Loading...
1/*
2 * INET An implementation of the TCP/IP protocol suite for the LINUX
3 * operating system. INET is implemented using the BSD Socket
4 * interface as the means of communication with the user level.
5 *
6 * Implementation of the Transmission Control Protocol(TCP).
7 *
8 * Authors: Ross Biro
9 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10 * Mark Evans, <evansmp@uhura.aston.ac.uk>
11 * Corey Minyard <wf-rch!minyard@relay.EU.net>
12 * Florian La Roche, <flla@stud.uni-sb.de>
13 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14 * Linus Torvalds, <torvalds@cs.helsinki.fi>
15 * Alan Cox, <gw4pts@gw4pts.ampr.org>
16 * Matthew Dillon, <dillon@apollo.west.oic.com>
17 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18 * Jorge Cwik, <jorge@laser.satlink.net>
19 */
20
21/*
22 * Changes: Pedro Roque : Retransmit queue handled by TCP.
23 * : Fragmentation on mtu decrease
24 * : Segment collapse on retransmit
25 * : AF independence
26 *
27 * Linus Torvalds : send_delayed_ack
28 * David S. Miller : Charge memory using the right skb
29 * during syn/ack processing.
30 * David S. Miller : Output engine completely rewritten.
31 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
32 * Cacophonix Gaul : draft-minshall-nagle-01
33 * J Hadi Salim : ECN support
34 *
35 */
36
37#define pr_fmt(fmt) "TCP: " fmt
38
39#include <net/tcp.h>
40
41#include <linux/compiler.h>
42#include <linux/gfp.h>
43#include <linux/module.h>
44
45/* People can turn this off for buggy TCP's found in printers etc. */
46int sysctl_tcp_retrans_collapse __read_mostly = 1;
47
48/* People can turn this on to work with those rare, broken TCPs that
49 * interpret the window field as a signed quantity.
50 */
51int sysctl_tcp_workaround_signed_windows __read_mostly = 0;
52
53/* Default TSQ limit of four TSO segments */
54int sysctl_tcp_limit_output_bytes __read_mostly = 262144;
55
56/* This limits the percentage of the congestion window which we
57 * will allow a single TSO frame to consume. Building TSO frames
58 * which are too large can cause TCP streams to be bursty.
59 */
60int sysctl_tcp_tso_win_divisor __read_mostly = 3;
61
62/* By default, RFC2861 behavior. */
63int sysctl_tcp_slow_start_after_idle __read_mostly = 1;
64
65static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
66 int push_one, gfp_t gfp);
67
68/* Account for new data that has been sent to the network. */
69static void tcp_event_new_data_sent(struct sock *sk, const struct sk_buff *skb)
70{
71 struct inet_connection_sock *icsk = inet_csk(sk);
72 struct tcp_sock *tp = tcp_sk(sk);
73 unsigned int prior_packets = tp->packets_out;
74
75 tcp_advance_send_head(sk, skb);
76 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
77
78 tp->packets_out += tcp_skb_pcount(skb);
79 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS ||
80 icsk->icsk_pending == ICSK_TIME_LOSS_PROBE) {
81 tcp_rearm_rto(sk);
82 }
83
84 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
85 tcp_skb_pcount(skb));
86}
87
88/* SND.NXT, if window was not shrunk.
89 * If window has been shrunk, what should we make? It is not clear at all.
90 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
91 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
92 * invalid. OK, let's make this for now:
93 */
94static inline __u32 tcp_acceptable_seq(const struct sock *sk)
95{
96 const struct tcp_sock *tp = tcp_sk(sk);
97
98 if (!before(tcp_wnd_end(tp), tp->snd_nxt))
99 return tp->snd_nxt;
100 else
101 return tcp_wnd_end(tp);
102}
103
104/* Calculate mss to advertise in SYN segment.
105 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
106 *
107 * 1. It is independent of path mtu.
108 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
109 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
110 * attached devices, because some buggy hosts are confused by
111 * large MSS.
112 * 4. We do not make 3, we advertise MSS, calculated from first
113 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
114 * This may be overridden via information stored in routing table.
115 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
116 * probably even Jumbo".
117 */
118static __u16 tcp_advertise_mss(struct sock *sk)
119{
120 struct tcp_sock *tp = tcp_sk(sk);
121 const struct dst_entry *dst = __sk_dst_get(sk);
122 int mss = tp->advmss;
123
124 if (dst) {
125 unsigned int metric = dst_metric_advmss(dst);
126
127 if (metric < mss) {
128 mss = metric;
129 tp->advmss = mss;
130 }
131 }
132
133 return (__u16)mss;
134}
135
136/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
137 * This is the first part of cwnd validation mechanism.
138 */
139void tcp_cwnd_restart(struct sock *sk, s32 delta)
140{
141 struct tcp_sock *tp = tcp_sk(sk);
142 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
143 u32 cwnd = tp->snd_cwnd;
144
145 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
146
147 tp->snd_ssthresh = tcp_current_ssthresh(sk);
148 restart_cwnd = min(restart_cwnd, cwnd);
149
150 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
151 cwnd >>= 1;
152 tp->snd_cwnd = max(cwnd, restart_cwnd);
153 tp->snd_cwnd_stamp = tcp_time_stamp;
154 tp->snd_cwnd_used = 0;
155}
156
157/* Congestion state accounting after a packet has been sent. */
158static void tcp_event_data_sent(struct tcp_sock *tp,
159 struct sock *sk)
160{
161 struct inet_connection_sock *icsk = inet_csk(sk);
162 const u32 now = tcp_time_stamp;
163
164 if (tcp_packets_in_flight(tp) == 0)
165 tcp_ca_event(sk, CA_EVENT_TX_START);
166
167 tp->lsndtime = now;
168
169 /* If it is a reply for ato after last received
170 * packet, enter pingpong mode.
171 */
172 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
173 icsk->icsk_ack.pingpong = 1;
174}
175
176/* Account for an ACK we sent. */
177static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts)
178{
179 tcp_dec_quickack_mode(sk, pkts);
180 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
181}
182
183
184u32 tcp_default_init_rwnd(u32 mss)
185{
186 /* Initial receive window should be twice of TCP_INIT_CWND to
187 * enable proper sending of new unsent data during fast recovery
188 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
189 * limit when mss is larger than 1460.
190 */
191 u32 init_rwnd = TCP_INIT_CWND * 2;
192
193 if (mss > 1460)
194 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
195 return init_rwnd;
196}
197
198/* Determine a window scaling and initial window to offer.
199 * Based on the assumption that the given amount of space
200 * will be offered. Store the results in the tp structure.
201 * NOTE: for smooth operation initial space offering should
202 * be a multiple of mss if possible. We assume here that mss >= 1.
203 * This MUST be enforced by all callers.
204 */
205void tcp_select_initial_window(int __space, __u32 mss,
206 __u32 *rcv_wnd, __u32 *window_clamp,
207 int wscale_ok, __u8 *rcv_wscale,
208 __u32 init_rcv_wnd)
209{
210 unsigned int space = (__space < 0 ? 0 : __space);
211
212 /* If no clamp set the clamp to the max possible scaled window */
213 if (*window_clamp == 0)
214 (*window_clamp) = (65535 << 14);
215 space = min(*window_clamp, space);
216
217 /* Quantize space offering to a multiple of mss if possible. */
218 if (space > mss)
219 space = (space / mss) * mss;
220
221 /* NOTE: offering an initial window larger than 32767
222 * will break some buggy TCP stacks. If the admin tells us
223 * it is likely we could be speaking with such a buggy stack
224 * we will truncate our initial window offering to 32K-1
225 * unless the remote has sent us a window scaling option,
226 * which we interpret as a sign the remote TCP is not
227 * misinterpreting the window field as a signed quantity.
228 */
229 if (sysctl_tcp_workaround_signed_windows)
230 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
231 else
232 (*rcv_wnd) = space;
233
234 (*rcv_wscale) = 0;
235 if (wscale_ok) {
236 /* Set window scaling on max possible window
237 * See RFC1323 for an explanation of the limit to 14
238 */
239 space = max_t(u32, sysctl_tcp_rmem[2], sysctl_rmem_max);
240 space = min_t(u32, space, *window_clamp);
241 while (space > 65535 && (*rcv_wscale) < 14) {
242 space >>= 1;
243 (*rcv_wscale)++;
244 }
245 }
246
247 if (mss > (1 << *rcv_wscale)) {
248 if (!init_rcv_wnd) /* Use default unless specified otherwise */
249 init_rcv_wnd = tcp_default_init_rwnd(mss);
250 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
251 }
252
253 /* Set the clamp no higher than max representable value */
254 (*window_clamp) = min(65535U << (*rcv_wscale), *window_clamp);
255}
256EXPORT_SYMBOL(tcp_select_initial_window);
257
258/* Chose a new window to advertise, update state in tcp_sock for the
259 * socket, and return result with RFC1323 scaling applied. The return
260 * value can be stuffed directly into th->window for an outgoing
261 * frame.
262 */
263static u16 tcp_select_window(struct sock *sk)
264{
265 struct tcp_sock *tp = tcp_sk(sk);
266 u32 old_win = tp->rcv_wnd;
267 u32 cur_win = tcp_receive_window(tp);
268 u32 new_win = __tcp_select_window(sk);
269
270 /* Never shrink the offered window */
271 if (new_win < cur_win) {
272 /* Danger Will Robinson!
273 * Don't update rcv_wup/rcv_wnd here or else
274 * we will not be able to advertise a zero
275 * window in time. --DaveM
276 *
277 * Relax Will Robinson.
278 */
279 if (new_win == 0)
280 NET_INC_STATS(sock_net(sk),
281 LINUX_MIB_TCPWANTZEROWINDOWADV);
282 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
283 }
284 tp->rcv_wnd = new_win;
285 tp->rcv_wup = tp->rcv_nxt;
286
287 /* Make sure we do not exceed the maximum possible
288 * scaled window.
289 */
290 if (!tp->rx_opt.rcv_wscale && sysctl_tcp_workaround_signed_windows)
291 new_win = min(new_win, MAX_TCP_WINDOW);
292 else
293 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
294
295 /* RFC1323 scaling applied */
296 new_win >>= tp->rx_opt.rcv_wscale;
297
298 /* If we advertise zero window, disable fast path. */
299 if (new_win == 0) {
300 tp->pred_flags = 0;
301 if (old_win)
302 NET_INC_STATS(sock_net(sk),
303 LINUX_MIB_TCPTOZEROWINDOWADV);
304 } else if (old_win == 0) {
305 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
306 }
307
308 return new_win;
309}
310
311/* Packet ECN state for a SYN-ACK */
312static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
313{
314 const struct tcp_sock *tp = tcp_sk(sk);
315
316 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
317 if (!(tp->ecn_flags & TCP_ECN_OK))
318 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
319 else if (tcp_ca_needs_ecn(sk))
320 INET_ECN_xmit(sk);
321}
322
323/* Packet ECN state for a SYN. */
324static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
325{
326 struct tcp_sock *tp = tcp_sk(sk);
327 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
328 tcp_ca_needs_ecn(sk);
329
330 if (!use_ecn) {
331 const struct dst_entry *dst = __sk_dst_get(sk);
332
333 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
334 use_ecn = true;
335 }
336
337 tp->ecn_flags = 0;
338
339 if (use_ecn) {
340 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
341 tp->ecn_flags = TCP_ECN_OK;
342 if (tcp_ca_needs_ecn(sk))
343 INET_ECN_xmit(sk);
344 }
345}
346
347static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
348{
349 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
350 /* tp->ecn_flags are cleared at a later point in time when
351 * SYN ACK is ultimatively being received.
352 */
353 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
354}
355
356static void
357tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
358{
359 if (inet_rsk(req)->ecn_ok)
360 th->ece = 1;
361}
362
363/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
364 * be sent.
365 */
366static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
367 int tcp_header_len)
368{
369 struct tcp_sock *tp = tcp_sk(sk);
370
371 if (tp->ecn_flags & TCP_ECN_OK) {
372 /* Not-retransmitted data segment: set ECT and inject CWR. */
373 if (skb->len != tcp_header_len &&
374 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
375 INET_ECN_xmit(sk);
376 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
377 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
378 tcp_hdr(skb)->cwr = 1;
379 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
380 }
381 } else if (!tcp_ca_needs_ecn(sk)) {
382 /* ACK or retransmitted segment: clear ECT|CE */
383 INET_ECN_dontxmit(sk);
384 }
385 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
386 tcp_hdr(skb)->ece = 1;
387 }
388}
389
390/* Constructs common control bits of non-data skb. If SYN/FIN is present,
391 * auto increment end seqno.
392 */
393static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
394{
395 skb->ip_summed = CHECKSUM_PARTIAL;
396 skb->csum = 0;
397
398 TCP_SKB_CB(skb)->tcp_flags = flags;
399 TCP_SKB_CB(skb)->sacked = 0;
400
401 tcp_skb_pcount_set(skb, 1);
402
403 TCP_SKB_CB(skb)->seq = seq;
404 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
405 seq++;
406 TCP_SKB_CB(skb)->end_seq = seq;
407}
408
409static inline bool tcp_urg_mode(const struct tcp_sock *tp)
410{
411 return tp->snd_una != tp->snd_up;
412}
413
414#define OPTION_SACK_ADVERTISE (1 << 0)
415#define OPTION_TS (1 << 1)
416#define OPTION_MD5 (1 << 2)
417#define OPTION_WSCALE (1 << 3)
418#define OPTION_FAST_OPEN_COOKIE (1 << 8)
419
420struct tcp_out_options {
421 u16 options; /* bit field of OPTION_* */
422 u16 mss; /* 0 to disable */
423 u8 ws; /* window scale, 0 to disable */
424 u8 num_sack_blocks; /* number of SACK blocks to include */
425 u8 hash_size; /* bytes in hash_location */
426 __u8 *hash_location; /* temporary pointer, overloaded */
427 __u32 tsval, tsecr; /* need to include OPTION_TS */
428 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
429};
430
431/* Write previously computed TCP options to the packet.
432 *
433 * Beware: Something in the Internet is very sensitive to the ordering of
434 * TCP options, we learned this through the hard way, so be careful here.
435 * Luckily we can at least blame others for their non-compliance but from
436 * inter-operability perspective it seems that we're somewhat stuck with
437 * the ordering which we have been using if we want to keep working with
438 * those broken things (not that it currently hurts anybody as there isn't
439 * particular reason why the ordering would need to be changed).
440 *
441 * At least SACK_PERM as the first option is known to lead to a disaster
442 * (but it may well be that other scenarios fail similarly).
443 */
444static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
445 struct tcp_out_options *opts)
446{
447 u16 options = opts->options; /* mungable copy */
448
449 if (unlikely(OPTION_MD5 & options)) {
450 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
451 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
452 /* overload cookie hash location */
453 opts->hash_location = (__u8 *)ptr;
454 ptr += 4;
455 }
456
457 if (unlikely(opts->mss)) {
458 *ptr++ = htonl((TCPOPT_MSS << 24) |
459 (TCPOLEN_MSS << 16) |
460 opts->mss);
461 }
462
463 if (likely(OPTION_TS & options)) {
464 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
465 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
466 (TCPOLEN_SACK_PERM << 16) |
467 (TCPOPT_TIMESTAMP << 8) |
468 TCPOLEN_TIMESTAMP);
469 options &= ~OPTION_SACK_ADVERTISE;
470 } else {
471 *ptr++ = htonl((TCPOPT_NOP << 24) |
472 (TCPOPT_NOP << 16) |
473 (TCPOPT_TIMESTAMP << 8) |
474 TCPOLEN_TIMESTAMP);
475 }
476 *ptr++ = htonl(opts->tsval);
477 *ptr++ = htonl(opts->tsecr);
478 }
479
480 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
481 *ptr++ = htonl((TCPOPT_NOP << 24) |
482 (TCPOPT_NOP << 16) |
483 (TCPOPT_SACK_PERM << 8) |
484 TCPOLEN_SACK_PERM);
485 }
486
487 if (unlikely(OPTION_WSCALE & options)) {
488 *ptr++ = htonl((TCPOPT_NOP << 24) |
489 (TCPOPT_WINDOW << 16) |
490 (TCPOLEN_WINDOW << 8) |
491 opts->ws);
492 }
493
494 if (unlikely(opts->num_sack_blocks)) {
495 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
496 tp->duplicate_sack : tp->selective_acks;
497 int this_sack;
498
499 *ptr++ = htonl((TCPOPT_NOP << 24) |
500 (TCPOPT_NOP << 16) |
501 (TCPOPT_SACK << 8) |
502 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
503 TCPOLEN_SACK_PERBLOCK)));
504
505 for (this_sack = 0; this_sack < opts->num_sack_blocks;
506 ++this_sack) {
507 *ptr++ = htonl(sp[this_sack].start_seq);
508 *ptr++ = htonl(sp[this_sack].end_seq);
509 }
510
511 tp->rx_opt.dsack = 0;
512 }
513
514 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
515 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
516 u8 *p = (u8 *)ptr;
517 u32 len; /* Fast Open option length */
518
519 if (foc->exp) {
520 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
521 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
522 TCPOPT_FASTOPEN_MAGIC);
523 p += TCPOLEN_EXP_FASTOPEN_BASE;
524 } else {
525 len = TCPOLEN_FASTOPEN_BASE + foc->len;
526 *p++ = TCPOPT_FASTOPEN;
527 *p++ = len;
528 }
529
530 memcpy(p, foc->val, foc->len);
531 if ((len & 3) == 2) {
532 p[foc->len] = TCPOPT_NOP;
533 p[foc->len + 1] = TCPOPT_NOP;
534 }
535 ptr += (len + 3) >> 2;
536 }
537}
538
539/* Compute TCP options for SYN packets. This is not the final
540 * network wire format yet.
541 */
542static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
543 struct tcp_out_options *opts,
544 struct tcp_md5sig_key **md5)
545{
546 struct tcp_sock *tp = tcp_sk(sk);
547 unsigned int remaining = MAX_TCP_OPTION_SPACE;
548 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
549
550#ifdef CONFIG_TCP_MD5SIG
551 *md5 = tp->af_specific->md5_lookup(sk, sk);
552 if (*md5) {
553 opts->options |= OPTION_MD5;
554 remaining -= TCPOLEN_MD5SIG_ALIGNED;
555 }
556#else
557 *md5 = NULL;
558#endif
559
560 /* We always get an MSS option. The option bytes which will be seen in
561 * normal data packets should timestamps be used, must be in the MSS
562 * advertised. But we subtract them from tp->mss_cache so that
563 * calculations in tcp_sendmsg are simpler etc. So account for this
564 * fact here if necessary. If we don't do this correctly, as a
565 * receiver we won't recognize data packets as being full sized when we
566 * should, and thus we won't abide by the delayed ACK rules correctly.
567 * SACKs don't matter, we never delay an ACK when we have any of those
568 * going out. */
569 opts->mss = tcp_advertise_mss(sk);
570 remaining -= TCPOLEN_MSS_ALIGNED;
571
572 if (likely(sysctl_tcp_timestamps && !*md5)) {
573 opts->options |= OPTION_TS;
574 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
575 opts->tsecr = tp->rx_opt.ts_recent;
576 remaining -= TCPOLEN_TSTAMP_ALIGNED;
577 }
578 if (likely(sysctl_tcp_window_scaling)) {
579 opts->ws = tp->rx_opt.rcv_wscale;
580 opts->options |= OPTION_WSCALE;
581 remaining -= TCPOLEN_WSCALE_ALIGNED;
582 }
583 if (likely(sysctl_tcp_sack)) {
584 opts->options |= OPTION_SACK_ADVERTISE;
585 if (unlikely(!(OPTION_TS & opts->options)))
586 remaining -= TCPOLEN_SACKPERM_ALIGNED;
587 }
588
589 if (fastopen && fastopen->cookie.len >= 0) {
590 u32 need = fastopen->cookie.len;
591
592 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
593 TCPOLEN_FASTOPEN_BASE;
594 need = (need + 3) & ~3U; /* Align to 32 bits */
595 if (remaining >= need) {
596 opts->options |= OPTION_FAST_OPEN_COOKIE;
597 opts->fastopen_cookie = &fastopen->cookie;
598 remaining -= need;
599 tp->syn_fastopen = 1;
600 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
601 }
602 }
603
604 return MAX_TCP_OPTION_SPACE - remaining;
605}
606
607/* Set up TCP options for SYN-ACKs. */
608static unsigned int tcp_synack_options(struct request_sock *req,
609 unsigned int mss, struct sk_buff *skb,
610 struct tcp_out_options *opts,
611 const struct tcp_md5sig_key *md5,
612 struct tcp_fastopen_cookie *foc)
613{
614 struct inet_request_sock *ireq = inet_rsk(req);
615 unsigned int remaining = MAX_TCP_OPTION_SPACE;
616
617#ifdef CONFIG_TCP_MD5SIG
618 if (md5) {
619 opts->options |= OPTION_MD5;
620 remaining -= TCPOLEN_MD5SIG_ALIGNED;
621
622 /* We can't fit any SACK blocks in a packet with MD5 + TS
623 * options. There was discussion about disabling SACK
624 * rather than TS in order to fit in better with old,
625 * buggy kernels, but that was deemed to be unnecessary.
626 */
627 ireq->tstamp_ok &= !ireq->sack_ok;
628 }
629#endif
630
631 /* We always send an MSS option. */
632 opts->mss = mss;
633 remaining -= TCPOLEN_MSS_ALIGNED;
634
635 if (likely(ireq->wscale_ok)) {
636 opts->ws = ireq->rcv_wscale;
637 opts->options |= OPTION_WSCALE;
638 remaining -= TCPOLEN_WSCALE_ALIGNED;
639 }
640 if (likely(ireq->tstamp_ok)) {
641 opts->options |= OPTION_TS;
642 opts->tsval = tcp_skb_timestamp(skb);
643 opts->tsecr = req->ts_recent;
644 remaining -= TCPOLEN_TSTAMP_ALIGNED;
645 }
646 if (likely(ireq->sack_ok)) {
647 opts->options |= OPTION_SACK_ADVERTISE;
648 if (unlikely(!ireq->tstamp_ok))
649 remaining -= TCPOLEN_SACKPERM_ALIGNED;
650 }
651 if (foc != NULL && foc->len >= 0) {
652 u32 need = foc->len;
653
654 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
655 TCPOLEN_FASTOPEN_BASE;
656 need = (need + 3) & ~3U; /* Align to 32 bits */
657 if (remaining >= need) {
658 opts->options |= OPTION_FAST_OPEN_COOKIE;
659 opts->fastopen_cookie = foc;
660 remaining -= need;
661 }
662 }
663
664 return MAX_TCP_OPTION_SPACE - remaining;
665}
666
667/* Compute TCP options for ESTABLISHED sockets. This is not the
668 * final wire format yet.
669 */
670static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
671 struct tcp_out_options *opts,
672 struct tcp_md5sig_key **md5)
673{
674 struct tcp_sock *tp = tcp_sk(sk);
675 unsigned int size = 0;
676 unsigned int eff_sacks;
677
678 opts->options = 0;
679
680#ifdef CONFIG_TCP_MD5SIG
681 *md5 = tp->af_specific->md5_lookup(sk, sk);
682 if (unlikely(*md5)) {
683 opts->options |= OPTION_MD5;
684 size += TCPOLEN_MD5SIG_ALIGNED;
685 }
686#else
687 *md5 = NULL;
688#endif
689
690 if (likely(tp->rx_opt.tstamp_ok)) {
691 opts->options |= OPTION_TS;
692 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
693 opts->tsecr = tp->rx_opt.ts_recent;
694 size += TCPOLEN_TSTAMP_ALIGNED;
695 }
696
697 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
698 if (unlikely(eff_sacks)) {
699 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
700 opts->num_sack_blocks =
701 min_t(unsigned int, eff_sacks,
702 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
703 TCPOLEN_SACK_PERBLOCK);
704 size += TCPOLEN_SACK_BASE_ALIGNED +
705 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
706 }
707
708 return size;
709}
710
711
712/* TCP SMALL QUEUES (TSQ)
713 *
714 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
715 * to reduce RTT and bufferbloat.
716 * We do this using a special skb destructor (tcp_wfree).
717 *
718 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
719 * needs to be reallocated in a driver.
720 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
721 *
722 * Since transmit from skb destructor is forbidden, we use a tasklet
723 * to process all sockets that eventually need to send more skbs.
724 * We use one tasklet per cpu, with its own queue of sockets.
725 */
726struct tsq_tasklet {
727 struct tasklet_struct tasklet;
728 struct list_head head; /* queue of tcp sockets */
729};
730static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
731
732static void tcp_tsq_handler(struct sock *sk)
733{
734 if ((1 << sk->sk_state) &
735 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
736 TCPF_CLOSE_WAIT | TCPF_LAST_ACK))
737 tcp_write_xmit(sk, tcp_current_mss(sk), tcp_sk(sk)->nonagle,
738 0, GFP_ATOMIC);
739}
740/*
741 * One tasklet per cpu tries to send more skbs.
742 * We run in tasklet context but need to disable irqs when
743 * transferring tsq->head because tcp_wfree() might
744 * interrupt us (non NAPI drivers)
745 */
746static void tcp_tasklet_func(unsigned long data)
747{
748 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
749 LIST_HEAD(list);
750 unsigned long flags;
751 struct list_head *q, *n;
752 struct tcp_sock *tp;
753 struct sock *sk;
754
755 local_irq_save(flags);
756 list_splice_init(&tsq->head, &list);
757 local_irq_restore(flags);
758
759 list_for_each_safe(q, n, &list) {
760 tp = list_entry(q, struct tcp_sock, tsq_node);
761 list_del(&tp->tsq_node);
762
763 sk = (struct sock *)tp;
764 bh_lock_sock(sk);
765
766 if (!sock_owned_by_user(sk)) {
767 tcp_tsq_handler(sk);
768 } else {
769 /* defer the work to tcp_release_cb() */
770 set_bit(TCP_TSQ_DEFERRED, &tp->tsq_flags);
771 }
772 bh_unlock_sock(sk);
773
774 clear_bit(TSQ_QUEUED, &tp->tsq_flags);
775 sk_free(sk);
776 }
777}
778
779#define TCP_DEFERRED_ALL ((1UL << TCP_TSQ_DEFERRED) | \
780 (1UL << TCP_WRITE_TIMER_DEFERRED) | \
781 (1UL << TCP_DELACK_TIMER_DEFERRED) | \
782 (1UL << TCP_MTU_REDUCED_DEFERRED))
783/**
784 * tcp_release_cb - tcp release_sock() callback
785 * @sk: socket
786 *
787 * called from release_sock() to perform protocol dependent
788 * actions before socket release.
789 */
790void tcp_release_cb(struct sock *sk)
791{
792 struct tcp_sock *tp = tcp_sk(sk);
793 unsigned long flags, nflags;
794
795 /* perform an atomic operation only if at least one flag is set */
796 do {
797 flags = tp->tsq_flags;
798 if (!(flags & TCP_DEFERRED_ALL))
799 return;
800 nflags = flags & ~TCP_DEFERRED_ALL;
801 } while (cmpxchg(&tp->tsq_flags, flags, nflags) != flags);
802
803 if (flags & (1UL << TCP_TSQ_DEFERRED))
804 tcp_tsq_handler(sk);
805
806 /* Here begins the tricky part :
807 * We are called from release_sock() with :
808 * 1) BH disabled
809 * 2) sk_lock.slock spinlock held
810 * 3) socket owned by us (sk->sk_lock.owned == 1)
811 *
812 * But following code is meant to be called from BH handlers,
813 * so we should keep BH disabled, but early release socket ownership
814 */
815 sock_release_ownership(sk);
816
817 if (flags & (1UL << TCP_WRITE_TIMER_DEFERRED)) {
818 tcp_write_timer_handler(sk);
819 __sock_put(sk);
820 }
821 if (flags & (1UL << TCP_DELACK_TIMER_DEFERRED)) {
822 tcp_delack_timer_handler(sk);
823 __sock_put(sk);
824 }
825 if (flags & (1UL << TCP_MTU_REDUCED_DEFERRED)) {
826 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
827 __sock_put(sk);
828 }
829}
830EXPORT_SYMBOL(tcp_release_cb);
831
832void __init tcp_tasklet_init(void)
833{
834 int i;
835
836 for_each_possible_cpu(i) {
837 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
838
839 INIT_LIST_HEAD(&tsq->head);
840 tasklet_init(&tsq->tasklet,
841 tcp_tasklet_func,
842 (unsigned long)tsq);
843 }
844}
845
846/*
847 * Write buffer destructor automatically called from kfree_skb.
848 * We can't xmit new skbs from this context, as we might already
849 * hold qdisc lock.
850 */
851void tcp_wfree(struct sk_buff *skb)
852{
853 struct sock *sk = skb->sk;
854 struct tcp_sock *tp = tcp_sk(sk);
855 int wmem;
856
857 /* Keep one reference on sk_wmem_alloc.
858 * Will be released by sk_free() from here or tcp_tasklet_func()
859 */
860 wmem = atomic_sub_return(skb->truesize - 1, &sk->sk_wmem_alloc);
861
862 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
863 * Wait until our queues (qdisc + devices) are drained.
864 * This gives :
865 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
866 * - chance for incoming ACK (processed by another cpu maybe)
867 * to migrate this flow (skb->ooo_okay will be eventually set)
868 */
869 if (wmem >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
870 goto out;
871
872 if (test_and_clear_bit(TSQ_THROTTLED, &tp->tsq_flags) &&
873 !test_and_set_bit(TSQ_QUEUED, &tp->tsq_flags)) {
874 unsigned long flags;
875 struct tsq_tasklet *tsq;
876
877 /* queue this socket to tasklet queue */
878 local_irq_save(flags);
879 tsq = this_cpu_ptr(&tsq_tasklet);
880 list_add(&tp->tsq_node, &tsq->head);
881 tasklet_schedule(&tsq->tasklet);
882 local_irq_restore(flags);
883 return;
884 }
885out:
886 sk_free(sk);
887}
888
889/* This routine actually transmits TCP packets queued in by
890 * tcp_do_sendmsg(). This is used by both the initial
891 * transmission and possible later retransmissions.
892 * All SKB's seen here are completely headerless. It is our
893 * job to build the TCP header, and pass the packet down to
894 * IP so it can do the same plus pass the packet off to the
895 * device.
896 *
897 * We are working here with either a clone of the original
898 * SKB, or a fresh unique copy made by the retransmit engine.
899 */
900static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
901 gfp_t gfp_mask)
902{
903 const struct inet_connection_sock *icsk = inet_csk(sk);
904 struct inet_sock *inet;
905 struct tcp_sock *tp;
906 struct tcp_skb_cb *tcb;
907 struct tcp_out_options opts;
908 unsigned int tcp_options_size, tcp_header_size;
909 struct tcp_md5sig_key *md5;
910 struct tcphdr *th;
911 int err;
912
913 BUG_ON(!skb || !tcp_skb_pcount(skb));
914
915 if (clone_it) {
916 skb_mstamp_get(&skb->skb_mstamp);
917
918 if (unlikely(skb_cloned(skb)))
919 skb = pskb_copy(skb, gfp_mask);
920 else
921 skb = skb_clone(skb, gfp_mask);
922 if (unlikely(!skb))
923 return -ENOBUFS;
924 }
925
926 inet = inet_sk(sk);
927 tp = tcp_sk(sk);
928 tcb = TCP_SKB_CB(skb);
929 memset(&opts, 0, sizeof(opts));
930
931 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
932 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
933 else
934 tcp_options_size = tcp_established_options(sk, skb, &opts,
935 &md5);
936 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
937
938 /* if no packet is in qdisc/device queue, then allow XPS to select
939 * another queue. We can be called from tcp_tsq_handler()
940 * which holds one reference to sk_wmem_alloc.
941 *
942 * TODO: Ideally, in-flight pure ACK packets should not matter here.
943 * One way to get this would be to set skb->truesize = 2 on them.
944 */
945 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
946
947 skb_push(skb, tcp_header_size);
948 skb_reset_transport_header(skb);
949
950 skb_orphan(skb);
951 skb->sk = sk;
952 skb->destructor = skb_is_tcp_pure_ack(skb) ? sock_wfree : tcp_wfree;
953 skb_set_hash_from_sk(skb, sk);
954 atomic_add(skb->truesize, &sk->sk_wmem_alloc);
955
956 /* Build TCP header and checksum it. */
957 th = tcp_hdr(skb);
958 th->source = inet->inet_sport;
959 th->dest = inet->inet_dport;
960 th->seq = htonl(tcb->seq);
961 th->ack_seq = htonl(tp->rcv_nxt);
962 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
963 tcb->tcp_flags);
964
965 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
966 /* RFC1323: The window in SYN & SYN/ACK segments
967 * is never scaled.
