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#include <linux/static_key.h>
45
46#include <trace/events/tcp.h>
47
48static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
49 int push_one, gfp_t gfp);
50
51/* Account for new data that has been sent to the network. */
52static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
53{
54 struct inet_connection_sock *icsk = inet_csk(sk);
55 struct tcp_sock *tp = tcp_sk(sk);
56 unsigned int prior_packets = tp->packets_out;
57
58 tp->snd_nxt = TCP_SKB_CB(skb)->end_seq;
59
60 __skb_unlink(skb, &sk->sk_write_queue);
61 tcp_rbtree_insert(&sk->tcp_rtx_queue, skb);
62
63 tp->packets_out += tcp_skb_pcount(skb);
64 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
65 tcp_rearm_rto(sk);
66
67 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
68 tcp_skb_pcount(skb));
69}
70
71/* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
72 * window scaling factor due to loss of precision.
73 * If window has been shrunk, what should we make? It is not clear at all.
74 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
75 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
76 * invalid. OK, let's make this for now:
77 */
78static inline __u32 tcp_acceptable_seq(const struct sock *sk)
79{
80 const struct tcp_sock *tp = tcp_sk(sk);
81
82 if (!before(tcp_wnd_end(tp), tp->snd_nxt) ||
83 (tp->rx_opt.wscale_ok &&
84 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
85 return tp->snd_nxt;
86 else
87 return tcp_wnd_end(tp);
88}
89
90/* Calculate mss to advertise in SYN segment.
91 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
92 *
93 * 1. It is independent of path mtu.
94 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
95 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
96 * attached devices, because some buggy hosts are confused by
97 * large MSS.
98 * 4. We do not make 3, we advertise MSS, calculated from first
99 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
100 * This may be overridden via information stored in routing table.
101 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
102 * probably even Jumbo".
103 */
104static __u16 tcp_advertise_mss(struct sock *sk)
105{
106 struct tcp_sock *tp = tcp_sk(sk);
107 const struct dst_entry *dst = __sk_dst_get(sk);
108 int mss = tp->advmss;
109
110 if (dst) {
111 unsigned int metric = dst_metric_advmss(dst);
112
113 if (metric < mss) {
114 mss = metric;
115 tp->advmss = mss;
116 }
117 }
118
119 return (__u16)mss;
120}
121
122/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
123 * This is the first part of cwnd validation mechanism.
124 */
125void tcp_cwnd_restart(struct sock *sk, s32 delta)
126{
127 struct tcp_sock *tp = tcp_sk(sk);
128 u32 restart_cwnd = tcp_init_cwnd(tp, __sk_dst_get(sk));
129 u32 cwnd = tp->snd_cwnd;
130
131 tcp_ca_event(sk, CA_EVENT_CWND_RESTART);
132
133 tp->snd_ssthresh = tcp_current_ssthresh(sk);
134 restart_cwnd = min(restart_cwnd, cwnd);
135
136 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
137 cwnd >>= 1;
138 tp->snd_cwnd = max(cwnd, restart_cwnd);
139 tp->snd_cwnd_stamp = tcp_jiffies32;
140 tp->snd_cwnd_used = 0;
141}
142
143/* Congestion state accounting after a packet has been sent. */
144static void tcp_event_data_sent(struct tcp_sock *tp,
145 struct sock *sk)
146{
147 struct inet_connection_sock *icsk = inet_csk(sk);
148 const u32 now = tcp_jiffies32;
149
150 if (tcp_packets_in_flight(tp) == 0)
151 tcp_ca_event(sk, CA_EVENT_TX_START);
152
153 tp->lsndtime = now;
154
155 /* If it is a reply for ato after last received
156 * packet, enter pingpong mode.
157 */
158 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
159 icsk->icsk_ack.pingpong = 1;
160}
161
162/* Account for an ACK we sent. */
163static inline void tcp_event_ack_sent(struct sock *sk, unsigned int pkts,
164 u32 rcv_nxt)
165{
166 struct tcp_sock *tp = tcp_sk(sk);
167
168 if (unlikely(tp->compressed_ack)) {
169 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
170 tp->compressed_ack);
171 tp->compressed_ack = 0;
172 if (hrtimer_try_to_cancel(&tp->compressed_ack_timer) == 1)
173 __sock_put(sk);
174 }
175
176 if (unlikely(rcv_nxt != tp->rcv_nxt))
177 return; /* Special ACK sent by DCTCP to reflect ECN */
178 tcp_dec_quickack_mode(sk, pkts);
179 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
180}
181
182
183u32 tcp_default_init_rwnd(u32 mss)
184{
185 /* Initial receive window should be twice of TCP_INIT_CWND to
186 * enable proper sending of new unsent data during fast recovery
187 * (RFC 3517, Section 4, NextSeg() rule (2)). Further place a
188 * limit when mss is larger than 1460.
189 */
190 u32 init_rwnd = TCP_INIT_CWND * 2;
191
192 if (mss > 1460)
193 init_rwnd = max((1460 * init_rwnd) / mss, 2U);
194 return init_rwnd;
195}
196
197/* Determine a window scaling and initial window to offer.
198 * Based on the assumption that the given amount of space
199 * will be offered. Store the results in the tp structure.
200 * NOTE: for smooth operation initial space offering should
201 * be a multiple of mss if possible. We assume here that mss >= 1.
202 * This MUST be enforced by all callers.
203 */
204void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
205 __u32 *rcv_wnd, __u32 *window_clamp,
206 int wscale_ok, __u8 *rcv_wscale,
207 __u32 init_rcv_wnd)
208{
209 unsigned int space = (__space < 0 ? 0 : __space);
210
211 /* If no clamp set the clamp to the max possible scaled window */
212 if (*window_clamp == 0)
213 (*window_clamp) = (U16_MAX << TCP_MAX_WSCALE);
214 space = min(*window_clamp, space);
215
216 /* Quantize space offering to a multiple of mss if possible. */
217 if (space > mss)
218 space = rounddown(space, mss);
219
220 /* NOTE: offering an initial window larger than 32767
221 * will break some buggy TCP stacks. If the admin tells us
222 * it is likely we could be speaking with such a buggy stack
223 * we will truncate our initial window offering to 32K-1
224 * unless the remote has sent us a window scaling option,
225 * which we interpret as a sign the remote TCP is not
226 * misinterpreting the window field as a signed quantity.
227 */
228 if (sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
229 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
230 else
231 (*rcv_wnd) = space;
232
233 (*rcv_wscale) = 0;
234 if (wscale_ok) {
235 /* Set window scaling on max possible window */
236 space = max_t(u32, space, sock_net(sk)->ipv4.sysctl_tcp_rmem[2]);
237 space = max_t(u32, space, sysctl_rmem_max);
238 space = min_t(u32, space, *window_clamp);
239 while (space > U16_MAX && (*rcv_wscale) < TCP_MAX_WSCALE) {
240 space >>= 1;
241 (*rcv_wscale)++;
242 }
243 }
244
245 if (!init_rcv_wnd) /* Use default unless specified otherwise */
246 init_rcv_wnd = tcp_default_init_rwnd(mss);
247 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
248
249 /* Set the clamp no higher than max representable value */
250 (*window_clamp) = min_t(__u32, U16_MAX << (*rcv_wscale), *window_clamp);
251}
252EXPORT_SYMBOL(tcp_select_initial_window);
253
254/* Chose a new window to advertise, update state in tcp_sock for the
255 * socket, and return result with RFC1323 scaling applied. The return
256 * value can be stuffed directly into th->window for an outgoing
257 * frame.
258 */
259static u16 tcp_select_window(struct sock *sk)
260{
261 struct tcp_sock *tp = tcp_sk(sk);
262 u32 old_win = tp->rcv_wnd;
263 u32 cur_win = tcp_receive_window(tp);
264 u32 new_win = __tcp_select_window(sk);
265
266 /* Never shrink the offered window */
267 if (new_win < cur_win) {
268 /* Danger Will Robinson!
269 * Don't update rcv_wup/rcv_wnd here or else
270 * we will not be able to advertise a zero
271 * window in time. --DaveM
272 *
273 * Relax Will Robinson.
274 */
275 if (new_win == 0)
276 NET_INC_STATS(sock_net(sk),
277 LINUX_MIB_TCPWANTZEROWINDOWADV);
278 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
279 }
280 tp->rcv_wnd = new_win;
281 tp->rcv_wup = tp->rcv_nxt;
282
283 /* Make sure we do not exceed the maximum possible
284 * scaled window.
285 */
286 if (!tp->rx_opt.rcv_wscale &&
287 sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows)
288 new_win = min(new_win, MAX_TCP_WINDOW);
289 else
290 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
291
292 /* RFC1323 scaling applied */
293 new_win >>= tp->rx_opt.rcv_wscale;
294
295 /* If we advertise zero window, disable fast path. */
296 if (new_win == 0) {
297 tp->pred_flags = 0;
298 if (old_win)
299 NET_INC_STATS(sock_net(sk),
300 LINUX_MIB_TCPTOZEROWINDOWADV);
301 } else if (old_win == 0) {
302 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPFROMZEROWINDOWADV);
303 }
304
305 return new_win;
306}
307
308/* Packet ECN state for a SYN-ACK */
309static void tcp_ecn_send_synack(struct sock *sk, struct sk_buff *skb)
310{
311 const struct tcp_sock *tp = tcp_sk(sk);
312
313 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_CWR;
314 if (!(tp->ecn_flags & TCP_ECN_OK))
315 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_ECE;
316 else if (tcp_ca_needs_ecn(sk) ||
317 tcp_bpf_ca_needs_ecn(sk))
318 INET_ECN_xmit(sk);
319}
320
321/* Packet ECN state for a SYN. */
322static void tcp_ecn_send_syn(struct sock *sk, struct sk_buff *skb)
323{
324 struct tcp_sock *tp = tcp_sk(sk);
325 bool bpf_needs_ecn = tcp_bpf_ca_needs_ecn(sk);
326 bool use_ecn = sock_net(sk)->ipv4.sysctl_tcp_ecn == 1 ||
327 tcp_ca_needs_ecn(sk) || bpf_needs_ecn;
328
329 if (!use_ecn) {
330 const struct dst_entry *dst = __sk_dst_get(sk);
331
332 if (dst && dst_feature(dst, RTAX_FEATURE_ECN))
333 use_ecn = true;
334 }
335
336 tp->ecn_flags = 0;
337
338 if (use_ecn) {
339 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ECE | TCPHDR_CWR;
340 tp->ecn_flags = TCP_ECN_OK;
341 if (tcp_ca_needs_ecn(sk) || bpf_needs_ecn)
342 INET_ECN_xmit(sk);
343 }
344}
345
346static void tcp_ecn_clear_syn(struct sock *sk, struct sk_buff *skb)
347{
348 if (sock_net(sk)->ipv4.sysctl_tcp_ecn_fallback)
349 /* tp->ecn_flags are cleared at a later point in time when
350 * SYN ACK is ultimatively being received.
351 */
352 TCP_SKB_CB(skb)->tcp_flags &= ~(TCPHDR_ECE | TCPHDR_CWR);
353}
354
355static void
356tcp_ecn_make_synack(const struct request_sock *req, struct tcphdr *th)
357{
358 if (inet_rsk(req)->ecn_ok)
359 th->ece = 1;
360}
361
362/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
363 * be sent.
364 */
365static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
366 struct tcphdr *th, int tcp_header_len)
367{
368 struct tcp_sock *tp = tcp_sk(sk);
369
370 if (tp->ecn_flags & TCP_ECN_OK) {
371 /* Not-retransmitted data segment: set ECT and inject CWR. */
372 if (skb->len != tcp_header_len &&
373 !before(TCP_SKB_CB(skb)->seq, tp->snd_nxt)) {
374 INET_ECN_xmit(sk);
375 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
376 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
377 th->cwr = 1;
378 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
379 }
380 } else if (!tcp_ca_needs_ecn(sk)) {
381 /* ACK or retransmitted segment: clear ECT|CE */
382 INET_ECN_dontxmit(sk);
383 }
384 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
385 th->ece = 1;
386 }
387}
388
389/* Constructs common control bits of non-data skb. If SYN/FIN is present,
390 * auto increment end seqno.
391 */
392static void tcp_init_nondata_skb(struct sk_buff *skb, u32 seq, u8 flags)
393{
394 skb->ip_summed = CHECKSUM_PARTIAL;
395
396 TCP_SKB_CB(skb)->tcp_flags = flags;
397 TCP_SKB_CB(skb)->sacked = 0;
398
399 tcp_skb_pcount_set(skb, 1);
400
401 TCP_SKB_CB(skb)->seq = seq;
402 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
403 seq++;
404 TCP_SKB_CB(skb)->end_seq = seq;
405}
406
407static inline bool tcp_urg_mode(const struct tcp_sock *tp)
408{
409 return tp->snd_una != tp->snd_up;
410}
411
412#define OPTION_SACK_ADVERTISE (1 << 0)
413#define OPTION_TS (1 << 1)
414#define OPTION_MD5 (1 << 2)
415#define OPTION_WSCALE (1 << 3)
416#define OPTION_FAST_OPEN_COOKIE (1 << 8)
417#define OPTION_SMC (1 << 9)
418
419static void smc_options_write(__be32 *ptr, u16 *options)
420{
421#if IS_ENABLED(CONFIG_SMC)
422 if (static_branch_unlikely(&tcp_have_smc)) {
423 if (unlikely(OPTION_SMC & *options)) {
424 *ptr++ = htonl((TCPOPT_NOP << 24) |
425 (TCPOPT_NOP << 16) |
426 (TCPOPT_EXP << 8) |
427 (TCPOLEN_EXP_SMC_BASE));
428 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
429 }
430 }
431#endif
432}
433
434struct tcp_out_options {
435 u16 options; /* bit field of OPTION_* */
436 u16 mss; /* 0 to disable */
437 u8 ws; /* window scale, 0 to disable */
438 u8 num_sack_blocks; /* number of SACK blocks to include */
439 u8 hash_size; /* bytes in hash_location */
440 __u8 *hash_location; /* temporary pointer, overloaded */
441 __u32 tsval, tsecr; /* need to include OPTION_TS */
442 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
443};
444
445/* Write previously computed TCP options to the packet.
