Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
9 * Authors: Ross Biro
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
20 */
21
22/*
23 * Changes: Pedro Roque : Retransmit queue handled by TCP.
24 * : Fragmentation on mtu decrease
25 * : Segment collapse on retransmit
26 * : AF independence
27 *
28 * Linus Torvalds : send_delayed_ack
29 * David S. Miller : Charge memory using the right skb
30 * during syn/ack processing.
31 * David S. Miller : Output engine completely rewritten.
32 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
33 * Cacophonix Gaul : draft-minshall-nagle-01
34 * J Hadi Salim : ECN support
35 *
36 */
37
38#define pr_fmt(fmt) "TCP: " fmt
39
40#include <net/tcp.h>
41#include <net/mptcp.h>
42
43#include <linux/compiler.h>
44#include <linux/gfp.h>
45#include <linux/module.h>
46#include <linux/static_key.h>
47
48#include <trace/events/tcp.h>
49
50/* Refresh clocks of a TCP socket,
51 * ensuring monotically increasing values.
52 */
53void tcp_mstamp_refresh(struct tcp_sock *tp)
54{
55 u64 val = tcp_clock_ns();
56
57 tp->tcp_clock_cache = val;
58 tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC);
59}
60
61static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
62 int push_one, gfp_t gfp);
63
64/* Account for new data that has been sent to the network. */
65static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
66{
67 struct inet_connection_sock *icsk = inet_csk(sk);
68 struct tcp_sock *tp = tcp_sk(sk);
69 unsigned int prior_packets = tp->packets_out;
70
71 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq);
72
73 __skb_unlink(skb, &sk->sk_write_queue);
74 tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
75
76 if (tp->highest_sack == NULL)
77 tp->highest_sack = skb;
78
79 tp->packets_out += tcp_skb_pcount(skb);
80 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
81 tcp_rearm_rto(sk);
82
83 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
84 tcp_skb_pcount(skb));
85}
86
87/* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
88 * window scaling factor due to loss of precision.
89 * If window has been shrunk, what should we make? It is not clear at all.
90 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
91 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
92 * invalid. OK, let's make this for now:
93 */
94static inline __u32 tcp_acceptable_seq(const struct sock *sk)
95{
96 const struct tcp_sock *tp = tcp_sk(sk);
97
98 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
99 (tp->rx_opt.wscale_ok &&
100 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
101 return tp->snd_nxt;
102 else
103 return tcp_wnd_end(tp);
104}
105
106/* Calculate mss to advertise in SYN segment.
107 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
108 *
109 * 1. It is independent of path mtu.
110 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
111 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
112 * attached devices, because some buggy hosts are confused by
113 * large MSS.
114 * 4. We do not make 3, we advertise MSS, calculated from first
115 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
116 * This may be overridden via information stored in routing table.
117 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
118 * probably even Jumbo".
119 */
120static __u16 tcp_advertise_mss(struct sock *sk)
121{
122 struct tcp_sock *tp = tcp_sk(sk);
123 const struct dst_entry *dst = __sk_dst_get(sk);
124 int mss = tp->advmss;
125
126 if (dst) {
127 unsigned int metric = dst_metric_advmss(dst);
128
129 if (metric < mss) {
130 mss = metric;
131 tp->advmss = mss;
132 }
133 }
134
135 return (__u16)mss;
136}
137
138/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
139 * This is the first part of cwnd validation mechanism.
140 */
141void tcp_cwnd_restart(struct sock *sk, s32 delta)
142{
143 struct tcp_sock *tp = tcp_sk(sk);
144 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
145 u32 cwnd = tp->snd_cwnd;
146
147 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
148
149 tp->snd_ssthresh = tcp_current_ssthresh(sk);
150 restart_cwnd = min(restart_cwnd, cwnd);
151
152 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
153 cwnd >>= 1;
154 tp->snd_cwnd = max(cwnd, restart_cwnd);
155 tp->snd_cwnd_stamp = tcp_jiffies32;
156 tp->snd_cwnd_used = 0;
157}
158
159/* Congestion state accounting after a packet has been sent. */
160static void tcp_event_data_sent(struct tcp_sock *tp,
161 struct sock *sk)
162{
163 struct inet_connection_sock *icsk = inet_csk(sk);
164 const u32 now = tcp_jiffies32;
165
166 if (tcp_packets_in_flight(tp) == 0)
167 tcp_ca_event(sk, CA_EVENT_TX_START);
168
169 /* If this is the first data packet sent in response to the
170 * previous received data,
171 * and it is a reply for ato after last received packet,
172 * increase pingpong count.
173 */
174 if (before(tp->lsndtime, icsk->icsk_ack.lrcvtime) &&
175 (u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
176 inet_csk_inc_pingpong_cnt(sk);
177
178 tp->lsndtime = now;
179}
180
181/* Account for an ACK we sent. */
182static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
183 u32 rcv_nxt)
184{
185 struct tcp_sock *tp = tcp_sk(sk);
186
187 if (unlikely(tp->compressed_ack)) {
188 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
189 tp->compressed_ack);
190 tp->compressed_ack = 0;
191 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
192 __sock_put(sk);
193 }
194
195 if (unlikely(rcv_nxt != tp->rcv_nxt))
196 return; /* Special ACK sent by DCTCP to reflect ECN */
197 tcp_dec_quickack_mode(sk, pkts);
198 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
199}
200
201/* Determine a window scaling and initial window to offer.
202 * Based on the assumption that the given amount of space
203 * will be offered. Store the results in the tp structure.
204 * NOTE: for smooth operation initial space offering should
205 * be a multiple of mss if possible. We assume here that mss >= 1.
206 * This MUST be enforced by all callers.
207 */
208void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
209 __u32 *rcv_wnd, __u32 *window_clamp,
210 int wscale_ok, __u8 *rcv_wscale,
211 __u32 init_rcv_wnd)
212{
213 unsigned int space = (__space < 0 ? 0 : __space);
214
215 /* If no clamp set the clamp to the max possible scaled window */
216 if (*window_clamp == 0)
217 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
218 space = min(*window_clamp, space);
219
220 /* Quantize space offering to a multiple of mss if possible. */
221 if (space > mss)
222 space = rounddown(space, mss);
223
224 /* NOTE: offering an initial window larger than 32767
225 * will break some buggy TCP stacks. If the admin tells us
226 * it is likely we could be speaking with such a buggy stack
227 * we will truncate our initial window offering to 32K-1
228 * unless the remote has sent us a window scaling option,
229 * which we interpret as a sign the remote TCP is not
230 * misinterpreting the window field as a signed quantity.
231 */
232 if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
233 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
234 else
235 (*rcv_wnd) = min_t(u32, space, U16_MAX);
236
237 if (init_rcv_wnd)
238 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
239
240 *rcv_wscale = 0;
241 if (wscale_ok) {
242 /* Set window scaling on max possible window */
243 space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
244 space = max_t(u32, space, sysctl_rmem_max);
245 space = min_t(u32, space, *window_clamp);
246 *rcv_wscale = clamp_t(int, ilog2(space) - 15,
247 0, TCP_MAX_WSCALE);
248 }
249 /* Set the clamp no higher than max representable value */
250 (*window_clamp) = min_t(__u32, U16_MAX << (*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 old_win = tp->rcv_wnd;
263 u32 cur_win = tcp_receive_window(tp);
264 u32 new_win = __tcp_select_window(sk);
265
266 /* Never shrink the offered window */
267 if (new_win < cur_win) {
268 /* Danger Will Robinson!
269 * Don't update rcv_wup/rcv_wnd here or else
270 * we will not be able to advertise a zero
271 * window in time. --DaveM
272 *
273 * Relax Will Robinson.
274 */
275 if (new_win == 0)
276 NET_INC_STATS(sock_net(sk),
277 LINUX_MIB_TCPWANTZEROWINDOWADV);
278 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
279 }
280 tp->rcv_wnd = new_win;
281 tp->rcv_wup = tp->rcv_nxt;
282
283 /* Make sure we do not exceed the maximum possible
284 * scaled window.
285 */
286 if (!tp->rx_opt.rcv_wscale &&
287 sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
288 new_win = min(new_win, MAX_TCP_WINDOW);
289 else
290 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
291
292 /* RFC1323 scaling applied */
293 new_win >>= tp->rx_opt.rcv_wscale;
294
295 /* If we advertise zero window, disable fast path. */
296 if (new_win == 0) {
297 tp->pred_flags = 0;
298 if (old_win)
299 NET_INC_STATS(sock_net(sk),
300 LINUX_MIB_TCPTOZEROWINDOWADV);
301 } else if (old_win == 0) {
302 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
303 }
304
305 return new_win;
306}
307
308/* Packet ECN state for a SYN-ACK */
309static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
310{
311 const struct tcp_sock *tp = tcp_sk(sk);
312
313 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
314 if (!(tp->ecn_flags & TCP_ECN_OK))
315 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
316 else if (tcp_ca_needs_ecn(sk) ||
317 tcp_bpf_ca_needs_ecn(sk))
318 INET_ECN_xmit(sk);
319}
320
321/* Packet ECN state for a SYN. */
322static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
323{
324 struct tcp_sock *tp = tcp_sk(sk);
325 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
326 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
327 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
328
329 if (!use_ecn) {
330 const struct dst_entry *dst = __sk_dst_get(sk);
331
332 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
333 use_ecn = true;
334 }
335
336 tp->ecn_flags = 0;
337
338 if (use_ecn) {
339 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
340 tp->ecn_flags = TCP_ECN_OK;
341 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
342 INET_ECN_xmit(sk);
343 }
344}
345
346static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
347{
348 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
349 /* tp->ecn_flags are cleared at a later point in time when
350 * SYN ACK is ultimatively being received.
351 */
352 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
353}
354
355static void
356tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
357{
358 if (inet_rsk(req)->ecn_ok)
359 th->ece = 1;
360}
361
362/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
363 * be sent.
364 */
365static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
366 struct tcphdr *th, int tcp_header_len)
367{
368 struct tcp_sock *tp = tcp_sk(sk);
369
370 if (tp->ecn_flags & TCP_ECN_OK) {
371 /* Not-retransmitted data segment: set ECT and inject CWR. */
372 if (skb->len != tcp_header_len &&
373 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
374 INET_ECN_xmit(sk);
375 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
376 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
377 th->cwr = 1;
378 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
379 }
380 } else if (!tcp_ca_needs_ecn(sk)) {
381 /* ACK or retransmitted segment: clear ECT|CE */
382 INET_ECN_dontxmit(sk);
383 }
384 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
385 th->ece = 1;
386 }
387}
388
389/* Constructs common control bits of non-data skb. If SYN/FIN is present,
390 * auto increment end seqno.
391 */
392static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
393{
394 skb->ip_summed = CHECKSUM_PARTIAL;
395
396 TCP_SKB_CB(skb)->tcp_flags = flags;
397 TCP_SKB_CB(skb)->sacked = 0;
398
399 tcp_skb_pcount_set(skb, 1);
400
401 TCP_SKB_CB(skb)->seq = seq;
402 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
403 seq++;
404 TCP_SKB_CB(skb)->end_seq = seq;
405}
406
407static inline bool tcp_urg_mode(const struct tcp_sock *tp)
408{
409 return tp->snd_una != tp->snd_up;
410}
411
412#define OPTION_SACK_ADVERTISE (1 << 0)
413#define OPTION_TS (1 << 1)
414#define OPTION_MD5 (1 << 2)
415#define OPTION_WSCALE (1 << 3)
416#define OPTION_FAST_OPEN_COOKIE (1 << 8)
417#define OPTION_SMC (1 << 9)
418#define OPTION_MPTCP (1 << 10)
419
420static void smc_options_write(__be32 *ptr, u16 *options)
421{
422#if IS_ENABLED(CONFIG_SMC)
423 if (static_branch_unlikely(&tcp_have_smc)) {
424 if (unlikely(OPTION_SMC & *options)) {
425 *ptr++ = htonl((TCPOPT_NOP << 24) |
426 (TCPOPT_NOP << 16) |
427 (TCPOPT_EXP << 8) |
428 (TCPOLEN_EXP_SMC_BASE));
429 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
430 }
431 }
432#endif
433}
434
435struct tcp_out_options {
436 u16 options; /* bit field of OPTION_* */
437 u16 mss; /* 0 to disable */
438 u8 ws; /* window scale, 0 to disable */
439 u8 num_sack_blocks; /* number of SACK blocks to include */
440 u8 hash_size; /* bytes in hash_location */
441 __u8 *hash_location; /* temporary pointer, overloaded */
442 __u32 tsval, tsecr; /* need to include OPTION_TS */
443 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
444 struct mptcp_out_options mptcp;
445};
446
447static void mptcp_options_write(__be32 *ptr, struct tcp_out_options *opts)
448{
449#if IS_ENABLED(CONFIG_MPTCP)
450 if (unlikely(OPTION_MPTCP & opts->options))
451 mptcp_write_options(ptr, &opts->mptcp);
452#endif
453}
454
455/* Write previously computed TCP options to the packet.
456 *
457 * Beware: Something in the Internet is very sensitive to the ordering of
458 * TCP options, we learned this through the hard way, so be careful here.
459 * Luckily we can at least blame others for their non-compliance but from
460 * inter-operability perspective it seems that we're somewhat stuck with
461 * the ordering which we have been using if we want to keep working with
462 * those broken things (not that it currently hurts anybody as there isn't
463 * particular reason why the ordering would need to be changed).
464 *
465 * At least SACK_PERM as the first option is known to lead to a disaster
466 * (but it may well be that other scenarios fail similarly).
467 */
468static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
469 struct tcp_out_options *opts)
470{
471 u16 options = opts->options; /* mungable copy */
472
473 if (unlikely(OPTION_MD5 & options)) {
474 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
475 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
476 /* overload cookie hash location */
477 opts->hash_location = (__u8 *)ptr;
478 ptr += 4;
479 }
480
481 if (unlikely(opts->mss)) {
482 *ptr++ = htonl((TCPOPT_MSS << 24) |
483 (TCPOLEN_MSS << 16) |
484 opts->mss);
485 }
486
487 if (likely(OPTION_TS & options)) {
488 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
489 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
490 (TCPOLEN_SACK_PERM << 16) |
491 (TCPOPT_TIMESTAMP << 8) |
492 TCPOLEN_TIMESTAMP);
493 options &= ~OPTION_SACK_ADVERTISE;
494 } else {
495 *ptr++ = htonl((TCPOPT_NOP << 24) |
496 (TCPOPT_NOP << 16) |
497 (TCPOPT_TIMESTAMP << 8) |
498 TCPOLEN_TIMESTAMP);
499 }
500 *ptr++ = htonl(opts->tsval);
501 *ptr++ = htonl(opts->tsecr);
502 }
503
504 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
505 *ptr++ = htonl((TCPOPT_NOP << 24) |
506 (TCPOPT_NOP << 16) |
507 (TCPOPT_SACK_PERM << 8) |
508 TCPOLEN_SACK_PERM);
509 }
510
511 if (unlikely(OPTION_WSCALE & options)) {
512 *ptr++ = htonl((TCPOPT_NOP << 24) |
513 (TCPOPT_WINDOW << 16) |
514 (TCPOLEN_WINDOW << 8) |
515 opts->ws);
516 }
517
518 if (unlikely(opts->num_sack_blocks)) {
519 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
520 tp->duplicate_sack : tp->selective_acks;
521 int this_sack;
522
523 *ptr++ = htonl((TCPOPT_NOP << 24) |
524 (TCPOPT_NOP << 16) |
525 (TCPOPT_SACK << 8) |
526 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
527 TCPOLEN_SACK_PERBLOCK)));
528
529 for (this_sack = 0; this_sack < opts->num_sack_blocks;
530 ++this_sack) {
531 *ptr++ = htonl(sp[this_sack].start_seq);
532 *ptr++ = htonl(sp[this_sack].end_seq);
533 }
534
535 tp->rx_opt.dsack = 0;
536 }
537
538 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
539 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
540 u8 *p = (u8 *)ptr;
541 u32 len; /* Fast Open option length */
542
543 if (foc->exp) {
544 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
545 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
546 TCPOPT_FASTOPEN_MAGIC);
547 p += TCPOLEN_EXP_FASTOPEN_BASE;
548 } else {
549 len = TCPOLEN_FASTOPEN_BASE + foc->len;
550 *p++ = TCPOPT_FASTOPEN;
551 *p++ = len;
552 }
553
554 memcpy(p, foc->val, foc->len);
555 if ((len & 3) == 2) {
556 p[foc->len] = TCPOPT_NOP;
557 p[foc->len + 1] = TCPOPT_NOP;
558 }
559 ptr += (len + 3) >> 2;
560 }
561
562 smc_options_write(ptr, &options);
563
564 mptcp_options_write(ptr, opts);
565}
566
567static void smc_set_option(const struct tcp_sock *tp,
568 struct tcp_out_options *opts,
569 unsigned int *remaining)
570{
571#if IS_ENABLED(CONFIG_SMC)
572 if (static_branch_unlikely(&tcp_have_smc)) {
573 if (tp->syn_smc) {
574 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
575 opts->options |= OPTION_SMC;
576 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
577 }
578 }
579 }
580#endif
581}
582
583static void smc_set_option_cond(const struct tcp_sock *tp,
584 const struct inet_request_sock *ireq,
585 struct tcp_out_options *opts,
586 unsigned int *remaining)
587{
588#if IS_ENABLED(CONFIG_SMC)
589 if (static_branch_unlikely(&tcp_have_smc)) {
590 if (tp->syn_smc && ireq->smc_ok) {
591 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
592 opts->options |= OPTION_SMC;
593 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
594 }
595 }
596 }
597#endif
598}
599
600static void mptcp_set_option_cond(const struct request_sock *req,
601 struct tcp_out_options *opts,
602 unsigned int *remaining)
603{
604 if (rsk_is_mptcp(req)) {
605 unsigned int size;
606
607 if (mptcp_synack_options(req, &size, &opts->mptcp)) {
608 if (*remaining >= size) {
609 opts->options |= OPTION_MPTCP;
610 *remaining -= size;
611 }
612 }
613 }
614}
615
616/* Compute TCP options for SYN packets. This is not the final
617 * network wire format yet.
618 */
619static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
620 struct tcp_out_options *opts,
621 struct tcp_md5sig_key **md5)
622{
623 struct tcp_sock *tp = tcp_sk(sk);
624 unsigned int remaining = MAX_TCP_OPTION_SPACE;
625 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
626
627 *md5 = NULL;
628#ifdef CONFIG_TCP_MD5SIG
629 if (static_branch_unlikely(&tcp_md5_needed) &&
630 rcu_access_pointer(tp->md5sig_info)) {
631 *md5 = tp->af_specific->md5_lookup(sk, sk);
632 if (*md5) {
633 opts->options |= OPTION_MD5;
634 remaining -= TCPOLEN_MD5SIG_ALIGNED;
635 }
636 }
637#endif
638
639 /* We always get an MSS option. The option bytes which will be seen in
640 * normal data packets should timestamps be used, must be in the MSS
641 * advertised. But we subtract them from tp->mss_cache so that
642 * calculations in tcp_sendmsg are simpler etc. So account for this
643 * fact here if necessary. If we don't do this correctly, as a
644 * receiver we won't recognize data packets as being full sized when we
645 * should, and thus we won't abide by the delayed ACK rules correctly.
646 * SACKs don't matter, we never delay an ACK when we have any of those
647 * going out. */
648 opts->mss = tcp_advertise_mss(sk);
649 remaining -= TCPOLEN_MSS_ALIGNED;
650
651 if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
652 opts->options |= OPTION_TS;
653 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
654 opts->tsecr = tp->rx_opt.ts_recent;
655 remaining -= TCPOLEN_TSTAMP_ALIGNED;
656 }
657 if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
658 opts->ws = tp->rx_opt.rcv_wscale;
659 opts->options |= OPTION_WSCALE;
660 remaining -= TCPOLEN_WSCALE_ALIGNED;
661 }
662 if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
663 opts->options |= OPTION_SACK_ADVERTISE;
664 if (unlikely(!(OPTION_TS & opts->options)))
665 remaining -= TCPOLEN_SACKPERM_ALIGNED;
666 }
667
668 if (fastopen && fastopen->cookie.len >= 0) {
669 u32 need = fastopen->cookie.len;
670
671 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
672 TCPOLEN_FASTOPEN_BASE;
673 need = (need + 3) & ~3U; /* Align to 32 bits */
674 if (remaining >= need) {
675 opts->options |= OPTION_FAST_OPEN_COOKIE;
676 opts->fastopen_cookie = &fastopen->cookie;
677 remaining -= need;
678 tp->syn_fastopen = 1;
679 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
680 }
681 }
682
683 smc_set_option(tp, opts, &remaining);
684
685 if (sk_is_mptcp(sk)) {
686 unsigned int size;
687
688 if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) {
689 opts->options |= OPTION_MPTCP;
690 remaining -= size;
691 }
692 }
693
694 return MAX_TCP_OPTION_SPACE - remaining;
695}
696
697/* Set up TCP options for SYN-ACKs. */
698static unsigned int tcp_synack_options(const struct sock *sk,
699 struct request_sock *req,
700 unsigned int mss, struct sk_buff *skb,
701 struct tcp_out_options *opts,
702 const struct tcp_md5sig_key *md5,
703 struct tcp_fastopen_cookie *foc,
704 enum tcp_synack_type synack_type)
705{
706 struct inet_request_sock *ireq = inet_rsk(req);
707 unsigned int remaining = MAX_TCP_OPTION_SPACE;
708
709#ifdef CONFIG_TCP_MD5SIG
710 if (md5) {
711 opts->options |= OPTION_MD5;
712 remaining -= TCPOLEN_MD5SIG_ALIGNED;
713
714 /* We can't fit any SACK blocks in a packet with MD5 + TS
715 * options. There was discussion about disabling SACK
716 * rather than TS in order to fit in better with old,
717 * buggy kernels, but that was deemed to be unnecessary.
718 */
719 if (synack_type != TCP_SYNACK_COOKIE)
720 ireq->tstamp_ok &= !ireq->sack_ok;
721 }
722#endif
723
724 /* We always send an MSS option. */
725 opts->mss = mss;
726 remaining -= TCPOLEN_MSS_ALIGNED;
727
728 if (likely(ireq->wscale_ok)) {
729 opts->ws = ireq->rcv_wscale;
730 opts->options |= OPTION_WSCALE;
731 remaining -= TCPOLEN_WSCALE_ALIGNED;
732 }
733 if (likely(ireq->tstamp_ok)) {
734 opts->options |= OPTION_TS;
735 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
736 opts->tsecr = req->ts_recent;
737 remaining -= TCPOLEN_TSTAMP_ALIGNED;
738 }
739 if (likely(ireq->sack_ok)) {
740 opts->options |= OPTION_SACK_ADVERTISE;
741 if (unlikely(!ireq->tstamp_ok))
742 remaining -= TCPOLEN_SACKPERM_ALIGNED;
743 }
744 if (foc != NULL && foc->len >= 0) {
745 u32 need = foc->len;
746
747 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
748 TCPOLEN_FASTOPEN_BASE;
749 need = (need + 3) & ~3U; /* Align to 32 bits */
750 if (remaining >= need) {
751 opts->options |= OPTION_FAST_OPEN_COOKIE;
752 opts->fastopen_cookie = foc;
753 remaining -= need;
754 }
755 }
756
757 mptcp_set_option_cond(req, opts, &remaining);
758
759 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
760
761 return MAX_TCP_OPTION_SPACE - remaining;
762}
763
764/* Compute TCP options for ESTABLISHED sockets. This is not the
765 * final wire format yet.
766 */
767static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
768 struct tcp_out_options *opts,
769 struct tcp_md5sig_key **md5)
770{
771 struct tcp_sock *tp = tcp_sk(sk);
772 unsigned int size = 0;
773 unsigned int eff_sacks;
774
775 opts->options = 0;
776
777 *md5 = NULL;
778#ifdef CONFIG_TCP_MD5SIG
779 if (static_branch_unlikely(&tcp_md5_needed) &&
780 rcu_access_pointer(tp->md5sig_info)) {
781 *md5 = tp->af_specific->md5_lookup(sk, sk);
782 if (*md5) {
783 opts->options |= OPTION_MD5;
784 size += TCPOLEN_MD5SIG_ALIGNED;
785 }
786 }
787#endif
788
789 if (likely(tp->rx_opt.tstamp_ok)) {
790 opts->options |= OPTION_TS;
791 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
792 opts->tsecr = tp->rx_opt.ts_recent;
793 size += TCPOLEN_TSTAMP_ALIGNED;
794 }
795
796 /* MPTCP options have precedence over SACK for the limited TCP
797 * option space because a MPTCP connection would be forced to
798 * fall back to regular TCP if a required multipath option is
799 * missing. SACK still gets a chance to use whatever space is
800 * left.
801 */
802 if (sk_is_mptcp(sk)) {
803 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
804 unsigned int opt_size = 0;
805
806 if (mptcp_established_options(sk, skb, &opt_size, remaining,
807 &opts->mptcp)) {
808 opts->options |= OPTION_MPTCP;
809 size += opt_size;
810 }
811 }
812
813 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
814 if (unlikely(eff_sacks)) {
815 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
816 if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED +
817 TCPOLEN_SACK_PERBLOCK))
818 return size;
819
820 opts->num_sack_blocks =
821 min_t(unsigned int, eff_sacks,
822 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
823 TCPOLEN_SACK_PERBLOCK);
824
825 size += TCPOLEN_SACK_BASE_ALIGNED +
826 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
827 }
828
829 return size;
830}
831
832
833/* TCP SMALL QUEUES (TSQ)
834 *
835 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
836 * to reduce RTT and bufferbloat.
837 * We do this using a special skb destructor (tcp_wfree).
838 *
839 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
840 * needs to be reallocated in a driver.
841 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
842 *
843 * Since transmit from skb destructor is forbidden, we use a tasklet
844 * to process all sockets that eventually need to send more skbs.
845 * We use one tasklet per cpu, with its own queue of sockets.
846 */
847struct tsq_tasklet {
848 struct tasklet_struct tasklet;
849 struct list_head head; /* queue of tcp sockets */
850};
851static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
852
853static void tcp_tsq_write(struct sock *sk)
854{
855 if ((1 << sk->sk_state) &
856 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
857 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
858 struct tcp_sock *tp = tcp_sk(sk);
859
860 if (tp->lost_out > tp->retrans_out &&
861 tp->snd_cwnd > tcp_packets_in_flight(tp)) {
862 tcp_mstamp_refresh(tp);
863 tcp_xmit_retransmit_queue(sk);
864 }
865
866 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
867 0, GFP_ATOMIC);
868 }
869}
870
871static void tcp_tsq_handler(struct sock *sk)
872{
873 bh_lock_sock(sk);
874 if (!sock_owned_by_user(sk))
875 tcp_tsq_write(sk);
876 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
877 sock_hold(sk);
878 bh_unlock_sock(sk);
879}
880/*
881 * One tasklet per cpu tries to send more skbs.