968 */
969 th->window = htons(min(tp->rcv_wnd, 65535U));
970 } else {
971 th->window = htons(tcp_select_window(sk));
972 }
973 th->check = 0;
974 th->urg_ptr = 0;
975
976 /* The urg_mode check is necessary during a below snd_una win probe */
977 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
978 if (before(tp->snd_up, tcb->seq + 0x10000)) {
979 th->urg_ptr = htons(tp->snd_up - tcb->seq);
980 th->urg = 1;
981 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
982 th->urg_ptr = htons(0xFFFF);
983 th->urg = 1;
984 }
985 }
986
987 tcp_options_write((__be32 *)(th + 1), tp, &opts);
988 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
989 if (likely((tcb->tcp_flags & TCPHDR_SYN) == 0))
990 tcp_ecn_send(sk, skb, tcp_header_size);
991
992#ifdef CONFIG_TCP_MD5SIG
993 /* Calculate the MD5 hash, as we have all we need now */
994 if (md5) {
995 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
996 tp->af_specific->calc_md5_hash(opts.hash_location,
997 md5, sk, skb);
998 }
999#endif
1000
1001 icsk->icsk_af_ops->send_check(sk, skb);
1002
1003 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1004 tcp_event_ack_sent(sk, tcp_skb_pcount(skb));
1005
1006 if (skb->len != tcp_header_size) {
1007 tcp_event_data_sent(tp, sk);
1008 tp->data_segs_out += tcp_skb_pcount(skb);
1009 }
1010
1011 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1012 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1013 tcp_skb_pcount(skb));
1014
1015 tp->segs_out += tcp_skb_pcount(skb);
1016 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1017 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1018 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1019
1020 /* Our usage of tstamp should remain private */
1021 skb->tstamp.tv64 = 0;
1022
1023 /* Cleanup our debris for IP stacks */
1024 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1025 sizeof(struct inet6_skb_parm)));
1026
1027 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1028
1029 if (likely(err <= 0))
1030 return err;
1031
1032 tcp_enter_cwr(sk);
1033
1034 return net_xmit_eval(err);
1035}
1036
1037/* This routine just queues the buffer for sending.
1038 *
1039 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1040 * otherwise socket can stall.
1041 */
1042static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1043{
1044 struct tcp_sock *tp = tcp_sk(sk);
1045
1046 /* Advance write_seq and place onto the write_queue. */
1047 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1048 __skb_header_release(skb);
1049 tcp_add_write_queue_tail(sk, skb);
1050 sk->sk_wmem_queued += skb->truesize;
1051 sk_mem_charge(sk, skb->truesize);
1052}
1053
1054/* Initialize TSO segments for a packet. */
1055static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1056{
1057 if (skb->len <= mss_now || skb->ip_summed == CHECKSUM_NONE) {
1058 /* Avoid the costly divide in the normal
1059 * non-TSO case.
1060 */
1061 tcp_skb_pcount_set(skb, 1);
1062 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1063 } else {
1064 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1065 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1066 }
1067}
1068
1069/* When a modification to fackets out becomes necessary, we need to check
1070 * skb is counted to fackets_out or not.
1071 */
1072static void tcp_adjust_fackets_out(struct sock *sk, const struct sk_buff *skb,
1073 int decr)
1074{
1075 struct tcp_sock *tp = tcp_sk(sk);
1076
1077 if (!tp->sacked_out || tcp_is_reno(tp))
1078 return;
1079
1080 if (after(tcp_highest_sack_seq(tp), TCP_SKB_CB(skb)->seq))
1081 tp->fackets_out -= decr;
1082}
1083
1084/* Pcount in the middle of the write queue got changed, we need to do various
1085 * tweaks to fix counters
1086 */
1087static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1088{
1089 struct tcp_sock *tp = tcp_sk(sk);
1090
1091 tp->packets_out -= decr;
1092
1093 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1094 tp->sacked_out -= decr;
1095 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1096 tp->retrans_out -= decr;
1097 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1098 tp->lost_out -= decr;
1099
1100 /* Reno case is special. Sigh... */
1101 if (tcp_is_reno(tp) && decr > 0)
1102 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1103
1104 tcp_adjust_fackets_out(sk, skb, decr);
1105
1106 if (tp->lost_skb_hint &&
1107 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1108 (tcp_is_fack(tp) || (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)))
1109 tp->lost_cnt_hint -= decr;
1110
1111 tcp_verify_left_out(tp);
1112}
1113
1114static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1115{
1116 struct skb_shared_info *shinfo = skb_shinfo(skb);
1117
1118 if (unlikely(shinfo->tx_flags & SKBTX_ANY_TSTAMP) &&
1119 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1120 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1121 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1122
1123 shinfo->tx_flags &= ~tsflags;
1124 shinfo2->tx_flags |= tsflags;
1125 swap(shinfo->tskey, shinfo2->tskey);
1126 }
1127}
1128
1129/* Function to create two new TCP segments. Shrinks the given segment
1130 * to the specified size and appends a new segment with the rest of the
1131 * packet to the list. This won't be called frequently, I hope.
1132 * Remember, these are still headerless SKBs at this point.
1133 */
1134int tcp_fragment(struct sock *sk, struct sk_buff *skb, u32 len,
1135 unsigned int mss_now, gfp_t gfp)
1136{
1137 struct tcp_sock *tp = tcp_sk(sk);
1138 struct sk_buff *buff;
1139 int nsize, old_factor;
1140 int nlen;
1141 u8 flags;
1142
1143 if (WARN_ON(len > skb->len))
1144 return -EINVAL;
1145
1146 nsize = skb_headlen(skb) - len;
1147 if (nsize < 0)
1148 nsize = 0;
1149
1150 if (skb_unclone(skb, gfp))
1151 return -ENOMEM;
1152
1153 /* Get a new skb... force flag on. */
1154 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1155 if (!buff)
1156 return -ENOMEM; /* We'll just try again later. */
1157
1158 sk->sk_wmem_queued += buff->truesize;
1159 sk_mem_charge(sk, buff->truesize);
1160 nlen = skb->len - len - nsize;
1161 buff->truesize += nlen;
1162 skb->truesize -= nlen;
1163
1164 /* Correct the sequence numbers. */
1165 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1166 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1167 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1168
1169 /* PSH and FIN should only be set in the second packet. */
1170 flags = TCP_SKB_CB(skb)->tcp_flags;
1171 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1172 TCP_SKB_CB(buff)->tcp_flags = flags;
1173 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1174
1175 if (!skb_shinfo(skb)->nr_frags && skb->ip_summed != CHECKSUM_PARTIAL) {
1176 /* Copy and checksum data tail into the new buffer. */
1177 buff->csum = csum_partial_copy_nocheck(skb->data + len,
1178 skb_put(buff, nsize),
1179 nsize, 0);
1180
1181 skb_trim(skb, len);
1182
1183 skb->csum = csum_block_sub(skb->csum, buff->csum, len);
1184 } else {
1185 skb->ip_summed = CHECKSUM_PARTIAL;
1186 skb_split(skb, buff, len);
1187 }
1188
1189 buff->ip_summed = skb->ip_summed;
1190
1191 buff->tstamp = skb->tstamp;
1192 tcp_fragment_tstamp(skb, buff);
1193
1194 old_factor = tcp_skb_pcount(skb);
1195
1196 /* Fix up tso_factor for both original and new SKB. */
1197 tcp_set_skb_tso_segs(skb, mss_now);
1198 tcp_set_skb_tso_segs(buff, mss_now);
1199
1200 /* If this packet has been sent out already, we must
1201 * adjust the various packet counters.
1202 */
1203 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1204 int diff = old_factor - tcp_skb_pcount(skb) -
1205 tcp_skb_pcount(buff);
1206
1207 if (diff)
1208 tcp_adjust_pcount(sk, skb, diff);
1209 }
1210
1211 /* Link BUFF into the send queue. */
1212 __skb_header_release(buff);
1213 tcp_insert_write_queue_after(skb, buff, sk);
1214
1215 return 0;
1216}
1217
1218/* This is similar to __pskb_pull_head() (it will go to core/skbuff.c
1219 * eventually). The difference is that pulled data not copied, but
1220 * immediately discarded.
1221 */
1222static void __pskb_trim_head(struct sk_buff *skb, int len)
1223{
1224 struct skb_shared_info *shinfo;
1225 int i, k, eat;
1226
1227 eat = min_t(int, len, skb_headlen(skb));
1228 if (eat) {
1229 __skb_pull(skb, eat);
1230 len -= eat;
1231 if (!len)
1232 return;
1233 }
1234 eat = len;
1235 k = 0;
1236 shinfo = skb_shinfo(skb);
1237 for (i = 0; i < shinfo->nr_frags; i++) {
1238 int size = skb_frag_size(&shinfo->frags[i]);
1239
1240 if (size <= eat) {
1241 skb_frag_unref(skb, i);
1242 eat -= size;
1243 } else {
1244 shinfo->frags[k] = shinfo->frags[i];
1245 if (eat) {
1246 shinfo->frags[k].page_offset += eat;
1247 skb_frag_size_sub(&shinfo->frags[k], eat);
1248 eat = 0;
1249 }
1250 k++;
1251 }
1252 }
1253 shinfo->nr_frags = k;
1254
1255 skb_reset_tail_pointer(skb);
1256 skb->data_len -= len;
1257 skb->len = skb->data_len;
1258}
1259
1260/* Remove acked data from a packet in the transmit queue. */
1261int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1262{
1263 if (skb_unclone(skb, GFP_ATOMIC))
1264 return -ENOMEM;
1265
1266 __pskb_trim_head(skb, len);
1267
1268 TCP_SKB_CB(skb)->seq += len;
1269 skb->ip_summed = CHECKSUM_PARTIAL;
1270
1271 skb->truesize -= len;
1272 sk->sk_wmem_queued -= len;
1273 sk_mem_uncharge(sk, len);
1274 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1275
1276 /* Any change of skb->len requires recalculation of tso factor. */
1277 if (tcp_skb_pcount(skb) > 1)
1278 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1279
1280 return 0;
1281}
1282
1283/* Calculate MSS not accounting any TCP options. */
1284static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1285{
1286 const struct tcp_sock *tp = tcp_sk(sk);
1287 const struct inet_connection_sock *icsk = inet_csk(sk);
1288 int mss_now;
1289
1290 /* Calculate base mss without TCP options:
1291 It is MMS_S - sizeof(tcphdr) of rfc1122
1292 */
1293 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1294
1295 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1296 if (icsk->icsk_af_ops->net_frag_header_len) {
1297 const struct dst_entry *dst = __sk_dst_get(sk);
1298
1299 if (dst && dst_allfrag(dst))
1300 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1301 }
1302
1303 /* Clamp it (mss_clamp does not include tcp options) */
1304 if (mss_now > tp->rx_opt.mss_clamp)
1305 mss_now = tp->rx_opt.mss_clamp;
1306
1307 /* Now subtract optional transport overhead */
1308 mss_now -= icsk->icsk_ext_hdr_len;
1309
1310 /* Then reserve room for full set of TCP options and 8 bytes of data */
1311 if (mss_now < 48)
1312 mss_now = 48;
1313 return mss_now;
1314}
1315
1316/* Calculate MSS. Not accounting for SACKs here. */
1317int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1318{
1319 /* Subtract TCP options size, not including SACKs */
1320 return __tcp_mtu_to_mss(sk, pmtu) -
1321 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1322}
1323
1324/* Inverse of above */
1325int tcp_mss_to_mtu(struct sock *sk, int mss)
1326{
1327 const struct tcp_sock *tp = tcp_sk(sk);
1328 const struct inet_connection_sock *icsk = inet_csk(sk);
1329 int mtu;
1330
1331 mtu = mss +
1332 tp->tcp_header_len +
1333 icsk->icsk_ext_hdr_len +
1334 icsk->icsk_af_ops->net_header_len;
1335
1336 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1337 if (icsk->icsk_af_ops->net_frag_header_len) {
1338 const struct dst_entry *dst = __sk_dst_get(sk);
1339
1340 if (dst && dst_allfrag(dst))
1341 mtu += icsk->icsk_af_ops->net_frag_header_len;
1342 }
1343 return mtu;
1344}
1345
1346/* MTU probing init per socket */
1347void tcp_mtup_init(struct sock *sk)
1348{
1349 struct tcp_sock *tp = tcp_sk(sk);
1350 struct inet_connection_sock *icsk = inet_csk(sk);
1351 struct net *net = sock_net(sk);
1352
1353 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1354 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1355 icsk->icsk_af_ops->net_header_len;
1356 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1357 icsk->icsk_mtup.probe_size = 0;
1358 if (icsk->icsk_mtup.enabled)
1359 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1360}
1361EXPORT_SYMBOL(tcp_mtup_init);
1362
1363/* This function synchronize snd mss to current pmtu/exthdr set.
1364
1365 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1366 for TCP options, but includes only bare TCP header.
1367
1368 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1369 It is minimum of user_mss and mss received with SYN.
1370 It also does not include TCP options.
1371
1372 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1373
1374 tp->mss_cache is current effective sending mss, including
1375 all tcp options except for SACKs. It is evaluated,
1376 taking into account current pmtu, but never exceeds
1377 tp->rx_opt.mss_clamp.
1378
1379 NOTE1. rfc1122 clearly states that advertised MSS
1380 DOES NOT include either tcp or ip options.
1381
1382 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1383 are READ ONLY outside this function. --ANK (980731)
1384 */
1385unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1386{
1387 struct tcp_sock *tp = tcp_sk(sk);
1388 struct inet_connection_sock *icsk = inet_csk(sk);
1389 int mss_now;
1390
1391 if (icsk->icsk_mtup.search_high > pmtu)
1392 icsk->icsk_mtup.search_high = pmtu;
1393
1394 mss_now = tcp_mtu_to_mss(sk, pmtu);
1395 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1396
1397 /* And store cached results */
1398 icsk->icsk_pmtu_cookie = pmtu;
1399 if (icsk->icsk_mtup.enabled)
1400 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1401 tp->mss_cache = mss_now;
1402
1403 return mss_now;
1404}
1405EXPORT_SYMBOL(tcp_sync_mss);
1406
1407/* Compute the current effective MSS, taking SACKs and IP options,
1408 * and even PMTU discovery events into account.
1409 */
1410unsigned int tcp_current_mss(struct sock *sk)
1411{
1412 const struct tcp_sock *tp = tcp_sk(sk);
1413 const struct dst_entry *dst = __sk_dst_get(sk);
1414 u32 mss_now;
1415 unsigned int header_len;
1416 struct tcp_out_options opts;
1417 struct tcp_md5sig_key *md5;
1418
1419 mss_now = tp->mss_cache;
1420
1421 if (dst) {
1422 u32 mtu = dst_mtu(dst);
1423 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1424 mss_now = tcp_sync_mss(sk, mtu);
1425 }
1426
1427 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1428 sizeof(struct tcphdr);
1429 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1430 * some common options. If this is an odd packet (because we have SACK
1431 * blocks etc) then our calculated header_len will be different, and
1432 * we have to adjust mss_now correspondingly */
1433 if (header_len != tp->tcp_header_len) {
1434 int delta = (int) header_len - tp->tcp_header_len;
1435 mss_now -= delta;
1436 }
1437
1438 return mss_now;
1439}
1440
1441/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1442 * As additional protections, we do not touch cwnd in retransmission phases,
1443 * and if application hit its sndbuf limit recently.
1444 */
1445static void tcp_cwnd_application_limited(struct sock *sk)
1446{
1447 struct tcp_sock *tp = tcp_sk(sk);
1448
1449 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1450 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1451 /* Limited by application or receiver window. */
1452 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1453 u32 win_used = max(tp->snd_cwnd_used, init_win);
1454 if (win_used < tp->snd_cwnd) {
1455 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1456 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1457 }
1458 tp->snd_cwnd_used = 0;
1459 }
1460 tp->snd_cwnd_stamp = tcp_time_stamp;
1461}
1462
1463static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1464{
1465 struct tcp_sock *tp = tcp_sk(sk);
1466
1467 /* Track the maximum number of outstanding packets in each
1468 * window, and remember whether we were cwnd-limited then.
1469 */
1470 if (!before(tp->snd_una, tp->max_packets_seq) ||
1471 tp->packets_out > tp->max_packets_out) {
1472 tp->max_packets_out = tp->packets_out;
1473 tp->max_packets_seq = tp->snd_nxt;
1474 tp->is_cwnd_limited = is_cwnd_limited;
1475 }
1476
1477 if (tcp_is_cwnd_limited(sk)) {
1478 /* Network is feed fully. */
1479 tp->snd_cwnd_used = 0;
1480 tp->snd_cwnd_stamp = tcp_time_stamp;
1481 } else {
1482 /* Network starves. */
1483 if (tp->packets_out > tp->snd_cwnd_used)
1484 tp->snd_cwnd_used = tp->packets_out;
1485
1486 if (sysctl_tcp_slow_start_after_idle &&
1487 (s32)(tcp_time_stamp - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto)
1488 tcp_cwnd_application_limited(sk);
1489 }
1490}
1491
1492/* Minshall's variant of the Nagle send check. */
1493static bool tcp_minshall_check(const struct tcp_sock *tp)
1494{
1495 return after(tp->snd_sml, tp->snd_una) &&
1496 !after(tp->snd_sml, tp->snd_nxt);
1497}
1498
1499/* Update snd_sml if this skb is under mss
1500 * Note that a TSO packet might end with a sub-mss segment
1501 * The test is really :
1502 * if ((skb->len % mss) != 0)
1503 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1504 * But we can avoid doing the divide again given we already have
1505 * skb_pcount = skb->len / mss_now
1506 */
1507static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1508 const struct sk_buff *skb)
1509{
1510 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1511 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1512}
1513
1514/* Return false, if packet can be sent now without violation Nagle's rules:
1515 * 1. It is full sized. (provided by caller in %partial bool)
1516 * 2. Or it contains FIN. (already checked by caller)
1517 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1518 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1519 * With Minshall's modification: all sent small packets are ACKed.
1520 */
1521static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1522 int nonagle)
1523{
1524 return partial &&
1525 ((nonagle & TCP_NAGLE_CORK) ||
1526 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1527}
1528
1529/* Return how many segs we'd like on a TSO packet,
1530 * to send one TSO packet per ms
1531 */
1532static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now)
1533{
1534 u32 bytes, segs;
1535
1536 bytes = min(sk->sk_pacing_rate >> 10,
1537 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1538
1539 /* Goal is to send at least one packet per ms,
1540 * not one big TSO packet every 100 ms.
1541 * This preserves ACK clocking and is consistent
1542 * with tcp_tso_should_defer() heuristic.
1543 */
1544 segs = max_t(u32, bytes / mss_now, sysctl_tcp_min_tso_segs);
1545
1546 return min_t(u32, segs, sk->sk_gso_max_segs);
1547}
1548
1549/* Returns the portion of skb which can be sent right away */
1550static unsigned int tcp_mss_split_point(const struct sock *sk,
1551 const struct sk_buff *skb,
1552 unsigned int mss_now,
1553 unsigned int max_segs,
1554 int nonagle)
1555{
1556 const struct tcp_sock *tp = tcp_sk(sk);
1557 u32 partial, needed, window, max_len;
1558
1559 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1560 max_len = mss_now * max_segs;
1561
1562 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1563 return max_len;
1564
1565 needed = min(skb->len, window);
1566
1567 if (max_len <= needed)
1568 return max_len;
1569
1570 partial = needed % mss_now;
1571 /* If last segment is not a full MSS, check if Nagle rules allow us
1572 * to include this last segment in this skb.
1573 * Otherwise, we'll split the skb at last MSS boundary
1574 */
1575 if (tcp_nagle_check(partial != 0, tp, nonagle))
1576 return needed - partial;
1577
1578 return needed;
1579}
1580
1581/* Can at least one segment of SKB be sent right now, according to the
1582 * congestion window rules? If so, return how many segments are allowed.
1583 */
1584static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1585 const struct sk_buff *skb)
1586{
1587 u32 in_flight, cwnd, halfcwnd;
1588
1589 /* Don't be strict about the congestion window for the final FIN. */
1590 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1591 tcp_skb_pcount(skb) == 1)
1592 return 1;
1593
1594 in_flight = tcp_packets_in_flight(tp);
1595 cwnd = tp->snd_cwnd;
1596 if (in_flight >= cwnd)
1597 return 0;
1598
1599 /* For better scheduling, ensure we have at least
1600 * 2 GSO packets in flight.
1601 */
1602 halfcwnd = max(cwnd >> 1, 1U);
1603 return min(halfcwnd, cwnd - in_flight);
1604}
1605
1606/* Initialize TSO state of a skb.
1607 * This must be invoked the first time we consider transmitting
1608 * SKB onto the wire.
1609 */
1610static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1611{
1612 int tso_segs = tcp_skb_pcount(skb);
1613
1614 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1615 tcp_set_skb_tso_segs(skb, mss_now);
1616 tso_segs = tcp_skb_pcount(skb);
1617 }
1618 return tso_segs;
1619}
1620
1621
1622/* Return true if the Nagle test allows this packet to be
1623 * sent now.
1624 */
1625static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1626 unsigned int cur_mss, int nonagle)
1627{
1628 /* Nagle rule does not apply to frames, which sit in the middle of the
1629 * write_queue (they have no chances to get new data).
1630 *
1631 * This is implemented in the callers, where they modify the 'nonagle'
1632 * argument based upon the location of SKB in the send queue.
1633 */
1634 if (nonagle & TCP_NAGLE_PUSH)
1635 return true;
1636
1637 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1638 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1639 return true;
1640
1641 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1642 return true;
1643
1644 return false;
1645}
1646
1647/* Does at least the first segment of SKB fit into the send window? */
1648static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1649 const struct sk_buff *skb,
1650 unsigned int cur_mss)
1651{
1652 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1653
1654 if (skb->len > cur_mss)
1655 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1656
1657 return !after(end_seq, tcp_wnd_end(tp));
1658}
1659
1660/* This checks if the data bearing packet SKB (usually tcp_send_head(sk))
1661 * should be put on the wire right now. If so, it returns the number of
1662 * packets allowed by the congestion window.
1663 */
1664static unsigned int tcp_snd_test(const struct sock *sk, struct sk_buff *skb,
1665 unsigned int cur_mss, int nonagle)
1666{
1667 const struct tcp_sock *tp = tcp_sk(sk);
1668 unsigned int cwnd_quota;
1669
1670 tcp_init_tso_segs(skb, cur_mss);
1671
1672 if (!tcp_nagle_test(tp, skb, cur_mss, nonagle))
1673 return 0;
1674
1675 cwnd_quota = tcp_cwnd_test(tp, skb);
1676 if (cwnd_quota && !tcp_snd_wnd_test(tp, skb, cur_mss))
1677 cwnd_quota = 0;
1678
1679 return cwnd_quota;
1680}
1681
1682/* Test if sending is allowed right now. */
1683bool tcp_may_send_now(struct sock *sk)
1684{
1685 const struct tcp_sock *tp = tcp_sk(sk);
1686 struct sk_buff *skb = tcp_send_head(sk);
1687
1688 return skb &&
1689 tcp_snd_test(sk, skb, tcp_current_mss(sk),
1690 (tcp_skb_is_last(sk, skb) ?
1691 tp->nonagle : TCP_NAGLE_PUSH));
1692}
1693
1694/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1695 * which is put after SKB on the list. It is very much like
1696 * tcp_fragment() except that it may make several kinds of assumptions
1697 * in order to speed up the splitting operation. In particular, we
1698 * know that all the data is in scatter-gather pages, and that the
1699 * packet has never been sent out before (and thus is not cloned).
1700 */
1701static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
1702 unsigned int mss_now, gfp_t gfp)
1703{
1704 struct sk_buff *buff;
1705 int nlen = skb->len - len;
1706 u8 flags;
1707
1708 /* All of a TSO frame must be composed of paged data. */
1709 if (skb->len != skb->data_len)
1710 return tcp_fragment(sk, skb, len, mss_now, gfp);
1711
1712 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1713 if (unlikely(!buff))
1714 return -ENOMEM;
1715
1716 sk->sk_wmem_queued += buff->truesize;
1717 sk_mem_charge(sk, buff->truesize);
1718 buff->truesize += nlen;
1719 skb->truesize -= nlen;
1720
1721 /* Correct the sequence numbers. */
1722 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1723 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1724 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1725
1726 /* PSH and FIN should only be set in the second packet. */
1727 flags = TCP_SKB_CB(skb)->tcp_flags;
1728 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1729 TCP_SKB_CB(buff)->tcp_flags = flags;
1730
1731 /* This packet was never sent out yet, so no SACK bits. */
1732 TCP_SKB_CB(buff)->sacked = 0;
1733
1734 buff->ip_summed = skb->ip_summed = CHECKSUM_PARTIAL;
1735 skb_split(skb, buff, len);
1736 tcp_fragment_tstamp(skb, buff);
1737
1738 /* Fix up tso_factor for both original and new SKB. */
1739 tcp_set_skb_tso_segs(skb, mss_now);
1740 tcp_set_skb_tso_segs(buff, mss_now);
1741
1742 /* Link BUFF into the send queue. */
1743 __skb_header_release(buff);
1744 tcp_insert_write_queue_after(skb, buff, sk);
1745
1746 return 0;
1747}
1748
1749/* Try to defer sending, if possible, in order to minimize the amount
1750 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1751 *
1752 * This algorithm is from John Heffner.
1753 */
1754static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1755 bool *is_cwnd_limited, u32 max_segs)
1756{
1757 const struct inet_connection_sock *icsk = inet_csk(sk);
1758 u32 age, send_win, cong_win, limit, in_flight;
1759 struct tcp_sock *tp = tcp_sk(sk);
1760 struct skb_mstamp now;
1761 struct sk_buff *head;
1762 int win_divisor;
1763
1764 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1765 goto send_now;
1766
1767 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1768 goto send_now;
1769
1770 /* Avoid bursty behavior by allowing defer
1771 * only if the last write was recent.
1772 */
1773 if ((s32)(tcp_time_stamp - tp->lsndtime) > 0)
1774 goto send_now;
1775
1776 in_flight = tcp_packets_in_flight(tp);
1777
1778 BUG_ON(tcp_skb_pcount(skb) <= 1 || (tp->snd_cwnd <= in_flight));
1779
1780 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1781
1782 /* From in_flight test above, we know that cwnd > in_flight. */
1783 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1784
1785 limit = min(send_win, cong_win);
1786
1787 /* If a full-sized TSO skb can be sent, do it. */
1788 if (limit >= max_segs * tp->mss_cache)
1789 goto send_now;
1790
1791 /* Middle in queue won't get any more data, full sendable already? */
1792 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1793 goto send_now;
1794
1795 win_divisor = ACCESS_ONCE(sysctl_tcp_tso_win_divisor);
1796 if (win_divisor) {
1797 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1798
1799 /* If at least some fraction of a window is available,
1800 * just use it.
1801 */
1802 chunk /= win_divisor;
1803 if (limit >= chunk)
1804 goto send_now;
1805 } else {
1806 /* Different approach, try not to defer past a single
1807 * ACK. Receiver should ACK every other full sized
1808 * frame, so if we have space for more than 3 frames
1809 * then send now.
1810 */
1811 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1812 goto send_now;
1813 }
1814
1815 head = tcp_write_queue_head(sk);
1816 skb_mstamp_get(&now);
1817 age = skb_mstamp_us_delta(&now, &head->skb_mstamp);
1818 /* If next ACK is likely to come too late (half srtt), do not defer */
1819 if (age < (tp->srtt_us >> 4))
1820 goto send_now;
1821
1822 /* Ok, it looks like it is advisable to defer. */
1823
1824 if (cong_win < send_win && cong_win <= skb->len)
1825 *is_cwnd_limited = true;
1826
1827 return true;
1828
1829send_now:
1830 return false;
1831}
1832
1833static inline void tcp_mtu_check_reprobe(struct sock *sk)
1834{
1835 struct inet_connection_sock *icsk = inet_csk(sk);
1836 struct tcp_sock *tp = tcp_sk(sk);
1837 struct net *net = sock_net(sk);
1838 u32 interval;
1839 s32 delta;
1840
1841 interval = net->ipv4.sysctl_tcp_probe_interval;
1842 delta = tcp_time_stamp - icsk->icsk_mtup.probe_timestamp;
1843 if (unlikely(delta >= interval * HZ)) {
1844 int mss = tcp_current_mss(sk);
1845
1846 /* Update current search range */
1847 icsk->icsk_mtup.probe_size = 0;
1848 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
1849 sizeof(struct tcphdr) +
1850 icsk->icsk_af_ops->net_header_len;
1851 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
1852
1853 /* Update probe time stamp */
1854 icsk->icsk_mtup.probe_timestamp = tcp_time_stamp;
1855 }
1856}
1857
1858/* Create a new MTU probe if we are ready.
1859 * MTU probe is regularly attempting to increase the path MTU by
1860 * deliberately sending larger packets. This discovers routing
1861 * changes resulting in larger path MTUs.
1862 *
1863 * Returns 0 if we should wait to probe (no cwnd available),
1864 * 1 if a probe was sent,
1865 * -1 otherwise
1866 */
1867static int tcp_mtu_probe(struct sock *sk)
1868{
1869 struct tcp_sock *tp = tcp_sk(sk);
1870 struct inet_connection_sock *icsk = inet_csk(sk);
1871 struct sk_buff *skb, *nskb, *next;
1872 struct net *net = sock_net(sk);
1873 int len;
1874 int probe_size;
1875 int size_needed;
1876 int copy;
1877 int mss_now;
1878 int interval;
1879
1880 /* Not currently probing/verifying,
1881 * not in recovery,
1882 * have enough cwnd, and
1883 * not SACKing (the variable headers throw things off) */
1884 if (!icsk->icsk_mtup.enabled ||
1885 icsk->icsk_mtup.probe_size ||
1886 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
1887 tp->snd_cwnd < 11 ||
1888 tp->rx_opt.num_sacks || tp->rx_opt.dsack)
1889 return -1;
1890
1891 /* Use binary search for probe_size between tcp_mss_base,
1892 * and current mss_clamp. if (search_high - search_low)
1893 * smaller than a threshold, backoff from probing.
1894 */
1895 mss_now = tcp_current_mss(sk);
1896 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
1897 icsk->icsk_mtup.search_low) >> 1);
1898 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
1899 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
1900 /* When misfortune happens, we are reprobing actively,
1901 * and then reprobe timer has expired. We stick with current
1902 * probing process by not resetting search range to its orignal.
1903 */
1904 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
1905 interval < net->ipv4.sysctl_tcp_probe_threshold) {
1906 /* Check whether enough time has elaplased for
1907 * another round of probing.
1908 */
1909 tcp_mtu_check_reprobe(sk);
1910 return -1;
1911 }
1912
1913 /* Have enough data in the send queue to probe? */
1914 if (tp->write_seq - tp->snd_nxt < size_needed)
1915 return -1;
1916
1917 if (tp->snd_wnd < size_needed)
1918 return -1;
1919 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
1920 return 0;
1921
1922 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
1923 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
1924 if (!tcp_packets_in_flight(tp))
1925 return -1;
1926 else
1927 return 0;
1928 }
1929
1930 /* We're allowed to probe. Build it now. */
1931 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
1932 if (!nskb)
1933 return -1;
1934 sk->sk_wmem_queued += nskb->truesize;
1935 sk_mem_charge(sk, nskb->truesize);
1936
1937 skb = tcp_send_head(sk);
1938
1939 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
1940 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
1941 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
1942 TCP_SKB_CB(nskb)->sacked = 0;
1943 nskb->csum = 0;
1944 nskb->ip_summed = skb->ip_summed;
1945
1946 tcp_insert_write_queue_before(nskb, skb, sk);
1947
1948 len = 0;
1949 tcp_for_write_queue_from_safe(skb, next, sk) {
1950 copy = min_t(int, skb->len, probe_size - len);
1951 if (nskb->ip_summed)
1952 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
1953 else
1954 nskb->csum = skb_copy_and_csum_bits(skb, 0,
1955 skb_put(nskb, copy),
1956 copy, nskb->csum);
1957
1958 if (skb->len <= copy) {
1959 /* We've eaten all the data from this skb.
1960 * Throw it away. */
1961 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
1962 tcp_unlink_write_queue(skb, sk);
1963 sk_wmem_free_skb(sk, skb);
1964 } else {
1965 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
1966 ~(TCPHDR_FIN|TCPHDR_PSH);
1967 if (!skb_shinfo(skb)->nr_frags) {
1968 skb_pull(skb, copy);
1969 if (skb->ip_summed != CHECKSUM_PARTIAL)
1970 skb->csum = csum_partial(skb->data,
1971 skb->len, 0);
1972 } else {
1973 __pskb_trim_head(skb, copy);
1974 tcp_set_skb_tso_segs(skb, mss_now);
1975 }
1976 TCP_SKB_CB(skb)->seq += copy;
1977 }
1978
1979 len += copy;
1980
1981 if (len >= probe_size)
1982 break;
1983 }
1984 tcp_init_tso_segs(nskb, nskb->len);
1985
1986 /* We're ready to send. If this fails, the probe will
1987 * be resegmented into mss-sized pieces by tcp_write_xmit().
1988 */
1989 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
1990 /* Decrement cwnd here because we are sending
1991 * effectively two packets. */
1992 tp->snd_cwnd--;
1993 tcp_event_new_data_sent(sk, nskb);
1994
1995 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
1996 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
1997 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
1998
1999 return 1;
2000 }
2001
2002 return -1;
2003}
2004
2005/* This routine writes packets to the network. It advances the
2006 * send_head. This happens as incoming acks open up the remote
2007 * window for us.