446 *
447 * Beware: Something in the Internet is very sensitive to the ordering of
448 * TCP options, we learned this through the hard way, so be careful here.
449 * Luckily we can at least blame others for their non-compliance but from
450 * inter-operability perspective it seems that we're somewhat stuck with
451 * the ordering which we have been using if we want to keep working with
452 * those broken things (not that it currently hurts anybody as there isn't
453 * particular reason why the ordering would need to be changed).
454 *
455 * At least SACK_PERM as the first option is known to lead to a disaster
456 * (but it may well be that other scenarios fail similarly).
457 */
458static void tcp_options_write(__be32 *ptr, struct tcp_sock *tp,
459 struct tcp_out_options *opts)
460{
461 u16 options = opts->options; /* mungable copy */
462
463 if (unlikely(OPTION_MD5 & options)) {
464 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
465 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
466 /* overload cookie hash location */
467 opts->hash_location = (__u8 *)ptr;
468 ptr += 4;
469 }
470
471 if (unlikely(opts->mss)) {
472 *ptr++ = htonl((TCPOPT_MSS << 24) |
473 (TCPOLEN_MSS << 16) |
474 opts->mss);
475 }
476
477 if (likely(OPTION_TS & options)) {
478 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
479 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
480 (TCPOLEN_SACK_PERM << 16) |
481 (TCPOPT_TIMESTAMP << 8) |
482 TCPOLEN_TIMESTAMP);
483 options &= ~OPTION_SACK_ADVERTISE;
484 } else {
485 *ptr++ = htonl((TCPOPT_NOP << 24) |
486 (TCPOPT_NOP << 16) |
487 (TCPOPT_TIMESTAMP << 8) |
488 TCPOLEN_TIMESTAMP);
489 }
490 *ptr++ = htonl(opts->tsval);
491 *ptr++ = htonl(opts->tsecr);
492 }
493
494 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
495 *ptr++ = htonl((TCPOPT_NOP << 24) |
496 (TCPOPT_NOP << 16) |
497 (TCPOPT_SACK_PERM << 8) |
498 TCPOLEN_SACK_PERM);
499 }
500
501 if (unlikely(OPTION_WSCALE & options)) {
502 *ptr++ = htonl((TCPOPT_NOP << 24) |
503 (TCPOPT_WINDOW << 16) |
504 (TCPOLEN_WINDOW << 8) |
505 opts->ws);
506 }
507
508 if (unlikely(opts->num_sack_blocks)) {
509 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
510 tp->duplicate_sack : tp->selective_acks;
511 int this_sack;
512
513 *ptr++ = htonl((TCPOPT_NOP << 24) |
514 (TCPOPT_NOP << 16) |
515 (TCPOPT_SACK << 8) |
516 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
517 TCPOLEN_SACK_PERBLOCK)));
518
519 for (this_sack = 0; this_sack < opts->num_sack_blocks;
520 ++this_sack) {
521 *ptr++ = htonl(sp[this_sack].start_seq);
522 *ptr++ = htonl(sp[this_sack].end_seq);
523 }
524
525 tp->rx_opt.dsack = 0;
526 }
527
528 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
529 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
530 u8 *p = (u8 *)ptr;
531 u32 len; /* Fast Open option length */
532
533 if (foc->exp) {
534 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
535 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
536 TCPOPT_FASTOPEN_MAGIC);
537 p += TCPOLEN_EXP_FASTOPEN_BASE;
538 } else {
539 len = TCPOLEN_FASTOPEN_BASE + foc->len;
540 *p++ = TCPOPT_FASTOPEN;
541 *p++ = len;
542 }
543
544 memcpy(p, foc->val, foc->len);
545 if ((len & 3) == 2) {
546 p[foc->len] = TCPOPT_NOP;
547 p[foc->len + 1] = TCPOPT_NOP;
548 }
549 ptr += (len + 3) >> 2;
550 }
551
552 smc_options_write(ptr, &options);
553}
554
555static void smc_set_option(const struct tcp_sock *tp,
556 struct tcp_out_options *opts,
557 unsigned int *remaining)
558{
559#if IS_ENABLED(CONFIG_SMC)
560 if (static_branch_unlikely(&tcp_have_smc)) {
561 if (tp->syn_smc) {
562 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
563 opts->options |= OPTION_SMC;
564 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
565 }
566 }
567 }
568#endif
569}
570
571static void smc_set_option_cond(const struct tcp_sock *tp,
572 const struct inet_request_sock *ireq,
573 struct tcp_out_options *opts,
574 unsigned int *remaining)
575{
576#if IS_ENABLED(CONFIG_SMC)
577 if (static_branch_unlikely(&tcp_have_smc)) {
578 if (tp->syn_smc && ireq->smc_ok) {
579 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
580 opts->options |= OPTION_SMC;
581 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
582 }
583 }
584 }
585#endif
586}
587
588/* Compute TCP options for SYN packets. This is not the final
589 * network wire format yet.
590 */
591static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
592 struct tcp_out_options *opts,
593 struct tcp_md5sig_key **md5)
594{
595 struct tcp_sock *tp = tcp_sk(sk);
596 unsigned int remaining = MAX_TCP_OPTION_SPACE;
597 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
598
599 *md5 = NULL;
600#ifdef CONFIG_TCP_MD5SIG
601 if (unlikely(rcu_access_pointer(tp->md5sig_info))) {
602 *md5 = tp->af_specific->md5_lookup(sk, sk);
603 if (*md5) {
604 opts->options |= OPTION_MD5;
605 remaining -= TCPOLEN_MD5SIG_ALIGNED;
606 }
607 }
608#endif
609
610 /* We always get an MSS option. The option bytes which will be seen in
611 * normal data packets should timestamps be used, must be in the MSS
612 * advertised. But we subtract them from tp->mss_cache so that
613 * calculations in tcp_sendmsg are simpler etc. So account for this
614 * fact here if necessary. If we don't do this correctly, as a
615 * receiver we won't recognize data packets as being full sized when we
616 * should, and thus we won't abide by the delayed ACK rules correctly.
617 * SACKs don't matter, we never delay an ACK when we have any of those
618 * going out. */
619 opts->mss = tcp_advertise_mss(sk);
620 remaining -= TCPOLEN_MSS_ALIGNED;
621
622 if (likely(sock_net(sk)->ipv4.sysctl_tcp_timestamps && !*md5)) {
623 opts->options |= OPTION_TS;
624 opts->tsval = tcp_skb_timestamp(skb) + tp->tsoffset;
625 opts->tsecr = tp->rx_opt.ts_recent;
626 remaining -= TCPOLEN_TSTAMP_ALIGNED;
627 }
628 if (likely(sock_net(sk)->ipv4.sysctl_tcp_window_scaling)) {
629 opts->ws = tp->rx_opt.rcv_wscale;
630 opts->options |= OPTION_WSCALE;
631 remaining -= TCPOLEN_WSCALE_ALIGNED;
632 }
633 if (likely(sock_net(sk)->ipv4.sysctl_tcp_sack)) {
634 opts->options |= OPTION_SACK_ADVERTISE;
635 if (unlikely(!(OPTION_TS & opts->options)))
636 remaining -= TCPOLEN_SACKPERM_ALIGNED;
637 }
638
639 if (fastopen && fastopen->cookie.len >= 0) {
640 u32 need = fastopen->cookie.len;
641
642 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
643 TCPOLEN_FASTOPEN_BASE;
644 need = (need + 3) & ~3U; /* Align to 32 bits */
645 if (remaining >= need) {
646 opts->options |= OPTION_FAST_OPEN_COOKIE;
647 opts->fastopen_cookie = &fastopen->cookie;
648 remaining -= need;
649 tp->syn_fastopen = 1;
650 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
651 }
652 }
653
654 smc_set_option(tp, opts, &remaining);
655
656 return MAX_TCP_OPTION_SPACE - remaining;
657}
658
659/* Set up TCP options for SYN-ACKs. */
660static unsigned int tcp_synack_options(const struct sock *sk,
661 struct request_sock *req,
662 unsigned int mss, struct sk_buff *skb,
663 struct tcp_out_options *opts,
664 const struct tcp_md5sig_key *md5,
665 struct tcp_fastopen_cookie *foc)
666{
667 struct inet_request_sock *ireq = inet_rsk(req);
668 unsigned int remaining = MAX_TCP_OPTION_SPACE;
669
670#ifdef CONFIG_TCP_MD5SIG
671 if (md5) {
672 opts->options |= OPTION_MD5;
673 remaining -= TCPOLEN_MD5SIG_ALIGNED;
674
675 /* We can't fit any SACK blocks in a packet with MD5 + TS
676 * options. There was discussion about disabling SACK
677 * rather than TS in order to fit in better with old,
678 * buggy kernels, but that was deemed to be unnecessary.
679 */
680 ireq->tstamp_ok &= !ireq->sack_ok;
681 }
682#endif
683
684 /* We always send an MSS option. */
685 opts->mss = mss;
686 remaining -= TCPOLEN_MSS_ALIGNED;
687
688 if (likely(ireq->wscale_ok)) {
689 opts->ws = ireq->rcv_wscale;
690 opts->options |= OPTION_WSCALE;
691 remaining -= TCPOLEN_WSCALE_ALIGNED;
692 }
693 if (likely(ireq->tstamp_ok)) {
694 opts->options |= OPTION_TS;
695 opts->tsval = tcp_skb_timestamp(skb) + tcp_rsk(req)->ts_off;
696 opts->tsecr = req->ts_recent;
697 remaining -= TCPOLEN_TSTAMP_ALIGNED;
698 }
699 if (likely(ireq->sack_ok)) {
700 opts->options |= OPTION_SACK_ADVERTISE;
701 if (unlikely(!ireq->tstamp_ok))
702 remaining -= TCPOLEN_SACKPERM_ALIGNED;
703 }
704 if (foc != NULL && foc->len >= 0) {
705 u32 need = foc->len;
706
707 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
708 TCPOLEN_FASTOPEN_BASE;
709 need = (need + 3) & ~3U; /* Align to 32 bits */
710 if (remaining >= need) {
711 opts->options |= OPTION_FAST_OPEN_COOKIE;
712 opts->fastopen_cookie = foc;
713 remaining -= need;
714 }
715 }
716
717 smc_set_option_cond(tcp_sk(sk), ireq, opts, &remaining);
718
719 return MAX_TCP_OPTION_SPACE - remaining;
720}
721
722/* Compute TCP options for ESTABLISHED sockets. This is not the
723 * final wire format yet.
724 */
725static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
726 struct tcp_out_options *opts,
727 struct tcp_md5sig_key **md5)
728{
729 struct tcp_sock *tp = tcp_sk(sk);
730 unsigned int size = 0;
731 unsigned int eff_sacks;
732
733 opts->options = 0;
734
735 *md5 = NULL;
736#ifdef CONFIG_TCP_MD5SIG
737 if (unlikely(rcu_access_pointer(tp->md5sig_info))) {
738 *md5 = tp->af_specific->md5_lookup(sk, sk);
739 if (*md5) {
740 opts->options |= OPTION_MD5;
741 size += TCPOLEN_MD5SIG_ALIGNED;
742 }
743 }
744#endif
745
746 if (likely(tp->rx_opt.tstamp_ok)) {
747 opts->options |= OPTION_TS;
748 opts->tsval = skb ? tcp_skb_timestamp(skb) + tp->tsoffset : 0;
749 opts->tsecr = tp->rx_opt.ts_recent;
750 size += TCPOLEN_TSTAMP_ALIGNED;
751 }
752
753 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
754 if (unlikely(eff_sacks)) {
755 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
756 opts->num_sack_blocks =
757 min_t(unsigned int, eff_sacks,
758 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
759 TCPOLEN_SACK_PERBLOCK);
760 size += TCPOLEN_SACK_BASE_ALIGNED +
761 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
762 }
763
764 return size;
765}
766
767
768/* TCP SMALL QUEUES (TSQ)
769 *
770 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
771 * to reduce RTT and bufferbloat.
772 * We do this using a special skb destructor (tcp_wfree).
773 *
774 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
775 * needs to be reallocated in a driver.
776 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
777 *
778 * Since transmit from skb destructor is forbidden, we use a tasklet
779 * to process all sockets that eventually need to send more skbs.
780 * We use one tasklet per cpu, with its own queue of sockets.
781 */
782struct tsq_tasklet {
783 struct tasklet_struct tasklet;
784 struct list_head head; /* queue of tcp sockets */
785};
786static DEFINE_PER_CPU(struct tsq_tasklet, tsq_tasklet);
787
788static void tcp_tsq_write(struct sock *sk)
789{
790 if ((1 << sk->sk_state) &
791 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
792 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
793 struct tcp_sock *tp = tcp_sk(sk);
794
795 if (tp->lost_out > tp->retrans_out &&
796 tp->snd_cwnd > tcp_packets_in_flight(tp)) {
797 tcp_mstamp_refresh(tp);
798 tcp_xmit_retransmit_queue(sk);
799 }
800
801 tcp_write_xmit(sk, tcp_current_mss(sk), tp->nonagle,
802 0, GFP_ATOMIC);
803 }
804}
805
806static void tcp_tsq_handler(struct sock *sk)
807{
808 bh_lock_sock(sk);
809 if (!sock_owned_by_user(sk))
810 tcp_tsq_write(sk);
811 else if (!test_and_set_bit(TCP_TSQ_DEFERRED, &sk->sk_tsq_flags))
812 sock_hold(sk);
813 bh_unlock_sock(sk);
814}
815/*
816 * One tasklet per cpu tries to send more skbs.
