Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
9 * Authors: Ross Biro
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
20 */
21
22/*
23 * Changes: Pedro Roque : Retransmit queue handled by TCP.
24 * : Fragmentation on mtu decrease
25 * : Segment collapse on retransmit
26 * : AF independence
27 *
28 * Linus Torvalds : send_delayed_ack
29 * David S. Miller : Charge memory using the right skb
30 * during syn/ack processing.
31 * David S. Miller : Output engine completely rewritten.
32 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
33 * Cacophonix Gaul : draft-minshall-nagle-01
34 * J Hadi Salim : ECN support
35 *
36 */
37
38#define pr_fmt(fmt) "TCP: " fmt
39
40#include <net/tcp.h>
41
42#include <linux/compiler.h>
43#include <linux/gfp.h>
44#include <linux/module.h>
45#include <linux/static_key.h>
46
47#include <trace/events/tcp.h>
48
49/* Refresh clocks of a TCP socket,
50 * ensuring monotically increasing values.
51 */
52void tcp_mstamp_refresh(struct tcp_sock *tp)
53{
54 u64 val = tcp_clock_ns();
55
56 tp->tcp_clock_cache = val;
57 tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC);
58}
59
60static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
61 int push_one, gfp_t gfp);
62
63/* Account for new data that has been sent to the network. */
64static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
65{
66 struct inet_connection_sock *icsk = inet_csk(sk);
67 struct tcp_sock *tp = tcp_sk(sk);
68 unsigned int prior_packets = tp->packets_out;
69
70 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq);
71
72 __skb_unlink(skb, &sk->sk_write_queue);
73 tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
74
75 tp->packets_out += tcp_skb_pcount(skb);
76 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
77 tcp_rearm_rto(sk);
78
79 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
80 tcp_skb_pcount(skb));
81}
82
83/* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
84 * window scaling factor due to loss of precision.
85 * If window has been shrunk, what should we make? It is not clear at all.
86 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
87 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
88 * invalid. OK, let's make this for now:
89 */
90static inline __u32 tcp_acceptable_seq(const struct sock *sk)
91{
92 const struct tcp_sock *tp = tcp_sk(sk);
93
94 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
95 (tp->rx_opt.wscale_ok &&
96 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
97 return tp->snd_nxt;
98 else
99 return tcp_wnd_end(tp);
100}
101
102/* Calculate mss to advertise in SYN segment.
103 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
104 *
105 * 1. It is independent of path mtu.
106 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
107 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
108 * attached devices, because some buggy hosts are confused by
109 * large MSS.
110 * 4. We do not make 3, we advertise MSS, calculated from first
111 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
112 * This may be overridden via information stored in routing table.
113 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
114 * probably even Jumbo".
115 */
116static __u16 tcp_advertise_mss(struct sock *sk)
117{
118 struct tcp_sock *tp = tcp_sk(sk);
119 const struct dst_entry *dst = __sk_dst_get(sk);
120 int mss = tp->advmss;
121
122 if (dst) {
123 unsigned int metric = dst_metric_advmss(dst);
124
125 if (metric < mss) {
126 mss = metric;
127 tp->advmss = mss;
128 }
129 }
130
131 return (__u16)mss;
132}
133
134/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
135 * This is the first part of cwnd validation mechanism.
136 */
137void tcp_cwnd_restart(struct sock *sk, s32 delta)
138{
139 struct tcp_sock *tp = tcp_sk(sk);
140 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
141 u32 cwnd = tp->snd_cwnd;
142
143 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
144
145 tp->snd_ssthresh = tcp_current_ssthresh(sk);
146 restart_cwnd = min(restart_cwnd, cwnd);
147
148 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
149 cwnd >>= 1;
150 tp->snd_cwnd = max(cwnd, restart_cwnd);
151 tp->snd_cwnd_stamp = tcp_jiffies32;
152 tp->snd_cwnd_used = 0;
153}
154
155/* Congestion state accounting after a packet has been sent. */
156static void tcp_event_data_sent(struct tcp_sock *tp,
157 struct sock *sk)
158{
159 struct inet_connection_sock *icsk = inet_csk(sk);
160 const u32 now = tcp_jiffies32;
161
162 if (tcp_packets_in_flight(tp) == 0)
163 tcp_ca_event(sk, CA_EVENT_TX_START);
164
165 /* If this is the first data packet sent in response to the
166 * previous received data,
167 * and it is a reply for ato after last received packet,
168 * increase pingpong count.
169 */
170 if (before(tp->lsndtime, icsk->icsk_ack.lrcvtime) &&
171 (u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
172 inet_csk_inc_pingpong_cnt(sk);
173
174 tp->lsndtime = now;
175}
176
177/* Account for an ACK we sent. */
178static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
179 u32 rcv_nxt)
180{
181 struct tcp_sock *tp = tcp_sk(sk);
182
183 if (unlikely(tp->compressed_ack > TCP_FASTRETRANS_THRESH)) {
184 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
185 tp->compressed_ack - TCP_FASTRETRANS_THRESH);
186 tp->compressed_ack = TCP_FASTRETRANS_THRESH;
187 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
188 __sock_put(sk);
189 }
190
191 if (unlikely(rcv_nxt != tp->rcv_nxt))
192 return; /* Special ACK sent by DCTCP to reflect ECN */
193 tcp_dec_quickack_mode(sk, pkts);
194 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
195}
196
197/* Determine a window scaling and initial window to offer.
198 * Based on the assumption that the given amount of space
199 * will be offered. Store the results in the tp structure.
200 * NOTE: for smooth operation initial space offering should
201 * be a multiple of mss if possible. We assume here that mss >= 1.
202 * This MUST be enforced by all callers.
203 */
204void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
205 __u32 *rcv_wnd, __u32 *window_clamp,
206 int wscale_ok, __u8 *rcv_wscale,
207 __u32 init_rcv_wnd)
208{
209 unsigned int space = (__space < 0 ? 0 : __space);
210
211 /* If no clamp set the clamp to the max possible scaled window */
212 if (*window_clamp == 0)
213 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
214 space = min(*window_clamp, space);
215
216 /* Quantize space offering to a multiple of mss if possible. */
217 if (space > mss)
218 space = rounddown(space, mss);
219
220 /* NOTE: offering an initial window larger than 32767
221 * will break some buggy TCP stacks. If the admin tells us
222 * it is likely we could be speaking with such a buggy stack
223 * we will truncate our initial window offering to 32K-1
224 * unless the remote has sent us a window scaling option,
225 * which we interpret as a sign the remote TCP is not
226 * misinterpreting the window field as a signed quantity.
227 */
228 if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
229 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
230 else
231 (*rcv_wnd) = min_t(u32, space, U16_MAX);
232
233 if (init_rcv_wnd)
234 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
235
236 *rcv_wscale = 0;
237 if (wscale_ok) {
238 /* Set window scaling on max possible window */
239 space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
240 space = max_t(u32, space, sysctl_rmem_max);
241 space = min_t(u32, space, *window_clamp);
242 *rcv_wscale = clamp_t(int, ilog2(space) - 15,
243 0, TCP_MAX_WSCALE);
244 }
245 /* Set the clamp no higher than max representable value */
246 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
247}
248EXPORT_SYMBOL(tcp_select_initial_window);
249
250/* Chose a new window to advertise, update state in tcp_sock for the
251 * socket, and return result with RFC1323 scaling applied. The return
252 * value can be stuffed directly into th->window for an outgoing
253 * frame.
254 */
255static u16 tcp_select_window(struct sock *sk)
256{
257 struct tcp_sock *tp = tcp_sk(sk);
258 u32 old_win = tp->rcv_wnd;
259 u32 cur_win = tcp_receive_window(tp);
260 u32 new_win = __tcp_select_window(sk);
261
262 /* Never shrink the offered window */
263 if (new_win < cur_win) {
264 /* Danger Will Robinson!
265 * Don't update rcv_wup/rcv_wnd here or else
266 * we will not be able to advertise a zero
267 * window in time. --DaveM
268 *
269 * Relax Will Robinson.
270 */
271 if (new_win == 0)
272 NET_INC_STATS(sock_net(sk),
273 LINUX_MIB_TCPWANTZEROWINDOWADV);
274 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
275 }
276 tp->rcv_wnd = new_win;
277 tp->rcv_wup = tp->rcv_nxt;
278
279 /* Make sure we do not exceed the maximum possible
280 * scaled window.
281 */
282 if (!tp->rx_opt.rcv_wscale &&
283 sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
284 new_win = min(new_win, MAX_TCP_WINDOW);
285 else
286 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
287
288 /* RFC1323 scaling applied */
289 new_win >>= tp->rx_opt.rcv_wscale;
290
291 /* If we advertise zero window, disable fast path. */
292 if (new_win == 0) {
293 tp->pred_flags = 0;
294 if (old_win)
295 NET_INC_STATS(sock_net(sk),
296 LINUX_MIB_TCPTOZEROWINDOWADV);
297 } else if (old_win == 0) {
298 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
299 }
300
301 return new_win;
302}
303
304/* Packet ECN state for a SYN-ACK */
305static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
306{
307 const struct tcp_sock *tp = tcp_sk(sk);
308
309 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
310 if (!(tp->ecn_flags & TCP_ECN_OK))
311 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
312 else if (tcp_ca_needs_ecn(sk) ||
313 tcp_bpf_ca_needs_ecn(sk))
314 INET_ECN_xmit(sk);
315}
316
317/* Packet ECN state for a SYN. */
318static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
319{
320 struct tcp_sock *tp = tcp_sk(sk);
321 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
322 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
323 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
324
325 if (!use_ecn) {
326 const struct dst_entry *dst = __sk_dst_get(sk);
327
328 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
329 use_ecn = true;
330 }
331
332 tp->ecn_flags = 0;
333
334 if (use_ecn) {
335 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
336 tp->ecn_flags = TCP_ECN_OK;
337 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
338 INET_ECN_xmit(sk);
339 }
340}
341
342static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
343{
344 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
345 /* tp->ecn_flags are cleared at a later point in time when
346 * SYN ACK is ultimatively being received.
347 */
348 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
349}
350
351static void
352tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
353{
354 if (inet_rsk(req)->ecn_ok)
355 th->ece = 1;
356}
357
358/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
359 * be sent.
360 */
361static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
362 struct tcphdr *th, int tcp_header_len)
363{
364 struct tcp_sock *tp = tcp_sk(sk);
365
366 if (tp->ecn_flags & TCP_ECN_OK) {
367 /* Not-retransmitted data segment: set ECT and inject CWR. */
368 if (skb->len != tcp_header_len &&
369 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
370 INET_ECN_xmit(sk);
371 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
372 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
373 th->cwr = 1;
374 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
375 }
376 } else if (!tcp_ca_needs_ecn(sk)) {
377 /* ACK or retransmitted segment: clear ECT|CE */
378 INET_ECN_dontxmit(sk);
379 }
380 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
381 th->ece = 1;
382 }
383}
384
385/* Constructs common control bits of non-data skb. If SYN/FIN is present,
386 * auto increment end seqno.
387 */
388static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
389{
390 skb->ip_summed = CHECKSUM_PARTIAL;
391
392 TCP_SKB_CB(skb)->tcp_flags = flags;
393 TCP_SKB_CB(skb)->sacked = 0;
394
395 tcp_skb_pcount_set(skb, 1);
396
397 TCP_SKB_CB(skb)->seq = seq;
398 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
399 seq++;
400 TCP_SKB_CB(skb)->end_seq = seq;
401}
402
403static inline bool tcp_urg_mode(const struct tcp_sock *tp)
404{
405 return tp->snd_una != tp->snd_up;
406}
407
408#define OPTION_SACK_ADVERTISE (1 << 0)
409#define OPTION_TS (1 << 1)
410#define OPTION_MD5 (1 << 2)
411#define OPTION_WSCALE (1 << 3)
412#define OPTION_FAST_OPEN_COOKIE (1 << 8)
413#define OPTION_SMC (1 << 9)
414
415static void smc_options_write(__be32 *ptr, u16 *options)
416{
417#if IS_ENABLED(CONFIG_SMC)
418 if (static_branch_unlikely(&tcp_have_smc)) {
419 if (unlikely(OPTION_SMC & *options)) {
420 *ptr++ = htonl((TCPOPT_NOP << 24) |
421 (TCPOPT_NOP << 16) |
422 (TCPOPT_EXP << 8) |
423 (TCPOLEN_EXP_SMC_BASE));
424 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
425 }
426 }
427#endif
428}
429
430struct tcp_out_options {
431 u16 options; /* bit field of OPTION_* */
432 u16 mss; /* 0 to disable */
433 u8 ws; /* window scale, 0 to disable */
434 u8 num_sack_blocks; /* number of SACK blocks to include */
435 u8 hash_size; /* bytes in hash_location */
436 __u8 *hash_location; /* temporary pointer, overloaded */
437 __u32 tsval, tsecr; /* need to include OPTION_TS */
438 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
439};
440
441/* Write previously computed TCP options to the packet.
442 *
443 * Beware: Something in the Internet is very sensitive to the ordering of
444 * TCP options, we learned this through the hard way, so be careful here.
445 * Luckily we can at least blame others for their non-compliance but from
446 * inter-operability perspective it seems that we're somewhat stuck with
447 * the ordering which we have been using if we want to keep working with
448 * those broken things (not that it currently hurts anybody as there isn't
449 * particular reason why the ordering would need to be changed).
450 *
451 * At least SACK_PERM as the first option is known to lead to a disaster
452 * (but it may well be that other scenarios fail similarly).
453 */
454static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
455 struct tcp_out_options *opts)
456{
457 u16 options = opts->options; /* mungable copy */
458
459 if (unlikely(OPTION_MD5 & options)) {
460 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
461 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
462 /* overload cookie hash location */
463 opts->hash_location = (__u8 *)ptr;
464 ptr += 4;
465 }
466
467 if (unlikely(opts->mss)) {
468 *ptr++ = htonl((TCPOPT_MSS << 24) |
469 (TCPOLEN_MSS << 16) |
470 opts->mss);
471 }
472
473 if (likely(OPTION_TS & options)) {
474 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
475 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
476 (TCPOLEN_SACK_PERM << 16) |
477 (TCPOPT_TIMESTAMP << 8) |
478 TCPOLEN_TIMESTAMP);
479 options &= ~OPTION_SACK_ADVERTISE;
480 } else {
481 *ptr++ = htonl((TCPOPT_NOP << 24) |
482 (TCPOPT_NOP << 16) |
483 (TCPOPT_TIMESTAMP << 8) |
484 TCPOLEN_TIMESTAMP);
485 }
486 *ptr++ = htonl(opts->tsval);
487 *ptr++ = htonl(opts->tsecr);
488 }
489
490 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
491 *ptr++ = htonl((TCPOPT_NOP << 24) |
492 (TCPOPT_NOP << 16) |
493 (TCPOPT_SACK_PERM << 8) |
494 TCPOLEN_SACK_PERM);
495 }
496
497 if (unlikely(OPTION_WSCALE & options)) {
498 *ptr++ = htonl((TCPOPT_NOP << 24) |
499 (TCPOPT_WINDOW << 16) |
500 (TCPOLEN_WINDOW << 8) |
501 opts->ws);
502 }
503
504 if (unlikely(opts->num_sack_blocks)) {
505 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
506 tp->duplicate_sack : tp->selective_acks;
507 int this_sack;
508
509 *ptr++ = htonl((TCPOPT_NOP << 24) |
510 (TCPOPT_NOP << 16) |
511 (TCPOPT_SACK << 8) |
512 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
513 TCPOLEN_SACK_PERBLOCK)));
514
515 for (this_sack = 0; this_sack < opts->num_sack_blocks;
516 ++this_sack) {
517 *ptr++ = htonl(sp[this_sack].start_seq);
518 *ptr++ = htonl(sp[this_sack].end_seq);
519 }
520
521 tp->rx_opt.dsack = 0;
522 }
523
524 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
525 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
526 u8 *p = (u8 *)ptr;
527 u32 len; /* Fast Open option length */
528
529 if (foc->exp) {
530 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
531 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
532 TCPOPT_FASTOPEN_MAGIC);
533 p += TCPOLEN_EXP_FASTOPEN_BASE;
534 } else {
535 len = TCPOLEN_FASTOPEN_BASE + foc->len;
536 *p++ = TCPOPT_FASTOPEN;
537 *p++ = len;
538 }
539
540 memcpy(p, foc->val, foc->len);
541 if ((len & 3) == 2) {
542 p[foc->len] = TCPOPT_NOP;
543 p[foc->len + 1] = TCPOPT_NOP;
544 }
545 ptr += (len + 3) >> 2;
546 }
547
548 smc_options_write(ptr, &options);
549}
550
551static void smc_set_option(const struct tcp_sock *tp,
552 struct tcp_out_options *opts,
553 unsigned int *remaining)
554{
555#if IS_ENABLED(CONFIG_SMC)
556 if (static_branch_unlikely(&tcp_have_smc)) {
557 if (tp->syn_smc) {
558 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
559 opts->options |= OPTION_SMC;
560 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
561 }
562 }
563 }
564#endif
565}
566
567static void smc_set_option_cond(const struct tcp_sock *tp,
568 const struct inet_request_sock *ireq,
569 struct tcp_out_options *opts,
570 unsigned int *remaining)
571{
572#if IS_ENABLED(CONFIG_SMC)
573 if (static_branch_unlikely(&tcp_have_smc)) {
574 if (tp->syn_smc && ireq->smc_ok) {
575 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
576 opts->options |= OPTION_SMC;
577 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
578 }
579 }
580 }
581#endif
582}
583
584/* Compute TCP options for SYN packets. This is not the final
585 * network wire format yet.
586 */
587static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
588 struct tcp_out_options *opts,
589 struct tcp_md5sig_key **md5)
590{
591 struct tcp_sock *tp = tcp_sk(sk);
592 unsigned int remaining = MAX_TCP_OPTION_SPACE;
593 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
594
595 *md5 = NULL;
596#ifdef CONFIG_TCP_MD5SIG
597 if (static_branch_unlikely(&tcp_md5_needed) &&
598 rcu_access_pointer(tp->md5sig_info)) {
599 *md5 = tp->af_specific->md5_lookup(sk, sk);
600 if (*md5) {
601 opts->options |= OPTION_MD5;
602 remaining -= TCPOLEN_MD5SIG_ALIGNED;
603 }
604 }
605#endif
606
607 /* We always get an MSS option. The option bytes which will be seen in
608 * normal data packets should timestamps be used, must be in the MSS
609 * advertised. But we subtract them from tp->mss_cache so that
610 * calculations in tcp_sendmsg are simpler etc. So account for this
611 * fact here if necessary. If we don't do this correctly, as a
612 * receiver we won't recognize data packets as being full sized when we
613 * should, and thus we won't abide by the delayed ACK rules correctly.
614 * SACKs don't matter, we never delay an ACK when we have any of those
615 * going out. */
616 opts->mss = tcp_advertise_mss(sk);
617 remaining -= TCPOLEN_MSS_ALIGNED;
618
619 if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
620 opts->options |= OPTION_TS;
621 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
622 opts->tsecr = tp->rx_opt.ts_recent;
623 remaining -= TCPOLEN_TSTAMP_ALIGNED;
624 }
625 if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
626 opts->ws = tp->rx_opt.rcv_wscale;
627 opts->options |= OPTION_WSCALE;
628 remaining -= TCPOLEN_WSCALE_ALIGNED;
629 }
630 if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
631 opts->options |= OPTION_SACK_ADVERTISE;
632 if (unlikely(!(OPTION_TS & opts->options)))
633 remaining -= TCPOLEN_SACKPERM_ALIGNED;
634 }
635
636 if (fastopen && fastopen->cookie.len >= 0) {
637 u32 need = fastopen->cookie.len;
638
639 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
640 TCPOLEN_FASTOPEN_BASE;
641 need = (need + 3) & ~3U; /* Align to 32 bits */
642 if (remaining >= need) {
643 opts->options |= OPTION_FAST_OPEN_COOKIE;
644 opts->fastopen_cookie = &fastopen->cookie;
645 remaining -= need;
646 tp->syn_fastopen = 1;
647 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
648 }
649 }
650
651 smc_set_option(tp, opts, &remaining);
652
653 return MAX_TCP_OPTION_SPACE - remaining;
654}
655
656/* Set up TCP options for SYN-ACKs. */
657static unsigned int tcp_synack_options(const struct sock *sk,
658 struct request_sock *req,
659 unsigned int mss, struct sk_buff *skb,
660 struct tcp_out_options *opts,
661 const struct tcp_md5sig_key *md5,
662 struct tcp_fastopen_cookie *foc)
663{
664 struct inet_request_sock *ireq = inet_rsk(req);
665 unsigned int remaining = MAX_TCP_OPTION_SPACE;
666
667#ifdef CONFIG_TCP_MD5SIG
668 if (md5) {
669 opts->options |= OPTION_MD5;
670 remaining -= TCPOLEN_MD5SIG_ALIGNED;
671
672 /* We can't fit any SACK blocks in a packet with MD5 + TS
673 * options. There was discussion about disabling SACK
674 * rather than TS in order to fit in better with old,
675 * buggy kernels, but that was deemed to be unnecessary.
676 */
677 ireq->tstamp_ok &= !ireq->sack_ok;
678 }
679#endif
680
681 /* We always send an MSS option. */
682 opts->mss = mss;
683 remaining -= TCPOLEN_MSS_ALIGNED;
684
685 if (likely(ireq->wscale_ok)) {
686 opts->ws = ireq->rcv_wscale;
687 opts->options |= OPTION_WSCALE;
688 remaining -= TCPOLEN_WSCALE_ALIGNED;
689 }
690 if (likely(ireq->tstamp_ok)) {
691 opts->options |= OPTION_TS;
692 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
693 opts->tsecr = req->ts_recent;
694 remaining -= TCPOLEN_TSTAMP_ALIGNED;
695 }
696 if (likely(ireq->sack_ok)) {
697 opts->options |= OPTION_SACK_ADVERTISE;
698 if (unlikely(!ireq->tstamp_ok))
699 remaining -= TCPOLEN_SACKPERM_ALIGNED;
700 }
701 if (foc != NULL && foc->len >= 0) {
702 u32 need = foc->len;
703
704 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
705 TCPOLEN_FASTOPEN_BASE;
706 need = (need + 3) & ~3U; /* Align to 32 bits */
707 if (remaining >= need) {
708 opts->options |= OPTION_FAST_OPEN_COOKIE;
709 opts->fastopen_cookie = foc;
710 remaining -= need;
711 }
712 }
713
714 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
715
716 return MAX_TCP_OPTION_SPACE - remaining;
717}
718
719/* Compute TCP options for ESTABLISHED sockets. This is not the
720 * final wire format yet.
721 */
722static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
723 struct tcp_out_options *opts,
724 struct tcp_md5sig_key **md5)
725{
726 struct tcp_sock *tp = tcp_sk(sk);
727 unsigned int size = 0;
728 unsigned int eff_sacks;
729
730 opts->options = 0;
731
732 *md5 = NULL;
733#ifdef CONFIG_TCP_MD5SIG
734 if (static_branch_unlikely(&tcp_md5_needed) &&
735 rcu_access_pointer(tp->md5sig_info)) {
736 *md5 = tp->af_specific->md5_lookup(sk, sk);
737 if (*md5) {
738 opts->options |= OPTION_MD5;
739 size += TCPOLEN_MD5SIG_ALIGNED;
740 }
741 }
742#endif
743
744 if (likely(tp->rx_opt.tstamp_ok)) {
745 opts->options |= OPTION_TS;
746 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
747 opts->tsecr = tp->rx_opt.ts_recent;
748 size += TCPOLEN_TSTAMP_ALIGNED;
749 }
750
751 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
752 if (unlikely(eff_sacks)) {
753 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
754 opts->num_sack_blocks =
755 min_t(unsigned int, eff_sacks,
756 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
757 TCPOLEN_SACK_PERBLOCK);
758 size += TCPOLEN_SACK_BASE_ALIGNED +
759 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
760 }
761
762 return size;
763}
764
765
766/* TCP SMALL QUEUES (TSQ)
767 *
768 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
769 * to reduce RTT and bufferbloat.
770 * We do this using a special skb destructor (tcp_wfree).
771 *
772 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
773 * needs to be reallocated in a driver.
774 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
775 *
776 * Since transmit from skb destructor is forbidden, we use a tasklet
777 * to process all sockets that eventually need to send more skbs.
778 * We use one tasklet per cpu, with its own queue of sockets.
779 */
780struct tsq_tasklet {
781 struct tasklet_struct tasklet;
782 struct list_head head; /* queue of tcp sockets */
783};
784static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
785
786static void tcp_tsq_write(struct sock *sk)
787{
788 if ((1 << sk->sk_state) &
789 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
790 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
791 struct tcp_sock *tp = tcp_sk(sk);
792
793 if (tp->lost_out > tp->retrans_out &&
794 tp->snd_cwnd > tcp_packets_in_flight(tp)) {
795 tcp_mstamp_refresh(tp);
796 tcp_xmit_retransmit_queue(sk);
797 }
798
799 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
800 0, GFP_ATOMIC);
801 }
802}
803
804static void tcp_tsq_handler(struct sock *sk)
805{
806 bh_lock_sock(sk);
807 if (!sock_owned_by_user(sk))
808 tcp_tsq_write(sk);
809 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
810 sock_hold(sk);
811 bh_unlock_sock(sk);
812}
813/*
814 * One tasklet per cpu tries to send more skbs.
815 * We run in tasklet context but need to disable irqs when
816 * transferring tsq->head because tcp_wfree() might
817 * interrupt us (non NAPI drivers)
818 */
819static void tcp_tasklet_func(unsigned long data)
820{
821 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
822 LIST_HEAD(list);
823 unsigned long flags;
824 struct list_head *q, *n;
825 struct tcp_sock *tp;
826 struct sock *sk;
827
828 local_irq_save(flags);
829 list_splice_init(&tsq->head, &list);
830 local_irq_restore(flags);
831
832 list_for_each_safe(q, n, &list) {
833 tp = list_entry(q, struct tcp_sock, tsq_node);
834 list_del(&tp->tsq_node);
835
836 sk = (struct sock *)tp;
837 smp_mb__before_atomic();
838 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
839
840 tcp_tsq_handler(sk);
841 sk_free(sk);
842 }
843}
844
845#define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
846 TCPF_WRITE_TIMER_DEFERRED | \
847 TCPF_DELACK_TIMER_DEFERRED | \
848 TCPF_MTU_REDUCED_DEFERRED)
849/**
850 * tcp_release_cb - tcp release_sock() callback
851 * @sk: socket
852 *
853 * called from release_sock() to perform protocol dependent
854 * actions before socket release.
855 */
856void tcp_release_cb(struct sock *sk)
857{
858 unsigned long flags, nflags;
859
860 /* perform an atomic operation only if at least one flag is set */
861 do {
862 flags = sk->sk_tsq_flags;
863 if (!(flags & TCP_DEFERRED_ALL))
864 return;
865 nflags = flags & ~TCP_DEFERRED_ALL;
866 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
867
868 if (flags & TCPF_TSQ_DEFERRED) {
869 tcp_tsq_write(sk);
870 __sock_put(sk);
871 }
872 /* Here begins the tricky part :
873 * We are called from release_sock() with :
874 * 1) BH disabled
875 * 2) sk_lock.slock spinlock held
876 * 3) socket owned by us (sk->sk_lock.owned == 1)
877 *
878 * But following code is meant to be called from BH handlers,
879 * so we should keep BH disabled, but early release socket ownership
880 */
881 sock_release_ownership(sk);
882
883 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
884 tcp_write_timer_handler(sk);
885 __sock_put(sk);
886 }
887 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
888 tcp_delack_timer_handler(sk);
889 __sock_put(sk);
890 }
891 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
892 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
893 __sock_put(sk);
894 }
895}
896EXPORT_SYMBOL(tcp_release_cb);
897
898void __init tcp_tasklet_init(void)
899{
900 int i;
901
902 for_each_possible_cpu(i) {
903 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
904
905 INIT_LIST_HEAD(&tsq->head);
906 tasklet_init(&tsq->tasklet,
907 tcp_tasklet_func,
908 (unsigned long)tsq);
909 }
910}
911
912/*
913 * Write buffer destructor automatically called from kfree_skb.
914 * We can't xmit new skbs from this context, as we might already
915 * hold qdisc lock.
916 */
917void tcp_wfree(struct sk_buff *skb)
918{
919 struct sock *sk = skb->sk;
920 struct tcp_sock *tp = tcp_sk(sk);
921 unsigned long flags, nval, oval;
922
923 /* Keep one reference on sk_wmem_alloc.
924 * Will be released by sk_free() from here or tcp_tasklet_func()
925 */
926 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
927
928 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
929 * Wait until our queues (qdisc + devices) are drained.
930 * This gives :
931 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
932 * - chance for incoming ACK (processed by another cpu maybe)
933 * to migrate this flow (skb->ooo_okay will be eventually set)
934 */
935 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
936 goto out;
937
938 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
939 struct tsq_tasklet *tsq;
940 bool empty;
941
942 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
943 goto out;
944
945 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
946 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
947 if (nval != oval)
948 continue;
949
950 /* queue this socket to tasklet queue */
951 local_irq_save(flags);
952 tsq = this_cpu_ptr(&tsq_tasklet);
953 empty = list_empty(&tsq->head);
954 list_add(&tp->tsq_node, &tsq->head);
955 if (empty)
956 tasklet_schedule(&tsq->tasklet);
957 local_irq_restore(flags);
958 return;
959 }
960out:
961 sk_free(sk);
962}
963
964/* Note: Called under soft irq.
965 * We can call TCP stack right away, unless socket is owned by user.
966 */
967enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
968{
969 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
970 struct sock *sk = (struct sock *)tp;
971
972 tcp_tsq_handler(sk);
973 sock_put(sk);
974
975 return HRTIMER_NORESTART;
976}
977
978static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
979 u64 prior_wstamp)
980{
981 struct tcp_sock *tp = tcp_sk(sk);
982
983 if (sk->sk_pacing_status != SK_PACING_NONE) {
984 unsigned long rate = sk->sk_pacing_rate;
985
986 /* Original sch_fq does not pace first 10 MSS
987 * Note that tp->data_segs_out overflows after 2^32 packets,
988 * this is a minor annoyance.
989 */
990 if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
991 u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
992 u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
993
994 /* take into account OS jitter */
995 len_ns -= min_t(u64, len_ns / 2, credit);
996 tp->tcp_wstamp_ns += len_ns;
997 }
998 }
999 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1000}
1001
1002/* This routine actually transmits TCP packets queued in by
1003 * tcp_do_sendmsg(). This is used by both the initial
1004 * transmission and possible later retransmissions.
1005 * All SKB's seen here are completely headerless. It is our
1006 * job to build the TCP header, and pass the packet down to
1007 * IP so it can do the same plus pass the packet off to the
1008 * device.
