 | 2878 | next_skb_size = next_skb->len; |
1da177e4 | 2879 | |
058dc334 | 2880 | BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1); |
a6963a6b | 2881 | |
f8071cde ED |
2882 | if (next_skb_size) { |
2883 | if (next_skb_size <= skb_availroom(skb)) | |
2884 | skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size), | |
2885 | next_skb_size); | |
3b4929f6 | 2886 | else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size)) |
f8071cde ED |
2887 | return false; |
2888 | } | |
2b7cda9c | 2889 | tcp_highest_sack_replace(sk, next_skb, skb); |
1da177e4 | 2890 | |
058dc334 IJ |
2891 | /* Update sequence range on original skb. */ |
2892 | TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq; | |
1da177e4 | 2893 | |
e6c7d085 | 2894 | /* Merge over control information. This moves PSH/FIN etc. over */ |
4de075e0 | 2895 | TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags; |
058dc334 IJ |
2896 | |
2897 | /* All done, get rid of second SKB and account for it so | |
2898 | * packet counting does not break. | |
2899 | */ | |
2900 | TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS; | |
a643b5d4 | 2901 | TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor; |
058dc334 IJ |
2902 | |
2903 | /* changed transmit queue under us so clear hints */ | |
ef9da47c IJ |
2904 | tcp_clear_retrans_hints_partial(tp); |
2905 | if (next_skb == tp->retransmit_skb_hint) | |
2906 | tp->retransmit_skb_hint = skb; | |
058dc334 | 2907 | |
797108d1 IJ |
2908 | tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb)); |
2909 | ||
082ac2d5 MKL |
2910 | tcp_skb_collapse_tstamp(skb, next_skb); |
2911 | ||
75c119af | 2912 | tcp_rtx_queue_unlink_and_free(next_skb, sk); |
f8071cde | 2913 | return true; |
1da177e4 LT |
2914 | } |
2915 | ||
67edfef7 | 2916 | /* Check if coalescing SKBs is legal. */ |
a2a385d6 | 2917 | static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb) |
4a17fc3a IJ |
2918 | { |
2919 | if (tcp_skb_pcount(skb) > 1) | |
a2a385d6 | 2920 | return false; |
4a17fc3a | 2921 | if (skb_cloned(skb)) |
a2a385d6 | 2922 | return false; |
2331ccc5 | 2923 | /* Some heuristics for collapsing over SACK'd could be invented */ |
4a17fc3a | 2924 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED) |
a2a385d6 | 2925 | return false; |
4a17fc3a | 2926 | |
a2a385d6 | 2927 | return true; |
4a17fc3a IJ |
2928 | } |
2929 | ||
67edfef7 AK |
2930 | /* Collapse packets in the retransmit queue to make to create |
2931 | * less packets on the wire. This is only done on retransmission. | |
2932 | */ | |
4a17fc3a IJ |
2933 | static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to, |
2934 | int space) | |
2935 | { | |
2936 | struct tcp_sock *tp = tcp_sk(sk); | |
2937 | struct sk_buff *skb = to, *tmp; | |
a2a385d6 | 2938 | bool first = true; |
4a17fc3a | 2939 | |
e0a1e5b5 | 2940 | if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse) |
4a17fc3a | 2941 | return; |
4de075e0 | 2942 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) |
4a17fc3a IJ |
2943 | return; |
2944 | ||
75c119af | 2945 | skb_rbtree_walk_from_safe(skb, tmp) { |
4a17fc3a IJ |
2946 | if (!tcp_can_collapse(sk, skb)) |
2947 | break; | |
2948 | ||
85712484 | 2949 | if (!tcp_skb_can_collapse(to, skb)) |
a643b5d4 MKL |
2950 | break; |
2951 | ||
4a17fc3a IJ |
2952 | space -= skb->len; |
2953 | ||
2954 | if (first) { | |
a2a385d6 | 2955 | first = false; |
4a17fc3a IJ |
2956 | continue; |
2957 | } | |
2958 | ||
2959 | if (space < 0) | |
2960 | break; | |
4a17fc3a IJ |
2961 | |
2962 | if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp))) | |
2963 | break; | |
2964 | ||
f8071cde ED |
2965 | if (!tcp_collapse_retrans(sk, to)) |
2966 | break; | |
4a17fc3a IJ |
2967 | } |
2968 | } | |
2969 | ||
1da177e4 LT |
2970 | /* This retransmits one SKB. Policy decisions and retransmit queue |
2971 | * state updates are done by the caller. Returns non-zero if an | |
2972 | * error occurred which prevented the send. | |
2973 | */ | |
10d3be56 | 2974 | int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) |
1da177e4 | 2975 | { |
5d424d5a | 2976 | struct inet_connection_sock *icsk = inet_csk(sk); |
10d3be56 | 2977 | struct tcp_sock *tp = tcp_sk(sk); |
7d227cd2 | 2978 | unsigned int cur_mss; |
10d3be56 ED |
2979 | int diff, len, err; |
2980 | ||
1da177e4 | 2981 | |
10d3be56 ED |
2982 | /* Inconclusive MTU probe */ |
2983 | if (icsk->icsk_mtup.probe_size) | |
5d424d5a | 2984 | icsk->icsk_mtup.probe_size = 0; |
5d424d5a | 2985 | |
1da177e4 | 2986 | /* Do not sent more than we queued. 1/4 is reserved for possible |
caa20d9a | 2987 | * copying overhead: fragmentation, tunneling, mangling etc. |
1da177e4 | 2988 | */ |
14afee4b | 2989 | if (refcount_read(&sk->sk_wmem_alloc) > |
ffb4d6c8 ED |
2990 | min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), |
2991 | sk->sk_sndbuf)) | |
1da177e4 LT |
2992 | return -EAGAIN; |
2993 | ||
1f3279ae ED |
2994 | if (skb_still_in_host_queue(sk, skb)) |
2995 | return -EBUSY; | |
2996 | ||
1da177e4 | 2997 | if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) { |
7f582b24 ED |
2998 | if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) { |
2999 | WARN_ON_ONCE(1); | |
3000 | return -EINVAL; | |
3001 | } | |
1da177e4 LT |
3002 | if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq)) |
3003 | return -ENOMEM; | |
3004 | } | |
3005 | ||
7d227cd2 SS |
3006 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) |
3007 | return -EHOSTUNREACH; /* Routing failure or similar. */ | |
3008 | ||
0c54b85f | 3009 | cur_mss = tcp_current_mss(sk); |
7d227cd2 | 3010 | |
1da177e4 LT |
3011 | /* If receiver has shrunk his window, and skb is out of |
3012 | * new window, do not retransmit it. The exception is the | |
3013 | * case, when window is shrunk to zero. In this case | |
3014 | * our retransmit serves as a zero window probe. | |
3015 | */ | |
9d4fb27d JP |
3016 | if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) && |
