Loading...
1/*
2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
8 * Authors: Ross Biro
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
19 */
20
21/*
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
25 * : AF independence
26 *
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
34 *
35 */
36
37#include <net/tcp.h>
38
39#include <linux/compiler.h>
40#include <linux/gfp.h>
41#include <linux/module.h>
42
43/* People can turn this off for buggy TCP's found in printers etc. */
44int sysctl_tcp_retrans_collapse __read_mostly = 1;
45
46/* People can turn this on to work with those rare, broken TCPs that
47 * interpret the window field as a signed quantity.
48 */
49int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
50
51/* This limits the percentage of the congestion window which we
52 * will allow a single TSO frame to consume. Building TSO frames
53 * which are too large can cause TCP streams to be bursty.
54 */
55int sysctl_tcp_tso_win_divisor __read_mostly = 3;
56
57int sysctl_tcp_mtu_probing __read_mostly = 0;
58int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS;
59
60/* By default, RFC2861 behavior. */
61int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
62
63int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */
64EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size);
65
66
67/* Account for new data that has been sent to the network. */
68static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
69{
70 struct tcp_sock *tp = tcp_sk(sk);
71 unsigned int prior_packets = tp->packets_out;
72
73 tcp_advance_send_head(sk, skb);
74 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
75
76 /* Don't override Nagle indefinitely with F-RTO */
77 if (tp->frto_counter == 2)
78 tp->frto_counter = 3;
79
80 tp->packets_out += tcp_skb_pcount(skb);
81 if (!prior_packets)
82 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
83 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
84}
85
86/* SND.NXT, if window was not shrunk.
87 * If window has been shrunk, what should we make? It is not clear at all.
88 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
89 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
90 * invalid. OK, let's make this for now:
91 */
92static inline __u32 tcp_acceptable_seq(struct sock *sk)
93{
94 struct tcp_sock *tp = tcp_sk(sk);
95
96 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
97 return tp->snd_nxt;
98 else
99 return tcp_wnd_end(tp);
100}
101
102/* Calculate mss to advertise in SYN segment.
103 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
104 *
105 * 1. It is independent of path mtu.
106 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
107 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
108 * attached devices, because some buggy hosts are confused by
109 * large MSS.
110 * 4. We do not make 3, we advertise MSS, calculated from first
111 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
112 * This may be overridden via information stored in routing table.
113 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
114 * probably even Jumbo".
115 */
116static __u16 tcp_advertise_mss(struct sock *sk)
117{
118 struct tcp_sock *tp = tcp_sk(sk);
119 struct dst_entry *dst = __sk_dst_get(sk);
120 int mss = tp->advmss;
121
122 if (dst) {
123 unsigned int metric = dst_metric_advmss(dst);
124
125 if (metric < mss) {
126 mss = metric;
127 tp->advmss = mss;
128 }
129 }
130
131 return (__u16)mss;
132}
133
134/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
135 * This is the first part of cwnd validation mechanism. */
136static void tcp_cwnd_restart(struct sock *sk, struct dst_entry *dst)
137{
138 struct tcp_sock *tp = tcp_sk(sk);
139 s32 delta = tcp_time_stamp - tp->lsndtime;
140 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
141 u32 cwnd = tp->snd_cwnd;
142
143 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
144
145 tp->snd_ssthresh = tcp_current_ssthresh(sk);
146 restart_cwnd = min(restart_cwnd, cwnd);
147
148 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
149 cwnd >>= 1;
150 tp->snd_cwnd = max(cwnd, restart_cwnd);
151 tp->snd_cwnd_stamp = tcp_time_stamp;
152 tp->snd_cwnd_used = 0;
153}
154
155/* Congestion state accounting after a packet has been sent. */
156static void tcp_event_data_sent(struct tcp_sock *tp,
157 struct sk_buff *skb, struct sock *sk)
158{
159 struct inet_connection_sock *icsk = inet_csk(sk);
160 const u32 now = tcp_time_stamp;
161
162 if (sysctl_tcp_slow_start_after_idle &&
163 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
164 tcp_cwnd_restart(sk, __sk_dst_get(sk));
165
166 tp->lsndtime = now;
167
168 /* If it is a reply for ato after last received
169 * packet, enter pingpong mode.
170 */
171 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
172 icsk->icsk_ack.pingpong = 1;
173}
174
175/* Account for an ACK we sent. */
176static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
177{
178 tcp_dec_quickack_mode(sk, pkts);
179 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
180}
181
182/* Determine a window scaling and initial window to offer.
183 * Based on the assumption that the given amount of space
184 * will be offered. Store the results in the tp structure.
185 * NOTE: for smooth operation initial space offering should
186 * be a multiple of mss if possible. We assume here that mss >= 1.
187 * This MUST be enforced by all callers.
188 */
189void tcp_select_initial_window(int __space, __u32 mss,
190 __u32 *rcv_wnd, __u32 *window_clamp,
191 int wscale_ok, __u8 *rcv_wscale,
192 __u32 init_rcv_wnd)
193{
194 unsigned int space = (__space < 0 ? 0 : __space);
195
196 /* If no clamp set the clamp to the max possible scaled window */
197 if (*window_clamp == 0)
198 (*window_clamp) = (65535 << 14);
199 space = min(*window_clamp, space);
200
201 /* Quantize space offering to a multiple of mss if possible. */
202 if (space > mss)
203 space = (space / mss) * mss;
204
205 /* NOTE: offering an initial window larger than 32767
206 * will break some buggy TCP stacks. If the admin tells us
207 * it is likely we could be speaking with such a buggy stack
208 * we will truncate our initial window offering to 32K-1
209 * unless the remote has sent us a window scaling option,
210 * which we interpret as a sign the remote TCP is not
211 * misinterpreting the window field as a signed quantity.
212 */
213 if (sysctl_tcp_workaround_signed_windows)
214 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
215 else
216 (*rcv_wnd) = space;
217
218 (*rcv_wscale) = 0;
219 if (wscale_ok) {
220 /* Set window scaling on max possible window
221 * See RFC1323 for an explanation of the limit to 14
222 */
223 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
224 space = min_t(u32, space, *window_clamp);
225 while (space > 65535 && (*rcv_wscale) < 14) {
226 space >>= 1;
227 (*rcv_wscale)++;
228 }
229 }
230
231 /* Set initial window to a value enough for senders starting with
232 * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place
233 * a limit on the initial window when mss is larger than 1460.
234 */
235 if (mss > (1 << *rcv_wscale)) {
236 int init_cwnd = TCP_DEFAULT_INIT_RCVWND;
237 if (mss > 1460)
238 init_cwnd =
239 max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
240 /* when initializing use the value from init_rcv_wnd
241 * rather than the default from above
242 */
243 if (init_rcv_wnd)
244 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
245 else
246 *rcv_wnd = min(*rcv_wnd, init_cwnd * mss);
247 }
248
249 /* Set the clamp no higher than max representable value */
250 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
251}
252EXPORT_SYMBOL(tcp_select_initial_window);
253
254/* Chose a new window to advertise, update state in tcp_sock for the
255 * socket, and return result with RFC1323 scaling applied. The return
256 * value can be stuffed directly into th->window for an outgoing
257 * frame.
258 */
259static u16 tcp_select_window(struct sock *sk)
260{
261 struct tcp_sock *tp = tcp_sk(sk);
262 u32 cur_win = tcp_receive_window(tp);
263 u32 new_win = __tcp_select_window(sk);
264
265 /* Never shrink the offered window */
266 if (new_win < cur_win) {
267 /* Danger Will Robinson!
268 * Don't update rcv_wup/rcv_wnd here or else
269 * we will not be able to advertise a zero
270 * window in time. --DaveM
271 *
272 * Relax Will Robinson.
273 */
274 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
275 }
276 tp->rcv_wnd = new_win;
277 tp->rcv_wup = tp->rcv_nxt;
278
279 /* Make sure we do not exceed the maximum possible
280 * scaled window.
281 */
282 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
283 new_win = min(new_win, MAX_TCP_WINDOW);
284 else
285 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
286
287 /* RFC1323 scaling applied */
288 new_win >>= tp->rx_opt.rcv_wscale;
289
290 /* If we advertise zero window, disable fast path. */
291 if (new_win == 0)
292 tp->pred_flags = 0;
293
294 return new_win;
295}
296
297/* Packet ECN state for a SYN-ACK */
298static inline void TCP_ECN_send_synack(struct tcp_sock *tp, struct sk_buff *skb)
299{
300 TCP_SKB_CB(skb)->flags &= ~TCPHDR_CWR;
301 if (!(tp->ecn_flags & TCP_ECN_OK))
302 TCP_SKB_CB(skb)->flags &= ~TCPHDR_ECE;
303}
304
305/* Packet ECN state for a SYN. */
306static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
307{
308 struct tcp_sock *tp = tcp_sk(sk);
309
310 tp->ecn_flags = 0;
311 if (sysctl_tcp_ecn == 1) {
312 TCP_SKB_CB(skb)->flags |= TCPHDR_ECE | TCPHDR_CWR;
313 tp->ecn_flags = TCP_ECN_OK;
314 }
315}
316
317static __inline__ void
318TCP_ECN_make_synack(struct request_sock *req, struct tcphdr *th)
319{
320 if (inet_rsk(req)->ecn_ok)
321 th->ece = 1;
322}
323
324/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
325 * be sent.
326 */
327static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
328 int tcp_header_len)
329{
330 struct tcp_sock *tp = tcp_sk(sk);
331
332 if (tp->ecn_flags & TCP_ECN_OK) {
333 /* Not-retransmitted data segment: set ECT and inject CWR. */
334 if (skb->len != tcp_header_len &&
335 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
336 INET_ECN_xmit(sk);
337 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
338 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
339 tcp_hdr(skb)->cwr = 1;
340 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
341 }
342 } else {
343 /* ACK or retransmitted segment: clear ECT|CE */
344 INET_ECN_dontxmit(sk);
345 }
346 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
347 tcp_hdr(skb)->ece = 1;
348 }
349}
350
351/* Constructs common control bits of non-data skb. If SYN/FIN is present,
352 * auto increment end seqno.
353 */
354static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
355{
356 skb->ip_summed = CHECKSUM_PARTIAL;
357 skb->csum = 0;
358
359 TCP_SKB_CB(skb)->flags = flags;
360 TCP_SKB_CB(skb)->sacked = 0;
361
362 skb_shinfo(skb)->gso_segs = 1;
363 skb_shinfo(skb)->gso_size = 0;
364 skb_shinfo(skb)->gso_type = 0;
365
366 TCP_SKB_CB(skb)->seq = seq;
367 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
368 seq++;
369 TCP_SKB_CB(skb)->end_seq = seq;
370}
371
372static inline int tcp_urg_mode(const struct tcp_sock *tp)
373{
374 return tp->snd_una != tp->snd_up;
375}
376
377#define OPTION_SACK_ADVERTISE (1 << 0)
378#define OPTION_TS (1 << 1)
379#define OPTION_MD5 (1 << 2)
380#define OPTION_WSCALE (1 << 3)
381#define OPTION_COOKIE_EXTENSION (1 << 4)
382
383struct tcp_out_options {
384 u8 options; /* bit field of OPTION_* */
385 u8 ws; /* window scale, 0 to disable */
386 u8 num_sack_blocks; /* number of SACK blocks to include */
387 u8 hash_size; /* bytes in hash_location */
388 u16 mss; /* 0 to disable */
389 __u32 tsval, tsecr; /* need to include OPTION_TS */
390 __u8 *hash_location; /* temporary pointer, overloaded */
391};
392
393/* The sysctl int routines are generic, so check consistency here.
394 */
395static u8 tcp_cookie_size_check(u8 desired)
396{
397 int cookie_size;
398
399 if (desired > 0)
400 /* previously specified */
401 return desired;
402
403 cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size);
404 if (cookie_size <= 0)
405 /* no default specified */
406 return 0;
407
408 if (cookie_size <= TCP_COOKIE_MIN)
409 /* value too small, specify minimum */
410 return TCP_COOKIE_MIN;
411
412 if (cookie_size >= TCP_COOKIE_MAX)
413 /* value too large, specify maximum */
414 return TCP_COOKIE_MAX;
415
416 if (cookie_size & 1)
417 /* 8-bit multiple, illegal, fix it */
418 cookie_size++;
419
420 return (u8)cookie_size;
421}
422
423/* Write previously computed TCP options to the packet.
424 *
425 * Beware: Something in the Internet is very sensitive to the ordering of
426 * TCP options, we learned this through the hard way, so be careful here.
427 * Luckily we can at least blame others for their non-compliance but from
428 * inter-operatibility perspective it seems that we're somewhat stuck with
429 * the ordering which we have been using if we want to keep working with
430 * those broken things (not that it currently hurts anybody as there isn't
431 * particular reason why the ordering would need to be changed).
432 *
433 * At least SACK_PERM as the first option is known to lead to a disaster
434 * (but it may well be that other scenarios fail similarly).
435 */
436static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
437 struct tcp_out_options *opts)
438{
439 u8 options = opts->options; /* mungable copy */
440
441 /* Having both authentication and cookies for security is redundant,
442 * and there's certainly not enough room. Instead, the cookie-less
443 * extension variant is proposed.
444 *
445 * Consider the pessimal case with authentication. The options
446 * could look like:
447 * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
448 */
449 if (unlikely(OPTION_MD5 & options)) {
450 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
451 *ptr++ = htonl((TCPOPT_COOKIE << 24) |
452 (TCPOLEN_COOKIE_BASE << 16) |
453 (TCPOPT_MD5SIG << 8) |
454 TCPOLEN_MD5SIG);
455 } else {
456 *ptr++ = htonl((TCPOPT_NOP << 24) |
457 (TCPOPT_NOP << 16) |
458 (TCPOPT_MD5SIG << 8) |
459 TCPOLEN_MD5SIG);
460 }
461 options &= ~OPTION_COOKIE_EXTENSION;
462 /* overload cookie hash location */
463 opts->hash_location = (__u8 *)ptr;
464 ptr += 4;
465 }
466
467 if (unlikely(opts->mss)) {
468 *ptr++ = htonl((TCPOPT_MSS << 24) |
469 (TCPOLEN_MSS << 16) |
470 opts->mss);
471 }
472
473 if (likely(OPTION_TS & options)) {
474 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
475 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
476 (TCPOLEN_SACK_PERM << 16) |
477 (TCPOPT_TIMESTAMP << 8) |
478 TCPOLEN_TIMESTAMP);
479 options &= ~OPTION_SACK_ADVERTISE;
480 } else {
481 *ptr++ = htonl((TCPOPT_NOP << 24) |
482 (TCPOPT_NOP << 16) |
483 (TCPOPT_TIMESTAMP << 8) |
484 TCPOLEN_TIMESTAMP);
485 }
486 *ptr++ = htonl(opts->tsval);
487 *ptr++ = htonl(opts->tsecr);
488 }
489
490 /* Specification requires after timestamp, so do it now.
491 *
492 * Consider the pessimal case without authentication. The options
493 * could look like:
494 * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
495 */
496 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
497 __u8 *cookie_copy = opts->hash_location;
498 u8 cookie_size = opts->hash_size;
499
500 /* 8-bit multiple handled in tcp_cookie_size_check() above,
501 * and elsewhere.
502 */
503 if (0x2 & cookie_size) {
504 __u8 *p = (__u8 *)ptr;
505
506 /* 16-bit multiple */
507 *p++ = TCPOPT_COOKIE;
508 *p++ = TCPOLEN_COOKIE_BASE + cookie_size;
509 *p++ = *cookie_copy++;
510 *p++ = *cookie_copy++;
511 ptr++;
512 cookie_size -= 2;
513 } else {
514 /* 32-bit multiple */
515 *ptr++ = htonl(((TCPOPT_NOP << 24) |
516 (TCPOPT_NOP << 16) |
517 (TCPOPT_COOKIE << 8) |
518 TCPOLEN_COOKIE_BASE) +
519 cookie_size);
520 }
521
522 if (cookie_size > 0) {
523 memcpy(ptr, cookie_copy, cookie_size);
524 ptr += (cookie_size / 4);
525 }
526 }
527
528 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
529 *ptr++ = htonl((TCPOPT_NOP << 24) |
530 (TCPOPT_NOP << 16) |
531 (TCPOPT_SACK_PERM << 8) |
532 TCPOLEN_SACK_PERM);
533 }
534
535 if (unlikely(OPTION_WSCALE & options)) {
536 *ptr++ = htonl((TCPOPT_NOP << 24) |
537 (TCPOPT_WINDOW << 16) |
538 (TCPOLEN_WINDOW << 8) |
539 opts->ws);
540 }
541
542 if (unlikely(opts->num_sack_blocks)) {
543 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
544 tp->duplicate_sack : tp->selective_acks;
545 int this_sack;
546
547 *ptr++ = htonl((TCPOPT_NOP << 24) |
548 (TCPOPT_NOP << 16) |
549 (TCPOPT_SACK << 8) |
550 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
551 TCPOLEN_SACK_PERBLOCK)));
552
553 for (this_sack = 0; this_sack < opts->num_sack_blocks;
554 ++this_sack) {
555 *ptr++ = htonl(sp[this_sack].start_seq);
556 *ptr++ = htonl(sp[this_sack].end_seq);
557 }
558
559 tp->rx_opt.dsack = 0;
560 }
561}
562
563/* Compute TCP options for SYN packets. This is not the final
564 * network wire format yet.
565 */
566static unsigned tcp_syn_options(struct sock *sk, struct sk_buff *skb,
567 struct tcp_out_options *opts,
568 struct tcp_md5sig_key **md5) {
569 struct tcp_sock *tp = tcp_sk(sk);
570 struct tcp_cookie_values *cvp = tp->cookie_values;
571 unsigned remaining = MAX_TCP_OPTION_SPACE;
572 u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ?
573 tcp_cookie_size_check(cvp->cookie_desired) :
574 0;
575
576#ifdef CONFIG_TCP_MD5SIG
577 *md5 = tp->af_specific->md5_lookup(sk, sk);
578 if (*md5) {
579 opts->options |= OPTION_MD5;
580 remaining -= TCPOLEN_MD5SIG_ALIGNED;
581 }
582#else
583 *md5 = NULL;
584#endif
585
586 /* We always get an MSS option. The option bytes which will be seen in
587 * normal data packets should timestamps be used, must be in the MSS
588 * advertised. But we subtract them from tp->mss_cache so that
589 * calculations in tcp_sendmsg are simpler etc. So account for this
590 * fact here if necessary. If we don't do this correctly, as a
591 * receiver we won't recognize data packets as being full sized when we
592 * should, and thus we won't abide by the delayed ACK rules correctly.
593 * SACKs don't matter, we never delay an ACK when we have any of those
594 * going out. */
595 opts->mss = tcp_advertise_mss(sk);
596 remaining -= TCPOLEN_MSS_ALIGNED;
597
598 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
599 opts->options |= OPTION_TS;
600 opts->tsval = TCP_SKB_CB(skb)->when;
601 opts->tsecr = tp->rx_opt.ts_recent;
602 remaining -= TCPOLEN_TSTAMP_ALIGNED;
603 }
604 if (likely(sysctl_tcp_window_scaling)) {
605 opts->ws = tp->rx_opt.rcv_wscale;
606 opts->options |= OPTION_WSCALE;
607 remaining -= TCPOLEN_WSCALE_ALIGNED;
608 }
609 if (likely(sysctl_tcp_sack)) {
610 opts->options |= OPTION_SACK_ADVERTISE;
611 if (unlikely(!(OPTION_TS & opts->options)))
612 remaining -= TCPOLEN_SACKPERM_ALIGNED;
613 }
614
615 /* Note that timestamps are required by the specification.
616 *
617 * Odd numbers of bytes are prohibited by the specification, ensuring
618 * that the cookie is 16-bit aligned, and the resulting cookie pair is
619 * 32-bit aligned.
620 */
621 if (*md5 == NULL &&
622 (OPTION_TS & opts->options) &&
623 cookie_size > 0) {
624 int need = TCPOLEN_COOKIE_BASE + cookie_size;
625
626 if (0x2 & need) {
627 /* 32-bit multiple */
628 need += 2; /* NOPs */
629
630 if (need > remaining) {
631 /* try shrinking cookie to fit */
632 cookie_size -= 2;
633 need -= 4;
634 }
635 }
636 while (need > remaining && TCP_COOKIE_MIN <= cookie_size) {
637 cookie_size -= 4;
638 need -= 4;
639 }
640 if (TCP_COOKIE_MIN <= cookie_size) {
641 opts->options |= OPTION_COOKIE_EXTENSION;
642 opts->hash_location = (__u8 *)&cvp->cookie_pair[0];
643 opts->hash_size = cookie_size;
644
645 /* Remember for future incarnations. */
646 cvp->cookie_desired = cookie_size;
647
648 if (cvp->cookie_desired != cvp->cookie_pair_size) {
649 /* Currently use random bytes as a nonce,
650 * assuming these are completely unpredictable
651 * by hostile users of the same system.
652 */
653 get_random_bytes(&cvp->cookie_pair[0],
654 cookie_size);
655 cvp->cookie_pair_size = cookie_size;
656 }
657
658 remaining -= need;
659 }
660 }
661 return MAX_TCP_OPTION_SPACE - remaining;
662}
663
664/* Set up TCP options for SYN-ACKs. */
665static unsigned tcp_synack_options(struct sock *sk,
666 struct request_sock *req,
667 unsigned mss, struct sk_buff *skb,
668 struct tcp_out_options *opts,
669 struct tcp_md5sig_key **md5,
670 struct tcp_extend_values *xvp)
671{
672 struct inet_request_sock *ireq = inet_rsk(req);
673 unsigned remaining = MAX_TCP_OPTION_SPACE;
674 u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ?
675 xvp->cookie_plus :
676 0;
677
678#ifdef CONFIG_TCP_MD5SIG
679 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
680 if (*md5) {
681 opts->options |= OPTION_MD5;
682 remaining -= TCPOLEN_MD5SIG_ALIGNED;
683
684 /* We can't fit any SACK blocks in a packet with MD5 + TS
685 * options. There was discussion about disabling SACK
686 * rather than TS in order to fit in better with old,
687 * buggy kernels, but that was deemed to be unnecessary.
688 */
689 ireq->tstamp_ok &= !ireq->sack_ok;
690 }
691#else
692 *md5 = NULL;
693#endif
694
695 /* We always send an MSS option. */
696 opts->mss = mss;
697 remaining -= TCPOLEN_MSS_ALIGNED;
698
699 if (likely(ireq->wscale_ok)) {
700 opts->ws = ireq->rcv_wscale;
701 opts->options |= OPTION_WSCALE;
702 remaining -= TCPOLEN_WSCALE_ALIGNED;
703 }
704 if (likely(ireq->tstamp_ok)) {
705 opts->options |= OPTION_TS;
706 opts->tsval = TCP_SKB_CB(skb)->when;
707 opts->tsecr = req->ts_recent;
708 remaining -= TCPOLEN_TSTAMP_ALIGNED;
709 }
710 if (likely(ireq->sack_ok)) {
711 opts->options |= OPTION_SACK_ADVERTISE;
712 if (unlikely(!ireq->tstamp_ok))
713 remaining -= TCPOLEN_SACKPERM_ALIGNED;
714 }
715
716 /* Similar rationale to tcp_syn_options() applies here, too.
717 * If the <SYN> options fit, the same options should fit now!
718 */
719 if (*md5 == NULL &&
720 ireq->tstamp_ok &&
721 cookie_plus > TCPOLEN_COOKIE_BASE) {
722 int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */
723
724 if (0x2 & need) {
725 /* 32-bit multiple */
726 need += 2; /* NOPs */
727 }
728 if (need <= remaining) {
729 opts->options |= OPTION_COOKIE_EXTENSION;
730 opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE;
731 remaining -= need;
732 } else {
733 /* There's no error return, so flag it. */
734 xvp->cookie_out_never = 1; /* true */
735 opts->hash_size = 0;
736 }
737 }
738 return MAX_TCP_OPTION_SPACE - remaining;
739}
740
741/* Compute TCP options for ESTABLISHED sockets. This is not the
742 * final wire format yet.
743 */
744static unsigned tcp_established_options(struct sock *sk, struct sk_buff *skb,
745 struct tcp_out_options *opts,
746 struct tcp_md5sig_key **md5) {
747 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
748 struct tcp_sock *tp = tcp_sk(sk);
749 unsigned size = 0;
750 unsigned int eff_sacks;
751
752#ifdef CONFIG_TCP_MD5SIG
753 *md5 = tp->af_specific->md5_lookup(sk, sk);
754 if (unlikely(*md5)) {
755 opts->options |= OPTION_MD5;
756 size += TCPOLEN_MD5SIG_ALIGNED;
757 }
758#else
759 *md5 = NULL;
760#endif
761
762 if (likely(tp->rx_opt.tstamp_ok)) {
763 opts->options |= OPTION_TS;
764 opts->tsval = tcb ? tcb->when : 0;
765 opts->tsecr = tp->rx_opt.ts_recent;
766 size += TCPOLEN_TSTAMP_ALIGNED;
767 }
768
769 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
770 if (unlikely(eff_sacks)) {
771 const unsigned remaining = MAX_TCP_OPTION_SPACE - size;
772 opts->num_sack_blocks =
773 min_t(unsigned, eff_sacks,
774 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
775 TCPOLEN_SACK_PERBLOCK);
776 size += TCPOLEN_SACK_BASE_ALIGNED +
777 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
778 }
779
780 return size;
781}
782
783/* This routine actually transmits TCP packets queued in by
784 * tcp_do_sendmsg(). This is used by both the initial
785 * transmission and possible later retransmissions.
