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 two TSO segments */
54int sysctl_tcp_limit_output_bytes __read_mostly = 131072;
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
62int sysctl_tcp_mtu_probing __read_mostly = 0;
63int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS;
64
65/* By default, RFC2861 behavior. */
66int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
67
68unsigned int sysctl_tcp_notsent_lowat __read_mostly = UINT_MAX;
69EXPORT_SYMBOL(sysctl_tcp_notsent_lowat);
70
71static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
72 int push_one, gfp_t gfp);
73
74/* Account for new data that has been sent to the network. */
75static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
76{
77 struct inet_connection_sock *icsk = inet_csk(sk);
78 struct tcp_sock *tp = tcp_sk(sk);
79 unsigned int prior_packets = tp->packets_out;
80
81 tcp_advance_send_head(sk, skb);
82 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
83
84 tp->packets_out += tcp_skb_pcount(skb);
85 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
86 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
87 tcp_rearm_rto(sk);
88 }
89
90 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
91 tcp_skb_pcount(skb));
92}
93
94/* SND.NXT, if window was not shrunk.
95 * If window has been shrunk, what should we make? It is not clear at all.
96 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
97 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
98 * invalid. OK, let's make this for now:
99 */
100static inline __u32 tcp_acceptable_seq(const struct sock *sk)
101{
102 const struct tcp_sock *tp = tcp_sk(sk);
103
104 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
105 return tp->snd_nxt;
106 else
107 return tcp_wnd_end(tp);
108}
109
110/* Calculate mss to advertise in SYN segment.
111 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
112 *
113 * 1. It is independent of path mtu.
114 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
115 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
116 * attached devices, because some buggy hosts are confused by
117 * large MSS.
118 * 4. We do not make 3, we advertise MSS, calculated from first
119 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
120 * This may be overridden via information stored in routing table.
121 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
122 * probably even Jumbo".
123 */
124static __u16 tcp_advertise_mss(struct sock *sk)
125{
126 struct tcp_sock *tp = tcp_sk(sk);
127 const struct dst_entry *dst = __sk_dst_get(sk);
128 int mss = tp->advmss;
129
130 if (dst) {
131 unsigned int metric = dst_metric_advmss(dst);
132
133 if (metric < mss) {
134 mss = metric;
135 tp->advmss = mss;
136 }
137 }
138
139 return (__u16)mss;
140}
141
142/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
143 * This is the first part of cwnd validation mechanism. */
144static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst)
145{
146 struct tcp_sock *tp = tcp_sk(sk);
147 s32 delta = tcp_time_stamp - tp->lsndtime;
148 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
149 u32 cwnd = tp->snd_cwnd;
150
151 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
152
153 tp->snd_ssthresh = tcp_current_ssthresh(sk);
154 restart_cwnd = min(restart_cwnd, cwnd);
155
156 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
157 cwnd >>= 1;
158 tp->snd_cwnd = max(cwnd, restart_cwnd);
159 tp->snd_cwnd_stamp = tcp_time_stamp;
160 tp->snd_cwnd_used = 0;
161}
162
163/* Congestion state accounting after a packet has been sent. */
164static void tcp_event_data_sent(struct tcp_sock *tp,
165 struct sock *sk)
166{
167 struct inet_connection_sock *icsk = inet_csk(sk);
168 const u32 now = tcp_time_stamp;
169 const struct dst_entry *dst = __sk_dst_get(sk);
170
171 if (sysctl_tcp_slow_start_after_idle &&
172 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
173 tcp_cwnd_restart(sk, __sk_dst_get(sk));
174
175 tp->lsndtime = now;
176
177 /* If it is a reply for ato after last received
178 * packet, enter pingpong mode.
179 */
180 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato &&
181 (!dst || !dst_metric(dst, RTAX_QUICKACK)))
182 icsk->icsk_ack.pingpong = 1;
183}
184
185/* Account for an ACK we sent. */
186static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
187{
188 tcp_dec_quickack_mode(sk, pkts);
189 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
190}
191
192
193u32 tcp_default_init_rwnd(u32 mss)
194{
195 /* Initial receive window should be twice of TCP_INIT_CWND to
196 * enable proper sending of new unsent data during fast recovery
197 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
198 * limit when mss is larger than 1460.
199 */
200 u32 init_rwnd = TCP_INIT_CWND * 2;
201
202 if (mss > 1460)
203 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
204 return init_rwnd;
205}
206
207/* Determine a window scaling and initial window to offer.
208 * Based on the assumption that the given amount of space
209 * will be offered. Store the results in the tp structure.
210 * NOTE: for smooth operation initial space offering should
211 * be a multiple of mss if possible. We assume here that mss >= 1.
212 * This MUST be enforced by all callers.
213 */
214void tcp_select_initial_window(int __space, __u32 mss,
215 __u32 *rcv_wnd, __u32 *window_clamp,
216 int wscale_ok, __u8 *rcv_wscale,
217 __u32 init_rcv_wnd)
218{
219 unsigned int space = (__space < 0 ? 0 : __space);
220
221 /* If no clamp set the clamp to the max possible scaled window */
222 if (*window_clamp == 0)
223 (*window_clamp) = (65535 << 14);
224 space = min(*window_clamp, space);
225
226 /* Quantize space offering to a multiple of mss if possible. */
227 if (space > mss)
228 space = (space / mss) * mss;
229
230 /* NOTE: offering an initial window larger than 32767
231 * will break some buggy TCP stacks. If the admin tells us
232 * it is likely we could be speaking with such a buggy stack
233 * we will truncate our initial window offering to 32K-1
234 * unless the remote has sent us a window scaling option,
235 * which we interpret as a sign the remote TCP is not
236 * misinterpreting the window field as a signed quantity.
237 */
238 if (sysctl_tcp_workaround_signed_windows)
239 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
240 else
241 (*rcv_wnd) = space;
242
243 (*rcv_wscale) = 0;
244 if (wscale_ok) {
245 /* Set window scaling on max possible window
246 * See RFC1323 for an explanation of the limit to 14
247 */
248 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
249 space = min_t(u32, space, *window_clamp);
250 while (space > 65535 && (*rcv_wscale) < 14) {
251 space >>= 1;
252 (*rcv_wscale)++;
253 }
254 }
255
256 if (mss > (1 << *rcv_wscale)) {
257 if (!init_rcv_wnd) /* Use default unless specified otherwise */
258 init_rcv_wnd = tcp_default_init_rwnd(mss);
259 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
260 }
261
262 /* Set the clamp no higher than max representable value */
263 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
264}
265EXPORT_SYMBOL(tcp_select_initial_window);
266
267/* Chose a new window to advertise, update state in tcp_sock for the
268 * socket, and return result with RFC1323 scaling applied. The return
269 * value can be stuffed directly into th->window for an outgoing
270 * frame.
271 */
272static u16 tcp_select_window(struct sock *sk)
273{
274 struct tcp_sock *tp = tcp_sk(sk);
275 u32 old_win = tp->rcv_wnd;
276 u32 cur_win = tcp_receive_window(tp);
277 u32 new_win = __tcp_select_window(sk);
278
279 /* Never shrink the offered window */
280 if (new_win < cur_win) {
281 /* Danger Will Robinson!
282 * Don't update rcv_wup/rcv_wnd here or else
283 * we will not be able to advertise a zero
284 * window in time. --DaveM
285 *
286 * Relax Will Robinson.
287 */
288 if (new_win == 0)
289 NET_INC_STATS(sock_net(sk),
290 LINUX_MIB_TCPWANTZEROWINDOWADV);
291 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
292 }
293 tp->rcv_wnd = new_win;
294 tp->rcv_wup = tp->rcv_nxt;
295
296 /* Make sure we do not exceed the maximum possible
297 * scaled window.
298 */
299 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
300 new_win = min(new_win, MAX_TCP_WINDOW);
301 else
302 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
303
304 /* RFC1323 scaling applied */
305 new_win >>= tp->rx_opt.rcv_wscale;
306
307 /* If we advertise zero window, disable fast path. */
308 if (new_win == 0) {
309 tp->pred_flags = 0;
310 if (old_win)
311 NET_INC_STATS(sock_net(sk),
312 LINUX_MIB_TCPTOZEROWINDOWADV);
313 } else if (old_win == 0) {
314 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
315 }
316
317 return new_win;
318}
319
320/* Packet ECN state for a SYN-ACK */
321static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb)
322{
323 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
324 if (!(tp->ecn_flags & TCP_ECN_OK))
325 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
326}
327
328/* Packet ECN state for a SYN. */
329static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
330{
331 struct tcp_sock *tp = tcp_sk(sk);
332
333 tp->ecn_flags = 0;
334 if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) {
335 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
336 tp->ecn_flags = TCP_ECN_OK;
337 }
338}
339
340static __inline__ void
341TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th)
342{
343 if (inet_rsk(req)->ecn_ok)
344 th->ece = 1;
345}
346
347/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
348 * be sent.
349 */
350static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
351 int tcp_header_len)
352{
353 struct tcp_sock *tp = tcp_sk(sk);
354
355 if (tp->ecn_flags & TCP_ECN_OK) {
356 /* Not-retransmitted data segment: set ECT and inject CWR. */
357 if (skb->len != tcp_header_len &&
358 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
359 INET_ECN_xmit(sk);
360 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
361 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
362 tcp_hdr(skb)->cwr = 1;
363 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
364 }
365 } else {
366 /* ACK or retransmitted segment: clear ECT|CE */
367 INET_ECN_dontxmit(sk);
368 }
369 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
370 tcp_hdr(skb)->ece = 1;
371 }
372}
373
374/* Constructs common control bits of non-data skb. If SYN/FIN is present,
375 * auto increment end seqno.
376 */
377static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
378{
379 struct skb_shared_info *shinfo = skb_shinfo(skb);
380
381 skb->ip_summed = CHECKSUM_PARTIAL;
382 skb->csum = 0;
383
384 TCP_SKB_CB(skb)->tcp_flags = flags;
385 TCP_SKB_CB(skb)->sacked = 0;
386
387 shinfo->gso_segs = 1;
388 shinfo->gso_size = 0;
389 shinfo->gso_type = 0;
390
391 TCP_SKB_CB(skb)->seq = seq;
392 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
393 seq++;
394 TCP_SKB_CB(skb)->end_seq = seq;
395}
396
397static inline bool tcp_urg_mode(const struct tcp_sock *tp)
398{
399 return tp->snd_una != tp->snd_up;
400}
401
402#define OPTION_SACK_ADVERTISE (1 << 0)
403#define OPTION_TS (1 << 1)
404#define OPTION_MD5 (1 << 2)
405#define OPTION_WSCALE (1 << 3)
406#define OPTION_FAST_OPEN_COOKIE (1 << 8)
407
408struct tcp_out_options {
409 u16 options; /* bit field of OPTION_* */
410 u16 mss; /* 0 to disable */
411 u8 ws; /* window scale, 0 to disable */
412 u8 num_sack_blocks; /* number of SACK blocks to include */
413 u8 hash_size; /* bytes in hash_location */
414 __u8 *hash_location; /* temporary pointer, overloaded */
415 __u32 tsval, tsecr; /* need to include OPTION_TS */
416 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
417};
418
419/* Write previously computed TCP options to the packet.
420 *
421 * Beware: Something in the Internet is very sensitive to the ordering of
422 * TCP options, we learned this through the hard way, so be careful here.
423 * Luckily we can at least blame others for their non-compliance but from
424 * inter-operability perspective it seems that we're somewhat stuck with
425 * the ordering which we have been using if we want to keep working with
426 * those broken things (not that it currently hurts anybody as there isn't
427 * particular reason why the ordering would need to be changed).
428 *
429 * At least SACK_PERM as the first option is known to lead to a disaster
430 * (but it may well be that other scenarios fail similarly).