817 * We run in tasklet context but need to disable irqs when
818 * transferring tsq->head because tcp_wfree() might
819 * interrupt us (non NAPI drivers)
820 */
821static void tcp_tasklet_func(unsigned long data)
822{
823 struct tsq_tasklet *tsq = (struct tsq_tasklet *)data;
824 LIST_HEAD(list);
825 unsigned long flags;
826 struct list_head *q, *n;
827 struct tcp_sock *tp;
828 struct sock *sk;
829
830 local_irq_save(flags);
831 list_splice_init(&tsq->head, &list);
832 local_irq_restore(flags);
833
834 list_for_each_safe(q, n, &list) {
835 tp = list_entry(q, struct tcp_sock, tsq_node);
836 list_del(&tp->tsq_node);
837
838 sk = (struct sock *)tp;
839 smp_mb__before_atomic();
840 clear_bit(TSQ_QUEUED, &sk->sk_tsq_flags);
841
842 tcp_tsq_handler(sk);
843 sk_free(sk);
844 }
845}
846
847#define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
848 TCPF_WRITE_TIMER_DEFERRED | \
849 TCPF_DELACK_TIMER_DEFERRED | \
850 TCPF_MTU_REDUCED_DEFERRED)
851/**
852 * tcp_release_cb - tcp release_sock() callback
853 * @sk: socket
854 *
855 * called from release_sock() to perform protocol dependent
856 * actions before socket release.
857 */
858void tcp_release_cb(struct sock *sk)
859{
860 unsigned long flags, nflags;
861
862 /* perform an atomic operation only if at least one flag is set */
863 do {
864 flags = sk->sk_tsq_flags;
865 if (!(flags & TCP_DEFERRED_ALL))
866 return;
867 nflags = flags & ~TCP_DEFERRED_ALL;
868 } while (cmpxchg(&sk->sk_tsq_flags, flags, nflags) != flags);
869
870 if (flags & TCPF_TSQ_DEFERRED) {
871 tcp_tsq_write(sk);
872 __sock_put(sk);
873 }
874 /* Here begins the tricky part :
875 * We are called from release_sock() with :
876 * 1) BH disabled
877 * 2) sk_lock.slock spinlock held
878 * 3) socket owned by us (sk->sk_lock.owned == 1)
879 *
880 * But following code is meant to be called from BH handlers,
881 * so we should keep BH disabled, but early release socket ownership
882 */
883 sock_release_ownership(sk);
884
885 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
886 tcp_write_timer_handler(sk);
887 __sock_put(sk);
888 }
889 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
890 tcp_delack_timer_handler(sk);
891 __sock_put(sk);
892 }
893 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
894 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
895 __sock_put(sk);
896 }
897}
898EXPORT_SYMBOL(tcp_release_cb);
899
900void __init tcp_tasklet_init(void)
901{
902 int i;
903
904 for_each_possible_cpu(i) {
905 struct tsq_tasklet *tsq = &per_cpu(tsq_tasklet, i);
906
907 INIT_LIST_HEAD(&tsq->head);
908 tasklet_init(&tsq->tasklet,
909 tcp_tasklet_func,
910 (unsigned long)tsq);
911 }
912}
913
914/*
915 * Write buffer destructor automatically called from kfree_skb.
916 * We can't xmit new skbs from this context, as we might already
917 * hold qdisc lock.
918 */
919void tcp_wfree(struct sk_buff *skb)
920{
921 struct sock *sk = skb->sk;
922 struct tcp_sock *tp = tcp_sk(sk);
923 unsigned long flags, nval, oval;
924
925 /* Keep one reference on sk_wmem_alloc.
926 * Will be released by sk_free() from here or tcp_tasklet_func()
927 */
928 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
929
930 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
931 * Wait until our queues (qdisc + devices) are drained.
932 * This gives :
933 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
934 * - chance for incoming ACK (processed by another cpu maybe)
935 * to migrate this flow (skb->ooo_okay will be eventually set)
936 */
937 if (refcount_read(&sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
938 goto out;
939
940 for (oval = READ_ONCE(sk->sk_tsq_flags);; oval = nval) {
941 struct tsq_tasklet *tsq;
942 bool empty;
943
944 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
945 goto out;
946
947 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
948 nval = cmpxchg(&sk->sk_tsq_flags, oval, nval);
949 if (nval != oval)
950 continue;
951
952 /* queue this socket to tasklet queue */
953 local_irq_save(flags);
954 tsq = this_cpu_ptr(&tsq_tasklet);
955 empty = list_empty(&tsq->head);
956 list_add(&tp->tsq_node, &tsq->head);
957 if (empty)
958 tasklet_schedule(&tsq->tasklet);
959 local_irq_restore(flags);
960 return;
961 }
962out:
963 sk_free(sk);
964}
965
966/* Note: Called under soft irq.
967 * We can call TCP stack right away, unless socket is owned by user.
968 */
969enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
970{
971 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
972 struct sock *sk = (struct sock *)tp;
973
974 tcp_tsq_handler(sk);
975 sock_put(sk);
976
977 return HRTIMER_NORESTART;
978}
979
980static void tcp_internal_pacing(struct sock *sk, const struct sk_buff *skb)
981{
982 u64 len_ns;
983 u32 rate;
984
985 if (!tcp_needs_internal_pacing(sk))
986 return;
987 rate = sk->sk_pacing_rate;
988 if (!rate || rate == ~0U)
989 return;
990
991 len_ns = (u64)skb->len * NSEC_PER_SEC;
992 do_div(len_ns, rate);
993 hrtimer_start(&tcp_sk(sk)->pacing_timer,
994 ktime_add_ns(ktime_get(), len_ns),
995 HRTIMER_MODE_ABS_PINNED_SOFT);
996 sock_hold(sk);
997}
998
999static void tcp_update_skb_after_send(struct tcp_sock *tp, struct sk_buff *skb)
1000{
1001 skb->skb_mstamp = tp->tcp_mstamp;
1002 list_move_tail(&skb->tcp_tsorted_anchor, &tp->tsorted_sent_queue);
1003}
1004
1005/* This routine actually transmits TCP packets queued in by
1006 * tcp_do_sendmsg(). This is used by both the initial
1007 * transmission and possible later retransmissions.
1008 * All SKB's seen here are completely headerless. It is our
1009 * job to build the TCP header, and pass the packet down to
1010 * IP so it can do the same plus pass the packet off to the
1011 * device.
1012 *
1013 * We are working here with either a clone of the original
1014 * SKB, or a fresh unique copy made by the retransmit engine.
1015 */
1016static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1017 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1018{
1019 const struct inet_connection_sock *icsk = inet_csk(sk);
1020 struct inet_sock *inet;
1021 struct tcp_sock *tp;
1022 struct tcp_skb_cb *tcb;
1023 struct tcp_out_options opts;
1024 unsigned int tcp_options_size, tcp_header_size;
1025 struct sk_buff *oskb = NULL;
1026 struct tcp_md5sig_key *md5;
1027 struct tcphdr *th;
1028 int err;
1029
1030 BUG_ON(!skb || !tcp_skb_pcount(skb));
1031 tp = tcp_sk(sk);
1032
1033 if (clone_it) {
1034 TCP_SKB_CB(skb)->tx.in_flight = TCP_SKB_CB(skb)->end_seq
1035 - tp->snd_una;
1036 oskb = skb;
1037
1038 tcp_skb_tsorted_save(oskb) {
1039 if (unlikely(skb_cloned(oskb)))
1040 skb = pskb_copy(oskb, gfp_mask);
1041 else
1042 skb = skb_clone(oskb, gfp_mask);
1043 } tcp_skb_tsorted_restore(oskb);
1044
1045 if (unlikely(!skb))
1046 return -ENOBUFS;
1047 }
1048 skb->skb_mstamp = tp->tcp_mstamp;
1049
1050 inet = inet_sk(sk);
1051 tcb = TCP_SKB_CB(skb);
1052 memset(&opts, 0, sizeof(opts));
1053
1054 if (unlikely(tcb->tcp_flags & TCPHDR_SYN))
1055 tcp_options_size = tcp_syn_options(sk, skb, &opts, &md5);
1056 else
1057 tcp_options_size = tcp_established_options(sk, skb, &opts,
1058 &md5);
1059 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1060
1061 /* if no packet is in qdisc/device queue, then allow XPS to select
1062 * another queue. We can be called from tcp_tsq_handler()
1063 * which holds one reference to sk.
1064 *
1065 * TODO: Ideally, in-flight pure ACK packets should not matter here.
1066 * One way to get this would be to set skb->truesize = 2 on them.
1067 */
1068 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1);
1069
1070 /* If we had to use memory reserve to allocate this skb,
1071 * this might cause drops if packet is looped back :
1072 * Other socket might not have SOCK_MEMALLOC.
1073 * Packets not looped back do not care about pfmemalloc.
1074 */
1075 skb->pfmemalloc = 0;
1076
1077 skb_push(skb, tcp_header_size);
1078 skb_reset_transport_header(skb);
1079
1080 skb_orphan(skb);
1081 skb->sk = sk;
1082 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1083 skb_set_hash_from_sk(skb, sk);
1084 refcount_add(skb->truesize, &sk->sk_wmem_alloc);
1085
1086 skb_set_dst_pending_confirm(skb, sk->sk_dst_pending_confirm);
1087
1088 /* Build TCP header and checksum it. */
1089 th = (struct tcphdr *)skb->data;
1090 th->source = inet->inet_sport;
1091 th->dest = inet->inet_dport;
1092 th->seq = htonl(tcb->seq);
1093 th->ack_seq = htonl(rcv_nxt);
1094 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1095 tcb->tcp_flags);
1096
1097 th->check = 0;
1098 th->urg_ptr = 0;
1099
1100 /* The urg_mode check is necessary during a below snd_una win probe */
1101 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1102 if (before(tp->snd_up, tcb->seq + 0x10000)) {
1103 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1104 th->urg = 1;
1105 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1106 th->urg_ptr = htons(0xFFFF);
1107 th->urg = 1;
1108 }
1109 }
1110
1111 tcp_options_write((__be32 *)(th + 1), tp, &opts);
1112 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1113 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1114 th->window = htons(tcp_select_window(sk));
1115 tcp_ecn_send(sk, skb, th, tcp_header_size);
1116 } else {
1117 /* RFC1323: The window in SYN & SYN/ACK segments
1118 * is never scaled.
1119 */
1120 th->window = htons(min(tp->rcv_wnd, 65535U));
1121 }
1122#ifdef CONFIG_TCP_MD5SIG
1123 /* Calculate the MD5 hash, as we have all we need now */
1124 if (md5) {
1125 sk_nocaps_add(sk, NETIF_F_GSO_MASK);
1126 tp->af_specific->calc_md5_hash(opts.hash_location,
1127 md5, sk, skb);
1128 }
1129#endif
1130
1131 icsk->icsk_af_ops->send_check(sk, skb);
1132
1133 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1134 tcp_event_ack_sent(sk, tcp_skb_pcount(skb), rcv_nxt);
1135
1136 if (skb->len != tcp_header_size) {
1137 tcp_event_data_sent(tp, sk);
1138 tp->data_segs_out += tcp_skb_pcount(skb);
1139 tp->bytes_sent += skb->len - tcp_header_size;
1140 tcp_internal_pacing(sk, skb);
1141 }
1142
1143 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1144 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1145 tcp_skb_pcount(skb));
1146
1147 tp->segs_out += tcp_skb_pcount(skb);
1148 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1149 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1150 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1151
1152 /* Our usage of tstamp should remain private */
1153 skb->tstamp = 0;
1154
1155 /* Cleanup our debris for IP stacks */
1156 memset(skb->cb, 0, max(sizeof(struct inet_skb_parm),
1157 sizeof(struct inet6_skb_parm)));
1158
1159 err = icsk->icsk_af_ops->queue_xmit(sk, skb, &inet->cork.fl);
1160
1161 if (unlikely(err > 0)) {
1162 tcp_enter_cwr(sk);
1163 err = net_xmit_eval(err);
1164 }
1165 if (!err && oskb) {
1166 tcp_update_skb_after_send(tp, oskb);
1167 tcp_rate_skb_sent(sk, oskb);
1168 }
1169 return err;
1170}
1171
1172static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1173 gfp_t gfp_mask)
1174{
1175 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1176 tcp_sk(sk)->rcv_nxt);
1177}
1178
1179/* This routine just queues the buffer for sending.
1180 *
1181 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1182 * otherwise socket can stall.
1183 */
1184static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1185{
1186 struct tcp_sock *tp = tcp_sk(sk);
1187
1188 /* Advance write_seq and place onto the write_queue. */
1189 tp->write_seq = TCP_SKB_CB(skb)->end_seq;
1190 __skb_header_release(skb);
1191 tcp_add_write_queue_tail(sk, skb);
1192 sk->sk_wmem_queued += skb->truesize;
1193 sk_mem_charge(sk, skb->truesize);
1194}
1195
1196/* Initialize TSO segments for a packet. */
1197static void tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1198{
1199 if (skb->len <= mss_now) {
1200 /* Avoid the costly divide in the normal
1201 * non-TSO case.