3017 | TCP_SKB_CB(skb)->seq != tp->snd_una) | |
1da177e4 LT |
3018 | return -EAGAIN; |
3019 | ||
10d3be56 ED |
3020 | len = cur_mss * segs; |
3021 | if (skb->len > len) { | |
75c119af ED |
3022 | if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len, |
3023 | cur_mss, GFP_ATOMIC)) | |
1da177e4 | 3024 | return -ENOMEM; /* We'll try again later. */ |
02276f3c | 3025 | } else { |
10d3be56 ED |
3026 | if (skb_unclone(skb, GFP_ATOMIC)) |
3027 | return -ENOMEM; | |
9eb9362e | 3028 | |
10d3be56 ED |
3029 | diff = tcp_skb_pcount(skb); |
3030 | tcp_set_skb_tso_segs(skb, cur_mss); | |
3031 | diff -= tcp_skb_pcount(skb); | |
3032 | if (diff) | |
3033 | tcp_adjust_pcount(sk, skb, diff); | |
3034 | if (skb->len < cur_mss) | |
3035 | tcp_retrans_try_collapse(sk, skb, cur_mss); | |
1da177e4 LT |
3036 | } |
3037 | ||
49213555 DB |
3038 | /* RFC3168, section 6.1.1.1. ECN fallback */ |
3039 | if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN) | |
3040 | tcp_ecn_clear_syn(sk, skb); | |
3041 | ||
678550c6 YC |
3042 | /* Update global and local TCP statistics. */ |
3043 | segs = tcp_skb_pcount(skb); | |
3044 | TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs); | |
3045 | if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN) | |
3046 | __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); | |
3047 | tp->total_retrans += segs; | |
fb31c9b9 | 3048 | tp->bytes_retrans += skb->len; |
678550c6 | 3049 | |
50bceae9 TG |
3050 | /* make sure skb->data is aligned on arches that require it |
3051 | * and check if ack-trimming & collapsing extended the headroom | |
3052 | * beyond what csum_start can cover. | |
3053 | */ | |
3054 | if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) || | |
3055 | skb_headroom(skb) >= 0xFFFF)) { | |
10a81980 ED |
3056 | struct sk_buff *nskb; |
3057 | ||
e2080072 ED |
3058 | tcp_skb_tsorted_save(skb) { |
3059 | nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC); | |
07f8e4d0 FW |
3060 | if (nskb) { |
3061 | nskb->dev = NULL; | |
3062 | err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC); | |
3063 | } else { | |
3064 | err = -ENOBUFS; | |
3065 | } | |
e2080072 ED |
3066 | } tcp_skb_tsorted_restore(skb); |
3067 | ||
5889e2c0 | 3068 | if (!err) { |
a7a25630 | 3069 | tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns); |
5889e2c0 YS |
3070 | tcp_rate_skb_sent(sk, skb); |
3071 | } | |
117632e6 | 3072 | } else { |
c84a5711 | 3073 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
117632e6 | 3074 | } |
c84a5711 | 3075 | |
7f12422c YC |
3076 | /* To avoid taking spuriously low RTT samples based on a timestamp |
3077 | * for a transmit that never happened, always mark EVER_RETRANS | |
3078 | */ | |
3079 | TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS; | |
3080 | ||
a31ad29e LB |
3081 | if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG)) |
3082 | tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB, | |
3083 | TCP_SKB_CB(skb)->seq, segs, err); | |
3084 | ||
fc9f3501 | 3085 | if (likely(!err)) { |
e086101b | 3086 | trace_tcp_retransmit_skb(sk, skb); |
678550c6 | 3087 | } else if (err != -EBUSY) { |
ec641b39 | 3088 | NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs); |
fc9f3501 | 3089 | } |
c84a5711 | 3090 | return err; |
93b174ad YC |
3091 | } |
3092 | ||
10d3be56 | 3093 | int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs) |
93b174ad YC |
3094 | { |
3095 | struct tcp_sock *tp = tcp_sk(sk); | |
10d3be56 | 3096 | int err = __tcp_retransmit_skb(sk, skb, segs); |
1da177e4 LT |
3097 | |
3098 | if (err == 0) { | |
1da177e4 | 3099 | #if FASTRETRANS_DEBUG > 0 |
056834d9 | 3100 | if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) { |
e87cc472 | 3101 | net_dbg_ratelimited("retrans_out leaked\n"); |
1da177e4 LT |
3102 | } |
3103 | #endif | |
3104 | TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS; | |
3105 | tp->retrans_out += tcp_skb_pcount(skb); | |
1da177e4 | 3106 | } |
6e08d5e3 | 3107 | |
7ae18975 YC |
3108 | /* Save stamp of the first (attempted) retransmit. */ |
3109 | if (!tp->retrans_stamp) | |
3110 | tp->retrans_stamp = tcp_skb_timestamp(skb); | |
3111 | ||
6e08d5e3 YC |
3112 | if (tp->undo_retrans < 0) |
3113 | tp->undo_retrans = 0; | |
3114 | tp->undo_retrans += tcp_skb_pcount(skb); | |
1da177e4 LT |
3115 | return err; |
3116 | } | |
3117 | ||
3118 | /* This gets called after a retransmit timeout, and the initially | |
3119 | * retransmitted data is acknowledged. It tries to continue | |
3120 | * resending the rest of the retransmit queue, until either | |
3121 | * we've sent it all or the congestion window limit is reached. | |
1da177e4 LT |
3122 | */ |
3123 | void tcp_xmit_retransmit_queue(struct sock *sk) | |
3124 | { | |
6687e988 | 3125 | const struct inet_connection_sock *icsk = inet_csk(sk); |
b9f1f1ce | 3126 | struct sk_buff *skb, *rtx_head, *hole = NULL; |
1da177e4 | 3127 | struct tcp_sock *tp = tcp_sk(sk); |
916e6d1a | 3128 | bool rearm_timer = false; |
840a3cbe | 3129 | u32 max_segs; |
61eb55f4 | 3130 | int mib_idx; |
6a438bbe | 3131 | |
45e77d31 IJ |
3132 | if (!tp->packets_out) |
3133 | return; | |
3134 | ||
b9f1f1ce ED |
3135 | rtx_head = tcp_rtx_queue_head(sk); |
3136 | skb = tp->retransmit_skb_hint ?: rtx_head; | |
ed6e7268 | 3137 | max_segs = tcp_tso_segs(sk, tcp_current_mss(sk)); |
75c119af | 3138 | skb_rbtree_walk_from(skb) { |
dca0aaf8 | 3139 | __u8 sacked; |
10d3be56 | 3140 | int segs; |
1da177e4 | 3141 | |
218af599 ED |
3142 | if (tcp_pacing_check(sk)) |
3143 | break; | |
3144 | ||
08ebd172 | 3145 | /* we could do better than to assign each time */ |
51456b29 | 3146 | if (!hole) |
0e1c54c2 | 3147 | tp->retransmit_skb_hint = skb; |
08ebd172 | 3148 | |
10d3be56 ED |
3149 | segs = tp->snd_cwnd - tcp_packets_in_flight(tp); |
3150 | if (segs <= 0) | |
916e6d1a | 3151 | break; |
dca0aaf8 | 3152 | sacked = TCP_SKB_CB(skb)->sacked; |
a3d2e9f8 ED |
3153 | /* In case tcp_shift_skb_data() have aggregated large skbs, |