431 */
432static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
433 struct tcp_out_options *opts)
434{
435 u16 options = opts->options; /* mungable copy */
436
437 if (unlikely(OPTION_MD5 & options)) {
438 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
439 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
440 /* overload cookie hash location */
441 opts->hash_location = (__u8 *)ptr;
442 ptr += 4;
443 }
444
445 if (unlikely(opts->mss)) {
446 *ptr++ = htonl((TCPOPT_MSS << 24) |
447 (TCPOLEN_MSS << 16) |
448 opts->mss);
449 }
450
451 if (likely(OPTION_TS & options)) {
452 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
453 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
454 (TCPOLEN_SACK_PERM << 16) |
455 (TCPOPT_TIMESTAMP << 8) |
456 TCPOLEN_TIMESTAMP);
457 options &= ~OPTION_SACK_ADVERTISE;
458 } else {
459 *ptr++ = htonl((TCPOPT_NOP << 24) |
460 (TCPOPT_NOP << 16) |
461 (TCPOPT_TIMESTAMP << 8) |
462 TCPOLEN_TIMESTAMP);
463 }
464 *ptr++ = htonl(opts->tsval);
465 *ptr++ = htonl(opts->tsecr);
466 }
467
468 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
469 *ptr++ = htonl((TCPOPT_NOP << 24) |
470 (TCPOPT_NOP << 16) |
471 (TCPOPT_SACK_PERM << 8) |
472 TCPOLEN_SACK_PERM);
473 }
474
475 if (unlikely(OPTION_WSCALE & options)) {
476 *ptr++ = htonl((TCPOPT_NOP << 24) |
477 (TCPOPT_WINDOW << 16) |
478 (TCPOLEN_WINDOW << 8) |
479 opts->ws);
480 }
481
482 if (unlikely(opts->num_sack_blocks)) {
483 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
484 tp->duplicate_sack : tp->selective_acks;
485 int this_sack;
486
487 *ptr++ = htonl((TCPOPT_NOP << 24) |
488 (TCPOPT_NOP << 16) |
489 (TCPOPT_SACK << 8) |
490 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
491 TCPOLEN_SACK_PERBLOCK)));
492
493 for (this_sack = 0; this_sack < opts->num_sack_blocks;
494 ++this_sack) {
495 *ptr++ = htonl(sp[this_sack].start_seq);
496 *ptr++ = htonl(sp[this_sack].end_seq);
497 }
498
499 tp->rx_opt.dsack = 0;
500 }
501
502 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
503 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
504
505 *ptr++ = htonl((TCPOPT_EXP << 24) |
506 ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) |
507 TCPOPT_FASTOPEN_MAGIC);
508
509 memcpy(ptr, foc->val, foc->len);
510 if ((foc->len & 3) == 2) {
511 u8 *align = ((u8 *)ptr) + foc->len;
512 align[0] = align[1] = TCPOPT_NOP;
513 }
514 ptr += (foc->len + 3) >> 2;
515 }
516}
517
518/* Compute TCP options for SYN packets. This is not the final
519 * network wire format yet.
520 */
521static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
522 struct tcp_out_options *opts,
523 struct tcp_md5sig_key **md5)
524{
525 struct tcp_sock *tp = tcp_sk(sk);
526 unsigned int remaining = MAX_TCP_OPTION_SPACE;
527 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
528
529#ifdef CONFIG_TCP_MD5SIG
530 *md5 = tp->af_specific->md5_lookup(sk, sk);
531 if (*md5) {
532 opts->options |= OPTION_MD5;
533 remaining -= TCPOLEN_MD5SIG_ALIGNED;
534 }
535#else
536 *md5 = NULL;
537#endif
538
539 /* We always get an MSS option. The option bytes which will be seen in
540 * normal data packets should timestamps be used, must be in the MSS
541 * advertised. But we subtract them from tp->mss_cache so that
542 * calculations in tcp_sendmsg are simpler etc. So account for this
543 * fact here if necessary. If we don't do this correctly, as a
544 * receiver we won't recognize data packets as being full sized when we
545 * should, and thus we won't abide by the delayed ACK rules correctly.
546 * SACKs don't matter, we never delay an ACK when we have any of those
547 * going out. */
548 opts->mss = tcp_advertise_mss(sk);
549 remaining -= TCPOLEN_MSS_ALIGNED;
550
551 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
552 opts->options |= OPTION_TS;
553 opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset;
554 opts->tsecr = tp->rx_opt.ts_recent;
555 remaining -= TCPOLEN_TSTAMP_ALIGNED;
556 }
557 if (likely(sysctl_tcp_window_scaling)) {
558 opts->ws = tp->rx_opt.rcv_wscale;
559 opts->options |= OPTION_WSCALE;
560 remaining -= TCPOLEN_WSCALE_ALIGNED;
561 }
562 if (likely(sysctl_tcp_sack)) {
563 opts->options |= OPTION_SACK_ADVERTISE;
564 if (unlikely(!(OPTION_TS & opts->options)))
565 remaining -= TCPOLEN_SACKPERM_ALIGNED;
566 }
567
568 if (fastopen && fastopen->cookie.len >= 0) {
569 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len;
570 need = (need + 3) & ~3U; /* Align to 32 bits */
571 if (remaining >= need) {
572 opts->options |= OPTION_FAST_OPEN_COOKIE;
573 opts->fastopen_cookie = &fastopen->cookie;
574 remaining -= need;
575 tp->syn_fastopen = 1;
576 }
577 }
578
579 return MAX_TCP_OPTION_SPACE - remaining;
580}
581
582/* Set up TCP options for SYN-ACKs. */
583static unsigned int tcp_synack_options(struct sock *sk,
584 struct request_sock *req,
585 unsigned int mss, struct sk_buff *skb,
586 struct tcp_out_options *opts,
587 struct tcp_md5sig_key **md5,
588 struct tcp_fastopen_cookie *foc)
589{
590 struct inet_request_sock *ireq = inet_rsk(req);
591 unsigned int remaining = MAX_TCP_OPTION_SPACE;
592
593#ifdef CONFIG_TCP_MD5SIG
594 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
595 if (*md5) {
596 opts->options |= OPTION_MD5;
597 remaining -= TCPOLEN_MD5SIG_ALIGNED;
598
599 /* We can't fit any SACK blocks in a packet with MD5 + TS
600 * options. There was discussion about disabling SACK
601 * rather than TS in order to fit in better with old,
602 * buggy kernels, but that was deemed to be unnecessary.
603 */
604 ireq->tstamp_ok &= !ireq->sack_ok;
605 }
606#else
607 *md5 = NULL;
608#endif
609
610 /* We always send an MSS option. */
611 opts->mss = mss;
612 remaining -= TCPOLEN_MSS_ALIGNED;
613
614 if (likely(ireq->wscale_ok)) {
615 opts->ws = ireq->rcv_wscale;
616 opts->options |= OPTION_WSCALE;
617 remaining -= TCPOLEN_WSCALE_ALIGNED;
618 }
619 if (likely(ireq->tstamp_ok)) {
620 opts->options |= OPTION_TS;
621 opts->tsval = TCP_SKB_CB(skb)->when;
622 opts->tsecr = req->ts_recent;
623 remaining -= TCPOLEN_TSTAMP_ALIGNED;
624 }
625 if (likely(ireq->sack_ok)) {
626 opts->options |= OPTION_SACK_ADVERTISE;
627 if (unlikely(!ireq->tstamp_ok))
628 remaining -= TCPOLEN_SACKPERM_ALIGNED;
629 }
630 if (foc != NULL) {
631 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
632 need = (need + 3) & ~3U; /* Align to 32 bits */
633 if (remaining >= need) {
634 opts->options |= OPTION_FAST_OPEN_COOKIE;
635 opts->fastopen_cookie = foc;
636 remaining -= need;
637 }
638 }
639
640 return MAX_TCP_OPTION_SPACE - remaining;
641}
642
643/* Compute TCP options for ESTABLISHED sockets. This is not the
644 * final wire format yet.
645 */
646static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
647 struct tcp_out_options *opts,
648 struct tcp_md5sig_key **md5)
649{
650 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
651 struct tcp_sock *tp = tcp_sk(sk);
652 unsigned int size = 0;
653 unsigned int eff_sacks;
654
655 opts->options = 0;
656
657#ifdef CONFIG_TCP_MD5SIG
658 *md5 = tp->af_specific->md5_lookup(sk, sk);
659 if (unlikely(*md5)) {
660 opts->options |= OPTION_MD5;
661 size += TCPOLEN_MD5SIG_ALIGNED;
662 }
663#else
664 *md5 = NULL;
665#endif
666
667 if (likely(tp->rx_opt.tstamp_ok)) {
668 opts->options |= OPTION_TS;
669 opts->tsval = tcb ? tcb->when + tp->tsoffset : 0;
670 opts->tsecr = tp->rx_opt.ts_recent;
671 size += TCPOLEN_TSTAMP_ALIGNED;
672 }
673
674 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
675 if (unlikely(eff_sacks)) {
676 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
677 opts->num_sack_blocks =
678 min_t(unsigned int, eff_sacks,
679 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
680 TCPOLEN_SACK_PERBLOCK);
681 size += TCPOLEN_SACK_BASE_ALIGNED +
682 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
683 }
684
685 return size;
686}
687
688
689/* TCP SMALL QUEUES (TSQ)
690 *
691 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
692 * to reduce RTT and bufferbloat.
693 * We do this using a special skb destructor (tcp_wfree).
694 *
695 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
696 * needs to be reallocated in a driver.
697 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
698 *
699 * Since transmit from skb destructor is forbidden, we use a tasklet
700 * to process all sockets that eventually need to send more skbs.
701 * We use one tasklet per cpu, with its own queue of sockets.
702 */
703struct tsq_tasklet {
704 struct tasklet_struct tasklet;
705 struct list_head head; /* queue of tcp sockets */
706};
707static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
708
709static void tcp_tsq_handler(struct sock *sk)
710{
711 if ((1 << sk->sk_state) &
712 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
713 TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
714 tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle,
715 0, GFP_ATOMIC);
716}
717/*
718 * One tasklet per cpu tries to send more skbs.
719 * We run in tasklet context but need to disable irqs when
720 * transferring tsq->head because tcp_wfree() might
721 * interrupt us (non NAPI drivers)
722 */
723static void tcp_tasklet_func(unsigned long data)
724{
725 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
726 LIST_HEAD(list);
727 unsigned long flags;
728 struct list_head *q, *n;
729 struct tcp_sock *tp;
730 struct sock *sk;
731
732 local_irq_save(flags);
733 list_splice_init(&tsq->head, &list);
734 local_irq_restore(flags);
735
736 list_for_each_safe(q, n, &list) {
737 tp = list_entry(q, struct tcp_sock, tsq_node);
738 list_del(&tp->tsq_node);
739
740 sk = (struct sock *)tp;
741 bh_lock_sock(sk);
742
743 if (!sock_owned_by_user(sk)) {
744 tcp_tsq_handler(sk);
745 } else {
746 /* defer the work to tcp_release_cb() */
747 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
748 }
749 bh_unlock_sock(sk);
750
751 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
752 sk_free(sk);
753 }
754}
755
756#define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
757 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
758 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
759 (1UL << TCP_MTU_REDUCED_DEFERRED))
760/**
761 * tcp_release_cb - tcp release_sock() callback
762 * @sk: socket
763 *
764 * called from release_sock() to perform protocol dependent
765 * actions before socket release.
766 */
767void tcp_release_cb(struct sock *sk)
768{
769 struct tcp_sock *tp = tcp_sk(sk);
770 unsigned long flags, nflags;
771
772 /* perform an atomic operation only if at least one flag is set */
773 do {
774 flags = tp->tsq_flags;
775 if (!(flags & TCP_DEFERRED_ALL))
776 return;
777 nflags = flags & ~TCP_DEFERRED_ALL;
778 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
779
780 if (flags & (1UL << TCP_TSQ_DEFERRED))
781 tcp_tsq_handler(sk);
782
783 /* Here begins the tricky part :
784 * We are called from release_sock() with :
785 * 1) BH disabled
786 * 2) sk_lock.slock spinlock held
787 * 3) socket owned by us (sk->sk_lock.owned == 1)
788 *
789 * But following code is meant to be called from BH handlers,
790 * so we should keep BH disabled, but early release socket ownership
791 */
792 sock_release_ownership(sk);
793
794 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
795 tcp_write_timer_handler(sk);
796 __sock_put(sk);
797 }
798 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
799 tcp_delack_timer_handler(sk);
800 __sock_put(sk);
801 }
802 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
803 sk->sk_prot->mtu_reduced(sk);
804 __sock_put(sk);
805 }
806}
807EXPORT_SYMBOL(tcp_release_cb);
808
809void __init tcp_tasklet_init(void)
810{
811 int i;
812
813 for_each_possible_cpu(i) {
814 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
815
816 INIT_LIST_HEAD(&tsq->head);
817 tasklet_init(&tsq->tasklet,
818 tcp_tasklet_func,
819 (unsigned long)tsq);
820 }
821}
822
823/*
824 * Write buffer destructor automatically called from kfree_skb.
825 * We can't xmit new skbs from this context, as we might already
826 * hold qdisc lock.
827 */
828void tcp_wfree(struct sk_buff *skb)
829{
830 struct sock *sk = skb->sk;
831 struct tcp_sock *tp = tcp_sk(sk);
832
833 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
834 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
835 unsigned long flags;
836 struct tsq_tasklet *tsq;
837
838 /* Keep a ref on socket.
