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