2008 *
2009 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2010 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2011 * account rare use of URG, this is not a big flaw.
2012 *
2013 * Send at most one packet when push_one > 0. Temporarily ignore
2014 * cwnd limit to force at most one packet out when push_one == 2.
2015
2016 * Returns true, if no segments are in flight and we have queued segments,
2017 * but cannot send anything now because of SWS or another problem.
2018 */
2019static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2020 int push_one, gfp_t gfp)
2021{
2022 struct tcp_sock *tp = tcp_sk(sk);
2023 struct sk_buff *skb;
2024 unsigned int tso_segs, sent_pkts;
2025 int cwnd_quota;
2026 int result;
2027 bool is_cwnd_limited = false;
2028 u32 max_segs;
2029
2030 sent_pkts = 0;
2031
2032 if (!push_one) {
2033 /* Do MTU probing. */
2034 result = tcp_mtu_probe(sk);
2035 if (!result) {
2036 return false;
2037 } else if (result > 0) {
2038 sent_pkts = 1;
2039 }
2040 }
2041
2042 max_segs = tcp_tso_autosize(sk, mss_now);
2043 while ((skb = tcp_send_head(sk))) {
2044 unsigned int limit;
2045
2046 tso_segs = tcp_init_tso_segs(skb, mss_now);
2047 BUG_ON(!tso_segs);
2048
2049 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2050 /* "skb_mstamp" is used as a start point for the retransmit timer */
2051 skb_mstamp_get(&skb->skb_mstamp);
2052 goto repair; /* Skip network transmission */
2053 }
2054
2055 cwnd_quota = tcp_cwnd_test(tp, skb);
2056 if (!cwnd_quota) {
2057 if (push_one == 2)
2058 /* Force out a loss probe pkt. */
2059 cwnd_quota = 1;
2060 else
2061 break;
2062 }
2063
2064 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now)))
2065 break;
2066
2067 if (tso_segs == 1) {
2068 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2069 (tcp_skb_is_last(sk, skb) ?
2070 nonagle : TCP_NAGLE_PUSH))))
2071 break;
2072 } else {
2073 if (!push_one &&
2074 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2075 max_segs))
2076 break;
2077 }
2078
2079 limit = mss_now;
2080 if (tso_segs > 1 && !tcp_urg_mode(tp))
2081 limit = tcp_mss_split_point(sk, skb, mss_now,
2082 min_t(unsigned int,
2083 cwnd_quota,
2084 max_segs),
2085 nonagle);
2086
2087 if (skb->len > limit &&
2088 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2089 break;
2090
2091 /* TCP Small Queues :
2092 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2093 * This allows for :
2094 * - better RTT estimation and ACK scheduling
2095 * - faster recovery
2096 * - high rates
2097 * Alas, some drivers / subsystems require a fair amount
2098 * of queued bytes to ensure line rate.
2099 * One example is wifi aggregation (802.11 AMPDU)
2100 */
2101 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> 10);
2102 limit = min_t(u32, limit, sysctl_tcp_limit_output_bytes);
2103
2104 if (atomic_read(&sk->sk_wmem_alloc) > limit) {
2105 set_bit(TSQ_THROTTLED, &tp->tsq_flags);
2106 /* It is possible TX completion already happened
2107 * before we set TSQ_THROTTLED, so we must
2108 * test again the condition.
2109 */
2110 smp_mb__after_atomic();
2111 if (atomic_read(&sk->sk_wmem_alloc) > limit)
2112 break;
2113 }
2114
2115 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2116 break;
2117
2118repair:
2119 /* Advance the send_head. This one is sent out.
2120 * This call will increment packets_out.
2121 */
2122 tcp_event_new_data_sent(sk, skb);
2123
2124 tcp_minshall_update(tp, mss_now, skb);
2125 sent_pkts += tcp_skb_pcount(skb);
2126
2127 if (push_one)
2128 break;
2129 }
2130
2131 if (likely(sent_pkts)) {
2132 if (tcp_in_cwnd_reduction(sk))
2133 tp->prr_out += sent_pkts;
2134
2135 /* Send one loss probe per tail loss episode. */
2136 if (push_one != 2)
2137 tcp_schedule_loss_probe(sk);
2138 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2139 tcp_cwnd_validate(sk, is_cwnd_limited);
2140 return false;
2141 }
2142 return !tp->packets_out && tcp_send_head(sk);
2143}
2144
2145bool tcp_schedule_loss_probe(struct sock *sk)
2146{
2147 struct inet_connection_sock *icsk = inet_csk(sk);
2148 struct tcp_sock *tp = tcp_sk(sk);
2149 u32 timeout, tlp_time_stamp, rto_time_stamp;
2150 u32 rtt = usecs_to_jiffies(tp->srtt_us >> 3);
2151
2152 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_EARLY_RETRANS))
2153 return false;
2154 /* No consecutive loss probes. */
2155 if (WARN_ON(icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)) {
2156 tcp_rearm_rto(sk);
2157 return false;
2158 }
2159 /* Don't do any loss probe on a Fast Open connection before 3WHS
2160 * finishes.
2161 */
2162 if (tp->fastopen_rsk)
2163 return false;
2164
2165 /* TLP is only scheduled when next timer event is RTO. */
2166 if (icsk->icsk_pending != ICSK_TIME_RETRANS)
2167 return false;
2168
2169 /* Schedule a loss probe in 2*RTT for SACK capable connections
2170 * in Open state, that are either limited by cwnd or application.
2171 */
2172 if (sysctl_tcp_early_retrans < 3 || !tp->packets_out ||
2173 !tcp_is_sack(tp) || inet_csk(sk)->icsk_ca_state != TCP_CA_Open)
2174 return false;
2175
2176 if ((tp->snd_cwnd > tcp_packets_in_flight(tp)) &&
2177 tcp_send_head(sk))
2178 return false;
2179
2180 /* Probe timeout is at least 1.5*rtt + TCP_DELACK_MAX to account
2181 * for delayed ack when there's one outstanding packet. If no RTT
2182 * sample is available then probe after TCP_TIMEOUT_INIT.
2183 */
2184 timeout = rtt << 1 ? : TCP_TIMEOUT_INIT;
2185 if (tp->packets_out == 1)
2186 timeout = max_t(u32, timeout,
2187 (rtt + (rtt >> 1) + TCP_DELACK_MAX));
2188 timeout = max_t(u32, timeout, msecs_to_jiffies(10));
2189
2190 /* If RTO is shorter, just schedule TLP in its place. */
2191 tlp_time_stamp = tcp_time_stamp + timeout;
2192 rto_time_stamp = (u32)inet_csk(sk)->icsk_timeout;
2193 if ((s32)(tlp_time_stamp - rto_time_stamp) > 0) {
2194 s32 delta = rto_time_stamp - tcp_time_stamp;
2195 if (delta > 0)
2196 timeout = delta;
2197 }
2198
2199 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2200 TCP_RTO_MAX);
2201 return true;
2202}
2203
2204/* Thanks to skb fast clones, we can detect if a prior transmit of
2205 * a packet is still in a qdisc or driver queue.
2206 * In this case, there is very little point doing a retransmit !
2207 * Note: This is called from BH context only.
2208 */
2209static bool skb_still_in_host_queue(const struct sock *sk,
2210 const struct sk_buff *skb)
2211{
2212 if (unlikely(skb_fclone_busy(sk, skb))) {
2213 NET_INC_STATS_BH(sock_net(sk),
2214 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2215 return true;
2216 }
2217 return false;
2218}
2219
2220/* When probe timeout (PTO) fires, try send a new segment if possible, else
2221 * retransmit the last segment.
2222 */
2223void tcp_send_loss_probe(struct sock *sk)
2224{
2225 struct tcp_sock *tp = tcp_sk(sk);
2226 struct sk_buff *skb;
2227 int pcount;
2228 int mss = tcp_current_mss(sk);
2229
2230 skb = tcp_send_head(sk);
2231 if (skb) {
2232 if (tcp_snd_wnd_test(tp, skb, mss)) {
2233 pcount = tp->packets_out;
2234 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2235 if (tp->packets_out > pcount)
2236 goto probe_sent;
2237 goto rearm_timer;
2238 }
2239 skb = tcp_write_queue_prev(sk, skb);
2240 } else {
2241 skb = tcp_write_queue_tail(sk);
2242 }
2243
2244 /* At most one outstanding TLP retransmission. */
2245 if (tp->tlp_high_seq)
2246 goto rearm_timer;
2247
2248 /* Retransmit last segment. */
2249 if (WARN_ON(!skb))
2250 goto rearm_timer;
2251
2252 if (skb_still_in_host_queue(sk, skb))
2253 goto rearm_timer;
2254
2255 pcount = tcp_skb_pcount(skb);
2256 if (WARN_ON(!pcount))
2257 goto rearm_timer;
2258
2259 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2260 if (unlikely(tcp_fragment(sk, skb, (pcount - 1) * mss, mss,
2261 GFP_ATOMIC)))
2262 goto rearm_timer;
2263 skb = tcp_write_queue_next(sk, skb);
2264 }
2265
2266 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2267 goto rearm_timer;
2268
2269 if (__tcp_retransmit_skb(sk, skb))
2270 goto rearm_timer;
2271
2272 /* Record snd_nxt for loss detection. */
2273 tp->tlp_high_seq = tp->snd_nxt;
2274
2275probe_sent:
2276 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2277 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2278 inet_csk(sk)->icsk_pending = 0;
2279rearm_timer:
2280 tcp_rearm_rto(sk);
2281}
2282
2283/* Push out any pending frames which were held back due to
2284 * TCP_CORK or attempt at coalescing tiny packets.
2285 * The socket must be locked by the caller.
2286 */
2287void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2288 int nonagle)
2289{
2290 /* If we are closed, the bytes will have to remain here.
2291 * In time closedown will finish, we empty the write queue and
2292 * all will be happy.
2293 */
2294 if (unlikely(sk->sk_state == TCP_CLOSE))
2295 return;
2296
2297 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2298 sk_gfp_mask(sk, GFP_ATOMIC)))
2299 tcp_check_probe_timer(sk);
2300}
2301
2302/* Send _single_ skb sitting at the send head. This function requires
2303 * true push pending frames to setup probe timer etc.
2304 */
2305void tcp_push_one(struct sock *sk, unsigned int mss_now)
2306{
2307 struct sk_buff *skb = tcp_send_head(sk);
2308
2309 BUG_ON(!skb || skb->len < mss_now);
2310
2311 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2312}
2313
2314/* This function returns the amount that we can raise the
2315 * usable window based on the following constraints
2316 *
2317 * 1. The window can never be shrunk once it is offered (RFC 793)
2318 * 2. We limit memory per socket
2319 *
2320 * RFC 1122:
2321 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2322 * RECV.NEXT + RCV.WIN fixed until:
2323 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2324 *
2325 * i.e. don't raise the right edge of the window until you can raise
2326 * it at least MSS bytes.
2327 *
2328 * Unfortunately, the recommended algorithm breaks header prediction,
2329 * since header prediction assumes th->window stays fixed.
2330 *
2331 * Strictly speaking, keeping th->window fixed violates the receiver
2332 * side SWS prevention criteria. The problem is that under this rule
2333 * a stream of single byte packets will cause the right side of the
2334 * window to always advance by a single byte.
2335 *
2336 * Of course, if the sender implements sender side SWS prevention
2337 * then this will not be a problem.
2338 *
2339 * BSD seems to make the following compromise:
2340 *
2341 * If the free space is less than the 1/4 of the maximum
2342 * space available and the free space is less than 1/2 mss,
2343 * then set the window to 0.
2344 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2345 * Otherwise, just prevent the window from shrinking
2346 * and from being larger than the largest representable value.
2347 *
2348 * This prevents incremental opening of the window in the regime
2349 * where TCP is limited by the speed of the reader side taking
2350 * data out of the TCP receive queue. It does nothing about
2351 * those cases where the window is constrained on the sender side
2352 * because the pipeline is full.
2353 *
2354 * BSD also seems to "accidentally" limit itself to windows that are a
2355 * multiple of MSS, at least until the free space gets quite small.
2356 * This would appear to be a side effect of the mbuf implementation.
2357 * Combining these two algorithms results in the observed behavior
2358 * of having a fixed window size at almost all times.
2359 *
2360 * Below we obtain similar behavior by forcing the offered window to
2361 * a multiple of the mss when it is feasible to do so.
2362 *
2363 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2364 * Regular options like TIMESTAMP are taken into account.
2365 */
2366u32 __tcp_select_window(struct sock *sk)
2367{
2368 struct inet_connection_sock *icsk = inet_csk(sk);
2369 struct tcp_sock *tp = tcp_sk(sk);
2370 /* MSS for the peer's data. Previous versions used mss_clamp
2371 * here. I don't know if the value based on our guesses
2372 * of peer's MSS is better for the performance. It's more correct
2373 * but may be worse for the performance because of rcv_mss
2374 * fluctuations. --SAW 1998/11/1
2375 */
2376 int mss = icsk->icsk_ack.rcv_mss;
2377 int free_space = tcp_space(sk);
2378 int allowed_space = tcp_full_space(sk);
2379 int full_space = min_t(int, tp->window_clamp, allowed_space);
2380 int window;
2381
2382 if (mss > full_space)
2383 mss = full_space;
2384
2385 if (free_space < (full_space >> 1)) {
2386 icsk->icsk_ack.quick = 0;
2387
2388 if (tcp_under_memory_pressure(sk))
2389 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2390 4U * tp->advmss);
2391
2392 /* free_space might become our new window, make sure we don't
2393 * increase it due to wscale.
2394 */
2395 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2396
2397 /* if free space is less than mss estimate, or is below 1/16th
2398 * of the maximum allowed, try to move to zero-window, else
2399 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2400 * new incoming data is dropped due to memory limits.
2401 * With large window, mss test triggers way too late in order
2402 * to announce zero window in time before rmem limit kicks in.
2403 */
2404 if (free_space < (allowed_space >> 4) || free_space < mss)
2405 return 0;
2406 }
2407
2408 if (free_space > tp->rcv_ssthresh)
2409 free_space = tp->rcv_ssthresh;
2410
2411 /* Don't do rounding if we are using window scaling, since the
2412 * scaled window will not line up with the MSS boundary anyway.
2413 */
2414 window = tp->rcv_wnd;
2415 if (tp->rx_opt.rcv_wscale) {
2416 window = free_space;
2417
2418 /* Advertise enough space so that it won't get scaled away.
2419 * Import case: prevent zero window announcement if
2420 * 1<<rcv_wscale > mss.
2421 */
2422 if (((window >> tp->rx_opt.rcv_wscale) << tp->rx_opt.rcv_wscale) != window)
2423 window = (((window >> tp->rx_opt.rcv_wscale) + 1)
2424 << tp->rx_opt.rcv_wscale);
2425 } else {
2426 /* Get the largest window that is a nice multiple of mss.
2427 * Window clamp already applied above.
2428 * If our current window offering is within 1 mss of the
2429 * free space we just keep it. This prevents the divide
2430 * and multiply from happening most of the time.
2431 * We also don't do any window rounding when the free space
2432 * is too small.
2433 */
2434 if (window <= free_space - mss || window > free_space)
2435 window = (free_space / mss) * mss;
2436 else if (mss == full_space &&
2437 free_space > window + (full_space >> 1))
2438 window = free_space;
2439 }
2440
2441 return window;
2442}
2443
2444void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2445 const struct sk_buff *next_skb)
2446{
2447 const struct skb_shared_info *next_shinfo = skb_shinfo(next_skb);
2448 u8 tsflags = next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2449
2450 if (unlikely(tsflags)) {
2451 struct skb_shared_info *shinfo = skb_shinfo(skb);
2452
2453 shinfo->tx_flags |= tsflags;
2454 shinfo->tskey = next_shinfo->tskey;
2455 }
2456}
2457
2458/* Collapses two adjacent SKB's during retransmission. */
2459static void tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2460{
2461 struct tcp_sock *tp = tcp_sk(sk);
2462 struct sk_buff *next_skb = tcp_write_queue_next(sk, skb);
2463 int skb_size, next_skb_size;
2464
2465 skb_size = skb->len;
2466 next_skb_size = next_skb->len;
2467
2468 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2469
2470 tcp_highest_sack_combine(sk, next_skb, skb);
2471
2472 tcp_unlink_write_queue(next_skb, sk);
2473
2474 skb_copy_from_linear_data(next_skb, skb_put(skb, next_skb_size),
2475 next_skb_size);
2476
2477 if (next_skb->ip_summed == CHECKSUM_PARTIAL)
2478 skb->ip_summed = CHECKSUM_PARTIAL;
2479
2480 if (skb->ip_summed != CHECKSUM_PARTIAL)
2481 skb->csum = csum_block_add(skb->csum, next_skb->csum, skb_size);
2482
2483 /* Update sequence range on original skb. */
2484 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2485
2486 /* Merge over control information. This moves PSH/FIN etc. over */
2487 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2488
2489 /* All done, get rid of second SKB and account for it so
2490 * packet counting does not break.
2491 */
2492 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2493
2494 /* changed transmit queue under us so clear hints */
2495 tcp_clear_retrans_hints_partial(tp);
2496 if (next_skb == tp->retransmit_skb_hint)
2497 tp->retransmit_skb_hint = skb;
2498
2499 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2500
2501 tcp_skb_collapse_tstamp(skb, next_skb);
2502
2503 sk_wmem_free_skb(sk, next_skb);
2504}
2505
2506/* Check if coalescing SKBs is legal. */
2507static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2508{
2509 if (tcp_skb_pcount(skb) > 1)
2510 return false;
2511 /* TODO: SACK collapsing could be used to remove this condition */
2512 if (skb_shinfo(skb)->nr_frags != 0)
2513 return false;
2514 if (skb_cloned(skb))
2515 return false;
2516 if (skb == tcp_send_head(sk))
2517 return false;
2518 /* Some heurestics for collapsing over SACK'd could be invented */
2519 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2520 return false;
2521
2522 return true;
2523}
2524
2525/* Collapse packets in the retransmit queue to make to create
2526 * less packets on the wire. This is only done on retransmission.
2527 */
2528static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2529 int space)
2530{
2531 struct tcp_sock *tp = tcp_sk(sk);
2532 struct sk_buff *skb = to, *tmp;
2533 bool first = true;
2534
2535 if (!sysctl_tcp_retrans_collapse)
2536 return;
2537 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2538 return;
2539
2540 tcp_for_write_queue_from_safe(skb, tmp, sk) {
2541 if (!tcp_can_collapse(sk, skb))
2542 break;
2543
2544 space -= skb->len;
2545
2546 if (first) {
2547 first = false;
2548 continue;
2549 }
2550
2551 if (space < 0)
2552 break;
2553 /* Punt if not enough space exists in the first SKB for
2554 * the data in the second
2555 */
2556 if (skb->len > skb_availroom(to))
2557 break;
2558
2559 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2560 break;
2561
2562 tcp_collapse_retrans(sk, to);
2563 }
2564}
2565
2566/* This retransmits one SKB. Policy decisions and retransmit queue
2567 * state updates are done by the caller. Returns non-zero if an
2568 * error occurred which prevented the send.
2569 */
2570int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2571{
2572 struct tcp_sock *tp = tcp_sk(sk);
2573 struct inet_connection_sock *icsk = inet_csk(sk);
2574 unsigned int cur_mss;
2575 int err;
2576
2577 /* Inconslusive MTU probe */
2578 if (icsk->icsk_mtup.probe_size) {
2579 icsk->icsk_mtup.probe_size = 0;
2580 }
2581
2582 /* Do not sent more than we queued. 1/4 is reserved for possible
2583 * copying overhead: fragmentation, tunneling, mangling etc.
2584 */
2585 if (atomic_read(&sk->sk_wmem_alloc) >
2586 min(sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2), sk->sk_sndbuf))
2587 return -EAGAIN;
2588
2589 if (skb_still_in_host_queue(sk, skb))
2590 return -EBUSY;
2591
2592 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2593 if (before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))
2594 BUG();
2595 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2596 return -ENOMEM;
2597 }
2598
2599 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2600 return -EHOSTUNREACH; /* Routing failure or similar. */
2601
2602 cur_mss = tcp_current_mss(sk);
2603
2604 /* If receiver has shrunk his window, and skb is out of
2605 * new window, do not retransmit it. The exception is the
2606 * case, when window is shrunk to zero. In this case
2607 * our retransmit serves as a zero window probe.
2608 */
2609 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2610 TCP_SKB_CB(skb)->seq != tp->snd_una)
2611 return -EAGAIN;
2612
2613 if (skb->len > cur_mss) {
2614 if (tcp_fragment(sk, skb, cur_mss, cur_mss, GFP_ATOMIC))
2615 return -ENOMEM; /* We'll try again later. */
2616 } else {
2617 int oldpcount = tcp_skb_pcount(skb);
2618
2619 if (unlikely(oldpcount > 1)) {
2620 if (skb_unclone(skb, GFP_ATOMIC))
2621 return -ENOMEM;
2622 tcp_init_tso_segs(skb, cur_mss);
2623 tcp_adjust_pcount(sk, skb, oldpcount - tcp_skb_pcount(skb));
2624 }
2625 }
2626
2627 /* RFC3168, section 6.1.1.1. ECN fallback */
2628 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2629 tcp_ecn_clear_syn(sk, skb);
2630
2631 tcp_retrans_try_collapse(sk, skb, cur_mss);
2632
2633 /* Make a copy, if the first transmission SKB clone we made
2634 * is still in somebody's hands, else make a clone.
2635 */
2636
2637 /* make sure skb->data is aligned on arches that require it
2638 * and check if ack-trimming & collapsing extended the headroom
2639 * beyond what csum_start can cover.
2640 */
2641 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2642 skb_headroom(skb) >= 0xFFFF)) {
2643 struct sk_buff *nskb;
2644
2645 skb_mstamp_get(&skb->skb_mstamp);
2646 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2647 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2648 -ENOBUFS;
2649 } else {
2650 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2651 }
2652
2653 if (likely(!err)) {
2654 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2655 /* Update global TCP statistics. */
2656 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
2657 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2658 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2659 tp->total_retrans++;
2660 }
2661 return err;
2662}
2663
2664int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb)
2665{
2666 struct tcp_sock *tp = tcp_sk(sk);
2667 int err = __tcp_retransmit_skb(sk, skb);
2668
2669 if (err == 0) {
2670#if FASTRETRANS_DEBUG > 0
2671 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2672 net_dbg_ratelimited("retrans_out leaked\n");
2673 }
2674#endif
2675 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2676 tp->retrans_out += tcp_skb_pcount(skb);
2677
2678 /* Save stamp of the first retransmit. */
2679 if (!tp->retrans_stamp)
2680 tp->retrans_stamp = tcp_skb_timestamp(skb);
2681
2682 } else if (err != -EBUSY) {
2683 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2684 }
2685
2686 if (tp->undo_retrans < 0)
2687 tp->undo_retrans = 0;
2688 tp->undo_retrans += tcp_skb_pcount(skb);
2689 return err;
2690}
2691
2692/* Check if we forward retransmits are possible in the current
2693 * window/congestion state.
2694 */
2695static bool tcp_can_forward_retransmit(struct sock *sk)
2696{
2697 const struct inet_connection_sock *icsk = inet_csk(sk);
2698 const struct tcp_sock *tp = tcp_sk(sk);
2699
2700 /* Forward retransmissions are possible only during Recovery. */
2701 if (icsk->icsk_ca_state != TCP_CA_Recovery)
2702 return false;
2703
2704 /* No forward retransmissions in Reno are possible. */
2705 if (tcp_is_reno(tp))
2706 return false;
2707
2708 /* Yeah, we have to make difficult choice between forward transmission
2709 * and retransmission... Both ways have their merits...
2710 *
2711 * For now we do not retransmit anything, while we have some new
2712 * segments to send. In the other cases, follow rule 3 for
2713 * NextSeg() specified in RFC3517.
2714 */
2715
2716 if (tcp_may_send_now(sk))
2717 return false;
2718
2719 return true;
2720}
2721
2722/* This gets called after a retransmit timeout, and the initially
2723 * retransmitted data is acknowledged. It tries to continue
2724 * resending the rest of the retransmit queue, until either
2725 * we've sent it all or the congestion window limit is reached.
2726 * If doing SACK, the first ACK which comes back for a timeout
2727 * based retransmit packet might feed us FACK information again.
2728 * If so, we use it to avoid unnecessarily retransmissions.
2729 */
2730void tcp_xmit_retransmit_queue(struct sock *sk)
2731{
2732 const struct inet_connection_sock *icsk = inet_csk(sk);
2733 struct tcp_sock *tp = tcp_sk(sk);
2734 struct sk_buff *skb;
2735 struct sk_buff *hole = NULL;
2736 u32 last_lost;
2737 int mib_idx;
2738 int fwd_rexmitting = 0;
2739
2740 if (!tp->packets_out)
2741 return;
2742
2743 if (!tp->lost_out)
2744 tp->retransmit_high = tp->snd_una;
2745
2746 if (tp->retransmit_skb_hint) {
2747 skb = tp->retransmit_skb_hint;
2748 last_lost = TCP_SKB_CB(skb)->end_seq;
2749 if (after(last_lost, tp->retransmit_high))
2750 last_lost = tp->retransmit_high;
2751 } else {
2752 skb = tcp_write_queue_head(sk);
2753 last_lost = tp->snd_una;
2754 }
2755
2756 tcp_for_write_queue_from(skb, sk) {
2757 __u8 sacked = TCP_SKB_CB(skb)->sacked;
2758
2759 if (skb == tcp_send_head(sk))
2760 break;
2761 /* we could do better than to assign each time */
2762 if (!hole)
2763 tp->retransmit_skb_hint = skb;
2764
2765 /* Assume this retransmit will generate
2766 * only one packet for congestion window
2767 * calculation purposes. This works because
2768 * tcp_retransmit_skb() will chop up the
2769 * packet to be MSS sized and all the
2770 * packet counting works out.
2771 */
2772 if (tcp_packets_in_flight(tp) >= tp->snd_cwnd)
2773 return;
2774
2775 if (fwd_rexmitting) {
2776begin_fwd:
2777 if (!before(TCP_SKB_CB(skb)->seq, tcp_highest_sack_seq(tp)))
2778 break;
2779 mib_idx = LINUX_MIB_TCPFORWARDRETRANS;
2780
2781 } else if (!before(TCP_SKB_CB(skb)->seq, tp->retransmit_high)) {
2782 tp->retransmit_high = last_lost;
2783 if (!tcp_can_forward_retransmit(sk))
2784 break;
2785 /* Backtrack if necessary to non-L'ed skb */
2786 if (hole) {
2787 skb = hole;
2788 hole = NULL;
2789 }
2790 fwd_rexmitting = 1;
2791 goto begin_fwd;
2792
2793 } else if (!(sacked & TCPCB_LOST)) {
2794 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2795 hole = skb;
2796 continue;
2797
2798 } else {
2799 last_lost = TCP_SKB_CB(skb)->end_seq;
2800 if (icsk->icsk_ca_state != TCP_CA_Loss)
2801 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2802 else
2803 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2804 }
2805
2806 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2807 continue;
2808
2809 if (tcp_retransmit_skb(sk, skb))
2810 return;
2811
2812 NET_INC_STATS_BH(sock_net(sk), mib_idx);
2813
2814 if (tcp_in_cwnd_reduction(sk))
2815 tp->prr_out += tcp_skb_pcount(skb);
2816
2817 if (skb == tcp_write_queue_head(sk))
2818 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
2819 inet_csk(sk)->icsk_rto,
2820 TCP_RTO_MAX);
2821 }
2822}
2823
2824/* We allow to exceed memory limits for FIN packets to expedite
2825 * connection tear down and (memory) recovery.
2826 * Otherwise tcp_send_fin() could be tempted to either delay FIN
2827 * or even be forced to close flow without any FIN.
2828 * In general, we want to allow one skb per socket to avoid hangs
2829 * with edge trigger epoll()
2830 */
2831void sk_forced_mem_schedule(struct sock *sk, int size)
2832{
2833 int amt;
2834
2835 if (size <= sk->sk_forward_alloc)
2836 return;
2837 amt = sk_mem_pages(size);
2838 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
2839 sk_memory_allocated_add(sk, amt);
2840
2841 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
2842 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
2843}
2844
2845/* Send a FIN. The caller locks the socket for us.
2846 * We should try to send a FIN packet really hard, but eventually give up.
2847 */
2848void tcp_send_fin(struct sock *sk)
2849{
2850 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
2851 struct tcp_sock *tp = tcp_sk(sk);
2852
2853 /* Optimization, tack on the FIN if we have one skb in write queue and
2854 * this skb was not yet sent, or we are under memory pressure.
2855 * Note: in the latter case, FIN packet will be sent after a timeout,
2856 * as TCP stack thinks it has already been transmitted.
2857 */
2858 if (tskb && (tcp_send_head(sk) || tcp_under_memory_pressure(sk))) {
2859coalesce:
2860 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
2861 TCP_SKB_CB(tskb)->end_seq++;
2862 tp->write_seq++;
2863 if (!tcp_send_head(sk)) {
2864 /* This means tskb was already sent.
2865 * Pretend we included the FIN on previous transmit.
2866 * We need to set tp->snd_nxt to the value it would have
2867 * if FIN had been sent. This is because retransmit path
2868 * does not change tp->snd_nxt.
2869 */
2870 tp->snd_nxt++;
2871 return;
2872 }
2873 } else {
2874 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
2875 if (unlikely(!skb)) {
2876 if (tskb)
2877 goto coalesce;
2878 return;
2879 }
2880 skb_reserve(skb, MAX_TCP_HEADER);
2881 sk_forced_mem_schedule(sk, skb->truesize);
2882 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
2883 tcp_init_nondata_skb(skb, tp->write_seq,
2884 TCPHDR_ACK | TCPHDR_FIN);
2885 tcp_queue_skb(sk, skb);
2886 }
2887 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
2888}
2889
2890/* We get here when a process closes a file descriptor (either due to
2891 * an explicit close() or as a byproduct of exit()'ing) and there
2892 * was unread data in the receive queue. This behavior is recommended
2893 * by RFC 2525, section 2.17. -DaveM
2894 */
2895void tcp_send_active_reset(struct sock *sk, gfp_t priority)
2896{
2897 struct sk_buff *skb;
2898
2899 /* NOTE: No TCP options attached and we never retransmit this. */
2900 skb = alloc_skb(MAX_TCP_HEADER, priority);
2901 if (!skb) {
2902 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2903 return;
2904 }
2905
2906 /* Reserve space for headers and prepare control bits. */
2907 skb_reserve(skb, MAX_TCP_HEADER);
2908 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
2909 TCPHDR_ACK | TCPHDR_RST);
2910 skb_mstamp_get(&skb->skb_mstamp);
2911 /* Send it off. */
2912 if (tcp_transmit_skb(sk, skb, 0, priority))
2913 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
2914
2915 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
2916}
2917
2918/* Send a crossed SYN-ACK during socket establishment.
2919 * WARNING: This routine must only be called when we have already sent
2920 * a SYN packet that crossed the incoming SYN that caused this routine
2921 * to get called. If this assumption fails then the initial rcv_wnd
2922 * and rcv_wscale values will not be correct.
2923 */
2924int tcp_send_synack(struct sock *sk)
2925{
2926 struct sk_buff *skb;
2927
2928 skb = tcp_write_queue_head(sk);
2929 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
2930 pr_debug("%s: wrong queue state\n", __func__);
2931 return -EFAULT;
2932 }
2933 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
2934 if (skb_cloned(skb)) {
2935 struct sk_buff *nskb = skb_copy(skb, GFP_ATOMIC);
2936 if (!nskb)
2937 return -ENOMEM;
2938 tcp_unlink_write_queue(skb, sk);
2939 __skb_header_release(nskb);
2940 __tcp_add_write_queue_head(sk, nskb);
2941 sk_wmem_free_skb(sk, skb);
2942 sk->sk_wmem_queued += nskb->truesize;
2943 sk_mem_charge(sk, nskb->truesize);
2944 skb = nskb;
2945 }
2946
2947 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
2948 tcp_ecn_send_synack(sk, skb);
2949 }
2950 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2951}
2952
2953/**
2954 * tcp_make_synack - Prepare a SYN-ACK.
2955 * sk: listener socket
2956 * dst: dst entry attached to the SYNACK
2957 * req: request_sock pointer
2958 *
2959 * Allocate one skb and build a SYNACK packet.
2960 * @dst is consumed : Caller should not use it again.