839 * This last ref will be released in tcp_tasklet_func()
840 */
841 atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc);
842
843 /* queue this socket to tasklet queue */
844 local_irq_save(flags);
845 tsq = &__get_cpu_var(tsq_tasklet);
846 list_add(&tp->tsq_node, &tsq->head);
847 tasklet_schedule(&tsq->tasklet);
848 local_irq_restore(flags);
849 } else {
850 sock_wfree(skb);
851 }
852}
853
854/* This routine actually transmits TCP packets queued in by
855 * tcp_do_sendmsg(). This is used by both the initial
856 * transmission and possible later retransmissions.
857 * All SKB's seen here are completely headerless. It is our
858 * job to build the TCP header, and pass the packet down to
859 * IP so it can do the same plus pass the packet off to the
860 * device.
861 *
862 * We are working here with either a clone of the original
863 * SKB, or a fresh unique copy made by the retransmit engine.
864 */
865static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
866 gfp_t gfp_mask)
867{
868 const struct inet_connection_sock *icsk = inet_csk(sk);
869 struct inet_sock *inet;
870 struct tcp_sock *tp;
871 struct tcp_skb_cb *tcb;
872 struct tcp_out_options opts;
873 unsigned int tcp_options_size, tcp_header_size;
874 struct tcp_md5sig_key *md5;
875 struct tcphdr *th;
876 int err;
877
878 BUG_ON(!skb || !tcp_skb_pcount(skb));
879
880 if (clone_it) {
881 const struct sk_buff *fclone = skb + 1;
882
883 skb_mstamp_get(&skb->skb_mstamp);
884
885 if (unlikely(skb->fclone == SKB_FCLONE_ORIG &&
886 fclone->fclone == SKB_FCLONE_CLONE))
887 NET_INC_STATS(sock_net(sk),
888 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
889
890 if (unlikely(skb_cloned(skb)))
891 skb = pskb_copy(skb, gfp_mask);
892 else
893 skb = skb_clone(skb, gfp_mask);
894 if (unlikely(!skb))
895 return -ENOBUFS;
896 /* Our usage of tstamp should remain private */
897 skb->tstamp.tv64 = 0;
898 }
899
900 inet = inet_sk(sk);
901 tp = tcp_sk(sk);
902 tcb = TCP_SKB_CB(skb);
903 memset(&opts, 0, sizeof(opts));
904
905 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
906 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
907 else
908 tcp_options_size = tcp_established_options(sk, skb, &opts,
909 &md5);
910 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
911
912 if (tcp_packets_in_flight(tp) == 0)
913 tcp_ca_event(sk, CA_EVENT_TX_START);
914
915 /* if no packet is in qdisc/device queue, then allow XPS to select
916 * another queue.
917 */
918 skb->ooo_okay = sk_wmem_alloc_get(sk) == 0;
919
920 skb_push(skb, tcp_header_size);
921 skb_reset_transport_header(skb);
922
923 skb_orphan(skb);
924 skb->sk = sk;
925 skb->destructor = tcp_wfree;
926 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
927
928 /* Build TCP header and checksum it. */
929 th = tcp_hdr(skb);
930 th->source = inet->inet_sport;
931 th->dest = inet->inet_dport;
932 th->seq = htonl(tcb->seq);
933 th->ack_seq = htonl(tp->rcv_nxt);
934 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
935 tcb->tcp_flags);
936
937 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
938 /* RFC1323: The window in SYN & SYN/ACK segments
939 * is never scaled.
940 */
941 th->window = htons(min(tp->rcv_wnd, 65535U));
942 } else {
943 th->window = htons(tcp_select_window(sk));
944 }
945 th->check = 0;
946 th->urg_ptr = 0;
947
948 /* The urg_mode check is necessary during a below snd_una win probe */
949 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
950 if (before(tp->snd_up, tcb->seq + 0x10000)) {
951 th->urg_ptr = htons(tp->snd_up - tcb->seq);
952 th->urg = 1;
953 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
954 th->urg_ptr = htons(0xFFFF);
955 th->urg = 1;
956 }
957 }
958
959 tcp_options_write((__be32 *)(th + 1), tp, &opts);
960 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
961 TCP_ECN_send(sk, skb, tcp_header_size);
962
963#ifdef CONFIG_TCP_MD5SIG
964 /* Calculate the MD5 hash, as we have all we need now */
965 if (md5) {
966 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
967 tp->af_specific->calc_md5_hash(opts.hash_location,
968 md5, sk, NULL, skb);
969 }
970#endif
971
972 icsk->icsk_af_ops->send_check(sk, skb);
973
974 if (likely(tcb->tcp_flags & TCPHDR_ACK))
975 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
976
977 if (skb->len != tcp_header_size)
978 tcp_event_data_sent(tp, sk);
979
980 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
981 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
982 tcp_skb_pcount(skb));
983
984 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
985 if (likely(err <= 0))
986 return err;
987
988 tcp_enter_cwr(sk, 1);
989
990 return net_xmit_eval(err);
991}
992
993/* This routine just queues the buffer for sending.
994 *
995 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
996 * otherwise socket can stall.
997 */
998static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
999{
1000 struct tcp_sock *tp = tcp_sk(sk);
1001
1002 /* Advance write_seq and place onto the write_queue. */
1003 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1004 skb_header_release(skb);
1005 tcp_add_write_queue_tail(sk, skb);
1006 sk->sk_wmem_queued += skb->truesize;
1007 sk_mem_charge(sk, skb->truesize);
1008}
1009
1010/* Initialize TSO segments for a packet. */
1011static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
1012 unsigned int mss_now)
1013{
1014 struct skb_shared_info *shinfo = skb_shinfo(skb);
1015
1016 /* Make sure we own this skb before messing gso_size/gso_segs */
1017 WARN_ON_ONCE(skb_cloned(skb));
1018
1019 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1020 /* Avoid the costly divide in the normal
1021 * non-TSO case.
1022 */
1023 shinfo->gso_segs = 1;
1024 shinfo->gso_size = 0;
1025 shinfo->gso_type = 0;
1026 } else {
1027 shinfo->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
1028 shinfo->gso_size = mss_now;
1029 shinfo->gso_type = sk->sk_gso_type;
1030 }
1031}
1032
1033/* When a modification to fackets out becomes necessary, we need to check
1034 * skb is counted to fackets_out or not.
1035 */
1036static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1037 int decr)
1038{
1039 struct tcp_sock *tp = tcp_sk(sk);
1040
1041 if (!tp->sacked_out || tcp_is_reno(tp))
1042 return;
1043
1044 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1045 tp->fackets_out -= decr;
1046}
1047
1048/* Pcount in the middle of the write queue got changed, we need to do various
1049 * tweaks to fix counters
1050 */
1051static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1052{
1053 struct tcp_sock *tp = tcp_sk(sk);
1054
1055 tp->packets_out -= decr;
1056
1057 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1058 tp->sacked_out -= decr;
1059 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1060 tp->retrans_out -= decr;
1061 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1062 tp->lost_out -= decr;
1063
1064 /* Reno case is special. Sigh... */
1065 if (tcp_is_reno(tp) && decr > 0)
1066 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1067
1068 tcp_adjust_fackets_out(sk, skb, decr);
1069
1070 if (tp->lost_skb_hint &&
1071 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1072 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1073 tp->lost_cnt_hint -= decr;
1074
1075 tcp_verify_left_out(tp);
1076}
1077
1078/* Function to create two new TCP segments. Shrinks the given segment
1079 * to the specified size and appends a new segment with the rest of the
1080 * packet to the list. This won't be called frequently, I hope.
1081 * Remember, these are still headerless SKBs at this point.
1082 */
1083int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1084 unsigned int mss_now)
1085{
1086 struct tcp_sock *tp = tcp_sk(sk);
1087 struct sk_buff *buff;
1088 int nsize, old_factor;
1089 int nlen;
1090 u8 flags;
1091
1092 if (WARN_ON(len > skb->len))
1093 return -EINVAL;
1094
1095 nsize = skb_headlen(skb) - len;
1096 if (nsize < 0)
1097 nsize = 0;
1098
1099 if (skb_unclone(skb, GFP_ATOMIC))
1100 return -ENOMEM;
1101
1102 /* Get a new skb... force flag on. */
1103 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1104 if (buff == NULL)
1105 return -ENOMEM; /* We'll just try again later. */
1106
1107 sk->sk_wmem_queued += buff->truesize;
1108 sk_mem_charge(sk, buff->truesize);
1109 nlen = skb->len - len - nsize;
1110 buff->truesize += nlen;
1111 skb->truesize -= nlen;
1112
1113 /* Correct the sequence numbers. */
1114 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1115 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1116 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1117
1118 /* PSH and FIN should only be set in the second packet. */
1119 flags = TCP_SKB_CB(skb)->tcp_flags;
1120 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1121 TCP_SKB_CB(buff)->tcp_flags = flags;
1122 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1123
1124 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1125 /* Copy and checksum data tail into the new buffer. */
1126 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1127 skb_put(buff, nsize),
1128 nsize, 0);
1129
1130 skb_trim(skb, len);
1131
1132 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1133 } else {
1134 skb->ip_summed = CHECKSUM_PARTIAL;
1135 skb_split(skb, buff, len);
1136 }
1137
1138 buff->ip_summed = skb->ip_summed;
1139
1140 /* Looks stupid, but our code really uses when of
1141 * skbs, which it never sent before. --ANK
1142 */
1143 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1144 buff->tstamp = skb->tstamp;
1145
1146 old_factor = tcp_skb_pcount(skb);
1147
1148 /* Fix up tso_factor for both original and new SKB. */
1149 tcp_set_skb_tso_segs(sk, skb, mss_now);
1150 tcp_set_skb_tso_segs(sk, buff, mss_now);
1151
1152 /* If this packet has been sent out already, we must
1153 * adjust the various packet counters.
1154 */
1155 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1156 int diff = old_factor - tcp_skb_pcount(skb) -
1157 tcp_skb_pcount(buff);
1158
1159 if (diff)
1160 tcp_adjust_pcount(sk, skb, diff);
1161 }
1162
1163 /* Link BUFF into the send queue. */
1164 skb_header_release(buff);
1165 tcp_insert_write_queue_after(skb, buff, sk);
1166
1167 return 0;
1168}
1169
1170/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1171 * eventually). The difference is that pulled data not copied, but
1172 * immediately discarded.
1173 */
1174static void __pskb_trim_head(struct sk_buff *skb, int len)
1175{
1176 struct skb_shared_info *shinfo;
1177 int i, k, eat;
1178
1179 eat = min_t(int, len, skb_headlen(skb));
1180 if (eat) {
1181 __skb_pull(skb, eat);
1182 len -= eat;
1183 if (!len)
1184 return;
1185 }
1186 eat = len;
1187 k = 0;
1188 shinfo = skb_shinfo(skb);
1189 for (i = 0; i < shinfo->nr_frags; i++) {
1190 int size = skb_frag_size(&shinfo->frags[i]);
1191
1192 if (size <= eat) {
1193 skb_frag_unref(skb, i);
1194 eat -= size;
1195 } else {
1196 shinfo->frags[k] = shinfo->frags[i];
1197 if (eat) {
1198 shinfo->frags[k].page_offset += eat;
1199 skb_frag_size_sub(&shinfo->frags[k], eat);
1200 eat = 0;
1201 }
1202 k++;
1203 }
1204 }
1205 shinfo->nr_frags = k;
1206
1207 skb_reset_tail_pointer(skb);
1208 skb->data_len -= len;
1209 skb->len = skb->data_len;
1210}
1211
1212/* Remove acked data from a packet in the transmit queue. */
1213int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1214{
1215 if (skb_unclone(skb, GFP_ATOMIC))
1216 return -ENOMEM;
1217
1218 __pskb_trim_head(skb, len);
1219
1220 TCP_SKB_CB(skb)->seq += len;
1221 skb->ip_summed = CHECKSUM_PARTIAL;
1222
1223 skb->truesize -= len;
1224 sk->sk_wmem_queued -= len;
1225 sk_mem_uncharge(sk, len);
1226 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1227
1228 /* Any change of skb->len requires recalculation of tso factor. */
1229 if (tcp_skb_pcount(skb) > 1)
1230 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1231
1232 return 0;
1233}
1234
1235/* Calculate MSS not accounting any TCP options. */
1236static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1237{
1238 const struct tcp_sock *tp = tcp_sk(sk);
1239 const struct inet_connection_sock *icsk = inet_csk(sk);
1240 int mss_now;
1241
1242 /* Calculate base mss without TCP options:
1243 It is MMS_S - sizeof(tcphdr) of rfc1122
1244 */
1245 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1246
1247 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1248 if (icsk->icsk_af_ops->net_frag_header_len) {
1249 const struct dst_entry *dst = __sk_dst_get(sk);
1250
1251 if (dst && dst_allfrag(dst))
1252 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1253 }
1254
1255 /* Clamp it (mss_clamp does not include tcp options) */
1256 if (mss_now > tp->rx_opt.mss_clamp)
1257 mss_now = tp->rx_opt.mss_clamp;
1258
1259 /* Now subtract optional transport overhead */
1260 mss_now -= icsk->icsk_ext_hdr_len;
1261
1262 /* Then reserve room for full set of TCP options and 8 bytes of data */
1263 if (mss_now < 48)
1264 mss_now = 48;
1265 return mss_now;
1266}
1267
1268/* Calculate MSS. Not accounting for SACKs here. */
1269int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1270{
1271 /* Subtract TCP options size, not including SACKs */
1272 return __tcp_mtu_to_mss(sk, pmtu) -
1273 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1274}
1275
1276/* Inverse of above */
1277int tcp_mss_to_mtu(struct sock *sk, int mss)
1278{
1279 const struct tcp_sock *tp = tcp_sk(sk);
1280 const struct inet_connection_sock *icsk = inet_csk(sk);
1281 int mtu;
1282
1283 mtu = mss +
1284 tp->tcp_header_len +
1285 icsk->icsk_ext_hdr_len +
1286 icsk->icsk_af_ops->net_header_len;
1287
1288 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1289 if (icsk->icsk_af_ops->net_frag_header_len) {
1290 const struct dst_entry *dst = __sk_dst_get(sk);
1291
1292 if (dst && dst_allfrag(dst))
1293 mtu += icsk->icsk_af_ops->net_frag_header_len;
1294 }
1295 return mtu;
1296}
1297
1298/* MTU probing init per socket */
1299void tcp_mtup_init(struct sock *sk)
1300{
1301 struct tcp_sock *tp = tcp_sk(sk);
1302 struct inet_connection_sock *icsk = inet_csk(sk);
1303
1304 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1305 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1306 icsk->icsk_af_ops->net_header_len;
1307 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1308 icsk->icsk_mtup.probe_size = 0;
1309}
1310EXPORT_SYMBOL(tcp_mtup_init);
1311
1312/* This function synchronize snd mss to current pmtu/exthdr set.