882 * We run in tasklet context but need to disable irqs when
883 * transferring tsq->head because tcp_wfree() might
884 * interrupt us (non NAPI drivers)
885 */
886static void tcp_tasklet_func(unsigned long data)
887{
888 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
889 LIST_HEAD(list);
890 unsigned long flags;
891 struct list_head *q, *n;
892 struct tcp_sock *tp;
893 struct sock *sk;
894
895 local_irq_save(flags);
896 list_splice_init(&tsq->head, &list);
897 local_irq_restore(flags);
898
899 list_for_each_safe(q, n, &list) {
900 tp = list_entry(q, struct tcp_sock, tsq_node);
901 list_del(&tp->tsq_node);
902
903 sk = (struct sock *)tp;
904 smp_mb__before_atomic();
905 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
906
907 tcp_tsq_handler(sk);
908 sk_free(sk);
909 }
910}
911
912#define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
913 TCPF_WRITE_TIMER_DEFERRED | \
914 TCPF_DELACK_TIMER_DEFERRED | \
915 TCPF_MTU_REDUCED_DEFERRED)
916/**
917 * tcp_release_cb - tcp release_sock() callback
918 * @sk: socket
919 *
920 * called from release_sock() to perform protocol dependent
921 * actions before socket release.
922 */
923void tcp_release_cb(struct sock *sk)
924{
925 unsigned long flags, nflags;
926
927 /* perform an atomic operation only if at least one flag is set */
928 do {
929 flags = sk->sk_tsq_flags;
930 if (!(flags & TCP_DEFERRED_ALL))
931 return;
932 nflags = flags & ~TCP_DEFERRED_ALL;
933 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
934
935 if (flags & TCPF_TSQ_DEFERRED) {
936 tcp_tsq_write(sk);
937 __sock_put(sk);
938 }
939 /* Here begins the tricky part :
940 * We are called from release_sock() with :
941 * 1) BH disabled
942 * 2) sk_lock.slock spinlock held
943 * 3) socket owned by us (sk->sk_lock.owned == 1)
944 *
945 * But following code is meant to be called from BH handlers,
946 * so we should keep BH disabled, but early release socket ownership
947 */
948 sock_release_ownership(sk);
949
950 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
951 tcp_write_timer_handler(sk);
952 __sock_put(sk);
953 }
954 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
955 tcp_delack_timer_handler(sk);
956 __sock_put(sk);
957 }
958 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
959 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
960 __sock_put(sk);
961 }
962}
963EXPORT_SYMBOL(tcp_release_cb);
964
965void __init tcp_tasklet_init(void)
966{
967 int i;
968
969 for_each_possible_cpu(i) {
970 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
971
972 INIT_LIST_HEAD(&tsq->head);
973 tasklet_init(&tsq->tasklet,
974 tcp_tasklet_func,
975 (unsigned long)tsq);
976 }
977}
978
979/*
980 * Write buffer destructor automatically called from kfree_skb.
981 * We can't xmit new skbs from this context, as we might already
982 * hold qdisc lock.
983 */
984void tcp_wfree(struct sk_buff *skb)
985{
986 struct sock *sk = skb->sk;
987 struct tcp_sock *tp = tcp_sk(sk);
988 unsigned long flags, nval, oval;
989
990 /* Keep one reference on sk_wmem_alloc.
991 * Will be released by sk_free() from here or tcp_tasklet_func()
992 */
993 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
994
995 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
996 * Wait until our queues (qdisc + devices) are drained.
997 * This gives :
998 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
999 * - chance for incoming ACK (processed by another cpu maybe)
1000 * to migrate this flow (skb->ooo_okay will be eventually set)
1001 */
1002 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
1003 goto out;
1004
1005 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
1006 struct tsq_tasklet *tsq;
1007 bool empty;
1008
1009 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
1010 goto out;
1011
1012 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
1013 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
1014 if (nval != oval)
1015 continue;
1016
1017 /* queue this socket to tasklet queue */
1018 local_irq_save(flags);
1019 tsq = this_cpu_ptr(&tsq_tasklet);
1020 empty = list_empty(&tsq->head);
1021 list_add(&tp->tsq_node, &tsq->head);
1022 if (empty)
1023 tasklet_schedule(&tsq->tasklet);
1024 local_irq_restore(flags);
1025 return;
1026 }
1027out:
1028 sk_free(sk);
1029}
1030
1031/* Note: Called under soft irq.
1032 * We can call TCP stack right away, unless socket is owned by user.
1033 */
1034enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
1035{
1036 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
1037 struct sock *sk = (struct sock *)tp;
1038
1039 tcp_tsq_handler(sk);
1040 sock_put(sk);
1041
1042 return HRTIMER_NORESTART;
1043}
1044
1045static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
1046 u64 prior_wstamp)
1047{
1048 struct tcp_sock *tp = tcp_sk(sk);
1049
1050 if (sk->sk_pacing_status != SK_PACING_NONE) {
1051 unsigned long rate = sk->sk_pacing_rate;
1052
1053 /* Original sch_fq does not pace first 10 MSS
1054 * Note that tp->data_segs_out overflows after 2^32 packets,
1055 * this is a minor annoyance.
1056 */
1057 if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
1058 u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
1059 u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
1060
1061 /* take into account OS jitter */
1062 len_ns -= min_t(u64, len_ns / 2, credit);
1063 tp->tcp_wstamp_ns += len_ns;
1064 }
1065 }
1066 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1067}
1068
1069INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1070INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1071INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb));
1072
1073/* This routine actually transmits TCP packets queued in by
1074 * tcp_do_sendmsg(). This is used by both the initial
1075 * transmission and possible later retransmissions.
1076 * All SKB's seen here are completely headerless. It is our
1077 * job to build the TCP header, and pass the packet down to
1078 * IP so it can do the same plus pass the packet off to the
1079 * device.
1080 *
1081 * We are working here with either a clone of the original
1082 * SKB, or a fresh unique copy made by the retransmit engine.
1083 */
1084static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1085 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1086{
1087 const struct inet_connection_sock *icsk = inet_csk(sk);
1088 struct inet_sock *inet;
1089 struct tcp_sock *tp;
1090 struct tcp_skb_cb *tcb;
1091 struct tcp_out_options opts;
1092 unsigned int tcp_options_size, tcp_header_size;
1093 struct sk_buff *oskb = NULL;
1094 struct tcp_md5sig_key *md5;
1095 struct tcphdr *th;
1096 u64 prior_wstamp;
1097 int err;
1098
1099 BUG_ON(!skb || !tcp_skb_pcount(skb));
1100 tp = tcp_sk(sk);
1101 prior_wstamp = tp->tcp_wstamp_ns;
1102 tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1103 skb->skb_mstamp_ns = tp->tcp_wstamp_ns;
1104 if (clone_it) {
1105 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1106 - tp->snd_una;
1107 oskb = skb;
1108
1109 tcp_skb_tsorted_save(oskb) {
1110 if (unlikely(skb_cloned(oskb)))
1111 skb = pskb_copy(oskb, gfp_mask);
1112 else
1113 skb = skb_clone(oskb, gfp_mask);
1114 } tcp_skb_tsorted_restore(oskb);
1115
1116 if (unlikely(!skb))
1117 return -ENOBUFS;
1118 /* retransmit skbs might have a non zero value in skb->dev
1119 * because skb->dev is aliased with skb->rbnode.rb_left
1120 */
1121 skb->dev = NULL;
1122 }
1123
1124 inet = inet_sk(sk);
1125 tcb = TCP_SKB_CB(skb);
1126 memset(&opts, 0, sizeof(opts));
1127
1128 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1129 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1130 } else {
1131 tcp_options_size = tcp_established_options(sk, skb, &opts,
1132 &md5);
1133 /* Force a PSH flag on all (GSO) packets to expedite GRO flush
1134 * at receiver : This slightly improve GRO performance.
1135 * Note that we do not force the PSH flag for non GSO packets,
1136 * because they might be sent under high congestion events,
1137 * and in this case it is better to delay the delivery of 1-MSS
1138 * packets and thus the corresponding ACK packet that would
1139 * release the following packet.
1140 */
1141 if (tcp_skb_pcount(skb) > 1)
1142 tcb->tcp_flags |= TCPHDR_PSH;
1143 }
1144 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1145
1146 /* if no packet is in qdisc/device queue, then allow XPS to select
1147 * another queue. We can be called from tcp_tsq_handler()
1148 * which holds one reference to sk.
1149 *
1150 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1151 * One way to get this would be to set skb->truesize = 2 on them.
1152 */
1153 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1154
1155 /* If we had to use memory reserve to allocate this skb,
1156 * this might cause drops if packet is looped back :
1157 * Other socket might not have SOCK_MEMALLOC.
1158 * Packets not looped back do not care about pfmemalloc.
1159 */
1160 skb->pfmemalloc = 0;
1161
1162 skb_push(skb, tcp_header_size);
1163 skb_reset_transport_header(skb);
1164
1165 skb_orphan(skb);
1166 skb->sk = sk;
1167 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1168 skb_set_hash_from_sk(skb, sk);
1169 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1170
1171 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1172
1173 /* Build TCP header and checksum it. */
1174 th = (struct tcphdr *)skb->data;
1175 th->source = inet->inet_sport;
1176 th->dest = inet->inet_dport;
1177 th->seq = htonl(tcb->seq);
1178 th->ack_seq = htonl(rcv_nxt);
1179 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1180 tcb->tcp_flags);
1181
1182 th->check = 0;
1183 th->urg_ptr = 0;
1184
1185 /* The urg_mode check is necessary during a below snd_una win probe */
1186 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1187 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1188 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1189 th->urg = 1;
1190 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1191 th->urg_ptr = htons(0xFFFF);
1192 th->urg = 1;
1193 }
1194 }
1195
1196 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1197 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1198 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1199 th->window = htons(tcp_select_window(sk));
1200 tcp_ecn_send(sk, skb, th, tcp_header_size);
1201 } else {
1202 /* RFC1323: The window in SYN & SYN/ACK segments
1203 * is never scaled.
1204 */
1205 th->window = htons(min(tp->rcv_wnd, 65535U));
1206 }
1207#ifdef CONFIG_TCP_MD5SIG
1208 /* Calculate the MD5 hash, as we have all we need now */
1209 if (md5) {
1210 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1211 tp->af_specific->calc_md5_hash(opts.hash_location,
1212 md5, sk, skb);
1213 }
1214#endif
1215
1216 INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check,
1217 tcp_v6_send_check, tcp_v4_send_check,
1218 sk, skb);
1219
1220 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1221 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1222
1223 if (skb->len != tcp_header_size) {
1224 tcp_event_data_sent(tp, sk);
1225 tp->data_segs_out += tcp_skb_pcount(skb);
1226 tp->bytes_sent += skb->len - tcp_header_size;
1227 }
1228
1229 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1230 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1231 tcp_skb_pcount(skb));
1232
1233 tp->segs_out += tcp_skb_pcount(skb);
1234 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1235 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1236 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1237
1238 /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1239
1240 /* Cleanup our debris for IP stacks */
1241 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1242 sizeof(struct inet6_skb_parm)));
1243
1244 tcp_add_tx_delay(skb, tp);
1245
1246 err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit,
1247 inet6_csk_xmit, ip_queue_xmit,
1248 sk, skb, &inet->cork.fl);
1249
1250 if (unlikely(err > 0)) {
1251 tcp_enter_cwr(sk);
1252 err = net_xmit_eval(err);
1253 }
1254 if (!err && oskb) {
1255 tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1256 tcp_rate_skb_sent(sk, oskb);
1257 }
1258 return err;
1259}
1260
1261static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1262 gfp_t gfp_mask)
1263{
1264 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1265 tcp_sk(sk)->rcv_nxt);
1266}
1267
1268/* This routine just queues the buffer for sending.
1269 *
1270 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1271 * otherwise socket can stall.
1272 */
1273static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1274{
1275 struct tcp_sock *tp = tcp_sk(sk);
1276
1277 /* Advance write_seq and place onto the write_queue. */
1278 WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1279 __skb_header_release(skb);
1280 tcp_add_write_queue_tail(sk, skb);
1281 sk_wmem_queued_add(sk, skb->truesize);
1282 sk_mem_charge(sk, skb->truesize);
1283}
1284
1285/* Initialize TSO segments for a packet. */
1286static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1287{
1288 if (skb->len <= mss_now) {
1289 /* Avoid the costly divide in the normal
1290 * non-TSO case.
1291 */
1292 tcp_skb_pcount_set(skb, 1);
1293 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1294 } else {
1295 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1296 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1297 }
1298}
1299
1300/* Pcount in the middle of the write queue got changed, we need to do various
1301 * tweaks to fix counters
1302 */
1303static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1304{
1305 struct tcp_sock *tp = tcp_sk(sk);
1306
1307 tp->packets_out -= decr;
1308
1309 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1310 tp->sacked_out -= decr;
1311 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1312 tp->retrans_out -= decr;
1313 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1314 tp->lost_out -= decr;
1315
1316 /* Reno case is special. Sigh... */
1317 if (tcp_is_reno(tp) && decr > 0)
1318 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1319
1320 if (tp->lost_skb_hint &&
1321 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1322 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1323 tp->lost_cnt_hint -= decr;
1324
1325 tcp_verify_left_out(tp);
1326}
1327
1328static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1329{
1330 return TCP_SKB_CB(skb)->txstamp_ack ||
1331 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1332}
1333
1334static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1335{
1336 struct skb_shared_info *shinfo = skb_shinfo(skb);
1337
1338 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1339 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1340 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1341 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1342
1343 shinfo->tx_flags &= ~tsflags;
1344 shinfo2->tx_flags |= tsflags;
1345 swap(shinfo->tskey, shinfo2->tskey);
1346 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1347 TCP_SKB_CB(skb)->txstamp_ack = 0;
1348 }
1349}
1350
1351static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1352{
1353 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1354 TCP_SKB_CB(skb)->eor = 0;
1355}
1356
1357/* Insert buff after skb on the write or rtx queue of sk. */
1358static void tcp_insert_write_queue_after(struct sk_buff *skb,
1359 struct sk_buff *buff,
1360 struct sock *sk,
1361 enum tcp_queue tcp_queue)
1362{
1363 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1364 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1365 else
1366 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1367}
1368
1369/* Function to create two new TCP segments. Shrinks the given segment
1370 * to the specified size and appends a new segment with the rest of the
1371 * packet to the list. This won't be called frequently, I hope.
1372 * Remember, these are still headerless SKBs at this point.
1373 */
1374int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1375 struct sk_buff *skb, u32 len,
1376 unsigned int mss_now, gfp_t gfp)
1377{
1378 struct tcp_sock *tp = tcp_sk(sk);
1379 struct sk_buff *buff;
1380 int nsize, old_factor;
1381 long limit;
1382 int nlen;
1383 u8 flags;
1384
1385 if (WARN_ON(len > skb->len))
1386 return -EINVAL;
1387
1388 nsize = skb_headlen(skb) - len;
1389 if (nsize < 0)
1390 nsize = 0;
1391
1392 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1393 * We need some allowance to not penalize applications setting small
1394 * SO_SNDBUF values.
1395 * Also allow first and last skb in retransmit queue to be split.
1396 */
1397 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_MAX_SIZE);
1398 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1399 tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1400 skb != tcp_rtx_queue_head(sk) &&
1401 skb != tcp_rtx_queue_tail(sk))) {
1402 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1403 return -ENOMEM;
1404 }
1405
1406 if (skb_unclone(skb, gfp))
1407 return -ENOMEM;
1408
1409 /* Get a new skb... force flag on. */
1410 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1411 if (!buff)
1412 return -ENOMEM; /* We'll just try again later. */
1413 skb_copy_decrypted(buff, skb);
1414
1415 sk_wmem_queued_add(sk, buff->truesize);
1416 sk_mem_charge(sk, buff->truesize);
1417 nlen = skb->len - len - nsize;
1418 buff->truesize += nlen;
1419 skb->truesize -= nlen;
1420
1421 /* Correct the sequence numbers. */
1422 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1423 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1424 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1425
1426 /* PSH and FIN should only be set in the second packet. */
1427 flags = TCP_SKB_CB(skb)->tcp_flags;
1428 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1429 TCP_SKB_CB(buff)->tcp_flags = flags;
1430 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1431 tcp_skb_fragment_eor(skb, buff);
1432
1433 skb_split(skb, buff, len);
1434
1435 buff->ip_summed = CHECKSUM_PARTIAL;
1436
1437 buff->tstamp = skb->tstamp;
1438 tcp_fragment_tstamp(skb, buff);
1439
1440 old_factor = tcp_skb_pcount(skb);
1441
1442 /* Fix up tso_factor for both original and new SKB. */
1443 tcp_set_skb_tso_segs(skb, mss_now);
1444 tcp_set_skb_tso_segs(buff, mss_now);
1445
1446 /* Update delivered info for the new segment */
1447 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1448
1449 /* If this packet has been sent out already, we must
1450 * adjust the various packet counters.
1451 */
1452 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1453 int diff = old_factor - tcp_skb_pcount(skb) -
1454 tcp_skb_pcount(buff);
1455
1456 if (diff)
1457 tcp_adjust_pcount(sk, skb, diff);
1458 }
1459
1460 /* Link BUFF into the send queue. */
1461 __skb_header_release(buff);
1462 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1463 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1464 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1465
1466 return 0;
1467}
1468
1469/* This is similar to __pskb_pull_tail(). The difference is that pulled
1470 * data is not copied, but immediately discarded.
1471 */
1472static int __pskb_trim_head(struct sk_buff *skb, int len)
1473{
1474 struct skb_shared_info *shinfo;
1475 int i, k, eat;
1476
1477 eat = min_t(int, len, skb_headlen(skb));
1478 if (eat) {
1479 __skb_pull(skb, eat);
1480 len -= eat;
1481 if (!len)
1482 return 0;
1483 }
1484 eat = len;
1485 k = 0;
1486 shinfo = skb_shinfo(skb);
1487 for (i = 0; i < shinfo->nr_frags; i++) {
1488 int size = skb_frag_size(&shinfo->frags[i]);
1489
1490 if (size <= eat) {
1491 skb_frag_unref(skb, i);
1492 eat -= size;
1493 } else {
1494 shinfo->frags[k] = shinfo->frags[i];
1495 if (eat) {
1496 skb_frag_off_add(&shinfo->frags[k], eat);
1497 skb_frag_size_sub(&shinfo->frags[k], eat);
1498 eat = 0;
1499 }
1500 k++;
1501 }
1502 }
1503 shinfo->nr_frags = k;
1504
1505 skb->data_len -= len;
1506 skb->len = skb->data_len;
1507 return len;
1508}
1509
1510/* Remove acked data from a packet in the transmit queue. */
1511int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1512{
1513 u32 delta_truesize;
1514
1515 if (skb_unclone(skb, GFP_ATOMIC))
1516 return -ENOMEM;
1517
1518 delta_truesize = __pskb_trim_head(skb, len);
1519
1520 TCP_SKB_CB(skb)->seq += len;
1521 skb->ip_summed = CHECKSUM_PARTIAL;
1522
1523 if (delta_truesize) {
1524 skb->truesize -= delta_truesize;
1525 sk_wmem_queued_add(sk, -delta_truesize);
1526 sk_mem_uncharge(sk, delta_truesize);
1527 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1528 }
1529
1530 /* Any change of skb->len requires recalculation of tso factor. */
1531 if (tcp_skb_pcount(skb) > 1)
1532 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1533
1534 return 0;
1535}
1536
1537/* Calculate MSS not accounting any TCP options. */
1538static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1539{
1540 const struct tcp_sock *tp = tcp_sk(sk);
1541 const struct inet_connection_sock *icsk = inet_csk(sk);
1542 int mss_now;
1543
1544 /* Calculate base mss without TCP options:
1545 It is MMS_S - sizeof(tcphdr) of rfc1122
1546 */
1547 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1548
1549 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1550 if (icsk->icsk_af_ops->net_frag_header_len) {
1551 const struct dst_entry *dst = __sk_dst_get(sk);
1552
1553 if (dst && dst_allfrag(dst))
1554 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1555 }
1556
1557 /* Clamp it (mss_clamp does not include tcp options) */
1558 if (mss_now > tp->rx_opt.mss_clamp)
1559 mss_now = tp->rx_opt.mss_clamp;
1560
1561 /* Now subtract optional transport overhead */
1562 mss_now -= icsk->icsk_ext_hdr_len;
1563
1564 /* Then reserve room for full set of TCP options and 8 bytes of data */
1565 mss_now = max(mss_now, sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss);
1566 return mss_now;
1567}
1568
1569/* Calculate MSS. Not accounting for SACKs here. */
1570int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1571{
1572 /* Subtract TCP options size, not including SACKs */
1573 return __tcp_mtu_to_mss(sk, pmtu) -
1574 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1575}
1576
1577/* Inverse of above */
1578int tcp_mss_to_mtu(struct sock *sk, int mss)
1579{
1580 const struct tcp_sock *tp = tcp_sk(sk);
1581 const struct inet_connection_sock *icsk = inet_csk(sk);
1582 int mtu;
1583
1584 mtu = mss +
1585 tp->tcp_header_len +
1586 icsk->icsk_ext_hdr_len +
1587 icsk->icsk_af_ops->net_header_len;
1588
1589 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1590 if (icsk->icsk_af_ops->net_frag_header_len) {
1591 const struct dst_entry *dst = __sk_dst_get(sk);
1592
1593 if (dst && dst_allfrag(dst))
1594 mtu += icsk->icsk_af_ops->net_frag_header_len;
1595 }
1596 return mtu;
1597}
1598EXPORT_SYMBOL(tcp_mss_to_mtu);
1599
1600/* MTU probing init per socket */
1601void tcp_mtup_init(struct sock *sk)
1602{
1603 struct tcp_sock *tp = tcp_sk(sk);
1604 struct inet_connection_sock *icsk = inet_csk(sk);
1605 struct net *net = sock_net(sk);
1606
1607 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1608 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1609 icsk->icsk_af_ops->net_header_len;
1610 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1611 icsk->icsk_mtup.probe_size = 0;
1612 if (icsk->icsk_mtup.enabled)
1613 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1614}
1615EXPORT_SYMBOL(tcp_mtup_init);
1616
1617/* This function synchronize snd mss to current pmtu/exthdr set.
1618
1619 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1620 for TCP options, but includes only bare TCP header.
1621
1622 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1623 It is minimum of user_mss and mss received with SYN.
1624 It also does not include TCP options.
1625
1626 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1627
1628 tp->mss_cache is current effective sending mss, including
1629 all tcp options except for SACKs. It is evaluated,
1630 taking into account current pmtu, but never exceeds
1631 tp->rx_opt.mss_clamp.
1632
1633 NOTE1. rfc1122 clearly states that advertised MSS
1634 DOES NOT include either tcp or ip options.
1635
1636 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1637 are READ ONLY outside this function. --ANK (980731)
1638 */
1639unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1640{
1641 struct tcp_sock *tp = tcp_sk(sk);
1642 struct inet_connection_sock *icsk = inet_csk(sk);
1643 int mss_now;
1644
1645 if (icsk->icsk_mtup.search_high > pmtu)
1646 icsk->icsk_mtup.search_high = pmtu;
1647
1648 mss_now = tcp_mtu_to_mss(sk, pmtu);
1649 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1650
1651 /* And store cached results */
1652 icsk->icsk_pmtu_cookie = pmtu;
1653 if (icsk->icsk_mtup.enabled)
1654 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1655 tp->mss_cache = mss_now;
1656
1657 return mss_now;
1658}
1659EXPORT_SYMBOL(tcp_sync_mss);
1660
1661/* Compute the current effective MSS, taking SACKs and IP options,
1662 * and even PMTU discovery events into account.
1663 */
1664unsigned int tcp_current_mss(struct sock *sk)
1665{
1666 const struct tcp_sock *tp = tcp_sk(sk);
1667 const struct dst_entry *dst = __sk_dst_get(sk);
1668 u32 mss_now;
1669 unsigned int header_len;
1670 struct tcp_out_options opts;
1671 struct tcp_md5sig_key *md5;
1672
1673 mss_now = tp->mss_cache;
1674
1675 if (dst) {
1676 u32 mtu = dst_mtu(dst);
1677 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1678 mss_now = tcp_sync_mss(sk, mtu);
1679 }
1680
1681 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1682 sizeof(struct tcphdr);
1683 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1684 * some common options. If this is an odd packet (because we have SACK
1685 * blocks etc) then our calculated header_len will be different, and
1686 * we have to adjust mss_now correspondingly */
1687 if (header_len != tp->tcp_header_len) {
1688 int delta = (int) header_len - tp->tcp_header_len;
1689 mss_now -= delta;
1690 }
1691
1692 return mss_now;
1693}
1694
1695/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1696 * As additional protections, we do not touch cwnd in retransmission phases,
1697 * and if application hit its sndbuf limit recently.
1698 */
1699static void tcp_cwnd_application_limited(struct sock *sk)
1700{
1701 struct tcp_sock *tp = tcp_sk(sk);
1702
1703 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1704 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1705 /* Limited by application or receiver window. */
1706 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1707 u32 win_used = max(tp->snd_cwnd_used, init_win);
1708 if (win_used < tp->snd_cwnd) {
1709 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1710 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1711 }
1712 tp->snd_cwnd_used = 0;
1713 }
1714 tp->snd_cwnd_stamp = tcp_jiffies32;
1715}
1716
1717static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1718{
1719 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1720 struct tcp_sock *tp = tcp_sk(sk);
1721
1722 /* Track the maximum number of outstanding packets in each
1723 * window, and remember whether we were cwnd-limited then.
