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#include <net/tcp.h>
38
39#include <linux/compiler.h>
40#include <linux/gfp.h>
41#include <linux/module.h>
42
43/* People can turn this off for buggy TCP's found in printers etc. */
44int sysctl_tcp_retrans_collapse __read_mostly = 1;
45
46/* People can turn this on to work with those rare, broken TCPs that
47 * interpret the window field as a signed quantity.
48 */
49int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
50
51/* This limits the percentage of the congestion window which we
52 * will allow a single TSO frame to consume. Building TSO frames
53 * which are too large can cause TCP streams to be bursty.
54 */
55int sysctl_tcp_tso_win_divisor __read_mostly = 3;
56
57int sysctl_tcp_mtu_probing __read_mostly = 0;
58int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS;
59
60/* By default, RFC2861 behavior. */
61int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
62
63int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */
64EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size);
65
66
67/* Account for new data that has been sent to the network. */
68static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
69{
70 struct tcp_sock *tp = tcp_sk(sk);
71 unsigned int prior_packets = tp->packets_out;
72
73 tcp_advance_send_head(sk, skb);
74 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
75
76 /* Don't override Nagle indefinitely with F-RTO */
77 if (tp->frto_counter == 2)
78 tp->frto_counter = 3;
79
80 tp->packets_out += tcp_skb_pcount(skb);
81 if (!prior_packets)
82 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
83 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
84}
85
86/* SND.NXT, if window was not shrunk.
87 * If window has been shrunk, what should we make? It is not clear at all.
88 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
89 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
90 * invalid. OK, let's make this for now:
91 */
92static inline __u32 tcp_acceptable_seq(struct sock *sk)
93{
94 struct tcp_sock *tp = tcp_sk(sk);
95
96 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
97 return tp->snd_nxt;
98 else
99 return tcp_wnd_end(tp);
100}
101
102/* Calculate mss to advertise in SYN segment.
103 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
104 *
105 * 1. It is independent of path mtu.
106 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
107 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
108 * attached devices, because some buggy hosts are confused by
109 * large MSS.
110 * 4. We do not make 3, we advertise MSS, calculated from first
111 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
112 * This may be overridden via information stored in routing table.
113 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
114 * probably even Jumbo".
115 */
116static __u16 tcp_advertise_mss(struct sock *sk)
117{
118 struct tcp_sock *tp = tcp_sk(sk);
119 struct dst_entry *dst = __sk_dst_get(sk);
120 int mss = tp->advmss;
121
122 if (dst) {
123 unsigned int metric = dst_metric_advmss(dst);
124
125 if (metric < mss) {
126 mss = metric;
127 tp->advmss = mss;
128 }
129 }
130
131 return (__u16)mss;
132}
133
134/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
135 * This is the first part of cwnd validation mechanism. */
136static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
137{
138 struct tcp_sock *tp = tcp_sk(sk);
139 s32 delta = tcp_time_stamp - tp->lsndtime;
140 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
141 u32 cwnd = tp->snd_cwnd;
142
143 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
144
145 tp->snd_ssthresh = tcp_current_ssthresh(sk);
146 restart_cwnd = min(restart_cwnd, cwnd);
147
148 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
149 cwnd >>= 1;
150 tp->snd_cwnd = max(cwnd, restart_cwnd);
151 tp->snd_cwnd_stamp = tcp_time_stamp;
152 tp->snd_cwnd_used = 0;
153}
154
155/* Congestion state accounting after a packet has been sent. */
156static void tcp_event_data_sent(struct tcp_sock *tp,
157 struct sk_buff *skb, struct sock *sk)
158{
159 struct inet_connection_sock *icsk = inet_csk(sk);
160 const u32 now = tcp_time_stamp;
161
162 if (sysctl_tcp_slow_start_after_idle &&
163 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
164 tcp_cwnd_restart(sk, __sk_dst_get(sk));
165
166 tp->lsndtime = now;
167
168 /* If it is a reply for ato after last received
169 * packet, enter pingpong mode.
170 */
171 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
172 icsk->icsk_ack.pingpong = 1;
173}
174
175/* Account for an ACK we sent. */
176static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
177{
178 tcp_dec_quickack_mode(sk, pkts);
179 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
180}
181
182/* Determine a window scaling and initial window to offer.
183 * Based on the assumption that the given amount of space
184 * will be offered. Store the results in the tp structure.
185 * NOTE: for smooth operation initial space offering should
186 * be a multiple of mss if possible. We assume here that mss >= 1.
187 * This MUST be enforced by all callers.
188 */
189void tcp_select_initial_window(int __space, __u32 mss,
190 __u32 *rcv_wnd, __u32 *window_clamp,
191 int wscale_ok, __u8 *rcv_wscale,
192 __u32 init_rcv_wnd)
193{
194 unsigned int space = (__space < 0 ? 0 : __space);
195
196 /* If no clamp set the clamp to the max possible scaled window */
197 if (*window_clamp == 0)
198 (*window_clamp) = (65535 << 14);
199 space = min(*window_clamp, space);
200
201 /* Quantize space offering to a multiple of mss if possible. */
202 if (space > mss)
203 space = (space / mss) * mss;
204
205 /* NOTE: offering an initial window larger than 32767
206 * will break some buggy TCP stacks. If the admin tells us
207 * it is likely we could be speaking with such a buggy stack
208 * we will truncate our initial window offering to 32K-1
209 * unless the remote has sent us a window scaling option,
210 * which we interpret as a sign the remote TCP is not
211 * misinterpreting the window field as a signed quantity.
212 */
213 if (sysctl_tcp_workaround_signed_windows)
214 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
215 else
216 (*rcv_wnd) = space;
217
218 (*rcv_wscale) = 0;
219 if (wscale_ok) {
220 /* Set window scaling on max possible window
221 * See RFC1323 for an explanation of the limit to 14
222 */
223 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
224 space = min_t(u32, space, *window_clamp);
225 while (space > 65535 && (*rcv_wscale) < 14) {
226 space >>= 1;
227 (*rcv_wscale)++;
228 }
229 }
230
231 /* Set initial window to a value enough for senders starting with
232 * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place
233 * a limit on the initial window when mss is larger than 1460.
234 */
235 if (mss > (1 << *rcv_wscale)) {
236 int init_cwnd = TCP_DEFAULT_INIT_RCVWND;
237 if (mss > 1460)
238 init_cwnd =
239 max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
240 /* when initializing use the value from init_rcv_wnd
241 * rather than the default from above
242 */
243 if (init_rcv_wnd)
244 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
245 else
246 *rcv_wnd = min(*rcv_wnd, init_cwnd * mss);
247 }
248
249 /* Set the clamp no higher than max representable value */
250 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
251}
252EXPORT_SYMBOL(tcp_select_initial_window);
253
254/* Chose a new window to advertise, update state in tcp_sock for the
255 * socket, and return result with RFC1323 scaling applied. The return
256 * value can be stuffed directly into th->window for an outgoing
257 * frame.
258 */
259static u16 tcp_select_window(struct sock *sk)
260{
261 struct tcp_sock *tp = tcp_sk(sk);
262 u32 cur_win = tcp_receive_window(tp);
263 u32 new_win = __tcp_select_window(sk);
264
265 /* Never shrink the offered window */
266 if (new_win < cur_win) {
267 /* Danger Will Robinson!
268 * Don't update rcv_wup/rcv_wnd here or else
269 * we will not be able to advertise a zero
270 * window in time. --DaveM
271 *
272 * Relax Will Robinson.
273 */
274 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
275 }
276 tp->rcv_wnd = new_win;
277 tp->rcv_wup = tp->rcv_nxt;
278
279 /* Make sure we do not exceed the maximum possible
280 * scaled window.
281 */
282 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
283 new_win = min(new_win, MAX_TCP_WINDOW);
284 else
285 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
286
287 /* RFC1323 scaling applied */
288 new_win >>= tp->rx_opt.rcv_wscale;
289
290 /* If we advertise zero window, disable fast path. */
291 if (new_win == 0)
292 tp->pred_flags = 0;
293
294 return new_win;
295}
296
297/* Packet ECN state for a SYN-ACK */
298static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb)
299{
300 TCP_SKB_CB(skb)->flags &= ~TCPHDR_CWR;
301 if (!(tp->ecn_flags & TCP_ECN_OK))
302 TCP_SKB_CB(skb)->flags &= ~TCPHDR_ECE;
303}
304
305/* Packet ECN state for a SYN. */
306static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
307{
308 struct tcp_sock *tp = tcp_sk(sk);
309
310 tp->ecn_flags = 0;
311 if (sysctl_tcp_ecn == 1) {
312 TCP_SKB_CB(skb)->flags |= TCPHDR_ECE | TCPHDR_CWR;
313 tp->ecn_flags = TCP_ECN_OK;
314 }
315}
316
317static __inline__ void
318TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th)
319{
320 if (inet_rsk(req)->ecn_ok)
321 th->ece = 1;
322}
323
324/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
325 * be sent.
326 */
327static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
328 int tcp_header_len)
329{
330 struct tcp_sock *tp = tcp_sk(sk);
331
332 if (tp->ecn_flags & TCP_ECN_OK) {
333 /* Not-retransmitted data segment: set ECT and inject CWR. */
334 if (skb->len != tcp_header_len &&
335 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
336 INET_ECN_xmit(sk);
337 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
338 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
339 tcp_hdr(skb)->cwr = 1;
340 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
341 }
342 } else {
343 /* ACK or retransmitted segment: clear ECT|CE */
344 INET_ECN_dontxmit(sk);
345 }
346 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
347 tcp_hdr(skb)->ece = 1;
348 }
349}
350
351/* Constructs common control bits of non-data skb. If SYN/FIN is present,
352 * auto increment end seqno.
353 */
354static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
355{
356 skb->ip_summed = CHECKSUM_PARTIAL;
357 skb->csum = 0;
358
359 TCP_SKB_CB(skb)->flags = flags;
360 TCP_SKB_CB(skb)->sacked = 0;
361
362 skb_shinfo(skb)->gso_segs = 1;
363 skb_shinfo(skb)->gso_size = 0;
364 skb_shinfo(skb)->gso_type = 0;
365
366 TCP_SKB_CB(skb)->seq = seq;
367 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
368 seq++;
369 TCP_SKB_CB(skb)->end_seq = seq;
370}
371
372static inline int tcp_urg_mode(const struct tcp_sock *tp)
373{
374 return tp->snd_una != tp->snd_up;
375}
376
377#define OPTION_SACK_ADVERTISE (1 << 0)
378#define OPTION_TS (1 << 1)
379#define OPTION_MD5 (1 << 2)
380#define OPTION_WSCALE (1 << 3)
381#define OPTION_COOKIE_EXTENSION (1 << 4)
382
383struct tcp_out_options {
384 u8 options; /* bit field of OPTION_* */
385 u8 ws; /* window scale, 0 to disable */
386 u8 num_sack_blocks; /* number of SACK blocks to include */
387 u8 hash_size; /* bytes in hash_location */
388 u16 mss; /* 0 to disable */
389 __u32 tsval, tsecr; /* need to include OPTION_TS */
390 __u8 *hash_location; /* temporary pointer, overloaded */
391};
392
393/* The sysctl int routines are generic, so check consistency here.
