1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
6 *
7 * Implementation of the Transmission Control Protocol(TCP).
8 *
9 * Authors: Ross Biro
10 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
11 * Mark Evans, <evansmp@uhura.aston.ac.uk>
12 * Corey Minyard <wf-rch!minyard@relay.EU.net>
13 * Florian La Roche, <flla@stud.uni-sb.de>
14 * Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
15 * Linus Torvalds, <torvalds@cs.helsinki.fi>
16 * Alan Cox, <gw4pts@gw4pts.ampr.org>
17 * Matthew Dillon, <dillon@apollo.west.oic.com>
18 * Arnt Gulbrandsen, <agulbra@nvg.unit.no>
19 * Jorge Cwik, <jorge@laser.satlink.net>
20 */
21
22/*
23 * Changes: Pedro Roque : Retransmit queue handled by TCP.
24 * : Fragmentation on mtu decrease
25 * : Segment collapse on retransmit
26 * : AF independence
27 *
28 * Linus Torvalds : send_delayed_ack
29 * David S. Miller : Charge memory using the right skb
30 * during syn/ack processing.
31 * David S. Miller : Output engine completely rewritten.
32 * Andrea Arcangeli: SYNACK carry ts_recent in tsecr.
33 * Cacophonix Gaul : draft-minshall-nagle-01
34 * J Hadi Salim : ECN support
35 *
36 */
37
38#define pr_fmt(fmt) "TCP: " fmt
39
40#include <net/tcp.h>
41#include <net/tcp_ecn.h>
42#include <net/mptcp.h>
43#include <net/proto_memory.h>
44#include <net/psp.h>
45
46#include <linux/compiler.h>
47#include <linux/gfp.h>
48#include <linux/module.h>
49#include <linux/static_key.h>
50#include <linux/skbuff_ref.h>
51
52#include <trace/events/tcp.h>
53
54/* Refresh clocks of a TCP socket,
55 * ensuring monotically increasing values.
56 */
57void tcp_mstamp_refresh(struct tcp_sock *tp)
58{
59 u64 val = tcp_clock_ns();
60
61 tp->tcp_clock_cache = val;
62 tp->tcp_mstamp = div_u64(dividend: val, NSEC_PER_USEC);
63}
64
65static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
66 int push_one, gfp_t gfp);
67
68/* Account for new data that has been sent to the network. */
69static void tcp_event_new_data_sent(struct sock *sk, struct sk_buff *skb)
70{
71 struct inet_connection_sock *icsk = inet_csk(sk);
72 struct tcp_sock *tp = tcp_sk(sk);
73 unsigned int prior_packets = tp->packets_out;
74
75 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(skb)->end_seq);
76
77 __skb_unlink(skb, list: &sk->sk_write_queue);
78 tcp_rbtree_insert(root: &sk->tcp_rtx_queue, skb);
79
80 if (tp->highest_sack == NULL)
81 tp->highest_sack = skb;
82
83 tp->packets_out += tcp_skb_pcount(skb);
84 if (!prior_packets || icsk->icsk_pending == ICSK_TIME_LOSS_PROBE)
85 tcp_rearm_rto(sk);
86
87 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT,
88 tcp_skb_pcount(skb));
89 tcp_check_space(sk);
90}
91
92/* SND.NXT, if window was not shrunk or the amount of shrunk was less than one
93 * window scaling factor due to loss of precision.
94 * If window has been shrunk, what should we make? It is not clear at all.
95 * Using SND.UNA we will fail to open window, SND.NXT is out of window. :-(
96 * Anything in between SND.UNA...SND.UNA+SND.WND also can be already
97 * invalid. OK, let's make this for now:
98 */
99static inline __u32 tcp_acceptable_seq(const struct sock *sk)
100{
101 const struct tcp_sock *tp = tcp_sk(sk);
102
103 if (!before(seq1: tcp_wnd_end(tp), seq2: tp->snd_nxt) ||
104 (tp->rx_opt.wscale_ok &&
105 ((tp->snd_nxt - tcp_wnd_end(tp)) < (1 << tp->rx_opt.rcv_wscale))))
106 return tp->snd_nxt;
107 else
108 return tcp_wnd_end(tp);
109}
110
111/* Calculate mss to advertise in SYN segment.
112 * RFC1122, RFC1063, draft-ietf-tcpimpl-pmtud-01 state that:
113 *
114 * 1. It is independent of path mtu.
115 * 2. Ideally, it is maximal possible segment size i.e. 65535-40.
116 * 3. For IPv4 it is reasonable to calculate it from maximal MTU of
117 * attached devices, because some buggy hosts are confused by
118 * large MSS.
119 * 4. We do not make 3, we advertise MSS, calculated from first
120 * hop device mtu, but allow to raise it to ip_rt_min_advmss.
121 * This may be overridden via information stored in routing table.
122 * 5. Value 65535 for MSS is valid in IPv6 and means "as large as possible,
123 * probably even Jumbo".
124 */
125static __u16 tcp_advertise_mss(struct sock *sk)
126{
127 struct tcp_sock *tp = tcp_sk(sk);
128 const struct dst_entry *dst = __sk_dst_get(sk);
129 int mss = tp->advmss;
130
131 if (dst) {
132 unsigned int metric = dst_metric_advmss(dst);
133
134 if (metric < mss) {
135 mss = metric;
136 tp->advmss = mss;
137 }
138 }
139
140 return (__u16)mss;
141}
142
143/* RFC2861. Reset CWND after idle period longer RTO to "restart window".
144 * This is the first part of cwnd validation mechanism.
145 */
146void tcp_cwnd_restart(struct sock *sk, s32 delta)
147{
148 struct tcp_sock *tp = tcp_sk(sk);
149 u32 restart_cwnd = tcp_init_cwnd(tp, dst: __sk_dst_get(sk));
150 u32 cwnd = tcp_snd_cwnd(tp);
151
152 tcp_ca_event(sk, event: CA_EVENT_CWND_RESTART);
153
154 tp->snd_ssthresh = tcp_current_ssthresh(sk);
155 restart_cwnd = min(restart_cwnd, cwnd);
156
157 while ((delta -= inet_csk(sk)->icsk_rto) > 0 && cwnd > restart_cwnd)
158 cwnd >>= 1;
159 tcp_snd_cwnd_set(tp, max(cwnd, restart_cwnd));
160 tp->snd_cwnd_stamp = tcp_jiffies32;
161 tp->snd_cwnd_used = 0;
162}
163
164/* Congestion state accounting after a packet has been sent. */
165static void tcp_event_data_sent(struct tcp_sock *tp,
166 struct sock *sk)
167{
168 struct inet_connection_sock *icsk = inet_csk(sk);
169 const u32 now = tcp_jiffies32;
170
171 if (tcp_packets_in_flight(tp) == 0)
172 tcp_ca_event(sk, event: CA_EVENT_TX_START);
173
174 tp->lsndtime = now;
175
176 /* If it is a reply for ato after last received
177 * packet, increase pingpong count.