3154 | * we need to make sure not sending too bigs TSO packets | |
3155 | */ | |
3156 | segs = min_t(int, segs, max_segs); | |
1da177e4 | 3157 | |
840a3cbe YC |
3158 | if (tp->retrans_out >= tp->lost_out) { |
3159 | break; | |
0e1c54c2 | 3160 | } else if (!(sacked & TCPCB_LOST)) { |
51456b29 | 3161 | if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED))) |
0e1c54c2 IJ |
3162 | hole = skb; |
3163 | continue; | |
1da177e4 | 3164 | |
0e1c54c2 IJ |
3165 | } else { |
3166 | if (icsk->icsk_ca_state != TCP_CA_Loss) | |
3167 | mib_idx = LINUX_MIB_TCPFASTRETRANS; | |
3168 | else | |
3169 | mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS; | |
3170 | } | |
1da177e4 | 3171 | |
0e1c54c2 | 3172 | if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS)) |
1da177e4 LT |
3173 | continue; |
3174 | ||
f9616c35 | 3175 | if (tcp_small_queue_check(sk, skb, 1)) |
916e6d1a | 3176 | break; |
f9616c35 | 3177 | |
10d3be56 | 3178 | if (tcp_retransmit_skb(sk, skb, segs)) |
916e6d1a | 3179 | break; |
24ab6bec | 3180 | |
de1d6578 | 3181 | NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb)); |
1da177e4 | 3182 | |
684bad11 | 3183 | if (tcp_in_cwnd_reduction(sk)) |
a262f0cd ND |
3184 | tp->prr_out += tcp_skb_pcount(skb); |
3185 | ||
75c119af | 3186 | if (skb == rtx_head && |
57dde7f7 | 3187 | icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT) |
916e6d1a ED |
3188 | rearm_timer = true; |
3189 | ||
1da177e4 | 3190 | } |
916e6d1a ED |
3191 | if (rearm_timer) |
3192 | tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS, | |
3193 | inet_csk(sk)->icsk_rto, | |
3194 | TCP_RTO_MAX); | |
1da177e4 LT |
3195 | } |
3196 | ||
d83769a5 ED |
3197 | /* We allow to exceed memory limits for FIN packets to expedite |
3198 | * connection tear down and (memory) recovery. | |
845704a5 ED |
3199 | * Otherwise tcp_send_fin() could be tempted to either delay FIN |
3200 | * or even be forced to close flow without any FIN. | |
a6c5ea4c ED |
3201 | * In general, we want to allow one skb per socket to avoid hangs |
3202 | * with edge trigger epoll() | |
d83769a5 | 3203 | */ |
a6c5ea4c | 3204 | void sk_forced_mem_schedule(struct sock *sk, int size) |
d83769a5 | 3205 | { |
e805605c | 3206 | int amt; |
d83769a5 ED |
3207 | |
3208 | if (size <= sk->sk_forward_alloc) | |
3209 | return; | |
3210 | amt = sk_mem_pages(size); | |
3211 | sk->sk_forward_alloc += amt * SK_MEM_QUANTUM; | |
e805605c JW |
3212 | sk_memory_allocated_add(sk, amt); |
3213 | ||
baac50bb JW |
3214 | if (mem_cgroup_sockets_enabled && sk->sk_memcg) |
3215 | mem_cgroup_charge_skmem(sk->sk_memcg, amt); | |
d83769a5 ED |
3216 | } |
3217 | ||
845704a5 ED |
3218 | /* Send a FIN. The caller locks the socket for us. |
3219 | * We should try to send a FIN packet really hard, but eventually give up. | |
1da177e4 LT |
3220 | */ |
3221 | void tcp_send_fin(struct sock *sk) | |
3222 | { | |
ee2aabd3 | 3223 | struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk); |
e905a9ed | 3224 | struct tcp_sock *tp = tcp_sk(sk); |
e905a9ed | 3225 | |
845704a5 ED |
3226 | /* Optimization, tack on the FIN if we have one skb in write queue and |
3227 | * this skb was not yet sent, or we are under memory pressure. | |
3228 | * Note: in the latter case, FIN packet will be sent after a timeout, | |
3229 | * as TCP stack thinks it has already been transmitted. | |
1da177e4 | 3230 | */ |
ee2aabd3 | 3231 | tskb = tail; |
75c119af ED |
3232 | if (!tskb && tcp_under_memory_pressure(sk)) |
3233 | tskb = skb_rb_last(&sk->tcp_rtx_queue); | |
3234 | ||
3235 | if (tskb) { | |
845704a5 ED |
3236 | TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN; |
3237 | TCP_SKB_CB(tskb)->end_seq++; | |
1da177e4 | 3238 | tp->write_seq++; |
ee2aabd3 | 3239 | if (!tail) { |
845704a5 ED |
3240 | /* This means tskb was already sent. |
3241 | * Pretend we included the FIN on previous transmit. | |
3242 | * We need to set tp->snd_nxt to the value it would have | |
3243 | * if FIN had been sent. This is because retransmit path | |
3244 | * does not change tp->snd_nxt. | |
3245 | */ | |
e0d694d6 | 3246 | WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1); |
845704a5 ED |
3247 | return; |
3248 | } | |
1da177e4 | 3249 | } else { |
845704a5 | 3250 | skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation); |
d1edc085 | 3251 | if (unlikely(!skb)) |
845704a5 | 3252 | return; |
d1edc085 | 3253 | |
e2080072 | 3254 | INIT_LIST_HEAD(&skb->tcp_tsorted_anchor); |
d83769a5 | 3255 | skb_reserve(skb, MAX_TCP_HEADER); |
a6c5ea4c | 3256 | sk_forced_mem_schedule(sk, skb->truesize); |
1da177e4 | 3257 | /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */ |
e870a8ef | 3258 | tcp_init_nondata_skb(skb, tp->write_seq, |
a3433f35 | 3259 | TCPHDR_ACK | TCPHDR_FIN); |
1da177e4 LT |
3260 | tcp_queue_skb(sk, skb); |
3261 | } | |
845704a5 | 3262 | __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF); |
1da177e4 LT |
3263 | } |
3264 | ||
3265 | /* We get here when a process closes a file descriptor (either due to | |
3266 | * an explicit close() or as a byproduct of exit()'ing) and there | |
3267 | * was unread data in the receive queue. This behavior is recommended | |
65bb723c | 3268 | * by RFC 2525, section 2.17. -DaveM |
1da177e4 | 3269 | */ |
dd0fc66f | 3270 | void tcp_send_active_reset(struct sock *sk, gfp_t priority) |
1da177e4 | 3271 | { |
1da177e4 LT |
3272 | struct sk_buff *skb; |
3273 | ||
7cc2b043 GF |
3274 | TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS); |
3275 | ||
1da177e4 LT |
3276 | /* NOTE: No TCP options attached and we never retransmit this. */ |
3277 | skb = alloc_skb(MAX_TCP_HEADER, priority); | |
3278 | if (!skb) { | |
4e673444 | 3279 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
1da177e4 LT |
3280 | return; |
3281 | } | |
3282 | ||
3283 | /* Reserve space for headers and prepare control bits. */ | |
3284 | skb_reserve(skb, MAX_TCP_HEADER); | |