394 */
395static u8 tcp_cookie_size_check(u8 desired)
396{
397 int cookie_size;
398
399 if (desired > 0)
400 /* previously specified */
401 return desired;
402
403 cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size);
404 if (cookie_size <= 0)
405 /* no default specified */
406 return 0;
407
408 if (cookie_size <= TCP_COOKIE_MIN)
409 /* value too small, specify minimum */
410 return TCP_COOKIE_MIN;
411
412 if (cookie_size >= TCP_COOKIE_MAX)
413 /* value too large, specify maximum */
414 return TCP_COOKIE_MAX;
415
416 if (cookie_size & 1)
417 /* 8-bit multiple, illegal, fix it */
418 cookie_size++;
419
420 return (u8)cookie_size;
421}
422
423/* Write previously computed TCP options to the packet.
424 *
425 * Beware: Something in the Internet is very sensitive to the ordering of
426 * TCP options, we learned this through the hard way, so be careful here.
427 * Luckily we can at least blame others for their non-compliance but from
428 * inter-operatibility perspective it seems that we're somewhat stuck with
429 * the ordering which we have been using if we want to keep working with
430 * those broken things (not that it currently hurts anybody as there isn't
431 * particular reason why the ordering would need to be changed).
432 *
433 * At least SACK_PERM as the first option is known to lead to a disaster
434 * (but it may well be that other scenarios fail similarly).
435 */
436static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
437 struct tcp_out_options *opts)
438{
439 u8 options = opts->options; /* mungable copy */
440
441 /* Having both authentication and cookies for security is redundant,
442 * and there's certainly not enough room. Instead, the cookie-less
443 * extension variant is proposed.
444 *
445 * Consider the pessimal case with authentication. The options
446 * could look like:
447 * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
448 */
449 if (unlikely(OPTION_MD5 & options)) {
450 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
451 *ptr++ = htonl((TCPOPT_COOKIE << 24) |
452 (TCPOLEN_COOKIE_BASE << 16) |
453 (TCPOPT_MD5SIG << 8) |
454 TCPOLEN_MD5SIG);
455 } else {
456 *ptr++ = htonl((TCPOPT_NOP << 24) |
457 (TCPOPT_NOP << 16) |
458 (TCPOPT_MD5SIG << 8) |
459 TCPOLEN_MD5SIG);
460 }
461 options &= ~OPTION_COOKIE_EXTENSION;
462 /* overload cookie hash location */
463 opts->hash_location = (__u8 *)ptr;
464 ptr += 4;
465 }
466
467 if (unlikely(opts->mss)) {
468 *ptr++ = htonl((TCPOPT_MSS << 24) |
469 (TCPOLEN_MSS << 16) |
470 opts->mss);
471 }
472
473 if (likely(OPTION_TS & options)) {
474 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
475 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
476 (TCPOLEN_SACK_PERM << 16) |
477 (TCPOPT_TIMESTAMP << 8) |
478 TCPOLEN_TIMESTAMP);
479 options &= ~OPTION_SACK_ADVERTISE;
480 } else {
481 *ptr++ = htonl((TCPOPT_NOP << 24) |
482 (TCPOPT_NOP << 16) |
483 (TCPOPT_TIMESTAMP << 8) |
484 TCPOLEN_TIMESTAMP);
485 }
486 *ptr++ = htonl(opts->tsval);
487 *ptr++ = htonl(opts->tsecr);
488 }
489
490 /* Specification requires after timestamp, so do it now.
491 *
492 * Consider the pessimal case without authentication. The options
493 * could look like:
494 * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
495 */
496 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
497 __u8 *cookie_copy = opts->hash_location;
498 u8 cookie_size = opts->hash_size;
499
500 /* 8-bit multiple handled in tcp_cookie_size_check() above,
501 * and elsewhere.
502 */
503 if (0x2 & cookie_size) {
504 __u8 *p = (__u8 *)ptr;
505
506 /* 16-bit multiple */
507 *p++ = TCPOPT_COOKIE;
508 *p++ = TCPOLEN_COOKIE_BASE + cookie_size;
509 *p++ = *cookie_copy++;
510 *p++ = *cookie_copy++;
511 ptr++;
512 cookie_size -= 2;
513 } else {
514 /* 32-bit multiple */
515 *ptr++ = htonl(((TCPOPT_NOP << 24) |
516 (TCPOPT_NOP << 16) |
517 (TCPOPT_COOKIE << 8) |
518 TCPOLEN_COOKIE_BASE) +
519 cookie_size);
520 }
521
522 if (cookie_size > 0) {
523 memcpy(ptr, cookie_copy, cookie_size);
524 ptr += (cookie_size / 4);
525 }
526 }
527
528 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
529 *ptr++ = htonl((TCPOPT_NOP << 24) |
530 (TCPOPT_NOP << 16) |
531 (TCPOPT_SACK_PERM << 8) |
532 TCPOLEN_SACK_PERM);
533 }
534
535 if (unlikely(OPTION_WSCALE & options)) {
536 *ptr++ = htonl((TCPOPT_NOP << 24) |
537 (TCPOPT_WINDOW << 16) |
538 (TCPOLEN_WINDOW << 8) |
539 opts->ws);
540 }
541
542 if (unlikely(opts->num_sack_blocks)) {
543 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
544 tp->duplicate_sack : tp->selective_acks;
545 int this_sack;
546
547 *ptr++ = htonl((TCPOPT_NOP << 24) |
548 (TCPOPT_NOP << 16) |
549 (TCPOPT_SACK << 8) |
550 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
551 TCPOLEN_SACK_PERBLOCK)));
552
553 for (this_sack = 0; this_sack < opts->num_sack_blocks;
554 ++this_sack) {
555 *ptr++ = htonl(sp[this_sack].start_seq);
556 *ptr++ = htonl(sp[this_sack].end_seq);
557 }
558
559 tp->rx_opt.dsack = 0;
560 }
561}
562
563/* Compute TCP options for SYN packets. This is not the final
564 * network wire format yet.
565 */
566static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb,
567 struct tcp_out_options *opts,
568 struct tcp_md5sig_key **md5) {
569 struct tcp_sock *tp = tcp_sk(sk);
570 struct tcp_cookie_values *cvp = tp->cookie_values;
571 unsigned remaining = MAX_TCP_OPTION_SPACE;
572 u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ?
573 tcp_cookie_size_check(cvp->cookie_desired) :
574 0;
575
576#ifdef CONFIG_TCP_MD5SIG
577 *md5 = tp->af_specific->md5_lookup(sk, sk);
578 if (*md5) {
579 opts->options |= OPTION_MD5;
580 remaining -= TCPOLEN_MD5SIG_ALIGNED;
581 }
582#else
583 *md5 = NULL;
584#endif
585
586 /* We always get an MSS option. The option bytes which will be seen in
587 * normal data packets should timestamps be used, must be in the MSS
588 * advertised. But we subtract them from tp->mss_cache so that
589 * calculations in tcp_sendmsg are simpler etc. So account for this
590 * fact here if necessary. If we don't do this correctly, as a
591 * receiver we won't recognize data packets as being full sized when we
592 * should, and thus we won't abide by the delayed ACK rules correctly.
593 * SACKs don't matter, we never delay an ACK when we have any of those
594 * going out. */
595 opts->mss = tcp_advertise_mss(sk);
596 remaining -= TCPOLEN_MSS_ALIGNED;
597
598 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
599 opts->options |= OPTION_TS;
600 opts->tsval = TCP_SKB_CB(skb)->when;
601 opts->tsecr = tp->rx_opt.ts_recent;
602 remaining -= TCPOLEN_TSTAMP_ALIGNED;
603 }
604 if (likely(sysctl_tcp_window_scaling)) {
605 opts->ws = tp->rx_opt.rcv_wscale;
606 opts->options |= OPTION_WSCALE;
607 remaining -= TCPOLEN_WSCALE_ALIGNED;
608 }
609 if (likely(sysctl_tcp_sack)) {
610 opts->options |= OPTION_SACK_ADVERTISE;
611 if (unlikely(!(OPTION_TS & opts->options)))
612 remaining -= TCPOLEN_SACKPERM_ALIGNED;
613 }
614
615 /* Note that timestamps are required by the specification.
616 *
617 * Odd numbers of bytes are prohibited by the specification, ensuring
618 * that the cookie is 16-bit aligned, and the resulting cookie pair is
619 * 32-bit aligned.
620 */
621 if (*md5 == NULL &&
622 (OPTION_TS & opts->options) &&
623 cookie_size > 0) {
624 int need = TCPOLEN_COOKIE_BASE + cookie_size;
625
626 if (0x2 & need) {
627 /* 32-bit multiple */
628 need += 2; /* NOPs */
629
630 if (need > remaining) {
631 /* try shrinking cookie to fit */
632 cookie_size -= 2;
633 need -= 4;
634 }
635 }
636 while (need > remaining && TCP_COOKIE_MIN <= cookie_size) {
637 cookie_size -= 4;
638 need -= 4;
639 }
640 if (TCP_COOKIE_MIN <= cookie_size) {
641 opts->options |= OPTION_COOKIE_EXTENSION;
642 opts->hash_location = (__u8 *)&cvp->cookie_pair[0];
643 opts->hash_size = cookie_size;
644
645 /* Remember for future incarnations. */
646 cvp->cookie_desired = cookie_size;
647
648 if (cvp->cookie_desired != cvp->cookie_pair_size) {
649 /* Currently use random bytes as a nonce,
650 * assuming these are completely unpredictable
651 * by hostile users of the same system.
652 */
653 get_random_bytes(&cvp->cookie_pair[0],
654 cookie_size);
655 cvp->cookie_pair_size = cookie_size;
656 }
657
658 remaining -= need;
659 }
660 }
661 return MAX_TCP_OPTION_SPACE - remaining;
662}
663
664/* Set up TCP options for SYN-ACKs. */
665static unsigned tcp_synack_options(struct sock *sk,
666 struct request_sock *req,
667 unsigned mss, struct sk_buff *skb,
668 struct tcp_out_options *opts,
669 struct tcp_md5sig_key **md5,
670 struct tcp_extend_values *xvp)
671{
672 struct inet_request_sock *ireq = inet_rsk(req);
673 unsigned remaining = MAX_TCP_OPTION_SPACE;
674 u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ?
675 xvp->cookie_plus :
676 0;
677
678#ifdef CONFIG_TCP_MD5SIG
679 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
680 if (*md5) {
681 opts->options |= OPTION_MD5;
682 remaining -= TCPOLEN_MD5SIG_ALIGNED;
683
684 /* We can't fit any SACK blocks in a packet with MD5 + TS
685 * options. There was discussion about disabling SACK
686 * rather than TS in order to fit in better with old,
687 * buggy kernels, but that was deemed to be unnecessary.
688 */
689 ireq->tstamp_ok &= !ireq->sack_ok;
690 }
691#else
692 *md5 = NULL;
693#endif
694
695 /* We always send an MSS option. */
696 opts->mss = mss;
697 remaining -= TCPOLEN_MSS_ALIGNED;
698
699 if (likely(ireq->wscale_ok)) {
700 opts->ws = ireq->rcv_wscale;
701 opts->options |= OPTION_WSCALE;
702 remaining -= TCPOLEN_WSCALE_ALIGNED;
703 }
704 if (likely(ireq->tstamp_ok)) {
705 opts->options |= OPTION_TS;
706 opts->tsval = TCP_SKB_CB(skb)->when;
707 opts->tsecr = req->ts_recent;
708 remaining -= TCPOLEN_TSTAMP_ALIGNED;
709 }
710 if (likely(ireq->sack_ok)) {
711 opts->options |= OPTION_SACK_ADVERTISE;
712 if (unlikely(!ireq->tstamp_ok))
713 remaining -= TCPOLEN_SACKPERM_ALIGNED;
714 }
715
716 /* Similar rationale to tcp_syn_options() applies here, too.