178 */
179 if ((u32)(now - icsk->icsk_ack.lrcvtime) < icsk->icsk_ack.ato)
180 inet_csk_inc_pingpong_cnt(sk);
181}
182
183/* Account for an ACK we sent. */
184static inline void tcp_event_ack_sent(struct sock *sk, u32 rcv_nxt)
185{
186 struct tcp_sock *tp = tcp_sk(sk);
187
188 if (unlikely(tp->compressed_ack)) {
189 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPACKCOMPRESSED,
190 tp->compressed_ack);
191 tp->compressed_ack = 0;
192 if (hrtimer_try_to_cancel(timer: &tp->compressed_ack_timer) == 1)
193 __sock_put(sk);
194 }
195
196 if (unlikely(rcv_nxt != tp->rcv_nxt))
197 return; /* Special ACK sent by DCTCP to reflect ECN */
198 tcp_dec_quickack_mode(sk);
199 inet_csk_clear_xmit_timer(sk, ICSK_TIME_DACK);
200}
201
202/* Determine a window scaling and initial window to offer.
203 * Based on the assumption that the given amount of space
204 * will be offered. Store the results in the tp structure.
205 * NOTE: for smooth operation initial space offering should
206 * be a multiple of mss if possible. We assume here that mss >= 1.
207 * This MUST be enforced by all callers.
208 */
209void tcp_select_initial_window(const struct sock *sk, int __space, __u32 mss,
210 __u32 *rcv_wnd, __u32 *__window_clamp,
211 int wscale_ok, __u8 *rcv_wscale,
212 __u32 init_rcv_wnd)
213{
214 unsigned int space = (__space < 0 ? 0 : __space);
215 u32 window_clamp = READ_ONCE(*__window_clamp);
216
217 /* If no clamp set the clamp to the max possible scaled window */
218 if (window_clamp == 0)
219 window_clamp = (U16_MAX << TCP_MAX_WSCALE);
220 space = min(window_clamp, space);
221
222 /* Quantize space offering to a multiple of mss if possible. */
223 if (space > mss)
224 space = rounddown(space, mss);
225
226 /* NOTE: offering an initial window larger than 32767
227 * will break some buggy TCP stacks. If the admin tells us
228 * it is likely we could be speaking with such a buggy stack
229 * we will truncate our initial window offering to 32K-1
230 * unless the remote has sent us a window scaling option,
231 * which we interpret as a sign the remote TCP is not
232 * misinterpreting the window field as a signed quantity.
233 */
234 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_workaround_signed_windows))
235 (*rcv_wnd) = min(space, MAX_TCP_WINDOW);
236 else
237 (*rcv_wnd) = space;
238
239 if (init_rcv_wnd)
240 *rcv_wnd = min(*rcv_wnd, init_rcv_wnd * mss);
241
242 *rcv_wscale = 0;
243 if (wscale_ok) {
244 /* Set window scaling on max possible window */
245 space = max_t(u32, space, READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_rmem[2]));
246 space = max_t(u32, space, READ_ONCE(sysctl_rmem_max));
247 space = min_t(u32, space, window_clamp);
248 *rcv_wscale = clamp_t(int, ilog2(space) - 15,
249 0, TCP_MAX_WSCALE);
250 }
251 /* Set the clamp no higher than max representable value */
252 WRITE_ONCE(*__window_clamp,
253 min_t(__u32, U16_MAX << (*rcv_wscale), window_clamp));
254}
255EXPORT_IPV6_MOD(tcp_select_initial_window);
256
257/* Chose a new window to advertise, update state in tcp_sock for the
258 * socket, and return result with RFC1323 scaling applied. The return
259 * value can be stuffed directly into th->window for an outgoing
260 * frame.
261 */
262static u16 tcp_select_window(struct sock *sk)
263{
264 struct tcp_sock *tp = tcp_sk(sk);
265 struct net *net = sock_net(sk);
266 u32 old_win = tp->rcv_wnd;
267 u32 cur_win, new_win;
268
269 /* Make the window 0 if we failed to queue the data because we
270 * are out of memory.
271 */
272 if (unlikely(inet_csk(sk)->icsk_ack.pending & ICSK_ACK_NOMEM)) {
273 tp->pred_flags = 0;
274 tp->rcv_wnd = 0;
275 tp->rcv_wup = tp->rcv_nxt;
276 return 0;
277 }
278
279 cur_win = tcp_receive_window(tp);
280 new_win = __tcp_select_window(sk);
281 if (new_win < cur_win) {
282 /* Danger Will Robinson!
283 * Don't update rcv_wup/rcv_wnd here or else
284 * we will not be able to advertise a zero
285 * window in time. --DaveM
286 *
287 * Relax Will Robinson.
288 */
289 if (!READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) || !tp->rx_opt.rcv_wscale) {
290 /* Never shrink the offered window */
291 if (new_win == 0)
292 NET_INC_STATS(net, LINUX_MIB_TCPWANTZEROWINDOWADV);
293 new_win = ALIGN(cur_win, 1 << tp->rx_opt.rcv_wscale);
294 }
295 }
296
297 tp->rcv_wnd = new_win;
298 tp->rcv_wup = tp->rcv_nxt;
299
300 /* Make sure we do not exceed the maximum possible
301 * scaled window.
302 */
303 if (!tp->rx_opt.rcv_wscale &&
304 READ_ONCE(net->ipv4.sysctl_tcp_workaround_signed_windows))
305 new_win = min(new_win, MAX_TCP_WINDOW);
306 else
307 new_win = min(new_win, (65535U << tp->rx_opt.rcv_wscale));
308
309 /* RFC1323 scaling applied */
310 new_win >>= tp->rx_opt.rcv_wscale;
311
312 /* If we advertise zero window, disable fast path. */
313 if (new_win == 0) {
314 tp->pred_flags = 0;
315 if (old_win)
316 NET_INC_STATS(net, LINUX_MIB_TCPTOZEROWINDOWADV);
317 } else if (old_win == 0) {
318 NET_INC_STATS(net, LINUX_MIB_TCPFROMZEROWINDOWADV);
319 }
320
321 return new_win;
322}
323
324/* Set up ECN state for a packet on a ESTABLISHED socket that is about to
325 * be sent.