e870a8ef | 3285 | tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk), |
a3433f35 | 3286 | TCPHDR_ACK | TCPHDR_RST); |
9a568de4 | 3287 | tcp_mstamp_refresh(tcp_sk(sk)); |
1da177e4 | 3288 | /* Send it off. */ |
dfb4b9dc | 3289 | if (tcp_transmit_skb(sk, skb, 0, priority)) |
4e673444 | 3290 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED); |
c24b14c4 SL |
3291 | |
3292 | /* skb of trace_tcp_send_reset() keeps the skb that caused RST, | |
3293 | * skb here is different to the troublesome skb, so use NULL | |
3294 | */ | |
3295 | trace_tcp_send_reset(sk, NULL); | |
1da177e4 LT |
3296 | } |
3297 | ||
67edfef7 AK |
3298 | /* Send a crossed SYN-ACK during socket establishment. |
3299 | * WARNING: This routine must only be called when we have already sent | |
1da177e4 LT |
3300 | * a SYN packet that crossed the incoming SYN that caused this routine |
3301 | * to get called. If this assumption fails then the initial rcv_wnd | |
3302 | * and rcv_wscale values will not be correct. | |
3303 | */ | |
3304 | int tcp_send_synack(struct sock *sk) | |
3305 | { | |
056834d9 | 3306 | struct sk_buff *skb; |
1da177e4 | 3307 | |
75c119af | 3308 | skb = tcp_rtx_queue_head(sk); |
51456b29 | 3309 | if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) { |
75c119af | 3310 | pr_err("%s: wrong queue state\n", __func__); |
1da177e4 LT |
3311 | return -EFAULT; |
3312 | } | |
4de075e0 | 3313 | if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) { |
1da177e4 | 3314 | if (skb_cloned(skb)) { |
e2080072 ED |
3315 | struct sk_buff *nskb; |
3316 | ||
3317 | tcp_skb_tsorted_save(skb) { | |
3318 | nskb = skb_copy(skb, GFP_ATOMIC); | |
3319 | } tcp_skb_tsorted_restore(skb); | |
51456b29 | 3320 | if (!nskb) |
1da177e4 | 3321 | return -ENOMEM; |
e2080072 | 3322 | INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor); |
2bec445f | 3323 | tcp_highest_sack_replace(sk, skb, nskb); |
75c119af | 3324 | tcp_rtx_queue_unlink_and_free(skb, sk); |
f4a775d1 | 3325 | __skb_header_release(nskb); |
75c119af | 3326 | tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb); |
ab4e846a | 3327 | sk_wmem_queued_add(sk, nskb->truesize); |
3ab224be | 3328 | sk_mem_charge(sk, nskb->truesize); |
1da177e4 LT |
3329 | skb = nskb; |
3330 | } | |
3331 | ||
4de075e0 | 3332 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK; |
735d3831 | 3333 | tcp_ecn_send_synack(sk, skb); |
1da177e4 | 3334 | } |
dfb4b9dc | 3335 | return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
1da177e4 LT |
3336 | } |
3337 | ||
4aea39c1 ED |
3338 | /** |
3339 | * tcp_make_synack - Prepare a SYN-ACK. | |
3340 | * sk: listener socket | |
3341 | * dst: dst entry attached to the SYNACK | |
3342 | * req: request_sock pointer | |
3628e3cb AL |
3343 | * foc: cookie for tcp fast open |
3344 | * synack_type: Type of synback to prepare | |
4aea39c1 ED |
3345 | * |
3346 | * Allocate one skb and build a SYNACK packet. | |
3347 | * @dst is consumed : Caller should not use it again. | |
3348 | */ | |
5d062de7 | 3349 | struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst, |
e6b4d113 | 3350 | struct request_sock *req, |
ca6fb065 | 3351 | struct tcp_fastopen_cookie *foc, |
b3d05147 | 3352 | enum tcp_synack_type synack_type) |
1da177e4 | 3353 | { |
2e6599cb | 3354 | struct inet_request_sock *ireq = inet_rsk(req); |
5d062de7 | 3355 | const struct tcp_sock *tp = tcp_sk(sk); |
80f03e27 | 3356 | struct tcp_md5sig_key *md5 = NULL; |
5d062de7 ED |
3357 | struct tcp_out_options opts; |
3358 | struct sk_buff *skb; | |
bd0388ae | 3359 | int tcp_header_size; |
5d062de7 | 3360 | struct tcphdr *th; |
f5fff5dc | 3361 | int mss; |
a842fe14 | 3362 | u64 now; |
1da177e4 | 3363 | |
ca6fb065 | 3364 | skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC); |
4aea39c1 ED |
3365 | if (unlikely(!skb)) { |
3366 | dst_release(dst); | |
1da177e4 | 3367 | return NULL; |
4aea39c1 | 3368 | } |
1da177e4 LT |
3369 | /* Reserve space for headers. */ |
3370 | skb_reserve(skb, MAX_TCP_HEADER); | |
3371 | ||
b3d05147 ED |
3372 | switch (synack_type) { |
3373 | case TCP_SYNACK_NORMAL: | |
9e17f8a4 | 3374 | skb_set_owner_w(skb, req_to_sk(req)); |
b3d05147 ED |
3375 | break; |
3376 | case TCP_SYNACK_COOKIE: | |
3377 | /* Under synflood, we do not attach skb to a socket, | |
3378 | * to avoid false sharing. | |
3379 | */ | |
3380 | break; | |
3381 | case TCP_SYNACK_FASTOPEN: | |
ca6fb065 ED |
3382 | /* sk is a const pointer, because we want to express multiple |
3383 | * cpu might call us concurrently. | |
3384 | * sk->sk_wmem_alloc in an atomic, we can promote to rw. | |
3385 | */ | |
3386 | skb_set_owner_w(skb, (struct sock *)sk); | |
b3d05147 | 3387 | break; |
ca6fb065 | 3388 | } |
4aea39c1 | 3389 | skb_dst_set(skb, dst); |
1da177e4 | 3390 | |
3541f9e8 | 3391 | mss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); |
f5fff5dc | 3392 | |
33ad798c | 3393 | memset(&opts, 0, sizeof(opts)); |
a842fe14 | 3394 | now = tcp_clock_ns(); |
8b5f12d0 | 3395 | #ifdef CONFIG_SYN_COOKIES |
f8ace8d9 | 3396 | if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok)) |
200ecef6 | 3397 | skb->skb_mstamp_ns = cookie_init_timestamp(req, now); |
8b5f12d0 FW |
3398 | else |
3399 | #endif | |
9e450c1e | 3400 | { |
a842fe14 | 3401 | skb->skb_mstamp_ns = now; |
9e450c1e YC |
3402 | if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */ |
3403 | tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb); | |
3404 | } | |
80f03e27 ED |
3405 | |
3406 | #ifdef CONFIG_TCP_MD5SIG | |
3407 | rcu_read_lock(); | |
fd3a154a | 3408 | md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req)); |
80f03e27 | 3409 | #endif |
58d607d3 | 3410 | skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4); |
60e2a778 | 3411 | tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5, |
e114e1e8 | 3412 | foc, synack_type) + sizeof(*th); |
cfb6eeb4 | 3413 | |
aa8223c7 ACM |
3414 | skb_push(skb, tcp_header_size); |
3415 | skb_reset_transport_header(skb); | |
1da177e4 | 3416 | |