1724 */
1725 if (!before(tp->snd_una, tp->max_packets_seq) ||
1726 tp->packets_out > tp->max_packets_out) {
1727 tp->max_packets_out = tp->packets_out;
1728 tp->max_packets_seq = tp->snd_nxt;
1729 tp->is_cwnd_limited = is_cwnd_limited;
1730 }
1731
1732 if (tcp_is_cwnd_limited(sk)) {
1733 /* Network is feed fully. */
1734 tp->snd_cwnd_used = 0;
1735 tp->snd_cwnd_stamp = tcp_jiffies32;
1736 } else {
1737 /* Network starves. */
1738 if (tp->packets_out > tp->snd_cwnd_used)
1739 tp->snd_cwnd_used = tp->packets_out;
1740
1741 if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle &&
1742 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1743 !ca_ops->cong_control)
1744 tcp_cwnd_application_limited(sk);
1745
1746 /* The following conditions together indicate the starvation
1747 * is caused by insufficient sender buffer:
1748 * 1) just sent some data (see tcp_write_xmit)
1749 * 2) not cwnd limited (this else condition)
1750 * 3) no more data to send (tcp_write_queue_empty())
1751 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1752 */
1753 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1754 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1755 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1756 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1757 }
1758}
1759
1760/* Minshall's variant of the Nagle send check. */
1761static bool tcp_minshall_check(const struct tcp_sock *tp)
1762{
1763 return after(tp->snd_sml, tp->snd_una) &&
1764 !after(tp->snd_sml, tp->snd_nxt);
1765}
1766
1767/* Update snd_sml if this skb is under mss
1768 * Note that a TSO packet might end with a sub-mss segment
1769 * The test is really :
1770 * if ((skb->len % mss) != 0)
1771 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1772 * But we can avoid doing the divide again given we already have
1773 * skb_pcount = skb->len / mss_now
1774 */
1775static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1776 const struct sk_buff *skb)
1777{
1778 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1779 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1780}
1781
1782/* Return false, if packet can be sent now without violation Nagle's rules:
1783 * 1. It is full sized. (provided by caller in %partial bool)
1784 * 2. Or it contains FIN. (already checked by caller)
1785 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1786 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1787 * With Minshall's modification: all sent small packets are ACKed.
1788 */
1789static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1790 int nonagle)
1791{
1792 return partial &&
1793 ((nonagle & TCP_NAGLE_CORK) ||
1794 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1795}
1796
1797/* Return how many segs we'd like on a TSO packet,
1798 * to send one TSO packet per ms
1799 */
1800static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1801 int min_tso_segs)
1802{
1803 u32 bytes, segs;
1804
1805 bytes = min_t(unsigned long,
1806 sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift),
1807 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1808
1809 /* Goal is to send at least one packet per ms,
1810 * not one big TSO packet every 100 ms.
1811 * This preserves ACK clocking and is consistent
1812 * with tcp_tso_should_defer() heuristic.
1813 */
1814 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1815
1816 return segs;
1817}
1818
1819/* Return the number of segments we want in the skb we are transmitting.
1820 * See if congestion control module wants to decide; otherwise, autosize.
1821 */
1822static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1823{
1824 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1825 u32 min_tso, tso_segs;
1826
1827 min_tso = ca_ops->min_tso_segs ?
1828 ca_ops->min_tso_segs(sk) :
1829 sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs;
1830
1831 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
1832 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1833}
1834
1835/* Returns the portion of skb which can be sent right away */
1836static unsigned int tcp_mss_split_point(const struct sock *sk,
1837 const struct sk_buff *skb,
1838 unsigned int mss_now,
1839 unsigned int max_segs,
1840 int nonagle)
1841{
1842 const struct tcp_sock *tp = tcp_sk(sk);
1843 u32 partial, needed, window, max_len;
1844
1845 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1846 max_len = mss_now * max_segs;
1847
1848 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1849 return max_len;
1850
1851 needed = min(skb->len, window);
1852
1853 if (max_len <= needed)
1854 return max_len;
1855
1856 partial = needed % mss_now;
1857 /* If last segment is not a full MSS, check if Nagle rules allow us
1858 * to include this last segment in this skb.
1859 * Otherwise, we'll split the skb at last MSS boundary
1860 */
1861 if (tcp_nagle_check(partial != 0, tp, nonagle))
1862 return needed - partial;
1863
1864 return needed;
1865}
1866
1867/* Can at least one segment of SKB be sent right now, according to the
1868 * congestion window rules? If so, return how many segments are allowed.
1869 */
1870static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1871 const struct sk_buff *skb)
1872{
1873 u32 in_flight, cwnd, halfcwnd;
1874
1875 /* Don't be strict about the congestion window for the final FIN. */
1876 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1877 tcp_skb_pcount(skb) == 1)
1878 return 1;
1879
1880 in_flight = tcp_packets_in_flight(tp);
1881 cwnd = tp->snd_cwnd;
1882 if (in_flight >= cwnd)
1883 return 0;
1884
1885 /* For better scheduling, ensure we have at least
1886 * 2 GSO packets in flight.
1887 */
1888 halfcwnd = max(cwnd >> 1, 1U);
1889 return min(halfcwnd, cwnd - in_flight);
1890}
1891
1892/* Initialize TSO state of a skb.
1893 * This must be invoked the first time we consider transmitting
1894 * SKB onto the wire.
1895 */
1896static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1897{
1898 int tso_segs = tcp_skb_pcount(skb);
1899
1900 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1901 tcp_set_skb_tso_segs(skb, mss_now);
1902 tso_segs = tcp_skb_pcount(skb);
1903 }
1904 return tso_segs;
1905}
1906
1907
1908/* Return true if the Nagle test allows this packet to be
1909 * sent now.
1910 */
1911static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1912 unsigned int cur_mss, int nonagle)
1913{
1914 /* Nagle rule does not apply to frames, which sit in the middle of the
1915 * write_queue (they have no chances to get new data).
1916 *
1917 * This is implemented in the callers, where they modify the 'nonagle'
1918 * argument based upon the location of SKB in the send queue.
1919 */
1920 if (nonagle & TCP_NAGLE_PUSH)
1921 return true;
1922
1923 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1924 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1925 return true;
1926
1927 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1928 return true;
1929
1930 return false;
1931}
1932
1933/* Does at least the first segment of SKB fit into the send window? */
1934static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1935 const struct sk_buff *skb,
1936 unsigned int cur_mss)
1937{
1938 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1939
1940 if (skb->len > cur_mss)
1941 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1942
1943 return !after(end_seq, tcp_wnd_end(tp));
1944}
1945
1946/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1947 * which is put after SKB on the list. It is very much like
1948 * tcp_fragment() except that it may make several kinds of assumptions
1949 * in order to speed up the splitting operation. In particular, we
1950 * know that all the data is in scatter-gather pages, and that the
1951 * packet has never been sent out before (and thus is not cloned).
1952 */
1953static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1954 unsigned int mss_now, gfp_t gfp)
1955{
1956 int nlen = skb->len - len;
1957 struct sk_buff *buff;
1958 u8 flags;
1959
1960 /* All of a TSO frame must be composed of paged data. */
1961 if (skb->len != skb->data_len)
1962 return tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
1963 skb, len, mss_now, gfp);
1964
1965 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1966 if (unlikely(!buff))
1967 return -ENOMEM;
1968 skb_copy_decrypted(buff, skb);
1969
1970 sk_wmem_queued_add(sk, buff->truesize);
1971 sk_mem_charge(sk, buff->truesize);
1972 buff->truesize += nlen;
1973 skb->truesize -= nlen;
1974
1975 /* Correct the sequence numbers. */
1976 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1977 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1978 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1979
1980 /* PSH and FIN should only be set in the second packet. */
1981 flags = TCP_SKB_CB(skb)->tcp_flags;
1982 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1983 TCP_SKB_CB(buff)->tcp_flags = flags;
1984
1985 /* This packet was never sent out yet, so no SACK bits. */
1986 TCP_SKB_CB(buff)->sacked = 0;
1987
1988 tcp_skb_fragment_eor(skb, buff);
1989
1990 buff->ip_summed = CHECKSUM_PARTIAL;
1991 skb_split(skb, buff, len);
1992 tcp_fragment_tstamp(skb, buff);
1993
1994 /* Fix up tso_factor for both original and new SKB. */
1995 tcp_set_skb_tso_segs(skb, mss_now);
1996 tcp_set_skb_tso_segs(buff, mss_now);
1997
1998 /* Link BUFF into the send queue. */
1999 __skb_header_release(buff);
2000 tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
2001
2002 return 0;
2003}
2004
2005/* Try to defer sending, if possible, in order to minimize the amount
2006 * of TSO splitting we do. View it as a kind of TSO Nagle test.
2007 *
2008 * This algorithm is from John Heffner.
2009 */
2010static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
2011 bool *is_cwnd_limited,
2012 bool *is_rwnd_limited,
2013 u32 max_segs)
2014{
2015 const struct inet_connection_sock *icsk = inet_csk(sk);
2016 u32 send_win, cong_win, limit, in_flight;
2017 struct tcp_sock *tp = tcp_sk(sk);
2018 struct sk_buff *head;
2019 int win_divisor;
2020 s64 delta;
2021
2022 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
2023 goto send_now;
2024
2025 /* Avoid bursty behavior by allowing defer
2026 * only if the last write was recent (1 ms).
2027 * Note that tp->tcp_wstamp_ns can be in the future if we have
2028 * packets waiting in a qdisc or device for EDT delivery.
2029 */
2030 delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
2031 if (delta > 0)
2032 goto send_now;
2033
2034 in_flight = tcp_packets_in_flight(tp);
2035
2036 BUG_ON(tcp_skb_pcount(skb) <= 1);
2037 BUG_ON(tp->snd_cwnd <= in_flight);
2038
2039 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2040
2041 /* From in_flight test above, we know that cwnd > in_flight. */
2042 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
2043
2044 limit = min(send_win, cong_win);
2045
2046 /* If a full-sized TSO skb can be sent, do it. */
2047 if (limit >= max_segs * tp->mss_cache)
2048 goto send_now;
2049
2050 /* Middle in queue won't get any more data, full sendable already? */
2051 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
2052 goto send_now;
2053
2054 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
2055 if (win_divisor) {
2056 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
2057
2058 /* If at least some fraction of a window is available,
2059 * just use it.
2060 */
2061 chunk /= win_divisor;
2062 if (limit >= chunk)
2063 goto send_now;
2064 } else {
2065 /* Different approach, try not to defer past a single
2066 * ACK. Receiver should ACK every other full sized
2067 * frame, so if we have space for more than 3 frames
2068 * then send now.
2069 */
2070 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
2071 goto send_now;
2072 }
2073
2074 /* TODO : use tsorted_sent_queue ? */
2075 head = tcp_rtx_queue_head(sk);
2076 if (!head)
2077 goto send_now;
2078 delta = tp->tcp_clock_cache - head->tstamp;
2079 /* If next ACK is likely to come too late (half srtt), do not defer */
2080 if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
2081 goto send_now;
2082
2083 /* Ok, it looks like it is advisable to defer.
2084 * Three cases are tracked :
2085 * 1) We are cwnd-limited
2086 * 2) We are rwnd-limited
2087 * 3) We are application limited.
2088 */
2089 if (cong_win < send_win) {
2090 if (cong_win <= skb->len) {
2091 *is_cwnd_limited = true;
2092 return true;
2093 }
2094 } else {
2095 if (send_win <= skb->len) {
2096 *is_rwnd_limited = true;
2097 return true;
2098 }
2099 }
2100
2101 /* If this packet won't get more data, do not wait. */
2102 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2103 TCP_SKB_CB(skb)->eor)
2104 goto send_now;
2105
2106 return true;
2107
2108send_now:
2109 return false;
2110}
2111
2112static inline void tcp_mtu_check_reprobe(struct sock *sk)
2113{
2114 struct inet_connection_sock *icsk = inet_csk(sk);
2115 struct tcp_sock *tp = tcp_sk(sk);
2116 struct net *net = sock_net(sk);
2117 u32 interval;
2118 s32 delta;
2119
2120 interval = net->ipv4.sysctl_tcp_probe_interval;
2121 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2122 if (unlikely(delta >= interval * HZ)) {
2123 int mss = tcp_current_mss(sk);
2124
2125 /* Update current search range */
2126 icsk->icsk_mtup.probe_size = 0;
2127 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2128 sizeof(struct tcphdr) +
2129 icsk->icsk_af_ops->net_header_len;
2130 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2131
2132 /* Update probe time stamp */
2133 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2134 }
2135}
2136
2137static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2138{
2139 struct sk_buff *skb, *next;
2140
2141 skb = tcp_send_head(sk);
2142 tcp_for_write_queue_from_safe(skb, next, sk) {
2143 if (len <= skb->len)
2144 break;
2145
2146 if (unlikely(TCP_SKB_CB(skb)->eor) || tcp_has_tx_tstamp(skb))
2147 return false;
2148
2149 len -= skb->len;
2150 }
2151
2152 return true;
2153}
2154
2155/* Create a new MTU probe if we are ready.
2156 * MTU probe is regularly attempting to increase the path MTU by
2157 * deliberately sending larger packets. This discovers routing
2158 * changes resulting in larger path MTUs.
2159 *
2160 * Returns 0 if we should wait to probe (no cwnd available),
2161 * 1 if a probe was sent,
2162 * -1 otherwise
2163 */
2164static int tcp_mtu_probe(struct sock *sk)
2165{
2166 struct inet_connection_sock *icsk = inet_csk(sk);
2167 struct tcp_sock *tp = tcp_sk(sk);
2168 struct sk_buff *skb, *nskb, *next;
2169 struct net *net = sock_net(sk);
2170 int probe_size;
2171 int size_needed;
2172 int copy, len;
2173 int mss_now;
2174 int interval;
2175
2176 /* Not currently probing/verifying,
2177 * not in recovery,
2178 * have enough cwnd, and
2179 * not SACKing (the variable headers throw things off)
2180 */
2181 if (likely(!icsk->icsk_mtup.enabled ||
2182 icsk->icsk_mtup.probe_size ||
2183 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2184 tp->snd_cwnd < 11 ||
2185 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2186 return -1;
2187
2188 /* Use binary search for probe_size between tcp_mss_base,
2189 * and current mss_clamp. if (search_high - search_low)
2190 * smaller than a threshold, backoff from probing.
2191 */
2192 mss_now = tcp_current_mss(sk);
2193 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2194 icsk->icsk_mtup.search_low) >> 1);
2195 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2196 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2197 /* When misfortune happens, we are reprobing actively,
2198 * and then reprobe timer has expired. We stick with current
2199 * probing process by not resetting search range to its orignal.
2200 */
2201 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2202 interval < net->ipv4.sysctl_tcp_probe_threshold) {
2203 /* Check whether enough time has elaplased for
2204 * another round of probing.
2205 */
2206 tcp_mtu_check_reprobe(sk);
2207 return -1;
2208 }
2209
2210 /* Have enough data in the send queue to probe? */
2211 if (tp->write_seq - tp->snd_nxt < size_needed)
2212 return -1;
2213
2214 if (tp->snd_wnd < size_needed)
2215 return -1;
2216 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2217 return 0;
2218
2219 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2220 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2221 if (!tcp_packets_in_flight(tp))
2222 return -1;
2223 else
2224 return 0;
2225 }
2226
2227 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2228 return -1;
2229
2230 /* We're allowed to probe. Build it now. */
2231 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2232 if (!nskb)
2233 return -1;
2234 sk_wmem_queued_add(sk, nskb->truesize);
2235 sk_mem_charge(sk, nskb->truesize);
2236
2237 skb = tcp_send_head(sk);
2238 skb_copy_decrypted(nskb, skb);
2239
2240 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2241 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2242 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2243 TCP_SKB_CB(nskb)->sacked = 0;
2244 nskb->csum = 0;
2245 nskb->ip_summed = CHECKSUM_PARTIAL;
2246
2247 tcp_insert_write_queue_before(nskb, skb, sk);
2248 tcp_highest_sack_replace(sk, skb, nskb);
2249
2250 len = 0;
2251 tcp_for_write_queue_from_safe(skb, next, sk) {
2252 copy = min_t(int, skb->len, probe_size - len);
2253 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2254
2255 if (skb->len <= copy) {
2256 /* We've eaten all the data from this skb.
2257 * Throw it away. */
2258 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2259 /* If this is the last SKB we copy and eor is set
2260 * we need to propagate it to the new skb.
2261 */
2262 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2263 tcp_skb_collapse_tstamp(nskb, skb);
2264 tcp_unlink_write_queue(skb, sk);
2265 sk_wmem_free_skb(sk, skb);
2266 } else {
2267 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2268 ~(TCPHDR_FIN|TCPHDR_PSH);
2269 if (!skb_shinfo(skb)->nr_frags) {
2270 skb_pull(skb, copy);
2271 } else {
2272 __pskb_trim_head(skb, copy);
2273 tcp_set_skb_tso_segs(skb, mss_now);
2274 }
2275 TCP_SKB_CB(skb)->seq += copy;
2276 }
2277
2278 len += copy;
2279
2280 if (len >= probe_size)
2281 break;
2282 }
2283 tcp_init_tso_segs(nskb, nskb->len);
2284
2285 /* We're ready to send. If this fails, the probe will
2286 * be resegmented into mss-sized pieces by tcp_write_xmit().
2287 */
2288 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2289 /* Decrement cwnd here because we are sending
2290 * effectively two packets. */
2291 tp->snd_cwnd--;
2292 tcp_event_new_data_sent(sk, nskb);
2293
2294 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2295 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2296 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2297
2298 return 1;
2299 }
2300
2301 return -1;
2302}
2303
2304static bool tcp_pacing_check(struct sock *sk)
2305{
2306 struct tcp_sock *tp = tcp_sk(sk);
2307
2308 if (!tcp_needs_internal_pacing(sk))
2309 return false;
2310
2311 if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2312 return false;
2313
2314 if (!hrtimer_is_queued(&tp->pacing_timer)) {
2315 hrtimer_start(&tp->pacing_timer,
2316 ns_to_ktime(tp->tcp_wstamp_ns),
2317 HRTIMER_MODE_ABS_PINNED_SOFT);
2318 sock_hold(sk);
2319 }
2320 return true;
2321}
2322
2323/* TCP Small Queues :
2324 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2325 * (These limits are doubled for retransmits)
2326 * This allows for :
2327 * - better RTT estimation and ACK scheduling
2328 * - faster recovery
2329 * - high rates
2330 * Alas, some drivers / subsystems require a fair amount
2331 * of queued bytes to ensure line rate.
2332 * One example is wifi aggregation (802.11 AMPDU)
2333 */
2334static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2335 unsigned int factor)
2336{
2337 unsigned long limit;
2338
2339 limit = max_t(unsigned long,
2340 2 * skb->truesize,
2341 sk->sk_pacing_rate >> READ_ONCE(sk->sk_pacing_shift));
2342 if (sk->sk_pacing_status == SK_PACING_NONE)
2343 limit = min_t(unsigned long, limit,
2344 sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
2345 limit <<= factor;
2346
2347 if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2348 tcp_sk(sk)->tcp_tx_delay) {
2349 u64 extra_bytes = (u64)sk->sk_pacing_rate * tcp_sk(sk)->tcp_tx_delay;
2350
2351 /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2352 * approximate our needs assuming an ~100% skb->truesize overhead.
2353 * USEC_PER_SEC is approximated by 2^20.
2354 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2355 */
2356 extra_bytes >>= (20 - 1);
2357 limit += extra_bytes;
2358 }
2359 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2360 /* Always send skb if rtx queue is empty.
2361 * No need to wait for TX completion to call us back,
2362 * after softirq/tasklet schedule.
2363 * This helps when TX completions are delayed too much.
2364 */
2365 if (tcp_rtx_queue_empty(sk))
2366 return false;
2367
2368 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2369 /* It is possible TX completion already happened
2370 * before we set TSQ_THROTTLED, so we must
2371 * test again the condition.
2372 */
2373 smp_mb__after_atomic();
2374 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2375 return true;
2376 }
2377 return false;
2378}
2379
2380static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2381{
2382 const u32 now = tcp_jiffies32;
2383 enum tcp_chrono old = tp->chrono_type;
2384
2385 if (old > TCP_CHRONO_UNSPEC)
2386 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2387 tp->chrono_start = now;
2388 tp->chrono_type = new;
2389}
2390
2391void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2392{
2393 struct tcp_sock *tp = tcp_sk(sk);
2394
2395 /* If there are multiple conditions worthy of tracking in a
2396 * chronograph then the highest priority enum takes precedence
2397 * over the other conditions. So that if something "more interesting"
2398 * starts happening, stop the previous chrono and start a new one.
2399 */
2400 if (type > tp->chrono_type)
2401 tcp_chrono_set(tp, type);
2402}
2403
2404void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2405{
2406 struct tcp_sock *tp = tcp_sk(sk);
2407
2408
2409 /* There are multiple conditions worthy of tracking in a
2410 * chronograph, so that the highest priority enum takes
2411 * precedence over the other conditions (see tcp_chrono_start).
2412 * If a condition stops, we only stop chrono tracking if
2413 * it's the "most interesting" or current chrono we are
2414 * tracking and starts busy chrono if we have pending data.
2415 */
2416 if (tcp_rtx_and_write_queues_empty(sk))
2417 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2418 else if (type == tp->chrono_type)
2419 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2420}
2421
2422/* This routine writes packets to the network. It advances the
2423 * send_head. This happens as incoming acks open up the remote
2424 * window for us.
2425 *
2426 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2427 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2428 * account rare use of URG, this is not a big flaw.
2429 *
2430 * Send at most one packet when push_one > 0. Temporarily ignore
2431 * cwnd limit to force at most one packet out when push_one == 2.
2432
2433 * Returns true, if no segments are in flight and we have queued segments,
2434 * but cannot send anything now because of SWS or another problem.
2435 */
2436static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2437 int push_one, gfp_t gfp)
2438{
2439 struct tcp_sock *tp = tcp_sk(sk);
2440 struct sk_buff *skb;
2441 unsigned int tso_segs, sent_pkts;
2442 int cwnd_quota;
2443 int result;
2444 bool is_cwnd_limited = false, is_rwnd_limited = false;
2445 u32 max_segs;
2446
2447 sent_pkts = 0;
2448
2449 tcp_mstamp_refresh(tp);
2450 if (!push_one) {
2451 /* Do MTU probing. */
2452 result = tcp_mtu_probe(sk);
2453 if (!result) {
2454 return false;
2455 } else if (result > 0) {
2456 sent_pkts = 1;
2457 }
2458 }
2459
2460 max_segs = tcp_tso_segs(sk, mss_now);
2461 while ((skb = tcp_send_head(sk))) {
2462 unsigned int limit;
2463
2464 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2465 /* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2466 skb->skb_mstamp_ns = tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2467 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2468 tcp_init_tso_segs(skb, mss_now);
2469 goto repair; /* Skip network transmission */
2470 }
2471
2472 if (tcp_pacing_check(sk))
2473 break;
2474
2475 tso_segs = tcp_init_tso_segs(skb, mss_now);
2476 BUG_ON(!tso_segs);
2477
2478 cwnd_quota = tcp_cwnd_test(tp, skb);
2479 if (!cwnd_quota) {
2480 if (push_one == 2)
2481 /* Force out a loss probe pkt. */
2482 cwnd_quota = 1;
2483 else
2484 break;
2485 }
2486
2487 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2488 is_rwnd_limited = true;
2489 break;
2490 }
2491
2492 if (tso_segs == 1) {
2493 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2494 (tcp_skb_is_last(sk, skb) ?
2495 nonagle : TCP_NAGLE_PUSH))))
2496 break;
2497 } else {
2498 if (!push_one &&
2499 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2500 &is_rwnd_limited, max_segs))
2501 break;
2502 }
2503
2504 limit = mss_now;
2505 if (tso_segs > 1 && !tcp_urg_mode(tp))
2506 limit = tcp_mss_split_point(sk, skb, mss_now,
2507 min_t(unsigned int,
2508 cwnd_quota,
2509 max_segs),
2510 nonagle);
2511
2512 if (skb->len > limit &&
2513 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2514 break;
2515
2516 if (tcp_small_queue_check(sk, skb, 0))
2517 break;
2518
2519 /* Argh, we hit an empty skb(), presumably a thread
2520 * is sleeping in sendmsg()/sk_stream_wait_memory().
2521 * We do not want to send a pure-ack packet and have
2522 * a strange looking rtx queue with empty packet(s).
2523 */
2524 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2525 break;
2526
2527 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2528 break;
2529
2530repair:
2531 /* Advance the send_head. This one is sent out.
2532 * This call will increment packets_out.
2533 */
2534 tcp_event_new_data_sent(sk, skb);
2535
2536 tcp_minshall_update(tp, mss_now, skb);
2537 sent_pkts += tcp_skb_pcount(skb);
2538
2539 if (push_one)
2540 break;
2541 }
2542
2543 if (is_rwnd_limited)
2544 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2545 else
2546 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2547
2548 if (likely(sent_pkts)) {
2549 if (tcp_in_cwnd_reduction(sk))
2550 tp->prr_out += sent_pkts;
2551
2552 /* Send one loss probe per tail loss episode. */
2553 if (push_one != 2)
2554 tcp_schedule_loss_probe(sk, false);
2555 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2556 tcp_cwnd_validate(sk, is_cwnd_limited);
2557 return false;
2558 }
2559 return !tp->packets_out && !tcp_write_queue_empty(sk);
2560}
2561
2562bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2563{
2564 struct inet_connection_sock *icsk = inet_csk(sk);
2565 struct tcp_sock *tp = tcp_sk(sk);
2566 u32 timeout, rto_delta_us;
2567 int early_retrans;
2568
2569 /* Don't do any loss probe on a Fast Open connection before 3WHS
2570 * finishes.
2571 */
2572 if (rcu_access_pointer(tp->fastopen_rsk))
2573 return false;
2574
2575 early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
2576 /* Schedule a loss probe in 2*RTT for SACK capable connections
2577 * not in loss recovery, that are either limited by cwnd or application.
2578 */
2579 if ((early_retrans != 3 && early_retrans != 4) ||
2580 !tp->packets_out || !tcp_is_sack(tp) ||
2581 (icsk->icsk_ca_state != TCP_CA_Open &&
2582 icsk->icsk_ca_state != TCP_CA_CWR))
2583 return false;
2584
2585 /* Probe timeout is 2*rtt. Add minimum RTO to account
2586 * for delayed ack when there's one outstanding packet. If no RTT
2587 * sample is available then probe after TCP_TIMEOUT_INIT.
2588 */
2589 if (tp->srtt_us) {
2590 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2591 if (tp->packets_out == 1)
2592 timeout += TCP_RTO_MIN;
2593 else
2594 timeout += TCP_TIMEOUT_MIN;
2595 } else {
2596 timeout = TCP_TIMEOUT_INIT;
2597 }
2598
2599 /* If the RTO formula yields an earlier time, then use that time. */
2600 rto_delta_us = advancing_rto ?