786 * All SKB's seen here are completely headerless. It is our
787 * job to build the TCP header, and pass the packet down to
788 * IP so it can do the same plus pass the packet off to the
789 * device.
790 *
791 * We are working here with either a clone of the original
792 * SKB, or a fresh unique copy made by the retransmit engine.
793 */
794static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
795 gfp_t gfp_mask)
796{
797 const struct inet_connection_sock *icsk = inet_csk(sk);
798 struct inet_sock *inet;
799 struct tcp_sock *tp;
800 struct tcp_skb_cb *tcb;
801 struct tcp_out_options opts;
802 unsigned tcp_options_size, tcp_header_size;
803 struct tcp_md5sig_key *md5;
804 struct tcphdr *th;
805 int err;
806
807 BUG_ON(!skb || !tcp_skb_pcount(skb));
808
809 /* If congestion control is doing timestamping, we must
810 * take such a timestamp before we potentially clone/copy.
811 */
812 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
813 __net_timestamp(skb);
814
815 if (likely(clone_it)) {
816 if (unlikely(skb_cloned(skb)))
817 skb = pskb_copy(skb, gfp_mask);
818 else
819 skb = skb_clone(skb, gfp_mask);
820 if (unlikely(!skb))
821 return -ENOBUFS;
822 }
823
824 inet = inet_sk(sk);
825 tp = tcp_sk(sk);
826 tcb = TCP_SKB_CB(skb);
827 memset(&opts, 0, sizeof(opts));
828
829 if (unlikely(tcb->flags & TCPHDR_SYN))
830 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
831 else
832 tcp_options_size = tcp_established_options(sk, skb, &opts,
833 &md5);
834 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
835
836 if (tcp_packets_in_flight(tp) == 0) {
837 tcp_ca_event(sk, CA_EVENT_TX_START);
838 skb->ooo_okay = 1;
839 } else
840 skb->ooo_okay = 0;
841
842 skb_push(skb, tcp_header_size);
843 skb_reset_transport_header(skb);
844 skb_set_owner_w(skb, sk);
845
846 /* Build TCP header and checksum it. */
847 th = tcp_hdr(skb);
848 th->source = inet->inet_sport;
849 th->dest = inet->inet_dport;
850 th->seq = htonl(tcb->seq);
851 th->ack_seq = htonl(tp->rcv_nxt);
852 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
853 tcb->flags);
854
855 if (unlikely(tcb->flags & TCPHDR_SYN)) {
856 /* RFC1323: The window in SYN & SYN/ACK segments
857 * is never scaled.
858 */
859 th->window = htons(min(tp->rcv_wnd, 65535U));
860 } else {
861 th->window = htons(tcp_select_window(sk));
862 }
863 th->check = 0;
864 th->urg_ptr = 0;
865
866 /* The urg_mode check is necessary during a below snd_una win probe */
867 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
868 if (before(tp->snd_up, tcb->seq + 0x10000)) {
869 th->urg_ptr = htons(tp->snd_up - tcb->seq);
870 th->urg = 1;
871 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
872 th->urg_ptr = htons(0xFFFF);
873 th->urg = 1;
874 }
875 }
876
877 tcp_options_write((__be32 *)(th + 1), tp, &opts);
878 if (likely((tcb->flags & TCPHDR_SYN) == 0))
879 TCP_ECN_send(sk, skb, tcp_header_size);
880
881#ifdef CONFIG_TCP_MD5SIG
882 /* Calculate the MD5 hash, as we have all we need now */
883 if (md5) {
884 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
885 tp->af_specific->calc_md5_hash(opts.hash_location,
886 md5, sk, NULL, skb);
887 }
888#endif
889
890 icsk->icsk_af_ops->send_check(sk, skb);
891
892 if (likely(tcb->flags & TCPHDR_ACK))
893 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
894
895 if (skb->len != tcp_header_size)
896 tcp_event_data_sent(tp, skb, sk);
897
898 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
899 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
900 tcp_skb_pcount(skb));
901
902 err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
903 if (likely(err <= 0))
904 return err;
905
906 tcp_enter_cwr(sk, 1);
907
908 return net_xmit_eval(err);
909}
910
911/* This routine just queues the buffer for sending.
912 *
913 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
914 * otherwise socket can stall.
915 */
916static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
917{
918 struct tcp_sock *tp = tcp_sk(sk);
919
920 /* Advance write_seq and place onto the write_queue. */
921 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
922 skb_header_release(skb);
923 tcp_add_write_queue_tail(sk, skb);
924 sk->sk_wmem_queued += skb->truesize;
925 sk_mem_charge(sk, skb->truesize);
926}
927
928/* Initialize TSO segments for a packet. */
929static void tcp_set_skb_tso_segs(struct sock *sk, struct sk_buff *skb,
930 unsigned int mss_now)
931{
932 if (skb->len <= mss_now || !sk_can_gso(sk) ||
933 skb->ip_summed == CHECKSUM_NONE) {
934 /* Avoid the costly divide in the normal
935 * non-TSO case.
936 */
937 skb_shinfo(skb)->gso_segs = 1;
938 skb_shinfo(skb)->gso_size = 0;
939 skb_shinfo(skb)->gso_type = 0;
940 } else {
941 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
942 skb_shinfo(skb)->gso_size = mss_now;
943 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
944 }
945}
946
947/* When a modification to fackets out becomes necessary, we need to check
948 * skb is counted to fackets_out or not.
949 */
950static void tcp_adjust_fackets_out(struct sock *sk, struct sk_buff *skb,
951 int decr)
952{
953 struct tcp_sock *tp = tcp_sk(sk);
954
955 if (!tp->sacked_out || tcp_is_reno(tp))
956 return;
957
958 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
959 tp->fackets_out -= decr;
960}
961
962/* Pcount in the middle of the write queue got changed, we need to do various
963 * tweaks to fix counters
964 */
965static void tcp_adjust_pcount(struct sock *sk, struct sk_buff *skb, int decr)
966{
967 struct tcp_sock *tp = tcp_sk(sk);
968
969 tp->packets_out -= decr;
970
971 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
972 tp->sacked_out -= decr;
973 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
974 tp->retrans_out -= decr;
975 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
976 tp->lost_out -= decr;
977
978 /* Reno case is special. Sigh... */
979 if (tcp_is_reno(tp) && decr > 0)
980 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
981
982 tcp_adjust_fackets_out(sk, skb, decr);
983
984 if (tp->lost_skb_hint &&
985 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
986 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
987 tp->lost_cnt_hint -= decr;
988
989 tcp_verify_left_out(tp);
990}
991
992/* Function to create two new TCP segments. Shrinks the given segment
993 * to the specified size and appends a new segment with the rest of the
994 * packet to the list. This won't be called frequently, I hope.
995 * Remember, these are still headerless SKBs at this point.
996 */
997int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
998 unsigned int mss_now)
999{
1000 struct tcp_sock *tp = tcp_sk(sk);
1001 struct sk_buff *buff;
1002 int nsize, old_factor;
1003 int nlen;
1004 u8 flags;
1005
1006 if (WARN_ON(len > skb->len))
1007 return -EINVAL;
1008
1009 nsize = skb_headlen(skb) - len;
1010 if (nsize < 0)
1011 nsize = 0;
1012
1013 if (skb_cloned(skb) &&
1014 skb_is_nonlinear(skb) &&
1015 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1016 return -ENOMEM;
1017
1018 /* Get a new skb... force flag on. */
1019 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1020 if (buff == NULL)
1021 return -ENOMEM; /* We'll just try again later. */
1022
1023 sk->sk_wmem_queued += buff->truesize;
1024 sk_mem_charge(sk, buff->truesize);
1025 nlen = skb->len - len - nsize;
1026 buff->truesize += nlen;
1027 skb->truesize -= nlen;
1028
1029 /* Correct the sequence numbers. */
1030 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1031 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1032 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1033
1034 /* PSH and FIN should only be set in the second packet. */
1035 flags = TCP_SKB_CB(skb)->flags;
1036 TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1037 TCP_SKB_CB(buff)->flags = flags;
1038 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1039
1040 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1041 /* Copy and checksum data tail into the new buffer. */
1042 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1043 skb_put(buff, nsize),
1044 nsize, 0);
1045
1046 skb_trim(skb, len);
1047
1048 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1049 } else {
1050 skb->ip_summed = CHECKSUM_PARTIAL;
1051 skb_split(skb, buff, len);
1052 }
1053
1054 buff->ip_summed = skb->ip_summed;
1055
1056 /* Looks stupid, but our code really uses when of
1057 * skbs, which it never sent before. --ANK
1058 */
1059 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1060 buff->tstamp = skb->tstamp;
1061
1062 old_factor = tcp_skb_pcount(skb);
1063
1064 /* Fix up tso_factor for both original and new SKB. */
1065 tcp_set_skb_tso_segs(sk, skb, mss_now);
1066 tcp_set_skb_tso_segs(sk, buff, mss_now);
1067
1068 /* If this packet has been sent out already, we must
1069 * adjust the various packet counters.
1070 */
1071 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1072 int diff = old_factor - tcp_skb_pcount(skb) -
1073 tcp_skb_pcount(buff);
1074
1075 if (diff)
1076 tcp_adjust_pcount(sk, skb, diff);
1077 }
1078
1079 /* Link BUFF into the send queue. */
1080 skb_header_release(buff);
1081 tcp_insert_write_queue_after(skb, buff, sk);
1082
1083 return 0;
1084}
1085
1086/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1087 * eventually). The difference is that pulled data not copied, but
1088 * immediately discarded.
1089 */
1090static void __pskb_trim_head(struct sk_buff *skb, int len)
1091{
1092 int i, k, eat;
1093
1094 eat = len;
1095 k = 0;
1096 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1097 if (skb_shinfo(skb)->frags[i].size <= eat) {
1098 put_page(skb_shinfo(skb)->frags[i].page);
1099 eat -= skb_shinfo(skb)->frags[i].size;
1100 } else {
1101 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1102 if (eat) {
1103 skb_shinfo(skb)->frags[k].page_offset += eat;
1104 skb_shinfo(skb)->frags[k].size -= eat;
1105 eat = 0;
1106 }
1107 k++;
1108 }
1109 }
1110 skb_shinfo(skb)->nr_frags = k;
1111
1112 skb_reset_tail_pointer(skb);
1113 skb->data_len -= len;
1114 skb->len = skb->data_len;
1115}
1116
1117/* Remove acked data from a packet in the transmit queue. */
1118int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1119{
1120 if (skb_cloned(skb) && pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1121 return -ENOMEM;
1122
1123 /* If len == headlen, we avoid __skb_pull to preserve alignment. */
1124 if (unlikely(len < skb_headlen(skb)))
1125 __skb_pull(skb, len);
1126 else
1127 __pskb_trim_head(skb, len - skb_headlen(skb));
1128
1129 TCP_SKB_CB(skb)->seq += len;
1130 skb->ip_summed = CHECKSUM_PARTIAL;
1131
1132 skb->truesize -= len;
1133 sk->sk_wmem_queued -= len;
1134 sk_mem_uncharge(sk, len);
1135 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1136
1137 /* Any change of skb->len requires recalculation of tso
1138 * factor and mss.
1139 */
1140 if (tcp_skb_pcount(skb) > 1)
1141 tcp_set_skb_tso_segs(sk, skb, tcp_current_mss(sk));
1142
1143 return 0;
1144}
1145
1146/* Calculate MSS. Not accounting for SACKs here. */
1147int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1148{
1149 struct tcp_sock *tp = tcp_sk(sk);
1150 struct inet_connection_sock *icsk = inet_csk(sk);
1151 int mss_now;
1152
1153 /* Calculate base mss without TCP options:
1154 It is MMS_S - sizeof(tcphdr) of rfc1122
1155 */
1156 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1157
1158 /* Clamp it (mss_clamp does not include tcp options) */
1159 if (mss_now > tp->rx_opt.mss_clamp)
1160 mss_now = tp->rx_opt.mss_clamp;
1161
1162 /* Now subtract optional transport overhead */
1163 mss_now -= icsk->icsk_ext_hdr_len;
1164
1165 /* Then reserve room for full set of TCP options and 8 bytes of data */
1166 if (mss_now < 48)
1167 mss_now = 48;
1168
1169 /* Now subtract TCP options size, not including SACKs */
1170 mss_now -= tp->tcp_header_len - sizeof(struct tcphdr);
1171
1172 return mss_now;
1173}
1174
1175/* Inverse of above */
1176int tcp_mss_to_mtu(struct sock *sk, int mss)
1177{
1178 struct tcp_sock *tp = tcp_sk(sk);
1179 struct inet_connection_sock *icsk = inet_csk(sk);
1180 int mtu;
1181
1182 mtu = mss +
1183 tp->tcp_header_len +
1184 icsk->icsk_ext_hdr_len +
1185 icsk->icsk_af_ops->net_header_len;
1186
1187 return mtu;
1188}
1189
1190/* MTU probing init per socket */
1191void tcp_mtup_init(struct sock *sk)
1192{
1193 struct tcp_sock *tp = tcp_sk(sk);
1194 struct inet_connection_sock *icsk = inet_csk(sk);
1195
1196 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1197 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1198 icsk->icsk_af_ops->net_header_len;
1199 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1200 icsk->icsk_mtup.probe_size = 0;
1201}
1202EXPORT_SYMBOL(tcp_mtup_init);
1203
1204/* This function synchronize snd mss to current pmtu/exthdr set.
1205
1206 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1207 for TCP options, but includes only bare TCP header.
1208
1209 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1210 It is minimum of user_mss and mss received with SYN.
1211 It also does not include TCP options.
1212
1213 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1214
1215 tp->mss_cache is current effective sending mss, including
1216 all tcp options except for SACKs. It is evaluated,
1217 taking into account current pmtu, but never exceeds
1218 tp->rx_opt.mss_clamp.
1219
1220 NOTE1. rfc1122 clearly states that advertised MSS
1221 DOES NOT include either tcp or ip options.
1222
1223 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1224 are READ ONLY outside this function. --ANK (980731)
1225 */
1226unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1227{
1228 struct tcp_sock *tp = tcp_sk(sk);
1229 struct inet_connection_sock *icsk = inet_csk(sk);
1230 int mss_now;
1231
1232 if (icsk->icsk_mtup.search_high > pmtu)
1233 icsk->icsk_mtup.search_high = pmtu;
1234
1235 mss_now = tcp_mtu_to_mss(sk, pmtu);
1236 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1237
1238 /* And store cached results */
1239 icsk->icsk_pmtu_cookie = pmtu;
1240 if (icsk->icsk_mtup.enabled)
1241 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1242 tp->mss_cache = mss_now;
1243
1244 return mss_now;
1245}
1246EXPORT_SYMBOL(tcp_sync_mss);
1247
1248/* Compute the current effective MSS, taking SACKs and IP options,
1249 * and even PMTU discovery events into account.
1250 */
1251unsigned int tcp_current_mss(struct sock *sk)
1252{
1253 struct tcp_sock *tp = tcp_sk(sk);
1254 struct dst_entry *dst = __sk_dst_get(sk);
1255 u32 mss_now;
1256 unsigned header_len;
1257 struct tcp_out_options opts;
1258 struct tcp_md5sig_key *md5;
1259
1260 mss_now = tp->mss_cache;
1261
1262 if (dst) {
1263 u32 mtu = dst_mtu(dst);
1264 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1265 mss_now = tcp_sync_mss(sk, mtu);
1266 }
1267
1268 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1269 sizeof(struct tcphdr);
1270 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1271 * some common options. If this is an odd packet (because we have SACK
1272 * blocks etc) then our calculated header_len will be different, and
1273 * we have to adjust mss_now correspondingly */
1274 if (header_len != tp->tcp_header_len) {
1275 int delta = (int) header_len - tp->tcp_header_len;
1276 mss_now -= delta;
1277 }
1278
1279 return mss_now;
1280}
1281
1282/* Congestion window validation. (RFC2861) */
1283static void tcp_cwnd_validate(struct sock *sk)
1284{
1285 struct tcp_sock *tp = tcp_sk(sk);
1286
1287 if (tp->packets_out >= tp->snd_cwnd) {
1288 /* Network is feed fully. */
1289 tp->snd_cwnd_used = 0;
1290 tp->snd_cwnd_stamp = tcp_time_stamp;
1291 } else {
1292 /* Network starves. */
1293 if (tp->packets_out > tp->snd_cwnd_used)
1294 tp->snd_cwnd_used = tp->packets_out;
1295
1296 if (sysctl_tcp_slow_start_after_idle &&
1297 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1298 tcp_cwnd_application_limited(sk);
1299 }
1300}
1301
1302/* Returns the portion of skb which can be sent right away without
1303 * introducing MSS oddities to segment boundaries. In rare cases where
1304 * mss_now != mss_cache, we will request caller to create a small skb
1305 * per input skb which could be mostly avoided here (if desired).
1306 *
1307 * We explicitly want to create a request for splitting write queue tail
1308 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1309 * thus all the complexity (cwnd_len is always MSS multiple which we
1310 * return whenever allowed by the other factors). Basically we need the
1311 * modulo only when the receiver window alone is the limiting factor or
1312 * when we would be allowed to send the split-due-to-Nagle skb fully.
1313 */
1314static unsigned int tcp_mss_split_point(struct sock *sk, struct sk_buff *skb,
1315 unsigned int mss_now, unsigned int cwnd)
1316{
1317 struct tcp_sock *tp = tcp_sk(sk);
1318 u32 needed, window, cwnd_len;
1319
1320 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1321 cwnd_len = mss_now * cwnd;
1322
1323 if (likely(cwnd_len <= window && skb != tcp_write_queue_tail(sk)))
1324 return cwnd_len;
1325
1326 needed = min(skb->len, window);
1327
1328 if (cwnd_len <= needed)
1329 return cwnd_len;
1330
1331 return needed - needed % mss_now;
1332}
1333
1334/* Can at least one segment of SKB be sent right now, according to the
1335 * congestion window rules? If so, return how many segments are allowed.
1336 */
1337static inline unsigned int tcp_cwnd_test(struct tcp_sock *tp,
1338 struct sk_buff *skb)
1339{
1340 u32 in_flight, cwnd;
1341
1342 /* Don't be strict about the congestion window for the final FIN. */
1343 if ((TCP_SKB_CB(skb)->flags & TCPHDR_FIN) && tcp_skb_pcount(skb) == 1)
1344 return 1;
1345
1346 in_flight = tcp_packets_in_flight(tp);
1347 cwnd = tp->snd_cwnd;
1348 if (in_flight < cwnd)
1349 return (cwnd - in_flight);
1350
1351 return 0;
1352}
1353
1354/* Initialize TSO state of a skb.
1355 * This must be invoked the first time we consider transmitting
1356 * SKB onto the wire.
1357 */
1358static int tcp_init_tso_segs(struct sock *sk, struct sk_buff *skb,
1359 unsigned int mss_now)
1360{
1361 int tso_segs = tcp_skb_pcount(skb);
1362
1363 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1364 tcp_set_skb_tso_segs(sk, skb, mss_now);
1365 tso_segs = tcp_skb_pcount(skb);
1366 }
1367 return tso_segs;
1368}
1369
1370/* Minshall's variant of the Nagle send check. */
1371static inline int tcp_minshall_check(const struct tcp_sock *tp)
1372{
1373 return after(tp->snd_sml, tp->snd_una) &&
1374 !after(tp->snd_sml, tp->snd_nxt);
1375}
1376
1377/* Return 0, if packet can be sent now without violation Nagle's rules:
1378 * 1. It is full sized.
1379 * 2. Or it contains FIN. (already checked by caller)
1380 * 3. Or TCP_NODELAY was set.
1381 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1382 * With Minshall's modification: all sent small packets are ACKed.
1383 */
1384static inline int tcp_nagle_check(const struct tcp_sock *tp,
1385 const struct sk_buff *skb,
1386 unsigned mss_now, int nonagle)
1387{
1388 return skb->len < mss_now &&
1389 ((nonagle & TCP_NAGLE_CORK) ||
1390 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1391}
1392
1393/* Return non-zero if the Nagle test allows this packet to be
1394 * sent now.
1395 */
1396static inline int tcp_nagle_test(struct tcp_sock *tp, struct sk_buff *skb,
1397 unsigned int cur_mss, int nonagle)
1398{
1399 /* Nagle rule does not apply to frames, which sit in the middle of the
1400 * write_queue (they have no chances to get new data).
1401 *
1402 * This is implemented in the callers, where they modify the 'nonagle'
1403 * argument based upon the location of SKB in the send queue.
1404 */
1405 if (nonagle & TCP_NAGLE_PUSH)
1406 return 1;
1407
1408 /* Don't use the nagle rule for urgent data (or for the final FIN).
1409 * Nagle can be ignored during F-RTO too (see RFC4138).
1410 */
1411 if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
1412 (TCP_SKB_CB(skb)->flags & TCPHDR_FIN))
1413 return 1;
1414
1415 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1416 return 1;
1417
1418 return 0;
1419}
1420
1421/* Does at least the first segment of SKB fit into the send window? */
1422static inline int tcp_snd_wnd_test(struct tcp_sock *tp, struct sk_buff *skb,
1423 unsigned int cur_mss)
1424{
1425 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1426
1427 if (skb->len > cur_mss)
1428 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1429
1430 return !after(end_seq, tcp_wnd_end(tp));
1431}
1432
1433/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1434 * should be put on the wire right now. If so, it returns the number of
1435 * packets allowed by the congestion window.
1436 */
1437static unsigned int tcp_snd_test(struct sock *sk, struct sk_buff *skb,
1438 unsigned int cur_mss, int nonagle)
1439{
1440 struct tcp_sock *tp = tcp_sk(sk);
1441 unsigned int cwnd_quota;
1442
1443 tcp_init_tso_segs(sk, skb, cur_mss);
1444
1445 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1446 return 0;
1447
1448 cwnd_quota = tcp_cwnd_test(tp, skb);
1449 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1450 cwnd_quota = 0;
1451
1452 return cwnd_quota;
1453}
1454
1455/* Test if sending is allowed right now. */
1456int tcp_may_send_now(struct sock *sk)
1457{
1458 struct tcp_sock *tp = tcp_sk(sk);
1459 struct sk_buff *skb = tcp_send_head(sk);
1460
1461 return skb &&
1462 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1463 (tcp_skb_is_last(sk, skb) ?
1464 tp->nonagle : TCP_NAGLE_PUSH));
1465}
1466
1467/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1468 * which is put after SKB on the list. It is very much like
1469 * tcp_fragment() except that it may make several kinds of assumptions
1470 * in order to speed up the splitting operation. In particular, we
1471 * know that all the data is in scatter-gather pages, and that the
1472 * packet has never been sent out before (and thus is not cloned).
1473 */
1474static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1475 unsigned int mss_now, gfp_t gfp)
1476{
1477 struct sk_buff *buff;
1478 int nlen = skb->len - len;
1479 u8 flags;
1480
1481 /* All of a TSO frame must be composed of paged data. */
1482 if (skb->len != skb->data_len)
1483 return tcp_fragment(sk, skb, len, mss_now);
1484
1485 buff = sk_stream_alloc_skb(sk, 0, gfp);
1486 if (unlikely(buff == NULL))
1487 return -ENOMEM;
1488
1489 sk->sk_wmem_queued += buff->truesize;
1490 sk_mem_charge(sk, buff->truesize);
1491 buff->truesize += nlen;
1492 skb->truesize -= nlen;
1493
1494 /* Correct the sequence numbers. */
1495 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1496 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1497 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1498
1499 /* PSH and FIN should only be set in the second packet. */
1500 flags = TCP_SKB_CB(skb)->flags;
1501 TCP_SKB_CB(skb)->flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1502 TCP_SKB_CB(buff)->flags = flags;
1503
1504 /* This packet was never sent out yet, so no SACK bits. */
1505 TCP_SKB_CB(buff)->sacked = 0;
1506
1507 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1508 skb_split(skb, buff, len);
1509
1510 /* Fix up tso_factor for both original and new SKB. */
1511 tcp_set_skb_tso_segs(sk, skb, mss_now);
1512 tcp_set_skb_tso_segs(sk, buff, mss_now);
1513
1514 /* Link BUFF into the send queue. */
1515 skb_header_release(buff);
1516 tcp_insert_write_queue_after(skb, buff, sk);
1517
1518 return 0;
1519}
1520
1521/* Try to defer sending, if possible, in order to minimize the amount
1522 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1523 *
1524 * This algorithm is from John Heffner.
1525 */
1526static int tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1527{
1528 struct tcp_sock *tp = tcp_sk(sk);
1529 const struct inet_connection_sock *icsk = inet_csk(sk);
1530 u32 send_win, cong_win, limit, in_flight;
1531 int win_divisor;
1532
1533 if (TCP_SKB_CB(skb)->flags & TCPHDR_FIN)
1534 goto send_now;
1535
1536 if (icsk->icsk_ca_state != TCP_CA_Open)
1537 goto send_now;
1538
1539 /* Defer for less than two clock ticks. */
1540 if (tp->tso_deferred &&
1541 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1542 goto send_now;
1543
1544 in_flight = tcp_packets_in_flight(tp);
1545
1546 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1547
1548 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1549
1550 /* From in_flight test above, we know that cwnd > in_flight. */
1551 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1552
1553 limit = min(send_win, cong_win);
1554
1555 /* If a full-sized TSO skb can be sent, do it. */
1556 if (limit >= sk->sk_gso_max_size)
1557 goto send_now;
1558
1559 /* Middle in queue won't get any more data, full sendable already? */
1560 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1561 goto send_now;
1562
1563 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1564 if (win_divisor) {
1565 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1566
1567 /* If at least some fraction of a window is available,
1568 * just use it.
