]> git.proxmox.com Git - mirror_ubuntu-hirsute-kernel.git/blame - net/ipv4/tcp_output.c
hns3: Another build fix.
[mirror_ubuntu-hirsute-kernel.git] / net / ipv4 / tcp_output.c
CommitLineData
1da177e4
LT
1/*
2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
02c30a84 8 * Authors: Ross Biro
1da177e4
LT
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
19 */
20
21/*
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
25 * : AF independence
26 *
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
34 *
35 */
36
91df42be
JP
37#define pr_fmt(fmt) "TCP: " fmt
38
1da177e4
LT
39#include <net/tcp.h>
40
41#include <linux/compiler.h>
5a0e3ad6 42#include <linux/gfp.h>
1da177e4 43#include <linux/module.h>
60e2a778 44#include <linux/static_key.h>
1da177e4 45
e086101b 46#include <trace/events/tcp.h>
35089bb2 47
9799ccb0
ED
48/* Refresh clocks of a TCP socket,
49 * ensuring monotically increasing values.
50 */
51void tcp_mstamp_refresh(struct tcp_sock *tp)
52{
53 u64 val = tcp_clock_ns();
54
55 /* departure time for next data packet */
56 if (val > tp->tcp_wstamp_ns)
57 tp->tcp_wstamp_ns = val;
58
59 val = div_u64(val, NSEC_PER_USEC);
60 if (val > tp->tcp_mstamp)
61 tp->tcp_mstamp = val;
62}
63
46d3ceab
ED
64static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
65 int push_one, gfp_t gfp);
519855c5 66
67edfef7 67/* Account for new data that has been sent to the network. */
75c119af 68static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
1da177e4 69{
6ba8a3b1 70 struct inet_connection_sock *icsk = inet_csk(sk);
9e412ba7 71 struct tcp_sock *tp = tcp_sk(sk);
66f5fe62 72 unsigned int prior_packets = tp->packets_out;
9e412ba7 73
1da177e4 74 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
8512430e 75
75c119af
ED
76 __skb_unlink(skb, &sk->sk_write_queue);
77 tcp_rbtree_insert(&sk->tcp_rtx_queue, 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 94static 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 */
120static __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 */
141void 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 160static 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
1da177e4
LT
169 tp->lsndtime = now;
170
171 /* If it is a reply for ato after last received
172 * packet, enter pingpong mode.
173 */
2251ae46
JM
174 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
175 icsk->icsk_ack.pingpong = 1;
1da177e4
LT
176}
177
67edfef7 178/* Account for an ACK we sent. */
27cde44a
YC
179static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
180 u32 rcv_nxt)
1da177e4 181{
5d9f4262
ED
182 struct tcp_sock *tp = tcp_sk(sk);
183
184 if (unlikely(tp->compressed_ack)) {
200d95f4
ED
185 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
186 tp->compressed_ack);
5d9f4262
ED
187 tp->compressed_ack = 0;
188 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
189 __sock_put(sk);
190 }
27cde44a
YC
191
192 if (unlikely(rcv_nxt != tp->rcv_nxt))
193 return; /* Special ACK sent by DCTCP to reflect ECN */
463c84b9
ACM
194 tcp_dec_quickack_mode(sk, pkts);
195 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
1da177e4
LT
196}
197
85f16525
YC
198
199u32 tcp_default_init_rwnd(u32 mss)
200{
201 /* Initial receive window should be twice of TCP_INIT_CWND to
9ef71e0c 202 * enable proper sending of new unsent data during fast recovery
85f16525
YC
203 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
204 * limit when mss is larger than 1460.
205 */
206 u32 init_rwnd = TCP_INIT_CWND * 2;
207
208 if (mss > 1460)
209 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
210 return init_rwnd;
211}
212
1da177e4
LT
213/* Determine a window scaling and initial window to offer.
214 * Based on the assumption that the given amount of space
215 * will be offered. Store the results in the tp structure.
216 * NOTE: for smooth operation initial space offering should
217 * be a multiple of mss if possible. We assume here that mss >= 1.
218 * This MUST be enforced by all callers.
219 */
ceef9ab6 220void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
1da177e4 221 __u32 *rcv_wnd, __u32 *window_clamp,
31d12926 222 int wscale_ok, __u8 *rcv_wscale,
223 __u32 init_rcv_wnd)
1da177e4
LT
224{
225 unsigned int space = (__space < 0 ? 0 : __space);
226
227 /* If no clamp set the clamp to the max possible scaled window */
228 if (*window_clamp == 0)
589c49cb 229 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
1da177e4
LT
230 space = min(*window_clamp, space);
231
232 /* Quantize space offering to a multiple of mss if possible. */
233 if (space > mss)
589c49cb 234 space = rounddown(space, mss);
1da177e4
LT
235
236 /* NOTE: offering an initial window larger than 32767
15d99e02
RJ
237 * will break some buggy TCP stacks. If the admin tells us
238 * it is likely we could be speaking with such a buggy stack
239 * we will truncate our initial window offering to 32K-1
240 * unless the remote has sent us a window scaling option,
241 * which we interpret as a sign the remote TCP is not
242 * misinterpreting the window field as a signed quantity.
1da177e4 243 */
ceef9ab6 244 if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
15d99e02
RJ
245 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
246 else
247 (*rcv_wnd) = space;
248
1da177e4
LT
249 (*rcv_wscale) = 0;
250 if (wscale_ok) {
589c49cb 251 /* Set window scaling on max possible window */
356d1833 252 space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
f626300a 253 space = max_t(u32, space, sysctl_rmem_max);
316c1592 254 space = min_t(u32, space, *window_clamp);
589c49cb 255 while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) {
1da177e4
LT
256 space >>= 1;
257 (*rcv_wscale)++;
258 }
259 }
260
c36207bd
WW
261 if (!init_rcv_wnd) /* Use default unless specified otherwise */
262 init_rcv_wnd = tcp_default_init_rwnd(mss);
263 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
1da177e4
LT
264
265 /* Set the clamp no higher than max representable value */
589c49cb 266 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
1da177e4 267}
4bc2f18b 268EXPORT_SYMBOL(tcp_select_initial_window);
1da177e4
LT
269
270/* Chose a new window to advertise, update state in tcp_sock for the
271 * socket, and return result with RFC1323 scaling applied. The return
272 * value can be stuffed directly into th->window for an outgoing
273 * frame.
274 */
40efc6fa 275static u16 tcp_select_window(struct sock *sk)
1da177e4
LT
276{
277 struct tcp_sock *tp = tcp_sk(sk);
8e165e20 278 u32 old_win = tp->rcv_wnd;
1da177e4
LT
279 u32 cur_win = tcp_receive_window(tp);
280 u32 new_win = __tcp_select_window(sk);
281
282 /* Never shrink the offered window */
2de979bd 283 if (new_win < cur_win) {
1da177e4
LT
284 /* Danger Will Robinson!
285 * Don't update rcv_wup/rcv_wnd here or else
286 * we will not be able to advertise a zero
287 * window in time. --DaveM
288 *
289 * Relax Will Robinson.
290 */
8e165e20
FW
291 if (new_win == 0)
292 NET_INC_STATS(sock_net(sk),
293 LINUX_MIB_TCPWANTZEROWINDOWADV);
607bfbf2 294 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
1da177e4
LT
295 }
296 tp->rcv_wnd = new_win;
297 tp->rcv_wup = tp->rcv_nxt;
298
299 /* Make sure we do not exceed the maximum possible
300 * scaled window.
301 */
ceef9ab6
ED
302 if (!tp->rx_opt.rcv_wscale &&
303 sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
1da177e4
LT
304 new_win = min(new_win, MAX_TCP_WINDOW);
305 else
306 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
307
308 /* RFC1323 scaling applied */
309 new_win >>= tp->rx_opt.rcv_wscale;
310
31770e34 311 /* If we advertise zero window, disable fast path. */
8e165e20 312 if (new_win == 0) {
31770e34 313 tp->pred_flags = 0;
8e165e20
FW
314 if (old_win)
315 NET_INC_STATS(sock_net(sk),
316 LINUX_MIB_TCPTOZEROWINDOWADV);
317 } else if (old_win == 0) {
318 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
319 }
1da177e4
LT
320
321 return new_win;
322}
323
67edfef7 324/* Packet ECN state for a SYN-ACK */
735d3831 325static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
bdf1ee5d 326{
30e502a3
DB
327 const struct tcp_sock *tp = tcp_sk(sk);
328
4de075e0 329 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
056834d9 330 if (!(tp->ecn_flags & TCP_ECN_OK))
4de075e0 331 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
91b5b21c
LB
332 else if (tcp_ca_needs_ecn(sk) ||
333 tcp_bpf_ca_needs_ecn(sk))
30e502a3 334 INET_ECN_xmit(sk);
bdf1ee5d
IJ
335}
336
67edfef7 337/* Packet ECN state for a SYN. */
735d3831 338static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
bdf1ee5d
IJ
339{
340 struct tcp_sock *tp = tcp_sk(sk);
91b5b21c 341 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
f7b3bec6 342 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
91b5b21c 343 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
f7b3bec6
FW
344
345 if (!use_ecn) {
346 const struct dst_entry *dst = __sk_dst_get(sk);
347
348 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
349 use_ecn = true;
350 }
bdf1ee5d
IJ
351
352 tp->ecn_flags = 0;
f7b3bec6
FW
353
354 if (use_ecn) {
4de075e0 355 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
bdf1ee5d 356 tp->ecn_flags = TCP_ECN_OK;
91b5b21c 357 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
30e502a3 358 INET_ECN_xmit(sk);
bdf1ee5d
IJ
359 }
360}
361
49213555
DB
362static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
363{
364 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
365 /* tp->ecn_flags are cleared at a later point in time when
366 * SYN ACK is ultimatively being received.
367 */
368 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
369}
370
735d3831 371static void
6ac705b1 372tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
bdf1ee5d 373{
6ac705b1 374 if (inet_rsk(req)->ecn_ok)
bdf1ee5d
IJ
375 th->ece = 1;
376}
377
67edfef7
AK
378/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
379 * be sent.
380 */
735d3831 381static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
ea1627c2 382 struct tcphdr *th, int tcp_header_len)
bdf1ee5d
IJ
383{
384 struct tcp_sock *tp = tcp_sk(sk);
385
386 if (tp->ecn_flags & TCP_ECN_OK) {
387 /* Not-retransmitted data segment: set ECT and inject CWR. */
388 if (skb->len != tcp_header_len &&
389 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
390 INET_ECN_xmit(sk);
056834d9 391 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
bdf1ee5d 392 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
ea1627c2 393 th->cwr = 1;
bdf1ee5d
IJ
394 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
395 }
30e502a3 396 } else if (!tcp_ca_needs_ecn(sk)) {
bdf1ee5d
IJ
397 /* ACK or retransmitted segment: clear ECT|CE */
398 INET_ECN_dontxmit(sk);
399 }
400 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
ea1627c2 401 th->ece = 1;
bdf1ee5d
IJ
402 }
403}
404
e870a8ef
IJ
405/* Constructs common control bits of non-data skb. If SYN/FIN is present,
406 * auto increment end seqno.
407 */
408static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
409{
2e8e18ef 410 skb->ip_summed = CHECKSUM_PARTIAL;
e870a8ef 411
4de075e0 412 TCP_SKB_CB(skb)->tcp_flags = flags;
e870a8ef
IJ
413 TCP_SKB_CB(skb)->sacked = 0;
414
cd7d8498 415 tcp_skb_pcount_set(skb, 1);
e870a8ef
IJ
416
417 TCP_SKB_CB(skb)->seq = seq;
a3433f35 418 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
e870a8ef
IJ
419 seq++;
420 TCP_SKB_CB(skb)->end_seq = seq;
421}
422
a2a385d6 423static inline bool tcp_urg_mode(const struct tcp_sock *tp)
33f5f57e
IJ
424{
425 return tp->snd_una != tp->snd_up;
426}
427
33ad798c
AL
428#define OPTION_SACK_ADVERTISE (1 << 0)
429#define OPTION_TS (1 << 1)
430#define OPTION_MD5 (1 << 2)
89e95a61 431#define OPTION_WSCALE (1 << 3)
2100c8d2 432#define OPTION_FAST_OPEN_COOKIE (1 << 8)
60e2a778
UB
433#define OPTION_SMC (1 << 9)
434
435static void smc_options_write(__be32 *ptr, u16 *options)
436{
437#if IS_ENABLED(CONFIG_SMC)
438 if (static_branch_unlikely(&tcp_have_smc)) {
439 if (unlikely(OPTION_SMC & *options)) {
440 *ptr++ = htonl((TCPOPT_NOP << 24) |
441 (TCPOPT_NOP << 16) |
442 (TCPOPT_EXP << 8) |
443 (TCPOLEN_EXP_SMC_BASE));
444 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
445 }
446 }
447#endif
448}
33ad798c
AL
449
450struct tcp_out_options {
2100c8d2
YC
451 u16 options; /* bit field of OPTION_* */
452 u16 mss; /* 0 to disable */
33ad798c
AL
453 u8 ws; /* window scale, 0 to disable */
454 u8 num_sack_blocks; /* number of SACK blocks to include */
bd0388ae 455 u8 hash_size; /* bytes in hash_location */
bd0388ae 456 __u8 *hash_location; /* temporary pointer, overloaded */
2100c8d2
YC
457 __u32 tsval, tsecr; /* need to include OPTION_TS */
458 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
33ad798c
AL
459};
460
67edfef7
AK
461/* Write previously computed TCP options to the packet.
462 *
463 * Beware: Something in the Internet is very sensitive to the ordering of
fd6149d3
IJ
464 * TCP options, we learned this through the hard way, so be careful here.
465 * Luckily we can at least blame others for their non-compliance but from
8e3bff96 466 * inter-operability perspective it seems that we're somewhat stuck with
fd6149d3
IJ
467 * the ordering which we have been using if we want to keep working with
468 * those broken things (not that it currently hurts anybody as there isn't
469 * particular reason why the ordering would need to be changed).
470 *
471 * At least SACK_PERM as the first option is known to lead to a disaster
472 * (but it may well be that other scenarios fail similarly).
473 */
33ad798c 474static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
bd0388ae
WAS
475 struct tcp_out_options *opts)
476{
2100c8d2 477 u16 options = opts->options; /* mungable copy */
bd0388ae 478
bd0388ae 479 if (unlikely(OPTION_MD5 & options)) {
1a2c6181
CP
480 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
481 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
bd0388ae
WAS
482 /* overload cookie hash location */
483 opts->hash_location = (__u8 *)ptr;
33ad798c 484 ptr += 4;
40efc6fa 485 }
33ad798c 486
fd6149d3
IJ
487 if (unlikely(opts->mss)) {
488 *ptr++ = htonl((TCPOPT_MSS << 24) |
489 (TCPOLEN_MSS << 16) |
490 opts->mss);
491 }
492
bd0388ae
WAS
493 if (likely(OPTION_TS & options)) {
494 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
33ad798c
AL
495 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
496 (TCPOLEN_SACK_PERM << 16) |
497 (TCPOPT_TIMESTAMP << 8) |
498 TCPOLEN_TIMESTAMP);
bd0388ae 499 options &= ~OPTION_SACK_ADVERTISE;
33ad798c
AL
500 } else {
501 *ptr++ = htonl((TCPOPT_NOP << 24) |
502 (TCPOPT_NOP << 16) |
503 (TCPOPT_TIMESTAMP << 8) |
504 TCPOLEN_TIMESTAMP);
505 }
506 *ptr++ = htonl(opts->tsval);
507 *ptr++ = htonl(opts->tsecr);
508 }
509
bd0388ae 510 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
33ad798c
AL
511 *ptr++ = htonl((TCPOPT_NOP << 24) |
512 (TCPOPT_NOP << 16) |
513 (TCPOPT_SACK_PERM << 8) |
514 TCPOLEN_SACK_PERM);
515 }
516
bd0388ae 517 if (unlikely(OPTION_WSCALE & options)) {
33ad798c
AL
518 *ptr++ = htonl((TCPOPT_NOP << 24) |
519 (TCPOPT_WINDOW << 16) |
520 (TCPOLEN_WINDOW << 8) |
521 opts->ws);
522 }
523
524 if (unlikely(opts->num_sack_blocks)) {
525 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
526 tp->duplicate_sack : tp->selective_acks;
40efc6fa
SH
527 int this_sack;
528
529 *ptr++ = htonl((TCPOPT_NOP << 24) |
530 (TCPOPT_NOP << 16) |
531 (TCPOPT_SACK << 8) |
33ad798c 532 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
40efc6fa 533 TCPOLEN_SACK_PERBLOCK)));
2de979bd 534
33ad798c
AL
535 for (this_sack = 0; this_sack < opts->num_sack_blocks;
536 ++this_sack) {
40efc6fa
SH
537 *ptr++ = htonl(sp[this_sack].start_seq);
538 *ptr++ = htonl(sp[this_sack].end_seq);
539 }
2de979bd 540
5861f8e5 541 tp->rx_opt.dsack = 0;
40efc6fa 542 }
2100c8d2
YC
543
544 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
545 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
7f9b838b
DL
546 u8 *p = (u8 *)ptr;
547 u32 len; /* Fast Open option length */
548
549 if (foc->exp) {
550 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
551 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
552 TCPOPT_FASTOPEN_MAGIC);
553 p += TCPOLEN_EXP_FASTOPEN_BASE;
554 } else {
555 len = TCPOLEN_FASTOPEN_BASE + foc->len;
556 *p++ = TCPOPT_FASTOPEN;
557 *p++ = len;
558 }
2100c8d2 559
7f9b838b
DL
560 memcpy(p, foc->val, foc->len);
561 if ((len & 3) == 2) {
562 p[foc->len] = TCPOPT_NOP;
563 p[foc->len + 1] = TCPOPT_NOP;
2100c8d2 564 }
7f9b838b 565 ptr += (len + 3) >> 2;
2100c8d2 566 }
60e2a778
UB
567
568 smc_options_write(ptr, &options);
569}
570
571static void smc_set_option(const struct tcp_sock *tp,
572 struct tcp_out_options *opts,
573 unsigned int *remaining)
574{
575#if IS_ENABLED(CONFIG_SMC)
576 if (static_branch_unlikely(&tcp_have_smc)) {
577 if (tp->syn_smc) {
578 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
579 opts->options |= OPTION_SMC;
580 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
581 }
582 }
583 }
584#endif
585}
586
587static void smc_set_option_cond(const struct tcp_sock *tp,
588 const struct inet_request_sock *ireq,
589 struct tcp_out_options *opts,
590 unsigned int *remaining)
591{
592#if IS_ENABLED(CONFIG_SMC)
593 if (static_branch_unlikely(&tcp_have_smc)) {
594 if (tp->syn_smc && ireq->smc_ok) {
595 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
596 opts->options |= OPTION_SMC;
597 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
598 }
599 }
600 }
601#endif
33ad798c
AL
602}
603
67edfef7
AK
604/* Compute TCP options for SYN packets. This is not the final
605 * network wire format yet.