717 * If the <SYN> options fit, the same options should fit now!
718 */
719 if (*md5 == NULL &&
720 ireq->tstamp_ok &&
721 cookie_plus > TCPOLEN_COOKIE_BASE) {
722 int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */
723
724 if (0x2 & need) {
725 /* 32-bit multiple */
726 need += 2; /* NOPs */
727 }
728 if (need <= remaining) {
729 opts->options |= OPTION_COOKIE_EXTENSION;
730 opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE;
731 remaining -= need;
732 } else {
733 /* There's no error return, so flag it. */
734 xvp->cookie_out_never = 1; /* true */
735 opts->hash_size = 0;
736 }
737 }
738 return MAX_TCP_OPTION_SPACE - remaining;
739}
740
741/* Compute TCP options for ESTABLISHED sockets. This is not the
742 * final wire format yet.
743 */
744static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb,
745 struct tcp_out_options *opts,
746 struct tcp_md5sig_key **md5) {
747 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
748 struct tcp_sock *tp = tcp_sk(sk);
749 unsigned size = 0;
750 unsigned int eff_sacks;
751
752#ifdef CONFIG_TCP_MD5SIG
753 *md5 = tp->af_specific->md5_lookup(sk, sk);
754 if (unlikely(*md5)) {
755 opts->options |= OPTION_MD5;
756 size += TCPOLEN_MD5SIG_ALIGNED;
757 }
758#else
759 *md5 = NULL;
760#endif
761
762 if (likely(tp->rx_opt.tstamp_ok)) {
763 opts->options |= OPTION_TS;
764 opts->tsval = tcb ? tcb->when : 0;
765 opts->tsecr = tp->rx_opt.ts_recent;
766 size += TCPOLEN_TSTAMP_ALIGNED;
767 }
768
769 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
770 if (unlikely(eff_sacks)) {
771 const unsigned remaining = MAX_TCP_OPTION_SPACE - size;
772 opts->num_sack_blocks =
773 min_t(unsigned, eff_sacks,
774 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
775 TCPOLEN_SACK_PERBLOCK);
776 size += TCPOLEN_SACK_BASE_ALIGNED +
777 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
778 }
779
780 return size;
781}
782
783/* This routine actually transmits TCP packets queued in by
784 * tcp_do_sendmsg(). This is used by both the initial
785 * transmission and possible later retransmissions.
786 * All SKB's seen here are completely headerless. It is our
787 * job to build the TCP header, and pass the packet down to
788 * IP so it can do the same plus pass the packet off to the
789 * device.
790 *
791 * We are working here with either a clone of the original
792 * SKB, or a fresh unique copy made by the retransmit engine.
793 */
794static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
795 gfp_t gfp_mask)
796{
797 const struct inet_connection_sock *icsk = inet_csk(sk);
798 struct inet_sock *inet;
799 struct tcp_sock *tp;
800 struct tcp_skb_cb *tcb;
801 struct tcp_out_options opts;
802 unsigned tcp_options_size, tcp_header_size;
803 struct tcp_md5sig_key *md5;
804 struct tcphdr *th;
805 int err;
806
807 BUG_ON(!skb || !tcp_skb_pcount(skb));
808
809 /* If congestion control is doing timestamping, we must
810 * take such a timestamp before we potentially clone/copy.
811 */
812 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
813 __net_timestamp(skb);
814
815 if (likely(clone_it)) {
816 if (unlikely(skb_cloned(skb)))
817 skb = pskb_copy(skb, gfp_mask);
818 else
819 skb = skb_clone(skb, gfp_mask);
820 if (unlikely(!skb))
821 return -ENOBUFS;
822 }
823
824 inet = inet_sk(sk);
825 tp = tcp_sk(sk);
826 tcb = TCP_SKB_CB(skb);
827 memset(&opts, 0, sizeof(opts));
828
829 if (unlikely(tcb->flags & TCPHDR_SYN))
830 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
831 else
832 tcp_options_size = tcp_established_options(sk, skb, &opts,
833 &md5);
834 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
835
836 if (tcp_packets_in_flight(tp) == 0) {
837 tcp_ca_event(sk, CA_EVENT_TX_START);
838 skb->ooo_okay = 1;
839 } else
840 skb->ooo_okay = 0;
841
842 skb_push(skb, tcp_header_size);
843 skb_reset_transport_header(skb);
844 skb_set_owner_w(skb, sk);
845
846 /* Build TCP header and checksum it. */
847 th = tcp_hdr(skb);
848 th->source = inet->inet_sport;
849 th->dest = inet->inet_dport;
850 th->seq = htonl(tcb->seq);
851 th->ack_seq = htonl(tp->rcv_nxt);
852 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
853 tcb->flags);
854
855 if (unlikely(tcb->flags & TCPHDR_SYN)) {
856 /* RFC1323: The window in SYN & SYN/ACK segments
857 * is never scaled.
858 */
859 th->window = htons(min(tp->rcv_wnd, 65535U));
860 } else {
861 th->window = htons(tcp_select_window(sk));
862 }
863 th->check = 0;
864 th->urg_ptr = 0;
865
866 /* The urg_mode check is necessary during a below snd_una win probe */
867 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
868 if (before(tp->snd_up, tcb->seq + 0x10000)) {
869 th->urg_ptr = htons(tp->snd_up - tcb->seq);
870 th->urg = 1;
871 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
872 th->urg_ptr = htons(0xFFFF);
873 th->urg = 1;
874 }
875 }
876
877 tcp_options_write((__be32 *)(th + 1), tp, &opts);
878 if (likely((tcb->flags & TCPHDR_SYN) == 0))
879 TCP_ECN_send(sk, skb, tcp_header_size);
880
881#ifdef CONFIG_TCP_MD5SIG
882 /* Calculate the MD5 hash, as we have all we need now */
883 if (md5) {
884 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
885 tp->af_specific->calc_md5_hash(opts.hash_location,
886 md5, sk, NULL, skb);
887 }
888#endif
889
890 icsk->icsk_af_ops->send_check(sk, skb);
891
892 if (likely(tcb->flags & TCPHDR_ACK))
893 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
894
895 if (skb->len != tcp_header_size)
896 tcp_event_data_sent(tp, skb, sk);
897
898 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
899 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
900 tcp_skb_pcount(skb));
901
902 err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
903 if (likely(err <= 0))
904 return err;
905
906 tcp_enter_cwr(sk, 1);
907
908 return net_xmit_eval(err);
909}
910
911/* This routine just queues the buffer for sending.
912 *
913 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
914 * otherwise socket can stall.
915 */
916static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
917{
918 struct tcp_sock *tp = tcp_sk(sk);
919
920 /* Advance write_seq and place onto the write_queue. */
921 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
922 skb_header_release(skb);
923 tcp_add_write_queue_tail(sk, skb);
924 sk->sk_wmem_queued += skb->truesize;
925 sk_mem_charge(sk, skb->truesize);
926}
927
928/* Initialize TSO segments for a packet. */
929static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb,
930 unsigned int mss_now)
931{
932 if (skb->len <= mss_now || !sk_can_gso(sk) ||
933 skb->ip_summed == CHECKSUM_NONE) {
934 /* Avoid the costly divide in the normal
935 * non-TSO case.
936 */
937 skb_shinfo(skb)->gso_segs = 1;
938 skb_shinfo(skb)->gso_size = 0;
939 skb_shinfo(skb)->gso_type = 0;
940 } else {
941 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
942 skb_shinfo(skb)->gso_size = mss_now;
943 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
944 }
945}
946
947/* When a modification to fackets out becomes necessary, we need to check
948 * skb is counted to fackets_out or not.
949 */
950static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
951 int decr)
952{
953 struct tcp_sock *tp = tcp_sk(sk);
954
955 if (!tp->sacked_out || tcp_is_reno(tp))
956 return;
957
958 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
959 tp->fackets_out -= decr;
960}
961
962/* Pcount in the middle of the write queue got changed, we need to do various
963 * tweaks to fix counters
964 */
965static void tcp_adjust_pcount(struct sock *sk, struct sk_buff *skb, int decr)
966{
967 struct tcp_sock *tp = tcp_sk(sk);
968
969 tp->packets_out -= decr;
970
971 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
972 tp->sacked_out -= decr;
973 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
974 tp->retrans_out -= decr;
975 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
976 tp->lost_out -= decr;
977
978 /* Reno case is special. Sigh... */
979 if (tcp_is_reno(tp) && decr > 0)
980 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
981
982 tcp_adjust_fackets_out(sk, skb, decr);
983
984 if (tp->lost_skb_hint &&
985 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
986 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
987 tp->lost_cnt_hint -= decr;
988
989 tcp_verify_left_out(tp);
990}
991
992/* Function to create two new TCP segments. Shrinks the given segment
993 * to the specified size and appends a new segment with the rest of the
994 * packet to the list. This won't be called frequently, I hope.
995 * Remember, these are still headerless SKBs at this point.
996 */
997int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
998 unsigned int mss_now)
999{
1000 struct tcp_sock *tp = tcp_sk(sk);
1001 struct sk_buff *buff;
1002 int nsize, old_factor;
1003 int nlen;
1004 u8 flags;
1005
1006 if (WARN_ON(len > skb->len))
1007 return -EINVAL;
1008
1009 nsize = skb_headlen(skb) - len;
1010 if (nsize < 0)
1011 nsize = 0;
1012
1013 if (skb_cloned(skb) &&
1014 skb_is_nonlinear(skb) &&
1015 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1016 return -ENOMEM;
1017
1018 /* Get a new skb... force flag on. */
1019 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1020 if (buff == NULL)
1021 return -ENOMEM; /* We'll just try again later. */
1022
1023 sk->sk_wmem_queued += buff->truesize;
1024 sk_mem_charge(sk, buff->truesize);
1025 nlen = skb->len - len - nsize;
1026 buff->truesize += nlen;
1027 skb->truesize -= nlen;
1028
1029 /* Correct the sequence numbers. */
1030 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1031 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1032 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1033
1034 /* PSH and FIN should only be set in the second packet. */
1035 flags = TCP_SKB_CB(skb)->flags;
1036 TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1037 TCP_SKB_CB(buff)->flags = flags;
1038 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1039
1040 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1041 /* Copy and checksum data tail into the new buffer. */
1042 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1043 skb_put(buff, nsize),
1044 nsize, 0);
1045
1046 skb_trim(skb, len);
1047
1048 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1049 } else {
1050 skb->ip_summed = CHECKSUM_PARTIAL;
1051 skb_split(skb, buff, len);
1052 }
1053
1054 buff->ip_summed = skb->ip_summed;
1055
1056 /* Looks stupid, but our code really uses when of
1057 * skbs, which it never sent before. --ANK
1058 */
1059 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1060 buff->tstamp = skb->tstamp;
1061
1062 old_factor = tcp_skb_pcount(skb);
1063
1064 /* Fix up tso_factor for both original and new SKB. */
1065 tcp_set_skb_tso_segs(sk, skb, mss_now);
1066 tcp_set_skb_tso_segs(sk, buff, mss_now);
1067
1068 /* If this packet has been sent out already, we must
1069 * adjust the various packet counters.
