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/*
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
45/* People can turn this off for buggy TCP's found in printers etc. */
46int sysctl_tcp_retrans_collapse __read_mostly = 1;
47
48/* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
50 */
51int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
52
53/* Default TSQ limit of four TSO segments */
54int sysctl_tcp_limit_output_bytes __read_mostly = 262144;
55
56/* This limits the percentage of the congestion window which we
57 * will allow a single TSO frame to consume. Building TSO frames
58 * which are too large can cause TCP streams to be bursty.
59 */
60int sysctl_tcp_tso_win_divisor __read_mostly = 3;
61
62/* By default, RFC2861 behavior. */
63int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
64
65static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
66 int push_one, gfp_t gfp);
67
68/* Account for new data that has been sent to the network. */
69static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
70{
71 struct inet_connection_sock *icsk = inet_csk(sk);
72 struct tcp_sock *tp = tcp_sk(sk);
73 unsigned int prior_packets = tp->packets_out;
74
75 tcp_advance_send_head(sk, skb);
76 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
77
78 tp->packets_out += tcp_skb_pcount(skb);
79 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
80 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
81 tcp_rearm_rto(sk);
82 }
83
84 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
85 tcp_skb_pcount(skb));
86}
87
88/* SND.NXT, if window was not shrunk.
89 * If window has been shrunk, what should we make? It is not clear at all.
90 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
91 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
92 * invalid. OK, let's make this for now:
93 */
94static inline __u32 tcp_acceptable_seq(const struct sock *sk)
95{
96 const struct tcp_sock *tp = tcp_sk(sk);
97
98 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
99 return tp->snd_nxt;
100 else
101 return tcp_wnd_end(tp);
102}
103
104/* Calculate mss to advertise in SYN segment.
105 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
106 *
107 * 1. It is independent of path mtu.
108 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
109 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
110 * attached devices, because some buggy hosts are confused by
111 * large MSS.
112 * 4. We do not make 3, we advertise MSS, calculated from first
113 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
114 * This may be overridden via information stored in routing table.
115 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
116 * probably even Jumbo".
117 */
118static __u16 tcp_advertise_mss(struct sock *sk)
119{
120 struct tcp_sock *tp = tcp_sk(sk);
121 const struct dst_entry *dst = __sk_dst_get(sk);
122 int mss = tp->advmss;
123
124 if (dst) {
125 unsigned int metric = dst_metric_advmss(dst);
126
127 if (metric < mss) {
128 mss = metric;
129 tp->advmss = mss;
130 }
131 }
132
133 return (__u16)mss;
134}
135
136/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
137 * This is the first part of cwnd validation mechanism.
138 */
139void tcp_cwnd_restart(struct sock *sk, s32 delta)
140{
141 struct tcp_sock *tp = tcp_sk(sk);
142 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
143 u32 cwnd = tp->snd_cwnd;
144
145 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
146
147 tp->snd_ssthresh = tcp_current_ssthresh(sk);
148 restart_cwnd = min(restart_cwnd, cwnd);
149
150 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
151 cwnd >>= 1;
152 tp->snd_cwnd = max(cwnd, restart_cwnd);
153 tp->snd_cwnd_stamp = tcp_time_stamp;
154 tp->snd_cwnd_used = 0;
155}
156
157/* Congestion state accounting after a packet has been sent. */
158static void tcp_event_data_sent(struct tcp_sock *tp,
159 struct sock *sk)
160{
161 struct inet_connection_sock *icsk = inet_csk(sk);
162 const u32 now = tcp_time_stamp;
163
164 if (tcp_packets_in_flight(tp) == 0)
165 tcp_ca_event(sk, CA_EVENT_TX_START);
166
167 tp->lsndtime = now;
168
169 /* If it is a reply for ato after last received
170 * packet, enter pingpong mode.
171 */
172 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
173 icsk->icsk_ack.pingpong = 1;
174}
175
176/* Account for an ACK we sent. */
177static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
178{
179 tcp_dec_quickack_mode(sk, pkts);
180 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
181}
182
183
184u32 tcp_default_init_rwnd(u32 mss)
185{
186 /* Initial receive window should be twice of TCP_INIT_CWND to
187 * enable proper sending of new unsent data during fast recovery
188 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
189 * limit when mss is larger than 1460.
190 */
191 u32 init_rwnd = TCP_INIT_CWND * 2;
192
193 if (mss > 1460)
194 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
195 return init_rwnd;
196}
197
198/* Determine a window scaling and initial window to offer.
199 * Based on the assumption that the given amount of space
200 * will be offered. Store the results in the tp structure.
201 * NOTE: for smooth operation initial space offering should
202 * be a multiple of mss if possible. We assume here that mss >= 1.
203 * This MUST be enforced by all callers.
204 */
205void tcp_select_initial_window(int __space, __u32 mss,
206 __u32 *rcv_wnd, __u32 *window_clamp,
207 int wscale_ok, __u8 *rcv_wscale,
208 __u32 init_rcv_wnd)
209{
210 unsigned int space = (__space < 0 ? 0 : __space);
211
212 /* If no clamp set the clamp to the max possible scaled window */
213 if (*window_clamp == 0)
214 (*window_clamp) = (65535 << 14);
215 space = min(*window_clamp, space);
216
217 /* Quantize space offering to a multiple of mss if possible. */
218 if (space > mss)
219 space = (space / mss) * mss;
220
221 /* NOTE: offering an initial window larger than 32767
222 * will break some buggy TCP stacks. If the admin tells us
223 * it is likely we could be speaking with such a buggy stack
224 * we will truncate our initial window offering to 32K-1
225 * unless the remote has sent us a window scaling option,
226 * which we interpret as a sign the remote TCP is not
227 * misinterpreting the window field as a signed quantity.
228 */
229 if (sysctl_tcp_workaround_signed_windows)
230 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
231 else
232 (*rcv_wnd) = space;
233
234 (*rcv_wscale) = 0;
235 if (wscale_ok) {
236 /* Set window scaling on max possible window
237 * See RFC1323 for an explanation of the limit to 14
238 */
239 space = max_t(u32, space, sysctl_tcp_rmem[2]);
240 space = max_t(u32, space, sysctl_rmem_max);
241 space = min_t(u32, space, *window_clamp);
242 while (space > 65535 && (*rcv_wscale) < 14) {
243 space >>= 1;
244 (*rcv_wscale)++;
245 }
246 }
247
248 if (mss > (1 << *rcv_wscale)) {
249 if (!init_rcv_wnd) /* Use default unless specified otherwise */
250 init_rcv_wnd = tcp_default_init_rwnd(mss);
251 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
252 }
253
254 /* Set the clamp no higher than max representable value */
255 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
256}
257EXPORT_SYMBOL(tcp_select_initial_window);
258
259/* Chose a new window to advertise, update state in tcp_sock for the
260 * socket, and return result with RFC1323 scaling applied. The return
261 * value can be stuffed directly into th->window for an outgoing
262 * frame.
263 */
264static u16 tcp_select_window(struct sock *sk)
265{
266 struct tcp_sock *tp = tcp_sk(sk);
267 u32 old_win = tp->rcv_wnd;
268 u32 cur_win = tcp_receive_window(tp);
269 u32 new_win = __tcp_select_window(sk);
270
271 /* Never shrink the offered window */
272 if (new_win < cur_win) {
273 /* Danger Will Robinson!
274 * Don't update rcv_wup/rcv_wnd here or else
275 * we will not be able to advertise a zero
276 * window in time. --DaveM
277 *
278 * Relax Will Robinson.
279 */
280 if (new_win == 0)
281 NET_INC_STATS(sock_net(sk),
282 LINUX_MIB_TCPWANTZEROWINDOWADV);
283 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
284 }
285 tp->rcv_wnd = new_win;
286 tp->rcv_wup = tp->rcv_nxt;
287
288 /* Make sure we do not exceed the maximum possible
289 * scaled window.
290 */
291 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
292 new_win = min(new_win, MAX_TCP_WINDOW);
293 else
294 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
295
296 /* RFC1323 scaling applied */
297 new_win >>= tp->rx_opt.rcv_wscale;
298
299 /* If we advertise zero window, disable fast path. */
300 if (new_win == 0) {
301 tp->pred_flags = 0;
302 if (old_win)
303 NET_INC_STATS(sock_net(sk),
304 LINUX_MIB_TCPTOZEROWINDOWADV);
305 } else if (old_win == 0) {
306 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
307 }
308
309 return new_win;
310}
311
312/* Packet ECN state for a SYN-ACK */
313static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
314{
315 const struct tcp_sock *tp = tcp_sk(sk);
316
317 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
318 if (!(tp->ecn_flags & TCP_ECN_OK))
319 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
320 else if (tcp_ca_needs_ecn(sk))
321 INET_ECN_xmit(sk);
322}
323
324/* Packet ECN state for a SYN. */
325static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
326{
327 struct tcp_sock *tp = tcp_sk(sk);
328 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
329 tcp_ca_needs_ecn(sk);
330
331 if (!use_ecn) {
332 const struct dst_entry *dst = __sk_dst_get(sk);
333
334 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
335 use_ecn = true;
336 }
337
338 tp->ecn_flags = 0;
339
340 if (use_ecn) {
341 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
342 tp->ecn_flags = TCP_ECN_OK;
343 if (tcp_ca_needs_ecn(sk))
344 INET_ECN_xmit(sk);
345 }
346}
347
348static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
349{
350 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
351 /* tp->ecn_flags are cleared at a later point in time when
352 * SYN ACK is ultimatively being received.
353 */
354 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
355}
356
357static void
358tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
359{
360 if (inet_rsk(req)->ecn_ok)
361 th->ece = 1;
362}
363
364/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
365 * be sent.
366 */
367static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
368 struct tcphdr *th, int tcp_header_len)
369{
370 struct tcp_sock *tp = tcp_sk(sk);
371
372 if (tp->ecn_flags & TCP_ECN_OK) {
373 /* Not-retransmitted data segment: set ECT and inject CWR. */
374 if (skb->len != tcp_header_len &&
375 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
376 INET_ECN_xmit(sk);
377 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
378 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
379 th->cwr = 1;
380 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
381 }
382 } else if (!tcp_ca_needs_ecn(sk)) {
383 /* ACK or retransmitted segment: clear ECT|CE */
384 INET_ECN_dontxmit(sk);
385 }
386 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
387 th->ece = 1;
388 }
389}
390
391/* Constructs common control bits of non-data skb. If SYN/FIN is present,
392 * auto increment end seqno.
393 */
394static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
395{
396 skb->ip_summed = CHECKSUM_PARTIAL;
397 skb->csum = 0;
398
399 TCP_SKB_CB(skb)->tcp_flags = flags;
400 TCP_SKB_CB(skb)->sacked = 0;
401
402 tcp_skb_pcount_set(skb, 1);
403
404 TCP_SKB_CB(skb)->seq = seq;
405 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
406 seq++;
407 TCP_SKB_CB(skb)->end_seq = seq;
408}
409
410static inline bool tcp_urg_mode(const struct tcp_sock *tp)
411{
412 return tp->snd_una != tp->snd_up;
413}
414
415#define OPTION_SACK_ADVERTISE (1 << 0)
416#define OPTION_TS (1 << 1)
417#define OPTION_MD5 (1 << 2)
418#define OPTION_WSCALE (1 << 3)
419#define OPTION_FAST_OPEN_COOKIE (1 << 8)
420
421struct tcp_out_options {
422 u16 options; /* bit field of OPTION_* */
423 u16 mss; /* 0 to disable */
424 u8 ws; /* window scale, 0 to disable */
425 u8 num_sack_blocks; /* number of SACK blocks to include */
426 u8 hash_size; /* bytes in hash_location */
427 __u8 *hash_location; /* temporary pointer, overloaded */
428 __u32 tsval, tsecr; /* need to include OPTION_TS */
429 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
430};
431
432/* Write previously computed TCP options to the packet.
433 *
434 * Beware: Something in the Internet is very sensitive to the ordering of
435 * TCP options, we learned this through the hard way, so be careful here.
436 * Luckily we can at least blame others for their non-compliance but from
437 * inter-operability perspective it seems that we're somewhat stuck with
438 * the ordering which we have been using if we want to keep working with
439 * those broken things (not that it currently hurts anybody as there isn't
440 * particular reason why the ordering would need to be changed).
441 *
442 * At least SACK_PERM as the first option is known to lead to a disaster
443 * (but it may well be that other scenarios fail similarly).
444 */
445static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
446 struct tcp_out_options *opts)
447{
448 u16 options = opts->options; /* mungable copy */
449
450 if (unlikely(OPTION_MD5 & options)) {
451 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
452 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
453 /* overload cookie hash location */
454 opts->hash_location = (__u8 *)ptr;
455 ptr += 4;
456 }
457
458 if (unlikely(opts->mss)) {
459 *ptr++ = htonl((TCPOPT_MSS << 24) |
460 (TCPOLEN_MSS << 16) |
461 opts->mss);
462 }
463
464 if (likely(OPTION_TS & options)) {
465 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
466 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
467 (TCPOLEN_SACK_PERM << 16) |
468 (TCPOPT_TIMESTAMP << 8) |
469 TCPOLEN_TIMESTAMP);
470 options &= ~OPTION_SACK_ADVERTISE;
471 } else {
472 *ptr++ = htonl((TCPOPT_NOP << 24) |
473 (TCPOPT_NOP << 16) |
474 (TCPOPT_TIMESTAMP << 8) |
475 TCPOLEN_TIMESTAMP);
476 }
477 *ptr++ = htonl(opts->tsval);
478 *ptr++ = htonl(opts->tsecr);
479 }
480
481 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
482 *ptr++ = htonl((TCPOPT_NOP << 24) |
483 (TCPOPT_NOP << 16) |
484 (TCPOPT_SACK_PERM << 8) |
485 TCPOLEN_SACK_PERM);
486 }
487
488 if (unlikely(OPTION_WSCALE & options)) {
489 *ptr++ = htonl((TCPOPT_NOP << 24) |
490 (TCPOPT_WINDOW << 16) |
491 (TCPOLEN_WINDOW << 8) |
492 opts->ws);
493 }
494
495 if (unlikely(opts->num_sack_blocks)) {
496 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
497 tp->duplicate_sack : tp->selective_acks;
498 int this_sack;
499
500 *ptr++ = htonl((TCPOPT_NOP << 24) |
501 (TCPOPT_NOP << 16) |
502 (TCPOPT_SACK << 8) |
503 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
504 TCPOLEN_SACK_PERBLOCK)));
505
506 for (this_sack = 0; this_sack < opts->num_sack_blocks;
507 ++this_sack) {
508 *ptr++ = htonl(sp[this_sack].start_seq);
509 *ptr++ = htonl(sp[this_sack].end_seq);
510 }
511
512 tp->rx_opt.dsack = 0;
513 }
514
515 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
516 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
517 u8 *p = (u8 *)ptr;
518 u32 len; /* Fast Open option length */
519
520 if (foc->exp) {
521 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
522 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
523 TCPOPT_FASTOPEN_MAGIC);
524 p += TCPOLEN_EXP_FASTOPEN_BASE;
525 } else {
526 len = TCPOLEN_FASTOPEN_BASE + foc->len;
527 *p++ = TCPOPT_FASTOPEN;
528 *p++ = len;
529 }
530
531 memcpy(p, foc->val, foc->len);
532 if ((len & 3) == 2) {
533 p[foc->len] = TCPOPT_NOP;
534 p[foc->len + 1] = TCPOPT_NOP;
535 }
536 ptr += (len + 3) >> 2;
537 }
538}
539
540/* Compute TCP options for SYN packets. This is not the final
541 * network wire format yet.