1313
1314 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1315 for TCP options, but includes only bare TCP header.
1316
1317 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1318 It is minimum of user_mss and mss received with SYN.
1319 It also does not include TCP options.
1320
1321 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1322
1323 tp->mss_cache is current effective sending mss, including
1324 all tcp options except for SACKs. It is evaluated,
1325 taking into account current pmtu, but never exceeds
1326 tp->rx_opt.mss_clamp.
1327
1328 NOTE1. rfc1122 clearly states that advertised MSS
1329 DOES NOT include either tcp or ip options.
1330
1331 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1332 are READ ONLY outside this function. --ANK (980731)
1333 */
1334unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1335{
1336 struct tcp_sock *tp = tcp_sk(sk);
1337 struct inet_connection_sock *icsk = inet_csk(sk);
1338 int mss_now;
1339
1340 if (icsk->icsk_mtup.search_high > pmtu)
1341 icsk->icsk_mtup.search_high = pmtu;
1342
1343 mss_now = tcp_mtu_to_mss(sk, pmtu);
1344 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1345
1346 /* And store cached results */
1347 icsk->icsk_pmtu_cookie = pmtu;
1348 if (icsk->icsk_mtup.enabled)
1349 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1350 tp->mss_cache = mss_now;
1351
1352 return mss_now;
1353}
1354EXPORT_SYMBOL(tcp_sync_mss);
1355
1356/* Compute the current effective MSS, taking SACKs and IP options,
1357 * and even PMTU discovery events into account.
1358 */
1359unsigned int tcp_current_mss(struct sock *sk)
1360{
1361 const struct tcp_sock *tp = tcp_sk(sk);
1362 const struct dst_entry *dst = __sk_dst_get(sk);
1363 u32 mss_now;
1364 unsigned int header_len;
1365 struct tcp_out_options opts;
1366 struct tcp_md5sig_key *md5;
1367
1368 mss_now = tp->mss_cache;
1369
1370 if (dst) {
1371 u32 mtu = dst_mtu(dst);
1372 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1373 mss_now = tcp_sync_mss(sk, mtu);
1374 }
1375
1376 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1377 sizeof(struct tcphdr);
1378 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1379 * some common options. If this is an odd packet (because we have SACK
1380 * blocks etc) then our calculated header_len will be different, and
1381 * we have to adjust mss_now correspondingly */
1382 if (header_len != tp->tcp_header_len) {
1383 int delta = (int) header_len - tp->tcp_header_len;
1384 mss_now -= delta;
1385 }
1386
1387 return mss_now;
1388}
1389
1390/* Congestion window validation. (RFC2861) */
1391static void tcp_cwnd_validate(struct sock *sk)
1392{
1393 struct tcp_sock *tp = tcp_sk(sk);
1394
1395 if (tp->packets_out >= tp->snd_cwnd) {
1396 /* Network is feed fully. */
1397 tp->snd_cwnd_used = 0;
1398 tp->snd_cwnd_stamp = tcp_time_stamp;
1399 } else {
1400 /* Network starves. */
1401 if (tp->packets_out > tp->snd_cwnd_used)
1402 tp->snd_cwnd_used = tp->packets_out;
1403
1404 if (sysctl_tcp_slow_start_after_idle &&
1405 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1406 tcp_cwnd_application_limited(sk);
1407 }
1408}
1409
1410/* Minshall's variant of the Nagle send check. */
1411static bool tcp_minshall_check(const struct tcp_sock *tp)
1412{
1413 return after(tp->snd_sml, tp->snd_una) &&
1414 !after(tp->snd_sml, tp->snd_nxt);
1415}
1416
1417/* Update snd_sml if this skb is under mss
1418 * Note that a TSO packet might end with a sub-mss segment
1419 * The test is really :
1420 * if ((skb->len % mss) != 0)
1421 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1422 * But we can avoid doing the divide again given we already have
1423 * skb_pcount = skb->len / mss_now
1424 */
1425static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1426 const struct sk_buff *skb)
1427{
1428 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1429 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1430}
1431
1432/* Return false, if packet can be sent now without violation Nagle's rules:
1433 * 1. It is full sized. (provided by caller in %partial bool)
1434 * 2. Or it contains FIN. (already checked by caller)
1435 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1436 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1437 * With Minshall's modification: all sent small packets are ACKed.
1438 */
1439static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1440 int nonagle)
1441{
1442 return partial &&
1443 ((nonagle & TCP_NAGLE_CORK) ||
1444 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1445}
1446/* Returns the portion of skb which can be sent right away */
1447static unsigned int tcp_mss_split_point(const struct sock *sk,
1448 const struct sk_buff *skb,
1449 unsigned int mss_now,
1450 unsigned int max_segs,
1451 int nonagle)
1452{
1453 const struct tcp_sock *tp = tcp_sk(sk);
1454 u32 partial, needed, window, max_len;
1455
1456 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1457 max_len = mss_now * max_segs;
1458
1459 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1460 return max_len;
1461
1462 needed = min(skb->len, window);
1463
1464 if (max_len <= needed)
1465 return max_len;
1466
1467 partial = needed % mss_now;
1468 /* If last segment is not a full MSS, check if Nagle rules allow us
1469 * to include this last segment in this skb.
1470 * Otherwise, we'll split the skb at last MSS boundary
1471 */
1472 if (tcp_nagle_check(partial != 0, tp, nonagle))
1473 return needed - partial;
1474
1475 return needed;
1476}
1477
1478/* Can at least one segment of SKB be sent right now, according to the
1479 * congestion window rules? If so, return how many segments are allowed.
1480 */
1481static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1482 const struct sk_buff *skb)
1483{
1484 u32 in_flight, cwnd;
1485
1486 /* Don't be strict about the congestion window for the final FIN. */
1487 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1488 tcp_skb_pcount(skb) == 1)
1489 return 1;
1490
1491 in_flight = tcp_packets_in_flight(tp);
1492 cwnd = tp->snd_cwnd;
1493 if (in_flight < cwnd)
1494 return (cwnd - in_flight);
1495
1496 return 0;
1497}
1498
1499/* Initialize TSO state of a skb.
1500 * This must be invoked the first time we consider transmitting
1501 * SKB onto the wire.
1502 */
1503static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1504 unsigned int mss_now)
1505{
1506 int tso_segs = tcp_skb_pcount(skb);
1507
1508 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1509 tcp_set_skb_tso_segs(sk, skb, mss_now);
1510 tso_segs = tcp_skb_pcount(skb);
1511 }
1512 return tso_segs;
1513}
1514
1515
1516/* Return true if the Nagle test allows this packet to be
1517 * sent now.
1518 */
1519static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1520 unsigned int cur_mss, int nonagle)
1521{
1522 /* Nagle rule does not apply to frames, which sit in the middle of the
1523 * write_queue (they have no chances to get new data).
1524 *
1525 * This is implemented in the callers, where they modify the 'nonagle'
1526 * argument based upon the location of SKB in the send queue.
1527 */
1528 if (nonagle & TCP_NAGLE_PUSH)
1529 return true;
1530
1531 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1532 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1533 return true;
1534
1535 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1536 return true;
1537
1538 return false;
1539}
1540
1541/* Does at least the first segment of SKB fit into the send window? */
1542static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1543 const struct sk_buff *skb,
1544 unsigned int cur_mss)
1545{
1546 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1547
1548 if (skb->len > cur_mss)
1549 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1550
1551 return !after(end_seq, tcp_wnd_end(tp));
1552}
1553
1554/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1555 * should be put on the wire right now. If so, it returns the number of
1556 * packets allowed by the congestion window.
1557 */
1558static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1559 unsigned int cur_mss, int nonagle)
1560{
1561 const struct tcp_sock *tp = tcp_sk(sk);
1562 unsigned int cwnd_quota;
1563
1564 tcp_init_tso_segs(sk, skb, cur_mss);
1565
1566 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1567 return 0;
1568
1569 cwnd_quota = tcp_cwnd_test(tp, skb);
1570 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1571 cwnd_quota = 0;
1572
1573 return cwnd_quota;
1574}
1575
1576/* Test if sending is allowed right now. */
1577bool tcp_may_send_now(struct sock *sk)
1578{
1579 const struct tcp_sock *tp = tcp_sk(sk);
1580 struct sk_buff *skb = tcp_send_head(sk);
1581
1582 return skb &&
1583 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1584 (tcp_skb_is_last(sk, skb) ?
1585 tp->nonagle : TCP_NAGLE_PUSH));
1586}
1587
1588/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1589 * which is put after SKB on the list. It is very much like
1590 * tcp_fragment() except that it may make several kinds of assumptions
1591 * in order to speed up the splitting operation. In particular, we
1592 * know that all the data is in scatter-gather pages, and that the
1593 * packet has never been sent out before (and thus is not cloned).
1594 */
1595static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1596 unsigned int mss_now, gfp_t gfp)
1597{
1598 struct sk_buff *buff;
1599 int nlen = skb->len - len;
1600 u8 flags;
1601
1602 /* All of a TSO frame must be composed of paged data. */
1603 if (skb->len != skb->data_len)
1604 return tcp_fragment(sk, skb, len, mss_now);
1605
1606 buff = sk_stream_alloc_skb(sk, 0, gfp);
1607 if (unlikely(buff == NULL))
1608 return -ENOMEM;
1609
1610 sk->sk_wmem_queued += buff->truesize;
1611 sk_mem_charge(sk, buff->truesize);
1612 buff->truesize += nlen;
1613 skb->truesize -= nlen;
1614
1615 /* Correct the sequence numbers. */
1616 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1617 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1618 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1619
1620 /* PSH and FIN should only be set in the second packet. */
1621 flags = TCP_SKB_CB(skb)->tcp_flags;
1622 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1623 TCP_SKB_CB(buff)->tcp_flags = flags;
1624
1625 /* This packet was never sent out yet, so no SACK bits. */
1626 TCP_SKB_CB(buff)->sacked = 0;
1627
1628 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1629 skb_split(skb, buff, len);
1630
1631 /* Fix up tso_factor for both original and new SKB. */
1632 tcp_set_skb_tso_segs(sk, skb, mss_now);
1633 tcp_set_skb_tso_segs(sk, buff, mss_now);
1634
1635 /* Link BUFF into the send queue. */
1636 skb_header_release(buff);
1637 tcp_insert_write_queue_after(skb, buff, sk);
1638
1639 return 0;
1640}
1641
1642/* Try to defer sending, if possible, in order to minimize the amount
1643 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1644 *
1645 * This algorithm is from John Heffner.