1569 */
1570 chunk /= win_divisor;
1571 if (limit >= chunk)
1572 goto send_now;
1573 } else {
1574 /* Different approach, try not to defer past a single
1575 * ACK. Receiver should ACK every other full sized
1576 * frame, so if we have space for more than 3 frames
1577 * then send now.
1578 */
1579 if (limit > tcp_max_burst(tp) * tp->mss_cache)
1580 goto send_now;
1581 }
1582
1583 /* Ok, it looks like it is advisable to defer. */
1584 tp->tso_deferred = 1 | (jiffies << 1);
1585
1586 return 1;
1587
1588send_now:
1589 tp->tso_deferred = 0;
1590 return 0;
1591}
1592
1593/* Create a new MTU probe if we are ready.
1594 * MTU probe is regularly attempting to increase the path MTU by
1595 * deliberately sending larger packets. This discovers routing
1596 * changes resulting in larger path MTUs.
1597 *
1598 * Returns 0 if we should wait to probe (no cwnd available),
1599 * 1 if a probe was sent,
1600 * -1 otherwise
1601 */
1602static int tcp_mtu_probe(struct sock *sk)
1603{
1604 struct tcp_sock *tp = tcp_sk(sk);
1605 struct inet_connection_sock *icsk = inet_csk(sk);
1606 struct sk_buff *skb, *nskb, *next;
1607 int len;
1608 int probe_size;
1609 int size_needed;
1610 int copy;
1611 int mss_now;
1612
1613 /* Not currently probing/verifying,
1614 * not in recovery,
1615 * have enough cwnd, and
1616 * not SACKing (the variable headers throw things off) */
1617 if (!icsk->icsk_mtup.enabled ||
1618 icsk->icsk_mtup.probe_size ||
1619 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1620 tp->snd_cwnd < 11 ||
1621 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1622 return -1;
1623
1624 /* Very simple search strategy: just double the MSS. */
1625 mss_now = tcp_current_mss(sk);
1626 probe_size = 2 * tp->mss_cache;
1627 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1628 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1629 /* TODO: set timer for probe_converge_event */
1630 return -1;
1631 }
1632
1633 /* Have enough data in the send queue to probe? */
1634 if (tp->write_seq - tp->snd_nxt < size_needed)
1635 return -1;
1636
1637 if (tp->snd_wnd < size_needed)
1638 return -1;
1639 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1640 return 0;
1641
1642 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1643 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1644 if (!tcp_packets_in_flight(tp))
1645 return -1;
1646 else
1647 return 0;
1648 }
1649
1650 /* We're allowed to probe. Build it now. */
1651 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1652 return -1;
1653 sk->sk_wmem_queued += nskb->truesize;
1654 sk_mem_charge(sk, nskb->truesize);
1655
1656 skb = tcp_send_head(sk);
1657
1658 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1659 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1660 TCP_SKB_CB(nskb)->flags = TCPHDR_ACK;
1661 TCP_SKB_CB(nskb)->sacked = 0;
1662 nskb->csum = 0;
1663 nskb->ip_summed = skb->ip_summed;
1664
1665 tcp_insert_write_queue_before(nskb, skb, sk);
1666
1667 len = 0;
1668 tcp_for_write_queue_from_safe(skb, next, sk) {
1669 copy = min_t(int, skb->len, probe_size - len);
1670 if (nskb->ip_summed)
1671 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1672 else
1673 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1674 skb_put(nskb, copy),
1675 copy, nskb->csum);
1676
1677 if (skb->len <= copy) {
1678 /* We've eaten all the data from this skb.
1679 * Throw it away. */
1680 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags;
1681 tcp_unlink_write_queue(skb, sk);
1682 sk_wmem_free_skb(sk, skb);
1683 } else {
1684 TCP_SKB_CB(nskb)->flags |= TCP_SKB_CB(skb)->flags &
1685 ~(TCPHDR_FIN|TCPHDR_PSH);
1686 if (!skb_shinfo(skb)->nr_frags) {
1687 skb_pull(skb, copy);
1688 if (skb->ip_summed != CHECKSUM_PARTIAL)
1689 skb->csum = csum_partial(skb->data,
1690 skb->len, 0);
1691 } else {
1692 __pskb_trim_head(skb, copy);
1693 tcp_set_skb_tso_segs(sk, skb, mss_now);
1694 }
1695 TCP_SKB_CB(skb)->seq += copy;
1696 }
1697
1698 len += copy;
1699
1700 if (len >= probe_size)
1701 break;
1702 }
1703 tcp_init_tso_segs(sk, nskb, nskb->len);
1704
1705 /* We're ready to send. If this fails, the probe will
1706 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1707 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1708 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1709 /* Decrement cwnd here because we are sending
1710 * effectively two packets. */
1711 tp->snd_cwnd--;
1712 tcp_event_new_data_sent(sk, nskb);
1713
1714 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1715 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1716 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1717
1718 return 1;
1719 }
1720
1721 return -1;
1722}
1723
1724/* This routine writes packets to the network. It advances the
1725 * send_head. This happens as incoming acks open up the remote
1726 * window for us.
1727 *
1728 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1729 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1730 * account rare use of URG, this is not a big flaw.
1731 *
1732 * Returns 1, if no segments are in flight and we have queued segments, but
1733 * cannot send anything now because of SWS or another problem.
1734 */
1735static int tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1736 int push_one, gfp_t gfp)
1737{
1738 struct tcp_sock *tp = tcp_sk(sk);
1739 struct sk_buff *skb;
1740 unsigned int tso_segs, sent_pkts;
1741 int cwnd_quota;
1742 int result;
1743
1744 sent_pkts = 0;
1745
1746 if (!push_one) {
1747 /* Do MTU probing. */
1748 result = tcp_mtu_probe(sk);
1749 if (!result) {
1750 return 0;
1751 } else if (result > 0) {
1752 sent_pkts = 1;
1753 }
1754 }
1755
1756 while ((skb = tcp_send_head(sk))) {
1757 unsigned int limit;
1758
1759 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1760 BUG_ON(!tso_segs);
1761
1762 cwnd_quota = tcp_cwnd_test(tp, skb);
1763 if (!cwnd_quota)
1764 break;
1765
1766 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1767 break;
1768
1769 if (tso_segs == 1) {
1770 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1771 (tcp_skb_is_last(sk, skb) ?
1772 nonagle : TCP_NAGLE_PUSH))))
1773 break;
1774 } else {
1775 if (!push_one && tcp_tso_should_defer(sk, skb))
1776 break;
1777 }
1778
1779 limit = mss_now;
1780 if (tso_segs > 1 && !tcp_urg_mode(tp))
1781 limit = tcp_mss_split_point(sk, skb, mss_now,
1782 cwnd_quota);
1783
1784 if (skb->len > limit &&
1785 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
1786 break;
1787
1788 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1789
1790 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1791 break;
1792
1793 /* Advance the send_head. This one is sent out.
1794 * This call will increment packets_out.
1795 */
1796 tcp_event_new_data_sent(sk, skb);
1797
1798 tcp_minshall_update(tp, mss_now, skb);
1799 sent_pkts++;
1800
1801 if (push_one)
1802 break;
1803 }
1804
1805 if (likely(sent_pkts)) {
1806 tcp_cwnd_validate(sk);
1807 return 0;
1808 }
1809 return !tp->packets_out && tcp_send_head(sk);
1810}
1811
1812/* Push out any pending frames which were held back due to
1813 * TCP_CORK or attempt at coalescing tiny packets.
1814 * The socket must be locked by the caller.
1815 */
1816void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
1817 int nonagle)
1818{
1819 /* If we are closed, the bytes will have to remain here.
1820 * In time closedown will finish, we empty the write queue and
1821 * all will be happy.
1822 */
1823 if (unlikely(sk->sk_state == TCP_CLOSE))
1824 return;
1825
1826 if (tcp_write_xmit(sk, cur_mss, nonagle, 0, GFP_ATOMIC))
1827 tcp_check_probe_timer(sk);
1828}
1829
1830/* Send _single_ skb sitting at the send head. This function requires
1831 * true push pending frames to setup probe timer etc.
1832 */
1833void tcp_push_one(struct sock *sk, unsigned int mss_now)
1834{
1835 struct sk_buff *skb = tcp_send_head(sk);
1836
1837 BUG_ON(!skb || skb->len < mss_now);
1838
1839 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
1840}
1841
1842/* This function returns the amount that we can raise the
1843 * usable window based on the following constraints
1844 *
1845 * 1. The window can never be shrunk once it is offered (RFC 793)
1846 * 2. We limit memory per socket
1847 *
1848 * RFC 1122:
1849 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
1850 * RECV.NEXT + RCV.WIN fixed until:
1851 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
1852 *
1853 * i.e. don't raise the right edge of the window until you can raise
1854 * it at least MSS bytes.
1855 *
1856 * Unfortunately, the recommended algorithm breaks header prediction,
1857 * since header prediction assumes th->window stays fixed.
1858 *
1859 * Strictly speaking, keeping th->window fixed violates the receiver
1860 * side SWS prevention criteria. The problem is that under this rule
1861 * a stream of single byte packets will cause the right side of the
1862 * window to always advance by a single byte.
1863 *
1864 * Of course, if the sender implements sender side SWS prevention
1865 * then this will not be a problem.
1866 *
1867 * BSD seems to make the following compromise:
1868 *
1869 * If the free space is less than the 1/4 of the maximum
1870 * space available and the free space is less than 1/2 mss,
1871 * then set the window to 0.
1872 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
1873 * Otherwise, just prevent the window from shrinking
1874 * and from being larger than the largest representable value.
1875 *
1876 * This prevents incremental opening of the window in the regime
1877 * where TCP is limited by the speed of the reader side taking
1878 * data out of the TCP receive queue. It does nothing about
1879 * those cases where the window is constrained on the sender side
1880 * because the pipeline is full.
1881 *
1882 * BSD also seems to "accidentally" limit itself to windows that are a
1883 * multiple of MSS, at least until the free space gets quite small.
1884 * This would appear to be a side effect of the mbuf implementation.
1885 * Combining these two algorithms results in the observed behavior
1886 * of having a fixed window size at almost all times.
1887 *
1888 * Below we obtain similar behavior by forcing the offered window to
1889 * a multiple of the mss when it is feasible to do so.
1890 *
1891 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
1892 * Regular options like TIMESTAMP are taken into account.
1893 */
1894u32 __tcp_select_window(struct sock *sk)
1895{
1896 struct inet_connection_sock *icsk = inet_csk(sk);
1897 struct tcp_sock *tp = tcp_sk(sk);
1898 /* MSS for the peer's data. Previous versions used mss_clamp
1899 * here. I don't know if the value based on our guesses
1900 * of peer's MSS is better for the performance. It's more correct
1901 * but may be worse for the performance because of rcv_mss
1902 * fluctuations. --SAW 1998/11/1
1903 */
1904 int mss = icsk->icsk_ack.rcv_mss;
1905 int free_space = tcp_space(sk);
1906 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
1907 int window;
1908
1909 if (mss > full_space)
1910 mss = full_space;
1911
1912 if (free_space < (full_space >> 1)) {
1913 icsk->icsk_ack.quick = 0;
1914
1915 if (tcp_memory_pressure)
1916 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
1917 4U * tp->advmss);
1918
1919 if (free_space < mss)
1920 return 0;
1921 }
1922
1923 if (free_space > tp->rcv_ssthresh)
1924 free_space = tp->rcv_ssthresh;
1925
1926 /* Don't do rounding if we are using window scaling, since the
1927 * scaled window will not line up with the MSS boundary anyway.
1928 */
1929 window = tp->rcv_wnd;
1930 if (tp->rx_opt.rcv_wscale) {
1931 window = free_space;
1932
1933 /* Advertise enough space so that it won't get scaled away.
1934 * Import case: prevent zero window announcement if
1935 * 1<<rcv_wscale > mss.
1936 */
1937 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
1938 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
1939 << tp->rx_opt.rcv_wscale);
1940 } else {
1941 /* Get the largest window that is a nice multiple of mss.
1942 * Window clamp already applied above.
1943 * If our current window offering is within 1 mss of the
1944 * free space we just keep it. This prevents the divide
1945 * and multiply from happening most of the time.
1946 * We also don't do any window rounding when the free space
1947 * is too small.
1948 */
1949 if (window <= free_space - mss || window > free_space)
1950 window = (free_space / mss) * mss;
1951 else if (mss == full_space &&
1952 free_space > window + (full_space >> 1))
1953 window = free_space;
1954 }
1955
1956 return window;
1957}
1958
1959/* Collapses two adjacent SKB's during retransmission. */
1960static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
1961{
1962 struct tcp_sock *tp = tcp_sk(sk);
1963 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
1964 int skb_size, next_skb_size;
1965
1966 skb_size = skb->len;
1967 next_skb_size = next_skb->len;
1968
1969 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
1970
1971 tcp_highest_sack_combine(sk, next_skb, skb);
1972
1973 tcp_unlink_write_queue(next_skb, sk);
1974
1975 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
1976 next_skb_size);
1977
1978 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
1979 skb->ip_summed = CHECKSUM_PARTIAL;
1980
1981 if (skb->ip_summed != CHECKSUM_PARTIAL)
1982 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
1983
1984 /* Update sequence range on original skb. */
1985 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
1986
1987 /* Merge over control information. This moves PSH/FIN etc. over */
1988 TCP_SKB_CB(skb)->flags |= TCP_SKB_CB(next_skb)->flags;
1989
1990 /* All done, get rid of second SKB and account for it so
1991 * packet counting does not break.
1992 */
1993 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
1994
1995 /* changed transmit queue under us so clear hints */
1996 tcp_clear_retrans_hints_partial(tp);
1997 if (next_skb == tp->retransmit_skb_hint)
1998 tp->retransmit_skb_hint = skb;
1999
2000 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2001
2002 sk_wmem_free_skb(sk, next_skb);
2003}
2004
2005/* Check if coalescing SKBs is legal. */
2006static int tcp_can_collapse(struct sock *sk, struct sk_buff *skb)
2007{
2008 if (tcp_skb_pcount(skb) > 1)
2009 return 0;
2010 /* TODO: SACK collapsing could be used to remove this condition */
2011 if (skb_shinfo(skb)->nr_frags != 0)
2012 return 0;
2013 if (skb_cloned(skb))
2014 return 0;
2015 if (skb == tcp_send_head(sk))
2016 return 0;
2017 /* Some heurestics for collapsing over SACK'd could be invented */
2018 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2019 return 0;
2020
2021 return 1;
2022}
2023
2024/* Collapse packets in the retransmit queue to make to create
2025 * less packets on the wire. This is only done on retransmission.
2026 */
2027static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2028 int space)
2029{
2030 struct tcp_sock *tp = tcp_sk(sk);
2031 struct sk_buff *skb = to, *tmp;
2032 int first = 1;
2033
2034 if (!sysctl_tcp_retrans_collapse)
2035 return;
2036 if (TCP_SKB_CB(skb)->flags & TCPHDR_SYN)
2037 return;
2038
2039 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2040 if (!tcp_can_collapse(sk, skb))
2041 break;
2042
2043 space -= skb->len;
2044
2045 if (first) {
2046 first = 0;
2047 continue;
2048 }
2049
2050 if (space < 0)
2051 break;
2052 /* Punt if not enough space exists in the first SKB for
2053 * the data in the second
2054 */
2055 if (skb->len > skb_tailroom(to))
2056 break;
2057
2058 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2059 break;
2060
2061 tcp_collapse_retrans(sk, to);
2062 }
2063}
2064
2065/* This retransmits one SKB. Policy decisions and retransmit queue
2066 * state updates are done by the caller. Returns non-zero if an
2067 * error occurred which prevented the send.
2068 */
2069int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2070{
2071 struct tcp_sock *tp = tcp_sk(sk);
2072 struct inet_connection_sock *icsk = inet_csk(sk);
2073 unsigned int cur_mss;
2074 int err;
2075
2076 /* Inconslusive MTU probe */
2077 if (icsk->icsk_mtup.probe_size) {
2078 icsk->icsk_mtup.probe_size = 0;
2079 }
2080
2081 /* Do not sent more than we queued. 1/4 is reserved for possible
2082 * copying overhead: fragmentation, tunneling, mangling etc.
2083 */
2084 if (atomic_read(&sk->sk_wmem_alloc) >
2085 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2086 return -EAGAIN;
2087
2088 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2089 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2090 BUG();
2091 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2092 return -ENOMEM;
2093 }
2094
2095 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2096 return -EHOSTUNREACH; /* Routing failure or similar. */
2097
2098 cur_mss = tcp_current_mss(sk);
2099
2100 /* If receiver has shrunk his window, and skb is out of
2101 * new window, do not retransmit it. The exception is the
2102 * case, when window is shrunk to zero. In this case
2103 * our retransmit serves as a zero window probe.
2104 */
2105 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2106 TCP_SKB_CB(skb)->seq != tp->snd_una)
2107 return -EAGAIN;
2108
2109 if (skb->len > cur_mss) {
2110 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2111 return -ENOMEM; /* We'll try again later. */
2112 } else {
2113 int oldpcount = tcp_skb_pcount(skb);
2114
2115 if (unlikely(oldpcount > 1)) {
2116 tcp_init_tso_segs(sk, skb, cur_mss);
2117 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2118 }
2119 }
2120
2121 tcp_retrans_try_collapse(sk, skb, cur_mss);
2122
2123 /* Some Solaris stacks overoptimize and ignore the FIN on a
2124 * retransmit when old data is attached. So strip it off
2125 * since it is cheap to do so and saves bytes on the network.
2126 */
2127 if (skb->len > 0 &&
2128 (TCP_SKB_CB(skb)->flags & TCPHDR_FIN) &&
2129 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2130 if (!pskb_trim(skb, 0)) {
2131 /* Reuse, even though it does some unnecessary work */
2132 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2133 TCP_SKB_CB(skb)->flags);
2134 skb->ip_summed = CHECKSUM_NONE;
2135 }
2136 }
2137
2138 /* Make a copy, if the first transmission SKB clone we made
2139 * is still in somebody's hands, else make a clone.
2140 */
2141 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2142
2143 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2144
2145 if (err == 0) {
2146 /* Update global TCP statistics. */
2147 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2148
2149 tp->total_retrans++;
2150
2151#if FASTRETRANS_DEBUG > 0
2152 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2153 if (net_ratelimit())
2154 printk(KERN_DEBUG "retrans_out leaked.\n");
2155 }
2156#endif
2157 if (!tp->retrans_out)
2158 tp->lost_retrans_low = tp->snd_nxt;
2159 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2160 tp->retrans_out += tcp_skb_pcount(skb);
2161
2162 /* Save stamp of the first retransmit. */
2163 if (!tp->retrans_stamp)
2164 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2165
2166 tp->undo_retrans += tcp_skb_pcount(skb);
2167
2168 /* snd_nxt is stored to detect loss of retransmitted segment,
2169 * see tcp_input.c tcp_sacktag_write_queue().
2170 */
2171 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2172 }
2173 return err;
2174}
2175
2176/* Check if we forward retransmits are possible in the current
2177 * window/congestion state.
2178 */
2179static int tcp_can_forward_retransmit(struct sock *sk)
2180{
2181 const struct inet_connection_sock *icsk = inet_csk(sk);
2182 struct tcp_sock *tp = tcp_sk(sk);
2183
2184 /* Forward retransmissions are possible only during Recovery. */
2185 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2186 return 0;
2187
2188 /* No forward retransmissions in Reno are possible. */
2189 if (tcp_is_reno(tp))
2190 return 0;
2191
2192 /* Yeah, we have to make difficult choice between forward transmission
2193 * and retransmission... Both ways have their merits...
2194 *
2195 * For now we do not retransmit anything, while we have some new
2196 * segments to send. In the other cases, follow rule 3 for
2197 * NextSeg() specified in RFC3517.
2198 */
2199
2200 if (tcp_may_send_now(sk))
2201 return 0;
2202
2203 return 1;
2204}
2205
2206/* This gets called after a retransmit timeout, and the initially
2207 * retransmitted data is acknowledged. It tries to continue
2208 * resending the rest of the retransmit queue, until either
2209 * we've sent it all or the congestion window limit is reached.
2210 * If doing SACK, the first ACK which comes back for a timeout
2211 * based retransmit packet might feed us FACK information again.
2212 * If so, we use it to avoid unnecessarily retransmissions.
2213 */
2214void tcp_xmit_retransmit_queue(struct sock *sk)
2215{
2216 const struct inet_connection_sock *icsk = inet_csk(sk);
2217 struct tcp_sock *tp = tcp_sk(sk);
2218 struct sk_buff *skb;
2219 struct sk_buff *hole = NULL;
2220 u32 last_lost;
2221 int mib_idx;
2222 int fwd_rexmitting = 0;
2223
2224 if (!tp->packets_out)
2225 return;
2226
2227 if (!tp->lost_out)
2228 tp->retransmit_high = tp->snd_una;
2229
2230 if (tp->retransmit_skb_hint) {
2231 skb = tp->retransmit_skb_hint;
2232 last_lost = TCP_SKB_CB(skb)->end_seq;
2233 if (after(last_lost, tp->retransmit_high))
2234 last_lost = tp->retransmit_high;
2235 } else {
2236 skb = tcp_write_queue_head(sk);
2237 last_lost = tp->snd_una;
2238 }
2239
2240 tcp_for_write_queue_from(skb, sk) {
2241 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2242
2243 if (skb == tcp_send_head(sk))
2244 break;
2245 /* we could do better than to assign each time */
2246 if (hole == NULL)
2247 tp->retransmit_skb_hint = skb;
2248
2249 /* Assume this retransmit will generate
2250 * only one packet for congestion window
2251 * calculation purposes. This works because
2252 * tcp_retransmit_skb() will chop up the
2253 * packet to be MSS sized and all the
2254 * packet counting works out.
2255 */
2256 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2257 return;
2258
2259 if (fwd_rexmitting) {
2260begin_fwd:
2261 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2262 break;
2263 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2264
2265 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2266 tp->retransmit_high = last_lost;
2267 if (!tcp_can_forward_retransmit(sk))
2268 break;
2269 /* Backtrack if necessary to non-L'ed skb */
2270 if (hole != NULL) {
2271 skb = hole;
2272 hole = NULL;
2273 }
2274 fwd_rexmitting = 1;
2275 goto begin_fwd;
2276
2277 } else if (!(sacked & TCPCB_LOST)) {
2278 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2279 hole = skb;
2280 continue;
2281
2282 } else {
2283 last_lost = TCP_SKB_CB(skb)->end_seq;
2284 if (icsk->icsk_ca_state != TCP_CA_Loss)
2285 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2286 else
2287 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2288 }
2289
2290 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2291 continue;
2292
2293 if (tcp_retransmit_skb(sk, skb))
2294 return;
2295 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2296
2297 if (skb == tcp_write_queue_head(sk))
2298 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2299 inet_csk(sk)->icsk_rto,
2300 TCP_RTO_MAX);
2301 }
2302}
2303
2304/* Send a fin. The caller locks the socket for us. This cannot be
2305 * allowed to fail queueing a FIN frame under any circumstances.
2306 */
2307void tcp_send_fin(struct sock *sk)
2308{
2309 struct tcp_sock *tp = tcp_sk(sk);
2310 struct sk_buff *skb = tcp_write_queue_tail(sk);
2311 int mss_now;
2312
2313 /* Optimization, tack on the FIN if we have a queue of
2314 * unsent frames. But be careful about outgoing SACKS
2315 * and IP options.
2316 */
2317 mss_now = tcp_current_mss(sk);
2318
2319 if (tcp_send_head(sk) != NULL) {
2320 TCP_SKB_CB(skb)->flags |= TCPHDR_FIN;
2321 TCP_SKB_CB(skb)->end_seq++;
2322 tp->write_seq++;
2323 } else {
2324 /* Socket is locked, keep trying until memory is available. */
2325 for (;;) {
2326 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2327 sk->sk_allocation);
2328 if (skb)
2329 break;
2330 yield();
2331 }
2332
2333 /* Reserve space for headers and prepare control bits. */
2334 skb_reserve(skb, MAX_TCP_HEADER);
2335 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2336 tcp_init_nondata_skb(skb, tp->write_seq,
2337 TCPHDR_ACK | TCPHDR_FIN);
2338 tcp_queue_skb(sk, skb);
2339 }
2340 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2341}
2342
2343/* We get here when a process closes a file descriptor (either due to
2344 * an explicit close() or as a byproduct of exit()'ing) and there
2345 * was unread data in the receive queue. This behavior is recommended
2346 * by RFC 2525, section 2.17. -DaveM
2347 */
2348void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2349{
2350 struct sk_buff *skb;
2351
2352 /* NOTE: No TCP options attached and we never retransmit this. */
2353 skb = alloc_skb(MAX_TCP_HEADER, priority);
2354 if (!skb) {
2355 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2356 return;
2357 }
2358
2359 /* Reserve space for headers and prepare control bits. */
2360 skb_reserve(skb, MAX_TCP_HEADER);
2361 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2362 TCPHDR_ACK | TCPHDR_RST);
2363 /* Send it off. */
2364 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2365 if (tcp_transmit_skb(sk, skb, 0, priority))
2366 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2367
2368 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2369}
2370
2371/* Send a crossed SYN-ACK during socket establishment.