1070 */
1071 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1072 int diff = old_factor - tcp_skb_pcount(skb) -
1073 tcp_skb_pcount(buff);
1074
1075 if (diff)
1076 tcp_adjust_pcount(sk, skb, diff);
1077 }
1078
1079 /* Link BUFF into the send queue. */
1080 skb_header_release(buff);
1081 tcp_insert_write_queue_after(skb, buff, sk);
1082
1083 return 0;
1084}
1085
1086/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1087 * eventually). The difference is that pulled data not copied, but
1088 * immediately discarded.
1089 */
1090static void __pskb_trim_head(struct sk_buff *skb, int len)
1091{
1092 int i, k, eat;
1093
1094 eat = len;
1095 k = 0;
1096 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1097 if (skb_shinfo(skb)->frags[i].size <= eat) {
1098 put_page(skb_shinfo(skb)->frags[i].page);
1099 eat -= skb_shinfo(skb)->frags[i].size;
1100 } else {
1101 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1102 if (eat) {
1103 skb_shinfo(skb)->frags[k].page_offset += eat;
1104 skb_shinfo(skb)->frags[k].size -= eat;
1105 eat = 0;
1106 }
1107 k++;
1108 }
1109 }
1110 skb_shinfo(skb)->nr_frags = k;
1111
1112 skb_reset_tail_pointer(skb);
1113 skb->data_len -= len;
1114 skb->len = skb->data_len;
1115}
1116
1117/* Remove acked data from a packet in the transmit queue. */
1118int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1119{
1120 if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1121 return -ENOMEM;
1122
1123 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
1124 if (unlikely(len < skb_headlen(skb)))
1125 __skb_pull(skb, len);
1126 else
1127 __pskb_trim_head(skb, len - skb_headlen(skb));
1128
1129 TCP_SKB_CB(skb)->seq += len;
1130 skb->ip_summed = CHECKSUM_PARTIAL;
1131
1132 skb->truesize -= len;
1133 sk->sk_wmem_queued -= len;
1134 sk_mem_uncharge(sk, len);
1135 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1136
1137 /* Any change of skb->len requires recalculation of tso
1138 * factor and mss.
1139 */
1140 if (tcp_skb_pcount(skb) > 1)
1141 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk));
1142
1143 return 0;
1144}
1145
1146/* Calculate MSS. Not accounting for SACKs here. */
1147int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1148{
1149 struct tcp_sock *tp = tcp_sk(sk);
1150 struct inet_connection_sock *icsk = inet_csk(sk);
1151 int mss_now;
1152
1153 /* Calculate base mss without TCP options:
1154 It is MMS_S - sizeof(tcphdr) of rfc1122
1155 */
1156 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1157
1158 /* Clamp it (mss_clamp does not include tcp options) */
1159 if (mss_now > tp->rx_opt.mss_clamp)
1160 mss_now = tp->rx_opt.mss_clamp;
1161
1162 /* Now subtract optional transport overhead */
1163 mss_now -= icsk->icsk_ext_hdr_len;
1164
1165 /* Then reserve room for full set of TCP options and 8 bytes of data */
1166 if (mss_now < 48)
1167 mss_now = 48;
1168
1169 /* Now subtract TCP options size, not including SACKs */
1170 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
1171
1172 return mss_now;
1173}
1174
1175/* Inverse of above */
1176int tcp_mss_to_mtu(struct sock *sk, int mss)
1177{
1178 struct tcp_sock *tp = tcp_sk(sk);
1179 struct inet_connection_sock *icsk = inet_csk(sk);
1180 int mtu;
1181
1182 mtu = mss +
1183 tp->tcp_header_len +
1184 icsk->icsk_ext_hdr_len +
1185 icsk->icsk_af_ops->net_header_len;
1186
1187 return mtu;
1188}
1189
1190/* MTU probing init per socket */
1191void tcp_mtup_init(struct sock *sk)
1192{
1193 struct tcp_sock *tp = tcp_sk(sk);
1194 struct inet_connection_sock *icsk = inet_csk(sk);
1195
1196 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1197 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1198 icsk->icsk_af_ops->net_header_len;
1199 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1200 icsk->icsk_mtup.probe_size = 0;
1201}
1202EXPORT_SYMBOL(tcp_mtup_init);
1203
1204/* This function synchronize snd mss to current pmtu/exthdr set.
1205
1206 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1207 for TCP options, but includes only bare TCP header.
1208
1209 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1210 It is minimum of user_mss and mss received with SYN.
1211 It also does not include TCP options.
1212
1213 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1214
1215 tp->mss_cache is current effective sending mss, including
1216 all tcp options except for SACKs. It is evaluated,
1217 taking into account current pmtu, but never exceeds
1218 tp->rx_opt.mss_clamp.
1219
1220 NOTE1. rfc1122 clearly states that advertised MSS
1221 DOES NOT include either tcp or ip options.
1222
1223 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1224 are READ ONLY outside this function. --ANK (980731)
1225 */
1226unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1227{
1228 struct tcp_sock *tp = tcp_sk(sk);
1229 struct inet_connection_sock *icsk = inet_csk(sk);
1230 int mss_now;
1231
1232 if (icsk->icsk_mtup.search_high > pmtu)
1233 icsk->icsk_mtup.search_high = pmtu;
1234
1235 mss_now = tcp_mtu_to_mss(sk, pmtu);
1236 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1237
1238 /* And store cached results */
1239 icsk->icsk_pmtu_cookie = pmtu;
1240 if (icsk->icsk_mtup.enabled)
1241 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1242 tp->mss_cache = mss_now;
1243
1244 return mss_now;
1245}
1246EXPORT_SYMBOL(tcp_sync_mss);
1247
1248/* Compute the current effective MSS, taking SACKs and IP options,
1249 * and even PMTU discovery events into account.
1250 */
1251unsigned int tcp_current_mss(struct sock *sk)
1252{
1253 struct tcp_sock *tp = tcp_sk(sk);
1254 struct dst_entry *dst = __sk_dst_get(sk);
1255 u32 mss_now;
1256 unsigned header_len;
1257 struct tcp_out_options opts;
1258 struct tcp_md5sig_key *md5;
1259
1260 mss_now = tp->mss_cache;
1261
1262 if (dst) {
1263 u32 mtu = dst_mtu(dst);
1264 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1265 mss_now = tcp_sync_mss(sk, mtu);
1266 }
1267
1268 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1269 sizeof(struct tcphdr);
1270 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1271 * some common options. If this is an odd packet (because we have SACK
1272 * blocks etc) then our calculated header_len will be different, and
1273 * we have to adjust mss_now correspondingly */
1274 if (header_len != tp->tcp_header_len) {
1275 int delta = (int) header_len - tp->tcp_header_len;
1276 mss_now -= delta;
1277 }
1278
1279 return mss_now;
1280}
1281
1282/* Congestion window validation. (RFC2861) */
1283static void tcp_cwnd_validate(struct sock *sk)
1284{
1285 struct tcp_sock *tp = tcp_sk(sk);
1286
1287 if (tp->packets_out >= tp->snd_cwnd) {
1288 /* Network is feed fully. */
1289 tp->snd_cwnd_used = 0;
1290 tp->snd_cwnd_stamp = tcp_time_stamp;
1291 } else {
1292 /* Network starves. */
1293 if (tp->packets_out > tp->snd_cwnd_used)
1294 tp->snd_cwnd_used = tp->packets_out;
1295
1296 if (sysctl_tcp_slow_start_after_idle &&
1297 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1298 tcp_cwnd_application_limited(sk);
1299 }
1300}
1301
1302/* Returns the portion of skb which can be sent right away without
1303 * introducing MSS oddities to segment boundaries. In rare cases where
1304 * mss_now != mss_cache, we will request caller to create a small skb
1305 * per input skb which could be mostly avoided here (if desired).
1306 *
1307 * We explicitly want to create a request for splitting write queue tail
1308 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1309 * thus all the complexity (cwnd_len is always MSS multiple which we
1310 * return whenever allowed by the other factors). Basically we need the
1311 * modulo only when the receiver window alone is the limiting factor or
1312 * when we would be allowed to send the split-due-to-Nagle skb fully.
1313 */
1314static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb,
1315 unsigned int mss_now, unsigned int cwnd)
1316{
1317 struct tcp_sock *tp = tcp_sk(sk);
1318 u32 needed, window, cwnd_len;
1319
1320 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1321 cwnd_len = mss_now * cwnd;
1322
1323 if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1324 return cwnd_len;
1325
1326 needed = min(skb->len, window);
1327
1328 if (cwnd_len <= needed)
1329 return cwnd_len;
1330
1331 return needed - needed % mss_now;
1332}
1333
1334/* Can at least one segment of SKB be sent right now, according to the
1335 * congestion window rules? If so, return how many segments are allowed.
1336 */
1337static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp,
1338 struct sk_buff *skb)
1339{
1340 u32 in_flight, cwnd;
1341
1342 /* Don't be strict about the congestion window for the final FIN. */
1343 if ((TCP_SKB_CB(skb)->flags & TCPHDR_FIN) && tcp_skb_pcount(skb) == 1)
1344 return 1;
1345
1346 in_flight = tcp_packets_in_flight(tp);
1347 cwnd = tp->snd_cwnd;
1348 if (in_flight < cwnd)
1349 return (cwnd - in_flight);
1350
1351 return 0;
1352}
1353
1354/* Initialize TSO state of a skb.
1355 * This must be invoked the first time we consider transmitting
1356 * SKB onto the wire.
1357 */
1358static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb,
1359 unsigned int mss_now)
1360{
1361 int tso_segs = tcp_skb_pcount(skb);
1362
1363 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1364 tcp_set_skb_tso_segs(sk, skb, mss_now);
1365 tso_segs = tcp_skb_pcount(skb);
1366 }
1367 return tso_segs;
1368}
1369
1370/* Minshall's variant of the Nagle send check. */
1371static inline int tcp_minshall_check(const struct tcp_sock *tp)
1372{
1373 return after(tp->snd_sml, tp->snd_una) &&
1374 !after(tp->snd_sml, tp->snd_nxt);
1375}
1376
1377/* Return 0, if packet can be sent now without violation Nagle's rules:
1378 * 1. It is full sized.
1379 * 2. Or it contains FIN. (already checked by caller)
1380 * 3. Or TCP_NODELAY was set.
1381 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1382 * With Minshall's modification: all sent small packets are ACKed.
1383 */
1384static inline int tcp_nagle_check(const struct tcp_sock *tp,
1385 const struct sk_buff *skb,
1386 unsigned mss_now, int nonagle)
1387{
1388 return skb->len < mss_now &&
1389 ((nonagle & TCP_NAGLE_CORK) ||
1390 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1391}
1392
1393/* Return non-zero if the Nagle test allows this packet to be
1394 * sent now.
1395 */
1396static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1397 unsigned int cur_mss, int nonagle)
1398{
1399 /* Nagle rule does not apply to frames, which sit in the middle of the
1400 * write_queue (they have no chances to get new data).
1401 *
1402 * This is implemented in the callers, where they modify the 'nonagle'
1403 * argument based upon the location of SKB in the send queue.
1404 */
1405 if (nonagle & TCP_NAGLE_PUSH)
1406 return 1;
1407
1408 /* Don't use the nagle rule for urgent data (or for the final FIN).
1409 * Nagle can be ignored during F-RTO too (see RFC4138).