326 */
327static void tcp_ecn_send(struct sock *sk, struct sk_buff *skb,
328 struct tcphdr *th, int tcp_header_len)
329{
330 struct tcp_sock *tp = tcp_sk(sk);
331
332 if (!tcp_ecn_mode_any(tp))
333 return;
334
335 if (tcp_ecn_mode_accecn(tp)) {
336 if (!tcp_accecn_ace_fail_recv(tp))
337 INET_ECN_xmit(sk);
338 tcp_accecn_set_ace(tp, skb, th);
339 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ACCECN;
340 } else {
341 /* Not-retransmitted data segment: set ECT and inject CWR. */
342 if (skb->len != tcp_header_len &&
343 !before(TCP_SKB_CB(skb)->seq, seq2: tp->snd_nxt)) {
344 INET_ECN_xmit(sk);
345 if (tp->ecn_flags & TCP_ECN_QUEUE_CWR) {
346 tp->ecn_flags &= ~TCP_ECN_QUEUE_CWR;
347 th->cwr = 1;
348 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
349 }
350 } else if (!tcp_ca_needs_ecn(sk)) {
351 /* ACK or retransmitted segment: clear ECT|CE */
352 INET_ECN_dontxmit(sk);
353 }
354 if (tp->ecn_flags & TCP_ECN_DEMAND_CWR)
355 th->ece = 1;
356 }
357}
358
359/* Constructs common control bits of non-data skb. If SYN/FIN is present,
360 * auto increment end seqno.
361 */
362static void tcp_init_nondata_skb(struct sk_buff *skb, struct sock *sk,
363 u32 seq, u16 flags)
364{
365 skb->ip_summed = CHECKSUM_PARTIAL;
366
367 TCP_SKB_CB(skb)->tcp_flags = flags;
368
369 tcp_skb_pcount_set(skb, segs: 1);
370 psp_enqueue_set_decrypted(sk, skb);
371
372 TCP_SKB_CB(skb)->seq = seq;
373 if (flags & (TCPHDR_SYN | TCPHDR_FIN))
374 seq++;
375 TCP_SKB_CB(skb)->end_seq = seq;
376}
377
378static inline bool tcp_urg_mode(const struct tcp_sock *tp)
379{
380 return tp->snd_una != tp->snd_up;
381}
382
383#define OPTION_SACK_ADVERTISE BIT(0)
384#define OPTION_TS BIT(1)
385#define OPTION_MD5 BIT(2)
386#define OPTION_WSCALE BIT(3)
387#define OPTION_FAST_OPEN_COOKIE BIT(8)
388#define OPTION_SMC BIT(9)
389#define OPTION_MPTCP BIT(10)
390#define OPTION_AO BIT(11)
391#define OPTION_ACCECN BIT(12)
392
393static void smc_options_write(__be32 *ptr, u16 *options)
394{
395#if IS_ENABLED(CONFIG_SMC)
396 if (static_branch_unlikely(&tcp_have_smc)) {
397 if (unlikely(OPTION_SMC & *options)) {
398 *ptr++ = htonl((TCPOPT_NOP << 24) |
399 (TCPOPT_NOP << 16) |
400 (TCPOPT_EXP << 8) |
401 (TCPOLEN_EXP_SMC_BASE));
402 *ptr++ = htonl(TCPOPT_SMC_MAGIC);
403 }
404 }
405#endif
406}
407
408struct tcp_out_options {
409 u16 options; /* bit field of OPTION_* */
410 u16 mss; /* 0 to disable */
411 u8 ws; /* window scale, 0 to disable */
412 u8 num_sack_blocks; /* number of SACK blocks to include */
413 u8 num_accecn_fields:7, /* number of AccECN fields needed */
414 use_synack_ecn_bytes:1; /* Use synack_ecn_bytes or not */
415 u8 hash_size; /* bytes in hash_location */
416 u8 bpf_opt_len; /* length of BPF hdr option */
417 __u8 *hash_location; /* temporary pointer, overloaded */
418 __u32 tsval, tsecr; /* need to include OPTION_TS */
419 struct tcp_fastopen_cookie *fastopen_cookie; /* Fast open cookie */
420 struct mptcp_out_options mptcp;
421};
422
423static void mptcp_options_write(struct tcphdr *th, __be32 *ptr,
424 struct tcp_sock *tp,
425 struct tcp_out_options *opts)
426{
427#if IS_ENABLED(CONFIG_MPTCP)
428 if (unlikely(OPTION_MPTCP & opts->options))
429 mptcp_write_options(th, ptr, tp, &opts->mptcp);
430#endif
431}
432
433#ifdef CONFIG_CGROUP_BPF
434static int bpf_skops_write_hdr_opt_arg0(struct sk_buff *skb,
435 enum tcp_synack_type synack_type)
436{
437 if (unlikely(!skb))
438 return BPF_WRITE_HDR_TCP_CURRENT_MSS;
439
440 if (unlikely(synack_type == TCP_SYNACK_COOKIE))
441 return BPF_WRITE_HDR_TCP_SYNACK_COOKIE;
442
443 return 0;
444}
445
446/* req, syn_skb and synack_type are used when writing synack */
447static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
448 struct request_sock *req,
449 struct sk_buff *syn_skb,
450 enum tcp_synack_type synack_type,
451 struct tcp_out_options *opts,
452 unsigned int *remaining)
453{
454 struct bpf_sock_ops_kern sock_ops;
455 int err;
456
457 if (likely(!BPF_SOCK_OPS_TEST_FLAG(tcp_sk(sk),
458 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG)) ||
459 !*remaining)
460 return;
461
462 /* *remaining has already been aligned to 4 bytes, so *remaining >= 4 */
463
464 /* init sock_ops */
465 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
466
467 sock_ops.op = BPF_SOCK_OPS_HDR_OPT_LEN_CB;
468
469 if (req) {
470 /* The listen "sk" cannot be passed here because
471 * it is not locked. It would not make too much
472 * sense to do bpf_setsockopt(listen_sk) based
473 * on individual connection request also.
474 *
475 * Thus, "req" is passed here and the cgroup-bpf-progs
476 * of the listen "sk" will be run.
477 *
478 * "req" is also used here for fastopen even the "sk" here is
479 * a fullsock "child" sk. It is to keep the behavior
480 * consistent between fastopen and non-fastopen on
481 * the bpf programming side.
