Merge tag 'ux500-fixes-for-v3.9' of git://git.kernel.org/pub/scm/linux/kernel/git...
[GitHub/mt8127/android_kernel_alcatel_ttab.git] / net / ipv4 / tcp_output.c
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. */
46 int 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 */
51 int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
52
53 /* Default TSQ limit of two TSO segments */
54 int 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 */
60 int sysctl_tcp_tso_win_divisor __read_mostly = 3;
61
62 int sysctl_tcp_mtu_probing __read_mostly = 0;
63 int sysctl_tcp_base_mss __read_mostly = TCP_BASE_MSS;
64
65 /* By default, RFC2861 behavior. */
66 int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
67
68 int sysctl_tcp_cookie_size __read_mostly = 0; /* TCP_COOKIE_MAX */
69 EXPORT_SYMBOL_GPL(sysctl_tcp_cookie_size);
70
71 static 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. */
75 static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
76 {
77 struct tcp_sock *tp = tcp_sk(sk);
78 unsigned int prior_packets = tp->packets_out;
79
80 tcp_advance_send_head(sk, skb);
81 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
82
83 /* Don't override Nagle indefinitely with F-RTO */
84 if (tp->frto_counter == 2)
85 tp->frto_counter = 3;
86
87 tp->packets_out += tcp_skb_pcount(skb);
88 if (!prior_packets || tp->early_retrans_delayed)
89 tcp_rearm_rto(sk);
90 }
91
92 /* SND.NXT, if window was not shrunk.
93 * If window has been shrunk, what should we make? It is not clear at all.
94 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
95 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
96 * invalid. OK, let's make this for now:
97 */
98 static inline __u32 tcp_acceptable_seq(const struct sock *sk)
99 {
100 const struct tcp_sock *tp = tcp_sk(sk);
101
102 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
103 return tp->snd_nxt;
104 else
105 return tcp_wnd_end(tp);
106 }
107
108 /* Calculate mss to advertise in SYN segment.
109 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
110 *
111 * 1. It is independent of path mtu.
112 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
113 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
114 * attached devices, because some buggy hosts are confused by
115 * large MSS.
116 * 4. We do not make 3, we advertise MSS, calculated from first
117 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
118 * This may be overridden via information stored in routing table.
119 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
120 * probably even Jumbo".
121 */
122 static __u16 tcp_advertise_mss(struct sock *sk)
123 {
124 struct tcp_sock *tp = tcp_sk(sk);
125 const struct dst_entry *dst = __sk_dst_get(sk);
126 int mss = tp->advmss;
127
128 if (dst) {
129 unsigned int metric = dst_metric_advmss(dst);
130
131 if (metric < mss) {
132 mss = metric;
133 tp->advmss = mss;
134 }
135 }
136
137 return (__u16)mss;
138 }
139
140 /* RFC2861. Reset CWND after idle period longer RTO to "restart window".
141 * This is the first part of cwnd validation mechanism. */
142 static void tcp_cwnd_restart(struct sock *sk, const struct dst_entry *dst)
143 {
144 struct tcp_sock *tp = tcp_sk(sk);
145 s32 delta = tcp_time_stamp - tp->lsndtime;
146 u32 restart_cwnd = tcp_init_cwnd(tp, dst);
147 u32 cwnd = tp->snd_cwnd;
148
149 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
150
151 tp->snd_ssthresh = tcp_current_ssthresh(sk);
152 restart_cwnd = min(restart_cwnd, cwnd);
153
154 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
155 cwnd >>= 1;
156 tp->snd_cwnd = max(cwnd, restart_cwnd);
157 tp->snd_cwnd_stamp = tcp_time_stamp;
158 tp->snd_cwnd_used = 0;
159 }
160
161 /* Congestion state accounting after a packet has been sent. */
162 static void tcp_event_data_sent(struct tcp_sock *tp,
163 struct sock *sk)
164 {
165 struct inet_connection_sock *icsk = inet_csk(sk);
166 const u32 now = tcp_time_stamp;
167
168 if (sysctl_tcp_slow_start_after_idle &&
169 (!tp->packets_out && (s32)(now - tp->lsndtime) > icsk->icsk_rto))
170 tcp_cwnd_restart(sk, __sk_dst_get(sk));
171
172 tp->lsndtime = now;
173
174 /* If it is a reply for ato after last received
175 * packet, enter pingpong mode.
176 */
177 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
178 icsk->icsk_ack.pingpong = 1;
179 }
180
181 /* Account for an ACK we sent. */
182 static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
183 {
184 tcp_dec_quickack_mode(sk, pkts);
185 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
186 }
187
188 /* Determine a window scaling and initial window to offer.
189 * Based on the assumption that the given amount of space
190 * will be offered. Store the results in the tp structure.
191 * NOTE: for smooth operation initial space offering should
192 * be a multiple of mss if possible. We assume here that mss >= 1.
193 * This MUST be enforced by all callers.
194 */
195 void tcp_select_initial_window(int __space, __u32 mss,
196 __u32 *rcv_wnd, __u32 *window_clamp,
197 int wscale_ok, __u8 *rcv_wscale,
198 __u32 init_rcv_wnd)
199 {
200 unsigned int space = (__space < 0 ? 0 : __space);
201
202 /* If no clamp set the clamp to the max possible scaled window */
203 if (*window_clamp == 0)
204 (*window_clamp) = (65535 << 14);
205 space = min(*window_clamp, space);
206
207 /* Quantize space offering to a multiple of mss if possible. */
208 if (space > mss)
209 space = (space / mss) * mss;
210
211 /* NOTE: offering an initial window larger than 32767
212 * will break some buggy TCP stacks. If the admin tells us
213 * it is likely we could be speaking with such a buggy stack
214 * we will truncate our initial window offering to 32K-1
215 * unless the remote has sent us a window scaling option,
216 * which we interpret as a sign the remote TCP is not
217 * misinterpreting the window field as a signed quantity.
218 */
219 if (sysctl_tcp_workaround_signed_windows)
220 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
221 else
222 (*rcv_wnd) = space;
223
224 (*rcv_wscale) = 0;
225 if (wscale_ok) {
226 /* Set window scaling on max possible window
227 * See RFC1323 for an explanation of the limit to 14
228 */
229 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
230 space = min_t(u32, space, *window_clamp);
231 while (space > 65535 && (*rcv_wscale) < 14) {
232 space >>= 1;
233 (*rcv_wscale)++;
234 }
235 }
236
237 /* Set initial window to a value enough for senders starting with
238 * initial congestion window of TCP_DEFAULT_INIT_RCVWND. Place
239 * a limit on the initial window when mss is larger than 1460.
240 */
241 if (mss > (1 << *rcv_wscale)) {
242 int init_cwnd = TCP_DEFAULT_INIT_RCVWND;
243 if (mss > 1460)
244 init_cwnd =
245 max_t(u32, (1460 * TCP_DEFAULT_INIT_RCVWND) / mss, 2);
246 /* when initializing use the value from init_rcv_wnd
247 * rather than the default from above
248 */
249 if (init_rcv_wnd)
250 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
251 else
252 *rcv_wnd = min(*rcv_wnd, init_cwnd * mss);
253 }
254
255 /* Set the clamp no higher than max representable value */
256 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
257 }
258 EXPORT_SYMBOL(tcp_select_initial_window);
259
260 /* Chose a new window to advertise, update state in tcp_sock for the
261 * socket, and return result with RFC1323 scaling applied. The return
262 * value can be stuffed directly into th->window for an outgoing
263 * frame.
264 */
265 static u16 tcp_select_window(struct sock *sk)
266 {
267 struct tcp_sock *tp = tcp_sk(sk);
268 u32 cur_win = tcp_receive_window(tp);
269 u32 new_win = __tcp_select_window(sk);
270
271 /* Never shrink the offered window */
272 if (new_win < cur_win) {
273 /* Danger Will Robinson!
274 * Don't update rcv_wup/rcv_wnd here or else
275 * we will not be able to advertise a zero
276 * window in time. --DaveM
277 *
278 * Relax Will Robinson.
279 */
280 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
281 }
282 tp->rcv_wnd = new_win;
283 tp->rcv_wup = tp->rcv_nxt;
284
285 /* Make sure we do not exceed the maximum possible
286 * scaled window.
287 */
288 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
289 new_win = min(new_win, MAX_TCP_WINDOW);
290 else
291 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
292
293 /* RFC1323 scaling applied */
294 new_win >>= tp->rx_opt.rcv_wscale;
295
296 /* If we advertise zero window, disable fast path. */
297 if (new_win == 0)
298 tp->pred_flags = 0;
299
300 return new_win;
301 }
302
303 /* Packet ECN state for a SYN-ACK */
304 static inline void TCP_ECN_send_synack(const struct tcp_sock *tp, struct sk_buff *skb)
305 {
306 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
307 if (!(tp->ecn_flags & TCP_ECN_OK))
308 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
309 }
310
311 /* Packet ECN state for a SYN. */
312 static inline void TCP_ECN_send_syn(struct sock *sk, struct sk_buff *skb)
313 {
314 struct tcp_sock *tp = tcp_sk(sk);
315
316 tp->ecn_flags = 0;
317 if (sock_net(sk)->ipv4.sysctl_tcp_ecn == 1) {
318 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
319 tp->ecn_flags = TCP_ECN_OK;
320 }
321 }
322
323 static __inline__ void
324 TCP_ECN_make_synack(const struct request_sock *req, struct tcphdr *th)
325 {
326 if (inet_rsk(req)->ecn_ok)
327 th->ece = 1;
328 }
329
330 /* Set up ECN state for a packet on a ESTABLISHED socket that is about to
331 * be sent.
332 */
333 static inline void TCP_ECN_send(struct sock *sk, struct sk_buff *skb,
334 int tcp_header_len)
335 {
336 struct tcp_sock *tp = tcp_sk(sk);
337
338 if (tp->ecn_flags & TCP_ECN_OK) {
339 /* Not-retransmitted data segment: set ECT and inject CWR. */
340 if (skb->len != tcp_header_len &&
341 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
342 INET_ECN_xmit(sk);
343 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
344 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
345 tcp_hdr(skb)->cwr = 1;
346 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
347 }
348 } else {
349 /* ACK or retransmitted segment: clear ECT|CE */
350 INET_ECN_dontxmit(sk);
351 }
352 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
353 tcp_hdr(skb)->ece = 1;
354 }
355 }
356
357 /* Constructs common control bits of non-data skb. If SYN/FIN is present,
358 * auto increment end seqno.
359 */
360 static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
361 {
362 skb->ip_summed = CHECKSUM_PARTIAL;
363 skb->csum = 0;
364
365 TCP_SKB_CB(skb)->tcp_flags = flags;
366 TCP_SKB_CB(skb)->sacked = 0;
367
368 skb_shinfo(skb)->gso_segs = 1;
369 skb_shinfo(skb)->gso_size = 0;
370 skb_shinfo(skb)->gso_type = 0;
371
372 TCP_SKB_CB(skb)->seq = seq;
373 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
374 seq++;
375 TCP_SKB_CB(skb)->end_seq = seq;
376 }
377
378 static inline bool tcp_urg_mode(const struct tcp_sock *tp)
379 {
380 return tp->snd_una != tp->snd_up;
381 }
382
383 #define OPTION_SACK_ADVERTISE (1 << 0)
384 #define OPTION_TS (1 << 1)
385 #define OPTION_MD5 (1 << 2)
386 #define OPTION_WSCALE (1 << 3)
387 #define OPTION_COOKIE_EXTENSION (1 << 4)
388 #define OPTION_FAST_OPEN_COOKIE (1 << 8)
389
390 struct tcp_out_options {
391 u16 options; /* bit field of OPTION_* */
392 u16 mss; /* 0 to disable */
393 u8 ws; /* window scale, 0 to disable */
394 u8 num_sack_blocks; /* number of SACK blocks to include */
395 u8 hash_size; /* bytes in hash_location */
396 __u8 *hash_location; /* temporary pointer, overloaded */
397 __u32 tsval, tsecr; /* need to include OPTION_TS */
398 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
399 };
400
401 /* The sysctl int routines are generic, so check consistency here.
402 */
403 static u8 tcp_cookie_size_check(u8 desired)
404 {
405 int cookie_size;
406
407 if (desired > 0)
408 /* previously specified */
409 return desired;
410
411 cookie_size = ACCESS_ONCE(sysctl_tcp_cookie_size);
412 if (cookie_size <= 0)
413 /* no default specified */
414 return 0;
415
416 if (cookie_size <= TCP_COOKIE_MIN)
417 /* value too small, specify minimum */
418 return TCP_COOKIE_MIN;
419
420 if (cookie_size >= TCP_COOKIE_MAX)
421 /* value too large, specify maximum */
422 return TCP_COOKIE_MAX;
423
424 if (cookie_size & 1)
425 /* 8-bit multiple, illegal, fix it */
426 cookie_size++;
427
428 return (u8)cookie_size;
429 }
430
431 /* Write previously computed TCP options to the packet.
432 *
433 * Beware: Something in the Internet is very sensitive to the ordering of
434 * TCP options, we learned this through the hard way, so be careful here.
435 * Luckily we can at least blame others for their non-compliance but from
436 * inter-operatibility perspective it seems that we're somewhat stuck with
437 * the ordering which we have been using if we want to keep working with
438 * those broken things (not that it currently hurts anybody as there isn't
439 * particular reason why the ordering would need to be changed).
440 *
441 * At least SACK_PERM as the first option is known to lead to a disaster
442 * (but it may well be that other scenarios fail similarly).
443 */
444 static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
445 struct tcp_out_options *opts)
446 {
447 u16 options = opts->options; /* mungable copy */
448
449 /* Having both authentication and cookies for security is redundant,
450 * and there's certainly not enough room. Instead, the cookie-less
451 * extension variant is proposed.