542 */
543static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
544 struct tcp_out_options *opts,
545 struct tcp_md5sig_key **md5)
546{
547 struct tcp_sock *tp = tcp_sk(sk);
548 unsigned int remaining = MAX_TCP_OPTION_SPACE;
549 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
550
551#ifdef CONFIG_TCP_MD5SIG
552 *md5 = tp->af_specific->md5_lookup(sk, sk);
553 if (*md5) {
554 opts->options |= OPTION_MD5;
555 remaining -= TCPOLEN_MD5SIG_ALIGNED;
556 }
557#else
558 *md5 = NULL;
559#endif
560
561 /* We always get an MSS option. The option bytes which will be seen in
562 * normal data packets should timestamps be used, must be in the MSS
563 * advertised. But we subtract them from tp->mss_cache so that
564 * calculations in tcp_sendmsg are simpler etc. So account for this
565 * fact here if necessary. If we don't do this correctly, as a
566 * receiver we won't recognize data packets as being full sized when we
567 * should, and thus we won't abide by the delayed ACK rules correctly.
568 * SACKs don't matter, we never delay an ACK when we have any of those
569 * going out. */
570 opts->mss = tcp_advertise_mss(sk);
571 remaining -= TCPOLEN_MSS_ALIGNED;
572
573 if (likely(sysctl_tcp_timestamps && !*md5)) {
574 opts->options |= OPTION_TS;
575 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
576 opts->tsecr = tp->rx_opt.ts_recent;
577 remaining -= TCPOLEN_TSTAMP_ALIGNED;
578 }
579 if (likely(sysctl_tcp_window_scaling)) {
580 opts->ws = tp->rx_opt.rcv_wscale;
581 opts->options |= OPTION_WSCALE;
582 remaining -= TCPOLEN_WSCALE_ALIGNED;
583 }
584 if (likely(sysctl_tcp_sack)) {
585 opts->options |= OPTION_SACK_ADVERTISE;
586 if (unlikely(!(OPTION_TS & opts->options)))
587 remaining -= TCPOLEN_SACKPERM_ALIGNED;
588 }
589
590 if (fastopen && fastopen->cookie.len >= 0) {
591 u32 need = fastopen->cookie.len;
592
593 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
594 TCPOLEN_FASTOPEN_BASE;
595 need = (need + 3) & ~3U; /* Align to 32 bits */
596 if (remaining >= need) {
597 opts->options |= OPTION_FAST_OPEN_COOKIE;
598 opts->fastopen_cookie = &fastopen->cookie;
599 remaining -= need;
600 tp->syn_fastopen = 1;
601 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
602 }
603 }
604
605 return MAX_TCP_OPTION_SPACE - remaining;
606}
607
608/* Set up TCP options for SYN-ACKs. */
609static unsigned int tcp_synack_options(struct request_sock *req,
610 unsigned int mss, struct sk_buff *skb,
611 struct tcp_out_options *opts,
612 const struct tcp_md5sig_key *md5,
613 struct tcp_fastopen_cookie *foc)
614{
615 struct inet_request_sock *ireq = inet_rsk(req);
616 unsigned int remaining = MAX_TCP_OPTION_SPACE;
617
618#ifdef CONFIG_TCP_MD5SIG
619 if (md5) {
620 opts->options |= OPTION_MD5;
621 remaining -= TCPOLEN_MD5SIG_ALIGNED;
622
623 /* We can't fit any SACK blocks in a packet with MD5 + TS
624 * options. There was discussion about disabling SACK
625 * rather than TS in order to fit in better with old,
626 * buggy kernels, but that was deemed to be unnecessary.
627 */
628 ireq->tstamp_ok &= !ireq->sack_ok;
629 }
630#endif
631
632 /* We always send an MSS option. */
633 opts->mss = mss;
634 remaining -= TCPOLEN_MSS_ALIGNED;
635
636 if (likely(ireq->wscale_ok)) {
637 opts->ws = ireq->rcv_wscale;
638 opts->options |= OPTION_WSCALE;
639 remaining -= TCPOLEN_WSCALE_ALIGNED;
640 }
641 if (likely(ireq->tstamp_ok)) {
642 opts->options |= OPTION_TS;
643 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
644 opts->tsecr = req->ts_recent;
645 remaining -= TCPOLEN_TSTAMP_ALIGNED;
646 }
647 if (likely(ireq->sack_ok)) {
648 opts->options |= OPTION_SACK_ADVERTISE;
649 if (unlikely(!ireq->tstamp_ok))
650 remaining -= TCPOLEN_SACKPERM_ALIGNED;
651 }
652 if (foc != NULL && foc->len >= 0) {
653 u32 need = foc->len;
654
655 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
656 TCPOLEN_FASTOPEN_BASE;
657 need = (need + 3) & ~3U; /* Align to 32 bits */
658 if (remaining >= need) {
659 opts->options |= OPTION_FAST_OPEN_COOKIE;
660 opts->fastopen_cookie = foc;
661 remaining -= need;
662 }
663 }
664
665 return MAX_TCP_OPTION_SPACE - remaining;
666}
667
668/* Compute TCP options for ESTABLISHED sockets. This is not the
669 * final wire format yet.
670 */
671static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
672 struct tcp_out_options *opts,
673 struct tcp_md5sig_key **md5)
674{
675 struct tcp_sock *tp = tcp_sk(sk);
676 unsigned int size = 0;
677 unsigned int eff_sacks;
678
679 opts->options = 0;
680
681#ifdef CONFIG_TCP_MD5SIG
682 *md5 = tp->af_specific->md5_lookup(sk, sk);
683 if (unlikely(*md5)) {
684 opts->options |= OPTION_MD5;
685 size += TCPOLEN_MD5SIG_ALIGNED;
686 }
687#else
688 *md5 = NULL;
689#endif
690
691 if (likely(tp->rx_opt.tstamp_ok)) {
692 opts->options |= OPTION_TS;
693 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
694 opts->tsecr = tp->rx_opt.ts_recent;
695 size += TCPOLEN_TSTAMP_ALIGNED;
696 }
697
698 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
699 if (unlikely(eff_sacks)) {
700 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
701 opts->num_sack_blocks =
702 min_t(unsigned int, eff_sacks,
703 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
704 TCPOLEN_SACK_PERBLOCK);
705 size += TCPOLEN_SACK_BASE_ALIGNED +
706 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
707 }
708
709 return size;
710}
711
712
713/* TCP SMALL QUEUES (TSQ)
714 *
715 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
716 * to reduce RTT and bufferbloat.
717 * We do this using a special skb destructor (tcp_wfree).
718 *
719 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
720 * needs to be reallocated in a driver.
721 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
722 *
723 * Since transmit from skb destructor is forbidden, we use a tasklet
724 * to process all sockets that eventually need to send more skbs.
725 * We use one tasklet per cpu, with its own queue of sockets.
726 */
727struct tsq_tasklet {
728 struct tasklet_struct tasklet;
729 struct list_head head; /* queue of tcp sockets */
730};
731static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
732
733static void tcp_tsq_handler(struct sock *sk)
734{
735 if ((1 << sk->sk_state) &
736 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
737 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
738 struct tcp_sock *tp = tcp_sk(sk);
739
740 if (tp->lost_out > tp->retrans_out &&
741 tp->snd_cwnd > tcp_packets_in_flight(tp))
742 tcp_xmit_retransmit_queue(sk);
743
744 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
745 0, GFP_ATOMIC);
746 }
747}
748/*
749 * One tasklet per cpu tries to send more skbs.
750 * We run in tasklet context but need to disable irqs when
751 * transferring tsq->head because tcp_wfree() might
752 * interrupt us (non NAPI drivers)
753 */
754static void tcp_tasklet_func(unsigned long data)
755{
756 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
757 LIST_HEAD(list);
758 unsigned long flags;
759 struct list_head *q, *n;
760 struct tcp_sock *tp;
761 struct sock *sk;
762
763 local_irq_save(flags);
764 list_splice_init(&tsq->head, &list);
765 local_irq_restore(flags);
766
767 list_for_each_safe(q, n, &list) {
768 tp = list_entry(q, struct tcp_sock, tsq_node);
769 list_del(&tp->tsq_node);
770
771 sk = (struct sock *)tp;
772 smp_mb__before_atomic();
773 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
774
775 if (!sk->sk_lock.owned &&
776 test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags)) {
777 bh_lock_sock(sk);
778 if (!sock_owned_by_user(sk)) {
779 clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
780 tcp_tsq_handler(sk);
781 }
782 bh_unlock_sock(sk);
783 }
784
785 sk_free(sk);
786 }
787}
788
789#define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
790 TCPF_WRITE_TIMER_DEFERRED | \
791 TCPF_DELACK_TIMER_DEFERRED | \
792 TCPF_MTU_REDUCED_DEFERRED)
793/**
794 * tcp_release_cb - tcp release_sock() callback
795 * @sk: socket
796 *
797 * called from release_sock() to perform protocol dependent
798 * actions before socket release.
799 */
800void tcp_release_cb(struct sock *sk)
801{
802 unsigned long flags, nflags;
803
804 /* perform an atomic operation only if at least one flag is set */
805 do {
806 flags = sk->sk_tsq_flags;
807 if (!(flags & TCP_DEFERRED_ALL))
808 return;
809 nflags = flags & ~TCP_DEFERRED_ALL;
810 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
811
812 if (flags & TCPF_TSQ_DEFERRED)
813 tcp_tsq_handler(sk);
814
815 /* Here begins the tricky part :
816 * We are called from release_sock() with :
817 * 1) BH disabled
818 * 2) sk_lock.slock spinlock held
819 * 3) socket owned by us (sk->sk_lock.owned == 1)
820 *
821 * But following code is meant to be called from BH handlers,
822 * so we should keep BH disabled, but early release socket ownership
823 */
824 sock_release_ownership(sk);
825
826 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
827 tcp_write_timer_handler(sk);
828 __sock_put(sk);
829 }
830 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
831 tcp_delack_timer_handler(sk);
832 __sock_put(sk);
833 }
834 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
835 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
836 __sock_put(sk);
837 }
838}
839EXPORT_SYMBOL(tcp_release_cb);
840
841void __init tcp_tasklet_init(void)
842{
843 int i;
844
845 for_each_possible_cpu(i) {
846 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
847
848 INIT_LIST_HEAD(&tsq->head);
849 tasklet_init(&tsq->tasklet,
850 tcp_tasklet_func,
851 (unsigned long)tsq);
852 }
853}
854
855/*
856 * Write buffer destructor automatically called from kfree_skb.
857 * We can't xmit new skbs from this context, as we might already
858 * hold qdisc lock.
859 */
860void tcp_wfree(struct sk_buff *skb)
861{
862 struct sock *sk = skb->sk;
863 struct tcp_sock *tp = tcp_sk(sk);
864 unsigned long flags, nval, oval;
865 int wmem;
866
867 /* Keep one reference on sk_wmem_alloc.
868 * Will be released by sk_free() from here or tcp_tasklet_func()
869 */
870 wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc);
871
872 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
873 * Wait until our queues (qdisc + devices) are drained.
874 * This gives :
875 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
876 * - chance for incoming ACK (processed by another cpu maybe)
877 * to migrate this flow (skb->ooo_okay will be eventually set)
878 */
879 if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
880 goto out;
881
882 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
883 struct tsq_tasklet *tsq;
884 bool empty;
885
886 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
887 goto out;
888
889 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED | TCPF_TSQ_DEFERRED;
890 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
891 if (nval != oval)
892 continue;
893
894 /* queue this socket to tasklet queue */
895 local_irq_save(flags);
896 tsq = this_cpu_ptr(&tsq_tasklet);
897 empty = list_empty(&tsq->head);
898 list_add(&tp->tsq_node, &tsq->head);
899 if (empty)
900 tasklet_schedule(&tsq->tasklet);
901 local_irq_restore(flags);
902 return;
903 }
904out:
905 sk_free(sk);
906}
907
908/* This routine actually transmits TCP packets queued in by
909 * tcp_do_sendmsg(). This is used by both the initial
910 * transmission and possible later retransmissions.
911 * All SKB's seen here are completely headerless. It is our
912 * job to build the TCP header, and pass the packet down to
913 * IP so it can do the same plus pass the packet off to the
914 * device.
915 *
916 * We are working here with either a clone of the original
917 * SKB, or a fresh unique copy made by the retransmit engine.
918 */
919static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
920 gfp_t gfp_mask)
921{
922 const struct inet_connection_sock *icsk = inet_csk(sk);
923 struct inet_sock *inet;
924 struct tcp_sock *tp;
925 struct tcp_skb_cb *tcb;
926 struct tcp_out_options opts;
927 unsigned int tcp_options_size, tcp_header_size;
928 struct tcp_md5sig_key *md5;
929 struct tcphdr *th;
930 int err;
931
932 BUG_ON(!skb || !tcp_skb_pcount(skb));
933 tp = tcp_sk(sk);
934
935 if (clone_it) {
936 skb_mstamp_get(&skb->skb_mstamp);
937 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
938 - tp->snd_una;
939 tcp_rate_skb_sent(sk, skb);
940
941 if (unlikely(skb_cloned(skb)))
942 skb = pskb_copy(skb, gfp_mask);
943 else
944 skb = skb_clone(skb, gfp_mask);
945 if (unlikely(!skb))
946 return -ENOBUFS;
947 }
948
949 inet = inet_sk(sk);
950 tcb = TCP_SKB_CB(skb);
951 memset(&opts, 0, sizeof(opts));
952
953 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
954 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
955 else
956 tcp_options_size = tcp_established_options(sk, skb, &opts,
957 &md5);
958 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
959
960 /* if no packet is in qdisc/device queue, then allow XPS to select
961 * another queue. We can be called from tcp_tsq_handler()
962 * which holds one reference to sk_wmem_alloc.