2601 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2602 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2603 if (rto_delta_us > 0)
2604 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2605
2606 tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX);
2607 return true;
2608}
2609
2610/* Thanks to skb fast clones, we can detect if a prior transmit of
2611 * a packet is still in a qdisc or driver queue.
2612 * In this case, there is very little point doing a retransmit !
2613 */
2614static bool skb_still_in_host_queue(const struct sock *sk,
2615 const struct sk_buff *skb)
2616{
2617 if (unlikely(skb_fclone_busy(sk, skb))) {
2618 NET_INC_STATS(sock_net(sk),
2619 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2620 return true;
2621 }
2622 return false;
2623}
2624
2625/* When probe timeout (PTO) fires, try send a new segment if possible, else
2626 * retransmit the last segment.
2627 */
2628void tcp_send_loss_probe(struct sock *sk)
2629{
2630 struct tcp_sock *tp = tcp_sk(sk);
2631 struct sk_buff *skb;
2632 int pcount;
2633 int mss = tcp_current_mss(sk);
2634
2635 /* At most one outstanding TLP */
2636 if (tp->tlp_high_seq)
2637 goto rearm_timer;
2638
2639 tp->tlp_retrans = 0;
2640 skb = tcp_send_head(sk);
2641 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2642 pcount = tp->packets_out;
2643 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2644 if (tp->packets_out > pcount)
2645 goto probe_sent;
2646 goto rearm_timer;
2647 }
2648 skb = skb_rb_last(&sk->tcp_rtx_queue);
2649 if (unlikely(!skb)) {
2650 WARN_ONCE(tp->packets_out,
2651 "invalid inflight: %u state %u cwnd %u mss %d\n",
2652 tp->packets_out, sk->sk_state, tp->snd_cwnd, mss);
2653 inet_csk(sk)->icsk_pending = 0;
2654 return;
2655 }
2656
2657 if (skb_still_in_host_queue(sk, skb))
2658 goto rearm_timer;
2659
2660 pcount = tcp_skb_pcount(skb);
2661 if (WARN_ON(!pcount))
2662 goto rearm_timer;
2663
2664 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2665 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2666 (pcount - 1) * mss, mss,
2667 GFP_ATOMIC)))
2668 goto rearm_timer;
2669 skb = skb_rb_next(skb);
2670 }
2671
2672 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2673 goto rearm_timer;
2674
2675 if (__tcp_retransmit_skb(sk, skb, 1))
2676 goto rearm_timer;
2677
2678 tp->tlp_retrans = 1;
2679
2680probe_sent:
2681 /* Record snd_nxt for loss detection. */
2682 tp->tlp_high_seq = tp->snd_nxt;
2683
2684 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2685 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2686 inet_csk(sk)->icsk_pending = 0;
2687rearm_timer:
2688 tcp_rearm_rto(sk);
2689}
2690
2691/* Push out any pending frames which were held back due to
2692 * TCP_CORK or attempt at coalescing tiny packets.
2693 * The socket must be locked by the caller.
2694 */
2695void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2696 int nonagle)
2697{
2698 /* If we are closed, the bytes will have to remain here.
2699 * In time closedown will finish, we empty the write queue and
2700 * all will be happy.
2701 */
2702 if (unlikely(sk->sk_state == TCP_CLOSE))
2703 return;
2704
2705 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2706 sk_gfp_mask(sk, GFP_ATOMIC)))
2707 tcp_check_probe_timer(sk);
2708}
2709
2710/* Send _single_ skb sitting at the send head. This function requires
2711 * true push pending frames to setup probe timer etc.
2712 */
2713void tcp_push_one(struct sock *sk, unsigned int mss_now)
2714{
2715 struct sk_buff *skb = tcp_send_head(sk);
2716
2717 BUG_ON(!skb || skb->len < mss_now);
2718
2719 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2720}
2721
2722/* This function returns the amount that we can raise the
2723 * usable window based on the following constraints
2724 *
2725 * 1. The window can never be shrunk once it is offered (RFC 793)
2726 * 2. We limit memory per socket
2727 *
2728 * RFC 1122:
2729 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2730 * RECV.NEXT + RCV.WIN fixed until:
2731 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2732 *
2733 * i.e. don't raise the right edge of the window until you can raise
2734 * it at least MSS bytes.
2735 *
2736 * Unfortunately, the recommended algorithm breaks header prediction,
2737 * since header prediction assumes th->window stays fixed.
2738 *
2739 * Strictly speaking, keeping th->window fixed violates the receiver
2740 * side SWS prevention criteria. The problem is that under this rule
2741 * a stream of single byte packets will cause the right side of the
2742 * window to always advance by a single byte.
2743 *
2744 * Of course, if the sender implements sender side SWS prevention
2745 * then this will not be a problem.
2746 *
2747 * BSD seems to make the following compromise:
2748 *
2749 * If the free space is less than the 1/4 of the maximum
2750 * space available and the free space is less than 1/2 mss,
2751 * then set the window to 0.
2752 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2753 * Otherwise, just prevent the window from shrinking
2754 * and from being larger than the largest representable value.
2755 *
2756 * This prevents incremental opening of the window in the regime
2757 * where TCP is limited by the speed of the reader side taking
2758 * data out of the TCP receive queue. It does nothing about
2759 * those cases where the window is constrained on the sender side
2760 * because the pipeline is full.
2761 *
2762 * BSD also seems to "accidentally" limit itself to windows that are a
2763 * multiple of MSS, at least until the free space gets quite small.
2764 * This would appear to be a side effect of the mbuf implementation.
2765 * Combining these two algorithms results in the observed behavior
2766 * of having a fixed window size at almost all times.
2767 *
2768 * Below we obtain similar behavior by forcing the offered window to
2769 * a multiple of the mss when it is feasible to do so.
2770 *
2771 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2772 * Regular options like TIMESTAMP are taken into account.
2773 */
2774u32 __tcp_select_window(struct sock *sk)
2775{
2776 struct inet_connection_sock *icsk = inet_csk(sk);
2777 struct tcp_sock *tp = tcp_sk(sk);
2778 /* MSS for the peer's data. Previous versions used mss_clamp
2779 * here. I don't know if the value based on our guesses
2780 * of peer's MSS is better for the performance. It's more correct
2781 * but may be worse for the performance because of rcv_mss
2782 * fluctuations. --SAW 1998/11/1
2783 */
2784 int mss = icsk->icsk_ack.rcv_mss;
2785 int free_space = tcp_space(sk);
2786 int allowed_space = tcp_full_space(sk);
2787 int full_space, window;
2788
2789 if (sk_is_mptcp(sk))
2790 mptcp_space(sk, &free_space, &allowed_space);
2791
2792 full_space = min_t(int, tp->window_clamp, allowed_space);
2793
2794 if (unlikely(mss > full_space)) {
2795 mss = full_space;
2796 if (mss <= 0)
2797 return 0;
2798 }
2799 if (free_space < (full_space >> 1)) {
2800 icsk->icsk_ack.quick = 0;
2801
2802 if (tcp_under_memory_pressure(sk))
2803 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2804 4U * tp->advmss);
2805
2806 /* free_space might become our new window, make sure we don't
2807 * increase it due to wscale.
2808 */
2809 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2810
2811 /* if free space is less than mss estimate, or is below 1/16th
2812 * of the maximum allowed, try to move to zero-window, else
2813 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2814 * new incoming data is dropped due to memory limits.
2815 * With large window, mss test triggers way too late in order
2816 * to announce zero window in time before rmem limit kicks in.
2817 */
2818 if (free_space < (allowed_space >> 4) || free_space < mss)
2819 return 0;
2820 }
2821
2822 if (free_space > tp->rcv_ssthresh)
2823 free_space = tp->rcv_ssthresh;
2824
2825 /* Don't do rounding if we are using window scaling, since the
2826 * scaled window will not line up with the MSS boundary anyway.
2827 */
2828 if (tp->rx_opt.rcv_wscale) {
2829 window = free_space;
2830
2831 /* Advertise enough space so that it won't get scaled away.
2832 * Import case: prevent zero window announcement if
2833 * 1<<rcv_wscale > mss.
2834 */
2835 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2836 } else {
2837 window = tp->rcv_wnd;
2838 /* Get the largest window that is a nice multiple of mss.
2839 * Window clamp already applied above.
2840 * If our current window offering is within 1 mss of the
2841 * free space we just keep it. This prevents the divide
2842 * and multiply from happening most of the time.
2843 * We also don't do any window rounding when the free space
2844 * is too small.
2845 */
2846 if (window <= free_space - mss || window > free_space)
2847 window = rounddown(free_space, mss);
2848 else if (mss == full_space &&
2849 free_space > window + (full_space >> 1))
2850 window = free_space;
2851 }
2852
2853 return window;
2854}
2855
2856void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2857 const struct sk_buff *next_skb)
2858{
2859 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2860 const struct skb_shared_info *next_shinfo =
2861 skb_shinfo(next_skb);
2862 struct skb_shared_info *shinfo = skb_shinfo(skb);
2863
2864 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2865 shinfo->tskey = next_shinfo->tskey;
2866 TCP_SKB_CB(skb)->txstamp_ack |=
2867 TCP_SKB_CB(next_skb)->txstamp_ack;
2868 }
2869}
2870
2871/* Collapses two adjacent SKB's during retransmission. */
2872static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2873{
2874 struct tcp_sock *tp = tcp_sk(sk);
2875 struct sk_buff *next_skb = skb_rb_next(skb);
2876 int next_skb_size;
2877
2878 next_skb_size = next_skb->len;
2879
2880 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2881
2882 if (next_skb_size) {
2883 if (next_skb_size <= skb_availroom(skb))
2884 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2885 next_skb_size);
2886 else if (!tcp_skb_shift(skb, next_skb, 1, next_skb_size))
2887 return false;
2888 }
2889 tcp_highest_sack_replace(sk, next_skb, skb);
2890
2891 /* Update sequence range on original skb. */
2892 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2893
2894 /* Merge over control information. This moves PSH/FIN etc. over */
2895 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2896
2897 /* All done, get rid of second SKB and account for it so
2898 * packet counting does not break.
2899 */
2900 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2901 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2902
2903 /* changed transmit queue under us so clear hints */
2904 tcp_clear_retrans_hints_partial(tp);
2905 if (next_skb == tp->retransmit_skb_hint)
2906 tp->retransmit_skb_hint = skb;
2907
2908 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2909
2910 tcp_skb_collapse_tstamp(skb, next_skb);
2911
2912 tcp_rtx_queue_unlink_and_free(next_skb, sk);
2913 return true;
2914}
2915
2916/* Check if coalescing SKBs is legal. */
2917static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2918{
2919 if (tcp_skb_pcount(skb) > 1)
2920 return false;
2921 if (skb_cloned(skb))
2922 return false;
2923 /* Some heuristics for collapsing over SACK'd could be invented */
2924 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2925 return false;
2926
2927 return true;
2928}
2929
2930/* Collapse packets in the retransmit queue to make to create
2931 * less packets on the wire. This is only done on retransmission.
2932 */
2933static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2934 int space)
2935{
2936 struct tcp_sock *tp = tcp_sk(sk);
2937 struct sk_buff *skb = to, *tmp;
2938 bool first = true;
2939
2940 if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
2941 return;
2942 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2943 return;
2944
2945 skb_rbtree_walk_from_safe(skb, tmp) {
2946 if (!tcp_can_collapse(sk, skb))
2947 break;
2948
2949 if (!tcp_skb_can_collapse(to, skb))
2950 break;
2951
2952 space -= skb->len;
2953
2954 if (first) {
2955 first = false;
2956 continue;
2957 }
2958
2959 if (space < 0)
2960 break;
2961
2962 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2963 break;
2964
2965 if (!tcp_collapse_retrans(sk, to))
2966 break;
2967 }
2968}
2969
2970/* This retransmits one SKB. Policy decisions and retransmit queue
2971 * state updates are done by the caller. Returns non-zero if an
2972 * error occurred which prevented the send.
2973 */
2974int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2975{
2976 struct inet_connection_sock *icsk = inet_csk(sk);
2977 struct tcp_sock *tp = tcp_sk(sk);
2978 unsigned int cur_mss;
2979 int diff, len, err;
2980
2981
2982 /* Inconclusive MTU probe */
2983 if (icsk->icsk_mtup.probe_size)
2984 icsk->icsk_mtup.probe_size = 0;
2985
2986 /* Do not sent more than we queued. 1/4 is reserved for possible
2987 * copying overhead: fragmentation, tunneling, mangling etc.
2988 */
2989 if (refcount_read(&sk->sk_wmem_alloc) >
2990 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2991 sk->sk_sndbuf))
2992 return -EAGAIN;
2993
2994 if (skb_still_in_host_queue(sk, skb))
2995 return -EBUSY;
2996
2997 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2998 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
2999 WARN_ON_ONCE(1);
3000 return -EINVAL;
3001 }
3002 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3003 return -ENOMEM;
3004 }
3005
3006 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3007 return -EHOSTUNREACH; /* Routing failure or similar. */
3008
3009 cur_mss = tcp_current_mss(sk);
3010
3011 /* If receiver has shrunk his window, and skb is out of
3012 * new window, do not retransmit it. The exception is the
3013 * case, when window is shrunk to zero. In this case
3014 * our retransmit serves as a zero window probe.
3015 */
3016 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
3017 TCP_SKB_CB(skb)->seq != tp->snd_una)
3018 return -EAGAIN;
3019
3020 len = cur_mss * segs;
3021 if (skb->len > len) {
3022 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
3023 cur_mss, GFP_ATOMIC))
3024 return -ENOMEM; /* We'll try again later. */
3025 } else {
3026 if (skb_unclone(skb, GFP_ATOMIC))
3027 return -ENOMEM;
3028
3029 diff = tcp_skb_pcount(skb);
3030 tcp_set_skb_tso_segs(skb, cur_mss);
3031 diff -= tcp_skb_pcount(skb);
3032 if (diff)
3033 tcp_adjust_pcount(sk, skb, diff);
3034 if (skb->len < cur_mss)
3035 tcp_retrans_try_collapse(sk, skb, cur_mss);
3036 }
3037
3038 /* RFC3168, section 6.1.1.1. ECN fallback */
3039 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
3040 tcp_ecn_clear_syn(sk, skb);
3041
3042 /* Update global and local TCP statistics. */
3043 segs = tcp_skb_pcount(skb);
3044 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
3045 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3046 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3047 tp->total_retrans += segs;
3048 tp->bytes_retrans += skb->len;
3049
3050 /* make sure skb->data is aligned on arches that require it
3051 * and check if ack-trimming & collapsing extended the headroom
3052 * beyond what csum_start can cover.
3053 */
3054 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
3055 skb_headroom(skb) >= 0xFFFF)) {
3056 struct sk_buff *nskb;
3057
3058 tcp_skb_tsorted_save(skb) {
3059 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
3060 if (nskb) {
3061 nskb->dev = NULL;
3062 err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC);
3063 } else {
3064 err = -ENOBUFS;
3065 }
3066 } tcp_skb_tsorted_restore(skb);
3067
3068 if (!err) {
3069 tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
3070 tcp_rate_skb_sent(sk, skb);
3071 }
3072 } else {
3073 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3074 }
3075
3076 /* To avoid taking spuriously low RTT samples based on a timestamp
3077 * for a transmit that never happened, always mark EVER_RETRANS
3078 */
3079 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
3080
3081 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
3082 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
3083 TCP_SKB_CB(skb)->seq, segs, err);
3084
3085 if (likely(!err)) {
3086 trace_tcp_retransmit_skb(sk, skb);
3087 } else if (err != -EBUSY) {
3088 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
3089 }
3090 return err;
3091}
3092
3093int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3094{
3095 struct tcp_sock *tp = tcp_sk(sk);
3096 int err = __tcp_retransmit_skb(sk, skb, segs);
3097
3098 if (err == 0) {
3099#if FASTRETRANS_DEBUG > 0
3100 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3101 net_dbg_ratelimited("retrans_out leaked\n");
3102 }
3103#endif
3104 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3105 tp->retrans_out += tcp_skb_pcount(skb);
3106 }
3107
3108 /* Save stamp of the first (attempted) retransmit. */
3109 if (!tp->retrans_stamp)
3110 tp->retrans_stamp = tcp_skb_timestamp(skb);
3111
3112 if (tp->undo_retrans < 0)
3113 tp->undo_retrans = 0;
3114 tp->undo_retrans += tcp_skb_pcount(skb);
3115 return err;
3116}
3117
3118/* This gets called after a retransmit timeout, and the initially
3119 * retransmitted data is acknowledged. It tries to continue
3120 * resending the rest of the retransmit queue, until either
3121 * we've sent it all or the congestion window limit is reached.
3122 */
3123void tcp_xmit_retransmit_queue(struct sock *sk)
3124{
3125 const struct inet_connection_sock *icsk = inet_csk(sk);
3126 struct sk_buff *skb, *rtx_head, *hole = NULL;
3127 struct tcp_sock *tp = tcp_sk(sk);
3128 bool rearm_timer = false;
3129 u32 max_segs;
3130 int mib_idx;
3131
3132 if (!tp->packets_out)
3133 return;
3134
3135 rtx_head = tcp_rtx_queue_head(sk);
3136 skb = tp->retransmit_skb_hint ?: rtx_head;
3137 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3138 skb_rbtree_walk_from(skb) {
3139 __u8 sacked;
3140 int segs;
3141
3142 if (tcp_pacing_check(sk))
3143 break;
3144
3145 /* we could do better than to assign each time */
3146 if (!hole)
3147 tp->retransmit_skb_hint = skb;
3148
3149 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
3150 if (segs <= 0)
3151 break;
3152 sacked = TCP_SKB_CB(skb)->sacked;
3153 /* In case tcp_shift_skb_data() have aggregated large skbs,
3154 * we need to make sure not sending too bigs TSO packets
3155 */
3156 segs = min_t(int, segs, max_segs);
3157
3158 if (tp->retrans_out >= tp->lost_out) {
3159 break;
3160 } else if (!(sacked & TCPCB_LOST)) {
3161 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3162 hole = skb;
3163 continue;
3164
3165 } else {
3166 if (icsk->icsk_ca_state != TCP_CA_Loss)
3167 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3168 else
3169 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3170 }
3171
3172 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3173 continue;
3174
3175 if (tcp_small_queue_check(sk, skb, 1))
3176 break;
3177
3178 if (tcp_retransmit_skb(sk, skb, segs))
3179 break;
3180
3181 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3182
3183 if (tcp_in_cwnd_reduction(sk))
3184 tp->prr_out += tcp_skb_pcount(skb);
3185
3186 if (skb == rtx_head &&
3187 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3188 rearm_timer = true;
3189
3190 }
3191 if (rearm_timer)
3192 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3193 inet_csk(sk)->icsk_rto,
3194 TCP_RTO_MAX);
3195}
3196
3197/* We allow to exceed memory limits for FIN packets to expedite
3198 * connection tear down and (memory) recovery.
3199 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3200 * or even be forced to close flow without any FIN.
3201 * In general, we want to allow one skb per socket to avoid hangs
3202 * with edge trigger epoll()
3203 */
3204void sk_forced_mem_schedule(struct sock *sk, int size)
3205{
3206 int amt;
3207
3208 if (size <= sk->sk_forward_alloc)
3209 return;
3210 amt = sk_mem_pages(size);
3211 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3212 sk_memory_allocated_add(sk, amt);
3213
3214 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3215 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3216}
3217
3218/* Send a FIN. The caller locks the socket for us.
3219 * We should try to send a FIN packet really hard, but eventually give up.
3220 */
3221void tcp_send_fin(struct sock *sk)
3222{
3223 struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk);
3224 struct tcp_sock *tp = tcp_sk(sk);
3225
3226 /* Optimization, tack on the FIN if we have one skb in write queue and
3227 * this skb was not yet sent, or we are under memory pressure.
3228 * Note: in the latter case, FIN packet will be sent after a timeout,
3229 * as TCP stack thinks it has already been transmitted.
3230 */
3231 tskb = tail;
3232 if (!tskb && tcp_under_memory_pressure(sk))
3233 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3234
3235 if (tskb) {
3236 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3237 TCP_SKB_CB(tskb)->end_seq++;
3238 tp->write_seq++;
3239 if (!tail) {
3240 /* This means tskb was already sent.
3241 * Pretend we included the FIN on previous transmit.
3242 * We need to set tp->snd_nxt to the value it would have
3243 * if FIN had been sent. This is because retransmit path
3244 * does not change tp->snd_nxt.
3245 */
3246 WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3247 return;
3248 }
3249 } else {
3250 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3251 if (unlikely(!skb))
3252 return;
3253
3254 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3255 skb_reserve(skb, MAX_TCP_HEADER);
3256 sk_forced_mem_schedule(sk, skb->truesize);
3257 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3258 tcp_init_nondata_skb(skb, tp->write_seq,
3259 TCPHDR_ACK | TCPHDR_FIN);
3260 tcp_queue_skb(sk, skb);
3261 }
3262 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3263}
3264
3265/* We get here when a process closes a file descriptor (either due to
3266 * an explicit close() or as a byproduct of exit()'ing) and there
3267 * was unread data in the receive queue. This behavior is recommended
3268 * by RFC 2525, section 2.17. -DaveM
3269 */
3270void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3271{
3272 struct sk_buff *skb;
3273
3274 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3275
3276 /* NOTE: No TCP options attached and we never retransmit this. */
3277 skb = alloc_skb(MAX_TCP_HEADER, priority);
3278 if (!skb) {
3279 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3280 return;
3281 }
3282
3283 /* Reserve space for headers and prepare control bits. */
3284 skb_reserve(skb, MAX_TCP_HEADER);
3285 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3286 TCPHDR_ACK | TCPHDR_RST);
3287 tcp_mstamp_refresh(tcp_sk(sk));
3288 /* Send it off. */
3289 if (tcp_transmit_skb(sk, skb, 0, priority))
3290 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3291
3292 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3293 * skb here is different to the troublesome skb, so use NULL
3294 */
3295 trace_tcp_send_reset(sk, NULL);
3296}
3297
3298/* Send a crossed SYN-ACK during socket establishment.
3299 * WARNING: This routine must only be called when we have already sent
3300 * a SYN packet that crossed the incoming SYN that caused this routine
3301 * to get called. If this assumption fails then the initial rcv_wnd
3302 * and rcv_wscale values will not be correct.
3303 */
3304int tcp_send_synack(struct sock *sk)
3305{
3306 struct sk_buff *skb;
3307
3308 skb = tcp_rtx_queue_head(sk);
3309 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3310 pr_err("%s: wrong queue state\n", __func__);
3311 return -EFAULT;
3312 }
3313 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3314 if (skb_cloned(skb)) {
3315 struct sk_buff *nskb;
3316
3317 tcp_skb_tsorted_save(skb) {
3318 nskb = skb_copy(skb, GFP_ATOMIC);
3319 } tcp_skb_tsorted_restore(skb);
3320 if (!nskb)
3321 return -ENOMEM;
3322 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3323 tcp_highest_sack_replace(sk, skb, nskb);
3324 tcp_rtx_queue_unlink_and_free(skb, sk);
3325 __skb_header_release(nskb);
3326 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3327 sk_wmem_queued_add(sk, nskb->truesize);
3328 sk_mem_charge(sk, nskb->truesize);
3329 skb = nskb;
3330 }
3331
3332 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3333 tcp_ecn_send_synack(sk, skb);
3334 }
3335 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3336}
3337
3338/**
3339 * tcp_make_synack - Prepare a SYN-ACK.
3340 * sk: listener socket
3341 * dst: dst entry attached to the SYNACK
3342 * req: request_sock pointer
3343 * foc: cookie for tcp fast open
3344 * synack_type: Type of synback to prepare
3345 *
3346 * Allocate one skb and build a SYNACK packet.
3347 * @dst is consumed : Caller should not use it again.
3348 */
3349struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3350 struct request_sock *req,
3351 struct tcp_fastopen_cookie *foc,
3352 enum tcp_synack_type synack_type)
3353{
3354 struct inet_request_sock *ireq = inet_rsk(req);
3355 const struct tcp_sock *tp = tcp_sk(sk);
3356 struct tcp_md5sig_key *md5 = NULL;
3357 struct tcp_out_options opts;
3358 struct sk_buff *skb;
3359 int tcp_header_size;
3360 struct tcphdr *th;
3361 int mss;
3362 u64 now;
3363
3364 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3365 if (unlikely(!skb)) {
3366 dst_release(dst);
3367 return NULL;
3368 }
3369 /* Reserve space for headers. */
3370 skb_reserve(skb, MAX_TCP_HEADER);
3371
3372 switch (synack_type) {
3373 case TCP_SYNACK_NORMAL:
3374 skb_set_owner_w(skb, req_to_sk(req));
3375 break;
3376 case TCP_SYNACK_COOKIE:
3377 /* Under synflood, we do not attach skb to a socket,
3378 * to avoid false sharing.
3379 */
3380 break;
3381 case TCP_SYNACK_FASTOPEN:
3382 /* sk is a const pointer, because we want to express multiple
3383 * cpu might call us concurrently.
3384 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3385 */
3386 skb_set_owner_w(skb, (struct sock *)sk);
3387 break;
3388 }
3389 skb_dst_set(skb, dst);
3390
3391 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3392
3393 memset(&opts, 0, sizeof(opts));
3394 now = tcp_clock_ns();
3395#ifdef CONFIG_SYN_COOKIES
3396 if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok))
3397 skb->skb_mstamp_ns = cookie_init_timestamp(req, now);
3398 else
3399#endif
3400 {
3401 skb->skb_mstamp_ns = now;
3402 if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3403 tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3404 }
3405
3406#ifdef CONFIG_TCP_MD5SIG
3407 rcu_read_lock();
3408 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3409#endif
3410 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3411 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3412 foc, synack_type) + sizeof(*th);
3413
3414 skb_push(skb, tcp_header_size);
3415 skb_reset_transport_header(skb);
3416
3417 th = (struct tcphdr *)skb->data;
3418 memset(th, 0, sizeof(struct tcphdr));
3419 th->syn = 1;
3420 th->ack = 1;
3421 tcp_ecn_make_synack(req, th);
3422 th->source = htons(ireq->ir_num);
3423 th->dest = ireq->ir_rmt_port;
3424 skb->mark = ireq->ir_mark;
3425 skb->ip_summed = CHECKSUM_PARTIAL;
3426 th->seq = htonl(tcp_rsk(req)->snt_isn);
3427 /* XXX data is queued and acked as is. No buffer/window check */
3428 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3429
3430 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3431 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3432 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3433 th->doff = (tcp_header_size >> 2);
3434 __TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3435
3436#ifdef CONFIG_TCP_MD5SIG
3437 /* Okay, we have all we need - do the md5 hash if needed */
3438 if (md5)
3439 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3440 md5, req_to_sk(req), skb);
3441 rcu_read_unlock();
3442#endif
3443
3444 skb->skb_mstamp_ns = now;
3445 tcp_add_tx_delay(skb, tp);
3446
3447 return skb;
3448}
3449EXPORT_SYMBOL(tcp_make_synack);
3450
3451static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3452{
3453 struct inet_connection_sock *icsk = inet_csk(sk);
3454 const struct tcp_congestion_ops *ca;
3455 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3456
3457 if (ca_key == TCP_CA_UNSPEC)
3458 return;
3459
3460 rcu_read_lock();
3461 ca = tcp_ca_find_key(ca_key);
3462 if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
3463 bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
3464 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3465 icsk->icsk_ca_ops = ca;
3466 }
3467 rcu_read_unlock();
3468}
3469
3470/* Do all connect socket setups that can be done AF independent. */
3471static void tcp_connect_init(struct sock *sk)
3472{
3473 const struct dst_entry *dst = __sk_dst_get(sk);
3474 struct tcp_sock *tp = tcp_sk(sk);
3475 __u8 rcv_wscale;
3476 u32 rcv_wnd;
3477
3478 /* We'll fix this up when we get a response from the other end.