452 *
453 * Consider the pessimal case with authentication. The options
454 * could look like:
455 * COOKIE|MD5(20) + MSS(4) + SACK|TS(12) + WSCALE(4) == 40
456 */
457 if (unlikely(OPTION_MD5 & options)) {
458 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
459 *ptr++ = htonl((TCPOPT_COOKIE << 24) |
460 (TCPOLEN_COOKIE_BASE << 16) |
461 (TCPOPT_MD5SIG << 8) |
462 TCPOLEN_MD5SIG);
463 } else {
464 *ptr++ = htonl((TCPOPT_NOP << 24) |
465 (TCPOPT_NOP << 16) |
466 (TCPOPT_MD5SIG << 8) |
467 TCPOLEN_MD5SIG);
468 }
469 options &= ~OPTION_COOKIE_EXTENSION;
470 /* overload cookie hash location */
471 opts->hash_location = (__u8 *)ptr;
472 ptr += 4;
473 }
474
475 if (unlikely(opts->mss)) {
476 *ptr++ = htonl((TCPOPT_MSS << 24) |
477 (TCPOLEN_MSS << 16) |
478 opts->mss);
479 }
480
481 if (likely(OPTION_TS & options)) {
482 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
483 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
484 (TCPOLEN_SACK_PERM << 16) |
485 (TCPOPT_TIMESTAMP << 8) |
486 TCPOLEN_TIMESTAMP);
487 options &= ~OPTION_SACK_ADVERTISE;
488 } else {
489 *ptr++ = htonl((TCPOPT_NOP << 24) |
490 (TCPOPT_NOP << 16) |
491 (TCPOPT_TIMESTAMP << 8) |
492 TCPOLEN_TIMESTAMP);
493 }
494 *ptr++ = htonl(opts->tsval);
495 *ptr++ = htonl(opts->tsecr);
496 }
497
498 /* Specification requires after timestamp, so do it now.
499 *
500 * Consider the pessimal case without authentication. The options
501 * could look like:
502 * MSS(4) + SACK|TS(12) + COOKIE(20) + WSCALE(4) == 40
503 */
504 if (unlikely(OPTION_COOKIE_EXTENSION & options)) {
505 __u8 *cookie_copy = opts->hash_location;
506 u8 cookie_size = opts->hash_size;
507
508 /* 8-bit multiple handled in tcp_cookie_size_check() above,
509 * and elsewhere.
510 */
511 if (0x2 & cookie_size) {
512 __u8 *p = (__u8 *)ptr;
513
514 /* 16-bit multiple */
515 *p++ = TCPOPT_COOKIE;
516 *p++ = TCPOLEN_COOKIE_BASE + cookie_size;
517 *p++ = *cookie_copy++;
518 *p++ = *cookie_copy++;
519 ptr++;
520 cookie_size -= 2;
521 } else {
522 /* 32-bit multiple */
523 *ptr++ = htonl(((TCPOPT_NOP << 24) |
524 (TCPOPT_NOP << 16) |
525 (TCPOPT_COOKIE << 8) |
526 TCPOLEN_COOKIE_BASE) +
527 cookie_size);
528 }
529
530 if (cookie_size > 0) {
531 memcpy(ptr, cookie_copy, cookie_size);
532 ptr += (cookie_size / 4);
533 }
534 }
535
536 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
537 *ptr++ = htonl((TCPOPT_NOP << 24) |
538 (TCPOPT_NOP << 16) |
539 (TCPOPT_SACK_PERM << 8) |
540 TCPOLEN_SACK_PERM);
541 }
542
543 if (unlikely(OPTION_WSCALE & options)) {
544 *ptr++ = htonl((TCPOPT_NOP << 24) |
545 (TCPOPT_WINDOW << 16) |
546 (TCPOLEN_WINDOW << 8) |
547 opts->ws);
548 }
549
550 if (unlikely(opts->num_sack_blocks)) {
551 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
552 tp->duplicate_sack : tp->selective_acks;
553 int this_sack;
554
555 *ptr++ = htonl((TCPOPT_NOP << 24) |
556 (TCPOPT_NOP << 16) |
557 (TCPOPT_SACK << 8) |
558 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
559 TCPOLEN_SACK_PERBLOCK)));
560
561 for (this_sack = 0; this_sack < opts->num_sack_blocks;
562 ++this_sack) {
563 *ptr++ = htonl(sp[this_sack].start_seq);
564 *ptr++ = htonl(sp[this_sack].end_seq);
565 }
566
567 tp->rx_opt.dsack = 0;
568 }
569
570 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
571 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
572
573 *ptr++ = htonl((TCPOPT_EXP << 24) |
574 ((TCPOLEN_EXP_FASTOPEN_BASE + foc->len) << 16) |
575 TCPOPT_FASTOPEN_MAGIC);
576
577 memcpy(ptr, foc->val, foc->len);
578 if ((foc->len & 3) == 2) {
579 u8 *align = ((u8 *)ptr) + foc->len;
580 align[0] = align[1] = TCPOPT_NOP;
581 }
582 ptr += (foc->len + 3) >> 2;
583 }
584 }
585
586 /* Compute TCP options for SYN packets. This is not the final
587 * network wire format yet.
588 */
589 static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
590 struct tcp_out_options *opts,
591 struct tcp_md5sig_key **md5)
592 {
593 struct tcp_sock *tp = tcp_sk(sk);
594 struct tcp_cookie_values *cvp = tp->cookie_values;
595 unsigned int remaining = MAX_TCP_OPTION_SPACE;
596 u8 cookie_size = (!tp->rx_opt.cookie_out_never && cvp != NULL) ?
597 tcp_cookie_size_check(cvp->cookie_desired) :
598 0;
599 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
600
601 #ifdef CONFIG_TCP_MD5SIG
602 *md5 = tp->af_specific->md5_lookup(sk, sk);
603 if (*md5) {
604 opts->options |= OPTION_MD5;
605 remaining -= TCPOLEN_MD5SIG_ALIGNED;
606 }
607 #else
608 *md5 = NULL;
609 #endif
610
611 /* We always get an MSS option. The option bytes which will be seen in
612 * normal data packets should timestamps be used, must be in the MSS
613 * advertised. But we subtract them from tp->mss_cache so that
614 * calculations in tcp_sendmsg are simpler etc. So account for this
615 * fact here if necessary. If we don't do this correctly, as a
616 * receiver we won't recognize data packets as being full sized when we
617 * should, and thus we won't abide by the delayed ACK rules correctly.
618 * SACKs don't matter, we never delay an ACK when we have any of those
619 * going out. */
620 opts->mss = tcp_advertise_mss(sk);
621 remaining -= TCPOLEN_MSS_ALIGNED;
622
623 if (likely(sysctl_tcp_timestamps && *md5 == NULL)) {
624 opts->options |= OPTION_TS;
625 opts->tsval = TCP_SKB_CB(skb)->when + tp->tsoffset;
626 opts->tsecr = tp->rx_opt.ts_recent;
627 remaining -= TCPOLEN_TSTAMP_ALIGNED;
628 }
629 if (likely(sysctl_tcp_window_scaling)) {
630 opts->ws = tp->rx_opt.rcv_wscale;
631 opts->options |= OPTION_WSCALE;
632 remaining -= TCPOLEN_WSCALE_ALIGNED;
633 }
634 if (likely(sysctl_tcp_sack)) {
635 opts->options |= OPTION_SACK_ADVERTISE;
636 if (unlikely(!(OPTION_TS & opts->options)))
637 remaining -= TCPOLEN_SACKPERM_ALIGNED;
638 }
639
640 if (fastopen && fastopen->cookie.len >= 0) {
641 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + fastopen->cookie.len;
642 need = (need + 3) & ~3U; /* Align to 32 bits */
643 if (remaining >= need) {
644 opts->options |= OPTION_FAST_OPEN_COOKIE;
645 opts->fastopen_cookie = &fastopen->cookie;
646 remaining -= need;
647 tp->syn_fastopen = 1;
648 }
649 }
650 /* Note that timestamps are required by the specification.
651 *
652 * Odd numbers of bytes are prohibited by the specification, ensuring
653 * that the cookie is 16-bit aligned, and the resulting cookie pair is
654 * 32-bit aligned.
655 */
656 if (*md5 == NULL &&
657 (OPTION_TS & opts->options) &&
658 cookie_size > 0) {
659 int need = TCPOLEN_COOKIE_BASE + cookie_size;
660
661 if (0x2 & need) {
662 /* 32-bit multiple */
663 need += 2; /* NOPs */
664
665 if (need > remaining) {
666 /* try shrinking cookie to fit */
667 cookie_size -= 2;
668 need -= 4;
669 }
670 }
671 while (need > remaining && TCP_COOKIE_MIN <= cookie_size) {
672 cookie_size -= 4;
673 need -= 4;
674 }
675 if (TCP_COOKIE_MIN <= cookie_size) {
676 opts->options |= OPTION_COOKIE_EXTENSION;
677 opts->hash_location = (__u8 *)&cvp->cookie_pair[0];
678 opts->hash_size = cookie_size;
679
680 /* Remember for future incarnations. */
681 cvp->cookie_desired = cookie_size;
682
683 if (cvp->cookie_desired != cvp->cookie_pair_size) {
684 /* Currently use random bytes as a nonce,
685 * assuming these are completely unpredictable
686 * by hostile users of the same system.
687 */
688 get_random_bytes(&cvp->cookie_pair[0],
689 cookie_size);
690 cvp->cookie_pair_size = cookie_size;
691 }
692
693 remaining -= need;
694 }
695 }
696 return MAX_TCP_OPTION_SPACE - remaining;
697 }
698
699 /* Set up TCP options for SYN-ACKs. */
700 static unsigned int tcp_synack_options(struct sock *sk,
701 struct request_sock *req,
702 unsigned int mss, struct sk_buff *skb,
703 struct tcp_out_options *opts,
704 struct tcp_md5sig_key **md5,
705 struct tcp_extend_values *xvp,
706 struct tcp_fastopen_cookie *foc)
707 {
708 struct inet_request_sock *ireq = inet_rsk(req);
709 unsigned int remaining = MAX_TCP_OPTION_SPACE;
710 u8 cookie_plus = (xvp != NULL && !xvp->cookie_out_never) ?
711 xvp->cookie_plus :
712 0;
713
714 #ifdef CONFIG_TCP_MD5SIG
715 *md5 = tcp_rsk(req)->af_specific->md5_lookup(sk, req);
716 if (*md5) {
717 opts->options |= OPTION_MD5;
718 remaining -= TCPOLEN_MD5SIG_ALIGNED;
719
720 /* We can't fit any SACK blocks in a packet with MD5 + TS
721 * options. There was discussion about disabling SACK
722 * rather than TS in order to fit in better with old,
723 * buggy kernels, but that was deemed to be unnecessary.
724 */
725 ireq->tstamp_ok &= !ireq->sack_ok;
726 }
727 #else
728 *md5 = NULL;
729 #endif
730
731 /* We always send an MSS option. */
732 opts->mss = mss;
733 remaining -= TCPOLEN_MSS_ALIGNED;
734
735 if (likely(ireq->wscale_ok)) {
736 opts->ws = ireq->rcv_wscale;
737 opts->options |= OPTION_WSCALE;
738 remaining -= TCPOLEN_WSCALE_ALIGNED;
739 }
740 if (likely(ireq->tstamp_ok)) {
741 opts->options |= OPTION_TS;
742 opts->tsval = TCP_SKB_CB(skb)->when;
743 opts->tsecr = req->ts_recent;
744 remaining -= TCPOLEN_TSTAMP_ALIGNED;
745 }
746 if (likely(ireq->sack_ok)) {
747 opts->options |= OPTION_SACK_ADVERTISE;
748 if (unlikely(!ireq->tstamp_ok))
749 remaining -= TCPOLEN_SACKPERM_ALIGNED;
750 }
751 if (foc != NULL) {
752 u32 need = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
753 need = (need + 3) & ~3U; /* Align to 32 bits */
754 if (remaining >= need) {
755 opts->options |= OPTION_FAST_OPEN_COOKIE;
756 opts->fastopen_cookie = foc;
757 remaining -= need;
758 }
759 }
760 /* Similar rationale to tcp_syn_options() applies here, too.
761 * If the <SYN> options fit, the same options should fit now!
762 */
763 if (*md5 == NULL &&
764 ireq->tstamp_ok &&
765 cookie_plus > TCPOLEN_COOKIE_BASE) {
766 int need = cookie_plus; /* has TCPOLEN_COOKIE_BASE */
767
768 if (0x2 & need) {
769 /* 32-bit multiple */
770 need += 2; /* NOPs */
771 }
772 if (need <= remaining) {
773 opts->options |= OPTION_COOKIE_EXTENSION;
774 opts->hash_size = cookie_plus - TCPOLEN_COOKIE_BASE;
775 remaining -= need;
776 } else {
777 /* There's no error return, so flag it. */
778 xvp->cookie_out_never = 1; /* true */
779 opts->hash_size = 0;
780 }
781 }
782 return MAX_TCP_OPTION_SPACE - remaining;
783 }
784
785 /* Compute TCP options for ESTABLISHED sockets. This is not the
786 * final wire format yet.