2961 */
2962struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
2963 struct request_sock *req,
2964 struct tcp_fastopen_cookie *foc,
2965 bool attach_req)
2966{
2967 struct inet_request_sock *ireq = inet_rsk(req);
2968 const struct tcp_sock *tp = tcp_sk(sk);
2969 struct tcp_md5sig_key *md5 = NULL;
2970 struct tcp_out_options opts;
2971 struct sk_buff *skb;
2972 int tcp_header_size;
2973 struct tcphdr *th;
2974 u16 user_mss;
2975 int mss;
2976
2977 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
2978 if (unlikely(!skb)) {
2979 dst_release(dst);
2980 return NULL;
2981 }
2982 /* Reserve space for headers. */
2983 skb_reserve(skb, MAX_TCP_HEADER);
2984
2985 if (attach_req) {
2986 skb_set_owner_w(skb, req_to_sk(req));
2987 } else {
2988 /* sk is a const pointer, because we want to express multiple
2989 * cpu might call us concurrently.
2990 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
2991 */
2992 skb_set_owner_w(skb, (struct sock *)sk);
2993 }
2994 skb_dst_set(skb, dst);
2995
2996 mss = dst_metric_advmss(dst);
2997 user_mss = READ_ONCE(tp->rx_opt.user_mss);
2998 if (user_mss && user_mss < mss)
2999 mss = user_mss;
3000
3001 memset(&opts, 0, sizeof(opts));
3002#ifdef CONFIG_SYN_COOKIES
3003 if (unlikely(req->cookie_ts))
3004 skb->skb_mstamp.stamp_jiffies = cookie_init_timestamp(req);
3005 else
3006#endif
3007 skb_mstamp_get(&skb->skb_mstamp);
3008
3009#ifdef CONFIG_TCP_MD5SIG
3010 rcu_read_lock();
3011 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3012#endif
3013 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3014 tcp_header_size = tcp_synack_options(req, mss, skb, &opts, md5, foc) +
3015 sizeof(*th);
3016
3017 skb_push(skb, tcp_header_size);
3018 skb_reset_transport_header(skb);
3019
3020 th = tcp_hdr(skb);
3021 memset(th, 0, sizeof(struct tcphdr));
3022 th->syn = 1;
3023 th->ack = 1;
3024 tcp_ecn_make_synack(req, th);
3025 th->source = htons(ireq->ir_num);
3026 th->dest = ireq->ir_rmt_port;
3027 /* Setting of flags are superfluous here for callers (and ECE is
3028 * not even correctly set)
3029 */
3030 tcp_init_nondata_skb(skb, tcp_rsk(req)->snt_isn,
3031 TCPHDR_SYN | TCPHDR_ACK);
3032
3033 th->seq = htonl(TCP_SKB_CB(skb)->seq);
3034 /* XXX data is queued and acked as is. No buffer/window check */
3035 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3036
3037 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3038 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3039 tcp_options_write((__be32 *)(th + 1), NULL, &opts);
3040 th->doff = (tcp_header_size >> 2);
3041 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_OUTSEGS);
3042
3043#ifdef CONFIG_TCP_MD5SIG
3044 /* Okay, we have all we need - do the md5 hash if needed */
3045 if (md5)
3046 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3047 md5, req_to_sk(req), skb);
3048 rcu_read_unlock();
3049#endif
3050
3051 /* Do not fool tcpdump (if any), clean our debris */
3052 skb->tstamp.tv64 = 0;
3053 return skb;
3054}
3055EXPORT_SYMBOL(tcp_make_synack);
3056
3057static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3058{
3059 struct inet_connection_sock *icsk = inet_csk(sk);
3060 const struct tcp_congestion_ops *ca;
3061 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3062
3063 if (ca_key == TCP_CA_UNSPEC)
3064 return;
3065
3066 rcu_read_lock();
3067 ca = tcp_ca_find_key(ca_key);
3068 if (likely(ca && try_module_get(ca->owner))) {
3069 module_put(icsk->icsk_ca_ops->owner);
3070 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3071 icsk->icsk_ca_ops = ca;
3072 }
3073 rcu_read_unlock();
3074}
3075
3076/* Do all connect socket setups that can be done AF independent. */
3077static void tcp_connect_init(struct sock *sk)
3078{
3079 const struct dst_entry *dst = __sk_dst_get(sk);
3080 struct tcp_sock *tp = tcp_sk(sk);
3081 __u8 rcv_wscale;
3082
3083 /* We'll fix this up when we get a response from the other end.
3084 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3085 */
3086 tp->tcp_header_len = sizeof(struct tcphdr) +
3087 (sysctl_tcp_timestamps ? TCPOLEN_TSTAMP_ALIGNED : 0);
3088
3089#ifdef CONFIG_TCP_MD5SIG
3090 if (tp->af_specific->md5_lookup(sk, sk))
3091 tp->tcp_header_len += TCPOLEN_MD5SIG_ALIGNED;
3092#endif
3093
3094 /* If user gave his TCP_MAXSEG, record it to clamp */
3095 if (tp->rx_opt.user_mss)
3096 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3097 tp->max_window = 0;
3098 tcp_mtup_init(sk);
3099 tcp_sync_mss(sk, dst_mtu(dst));
3100
3101 tcp_ca_dst_init(sk, dst);
3102
3103 if (!tp->window_clamp)
3104 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3105 tp->advmss = dst_metric_advmss(dst);
3106 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->advmss)
3107 tp->advmss = tp->rx_opt.user_mss;
3108
3109 tcp_initialize_rcv_mss(sk);
3110
3111 /* limit the window selection if the user enforce a smaller rx buffer */
3112 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3113 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3114 tp->window_clamp = tcp_full_space(sk);
3115
3116 tcp_select_initial_window(tcp_full_space(sk),
3117 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3118 &tp->rcv_wnd,
3119 &tp->window_clamp,
3120 sysctl_tcp_window_scaling,
3121 &rcv_wscale,
3122 dst_metric(dst, RTAX_INITRWND));
3123
3124 tp->rx_opt.rcv_wscale = rcv_wscale;
3125 tp->rcv_ssthresh = tp->rcv_wnd;
3126
3127 sk->sk_err = 0;
3128 sock_reset_flag(sk, SOCK_DONE);
3129 tp->snd_wnd = 0;
3130 tcp_init_wl(tp, 0);
3131 tp->snd_una = tp->write_seq;
3132 tp->snd_sml = tp->write_seq;
3133 tp->snd_up = tp->write_seq;
3134 tp->snd_nxt = tp->write_seq;
3135
3136 if (likely(!tp->repair))
3137 tp->rcv_nxt = 0;
3138 else
3139 tp->rcv_tstamp = tcp_time_stamp;
3140 tp->rcv_wup = tp->rcv_nxt;
3141 tp->copied_seq = tp->rcv_nxt;
3142
3143 inet_csk(sk)->icsk_rto = TCP_TIMEOUT_INIT;
3144 inet_csk(sk)->icsk_retransmits = 0;
3145 tcp_clear_retrans(tp);
3146}
3147
3148static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3149{
3150 struct tcp_sock *tp = tcp_sk(sk);
3151 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3152
3153 tcb->end_seq += skb->len;
3154 __skb_header_release(skb);
3155 __tcp_add_write_queue_tail(sk, skb);
3156 sk->sk_wmem_queued += skb->truesize;
3157 sk_mem_charge(sk, skb->truesize);
3158 tp->write_seq = tcb->end_seq;
3159 tp->packets_out += tcp_skb_pcount(skb);
3160}
3161
3162/* Build and send a SYN with data and (cached) Fast Open cookie. However,
3163 * queue a data-only packet after the regular SYN, such that regular SYNs
3164 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3165 * only the SYN sequence, the data are retransmitted in the first ACK.
3166 * If cookie is not cached or other error occurs, falls back to send a
3167 * regular SYN with Fast Open cookie request option.
3168 */
3169static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3170{
3171 struct tcp_sock *tp = tcp_sk(sk);
3172 struct tcp_fastopen_request *fo = tp->fastopen_req;
3173 int syn_loss = 0, space, err = 0;
3174 unsigned long last_syn_loss = 0;
3175 struct sk_buff *syn_data;
3176
3177 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3178 tcp_fastopen_cache_get(sk, &tp->rx_opt.mss_clamp, &fo->cookie,
3179 &syn_loss, &last_syn_loss);
3180 /* Recurring FO SYN losses: revert to regular handshake temporarily */
3181 if (syn_loss > 1 &&
3182 time_before(jiffies, last_syn_loss + (60*HZ << syn_loss))) {
3183 fo->cookie.len = -1;
3184 goto fallback;
3185 }
3186
3187 if (sysctl_tcp_fastopen & TFO_CLIENT_NO_COOKIE)
3188 fo->cookie.len = -1;
3189 else if (fo->cookie.len <= 0)
3190 goto fallback;
3191
3192 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3193 * user-MSS. Reserve maximum option space for middleboxes that add
3194 * private TCP options. The cost is reduced data space in SYN :(
3195 */
3196 if (tp->rx_opt.user_mss && tp->rx_opt.user_mss < tp->rx_opt.mss_clamp)
3197 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3198 space = __tcp_mtu_to_mss(sk, inet_csk(sk)->icsk_pmtu_cookie) -
3199 MAX_TCP_OPTION_SPACE;
3200
3201 space = min_t(size_t, space, fo->size);
3202
3203 /* limit to order-0 allocations */
3204 space = min_t(size_t, space, SKB_MAX_HEAD(MAX_TCP_HEADER));
3205
3206 syn_data = sk_stream_alloc_skb(sk, space, sk->sk_allocation, false);
3207 if (!syn_data)
3208 goto fallback;
3209 syn_data->ip_summed = CHECKSUM_PARTIAL;
3210 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3211 if (space) {
3212 int copied = copy_from_iter(skb_put(syn_data, space), space,
3213 &fo->data->msg_iter);
3214 if (unlikely(!copied)) {
3215 kfree_skb(syn_data);
3216 goto fallback;
3217 }
3218 if (copied != space) {
3219 skb_trim(syn_data, copied);
3220 space = copied;
3221 }
3222 }
3223 /* No more data pending in inet_wait_for_connect() */
3224 if (space == fo->size)
3225 fo->data = NULL;
3226 fo->copied = space;
3227
3228 tcp_connect_queue_skb(sk, syn_data);
3229
3230 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3231
3232 syn->skb_mstamp = syn_data->skb_mstamp;
3233
3234 /* Now full SYN+DATA was cloned and sent (or not),
3235 * remove the SYN from the original skb (syn_data)
3236 * we keep in write queue in case of a retransmit, as we
3237 * also have the SYN packet (with no data) in the same queue.
3238 */
3239 TCP_SKB_CB(syn_data)->seq++;
3240 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
3241 if (!err) {
3242 tp->syn_data = (fo->copied > 0);
3243 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
3244 goto done;
3245 }
3246
3247fallback:
3248 /* Send a regular SYN with Fast Open cookie request option */
3249 if (fo->cookie.len > 0)
3250 fo->cookie.len = 0;
3251 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
3252 if (err)
3253 tp->syn_fastopen = 0;
3254done:
3255 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
3256 return err;
3257}
3258
3259/* Build a SYN and send it off. */
3260int tcp_connect(struct sock *sk)
3261{
3262 struct tcp_sock *tp = tcp_sk(sk);
3263 struct sk_buff *buff;
3264 int err;
3265
3266 tcp_connect_init(sk);
3267
3268 if (unlikely(tp->repair)) {
3269 tcp_finish_connect(sk, NULL);
3270 return 0;
3271 }
3272
3273 buff = sk_stream_alloc_skb(sk, 0, sk->sk_allocation, true);
3274 if (unlikely(!buff))
3275 return -ENOBUFS;
3276
3277 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
3278 tp->retrans_stamp = tcp_time_stamp;
3279 tcp_connect_queue_skb(sk, buff);
3280 tcp_ecn_send_syn(sk, buff);
3281
3282 /* Send off SYN; include data in Fast Open. */
3283 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
3284 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
3285 if (err == -ECONNREFUSED)
3286 return err;
3287
3288 /* We change tp->snd_nxt after the tcp_transmit_skb() call
3289 * in order to make this packet get counted in tcpOutSegs.
3290 */
3291 tp->snd_nxt = tp->write_seq;
3292 tp->pushed_seq = tp->write_seq;
3293 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
3294
3295 /* Timer for repeating the SYN until an answer. */
3296 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3297 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
3298 return 0;
3299}
3300EXPORT_SYMBOL(tcp_connect);
3301
3302/* Send out a delayed ack, the caller does the policy checking
3303 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
3304 * for details.
3305 */
3306void tcp_send_delayed_ack(struct sock *sk)
3307{
3308 struct inet_connection_sock *icsk = inet_csk(sk);
3309 int ato = icsk->icsk_ack.ato;
3310 unsigned long timeout;
3311
3312 tcp_ca_event(sk, CA_EVENT_DELAYED_ACK);
3313
3314 if (ato > TCP_DELACK_MIN) {
3315 const struct tcp_sock *tp = tcp_sk(sk);
3316 int max_ato = HZ / 2;
3317
3318 if (icsk->icsk_ack.pingpong ||
3319 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
3320 max_ato = TCP_DELACK_MAX;
3321
3322 /* Slow path, intersegment interval is "high". */
3323
3324 /* If some rtt estimate is known, use it to bound delayed ack.
3325 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
3326 * directly.
3327 */
3328 if (tp->srtt_us) {
3329 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
3330 TCP_DELACK_MIN);
3331
3332 if (rtt < max_ato)
3333 max_ato = rtt;
3334 }
3335
3336 ato = min(ato, max_ato);
3337 }
3338
3339 /* Stay within the limit we were given */
3340 timeout = jiffies + ato;
3341
3342 /* Use new timeout only if there wasn't a older one earlier. */
3343 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
3344 /* If delack timer was blocked or is about to expire,
3345 * send ACK now.
3346 */
3347 if (icsk->icsk_ack.blocked ||
3348 time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
3349 tcp_send_ack(sk);
3350 return;
3351 }
3352
3353 if (!time_before(timeout, icsk->icsk_ack.timeout))
3354 timeout = icsk->icsk_ack.timeout;
3355 }
3356 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
3357 icsk->icsk_ack.timeout = timeout;
3358 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
3359}
3360
3361/* This routine sends an ack and also updates the window. */
3362void tcp_send_ack(struct sock *sk)
3363{
3364 struct sk_buff *buff;
3365
3366 /* If we have been reset, we may not send again. */
3367 if (sk->sk_state == TCP_CLOSE)
3368 return;
3369
3370 tcp_ca_event(sk, CA_EVENT_NON_DELAYED_ACK);
3371
3372 /* We are not putting this on the write queue, so
3373 * tcp_transmit_skb() will set the ownership to this
3374 * sock.
3375 */
3376 buff = alloc_skb(MAX_TCP_HEADER,
3377 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3378 if (unlikely(!buff)) {
3379 inet_csk_schedule_ack(sk);
3380 inet_csk(sk)->icsk_ack.ato = TCP_ATO_MIN;
3381 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK,
3382 TCP_DELACK_MAX, TCP_RTO_MAX);
3383 return;
3384 }
3385
3386 /* Reserve space for headers and prepare control bits. */
3387 skb_reserve(buff, MAX_TCP_HEADER);
3388 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
3389
3390 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
3391 * too much.
3392 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
3393 * We also avoid tcp_wfree() overhead (cache line miss accessing
3394 * tp->tsq_flags) by using regular sock_wfree()
3395 */
3396 skb_set_tcp_pure_ack(buff);
3397
3398 /* Send it off, this clears delayed acks for us. */
3399 skb_mstamp_get(&buff->skb_mstamp);
3400 tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0);
3401}
3402EXPORT_SYMBOL_GPL(tcp_send_ack);
3403
3404/* This routine sends a packet with an out of date sequence
3405 * number. It assumes the other end will try to ack it.
3406 *
3407 * Question: what should we make while urgent mode?
3408 * 4.4BSD forces sending single byte of data. We cannot send
3409 * out of window data, because we have SND.NXT==SND.MAX...
3410 *
3411 * Current solution: to send TWO zero-length segments in urgent mode:
3412 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
3413 * out-of-date with SND.UNA-1 to probe window.
3414 */
3415static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
3416{
3417 struct tcp_sock *tp = tcp_sk(sk);
3418 struct sk_buff *skb;
3419
3420 /* We don't queue it, tcp_transmit_skb() sets ownership. */
3421 skb = alloc_skb(MAX_TCP_HEADER,
3422 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
3423 if (!skb)
3424 return -1;
3425
3426 /* Reserve space for headers and set control bits. */
3427 skb_reserve(skb, MAX_TCP_HEADER);
3428 /* Use a previous sequence. This should cause the other
3429 * end to send an ack. Don't queue or clone SKB, just
3430 * send it.
3431 */
3432 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
3433 skb_mstamp_get(&skb->skb_mstamp);
3434 NET_INC_STATS(sock_net(sk), mib);
3435 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
3436}
3437
3438void tcp_send_window_probe(struct sock *sk)
3439{
3440 if (sk->sk_state == TCP_ESTABLISHED) {
3441 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
3442 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
3443 }
3444}
3445
3446/* Initiate keepalive or window probe from timer. */
3447int tcp_write_wakeup(struct sock *sk, int mib)
3448{
3449 struct tcp_sock *tp = tcp_sk(sk);
3450 struct sk_buff *skb;
3451
3452 if (sk->sk_state == TCP_CLOSE)
3453 return -1;
3454
3455 skb = tcp_send_head(sk);
3456 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
3457 int err;
3458 unsigned int mss = tcp_current_mss(sk);
3459 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3460
3461 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
3462 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
3463
3464 /* We are probing the opening of a window
3465 * but the window size is != 0
3466 * must have been a result SWS avoidance ( sender )
3467 */
3468 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
3469 skb->len > mss) {
3470 seg_size = min(seg_size, mss);
3471 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3472 if (tcp_fragment(sk, skb, seg_size, mss, GFP_ATOMIC))
3473 return -1;
3474 } else if (!tcp_skb_pcount(skb))
3475 tcp_set_skb_tso_segs(skb, mss);
3476
3477 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
3478 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3479 if (!err)
3480 tcp_event_new_data_sent(sk, skb);
3481 return err;
3482 } else {
3483 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
3484 tcp_xmit_probe_skb(sk, 1, mib);
3485 return tcp_xmit_probe_skb(sk, 0, mib);
3486 }
3487}
3488
3489/* A window probe timeout has occurred. If window is not closed send
3490 * a partial packet else a zero probe.
3491 */
3492void tcp_send_probe0(struct sock *sk)
3493{
3494 struct inet_connection_sock *icsk = inet_csk(sk);
3495 struct tcp_sock *tp = tcp_sk(sk);
3496 struct net *net = sock_net(sk);
3497 unsigned long probe_max;
3498 int err;
3499
3500 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
3501
3502 if (tp->packets_out || !tcp_send_head(sk)) {
3503 /* Cancel probe timer, if it is not required. */
3504 icsk->icsk_probes_out = 0;
3505 icsk->icsk_backoff = 0;
3506 return;
3507 }
3508
3509 if (err <= 0) {
3510 if (icsk->icsk_backoff < net->ipv4.sysctl_tcp_retries2)
3511 icsk->icsk_backoff++;
3512 icsk->icsk_probes_out++;
3513 probe_max = TCP_RTO_MAX;
3514 } else {
3515 /* If packet was not sent due to local congestion,
3516 * do not backoff and do not remember icsk_probes_out.
3517 * Let local senders to fight for local resources.
3518 *
3519 * Use accumulated backoff yet.
3520 */
3521 if (!icsk->icsk_probes_out)
3522 icsk->icsk_probes_out = 1;
3523 probe_max = TCP_RESOURCE_PROBE_INTERVAL;
3524 }
3525 inet_csk_reset_xmit_timer(sk, ICSK_TIME_PROBE0,
3526 tcp_probe0_when(sk, probe_max),
3527 TCP_RTO_MAX);
3528}
3529
3530int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
3531{
3532 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
3533 struct flowi fl;
3534 int res;
3535
3536 tcp_rsk(req)->txhash = net_tx_rndhash();
3537 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, true);
3538 if (!res) {
3539 TCP_INC_STATS_BH(sock_net(sk), TCP_MIB_RETRANSSEGS);
3540 NET_INC_STATS_BH(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3541 }
3542 return res;
3543}
3544EXPORT_SYMBOL(tcp_rtx_synack);
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
9 * Authors: Ross Biro
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
20 */
21
22/*
23 * Changes: Pedro Roque : Retransmit queue handled by TCP.
24 * : Fragmentation on mtu decrease
25 * : Segment collapse on retransmit
26 * : AF independence
27 *
28 * Linus Torvalds : send_delayed_ack
29 * David S. Miller : Charge memory using the right skb
30 * during syn/ack processing.
31 * David S. Miller : Output engine completely rewritten.
32 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
33 * Cacophonix Gaul : draft-minshall-nagle-01
34 * J Hadi Salim : ECN support
35 *
36 */
37
38#define pr_fmt(fmt) "TCP: " fmt
39
40#include <net/tcp.h>
41#include <net/mptcp.h>
42
43#include <linux/compiler.h>
44#include <linux/gfp.h>
45#include <linux/module.h>
46#include <linux/static_key.h>
47
48#include <trace/events/tcp.h>
49
50/* Refresh clocks of a TCP socket,
51 * ensuring monotically increasing values.
52 */
53void tcp_mstamp_refresh(struct tcp_sock *tp)
54{
55 u64 val = tcp_clock_ns();
56
57 tp->tcp_clock_cache = val;
58 tp->tcp_mstamp = div_u64(val, NSEC_PER_USEC);
59}
60
61static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
62 int push_one, gfp_t gfp);
63
64/* Account for new data that has been sent to the network. */
65static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
66{
67 struct inet_connection_sock *icsk = inet_csk(sk);
68 struct tcp_sock *tp = tcp_sk(sk);
69 unsigned int prior_packets = tp->packets_out;
70
71 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq);
72
73 __skb_unlink(skb, &sk->sk_write_queue);
74 tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
75
76 if (tp->highest_sack == NULL)
77 tp->highest_sack = skb;
78
79 tp->packets_out += tcp_skb_pcount(skb);
80 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
81 tcp_rearm_rto(sk);
82
83 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
84 tcp_skb_pcount(skb));
85 tcp_check_space(sk);
86}
87
88/* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
89 * window scaling factor due to loss of precision.
90 * If window has been shrunk, what should we make? It is not clear at all.
91 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
92 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
93 * invalid. OK, let's make this for now:
94 */
95static inline __u32 tcp_acceptable_seq(const struct sock *sk)
96{
97 const struct tcp_sock *tp = tcp_sk(sk);
98
99 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
100 (tp->rx_opt.wscale_ok &&
101 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
102 return tp->snd_nxt;
103 else
104 return tcp_wnd_end(tp);
105}
106
107/* Calculate mss to advertise in SYN segment.
108 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
109 *
110 * 1. It is independent of path mtu.
111 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
112 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
113 * attached devices, because some buggy hosts are confused by
114 * large MSS.
115 * 4. We do not make 3, we advertise MSS, calculated from first
116 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
117 * This may be overridden via information stored in routing table.
118 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
119 * probably even Jumbo".
120 */
121static __u16 tcp_advertise_mss(struct sock *sk)
122{
123 struct tcp_sock *tp = tcp_sk(sk);
124 const struct dst_entry *dst = __sk_dst_get(sk);
125 int mss = tp->advmss;
126
127 if (dst) {
128 unsigned int metric = dst_metric_advmss(dst);
129
130 if (metric < mss) {
131 mss = metric;
132 tp->advmss = mss;
133 }
134 }
135
136 return (__u16)mss;
137}
138
139/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
140 * This is the first part of cwnd validation mechanism.
141 */
142void tcp_cwnd_restart(struct sock *sk, s32 delta)
143{
144 struct tcp_sock *tp = tcp_sk(sk);
145 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
146 u32 cwnd = tcp_snd_cwnd(tp);
147
148 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
149
150 tp->snd_ssthresh = tcp_current_ssthresh(sk);
151 restart_cwnd = min(restart_cwnd, cwnd);
152
153 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
154 cwnd >>= 1;
155 tcp_snd_cwnd_set(tp, max(cwnd, restart_cwnd));
156 tp->snd_cwnd_stamp = tcp_jiffies32;
157 tp->snd_cwnd_used = 0;
158}
159
160/* Congestion state accounting after a packet has been sent. */
161static void tcp_event_data_sent(struct tcp_sock *tp,
162 struct sock *sk)
163{
164 struct inet_connection_sock *icsk = inet_csk(sk);
165 const u32 now = tcp_jiffies32;
166
167 if (tcp_packets_in_flight(tp) == 0)
168 tcp_ca_event(sk, CA_EVENT_TX_START);
169
170 tp->lsndtime = now;
171
172 /* If it is a reply for ato after last received
173 * packet, increase pingpong count.
174 */
175 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
176 inet_csk_inc_pingpong_cnt(sk);
177}
178
179/* Account for an ACK we sent. */
180static inline void tcp_event_ack_sent(struct sock *sk, u32 rcv_nxt)
181{
182 struct tcp_sock *tp = tcp_sk(sk);
183
184 if (unlikely(tp->compressed_ack)) {
185 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
186 tp->compressed_ack);
187 tp->compressed_ack = 0;
188 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
189 __sock_put(sk);
190 }
191
192 if (unlikely(rcv_nxt != tp->rcv_nxt))
193 return; /* Special ACK sent by DCTCP to reflect ECN */
194 tcp_dec_quickack_mode(sk);
195 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
196}
197
198/* Determine a window scaling and initial window to offer.
199 * Based on the assumption that the given amount of space
200 * will be offered. Store the results in the tp structure.
201 * NOTE: for smooth operation initial space offering should
202 * be a multiple of mss if possible. We assume here that mss >= 1.
203 * This MUST be enforced by all callers.
204 */
205void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
206 __u32 *rcv_wnd, __u32 *window_clamp,
207 int wscale_ok, __u8 *rcv_wscale,
208 __u32 init_rcv_wnd)
209{
210 unsigned int space = (__space < 0 ? 0 : __space);
211
212 /* If no clamp set the clamp to the max possible scaled window */
213 if (*window_clamp == 0)
214 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
215 space = min(*window_clamp, space);
216
217 /* Quantize space offering to a multiple of mss if possible. */
218 if (space > mss)
219 space = rounddown(space, mss);
220
221 /* NOTE: offering an initial window larger than 32767
222 * will break some buggy TCP stacks. If the admin tells us
223 * it is likely we could be speaking with such a buggy stack
224 * we will truncate our initial window offering to 32K-1
225 * unless the remote has sent us a window scaling option,
226 * which we interpret as a sign the remote TCP is not
227 * misinterpreting the window field as a signed quantity.
228 */
229 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows))
230 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
231 else
232 (*rcv_wnd) = min_t(u32, space, U16_MAX);
233
234 if (init_rcv_wnd)
235 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
236
237 *rcv_wscale = 0;
238 if (wscale_ok) {
239 /* Set window scaling on max possible window */
240 space = max_t(u32, space, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
241 space = max_t(u32, space, READ_ONCE(sysctl_rmem_max));
242 space = min_t(u32, space, *window_clamp);
243 *rcv_wscale = clamp_t(int, ilog2(space) - 15,
244 0, TCP_MAX_WSCALE);
245 }
246 /* Set the clamp no higher than max representable value */
247 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
248}
249EXPORT_SYMBOL(tcp_select_initial_window);
250
251/* Chose a new window to advertise, update state in tcp_sock for the
252 * socket, and return result with RFC1323 scaling applied. The return
253 * value can be stuffed directly into th->window for an outgoing
254 * frame.
255 */
256static u16 tcp_select_window(struct sock *sk)
257{
258 struct tcp_sock *tp = tcp_sk(sk);
259 struct net *net = sock_net(sk);
260 u32 old_win = tp->rcv_wnd;
261 u32 cur_win, new_win;
262
263 /* Make the window 0 if we failed to queue the data because we
264 * are out of memory. The window is temporary, so we don't store
265 * it on the socket.
266 */
267 if (unlikely(inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOMEM))
268 return 0;
269
270 cur_win = tcp_receive_window(tp);
271 new_win = __tcp_select_window(sk);
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 if (!READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) || !tp->rx_opt.rcv_wscale) {
281 /* Never shrink the offered window */
282 if (new_win == 0)
283 NET_INC_STATS(net, LINUX_MIB_TCPWANTZEROWINDOWADV);
284 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
285 }
286 }
287
288 tp->rcv_wnd = new_win;
289 tp->rcv_wup = tp->rcv_nxt;
290
291 /* Make sure we do not exceed the maximum possible
292 * scaled window.
293 */
294 if (!tp->rx_opt.rcv_wscale &&
295 READ_ONCE(net->ipv4.sysctl_tcp_workaround_signed_windows))
296 new_win = min(new_win, MAX_TCP_WINDOW);
297 else
298 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
299
300 /* RFC1323 scaling applied */
301 new_win >>= tp->rx_opt.rcv_wscale;
302
303 /* If we advertise zero window, disable fast path. */
304 if (new_win == 0) {
305 tp->pred_flags = 0;
306 if (old_win)
307 NET_INC_STATS(net, LINUX_MIB_TCPTOZEROWINDOWADV);
308 } else if (old_win == 0) {
309 NET_INC_STATS(net, LINUX_MIB_TCPFROMZEROWINDOWADV);
310 }
311
312 return new_win;
313}
314
315/* Packet ECN state for a SYN-ACK */
316static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
317{
318 const struct tcp_sock *tp = tcp_sk(sk);
319
320 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
321 if (!(tp->ecn_flags & TCP_ECN_OK))
322 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
323 else if (tcp_ca_needs_ecn(sk) ||
324 tcp_bpf_ca_needs_ecn(sk))
325 INET_ECN_xmit(sk);
326}
327
328/* Packet ECN state for a SYN. */
329static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
330{
331 struct tcp_sock *tp = tcp_sk(sk);
332 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
333 bool use_ecn = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn) == 1 ||
334 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
335
336 if (!use_ecn) {
337 const struct dst_entry *dst = __sk_dst_get(sk);
338
339 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
340 use_ecn = true;
341 }
342
343 tp->ecn_flags = 0;
344
345 if (use_ecn) {
346 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
347 tp->ecn_flags = TCP_ECN_OK;
348 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
349 INET_ECN_xmit(sk);
350 }
351}
352
353static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
354{
355 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback))
356 /* tp->ecn_flags are cleared at a later point in time when
357 * SYN ACK is ultimatively being received.
358 */
359 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
360}
361
362static void
363tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
364{
365 if (inet_rsk(req)->ecn_ok)
366 th->ece = 1;
367}
368
369/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
370 * be sent.
371 */
372static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
373 struct tcphdr *th, int tcp_header_len)
374{
375 struct tcp_sock *tp = tcp_sk(sk);
376
377 if (tp->ecn_flags & TCP_ECN_OK) {
378 /* Not-retransmitted data segment: set ECT and inject CWR. */
379 if (skb->len != tcp_header_len &&
380 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
381 INET_ECN_xmit(sk);
382 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
383 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
384 th->cwr = 1;
385 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
386 }
387 } else if (!tcp_ca_needs_ecn(sk)) {
388 /* ACK or retransmitted segment: clear ECT|CE */
389 INET_ECN_dontxmit(sk);
390 }
391 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
392 th->ece = 1;
393 }
394}
395
396/* Constructs common control bits of non-data skb. If SYN/FIN is present,
397 * auto increment end seqno.
398 */
399static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
400{
401 skb->ip_summed = CHECKSUM_PARTIAL;
402
403 TCP_SKB_CB(skb)->tcp_flags = flags;
404
405 tcp_skb_pcount_set(skb, 1);
406
407 TCP_SKB_CB(skb)->seq = seq;
408 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
409 seq++;
410 TCP_SKB_CB(skb)->end_seq = seq;
411}
412
413static inline bool tcp_urg_mode(const struct tcp_sock *tp)
414{
415 return tp->snd_una != tp->snd_up;
416}
417
418#define OPTION_SACK_ADVERTISE BIT(0)
419#define OPTION_TS BIT(1)
420#define OPTION_MD5 BIT(2)
421#define OPTION_WSCALE BIT(3)
422#define OPTION_FAST_OPEN_COOKIE BIT(8)
423#define OPTION_SMC BIT(9)
424#define OPTION_MPTCP BIT(10)
425#define OPTION_AO BIT(11)
426
427static void smc_options_write(__be32 *ptr, u16 *options)
428{
429#if IS_ENABLED(CONFIG_SMC)
430 if (static_branch_unlikely(&tcp_have_smc)) {
431 if (unlikely(OPTION_SMC & *options)) {
432 *ptr++ = htonl((TCPOPT_NOP << 24) |
433 (TCPOPT_NOP << 16) |
434 (TCPOPT_EXP << 8) |
435 (TCPOLEN_EXP_SMC_BASE));
436 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
437 }
438 }
439#endif
440}
441
442struct tcp_out_options {
443 u16 options; /* bit field of OPTION_* */
444 u16 mss; /* 0 to disable */
445 u8 ws; /* window scale, 0 to disable */
446 u8 num_sack_blocks; /* number of SACK blocks to include */
447 u8 hash_size; /* bytes in hash_location */
448 u8 bpf_opt_len; /* length of BPF hdr option */
449 __u8 *hash_location; /* temporary pointer, overloaded */
450 __u32 tsval, tsecr; /* need to include OPTION_TS */
451 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
452 struct mptcp_out_options mptcp;
453};
454
455static void mptcp_options_write(struct tcphdr *th, __be32 *ptr,
456 struct tcp_sock *tp,
457 struct tcp_out_options *opts)
458{
459#if IS_ENABLED(CONFIG_MPTCP)
460 if (unlikely(OPTION_MPTCP & opts->options))
461 mptcp_write_options(th, ptr, tp, &opts->mptcp);
462#endif
463}
464
465#ifdef CONFIG_CGROUP_BPF
466static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb,
467 enum tcp_synack_type synack_type)
468{
469 if (unlikely(!skb))
470 return BPF_WRITE_HDR_TCP_CURRENT_MSS;
471
472 if (unlikely(synack_type == TCP_SYNACK_COOKIE))
473 return BPF_WRITE_HDR_TCP_SYNACK_COOKIE;
474
475 return 0;
476}
477
478/* req, syn_skb and synack_type are used when writing synack */
479static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
480 struct request_sock *req,
481 struct sk_buff *syn_skb,
482 enum tcp_synack_type synack_type,
483 struct tcp_out_options *opts,
484 unsigned int *remaining)
485{
486 struct bpf_sock_ops_kern sock_ops;
487 int err;
488
489 if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
490 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) ||
491 !*remaining)
492 return;
493
494 /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */
495
496 /* init sock_ops */
497 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
498
499 sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB;
500
501 if (req) {
502 /* The listen "sk" cannot be passed here because
503 * it is not locked. It would not make too much
504 * sense to do bpf_setsockopt(listen_sk) based
505 * on individual connection request also.