1646 */
1647static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1648{
1649 struct tcp_sock *tp = tcp_sk(sk);
1650 const struct inet_connection_sock *icsk = inet_csk(sk);
1651 u32 send_win, cong_win, limit, in_flight;
1652 int win_divisor;
1653
1654 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1655 goto send_now;
1656
1657 if (icsk->icsk_ca_state != TCP_CA_Open)
1658 goto send_now;
1659
1660 /* Defer for less than two clock ticks. */
1661 if (tp->tso_deferred &&
1662 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1663 goto send_now;
1664
1665 in_flight = tcp_packets_in_flight(tp);
1666
1667 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1668
1669 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1670
1671 /* From in_flight test above, we know that cwnd > in_flight. */
1672 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1673
1674 limit = min(send_win, cong_win);
1675
1676 /* If a full-sized TSO skb can be sent, do it. */
1677 if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
1678 tp->xmit_size_goal_segs * tp->mss_cache))
1679 goto send_now;
1680
1681 /* Middle in queue won't get any more data, full sendable already? */
1682 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1683 goto send_now;
1684
1685 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1686 if (win_divisor) {
1687 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1688
1689 /* If at least some fraction of a window is available,
1690 * just use it.
1691 */
1692 chunk /= win_divisor;
1693 if (limit >= chunk)
1694 goto send_now;
1695 } else {
1696 /* Different approach, try not to defer past a single
1697 * ACK. Receiver should ACK every other full sized
1698 * frame, so if we have space for more than 3 frames
1699 * then send now.
1700 */
1701 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1702 goto send_now;
1703 }
1704
1705 /* Ok, it looks like it is advisable to defer.
1706 * Do not rearm the timer if already set to not break TCP ACK clocking.
1707 */
1708 if (!tp->tso_deferred)
1709 tp->tso_deferred = 1 | (jiffies << 1);
1710
1711 return true;
1712
1713send_now:
1714 tp->tso_deferred = 0;
1715 return false;
1716}
1717
1718/* Create a new MTU probe if we are ready.
1719 * MTU probe is regularly attempting to increase the path MTU by
1720 * deliberately sending larger packets. This discovers routing
1721 * changes resulting in larger path MTUs.
1722 *
1723 * Returns 0 if we should wait to probe (no cwnd available),
1724 * 1 if a probe was sent,
1725 * -1 otherwise
1726 */
1727static int tcp_mtu_probe(struct sock *sk)
1728{
1729 struct tcp_sock *tp = tcp_sk(sk);
1730 struct inet_connection_sock *icsk = inet_csk(sk);
1731 struct sk_buff *skb, *nskb, *next;
1732 int len;
1733 int probe_size;
1734 int size_needed;
1735 int copy;
1736 int mss_now;
1737
1738 /* Not currently probing/verifying,
1739 * not in recovery,
1740 * have enough cwnd, and
1741 * not SACKing (the variable headers throw things off) */
1742 if (!icsk->icsk_mtup.enabled ||
1743 icsk->icsk_mtup.probe_size ||
1744 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1745 tp->snd_cwnd < 11 ||
1746 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1747 return -1;
1748
1749 /* Very simple search strategy: just double the MSS. */
1750 mss_now = tcp_current_mss(sk);
1751 probe_size = 2 * tp->mss_cache;
1752 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1753 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1754 /* TODO: set timer for probe_converge_event */
1755 return -1;
1756 }
1757
1758 /* Have enough data in the send queue to probe? */
1759 if (tp->write_seq - tp->snd_nxt < size_needed)
1760 return -1;
1761
1762 if (tp->snd_wnd < size_needed)
1763 return -1;
1764 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1765 return 0;
1766
1767 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1768 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1769 if (!tcp_packets_in_flight(tp))
1770 return -1;
1771 else
1772 return 0;
1773 }
1774
1775 /* We're allowed to probe. Build it now. */
1776 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1777 return -1;
1778 sk->sk_wmem_queued += nskb->truesize;
1779 sk_mem_charge(sk, nskb->truesize);
1780
1781 skb = tcp_send_head(sk);
1782
1783 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1784 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1785 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1786 TCP_SKB_CB(nskb)->sacked = 0;
1787 nskb->csum = 0;
1788 nskb->ip_summed = skb->ip_summed;
1789
1790 tcp_insert_write_queue_before(nskb, skb, sk);
1791
1792 len = 0;
1793 tcp_for_write_queue_from_safe(skb, next, sk) {
1794 copy = min_t(int, skb->len, probe_size - len);
1795 if (nskb->ip_summed)
1796 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1797 else
1798 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1799 skb_put(nskb, copy),
1800 copy, nskb->csum);
1801
1802 if (skb->len <= copy) {
1803 /* We've eaten all the data from this skb.
1804 * Throw it away. */
1805 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1806 tcp_unlink_write_queue(skb, sk);
1807 sk_wmem_free_skb(sk, skb);
1808 } else {
1809 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1810 ~(TCPHDR_FIN|TCPHDR_PSH);
1811 if (!skb_shinfo(skb)->nr_frags) {
1812 skb_pull(skb, copy);
1813 if (skb->ip_summed != CHECKSUM_PARTIAL)
1814 skb->csum = csum_partial(skb->data,
1815 skb->len, 0);
1816 } else {
1817 __pskb_trim_head(skb, copy);
1818 tcp_set_skb_tso_segs(sk, skb, mss_now);
1819 }
1820 TCP_SKB_CB(skb)->seq += copy;
1821 }
1822
1823 len += copy;
1824
1825 if (len >= probe_size)
1826 break;
1827 }
1828 tcp_init_tso_segs(sk, nskb, nskb->len);
1829
1830 /* We're ready to send. If this fails, the probe will
1831 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1832 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1833 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1834 /* Decrement cwnd here because we are sending
1835 * effectively two packets. */
1836 tp->snd_cwnd--;
1837 tcp_event_new_data_sent(sk, nskb);
1838
1839 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1840 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1841 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1842
1843 return 1;
1844 }
1845
1846 return -1;
1847}
1848
1849/* This routine writes packets to the network. It advances the
1850 * send_head. This happens as incoming acks open up the remote
1851 * window for us.
1852 *
1853 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1854 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1855 * account rare use of URG, this is not a big flaw.
1856 *
1857 * Send at most one packet when push_one > 0. Temporarily ignore
1858 * cwnd limit to force at most one packet out when push_one == 2.
1859
1860 * Returns true, if no segments are in flight and we have queued segments,
1861 * but cannot send anything now because of SWS or another problem.
1862 */
1863static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1864 int push_one, gfp_t gfp)
1865{
1866 struct tcp_sock *tp = tcp_sk(sk);
1867 struct sk_buff *skb;
1868 unsigned int tso_segs, sent_pkts;
1869 int cwnd_quota;
1870 int result;
1871
1872 sent_pkts = 0;
1873
1874 if (!push_one) {
1875 /* Do MTU probing. */
1876 result = tcp_mtu_probe(sk);
1877 if (!result) {
1878 return false;
1879 } else if (result > 0) {
1880 sent_pkts = 1;
1881 }
1882 }
1883
1884 while ((skb = tcp_send_head(sk))) {
1885 unsigned int limit;
1886
1887 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1888 BUG_ON(!tso_segs);
1889
1890 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE)
1891 goto repair; /* Skip network transmission */
1892
1893 cwnd_quota = tcp_cwnd_test(tp, skb);
1894 if (!cwnd_quota) {
1895 if (push_one == 2)
1896 /* Force out a loss probe pkt. */
1897 cwnd_quota = 1;
1898 else
1899 break;
1900 }
1901
1902 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
1903 break;
1904
1905 if (tso_segs == 1) {
1906 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
1907 (tcp_skb_is_last(sk, skb) ?
1908 nonagle : TCP_NAGLE_PUSH))))
1909 break;
1910 } else {
1911 if (!push_one && tcp_tso_should_defer(sk, skb))
1912 break;
1913 }
1914
1915 /* TCP Small Queues :
1916 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
1917 * This allows for :
1918 * - better RTT estimation and ACK scheduling
1919 * - faster recovery
1920 * - high rates
1921 * Alas, some drivers / subsystems require a fair amount
1922 * of queued bytes to ensure line rate.
1923 * One example is wifi aggregation (802.11 AMPDU)
1924 */
1925 limit = max_t(unsigned int, sysctl_tcp_limit_output_bytes,
1926 sk->sk_pacing_rate >> 10);
1927
1928 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
1929 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
1930 /* It is possible TX completion already happened
1931 * before we set TSQ_THROTTLED, so we must
1932 * test again the condition.
1933 * We abuse smp_mb__after_clear_bit() because
1934 * there is no smp_mb__after_set_bit() yet
1935 */
1936 smp_mb__after_clear_bit();
1937 if (atomic_read(&sk->sk_wmem_alloc) > limit)
1938 break;
1939 }
1940
1941 limit = mss_now;
1942 if (tso_segs > 1 && !tcp_urg_mode(tp))
1943 limit = tcp_mss_split_point(sk, skb, mss_now,
1944 min_t(unsigned int,
1945 cwnd_quota,
1946 sk->sk_gso_max_segs),
1947 nonagle);
1948
1949 if (skb->len > limit &&
1950 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
1951 break;
1952
1953 TCP_SKB_CB(skb)->when = tcp_time_stamp;
1954
1955 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
1956 break;
1957
1958repair:
1959 /* Advance the send_head. This one is sent out.
1960 * This call will increment packets_out.
1961 */
1962 tcp_event_new_data_sent(sk, skb);
1963
1964 tcp_minshall_update(tp, mss_now, skb);
1965 sent_pkts += tcp_skb_pcount(skb);
1966
1967 if (push_one)
1968 break;
1969 }
1970
1971 if (likely(sent_pkts)) {
1972 if (tcp_in_cwnd_reduction(sk))
1973 tp->prr_out += sent_pkts;
1974
1975 /* Send one loss probe per tail loss episode. */
1976 if (push_one != 2)
1977 tcp_schedule_loss_probe(sk);
1978 tcp_cwnd_validate(sk);
1979 return false;
1980 }
1981 return (push_one == 2) || (!tp->packets_out && tcp_send_head(sk));
1982}
1983
1984bool tcp_schedule_loss_probe(struct sock *sk)
1985{
1986 struct inet_connection_sock *icsk = inet_csk(sk);
1987 struct tcp_sock *tp = tcp_sk(sk);
1988 u32 timeout, tlp_time_stamp, rto_time_stamp;
1989 u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
1990
1991 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
1992 return false;
1993 /* No consecutive loss probes. */
1994 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
1995 tcp_rearm_rto(sk);
1996 return false;
1997 }
1998 /* Don't do any loss probe on a Fast Open connection before 3WHS
1999 * finishes.
2000 */
2001 if (sk->sk_state == TCP_SYN_RECV)
2002 return false;
2003
2004 /* TLP is only scheduled when next timer event is RTO. */
2005 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2006 return false;
2007
2008 /* Schedule a loss probe in 2*RTT for SACK capable connections
2009 * in Open state, that are either limited by cwnd or application.
2010 */
2011 if (sysctl_tcp_early_retrans < 3 || !tp->srtt_us || !tp->packets_out ||
2012 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2013 return false;
2014
2015 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2016 tcp_send_head(sk))
2017 return false;
2018
2019 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2020 * for delayed ack when there's one outstanding packet.
2021 */
2022 timeout = rtt << 1;
2023 if (tp->packets_out == 1)
2024 timeout = max_t(u32, timeout,
2025 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2026 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2027
2028 /* If RTO is shorter, just schedule TLP in its place. */
2029 tlp_time_stamp = tcp_time_stamp + timeout;
2030 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2031 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2032 s32 delta = rto_time_stamp - tcp_time_stamp;
2033 if (delta > 0)
2034 timeout = delta;
2035 }
2036
2037 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2038 TCP_RTO_MAX);
2039 return true;
2040}
2041
2042/* When probe timeout (PTO) fires, send a new segment if one exists, else
2043 * retransmit the last segment.