3479 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3480 */
3481 tp->tcp_header_len = sizeof(struct tcphdr);
3482 if (sock_net(sk)->ipv4.sysctl_tcp_timestamps)
3483 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3484
3485#ifdef CONFIG_TCP_MD5SIG
3486 if (tp->af_specific->md5_lookup(sk, sk))
3487 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3488#endif
3489
3490 /* If user gave his TCP_MAXSEG, record it to clamp */
3491 if (tp->rx_opt.user_mss)
3492 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3493 tp->max_window = 0;
3494 tcp_mtup_init(sk);
3495 tcp_sync_mss(sk, dst_mtu(dst));
3496
3497 tcp_ca_dst_init(sk, dst);
3498
3499 if (!tp->window_clamp)
3500 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3501 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3502
3503 tcp_initialize_rcv_mss(sk);
3504
3505 /* limit the window selection if the user enforce a smaller rx buffer */
3506 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3507 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3508 tp->window_clamp = tcp_full_space(sk);
3509
3510 rcv_wnd = tcp_rwnd_init_bpf(sk);
3511 if (rcv_wnd == 0)
3512 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3513
3514 tcp_select_initial_window(sk, tcp_full_space(sk),
3515 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3516 &tp->rcv_wnd,
3517 &tp->window_clamp,
3518 sock_net(sk)->ipv4.sysctl_tcp_window_scaling,
3519 &rcv_wscale,
3520 rcv_wnd);
3521
3522 tp->rx_opt.rcv_wscale = rcv_wscale;
3523 tp->rcv_ssthresh = tp->rcv_wnd;
3524
3525 sk->sk_err = 0;
3526 sock_reset_flag(sk, SOCK_DONE);
3527 tp->snd_wnd = 0;
3528 tcp_init_wl(tp, 0);
3529 tcp_write_queue_purge(sk);
3530 tp->snd_una = tp->write_seq;
3531 tp->snd_sml = tp->write_seq;
3532 tp->snd_up = tp->write_seq;
3533 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3534
3535 if (likely(!tp->repair))
3536 tp->rcv_nxt = 0;
3537 else
3538 tp->rcv_tstamp = tcp_jiffies32;
3539 tp->rcv_wup = tp->rcv_nxt;
3540 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3541
3542 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3543 inet_csk(sk)->icsk_retransmits = 0;
3544 tcp_clear_retrans(tp);
3545}
3546
3547static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3548{
3549 struct tcp_sock *tp = tcp_sk(sk);
3550 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3551
3552 tcb->end_seq += skb->len;
3553 __skb_header_release(skb);
3554 sk_wmem_queued_add(sk, skb->truesize);
3555 sk_mem_charge(sk, skb->truesize);
3556 WRITE_ONCE(tp->write_seq, tcb->end_seq);
3557 tp->packets_out += tcp_skb_pcount(skb);
3558}
3559
3560/* Build and send a SYN with data and (cached) Fast Open cookie. However,
3561 * queue a data-only packet after the regular SYN, such that regular SYNs
3562 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3563 * only the SYN sequence, the data are retransmitted in the first ACK.
3564 * If cookie is not cached or other error occurs, falls back to send a
3565 * regular SYN with Fast Open cookie request option.
3566 */
3567static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3568{
3569 struct tcp_sock *tp = tcp_sk(sk);
3570 struct tcp_fastopen_request *fo = tp->fastopen_req;
3571 int space, err = 0;
3572 struct sk_buff *syn_data;
3573
3574 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3575 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3576 goto fallback;
3577
3578 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3579 * user-MSS. Reserve maximum option space for middleboxes that add
3580 * private TCP options. The cost is reduced data space in SYN :(
3581 */
3582 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3583
3584 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3585 MAX_TCP_OPTION_SPACE;
3586
3587 space = min_t(size_t, space, fo->size);
3588
3589 /* limit to order-0 allocations */
3590 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3591
3592 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3593 if (!syn_data)
3594 goto fallback;
3595 syn_data->ip_summed = CHECKSUM_PARTIAL;
3596 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3597 if (space) {
3598 int copied = copy_from_iter(skb_put(syn_data, space), space,
3599 &fo->data->msg_iter);
3600 if (unlikely(!copied)) {
3601 tcp_skb_tsorted_anchor_cleanup(syn_data);
3602 kfree_skb(syn_data);
3603 goto fallback;
3604 }
3605 if (copied != space) {
3606 skb_trim(syn_data, copied);
3607 space = copied;
3608 }
3609 skb_zcopy_set(syn_data, fo->uarg, NULL);
3610 }
3611 /* No more data pending in inet_wait_for_connect() */
3612 if (space == fo->size)
3613 fo->data = NULL;
3614 fo->copied = space;
3615
3616 tcp_connect_queue_skb(sk, syn_data);
3617 if (syn_data->len)
3618 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3619
3620 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3621
3622 syn->skb_mstamp_ns = syn_data->skb_mstamp_ns;
3623
3624 /* Now full SYN+DATA was cloned and sent (or not),
3625 * remove the SYN from the original skb (syn_data)
3626 * we keep in write queue in case of a retransmit, as we
3627 * also have the SYN packet (with no data) in the same queue.
3628 */
3629 TCP_SKB_CB(syn_data)->seq++;
3630 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3631 if (!err) {
3632 tp->syn_data = (fo->copied > 0);
3633 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
3634 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3635 goto done;
3636 }
3637
3638 /* data was not sent, put it in write_queue */
3639 __skb_queue_tail(&sk->sk_write_queue, syn_data);
3640 tp->packets_out -= tcp_skb_pcount(syn_data);
3641
3642fallback:
3643 /* Send a regular SYN with Fast Open cookie request option */
3644 if (fo->cookie.len > 0)
3645 fo->cookie.len = 0;
3646 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3647 if (err)
3648 tp->syn_fastopen = 0;
3649done:
3650 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3651 return err;
3652}
3653
3654/* Build a SYN and send it off. */
3655int tcp_connect(struct sock *sk)
3656{
3657 struct tcp_sock *tp = tcp_sk(sk);
3658 struct sk_buff *buff;
3659 int err;
3660
3661 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
3662
3663 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3664 return -EHOSTUNREACH; /* Routing failure or similar. */
3665
3666 tcp_connect_init(sk);
3667
3668 if (unlikely(tp->repair)) {
3669 tcp_finish_connect(sk, NULL);
3670 return 0;
3671 }
3672
3673 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3674 if (unlikely(!buff))
3675 return -ENOBUFS;
3676
3677 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3678 tcp_mstamp_refresh(tp);
3679 tp->retrans_stamp = tcp_time_stamp(tp);
3680 tcp_connect_queue_skb(sk, buff);
3681 tcp_ecn_send_syn(sk, buff);
3682 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
3683
3684 /* Send off SYN; include data in Fast Open. */
3685 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3686 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3687 if (err == -ECONNREFUSED)
3688 return err;
3689
3690 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3691 * in order to make this packet get counted in tcpOutSegs.
3692 */
3693 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3694 tp->pushed_seq = tp->write_seq;
3695 buff = tcp_send_head(sk);
3696 if (unlikely(buff)) {
3697 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
3698 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
3699 }
3700 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3701
3702 /* Timer for repeating the SYN until an answer. */
3703 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3704 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3705 return 0;
3706}
3707EXPORT_SYMBOL(tcp_connect);
3708
3709/* Send out a delayed ack, the caller does the policy checking
3710 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3711 * for details.
3712 */
3713void tcp_send_delayed_ack(struct sock *sk)
3714{
3715 struct inet_connection_sock *icsk = inet_csk(sk);
3716 int ato = icsk->icsk_ack.ato;
3717 unsigned long timeout;
3718
3719 if (ato > TCP_DELACK_MIN) {
3720 const struct tcp_sock *tp = tcp_sk(sk);
3721 int max_ato = HZ / 2;
3722
3723 if (inet_csk_in_pingpong_mode(sk) ||
3724 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3725 max_ato = TCP_DELACK_MAX;
3726
3727 /* Slow path, intersegment interval is "high". */
3728
3729 /* If some rtt estimate is known, use it to bound delayed ack.
3730 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3731 * directly.
3732 */
3733 if (tp->srtt_us) {
3734 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3735 TCP_DELACK_MIN);
3736
3737 if (rtt < max_ato)
3738 max_ato = rtt;
3739 }
3740
3741 ato = min(ato, max_ato);
3742 }
3743
3744 /* Stay within the limit we were given */
3745 timeout = jiffies + ato;
3746
3747 /* Use new timeout only if there wasn't a older one earlier. */
3748 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3749 /* If delack timer was blocked or is about to expire,
3750 * send ACK now.
3751 */
3752 if (icsk->icsk_ack.blocked ||
3753 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3754 tcp_send_ack(sk);
3755 return;
3756 }
3757
3758 if (!time_before(timeout, icsk->icsk_ack.timeout))
3759 timeout = icsk->icsk_ack.timeout;
3760 }
3761 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3762 icsk->icsk_ack.timeout = timeout;
3763 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3764}
3765
3766/* This routine sends an ack and also updates the window. */
3767void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
3768{
3769 struct sk_buff *buff;
3770
3771 /* If we have been reset, we may not send again. */
3772 if (sk->sk_state == TCP_CLOSE)
3773 return;
3774
3775 /* We are not putting this on the write queue, so
3776 * tcp_transmit_skb() will set the ownership to this
3777 * sock.
3778 */
3779 buff = alloc_skb(MAX_TCP_HEADER,
3780 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3781 if (unlikely(!buff)) {
3782 inet_csk_schedule_ack(sk);
3783 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3784 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3785 TCP_DELACK_MAX, TCP_RTO_MAX);
3786 return;
3787 }
3788
3789 /* Reserve space for headers and prepare control bits. */
3790 skb_reserve(buff, MAX_TCP_HEADER);
3791 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3792
3793 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3794 * too much.
3795 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3796 */
3797 skb_set_tcp_pure_ack(buff);
3798
3799 /* Send it off, this clears delayed acks for us. */
3800 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
3801}
3802EXPORT_SYMBOL_GPL(__tcp_send_ack);
3803
3804void tcp_send_ack(struct sock *sk)
3805{
3806 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
3807}
3808
3809/* This routine sends a packet with an out of date sequence
3810 * number. It assumes the other end will try to ack it.
3811 *
3812 * Question: what should we make while urgent mode?
3813 * 4.4BSD forces sending single byte of data. We cannot send
3814 * out of window data, because we have SND.NXT==SND.MAX...
3815 *
3816 * Current solution: to send TWO zero-length segments in urgent mode:
3817 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3818 * out-of-date with SND.UNA-1 to probe window.
3819 */
3820static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3821{
3822 struct tcp_sock *tp = tcp_sk(sk);
3823 struct sk_buff *skb;
3824
3825 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3826 skb = alloc_skb(MAX_TCP_HEADER,
3827 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3828 if (!skb)
3829 return -1;
3830
3831 /* Reserve space for headers and set control bits. */
3832 skb_reserve(skb, MAX_TCP_HEADER);
3833 /* Use a previous sequence. This should cause the other
3834 * end to send an ack. Don't queue or clone SKB, just
3835 * send it.
3836 */
3837 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3838 NET_INC_STATS(sock_net(sk), mib);
3839 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3840}
3841
3842/* Called from setsockopt( ... TCP_REPAIR ) */
3843void tcp_send_window_probe(struct sock *sk)
3844{
3845 if (sk->sk_state == TCP_ESTABLISHED) {
3846 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3847 tcp_mstamp_refresh(tcp_sk(sk));
3848 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3849 }
3850}
3851
3852/* Initiate keepalive or window probe from timer. */
3853int tcp_write_wakeup(struct sock *sk, int mib)
3854{
3855 struct tcp_sock *tp = tcp_sk(sk);
3856 struct sk_buff *skb;
3857
3858 if (sk->sk_state == TCP_CLOSE)
3859 return -1;
3860
3861 skb = tcp_send_head(sk);
3862 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3863 int err;
3864 unsigned int mss = tcp_current_mss(sk);
3865 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3866
3867 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3868 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3869
3870 /* We are probing the opening of a window
3871 * but the window size is != 0
3872 * must have been a result SWS avoidance ( sender )
3873 */
3874 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3875 skb->len > mss) {
3876 seg_size = min(seg_size, mss);
3877 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3878 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
3879 skb, seg_size, mss, GFP_ATOMIC))
3880 return -1;
3881 } else if (!tcp_skb_pcount(skb))
3882 tcp_set_skb_tso_segs(skb, mss);
3883
3884 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3885 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3886 if (!err)
3887 tcp_event_new_data_sent(sk, skb);
3888 return err;
3889 } else {
3890 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3891 tcp_xmit_probe_skb(sk, 1, mib);
3892 return tcp_xmit_probe_skb(sk, 0, mib);
3893 }
3894}
3895
3896/* A window probe timeout has occurred. If window is not closed send
3897 * a partial packet else a zero probe.
3898 */
3899void tcp_send_probe0(struct sock *sk)
3900{
3901 struct inet_connection_sock *icsk = inet_csk(sk);
3902 struct tcp_sock *tp = tcp_sk(sk);
3903 struct net *net = sock_net(sk);
3904 unsigned long timeout;
3905 int err;
3906
3907 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3908
3909 if (tp->packets_out || tcp_write_queue_empty(sk)) {
3910 /* Cancel probe timer, if it is not required. */
3911 icsk->icsk_probes_out = 0;
3912 icsk->icsk_backoff = 0;
3913 return;
3914 }
3915
3916 icsk->icsk_probes_out++;
3917 if (err <= 0) {
3918 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3919 icsk->icsk_backoff++;
3920 timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
3921 } else {
3922 /* If packet was not sent due to local congestion,
3923 * Let senders fight for local resources conservatively.
3924 */
3925 timeout = TCP_RESOURCE_PROBE_INTERVAL;
3926 }
3927 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX);
3928}
3929
3930int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3931{
3932 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3933 struct flowi fl;
3934 int res;
3935
3936 tcp_rsk(req)->txhash = net_tx_rndhash();
3937 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL);
3938 if (!res) {
3939 __TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
3940 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3941 if (unlikely(tcp_passive_fastopen(sk)))
3942 tcp_sk(sk)->total_retrans++;
3943 trace_tcp_retransmit_synack(sk, req);
3944 }
3945 return res;
3946}
3947EXPORT_SYMBOL(tcp_rtx_synack);
1/*
2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
8 * Authors: Ross Biro
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
19 */
20
21/*
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
25 * : AF independence
26 *
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
34 *
35 */
36
37#define pr_fmt(fmt) "TCP: " fmt
38
39#include <net/tcp.h>
40
41#include <linux/compiler.h>
42#include <linux/gfp.h>
43#include <linux/module.h>
44
45/* People can turn this off for buggy TCP's found in printers etc. */
46int sysctl_tcp_retrans_collapse __read_mostly = 1;
47
48/* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
50 */
51int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
52
53/* This limits the percentage of the congestion window which we
54 * will allow a single TSO frame to consume. Building TSO frames
55 * which are too large can cause TCP streams to be bursty.
56 */
57int sysctl_tcp_tso_win_divisor __read_mostly = 3;
58
59int sysctl_tcp_mtu_probing __read_mostly = 0;
60int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS;
61
62/* By default, RFC2861 behavior. */
63int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
64
65int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */
66EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size);
67
68
69/* Account for new data that has been sent to the network. */
70static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
71{
72 struct tcp_sock *tp = tcp_sk(sk);
73 unsigned int prior_packets = tp->packets_out;
74
75 tcp_advance_send_head(sk, skb);
76 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
77
78 /* Don't override Nagle indefinitely with F-RTO */
79 if (tp->frto_counter == 2)
80 tp->frto_counter = 3;
81
82 tp->packets_out += tcp_skb_pcount(skb);
83 if (!prior_packets || tp->early_retrans_delayed)
84 tcp_rearm_rto(sk);
85}
86
87/* SND.NXT, if window was not shrunk.
88 * If window has been shrunk, what should we make? It is not clear at all.
89 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
90 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
91 * invalid. OK, let's make this for now:
92 */
93static inline __u32 tcp_acceptable_seq(const struct sock *sk)
94{
95 const struct tcp_sock *tp = tcp_sk(sk);
96
97 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
98 return tp->snd_nxt;
99 else
100 return tcp_wnd_end(tp);
101}
102
103/* Calculate mss to advertise in SYN segment.
104 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
105 *
106 * 1. It is independent of path mtu.
107 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
108 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
109 * attached devices, because some buggy hosts are confused by
110 * large MSS.
111 * 4. We do not make 3, we advertise MSS, calculated from first
112 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
113 * This may be overridden via information stored in routing table.
114 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
115 * probably even Jumbo".
116 */
117static __u16 tcp_advertise_mss(struct sock *sk)
118{
119 struct tcp_sock *tp = tcp_sk(sk);
120 const struct dst_entry *dst = __sk_dst_get(sk);
121 int mss = tp->advmss;
122
123 if (dst) {
124 unsigned int metric = dst_metric_advmss(dst);
125
126 if (metric < mss) {
127 mss = metric;
128 tp->advmss = mss;
129 }
130 }
131
132 return (__u16)mss;
133}
134
135/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
136 * This is the first part of cwnd validation mechanism. */
137static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst)
138{
139 struct tcp_sock *tp = tcp_sk(sk);
140 s32 delta = tcp_time_stamp - tp->lsndtime;
141 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
142 u32 cwnd = tp->snd_cwnd;
143
144 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
145
146 tp->snd_ssthresh = tcp_current_ssthresh(sk);
147 restart_cwnd = min(restart_cwnd, cwnd);
148
149 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
150 cwnd >>= 1;
151 tp->snd_cwnd = max(cwnd, restart_cwnd);
152 tp->snd_cwnd_stamp = tcp_time_stamp;
153 tp->snd_cwnd_used = 0;
154}
155
156/* Congestion state accounting after a packet has been sent. */
157static void tcp_event_data_sent(struct tcp_sock *tp,
158 struct sock *sk)
159{
160 struct inet_connection_sock *icsk = inet_csk(sk);
161 const u32 now = tcp_time_stamp;
162
163 if (sysctl_tcp_slow_start_after_idle &&
164 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
165 tcp_cwnd_restart(sk, __sk_dst_get(sk));
166
167 tp->lsndtime = now;
168
169 /* If it is a reply for ato after last received
170 * packet, enter pingpong mode.
171 */
172 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
173 icsk->icsk_ack.pingpong = 1;
174}
175
176/* Account for an ACK we sent. */
177static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
178{
179 tcp_dec_quickack_mode(sk, pkts);
180 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
181}
182
183/* Determine a window scaling and initial window to offer.
184 * Based on the assumption that the given amount of space
185 * will be offered. Store the results in the tp structure.
186 * NOTE: for smooth operation initial space offering should
187 * be a multiple of mss if possible. We assume here that mss >= 1.
188 * This MUST be enforced by all callers.
189 */
190void tcp_select_initial_window(int __space, __u32 mss,
191 __u32 *rcv_wnd, __u32 *window_clamp,
192 int wscale_ok, __u8 *rcv_wscale,
193 __u32 init_rcv_wnd)
194{
195 unsigned int space = (__space < 0 ? 0 : __space);
196
197 /* If no clamp set the clamp to the max possible scaled window */
198 if (*window_clamp == 0)
199 (*window_clamp) = (65535 << 14);
200 space = min(*window_clamp, space);
201
202 /* Quantize space offering to a multiple of mss if possible. */
203 if (space > mss)
204 space = (space / mss) * mss;
205
206 /* NOTE: offering an initial window larger than 32767
207 * will break some buggy TCP stacks. If the admin tells us
208 * it is likely we could be speaking with such a buggy stack
209 * we will truncate our initial window offering to 32K-1
210 * unless the remote has sent us a window scaling option,
211 * which we interpret as a sign the remote TCP is not
212 * misinterpreting the window field as a signed quantity.
213 */
214 if (sysctl_tcp_workaround_signed_windows)
215 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
216 else
217 (*rcv_wnd) = space;
218
219 (*rcv_wscale) = 0;
220 if (wscale_ok) {
221 /* Set window scaling on max possible window
222 * See RFC1323 for an explanation of the limit to 14
223 */
224 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
225 space = min_t(u32, space, *window_clamp);
226 while (space > 65535 && (*rcv_wscale) < 14) {
227 space >>= 1;
228 (*rcv_wscale)++;
229 }
230 }
231
232 /* Set initial window to a value enough for senders starting with
233 * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place
234 * a limit on the initial window when mss is larger than 1460.
235 */
236 if (mss > (1 << *rcv_wscale)) {
237 int init_cwnd = TCP_DEFAULT_INIT_RCVWND;
238 if (mss > 1460)
239 init_cwnd =
240 max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
241 /* when initializing use the value from init_rcv_wnd
242 * rather than the default from above
243 */
244 if (init_rcv_wnd)
245 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
246 else
247 *rcv_wnd = min(*rcv_wnd, init_cwnd * mss);
248 }
249
250 /* Set the clamp no higher than max representable value */
251 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
252}
253EXPORT_SYMBOL(tcp_select_initial_window);
254
255/* Chose a new window to advertise, update state in tcp_sock for the
256 * socket, and return result with RFC1323 scaling applied. The return
257 * value can be stuffed directly into th->window for an outgoing
258 * frame.
259 */
260static u16 tcp_select_window(struct sock *sk)
261{
262 struct tcp_sock *tp = tcp_sk(sk);
263 u32 cur_win = tcp_receive_window(tp);
264 u32 new_win = __tcp_select_window(sk);
265
266 /* Never shrink the offered window */
267 if (new_win < cur_win) {
268 /* Danger Will Robinson!
269 * Don't update rcv_wup/rcv_wnd here or else
270 * we will not be able to advertise a zero
271 * window in time. --DaveM
272 *
273 * Relax Will Robinson.
274 */
275 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
276 }
277 tp->rcv_wnd = new_win;
278 tp->rcv_wup = tp->rcv_nxt;
279
280 /* Make sure we do not exceed the maximum possible
281 * scaled window.
282 */
283 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
284 new_win = min(new_win, MAX_TCP_WINDOW);
285 else
286 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
287
288 /* RFC1323 scaling applied */
289 new_win >>= tp->rx_opt.rcv_wscale;
290
291 /* If we advertise zero window, disable fast path. */
292 if (new_win == 0)
293 tp->pred_flags = 0;
294
295 return new_win;
296}
297
298/* Packet ECN state for a SYN-ACK */
299static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb)
300{
301 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
302 if (!(tp->ecn_flags & TCP_ECN_OK))
303 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
304}
305
306/* Packet ECN state for a SYN. */
307static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
308{
309 struct tcp_sock *tp = tcp_sk(sk);
310
311 tp->ecn_flags = 0;
312 if (sysctl_tcp_ecn == 1) {
313 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
314 tp->ecn_flags = TCP_ECN_OK;
315 }
316}
317
318static __inline__ void
319TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th)
320{
321 if (inet_rsk(req)->ecn_ok)
322 th->ece = 1;
323}
324
325/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
326 * be sent.
327 */
328static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
329 int tcp_header_len)
330{
331 struct tcp_sock *tp = tcp_sk(sk);
332
333 if (tp->ecn_flags & TCP_ECN_OK) {
334 /* Not-retransmitted data segment: set ECT and inject CWR. */
335 if (skb->len != tcp_header_len &&
336 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
337 INET_ECN_xmit(sk);
338 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
339 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
340 tcp_hdr(skb)->cwr = 1;
341 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
342 }
343 } else {
344 /* ACK or retransmitted segment: clear ECT|CE */
345 INET_ECN_dontxmit(sk);
346 }
347 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
348 tcp_hdr(skb)->ece = 1;
349 }
350}
351
352/* Constructs common control bits of non-data skb. If SYN/FIN is present,
353 * auto increment end seqno.
354 */
355static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
356{
357 skb->ip_summed = CHECKSUM_PARTIAL;
358 skb->csum = 0;
359
360 TCP_SKB_CB(skb)->tcp_flags = flags;
361 TCP_SKB_CB(skb)->sacked = 0;
362
363 skb_shinfo(skb)->gso_segs = 1;
364 skb_shinfo(skb)->gso_size = 0;
365 skb_shinfo(skb)->gso_type = 0;
366
367 TCP_SKB_CB(skb)->seq = seq;
368 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
369 seq++;
370 TCP_SKB_CB(skb)->end_seq = seq;
371}
372
373static inline bool tcp_urg_mode(const struct tcp_sock *tp)
374{
375 return tp->snd_una != tp->snd_up;
376}
377
378#define OPTION_SACK_ADVERTISE (1 << 0)
379#define OPTION_TS (1 << 1)
380#define OPTION_MD5 (1 << 2)
381#define OPTION_WSCALE (1 << 3)
382#define OPTION_COOKIE_EXTENSION (1 << 4)
383
384struct tcp_out_options {
385 u8 options; /* bit field of OPTION_* */
386 u8 ws; /* window scale, 0 to disable */
387 u8 num_sack_blocks; /* number of SACK blocks to include */
388 u8 hash_size; /* bytes in hash_location */
389 u16 mss; /* 0 to disable */
390 __u32 tsval, tsecr; /* need to include OPTION_TS */
391 __u8 *hash_location; /* temporary pointer, overloaded */
392};
393
394/* The sysctl int routines are generic, so check consistency here.