1202 */
1203 tcp_skb_pcount_set(skb, 1);
1204 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1205 } else {
1206 tcp_skb_pcount_set(skb, DIV_ROUND_UP(skb->len, mss_now));
1207 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1208 }
1209}
1210
1211/* Pcount in the middle of the write queue got changed, we need to do various
1212 * tweaks to fix counters
1213 */
1214static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1215{
1216 struct tcp_sock *tp = tcp_sk(sk);
1217
1218 tp->packets_out -= decr;
1219
1220 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1221 tp->sacked_out -= decr;
1222 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1223 tp->retrans_out -= decr;
1224 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1225 tp->lost_out -= decr;
1226
1227 /* Reno case is special. Sigh... */
1228 if (tcp_is_reno(tp) && decr > 0)
1229 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1230
1231 if (tp->lost_skb_hint &&
1232 before(TCP_SKB_CB(skb)->seq, TCP_SKB_CB(tp->lost_skb_hint)->seq) &&
1233 (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED))
1234 tp->lost_cnt_hint -= decr;
1235
1236 tcp_verify_left_out(tp);
1237}
1238
1239static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1240{
1241 return TCP_SKB_CB(skb)->txstamp_ack ||
1242 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1243}
1244
1245static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1246{
1247 struct skb_shared_info *shinfo = skb_shinfo(skb);
1248
1249 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1250 !before(shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1251 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1252 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1253
1254 shinfo->tx_flags &= ~tsflags;
1255 shinfo2->tx_flags |= tsflags;
1256 swap(shinfo->tskey, shinfo2->tskey);
1257 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1258 TCP_SKB_CB(skb)->txstamp_ack = 0;
1259 }
1260}
1261
1262static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1263{
1264 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1265 TCP_SKB_CB(skb)->eor = 0;
1266}
1267
1268/* Insert buff after skb on the write or rtx queue of sk. */
1269static void tcp_insert_write_queue_after(struct sk_buff *skb,
1270 struct sk_buff *buff,
1271 struct sock *sk,
1272 enum tcp_queue tcp_queue)
1273{
1274 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1275 __skb_queue_after(&sk->sk_write_queue, skb, buff);
1276 else
1277 tcp_rbtree_insert(&sk->tcp_rtx_queue, buff);
1278}
1279
1280/* Function to create two new TCP segments. Shrinks the given segment
1281 * to the specified size and appends a new segment with the rest of the
1282 * packet to the list. This won't be called frequently, I hope.
1283 * Remember, these are still headerless SKBs at this point.
1284 */
1285int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1286 struct sk_buff *skb, u32 len,
1287 unsigned int mss_now, gfp_t gfp)
1288{
1289 struct tcp_sock *tp = tcp_sk(sk);
1290 struct sk_buff *buff;
1291 int nsize, old_factor;
1292 int nlen;
1293 u8 flags;
1294
1295 if (WARN_ON(len > skb->len))
1296 return -EINVAL;
1297
1298 nsize = skb_headlen(skb) - len;
1299 if (nsize < 0)
1300 nsize = 0;
1301
1302 if (skb_unclone(skb, gfp))
1303 return -ENOMEM;
1304
1305 /* Get a new skb... force flag on. */
1306 buff = sk_stream_alloc_skb(sk, nsize, gfp, true);
1307 if (!buff)
1308 return -ENOMEM; /* We'll just try again later. */
1309
1310 sk->sk_wmem_queued += buff->truesize;
1311 sk_mem_charge(sk, buff->truesize);
1312 nlen = skb->len - len - nsize;
1313 buff->truesize += nlen;
1314 skb->truesize -= nlen;
1315
1316 /* Correct the sequence numbers. */
1317 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1318 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1319 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1320
1321 /* PSH and FIN should only be set in the second packet. */
1322 flags = TCP_SKB_CB(skb)->tcp_flags;
1323 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1324 TCP_SKB_CB(buff)->tcp_flags = flags;
1325 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1326 tcp_skb_fragment_eor(skb, buff);
1327
1328 skb_split(skb, buff, len);
1329
1330 buff->ip_summed = CHECKSUM_PARTIAL;
1331
1332 buff->tstamp = skb->tstamp;
1333 tcp_fragment_tstamp(skb, buff);
1334
1335 old_factor = tcp_skb_pcount(skb);
1336
1337 /* Fix up tso_factor for both original and new SKB. */
1338 tcp_set_skb_tso_segs(skb, mss_now);
1339 tcp_set_skb_tso_segs(buff, mss_now);
1340
1341 /* Update delivered info for the new segment */
1342 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1343
1344 /* If this packet has been sent out already, we must
1345 * adjust the various packet counters.
1346 */
1347 if (!before(tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1348 int diff = old_factor - tcp_skb_pcount(skb) -
1349 tcp_skb_pcount(buff);
1350
1351 if (diff)
1352 tcp_adjust_pcount(sk, skb, diff);
1353 }
1354
1355 /* Link BUFF into the send queue. */
1356 __skb_header_release(buff);
1357 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1358 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1359 list_add(&buff->tcp_tsorted_anchor, &skb->tcp_tsorted_anchor);
1360
1361 return 0;
1362}
1363
1364/* This is similar to __pskb_pull_tail(). The difference is that pulled
1365 * data is not copied, but immediately discarded.
1366 */
1367static int __pskb_trim_head(struct sk_buff *skb, int len)
1368{
1369 struct skb_shared_info *shinfo;
1370 int i, k, eat;
1371
1372 eat = min_t(int, len, skb_headlen(skb));
1373 if (eat) {
1374 __skb_pull(skb, eat);
1375 len -= eat;
1376 if (!len)
1377 return 0;
1378 }
1379 eat = len;
1380 k = 0;
1381 shinfo = skb_shinfo(skb);
1382 for (i = 0; i < shinfo->nr_frags; i++) {
1383 int size = skb_frag_size(&shinfo->frags[i]);
1384
1385 if (size <= eat) {
1386 skb_frag_unref(skb, i);
1387 eat -= size;
1388 } else {
1389 shinfo->frags[k] = shinfo->frags[i];
1390 if (eat) {
1391 shinfo->frags[k].page_offset += eat;
1392 skb_frag_size_sub(&shinfo->frags[k], eat);
1393 eat = 0;
1394 }
1395 k++;
1396 }
1397 }
1398 shinfo->nr_frags = k;
1399
1400 skb->data_len -= len;
1401 skb->len = skb->data_len;
1402 return len;
1403}
1404
1405/* Remove acked data from a packet in the transmit queue. */
1406int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1407{
1408 u32 delta_truesize;
1409
1410 if (skb_unclone(skb, GFP_ATOMIC))
1411 return -ENOMEM;
1412
1413 delta_truesize = __pskb_trim_head(skb, len);
1414
1415 TCP_SKB_CB(skb)->seq += len;
1416 skb->ip_summed = CHECKSUM_PARTIAL;
1417
1418 if (delta_truesize) {
1419 skb->truesize -= delta_truesize;
1420 sk->sk_wmem_queued -= delta_truesize;
1421 sk_mem_uncharge(sk, delta_truesize);
1422 sock_set_flag(sk, SOCK_QUEUE_SHRUNK);
1423 }
1424
1425 /* Any change of skb->len requires recalculation of tso factor. */
1426 if (tcp_skb_pcount(skb) > 1)
1427 tcp_set_skb_tso_segs(skb, tcp_skb_mss(skb));
1428
1429 return 0;
1430}
1431
1432/* Calculate MSS not accounting any TCP options. */
1433static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1434{
1435 const struct tcp_sock *tp = tcp_sk(sk);
1436 const struct inet_connection_sock *icsk = inet_csk(sk);
1437 int mss_now;
1438
1439 /* Calculate base mss without TCP options:
1440 It is MMS_S - sizeof(tcphdr) of rfc1122
1441 */
1442 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1443
1444 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1445 if (icsk->icsk_af_ops->net_frag_header_len) {
1446 const struct dst_entry *dst = __sk_dst_get(sk);
1447
1448 if (dst && dst_allfrag(dst))
1449 mss_now -= icsk->icsk_af_ops->net_frag_header_len;
1450 }
1451
1452 /* Clamp it (mss_clamp does not include tcp options) */
1453 if (mss_now > tp->rx_opt.mss_clamp)
1454 mss_now = tp->rx_opt.mss_clamp;
1455
1456 /* Now subtract optional transport overhead */
1457 mss_now -= icsk->icsk_ext_hdr_len;
1458
1459 /* Then reserve room for full set of TCP options and 8 bytes of data */
1460 if (mss_now < 48)
1461 mss_now = 48;
1462 return mss_now;
1463}
1464
1465/* Calculate MSS. Not accounting for SACKs here. */
1466int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1467{
1468 /* Subtract TCP options size, not including SACKs */
1469 return __tcp_mtu_to_mss(sk, pmtu) -
1470 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1471}
1472
1473/* Inverse of above */
1474int tcp_mss_to_mtu(struct sock *sk, int mss)
1475{
1476 const struct tcp_sock *tp = tcp_sk(sk);
1477 const struct inet_connection_sock *icsk = inet_csk(sk);
1478 int mtu;
1479
1480 mtu = mss +
1481 tp->tcp_header_len +
1482 icsk->icsk_ext_hdr_len +
1483 icsk->icsk_af_ops->net_header_len;
1484
1485 /* IPv6 adds a frag_hdr in case RTAX_FEATURE_ALLFRAG is set */
1486 if (icsk->icsk_af_ops->net_frag_header_len) {
1487 const struct dst_entry *dst = __sk_dst_get(sk);
1488
1489 if (dst && dst_allfrag(dst))
1490 mtu += icsk->icsk_af_ops->net_frag_header_len;
1491 }
1492 return mtu;
1493}
1494EXPORT_SYMBOL(tcp_mss_to_mtu);
1495
1496/* MTU probing init per socket */
1497void tcp_mtup_init(struct sock *sk)
1498{
1499 struct tcp_sock *tp = tcp_sk(sk);
1500 struct inet_connection_sock *icsk = inet_csk(sk);
1501 struct net *net = sock_net(sk);
1502
1503 icsk->icsk_mtup.enabled = net->ipv4.sysctl_tcp_mtu_probing > 1;
1504 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1505 icsk->icsk_af_ops->net_header_len;
1506 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, net->ipv4.sysctl_tcp_base_mss);
1507 icsk->icsk_mtup.probe_size = 0;
1508 if (icsk->icsk_mtup.enabled)
1509 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1510}
1511EXPORT_SYMBOL(tcp_mtup_init);
1512
1513/* This function synchronize snd mss to current pmtu/exthdr set.
1514
1515 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1516 for TCP options, but includes only bare TCP header.
1517
1518 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1519 It is minimum of user_mss and mss received with SYN.
1520 It also does not include TCP options.
1521
1522 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1523
1524 tp->mss_cache is current effective sending mss, including
1525 all tcp options except for SACKs. It is evaluated,
1526 taking into account current pmtu, but never exceeds
1527 tp->rx_opt.mss_clamp.
1528
1529 NOTE1. rfc1122 clearly states that advertised MSS
1530 DOES NOT include either tcp or ip options.
1531
1532 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1533 are READ ONLY outside this function. --ANK (980731)
1534 */
1535unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1536{
1537 struct tcp_sock *tp = tcp_sk(sk);
1538 struct inet_connection_sock *icsk = inet_csk(sk);
1539 int mss_now;
1540
1541 if (icsk->icsk_mtup.search_high > pmtu)
1542 icsk->icsk_mtup.search_high = pmtu;
1543
1544 mss_now = tcp_mtu_to_mss(sk, pmtu);
1545 mss_now = tcp_bound_to_half_wnd(tp, mss_now);
1546
1547 /* And store cached results */
1548 icsk->icsk_pmtu_cookie = pmtu;
1549 if (icsk->icsk_mtup.enabled)
1550 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
1551 tp->mss_cache = mss_now;
1552
1553 return mss_now;
1554}
1555EXPORT_SYMBOL(tcp_sync_mss);
1556
1557/* Compute the current effective MSS, taking SACKs and IP options,
1558 * and even PMTU discovery events into account.
1559 */
1560unsigned int tcp_current_mss(struct sock *sk)
1561{
1562 const struct tcp_sock *tp = tcp_sk(sk);
1563 const struct dst_entry *dst = __sk_dst_get(sk);
1564 u32 mss_now;
1565 unsigned int header_len;
1566 struct tcp_out_options opts;
1567 struct tcp_md5sig_key *md5;
1568
1569 mss_now = tp->mss_cache;
1570
1571 if (dst) {
1572 u32 mtu = dst_mtu(dst);
1573 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
1574 mss_now = tcp_sync_mss(sk, mtu);
1575 }
1576
1577 header_len = tcp_established_options(sk, NULL, &opts, &md5) +
1578 sizeof(struct tcphdr);
1579 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
1580 * some common options. If this is an odd packet (because we have SACK
1581 * blocks etc) then our calculated header_len will be different, and
1582 * we have to adjust mss_now correspondingly */
1583 if (header_len != tp->tcp_header_len) {
1584 int delta = (int) header_len - tp->tcp_header_len;
1585 mss_now -= delta;
1586 }
1587
1588 return mss_now;
1589}
1590
1591/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
1592 * As additional protections, we do not touch cwnd in retransmission phases,
1593 * and if application hit its sndbuf limit recently.
1594 */
1595static void tcp_cwnd_application_limited(struct sock *sk)
1596{
1597 struct tcp_sock *tp = tcp_sk(sk);
1598
1599 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
1600 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
1601 /* Limited by application or receiver window. */
1602 u32 init_win = tcp_init_cwnd(tp, __sk_dst_get(sk));
1603 u32 win_used = max(tp->snd_cwnd_used, init_win);
1604 if (win_used < tp->snd_cwnd) {
1605 tp->snd_ssthresh = tcp_current_ssthresh(sk);
1606 tp->snd_cwnd = (tp->snd_cwnd + win_used) >> 1;
1607 }
1608 tp->snd_cwnd_used = 0;
1609 }
1610 tp->snd_cwnd_stamp = tcp_jiffies32;
1611}
1612
1613static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
1614{
1615 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1616 struct tcp_sock *tp = tcp_sk(sk);
1617
1618 /* Track the maximum number of outstanding packets in each
1619 * window, and remember whether we were cwnd-limited then.