2372 * WARNING: This routine must only be called when we have already sent
2373 * a SYN packet that crossed the incoming SYN that caused this routine
2374 * to get called. If this assumption fails then the initial rcv_wnd
2375 * and rcv_wscale values will not be correct.
2376 */
2377int tcp_send_synack(struct sock *sk)
2378{
2379 struct sk_buff *skb;
2380
2381 skb = tcp_write_queue_head(sk);
2382 if (skb == NULL || !(TCP_SKB_CB(skb)->flags & TCPHDR_SYN)) {
2383 printk(KERN_DEBUG "tcp_send_synack: wrong queue state\n");
2384 return -EFAULT;
2385 }
2386 if (!(TCP_SKB_CB(skb)->flags & TCPHDR_ACK)) {
2387 if (skb_cloned(skb)) {
2388 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2389 if (nskb == NULL)
2390 return -ENOMEM;
2391 tcp_unlink_write_queue(skb, sk);
2392 skb_header_release(nskb);
2393 __tcp_add_write_queue_head(sk, nskb);
2394 sk_wmem_free_skb(sk, skb);
2395 sk->sk_wmem_queued += nskb->truesize;
2396 sk_mem_charge(sk, nskb->truesize);
2397 skb = nskb;
2398 }
2399
2400 TCP_SKB_CB(skb)->flags |= TCPHDR_ACK;
2401 TCP_ECN_send_synack(tcp_sk(sk), skb);
2402 }
2403 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2404 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2405}
2406
2407/* Prepare a SYN-ACK. */
2408struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2409 struct request_sock *req,
2410 struct request_values *rvp)
2411{
2412 struct tcp_out_options opts;
2413 struct tcp_extend_values *xvp = tcp_xv(rvp);
2414 struct inet_request_sock *ireq = inet_rsk(req);
2415 struct tcp_sock *tp = tcp_sk(sk);
2416 const struct tcp_cookie_values *cvp = tp->cookie_values;
2417 struct tcphdr *th;
2418 struct sk_buff *skb;
2419 struct tcp_md5sig_key *md5;
2420 int tcp_header_size;
2421 int mss;
2422 int s_data_desired = 0;
2423
2424 if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired)
2425 s_data_desired = cvp->s_data_desired;
2426 skb = sock_wmalloc(sk, MAX_TCP_HEADER + 15 + s_data_desired, 1, GFP_ATOMIC);
2427 if (skb == NULL)
2428 return NULL;
2429
2430 /* Reserve space for headers. */
2431 skb_reserve(skb, MAX_TCP_HEADER);
2432
2433 skb_dst_set(skb, dst_clone(dst));
2434
2435 mss = dst_metric_advmss(dst);
2436 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2437 mss = tp->rx_opt.user_mss;
2438
2439 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2440 __u8 rcv_wscale;
2441 /* Set this up on the first call only */
2442 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2443
2444 /* limit the window selection if the user enforce a smaller rx buffer */
2445 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2446 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2447 req->window_clamp = tcp_full_space(sk);
2448
2449 /* tcp_full_space because it is guaranteed to be the first packet */
2450 tcp_select_initial_window(tcp_full_space(sk),
2451 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2452 &req->rcv_wnd,
2453 &req->window_clamp,
2454 ireq->wscale_ok,
2455 &rcv_wscale,
2456 dst_metric(dst, RTAX_INITRWND));
2457 ireq->rcv_wscale = rcv_wscale;
2458 }
2459
2460 memset(&opts, 0, sizeof(opts));
2461#ifdef CONFIG_SYN_COOKIES
2462 if (unlikely(req->cookie_ts))
2463 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2464 else
2465#endif
2466 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2467 tcp_header_size = tcp_synack_options(sk, req, mss,
2468 skb, &opts, &md5, xvp)
2469 + sizeof(*th);
2470
2471 skb_push(skb, tcp_header_size);
2472 skb_reset_transport_header(skb);
2473
2474 th = tcp_hdr(skb);
2475 memset(th, 0, sizeof(struct tcphdr));
2476 th->syn = 1;
2477 th->ack = 1;
2478 TCP_ECN_make_synack(req, th);
2479 th->source = ireq->loc_port;
2480 th->dest = ireq->rmt_port;
2481 /* Setting of flags are superfluous here for callers (and ECE is
2482 * not even correctly set)
2483 */
2484 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2485 TCPHDR_SYN | TCPHDR_ACK);
2486
2487 if (OPTION_COOKIE_EXTENSION & opts.options) {
2488 if (s_data_desired) {
2489 u8 *buf = skb_put(skb, s_data_desired);
2490
2491 /* copy data directly from the listening socket. */
2492 memcpy(buf, cvp->s_data_payload, s_data_desired);
2493 TCP_SKB_CB(skb)->end_seq += s_data_desired;
2494 }
2495
2496 if (opts.hash_size > 0) {
2497 __u32 workspace[SHA_WORKSPACE_WORDS];
2498 u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS];
2499 u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1];
2500
2501 /* Secret recipe depends on the Timestamp, (future)
2502 * Sequence and Acknowledgment Numbers, Initiator
2503 * Cookie, and others handled by IP variant caller.
2504 */
2505 *tail-- ^= opts.tsval;
2506 *tail-- ^= tcp_rsk(req)->rcv_isn + 1;
2507 *tail-- ^= TCP_SKB_CB(skb)->seq + 1;
2508
2509 /* recommended */
2510 *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source);
2511 *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */
2512
2513 sha_transform((__u32 *)&xvp->cookie_bakery[0],
2514 (char *)mess,
2515 &workspace[0]);
2516 opts.hash_location =
2517 (__u8 *)&xvp->cookie_bakery[0];
2518 }
2519 }
2520
2521 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2522 th->ack_seq = htonl(tcp_rsk(req)->rcv_isn + 1);
2523
2524 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2525 th->window = htons(min(req->rcv_wnd, 65535U));
2526 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2527 th->doff = (tcp_header_size >> 2);
2528 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2529
2530#ifdef CONFIG_TCP_MD5SIG
2531 /* Okay, we have all we need - do the md5 hash if needed */
2532 if (md5) {
2533 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2534 md5, NULL, req, skb);
2535 }
2536#endif
2537
2538 return skb;
2539}
2540EXPORT_SYMBOL(tcp_make_synack);
2541
2542/* Do all connect socket setups that can be done AF independent. */
2543static void tcp_connect_init(struct sock *sk)
2544{
2545 struct dst_entry *dst = __sk_dst_get(sk);
2546 struct tcp_sock *tp = tcp_sk(sk);
2547 __u8 rcv_wscale;
2548
2549 /* We'll fix this up when we get a response from the other end.
2550 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2551 */
2552 tp->tcp_header_len = sizeof(struct tcphdr) +
2553 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2554
2555#ifdef CONFIG_TCP_MD5SIG
2556 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2557 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2558#endif
2559
2560 /* If user gave his TCP_MAXSEG, record it to clamp */
2561 if (tp->rx_opt.user_mss)
2562 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2563 tp->max_window = 0;
2564 tcp_mtup_init(sk);
2565 tcp_sync_mss(sk, dst_mtu(dst));
2566
2567 if (!tp->window_clamp)
2568 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2569 tp->advmss = dst_metric_advmss(dst);
2570 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2571 tp->advmss = tp->rx_opt.user_mss;
2572
2573 tcp_initialize_rcv_mss(sk);
2574
2575 /* limit the window selection if the user enforce a smaller rx buffer */
2576 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2577 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2578 tp->window_clamp = tcp_full_space(sk);
2579
2580 tcp_select_initial_window(tcp_full_space(sk),
2581 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2582 &tp->rcv_wnd,
2583 &tp->window_clamp,
2584 sysctl_tcp_window_scaling,
2585 &rcv_wscale,
2586 dst_metric(dst, RTAX_INITRWND));
2587
2588 tp->rx_opt.rcv_wscale = rcv_wscale;
2589 tp->rcv_ssthresh = tp->rcv_wnd;
2590
2591 sk->sk_err = 0;
2592 sock_reset_flag(sk, SOCK_DONE);
2593 tp->snd_wnd = 0;
2594 tcp_init_wl(tp, 0);
2595 tp->snd_una = tp->write_seq;
2596 tp->snd_sml = tp->write_seq;
2597 tp->snd_up = tp->write_seq;
2598 tp->rcv_nxt = 0;
2599 tp->rcv_wup = 0;
2600 tp->copied_seq = 0;
2601
2602 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2603 inet_csk(sk)->icsk_retransmits = 0;
2604 tcp_clear_retrans(tp);
2605}
2606
2607/* Build a SYN and send it off. */
2608int tcp_connect(struct sock *sk)
2609{
2610 struct tcp_sock *tp = tcp_sk(sk);
2611 struct sk_buff *buff;
2612 int err;
2613
2614 tcp_connect_init(sk);
2615
2616 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
2617 if (unlikely(buff == NULL))
2618 return -ENOBUFS;
2619
2620 /* Reserve space for headers. */
2621 skb_reserve(buff, MAX_TCP_HEADER);
2622
2623 tp->snd_nxt = tp->write_seq;
2624 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
2625 TCP_ECN_send_syn(sk, buff);
2626
2627 /* Send it off. */
2628 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2629 tp->retrans_stamp = TCP_SKB_CB(buff)->when;
2630 skb_header_release(buff);
2631 __tcp_add_write_queue_tail(sk, buff);
2632 sk->sk_wmem_queued += buff->truesize;
2633 sk_mem_charge(sk, buff->truesize);
2634 tp->packets_out += tcp_skb_pcount(buff);
2635 err = tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
2636 if (err == -ECONNREFUSED)
2637 return err;
2638
2639 /* We change tp->snd_nxt after the tcp_transmit_skb() call
2640 * in order to make this packet get counted in tcpOutSegs.
2641 */
2642 tp->snd_nxt = tp->write_seq;
2643 tp->pushed_seq = tp->write_seq;
2644 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
2645
2646 /* Timer for repeating the SYN until an answer. */
2647 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2648 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
2649 return 0;
2650}
2651EXPORT_SYMBOL(tcp_connect);
2652
2653/* Send out a delayed ack, the caller does the policy checking
2654 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
2655 * for details.
2656 */
2657void tcp_send_delayed_ack(struct sock *sk)
2658{
2659 struct inet_connection_sock *icsk = inet_csk(sk);
2660 int ato = icsk->icsk_ack.ato;
2661 unsigned long timeout;
2662
2663 if (ato > TCP_DELACK_MIN) {
2664 const struct tcp_sock *tp = tcp_sk(sk);
2665 int max_ato = HZ / 2;
2666
2667 if (icsk->icsk_ack.pingpong ||
2668 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
2669 max_ato = TCP_DELACK_MAX;
2670
2671 /* Slow path, intersegment interval is "high". */
2672
2673 /* If some rtt estimate is known, use it to bound delayed ack.
2674 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
2675 * directly.
2676 */
2677 if (tp->srtt) {
2678 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
2679
2680 if (rtt < max_ato)
2681 max_ato = rtt;
2682 }
2683
2684 ato = min(ato, max_ato);
2685 }
2686
2687 /* Stay within the limit we were given */
2688 timeout = jiffies + ato;
2689
2690 /* Use new timeout only if there wasn't a older one earlier. */
2691 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
2692 /* If delack timer was blocked or is about to expire,
2693 * send ACK now.
2694 */
2695 if (icsk->icsk_ack.blocked ||
2696 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
2697 tcp_send_ack(sk);
2698 return;
2699 }
2700
2701 if (!time_before(timeout, icsk->icsk_ack.timeout))
2702 timeout = icsk->icsk_ack.timeout;
2703 }
2704 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
2705 icsk->icsk_ack.timeout = timeout;
2706 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
2707}
2708
2709/* This routine sends an ack and also updates the window. */
2710void tcp_send_ack(struct sock *sk)
2711{
2712 struct sk_buff *buff;
2713
2714 /* If we have been reset, we may not send again. */
2715 if (sk->sk_state == TCP_CLOSE)
2716 return;
2717
2718 /* We are not putting this on the write queue, so
2719 * tcp_transmit_skb() will set the ownership to this
2720 * sock.
2721 */
2722 buff = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2723 if (buff == NULL) {
2724 inet_csk_schedule_ack(sk);
2725 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
2726 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
2727 TCP_DELACK_MAX, TCP_RTO_MAX);
2728 return;
2729 }
2730
2731 /* Reserve space for headers and prepare control bits. */
2732 skb_reserve(buff, MAX_TCP_HEADER);
2733 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
2734
2735 /* Send it off, this clears delayed acks for us. */
2736 TCP_SKB_CB(buff)->when = tcp_time_stamp;
2737 tcp_transmit_skb(sk, buff, 0, GFP_ATOMIC);
2738}
2739
2740/* This routine sends a packet with an out of date sequence
2741 * number. It assumes the other end will try to ack it.
2742 *
2743 * Question: what should we make while urgent mode?
2744 * 4.4BSD forces sending single byte of data. We cannot send
2745 * out of window data, because we have SND.NXT==SND.MAX...
2746 *
2747 * Current solution: to send TWO zero-length segments in urgent mode:
2748 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
2749 * out-of-date with SND.UNA-1 to probe window.
2750 */
2751static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
2752{
2753 struct tcp_sock *tp = tcp_sk(sk);
2754 struct sk_buff *skb;
2755
2756 /* We don't queue it, tcp_transmit_skb() sets ownership. */
2757 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2758 if (skb == NULL)
2759 return -1;
2760
2761 /* Reserve space for headers and set control bits. */
2762 skb_reserve(skb, MAX_TCP_HEADER);
2763 /* Use a previous sequence. This should cause the other
2764 * end to send an ack. Don't queue or clone SKB, just
2765 * send it.
2766 */
2767 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
2768 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2769 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
2770}
2771
2772/* Initiate keepalive or window probe from timer. */
2773int tcp_write_wakeup(struct sock *sk)
2774{
2775 struct tcp_sock *tp = tcp_sk(sk);
2776 struct sk_buff *skb;
2777
2778 if (sk->sk_state == TCP_CLOSE)
2779 return -1;
2780
2781 if ((skb = tcp_send_head(sk)) != NULL &&
2782 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
2783 int err;
2784 unsigned int mss = tcp_current_mss(sk);
2785 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2786
2787 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
2788 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
2789
2790 /* We are probing the opening of a window
2791 * but the window size is != 0
2792 * must have been a result SWS avoidance ( sender )
2793 */
2794 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
2795 skb->len > mss) {
2796 seg_size = min(seg_size, mss);
2797 TCP_SKB_CB(skb)->flags |= TCPHDR_PSH;
2798 if (tcp_fragment(sk, skb, seg_size, mss))
2799 return -1;
2800 } else if (!tcp_skb_pcount(skb))
2801 tcp_set_skb_tso_segs(sk, skb, mss);
2802
2803 TCP_SKB_CB(skb)->flags |= TCPHDR_PSH;
2804 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2805 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2806 if (!err)
2807 tcp_event_new_data_sent(sk, skb);
2808 return err;
2809 } else {
2810 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
2811 tcp_xmit_probe_skb(sk, 1);
2812 return tcp_xmit_probe_skb(sk, 0);
2813 }
2814}
2815
2816/* A window probe timeout has occurred. If window is not closed send
2817 * a partial packet else a zero probe.
2818 */
2819void tcp_send_probe0(struct sock *sk)
2820{
2821 struct inet_connection_sock *icsk = inet_csk(sk);
2822 struct tcp_sock *tp = tcp_sk(sk);
2823 int err;
2824
2825 err = tcp_write_wakeup(sk);
2826
2827 if (tp->packets_out || !tcp_send_head(sk)) {
2828 /* Cancel probe timer, if it is not required. */
2829 icsk->icsk_probes_out = 0;
2830 icsk->icsk_backoff = 0;
2831 return;
2832 }
2833
2834 if (err <= 0) {
2835 if (icsk->icsk_backoff < sysctl_tcp_retries2)
2836 icsk->icsk_backoff++;
2837 icsk->icsk_probes_out++;
2838 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2839 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
2840 TCP_RTO_MAX);
2841 } else {
2842 /* If packet was not sent due to local congestion,
2843 * do not backoff and do not remember icsk_probes_out.
2844 * Let local senders to fight for local resources.
2845 *
2846 * Use accumulated backoff yet.
2847 */
2848 if (!icsk->icsk_probes_out)
2849 icsk->icsk_probes_out = 1;
2850 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
2851 min(icsk->icsk_rto << icsk->icsk_backoff,
2852 TCP_RESOURCE_PROBE_INTERVAL),
2853 TCP_RTO_MAX);
2854 }
2855}
1/*
2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
8 * Authors: Ross Biro
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
19 */
20
21/*
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
25 * : AF independence
26 *
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
34 *
35 */
36
37#define pr_fmt(fmt) "TCP: " fmt
38
39#include <net/tcp.h>
40
41#include <linux/compiler.h>
42#include <linux/gfp.h>
43#include <linux/module.h>
44
45/* People can turn this off for buggy TCP's found in printers etc. */
46int sysctl_tcp_retrans_collapse __read_mostly = 1;
47
48/* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
50 */
51int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
52
53/* Default TSQ limit of four TSO segments */
54int sysctl_tcp_limit_output_bytes __read_mostly = 262144;
55
56/* This limits the percentage of the congestion window which we
57 * will allow a single TSO frame to consume. Building TSO frames
58 * which are too large can cause TCP streams to be bursty.
59 */
60int sysctl_tcp_tso_win_divisor __read_mostly = 3;
61
62/* By default, RFC2861 behavior. */
63int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
64
65static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
66 int push_one, gfp_t gfp);
67
68/* Account for new data that has been sent to the network. */
69static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
70{
71 struct inet_connection_sock *icsk = inet_csk(sk);
72 struct tcp_sock *tp = tcp_sk(sk);
73 unsigned int prior_packets = tp->packets_out;
74
75 tcp_advance_send_head(sk, skb);
76 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
77
78 tp->packets_out += tcp_skb_pcount(skb);
79 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
80 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
81 tcp_rearm_rto(sk);
82 }
83
84 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
85 tcp_skb_pcount(skb));
86}
87
88/* SND.NXT, if window was not shrunk.
89 * If window has been shrunk, what should we make? It is not clear at all.
90 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
91 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
92 * invalid. OK, let's make this for now:
93 */
94static inline __u32 tcp_acceptable_seq(const struct sock *sk)
95{
96 const struct tcp_sock *tp = tcp_sk(sk);
97
98 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
99 return tp->snd_nxt;
100 else
101 return tcp_wnd_end(tp);
102}
103
104/* Calculate mss to advertise in SYN segment.
105 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
106 *
107 * 1. It is independent of path mtu.
108 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
109 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
110 * attached devices, because some buggy hosts are confused by
111 * large MSS.
112 * 4. We do not make 3, we advertise MSS, calculated from first
113 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
114 * This may be overridden via information stored in routing table.
115 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
116 * probably even Jumbo".
117 */
118static __u16 tcp_advertise_mss(struct sock *sk)
119{
120 struct tcp_sock *tp = tcp_sk(sk);
121 const struct dst_entry *dst = __sk_dst_get(sk);
122 int mss = tp->advmss;
123
124 if (dst) {
125 unsigned int metric = dst_metric_advmss(dst);
126
127 if (metric < mss) {
128 mss = metric;
129 tp->advmss = mss;
130 }
131 }
132
133 return (__u16)mss;
134}
135
136/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
137 * This is the first part of cwnd validation mechanism.
138 */
139void tcp_cwnd_restart(struct sock *sk, s32 delta)
140{
141 struct tcp_sock *tp = tcp_sk(sk);
142 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
143 u32 cwnd = tp->snd_cwnd;
144
145 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
146
147 tp->snd_ssthresh = tcp_current_ssthresh(sk);
148 restart_cwnd = min(restart_cwnd, cwnd);
149
150 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
151 cwnd >>= 1;
152 tp->snd_cwnd = max(cwnd, restart_cwnd);
153 tp->snd_cwnd_stamp = tcp_time_stamp;
154 tp->snd_cwnd_used = 0;
155}
156
157/* Congestion state accounting after a packet has been sent. */
158static void tcp_event_data_sent(struct tcp_sock *tp,
159 struct sock *sk)
160{
161 struct inet_connection_sock *icsk = inet_csk(sk);
162 const u32 now = tcp_time_stamp;
163
164 if (tcp_packets_in_flight(tp) == 0)
165 tcp_ca_event(sk, CA_EVENT_TX_START);
166
167 tp->lsndtime = now;
168
169 /* If it is a reply for ato after last received
170 * packet, enter pingpong mode.
171 */
172 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
173 icsk->icsk_ack.pingpong = 1;
174}
175
176/* Account for an ACK we sent. */
177static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
178{
179 tcp_dec_quickack_mode(sk, pkts);
180 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
181}
182
183
184u32 tcp_default_init_rwnd(u32 mss)
185{
186 /* Initial receive window should be twice of TCP_INIT_CWND to
187 * enable proper sending of new unsent data during fast recovery
188 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
189 * limit when mss is larger than 1460.
190 */
191 u32 init_rwnd = TCP_INIT_CWND * 2;
192
193 if (mss > 1460)
194 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
195 return init_rwnd;
196}
197
198/* Determine a window scaling and initial window to offer.
199 * Based on the assumption that the given amount of space
200 * will be offered. Store the results in the tp structure.
201 * NOTE: for smooth operation initial space offering should
202 * be a multiple of mss if possible. We assume here that mss >= 1.
203 * This MUST be enforced by all callers.
204 */
205void tcp_select_initial_window(int __space, __u32 mss,
206 __u32 *rcv_wnd, __u32 *window_clamp,
207 int wscale_ok, __u8 *rcv_wscale,
208 __u32 init_rcv_wnd)
209{
210 unsigned int space = (__space < 0 ? 0 : __space);
211
212 /* If no clamp set the clamp to the max possible scaled window */
213 if (*window_clamp == 0)
214 (*window_clamp) = (65535 << 14);
215 space = min(*window_clamp, space);
216
217 /* Quantize space offering to a multiple of mss if possible. */
218 if (space > mss)
219 space = (space / mss) * mss;
220
221 /* NOTE: offering an initial window larger than 32767
222 * will break some buggy TCP stacks. If the admin tells us
223 * it is likely we could be speaking with such a buggy stack
224 * we will truncate our initial window offering to 32K-1
225 * unless the remote has sent us a window scaling option,
226 * which we interpret as a sign the remote TCP is not
227 * misinterpreting the window field as a signed quantity.
228 */
229 if (sysctl_tcp_workaround_signed_windows)
230 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
231 else
232 (*rcv_wnd) = space;
233
234 (*rcv_wscale) = 0;
235 if (wscale_ok) {
236 /* Set window scaling on max possible window
237 * See RFC1323 for an explanation of the limit to 14
238 */
239 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
240 space = min_t(u32, space, *window_clamp);
241 while (space > 65535 && (*rcv_wscale) < 14) {
242 space >>= 1;
243 (*rcv_wscale)++;
244 }
245 }
246
247 if (mss > (1 << *rcv_wscale)) {
248 if (!init_rcv_wnd) /* Use default unless specified otherwise */
249 init_rcv_wnd = tcp_default_init_rwnd(mss);
250 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
251 }
252
253 /* Set the clamp no higher than max representable value */
254 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
255}
256EXPORT_SYMBOL(tcp_select_initial_window);
257
258/* Chose a new window to advertise, update state in tcp_sock for the
259 * socket, and return result with RFC1323 scaling applied. The return
260 * value can be stuffed directly into th->window for an outgoing
261 * frame.
262 */
263static u16 tcp_select_window(struct sock *sk)
264{
265 struct tcp_sock *tp = tcp_sk(sk);
266 u32 old_win = tp->rcv_wnd;
267 u32 cur_win = tcp_receive_window(tp);
268 u32 new_win = __tcp_select_window(sk);
269
270 /* Never shrink the offered window */
271 if (new_win < cur_win) {
272 /* Danger Will Robinson!
273 * Don't update rcv_wup/rcv_wnd here or else
274 * we will not be able to advertise a zero
275 * window in time. --DaveM
276 *
277 * Relax Will Robinson.
278 */
279 if (new_win == 0)
280 NET_INC_STATS(sock_net(sk),
281 LINUX_MIB_TCPWANTZEROWINDOWADV);
282 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
283 }
284 tp->rcv_wnd = new_win;
285 tp->rcv_wup = tp->rcv_nxt;
286
287 /* Make sure we do not exceed the maximum possible
288 * scaled window.