787 */
788 static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
789 struct tcp_out_options *opts,
790 struct tcp_md5sig_key **md5)
791 {
792 struct tcp_skb_cb *tcb = skb ? TCP_SKB_CB(skb) : NULL;
793 struct tcp_sock *tp = tcp_sk(sk);
794 unsigned int size = 0;
795 unsigned int eff_sacks;
796
797 #ifdef CONFIG_TCP_MD5SIG
798 *md5 = tp->af_specific->md5_lookup(sk, sk);
799 if (unlikely(*md5)) {
800 opts->options |= OPTION_MD5;
801 size += TCPOLEN_MD5SIG_ALIGNED;
802 }
803 #else
804 *md5 = NULL;
805 #endif
806
807 if (likely(tp->rx_opt.tstamp_ok)) {
808 opts->options |= OPTION_TS;
809 opts->tsval = tcb ? tcb->when + tp->tsoffset : 0;
810 opts->tsecr = tp->rx_opt.ts_recent;
811 size += TCPOLEN_TSTAMP_ALIGNED;
812 }
813
814 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
815 if (unlikely(eff_sacks)) {
816 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
817 opts->num_sack_blocks =
818 min_t(unsigned int, eff_sacks,
819 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
820 TCPOLEN_SACK_PERBLOCK);
821 size += TCPOLEN_SACK_BASE_ALIGNED +
822 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
823 }
824
825 return size;
826 }
827
828
829 /* TCP SMALL QUEUES (TSQ)
830 *
831 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
832 * to reduce RTT and bufferbloat.
833 * We do this using a special skb destructor (tcp_wfree).
834 *
835 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
836 * needs to be reallocated in a driver.
837 * The invariant being skb->truesize substracted from sk->sk_wmem_alloc
838 *
839 * Since transmit from skb destructor is forbidden, we use a tasklet
840 * to process all sockets that eventually need to send more skbs.
841 * We use one tasklet per cpu, with its own queue of sockets.
842 */
843 struct tsq_tasklet {
844 struct tasklet_struct tasklet;
845 struct list_head head; /* queue of tcp sockets */
846 };
847 static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
848
849 static void tcp_tsq_handler(struct sock *sk)
850 {
851 if ((1 << sk->sk_state) &
852 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
853 TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
854 tcp_write_xmit(sk, tcp_current_mss(sk), 0, 0, GFP_ATOMIC);
855 }
856 /*
857 * One tasklest per cpu tries to send more skbs.
858 * We run in tasklet context but need to disable irqs when
859 * transfering tsq->head because tcp_wfree() might
860 * interrupt us (non NAPI drivers)
861 */
862 static void tcp_tasklet_func(unsigned long data)
863 {
864 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
865 LIST_HEAD(list);
866 unsigned long flags;
867 struct list_head *q, *n;
868 struct tcp_sock *tp;
869 struct sock *sk;
870
871 local_irq_save(flags);
872 list_splice_init(&tsq->head, &list);
873 local_irq_restore(flags);
874
875 list_for_each_safe(q, n, &list) {
876 tp = list_entry(q, struct tcp_sock, tsq_node);
877 list_del(&tp->tsq_node);
878
879 sk = (struct sock *)tp;
880 bh_lock_sock(sk);
881
882 if (!sock_owned_by_user(sk)) {
883 tcp_tsq_handler(sk);
884 } else {
885 /* defer the work to tcp_release_cb() */
886 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
887 }
888 bh_unlock_sock(sk);
889
890 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
891 sk_free(sk);
892 }
893 }
894
895 #define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
896 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
897 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
898 (1UL << TCP_MTU_REDUCED_DEFERRED))
899 /**
900 * tcp_release_cb - tcp release_sock() callback
901 * @sk: socket
902 *
903 * called from release_sock() to perform protocol dependent
904 * actions before socket release.
905 */
906 void tcp_release_cb(struct sock *sk)
907 {
908 struct tcp_sock *tp = tcp_sk(sk);
909 unsigned long flags, nflags;
910
911 /* perform an atomic operation only if at least one flag is set */
912 do {
913 flags = tp->tsq_flags;
914 if (!(flags & TCP_DEFERRED_ALL))
915 return;
916 nflags = flags & ~TCP_DEFERRED_ALL;
917 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
918
919 if (flags & (1UL << TCP_TSQ_DEFERRED))
920 tcp_tsq_handler(sk);
921
922 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
923 tcp_write_timer_handler(sk);
924 __sock_put(sk);
925 }
926 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
927 tcp_delack_timer_handler(sk);
928 __sock_put(sk);
929 }
930 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
931 sk->sk_prot->mtu_reduced(sk);
932 __sock_put(sk);
933 }
934 }
935 EXPORT_SYMBOL(tcp_release_cb);
936
937 void __init tcp_tasklet_init(void)
938 {
939 int i;
940
941 for_each_possible_cpu(i) {
942 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
943
944 INIT_LIST_HEAD(&tsq->head);
945 tasklet_init(&tsq->tasklet,
946 tcp_tasklet_func,
947 (unsigned long)tsq);
948 }
949 }
950
951 /*
952 * Write buffer destructor automatically called from kfree_skb.
953 * We cant xmit new skbs from this context, as we might already
954 * hold qdisc lock.
955 */
956 static void tcp_wfree(struct sk_buff *skb)
957 {
958 struct sock *sk = skb->sk;
959 struct tcp_sock *tp = tcp_sk(sk);
960
961 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
962 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
963 unsigned long flags;
964 struct tsq_tasklet *tsq;
965
966 /* Keep a ref on socket.
967 * This last ref will be released in tcp_tasklet_func()
968 */
969 atomic_sub(skb->truesize - 1, &sk->sk_wmem_alloc);
970
971 /* queue this socket to tasklet queue */
972 local_irq_save(flags);
973 tsq = &__get_cpu_var(tsq_tasklet);
974 list_add(&tp->tsq_node, &tsq->head);
975 tasklet_schedule(&tsq->tasklet);
976 local_irq_restore(flags);
977 } else {
978 sock_wfree(skb);
979 }
980 }
981
982 /* This routine actually transmits TCP packets queued in by
983 * tcp_do_sendmsg(). This is used by both the initial
984 * transmission and possible later retransmissions.
985 * All SKB's seen here are completely headerless. It is our
986 * job to build the TCP header, and pass the packet down to
987 * IP so it can do the same plus pass the packet off to the
988 * device.
989 *
990 * We are working here with either a clone of the original
991 * SKB, or a fresh unique copy made by the retransmit engine.
992 */
993 static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
994 gfp_t gfp_mask)
995 {
996 const struct inet_connection_sock *icsk = inet_csk(sk);
997 struct inet_sock *inet;
998 struct tcp_sock *tp;
999 struct tcp_skb_cb *tcb;
1000 struct tcp_out_options opts;
1001 unsigned int tcp_options_size, tcp_header_size;
1002 struct tcp_md5sig_key *md5;
1003 struct tcphdr *th;
1004 int err;
1005
1006 BUG_ON(!skb || !tcp_skb_pcount(skb));
1007
1008 /* If congestion control is doing timestamping, we must
1009 * take such a timestamp before we potentially clone/copy.
1010 */
1011 if (icsk->icsk_ca_ops->flags & TCP_CONG_RTT_STAMP)
1012 __net_timestamp(skb);
1013
1014 if (likely(clone_it)) {
1015 if (unlikely(skb_cloned(skb)))
1016 skb = pskb_copy(skb, gfp_mask);
1017 else
1018 skb = skb_clone(skb, gfp_mask);
1019 if (unlikely(!skb))
1020 return -ENOBUFS;
1021 }
1022
1023 inet = inet_sk(sk);
1024 tp = tcp_sk(sk);
1025 tcb = TCP_SKB_CB(skb);
1026 memset(&opts, 0, sizeof(opts));
1027
1028 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
1029 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1030 else
1031 tcp_options_size = tcp_established_options(sk, skb, &opts,
1032 &md5);
1033 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1034
1035 if (tcp_packets_in_flight(tp) == 0) {
1036 tcp_ca_event(sk, CA_EVENT_TX_START);
1037 skb->ooo_okay = 1;
1038 } else
1039 skb->ooo_okay = 0;
1040
1041 skb_push(skb, tcp_header_size);
1042 skb_reset_transport_header(skb);
1043
1044 skb_orphan(skb);
1045 skb->sk = sk;
1046 skb->destructor = (sysctl_tcp_limit_output_bytes > 0) ?
1047 tcp_wfree : sock_wfree;
1048 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
1049
1050 /* Build TCP header and checksum it. */
1051 th = tcp_hdr(skb);
1052 th->source = inet->inet_sport;
1053 th->dest = inet->inet_dport;
1054 th->seq = htonl(tcb->seq);
1055 th->ack_seq = htonl(tp->rcv_nxt);
1056 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1057 tcb->tcp_flags);
1058
1059 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1060 /* RFC1323: The window in SYN & SYN/ACK segments
1061 * is never scaled.
1062 */
1063 th->window = htons(min(tp->rcv_wnd, 65535U));
1064 } else {
1065 th->window = htons(tcp_select_window(sk));
1066 }
1067 th->check = 0;
1068 th->urg_ptr = 0;
1069
1070 /* The urg_mode check is necessary during a below snd_una win probe */
1071 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1072 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1073 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1074 th->urg = 1;
1075 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1076 th->urg_ptr = htons(0xFFFF);
1077 th->urg = 1;
1078 }
1079 }
1080
1081 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1082 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
1083 TCP_ECN_send(sk, skb, tcp_header_size);
1084
1085 #ifdef CONFIG_TCP_MD5SIG
1086 /* Calculate the MD5 hash, as we have all we need now */
1087 if (md5) {
1088 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1089 tp->af_specific->calc_md5_hash(opts.hash_location,
1090 md5, sk, NULL, skb);
1091 }
1092 #endif
1093
1094 icsk->icsk_af_ops->send_check(sk, skb);
1095
1096 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1097 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1098
1099 if (skb->len != tcp_header_size)
1100 tcp_event_data_sent(tp, sk);
1101
1102 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1103 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1104 tcp_skb_pcount(skb));
1105
1106 err = icsk->icsk_af_ops->queue_xmit(skb, &inet->cork.fl);
1107 if (likely(err <= 0))
1108 return err;
1109
1110 tcp_enter_cwr(sk, 1);
1111
1112 return net_xmit_eval(err);
1113 }
1114
1115 /* This routine just queues the buffer for sending.
1116 *
1117 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1118 * otherwise socket can stall.
1119 */
1120 static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1121 {
1122 struct tcp_sock *tp = tcp_sk(sk);
1123
1124 /* Advance write_seq and place onto the write_queue. */
1125 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1126 skb_header_release(skb);
1127 tcp_add_write_queue_tail(sk, skb);
1128 sk->sk_wmem_queued += skb->truesize;
1129 sk_mem_charge(sk, skb->truesize);
1130 }
1131
1132 /* Initialize TSO segments for a packet. */
1133 static void tcp_set_skb_tso_segs(const struct sock *sk, struct sk_buff *skb,
1134 unsigned int mss_now)
1135 {
1136 if (skb->len <= mss_now || !sk_can_gso(sk) ||
1137 skb->ip_summed == CHECKSUM_NONE) {
1138 /* Avoid the costly divide in the normal
1139 * non-TSO case.
1140 */
1141 skb_shinfo(skb)->gso_segs = 1;
1142 skb_shinfo(skb)->gso_size = 0;
1143 skb_shinfo(skb)->gso_type = 0;
1144 } else {
1145 skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss_now);
1146 skb_shinfo(skb)->gso_size = mss_now;
1147 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1148 }
1149 }
1150
1151 /* When a modification to fackets out becomes necessary, we need to check
1152 * skb is counted to fackets_out or not.
1153 */
1154 static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1155 int decr)
1156 {
1157 struct tcp_sock *tp = tcp_sk(sk);
1158
1159 if (!tp->sacked_out || tcp_is_reno(tp))
1160 return;
1161
1162 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1163 tp->fackets_out -= decr;
1164 }
1165
1166 /* Pcount in the middle of the write queue got changed, we need to do various
1167 * tweaks to fix counters
1168 */
1169 static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1170 {
1171 struct tcp_sock *tp = tcp_sk(sk);
1172
1173 tp->packets_out -= decr;
1174
1175 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1176 tp->sacked_out -= decr;
1177 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1178 tp->retrans_out -= decr;
1179 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1180 tp->lost_out -= decr;
1181
1182 /* Reno case is special. Sigh... */
1183 if (tcp_is_reno(tp) && decr > 0)
1184 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1185
1186 tcp_adjust_fackets_out(sk, skb, decr);
1187
1188 if (tp->lost_skb_hint &&
1189 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1190 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1191 tp->lost_cnt_hint -= decr;
1192
1193 tcp_verify_left_out(tp);
1194 }
1195
1196 /* Function to create two new TCP segments. Shrinks the given segment
1197 * to the specified size and appends a new segment with the rest of the
1198 * packet to the list. This won't be called frequently, I hope.
1199 * Remember, these are still headerless SKBs at this point.