1009 *
1010 * We are working here with either a clone of the original
1011 * SKB, or a fresh unique copy made by the retransmit engine.
1012 */
1013static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1014 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1015{
1016 const struct inet_connection_sock *icsk = inet_csk(sk);
1017 struct inet_sock *inet;
1018 struct tcp_sock *tp;
1019 struct tcp_skb_cb *tcb;
1020 struct tcp_out_options opts;
1021 unsigned int tcp_options_size, tcp_header_size;
1022 struct sk_buff *oskb = NULL;
1023 struct tcp_md5sig_key *md5;
1024 struct tcphdr *th;
1025 u64 prior_wstamp;
1026 int err;
1027
1028 BUG_ON(!skb || !tcp_skb_pcount(skb));
1029 tp = tcp_sk(sk);
1030 prior_wstamp = tp->tcp_wstamp_ns;
1031 tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1032 skb->skb_mstamp_ns = tp->tcp_wstamp_ns;
1033 if (clone_it) {
1034 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1035 - tp->snd_una;
1036 oskb = skb;
1037
1038 tcp_skb_tsorted_save(oskb) {
1039 if (unlikely(skb_cloned(oskb)))
1040 skb = pskb_copy(oskb, gfp_mask);
1041 else
1042 skb = skb_clone(oskb, gfp_mask);
1043 } tcp_skb_tsorted_restore(oskb);
1044
1045 if (unlikely(!skb))
1046 return -ENOBUFS;
1047 }
1048
1049 inet = inet_sk(sk);
1050 tcb = TCP_SKB_CB(skb);
1051 memset(&opts, 0, sizeof(opts));
1052
1053 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1054 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1055 } else {
1056 tcp_options_size = tcp_established_options(sk, skb, &opts,
1057 &md5);
1058 /* Force a PSH flag on all (GSO) packets to expedite GRO flush
1059 * at receiver : This slightly improve GRO performance.
1060 * Note that we do not force the PSH flag for non GSO packets,
1061 * because they might be sent under high congestion events,
1062 * and in this case it is better to delay the delivery of 1-MSS
1063 * packets and thus the corresponding ACK packet that would
1064 * release the following packet.
1065 */
1066 if (tcp_skb_pcount(skb) > 1)
1067 tcb->tcp_flags |= TCPHDR_PSH;
1068 }
1069 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1070
1071 /* if no packet is in qdisc/device queue, then allow XPS to select
1072 * another queue. We can be called from tcp_tsq_handler()
1073 * which holds one reference to sk.
1074 *
1075 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1076 * One way to get this would be to set skb->truesize = 2 on them.
1077 */
1078 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1079
1080 /* If we had to use memory reserve to allocate this skb,
1081 * this might cause drops if packet is looped back :
1082 * Other socket might not have SOCK_MEMALLOC.
1083 * Packets not looped back do not care about pfmemalloc.
1084 */
1085 skb->pfmemalloc = 0;
1086
1087 skb_push(skb, tcp_header_size);
1088 skb_reset_transport_header(skb);
1089
1090 skb_orphan(skb);
1091 skb->sk = sk;
1092 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1093 skb_set_hash_from_sk(skb, sk);
1094 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1095
1096 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1097
1098 /* Build TCP header and checksum it. */
1099 th = (struct tcphdr *)skb->data;
1100 th->source = inet->inet_sport;
1101 th->dest = inet->inet_dport;
1102 th->seq = htonl(tcb->seq);
1103 th->ack_seq = htonl(rcv_nxt);
1104 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1105 tcb->tcp_flags);
1106
1107 th->check = 0;
1108 th->urg_ptr = 0;
1109
1110 /* The urg_mode check is necessary during a below snd_una win probe */
1111 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1112 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1113 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1114 th->urg = 1;
1115 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1116 th->urg_ptr = htons(0xFFFF);
1117 th->urg = 1;
1118 }
1119 }
1120
1121 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1122 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1123 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1124 th->window = htons(tcp_select_window(sk));
1125 tcp_ecn_send(sk, skb, th, tcp_header_size);
1126 } else {
1127 /* RFC1323: The window in SYN & SYN/ACK segments
1128 * is never scaled.
1129 */
1130 th->window = htons(min(tp->rcv_wnd, 65535U));
1131 }
1132#ifdef CONFIG_TCP_MD5SIG
1133 /* Calculate the MD5 hash, as we have all we need now */
1134 if (md5) {
1135 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1136 tp->af_specific->calc_md5_hash(opts.hash_location,
1137 md5, sk, skb);
1138 }
1139#endif
1140
1141 icsk->icsk_af_ops->send_check(sk, skb);
1142
1143 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1144 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1145
1146 if (skb->len != tcp_header_size) {
1147 tcp_event_data_sent(tp, sk);
1148 tp->data_segs_out += tcp_skb_pcount(skb);
1149 tp->bytes_sent += skb->len - tcp_header_size;
1150 }
1151
1152 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1153 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1154 tcp_skb_pcount(skb));
1155
1156 tp->segs_out += tcp_skb_pcount(skb);
1157 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1158 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1159 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1160
1161 /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1162
1163 /* Cleanup our debris for IP stacks */
1164 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1165 sizeof(struct inet6_skb_parm)));
1166
1167 tcp_add_tx_delay(skb, tp);
1168
1169 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1170
1171 if (unlikely(err > 0)) {
1172 tcp_enter_cwr(sk);
1173 err = net_xmit_eval(err);
1174 }
1175 if (!err && oskb) {
1176 tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1177 tcp_rate_skb_sent(sk, oskb);
1178 }
1179 return err;
1180}
1181
1182static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1183 gfp_t gfp_mask)
1184{
1185 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1186 tcp_sk(sk)->rcv_nxt);
1187}
1188
1189/* This routine just queues the buffer for sending.
1190 *
1191 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1192 * otherwise socket can stall.
1193 */
1194static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1195{
1196 struct tcp_sock *tp = tcp_sk(sk);
1197
1198 /* Advance write_seq and place onto the write_queue. */
1199 WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1200 __skb_header_release(skb);
1201 tcp_add_write_queue_tail(sk, skb);
1202 sk_wmem_queued_add(sk, skb->truesize);
1203 sk_mem_charge(sk, skb->truesize);
1204}
1205
1206/* Initialize TSO segments for a packet. */
1207static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1208{
1209 if (skb->len <= mss_now) {
1210 /* Avoid the costly divide in the normal
1211 * non-TSO case.
1212 */
1213 tcp_skb_pcount_set(skb, 1);
1214 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1215 } else {
1216 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1217 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1218 }
1219}
1220
1221/* Pcount in the middle of the write queue got changed, we need to do various
1222 * tweaks to fix counters
1223 */
1224static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1225{
1226 struct tcp_sock *tp = tcp_sk(sk);
1227
1228 tp->packets_out -= decr;
1229
1230 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1231 tp->sacked_out -= decr;
1232 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1233 tp->retrans_out -= decr;
1234 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1235 tp->lost_out -= decr;
1236
1237 /* Reno case is special. Sigh... */
1238 if (tcp_is_reno(tp) && decr > 0)
1239 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1240
1241 if (tp->lost_skb_hint &&
1242 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1243 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1244 tp->lost_cnt_hint -= decr;
1245
1246 tcp_verify_left_out(tp);
1247}
1248
1249static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1250{
1251 return TCP_SKB_CB(skb)->txstamp_ack ||
1252 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1253}
1254
1255static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1256{
1257 struct skb_shared_info *shinfo = skb_shinfo(skb);
1258
1259 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1260 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1261 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1262 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1263
1264 shinfo->tx_flags &= ~tsflags;
1265 shinfo2->tx_flags |= tsflags;
1266 swap(shinfo->tskey, shinfo2->tskey);
1267 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1268 TCP_SKB_CB(skb)->txstamp_ack = 0;
1269 }
1270}
1271
1272static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1273{
1274 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1275 TCP_SKB_CB(skb)->eor = 0;
1276}
1277
1278/* Insert buff after skb on the write or rtx queue of sk. */
1279static void tcp_insert_write_queue_after(struct sk_buff *skb,
1280 struct sk_buff *buff,
1281 struct sock *sk,
1282 enum tcp_queue tcp_queue)
1283{
1284 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1285 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1286 else
1287 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1288}
1289
1290/* Function to create two new TCP segments. Shrinks the given segment
1291 * to the specified size and appends a new segment with the rest of the
1292 * packet to the list. This won't be called frequently, I hope.
1293 * Remember, these are still headerless SKBs at this point.
1294 */
1295int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1296 struct sk_buff *skb, u32 len,
1297 unsigned int mss_now, gfp_t gfp)
1298{
1299 struct tcp_sock *tp = tcp_sk(sk);
1300 struct sk_buff *buff;
1301 int nsize, old_factor;
1302 long limit;
1303 int nlen;
1304 u8 flags;
1305
1306 if (WARN_ON(len > skb->len))
1307 return -EINVAL;
1308
1309 nsize = skb_headlen(skb) - len;
1310 if (nsize < 0)
1311 nsize = 0;
1312
1313 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1314 * We need some allowance to not penalize applications setting small
1315 * SO_SNDBUF values.
1316 * Also allow first and last skb in retransmit queue to be split.
1317 */
1318 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE);
1319 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1320 tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1321 skb != tcp_rtx_queue_head(sk) &&
1322 skb != tcp_rtx_queue_tail(sk))) {
1323 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1324 return -ENOMEM;
1325 }
1326
1327 if (skb_unclone(skb, gfp))
1328 return -ENOMEM;
1329
1330 /* Get a new skb... force flag on. */
1331 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1332 if (!buff)
1333 return -ENOMEM; /* We'll just try again later. */
1334 skb_copy_decrypted(buff, skb);
1335
1336 sk_wmem_queued_add(sk, buff->truesize);
1337 sk_mem_charge(sk, buff->truesize);
1338 nlen = skb->len - len - nsize;
1339 buff->truesize += nlen;
1340 skb->truesize -= nlen;
1341
1342 /* Correct the sequence numbers. */
1343 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1344 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1345 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1346
1347 /* PSH and FIN should only be set in the second packet. */
1348 flags = TCP_SKB_CB(skb)->tcp_flags;
1349 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1350 TCP_SKB_CB(buff)->tcp_flags = flags;
1351 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1352 tcp_skb_fragment_eor(skb, buff);
1353
1354 skb_split(skb, buff, len);
1355
1356 buff->ip_summed = CHECKSUM_PARTIAL;
1357
1358 buff->tstamp = skb->tstamp;
1359 tcp_fragment_tstamp(skb, buff);
1360
1361 old_factor = tcp_skb_pcount(skb);
1362
1363 /* Fix up tso_factor for both original and new SKB. */
1364 tcp_set_skb_tso_segs(skb, mss_now);
1365 tcp_set_skb_tso_segs(buff, mss_now);
1366
1367 /* Update delivered info for the new segment */
1368 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1369
1370 /* If this packet has been sent out already, we must
1371 * adjust the various packet counters.
1372 */
1373 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1374 int diff = old_factor - tcp_skb_pcount(skb) -
1375 tcp_skb_pcount(buff);
1376
1377 if (diff)
1378 tcp_adjust_pcount(sk, skb, diff);
1379 }
1380
1381 /* Link BUFF into the send queue. */
1382 __skb_header_release(buff);
1383 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1384 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1385 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1386
1387 return 0;
1388}
1389
1390/* This is similar to __pskb_pull_tail(). The difference is that pulled
1391 * data is not copied, but immediately discarded.
1392 */
1393static int __pskb_trim_head(struct sk_buff *skb, int len)
1394{
1395 struct skb_shared_info *shinfo;
1396 int i, k, eat;
1397
1398 eat = min_t(int, len, skb_headlen(skb));
1399 if (eat) {
1400 __skb_pull(skb, eat);
1401 len -= eat;
1402 if (!len)
1403 return 0;
1404 }
1405 eat = len;
1406 k = 0;
1407 shinfo = skb_shinfo(skb);
1408 for (i = 0; i < shinfo->nr_frags; i++) {
1409 int size = skb_frag_size(&shinfo->frags[i]);
1410
1411 if (size <= eat) {
1412 skb_frag_unref(skb, i);
1413 eat -= size;
1414 } else {
1415 shinfo->frags[k] = shinfo->frags[i];
1416 if (eat) {
1417 skb_frag_off_add(&shinfo->frags[k], eat);
1418 skb_frag_size_sub(&shinfo->frags[k], eat);
1419 eat = 0;
1420 }
1421 k++;
1422 }
1423 }
1424 shinfo->nr_frags = k;
1425
1426 skb->data_len -= len;
1427 skb->len = skb->data_len;
1428 return len;
1429}
1430
1431/* Remove acked data from a packet in the transmit queue. */
1432int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1433{
1434 u32 delta_truesize;
1435
1436 if (skb_unclone(skb, GFP_ATOMIC))
1437 return -ENOMEM;
1438
1439 delta_truesize = __pskb_trim_head(skb, len);
1440
1441 TCP_SKB_CB(skb)->seq += len;
1442 skb->ip_summed = CHECKSUM_PARTIAL;
1443
1444 if (delta_truesize) {
1445 skb->truesize -= delta_truesize;
1446 sk_wmem_queued_add(sk, -delta_truesize);
1447 sk_mem_uncharge(sk, delta_truesize);
1448 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1449 }
1450
1451 /* Any change of skb->len requires recalculation of tso factor. */
1452 if (tcp_skb_pcount(skb) > 1)
1453 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1454
1455 return 0;
1456}
1457
1458/* Calculate MSS not accounting any TCP options. */
1459static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1460{
1461 const struct tcp_sock *tp = tcp_sk(sk);
1462 const struct inet_connection_sock *icsk = inet_csk(sk);
1463 int mss_now;
1464
1465 /* Calculate base mss without TCP options:
1466 It is MMS_S - sizeof(tcphdr) of rfc1122
1467 */
1468 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1469
1470 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1471 if (icsk->icsk_af_ops->net_frag_header_len) {
1472 const struct dst_entry *dst = __sk_dst_get(sk);
1473
1474 if (dst && dst_allfrag(dst))
1475 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1476 }
1477
1478 /* Clamp it (mss_clamp does not include tcp options) */
1479 if (mss_now > tp->rx_opt.mss_clamp)
1480 mss_now = tp->rx_opt.mss_clamp;
1481
1482 /* Now subtract optional transport overhead */
1483 mss_now -= icsk->icsk_ext_hdr_len;
1484
1485 /* Then reserve room for full set of TCP options and 8 bytes of data */
1486 mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss);
1487 return mss_now;
1488}
1489
1490/* Calculate MSS. Not accounting for SACKs here. */
1491int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1492{
1493 /* Subtract TCP options size, not including SACKs */
1494 return __tcp_mtu_to_mss(sk, pmtu) -
1495 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1496}
1497
1498/* Inverse of above */
1499int tcp_mss_to_mtu(struct sock *sk, int mss)
1500{
1501 const struct tcp_sock *tp = tcp_sk(sk);
1502 const struct inet_connection_sock *icsk = inet_csk(sk);
1503 int mtu;
1504
1505 mtu = mss +
1506 tp->tcp_header_len +
1507 icsk->icsk_ext_hdr_len +
1508 icsk->icsk_af_ops->net_header_len;
1509
1510 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1511 if (icsk->icsk_af_ops->net_frag_header_len) {
1512 const struct dst_entry *dst = __sk_dst_get(sk);
1513
1514 if (dst && dst_allfrag(dst))
1515 mtu += icsk->icsk_af_ops->net_frag_header_len;
1516 }
1517 return mtu;
1518}
1519EXPORT_SYMBOL(tcp_mss_to_mtu);
1520
1521/* MTU probing init per socket */
1522void tcp_mtup_init(struct sock *sk)
1523{
1524 struct tcp_sock *tp = tcp_sk(sk);
1525 struct inet_connection_sock *icsk = inet_csk(sk);
1526 struct net *net = sock_net(sk);
1527
1528 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1529 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1530 icsk->icsk_af_ops->net_header_len;
1531 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1532 icsk->icsk_mtup.probe_size = 0;
1533 if (icsk->icsk_mtup.enabled)
1534 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1535}
1536EXPORT_SYMBOL(tcp_mtup_init);
1537
1538/* This function synchronize snd mss to current pmtu/exthdr set.
1539
1540 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1541 for TCP options, but includes only bare TCP header.
1542
1543 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1544 It is minimum of user_mss and mss received with SYN.
1545 It also does not include TCP options.
1546
1547 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1548
1549 tp->mss_cache is current effective sending mss, including
1550 all tcp options except for SACKs. It is evaluated,
1551 taking into account current pmtu, but never exceeds
1552 tp->rx_opt.mss_clamp.
1553
1554 NOTE1. rfc1122 clearly states that advertised MSS
1555 DOES NOT include either tcp or ip options.
1556
1557 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1558 are READ ONLY outside this function. --ANK (980731)
1559 */
1560unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1561{
1562 struct tcp_sock *tp = tcp_sk(sk);
1563 struct inet_connection_sock *icsk = inet_csk(sk);
1564 int mss_now;
1565
1566 if (icsk->icsk_mtup.search_high > pmtu)
1567 icsk->icsk_mtup.search_high = pmtu;
1568
1569 mss_now = tcp_mtu_to_mss(sk, pmtu);
1570 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1571
1572 /* And store cached results */
1573 icsk->icsk_pmtu_cookie = pmtu;
1574 if (icsk->icsk_mtup.enabled)
1575 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1576 tp->mss_cache = mss_now;
1577
1578 return mss_now;
1579}
1580EXPORT_SYMBOL(tcp_sync_mss);
1581
1582/* Compute the current effective MSS, taking SACKs and IP options,
1583 * and even PMTU discovery events into account.
1584 */
1585unsigned int tcp_current_mss(struct sock *sk)
1586{
1587 const struct tcp_sock *tp = tcp_sk(sk);
1588 const struct dst_entry *dst = __sk_dst_get(sk);
1589 u32 mss_now;
1590 unsigned int header_len;
1591 struct tcp_out_options opts;
1592 struct tcp_md5sig_key *md5;
1593
1594 mss_now = tp->mss_cache;
1595
1596 if (dst) {
1597 u32 mtu = dst_mtu(dst);
1598 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1599 mss_now = tcp_sync_mss(sk, mtu);
1600 }
1601
1602 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1603 sizeof(struct tcphdr);
1604 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1605 * some common options. If this is an odd packet (because we have SACK
1606 * blocks etc) then our calculated header_len will be different, and
1607 * we have to adjust mss_now correspondingly */
1608 if (header_len != tp->tcp_header_len) {
1609 int delta = (int) header_len - tp->tcp_header_len;
1610 mss_now -= delta;
1611 }
1612
1613 return mss_now;
1614}
1615
1616/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1617 * As additional protections, we do not touch cwnd in retransmission phases,
1618 * and if application hit its sndbuf limit recently.
1619 */
1620static void tcp_cwnd_application_limited(struct sock *sk)
1621{
1622 struct tcp_sock *tp = tcp_sk(sk);
1623
1624 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1625 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1626 /* Limited by application or receiver window. */
1627 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1628 u32 win_used = max(tp->snd_cwnd_used, init_win);
1629 if (win_used < tp->snd_cwnd) {
1630 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1631 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1632 }
1633 tp->snd_cwnd_used = 0;
1634 }
1635 tp->snd_cwnd_stamp = tcp_jiffies32;
1636}
1637
1638static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1639{
1640 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1641 struct tcp_sock *tp = tcp_sk(sk);
1642
1643 /* Track the maximum number of outstanding packets in each
1644 * window, and remember whether we were cwnd-limited then.
1645 */
1646 if (!before(tp->snd_una, tp->max_packets_seq) ||
1647 tp->packets_out > tp->max_packets_out) {
1648 tp->max_packets_out = tp->packets_out;
1649 tp->max_packets_seq = tp->snd_nxt;
1650 tp->is_cwnd_limited = is_cwnd_limited;
1651 }
1652
1653 if (tcp_is_cwnd_limited(sk)) {
1654 /* Network is feed fully. */
1655 tp->snd_cwnd_used = 0;
1656 tp->snd_cwnd_stamp = tcp_jiffies32;
1657 } else {
1658 /* Network starves. */
1659 if (tp->packets_out > tp->snd_cwnd_used)
1660 tp->snd_cwnd_used = tp->packets_out;
1661
1662 if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle &&
1663 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1664 !ca_ops->cong_control)
1665 tcp_cwnd_application_limited(sk);
1666
1667 /* The following conditions together indicate the starvation
1668 * is caused by insufficient sender buffer:
1669 * 1) just sent some data (see tcp_write_xmit)
1670 * 2) not cwnd limited (this else condition)
1671 * 3) no more data to send (tcp_write_queue_empty())
1672 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1673 */
1674 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1675 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1676 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1677 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1678 }
1679}
1680
1681/* Minshall's variant of the Nagle send check. */
1682static bool tcp_minshall_check(const struct tcp_sock *tp)
1683{
1684 return after(tp->snd_sml, tp->snd_una) &&
1685 !after(tp->snd_sml, tp->snd_nxt);
1686}
1687
1688/* Update snd_sml if this skb is under mss
1689 * Note that a TSO packet might end with a sub-mss segment
1690 * The test is really :
1691 * if ((skb->len % mss) != 0)
1692 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1693 * But we can avoid doing the divide again given we already have
1694 * skb_pcount = skb->len / mss_now
1695 */
1696static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1697 const struct sk_buff *skb)
1698{
1699 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1700 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1701}
1702
1703/* Return false, if packet can be sent now without violation Nagle's rules:
1704 * 1. It is full sized. (provided by caller in %partial bool)
1705 * 2. Or it contains FIN. (already checked by caller)
1706 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1707 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1708 * With Minshall's modification: all sent small packets are ACKed.
1709 */
1710static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1711 int nonagle)
1712{
1713 return partial &&
1714 ((nonagle & TCP_NAGLE_CORK) ||
1715 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1716}
1717
1718/* Return how many segs we'd like on a TSO packet,
1719 * to send one TSO packet per ms
1720 */
1721static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1722 int min_tso_segs)
1723{
1724 u32 bytes, segs;
1725
1726 bytes = min_t(unsigned long,
1727 sk->sk_pacing_rate >> sk->sk_pacing_shift,
1728 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1729
1730 /* Goal is to send at least one packet per ms,
1731 * not one big TSO packet every 100 ms.
1732 * This preserves ACK clocking and is consistent
1733 * with tcp_tso_should_defer() heuristic.
1734 */
1735 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1736
1737 return segs;
1738}
1739
1740/* Return the number of segments we want in the skb we are transmitting.
1741 * See if congestion control module wants to decide; otherwise, autosize.
1742 */
1743static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1744{
1745 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1746 u32 min_tso, tso_segs;
1747
1748 min_tso = ca_ops->min_tso_segs ?
1749 ca_ops->min_tso_segs(sk) :
1750 sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs;
1751
1752 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
1753 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1754}
1755
1756/* Returns the portion of skb which can be sent right away */
1757static unsigned int tcp_mss_split_point(const struct sock *sk,
1758 const struct sk_buff *skb,
1759 unsigned int mss_now,
1760 unsigned int max_segs,
1761 int nonagle)
1762{
1763 const struct tcp_sock *tp = tcp_sk(sk);
1764 u32 partial, needed, window, max_len;
1765
1766 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1767 max_len = mss_now * max_segs;
1768
1769 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1770 return max_len;
1771
1772 needed = min(skb->len, window);
1773
1774 if (max_len <= needed)
1775 return max_len;
1776
1777 partial = needed % mss_now;
1778 /* If last segment is not a full MSS, check if Nagle rules allow us
1779 * to include this last segment in this skb.
1780 * Otherwise, we'll split the skb at last MSS boundary
1781 */
1782 if (tcp_nagle_check(partial != 0, tp, nonagle))
1783 return needed - partial;
1784
1785 return needed;
1786}
1787
1788/* Can at least one segment of SKB be sent right now, according to the
1789 * congestion window rules? If so, return how many segments are allowed.
1790 */
1791static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1792 const struct sk_buff *skb)
1793{
1794 u32 in_flight, cwnd, halfcwnd;
1795
1796 /* Don't be strict about the congestion window for the final FIN. */
1797 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1798 tcp_skb_pcount(skb) == 1)
1799 return 1;
1800
1801 in_flight = tcp_packets_in_flight(tp);
1802 cwnd = tp->snd_cwnd;
1803 if (in_flight >= cwnd)
1804 return 0;
1805
1806 /* For better scheduling, ensure we have at least
1807 * 2 GSO packets in flight.
1808 */
1809 halfcwnd = max(cwnd >> 1, 1U);
1810 return min(halfcwnd, cwnd - in_flight);
1811}
1812
1813/* Initialize TSO state of a skb.
1814 * This must be invoked the first time we consider transmitting
1815 * SKB onto the wire.
1816 */
1817static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1818{
1819 int tso_segs = tcp_skb_pcount(skb);
1820
1821 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1822 tcp_set_skb_tso_segs(skb, mss_now);
1823 tso_segs = tcp_skb_pcount(skb);
1824 }
1825 return tso_segs;
1826}
1827
1828
1829/* Return true if the Nagle test allows this packet to be
1830 * sent now.
1831 */
1832static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1833 unsigned int cur_mss, int nonagle)
1834{
1835 /* Nagle rule does not apply to frames, which sit in the middle of the
1836 * write_queue (they have no chances to get new data).
1837 *
1838 * This is implemented in the callers, where they modify the 'nonagle'
1839 * argument based upon the location of SKB in the send queue.
1840 */
1841 if (nonagle & TCP_NAGLE_PUSH)
1842 return true;
1843
1844 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1845 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1846 return true;
1847
1848 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1849 return true;
1850
1851 return false;
1852}
1853
1854/* Does at least the first segment of SKB fit into the send window? */
1855static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1856 const struct sk_buff *skb,
1857 unsigned int cur_mss)
1858{
1859 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1860
1861 if (skb->len > cur_mss)
1862 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1863
1864 return !after(end_seq, tcp_wnd_end(tp));
1865}
1866
1867/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1868 * which is put after SKB on the list. It is very much like
1869 * tcp_fragment() except that it may make several kinds of assumptions
1870 * in order to speed up the splitting operation. In particular, we
1871 * know that all the data is in scatter-gather pages, and that the
1872 * packet has never been sent out before (and thus is not cloned).
1873 */
1874static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1875 unsigned int mss_now, gfp_t gfp)
1876{
1877 int nlen = skb->len - len;
1878 struct sk_buff *buff;
1879 u8 flags;
1880
1881 /* All of a TSO frame must be composed of paged data. */
1882 if (skb->len != skb->data_len)
1883 return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
1884 skb, len, mss_now, gfp);
1885
1886 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1887 if (unlikely(!buff))
1888 return -ENOMEM;
1889 skb_copy_decrypted(buff, skb);
1890
1891 sk_wmem_queued_add(sk, buff->truesize);
1892 sk_mem_charge(sk, buff->truesize);
1893 buff->truesize += nlen;
1894 skb->truesize -= nlen;
1895
1896 /* Correct the sequence numbers. */
1897 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1898 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1899 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1900
1901 /* PSH and FIN should only be set in the second packet. */
1902 flags = TCP_SKB_CB(skb)->tcp_flags;
1903 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1904 TCP_SKB_CB(buff)->tcp_flags = flags;
1905
1906 /* This packet was never sent out yet, so no SACK bits. */
1907 TCP_SKB_CB(buff)->sacked = 0;
1908
1909 tcp_skb_fragment_eor(skb, buff);
1910
1911 buff->ip_summed = CHECKSUM_PARTIAL;
1912 skb_split(skb, buff, len);
1913 tcp_fragment_tstamp(skb, buff);
1914
1915 /* Fix up tso_factor for both original and new SKB. */
1916 tcp_set_skb_tso_segs(skb, mss_now);
1917 tcp_set_skb_tso_segs(buff, mss_now);
1918
1919 /* Link BUFF into the send queue. */
1920 __skb_header_release(buff);
1921 tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
1922
1923 return 0;
1924}
1925
1926/* Try to defer sending, if possible, in order to minimize the amount
1927 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1928 *
1929 * This algorithm is from John Heffner.
1930 */
1931static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1932 bool *is_cwnd_limited,
1933 bool *is_rwnd_limited,
1934 u32 max_segs)
1935{
1936 const struct inet_connection_sock *icsk = inet_csk(sk);
1937 u32 send_win, cong_win, limit, in_flight;
1938 struct tcp_sock *tp = tcp_sk(sk);
1939 struct sk_buff *head;
1940 int win_divisor;
1941 s64 delta;
1942
1943 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1944 goto send_now;
1945
1946 /* Avoid bursty behavior by allowing defer
1947 * only if the last write was recent (1 ms).
1948 * Note that tp->tcp_wstamp_ns can be in the future if we have
1949 * packets waiting in a qdisc or device for EDT delivery.
1950 */
1951 delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
1952 if (delta > 0)
1953 goto send_now;
1954
1955 in_flight = tcp_packets_in_flight(tp);
1956
1957 BUG_ON(tcp_skb_pcount(skb) <= 1);
1958 BUG_ON(tp->snd_cwnd <= in_flight);
1959
1960 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1961
1962 /* From in_flight test above, we know that cwnd > in_flight. */
1963 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1964
1965 limit = min(send_win, cong_win);
1966
1967 /* If a full-sized TSO skb can be sent, do it. */
1968 if (limit >= max_segs * tp->mss_cache)
1969 goto send_now;
1970
1971 /* Middle in queue won't get any more data, full sendable already? */
1972 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1973 goto send_now;
1974
1975 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
1976 if (win_divisor) {
1977 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1978
1979 /* If at least some fraction of a window is available,
1980 * just use it.
1981 */
1982 chunk /= win_divisor;
1983 if (limit >= chunk)
1984 goto send_now;
1985 } else {
1986 /* Different approach, try not to defer past a single
1987 * ACK. Receiver should ACK every other full sized
1988 * frame, so if we have space for more than 3 frames
1989 * then send now.
1990 */
1991 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1992 goto send_now;
1993 }
1994
1995 /* TODO : use tsorted_sent_queue ? */
1996 head = tcp_rtx_queue_head(sk);
1997 if (!head)
1998 goto send_now;
1999 delta = tp->tcp_clock_cache - head->tstamp;
2000 /* If next ACK is likely to come too late (half srtt), do not defer */
2001 if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
2002 goto send_now;
2003
2004 /* Ok, it looks like it is advisable to defer.
2005 * Three cases are tracked :
2006 * 1) We are cwnd-limited
2007 * 2) We are rwnd-limited
2008 * 3) We are application limited.