482 */
483 sock_ops.sk = (struct sock *)req;
484 sock_ops.syn_skb = syn_skb;
485 } else {
486 sock_owned_by_me(sk);
487
488 sock_ops.is_fullsock = 1;
489 sock_ops.is_locked_tcp_sock = 1;
490 sock_ops.sk = sk;
491 }
492
493 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
494 sock_ops.remaining_opt_len = *remaining;
495 /* tcp_current_mss() does not pass a skb */
496 if (skb)
497 bpf_skops_init_skb(&sock_ops, skb, 0);
498
499 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
500
501 if (err || sock_ops.remaining_opt_len == *remaining)
502 return;
503
504 opts->bpf_opt_len = *remaining - sock_ops.remaining_opt_len;
505 /* round up to 4 bytes */
506 opts->bpf_opt_len = (opts->bpf_opt_len + 3) & ~3;
507
508 *remaining -= opts->bpf_opt_len;
509}
510
511static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
512 struct request_sock *req,
513 struct sk_buff *syn_skb,
514 enum tcp_synack_type synack_type,
515 struct tcp_out_options *opts)
516{
517 u8 first_opt_off, nr_written, max_opt_len = opts->bpf_opt_len;
518 struct bpf_sock_ops_kern sock_ops;
519 int err;
520
521 if (likely(!max_opt_len))
522 return;
523
524 memset(&sock_ops, 0, offsetof(struct bpf_sock_ops_kern, temp));
525
526 sock_ops.op = BPF_SOCK_OPS_WRITE_HDR_OPT_CB;
527
528 if (req) {
529 sock_ops.sk = (struct sock *)req;
530 sock_ops.syn_skb = syn_skb;
531 } else {
532 sock_owned_by_me(sk);
533
534 sock_ops.is_fullsock = 1;
535 sock_ops.is_locked_tcp_sock = 1;
536 sock_ops.sk = sk;
537 }
538
539 sock_ops.args[0] = bpf_skops_write_hdr_opt_arg0(skb, synack_type);
540 sock_ops.remaining_opt_len = max_opt_len;
541 first_opt_off = tcp_hdrlen(skb) - max_opt_len;
542 bpf_skops_init_skb(&sock_ops, skb, first_opt_off);
543
544 err = BPF_CGROUP_RUN_PROG_SOCK_OPS_SK(&sock_ops, sk);
545
546 if (err)
547 nr_written = 0;
548 else
549 nr_written = max_opt_len - sock_ops.remaining_opt_len;
550
551 if (nr_written < max_opt_len)
552 memset(skb->data + first_opt_off + nr_written, TCPOPT_NOP,
553 max_opt_len - nr_written);
554}
555#else
556static void bpf_skops_hdr_opt_len(struct sock *sk, struct sk_buff *skb,
557 struct request_sock *req,
558 struct sk_buff *syn_skb,
559 enum tcp_synack_type synack_type,
560 struct tcp_out_options *opts,
561 unsigned int *remaining)
562{
563}
564
565static void bpf_skops_write_hdr_opt(struct sock *sk, struct sk_buff *skb,
566 struct request_sock *req,
567 struct sk_buff *syn_skb,
568 enum tcp_synack_type synack_type,
569 struct tcp_out_options *opts)
570{
571}
572#endif
573
574static __be32 *process_tcp_ao_options(struct tcp_sock *tp,
575 const struct tcp_request_sock *tcprsk,
576 struct tcp_out_options *opts,
577 struct tcp_key *key, __be32 *ptr)
578{
579#ifdef CONFIG_TCP_AO
580 u8 maclen = tcp_ao_maclen(key->ao_key);
581
582 if (tcprsk) {
583 u8 aolen = maclen + sizeof(struct tcp_ao_hdr);
584
585 *ptr++ = htonl((TCPOPT_AO << 24) | (aolen << 16) |
586 (tcprsk->ao_keyid << 8) |
587 (tcprsk->ao_rcv_next));
588 } else {
589 struct tcp_ao_key *rnext_key;
590 struct tcp_ao_info *ao_info;
591
592 ao_info = rcu_dereference_check(tp->ao_info,
593 lockdep_sock_is_held(&tp->inet_conn.icsk_inet.sk));
594 rnext_key = READ_ONCE(ao_info->rnext_key);
595 if (WARN_ON_ONCE(!rnext_key))
596 return ptr;
597 *ptr++ = htonl((TCPOPT_AO << 24) |
598 (tcp_ao_len(key->ao_key) << 16) |
599 (key->ao_key->sndid << 8) |
600 (rnext_key->rcvid));
601 }
602 opts->hash_location = (__u8 *)ptr;
603 ptr += maclen / sizeof(*ptr);
604 if (unlikely(maclen % sizeof(*ptr))) {
605 memset(ptr, TCPOPT_NOP, sizeof(*ptr));
606 ptr++;
607 }
608#endif
609 return ptr;
610}
611
612/* Initial values for AccECN option, ordered is based on ECN field bits
613 * similar to received_ecn_bytes. Used for SYN/ACK AccECN option.
614 */
615static const u32 synack_ecn_bytes[3] = { 0, 0, 0 };
616
617/* Write previously computed TCP options to the packet.
618 *
619 * Beware: Something in the Internet is very sensitive to the ordering of
620 * TCP options, we learned this through the hard way, so be careful here.
621 * Luckily we can at least blame others for their non-compliance but from
622 * inter-operability perspective it seems that we're somewhat stuck with
623 * the ordering which we have been using if we want to keep working with
624 * those broken things (not that it currently hurts anybody as there isn't
625 * particular reason why the ordering would need to be changed).
626 *
627 * At least SACK_PERM as the first option is known to lead to a disaster
628 * (but it may well be that other scenarios fail similarly).
629 */
630static void tcp_options_write(struct tcphdr *th, struct tcp_sock *tp,
631 const struct tcp_request_sock *tcprsk,
632 struct tcp_out_options *opts,
633 struct tcp_key *key)
634{
635 u8 leftover_highbyte = TCPOPT_NOP; /* replace 1st NOP if avail */
636 u8 leftover_lowbyte = TCPOPT_NOP; /* replace 2nd NOP in succession */
637 __be32 *ptr = (__be32 *)(th + 1);
638 u16 options = opts->options; /* mungable copy */
639
640 if (tcp_key_is_md5(key)) {
641 *ptr++ = htonl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) |
642 (TCPOPT_MD5SIG << 8) | TCPOLEN_MD5SIG);
643 /* overload cookie hash location */
644 opts->hash_location = (__u8 *)ptr;
645 ptr += 4;
646 } else if (tcp_key_is_ao(key)) {
647 ptr = process_tcp_ao_options(tp, tcprsk, opts, key, ptr);
648 }
649 if (unlikely(opts->mss)) {
650 *ptr++ = htonl((TCPOPT_MSS << 24) |
651 (TCPOLEN_MSS << 16) |
652 opts->mss);
653 }
654
655 if (likely(OPTION_TS & options)) {
656 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
657 *ptr++ = htonl((TCPOPT_SACK_PERM << 24) |
658 (TCPOLEN_SACK_PERM << 16) |
659 (TCPOPT_TIMESTAMP << 8) |
660 TCPOLEN_TIMESTAMP);
661 options &= ~OPTION_SACK_ADVERTISE;
662 } else {
663 *ptr++ = htonl((TCPOPT_NOP << 24) |
664 (TCPOPT_NOP << 16) |
665 (TCPOPT_TIMESTAMP << 8) |
666 TCPOLEN_TIMESTAMP);
667 }
668 *ptr++ = htonl(opts->tsval);
669 *ptr++ = htonl(opts->tsecr);
670 }
671
672 if (OPTION_ACCECN & options) {
673 const u32 *ecn_bytes = opts->use_synack_ecn_bytes ?