2044 */
2045void tcp_send_loss_probe(struct sock *sk)
2046{
2047 struct tcp_sock *tp = tcp_sk(sk);
2048 struct sk_buff *skb;
2049 int pcount;
2050 int mss = tcp_current_mss(sk);
2051 int err = -1;
2052
2053 if (tcp_send_head(sk) != NULL) {
2054 err = tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2055 goto rearm_timer;
2056 }
2057
2058 /* At most one outstanding TLP retransmission. */
2059 if (tp->tlp_high_seq)
2060 goto rearm_timer;
2061
2062 /* Retransmit last segment. */
2063 skb = tcp_write_queue_tail(sk);
2064 if (WARN_ON(!skb))
2065 goto rearm_timer;
2066
2067 pcount = tcp_skb_pcount(skb);
2068 if (WARN_ON(!pcount))
2069 goto rearm_timer;
2070
2071 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2072 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss)))
2073 goto rearm_timer;
2074 skb = tcp_write_queue_tail(sk);
2075 }
2076
2077 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2078 goto rearm_timer;
2079
2080 /* Probe with zero data doesn't trigger fast recovery. */
2081 if (skb->len > 0)
2082 err = __tcp_retransmit_skb(sk, skb);
2083
2084 /* Record snd_nxt for loss detection. */
2085 if (likely(!err))
2086 tp->tlp_high_seq = tp->snd_nxt;
2087
2088rearm_timer:
2089 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2090 inet_csk(sk)->icsk_rto,
2091 TCP_RTO_MAX);
2092
2093 if (likely(!err))
2094 NET_INC_STATS_BH(sock_net(sk),
2095 LINUX_MIB_TCPLOSSPROBES);
2096}
2097
2098/* Push out any pending frames which were held back due to
2099 * TCP_CORK or attempt at coalescing tiny packets.
2100 * The socket must be locked by the caller.
2101 */
2102void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2103 int nonagle)
2104{
2105 /* If we are closed, the bytes will have to remain here.
2106 * In time closedown will finish, we empty the write queue and
2107 * all will be happy.
2108 */
2109 if (unlikely(sk->sk_state == TCP_CLOSE))
2110 return;
2111
2112 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2113 sk_gfp_atomic(sk, GFP_ATOMIC)))
2114 tcp_check_probe_timer(sk);
2115}
2116
2117/* Send _single_ skb sitting at the send head. This function requires
2118 * true push pending frames to setup probe timer etc.
2119 */
2120void tcp_push_one(struct sock *sk, unsigned int mss_now)
2121{
2122 struct sk_buff *skb = tcp_send_head(sk);
2123
2124 BUG_ON(!skb || skb->len < mss_now);
2125
2126 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2127}
2128
2129/* This function returns the amount that we can raise the
2130 * usable window based on the following constraints
2131 *
2132 * 1. The window can never be shrunk once it is offered (RFC 793)
2133 * 2. We limit memory per socket
2134 *
2135 * RFC 1122:
2136 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2137 * RECV.NEXT + RCV.WIN fixed until:
2138 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2139 *
2140 * i.e. don't raise the right edge of the window until you can raise
2141 * it at least MSS bytes.
2142 *
2143 * Unfortunately, the recommended algorithm breaks header prediction,
2144 * since header prediction assumes th->window stays fixed.
2145 *
2146 * Strictly speaking, keeping th->window fixed violates the receiver
2147 * side SWS prevention criteria. The problem is that under this rule
2148 * a stream of single byte packets will cause the right side of the
2149 * window to always advance by a single byte.
2150 *
2151 * Of course, if the sender implements sender side SWS prevention
2152 * then this will not be a problem.
2153 *
2154 * BSD seems to make the following compromise:
2155 *
2156 * If the free space is less than the 1/4 of the maximum
2157 * space available and the free space is less than 1/2 mss,
2158 * then set the window to 0.
2159 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2160 * Otherwise, just prevent the window from shrinking
2161 * and from being larger than the largest representable value.
2162 *
2163 * This prevents incremental opening of the window in the regime
2164 * where TCP is limited by the speed of the reader side taking
2165 * data out of the TCP receive queue. It does nothing about
2166 * those cases where the window is constrained on the sender side
2167 * because the pipeline is full.
2168 *
2169 * BSD also seems to "accidentally" limit itself to windows that are a
2170 * multiple of MSS, at least until the free space gets quite small.
2171 * This would appear to be a side effect of the mbuf implementation.
2172 * Combining these two algorithms results in the observed behavior
2173 * of having a fixed window size at almost all times.
2174 *
2175 * Below we obtain similar behavior by forcing the offered window to
2176 * a multiple of the mss when it is feasible to do so.
2177 *
2178 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2179 * Regular options like TIMESTAMP are taken into account.
2180 */
2181u32 __tcp_select_window(struct sock *sk)
2182{
2183 struct inet_connection_sock *icsk = inet_csk(sk);
2184 struct tcp_sock *tp = tcp_sk(sk);
2185 /* MSS for the peer's data. Previous versions used mss_clamp
2186 * here. I don't know if the value based on our guesses
2187 * of peer's MSS is better for the performance. It's more correct
2188 * but may be worse for the performance because of rcv_mss
2189 * fluctuations. --SAW 1998/11/1
2190 */
2191 int mss = icsk->icsk_ack.rcv_mss;
2192 int free_space = tcp_space(sk);
2193 int allowed_space = tcp_full_space(sk);
2194 int full_space = min_t(int, tp->window_clamp, allowed_space);
2195 int window;
2196
2197 if (mss > full_space)
2198 mss = full_space;
2199
2200 if (free_space < (full_space >> 1)) {
2201 icsk->icsk_ack.quick = 0;
2202
2203 if (sk_under_memory_pressure(sk))
2204 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2205 4U * tp->advmss);
2206
2207 /* free_space might become our new window, make sure we don't
2208 * increase it due to wscale.
2209 */
2210 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2211
2212 /* if free space is less than mss estimate, or is below 1/16th
2213 * of the maximum allowed, try to move to zero-window, else
2214 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2215 * new incoming data is dropped due to memory limits.
2216 * With large window, mss test triggers way too late in order
2217 * to announce zero window in time before rmem limit kicks in.
2218 */
2219 if (free_space < (allowed_space >> 4) || free_space < mss)
2220 return 0;
2221 }
2222
2223 if (free_space > tp->rcv_ssthresh)
2224 free_space = tp->rcv_ssthresh;
2225
2226 /* Don't do rounding if we are using window scaling, since the
2227 * scaled window will not line up with the MSS boundary anyway.
2228 */
2229 window = tp->rcv_wnd;
2230 if (tp->rx_opt.rcv_wscale) {
2231 window = free_space;
2232
2233 /* Advertise enough space so that it won't get scaled away.
2234 * Import case: prevent zero window announcement if
2235 * 1<<rcv_wscale > mss.
2236 */
2237 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2238 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2239 << tp->rx_opt.rcv_wscale);
2240 } else {
2241 /* Get the largest window that is a nice multiple of mss.
2242 * Window clamp already applied above.
2243 * If our current window offering is within 1 mss of the
2244 * free space we just keep it. This prevents the divide
2245 * and multiply from happening most of the time.
2246 * We also don't do any window rounding when the free space
2247 * is too small.
2248 */
2249 if (window <= free_space - mss || window > free_space)
2250 window = (free_space / mss) * mss;
2251 else if (mss == full_space &&
2252 free_space > window + (full_space >> 1))
2253 window = free_space;
2254 }
2255
2256 return window;
2257}
2258
2259/* Collapses two adjacent SKB's during retransmission. */
2260static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2261{
2262 struct tcp_sock *tp = tcp_sk(sk);
2263 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2264 int skb_size, next_skb_size;
2265
2266 skb_size = skb->len;
2267 next_skb_size = next_skb->len;
2268
2269 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2270
2271 tcp_highest_sack_combine(sk, next_skb, skb);
2272
2273 tcp_unlink_write_queue(next_skb, sk);
2274
2275 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2276 next_skb_size);
2277
2278 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2279 skb->ip_summed = CHECKSUM_PARTIAL;
2280
2281 if (skb->ip_summed != CHECKSUM_PARTIAL)
2282 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2283
2284 /* Update sequence range on original skb. */
2285 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2286
2287 /* Merge over control information. This moves PSH/FIN etc. over */
2288 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2289
2290 /* All done, get rid of second SKB and account for it so
2291 * packet counting does not break.
2292 */
2293 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2294
2295 /* changed transmit queue under us so clear hints */
2296 tcp_clear_retrans_hints_partial(tp);
2297 if (next_skb == tp->retransmit_skb_hint)
2298 tp->retransmit_skb_hint = skb;
2299
2300 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2301
2302 sk_wmem_free_skb(sk, next_skb);
2303}
2304
2305/* Check if coalescing SKBs is legal. */
2306static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2307{
2308 if (tcp_skb_pcount(skb) > 1)
2309 return false;
2310 /* TODO: SACK collapsing could be used to remove this condition */
2311 if (skb_shinfo(skb)->nr_frags != 0)
2312 return false;
2313 if (skb_cloned(skb))
2314 return false;
2315 if (skb == tcp_send_head(sk))
2316 return false;
2317 /* Some heurestics for collapsing over SACK'd could be invented */
2318 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2319 return false;
2320
2321 return true;
2322}
2323
2324/* Collapse packets in the retransmit queue to make to create
2325 * less packets on the wire. This is only done on retransmission.
2326 */
2327static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2328 int space)
2329{
2330 struct tcp_sock *tp = tcp_sk(sk);
2331 struct sk_buff *skb = to, *tmp;
2332 bool first = true;
2333
2334 if (!sysctl_tcp_retrans_collapse)
2335 return;
2336 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2337 return;
2338
2339 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2340 if (!tcp_can_collapse(sk, skb))
2341 break;
2342
2343 space -= skb->len;
2344
2345 if (first) {
2346 first = false;
2347 continue;
2348 }
2349
2350 if (space < 0)
2351 break;
2352 /* Punt if not enough space exists in the first SKB for
2353 * the data in the second
2354 */
2355 if (skb->len > skb_availroom(to))
2356 break;
2357
2358 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2359 break;
2360
2361 tcp_collapse_retrans(sk, to);
2362 }
2363}
2364
2365/* This retransmits one SKB. Policy decisions and retransmit queue
2366 * state updates are done by the caller. Returns non-zero if an
2367 * error occurred which prevented the send.
2368 */
2369int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2370{
2371 struct tcp_sock *tp = tcp_sk(sk);
2372 struct inet_connection_sock *icsk = inet_csk(sk);
2373 unsigned int cur_mss;
2374 int err;
2375
2376 /* Inconslusive MTU probe */
2377 if (icsk->icsk_mtup.probe_size) {
2378 icsk->icsk_mtup.probe_size = 0;
2379 }
2380
2381 /* Do not sent more than we queued. 1/4 is reserved for possible
2382 * copying overhead: fragmentation, tunneling, mangling etc.
2383 */
2384 if (atomic_read(&sk->sk_wmem_alloc) >
2385 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2386 return -EAGAIN;
2387
2388 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2389 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2390 BUG();
2391 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2392 return -ENOMEM;
2393 }
2394
2395 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2396 return -EHOSTUNREACH; /* Routing failure or similar. */
2397
2398 cur_mss = tcp_current_mss(sk);
2399
2400 /* If receiver has shrunk his window, and skb is out of
2401 * new window, do not retransmit it. The exception is the
2402 * case, when window is shrunk to zero. In this case
2403 * our retransmit serves as a zero window probe.
2404 */
2405 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2406 TCP_SKB_CB(skb)->seq != tp->snd_una)
2407 return -EAGAIN;
2408
2409 if (skb->len > cur_mss) {
2410 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2411 return -ENOMEM; /* We'll try again later. */
2412 } else {
2413 int oldpcount = tcp_skb_pcount(skb);
2414
2415 if (unlikely(oldpcount > 1)) {
2416 if (skb_unclone(skb, GFP_ATOMIC))
2417 return -ENOMEM;
2418 tcp_init_tso_segs(sk, skb, cur_mss);
2419 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2420 }
2421 }
2422
2423 tcp_retrans_try_collapse(sk, skb, cur_mss);
2424
2425 /* Make a copy, if the first transmission SKB clone we made
2426 * is still in somebody's hands, else make a clone.
2427 */
2428 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2429
2430 /* make sure skb->data is aligned on arches that require it
2431 * and check if ack-trimming & collapsing extended the headroom
2432 * beyond what csum_start can cover.
2433 */
2434 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2435 skb_headroom(skb) >= 0xFFFF)) {
2436 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2437 GFP_ATOMIC);
2438 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2439 -ENOBUFS;
2440 } else {
2441 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2442 }
2443
2444 if (likely(!err)) {
2445 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2446 /* Update global TCP statistics. */
2447 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2448 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2449 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2450 tp->total_retrans++;
2451 }
2452 return err;
2453}
2454
2455int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2456{
2457 struct tcp_sock *tp = tcp_sk(sk);
2458 int err = __tcp_retransmit_skb(sk, skb);
2459
2460 if (err == 0) {
2461#if FASTRETRANS_DEBUG > 0
2462 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2463 net_dbg_ratelimited("retrans_out leaked\n");
2464 }
2465#endif
2466 if (!tp->retrans_out)
2467 tp->lost_retrans_low = tp->snd_nxt;
2468 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2469 tp->retrans_out += tcp_skb_pcount(skb);
2470
2471 /* Save stamp of the first retransmit. */
2472 if (!tp->retrans_stamp)
2473 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2474
2475 tp->undo_retrans += tcp_skb_pcount(skb);
2476
2477 /* snd_nxt is stored to detect loss of retransmitted segment,
2478 * see tcp_input.c tcp_sacktag_write_queue().