ea1627c2 | 3417 | th = (struct tcphdr *)skb->data; |
1da177e4 LT |
3418 | memset(th, 0, sizeof(struct tcphdr)); |
3419 | th->syn = 1; | |
3420 | th->ack = 1; | |
6ac705b1 | 3421 | tcp_ecn_make_synack(req, th); |
b44084c2 | 3422 | th->source = htons(ireq->ir_num); |
634fb979 | 3423 | th->dest = ireq->ir_rmt_port; |
e05a90ec | 3424 | skb->mark = ireq->ir_mark; |
3b117750 ED |
3425 | skb->ip_summed = CHECKSUM_PARTIAL; |
3426 | th->seq = htonl(tcp_rsk(req)->snt_isn); | |
8336886f JC |
3427 | /* XXX data is queued and acked as is. No buffer/window check */ |
3428 | th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt); | |
1da177e4 LT |
3429 | |
3430 | /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */ | |
ed53d0ab | 3431 | th->window = htons(min(req->rsk_rcv_wnd, 65535U)); |
5d062de7 | 3432 | tcp_options_write((__be32 *)(th + 1), NULL, &opts); |
1da177e4 | 3433 | th->doff = (tcp_header_size >> 2); |
90bbcc60 | 3434 | __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS); |
cfb6eeb4 YH |
3435 | |
3436 | #ifdef CONFIG_TCP_MD5SIG | |
3437 | /* Okay, we have all we need - do the md5 hash if needed */ | |
80f03e27 | 3438 | if (md5) |
bd0388ae | 3439 | tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location, |
39f8e58e | 3440 | md5, req_to_sk(req), skb); |
80f03e27 | 3441 | rcu_read_unlock(); |
cfb6eeb4 YH |
3442 | #endif |
3443 | ||
a842fe14 ED |
3444 | skb->skb_mstamp_ns = now; |
3445 | tcp_add_tx_delay(skb, tp); | |
3446 | ||
1da177e4 LT |
3447 | return skb; |
3448 | } | |
4bc2f18b | 3449 | EXPORT_SYMBOL(tcp_make_synack); |
1da177e4 | 3450 | |
81164413 DB |
3451 | static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst) |
3452 | { | |
3453 | struct inet_connection_sock *icsk = inet_csk(sk); | |
3454 | const struct tcp_congestion_ops *ca; | |
3455 | u32 ca_key = dst_metric(dst, RTAX_CC_ALGO); | |
3456 | ||
3457 | if (ca_key == TCP_CA_UNSPEC) | |
3458 | return; | |
3459 | ||
3460 | rcu_read_lock(); | |
3461 | ca = tcp_ca_find_key(ca_key); | |
0baf26b0 MKL |
3462 | if (likely(ca && bpf_try_module_get(ca, ca->owner))) { |
3463 | bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner); | |
81164413 DB |
3464 | icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst); |
3465 | icsk->icsk_ca_ops = ca; | |
3466 | } | |
3467 | rcu_read_unlock(); | |
3468 | } | |
3469 | ||
67edfef7 | 3470 | /* Do all connect socket setups that can be done AF independent. */ |
f7e56a76 | 3471 | static void tcp_connect_init(struct sock *sk) |
1da177e4 | 3472 | { |
cf533ea5 | 3473 | const struct dst_entry *dst = __sk_dst_get(sk); |
1da177e4 LT |
3474 | struct tcp_sock *tp = tcp_sk(sk); |
3475 | __u8 rcv_wscale; | |
13d3b1eb | 3476 | u32 rcv_wnd; |
1da177e4 LT |
3477 | |
3478 | /* We'll fix this up when we get a response from the other end. | |
3479 | * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT. | |
3480 | */ | |
5d2ed052 ED |
3481 | tp->tcp_header_len = sizeof(struct tcphdr); |
3482 | if (sock_net(sk)->ipv4.sysctl_tcp_timestamps) | |
3483 | tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED; | |
1da177e4 | 3484 | |
cfb6eeb4 | 3485 | #ifdef CONFIG_TCP_MD5SIG |
00db4124 | 3486 | if (tp->af_specific->md5_lookup(sk, sk)) |
cfb6eeb4 YH |
3487 | tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED; |
3488 | #endif | |
3489 | ||
1da177e4 LT |
3490 | /* If user gave his TCP_MAXSEG, record it to clamp */ |
3491 | if (tp->rx_opt.user_mss) | |
3492 | tp->rx_opt.mss_clamp = tp->rx_opt.user_mss; | |
3493 | tp->max_window = 0; | |
5d424d5a | 3494 | tcp_mtup_init(sk); |
1da177e4 LT |
3495 | tcp_sync_mss(sk, dst_mtu(dst)); |
3496 | ||
81164413 DB |
3497 | tcp_ca_dst_init(sk, dst); |
3498 | ||
1da177e4 LT |
3499 | if (!tp->window_clamp) |
3500 | tp->window_clamp = dst_metric(dst, RTAX_WINDOW); | |
3541f9e8 | 3501 | tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst)); |
f5fff5dc | 3502 | |
1da177e4 | 3503 | tcp_initialize_rcv_mss(sk); |
1da177e4 | 3504 | |
e88c64f0 HPP |
3505 | /* limit the window selection if the user enforce a smaller rx buffer */ |
3506 | if (sk->sk_userlocks & SOCK_RCVBUF_LOCK && | |
3507 | (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0)) | |
3508 | tp->window_clamp = tcp_full_space(sk); | |
3509 | ||
13d3b1eb LB |
3510 | rcv_wnd = tcp_rwnd_init_bpf(sk); |
3511 | if (rcv_wnd == 0) | |
3512 | rcv_wnd = dst_metric(dst, RTAX_INITRWND); | |
3513 | ||
ceef9ab6 | 3514 | tcp_select_initial_window(sk, tcp_full_space(sk), |
1da177e4 LT |
3515 | tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0), |
3516 | &tp->rcv_wnd, | |
3517 | &tp->window_clamp, | |
9bb37ef0 | 3518 | sock_net(sk)->ipv4.sysctl_tcp_window_scaling, |
31d12926 | 3519 | &rcv_wscale, |
13d3b1eb | 3520 | rcv_wnd); |
1da177e4 LT |
3521 | |
3522 | tp->rx_opt.rcv_wscale = rcv_wscale; | |
3523 | tp->rcv_ssthresh = tp->rcv_wnd; | |
3524 | ||
3525 | sk->sk_err = 0; | |
3526 | sock_reset_flag(sk, SOCK_DONE); | |
3527 | tp->snd_wnd = 0; | |
ee7537b6 | 3528 | tcp_init_wl(tp, 0); |
7f582b24 | 3529 | tcp_write_queue_purge(sk); |
1da177e4 LT |
3530 | tp->snd_una = tp->write_seq; |
3531 | tp->snd_sml = tp->write_seq; | |
33f5f57e | 3532 | tp->snd_up = tp->write_seq; |
e0d694d6 | 3533 | WRITE_ONCE(tp->snd_nxt, tp->write_seq); |
ee995283 PE |
3534 | |
3535 | if (likely(!tp->repair)) | |
3536 | tp->rcv_nxt = 0; | |
c7781a6e | 3537 | else |
70eabf0e | 3538 | tp->rcv_tstamp = tcp_jiffies32; |
ee995283 | 3539 | tp->rcv_wup = tp->rcv_nxt; |
7db48e98 | 3540 | WRITE_ONCE(tp->copied_seq, tp->rcv_nxt); |
1da177e4 | 3541 | |
8550f328 | 3542 | inet_csk(sk)->icsk_rto = tcp_timeout_init(sk); |
463c84b9 | 3543 | inet_csk(sk)->icsk_retransmits = 0; |
1da177e4 LT |
3544 | tcp_clear_retrans(tp); |
3545 | } | |
3546 | ||
783237e8 YC |
3547 | static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb) |
3548 | { | |
3549 | struct tcp_sock *tp = tcp_sk(sk); | |
3550 | struct tcp_skb_cb *tcb = TCP_SKB_CB(skb); | |
3551 | ||