1200 */
1201 int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1202 unsigned int mss_now)
1203 {
1204 struct tcp_sock *tp = tcp_sk(sk);
1205 struct sk_buff *buff;
1206 int nsize, old_factor;
1207 int nlen;
1208 u8 flags;
1209
1210 if (WARN_ON(len > skb->len))
1211 return -EINVAL;
1212
1213 nsize = skb_headlen(skb) - len;
1214 if (nsize < 0)
1215 nsize = 0;
1216
1217 if (skb_cloned(skb) &&
1218 skb_is_nonlinear(skb) &&
1219 pskb_expand_head(skb, 0, 0, GFP_ATOMIC))
1220 return -ENOMEM;
1221
1222 /* Get a new skb... force flag on. */
1223 buff = sk_stream_alloc_skb(sk, nsize, GFP_ATOMIC);
1224 if (buff == NULL)
1225 return -ENOMEM; /* We'll just try again later. */
1226
1227 sk->sk_wmem_queued += buff->truesize;
1228 sk_mem_charge(sk, buff->truesize);
1229 nlen = skb->len - len - nsize;
1230 buff->truesize += nlen;
1231 skb->truesize -= nlen;
1232
1233 /* Correct the sequence numbers. */
1234 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1235 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1236 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1237
1238 /* PSH and FIN should only be set in the second packet. */
1239 flags = TCP_SKB_CB(skb)->tcp_flags;
1240 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1241 TCP_SKB_CB(buff)->tcp_flags = flags;
1242 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1243
1244 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1245 /* Copy and checksum data tail into the new buffer. */
1246 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1247 skb_put(buff, nsize),
1248 nsize, 0);
1249
1250 skb_trim(skb, len);
1251
1252 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1253 } else {
1254 skb->ip_summed = CHECKSUM_PARTIAL;
1255 skb_split(skb, buff, len);
1256 }
1257
1258 buff->ip_summed = skb->ip_summed;
1259
1260 /* Looks stupid, but our code really uses when of
1261 * skbs, which it never sent before. --ANK
1262 */
1263 TCP_SKB_CB(buff)->when = TCP_SKB_CB(skb)->when;
1264 buff->tstamp = skb->tstamp;
1265
1266 old_factor = tcp_skb_pcount(skb);
1267
1268 /* Fix up tso_factor for both original and new SKB. */
1269 tcp_set_skb_tso_segs(sk, skb, mss_now);
1270 tcp_set_skb_tso_segs(sk, buff, mss_now);
1271
1272 /* If this packet has been sent out already, we must
1273 * adjust the various packet counters.
1274 */
1275 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1276 int diff = old_factor - tcp_skb_pcount(skb) -
1277 tcp_skb_pcount(buff);
1278
1279 if (diff)
1280 tcp_adjust_pcount(sk, skb, diff);
1281 }
1282
1283 /* Link BUFF into the send queue. */
1284 skb_header_release(buff);
1285 tcp_insert_write_queue_after(skb, buff, sk);
1286
1287 return 0;
1288 }
1289
1290 /* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1291 * eventually). The difference is that pulled data not copied, but
1292 * immediately discarded.
1293 */
1294 static void __pskb_trim_head(struct sk_buff *skb, int len)
1295 {
1296 int i, k, eat;
1297
1298 eat = min_t(int, len, skb_headlen(skb));
1299 if (eat) {
1300 __skb_pull(skb, eat);
1301 len -= eat;
1302 if (!len)
1303 return;
1304 }
1305 eat = len;
1306 k = 0;
1307 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
1308 int size = skb_frag_size(&skb_shinfo(skb)->frags[i]);
1309
1310 if (size <= eat) {
1311 skb_frag_unref(skb, i);
1312 eat -= size;
1313 } else {
1314 skb_shinfo(skb)->frags[k] = skb_shinfo(skb)->frags[i];
1315 if (eat) {
1316 skb_shinfo(skb)->frags[k].page_offset += eat;
1317 skb_frag_size_sub(&skb_shinfo(skb)->frags[k], eat);
1318 eat = 0;
1319 }
1320 k++;
1321 }
1322 }
1323 skb_shinfo(skb)->nr_frags = k;
1324
1325 skb_reset_tail_pointer(skb);
1326 skb->data_len -= len;
1327 skb->len = skb->data_len;
1328 }
1329
1330 /* Remove acked data from a packet in the transmit queue. */
1331 int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1332 {
1333 if (skb_unclone(skb, GFP_ATOMIC))
1334 return -ENOMEM;
1335
1336 __pskb_trim_head(skb, len);
1337
1338 TCP_SKB_CB(skb)->seq += len;
1339 skb->ip_summed = CHECKSUM_PARTIAL;
1340
1341 skb->truesize -= len;
1342 sk->sk_wmem_queued -= len;
1343 sk_mem_uncharge(sk, len);
1344 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1345
1346 /* Any change of skb->len requires recalculation of tso factor. */
1347 if (tcp_skb_pcount(skb) > 1)
1348 tcp_set_skb_tso_segs(sk, skb, tcp_skb_mss(skb));
1349
1350 return 0;
1351 }
1352
1353 /* Calculate MSS not accounting any TCP options. */
1354 static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1355 {
1356 const struct tcp_sock *tp = tcp_sk(sk);
1357 const struct inet_connection_sock *icsk = inet_csk(sk);
1358 int mss_now;
1359
1360 /* Calculate base mss without TCP options:
1361 It is MMS_S - sizeof(tcphdr) of rfc1122
1362 */
1363 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1364
1365 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1366 if (icsk->icsk_af_ops->net_frag_header_len) {
1367 const struct dst_entry *dst = __sk_dst_get(sk);
1368
1369 if (dst && dst_allfrag(dst))
1370 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1371 }
1372
1373 /* Clamp it (mss_clamp does not include tcp options) */
1374 if (mss_now > tp->rx_opt.mss_clamp)
1375 mss_now = tp->rx_opt.mss_clamp;
1376
1377 /* Now subtract optional transport overhead */
1378 mss_now -= icsk->icsk_ext_hdr_len;
1379
1380 /* Then reserve room for full set of TCP options and 8 bytes of data */
1381 if (mss_now < 48)
1382 mss_now = 48;
1383 return mss_now;
1384 }
1385
1386 /* Calculate MSS. Not accounting for SACKs here. */
1387 int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1388 {
1389 /* Subtract TCP options size, not including SACKs */
1390 return __tcp_mtu_to_mss(sk, pmtu) -
1391 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1392 }
1393
1394 /* Inverse of above */
1395 int tcp_mss_to_mtu(struct sock *sk, int mss)
1396 {
1397 const struct tcp_sock *tp = tcp_sk(sk);
1398 const struct inet_connection_sock *icsk = inet_csk(sk);
1399 int mtu;
1400
1401 mtu = mss +
1402 tp->tcp_header_len +
1403 icsk->icsk_ext_hdr_len +
1404 icsk->icsk_af_ops->net_header_len;
1405
1406 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1407 if (icsk->icsk_af_ops->net_frag_header_len) {
1408 const struct dst_entry *dst = __sk_dst_get(sk);
1409
1410 if (dst && dst_allfrag(dst))
1411 mtu += icsk->icsk_af_ops->net_frag_header_len;
1412 }
1413 return mtu;
1414 }
1415
1416 /* MTU probing init per socket */
1417 void tcp_mtup_init(struct sock *sk)
1418 {
1419 struct tcp_sock *tp = tcp_sk(sk);
1420 struct inet_connection_sock *icsk = inet_csk(sk);
1421
1422 icsk->icsk_mtup.enabled = sysctl_tcp_mtu_probing > 1;
1423 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1424 icsk->icsk_af_ops->net_header_len;
1425 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, sysctl_tcp_base_mss);
1426 icsk->icsk_mtup.probe_size = 0;
1427 }
1428 EXPORT_SYMBOL(tcp_mtup_init);
1429
1430 /* This function synchronize snd mss to current pmtu/exthdr set.
1431
1432 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1433 for TCP options, but includes only bare TCP header.
1434
1435 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1436 It is minimum of user_mss and mss received with SYN.
1437 It also does not include TCP options.
1438
1439 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1440
1441 tp->mss_cache is current effective sending mss, including
1442 all tcp options except for SACKs. It is evaluated,
1443 taking into account current pmtu, but never exceeds
1444 tp->rx_opt.mss_clamp.
1445
1446 NOTE1. rfc1122 clearly states that advertised MSS
1447 DOES NOT include either tcp or ip options.
1448
1449 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1450 are READ ONLY outside this function. --ANK (980731)
1451 */
1452 unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1453 {
1454 struct tcp_sock *tp = tcp_sk(sk);
1455 struct inet_connection_sock *icsk = inet_csk(sk);
1456 int mss_now;
1457
1458 if (icsk->icsk_mtup.search_high > pmtu)
1459 icsk->icsk_mtup.search_high = pmtu;
1460
1461 mss_now = tcp_mtu_to_mss(sk, pmtu);
1462 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1463
1464 /* And store cached results */
1465 icsk->icsk_pmtu_cookie = pmtu;
1466 if (icsk->icsk_mtup.enabled)
1467 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1468 tp->mss_cache = mss_now;
1469
1470 return mss_now;
1471 }
1472 EXPORT_SYMBOL(tcp_sync_mss);
1473
1474 /* Compute the current effective MSS, taking SACKs and IP options,
1475 * and even PMTU discovery events into account.
1476 */
1477 unsigned int tcp_current_mss(struct sock *sk)
1478 {
1479 const struct tcp_sock *tp = tcp_sk(sk);
1480 const struct dst_entry *dst = __sk_dst_get(sk);
1481 u32 mss_now;
1482 unsigned int header_len;
1483 struct tcp_out_options opts;
1484 struct tcp_md5sig_key *md5;
1485
1486 mss_now = tp->mss_cache;
1487
1488 if (dst) {
1489 u32 mtu = dst_mtu(dst);
1490 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1491 mss_now = tcp_sync_mss(sk, mtu);
1492 }
1493
1494 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1495 sizeof(struct tcphdr);
1496 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1497 * some common options. If this is an odd packet (because we have SACK
1498 * blocks etc) then our calculated header_len will be different, and
1499 * we have to adjust mss_now correspondingly */
1500 if (header_len != tp->tcp_header_len) {
1501 int delta = (int) header_len - tp->tcp_header_len;
1502 mss_now -= delta;
1503 }
1504
1505 return mss_now;
1506 }
1507
1508 /* Congestion window validation. (RFC2861) */
1509 static void tcp_cwnd_validate(struct sock *sk)
1510 {
1511 struct tcp_sock *tp = tcp_sk(sk);
1512
1513 if (tp->packets_out >= tp->snd_cwnd) {
1514 /* Network is feed fully. */
1515 tp->snd_cwnd_used = 0;
1516 tp->snd_cwnd_stamp = tcp_time_stamp;
1517 } else {
1518 /* Network starves. */
1519 if (tp->packets_out > tp->snd_cwnd_used)
1520 tp->snd_cwnd_used = tp->packets_out;
1521
1522 if (sysctl_tcp_slow_start_after_idle &&
1523 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1524 tcp_cwnd_application_limited(sk);
1525 }
1526 }
1527
1528 /* Returns the portion of skb which can be sent right away without
1529 * introducing MSS oddities to segment boundaries. In rare cases where
1530 * mss_now != mss_cache, we will request caller to create a small skb
1531 * per input skb which could be mostly avoided here (if desired).
1532 *
1533 * We explicitly want to create a request for splitting write queue tail
1534 * to a small skb for Nagle purposes while avoiding unnecessary modulos,
1535 * thus all the complexity (cwnd_len is always MSS multiple which we
1536 * return whenever allowed by the other factors). Basically we need the
1537 * modulo only when the receiver window alone is the limiting factor or
1538 * when we would be allowed to send the split-due-to-Nagle skb fully.
1539 */
1540 static unsigned int tcp_mss_split_point(const struct sock *sk, const struct sk_buff *skb,
1541 unsigned int mss_now, unsigned int max_segs)
1542 {
1543 const struct tcp_sock *tp = tcp_sk(sk);
1544 u32 needed, window, max_len;
1545
1546 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1547 max_len = mss_now * max_segs;
1548
1549 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1550 return max_len;
1551
1552 needed = min(skb->len, window);
1553
1554 if (max_len <= needed)
1555 return max_len;
1556
1557 return needed - needed % mss_now;
1558 }
1559
1560 /* Can at least one segment of SKB be sent right now, according to the
1561 * congestion window rules? If so, return how many segments are allowed.
1562 */
1563 static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1564 const struct sk_buff *skb)
1565 {
1566 u32 in_flight, cwnd;
1567
1568 /* Don't be strict about the congestion window for the final FIN. */
1569 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1570 tcp_skb_pcount(skb) == 1)
1571 return 1;
1572
1573 in_flight = tcp_packets_in_flight(tp);
1574 cwnd = tp->snd_cwnd;
1575 if (in_flight < cwnd)
1576 return (cwnd - in_flight);
1577
1578 return 0;
1579 }
1580
1581 /* Initialize TSO state of a skb.
1582 * This must be invoked the first time we consider transmitting
1583 * SKB onto the wire.
1584 */
1585 static int tcp_init_tso_segs(const struct sock *sk, struct sk_buff *skb,
1586 unsigned int mss_now)
1587 {
1588 int tso_segs = tcp_skb_pcount(skb);
1589
1590 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1591 tcp_set_skb_tso_segs(sk, skb, mss_now);
1592 tso_segs = tcp_skb_pcount(skb);
1593 }
1594 return tso_segs;
1595 }
1596
1597 /* Minshall's variant of the Nagle send check. */
1598 static inline bool tcp_minshall_check(const struct tcp_sock *tp)
1599 {
1600 return after(tp->snd_sml, tp->snd_una) &&
1601 !after(tp->snd_sml, tp->snd_nxt);
1602 }
1603
1604 /* Return false, if packet can be sent now without violation Nagle's rules:
1605 * 1. It is full sized.
1606 * 2. Or it contains FIN. (already checked by caller)
1607 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1608 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1609 * With Minshall's modification: all sent small packets are ACKed.
1610 */
1611 static inline bool tcp_nagle_check(const struct tcp_sock *tp,
1612 const struct sk_buff *skb,
1613 unsigned int mss_now, int nonagle)
1614 {
1615 return skb->len < mss_now &&
1616 ((nonagle & TCP_NAGLE_CORK) ||
1617 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1618 }
1619
1620 /* Return true if the Nagle test allows this packet to be
1621 * sent now.
1622 */
1623 static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1624 unsigned int cur_mss, int nonagle)
1625 {
1626 /* Nagle rule does not apply to frames, which sit in the middle of the
1627 * write_queue (they have no chances to get new data).
1628 *
1629 * This is implemented in the callers, where they modify the 'nonagle'
1630 * argument based upon the location of SKB in the send queue.