606 */
95c96174 607static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
33ad798c 608 struct tcp_out_options *opts,
cf533ea5
ED
609 struct tcp_md5sig_key **md5)
610{
33ad798c 611 struct tcp_sock *tp = tcp_sk(sk);
95c96174 612 unsigned int remaining = MAX_TCP_OPTION_SPACE;
783237e8 613 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
33ad798c 614
8c2320e8 615 *md5 = NULL;
cfb6eeb4 616#ifdef CONFIG_TCP_MD5SIG
8c2320e8
ED
617 if (unlikely(rcu_access_pointer(tp->md5sig_info))) {
618 *md5 = tp->af_specific->md5_lookup(sk, sk);
619 if (*md5) {
620 opts->options |= OPTION_MD5;
621 remaining -= TCPOLEN_MD5SIG_ALIGNED;
622 }
cfb6eeb4
YH
623 }
624#endif
33ad798c
AL
625
626 /* We always get an MSS option. The option bytes which will be seen in
627 * normal data packets should timestamps be used, must be in the MSS
628 * advertised. But we subtract them from tp->mss_cache so that
629 * calculations in tcp_sendmsg are simpler etc. So account for this
630 * fact here if necessary. If we don't do this correctly, as a
631 * receiver we won't recognize data packets as being full sized when we
632 * should, and thus we won't abide by the delayed ACK rules correctly.
633 * SACKs don't matter, we never delay an ACK when we have any of those
634 * going out. */
635 opts->mss = tcp_advertise_mss(sk);
bd0388ae 636 remaining -= TCPOLEN_MSS_ALIGNED;
33ad798c 637
5d2ed052 638 if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
33ad798c 639 opts->options |= OPTION_TS;
7faee5c0 640 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
33ad798c 641 opts->tsecr = tp->rx_opt.ts_recent;
bd0388ae 642 remaining -= TCPOLEN_TSTAMP_ALIGNED;
33ad798c 643 }
9bb37ef0 644 if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
33ad798c 645 opts->ws = tp->rx_opt.rcv_wscale;
89e95a61 646 opts->options |= OPTION_WSCALE;
bd0388ae 647 remaining -= TCPOLEN_WSCALE_ALIGNED;
33ad798c 648 }
f9301034 649 if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
33ad798c 650 opts->options |= OPTION_SACK_ADVERTISE;
b32d1310 651 if (unlikely(!(OPTION_TS & opts->options)))
bd0388ae 652 remaining -= TCPOLEN_SACKPERM_ALIGNED;
33ad798c
AL
653 }
654
783237e8 655 if (fastopen && fastopen->cookie.len >= 0) {
2646c831
DL
656 u32 need = fastopen->cookie.len;
657
658 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
659 TCPOLEN_FASTOPEN_BASE;
783237e8
YC
660 need = (need + 3) & ~3U; /* Align to 32 bits */
661 if (remaining >= need) {
662 opts->options |= OPTION_FAST_OPEN_COOKIE;
663 opts->fastopen_cookie = &fastopen->cookie;
664 remaining -= need;
665 tp->syn_fastopen = 1;
2646c831 666 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
783237e8
YC
667 }
668 }
bd0388ae 669
60e2a778
UB
670 smc_set_option(tp, opts, &remaining);
671
bd0388ae 672 return MAX_TCP_OPTION_SPACE - remaining;
40efc6fa
SH
673}
674
67edfef7 675/* Set up TCP options for SYN-ACKs. */
60e2a778
UB
676static unsigned int tcp_synack_options(const struct sock *sk,
677 struct request_sock *req,
37bfbdda
ED
678 unsigned int mss, struct sk_buff *skb,
679 struct tcp_out_options *opts,
680 const struct tcp_md5sig_key *md5,
681 struct tcp_fastopen_cookie *foc)
4957faad 682{
33ad798c 683 struct inet_request_sock *ireq = inet_rsk(req);
95c96174 684 unsigned int remaining = MAX_TCP_OPTION_SPACE;
33ad798c 685
cfb6eeb4 686#ifdef CONFIG_TCP_MD5SIG
80f03e27 687 if (md5) {
33ad798c 688 opts->options |= OPTION_MD5;
4957faad
WAS
689 remaining -= TCPOLEN_MD5SIG_ALIGNED;
690
691 /* We can't fit any SACK blocks in a packet with MD5 + TS
692 * options. There was discussion about disabling SACK
693 * rather than TS in order to fit in better with old,
694 * buggy kernels, but that was deemed to be unnecessary.
695 */
de213e5e 696 ireq->tstamp_ok &= !ireq->sack_ok;
cfb6eeb4
YH
697 }
698#endif
33ad798c 699
4957faad 700 /* We always send an MSS option. */
33ad798c 701 opts->mss = mss;
4957faad 702 remaining -= TCPOLEN_MSS_ALIGNED;
33ad798c
AL
703
704 if (likely(ireq->wscale_ok)) {
705 opts->ws = ireq->rcv_wscale;
89e95a61 706 opts->options |= OPTION_WSCALE;
4957faad 707 remaining -= TCPOLEN_WSCALE_ALIGNED;
33ad798c 708 }
de213e5e 709 if (likely(ireq->tstamp_ok)) {
33ad798c 710 opts->options |= OPTION_TS;
95a22cae 711 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
33ad798c 712 opts->tsecr = req->ts_recent;
4957faad 713 remaining -= TCPOLEN_TSTAMP_ALIGNED;
33ad798c
AL
714 }
715 if (likely(ireq->sack_ok)) {
716 opts->options |= OPTION_SACK_ADVERTISE;
de213e5e 717 if (unlikely(!ireq->tstamp_ok))
4957faad 718 remaining -= TCPOLEN_SACKPERM_ALIGNED;
33ad798c 719 }
7f9b838b
DL
720 if (foc != NULL && foc->len >= 0) {
721 u32 need = foc->len;
722
723 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
724 TCPOLEN_FASTOPEN_BASE;
8336886f
JC
725 need = (need + 3) & ~3U; /* Align to 32 bits */
726 if (remaining >= need) {
727 opts->options |= OPTION_FAST_OPEN_COOKIE;
728 opts->fastopen_cookie = foc;
729 remaining -= need;
730 }
731 }
1a2c6181 732
60e2a778
UB
733 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
734
4957faad 735 return MAX_TCP_OPTION_SPACE - remaining;
33ad798c
AL
736}
737
67edfef7
AK
738/* Compute TCP options for ESTABLISHED sockets. This is not the
739 * final wire format yet.
740 */
95c96174 741static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
33ad798c 742 struct tcp_out_options *opts,
cf533ea5
ED
743 struct tcp_md5sig_key **md5)
744{
33ad798c 745 struct tcp_sock *tp = tcp_sk(sk);
95c96174 746 unsigned int size = 0;
cabeccbd 747 unsigned int eff_sacks;
33ad798c 748
5843ef42
AK
749 opts->options = 0;
750
8c2320e8 751 *md5 = NULL;
33ad798c 752#ifdef CONFIG_TCP_MD5SIG
8c2320e8
ED
753 if (unlikely(rcu_access_pointer(tp->md5sig_info))) {
754 *md5 = tp->af_specific->md5_lookup(sk, sk);
755 if (*md5) {
756 opts->options |= OPTION_MD5;
757 size += TCPOLEN_MD5SIG_ALIGNED;
758 }
33ad798c 759 }
33ad798c
AL
760#endif
761
762 if (likely(tp->rx_opt.tstamp_ok)) {
763 opts->options |= OPTION_TS;
7faee5c0 764 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
33ad798c
AL
765 opts->tsecr = tp->rx_opt.ts_recent;
766 size += TCPOLEN_TSTAMP_ALIGNED;
767 }
768
cabeccbd
IJ
769 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
770 if (unlikely(eff_sacks)) {
95c96174 771 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
33ad798c 772 opts->num_sack_blocks =
95c96174 773 min_t(unsigned int, eff_sacks,
33ad798c
AL
774 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
775 TCPOLEN_SACK_PERBLOCK);
776 size += TCPOLEN_SACK_BASE_ALIGNED +
777 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
778 }
779
780 return size;
40efc6fa 781}
1da177e4 782
46d3ceab
ED
783
784/* TCP SMALL QUEUES (TSQ)
785 *
786 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
787 * to reduce RTT and bufferbloat.
788 * We do this using a special skb destructor (tcp_wfree).
789 *
790 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
791 * needs to be reallocated in a driver.
8e3bff96 792 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
46d3ceab
ED
793 *
794 * Since transmit from skb destructor is forbidden, we use a tasklet
795 * to process all sockets that eventually need to send more skbs.
796 * We use one tasklet per cpu, with its own queue of sockets.
797 */
798struct tsq_tasklet {
799 struct tasklet_struct tasklet;
800 struct list_head head; /* queue of tcp sockets */
801};
802static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
803
73a6bab5 804static void tcp_tsq_write(struct sock *sk)
6f458dfb
ED
805{
806 if ((1 << sk->sk_state) &
807 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
f9616c35
ED
808 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
809 struct tcp_sock *tp = tcp_sk(sk);
810
811 if (tp->lost_out > tp->retrans_out &&
3a91d29f
KD
812 tp->snd_cwnd > tcp_packets_in_flight(tp)) {
813 tcp_mstamp_refresh(tp);
f9616c35 814 tcp_xmit_retransmit_queue(sk);
3a91d29f 815 }
f9616c35
ED
816
817 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
bf06200e 818 0, GFP_ATOMIC);
f9616c35 819 }
6f458dfb 820}
73a6bab5
ED
821
822static void tcp_tsq_handler(struct sock *sk)
823{
824 bh_lock_sock(sk);
825 if (!sock_owned_by_user(sk))
826 tcp_tsq_write(sk);
827 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
828 sock_hold(sk);
829 bh_unlock_sock(sk);
830}
46d3ceab 831/*
8e3bff96 832 * One tasklet per cpu tries to send more skbs.
46d3ceab 833 * We run in tasklet context but need to disable irqs when
8e3bff96 834 * transferring tsq->head because tcp_wfree() might
46d3ceab
ED
835 * interrupt us (non NAPI drivers)
836 */
837static void tcp_tasklet_func(unsigned long data)
838{
839 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
840 LIST_HEAD(list);
841 unsigned long flags;
842 struct list_head *q, *n;
843 struct tcp_sock *tp;
844 struct sock *sk;
845
846 local_irq_save(flags);
847 list_splice_init(&tsq->head, &list);
848 local_irq_restore(flags);
849
850 list_for_each_safe(q, n, &list) {
851 tp = list_entry(q, struct tcp_sock, tsq_node);
852 list_del(&tp->tsq_node);
853
854 sk = (struct sock *)tp;
0a9648f1 855 smp_mb__before_atomic();
7aa5470c
ED
856 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
857
73a6bab5 858 tcp_tsq_handler(sk);
46d3ceab
ED
859 sk_free(sk);
860 }
861}
862
40fc3423
ED
863#define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
864 TCPF_WRITE_TIMER_DEFERRED | \
865 TCPF_DELACK_TIMER_DEFERRED | \
866 TCPF_MTU_REDUCED_DEFERRED)
46d3ceab
ED
867/**
868 * tcp_release_cb - tcp release_sock() callback
869 * @sk: socket
870 *
871 * called from release_sock() to perform protocol dependent
872 * actions before socket release.
873 */
874void tcp_release_cb(struct sock *sk)
875{
6f458dfb 876 unsigned long flags, nflags;
46d3ceab 877
6f458dfb
ED
878 /* perform an atomic operation only if at least one flag is set */
879 do {
7aa5470c 880 flags = sk->sk_tsq_flags;
6f458dfb
ED
881 if (!(flags & TCP_DEFERRED_ALL))
882 return;
883 nflags = flags & ~TCP_DEFERRED_ALL;
7aa5470c 884 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
6f458dfb 885
73a6bab5
ED
886 if (flags & TCPF_TSQ_DEFERRED) {
887 tcp_tsq_write(sk);
888 __sock_put(sk);
889 }
c3f9b018
ED
890 /* Here begins the tricky part :
891 * We are called from release_sock() with :
892 * 1) BH disabled
893 * 2) sk_lock.slock spinlock held
894 * 3) socket owned by us (sk->sk_lock.owned == 1)
895 *
896 * But following code is meant to be called from BH handlers,
897 * so we should keep BH disabled, but early release socket ownership
898 */
899 sock_release_ownership(sk);
900
40fc3423 901 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
6f458dfb 902 tcp_write_timer_handler(sk);
144d56e9
ED
903 __sock_put(sk);
904 }
40fc3423 905 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
6f458dfb 906 tcp_delack_timer_handler(sk);
144d56e9
ED
907 __sock_put(sk);
908 }
40fc3423 909 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
4fab9071 910 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
144d56e9
ED
911 __sock_put(sk);
912 }
46d3ceab
ED
913}
914EXPORT_SYMBOL(tcp_release_cb);
915
916void __init tcp_tasklet_init(void)
917{
918 int i;
919
920 for_each_possible_cpu(i) {
921 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
922
923 INIT_LIST_HEAD(&tsq->head);
924 tasklet_init(&tsq->tasklet,
925 tcp_tasklet_func,
926 (unsigned long)tsq);
927 }
928}
929
930/*
931 * Write buffer destructor automatically called from kfree_skb.
8e3bff96 932 * We can't xmit new skbs from this context, as we might already
46d3ceab
ED
933 * hold qdisc lock.
934 */
d6a4a104 935void tcp_wfree(struct sk_buff *skb)
46d3ceab
ED
936{
937 struct sock *sk = skb->sk;
938 struct tcp_sock *tp = tcp_sk(sk);
408f0a6c 939 unsigned long flags, nval, oval;
9b462d02
ED
940
941 /* Keep one reference on sk_wmem_alloc.
942 * Will be released by sk_free() from here or tcp_tasklet_func()
943 */
14afee4b 944 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
9b462d02
ED
945
946 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
947 * Wait until our queues (qdisc + devices) are drained.
948 * This gives :
949 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
950 * - chance for incoming ACK (processed by another cpu maybe)
951 * to migrate this flow (skb->ooo_okay will be eventually set)
952 */
14afee4b 953 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
9b462d02 954 goto out;
46d3ceab 955
7aa5470c 956 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
46d3ceab 957 struct tsq_tasklet *tsq;
a9b204d1 958 bool empty;
46d3ceab 959
408f0a6c
ED
960 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
961 goto out;
962
73a6bab5 963 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
7aa5470c 964 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
408f0a6c
ED
965 if (nval != oval)
966 continue;
967
46d3ceab
ED
968 /* queue this socket to tasklet queue */
969 local_irq_save(flags);
903ceff7 970 tsq = this_cpu_ptr(&tsq_tasklet);
a9b204d1 971 empty = list_empty(&tsq->head);
46d3ceab 972 list_add(&tp->tsq_node, &tsq->head);
a9b204d1
ED
973 if (empty)
974 tasklet_schedule(&tsq->tasklet);
46d3ceab 975 local_irq_restore(flags);
9b462d02 976 return;
46d3ceab 977 }
9b462d02
ED
978out:
979 sk_free(sk);
46d3ceab
ED
980}
981
73a6bab5
ED
982/* Note: Called under soft irq.
983 * We can call TCP stack right away, unless socket is owned by user.
218af599
ED
984 */
985enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
986{
987 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
988 struct sock *sk = (struct sock *)tp;
218af599 989
73a6bab5
ED
990 tcp_tsq_handler(sk);
991 sock_put(sk);
218af599 992
218af599
ED
993 return HRTIMER_NORESTART;
994}
995
c092dd5f 996static void tcp_internal_pacing(struct sock *sk)
218af599 997{
218af599
ED
998 if (!tcp_needs_internal_pacing(sk))
999 return;
218af599 1000 hrtimer_start(&tcp_sk(sk)->pacing_timer,
c092dd5f 1001 ns_to_ktime(tcp_sk(sk)->tcp_wstamp_ns),
73a6bab5
ED
1002 HRTIMER_MODE_ABS_PINNED_SOFT);
1003 sock_hold(sk);
218af599
ED
1004}
1005
ab408b6d 1006static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb)
e2080072 1007{
ab408b6d
ED
1008 struct tcp_sock *tp = tcp_sk(sk);
1009
d3edd06e 1010 skb->skb_mstamp_ns = tp->tcp_wstamp_ns;
ab408b6d
ED
1011 if (sk->sk_pacing_status != SK_PACING_NONE) {
1012 u32 rate = sk->sk_pacing_rate;
1013
1014 /* Original sch_fq does not pace first 10 MSS
1015 * Note that tp->data_segs_out overflows after 2^32 packets,
1016 * this is a minor annoyance.
1017 */
1018 if (rate != ~0U && rate && tp->data_segs_out >= 10) {
1019 tp->tcp_wstamp_ns += div_u64((u64)skb->len * NSEC_PER_SEC, rate);
c092dd5f
ED
1020
1021 tcp_internal_pacing(sk);
ab408b6d
ED
1022 }
1023 }
e2080072
ED
1024 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1025}
1026
1da177e4
LT
1027/* This routine actually transmits TCP packets queued in by
1028 * tcp_do_sendmsg(). This is used by both the initial
1029 * transmission and possible later retransmissions.
1030 * All SKB's seen here are completely headerless. It is our
1031 * job to build the TCP header, and pass the packet down to
1032 * IP so it can do the same plus pass the packet off to the
1033 * device.
1034 *
1035 * We are working here with either a clone of the original
1036 * SKB, or a fresh unique copy made by the retransmit engine.