3552 | tcb->end_seq += skb->len; | |
f4a775d1 | 3553 | __skb_header_release(skb); |
ab4e846a | 3554 | sk_wmem_queued_add(sk, skb->truesize); |
783237e8 | 3555 | sk_mem_charge(sk, skb->truesize); |
0f317464 | 3556 | WRITE_ONCE(tp->write_seq, tcb->end_seq); |
783237e8 YC |
3557 | tp->packets_out += tcp_skb_pcount(skb); |
3558 | } | |
3559 | ||
3560 | /* Build and send a SYN with data and (cached) Fast Open cookie. However, | |
3561 | * queue a data-only packet after the regular SYN, such that regular SYNs | |
3562 | * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges | |
3563 | * only the SYN sequence, the data are retransmitted in the first ACK. | |
3564 | * If cookie is not cached or other error occurs, falls back to send a | |
3565 | * regular SYN with Fast Open cookie request option. | |
3566 | */ | |
3567 | static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn) | |
3568 | { | |
3569 | struct tcp_sock *tp = tcp_sk(sk); | |
3570 | struct tcp_fastopen_request *fo = tp->fastopen_req; | |
065263f4 | 3571 | int space, err = 0; |
355a901e | 3572 | struct sk_buff *syn_data; |
aab48743 | 3573 | |
67da22d2 | 3574 | tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */ |
065263f4 | 3575 | if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie)) |
783237e8 YC |
3576 | goto fallback; |
3577 | ||
3578 | /* MSS for SYN-data is based on cached MSS and bounded by PMTU and | |
3579 | * user-MSS. Reserve maximum option space for middleboxes that add | |
3580 | * private TCP options. The cost is reduced data space in SYN :( | |
3581 | */ | |
3541f9e8 ED |
3582 | tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp); |
3583 | ||
1b63edd6 | 3584 | space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) - |
783237e8 YC |
3585 | MAX_TCP_OPTION_SPACE; |
3586 | ||
f5ddcbbb ED |
3587 | space = min_t(size_t, space, fo->size); |
3588 | ||
3589 | /* limit to order-0 allocations */ | |
3590 | space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER)); | |
3591 | ||
eb934478 | 3592 | syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false); |
355a901e | 3593 | if (!syn_data) |
783237e8 | 3594 | goto fallback; |
355a901e ED |
3595 | syn_data->ip_summed = CHECKSUM_PARTIAL; |
3596 | memcpy(syn_data->cb, syn->cb, sizeof(syn->cb)); | |
07e100f9 ED |
3597 | if (space) { |
3598 | int copied = copy_from_iter(skb_put(syn_data, space), space, | |
3599 | &fo->data->msg_iter); | |
3600 | if (unlikely(!copied)) { | |
ba233b34 | 3601 | tcp_skb_tsorted_anchor_cleanup(syn_data); |
07e100f9 ED |
3602 | kfree_skb(syn_data); |
3603 | goto fallback; | |
3604 | } | |
3605 | if (copied != space) { | |
3606 | skb_trim(syn_data, copied); | |
3607 | space = copied; | |
3608 | } | |
f859a448 | 3609 | skb_zcopy_set(syn_data, fo->uarg, NULL); |
57be5bda | 3610 | } |
355a901e ED |
3611 | /* No more data pending in inet_wait_for_connect() */ |
3612 | if (space == fo->size) | |
3613 | fo->data = NULL; | |
3614 | fo->copied = space; | |
783237e8 | 3615 | |
355a901e | 3616 | tcp_connect_queue_skb(sk, syn_data); |
0f87230d FY |
3617 | if (syn_data->len) |
3618 | tcp_chrono_start(sk, TCP_CHRONO_BUSY); | |
783237e8 | 3619 | |
355a901e | 3620 | err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation); |
783237e8 | 3621 | |
d3edd06e | 3622 | syn->skb_mstamp_ns = syn_data->skb_mstamp_ns; |
431a9124 | 3623 | |
355a901e ED |
3624 | /* Now full SYN+DATA was cloned and sent (or not), |
3625 | * remove the SYN from the original skb (syn_data) | |
3626 | * we keep in write queue in case of a retransmit, as we | |
3627 | * also have the SYN packet (with no data) in the same queue. | |
3628 | */ | |
3629 | TCP_SKB_CB(syn_data)->seq++; | |
3630 | TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH; | |
3631 | if (!err) { | |
67da22d2 | 3632 | tp->syn_data = (fo->copied > 0); |
75c119af | 3633 | tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data); |
f19c29e3 | 3634 | NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT); |
783237e8 YC |
3635 | goto done; |
3636 | } | |
783237e8 | 3637 | |
75c119af ED |
3638 | /* data was not sent, put it in write_queue */ |
3639 | __skb_queue_tail(&sk->sk_write_queue, syn_data); | |
b5b7db8d ED |
3640 | tp->packets_out -= tcp_skb_pcount(syn_data); |
3641 | ||
783237e8 YC |
3642 | fallback: |
3643 | /* Send a regular SYN with Fast Open cookie request option */ | |
3644 | if (fo->cookie.len > 0) | |
3645 | fo->cookie.len = 0; | |
3646 | err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation); | |
3647 | if (err) | |
3648 | tp->syn_fastopen = 0; | |
783237e8 YC |
3649 | done: |
3650 | fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */ | |
3651 | return err; | |
3652 | } | |
3653 | ||
67edfef7 | 3654 | /* Build a SYN and send it off. */ |
1da177e4 LT |
3655 | int tcp_connect(struct sock *sk) |
3656 | { | |
3657 | struct tcp_sock *tp = tcp_sk(sk); | |
3658 | struct sk_buff *buff; | |
ee586811 | 3659 | int err; |
1da177e4 | 3660 | |
de525be2 | 3661 | tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL); |
8ba60924 ED |
3662 | |
3663 | if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) | |
3664 | return -EHOSTUNREACH; /* Routing failure or similar. */ | |
3665 | ||
1da177e4 LT |
3666 | tcp_connect_init(sk); |
3667 | ||
2b916477 AV |
3668 | if (unlikely(tp->repair)) { |
3669 | tcp_finish_connect(sk, NULL); | |
3670 | return 0; | |
3671 | } | |
3672 | ||
eb934478 | 3673 | buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true); |
355a901e | 3674 | if (unlikely(!buff)) |
1da177e4 LT |
3675 | return -ENOBUFS; |
3676 | ||
a3433f35 | 3677 | tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN); |
9a568de4 ED |
3678 | tcp_mstamp_refresh(tp); |
3679 | tp->retrans_stamp = tcp_time_stamp(tp); | |
783237e8 | 3680 | tcp_connect_queue_skb(sk, buff); |
735d3831 | 3681 | tcp_ecn_send_syn(sk, buff); |
75c119af | 3682 | tcp_rbtree_insert(&sk->tcp_rtx_queue, buff); |
1da177e4 | 3683 | |