1410 */
1411 if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
1412 (TCP_SKB_CB(skb)->flags & TCPHDR_FIN))
1413 return 1;
1414
1415 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1416 return 1;
1417
1418 return 0;
1419}
1420
1421/* Does at least the first segment of SKB fit into the send window? */
1422static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb,
1423 unsigned int cur_mss)
1424{
1425 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1426
1427 if (skb->len > cur_mss)
1428 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1429
1430 return !after(end_seq, tcp_wnd_end(tp));
1431}
1432
1433/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1434 * should be put on the wire right now. If so, it returns the number of
1435 * packets allowed by the congestion window.
1436 */
1437static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1438 unsigned int cur_mss, int nonagle)
1439{
1440 struct tcp_sock *tp = tcp_sk(sk);
1441 unsigned int cwnd_quota;
1442
1443 tcp_init_tso_segs(sk, skb, cur_mss);
1444
1445 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1446 return 0;
1447
1448 cwnd_quota = tcp_cwnd_test(tp, skb);
1449 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1450 cwnd_quota = 0;
1451
1452 return cwnd_quota;
1453}
1454
1455/* Test if sending is allowed right now. */
1456int tcp_may_send_now(struct sock *sk)
1457{
1458 struct tcp_sock *tp = tcp_sk(sk);
1459 struct sk_buff *skb = tcp_send_head(sk);
1460
1461 return skb &&
1462 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1463 (tcp_skb_is_last(sk, skb) ?
1464 tp->nonagle : TCP_NAGLE_PUSH));
1465}
1466
1467/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1468 * which is put after SKB on the list. It is very much like
1469 * tcp_fragment() except that it may make several kinds of assumptions
1470 * in order to speed up the splitting operation. In particular, we
1471 * know that all the data is in scatter-gather pages, and that the
1472 * packet has never been sent out before (and thus is not cloned).
1473 */
1474static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1475 unsigned int mss_now, gfp_t gfp)
1476{
1477 struct sk_buff *buff;
1478 int nlen = skb->len - len;
1479 u8 flags;
1480
1481 /* All of a TSO frame must be composed of paged data. */
1482 if (skb->len != skb->data_len)
1483 return tcp_fragment(sk, skb, len, mss_now);
1484
1485 buff = sk_stream_alloc_skb(sk, 0, gfp);
1486 if (unlikely(buff == NULL))
1487 return -ENOMEM;
1488
1489 sk->sk_wmem_queued += buff->truesize;
1490 sk_mem_charge(sk, buff->truesize);
1491 buff->truesize += nlen;
1492 skb->truesize -= nlen;
1493
1494 /* Correct the sequence numbers. */
1495 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1496 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1497 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1498
1499 /* PSH and FIN should only be set in the second packet. */
1500 flags = TCP_SKB_CB(skb)->flags;
1501 TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1502 TCP_SKB_CB(buff)->flags = flags;
1503
1504 /* This packet was never sent out yet, so no SACK bits. */
1505 TCP_SKB_CB(buff)->sacked = 0;
1506
1507 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1508 skb_split(skb, buff, len);
1509
1510 /* Fix up tso_factor for both original and new SKB. */
1511 tcp_set_skb_tso_segs(sk, skb, mss_now);
1512 tcp_set_skb_tso_segs(sk, buff, mss_now);
1513
1514 /* Link BUFF into the send queue. */
1515 skb_header_release(buff);
1516 tcp_insert_write_queue_after(skb, buff, sk);
1517
1518 return 0;
1519}
1520
1521/* Try to defer sending, if possible, in order to minimize the amount
1522 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1523 *
1524 * This algorithm is from John Heffner.
1525 */
1526static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1527{
1528 struct tcp_sock *tp = tcp_sk(sk);
1529 const struct inet_connection_sock *icsk = inet_csk(sk);
1530 u32 send_win, cong_win, limit, in_flight;
1531 int win_divisor;
1532
1533 if (TCP_SKB_CB(skb)->flags & TCPHDR_FIN)
1534 goto send_now;
1535
1536 if (icsk->icsk_ca_state != TCP_CA_Open)
1537 goto send_now;
1538
1539 /* Defer for less than two clock ticks. */
1540 if (tp->tso_deferred &&
1541 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1542 goto send_now;
1543
1544 in_flight = tcp_packets_in_flight(tp);
1545
1546 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1547
1548 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1549
1550 /* From in_flight test above, we know that cwnd > in_flight. */
1551 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1552
1553 limit = min(send_win, cong_win);
1554
1555 /* If a full-sized TSO skb can be sent, do it. */
1556 if (limit >= sk->sk_gso_max_size)
1557 goto send_now;
1558
1559 /* Middle in queue won't get any more data, full sendable already? */
1560 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1561 goto send_now;
1562
1563 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1564 if (win_divisor) {
1565 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1566
1567 /* If at least some fraction of a window is available,
1568 * just use it.
1569 */
1570 chunk /= win_divisor;
1571 if (limit >= chunk)
1572 goto send_now;
1573 } else {
1574 /* Different approach, try not to defer past a single
1575 * ACK. Receiver should ACK every other full sized
1576 * frame, so if we have space for more than 3 frames
1577 * then send now.
1578 */
1579 if (limit > tcp_max_burst(tp) * tp->mss_cache)
1580 goto send_now;
1581 }
1582
1583 /* Ok, it looks like it is advisable to defer. */
1584 tp->tso_deferred = 1 | (jiffies << 1);
1585
1586 return 1;
1587
1588send_now:
1589 tp->tso_deferred = 0;
1590 return 0;
1591}
1592
1593/* Create a new MTU probe if we are ready.
1594 * MTU probe is regularly attempting to increase the path MTU by
1595 * deliberately sending larger packets. This discovers routing
1596 * changes resulting in larger path MTUs.
1597 *
1598 * Returns 0 if we should wait to probe (no cwnd available),
1599 * 1 if a probe was sent,
1600 * -1 otherwise
1601 */
1602static int tcp_mtu_probe(struct sock *sk)
1603{
1604 struct tcp_sock *tp = tcp_sk(sk);
1605 struct inet_connection_sock *icsk = inet_csk(sk);
1606 struct sk_buff *skb, *nskb, *next;
1607 int len;
1608 int probe_size;
1609 int size_needed;
1610 int copy;
1611 int mss_now;
1612
1613 /* Not currently probing/verifying,
1614 * not in recovery,
1615 * have enough cwnd, and
1616 * not SACKing (the variable headers throw things off) */
1617 if (!icsk->icsk_mtup.enabled ||
1618 icsk->icsk_mtup.probe_size ||
1619 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1620 tp->snd_cwnd < 11 ||
1621 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1622 return -1;
1623
1624 /* Very simple search strategy: just double the MSS. */
1625 mss_now = tcp_current_mss(sk);
1626 probe_size = 2 * tp->mss_cache;
1627 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1628 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1629 /* TODO: set timer for probe_converge_event */
1630 return -1;
1631 }
1632
1633 /* Have enough data in the send queue to probe? */
1634 if (tp->write_seq - tp->snd_nxt < size_needed)
1635 return -1;
1636
1637 if (tp->snd_wnd < size_needed)
1638 return -1;
1639 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1640 return 0;
1641
1642 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1643 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1644 if (!tcp_packets_in_flight(tp))
1645 return -1;
1646 else
1647 return 0;
1648 }
1649
1650 /* We're allowed to probe. Build it now. */
1651 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1652 return -1;
1653 sk->sk_wmem_queued += nskb->truesize;
1654 sk_mem_charge(sk, nskb->truesize);
1655
1656 skb = tcp_send_head(sk);
1657
1658 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1659 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1660 TCP_SKB_CB(nskb)->flags = TCPHDR_ACK;
1661 TCP_SKB_CB(nskb)->sacked = 0;
1662 nskb->csum = 0;
1663 nskb->ip_summed = skb->ip_summed;
1664
1665 tcp_insert_write_queue_before(nskb, skb, sk);
1666
1667 len = 0;
1668 tcp_for_write_queue_from_safe(skb, next, sk) {
1669 copy = min_t(int, skb->len, probe_size - len);
1670 if (nskb->ip_summed)
1671 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1672 else
1673 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1674 skb_put(nskb, copy),
1675 copy, nskb->csum);
1676
1677 if (skb->len <= copy) {
1678 /* We've eaten all the data from this skb.
1679 * Throw it away. */
1680 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
1681 tcp_unlink_write_queue(skb, sk);
1682 sk_wmem_free_skb(sk, skb);
1683 } else {
1684 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
1685 ~(TCPHDR_FIN|TCPHDR_PSH);
1686 if (!skb_shinfo(skb)->nr_frags) {
1687 skb_pull(skb, copy);
1688 if (skb->ip_summed != CHECKSUM_PARTIAL)
1689 skb->csum = csum_partial(skb->data,
1690 skb->len, 0);
1691 } else {
1692 __pskb_trim_head(skb, copy);
1693 tcp_set_skb_tso_segs(sk, skb, mss_now);
1694 }
1695 TCP_SKB_CB(skb)->seq += copy;
1696 }
1697
1698 len += copy;
1699
1700 if (len >= probe_size)
1701 break;
1702 }
1703 tcp_init_tso_segs(sk, nskb, nskb->len);
1704
1705 /* We're ready to send. If this fails, the probe will
1706 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1707 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1708 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1709 /* Decrement cwnd here because we are sending
1710 * effectively two packets. */
1711 tp->snd_cwnd--;
1712 tcp_event_new_data_sent(sk, nskb);
1713
1714 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1715 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1716 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1717
1718 return 1;
1719 }
1720
1721 return -1;
1722}
1723
1724/* This routine writes packets to the network. It advances the
1725 * send_head. This happens as incoming acks open up the remote
1726 * window for us.
1727 *
1728 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1729 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1730 * account rare use of URG, this is not a big flaw.
1731 *
1732 * Returns 1, if no segments are in flight and we have queued segments, but
1733 * cannot send anything now because of SWS or another problem.
1734 */
1735static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1736 int push_one, gfp_t gfp)
1737{
1738 struct tcp_sock *tp = tcp_sk(sk);
1739 struct sk_buff *skb;
1740 unsigned int tso_segs, sent_pkts;
1741 int cwnd_quota;
1742 int result;
1743
1744 sent_pkts = 0;
1745
1746 if (!push_one) {
1747 /* Do MTU probing. */
1748 result = tcp_mtu_probe(sk);
1749 if (!result) {
1750 return 0;
1751 } else if (result > 0) {
1752 sent_pkts = 1;
1753 }
1754 }
1755
1756 while ((skb = tcp_send_head(sk))) {
1757 unsigned int limit;
1758
1759 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1760 BUG_ON(!tso_segs);
1761
1762 cwnd_quota = tcp_cwnd_test(tp, skb);
1763 if (!cwnd_quota)
1764 break;
1765
1766 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1767 break;
1768
1769 if (tso_segs == 1) {
1770 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1771 (tcp_skb_is_last(sk, skb) ?
1772 nonagle : TCP_NAGLE_PUSH))))
1773 break;
1774 } else {
1775 if (!push_one && tcp_tso_should_defer(sk, skb))
1776 break;
1777 }
1778
1779 limit = mss_now;
1780 if (tso_segs > 1 && !tcp_urg_mode(tp))
1781 limit = tcp_mss_split_point(sk, skb, mss_now,
1782 cwnd_quota);
1783
1784 if (skb->len > limit &&
1785 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
1786 break;
1787
1788 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1789
1790 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1791 break;
1792
1793 /* Advance the send_head. This one is sent out.
1794 * This call will increment packets_out.