674 synack_ecn_bytes :
675 tp->received_ecn_bytes;
676 const u8 ect0_idx = INET_ECN_ECT_0 - 1;
677 const u8 ect1_idx = INET_ECN_ECT_1 - 1;
678 const u8 ce_idx = INET_ECN_CE - 1;
679 u32 e0b;
680 u32 e1b;
681 u32 ceb;
682 u8 len;
683
684 e0b = ecn_bytes[ect0_idx] + TCP_ACCECN_E0B_INIT_OFFSET;
685 e1b = ecn_bytes[ect1_idx] + TCP_ACCECN_E1B_INIT_OFFSET;
686 ceb = ecn_bytes[ce_idx] + TCP_ACCECN_CEB_INIT_OFFSET;
687 len = TCPOLEN_ACCECN_BASE +
688 opts->num_accecn_fields * TCPOLEN_ACCECN_PERFIELD;
689
690 if (opts->num_accecn_fields == 2) {
691 *ptr++ = htonl((TCPOPT_ACCECN1 << 24) | (len << 16) |
692 ((e1b >> 8) & 0xffff));
693 *ptr++ = htonl(((e1b & 0xff) << 24) |
694 (ceb & 0xffffff));
695 } else if (opts->num_accecn_fields == 1) {
696 *ptr++ = htonl((TCPOPT_ACCECN1 << 24) | (len << 16) |
697 ((e1b >> 8) & 0xffff));
698 leftover_highbyte = e1b & 0xff;
699 leftover_lowbyte = TCPOPT_NOP;
700 } else if (opts->num_accecn_fields == 0) {
701 leftover_highbyte = TCPOPT_ACCECN1;
702 leftover_lowbyte = len;
703 } else if (opts->num_accecn_fields == 3) {
704 *ptr++ = htonl((TCPOPT_ACCECN1 << 24) | (len << 16) |
705 ((e1b >> 8) & 0xffff));
706 *ptr++ = htonl(((e1b & 0xff) << 24) |
707 (ceb & 0xffffff));
708 *ptr++ = htonl(((e0b & 0xffffff) << 8) |
709 TCPOPT_NOP);
710 }
711 if (tp) {
712 tp->accecn_minlen = 0;
713 tp->accecn_opt_tstamp = tp->tcp_mstamp;
714 if (tp->accecn_opt_demand)
715 tp->accecn_opt_demand--;
716 }
717 }
718
719 if (unlikely(OPTION_SACK_ADVERTISE & options)) {
720 *ptr++ = htonl((leftover_highbyte << 24) |
721 (leftover_lowbyte << 16) |
722 (TCPOPT_SACK_PERM << 8) |
723 TCPOLEN_SACK_PERM);
724 leftover_highbyte = TCPOPT_NOP;
725 leftover_lowbyte = TCPOPT_NOP;
726 }
727
728 if (unlikely(OPTION_WSCALE & options)) {
729 u8 highbyte = TCPOPT_NOP;
730
731 /* Do not split the leftover 2-byte to fit into a single
732 * NOP, i.e., replace this NOP only when 1 byte is leftover
733 * within leftover_highbyte.
734 */
735 if (unlikely(leftover_highbyte != TCPOPT_NOP &&
736 leftover_lowbyte == TCPOPT_NOP)) {
737 highbyte = leftover_highbyte;
738 leftover_highbyte = TCPOPT_NOP;
739 }
740 *ptr++ = htonl((highbyte << 24) |
741 (TCPOPT_WINDOW << 16) |
742 (TCPOLEN_WINDOW << 8) |
743 opts->ws);
744 }
745
746 if (unlikely(opts->num_sack_blocks)) {
747 struct tcp_sack_block *sp = tp->rx_opt.dsack ?
748 tp->duplicate_sack : tp->selective_acks;
749 int this_sack;
750
751 *ptr++ = htonl((leftover_highbyte << 24) |
752 (leftover_lowbyte << 16) |
753 (TCPOPT_SACK << 8) |
754 (TCPOLEN_SACK_BASE + (opts->num_sack_blocks *
755 TCPOLEN_SACK_PERBLOCK)));
756 leftover_highbyte = TCPOPT_NOP;
757 leftover_lowbyte = TCPOPT_NOP;
758
759 for (this_sack = 0; this_sack < opts->num_sack_blocks;
760 ++this_sack) {
761 *ptr++ = htonl(sp[this_sack].start_seq);
762 *ptr++ = htonl(sp[this_sack].end_seq);
763 }
764
765 tp->rx_opt.dsack = 0;
766 } else if (unlikely(leftover_highbyte != TCPOPT_NOP ||
767 leftover_lowbyte != TCPOPT_NOP)) {
768 *ptr++ = htonl((leftover_highbyte << 24) |
769 (leftover_lowbyte << 16) |
770 (TCPOPT_NOP << 8) |
771 TCPOPT_NOP);
772 leftover_highbyte = TCPOPT_NOP;
773 leftover_lowbyte = TCPOPT_NOP;
774 }
775
776 if (unlikely(OPTION_FAST_OPEN_COOKIE & options)) {
777 struct tcp_fastopen_cookie *foc = opts->fastopen_cookie;
778 u8 *p = (u8 *)ptr;
779 u32 len; /* Fast Open option length */
780
781 if (foc->exp) {
782 len = TCPOLEN_EXP_FASTOPEN_BASE + foc->len;
783 *ptr = htonl((TCPOPT_EXP << 24) | (len << 16) |
784 TCPOPT_FASTOPEN_MAGIC);
785 p += TCPOLEN_EXP_FASTOPEN_BASE;
786 } else {
787 len = TCPOLEN_FASTOPEN_BASE + foc->len;
788 *p++ = TCPOPT_FASTOPEN;
789 *p++ = len;
790 }
791
792 memcpy(to: p, from: foc->val, len: foc->len);
793 if ((len & 3) == 2) {
794 p[foc->len] = TCPOPT_NOP;
795 p[foc->len + 1] = TCPOPT_NOP;
796 }
797 ptr += (len + 3) >> 2;
798 }
799
800 smc_options_write(ptr, options: &options);
801
802 mptcp_options_write(th, ptr, tp, opts);
803}
804
805static void smc_set_option(const struct tcp_sock *tp,
806 struct tcp_out_options *opts,
807 unsigned int *remaining)
808{
809#if IS_ENABLED(CONFIG_SMC)
810 if (static_branch_unlikely(&tcp_have_smc)) {
811 if (tp->syn_smc) {
812 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
813 opts->options |= OPTION_SMC;
814 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
815 }
816 }
817 }
818#endif
819}
820
821static void smc_set_option_cond(const struct tcp_sock *tp,
822 const struct inet_request_sock *ireq,
823 struct tcp_out_options *opts,
824 unsigned int *remaining)
825{
826#if IS_ENABLED(CONFIG_SMC)
827 if (static_branch_unlikely(&tcp_have_smc)) {
828 if (tp->syn_smc && ireq->smc_ok) {
829 if (*remaining >= TCPOLEN_EXP_SMC_BASE_ALIGNED) {
830 opts->options |= OPTION_SMC;
831 *remaining -= TCPOLEN_EXP_SMC_BASE_ALIGNED;
832 }
833 }
834 }
835#endif
836}
837
838static void mptcp_set_option_cond(const struct request_sock *req,
839 struct tcp_out_options *opts,
840 unsigned int *remaining)
841{
842 if (rsk_is_mptcp(req)) {
843 unsigned int size;
844
845 if (mptcp_synack_options(req, size: &size, opts: &opts->mptcp)) {
846 if (*remaining >= size) {
847 opts->options |= OPTION_MPTCP;
848 *remaining -= size;
849 }
850 }
851 }
852}
853
854static u32 tcp_synack_options_combine_saving(struct tcp_out_options *opts)
855{
856 /* How much there's room for combining with the alignment padding? */
857 if ((opts->options & (OPTION_SACK_ADVERTISE | OPTION_TS)) ==
858 OPTION_SACK_ADVERTISE)
859 return 2;
860 else if (opts->options & OPTION_WSCALE)
861 return 1;
862 return 0;
863}
864
865/* Calculates how long AccECN option will fit to @remaining option space.