2479 */
2480 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2481 } else {
2482 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2483 }
2484 return err;
2485}
2486
2487/* Check if we forward retransmits are possible in the current
2488 * window/congestion state.
2489 */
2490static bool tcp_can_forward_retransmit(struct sock *sk)
2491{
2492 const struct inet_connection_sock *icsk = inet_csk(sk);
2493 const struct tcp_sock *tp = tcp_sk(sk);
2494
2495 /* Forward retransmissions are possible only during Recovery. */
2496 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2497 return false;
2498
2499 /* No forward retransmissions in Reno are possible. */
2500 if (tcp_is_reno(tp))
2501 return false;
2502
2503 /* Yeah, we have to make difficult choice between forward transmission
2504 * and retransmission... Both ways have their merits...
2505 *
2506 * For now we do not retransmit anything, while we have some new
2507 * segments to send. In the other cases, follow rule 3 for
2508 * NextSeg() specified in RFC3517.
2509 */
2510
2511 if (tcp_may_send_now(sk))
2512 return false;
2513
2514 return true;
2515}
2516
2517/* This gets called after a retransmit timeout, and the initially
2518 * retransmitted data is acknowledged. It tries to continue
2519 * resending the rest of the retransmit queue, until either
2520 * we've sent it all or the congestion window limit is reached.
2521 * If doing SACK, the first ACK which comes back for a timeout
2522 * based retransmit packet might feed us FACK information again.
2523 * If so, we use it to avoid unnecessarily retransmissions.
2524 */
2525void tcp_xmit_retransmit_queue(struct sock *sk)
2526{
2527 const struct inet_connection_sock *icsk = inet_csk(sk);
2528 struct tcp_sock *tp = tcp_sk(sk);
2529 struct sk_buff *skb;
2530 struct sk_buff *hole = NULL;
2531 u32 last_lost;
2532 int mib_idx;
2533 int fwd_rexmitting = 0;
2534
2535 if (!tp->packets_out)
2536 return;
2537
2538 if (!tp->lost_out)
2539 tp->retransmit_high = tp->snd_una;
2540
2541 if (tp->retransmit_skb_hint) {
2542 skb = tp->retransmit_skb_hint;
2543 last_lost = TCP_SKB_CB(skb)->end_seq;
2544 if (after(last_lost, tp->retransmit_high))
2545 last_lost = tp->retransmit_high;
2546 } else {
2547 skb = tcp_write_queue_head(sk);
2548 last_lost = tp->snd_una;
2549 }
2550
2551 tcp_for_write_queue_from(skb, sk) {
2552 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2553
2554 if (skb == tcp_send_head(sk))
2555 break;
2556 /* we could do better than to assign each time */
2557 if (hole == NULL)
2558 tp->retransmit_skb_hint = skb;
2559
2560 /* Assume this retransmit will generate
2561 * only one packet for congestion window
2562 * calculation purposes. This works because
2563 * tcp_retransmit_skb() will chop up the
2564 * packet to be MSS sized and all the
2565 * packet counting works out.
2566 */
2567 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2568 return;
2569
2570 if (fwd_rexmitting) {
2571begin_fwd:
2572 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2573 break;
2574 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2575
2576 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2577 tp->retransmit_high = last_lost;
2578 if (!tcp_can_forward_retransmit(sk))
2579 break;
2580 /* Backtrack if necessary to non-L'ed skb */
2581 if (hole != NULL) {
2582 skb = hole;
2583 hole = NULL;
2584 }
2585 fwd_rexmitting = 1;
2586 goto begin_fwd;
2587
2588 } else if (!(sacked & TCPCB_LOST)) {
2589 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2590 hole = skb;
2591 continue;
2592
2593 } else {
2594 last_lost = TCP_SKB_CB(skb)->end_seq;
2595 if (icsk->icsk_ca_state != TCP_CA_Loss)
2596 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2597 else
2598 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2599 }
2600
2601 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2602 continue;
2603
2604 if (tcp_retransmit_skb(sk, skb))
2605 return;
2606
2607 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2608
2609 if (tcp_in_cwnd_reduction(sk))
2610 tp->prr_out += tcp_skb_pcount(skb);
2611
2612 if (skb == tcp_write_queue_head(sk))
2613 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2614 inet_csk(sk)->icsk_rto,
2615 TCP_RTO_MAX);
2616 }
2617}
2618
2619/* Send a fin. The caller locks the socket for us. This cannot be
2620 * allowed to fail queueing a FIN frame under any circumstances.
2621 */
2622void tcp_send_fin(struct sock *sk)
2623{
2624 struct tcp_sock *tp = tcp_sk(sk);
2625 struct sk_buff *skb = tcp_write_queue_tail(sk);
2626 int mss_now;
2627
2628 /* Optimization, tack on the FIN if we have a queue of
2629 * unsent frames. But be careful about outgoing SACKS
2630 * and IP options.
2631 */
2632 mss_now = tcp_current_mss(sk);
2633
2634 if (tcp_send_head(sk) != NULL) {
2635 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2636 TCP_SKB_CB(skb)->end_seq++;
2637 tp->write_seq++;
2638 } else {
2639 /* Socket is locked, keep trying until memory is available. */
2640 for (;;) {
2641 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2642 sk->sk_allocation);
2643 if (skb)
2644 break;
2645 yield();
2646 }
2647
2648 /* Reserve space for headers and prepare control bits. */
2649 skb_reserve(skb, MAX_TCP_HEADER);
2650 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2651 tcp_init_nondata_skb(skb, tp->write_seq,
2652 TCPHDR_ACK | TCPHDR_FIN);
2653 tcp_queue_skb(sk, skb);
2654 }
2655 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2656}
2657
2658/* We get here when a process closes a file descriptor (either due to
2659 * an explicit close() or as a byproduct of exit()'ing) and there
2660 * was unread data in the receive queue. This behavior is recommended
2661 * by RFC 2525, section 2.17. -DaveM
2662 */
2663void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2664{
2665 struct sk_buff *skb;
2666
2667 /* NOTE: No TCP options attached and we never retransmit this. */
2668 skb = alloc_skb(MAX_TCP_HEADER, priority);
2669 if (!skb) {
2670 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2671 return;
2672 }
2673
2674 /* Reserve space for headers and prepare control bits. */
2675 skb_reserve(skb, MAX_TCP_HEADER);
2676 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2677 TCPHDR_ACK | TCPHDR_RST);
2678 /* Send it off. */
2679 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2680 if (tcp_transmit_skb(sk, skb, 0, priority))
2681 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2682
2683 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2684}
2685
2686/* Send a crossed SYN-ACK during socket establishment.
2687 * WARNING: This routine must only be called when we have already sent
2688 * a SYN packet that crossed the incoming SYN that caused this routine
2689 * to get called. If this assumption fails then the initial rcv_wnd
2690 * and rcv_wscale values will not be correct.
2691 */
2692int tcp_send_synack(struct sock *sk)
2693{
2694 struct sk_buff *skb;
2695
2696 skb = tcp_write_queue_head(sk);
2697 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2698 pr_debug("%s: wrong queue state\n", __func__);
2699 return -EFAULT;
2700 }
2701 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2702 if (skb_cloned(skb)) {
2703 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2704 if (nskb == NULL)
2705 return -ENOMEM;
2706 tcp_unlink_write_queue(skb, sk);
2707 skb_header_release(nskb);
2708 __tcp_add_write_queue_head(sk, nskb);
2709 sk_wmem_free_skb(sk, skb);
2710 sk->sk_wmem_queued += nskb->truesize;
2711 sk_mem_charge(sk, nskb->truesize);
2712 skb = nskb;
2713 }
2714
2715 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2716 TCP_ECN_send_synack(tcp_sk(sk), skb);
2717 }
2718 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2719 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2720}
2721
2722/**
2723 * tcp_make_synack - Prepare a SYN-ACK.
2724 * sk: listener socket
2725 * dst: dst entry attached to the SYNACK
2726 * req: request_sock pointer
2727 *
2728 * Allocate one skb and build a SYNACK packet.
2729 * @dst is consumed : Caller should not use it again.
2730 */
2731struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2732 struct request_sock *req,
2733 struct tcp_fastopen_cookie *foc)
2734{
2735 struct tcp_out_options opts;
2736 struct inet_request_sock *ireq = inet_rsk(req);
2737 struct tcp_sock *tp = tcp_sk(sk);
2738 struct tcphdr *th;
2739 struct sk_buff *skb;
2740 struct tcp_md5sig_key *md5;
2741 int tcp_header_size;
2742 int mss;
2743
2744 skb = sock_wmalloc(sk, MAX_TCP_HEADER, 1, GFP_ATOMIC);
2745 if (unlikely(!skb)) {
2746 dst_release(dst);
2747 return NULL;
2748 }
2749 /* Reserve space for headers. */
2750 skb_reserve(skb, MAX_TCP_HEADER);
2751
2752 skb_dst_set(skb, dst);
2753 security_skb_owned_by(skb, sk);
2754
2755 mss = dst_metric_advmss(dst);
2756 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2757 mss = tp->rx_opt.user_mss;
2758
2759 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2760 __u8 rcv_wscale;
2761 /* Set this up on the first call only */
2762 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2763
2764 /* limit the window selection if the user enforce a smaller rx buffer */
2765 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2766 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2767 req->window_clamp = tcp_full_space(sk);
2768
2769 /* tcp_full_space because it is guaranteed to be the first packet */
2770 tcp_select_initial_window(tcp_full_space(sk),
2771 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2772 &req->rcv_wnd,
2773 &req->window_clamp,
2774 ireq->wscale_ok,
2775 &rcv_wscale,
2776 dst_metric(dst, RTAX_INITRWND));
2777 ireq->rcv_wscale = rcv_wscale;
2778 }
2779
2780 memset(&opts, 0, sizeof(opts));
2781#ifdef CONFIG_SYN_COOKIES
2782 if (unlikely(req->cookie_ts))
2783 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2784 else
2785#endif
2786 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2787 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, &md5,
2788 foc) + sizeof(*th);
2789
2790 skb_push(skb, tcp_header_size);
2791 skb_reset_transport_header(skb);
2792
2793 th = tcp_hdr(skb);
2794 memset(th, 0, sizeof(struct tcphdr));
2795 th->syn = 1;
2796 th->ack = 1;
2797 TCP_ECN_make_synack(req, th);
2798 th->source = htons(ireq->ir_num);
2799 th->dest = ireq->ir_rmt_port;
2800 /* Setting of flags are superfluous here for callers (and ECE is
2801 * not even correctly set)
2802 */
2803 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2804 TCPHDR_SYN | TCPHDR_ACK);
2805
2806 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2807 /* XXX data is queued and acked as is. No buffer/window check */
2808 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
2809
2810 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2811 th->window = htons(min(req->rcv_wnd, 65535U));
2812 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2813 th->doff = (tcp_header_size >> 2);
2814 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS);
2815
2816#ifdef CONFIG_TCP_MD5SIG
2817 /* Okay, we have all we need - do the md5 hash if needed */
2818 if (md5) {
2819 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2820 md5, NULL, req, skb);
2821 }
2822#endif
2823
2824 return skb;
2825}
2826EXPORT_SYMBOL(tcp_make_synack);
2827
2828/* Do all connect socket setups that can be done AF independent. */
2829static void tcp_connect_init(struct sock *sk)
2830{
2831 const struct dst_entry *dst = __sk_dst_get(sk);
2832 struct tcp_sock *tp = tcp_sk(sk);
2833 __u8 rcv_wscale;
2834
2835 /* We'll fix this up when we get a response from the other end.