1795 */
1796 tcp_event_new_data_sent(sk, skb);
1797
1798 tcp_minshall_update(tp, mss_now, skb);
1799 sent_pkts++;
1800
1801 if (push_one)
1802 break;
1803 }
1804
1805 if (likely(sent_pkts)) {
1806 tcp_cwnd_validate(sk);
1807 return 0;
1808 }
1809 return !tp->packets_out && tcp_send_head(sk);
1810}
1811
1812/* Push out any pending frames which were held back due to
1813 * TCP_CORK or attempt at coalescing tiny packets.
1814 * The socket must be locked by the caller.
1815 */
1816void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1817 int nonagle)
1818{
1819 /* If we are closed, the bytes will have to remain here.
1820 * In time closedown will finish, we empty the write queue and
1821 * all will be happy.
1822 */
1823 if (unlikely(sk->sk_state == TCP_CLOSE))
1824 return;
1825
1826 if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC))
1827 tcp_check_probe_timer(sk);
1828}
1829
1830/* Send _single_ skb sitting at the send head. This function requires
1831 * true push pending frames to setup probe timer etc.
1832 */
1833void tcp_push_one(struct sock *sk, unsigned int mss_now)
1834{
1835 struct sk_buff *skb = tcp_send_head(sk);
1836
1837 BUG_ON(!skb || skb->len < mss_now);
1838
1839 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
1840}
1841
1842/* This function returns the amount that we can raise the
1843 * usable window based on the following constraints
1844 *
1845 * 1. The window can never be shrunk once it is offered (RFC 793)
1846 * 2. We limit memory per socket
1847 *
1848 * RFC 1122:
1849 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1850 * RECV.NEXT + RCV.WIN fixed until:
1851 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1852 *
1853 * i.e. don't raise the right edge of the window until you can raise
1854 * it at least MSS bytes.
1855 *
1856 * Unfortunately, the recommended algorithm breaks header prediction,
1857 * since header prediction assumes th->window stays fixed.
1858 *
1859 * Strictly speaking, keeping th->window fixed violates the receiver
1860 * side SWS prevention criteria. The problem is that under this rule
1861 * a stream of single byte packets will cause the right side of the
1862 * window to always advance by a single byte.
1863 *
1864 * Of course, if the sender implements sender side SWS prevention
1865 * then this will not be a problem.
1866 *
1867 * BSD seems to make the following compromise:
1868 *
1869 * If the free space is less than the 1/4 of the maximum
1870 * space available and the free space is less than 1/2 mss,
1871 * then set the window to 0.
1872 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1873 * Otherwise, just prevent the window from shrinking
1874 * and from being larger than the largest representable value.
1875 *
1876 * This prevents incremental opening of the window in the regime
1877 * where TCP is limited by the speed of the reader side taking
1878 * data out of the TCP receive queue. It does nothing about
1879 * those cases where the window is constrained on the sender side
1880 * because the pipeline is full.
1881 *
1882 * BSD also seems to "accidentally" limit itself to windows that are a
1883 * multiple of MSS, at least until the free space gets quite small.
1884 * This would appear to be a side effect of the mbuf implementation.
1885 * Combining these two algorithms results in the observed behavior
1886 * of having a fixed window size at almost all times.
1887 *
1888 * Below we obtain similar behavior by forcing the offered window to
1889 * a multiple of the mss when it is feasible to do so.
1890 *
1891 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1892 * Regular options like TIMESTAMP are taken into account.
1893 */
1894u32 __tcp_select_window(struct sock *sk)
1895{
1896 struct inet_connection_sock *icsk = inet_csk(sk);
1897 struct tcp_sock *tp = tcp_sk(sk);
1898 /* MSS for the peer's data. Previous versions used mss_clamp
1899 * here. I don't know if the value based on our guesses
1900 * of peer's MSS is better for the performance. It's more correct
1901 * but may be worse for the performance because of rcv_mss
1902 * fluctuations. --SAW 1998/11/1
1903 */
1904 int mss = icsk->icsk_ack.rcv_mss;
1905 int free_space = tcp_space(sk);
1906 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1907 int window;
1908
1909 if (mss > full_space)
1910 mss = full_space;
1911
1912 if (free_space < (full_space >> 1)) {
1913 icsk->icsk_ack.quick = 0;
1914
1915 if (tcp_memory_pressure)
1916 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1917 4U * tp->advmss);
1918
1919 if (free_space < mss)
1920 return 0;
1921 }
1922
1923 if (free_space > tp->rcv_ssthresh)
1924 free_space = tp->rcv_ssthresh;
1925
1926 /* Don't do rounding if we are using window scaling, since the
1927 * scaled window will not line up with the MSS boundary anyway.
1928 */
1929 window = tp->rcv_wnd;
1930 if (tp->rx_opt.rcv_wscale) {
1931 window = free_space;
1932
1933 /* Advertise enough space so that it won't get scaled away.
1934 * Import case: prevent zero window announcement if
1935 * 1<<rcv_wscale > mss.
1936 */
1937 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1938 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1939 << tp->rx_opt.rcv_wscale);
1940 } else {
1941 /* Get the largest window that is a nice multiple of mss.
1942 * Window clamp already applied above.
1943 * If our current window offering is within 1 mss of the
1944 * free space we just keep it. This prevents the divide
1945 * and multiply from happening most of the time.
1946 * We also don't do any window rounding when the free space
1947 * is too small.
1948 */
1949 if (window <= free_space - mss || window > free_space)
1950 window = (free_space / mss) * mss;
1951 else if (mss == full_space &&
1952 free_space > window + (full_space >> 1))
1953 window = free_space;
1954 }
1955
1956 return window;
1957}
1958
1959/* Collapses two adjacent SKB's during retransmission. */
1960static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
1961{
1962 struct tcp_sock *tp = tcp_sk(sk);
1963 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
1964 int skb_size, next_skb_size;
1965
1966 skb_size = skb->len;
1967 next_skb_size = next_skb->len;
1968
1969 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
1970
1971 tcp_highest_sack_combine(sk, next_skb, skb);
1972
1973 tcp_unlink_write_queue(next_skb, sk);
1974
1975 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
1976 next_skb_size);
1977
1978 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1979 skb->ip_summed = CHECKSUM_PARTIAL;
1980
1981 if (skb->ip_summed != CHECKSUM_PARTIAL)
1982 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1983
1984 /* Update sequence range on original skb. */
1985 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1986
1987 /* Merge over control information. This moves PSH/FIN etc. over */
1988 TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(next_skb)->flags;
1989
1990 /* All done, get rid of second SKB and account for it so
1991 * packet counting does not break.
1992 */
1993 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
1994
1995 /* changed transmit queue under us so clear hints */
1996 tcp_clear_retrans_hints_partial(tp);
1997 if (next_skb == tp->retransmit_skb_hint)
1998 tp->retransmit_skb_hint = skb;
1999
2000 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2001
2002 sk_wmem_free_skb(sk, next_skb);
2003}
2004
2005/* Check if coalescing SKBs is legal. */
2006static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb)
2007{
2008 if (tcp_skb_pcount(skb) > 1)
2009 return 0;
2010 /* TODO: SACK collapsing could be used to remove this condition */
2011 if (skb_shinfo(skb)->nr_frags != 0)
2012 return 0;
2013 if (skb_cloned(skb))
2014 return 0;
2015 if (skb == tcp_send_head(sk))
2016 return 0;
2017 /* Some heurestics for collapsing over SACK'd could be invented */
2018 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2019 return 0;
2020
2021 return 1;
2022}
2023
2024/* Collapse packets in the retransmit queue to make to create
2025 * less packets on the wire. This is only done on retransmission.
2026 */
2027static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2028 int space)
2029{
2030 struct tcp_sock *tp = tcp_sk(sk);
2031 struct sk_buff *skb = to, *tmp;
2032 int first = 1;
2033
2034 if (!sysctl_tcp_retrans_collapse)
2035 return;
2036 if (TCP_SKB_CB(skb)->flags & TCPHDR_SYN)
2037 return;
2038
2039 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2040 if (!tcp_can_collapse(sk, skb))
2041 break;
2042
2043 space -= skb->len;
2044
2045 if (first) {
2046 first = 0;
2047 continue;
2048 }
2049
2050 if (space < 0)
2051 break;
2052 /* Punt if not enough space exists in the first SKB for
2053 * the data in the second
2054 */
2055 if (skb->len > skb_tailroom(to))
2056 break;
2057
2058 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2059 break;
2060
2061 tcp_collapse_retrans(sk, to);
2062 }
2063}
2064
2065/* This retransmits one SKB. Policy decisions and retransmit queue
2066 * state updates are done by the caller. Returns non-zero if an
2067 * error occurred which prevented the send.
2068 */
2069int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2070{
2071 struct tcp_sock *tp = tcp_sk(sk);
2072 struct inet_connection_sock *icsk = inet_csk(sk);
2073 unsigned int cur_mss;
2074 int err;
2075
2076 /* Inconslusive MTU probe */
2077 if (icsk->icsk_mtup.probe_size) {
2078 icsk->icsk_mtup.probe_size = 0;
2079 }
2080
2081 /* Do not sent more than we queued. 1/4 is reserved for possible
2082 * copying overhead: fragmentation, tunneling, mangling etc.
2083 */
2084 if (atomic_read(&sk->sk_wmem_alloc) >
2085 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2086 return -EAGAIN;
2087
2088 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2089 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2090 BUG();
2091 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2092 return -ENOMEM;
2093 }
2094
2095 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2096 return -EHOSTUNREACH; /* Routing failure or similar. */
2097
2098 cur_mss = tcp_current_mss(sk);
2099
2100 /* If receiver has shrunk his window, and skb is out of
2101 * new window, do not retransmit it. The exception is the
2102 * case, when window is shrunk to zero. In this case
2103 * our retransmit serves as a zero window probe.
2104 */
2105 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2106 TCP_SKB_CB(skb)->seq != tp->snd_una)
2107 return -EAGAIN;
2108
2109 if (skb->len > cur_mss) {
2110 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2111 return -ENOMEM; /* We'll try again later. */
2112 } else {
2113 int oldpcount = tcp_skb_pcount(skb);
2114
2115 if (unlikely(oldpcount > 1)) {
2116 tcp_init_tso_segs(sk, skb, cur_mss);
2117 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2118 }
2119 }
2120
2121 tcp_retrans_try_collapse(sk, skb, cur_mss);
2122
2123 /* Some Solaris stacks overoptimize and ignore the FIN on a
2124 * retransmit when old data is attached. So strip it off
2125 * since it is cheap to do so and saves bytes on the network.
2126 */
2127 if (skb->len > 0 &&
2128 (TCP_SKB_CB(skb)->flags & TCPHDR_FIN) &&
2129 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2130 if (!pskb_trim(skb, 0)) {
2131 /* Reuse, even though it does some unnecessary work */
2132 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2133 TCP_SKB_CB(skb)->flags);
2134 skb->ip_summed = CHECKSUM_NONE;
2135 }
2136 }
2137
2138 /* Make a copy, if the first transmission SKB clone we made
2139 * is still in somebody's hands, else make a clone.