963 *
964 * TODO: Ideally, in-flight pure ACK packets should not matter here.
965 * One way to get this would be to set skb->truesize = 2 on them.
966 */
967 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
968
969 skb_push(skb, tcp_header_size);
970 skb_reset_transport_header(skb);
971
972 skb_orphan(skb);
973 skb->sk = sk;
974 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
975 skb_set_hash_from_sk(skb, sk);
976 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
977
978 /* Build TCP header and checksum it. */
979 th = (struct tcphdr *)skb->data;
980 th->source = inet->inet_sport;
981 th->dest = inet->inet_dport;
982 th->seq = htonl(tcb->seq);
983 th->ack_seq = htonl(tp->rcv_nxt);
984 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
985 tcb->tcp_flags);
986
987 th->check = 0;
988 th->urg_ptr = 0;
989
990 /* The urg_mode check is necessary during a below snd_una win probe */
991 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
992 if (before(tp->snd_up, tcb->seq + 0x10000)) {
993 th->urg_ptr = htons(tp->snd_up - tcb->seq);
994 th->urg = 1;
995 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
996 th->urg_ptr = htons(0xFFFF);
997 th->urg = 1;
998 }
999 }
1000
1001 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1002 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1003 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1004 th->window = htons(tcp_select_window(sk));
1005 tcp_ecn_send(sk, skb, th, tcp_header_size);
1006 } else {
1007 /* RFC1323: The window in SYN & SYN/ACK segments
1008 * is never scaled.
1009 */
1010 th->window = htons(min(tp->rcv_wnd, 65535U));
1011 }
1012#ifdef CONFIG_TCP_MD5SIG
1013 /* Calculate the MD5 hash, as we have all we need now */
1014 if (md5) {
1015 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1016 tp->af_specific->calc_md5_hash(opts.hash_location,
1017 md5, sk, skb);
1018 }
1019#endif
1020
1021 icsk->icsk_af_ops->send_check(sk, skb);
1022
1023 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1024 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1025
1026 if (skb->len != tcp_header_size) {
1027 tcp_event_data_sent(tp, sk);
1028 tp->data_segs_out += tcp_skb_pcount(skb);
1029 }
1030
1031 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1032 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1033 tcp_skb_pcount(skb));
1034
1035 tp->segs_out += tcp_skb_pcount(skb);
1036 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1037 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1038 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1039
1040 /* Our usage of tstamp should remain private */
1041 skb->tstamp = 0;
1042
1043 /* Cleanup our debris for IP stacks */
1044 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1045 sizeof(struct inet6_skb_parm)));
1046
1047 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1048
1049 if (likely(err <= 0))
1050 return err;
1051
1052 tcp_enter_cwr(sk);
1053
1054 return net_xmit_eval(err);
1055}
1056
1057/* This routine just queues the buffer for sending.
1058 *
1059 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1060 * otherwise socket can stall.
1061 */
1062static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1063{
1064 struct tcp_sock *tp = tcp_sk(sk);
1065
1066 /* Advance write_seq and place onto the write_queue. */
1067 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1068 __skb_header_release(skb);
1069 tcp_add_write_queue_tail(sk, skb);
1070 sk->sk_wmem_queued += skb->truesize;
1071 sk_mem_charge(sk, skb->truesize);
1072}
1073
1074/* Initialize TSO segments for a packet. */
1075static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1076{
1077 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1078 /* Avoid the costly divide in the normal
1079 * non-TSO case.
1080 */
1081 tcp_skb_pcount_set(skb, 1);
1082 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1083 } else {
1084 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1085 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1086 }
1087}
1088
1089/* When a modification to fackets out becomes necessary, we need to check
1090 * skb is counted to fackets_out or not.
1091 */
1092static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1093 int decr)
1094{
1095 struct tcp_sock *tp = tcp_sk(sk);
1096
1097 if (!tp->sacked_out || tcp_is_reno(tp))
1098 return;
1099
1100 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1101 tp->fackets_out -= decr;
1102}
1103
1104/* Pcount in the middle of the write queue got changed, we need to do various
1105 * tweaks to fix counters
1106 */
1107static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1108{
1109 struct tcp_sock *tp = tcp_sk(sk);
1110
1111 tp->packets_out -= decr;
1112
1113 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1114 tp->sacked_out -= decr;
1115 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1116 tp->retrans_out -= decr;
1117 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1118 tp->lost_out -= decr;
1119
1120 /* Reno case is special. Sigh... */
1121 if (tcp_is_reno(tp) && decr > 0)
1122 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1123
1124 tcp_adjust_fackets_out(sk, skb, decr);
1125
1126 if (tp->lost_skb_hint &&
1127 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1128 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1129 tp->lost_cnt_hint -= decr;
1130
1131 tcp_verify_left_out(tp);
1132}
1133
1134static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1135{
1136 return TCP_SKB_CB(skb)->txstamp_ack ||
1137 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1138}
1139
1140static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1141{
1142 struct skb_shared_info *shinfo = skb_shinfo(skb);
1143
1144 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1145 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1146 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1147 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1148
1149 shinfo->tx_flags &= ~tsflags;
1150 shinfo2->tx_flags |= tsflags;
1151 swap(shinfo->tskey, shinfo2->tskey);
1152 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1153 TCP_SKB_CB(skb)->txstamp_ack = 0;
1154 }
1155}
1156
1157static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1158{
1159 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1160 TCP_SKB_CB(skb)->eor = 0;
1161}
1162
1163/* Function to create two new TCP segments. Shrinks the given segment
1164 * to the specified size and appends a new segment with the rest of the
1165 * packet to the list. This won't be called frequently, I hope.
1166 * Remember, these are still headerless SKBs at this point.
1167 */
1168int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1169 unsigned int mss_now, gfp_t gfp)
1170{
1171 struct tcp_sock *tp = tcp_sk(sk);
1172 struct sk_buff *buff;
1173 int nsize, old_factor;
1174 int nlen;
1175 u8 flags;
1176
1177 if (WARN_ON(len > skb->len))
1178 return -EINVAL;
1179
1180 nsize = skb_headlen(skb) - len;
1181 if (nsize < 0)
1182 nsize = 0;
1183
1184 if (skb_unclone(skb, gfp))
1185 return -ENOMEM;
1186
1187 /* Get a new skb... force flag on. */
1188 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1189 if (!buff)
1190 return -ENOMEM; /* We'll just try again later. */
1191
1192 sk->sk_wmem_queued += buff->truesize;
1193 sk_mem_charge(sk, buff->truesize);
1194 nlen = skb->len - len - nsize;
1195 buff->truesize += nlen;
1196 skb->truesize -= nlen;
1197
1198 /* Correct the sequence numbers. */
1199 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1200 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1201 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1202
1203 /* PSH and FIN should only be set in the second packet. */
1204 flags = TCP_SKB_CB(skb)->tcp_flags;
1205 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1206 TCP_SKB_CB(buff)->tcp_flags = flags;
1207 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1208 tcp_skb_fragment_eor(skb, buff);
1209
1210 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1211 /* Copy and checksum data tail into the new buffer. */
1212 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1213 skb_put(buff, nsize),
1214 nsize, 0);
1215
1216 skb_trim(skb, len);
1217
1218 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1219 } else {
1220 skb->ip_summed = CHECKSUM_PARTIAL;
1221 skb_split(skb, buff, len);
1222 }
1223
1224 buff->ip_summed = skb->ip_summed;
1225
1226 buff->tstamp = skb->tstamp;
1227 tcp_fragment_tstamp(skb, buff);
1228
1229 old_factor = tcp_skb_pcount(skb);
1230
1231 /* Fix up tso_factor for both original and new SKB. */
1232 tcp_set_skb_tso_segs(skb, mss_now);
1233 tcp_set_skb_tso_segs(buff, mss_now);
1234
1235 /* Update delivered info for the new segment */
1236 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1237
1238 /* If this packet has been sent out already, we must
1239 * adjust the various packet counters.
1240 */
1241 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1242 int diff = old_factor - tcp_skb_pcount(skb) -
1243 tcp_skb_pcount(buff);
1244
1245 if (diff)
1246 tcp_adjust_pcount(sk, skb, diff);
1247 }
1248
1249 /* Link BUFF into the send queue. */
1250 __skb_header_release(buff);
1251 tcp_insert_write_queue_after(skb, buff, sk);
1252
1253 return 0;
1254}
1255
1256/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1257 * eventually). The difference is that pulled data not copied, but
1258 * immediately discarded.
1259 */
1260static void __pskb_trim_head(struct sk_buff *skb, int len)
1261{
1262 struct skb_shared_info *shinfo;
1263 int i, k, eat;
1264
1265 eat = min_t(int, len, skb_headlen(skb));
1266 if (eat) {
1267 __skb_pull(skb, eat);
1268 len -= eat;
1269 if (!len)
1270 return;
1271 }
1272 eat = len;
1273 k = 0;
1274 shinfo = skb_shinfo(skb);
1275 for (i = 0; i < shinfo->nr_frags; i++) {
1276 int size = skb_frag_size(&shinfo->frags[i]);
1277
1278 if (size <= eat) {
1279 skb_frag_unref(skb, i);
1280 eat -= size;
1281 } else {
1282 shinfo->frags[k] = shinfo->frags[i];
1283 if (eat) {
1284 shinfo->frags[k].page_offset += eat;
1285 skb_frag_size_sub(&shinfo->frags[k], eat);
1286 eat = 0;
1287 }
1288 k++;
1289 }
1290 }
1291 shinfo->nr_frags = k;
1292
1293 skb_reset_tail_pointer(skb);
1294 skb->data_len -= len;
1295 skb->len = skb->data_len;
1296}
1297
1298/* Remove acked data from a packet in the transmit queue. */
1299int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1300{
1301 if (skb_unclone(skb, GFP_ATOMIC))
1302 return -ENOMEM;
1303
1304 __pskb_trim_head(skb, len);
1305
1306 TCP_SKB_CB(skb)->seq += len;
1307 skb->ip_summed = CHECKSUM_PARTIAL;
1308
1309 skb->truesize -= len;
1310 sk->sk_wmem_queued -= len;
1311 sk_mem_uncharge(sk, len);
1312 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1313
1314 /* Any change of skb->len requires recalculation of tso factor. */
1315 if (tcp_skb_pcount(skb) > 1)
1316 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1317
1318 return 0;
1319}
1320
1321/* Calculate MSS not accounting any TCP options. */
1322static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1323{
1324 const struct tcp_sock *tp = tcp_sk(sk);
1325 const struct inet_connection_sock *icsk = inet_csk(sk);
1326 int mss_now;
1327
1328 /* Calculate base mss without TCP options:
1329 It is MMS_S - sizeof(tcphdr) of rfc1122
1330 */
1331 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1332
1333 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1334 if (icsk->icsk_af_ops->net_frag_header_len) {
1335 const struct dst_entry *dst = __sk_dst_get(sk);
1336
1337 if (dst && dst_allfrag(dst))
1338 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1339 }
1340
1341 /* Clamp it (mss_clamp does not include tcp options) */
1342 if (mss_now > tp->rx_opt.mss_clamp)
1343 mss_now = tp->rx_opt.mss_clamp;
1344
1345 /* Now subtract optional transport overhead */
1346 mss_now -= icsk->icsk_ext_hdr_len;
1347
1348 /* Then reserve room for full set of TCP options and 8 bytes of data */
1349 if (mss_now < 48)
1350 mss_now = 48;
1351 return mss_now;
1352}
1353
1354/* Calculate MSS. Not accounting for SACKs here. */
1355int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1356{
1357 /* Subtract TCP options size, not including SACKs */
1358 return __tcp_mtu_to_mss(sk, pmtu) -
1359 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1360}
1361
1362/* Inverse of above */
1363int tcp_mss_to_mtu(struct sock *sk, int mss)
1364{
1365 const struct tcp_sock *tp = tcp_sk(sk);
1366 const struct inet_connection_sock *icsk = inet_csk(sk);
1367 int mtu;
1368
1369 mtu = mss +
1370 tp->tcp_header_len +
1371 icsk->icsk_ext_hdr_len +
1372 icsk->icsk_af_ops->net_header_len;
1373
1374 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1375 if (icsk->icsk_af_ops->net_frag_header_len) {
1376 const struct dst_entry *dst = __sk_dst_get(sk);
1377
1378 if (dst && dst_allfrag(dst))
1379 mtu += icsk->icsk_af_ops->net_frag_header_len;
1380 }
1381 return mtu;
1382}
1383EXPORT_SYMBOL(tcp_mss_to_mtu);
1384
1385/* MTU probing init per socket */
1386void tcp_mtup_init(struct sock *sk)
1387{
1388 struct tcp_sock *tp = tcp_sk(sk);
1389 struct inet_connection_sock *icsk = inet_csk(sk);
1390 struct net *net = sock_net(sk);
1391
1392 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1393 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1394 icsk->icsk_af_ops->net_header_len;
1395 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1396 icsk->icsk_mtup.probe_size = 0;
1397 if (icsk->icsk_mtup.enabled)
1398 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1399}
1400EXPORT_SYMBOL(tcp_mtup_init);
1401
1402/* This function synchronize snd mss to current pmtu/exthdr set.
1403
1404 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1405 for TCP options, but includes only bare TCP header.
1406
1407 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1408 It is minimum of user_mss and mss received with SYN.
1409 It also does not include TCP options.
1410
1411 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1412
1413 tp->mss_cache is current effective sending mss, including
1414 all tcp options except for SACKs. It is evaluated,
1415 taking into account current pmtu, but never exceeds
1416 tp->rx_opt.mss_clamp.