783237e8 YC |
3684 | /* Send off SYN; include data in Fast Open. */ |
3685 | err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) : | |
3686 | tcp_transmit_skb(sk, buff, 1, sk->sk_allocation); | |
ee586811 EP |
3687 | if (err == -ECONNREFUSED) |
3688 | return err; | |
bd37a088 WY |
3689 | |
3690 | /* We change tp->snd_nxt after the tcp_transmit_skb() call | |
3691 | * in order to make this packet get counted in tcpOutSegs. | |
3692 | */ | |
e0d694d6 | 3693 | WRITE_ONCE(tp->snd_nxt, tp->write_seq); |
bd37a088 | 3694 | tp->pushed_seq = tp->write_seq; |
b5b7db8d ED |
3695 | buff = tcp_send_head(sk); |
3696 | if (unlikely(buff)) { | |
e0d694d6 | 3697 | WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq); |
b5b7db8d ED |
3698 | tp->pushed_seq = TCP_SKB_CB(buff)->seq; |
3699 | } | |
81cc8a75 | 3700 | TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS); |
1da177e4 LT |
3701 | |
3702 | /* Timer for repeating the SYN until an answer. */ | |
3f421baa ACM |
3703 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, |
3704 | inet_csk(sk)->icsk_rto, TCP_RTO_MAX); | |
1da177e4 LT |
3705 | return 0; |
3706 | } | |
4bc2f18b | 3707 | EXPORT_SYMBOL(tcp_connect); |
1da177e4 LT |
3708 | |
3709 | /* Send out a delayed ack, the caller does the policy checking | |
3710 | * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check() | |
3711 | * for details. | |
3712 | */ | |
3713 | void tcp_send_delayed_ack(struct sock *sk) | |
3714 | { | |
463c84b9 ACM |
3715 | struct inet_connection_sock *icsk = inet_csk(sk); |
3716 | int ato = icsk->icsk_ack.ato; | |
1da177e4 LT |
3717 | unsigned long timeout; |
3718 | ||
3719 | if (ato > TCP_DELACK_MIN) { | |
463c84b9 | 3720 | const struct tcp_sock *tp = tcp_sk(sk); |
056834d9 | 3721 | int max_ato = HZ / 2; |
1da177e4 | 3722 | |
31954cd8 | 3723 | if (inet_csk_in_pingpong_mode(sk) || |
056834d9 | 3724 | (icsk->icsk_ack.pending & ICSK_ACK_PUSHED)) |
1da177e4 LT |
3725 | max_ato = TCP_DELACK_MAX; |
3726 | ||
3727 | /* Slow path, intersegment interval is "high". */ | |
3728 | ||
3729 | /* If some rtt estimate is known, use it to bound delayed ack. | |
463c84b9 | 3730 | * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements |
1da177e4 LT |
3731 | * directly. |
3732 | */ | |
740b0f18 ED |
3733 | if (tp->srtt_us) { |
3734 | int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3), | |
3735 | TCP_DELACK_MIN); | |
1da177e4 LT |
3736 | |
3737 | if (rtt < max_ato) | |
3738 | max_ato = rtt; | |
3739 | } | |
3740 | ||
3741 | ato = min(ato, max_ato); | |
3742 | } | |
3743 | ||
2b8ee4f0 MKL |
3744 | ato = min_t(u32, ato, inet_csk(sk)->icsk_delack_max); |
3745 | ||
1da177e4 LT |
3746 | /* Stay within the limit we were given */ |
3747 | timeout = jiffies + ato; | |
3748 | ||
3749 | /* Use new timeout only if there wasn't a older one earlier. */ | |
463c84b9 | 3750 | if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) { |
1da177e4 LT |
3751 | /* If delack timer was blocked or is about to expire, |
3752 | * send ACK now. | |
3753 | */ | |
463c84b9 ACM |
3754 | if (icsk->icsk_ack.blocked || |
3755 | time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) { | |
1da177e4 LT |
3756 | tcp_send_ack(sk); |
3757 | return; | |
3758 | } | |
3759 | ||
463c84b9 ACM |
3760 | if (!time_before(timeout, icsk->icsk_ack.timeout)) |
3761 | timeout = icsk->icsk_ack.timeout; | |
1da177e4 | 3762 | } |
463c84b9 ACM |
3763 | icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER; |
3764 | icsk->icsk_ack.timeout = timeout; | |
3765 | sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout); | |
1da177e4 LT |
3766 | } |
3767 | ||
3768 | /* This routine sends an ack and also updates the window. */ | |
2987babb | 3769 | void __tcp_send_ack(struct sock *sk, u32 rcv_nxt) |
1da177e4 | 3770 | { |
058dc334 | 3771 | struct sk_buff *buff; |
1da177e4 | 3772 | |
058dc334 IJ |
3773 | /* If we have been reset, we may not send again. */ |
3774 | if (sk->sk_state == TCP_CLOSE) | |
3775 | return; | |
1da177e4 | 3776 | |
058dc334 IJ |
3777 | /* We are not putting this on the write queue, so |
3778 | * tcp_transmit_skb() will set the ownership to this | |
3779 | * sock. | |
3780 | */ | |
7450aaf6 ED |
3781 | buff = alloc_skb(MAX_TCP_HEADER, |
3782 | sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); | |
3783 | if (unlikely(!buff)) { | |
058dc334 IJ |
3784 | inet_csk_schedule_ack(sk); |
3785 | inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN; | |
3786 | inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, | |
3787 | TCP_DELACK_MAX, TCP_RTO_MAX); | |
3788 | return; | |
1da177e4 | 3789 | } |
058dc334 IJ |
3790 | |
3791 | /* Reserve space for headers and prepare control bits. */ | |
3792 | skb_reserve(buff, MAX_TCP_HEADER); | |
a3433f35 | 3793 | tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK); |
058dc334 | 3794 | |
98781965 ED |
3795 | /* We do not want pure acks influencing TCP Small Queues or fq/pacing |
3796 | * too much. | |
3797 | * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784 | |
98781965 ED |
3798 | */ |
3799 | skb_set_tcp_pure_ack(buff); | |
3800 | ||
058dc334 | 3801 | /* Send it off, this clears delayed acks for us. */ |
2987babb YC |
3802 | __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt); |
3803 | } | |
27cde44a | 3804 | EXPORT_SYMBOL_GPL(__tcp_send_ack); |
2987babb YC |
3805 | |
3806 | void tcp_send_ack(struct sock *sk) | |
3807 | { | |
3808 | __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt); | |
1da177e4 LT |
3809 | } |
3810 | ||
3811 | /* This routine sends a packet with an out of date sequence | |
3812 | * number. It assumes the other end will try to ack it. | |
3813 | * | |
3814 | * Question: what should we make while urgent mode? | |
3815 | * 4.4BSD forces sending single byte of data. We cannot send | |
3816 | * out of window data, because we have SND.NXT==SND.MAX... | |
3817 | * | |
3818 | * Current solution: to send TWO zero-length segments in urgent mode: | |