866 *
867 * AccECN option can sometimes replace NOPs used for alignment of other
868 * TCP options (up to @max_combine_saving available).
869 *
870 * Only solutions with at least @required AccECN fields are accepted.
871 *
872 * Returns: The size of the AccECN option excluding space repurposed from
873 * the alignment of the other options.
874 */
875static int tcp_options_fit_accecn(struct tcp_out_options *opts, int required,
876 int remaining)
877{
878 int size = TCP_ACCECN_MAXSIZE;
879 int sack_blocks_reduce = 0;
880 int max_combine_saving;
881 int rem = remaining;
882 int align_size;
883
884 if (opts->use_synack_ecn_bytes)
885 max_combine_saving = tcp_synack_options_combine_saving(opts);
886 else
887 max_combine_saving = opts->num_sack_blocks > 0 ? 2 : 0;
888 opts->num_accecn_fields = TCP_ACCECN_NUMFIELDS;
889 while (opts->num_accecn_fields >= required) {
890 /* Pad to dword if cannot combine */
891 if ((size & 0x3) > max_combine_saving)
892 align_size = ALIGN(size, 4);
893 else
894 align_size = ALIGN_DOWN(size, 4);
895
896 if (rem >= align_size) {
897 size = align_size;
898 break;
899 } else if (opts->num_accecn_fields == required &&
900 opts->num_sack_blocks > 2 &&
901 required > 0) {
902 /* Try to fit the option by removing one SACK block */
903 opts->num_sack_blocks--;
904 sack_blocks_reduce++;
905 rem = rem + TCPOLEN_SACK_PERBLOCK;
906
907 opts->num_accecn_fields = TCP_ACCECN_NUMFIELDS;
908 size = TCP_ACCECN_MAXSIZE;
909 continue;
910 }
911
912 opts->num_accecn_fields--;
913 size -= TCPOLEN_ACCECN_PERFIELD;
914 }
915 if (sack_blocks_reduce > 0) {
916 if (opts->num_accecn_fields >= required)
917 size -= sack_blocks_reduce * TCPOLEN_SACK_PERBLOCK;
918 else
919 opts->num_sack_blocks += sack_blocks_reduce;
920 }
921 if (opts->num_accecn_fields < required)
922 return 0;
923
924 opts->options |= OPTION_ACCECN;
925 return size;
926}
927
928/* Compute TCP options for SYN packets. This is not the final
929 * network wire format yet.
930 */
931static unsigned int tcp_syn_options(struct sock *sk, struct sk_buff *skb,
932 struct tcp_out_options *opts,
933 struct tcp_key *key)
934{
935 struct tcp_sock *tp = tcp_sk(sk);
936 unsigned int remaining = MAX_TCP_OPTION_SPACE;
937 struct tcp_fastopen_request *fastopen = tp->fastopen_req;
938 bool timestamps;
939
940 /* Better than switch (key.type) as it has static branches */
941 if (tcp_key_is_md5(key)) {
942 timestamps = false;
943 opts->options |= OPTION_MD5;
944 remaining -= TCPOLEN_MD5SIG_ALIGNED;
945 } else {
946 timestamps = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps);
947 if (tcp_key_is_ao(key)) {
948 opts->options |= OPTION_AO;
949 remaining -= tcp_ao_len_aligned(key: key->ao_key);
950 }
951 }
952
953 /* We always get an MSS option. The option bytes which will be seen in
954 * normal data packets should timestamps be used, must be in the MSS
955 * advertised. But we subtract them from tp->mss_cache so that
956 * calculations in tcp_sendmsg are simpler etc. So account for this
957 * fact here if necessary. If we don't do this correctly, as a
958 * receiver we won't recognize data packets as being full sized when we
959 * should, and thus we won't abide by the delayed ACK rules correctly.
960 * SACKs don't matter, we never delay an ACK when we have any of those
961 * going out. */
962 opts->mss = tcp_advertise_mss(sk);
963 remaining -= TCPOLEN_MSS_ALIGNED;
964
965 if (likely(timestamps)) {
966 opts->options |= OPTION_TS;
967 opts->tsval = tcp_skb_timestamp_ts(usec_ts: tp->tcp_usec_ts, skb) + tp->tsoffset;
968 opts->tsecr = tp->rx_opt.ts_recent;
969 remaining -= TCPOLEN_TSTAMP_ALIGNED;
970 }
971 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling))) {
972 opts->ws = tp->rx_opt.rcv_wscale;
973 opts->options |= OPTION_WSCALE;
974 remaining -= TCPOLEN_WSCALE_ALIGNED;
975 }
976 if (likely(READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_sack))) {
977 opts->options |= OPTION_SACK_ADVERTISE;
978 if (unlikely(!(OPTION_TS & opts->options)))
979 remaining -= TCPOLEN_SACKPERM_ALIGNED;
980 }
981
982 if (fastopen && fastopen->cookie.len >= 0) {
983 u32 need = fastopen->cookie.len;
984
985 need += fastopen->cookie.exp ? TCPOLEN_EXP_FASTOPEN_BASE :
986 TCPOLEN_FASTOPEN_BASE;
987 need = (need + 3) & ~3U; /* Align to 32 bits */
988 if (remaining >= need) {
989 opts->options |= OPTION_FAST_OPEN_COOKIE;
990 opts->fastopen_cookie = &fastopen->cookie;
991 remaining -= need;
992 tp->syn_fastopen = 1;
993 tp->syn_fastopen_exp = fastopen->cookie.exp ? 1 : 0;
994 }
995 }
996
997 smc_set_option(tp, opts, remaining: &remaining);
998
999 if (sk_is_mptcp(sk)) {
1000 unsigned int size;
1001
1002 if (mptcp_syn_options(sk, skb, size: &size, opts: &opts->mptcp)) {
1003 if (remaining >= size) {
1004 opts->options |= OPTION_MPTCP;
1005 remaining -= size;
1006 }
1007 }
1008 }
1009
1010 /* Simultaneous open SYN/ACK needs AccECN option but not SYN.