1620 */
1621 if (!before(tp->snd_una, tp->max_packets_seq) ||
1622 tp->packets_out > tp->max_packets_out) {
1623 tp->max_packets_out = tp->packets_out;
1624 tp->max_packets_seq = tp->snd_nxt;
1625 tp->is_cwnd_limited = is_cwnd_limited;
1626 }
1627
1628 if (tcp_is_cwnd_limited(sk)) {
1629 /* Network is feed fully. */
1630 tp->snd_cwnd_used = 0;
1631 tp->snd_cwnd_stamp = tcp_jiffies32;
1632 } else {
1633 /* Network starves. */
1634 if (tp->packets_out > tp->snd_cwnd_used)
1635 tp->snd_cwnd_used = tp->packets_out;
1636
1637 if (sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle &&
1638 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
1639 !ca_ops->cong_control)
1640 tcp_cwnd_application_limited(sk);
1641
1642 /* The following conditions together indicate the starvation
1643 * is caused by insufficient sender buffer:
1644 * 1) just sent some data (see tcp_write_xmit)
1645 * 2) not cwnd limited (this else condition)
1646 * 3) no more data to send (tcp_write_queue_empty())
1647 * 4) application is hitting buffer limit (SOCK_NOSPACE)
1648 */
1649 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
1650 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
1651 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
1652 tcp_chrono_start(sk, TCP_CHRONO_SNDBUF_LIMITED);
1653 }
1654}
1655
1656/* Minshall's variant of the Nagle send check. */
1657static bool tcp_minshall_check(const struct tcp_sock *tp)
1658{
1659 return after(tp->snd_sml, tp->snd_una) &&
1660 !after(tp->snd_sml, tp->snd_nxt);
1661}
1662
1663/* Update snd_sml if this skb is under mss
1664 * Note that a TSO packet might end with a sub-mss segment
1665 * The test is really :
1666 * if ((skb->len % mss) != 0)
1667 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1668 * But we can avoid doing the divide again given we already have
1669 * skb_pcount = skb->len / mss_now
1670 */
1671static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
1672 const struct sk_buff *skb)
1673{
1674 if (skb->len < tcp_skb_pcount(skb) * mss_now)
1675 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
1676}
1677
1678/* Return false, if packet can be sent now without violation Nagle's rules:
1679 * 1. It is full sized. (provided by caller in %partial bool)
1680 * 2. Or it contains FIN. (already checked by caller)
1681 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
1682 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
1683 * With Minshall's modification: all sent small packets are ACKed.
1684 */
1685static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
1686 int nonagle)
1687{
1688 return partial &&
1689 ((nonagle & TCP_NAGLE_CORK) ||
1690 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
1691}
1692
1693/* Return how many segs we'd like on a TSO packet,
1694 * to send one TSO packet per ms
1695 */
1696static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
1697 int min_tso_segs)
1698{
1699 u32 bytes, segs;
1700
1701 bytes = min(sk->sk_pacing_rate >> sk->sk_pacing_shift,
1702 sk->sk_gso_max_size - 1 - MAX_TCP_HEADER);
1703
1704 /* Goal is to send at least one packet per ms,
1705 * not one big TSO packet every 100 ms.
1706 * This preserves ACK clocking and is consistent
1707 * with tcp_tso_should_defer() heuristic.
1708 */
1709 segs = max_t(u32, bytes / mss_now, min_tso_segs);
1710
1711 return segs;
1712}
1713
1714/* Return the number of segments we want in the skb we are transmitting.
1715 * See if congestion control module wants to decide; otherwise, autosize.
1716 */
1717static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
1718{
1719 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
1720 u32 min_tso, tso_segs;
1721
1722 min_tso = ca_ops->min_tso_segs ?
1723 ca_ops->min_tso_segs(sk) :
1724 sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs;
1725
1726 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso);
1727 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
1728}
1729
1730/* Returns the portion of skb which can be sent right away */
1731static unsigned int tcp_mss_split_point(const struct sock *sk,
1732 const struct sk_buff *skb,
1733 unsigned int mss_now,
1734 unsigned int max_segs,
1735 int nonagle)
1736{
1737 const struct tcp_sock *tp = tcp_sk(sk);
1738 u32 partial, needed, window, max_len;
1739
1740 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1741 max_len = mss_now * max_segs;
1742
1743 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
1744 return max_len;
1745
1746 needed = min(skb->len, window);
1747
1748 if (max_len <= needed)
1749 return max_len;
1750
1751 partial = needed % mss_now;
1752 /* If last segment is not a full MSS, check if Nagle rules allow us
1753 * to include this last segment in this skb.
1754 * Otherwise, we'll split the skb at last MSS boundary
1755 */
1756 if (tcp_nagle_check(partial != 0, tp, nonagle))
1757 return needed - partial;
1758
1759 return needed;
1760}
1761
1762/* Can at least one segment of SKB be sent right now, according to the
1763 * congestion window rules? If so, return how many segments are allowed.
1764 */
1765static inline unsigned int tcp_cwnd_test(const struct tcp_sock *tp,
1766 const struct sk_buff *skb)
1767{
1768 u32 in_flight, cwnd, halfcwnd;
1769
1770 /* Don't be strict about the congestion window for the final FIN. */
1771 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) &&
1772 tcp_skb_pcount(skb) == 1)
1773 return 1;
1774
1775 in_flight = tcp_packets_in_flight(tp);
1776 cwnd = tp->snd_cwnd;
1777 if (in_flight >= cwnd)
1778 return 0;
1779
1780 /* For better scheduling, ensure we have at least
1781 * 2 GSO packets in flight.
1782 */
1783 halfcwnd = max(cwnd >> 1, 1U);
1784 return min(halfcwnd, cwnd - in_flight);
1785}
1786
1787/* Initialize TSO state of a skb.
1788 * This must be invoked the first time we consider transmitting
1789 * SKB onto the wire.
1790 */
1791static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1792{
1793 int tso_segs = tcp_skb_pcount(skb);
1794
1795 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now)) {
1796 tcp_set_skb_tso_segs(skb, mss_now);
1797 tso_segs = tcp_skb_pcount(skb);
1798 }
1799 return tso_segs;
1800}
1801
1802
1803/* Return true if the Nagle test allows this packet to be
1804 * sent now.
1805 */
1806static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
1807 unsigned int cur_mss, int nonagle)
1808{
1809 /* Nagle rule does not apply to frames, which sit in the middle of the
1810 * write_queue (they have no chances to get new data).
1811 *
1812 * This is implemented in the callers, where they modify the 'nonagle'
1813 * argument based upon the location of SKB in the send queue.
1814 */
1815 if (nonagle & TCP_NAGLE_PUSH)
1816 return true;
1817
1818 /* Don't use the nagle rule for urgent data (or for the final FIN). */
1819 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
1820 return true;
1821
1822 if (!tcp_nagle_check(skb->len < cur_mss, tp, nonagle))
1823 return true;
1824
1825 return false;
1826}
1827
1828/* Does at least the first segment of SKB fit into the send window? */
1829static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
1830 const struct sk_buff *skb,
1831 unsigned int cur_mss)
1832{
1833 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
1834
1835 if (skb->len > cur_mss)
1836 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
1837
1838 return !after(end_seq, tcp_wnd_end(tp));
1839}
1840
1841/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
1842 * which is put after SKB on the list. It is very much like
1843 * tcp_fragment() except that it may make several kinds of assumptions
1844 * in order to speed up the splitting operation. In particular, we
1845 * know that all the data is in scatter-gather pages, and that the
1846 * packet has never been sent out before (and thus is not cloned).
1847 */
1848static int tso_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1849 struct sk_buff *skb, unsigned int len,
1850 unsigned int mss_now, gfp_t gfp)
1851{
1852 struct sk_buff *buff;
1853 int nlen = skb->len - len;
1854 u8 flags;
1855
1856 /* All of a TSO frame must be composed of paged data. */
1857 if (skb->len != skb->data_len)
1858 return tcp_fragment(sk, tcp_queue, skb, len, mss_now, gfp);
1859
1860 buff = sk_stream_alloc_skb(sk, 0, gfp, true);
1861 if (unlikely(!buff))
1862 return -ENOMEM;
1863
1864 sk->sk_wmem_queued += buff->truesize;
1865 sk_mem_charge(sk, buff->truesize);
1866 buff->truesize += nlen;
1867 skb->truesize -= nlen;
1868
1869 /* Correct the sequence numbers. */
1870 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1871 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1872 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1873
1874 /* PSH and FIN should only be set in the second packet. */
1875 flags = TCP_SKB_CB(skb)->tcp_flags;
1876 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1877 TCP_SKB_CB(buff)->tcp_flags = flags;
1878
1879 /* This packet was never sent out yet, so no SACK bits. */
1880 TCP_SKB_CB(buff)->sacked = 0;
1881
1882 tcp_skb_fragment_eor(skb, buff);
1883
1884 buff->ip_summed = CHECKSUM_PARTIAL;
1885 skb_split(skb, buff, len);
1886 tcp_fragment_tstamp(skb, buff);
1887
1888 /* Fix up tso_factor for both original and new SKB. */
1889 tcp_set_skb_tso_segs(skb, mss_now);
1890 tcp_set_skb_tso_segs(buff, mss_now);
1891
1892 /* Link BUFF into the send queue. */
1893 __skb_header_release(buff);
1894 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1895
1896 return 0;
1897}
1898
1899/* Try to defer sending, if possible, in order to minimize the amount
1900 * of TSO splitting we do. View it as a kind of TSO Nagle test.
1901 *
1902 * This algorithm is from John Heffner.
1903 */
1904static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
1905 bool *is_cwnd_limited, u32 max_segs)
1906{
1907 const struct inet_connection_sock *icsk = inet_csk(sk);
1908 u32 age, send_win, cong_win, limit, in_flight;
1909 struct tcp_sock *tp = tcp_sk(sk);
1910 struct sk_buff *head;
1911 int win_divisor;
1912
1913 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN)
1914 goto send_now;
1915
1916 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
1917 goto send_now;
1918
1919 /* Avoid bursty behavior by allowing defer
1920 * only if the last write was recent.
1921 */
1922 if ((s32)(tcp_jiffies32 - tp->lsndtime) > 0)
1923 goto send_now;
1924
1925 in_flight = tcp_packets_in_flight(tp);
1926
1927 BUG_ON(tcp_skb_pcount(skb) <= 1);
1928 BUG_ON(tp->snd_cwnd <= in_flight);
1929
1930 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
1931
1932 /* From in_flight test above, we know that cwnd > in_flight. */
1933 cong_win = (tp->snd_cwnd - in_flight) * tp->mss_cache;
1934
1935 limit = min(send_win, cong_win);
1936
1937 /* If a full-sized TSO skb can be sent, do it. */
1938 if (limit >= max_segs * tp->mss_cache)
1939 goto send_now;
1940
1941 /* Middle in queue won't get any more data, full sendable already? */
1942 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
1943 goto send_now;
1944
1945 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
1946 if (win_divisor) {
1947 u32 chunk = min(tp->snd_wnd, tp->snd_cwnd * tp->mss_cache);
1948
1949 /* If at least some fraction of a window is available,
1950 * just use it.
1951 */
1952 chunk /= win_divisor;
1953 if (limit >= chunk)
1954 goto send_now;
1955 } else {
1956 /* Different approach, try not to defer past a single
1957 * ACK. Receiver should ACK every other full sized
1958 * frame, so if we have space for more than 3 frames
1959 * then send now.
1960 */
1961 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
1962 goto send_now;
1963 }
1964
1965 /* TODO : use tsorted_sent_queue ? */
1966 head = tcp_rtx_queue_head(sk);
1967 if (!head)
1968 goto send_now;
1969 age = tcp_stamp_us_delta(tp->tcp_mstamp, head->skb_mstamp);
1970 /* If next ACK is likely to come too late (half srtt), do not defer */
1971 if (age < (tp->srtt_us >> 4))
1972 goto send_now;
1973
1974 /* Ok, it looks like it is advisable to defer. */
1975
1976 if (cong_win < send_win && cong_win <= skb->len)
1977 *is_cwnd_limited = true;
1978
1979 return true;
1980
1981send_now:
1982 return false;
1983}
1984
1985static inline void tcp_mtu_check_reprobe(struct sock *sk)
1986{
1987 struct inet_connection_sock *icsk = inet_csk(sk);
1988 struct tcp_sock *tp = tcp_sk(sk);
1989 struct net *net = sock_net(sk);
1990 u32 interval;
1991 s32 delta;
1992
1993 interval = net->ipv4.sysctl_tcp_probe_interval;
1994 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
1995 if (unlikely(delta >= interval * HZ)) {
1996 int mss = tcp_current_mss(sk);
1997
1998 /* Update current search range */
1999 icsk->icsk_mtup.probe_size = 0;
2000 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2001 sizeof(struct tcphdr) +
2002 icsk->icsk_af_ops->net_header_len;
2003 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2004
2005 /* Update probe time stamp */
2006 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2007 }
2008}
2009
2010static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2011{
2012 struct sk_buff *skb, *next;
2013
2014 skb = tcp_send_head(sk);
2015 tcp_for_write_queue_from_safe(skb, next, sk) {
2016 if (len <= skb->len)
2017 break;
2018
2019 if (unlikely(TCP_SKB_CB(skb)->eor))
2020 return false;
2021
2022 len -= skb->len;
2023 }
2024
2025 return true;
2026}
2027
2028/* Create a new MTU probe if we are ready.
2029 * MTU probe is regularly attempting to increase the path MTU by
2030 * deliberately sending larger packets. This discovers routing
2031 * changes resulting in larger path MTUs.
2032 *
2033 * Returns 0 if we should wait to probe (no cwnd available),
2034 * 1 if a probe was sent,
2035 * -1 otherwise
2036 */
2037static int tcp_mtu_probe(struct sock *sk)
2038{
2039 struct inet_connection_sock *icsk = inet_csk(sk);
2040 struct tcp_sock *tp = tcp_sk(sk);
2041 struct sk_buff *skb, *nskb, *next;
2042 struct net *net = sock_net(sk);
2043 int probe_size;
2044 int size_needed;
2045 int copy, len;
2046 int mss_now;
2047 int interval;
2048
2049 /* Not currently probing/verifying,
2050 * not in recovery,
2051 * have enough cwnd, and
2052 * not SACKing (the variable headers throw things off)
2053 */
2054 if (likely(!icsk->icsk_mtup.enabled ||
2055 icsk->icsk_mtup.probe_size ||
2056 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2057 tp->snd_cwnd < 11 ||
2058 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2059 return -1;
2060
2061 /* Use binary search for probe_size between tcp_mss_base,
2062 * and current mss_clamp. if (search_high - search_low)
2063 * smaller than a threshold, backoff from probing.