1631 */
1632 if (nonagle & TCP_NAGLE_PUSH)
1633 return true;
1634
1635 /* Don't use the nagle rule for urgent data (or for the final FIN).
1636 * Nagle can be ignored during F-RTO too (see RFC4138).
1637 */
1638 if (tcp_urg_mode(tp) || (tp->frto_counter == 2) ||
1639 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1640 return true;
1641
1642 if (!tcp_nagle_check(tp, skb, cur_mss, nonagle))
1643 return true;
1644
1645 return false;
1646 }
1647
1648 /* Does at least the first segment of SKB fit into the send window? */
1649 static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1650 const struct sk_buff *skb,
1651 unsigned int cur_mss)
1652 {
1653 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1654
1655 if (skb->len > cur_mss)
1656 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1657
1658 return !after(end_seq, tcp_wnd_end(tp));
1659 }
1660
1661 /* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1662 * should be put on the wire right now. If so, it returns the number of
1663 * packets allowed by the congestion window.
1664 */
1665 static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1666 unsigned int cur_mss, int nonagle)
1667 {
1668 const struct tcp_sock *tp = tcp_sk(sk);
1669 unsigned int cwnd_quota;
1670
1671 tcp_init_tso_segs(sk, skb, cur_mss);
1672
1673 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1674 return 0;
1675
1676 cwnd_quota = tcp_cwnd_test(tp, skb);
1677 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1678 cwnd_quota = 0;
1679
1680 return cwnd_quota;
1681 }
1682
1683 /* Test if sending is allowed right now. */
1684 bool tcp_may_send_now(struct sock *sk)
1685 {
1686 const struct tcp_sock *tp = tcp_sk(sk);
1687 struct sk_buff *skb = tcp_send_head(sk);
1688
1689 return skb &&
1690 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1691 (tcp_skb_is_last(sk, skb) ?
1692 tp->nonagle : TCP_NAGLE_PUSH));
1693 }
1694
1695 /* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1696 * which is put after SKB on the list. It is very much like
1697 * tcp_fragment() except that it may make several kinds of assumptions
1698 * in order to speed up the splitting operation. In particular, we
1699 * know that all the data is in scatter-gather pages, and that the
1700 * packet has never been sent out before (and thus is not cloned).
1701 */
1702 static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1703 unsigned int mss_now, gfp_t gfp)
1704 {
1705 struct sk_buff *buff;
1706 int nlen = skb->len - len;
1707 u8 flags;
1708
1709 /* All of a TSO frame must be composed of paged data. */
1710 if (skb->len != skb->data_len)
1711 return tcp_fragment(sk, skb, len, mss_now);
1712
1713 buff = sk_stream_alloc_skb(sk, 0, gfp);
1714 if (unlikely(buff == NULL))
1715 return -ENOMEM;
1716
1717 sk->sk_wmem_queued += buff->truesize;
1718 sk_mem_charge(sk, buff->truesize);
1719 buff->truesize += nlen;
1720 skb->truesize -= nlen;
1721
1722 /* Correct the sequence numbers. */
1723 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1724 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1725 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1726
1727 /* PSH and FIN should only be set in the second packet. */
1728 flags = TCP_SKB_CB(skb)->tcp_flags;
1729 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1730 TCP_SKB_CB(buff)->tcp_flags = flags;
1731
1732 /* This packet was never sent out yet, so no SACK bits. */
1733 TCP_SKB_CB(buff)->sacked = 0;
1734
1735 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1736 skb_split(skb, buff, len);
1737
1738 /* Fix up tso_factor for both original and new SKB. */
1739 tcp_set_skb_tso_segs(sk, skb, mss_now);
1740 tcp_set_skb_tso_segs(sk, buff, mss_now);
1741
1742 /* Link BUFF into the send queue. */
1743 skb_header_release(buff);
1744 tcp_insert_write_queue_after(skb, buff, sk);
1745
1746 return 0;
1747 }
1748
1749 /* Try to defer sending, if possible, in order to minimize the amount
1750 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1751 *
1752 * This algorithm is from John Heffner.
1753 */
1754 static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb)
1755 {
1756 struct tcp_sock *tp = tcp_sk(sk);
1757 const struct inet_connection_sock *icsk = inet_csk(sk);
1758 u32 send_win, cong_win, limit, in_flight;
1759 int win_divisor;
1760
1761 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1762 goto send_now;
1763
1764 if (icsk->icsk_ca_state != TCP_CA_Open)
1765 goto send_now;
1766
1767 /* Defer for less than two clock ticks. */
1768 if (tp->tso_deferred &&
1769 (((u32)jiffies << 1) >> 1) - (tp->tso_deferred >> 1) > 1)
1770 goto send_now;
1771
1772 in_flight = tcp_packets_in_flight(tp);
1773
1774 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1775
1776 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1777
1778 /* From in_flight test above, we know that cwnd > in_flight. */
1779 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1780
1781 limit = min(send_win, cong_win);
1782
1783 /* If a full-sized TSO skb can be sent, do it. */
1784 if (limit >= min_t(unsigned int, sk->sk_gso_max_size,
1785 sk->sk_gso_max_segs * tp->mss_cache))
1786 goto send_now;
1787
1788 /* Middle in queue won't get any more data, full sendable already? */
1789 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1790 goto send_now;
1791
1792 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1793 if (win_divisor) {
1794 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1795
1796 /* If at least some fraction of a window is available,
1797 * just use it.
1798 */
1799 chunk /= win_divisor;
1800 if (limit >= chunk)
1801 goto send_now;
1802 } else {
1803 /* Different approach, try not to defer past a single
1804 * ACK. Receiver should ACK every other full sized
1805 * frame, so if we have space for more than 3 frames
1806 * then send now.
1807 */
1808 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1809 goto send_now;
1810 }
1811
1812 /* Ok, it looks like it is advisable to defer. */
1813 tp->tso_deferred = 1 | (jiffies << 1);
1814
1815 return true;
1816
1817 send_now:
1818 tp->tso_deferred = 0;
1819 return false;
1820 }
1821
1822 /* Create a new MTU probe if we are ready.
1823 * MTU probe is regularly attempting to increase the path MTU by
1824 * deliberately sending larger packets. This discovers routing
1825 * changes resulting in larger path MTUs.
1826 *
1827 * Returns 0 if we should wait to probe (no cwnd available),
1828 * 1 if a probe was sent,
1829 * -1 otherwise
1830 */
1831 static int tcp_mtu_probe(struct sock *sk)
1832 {
1833 struct tcp_sock *tp = tcp_sk(sk);
1834 struct inet_connection_sock *icsk = inet_csk(sk);
1835 struct sk_buff *skb, *nskb, *next;
1836 int len;
1837 int probe_size;
1838 int size_needed;
1839 int copy;
1840 int mss_now;
1841
1842 /* Not currently probing/verifying,
1843 * not in recovery,
1844 * have enough cwnd, and
1845 * not SACKing (the variable headers throw things off) */
1846 if (!icsk->icsk_mtup.enabled ||
1847 icsk->icsk_mtup.probe_size ||
1848 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1849 tp->snd_cwnd < 11 ||
1850 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1851 return -1;
1852
1853 /* Very simple search strategy: just double the MSS. */
1854 mss_now = tcp_current_mss(sk);
1855 probe_size = 2 * tp->mss_cache;
1856 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1857 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high)) {
1858 /* TODO: set timer for probe_converge_event */
1859 return -1;
1860 }
1861
1862 /* Have enough data in the send queue to probe? */
1863 if (tp->write_seq - tp->snd_nxt < size_needed)
1864 return -1;
1865
1866 if (tp->snd_wnd < size_needed)
1867 return -1;
1868 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1869 return 0;
1870
1871 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1872 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1873 if (!tcp_packets_in_flight(tp))
1874 return -1;
1875 else
1876 return 0;
1877 }
1878
1879 /* We're allowed to probe. Build it now. */
1880 if ((nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC)) == NULL)
1881 return -1;
1882 sk->sk_wmem_queued += nskb->truesize;
1883 sk_mem_charge(sk, nskb->truesize);
1884
1885 skb = tcp_send_head(sk);
1886
1887 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1888 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1889 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1890 TCP_SKB_CB(nskb)->sacked = 0;
1891 nskb->csum = 0;
1892 nskb->ip_summed = skb->ip_summed;
1893
1894 tcp_insert_write_queue_before(nskb, skb, sk);
1895
1896 len = 0;
1897 tcp_for_write_queue_from_safe(skb, next, sk) {
1898 copy = min_t(int, skb->len, probe_size - len);
1899 if (nskb->ip_summed)
1900 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1901 else
1902 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1903 skb_put(nskb, copy),
1904 copy, nskb->csum);
1905
1906 if (skb->len <= copy) {
1907 /* We've eaten all the data from this skb.
1908 * Throw it away. */
1909 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1910 tcp_unlink_write_queue(skb, sk);
1911 sk_wmem_free_skb(sk, skb);
1912 } else {
1913 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1914 ~(TCPHDR_FIN|TCPHDR_PSH);
1915 if (!skb_shinfo(skb)->nr_frags) {
1916 skb_pull(skb, copy);
1917 if (skb->ip_summed != CHECKSUM_PARTIAL)
1918 skb->csum = csum_partial(skb->data,
1919 skb->len, 0);
1920 } else {
1921 __pskb_trim_head(skb, copy);
1922 tcp_set_skb_tso_segs(sk, skb, mss_now);
1923 }
1924 TCP_SKB_CB(skb)->seq += copy;
1925 }
1926
1927 len += copy;
1928
1929 if (len >= probe_size)
1930 break;
1931 }
1932 tcp_init_tso_segs(sk, nskb, nskb->len);
1933
1934 /* We're ready to send. If this fails, the probe will
1935 * be resegmented into mss-sized pieces by tcp_write_xmit(). */
1936 TCP_SKB_CB(nskb)->when = tcp_time_stamp;
1937 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1938 /* Decrement cwnd here because we are sending
1939 * effectively two packets. */
1940 tp->snd_cwnd--;
1941 tcp_event_new_data_sent(sk, nskb);
1942
1943 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1944 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1945 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1946
1947 return 1;
1948 }
1949
1950 return -1;
1951 }
1952
1953 /* This routine writes packets to the network. It advances the
1954 * send_head. This happens as incoming acks open up the remote
1955 * window for us.
1956 *
1957 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
1958 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
1959 * account rare use of URG, this is not a big flaw.
1960 *
1961 * Returns true, if no segments are in flight and we have queued segments,
1962 * but cannot send anything now because of SWS or another problem.
1963 */
1964 static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
1965 int push_one, gfp_t gfp)
1966 {
1967 struct tcp_sock *tp = tcp_sk(sk);
1968 struct sk_buff *skb;
1969 unsigned int tso_segs, sent_pkts;
1970 int cwnd_quota;
1971 int result;
1972
1973 sent_pkts = 0;
1974
1975 if (!push_one) {
1976 /* Do MTU probing. */
1977 result = tcp_mtu_probe(sk);
1978 if (!result) {
1979 return false;
1980 } else if (result > 0) {
1981 sent_pkts = 1;
1982 }
1983 }
1984
1985 while ((skb = tcp_send_head(sk))) {
1986 unsigned int limit;
1987
1988
1989 tso_segs = tcp_init_tso_segs(sk, skb, mss_now);
1990 BUG_ON(!tso_segs);
1991
1992 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE)
1993 goto repair; /* Skip network transmission */
1994
1995 cwnd_quota = tcp_cwnd_test(tp, skb);
1996 if (!cwnd_quota)
1997 break;
1998
1999 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
2000 break;
2001
2002 if (tso_segs == 1) {
2003 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2004 (tcp_skb_is_last(sk, skb) ?
2005 nonagle : TCP_NAGLE_PUSH))))
2006 break;
2007 } else {
2008 if (!push_one && tcp_tso_should_defer(sk, skb))
2009 break;
2010 }
2011
2012 /* TSQ : sk_wmem_alloc accounts skb truesize,
2013 * including skb overhead. But thats OK.
2014 */
2015 if (atomic_read(&sk->sk_wmem_alloc) >= sysctl_tcp_limit_output_bytes) {
2016 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
2017 break;
2018 }
2019 limit = mss_now;
2020 if (tso_segs > 1 && !tcp_urg_mode(tp))
2021 limit = tcp_mss_split_point(sk, skb, mss_now,
2022 min_t(unsigned int,
2023 cwnd_quota,
2024 sk->sk_gso_max_segs));
2025
2026 if (skb->len > limit &&
2027 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2028 break;
2029
2030 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2031
2032 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2033 break;
2034
2035 repair:
2036 /* Advance the send_head. This one is sent out.
2037 * This call will increment packets_out.
2038 */
2039 tcp_event_new_data_sent(sk, skb);
2040
2041 tcp_minshall_update(tp, mss_now, skb);
2042 sent_pkts += tcp_skb_pcount(skb);
2043
2044 if (push_one)
2045 break;
2046 }
2047
2048 if (likely(sent_pkts)) {
2049 if (tcp_in_cwnd_reduction(sk))
2050 tp->prr_out += sent_pkts;
2051 tcp_cwnd_validate(sk);
2052 return false;
2053 }
2054 return !tp->packets_out && tcp_send_head(sk);
2055 }
2056
2057 /* Push out any pending frames which were held back due to
2058 * TCP_CORK or attempt at coalescing tiny packets.
2059 * The socket must be locked by the caller.
2060 */
2061 void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2062 int nonagle)
2063 {
2064 /* If we are closed, the bytes will have to remain here.
2065 * In time closedown will finish, we empty the write queue and
2066 * all will be happy.
2067 */
2068 if (unlikely(sk->sk_state == TCP_CLOSE))
2069 return;
2070
2071 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2072 sk_gfp_atomic(sk, GFP_ATOMIC)))
2073 tcp_check_probe_timer(sk);
2074 }
2075
2076 /* Send _single_ skb sitting at the send head. This function requires
2077 * true push pending frames to setup probe timer etc.