289 */
290 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
291 new_win = min(new_win, MAX_TCP_WINDOW);
292 else
293 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
294
295 /* RFC1323 scaling applied */
296 new_win >>= tp->rx_opt.rcv_wscale;
297
298 /* If we advertise zero window, disable fast path. */
299 if (new_win == 0) {
300 tp->pred_flags = 0;
301 if (old_win)
302 NET_INC_STATS(sock_net(sk),
303 LINUX_MIB_TCPTOZEROWINDOWADV);
304 } else if (old_win == 0) {
305 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
306 }
307
308 return new_win;
309}
310
311/* Packet ECN state for a SYN-ACK */
312static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
313{
314 const struct tcp_sock *tp = tcp_sk(sk);
315
316 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
317 if (!(tp->ecn_flags & TCP_ECN_OK))
318 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
319 else if (tcp_ca_needs_ecn(sk))
320 INET_ECN_xmit(sk);
321}
322
323/* Packet ECN state for a SYN. */
324static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
325{
326 struct tcp_sock *tp = tcp_sk(sk);
327 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
328 tcp_ca_needs_ecn(sk);
329
330 if (!use_ecn) {
331 const struct dst_entry *dst = __sk_dst_get(sk);
332
333 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
334 use_ecn = true;
335 }
336
337 tp->ecn_flags = 0;
338
339 if (use_ecn) {
340 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
341 tp->ecn_flags = TCP_ECN_OK;
342 if (tcp_ca_needs_ecn(sk))
343 INET_ECN_xmit(sk);
344 }
345}
346
347static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
348{
349 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
350 /* tp->ecn_flags are cleared at a later point in time when
351 * SYN ACK is ultimatively being received.
352 */
353 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
354}
355
356static void
357tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
358{
359 if (inet_rsk(req)->ecn_ok)
360 th->ece = 1;
361}
362
363/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
364 * be sent.
365 */
366static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
367 int tcp_header_len)
368{
369 struct tcp_sock *tp = tcp_sk(sk);
370
371 if (tp->ecn_flags & TCP_ECN_OK) {
372 /* Not-retransmitted data segment: set ECT and inject CWR. */
373 if (skb->len != tcp_header_len &&
374 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
375 INET_ECN_xmit(sk);
376 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
377 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
378 tcp_hdr(skb)->cwr = 1;
379 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
380 }
381 } else if (!tcp_ca_needs_ecn(sk)) {
382 /* ACK or retransmitted segment: clear ECT|CE */
383 INET_ECN_dontxmit(sk);
384 }
385 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
386 tcp_hdr(skb)->ece = 1;
387 }
388}
389
390/* Constructs common control bits of non-data skb. If SYN/FIN is present,
391 * auto increment end seqno.
392 */
393static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
394{
395 skb->ip_summed = CHECKSUM_PARTIAL;
396 skb->csum = 0;
397
398 TCP_SKB_CB(skb)->tcp_flags = flags;
399 TCP_SKB_CB(skb)->sacked = 0;
400
401 tcp_skb_pcount_set(skb, 1);
402
403 TCP_SKB_CB(skb)->seq = seq;
404 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
405 seq++;
406 TCP_SKB_CB(skb)->end_seq = seq;
407}
408
409static inline bool tcp_urg_mode(const struct tcp_sock *tp)
410{
411 return tp->snd_una != tp->snd_up;
412}
413
414#define OPTION_SACK_ADVERTISE (1 << 0)
415#define OPTION_TS (1 << 1)
416#define OPTION_MD5 (1 << 2)
417#define OPTION_WSCALE (1 << 3)
418#define OPTION_FAST_OPEN_COOKIE (1 << 8)
419
420struct tcp_out_options {
421 u16 options; /* bit field of OPTION_* */
422 u16 mss; /* 0 to disable */
423 u8 ws; /* window scale, 0 to disable */
424 u8 num_sack_blocks; /* number of SACK blocks to include */
425 u8 hash_size; /* bytes in hash_location */
426 __u8 *hash_location; /* temporary pointer, overloaded */
427 __u32 tsval, tsecr; /* need to include OPTION_TS */
428 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
429};
430
431/* Write previously computed TCP options to the packet.
432 *
433 * Beware: Something in the Internet is very sensitive to the ordering of
434 * TCP options, we learned this through the hard way, so be careful here.
435 * Luckily we can at least blame others for their non-compliance but from
436 * inter-operability perspective it seems that we're somewhat stuck with
437 * the ordering which we have been using if we want to keep working with
438 * those broken things (not that it currently hurts anybody as there isn't
439 * particular reason why the ordering would need to be changed).
440 *
441 * At least SACK_PERM as the first option is known to lead to a disaster
442 * (but it may well be that other scenarios fail similarly).
443 */
444static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
445 struct tcp_out_options *opts)
446{
447 u16 options = opts->options; /* mungable copy */
448
449 if (unlikely(OPTION_MD5 & options)) {
450 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
451 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
452 /* overload cookie hash location */
453 opts->hash_location = (__u8 *)ptr;
454 ptr += 4;
455 }
456
457 if (unlikely(opts->mss)) {
458 *ptr++ = htonl((TCPOPT_MSS << 24) |
459 (TCPOLEN_MSS << 16) |
460 opts->mss);
461 }
462
463 if (likely(OPTION_TS & options)) {
464 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
465 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
466 (TCPOLEN_SACK_PERM << 16) |
467 (TCPOPT_TIMESTAMP << 8) |
468 TCPOLEN_TIMESTAMP);
469 options &= ~OPTION_SACK_ADVERTISE;
470 } else {
471 *ptr++ = htonl((TCPOPT_NOP << 24) |
472 (TCPOPT_NOP << 16) |
473 (TCPOPT_TIMESTAMP << 8) |
474 TCPOLEN_TIMESTAMP);
475 }
476 *ptr++ = htonl(opts->tsval);
477 *ptr++ = htonl(opts->tsecr);
478 }
479
480 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
481 *ptr++ = htonl((TCPOPT_NOP << 24) |
482 (TCPOPT_NOP << 16) |
483 (TCPOPT_SACK_PERM << 8) |
484 TCPOLEN_SACK_PERM);
485 }
486
487 if (unlikely(OPTION_WSCALE & options)) {
488 *ptr++ = htonl((TCPOPT_NOP << 24) |
489 (TCPOPT_WINDOW << 16) |
490 (TCPOLEN_WINDOW << 8) |
491 opts->ws);
492 }
493
494 if (unlikely(opts->num_sack_blocks)) {
495 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
496 tp->duplicate_sack : tp->selective_acks;
497 int this_sack;
498
499 *ptr++ = htonl((TCPOPT_NOP << 24) |
500 (TCPOPT_NOP << 16) |
501 (TCPOPT_SACK << 8) |
502 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
503 TCPOLEN_SACK_PERBLOCK)));
504
505 for (this_sack = 0; this_sack < opts->num_sack_blocks;
506 ++this_sack) {
507 *ptr++ = htonl(sp[this_sack].start_seq);
508 *ptr++ = htonl(sp[this_sack].end_seq);
509 }
510
511 tp->rx_opt.dsack = 0;
512 }
513
514 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
515 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
516 u8 *p = (u8 *)ptr;
517 u32 len; /* Fast Open option length */
518
519 if (foc->exp) {
520 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
521 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
522 TCPOPT_FASTOPEN_MAGIC);
523 p += TCPOLEN_EXP_FASTOPEN_BASE;
524 } else {
525 len = TCPOLEN_FASTOPEN_BASE + foc->len;
526 *p++ = TCPOPT_FASTOPEN;
527 *p++ = len;
528 }
529
530 memcpy(p, foc->val, foc->len);
531 if ((len & 3) == 2) {
532 p[foc->len] = TCPOPT_NOP;
533 p[foc->len + 1] = TCPOPT_NOP;
534 }
535 ptr += (len + 3) >> 2;
536 }
537}
538
539/* Compute TCP options for SYN packets. This is not the final
540 * network wire format yet.
541 */
542static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
543 struct tcp_out_options *opts,
544 struct tcp_md5sig_key **md5)
545{
546 struct tcp_sock *tp = tcp_sk(sk);
547 unsigned int remaining = MAX_TCP_OPTION_SPACE;
548 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
549
550#ifdef CONFIG_TCP_MD5SIG
551 *md5 = tp->af_specific->md5_lookup(sk, sk);
552 if (*md5) {
553 opts->options |= OPTION_MD5;
554 remaining -= TCPOLEN_MD5SIG_ALIGNED;
555 }
556#else
557 *md5 = NULL;
558#endif
559
560 /* We always get an MSS option. The option bytes which will be seen in
561 * normal data packets should timestamps be used, must be in the MSS
562 * advertised. But we subtract them from tp->mss_cache so that
563 * calculations in tcp_sendmsg are simpler etc. So account for this
564 * fact here if necessary. If we don't do this correctly, as a
565 * receiver we won't recognize data packets as being full sized when we
566 * should, and thus we won't abide by the delayed ACK rules correctly.
567 * SACKs don't matter, we never delay an ACK when we have any of those
568 * going out. */
569 opts->mss = tcp_advertise_mss(sk);
570 remaining -= TCPOLEN_MSS_ALIGNED;
571
572 if (likely(sysctl_tcp_timestamps && !*md5)) {
573 opts->options |= OPTION_TS;
574 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
575 opts->tsecr = tp->rx_opt.ts_recent;
576 remaining -= TCPOLEN_TSTAMP_ALIGNED;
577 }
578 if (likely(sysctl_tcp_window_scaling)) {
579 opts->ws = tp->rx_opt.rcv_wscale;
580 opts->options |= OPTION_WSCALE;
581 remaining -= TCPOLEN_WSCALE_ALIGNED;
582 }
583 if (likely(sysctl_tcp_sack)) {
584 opts->options |= OPTION_SACK_ADVERTISE;
585 if (unlikely(!(OPTION_TS & opts->options)))
586 remaining -= TCPOLEN_SACKPERM_ALIGNED;
587 }
588
589 if (fastopen && fastopen->cookie.len >= 0) {
590 u32 need = fastopen->cookie.len;
591
592 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
593 TCPOLEN_FASTOPEN_BASE;
594 need = (need + 3) & ~3U; /* Align to 32 bits */
595 if (remaining >= need) {
596 opts->options |= OPTION_FAST_OPEN_COOKIE;
597 opts->fastopen_cookie = &fastopen->cookie;
598 remaining -= need;
599 tp->syn_fastopen = 1;
600 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
601 }
602 }
603
604 return MAX_TCP_OPTION_SPACE - remaining;
605}
606
607/* Set up TCP options for SYN-ACKs. */
608static unsigned int tcp_synack_options(struct request_sock *req,
609 unsigned int mss, struct sk_buff *skb,
610 struct tcp_out_options *opts,
611 const struct tcp_md5sig_key *md5,
612 struct tcp_fastopen_cookie *foc)
613{
614 struct inet_request_sock *ireq = inet_rsk(req);
615 unsigned int remaining = MAX_TCP_OPTION_SPACE;
616
617#ifdef CONFIG_TCP_MD5SIG
618 if (md5) {
619 opts->options |= OPTION_MD5;
620 remaining -= TCPOLEN_MD5SIG_ALIGNED;
621
622 /* We can't fit any SACK blocks in a packet with MD5 + TS
623 * options. There was discussion about disabling SACK
624 * rather than TS in order to fit in better with old,
625 * buggy kernels, but that was deemed to be unnecessary.
626 */
627 ireq->tstamp_ok &= !ireq->sack_ok;
628 }
629#endif
630
631 /* We always send an MSS option. */
632 opts->mss = mss;
633 remaining -= TCPOLEN_MSS_ALIGNED;
634
635 if (likely(ireq->wscale_ok)) {
636 opts->ws = ireq->rcv_wscale;
637 opts->options |= OPTION_WSCALE;
638 remaining -= TCPOLEN_WSCALE_ALIGNED;
639 }
640 if (likely(ireq->tstamp_ok)) {
641 opts->options |= OPTION_TS;
642 opts->tsval = tcp_skb_timestamp(skb);
643 opts->tsecr = req->ts_recent;
644 remaining -= TCPOLEN_TSTAMP_ALIGNED;
645 }
646 if (likely(ireq->sack_ok)) {
647 opts->options |= OPTION_SACK_ADVERTISE;
648 if (unlikely(!ireq->tstamp_ok))
649 remaining -= TCPOLEN_SACKPERM_ALIGNED;
650 }
651 if (foc != NULL && foc->len >= 0) {
652 u32 need = foc->len;
653
654 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
655 TCPOLEN_FASTOPEN_BASE;
656 need = (need + 3) & ~3U; /* Align to 32 bits */
657 if (remaining >= need) {
658 opts->options |= OPTION_FAST_OPEN_COOKIE;
659 opts->fastopen_cookie = foc;
660 remaining -= need;
661 }
662 }
663
664 return MAX_TCP_OPTION_SPACE - remaining;
665}
666
667/* Compute TCP options for ESTABLISHED sockets. This is not the
668 * final wire format yet.
669 */
670static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
671 struct tcp_out_options *opts,
672 struct tcp_md5sig_key **md5)
673{
674 struct tcp_sock *tp = tcp_sk(sk);
675 unsigned int size = 0;
676 unsigned int eff_sacks;
677
678 opts->options = 0;
679
680#ifdef CONFIG_TCP_MD5SIG
681 *md5 = tp->af_specific->md5_lookup(sk, sk);
682 if (unlikely(*md5)) {
683 opts->options |= OPTION_MD5;
684 size += TCPOLEN_MD5SIG_ALIGNED;
685 }
686#else
687 *md5 = NULL;
688#endif
689
690 if (likely(tp->rx_opt.tstamp_ok)) {
691 opts->options |= OPTION_TS;
692 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
693 opts->tsecr = tp->rx_opt.ts_recent;
694 size += TCPOLEN_TSTAMP_ALIGNED;
695 }
696
697 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
698 if (unlikely(eff_sacks)) {
699 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
700 opts->num_sack_blocks =
701 min_t(unsigned int, eff_sacks,
702 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
703 TCPOLEN_SACK_PERBLOCK);
704 size += TCPOLEN_SACK_BASE_ALIGNED +
705 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
706 }
707
708 return size;
709}
710
711
712/* TCP SMALL QUEUES (TSQ)
713 *
714 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
715 * to reduce RTT and bufferbloat.
716 * We do this using a special skb destructor (tcp_wfree).
717 *
718 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
719 * needs to be reallocated in a driver.
720 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
721 *
722 * Since transmit from skb destructor is forbidden, we use a tasklet
723 * to process all sockets that eventually need to send more skbs.
724 * We use one tasklet per cpu, with its own queue of sockets.
725 */
726struct tsq_tasklet {
727 struct tasklet_struct tasklet;
728 struct list_head head; /* queue of tcp sockets */
729};
730static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
731
732static void tcp_tsq_handler(struct sock *sk)
733{
734 if ((1 << sk->sk_state) &
735 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
736 TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
737 tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle,
738 0, GFP_ATOMIC);
739}
740/*
741 * One tasklet per cpu tries to send more skbs.
742 * We run in tasklet context but need to disable irqs when
743 * transferring tsq->head because tcp_wfree() might
744 * interrupt us (non NAPI drivers)
745 */
746static void tcp_tasklet_func(unsigned long data)
747{
748 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
749 LIST_HEAD(list);
750 unsigned long flags;
751 struct list_head *q, *n;
752 struct tcp_sock *tp;
753 struct sock *sk;
754
755 local_irq_save(flags);
756 list_splice_init(&tsq->head, &list);
757 local_irq_restore(flags);
758
759 list_for_each_safe(q, n, &list) {
760 tp = list_entry(q, struct tcp_sock, tsq_node);
761 list_del(&tp->tsq_node);
762
763 sk = (struct sock *)tp;
764 bh_lock_sock(sk);
765
766 if (!sock_owned_by_user(sk)) {
767 tcp_tsq_handler(sk);
768 } else {
769 /* defer the work to tcp_release_cb() */
770 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
771 }
772 bh_unlock_sock(sk);
773
774 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
775 sk_free(sk);
776 }
777}
778
779#define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
780 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
781 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
782 (1UL << TCP_MTU_REDUCED_DEFERRED))
783/**
784 * tcp_release_cb - tcp release_sock() callback
785 * @sk: socket
786 *
787 * called from release_sock() to perform protocol dependent
788 * actions before socket release.
789 */
790void tcp_release_cb(struct sock *sk)
791{
792 struct tcp_sock *tp = tcp_sk(sk);
793 unsigned long flags, nflags;
794
795 /* perform an atomic operation only if at least one flag is set */
796 do {
797 flags = tp->tsq_flags;
798 if (!(flags & TCP_DEFERRED_ALL))
799 return;
800 nflags = flags & ~TCP_DEFERRED_ALL;
801 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
802
803 if (flags & (1UL << TCP_TSQ_DEFERRED))
804 tcp_tsq_handler(sk);
805
806 /* Here begins the tricky part :
807 * We are called from release_sock() with :
808 * 1) BH disabled
809 * 2) sk_lock.slock spinlock held
810 * 3) socket owned by us (sk->sk_lock.owned == 1)
811 *
812 * But following code is meant to be called from BH handlers,
813 * so we should keep BH disabled, but early release socket ownership
814 */
815 sock_release_ownership(sk);
816
817 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
818 tcp_write_timer_handler(sk);
819 __sock_put(sk);
820 }
821 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
822 tcp_delack_timer_handler(sk);
823 __sock_put(sk);
824 }
825 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
826 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
827 __sock_put(sk);
828 }
829}
830EXPORT_SYMBOL(tcp_release_cb);
831
832void __init tcp_tasklet_init(void)
833{
834 int i;
835
836 for_each_possible_cpu(i) {
837 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
838
839 INIT_LIST_HEAD(&tsq->head);
840 tasklet_init(&tsq->tasklet,
841 tcp_tasklet_func,
842 (unsigned long)tsq);
843 }
844}
845
846/*
847 * Write buffer destructor automatically called from kfree_skb.
848 * We can't xmit new skbs from this context, as we might already
849 * hold qdisc lock.
850 */
851void tcp_wfree(struct sk_buff *skb)
852{
853 struct sock *sk = skb->sk;
854 struct tcp_sock *tp = tcp_sk(sk);
855 int wmem;
856
857 /* Keep one reference on sk_wmem_alloc.
858 * Will be released by sk_free() from here or tcp_tasklet_func()
859 */
860 wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc);
861
862 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
863 * Wait until our queues (qdisc + devices) are drained.
864 * This gives :
865 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
866 * - chance for incoming ACK (processed by another cpu maybe)
867 * to migrate this flow (skb->ooo_okay will be eventually set)
868 */
869 if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
870 goto out;
871
872 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
873 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
874 unsigned long flags;
875 struct tsq_tasklet *tsq;
876
877 /* queue this socket to tasklet queue */
878 local_irq_save(flags);
879 tsq = this_cpu_ptr(&tsq_tasklet);
880 list_add(&tp->tsq_node, &tsq->head);
881 tasklet_schedule(&tsq->tasklet);
882 local_irq_restore(flags);
883 return;
884 }
885out:
886 sk_free(sk);
887}
888
889/* This routine actually transmits TCP packets queued in by
890 * tcp_do_sendmsg(). This is used by both the initial
891 * transmission and possible later retransmissions.
892 * All SKB's seen here are completely headerless. It is our
893 * job to build the TCP header, and pass the packet down to
894 * IP so it can do the same plus pass the packet off to the
895 * device.
896 *
897 * We are working here with either a clone of the original
898 * SKB, or a fresh unique copy made by the retransmit engine.
899 */
900static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
901 gfp_t gfp_mask)
902{
903 const struct inet_connection_sock *icsk = inet_csk(sk);
904 struct inet_sock *inet;
905 struct tcp_sock *tp;
906 struct tcp_skb_cb *tcb;
907 struct tcp_out_options opts;
908 unsigned int tcp_options_size, tcp_header_size;
909 struct tcp_md5sig_key *md5;
910 struct tcphdr *th;
911 int err;
912
913 BUG_ON(!skb || !tcp_skb_pcount(skb));
914
915 if (clone_it) {
916 skb_mstamp_get(&skb->skb_mstamp);
917
918 if (unlikely(skb_cloned(skb)))
919 skb = pskb_copy(skb, gfp_mask);
920 else
921 skb = skb_clone(skb, gfp_mask);
922 if (unlikely(!skb))
923 return -ENOBUFS;
924 }
925
926 inet = inet_sk(sk);
927 tp = tcp_sk(sk);
928 tcb = TCP_SKB_CB(skb);
929 memset(&opts, 0, sizeof(opts));
930
931 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
932 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
933 else
934 tcp_options_size = tcp_established_options(sk, skb, &opts,
935 &md5);
936 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
937
938 /* if no packet is in qdisc/device queue, then allow XPS to select
939 * another queue. We can be called from tcp_tsq_handler()
940 * which holds one reference to sk_wmem_alloc.
941 *
942 * TODO: Ideally, in-flight pure ACK packets should not matter here.
943 * One way to get this would be to set skb->truesize = 2 on them.
944 */
945 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
946
947 skb_push(skb, tcp_header_size);
948 skb_reset_transport_header(skb);
949
950 skb_orphan(skb);
951 skb->sk = sk;
952 skb->destructor = skb_is_tcp_pure_ack(skb) ? sock_wfree : tcp_wfree;
953 skb_set_hash_from_sk(skb, sk);
954 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
955
956 /* Build TCP header and checksum it. */
957 th = tcp_hdr(skb);
958 th->source = inet->inet_sport;
959 th->dest = inet->inet_dport;
960 th->seq = htonl(tcb->seq);
961 th->ack_seq = htonl(tp->rcv_nxt);
962 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
963 tcb->tcp_flags);
964
965 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
966 /* RFC1323: The window in SYN & SYN/ACK segments
967 * is never scaled.
968 */
969 th->window = htons(min(tp->rcv_wnd, 65535U));
970 } else {
971 th->window = htons(tcp_select_window(sk));
972 }
973 th->check = 0;
974 th->urg_ptr = 0;
975
976 /* The urg_mode check is necessary during a below snd_una win probe */
977 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
978 if (before(tp->snd_up, tcb->seq + 0x10000)) {
979 th->urg_ptr = htons(tp->snd_up - tcb->seq);
980 th->urg = 1;
981 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
982 th->urg_ptr = htons(0xFFFF);
983 th->urg = 1;
984 }
985 }
986
987 tcp_options_write((__be32 *)(th + 1), tp, &opts);
988 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
989 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
990 tcp_ecn_send(sk, skb, tcp_header_size);
991
992#ifdef CONFIG_TCP_MD5SIG
993 /* Calculate the MD5 hash, as we have all we need now */
994 if (md5) {
995 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
996 tp->af_specific->calc_md5_hash(opts.hash_location,
997 md5, sk, skb);
998 }
999#endif
1000
1001 icsk->icsk_af_ops->send_check(sk, skb);
1002
1003 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1004 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1005
1006 if (skb->len != tcp_header_size) {
1007 tcp_event_data_sent(tp, sk);
1008 tp->data_segs_out += tcp_skb_pcount(skb);
1009 }
1010
1011 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1012 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1013 tcp_skb_pcount(skb));
1014
1015 tp->segs_out += tcp_skb_pcount(skb);
1016 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1017 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1018 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1019
1020 /* Our usage of tstamp should remain private */
1021 skb->tstamp.tv64 = 0;
1022
1023 /* Cleanup our debris for IP stacks */
1024 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1025 sizeof(struct inet6_skb_parm)));
1026
1027 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1028
1029 if (likely(err <= 0))
1030 return err;
1031
1032 tcp_enter_cwr(sk);
1033
1034 return net_xmit_eval(err);
1035}
1036
1037/* This routine just queues the buffer for sending.
1038 *
1039 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1040 * otherwise socket can stall.
1041 */
1042static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1043{
1044 struct tcp_sock *tp = tcp_sk(sk);
1045
1046 /* Advance write_seq and place onto the write_queue. */
1047 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1048 __skb_header_release(skb);
1049 tcp_add_write_queue_tail(sk, skb);
1050 sk->sk_wmem_queued += skb->truesize;
1051 sk_mem_charge(sk, skb->truesize);
1052}
1053
1054/* Initialize TSO segments for a packet. */
1055static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1056{
1057 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1058 /* Avoid the costly divide in the normal
1059 * non-TSO case.
1060 */
1061 tcp_skb_pcount_set(skb, 1);
1062 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1063 } else {
1064 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1065 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1066 }
1067}
1068
1069/* When a modification to fackets out becomes necessary, we need to check
1070 * skb is counted to fackets_out or not.
1071 */
1072static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1073 int decr)
1074{
1075 struct tcp_sock *tp = tcp_sk(sk);
1076
1077 if (!tp->sacked_out || tcp_is_reno(tp))
1078 return;
1079
1080 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1081 tp->fackets_out -= decr;
1082}
1083
1084/* Pcount in the middle of the write queue got changed, we need to do various
1085 * tweaks to fix counters
1086 */
1087static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1088{
1089 struct tcp_sock *tp = tcp_sk(sk);
1090
1091 tp->packets_out -= decr;
1092
1093 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1094 tp->sacked_out -= decr;
1095 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1096 tp->retrans_out -= decr;
1097 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1098 tp->lost_out -= decr;
1099
1100 /* Reno case is special. Sigh... */
1101 if (tcp_is_reno(tp) && decr > 0)
1102 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1103
1104 tcp_adjust_fackets_out(sk, skb, decr);
1105
1106 if (tp->lost_skb_hint &&
1107 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1108 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1109 tp->lost_cnt_hint -= decr;
1110
1111 tcp_verify_left_out(tp);
1112}
1113
1114static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1115{
1116 struct skb_shared_info *shinfo = skb_shinfo(skb);
1117
1118 if (unlikely(shinfo->tx_flags & SKBTX_ANY_TSTAMP) &&
1119 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1120 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1121 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1122
1123 shinfo->tx_flags &= ~tsflags;
1124 shinfo2->tx_flags |= tsflags;
1125 swap(shinfo->tskey, shinfo2->tskey);
1126 }
1127}
1128
1129/* Function to create two new TCP segments. Shrinks the given segment
1130 * to the specified size and appends a new segment with the rest of the
1131 * packet to the list. This won't be called frequently, I hope.