395 */
396static u8 tcp_cookie_size_check(u8 desired)
397{
398 int cookie_size;
399
400 if (desired > 0)
401 /* previously specified */
402 return desired;
403
404 cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size);
405 if (cookie_size <= 0)
406 /* no default specified */
407 return 0;
408
409 if (cookie_size <= TCP_COOKIE_MIN)
410 /* value too small, specify minimum */
411 return TCP_COOKIE_MIN;
412
413 if (cookie_size >= TCP_COOKIE_MAX)
414 /* value too large, specify maximum */
415 return TCP_COOKIE_MAX;
416
417 if (cookie_size & 1)
418 /* 8-bit multiple, illegal, fix it */
419 cookie_size++;
420
421 return (u8)cookie_size;
422}
423
424/* Write previously computed TCP options to the packet.
425 *
426 * Beware: Something in the Internet is very sensitive to the ordering of
427 * TCP options, we learned this through the hard way, so be careful here.
428 * Luckily we can at least blame others for their non-compliance but from
429 * inter-operatibility perspective it seems that we're somewhat stuck with
430 * the ordering which we have been using if we want to keep working with
431 * those broken things (not that it currently hurts anybody as there isn't
432 * particular reason why the ordering would need to be changed).
433 *
434 * At least SACK_PERM as the first option is known to lead to a disaster
435 * (but it may well be that other scenarios fail similarly).
436 */
437static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
438 struct tcp_out_options *opts)
439{
440 u8 options = opts->options; /* mungable copy */
441
442 /* Having both authentication and cookies for security is redundant,
443 * and there's certainly not enough room. Instead, the cookie-less
444 * extension variant is proposed.
445 *
446 * Consider the pessimal case with authentication. The options
447 * could look like:
448 * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
449 */
450 if (unlikely(OPTION_MD5 & options)) {
451 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
452 *ptr++ = htonl((TCPOPT_COOKIE << 24) |
453 (TCPOLEN_COOKIE_BASE << 16) |
454 (TCPOPT_MD5SIG << 8) |
455 TCPOLEN_MD5SIG);
456 } else {
457 *ptr++ = htonl((TCPOPT_NOP << 24) |
458 (TCPOPT_NOP << 16) |
459 (TCPOPT_MD5SIG << 8) |
460 TCPOLEN_MD5SIG);
461 }
462 options &= ~OPTION_COOKIE_EXTENSION;
463 /* overload cookie hash location */
464 opts->hash_location = (__u8 *)ptr;
465 ptr += 4;
466 }
467
468 if (unlikely(opts->mss)) {
469 *ptr++ = htonl((TCPOPT_MSS << 24) |
470 (TCPOLEN_MSS << 16) |
471 opts->mss);
472 }
473
474 if (likely(OPTION_TS & options)) {
475 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
476 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
477 (TCPOLEN_SACK_PERM << 16) |
478 (TCPOPT_TIMESTAMP << 8) |
479 TCPOLEN_TIMESTAMP);
480 options &= ~OPTION_SACK_ADVERTISE;
481 } else {
482 *ptr++ = htonl((TCPOPT_NOP << 24) |
483 (TCPOPT_NOP << 16) |
484 (TCPOPT_TIMESTAMP << 8) |
485 TCPOLEN_TIMESTAMP);
486 }
487 *ptr++ = htonl(opts->tsval);
488 *ptr++ = htonl(opts->tsecr);
489 }
490
491 /* Specification requires after timestamp, so do it now.
492 *
493 * Consider the pessimal case without authentication. The options
494 * could look like:
495 * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
496 */
497 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
498 __u8 *cookie_copy = opts->hash_location;
499 u8 cookie_size = opts->hash_size;
500
501 /* 8-bit multiple handled in tcp_cookie_size_check() above,
502 * and elsewhere.
503 */
504 if (0x2 & cookie_size) {
505 __u8 *p = (__u8 *)ptr;
506
507 /* 16-bit multiple */
508 *p++ = TCPOPT_COOKIE;
509 *p++ = TCPOLEN_COOKIE_BASE + cookie_size;
510 *p++ = *cookie_copy++;
511 *p++ = *cookie_copy++;
512 ptr++;
513 cookie_size -= 2;
514 } else {
515 /* 32-bit multiple */
516 *ptr++ = htonl(((TCPOPT_NOP << 24) |
517 (TCPOPT_NOP << 16) |
518 (TCPOPT_COOKIE << 8) |
519 TCPOLEN_COOKIE_BASE) +
520 cookie_size);
521 }
522
523 if (cookie_size > 0) {
524 memcpy(ptr, cookie_copy, cookie_size);
525 ptr += (cookie_size / 4);
526 }
527 }
528
529 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
530 *ptr++ = htonl((TCPOPT_NOP << 24) |
531 (TCPOPT_NOP << 16) |
532 (TCPOPT_SACK_PERM << 8) |
533 TCPOLEN_SACK_PERM);
534 }
535
536 if (unlikely(OPTION_WSCALE & options)) {
537 *ptr++ = htonl((TCPOPT_NOP << 24) |
538 (TCPOPT_WINDOW << 16) |
539 (TCPOLEN_WINDOW << 8) |
540 opts->ws);
541 }
542
543 if (unlikely(opts->num_sack_blocks)) {
544 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
545 tp->duplicate_sack : tp->selective_acks;
546 int this_sack;
547
548 *ptr++ = htonl((TCPOPT_NOP << 24) |
549 (TCPOPT_NOP << 16) |
550 (TCPOPT_SACK << 8) |
551 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
552 TCPOLEN_SACK_PERBLOCK)));
553
554 for (this_sack = 0; this_sack < opts->num_sack_blocks;
555 ++this_sack) {
556 *ptr++ = htonl(sp[this_sack].start_seq);
557 *ptr++ = htonl(sp[this_sack].end_seq);
558 }
559
560 tp->rx_opt.dsack = 0;
561 }
562}
563
564/* Compute TCP options for SYN packets. This is not the final
565 * network wire format yet.
566 */
567static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
568 struct tcp_out_options *opts,
569 struct tcp_md5sig_key **md5)
570{
571 struct tcp_sock *tp = tcp_sk(sk);
572 struct tcp_cookie_values *cvp = tp->cookie_values;
573 unsigned int remaining = MAX_TCP_OPTION_SPACE;
574 u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ?
575 tcp_cookie_size_check(cvp->cookie_desired) :
576 0;
577
578#ifdef CONFIG_TCP_MD5SIG
579 *md5 = tp->af_specific->md5_lookup(sk, sk);
580 if (*md5) {
581 opts->options |= OPTION_MD5;
582 remaining -= TCPOLEN_MD5SIG_ALIGNED;
583 }
584#else
585 *md5 = NULL;
586#endif
587
588 /* We always get an MSS option. The option bytes which will be seen in
589 * normal data packets should timestamps be used, must be in the MSS
590 * advertised. But we subtract them from tp->mss_cache so that
591 * calculations in tcp_sendmsg are simpler etc. So account for this
592 * fact here if necessary. If we don't do this correctly, as a
593 * receiver we won't recognize data packets as being full sized when we
594 * should, and thus we won't abide by the delayed ACK rules correctly.
595 * SACKs don't matter, we never delay an ACK when we have any of those
596 * going out. */
597 opts->mss = tcp_advertise_mss(sk);
598 remaining -= TCPOLEN_MSS_ALIGNED;
599
600 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
601 opts->options |= OPTION_TS;
602 opts->tsval = TCP_SKB_CB(skb)->when;
603 opts->tsecr = tp->rx_opt.ts_recent;
604 remaining -= TCPOLEN_TSTAMP_ALIGNED;
605 }
606 if (likely(sysctl_tcp_window_scaling)) {
607 opts->ws = tp->rx_opt.rcv_wscale;
608 opts->options |= OPTION_WSCALE;
609 remaining -= TCPOLEN_WSCALE_ALIGNED;
610 }
611 if (likely(sysctl_tcp_sack)) {
612 opts->options |= OPTION_SACK_ADVERTISE;
613 if (unlikely(!(OPTION_TS & opts->options)))
614 remaining -= TCPOLEN_SACKPERM_ALIGNED;
615 }
616
617 /* Note that timestamps are required by the specification.
618 *
619 * Odd numbers of bytes are prohibited by the specification, ensuring
620 * that the cookie is 16-bit aligned, and the resulting cookie pair is
621 * 32-bit aligned.
622 */
623 if (*md5 == NULL &&
624 (OPTION_TS & opts->options) &&
625 cookie_size > 0) {
626 int need = TCPOLEN_COOKIE_BASE + cookie_size;
627
628 if (0x2 & need) {
629 /* 32-bit multiple */
630 need += 2; /* NOPs */
631
632 if (need > remaining) {
633 /* try shrinking cookie to fit */
634 cookie_size -= 2;
635 need -= 4;
636 }
637 }
638 while (need > remaining && TCP_COOKIE_MIN <= cookie_size) {
639 cookie_size -= 4;
640 need -= 4;
641 }
642 if (TCP_COOKIE_MIN <= cookie_size) {
643 opts->options |= OPTION_COOKIE_EXTENSION;
644 opts->hash_location = (__u8 *)&cvp->cookie_pair[0];
645 opts->hash_size = cookie_size;
646
647 /* Remember for future incarnations. */
648 cvp->cookie_desired = cookie_size;
649
650 if (cvp->cookie_desired != cvp->cookie_pair_size) {
651 /* Currently use random bytes as a nonce,
652 * assuming these are completely unpredictable
653 * by hostile users of the same system.
654 */
655 get_random_bytes(&cvp->cookie_pair[0],
656 cookie_size);
657 cvp->cookie_pair_size = cookie_size;
658 }
659
660 remaining -= need;
661 }
662 }
663 return MAX_TCP_OPTION_SPACE - remaining;
664}
665
666/* Set up TCP options for SYN-ACKs. */
667static unsigned int tcp_synack_options(struct sock *sk,
668 struct request_sock *req,
669 unsigned int mss, struct sk_buff *skb,
670 struct tcp_out_options *opts,
671 struct tcp_md5sig_key **md5,
672 struct tcp_extend_values *xvp)
673{
674 struct inet_request_sock *ireq = inet_rsk(req);
675 unsigned int remaining = MAX_TCP_OPTION_SPACE;
676 u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ?
677 xvp->cookie_plus :
678 0;
679
680#ifdef CONFIG_TCP_MD5SIG
681 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
682 if (*md5) {
683 opts->options |= OPTION_MD5;
684 remaining -= TCPOLEN_MD5SIG_ALIGNED;
685
686 /* We can't fit any SACK blocks in a packet with MD5 + TS
687 * options. There was discussion about disabling SACK
688 * rather than TS in order to fit in better with old,
689 * buggy kernels, but that was deemed to be unnecessary.
690 */
691 ireq->tstamp_ok &= !ireq->sack_ok;
692 }
693#else
694 *md5 = NULL;
695#endif
696
697 /* We always send an MSS option. */
698 opts->mss = mss;
699 remaining -= TCPOLEN_MSS_ALIGNED;
700
701 if (likely(ireq->wscale_ok)) {
702 opts->ws = ireq->rcv_wscale;
703 opts->options |= OPTION_WSCALE;
704 remaining -= TCPOLEN_WSCALE_ALIGNED;
705 }
706 if (likely(ireq->tstamp_ok)) {
707 opts->options |= OPTION_TS;
708 opts->tsval = TCP_SKB_CB(skb)->when;
709 opts->tsecr = req->ts_recent;
710 remaining -= TCPOLEN_TSTAMP_ALIGNED;
711 }
712 if (likely(ireq->sack_ok)) {
713 opts->options |= OPTION_SACK_ADVERTISE;
714 if (unlikely(!ireq->tstamp_ok))
715 remaining -= TCPOLEN_SACKPERM_ALIGNED;
716 }
717
718 /* Similar rationale to tcp_syn_options() applies here, too.
719 * If the <SYN> options fit, the same options should fit now!
720 */
721 if (*md5 == NULL &&
722 ireq->tstamp_ok &&
723 cookie_plus > TCPOLEN_COOKIE_BASE) {
724 int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */
725
726 if (0x2 & need) {
727 /* 32-bit multiple */
728 need += 2; /* NOPs */
729 }
730 if (need <= remaining) {
731 opts->options |= OPTION_COOKIE_EXTENSION;
732 opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE;
733 remaining -= need;
734 } else {
735 /* There's no error return, so flag it. */
736 xvp->cookie_out_never = 1; /* true */
737 opts->hash_size = 0;
738 }
739 }
740 return MAX_TCP_OPTION_SPACE - remaining;
741}
742
743/* Compute TCP options for ESTABLISHED sockets. This is not the
744 * final wire format yet.
745 */
746static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
747 struct tcp_out_options *opts,
748 struct tcp_md5sig_key **md5)
749{
750 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
751 struct tcp_sock *tp = tcp_sk(sk);
752 unsigned int size = 0;
753 unsigned int eff_sacks;
754
755#ifdef CONFIG_TCP_MD5SIG
756 *md5 = tp->af_specific->md5_lookup(sk, sk);
757 if (unlikely(*md5)) {
758 opts->options |= OPTION_MD5;
759 size += TCPOLEN_MD5SIG_ALIGNED;
760 }
761#else
762 *md5 = NULL;
763#endif
764
765 if (likely(tp->rx_opt.tstamp_ok)) {
766 opts->options |= OPTION_TS;
767 opts->tsval = tcb ? tcb->when : 0;
768 opts->tsecr = tp->rx_opt.ts_recent;
769 size += TCPOLEN_TSTAMP_ALIGNED;
770 }
771
772 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
773 if (unlikely(eff_sacks)) {
774 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
775 opts->num_sack_blocks =
776 min_t(unsigned int, eff_sacks,
777 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
778 TCPOLEN_SACK_PERBLOCK);
779 size += TCPOLEN_SACK_BASE_ALIGNED +
780 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
781 }
782
783 return size;
784}
785
786/* This routine actually transmits TCP packets queued in by
787 * tcp_do_sendmsg(). This is used by both the initial
788 * transmission and possible later retransmissions.
789 * All SKB's seen here are completely headerless. It is our
790 * job to build the TCP header, and pass the packet down to
791 * IP so it can do the same plus pass the packet off to the
792 * device.
793 *
794 * We are working here with either a clone of the original
795 * SKB, or a fresh unique copy made by the retransmit engine.
796 */
797static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
798 gfp_t gfp_mask)
799{
800 const struct inet_connection_sock *icsk = inet_csk(sk);
801 struct inet_sock *inet;
802 struct tcp_sock *tp;
803 struct tcp_skb_cb *tcb;
804 struct tcp_out_options opts;
805 unsigned int tcp_options_size, tcp_header_size;
806 struct tcp_md5sig_key *md5;
807 struct tcphdr *th;
808 int err;
809
810 BUG_ON(!skb || !tcp_skb_pcount(skb));
811
812 /* If congestion control is doing timestamping, we must
813 * take such a timestamp before we potentially clone/copy.
814 */
815 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
816 __net_timestamp(skb);
817
818 if (likely(clone_it)) {
819 if (unlikely(skb_cloned(skb)))
820 skb = pskb_copy(skb, gfp_mask);
821 else
822 skb = skb_clone(skb, gfp_mask);
823 if (unlikely(!skb))
824 return -ENOBUFS;
825 }
826
827 inet = inet_sk(sk);
828 tp = tcp_sk(sk);
829 tcb = TCP_SKB_CB(skb);
830 memset(&opts, 0, sizeof(opts));
831
832 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
833 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
834 else
835 tcp_options_size = tcp_established_options(sk, skb, &opts,
836 &md5);
837 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
838
839 if (tcp_packets_in_flight(tp) == 0) {
840 tcp_ca_event(sk, CA_EVENT_TX_START);
841 skb->ooo_okay = 1;
842 } else
843 skb->ooo_okay = 0;
844
845 skb_push(skb, tcp_header_size);
846 skb_reset_transport_header(skb);
847 skb_set_owner_w(skb, sk);
848
849 /* Build TCP header and checksum it. */
850 th = tcp_hdr(skb);
851 th->source = inet->inet_sport;
852 th->dest = inet->inet_dport;
853 th->seq = htonl(tcb->seq);
854 th->ack_seq = htonl(tp->rcv_nxt);
855 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
856 tcb->tcp_flags);
857
858 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
859 /* RFC1323: The window in SYN & SYN/ACK segments
860 * is never scaled.
861 */
862 th->window = htons(min(tp->rcv_wnd, 65535U));
863 } else {
864 th->window = htons(tcp_select_window(sk));
865 }
866 th->check = 0;
867 th->urg_ptr = 0;
868
869 /* The urg_mode check is necessary during a below snd_una win probe */
870 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
871 if (before(tp->snd_up, tcb->seq + 0x10000)) {
872 th->urg_ptr = htons(tp->snd_up - tcb->seq);
873 th->urg = 1;
874 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
875 th->urg_ptr = htons(0xFFFF);
876 th->urg = 1;
877 }
878 }
879
880 tcp_options_write((__be32 *)(th + 1), tp, &opts);
881 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
882 TCP_ECN_send(sk, skb, tcp_header_size);
883
884#ifdef CONFIG_TCP_MD5SIG
885 /* Calculate the MD5 hash, as we have all we need now */
886 if (md5) {
887 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
888 tp->af_specific->calc_md5_hash(opts.hash_location,
889 md5, sk, NULL, skb);
890 }
891#endif
892
893 icsk->icsk_af_ops->send_check(sk, skb);
894
895 if (likely(tcb->tcp_flags & TCPHDR_ACK))
896 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
897
898 if (skb->len != tcp_header_size)
899 tcp_event_data_sent(tp, sk);
900
901 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
902 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
903 tcp_skb_pcount(skb));
904
905 err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
906 if (likely(err <= 0))
907 return err;
908
909 tcp_enter_cwr(sk, 1);
910
911 return net_xmit_eval(err);
912}
913
914/* This routine just queues the buffer for sending.
915 *
916 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
917 * otherwise socket can stall.
918 */
919static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
920{
921 struct tcp_sock *tp = tcp_sk(sk);
922
923 /* Advance write_seq and place onto the write_queue. */
924 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
925 skb_header_release(skb);
926 tcp_add_write_queue_tail(sk, skb);
927 sk->sk_wmem_queued += skb->truesize;
928 sk_mem_charge(sk, skb->truesize);
929}
930
931/* Initialize TSO segments for a packet. */
932static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
933 unsigned int mss_now)
934{
935 if (skb->len <= mss_now || !sk_can_gso(sk) ||
936 skb->ip_summed == CHECKSUM_NONE) {
937 /* Avoid the costly divide in the normal
938 * non-TSO case.
939 */
940 skb_shinfo(skb)->gso_segs = 1;
941 skb_shinfo(skb)->gso_size = 0;
942 skb_shinfo(skb)->gso_type = 0;
943 } else {
944 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
945 skb_shinfo(skb)->gso_size = mss_now;
946 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
947 }
948}
949
950/* When a modification to fackets out becomes necessary, we need to check
951 * skb is counted to fackets_out or not.
952 */
953static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
954 int decr)
955{
956 struct tcp_sock *tp = tcp_sk(sk);
957
958 if (!tp->sacked_out || tcp_is_reno(tp))
959 return;
960
961 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
962 tp->fackets_out -= decr;
963}
964
965/* Pcount in the middle of the write queue got changed, we need to do various
966 * tweaks to fix counters
967 */
968static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
969{
970 struct tcp_sock *tp = tcp_sk(sk);
971
972 tp->packets_out -= decr;
973
974 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
975 tp->sacked_out -= decr;
976 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
977 tp->retrans_out -= decr;
978 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
979 tp->lost_out -= decr;
980
981 /* Reno case is special. Sigh... */
982 if (tcp_is_reno(tp) && decr > 0)
983 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
984
985 tcp_adjust_fackets_out(sk, skb, decr);
986
987 if (tp->lost_skb_hint &&
988 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
989 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
990 tp->lost_cnt_hint -= decr;
991
992 tcp_verify_left_out(tp);
993}
994
995/* Function to create two new TCP segments. Shrinks the given segment
996 * to the specified size and appends a new segment with the rest of the
997 * packet to the list. This won't be called frequently, I hope.
998 * Remember, these are still headerless SKBs at this point.
999 */
1000int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1001 unsigned int mss_now)
1002{
1003 struct tcp_sock *tp = tcp_sk(sk);
1004 struct sk_buff *buff;
1005 int nsize, old_factor;
1006 int nlen;
1007 u8 flags;
1008
1009 if (WARN_ON(len > skb->len))
1010 return -EINVAL;
1011
1012 nsize = skb_headlen(skb) - len;
1013 if (nsize < 0)
1014 nsize = 0;
1015
1016 if (skb_cloned(skb) &&
1017 skb_is_nonlinear(skb) &&
1018 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1019 return -ENOMEM;
1020
1021 /* Get a new skb... force flag on. */
1022 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1023 if (buff == NULL)
1024 return -ENOMEM; /* We'll just try again later. */
1025
1026 sk->sk_wmem_queued += buff->truesize;
1027 sk_mem_charge(sk, buff->truesize);
1028 nlen = skb->len - len - nsize;
1029 buff->truesize += nlen;
1030 skb->truesize -= nlen;
1031
1032 /* Correct the sequence numbers. */
1033 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1034 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1035 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1036
1037 /* PSH and FIN should only be set in the second packet. */
1038 flags = TCP_SKB_CB(skb)->tcp_flags;
1039 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1040 TCP_SKB_CB(buff)->tcp_flags = flags;
1041 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1042
1043 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1044 /* Copy and checksum data tail into the new buffer. */
1045 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1046 skb_put(buff, nsize),
1047 nsize, 0);
1048
1049 skb_trim(skb, len);
1050
1051 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1052 } else {
1053 skb->ip_summed = CHECKSUM_PARTIAL;
1054 skb_split(skb, buff, len);
1055 }
1056
1057 buff->ip_summed = skb->ip_summed;
1058
1059 /* Looks stupid, but our code really uses when of
1060 * skbs, which it never sent before. --ANK
1061 */
1062 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1063 buff->tstamp = skb->tstamp;
1064
1065 old_factor = tcp_skb_pcount(skb);
1066
1067 /* Fix up tso_factor for both original and new SKB. */
1068 tcp_set_skb_tso_segs(sk, skb, mss_now);
1069 tcp_set_skb_tso_segs(sk, buff, mss_now);
1070
1071 /* If this packet has been sent out already, we must
1072 * adjust the various packet counters.
1073 */
1074 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1075 int diff = old_factor - tcp_skb_pcount(skb) -
1076 tcp_skb_pcount(buff);
1077
1078 if (diff)
1079 tcp_adjust_pcount(sk, skb, diff);
1080 }
1081
1082 /* Link BUFF into the send queue. */
1083 skb_header_release(buff);
1084 tcp_insert_write_queue_after(skb, buff, sk);
1085
1086 return 0;
1087}
1088
1089/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1090 * eventually). The difference is that pulled data not copied, but
1091 * immediately discarded.
1092 */
1093static void __pskb_trim_head(struct sk_buff *skb, int len)
1094{
1095 int i, k, eat;
1096
1097 eat = min_t(int, len, skb_headlen(skb));
1098 if (eat) {
1099 __skb_pull(skb, eat);
1100 skb->avail_size -= eat;
1101 len -= eat;
1102 if (!len)
1103 return;
1104 }
1105 eat = len;
1106 k = 0;
1107 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1108 int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
1109
1110 if (size <= eat) {
1111 skb_frag_unref(skb, i);
1112 eat -= size;
1113 } else {
1114 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1115 if (eat) {
1116 skb_shinfo(skb)->frags[k].page_offset += eat;
1117 skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
1118 eat = 0;
1119 }
1120 k++;
1121 }
1122 }
1123 skb_shinfo(skb)->nr_frags = k;
1124
1125 skb_reset_tail_pointer(skb);
1126 skb->data_len -= len;
1127 skb->len = skb->data_len;
1128}
1129
1130/* Remove acked data from a packet in the transmit queue. */
1131int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1132{
1133 if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1134 return -ENOMEM;
1135
1136 __pskb_trim_head(skb, len);
1137
1138 TCP_SKB_CB(skb)->seq += len;
1139 skb->ip_summed = CHECKSUM_PARTIAL;
1140
1141 skb->truesize -= len;
1142 sk->sk_wmem_queued -= len;
1143 sk_mem_uncharge(sk, len);
1144 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1145
1146 /* Any change of skb->len requires recalculation of tso factor. */
1147 if (tcp_skb_pcount(skb) > 1)
1148 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1149
1150 return 0;
1151}
1152
1153/* Calculate MSS. Not accounting for SACKs here. */
1154int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1155{
1156 const struct tcp_sock *tp = tcp_sk(sk);
1157 const struct inet_connection_sock *icsk = inet_csk(sk);
1158 int mss_now;
1159
1160 /* Calculate base mss without TCP options:
1161 It is MMS_S - sizeof(tcphdr) of rfc1122
1162 */
1163 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1164
1165 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1166 if (icsk->icsk_af_ops->net_frag_header_len) {
1167 const struct dst_entry *dst = __sk_dst_get(sk);
1168
1169 if (dst && dst_allfrag(dst))
1170 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1171 }
1172
1173 /* Clamp it (mss_clamp does not include tcp options) */
1174 if (mss_now > tp->rx_opt.mss_clamp)
1175 mss_now = tp->rx_opt.mss_clamp;
1176
1177 /* Now subtract optional transport overhead */
1178 mss_now -= icsk->icsk_ext_hdr_len;
1179
1180 /* Then reserve room for full set of TCP options and 8 bytes of data */
1181 if (mss_now < 48)
1182 mss_now = 48;
1183
1184 /* Now subtract TCP options size, not including SACKs */
1185 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
1186
1187 return mss_now;
1188}
1189
1190/* Inverse of above */
1191int tcp_mss_to_mtu(struct sock *sk, int mss)
1192{
1193 const struct tcp_sock *tp = tcp_sk(sk);
1194 const struct inet_connection_sock *icsk = inet_csk(sk);
1195 int mtu;
1196
1197 mtu = mss +
1198 tp->tcp_header_len +
1199 icsk->icsk_ext_hdr_len +
1200 icsk->icsk_af_ops->net_header_len;
1201
1202 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1203 if (icsk->icsk_af_ops->net_frag_header_len) {
1204 const struct dst_entry *dst = __sk_dst_get(sk);
1205
1206 if (dst && dst_allfrag(dst))
1207 mtu += icsk->icsk_af_ops->net_frag_header_len;
1208 }
1209 return mtu;
1210}
1211
1212/* MTU probing init per socket */
1213void tcp_mtup_init(struct sock *sk)
1214{
1215 struct tcp_sock *tp = tcp_sk(sk);
1216 struct inet_connection_sock *icsk = inet_csk(sk);
1217
1218 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1219 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1220 icsk->icsk_af_ops->net_header_len;
1221 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1222 icsk->icsk_mtup.probe_size = 0;
1223}
1224EXPORT_SYMBOL(tcp_mtup_init);
1225
1226/* This function synchronize snd mss to current pmtu/exthdr set.