2078 */
2079 void tcp_push_one(struct sock *sk, unsigned int mss_now)
2080 {
2081 struct sk_buff *skb = tcp_send_head(sk);
2082
2083 BUG_ON(!skb || skb->len < mss_now);
2084
2085 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2086 }
2087
2088 /* This function returns the amount that we can raise the
2089 * usable window based on the following constraints
2090 *
2091 * 1. The window can never be shrunk once it is offered (RFC 793)
2092 * 2. We limit memory per socket
2093 *
2094 * RFC 1122:
2095 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2096 * RECV.NEXT + RCV.WIN fixed until:
2097 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2098 *
2099 * i.e. don't raise the right edge of the window until you can raise
2100 * it at least MSS bytes.
2101 *
2102 * Unfortunately, the recommended algorithm breaks header prediction,
2103 * since header prediction assumes th->window stays fixed.
2104 *
2105 * Strictly speaking, keeping th->window fixed violates the receiver
2106 * side SWS prevention criteria. The problem is that under this rule
2107 * a stream of single byte packets will cause the right side of the
2108 * window to always advance by a single byte.
2109 *
2110 * Of course, if the sender implements sender side SWS prevention
2111 * then this will not be a problem.
2112 *
2113 * BSD seems to make the following compromise:
2114 *
2115 * If the free space is less than the 1/4 of the maximum
2116 * space available and the free space is less than 1/2 mss,
2117 * then set the window to 0.
2118 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2119 * Otherwise, just prevent the window from shrinking
2120 * and from being larger than the largest representable value.
2121 *
2122 * This prevents incremental opening of the window in the regime
2123 * where TCP is limited by the speed of the reader side taking
2124 * data out of the TCP receive queue. It does nothing about
2125 * those cases where the window is constrained on the sender side
2126 * because the pipeline is full.
2127 *
2128 * BSD also seems to "accidentally" limit itself to windows that are a
2129 * multiple of MSS, at least until the free space gets quite small.
2130 * This would appear to be a side effect of the mbuf implementation.
2131 * Combining these two algorithms results in the observed behavior
2132 * of having a fixed window size at almost all times.
2133 *
2134 * Below we obtain similar behavior by forcing the offered window to
2135 * a multiple of the mss when it is feasible to do so.
2136 *
2137 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2138 * Regular options like TIMESTAMP are taken into account.
2139 */
2140 u32 __tcp_select_window(struct sock *sk)
2141 {
2142 struct inet_connection_sock *icsk = inet_csk(sk);
2143 struct tcp_sock *tp = tcp_sk(sk);
2144 /* MSS for the peer's data. Previous versions used mss_clamp
2145 * here. I don't know if the value based on our guesses
2146 * of peer's MSS is better for the performance. It's more correct
2147 * but may be worse for the performance because of rcv_mss
2148 * fluctuations. --SAW 1998/11/1
2149 */
2150 int mss = icsk->icsk_ack.rcv_mss;
2151 int free_space = tcp_space(sk);
2152 int full_space = min_t(int, tp->window_clamp, tcp_full_space(sk));
2153 int window;
2154
2155 if (mss > full_space)
2156 mss = full_space;
2157
2158 if (free_space < (full_space >> 1)) {
2159 icsk->icsk_ack.quick = 0;
2160
2161 if (sk_under_memory_pressure(sk))
2162 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2163 4U * tp->advmss);
2164
2165 if (free_space < mss)
2166 return 0;
2167 }
2168
2169 if (free_space > tp->rcv_ssthresh)
2170 free_space = tp->rcv_ssthresh;
2171
2172 /* Don't do rounding if we are using window scaling, since the
2173 * scaled window will not line up with the MSS boundary anyway.
2174 */
2175 window = tp->rcv_wnd;
2176 if (tp->rx_opt.rcv_wscale) {
2177 window = free_space;
2178
2179 /* Advertise enough space so that it won't get scaled away.
2180 * Import case: prevent zero window announcement if
2181 * 1<<rcv_wscale > mss.
2182 */
2183 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2184 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2185 << tp->rx_opt.rcv_wscale);
2186 } else {
2187 /* Get the largest window that is a nice multiple of mss.
2188 * Window clamp already applied above.
2189 * If our current window offering is within 1 mss of the
2190 * free space we just keep it. This prevents the divide
2191 * and multiply from happening most of the time.
2192 * We also don't do any window rounding when the free space
2193 * is too small.
2194 */
2195 if (window <= free_space - mss || window > free_space)
2196 window = (free_space / mss) * mss;
2197 else if (mss == full_space &&
2198 free_space > window + (full_space >> 1))
2199 window = free_space;
2200 }
2201
2202 return window;
2203 }
2204
2205 /* Collapses two adjacent SKB's during retransmission. */
2206 static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2207 {
2208 struct tcp_sock *tp = tcp_sk(sk);
2209 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2210 int skb_size, next_skb_size;
2211
2212 skb_size = skb->len;
2213 next_skb_size = next_skb->len;
2214
2215 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2216
2217 tcp_highest_sack_combine(sk, next_skb, skb);
2218
2219 tcp_unlink_write_queue(next_skb, sk);
2220
2221 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2222 next_skb_size);
2223
2224 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2225 skb->ip_summed = CHECKSUM_PARTIAL;
2226
2227 if (skb->ip_summed != CHECKSUM_PARTIAL)
2228 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2229
2230 /* Update sequence range on original skb. */
2231 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2232
2233 /* Merge over control information. This moves PSH/FIN etc. over */
2234 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2235
2236 /* All done, get rid of second SKB and account for it so
2237 * packet counting does not break.
2238 */
2239 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2240
2241 /* changed transmit queue under us so clear hints */
2242 tcp_clear_retrans_hints_partial(tp);
2243 if (next_skb == tp->retransmit_skb_hint)
2244 tp->retransmit_skb_hint = skb;
2245
2246 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2247
2248 sk_wmem_free_skb(sk, next_skb);
2249 }
2250
2251 /* Check if coalescing SKBs is legal. */
2252 static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2253 {
2254 if (tcp_skb_pcount(skb) > 1)
2255 return false;
2256 /* TODO: SACK collapsing could be used to remove this condition */
2257 if (skb_shinfo(skb)->nr_frags != 0)
2258 return false;
2259 if (skb_cloned(skb))
2260 return false;
2261 if (skb == tcp_send_head(sk))
2262 return false;
2263 /* Some heurestics for collapsing over SACK'd could be invented */
2264 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2265 return false;
2266
2267 return true;
2268 }
2269
2270 /* Collapse packets in the retransmit queue to make to create
2271 * less packets on the wire. This is only done on retransmission.
2272 */
2273 static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2274 int space)
2275 {
2276 struct tcp_sock *tp = tcp_sk(sk);
2277 struct sk_buff *skb = to, *tmp;
2278 bool first = true;
2279
2280 if (!sysctl_tcp_retrans_collapse)
2281 return;
2282 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2283 return;
2284
2285 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2286 if (!tcp_can_collapse(sk, skb))
2287 break;
2288
2289 space -= skb->len;
2290
2291 if (first) {
2292 first = false;
2293 continue;
2294 }
2295
2296 if (space < 0)
2297 break;
2298 /* Punt if not enough space exists in the first SKB for
2299 * the data in the second
2300 */
2301 if (skb->len > skb_availroom(to))
2302 break;
2303
2304 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2305 break;
2306
2307 tcp_collapse_retrans(sk, to);
2308 }
2309 }
2310
2311 /* This retransmits one SKB. Policy decisions and retransmit queue
2312 * state updates are done by the caller. Returns non-zero if an
2313 * error occurred which prevented the send.
2314 */
2315 int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2316 {
2317 struct tcp_sock *tp = tcp_sk(sk);
2318 struct inet_connection_sock *icsk = inet_csk(sk);
2319 unsigned int cur_mss;
2320
2321 /* Inconslusive MTU probe */
2322 if (icsk->icsk_mtup.probe_size) {
2323 icsk->icsk_mtup.probe_size = 0;
2324 }
2325
2326 /* Do not sent more than we queued. 1/4 is reserved for possible
2327 * copying overhead: fragmentation, tunneling, mangling etc.
2328 */
2329 if (atomic_read(&sk->sk_wmem_alloc) >
2330 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2331 return -EAGAIN;
2332
2333 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2334 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2335 BUG();
2336 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2337 return -ENOMEM;
2338 }
2339
2340 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2341 return -EHOSTUNREACH; /* Routing failure or similar. */
2342
2343 cur_mss = tcp_current_mss(sk);
2344
2345 /* If receiver has shrunk his window, and skb is out of
2346 * new window, do not retransmit it. The exception is the
2347 * case, when window is shrunk to zero. In this case
2348 * our retransmit serves as a zero window probe.
2349 */
2350 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2351 TCP_SKB_CB(skb)->seq != tp->snd_una)
2352 return -EAGAIN;
2353
2354 if (skb->len > cur_mss) {
2355 if (tcp_fragment(sk, skb, cur_mss, cur_mss))
2356 return -ENOMEM; /* We'll try again later. */
2357 } else {
2358 int oldpcount = tcp_skb_pcount(skb);
2359
2360 if (unlikely(oldpcount > 1)) {
2361 tcp_init_tso_segs(sk, skb, cur_mss);
2362 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2363 }
2364 }
2365
2366 tcp_retrans_try_collapse(sk, skb, cur_mss);
2367
2368 /* Some Solaris stacks overoptimize and ignore the FIN on a
2369 * retransmit when old data is attached. So strip it off
2370 * since it is cheap to do so and saves bytes on the network.
2371 */
2372 if (skb->len > 0 &&
2373 (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2374 tp->snd_una == (TCP_SKB_CB(skb)->end_seq - 1)) {
2375 if (!pskb_trim(skb, 0)) {
2376 /* Reuse, even though it does some unnecessary work */
2377 tcp_init_nondata_skb(skb, TCP_SKB_CB(skb)->end_seq - 1,
2378 TCP_SKB_CB(skb)->tcp_flags);
2379 skb->ip_summed = CHECKSUM_NONE;
2380 }
2381 }
2382
2383 /* Make a copy, if the first transmission SKB clone we made
2384 * is still in somebody's hands, else make a clone.
2385 */
2386 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2387
2388 /* make sure skb->data is aligned on arches that require it */
2389 if (unlikely(NET_IP_ALIGN && ((unsigned long)skb->data & 3))) {
2390 struct sk_buff *nskb = __pskb_copy(skb, MAX_TCP_HEADER,
2391 GFP_ATOMIC);
2392 return nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2393 -ENOBUFS;
2394 } else {
2395 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2396 }
2397 }
2398
2399 int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2400 {
2401 struct tcp_sock *tp = tcp_sk(sk);
2402 int err = __tcp_retransmit_skb(sk, skb);
2403
2404 if (err == 0) {
2405 /* Update global TCP statistics. */
2406 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2407
2408 tp->total_retrans++;
2409
2410 #if FASTRETRANS_DEBUG > 0
2411 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2412 net_dbg_ratelimited("retrans_out leaked\n");
2413 }
2414 #endif
2415 if (!tp->retrans_out)
2416 tp->lost_retrans_low = tp->snd_nxt;
2417 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2418 tp->retrans_out += tcp_skb_pcount(skb);
2419
2420 /* Save stamp of the first retransmit. */
2421 if (!tp->retrans_stamp)
2422 tp->retrans_stamp = TCP_SKB_CB(skb)->when;
2423
2424 tp->undo_retrans += tcp_skb_pcount(skb);
2425
2426 /* snd_nxt is stored to detect loss of retransmitted segment,
2427 * see tcp_input.c tcp_sacktag_write_queue().
2428 */
2429 TCP_SKB_CB(skb)->ack_seq = tp->snd_nxt;
2430 }
2431 return err;
2432 }
2433
2434 /* Check if we forward retransmits are possible in the current
2435 * window/congestion state.
2436 */
2437 static bool tcp_can_forward_retransmit(struct sock *sk)
2438 {
2439 const struct inet_connection_sock *icsk = inet_csk(sk);
2440 const struct tcp_sock *tp = tcp_sk(sk);
2441
2442 /* Forward retransmissions are possible only during Recovery. */
2443 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2444 return false;
2445
2446 /* No forward retransmissions in Reno are possible. */
2447 if (tcp_is_reno(tp))
2448 return false;
2449
2450 /* Yeah, we have to make difficult choice between forward transmission
2451 * and retransmission... Both ways have their merits...
2452 *
2453 * For now we do not retransmit anything, while we have some new
2454 * segments to send. In the other cases, follow rule 3 for
2455 * NextSeg() specified in RFC3517.
2456 */
2457
2458 if (tcp_may_send_now(sk))
2459 return false;
2460
2461 return true;
2462 }
2463
2464 /* This gets called after a retransmit timeout, and the initially
2465 * retransmitted data is acknowledged. It tries to continue
2466 * resending the rest of the retransmit queue, until either
2467 * we've sent it all or the congestion window limit is reached.
2468 * If doing SACK, the first ACK which comes back for a timeout
2469 * based retransmit packet might feed us FACK information again.
2470 * If so, we use it to avoid unnecessarily retransmissions.