3819 | * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is | |
3820 | * out-of-date with SND.UNA-1 to probe window. | |
3821 | */ | |
e520af48 | 3822 | static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib) |
1da177e4 LT |
3823 | { |
3824 | struct tcp_sock *tp = tcp_sk(sk); | |
3825 | struct sk_buff *skb; | |
3826 | ||
3827 | /* We don't queue it, tcp_transmit_skb() sets ownership. */ | |
7450aaf6 ED |
3828 | skb = alloc_skb(MAX_TCP_HEADER, |
3829 | sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN)); | |
51456b29 | 3830 | if (!skb) |
1da177e4 LT |
3831 | return -1; |
3832 | ||
3833 | /* Reserve space for headers and set control bits. */ | |
3834 | skb_reserve(skb, MAX_TCP_HEADER); | |
1da177e4 LT |
3835 | /* Use a previous sequence. This should cause the other |
3836 | * end to send an ack. Don't queue or clone SKB, just | |
3837 | * send it. | |
3838 | */ | |
a3433f35 | 3839 | tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK); |
e2e8009f | 3840 | NET_INC_STATS(sock_net(sk), mib); |
7450aaf6 | 3841 | return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0); |
1da177e4 LT |
3842 | } |
3843 | ||
385e2070 | 3844 | /* Called from setsockopt( ... TCP_REPAIR ) */ |
ee995283 PE |
3845 | void tcp_send_window_probe(struct sock *sk) |
3846 | { | |
3847 | if (sk->sk_state == TCP_ESTABLISHED) { | |
3848 | tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1; | |
9a568de4 | 3849 | tcp_mstamp_refresh(tcp_sk(sk)); |
e520af48 | 3850 | tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE); |
ee995283 PE |
3851 | } |
3852 | } | |
3853 | ||
67edfef7 | 3854 | /* Initiate keepalive or window probe from timer. */ |
e520af48 | 3855 | int tcp_write_wakeup(struct sock *sk, int mib) |
1da177e4 | 3856 | { |
058dc334 IJ |
3857 | struct tcp_sock *tp = tcp_sk(sk); |
3858 | struct sk_buff *skb; | |
1da177e4 | 3859 | |
058dc334 IJ |
3860 | if (sk->sk_state == TCP_CLOSE) |
3861 | return -1; | |
3862 | ||
00db4124 IM |
3863 | skb = tcp_send_head(sk); |
3864 | if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) { | |
058dc334 | 3865 | int err; |
0c54b85f | 3866 | unsigned int mss = tcp_current_mss(sk); |
058dc334 IJ |
3867 | unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq; |
3868 | ||
3869 | if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq)) | |
3870 | tp->pushed_seq = TCP_SKB_CB(skb)->end_seq; | |
3871 | ||
3872 | /* We are probing the opening of a window | |
3873 | * but the window size is != 0 | |
3874 | * must have been a result SWS avoidance ( sender ) | |
3875 | */ | |
3876 | if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq || | |
3877 | skb->len > mss) { | |
3878 | seg_size = min(seg_size, mss); | |
4de075e0 | 3879 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
75c119af ED |
3880 | if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE, |
3881 | skb, seg_size, mss, GFP_ATOMIC)) | |
058dc334 IJ |
3882 | return -1; |
3883 | } else if (!tcp_skb_pcount(skb)) | |
5bbb432c | 3884 | tcp_set_skb_tso_segs(skb, mss); |
058dc334 | 3885 | |
4de075e0 | 3886 | TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH; |
058dc334 IJ |
3887 | err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC); |
3888 | if (!err) | |
3889 | tcp_event_new_data_sent(sk, skb); | |
3890 | return err; | |
3891 | } else { | |
33f5f57e | 3892 | if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF)) |
e520af48 ED |
3893 | tcp_xmit_probe_skb(sk, 1, mib); |
3894 | return tcp_xmit_probe_skb(sk, 0, mib); | |
1da177e4 | 3895 | } |
1da177e4 LT |
3896 | } |
3897 | ||
3898 | /* A window probe timeout has occurred. If window is not closed send | |
3899 | * a partial packet else a zero probe. | |
3900 | */ | |
3901 | void tcp_send_probe0(struct sock *sk) | |
3902 | { | |
463c84b9 | 3903 | struct inet_connection_sock *icsk = inet_csk(sk); |
1da177e4 | 3904 | struct tcp_sock *tp = tcp_sk(sk); |
c6214a97 | 3905 | struct net *net = sock_net(sk); |
c1d5674f | 3906 | unsigned long timeout; |
1da177e4 LT |
3907 | int err; |
3908 | ||
e520af48 | 3909 | err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE); |
1da177e4 | 3910 | |
75c119af | 3911 | if (tp->packets_out || tcp_write_queue_empty(sk)) { |
1da177e4 | 3912 | /* Cancel probe timer, if it is not required. */ |
6687e988 | 3913 | icsk->icsk_probes_out = 0; |
463c84b9 | 3914 | icsk->icsk_backoff = 0; |
1da177e4 LT |
3915 | return; |
3916 | } | |
3917 | ||
c1d5674f | 3918 | icsk->icsk_probes_out++; |
1da177e4 | 3919 | if (err <= 0) { |
c6214a97 | 3920 | if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2) |
463c84b9 | 3921 | icsk->icsk_backoff++; |
c1d5674f | 3922 | timeout = tcp_probe0_when(sk, TCP_RTO_MAX); |
1da177e4 LT |
3923 | } else { |
3924 | /* If packet was not sent due to local congestion, | |
c1d5674f | 3925 | * Let senders fight for local resources conservatively. |
1da177e4 | 3926 | */ |
c1d5674f YC |
3927 | timeout = TCP_RESOURCE_PROBE_INTERVAL; |
3928 | } | |
8dc242ad | 3929 | tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX); |
1da177e4 | 3930 | } |
5db92c99 | 3931 | |
ea3bea3a | 3932 | int tcp_rtx_synack(const struct sock *sk, struct request_sock *req) |
5db92c99 OP |
3933 | { |
3934 | const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific; | |
3935 | struct flowi fl; | |
3936 | int res; | |
3937 | ||
58d607d3 | 3938 | tcp_rsk(req)->txhash = net_tx_rndhash(); |
b3d05147 | 3939 | res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL); |
5db92c99 | 3940 | if (!res) { |
90bbcc60 | 3941 | __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS); |
02a1d6e7 | 3942 | __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS); |
7e32b443 YC |
3943 | if (unlikely(tcp_passive_fastopen(sk))) |
3944 | tcp_sk(sk)->total_retrans++; | |
cf34ce3d | 3945 | trace_tcp_retransmit_synack(sk, req); |
5db92c99 OP |
3946 | } |
3947 | return res; | |
3948 | } | |
3949 | EXPORT_SYMBOL(tcp_rtx_synack); |