506 *
507 * Thus, "req" is passed here and the cgroup-bpf-progs
508 * of the listen "sk" will be run.
509 *
510 * "req" is also used here for fastopen even the "sk" here is
511 * a fullsock "child" sk. It is to keep the behavior
512 * consistent between fastopen and non-fastopen on
513 * the bpf programming side.
514 */
515 sock_ops.sk = (struct sock *)req;
516 sock_ops.syn_skb = syn_skb;
517 } else {
518 sock_owned_by_me(sk);
519
520 sock_ops.is_fullsock = 1;
521 sock_ops.sk = sk;
522 }
523
524 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
525 sock_ops.remaining_opt_len = *remaining;
526 /* tcp_current_mss() does not pass a skb */
527 if (skb)
528 bpf_skops_init_skb(&sock_ops, skb, 0);
529
530 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
531
532 if (err || sock_ops.remaining_opt_len == *remaining)
533 return;
534
535 opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len;
536 /* round up to 4 bytes */
537 opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3;
538
539 *remaining -= opts->bpf_opt_len;
540}
541
542static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
543 struct request_sock *req,
544 struct sk_buff *syn_skb,
545 enum tcp_synack_type synack_type,
546 struct tcp_out_options *opts)
547{
548 u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len;
549 struct bpf_sock_ops_kern sock_ops;
550 int err;
551
552 if (likely(!max_opt_len))
553 return;
554
555 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
556
557 sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB;
558
559 if (req) {
560 sock_ops.sk = (struct sock *)req;
561 sock_ops.syn_skb = syn_skb;
562 } else {
563 sock_owned_by_me(sk);
564
565 sock_ops.is_fullsock = 1;
566 sock_ops.sk = sk;
567 }
568
569 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
570 sock_ops.remaining_opt_len = max_opt_len;
571 first_opt_off = tcp_hdrlen(skb) - max_opt_len;
572 bpf_skops_init_skb(&sock_ops, skb, first_opt_off);
573
574 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
575
576 if (err)
577 nr_written = 0;
578 else
579 nr_written = max_opt_len - sock_ops.remaining_opt_len;
580
581 if (nr_written < max_opt_len)
582 memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP,
583 max_opt_len - nr_written);
584}
585#else
586static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
587 struct request_sock *req,
588 struct sk_buff *syn_skb,
589 enum tcp_synack_type synack_type,
590 struct tcp_out_options *opts,
591 unsigned int *remaining)
592{
593}
594
595static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
596 struct request_sock *req,
597 struct sk_buff *syn_skb,
598 enum tcp_synack_type synack_type,
599 struct tcp_out_options *opts)
600{
601}
602#endif
603
604static __be32 *process_tcp_ao_options(struct tcp_sock *tp,
605 const struct tcp_request_sock *tcprsk,
606 struct tcp_out_options *opts,
607 struct tcp_key *key, __be32 *ptr)
608{
609#ifdef CONFIG_TCP_AO
610 u8 maclen = tcp_ao_maclen(key->ao_key);
611
612 if (tcprsk) {
613 u8 aolen = maclen + sizeof(struct tcp_ao_hdr);
614
615 *ptr++ = htonl((TCPOPT_AO << 24) | (aolen << 16) |
616 (tcprsk->ao_keyid << 8) |
617 (tcprsk->ao_rcv_next));
618 } else {
619 struct tcp_ao_key *rnext_key;
620 struct tcp_ao_info *ao_info;
621
622 ao_info = rcu_dereference_check(tp->ao_info,
623 lockdep_sock_is_held(&tp->inet_conn.icsk_inet.sk));
624 rnext_key = READ_ONCE(ao_info->rnext_key);
625 if (WARN_ON_ONCE(!rnext_key))
626 return ptr;
627 *ptr++ = htonl((TCPOPT_AO << 24) |
628 (tcp_ao_len(key->ao_key) << 16) |
629 (key->ao_key->sndid << 8) |
630 (rnext_key->rcvid));
631 }
632 opts->hash_location = (__u8 *)ptr;
633 ptr += maclen / sizeof(*ptr);
634 if (unlikely(maclen % sizeof(*ptr))) {
635 memset(ptr, TCPOPT_NOP, sizeof(*ptr));
636 ptr++;
637 }
638#endif
639 return ptr;
640}
641
642/* Write previously computed TCP options to the packet.
643 *
644 * Beware: Something in the Internet is very sensitive to the ordering of
645 * TCP options, we learned this through the hard way, so be careful here.
646 * Luckily we can at least blame others for their non-compliance but from
647 * inter-operability perspective it seems that we're somewhat stuck with
648 * the ordering which we have been using if we want to keep working with
649 * those broken things (not that it currently hurts anybody as there isn't
650 * particular reason why the ordering would need to be changed).
651 *
652 * At least SACK_PERM as the first option is known to lead to a disaster
653 * (but it may well be that other scenarios fail similarly).
654 */
655static void tcp_options_write(struct tcphdr *th, struct tcp_sock *tp,
656 const struct tcp_request_sock *tcprsk,
657 struct tcp_out_options *opts,
658 struct tcp_key *key)
659{
660 __be32 *ptr = (__be32 *)(th + 1);
661 u16 options = opts->options; /* mungable copy */
662
663 if (tcp_key_is_md5(key)) {
664 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
665 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
666 /* overload cookie hash location */
667 opts->hash_location = (__u8 *)ptr;
668 ptr += 4;
669 } else if (tcp_key_is_ao(key)) {
670 ptr = process_tcp_ao_options(tp, tcprsk, opts, key, ptr);
671 }
672 if (unlikely(opts->mss)) {
673 *ptr++ = htonl((TCPOPT_MSS << 24) |
674 (TCPOLEN_MSS << 16) |
675 opts->mss);
676 }
677
678 if (likely(OPTION_TS & options)) {
679 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
680 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
681 (TCPOLEN_SACK_PERM << 16) |
682 (TCPOPT_TIMESTAMP << 8) |
683 TCPOLEN_TIMESTAMP);
684 options &= ~OPTION_SACK_ADVERTISE;
685 } else {
686 *ptr++ = htonl((TCPOPT_NOP << 24) |
687 (TCPOPT_NOP << 16) |
688 (TCPOPT_TIMESTAMP << 8) |
689 TCPOLEN_TIMESTAMP);
690 }
691 *ptr++ = htonl(opts->tsval);
692 *ptr++ = htonl(opts->tsecr);
693 }
694
695 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
696 *ptr++ = htonl((TCPOPT_NOP << 24) |
697 (TCPOPT_NOP << 16) |
698 (TCPOPT_SACK_PERM << 8) |
699 TCPOLEN_SACK_PERM);
700 }
701
702 if (unlikely(OPTION_WSCALE & options)) {
703 *ptr++ = htonl((TCPOPT_NOP << 24) |
704 (TCPOPT_WINDOW << 16) |
705 (TCPOLEN_WINDOW << 8) |
706 opts->ws);
707 }
708
709 if (unlikely(opts->num_sack_blocks)) {
710 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
711 tp->duplicate_sack : tp->selective_acks;
712 int this_sack;
713
714 *ptr++ = htonl((TCPOPT_NOP << 24) |
715 (TCPOPT_NOP << 16) |
716 (TCPOPT_SACK << 8) |
717 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
718 TCPOLEN_SACK_PERBLOCK)));
719
720 for (this_sack = 0; this_sack < opts->num_sack_blocks;
721 ++this_sack) {
722 *ptr++ = htonl(sp[this_sack].start_seq);
723 *ptr++ = htonl(sp[this_sack].end_seq);
724 }
725
726 tp->rx_opt.dsack = 0;
727 }
728
729 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
730 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
731 u8 *p = (u8 *)ptr;
732 u32 len; /* Fast Open option length */
733
734 if (foc->exp) {
735 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
736 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
737 TCPOPT_FASTOPEN_MAGIC);
738 p += TCPOLEN_EXP_FASTOPEN_BASE;
739 } else {
740 len = TCPOLEN_FASTOPEN_BASE + foc->len;
741 *p++ = TCPOPT_FASTOPEN;
742 *p++ = len;
743 }
744
745 memcpy(p, foc->val, foc->len);
746 if ((len & 3) == 2) {
747 p[foc->len] = TCPOPT_NOP;
748 p[foc->len + 1] = TCPOPT_NOP;
749 }
750 ptr += (len + 3) >> 2;
751 }
752
753 smc_options_write(ptr, &options);
754
755 mptcp_options_write(th, ptr, tp, opts);
756}
757
758static void smc_set_option(const struct tcp_sock *tp,
759 struct tcp_out_options *opts,
760 unsigned int *remaining)
761{
762#if IS_ENABLED(CONFIG_SMC)
763 if (static_branch_unlikely(&tcp_have_smc)) {
764 if (tp->syn_smc) {
765 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
766 opts->options |= OPTION_SMC;
767 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
768 }
769 }
770 }
771#endif
772}
773
774static void smc_set_option_cond(const struct tcp_sock *tp,
775 const struct inet_request_sock *ireq,
776 struct tcp_out_options *opts,
777 unsigned int *remaining)
778{
779#if IS_ENABLED(CONFIG_SMC)
780 if (static_branch_unlikely(&tcp_have_smc)) {
781 if (tp->syn_smc && ireq->smc_ok) {
782 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
783 opts->options |= OPTION_SMC;
784 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
785 }
786 }
787 }
788#endif
789}
790
791static void mptcp_set_option_cond(const struct request_sock *req,
792 struct tcp_out_options *opts,
793 unsigned int *remaining)
794{
795 if (rsk_is_mptcp(req)) {
796 unsigned int size;
797
798 if (mptcp_synack_options(req, &size, &opts->mptcp)) {
799 if (*remaining >= size) {
800 opts->options |= OPTION_MPTCP;
801 *remaining -= size;
802 }
803 }
804 }
805}
806
807/* Compute TCP options for SYN packets. This is not the final
808 * network wire format yet.
809 */
810static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
811 struct tcp_out_options *opts,
812 struct tcp_key *key)
813{
814 struct tcp_sock *tp = tcp_sk(sk);
815 unsigned int remaining = MAX_TCP_OPTION_SPACE;
816 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
817 bool timestamps;
818
819 /* Better than switch (key.type) as it has static branches */
820 if (tcp_key_is_md5(key)) {
821 timestamps = false;
822 opts->options |= OPTION_MD5;
823 remaining -= TCPOLEN_MD5SIG_ALIGNED;
824 } else {
825 timestamps = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps);
826 if (tcp_key_is_ao(key)) {
827 opts->options |= OPTION_AO;
828 remaining -= tcp_ao_len_aligned(key->ao_key);
829 }
830 }
831
832 /* We always get an MSS option. The option bytes which will be seen in
833 * normal data packets should timestamps be used, must be in the MSS
834 * advertised. But we subtract them from tp->mss_cache so that
835 * calculations in tcp_sendmsg are simpler etc. So account for this
836 * fact here if necessary. If we don't do this correctly, as a
837 * receiver we won't recognize data packets as being full sized when we
838 * should, and thus we won't abide by the delayed ACK rules correctly.
839 * SACKs don't matter, we never delay an ACK when we have any of those
840 * going out. */
841 opts->mss = tcp_advertise_mss(sk);
842 remaining -= TCPOLEN_MSS_ALIGNED;
843
844 if (likely(timestamps)) {
845 opts->options |= OPTION_TS;
846 opts->tsval = tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb) + tp->tsoffset;
847 opts->tsecr = tp->rx_opt.ts_recent;
848 remaining -= TCPOLEN_TSTAMP_ALIGNED;
849 }
850 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling))) {
851 opts->ws = tp->rx_opt.rcv_wscale;
852 opts->options |= OPTION_WSCALE;
853 remaining -= TCPOLEN_WSCALE_ALIGNED;
854 }
855 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_sack))) {
856 opts->options |= OPTION_SACK_ADVERTISE;
857 if (unlikely(!(OPTION_TS & opts->options)))
858 remaining -= TCPOLEN_SACKPERM_ALIGNED;
859 }
860
861 if (fastopen && fastopen->cookie.len >= 0) {
862 u32 need = fastopen->cookie.len;
863
864 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
865 TCPOLEN_FASTOPEN_BASE;
866 need = (need + 3) & ~3U; /* Align to 32 bits */
867 if (remaining >= need) {
868 opts->options |= OPTION_FAST_OPEN_COOKIE;
869 opts->fastopen_cookie = &fastopen->cookie;
870 remaining -= need;
871 tp->syn_fastopen = 1;
872 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
873 }
874 }
875
876 smc_set_option(tp, opts, &remaining);
877
878 if (sk_is_mptcp(sk)) {
879 unsigned int size;
880
881 if (mptcp_syn_options(sk, skb, &size, &opts->mptcp)) {
882 opts->options |= OPTION_MPTCP;
883 remaining -= size;
884 }
885 }
886
887 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
888
889 return MAX_TCP_OPTION_SPACE - remaining;
890}
891
892/* Set up TCP options for SYN-ACKs. */
893static unsigned int tcp_synack_options(const struct sock *sk,
894 struct request_sock *req,
895 unsigned int mss, struct sk_buff *skb,
896 struct tcp_out_options *opts,
897 const struct tcp_key *key,
898 struct tcp_fastopen_cookie *foc,
899 enum tcp_synack_type synack_type,
900 struct sk_buff *syn_skb)
901{
902 struct inet_request_sock *ireq = inet_rsk(req);
903 unsigned int remaining = MAX_TCP_OPTION_SPACE;
904
905 if (tcp_key_is_md5(key)) {
906 opts->options |= OPTION_MD5;
907 remaining -= TCPOLEN_MD5SIG_ALIGNED;
908
909 /* We can't fit any SACK blocks in a packet with MD5 + TS
910 * options. There was discussion about disabling SACK
911 * rather than TS in order to fit in better with old,
912 * buggy kernels, but that was deemed to be unnecessary.
913 */
914 if (synack_type != TCP_SYNACK_COOKIE)
915 ireq->tstamp_ok &= !ireq->sack_ok;
916 } else if (tcp_key_is_ao(key)) {
917 opts->options |= OPTION_AO;
918 remaining -= tcp_ao_len_aligned(key->ao_key);
919 ireq->tstamp_ok &= !ireq->sack_ok;
920 }
921
922 /* We always send an MSS option. */
923 opts->mss = mss;
924 remaining -= TCPOLEN_MSS_ALIGNED;
925
926 if (likely(ireq->wscale_ok)) {
927 opts->ws = ireq->rcv_wscale;
928 opts->options |= OPTION_WSCALE;
929 remaining -= TCPOLEN_WSCALE_ALIGNED;
930 }
931 if (likely(ireq->tstamp_ok)) {
932 opts->options |= OPTION_TS;
933 opts->tsval = tcp_skb_timestamp_ts(tcp_rsk(req)->req_usec_ts, skb) +
934 tcp_rsk(req)->ts_off;
935 opts->tsecr = READ_ONCE(req->ts_recent);
936 remaining -= TCPOLEN_TSTAMP_ALIGNED;
937 }
938 if (likely(ireq->sack_ok)) {
939 opts->options |= OPTION_SACK_ADVERTISE;
940 if (unlikely(!ireq->tstamp_ok))
941 remaining -= TCPOLEN_SACKPERM_ALIGNED;
942 }
943 if (foc != NULL && foc->len >= 0) {
944 u32 need = foc->len;
945
946 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
947 TCPOLEN_FASTOPEN_BASE;
948 need = (need + 3) & ~3U; /* Align to 32 bits */
949 if (remaining >= need) {
950 opts->options |= OPTION_FAST_OPEN_COOKIE;
951 opts->fastopen_cookie = foc;
952 remaining -= need;
953 }
954 }
955
956 mptcp_set_option_cond(req, opts, &remaining);
957
958 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
959
960 bpf_skops_hdr_opt_len((struct sock *)sk, skb, req, syn_skb,
961 synack_type, opts, &remaining);
962
963 return MAX_TCP_OPTION_SPACE - remaining;
964}
965
966/* Compute TCP options for ESTABLISHED sockets. This is not the
967 * final wire format yet.
968 */
969static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
970 struct tcp_out_options *opts,
971 struct tcp_key *key)
972{
973 struct tcp_sock *tp = tcp_sk(sk);
974 unsigned int size = 0;
975 unsigned int eff_sacks;
976
977 opts->options = 0;
978
979 /* Better than switch (key.type) as it has static branches */
980 if (tcp_key_is_md5(key)) {
981 opts->options |= OPTION_MD5;
982 size += TCPOLEN_MD5SIG_ALIGNED;
983 } else if (tcp_key_is_ao(key)) {
984 opts->options |= OPTION_AO;
985 size += tcp_ao_len_aligned(key->ao_key);
986 }
987
988 if (likely(tp->rx_opt.tstamp_ok)) {
989 opts->options |= OPTION_TS;
990 opts->tsval = skb ? tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb) +
991 tp->tsoffset : 0;
992 opts->tsecr = tp->rx_opt.ts_recent;
993 size += TCPOLEN_TSTAMP_ALIGNED;
994 }
995
996 /* MPTCP options have precedence over SACK for the limited TCP
997 * option space because a MPTCP connection would be forced to
998 * fall back to regular TCP if a required multipath option is
999 * missing. SACK still gets a chance to use whatever space is
1000 * left.
1001 */
1002 if (sk_is_mptcp(sk)) {
1003 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
1004 unsigned int opt_size = 0;
1005
1006 if (mptcp_established_options(sk, skb, &opt_size, remaining,
1007 &opts->mptcp)) {
1008 opts->options |= OPTION_MPTCP;
1009 size += opt_size;
1010 }
1011 }
1012
1013 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
1014 if (unlikely(eff_sacks)) {
1015 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
1016 if (unlikely(remaining < TCPOLEN_SACK_BASE_ALIGNED +
1017 TCPOLEN_SACK_PERBLOCK))
1018 return size;
1019
1020 opts->num_sack_blocks =
1021 min_t(unsigned int, eff_sacks,
1022 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
1023 TCPOLEN_SACK_PERBLOCK);
1024
1025 size += TCPOLEN_SACK_BASE_ALIGNED +
1026 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
1027 }
1028
1029 if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp,
1030 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) {
1031 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
1032
1033 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, 0, opts, &remaining);
1034
1035 size = MAX_TCP_OPTION_SPACE - remaining;
1036 }
1037
1038 return size;
1039}
1040
1041
1042/* TCP SMALL QUEUES (TSQ)
1043 *
1044 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
1045 * to reduce RTT and bufferbloat.
1046 * We do this using a special skb destructor (tcp_wfree).
1047 *
1048 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
1049 * needs to be reallocated in a driver.
1050 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
1051 *
1052 * Since transmit from skb destructor is forbidden, we use a tasklet
1053 * to process all sockets that eventually need to send more skbs.
1054 * We use one tasklet per cpu, with its own queue of sockets.
1055 */
1056struct tsq_tasklet {
1057 struct tasklet_struct tasklet;
1058 struct list_head head; /* queue of tcp sockets */
1059};
1060static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
1061
1062static void tcp_tsq_write(struct sock *sk)
1063{
1064 if ((1 << sk->sk_state) &
1065 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
1066 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
1067 struct tcp_sock *tp = tcp_sk(sk);
1068
1069 if (tp->lost_out > tp->retrans_out &&
1070 tcp_snd_cwnd(tp) > tcp_packets_in_flight(tp)) {
1071 tcp_mstamp_refresh(tp);
1072 tcp_xmit_retransmit_queue(sk);
1073 }
1074
1075 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
1076 0, GFP_ATOMIC);
1077 }
1078}
1079
1080static void tcp_tsq_handler(struct sock *sk)
1081{
1082 bh_lock_sock(sk);
1083 if (!sock_owned_by_user(sk))
1084 tcp_tsq_write(sk);
1085 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
1086 sock_hold(sk);
1087 bh_unlock_sock(sk);
1088}
1089/*
1090 * One tasklet per cpu tries to send more skbs.
1091 * We run in tasklet context but need to disable irqs when
1092 * transferring tsq->head because tcp_wfree() might
1093 * interrupt us (non NAPI drivers)
1094 */
1095static void tcp_tasklet_func(struct tasklet_struct *t)
1096{
1097 struct tsq_tasklet *tsq = from_tasklet(tsq, t, tasklet);
1098 LIST_HEAD(list);
1099 unsigned long flags;
1100 struct list_head *q, *n;
1101 struct tcp_sock *tp;
1102 struct sock *sk;
1103
1104 local_irq_save(flags);
1105 list_splice_init(&tsq->head, &list);
1106 local_irq_restore(flags);
1107
1108 list_for_each_safe(q, n, &list) {
1109 tp = list_entry(q, struct tcp_sock, tsq_node);
1110 list_del(&tp->tsq_node);
1111
1112 sk = (struct sock *)tp;
1113 smp_mb__before_atomic();
1114 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
1115
1116 tcp_tsq_handler(sk);
1117 sk_free(sk);
1118 }
1119}
1120
1121#define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
1122 TCPF_WRITE_TIMER_DEFERRED | \
1123 TCPF_DELACK_TIMER_DEFERRED | \
1124 TCPF_MTU_REDUCED_DEFERRED | \
1125 TCPF_ACK_DEFERRED)
1126/**
1127 * tcp_release_cb - tcp release_sock() callback
1128 * @sk: socket
1129 *
1130 * called from release_sock() to perform protocol dependent
1131 * actions before socket release.
1132 */
1133void tcp_release_cb(struct sock *sk)
1134{
1135 unsigned long flags = smp_load_acquire(&sk->sk_tsq_flags);
1136 unsigned long nflags;
1137
1138 /* perform an atomic operation only if at least one flag is set */
1139 do {
1140 if (!(flags & TCP_DEFERRED_ALL))
1141 return;
1142 nflags = flags & ~TCP_DEFERRED_ALL;
1143 } while (!try_cmpxchg(&sk->sk_tsq_flags, &flags, nflags));
1144
1145 if (flags & TCPF_TSQ_DEFERRED) {
1146 tcp_tsq_write(sk);
1147 __sock_put(sk);
1148 }
1149
1150 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
1151 tcp_write_timer_handler(sk);
1152 __sock_put(sk);
1153 }
1154 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
1155 tcp_delack_timer_handler(sk);
1156 __sock_put(sk);
1157 }
1158 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
1159 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
1160 __sock_put(sk);
1161 }
1162 if ((flags & TCPF_ACK_DEFERRED) && inet_csk_ack_scheduled(sk))
1163 tcp_send_ack(sk);
1164}
1165EXPORT_SYMBOL(tcp_release_cb);
1166
1167void __init tcp_tasklet_init(void)
1168{
1169 int i;
1170
1171 for_each_possible_cpu(i) {
1172 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
1173
1174 INIT_LIST_HEAD(&tsq->head);
1175 tasklet_setup(&tsq->tasklet, tcp_tasklet_func);
1176 }
1177}
1178
1179/*
1180 * Write buffer destructor automatically called from kfree_skb.
1181 * We can't xmit new skbs from this context, as we might already
1182 * hold qdisc lock.
1183 */
1184void tcp_wfree(struct sk_buff *skb)
1185{
1186 struct sock *sk = skb->sk;
1187 struct tcp_sock *tp = tcp_sk(sk);
1188 unsigned long flags, nval, oval;
1189 struct tsq_tasklet *tsq;
1190 bool empty;
1191
1192 /* Keep one reference on sk_wmem_alloc.
1193 * Will be released by sk_free() from here or tcp_tasklet_func()
1194 */
1195 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
1196
1197 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
1198 * Wait until our queues (qdisc + devices) are drained.
1199 * This gives :
1200 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
1201 * - chance for incoming ACK (processed by another cpu maybe)
1202 * to migrate this flow (skb->ooo_okay will be eventually set)
1203 */
1204 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
1205 goto out;
1206
1207 oval = smp_load_acquire(&sk->sk_tsq_flags);
1208 do {
1209 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
1210 goto out;
1211
1212 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
1213 } while (!try_cmpxchg(&sk->sk_tsq_flags, &oval, nval));
1214
1215 /* queue this socket to tasklet queue */
1216 local_irq_save(flags);
1217 tsq = this_cpu_ptr(&tsq_tasklet);
1218 empty = list_empty(&tsq->head);
1219 list_add(&tp->tsq_node, &tsq->head);
1220 if (empty)
1221 tasklet_schedule(&tsq->tasklet);
1222 local_irq_restore(flags);
1223 return;
1224out:
1225 sk_free(sk);
1226}
1227
1228/* Note: Called under soft irq.
1229 * We can call TCP stack right away, unless socket is owned by user.
1230 */
1231enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
1232{
1233 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
1234 struct sock *sk = (struct sock *)tp;
1235
1236 tcp_tsq_handler(sk);
1237 sock_put(sk);
1238
1239 return HRTIMER_NORESTART;
1240}
1241
1242static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
1243 u64 prior_wstamp)
1244{
1245 struct tcp_sock *tp = tcp_sk(sk);
1246
1247 if (sk->sk_pacing_status != SK_PACING_NONE) {
1248 unsigned long rate = READ_ONCE(sk->sk_pacing_rate);
1249
1250 /* Original sch_fq does not pace first 10 MSS
1251 * Note that tp->data_segs_out overflows after 2^32 packets,
1252 * this is a minor annoyance.
1253 */
1254 if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
1255 u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
1256 u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
1257
1258 /* take into account OS jitter */
1259 len_ns -= min_t(u64, len_ns / 2, credit);
1260 tp->tcp_wstamp_ns += len_ns;
1261 }
1262 }
1263 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1264}
1265
1266INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1267INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1268INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb));
1269
1270/* This routine actually transmits TCP packets queued in by
1271 * tcp_do_sendmsg(). This is used by both the initial
1272 * transmission and possible later retransmissions.
1273 * All SKB's seen here are completely headerless. It is our
1274 * job to build the TCP header, and pass the packet down to
1275 * IP so it can do the same plus pass the packet off to the
1276 * device.
1277 *
1278 * We are working here with either a clone of the original
1279 * SKB, or a fresh unique copy made by the retransmit engine.
1280 */
1281static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1282 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1283{
1284 const struct inet_connection_sock *icsk = inet_csk(sk);
1285 struct inet_sock *inet;
1286 struct tcp_sock *tp;
1287 struct tcp_skb_cb *tcb;
1288 struct tcp_out_options opts;
1289 unsigned int tcp_options_size, tcp_header_size;
1290 struct sk_buff *oskb = NULL;
1291 struct tcp_key key;
1292 struct tcphdr *th;
1293 u64 prior_wstamp;
1294 int err;
1295
1296 BUG_ON(!skb || !tcp_skb_pcount(skb));
1297 tp = tcp_sk(sk);
1298 prior_wstamp = tp->tcp_wstamp_ns;
1299 tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1300 skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true);
1301 if (clone_it) {
1302 oskb = skb;
1303
1304 tcp_skb_tsorted_save(oskb) {
1305 if (unlikely(skb_cloned(oskb)))
1306 skb = pskb_copy(oskb, gfp_mask);
1307 else
1308 skb = skb_clone(oskb, gfp_mask);
1309 } tcp_skb_tsorted_restore(oskb);
1310
1311 if (unlikely(!skb))
1312 return -ENOBUFS;
1313 /* retransmit skbs might have a non zero value in skb->dev
1314 * because skb->dev is aliased with skb->rbnode.rb_left
1315 */
1316 skb->dev = NULL;
1317 }
1318
1319 inet = inet_sk(sk);
1320 tcb = TCP_SKB_CB(skb);
1321 memset(&opts, 0, sizeof(opts));
1322
1323 tcp_get_current_key(sk, &key);
1324 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1325 tcp_options_size = tcp_syn_options(sk, skb, &opts, &key);
1326 } else {
1327 tcp_options_size = tcp_established_options(sk, skb, &opts, &key);
1328 /* Force a PSH flag on all (GSO) packets to expedite GRO flush
1329 * at receiver : This slightly improve GRO performance.
1330 * Note that we do not force the PSH flag for non GSO packets,
1331 * because they might be sent under high congestion events,
1332 * and in this case it is better to delay the delivery of 1-MSS
1333 * packets and thus the corresponding ACK packet that would
1334 * release the following packet.
1335 */
1336 if (tcp_skb_pcount(skb) > 1)
1337 tcb->tcp_flags |= TCPHDR_PSH;
1338 }
1339 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1340
1341 /* We set skb->ooo_okay to one if this packet can select
1342 * a different TX queue than prior packets of this flow,
1343 * to avoid self inflicted reorders.
1344 * The 'other' queue decision is based on current cpu number
1345 * if XPS is enabled, or sk->sk_txhash otherwise.
1346 * We can switch to another (and better) queue if:
1347 * 1) No packet with payload is in qdisc/device queues.
1348 * Delays in TX completion can defeat the test
1349 * even if packets were already sent.
1350 * 2) Or rtx queue is empty.
1351 * This mitigates above case if ACK packets for
1352 * all prior packets were already processed.
1353 */
1354 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1) ||
1355 tcp_rtx_queue_empty(sk);
1356
1357 /* If we had to use memory reserve to allocate this skb,
1358 * this might cause drops if packet is looped back :
1359 * Other socket might not have SOCK_MEMALLOC.
1360 * Packets not looped back do not care about pfmemalloc.
1361 */
1362 skb->pfmemalloc = 0;
1363
1364 skb_push(skb, tcp_header_size);
1365 skb_reset_transport_header(skb);
1366
1367 skb_orphan(skb);
1368 skb->sk = sk;
1369 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1370 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1371
1372 skb_set_dst_pending_confirm(skb, READ_ONCE(sk->sk_dst_pending_confirm));
1373
1374 /* Build TCP header and checksum it. */
1375 th = (struct tcphdr *)skb->data;
1376 th->source = inet->inet_sport;
1377 th->dest = inet->inet_dport;
1378 th->seq = htonl(tcb->seq);
1379 th->ack_seq = htonl(rcv_nxt);
1380 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1381 tcb->tcp_flags);
1382
1383 th->check = 0;
1384 th->urg_ptr = 0;
1385
1386 /* The urg_mode check is necessary during a below snd_una win probe */
1387 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1388 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1389 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1390 th->urg = 1;
1391 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1392 th->urg_ptr = htons(0xFFFF);
1393 th->urg = 1;
1394 }
1395 }
1396
1397 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1398 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1399 th->window = htons(tcp_select_window(sk));
1400 tcp_ecn_send(sk, skb, th, tcp_header_size);
1401 } else {
1402 /* RFC1323: The window in SYN & SYN/ACK segments
1403 * is never scaled.