1132 * Remember, these are still headerless SKBs at this point.
1133 */
1134int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1135 unsigned int mss_now, gfp_t gfp)
1136{
1137 struct tcp_sock *tp = tcp_sk(sk);
1138 struct sk_buff *buff;
1139 int nsize, old_factor;
1140 int nlen;
1141 u8 flags;
1142
1143 if (WARN_ON(len > skb->len))
1144 return -EINVAL;
1145
1146 nsize = skb_headlen(skb) - len;
1147 if (nsize < 0)
1148 nsize = 0;
1149
1150 if (skb_unclone(skb, gfp))
1151 return -ENOMEM;
1152
1153 /* Get a new skb... force flag on. */
1154 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1155 if (!buff)
1156 return -ENOMEM; /* We'll just try again later. */
1157
1158 sk->sk_wmem_queued += buff->truesize;
1159 sk_mem_charge(sk, buff->truesize);
1160 nlen = skb->len - len - nsize;
1161 buff->truesize += nlen;
1162 skb->truesize -= nlen;
1163
1164 /* Correct the sequence numbers. */
1165 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1166 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1167 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1168
1169 /* PSH and FIN should only be set in the second packet. */
1170 flags = TCP_SKB_CB(skb)->tcp_flags;
1171 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1172 TCP_SKB_CB(buff)->tcp_flags = flags;
1173 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1174
1175 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1176 /* Copy and checksum data tail into the new buffer. */
1177 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1178 skb_put(buff, nsize),
1179 nsize, 0);
1180
1181 skb_trim(skb, len);
1182
1183 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1184 } else {
1185 skb->ip_summed = CHECKSUM_PARTIAL;
1186 skb_split(skb, buff, len);
1187 }
1188
1189 buff->ip_summed = skb->ip_summed;
1190
1191 buff->tstamp = skb->tstamp;
1192 tcp_fragment_tstamp(skb, buff);
1193
1194 old_factor = tcp_skb_pcount(skb);
1195
1196 /* Fix up tso_factor for both original and new SKB. */
1197 tcp_set_skb_tso_segs(skb, mss_now);
1198 tcp_set_skb_tso_segs(buff, mss_now);
1199
1200 /* If this packet has been sent out already, we must
1201 * adjust the various packet counters.
1202 */
1203 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1204 int diff = old_factor - tcp_skb_pcount(skb) -
1205 tcp_skb_pcount(buff);
1206
1207 if (diff)
1208 tcp_adjust_pcount(sk, skb, diff);
1209 }
1210
1211 /* Link BUFF into the send queue. */
1212 __skb_header_release(buff);
1213 tcp_insert_write_queue_after(skb, buff, sk);
1214
1215 return 0;
1216}
1217
1218/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1219 * eventually). The difference is that pulled data not copied, but
1220 * immediately discarded.
1221 */
1222static void __pskb_trim_head(struct sk_buff *skb, int len)
1223{
1224 struct skb_shared_info *shinfo;
1225 int i, k, eat;
1226
1227 eat = min_t(int, len, skb_headlen(skb));
1228 if (eat) {
1229 __skb_pull(skb, eat);
1230 len -= eat;
1231 if (!len)
1232 return;
1233 }
1234 eat = len;
1235 k = 0;
1236 shinfo = skb_shinfo(skb);
1237 for (i = 0; i < shinfo->nr_frags; i++) {
1238 int size = skb_frag_size(&shinfo->frags[i]);
1239
1240 if (size <= eat) {
1241 skb_frag_unref(skb, i);
1242 eat -= size;
1243 } else {
1244 shinfo->frags[k] = shinfo->frags[i];
1245 if (eat) {
1246 shinfo->frags[k].page_offset += eat;
1247 skb_frag_size_sub(&shinfo->frags[k], eat);
1248 eat = 0;
1249 }
1250 k++;
1251 }
1252 }
1253 shinfo->nr_frags = k;
1254
1255 skb_reset_tail_pointer(skb);
1256 skb->data_len -= len;
1257 skb->len = skb->data_len;
1258}
1259
1260/* Remove acked data from a packet in the transmit queue. */
1261int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1262{
1263 if (skb_unclone(skb, GFP_ATOMIC))
1264 return -ENOMEM;
1265
1266 __pskb_trim_head(skb, len);
1267
1268 TCP_SKB_CB(skb)->seq += len;
1269 skb->ip_summed = CHECKSUM_PARTIAL;
1270
1271 skb->truesize -= len;
1272 sk->sk_wmem_queued -= len;
1273 sk_mem_uncharge(sk, len);
1274 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1275
1276 /* Any change of skb->len requires recalculation of tso factor. */
1277 if (tcp_skb_pcount(skb) > 1)
1278 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1279
1280 return 0;
1281}
1282
1283/* Calculate MSS not accounting any TCP options. */
1284static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1285{
1286 const struct tcp_sock *tp = tcp_sk(sk);
1287 const struct inet_connection_sock *icsk = inet_csk(sk);
1288 int mss_now;
1289
1290 /* Calculate base mss without TCP options:
1291 It is MMS_S - sizeof(tcphdr) of rfc1122
1292 */
1293 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1294
1295 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1296 if (icsk->icsk_af_ops->net_frag_header_len) {
1297 const struct dst_entry *dst = __sk_dst_get(sk);
1298
1299 if (dst && dst_allfrag(dst))
1300 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1301 }
1302
1303 /* Clamp it (mss_clamp does not include tcp options) */
1304 if (mss_now > tp->rx_opt.mss_clamp)
1305 mss_now = tp->rx_opt.mss_clamp;
1306
1307 /* Now subtract optional transport overhead */
1308 mss_now -= icsk->icsk_ext_hdr_len;
1309
1310 /* Then reserve room for full set of TCP options and 8 bytes of data */
1311 if (mss_now < 48)
1312 mss_now = 48;
1313 return mss_now;
1314}
1315
1316/* Calculate MSS. Not accounting for SACKs here. */
1317int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1318{
1319 /* Subtract TCP options size, not including SACKs */
1320 return __tcp_mtu_to_mss(sk, pmtu) -
1321 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1322}
1323
1324/* Inverse of above */
1325int tcp_mss_to_mtu(struct sock *sk, int mss)
1326{
1327 const struct tcp_sock *tp = tcp_sk(sk);
1328 const struct inet_connection_sock *icsk = inet_csk(sk);
1329 int mtu;
1330
1331 mtu = mss +
1332 tp->tcp_header_len +
1333 icsk->icsk_ext_hdr_len +
1334 icsk->icsk_af_ops->net_header_len;
1335
1336 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1337 if (icsk->icsk_af_ops->net_frag_header_len) {
1338 const struct dst_entry *dst = __sk_dst_get(sk);
1339
1340 if (dst && dst_allfrag(dst))
1341 mtu += icsk->icsk_af_ops->net_frag_header_len;
1342 }
1343 return mtu;
1344}
1345
1346/* MTU probing init per socket */
1347void tcp_mtup_init(struct sock *sk)
1348{
1349 struct tcp_sock *tp = tcp_sk(sk);
1350 struct inet_connection_sock *icsk = inet_csk(sk);
1351 struct net *net = sock_net(sk);
1352
1353 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1354 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1355 icsk->icsk_af_ops->net_header_len;
1356 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1357 icsk->icsk_mtup.probe_size = 0;
1358 if (icsk->icsk_mtup.enabled)
1359 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1360}
1361EXPORT_SYMBOL(tcp_mtup_init);
1362
1363/* This function synchronize snd mss to current pmtu/exthdr set.
1364
1365 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1366 for TCP options, but includes only bare TCP header.
1367
1368 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1369 It is minimum of user_mss and mss received with SYN.
1370 It also does not include TCP options.
1371
1372 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1373
1374 tp->mss_cache is current effective sending mss, including
1375 all tcp options except for SACKs. It is evaluated,
1376 taking into account current pmtu, but never exceeds
1377 tp->rx_opt.mss_clamp.
1378
1379 NOTE1. rfc1122 clearly states that advertised MSS
1380 DOES NOT include either tcp or ip options.
1381
1382 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1383 are READ ONLY outside this function. --ANK (980731)
1384 */
1385unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1386{
1387 struct tcp_sock *tp = tcp_sk(sk);
1388 struct inet_connection_sock *icsk = inet_csk(sk);
1389 int mss_now;
1390
1391 if (icsk->icsk_mtup.search_high > pmtu)
1392 icsk->icsk_mtup.search_high = pmtu;
1393
1394 mss_now = tcp_mtu_to_mss(sk, pmtu);
1395 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1396
1397 /* And store cached results */
1398 icsk->icsk_pmtu_cookie = pmtu;
1399 if (icsk->icsk_mtup.enabled)
1400 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1401 tp->mss_cache = mss_now;
1402
1403 return mss_now;
1404}
1405EXPORT_SYMBOL(tcp_sync_mss);
1406
1407/* Compute the current effective MSS, taking SACKs and IP options,
1408 * and even PMTU discovery events into account.
1409 */
1410unsigned int tcp_current_mss(struct sock *sk)
1411{
1412 const struct tcp_sock *tp = tcp_sk(sk);
1413 const struct dst_entry *dst = __sk_dst_get(sk);
1414 u32 mss_now;
1415 unsigned int header_len;
1416 struct tcp_out_options opts;
1417 struct tcp_md5sig_key *md5;
1418
1419 mss_now = tp->mss_cache;
1420
1421 if (dst) {
1422 u32 mtu = dst_mtu(dst);
1423 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1424 mss_now = tcp_sync_mss(sk, mtu);
1425 }
1426
1427 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1428 sizeof(struct tcphdr);
1429 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1430 * some common options. If this is an odd packet (because we have SACK
1431 * blocks etc) then our calculated header_len will be different, and
1432 * we have to adjust mss_now correspondingly */
1433 if (header_len != tp->tcp_header_len) {
1434 int delta = (int) header_len - tp->tcp_header_len;
1435 mss_now -= delta;
1436 }
1437
1438 return mss_now;
1439}
1440
1441/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1442 * As additional protections, we do not touch cwnd in retransmission phases,
1443 * and if application hit its sndbuf limit recently.
1444 */
1445static void tcp_cwnd_application_limited(struct sock *sk)
1446{
1447 struct tcp_sock *tp = tcp_sk(sk);
1448
1449 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1450 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1451 /* Limited by application or receiver window. */
1452 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1453 u32 win_used = max(tp->snd_cwnd_used, init_win);
1454 if (win_used < tp->snd_cwnd) {
1455 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1456 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1457 }
1458 tp->snd_cwnd_used = 0;
1459 }
1460 tp->snd_cwnd_stamp = tcp_time_stamp;
1461}
1462
1463static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1464{
1465 struct tcp_sock *tp = tcp_sk(sk);
1466
1467 /* Track the maximum number of outstanding packets in each
1468 * window, and remember whether we were cwnd-limited then.
1469 */
1470 if (!before(tp->snd_una, tp->max_packets_seq) ||
1471 tp->packets_out > tp->max_packets_out) {
1472 tp->max_packets_out = tp->packets_out;
1473 tp->max_packets_seq = tp->snd_nxt;
1474 tp->is_cwnd_limited = is_cwnd_limited;
1475 }
1476
1477 if (tcp_is_cwnd_limited(sk)) {
1478 /* Network is feed fully. */
1479 tp->snd_cwnd_used = 0;
1480 tp->snd_cwnd_stamp = tcp_time_stamp;
1481 } else {
1482 /* Network starves. */
1483 if (tp->packets_out > tp->snd_cwnd_used)
1484 tp->snd_cwnd_used = tp->packets_out;
1485
1486 if (sysctl_tcp_slow_start_after_idle &&
1487 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1488 tcp_cwnd_application_limited(sk);
1489 }
1490}
1491
1492/* Minshall's variant of the Nagle send check. */
1493static bool tcp_minshall_check(const struct tcp_sock *tp)
1494{
1495 return after(tp->snd_sml, tp->snd_una) &&
1496 !after(tp->snd_sml, tp->snd_nxt);
1497}
1498
1499/* Update snd_sml if this skb is under mss
1500 * Note that a TSO packet might end with a sub-mss segment
1501 * The test is really :
1502 * if ((skb->len % mss) != 0)
1503 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1504 * But we can avoid doing the divide again given we already have
1505 * skb_pcount = skb->len / mss_now
1506 */
1507static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1508 const struct sk_buff *skb)
1509{
1510 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1511 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1512}
1513
1514/* Return false, if packet can be sent now without violation Nagle's rules:
1515 * 1. It is full sized. (provided by caller in %partial bool)
1516 * 2. Or it contains FIN. (already checked by caller)
1517 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1518 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1519 * With Minshall's modification: all sent small packets are ACKed.
1520 */
1521static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1522 int nonagle)
1523{
1524 return partial &&
1525 ((nonagle & TCP_NAGLE_CORK) ||
1526 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1527}
1528
1529/* Return how many segs we'd like on a TSO packet,
1530 * to send one TSO packet per ms
1531 */
1532static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now)
1533{
1534 u32 bytes, segs;
1535
1536 bytes = min(sk->sk_pacing_rate >> 10,
1537 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1538
1539 /* Goal is to send at least one packet per ms,
1540 * not one big TSO packet every 100 ms.
1541 * This preserves ACK clocking and is consistent
1542 * with tcp_tso_should_defer() heuristic.
1543 */
1544 segs = max_t(u32, bytes / mss_now, sysctl_tcp_min_tso_segs);
1545
1546 return min_t(u32, segs, sk->sk_gso_max_segs);
1547}
1548
1549/* Returns the portion of skb which can be sent right away */
1550static unsigned int tcp_mss_split_point(const struct sock *sk,
1551 const struct sk_buff *skb,
1552 unsigned int mss_now,
1553 unsigned int max_segs,
1554 int nonagle)
1555{
1556 const struct tcp_sock *tp = tcp_sk(sk);
1557 u32 partial, needed, window, max_len;
1558
1559 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1560 max_len = mss_now * max_segs;
1561
1562 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1563 return max_len;
1564
1565 needed = min(skb->len, window);
1566
1567 if (max_len <= needed)
1568 return max_len;
1569
1570 partial = needed % mss_now;
1571 /* If last segment is not a full MSS, check if Nagle rules allow us
1572 * to include this last segment in this skb.
1573 * Otherwise, we'll split the skb at last MSS boundary
1574 */
1575 if (tcp_nagle_check(partial != 0, tp, nonagle))
1576 return needed - partial;
1577
1578 return needed;
1579}
1580
1581/* Can at least one segment of SKB be sent right now, according to the
1582 * congestion window rules? If so, return how many segments are allowed.
1583 */
1584static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1585 const struct sk_buff *skb)
1586{
1587 u32 in_flight, cwnd, halfcwnd;
1588
1589 /* Don't be strict about the congestion window for the final FIN. */
1590 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1591 tcp_skb_pcount(skb) == 1)
1592 return 1;
1593
1594 in_flight = tcp_packets_in_flight(tp);
1595 cwnd = tp->snd_cwnd;
1596 if (in_flight >= cwnd)
1597 return 0;
1598
1599 /* For better scheduling, ensure we have at least
1600 * 2 GSO packets in flight.
1601 */
1602 halfcwnd = max(cwnd >> 1, 1U);
1603 return min(halfcwnd, cwnd - in_flight);
1604}
1605
1606/* Initialize TSO state of a skb.
1607 * This must be invoked the first time we consider transmitting
1608 * SKB onto the wire.
1609 */
1610static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1611{
1612 int tso_segs = tcp_skb_pcount(skb);
1613
1614 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1615 tcp_set_skb_tso_segs(skb, mss_now);
1616 tso_segs = tcp_skb_pcount(skb);
1617 }
1618 return tso_segs;
1619}
1620
1621
1622/* Return true if the Nagle test allows this packet to be
1623 * sent now.
1624 */
1625static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1626 unsigned int cur_mss, int nonagle)
1627{
1628 /* Nagle rule does not apply to frames, which sit in the middle of the
1629 * write_queue (they have no chances to get new data).
1630 *
1631 * This is implemented in the callers, where they modify the 'nonagle'
1632 * argument based upon the location of SKB in the send queue.
1633 */
1634 if (nonagle & TCP_NAGLE_PUSH)
1635 return true;
1636
1637 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1638 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1639 return true;
1640
1641 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1642 return true;
1643
1644 return false;
1645}
1646
1647/* Does at least the first segment of SKB fit into the send window? */
1648static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1649 const struct sk_buff *skb,
1650 unsigned int cur_mss)
1651{
1652 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1653
1654 if (skb->len > cur_mss)
1655 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1656
1657 return !after(end_seq, tcp_wnd_end(tp));
1658}
1659
1660/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1661 * should be put on the wire right now. If so, it returns the number of
1662 * packets allowed by the congestion window.
1663 */
1664static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1665 unsigned int cur_mss, int nonagle)
1666{
1667 const struct tcp_sock *tp = tcp_sk(sk);
1668 unsigned int cwnd_quota;
1669
1670 tcp_init_tso_segs(skb, cur_mss);
1671
1672 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1673 return 0;
1674
1675 cwnd_quota = tcp_cwnd_test(tp, skb);
1676 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1677 cwnd_quota = 0;
1678
1679 return cwnd_quota;
1680}
1681
1682/* Test if sending is allowed right now. */
1683bool tcp_may_send_now(struct sock *sk)
1684{
1685 const struct tcp_sock *tp = tcp_sk(sk);
1686 struct sk_buff *skb = tcp_send_head(sk);
1687
1688 return skb &&
1689 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1690 (tcp_skb_is_last(sk, skb) ?
1691 tp->nonagle : TCP_NAGLE_PUSH));
1692}
1693
1694/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1695 * which is put after SKB on the list. It is very much like
1696 * tcp_fragment() except that it may make several kinds of assumptions
1697 * in order to speed up the splitting operation. In particular, we
1698 * know that all the data is in scatter-gather pages, and that the
1699 * packet has never been sent out before (and thus is not cloned).
1700 */
1701static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1702 unsigned int mss_now, gfp_t gfp)
1703{
1704 struct sk_buff *buff;
1705 int nlen = skb->len - len;
1706 u8 flags;
1707
1708 /* All of a TSO frame must be composed of paged data. */
1709 if (skb->len != skb->data_len)
1710 return tcp_fragment(sk, skb, len, mss_now, gfp);
1711
1712 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1713 if (unlikely(!buff))
1714 return -ENOMEM;
1715
1716 sk->sk_wmem_queued += buff->truesize;
1717 sk_mem_charge(sk, buff->truesize);
1718 buff->truesize += nlen;
1719 skb->truesize -= nlen;
1720
1721 /* Correct the sequence numbers. */
1722 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1723 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1724 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1725
1726 /* PSH and FIN should only be set in the second packet. */
1727 flags = TCP_SKB_CB(skb)->tcp_flags;
1728 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1729 TCP_SKB_CB(buff)->tcp_flags = flags;
1730
1731 /* This packet was never sent out yet, so no SACK bits. */
1732 TCP_SKB_CB(buff)->sacked = 0;
1733
1734 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1735 skb_split(skb, buff, len);
1736 tcp_fragment_tstamp(skb, buff);
1737
1738 /* Fix up tso_factor for both original and new SKB. */
1739 tcp_set_skb_tso_segs(skb, mss_now);
1740 tcp_set_skb_tso_segs(buff, mss_now);
1741
1742 /* Link BUFF into the send queue. */
1743 __skb_header_release(buff);
1744 tcp_insert_write_queue_after(skb, buff, sk);
1745
1746 return 0;
1747}
1748
1749/* Try to defer sending, if possible, in order to minimize the amount
1750 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1751 *
1752 * This algorithm is from John Heffner.
1753 */
1754static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1755 bool *is_cwnd_limited, u32 max_segs)
1756{
1757 const struct inet_connection_sock *icsk = inet_csk(sk);
1758 u32 age, send_win, cong_win, limit, in_flight;
1759 struct tcp_sock *tp = tcp_sk(sk);
1760 struct skb_mstamp now;
1761 struct sk_buff *head;
1762 int win_divisor;
1763
1764 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1765 goto send_now;
1766
1767 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1768 goto send_now;
1769
1770 /* Avoid bursty behavior by allowing defer
1771 * only if the last write was recent.
1772 */
1773 if ((s32)(tcp_time_stamp - tp->lsndtime) > 0)
1774 goto send_now;
1775
1776 in_flight = tcp_packets_in_flight(tp);
1777
1778 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1779
1780 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1781
1782 /* From in_flight test above, we know that cwnd > in_flight. */
1783 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1784
1785 limit = min(send_win, cong_win);
1786
1787 /* If a full-sized TSO skb can be sent, do it. */
1788 if (limit >= max_segs * tp->mss_cache)
1789 goto send_now;
1790
1791 /* Middle in queue won't get any more data, full sendable already? */
1792 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1793 goto send_now;
1794
1795 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1796 if (win_divisor) {
1797 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1798
1799 /* If at least some fraction of a window is available,
1800 * just use it.
1801 */
1802 chunk /= win_divisor;
1803 if (limit >= chunk)
1804 goto send_now;
1805 } else {
1806 /* Different approach, try not to defer past a single
1807 * ACK. Receiver should ACK every other full sized
1808 * frame, so if we have space for more than 3 frames
1809 * then send now.
1810 */
1811 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1812 goto send_now;
1813 }
1814
1815 head = tcp_write_queue_head(sk);
1816 skb_mstamp_get(&now);
1817 age = skb_mstamp_us_delta(&now, &head->skb_mstamp);
1818 /* If next ACK is likely to come too late (half srtt), do not defer */
1819 if (age < (tp->srtt_us >> 4))
1820 goto send_now;
1821
1822 /* Ok, it looks like it is advisable to defer. */
1823
1824 if (cong_win < send_win && cong_win <= skb->len)
1825 *is_cwnd_limited = true;
1826
1827 return true;
1828
1829send_now:
1830 return false;
1831}
1832
1833static inline void tcp_mtu_check_reprobe(struct sock *sk)
1834{
1835 struct inet_connection_sock *icsk = inet_csk(sk);
1836 struct tcp_sock *tp = tcp_sk(sk);
1837 struct net *net = sock_net(sk);
1838 u32 interval;
1839 s32 delta;
1840
1841 interval = net->ipv4.sysctl_tcp_probe_interval;
1842 delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp;
1843 if (unlikely(delta >= interval * HZ)) {
1844 int mss = tcp_current_mss(sk);
1845
1846 /* Update current search range */
1847 icsk->icsk_mtup.probe_size = 0;
1848 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1849 sizeof(struct tcphdr) +
1850 icsk->icsk_af_ops->net_header_len;
1851 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
1852
1853 /* Update probe time stamp */
1854 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1855 }
1856}
1857
1858/* Create a new MTU probe if we are ready.
1859 * MTU probe is regularly attempting to increase the path MTU by
1860 * deliberately sending larger packets. This discovers routing
1861 * changes resulting in larger path MTUs.
1862 *
1863 * Returns 0 if we should wait to probe (no cwnd available),
1864 * 1 if a probe was sent,
1865 * -1 otherwise
1866 */
1867static int tcp_mtu_probe(struct sock *sk)
1868{
1869 struct tcp_sock *tp = tcp_sk(sk);
1870 struct inet_connection_sock *icsk = inet_csk(sk);
1871 struct sk_buff *skb, *nskb, *next;
1872 struct net *net = sock_net(sk);
1873 int len;
1874 int probe_size;
1875 int size_needed;
1876 int copy;
1877 int mss_now;
1878 int interval;
1879
1880 /* Not currently probing/verifying,
1881 * not in recovery,
1882 * have enough cwnd, and
1883 * not SACKing (the variable headers throw things off) */
1884 if (!icsk->icsk_mtup.enabled ||
1885 icsk->icsk_mtup.probe_size ||
1886 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1887 tp->snd_cwnd < 11 ||
1888 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1889 return -1;
1890
1891 /* Use binary search for probe_size between tcp_mss_base,
1892 * and current mss_clamp. if (search_high - search_low)
1893 * smaller than a threshold, backoff from probing.
1894 */
1895 mss_now = tcp_current_mss(sk);
1896 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
1897 icsk->icsk_mtup.search_low) >> 1);
1898 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1899 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
1900 /* When misfortune happens, we are reprobing actively,
1901 * and then reprobe timer has expired. We stick with current
1902 * probing process by not resetting search range to its orignal.
1903 */
1904 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
1905 interval < net->ipv4.sysctl_tcp_probe_threshold) {
1906 /* Check whether enough time has elaplased for
1907 * another round of probing.