1037 */
2987babb
YC
1038static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1039 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1da177e4 1040{
dfb4b9dc
DM
1041 const struct inet_connection_sock *icsk = inet_csk(sk);
1042 struct inet_sock *inet;
1043 struct tcp_sock *tp;
1044 struct tcp_skb_cb *tcb;
33ad798c 1045 struct tcp_out_options opts;
95c96174 1046 unsigned int tcp_options_size, tcp_header_size;
8c72c65b 1047 struct sk_buff *oskb = NULL;
cfb6eeb4 1048 struct tcp_md5sig_key *md5;
dfb4b9dc 1049 struct tcphdr *th;
dfb4b9dc
DM
1050 int err;
1051
1052 BUG_ON(!skb || !tcp_skb_pcount(skb));
6f094b9e 1053 tp = tcp_sk(sk);
dfb4b9dc 1054
ccdbb6e9 1055 if (clone_it) {
6f094b9e
LB
1056 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1057 - tp->snd_una;
8c72c65b 1058 oskb = skb;
e2080072
ED
1059
1060 tcp_skb_tsorted_save(oskb) {
1061 if (unlikely(skb_cloned(oskb)))
1062 skb = pskb_copy(oskb, gfp_mask);
1063 else
1064 skb = skb_clone(oskb, gfp_mask);
1065 } tcp_skb_tsorted_restore(oskb);
1066
dfb4b9dc
DM
1067 if (unlikely(!skb))
1068 return -ENOBUFS;
1069 }
d3edd06e 1070 skb->skb_mstamp_ns = tp->tcp_wstamp_ns;
1da177e4 1071
dfb4b9dc 1072 inet = inet_sk(sk);
dfb4b9dc 1073 tcb = TCP_SKB_CB(skb);
33ad798c 1074 memset(&opts, 0, sizeof(opts));
1da177e4 1075
4de075e0 1076 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
33ad798c
AL
1077 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1078 else
1079 tcp_options_size = tcp_established_options(sk, skb, &opts,
1080 &md5);
1081 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
e905a9ed 1082
547669d4 1083 /* if no packet is in qdisc/device queue, then allow XPS to select
b2532eb9 1084 * another queue. We can be called from tcp_tsq_handler()
73a6bab5 1085 * which holds one reference to sk.
b2532eb9
ED
1086 *
1087 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1088 * One way to get this would be to set skb->truesize = 2 on them.
547669d4 1089 */
b2532eb9 1090 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
dfb4b9dc 1091
38ab52e8
ED
1092 /* If we had to use memory reserve to allocate this skb,
1093 * this might cause drops if packet is looped back :
1094 * Other socket might not have SOCK_MEMALLOC.
1095 * Packets not looped back do not care about pfmemalloc.
1096 */
1097 skb->pfmemalloc = 0;
1098
aa8223c7
ACM
1099 skb_push(skb, tcp_header_size);
1100 skb_reset_transport_header(skb);
46d3ceab
ED
1101
1102 skb_orphan(skb);
1103 skb->sk = sk;
1d2077ac 1104 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
b73c3d0e 1105 skb_set_hash_from_sk(skb, sk);
14afee4b 1106 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
dfb4b9dc 1107
c3a2e837
JA
1108 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1109
dfb4b9dc 1110 /* Build TCP header and checksum it. */
ea1627c2 1111 th = (struct tcphdr *)skb->data;
c720c7e8
ED
1112 th->source = inet->inet_sport;
1113 th->dest = inet->inet_dport;
dfb4b9dc 1114 th->seq = htonl(tcb->seq);
2987babb 1115 th->ack_seq = htonl(rcv_nxt);
df7a3b07 1116 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
4de075e0 1117 tcb->tcp_flags);
dfb4b9dc 1118
dfb4b9dc
DM
1119 th->check = 0;
1120 th->urg_ptr = 0;
1da177e4 1121
33f5f57e 1122 /* The urg_mode check is necessary during a below snd_una win probe */
7691367d
HX
1123 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1124 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1125 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1126 th->urg = 1;
1127 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
0eae88f3 1128 th->urg_ptr = htons(0xFFFF);
7691367d
HX
1129 th->urg = 1;
1130 }
dfb4b9dc 1131 }
1da177e4 1132
bd0388ae 1133 tcp_options_write((__be32 *)(th + 1), tp, &opts);
51466a75 1134 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
ea1627c2
ED
1135 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1136 th->window = htons(tcp_select_window(sk));
1137 tcp_ecn_send(sk, skb, th, tcp_header_size);
1138 } else {
1139 /* RFC1323: The window in SYN & SYN/ACK segments
1140 * is never scaled.
1141 */
1142 th->window = htons(min(tp->rcv_wnd, 65535U));
1143 }
cfb6eeb4
YH
1144#ifdef CONFIG_TCP_MD5SIG
1145 /* Calculate the MD5 hash, as we have all we need now */
1146 if (md5) {
a465419b 1147 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
bd0388ae 1148 tp->af_specific->calc_md5_hash(opts.hash_location,
39f8e58e 1149 md5, sk, skb);
cfb6eeb4
YH
1150 }
1151#endif
1152
bb296246 1153 icsk->icsk_af_ops->send_check(sk, skb);
1da177e4 1154
4de075e0 1155 if (likely(tcb->tcp_flags & TCPHDR_ACK))
27cde44a 1156 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1da177e4 1157
a44d6eac 1158 if (skb->len != tcp_header_size) {
cf533ea5 1159 tcp_event_data_sent(tp, sk);
a44d6eac 1160 tp->data_segs_out += tcp_skb_pcount(skb);
ba113c3a 1161 tp->bytes_sent += skb->len - tcp_header_size;
a44d6eac 1162 }
1da177e4 1163
bd37a088 1164 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
aa2ea058
TH
1165 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1166 tcp_skb_pcount(skb));
1da177e4 1167
2efd055c 1168 tp->segs_out += tcp_skb_pcount(skb);
f69ad292 1169 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
cd7d8498 1170 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
f69ad292 1171 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
cd7d8498 1172
d3edd06e 1173 /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
971f10ec
ED
1174
1175 /* Cleanup our debris for IP stacks */
1176 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1177 sizeof(struct inet6_skb_parm)));
1178
b0270e91 1179 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
7faee5c0 1180
8c72c65b
ED
1181 if (unlikely(err > 0)) {
1182 tcp_enter_cwr(sk);
1183 err = net_xmit_eval(err);
1184 }
fc225799 1185 if (!err && oskb) {
ab408b6d 1186 tcp_update_skb_after_send(sk, oskb);
fc225799
ED
1187 tcp_rate_skb_sent(sk, oskb);
1188 }
8c72c65b 1189 return err;
1da177e4
LT
1190}
1191
2987babb
YC
1192static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1193 gfp_t gfp_mask)
1194{
1195 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1196 tcp_sk(sk)->rcv_nxt);
1197}
1198
67edfef7 1199/* This routine just queues the buffer for sending.
1da177e4
LT
1200 *
1201 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1202 * otherwise socket can stall.
1203 */
1204static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1205{
1206 struct tcp_sock *tp = tcp_sk(sk);
1207
1208 /* Advance write_seq and place onto the write_queue. */
1209 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
f4a775d1 1210 __skb_header_release(skb);
fe067e8a 1211 tcp_add_write_queue_tail(sk, skb);
3ab224be
HA
1212 sk->sk_wmem_queued += skb->truesize;
1213 sk_mem_charge(sk, skb->truesize);
1da177e4
LT
1214}
1215
67edfef7 1216/* Initialize TSO segments for a packet. */
5bbb432c 1217static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
f6302d1d 1218{
4a64fd6c 1219 if (skb->len <= mss_now) {
f6302d1d
DM
1220 /* Avoid the costly divide in the normal
1221 * non-TSO case.
1222 */
cd7d8498 1223 tcp_skb_pcount_set(skb, 1);
f69ad292 1224 TCP_SKB_CB(skb)->tcp_gso_size = 0;
f6302d1d 1225 } else {
cd7d8498 1226 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
f69ad292 1227 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1da177e4
LT
1228 }
1229}
1230
797108d1
IJ
1231/* Pcount in the middle of the write queue got changed, we need to do various
1232 * tweaks to fix counters
1233 */
cf533ea5 1234static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
797108d1
IJ
1235{
1236 struct tcp_sock *tp = tcp_sk(sk);
1237
1238 tp->packets_out -= decr;
1239
1240 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1241 tp->sacked_out -= decr;
1242 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1243 tp->retrans_out -= decr;
1244 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1245 tp->lost_out -= decr;
1246
1247 /* Reno case is special. Sigh... */
1248 if (tcp_is_reno(tp) && decr > 0)
1249 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1250
797108d1
IJ
1251 if (tp->lost_skb_hint &&
1252 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
713bafea 1253 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
797108d1
IJ
1254 tp->lost_cnt_hint -= decr;
1255
1256 tcp_verify_left_out(tp);
1257}
1258
0a2cf20c
SHY
1259static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1260{
1261 return TCP_SKB_CB(skb)->txstamp_ack ||
1262 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1263}
1264
490cc7d0
WB
1265static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1266{
1267 struct skb_shared_info *shinfo = skb_shinfo(skb);
1268
0a2cf20c 1269 if (unlikely(tcp_has_tx_tstamp(skb)) &&
490cc7d0
WB
1270 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1271 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1272 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1273
1274 shinfo->tx_flags &= ~tsflags;
1275 shinfo2->tx_flags |= tsflags;
1276 swap(shinfo->tskey, shinfo2->tskey);
b51e13fa
MKL
1277 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1278 TCP_SKB_CB(skb)->txstamp_ack = 0;
490cc7d0
WB
1279 }
1280}
1281
a166140e
MKL
1282static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1283{
1284 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1285 TCP_SKB_CB(skb)->eor = 0;
1286}
1287
75c119af
ED
1288/* Insert buff after skb on the write or rtx queue of sk. */
1289static void tcp_insert_write_queue_after(struct sk_buff *skb,
1290 struct sk_buff *buff,
1291 struct sock *sk,
1292 enum tcp_queue tcp_queue)
1293{
1294 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1295 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1296 else
1297 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1298}
1299
1da177e4
LT
1300/* Function to create two new TCP segments. Shrinks the given segment
1301 * to the specified size and appends a new segment with the rest of the
e905a9ed 1302 * packet to the list. This won't be called frequently, I hope.
1da177e4
LT
1303 * Remember, these are still headerless SKBs at this point.
1304 */
75c119af
ED
1305int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1306 struct sk_buff *skb, u32 len,
6cc55e09 1307 unsigned int mss_now, gfp_t gfp)
1da177e4
LT
1308{
1309 struct tcp_sock *tp = tcp_sk(sk);
1310 struct sk_buff *buff;
6475be16 1311 int nsize, old_factor;
b60b49ea 1312 int nlen;
9ce01461 1313 u8 flags;
1da177e4 1314
2fceec13
IJ
1315 if (WARN_ON(len > skb->len))
1316 return -EINVAL;
6a438bbe 1317
1da177e4
LT
1318 nsize = skb_headlen(skb) - len;
1319 if (nsize < 0)
1320 nsize = 0;
1321
6cc55e09 1322 if (skb_unclone(skb, gfp))
1da177e4
LT
1323 return -ENOMEM;
1324
1325 /* Get a new skb... force flag on. */
eb934478 1326 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
51456b29 1327 if (!buff)
1da177e4 1328 return -ENOMEM; /* We'll just try again later. */
ef5cb973 1329
3ab224be
HA
1330 sk->sk_wmem_queued += buff->truesize;
1331 sk_mem_charge(sk, buff->truesize);
b60b49ea
HX
1332 nlen = skb->len - len - nsize;
1333 buff->truesize += nlen;
1334 skb->truesize -= nlen;
1da177e4
LT
1335
1336 /* Correct the sequence numbers. */
1337 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1338 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1339 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1340
1341 /* PSH and FIN should only be set in the second packet. */
4de075e0
ED
1342 flags = TCP_SKB_CB(skb)->tcp_flags;
1343 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1344 TCP_SKB_CB(buff)->tcp_flags = flags;
e14c3caf 1345 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
a166140e 1346 tcp_skb_fragment_eor(skb, buff);
1da177e4 1347
98be9b12 1348 skb_split(skb, buff, len);
1da177e4 1349
98be9b12 1350 buff->ip_summed = CHECKSUM_PARTIAL;
1da177e4 1351
a61bbcf2 1352 buff->tstamp = skb->tstamp;
490cc7d0 1353 tcp_fragment_tstamp(skb, buff);
1da177e4 1354
6475be16
DM
1355 old_factor = tcp_skb_pcount(skb);
1356
1da177e4 1357 /* Fix up tso_factor for both original and new SKB. */
5bbb432c
ED
1358 tcp_set_skb_tso_segs(skb, mss_now);
1359 tcp_set_skb_tso_segs(buff, mss_now);
1da177e4 1360
b9f64820
YC
1361 /* Update delivered info for the new segment */
1362 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1363
6475be16
DM
1364 /* If this packet has been sent out already, we must
1365 * adjust the various packet counters.
1366 */
cf0b450c 1367 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
6475be16
DM
1368 int diff = old_factor - tcp_skb_pcount(skb) -
1369 tcp_skb_pcount(buff);
1da177e4 1370
797108d1
IJ
1371 if (diff)
1372 tcp_adjust_pcount(sk, skb, diff);
1da177e4
LT
1373 }
1374
1375 /* Link BUFF into the send queue. */
f4a775d1 1376 __skb_header_release(buff);
75c119af 1377 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
f67971e6
ED
1378 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1379 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1da177e4
LT
1380
1381 return 0;
1382}
1383
f4d01666
ED
1384/* This is similar to __pskb_pull_tail(). The difference is that pulled
1385 * data is not copied, but immediately discarded.
1da177e4 1386 */
7162fb24 1387static int __pskb_trim_head(struct sk_buff *skb, int len)
1da177e4 1388{
7b7fc97a 1389 struct skb_shared_info *shinfo;
1da177e4
LT
1390 int i, k, eat;
1391
4fa48bf3
ED
1392 eat = min_t(int, len, skb_headlen(skb));
1393 if (eat) {
1394 __skb_pull(skb, eat);
1395 len -= eat;
1396 if (!len)
7162fb24 1397 return 0;
4fa48bf3 1398 }
1da177e4
LT
1399 eat = len;
1400 k = 0;
7b7fc97a
ED
1401 shinfo = skb_shinfo(skb);
1402 for (i = 0; i < shinfo->nr_frags; i++) {
1403 int size = skb_frag_size(&shinfo->frags[i]);
9e903e08
ED
1404
1405 if (size <= eat) {
aff65da0 1406 skb_frag_unref(skb, i);
9e903e08 1407 eat -= size;
1da177e4 1408 } else {
7b7fc97a 1409 shinfo->frags[k] = shinfo->frags[i];
1da177e4 1410 if (eat) {
7b7fc97a
ED
1411 shinfo->frags[k].page_offset += eat;
1412 skb_frag_size_sub(&shinfo->frags[k], eat);
1da177e4
LT
1413 eat = 0;
1414 }
1415 k++;
1416 }
1417 }
7b7fc97a 1418 shinfo->nr_frags = k;
1da177e4 1419
1da177e4
LT
1420 skb->data_len -= len;
1421 skb->len = skb->data_len;
7162fb24 1422 return len;
1da177e4
LT
1423}
1424
67edfef7 1425/* Remove acked data from a packet in the transmit queue. */
1da177e4
LT
1426int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1427{
7162fb24
ED
1428 u32 delta_truesize;
1429
14bbd6a5 1430 if (skb_unclone(skb, GFP_ATOMIC))
1da177e4
LT
1431 return -ENOMEM;
1432
7162fb24 1433 delta_truesize = __pskb_trim_head(skb, len);
1da177e4
LT
1434
1435 TCP_SKB_CB(skb)->seq += len;
84fa7933 1436 skb->ip_summed = CHECKSUM_PARTIAL;
1da177e4 1437
7162fb24
ED
1438 if (delta_truesize) {
1439 skb->truesize -= delta_truesize;
1440 sk->sk_wmem_queued -= delta_truesize;
1441 sk_mem_uncharge(sk, delta_truesize);
1442 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1443 }
1da177e4 1444
5b35e1e6 1445 /* Any change of skb->len requires recalculation of tso factor. */
1da177e4 1446 if (tcp_skb_pcount(skb) > 1)
5bbb432c 1447 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1da177e4
LT
1448
1449 return 0;
1450}
1451
1b63edd6
YC
1452/* Calculate MSS not accounting any TCP options. */
1453static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
5d424d5a 1454{
cf533ea5
ED
1455 const struct tcp_sock *tp = tcp_sk(sk);
1456 const struct inet_connection_sock *icsk = inet_csk(sk);
5d424d5a
JH
1457 int mss_now;
1458
1459 /* Calculate base mss without TCP options:
1460 It is MMS_S - sizeof(tcphdr) of rfc1122
1461 */
1462 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1463
67469601
ED
1464 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1465 if (icsk->icsk_af_ops->net_frag_header_len) {
1466 const struct dst_entry *dst = __sk_dst_get(sk);
1467
1468 if (dst && dst_allfrag(dst))
1469 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1470 }
1471
5d424d5a
JH
1472 /* Clamp it (mss_clamp does not include tcp options) */
1473 if (mss_now > tp->rx_opt.mss_clamp)
1474 mss_now = tp->rx_opt.mss_clamp;
1475
1476 /* Now subtract optional transport overhead */
1477 mss_now -= icsk->icsk_ext_hdr_len;
1478
1479 /* Then reserve room for full set of TCP options and 8 bytes of data */
1480 if (mss_now < 48)
1481 mss_now = 48;
5d424d5a
JH
1482 return mss_now;
1483}
1484
1b63edd6
YC
1485/* Calculate MSS. Not accounting for SACKs here. */
1486int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1487{
1488 /* Subtract TCP options size, not including SACKs */
1489 return __tcp_mtu_to_mss(sk, pmtu) -
1490 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1491}
1492
5d424d5a 1493/* Inverse of above */
67469601 1494int tcp_mss_to_mtu(struct sock *sk, int mss)
5d424d5a 1495{
cf533ea5
ED
1496 const struct tcp_sock *tp = tcp_sk(sk);
1497 const struct inet_connection_sock *icsk = inet_csk(sk);
5d424d5a
JH
1498 int mtu;
1499
1500 mtu = mss +
1501 tp->tcp_header_len +
1502 icsk->icsk_ext_hdr_len +
1503 icsk->icsk_af_ops->net_header_len;
1504
67469601
ED
1505 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1506 if (icsk->icsk_af_ops->net_frag_header_len) {
1507 const struct dst_entry *dst = __sk_dst_get(sk);
1508
1509 if (dst && dst_allfrag(dst))
1510 mtu += icsk->icsk_af_ops->net_frag_header_len;
1511 }
5d424d5a
JH
1512 return mtu;
1513}
556c6b46 1514EXPORT_SYMBOL(tcp_mss_to_mtu);
5d424d5a 1515
67edfef7 1516/* MTU probing init per socket */
5d424d5a
JH
1517void tcp_mtup_init(struct sock *sk)
1518{
1519 struct tcp_sock *tp = tcp_sk(sk);
1520 struct inet_connection_sock *icsk = inet_csk(sk);
b0f9ca53 1521 struct net *net = sock_net(sk);
5d424d5a 1522
b0f9ca53 1523 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
5d424d5a 1524 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
e905a9ed 1525 icsk->icsk_af_ops->net_header_len;
b0f9ca53 1526 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
5d424d5a 1527 icsk->icsk_mtup.probe_size = 0;
05cbc0db 1528 if (icsk->icsk_mtup.enabled)
c74df29a 1529 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
5d424d5a 1530}
4bc2f18b 1531EXPORT_SYMBOL(tcp_mtup_init);
5d424d5a 1532
1da177e4
LT
1533/* This function synchronize snd mss to current pmtu/exthdr set.