1404 */
1405 th->window = htons(min(tp->rcv_wnd, 65535U));
1406 }
1407
1408 tcp_options_write(th, tp, NULL, &opts, &key);
1409
1410 if (tcp_key_is_md5(&key)) {
1411#ifdef CONFIG_TCP_MD5SIG
1412 /* Calculate the MD5 hash, as we have all we need now */
1413 sk_gso_disable(sk);
1414 tp->af_specific->calc_md5_hash(opts.hash_location,
1415 key.md5_key, sk, skb);
1416#endif
1417 } else if (tcp_key_is_ao(&key)) {
1418 int err;
1419
1420 err = tcp_ao_transmit_skb(sk, skb, key.ao_key, th,
1421 opts.hash_location);
1422 if (err) {
1423 kfree_skb_reason(skb, SKB_DROP_REASON_NOT_SPECIFIED);
1424 return -ENOMEM;
1425 }
1426 }
1427
1428 /* BPF prog is the last one writing header option */
1429 bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, 0, &opts);
1430
1431 INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check,
1432 tcp_v6_send_check, tcp_v4_send_check,
1433 sk, skb);
1434
1435 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1436 tcp_event_ack_sent(sk, rcv_nxt);
1437
1438 if (skb->len != tcp_header_size) {
1439 tcp_event_data_sent(tp, sk);
1440 tp->data_segs_out += tcp_skb_pcount(skb);
1441 tp->bytes_sent += skb->len - tcp_header_size;
1442 }
1443
1444 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1445 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1446 tcp_skb_pcount(skb));
1447
1448 tp->segs_out += tcp_skb_pcount(skb);
1449 skb_set_hash_from_sk(skb, sk);
1450 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1451 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1452 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1453
1454 /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1455
1456 /* Cleanup our debris for IP stacks */
1457 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1458 sizeof(struct inet6_skb_parm)));
1459
1460 tcp_add_tx_delay(skb, tp);
1461
1462 err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit,
1463 inet6_csk_xmit, ip_queue_xmit,
1464 sk, skb, &inet->cork.fl);
1465
1466 if (unlikely(err > 0)) {
1467 tcp_enter_cwr(sk);
1468 err = net_xmit_eval(err);
1469 }
1470 if (!err && oskb) {
1471 tcp_update_skb_after_send(sk, oskb, prior_wstamp);
1472 tcp_rate_skb_sent(sk, oskb);
1473 }
1474 return err;
1475}
1476
1477static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1478 gfp_t gfp_mask)
1479{
1480 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1481 tcp_sk(sk)->rcv_nxt);
1482}
1483
1484/* This routine just queues the buffer for sending.
1485 *
1486 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1487 * otherwise socket can stall.
1488 */
1489static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1490{
1491 struct tcp_sock *tp = tcp_sk(sk);
1492
1493 /* Advance write_seq and place onto the write_queue. */
1494 WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1495 __skb_header_release(skb);
1496 tcp_add_write_queue_tail(sk, skb);
1497 sk_wmem_queued_add(sk, skb->truesize);
1498 sk_mem_charge(sk, skb->truesize);
1499}
1500
1501/* Initialize TSO segments for a packet. */
1502static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1503{
1504 if (skb->len <= mss_now) {
1505 /* Avoid the costly divide in the normal
1506 * non-TSO case.
1507 */
1508 tcp_skb_pcount_set(skb, 1);
1509 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1510 } else {
1511 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1512 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1513 }
1514}
1515
1516/* Pcount in the middle of the write queue got changed, we need to do various
1517 * tweaks to fix counters
1518 */
1519static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1520{
1521 struct tcp_sock *tp = tcp_sk(sk);
1522
1523 tp->packets_out -= decr;
1524
1525 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1526 tp->sacked_out -= decr;
1527 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1528 tp->retrans_out -= decr;
1529 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1530 tp->lost_out -= decr;
1531
1532 /* Reno case is special. Sigh... */
1533 if (tcp_is_reno(tp) && decr > 0)
1534 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1535
1536 if (tp->lost_skb_hint &&
1537 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1538 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1539 tp->lost_cnt_hint -= decr;
1540
1541 tcp_verify_left_out(tp);
1542}
1543
1544static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1545{
1546 return TCP_SKB_CB(skb)->txstamp_ack ||
1547 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1548}
1549
1550static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1551{
1552 struct skb_shared_info *shinfo = skb_shinfo(skb);
1553
1554 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1555 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1556 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1557 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1558
1559 shinfo->tx_flags &= ~tsflags;
1560 shinfo2->tx_flags |= tsflags;
1561 swap(shinfo->tskey, shinfo2->tskey);
1562 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1563 TCP_SKB_CB(skb)->txstamp_ack = 0;
1564 }
1565}
1566
1567static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1568{
1569 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1570 TCP_SKB_CB(skb)->eor = 0;
1571}
1572
1573/* Insert buff after skb on the write or rtx queue of sk. */
1574static void tcp_insert_write_queue_after(struct sk_buff *skb,
1575 struct sk_buff *buff,
1576 struct sock *sk,
1577 enum tcp_queue tcp_queue)
1578{
1579 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1580 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1581 else
1582 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1583}
1584
1585/* Function to create two new TCP segments. Shrinks the given segment
1586 * to the specified size and appends a new segment with the rest of the
1587 * packet to the list. This won't be called frequently, I hope.
1588 * Remember, these are still headerless SKBs at this point.
1589 */
1590int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1591 struct sk_buff *skb, u32 len,
1592 unsigned int mss_now, gfp_t gfp)
1593{
1594 struct tcp_sock *tp = tcp_sk(sk);
1595 struct sk_buff *buff;
1596 int old_factor;
1597 long limit;
1598 int nlen;
1599 u8 flags;
1600
1601 if (WARN_ON(len > skb->len))
1602 return -EINVAL;
1603
1604 DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb));
1605
1606 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1607 * We need some allowance to not penalize applications setting small
1608 * SO_SNDBUF values.
1609 * Also allow first and last skb in retransmit queue to be split.
1610 */
1611 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_LEGACY_MAX_SIZE);
1612 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1613 tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1614 skb != tcp_rtx_queue_head(sk) &&
1615 skb != tcp_rtx_queue_tail(sk))) {
1616 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1617 return -ENOMEM;
1618 }
1619
1620 if (skb_unclone_keeptruesize(skb, gfp))
1621 return -ENOMEM;
1622
1623 /* Get a new skb... force flag on. */
1624 buff = tcp_stream_alloc_skb(sk, gfp, true);
1625 if (!buff)
1626 return -ENOMEM; /* We'll just try again later. */
1627 skb_copy_decrypted(buff, skb);
1628 mptcp_skb_ext_copy(buff, skb);
1629
1630 sk_wmem_queued_add(sk, buff->truesize);
1631 sk_mem_charge(sk, buff->truesize);
1632 nlen = skb->len - len;
1633 buff->truesize += nlen;
1634 skb->truesize -= nlen;
1635
1636 /* Correct the sequence numbers. */
1637 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1638 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1639 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1640
1641 /* PSH and FIN should only be set in the second packet. */
1642 flags = TCP_SKB_CB(skb)->tcp_flags;
1643 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1644 TCP_SKB_CB(buff)->tcp_flags = flags;
1645 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1646 tcp_skb_fragment_eor(skb, buff);
1647
1648 skb_split(skb, buff, len);
1649
1650 skb_set_delivery_time(buff, skb->tstamp, true);
1651 tcp_fragment_tstamp(skb, buff);
1652
1653 old_factor = tcp_skb_pcount(skb);
1654
1655 /* Fix up tso_factor for both original and new SKB. */
1656 tcp_set_skb_tso_segs(skb, mss_now);
1657 tcp_set_skb_tso_segs(buff, mss_now);
1658
1659 /* Update delivered info for the new segment */
1660 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1661
1662 /* If this packet has been sent out already, we must
1663 * adjust the various packet counters.
1664 */
1665 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1666 int diff = old_factor - tcp_skb_pcount(skb) -
1667 tcp_skb_pcount(buff);
1668
1669 if (diff)
1670 tcp_adjust_pcount(sk, skb, diff);
1671 }
1672
1673 /* Link BUFF into the send queue. */
1674 __skb_header_release(buff);
1675 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1676 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1677 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1678
1679 return 0;
1680}
1681
1682/* This is similar to __pskb_pull_tail(). The difference is that pulled
1683 * data is not copied, but immediately discarded.
1684 */
1685static int __pskb_trim_head(struct sk_buff *skb, int len)
1686{
1687 struct skb_shared_info *shinfo;
1688 int i, k, eat;
1689
1690 DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb));
1691 eat = len;
1692 k = 0;
1693 shinfo = skb_shinfo(skb);
1694 for (i = 0; i < shinfo->nr_frags; i++) {
1695 int size = skb_frag_size(&shinfo->frags[i]);
1696
1697 if (size <= eat) {
1698 skb_frag_unref(skb, i);
1699 eat -= size;
1700 } else {
1701 shinfo->frags[k] = shinfo->frags[i];
1702 if (eat) {
1703 skb_frag_off_add(&shinfo->frags[k], eat);
1704 skb_frag_size_sub(&shinfo->frags[k], eat);
1705 eat = 0;
1706 }
1707 k++;
1708 }
1709 }
1710 shinfo->nr_frags = k;
1711
1712 skb->data_len -= len;
1713 skb->len = skb->data_len;
1714 return len;
1715}
1716
1717/* Remove acked data from a packet in the transmit queue. */
1718int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1719{
1720 u32 delta_truesize;
1721
1722 if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
1723 return -ENOMEM;
1724
1725 delta_truesize = __pskb_trim_head(skb, len);
1726
1727 TCP_SKB_CB(skb)->seq += len;
1728
1729 skb->truesize -= delta_truesize;
1730 sk_wmem_queued_add(sk, -delta_truesize);
1731 if (!skb_zcopy_pure(skb))
1732 sk_mem_uncharge(sk, delta_truesize);
1733
1734 /* Any change of skb->len requires recalculation of tso factor. */
1735 if (tcp_skb_pcount(skb) > 1)
1736 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1737
1738 return 0;
1739}
1740
1741/* Calculate MSS not accounting any TCP options. */
1742static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1743{
1744 const struct tcp_sock *tp = tcp_sk(sk);
1745 const struct inet_connection_sock *icsk = inet_csk(sk);
1746 int mss_now;
1747
1748 /* Calculate base mss without TCP options:
1749 It is MMS_S - sizeof(tcphdr) of rfc1122
1750 */
1751 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1752
1753 /* Clamp it (mss_clamp does not include tcp options) */
1754 if (mss_now > tp->rx_opt.mss_clamp)
1755 mss_now = tp->rx_opt.mss_clamp;
1756
1757 /* Now subtract optional transport overhead */
1758 mss_now -= icsk->icsk_ext_hdr_len;
1759
1760 /* Then reserve room for full set of TCP options and 8 bytes of data */
1761 mss_now = max(mss_now,
1762 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss));
1763 return mss_now;
1764}
1765
1766/* Calculate MSS. Not accounting for SACKs here. */
1767int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1768{
1769 /* Subtract TCP options size, not including SACKs */
1770 return __tcp_mtu_to_mss(sk, pmtu) -
1771 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1772}
1773EXPORT_SYMBOL(tcp_mtu_to_mss);
1774
1775/* Inverse of above */
1776int tcp_mss_to_mtu(struct sock *sk, int mss)
1777{
1778 const struct tcp_sock *tp = tcp_sk(sk);
1779 const struct inet_connection_sock *icsk = inet_csk(sk);
1780
1781 return mss +
1782 tp->tcp_header_len +
1783 icsk->icsk_ext_hdr_len +
1784 icsk->icsk_af_ops->net_header_len;
1785}
1786EXPORT_SYMBOL(tcp_mss_to_mtu);
1787
1788/* MTU probing init per socket */
1789void tcp_mtup_init(struct sock *sk)
1790{
1791 struct tcp_sock *tp = tcp_sk(sk);
1792 struct inet_connection_sock *icsk = inet_csk(sk);
1793 struct net *net = sock_net(sk);
1794
1795 icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1;
1796 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1797 icsk->icsk_af_ops->net_header_len;
1798 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss));
1799 icsk->icsk_mtup.probe_size = 0;
1800 if (icsk->icsk_mtup.enabled)
1801 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1802}
1803EXPORT_SYMBOL(tcp_mtup_init);
1804
1805/* This function synchronize snd mss to current pmtu/exthdr set.
1806
1807 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1808 for TCP options, but includes only bare TCP header.
1809
1810 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1811 It is minimum of user_mss and mss received with SYN.
1812 It also does not include TCP options.
1813
1814 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1815
1816 tp->mss_cache is current effective sending mss, including
1817 all tcp options except for SACKs. It is evaluated,
1818 taking into account current pmtu, but never exceeds
1819 tp->rx_opt.mss_clamp.
1820
1821 NOTE1. rfc1122 clearly states that advertised MSS
1822 DOES NOT include either tcp or ip options.
1823
1824 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1825 are READ ONLY outside this function. --ANK (980731)
1826 */
1827unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1828{
1829 struct tcp_sock *tp = tcp_sk(sk);
1830 struct inet_connection_sock *icsk = inet_csk(sk);
1831 int mss_now;
1832
1833 if (icsk->icsk_mtup.search_high > pmtu)
1834 icsk->icsk_mtup.search_high = pmtu;
1835
1836 mss_now = tcp_mtu_to_mss(sk, pmtu);
1837 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1838
1839 /* And store cached results */
1840 icsk->icsk_pmtu_cookie = pmtu;
1841 if (icsk->icsk_mtup.enabled)
1842 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1843 tp->mss_cache = mss_now;
1844
1845 return mss_now;
1846}
1847EXPORT_SYMBOL(tcp_sync_mss);
1848
1849/* Compute the current effective MSS, taking SACKs and IP options,
1850 * and even PMTU discovery events into account.
1851 */
1852unsigned int tcp_current_mss(struct sock *sk)
1853{
1854 const struct tcp_sock *tp = tcp_sk(sk);
1855 const struct dst_entry *dst = __sk_dst_get(sk);
1856 u32 mss_now;
1857 unsigned int header_len;
1858 struct tcp_out_options opts;
1859 struct tcp_key key;
1860
1861 mss_now = tp->mss_cache;
1862
1863 if (dst) {
1864 u32 mtu = dst_mtu(dst);
1865 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1866 mss_now = tcp_sync_mss(sk, mtu);
1867 }
1868 tcp_get_current_key(sk, &key);
1869 header_len = tcp_established_options(sk, NULL, &opts, &key) +
1870 sizeof(struct tcphdr);
1871 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1872 * some common options. If this is an odd packet (because we have SACK
1873 * blocks etc) then our calculated header_len will be different, and
1874 * we have to adjust mss_now correspondingly */
1875 if (header_len != tp->tcp_header_len) {
1876 int delta = (int) header_len - tp->tcp_header_len;
1877 mss_now -= delta;
1878 }
1879
1880 return mss_now;
1881}
1882
1883/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1884 * As additional protections, we do not touch cwnd in retransmission phases,
1885 * and if application hit its sndbuf limit recently.
1886 */
1887static void tcp_cwnd_application_limited(struct sock *sk)
1888{
1889 struct tcp_sock *tp = tcp_sk(sk);
1890
1891 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1892 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1893 /* Limited by application or receiver window. */
1894 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1895 u32 win_used = max(tp->snd_cwnd_used, init_win);
1896 if (win_used < tcp_snd_cwnd(tp)) {
1897 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1898 tcp_snd_cwnd_set(tp, (tcp_snd_cwnd(tp) + win_used) >> 1);
1899 }
1900 tp->snd_cwnd_used = 0;
1901 }
1902 tp->snd_cwnd_stamp = tcp_jiffies32;
1903}
1904
1905static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1906{
1907 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1908 struct tcp_sock *tp = tcp_sk(sk);
1909
1910 /* Track the strongest available signal of the degree to which the cwnd
1911 * is fully utilized. If cwnd-limited then remember that fact for the
1912 * current window. If not cwnd-limited then track the maximum number of
1913 * outstanding packets in the current window. (If cwnd-limited then we
1914 * chose to not update tp->max_packets_out to avoid an extra else
1915 * clause with no functional impact.)
1916 */
1917 if (!before(tp->snd_una, tp->cwnd_usage_seq) ||
1918 is_cwnd_limited ||
1919 (!tp->is_cwnd_limited &&
1920 tp->packets_out > tp->max_packets_out)) {
1921 tp->is_cwnd_limited = is_cwnd_limited;
1922 tp->max_packets_out = tp->packets_out;
1923 tp->cwnd_usage_seq = tp->snd_nxt;
1924 }
1925
1926 if (tcp_is_cwnd_limited(sk)) {
1927 /* Network is feed fully. */
1928 tp->snd_cwnd_used = 0;
1929 tp->snd_cwnd_stamp = tcp_jiffies32;
1930 } else {
1931 /* Network starves. */
1932 if (tp->packets_out > tp->snd_cwnd_used)
1933 tp->snd_cwnd_used = tp->packets_out;
1934
1935 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) &&
1936 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1937 !ca_ops->cong_control)
1938 tcp_cwnd_application_limited(sk);
1939
1940 /* The following conditions together indicate the starvation
1941 * is caused by insufficient sender buffer:
1942 * 1) just sent some data (see tcp_write_xmit)
1943 * 2) not cwnd limited (this else condition)
1944 * 3) no more data to send (tcp_write_queue_empty())
1945 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1946 */
1947 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1948 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1949 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1950 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1951 }
1952}
1953
1954/* Minshall's variant of the Nagle send check. */
1955static bool tcp_minshall_check(const struct tcp_sock *tp)
1956{
1957 return after(tp->snd_sml, tp->snd_una) &&
1958 !after(tp->snd_sml, tp->snd_nxt);
1959}
1960
1961/* Update snd_sml if this skb is under mss
1962 * Note that a TSO packet might end with a sub-mss segment
1963 * The test is really :
1964 * if ((skb->len % mss) != 0)
1965 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1966 * But we can avoid doing the divide again given we already have
1967 * skb_pcount = skb->len / mss_now
1968 */
1969static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1970 const struct sk_buff *skb)
1971{
1972 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1973 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1974}
1975
1976/* Return false, if packet can be sent now without violation Nagle's rules:
1977 * 1. It is full sized. (provided by caller in %partial bool)
1978 * 2. Or it contains FIN. (already checked by caller)
1979 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1980 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1981 * With Minshall's modification: all sent small packets are ACKed.
1982 */
1983static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1984 int nonagle)
1985{
1986 return partial &&
1987 ((nonagle & TCP_NAGLE_CORK) ||
1988 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1989}
1990
1991/* Return how many segs we'd like on a TSO packet,
1992 * depending on current pacing rate, and how close the peer is.
1993 *
1994 * Rationale is:
1995 * - For close peers, we rather send bigger packets to reduce
1996 * cpu costs, because occasional losses will be repaired fast.
1997 * - For long distance/rtt flows, we would like to get ACK clocking
1998 * with 1 ACK per ms.
1999 *
2000 * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting
2001 * in bigger TSO bursts. We we cut the RTT-based allowance in half
2002 * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance
2003 * is below 1500 bytes after 6 * ~500 usec = 3ms.
2004 */
2005static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
2006 int min_tso_segs)
2007{
2008 unsigned long bytes;
2009 u32 r;
2010
2011 bytes = READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift);
2012
2013 r = tcp_min_rtt(tcp_sk(sk)) >> READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_rtt_log);
2014 if (r < BITS_PER_TYPE(sk->sk_gso_max_size))
2015 bytes += sk->sk_gso_max_size >> r;
2016
2017 bytes = min_t(unsigned long, bytes, sk->sk_gso_max_size);
2018
2019 return max_t(u32, bytes / mss_now, min_tso_segs);
2020}
2021
2022/* Return the number of segments we want in the skb we are transmitting.
2023 * See if congestion control module wants to decide; otherwise, autosize.
2024 */
2025static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
2026{
2027 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2028 u32 min_tso, tso_segs;
2029
2030 min_tso = ca_ops->min_tso_segs ?
2031 ca_ops->min_tso_segs(sk) :
2032 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs);
2033
2034 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
2035 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
2036}
2037
2038/* Returns the portion of skb which can be sent right away */
2039static unsigned int tcp_mss_split_point(const struct sock *sk,
2040 const struct sk_buff *skb,
2041 unsigned int mss_now,
2042 unsigned int max_segs,
2043 int nonagle)
2044{
2045 const struct tcp_sock *tp = tcp_sk(sk);
2046 u32 partial, needed, window, max_len;
2047
2048 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2049 max_len = mss_now * max_segs;
2050
2051 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
2052 return max_len;
2053
2054 needed = min(skb->len, window);
2055
2056 if (max_len <= needed)
2057 return max_len;
2058
2059 partial = needed % mss_now;
2060 /* If last segment is not a full MSS, check if Nagle rules allow us
2061 * to include this last segment in this skb.
2062 * Otherwise, we'll split the skb at last MSS boundary
2063 */
2064 if (tcp_nagle_check(partial != 0, tp, nonagle))
2065 return needed - partial;
2066
2067 return needed;
2068}
2069
2070/* Can at least one segment of SKB be sent right now, according to the
2071 * congestion window rules? If so, return how many segments are allowed.
2072 */
2073static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
2074 const struct sk_buff *skb)
2075{
2076 u32 in_flight, cwnd, halfcwnd;
2077
2078 /* Don't be strict about the congestion window for the final FIN. */
2079 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
2080 tcp_skb_pcount(skb) == 1)
2081 return 1;
2082
2083 in_flight = tcp_packets_in_flight(tp);
2084 cwnd = tcp_snd_cwnd(tp);
2085 if (in_flight >= cwnd)
2086 return 0;
2087
2088 /* For better scheduling, ensure we have at least
2089 * 2 GSO packets in flight.
2090 */
2091 halfcwnd = max(cwnd >> 1, 1U);
2092 return min(halfcwnd, cwnd - in_flight);
2093}
2094
2095/* Initialize TSO state of a skb.
2096 * This must be invoked the first time we consider transmitting
2097 * SKB onto the wire.
2098 */
2099static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
2100{
2101 int tso_segs = tcp_skb_pcount(skb);
2102
2103 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
2104 tcp_set_skb_tso_segs(skb, mss_now);
2105 tso_segs = tcp_skb_pcount(skb);
2106 }
2107 return tso_segs;
2108}
2109
2110
2111/* Return true if the Nagle test allows this packet to be
2112 * sent now.
2113 */
2114static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
2115 unsigned int cur_mss, int nonagle)
2116{
2117 /* Nagle rule does not apply to frames, which sit in the middle of the
2118 * write_queue (they have no chances to get new data).
2119 *
2120 * This is implemented in the callers, where they modify the 'nonagle'
2121 * argument based upon the location of SKB in the send queue.
2122 */
2123 if (nonagle & TCP_NAGLE_PUSH)
2124 return true;
2125
2126 /* Don't use the nagle rule for urgent data (or for the final FIN). */
2127 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
2128 return true;
2129
2130 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
2131 return true;
2132
2133 return false;
2134}
2135
2136/* Does at least the first segment of SKB fit into the send window? */
2137static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
2138 const struct sk_buff *skb,
2139 unsigned int cur_mss)
2140{
2141 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2142
2143 if (skb->len > cur_mss)
2144 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
2145
2146 return !after(end_seq, tcp_wnd_end(tp));
2147}
2148
2149/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
2150 * which is put after SKB on the list. It is very much like
2151 * tcp_fragment() except that it may make several kinds of assumptions
2152 * in order to speed up the splitting operation. In particular, we
2153 * know that all the data is in scatter-gather pages, and that the
2154 * packet has never been sent out before (and thus is not cloned).
2155 */
2156static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
2157 unsigned int mss_now, gfp_t gfp)
2158{
2159 int nlen = skb->len - len;
2160 struct sk_buff *buff;
2161 u8 flags;
2162
2163 /* All of a TSO frame must be composed of paged data. */
2164 DEBUG_NET_WARN_ON_ONCE(skb->len != skb->data_len);
2165
2166 buff = tcp_stream_alloc_skb(sk, gfp, true);
2167 if (unlikely(!buff))
2168 return -ENOMEM;
2169 skb_copy_decrypted(buff, skb);
2170 mptcp_skb_ext_copy(buff, skb);
2171
2172 sk_wmem_queued_add(sk, buff->truesize);
2173 sk_mem_charge(sk, buff->truesize);
2174 buff->truesize += nlen;
2175 skb->truesize -= nlen;
2176
2177 /* Correct the sequence numbers. */
2178 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
2179 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
2180 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
2181
2182 /* PSH and FIN should only be set in the second packet. */
2183 flags = TCP_SKB_CB(skb)->tcp_flags;
2184 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
2185 TCP_SKB_CB(buff)->tcp_flags = flags;
2186
2187 tcp_skb_fragment_eor(skb, buff);
2188
2189 skb_split(skb, buff, len);
2190 tcp_fragment_tstamp(skb, buff);
2191
2192 /* Fix up tso_factor for both original and new SKB. */
2193 tcp_set_skb_tso_segs(skb, mss_now);
2194 tcp_set_skb_tso_segs(buff, mss_now);
2195
2196 /* Link BUFF into the send queue. */
2197 __skb_header_release(buff);
2198 tcp_insert_write_queue_after(skb, buff, sk, TCP_FRAG_IN_WRITE_QUEUE);
2199
2200 return 0;
2201}
2202
2203/* Try to defer sending, if possible, in order to minimize the amount
2204 * of TSO splitting we do. View it as a kind of TSO Nagle test.
2205 *
2206 * This algorithm is from John Heffner.
2207 */
2208static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
2209 bool *is_cwnd_limited,
2210 bool *is_rwnd_limited,
2211 u32 max_segs)
2212{
2213 const struct inet_connection_sock *icsk = inet_csk(sk);
2214 u32 send_win, cong_win, limit, in_flight;
2215 struct tcp_sock *tp = tcp_sk(sk);
2216 struct sk_buff *head;
2217 int win_divisor;
2218 s64 delta;
2219
2220 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
2221 goto send_now;
2222
2223 /* Avoid bursty behavior by allowing defer
2224 * only if the last write was recent (1 ms).
2225 * Note that tp->tcp_wstamp_ns can be in the future if we have
2226 * packets waiting in a qdisc or device for EDT delivery.
2227 */
2228 delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
2229 if (delta > 0)
2230 goto send_now;
2231
2232 in_flight = tcp_packets_in_flight(tp);
2233
2234 BUG_ON(tcp_skb_pcount(skb) <= 1);
2235 BUG_ON(tcp_snd_cwnd(tp) <= in_flight);
2236
2237 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2238
2239 /* From in_flight test above, we know that cwnd > in_flight. */
2240 cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache;
2241
2242 limit = min(send_win, cong_win);
2243
2244 /* If a full-sized TSO skb can be sent, do it. */
2245 if (limit >= max_segs * tp->mss_cache)
2246 goto send_now;
2247
2248 /* Middle in queue won't get any more data, full sendable already? */
2249 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
2250 goto send_now;
2251
2252 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
2253 if (win_divisor) {
2254 u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache);
2255
2256 /* If at least some fraction of a window is available,
2257 * just use it.
2258 */
2259 chunk /= win_divisor;
2260 if (limit >= chunk)
2261 goto send_now;
2262 } else {
2263 /* Different approach, try not to defer past a single
2264 * ACK. Receiver should ACK every other full sized
2265 * frame, so if we have space for more than 3 frames
2266 * then send now.
2267 */
2268 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
2269 goto send_now;
2270 }
2271
2272 /* TODO : use tsorted_sent_queue ? */
2273 head = tcp_rtx_queue_head(sk);
2274 if (!head)
2275 goto send_now;
2276 delta = tp->tcp_clock_cache - head->tstamp;
2277 /* If next ACK is likely to come too late (half srtt), do not defer */
2278 if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
2279 goto send_now;
2280
2281 /* Ok, it looks like it is advisable to defer.
2282 * Three cases are tracked :
2283 * 1) We are cwnd-limited
2284 * 2) We are rwnd-limited
2285 * 3) We are application limited.
2286 */
2287 if (cong_win < send_win) {
2288 if (cong_win <= skb->len) {
2289 *is_cwnd_limited = true;
2290 return true;
2291 }
2292 } else {
2293 if (send_win <= skb->len) {
2294 *is_rwnd_limited = true;
2295 return true;
2296 }
2297 }
2298
2299 /* If this packet won't get more data, do not wait. */
2300 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2301 TCP_SKB_CB(skb)->eor)
2302 goto send_now;
2303
2304 return true;
2305
2306send_now:
2307 return false;
2308}
2309
2310static inline void tcp_mtu_check_reprobe(struct sock *sk)
2311{
2312 struct inet_connection_sock *icsk = inet_csk(sk);
2313 struct tcp_sock *tp = tcp_sk(sk);
2314 struct net *net = sock_net(sk);
2315 u32 interval;
2316 s32 delta;
2317
2318 interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval);
2319 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2320 if (unlikely(delta >= interval * HZ)) {
2321 int mss = tcp_current_mss(sk);
2322
2323 /* Update current search range */
2324 icsk->icsk_mtup.probe_size = 0;
2325 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2326 sizeof(struct tcphdr) +
2327 icsk->icsk_af_ops->net_header_len;
2328 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2329
2330 /* Update probe time stamp */
2331 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2332 }
2333}
2334
2335static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2336{
2337 struct sk_buff *skb, *next;
2338
2339 skb = tcp_send_head(sk);
2340 tcp_for_write_queue_from_safe(skb, next, sk) {
2341 if (len <= skb->len)
2342 break;
2343
2344 if (unlikely(TCP_SKB_CB(skb)->eor) ||
2345 tcp_has_tx_tstamp(skb) ||
2346 !skb_pure_zcopy_same(skb, next))
2347 return false;
2348
2349 len -= skb->len;
2350 }
2351
2352 return true;
2353}
2354
2355static int tcp_clone_payload(struct sock *sk, struct sk_buff *to,
2356 int probe_size)
2357{
2358 skb_frag_t *lastfrag = NULL, *fragto = skb_shinfo(to)->frags;
2359 int i, todo, len = 0, nr_frags = 0;
2360 const struct sk_buff *skb;
2361
2362 if (!sk_wmem_schedule(sk, to->truesize + probe_size))
2363 return -ENOMEM;
2364
2365 skb_queue_walk(&sk->sk_write_queue, skb) {
2366 const skb_frag_t *fragfrom = skb_shinfo(skb)->frags;
2367
2368 if (skb_headlen(skb))
2369 return -EINVAL;
2370
2371 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, fragfrom++) {
2372 if (len >= probe_size)
2373 goto commit;
2374 todo = min_t(int, skb_frag_size(fragfrom),
2375 probe_size - len);
2376 len += todo;
2377 if (lastfrag &&
2378 skb_frag_page(fragfrom) == skb_frag_page(lastfrag) &&
2379 skb_frag_off(fragfrom) == skb_frag_off(lastfrag) +
2380 skb_frag_size(lastfrag)) {
2381 skb_frag_size_add(lastfrag, todo);
2382 continue;
2383 }
2384 if (unlikely(nr_frags == MAX_SKB_FRAGS))
2385 return -E2BIG;
2386 skb_frag_page_copy(fragto, fragfrom);
2387 skb_frag_off_copy(fragto, fragfrom);
2388 skb_frag_size_set(fragto, todo);
2389 nr_frags++;
2390 lastfrag = fragto++;
2391 }
2392 }
2393commit:
2394 WARN_ON_ONCE(len != probe_size);
2395 for (i = 0; i < nr_frags; i++)
2396 skb_frag_ref(to, i);
2397
2398 skb_shinfo(to)->nr_frags = nr_frags;
2399 to->truesize += probe_size;
2400 to->len += probe_size;
2401 to->data_len += probe_size;
2402 __skb_header_release(to);
2403 return 0;
2404}
2405
2406/* Create a new MTU probe if we are ready.
2407 * MTU probe is regularly attempting to increase the path MTU by
2408 * deliberately sending larger packets. This discovers routing
2409 * changes resulting in larger path MTUs.
2410 *
2411 * Returns 0 if we should wait to probe (no cwnd available),
2412 * 1 if a probe was sent,
2413 * -1 otherwise
2414 */
2415static int tcp_mtu_probe(struct sock *sk)
2416{
2417 struct inet_connection_sock *icsk = inet_csk(sk);
2418 struct tcp_sock *tp = tcp_sk(sk);
2419 struct sk_buff *skb, *nskb, *next;
2420 struct net *net = sock_net(sk);
2421 int probe_size;
2422 int size_needed;
2423 int copy, len;
2424 int mss_now;
2425 int interval;
2426
2427 /* Not currently probing/verifying,
2428 * not in recovery,
2429 * have enough cwnd, and
2430 * not SACKing (the variable headers throw things off)
2431 */
2432 if (likely(!icsk->icsk_mtup.enabled ||
2433 icsk->icsk_mtup.probe_size ||
2434 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2435 tcp_snd_cwnd(tp) < 11 ||
2436 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2437 return -1;
2438
2439 /* Use binary search for probe_size between tcp_mss_base,
2440 * and current mss_clamp. if (search_high - search_low)
2441 * smaller than a threshold, backoff from probing.
2442 */
2443 mss_now = tcp_current_mss(sk);
2444 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2445 icsk->icsk_mtup.search_low) >> 1);
2446 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2447 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2448 /* When misfortune happens, we are reprobing actively,
2449 * and then reprobe timer has expired. We stick with current
2450 * probing process by not resetting search range to its orignal.
2451 */
2452 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2453 interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) {
2454 /* Check whether enough time has elaplased for
2455 * another round of probing.