2009 */
2010 if (cong_win < send_win) {
2011 if (cong_win <= skb->len) {
2012 *is_cwnd_limited = true;
2013 return true;
2014 }
2015 } else {
2016 if (send_win <= skb->len) {
2017 *is_rwnd_limited = true;
2018 return true;
2019 }
2020 }
2021
2022 /* If this packet won't get more data, do not wait. */
2023 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2024 TCP_SKB_CB(skb)->eor)
2025 goto send_now;
2026
2027 return true;
2028
2029send_now:
2030 return false;
2031}
2032
2033static inline void tcp_mtu_check_reprobe(struct sock *sk)
2034{
2035 struct inet_connection_sock *icsk = inet_csk(sk);
2036 struct tcp_sock *tp = tcp_sk(sk);
2037 struct net *net = sock_net(sk);
2038 u32 interval;
2039 s32 delta;
2040
2041 interval = net->ipv4.sysctl_tcp_probe_interval;
2042 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2043 if (unlikely(delta >= interval * HZ)) {
2044 int mss = tcp_current_mss(sk);
2045
2046 /* Update current search range */
2047 icsk->icsk_mtup.probe_size = 0;
2048 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2049 sizeof(struct tcphdr) +
2050 icsk->icsk_af_ops->net_header_len;
2051 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2052
2053 /* Update probe time stamp */
2054 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2055 }
2056}
2057
2058static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2059{
2060 struct sk_buff *skb, *next;
2061
2062 skb = tcp_send_head(sk);
2063 tcp_for_write_queue_from_safe(skb, next, sk) {
2064 if (len <= skb->len)
2065 break;
2066
2067 if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb))
2068 return false;
2069
2070 len -= skb->len;
2071 }
2072
2073 return true;
2074}
2075
2076/* Create a new MTU probe if we are ready.
2077 * MTU probe is regularly attempting to increase the path MTU by
2078 * deliberately sending larger packets. This discovers routing
2079 * changes resulting in larger path MTUs.
2080 *
2081 * Returns 0 if we should wait to probe (no cwnd available),
2082 * 1 if a probe was sent,
2083 * -1 otherwise
2084 */
2085static int tcp_mtu_probe(struct sock *sk)
2086{
2087 struct inet_connection_sock *icsk = inet_csk(sk);
2088 struct tcp_sock *tp = tcp_sk(sk);
2089 struct sk_buff *skb, *nskb, *next;
2090 struct net *net = sock_net(sk);
2091 int probe_size;
2092 int size_needed;
2093 int copy, len;
2094 int mss_now;
2095 int interval;
2096
2097 /* Not currently probing/verifying,
2098 * not in recovery,
2099 * have enough cwnd, and
2100 * not SACKing (the variable headers throw things off)
2101 */
2102 if (likely(!icsk->icsk_mtup.enabled ||
2103 icsk->icsk_mtup.probe_size ||
2104 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2105 tp->snd_cwnd < 11 ||
2106 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2107 return -1;
2108
2109 /* Use binary search for probe_size between tcp_mss_base,
2110 * and current mss_clamp. if (search_high - search_low)
2111 * smaller than a threshold, backoff from probing.
2112 */
2113 mss_now = tcp_current_mss(sk);
2114 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2115 icsk->icsk_mtup.search_low) >> 1);
2116 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2117 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2118 /* When misfortune happens, we are reprobing actively,
2119 * and then reprobe timer has expired. We stick with current
2120 * probing process by not resetting search range to its orignal.
2121 */
2122 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2123 interval < net->ipv4.sysctl_tcp_probe_threshold) {
2124 /* Check whether enough time has elaplased for
2125 * another round of probing.
2126 */
2127 tcp_mtu_check_reprobe(sk);
2128 return -1;
2129 }
2130
2131 /* Have enough data in the send queue to probe? */
2132 if (tp->write_seq - tp->snd_nxt < size_needed)
2133 return -1;
2134
2135 if (tp->snd_wnd < size_needed)
2136 return -1;
2137 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2138 return 0;
2139
2140 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2141 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2142 if (!tcp_packets_in_flight(tp))
2143 return -1;
2144 else
2145 return 0;
2146 }
2147
2148 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2149 return -1;
2150
2151 /* We're allowed to probe. Build it now. */
2152 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2153 if (!nskb)
2154 return -1;
2155 sk_wmem_queued_add(sk, nskb->truesize);
2156 sk_mem_charge(sk, nskb->truesize);
2157
2158 skb = tcp_send_head(sk);
2159 skb_copy_decrypted(nskb, skb);
2160
2161 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2162 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2163 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2164 TCP_SKB_CB(nskb)->sacked = 0;
2165 nskb->csum = 0;
2166 nskb->ip_summed = CHECKSUM_PARTIAL;
2167
2168 tcp_insert_write_queue_before(nskb, skb, sk);
2169 tcp_highest_sack_replace(sk, skb, nskb);
2170
2171 len = 0;
2172 tcp_for_write_queue_from_safe(skb, next, sk) {
2173 copy = min_t(int, skb->len, probe_size - len);
2174 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2175
2176 if (skb->len <= copy) {
2177 /* We've eaten all the data from this skb.
2178 * Throw it away. */
2179 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2180 /* If this is the last SKB we copy and eor is set
2181 * we need to propagate it to the new skb.
2182 */
2183 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2184 tcp_skb_collapse_tstamp(nskb, skb);
2185 tcp_unlink_write_queue(skb, sk);
2186 sk_wmem_free_skb(sk, skb);
2187 } else {
2188 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2189 ~(TCPHDR_FIN|TCPHDR_PSH);
2190 if (!skb_shinfo(skb)->nr_frags) {
2191 skb_pull(skb, copy);
2192 } else {
2193 __pskb_trim_head(skb, copy);
2194 tcp_set_skb_tso_segs(skb, mss_now);
2195 }
2196 TCP_SKB_CB(skb)->seq += copy;
2197 }
2198
2199 len += copy;
2200
2201 if (len >= probe_size)
2202 break;
2203 }
2204 tcp_init_tso_segs(nskb, nskb->len);
2205
2206 /* We're ready to send. If this fails, the probe will
2207 * be resegmented into mss-sized pieces by tcp_write_xmit().
2208 */
2209 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2210 /* Decrement cwnd here because we are sending
2211 * effectively two packets. */
2212 tp->snd_cwnd--;
2213 tcp_event_new_data_sent(sk, nskb);
2214
2215 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2216 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2217 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2218
2219 return 1;
2220 }
2221
2222 return -1;
2223}
2224
2225static bool tcp_pacing_check(struct sock *sk)
2226{
2227 struct tcp_sock *tp = tcp_sk(sk);
2228
2229 if (!tcp_needs_internal_pacing(sk))
2230 return false;
2231
2232 if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2233 return false;
2234
2235 if (!hrtimer_is_queued(&tp->pacing_timer)) {
2236 hrtimer_start(&tp->pacing_timer,
2237 ns_to_ktime(tp->tcp_wstamp_ns),
2238 HRTIMER_MODE_ABS_PINNED_SOFT);
2239 sock_hold(sk);
2240 }
2241 return true;
2242}
2243
2244/* TCP Small Queues :
2245 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2246 * (These limits are doubled for retransmits)
2247 * This allows for :
2248 * - better RTT estimation and ACK scheduling
2249 * - faster recovery
2250 * - high rates
2251 * Alas, some drivers / subsystems require a fair amount
2252 * of queued bytes to ensure line rate.
2253 * One example is wifi aggregation (802.11 AMPDU)
2254 */
2255static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2256 unsigned int factor)
2257{
2258 unsigned long limit;
2259
2260 limit = max_t(unsigned long,
2261 2 * skb->truesize,
2262 sk->sk_pacing_rate >> sk->sk_pacing_shift);
2263 if (sk->sk_pacing_status == SK_PACING_NONE)
2264 limit = min_t(unsigned long, limit,
2265 sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
2266 limit <<= factor;
2267
2268 if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2269 tcp_sk(sk)->tcp_tx_delay) {
2270 u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay;
2271
2272 /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2273 * approximate our needs assuming an ~100% skb->truesize overhead.
2274 * USEC_PER_SEC is approximated by 2^20.
2275 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2276 */
2277 extra_bytes >>= (20 - 1);
2278 limit += extra_bytes;
2279 }
2280 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2281 /* Always send skb if rtx queue is empty.
2282 * No need to wait for TX completion to call us back,
2283 * after softirq/tasklet schedule.
2284 * This helps when TX completions are delayed too much.
2285 */
2286 if (tcp_rtx_queue_empty(sk))
2287 return false;
2288
2289 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2290 /* It is possible TX completion already happened
2291 * before we set TSQ_THROTTLED, so we must
2292 * test again the condition.
2293 */
2294 smp_mb__after_atomic();
2295 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2296 return true;
2297 }
2298 return false;
2299}
2300
2301static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2302{
2303 const u32 now = tcp_jiffies32;
2304 enum tcp_chrono old = tp->chrono_type;
2305
2306 if (old > TCP_CHRONO_UNSPEC)
2307 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2308 tp->chrono_start = now;
2309 tp->chrono_type = new;
2310}
2311
2312void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2313{
2314 struct tcp_sock *tp = tcp_sk(sk);
2315
2316 /* If there are multiple conditions worthy of tracking in a
2317 * chronograph then the highest priority enum takes precedence
2318 * over the other conditions. So that if something "more interesting"
2319 * starts happening, stop the previous chrono and start a new one.
2320 */
2321 if (type > tp->chrono_type)
2322 tcp_chrono_set(tp, type);
2323}
2324
2325void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2326{
2327 struct tcp_sock *tp = tcp_sk(sk);
2328
2329
2330 /* There are multiple conditions worthy of tracking in a
2331 * chronograph, so that the highest priority enum takes
2332 * precedence over the other conditions (see tcp_chrono_start).
2333 * If a condition stops, we only stop chrono tracking if
2334 * it's the "most interesting" or current chrono we are
2335 * tracking and starts busy chrono if we have pending data.
2336 */
2337 if (tcp_rtx_and_write_queues_empty(sk))
2338 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2339 else if (type == tp->chrono_type)
2340 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2341}
2342
2343/* This routine writes packets to the network. It advances the
2344 * send_head. This happens as incoming acks open up the remote
2345 * window for us.
2346 *
2347 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2348 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2349 * account rare use of URG, this is not a big flaw.
2350 *
2351 * Send at most one packet when push_one > 0. Temporarily ignore
2352 * cwnd limit to force at most one packet out when push_one == 2.
2353
2354 * Returns true, if no segments are in flight and we have queued segments,
2355 * but cannot send anything now because of SWS or another problem.
2356 */
2357static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2358 int push_one, gfp_t gfp)
2359{
2360 struct tcp_sock *tp = tcp_sk(sk);
2361 struct sk_buff *skb;
2362 unsigned int tso_segs, sent_pkts;
2363 int cwnd_quota;
2364 int result;
2365 bool is_cwnd_limited = false, is_rwnd_limited = false;
2366 u32 max_segs;
2367
2368 sent_pkts = 0;
2369
2370 tcp_mstamp_refresh(tp);
2371 if (!push_one) {
2372 /* Do MTU probing. */
2373 result = tcp_mtu_probe(sk);
2374 if (!result) {
2375 return false;
2376 } else if (result > 0) {
2377 sent_pkts = 1;
2378 }
2379 }
2380
2381 max_segs = tcp_tso_segs(sk, mss_now);
2382 while ((skb = tcp_send_head(sk))) {
2383 unsigned int limit;
2384
2385 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2386 /* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2387 skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2388 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2389 tcp_init_tso_segs(skb, mss_now);
2390 goto repair; /* Skip network transmission */
2391 }
2392
2393 if (tcp_pacing_check(sk))
2394 break;
2395
2396 tso_segs = tcp_init_tso_segs(skb, mss_now);
2397 BUG_ON(!tso_segs);
2398
2399 cwnd_quota = tcp_cwnd_test(tp, skb);
2400 if (!cwnd_quota) {
2401 if (push_one == 2)
2402 /* Force out a loss probe pkt. */
2403 cwnd_quota = 1;
2404 else
2405 break;
2406 }
2407
2408 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2409 is_rwnd_limited = true;
2410 break;
2411 }
2412
2413 if (tso_segs == 1) {
2414 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2415 (tcp_skb_is_last(sk, skb) ?
2416 nonagle : TCP_NAGLE_PUSH))))
2417 break;
2418 } else {
2419 if (!push_one &&
2420 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2421 &is_rwnd_limited, max_segs))
2422 break;
2423 }
2424
2425 limit = mss_now;
2426 if (tso_segs > 1 && !tcp_urg_mode(tp))
2427 limit = tcp_mss_split_point(sk, skb, mss_now,
2428 min_t(unsigned int,
2429 cwnd_quota,
2430 max_segs),
2431 nonagle);
2432
2433 if (skb->len > limit &&
2434 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2435 break;
2436
2437 if (tcp_small_queue_check(sk, skb, 0))
2438 break;
2439
2440 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2441 break;
2442
2443repair:
2444 /* Advance the send_head. This one is sent out.
2445 * This call will increment packets_out.
2446 */
2447 tcp_event_new_data_sent(sk, skb);
2448
2449 tcp_minshall_update(tp, mss_now, skb);
2450 sent_pkts += tcp_skb_pcount(skb);
2451
2452 if (push_one)
2453 break;
2454 }
2455
2456 if (is_rwnd_limited)
2457 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2458 else
2459 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2460
2461 if (likely(sent_pkts)) {
2462 if (tcp_in_cwnd_reduction(sk))
2463 tp->prr_out += sent_pkts;
2464
2465 /* Send one loss probe per tail loss episode. */
2466 if (push_one != 2)
2467 tcp_schedule_loss_probe(sk, false);
2468 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2469 tcp_cwnd_validate(sk, is_cwnd_limited);
2470 return false;
2471 }
2472 return !tp->packets_out && !tcp_write_queue_empty(sk);
2473}
2474
2475bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2476{
2477 struct inet_connection_sock *icsk = inet_csk(sk);
2478 struct tcp_sock *tp = tcp_sk(sk);
2479 u32 timeout, rto_delta_us;
2480 int early_retrans;
2481
2482 /* Don't do any loss probe on a Fast Open connection before 3WHS
2483 * finishes.
2484 */
2485 if (rcu_access_pointer(tp->fastopen_rsk))
2486 return false;
2487
2488 early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
2489 /* Schedule a loss probe in 2*RTT for SACK capable connections
2490 * not in loss recovery, that are either limited by cwnd or application.
2491 */
2492 if ((early_retrans != 3 && early_retrans != 4) ||
2493 !tp->packets_out || !tcp_is_sack(tp) ||
2494 (icsk->icsk_ca_state != TCP_CA_Open &&
2495 icsk->icsk_ca_state != TCP_CA_CWR))
2496 return false;
2497
2498 /* Probe timeout is 2*rtt. Add minimum RTO to account
2499 * for delayed ack when there's one outstanding packet. If no RTT
2500 * sample is available then probe after TCP_TIMEOUT_INIT.
2501 */
2502 if (tp->srtt_us) {
2503 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2504 if (tp->packets_out == 1)
2505 timeout += TCP_RTO_MIN;
2506 else
2507 timeout += TCP_TIMEOUT_MIN;
2508 } else {
2509 timeout = TCP_TIMEOUT_INIT;
2510 }
2511
2512 /* If the RTO formula yields an earlier time, then use that time. */
2513 rto_delta_us = advancing_rto ?
2514 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2515 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2516 if (rto_delta_us > 0)
2517 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2518
2519 tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2520 TCP_RTO_MAX, NULL);
2521 return true;
2522}
2523
2524/* Thanks to skb fast clones, we can detect if a prior transmit of
2525 * a packet is still in a qdisc or driver queue.
2526 * In this case, there is very little point doing a retransmit !
2527 */
2528static bool skb_still_in_host_queue(const struct sock *sk,
2529 const struct sk_buff *skb)
2530{
2531 if (unlikely(skb_fclone_busy(sk, skb))) {
2532 NET_INC_STATS(sock_net(sk),
2533 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2534 return true;
2535 }
2536 return false;
2537}
2538
2539/* When probe timeout (PTO) fires, try send a new segment if possible, else
2540 * retransmit the last segment.
2541 */
2542void tcp_send_loss_probe(struct sock *sk)
2543{
2544 struct tcp_sock *tp = tcp_sk(sk);
2545 struct sk_buff *skb;
2546 int pcount;
2547 int mss = tcp_current_mss(sk);
2548
2549 skb = tcp_send_head(sk);
2550 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2551 pcount = tp->packets_out;
2552 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2553 if (tp->packets_out > pcount)
2554 goto probe_sent;
2555 goto rearm_timer;
2556 }
2557 skb = skb_rb_last(&sk->tcp_rtx_queue);
2558 if (unlikely(!skb)) {
2559 WARN_ONCE(tp->packets_out,
2560 "invalid inflight: %u state %u cwnd %u mss %d\n",
2561 tp->packets_out, sk->sk_state, tp->snd_cwnd, mss);
2562 inet_csk(sk)->icsk_pending = 0;
2563 return;
2564 }
2565
2566 /* At most one outstanding TLP retransmission. */
2567 if (tp->tlp_high_seq)
2568 goto rearm_timer;
2569
2570 if (skb_still_in_host_queue(sk, skb))
2571 goto rearm_timer;
2572
2573 pcount = tcp_skb_pcount(skb);
2574 if (WARN_ON(!pcount))
2575 goto rearm_timer;
2576
2577 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2578 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2579 (pcount - 1) * mss, mss,
2580 GFP_ATOMIC)))
2581 goto rearm_timer;
2582 skb = skb_rb_next(skb);
2583 }
2584
2585 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2586 goto rearm_timer;
2587
2588 if (__tcp_retransmit_skb(sk, skb, 1))
2589 goto rearm_timer;
2590
2591 /* Record snd_nxt for loss detection. */
2592 tp->tlp_high_seq = tp->snd_nxt;
2593
2594probe_sent:
2595 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2596 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2597 inet_csk(sk)->icsk_pending = 0;
2598rearm_timer:
2599 tcp_rearm_rto(sk);
2600}
2601
2602/* Push out any pending frames which were held back due to
2603 * TCP_CORK or attempt at coalescing tiny packets.
2604 * The socket must be locked by the caller.
2605 */
2606void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2607 int nonagle)
2608{
2609 /* If we are closed, the bytes will have to remain here.
2610 * In time closedown will finish, we empty the write queue and
2611 * all will be happy.
2612 */
2613 if (unlikely(sk->sk_state == TCP_CLOSE))
2614 return;
2615
2616 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2617 sk_gfp_mask(sk, GFP_ATOMIC)))
2618 tcp_check_probe_timer(sk);
2619}
2620
2621/* Send _single_ skb sitting at the send head. This function requires
2622 * true push pending frames to setup probe timer etc.
2623 */
2624void tcp_push_one(struct sock *sk, unsigned int mss_now)
2625{
2626 struct sk_buff *skb = tcp_send_head(sk);
2627
2628 BUG_ON(!skb || skb->len < mss_now);
2629
2630 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2631}
2632
2633/* This function returns the amount that we can raise the
2634 * usable window based on the following constraints
2635 *
2636 * 1. The window can never be shrunk once it is offered (RFC 793)
2637 * 2. We limit memory per socket
2638 *
2639 * RFC 1122:
2640 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2641 * RECV.NEXT + RCV.WIN fixed until:
2642 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2643 *
2644 * i.e. don't raise the right edge of the window until you can raise
2645 * it at least MSS bytes.
2646 *
2647 * Unfortunately, the recommended algorithm breaks header prediction,
2648 * since header prediction assumes th->window stays fixed.
2649 *
2650 * Strictly speaking, keeping th->window fixed violates the receiver
2651 * side SWS prevention criteria. The problem is that under this rule
2652 * a stream of single byte packets will cause the right side of the
2653 * window to always advance by a single byte.
2654 *
2655 * Of course, if the sender implements sender side SWS prevention
2656 * then this will not be a problem.
2657 *
2658 * BSD seems to make the following compromise:
2659 *
2660 * If the free space is less than the 1/4 of the maximum
2661 * space available and the free space is less than 1/2 mss,
2662 * then set the window to 0.
2663 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2664 * Otherwise, just prevent the window from shrinking
2665 * and from being larger than the largest representable value.
2666 *
2667 * This prevents incremental opening of the window in the regime
2668 * where TCP is limited by the speed of the reader side taking
2669 * data out of the TCP receive queue. It does nothing about
2670 * those cases where the window is constrained on the sender side
2671 * because the pipeline is full.
2672 *
2673 * BSD also seems to "accidentally" limit itself to windows that are a
2674 * multiple of MSS, at least until the free space gets quite small.
2675 * This would appear to be a side effect of the mbuf implementation.
2676 * Combining these two algorithms results in the observed behavior
2677 * of having a fixed window size at almost all times.
2678 *
2679 * Below we obtain similar behavior by forcing the offered window to
2680 * a multiple of the mss when it is feasible to do so.
2681 *
2682 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2683 * Regular options like TIMESTAMP are taken into account.
2684 */
2685u32 __tcp_select_window(struct sock *sk)
2686{
2687 struct inet_connection_sock *icsk = inet_csk(sk);
2688 struct tcp_sock *tp = tcp_sk(sk);
2689 /* MSS for the peer's data. Previous versions used mss_clamp
2690 * here. I don't know if the value based on our guesses
2691 * of peer's MSS is better for the performance. It's more correct
2692 * but may be worse for the performance because of rcv_mss
2693 * fluctuations. --SAW 1998/11/1
2694 */
2695 int mss = icsk->icsk_ack.rcv_mss;
2696 int free_space = tcp_space(sk);
2697 int allowed_space = tcp_full_space(sk);
2698 int full_space = min_t(int, tp->window_clamp, allowed_space);
2699 int window;
2700
2701 if (unlikely(mss > full_space)) {
2702 mss = full_space;
2703 if (mss <= 0)
2704 return 0;
2705 }
2706 if (free_space < (full_space >> 1)) {
2707 icsk->icsk_ack.quick = 0;
2708
2709 if (tcp_under_memory_pressure(sk))
2710 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2711 4U * tp->advmss);
2712
2713 /* free_space might become our new window, make sure we don't
2714 * increase it due to wscale.
2715 */
2716 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2717
2718 /* if free space is less than mss estimate, or is below 1/16th
2719 * of the maximum allowed, try to move to zero-window, else
2720 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2721 * new incoming data is dropped due to memory limits.
2722 * With large window, mss test triggers way too late in order
2723 * to announce zero window in time before rmem limit kicks in.
2724 */
2725 if (free_space < (allowed_space >> 4) || free_space < mss)
2726 return 0;
2727 }
2728
2729 if (free_space > tp->rcv_ssthresh)
2730 free_space = tp->rcv_ssthresh;
2731
2732 /* Don't do rounding if we are using window scaling, since the
2733 * scaled window will not line up with the MSS boundary anyway.
2734 */
2735 if (tp->rx_opt.rcv_wscale) {
2736 window = free_space;
2737
2738 /* Advertise enough space so that it won't get scaled away.
2739 * Import case: prevent zero window announcement if
2740 * 1<<rcv_wscale > mss.
2741 */
2742 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2743 } else {
2744 window = tp->rcv_wnd;
2745 /* Get the largest window that is a nice multiple of mss.
2746 * Window clamp already applied above.
2747 * If our current window offering is within 1 mss of the
2748 * free space we just keep it. This prevents the divide
2749 * and multiply from happening most of the time.
2750 * We also don't do any window rounding when the free space
2751 * is too small.
2752 */
2753 if (window <= free_space - mss || window > free_space)
2754 window = rounddown(free_space, mss);
2755 else if (mss == full_space &&
2756 free_space > window + (full_space >> 1))
2757 window = free_space;
2758 }
2759
2760 return window;
2761}
2762
2763void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2764 const struct sk_buff *next_skb)
2765{
2766 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2767 const struct skb_shared_info *next_shinfo =
2768 skb_shinfo(next_skb);
2769 struct skb_shared_info *shinfo = skb_shinfo(skb);
2770
2771 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2772 shinfo->tskey = next_shinfo->tskey;
2773 TCP_SKB_CB(skb)->txstamp_ack |=
2774 TCP_SKB_CB(next_skb)->txstamp_ack;
2775 }
2776}
2777
2778/* Collapses two adjacent SKB's during retransmission. */
2779static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2780{
2781 struct tcp_sock *tp = tcp_sk(sk);
2782 struct sk_buff *next_skb = skb_rb_next(skb);
2783 int next_skb_size;
2784
2785 next_skb_size = next_skb->len;
2786
2787 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2788
2789 if (next_skb_size) {
2790 if (next_skb_size <= skb_availroom(skb))
2791 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2792 next_skb_size);
2793 else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size))
2794 return false;
2795 }
2796 tcp_highest_sack_replace(sk, next_skb, skb);
2797
2798 /* Update sequence range on original skb. */
2799 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2800
2801 /* Merge over control information. This moves PSH/FIN etc. over */
2802 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2803
2804 /* All done, get rid of second SKB and account for it so
2805 * packet counting does not break.
2806 */
2807 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2808 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2809
2810 /* changed transmit queue under us so clear hints */
2811 tcp_clear_retrans_hints_partial(tp);
2812 if (next_skb == tp->retransmit_skb_hint)
2813 tp->retransmit_skb_hint = skb;
2814
2815 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2816
2817 tcp_skb_collapse_tstamp(skb, next_skb);
2818
2819 tcp_rtx_queue_unlink_and_free(next_skb, sk);
2820 return true;
2821}
2822
2823/* Check if coalescing SKBs is legal. */
2824static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2825{
2826 if (tcp_skb_pcount(skb) > 1)
2827 return false;
2828 if (skb_cloned(skb))
2829 return false;
2830 /* Some heuristics for collapsing over SACK'd could be invented */
2831 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2832 return false;
2833
2834 return true;
2835}
2836
2837/* Collapse packets in the retransmit queue to make to create
2838 * less packets on the wire. This is only done on retransmission.
2839 */
2840static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2841 int space)
2842{
2843 struct tcp_sock *tp = tcp_sk(sk);
2844 struct sk_buff *skb = to, *tmp;
2845 bool first = true;
2846
2847 if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
2848 return;
2849 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2850 return;
2851
2852 skb_rbtree_walk_from_safe(skb, tmp) {
2853 if (!tcp_can_collapse(sk, skb))
2854 break;
2855
2856 if (!tcp_skb_can_collapse_to(to))
2857 break;
2858
2859 space -= skb->len;
2860
2861 if (first) {
2862 first = false;
2863 continue;
2864 }
2865
2866 if (space < 0)
2867 break;
2868
2869 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2870 break;
2871
2872 if (!tcp_collapse_retrans(sk, to))
2873 break;
2874 }
2875}
2876
2877/* This retransmits one SKB. Policy decisions and retransmit queue
2878 * state updates are done by the caller. Returns non-zero if an
2879 * error occurred which prevented the send.
2880 */
2881int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2882{
2883 struct inet_connection_sock *icsk = inet_csk(sk);
2884 struct tcp_sock *tp = tcp_sk(sk);
2885 unsigned int cur_mss;
2886 int diff, len, err;
2887
2888
2889 /* Inconclusive MTU probe */
2890 if (icsk->icsk_mtup.probe_size)
2891 icsk->icsk_mtup.probe_size = 0;
2892
2893 /* Do not sent more than we queued. 1/4 is reserved for possible
2894 * copying overhead: fragmentation, tunneling, mangling etc.
2895 */
2896 if (refcount_read(&sk->sk_wmem_alloc) >
2897 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2898 sk->sk_sndbuf))
2899 return -EAGAIN;
2900
2901 if (skb_still_in_host_queue(sk, skb))
2902 return -EBUSY;
2903
2904 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2905 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
2906 WARN_ON_ONCE(1);
2907 return -EINVAL;
2908 }
2909 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2910 return -ENOMEM;
2911 }
2912
2913 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2914 return -EHOSTUNREACH; /* Routing failure or similar. */
2915
2916 cur_mss = tcp_current_mss(sk);
2917
2918 /* If receiver has shrunk his window, and skb is out of
2919 * new window, do not retransmit it. The exception is the
2920 * case, when window is shrunk to zero. In this case
2921 * our retransmit serves as a zero window probe.
2922 */
2923 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2924 TCP_SKB_CB(skb)->seq != tp->snd_una)
2925 return -EAGAIN;
2926
2927 len = cur_mss * segs;
2928 if (skb->len > len) {
2929 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
2930 cur_mss, GFP_ATOMIC))
2931 return -ENOMEM; /* We'll try again later. */
2932 } else {
2933 if (skb_unclone(skb, GFP_ATOMIC))
2934 return -ENOMEM;
2935
2936 diff = tcp_skb_pcount(skb);
2937 tcp_set_skb_tso_segs(skb, cur_mss);
2938 diff -= tcp_skb_pcount(skb);
2939 if (diff)
2940 tcp_adjust_pcount(sk, skb, diff);
2941 if (skb->len < cur_mss)
2942 tcp_retrans_try_collapse(sk, skb, cur_mss);
2943 }
2944
2945 /* RFC3168, section 6.1.1.1. ECN fallback */
2946 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2947 tcp_ecn_clear_syn(sk, skb);
2948
2949 /* Update global and local TCP statistics. */
2950 segs = tcp_skb_pcount(skb);
2951 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2952 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2953 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2954 tp->total_retrans += segs;
2955 tp->bytes_retrans += skb->len;
2956
2957 /* make sure skb->data is aligned on arches that require it
2958 * and check if ack-trimming & collapsing extended the headroom
2959 * beyond what csum_start can cover.
2960 */
2961 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2962 skb_headroom(skb) >= 0xFFFF)) {
2963 struct sk_buff *nskb;
2964
2965 tcp_skb_tsorted_save(skb) {
2966 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2967 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2968 -ENOBUFS;
2969 } tcp_skb_tsorted_restore(skb);
2970
2971 if (!err) {
2972 tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
2973 tcp_rate_skb_sent(sk, skb);
2974 }
2975 } else {
2976 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2977 }
2978
2979 /* To avoid taking spuriously low RTT samples based on a timestamp
2980 * for a transmit that never happened, always mark EVER_RETRANS
2981 */
2982 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2983
2984 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
2985 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
2986 TCP_SKB_CB(skb)->seq, segs, err);
2987
2988 if (likely(!err)) {
2989 trace_tcp_retransmit_skb(sk, skb);
2990 } else if (err != -EBUSY) {
2991 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
2992 }
2993 return err;
2994}
2995
2996int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2997{
2998 struct tcp_sock *tp = tcp_sk(sk);
2999 int err = __tcp_retransmit_skb(sk, skb, segs);
3000
3001 if (err == 0) {
3002#if FASTRETRANS_DEBUG > 0
3003 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3004 net_dbg_ratelimited("retrans_out leaked\n");
3005 }
3006#endif
3007 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3008 tp->retrans_out += tcp_skb_pcount(skb);
3009 }
3010
3011 /* Save stamp of the first (attempted) retransmit. */
3012 if (!tp->retrans_stamp)
3013 tp->retrans_stamp = tcp_skb_timestamp(skb);
3014
3015 if (tp->undo_retrans < 0)
3016 tp->undo_retrans = 0;
3017 tp->undo_retrans += tcp_skb_pcount(skb);
3018 return err;
3019}
3020
3021/* This gets called after a retransmit timeout, and the initially
3022 * retransmitted data is acknowledged. It tries to continue
3023 * resending the rest of the retransmit queue, until either
3024 * we've sent it all or the congestion window limit is reached.