1011 * It is attempted to negotiate the use of AccECN also on the first
1012 * retransmitted SYN, as mentioned in "3.1.4.1. Retransmitted SYNs"
1013 * of AccECN draft.
1014 */
1015 if (unlikely((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK) &&
1016 tcp_ecn_mode_accecn(tp) &&
1017 inet_csk(sk)->icsk_retransmits < 2 &&
1018 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_option) &&
1019 remaining >= TCPOLEN_ACCECN_BASE)) {
1020 opts->use_synack_ecn_bytes = 1;
1021 remaining -= tcp_options_fit_accecn(opts, required: 0, remaining);
1022 }
1023
1024 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, synack_type: 0, opts, remaining: &remaining);
1025
1026 return MAX_TCP_OPTION_SPACE - remaining;
1027}
1028
1029/* Set up TCP options for SYN-ACKs. */
1030static unsigned int tcp_synack_options(const struct sock *sk,
1031 struct request_sock *req,
1032 unsigned int mss, struct sk_buff *skb,
1033 struct tcp_out_options *opts,
1034 const struct tcp_key *key,
1035 struct tcp_fastopen_cookie *foc,
1036 enum tcp_synack_type synack_type,
1037 struct sk_buff *syn_skb)
1038{
1039 struct inet_request_sock *ireq = inet_rsk(sk: req);
1040 unsigned int remaining = MAX_TCP_OPTION_SPACE;
1041 struct tcp_request_sock *treq = tcp_rsk(req);
1042
1043 if (tcp_key_is_md5(key)) {
1044 opts->options |= OPTION_MD5;
1045 remaining -= TCPOLEN_MD5SIG_ALIGNED;
1046
1047 /* We can't fit any SACK blocks in a packet with MD5 + TS
1048 * options. There was discussion about disabling SACK
1049 * rather than TS in order to fit in better with old,
1050 * buggy kernels, but that was deemed to be unnecessary.
1051 */
1052 if (synack_type != TCP_SYNACK_COOKIE)
1053 ireq->tstamp_ok &= !ireq->sack_ok;
1054 } else if (tcp_key_is_ao(key)) {
1055 opts->options |= OPTION_AO;
1056 remaining -= tcp_ao_len_aligned(key: key->ao_key);
1057 ireq->tstamp_ok &= !ireq->sack_ok;
1058 }
1059
1060 /* We always send an MSS option. */
1061 opts->mss = mss;
1062 remaining -= TCPOLEN_MSS_ALIGNED;
1063
1064 if (likely(ireq->wscale_ok)) {
1065 opts->ws = ireq->rcv_wscale;
1066 opts->options |= OPTION_WSCALE;
1067 remaining -= TCPOLEN_WSCALE_ALIGNED;
1068 }
1069 if (likely(ireq->tstamp_ok)) {
1070 opts->options |= OPTION_TS;
1071 opts->tsval = tcp_skb_timestamp_ts(usec_ts: tcp_rsk(req)->req_usec_ts, skb) +
1072 tcp_rsk(req)->ts_off;
1073 if (!tcp_rsk(req)->snt_tsval_first) {
1074 if (!opts->tsval)
1075 opts->tsval = ~0U;
1076 tcp_rsk(req)->snt_tsval_first = opts->tsval;
1077 }
1078 WRITE_ONCE(tcp_rsk(req)->snt_tsval_last, opts->tsval);
1079 opts->tsecr = req->ts_recent;
1080 remaining -= TCPOLEN_TSTAMP_ALIGNED;
1081 }
1082 if (likely(ireq->sack_ok)) {
1083 opts->options |= OPTION_SACK_ADVERTISE;
1084 if (unlikely(!ireq->tstamp_ok))
1085 remaining -= TCPOLEN_SACKPERM_ALIGNED;
1086 }
1087 if (foc != NULL && foc->len >= 0) {
1088 u32 need = foc->len;
1089
1090 need += foc->exp ? TCPOLEN_EXP_FASTOPEN_BASE :
1091 TCPOLEN_FASTOPEN_BASE;
1092 need = (need + 3) & ~3U; /* Align to 32 bits */
1093 if (remaining >= need) {
1094 opts->options |= OPTION_FAST_OPEN_COOKIE;
1095 opts->fastopen_cookie = foc;
1096 remaining -= need;
1097 }
1098 }
1099
1100 mptcp_set_option_cond(req, opts, remaining: &remaining);
1101
1102 smc_set_option_cond(tcp_sk(sk), ireq, opts, remaining: &remaining);
1103
1104 if (treq->accecn_ok &&
1105 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_option) &&
1106 req->num_timeout < 1 && remaining >= TCPOLEN_ACCECN_BASE) {
1107 opts->use_synack_ecn_bytes = 1;
1108 remaining -= tcp_options_fit_accecn(opts, required: 0, remaining);
1109 }
1110
1111 bpf_skops_hdr_opt_len(sk: (struct sock *)sk, skb, req, syn_skb,
1112 synack_type, opts, remaining: &remaining);
1113
1114 return MAX_TCP_OPTION_SPACE - remaining;
1115}
1116
1117/* Compute TCP options for ESTABLISHED sockets. This is not the
1118 * final wire format yet.