2456 */
2457 tcp_mtu_check_reprobe(sk);
2458 return -1;
2459 }
2460
2461 /* Have enough data in the send queue to probe? */
2462 if (tp->write_seq - tp->snd_nxt < size_needed)
2463 return -1;
2464
2465 if (tp->snd_wnd < size_needed)
2466 return -1;
2467 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2468 return 0;
2469
2470 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2471 if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) {
2472 if (!tcp_packets_in_flight(tp))
2473 return -1;
2474 else
2475 return 0;
2476 }
2477
2478 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2479 return -1;
2480
2481 /* We're allowed to probe. Build it now. */
2482 nskb = tcp_stream_alloc_skb(sk, GFP_ATOMIC, false);
2483 if (!nskb)
2484 return -1;
2485
2486 /* build the payload, and be prepared to abort if this fails. */
2487 if (tcp_clone_payload(sk, nskb, probe_size)) {
2488 tcp_skb_tsorted_anchor_cleanup(nskb);
2489 consume_skb(nskb);
2490 return -1;
2491 }
2492 sk_wmem_queued_add(sk, nskb->truesize);
2493 sk_mem_charge(sk, nskb->truesize);
2494
2495 skb = tcp_send_head(sk);
2496 skb_copy_decrypted(nskb, skb);
2497 mptcp_skb_ext_copy(nskb, skb);
2498
2499 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2500 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2501 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2502
2503 tcp_insert_write_queue_before(nskb, skb, sk);
2504 tcp_highest_sack_replace(sk, skb, nskb);
2505
2506 len = 0;
2507 tcp_for_write_queue_from_safe(skb, next, sk) {
2508 copy = min_t(int, skb->len, probe_size - len);
2509
2510 if (skb->len <= copy) {
2511 /* We've eaten all the data from this skb.
2512 * Throw it away. */
2513 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2514 /* If this is the last SKB we copy and eor is set
2515 * we need to propagate it to the new skb.
2516 */
2517 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2518 tcp_skb_collapse_tstamp(nskb, skb);
2519 tcp_unlink_write_queue(skb, sk);
2520 tcp_wmem_free_skb(sk, skb);
2521 } else {
2522 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2523 ~(TCPHDR_FIN|TCPHDR_PSH);
2524 __pskb_trim_head(skb, copy);
2525 tcp_set_skb_tso_segs(skb, mss_now);
2526 TCP_SKB_CB(skb)->seq += copy;
2527 }
2528
2529 len += copy;
2530
2531 if (len >= probe_size)
2532 break;
2533 }
2534 tcp_init_tso_segs(nskb, nskb->len);
2535
2536 /* We're ready to send. If this fails, the probe will
2537 * be resegmented into mss-sized pieces by tcp_write_xmit().
2538 */
2539 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2540 /* Decrement cwnd here because we are sending
2541 * effectively two packets. */
2542 tcp_snd_cwnd_set(tp, tcp_snd_cwnd(tp) - 1);
2543 tcp_event_new_data_sent(sk, nskb);
2544
2545 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2546 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2547 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2548
2549 return 1;
2550 }
2551
2552 return -1;
2553}
2554
2555static bool tcp_pacing_check(struct sock *sk)
2556{
2557 struct tcp_sock *tp = tcp_sk(sk);
2558
2559 if (!tcp_needs_internal_pacing(sk))
2560 return false;
2561
2562 if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2563 return false;
2564
2565 if (!hrtimer_is_queued(&tp->pacing_timer)) {
2566 hrtimer_start(&tp->pacing_timer,
2567 ns_to_ktime(tp->tcp_wstamp_ns),
2568 HRTIMER_MODE_ABS_PINNED_SOFT);
2569 sock_hold(sk);
2570 }
2571 return true;
2572}
2573
2574static bool tcp_rtx_queue_empty_or_single_skb(const struct sock *sk)
2575{
2576 const struct rb_node *node = sk->tcp_rtx_queue.rb_node;
2577
2578 /* No skb in the rtx queue. */
2579 if (!node)
2580 return true;
2581
2582 /* Only one skb in rtx queue. */
2583 return !node->rb_left && !node->rb_right;
2584}
2585
2586/* TCP Small Queues :
2587 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2588 * (These limits are doubled for retransmits)
2589 * This allows for :
2590 * - better RTT estimation and ACK scheduling
2591 * - faster recovery
2592 * - high rates
2593 * Alas, some drivers / subsystems require a fair amount
2594 * of queued bytes to ensure line rate.
2595 * One example is wifi aggregation (802.11 AMPDU)
2596 */
2597static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2598 unsigned int factor)
2599{
2600 unsigned long limit;
2601
2602 limit = max_t(unsigned long,
2603 2 * skb->truesize,
2604 READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift));
2605 if (sk->sk_pacing_status == SK_PACING_NONE)
2606 limit = min_t(unsigned long, limit,
2607 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes));
2608 limit <<= factor;
2609
2610 if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2611 tcp_sk(sk)->tcp_tx_delay) {
2612 u64 extra_bytes = (u64)READ_ONCE(sk->sk_pacing_rate) *
2613 tcp_sk(sk)->tcp_tx_delay;
2614
2615 /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2616 * approximate our needs assuming an ~100% skb->truesize overhead.
2617 * USEC_PER_SEC is approximated by 2^20.
2618 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2619 */
2620 extra_bytes >>= (20 - 1);
2621 limit += extra_bytes;
2622 }
2623 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2624 /* Always send skb if rtx queue is empty or has one skb.
2625 * No need to wait for TX completion to call us back,
2626 * after softirq/tasklet schedule.
2627 * This helps when TX completions are delayed too much.
2628 */
2629 if (tcp_rtx_queue_empty_or_single_skb(sk))
2630 return false;
2631
2632 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2633 /* It is possible TX completion already happened
2634 * before we set TSQ_THROTTLED, so we must
2635 * test again the condition.
2636 */
2637 smp_mb__after_atomic();
2638 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2639 return true;
2640 }
2641 return false;
2642}
2643
2644static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2645{
2646 const u32 now = tcp_jiffies32;
2647 enum tcp_chrono old = tp->chrono_type;
2648
2649 if (old > TCP_CHRONO_UNSPEC)
2650 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2651 tp->chrono_start = now;
2652 tp->chrono_type = new;
2653}
2654
2655void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2656{
2657 struct tcp_sock *tp = tcp_sk(sk);
2658
2659 /* If there are multiple conditions worthy of tracking in a
2660 * chronograph then the highest priority enum takes precedence
2661 * over the other conditions. So that if something "more interesting"
2662 * starts happening, stop the previous chrono and start a new one.
2663 */
2664 if (type > tp->chrono_type)
2665 tcp_chrono_set(tp, type);
2666}
2667
2668void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2669{
2670 struct tcp_sock *tp = tcp_sk(sk);
2671
2672
2673 /* There are multiple conditions worthy of tracking in a
2674 * chronograph, so that the highest priority enum takes
2675 * precedence over the other conditions (see tcp_chrono_start).
2676 * If a condition stops, we only stop chrono tracking if
2677 * it's the "most interesting" or current chrono we are
2678 * tracking and starts busy chrono if we have pending data.
2679 */
2680 if (tcp_rtx_and_write_queues_empty(sk))
2681 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2682 else if (type == tp->chrono_type)
2683 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2684}
2685
2686/* This routine writes packets to the network. It advances the
2687 * send_head. This happens as incoming acks open up the remote
2688 * window for us.
2689 *
2690 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2691 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2692 * account rare use of URG, this is not a big flaw.
2693 *
2694 * Send at most one packet when push_one > 0. Temporarily ignore
2695 * cwnd limit to force at most one packet out when push_one == 2.
2696
2697 * Returns true, if no segments are in flight and we have queued segments,
2698 * but cannot send anything now because of SWS or another problem.
2699 */
2700static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2701 int push_one, gfp_t gfp)
2702{
2703 struct tcp_sock *tp = tcp_sk(sk);
2704 struct sk_buff *skb;
2705 unsigned int tso_segs, sent_pkts;
2706 int cwnd_quota;
2707 int result;
2708 bool is_cwnd_limited = false, is_rwnd_limited = false;
2709 u32 max_segs;
2710
2711 sent_pkts = 0;
2712
2713 tcp_mstamp_refresh(tp);
2714 if (!push_one) {
2715 /* Do MTU probing. */
2716 result = tcp_mtu_probe(sk);
2717 if (!result) {
2718 return false;
2719 } else if (result > 0) {
2720 sent_pkts = 1;
2721 }
2722 }
2723
2724 max_segs = tcp_tso_segs(sk, mss_now);
2725 while ((skb = tcp_send_head(sk))) {
2726 unsigned int limit;
2727
2728 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2729 /* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2730 tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2731 skb_set_delivery_time(skb, tp->tcp_wstamp_ns, true);
2732 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
2733 tcp_init_tso_segs(skb, mss_now);
2734 goto repair; /* Skip network transmission */
2735 }
2736
2737 if (tcp_pacing_check(sk))
2738 break;
2739
2740 tso_segs = tcp_init_tso_segs(skb, mss_now);
2741 BUG_ON(!tso_segs);
2742
2743 cwnd_quota = tcp_cwnd_test(tp, skb);
2744 if (!cwnd_quota) {
2745 if (push_one == 2)
2746 /* Force out a loss probe pkt. */
2747 cwnd_quota = 1;
2748 else
2749 break;
2750 }
2751
2752 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2753 is_rwnd_limited = true;
2754 break;
2755 }
2756
2757 if (tso_segs == 1) {
2758 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2759 (tcp_skb_is_last(sk, skb) ?
2760 nonagle : TCP_NAGLE_PUSH))))
2761 break;
2762 } else {
2763 if (!push_one &&
2764 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2765 &is_rwnd_limited, max_segs))
2766 break;
2767 }
2768
2769 limit = mss_now;
2770 if (tso_segs > 1 && !tcp_urg_mode(tp))
2771 limit = tcp_mss_split_point(sk, skb, mss_now,
2772 min_t(unsigned int,
2773 cwnd_quota,
2774 max_segs),
2775 nonagle);
2776
2777 if (skb->len > limit &&
2778 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2779 break;
2780
2781 if (tcp_small_queue_check(sk, skb, 0))
2782 break;
2783
2784 /* Argh, we hit an empty skb(), presumably a thread
2785 * is sleeping in sendmsg()/sk_stream_wait_memory().
2786 * We do not want to send a pure-ack packet and have
2787 * a strange looking rtx queue with empty packet(s).
2788 */
2789 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2790 break;
2791
2792 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2793 break;
2794
2795repair:
2796 /* Advance the send_head. This one is sent out.
2797 * This call will increment packets_out.
2798 */
2799 tcp_event_new_data_sent(sk, skb);
2800
2801 tcp_minshall_update(tp, mss_now, skb);
2802 sent_pkts += tcp_skb_pcount(skb);
2803
2804 if (push_one)
2805 break;
2806 }
2807
2808 if (is_rwnd_limited)
2809 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2810 else
2811 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2812
2813 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp));
2814 if (likely(sent_pkts || is_cwnd_limited))
2815 tcp_cwnd_validate(sk, is_cwnd_limited);
2816
2817 if (likely(sent_pkts)) {
2818 if (tcp_in_cwnd_reduction(sk))
2819 tp->prr_out += sent_pkts;
2820
2821 /* Send one loss probe per tail loss episode. */
2822 if (push_one != 2)
2823 tcp_schedule_loss_probe(sk, false);
2824 return false;
2825 }
2826 return !tp->packets_out && !tcp_write_queue_empty(sk);
2827}
2828
2829bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2830{
2831 struct inet_connection_sock *icsk = inet_csk(sk);
2832 struct tcp_sock *tp = tcp_sk(sk);
2833 u32 timeout, timeout_us, rto_delta_us;
2834 int early_retrans;
2835
2836 /* Don't do any loss probe on a Fast Open connection before 3WHS
2837 * finishes.
2838 */
2839 if (rcu_access_pointer(tp->fastopen_rsk))
2840 return false;
2841
2842 early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans);
2843 /* Schedule a loss probe in 2*RTT for SACK capable connections
2844 * not in loss recovery, that are either limited by cwnd or application.
2845 */
2846 if ((early_retrans != 3 && early_retrans != 4) ||
2847 !tp->packets_out || !tcp_is_sack(tp) ||
2848 (icsk->icsk_ca_state != TCP_CA_Open &&
2849 icsk->icsk_ca_state != TCP_CA_CWR))
2850 return false;
2851
2852 /* Probe timeout is 2*rtt. Add minimum RTO to account
2853 * for delayed ack when there's one outstanding packet. If no RTT
2854 * sample is available then probe after TCP_TIMEOUT_INIT.
2855 */
2856 if (tp->srtt_us) {
2857 timeout_us = tp->srtt_us >> 2;
2858 if (tp->packets_out == 1)
2859 timeout_us += tcp_rto_min_us(sk);
2860 else
2861 timeout_us += TCP_TIMEOUT_MIN_US;
2862 timeout = usecs_to_jiffies(timeout_us);
2863 } else {
2864 timeout = TCP_TIMEOUT_INIT;
2865 }
2866
2867 /* If the RTO formula yields an earlier time, then use that time. */
2868 rto_delta_us = advancing_rto ?
2869 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2870 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2871 if (rto_delta_us > 0)
2872 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2873
2874 tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout, TCP_RTO_MAX);
2875 return true;
2876}
2877
2878/* Thanks to skb fast clones, we can detect if a prior transmit of
2879 * a packet is still in a qdisc or driver queue.
2880 * In this case, there is very little point doing a retransmit !
2881 */
2882static bool skb_still_in_host_queue(struct sock *sk,
2883 const struct sk_buff *skb)
2884{
2885 if (unlikely(skb_fclone_busy(sk, skb))) {
2886 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2887 smp_mb__after_atomic();
2888 if (skb_fclone_busy(sk, skb)) {
2889 NET_INC_STATS(sock_net(sk),
2890 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2891 return true;
2892 }
2893 }
2894 return false;
2895}
2896
2897/* When probe timeout (PTO) fires, try send a new segment if possible, else
2898 * retransmit the last segment.
2899 */
2900void tcp_send_loss_probe(struct sock *sk)
2901{
2902 struct tcp_sock *tp = tcp_sk(sk);
2903 struct sk_buff *skb;
2904 int pcount;
2905 int mss = tcp_current_mss(sk);
2906
2907 /* At most one outstanding TLP */
2908 if (tp->tlp_high_seq)
2909 goto rearm_timer;
2910
2911 tp->tlp_retrans = 0;
2912 skb = tcp_send_head(sk);
2913 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2914 pcount = tp->packets_out;
2915 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2916 if (tp->packets_out > pcount)
2917 goto probe_sent;
2918 goto rearm_timer;
2919 }
2920 skb = skb_rb_last(&sk->tcp_rtx_queue);
2921 if (unlikely(!skb)) {
2922 WARN_ONCE(tp->packets_out,
2923 "invalid inflight: %u state %u cwnd %u mss %d\n",
2924 tp->packets_out, sk->sk_state, tcp_snd_cwnd(tp), mss);
2925 inet_csk(sk)->icsk_pending = 0;
2926 return;
2927 }
2928
2929 if (skb_still_in_host_queue(sk, skb))
2930 goto rearm_timer;
2931
2932 pcount = tcp_skb_pcount(skb);
2933 if (WARN_ON(!pcount))
2934 goto rearm_timer;
2935
2936 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2937 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2938 (pcount - 1) * mss, mss,
2939 GFP_ATOMIC)))
2940 goto rearm_timer;
2941 skb = skb_rb_next(skb);
2942 }
2943
2944 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2945 goto rearm_timer;
2946
2947 if (__tcp_retransmit_skb(sk, skb, 1))
2948 goto rearm_timer;
2949
2950 tp->tlp_retrans = 1;
2951
2952probe_sent:
2953 /* Record snd_nxt for loss detection. */
2954 tp->tlp_high_seq = tp->snd_nxt;
2955
2956 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2957 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2958 inet_csk(sk)->icsk_pending = 0;
2959rearm_timer:
2960 tcp_rearm_rto(sk);
2961}
2962
2963/* Push out any pending frames which were held back due to
2964 * TCP_CORK or attempt at coalescing tiny packets.
2965 * The socket must be locked by the caller.
2966 */
2967void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2968 int nonagle)
2969{
2970 /* If we are closed, the bytes will have to remain here.
2971 * In time closedown will finish, we empty the write queue and
2972 * all will be happy.
2973 */
2974 if (unlikely(sk->sk_state == TCP_CLOSE))
2975 return;
2976
2977 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2978 sk_gfp_mask(sk, GFP_ATOMIC)))
2979 tcp_check_probe_timer(sk);
2980}
2981
2982/* Send _single_ skb sitting at the send head. This function requires
2983 * true push pending frames to setup probe timer etc.
2984 */
2985void tcp_push_one(struct sock *sk, unsigned int mss_now)
2986{
2987 struct sk_buff *skb = tcp_send_head(sk);
2988
2989 BUG_ON(!skb || skb->len < mss_now);
2990
2991 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2992}
2993
2994/* This function returns the amount that we can raise the
2995 * usable window based on the following constraints
2996 *
2997 * 1. The window can never be shrunk once it is offered (RFC 793)
2998 * 2. We limit memory per socket
2999 *
3000 * RFC 1122:
3001 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
3002 * RECV.NEXT + RCV.WIN fixed until:
3003 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
3004 *
3005 * i.e. don't raise the right edge of the window until you can raise
3006 * it at least MSS bytes.
3007 *
3008 * Unfortunately, the recommended algorithm breaks header prediction,
3009 * since header prediction assumes th->window stays fixed.
3010 *
3011 * Strictly speaking, keeping th->window fixed violates the receiver
3012 * side SWS prevention criteria. The problem is that under this rule
3013 * a stream of single byte packets will cause the right side of the
3014 * window to always advance by a single byte.
3015 *
3016 * Of course, if the sender implements sender side SWS prevention
3017 * then this will not be a problem.
3018 *
3019 * BSD seems to make the following compromise:
3020 *
3021 * If the free space is less than the 1/4 of the maximum
3022 * space available and the free space is less than 1/2 mss,
3023 * then set the window to 0.
3024 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
3025 * Otherwise, just prevent the window from shrinking
3026 * and from being larger than the largest representable value.
3027 *
3028 * This prevents incremental opening of the window in the regime
3029 * where TCP is limited by the speed of the reader side taking
3030 * data out of the TCP receive queue. It does nothing about
3031 * those cases where the window is constrained on the sender side
3032 * because the pipeline is full.
3033 *
3034 * BSD also seems to "accidentally" limit itself to windows that are a
3035 * multiple of MSS, at least until the free space gets quite small.
3036 * This would appear to be a side effect of the mbuf implementation.
3037 * Combining these two algorithms results in the observed behavior
3038 * of having a fixed window size at almost all times.
3039 *
3040 * Below we obtain similar behavior by forcing the offered window to
3041 * a multiple of the mss when it is feasible to do so.
3042 *
3043 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
3044 * Regular options like TIMESTAMP are taken into account.
3045 */
3046u32 __tcp_select_window(struct sock *sk)
3047{
3048 struct inet_connection_sock *icsk = inet_csk(sk);
3049 struct tcp_sock *tp = tcp_sk(sk);
3050 struct net *net = sock_net(sk);
3051 /* MSS for the peer's data. Previous versions used mss_clamp
3052 * here. I don't know if the value based on our guesses
3053 * of peer's MSS is better for the performance. It's more correct
3054 * but may be worse for the performance because of rcv_mss
3055 * fluctuations. --SAW 1998/11/1
3056 */
3057 int mss = icsk->icsk_ack.rcv_mss;
3058 int free_space = tcp_space(sk);
3059 int allowed_space = tcp_full_space(sk);
3060 int full_space, window;
3061
3062 if (sk_is_mptcp(sk))
3063 mptcp_space(sk, &free_space, &allowed_space);
3064
3065 full_space = min_t(int, tp->window_clamp, allowed_space);
3066
3067 if (unlikely(mss > full_space)) {
3068 mss = full_space;
3069 if (mss <= 0)
3070 return 0;
3071 }
3072
3073 /* Only allow window shrink if the sysctl is enabled and we have
3074 * a non-zero scaling factor in effect.
3075 */
3076 if (READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) && tp->rx_opt.rcv_wscale)
3077 goto shrink_window_allowed;
3078
3079 /* do not allow window to shrink */
3080
3081 if (free_space < (full_space >> 1)) {
3082 icsk->icsk_ack.quick = 0;
3083
3084 if (tcp_under_memory_pressure(sk))
3085 tcp_adjust_rcv_ssthresh(sk);
3086
3087 /* free_space might become our new window, make sure we don't
3088 * increase it due to wscale.
3089 */
3090 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
3091
3092 /* if free space is less than mss estimate, or is below 1/16th
3093 * of the maximum allowed, try to move to zero-window, else
3094 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
3095 * new incoming data is dropped due to memory limits.
3096 * With large window, mss test triggers way too late in order
3097 * to announce zero window in time before rmem limit kicks in.
3098 */
3099 if (free_space < (allowed_space >> 4) || free_space < mss)
3100 return 0;
3101 }
3102
3103 if (free_space > tp->rcv_ssthresh)
3104 free_space = tp->rcv_ssthresh;
3105
3106 /* Don't do rounding if we are using window scaling, since the
3107 * scaled window will not line up with the MSS boundary anyway.
3108 */
3109 if (tp->rx_opt.rcv_wscale) {
3110 window = free_space;
3111
3112 /* Advertise enough space so that it won't get scaled away.
3113 * Import case: prevent zero window announcement if
3114 * 1<<rcv_wscale > mss.
3115 */
3116 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
3117 } else {
3118 window = tp->rcv_wnd;
3119 /* Get the largest window that is a nice multiple of mss.
3120 * Window clamp already applied above.
3121 * If our current window offering is within 1 mss of the
3122 * free space we just keep it. This prevents the divide
3123 * and multiply from happening most of the time.
3124 * We also don't do any window rounding when the free space
3125 * is too small.
3126 */
3127 if (window <= free_space - mss || window > free_space)
3128 window = rounddown(free_space, mss);
3129 else if (mss == full_space &&
3130 free_space > window + (full_space >> 1))
3131 window = free_space;
3132 }
3133
3134 return window;
3135
3136shrink_window_allowed:
3137 /* new window should always be an exact multiple of scaling factor */
3138 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
3139
3140 if (free_space < (full_space >> 1)) {
3141 icsk->icsk_ack.quick = 0;
3142
3143 if (tcp_under_memory_pressure(sk))
3144 tcp_adjust_rcv_ssthresh(sk);
3145
3146 /* if free space is too low, return a zero window */
3147 if (free_space < (allowed_space >> 4) || free_space < mss ||
3148 free_space < (1 << tp->rx_opt.rcv_wscale))
3149 return 0;
3150 }
3151
3152 if (free_space > tp->rcv_ssthresh) {
3153 free_space = tp->rcv_ssthresh;
3154 /* new window should always be an exact multiple of scaling factor
3155 *
3156 * For this case, we ALIGN "up" (increase free_space) because
3157 * we know free_space is not zero here, it has been reduced from
3158 * the memory-based limit, and rcv_ssthresh is not a hard limit
3159 * (unlike sk_rcvbuf).
3160 */
3161 free_space = ALIGN(free_space, (1 << tp->rx_opt.rcv_wscale));
3162 }
3163
3164 return free_space;
3165}
3166
3167void tcp_skb_collapse_tstamp(struct sk_buff *skb,
3168 const struct sk_buff *next_skb)
3169{
3170 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
3171 const struct skb_shared_info *next_shinfo =
3172 skb_shinfo(next_skb);
3173 struct skb_shared_info *shinfo = skb_shinfo(skb);
3174
3175 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
3176 shinfo->tskey = next_shinfo->tskey;
3177 TCP_SKB_CB(skb)->txstamp_ack |=
3178 TCP_SKB_CB(next_skb)->txstamp_ack;
3179 }
3180}
3181
3182/* Collapses two adjacent SKB's during retransmission. */
3183static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
3184{
3185 struct tcp_sock *tp = tcp_sk(sk);
3186 struct sk_buff *next_skb = skb_rb_next(skb);
3187 int next_skb_size;
3188
3189 next_skb_size = next_skb->len;
3190
3191 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
3192
3193 if (next_skb_size && !tcp_skb_shift(skb, next_skb, 1, next_skb_size))
3194 return false;
3195
3196 tcp_highest_sack_replace(sk, next_skb, skb);
3197
3198 /* Update sequence range on original skb. */
3199 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
3200
3201 /* Merge over control information. This moves PSH/FIN etc. over */
3202 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
3203
3204 /* All done, get rid of second SKB and account for it so
3205 * packet counting does not break.
3206 */
3207 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
3208 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
3209
3210 /* changed transmit queue under us so clear hints */
3211 tcp_clear_retrans_hints_partial(tp);
3212 if (next_skb == tp->retransmit_skb_hint)
3213 tp->retransmit_skb_hint = skb;
3214
3215 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
3216
3217 tcp_skb_collapse_tstamp(skb, next_skb);
3218
3219 tcp_rtx_queue_unlink_and_free(next_skb, sk);
3220 return true;
3221}
3222
3223/* Check if coalescing SKBs is legal. */
3224static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
3225{
3226 if (tcp_skb_pcount(skb) > 1)
3227 return false;
3228 if (skb_cloned(skb))
3229 return false;
3230 /* Some heuristics for collapsing over SACK'd could be invented */
3231 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3232 return false;
3233
3234 return true;
3235}
3236
3237/* Collapse packets in the retransmit queue to make to create
3238 * less packets on the wire. This is only done on retransmission.
3239 */
3240static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
3241 int space)
3242{
3243 struct tcp_sock *tp = tcp_sk(sk);
3244 struct sk_buff *skb = to, *tmp;
3245 bool first = true;
3246
3247 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse))
3248 return;
3249 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3250 return;
3251
3252 skb_rbtree_walk_from_safe(skb, tmp) {
3253 if (!tcp_can_collapse(sk, skb))
3254 break;
3255
3256 if (!tcp_skb_can_collapse(to, skb))
3257 break;
3258
3259 space -= skb->len;
3260
3261 if (first) {
3262 first = false;
3263 continue;
3264 }
3265
3266 if (space < 0)
3267 break;
3268
3269 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
3270 break;
3271
3272 if (!tcp_collapse_retrans(sk, to))
3273 break;
3274 }
3275}
3276
3277/* This retransmits one SKB. Policy decisions and retransmit queue
3278 * state updates are done by the caller. Returns non-zero if an
3279 * error occurred which prevented the send.
3280 */
3281int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3282{
3283 struct inet_connection_sock *icsk = inet_csk(sk);
3284 struct tcp_sock *tp = tcp_sk(sk);
3285 unsigned int cur_mss;
3286 int diff, len, err;
3287 int avail_wnd;
3288
3289 /* Inconclusive MTU probe */
3290 if (icsk->icsk_mtup.probe_size)
3291 icsk->icsk_mtup.probe_size = 0;
3292
3293 if (skb_still_in_host_queue(sk, skb))
3294 return -EBUSY;
3295
3296start:
3297 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
3298 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3299 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN;
3300 TCP_SKB_CB(skb)->seq++;
3301 goto start;
3302 }
3303 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
3304 WARN_ON_ONCE(1);
3305 return -EINVAL;
3306 }
3307 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
3308 return -ENOMEM;
3309 }
3310
3311 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
3312 return -EHOSTUNREACH; /* Routing failure or similar. */
3313
3314 cur_mss = tcp_current_mss(sk);
3315 avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3316
3317 /* If receiver has shrunk his window, and skb is out of
3318 * new window, do not retransmit it. The exception is the
3319 * case, when window is shrunk to zero. In this case
3320 * our retransmit of one segment serves as a zero window probe.
3321 */
3322 if (avail_wnd <= 0) {
3323 if (TCP_SKB_CB(skb)->seq != tp->snd_una)
3324 return -EAGAIN;
3325 avail_wnd = cur_mss;
3326 }
3327
3328 len = cur_mss * segs;
3329 if (len > avail_wnd) {
3330 len = rounddown(avail_wnd, cur_mss);
3331 if (!len)
3332 len = avail_wnd;
3333 }
3334 if (skb->len > len) {
3335 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
3336 cur_mss, GFP_ATOMIC))
3337 return -ENOMEM; /* We'll try again later. */
3338 } else {
3339 if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
3340 return -ENOMEM;
3341
3342 diff = tcp_skb_pcount(skb);
3343 tcp_set_skb_tso_segs(skb, cur_mss);
3344 diff -= tcp_skb_pcount(skb);
3345 if (diff)
3346 tcp_adjust_pcount(sk, skb, diff);
3347 avail_wnd = min_t(int, avail_wnd, cur_mss);
3348 if (skb->len < avail_wnd)
3349 tcp_retrans_try_collapse(sk, skb, avail_wnd);
3350 }
3351
3352 /* RFC3168, section 6.1.1.1. ECN fallback */
3353 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
3354 tcp_ecn_clear_syn(sk, skb);
3355
3356 /* Update global and local TCP statistics. */
3357 segs = tcp_skb_pcount(skb);
3358 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
3359 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3360 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3361 tp->total_retrans += segs;
3362 tp->bytes_retrans += skb->len;
3363
3364 /* make sure skb->data is aligned on arches that require it
3365 * and check if ack-trimming & collapsing extended the headroom
3366 * beyond what csum_start can cover.
3367 */
3368 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
3369 skb_headroom(skb) >= 0xFFFF)) {
3370 struct sk_buff *nskb;
3371
3372 tcp_skb_tsorted_save(skb) {
3373 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
3374 if (nskb) {
3375 nskb->dev = NULL;
3376 err = tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC);
3377 } else {
3378 err = -ENOBUFS;
3379 }
3380 } tcp_skb_tsorted_restore(skb);
3381
3382 if (!err) {
3383 tcp_update_skb_after_send(sk, skb, tp->tcp_wstamp_ns);
3384 tcp_rate_skb_sent(sk, skb);
3385 }
3386 } else {
3387 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3388 }
3389
3390 /* To avoid taking spuriously low RTT samples based on a timestamp
3391 * for a transmit that never happened, always mark EVER_RETRANS
3392 */
3393 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
3394
3395 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
3396 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
3397 TCP_SKB_CB(skb)->seq, segs, err);
3398
3399 if (likely(!err)) {
3400 trace_tcp_retransmit_skb(sk, skb);
3401 } else if (err != -EBUSY) {
3402 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
3403 }
3404 return err;
3405}
3406
3407int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3408{
3409 struct tcp_sock *tp = tcp_sk(sk);
3410 int err = __tcp_retransmit_skb(sk, skb, segs);
3411
3412 if (err == 0) {
3413#if FASTRETRANS_DEBUG > 0
3414 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3415 net_dbg_ratelimited("retrans_out leaked\n");
3416 }
3417#endif
3418 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3419 tp->retrans_out += tcp_skb_pcount(skb);
3420 }
3421
3422 /* Save stamp of the first (attempted) retransmit. */
3423 if (!tp->retrans_stamp)
3424 tp->retrans_stamp = tcp_skb_timestamp_ts(tp->tcp_usec_ts, skb);
3425
3426 if (tp->undo_retrans < 0)
3427 tp->undo_retrans = 0;
3428 tp->undo_retrans += tcp_skb_pcount(skb);
3429 return err;
3430}
3431
3432/* This gets called after a retransmit timeout, and the initially
3433 * retransmitted data is acknowledged. It tries to continue
3434 * resending the rest of the retransmit queue, until either
3435 * we've sent it all or the congestion window limit is reached.
3436 */
3437void tcp_xmit_retransmit_queue(struct sock *sk)
3438{
3439 const struct inet_connection_sock *icsk = inet_csk(sk);
3440 struct sk_buff *skb, *rtx_head, *hole = NULL;
3441 struct tcp_sock *tp = tcp_sk(sk);
3442 bool rearm_timer = false;
3443 u32 max_segs;
3444 int mib_idx;
3445
3446 if (!tp->packets_out)
3447 return;
3448
3449 rtx_head = tcp_rtx_queue_head(sk);
3450 skb = tp->retransmit_skb_hint ?: rtx_head;
3451 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
3452 skb_rbtree_walk_from(skb) {
3453 __u8 sacked;
3454 int segs;
3455
3456 if (tcp_pacing_check(sk))
3457 break;
3458
3459 /* we could do better than to assign each time */
3460 if (!hole)
3461 tp->retransmit_skb_hint = skb;
3462
3463 segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp);
3464 if (segs <= 0)
3465 break;
3466 sacked = TCP_SKB_CB(skb)->sacked;
3467 /* In case tcp_shift_skb_data() have aggregated large skbs,
3468 * we need to make sure not sending too bigs TSO packets
3469 */
3470 segs = min_t(int, segs, max_segs);
3471
3472 if (tp->retrans_out >= tp->lost_out) {
3473 break;
3474 } else if (!(sacked & TCPCB_LOST)) {
3475 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3476 hole = skb;
3477 continue;
3478
3479 } else {
3480 if (icsk->icsk_ca_state != TCP_CA_Loss)
3481 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3482 else
3483 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3484 }
3485
3486 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3487 continue;
3488
3489 if (tcp_small_queue_check(sk, skb, 1))
3490 break;
3491
3492 if (tcp_retransmit_skb(sk, skb, segs))
3493 break;
3494
3495 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3496
3497 if (tcp_in_cwnd_reduction(sk))
3498 tp->prr_out += tcp_skb_pcount(skb);
3499
3500 if (skb == rtx_head &&
3501 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3502 rearm_timer = true;
3503
3504 }
3505 if (rearm_timer)
3506 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3507 inet_csk(sk)->icsk_rto,
3508 TCP_RTO_MAX);
3509}
3510
3511/* We allow to exceed memory limits for FIN packets to expedite
3512 * connection tear down and (memory) recovery.