3025 */
3026void tcp_xmit_retransmit_queue(struct sock *sk)
3027{
3028 const struct inet_connection_sock *icsk = inet_csk(sk);
3029 struct sk_buff *skb, *rtx_head, *hole = NULL;
3030 struct tcp_sock *tp = tcp_sk(sk);
3031 u32 max_segs;
3032 int mib_idx;
3033
3034 if (!tp->packets_out)
3035 return;
3036
3037 rtx_head = tcp_rtx_queue_head(sk);
3038 skb = tp->retransmit_skb_hint ?: rtx_head;
3039 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3040 skb_rbtree_walk_from(skb) {
3041 __u8 sacked;
3042 int segs;
3043
3044 if (tcp_pacing_check(sk))
3045 break;
3046
3047 /* we could do better than to assign each time */
3048 if (!hole)
3049 tp->retransmit_skb_hint = skb;
3050
3051 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
3052 if (segs <= 0)
3053 return;
3054 sacked = TCP_SKB_CB(skb)->sacked;
3055 /* In case tcp_shift_skb_data() have aggregated large skbs,
3056 * we need to make sure not sending too bigs TSO packets
3057 */
3058 segs = min_t(int, segs, max_segs);
3059
3060 if (tp->retrans_out >= tp->lost_out) {
3061 break;
3062 } else if (!(sacked & TCPCB_LOST)) {
3063 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3064 hole = skb;
3065 continue;
3066
3067 } else {
3068 if (icsk->icsk_ca_state != TCP_CA_Loss)
3069 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3070 else
3071 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3072 }
3073
3074 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3075 continue;
3076
3077 if (tcp_small_queue_check(sk, skb, 1))
3078 return;
3079
3080 if (tcp_retransmit_skb(sk, skb, segs))
3081 return;
3082
3083 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3084
3085 if (tcp_in_cwnd_reduction(sk))
3086 tp->prr_out += tcp_skb_pcount(skb);
3087
3088 if (skb == rtx_head &&
3089 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3090 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3091 inet_csk(sk)->icsk_rto,
3092 TCP_RTO_MAX,
3093 skb);
3094 }
3095}
3096
3097/* We allow to exceed memory limits for FIN packets to expedite
3098 * connection tear down and (memory) recovery.
3099 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3100 * or even be forced to close flow without any FIN.
3101 * In general, we want to allow one skb per socket to avoid hangs
3102 * with edge trigger epoll()
3103 */
3104void sk_forced_mem_schedule(struct sock *sk, int size)
3105{
3106 int amt;
3107
3108 if (size <= sk->sk_forward_alloc)
3109 return;
3110 amt = sk_mem_pages(size);
3111 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3112 sk_memory_allocated_add(sk, amt);
3113
3114 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3115 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3116}
3117
3118/* Send a FIN. The caller locks the socket for us.
3119 * We should try to send a FIN packet really hard, but eventually give up.
3120 */
3121void tcp_send_fin(struct sock *sk)
3122{
3123 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
3124 struct tcp_sock *tp = tcp_sk(sk);
3125
3126 /* Optimization, tack on the FIN if we have one skb in write queue and
3127 * this skb was not yet sent, or we are under memory pressure.
3128 * Note: in the latter case, FIN packet will be sent after a timeout,
3129 * as TCP stack thinks it has already been transmitted.
3130 */
3131 if (!tskb && tcp_under_memory_pressure(sk))
3132 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3133
3134 if (tskb) {
3135 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3136 TCP_SKB_CB(tskb)->end_seq++;
3137 tp->write_seq++;
3138 if (tcp_write_queue_empty(sk)) {
3139 /* This means tskb was already sent.
3140 * Pretend we included the FIN on previous transmit.
3141 * We need to set tp->snd_nxt to the value it would have
3142 * if FIN had been sent. This is because retransmit path
3143 * does not change tp->snd_nxt.
3144 */
3145 WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3146 return;
3147 }
3148 } else {
3149 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3150 if (unlikely(!skb))
3151 return;
3152
3153 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3154 skb_reserve(skb, MAX_TCP_HEADER);
3155 sk_forced_mem_schedule(sk, skb->truesize);
3156 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3157 tcp_init_nondata_skb(skb, tp->write_seq,
3158 TCPHDR_ACK | TCPHDR_FIN);
3159 tcp_queue_skb(sk, skb);
3160 }
3161 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3162}
3163
3164/* We get here when a process closes a file descriptor (either due to
3165 * an explicit close() or as a byproduct of exit()'ing) and there
3166 * was unread data in the receive queue. This behavior is recommended
3167 * by RFC 2525, section 2.17. -DaveM
3168 */
3169void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3170{
3171 struct sk_buff *skb;
3172
3173 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3174
3175 /* NOTE: No TCP options attached and we never retransmit this. */
3176 skb = alloc_skb(MAX_TCP_HEADER, priority);
3177 if (!skb) {
3178 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3179 return;
3180 }
3181
3182 /* Reserve space for headers and prepare control bits. */
3183 skb_reserve(skb, MAX_TCP_HEADER);
3184 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3185 TCPHDR_ACK | TCPHDR_RST);
3186 tcp_mstamp_refresh(tcp_sk(sk));
3187 /* Send it off. */
3188 if (tcp_transmit_skb(sk, skb, 0, priority))
3189 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3190
3191 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3192 * skb here is different to the troublesome skb, so use NULL
3193 */
3194 trace_tcp_send_reset(sk, NULL);
3195}
3196
3197/* Send a crossed SYN-ACK during socket establishment.
3198 * WARNING: This routine must only be called when we have already sent
3199 * a SYN packet that crossed the incoming SYN that caused this routine
3200 * to get called. If this assumption fails then the initial rcv_wnd
3201 * and rcv_wscale values will not be correct.
3202 */
3203int tcp_send_synack(struct sock *sk)
3204{
3205 struct sk_buff *skb;
3206
3207 skb = tcp_rtx_queue_head(sk);
3208 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3209 pr_err("%s: wrong queue state\n", __func__);
3210 return -EFAULT;
3211 }
3212 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3213 if (skb_cloned(skb)) {
3214 struct sk_buff *nskb;
3215
3216 tcp_skb_tsorted_save(skb) {
3217 nskb = skb_copy(skb, GFP_ATOMIC);
3218 } tcp_skb_tsorted_restore(skb);
3219 if (!nskb)
3220 return -ENOMEM;
3221 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3222 tcp_rtx_queue_unlink_and_free(skb, sk);
3223 __skb_header_release(nskb);
3224 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3225 sk_wmem_queued_add(sk, nskb->truesize);
3226 sk_mem_charge(sk, nskb->truesize);
3227 skb = nskb;
3228 }
3229
3230 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3231 tcp_ecn_send_synack(sk, skb);
3232 }
3233 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3234}
3235
3236/**
3237 * tcp_make_synack - Prepare a SYN-ACK.
3238 * sk: listener socket
3239 * dst: dst entry attached to the SYNACK
3240 * req: request_sock pointer
3241 *
3242 * Allocate one skb and build a SYNACK packet.
3243 * @dst is consumed : Caller should not use it again.
3244 */
3245struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3246 struct request_sock *req,
3247 struct tcp_fastopen_cookie *foc,
3248 enum tcp_synack_type synack_type)
3249{
3250 struct inet_request_sock *ireq = inet_rsk(req);
3251 const struct tcp_sock *tp = tcp_sk(sk);
3252 struct tcp_md5sig_key *md5 = NULL;
3253 struct tcp_out_options opts;
3254 struct sk_buff *skb;
3255 int tcp_header_size;
3256 struct tcphdr *th;
3257 int mss;
3258 u64 now;
3259
3260 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3261 if (unlikely(!skb)) {
3262 dst_release(dst);
3263 return NULL;
3264 }
3265 /* Reserve space for headers. */
3266 skb_reserve(skb, MAX_TCP_HEADER);
3267
3268 switch (synack_type) {
3269 case TCP_SYNACK_NORMAL:
3270 skb_set_owner_w(skb, req_to_sk(req));
3271 break;
3272 case TCP_SYNACK_COOKIE:
3273 /* Under synflood, we do not attach skb to a socket,
3274 * to avoid false sharing.
3275 */
3276 break;
3277 case TCP_SYNACK_FASTOPEN:
3278 /* sk is a const pointer, because we want to express multiple
3279 * cpu might call us concurrently.
3280 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3281 */
3282 skb_set_owner_w(skb, (struct sock *)sk);
3283 break;
3284 }
3285 skb_dst_set(skb, dst);
3286
3287 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3288
3289 memset(&opts, 0, sizeof(opts));
3290 now = tcp_clock_ns();
3291#ifdef CONFIG_SYN_COOKIES
3292 if (unlikely(req->cookie_ts))
3293 skb->skb_mstamp_ns = cookie_init_timestamp(req);
3294 else
3295#endif
3296 {
3297 skb->skb_mstamp_ns = now;
3298 if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3299 tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3300 }
3301
3302#ifdef CONFIG_TCP_MD5SIG
3303 rcu_read_lock();
3304 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3305#endif
3306 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3307 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3308 foc) + sizeof(*th);
3309
3310 skb_push(skb, tcp_header_size);
3311 skb_reset_transport_header(skb);
3312
3313 th = (struct tcphdr *)skb->data;
3314 memset(th, 0, sizeof(struct tcphdr));
3315 th->syn = 1;
3316 th->ack = 1;
3317 tcp_ecn_make_synack(req, th);
3318 th->source = htons(ireq->ir_num);
3319 th->dest = ireq->ir_rmt_port;
3320 skb->mark = ireq->ir_mark;
3321 skb->ip_summed = CHECKSUM_PARTIAL;
3322 th->seq = htonl(tcp_rsk(req)->snt_isn);
3323 /* XXX data is queued and acked as is. No buffer/window check */
3324 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3325
3326 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3327 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3328 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3329 th->doff = (tcp_header_size >> 2);
3330 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3331
3332#ifdef CONFIG_TCP_MD5SIG
3333 /* Okay, we have all we need - do the md5 hash if needed */
3334 if (md5)
3335 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3336 md5, req_to_sk(req), skb);
3337 rcu_read_unlock();
3338#endif
3339
3340 skb->skb_mstamp_ns = now;
3341 tcp_add_tx_delay(skb, tp);
3342
3343 return skb;
3344}
3345EXPORT_SYMBOL(tcp_make_synack);
3346
3347static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3348{
3349 struct inet_connection_sock *icsk = inet_csk(sk);
3350 const struct tcp_congestion_ops *ca;
3351 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3352
3353 if (ca_key == TCP_CA_UNSPEC)
3354 return;
3355
3356 rcu_read_lock();
3357 ca = tcp_ca_find_key(ca_key);
3358 if (likely(ca && try_module_get(ca->owner))) {
3359 module_put(icsk->icsk_ca_ops->owner);
3360 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3361 icsk->icsk_ca_ops = ca;
3362 }
3363 rcu_read_unlock();
3364}
3365
3366/* Do all connect socket setups that can be done AF independent. */
3367static void tcp_connect_init(struct sock *sk)
3368{
3369 const struct dst_entry *dst = __sk_dst_get(sk);
3370 struct tcp_sock *tp = tcp_sk(sk);
3371 __u8 rcv_wscale;
3372 u32 rcv_wnd;
3373
3374 /* We'll fix this up when we get a response from the other end.
3375 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3376 */
3377 tp->tcp_header_len = sizeof(struct tcphdr);
3378 if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3379 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3380
3381#ifdef CONFIG_TCP_MD5SIG
3382 if (tp->af_specific->md5_lookup(sk, sk))
3383 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3384#endif
3385
3386 /* If user gave his TCP_MAXSEG, record it to clamp */
3387 if (tp->rx_opt.user_mss)
3388 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3389 tp->max_window = 0;
3390 tcp_mtup_init(sk);
3391 tcp_sync_mss(sk, dst_mtu(dst));
3392
3393 tcp_ca_dst_init(sk, dst);
3394
3395 if (!tp->window_clamp)
3396 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3397 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3398
3399 tcp_initialize_rcv_mss(sk);
3400
3401 /* limit the window selection if the user enforce a smaller rx buffer */
3402 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3403 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3404 tp->window_clamp = tcp_full_space(sk);
3405
3406 rcv_wnd = tcp_rwnd_init_bpf(sk);
3407 if (rcv_wnd == 0)
3408 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3409
3410 tcp_select_initial_window(sk, tcp_full_space(sk),
3411 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3412 &tp->rcv_wnd,
3413 &tp->window_clamp,
3414 sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3415 &rcv_wscale,
3416 rcv_wnd);
3417
3418 tp->rx_opt.rcv_wscale = rcv_wscale;
3419 tp->rcv_ssthresh = tp->rcv_wnd;
3420
3421 sk->sk_err = 0;
3422 sock_reset_flag(sk, SOCK_DONE);
3423 tp->snd_wnd = 0;
3424 tcp_init_wl(tp, 0);
3425 tcp_write_queue_purge(sk);
3426 tp->snd_una = tp->write_seq;
3427 tp->snd_sml = tp->write_seq;
3428 tp->snd_up = tp->write_seq;
3429 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3430
3431 if (likely(!tp->repair))
3432 tp->rcv_nxt = 0;
3433 else
3434 tp->rcv_tstamp = tcp_jiffies32;
3435 tp->rcv_wup = tp->rcv_nxt;
3436 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3437
3438 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3439 inet_csk(sk)->icsk_retransmits = 0;
3440 tcp_clear_retrans(tp);
3441}
3442
3443static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3444{
3445 struct tcp_sock *tp = tcp_sk(sk);
3446 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3447
3448 tcb->end_seq += skb->len;
3449 __skb_header_release(skb);
3450 sk_wmem_queued_add(sk, skb->truesize);
3451 sk_mem_charge(sk, skb->truesize);
3452 WRITE_ONCE(tp->write_seq, tcb->end_seq);
3453 tp->packets_out += tcp_skb_pcount(skb);
3454}
3455
3456/* Build and send a SYN with data and (cached) Fast Open cookie. However,
3457 * queue a data-only packet after the regular SYN, such that regular SYNs
3458 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3459 * only the SYN sequence, the data are retransmitted in the first ACK.
3460 * If cookie is not cached or other error occurs, falls back to send a
3461 * regular SYN with Fast Open cookie request option.
3462 */
3463static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3464{
3465 struct tcp_sock *tp = tcp_sk(sk);
3466 struct tcp_fastopen_request *fo = tp->fastopen_req;
3467 int space, err = 0;
3468 struct sk_buff *syn_data;
3469
3470 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3471 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3472 goto fallback;
3473
3474 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3475 * user-MSS. Reserve maximum option space for middleboxes that add
3476 * private TCP options. The cost is reduced data space in SYN :(
3477 */
3478 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3479
3480 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3481 MAX_TCP_OPTION_SPACE;
3482
3483 space = min_t(size_t, space, fo->size);
3484
3485 /* limit to order-0 allocations */
3486 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3487
3488 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3489 if (!syn_data)
3490 goto fallback;
3491 syn_data->ip_summed = CHECKSUM_PARTIAL;
3492 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3493 if (space) {
3494 int copied = copy_from_iter(skb_put(syn_data, space), space,
3495 &fo->data->msg_iter);
3496 if (unlikely(!copied)) {
3497 tcp_skb_tsorted_anchor_cleanup(syn_data);
3498 kfree_skb(syn_data);
3499 goto fallback;
3500 }
3501 if (copied != space) {
3502 skb_trim(syn_data, copied);
3503 space = copied;
3504 }
3505 skb_zcopy_set(syn_data, fo->uarg, NULL);
3506 }
3507 /* No more data pending in inet_wait_for_connect() */
3508 if (space == fo->size)
3509 fo->data = NULL;
3510 fo->copied = space;
3511
3512 tcp_connect_queue_skb(sk, syn_data);
3513 if (syn_data->len)
3514 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3515
3516 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3517
3518 syn->skb_mstamp_ns = syn_data->skb_mstamp_ns;
3519
3520 /* Now full SYN+DATA was cloned and sent (or not),
3521 * remove the SYN from the original skb (syn_data)
3522 * we keep in write queue in case of a retransmit, as we
3523 * also have the SYN packet (with no data) in the same queue.
3524 */
3525 TCP_SKB_CB(syn_data)->seq++;
3526 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3527 if (!err) {
3528 tp->syn_data = (fo->copied > 0);
3529 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3530 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3531 goto done;
3532 }
3533
3534 /* data was not sent, put it in write_queue */
3535 __skb_queue_tail(&sk->sk_write_queue, syn_data);
3536 tp->packets_out -= tcp_skb_pcount(syn_data);
3537
3538fallback:
3539 /* Send a regular SYN with Fast Open cookie request option */
3540 if (fo->cookie.len > 0)
3541 fo->cookie.len = 0;
3542 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3543 if (err)
3544 tp->syn_fastopen = 0;
3545done:
3546 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3547 return err;
3548}
3549
3550/* Build a SYN and send it off. */
3551int tcp_connect(struct sock *sk)
3552{
3553 struct tcp_sock *tp = tcp_sk(sk);
3554 struct sk_buff *buff;
3555 int err;
3556
3557 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3558
3559 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3560 return -EHOSTUNREACH; /* Routing failure or similar. */
3561
3562 tcp_connect_init(sk);
3563
3564 if (unlikely(tp->repair)) {
3565 tcp_finish_connect(sk, NULL);
3566 return 0;
3567 }
3568
3569 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3570 if (unlikely(!buff))
3571 return -ENOBUFS;
3572
3573 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3574 tcp_mstamp_refresh(tp);
3575 tp->retrans_stamp = tcp_time_stamp(tp);
3576 tcp_connect_queue_skb(sk, buff);
3577 tcp_ecn_send_syn(sk, buff);
3578 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3579
3580 /* Send off SYN; include data in Fast Open. */
3581 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3582 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3583 if (err == -ECONNREFUSED)
3584 return err;
3585
3586 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3587 * in order to make this packet get counted in tcpOutSegs.
3588 */
3589 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3590 tp->pushed_seq = tp->write_seq;
3591 buff = tcp_send_head(sk);
3592 if (unlikely(buff)) {
3593 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
3594 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3595 }
3596 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3597
3598 /* Timer for repeating the SYN until an answer. */
3599 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3600 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3601 return 0;
3602}
3603EXPORT_SYMBOL(tcp_connect);
3604
3605/* Send out a delayed ack, the caller does the policy checking
3606 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3607 * for details.
3608 */
3609void tcp_send_delayed_ack(struct sock *sk)
3610{
3611 struct inet_connection_sock *icsk = inet_csk(sk);
3612 int ato = icsk->icsk_ack.ato;
3613 unsigned long timeout;
3614
3615 if (ato > TCP_DELACK_MIN) {
3616 const struct tcp_sock *tp = tcp_sk(sk);
3617 int max_ato = HZ / 2;
3618
3619 if (inet_csk_in_pingpong_mode(sk) ||
3620 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3621 max_ato = TCP_DELACK_MAX;
3622
3623 /* Slow path, intersegment interval is "high". */
3624
3625 /* If some rtt estimate is known, use it to bound delayed ack.
3626 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3627 * directly.
3628 */
3629 if (tp->srtt_us) {
3630 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3631 TCP_DELACK_MIN);
3632
3633 if (rtt < max_ato)
3634 max_ato = rtt;
3635 }
3636
3637 ato = min(ato, max_ato);
3638 }
3639
3640 /* Stay within the limit we were given */
3641 timeout = jiffies + ato;
3642
3643 /* Use new timeout only if there wasn't a older one earlier. */
3644 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3645 /* If delack timer was blocked or is about to expire,
3646 * send ACK now.
3647 */
3648 if (icsk->icsk_ack.blocked ||
3649 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3650 tcp_send_ack(sk);
3651 return;
3652 }
3653
3654 if (!time_before(timeout, icsk->icsk_ack.timeout))
3655 timeout = icsk->icsk_ack.timeout;
3656 }
3657 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3658 icsk->icsk_ack.timeout = timeout;
3659 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3660}
3661
3662/* This routine sends an ack and also updates the window. */
3663void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3664{
3665 struct sk_buff *buff;
3666
3667 /* If we have been reset, we may not send again. */
3668 if (sk->sk_state == TCP_CLOSE)
3669 return;
3670
3671 /* We are not putting this on the write queue, so
3672 * tcp_transmit_skb() will set the ownership to this
3673 * sock.
3674 */
3675 buff = alloc_skb(MAX_TCP_HEADER,
3676 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3677 if (unlikely(!buff)) {
3678 inet_csk_schedule_ack(sk);
3679 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3680 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3681 TCP_DELACK_MAX, TCP_RTO_MAX);
3682 return;
3683 }
3684
3685 /* Reserve space for headers and prepare control bits. */
3686 skb_reserve(buff, MAX_TCP_HEADER);
3687 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3688
3689 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3690 * too much.
3691 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3692 */
3693 skb_set_tcp_pure_ack(buff);
3694
3695 /* Send it off, this clears delayed acks for us. */
3696 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3697}
3698EXPORT_SYMBOL_GPL(__tcp_send_ack);
3699
3700void tcp_send_ack(struct sock *sk)
3701{
3702 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3703}
3704
3705/* This routine sends a packet with an out of date sequence
3706 * number. It assumes the other end will try to ack it.
3707 *
3708 * Question: what should we make while urgent mode?
3709 * 4.4BSD forces sending single byte of data. We cannot send
3710 * out of window data, because we have SND.NXT==SND.MAX...
3711 *
3712 * Current solution: to send TWO zero-length segments in urgent mode:
3713 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3714 * out-of-date with SND.UNA-1 to probe window.
3715 */
3716static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3717{
3718 struct tcp_sock *tp = tcp_sk(sk);
3719 struct sk_buff *skb;
3720
3721 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3722 skb = alloc_skb(MAX_TCP_HEADER,
3723 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3724 if (!skb)
3725 return -1;
3726
3727 /* Reserve space for headers and set control bits. */
3728 skb_reserve(skb, MAX_TCP_HEADER);
3729 /* Use a previous sequence. This should cause the other
3730 * end to send an ack. Don't queue or clone SKB, just
3731 * send it.
3732 */
3733 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3734 NET_INC_STATS(sock_net(sk), mib);
3735 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3736}
3737
3738/* Called from setsockopt( ... TCP_REPAIR ) */
3739void tcp_send_window_probe(struct sock *sk)
3740{
3741 if (sk->sk_state == TCP_ESTABLISHED) {
3742 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3743 tcp_mstamp_refresh(tcp_sk(sk));
3744 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3745 }
3746}
3747
3748/* Initiate keepalive or window probe from timer. */
3749int tcp_write_wakeup(struct sock *sk, int mib)
3750{
3751 struct tcp_sock *tp = tcp_sk(sk);
3752 struct sk_buff *skb;
3753
3754 if (sk->sk_state == TCP_CLOSE)
3755 return -1;
3756
3757 skb = tcp_send_head(sk);
3758 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3759 int err;
3760 unsigned int mss = tcp_current_mss(sk);
3761 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3762
3763 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3764 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3765
3766 /* We are probing the opening of a window
3767 * but the window size is != 0
3768 * must have been a result SWS avoidance ( sender )
3769 */
3770 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3771 skb->len > mss) {
3772 seg_size = min(seg_size, mss);
3773 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3774 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
3775 skb, seg_size, mss, GFP_ATOMIC))
3776 return -1;
3777 } else if (!tcp_skb_pcount(skb))
3778 tcp_set_skb_tso_segs(skb, mss);
3779
3780 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3781 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3782 if (!err)
3783 tcp_event_new_data_sent(sk, skb);
3784 return err;
3785 } else {
3786 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3787 tcp_xmit_probe_skb(sk, 1, mib);
3788 return tcp_xmit_probe_skb(sk, 0, mib);
3789 }
3790}
3791
3792/* A window probe timeout has occurred. If window is not closed send
3793 * a partial packet else a zero probe.
3794 */
3795void tcp_send_probe0(struct sock *sk)
3796{
3797 struct inet_connection_sock *icsk = inet_csk(sk);
3798 struct tcp_sock *tp = tcp_sk(sk);
3799 struct net *net = sock_net(sk);
3800 unsigned long timeout;
3801 int err;
3802
3803 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3804
3805 if (tp->packets_out || tcp_write_queue_empty(sk)) {
3806 /* Cancel probe timer, if it is not required. */
3807 icsk->icsk_probes_out = 0;
3808 icsk->icsk_backoff = 0;
3809 return;
3810 }
3811
3812 icsk->icsk_probes_out++;
3813 if (err <= 0) {
3814 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3815 icsk->icsk_backoff++;
3816 timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
3817 } else {
3818 /* If packet was not sent due to local congestion,
3819 * Let senders fight for local resources conservatively.
3820 */
3821 timeout = TCP_RESOURCE_PROBE_INTERVAL;
3822 }
3823 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX, NULL);
3824}
3825
3826int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3827{
3828 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3829 struct flowi fl;
3830 int res;
3831
3832 tcp_rsk(req)->txhash = net_tx_rndhash();
3833 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3834 if (!res) {
3835 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3836 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3837 if (unlikely(tcp_passive_fastopen(sk)))
3838 tcp_sk(sk)->total_retrans++;
3839 trace_tcp_retransmit_synack(sk, req);
3840 }
3841 return res;
3842}
3843EXPORT_SYMBOL(tcp_rtx_synack);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
9 * Authors: Ross Biro
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
20 */
21
22/*
23 * Changes: Pedro Roque : Retransmit queue handled by TCP.
24 * : Fragmentation on mtu decrease
25 * : Segment collapse on retransmit
26 * : AF independence
27 *
28 * Linus Torvalds : send_delayed_ack
29 * David S. Miller : Charge memory using the right skb
30 * during syn/ack processing.
31 * David S. Miller : Output engine completely rewritten.
32 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
33 * Cacophonix Gaul : draft-minshall-nagle-01
34 * J Hadi Salim : ECN support
35 *
36 */
37
38#define pr_fmt(fmt) "TCP: " fmt
39
40#include <net/tcp.h>
41#include <net/mptcp.h>
42
43#include <linux/compiler.h>
44#include <linux/gfp.h>
45#include <linux/module.h>
46#include <linux/static_key.h>
47
48#include <trace/events/tcp.h>
49
50/* Refresh clocks of a TCP socket,
51 * ensuring monotically increasing values.
52 */
53void tcp_mstamp_refresh(struct tcp_sock *tp)
54{
55 u64 val = tcp_clock_ns();
56
57 tp->tcp_clock_cache = val;
58 tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC);
59}
60
61static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
62 int push_one, gfp_t gfp);
63
64/* Account for new data that has been sent to the network. */
65static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
66{
67 struct inet_connection_sock *icsk = inet_csk(sk);
68 struct tcp_sock *tp = tcp_sk(sk);
69 unsigned int prior_packets = tp->packets_out;
70
71 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq);
72
73 __skb_unlink(skb, &sk->sk_write_queue);
74 tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
75
76 if (tp->highest_sack == NULL)
77 tp->highest_sack = skb;
78
79 tp->packets_out += tcp_skb_pcount(skb);
80 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
81 tcp_rearm_rto(sk);
82
83 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
84 tcp_skb_pcount(skb));
85 tcp_check_space(sk);
86}
87
88/* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
89 * window scaling factor due to loss of precision.
90 * If window has been shrunk, what should we make? It is not clear at all.
91 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
92 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
93 * invalid. OK, let's make this for now:
94 */
95static inline __u32 tcp_acceptable_seq(const struct sock *sk)
96{
97 const struct tcp_sock *tp = tcp_sk(sk);
98
99 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
100 (tp->rx_opt.wscale_ok &&
101 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
102 return tp->snd_nxt;
103 else
104 return tcp_wnd_end(tp);
105}
106
107/* Calculate mss to advertise in SYN segment.
108 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
109 *
110 * 1. It is independent of path mtu.
111 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
112 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
113 * attached devices, because some buggy hosts are confused by
114 * large MSS.
115 * 4. We do not make 3, we advertise MSS, calculated from first
116 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
117 * This may be overridden via information stored in routing table.
118 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
119 * probably even Jumbo".
120 */
121static __u16 tcp_advertise_mss(struct sock *sk)
122{
123 struct tcp_sock *tp = tcp_sk(sk);
124 const struct dst_entry *dst = __sk_dst_get(sk);
125 int mss = tp->advmss;
126
127 if (dst) {
128 unsigned int metric = dst_metric_advmss(dst);
129
130 if (metric < mss) {
131 mss = metric;
132 tp->advmss = mss;
133 }
134 }
135
136 return (__u16)mss;
137}
138
139/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
140 * This is the first part of cwnd validation mechanism.
141 */
142void tcp_cwnd_restart(struct sock *sk, s32 delta)
143{
144 struct tcp_sock *tp = tcp_sk(sk);
145 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
146 u32 cwnd = tcp_snd_cwnd(tp);
147
148 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
149
150 tp->snd_ssthresh = tcp_current_ssthresh(sk);
151 restart_cwnd = min(restart_cwnd, cwnd);
152
153 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
154 cwnd >>= 1;
155 tcp_snd_cwnd_set(tp, max(cwnd, restart_cwnd));
156 tp->snd_cwnd_stamp = tcp_jiffies32;
157 tp->snd_cwnd_used = 0;
158}
159
160/* Congestion state accounting after a packet has been sent. */
161static void tcp_event_data_sent(struct tcp_sock *tp,
162 struct sock *sk)
163{
164 struct inet_connection_sock *icsk = inet_csk(sk);
165 const u32 now = tcp_jiffies32;
166
167 if (tcp_packets_in_flight(tp) == 0)
168 tcp_ca_event(sk, CA_EVENT_TX_START);
169
170 tp->lsndtime = now;
171
172 /* If it is a reply for ato after last received
173 * packet, enter pingpong mode.
174 */
175 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
176 inet_csk_enter_pingpong_mode(sk);
177}
178
179/* Account for an ACK we sent. */
180static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
181 u32 rcv_nxt)
182{
183 struct tcp_sock *tp = tcp_sk(sk);
184
185 if (unlikely(tp->compressed_ack)) {
186 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
187 tp->compressed_ack);
188 tp->compressed_ack = 0;
189 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
190 __sock_put(sk);
191 }
192
193 if (unlikely(rcv_nxt != tp->rcv_nxt))
194 return; /* Special ACK sent by DCTCP to reflect ECN */
195 tcp_dec_quickack_mode(sk, pkts);
196 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
197}
198
199/* Determine a window scaling and initial window to offer.
200 * Based on the assumption that the given amount of space
201 * will be offered. Store the results in the tp structure.
202 * NOTE: for smooth operation initial space offering should
203 * be a multiple of mss if possible. We assume here that mss >= 1.
204 * This MUST be enforced by all callers.