2140 */
2141 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2142
2143 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2144
2145 if (err == 0) {
2146 /* Update global TCP statistics. */
2147 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2148
2149 tp->total_retrans++;
2150
2151#if FASTRETRANS_DEBUG > 0
2152 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2153 if (net_ratelimit())
2154 printk(KERN_DEBUG "retrans_out leaked.\n");
2155 }
2156#endif
2157 if (!tp->retrans_out)
2158 tp->lost_retrans_low = tp->snd_nxt;
2159 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2160 tp->retrans_out += tcp_skb_pcount(skb);
2161
2162 /* Save stamp of the first retransmit. */
2163 if (!tp->retrans_stamp)
2164 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2165
2166 tp->undo_retrans += tcp_skb_pcount(skb);
2167
2168 /* snd_nxt is stored to detect loss of retransmitted segment,
2169 * see tcp_input.c tcp_sacktag_write_queue().
2170 */
2171 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2172 }
2173 return err;
2174}
2175
2176/* Check if we forward retransmits are possible in the current
2177 * window/congestion state.
2178 */
2179static int tcp_can_forward_retransmit(struct sock *sk)
2180{
2181 const struct inet_connection_sock *icsk = inet_csk(sk);
2182 struct tcp_sock *tp = tcp_sk(sk);
2183
2184 /* Forward retransmissions are possible only during Recovery. */
2185 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2186 return 0;
2187
2188 /* No forward retransmissions in Reno are possible. */
2189 if (tcp_is_reno(tp))
2190 return 0;
2191
2192 /* Yeah, we have to make difficult choice between forward transmission
2193 * and retransmission... Both ways have their merits...
2194 *
2195 * For now we do not retransmit anything, while we have some new
2196 * segments to send. In the other cases, follow rule 3 for
2197 * NextSeg() specified in RFC3517.
2198 */
2199
2200 if (tcp_may_send_now(sk))
2201 return 0;
2202
2203 return 1;
2204}
2205
2206/* This gets called after a retransmit timeout, and the initially
2207 * retransmitted data is acknowledged. It tries to continue
2208 * resending the rest of the retransmit queue, until either
2209 * we've sent it all or the congestion window limit is reached.
2210 * If doing SACK, the first ACK which comes back for a timeout
2211 * based retransmit packet might feed us FACK information again.
2212 * If so, we use it to avoid unnecessarily retransmissions.
2213 */
2214void tcp_xmit_retransmit_queue(struct sock *sk)
2215{
2216 const struct inet_connection_sock *icsk = inet_csk(sk);
2217 struct tcp_sock *tp = tcp_sk(sk);
2218 struct sk_buff *skb;
2219 struct sk_buff *hole = NULL;
2220 u32 last_lost;
2221 int mib_idx;
2222 int fwd_rexmitting = 0;
2223
2224 if (!tp->packets_out)
2225 return;
2226
2227 if (!tp->lost_out)
2228 tp->retransmit_high = tp->snd_una;
2229
2230 if (tp->retransmit_skb_hint) {
2231 skb = tp->retransmit_skb_hint;
2232 last_lost = TCP_SKB_CB(skb)->end_seq;
2233 if (after(last_lost, tp->retransmit_high))
2234 last_lost = tp->retransmit_high;
2235 } else {
2236 skb = tcp_write_queue_head(sk);
2237 last_lost = tp->snd_una;
2238 }
2239
2240 tcp_for_write_queue_from(skb, sk) {
2241 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2242
2243 if (skb == tcp_send_head(sk))
2244 break;
2245 /* we could do better than to assign each time */
2246 if (hole == NULL)
2247 tp->retransmit_skb_hint = skb;
2248
2249 /* Assume this retransmit will generate
2250 * only one packet for congestion window
2251 * calculation purposes. This works because
2252 * tcp_retransmit_skb() will chop up the
2253 * packet to be MSS sized and all the
2254 * packet counting works out.
2255 */
2256 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2257 return;
2258
2259 if (fwd_rexmitting) {
2260begin_fwd:
2261 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2262 break;
2263 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2264
2265 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2266 tp->retransmit_high = last_lost;
2267 if (!tcp_can_forward_retransmit(sk))
2268 break;
2269 /* Backtrack if necessary to non-L'ed skb */
2270 if (hole != NULL) {
2271 skb = hole;
2272 hole = NULL;
2273 }
2274 fwd_rexmitting = 1;
2275 goto begin_fwd;
2276
2277 } else if (!(sacked & TCPCB_LOST)) {
2278 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2279 hole = skb;
2280 continue;
2281
2282 } else {
2283 last_lost = TCP_SKB_CB(skb)->end_seq;
2284 if (icsk->icsk_ca_state != TCP_CA_Loss)
2285 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2286 else
2287 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2288 }
2289
2290 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2291 continue;
2292
2293 if (tcp_retransmit_skb(sk, skb))
2294 return;
2295 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2296
2297 if (skb == tcp_write_queue_head(sk))
2298 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2299 inet_csk(sk)->icsk_rto,
2300 TCP_RTO_MAX);
2301 }
2302}
2303
2304/* Send a fin. The caller locks the socket for us. This cannot be
2305 * allowed to fail queueing a FIN frame under any circumstances.
2306 */
2307void tcp_send_fin(struct sock *sk)
2308{
2309 struct tcp_sock *tp = tcp_sk(sk);
2310 struct sk_buff *skb = tcp_write_queue_tail(sk);
2311 int mss_now;
2312
2313 /* Optimization, tack on the FIN if we have a queue of
2314 * unsent frames. But be careful about outgoing SACKS
2315 * and IP options.
2316 */
2317 mss_now = tcp_current_mss(sk);
2318
2319 if (tcp_send_head(sk) != NULL) {
2320 TCP_SKB_CB(skb)->flags |= TCPHDR_FIN;
2321 TCP_SKB_CB(skb)->end_seq++;
2322 tp->write_seq++;
2323 } else {
2324 /* Socket is locked, keep trying until memory is available. */
2325 for (;;) {
2326 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2327 sk->sk_allocation);
2328 if (skb)
2329 break;
2330 yield();
2331 }
2332
2333 /* Reserve space for headers and prepare control bits. */
2334 skb_reserve(skb, MAX_TCP_HEADER);
2335 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2336 tcp_init_nondata_skb(skb, tp->write_seq,
2337 TCPHDR_ACK | TCPHDR_FIN);
2338 tcp_queue_skb(sk, skb);
2339 }
2340 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2341}
2342
2343/* We get here when a process closes a file descriptor (either due to
2344 * an explicit close() or as a byproduct of exit()'ing) and there
2345 * was unread data in the receive queue. This behavior is recommended
2346 * by RFC 2525, section 2.17. -DaveM
2347 */
2348void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2349{
2350 struct sk_buff *skb;
2351
2352 /* NOTE: No TCP options attached and we never retransmit this. */
2353 skb = alloc_skb(MAX_TCP_HEADER, priority);
2354 if (!skb) {
2355 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2356 return;
2357 }
2358
2359 /* Reserve space for headers and prepare control bits. */
2360 skb_reserve(skb, MAX_TCP_HEADER);
2361 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2362 TCPHDR_ACK | TCPHDR_RST);
2363 /* Send it off. */
2364 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2365 if (tcp_transmit_skb(sk, skb, 0, priority))
2366 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2367
2368 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2369}
2370
2371/* Send a crossed SYN-ACK during socket establishment.
2372 * WARNING: This routine must only be called when we have already sent
2373 * a SYN packet that crossed the incoming SYN that caused this routine
2374 * to get called. If this assumption fails then the initial rcv_wnd
2375 * and rcv_wscale values will not be correct.
2376 */
2377int tcp_send_synack(struct sock *sk)
2378{
2379 struct sk_buff *skb;
2380
2381 skb = tcp_write_queue_head(sk);
2382 if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPHDR_SYN)) {
2383 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2384 return -EFAULT;
2385 }
2386 if (!(TCP_SKB_CB(skb)->flags & TCPHDR_ACK)) {
2387 if (skb_cloned(skb)) {
2388 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2389 if (nskb == NULL)
2390 return -ENOMEM;
2391 tcp_unlink_write_queue(skb, sk);
2392 skb_header_release(nskb);
2393 __tcp_add_write_queue_head(sk, nskb);
2394 sk_wmem_free_skb(sk, skb);
2395 sk->sk_wmem_queued += nskb->truesize;
2396 sk_mem_charge(sk, nskb->truesize);
2397 skb = nskb;
2398 }
2399
2400 TCP_SKB_CB(skb)->flags |= TCPHDR_ACK;
2401 TCP_ECN_send_synack(tcp_sk(sk), skb);
2402 }
2403 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2404 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2405}
2406
2407/* Prepare a SYN-ACK. */
2408struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2409 struct request_sock *req,
2410 struct request_values *rvp)
2411{
2412 struct tcp_out_options opts;
2413 struct tcp_extend_values *xvp = tcp_xv(rvp);
2414 struct inet_request_sock *ireq = inet_rsk(req);
2415 struct tcp_sock *tp = tcp_sk(sk);
2416 const struct tcp_cookie_values *cvp = tp->cookie_values;
2417 struct tcphdr *th;
2418 struct sk_buff *skb;
2419 struct tcp_md5sig_key *md5;
2420 int tcp_header_size;
2421 int mss;
2422 int s_data_desired = 0;
2423
2424 if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired)
2425 s_data_desired = cvp->s_data_desired;
2426 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15 + s_data_desired, 1, GFP_ATOMIC);
2427 if (skb == NULL)
2428 return NULL;
2429
2430 /* Reserve space for headers. */
2431 skb_reserve(skb, MAX_TCP_HEADER);
2432
2433 skb_dst_set(skb, dst_clone(dst));
2434
2435 mss = dst_metric_advmss(dst);
2436 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2437 mss = tp->rx_opt.user_mss;
2438
2439 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2440 __u8 rcv_wscale;
2441 /* Set this up on the first call only */
2442 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2443
2444 /* limit the window selection if the user enforce a smaller rx buffer */
2445 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2446 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2447 req->window_clamp = tcp_full_space(sk);
2448
2449 /* tcp_full_space because it is guaranteed to be the first packet */
2450 tcp_select_initial_window(tcp_full_space(sk),
2451 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2452 &req->rcv_wnd,
2453 &req->window_clamp,
2454 ireq->wscale_ok,
2455 &rcv_wscale,
2456 dst_metric(dst, RTAX_INITRWND));
2457 ireq->rcv_wscale = rcv_wscale;
2458 }
2459
2460 memset(&opts, 0, sizeof(opts));
2461#ifdef CONFIG_SYN_COOKIES
2462 if (unlikely(req->cookie_ts))
2463 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2464 else
2465#endif
2466 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2467 tcp_header_size = tcp_synack_options(sk, req, mss,
2468 skb, &opts, &md5, xvp)
2469 + sizeof(*th);
2470
2471 skb_push(skb, tcp_header_size);
2472 skb_reset_transport_header(skb);
2473
2474 th = tcp_hdr(skb);
2475 memset(th, 0, sizeof(struct tcphdr));
2476 th->syn = 1;
2477 th->ack = 1;
2478 TCP_ECN_make_synack(req, th);
2479 th->source = ireq->loc_port;
2480 th->dest = ireq->rmt_port;
2481 /* Setting of flags are superfluous here for callers (and ECE is
2482 * not even correctly set)
2483 */
2484 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2485 TCPHDR_SYN | TCPHDR_ACK);
2486
2487 if (OPTION_COOKIE_EXTENSION & opts.options) {
2488 if (s_data_desired) {
2489 u8 *buf = skb_put(skb, s_data_desired);
2490
2491 /* copy data directly from the listening socket. */
2492 memcpy(buf, cvp->s_data_payload, s_data_desired);
2493 TCP_SKB_CB(skb)->end_seq += s_data_desired;
2494 }
2495
2496 if (opts.hash_size > 0) {
2497 __u32 workspace[SHA_WORKSPACE_WORDS];
2498 u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS];
2499 u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1];
2500
2501 /* Secret recipe depends on the Timestamp, (future)
2502 * Sequence and Acknowledgment Numbers, Initiator
2503 * Cookie, and others handled by IP variant caller.