1534
1535 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1536 for TCP options, but includes only bare TCP header.
1537
1538 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
caa20d9a 1539 It is minimum of user_mss and mss received with SYN.
1da177e4
LT
1540 It also does not include TCP options.
1541
d83d8461 1542 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1da177e4
LT
1543
1544 tp->mss_cache is current effective sending mss, including
1545 all tcp options except for SACKs. It is evaluated,
1546 taking into account current pmtu, but never exceeds
1547 tp->rx_opt.mss_clamp.
1548
1549 NOTE1. rfc1122 clearly states that advertised MSS
1550 DOES NOT include either tcp or ip options.
1551
d83d8461
ACM
1552 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1553 are READ ONLY outside this function. --ANK (980731)
1da177e4 1554 */
1da177e4
LT
1555unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1556{
1557 struct tcp_sock *tp = tcp_sk(sk);
d83d8461 1558 struct inet_connection_sock *icsk = inet_csk(sk);
5d424d5a 1559 int mss_now;
1da177e4 1560
5d424d5a
JH
1561 if (icsk->icsk_mtup.search_high > pmtu)
1562 icsk->icsk_mtup.search_high = pmtu;
1da177e4 1563
5d424d5a 1564 mss_now = tcp_mtu_to_mss(sk, pmtu);
409d22b4 1565 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1da177e4
LT
1566
1567 /* And store cached results */
d83d8461 1568 icsk->icsk_pmtu_cookie = pmtu;
5d424d5a
JH
1569 if (icsk->icsk_mtup.enabled)
1570 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
c1b4a7e6 1571 tp->mss_cache = mss_now;
1da177e4
LT
1572
1573 return mss_now;
1574}
4bc2f18b 1575EXPORT_SYMBOL(tcp_sync_mss);
1da177e4
LT
1576
1577/* Compute the current effective MSS, taking SACKs and IP options,
1578 * and even PMTU discovery events into account.
1da177e4 1579 */
0c54b85f 1580unsigned int tcp_current_mss(struct sock *sk)
1da177e4 1581{
cf533ea5
ED
1582 const struct tcp_sock *tp = tcp_sk(sk);
1583 const struct dst_entry *dst = __sk_dst_get(sk);
c1b4a7e6 1584 u32 mss_now;
95c96174 1585 unsigned int header_len;
33ad798c
AL
1586 struct tcp_out_options opts;
1587 struct tcp_md5sig_key *md5;
c1b4a7e6
DM
1588
1589 mss_now = tp->mss_cache;
1590
1da177e4
LT
1591 if (dst) {
1592 u32 mtu = dst_mtu(dst);
d83d8461 1593 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1da177e4
LT
1594 mss_now = tcp_sync_mss(sk, mtu);
1595 }
1596
33ad798c
AL
1597 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1598 sizeof(struct tcphdr);
1599 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1600 * some common options. If this is an odd packet (because we have SACK
1601 * blocks etc) then our calculated header_len will be different, and
1602 * we have to adjust mss_now correspondingly */
1603 if (header_len != tp->tcp_header_len) {
1604 int delta = (int) header_len - tp->tcp_header_len;
1605 mss_now -= delta;
1606 }
cfb6eeb4 1607
1da177e4
LT
1608 return mss_now;
1609}
1610
86fd14ad
WP
1611/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1612 * As additional protections, we do not touch cwnd in retransmission phases,
1613 * and if application hit its sndbuf limit recently.
1614 */
1615static void tcp_cwnd_application_limited(struct sock *sk)
1616{
1617 struct tcp_sock *tp = tcp_sk(sk);
1618
1619 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1620 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1621 /* Limited by application or receiver window. */
1622 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1623 u32 win_used = max(tp->snd_cwnd_used, init_win);
1624 if (win_used < tp->snd_cwnd) {
1625 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1626 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1627 }
1628 tp->snd_cwnd_used = 0;
1629 }
c2203cf7 1630 tp->snd_cwnd_stamp = tcp_jiffies32;
86fd14ad
WP
1631}
1632
ca8a2263 1633static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
a762a980 1634{
1b1fc3fd 1635 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
9e412ba7 1636 struct tcp_sock *tp = tcp_sk(sk);
a762a980 1637
ca8a2263
NC
1638 /* Track the maximum number of outstanding packets in each
1639 * window, and remember whether we were cwnd-limited then.
1640 */
1641 if (!before(tp->snd_una, tp->max_packets_seq) ||
1642 tp->packets_out > tp->max_packets_out) {
1643 tp->max_packets_out = tp->packets_out;
1644 tp->max_packets_seq = tp->snd_nxt;
1645 tp->is_cwnd_limited = is_cwnd_limited;
1646 }
e114a710 1647
24901551 1648 if (tcp_is_cwnd_limited(sk)) {
a762a980
DM
1649 /* Network is feed fully. */
1650 tp->snd_cwnd_used = 0;
c2203cf7 1651 tp->snd_cwnd_stamp = tcp_jiffies32;
a762a980
DM
1652 } else {
1653 /* Network starves. */
1654 if (tp->packets_out > tp->snd_cwnd_used)
1655 tp->snd_cwnd_used = tp->packets_out;
1656
b510f0d2 1657 if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle &&
c2203cf7 1658 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1b1fc3fd 1659 !ca_ops->cong_control)
a762a980 1660 tcp_cwnd_application_limited(sk);
b0f71bd3
FY
1661
1662 /* The following conditions together indicate the starvation
1663 * is caused by insufficient sender buffer:
1664 * 1) just sent some data (see tcp_write_xmit)
1665 * 2) not cwnd limited (this else condition)
75c119af 1666 * 3) no more data to send (tcp_write_queue_empty())
b0f71bd3
FY
1667 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1668 */
75c119af 1669 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
b0f71bd3
FY
1670 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1671 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1672 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
a762a980
DM
1673 }
1674}
1675
d4589926
ED
1676/* Minshall's variant of the Nagle send check. */
1677static bool tcp_minshall_check(const struct tcp_sock *tp)
1678{
1679 return after(tp->snd_sml, tp->snd_una) &&
1680 !after(tp->snd_sml, tp->snd_nxt);
1681}
1682
1683/* Update snd_sml if this skb is under mss
1684 * Note that a TSO packet might end with a sub-mss segment
1685 * The test is really :
1686 * if ((skb->len % mss) != 0)
1687 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1688 * But we can avoid doing the divide again given we already have
1689 * skb_pcount = skb->len / mss_now
0e3a4803 1690 */
d4589926
ED
1691static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1692 const struct sk_buff *skb)
1693{
1694 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1695 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1696}
1697
1698/* Return false, if packet can be sent now without violation Nagle's rules:
1699 * 1. It is full sized. (provided by caller in %partial bool)
1700 * 2. Or it contains FIN. (already checked by caller)
1701 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1702 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1703 * With Minshall's modification: all sent small packets are ACKed.
1704 */
1705static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
cc93fc51 1706 int nonagle)
d4589926
ED
1707{
1708 return partial &&
1709 ((nonagle & TCP_NAGLE_CORK) ||
1710 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1711}
605ad7f1
ED
1712
1713/* Return how many segs we'd like on a TSO packet,
1714 * to send one TSO packet per ms
1715 */
dcb8c9b4
ED
1716static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1717 int min_tso_segs)
605ad7f1
ED
1718{
1719 u32 bytes, segs;
1720
3a9b76fd 1721 bytes = min(sk->sk_pacing_rate >> sk->sk_pacing_shift,
605ad7f1
ED
1722 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1723
1724 /* Goal is to send at least one packet per ms,
1725 * not one big TSO packet every 100 ms.
1726 * This preserves ACK clocking and is consistent
1727 * with tcp_tso_should_defer() heuristic.
1728 */
1b3878ca 1729 segs = max_t(u32, bytes / mss_now, min_tso_segs);
605ad7f1 1730
350c9f48 1731 return segs;
605ad7f1
ED
1732}
1733
ed6e7268
NC
1734/* Return the number of segments we want in the skb we are transmitting.
1735 * See if congestion control module wants to decide; otherwise, autosize.
1736 */
1737static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1738{
1739 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
dcb8c9b4 1740 u32 min_tso, tso_segs;
ed6e7268 1741
dcb8c9b4
ED
1742 min_tso = ca_ops->min_tso_segs ?
1743 ca_ops->min_tso_segs(sk) :
1744 sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs;
1745
1746 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
350c9f48 1747 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
ed6e7268
NC
1748}
1749
d4589926
ED
1750/* Returns the portion of skb which can be sent right away */
1751static unsigned int tcp_mss_split_point(const struct sock *sk,
1752 const struct sk_buff *skb,
1753 unsigned int mss_now,
1754 unsigned int max_segs,
1755 int nonagle)
c1b4a7e6 1756{
cf533ea5 1757 const struct tcp_sock *tp = tcp_sk(sk);
d4589926 1758 u32 partial, needed, window, max_len;
c1b4a7e6 1759
90840def 1760 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1485348d 1761 max_len = mss_now * max_segs;
0e3a4803 1762
1485348d
BH
1763 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1764 return max_len;
0e3a4803 1765
5ea3a748
IJ
1766 needed = min(skb->len, window);
1767
1485348d
BH
1768 if (max_len <= needed)
1769 return max_len;
0e3a4803 1770
d4589926
ED
1771 partial = needed % mss_now;
1772 /* If last segment is not a full MSS, check if Nagle rules allow us
1773 * to include this last segment in this skb.
1774 * Otherwise, we'll split the skb at last MSS boundary
1775 */
cc93fc51 1776 if (tcp_nagle_check(partial != 0, tp, nonagle))
d4589926
ED
1777 return needed - partial;
1778
1779 return needed;
c1b4a7e6
DM
1780}
1781
1782/* Can at least one segment of SKB be sent right now, according to the
1783 * congestion window rules? If so, return how many segments are allowed.
1784 */
cf533ea5
ED
1785static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1786 const struct sk_buff *skb)
c1b4a7e6 1787{
d649a7a8 1788 u32 in_flight, cwnd, halfcwnd;
c1b4a7e6
DM
1789
1790 /* Don't be strict about the congestion window for the final FIN. */
4de075e0
ED
1791 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1792 tcp_skb_pcount(skb) == 1)
c1b4a7e6
DM
1793 return 1;
1794
1795 in_flight = tcp_packets_in_flight(tp);
1796 cwnd = tp->snd_cwnd;
d649a7a8
ED
1797 if (in_flight >= cwnd)
1798 return 0;
c1b4a7e6 1799
d649a7a8
ED
1800 /* For better scheduling, ensure we have at least
1801 * 2 GSO packets in flight.
1802 */
1803 halfcwnd = max(cwnd >> 1, 1U);
1804 return min(halfcwnd, cwnd - in_flight);
c1b4a7e6
DM
1805}
1806
b595076a 1807/* Initialize TSO state of a skb.
67edfef7 1808 * This must be invoked the first time we consider transmitting
c1b4a7e6
DM
1809 * SKB onto the wire.
1810 */
5bbb432c 1811static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
c1b4a7e6
DM
1812{
1813 int tso_segs = tcp_skb_pcount(skb);
1814
f8269a49 1815 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
5bbb432c 1816 tcp_set_skb_tso_segs(skb, mss_now);
c1b4a7e6
DM
1817 tso_segs = tcp_skb_pcount(skb);
1818 }
1819 return tso_segs;
1820}
1821
c1b4a7e6 1822
a2a385d6 1823/* Return true if the Nagle test allows this packet to be
c1b4a7e6
DM
1824 * sent now.
1825 */
a2a385d6
ED
1826static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1827 unsigned int cur_mss, int nonagle)
c1b4a7e6
DM
1828{
1829 /* Nagle rule does not apply to frames, which sit in the middle of the
1830 * write_queue (they have no chances to get new data).
1831 *
1832 * This is implemented in the callers, where they modify the 'nonagle'
1833 * argument based upon the location of SKB in the send queue.
1834 */
1835 if (nonagle & TCP_NAGLE_PUSH)
a2a385d6 1836 return true;
c1b4a7e6 1837
9b44190d
YC
1838 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1839 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
a2a385d6 1840 return true;
c1b4a7e6 1841
cc93fc51 1842 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
a2a385d6 1843 return true;
c1b4a7e6 1844
a2a385d6 1845 return false;
c1b4a7e6
DM
1846}
1847
1848/* Does at least the first segment of SKB fit into the send window? */
a2a385d6
ED
1849static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1850 const struct sk_buff *skb,
1851 unsigned int cur_mss)
c1b4a7e6
DM
1852{
1853 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1854
1855 if (skb->len > cur_mss)
1856 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1857
90840def 1858 return !after(end_seq, tcp_wnd_end(tp));
c1b4a7e6
DM
1859}
1860
c1b4a7e6
DM
1861/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1862 * which is put after SKB on the list. It is very much like
1863 * tcp_fragment() except that it may make several kinds of assumptions
1864 * in order to speed up the splitting operation. In particular, we
1865 * know that all the data is in scatter-gather pages, and that the
1866 * packet has never been sent out before (and thus is not cloned).
1867 */
75c119af
ED
1868static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1869 struct sk_buff *skb, unsigned int len,
c4ead4c5 1870 unsigned int mss_now, gfp_t gfp)
c1b4a7e6
DM
1871{
1872 struct sk_buff *buff;
1873 int nlen = skb->len - len;
9ce01461 1874 u8 flags;
c1b4a7e6
DM
1875
1876 /* All of a TSO frame must be composed of paged data. */
c8ac3774 1877 if (skb->len != skb->data_len)
75c119af 1878 return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp);
c1b4a7e6 1879
eb934478 1880 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
51456b29 1881 if (unlikely(!buff))
c1b4a7e6
DM
1882 return -ENOMEM;
1883
3ab224be
HA
1884 sk->sk_wmem_queued += buff->truesize;
1885 sk_mem_charge(sk, buff->truesize);
b60b49ea 1886 buff->truesize += nlen;
c1b4a7e6
DM
1887 skb->truesize -= nlen;
1888
1889 /* Correct the sequence numbers. */
1890 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1891 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1892 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1893
1894 /* PSH and FIN should only be set in the second packet. */
4de075e0
ED
1895 flags = TCP_SKB_CB(skb)->tcp_flags;
1896 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1897 TCP_SKB_CB(buff)->tcp_flags = flags;
c1b4a7e6
DM
1898
1899 /* This packet was never sent out yet, so no SACK bits. */
1900 TCP_SKB_CB(buff)->sacked = 0;
1901
a166140e
MKL
1902 tcp_skb_fragment_eor(skb, buff);
1903
98be9b12 1904 buff->ip_summed = CHECKSUM_PARTIAL;
c1b4a7e6 1905 skb_split(skb, buff, len);
490cc7d0 1906 tcp_fragment_tstamp(skb, buff);
c1b4a7e6
DM
1907
1908 /* Fix up tso_factor for both original and new SKB. */
5bbb432c
ED
1909 tcp_set_skb_tso_segs(skb, mss_now);
1910 tcp_set_skb_tso_segs(buff, mss_now);
c1b4a7e6
DM
1911
1912 /* Link BUFF into the send queue. */
f4a775d1 1913 __skb_header_release(buff);
75c119af 1914 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
c1b4a7e6
DM
1915
1916 return 0;
1917}
1918
1919/* Try to defer sending, if possible, in order to minimize the amount
1920 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1921 *
1922 * This algorithm is from John Heffner.
1923 */
ca8a2263 1924static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
605ad7f1 1925 bool *is_cwnd_limited, u32 max_segs)
c1b4a7e6 1926{
6687e988 1927 const struct inet_connection_sock *icsk = inet_csk(sk);
50c8339e
ED
1928 u32 age, send_win, cong_win, limit, in_flight;
1929 struct tcp_sock *tp = tcp_sk(sk);
50c8339e 1930 struct sk_buff *head;
ad9f4f50 1931 int win_divisor;
c1b4a7e6 1932
4de075e0 1933 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
ae8064ac 1934 goto send_now;
c1b4a7e6 1935
99d7662a 1936 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
ae8064ac
JH
1937 goto send_now;
1938
5f852eb5
ED
1939 /* Avoid bursty behavior by allowing defer
1940 * only if the last write was recent.
1941 */
d635fbe2 1942 if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0)
ae8064ac 1943 goto send_now;
908a75c1 1944
c1b4a7e6
DM
1945 in_flight = tcp_packets_in_flight(tp);
1946
c8c9aeb5
SB
1947 BUG_ON(tcp_skb_pcount(skb) <= 1);
1948 BUG_ON(tp->snd_cwnd <= in_flight);
c1b4a7e6 1949
90840def 1950 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
c1b4a7e6
DM
1951
1952 /* From in_flight test above, we know that cwnd > in_flight. */
1953 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1954
1955 limit = min(send_win, cong_win);
1956
ba244fe9 1957 /* If a full-sized TSO skb can be sent, do it. */
605ad7f1 1958 if (limit >= max_segs * tp->mss_cache)
ae8064ac 1959 goto send_now;
ba244fe9 1960
62ad2761
IJ
1961 /* Middle in queue won't get any more data, full sendable already? */
1962 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1963 goto send_now;
1964
5bbcc0f5 1965 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
ad9f4f50 1966 if (win_divisor) {
c1b4a7e6
DM
1967 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1968
1969 /* If at least some fraction of a window is available,
1970 * just use it.