2471 */
2472 void tcp_xmit_retransmit_queue(struct sock *sk)
2473 {
2474 const struct inet_connection_sock *icsk = inet_csk(sk);
2475 struct tcp_sock *tp = tcp_sk(sk);
2476 struct sk_buff *skb;
2477 struct sk_buff *hole = NULL;
2478 u32 last_lost;
2479 int mib_idx;
2480 int fwd_rexmitting = 0;
2481
2482 if (!tp->packets_out)
2483 return;
2484
2485 if (!tp->lost_out)
2486 tp->retransmit_high = tp->snd_una;
2487
2488 if (tp->retransmit_skb_hint) {
2489 skb = tp->retransmit_skb_hint;
2490 last_lost = TCP_SKB_CB(skb)->end_seq;
2491 if (after(last_lost, tp->retransmit_high))
2492 last_lost = tp->retransmit_high;
2493 } else {
2494 skb = tcp_write_queue_head(sk);
2495 last_lost = tp->snd_una;
2496 }
2497
2498 tcp_for_write_queue_from(skb, sk) {
2499 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2500
2501 if (skb == tcp_send_head(sk))
2502 break;
2503 /* we could do better than to assign each time */
2504 if (hole == NULL)
2505 tp->retransmit_skb_hint = skb;
2506
2507 /* Assume this retransmit will generate
2508 * only one packet for congestion window
2509 * calculation purposes. This works because
2510 * tcp_retransmit_skb() will chop up the
2511 * packet to be MSS sized and all the
2512 * packet counting works out.
2513 */
2514 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2515 return;
2516
2517 if (fwd_rexmitting) {
2518 begin_fwd:
2519 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2520 break;
2521 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2522
2523 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2524 tp->retransmit_high = last_lost;
2525 if (!tcp_can_forward_retransmit(sk))
2526 break;
2527 /* Backtrack if necessary to non-L'ed skb */
2528 if (hole != NULL) {
2529 skb = hole;
2530 hole = NULL;
2531 }
2532 fwd_rexmitting = 1;
2533 goto begin_fwd;
2534
2535 } else if (!(sacked & TCPCB_LOST)) {
2536 if (hole == NULL && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2537 hole = skb;
2538 continue;
2539
2540 } else {
2541 last_lost = TCP_SKB_CB(skb)->end_seq;
2542 if (icsk->icsk_ca_state != TCP_CA_Loss)
2543 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2544 else
2545 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2546 }
2547
2548 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2549 continue;
2550
2551 if (tcp_retransmit_skb(sk, skb)) {
2552 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2553 return;
2554 }
2555 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2556
2557 if (tcp_in_cwnd_reduction(sk))
2558 tp->prr_out += tcp_skb_pcount(skb);
2559
2560 if (skb == tcp_write_queue_head(sk))
2561 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2562 inet_csk(sk)->icsk_rto,
2563 TCP_RTO_MAX);
2564 }
2565 }
2566
2567 /* Send a fin. The caller locks the socket for us. This cannot be
2568 * allowed to fail queueing a FIN frame under any circumstances.
2569 */
2570 void tcp_send_fin(struct sock *sk)
2571 {
2572 struct tcp_sock *tp = tcp_sk(sk);
2573 struct sk_buff *skb = tcp_write_queue_tail(sk);
2574 int mss_now;
2575
2576 /* Optimization, tack on the FIN if we have a queue of
2577 * unsent frames. But be careful about outgoing SACKS
2578 * and IP options.
2579 */
2580 mss_now = tcp_current_mss(sk);
2581
2582 if (tcp_send_head(sk) != NULL) {
2583 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_FIN;
2584 TCP_SKB_CB(skb)->end_seq++;
2585 tp->write_seq++;
2586 } else {
2587 /* Socket is locked, keep trying until memory is available. */
2588 for (;;) {
2589 skb = alloc_skb_fclone(MAX_TCP_HEADER,
2590 sk->sk_allocation);
2591 if (skb)
2592 break;
2593 yield();
2594 }
2595
2596 /* Reserve space for headers and prepare control bits. */
2597 skb_reserve(skb, MAX_TCP_HEADER);
2598 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2599 tcp_init_nondata_skb(skb, tp->write_seq,
2600 TCPHDR_ACK | TCPHDR_FIN);
2601 tcp_queue_skb(sk, skb);
2602 }
2603 __tcp_push_pending_frames(sk, mss_now, TCP_NAGLE_OFF);
2604 }
2605
2606 /* We get here when a process closes a file descriptor (either due to
2607 * an explicit close() or as a byproduct of exit()'ing) and there
2608 * was unread data in the receive queue. This behavior is recommended
2609 * by RFC 2525, section 2.17. -DaveM
2610 */
2611 void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2612 {
2613 struct sk_buff *skb;
2614
2615 /* NOTE: No TCP options attached and we never retransmit this. */
2616 skb = alloc_skb(MAX_TCP_HEADER, priority);
2617 if (!skb) {
2618 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2619 return;
2620 }
2621
2622 /* Reserve space for headers and prepare control bits. */
2623 skb_reserve(skb, MAX_TCP_HEADER);
2624 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2625 TCPHDR_ACK | TCPHDR_RST);
2626 /* Send it off. */
2627 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2628 if (tcp_transmit_skb(sk, skb, 0, priority))
2629 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2630
2631 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2632 }
2633
2634 /* Send a crossed SYN-ACK during socket establishment.
2635 * WARNING: This routine must only be called when we have already sent
2636 * a SYN packet that crossed the incoming SYN that caused this routine
2637 * to get called. If this assumption fails then the initial rcv_wnd
2638 * and rcv_wscale values will not be correct.
2639 */
2640 int tcp_send_synack(struct sock *sk)
2641 {
2642 struct sk_buff *skb;
2643
2644 skb = tcp_write_queue_head(sk);
2645 if (skb == NULL || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2646 pr_debug("%s: wrong queue state\n", __func__);
2647 return -EFAULT;
2648 }
2649 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2650 if (skb_cloned(skb)) {
2651 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2652 if (nskb == NULL)
2653 return -ENOMEM;
2654 tcp_unlink_write_queue(skb, sk);
2655 skb_header_release(nskb);
2656 __tcp_add_write_queue_head(sk, nskb);
2657 sk_wmem_free_skb(sk, skb);
2658 sk->sk_wmem_queued += nskb->truesize;
2659 sk_mem_charge(sk, nskb->truesize);
2660 skb = nskb;
2661 }
2662
2663 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2664 TCP_ECN_send_synack(tcp_sk(sk), skb);
2665 }
2666 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2667 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2668 }
2669
2670 /**
2671 * tcp_make_synack - Prepare a SYN-ACK.
2672 * sk: listener socket
2673 * dst: dst entry attached to the SYNACK
2674 * req: request_sock pointer
2675 * rvp: request_values pointer
2676 *
2677 * Allocate one skb and build a SYNACK packet.
2678 * @dst is consumed : Caller should not use it again.
2679 */
2680 struct sk_buff *tcp_make_synack(struct sock *sk, struct dst_entry *dst,
2681 struct request_sock *req,
2682 struct request_values *rvp,
2683 struct tcp_fastopen_cookie *foc)
2684 {
2685 struct tcp_out_options opts;
2686 struct tcp_extend_values *xvp = tcp_xv(rvp);
2687 struct inet_request_sock *ireq = inet_rsk(req);
2688 struct tcp_sock *tp = tcp_sk(sk);
2689 const struct tcp_cookie_values *cvp = tp->cookie_values;
2690 struct tcphdr *th;
2691 struct sk_buff *skb;
2692 struct tcp_md5sig_key *md5;
2693 int tcp_header_size;
2694 int mss;
2695 int s_data_desired = 0;
2696
2697 if (cvp != NULL && cvp->s_data_constant && cvp->s_data_desired)
2698 s_data_desired = cvp->s_data_desired;
2699 skb = alloc_skb(MAX_TCP_HEADER + 15 + s_data_desired,
2700 sk_gfp_atomic(sk, GFP_ATOMIC));
2701 if (unlikely(!skb)) {
2702 dst_release(dst);
2703 return NULL;
2704 }
2705 /* Reserve space for headers. */
2706 skb_reserve(skb, MAX_TCP_HEADER);
2707
2708 skb_dst_set(skb, dst);
2709
2710 mss = dst_metric_advmss(dst);
2711 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < mss)
2712 mss = tp->rx_opt.user_mss;
2713
2714 if (req->rcv_wnd == 0) { /* ignored for retransmitted syns */
2715 __u8 rcv_wscale;
2716 /* Set this up on the first call only */
2717 req->window_clamp = tp->window_clamp ? : dst_metric(dst, RTAX_WINDOW);
2718
2719 /* limit the window selection if the user enforce a smaller rx buffer */
2720 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2721 (req->window_clamp > tcp_full_space(sk) || req->window_clamp == 0))
2722 req->window_clamp = tcp_full_space(sk);
2723
2724 /* tcp_full_space because it is guaranteed to be the first packet */
2725 tcp_select_initial_window(tcp_full_space(sk),
2726 mss - (ireq->tstamp_ok ? TCPOLEN_TSTAMP_ALIGNED : 0),
2727 &req->rcv_wnd,
2728 &req->window_clamp,
2729 ireq->wscale_ok,
2730 &rcv_wscale,
2731 dst_metric(dst, RTAX_INITRWND));
2732 ireq->rcv_wscale = rcv_wscale;
2733 }
2734
2735 memset(&opts, 0, sizeof(opts));
2736 #ifdef CONFIG_SYN_COOKIES
2737 if (unlikely(req->cookie_ts))
2738 TCP_SKB_CB(skb)->when = cookie_init_timestamp(req);
2739 else
2740 #endif
2741 TCP_SKB_CB(skb)->when = tcp_time_stamp;
2742 tcp_header_size = tcp_synack_options(sk, req, mss,
2743 skb, &opts, &md5, xvp, foc)
2744 + sizeof(*th);
2745
2746 skb_push(skb, tcp_header_size);
2747 skb_reset_transport_header(skb);
2748
2749 th = tcp_hdr(skb);
2750 memset(th, 0, sizeof(struct tcphdr));
2751 th->syn = 1;
2752 th->ack = 1;
2753 TCP_ECN_make_synack(req, th);
2754 th->source = ireq->loc_port;
2755 th->dest = ireq->rmt_port;
2756 /* Setting of flags are superfluous here for callers (and ECE is
2757 * not even correctly set)
2758 */
2759 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
2760 TCPHDR_SYN | TCPHDR_ACK);
2761
2762 if (OPTION_COOKIE_EXTENSION & opts.options) {
2763 if (s_data_desired) {
2764 u8 *buf = skb_put(skb, s_data_desired);
2765
2766 /* copy data directly from the listening socket. */
2767 memcpy(buf, cvp->s_data_payload, s_data_desired);
2768 TCP_SKB_CB(skb)->end_seq += s_data_desired;
2769 }
2770
2771 if (opts.hash_size > 0) {
2772 __u32 workspace[SHA_WORKSPACE_WORDS];
2773 u32 *mess = &xvp->cookie_bakery[COOKIE_DIGEST_WORDS];
2774 u32 *tail = &mess[COOKIE_MESSAGE_WORDS-1];
2775
2776 /* Secret recipe depends on the Timestamp, (future)
2777 * Sequence and Acknowledgment Numbers, Initiator
2778 * Cookie, and others handled by IP variant caller.
2779 */
2780 *tail-- ^= opts.tsval;
2781 *tail-- ^= tcp_rsk(req)->rcv_isn + 1;
2782 *tail-- ^= TCP_SKB_CB(skb)->seq + 1;
2783
2784 /* recommended */
2785 *tail-- ^= (((__force u32)th->dest << 16) | (__force u32)th->source);
2786 *tail-- ^= (u32)(unsigned long)cvp; /* per sockopt */
2787
2788 sha_transform((__u32 *)&xvp->cookie_bakery[0],
2789 (char *)mess,
2790 &workspace[0]);
2791 opts.hash_location =
2792 (__u8 *)&xvp->cookie_bakery[0];
2793 }
2794 }
2795
2796 th->seq = htonl(TCP_SKB_CB(skb)->seq);
2797 /* XXX data is queued and acked as is. No buffer/window check */
2798 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
2799
2800 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
2801 th->window = htons(min(req->rcv_wnd, 65535U));
2802 tcp_options_write((__be32 *)(th + 1), tp, &opts);
2803 th->doff = (tcp_header_size >> 2);
2804 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS, tcp_skb_pcount(skb));
2805
2806 #ifdef CONFIG_TCP_MD5SIG
2807 /* Okay, we have all we need - do the md5 hash if needed */
2808 if (md5) {
2809 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
2810 md5, NULL, req, skb);
2811 }
2812 #endif
2813
2814 return skb;
2815 }
2816 EXPORT_SYMBOL(tcp_make_synack);
2817
2818 /* Do all connect socket setups that can be done AF independent. */
2819 void tcp_connect_init(struct sock *sk)
2820 {
2821 const struct dst_entry *dst = __sk_dst_get(sk);
2822 struct tcp_sock *tp = tcp_sk(sk);
2823 __u8 rcv_wscale;
2824
2825 /* We'll fix this up when we get a response from the other end.