2064 */
2065 mss_now = tcp_current_mss(sk);
2066 probe_size = tcp_mtu_to_mss(sk, (icsk->icsk_mtup.search_high +
2067 icsk->icsk_mtup.search_low) >> 1);
2068 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2069 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2070 /* When misfortune happens, we are reprobing actively,
2071 * and then reprobe timer has expired. We stick with current
2072 * probing process by not resetting search range to its orignal.
2073 */
2074 if (probe_size > tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_high) ||
2075 interval < net->ipv4.sysctl_tcp_probe_threshold) {
2076 /* Check whether enough time has elaplased for
2077 * another round of probing.
2078 */
2079 tcp_mtu_check_reprobe(sk);
2080 return -1;
2081 }
2082
2083 /* Have enough data in the send queue to probe? */
2084 if (tp->write_seq - tp->snd_nxt < size_needed)
2085 return -1;
2086
2087 if (tp->snd_wnd < size_needed)
2088 return -1;
2089 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2090 return 0;
2091
2092 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2093 if (tcp_packets_in_flight(tp) + 2 > tp->snd_cwnd) {
2094 if (!tcp_packets_in_flight(tp))
2095 return -1;
2096 else
2097 return 0;
2098 }
2099
2100 if (!tcp_can_coalesce_send_queue_head(sk, probe_size))
2101 return -1;
2102
2103 /* We're allowed to probe. Build it now. */
2104 nskb = sk_stream_alloc_skb(sk, probe_size, GFP_ATOMIC, false);
2105 if (!nskb)
2106 return -1;
2107 sk->sk_wmem_queued += nskb->truesize;
2108 sk_mem_charge(sk, nskb->truesize);
2109
2110 skb = tcp_send_head(sk);
2111
2112 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2113 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2114 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2115 TCP_SKB_CB(nskb)->sacked = 0;
2116 nskb->csum = 0;
2117 nskb->ip_summed = CHECKSUM_PARTIAL;
2118
2119 tcp_insert_write_queue_before(nskb, skb, sk);
2120 tcp_highest_sack_replace(sk, skb, nskb);
2121
2122 len = 0;
2123 tcp_for_write_queue_from_safe(skb, next, sk) {
2124 copy = min_t(int, skb->len, probe_size - len);
2125 skb_copy_bits(skb, 0, skb_put(nskb, copy), copy);
2126
2127 if (skb->len <= copy) {
2128 /* We've eaten all the data from this skb.
2129 * Throw it away. */
2130 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags;
2131 /* If this is the last SKB we copy and eor is set
2132 * we need to propagate it to the new skb.
2133 */
2134 TCP_SKB_CB(nskb)->eor = TCP_SKB_CB(skb)->eor;
2135 tcp_unlink_write_queue(skb, sk);
2136 sk_wmem_free_skb(sk, skb);
2137 } else {
2138 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2139 ~(TCPHDR_FIN|TCPHDR_PSH);
2140 if (!skb_shinfo(skb)->nr_frags) {
2141 skb_pull(skb, copy);
2142 } else {
2143 __pskb_trim_head(skb, copy);
2144 tcp_set_skb_tso_segs(skb, mss_now);
2145 }
2146 TCP_SKB_CB(skb)->seq += copy;
2147 }
2148
2149 len += copy;
2150
2151 if (len >= probe_size)
2152 break;
2153 }
2154 tcp_init_tso_segs(nskb, nskb->len);
2155
2156 /* We're ready to send. If this fails, the probe will
2157 * be resegmented into mss-sized pieces by tcp_write_xmit().
2158 */
2159 if (!tcp_transmit_skb(sk, nskb, 1, GFP_ATOMIC)) {
2160 /* Decrement cwnd here because we are sending
2161 * effectively two packets. */
2162 tp->snd_cwnd--;
2163 tcp_event_new_data_sent(sk, nskb);
2164
2165 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2166 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2167 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2168
2169 return 1;
2170 }
2171
2172 return -1;
2173}
2174
2175static bool tcp_pacing_check(const struct sock *sk)
2176{
2177 return tcp_needs_internal_pacing(sk) &&
2178 hrtimer_is_queued(&tcp_sk(sk)->pacing_timer);
2179}
2180
2181/* TCP Small Queues :
2182 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2183 * (These limits are doubled for retransmits)
2184 * This allows for :
2185 * - better RTT estimation and ACK scheduling
2186 * - faster recovery
2187 * - high rates
2188 * Alas, some drivers / subsystems require a fair amount
2189 * of queued bytes to ensure line rate.
2190 * One example is wifi aggregation (802.11 AMPDU)
2191 */
2192static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2193 unsigned int factor)
2194{
2195 unsigned int limit;
2196
2197 limit = max(2 * skb->truesize, sk->sk_pacing_rate >> sk->sk_pacing_shift);
2198 limit = min_t(u32, limit,
2199 sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes);
2200 limit <<= factor;
2201
2202 if (refcount_read(&sk->sk_wmem_alloc) > limit) {
2203 /* Always send skb if rtx queue is empty.
2204 * No need to wait for TX completion to call us back,
2205 * after softirq/tasklet schedule.
2206 * This helps when TX completions are delayed too much.
2207 */
2208 if (tcp_rtx_queue_empty(sk))
2209 return false;
2210
2211 set_bit(TSQ_THROTTLED, &sk->sk_tsq_flags);
2212 /* It is possible TX completion already happened
2213 * before we set TSQ_THROTTLED, so we must
2214 * test again the condition.
2215 */
2216 smp_mb__after_atomic();
2217 if (refcount_read(&sk->sk_wmem_alloc) > limit)
2218 return true;
2219 }
2220 return false;
2221}
2222
2223static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2224{
2225 const u32 now = tcp_jiffies32;
2226 enum tcp_chrono old = tp->chrono_type;
2227
2228 if (old > TCP_CHRONO_UNSPEC)
2229 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2230 tp->chrono_start = now;
2231 tp->chrono_type = new;
2232}
2233
2234void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2235{
2236 struct tcp_sock *tp = tcp_sk(sk);
2237
2238 /* If there are multiple conditions worthy of tracking in a
2239 * chronograph then the highest priority enum takes precedence
2240 * over the other conditions. So that if something "more interesting"
2241 * starts happening, stop the previous chrono and start a new one.
2242 */
2243 if (type > tp->chrono_type)
2244 tcp_chrono_set(tp, type);
2245}
2246
2247void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2248{
2249 struct tcp_sock *tp = tcp_sk(sk);
2250
2251
2252 /* There are multiple conditions worthy of tracking in a
2253 * chronograph, so that the highest priority enum takes
2254 * precedence over the other conditions (see tcp_chrono_start).
2255 * If a condition stops, we only stop chrono tracking if
2256 * it's the "most interesting" or current chrono we are
2257 * tracking and starts busy chrono if we have pending data.
2258 */
2259 if (tcp_rtx_and_write_queues_empty(sk))
2260 tcp_chrono_set(tp, TCP_CHRONO_UNSPEC);
2261 else if (type == tp->chrono_type)
2262 tcp_chrono_set(tp, TCP_CHRONO_BUSY);
2263}
2264
2265/* This routine writes packets to the network. It advances the
2266 * send_head. This happens as incoming acks open up the remote
2267 * window for us.
2268 *
2269 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2270 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2271 * account rare use of URG, this is not a big flaw.
2272 *
2273 * Send at most one packet when push_one > 0. Temporarily ignore
2274 * cwnd limit to force at most one packet out when push_one == 2.
2275
2276 * Returns true, if no segments are in flight and we have queued segments,
2277 * but cannot send anything now because of SWS or another problem.
2278 */
2279static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2280 int push_one, gfp_t gfp)
2281{
2282 struct tcp_sock *tp = tcp_sk(sk);
2283 struct sk_buff *skb;
2284 unsigned int tso_segs, sent_pkts;
2285 int cwnd_quota;
2286 int result;
2287 bool is_cwnd_limited = false, is_rwnd_limited = false;
2288 u32 max_segs;
2289
2290 sent_pkts = 0;
2291
2292 tcp_mstamp_refresh(tp);
2293 if (!push_one) {
2294 /* Do MTU probing. */
2295 result = tcp_mtu_probe(sk);
2296 if (!result) {
2297 return false;
2298 } else if (result > 0) {
2299 sent_pkts = 1;
2300 }
2301 }
2302
2303 max_segs = tcp_tso_segs(sk, mss_now);
2304 while ((skb = tcp_send_head(sk))) {
2305 unsigned int limit;
2306
2307 if (tcp_pacing_check(sk))
2308 break;
2309
2310 tso_segs = tcp_init_tso_segs(skb, mss_now);
2311 BUG_ON(!tso_segs);
2312
2313 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2314 /* "skb_mstamp" is used as a start point for the retransmit timer */
2315 tcp_update_skb_after_send(tp, skb);
2316 goto repair; /* Skip network transmission */
2317 }
2318
2319 cwnd_quota = tcp_cwnd_test(tp, skb);
2320 if (!cwnd_quota) {
2321 if (push_one == 2)
2322 /* Force out a loss probe pkt. */
2323 cwnd_quota = 1;
2324 else
2325 break;
2326 }
2327
2328 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2329 is_rwnd_limited = true;
2330 break;
2331 }
2332
2333 if (tso_segs == 1) {
2334 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2335 (tcp_skb_is_last(sk, skb) ?
2336 nonagle : TCP_NAGLE_PUSH))))
2337 break;
2338 } else {
2339 if (!push_one &&
2340 tcp_tso_should_defer(sk, skb, &is_cwnd_limited,
2341 max_segs))
2342 break;
2343 }
2344
2345 limit = mss_now;
2346 if (tso_segs > 1 && !tcp_urg_mode(tp))
2347 limit = tcp_mss_split_point(sk, skb, mss_now,
2348 min_t(unsigned int,
2349 cwnd_quota,
2350 max_segs),
2351 nonagle);
2352
2353 if (skb->len > limit &&
2354 unlikely(tso_fragment(sk, TCP_FRAG_IN_WRITE_QUEUE,
2355 skb, limit, mss_now, gfp)))
2356 break;
2357
2358 if (tcp_small_queue_check(sk, skb, 0))
2359 break;
2360
2361 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2362 break;
2363
2364repair:
2365 /* Advance the send_head. This one is sent out.
2366 * This call will increment packets_out.
2367 */
2368 tcp_event_new_data_sent(sk, skb);
2369
2370 tcp_minshall_update(tp, mss_now, skb);
2371 sent_pkts += tcp_skb_pcount(skb);
2372
2373 if (push_one)
2374 break;
2375 }
2376
2377 if (is_rwnd_limited)
2378 tcp_chrono_start(sk, TCP_CHRONO_RWND_LIMITED);
2379 else
2380 tcp_chrono_stop(sk, TCP_CHRONO_RWND_LIMITED);
2381
2382 if (likely(sent_pkts)) {
2383 if (tcp_in_cwnd_reduction(sk))
2384 tp->prr_out += sent_pkts;
2385
2386 /* Send one loss probe per tail loss episode. */
2387 if (push_one != 2)
2388 tcp_schedule_loss_probe(sk, false);
2389 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tp->snd_cwnd);
2390 tcp_cwnd_validate(sk, is_cwnd_limited);
2391 return false;
2392 }
2393 return !tp->packets_out && !tcp_write_queue_empty(sk);
2394}
2395
2396bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
2397{
2398 struct inet_connection_sock *icsk = inet_csk(sk);
2399 struct tcp_sock *tp = tcp_sk(sk);
2400 u32 timeout, rto_delta_us;
2401 int early_retrans;
2402
2403 /* Don't do any loss probe on a Fast Open connection before 3WHS
2404 * finishes.
2405 */
2406 if (tp->fastopen_rsk)
2407 return false;
2408
2409 early_retrans = sock_net(sk)->ipv4.sysctl_tcp_early_retrans;
2410 /* Schedule a loss probe in 2*RTT for SACK capable connections
2411 * not in loss recovery, that are either limited by cwnd or application.
2412 */
2413 if ((early_retrans != 3 && early_retrans != 4) ||
2414 !tp->packets_out || !tcp_is_sack(tp) ||
2415 (icsk->icsk_ca_state != TCP_CA_Open &&
2416 icsk->icsk_ca_state != TCP_CA_CWR))
2417 return false;
2418
2419 /* Probe timeout is 2*rtt. Add minimum RTO to account
2420 * for delayed ack when there's one outstanding packet. If no RTT
2421 * sample is available then probe after TCP_TIMEOUT_INIT.
2422 */
2423 if (tp->srtt_us) {
2424 timeout = usecs_to_jiffies(tp->srtt_us >> 2);
2425 if (tp->packets_out == 1)
2426 timeout += TCP_RTO_MIN;
2427 else
2428 timeout += TCP_TIMEOUT_MIN;
2429 } else {
2430 timeout = TCP_TIMEOUT_INIT;
2431 }
2432
2433 /* If the RTO formula yields an earlier time, then use that time. */
2434 rto_delta_us = advancing_rto ?
2435 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
2436 tcp_rto_delta_us(sk); /* How far in future is RTO? */
2437 if (rto_delta_us > 0)
2438 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
2439
2440 inet_csk_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, timeout,
2441 TCP_RTO_MAX);
2442 return true;
2443}
2444
2445/* Thanks to skb fast clones, we can detect if a prior transmit of
2446 * a packet is still in a qdisc or driver queue.