1908 */
1909 tcp_mtu_check_reprobe(sk);
1910 return -1;
1911 }
1912
1913 /* Have enough data in the send queue to probe? */
1914 if (tp->write_seq - tp->snd_nxt < size_needed)
1915 return -1;
1916
1917 if (tp->snd_wnd < size_needed)
1918 return -1;
1919 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1920 return 0;
1921
1922 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1923 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1924 if (!tcp_packets_in_flight(tp))
1925 return -1;
1926 else
1927 return 0;
1928 }
1929
1930 /* We're allowed to probe. Build it now. */
1931 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
1932 if (!nskb)
1933 return -1;
1934 sk->sk_wmem_queued += nskb->truesize;
1935 sk_mem_charge(sk, nskb->truesize);
1936
1937 skb = tcp_send_head(sk);
1938
1939 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1940 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1941 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1942 TCP_SKB_CB(nskb)->sacked = 0;
1943 nskb->csum = 0;
1944 nskb->ip_summed = skb->ip_summed;
1945
1946 tcp_insert_write_queue_before(nskb, skb, sk);
1947
1948 len = 0;
1949 tcp_for_write_queue_from_safe(skb, next, sk) {
1950 copy = min_t(int, skb->len, probe_size - len);
1951 if (nskb->ip_summed)
1952 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1953 else
1954 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1955 skb_put(nskb, copy),
1956 copy, nskb->csum);
1957
1958 if (skb->len <= copy) {
1959 /* We've eaten all the data from this skb.
1960 * Throw it away. */
1961 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1962 tcp_unlink_write_queue(skb, sk);
1963 sk_wmem_free_skb(sk, skb);
1964 } else {
1965 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1966 ~(TCPHDR_FIN|TCPHDR_PSH);
1967 if (!skb_shinfo(skb)->nr_frags) {
1968 skb_pull(skb, copy);
1969 if (skb->ip_summed != CHECKSUM_PARTIAL)
1970 skb->csum = csum_partial(skb->data,
1971 skb->len, 0);
1972 } else {
1973 __pskb_trim_head(skb, copy);
1974 tcp_set_skb_tso_segs(skb, mss_now);
1975 }
1976 TCP_SKB_CB(skb)->seq += copy;
1977 }
1978
1979 len += copy;
1980
1981 if (len >= probe_size)
1982 break;
1983 }
1984 tcp_init_tso_segs(nskb, nskb->len);
1985
1986 /* We're ready to send. If this fails, the probe will
1987 * be resegmented into mss-sized pieces by tcp_write_xmit().
1988 */
1989 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1990 /* Decrement cwnd here because we are sending
1991 * effectively two packets. */
1992 tp->snd_cwnd--;
1993 tcp_event_new_data_sent(sk, nskb);
1994
1995 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1996 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1997 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1998
1999 return 1;
2000 }
2001
2002 return -1;
2003}
2004
2005/* This routine writes packets to the network. It advances the
2006 * send_head. This happens as incoming acks open up the remote
2007 * window for us.
2008 *
2009 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2010 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2011 * account rare use of URG, this is not a big flaw.
2012 *
2013 * Send at most one packet when push_one > 0. Temporarily ignore
2014 * cwnd limit to force at most one packet out when push_one == 2.
2015
2016 * Returns true, if no segments are in flight and we have queued segments,
2017 * but cannot send anything now because of SWS or another problem.
2018 */
2019static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2020 int push_one, gfp_t gfp)
2021{
2022 struct tcp_sock *tp = tcp_sk(sk);
2023 struct sk_buff *skb;
2024 unsigned int tso_segs, sent_pkts;
2025 int cwnd_quota;
2026 int result;
2027 bool is_cwnd_limited = false;
2028 u32 max_segs;
2029
2030 sent_pkts = 0;
2031
2032 if (!push_one) {
2033 /* Do MTU probing. */
2034 result = tcp_mtu_probe(sk);
2035 if (!result) {
2036 return false;
2037 } else if (result > 0) {
2038 sent_pkts = 1;
2039 }
2040 }
2041
2042 max_segs = tcp_tso_autosize(sk, mss_now);
2043 while ((skb = tcp_send_head(sk))) {
2044 unsigned int limit;
2045
2046 tso_segs = tcp_init_tso_segs(skb, mss_now);
2047 BUG_ON(!tso_segs);
2048
2049 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2050 /* "skb_mstamp" is used as a start point for the retransmit timer */
2051 skb_mstamp_get(&skb->skb_mstamp);
2052 goto repair; /* Skip network transmission */
2053 }
2054
2055 cwnd_quota = tcp_cwnd_test(tp, skb);
2056 if (!cwnd_quota) {
2057 if (push_one == 2)
2058 /* Force out a loss probe pkt. */
2059 cwnd_quota = 1;
2060 else
2061 break;
2062 }
2063
2064 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
2065 break;
2066
2067 if (tso_segs == 1) {
2068 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2069 (tcp_skb_is_last(sk, skb) ?
2070 nonagle : TCP_NAGLE_PUSH))))
2071 break;
2072 } else {
2073 if (!push_one &&
2074 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2075 max_segs))
2076 break;
2077 }
2078
2079 limit = mss_now;
2080 if (tso_segs > 1 && !tcp_urg_mode(tp))
2081 limit = tcp_mss_split_point(sk, skb, mss_now,
2082 min_t(unsigned int,
2083 cwnd_quota,
2084 max_segs),
2085 nonagle);
2086
2087 if (skb->len > limit &&
2088 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2089 break;
2090
2091 /* TCP Small Queues :
2092 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2093 * This allows for :
2094 * - better RTT estimation and ACK scheduling
2095 * - faster recovery
2096 * - high rates
2097 * Alas, some drivers / subsystems require a fair amount
2098 * of queued bytes to ensure line rate.
2099 * One example is wifi aggregation (802.11 AMPDU)
2100 */
2101 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2102 limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2103
2104 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
2105 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
2106 /* It is possible TX completion already happened
2107 * before we set TSQ_THROTTLED, so we must
2108 * test again the condition.
2109 */
2110 smp_mb__after_atomic();
2111 if (atomic_read(&sk->sk_wmem_alloc) > limit)
2112 break;
2113 }
2114
2115 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2116 break;
2117
2118repair:
2119 /* Advance the send_head. This one is sent out.
2120 * This call will increment packets_out.
2121 */
2122 tcp_event_new_data_sent(sk, skb);
2123
2124 tcp_minshall_update(tp, mss_now, skb);
2125 sent_pkts += tcp_skb_pcount(skb);
2126
2127 if (push_one)
2128 break;
2129 }
2130
2131 if (likely(sent_pkts)) {
2132 if (tcp_in_cwnd_reduction(sk))
2133 tp->prr_out += sent_pkts;
2134
2135 /* Send one loss probe per tail loss episode. */
2136 if (push_one != 2)
2137 tcp_schedule_loss_probe(sk);
2138 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2139 tcp_cwnd_validate(sk, is_cwnd_limited);
2140 return false;
2141 }
2142 return !tp->packets_out && tcp_send_head(sk);
2143}
2144
2145bool tcp_schedule_loss_probe(struct sock *sk)
2146{
2147 struct inet_connection_sock *icsk = inet_csk(sk);
2148 struct tcp_sock *tp = tcp_sk(sk);
2149 u32 timeout, tlp_time_stamp, rto_time_stamp;
2150 u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
2151
2152 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2153 return false;
2154 /* No consecutive loss probes. */
2155 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2156 tcp_rearm_rto(sk);
2157 return false;
2158 }
2159 /* Don't do any loss probe on a Fast Open connection before 3WHS
2160 * finishes.
2161 */
2162 if (tp->fastopen_rsk)
2163 return false;
2164
2165 /* TLP is only scheduled when next timer event is RTO. */
2166 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2167 return false;
2168
2169 /* Schedule a loss probe in 2*RTT for SACK capable connections
2170 * in Open state, that are either limited by cwnd or application.
2171 */
2172 if (sysctl_tcp_early_retrans < 3 || !tp->packets_out ||
2173 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2174 return false;
2175
2176 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2177 tcp_send_head(sk))
2178 return false;
2179
2180 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2181 * for delayed ack when there's one outstanding packet. If no RTT
2182 * sample is available then probe after TCP_TIMEOUT_INIT.
2183 */
2184 timeout = rtt << 1 ? : TCP_TIMEOUT_INIT;
2185 if (tp->packets_out == 1)
2186 timeout = max_t(u32, timeout,
2187 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2188 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2189
2190 /* If RTO is shorter, just schedule TLP in its place. */
2191 tlp_time_stamp = tcp_time_stamp + timeout;
2192 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2193 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2194 s32 delta = rto_time_stamp - tcp_time_stamp;
2195 if (delta > 0)
2196 timeout = delta;
2197 }
2198
2199 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2200 TCP_RTO_MAX);
2201 return true;
2202}
2203
2204/* Thanks to skb fast clones, we can detect if a prior transmit of
2205 * a packet is still in a qdisc or driver queue.
2206 * In this case, there is very little point doing a retransmit !
2207 * Note: This is called from BH context only.
2208 */
2209static bool skb_still_in_host_queue(const struct sock *sk,
2210 const struct sk_buff *skb)
2211{
2212 if (unlikely(skb_fclone_busy(sk, skb))) {
2213 NET_INC_STATS_BH(sock_net(sk),
2214 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2215 return true;
2216 }
2217 return false;
2218}
2219
2220/* When probe timeout (PTO) fires, try send a new segment if possible, else
2221 * retransmit the last segment.
2222 */
2223void tcp_send_loss_probe(struct sock *sk)
2224{
2225 struct tcp_sock *tp = tcp_sk(sk);
2226 struct sk_buff *skb;
2227 int pcount;
2228 int mss = tcp_current_mss(sk);
2229
2230 skb = tcp_send_head(sk);
2231 if (skb) {
2232 if (tcp_snd_wnd_test(tp, skb, mss)) {
2233 pcount = tp->packets_out;
2234 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2235 if (tp->packets_out > pcount)
2236 goto probe_sent;
2237 goto rearm_timer;
2238 }
2239 skb = tcp_write_queue_prev(sk, skb);
2240 } else {
2241 skb = tcp_write_queue_tail(sk);
2242 }
2243
2244 /* At most one outstanding TLP retransmission. */
2245 if (tp->tlp_high_seq)
2246 goto rearm_timer;
2247
2248 /* Retransmit last segment. */
2249 if (WARN_ON(!skb))
2250 goto rearm_timer;
2251
2252 if (skb_still_in_host_queue(sk, skb))
2253 goto rearm_timer;
2254
2255 pcount = tcp_skb_pcount(skb);
2256 if (WARN_ON(!pcount))
2257 goto rearm_timer;
2258
2259 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2260 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2261 GFP_ATOMIC)))
2262 goto rearm_timer;
2263 skb = tcp_write_queue_next(sk, skb);
2264 }
2265
2266 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2267 goto rearm_timer;
2268
2269 if (__tcp_retransmit_skb(sk, skb))
2270 goto rearm_timer;
2271
2272 /* Record snd_nxt for loss detection. */
2273 tp->tlp_high_seq = tp->snd_nxt;
2274
2275probe_sent:
2276 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2277 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2278 inet_csk(sk)->icsk_pending = 0;
2279rearm_timer:
2280 tcp_rearm_rto(sk);
2281}
2282
2283/* Push out any pending frames which were held back due to
2284 * TCP_CORK or attempt at coalescing tiny packets.
2285 * The socket must be locked by the caller.
2286 */
2287void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2288 int nonagle)
2289{
2290 /* If we are closed, the bytes will have to remain here.
2291 * In time closedown will finish, we empty the write queue and
2292 * all will be happy.
2293 */
2294 if (unlikely(sk->sk_state == TCP_CLOSE))
2295 return;
2296
2297 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2298 sk_gfp_mask(sk, GFP_ATOMIC)))
2299 tcp_check_probe_timer(sk);
2300}
2301
2302/* Send _single_ skb sitting at the send head. This function requires
2303 * true push pending frames to setup probe timer etc.
2304 */
2305void tcp_push_one(struct sock *sk, unsigned int mss_now)
2306{
2307 struct sk_buff *skb = tcp_send_head(sk);
2308
2309 BUG_ON(!skb || skb->len < mss_now);
2310
2311 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2312}
2313
2314/* This function returns the amount that we can raise the
2315 * usable window based on the following constraints
2316 *
2317 * 1. The window can never be shrunk once it is offered (RFC 793)
2318 * 2. We limit memory per socket
2319 *
2320 * RFC 1122:
2321 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2322 * RECV.NEXT + RCV.WIN fixed until:
2323 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2324 *
2325 * i.e. don't raise the right edge of the window until you can raise
2326 * it at least MSS bytes.
2327 *
2328 * Unfortunately, the recommended algorithm breaks header prediction,
2329 * since header prediction assumes th->window stays fixed.
2330 *
2331 * Strictly speaking, keeping th->window fixed violates the receiver
2332 * side SWS prevention criteria. The problem is that under this rule
2333 * a stream of single byte packets will cause the right side of the
2334 * window to always advance by a single byte.
2335 *
2336 * Of course, if the sender implements sender side SWS prevention
2337 * then this will not be a problem.
2338 *
2339 * BSD seems to make the following compromise:
2340 *
2341 * If the free space is less than the 1/4 of the maximum
2342 * space available and the free space is less than 1/2 mss,
2343 * then set the window to 0.
2344 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2345 * Otherwise, just prevent the window from shrinking
2346 * and from being larger than the largest representable value.
2347 *
2348 * This prevents incremental opening of the window in the regime
2349 * where TCP is limited by the speed of the reader side taking
2350 * data out of the TCP receive queue. It does nothing about
2351 * those cases where the window is constrained on the sender side
2352 * because the pipeline is full.
2353 *
2354 * BSD also seems to "accidentally" limit itself to windows that are a
2355 * multiple of MSS, at least until the free space gets quite small.
2356 * This would appear to be a side effect of the mbuf implementation.
2357 * Combining these two algorithms results in the observed behavior
2358 * of having a fixed window size at almost all times.
2359 *
2360 * Below we obtain similar behavior by forcing the offered window to
2361 * a multiple of the mss when it is feasible to do so.
2362 *
2363 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2364 * Regular options like TIMESTAMP are taken into account.
2365 */
2366u32 __tcp_select_window(struct sock *sk)
2367{
2368 struct inet_connection_sock *icsk = inet_csk(sk);
2369 struct tcp_sock *tp = tcp_sk(sk);
2370 /* MSS for the peer's data. Previous versions used mss_clamp
2371 * here. I don't know if the value based on our guesses
2372 * of peer's MSS is better for the performance. It's more correct
2373 * but may be worse for the performance because of rcv_mss
2374 * fluctuations. --SAW 1998/11/1
2375 */
2376 int mss = icsk->icsk_ack.rcv_mss;
2377 int free_space = tcp_space(sk);
2378 int allowed_space = tcp_full_space(sk);
2379 int full_space = min_t(int, tp->window_clamp, allowed_space);
2380 int window;
2381
2382 if (mss > full_space)
2383 mss = full_space;
2384
2385 if (free_space < (full_space >> 1)) {
2386 icsk->icsk_ack.quick = 0;
2387
2388 if (tcp_under_memory_pressure(sk))
2389 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2390 4U * tp->advmss);
2391
2392 /* free_space might become our new window, make sure we don't
2393 * increase it due to wscale.
2394 */
2395 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2396
2397 /* if free space is less than mss estimate, or is below 1/16th
2398 * of the maximum allowed, try to move to zero-window, else
2399 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2400 * new incoming data is dropped due to memory limits.
2401 * With large window, mss test triggers way too late in order
2402 * to announce zero window in time before rmem limit kicks in.
2403 */
2404 if (free_space < (allowed_space >> 4) || free_space < mss)
2405 return 0;
2406 }
2407
2408 if (free_space > tp->rcv_ssthresh)
2409 free_space = tp->rcv_ssthresh;
2410
2411 /* Don't do rounding if we are using window scaling, since the
2412 * scaled window will not line up with the MSS boundary anyway.
2413 */
2414 window = tp->rcv_wnd;
2415 if (tp->rx_opt.rcv_wscale) {
2416 window = free_space;
2417
2418 /* Advertise enough space so that it won't get scaled away.
2419 * Import case: prevent zero window announcement if
2420 * 1<<rcv_wscale > mss.
2421 */
2422 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2423 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2424 << tp->rx_opt.rcv_wscale);
2425 } else {
2426 /* Get the largest window that is a nice multiple of mss.
2427 * Window clamp already applied above.
2428 * If our current window offering is within 1 mss of the
2429 * free space we just keep it. This prevents the divide
2430 * and multiply from happening most of the time.
2431 * We also don't do any window rounding when the free space
2432 * is too small.
2433 */
2434 if (window <= free_space - mss || window > free_space)
2435 window = (free_space / mss) * mss;
2436 else if (mss == full_space &&
2437 free_space > window + (full_space >> 1))
2438 window = free_space;
2439 }
2440
2441 return window;
2442}
2443
2444void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2445 const struct sk_buff *next_skb)
2446{
2447 const struct skb_shared_info *next_shinfo = skb_shinfo(next_skb);
2448 u8 tsflags = next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2449
2450 if (unlikely(tsflags)) {
2451 struct skb_shared_info *shinfo = skb_shinfo(skb);
2452
2453 shinfo->tx_flags |= tsflags;
2454 shinfo->tskey = next_shinfo->tskey;
2455 }
2456}
2457
2458/* Collapses two adjacent SKB's during retransmission. */
2459static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2460{
2461 struct tcp_sock *tp = tcp_sk(sk);
2462 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2463 int skb_size, next_skb_size;
2464
2465 skb_size = skb->len;
2466 next_skb_size = next_skb->len;
2467
2468 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2469
2470 tcp_highest_sack_combine(sk, next_skb, skb);
2471
2472 tcp_unlink_write_queue(next_skb, sk);
2473
2474 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2475 next_skb_size);
2476
2477 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2478 skb->ip_summed = CHECKSUM_PARTIAL;
2479
2480 if (skb->ip_summed != CHECKSUM_PARTIAL)
2481 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2482
2483 /* Update sequence range on original skb. */
2484 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2485
2486 /* Merge over control information. This moves PSH/FIN etc. over */
2487 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2488
2489 /* All done, get rid of second SKB and account for it so
2490 * packet counting does not break.
2491 */
2492 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2493
2494 /* changed transmit queue under us so clear hints */
2495 tcp_clear_retrans_hints_partial(tp);
2496 if (next_skb == tp->retransmit_skb_hint)
2497 tp->retransmit_skb_hint = skb;
2498
2499 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2500
2501 tcp_skb_collapse_tstamp(skb, next_skb);
2502
2503 sk_wmem_free_skb(sk, next_skb);
2504}
2505
2506/* Check if coalescing SKBs is legal. */
2507static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2508{
2509 if (tcp_skb_pcount(skb) > 1)
2510 return false;
2511 /* TODO: SACK collapsing could be used to remove this condition */
2512 if (skb_shinfo(skb)->nr_frags != 0)
2513 return false;
2514 if (skb_cloned(skb))
2515 return false;
2516 if (skb == tcp_send_head(sk))
2517 return false;
2518 /* Some heurestics for collapsing over SACK'd could be invented */
2519 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2520 return false;
2521
2522 return true;
2523}
2524
2525/* Collapse packets in the retransmit queue to make to create
2526 * less packets on the wire. This is only done on retransmission.
2527 */
2528static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2529 int space)
2530{
2531 struct tcp_sock *tp = tcp_sk(sk);
2532 struct sk_buff *skb = to, *tmp;
2533 bool first = true;
2534
2535 if (!sysctl_tcp_retrans_collapse)
2536 return;
2537 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2538 return;
2539
2540 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2541 if (!tcp_can_collapse(sk, skb))
2542 break;
2543
2544 space -= skb->len;
2545
2546 if (first) {
2547 first = false;
2548 continue;
2549 }
2550
2551 if (space < 0)
2552 break;
2553 /* Punt if not enough space exists in the first SKB for
2554 * the data in the second
2555 */
2556 if (skb->len > skb_availroom(to))
2557 break;
2558
2559 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2560 break;
2561
2562 tcp_collapse_retrans(sk, to);
2563 }
2564}
2565
2566/* This retransmits one SKB. Policy decisions and retransmit queue
2567 * state updates are done by the caller. Returns non-zero if an
2568 * error occurred which prevented the send.
2569 */
2570int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2571{
2572 struct tcp_sock *tp = tcp_sk(sk);
2573 struct inet_connection_sock *icsk = inet_csk(sk);
2574 unsigned int cur_mss;
2575 int err;
2576
2577 /* Inconslusive MTU probe */
2578 if (icsk->icsk_mtup.probe_size) {
2579 icsk->icsk_mtup.probe_size = 0;
2580 }
2581
2582 /* Do not sent more than we queued. 1/4 is reserved for possible
2583 * copying overhead: fragmentation, tunneling, mangling etc.
2584 */
2585 if (atomic_read(&sk->sk_wmem_alloc) >
2586 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2587 return -EAGAIN;
2588
2589 if (skb_still_in_host_queue(sk, skb))
2590 return -EBUSY;
2591
2592 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2593 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2594 BUG();
2595 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2596 return -ENOMEM;
2597 }
2598
2599 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2600 return -EHOSTUNREACH; /* Routing failure or similar. */
2601
2602 cur_mss = tcp_current_mss(sk);
2603
2604 /* If receiver has shrunk his window, and skb is out of
2605 * new window, do not retransmit it. The exception is the
2606 * case, when window is shrunk to zero. In this case
2607 * our retransmit serves as a zero window probe.
2608 */
2609 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2610 TCP_SKB_CB(skb)->seq != tp->snd_una)
2611 return -EAGAIN;
2612
2613 if (skb->len > cur_mss) {
2614 if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC))
2615 return -ENOMEM; /* We'll try again later. */
2616 } else {
2617 int oldpcount = tcp_skb_pcount(skb);
2618
2619 if (unlikely(oldpcount > 1)) {
2620 if (skb_unclone(skb, GFP_ATOMIC))
2621 return -ENOMEM;
2622 tcp_init_tso_segs(skb, cur_mss);
2623 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2624 }
2625 }
2626
2627 /* RFC3168, section 6.1.1.1. ECN fallback */
2628 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2629 tcp_ecn_clear_syn(sk, skb);
2630
2631 tcp_retrans_try_collapse(sk, skb, cur_mss);
2632
2633 /* Make a copy, if the first transmission SKB clone we made
2634 * is still in somebody's hands, else make a clone.
2635 */
2636
2637 /* make sure skb->data is aligned on arches that require it
2638 * and check if ack-trimming & collapsing extended the headroom
2639 * beyond what csum_start can cover.
2640 */
2641 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2642 skb_headroom(skb) >= 0xFFFF)) {
2643 struct sk_buff *nskb;
2644
2645 skb_mstamp_get(&skb->skb_mstamp);
2646 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2647 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2648 -ENOBUFS;
2649 } else {
2650 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2651 }
2652
2653 if (likely(!err)) {
2654 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2655 /* Update global TCP statistics. */
2656 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2657 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2658 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2659 tp->total_retrans++;
2660 }
2661 return err;
2662}
2663
2664int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2665{
2666 struct tcp_sock *tp = tcp_sk(sk);
2667 int err = __tcp_retransmit_skb(sk, skb);
2668
2669 if (err == 0) {
2670#if FASTRETRANS_DEBUG > 0
2671 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2672 net_dbg_ratelimited("retrans_out leaked\n");
2673 }
2674#endif
2675 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2676 tp->retrans_out += tcp_skb_pcount(skb);
2677
2678 /* Save stamp of the first retransmit. */
2679 if (!tp->retrans_stamp)
2680 tp->retrans_stamp = tcp_skb_timestamp(skb);
2681
2682 } else if (err != -EBUSY) {
2683 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2684 }
2685
2686 if (tp->undo_retrans < 0)
2687 tp->undo_retrans = 0;
2688 tp->undo_retrans += tcp_skb_pcount(skb);
2689 return err;
2690}
2691
2692/* Check if we forward retransmits are possible in the current
2693 * window/congestion state.
2694 */
2695static bool tcp_can_forward_retransmit(struct sock *sk)
2696{
2697 const struct inet_connection_sock *icsk = inet_csk(sk);
2698 const struct tcp_sock *tp = tcp_sk(sk);
2699
2700 /* Forward retransmissions are possible only during Recovery. */
2701 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2702 return false;
2703
2704 /* No forward retransmissions in Reno are possible. */
2705 if (tcp_is_reno(tp))
2706 return false;
2707
2708 /* Yeah, we have to make difficult choice between forward transmission
2709 * and retransmission... Both ways have their merits...
2710 *
2711 * For now we do not retransmit anything, while we have some new
2712 * segments to send. In the other cases, follow rule 3 for
2713 * NextSeg() specified in RFC3517.
2714 */
2715
2716 if (tcp_may_send_now(sk))
2717 return false;
2718
2719 return true;
2720}
2721
2722/* This gets called after a retransmit timeout, and the initially
2723 * retransmitted data is acknowledged. It tries to continue
2724 * resending the rest of the retransmit queue, until either
2725 * we've sent it all or the congestion window limit is reached.
2726 * If doing SACK, the first ACK which comes back for a timeout
2727 * based retransmit packet might feed us FACK information again.
2728 * If so, we use it to avoid unnecessarily retransmissions.