205 */
206void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
207 __u32 *rcv_wnd, __u32 *window_clamp,
208 int wscale_ok, __u8 *rcv_wscale,
209 __u32 init_rcv_wnd)
210{
211 unsigned int space = (__space < 0 ? 0 : __space);
212
213 /* If no clamp set the clamp to the max possible scaled window */
214 if (*window_clamp == 0)
215 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
216 space = min(*window_clamp, space);
217
218 /* Quantize space offering to a multiple of mss if possible. */
219 if (space > mss)
220 space = rounddown(space, mss);
221
222 /* NOTE: offering an initial window larger than 32767
223 * will break some buggy TCP stacks. If the admin tells us
224 * it is likely we could be speaking with such a buggy stack
225 * we will truncate our initial window offering to 32K-1
226 * unless the remote has sent us a window scaling option,
227 * which we interpret as a sign the remote TCP is not
228 * misinterpreting the window field as a signed quantity.
229 */
230 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows))
231 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
232 else
233 (*rcv_wnd) = min_t(u32, space, U16_MAX);
234
235 if (init_rcv_wnd)
236 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
237
238 *rcv_wscale = 0;
239 if (wscale_ok) {
240 /* Set window scaling on max possible window */
241 space = max_t(u32, space, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
242 space = max_t(u32, space, READ_ONCE(sysctl_rmem_max));
243 space = min_t(u32, space, *window_clamp);
244 *rcv_wscale = clamp_t(int, ilog2(space) - 15,
245 0, TCP_MAX_WSCALE);
246 }
247 /* Set the clamp no higher than max representable value */
248 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
249}
250EXPORT_SYMBOL(tcp_select_initial_window);
251
252/* Chose a new window to advertise, update state in tcp_sock for the
253 * socket, and return result with RFC1323 scaling applied. The return
254 * value can be stuffed directly into th->window for an outgoing
255 * frame.
256 */
257static u16 tcp_select_window(struct sock *sk)
258{
259 struct tcp_sock *tp = tcp_sk(sk);
260 u32 old_win = tp->rcv_wnd;
261 u32 cur_win = tcp_receive_window(tp);
262 u32 new_win = __tcp_select_window(sk);
263
264 /* Never shrink the offered window */
265 if (new_win < cur_win) {
266 /* Danger Will Robinson!
267 * Don't update rcv_wup/rcv_wnd here or else
268 * we will not be able to advertise a zero
269 * window in time. --DaveM
270 *
271 * Relax Will Robinson.
272 */
273 if (new_win == 0)
274 NET_INC_STATS(sock_net(sk),
275 LINUX_MIB_TCPWANTZEROWINDOWADV);
276 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
277 }
278 tp->rcv_wnd = new_win;
279 tp->rcv_wup = tp->rcv_nxt;
280
281 /* Make sure we do not exceed the maximum possible
282 * scaled window.
283 */
284 if (!tp->rx_opt.rcv_wscale &&
285 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows))
286 new_win = min(new_win, MAX_TCP_WINDOW);
287 else
288 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
289
290 /* RFC1323 scaling applied */
291 new_win >>= tp->rx_opt.rcv_wscale;
292
293 /* If we advertise zero window, disable fast path. */
294 if (new_win == 0) {
295 tp->pred_flags = 0;
296 if (old_win)
297 NET_INC_STATS(sock_net(sk),
298 LINUX_MIB_TCPTOZEROWINDOWADV);
299 } else if (old_win == 0) {
300 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
301 }
302
303 return new_win;
304}
305
306/* Packet ECN state for a SYN-ACK */
307static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
308{
309 const struct tcp_sock *tp = tcp_sk(sk);
310
311 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
312 if (!(tp->ecn_flags & TCP_ECN_OK))
313 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
314 else if (tcp_ca_needs_ecn(sk) ||
315 tcp_bpf_ca_needs_ecn(sk))
316 INET_ECN_xmit(sk);
317}
318
319/* Packet ECN state for a SYN. */
320static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
321{
322 struct tcp_sock *tp = tcp_sk(sk);
323 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
324 bool use_ecn = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn) == 1 ||
325 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
326
327 if (!use_ecn) {
328 const struct dst_entry *dst = __sk_dst_get(sk);
329
330 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
331 use_ecn = true;
332 }
333
334 tp->ecn_flags = 0;
335
336 if (use_ecn) {
337 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
338 tp->ecn_flags = TCP_ECN_OK;
339 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
340 INET_ECN_xmit(sk);
341 }
342}
343
344static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
345{
346 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback))
347 /* tp->ecn_flags are cleared at a later point in time when
348 * SYN ACK is ultimatively being received.
349 */
350 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
351}
352
353static void
354tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
355{
356 if (inet_rsk(req)->ecn_ok)
357 th->ece = 1;
358}
359
360/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
361 * be sent.
362 */
363static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
364 struct tcphdr *th, int tcp_header_len)
365{
366 struct tcp_sock *tp = tcp_sk(sk);
367
368 if (tp->ecn_flags & TCP_ECN_OK) {
369 /* Not-retransmitted data segment: set ECT and inject CWR. */
370 if (skb->len != tcp_header_len &&
371 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
372 INET_ECN_xmit(sk);
373 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
374 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
375 th->cwr = 1;
376 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
377 }
378 } else if (!tcp_ca_needs_ecn(sk)) {
379 /* ACK or retransmitted segment: clear ECT|CE */
380 INET_ECN_dontxmit(sk);
381 }
382 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
383 th->ece = 1;
384 }
385}
386
387/* Constructs common control bits of non-data skb. If SYN/FIN is present,
388 * auto increment end seqno.
389 */
390static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
391{
392 skb->ip_summed = CHECKSUM_PARTIAL;
393
394 TCP_SKB_CB(skb)->tcp_flags = flags;
395
396 tcp_skb_pcount_set(skb, 1);
397
398 TCP_SKB_CB(skb)->seq = seq;
399 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
400 seq++;
401 TCP_SKB_CB(skb)->end_seq = seq;
402}
403
404static inline bool tcp_urg_mode(const struct tcp_sock *tp)
405{
406 return tp->snd_una != tp->snd_up;
407}
408
409#define OPTION_SACK_ADVERTISE BIT(0)
410#define OPTION_TS BIT(1)
411#define OPTION_MD5 BIT(2)
412#define OPTION_WSCALE BIT(3)
413#define OPTION_FAST_OPEN_COOKIE BIT(8)
414#define OPTION_SMC BIT(9)
415#define OPTION_MPTCP BIT(10)
416
417static void smc_options_write(__be32 *ptr, u16 *options)
418{
419#if IS_ENABLED(CONFIG_SMC)
420 if (static_branch_unlikely(&tcp_have_smc)) {
421 if (unlikely(OPTION_SMC & *options)) {
422 *ptr++ = htonl((TCPOPT_NOP << 24) |
423 (TCPOPT_NOP << 16) |
424 (TCPOPT_EXP << 8) |
425 (TCPOLEN_EXP_SMC_BASE));
426 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
427 }
428 }
429#endif
430}
431
432struct tcp_out_options {
433 u16 options; /* bit field of OPTION_* */
434 u16 mss; /* 0 to disable */
435 u8 ws; /* window scale, 0 to disable */
436 u8 num_sack_blocks; /* number of SACK blocks to include */
437 u8 hash_size; /* bytes in hash_location */
438 u8 bpf_opt_len; /* length of BPF hdr option */
439 __u8 *hash_location; /* temporary pointer, overloaded */
440 __u32 tsval, tsecr; /* need to include OPTION_TS */
441 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
442 struct mptcp_out_options mptcp;
443};
444
445static void mptcp_options_write(struct tcphdr *th, __be32 *ptr,
446 struct tcp_sock *tp,
447 struct tcp_out_options *opts)
448{
449#if IS_ENABLED(CONFIG_MPTCP)
450 if (unlikely(OPTION_MPTCP & opts->options))
451 mptcp_write_options(th, ptr, tp, &opts->mptcp);
452#endif
453}
454
455#ifdef CONFIG_CGROUP_BPF
456static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb,
457 enum tcp_synack_type synack_type)
458{
459 if (unlikely(!skb))
460 return BPF_WRITE_HDR_TCP_CURRENT_MSS;
461
462 if (unlikely(synack_type == TCP_SYNACK_COOKIE))
463 return BPF_WRITE_HDR_TCP_SYNACK_COOKIE;
464
465 return 0;
466}
467
468/* req, syn_skb and synack_type are used when writing synack */
469static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
470 struct request_sock *req,
471 struct sk_buff *syn_skb,
472 enum tcp_synack_type synack_type,
473 struct tcp_out_options *opts,
474 unsigned int *remaining)
475{
476 struct bpf_sock_ops_kern sock_ops;
477 int err;
478
479 if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
480 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) ||
481 !*remaining)
482 return;
483
484 /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */
485
486 /* init sock_ops */
487 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
488
489 sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB;
490
491 if (req) {
492 /* The listen "sk" cannot be passed here because
493 * it is not locked. It would not make too much
494 * sense to do bpf_setsockopt(listen_sk) based
495 * on individual connection request also.
496 *
497 * Thus, "req" is passed here and the cgroup-bpf-progs
498 * of the listen "sk" will be run.
499 *
500 * "req" is also used here for fastopen even the "sk" here is
501 * a fullsock "child" sk. It is to keep the behavior
502 * consistent between fastopen and non-fastopen on
503 * the bpf programming side.
504 */
505 sock_ops.sk = (struct sock *)req;
506 sock_ops.syn_skb = syn_skb;
507 } else {
508 sock_owned_by_me(sk);
509
510 sock_ops.is_fullsock = 1;
511 sock_ops.sk = sk;
512 }
513
514 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
515 sock_ops.remaining_opt_len = *remaining;
516 /* tcp_current_mss() does not pass a skb */
517 if (skb)
518 bpf_skops_init_skb(&sock_ops, skb, 0);
519
520 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
521
522 if (err || sock_ops.remaining_opt_len == *remaining)
523 return;
524
525 opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len;
526 /* round up to 4 bytes */
527 opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3;
528
529 *remaining -= opts->bpf_opt_len;
530}
531
532static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
533 struct request_sock *req,
534 struct sk_buff *syn_skb,
535 enum tcp_synack_type synack_type,
536 struct tcp_out_options *opts)
537{
538 u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len;
539 struct bpf_sock_ops_kern sock_ops;
540 int err;
541
542 if (likely(!max_opt_len))
543 return;
544
545 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
546
547 sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB;
548
549 if (req) {
550 sock_ops.sk = (struct sock *)req;
551 sock_ops.syn_skb = syn_skb;
552 } else {
553 sock_owned_by_me(sk);
554
555 sock_ops.is_fullsock = 1;
556 sock_ops.sk = sk;
557 }
558
559 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
560 sock_ops.remaining_opt_len = max_opt_len;
561 first_opt_off = tcp_hdrlen(skb) - max_opt_len;
562 bpf_skops_init_skb(&sock_ops, skb, first_opt_off);
563
564 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
565
566 if (err)
567 nr_written = 0;
568 else
569 nr_written = max_opt_len - sock_ops.remaining_opt_len;
570
571 if (nr_written < max_opt_len)
572 memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP,
573 max_opt_len - nr_written);
574}
575#else
576static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
577 struct request_sock *req,
578 struct sk_buff *syn_skb,
579 enum tcp_synack_type synack_type,
580 struct tcp_out_options *opts,
581 unsigned int *remaining)
582{
583}
584
585static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
586 struct request_sock *req,
587 struct sk_buff *syn_skb,
588 enum tcp_synack_type synack_type,
589 struct tcp_out_options *opts)
590{
591}
592#endif
593
594/* Write previously computed TCP options to the packet.
595 *
596 * Beware: Something in the Internet is very sensitive to the ordering of
597 * TCP options, we learned this through the hard way, so be careful here.
598 * Luckily we can at least blame others for their non-compliance but from
599 * inter-operability perspective it seems that we're somewhat stuck with
600 * the ordering which we have been using if we want to keep working with
601 * those broken things (not that it currently hurts anybody as there isn't
602 * particular reason why the ordering would need to be changed).
603 *
604 * At least SACK_PERM as the first option is known to lead to a disaster
605 * (but it may well be that other scenarios fail similarly).
606 */
607static void tcp_options_write(struct tcphdr *th, struct tcp_sock *tp,
608 struct tcp_out_options *opts)
609{
610 __be32 *ptr = (__be32 *)(th + 1);
611 u16 options = opts->options; /* mungable copy */
612
613 if (unlikely(OPTION_MD5 & options)) {
614 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
615 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
616 /* overload cookie hash location */
617 opts->hash_location = (__u8 *)ptr;
618 ptr += 4;
619 }
620
621 if (unlikely(opts->mss)) {
622 *ptr++ = htonl((TCPOPT_MSS << 24) |
623 (TCPOLEN_MSS << 16) |
624 opts->mss);
625 }
626
627 if (likely(OPTION_TS & options)) {
628 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
629 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
630 (TCPOLEN_SACK_PERM << 16) |
631 (TCPOPT_TIMESTAMP << 8) |
632 TCPOLEN_TIMESTAMP);
633 options &= ~OPTION_SACK_ADVERTISE;
634 } else {
635 *ptr++ = htonl((TCPOPT_NOP << 24) |
636 (TCPOPT_NOP << 16) |
637 (TCPOPT_TIMESTAMP << 8) |
638 TCPOLEN_TIMESTAMP);
639 }
640 *ptr++ = htonl(opts->tsval);
641 *ptr++ = htonl(opts->tsecr);
642 }
643
644 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
645 *ptr++ = htonl((TCPOPT_NOP << 24) |
646 (TCPOPT_NOP << 16) |
647 (TCPOPT_SACK_PERM << 8) |
648 TCPOLEN_SACK_PERM);
649 }
650
651 if (unlikely(OPTION_WSCALE & options)) {
652 *ptr++ = htonl((TCPOPT_NOP << 24) |
653 (TCPOPT_WINDOW << 16) |
654 (TCPOLEN_WINDOW << 8) |
655 opts->ws);
656 }
657
658 if (unlikely(opts->num_sack_blocks)) {
659 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
660 tp->duplicate_sack : tp->selective_acks;
661 int this_sack;
662
663 *ptr++ = htonl((TCPOPT_NOP << 24) |
664 (TCPOPT_NOP << 16) |
665 (TCPOPT_SACK << 8) |
666 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
667 TCPOLEN_SACK_PERBLOCK)));
668
669 for (this_sack = 0; this_sack < opts->num_sack_blocks;
670 ++this_sack) {
671 *ptr++ = htonl(sp[this_sack].start_seq);
672 *ptr++ = htonl(sp[this_sack].end_seq);
673 }
674
675 tp->rx_opt.dsack = 0;
676 }
677
678 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
679 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
680 u8 *p = (u8 *)ptr;
681 u32 len; /* Fast Open option length */
682
683 if (foc->exp) {
684 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
685 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
686 TCPOPT_FASTOPEN_MAGIC);
687 p += TCPOLEN_EXP_FASTOPEN_BASE;
688 } else {
689 len = TCPOLEN_FASTOPEN_BASE + foc->len;
690 *p++ = TCPOPT_FASTOPEN;
691 *p++ = len;
692 }
693
694 memcpy(p, foc->val, foc->len);
695 if ((len & 3) == 2) {
696 p[foc->len] = TCPOPT_NOP;
697 p[foc->len + 1] = TCPOPT_NOP;
698 }
699 ptr += (len + 3) >> 2;
700 }
701
702 smc_options_write(ptr, &options);
703
704 mptcp_options_write(th, ptr, tp, opts);
705}
706
707static void smc_set_option(const struct tcp_sock *tp,
708 struct tcp_out_options *opts,
709 unsigned int *remaining)
710{
711#if IS_ENABLED(CONFIG_SMC)
712 if (static_branch_unlikely(&tcp_have_smc)) {
713 if (tp->syn_smc) {
714 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
715 opts->options |= OPTION_SMC;
716 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
717 }
718 }
719 }
720#endif
721}
722
723static void smc_set_option_cond(const struct tcp_sock *tp,
724 const struct inet_request_sock *ireq,
725 struct tcp_out_options *opts,
726 unsigned int *remaining)
727{
728#if IS_ENABLED(CONFIG_SMC)
729 if (static_branch_unlikely(&tcp_have_smc)) {
730 if (tp->syn_smc && ireq->smc_ok) {
731 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
732 opts->options |= OPTION_SMC;
733 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
734 }
735 }
736 }
737#endif
738}
739
740static void mptcp_set_option_cond(const struct request_sock *req,
741 struct tcp_out_options *opts,
742 unsigned int *remaining)
743{
744 if (rsk_is_mptcp(req)) {
745 unsigned int size;
746
747 if (mptcp_synack_options(req, &size, &opts->mptcp)) {
748 if (*remaining >= size) {
749 opts->options |= OPTION_MPTCP;
750 *remaining -= size;
751 }
752 }
753 }
754}
755
756/* Compute TCP options for SYN packets. This is not the final
757 * network wire format yet.
758 */
759static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
760 struct tcp_out_options *opts,
761 struct tcp_md5sig_key **md5)
762{
763 struct tcp_sock *tp = tcp_sk(sk);
764 unsigned int remaining = MAX_TCP_OPTION_SPACE;
765 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
766
767 *md5 = NULL;
768#ifdef CONFIG_TCP_MD5SIG
769 if (static_branch_unlikely(&tcp_md5_needed.key) &&
770 rcu_access_pointer(tp->md5sig_info)) {
771 *md5 = tp->af_specific->md5_lookup(sk, sk);
772 if (*md5) {
773 opts->options |= OPTION_MD5;
774 remaining -= TCPOLEN_MD5SIG_ALIGNED;
775 }
776 }
777#endif
778
779 /* We always get an MSS option. The option bytes which will be seen in
780 * normal data packets should timestamps be used, must be in the MSS
781 * advertised. But we subtract them from tp->mss_cache so that
782 * calculations in tcp_sendmsg are simpler etc. So account for this
783 * fact here if necessary. If we don't do this correctly, as a
784 * receiver we won't recognize data packets as being full sized when we
785 * should, and thus we won't abide by the delayed ACK rules correctly.
786 * SACKs don't matter, we never delay an ACK when we have any of those
787 * going out. */
788 opts->mss = tcp_advertise_mss(sk);
789 remaining -= TCPOLEN_MSS_ALIGNED;
790
791 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps) && !*md5)) {
792 opts->options |= OPTION_TS;
793 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
794 opts->tsecr = tp->rx_opt.ts_recent;
795 remaining -= TCPOLEN_TSTAMP_ALIGNED;
796 }
797 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling))) {
798 opts->ws = tp->rx_opt.rcv_wscale;
799 opts->options |= OPTION_WSCALE;
800 remaining -= TCPOLEN_WSCALE_ALIGNED;
801 }
802 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_sack))) {
803 opts->options |= OPTION_SACK_ADVERTISE;
804 if (unlikely(!(OPTION_TS & opts->options)))
805 remaining -= TCPOLEN_SACKPERM_ALIGNED;
806 }
807
808 if (fastopen && fastopen->cookie.len >= 0) {
809 u32 need = fastopen->cookie.len;
810
811 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
812 TCPOLEN_FASTOPEN_BASE;
813 need = (need + 3) & ~3U; /* Align to 32 bits */
814 if (remaining >= need) {
815 opts->options |= OPTION_FAST_OPEN_COOKIE;
816 opts->fastopen_cookie = &fastopen->cookie;
817 remaining -= need;
818 tp->syn_fastopen = 1;
819 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
820 }
821 }
822
823 smc_set_option(tp, opts, &remaining);
824
825 if (sk_is_mptcp(sk)) {
826 unsigned int size;
827
828 if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) {
829 opts->options |= OPTION_MPTCP;
830 remaining -= size;
831 }
832 }
833
834 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
835
836 return MAX_TCP_OPTION_SPACE - remaining;
837}
838
839/* Set up TCP options for SYN-ACKs. */
840static unsigned int tcp_synack_options(const struct sock *sk,
841 struct request_sock *req,
842 unsigned int mss, struct sk_buff *skb,
843 struct tcp_out_options *opts,
844 const struct tcp_md5sig_key *md5,
845 struct tcp_fastopen_cookie *foc,
846 enum tcp_synack_type synack_type,
847 struct sk_buff *syn_skb)
848{
849 struct inet_request_sock *ireq = inet_rsk(req);
850 unsigned int remaining = MAX_TCP_OPTION_SPACE;
851
852#ifdef CONFIG_TCP_MD5SIG
853 if (md5) {
854 opts->options |= OPTION_MD5;
855 remaining -= TCPOLEN_MD5SIG_ALIGNED;
856
857 /* We can't fit any SACK blocks in a packet with MD5 + TS
858 * options. There was discussion about disabling SACK
859 * rather than TS in order to fit in better with old,
860 * buggy kernels, but that was deemed to be unnecessary.
861 */
862 if (synack_type != TCP_SYNACK_COOKIE)
863 ireq->tstamp_ok &= !ireq->sack_ok;
864 }
865#endif
866
867 /* We always send an MSS option. */
868 opts->mss = mss;
869 remaining -= TCPOLEN_MSS_ALIGNED;
870
871 if (likely(ireq->wscale_ok)) {
872 opts->ws = ireq->rcv_wscale;
873 opts->options |= OPTION_WSCALE;
874 remaining -= TCPOLEN_WSCALE_ALIGNED;
875 }
876 if (likely(ireq->tstamp_ok)) {
877 opts->options |= OPTION_TS;
878 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
879 opts->tsecr = req->ts_recent;
880 remaining -= TCPOLEN_TSTAMP_ALIGNED;
881 }
882 if (likely(ireq->sack_ok)) {
883 opts->options |= OPTION_SACK_ADVERTISE;
884 if (unlikely(!ireq->tstamp_ok))
885 remaining -= TCPOLEN_SACKPERM_ALIGNED;
886 }
887 if (foc != NULL && foc->len >= 0) {
888 u32 need = foc->len;
889
890 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
891 TCPOLEN_FASTOPEN_BASE;
892 need = (need + 3) & ~3U; /* Align to 32 bits */
893 if (remaining >= need) {
894 opts->options |= OPTION_FAST_OPEN_COOKIE;
895 opts->fastopen_cookie = foc;
896 remaining -= need;
897 }
898 }
899
900 mptcp_set_option_cond(req, opts, &remaining);
901
902 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
903
904 bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb,
905 synack_type, opts, &remaining);
906
907 return MAX_TCP_OPTION_SPACE - remaining;
908}
909
910/* Compute TCP options for ESTABLISHED sockets. This is not the
911 * final wire format yet.
912 */
913static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
914 struct tcp_out_options *opts,
915 struct tcp_md5sig_key **md5)
916{
917 struct tcp_sock *tp = tcp_sk(sk);
918 unsigned int size = 0;
919 unsigned int eff_sacks;
920
921 opts->options = 0;
922
923 *md5 = NULL;
924#ifdef CONFIG_TCP_MD5SIG
925 if (static_branch_unlikely(&tcp_md5_needed.key) &&
926 rcu_access_pointer(tp->md5sig_info)) {
927 *md5 = tp->af_specific->md5_lookup(sk, sk);
928 if (*md5) {
929 opts->options |= OPTION_MD5;
930 size += TCPOLEN_MD5SIG_ALIGNED;
931 }
932 }
933#endif
934
935 if (likely(tp->rx_opt.tstamp_ok)) {
936 opts->options |= OPTION_TS;
937 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
938 opts->tsecr = tp->rx_opt.ts_recent;
939 size += TCPOLEN_TSTAMP_ALIGNED;
940 }
941
942 /* MPTCP options have precedence over SACK for the limited TCP
943 * option space because a MPTCP connection would be forced to
944 * fall back to regular TCP if a required multipath option is
945 * missing. SACK still gets a chance to use whatever space is
946 * left.
947 */
948 if (sk_is_mptcp(sk)) {
949 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
950 unsigned int opt_size = 0;
951
952 if (mptcp_established_options(sk, skb, &opt_size, remaining,
953 &opts->mptcp)) {
954 opts->options |= OPTION_MPTCP;
955 size += opt_size;
956 }
957 }
958
959 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
960 if (unlikely(eff_sacks)) {
961 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
962 if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED +
963 TCPOLEN_SACK_PERBLOCK))
964 return size;
965
966 opts->num_sack_blocks =
967 min_t(unsigned int, eff_sacks,
968 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
969 TCPOLEN_SACK_PERBLOCK);
970
971 size += TCPOLEN_SACK_BASE_ALIGNED +
972 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
973 }
974
975 if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp,
976 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) {
977 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
978
979 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
980
981 size = MAX_TCP_OPTION_SPACE - remaining;
982 }
983
984 return size;
985}
986
987
988/* TCP SMALL QUEUES (TSQ)
989 *
990 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
991 * to reduce RTT and bufferbloat.
992 * We do this using a special skb destructor (tcp_wfree).
993 *
994 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
995 * needs to be reallocated in a driver.
996 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
997 *
998 * Since transmit from skb destructor is forbidden, we use a tasklet
999 * to process all sockets that eventually need to send more skbs.
1000 * We use one tasklet per cpu, with its own queue of sockets.
1001 */
1002struct tsq_tasklet {
1003 struct tasklet_struct tasklet;
1004 struct list_head head; /* queue of tcp sockets */
1005};
1006static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
1007
1008static void tcp_tsq_write(struct sock *sk)
1009{
1010 if ((1 << sk->sk_state) &
1011 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
1012 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
1013 struct tcp_sock *tp = tcp_sk(sk);
1014
1015 if (tp->lost_out > tp->retrans_out &&
1016 tcp_snd_cwnd(tp) > tcp_packets_in_flight(tp)) {
1017 tcp_mstamp_refresh(tp);
1018 tcp_xmit_retransmit_queue(sk);
1019 }
1020
1021 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
1022 0, GFP_ATOMIC);
1023 }
1024}
1025
1026static void tcp_tsq_handler(struct sock *sk)
1027{
1028 bh_lock_sock(sk);
1029 if (!sock_owned_by_user(sk))
1030 tcp_tsq_write(sk);
1031 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
1032 sock_hold(sk);
1033 bh_unlock_sock(sk);
1034}
1035/*
1036 * One tasklet per cpu tries to send more skbs.
1037 * We run in tasklet context but need to disable irqs when
1038 * transferring tsq->head because tcp_wfree() might
1039 * interrupt us (non NAPI drivers)
1040 */
1041static void tcp_tasklet_func(struct tasklet_struct *t)
1042{
1043 struct tsq_tasklet *tsq = from_tasklet(tsq, t, tasklet);
1044 LIST_HEAD(list);
1045 unsigned long flags;
1046 struct list_head *q, *n;
1047 struct tcp_sock *tp;
1048 struct sock *sk;
1049
1050 local_irq_save(flags);
1051 list_splice_init(&tsq->head, &list);
1052 local_irq_restore(flags);
1053
1054 list_for_each_safe(q, n, &list) {
1055 tp = list_entry(q, struct tcp_sock, tsq_node);
1056 list_del(&tp->tsq_node);
1057
1058 sk = (struct sock *)tp;
1059 smp_mb__before_atomic();
1060 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
1061
1062 tcp_tsq_handler(sk);
1063 sk_free(sk);
1064 }
1065}
1066
1067#define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
1068 TCPF_WRITE_TIMER_DEFERRED | \
1069 TCPF_DELACK_TIMER_DEFERRED | \
1070 TCPF_MTU_REDUCED_DEFERRED)
1071/**
1072 * tcp_release_cb - tcp release_sock() callback
1073 * @sk: socket
1074 *
1075 * called from release_sock() to perform protocol dependent
1076 * actions before socket release.
1077 */
1078void tcp_release_cb(struct sock *sk)
1079{
1080 unsigned long flags = smp_load_acquire(&sk->sk_tsq_flags);
1081 unsigned long nflags;
1082
1083 /* perform an atomic operation only if at least one flag is set */
1084 do {
1085 if (!(flags & TCP_DEFERRED_ALL))
1086 return;
1087 nflags = flags & ~TCP_DEFERRED_ALL;
1088 } while (!try_cmpxchg(&sk->sk_tsq_flags, &flags, nflags));
1089
1090 if (flags & TCPF_TSQ_DEFERRED) {
1091 tcp_tsq_write(sk);
1092 __sock_put(sk);
1093 }
1094 /* Here begins the tricky part :
1095 * We are called from release_sock() with :
1096 * 1) BH disabled
1097 * 2) sk_lock.slock spinlock held
1098 * 3) socket owned by us (sk->sk_lock.owned == 1)
1099 *
1100 * But following code is meant to be called from BH handlers,
1101 * so we should keep BH disabled, but early release socket ownership
1102 */
1103 sock_release_ownership(sk);
1104
1105 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
1106 tcp_write_timer_handler(sk);
1107 __sock_put(sk);
1108 }
1109 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
1110 tcp_delack_timer_handler(sk);
1111 __sock_put(sk);
1112 }
1113 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
1114 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
1115 __sock_put(sk);
1116 }
1117}
1118EXPORT_SYMBOL(tcp_release_cb);
1119
1120void __init tcp_tasklet_init(void)
1121{
1122 int i;
1123
1124 for_each_possible_cpu(i) {
1125 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
1126
1127 INIT_LIST_HEAD(&tsq->head);
1128 tasklet_setup(&tsq->tasklet, tcp_tasklet_func);
1129 }
1130}
1131
1132/*
1133 * Write buffer destructor automatically called from kfree_skb.
1134 * We can't xmit new skbs from this context, as we might already
1135 * hold qdisc lock.
1136 */
1137void tcp_wfree(struct sk_buff *skb)
1138{
1139 struct sock *sk = skb->sk;
1140 struct tcp_sock *tp = tcp_sk(sk);
1141 unsigned long flags, nval, oval;
1142 struct tsq_tasklet *tsq;
1143 bool empty;
1144
1145 /* Keep one reference on sk_wmem_alloc.
1146 * Will be released by sk_free() from here or tcp_tasklet_func()
1147 */
1148 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
1149
1150 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
1151 * Wait until our queues (qdisc + devices) are drained.
1152 * This gives :
1153 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
1154 * - chance for incoming ACK (processed by another cpu maybe)
1155 * to migrate this flow (skb->ooo_okay will be eventually set)
1156 */
1157 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
1158 goto out;
1159
1160 oval = smp_load_acquire(&sk->sk_tsq_flags);
1161 do {
1162 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
1163 goto out;
1164
1165 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
1166 } while (!try_cmpxchg(&sk->sk_tsq_flags, &oval, nval));
1167
1168 /* queue this socket to tasklet queue */
1169 local_irq_save(flags);
1170 tsq = this_cpu_ptr(&tsq_tasklet);
1171 empty = list_empty(&tsq->head);
1172 list_add(&tp->tsq_node, &tsq->head);
1173 if (empty)
1174 tasklet_schedule(&tsq->tasklet);
1175 local_irq_restore(flags);
1176 return;
1177out:
1178 sk_free(sk);
1179}
1180
1181/* Note: Called under soft irq.