1227
1228 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1229 for TCP options, but includes only bare TCP header.
1230
1231 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1232 It is minimum of user_mss and mss received with SYN.
1233 It also does not include TCP options.
1234
1235 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1236
1237 tp->mss_cache is current effective sending mss, including
1238 all tcp options except for SACKs. It is evaluated,
1239 taking into account current pmtu, but never exceeds
1240 tp->rx_opt.mss_clamp.
1241
1242 NOTE1. rfc1122 clearly states that advertised MSS
1243 DOES NOT include either tcp or ip options.
1244
1245 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1246 are READ ONLY outside this function. --ANK (980731)
1247 */
1248unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1249{
1250 struct tcp_sock *tp = tcp_sk(sk);
1251 struct inet_connection_sock *icsk = inet_csk(sk);
1252 int mss_now;
1253
1254 if (icsk->icsk_mtup.search_high > pmtu)
1255 icsk->icsk_mtup.search_high = pmtu;
1256
1257 mss_now = tcp_mtu_to_mss(sk, pmtu);
1258 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1259
1260 /* And store cached results */
1261 icsk->icsk_pmtu_cookie = pmtu;
1262 if (icsk->icsk_mtup.enabled)
1263 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1264 tp->mss_cache = mss_now;
1265
1266 return mss_now;
1267}
1268EXPORT_SYMBOL(tcp_sync_mss);
1269
1270/* Compute the current effective MSS, taking SACKs and IP options,
1271 * and even PMTU discovery events into account.
1272 */
1273unsigned int tcp_current_mss(struct sock *sk)
1274{
1275 const struct tcp_sock *tp = tcp_sk(sk);
1276 const struct dst_entry *dst = __sk_dst_get(sk);
1277 u32 mss_now;
1278 unsigned int header_len;
1279 struct tcp_out_options opts;
1280 struct tcp_md5sig_key *md5;
1281
1282 mss_now = tp->mss_cache;
1283
1284 if (dst) {
1285 u32 mtu = dst_mtu(dst);
1286 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1287 mss_now = tcp_sync_mss(sk, mtu);
1288 }
1289
1290 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1291 sizeof(struct tcphdr);
1292 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1293 * some common options. If this is an odd packet (because we have SACK
1294 * blocks etc) then our calculated header_len will be different, and
1295 * we have to adjust mss_now correspondingly */
1296 if (header_len != tp->tcp_header_len) {
1297 int delta = (int) header_len - tp->tcp_header_len;
1298 mss_now -= delta;
1299 }
1300
1301 return mss_now;
1302}
1303
1304/* Congestion window validation. (RFC2861) */
1305static void tcp_cwnd_validate(struct sock *sk)
1306{
1307 struct tcp_sock *tp = tcp_sk(sk);
1308
1309 if (tp->packets_out >= tp->snd_cwnd) {
1310 /* Network is feed fully. */
1311 tp->snd_cwnd_used = 0;
1312 tp->snd_cwnd_stamp = tcp_time_stamp;
1313 } else {
1314 /* Network starves. */
1315 if (tp->packets_out > tp->snd_cwnd_used)
1316 tp->snd_cwnd_used = tp->packets_out;
1317
1318 if (sysctl_tcp_slow_start_after_idle &&
1319 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1320 tcp_cwnd_application_limited(sk);
1321 }
1322}
1323
1324/* Returns the portion of skb which can be sent right away without
1325 * introducing MSS oddities to segment boundaries. In rare cases where
1326 * mss_now != mss_cache, we will request caller to create a small skb
1327 * per input skb which could be mostly avoided here (if desired).
1328 *
1329 * We explicitly want to create a request for splitting write queue tail
1330 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1331 * thus all the complexity (cwnd_len is always MSS multiple which we
1332 * return whenever allowed by the other factors). Basically we need the
1333 * modulo only when the receiver window alone is the limiting factor or
1334 * when we would be allowed to send the split-due-to-Nagle skb fully.
1335 */
1336static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb,
1337 unsigned int mss_now, unsigned int max_segs)
1338{
1339 const struct tcp_sock *tp = tcp_sk(sk);
1340 u32 needed, window, max_len;
1341
1342 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1343 max_len = mss_now * max_segs;
1344
1345 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1346 return max_len;
1347
1348 needed = min(skb->len, window);
1349
1350 if (max_len <= needed)
1351 return max_len;
1352
1353 return needed - needed % mss_now;
1354}
1355
1356/* Can at least one segment of SKB be sent right now, according to the
1357 * congestion window rules? If so, return how many segments are allowed.
1358 */
1359static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1360 const struct sk_buff *skb)
1361{
1362 u32 in_flight, cwnd;
1363
1364 /* Don't be strict about the congestion window for the final FIN. */
1365 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1366 tcp_skb_pcount(skb) == 1)
1367 return 1;
1368
1369 in_flight = tcp_packets_in_flight(tp);
1370 cwnd = tp->snd_cwnd;
1371 if (in_flight < cwnd)
1372 return (cwnd - in_flight);
1373
1374 return 0;
1375}
1376
1377/* Initialize TSO state of a skb.
1378 * This must be invoked the first time we consider transmitting
1379 * SKB onto the wire.
1380 */
1381static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1382 unsigned int mss_now)
1383{
1384 int tso_segs = tcp_skb_pcount(skb);
1385
1386 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1387 tcp_set_skb_tso_segs(sk, skb, mss_now);
1388 tso_segs = tcp_skb_pcount(skb);
1389 }
1390 return tso_segs;
1391}
1392
1393/* Minshall's variant of the Nagle send check. */
1394static inline bool tcp_minshall_check(const struct tcp_sock *tp)
1395{
1396 return after(tp->snd_sml, tp->snd_una) &&
1397 !after(tp->snd_sml, tp->snd_nxt);
1398}
1399
1400/* Return false, if packet can be sent now without violation Nagle's rules:
1401 * 1. It is full sized.
1402 * 2. Or it contains FIN. (already checked by caller)
1403 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1404 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1405 * With Minshall's modification: all sent small packets are ACKed.
1406 */
1407static inline bool tcp_nagle_check(const struct tcp_sock *tp,
1408 const struct sk_buff *skb,
1409 unsigned int mss_now, int nonagle)
1410{
1411 return skb->len < mss_now &&
1412 ((nonagle & TCP_NAGLE_CORK) ||
1413 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1414}
1415
1416/* Return true if the Nagle test allows this packet to be
1417 * sent now.
1418 */
1419static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1420 unsigned int cur_mss, int nonagle)
1421{
1422 /* Nagle rule does not apply to frames, which sit in the middle of the
1423 * write_queue (they have no chances to get new data).
1424 *
1425 * This is implemented in the callers, where they modify the 'nonagle'
1426 * argument based upon the location of SKB in the send queue.
1427 */
1428 if (nonagle & TCP_NAGLE_PUSH)
1429 return true;
1430
1431 /* Don't use the nagle rule for urgent data (or for the final FIN).
1432 * Nagle can be ignored during F-RTO too (see RFC4138).
1433 */
1434 if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
1435 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1436 return true;
1437
1438 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1439 return true;
1440
1441 return false;
1442}
1443
1444/* Does at least the first segment of SKB fit into the send window? */
1445static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1446 const struct sk_buff *skb,
1447 unsigned int cur_mss)
1448{
1449 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1450
1451 if (skb->len > cur_mss)
1452 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1453
1454 return !after(end_seq, tcp_wnd_end(tp));
1455}
1456
1457/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1458 * should be put on the wire right now. If so, it returns the number of
1459 * packets allowed by the congestion window.
1460 */
1461static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1462 unsigned int cur_mss, int nonagle)
1463{
1464 const struct tcp_sock *tp = tcp_sk(sk);
1465 unsigned int cwnd_quota;
1466
1467 tcp_init_tso_segs(sk, skb, cur_mss);
1468
1469 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1470 return 0;
1471
1472 cwnd_quota = tcp_cwnd_test(tp, skb);
1473 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1474 cwnd_quota = 0;
1475
1476 return cwnd_quota;
1477}
1478
1479/* Test if sending is allowed right now. */
1480bool tcp_may_send_now(struct sock *sk)
1481{
1482 const struct tcp_sock *tp = tcp_sk(sk);
1483 struct sk_buff *skb = tcp_send_head(sk);
1484
1485 return skb &&
1486 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1487 (tcp_skb_is_last(sk, skb) ?
1488 tp->nonagle : TCP_NAGLE_PUSH));
1489}
1490
1491/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1492 * which is put after SKB on the list. It is very much like
1493 * tcp_fragment() except that it may make several kinds of assumptions
1494 * in order to speed up the splitting operation. In particular, we
1495 * know that all the data is in scatter-gather pages, and that the
1496 * packet has never been sent out before (and thus is not cloned).
1497 */
1498static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1499 unsigned int mss_now, gfp_t gfp)
1500{
1501 struct sk_buff *buff;
1502 int nlen = skb->len - len;
1503 u8 flags;
1504
1505 /* All of a TSO frame must be composed of paged data. */
1506 if (skb->len != skb->data_len)
1507 return tcp_fragment(sk, skb, len, mss_now);
1508
1509 buff = sk_stream_alloc_skb(sk, 0, gfp);
1510 if (unlikely(buff == NULL))
1511 return -ENOMEM;
1512
1513 sk->sk_wmem_queued += buff->truesize;
1514 sk_mem_charge(sk, buff->truesize);
1515 buff->truesize += nlen;
1516 skb->truesize -= nlen;
1517
1518 /* Correct the sequence numbers. */
1519 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1520 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1521 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1522
1523 /* PSH and FIN should only be set in the second packet. */
1524 flags = TCP_SKB_CB(skb)->tcp_flags;
1525 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1526 TCP_SKB_CB(buff)->tcp_flags = flags;
1527
1528 /* This packet was never sent out yet, so no SACK bits. */
1529 TCP_SKB_CB(buff)->sacked = 0;
1530
1531 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1532 skb_split(skb, buff, len);
1533
1534 /* Fix up tso_factor for both original and new SKB. */
1535 tcp_set_skb_tso_segs(sk, skb, mss_now);
1536 tcp_set_skb_tso_segs(sk, buff, mss_now);
1537
1538 /* Link BUFF into the send queue. */
1539 skb_header_release(buff);
1540 tcp_insert_write_queue_after(skb, buff, sk);
1541
1542 return 0;
1543}
1544
1545/* Try to defer sending, if possible, in order to minimize the amount
1546 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1547 *
1548 * This algorithm is from John Heffner.
1549 */
1550static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1551{
1552 struct tcp_sock *tp = tcp_sk(sk);
1553 const struct inet_connection_sock *icsk = inet_csk(sk);
1554 u32 send_win, cong_win, limit, in_flight;
1555 int win_divisor;
1556
1557 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1558 goto send_now;
1559
1560 if (icsk->icsk_ca_state != TCP_CA_Open)
1561 goto send_now;
1562
1563 /* Defer for less than two clock ticks. */
1564 if (tp->tso_deferred &&
1565 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1566 goto send_now;
1567
1568 in_flight = tcp_packets_in_flight(tp);
1569
1570 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1571
1572 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1573
1574 /* From in_flight test above, we know that cwnd > in_flight. */
1575 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1576
1577 limit = min(send_win, cong_win);
1578
1579 /* If a full-sized TSO skb can be sent, do it. */
1580 if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
1581 sk->sk_gso_max_segs * tp->mss_cache))
1582 goto send_now;
1583
1584 /* Middle in queue won't get any more data, full sendable already? */
1585 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1586 goto send_now;
1587
1588 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1589 if (win_divisor) {
1590 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1591
1592 /* If at least some fraction of a window is available,
1593 * just use it.
1594 */
1595 chunk /= win_divisor;
1596 if (limit >= chunk)
1597 goto send_now;
1598 } else {
1599 /* Different approach, try not to defer past a single
1600 * ACK. Receiver should ACK every other full sized
1601 * frame, so if we have space for more than 3 frames
1602 * then send now.
1603 */
1604 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1605 goto send_now;
1606 }
1607
1608 /* Ok, it looks like it is advisable to defer. */
1609 tp->tso_deferred = 1 | (jiffies << 1);
1610
1611 return true;
1612
1613send_now:
1614 tp->tso_deferred = 0;
1615 return false;
1616}
1617
1618/* Create a new MTU probe if we are ready.
1619 * MTU probe is regularly attempting to increase the path MTU by
1620 * deliberately sending larger packets. This discovers routing
1621 * changes resulting in larger path MTUs.
1622 *
1623 * Returns 0 if we should wait to probe (no cwnd available),
1624 * 1 if a probe was sent,
1625 * -1 otherwise
1626 */
1627static int tcp_mtu_probe(struct sock *sk)
1628{
1629 struct tcp_sock *tp = tcp_sk(sk);
1630 struct inet_connection_sock *icsk = inet_csk(sk);
1631 struct sk_buff *skb, *nskb, *next;
1632 int len;
1633 int probe_size;
1634 int size_needed;
1635 int copy;
1636 int mss_now;
1637
1638 /* Not currently probing/verifying,
1639 * not in recovery,
1640 * have enough cwnd, and
1641 * not SACKing (the variable headers throw things off) */
1642 if (!icsk->icsk_mtup.enabled ||
1643 icsk->icsk_mtup.probe_size ||
1644 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1645 tp->snd_cwnd < 11 ||
1646 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1647 return -1;
1648
1649 /* Very simple search strategy: just double the MSS. */
1650 mss_now = tcp_current_mss(sk);
1651 probe_size = 2 * tp->mss_cache;
1652 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1653 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1654 /* TODO: set timer for probe_converge_event */
1655 return -1;
1656 }
1657
1658 /* Have enough data in the send queue to probe? */
1659 if (tp->write_seq - tp->snd_nxt < size_needed)
1660 return -1;
1661
1662 if (tp->snd_wnd < size_needed)
1663 return -1;
1664 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1665 return 0;
1666
1667 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1668 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1669 if (!tcp_packets_in_flight(tp))
1670 return -1;
1671 else
1672 return 0;
1673 }
1674
1675 /* We're allowed to probe. Build it now. */
1676 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1677 return -1;
1678 sk->sk_wmem_queued += nskb->truesize;
1679 sk_mem_charge(sk, nskb->truesize);
1680
1681 skb = tcp_send_head(sk);
1682
1683 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1684 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1685 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1686 TCP_SKB_CB(nskb)->sacked = 0;
1687 nskb->csum = 0;
1688 nskb->ip_summed = skb->ip_summed;
1689
1690 tcp_insert_write_queue_before(nskb, skb, sk);
1691
1692 len = 0;
1693 tcp_for_write_queue_from_safe(skb, next, sk) {
1694 copy = min_t(int, skb->len, probe_size - len);
1695 if (nskb->ip_summed)
1696 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1697 else
1698 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1699 skb_put(nskb, copy),
1700 copy, nskb->csum);
1701
1702 if (skb->len <= copy) {
1703 /* We've eaten all the data from this skb.
1704 * Throw it away. */
1705 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1706 tcp_unlink_write_queue(skb, sk);
1707 sk_wmem_free_skb(sk, skb);
1708 } else {
1709 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1710 ~(TCPHDR_FIN|TCPHDR_PSH);
1711 if (!skb_shinfo(skb)->nr_frags) {
1712 skb_pull(skb, copy);
1713 if (skb->ip_summed != CHECKSUM_PARTIAL)
1714 skb->csum = csum_partial(skb->data,
1715 skb->len, 0);
1716 } else {
1717 __pskb_trim_head(skb, copy);
1718 tcp_set_skb_tso_segs(sk, skb, mss_now);
1719 }
1720 TCP_SKB_CB(skb)->seq += copy;
1721 }
1722
1723 len += copy;
1724
1725 if (len >= probe_size)
1726 break;
1727 }
1728 tcp_init_tso_segs(sk, nskb, nskb->len);
1729
1730 /* We're ready to send. If this fails, the probe will
1731 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1732 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1733 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1734 /* Decrement cwnd here because we are sending
1735 * effectively two packets. */
1736 tp->snd_cwnd--;
1737 tcp_event_new_data_sent(sk, nskb);
1738
1739 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1740 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1741 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1742
1743 return 1;
1744 }
1745
1746 return -1;
1747}
1748
1749/* This routine writes packets to the network. It advances the
1750 * send_head. This happens as incoming acks open up the remote
1751 * window for us.
1752 *
1753 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1754 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1755 * account rare use of URG, this is not a big flaw.
1756 *
1757 * Returns true, if no segments are in flight and we have queued segments,
1758 * but cannot send anything now because of SWS or another problem.
1759 */
1760static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1761 int push_one, gfp_t gfp)
1762{
1763 struct tcp_sock *tp = tcp_sk(sk);
1764 struct sk_buff *skb;
1765 unsigned int tso_segs, sent_pkts;
1766 int cwnd_quota;
1767 int result;
1768
1769 sent_pkts = 0;
1770
1771 if (!push_one) {
1772 /* Do MTU probing. */
1773 result = tcp_mtu_probe(sk);
1774 if (!result) {
1775 return false;
1776 } else if (result > 0) {
1777 sent_pkts = 1;
1778 }
1779 }
1780
1781 while ((skb = tcp_send_head(sk))) {
1782 unsigned int limit;
1783
1784 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1785 BUG_ON(!tso_segs);
1786
1787 cwnd_quota = tcp_cwnd_test(tp, skb);
1788 if (!cwnd_quota)
1789 break;
1790
1791 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1792 break;
1793
1794 if (tso_segs == 1) {
1795 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1796 (tcp_skb_is_last(sk, skb) ?
1797 nonagle : TCP_NAGLE_PUSH))))
1798 break;
1799 } else {
1800 if (!push_one && tcp_tso_should_defer(sk, skb))
1801 break;
1802 }
1803
1804 limit = mss_now;
1805 if (tso_segs > 1 && !tcp_urg_mode(tp))
1806 limit = tcp_mss_split_point(sk, skb, mss_now,
1807 min_t(unsigned int,
1808 cwnd_quota,
1809 sk->sk_gso_max_segs));
1810
1811 if (skb->len > limit &&
1812 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
1813 break;
1814
1815 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1816
1817 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1818 break;
1819
1820 /* Advance the send_head. This one is sent out.
1821 * This call will increment packets_out.
1822 */
1823 tcp_event_new_data_sent(sk, skb);
1824
1825 tcp_minshall_update(tp, mss_now, skb);
1826 sent_pkts += tcp_skb_pcount(skb);
1827
1828 if (push_one)
1829 break;
1830 }
1831 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery)
1832 tp->prr_out += sent_pkts;
1833
1834 if (likely(sent_pkts)) {
1835 tcp_cwnd_validate(sk);
1836 return false;
1837 }
1838 return !tp->packets_out && tcp_send_head(sk);
1839}
1840
1841/* Push out any pending frames which were held back due to
1842 * TCP_CORK or attempt at coalescing tiny packets.
1843 * The socket must be locked by the caller.
1844 */
1845void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1846 int nonagle)
1847{
1848 /* If we are closed, the bytes will have to remain here.
1849 * In time closedown will finish, we empty the write queue and
1850 * all will be happy.
1851 */
1852 if (unlikely(sk->sk_state == TCP_CLOSE))
1853 return;
1854
1855 if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC))
1856 tcp_check_probe_timer(sk);
1857}
1858
1859/* Send _single_ skb sitting at the send head. This function requires
1860 * true push pending frames to setup probe timer etc.
1861 */
1862void tcp_push_one(struct sock *sk, unsigned int mss_now)
1863{
1864 struct sk_buff *skb = tcp_send_head(sk);
1865
1866 BUG_ON(!skb || skb->len < mss_now);
1867
1868 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
1869}
1870
1871/* This function returns the amount that we can raise the
1872 * usable window based on the following constraints
1873 *
1874 * 1. The window can never be shrunk once it is offered (RFC 793)
1875 * 2. We limit memory per socket
1876 *
1877 * RFC 1122:
1878 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1879 * RECV.NEXT + RCV.WIN fixed until:
1880 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1881 *
1882 * i.e. don't raise the right edge of the window until you can raise
1883 * it at least MSS bytes.
1884 *
1885 * Unfortunately, the recommended algorithm breaks header prediction,
1886 * since header prediction assumes th->window stays fixed.
1887 *
1888 * Strictly speaking, keeping th->window fixed violates the receiver
1889 * side SWS prevention criteria. The problem is that under this rule
1890 * a stream of single byte packets will cause the right side of the
1891 * window to always advance by a single byte.
1892 *
1893 * Of course, if the sender implements sender side SWS prevention
1894 * then this will not be a problem.
1895 *
1896 * BSD seems to make the following compromise:
1897 *
1898 * If the free space is less than the 1/4 of the maximum
1899 * space available and the free space is less than 1/2 mss,
1900 * then set the window to 0.
1901 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1902 * Otherwise, just prevent the window from shrinking
1903 * and from being larger than the largest representable value.
1904 *
1905 * This prevents incremental opening of the window in the regime
1906 * where TCP is limited by the speed of the reader side taking
1907 * data out of the TCP receive queue. It does nothing about
1908 * those cases where the window is constrained on the sender side
1909 * because the pipeline is full.
1910 *
1911 * BSD also seems to "accidentally" limit itself to windows that are a
1912 * multiple of MSS, at least until the free space gets quite small.
1913 * This would appear to be a side effect of the mbuf implementation.
1914 * Combining these two algorithms results in the observed behavior
1915 * of having a fixed window size at almost all times.
1916 *
1917 * Below we obtain similar behavior by forcing the offered window to
1918 * a multiple of the mss when it is feasible to do so.
1919 *
1920 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1921 * Regular options like TIMESTAMP are taken into account.
1922 */
1923u32 __tcp_select_window(struct sock *sk)
1924{
1925 struct inet_connection_sock *icsk = inet_csk(sk);
1926 struct tcp_sock *tp = tcp_sk(sk);
1927 /* MSS for the peer's data. Previous versions used mss_clamp
1928 * here. I don't know if the value based on our guesses
1929 * of peer's MSS is better for the performance. It's more correct
1930 * but may be worse for the performance because of rcv_mss
1931 * fluctuations. --SAW 1998/11/1
1932 */
1933 int mss = icsk->icsk_ack.rcv_mss;
1934 int free_space = tcp_space(sk);
1935 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1936 int window;
1937
1938 if (mss > full_space)
1939 mss = full_space;
1940
1941 if (free_space < (full_space >> 1)) {
1942 icsk->icsk_ack.quick = 0;
1943
1944 if (sk_under_memory_pressure(sk))
1945 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1946 4U * tp->advmss);
1947
1948 if (free_space < mss)
1949 return 0;
1950 }
1951
1952 if (free_space > tp->rcv_ssthresh)
1953 free_space = tp->rcv_ssthresh;
1954
1955 /* Don't do rounding if we are using window scaling, since the
1956 * scaled window will not line up with the MSS boundary anyway.
1957 */
1958 window = tp->rcv_wnd;
1959 if (tp->rx_opt.rcv_wscale) {
1960 window = free_space;
1961
1962 /* Advertise enough space so that it won't get scaled away.
1963 * Import case: prevent zero window announcement if
1964 * 1<<rcv_wscale > mss.
1965 */
1966 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1967 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1968 << tp->rx_opt.rcv_wscale);
1969 } else {
1970 /* Get the largest window that is a nice multiple of mss.
1971 * Window clamp already applied above.
1972 * If our current window offering is within 1 mss of the
1973 * free space we just keep it. This prevents the divide
1974 * and multiply from happening most of the time.
1975 * We also don't do any window rounding when the free space
1976 * is too small.
1977 */
1978 if (window <= free_space - mss || window > free_space)
1979 window = (free_space / mss) * mss;
1980 else if (mss == full_space &&
1981 free_space > window + (full_space >> 1))
1982 window = free_space;
1983 }
1984
1985 return window;
1986}
1987
1988/* Collapses two adjacent SKB's during retransmission. */
1989static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
1990{
1991 struct tcp_sock *tp = tcp_sk(sk);
1992 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
1993 int skb_size, next_skb_size;
1994
1995 skb_size = skb->len;
1996 next_skb_size = next_skb->len;
1997
1998 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
1999
2000 tcp_highest_sack_combine(sk, next_skb, skb);
2001
2002 tcp_unlink_write_queue(next_skb, sk);
2003
2004 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2005 next_skb_size);
2006
2007 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2008 skb->ip_summed = CHECKSUM_PARTIAL;
2009
2010 if (skb->ip_summed != CHECKSUM_PARTIAL)
2011 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2012
2013 /* Update sequence range on original skb. */
2014 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2015
2016 /* Merge over control information. This moves PSH/FIN etc. over */
2017 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2018
2019 /* All done, get rid of second SKB and account for it so
2020 * packet counting does not break.
2021 */
2022 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2023
2024 /* changed transmit queue under us so clear hints */
2025 tcp_clear_retrans_hints_partial(tp);
2026 if (next_skb == tp->retransmit_skb_hint)
2027 tp->retransmit_skb_hint = skb;
2028
2029 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2030
2031 sk_wmem_free_skb(sk, next_skb);
2032}
2033
2034/* Check if coalescing SKBs is legal. */
2035static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2036{
2037 if (tcp_skb_pcount(skb) > 1)
2038 return false;
2039 /* TODO: SACK collapsing could be used to remove this condition */
2040 if (skb_shinfo(skb)->nr_frags != 0)
2041 return false;
2042 if (skb_cloned(skb))
2043 return false;
2044 if (skb == tcp_send_head(sk))
2045 return false;
2046 /* Some heurestics for collapsing over SACK'd could be invented */
2047 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2048 return false;
2049
2050 return true;
2051}
2052
2053/* Collapse packets in the retransmit queue to make to create
2054 * less packets on the wire. This is only done on retransmission.