2504 */
2505 *tail-- ^= opts.tsval;
2506 *tail-- ^= tcp_rsk(req)->rcv_isn + 1;
2507 *tail-- ^= TCP_SKB_CB(skb)->seq + 1;
2508
2509 /* recommended */
2510 *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source);
2511 *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */
2512
2513 sha_transform((__u32 *)&xvp->cookie_bakery[0],
2514 (char *)mess,
2515 &workspace[0]);
2516 opts.hash_location =
2517 (__u8 *)&xvp->cookie_bakery[0];
2518 }
2519 }
2520
2521 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2522 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2523
2524 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2525 th->window = htons(min(req->rcv_wnd, 65535U));
2526 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2527 th->doff = (tcp_header_size >> 2);
2528 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2529
2530#ifdef CONFIG_TCP_MD5SIG
2531 /* Okay, we have all we need - do the md5 hash if needed */
2532 if (md5) {
2533 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2534 md5, NULL, req, skb);
2535 }
2536#endif
2537
2538 return skb;
2539}
2540EXPORT_SYMBOL(tcp_make_synack);
2541
2542/* Do all connect socket setups that can be done AF independent. */
2543static void tcp_connect_init(struct sock *sk)
2544{
2545 struct dst_entry *dst = __sk_dst_get(sk);
2546 struct tcp_sock *tp = tcp_sk(sk);
2547 __u8 rcv_wscale;
2548
2549 /* We'll fix this up when we get a response from the other end.
2550 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2551 */
2552 tp->tcp_header_len = sizeof(struct tcphdr) +
2553 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2554
2555#ifdef CONFIG_TCP_MD5SIG
2556 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2557 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2558#endif
2559
2560 /* If user gave his TCP_MAXSEG, record it to clamp */
2561 if (tp->rx_opt.user_mss)
2562 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2563 tp->max_window = 0;
2564 tcp_mtup_init(sk);
2565 tcp_sync_mss(sk, dst_mtu(dst));
2566
2567 if (!tp->window_clamp)
2568 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2569 tp->advmss = dst_metric_advmss(dst);
2570 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2571 tp->advmss = tp->rx_opt.user_mss;
2572
2573 tcp_initialize_rcv_mss(sk);
2574
2575 /* limit the window selection if the user enforce a smaller rx buffer */
2576 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2577 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2578 tp->window_clamp = tcp_full_space(sk);
2579
2580 tcp_select_initial_window(tcp_full_space(sk),
2581 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2582 &tp->rcv_wnd,
2583 &tp->window_clamp,
2584 sysctl_tcp_window_scaling,
2585 &rcv_wscale,
2586 dst_metric(dst, RTAX_INITRWND));
2587
2588 tp->rx_opt.rcv_wscale = rcv_wscale;
2589 tp->rcv_ssthresh = tp->rcv_wnd;
2590
2591 sk->sk_err = 0;
2592 sock_reset_flag(sk, SOCK_DONE);
2593 tp->snd_wnd = 0;
2594 tcp_init_wl(tp, 0);
2595 tp->snd_una = tp->write_seq;
2596 tp->snd_sml = tp->write_seq;
2597 tp->snd_up = tp->write_seq;
2598 tp->rcv_nxt = 0;
2599 tp->rcv_wup = 0;
2600 tp->copied_seq = 0;
2601
2602 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2603 inet_csk(sk)->icsk_retransmits = 0;
2604 tcp_clear_retrans(tp);
2605}
2606
2607/* Build a SYN and send it off. */
2608int tcp_connect(struct sock *sk)
2609{
2610 struct tcp_sock *tp = tcp_sk(sk);
2611 struct sk_buff *buff;
2612 int err;
2613
2614 tcp_connect_init(sk);
2615
2616 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2617 if (unlikely(buff == NULL))
2618 return -ENOBUFS;
2619
2620 /* Reserve space for headers. */
2621 skb_reserve(buff, MAX_TCP_HEADER);
2622
2623 tp->snd_nxt = tp->write_seq;
2624 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
2625 TCP_ECN_send_syn(sk, buff);
2626
2627 /* Send it off. */
2628 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2629 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2630 skb_header_release(buff);
2631 __tcp_add_write_queue_tail(sk, buff);
2632 sk->sk_wmem_queued += buff->truesize;
2633 sk_mem_charge(sk, buff->truesize);
2634 tp->packets_out += tcp_skb_pcount(buff);
2635 err = tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2636 if (err == -ECONNREFUSED)
2637 return err;
2638
2639 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2640 * in order to make this packet get counted in tcpOutSegs.
2641 */
2642 tp->snd_nxt = tp->write_seq;
2643 tp->pushed_seq = tp->write_seq;
2644 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2645
2646 /* Timer for repeating the SYN until an answer. */
2647 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2648 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2649 return 0;
2650}
2651EXPORT_SYMBOL(tcp_connect);
2652
2653/* Send out a delayed ack, the caller does the policy checking
2654 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2655 * for details.
2656 */
2657void tcp_send_delayed_ack(struct sock *sk)
2658{
2659 struct inet_connection_sock *icsk = inet_csk(sk);
2660 int ato = icsk->icsk_ack.ato;
2661 unsigned long timeout;
2662
2663 if (ato > TCP_DELACK_MIN) {
2664 const struct tcp_sock *tp = tcp_sk(sk);
2665 int max_ato = HZ / 2;
2666
2667 if (icsk->icsk_ack.pingpong ||
2668 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2669 max_ato = TCP_DELACK_MAX;
2670
2671 /* Slow path, intersegment interval is "high". */
2672
2673 /* If some rtt estimate is known, use it to bound delayed ack.
2674 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2675 * directly.
2676 */
2677 if (tp->srtt) {
2678 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
2679
2680 if (rtt < max_ato)
2681 max_ato = rtt;
2682 }
2683
2684 ato = min(ato, max_ato);
2685 }
2686
2687 /* Stay within the limit we were given */
2688 timeout = jiffies + ato;
2689
2690 /* Use new timeout only if there wasn't a older one earlier. */
2691 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2692 /* If delack timer was blocked or is about to expire,
2693 * send ACK now.
2694 */
2695 if (icsk->icsk_ack.blocked ||
2696 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2697 tcp_send_ack(sk);
2698 return;
2699 }
2700
2701 if (!time_before(timeout, icsk->icsk_ack.timeout))
2702 timeout = icsk->icsk_ack.timeout;
2703 }
2704 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2705 icsk->icsk_ack.timeout = timeout;
2706 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2707}
2708
2709/* This routine sends an ack and also updates the window. */
2710void tcp_send_ack(struct sock *sk)
2711{
2712 struct sk_buff *buff;
2713
2714 /* If we have been reset, we may not send again. */
2715 if (sk->sk_state == TCP_CLOSE)
2716 return;
2717
2718 /* We are not putting this on the write queue, so
2719 * tcp_transmit_skb() will set the ownership to this
2720 * sock.
2721 */
2722 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2723 if (buff == NULL) {
2724 inet_csk_schedule_ack(sk);
2725 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2726 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2727 TCP_DELACK_MAX, TCP_RTO_MAX);
2728 return;
2729 }
2730
2731 /* Reserve space for headers and prepare control bits. */
2732 skb_reserve(buff, MAX_TCP_HEADER);
2733 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
2734
2735 /* Send it off, this clears delayed acks for us. */
2736 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2737 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2738}
2739
2740/* This routine sends a packet with an out of date sequence
2741 * number. It assumes the other end will try to ack it.
2742 *
2743 * Question: what should we make while urgent mode?
2744 * 4.4BSD forces sending single byte of data. We cannot send
2745 * out of window data, because we have SND.NXT==SND.MAX...
2746 *
2747 * Current solution: to send TWO zero-length segments in urgent mode:
2748 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2749 * out-of-date with SND.UNA-1 to probe window.
2750 */
2751static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2752{
2753 struct tcp_sock *tp = tcp_sk(sk);
2754 struct sk_buff *skb;
2755
2756 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2757 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2758 if (skb == NULL)
2759 return -1;
2760
2761 /* Reserve space for headers and set control bits. */
2762 skb_reserve(skb, MAX_TCP_HEADER);
2763 /* Use a previous sequence. This should cause the other
2764 * end to send an ack. Don't queue or clone SKB, just
2765 * send it.
2766 */
2767 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
2768 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2769 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2770}
2771
2772/* Initiate keepalive or window probe from timer. */
2773int tcp_write_wakeup(struct sock *sk)
2774{
2775 struct tcp_sock *tp = tcp_sk(sk);
2776 struct sk_buff *skb;
2777
2778 if (sk->sk_state == TCP_CLOSE)
2779 return -1;
2780
2781 if ((skb = tcp_send_head(sk)) != NULL &&
2782 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2783 int err;
2784 unsigned int mss = tcp_current_mss(sk);
2785 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2786
2787 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2788 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2789
2790 /* We are probing the opening of a window
2791 * but the window size is != 0
2792 * must have been a result SWS avoidance ( sender )
2793 */
2794 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2795 skb->len > mss) {
2796 seg_size = min(seg_size, mss);
2797 TCP_SKB_CB(skb)->flags |= TCPHDR_PSH;
2798 if (tcp_fragment(sk, skb, seg_size, mss))
2799 return -1;
2800 } else if (!tcp_skb_pcount(skb))
2801 tcp_set_skb_tso_segs(sk, skb, mss);
2802
2803 TCP_SKB_CB(skb)->flags |= TCPHDR_PSH;
2804 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2805 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2806 if (!err)
2807 tcp_event_new_data_sent(sk, skb);
2808 return err;
2809 } else {
2810 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
2811 tcp_xmit_probe_skb(sk, 1);
2812 return tcp_xmit_probe_skb(sk, 0);
2813 }
2814}
2815
2816/* A window probe timeout has occurred. If window is not closed send
2817 * a partial packet else a zero probe.
2818 */
2819void tcp_send_probe0(struct sock *sk)
2820{
2821 struct inet_connection_sock *icsk = inet_csk(sk);
2822 struct tcp_sock *tp = tcp_sk(sk);
2823 int err;
2824
2825 err = tcp_write_wakeup(sk);
2826
2827 if (tp->packets_out || !tcp_send_head(sk)) {
2828 /* Cancel probe timer, if it is not required. */
2829 icsk->icsk_probes_out = 0;
2830 icsk->icsk_backoff = 0;
2831 return;
2832 }
2833
2834 if (err <= 0) {
2835 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2836 icsk->icsk_backoff++;
2837 icsk->icsk_probes_out++;
2838 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2839 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2840 TCP_RTO_MAX);
2841 } else {
2842 /* If packet was not sent due to local congestion,
2843 * do not backoff and do not remember icsk_probes_out.
2844 * Let local senders to fight for local resources.
2845 *
2846 * Use accumulated backoff yet.
2847 */
2848 if (!icsk->icsk_probes_out)
2849 icsk->icsk_probes_out = 1;
2850 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2851 min(icsk->icsk_rto << icsk->icsk_backoff,
2852 TCP_RESOURCE_PROBE_INTERVAL),
2853 TCP_RTO_MAX);
2854 }
2855}