1971 */
ad9f4f50 1972 chunk /= win_divisor;
c1b4a7e6 1973 if (limit >= chunk)
ae8064ac 1974 goto send_now;
c1b4a7e6
DM
1975 } else {
1976 /* Different approach, try not to defer past a single
1977 * ACK. Receiver should ACK every other full sized
1978 * frame, so if we have space for more than 3 frames
1979 * then send now.
1980 */
6b5a5c0d 1981 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
ae8064ac 1982 goto send_now;
c1b4a7e6
DM
1983 }
1984
75c119af
ED
1985 /* TODO : use tsorted_sent_queue ? */
1986 head = tcp_rtx_queue_head(sk);
1987 if (!head)
1988 goto send_now;
2fd66ffb 1989 age = tcp_stamp_us_delta(tp->tcp_mstamp, tcp_skb_timestamp_us(head));
50c8339e
ED
1990 /* If next ACK is likely to come too late (half srtt), do not defer */
1991 if (age < (tp->srtt_us >> 4))
1992 goto send_now;
1993
5f852eb5 1994 /* Ok, it looks like it is advisable to defer. */
ae8064ac 1995
d2e1339f 1996 if (cong_win < send_win && cong_win <= skb->len)
ca8a2263
NC
1997 *is_cwnd_limited = true;
1998
a2a385d6 1999 return true;
ae8064ac
JH
2000
2001send_now:
a2a385d6 2002 return false;
c1b4a7e6
DM
2003}
2004
05cbc0db
FD
2005static inline void tcp_mtu_check_reprobe(struct sock *sk)
2006{
2007 struct inet_connection_sock *icsk = inet_csk(sk);
2008 struct tcp_sock *tp = tcp_sk(sk);
2009 struct net *net = sock_net(sk);
2010 u32 interval;
2011 s32 delta;
2012
2013 interval = net->ipv4.sysctl_tcp_probe_interval;
c74df29a 2014 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
05cbc0db
FD
2015 if (unlikely(delta >= interval * HZ)) {
2016 int mss = tcp_current_mss(sk);
2017
2018 /* Update current search range */
2019 icsk->icsk_mtup.probe_size = 0;
2020 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2021 sizeof(struct tcphdr) +
2022 icsk->icsk_af_ops->net_header_len;
2023 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2024
2025 /* Update probe time stamp */
c74df29a 2026 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
05cbc0db
FD
2027 }
2028}
2029
808cf9e3
IL
2030static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2031{
2032 struct sk_buff *skb, *next;
2033
2034 skb = tcp_send_head(sk);
2035 tcp_for_write_queue_from_safe(skb, next, sk) {
2036 if (len <= skb->len)
2037 break;
2038
2039 if (unlikely(TCP_SKB_CB(skb)->eor))
2040 return false;
2041
2042 len -= skb->len;
2043 }
2044
2045 return true;
2046}
2047
5d424d5a 2048/* Create a new MTU probe if we are ready.
67edfef7
AK
2049 * MTU probe is regularly attempting to increase the path MTU by
2050 * deliberately sending larger packets. This discovers routing
2051 * changes resulting in larger path MTUs.
2052 *
5d424d5a
JH
2053 * Returns 0 if we should wait to probe (no cwnd available),
2054 * 1 if a probe was sent,
056834d9
IJ
2055 * -1 otherwise
2056 */
5d424d5a
JH
2057static int tcp_mtu_probe(struct sock *sk)
2058{
5d424d5a 2059 struct inet_connection_sock *icsk = inet_csk(sk);
12a59abc 2060 struct tcp_sock *tp = tcp_sk(sk);
5d424d5a 2061 struct sk_buff *skb, *nskb, *next;
6b58e0a5 2062 struct net *net = sock_net(sk);
5d424d5a 2063 int probe_size;
91cc17c0 2064 int size_needed;
12a59abc 2065 int copy, len;
5d424d5a 2066 int mss_now;
6b58e0a5 2067 int interval;
5d424d5a
JH
2068
2069 /* Not currently probing/verifying,
2070 * not in recovery,
2071 * have enough cwnd, and
12a59abc
ED
2072 * not SACKing (the variable headers throw things off)
2073 */
2074 if (likely(!icsk->icsk_mtup.enabled ||
2075 icsk->icsk_mtup.probe_size ||
2076 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2077 tp->snd_cwnd < 11 ||
2078 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
5d424d5a
JH
2079 return -1;
2080
6b58e0a5
FD
2081 /* Use binary search for probe_size between tcp_mss_base,
2082 * and current mss_clamp. if (search_high - search_low)
2083 * smaller than a threshold, backoff from probing.
2084 */
0c54b85f 2085 mss_now = tcp_current_mss(sk);
6b58e0a5
FD
2086 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2087 icsk->icsk_mtup.search_low) >> 1);
91cc17c0 2088 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
6b58e0a5 2089 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
05cbc0db
FD
2090 /* When misfortune happens, we are reprobing actively,
2091 * and then reprobe timer has expired. We stick with current
2092 * probing process by not resetting search range to its orignal.
2093 */
6b58e0a5 2094 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
05cbc0db
FD
2095 interval < net->ipv4.sysctl_tcp_probe_threshold) {
2096 /* Check whether enough time has elaplased for
2097 * another round of probing.
2098 */
2099 tcp_mtu_check_reprobe(sk);
5d424d5a
JH
2100 return -1;
2101 }
2102
2103 /* Have enough data in the send queue to probe? */
7f9c33e5 2104 if (tp->write_seq - tp->snd_nxt < size_needed)
5d424d5a
JH
2105 return -1;
2106
91cc17c0
IJ
2107 if (tp->snd_wnd < size_needed)
2108 return -1;
90840def 2109 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
91cc17c0 2110 return 0;
5d424d5a 2111
d67c58e9
IJ
2112 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2113 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2114 if (!tcp_packets_in_flight(tp))
5d424d5a
JH
2115 return -1;
2116 else
2117 return 0;
2118 }
2119
808cf9e3
IL
2120 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2121 return -1;
2122
5d424d5a 2123 /* We're allowed to probe. Build it now. */
eb934478 2124 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
51456b29 2125 if (!nskb)
5d424d5a 2126 return -1;
3ab224be
HA
2127 sk->sk_wmem_queued += nskb->truesize;
2128 sk_mem_charge(sk, nskb->truesize);
5d424d5a 2129
fe067e8a 2130 skb = tcp_send_head(sk);
5d424d5a
JH
2131
2132 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2133 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
4de075e0 2134 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
5d424d5a
JH
2135 TCP_SKB_CB(nskb)->sacked = 0;
2136 nskb->csum = 0;
98be9b12 2137 nskb->ip_summed = CHECKSUM_PARTIAL;
5d424d5a 2138
50c4817e 2139 tcp_insert_write_queue_before(nskb, skb, sk);
2b7cda9c 2140 tcp_highest_sack_replace(sk, skb, nskb);
50c4817e 2141
5d424d5a 2142 len = 0;
234b6860 2143 tcp_for_write_queue_from_safe(skb, next, sk) {
5d424d5a 2144 copy = min_t(int, skb->len, probe_size - len);
98be9b12 2145 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
5d424d5a
JH
2146
2147 if (skb->len <= copy) {
2148 /* We've eaten all the data from this skb.
2149 * Throw it away. */
4de075e0 2150 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
808cf9e3
IL
2151 /* If this is the last SKB we copy and eor is set
2152 * we need to propagate it to the new skb.
2153 */
2154 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
fe067e8a 2155 tcp_unlink_write_queue(skb, sk);
3ab224be 2156 sk_wmem_free_skb(sk, skb);
5d424d5a 2157 } else {
4de075e0 2158 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
a3433f35 2159 ~(TCPHDR_FIN|TCPHDR_PSH);
5d424d5a
JH
2160 if (!skb_shinfo(skb)->nr_frags) {
2161 skb_pull(skb, copy);
5d424d5a
JH
2162 } else {
2163 __pskb_trim_head(skb, copy);
5bbb432c 2164 tcp_set_skb_tso_segs(skb, mss_now);
5d424d5a
JH
2165 }
2166 TCP_SKB_CB(skb)->seq += copy;
2167 }
2168
2169 len += copy;
234b6860
IJ
2170
2171 if (len >= probe_size)
2172 break;
5d424d5a 2173 }
5bbb432c 2174 tcp_init_tso_segs(nskb, nskb->len);
5d424d5a
JH
2175
2176 /* We're ready to send. If this fails, the probe will
7faee5c0
ED
2177 * be resegmented into mss-sized pieces by tcp_write_xmit().
2178 */
5d424d5a
JH
2179 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2180 /* Decrement cwnd here because we are sending
056834d9 2181 * effectively two packets. */
5d424d5a 2182 tp->snd_cwnd--;
66f5fe62 2183 tcp_event_new_data_sent(sk, nskb);
5d424d5a
JH
2184
2185 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
0e7b1368
JH
2186 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2187 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
5d424d5a
JH
2188
2189 return 1;
2190 }
2191
2192 return -1;
2193}
2194
218af599
ED
2195static bool tcp_pacing_check(const struct sock *sk)
2196{
2197 return tcp_needs_internal_pacing(sk) &&
73a6bab5 2198 hrtimer_is_queued(&tcp_sk(sk)->pacing_timer);
218af599
ED
2199}
2200
f9616c35
ED
2201/* TCP Small Queues :
2202 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2203 * (These limits are doubled for retransmits)
2204 * This allows for :
2205 * - better RTT estimation and ACK scheduling
2206 * - faster recovery
2207 * - high rates
2208 * Alas, some drivers / subsystems require a fair amount
2209 * of queued bytes to ensure line rate.
2210 * One example is wifi aggregation (802.11 AMPDU)
2211 */
2212static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2213 unsigned int factor)
2214{
2215 unsigned int limit;
2216
3a9b76fd 2217 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> sk->sk_pacing_shift);
9184d8bb
ED
2218 limit = min_t(u32, limit,
2219 sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
f9616c35
ED
2220 limit <<= factor;
2221
14afee4b 2222 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
75c119af 2223 /* Always send skb if rtx queue is empty.
75eefc6c
ED
2224 * No need to wait for TX completion to call us back,
2225 * after softirq/tasklet schedule.
2226 * This helps when TX completions are delayed too much.
2227 */
75c119af 2228 if (tcp_rtx_queue_empty(sk))
75eefc6c
ED
2229 return false;
2230
7aa5470c 2231 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
f9616c35
ED
2232 /* It is possible TX completion already happened
2233 * before we set TSQ_THROTTLED, so we must
2234 * test again the condition.
2235 */
2236 smp_mb__after_atomic();
14afee4b 2237 if (refcount_read(&sk->sk_wmem_alloc) > limit)
f9616c35
ED
2238 return true;
2239 }
2240 return false;
2241}
2242
05b055e8
FY
2243static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2244{
628174cc 2245 const u32 now = tcp_jiffies32;
efe967cd 2246 enum tcp_chrono old = tp->chrono_type;
05b055e8 2247
efe967cd
AB
2248 if (old > TCP_CHRONO_UNSPEC)
2249 tp->chrono_stat[old - 1] += now - tp->chrono_start;
05b055e8
FY
2250 tp->chrono_start = now;
2251 tp->chrono_type = new;
2252}
2253
2254void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2255{
2256 struct tcp_sock *tp = tcp_sk(sk);
2257
2258 /* If there are multiple conditions worthy of tracking in a
0f87230d
FY
2259 * chronograph then the highest priority enum takes precedence
2260 * over the other conditions. So that if something "more interesting"
05b055e8
FY
2261 * starts happening, stop the previous chrono and start a new one.
2262 */
2263 if (type > tp->chrono_type)
2264 tcp_chrono_set(tp, type);
2265}
2266
2267void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2268{
2269 struct tcp_sock *tp = tcp_sk(sk);
2270
0f87230d
FY
2271
2272 /* There are multiple conditions worthy of tracking in a
2273 * chronograph, so that the highest priority enum takes
2274 * precedence over the other conditions (see tcp_chrono_start).
2275 * If a condition stops, we only stop chrono tracking if
2276 * it's the "most interesting" or current chrono we are
2277 * tracking and starts busy chrono if we have pending data.
2278 */
75c119af 2279 if (tcp_rtx_and_write_queues_empty(sk))
0f87230d
FY
2280 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2281 else if (type == tp->chrono_type)
2282 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
05b055e8
FY
2283}
2284
1da177e4
LT
2285/* This routine writes packets to the network. It advances the
2286 * send_head. This happens as incoming acks open up the remote
2287 * window for us.
2288 *
f8269a49
IJ
2289 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2290 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2291 * account rare use of URG, this is not a big flaw.
2292 *
6ba8a3b1
ND
2293 * Send at most one packet when push_one > 0. Temporarily ignore
2294 * cwnd limit to force at most one packet out when push_one == 2.
2295
a2a385d6
ED
2296 * Returns true, if no segments are in flight and we have queued segments,
2297 * but cannot send anything now because of SWS or another problem.
1da177e4 2298 */
a2a385d6
ED
2299static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2300 int push_one, gfp_t gfp)
1da177e4
LT
2301{
2302 struct tcp_sock *tp = tcp_sk(sk);
92df7b51 2303 struct sk_buff *skb;
c1b4a7e6
DM
2304 unsigned int tso_segs, sent_pkts;
2305 int cwnd_quota;
5d424d5a 2306 int result;
5615f886 2307 bool is_cwnd_limited = false, is_rwnd_limited = false;
605ad7f1 2308 u32 max_segs;
1da177e4 2309
92df7b51 2310 sent_pkts = 0;
5d424d5a 2311
ee1836ae 2312 tcp_mstamp_refresh(tp);
d5dd9175
IJ
2313 if (!push_one) {
2314 /* Do MTU probing. */
2315 result = tcp_mtu_probe(sk);
2316 if (!result) {
a2a385d6 2317 return false;
d5dd9175
IJ
2318 } else if (result > 0) {
2319 sent_pkts = 1;
2320 }
5d424d5a
JH
2321 }
2322
ed6e7268 2323 max_segs = tcp_tso_segs(sk, mss_now);
fe067e8a 2324 while ((skb = tcp_send_head(sk))) {
c8ac3774
HX
2325 unsigned int limit;
2326
218af599
ED
2327 if (tcp_pacing_check(sk))
2328 break;
2329
5bbb432c 2330 tso_segs = tcp_init_tso_segs(skb, mss_now);
c1b4a7e6 2331 BUG_ON(!tso_segs);
aa93466b 2332
9d186cac 2333 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
7faee5c0 2334 /* "skb_mstamp" is used as a start point for the retransmit timer */
ab408b6d 2335 tcp_update_skb_after_send(sk, skb);
ec342325 2336 goto repair; /* Skip network transmission */
9d186cac 2337 }
ec342325 2338
b68e9f85 2339 cwnd_quota = tcp_cwnd_test(tp, skb);
6ba8a3b1
ND
2340 if (!cwnd_quota) {
2341 if (push_one == 2)
2342 /* Force out a loss probe pkt. */
2343 cwnd_quota = 1;
2344 else
2345 break;
2346 }
b68e9f85 2347
5615f886
FY
2348 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2349 is_rwnd_limited = true;
b68e9f85 2350 break;
5615f886 2351 }
b68e9f85 2352
d6a4e26a 2353 if (tso_segs == 1) {
c1b4a7e6
DM
2354 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2355 (tcp_skb_is_last(sk, skb) ?
2356 nonagle : TCP_NAGLE_PUSH))))
2357 break;
2358 } else {
ca8a2263 2359 if (!push_one &&
605ad7f1
ED
2360 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2361 max_segs))
c1b4a7e6
DM
2362 break;
2363 }
aa93466b 2364
605ad7f1 2365 limit = mss_now;
d6a4e26a 2366 if (tso_segs > 1 && !tcp_urg_mode(tp))
605ad7f1
ED
2367 limit = tcp_mss_split_point(sk, skb, mss_now,
2368 min_t(unsigned int,
2369 cwnd_quota,
2370 max_segs),
2371 nonagle);
2372
2373 if (skb->len > limit &&
75c119af
ED
2374 unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
2375 skb, limit, mss_now, gfp)))
605ad7f1
ED
2376 break;
2377
f9616c35
ED
2378 if (tcp_small_queue_check(sk, skb, 0))
2379 break;
c9eeec26 2380
d5dd9175 2381 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
92df7b51 2382 break;
1da177e4 2383
ec342325 2384repair:
92df7b51
DM
2385 /* Advance the send_head. This one is sent out.
2386 * This call will increment packets_out.
2387 */
66f5fe62 2388 tcp_event_new_data_sent(sk, skb);
1da177e4 2389
92df7b51 2390 tcp_minshall_update(tp, mss_now, skb);
a262f0cd 2391 sent_pkts += tcp_skb_pcount(skb);
d5dd9175
IJ
2392
2393 if (push_one)
2394 break;
92df7b51 2395 }
1da177e4 2396
5615f886
FY
2397 if (is_rwnd_limited)
2398 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2399 else
2400 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2401
aa93466b 2402 if (likely(sent_pkts)) {
684bad11
YC
2403 if (tcp_in_cwnd_reduction(sk))
2404 tp->prr_out += sent_pkts;
6ba8a3b1
ND
2405
2406 /* Send one loss probe per tail loss episode. */
2407 if (push_one != 2)
ed66dfaf 2408 tcp_schedule_loss_probe(sk, false);
d2e1339f 2409 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
ca8a2263 2410 tcp_cwnd_validate(sk, is_cwnd_limited);
a2a385d6 2411 return false;
1da177e4 2412 }
75c119af 2413 return !tp->packets_out && !tcp_write_queue_empty(sk);
6ba8a3b1
ND
2414}
2415
ed66dfaf 2416bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
6ba8a3b1
ND
2417{
2418 struct inet_connection_sock *icsk = inet_csk(sk);
2419 struct tcp_sock *tp = tcp_sk(sk);
a2815817 2420 u32 timeout, rto_delta_us;
2ae21cf5 2421 int early_retrans;
6ba8a3b1 2422
6ba8a3b1
ND
2423 /* Don't do any loss probe on a Fast Open connection before 3WHS
2424 * finishes.
2425 */
f9b99582 2426 if (tp->fastopen_rsk)
6ba8a3b1
ND
2427 return false;
2428
2ae21cf5 2429 early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
6ba8a3b1 2430 /* Schedule a loss probe in 2*RTT for SACK capable connections
b4f70c3d 2431 * not in loss recovery, that are either limited by cwnd or application.