2055 */
2056static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2057 int space)
2058{
2059 struct tcp_sock *tp = tcp_sk(sk);
2060 struct sk_buff *skb = to, *tmp;
2061 bool first = true;
2062
2063 if (!sysctl_tcp_retrans_collapse)
2064 return;
2065 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2066 return;
2067
2068 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2069 if (!tcp_can_collapse(sk, skb))
2070 break;
2071
2072 space -= skb->len;
2073
2074 if (first) {
2075 first = false;
2076 continue;
2077 }
2078
2079 if (space < 0)
2080 break;
2081 /* Punt if not enough space exists in the first SKB for
2082 * the data in the second
2083 */
2084 if (skb->len > skb_availroom(to))
2085 break;
2086
2087 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2088 break;
2089
2090 tcp_collapse_retrans(sk, to);
2091 }
2092}
2093
2094/* This retransmits one SKB. Policy decisions and retransmit queue
2095 * state updates are done by the caller. Returns non-zero if an
2096 * error occurred which prevented the send.
2097 */
2098int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2099{
2100 struct tcp_sock *tp = tcp_sk(sk);
2101 struct inet_connection_sock *icsk = inet_csk(sk);
2102 unsigned int cur_mss;
2103 int err;
2104
2105 /* Inconslusive MTU probe */
2106 if (icsk->icsk_mtup.probe_size) {
2107 icsk->icsk_mtup.probe_size = 0;
2108 }
2109
2110 /* Do not sent more than we queued. 1/4 is reserved for possible
2111 * copying overhead: fragmentation, tunneling, mangling etc.
2112 */
2113 if (atomic_read(&sk->sk_wmem_alloc) >
2114 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2115 return -EAGAIN;
2116
2117 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2118 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2119 BUG();
2120 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2121 return -ENOMEM;
2122 }
2123
2124 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2125 return -EHOSTUNREACH; /* Routing failure or similar. */
2126
2127 cur_mss = tcp_current_mss(sk);
2128
2129 /* If receiver has shrunk his window, and skb is out of
2130 * new window, do not retransmit it. The exception is the
2131 * case, when window is shrunk to zero. In this case
2132 * our retransmit serves as a zero window probe.
2133 */
2134 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2135 TCP_SKB_CB(skb)->seq != tp->snd_una)
2136 return -EAGAIN;
2137
2138 if (skb->len > cur_mss) {
2139 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2140 return -ENOMEM; /* We'll try again later. */
2141 } else {
2142 int oldpcount = tcp_skb_pcount(skb);
2143
2144 if (unlikely(oldpcount > 1)) {
2145 tcp_init_tso_segs(sk, skb, cur_mss);
2146 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2147 }
2148 }
2149
2150 tcp_retrans_try_collapse(sk, skb, cur_mss);
2151
2152 /* Some Solaris stacks overoptimize and ignore the FIN on a
2153 * retransmit when old data is attached. So strip it off
2154 * since it is cheap to do so and saves bytes on the network.
2155 */
2156 if (skb->len > 0 &&
2157 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2158 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2159 if (!pskb_trim(skb, 0)) {
2160 /* Reuse, even though it does some unnecessary work */
2161 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2162 TCP_SKB_CB(skb)->tcp_flags);
2163 skb->ip_summed = CHECKSUM_NONE;
2164 }
2165 }
2166
2167 /* Make a copy, if the first transmission SKB clone we made
2168 * is still in somebody's hands, else make a clone.
2169 */
2170 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2171
2172 /* make sure skb->data is aligned on arches that require it */
2173 if (unlikely(NET_IP_ALIGN && ((unsigned long)skb->data & 3))) {
2174 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2175 GFP_ATOMIC);
2176 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2177 -ENOBUFS;
2178 } else {
2179 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2180 }
2181
2182 if (err == 0) {
2183 /* Update global TCP statistics. */
2184 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2185
2186 tp->total_retrans++;
2187
2188#if FASTRETRANS_DEBUG > 0
2189 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2190 net_dbg_ratelimited("retrans_out leaked\n");
2191 }
2192#endif
2193 if (!tp->retrans_out)
2194 tp->lost_retrans_low = tp->snd_nxt;
2195 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2196 tp->retrans_out += tcp_skb_pcount(skb);
2197
2198 /* Save stamp of the first retransmit. */
2199 if (!tp->retrans_stamp)
2200 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2201
2202 tp->undo_retrans += tcp_skb_pcount(skb);
2203
2204 /* snd_nxt is stored to detect loss of retransmitted segment,
2205 * see tcp_input.c tcp_sacktag_write_queue().
2206 */
2207 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2208 }
2209 return err;
2210}
2211
2212/* Check if we forward retransmits are possible in the current
2213 * window/congestion state.
2214 */
2215static bool tcp_can_forward_retransmit(struct sock *sk)
2216{
2217 const struct inet_connection_sock *icsk = inet_csk(sk);
2218 const struct tcp_sock *tp = tcp_sk(sk);
2219
2220 /* Forward retransmissions are possible only during Recovery. */
2221 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2222 return false;
2223
2224 /* No forward retransmissions in Reno are possible. */
2225 if (tcp_is_reno(tp))
2226 return false;
2227
2228 /* Yeah, we have to make difficult choice between forward transmission
2229 * and retransmission... Both ways have their merits...
2230 *
2231 * For now we do not retransmit anything, while we have some new
2232 * segments to send. In the other cases, follow rule 3 for
2233 * NextSeg() specified in RFC3517.
2234 */
2235
2236 if (tcp_may_send_now(sk))
2237 return false;
2238
2239 return true;
2240}
2241
2242/* This gets called after a retransmit timeout, and the initially
2243 * retransmitted data is acknowledged. It tries to continue
2244 * resending the rest of the retransmit queue, until either
2245 * we've sent it all or the congestion window limit is reached.
2246 * If doing SACK, the first ACK which comes back for a timeout
2247 * based retransmit packet might feed us FACK information again.
2248 * If so, we use it to avoid unnecessarily retransmissions.
2249 */
2250void tcp_xmit_retransmit_queue(struct sock *sk)
2251{
2252 const struct inet_connection_sock *icsk = inet_csk(sk);
2253 struct tcp_sock *tp = tcp_sk(sk);
2254 struct sk_buff *skb;
2255 struct sk_buff *hole = NULL;
2256 u32 last_lost;
2257 int mib_idx;
2258 int fwd_rexmitting = 0;
2259
2260 if (!tp->packets_out)
2261 return;
2262
2263 if (!tp->lost_out)
2264 tp->retransmit_high = tp->snd_una;
2265
2266 if (tp->retransmit_skb_hint) {
2267 skb = tp->retransmit_skb_hint;
2268 last_lost = TCP_SKB_CB(skb)->end_seq;
2269 if (after(last_lost, tp->retransmit_high))
2270 last_lost = tp->retransmit_high;
2271 } else {
2272 skb = tcp_write_queue_head(sk);
2273 last_lost = tp->snd_una;
2274 }
2275
2276 tcp_for_write_queue_from(skb, sk) {
2277 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2278
2279 if (skb == tcp_send_head(sk))
2280 break;
2281 /* we could do better than to assign each time */
2282 if (hole == NULL)
2283 tp->retransmit_skb_hint = skb;
2284
2285 /* Assume this retransmit will generate
2286 * only one packet for congestion window
2287 * calculation purposes. This works because
2288 * tcp_retransmit_skb() will chop up the
2289 * packet to be MSS sized and all the
2290 * packet counting works out.
2291 */
2292 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2293 return;
2294
2295 if (fwd_rexmitting) {
2296begin_fwd:
2297 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2298 break;
2299 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2300
2301 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2302 tp->retransmit_high = last_lost;
2303 if (!tcp_can_forward_retransmit(sk))
2304 break;
2305 /* Backtrack if necessary to non-L'ed skb */
2306 if (hole != NULL) {
2307 skb = hole;
2308 hole = NULL;
2309 }
2310 fwd_rexmitting = 1;
2311 goto begin_fwd;
2312
2313 } else if (!(sacked & TCPCB_LOST)) {
2314 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2315 hole = skb;
2316 continue;
2317
2318 } else {
2319 last_lost = TCP_SKB_CB(skb)->end_seq;
2320 if (icsk->icsk_ca_state != TCP_CA_Loss)
2321 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2322 else
2323 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2324 }
2325
2326 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2327 continue;
2328
2329 if (tcp_retransmit_skb(sk, skb)) {
2330 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2331 return;
2332 }
2333 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2334
2335 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Recovery)
2336 tp->prr_out += tcp_skb_pcount(skb);
2337
2338 if (skb == tcp_write_queue_head(sk))
2339 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2340 inet_csk(sk)->icsk_rto,
2341 TCP_RTO_MAX);
2342 }
2343}
2344
2345/* Send a fin. The caller locks the socket for us. This cannot be
2346 * allowed to fail queueing a FIN frame under any circumstances.
2347 */
2348void tcp_send_fin(struct sock *sk)
2349{
2350 struct tcp_sock *tp = tcp_sk(sk);
2351 struct sk_buff *skb = tcp_write_queue_tail(sk);
2352 int mss_now;
2353
2354 /* Optimization, tack on the FIN if we have a queue of
2355 * unsent frames. But be careful about outgoing SACKS
2356 * and IP options.
2357 */
2358 mss_now = tcp_current_mss(sk);
2359
2360 if (tcp_send_head(sk) != NULL) {
2361 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2362 TCP_SKB_CB(skb)->end_seq++;
2363 tp->write_seq++;
2364 } else {
2365 /* Socket is locked, keep trying until memory is available. */
2366 for (;;) {
2367 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2368 sk->sk_allocation);
2369 if (skb)
2370 break;
2371 yield();
2372 }
2373
2374 /* Reserve space for headers and prepare control bits. */
2375 skb_reserve(skb, MAX_TCP_HEADER);
2376 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2377 tcp_init_nondata_skb(skb, tp->write_seq,
2378 TCPHDR_ACK | TCPHDR_FIN);
2379 tcp_queue_skb(sk, skb);
2380 }
2381 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2382}
2383
2384/* We get here when a process closes a file descriptor (either due to
2385 * an explicit close() or as a byproduct of exit()'ing) and there
2386 * was unread data in the receive queue. This behavior is recommended
2387 * by RFC 2525, section 2.17. -DaveM
2388 */
2389void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2390{
2391 struct sk_buff *skb;
2392
2393 /* NOTE: No TCP options attached and we never retransmit this. */
2394 skb = alloc_skb(MAX_TCP_HEADER, priority);
2395 if (!skb) {
2396 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2397 return;
2398 }
2399
2400 /* Reserve space for headers and prepare control bits. */
2401 skb_reserve(skb, MAX_TCP_HEADER);
2402 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2403 TCPHDR_ACK | TCPHDR_RST);
2404 /* Send it off. */
2405 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2406 if (tcp_transmit_skb(sk, skb, 0, priority))
2407 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2408
2409 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2410}
2411
2412/* Send a crossed SYN-ACK during socket establishment.
2413 * WARNING: This routine must only be called when we have already sent
2414 * a SYN packet that crossed the incoming SYN that caused this routine
2415 * to get called. If this assumption fails then the initial rcv_wnd
2416 * and rcv_wscale values will not be correct.
2417 */
2418int tcp_send_synack(struct sock *sk)
2419{
2420 struct sk_buff *skb;
2421
2422 skb = tcp_write_queue_head(sk);
2423 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2424 pr_debug("%s: wrong queue state\n", __func__);
2425 return -EFAULT;
2426 }
2427 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2428 if (skb_cloned(skb)) {
2429 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2430 if (nskb == NULL)
2431 return -ENOMEM;
2432 tcp_unlink_write_queue(skb, sk);
2433 skb_header_release(nskb);
2434 __tcp_add_write_queue_head(sk, nskb);
2435 sk_wmem_free_skb(sk, skb);
2436 sk->sk_wmem_queued += nskb->truesize;
2437 sk_mem_charge(sk, nskb->truesize);
2438 skb = nskb;
2439 }
2440
2441 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2442 TCP_ECN_send_synack(tcp_sk(sk), skb);
2443 }
2444 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2445 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2446}
2447
2448/* Prepare a SYN-ACK. */
2449struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2450 struct request_sock *req,
2451 struct request_values *rvp)
2452{
2453 struct tcp_out_options opts;
2454 struct tcp_extend_values *xvp = tcp_xv(rvp);
2455 struct inet_request_sock *ireq = inet_rsk(req);
2456 struct tcp_sock *tp = tcp_sk(sk);
2457 const struct tcp_cookie_values *cvp = tp->cookie_values;
2458 struct tcphdr *th;
2459 struct sk_buff *skb;
2460 struct tcp_md5sig_key *md5;
2461 int tcp_header_size;
2462 int mss;
2463 int s_data_desired = 0;
2464
2465 if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired)
2466 s_data_desired = cvp->s_data_desired;
2467 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15 + s_data_desired, 1, GFP_ATOMIC);
2468 if (skb == NULL)
2469 return NULL;
2470
2471 /* Reserve space for headers. */
2472 skb_reserve(skb, MAX_TCP_HEADER);
2473
2474 skb_dst_set(skb, dst_clone(dst));
2475
2476 mss = dst_metric_advmss(dst);
2477 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2478 mss = tp->rx_opt.user_mss;
2479
2480 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2481 __u8 rcv_wscale;
2482 /* Set this up on the first call only */
2483 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2484
2485 /* limit the window selection if the user enforce a smaller rx buffer */
2486 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2487 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2488 req->window_clamp = tcp_full_space(sk);
2489
2490 /* tcp_full_space because it is guaranteed to be the first packet */
2491 tcp_select_initial_window(tcp_full_space(sk),
2492 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2493 &req->rcv_wnd,
2494 &req->window_clamp,
2495 ireq->wscale_ok,
2496 &rcv_wscale,
2497 dst_metric(dst, RTAX_INITRWND));
2498 ireq->rcv_wscale = rcv_wscale;
2499 }
2500
2501 memset(&opts, 0, sizeof(opts));
2502#ifdef CONFIG_SYN_COOKIES
2503 if (unlikely(req->cookie_ts))
2504 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2505 else
2506#endif
2507 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2508 tcp_header_size = tcp_synack_options(sk, req, mss,
2509 skb, &opts, &md5, xvp)
2510 + sizeof(*th);
2511
2512 skb_push(skb, tcp_header_size);
2513 skb_reset_transport_header(skb);
2514
2515 th = tcp_hdr(skb);
2516 memset(th, 0, sizeof(struct tcphdr));
2517 th->syn = 1;
2518 th->ack = 1;
2519 TCP_ECN_make_synack(req, th);
2520 th->source = ireq->loc_port;
2521 th->dest = ireq->rmt_port;
2522 /* Setting of flags are superfluous here for callers (and ECE is
2523 * not even correctly set)
2524 */
2525 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2526 TCPHDR_SYN | TCPHDR_ACK);
2527
2528 if (OPTION_COOKIE_EXTENSION & opts.options) {
2529 if (s_data_desired) {
2530 u8 *buf = skb_put(skb, s_data_desired);
2531
2532 /* copy data directly from the listening socket. */
2533 memcpy(buf, cvp->s_data_payload, s_data_desired);
2534 TCP_SKB_CB(skb)->end_seq += s_data_desired;
2535 }
2536
2537 if (opts.hash_size > 0) {
2538 __u32 workspace[SHA_WORKSPACE_WORDS];
2539 u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS];
2540 u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1];
2541
2542 /* Secret recipe depends on the Timestamp, (future)
2543 * Sequence and Acknowledgment Numbers, Initiator
2544 * Cookie, and others handled by IP variant caller.
2545 */
2546 *tail-- ^= opts.tsval;
2547 *tail-- ^= tcp_rsk(req)->rcv_isn + 1;
2548 *tail-- ^= TCP_SKB_CB(skb)->seq + 1;
2549
2550 /* recommended */
2551 *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source);
2552 *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */
2553
2554 sha_transform((__u32 *)&xvp->cookie_bakery[0],
2555 (char *)mess,
2556 &workspace[0]);
2557 opts.hash_location =
2558 (__u8 *)&xvp->cookie_bakery[0];
2559 }
2560 }
2561
2562 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2563 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2564
2565 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2566 th->window = htons(min(req->rcv_wnd, 65535U));
2567 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2568 th->doff = (tcp_header_size >> 2);
2569 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2570
2571#ifdef CONFIG_TCP_MD5SIG
2572 /* Okay, we have all we need - do the md5 hash if needed */
2573 if (md5) {
2574 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2575 md5, NULL, req, skb);
2576 }
2577#endif
2578
2579 return skb;
2580}
2581EXPORT_SYMBOL(tcp_make_synack);
2582
2583/* Do all connect socket setups that can be done AF independent. */
2584void tcp_connect_init(struct sock *sk)
2585{
2586 const struct dst_entry *dst = __sk_dst_get(sk);
2587 struct tcp_sock *tp = tcp_sk(sk);
2588 __u8 rcv_wscale;
2589
2590 /* We'll fix this up when we get a response from the other end.
2591 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2592 */
2593 tp->tcp_header_len = sizeof(struct tcphdr) +
2594 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2595
2596#ifdef CONFIG_TCP_MD5SIG
2597 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2598 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2599#endif
2600
2601 /* If user gave his TCP_MAXSEG, record it to clamp */
2602 if (tp->rx_opt.user_mss)
2603 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2604 tp->max_window = 0;
2605 tcp_mtup_init(sk);
2606 tcp_sync_mss(sk, dst_mtu(dst));
2607
2608 if (!tp->window_clamp)
2609 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2610 tp->advmss = dst_metric_advmss(dst);
2611 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2612 tp->advmss = tp->rx_opt.user_mss;
2613
2614 tcp_initialize_rcv_mss(sk);
2615
2616 /* limit the window selection if the user enforce a smaller rx buffer */
2617 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2618 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2619 tp->window_clamp = tcp_full_space(sk);
2620
2621 tcp_select_initial_window(tcp_full_space(sk),
2622 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2623 &tp->rcv_wnd,
2624 &tp->window_clamp,
2625 sysctl_tcp_window_scaling,
2626 &rcv_wscale,
2627 dst_metric(dst, RTAX_INITRWND));
2628
2629 tp->rx_opt.rcv_wscale = rcv_wscale;
2630 tp->rcv_ssthresh = tp->rcv_wnd;
2631
2632 sk->sk_err = 0;
2633 sock_reset_flag(sk, SOCK_DONE);
2634 tp->snd_wnd = 0;
2635 tcp_init_wl(tp, 0);
2636 tp->snd_una = tp->write_seq;
2637 tp->snd_sml = tp->write_seq;
2638 tp->snd_up = tp->write_seq;
2639 tp->snd_nxt = tp->write_seq;
2640
2641 if (likely(!tp->repair))
2642 tp->rcv_nxt = 0;
2643 tp->rcv_wup = tp->rcv_nxt;
2644 tp->copied_seq = tp->rcv_nxt;
2645
2646 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2647 inet_csk(sk)->icsk_retransmits = 0;
2648 tcp_clear_retrans(tp);
2649}
2650
2651/* Build a SYN and send it off. */
2652int tcp_connect(struct sock *sk)
2653{
2654 struct tcp_sock *tp = tcp_sk(sk);
2655 struct sk_buff *buff;
2656 int err;
2657
2658 tcp_connect_init(sk);
2659
2660 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2661 if (unlikely(buff == NULL))
2662 return -ENOBUFS;
2663
2664 /* Reserve space for headers. */
2665 skb_reserve(buff, MAX_TCP_HEADER);
2666
2667 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
2668 TCP_ECN_send_syn(sk, buff);
2669
2670 /* Send it off. */
2671 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2672 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2673 skb_header_release(buff);
2674 __tcp_add_write_queue_tail(sk, buff);
2675 sk->sk_wmem_queued += buff->truesize;
2676 sk_mem_charge(sk, buff->truesize);
2677 tp->packets_out += tcp_skb_pcount(buff);
2678 err = tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2679 if (err == -ECONNREFUSED)
2680 return err;
2681
2682 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2683 * in order to make this packet get counted in tcpOutSegs.
2684 */
2685 tp->snd_nxt = tp->write_seq;
2686 tp->pushed_seq = tp->write_seq;
2687 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2688
2689 /* Timer for repeating the SYN until an answer. */
2690 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2691 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2692 return 0;
2693}
2694EXPORT_SYMBOL(tcp_connect);
2695
2696/* Send out a delayed ack, the caller does the policy checking
2697 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2698 * for details.
2699 */
2700void tcp_send_delayed_ack(struct sock *sk)
2701{
2702 struct inet_connection_sock *icsk = inet_csk(sk);
2703 int ato = icsk->icsk_ack.ato;
2704 unsigned long timeout;
2705
2706 if (ato > TCP_DELACK_MIN) {
2707 const struct tcp_sock *tp = tcp_sk(sk);
2708 int max_ato = HZ / 2;
2709
2710 if (icsk->icsk_ack.pingpong ||
2711 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2712 max_ato = TCP_DELACK_MAX;
2713
2714 /* Slow path, intersegment interval is "high". */
2715
2716 /* If some rtt estimate is known, use it to bound delayed ack.
2717 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2718 * directly.
2719 */
2720 if (tp->srtt) {
2721 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
2722
2723 if (rtt < max_ato)
2724 max_ato = rtt;
2725 }
2726
2727 ato = min(ato, max_ato);
2728 }
2729
2730 /* Stay within the limit we were given */
2731 timeout = jiffies + ato;
2732
2733 /* Use new timeout only if there wasn't a older one earlier. */
2734 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2735 /* If delack timer was blocked or is about to expire,
2736 * send ACK now.
2737 */
2738 if (icsk->icsk_ack.blocked ||
2739 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2740 tcp_send_ack(sk);
2741 return;
2742 }
2743
2744 if (!time_before(timeout, icsk->icsk_ack.timeout))
2745 timeout = icsk->icsk_ack.timeout;
2746 }
2747 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2748 icsk->icsk_ack.timeout = timeout;
2749 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2750}
2751
2752/* This routine sends an ack and also updates the window. */
2753void tcp_send_ack(struct sock *sk)
2754{
2755 struct sk_buff *buff;
2756
2757 /* If we have been reset, we may not send again. */
2758 if (sk->sk_state == TCP_CLOSE)
2759 return;
2760
2761 /* We are not putting this on the write queue, so
2762 * tcp_transmit_skb() will set the ownership to this
2763 * sock.
2764 */
2765 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2766 if (buff == NULL) {
2767 inet_csk_schedule_ack(sk);
2768 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2769 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2770 TCP_DELACK_MAX, TCP_RTO_MAX);
2771 return;
2772 }
2773
2774 /* Reserve space for headers and prepare control bits. */
2775 skb_reserve(buff, MAX_TCP_HEADER);
2776 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
2777
2778 /* Send it off, this clears delayed acks for us. */
2779 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2780 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2781}
2782
2783/* This routine sends a packet with an out of date sequence
2784 * number. It assumes the other end will try to ack it.
2785 *
2786 * Question: what should we make while urgent mode?
2787 * 4.4BSD forces sending single byte of data. We cannot send
2788 * out of window data, because we have SND.NXT==SND.MAX...
2789 *
2790 * Current solution: to send TWO zero-length segments in urgent mode:
2791 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2792 * out-of-date with SND.UNA-1 to probe window.
2793 */
2794static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2795{
2796 struct tcp_sock *tp = tcp_sk(sk);
2797 struct sk_buff *skb;
2798
2799 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2800 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2801 if (skb == NULL)
2802 return -1;
2803
2804 /* Reserve space for headers and set control bits. */
2805 skb_reserve(skb, MAX_TCP_HEADER);
2806 /* Use a previous sequence. This should cause the other
2807 * end to send an ack. Don't queue or clone SKB, just
2808 * send it.
2809 */
2810 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
2811 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2812 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2813}
2814
2815void tcp_send_window_probe(struct sock *sk)
2816{
2817 if (sk->sk_state == TCP_ESTABLISHED) {
2818 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
2819 tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq;
2820 tcp_xmit_probe_skb(sk, 0);
2821 }
2822}
2823
2824/* Initiate keepalive or window probe from timer. */
2825int tcp_write_wakeup(struct sock *sk)
2826{
2827 struct tcp_sock *tp = tcp_sk(sk);
2828 struct sk_buff *skb;
2829
2830 if (sk->sk_state == TCP_CLOSE)
2831 return -1;
2832
2833 if ((skb = tcp_send_head(sk)) != NULL &&
2834 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2835 int err;
2836 unsigned int mss = tcp_current_mss(sk);
2837 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2838
2839 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2840 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2841
2842 /* We are probing the opening of a window
2843 * but the window size is != 0
2844 * must have been a result SWS avoidance ( sender )
2845 */
2846 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2847 skb->len > mss) {
2848 seg_size = min(seg_size, mss);
2849 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
2850 if (tcp_fragment(sk, skb, seg_size, mss))
2851 return -1;
2852 } else if (!tcp_skb_pcount(skb))
2853 tcp_set_skb_tso_segs(sk, skb, mss);
2854
2855 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
2856 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2857 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2858 if (!err)
2859 tcp_event_new_data_sent(sk, skb);
2860 return err;
2861 } else {
2862 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
2863 tcp_xmit_probe_skb(sk, 1);
2864 return tcp_xmit_probe_skb(sk, 0);
2865 }
2866}
2867
2868/* A window probe timeout has occurred. If window is not closed send
2869 * a partial packet else a zero probe.
2870 */
2871void tcp_send_probe0(struct sock *sk)
2872{
2873 struct inet_connection_sock *icsk = inet_csk(sk);
2874 struct tcp_sock *tp = tcp_sk(sk);
2875 int err;
2876
2877 err = tcp_write_wakeup(sk);
2878
2879 if (tp->packets_out || !tcp_send_head(sk)) {
2880 /* Cancel probe timer, if it is not required. */
2881 icsk->icsk_probes_out = 0;
2882 icsk->icsk_backoff = 0;
2883 return;
2884 }
2885
2886 if (err <= 0) {
2887 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2888 icsk->icsk_backoff++;
2889 icsk->icsk_probes_out++;
2890 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2891 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2892 TCP_RTO_MAX);
2893 } else {
2894 /* If packet was not sent due to local congestion,
2895 * do not backoff and do not remember icsk_probes_out.
2896 * Let local senders to fight for local resources.
2897 *
2898 * Use accumulated backoff yet.
2899 */
2900 if (!icsk->icsk_probes_out)
2901 icsk->icsk_probes_out = 1;
2902 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2903 min(icsk->icsk_rto << icsk->icsk_backoff,
2904 TCP_RESOURCE_PROBE_INTERVAL),
2905 TCP_RTO_MAX);
2906 }
2907}