2447 * In this case, there is very little point doing a retransmit !
2448 */
2449static bool skb_still_in_host_queue(const struct sock *sk,
2450 const struct sk_buff *skb)
2451{
2452 if (unlikely(skb_fclone_busy(sk, skb))) {
2453 NET_INC_STATS(sock_net(sk),
2454 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
2455 return true;
2456 }
2457 return false;
2458}
2459
2460/* When probe timeout (PTO) fires, try send a new segment if possible, else
2461 * retransmit the last segment.
2462 */
2463void tcp_send_loss_probe(struct sock *sk)
2464{
2465 struct tcp_sock *tp = tcp_sk(sk);
2466 struct sk_buff *skb;
2467 int pcount;
2468 int mss = tcp_current_mss(sk);
2469
2470 skb = tcp_send_head(sk);
2471 if (skb && tcp_snd_wnd_test(tp, skb, mss)) {
2472 pcount = tp->packets_out;
2473 tcp_write_xmit(sk, mss, TCP_NAGLE_OFF, 2, GFP_ATOMIC);
2474 if (tp->packets_out > pcount)
2475 goto probe_sent;
2476 goto rearm_timer;
2477 }
2478 skb = skb_rb_last(&sk->tcp_rtx_queue);
2479
2480 /* At most one outstanding TLP retransmission. */
2481 if (tp->tlp_high_seq)
2482 goto rearm_timer;
2483
2484 /* Retransmit last segment. */
2485 if (WARN_ON(!skb))
2486 goto rearm_timer;
2487
2488 if (skb_still_in_host_queue(sk, skb))
2489 goto rearm_timer;
2490
2491 pcount = tcp_skb_pcount(skb);
2492 if (WARN_ON(!pcount))
2493 goto rearm_timer;
2494
2495 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
2496 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
2497 (pcount - 1) * mss, mss,
2498 GFP_ATOMIC)))
2499 goto rearm_timer;
2500 skb = skb_rb_next(skb);
2501 }
2502
2503 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
2504 goto rearm_timer;
2505
2506 if (__tcp_retransmit_skb(sk, skb, 1))
2507 goto rearm_timer;
2508
2509 /* Record snd_nxt for loss detection. */
2510 tp->tlp_high_seq = tp->snd_nxt;
2511
2512probe_sent:
2513 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
2514 /* Reset s.t. tcp_rearm_rto will restart timer from now */
2515 inet_csk(sk)->icsk_pending = 0;
2516rearm_timer:
2517 tcp_rearm_rto(sk);
2518}
2519
2520/* Push out any pending frames which were held back due to
2521 * TCP_CORK or attempt at coalescing tiny packets.
2522 * The socket must be locked by the caller.
2523 */
2524void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
2525 int nonagle)
2526{
2527 /* If we are closed, the bytes will have to remain here.
2528 * In time closedown will finish, we empty the write queue and
2529 * all will be happy.
2530 */
2531 if (unlikely(sk->sk_state == TCP_CLOSE))
2532 return;
2533
2534 if (tcp_write_xmit(sk, cur_mss, nonagle, 0,
2535 sk_gfp_mask(sk, GFP_ATOMIC)))
2536 tcp_check_probe_timer(sk);
2537}
2538
2539/* Send _single_ skb sitting at the send head. This function requires
2540 * true push pending frames to setup probe timer etc.
2541 */
2542void tcp_push_one(struct sock *sk, unsigned int mss_now)
2543{
2544 struct sk_buff *skb = tcp_send_head(sk);
2545
2546 BUG_ON(!skb || skb->len < mss_now);
2547
2548 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, 1, sk->sk_allocation);
2549}
2550
2551/* This function returns the amount that we can raise the
2552 * usable window based on the following constraints
2553 *
2554 * 1. The window can never be shrunk once it is offered (RFC 793)
2555 * 2. We limit memory per socket
2556 *
2557 * RFC 1122:
2558 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
2559 * RECV.NEXT + RCV.WIN fixed until:
2560 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
2561 *
2562 * i.e. don't raise the right edge of the window until you can raise
2563 * it at least MSS bytes.
2564 *
2565 * Unfortunately, the recommended algorithm breaks header prediction,
2566 * since header prediction assumes th->window stays fixed.
2567 *
2568 * Strictly speaking, keeping th->window fixed violates the receiver
2569 * side SWS prevention criteria. The problem is that under this rule
2570 * a stream of single byte packets will cause the right side of the
2571 * window to always advance by a single byte.
2572 *
2573 * Of course, if the sender implements sender side SWS prevention
2574 * then this will not be a problem.
2575 *
2576 * BSD seems to make the following compromise:
2577 *
2578 * If the free space is less than the 1/4 of the maximum
2579 * space available and the free space is less than 1/2 mss,
2580 * then set the window to 0.
2581 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
2582 * Otherwise, just prevent the window from shrinking
2583 * and from being larger than the largest representable value.
2584 *
2585 * This prevents incremental opening of the window in the regime
2586 * where TCP is limited by the speed of the reader side taking
2587 * data out of the TCP receive queue. It does nothing about
2588 * those cases where the window is constrained on the sender side
2589 * because the pipeline is full.
2590 *
2591 * BSD also seems to "accidentally" limit itself to windows that are a
2592 * multiple of MSS, at least until the free space gets quite small.
2593 * This would appear to be a side effect of the mbuf implementation.
2594 * Combining these two algorithms results in the observed behavior
2595 * of having a fixed window size at almost all times.
2596 *
2597 * Below we obtain similar behavior by forcing the offered window to
2598 * a multiple of the mss when it is feasible to do so.
2599 *
2600 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
2601 * Regular options like TIMESTAMP are taken into account.
2602 */
2603u32 __tcp_select_window(struct sock *sk)
2604{
2605 struct inet_connection_sock *icsk = inet_csk(sk);
2606 struct tcp_sock *tp = tcp_sk(sk);
2607 /* MSS for the peer's data. Previous versions used mss_clamp
2608 * here. I don't know if the value based on our guesses
2609 * of peer's MSS is better for the performance. It's more correct
2610 * but may be worse for the performance because of rcv_mss
2611 * fluctuations. --SAW 1998/11/1
2612 */
2613 int mss = icsk->icsk_ack.rcv_mss;
2614 int free_space = tcp_space(sk);
2615 int allowed_space = tcp_full_space(sk);
2616 int full_space = min_t(int, tp->window_clamp, allowed_space);
2617 int window;
2618
2619 if (unlikely(mss > full_space)) {
2620 mss = full_space;
2621 if (mss <= 0)
2622 return 0;
2623 }
2624 if (free_space < (full_space >> 1)) {
2625 icsk->icsk_ack.quick = 0;
2626
2627 if (tcp_under_memory_pressure(sk))
2628 tp->rcv_ssthresh = min(tp->rcv_ssthresh,
2629 4U * tp->advmss);
2630
2631 /* free_space might become our new window, make sure we don't
2632 * increase it due to wscale.
2633 */
2634 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
2635
2636 /* if free space is less than mss estimate, or is below 1/16th
2637 * of the maximum allowed, try to move to zero-window, else
2638 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
2639 * new incoming data is dropped due to memory limits.
2640 * With large window, mss test triggers way too late in order
2641 * to announce zero window in time before rmem limit kicks in.
2642 */
2643 if (free_space < (allowed_space >> 4) || free_space < mss)
2644 return 0;
2645 }
2646
2647 if (free_space > tp->rcv_ssthresh)
2648 free_space = tp->rcv_ssthresh;
2649
2650 /* Don't do rounding if we are using window scaling, since the
2651 * scaled window will not line up with the MSS boundary anyway.
2652 */
2653 if (tp->rx_opt.rcv_wscale) {
2654 window = free_space;
2655
2656 /* Advertise enough space so that it won't get scaled away.
2657 * Import case: prevent zero window announcement if
2658 * 1<<rcv_wscale > mss.
2659 */
2660 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
2661 } else {
2662 window = tp->rcv_wnd;
2663 /* Get the largest window that is a nice multiple of mss.
2664 * Window clamp already applied above.
2665 * If our current window offering is within 1 mss of the
2666 * free space we just keep it. This prevents the divide
2667 * and multiply from happening most of the time.
2668 * We also don't do any window rounding when the free space
2669 * is too small.
2670 */
2671 if (window <= free_space - mss || window > free_space)
2672 window = rounddown(free_space, mss);
2673 else if (mss == full_space &&
2674 free_space > window + (full_space >> 1))
2675 window = free_space;
2676 }
2677
2678 return window;
2679}
2680
2681void tcp_skb_collapse_tstamp(struct sk_buff *skb,
2682 const struct sk_buff *next_skb)
2683{
2684 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
2685 const struct skb_shared_info *next_shinfo =
2686 skb_shinfo(next_skb);
2687 struct skb_shared_info *shinfo = skb_shinfo(skb);
2688
2689 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
2690 shinfo->tskey = next_shinfo->tskey;
2691 TCP_SKB_CB(skb)->txstamp_ack |=
2692 TCP_SKB_CB(next_skb)->txstamp_ack;
2693 }
2694}
2695
2696/* Collapses two adjacent SKB's during retransmission. */
2697static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
2698{
2699 struct tcp_sock *tp = tcp_sk(sk);
2700 struct sk_buff *next_skb = skb_rb_next(skb);
2701 int next_skb_size;
2702
2703 next_skb_size = next_skb->len;
2704
2705 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
2706
2707 if (next_skb_size) {
2708 if (next_skb_size <= skb_availroom(skb))
2709 skb_copy_bits(next_skb, 0, skb_put(skb, next_skb_size),
2710 next_skb_size);
2711 else if (!skb_shift(skb, next_skb, next_skb_size))
2712 return false;
2713 }
2714 tcp_highest_sack_replace(sk, next_skb, skb);
2715
2716 /* Update sequence range on original skb. */
2717 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2718
2719 /* Merge over control information. This moves PSH/FIN etc. over */
2720 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
2721
2722 /* All done, get rid of second SKB and account for it so
2723 * packet counting does not break.
2724 */
2725 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
2726 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
2727
2728 /* changed transmit queue under us so clear hints */
2729 tcp_clear_retrans_hints_partial(tp);
2730 if (next_skb == tp->retransmit_skb_hint)
2731 tp->retransmit_skb_hint = skb;
2732
2733 tcp_adjust_pcount(sk, next_skb, tcp_skb_pcount(next_skb));
2734
2735 tcp_skb_collapse_tstamp(skb, next_skb);
2736
2737 tcp_rtx_queue_unlink_and_free(next_skb, sk);
2738 return true;
2739}
2740
2741/* Check if coalescing SKBs is legal. */
2742static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
2743{
2744 if (tcp_skb_pcount(skb) > 1)
2745 return false;
2746 if (skb_cloned(skb))
2747 return false;
2748 /* Some heuristics for collapsing over SACK'd could be invented */
2749 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
2750 return false;
2751
2752 return true;
2753}
2754
2755/* Collapse packets in the retransmit queue to make to create
2756 * less packets on the wire. This is only done on retransmission.
2757 */
2758static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
2759 int space)
2760{
2761 struct tcp_sock *tp = tcp_sk(sk);
2762 struct sk_buff *skb = to, *tmp;
2763 bool first = true;
2764
2765 if (!sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse)
2766 return;
2767 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2768 return;
2769
2770 skb_rbtree_walk_from_safe(skb, tmp) {
2771 if (!tcp_can_collapse(sk, skb))
2772 break;
2773
2774 if (!tcp_skb_can_collapse_to(to))
2775 break;
2776
2777 space -= skb->len;
2778
2779 if (first) {
2780 first = false;
2781 continue;
2782 }
2783
2784 if (space < 0)
2785 break;
2786
2787 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
2788 break;
2789
2790 if (!tcp_collapse_retrans(sk, to))
2791 break;
2792 }
2793}
2794
2795/* This retransmits one SKB. Policy decisions and retransmit queue
2796 * state updates are done by the caller. Returns non-zero if an
2797 * error occurred which prevented the send.
2798 */
2799int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2800{
2801 struct inet_connection_sock *icsk = inet_csk(sk);
2802 struct tcp_sock *tp = tcp_sk(sk);
2803 unsigned int cur_mss;
2804 int diff, len, err;
2805
2806
2807 /* Inconclusive MTU probe */
2808 if (icsk->icsk_mtup.probe_size)
2809 icsk->icsk_mtup.probe_size = 0;
2810
2811 /* Do not sent more than we queued. 1/4 is reserved for possible
2812 * copying overhead: fragmentation, tunneling, mangling etc.
2813 */
2814 if (refcount_read(&sk->sk_wmem_alloc) >
2815 min_t(u32, sk->sk_wmem_queued + (sk->sk_wmem_queued >> 2),
2816 sk->sk_sndbuf))
2817 return -EAGAIN;
2818
2819 if (skb_still_in_host_queue(sk, skb))
2820 return -EBUSY;
2821
2822 if (before(TCP_SKB_CB(skb)->seq, tp->snd_una)) {
2823 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
2824 WARN_ON_ONCE(1);
2825 return -EINVAL;
2826 }
2827 if (tcp_trim_head(sk, skb, tp->snd_una - TCP_SKB_CB(skb)->seq))
2828 return -ENOMEM;
2829 }
2830
2831 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
2832 return -EHOSTUNREACH; /* Routing failure or similar. */
2833
2834 cur_mss = tcp_current_mss(sk);
2835
2836 /* If receiver has shrunk his window, and skb is out of
2837 * new window, do not retransmit it. The exception is the
2838 * case, when window is shrunk to zero. In this case
2839 * our retransmit serves as a zero window probe.