6ba8a3b1 2432 */
2ae21cf5 2433 if ((early_retrans != 3 && early_retrans != 4) ||
bec41a11 2434 !tp->packets_out || !tcp_is_sack(tp) ||
b4f70c3d
NC
2435 (icsk->icsk_ca_state != TCP_CA_Open &&
2436 icsk->icsk_ca_state != TCP_CA_CWR))
6ba8a3b1
ND
2437 return false;
2438
bb4d991a 2439 /* Probe timeout is 2*rtt. Add minimum RTO to account
f9b99582
YC
2440 * for delayed ack when there's one outstanding packet. If no RTT
2441 * sample is available then probe after TCP_TIMEOUT_INIT.
6ba8a3b1 2442 */
bb4d991a
YC
2443 if (tp->srtt_us) {
2444 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2445 if (tp->packets_out == 1)
2446 timeout += TCP_RTO_MIN;
2447 else
2448 timeout += TCP_TIMEOUT_MIN;
2449 } else {
2450 timeout = TCP_TIMEOUT_INIT;
2451 }
6ba8a3b1 2452
a2815817 2453 /* If the RTO formula yields an earlier time, then use that time. */
ed66dfaf
NC
2454 rto_delta_us = advancing_rto ?
2455 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2456 tcp_rto_delta_us(sk); /* How far in future is RTO? */
a2815817
NC
2457 if (rto_delta_us > 0)
2458 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
6ba8a3b1
ND
2459
2460 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2461 TCP_RTO_MAX);
2462 return true;
2463}
2464
1f3279ae
ED
2465/* Thanks to skb fast clones, we can detect if a prior transmit of
2466 * a packet is still in a qdisc or driver queue.
2467 * In this case, there is very little point doing a retransmit !
1f3279ae
ED
2468 */
2469static bool skb_still_in_host_queue(const struct sock *sk,
2470 const struct sk_buff *skb)
2471{
39bb5e62 2472 if (unlikely(skb_fclone_busy(sk, skb))) {
c10d9310
ED
2473 NET_INC_STATS(sock_net(sk),
2474 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
1f3279ae
ED
2475 return true;
2476 }
2477 return false;
2478}
2479
b340b264 2480/* When probe timeout (PTO) fires, try send a new segment if possible, else
6ba8a3b1
ND
2481 * retransmit the last segment.
2482 */
2483void tcp_send_loss_probe(struct sock *sk)
2484{
9b717a8d 2485 struct tcp_sock *tp = tcp_sk(sk);
6ba8a3b1
ND
2486 struct sk_buff *skb;
2487 int pcount;
2488 int mss = tcp_current_mss(sk);
6ba8a3b1 2489
b340b264 2490 skb = tcp_send_head(sk);
75c119af
ED
2491 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2492 pcount = tp->packets_out;
2493 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2494 if (tp->packets_out > pcount)
2495 goto probe_sent;
2496 goto rearm_timer;
6ba8a3b1 2497 }
75c119af 2498 skb = skb_rb_last(&sk->tcp_rtx_queue);
6ba8a3b1 2499
9b717a8d
ND
2500 /* At most one outstanding TLP retransmission. */
2501 if (tp->tlp_high_seq)
2502 goto rearm_timer;
2503
6ba8a3b1 2504 /* Retransmit last segment. */
6ba8a3b1
ND
2505 if (WARN_ON(!skb))
2506 goto rearm_timer;
2507
1f3279ae
ED
2508 if (skb_still_in_host_queue(sk, skb))
2509 goto rearm_timer;
2510
6ba8a3b1
ND
2511 pcount = tcp_skb_pcount(skb);
2512 if (WARN_ON(!pcount))
2513 goto rearm_timer;
2514
2515 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
75c119af
ED
2516 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2517 (pcount - 1) * mss, mss,
6cc55e09 2518 GFP_ATOMIC)))
6ba8a3b1 2519 goto rearm_timer;
75c119af 2520 skb = skb_rb_next(skb);
6ba8a3b1
ND
2521 }
2522
2523 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2524 goto rearm_timer;
2525
10d3be56 2526 if (__tcp_retransmit_skb(sk, skb, 1))
b340b264 2527 goto rearm_timer;
6ba8a3b1 2528
9b717a8d 2529 /* Record snd_nxt for loss detection. */
b340b264 2530 tp->tlp_high_seq = tp->snd_nxt;
9b717a8d 2531
b340b264 2532probe_sent:
c10d9310 2533 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
b340b264
YC
2534 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2535 inet_csk(sk)->icsk_pending = 0;
6ba8a3b1 2536rearm_timer:
fcd16c0a 2537 tcp_rearm_rto(sk);
1da177e4
LT
2538}
2539
a762a980
DM
2540/* Push out any pending frames which were held back due to
2541 * TCP_CORK or attempt at coalescing tiny packets.
2542 * The socket must be locked by the caller.
2543 */
9e412ba7
IJ
2544void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2545 int nonagle)
a762a980 2546{
726e07a8
IJ
2547 /* If we are closed, the bytes will have to remain here.
2548 * In time closedown will finish, we empty the write queue and
2549 * all will be happy.
2550 */
2551 if (unlikely(sk->sk_state == TCP_CLOSE))
2552 return;
2553
99a1dec7 2554 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
7450aaf6 2555 sk_gfp_mask(sk, GFP_ATOMIC)))
726e07a8 2556 tcp_check_probe_timer(sk);
a762a980
DM
2557}
2558
c1b4a7e6
DM
2559/* Send _single_ skb sitting at the send head. This function requires
2560 * true push pending frames to setup probe timer etc.
2561 */
2562void tcp_push_one(struct sock *sk, unsigned int mss_now)
2563{
fe067e8a 2564 struct sk_buff *skb = tcp_send_head(sk);
c1b4a7e6
DM
2565
2566 BUG_ON(!skb || skb->len < mss_now);
2567
d5dd9175 2568 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
c1b4a7e6
DM
2569}
2570
1da177e4
LT
2571/* This function returns the amount that we can raise the
2572 * usable window based on the following constraints
e905a9ed 2573 *
1da177e4
LT
2574 * 1. The window can never be shrunk once it is offered (RFC 793)
2575 * 2. We limit memory per socket
2576 *
2577 * RFC 1122:
2578 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2579 * RECV.NEXT + RCV.WIN fixed until:
2580 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2581 *
2582 * i.e. don't raise the right edge of the window until you can raise
2583 * it at least MSS bytes.
2584 *
2585 * Unfortunately, the recommended algorithm breaks header prediction,
2586 * since header prediction assumes th->window stays fixed.
2587 *
2588 * Strictly speaking, keeping th->window fixed violates the receiver
2589 * side SWS prevention criteria. The problem is that under this rule
2590 * a stream of single byte packets will cause the right side of the
2591 * window to always advance by a single byte.
e905a9ed 2592 *
1da177e4
LT
2593 * Of course, if the sender implements sender side SWS prevention
2594 * then this will not be a problem.
e905a9ed 2595 *
1da177e4 2596 * BSD seems to make the following compromise:
e905a9ed 2597 *
1da177e4
LT
2598 * If the free space is less than the 1/4 of the maximum
2599 * space available and the free space is less than 1/2 mss,
2600 * then set the window to 0.
2601 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2602 * Otherwise, just prevent the window from shrinking
2603 * and from being larger than the largest representable value.
2604 *
2605 * This prevents incremental opening of the window in the regime
2606 * where TCP is limited by the speed of the reader side taking
2607 * data out of the TCP receive queue. It does nothing about
2608 * those cases where the window is constrained on the sender side
2609 * because the pipeline is full.
2610 *
2611 * BSD also seems to "accidentally" limit itself to windows that are a
2612 * multiple of MSS, at least until the free space gets quite small.
2613 * This would appear to be a side effect of the mbuf implementation.
2614 * Combining these two algorithms results in the observed behavior
2615 * of having a fixed window size at almost all times.
2616 *
2617 * Below we obtain similar behavior by forcing the offered window to
2618 * a multiple of the mss when it is feasible to do so.
2619 *
2620 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2621 * Regular options like TIMESTAMP are taken into account.
2622 */
2623u32 __tcp_select_window(struct sock *sk)
2624{
463c84b9 2625 struct inet_connection_sock *icsk = inet_csk(sk);
1da177e4 2626 struct tcp_sock *tp = tcp_sk(sk);
caa20d9a 2627 /* MSS for the peer's data. Previous versions used mss_clamp
1da177e4
LT
2628 * here. I don't know if the value based on our guesses
2629 * of peer's MSS is better for the performance. It's more correct
2630 * but may be worse for the performance because of rcv_mss
2631 * fluctuations. --SAW 1998/11/1
2632 */
463c84b9 2633 int mss = icsk->icsk_ack.rcv_mss;
1da177e4 2634 int free_space = tcp_space(sk);
86c1a045
FW
2635 int allowed_space = tcp_full_space(sk);
2636 int full_space = min_t(int, tp->window_clamp, allowed_space);
1da177e4
LT
2637 int window;
2638
06425c30 2639 if (unlikely(mss > full_space)) {
e905a9ed 2640 mss = full_space;
06425c30
ED
2641 if (mss <= 0)
2642 return 0;
2643 }
b92edbe0 2644 if (free_space < (full_space >> 1)) {
463c84b9 2645 icsk->icsk_ack.quick = 0;
1da177e4 2646
b8da51eb 2647 if (tcp_under_memory_pressure(sk))
056834d9
IJ
2648 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2649 4U * tp->advmss);
1da177e4 2650
86c1a045
FW
2651 /* free_space might become our new window, make sure we don't
2652 * increase it due to wscale.
2653 */
2654 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2655
2656 /* if free space is less than mss estimate, or is below 1/16th
2657 * of the maximum allowed, try to move to zero-window, else
2658 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2659 * new incoming data is dropped due to memory limits.
2660 * With large window, mss test triggers way too late in order
2661 * to announce zero window in time before rmem limit kicks in.
2662 */
2663 if (free_space < (allowed_space >> 4) || free_space < mss)
1da177e4
LT
2664 return 0;
2665 }
2666
2667 if (free_space > tp->rcv_ssthresh)
2668 free_space = tp->rcv_ssthresh;
2669
2670 /* Don't do rounding if we are using window scaling, since the
2671 * scaled window will not line up with the MSS boundary anyway.
2672 */
1da177e4
LT
2673 if (tp->rx_opt.rcv_wscale) {
2674 window = free_space;
2675
2676 /* Advertise enough space so that it won't get scaled away.
2677 * Import case: prevent zero window announcement if
2678 * 1<<rcv_wscale > mss.
2679 */
1935299d 2680 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
1da177e4 2681 } else {
1935299d 2682 window = tp->rcv_wnd;
1da177e4
LT
2683 /* Get the largest window that is a nice multiple of mss.
2684 * Window clamp already applied above.
2685 * If our current window offering is within 1 mss of the
2686 * free space we just keep it. This prevents the divide
2687 * and multiply from happening most of the time.
2688 * We also don't do any window rounding when the free space
2689 * is too small.
2690 */
2691 if (window <= free_space - mss || window > free_space)
1935299d 2692 window = rounddown(free_space, mss);
84565070 2693 else if (mss == full_space &&
b92edbe0 2694 free_space > window + (full_space >> 1))
84565070 2695 window = free_space;
1da177e4
LT
2696 }
2697
2698 return window;
2699}
2700
cfea5a68
MKL
2701void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2702 const struct sk_buff *next_skb)
082ac2d5 2703{
0a2cf20c
SHY
2704 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2705 const struct skb_shared_info *next_shinfo =
2706 skb_shinfo(next_skb);
082ac2d5
MKL
2707 struct skb_shared_info *shinfo = skb_shinfo(skb);
2708
0a2cf20c 2709 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
082ac2d5 2710 shinfo->tskey = next_shinfo->tskey;
2de8023e
MKL
2711 TCP_SKB_CB(skb)->txstamp_ack |=
2712 TCP_SKB_CB(next_skb)->txstamp_ack;
082ac2d5
MKL
2713 }
2714}
2715
4a17fc3a 2716/* Collapses two adjacent SKB's during retransmission. */
f8071cde 2717static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
1da177e4
LT
2718{
2719 struct tcp_sock *tp = tcp_sk(sk);
75c119af 2720 struct sk_buff *next_skb = skb_rb_next(skb);
13dde04f 2721 int next_skb_size;
1da177e4 2722
058dc334 2723 next_skb_size = next_skb->len;
1da177e4 2724
058dc334 2725 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
a6963a6b 2726
f8071cde
ED
2727 if (next_skb_size) {
2728 if (next_skb_size <= skb_availroom(skb))
2729 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2730 next_skb_size);
2731 else if (!skb_shift(skb, next_skb, next_skb_size))
2732 return false;
2733 }
2b7cda9c 2734 tcp_highest_sack_replace(sk, next_skb, skb);
1da177e4 2735
058dc334
IJ
2736 /* Update sequence range on original skb. */
2737 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1da177e4 2738
e6c7d085 2739 /* Merge over control information. This moves PSH/FIN etc. over */
4de075e0 2740 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
058dc334
IJ
2741
2742 /* All done, get rid of second SKB and account for it so
2743 * packet counting does not break.
2744 */
2745 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
a643b5d4 2746 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
058dc334
IJ
2747
2748 /* changed transmit queue under us so clear hints */
ef9da47c
IJ
2749 tcp_clear_retrans_hints_partial(tp);
2750 if (next_skb == tp->retransmit_skb_hint)
2751 tp->retransmit_skb_hint = skb;
058dc334 2752
797108d1
IJ
2753 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2754
082ac2d5
MKL
2755 tcp_skb_collapse_tstamp(skb, next_skb);
2756
75c119af 2757 tcp_rtx_queue_unlink_and_free(next_skb, sk);
f8071cde 2758 return true;
1da177e4
LT
2759}
2760
67edfef7 2761/* Check if coalescing SKBs is legal. */
a2a385d6 2762static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
4a17fc3a
IJ
2763{
2764 if (tcp_skb_pcount(skb) > 1)
a2a385d6 2765 return false;
4a17fc3a 2766 if (skb_cloned(skb))
a2a385d6 2767 return false;
2331ccc5 2768 /* Some heuristics for collapsing over SACK'd could be invented */
4a17fc3a 2769 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
a2a385d6 2770 return false;
4a17fc3a 2771
a2a385d6 2772 return true;
4a17fc3a
IJ
2773}
2774
67edfef7
AK
2775/* Collapse packets in the retransmit queue to make to create
2776 * less packets on the wire. This is only done on retransmission.
2777 */
4a17fc3a
IJ
2778static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2779 int space)
2780{
2781 struct tcp_sock *tp = tcp_sk(sk);
2782 struct sk_buff *skb = to, *tmp;
a2a385d6 2783 bool first = true;
4a17fc3a 2784
e0a1e5b5 2785 if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
4a17fc3a 2786 return;
4de075e0 2787 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
4a17fc3a
IJ
2788 return;
2789
75c119af 2790 skb_rbtree_walk_from_safe(skb, tmp) {
4a17fc3a
IJ
2791 if (!tcp_can_collapse(sk, skb))
2792 break;
2793
a643b5d4
MKL
2794 if (!tcp_skb_can_collapse_to(to))
2795 break;
2796
4a17fc3a
IJ
2797 space -= skb->len;
2798
2799 if (first) {
a2a385d6 2800 first = false;
4a17fc3a
IJ
2801 continue;
2802 }
2803
2804 if (space < 0)
2805 break;
4a17fc3a
IJ
2806
2807 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2808 break;
2809
f8071cde
ED
2810 if (!tcp_collapse_retrans(sk, to))
2811 break;
4a17fc3a
IJ
2812 }
2813}
2814
1da177e4
LT
2815/* This retransmits one SKB. Policy decisions and retransmit queue
2816 * state updates are done by the caller. Returns non-zero if an
2817 * error occurred which prevented the send.
2818 */
10d3be56 2819int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
1da177e4 2820{
5d424d5a 2821 struct inet_connection_sock *icsk = inet_csk(sk);
10d3be56 2822 struct tcp_sock *tp = tcp_sk(sk);
7d227cd2 2823 unsigned int cur_mss;
10d3be56
ED
2824 int diff, len, err;
2825
1da177e4 2826
10d3be56
ED
2827 /* Inconclusive MTU probe */
2828 if (icsk->icsk_mtup.probe_size)
5d424d5a 2829 icsk->icsk_mtup.probe_size = 0;
5d424d5a 2830
1da177e4 2831 /* Do not sent more than we queued. 1/4 is reserved for possible
caa20d9a 2832 * copying overhead: fragmentation, tunneling, mangling etc.
1da177e4 2833 */
14afee4b 2834 if (refcount_read(&sk->sk_wmem_alloc) >
ffb4d6c8
ED
2835 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2836 sk->sk_sndbuf))
1da177e4
LT
2837 return -EAGAIN;
2838
1f3279ae
ED
2839 if (skb_still_in_host_queue(sk, skb))
2840 return -EBUSY;
2841
1da177e4 2842 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
7f582b24
ED
2843 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
2844 WARN_ON_ONCE(1);
2845 return -EINVAL;
2846 }
1da177e4
LT
2847 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2848 return -ENOMEM;
2849 }
2850
7d227cd2
SS
2851 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2852 return -EHOSTUNREACH; /* Routing failure or similar. */
2853
0c54b85f 2854 cur_mss = tcp_current_mss(sk);
7d227cd2 2855
1da177e4
LT
2856 /* If receiver has shrunk his window, and skb is out of
2857 * new window, do not retransmit it. The exception is the
2858 * case, when window is shrunk to zero. In this case
2859 * our retransmit serves as a zero window probe.