3513 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3514 * or even be forced to close flow without any FIN.
3515 * In general, we want to allow one skb per socket to avoid hangs
3516 * with edge trigger epoll()
3517 */
3518void sk_forced_mem_schedule(struct sock *sk, int size)
3519{
3520 int delta, amt;
3521
3522 delta = size - sk->sk_forward_alloc;
3523 if (delta <= 0)
3524 return;
3525 amt = sk_mem_pages(delta);
3526 sk_forward_alloc_add(sk, amt << PAGE_SHIFT);
3527 sk_memory_allocated_add(sk, amt);
3528
3529 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3530 mem_cgroup_charge_skmem(sk->sk_memcg, amt,
3531 gfp_memcg_charge() | __GFP_NOFAIL);
3532}
3533
3534/* Send a FIN. The caller locks the socket for us.
3535 * We should try to send a FIN packet really hard, but eventually give up.
3536 */
3537void tcp_send_fin(struct sock *sk)
3538{
3539 struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk);
3540 struct tcp_sock *tp = tcp_sk(sk);
3541
3542 /* Optimization, tack on the FIN if we have one skb in write queue and
3543 * this skb was not yet sent, or we are under memory pressure.
3544 * Note: in the latter case, FIN packet will be sent after a timeout,
3545 * as TCP stack thinks it has already been transmitted.
3546 */
3547 tskb = tail;
3548 if (!tskb && tcp_under_memory_pressure(sk))
3549 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3550
3551 if (tskb) {
3552 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3553 TCP_SKB_CB(tskb)->end_seq++;
3554 tp->write_seq++;
3555 if (!tail) {
3556 /* This means tskb was already sent.
3557 * Pretend we included the FIN on previous transmit.
3558 * We need to set tp->snd_nxt to the value it would have
3559 * if FIN had been sent. This is because retransmit path
3560 * does not change tp->snd_nxt.
3561 */
3562 WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3563 return;
3564 }
3565 } else {
3566 skb = alloc_skb_fclone(MAX_TCP_HEADER,
3567 sk_gfp_mask(sk, GFP_ATOMIC |
3568 __GFP_NOWARN));
3569 if (unlikely(!skb))
3570 return;
3571
3572 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3573 skb_reserve(skb, MAX_TCP_HEADER);
3574 sk_forced_mem_schedule(sk, skb->truesize);
3575 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3576 tcp_init_nondata_skb(skb, tp->write_seq,
3577 TCPHDR_ACK | TCPHDR_FIN);
3578 tcp_queue_skb(sk, skb);
3579 }
3580 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3581}
3582
3583/* We get here when a process closes a file descriptor (either due to
3584 * an explicit close() or as a byproduct of exit()'ing) and there
3585 * was unread data in the receive queue. This behavior is recommended
3586 * by RFC 2525, section 2.17. -DaveM
3587 */
3588void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3589{
3590 struct sk_buff *skb;
3591
3592 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3593
3594 /* NOTE: No TCP options attached and we never retransmit this. */
3595 skb = alloc_skb(MAX_TCP_HEADER, priority);
3596 if (!skb) {
3597 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3598 return;
3599 }
3600
3601 /* Reserve space for headers and prepare control bits. */
3602 skb_reserve(skb, MAX_TCP_HEADER);
3603 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3604 TCPHDR_ACK | TCPHDR_RST);
3605 tcp_mstamp_refresh(tcp_sk(sk));
3606 /* Send it off. */
3607 if (tcp_transmit_skb(sk, skb, 0, priority))
3608 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3609
3610 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3611 * skb here is different to the troublesome skb, so use NULL
3612 */
3613 trace_tcp_send_reset(sk, NULL);
3614}
3615
3616/* Send a crossed SYN-ACK during socket establishment.
3617 * WARNING: This routine must only be called when we have already sent
3618 * a SYN packet that crossed the incoming SYN that caused this routine
3619 * to get called. If this assumption fails then the initial rcv_wnd
3620 * and rcv_wscale values will not be correct.
3621 */
3622int tcp_send_synack(struct sock *sk)
3623{
3624 struct sk_buff *skb;
3625
3626 skb = tcp_rtx_queue_head(sk);
3627 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3628 pr_err("%s: wrong queue state\n", __func__);
3629 return -EFAULT;
3630 }
3631 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3632 if (skb_cloned(skb)) {
3633 struct sk_buff *nskb;
3634
3635 tcp_skb_tsorted_save(skb) {
3636 nskb = skb_copy(skb, GFP_ATOMIC);
3637 } tcp_skb_tsorted_restore(skb);
3638 if (!nskb)
3639 return -ENOMEM;
3640 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3641 tcp_highest_sack_replace(sk, skb, nskb);
3642 tcp_rtx_queue_unlink_and_free(skb, sk);
3643 __skb_header_release(nskb);
3644 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3645 sk_wmem_queued_add(sk, nskb->truesize);
3646 sk_mem_charge(sk, nskb->truesize);
3647 skb = nskb;
3648 }
3649
3650 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3651 tcp_ecn_send_synack(sk, skb);
3652 }
3653 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3654}
3655
3656/**
3657 * tcp_make_synack - Allocate one skb and build a SYNACK packet.
3658 * @sk: listener socket
3659 * @dst: dst entry attached to the SYNACK. It is consumed and caller
3660 * should not use it again.
3661 * @req: request_sock pointer
3662 * @foc: cookie for tcp fast open
3663 * @synack_type: Type of synack to prepare
3664 * @syn_skb: SYN packet just received. It could be NULL for rtx case.
3665 */
3666struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3667 struct request_sock *req,
3668 struct tcp_fastopen_cookie *foc,
3669 enum tcp_synack_type synack_type,
3670 struct sk_buff *syn_skb)
3671{
3672 struct inet_request_sock *ireq = inet_rsk(req);
3673 const struct tcp_sock *tp = tcp_sk(sk);
3674 struct tcp_out_options opts;
3675 struct tcp_key key = {};
3676 struct sk_buff *skb;
3677 int tcp_header_size;
3678 struct tcphdr *th;
3679 int mss;
3680 u64 now;
3681
3682 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3683 if (unlikely(!skb)) {
3684 dst_release(dst);
3685 return NULL;
3686 }
3687 /* Reserve space for headers. */
3688 skb_reserve(skb, MAX_TCP_HEADER);
3689
3690 switch (synack_type) {
3691 case TCP_SYNACK_NORMAL:
3692 skb_set_owner_w(skb, req_to_sk(req));
3693 break;
3694 case TCP_SYNACK_COOKIE:
3695 /* Under synflood, we do not attach skb to a socket,
3696 * to avoid false sharing.
3697 */
3698 break;
3699 case TCP_SYNACK_FASTOPEN:
3700 /* sk is a const pointer, because we want to express multiple
3701 * cpu might call us concurrently.
3702 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3703 */
3704 skb_set_owner_w(skb, (struct sock *)sk);
3705 break;
3706 }
3707 skb_dst_set(skb, dst);
3708
3709 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3710
3711 memset(&opts, 0, sizeof(opts));
3712 now = tcp_clock_ns();
3713#ifdef CONFIG_SYN_COOKIES
3714 if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok))
3715 skb_set_delivery_time(skb, cookie_init_timestamp(req, now),
3716 true);
3717 else
3718#endif
3719 {
3720 skb_set_delivery_time(skb, now, true);
3721 if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3722 tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3723 }
3724
3725#if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
3726 rcu_read_lock();
3727#endif
3728 if (tcp_rsk_used_ao(req)) {
3729#ifdef CONFIG_TCP_AO
3730 struct tcp_ao_key *ao_key = NULL;
3731 u8 keyid = tcp_rsk(req)->ao_keyid;
3732
3733 ao_key = tcp_sk(sk)->af_specific->ao_lookup(sk, req_to_sk(req),
3734 keyid, -1);
3735 /* If there is no matching key - avoid sending anything,
3736 * especially usigned segments. It could try harder and lookup
3737 * for another peer-matching key, but the peer has requested
3738 * ao_keyid (RFC5925 RNextKeyID), so let's keep it simple here.
3739 */
3740 if (unlikely(!ao_key)) {
3741 rcu_read_unlock();
3742 kfree_skb(skb);
3743 net_warn_ratelimited("TCP-AO: the keyid %u from SYN packet is not present - not sending SYNACK\n",
3744 keyid);
3745 return NULL;
3746 }
3747 key.ao_key = ao_key;
3748 key.type = TCP_KEY_AO;
3749#endif
3750 } else {
3751#ifdef CONFIG_TCP_MD5SIG
3752 key.md5_key = tcp_rsk(req)->af_specific->req_md5_lookup(sk,
3753 req_to_sk(req));
3754 if (key.md5_key)
3755 key.type = TCP_KEY_MD5;
3756#endif
3757 }
3758 skb_set_hash(skb, READ_ONCE(tcp_rsk(req)->txhash), PKT_HASH_TYPE_L4);
3759 /* bpf program will be interested in the tcp_flags */
3760 TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK;
3761 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts,
3762 &key, foc, synack_type, syn_skb)
3763 + sizeof(*th);
3764
3765 skb_push(skb, tcp_header_size);
3766 skb_reset_transport_header(skb);
3767
3768 th = (struct tcphdr *)skb->data;
3769 memset(th, 0, sizeof(struct tcphdr));
3770 th->syn = 1;
3771 th->ack = 1;
3772 tcp_ecn_make_synack(req, th);
3773 th->source = htons(ireq->ir_num);
3774 th->dest = ireq->ir_rmt_port;
3775 skb->mark = ireq->ir_mark;
3776 skb->ip_summed = CHECKSUM_PARTIAL;
3777 th->seq = htonl(tcp_rsk(req)->snt_isn);
3778 /* XXX data is queued and acked as is. No buffer/window check */
3779 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3780
3781 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3782 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3783 tcp_options_write(th, NULL, tcp_rsk(req), &opts, &key);
3784 th->doff = (tcp_header_size >> 2);
3785 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3786
3787 /* Okay, we have all we need - do the md5 hash if needed */
3788 if (tcp_key_is_md5(&key)) {
3789#ifdef CONFIG_TCP_MD5SIG
3790 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
3791 key.md5_key, req_to_sk(req), skb);
3792#endif
3793 } else if (tcp_key_is_ao(&key)) {
3794#ifdef CONFIG_TCP_AO
3795 tcp_rsk(req)->af_specific->ao_synack_hash(opts.hash_location,
3796 key.ao_key, req, skb,
3797 opts.hash_location - (u8 *)th, 0);
3798#endif
3799 }
3800#if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
3801 rcu_read_unlock();
3802#endif
3803
3804 bpf_skops_write_hdr_opt((struct sock *)sk, skb, req, syn_skb,
3805 synack_type, &opts);
3806
3807 skb_set_delivery_time(skb, now, true);
3808 tcp_add_tx_delay(skb, tp);
3809
3810 return skb;
3811}
3812EXPORT_SYMBOL(tcp_make_synack);
3813
3814static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
3815{
3816 struct inet_connection_sock *icsk = inet_csk(sk);
3817 const struct tcp_congestion_ops *ca;
3818 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
3819
3820 if (ca_key == TCP_CA_UNSPEC)
3821 return;
3822
3823 rcu_read_lock();
3824 ca = tcp_ca_find_key(ca_key);
3825 if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
3826 bpf_module_put(icsk->icsk_ca_ops, icsk->icsk_ca_ops->owner);
3827 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
3828 icsk->icsk_ca_ops = ca;
3829 }
3830 rcu_read_unlock();
3831}
3832
3833/* Do all connect socket setups that can be done AF independent. */
3834static void tcp_connect_init(struct sock *sk)
3835{
3836 const struct dst_entry *dst = __sk_dst_get(sk);
3837 struct tcp_sock *tp = tcp_sk(sk);
3838 __u8 rcv_wscale;
3839 u32 rcv_wnd;
3840
3841 /* We'll fix this up when we get a response from the other end.
3842 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
3843 */
3844 tp->tcp_header_len = sizeof(struct tcphdr);
3845 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps))
3846 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
3847
3848 tcp_ao_connect_init(sk);
3849
3850 /* If user gave his TCP_MAXSEG, record it to clamp */
3851 if (tp->rx_opt.user_mss)
3852 tp->rx_opt.mss_clamp = tp->rx_opt.user_mss;
3853 tp->max_window = 0;
3854 tcp_mtup_init(sk);
3855 tcp_sync_mss(sk, dst_mtu(dst));
3856
3857 tcp_ca_dst_init(sk, dst);
3858
3859 if (!tp->window_clamp)
3860 tp->window_clamp = dst_metric(dst, RTAX_WINDOW);
3861 tp->advmss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3862
3863 tcp_initialize_rcv_mss(sk);
3864
3865 /* limit the window selection if the user enforce a smaller rx buffer */
3866 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
3867 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
3868 tp->window_clamp = tcp_full_space(sk);
3869
3870 rcv_wnd = tcp_rwnd_init_bpf(sk);
3871 if (rcv_wnd == 0)
3872 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
3873
3874 tcp_select_initial_window(sk, tcp_full_space(sk),
3875 tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
3876 &tp->rcv_wnd,
3877 &tp->window_clamp,
3878 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling),
3879 &rcv_wscale,
3880 rcv_wnd);
3881
3882 tp->rx_opt.rcv_wscale = rcv_wscale;
3883 tp->rcv_ssthresh = tp->rcv_wnd;
3884
3885 WRITE_ONCE(sk->sk_err, 0);
3886 sock_reset_flag(sk, SOCK_DONE);
3887 tp->snd_wnd = 0;
3888 tcp_init_wl(tp, 0);
3889 tcp_write_queue_purge(sk);
3890 tp->snd_una = tp->write_seq;
3891 tp->snd_sml = tp->write_seq;
3892 tp->snd_up = tp->write_seq;
3893 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
3894
3895 if (likely(!tp->repair))
3896 tp->rcv_nxt = 0;
3897 else
3898 tp->rcv_tstamp = tcp_jiffies32;
3899 tp->rcv_wup = tp->rcv_nxt;
3900 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
3901
3902 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
3903 inet_csk(sk)->icsk_retransmits = 0;
3904 tcp_clear_retrans(tp);
3905}
3906
3907static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
3908{
3909 struct tcp_sock *tp = tcp_sk(sk);
3910 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
3911
3912 tcb->end_seq += skb->len;
3913 __skb_header_release(skb);
3914 sk_wmem_queued_add(sk, skb->truesize);
3915 sk_mem_charge(sk, skb->truesize);
3916 WRITE_ONCE(tp->write_seq, tcb->end_seq);
3917 tp->packets_out += tcp_skb_pcount(skb);
3918}
3919
3920/* Build and send a SYN with data and (cached) Fast Open cookie. However,
3921 * queue a data-only packet after the regular SYN, such that regular SYNs
3922 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
3923 * only the SYN sequence, the data are retransmitted in the first ACK.
3924 * If cookie is not cached or other error occurs, falls back to send a
3925 * regular SYN with Fast Open cookie request option.
3926 */
3927static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
3928{
3929 struct inet_connection_sock *icsk = inet_csk(sk);
3930 struct tcp_sock *tp = tcp_sk(sk);
3931 struct tcp_fastopen_request *fo = tp->fastopen_req;
3932 struct page_frag *pfrag = sk_page_frag(sk);
3933 struct sk_buff *syn_data;
3934 int space, err = 0;
3935
3936 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
3937 if (!tcp_fastopen_cookie_check(sk, &tp->rx_opt.mss_clamp, &fo->cookie))
3938 goto fallback;
3939
3940 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
3941 * user-MSS. Reserve maximum option space for middleboxes that add
3942 * private TCP options. The cost is reduced data space in SYN :(
3943 */
3944 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, tp->rx_opt.mss_clamp);
3945 /* Sync mss_cache after updating the mss_clamp */
3946 tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
3947
3948 space = __tcp_mtu_to_mss(sk, icsk->icsk_pmtu_cookie) -
3949 MAX_TCP_OPTION_SPACE;
3950
3951 space = min_t(size_t, space, fo->size);
3952
3953 if (space &&
3954 !skb_page_frag_refill(min_t(size_t, space, PAGE_SIZE),
3955 pfrag, sk->sk_allocation))
3956 goto fallback;
3957 syn_data = tcp_stream_alloc_skb(sk, sk->sk_allocation, false);
3958 if (!syn_data)
3959 goto fallback;
3960 memcpy(syn_data->cb, syn->cb, sizeof(syn->cb));
3961 if (space) {
3962 space = min_t(size_t, space, pfrag->size - pfrag->offset);
3963 space = tcp_wmem_schedule(sk, space);
3964 }
3965 if (space) {
3966 space = copy_page_from_iter(pfrag->page, pfrag->offset,
3967 space, &fo->data->msg_iter);
3968 if (unlikely(!space)) {
3969 tcp_skb_tsorted_anchor_cleanup(syn_data);
3970 kfree_skb(syn_data);
3971 goto fallback;
3972 }
3973 skb_fill_page_desc(syn_data, 0, pfrag->page,
3974 pfrag->offset, space);
3975 page_ref_inc(pfrag->page);
3976 pfrag->offset += space;
3977 skb_len_add(syn_data, space);
3978 skb_zcopy_set(syn_data, fo->uarg, NULL);
3979 }
3980 /* No more data pending in inet_wait_for_connect() */
3981 if (space == fo->size)
3982 fo->data = NULL;
3983 fo->copied = space;
3984
3985 tcp_connect_queue_skb(sk, syn_data);
3986 if (syn_data->len)
3987 tcp_chrono_start(sk, TCP_CHRONO_BUSY);
3988
3989 err = tcp_transmit_skb(sk, syn_data, 1, sk->sk_allocation);
3990
3991 skb_set_delivery_time(syn, syn_data->skb_mstamp_ns, true);
3992
3993 /* Now full SYN+DATA was cloned and sent (or not),
3994 * remove the SYN from the original skb (syn_data)
3995 * we keep in write queue in case of a retransmit, as we
3996 * also have the SYN packet (with no data) in the same queue.
3997 */
3998 TCP_SKB_CB(syn_data)->seq++;
3999 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
4000 if (!err) {
4001 tp->syn_data = (fo->copied > 0);
4002 tcp_rbtree_insert(&sk->tcp_rtx_queue, syn_data);
4003 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
4004 goto done;
4005 }
4006
4007 /* data was not sent, put it in write_queue */
4008 __skb_queue_tail(&sk->sk_write_queue, syn_data);
4009 tp->packets_out -= tcp_skb_pcount(syn_data);
4010
4011fallback:
4012 /* Send a regular SYN with Fast Open cookie request option */
4013 if (fo->cookie.len > 0)
4014 fo->cookie.len = 0;
4015 err = tcp_transmit_skb(sk, syn, 1, sk->sk_allocation);
4016 if (err)
4017 tp->syn_fastopen = 0;
4018done:
4019 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
4020 return err;
4021}
4022
4023/* Build a SYN and send it off. */
4024int tcp_connect(struct sock *sk)
4025{
4026 struct tcp_sock *tp = tcp_sk(sk);
4027 struct sk_buff *buff;
4028 int err;
4029
4030 tcp_call_bpf(sk, BPF_SOCK_OPS_TCP_CONNECT_CB, 0, NULL);
4031
4032#if defined(CONFIG_TCP_MD5SIG) && defined(CONFIG_TCP_AO)
4033 /* Has to be checked late, after setting daddr/saddr/ops.
4034 * Return error if the peer has both a md5 and a tcp-ao key
4035 * configured as this is ambiguous.
4036 */
4037 if (unlikely(rcu_dereference_protected(tp->md5sig_info,
4038 lockdep_sock_is_held(sk)))) {
4039 bool needs_ao = !!tp->af_specific->ao_lookup(sk, sk, -1, -1);
4040 bool needs_md5 = !!tp->af_specific->md5_lookup(sk, sk);
4041 struct tcp_ao_info *ao_info;
4042
4043 ao_info = rcu_dereference_check(tp->ao_info,
4044 lockdep_sock_is_held(sk));
4045 if (ao_info) {
4046 /* This is an extra check: tcp_ao_required() in
4047 * tcp_v{4,6}_parse_md5_keys() should prevent adding
4048 * md5 keys on ao_required socket.
4049 */
4050 needs_ao |= ao_info->ao_required;
4051 WARN_ON_ONCE(ao_info->ao_required && needs_md5);
4052 }
4053 if (needs_md5 && needs_ao)
4054 return -EKEYREJECTED;
4055
4056 /* If we have a matching md5 key and no matching tcp-ao key
4057 * then free up ao_info if allocated.
4058 */
4059 if (needs_md5) {
4060 tcp_ao_destroy_sock(sk, false);
4061 } else if (needs_ao) {
4062 tcp_clear_md5_list(sk);
4063 kfree(rcu_replace_pointer(tp->md5sig_info, NULL,
4064 lockdep_sock_is_held(sk)));
4065 }
4066 }
4067#endif
4068#ifdef CONFIG_TCP_AO
4069 if (unlikely(rcu_dereference_protected(tp->ao_info,
4070 lockdep_sock_is_held(sk)))) {
4071 /* Don't allow connecting if ao is configured but no
4072 * matching key is found.
4073 */
4074 if (!tp->af_specific->ao_lookup(sk, sk, -1, -1))
4075 return -EKEYREJECTED;
4076 }
4077#endif
4078
4079 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
4080 return -EHOSTUNREACH; /* Routing failure or similar. */
4081
4082 tcp_connect_init(sk);
4083
4084 if (unlikely(tp->repair)) {
4085 tcp_finish_connect(sk, NULL);
4086 return 0;
4087 }
4088
4089 buff = tcp_stream_alloc_skb(sk, sk->sk_allocation, true);
4090 if (unlikely(!buff))
4091 return -ENOBUFS;
4092
4093 tcp_init_nondata_skb(buff, tp->write_seq++, TCPHDR_SYN);
4094 tcp_mstamp_refresh(tp);
4095 tp->retrans_stamp = tcp_time_stamp_ts(tp);
4096 tcp_connect_queue_skb(sk, buff);
4097 tcp_ecn_send_syn(sk, buff);
4098 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
4099
4100 /* Send off SYN; include data in Fast Open. */
4101 err = tp->fastopen_req ? tcp_send_syn_data(sk, buff) :
4102 tcp_transmit_skb(sk, buff, 1, sk->sk_allocation);
4103 if (err == -ECONNREFUSED)
4104 return err;
4105
4106 /* We change tp->snd_nxt after the tcp_transmit_skb() call
4107 * in order to make this packet get counted in tcpOutSegs.
4108 */
4109 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
4110 tp->pushed_seq = tp->write_seq;
4111 buff = tcp_send_head(sk);
4112 if (unlikely(buff)) {
4113 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
4114 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
4115 }
4116 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
4117
4118 /* Timer for repeating the SYN until an answer. */
4119 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
4120 inet_csk(sk)->icsk_rto, TCP_RTO_MAX);
4121 return 0;
4122}
4123EXPORT_SYMBOL(tcp_connect);
4124
4125u32 tcp_delack_max(const struct sock *sk)
4126{
4127 const struct dst_entry *dst = __sk_dst_get(sk);
4128 u32 delack_max = inet_csk(sk)->icsk_delack_max;
4129
4130 if (dst && dst_metric_locked(dst, RTAX_RTO_MIN)) {
4131 u32 rto_min = dst_metric_rtt(dst, RTAX_RTO_MIN);
4132 u32 delack_from_rto_min = max_t(int, 1, rto_min - 1);
4133
4134 delack_max = min_t(u32, delack_max, delack_from_rto_min);
4135 }
4136 return delack_max;
4137}
4138
4139/* Send out a delayed ack, the caller does the policy checking
4140 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
4141 * for details.
4142 */
4143void tcp_send_delayed_ack(struct sock *sk)
4144{
4145 struct inet_connection_sock *icsk = inet_csk(sk);
4146 int ato = icsk->icsk_ack.ato;
4147 unsigned long timeout;
4148
4149 if (ato > TCP_DELACK_MIN) {
4150 const struct tcp_sock *tp = tcp_sk(sk);
4151 int max_ato = HZ / 2;
4152
4153 if (inet_csk_in_pingpong_mode(sk) ||
4154 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
4155 max_ato = TCP_DELACK_MAX;
4156
4157 /* Slow path, intersegment interval is "high". */
4158
4159 /* If some rtt estimate is known, use it to bound delayed ack.
4160 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
4161 * directly.
4162 */
4163 if (tp->srtt_us) {
4164 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
4165 TCP_DELACK_MIN);
4166
4167 if (rtt < max_ato)
4168 max_ato = rtt;
4169 }
4170
4171 ato = min(ato, max_ato);
4172 }
4173
4174 ato = min_t(u32, ato, tcp_delack_max(sk));
4175
4176 /* Stay within the limit we were given */
4177 timeout = jiffies + ato;
4178
4179 /* Use new timeout only if there wasn't a older one earlier. */
4180 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
4181 /* If delack timer is about to expire, send ACK now. */
4182 if (time_before_eq(icsk->icsk_ack.timeout, jiffies + (ato >> 2))) {
4183 tcp_send_ack(sk);
4184 return;
4185 }
4186
4187 if (!time_before(timeout, icsk->icsk_ack.timeout))
4188 timeout = icsk->icsk_ack.timeout;
4189 }
4190 icsk->icsk_ack.pending |= ICSK_ACK_SCHED | ICSK_ACK_TIMER;
4191 icsk->icsk_ack.timeout = timeout;
4192 sk_reset_timer(sk, &icsk->icsk_delack_timer, timeout);
4193}
4194
4195/* This routine sends an ack and also updates the window. */
4196void __tcp_send_ack(struct sock *sk, u32 rcv_nxt)
4197{
4198 struct sk_buff *buff;
4199
4200 /* If we have been reset, we may not send again. */
4201 if (sk->sk_state == TCP_CLOSE)
4202 return;
4203
4204 /* We are not putting this on the write queue, so
4205 * tcp_transmit_skb() will set the ownership to this
4206 * sock.
4207 */
4208 buff = alloc_skb(MAX_TCP_HEADER,
4209 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4210 if (unlikely(!buff)) {
4211 struct inet_connection_sock *icsk = inet_csk(sk);
4212 unsigned long delay;
4213
4214 delay = TCP_DELACK_MAX << icsk->icsk_ack.retry;
4215 if (delay < TCP_RTO_MAX)
4216 icsk->icsk_ack.retry++;
4217 inet_csk_schedule_ack(sk);
4218 icsk->icsk_ack.ato = TCP_ATO_MIN;
4219 inet_csk_reset_xmit_timer(sk, ICSK_TIME_DACK, delay, TCP_RTO_MAX);
4220 return;
4221 }
4222
4223 /* Reserve space for headers and prepare control bits. */
4224 skb_reserve(buff, MAX_TCP_HEADER);
4225 tcp_init_nondata_skb(buff, tcp_acceptable_seq(sk), TCPHDR_ACK);
4226
4227 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
4228 * too much.
4229 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
4230 */
4231 skb_set_tcp_pure_ack(buff);
4232
4233 /* Send it off, this clears delayed acks for us. */
4234 __tcp_transmit_skb(sk, buff, 0, (__force gfp_t)0, rcv_nxt);
4235}
4236EXPORT_SYMBOL_GPL(__tcp_send_ack);
4237
4238void tcp_send_ack(struct sock *sk)
4239{
4240 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt);
4241}
4242
4243/* This routine sends a packet with an out of date sequence
4244 * number. It assumes the other end will try to ack it.
4245 *
4246 * Question: what should we make while urgent mode?
4247 * 4.4BSD forces sending single byte of data. We cannot send
4248 * out of window data, because we have SND.NXT==SND.MAX...
4249 *
4250 * Current solution: to send TWO zero-length segments in urgent mode:
4251 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
4252 * out-of-date with SND.UNA-1 to probe window.
4253 */
4254static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
4255{
4256 struct tcp_sock *tp = tcp_sk(sk);
4257 struct sk_buff *skb;
4258
4259 /* We don't queue it, tcp_transmit_skb() sets ownership. */
4260 skb = alloc_skb(MAX_TCP_HEADER,
4261 sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4262 if (!skb)
4263 return -1;
4264
4265 /* Reserve space for headers and set control bits. */
4266 skb_reserve(skb, MAX_TCP_HEADER);
4267 /* Use a previous sequence. This should cause the other
4268 * end to send an ack. Don't queue or clone SKB, just
4269 * send it.
4270 */
4271 tcp_init_nondata_skb(skb, tp->snd_una - !urgent, TCPHDR_ACK);
4272 NET_INC_STATS(sock_net(sk), mib);
4273 return tcp_transmit_skb(sk, skb, 0, (__force gfp_t)0);
4274}
4275
4276/* Called from setsockopt( ... TCP_REPAIR ) */
4277void tcp_send_window_probe(struct sock *sk)
4278{
4279 if (sk->sk_state == TCP_ESTABLISHED) {
4280 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
4281 tcp_mstamp_refresh(tcp_sk(sk));
4282 tcp_xmit_probe_skb(sk, 0, LINUX_MIB_TCPWINPROBE);
4283 }
4284}
4285
4286/* Initiate keepalive or window probe from timer. */
4287int tcp_write_wakeup(struct sock *sk, int mib)
4288{
4289 struct tcp_sock *tp = tcp_sk(sk);
4290 struct sk_buff *skb;
4291
4292 if (sk->sk_state == TCP_CLOSE)
4293 return -1;
4294
4295 skb = tcp_send_head(sk);
4296 if (skb && before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp))) {
4297 int err;
4298 unsigned int mss = tcp_current_mss(sk);
4299 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
4300
4301 if (before(tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
4302 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
4303
4304 /* We are probing the opening of a window
4305 * but the window size is != 0
4306 * must have been a result SWS avoidance ( sender )
4307 */
4308 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
4309 skb->len > mss) {
4310 seg_size = min(seg_size, mss);
4311 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4312 if (tcp_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
4313 skb, seg_size, mss, GFP_ATOMIC))
4314 return -1;
4315 } else if (!tcp_skb_pcount(skb))
4316 tcp_set_skb_tso_segs(skb, mss);
4317
4318 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4319 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
4320 if (!err)
4321 tcp_event_new_data_sent(sk, skb);
4322 return err;
4323 } else {
4324 if (between(tp->snd_up, tp->snd_una + 1, tp->snd_una + 0xFFFF))
4325 tcp_xmit_probe_skb(sk, 1, mib);
4326 return tcp_xmit_probe_skb(sk, 0, mib);
4327 }
4328}
4329
4330/* A window probe timeout has occurred. If window is not closed send
4331 * a partial packet else a zero probe.
4332 */
4333void tcp_send_probe0(struct sock *sk)
4334{
4335 struct inet_connection_sock *icsk = inet_csk(sk);
4336 struct tcp_sock *tp = tcp_sk(sk);
4337 struct net *net = sock_net(sk);
4338 unsigned long timeout;
4339 int err;
4340
4341 err = tcp_write_wakeup(sk, LINUX_MIB_TCPWINPROBE);
4342
4343 if (tp->packets_out || tcp_write_queue_empty(sk)) {
4344 /* Cancel probe timer, if it is not required. */
4345 icsk->icsk_probes_out = 0;
4346 icsk->icsk_backoff = 0;
4347 icsk->icsk_probes_tstamp = 0;
4348 return;
4349 }
4350
4351 icsk->icsk_probes_out++;
4352 if (err <= 0) {
4353 if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2))
4354 icsk->icsk_backoff++;
4355 timeout = tcp_probe0_when(sk, TCP_RTO_MAX);
4356 } else {
4357 /* If packet was not sent due to local congestion,
4358 * Let senders fight for local resources conservatively.
4359 */
4360 timeout = TCP_RESOURCE_PROBE_INTERVAL;
4361 }
4362
4363 timeout = tcp_clamp_probe0_to_user_timeout(sk, timeout);
4364 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, timeout, TCP_RTO_MAX);
4365}
4366
4367int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
4368{
4369 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
4370 struct flowi fl;
4371 int res;
4372
4373 /* Paired with WRITE_ONCE() in sock_setsockopt() */
4374 if (READ_ONCE(sk->sk_txrehash) == SOCK_TXREHASH_ENABLED)
4375 WRITE_ONCE(tcp_rsk(req)->txhash, net_tx_rndhash());
4376 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL,
4377 NULL);
4378 if (!res) {
4379 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
4380 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
4381 if (unlikely(tcp_passive_fastopen(sk))) {
4382 /* sk has const attribute because listeners are lockless.
4383 * However in this case, we are dealing with a passive fastopen
4384 * socket thus we can change total_retrans value.
4385 */
4386 tcp_sk_rw(sk)->total_retrans++;
4387 }
4388 trace_tcp_retransmit_synack(sk, req);
4389 }
4390 return res;
4391}
4392EXPORT_SYMBOL(tcp_rtx_synack);