1417
1418 NOTE1. rfc1122 clearly states that advertised MSS
1419 DOES NOT include either tcp or ip options.
1420
1421 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1422 are READ ONLY outside this function. --ANK (980731)
1423 */
1424unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1425{
1426 struct tcp_sock *tp = tcp_sk(sk);
1427 struct inet_connection_sock *icsk = inet_csk(sk);
1428 int mss_now;
1429
1430 if (icsk->icsk_mtup.search_high > pmtu)
1431 icsk->icsk_mtup.search_high = pmtu;
1432
1433 mss_now = tcp_mtu_to_mss(sk, pmtu);
1434 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1435
1436 /* And store cached results */
1437 icsk->icsk_pmtu_cookie = pmtu;
1438 if (icsk->icsk_mtup.enabled)
1439 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1440 tp->mss_cache = mss_now;
1441
1442 return mss_now;
1443}
1444EXPORT_SYMBOL(tcp_sync_mss);
1445
1446/* Compute the current effective MSS, taking SACKs and IP options,
1447 * and even PMTU discovery events into account.
1448 */
1449unsigned int tcp_current_mss(struct sock *sk)
1450{
1451 const struct tcp_sock *tp = tcp_sk(sk);
1452 const struct dst_entry *dst = __sk_dst_get(sk);
1453 u32 mss_now;
1454 unsigned int header_len;
1455 struct tcp_out_options opts;
1456 struct tcp_md5sig_key *md5;
1457
1458 mss_now = tp->mss_cache;
1459
1460 if (dst) {
1461 u32 mtu = dst_mtu(dst);
1462 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1463 mss_now = tcp_sync_mss(sk, mtu);
1464 }
1465
1466 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1467 sizeof(struct tcphdr);
1468 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1469 * some common options. If this is an odd packet (because we have SACK
1470 * blocks etc) then our calculated header_len will be different, and
1471 * we have to adjust mss_now correspondingly */
1472 if (header_len != tp->tcp_header_len) {
1473 int delta = (int) header_len - tp->tcp_header_len;
1474 mss_now -= delta;
1475 }
1476
1477 return mss_now;
1478}
1479
1480/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1481 * As additional protections, we do not touch cwnd in retransmission phases,
1482 * and if application hit its sndbuf limit recently.
1483 */
1484static void tcp_cwnd_application_limited(struct sock *sk)
1485{
1486 struct tcp_sock *tp = tcp_sk(sk);
1487
1488 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1489 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1490 /* Limited by application or receiver window. */
1491 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1492 u32 win_used = max(tp->snd_cwnd_used, init_win);
1493 if (win_used < tp->snd_cwnd) {
1494 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1495 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1496 }
1497 tp->snd_cwnd_used = 0;
1498 }
1499 tp->snd_cwnd_stamp = tcp_time_stamp;
1500}
1501
1502static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1503{
1504 struct tcp_sock *tp = tcp_sk(sk);
1505
1506 /* Track the maximum number of outstanding packets in each
1507 * window, and remember whether we were cwnd-limited then.
1508 */
1509 if (!before(tp->snd_una, tp->max_packets_seq) ||
1510 tp->packets_out > tp->max_packets_out) {
1511 tp->max_packets_out = tp->packets_out;
1512 tp->max_packets_seq = tp->snd_nxt;
1513 tp->is_cwnd_limited = is_cwnd_limited;
1514 }
1515
1516 if (tcp_is_cwnd_limited(sk)) {
1517 /* Network is feed fully. */
1518 tp->snd_cwnd_used = 0;
1519 tp->snd_cwnd_stamp = tcp_time_stamp;
1520 } else {
1521 /* Network starves. */
1522 if (tp->packets_out > tp->snd_cwnd_used)
1523 tp->snd_cwnd_used = tp->packets_out;
1524
1525 if (sysctl_tcp_slow_start_after_idle &&
1526 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1527 tcp_cwnd_application_limited(sk);
1528
1529 /* The following conditions together indicate the starvation
1530 * is caused by insufficient sender buffer:
1531 * 1) just sent some data (see tcp_write_xmit)
1532 * 2) not cwnd limited (this else condition)
1533 * 3) no more data to send (null tcp_send_head )
1534 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1535 */
1536 if (!tcp_send_head(sk) && sk->sk_socket &&
1537 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1538 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1539 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1540 }
1541}
1542
1543/* Minshall's variant of the Nagle send check. */
1544static bool tcp_minshall_check(const struct tcp_sock *tp)
1545{
1546 return after(tp->snd_sml, tp->snd_una) &&
1547 !after(tp->snd_sml, tp->snd_nxt);
1548}
1549
1550/* Update snd_sml if this skb is under mss
1551 * Note that a TSO packet might end with a sub-mss segment
1552 * The test is really :
1553 * if ((skb->len % mss) != 0)
1554 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1555 * But we can avoid doing the divide again given we already have
1556 * skb_pcount = skb->len / mss_now
1557 */
1558static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1559 const struct sk_buff *skb)
1560{
1561 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1562 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1563}
1564
1565/* Return false, if packet can be sent now without violation Nagle's rules:
1566 * 1. It is full sized. (provided by caller in %partial bool)
1567 * 2. Or it contains FIN. (already checked by caller)
1568 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1569 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1570 * With Minshall's modification: all sent small packets are ACKed.
1571 */
1572static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1573 int nonagle)
1574{
1575 return partial &&
1576 ((nonagle & TCP_NAGLE_CORK) ||
1577 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1578}
1579
1580/* Return how many segs we'd like on a TSO packet,
1581 * to send one TSO packet per ms
1582 */
1583u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1584 int min_tso_segs)
1585{
1586 u32 bytes, segs;
1587
1588 bytes = min(sk->sk_pacing_rate >> 10,
1589 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1590
1591 /* Goal is to send at least one packet per ms,
1592 * not one big TSO packet every 100 ms.
1593 * This preserves ACK clocking and is consistent
1594 * with tcp_tso_should_defer() heuristic.
1595 */
1596 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1597
1598 return min_t(u32, segs, sk->sk_gso_max_segs);
1599}
1600EXPORT_SYMBOL(tcp_tso_autosize);
1601
1602/* Return the number of segments we want in the skb we are transmitting.
1603 * See if congestion control module wants to decide; otherwise, autosize.
1604 */
1605static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1606{
1607 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1608 u32 tso_segs = ca_ops->tso_segs_goal ? ca_ops->tso_segs_goal(sk) : 0;
1609
1610 return tso_segs ? :
1611 tcp_tso_autosize(sk, mss_now, sysctl_tcp_min_tso_segs);
1612}
1613
1614/* Returns the portion of skb which can be sent right away */
1615static unsigned int tcp_mss_split_point(const struct sock *sk,
1616 const struct sk_buff *skb,
1617 unsigned int mss_now,
1618 unsigned int max_segs,
1619 int nonagle)
1620{
1621 const struct tcp_sock *tp = tcp_sk(sk);
1622 u32 partial, needed, window, max_len;
1623
1624 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1625 max_len = mss_now * max_segs;
1626
1627 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1628 return max_len;
1629
1630 needed = min(skb->len, window);
1631
1632 if (max_len <= needed)
1633 return max_len;
1634
1635 partial = needed % mss_now;
1636 /* If last segment is not a full MSS, check if Nagle rules allow us
1637 * to include this last segment in this skb.
1638 * Otherwise, we'll split the skb at last MSS boundary
1639 */
1640 if (tcp_nagle_check(partial != 0, tp, nonagle))
1641 return needed - partial;
1642
1643 return needed;
1644}
1645
1646/* Can at least one segment of SKB be sent right now, according to the
1647 * congestion window rules? If so, return how many segments are allowed.
1648 */
1649static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1650 const struct sk_buff *skb)
1651{
1652 u32 in_flight, cwnd, halfcwnd;
1653
1654 /* Don't be strict about the congestion window for the final FIN. */
1655 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1656 tcp_skb_pcount(skb) == 1)
1657 return 1;
1658
1659 in_flight = tcp_packets_in_flight(tp);
1660 cwnd = tp->snd_cwnd;
1661 if (in_flight >= cwnd)
1662 return 0;
1663
1664 /* For better scheduling, ensure we have at least
1665 * 2 GSO packets in flight.
1666 */
1667 halfcwnd = max(cwnd >> 1, 1U);
1668 return min(halfcwnd, cwnd - in_flight);
1669}
1670
1671/* Initialize TSO state of a skb.
1672 * This must be invoked the first time we consider transmitting
1673 * SKB onto the wire.
1674 */
1675static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1676{
1677 int tso_segs = tcp_skb_pcount(skb);
1678
1679 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1680 tcp_set_skb_tso_segs(skb, mss_now);
1681 tso_segs = tcp_skb_pcount(skb);
1682 }
1683 return tso_segs;
1684}
1685
1686
1687/* Return true if the Nagle test allows this packet to be
1688 * sent now.
1689 */
1690static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1691 unsigned int cur_mss, int nonagle)
1692{
1693 /* Nagle rule does not apply to frames, which sit in the middle of the
1694 * write_queue (they have no chances to get new data).
1695 *
1696 * This is implemented in the callers, where they modify the 'nonagle'
1697 * argument based upon the location of SKB in the send queue.
1698 */
1699 if (nonagle & TCP_NAGLE_PUSH)
1700 return true;
1701
1702 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1703 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1704 return true;
1705
1706 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1707 return true;
1708
1709 return false;
1710}
1711
1712/* Does at least the first segment of SKB fit into the send window? */
1713static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1714 const struct sk_buff *skb,
1715 unsigned int cur_mss)
1716{
1717 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1718
1719 if (skb->len > cur_mss)
1720 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1721
1722 return !after(end_seq, tcp_wnd_end(tp));
1723}
1724
1725/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1726 * should be put on the wire right now. If so, it returns the number of
1727 * packets allowed by the congestion window.
1728 */
1729static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1730 unsigned int cur_mss, int nonagle)
1731{
1732 const struct tcp_sock *tp = tcp_sk(sk);
1733 unsigned int cwnd_quota;
1734
1735 tcp_init_tso_segs(skb, cur_mss);
1736
1737 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1738 return 0;
1739
1740 cwnd_quota = tcp_cwnd_test(tp, skb);
1741 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1742 cwnd_quota = 0;
1743
1744 return cwnd_quota;
1745}
1746
1747/* Test if sending is allowed right now. */
1748bool tcp_may_send_now(struct sock *sk)
1749{
1750 const struct tcp_sock *tp = tcp_sk(sk);
1751 struct sk_buff *skb = tcp_send_head(sk);
1752
1753 return skb &&
1754 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1755 (tcp_skb_is_last(sk, skb) ?
1756 tp->nonagle : TCP_NAGLE_PUSH));
1757}
1758
1759/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1760 * which is put after SKB on the list. It is very much like
1761 * tcp_fragment() except that it may make several kinds of assumptions
1762 * in order to speed up the splitting operation. In particular, we
1763 * know that all the data is in scatter-gather pages, and that the
1764 * packet has never been sent out before (and thus is not cloned).
1765 */
1766static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1767 unsigned int mss_now, gfp_t gfp)
1768{
1769 struct sk_buff *buff;
1770 int nlen = skb->len - len;
1771 u8 flags;
1772
1773 /* All of a TSO frame must be composed of paged data. */
1774 if (skb->len != skb->data_len)
1775 return tcp_fragment(sk, skb, len, mss_now, gfp);
1776
1777 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1778 if (unlikely(!buff))
1779 return -ENOMEM;
1780
1781 sk->sk_wmem_queued += buff->truesize;
1782 sk_mem_charge(sk, buff->truesize);
1783 buff->truesize += nlen;
1784 skb->truesize -= nlen;
1785
1786 /* Correct the sequence numbers. */
1787 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1788 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1789 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1790
1791 /* PSH and FIN should only be set in the second packet. */
1792 flags = TCP_SKB_CB(skb)->tcp_flags;
1793 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1794 TCP_SKB_CB(buff)->tcp_flags = flags;
1795
1796 /* This packet was never sent out yet, so no SACK bits. */
1797 TCP_SKB_CB(buff)->sacked = 0;
1798
1799 tcp_skb_fragment_eor(skb, buff);
1800
1801 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1802 skb_split(skb, buff, len);
1803 tcp_fragment_tstamp(skb, buff);
1804
1805 /* Fix up tso_factor for both original and new SKB. */
1806 tcp_set_skb_tso_segs(skb, mss_now);
1807 tcp_set_skb_tso_segs(buff, mss_now);
1808
1809 /* Link BUFF into the send queue. */
1810 __skb_header_release(buff);
1811 tcp_insert_write_queue_after(skb, buff, sk);
1812
1813 return 0;
1814}
1815
1816/* Try to defer sending, if possible, in order to minimize the amount
1817 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1818 *
1819 * This algorithm is from John Heffner.
1820 */
1821static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1822 bool *is_cwnd_limited, u32 max_segs)
1823{
1824 const struct inet_connection_sock *icsk = inet_csk(sk);
1825 u32 age, send_win, cong_win, limit, in_flight;
1826 struct tcp_sock *tp = tcp_sk(sk);
1827 struct skb_mstamp now;
1828 struct sk_buff *head;
1829 int win_divisor;
1830
1831 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1832 goto send_now;
1833
1834 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1835 goto send_now;
1836
1837 /* Avoid bursty behavior by allowing defer
1838 * only if the last write was recent.
1839 */
1840 if ((s32)(tcp_time_stamp - tp->lsndtime) > 0)
1841 goto send_now;
1842
1843 in_flight = tcp_packets_in_flight(tp);
1844
1845 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1846
1847 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1848
1849 /* From in_flight test above, we know that cwnd > in_flight. */
1850 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1851
1852 limit = min(send_win, cong_win);
1853
1854 /* If a full-sized TSO skb can be sent, do it. */
1855 if (limit >= max_segs * tp->mss_cache)
1856 goto send_now;
1857
1858 /* Middle in queue won't get any more data, full sendable already? */
1859 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1860 goto send_now;
1861
1862 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1863 if (win_divisor) {
1864 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1865
1866 /* If at least some fraction of a window is available,
1867 * just use it.
1868 */
1869 chunk /= win_divisor;
1870 if (limit >= chunk)
1871 goto send_now;
1872 } else {
1873 /* Different approach, try not to defer past a single
1874 * ACK. Receiver should ACK every other full sized
1875 * frame, so if we have space for more than 3 frames
1876 * then send now.