2860 */
9d4fb27d
JP
2861 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2862 TCP_SKB_CB(skb)->seq != tp->snd_una)
1da177e4
LT
2863 return -EAGAIN;
2864
10d3be56
ED
2865 len = cur_mss * segs;
2866 if (skb->len > len) {
75c119af
ED
2867 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
2868 cur_mss, GFP_ATOMIC))
1da177e4 2869 return -ENOMEM; /* We'll try again later. */
02276f3c 2870 } else {
10d3be56
ED
2871 if (skb_unclone(skb, GFP_ATOMIC))
2872 return -ENOMEM;
9eb9362e 2873
10d3be56
ED
2874 diff = tcp_skb_pcount(skb);
2875 tcp_set_skb_tso_segs(skb, cur_mss);
2876 diff -= tcp_skb_pcount(skb);
2877 if (diff)
2878 tcp_adjust_pcount(sk, skb, diff);
2879 if (skb->len < cur_mss)
2880 tcp_retrans_try_collapse(sk, skb, cur_mss);
1da177e4
LT
2881 }
2882
49213555
DB
2883 /* RFC3168, section 6.1.1.1. ECN fallback */
2884 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2885 tcp_ecn_clear_syn(sk, skb);
2886
678550c6
YC
2887 /* Update global and local TCP statistics. */
2888 segs = tcp_skb_pcount(skb);
2889 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2890 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2891 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2892 tp->total_retrans += segs;
fb31c9b9 2893 tp->bytes_retrans += skb->len;
678550c6 2894
50bceae9
TG
2895 /* make sure skb->data is aligned on arches that require it
2896 * and check if ack-trimming & collapsing extended the headroom
2897 * beyond what csum_start can cover.
2898 */
2899 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2900 skb_headroom(skb) >= 0xFFFF)) {
10a81980
ED
2901 struct sk_buff *nskb;
2902
e2080072
ED
2903 tcp_skb_tsorted_save(skb) {
2904 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2905 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2906 -ENOBUFS;
2907 } tcp_skb_tsorted_restore(skb);
2908
5889e2c0 2909 if (!err) {
ab408b6d 2910 tcp_update_skb_after_send(sk, skb);
5889e2c0
YS
2911 tcp_rate_skb_sent(sk, skb);
2912 }
117632e6 2913 } else {
c84a5711 2914 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
117632e6 2915 }
c84a5711 2916
a31ad29e
LB
2917 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
2918 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
2919 TCP_SKB_CB(skb)->seq, segs, err);
2920
fc9f3501 2921 if (likely(!err)) {
c84a5711 2922 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
e086101b 2923 trace_tcp_retransmit_skb(sk, skb);
678550c6
YC
2924 } else if (err != -EBUSY) {
2925 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
fc9f3501 2926 }
c84a5711 2927 return err;
93b174ad
YC
2928}
2929
10d3be56 2930int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
93b174ad
YC
2931{
2932 struct tcp_sock *tp = tcp_sk(sk);
10d3be56 2933 int err = __tcp_retransmit_skb(sk, skb, segs);
1da177e4
LT
2934
2935 if (err == 0) {
1da177e4 2936#if FASTRETRANS_DEBUG > 0
056834d9 2937 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
e87cc472 2938 net_dbg_ratelimited("retrans_out leaked\n");
1da177e4
LT
2939 }
2940#endif
2941 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2942 tp->retrans_out += tcp_skb_pcount(skb);
2943
2944 /* Save stamp of the first retransmit. */
2945 if (!tp->retrans_stamp)
7faee5c0 2946 tp->retrans_stamp = tcp_skb_timestamp(skb);
1da177e4 2947
1da177e4 2948 }
6e08d5e3
YC
2949
2950 if (tp->undo_retrans < 0)
2951 tp->undo_retrans = 0;
2952 tp->undo_retrans += tcp_skb_pcount(skb);
1da177e4
LT
2953 return err;
2954}
2955
2956/* This gets called after a retransmit timeout, and the initially
2957 * retransmitted data is acknowledged. It tries to continue
2958 * resending the rest of the retransmit queue, until either
2959 * we've sent it all or the congestion window limit is reached.
1da177e4
LT
2960 */
2961void tcp_xmit_retransmit_queue(struct sock *sk)
2962{
6687e988 2963 const struct inet_connection_sock *icsk = inet_csk(sk);
b9f1f1ce 2964 struct sk_buff *skb, *rtx_head, *hole = NULL;
1da177e4 2965 struct tcp_sock *tp = tcp_sk(sk);
840a3cbe 2966 u32 max_segs;
61eb55f4 2967 int mib_idx;
6a438bbe 2968
45e77d31
IJ
2969 if (!tp->packets_out)
2970 return;
2971
b9f1f1ce
ED
2972 rtx_head = tcp_rtx_queue_head(sk);
2973 skb = tp->retransmit_skb_hint ?: rtx_head;
ed6e7268 2974 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
75c119af 2975 skb_rbtree_walk_from(skb) {
dca0aaf8 2976 __u8 sacked;
10d3be56 2977 int segs;
1da177e4 2978
218af599
ED
2979 if (tcp_pacing_check(sk))
2980 break;
2981
08ebd172 2982 /* we could do better than to assign each time */
51456b29 2983 if (!hole)
0e1c54c2 2984 tp->retransmit_skb_hint = skb;
08ebd172 2985
10d3be56
ED
2986 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
2987 if (segs <= 0)
08ebd172 2988 return;
dca0aaf8 2989 sacked = TCP_SKB_CB(skb)->sacked;
a3d2e9f8
ED
2990 /* In case tcp_shift_skb_data() have aggregated large skbs,
2991 * we need to make sure not sending too bigs TSO packets
2992 */
2993 segs = min_t(int, segs, max_segs);
1da177e4 2994
840a3cbe
YC
2995 if (tp->retrans_out >= tp->lost_out) {
2996 break;
0e1c54c2 2997 } else if (!(sacked & TCPCB_LOST)) {
51456b29 2998 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
0e1c54c2
IJ
2999 hole = skb;
3000 continue;
1da177e4 3001
0e1c54c2
IJ
3002 } else {
3003 if (icsk->icsk_ca_state != TCP_CA_Loss)
3004 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3005 else
3006 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3007 }
1da177e4 3008
0e1c54c2 3009 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
1da177e4
LT
3010 continue;
3011
f9616c35
ED
3012 if (tcp_small_queue_check(sk, skb, 1))
3013 return;
3014
10d3be56 3015 if (tcp_retransmit_skb(sk, skb, segs))
0e1c54c2 3016 return;
24ab6bec 3017
de1d6578 3018 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
1da177e4 3019
684bad11 3020 if (tcp_in_cwnd_reduction(sk))
a262f0cd
ND
3021 tp->prr_out += tcp_skb_pcount(skb);
3022
75c119af 3023 if (skb == rtx_head &&
57dde7f7 3024 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3f421baa
ACM
3025 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3026 inet_csk(sk)->icsk_rto,
3027 TCP_RTO_MAX);
1da177e4
LT
3028 }
3029}
3030
d83769a5
ED
3031/* We allow to exceed memory limits for FIN packets to expedite
3032 * connection tear down and (memory) recovery.
845704a5
ED
3033 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3034 * or even be forced to close flow without any FIN.
a6c5ea4c
ED
3035 * In general, we want to allow one skb per socket to avoid hangs
3036 * with edge trigger epoll()
d83769a5 3037 */
a6c5ea4c 3038void sk_forced_mem_schedule(struct sock *sk, int size)
d83769a5 3039{
e805605c 3040 int amt;
d83769a5
ED
3041
3042 if (size <= sk->sk_forward_alloc)
3043 return;
3044 amt = sk_mem_pages(size);
3045 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
e805605c
JW
3046 sk_memory_allocated_add(sk, amt);
3047
baac50bb
JW
3048 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3049 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
d83769a5
ED
3050}
3051
845704a5
ED
3052/* Send a FIN. The caller locks the socket for us.
3053 * We should try to send a FIN packet really hard, but eventually give up.
1da177e4
LT
3054 */
3055void tcp_send_fin(struct sock *sk)
3056{
845704a5 3057 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
e905a9ed 3058 struct tcp_sock *tp = tcp_sk(sk);
e905a9ed 3059
845704a5
ED
3060 /* Optimization, tack on the FIN if we have one skb in write queue and
3061 * this skb was not yet sent, or we are under memory pressure.
3062 * Note: in the latter case, FIN packet will be sent after a timeout,
3063 * as TCP stack thinks it has already been transmitted.
1da177e4 3064 */
75c119af
ED
3065 if (!tskb && tcp_under_memory_pressure(sk))
3066 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3067
3068 if (tskb) {
845704a5
ED
3069coalesce:
3070 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3071 TCP_SKB_CB(tskb)->end_seq++;
1da177e4 3072 tp->write_seq++;
75c119af 3073 if (tcp_write_queue_empty(sk)) {
845704a5
ED
3074 /* This means tskb was already sent.
3075 * Pretend we included the FIN on previous transmit.
3076 * We need to set tp->snd_nxt to the value it would have
3077 * if FIN had been sent. This is because retransmit path
3078 * does not change tp->snd_nxt.
3079 */
3080 tp->snd_nxt++;
3081 return;
3082 }
1da177e4 3083 } else {
845704a5
ED
3084 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3085 if (unlikely(!skb)) {
3086 if (tskb)
3087 goto coalesce;
3088 return;
1da177e4 3089 }
e2080072 3090 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
d83769a5 3091 skb_reserve(skb, MAX_TCP_HEADER);
a6c5ea4c 3092 sk_forced_mem_schedule(sk, skb->truesize);
1da177e4 3093 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
e870a8ef 3094 tcp_init_nondata_skb(skb, tp->write_seq,
a3433f35 3095 TCPHDR_ACK | TCPHDR_FIN);
1da177e4
LT
3096 tcp_queue_skb(sk, skb);
3097 }
845704a5 3098 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
1da177e4
LT
3099}
3100
3101/* We get here when a process closes a file descriptor (either due to
3102 * an explicit close() or as a byproduct of exit()'ing) and there
3103 * was unread data in the receive queue. This behavior is recommended
65bb723c 3104 * by RFC 2525, section 2.17. -DaveM
1da177e4 3105 */
dd0fc66f 3106void tcp_send_active_reset(struct sock *sk, gfp_t priority)
1da177e4 3107{
1da177e4
LT
3108 struct sk_buff *skb;
3109
7cc2b043
GF
3110 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3111
1da177e4
LT
3112 /* NOTE: No TCP options attached and we never retransmit this. */
3113 skb = alloc_skb(MAX_TCP_HEADER, priority);
3114 if (!skb) {
4e673444 3115 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
1da177e4
LT
3116 return;
3117 }
3118
3119 /* Reserve space for headers and prepare control bits. */
3120 skb_reserve(skb, MAX_TCP_HEADER);
e870a8ef 3121 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
a3433f35 3122 TCPHDR_ACK | TCPHDR_RST);
9a568de4 3123 tcp_mstamp_refresh(tcp_sk(sk));
1da177e4 3124 /* Send it off. */
dfb4b9dc 3125 if (tcp_transmit_skb(sk, skb, 0, priority))
4e673444 3126 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
c24b14c4
SL
3127
3128 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3129 * skb here is different to the troublesome skb, so use NULL
3130 */
3131 trace_tcp_send_reset(sk, NULL);
1da177e4
LT
3132}
3133
67edfef7
AK
3134/* Send a crossed SYN-ACK during socket establishment.
3135 * WARNING: This routine must only be called when we have already sent
1da177e4
LT
3136 * a SYN packet that crossed the incoming SYN that caused this routine
3137 * to get called. If this assumption fails then the initial rcv_wnd
3138 * and rcv_wscale values will not be correct.
3139 */
3140int tcp_send_synack(struct sock *sk)
3141{
056834d9 3142 struct sk_buff *skb;
1da177e4 3143
75c119af 3144 skb = tcp_rtx_queue_head(sk);
51456b29 3145 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
75c119af 3146 pr_err("%s: wrong queue state\n", __func__);
1da177e4
LT
3147 return -EFAULT;
3148 }
4de075e0 3149 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
1da177e4 3150 if (skb_cloned(skb)) {
e2080072
ED
3151 struct sk_buff *nskb;
3152
3153 tcp_skb_tsorted_save(skb) {
3154 nskb = skb_copy(skb, GFP_ATOMIC);
3155 } tcp_skb_tsorted_restore(skb);
51456b29 3156 if (!nskb)
1da177e4 3157 return -ENOMEM;
e2080072 3158 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
75c119af 3159 tcp_rtx_queue_unlink_and_free(skb, sk);
f4a775d1 3160 __skb_header_release(nskb);
75c119af 3161 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3ab224be
HA
3162 sk->sk_wmem_queued += nskb->truesize;
3163 sk_mem_charge(sk, nskb->truesize);
1da177e4
LT
3164 skb = nskb;
3165 }
3166
4de075e0 3167 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
735d3831 3168 tcp_ecn_send_synack(sk, skb);
1da177e4 3169 }
dfb4b9dc 3170 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
1da177e4
LT
3171}
3172
4aea39c1
ED
3173/**
3174 * tcp_make_synack - Prepare a SYN-ACK.
3175 * sk: listener socket
3176 * dst: dst entry attached to the SYNACK
3177 * req: request_sock pointer
4aea39c1
ED
3178 *
3179 * Allocate one skb and build a SYNACK packet.
3180 * @dst is consumed : Caller should not use it again.
3181 */
5d062de7 3182struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
e6b4d113 3183 struct request_sock *req,
ca6fb065 3184 struct tcp_fastopen_cookie *foc,
b3d05147 3185 enum tcp_synack_type synack_type)
1da177e4 3186{
2e6599cb 3187 struct inet_request_sock *ireq = inet_rsk(req);
5d062de7 3188 const struct tcp_sock *tp = tcp_sk(sk);
80f03e27 3189 struct tcp_md5sig_key *md5 = NULL;
5d062de7
ED
3190 struct tcp_out_options opts;
3191 struct sk_buff *skb;
bd0388ae 3192 int tcp_header_size;
5d062de7 3193 struct tcphdr *th;
f5fff5dc 3194 int mss;
1da177e4 3195
ca6fb065 3196 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
4aea39c1
ED
3197 if (unlikely(!skb)) {
3198 dst_release(dst);
1da177e4 3199 return NULL;
4aea39c1 3200 }
1da177e4
LT
3201 /* Reserve space for headers. */
3202 skb_reserve(skb, MAX_TCP_HEADER);
3203
b3d05147
ED
3204 switch (synack_type) {
3205 case TCP_SYNACK_NORMAL:
9e17f8a4 3206 skb_set_owner_w(skb, req_to_sk(req));
b3d05147
ED
3207 break;
3208 case TCP_SYNACK_COOKIE:
3209 /* Under synflood, we do not attach skb to a socket,
3210 * to avoid false sharing.
3211 */
3212 break;
3213 case TCP_SYNACK_FASTOPEN:
ca6fb065
ED
3214 /* sk is a const pointer, because we want to express multiple
3215 * cpu might call us concurrently.
3216 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3217 */
3218 skb_set_owner_w(skb, (struct sock *)sk);
b3d05147 3219 break;
ca6fb065 3220 }
4aea39c1 3221 skb_dst_set(skb, dst);
1da177e4 3222
3541f9e8 3223 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
f5fff5dc 3224
33ad798c 3225 memset(&opts, 0, sizeof(opts));
8b5f12d0
FW
3226#ifdef CONFIG_SYN_COOKIES
3227 if (unlikely(req->cookie_ts))
d3edd06e 3228 skb->skb_mstamp_ns = cookie_init_timestamp(req);
8b5f12d0
FW
3229 else
3230#endif
d3edd06e 3231 skb->skb_mstamp_ns = tcp_clock_ns();
80f03e27
ED
3232
3233#ifdef CONFIG_TCP_MD5SIG
3234 rcu_read_lock();
fd3a154a 3235 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
80f03e27 3236#endif
58d607d3 3237 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
60e2a778
UB
3238 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3239 foc) + sizeof(*th);
cfb6eeb4 3240
aa8223c7
ACM
3241 skb_push(skb, tcp_header_size);
3242 skb_reset_transport_header(skb);
1da177e4 3243
ea1627c2 3244 th = (struct tcphdr *)skb->data;
1da177e4
LT
3245 memset(th, 0, sizeof(struct tcphdr));
3246 th->syn = 1;
3247 th->ack = 1;
6ac705b1 3248 tcp_ecn_make_synack(req, th);
b44084c2 3249 th->source = htons(ireq->ir_num);
634fb979 3250 th->dest = ireq->ir_rmt_port;
e05a90ec 3251 skb->mark = ireq->ir_mark;
3b117750
ED
3252 skb->ip_summed = CHECKSUM_PARTIAL;
3253 th->seq = htonl(tcp_rsk(req)->snt_isn);
8336886f
JC
3254 /* XXX data is queued and acked as is. No buffer/window check */
3255 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
1da177e4
LT
3256
3257 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
ed53d0ab 3258 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
5d062de7 3259 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
1da177e4 3260 th->doff = (tcp_header_size >> 2);
90bbcc60 3261 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
cfb6eeb4
YH
3262
3263#ifdef CONFIG_TCP_MD5SIG
3264 /* Okay, we have all we need - do the md5 hash if needed */
80f03e27 3265 if (md5)
bd0388ae 3266 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
39f8e58e 3267 md5, req_to_sk(req), skb);
80f03e27 3268 rcu_read_unlock();
cfb6eeb4
YH
3269#endif
3270
b50edd78 3271 /* Do not fool tcpdump (if any), clean our debris */
2456e855 3272 skb->tstamp = 0;
1da177e4
LT
3273 return skb;
3274}
4bc2f18b 3275EXPORT_SYMBOL(tcp_make_synack);
1da177e4 3276
81164413
DB
3277static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3278{
3279 struct inet_connection_sock *icsk = inet_csk(sk);
3280 const struct tcp_congestion_ops *ca;
3281 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3282
3283 if (ca_key == TCP_CA_UNSPEC)
3284 return;
3285
3286 rcu_read_lock();
3287 ca = tcp_ca_find_key(ca_key);
3288 if (likely(ca && try_module_get(ca->owner))) {
3289 module_put(icsk->icsk_ca_ops->owner);
3290 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3291 icsk->icsk_ca_ops = ca;
3292 }
3293 rcu_read_unlock();
3294}
3295
67edfef7 3296/* Do all connect socket setups that can be done AF independent. */
f7e56a76 3297static void tcp_connect_init(struct sock *sk)
1da177e4 3298{
cf533ea5 3299 const struct dst_entry *dst = __sk_dst_get(sk);
1da177e4
LT
3300 struct tcp_sock *tp = tcp_sk(sk);
3301 __u8 rcv_wscale;
13d3b1eb 3302 u32 rcv_wnd;
1da177e4
LT
3303
3304 /* We'll fix this up when we get a response from the other end.