1182 * We can call TCP stack right away, unless socket is owned by user.
1183 */
1184enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
1185{
1186 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
1187 struct sock *sk = (struct sock *)tp;
1188
1189 tcp_tsq_handler(sk);
1190 sock_put(sk);
1191
1192 return HRTIMER_NORESTART;
1193}
1194
1195static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
1196 u64 prior_wstamp)
1197{
1198 struct tcp_sock *tp = tcp_sk(sk);
1199
1200 if (sk->sk_pacing_status != SK_PACING_NONE) {
1201 unsigned long rate = sk->sk_pacing_rate;
1202
1203 /* Original sch_fq does not pace first 10 MSS
1204 * Note that tp->data_segs_out overflows after 2^32 packets,
1205 * this is a minor annoyance.
1206 */
1207 if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
1208 u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
1209 u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
1210
1211 /* take into account OS jitter */
1212 len_ns -= min_t(u64, len_ns / 2, credit);
1213 tp->tcp_wstamp_ns += len_ns;
1214 }
1215 }
1216 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1217}
1218
1219INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1220INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1221INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb));
1222
1223/* This routine actually transmits TCP packets queued in by
1224 * tcp_do_sendmsg(). This is used by both the initial
1225 * transmission and possible later retransmissions.
1226 * All SKB's seen here are completely headerless. It is our
1227 * job to build the TCP header, and pass the packet down to
1228 * IP so it can do the same plus pass the packet off to the
1229 * device.
1230 *
1231 * We are working here with either a clone of the original
1232 * SKB, or a fresh unique copy made by the retransmit engine.
1233 */
1234static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1235 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1236{
1237 const struct inet_connection_sock *icsk = inet_csk(sk);
1238 struct inet_sock *inet;
1239 struct tcp_sock *tp;
1240 struct tcp_skb_cb *tcb;
1241 struct tcp_out_options opts;
1242 unsigned int tcp_options_size, tcp_header_size;
1243 struct sk_buff *oskb = NULL;
1244 struct tcp_md5sig_key *md5;
1245 struct tcphdr *th;
1246 u64 prior_wstamp;
1247 int err;
1248
1249 BUG_ON(!skb || !tcp_skb_pcount(skb));
1250 tp = tcp_sk(sk);
1251 prior_wstamp = tp->tcp_wstamp_ns;
1252 tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1253 skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true);
1254 if (clone_it) {
1255 oskb = skb;
1256
1257 tcp_skb_tsorted_save(oskb) {
1258 if (unlikely(skb_cloned(oskb)))
1259 skb = pskb_copy(oskb, gfp_mask);
1260 else
1261 skb = skb_clone(oskb, gfp_mask);
1262 } tcp_skb_tsorted_restore(oskb);
1263
1264 if (unlikely(!skb))
1265 return -ENOBUFS;
1266 /* retransmit skbs might have a non zero value in skb->dev
1267 * because skb->dev is aliased with skb->rbnode.rb_left
1268 */
1269 skb->dev = NULL;
1270 }
1271
1272 inet = inet_sk(sk);
1273 tcb = TCP_SKB_CB(skb);
1274 memset(&opts, 0, sizeof(opts));
1275
1276 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1277 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1278 } else {
1279 tcp_options_size = tcp_established_options(sk, skb, &opts,
1280 &md5);
1281 /* Force a PSH flag on all (GSO) packets to expedite GRO flush
1282 * at receiver : This slightly improve GRO performance.
1283 * Note that we do not force the PSH flag for non GSO packets,
1284 * because they might be sent under high congestion events,
1285 * and in this case it is better to delay the delivery of 1-MSS
1286 * packets and thus the corresponding ACK packet that would
1287 * release the following packet.
1288 */
1289 if (tcp_skb_pcount(skb) > 1)
1290 tcb->tcp_flags |= TCPHDR_PSH;
1291 }
1292 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1293
1294 /* if no packet is in qdisc/device queue, then allow XPS to select
1295 * another queue. We can be called from tcp_tsq_handler()
1296 * which holds one reference to sk.
1297 *
1298 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1299 * One way to get this would be to set skb->truesize = 2 on them.
1300 */
1301 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1302
1303 /* If we had to use memory reserve to allocate this skb,
1304 * this might cause drops if packet is looped back :
1305 * Other socket might not have SOCK_MEMALLOC.
1306 * Packets not looped back do not care about pfmemalloc.
1307 */
1308 skb->pfmemalloc = 0;
1309
1310 skb_push(skb, tcp_header_size);
1311 skb_reset_transport_header(skb);
1312
1313 skb_orphan(skb);
1314 skb->sk = sk;
1315 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1316 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1317
1318 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1319
1320 /* Build TCP header and checksum it. */
1321 th = (struct tcphdr *)skb->data;
1322 th->source = inet->inet_sport;
1323 th->dest = inet->inet_dport;
1324 th->seq = htonl(tcb->seq);
1325 th->ack_seq = htonl(rcv_nxt);
1326 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1327 tcb->tcp_flags);
1328
1329 th->check = 0;
1330 th->urg_ptr = 0;
1331
1332 /* The urg_mode check is necessary during a below snd_una win probe */
1333 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1334 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1335 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1336 th->urg = 1;
1337 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1338 th->urg_ptr = htons(0xFFFF);
1339 th->urg = 1;
1340 }
1341 }
1342
1343 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1344 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1345 th->window = htons(tcp_select_window(sk));
1346 tcp_ecn_send(sk, skb, th, tcp_header_size);
1347 } else {
1348 /* RFC1323: The window in SYN & SYN/ACK segments
1349 * is never scaled.
1350 */
1351 th->window = htons(min(tp->rcv_wnd, 65535U));
1352 }
1353
1354 tcp_options_write(th, tp, &opts);
1355
1356#ifdef CONFIG_TCP_MD5SIG
1357 /* Calculate the MD5 hash, as we have all we need now */
1358 if (md5) {
1359 sk_gso_disable(sk);
1360 tp->af_specific->calc_md5_hash(opts.hash_location,
1361 md5, sk, skb);
1362 }
1363#endif
1364
1365 /* BPF prog is the last one writing header option */
1366 bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts);
1367
1368 INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check,
1369 tcp_v6_send_check, tcp_v4_send_check,
1370 sk, skb);
1371
1372 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1373 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1374
1375 if (skb->len != tcp_header_size) {
1376 tcp_event_data_sent(tp, sk);
1377 tp->data_segs_out += tcp_skb_pcount(skb);
1378 tp->bytes_sent += skb->len - tcp_header_size;
1379 }
1380
1381 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1382 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1383 tcp_skb_pcount(skb));
1384
1385 tp->segs_out += tcp_skb_pcount(skb);
1386 skb_set_hash_from_sk(skb, sk);
1387 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1388 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1389 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1390
1391 /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1392
1393 /* Cleanup our debris for IP stacks */
1394 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1395 sizeof(struct inet6_skb_parm)));
1396
1397 tcp_add_tx_delay(skb, tp);
1398
1399 err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit,
1400 inet6_csk_xmit, ip_queue_xmit,
1401 sk, skb, &inet->cork.fl);
1402
1403 if (unlikely(err > 0)) {
1404 tcp_enter_cwr(sk);
1405 err = net_xmit_eval(err);
1406 }
1407 if (!err && oskb) {
1408 tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1409 tcp_rate_skb_sent(sk, oskb);
1410 }
1411 return err;
1412}
1413
1414static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1415 gfp_t gfp_mask)
1416{
1417 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1418 tcp_sk(sk)->rcv_nxt);
1419}
1420
1421/* This routine just queues the buffer for sending.
1422 *
1423 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1424 * otherwise socket can stall.
1425 */
1426static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1427{
1428 struct tcp_sock *tp = tcp_sk(sk);
1429
1430 /* Advance write_seq and place onto the write_queue. */
1431 WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1432 __skb_header_release(skb);
1433 tcp_add_write_queue_tail(sk, skb);
1434 sk_wmem_queued_add(sk, skb->truesize);
1435 sk_mem_charge(sk, skb->truesize);
1436}
1437
1438/* Initialize TSO segments for a packet. */
1439static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1440{
1441 if (skb->len <= mss_now) {
1442 /* Avoid the costly divide in the normal
1443 * non-TSO case.
1444 */
1445 tcp_skb_pcount_set(skb, 1);
1446 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1447 } else {
1448 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1449 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1450 }
1451}
1452
1453/* Pcount in the middle of the write queue got changed, we need to do various
1454 * tweaks to fix counters
1455 */
1456static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1457{
1458 struct tcp_sock *tp = tcp_sk(sk);
1459
1460 tp->packets_out -= decr;
1461
1462 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1463 tp->sacked_out -= decr;
1464 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1465 tp->retrans_out -= decr;
1466 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1467 tp->lost_out -= decr;
1468
1469 /* Reno case is special. Sigh... */
1470 if (tcp_is_reno(tp) && decr > 0)
1471 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1472
1473 if (tp->lost_skb_hint &&
1474 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1475 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1476 tp->lost_cnt_hint -= decr;
1477
1478 tcp_verify_left_out(tp);
1479}
1480
1481static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1482{
1483 return TCP_SKB_CB(skb)->txstamp_ack ||
1484 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1485}
1486
1487static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1488{
1489 struct skb_shared_info *shinfo = skb_shinfo(skb);
1490
1491 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1492 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1493 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1494 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1495
1496 shinfo->tx_flags &= ~tsflags;
1497 shinfo2->tx_flags |= tsflags;
1498 swap(shinfo->tskey, shinfo2->tskey);
1499 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1500 TCP_SKB_CB(skb)->txstamp_ack = 0;
1501 }
1502}
1503
1504static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1505{
1506 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1507 TCP_SKB_CB(skb)->eor = 0;
1508}
1509
1510/* Insert buff after skb on the write or rtx queue of sk. */
1511static void tcp_insert_write_queue_after(struct sk_buff *skb,
1512 struct sk_buff *buff,
1513 struct sock *sk,
1514 enum tcp_queue tcp_queue)
1515{
1516 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1517 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1518 else
1519 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1520}
1521
1522/* Function to create two new TCP segments. Shrinks the given segment
1523 * to the specified size and appends a new segment with the rest of the
1524 * packet to the list. This won't be called frequently, I hope.
1525 * Remember, these are still headerless SKBs at this point.
1526 */
1527int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1528 struct sk_buff *skb, u32 len,
1529 unsigned int mss_now, gfp_t gfp)
1530{
1531 struct tcp_sock *tp = tcp_sk(sk);
1532 struct sk_buff *buff;
1533 int nsize, old_factor;
1534 long limit;
1535 int nlen;
1536 u8 flags;
1537
1538 if (WARN_ON(len > skb->len))
1539 return -EINVAL;
1540
1541 nsize = skb_headlen(skb) - len;
1542 if (nsize < 0)
1543 nsize = 0;
1544
1545 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1546 * We need some allowance to not penalize applications setting small
1547 * SO_SNDBUF values.
1548 * Also allow first and last skb in retransmit queue to be split.
1549 */
1550 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_LEGACY_MAX_SIZE);
1551 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1552 tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1553 skb != tcp_rtx_queue_head(sk) &&
1554 skb != tcp_rtx_queue_tail(sk))) {
1555 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1556 return -ENOMEM;
1557 }
1558
1559 if (skb_unclone_keeptruesize(skb, gfp))
1560 return -ENOMEM;
1561
1562 /* Get a new skb... force flag on. */
1563 buff = tcp_stream_alloc_skb(sk, nsize, gfp, true);
1564 if (!buff)
1565 return -ENOMEM; /* We'll just try again later. */
1566 skb_copy_decrypted(buff, skb);
1567 mptcp_skb_ext_copy(buff, skb);
1568
1569 sk_wmem_queued_add(sk, buff->truesize);
1570 sk_mem_charge(sk, buff->truesize);
1571 nlen = skb->len - len - nsize;
1572 buff->truesize += nlen;
1573 skb->truesize -= nlen;
1574
1575 /* Correct the sequence numbers. */
1576 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1577 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1578 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1579
1580 /* PSH and FIN should only be set in the second packet. */
1581 flags = TCP_SKB_CB(skb)->tcp_flags;
1582 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1583 TCP_SKB_CB(buff)->tcp_flags = flags;
1584 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1585 tcp_skb_fragment_eor(skb, buff);
1586
1587 skb_split(skb, buff, len);
1588
1589 skb_set_delivery_time(buff, skb->tstamp, true);
1590 tcp_fragment_tstamp(skb, buff);
1591
1592 old_factor = tcp_skb_pcount(skb);
1593
1594 /* Fix up tso_factor for both original and new SKB. */
1595 tcp_set_skb_tso_segs(skb, mss_now);
1596 tcp_set_skb_tso_segs(buff, mss_now);
1597
1598 /* Update delivered info for the new segment */
1599 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1600
1601 /* If this packet has been sent out already, we must
1602 * adjust the various packet counters.
1603 */
1604 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1605 int diff = old_factor - tcp_skb_pcount(skb) -
1606 tcp_skb_pcount(buff);
1607
1608 if (diff)
1609 tcp_adjust_pcount(sk, skb, diff);
1610 }
1611
1612 /* Link BUFF into the send queue. */
1613 __skb_header_release(buff);
1614 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1615 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1616 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1617
1618 return 0;
1619}
1620
1621/* This is similar to __pskb_pull_tail(). The difference is that pulled
1622 * data is not copied, but immediately discarded.
1623 */
1624static int __pskb_trim_head(struct sk_buff *skb, int len)
1625{
1626 struct skb_shared_info *shinfo;
1627 int i, k, eat;
1628
1629 eat = min_t(int, len, skb_headlen(skb));
1630 if (eat) {
1631 __skb_pull(skb, eat);
1632 len -= eat;
1633 if (!len)
1634 return 0;
1635 }
1636 eat = len;
1637 k = 0;
1638 shinfo = skb_shinfo(skb);
1639 for (i = 0; i < shinfo->nr_frags; i++) {
1640 int size = skb_frag_size(&shinfo->frags[i]);
1641
1642 if (size <= eat) {
1643 skb_frag_unref(skb, i);
1644 eat -= size;
1645 } else {
1646 shinfo->frags[k] = shinfo->frags[i];
1647 if (eat) {
1648 skb_frag_off_add(&shinfo->frags[k], eat);
1649 skb_frag_size_sub(&shinfo->frags[k], eat);
1650 eat = 0;
1651 }
1652 k++;
1653 }
1654 }
1655 shinfo->nr_frags = k;
1656
1657 skb->data_len -= len;
1658 skb->len = skb->data_len;
1659 return len;
1660}
1661
1662/* Remove acked data from a packet in the transmit queue. */
1663int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1664{
1665 u32 delta_truesize;
1666
1667 if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
1668 return -ENOMEM;
1669
1670 delta_truesize = __pskb_trim_head(skb, len);
1671
1672 TCP_SKB_CB(skb)->seq += len;
1673
1674 if (delta_truesize) {
1675 skb->truesize -= delta_truesize;
1676 sk_wmem_queued_add(sk, -delta_truesize);
1677 if (!skb_zcopy_pure(skb))
1678 sk_mem_uncharge(sk, delta_truesize);
1679 }
1680
1681 /* Any change of skb->len requires recalculation of tso factor. */
1682 if (tcp_skb_pcount(skb) > 1)
1683 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1684
1685 return 0;
1686}
1687
1688/* Calculate MSS not accounting any TCP options. */
1689static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1690{
1691 const struct tcp_sock *tp = tcp_sk(sk);
1692 const struct inet_connection_sock *icsk = inet_csk(sk);
1693 int mss_now;
1694
1695 /* Calculate base mss without TCP options:
1696 It is MMS_S - sizeof(tcphdr) of rfc1122
1697 */
1698 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1699
1700 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1701 if (icsk->icsk_af_ops->net_frag_header_len) {
1702 const struct dst_entry *dst = __sk_dst_get(sk);
1703
1704 if (dst && dst_allfrag(dst))
1705 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1706 }
1707
1708 /* Clamp it (mss_clamp does not include tcp options) */
1709 if (mss_now > tp->rx_opt.mss_clamp)
1710 mss_now = tp->rx_opt.mss_clamp;
1711
1712 /* Now subtract optional transport overhead */
1713 mss_now -= icsk->icsk_ext_hdr_len;
1714
1715 /* Then reserve room for full set of TCP options and 8 bytes of data */
1716 mss_now = max(mss_now,
1717 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss));
1718 return mss_now;
1719}
1720
1721/* Calculate MSS. Not accounting for SACKs here. */
1722int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1723{
1724 /* Subtract TCP options size, not including SACKs */
1725 return __tcp_mtu_to_mss(sk, pmtu) -
1726 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1727}
1728EXPORT_SYMBOL(tcp_mtu_to_mss);
1729
1730/* Inverse of above */
1731int tcp_mss_to_mtu(struct sock *sk, int mss)
1732{
1733 const struct tcp_sock *tp = tcp_sk(sk);
1734 const struct inet_connection_sock *icsk = inet_csk(sk);
1735 int mtu;
1736
1737 mtu = mss +
1738 tp->tcp_header_len +
1739 icsk->icsk_ext_hdr_len +
1740 icsk->icsk_af_ops->net_header_len;
1741
1742 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1743 if (icsk->icsk_af_ops->net_frag_header_len) {
1744 const struct dst_entry *dst = __sk_dst_get(sk);
1745
1746 if (dst && dst_allfrag(dst))
1747 mtu += icsk->icsk_af_ops->net_frag_header_len;
1748 }
1749 return mtu;
1750}
1751EXPORT_SYMBOL(tcp_mss_to_mtu);
1752
1753/* MTU probing init per socket */
1754void tcp_mtup_init(struct sock *sk)
1755{
1756 struct tcp_sock *tp = tcp_sk(sk);
1757 struct inet_connection_sock *icsk = inet_csk(sk);
1758 struct net *net = sock_net(sk);
1759
1760 icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1;
1761 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1762 icsk->icsk_af_ops->net_header_len;
1763 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss));
1764 icsk->icsk_mtup.probe_size = 0;
1765 if (icsk->icsk_mtup.enabled)
1766 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1767}
1768EXPORT_SYMBOL(tcp_mtup_init);
1769
1770/* This function synchronize snd mss to current pmtu/exthdr set.
1771
1772 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1773 for TCP options, but includes only bare TCP header.
1774
1775 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1776 It is minimum of user_mss and mss received with SYN.
1777 It also does not include TCP options.
1778
1779 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1780
1781 tp->mss_cache is current effective sending mss, including
1782 all tcp options except for SACKs. It is evaluated,
1783 taking into account current pmtu, but never exceeds
1784 tp->rx_opt.mss_clamp.
1785
1786 NOTE1. rfc1122 clearly states that advertised MSS
1787 DOES NOT include either tcp or ip options.
1788
1789 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1790 are READ ONLY outside this function. --ANK (980731)
1791 */
1792unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1793{
1794 struct tcp_sock *tp = tcp_sk(sk);
1795 struct inet_connection_sock *icsk = inet_csk(sk);
1796 int mss_now;
1797
1798 if (icsk->icsk_mtup.search_high > pmtu)
1799 icsk->icsk_mtup.search_high = pmtu;
1800
1801 mss_now = tcp_mtu_to_mss(sk, pmtu);
1802 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1803
1804 /* And store cached results */
1805 icsk->icsk_pmtu_cookie = pmtu;
1806 if (icsk->icsk_mtup.enabled)
1807 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1808 tp->mss_cache = mss_now;
1809
1810 return mss_now;
1811}
1812EXPORT_SYMBOL(tcp_sync_mss);
1813
1814/* Compute the current effective MSS, taking SACKs and IP options,
1815 * and even PMTU discovery events into account.
1816 */
1817unsigned int tcp_current_mss(struct sock *sk)
1818{
1819 const struct tcp_sock *tp = tcp_sk(sk);
1820 const struct dst_entry *dst = __sk_dst_get(sk);
1821 u32 mss_now;
1822 unsigned int header_len;
1823 struct tcp_out_options opts;
1824 struct tcp_md5sig_key *md5;
1825
1826 mss_now = tp->mss_cache;
1827
1828 if (dst) {
1829 u32 mtu = dst_mtu(dst);
1830 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1831 mss_now = tcp_sync_mss(sk, mtu);
1832 }
1833
1834 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1835 sizeof(struct tcphdr);
1836 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1837 * some common options. If this is an odd packet (because we have SACK
1838 * blocks etc) then our calculated header_len will be different, and
1839 * we have to adjust mss_now correspondingly */
1840 if (header_len != tp->tcp_header_len) {
1841 int delta = (int) header_len - tp->tcp_header_len;
1842 mss_now -= delta;
1843 }
1844
1845 return mss_now;
1846}
1847
1848/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1849 * As additional protections, we do not touch cwnd in retransmission phases,
1850 * and if application hit its sndbuf limit recently.
1851 */
1852static void tcp_cwnd_application_limited(struct sock *sk)
1853{
1854 struct tcp_sock *tp = tcp_sk(sk);
1855
1856 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1857 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1858 /* Limited by application or receiver window. */
1859 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1860 u32 win_used = max(tp->snd_cwnd_used, init_win);
1861 if (win_used < tcp_snd_cwnd(tp)) {
1862 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1863 tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1);
1864 }
1865 tp->snd_cwnd_used = 0;
1866 }
1867 tp->snd_cwnd_stamp = tcp_jiffies32;
1868}
1869
1870static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1871{
1872 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1873 struct tcp_sock *tp = tcp_sk(sk);
1874
1875 /* Track the strongest available signal of the degree to which the cwnd
1876 * is fully utilized. If cwnd-limited then remember that fact for the
1877 * current window. If not cwnd-limited then track the maximum number of
1878 * outstanding packets in the current window. (If cwnd-limited then we
1879 * chose to not update tp->max_packets_out to avoid an extra else
1880 * clause with no functional impact.)
1881 */
1882 if (!before(tp->snd_una, tp->cwnd_usage_seq) ||
1883 is_cwnd_limited ||
1884 (!tp->is_cwnd_limited &&
1885 tp->packets_out > tp->max_packets_out)) {
1886 tp->is_cwnd_limited = is_cwnd_limited;
1887 tp->max_packets_out = tp->packets_out;
1888 tp->cwnd_usage_seq = tp->snd_nxt;
1889 }
1890
1891 if (tcp_is_cwnd_limited(sk)) {
1892 /* Network is feed fully. */
1893 tp->snd_cwnd_used = 0;
1894 tp->snd_cwnd_stamp = tcp_jiffies32;
1895 } else {
1896 /* Network starves. */
1897 if (tp->packets_out > tp->snd_cwnd_used)
1898 tp->snd_cwnd_used = tp->packets_out;
1899
1900 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) &&
1901 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1902 !ca_ops->cong_control)
1903 tcp_cwnd_application_limited(sk);
1904
1905 /* The following conditions together indicate the starvation
1906 * is caused by insufficient sender buffer:
1907 * 1) just sent some data (see tcp_write_xmit)
1908 * 2) not cwnd limited (this else condition)
1909 * 3) no more data to send (tcp_write_queue_empty())
1910 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1911 */
1912 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1913 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1914 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1915 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1916 }
1917}
1918
1919/* Minshall's variant of the Nagle send check. */
1920static bool tcp_minshall_check(const struct tcp_sock *tp)
1921{
1922 return after(tp->snd_sml, tp->snd_una) &&
1923 !after(tp->snd_sml, tp->snd_nxt);
1924}
1925
1926/* Update snd_sml if this skb is under mss
1927 * Note that a TSO packet might end with a sub-mss segment
1928 * The test is really :
1929 * if ((skb->len % mss) != 0)
1930 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1931 * But we can avoid doing the divide again given we already have
1932 * skb_pcount = skb->len / mss_now
1933 */
1934static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1935 const struct sk_buff *skb)
1936{
1937 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1938 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1939}
1940
1941/* Return false, if packet can be sent now without violation Nagle's rules:
1942 * 1. It is full sized. (provided by caller in %partial bool)
1943 * 2. Or it contains FIN. (already checked by caller)
1944 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1945 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1946 * With Minshall's modification: all sent small packets are ACKed.
1947 */
1948static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1949 int nonagle)
1950{
1951 return partial &&
1952 ((nonagle & TCP_NAGLE_CORK) ||
1953 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1954}
1955
1956/* Return how many segs we'd like on a TSO packet,
1957 * depending on current pacing rate, and how close the peer is.
1958 *
1959 * Rationale is:
1960 * - For close peers, we rather send bigger packets to reduce
1961 * cpu costs, because occasional losses will be repaired fast.
1962 * - For long distance/rtt flows, we would like to get ACK clocking
1963 * with 1 ACK per ms.
1964 *
1965 * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting
1966 * in bigger TSO bursts. We we cut the RTT-based allowance in half
1967 * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance
1968 * is below 1500 bytes after 6 * ~500 usec = 3ms.
1969 */
1970static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1971 int min_tso_segs)
1972{
1973 unsigned long bytes;
1974 u32 r;
1975
1976 bytes = sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift);
1977
1978 r = tcp_min_rtt(tcp_sk(sk)) >> READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_rtt_log);
1979 if (r < BITS_PER_TYPE(sk->sk_gso_max_size))
1980 bytes += sk->sk_gso_max_size >> r;
1981
1982 bytes = min_t(unsigned long, bytes, sk->sk_gso_max_size);
1983
1984 return max_t(u32, bytes / mss_now, min_tso_segs);
1985}
1986
1987/* Return the number of segments we want in the skb we are transmitting.
1988 * See if congestion control module wants to decide; otherwise, autosize.
1989 */
1990static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1991{
1992 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1993 u32 min_tso, tso_segs;
1994
1995 min_tso = ca_ops->min_tso_segs ?
1996 ca_ops->min_tso_segs(sk) :
1997 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs);
1998
1999 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
2000 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
2001}
2002
2003/* Returns the portion of skb which can be sent right away */
2004static unsigned int tcp_mss_split_point(const struct sock *sk,
2005 const struct sk_buff *skb,
2006 unsigned int mss_now,
2007 unsigned int max_segs,
2008 int nonagle)
2009{
2010 const struct tcp_sock *tp = tcp_sk(sk);
2011 u32 partial, needed, window, max_len;
2012
2013 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2014 max_len = mss_now * max_segs;
2015
2016 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
2017 return max_len;
2018
2019 needed = min(skb->len, window);
2020
2021 if (max_len <= needed)
2022 return max_len;
2023
2024 partial = needed % mss_now;
2025 /* If last segment is not a full MSS, check if Nagle rules allow us
2026 * to include this last segment in this skb.
2027 * Otherwise, we'll split the skb at last MSS boundary
2028 */
2029 if (tcp_nagle_check(partial != 0, tp, nonagle))
2030 return needed - partial;
2031
2032 return needed;
2033}
2034
2035/* Can at least one segment of SKB be sent right now, according to the
2036 * congestion window rules? If so, return how many segments are allowed.
2037 */
2038static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
2039 const struct sk_buff *skb)
2040{
2041 u32 in_flight, cwnd, halfcwnd;
2042
2043 /* Don't be strict about the congestion window for the final FIN. */
2044 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2045 tcp_skb_pcount(skb) == 1)
2046 return 1;
2047
2048 in_flight = tcp_packets_in_flight(tp);
2049 cwnd = tcp_snd_cwnd(tp);
2050 if (in_flight >= cwnd)
2051 return 0;
2052
2053 /* For better scheduling, ensure we have at least
2054 * 2 GSO packets in flight.
2055 */
2056 halfcwnd = max(cwnd >> 1, 1U);
2057 return min(halfcwnd, cwnd - in_flight);
2058}
2059
2060/* Initialize TSO state of a skb.
2061 * This must be invoked the first time we consider transmitting
2062 * SKB onto the wire.
2063 */
2064static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
2065{
2066 int tso_segs = tcp_skb_pcount(skb);
2067
2068 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
2069 tcp_set_skb_tso_segs(skb, mss_now);
2070 tso_segs = tcp_skb_pcount(skb);
2071 }
2072 return tso_segs;
2073}
2074
2075
2076/* Return true if the Nagle test allows this packet to be
2077 * sent now.
2078 */
2079static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
2080 unsigned int cur_mss, int nonagle)
2081{
2082 /* Nagle rule does not apply to frames, which sit in the middle of the
2083 * write_queue (they have no chances to get new data).
2084 *
2085 * This is implemented in the callers, where they modify the 'nonagle'
2086 * argument based upon the location of SKB in the send queue.
2087 */
2088 if (nonagle & TCP_NAGLE_PUSH)
2089 return true;
2090
2091 /* Don't use the nagle rule for urgent data (or for the final FIN). */
2092 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
2093 return true;
2094
2095 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
2096 return true;
2097
2098 return false;
2099}
2100
2101/* Does at least the first segment of SKB fit into the send window? */
2102static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
2103 const struct sk_buff *skb,
2104 unsigned int cur_mss)
2105{
2106 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2107
2108 if (skb->len > cur_mss)
2109 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
2110
2111 return !after(end_seq, tcp_wnd_end(tp));
2112}
2113
2114/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
2115 * which is put after SKB on the list. It is very much like
2116 * tcp_fragment() except that it may make several kinds of assumptions
2117 * in order to speed up the splitting operation. In particular, we
2118 * know that all the data is in scatter-gather pages, and that the
2119 * packet has never been sent out before (and thus is not cloned).
2120 */
2121static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
2122 unsigned int mss_now, gfp_t gfp)
2123{
2124 int nlen = skb->len - len;
2125 struct sk_buff *buff;
2126 u8 flags;
2127
2128 /* All of a TSO frame must be composed of paged data. */
2129 if (skb->len != skb->data_len)
2130 return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
2131 skb, len, mss_now, gfp);
2132
2133 buff = tcp_stream_alloc_skb(sk, 0, gfp, true);
2134 if (unlikely(!buff))
2135 return -ENOMEM;
2136 skb_copy_decrypted(buff, skb);
2137 mptcp_skb_ext_copy(buff, skb);
2138
2139 sk_wmem_queued_add(sk, buff->truesize);
2140 sk_mem_charge(sk, buff->truesize);
2141 buff->truesize += nlen;
2142 skb->truesize -= nlen;
2143
2144 /* Correct the sequence numbers. */
2145 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
2146 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
2147 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
2148
2149 /* PSH and FIN should only be set in the second packet. */
2150 flags = TCP_SKB_CB(skb)->tcp_flags;
2151 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
2152 TCP_SKB_CB(buff)->tcp_flags = flags;
2153
2154 tcp_skb_fragment_eor(skb, buff);
2155
2156 skb_split(skb, buff, len);
2157 tcp_fragment_tstamp(skb, buff);
2158
2159 /* Fix up tso_factor for both original and new SKB. */
2160 tcp_set_skb_tso_segs(skb, mss_now);
2161 tcp_set_skb_tso_segs(buff, mss_now);
2162
2163 /* Link BUFF into the send queue. */
2164 __skb_header_release(buff);
2165 tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
2166
2167 return 0;
2168}
2169
2170/* Try to defer sending, if possible, in order to minimize the amount
2171 * of TSO splitting we do. View it as a kind of TSO Nagle test.