1877 */
1878 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1879 goto send_now;
1880 }
1881
1882 head = tcp_write_queue_head(sk);
1883 skb_mstamp_get(&now);
1884 age = skb_mstamp_us_delta(&now, &head->skb_mstamp);
1885 /* If next ACK is likely to come too late (half srtt), do not defer */
1886 if (age < (tp->srtt_us >> 4))
1887 goto send_now;
1888
1889 /* Ok, it looks like it is advisable to defer. */
1890
1891 if (cong_win < send_win && cong_win <= skb->len)
1892 *is_cwnd_limited = true;
1893
1894 return true;
1895
1896send_now:
1897 return false;
1898}
1899
1900static inline void tcp_mtu_check_reprobe(struct sock *sk)
1901{
1902 struct inet_connection_sock *icsk = inet_csk(sk);
1903 struct tcp_sock *tp = tcp_sk(sk);
1904 struct net *net = sock_net(sk);
1905 u32 interval;
1906 s32 delta;
1907
1908 interval = net->ipv4.sysctl_tcp_probe_interval;
1909 delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp;
1910 if (unlikely(delta >= interval * HZ)) {
1911 int mss = tcp_current_mss(sk);
1912
1913 /* Update current search range */
1914 icsk->icsk_mtup.probe_size = 0;
1915 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1916 sizeof(struct tcphdr) +
1917 icsk->icsk_af_ops->net_header_len;
1918 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
1919
1920 /* Update probe time stamp */
1921 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1922 }
1923}
1924
1925/* Create a new MTU probe if we are ready.
1926 * MTU probe is regularly attempting to increase the path MTU by
1927 * deliberately sending larger packets. This discovers routing
1928 * changes resulting in larger path MTUs.
1929 *
1930 * Returns 0 if we should wait to probe (no cwnd available),
1931 * 1 if a probe was sent,
1932 * -1 otherwise
1933 */
1934static int tcp_mtu_probe(struct sock *sk)
1935{
1936 struct inet_connection_sock *icsk = inet_csk(sk);
1937 struct tcp_sock *tp = tcp_sk(sk);
1938 struct sk_buff *skb, *nskb, *next;
1939 struct net *net = sock_net(sk);
1940 int probe_size;
1941 int size_needed;
1942 int copy, len;
1943 int mss_now;
1944 int interval;
1945
1946 /* Not currently probing/verifying,
1947 * not in recovery,
1948 * have enough cwnd, and
1949 * not SACKing (the variable headers throw things off)
1950 */
1951 if (likely(!icsk->icsk_mtup.enabled ||
1952 icsk->icsk_mtup.probe_size ||
1953 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1954 tp->snd_cwnd < 11 ||
1955 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
1956 return -1;
1957
1958 /* Use binary search for probe_size between tcp_mss_base,
1959 * and current mss_clamp. if (search_high - search_low)
1960 * smaller than a threshold, backoff from probing.
1961 */
1962 mss_now = tcp_current_mss(sk);
1963 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
1964 icsk->icsk_mtup.search_low) >> 1);
1965 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1966 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
1967 /* When misfortune happens, we are reprobing actively,
1968 * and then reprobe timer has expired. We stick with current
1969 * probing process by not resetting search range to its orignal.
1970 */
1971 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
1972 interval < net->ipv4.sysctl_tcp_probe_threshold) {
1973 /* Check whether enough time has elaplased for
1974 * another round of probing.
1975 */
1976 tcp_mtu_check_reprobe(sk);
1977 return -1;
1978 }
1979
1980 /* Have enough data in the send queue to probe? */
1981 if (tp->write_seq - tp->snd_nxt < size_needed)
1982 return -1;
1983
1984 if (tp->snd_wnd < size_needed)
1985 return -1;
1986 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1987 return 0;
1988
1989 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1990 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1991 if (!tcp_packets_in_flight(tp))
1992 return -1;
1993 else
1994 return 0;
1995 }
1996
1997 /* We're allowed to probe. Build it now. */
1998 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
1999 if (!nskb)
2000 return -1;
2001 sk->sk_wmem_queued += nskb->truesize;
2002 sk_mem_charge(sk, nskb->truesize);
2003
2004 skb = tcp_send_head(sk);
2005
2006 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2007 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2008 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2009 TCP_SKB_CB(nskb)->sacked = 0;
2010 nskb->csum = 0;
2011 nskb->ip_summed = skb->ip_summed;
2012
2013 tcp_insert_write_queue_before(nskb, skb, sk);
2014
2015 len = 0;
2016 tcp_for_write_queue_from_safe(skb, next, sk) {
2017 copy = min_t(int, skb->len, probe_size - len);
2018 if (nskb->ip_summed) {
2019 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2020 } else {
2021 __wsum csum = skb_copy_and_csum_bits(skb, 0,
2022 skb_put(nskb, copy),
2023 copy, 0);
2024 nskb->csum = csum_block_add(nskb->csum, csum, len);
2025 }
2026
2027 if (skb->len <= copy) {
2028 /* We've eaten all the data from this skb.
2029 * Throw it away. */
2030 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2031 tcp_unlink_write_queue(skb, sk);
2032 sk_wmem_free_skb(sk, skb);
2033 } else {
2034 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2035 ~(TCPHDR_FIN|TCPHDR_PSH);
2036 if (!skb_shinfo(skb)->nr_frags) {
2037 skb_pull(skb, copy);
2038 if (skb->ip_summed != CHECKSUM_PARTIAL)
2039 skb->csum = csum_partial(skb->data,
2040 skb->len, 0);
2041 } else {
2042 __pskb_trim_head(skb, copy);
2043 tcp_set_skb_tso_segs(skb, mss_now);
2044 }
2045 TCP_SKB_CB(skb)->seq += copy;
2046 }
2047
2048 len += copy;
2049
2050 if (len >= probe_size)
2051 break;
2052 }
2053 tcp_init_tso_segs(nskb, nskb->len);
2054
2055 /* We're ready to send. If this fails, the probe will
2056 * be resegmented into mss-sized pieces by tcp_write_xmit().
2057 */
2058 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2059 /* Decrement cwnd here because we are sending
2060 * effectively two packets. */
2061 tp->snd_cwnd--;
2062 tcp_event_new_data_sent(sk, nskb);
2063
2064 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2065 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2066 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2067
2068 return 1;
2069 }
2070
2071 return -1;
2072}
2073
2074/* TCP Small Queues :
2075 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2076 * (These limits are doubled for retransmits)
2077 * This allows for :
2078 * - better RTT estimation and ACK scheduling
2079 * - faster recovery
2080 * - high rates
2081 * Alas, some drivers / subsystems require a fair amount
2082 * of queued bytes to ensure line rate.
2083 * One example is wifi aggregation (802.11 AMPDU)
2084 */
2085static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2086 unsigned int factor)
2087{
2088 unsigned int limit;
2089
2090 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2091 limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2092 limit <<= factor;
2093
2094 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
2095 /* Always send the 1st or 2nd skb in write queue.
2096 * No need to wait for TX completion to call us back,
2097 * after softirq/tasklet schedule.
2098 * This helps when TX completions are delayed too much.
2099 */
2100 if (skb == sk->sk_write_queue.next ||
2101 skb->prev == sk->sk_write_queue.next)
2102 return false;
2103
2104 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2105 /* It is possible TX completion already happened
2106 * before we set TSQ_THROTTLED, so we must
2107 * test again the condition.
2108 */
2109 smp_mb__after_atomic();
2110 if (atomic_read(&sk->sk_wmem_alloc) > limit)
2111 return true;
2112 }
2113 return false;
2114}
2115
2116static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2117{
2118 const u32 now = tcp_time_stamp;
2119
2120 if (tp->chrono_type > TCP_CHRONO_UNSPEC)
2121 tp->chrono_stat[tp->chrono_type - 1] += now - tp->chrono_start;
2122 tp->chrono_start = now;
2123 tp->chrono_type = new;
2124}
2125
2126void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2127{
2128 struct tcp_sock *tp = tcp_sk(sk);
2129
2130 /* If there are multiple conditions worthy of tracking in a
2131 * chronograph then the highest priority enum takes precedence
2132 * over the other conditions. So that if something "more interesting"
2133 * starts happening, stop the previous chrono and start a new one.
2134 */
2135 if (type > tp->chrono_type)
2136 tcp_chrono_set(tp, type);
2137}
2138
2139void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2140{
2141 struct tcp_sock *tp = tcp_sk(sk);
2142
2143
2144 /* There are multiple conditions worthy of tracking in a
2145 * chronograph, so that the highest priority enum takes
2146 * precedence over the other conditions (see tcp_chrono_start).
2147 * If a condition stops, we only stop chrono tracking if
2148 * it's the "most interesting" or current chrono we are
2149 * tracking and starts busy chrono if we have pending data.
2150 */
2151 if (tcp_write_queue_empty(sk))
2152 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2153 else if (type == tp->chrono_type)
2154 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2155}
2156
2157/* This routine writes packets to the network. It advances the
2158 * send_head. This happens as incoming acks open up the remote
2159 * window for us.
2160 *
2161 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2162 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2163 * account rare use of URG, this is not a big flaw.
2164 *
2165 * Send at most one packet when push_one > 0. Temporarily ignore
2166 * cwnd limit to force at most one packet out when push_one == 2.
2167
2168 * Returns true, if no segments are in flight and we have queued segments,
2169 * but cannot send anything now because of SWS or another problem.
2170 */
2171static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2172 int push_one, gfp_t gfp)
2173{
2174 struct tcp_sock *tp = tcp_sk(sk);
2175 struct sk_buff *skb;
2176 unsigned int tso_segs, sent_pkts;
2177 int cwnd_quota;
2178 int result;
2179 bool is_cwnd_limited = false, is_rwnd_limited = false;
2180 u32 max_segs;
2181
2182 sent_pkts = 0;
2183
2184 if (!push_one) {
2185 /* Do MTU probing. */
2186 result = tcp_mtu_probe(sk);
2187 if (!result) {
2188 return false;
2189 } else if (result > 0) {
2190 sent_pkts = 1;
2191 }
2192 }
2193
2194 max_segs = tcp_tso_segs(sk, mss_now);
2195 while ((skb = tcp_send_head(sk))) {
2196 unsigned int limit;
2197
2198 tso_segs = tcp_init_tso_segs(skb, mss_now);
2199 BUG_ON(!tso_segs);
2200
2201 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2202 /* "skb_mstamp" is used as a start point for the retransmit timer */
2203 skb_mstamp_get(&skb->skb_mstamp);
2204 goto repair; /* Skip network transmission */
2205 }
2206
2207 cwnd_quota = tcp_cwnd_test(tp, skb);
2208 if (!cwnd_quota) {
2209 if (push_one == 2)
2210 /* Force out a loss probe pkt. */
2211 cwnd_quota = 1;
2212 else
2213 break;
2214 }
2215
2216 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2217 is_rwnd_limited = true;
2218 break;
2219 }
2220
2221 if (tso_segs == 1) {
2222 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2223 (tcp_skb_is_last(sk, skb) ?
2224 nonagle : TCP_NAGLE_PUSH))))
2225 break;
2226 } else {
2227 if (!push_one &&
2228 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2229 max_segs))
2230 break;
2231 }
2232
2233 limit = mss_now;
2234 if (tso_segs > 1 && !tcp_urg_mode(tp))
2235 limit = tcp_mss_split_point(sk, skb, mss_now,
2236 min_t(unsigned int,
2237 cwnd_quota,
2238 max_segs),
2239 nonagle);
2240
2241 if (skb->len > limit &&
2242 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2243 break;
2244
2245 if (test_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
2246 clear_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags);
2247 if (tcp_small_queue_check(sk, skb, 0))
2248 break;
2249
2250 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2251 break;
2252
2253repair:
2254 /* Advance the send_head. This one is sent out.
2255 * This call will increment packets_out.
2256 */
2257 tcp_event_new_data_sent(sk, skb);
2258
2259 tcp_minshall_update(tp, mss_now, skb);
2260 sent_pkts += tcp_skb_pcount(skb);
2261
2262 if (push_one)
2263 break;
2264 }
2265
2266 if (is_rwnd_limited)
2267 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2268 else
2269 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2270
2271 if (likely(sent_pkts)) {
2272 if (tcp_in_cwnd_reduction(sk))
2273 tp->prr_out += sent_pkts;
2274
2275 /* Send one loss probe per tail loss episode. */
2276 if (push_one != 2)
2277 tcp_schedule_loss_probe(sk);
2278 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2279 tcp_cwnd_validate(sk, is_cwnd_limited);
2280 return false;
2281 }
2282 return !tp->packets_out && tcp_send_head(sk);
2283}
2284
2285bool tcp_schedule_loss_probe(struct sock *sk)
2286{
2287 struct inet_connection_sock *icsk = inet_csk(sk);
2288 struct tcp_sock *tp = tcp_sk(sk);
2289 u32 timeout, tlp_time_stamp, rto_time_stamp;
2290 u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
2291
2292 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2293 return false;
2294 /* No consecutive loss probes. */
2295 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2296 tcp_rearm_rto(sk);
2297 return false;
2298 }
2299 /* Don't do any loss probe on a Fast Open connection before 3WHS
2300 * finishes.
2301 */
2302 if (tp->fastopen_rsk)
2303 return false;
2304
2305 /* TLP is only scheduled when next timer event is RTO. */
2306 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2307 return false;
2308
2309 /* Schedule a loss probe in 2*RTT for SACK capable connections
2310 * in Open state, that are either limited by cwnd or application.
2311 */
2312 if (sysctl_tcp_early_retrans < 3 || !tp->packets_out ||
2313 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2314 return false;
2315
2316 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2317 tcp_send_head(sk))
2318 return false;
2319
2320 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2321 * for delayed ack when there's one outstanding packet. If no RTT
2322 * sample is available then probe after TCP_TIMEOUT_INIT.
2323 */
2324 timeout = rtt << 1 ? : TCP_TIMEOUT_INIT;
2325 if (tp->packets_out == 1)
2326 timeout = max_t(u32, timeout,
2327 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2328 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2329
2330 /* If RTO is shorter, just schedule TLP in its place. */
2331 tlp_time_stamp = tcp_time_stamp + timeout;
2332 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2333 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2334 s32 delta = rto_time_stamp - tcp_time_stamp;
2335 if (delta > 0)
2336 timeout = delta;
2337 }
2338
2339 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2340 TCP_RTO_MAX);
2341 return true;
2342}
2343
2344/* Thanks to skb fast clones, we can detect if a prior transmit of
2345 * a packet is still in a qdisc or driver queue.
2346 * In this case, there is very little point doing a retransmit !