2826 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
2827 */
2828 tp->tcp_header_len = sizeof(struct tcphdr) +
2829 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
2830
2831 #ifdef CONFIG_TCP_MD5SIG
2832 if (tp->af_specific->md5_lookup(sk, sk) != NULL)
2833 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
2834 #endif
2835
2836 /* If user gave his TCP_MAXSEG, record it to clamp */
2837 if (tp->rx_opt.user_mss)
2838 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2839 tp->max_window = 0;
2840 tcp_mtup_init(sk);
2841 tcp_sync_mss(sk, dst_mtu(dst));
2842
2843 if (!tp->window_clamp)
2844 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
2845 tp->advmss = dst_metric_advmss(dst);
2846 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
2847 tp->advmss = tp->rx_opt.user_mss;
2848
2849 tcp_initialize_rcv_mss(sk);
2850
2851 /* limit the window selection if the user enforce a smaller rx buffer */
2852 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
2853 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
2854 tp->window_clamp = tcp_full_space(sk);
2855
2856 tcp_select_initial_window(tcp_full_space(sk),
2857 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
2858 &tp->rcv_wnd,
2859 &tp->window_clamp,
2860 sysctl_tcp_window_scaling,
2861 &rcv_wscale,
2862 dst_metric(dst, RTAX_INITRWND));
2863
2864 tp->rx_opt.rcv_wscale = rcv_wscale;
2865 tp->rcv_ssthresh = tp->rcv_wnd;
2866
2867 sk->sk_err = 0;
2868 sock_reset_flag(sk, SOCK_DONE);
2869 tp->snd_wnd = 0;
2870 tcp_init_wl(tp, 0);
2871 tp->snd_una = tp->write_seq;
2872 tp->snd_sml = tp->write_seq;
2873 tp->snd_up = tp->write_seq;
2874 tp->snd_nxt = tp->write_seq;
2875
2876 if (likely(!tp->repair))
2877 tp->rcv_nxt = 0;
2878 tp->rcv_wup = tp->rcv_nxt;
2879 tp->copied_seq = tp->rcv_nxt;
2880
2881 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
2882 inet_csk(sk)->icsk_retransmits = 0;
2883 tcp_clear_retrans(tp);
2884 }
2885
2886 static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
2887 {
2888 struct tcp_sock *tp = tcp_sk(sk);
2889 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
2890
2891 tcb->end_seq += skb->len;
2892 skb_header_release(skb);
2893 __tcp_add_write_queue_tail(sk, skb);
2894 sk->sk_wmem_queued += skb->truesize;
2895 sk_mem_charge(sk, skb->truesize);
2896 tp->write_seq = tcb->end_seq;
2897 tp->packets_out += tcp_skb_pcount(skb);
2898 }
2899
2900 /* Build and send a SYN with data and (cached) Fast Open cookie. However,
2901 * queue a data-only packet after the regular SYN, such that regular SYNs
2902 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
2903 * only the SYN sequence, the data are retransmitted in the first ACK.
2904 * If cookie is not cached or other error occurs, falls back to send a
2905 * regular SYN with Fast Open cookie request option.
2906 */
2907 static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
2908 {
2909 struct tcp_sock *tp = tcp_sk(sk);
2910 struct tcp_fastopen_request *fo = tp->fastopen_req;
2911 int syn_loss = 0, space, i, err = 0, iovlen = fo->data->msg_iovlen;
2912 struct sk_buff *syn_data = NULL, *data;
2913 unsigned long last_syn_loss = 0;
2914
2915 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
2916 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
2917 &syn_loss, &last_syn_loss);
2918 /* Recurring FO SYN losses: revert to regular handshake temporarily */
2919 if (syn_loss > 1 &&
2920 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
2921 fo->cookie.len = -1;
2922 goto fallback;
2923 }
2924
2925 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
2926 fo->cookie.len = -1;
2927 else if (fo->cookie.len <= 0)
2928 goto fallback;
2929
2930 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
2931 * user-MSS. Reserve maximum option space for middleboxes that add
2932 * private TCP options. The cost is reduced data space in SYN :(
2933 */
2934 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
2935 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
2936 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
2937 MAX_TCP_OPTION_SPACE;
2938
2939 syn_data = skb_copy_expand(syn, skb_headroom(syn), space,
2940 sk->sk_allocation);
2941 if (syn_data == NULL)
2942 goto fallback;
2943
2944 for (i = 0; i < iovlen && syn_data->len < space; ++i) {
2945 struct iovec *iov = &fo->data->msg_iov[i];
2946 unsigned char __user *from = iov->iov_base;
2947 int len = iov->iov_len;
2948
2949 if (syn_data->len + len > space)
2950 len = space - syn_data->len;
2951 else if (i + 1 == iovlen)
2952 /* No more data pending in inet_wait_for_connect() */
2953 fo->data = NULL;
2954
2955 if (skb_add_data(syn_data, from, len))
2956 goto fallback;
2957 }
2958
2959 /* Queue a data-only packet after the regular SYN for retransmission */
2960 data = pskb_copy(syn_data, sk->sk_allocation);
2961 if (data == NULL)
2962 goto fallback;
2963 TCP_SKB_CB(data)->seq++;
2964 TCP_SKB_CB(data)->tcp_flags &= ~TCPHDR_SYN;
2965 TCP_SKB_CB(data)->tcp_flags = (TCPHDR_ACK|TCPHDR_PSH);
2966 tcp_connect_queue_skb(sk, data);
2967 fo->copied = data->len;
2968
2969 if (tcp_transmit_skb(sk, syn_data, 0, sk->sk_allocation) == 0) {
2970 tp->syn_data = (fo->copied > 0);
2971 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFASTOPENACTIVE);
2972 goto done;
2973 }
2974 syn_data = NULL;
2975
2976 fallback:
2977 /* Send a regular SYN with Fast Open cookie request option */
2978 if (fo->cookie.len > 0)
2979 fo->cookie.len = 0;
2980 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
2981 if (err)
2982 tp->syn_fastopen = 0;
2983 kfree_skb(syn_data);
2984 done:
2985 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
2986 return err;
2987 }
2988
2989 /* Build a SYN and send it off. */
2990 int tcp_connect(struct sock *sk)
2991 {
2992 struct tcp_sock *tp = tcp_sk(sk);
2993 struct sk_buff *buff;
2994 int err;
2995
2996 tcp_connect_init(sk);
2997
2998 if (unlikely(tp->repair)) {
2999 tcp_finish_connect(sk, NULL);
3000 return 0;
3001 }
3002
3003 buff = alloc_skb_fclone(MAX_TCP_HEADER + 15, sk->sk_allocation);
3004 if (unlikely(buff == NULL))
3005 return -ENOBUFS;
3006
3007 /* Reserve space for headers. */
3008 skb_reserve(buff, MAX_TCP_HEADER);
3009
3010 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3011 tp->retrans_stamp = TCP_SKB_CB(buff)->when = tcp_time_stamp;
3012 tcp_connect_queue_skb(sk, buff);
3013 TCP_ECN_send_syn(sk, buff);
3014
3015 /* Send off SYN; include data in Fast Open. */
3016 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3017 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3018 if (err == -ECONNREFUSED)
3019 return err;
3020
3021 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3022 * in order to make this packet get counted in tcpOutSegs.
3023 */
3024 tp->snd_nxt = tp->write_seq;
3025 tp->pushed_seq = tp->write_seq;
3026 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3027
3028 /* Timer for repeating the SYN until an answer. */
3029 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3030 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3031 return 0;
3032 }
3033 EXPORT_SYMBOL(tcp_connect);
3034
3035 /* Send out a delayed ack, the caller does the policy checking
3036 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3037 * for details.
3038 */
3039 void tcp_send_delayed_ack(struct sock *sk)
3040 {
3041 struct inet_connection_sock *icsk = inet_csk(sk);
3042 int ato = icsk->icsk_ack.ato;
3043 unsigned long timeout;
3044
3045 if (ato > TCP_DELACK_MIN) {
3046 const struct tcp_sock *tp = tcp_sk(sk);
3047 int max_ato = HZ / 2;
3048
3049 if (icsk->icsk_ack.pingpong ||
3050 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3051 max_ato = TCP_DELACK_MAX;
3052
3053 /* Slow path, intersegment interval is "high". */
3054
3055 /* If some rtt estimate is known, use it to bound delayed ack.
3056 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3057 * directly.
3058 */
3059 if (tp->srtt) {
3060 int rtt = max(tp->srtt >> 3, TCP_DELACK_MIN);
3061
3062 if (rtt < max_ato)
3063 max_ato = rtt;
3064 }
3065
3066 ato = min(ato, max_ato);
3067 }
3068
3069 /* Stay within the limit we were given */
3070 timeout = jiffies + ato;
3071
3072 /* Use new timeout only if there wasn't a older one earlier. */
3073 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3074 /* If delack timer was blocked or is about to expire,
3075 * send ACK now.
3076 */
3077 if (icsk->icsk_ack.blocked ||
3078 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3079 tcp_send_ack(sk);
3080 return;
3081 }
3082
3083 if (!time_before(timeout, icsk->icsk_ack.timeout))
3084 timeout = icsk->icsk_ack.timeout;
3085 }
3086 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3087 icsk->icsk_ack.timeout = timeout;
3088 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3089 }
3090
3091 /* This routine sends an ack and also updates the window. */
3092 void tcp_send_ack(struct sock *sk)
3093 {
3094 struct sk_buff *buff;
3095
3096 /* If we have been reset, we may not send again. */
3097 if (sk->sk_state == TCP_CLOSE)
3098 return;
3099
3100 /* We are not putting this on the write queue, so
3101 * tcp_transmit_skb() will set the ownership to this
3102 * sock.
3103 */
3104 buff = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3105 if (buff == NULL) {
3106 inet_csk_schedule_ack(sk);
3107 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3108 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3109 TCP_DELACK_MAX, TCP_RTO_MAX);
3110 return;
3111 }
3112
3113 /* Reserve space for headers and prepare control bits. */
3114 skb_reserve(buff, MAX_TCP_HEADER);
3115 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3116
3117 /* Send it off, this clears delayed acks for us. */
3118 TCP_SKB_CB(buff)->when = tcp_time_stamp;
3119 tcp_transmit_skb(sk, buff, 0, sk_gfp_atomic(sk, GFP_ATOMIC));
3120 }
3121
3122 /* This routine sends a packet with an out of date sequence
3123 * number. It assumes the other end will try to ack it.
3124 *
3125 * Question: what should we make while urgent mode?
3126 * 4.4BSD forces sending single byte of data. We cannot send
3127 * out of window data, because we have SND.NXT==SND.MAX...
3128 *
3129 * Current solution: to send TWO zero-length segments in urgent mode:
3130 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3131 * out-of-date with SND.UNA-1 to probe window.
3132 */
3133 static int tcp_xmit_probe_skb(struct sock *sk, int urgent)
3134 {
3135 struct tcp_sock *tp = tcp_sk(sk);
3136 struct sk_buff *skb;
3137
3138 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3139 skb = alloc_skb(MAX_TCP_HEADER, sk_gfp_atomic(sk, GFP_ATOMIC));
3140 if (skb == NULL)
3141 return -1;
3142
3143 /* Reserve space for headers and set control bits. */
3144 skb_reserve(skb, MAX_TCP_HEADER);
3145 /* Use a previous sequence. This should cause the other
3146 * end to send an ack. Don't queue or clone SKB, just
3147 * send it.
3148 */
3149 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3150 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3151 return tcp_transmit_skb(sk, skb, 0, GFP_ATOMIC);
3152 }
3153
3154 void tcp_send_window_probe(struct sock *sk)
3155 {
3156 if (sk->sk_state == TCP_ESTABLISHED) {
3157 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3158 tcp_sk(sk)->snd_nxt = tcp_sk(sk)->write_seq;
3159 tcp_xmit_probe_skb(sk, 0);
3160 }
3161 }
3162
3163 /* Initiate keepalive or window probe from timer. */
3164 int tcp_write_wakeup(struct sock *sk)
3165 {
3166 struct tcp_sock *tp = tcp_sk(sk);
3167 struct sk_buff *skb;
3168
3169 if (sk->sk_state == TCP_CLOSE)
3170 return -1;
3171
3172 if ((skb = tcp_send_head(sk)) != NULL &&
3173 before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3174 int err;
3175 unsigned int mss = tcp_current_mss(sk);
3176 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3177
3178 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3179 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3180
3181 /* We are probing the opening of a window
3182 * but the window size is != 0
3183 * must have been a result SWS avoidance ( sender )
3184 */
3185 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3186 skb->len > mss) {
3187 seg_size = min(seg_size, mss);
3188 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3189 if (tcp_fragment(sk, skb, seg_size, mss))
3190 return -1;
3191 } else if (!tcp_skb_pcount(skb))
3192 tcp_set_skb_tso_segs(sk, skb, mss);
3193
3194 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3195 TCP_SKB_CB(skb)->when = tcp_time_stamp;
3196 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3197 if (!err)
3198 tcp_event_new_data_sent(sk, skb);
3199 return err;
3200 } else {
3201 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3202 tcp_xmit_probe_skb(sk, 1);
3203 return tcp_xmit_probe_skb(sk, 0);
3204 }
3205 }
3206
3207 /* A window probe timeout has occurred. If window is not closed send
3208 * a partial packet else a zero probe.
3209 */
3210 void tcp_send_probe0(struct sock *sk)
3211 {
3212 struct inet_connection_sock *icsk = inet_csk(sk);
3213 struct tcp_sock *tp = tcp_sk(sk);
3214 int err;
3215
3216 err = tcp_write_wakeup(sk);
3217
3218 if (tp->packets_out || !tcp_send_head(sk)) {
3219 /* Cancel probe timer, if it is not required. */
3220 icsk->icsk_probes_out = 0;
3221 icsk->icsk_backoff = 0;
3222 return;
3223 }
3224
3225 if (err <= 0) {
3226 if (icsk->icsk_backoff < sysctl_tcp_retries2)
3227 icsk->icsk_backoff++;
3228 icsk->icsk_probes_out++;
3229 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3230 min(icsk->icsk_rto << icsk->icsk_backoff, TCP_RTO_MAX),
3231 TCP_RTO_MAX);
3232 } else {
3233 /* If packet was not sent due to local congestion,
3234 * do not backoff and do not remember icsk_probes_out.
3235 * Let local senders to fight for local resources.
3236 *
3237 * Use accumulated backoff yet.
3238 */
3239 if (!icsk->icsk_probes_out)
3240 icsk->icsk_probes_out = 1;
3241 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3242 min(icsk->icsk_rto << icsk->icsk_backoff,
3243 TCP_RESOURCE_PROBE_INTERVAL),
3244 TCP_RTO_MAX);
3245 }
3246 }