2840 */
2841 if (!before(TCP_SKB_CB(skb)->seq, tcp_wnd_end(tp)) &&
2842 TCP_SKB_CB(skb)->seq != tp->snd_una)
2843 return -EAGAIN;
2844
2845 len = cur_mss * segs;
2846 if (skb->len > len) {
2847 if (tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb, len,
2848 cur_mss, GFP_ATOMIC))
2849 return -ENOMEM; /* We'll try again later. */
2850 } else {
2851 if (skb_unclone(skb, GFP_ATOMIC))
2852 return -ENOMEM;
2853
2854 diff = tcp_skb_pcount(skb);
2855 tcp_set_skb_tso_segs(skb, cur_mss);
2856 diff -= tcp_skb_pcount(skb);
2857 if (diff)
2858 tcp_adjust_pcount(sk, skb, diff);
2859 if (skb->len < cur_mss)
2860 tcp_retrans_try_collapse(sk, skb, cur_mss);
2861 }
2862
2863 /* RFC3168, section 6.1.1.1. ECN fallback */
2864 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
2865 tcp_ecn_clear_syn(sk, skb);
2866
2867 /* Update global and local TCP statistics. */
2868 segs = tcp_skb_pcount(skb);
2869 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
2870 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
2871 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
2872 tp->total_retrans += segs;
2873 tp->bytes_retrans += skb->len;
2874
2875 /* make sure skb->data is aligned on arches that require it
2876 * and check if ack-trimming & collapsing extended the headroom
2877 * beyond what csum_start can cover.
2878 */
2879 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
2880 skb_headroom(skb) >= 0xFFFF)) {
2881 struct sk_buff *nskb;
2882
2883 tcp_skb_tsorted_save(skb) {
2884 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
2885 err = nskb ? tcp_transmit_skb(sk, nskb, 0, GFP_ATOMIC) :
2886 -ENOBUFS;
2887 } tcp_skb_tsorted_restore(skb);
2888
2889 if (!err) {
2890 tcp_update_skb_after_send(tp, skb);
2891 tcp_rate_skb_sent(sk, skb);
2892 }
2893 } else {
2894 err = tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
2895 }
2896
2897 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
2898 tcp_call_bpf_3arg(sk, BPF_SOCK_OPS_RETRANS_CB,
2899 TCP_SKB_CB(skb)->seq, segs, err);
2900
2901 if (likely(!err)) {
2902 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
2903 trace_tcp_retransmit_skb(sk, skb);
2904 } else if (err != -EBUSY) {
2905 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL);
2906 }
2907 return err;
2908}
2909
2910int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
2911{
2912 struct tcp_sock *tp = tcp_sk(sk);
2913 int err = __tcp_retransmit_skb(sk, skb, segs);
2914
2915 if (err == 0) {
2916#if FASTRETRANS_DEBUG > 0
2917 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
2918 net_dbg_ratelimited("retrans_out leaked\n");
2919 }
2920#endif
2921 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
2922 tp->retrans_out += tcp_skb_pcount(skb);
2923
2924 /* Save stamp of the first retransmit. */
2925 if (!tp->retrans_stamp)
2926 tp->retrans_stamp = tcp_skb_timestamp(skb);
2927
2928 }
2929
2930 if (tp->undo_retrans < 0)
2931 tp->undo_retrans = 0;
2932 tp->undo_retrans += tcp_skb_pcount(skb);
2933 return err;
2934}
2935
2936/* This gets called after a retransmit timeout, and the initially
2937 * retransmitted data is acknowledged. It tries to continue
2938 * resending the rest of the retransmit queue, until either
2939 * we've sent it all or the congestion window limit is reached.
2940 */
2941void tcp_xmit_retransmit_queue(struct sock *sk)
2942{
2943 const struct inet_connection_sock *icsk = inet_csk(sk);
2944 struct sk_buff *skb, *rtx_head, *hole = NULL;
2945 struct tcp_sock *tp = tcp_sk(sk);
2946 u32 max_segs;
2947 int mib_idx;
2948
2949 if (!tp->packets_out)
2950 return;
2951
2952 rtx_head = tcp_rtx_queue_head(sk);
2953 skb = tp->retransmit_skb_hint ?: rtx_head;
2954 max_segs = tcp_tso_segs(sk, tcp_current_mss(sk));
2955 skb_rbtree_walk_from(skb) {
2956 __u8 sacked;
2957 int segs;
2958
2959 if (tcp_pacing_check(sk))
2960 break;
2961
2962 /* we could do better than to assign each time */
2963 if (!hole)
2964 tp->retransmit_skb_hint = skb;
2965
2966 segs = tp->snd_cwnd - tcp_packets_in_flight(tp);
2967 if (segs <= 0)
2968 return;
2969 sacked = TCP_SKB_CB(skb)->sacked;
2970 /* In case tcp_shift_skb_data() have aggregated large skbs,
2971 * we need to make sure not sending too bigs TSO packets
2972 */
2973 segs = min_t(int, segs, max_segs);
2974
2975 if (tp->retrans_out >= tp->lost_out) {
2976 break;
2977 } else if (!(sacked & TCPCB_LOST)) {
2978 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
2979 hole = skb;
2980 continue;
2981
2982 } else {
2983 if (icsk->icsk_ca_state != TCP_CA_Loss)
2984 mib_idx = LINUX_MIB_TCPFASTRETRANS;
2985 else
2986 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
2987 }
2988
2989 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
2990 continue;
2991
2992 if (tcp_small_queue_check(sk, skb, 1))
2993 return;
2994
2995 if (tcp_retransmit_skb(sk, skb, segs))
2996 return;
2997
2998 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
2999
3000 if (tcp_in_cwnd_reduction(sk))
3001 tp->prr_out += tcp_skb_pcount(skb);
3002
3003 if (skb == rtx_head &&
3004 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3005 inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3006 inet_csk(sk)->icsk_rto,
3007 TCP_RTO_MAX);
3008 }
3009}
3010
3011/* We allow to exceed memory limits for FIN packets to expedite
3012 * connection tear down and (memory) recovery.
3013 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3014 * or even be forced to close flow without any FIN.
3015 * In general, we want to allow one skb per socket to avoid hangs
3016 * with edge trigger epoll()
3017 */
3018void sk_forced_mem_schedule(struct sock *sk, int size)
3019{
3020 int amt;
3021
3022 if (size <= sk->sk_forward_alloc)
3023 return;
3024 amt = sk_mem_pages(size);
3025 sk->sk_forward_alloc += amt * SK_MEM_QUANTUM;
3026 sk_memory_allocated_add(sk, amt);
3027
3028 if (mem_cgroup_sockets_enabled && sk->sk_memcg)
3029 mem_cgroup_charge_skmem(sk->sk_memcg, amt);
3030}
3031
3032/* Send a FIN. The caller locks the socket for us.
3033 * We should try to send a FIN packet really hard, but eventually give up.
3034 */
3035void tcp_send_fin(struct sock *sk)
3036{
3037 struct sk_buff *skb, *tskb = tcp_write_queue_tail(sk);
3038 struct tcp_sock *tp = tcp_sk(sk);
3039
3040 /* Optimization, tack on the FIN if we have one skb in write queue and
3041 * this skb was not yet sent, or we are under memory pressure.
3042 * Note: in the latter case, FIN packet will be sent after a timeout,
3043 * as TCP stack thinks it has already been transmitted.
3044 */
3045 if (!tskb && tcp_under_memory_pressure(sk))
3046 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3047
3048 if (tskb) {
3049coalesce:
3050 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3051 TCP_SKB_CB(tskb)->end_seq++;
3052 tp->write_seq++;
3053 if (tcp_write_queue_empty(sk)) {
3054 /* This means tskb was already sent.
3055 * Pretend we included the FIN on previous transmit.
3056 * We need to set tp->snd_nxt to the value it would have
3057 * if FIN had been sent. This is because retransmit path
3058 * does not change tp->snd_nxt.
3059 */
3060 tp->snd_nxt++;
3061 return;
3062 }
3063 } else {
3064 skb = alloc_skb_fclone(MAX_TCP_HEADER, sk->sk_allocation);
3065 if (unlikely(!skb)) {
3066 if (tskb)
3067 goto coalesce;
3068 return;
3069 }
3070 INIT_LIST_HEAD(&skb->tcp_tsorted_anchor);
3071 skb_reserve(skb, MAX_TCP_HEADER);
3072 sk_forced_mem_schedule(sk, skb->truesize);
3073 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3074 tcp_init_nondata_skb(skb, tp->write_seq,
3075 TCPHDR_ACK | TCPHDR_FIN);
3076 tcp_queue_skb(sk, skb);
3077 }
3078 __tcp_push_pending_frames(sk, tcp_current_mss(sk), TCP_NAGLE_OFF);
3079}
3080
3081/* We get here when a process closes a file descriptor (either due to
3082 * an explicit close() or as a byproduct of exit()'ing) and there
3083 * was unread data in the receive queue. This behavior is recommended
3084 * by RFC 2525, section 2.17. -DaveM
3085 */
3086void tcp_send_active_reset(struct sock *sk, gfp_t priority)
3087{
3088 struct sk_buff *skb;
3089
3090 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3091
3092 /* NOTE: No TCP options attached and we never retransmit this. */
3093 skb = alloc_skb(MAX_TCP_HEADER, priority);
3094 if (!skb) {
3095 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3096 return;
3097 }
3098
3099 /* Reserve space for headers and prepare control bits. */
3100 skb_reserve(skb, MAX_TCP_HEADER);
3101 tcp_init_nondata_skb(skb, tcp_acceptable_seq(sk),
3102 TCPHDR_ACK | TCPHDR_RST);
3103 tcp_mstamp_refresh(tcp_sk(sk));
3104 /* Send it off. */
3105 if (tcp_transmit_skb(sk, skb, 0, priority))
3106 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3107
3108 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3109 * skb here is different to the troublesome skb, so use NULL
3110 */
3111 trace_tcp_send_reset(sk, NULL);
3112}
3113
3114/* Send a crossed SYN-ACK during socket establishment.
3115 * WARNING: This routine must only be called when we have already sent
3116 * a SYN packet that crossed the incoming SYN that caused this routine
3117 * to get called. If this assumption fails then the initial rcv_wnd
3118 * and rcv_wscale values will not be correct.
3119 */
3120int tcp_send_synack(struct sock *sk)
3121{
3122 struct sk_buff *skb;
3123
3124 skb = tcp_rtx_queue_head(sk);
3125 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3126 pr_err("%s: wrong queue state\n", __func__);
3127 return -EFAULT;
3128 }
3129 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3130 if (skb_cloned(skb)) {
3131 struct sk_buff *nskb;
3132
3133 tcp_skb_tsorted_save(skb) {
3134 nskb = skb_copy(skb, GFP_ATOMIC);
3135 } tcp_skb_tsorted_restore(skb);
3136 if (!nskb)
3137 return -ENOMEM;
3138 INIT_LIST_HEAD(&nskb->tcp_tsorted_anchor);
3139 tcp_rtx_queue_unlink_and_free(skb, sk);
3140 __skb_header_release(nskb);
3141 tcp_rbtree_insert(&sk->tcp_rtx_queue, nskb);
3142 sk->sk_wmem_queued += nskb->truesize;
3143 sk_mem_charge(sk, nskb->truesize);
3144 skb = nskb;
3145 }
3146
3147 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3148 tcp_ecn_send_synack(sk, skb);
3149 }
3150 return tcp_transmit_skb(sk, skb, 1, GFP_ATOMIC);
3151}
3152
3153/**
3154 * tcp_make_synack - Prepare a SYN-ACK.
3155 * sk: listener socket
3156 * dst: dst entry attached to the SYNACK
3157 * req: request_sock pointer
3158 *
3159 * Allocate one skb and build a SYNACK packet.
3160 * @dst is consumed : Caller should not use it again.
3161 */
3162struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3163 struct request_sock *req,
3164 struct tcp_fastopen_cookie *foc,
3165 enum tcp_synack_type synack_type)
3166{
3167 struct inet_request_sock *ireq = inet_rsk(req);
3168 const struct tcp_sock *tp = tcp_sk(sk);
3169 struct tcp_md5sig_key *md5 = NULL;
3170 struct tcp_out_options opts;
3171 struct sk_buff *skb;
3172 int tcp_header_size;
3173 struct tcphdr *th;
3174 int mss;
3175
3176 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3177 if (unlikely(!skb)) {
3178 dst_release(dst);
3179 return NULL;
3180 }
3181 /* Reserve space for headers. */
3182 skb_reserve(skb, MAX_TCP_HEADER);
3183
3184 switch (synack_type) {
3185 case TCP_SYNACK_NORMAL:
3186 skb_set_owner_w(skb, req_to_sk(req));
3187 break;
3188 case TCP_SYNACK_COOKIE:
3189 /* Under synflood, we do not attach skb to a socket,
3190 * to avoid false sharing.
3191 */
3192 break;
3193 case TCP_SYNACK_FASTOPEN:
3194 /* sk is a const pointer, because we want to express multiple
3195 * cpu might call us concurrently.
3196 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3197 */
3198 skb_set_owner_w(skb, (struct sock *)sk);
3199 break;
3200 }
3201 skb_dst_set(skb, dst);
3202
3203 mss = tcp_mss_clamp(tp, dst_metric_advmss(dst));
3204
3205 memset(&opts, 0, sizeof(opts));
3206#ifdef CONFIG_SYN_COOKIES
3207 if (unlikely(req->cookie_ts))
3208 skb->skb_mstamp = cookie_init_timestamp(req);
3209 else
3210#endif
3211 skb->skb_mstamp = tcp_clock_us();
3212
3213#ifdef CONFIG_TCP_MD5SIG
3214 rcu_read_lock();
3215 md5 = tcp_rsk(req)->af_specific->req_md5_lookup(sk, req_to_sk(req));
3216#endif
3217 skb_set_hash(skb, tcp_rsk(req)->txhash, PKT_HASH_TYPE_L4);
3218 tcp_header_size = tcp_synack_options(sk, req, mss, skb, &opts, md5,
3219 foc) + sizeof(*th);
3220<