2347 */
2348static bool skb_still_in_host_queue(const struct sock *sk,
2349 const struct sk_buff *skb)
2350{
2351 if (unlikely(skb_fclone_busy(sk, skb))) {
2352 NET_INC_STATS(sock_net(sk),
2353 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2354 return true;
2355 }
2356 return false;
2357}
2358
2359/* When probe timeout (PTO) fires, try send a new segment if possible, else
2360 * retransmit the last segment.
2361 */
2362void tcp_send_loss_probe(struct sock *sk)
2363{
2364 struct tcp_sock *tp = tcp_sk(sk);
2365 struct sk_buff *skb;
2366 int pcount;
2367 int mss = tcp_current_mss(sk);
2368
2369 skb = tcp_send_head(sk);
2370 if (skb) {
2371 if (tcp_snd_wnd_test(tp, skb, mss)) {
2372 pcount = tp->packets_out;
2373 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2374 if (tp->packets_out > pcount)
2375 goto probe_sent;
2376 goto rearm_timer;
2377 }
2378 skb = tcp_write_queue_prev(sk, skb);
2379 } else {
2380 skb = tcp_write_queue_tail(sk);
2381 }
2382
2383 /* At most one outstanding TLP retransmission. */
2384 if (tp->tlp_high_seq)
2385 goto rearm_timer;
2386
2387 /* Retransmit last segment. */
2388 if (WARN_ON(!skb))
2389 goto rearm_timer;
2390
2391 if (skb_still_in_host_queue(sk, skb))
2392 goto rearm_timer;
2393
2394 pcount = tcp_skb_pcount(skb);
2395 if (WARN_ON(!pcount))
2396 goto rearm_timer;
2397
2398 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2399 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2400 GFP_ATOMIC)))
2401 goto rearm_timer;
2402 skb = tcp_write_queue_next(sk, skb);
2403 }
2404
2405 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2406 goto rearm_timer;
2407
2408 if (__tcp_retransmit_skb(sk, skb, 1))
2409 goto rearm_timer;
2410
2411 /* Record snd_nxt for loss detection. */
2412 tp->tlp_high_seq = tp->snd_nxt;
2413
2414probe_sent:
2415 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2416 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2417 inet_csk(sk)->icsk_pending = 0;
2418rearm_timer:
2419 tcp_rearm_rto(sk);
2420}
2421
2422/* Push out any pending frames which were held back due to
2423 * TCP_CORK or attempt at coalescing tiny packets.
2424 * The socket must be locked by the caller.
2425 */
2426void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2427 int nonagle)
2428{
2429 /* If we are closed, the bytes will have to remain here.
2430 * In time closedown will finish, we empty the write queue and
2431 * all will be happy.
2432 */
2433 if (unlikely(sk->sk_state == TCP_CLOSE))
2434 return;
2435
2436 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2437 sk_gfp_mask(sk, GFP_ATOMIC)))
2438 tcp_check_probe_timer(sk);
2439}
2440
2441/* Send _single_ skb sitting at the send head. This function requires
2442 * true push pending frames to setup probe timer etc.
2443 */
2444void tcp_push_one(struct sock *sk, unsigned int mss_now)
2445{
2446 struct sk_buff *skb = tcp_send_head(sk);
2447
2448 BUG_ON(!skb || skb->len < mss_now);
2449
2450 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2451}
2452
2453/* This function returns the amount that we can raise the
2454 * usable window based on the following constraints
2455 *
2456 * 1. The window can never be shrunk once it is offered (RFC 793)
2457 * 2. We limit memory per socket
2458 *
2459 * RFC 1122:
2460 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2461 * RECV.NEXT + RCV.WIN fixed until:
2462 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2463 *
2464 * i.e. don't raise the right edge of the window until you can raise
2465 * it at least MSS bytes.
2466 *
2467 * Unfortunately, the recommended algorithm breaks header prediction,
2468 * since header prediction assumes th->window stays fixed.
2469 *
2470 * Strictly speaking, keeping th->window fixed violates the receiver
2471 * side SWS prevention criteria. The problem is that under this rule
2472 * a stream of single byte packets will cause the right side of the
2473 * window to always advance by a single byte.
2474 *
2475 * Of course, if the sender implements sender side SWS prevention
2476 * then this will not be a problem.
2477 *
2478 * BSD seems to make the following compromise:
2479 *
2480 * If the free space is less than the 1/4 of the maximum
2481 * space available and the free space is less than 1/2 mss,
2482 * then set the window to 0.
2483 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2484 * Otherwise, just prevent the window from shrinking
2485 * and from being larger than the largest representable value.
2486 *
2487 * This prevents incremental opening of the window in the regime
2488 * where TCP is limited by the speed of the reader side taking
2489 * data out of the TCP receive queue. It does nothing about
2490 * those cases where the window is constrained on the sender side
2491 * because the pipeline is full.
2492 *
2493 * BSD also seems to "accidentally" limit itself to windows that are a
2494 * multiple of MSS, at least until the free space gets quite small.
2495 * This would appear to be a side effect of the mbuf implementation.
2496 * Combining these two algorithms results in the observed behavior
2497 * of having a fixed window size at almost all times.
2498 *
2499 * Below we obtain similar behavior by forcing the offered window to
2500 * a multiple of the mss when it is feasible to do so.
2501 *
2502 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2503 * Regular options like TIMESTAMP are taken into account.
2504 */
2505u32 __tcp_select_window(struct sock *sk)
2506{
2507 struct inet_connection_sock *icsk = inet_csk(sk);
2508 struct tcp_sock *tp = tcp_sk(sk);
2509 /* MSS for the peer's data. Previous versions used mss_clamp
2510 * here. I don't know if the value based on our guesses
2511 * of peer's MSS is better for the performance. It's more correct
2512 * but may be worse for the performance because of rcv_mss
2513 * fluctuations. --SAW 1998/11/1
2514 */
2515 int mss = icsk->icsk_ack.rcv_mss;
2516 int free_space = tcp_space(sk);
2517 int allowed_space = tcp_full_space(sk);
2518 int full_space = min_t(int, tp->window_clamp, allowed_space);
2519 int window;
2520
2521 if (unlikely(mss > full_space)) {
2522 mss = full_space;
2523 if (mss <= 0)
2524 return 0;
2525 }
2526 if (free_space < (full_space >> 1)) {
2527 icsk->icsk_ack.quick = 0;
2528
2529 if (tcp_under_memory_pressure(sk))
2530 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2531 4U * tp->advmss);
2532
2533 /* free_space might become our new window, make sure we don't
2534 * increase it due to wscale.
2535 */
2536 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2537
2538 /* if free space is less than mss estimate, or is below 1/16th
2539 * of the maximum allowed, try to move to zero-window, else
2540 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2541 * new incoming data is dropped due to memory limits.
2542 * With large window, mss test triggers way too late in order
2543 * to announce zero window in time before rmem limit kicks in.
2544 */
2545 if (free_space < (allowed_space >> 4) || free_space < mss)
2546 return 0;
2547 }
2548
2549 if (free_space > tp->rcv_ssthresh)
2550 free_space = tp->rcv_ssthresh;
2551
2552 /* Don't do rounding if we are using window scaling, since the
2553 * scaled window will not line up with the MSS boundary anyway.
2554 */
2555 window = tp->rcv_wnd;
2556 if (tp->rx_opt.rcv_wscale) {
2557 window = free_space;
2558
2559 /* Advertise enough space so that it won't get scaled away.
2560 * Import case: prevent zero window announcement if
2561 * 1<<rcv_wscale > mss.
2562 */
2563 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2564 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2565 << tp->rx_opt.rcv_wscale);
2566 } else {
2567 /* Get the largest window that is a nice multiple of mss.
2568 * Window clamp already applied above.
2569 * If our current window offering is within 1 mss of the
2570 * free space we just keep it. This prevents the divide
2571 * and multiply from happening most of the time.
2572 * We also don't do any window rounding when the free space
2573 * is too small.
2574 */
2575 if (window <= free_space - mss || window > free_space)
2576 window = (free_space / mss) * mss;
2577 else if (mss == full_space &&
2578 free_space > window + (full_space >> 1))
2579 window = free_space;
2580 }
2581
2582 return window;
2583}
2584
2585void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2586 const struct sk_buff *next_skb)
2587{
2588 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2589 const struct skb_shared_info *next_shinfo =
2590 skb_shinfo(next_skb);
2591 struct skb_shared_info *shinfo = skb_shinfo(skb);
2592
2593 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2594 shinfo->tskey = next_shinfo->tskey;
2595 TCP_SKB_CB(skb)->txstamp_ack |=
2596 TCP_SKB_CB(next_skb)->txstamp_ack;
2597 }
2598}
2599
2600/* Collapses two adjacent SKB's during retransmission. */
2601static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2602{
2603 struct tcp_sock *tp = tcp_sk(sk);
2604 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2605 int skb_size, next_skb_size;
2606
2607 skb_size = skb->len;
2608 next_skb_size = next_skb->len;
2609
2610 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2611
2612 if (next_skb_size) {
2613 if (next_skb_size <= skb_availroom(skb))
2614 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2615 next_skb_size);
2616 else if (!skb_shift(skb, next_skb, next_skb_size))
2617 return false;
2618 }
2619 tcp_highest_sack_combine(sk, next_skb, skb);
2620
2621 tcp_unlink_write_queue(next_skb, sk);
2622
2623 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2624 skb->ip_summed = CHECKSUM_PARTIAL;
2625
2626 if (skb->ip_summed != CHECKSUM_PARTIAL)
2627 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2628
2629 /* Update sequence range on original skb. */
2630 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2631
2632 /* Merge over control information. This moves PSH/FIN etc. over */
2633 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2634
2635 /* All done, get rid of second SKB and account for it so
2636 * packet counting does not break.
2637 */
2638 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2639 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2640
2641 /* changed transmit queue under us so clear hints */
2642 tcp_clear_retrans_hints_partial(tp);
2643 if (next_skb == tp->retransmit_skb_hint)
2644 tp->retransmit_skb_hint = skb;
2645
2646 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2647
2648 tcp_skb_collapse_tstamp(skb, next_skb);
2649
2650 sk_wmem_free_skb(sk, next_skb);
2651 return true;
2652}
2653
2654/* Check if coalescing SKBs is legal. */
2655static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2656{
2657 if (tcp_skb_pcount(skb) > 1)
2658 return false;
2659 if (skb_cloned(skb))
2660 return false;
2661 if (skb == tcp_send_head(sk))
2662 return false;
2663 /* Some heuristics for collapsing over SACK'd could be invented */
2664 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2665 return false;
2666
2667 return true;
2668}
2669
2670/* Collapse packets in the retransmit queue to make to create
2671 * less packets on the wire. This is only done on retransmission.
2672 */
2673static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2674 int space)
2675{
2676 struct tcp_sock *tp = tcp_sk(sk);
2677 struct sk_buff *skb = to, *tmp;
2678 bool first = true;
2679
2680 if (!sysctl_tcp_retrans_collapse)
2681 return;
2682 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2683 return;
2684
2685 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2686 if (!tcp_can_collapse(sk, skb))
2687 break;
2688
2689 if (!tcp_skb_can_collapse_to(to))
2690 break;
2691
2692 space -= skb->len;
2693
2694 if (first) {
2695 first = false;
2696 continue;
2697 }
2698
2699 if (space < 0)
2700 break;
2701
2702 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2703 break;
2704
2705 if (!tcp_collapse_retrans(sk, to))
2706 break;
2707 }
2708}
2709
2710/* This retransmits one SKB. Policy decisions and retransmit queue
2711 * state updates are done by the caller. Returns non-zero if an
2712 * error occurred which prevented the send.
2713 */
2714int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2715{
2716 struct inet_connection_sock *icsk = inet_csk(sk);
2717 struct tcp_sock *tp = tcp_sk(sk);
2718 unsigned int cur_mss;
2719 int diff, len, err;
2720
2721
2722 /* Inconclusive MTU probe */
2723 if (icsk->icsk_mtup.probe_size)
2724 icsk->icsk_mtup.probe_size = 0;
2725
2726 /* Do not sent more than we queued. 1/4 is reserved for possible
2727 * copying overhead: fragmentation, tunneling, mangling etc.
2728 */
2729 if (atomic_read(&sk->sk_wmem_alloc) >
2730 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2731 sk->sk_sndbuf))
2732 return -EAGAIN;
2733
2734 if (skb_still_in_host_queue(sk, skb))
2735 return -EBUSY;
2736
2737 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2738 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2739 BUG();
2740 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2741 return -ENOMEM;
2742 }
2743
2744 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2745 return -EHOSTUNREACH; /* Routing failure or similar. */
2746
2747 cur_mss = tcp_current_mss(sk);
2748
2749 /* If receiver has shrunk his window, and skb is out of
2750 * new window, do not retransmit it. The exception is the
2751 * case, when window is shrunk to zero. In this case
2752 * our retransmit serves as a zero window probe.
2753 */
2754 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2755 TCP_SKB_CB(skb)->seq != tp->snd_una)
2756 return -EAGAIN;
2757
2758 len = cur_mss * segs;
2759 if (skb->len > len) {
2760 if (tcp_fragment(sk, skb, len, cur_mss, GFP_ATOMIC))
2761 return -ENOMEM; /* We'll try again later. */
2762 } else {
2763 if (skb_unclone(skb, GFP_ATOMIC))
2764 return -ENOMEM;
2765
2766 diff = tcp_skb_pcount(skb);
2767 tcp_set_skb_tso_segs(skb, cur_mss);
2768 diff -= tcp_skb_pcount(skb);
2769 if (diff)
2770 tcp_adjust_pcount(sk, skb, diff);
2771 if (skb->len < cur_mss)
2772 tcp_retrans_try_collapse(sk, skb, cur_mss);
2773 }
2774
2775 /* RFC3168, section 6.1.1.1. ECN fallback */
2776 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2777 tcp_ecn_clear_syn(sk, skb);
2778
2779 /* make sure skb->data is aligned on arches that require it
2780 * and check if ack-trimming & collapsing extended the headroom
2781 * beyond what csum_start can cover.