2172 *
2173 * This algorithm is from John Heffner.
2174 */
2175static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
2176 bool *is_cwnd_limited,
2177 bool *is_rwnd_limited,
2178 u32 max_segs)
2179{
2180 const struct inet_connection_sock *icsk = inet_csk(sk);
2181 u32 send_win, cong_win, limit, in_flight;
2182 struct tcp_sock *tp = tcp_sk(sk);
2183 struct sk_buff *head;
2184 int win_divisor;
2185 s64 delta;
2186
2187 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
2188 goto send_now;
2189
2190 /* Avoid bursty behavior by allowing defer
2191 * only if the last write was recent (1 ms).
2192 * Note that tp->tcp_wstamp_ns can be in the future if we have
2193 * packets waiting in a qdisc or device for EDT delivery.
2194 */
2195 delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
2196 if (delta > 0)
2197 goto send_now;
2198
2199 in_flight = tcp_packets_in_flight(tp);
2200
2201 BUG_ON(tcp_skb_pcount(skb) <= 1);
2202 BUG_ON(tcp_snd_cwnd(tp) <= in_flight);
2203
2204 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2205
2206 /* From in_flight test above, we know that cwnd > in_flight. */
2207 cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache;
2208
2209 limit = min(send_win, cong_win);
2210
2211 /* If a full-sized TSO skb can be sent, do it. */
2212 if (limit >= max_segs * tp->mss_cache)
2213 goto send_now;
2214
2215 /* Middle in queue won't get any more data, full sendable already? */
2216 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
2217 goto send_now;
2218
2219 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
2220 if (win_divisor) {
2221 u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache);
2222
2223 /* If at least some fraction of a window is available,
2224 * just use it.
2225 */
2226 chunk /= win_divisor;
2227 if (limit >= chunk)
2228 goto send_now;
2229 } else {
2230 /* Different approach, try not to defer past a single
2231 * ACK. Receiver should ACK every other full sized
2232 * frame, so if we have space for more than 3 frames
2233 * then send now.
2234 */
2235 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
2236 goto send_now;
2237 }
2238
2239 /* TODO : use tsorted_sent_queue ? */
2240 head = tcp_rtx_queue_head(sk);
2241 if (!head)
2242 goto send_now;
2243 delta = tp->tcp_clock_cache - head->tstamp;
2244 /* If next ACK is likely to come too late (half srtt), do not defer */
2245 if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
2246 goto send_now;
2247
2248 /* Ok, it looks like it is advisable to defer.
2249 * Three cases are tracked :
2250 * 1) We are cwnd-limited
2251 * 2) We are rwnd-limited
2252 * 3) We are application limited.
2253 */
2254 if (cong_win < send_win) {
2255 if (cong_win <= skb->len) {
2256 *is_cwnd_limited = true;
2257 return true;
2258 }
2259 } else {
2260 if (send_win <= skb->len) {
2261 *is_rwnd_limited = true;
2262 return true;
2263 }
2264 }
2265
2266 /* If this packet won't get more data, do not wait. */
2267 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2268 TCP_SKB_CB(skb)->eor)
2269 goto send_now;
2270
2271 return true;
2272
2273send_now:
2274 return false;
2275}
2276
2277static inline void tcp_mtu_check_reprobe(struct sock *sk)
2278{
2279 struct inet_connection_sock *icsk = inet_csk(sk);
2280 struct tcp_sock *tp = tcp_sk(sk);
2281 struct net *net = sock_net(sk);
2282 u32 interval;
2283 s32 delta;
2284
2285 interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval);
2286 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2287 if (unlikely(delta >= interval * HZ)) {
2288 int mss = tcp_current_mss(sk);
2289
2290 /* Update current search range */
2291 icsk->icsk_mtup.probe_size = 0;
2292 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2293 sizeof(struct tcphdr) +
2294 icsk->icsk_af_ops->net_header_len;
2295 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2296
2297 /* Update probe time stamp */
2298 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2299 }
2300}
2301
2302static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2303{
2304 struct sk_buff *skb, *next;
2305
2306 skb = tcp_send_head(sk);
2307 tcp_for_write_queue_from_safe(skb, next, sk) {
2308 if (len <= skb->len)
2309 break;
2310
2311 if (unlikely(TCP_SKB_CB(skb)->eor) ||
2312 tcp_has_tx_tstamp(skb) ||
2313 !skb_pure_zcopy_same(skb, next))
2314 return false;
2315
2316 len -= skb->len;
2317 }
2318
2319 return true;
2320}
2321
2322/* Create a new MTU probe if we are ready.
2323 * MTU probe is regularly attempting to increase the path MTU by
2324 * deliberately sending larger packets. This discovers routing
2325 * changes resulting in larger path MTUs.
2326 *
2327 * Returns 0 if we should wait to probe (no cwnd available),
2328 * 1 if a probe was sent,
2329 * -1 otherwise
2330 */
2331static int tcp_mtu_probe(struct sock *sk)
2332{
2333 struct inet_connection_sock *icsk = inet_csk(sk);
2334 struct tcp_sock *tp = tcp_sk(sk);
2335 struct sk_buff *skb, *nskb, *next;
2336 struct net *net = sock_net(sk);
2337 int probe_size;
2338 int size_needed;
2339 int copy, len;
2340 int mss_now;
2341 int interval;
2342
2343 /* Not currently probing/verifying,
2344 * not in recovery,
2345 * have enough cwnd, and
2346 * not SACKing (the variable headers throw things off)
2347 */
2348 if (likely(!icsk->icsk_mtup.enabled ||
2349 icsk->icsk_mtup.probe_size ||
2350 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2351 tcp_snd_cwnd(tp) < 11 ||
2352 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2353 return -1;
2354
2355 /* Use binary search for probe_size between tcp_mss_base,
2356 * and current mss_clamp. if (search_high - search_low)
2357 * smaller than a threshold, backoff from probing.
2358 */
2359 mss_now = tcp_current_mss(sk);
2360 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2361 icsk->icsk_mtup.search_low) >> 1);
2362 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2363 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2364 /* When misfortune happens, we are reprobing actively,
2365 * and then reprobe timer has expired. We stick with current
2366 * probing process by not resetting search range to its orignal.
2367 */
2368 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2369 interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) {
2370 /* Check whether enough time has elaplased for
2371 * another round of probing.
2372 */
2373 tcp_mtu_check_reprobe(sk);
2374 return -1;
2375 }
2376
2377 /* Have enough data in the send queue to probe? */
2378 if (tp->write_seq - tp->snd_nxt < size_needed)
2379 return -1;
2380
2381 if (tp->snd_wnd < size_needed)
2382 return -1;
2383 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2384 return 0;
2385
2386 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2387 if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) {
2388 if (!tcp_packets_in_flight(tp))
2389 return -1;
2390 else
2391 return 0;
2392 }
2393
2394 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2395 return -1;
2396
2397 /* We're allowed to probe. Build it now. */
2398 nskb = tcp_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2399 if (!nskb)
2400 return -1;
2401 sk_wmem_queued_add(sk, nskb->truesize);
2402 sk_mem_charge(sk, nskb->truesize);
2403
2404 skb = tcp_send_head(sk);
2405 skb_copy_decrypted(nskb, skb);
2406 mptcp_skb_ext_copy(nskb, skb);
2407
2408 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2409 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2410 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2411
2412 tcp_insert_write_queue_before(nskb, skb, sk);
2413 tcp_highest_sack_replace(sk, skb, nskb);
2414
2415 len = 0;
2416 tcp_for_write_queue_from_safe(skb, next, sk) {
2417 copy = min_t(int, skb->len, probe_size - len);
2418 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2419
2420 if (skb->len <= copy) {
2421 /* We've eaten all the data from this skb.
2422 * Throw it away. */
2423 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2424 /* If this is the last SKB we copy and eor is set
2425 * we need to propagate it to the new skb.
2426 */
2427 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2428 tcp_skb_collapse_tstamp(nskb, skb);
2429 tcp_unlink_write_queue(skb, sk);
2430 tcp_wmem_free_skb(sk, skb);
2431 } else {
2432 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2433 ~(TCPHDR_FIN|TCPHDR_PSH);
2434 if (!skb_shinfo(skb)->nr_frags) {
2435 skb_pull(skb, copy);
2436 } else {
2437 __pskb_trim_head(skb, copy);
2438 tcp_set_skb_tso_segs(skb, mss_now);
2439 }
2440 TCP_SKB_CB(skb)->seq += copy;
2441 }
2442
2443 len += copy;
2444
2445 if (len >= probe_size)
2446 break;
2447 }
2448 tcp_init_tso_segs(nskb, nskb->len);
2449
2450 /* We're ready to send. If this fails, the probe will
2451 * be resegmented into mss-sized pieces by tcp_write_xmit().
2452 */
2453 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2454 /* Decrement cwnd here because we are sending
2455 * effectively two packets. */
2456 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1);
2457 tcp_event_new_data_sent(sk, nskb);
2458
2459 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2460 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2461 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2462
2463 return 1;
2464 }
2465
2466 return -1;
2467}
2468
2469static bool tcp_pacing_check(struct sock *sk)
2470{
2471 struct tcp_sock *tp = tcp_sk(sk);
2472
2473 if (!tcp_needs_internal_pacing(sk))
2474 return false;
2475
2476 if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2477 return false;
2478
2479 if (!hrtimer_is_queued(&tp->pacing_timer)) {
2480 hrtimer_start(&tp->pacing_timer,
2481 ns_to_ktime(tp->tcp_wstamp_ns),
2482 HRTIMER_MODE_ABS_PINNED_SOFT);
2483 sock_hold(sk);
2484 }
2485 return true;
2486}
2487
2488/* TCP Small Queues :
2489 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2490 * (These limits are doubled for retransmits)
2491 * This allows for :
2492 * - better RTT estimation and ACK scheduling
2493 * - faster recovery
2494 * - high rates
2495 * Alas, some drivers / subsystems require a fair amount
2496 * of queued bytes to ensure line rate.
2497 * One example is wifi aggregation (802.11 AMPDU)
2498 */
2499static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2500 unsigned int factor)
2501{
2502 unsigned long limit;
2503
2504 limit = max_t(unsigned long,
2505 2 * skb->truesize,
2506 sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift));
2507 if (sk->sk_pacing_status == SK_PACING_NONE)
2508 limit = min_t(unsigned long, limit,
2509 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes));
2510 limit <<= factor;
2511
2512 if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2513 tcp_sk(sk)->tcp_tx_delay) {
2514 u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay;
2515
2516 /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2517 * approximate our needs assuming an ~100% skb->truesize overhead.
2518 * USEC_PER_SEC is approximated by 2^20.
2519 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2520 */
2521 extra_bytes >>= (20 - 1);
2522 limit += extra_bytes;
2523 }
2524 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2525 /* Always send skb if rtx queue is empty.
2526 * No need to wait for TX completion to call us back,
2527 * after softirq/tasklet schedule.
2528 * This helps when TX completions are delayed too much.
2529 */
2530 if (tcp_rtx_queue_empty(sk))
2531 return false;
2532
2533 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2534 /* It is possible TX completion already happened
2535 * before we set TSQ_THROTTLED, so we must
2536 * test again the condition.
2537 */
2538 smp_mb__after_atomic();
2539 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2540 return true;
2541 }
2542 return false;
2543}
2544
2545static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2546{
2547 const u32 now = tcp_jiffies32;
2548 enum tcp_chrono old = tp->chrono_type;
2549
2550 if (old > TCP_CHRONO_UNSPEC)
2551 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2552 tp->chrono_start = now;
2553 tp->chrono_type = new;
2554}
2555
2556void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2557{
2558 struct tcp_sock *tp = tcp_sk(sk);
2559
2560 /* If there are multiple conditions worthy of tracking in a
2561 * chronograph then the highest priority enum takes precedence
2562 * over the other conditions. So that if something "more interesting"
2563 * starts happening, stop the previous chrono and start a new one.
2564 */
2565 if (type > tp->chrono_type)
2566 tcp_chrono_set(tp, type);
2567}
2568
2569void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2570{
2571 struct tcp_sock *tp = tcp_sk(sk);
2572
2573
2574 /* There are multiple conditions worthy of tracking in a
2575 * chronograph, so that the highest priority enum takes
2576 * precedence over the other conditions (see tcp_chrono_start).
2577 * If a condition stops, we only stop chrono tracking if
2578 * it's the "most interesting" or current chrono we are
2579 * tracking and starts busy chrono if we have pending data.
2580 */
2581 if (tcp_rtx_and_write_queues_empty(sk))
2582 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2583 else if (type == tp->chrono_type)
2584 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2585}
2586
2587/* This routine writes packets to the network. It advances the
2588 * send_head. This happens as incoming acks open up the remote
2589 * window for us.
2590 *
2591 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2592 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2593 * account rare use of URG, this is not a big flaw.
2594 *
2595 * Send at most one packet when push_one > 0. Temporarily ignore
2596 * cwnd limit to force at most one packet out when push_one == 2.
2597
2598 * Returns true, if no segments are in flight and we have queued segments,
2599 * but cannot send anything now because of SWS or another problem.
2600 */
2601static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2602 int push_one, gfp_t gfp)
2603{
2604 struct tcp_sock *tp = tcp_sk(sk);
2605 struct sk_buff *skb;
2606 unsigned int tso_segs, sent_pkts;
2607 int cwnd_quota;
2608 int result;
2609 bool is_cwnd_limited = false, is_rwnd_limited = false;
2610 u32 max_segs;
2611
2612 sent_pkts = 0;
2613
2614 tcp_mstamp_refresh(tp);
2615 if (!push_one) {
2616 /* Do MTU probing. */
2617 result = tcp_mtu_probe(sk);
2618 if (!result) {
2619 return false;
2620 } else if (result > 0) {
2621 sent_pkts = 1;
2622 }
2623 }
2624
2625 max_segs = tcp_tso_segs(sk, mss_now);
2626 while ((skb = tcp_send_head(sk))) {
2627 unsigned int limit;
2628
2629 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2630 /* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2631 tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2632 skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true);
2633 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2634 tcp_init_tso_segs(skb, mss_now);
2635 goto repair; /* Skip network transmission */
2636 }
2637
2638 if (tcp_pacing_check(sk))
2639 break;
2640
2641 tso_segs = tcp_init_tso_segs(skb, mss_now);
2642 BUG_ON(!tso_segs);
2643
2644 cwnd_quota = tcp_cwnd_test(tp, skb);
2645 if (!cwnd_quota) {
2646 if (push_one == 2)
2647 /* Force out a loss probe pkt. */
2648 cwnd_quota = 1;
2649 else
2650 break;
2651 }
2652
2653 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2654 is_rwnd_limited = true;
2655 break;
2656 }
2657
2658 if (tso_segs == 1) {
2659 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2660 (tcp_skb_is_last(sk, skb) ?
2661 nonagle : TCP_NAGLE_PUSH))))
2662 break;
2663 } else {
2664 if (!push_one &&
2665 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2666 &is_rwnd_limited, max_segs))
2667 break;
2668 }
2669
2670 limit = mss_now;
2671 if (tso_segs > 1 && !tcp_urg_mode(tp))
2672 limit = tcp_mss_split_point(sk, skb, mss_now,
2673 min_t(unsigned int,
2674 cwnd_quota,
2675 max_segs),
2676 nonagle);
2677
2678 if (skb->len > limit &&
2679 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2680 break;
2681
2682 if (tcp_small_queue_check(sk, skb, 0))
2683 break;
2684
2685 /* Argh, we hit an empty skb(), presumably a thread
2686 * is sleeping in sendmsg()/sk_stream_wait_memory().
2687 * We do not want to send a pure-ack packet and have
2688 * a strange looking rtx queue with empty packet(s).
2689 */
2690 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2691 break;
2692
2693 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2694 break;
2695
2696repair:
2697 /* Advance the send_head. This one is sent out.
2698 * This call will increment packets_out.
2699 */
2700 tcp_event_new_data_sent(sk, skb);
2701
2702 tcp_minshall_update(tp, mss_now, skb);
2703 sent_pkts += tcp_skb_pcount(skb);
2704
2705 if (push_one)
2706 break;
2707 }
2708
2709 if (is_rwnd_limited)
2710 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2711 else
2712 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2713
2714 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp));
2715 if (likely(sent_pkts || is_cwnd_limited))
2716 tcp_cwnd_validate(sk, is_cwnd_limited);
2717
2718 if (likely(sent_pkts)) {
2719 if (tcp_in_cwnd_reduction(sk))
2720 tp->prr_out += sent_pkts;
2721
2722 /* Send one loss probe per tail loss episode. */
2723 if (push_one != 2)
2724 tcp_schedule_loss_probe(sk, false);
2725 return false;
2726 }
2727 return !tp->packets_out && !tcp_write_queue_empty(sk);
2728}
2729
2730bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2731{
2732 struct inet_connection_sock *icsk = inet_csk(sk);
2733 struct tcp_sock *tp = tcp_sk(sk);
2734 u32 timeout, rto_delta_us;
2735 int early_retrans;
2736
2737 /* Don't do any loss probe on a Fast Open connection before 3WHS
2738 * finishes.
2739 */
2740 if (rcu_access_pointer(tp->fastopen_rsk))
2741 return false;
2742
2743 early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans);
2744 /* Schedule a loss probe in 2*RTT for SACK capable connections
2745 * not in loss recovery, that are either limited by cwnd or application.
2746 */
2747 if ((early_retrans != 3 && early_retrans != 4) ||
2748 !tp->packets_out || !tcp_is_sack(tp) ||
2749 (icsk->icsk_ca_state != TCP_CA_Open &&
2750 icsk->icsk_ca_state != TCP_CA_CWR))
2751 return false;
2752
2753 /* Probe timeout is 2*rtt. Add minimum RTO to account
2754 * for delayed ack when there's one outstanding packet. If no RTT
2755 * sample is available then probe after TCP_TIMEOUT_INIT.
2756 */
2757 if (tp->srtt_us) {
2758 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2759 if (tp->packets_out == 1)
2760 timeout += TCP_RTO_MIN;
2761 else
2762 timeout += TCP_TIMEOUT_MIN;
2763 } else {
2764 timeout = TCP_TIMEOUT_INIT;
2765 }
2766
2767 /* If the RTO formula yields an earlier time, then use that time. */
2768 rto_delta_us = advancing_rto ?
2769 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2770 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2771 if (rto_delta_us > 0)
2772 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2773
2774 tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX);
2775 return true;
2776}
2777
2778/* Thanks to skb fast clones, we can detect if a prior transmit of
2779 * a packet is still in a qdisc or driver queue.
2780 * In this case, there is very little point doing a retransmit !
2781 */
2782static bool skb_still_in_host_queue(struct sock *sk,
2783 const struct sk_buff *skb)
2784{
2785 if (unlikely(skb_fclone_busy(sk, skb))) {
2786 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2787 smp_mb__after_atomic();
2788 if (skb_fclone_busy(sk, skb)) {
2789 NET_INC_STATS(sock_net(sk),
2790 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2791 return true;
2792 }
2793 }
2794 return false;
2795}
2796
2797/* When probe timeout (PTO) fires, try send a new segment if possible, else
2798 * retransmit the last segment.
2799 */
2800void tcp_send_loss_probe(struct sock *sk)
2801{
2802 struct tcp_sock *tp = tcp_sk(sk);
2803 struct sk_buff *skb;
2804 int pcount;
2805 int mss = tcp_current_mss(sk);
2806
2807 /* At most one outstanding TLP */
2808 if (tp->tlp_high_seq)
2809 goto rearm_timer;
2810
2811 tp->tlp_retrans = 0;
2812 skb = tcp_send_head(sk);
2813 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2814 pcount = tp->packets_out;
2815 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2816 if (tp->packets_out > pcount)
2817 goto probe_sent;
2818 goto rearm_timer;
2819 }
2820 skb = skb_rb_last(&sk->tcp_rtx_queue);
2821 if (unlikely(!skb)) {
2822 WARN_ONCE(tp->packets_out,
2823 "invalid inflight: %u state %u cwnd %u mss %d\n",
2824 tp->packets_out, sk->sk_state, tcp_snd_cwnd(tp), mss);
2825 inet_csk(sk)->icsk_pending = 0;
2826 return;
2827 }
2828
2829 if (skb_still_in_host_queue(sk, skb))
2830 goto rearm_timer;
2831
2832 pcount = tcp_skb_pcount(skb);
2833 if (WARN_ON(!pcount))
2834 goto rearm_timer;
2835
2836 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2837 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2838 (pcount - 1) * mss, mss,
2839 GFP_ATOMIC)))
2840 goto rearm_timer;
2841 skb = skb_rb_next(skb);
2842 }
2843
2844 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2845 goto rearm_timer;
2846
2847 if (__tcp_retransmit_skb(sk, skb, 1))
2848 goto rearm_timer;
2849
2850 tp->tlp_retrans = 1;
2851
2852probe_sent:
2853 /* Record snd_nxt for loss detection. */
2854 tp->tlp_high_seq = tp->snd_nxt;
2855
2856 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2857 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2858 inet_csk(sk)->icsk_pending = 0;
2859rearm_timer:
2860 tcp_rearm_rto(sk);
2861}
2862
2863/* Push out any pending frames which were held back due to
2864 * TCP_CORK or attempt at coalescing tiny packets.
2865 * The socket must be locked by the caller.
2866 */
2867void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2868 int nonagle)
2869{
2870 /* If we are closed, the bytes will have to remain here.
2871 * In time closedown will finish, we empty the write queue and
2872 * all will be happy.
2873 */
2874 if (unlikely(sk->sk_state == TCP_CLOSE))
2875 return;
2876
2877 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2878 sk_gfp_mask(sk, GFP_ATOMIC)))
2879 tcp_check_probe_timer(sk);
2880}
2881
2882/* Send _single_ skb sitting at the send head. This function requires
2883 * true push pending frames to setup probe timer etc.
2884 */
2885void tcp_push_one(struct sock *sk, unsigned int mss_now)
2886{
2887 struct sk_buff *skb = tcp_send_head(sk);
2888
2889 BUG_ON(!skb || skb->len < mss_now);
2890
2891 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2892}
2893
2894/* This function returns the amount that we can raise the
2895 * usable window based on the following constraints
2896 *
2897 * 1. The window can never be shrunk once it is offered (RFC 793)
2898 * 2. We limit memory per socket
2899 *
2900 * RFC 1122:
2901 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2902 * RECV.NEXT + RCV.WIN fixed until:
2903 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2904 *
2905 * i.e. don't raise the right edge of the window until you can raise
2906 * it at least MSS bytes.
2907 *
2908 * Unfortunately, the recommended algorithm breaks header prediction,
2909 * since header prediction assumes th->window stays fixed.
2910 *
2911 * Strictly speaking, keeping th->window fixed violates the receiver
2912 * side SWS prevention criteria. The problem is that under this rule
2913 * a stream of single byte packets will cause the right side of the
2914 * window to always advance by a single byte.
2915 *
2916 * Of course, if the sender implements sender side SWS prevention
2917 * then this will not be a problem.
2918 *
2919 * BSD seems to make the following compromise:
2920 *
2921 * If the free space is less than the 1/4 of the maximum
2922 * space available and the free space is less than 1/2 mss,
2923 * then set the window to 0.
2924 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2925 * Otherwise, just prevent the window from shrinking
2926 * and from being larger than the largest representable value.
2927 *
2928 * This prevents incremental opening of the window in the regime
2929 * where TCP is limited by the speed of the reader side taking
2930 * data out of the TCP receive queue. It does nothing about
2931 * those cases where the window is constrained on the sender side
2932 * because the pipeline is full.
2933 *
2934 * BSD also seems to "accidentally" limit itself to windows that are a
2935 * multiple of MSS, at least until the free space gets quite small.
2936 * This would appear to be a side effect of the mbuf implementation.
2937 * Combining these two algorithms results in the observed behavior
2938 * of having a fixed window size at almost all times.
2939 *
2940 * Below we obtain similar behavior by forcing the offered window to
2941 * a multiple of the mss when it is feasible to do so.
2942 *
2943 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2944 * Regular options like TIMESTAMP are taken into account.
2945 */
2946u32 __tcp_select_window(struct sock *sk)
2947{
2948 struct inet_connection_sock *icsk = inet_csk(sk);
2949 struct tcp_sock *tp = tcp_sk(sk);
2950 /* MSS for the peer's data. Previous versions used mss_clamp
2951 * here. I don't know if the value based on our guesses
2952 * of peer's MSS is better for the performance. It's more correct
2953 * but may be worse for the performance because of rcv_mss
2954 * fluctuations. --SAW 1998/11/1
2955 */
2956 int mss = icsk->icsk_ack.rcv_mss;
2957 int free_space = tcp_space(sk);
2958 int allowed_space = tcp_full_space(sk);
2959 int full_space, window;
2960
2961 if (sk_is_mptcp(sk))
2962 mptcp_space(sk, &free_space, &allowed_space);
2963
2964 full_space = min_t(int, tp->window_clamp, allowed_space);
2965
2966 if (unlikely(mss > full_space)) {
2967 mss = full_space;
2968 if (mss <= 0)
2969 return 0;
2970 }
2971 if (free_space < (full_space >> 1)) {
2972 icsk->icsk_ack.quick = 0;
2973
2974 if (tcp_under_memory_pressure(sk))
2975 tcp_adjust_rcv_ssthresh(sk);
2976
2977 /* free_space might become our new window, make sure we don't
2978 * increase it due to wscale.
2979 */
2980 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2981
2982 /* if free space is less than mss estimate, or is below 1/16th
2983 * of the maximum allowed, try to move to zero-window, else
2984 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2985 * new incoming data is dropped due to memory limits.
2986 * With large window, mss test triggers way too late in order
2987 * to announce zero window in time before rmem limit kicks in.
2988 */
2989 if (free_space < (allowed_space >> 4) || free_space < mss)
2990 return 0;
2991 }
2992
2993 if (free_space > tp->rcv_ssthresh)
2994 free_space = tp->rcv_ssthresh;
2995
2996 /* Don't do rounding if we are using window scaling, since the
2997 * scaled window will not line up with the MSS boundary anyway.
2998 */
2999 if (tp->rx_opt.rcv_wscale) {
3000 window = free_space;
3001
3002 /* Advertise enough space so that it won't get scaled away.
3003 * Import case: prevent zero window announcement if
3004 * 1<<rcv_wscale > mss.
3005 */
3006 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
3007 } else {
3008 window = tp->rcv_wnd;
3009 /* Get the largest window that is a nice multiple of mss.
3010 * Window clamp already applied above.
3011 * If our current window offering is within 1 mss of the
3012 * free space we just keep it. This prevents the divide
3013 * and multiply from happening most of the time.
3014 * We also don't do any window rounding when the free space
3015 * is too small.
3016 */
3017 if (window <= free_space - mss || window > free_space)
3018 window = rounddown(free_space, mss);
3019 else if (mss == full_space &&
3020 free_space > window + (full_space >> 1))
3021 window = free_space;
3022 }
3023
3024 return window;
3025}
3026
3027void tcp_skb_collapse_tstamp(struct sk_buff *skb,
3028 const struct sk_buff *next_skb)
3029{
3030 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
3031 const struct skb_shared_info *next_shinfo =
3032 skb_shinfo(next_skb);
3033 struct skb_shared_info *shinfo = skb_shinfo(skb);
3034
3035 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
3036 shinfo->tskey = next_shinfo->tskey;
3037 TCP_SKB_CB(skb)->txstamp_ack |=
3038 TCP_SKB_CB(next_skb)->txstamp_ack;
3039 }
3040}
3041
3042/* Collapses two adjacent SKB's during retransmission. */
3043static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
3044{
3045 struct tcp_sock *tp = tcp_sk(sk);
3046 struct sk_buff *next_skb = skb_rb_next(skb);
3047 int next_skb_size;
3048
3049 next_skb_size = next_skb->len;
3050
3051 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
3052
3053 if (next_skb_size && !tcp_skb_shift(skb, next_skb, 1, next_skb_size))
3054 return false;
3055
3056 tcp_highest_sack_replace(sk, next_skb, skb);
3057
3058 /* Update sequence range on original skb. */
3059 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
3060
3061 /* Merge over control information. This moves PSH/FIN etc. over */
3062 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
3063
3064 /* All done, get rid of second SKB and account for it so
3065 * packet counting does not break.
3066 */
3067 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
3068 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
3069
3070 /* changed transmit queue under us so clear hints */
3071 tcp_clear_retrans_hints_partial(tp);
3072 if (next_skb == tp->retransmit_skb_hint)
3073 tp->retransmit_skb_hint = skb;
3074
3075 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
3076
3077 tcp_skb_collapse_tstamp(skb, next_skb);
3078
3079 tcp_rtx_queue_unlink_and_free(next_skb, sk);
3080 return true;
3081}
3082
3083/* Check if coalescing SKBs is legal. */
3084static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
3085{
3086 if (tcp_skb_pcount(skb) > 1)
3087 return false;
3088 if (skb_cloned(skb))
3089 return false;
3090 /* Some heuristics for collapsing over SACK'd could be invented */
3091 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3092 return false;
3093
3094 return true;
3095}
3096
3097/* Collapse packets in the retransmit queue to make to create
3098 * less packets on the wire. This is only done on retransmission.
3099 */
3100static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
3101 int space)
3102{
3103 struct tcp_sock *tp = tcp_sk(sk);
3104 struct sk_buff *skb = to, *tmp;
3105 bool first = true;
3106
3107 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse))
3108 return;
3109 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3110 return;
3111
3112 skb_rbtree_walk_from_safe(skb, tmp) {
3113 if (!tcp_can_collapse(sk, skb))
3114 break;
3115
3116 if (!tcp_skb_can_collapse(to, skb))
3117 break;
3118
3119 space -= skb->len;
3120
3121 if (first) {
3122 first = false;
3123 continue;
3124 }
3125
3126 if (space < 0)
3127 break;
3128
3129 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
3130 break;
3131
3132 if (!tcp_collapse_retrans(sk, to))
3133 break;
3134 }
3135}
3136
3137/* This retransmits one SKB. Policy decisions and retransmit queue
3138 * state updates are done by the caller. Returns non-zero if an
3139 * error occurred which prevented the send.
3140 */
3141int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3142{
3143 struct inet_connection_sock *icsk = inet_csk(sk);
3144 struct tcp_sock *tp = tcp_sk(sk);
3145 unsigned int cur_mss;
3146 int diff, len, err;
3147 int avail_wnd;
3148
3149 /* Inconclusive MTU probe */
3150 if (icsk->icsk_mtup.probe_size)
3151 icsk->icsk_mtup.probe_size = 0;
3152
3153 if (skb_still_in_host_queue(sk, skb))
3154 return -EBUSY;
3155
3156 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
3157 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
3158 WARN_ON_ONCE(1);
3159 return -EINVAL;
3160 }
3161 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3162 return -ENOMEM;
3163 }
3164
3165 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3166 return -EHOSTUNREACH; /* Routing failure or similar. */
3167
3168 cur_mss = tcp_current_mss(sk);
3169 avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3170
3171 /* If receiver has shrunk his window, and skb is out of
3172 * new window, do not retransmit it. The exception is the
3173 * case, when window is shrunk to zero. In this case
3174 * our retransmit of one segment serves as a zero window probe.