2729 */
2730void tcp_xmit_retransmit_queue(struct sock *sk)
2731{
2732 const struct inet_connection_sock *icsk = inet_csk(sk);
2733 struct tcp_sock *tp = tcp_sk(sk);
2734 struct sk_buff *skb;
2735 struct sk_buff *hole = NULL;
2736 u32 last_lost;
2737 int mib_idx;
2738 int fwd_rexmitting = 0;
2739
2740 if (!tp->packets_out)
2741 return;
2742
2743 if (!tp->lost_out)
2744 tp->retransmit_high = tp->snd_una;
2745
2746 if (tp->retransmit_skb_hint) {
2747 skb = tp->retransmit_skb_hint;
2748 last_lost = TCP_SKB_CB(skb)->end_seq;
2749 if (after(last_lost, tp->retransmit_high))
2750 last_lost = tp->retransmit_high;
2751 } else {
2752 skb = tcp_write_queue_head(sk);
2753 last_lost = tp->snd_una;
2754 }
2755
2756 tcp_for_write_queue_from(skb, sk) {
2757 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2758
2759 if (skb == tcp_send_head(sk))
2760 break;
2761 /* we could do better than to assign each time */
2762 if (!hole)
2763 tp->retransmit_skb_hint = skb;
2764
2765 /* Assume this retransmit will generate
2766 * only one packet for congestion window
2767 * calculation purposes. This works because
2768 * tcp_retransmit_skb() will chop up the
2769 * packet to be MSS sized and all the
2770 * packet counting works out.
2771 */
2772 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2773 return;
2774
2775 if (fwd_rexmitting) {
2776begin_fwd:
2777 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2778 break;
2779 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2780
2781 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2782 tp->retransmit_high = last_lost;
2783 if (!tcp_can_forward_retransmit(sk))
2784 break;
2785 /* Backtrack if necessary to non-L'ed skb */
2786 if (hole) {
2787 skb = hole;
2788 hole = NULL;
2789 }
2790 fwd_rexmitting = 1;
2791 goto begin_fwd;
2792
2793 } else if (!(sacked & TCPCB_LOST)) {
2794 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2795 hole = skb;
2796 continue;
2797
2798 } else {
2799 last_lost = TCP_SKB_CB(skb)->end_seq;
2800 if (icsk->icsk_ca_state != TCP_CA_Loss)
2801 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2802 else
2803 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2804 }
2805
2806 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2807 continue;
2808
2809 if (tcp_retransmit_skb(sk, skb))
2810 return;
2811
2812 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2813
2814 if (tcp_in_cwnd_reduction(sk))
2815 tp->prr_out += tcp_skb_pcount(skb);
2816
2817 if (skb == tcp_write_queue_head(sk))
2818 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2819 inet_csk(sk)->icsk_rto,
2820 TCP_RTO_MAX);
2821 }
2822}
2823
2824/* We allow to exceed memory limits for FIN packets to expedite
2825 * connection tear down and (memory) recovery.
2826 * Otherwise tcp_send_fin() could be tempted to either delay FIN
2827 * or even be forced to close flow without any FIN.
2828 * In general, we want to allow one skb per socket to avoid hangs
2829 * with edge trigger epoll()
2830 */
2831void sk_forced_mem_schedule(struct sock *sk, int size)
2832{
2833 int amt;
2834
2835 if (size <= sk->sk_forward_alloc)
2836 return;
2837 amt = sk_mem_pages(size);
2838 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
2839 sk_memory_allocated_add(sk, amt);
2840
2841 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
2842 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
2843}
2844
2845/* Send a FIN. The caller locks the socket for us.
2846 * We should try to send a FIN packet really hard, but eventually give up.
2847 */
2848void tcp_send_fin(struct sock *sk)
2849{
2850 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
2851 struct tcp_sock *tp = tcp_sk(sk);
2852
2853 /* Optimization, tack on the FIN if we have one skb in write queue and
2854 * this skb was not yet sent, or we are under memory pressure.
2855 * Note: in the latter case, FIN packet will be sent after a timeout,
2856 * as TCP stack thinks it has already been transmitted.
2857 */
2858 if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
2859coalesce:
2860 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
2861 TCP_SKB_CB(tskb)->end_seq++;
2862 tp->write_seq++;
2863 if (!tcp_send_head(sk)) {
2864 /* This means tskb was already sent.
2865 * Pretend we included the FIN on previous transmit.
2866 * We need to set tp->snd_nxt to the value it would have
2867 * if FIN had been sent. This is because retransmit path
2868 * does not change tp->snd_nxt.
2869 */
2870 tp->snd_nxt++;
2871 return;
2872 }
2873 } else {
2874 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
2875 if (unlikely(!skb)) {
2876 if (tskb)
2877 goto coalesce;
2878 return;
2879 }
2880 skb_reserve(skb, MAX_TCP_HEADER);
2881 sk_forced_mem_schedule(sk, skb->truesize);
2882 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2883 tcp_init_nondata_skb(skb, tp->write_seq,
2884 TCPHDR_ACK | TCPHDR_FIN);
2885 tcp_queue_skb(sk, skb);
2886 }
2887 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
2888}
2889
2890/* We get here when a process closes a file descriptor (either due to
2891 * an explicit close() or as a byproduct of exit()'ing) and there
2892 * was unread data in the receive queue. This behavior is recommended
2893 * by RFC 2525, section 2.17. -DaveM
2894 */
2895void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2896{
2897 struct sk_buff *skb;
2898
2899 /* NOTE: No TCP options attached and we never retransmit this. */
2900 skb = alloc_skb(MAX_TCP_HEADER, priority);
2901 if (!skb) {
2902 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2903 return;
2904 }
2905
2906 /* Reserve space for headers and prepare control bits. */
2907 skb_reserve(skb, MAX_TCP_HEADER);
2908 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2909 TCPHDR_ACK | TCPHDR_RST);
2910 skb_mstamp_get(&skb->skb_mstamp);
2911 /* Send it off. */
2912 if (tcp_transmit_skb(sk, skb, 0, priority))
2913 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2914
2915 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2916}
2917
2918/* Send a crossed SYN-ACK during socket establishment.
2919 * WARNING: This routine must only be called when we have already sent
2920 * a SYN packet that crossed the incoming SYN that caused this routine
2921 * to get called. If this assumption fails then the initial rcv_wnd
2922 * and rcv_wscale values will not be correct.
2923 */
2924int tcp_send_synack(struct sock *sk)
2925{
2926 struct sk_buff *skb;
2927
2928 skb = tcp_write_queue_head(sk);
2929 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2930 pr_debug("%s: wrong queue state\n", __func__);
2931 return -EFAULT;
2932 }
2933 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2934 if (skb_cloned(skb)) {
2935 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2936 if (!nskb)
2937 return -ENOMEM;
2938 tcp_unlink_write_queue(skb, sk);
2939 __skb_header_release(nskb);
2940 __tcp_add_write_queue_head(sk, nskb);
2941 sk_wmem_free_skb(sk, skb);
2942 sk->sk_wmem_queued += nskb->truesize;
2943 sk_mem_charge(sk, nskb->truesize);
2944 skb = nskb;
2945 }
2946
2947 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2948 tcp_ecn_send_synack(sk, skb);
2949 }
2950 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2951}
2952
2953/**
2954 * tcp_make_synack - Prepare a SYN-ACK.
2955 * sk: listener socket
2956 * dst: dst entry attached to the SYNACK
2957 * req: request_sock pointer
2958 *
2959 * Allocate one skb and build a SYNACK packet.
2960 * @dst is consumed : Caller should not use it again.
2961 */
2962struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
2963 struct request_sock *req,
2964 struct tcp_fastopen_cookie *foc,
2965 bool attach_req)
2966{
2967 struct inet_request_sock *ireq = inet_rsk(req);
2968 const struct tcp_sock *tp = tcp_sk(sk);
2969 struct tcp_md5sig_key *md5 = NULL;
2970 struct tcp_out_options opts;
2971 struct sk_buff *skb;
2972 int tcp_header_size;
2973 struct tcphdr *th;
2974 u16 user_mss;
2975 int mss;
2976
2977 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2978 if (unlikely(!skb)) {
2979 dst_release(dst);
2980 return NULL;
2981 }
2982 /* Reserve space for headers. */
2983 skb_reserve(skb, MAX_TCP_HEADER);
2984
2985 if (attach_req) {
2986 skb_set_owner_w(skb, req_to_sk(req));
2987 } else {
2988 /* sk is a const pointer, because we want to express multiple
2989 * cpu might call us concurrently.
2990 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
2991 */
2992 skb_set_owner_w(skb, (struct sock *)sk);
2993 }
2994 skb_dst_set(skb, dst);
2995
2996 mss = dst_metric_advmss(dst);
2997 user_mss = READ_ONCE(tp->rx_opt.user_mss);
2998 if (user_mss && user_mss < mss)
2999 mss = user_mss;
3000
3001 memset(&opts, 0, sizeof(opts));
3002#ifdef CONFIG_SYN_COOKIES
3003 if (unlikely(req->cookie_ts))
3004 skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
3005 else
3006#endif
3007 skb_mstamp_get(&skb->skb_mstamp);
3008
3009#ifdef CONFIG_TCP_MD5SIG
3010 rcu_read_lock();
3011 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3012#endif
3013 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3014 tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) +
3015 sizeof(*th);
3016
3017 skb_push(skb, tcp_header_size);
3018 skb_reset_transport_header(skb);
3019
3020 th = tcp_hdr(skb);
3021 memset(th, 0, sizeof(struct tcphdr));
3022 th->syn = 1;
3023 th->ack = 1;
3024 tcp_ecn_make_synack(req, th);
3025 th->source = htons(ireq->ir_num);
3026 th->dest = ireq->ir_rmt_port;
3027 /* Setting of flags are superfluous here for callers (and ECE is
3028 * not even correctly set)
3029 */
3030 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
3031 TCPHDR_SYN | TCPHDR_ACK);
3032
3033 th->seq = htonl(TCP_SKB_CB(skb)->seq);
3034 /* XXX data is queued and acked as is. No buffer/window check */
3035 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3036
3037 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3038 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3039 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3040 th->doff = (tcp_header_size >> 2);
3041 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS);
3042
3043#ifdef CONFIG_TCP_MD5SIG
3044 /* Okay, we have all we need - do the md5 hash if needed */
3045 if (md5)
3046 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3047 md5, req_to_sk(req), skb);
3048 rcu_read_unlock();
3049#endif
3050
3051 /* Do not fool tcpdump (if any), clean our debris */
3052 skb->tstamp.tv64 = 0;
3053 return skb;
3054}
3055EXPORT_SYMBOL(tcp_make_synack);
3056
3057static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3058{
3059 struct inet_connection_sock *icsk = inet_csk(sk);
3060 const struct tcp_congestion_ops *ca;
3061 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3062
3063 if (ca_key == TCP_CA_UNSPEC)
3064 return;
3065
3066 rcu_read_lock();
3067 ca = tcp_ca_find_key(ca_key);
3068 if (likely(ca && try_module_get(ca->owner))) {
3069 module_put(icsk->icsk_ca_ops->owner);
3070 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3071 icsk->icsk_ca_ops = ca;
3072 }
3073 rcu_read_unlock();
3074}
3075
3076/* Do all connect socket setups that can be done AF independent. */
3077static void tcp_connect_init(struct sock *sk)
3078{
3079 const struct dst_entry *dst = __sk_dst_get(sk);
3080 struct tcp_sock *tp = tcp_sk(sk);
3081 __u8 rcv_wscale;
3082
3083 /* We'll fix this up when we get a response from the other end.
3084 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3085 */
3086 tp->tcp_header_len = sizeof(struct tcphdr) +
3087 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
3088
3089#ifdef CONFIG_TCP_MD5SIG
3090 if (tp->af_specific->md5_lookup(sk, sk))
3091 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3092#endif
3093
3094 /* If user gave his TCP_MAXSEG, record it to clamp */
3095 if (tp->rx_opt.user_mss)
3096 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3097 tp->max_window = 0;
3098 tcp_mtup_init(sk);
3099 tcp_sync_mss(sk, dst_mtu(dst));
3100
3101 tcp_ca_dst_init(sk, dst);
3102
3103 if (!tp->window_clamp)
3104 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3105 tp->advmss = dst_metric_advmss(dst);
3106 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
3107 tp->advmss = tp->rx_opt.user_mss;
3108
3109 tcp_initialize_rcv_mss(sk);
3110
3111 /* limit the window selection if the user enforce a smaller rx buffer */
3112 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3113 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3114 tp->window_clamp = tcp_full_space(sk);
3115
3116 tcp_select_initial_window(tcp_full_space(sk),
3117 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3118 &tp->rcv_wnd,
3119 &tp->window_clamp,
3120 sysctl_tcp_window_scaling,
3121 &rcv_wscale,
3122 dst_metric(dst, RTAX_INITRWND));
3123
3124 tp->rx_opt.rcv_wscale = rcv_wscale;
3125 tp->rcv_ssthresh = tp->rcv_wnd;
3126
3127 sk->sk_err = 0;
3128 sock_reset_flag(sk, SOCK_DONE);
3129 tp->snd_wnd = 0;
3130 tcp_init_wl(tp, 0);
3131 tp->snd_una = tp->write_seq;
3132 tp->snd_sml = tp->write_seq;
3133 tp->snd_up = tp->write_seq;
3134 tp->snd_nxt = tp->write_seq;
3135
3136 if (likely(!tp->repair))
3137 tp->rcv_nxt = 0;
3138 else
3139 tp->rcv_tstamp = tcp_time_stamp;
3140 tp->rcv_wup = tp->rcv_nxt;
3141 tp->copied_seq = tp->rcv_nxt;
3142
3143 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
3144 inet_csk(sk)->icsk_retransmits = 0;
3145 tcp_clear_retrans(tp);
3146}
3147
3148static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3149{
3150 struct tcp_sock *tp = tcp_sk(sk);
3151 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3152
3153 tcb->end_seq += skb->len;
3154 __skb_header_release(skb);
3155 __tcp_add_write_queue_tail(sk, skb);
3156 sk->sk_wmem_queued += skb->truesize;
3157 sk_mem_charge(sk, skb->truesize);
3158 tp->write_seq = tcb->end_seq;
3159 tp->packets_out += tcp_skb_pcount(skb);
3160}
3161
3162/* Build and send a SYN with data and (cached) Fast Open cookie. However,
3163 * queue a data-only packet after the regular SYN, such that regular SYNs
3164 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3165 * only the SYN sequence, the data are retransmitted in the first ACK.
3166 * If cookie is not cached or other error occurs, falls back to send a
3167 * regular SYN with Fast Open cookie request option.
3168 */
3169static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3170{
3171 struct tcp_sock *tp = tcp_sk(sk);
3172 struct tcp_fastopen_request *fo = tp->fastopen_req;
3173 int syn_loss = 0, space, err = 0;
3174 unsigned long last_syn_loss = 0;
3175 struct sk_buff *syn_data;
3176
3177 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3178 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
3179 &syn_loss, &last_syn_loss);
3180 /* Recurring FO SYN losses: revert to regular handshake temporarily */
3181 if (syn_loss > 1 &&
3182 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
3183 fo->cookie.len = -1;
3184 goto fallback;
3185 }
3186
3187 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
3188 fo->cookie.len = -1;
3189 else if (fo->cookie.len <= 0)
3190 goto fallback;
3191
3192 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3193 * user-MSS. Reserve maximum option space for middleboxes that add
3194 * private TCP options. The cost is reduced data space in SYN :(
3195 */
3196 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
3197 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3198 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3199 MAX_TCP_OPTION_SPACE;
3200
3201 space = min_t(size_t, space, fo->size);
3202
3203 /* limit to order-0 allocations */
3204 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3205
3206 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3207 if (!syn_data)
3208 goto fallback;
3209 syn_data->ip_summed = CHECKSUM_PARTIAL;
3210 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3211 if (space) {
3212 int copied = copy_from_iter(skb_put(syn_data, space), space,
3213 &fo->data->msg_iter);
3214 if (unlikely(!copied)) {
3215 kfree_skb(syn_data);
3216 goto fallback;
3217 }
3218 if (copied != space) {
3219 skb_trim(syn_data, copied);
3220 space = copied;
3221 }
3222 }
3223 /* No more data pending in inet_wait_for_connect() */
3224 if (space == fo->size)
3225 fo->data = NULL;
3226 fo->copied = space;
3227
3228 tcp_connect_queue_skb(sk, syn_data);
3229
3230 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3231
3232 syn->skb_mstamp = syn_data->skb_mstamp;
3233
3234 /* Now full SYN+DATA was cloned and sent (or not),
3235 * remove the SYN from the original skb (syn_data)
3236 * we keep in write queue in case of a retransmit, as we
3237 * also have the SYN packet (with no data) in the same queue.
3238 */
3239 TCP_SKB_CB(syn_data)->seq++;
3240 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3241 if (!err) {
3242 tp->syn_data = (fo->copied > 0);
3243 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3244 goto done;
3245 }
3246
3247fallback:
3248 /* Send a regular SYN with Fast Open cookie request option */
3249 if (fo->cookie.len > 0)
3250 fo->cookie.len = 0;
3251 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3252 if (err)
3253 tp->syn_fastopen = 0;
3254done:
3255 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3256 return err;
3257}
3258
3259/* Build a SYN and send it off. */
3260int tcp_connect(struct sock *sk)
3261{
3262 struct tcp_sock *tp = tcp_sk(sk);
3263 struct sk_buff *buff;
3264 int err;
3265
3266 tcp_connect_init(sk);
3267
3268 if (unlikely(tp->repair)) {
3269 tcp_finish_connect(sk, NULL);
3270 return 0;
3271 }
3272
3273 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3274 if (unlikely(!buff))
3275 return -ENOBUFS;
3276
3277 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3278 tp->retrans_stamp = tcp_time_stamp;
3279 tcp_connect_queue_skb(sk, buff);
3280 tcp_ecn_send_syn(sk, buff);
3281
3282 /* Send off SYN; include data in Fast Open. */
3283 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3284 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3285 if (err == -ECONNREFUSED)
3286 return err;
3287
3288 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3289 * in order to make this packet get counted in tcpOutSegs.
3290 */
3291 tp->snd_nxt = tp->write_seq;
3292 tp->pushed_seq = tp->write_seq;
3293 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3294
3295 /* Timer for repeating the SYN until an answer. */
3296 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3297 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3298 return 0;
3299}
3300EXPORT_SYMBOL(tcp_connect);
3301
3302/* Send out a delayed ack, the caller does the policy checking
3303 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3304 * for details.
3305 */
3306void tcp_send_delayed_ack(struct sock *sk)
3307{
3308 struct inet_connection_sock *icsk = inet_csk(sk);
3309 int ato = icsk->icsk_ack.ato;
3310 unsigned long timeout;
3311
3312 tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3313
3314 if (ato > TCP_DELACK_MIN) {
3315 const struct tcp_sock *tp = tcp_sk(sk);
3316 int max_ato = HZ / 2;
3317
3318 if (icsk->icsk_ack.pingpong ||
3319 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3320 max_ato = TCP_DELACK_MAX;
3321
3322 /* Slow path, intersegment interval is "high". */
3323
3324 /* If some rtt estimate is known, use it to bound delayed ack.
3325 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3326 * directly.
3327 */
3328 if (tp->srtt_us) {
3329 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3330 TCP_DELACK_MIN);
3331
3332 if (rtt < max_ato)
3333 max_ato = rtt;
3334 }
3335
3336 ato = min(ato, max_ato);
3337 }
3338
3339 /* Stay within the limit we were given */
3340 timeout = jiffies + ato;
3341
3342 /* Use new timeout only if there wasn't a older one earlier. */
3343 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3344 /* If delack timer was blocked or is about to expire,
3345 * send ACK now.
3346 */
3347 if (icsk->icsk_ack.blocked ||
3348 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3349 tcp_send_ack(sk);
3350 return;
3351 }
3352
3353 if (!time_before(timeout, icsk->icsk_ack.timeout))
3354 timeout = icsk->icsk_ack.timeout;
3355 }
3356 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3357 icsk->icsk_ack.timeout = timeout;
3358 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3359}
3360
3361/* This routine sends an ack and also updates the window. */
3362void tcp_send_ack(struct sock *sk)
3363{
3364 struct sk_buff *buff;
3365
3366 /* If we have been reset, we may not send again. */
3367 if (sk->sk_state == TCP_CLOSE)
3368 return;
3369
3370 tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3371
3372 /* We are not putting this on the write queue, so
3373 * tcp_transmit_skb() will set the ownership to this
3374 * sock.
3375 */
3376 buff = alloc_skb(MAX_TCP_HEADER,
3377 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3378 if (unlikely(!buff)) {
3379 inet_csk_schedule_ack(sk);
3380 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3381 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3382 TCP_DELACK_MAX, TCP_RTO_MAX);
3383 return;
3384 }
3385
3386 /* Reserve space for headers and prepare control bits. */
3387 skb_reserve(buff, MAX_TCP_HEADER);
3388 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3389
3390 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3391 * too much.
3392 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3393 * We also avoid tcp_wfree() overhead (cache line miss accessing
3394 * tp->tsq_flags) by using regular sock_wfree()
3395 */
3396 skb_set_tcp_pure_ack(buff);
3397
3398 /* Send it off, this clears delayed acks for us. */
3399 skb_mstamp_get(&buff->skb_mstamp);
3400 tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0);
3401}
3402EXPORT_SYMBOL_GPL(tcp_send_ack);
3403
3404/* This routine sends a packet with an out of date sequence
3405 * number. It assumes the other end will try to ack it.
3406 *
3407 * Question: what should we make while urgent mode?
3408 * 4.4BSD forces sending single byte of data. We cannot send
3409 * out of window data, because we have SND.NXT==SND.MAX...
3410 *
3411 * Current solution: to send TWO zero-length segments in urgent mode:
3412 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3413 * out-of-date with SND.UNA-1 to probe window.
3414 */
3415static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3416{
3417 struct tcp_sock *tp = tcp_sk(sk);
3418 struct sk_buff *skb;
3419
3420 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3421 skb = alloc_skb(MAX_TCP_HEADER,
3422 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3423 if (!skb)
3424 return -1;
3425
3426 /* Reserve space for headers and set control bits. */
3427 skb_reserve(skb, MAX_TCP_HEADER);
3428 /* Use a previous sequence. This should cause the other
3429 * end to send an ack. Don't queue or clone SKB, just
3430 * send it.
3431 */
3432 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3433 skb_mstamp_get(&skb->skb_mstamp);
3434 NET_INC_STATS(sock_net(sk), mib);
3435 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3436}
3437
3438void tcp_send_window_probe(struct sock *sk)
3439{
3440 if (sk->sk_state == TCP_ESTABLISHED) {
3441 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3442 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3443 }
3444}
3445
3446/* Initiate keepalive or window probe from timer. */
3447int tcp_write_wakeup(struct sock *sk, int mib)
3448{
3449 struct tcp_sock *tp = tcp_sk(sk);
3450 struct sk_buff *skb;
3451
3452 if (sk->sk_state == TCP_CLOSE)
3453 return -1;
3454
3455 skb = tcp_send_head(sk);
3456 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3457 int err;
3458 unsigned int mss = tcp_current_mss(sk);
3459 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3460
3461 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3462 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3463
3464 /* We are probing the opening of a window
3465 * but the window size is != 0
3466 * must have been a result SWS avoidance ( sender )
3467 */
3468 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3469 skb->len > mss) {
3470 seg_size = min(seg_size, mss);
3471 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3472 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3473 return -1;
3474 } else if (!tcp_skb_pcount(skb))
3475 tcp_set_skb_tso_segs(skb, mss);
3476
3477 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3478 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3479 if (!err)
3480 tcp_event_new_data_sent(sk, skb);
3481 return err;
3482 } else {
3483 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3484 tcp_xmit_probe_skb(sk, 1, mib);
3485 return tcp_xmit_probe_skb(sk, 0, mib);
3486 }
3487}
3488
3489/* A window probe timeout has occurred. If window is not closed send
3490 * a partial packet else a zero probe.
3491 */
3492void tcp_send_probe0(struct sock *sk)
3493{
3494 struct inet_connection_sock *icsk = inet_csk(sk);
3495 struct tcp_sock *tp = tcp_sk(sk);
3496 struct net *net = sock_net(sk);
3497 unsigned long probe_max;
3498 int err;
3499
3500 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3501
3502 if (tp->packets_out || !tcp_send_head(sk)) {
3503 /* Cancel probe timer, if it is not required. */
3504 icsk->icsk_probes_out = 0;
3505 icsk->icsk_backoff = 0;
3506 return;
3507 }
3508
3509 if (err <= 0) {
3510 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3511 icsk->icsk_backoff++;
3512 icsk->icsk_probes_out++;
3513 probe_max = TCP_RTO_MAX;
3514 } else {
3515 /* If packet was not sent due to local congestion,
3516 * do not backoff and do not remember icsk_probes_out.
3517 * Let local senders to fight for local resources.
3518 *
3519 * Use accumulated backoff yet.
3520 */
3521 if (!icsk->icsk_probes_out)
3522 icsk->icsk_probes_out = 1;
3523 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3524 }
3525 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3526 tcp_probe0_when(sk, probe_max),
3527 TCP_RTO_MAX);
3528}
3529
3530int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3531{
3532 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3533 struct flowi fl;
3534 int res;
3535
3536 tcp_rsk(req)->txhash = net_tx_rndhash();
3537 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, true);
3538 if (!res) {
3539 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_RETRANSSEGS);
3540 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3541 }
3542 return res;
3543}
3544EXPORT_SYMBOL(tcp_rtx_synack);