2782 */
2783 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2784 skb_headroom(skb) >= 0xFFFF)) {
2785 struct sk_buff *nskb;
2786
2787 skb_mstamp_get(&skb->skb_mstamp);
2788 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2789 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2790 -ENOBUFS;
2791 } else {
2792 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2793 }
2794
2795 if (likely(!err)) {
2796 segs = tcp_skb_pcount(skb);
2797
2798 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2799 /* Update global TCP statistics. */
2800 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2801 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2802 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2803 tp->total_retrans += segs;
2804 }
2805 return err;
2806}
2807
2808int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2809{
2810 struct tcp_sock *tp = tcp_sk(sk);
2811 int err = __tcp_retransmit_skb(sk, skb, segs);
2812
2813 if (err == 0) {
2814#if FASTRETRANS_DEBUG > 0
2815 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2816 net_dbg_ratelimited("retrans_out leaked\n");
2817 }
2818#endif
2819 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2820 tp->retrans_out += tcp_skb_pcount(skb);
2821
2822 /* Save stamp of the first retransmit. */
2823 if (!tp->retrans_stamp)
2824 tp->retrans_stamp = tcp_skb_timestamp(skb);
2825
2826 } else if (err != -EBUSY) {
2827 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2828 }
2829
2830 if (tp->undo_retrans < 0)
2831 tp->undo_retrans = 0;
2832 tp->undo_retrans += tcp_skb_pcount(skb);
2833 return err;
2834}
2835
2836/* Check if we forward retransmits are possible in the current
2837 * window/congestion state.
2838 */
2839static bool tcp_can_forward_retransmit(struct sock *sk)
2840{
2841 const struct inet_connection_sock *icsk = inet_csk(sk);
2842 const struct tcp_sock *tp = tcp_sk(sk);
2843
2844 /* Forward retransmissions are possible only during Recovery. */
2845 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2846 return false;
2847
2848 /* No forward retransmissions in Reno are possible. */
2849 if (tcp_is_reno(tp))
2850 return false;
2851
2852 /* Yeah, we have to make difficult choice between forward transmission
2853 * and retransmission... Both ways have their merits...
2854 *
2855 * For now we do not retransmit anything, while we have some new
2856 * segments to send. In the other cases, follow rule 3 for
2857 * NextSeg() specified in RFC3517.
2858 */
2859
2860 if (tcp_may_send_now(sk))
2861 return false;
2862
2863 return true;
2864}
2865
2866/* This gets called after a retransmit timeout, and the initially
2867 * retransmitted data is acknowledged. It tries to continue
2868 * resending the rest of the retransmit queue, until either
2869 * we've sent it all or the congestion window limit is reached.
2870 * If doing SACK, the first ACK which comes back for a timeout
2871 * based retransmit packet might feed us FACK information again.
2872 * If so, we use it to avoid unnecessarily retransmissions.
2873 */
2874void tcp_xmit_retransmit_queue(struct sock *sk)
2875{
2876 const struct inet_connection_sock *icsk = inet_csk(sk);
2877 struct tcp_sock *tp = tcp_sk(sk);
2878 struct sk_buff *skb;
2879 struct sk_buff *hole = NULL;
2880 u32 max_segs, last_lost;
2881 int mib_idx;
2882 int fwd_rexmitting = 0;
2883
2884 if (!tp->packets_out)
2885 return;
2886
2887 if (!tp->lost_out)
2888 tp->retransmit_high = tp->snd_una;
2889
2890 if (tp->retransmit_skb_hint) {
2891 skb = tp->retransmit_skb_hint;
2892 last_lost = TCP_SKB_CB(skb)->end_seq;
2893 if (after(last_lost, tp->retransmit_high))
2894 last_lost = tp->retransmit_high;
2895 } else {
2896 skb = tcp_write_queue_head(sk);
2897 last_lost = tp->snd_una;
2898 }
2899
2900 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
2901 tcp_for_write_queue_from(skb, sk) {
2902 __u8 sacked;
2903 int segs;
2904
2905 if (skb == tcp_send_head(sk))
2906 break;
2907 /* we could do better than to assign each time */
2908 if (!hole)
2909 tp->retransmit_skb_hint = skb;
2910
2911 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
2912 if (segs <= 0)
2913 return;
2914 sacked = TCP_SKB_CB(skb)->sacked;
2915 /* In case tcp_shift_skb_data() have aggregated large skbs,
2916 * we need to make sure not sending too bigs TSO packets
2917 */
2918 segs = min_t(int, segs, max_segs);
2919
2920 if (fwd_rexmitting) {
2921begin_fwd:
2922 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2923 break;
2924 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2925
2926 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2927 tp->retransmit_high = last_lost;
2928 if (!tcp_can_forward_retransmit(sk))
2929 break;
2930 /* Backtrack if necessary to non-L'ed skb */
2931 if (hole) {
2932 skb = hole;
2933 hole = NULL;
2934 }
2935 fwd_rexmitting = 1;
2936 goto begin_fwd;
2937
2938 } else if (!(sacked & TCPCB_LOST)) {
2939 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2940 hole = skb;
2941 continue;
2942
2943 } else {
2944 last_lost = TCP_SKB_CB(skb)->end_seq;
2945 if (icsk->icsk_ca_state != TCP_CA_Loss)
2946 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2947 else
2948 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2949 }
2950
2951 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2952 continue;
2953
2954 if (tcp_small_queue_check(sk, skb, 1))
2955 return;
2956
2957 if (tcp_retransmit_skb(sk, skb, segs))
2958 return;
2959
2960 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
2961
2962 if (tcp_in_cwnd_reduction(sk))
2963 tp->prr_out += tcp_skb_pcount(skb);
2964
2965 if (skb == tcp_write_queue_head(sk))
2966 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2967 inet_csk(sk)->icsk_rto,
2968 TCP_RTO_MAX);
2969 }
2970}
2971
2972/* We allow to exceed memory limits for FIN packets to expedite
2973 * connection tear down and (memory) recovery.
2974 * Otherwise tcp_send_fin() could be tempted to either delay FIN
2975 * or even be forced to close flow without any FIN.
2976 * In general, we want to allow one skb per socket to avoid hangs
2977 * with edge trigger epoll()
2978 */
2979void sk_forced_mem_schedule(struct sock *sk, int size)
2980{
2981 int amt;
2982
2983 if (size <= sk->sk_forward_alloc)
2984 return;
2985 amt = sk_mem_pages(size);
2986 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
2987 sk_memory_allocated_add(sk, amt);
2988
2989 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
2990 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
2991}
2992
2993/* Send a FIN. The caller locks the socket for us.
2994 * We should try to send a FIN packet really hard, but eventually give up.
2995 */
2996void tcp_send_fin(struct sock *sk)
2997{
2998 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
2999 struct tcp_sock *tp = tcp_sk(sk);
3000
3001 /* Optimization, tack on the FIN if we have one skb in write queue and
3002 * this skb was not yet sent, or we are under memory pressure.
3003 * Note: in the latter case, FIN packet will be sent after a timeout,
3004 * as TCP stack thinks it has already been transmitted.
3005 */
3006 if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
3007coalesce:
3008 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3009 TCP_SKB_CB(tskb)->end_seq++;
3010 tp->write_seq++;
3011 if (!tcp_send_head(sk)) {
3012 /* This means tskb was already sent.
3013 * Pretend we included the FIN on previous transmit.
3014 * We need to set tp->snd_nxt to the value it would have
3015 * if FIN had been sent. This is because retransmit path
3016 * does not change tp->snd_nxt.
3017 */
3018 tp->snd_nxt++;
3019 return;
3020 }
3021 } else {
3022 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3023 if (unlikely(!skb)) {
3024 if (tskb)
3025 goto coalesce;
3026 return;
3027 }
3028 skb_reserve(skb, MAX_TCP_HEADER);
3029 sk_forced_mem_schedule(sk, skb->truesize);
3030 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3031 tcp_init_nondata_skb(skb, tp->write_seq,
3032 TCPHDR_ACK | TCPHDR_FIN);
3033 tcp_queue_skb(sk, skb);
3034 }
3035 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3036}
3037
3038/* We get here when a process closes a file descriptor (either due to
3039 * an explicit close() or as a byproduct of exit()'ing) and there
3040 * was unread data in the receive queue. This behavior is recommended
3041 * by RFC 2525, section 2.17. -DaveM
3042 */
3043void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3044{
3045 struct sk_buff *skb;
3046
3047 /* NOTE: No TCP options attached and we never retransmit this. */
3048 skb = alloc_skb(MAX_TCP_HEADER, priority);
3049 if (!skb) {
3050 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3051 return;
3052 }
3053
3054 /* Reserve space for headers and prepare control bits. */
3055 skb_reserve(skb, MAX_TCP_HEADER);
3056 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3057 TCPHDR_ACK | TCPHDR_RST);
3058 skb_mstamp_get(&skb->skb_mstamp);
3059 /* Send it off. */
3060 if (tcp_transmit_skb(sk, skb, 0, priority))
3061 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3062
3063 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3064}
3065
3066/* Send a crossed SYN-ACK during socket establishment.
3067 * WARNING: This routine must only be called when we have already sent
3068 * a SYN packet that crossed the incoming SYN that caused this routine
3069 * to get called. If this assumption fails then the initial rcv_wnd
3070 * and rcv_wscale values will not be correct.
3071 */
3072int tcp_send_synack(struct sock *sk)
3073{
3074 struct sk_buff *skb;
3075
3076 skb = tcp_write_queue_head(sk);
3077 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3078 pr_debug("%s: wrong queue state\n", __func__);
3079 return -EFAULT;
3080 }
3081 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3082 if (skb_cloned(skb)) {
3083 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
3084 if (!nskb)
3085 return -ENOMEM;
3086 tcp_unlink_write_queue(skb, sk);
3087 __skb_header_release(nskb);
3088 __tcp_add_write_queue_head(sk, nskb);
3089 sk_wmem_free_skb(sk, skb);
3090 sk->sk_wmem_queued += nskb->truesize;
3091 sk_mem_charge(sk, nskb->truesize);
3092 skb = nskb;
3093 }
3094
3095 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3096 tcp_ecn_send_synack(sk, skb);
3097 }
3098 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3099}
3100
3101/**
3102 * tcp_make_synack - Prepare a SYN-ACK.
3103 * sk: listener socket
3104 * dst: dst entry attached to the SYNACK
3105 * req: request_sock pointer
3106 *
3107 * Allocate one skb and build a SYNACK packet.
3108 * @dst is consumed : Caller should not use it again.
3109 */
3110struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3111 struct request_sock *req,
3112 struct tcp_fastopen_cookie *foc,
3113 enum tcp_synack_type synack_type)
3114{
3115 struct inet_request_sock *ireq = inet_rsk(req);
3116 const struct tcp_sock *tp = tcp_sk(sk);
3117 struct tcp_md5sig_key *md5 = NULL;
3118 struct tcp_out_options opts;
3119 struct sk_buff *skb;
3120 int tcp_header_size;
3121 struct tcphdr *th;
3122 u16 user_mss;
3123 int mss;
3124
3125 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3126 if (unlikely(!skb)) {
3127 dst_release(dst);
3128 return NULL;
3129 }
3130 /* Reserve space for headers. */
3131 skb_reserve(skb, MAX_TCP_HEADER);
3132
3133 switch (synack_type) {
3134 case TCP_SYNACK_NORMAL:
3135 skb_set_owner_w(skb, req_to_sk(req));
3136 break;
3137 case TCP_SYNACK_COOKIE:
3138 /* Under synflood, we do not attach skb to a socket,
3139 * to avoid false sharing.
3140 */
3141 break;
3142 case TCP_SYNACK_FASTOPEN:
3143 /* sk is a const pointer, because we want to express multiple
3144 * cpu might call us concurrently.
3145 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3146 */
3147 skb_set_owner_w(skb, (struct sock *)sk);
3148 break;
3149 }
3150 skb_dst_set(skb, dst);
3151
3152 mss = dst_metric_advmss(dst);
3153 user_mss = READ_ONCE(tp->rx_opt.user_mss);
3154 if (user_mss && user_mss < mss)
3155 mss = user_mss;
3156
3157 memset(&opts, 0, sizeof(opts));
3158#ifdef CONFIG_SYN_COOKIES
3159 if (unlikely(req->cookie_ts))
3160 skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
3161 else
3162#endif
3163 skb_mstamp_get(&skb->skb_mstamp);
3164
3165#ifdef CONFIG_TCP_MD5SIG
3166 rcu_read_lock();
3167 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3168#endif
3169 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3170 tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) +
3171 sizeof(*th);
3172
3173 skb_push(skb, tcp_header_size);
3174 skb_reset_transport_header(skb);
3175
3176 th = (struct tcphdr *)skb->data;
3177 memset(th, 0, sizeof(struct tcphdr));
3178 th->syn = 1;
3179 th->ack = 1;
3180 tcp_ecn_make_synack(req, th);
3181 th->source = htons(ireq->ir_num);
3182 th->dest = ireq->ir_rmt_port;
3183 /* Setting of flags are superfluous here for callers (and ECE is
3184 * not even correctly set)
3185 */
3186 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
3187 TCPHDR_SYN | TCPHDR_ACK);
3188
3189 th->seq = htonl(TCP_SKB_CB(skb)->seq);
3190 /* XXX data is queued and acked as is. No buffer/window check */
3191 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3192
3193 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3194 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3195 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3196 th->doff = (tcp_header_size >> 2);
3197 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3198
3199#ifdef CONFIG_TCP_MD5SIG
3200 /* Okay, we have all we need - do the md5 hash if needed */
3201 if (md5)
3202 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3203 md5, req_to_sk(req), skb);
3204 rcu_read_unlock();
3205#endif
3206
3207 /* Do not fool tcpdump (if any), clean our debris */
3208 skb->tstamp = 0;
3209 return skb;
3210}
3211EXPORT_SYMBOL(tcp_make_synack);
3212
3213static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3214{
3215 struct inet_connection_sock *icsk = inet_csk(sk);
3216 const struct tcp_congestion_ops *ca;
3217 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3218
3219 if (ca_key == TCP_CA_UNSPEC)
3220 return;
3221
3222 rcu_read_lock();
3223 ca = tcp_ca_find_key(ca_key);
3224 if (likely(ca && try_module_get(ca->owner))) {
3225 module_put(icsk->icsk_ca_ops->owner);
3226 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3227 icsk->icsk_ca_ops = ca;
3228 }
3229 rcu_read_unlock();
3230}
3231
3232/* Do all connect socket setups that can be done AF independent. */
3233static void tcp_connect_init(struct sock *sk)
3234{
3235 const struct dst_entry *dst = __sk_dst_get(sk);
3236 struct tcp_sock *tp = tcp_sk(sk);
3237 __u8 rcv_wscale;
3238
3239 /* We'll fix this up when we get a response from the other end.