3305 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3306 */
5d2ed052
ED
3307 tp->tcp_header_len = sizeof(struct tcphdr);
3308 if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3309 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
1da177e4 3310
cfb6eeb4 3311#ifdef CONFIG_TCP_MD5SIG
00db4124 3312 if (tp->af_specific->md5_lookup(sk, sk))
cfb6eeb4
YH
3313 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3314#endif
3315
1da177e4
LT
3316 /* If user gave his TCP_MAXSEG, record it to clamp */
3317 if (tp->rx_opt.user_mss)
3318 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3319 tp->max_window = 0;
5d424d5a 3320 tcp_mtup_init(sk);
1da177e4
LT
3321 tcp_sync_mss(sk, dst_mtu(dst));
3322
81164413
DB
3323 tcp_ca_dst_init(sk, dst);
3324
1da177e4
LT
3325 if (!tp->window_clamp)
3326 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3541f9e8 3327 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
f5fff5dc 3328
1da177e4 3329 tcp_initialize_rcv_mss(sk);
1da177e4 3330
e88c64f0
HPP
3331 /* limit the window selection if the user enforce a smaller rx buffer */
3332 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3333 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3334 tp->window_clamp = tcp_full_space(sk);
3335
13d3b1eb
LB
3336 rcv_wnd = tcp_rwnd_init_bpf(sk);
3337 if (rcv_wnd == 0)
3338 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3339
ceef9ab6 3340 tcp_select_initial_window(sk, tcp_full_space(sk),
1da177e4
LT
3341 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3342 &tp->rcv_wnd,
3343 &tp->window_clamp,
9bb37ef0 3344 sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
31d12926 3345 &rcv_wscale,
13d3b1eb 3346 rcv_wnd);
1da177e4
LT
3347
3348 tp->rx_opt.rcv_wscale = rcv_wscale;
3349 tp->rcv_ssthresh = tp->rcv_wnd;
3350
3351 sk->sk_err = 0;
3352 sock_reset_flag(sk, SOCK_DONE);
3353 tp->snd_wnd = 0;
ee7537b6 3354 tcp_init_wl(tp, 0);
7f582b24 3355 tcp_write_queue_purge(sk);
1da177e4
LT
3356 tp->snd_una = tp->write_seq;
3357 tp->snd_sml = tp->write_seq;
33f5f57e 3358 tp->snd_up = tp->write_seq;
370816ae 3359 tp->snd_nxt = tp->write_seq;
ee995283
PE
3360
3361 if (likely(!tp->repair))
3362 tp->rcv_nxt = 0;
c7781a6e 3363 else
70eabf0e 3364 tp->rcv_tstamp = tcp_jiffies32;
ee995283
PE
3365 tp->rcv_wup = tp->rcv_nxt;
3366 tp->copied_seq = tp->rcv_nxt;
1da177e4 3367
8550f328 3368 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
463c84b9 3369 inet_csk(sk)->icsk_retransmits = 0;
1da177e4
LT
3370 tcp_clear_retrans(tp);
3371}
3372
783237e8
YC
3373static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3374{
3375 struct tcp_sock *tp = tcp_sk(sk);
3376 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3377
3378 tcb->end_seq += skb->len;
f4a775d1 3379 __skb_header_release(skb);
783237e8
YC
3380 sk->sk_wmem_queued += skb->truesize;
3381 sk_mem_charge(sk, skb->truesize);
3382 tp->write_seq = tcb->end_seq;
3383 tp->packets_out += tcp_skb_pcount(skb);
3384}
3385
3386/* Build and send a SYN with data and (cached) Fast Open cookie. However,
3387 * queue a data-only packet after the regular SYN, such that regular SYNs
3388 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3389 * only the SYN sequence, the data are retransmitted in the first ACK.
3390 * If cookie is not cached or other error occurs, falls back to send a
3391 * regular SYN with Fast Open cookie request option.
3392 */
3393static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3394{
3395 struct tcp_sock *tp = tcp_sk(sk);
3396 struct tcp_fastopen_request *fo = tp->fastopen_req;
065263f4 3397 int space, err = 0;
355a901e 3398 struct sk_buff *syn_data;
aab48743 3399
67da22d2 3400 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
065263f4 3401 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
783237e8
YC
3402 goto fallback;
3403
3404 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3405 * user-MSS. Reserve maximum option space for middleboxes that add
3406 * private TCP options. The cost is reduced data space in SYN :(
3407 */
3541f9e8
ED
3408 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3409
1b63edd6 3410 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
783237e8
YC
3411 MAX_TCP_OPTION_SPACE;
3412
f5ddcbbb
ED
3413 space = min_t(size_t, space, fo->size);
3414
3415 /* limit to order-0 allocations */
3416 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3417
eb934478 3418 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
355a901e 3419 if (!syn_data)
783237e8 3420 goto fallback;
355a901e
ED
3421 syn_data->ip_summed = CHECKSUM_PARTIAL;
3422 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
07e100f9
ED
3423 if (space) {
3424 int copied = copy_from_iter(skb_put(syn_data, space), space,
3425 &fo->data->msg_iter);
3426 if (unlikely(!copied)) {
ba233b34 3427 tcp_skb_tsorted_anchor_cleanup(syn_data);
07e100f9
ED
3428 kfree_skb(syn_data);
3429 goto fallback;
3430 }
3431 if (copied != space) {
3432 skb_trim(syn_data, copied);
3433 space = copied;
3434 }
57be5bda 3435 }
355a901e
ED
3436 /* No more data pending in inet_wait_for_connect() */
3437 if (space == fo->size)
3438 fo->data = NULL;
3439 fo->copied = space;
783237e8 3440
355a901e 3441 tcp_connect_queue_skb(sk, syn_data);
0f87230d
FY
3442 if (syn_data->len)
3443 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
783237e8 3444
355a901e 3445 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
783237e8 3446
d3edd06e 3447 syn->skb_mstamp_ns = syn_data->skb_mstamp_ns;
431a9124 3448
355a901e
ED
3449 /* Now full SYN+DATA was cloned and sent (or not),
3450 * remove the SYN from the original skb (syn_data)
3451 * we keep in write queue in case of a retransmit, as we
3452 * also have the SYN packet (with no data) in the same queue.
3453 */
3454 TCP_SKB_CB(syn_data)->seq++;
3455 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3456 if (!err) {
67da22d2 3457 tp->syn_data = (fo->copied > 0);
75c119af 3458 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
f19c29e3 3459 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
783237e8
YC
3460 goto done;
3461 }
783237e8 3462
75c119af
ED
3463 /* data was not sent, put it in write_queue */
3464 __skb_queue_tail(&sk->sk_write_queue, syn_data);
b5b7db8d
ED
3465 tp->packets_out -= tcp_skb_pcount(syn_data);
3466
783237e8
YC
3467fallback:
3468 /* Send a regular SYN with Fast Open cookie request option */
3469 if (fo->cookie.len > 0)
3470 fo->cookie.len = 0;
3471 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3472 if (err)
3473 tp->syn_fastopen = 0;
783237e8
YC
3474done:
3475 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3476 return err;
3477}
3478
67edfef7 3479/* Build a SYN and send it off. */
1da177e4
LT
3480int tcp_connect(struct sock *sk)
3481{
3482 struct tcp_sock *tp = tcp_sk(sk);
3483 struct sk_buff *buff;
ee586811 3484 int err;
1da177e4 3485
de525be2 3486 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
8ba60924
ED
3487
3488 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3489 return -EHOSTUNREACH; /* Routing failure or similar. */
3490
1da177e4
LT
3491 tcp_connect_init(sk);
3492
2b916477
AV
3493 if (unlikely(tp->repair)) {
3494 tcp_finish_connect(sk, NULL);
3495 return 0;
3496 }
3497
eb934478 3498 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
355a901e 3499 if (unlikely(!buff))
1da177e4
LT
3500 return -ENOBUFS;
3501
a3433f35 3502 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
9a568de4
ED
3503 tcp_mstamp_refresh(tp);
3504 tp->retrans_stamp = tcp_time_stamp(tp);
783237e8 3505 tcp_connect_queue_skb(sk, buff);
735d3831 3506 tcp_ecn_send_syn(sk, buff);
75c119af 3507 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1da177e4 3508
783237e8
YC
3509 /* Send off SYN; include data in Fast Open. */
3510 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3511 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
ee586811
EP
3512 if (err == -ECONNREFUSED)
3513 return err;
bd37a088
WY
3514
3515 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3516 * in order to make this packet get counted in tcpOutSegs.
3517 */
3518 tp->snd_nxt = tp->write_seq;
3519 tp->pushed_seq = tp->write_seq;
b5b7db8d
ED
3520 buff = tcp_send_head(sk);
3521 if (unlikely(buff)) {
3522 tp->snd_nxt = TCP_SKB_CB(buff)->seq;
3523 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3524 }
81cc8a75 3525 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
1da177e4
LT
3526
3527 /* Timer for repeating the SYN until an answer. */
3f421baa
ACM
3528 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3529 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
1da177e4
LT
3530 return 0;
3531}
4bc2f18b 3532EXPORT_SYMBOL(tcp_connect);
1da177e4
LT
3533
3534/* Send out a delayed ack, the caller does the policy checking
3535 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3536 * for details.
3537 */
3538void tcp_send_delayed_ack(struct sock *sk)
3539{
463c84b9
ACM
3540 struct inet_connection_sock *icsk = inet_csk(sk);
3541 int ato = icsk->icsk_ack.ato;
1da177e4
LT
3542 unsigned long timeout;
3543
3544 if (ato > TCP_DELACK_MIN) {
463c84b9 3545 const struct tcp_sock *tp = tcp_sk(sk);
056834d9 3546 int max_ato = HZ / 2;
1da177e4 3547
056834d9
IJ
3548 if (icsk->icsk_ack.pingpong ||
3549 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
1da177e4
LT
3550 max_ato = TCP_DELACK_MAX;
3551
3552 /* Slow path, intersegment interval is "high". */
3553
3554 /* If some rtt estimate is known, use it to bound delayed ack.
463c84b9 3555 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
1da177e4
LT
3556 * directly.
3557 */
740b0f18
ED
3558 if (tp->srtt_us) {
3559 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3560 TCP_DELACK_MIN);
1da177e4
LT
3561
3562 if (rtt < max_ato)
3563 max_ato = rtt;
3564 }
3565
3566 ato = min(ato, max_ato);
3567 }
3568
3569 /* Stay within the limit we were given */
3570 timeout = jiffies + ato;
3571
3572 /* Use new timeout only if there wasn't a older one earlier. */
463c84b9 3573 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
1da177e4
LT
3574 /* If delack timer was blocked or is about to expire,
3575 * send ACK now.
3576 */
463c84b9
ACM
3577 if (icsk->icsk_ack.blocked ||
3578 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
1da177e4
LT
3579 tcp_send_ack(sk);
3580 return;
3581 }
3582
463c84b9
ACM
3583 if (!time_before(timeout, icsk->icsk_ack.timeout))
3584 timeout = icsk->icsk_ack.timeout;
1da177e4 3585 }
463c84b9
ACM
3586 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3587 icsk->icsk_ack.timeout = timeout;
3588 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
1da177e4
LT
3589}
3590
3591/* This routine sends an ack and also updates the window. */
2987babb 3592void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
1da177e4 3593{
058dc334 3594 struct sk_buff *buff;
1da177e4 3595
058dc334
IJ
3596 /* If we have been reset, we may not send again. */
3597 if (sk->sk_state == TCP_CLOSE)
3598 return;
1da177e4 3599
058dc334
IJ
3600 /* We are not putting this on the write queue, so
3601 * tcp_transmit_skb() will set the ownership to this
3602 * sock.
3603 */
7450aaf6
ED
3604 buff = alloc_skb(MAX_TCP_HEADER,
3605 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3606 if (unlikely(!buff)) {
058dc334
IJ
3607 inet_csk_schedule_ack(sk);
3608 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3609 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3610 TCP_DELACK_MAX, TCP_RTO_MAX);
3611 return;
1da177e4 3612 }
058dc334
IJ
3613
3614 /* Reserve space for headers and prepare control bits. */
3615 skb_reserve(buff, MAX_TCP_HEADER);
a3433f35 3616 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
058dc334 3617
98781965
ED
3618 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3619 * too much.
3620 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
98781965
ED
3621 */
3622 skb_set_tcp_pure_ack(buff);
3623
058dc334 3624 /* Send it off, this clears delayed acks for us. */
2987babb
YC
3625 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3626}
27cde44a 3627EXPORT_SYMBOL_GPL(__tcp_send_ack);
2987babb
YC
3628
3629void tcp_send_ack(struct sock *sk)
3630{
3631 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
1da177e4
LT
3632}
3633
3634/* This routine sends a packet with an out of date sequence
3635 * number. It assumes the other end will try to ack it.
3636 *
3637 * Question: what should we make while urgent mode?
3638 * 4.4BSD forces sending single byte of data. We cannot send
3639 * out of window data, because we have SND.NXT==SND.MAX...
3640 *
3641 * Current solution: to send TWO zero-length segments in urgent mode:
3642 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3643 * out-of-date with SND.UNA-1 to probe window.
3644 */
e520af48 3645static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
1da177e4
LT
3646{
3647 struct tcp_sock *tp = tcp_sk(sk);
3648 struct sk_buff *skb;
3649
3650 /* We don't queue it, tcp_transmit_skb() sets ownership. */
7450aaf6
ED
3651 skb = alloc_skb(MAX_TCP_HEADER,
3652 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
51456b29 3653 if (!skb)
1da177e4
LT
3654 return -1;
3655
3656 /* Reserve space for headers and set control bits. */
3657 skb_reserve(skb, MAX_TCP_HEADER);
1da177e4
LT
3658 /* Use a previous sequence. This should cause the other
3659 * end to send an ack. Don't queue or clone SKB, just
3660 * send it.
3661 */
a3433f35 3662 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
e2e8009f 3663 NET_INC_STATS(sock_net(sk), mib);
7450aaf6 3664 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
1da177e4
LT
3665}
3666
385e2070 3667/* Called from setsockopt( ... TCP_REPAIR ) */
ee995283
PE
3668void tcp_send_window_probe(struct sock *sk)
3669{
3670 if (sk->sk_state == TCP_ESTABLISHED) {
3671 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
9a568de4 3672 tcp_mstamp_refresh(tcp_sk(sk));
e520af48 3673 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
ee995283
PE
3674 }
3675}
3676
67edfef7 3677/* Initiate keepalive or window probe from timer. */
e520af48 3678int tcp_write_wakeup(struct sock *sk, int mib)
1da177e4 3679{
058dc334
IJ
3680 struct tcp_sock *tp = tcp_sk(sk);
3681 struct sk_buff *skb;
1da177e4 3682
058dc334
IJ
3683 if (sk->sk_state == TCP_CLOSE)
3684 return -1;
3685
00db4124
IM
3686 skb = tcp_send_head(sk);
3687 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
058dc334 3688 int err;
0c54b85f 3689 unsigned int mss = tcp_current_mss(sk);
058dc334
IJ
3690 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3691
3692 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3693 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3694
3695 /* We are probing the opening of a window
3696 * but the window size is != 0
3697 * must have been a result SWS avoidance ( sender )
3698 */
3699 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3700 skb->len > mss) {
3701 seg_size = min(seg_size, mss);
4de075e0 3702 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
75c119af
ED
3703 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
3704 skb, seg_size, mss, GFP_ATOMIC))
058dc334
IJ
3705 return -1;
3706 } else if (!tcp_skb_pcount(skb))
5bbb432c 3707 tcp_set_skb_tso_segs(skb, mss);
058dc334 3708
4de075e0 3709 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
058dc334
IJ
3710 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3711 if (!err)
3712 tcp_event_new_data_sent(sk, skb);
3713 return err;
3714 } else {
33f5f57e 3715 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
e520af48
ED
3716 tcp_xmit_probe_skb(sk, 1, mib);
3717 return tcp_xmit_probe_skb(sk, 0, mib);
1da177e4 3718 }
1da177e4
LT
3719}
3720
3721/* A window probe timeout has occurred. If window is not closed send
3722 * a partial packet else a zero probe.
3723 */
3724void tcp_send_probe0(struct sock *sk)
3725{
463c84b9 3726 struct inet_connection_sock *icsk = inet_csk(sk);
1da177e4 3727 struct tcp_sock *tp = tcp_sk(sk);
c6214a97 3728 struct net *net = sock_net(sk);
fcdd1cf4 3729 unsigned long probe_max;
1da177e4
LT
3730 int err;
3731
e520af48 3732 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
1da177e4 3733
75c119af 3734 if (tp->packets_out || tcp_write_queue_empty(sk)) {
1da177e4 3735 /* Cancel probe timer, if it is not required. */
6687e988 3736 icsk->icsk_probes_out = 0;
463c84b9 3737 icsk->icsk_backoff = 0;
1da177e4
LT
3738 return;
3739 }
3740
3741 if (err <= 0) {
c6214a97 3742 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
463c84b9 3743 icsk->icsk_backoff++;
6687e988 3744 icsk->icsk_probes_out++;
fcdd1cf4 3745 probe_max = TCP_RTO_MAX;
1da177e4
LT
3746 } else {
3747 /* If packet was not sent due to local congestion,
6687e988 3748 * do not backoff and do not remember icsk_probes_out.
1da177e4
LT
3749 * Let local senders to fight for local resources.
3750 *
3751 * Use accumulated backoff yet.
3752 */
6687e988
ACM
3753 if (!icsk->icsk_probes_out)
3754 icsk->icsk_probes_out = 1;
fcdd1cf4 3755 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
1da177e4 3756 }
fcdd1cf4 3757 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
21c8fe99 3758 tcp_probe0_when(sk, probe_max),
fcdd1cf4 3759 TCP_RTO_MAX);
1da177e4 3760}
5db92c99 3761
ea3bea3a 3762int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
5db92c99
OP
3763{
3764 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3765 struct flowi fl;
3766 int res;
3767
58d607d3 3768 tcp_rsk(req)->txhash = net_tx_rndhash();
b3d05147 3769 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
5db92c99 3770 if (!res) {
90bbcc60 3771 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
02a1d6e7 3772 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
7e32b443
YC
3773 if (unlikely(tcp_passive_fastopen(sk)))
3774 tcp_sk(sk)->total_retrans++;
cf34ce3d 3775 trace_tcp_retransmit_synack(sk, req);
5db92c99
OP
3776 }
3777 return res;
3778}
3779EXPORT_SYMBOL(tcp_rtx_synack);