2836 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2837 */
2838 tp->tcp_header_len = sizeof(struct tcphdr) +
2839 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2840
2841#ifdef CONFIG_TCP_MD5SIG
2842 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2843 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2844#endif
2845
2846 /* If user gave his TCP_MAXSEG, record it to clamp */
2847 if (tp->rx_opt.user_mss)
2848 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2849 tp->max_window = 0;
2850 tcp_mtup_init(sk);
2851 tcp_sync_mss(sk, dst_mtu(dst));
2852
2853 if (!tp->window_clamp)
2854 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2855 tp->advmss = dst_metric_advmss(dst);
2856 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2857 tp->advmss = tp->rx_opt.user_mss;
2858
2859 tcp_initialize_rcv_mss(sk);
2860
2861 /* limit the window selection if the user enforce a smaller rx buffer */
2862 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2863 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2864 tp->window_clamp = tcp_full_space(sk);
2865
2866 tcp_select_initial_window(tcp_full_space(sk),
2867 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2868 &tp->rcv_wnd,
2869 &tp->window_clamp,
2870 sysctl_tcp_window_scaling,
2871 &rcv_wscale,
2872 dst_metric(dst, RTAX_INITRWND));
2873
2874 tp->rx_opt.rcv_wscale = rcv_wscale;
2875 tp->rcv_ssthresh = tp->rcv_wnd;
2876
2877 sk->sk_err = 0;
2878 sock_reset_flag(sk, SOCK_DONE);
2879 tp->snd_wnd = 0;
2880 tcp_init_wl(tp, 0);
2881 tp->snd_una = tp->write_seq;
2882 tp->snd_sml = tp->write_seq;
2883 tp->snd_up = tp->write_seq;
2884 tp->snd_nxt = tp->write_seq;
2885
2886 if (likely(!tp->repair))
2887 tp->rcv_nxt = 0;
2888 else
2889 tp->rcv_tstamp = tcp_time_stamp;
2890 tp->rcv_wup = tp->rcv_nxt;
2891 tp->copied_seq = tp->rcv_nxt;
2892
2893 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2894 inet_csk(sk)->icsk_retransmits = 0;
2895 tcp_clear_retrans(tp);
2896}
2897
2898static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
2899{
2900 struct tcp_sock *tp = tcp_sk(sk);
2901 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
2902
2903 tcb->end_seq += skb->len;
2904 skb_header_release(skb);
2905 __tcp_add_write_queue_tail(sk, skb);
2906 sk->sk_wmem_queued += skb->truesize;
2907 sk_mem_charge(sk, skb->truesize);
2908 tp->write_seq = tcb->end_seq;
2909 tp->packets_out += tcp_skb_pcount(skb);
2910}
2911
2912/* Build and send a SYN with data and (cached) Fast Open cookie. However,
2913 * queue a data-only packet after the regular SYN, such that regular SYNs
2914 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2915 * only the SYN sequence, the data are retransmitted in the first ACK.
2916 * If cookie is not cached or other error occurs, falls back to send a
2917 * regular SYN with Fast Open cookie request option.
2918 */
2919static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
2920{
2921 struct tcp_sock *tp = tcp_sk(sk);
2922 struct tcp_fastopen_request *fo = tp->fastopen_req;
2923 int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
2924 struct sk_buff *syn_data = NULL, *data;
2925 unsigned long last_syn_loss = 0;
2926
2927 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
2928 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
2929 &syn_loss, &last_syn_loss);
2930 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2931 if (syn_loss > 1 &&
2932 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
2933 fo->cookie.len = -1;
2934 goto fallback;
2935 }
2936
2937 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
2938 fo->cookie.len = -1;
2939 else if (fo->cookie.len <= 0)
2940 goto fallback;
2941
2942 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2943 * user-MSS. Reserve maximum option space for middleboxes that add
2944 * private TCP options. The cost is reduced data space in SYN :(
2945 */
2946 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
2947 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2948 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
2949 MAX_TCP_OPTION_SPACE;
2950
2951 space = min_t(size_t, space, fo->size);
2952
2953 /* limit to order-0 allocations */
2954 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
2955
2956 syn_data = skb_copy_expand(syn, MAX_TCP_HEADER, space,
2957 sk->sk_allocation);
2958 if (syn_data == NULL)
2959 goto fallback;
2960
2961 for (i = 0; i < iovlen && syn_data->len < space; ++i) {
2962 struct iovec *iov = &fo->data->msg_iov[i];
2963 unsigned char __user *from = iov->iov_base;
2964 int len = iov->iov_len;
2965
2966 if (syn_data->len + len > space)
2967 len = space - syn_data->len;
2968 else if (i + 1 == iovlen)
2969 /* No more data pending in inet_wait_for_connect() */
2970 fo->data = NULL;
2971
2972 if (skb_add_data(syn_data, from, len))
2973 goto fallback;
2974 }
2975
2976 /* Queue a data-only packet after the regular SYN for retransmission */
2977 data = pskb_copy(syn_data, sk->sk_allocation);
2978 if (data == NULL)
2979 goto fallback;
2980 TCP_SKB_CB(data)->seq++;
2981 TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
2982 TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
2983 tcp_connect_queue_skb(sk, data);
2984 fo->copied = data->len;
2985
2986 /* syn_data is about to be sent, we need to take current time stamps
2987 * for the packets that are in write queue : SYN packet and DATA
2988 */
2989 skb_mstamp_get(&syn->skb_mstamp);
2990 data->skb_mstamp = syn->skb_mstamp;
2991
2992 if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
2993 tp->syn_data = (fo->copied > 0);
2994 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
2995 goto done;
2996 }
2997 syn_data = NULL;
2998
2999fallback:
3000 /* Send a regular SYN with Fast Open cookie request option */
3001 if (fo->cookie.len > 0)
3002 fo->cookie.len = 0;
3003 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3004 if (err)
3005 tp->syn_fastopen = 0;
3006 kfree_skb(syn_data);
3007done:
3008 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3009 return err;
3010}
3011
3012/* Build a SYN and send it off. */
3013int tcp_connect(struct sock *sk)
3014{
3015 struct tcp_sock *tp = tcp_sk(sk);
3016 struct sk_buff *buff;
3017 int err;
3018
3019 tcp_connect_init(sk);
3020
3021 if (unlikely(tp->repair)) {
3022 tcp_finish_connect(sk, NULL);
3023 return 0;
3024 }
3025
3026 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
3027 if (unlikely(buff == NULL))
3028 return -ENOBUFS;
3029
3030 /* Reserve space for headers. */
3031 skb_reserve(buff, MAX_TCP_HEADER);
3032
3033 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3034 tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp;
3035 tcp_connect_queue_skb(sk, buff);
3036 TCP_ECN_send_syn(sk, buff);
3037
3038 /* Send off SYN; include data in Fast Open. */
3039 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3040 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3041 if (err == -ECONNREFUSED)
3042 return err;
3043
3044 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3045 * in order to make this packet get counted in tcpOutSegs.
3046 */
3047 tp->snd_nxt = tp->write_seq;
3048 tp->pushed_seq = tp->write_seq;
3049 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3050
3051 /* Timer for repeating the SYN until an answer. */
3052 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3053 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3054 return 0;
3055}
3056EXPORT_SYMBOL(tcp_connect);
3057
3058/* Send out a delayed ack, the caller does the policy checking
3059 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3060 * for details.
3061 */
3062void tcp_send_delayed_ack(struct sock *sk)
3063{
3064 struct inet_connection_sock *icsk = inet_csk(sk);
3065 int ato = icsk->icsk_ack.ato;
3066 unsigned long timeout;
3067
3068 if (ato > TCP_DELACK_MIN) {
3069 const struct tcp_sock *tp = tcp_sk(sk);
3070 int max_ato = HZ / 2;
3071
3072 if (icsk->icsk_ack.pingpong ||
3073 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3074 max_ato = TCP_DELACK_MAX;
3075
3076 /* Slow path, intersegment interval is "high". */
3077
3078 /* If some rtt estimate is known, use it to bound delayed ack.
3079 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3080 * directly.
3081 */
3082 if (tp->srtt_us) {
3083 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3084 TCP_DELACK_MIN);
3085
3086 if (rtt < max_ato)
3087 max_ato = rtt;
3088 }
3089
3090 ato = min(ato, max_ato);
3091 }
3092
3093 /* Stay within the limit we were given */
3094 timeout = jiffies + ato;
3095
3096 /* Use new timeout only if there wasn't a older one earlier. */
3097 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3098 /* If delack timer was blocked or is about to expire,
3099 * send ACK now.
3100 */
3101 if (icsk->icsk_ack.blocked ||
3102 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3103 tcp_send_ack(sk);
3104 return;
3105 }
3106
3107 if (!time_before(timeout, icsk->icsk_ack.timeout))
3108 timeout = icsk->icsk_ack.timeout;
3109 }
3110 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3111 icsk->icsk_ack.timeout = timeout;
3112 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3113}
3114
3115/* This routine sends an ack and also updates the window. */
3116void tcp_send_ack(struct sock *sk)
3117{
3118 struct sk_buff *buff;
3119
3120 /* If we have been reset, we may not send again. */
3121 if (sk->sk_state == TCP_CLOSE)
3122 return;
3123
3124 /* We are not putting this on the write queue, so
3125 * tcp_transmit_skb() will set the ownership to this
3126 * sock.
3127 */
3128 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3129 if (buff == NULL) {
3130 inet_csk_schedule_ack(sk);
3131 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3132 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3133 TCP_DELACK_MAX, TCP_RTO_MAX);
3134 return;
3135 }
3136
3137 /* Reserve space for headers and prepare control bits. */
3138 skb_reserve(buff, MAX_TCP_HEADER);
3139 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3140
3141 /* Send it off, this clears delayed acks for us. */
3142 TCP_SKB_CB(buff)->when = tcp_time_stamp;
3143 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
3144}
3145
3146/* This routine sends a packet with an out of date sequence
3147 * number. It assumes the other end will try to ack it.
3148 *
3149 * Question: what should we make while urgent mode?
3150 * 4.4BSD forces sending single byte of data. We cannot send
3151 * out of window data, because we have SND.NXT==SND.MAX...
3152 *
3153 * Current solution: to send TWO zero-length segments in urgent mode:
3154 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3155 * out-of-date with SND.UNA-1 to probe window.
3156 */
3157static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3158{
3159 struct tcp_sock *tp = tcp_sk(sk);
3160 struct sk_buff *skb;
3161
3162 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3163 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3164 if (skb == NULL)
3165 return -1;
3166
3167 /* Reserve space for headers and set control bits. */
3168 skb_reserve(skb, MAX_TCP_HEADER);
3169 /* Use a previous sequence. This should cause the other
3170 * end to send an ack. Don't queue or clone SKB, just
3171 * send it.
3172 */
3173 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3174 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3175 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3176}
3177
3178void tcp_send_window_probe(struct sock *sk)
3179{
3180 if (sk->sk_state == TCP_ESTABLISHED) {
3181 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3182 tcp_xmit_probe_skb(sk, 0);
3183 }
3184}
3185
3186/* Initiate keepalive or window probe from timer. */
3187int tcp_write_wakeup(struct sock *sk)
3188{
3189 struct tcp_sock *tp = tcp_sk(sk);
3190 struct sk_buff *skb;
3191
3192 if (sk->sk_state == TCP_CLOSE)
3193 return -1;
3194
3195 if ((skb = tcp_send_head(sk)) != NULL &&
3196 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3197 int err;
3198 unsigned int mss = tcp_current_mss(sk);
3199 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3200
3201 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3202 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3203
3204 /* We are probing the opening of a window
3205 * but the window size is != 0
3206 * must have been a result SWS avoidance ( sender )
3207 */
3208 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3209 skb->len > mss) {
3210 seg_size = min(seg_size, mss);
3211 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3212 if (tcp_fragment(sk, skb, seg_size, mss))
3213 return -1;
3214 } else if (!tcp_skb_pcount(skb))
3215 tcp_set_skb_tso_segs(sk, skb, mss);
3216
3217 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3218 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3219 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3220 if (!err)
3221 tcp_event_new_data_sent(sk, skb);
3222 return err;
3223 } else {
3224 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3225 tcp_xmit_probe_skb(sk, 1);
3226 return tcp_xmit_probe_skb(sk, 0);
3227 }
3228}
3229
3230/* A window probe timeout has occurred. If window is not closed send
3231 * a partial packet else a zero probe.
3232 */
3233void tcp_send_probe0(struct sock *sk)
3234{
3235 struct inet_connection_sock *icsk = inet_csk(sk);
3236 struct tcp_sock *tp = tcp_sk(sk);
3237 int err;
3238
3239 err = tcp_write_wakeup(sk);
3240
3241 if (tp->packets_out || !tcp_send_head(sk)) {
3242 /* Cancel probe timer, if it is not required. */
3243 icsk->icsk_probes_out = 0;
3244 icsk->icsk_backoff = 0;
3245 return;
3246 }
3247
3248 if (err <= 0) {
3249 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3250 icsk->icsk_backoff++;
3251 icsk->icsk_probes_out++;
3252 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3253 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3254 TCP_RTO_MAX);
3255 } else {
3256 /* If packet was not sent due to local congestion,
3257 * do not backoff and do not remember icsk_probes_out.
3258 * Let local senders to fight for local resources.
3259 *
3260 * Use accumulated backoff yet.
3261 */
3262 if (!icsk->icsk_probes_out)
3263 icsk->icsk_probes_out = 1;
3264 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3265 min(icsk->icsk_rto << icsk->icsk_backoff,
3266 TCP_RESOURCE_PROBE_INTERVAL),
3267 TCP_RTO_MAX);
3268 }
3269}