3240 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3241 */
3242 tp->tcp_header_len = sizeof(struct tcphdr) +
3243 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
3244
3245#ifdef CONFIG_TCP_MD5SIG
3246 if (tp->af_specific->md5_lookup(sk, sk))
3247 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3248#endif
3249
3250 /* If user gave his TCP_MAXSEG, record it to clamp */
3251 if (tp->rx_opt.user_mss)
3252 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3253 tp->max_window = 0;
3254 tcp_mtup_init(sk);
3255 tcp_sync_mss(sk, dst_mtu(dst));
3256
3257 tcp_ca_dst_init(sk, dst);
3258
3259 if (!tp->window_clamp)
3260 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3261 tp->advmss = dst_metric_advmss(dst);
3262 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
3263 tp->advmss = tp->rx_opt.user_mss;
3264
3265 tcp_initialize_rcv_mss(sk);
3266
3267 /* limit the window selection if the user enforce a smaller rx buffer */
3268 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3269 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3270 tp->window_clamp = tcp_full_space(sk);
3271
3272 tcp_select_initial_window(tcp_full_space(sk),
3273 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3274 &tp->rcv_wnd,
3275 &tp->window_clamp,
3276 sysctl_tcp_window_scaling,
3277 &rcv_wscale,
3278 dst_metric(dst, RTAX_INITRWND));
3279
3280 tp->rx_opt.rcv_wscale = rcv_wscale;
3281 tp->rcv_ssthresh = tp->rcv_wnd;
3282
3283 sk->sk_err = 0;
3284 sock_reset_flag(sk, SOCK_DONE);
3285 tp->snd_wnd = 0;
3286 tcp_init_wl(tp, 0);
3287 tp->snd_una = tp->write_seq;
3288 tp->snd_sml = tp->write_seq;
3289 tp->snd_up = tp->write_seq;
3290 tp->snd_nxt = tp->write_seq;
3291
3292 if (likely(!tp->repair))
3293 tp->rcv_nxt = 0;
3294 else
3295 tp->rcv_tstamp = tcp_time_stamp;
3296 tp->rcv_wup = tp->rcv_nxt;
3297 tp->copied_seq = tp->rcv_nxt;
3298
3299 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
3300 inet_csk(sk)->icsk_retransmits = 0;
3301 tcp_clear_retrans(tp);
3302}
3303
3304static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3305{
3306 struct tcp_sock *tp = tcp_sk(sk);
3307 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3308
3309 tcb->end_seq += skb->len;
3310 __skb_header_release(skb);
3311 __tcp_add_write_queue_tail(sk, skb);
3312 sk->sk_wmem_queued += skb->truesize;
3313 sk_mem_charge(sk, skb->truesize);
3314 tp->write_seq = tcb->end_seq;
3315 tp->packets_out += tcp_skb_pcount(skb);
3316}
3317
3318/* Build and send a SYN with data and (cached) Fast Open cookie. However,
3319 * queue a data-only packet after the regular SYN, such that regular SYNs
3320 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3321 * only the SYN sequence, the data are retransmitted in the first ACK.
3322 * If cookie is not cached or other error occurs, falls back to send a
3323 * regular SYN with Fast Open cookie request option.
3324 */
3325static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3326{
3327 struct tcp_sock *tp = tcp_sk(sk);
3328 struct tcp_fastopen_request *fo = tp->fastopen_req;
3329 int syn_loss = 0, space, err = 0;
3330 unsigned long last_syn_loss = 0;
3331 struct sk_buff *syn_data;
3332
3333 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3334 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
3335 &syn_loss, &last_syn_loss);
3336 /* Recurring FO SYN losses: revert to regular handshake temporarily */
3337 if (syn_loss > 1 &&
3338 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
3339 fo->cookie.len = -1;
3340 goto fallback;
3341 }
3342
3343 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
3344 fo->cookie.len = -1;
3345 else if (fo->cookie.len <= 0)
3346 goto fallback;
3347
3348 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3349 * user-MSS. Reserve maximum option space for middleboxes that add
3350 * private TCP options. The cost is reduced data space in SYN :(
3351 */
3352 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
3353 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3354 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3355 MAX_TCP_OPTION_SPACE;
3356
3357 space = min_t(size_t, space, fo->size);
3358
3359 /* limit to order-0 allocations */
3360 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3361
3362 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3363 if (!syn_data)
3364 goto fallback;
3365 syn_data->ip_summed = CHECKSUM_PARTIAL;
3366 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3367 if (space) {
3368 int copied = copy_from_iter(skb_put(syn_data, space), space,
3369 &fo->data->msg_iter);
3370 if (unlikely(!copied)) {
3371 kfree_skb(syn_data);
3372 goto fallback;
3373 }
3374 if (copied != space) {
3375 skb_trim(syn_data, copied);
3376 space = copied;
3377 }
3378 }
3379 /* No more data pending in inet_wait_for_connect() */
3380 if (space == fo->size)
3381 fo->data = NULL;
3382 fo->copied = space;
3383
3384 tcp_connect_queue_skb(sk, syn_data);
3385 if (syn_data->len)
3386 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3387
3388 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3389
3390 syn->skb_mstamp = syn_data->skb_mstamp;
3391
3392 /* Now full SYN+DATA was cloned and sent (or not),
3393 * remove the SYN from the original skb (syn_data)
3394 * we keep in write queue in case of a retransmit, as we
3395 * also have the SYN packet (with no data) in the same queue.
3396 */
3397 TCP_SKB_CB(syn_data)->seq++;
3398 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3399 if (!err) {
3400 tp->syn_data = (fo->copied > 0);
3401 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3402 goto done;
3403 }
3404
3405fallback:
3406 /* Send a regular SYN with Fast Open cookie request option */
3407 if (fo->cookie.len > 0)
3408 fo->cookie.len = 0;
3409 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3410 if (err)
3411 tp->syn_fastopen = 0;
3412done:
3413 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3414 return err;
3415}
3416
3417/* Build a SYN and send it off. */
3418int tcp_connect(struct sock *sk)
3419{
3420 struct tcp_sock *tp = tcp_sk(sk);
3421 struct sk_buff *buff;
3422 int err;
3423
3424 tcp_connect_init(sk);
3425
3426 if (unlikely(tp->repair)) {
3427 tcp_finish_connect(sk, NULL);
3428 return 0;
3429 }
3430
3431 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3432 if (unlikely(!buff))
3433 return -ENOBUFS;
3434
3435 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3436 tp->retrans_stamp = tcp_time_stamp;
3437 tcp_connect_queue_skb(sk, buff);
3438 tcp_ecn_send_syn(sk, buff);
3439
3440 /* Send off SYN; include data in Fast Open. */
3441 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3442 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3443 if (err == -ECONNREFUSED)
3444 return err;
3445
3446 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3447 * in order to make this packet get counted in tcpOutSegs.
3448 */
3449 tp->snd_nxt = tp->write_seq;
3450 tp->pushed_seq = tp->write_seq;
3451 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3452
3453 /* Timer for repeating the SYN until an answer. */
3454 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3455 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3456 return 0;
3457}
3458EXPORT_SYMBOL(tcp_connect);
3459
3460/* Send out a delayed ack, the caller does the policy checking
3461 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3462 * for details.
3463 */
3464void tcp_send_delayed_ack(struct sock *sk)
3465{
3466 struct inet_connection_sock *icsk = inet_csk(sk);
3467 int ato = icsk->icsk_ack.ato;
3468 unsigned long timeout;
3469
3470 tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3471
3472 if (ato > TCP_DELACK_MIN) {
3473 const struct tcp_sock *tp = tcp_sk(sk);
3474 int max_ato = HZ / 2;
3475
3476 if (icsk->icsk_ack.pingpong ||
3477 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3478 max_ato = TCP_DELACK_MAX;
3479
3480 /* Slow path, intersegment interval is "high". */
3481
3482 /* If some rtt estimate is known, use it to bound delayed ack.
3483 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3484 * directly.
3485 */
3486 if (tp->srtt_us) {
3487 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3488 TCP_DELACK_MIN);
3489
3490 if (rtt < max_ato)
3491 max_ato = rtt;
3492 }
3493
3494 ato = min(ato, max_ato);
3495 }
3496
3497 /* Stay within the limit we were given */
3498 timeout = jiffies + ato;
3499
3500 /* Use new timeout only if there wasn't a older one earlier. */
3501 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3502 /* If delack timer was blocked or is about to expire,
3503 * send ACK now.
3504 */
3505 if (icsk->icsk_ack.blocked ||
3506 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3507 tcp_send_ack(sk);
3508 return;
3509 }
3510
3511 if (!time_before(timeout, icsk->icsk_ack.timeout))
3512 timeout = icsk->icsk_ack.timeout;
3513 }
3514 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3515 icsk->icsk_ack.timeout = timeout;
3516 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3517}
3518
3519/* This routine sends an ack and also updates the window. */
3520void tcp_send_ack(struct sock *sk)
3521{
3522 struct sk_buff *buff;
3523
3524 /* If we have been reset, we may not send again. */
3525 if (sk->sk_state == TCP_CLOSE)
3526 return;
3527
3528 tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3529
3530 /* We are not putting this on the write queue, so
3531 * tcp_transmit_skb() will set the ownership to this
3532 * sock.
3533 */
3534 buff = alloc_skb(MAX_TCP_HEADER,
3535 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3536 if (unlikely(!buff)) {
3537 inet_csk_schedule_ack(sk);
3538 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3539 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3540 TCP_DELACK_MAX, TCP_RTO_MAX);
3541 return;
3542 }
3543
3544 /* Reserve space for headers and prepare control bits. */
3545 skb_reserve(buff, MAX_TCP_HEADER);
3546 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3547
3548 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3549 * too much.
3550 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3551 */
3552 skb_set_tcp_pure_ack(buff);
3553
3554 /* Send it off, this clears delayed acks for us. */
3555 skb_mstamp_get(&buff->skb_mstamp);
3556 tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0);
3557}
3558EXPORT_SYMBOL_GPL(tcp_send_ack);
3559
3560/* This routine sends a packet with an out of date sequence
3561 * number. It assumes the other end will try to ack it.
3562 *
3563 * Question: what should we make while urgent mode?
3564 * 4.4BSD forces sending single byte of data. We cannot send
3565 * out of window data, because we have SND.NXT==SND.MAX...
3566 *
3567 * Current solution: to send TWO zero-length segments in urgent mode:
3568 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3569 * out-of-date with SND.UNA-1 to probe window.
3570 */
3571static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3572{
3573 struct tcp_sock *tp = tcp_sk(sk);
3574 struct sk_buff *skb;
3575
3576 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3577 skb = alloc_skb(MAX_TCP_HEADER,
3578 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3579 if (!skb)
3580 return -1;
3581
3582 /* Reserve space for headers and set control bits. */
3583 skb_reserve(skb, MAX_TCP_HEADER);
3584 /* Use a previous sequence. This should cause the other
3585 * end to send an ack. Don't queue or clone SKB, just
3586 * send it.
3587 */
3588 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3589 skb_mstamp_get(&skb->skb_mstamp);
3590 NET_INC_STATS(sock_net(sk), mib);
3591 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3592}
3593
3594void tcp_send_window_probe(struct sock *sk)
3595{
3596 if (sk->sk_state == TCP_ESTABLISHED) {
3597 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3598 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3599 }
3600}
3601
3602/* Initiate keepalive or window probe from timer. */
3603int tcp_write_wakeup(struct sock *sk, int mib)
3604{
3605 struct tcp_sock *tp = tcp_sk(sk);
3606 struct sk_buff *skb;
3607
3608 if (sk->sk_state == TCP_CLOSE)
3609 return -1;
3610
3611 skb = tcp_send_head(sk);
3612 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3613 int err;
3614 unsigned int mss = tcp_current_mss(sk);
3615 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3616
3617 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3618 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3619
3620 /* We are probing the opening of a window
3621 * but the window size is != 0
3622 * must have been a result SWS avoidance ( sender )
3623 */
3624 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3625 skb->len > mss) {
3626 seg_size = min(seg_size, mss);
3627 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3628 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3629 return -1;
3630 } else if (!tcp_skb_pcount(skb))
3631 tcp_set_skb_tso_segs(skb, mss);
3632
3633 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3634 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3635 if (!err)
3636 tcp_event_new_data_sent(sk, skb);
3637 return err;
3638 } else {
3639 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3640 tcp_xmit_probe_skb(sk, 1, mib);
3641 return tcp_xmit_probe_skb(sk, 0, mib);
3642 }
3643}
3644
3645/* A window probe timeout has occurred. If window is not closed send
3646 * a partial packet else a zero probe.
3647 */
3648void tcp_send_probe0(struct sock *sk)
3649{
3650 struct inet_connection_sock *icsk = inet_csk(sk);
3651 struct tcp_sock *tp = tcp_sk(sk);
3652 struct net *net = sock_net(sk);
3653 unsigned long probe_max;
3654 int err;
3655
3656 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3657
3658 if (tp->packets_out || !tcp_send_head(sk)) {
3659 /* Cancel probe timer, if it is not required. */
3660 icsk->icsk_probes_out = 0;
3661 icsk->icsk_backoff = 0;
3662 return;
3663 }
3664
3665 if (err <= 0) {
3666 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3667 icsk->icsk_backoff++;
3668 icsk->icsk_probes_out++;
3669 probe_max = TCP_RTO_MAX;
3670 } else {
3671 /* If packet was not sent due to local congestion,
3672 * do not backoff and do not remember icsk_probes_out.
3673 * Let local senders to fight for local resources.
3674 *
3675 * Use accumulated backoff yet.
3676 */
3677 if (!icsk->icsk_probes_out)
3678 icsk->icsk_probes_out = 1;
3679 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3680 }
3681 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3682 tcp_probe0_when(sk, probe_max),
3683 TCP_RTO_MAX);
3684}
3685
3686int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3687{
3688 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3689 struct flowi fl;
3690 int res;
3691
3692 tcp_rsk(req)->txhash = net_tx_rndhash();
3693 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3694 if (!res) {
3695 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3696 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3697 if (unlikely(tcp_passive_fastopen(sk)))
3698 tcp_sk(sk)->total_retrans++;
3699 }
3700 return res;
3701}
3702EXPORT_SYMBOL(tcp_rtx_synack);