3175 */
3176 if (avail_wnd <= 0) {
3177 if (TCP_SKB_CB(skb)->seq != tp->snd_una)
3178 return -EAGAIN;
3179 avail_wnd = cur_mss;
3180 }
3181
3182 len = cur_mss * segs;
3183 if (len > avail_wnd) {
3184 len = rounddown(avail_wnd, cur_mss);
3185 if (!len)
3186 len = avail_wnd;
3187 }
3188 if (skb->len > len) {
3189 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
3190 cur_mss, GFP_ATOMIC))
3191 return -ENOMEM; /* We'll try again later. */
3192 } else {
3193 if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
3194 return -ENOMEM;
3195
3196 diff = tcp_skb_pcount(skb);
3197 tcp_set_skb_tso_segs(skb, cur_mss);
3198 diff -= tcp_skb_pcount(skb);
3199 if (diff)
3200 tcp_adjust_pcount(sk, skb, diff);
3201 avail_wnd = min_t(int, avail_wnd, cur_mss);
3202 if (skb->len < avail_wnd)
3203 tcp_retrans_try_collapse(sk, skb, avail_wnd);
3204 }
3205
3206 /* RFC3168, section 6.1.1.1. ECN fallback */
3207 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
3208 tcp_ecn_clear_syn(sk, skb);
3209
3210 /* Update global and local TCP statistics. */
3211 segs = tcp_skb_pcount(skb);
3212 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
3213 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3214 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3215 tp->total_retrans += segs;
3216 tp->bytes_retrans += skb->len;
3217
3218 /* make sure skb->data is aligned on arches that require it
3219 * and check if ack-trimming & collapsing extended the headroom
3220 * beyond what csum_start can cover.
3221 */
3222 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
3223 skb_headroom(skb) >= 0xFFFF)) {
3224 struct sk_buff *nskb;
3225
3226 tcp_skb_tsorted_save(skb) {
3227 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
3228 if (nskb) {
3229 nskb->dev = NULL;
3230 err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC);
3231 } else {
3232 err = -ENOBUFS;
3233 }
3234 } tcp_skb_tsorted_restore(skb);
3235
3236 if (!err) {
3237 tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
3238 tcp_rate_skb_sent(sk, skb);
3239 }
3240 } else {
3241 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3242 }
3243
3244 /* To avoid taking spuriously low RTT samples based on a timestamp
3245 * for a transmit that never happened, always mark EVER_RETRANS
3246 */
3247 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
3248
3249 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
3250 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
3251 TCP_SKB_CB(skb)->seq, segs, err);
3252
3253 if (likely(!err)) {
3254 trace_tcp_retransmit_skb(sk, skb);
3255 } else if (err != -EBUSY) {
3256 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
3257 }
3258 return err;
3259}
3260
3261int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3262{
3263 struct tcp_sock *tp = tcp_sk(sk);
3264 int err = __tcp_retransmit_skb(sk, skb, segs);
3265
3266 if (err == 0) {
3267#if FASTRETRANS_DEBUG > 0
3268 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3269 net_dbg_ratelimited("retrans_out leaked\n");
3270 }
3271#endif
3272 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3273 tp->retrans_out += tcp_skb_pcount(skb);
3274 }
3275
3276 /* Save stamp of the first (attempted) retransmit. */
3277 if (!tp->retrans_stamp)
3278 tp->retrans_stamp = tcp_skb_timestamp(skb);
3279
3280 if (tp->undo_retrans < 0)
3281 tp->undo_retrans = 0;
3282 tp->undo_retrans += tcp_skb_pcount(skb);
3283 return err;
3284}
3285
3286/* This gets called after a retransmit timeout, and the initially
3287 * retransmitted data is acknowledged. It tries to continue
3288 * resending the rest of the retransmit queue, until either
3289 * we've sent it all or the congestion window limit is reached.
3290 */
3291void tcp_xmit_retransmit_queue(struct sock *sk)
3292{
3293 const struct inet_connection_sock *icsk = inet_csk(sk);
3294 struct sk_buff *skb, *rtx_head, *hole = NULL;
3295 struct tcp_sock *tp = tcp_sk(sk);
3296 bool rearm_timer = false;
3297 u32 max_segs;
3298 int mib_idx;
3299
3300 if (!tp->packets_out)
3301 return;
3302
3303 rtx_head = tcp_rtx_queue_head(sk);
3304 skb = tp->retransmit_skb_hint ?: rtx_head;
3305 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3306 skb_rbtree_walk_from(skb) {
3307 __u8 sacked;
3308 int segs;
3309
3310 if (tcp_pacing_check(sk))
3311 break;
3312
3313 /* we could do better than to assign each time */
3314 if (!hole)
3315 tp->retransmit_skb_hint = skb;
3316
3317 segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp);
3318 if (segs <= 0)
3319 break;
3320 sacked = TCP_SKB_CB(skb)->sacked;
3321 /* In case tcp_shift_skb_data() have aggregated large skbs,
3322 * we need to make sure not sending too bigs TSO packets
3323 */
3324 segs = min_t(int, segs, max_segs);
3325
3326 if (tp->retrans_out >= tp->lost_out) {
3327 break;
3328 } else if (!(sacked & TCPCB_LOST)) {
3329 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3330 hole = skb;
3331 continue;
3332
3333 } else {
3334 if (icsk->icsk_ca_state != TCP_CA_Loss)
3335 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3336 else
3337 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3338 }
3339
3340 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3341 continue;
3342
3343 if (tcp_small_queue_check(sk, skb, 1))
3344 break;
3345
3346 if (tcp_retransmit_skb(sk, skb, segs))
3347 break;
3348
3349 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3350
3351 if (tcp_in_cwnd_reduction(sk))
3352 tp->prr_out += tcp_skb_pcount(skb);
3353
3354 if (skb == rtx_head &&
3355 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3356 rearm_timer = true;
3357
3358 }
3359 if (rearm_timer)
3360 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3361 inet_csk(sk)->icsk_rto,
3362 TCP_RTO_MAX);
3363}
3364
3365/* We allow to exceed memory limits for FIN packets to expedite
3366 * connection tear down and (memory) recovery.
3367 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3368 * or even be forced to close flow without any FIN.
3369 * In general, we want to allow one skb per socket to avoid hangs
3370 * with edge trigger epoll()
3371 */
3372void sk_forced_mem_schedule(struct sock *sk, int size)
3373{
3374 int delta, amt;
3375
3376 delta = size - sk->sk_forward_alloc;
3377 if (delta <= 0)
3378 return;
3379 amt = sk_mem_pages(delta);
3380 sk->sk_forward_alloc += amt << PAGE_SHIFT;
3381 sk_memory_allocated_add(sk, amt);
3382
3383 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3384 mem_cgroup_charge_skmem(sk->sk_memcg, amt,
3385 gfp_memcg_charge() | __GFP_NOFAIL);
3386}
3387
3388/* Send a FIN. The caller locks the socket for us.
3389 * We should try to send a FIN packet really hard, but eventually give up.
3390 */
3391void tcp_send_fin(struct sock *sk)
3392{
3393 struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk);
3394 struct tcp_sock *tp = tcp_sk(sk);
3395
3396 /* Optimization, tack on the FIN if we have one skb in write queue and
3397 * this skb was not yet sent, or we are under memory pressure.
3398 * Note: in the latter case, FIN packet will be sent after a timeout,
3399 * as TCP stack thinks it has already been transmitted.
3400 */
3401 tskb = tail;
3402 if (!tskb && tcp_under_memory_pressure(sk))
3403 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3404
3405 if (tskb) {
3406 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3407 TCP_SKB_CB(tskb)->end_seq++;
3408 tp->write_seq++;
3409 if (!tail) {
3410 /* This means tskb was already sent.
3411 * Pretend we included the FIN on previous transmit.
3412 * We need to set tp->snd_nxt to the value it would have
3413 * if FIN had been sent. This is because retransmit path
3414 * does not change tp->snd_nxt.
3415 */
3416 WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3417 return;
3418 }
3419 } else {
3420 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3421 if (unlikely(!skb))
3422 return;
3423
3424 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3425 skb_reserve(skb, MAX_TCP_HEADER);
3426 sk_forced_mem_schedule(sk, skb->truesize);
3427 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3428 tcp_init_nondata_skb(skb, tp->write_seq,
3429 TCPHDR_ACK | TCPHDR_FIN);
3430 tcp_queue_skb(sk, skb);
3431 }
3432 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3433}
3434
3435/* We get here when a process closes a file descriptor (either due to
3436 * an explicit close() or as a byproduct of exit()'ing) and there
3437 * was unread data in the receive queue. This behavior is recommended
3438 * by RFC 2525, section 2.17. -DaveM
3439 */
3440void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3441{
3442 struct sk_buff *skb;
3443
3444 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3445
3446 /* NOTE: No TCP options attached and we never retransmit this. */
3447 skb = alloc_skb(MAX_TCP_HEADER, priority);
3448 if (!skb) {
3449 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3450 return;
3451 }
3452
3453 /* Reserve space for headers and prepare control bits. */
3454 skb_reserve(skb, MAX_TCP_HEADER);
3455 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3456 TCPHDR_ACK | TCPHDR_RST);
3457 tcp_mstamp_refresh(tcp_sk(sk));
3458 /* Send it off. */
3459 if (tcp_transmit_skb(sk, skb, 0, priority))
3460 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3461
3462 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3463 * skb here is different to the troublesome skb, so use NULL
3464 */
3465 trace_tcp_send_reset(sk, NULL);
3466}
3467
3468/* Send a crossed SYN-ACK during socket establishment.
3469 * WARNING: This routine must only be called when we have already sent
3470 * a SYN packet that crossed the incoming SYN that caused this routine
3471 * to get called. If this assumption fails then the initial rcv_wnd
3472 * and rcv_wscale values will not be correct.
3473 */
3474int tcp_send_synack(struct sock *sk)
3475{
3476 struct sk_buff *skb;
3477
3478 skb = tcp_rtx_queue_head(sk);
3479 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3480 pr_err("%s: wrong queue state\n", __func__);
3481 return -EFAULT;
3482 }
3483 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3484 if (skb_cloned(skb)) {
3485 struct sk_buff *nskb;
3486
3487 tcp_skb_tsorted_save(skb) {
3488 nskb = skb_copy(skb, GFP_ATOMIC);
3489 } tcp_skb_tsorted_restore(skb);
3490 if (!nskb)
3491 return -ENOMEM;
3492 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3493 tcp_highest_sack_replace(sk, skb, nskb);
3494 tcp_rtx_queue_unlink_and_free(skb, sk);
3495 __skb_header_release(nskb);
3496 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3497 sk_wmem_queued_add(sk, nskb->truesize);
3498 sk_mem_charge(sk, nskb->truesize);
3499 skb = nskb;
3500 }
3501
3502 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3503 tcp_ecn_send_synack(sk, skb);
3504 }
3505 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3506}
3507
3508/**
3509 * tcp_make_synack - Allocate one skb and build a SYNACK packet.
3510 * @sk: listener socket
3511 * @dst: dst entry attached to the SYNACK. It is consumed and caller
3512 * should not use it again.
3513 * @req: request_sock pointer
3514 * @foc: cookie for tcp fast open
3515 * @synack_type: Type of synack to prepare
3516 * @syn_skb: SYN packet just received. It could be NULL for rtx case.
3517 */
3518struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3519 struct request_sock *req,
3520 struct tcp_fastopen_cookie *foc,
3521 enum tcp_synack_type synack_type,
3522 struct sk_buff *syn_skb)
3523{
3524 struct inet_request_sock *ireq = inet_rsk(req);
3525 const struct tcp_sock *tp = tcp_sk(sk);
3526 struct tcp_md5sig_key *md5 = NULL;
3527 struct tcp_out_options opts;
3528 struct sk_buff *skb;
3529 int tcp_header_size;
3530 struct tcphdr *th;
3531 int mss;
3532 u64 now;
3533
3534 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3535 if (unlikely(!skb)) {
3536 dst_release(dst);
3537 return NULL;
3538 }
3539 /* Reserve space for headers. */
3540 skb_reserve(skb, MAX_TCP_HEADER);
3541
3542 switch (synack_type) {
3543 case TCP_SYNACK_NORMAL:
3544 skb_set_owner_w(skb, req_to_sk(req));
3545 break;
3546 case TCP_SYNACK_COOKIE:
3547 /* Under synflood, we do not attach skb to a socket,
3548 * to avoid false sharing.
3549 */
3550 break;
3551 case TCP_SYNACK_FASTOPEN:
3552 /* sk is a const pointer, because we want to express multiple
3553 * cpu might call us concurrently.
3554 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3555 */
3556 skb_set_owner_w(skb, (struct sock *)sk);
3557 break;
3558 }
3559 skb_dst_set(skb, dst);
3560
3561 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3562
3563 memset(&opts, 0, sizeof(opts));
3564 now = tcp_clock_ns();
3565#ifdef CONFIG_SYN_COOKIES
3566 if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok))
3567 skb_set_delivery_time(skb, cookie_init_timestamp(req, now),
3568 true);
3569 else
3570#endif
3571 {
3572 skb_set_delivery_time(skb, now, true);
3573 if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3574 tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3575 }
3576
3577#ifdef CONFIG_TCP_MD5SIG
3578 rcu_read_lock();
3579 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3580#endif
3581 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3582 /* bpf program will be interested in the tcp_flags */
3583 TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK;
3584 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3585 foc, synack_type,
3586 syn_skb) + sizeof(*th);
3587
3588 skb_push(skb, tcp_header_size);
3589 skb_reset_transport_header(skb);
3590
3591 th = (struct tcphdr *)skb->data;
3592 memset(th, 0, sizeof(struct tcphdr));
3593 th->syn = 1;
3594 th->ack = 1;
3595 tcp_ecn_make_synack(req, th);
3596 th->source = htons(ireq->ir_num);
3597 th->dest = ireq->ir_rmt_port;
3598 skb->mark = ireq->ir_mark;
3599 skb->ip_summed = CHECKSUM_PARTIAL;
3600 th->seq = htonl(tcp_rsk(req)->snt_isn);
3601 /* XXX data is queued and acked as is. No buffer/window check */
3602 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3603
3604 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3605 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3606 tcp_options_write(th, NULL, &opts);
3607 th->doff = (tcp_header_size >> 2);
3608 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3609
3610#ifdef CONFIG_TCP_MD5SIG
3611 /* Okay, we have all we need - do the md5 hash if needed */
3612 if (md5)
3613 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3614 md5, req_to_sk(req), skb);
3615 rcu_read_unlock();
3616#endif
3617
3618 bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb,
3619 synack_type, &opts);
3620
3621 skb_set_delivery_time(skb, now, true);
3622 tcp_add_tx_delay(skb, tp);
3623
3624 return skb;
3625}
3626EXPORT_SYMBOL(tcp_make_synack);
3627
3628static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3629{
3630 struct inet_connection_sock *icsk = inet_csk(sk);
3631 const struct tcp_congestion_ops *ca;
3632 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3633
3634 if (ca_key == TCP_CA_UNSPEC)
3635 return;
3636
3637 rcu_read_lock();
3638 ca = tcp_ca_find_key(ca_key);
3639 if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
3640 bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
3641 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3642 icsk->icsk_ca_ops = ca;
3643 }
3644 rcu_read_unlock();
3645}
3646
3647/* Do all connect socket setups that can be done AF independent. */
3648static void tcp_connect_init(struct sock *sk)
3649{
3650 const struct dst_entry *dst = __sk_dst_get(sk);
3651 struct tcp_sock *tp = tcp_sk(sk);
3652 __u8 rcv_wscale;
3653 u32 rcv_wnd;
3654
3655 /* We'll fix this up when we get a response from the other end.
3656 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3657 */
3658 tp->tcp_header_len = sizeof(struct tcphdr);
3659 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps))
3660 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3661
3662#ifdef CONFIG_TCP_MD5SIG
3663 if (tp->af_specific->md5_lookup(sk, sk))
3664 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3665#endif
3666
3667 /* If user gave his TCP_MAXSEG, record it to clamp */
3668 if (tp->rx_opt.user_mss)
3669 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3670 tp->max_window = 0;
3671 tcp_mtup_init(sk);
3672 tcp_sync_mss(sk, dst_mtu(dst));
3673
3674 tcp_ca_dst_init(sk, dst);
3675
3676 if (!tp->window_clamp)
3677 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3678 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3679
3680 tcp_initialize_rcv_mss(sk);
3681
3682 /* limit the window selection if the user enforce a smaller rx buffer */
3683 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3684 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3685 tp->window_clamp = tcp_full_space(sk);
3686
3687 rcv_wnd = tcp_rwnd_init_bpf(sk);
3688 if (rcv_wnd == 0)
3689 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3690
3691 tcp_select_initial_window(sk, tcp_full_space(sk),
3692 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3693 &tp->rcv_wnd,
3694 &tp->window_clamp,
3695 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling),
3696 &rcv_wscale,
3697 rcv_wnd);
3698
3699 tp->rx_opt.rcv_wscale = rcv_wscale;
3700 tp->rcv_ssthresh = tp->rcv_wnd;
3701
3702 sk->sk_err = 0;
3703 sock_reset_flag(sk, SOCK_DONE);
3704 tp->snd_wnd = 0;
3705 tcp_init_wl(tp, 0);
3706 tcp_write_queue_purge(sk);
3707 tp->snd_una = tp->write_seq;
3708 tp->snd_sml = tp->write_seq;
3709 tp->snd_up = tp->write_seq;
3710 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3711
3712 if (likely(!tp->repair))
3713 tp->rcv_nxt = 0;
3714 else
3715 tp->rcv_tstamp = tcp_jiffies32;
3716 tp->rcv_wup = tp->rcv_nxt;
3717 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3718
3719 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3720 inet_csk(sk)->icsk_retransmits = 0;
3721 tcp_clear_retrans(tp);
3722}
3723
3724static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3725{
3726 struct tcp_sock *tp = tcp_sk(sk);
3727 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3728
3729 tcb->end_seq += skb->len;
3730 __skb_header_release(skb);
3731 sk_wmem_queued_add(sk, skb->truesize);
3732 sk_mem_charge(sk, skb->truesize);
3733 WRITE_ONCE(tp->write_seq, tcb->end_seq);
3734 tp->packets_out += tcp_skb_pcount(skb);
3735}
3736
3737/* Build and send a SYN with data and (cached) Fast Open cookie. However,
3738 * queue a data-only packet after the regular SYN, such that regular SYNs
3739 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3740 * only the SYN sequence, the data are retransmitted in the first ACK.
3741 * If cookie is not cached or other error occurs, falls back to send a
3742 * regular SYN with Fast Open cookie request option.
3743 */
3744static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3745{
3746 struct inet_connection_sock *icsk = inet_csk(sk);
3747 struct tcp_sock *tp = tcp_sk(sk);
3748 struct tcp_fastopen_request *fo = tp->fastopen_req;
3749 int space, err = 0;
3750 struct sk_buff *syn_data;
3751
3752 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3753 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3754 goto fallback;
3755
3756 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3757 * user-MSS. Reserve maximum option space for middleboxes that add
3758 * private TCP options. The cost is reduced data space in SYN :(
3759 */
3760 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3761 /* Sync mss_cache after updating the mss_clamp */
3762 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
3763
3764 space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) -
3765 MAX_TCP_OPTION_SPACE;
3766
3767 space = min_t(size_t, space, fo->size);
3768
3769 /* limit to order-0 allocations */
3770 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3771
3772 syn_data = tcp_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3773 if (!syn_data)
3774 goto fallback;
3775 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3776 if (space) {
3777 int copied = copy_from_iter(skb_put(syn_data, space), space,
3778 &fo->data->msg_iter);
3779 if (unlikely(!copied)) {
3780 tcp_skb_tsorted_anchor_cleanup(syn_data);
3781 kfree_skb(syn_data);
3782 goto fallback;
3783 }
3784 if (copied != space) {
3785 skb_trim(syn_data, copied);
3786 space = copied;
3787 }
3788 skb_zcopy_set(syn_data, fo->uarg, NULL);
3789 }
3790 /* No more data pending in inet_wait_for_connect() */
3791 if (space == fo->size)
3792 fo->data = NULL;
3793 fo->copied = space;
3794
3795 tcp_connect_queue_skb(sk, syn_data);
3796 if (syn_data->len)
3797 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3798
3799 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3800
3801 skb_set_delivery_time(syn, syn_data->skb_mstamp_ns, true);
3802
3803 /* Now full SYN+DATA was cloned and sent (or not),
3804 * remove the SYN from the original skb (syn_data)
3805 * we keep in write queue in case of a retransmit, as we
3806 * also have the SYN packet (with no data) in the same queue.
3807 */
3808 TCP_SKB_CB(syn_data)->seq++;
3809 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3810 if (!err) {
3811 tp->syn_data = (fo->copied > 0);
3812 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3813 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3814 goto done;
3815 }
3816
3817 /* data was not sent, put it in write_queue */
3818 __skb_queue_tail(&sk->sk_write_queue, syn_data);
3819 tp->packets_out -= tcp_skb_pcount(syn_data);
3820
3821fallback:
3822 /* Send a regular SYN with Fast Open cookie request option */
3823 if (fo->cookie.len > 0)
3824 fo->cookie.len = 0;
3825 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3826 if (err)
3827 tp->syn_fastopen = 0;
3828done:
3829 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3830 return err;
3831}
3832
3833/* Build a SYN and send it off. */
3834int tcp_connect(struct sock *sk)
3835{
3836 struct tcp_sock *tp = tcp_sk(sk);
3837 struct sk_buff *buff;
3838 int err;
3839
3840 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3841
3842 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3843 return -EHOSTUNREACH; /* Routing failure or similar. */
3844
3845 tcp_connect_init(sk);
3846
3847 if (unlikely(tp->repair)) {
3848 tcp_finish_connect(sk, NULL);
3849 return 0;
3850 }
3851
3852 buff = tcp_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3853 if (unlikely(!buff))
3854 return -ENOBUFS;
3855
3856 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3857 tcp_mstamp_refresh(tp);
3858 tp->retrans_stamp = tcp_time_stamp(tp);
3859 tcp_connect_queue_skb(sk, buff);
3860 tcp_ecn_send_syn(sk, buff);
3861 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3862
3863 /* Send off SYN; include data in Fast Open. */
3864 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3865 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3866 if (err == -ECONNREFUSED)
3867 return err;
3868
3869 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3870 * in order to make this packet get counted in tcpOutSegs.
3871 */
3872 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3873 tp->pushed_seq = tp->write_seq;
3874 buff = tcp_send_head(sk);
3875 if (unlikely(buff)) {
3876 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
3877 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3878 }
3879 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3880
3881 /* Timer for repeating the SYN until an answer. */
3882 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3883 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3884 return 0;
3885}
3886EXPORT_SYMBOL(tcp_connect);
3887
3888/* Send out a delayed ack, the caller does the policy checking
3889 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3890 * for details.
3891 */
3892void tcp_send_delayed_ack(struct sock *sk)
3893{
3894 struct inet_connection_sock *icsk = inet_csk(sk);
3895 int ato = icsk->icsk_ack.ato;
3896 unsigned long timeout;
3897
3898 if (ato > TCP_DELACK_MIN) {
3899 const struct tcp_sock *tp = tcp_sk(sk);
3900 int max_ato = HZ / 2;
3901
3902 if (inet_csk_in_pingpong_mode(sk) ||
3903 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3904 max_ato = TCP_DELACK_MAX;
3905
3906 /* Slow path, intersegment interval is "high". */
3907
3908 /* If some rtt estimate is known, use it to bound delayed ack.
3909 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3910 * directly.
3911 */
3912 if (tp->srtt_us) {
3913 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3914 TCP_DELACK_MIN);
3915
3916 if (rtt < max_ato)
3917 max_ato = rtt;
3918 }
3919
3920 ato = min(ato, max_ato);
3921 }
3922
3923 ato = min_t(u32, ato, inet_csk(sk)->icsk_delack_max);
3924
3925 /* Stay within the limit we were given */
3926 timeout = jiffies + ato;
3927
3928 /* Use new timeout only if there wasn't a older one earlier. */
3929 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3930 /* If delack timer is about to expire, send ACK now. */
3931 if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3932 tcp_send_ack(sk);
3933 return;
3934 }
3935
3936 if (!time_before(timeout, icsk->icsk_ack.timeout))
3937 timeout = icsk->icsk_ack.timeout;
3938 }
3939 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3940 icsk->icsk_ack.timeout = timeout;
3941 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3942}
3943
3944/* This routine sends an ack and also updates the window. */
3945void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3946{
3947 struct sk_buff *buff;
3948
3949 /* If we have been reset, we may not send again. */
3950 if (sk->sk_state == TCP_CLOSE)
3951 return;
3952
3953 /* We are not putting this on the write queue, so
3954 * tcp_transmit_skb() will set the ownership to this
3955 * sock.
3956 */
3957 buff = alloc_skb(MAX_TCP_HEADER,
3958 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3959 if (unlikely(!buff)) {
3960 struct inet_connection_sock *icsk = inet_csk(sk);
3961 unsigned long delay;
3962
3963 delay = TCP_DELACK_MAX << icsk->icsk_ack.retry;
3964 if (delay < TCP_RTO_MAX)
3965 icsk->icsk_ack.retry++;
3966 inet_csk_schedule_ack(sk);
3967 icsk->icsk_ack.ato = TCP_ATO_MIN;
3968 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX);
3969 return;
3970 }
3971
3972 /* Reserve space for headers and prepare control bits. */
3973 skb_reserve(buff, MAX_TCP_HEADER);
3974 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3975
3976 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3977 * too much.
3978 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3979 */
3980 skb_set_tcp_pure_ack(buff);
3981
3982 /* Send it off, this clears delayed acks for us. */
3983 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3984}
3985EXPORT_SYMBOL_GPL(__tcp_send_ack);
3986
3987void tcp_send_ack(struct sock *sk)
3988{
3989 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3990}
3991
3992/* This routine sends a packet with an out of date sequence
3993 * number. It assumes the other end will try to ack it.
3994 *
3995 * Question: what should we make while urgent mode?
3996 * 4.4BSD forces sending single byte of data. We cannot send
3997 * out of window data, because we have SND.NXT==SND.MAX...
3998 *
3999 * Current solution: to send TWO zero-length segments in urgent mode:
4000 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
4001 * out-of-date with SND.UNA-1 to probe window.
4002 */
4003static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
4004{
4005 struct tcp_sock *tp = tcp_sk(sk);
4006 struct sk_buff *skb;
4007
4008 /* We don't queue it, tcp_transmit_skb() sets ownership. */
4009 skb = alloc_skb(MAX_TCP_HEADER,
4010 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4011 if (!skb)
4012 return -1;
4013
4014 /* Reserve space for headers and set control bits. */
4015 skb_reserve(skb, MAX_TCP_HEADER);
4016 /* Use a previous sequence. This should cause the other
4017 * end to send an ack. Don't queue or clone SKB, just
4018 * send it.
4019 */
4020 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
4021 NET_INC_STATS(sock_net(sk), mib);
4022 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
4023}
4024
4025/* Called from setsockopt( ... TCP_REPAIR ) */
4026void tcp_send_window_probe(struct sock *sk)
4027{
4028 if (sk->sk_state == TCP_ESTABLISHED) {
4029 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
4030 tcp_mstamp_refresh(tcp_sk(sk));
4031 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
4032 }
4033}
4034
4035/* Initiate keepalive or window probe from timer. */
4036int tcp_write_wakeup(struct sock *sk, int mib)
4037{
4038 struct tcp_sock *tp = tcp_sk(sk);
4039 struct sk_buff *skb;
4040
4041 if (sk->sk_state == TCP_CLOSE)
4042 return -1;
4043
4044 skb = tcp_send_head(sk);
4045 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
4046 int err;
4047 unsigned int mss = tcp_current_mss(sk);
4048 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
4049
4050 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
4051 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
4052
4053 /* We are probing the opening of a window
4054 * but the window size is != 0
4055 * must have been a result SWS avoidance ( sender )
4056 */
4057 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
4058 skb->len > mss) {
4059 seg_size = min(seg_size, mss);
4060 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4061 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
4062 skb, seg_size, mss, GFP_ATOMIC))
4063 return -1;
4064 } else if (!tcp_skb_pcount(skb))
4065 tcp_set_skb_tso_segs(skb, mss);
4066
4067 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4068 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
4069 if (!err)
4070 tcp_event_new_data_sent(sk, skb);
4071 return err;
4072 } else {
4073 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
4074 tcp_xmit_probe_skb(sk, 1, mib);
4075 return tcp_xmit_probe_skb(sk, 0, mib);
4076 }
4077}
4078
4079/* A window probe timeout has occurred. If window is not closed send
4080 * a partial packet else a zero probe.
4081 */
4082void tcp_send_probe0(struct sock *sk)
4083{
4084 struct inet_connection_sock *icsk = inet_csk(sk);
4085 struct tcp_sock *tp = tcp_sk(sk);
4086 struct net *net = sock_net(sk);
4087 unsigned long timeout;
4088 int err;
4089
4090 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
4091
4092 if (tp->packets_out || tcp_write_queue_empty(sk)) {
4093 /* Cancel probe timer, if it is not required. */
4094 icsk->icsk_probes_out = 0;
4095 icsk->icsk_backoff = 0;
4096 icsk->icsk_probes_tstamp = 0;
4097 return;
4098 }
4099
4100 icsk->icsk_probes_out++;
4101 if (err <= 0) {
4102 if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2))
4103 icsk->icsk_backoff++;
4104 timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
4105 } else {
4106 /* If packet was not sent due to local congestion,
4107 * Let senders fight for local resources conservatively.
4108 */
4109 timeout = TCP_RESOURCE_PROBE_INTERVAL;
4110 }
4111
4112 timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout);
4113 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX);
4114}
4115
4116int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
4117{
4118 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
4119 struct flowi fl;
4120 int res;
4121
4122 /* Paired with WRITE_ONCE() in sock_setsockopt() */
4123 if (READ_ONCE(sk->sk_txrehash) == SOCK_TXREHASH_ENABLED)
4124 tcp_rsk(req)->txhash = net_tx_rndhash();
4125 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL,
4126 NULL);
4127 if (!res) {
4128 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
4129 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
4130 if (unlikely(tcp_passive_fastopen(sk)))
4131 tcp_sk(sk)->total_retrans++;
4132 trace_tcp_retransmit_synack(sk, req);
4133 }
4134 return res;
4135}
4136EXPORT_SYMBOL(tcp_rtx_synack);