1119 */
1120static unsigned int tcp_established_options(struct sock *sk, struct sk_buff *skb,
1121 struct tcp_out_options *opts,
1122 struct tcp_key *key)
1123{
1124 struct tcp_sock *tp = tcp_sk(sk);
1125 unsigned int size = 0;
1126 unsigned int eff_sacks;
1127
1128 opts->options = 0;
1129
1130 /* Better than switch (key.type) as it has static branches */
1131 if (tcp_key_is_md5(key)) {
1132 opts->options |= OPTION_MD5;
1133 size += TCPOLEN_MD5SIG_ALIGNED;
1134 } else if (tcp_key_is_ao(key)) {
1135 opts->options |= OPTION_AO;
1136 size += tcp_ao_len_aligned(key: key->ao_key);
1137 }
1138
1139 if (likely(tp->rx_opt.tstamp_ok)) {
1140 opts->options |= OPTION_TS;
1141 opts->tsval = skb ? tcp_skb_timestamp_ts(usec_ts: tp->tcp_usec_ts, skb) +
1142 tp->tsoffset : 0;
1143 opts->tsecr = tp->rx_opt.ts_recent;
1144 size += TCPOLEN_TSTAMP_ALIGNED;
1145 }
1146
1147 /* MPTCP options have precedence over SACK for the limited TCP
1148 * option space because a MPTCP connection would be forced to
1149 * fall back to regular TCP if a required multipath option is
1150 * missing. SACK still gets a chance to use whatever space is
1151 * left.
1152 */
1153 if (sk_is_mptcp(sk)) {
1154 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
1155 unsigned int opt_size = 0;
1156
1157 if (mptcp_established_options(sk, skb, size: &opt_size, remaining,
1158 opts: &opts->mptcp)) {
1159 opts->options |= OPTION_MPTCP;
1160 size += opt_size;
1161 }
1162 }
1163
1164 eff_sacks = tp->rx_opt.num_sacks + tp->rx_opt.dsack;
1165 if (unlikely(eff_sacks)) {
1166 const unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
1167 if (likely(remaining >= TCPOLEN_SACK_BASE_ALIGNED +
1168 TCPOLEN_SACK_PERBLOCK)) {
1169 opts->num_sack_blocks =
1170 min_t(unsigned int, eff_sacks,
1171 (remaining - TCPOLEN_SACK_BASE_ALIGNED) /
1172 TCPOLEN_SACK_PERBLOCK);
1173
1174 size += TCPOLEN_SACK_BASE_ALIGNED +
1175 opts->num_sack_blocks * TCPOLEN_SACK_PERBLOCK;
1176 } else {
1177 opts->num_sack_blocks = 0;
1178 }
1179 } else {
1180 opts->num_sack_blocks = 0;
1181 }
1182
1183 if (tcp_ecn_mode_accecn(tp)) {
1184 int ecn_opt = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_ecn_option);
1185
1186 if (ecn_opt && tp->saw_accecn_opt && !tcp_accecn_opt_fail_send(tp) &&
1187 (ecn_opt >= TCP_ACCECN_OPTION_FULL || tp->accecn_opt_demand ||
1188 tcp_accecn_option_beacon_check(sk))) {
1189 opts->use_synack_ecn_bytes = 0;
1190 size += tcp_options_fit_accecn(opts, required: tp->accecn_minlen,
1191 MAX_TCP_OPTION_SPACE - size);
1192 }
1193 }
1194
1195 if (unlikely(BPF_SOCK_OPS_TEST_FLAG(tp,
1196 BPF_SOCK_OPS_WRITE_HDR_OPT_CB_FLAG))) {
1197 unsigned int remaining = MAX_TCP_OPTION_SPACE - size;
1198
1199 bpf_skops_hdr_opt_len(sk, skb, NULL, NULL, synack_type: 0, opts, remaining: &remaining);
1200
1201 size = MAX_TCP_OPTION_SPACE - remaining;
1202 }
1203
1204 return size;
1205}
1206
1207
1208/* TCP SMALL QUEUES (TSQ)
1209 *
1210 * TSQ goal is to keep small amount of skbs per tcp flow in tx queues (qdisc+dev)
1211 * to reduce RTT and bufferbloat.
1212 * We do this using a special skb destructor (tcp_wfree).
1213 *
1214 * Its important tcp_wfree() can be replaced by sock_wfree() in the event skb
1215 * needs to be reallocated in a driver.
1216 * The invariant being skb->truesize subtracted from sk->sk_wmem_alloc
1217 *
1218 * Since transmit from skb destructor is forbidden, we use a BH work item
1219 * to process all sockets that eventually need to send more skbs.
1220 * We use one work item per cpu, with its own queue of sockets.
1221 */
1222struct tsq_work {
1223 struct work_struct work;
1224 struct list_head head; /* queue of tcp sockets */
1225};
1226static DEFINE_PER_CPU(struct tsq_work, tsq_work);
1227
1228static void tcp_tsq_write(struct sock *sk)
1229{
1230 if ((1 << sk->sk_state) &
1231 (TCPF_ESTABLISHED | TCPF_FIN_WAIT1 | TCPF_CLOSING |
1232 TCPF_CLOSE_WAIT | TCPF_LAST_ACK)) {
1233 struct tcp_sock *tp = tcp_sk(sk);
1234
1235 if (tp->lost_out > tp->retrans_out &&
1236 tcp_snd_cwnd(tp) > tcp_packets_in_flight(tp)) {
1237 tcp_mstamp_refresh(tp);
1238 tcp_xmit_retransmit_queue(sk);
1239 }
1240
1241 tcp_write_xmit(sk, mss_now: tcp_current_mss(sk), nonagle: tp->nonagle,
1242 push_one: 0, GFP_ATOMIC);
1243 }
1244}
1245
1246static void tcp_tsq_handler(struct sock *sk)
1247{
1248 bh_lock_sock(sk);
1249 if (!sock_owned_by_user(sk))
1250 tcp_tsq_write(sk);
1251 else if (!test_and_set_bit(nr: TCP_TSQ_DEFERRED, addr: &sk->sk_tsq_flags))
1252 sock_hold(sk);
1253 bh_unlock_sock(sk);
1254}
1255/*
1256 * One work item per cpu tries to send more skbs.
1257 * We run in BH context but need to disable irqs when
1258 * transferring tsq->head because tcp_wfree() might
1259 * interrupt us (non NAPI drivers)
1260 */
1261static void tcp_tsq_workfn(struct work_struct *work)
1262{
1263 struct tsq_work *tsq = container_of(work, struct tsq_work, work);
1264 LIST_HEAD(list);
1265 unsigned long flags;
1266 struct list_head *q, *n;
1267 struct tcp_sock *tp;
1268 struct sock *sk;
1269
1270 local_irq_save(flags);
1271 list_splice_init(list: &tsq->head, head: &list);
1272 local_irq_restore(flags);
1273
1274 list_for_each_safe(q, n, &list) {
1275 tp = list_entry(q, struct tcp_sock, tsq_node);
1276 list_del(entry: &tp->tsq_node);
1277
1278 sk = (struct sock *)tp;
1279 smp_mb__before_atomic();
1280 clear_bit(nr: TSQ_QUEUED, addr: &sk->sk_tsq_flags);
1281
1282 tcp_tsq_handler(sk);
1283 sk_free(sk);
1284 }
1285}
1286
1287#define TCP_DEFERRED_ALL (TCPF_TSQ_DEFERRED | \
1288 TCPF_WRITE_TIMER_DEFERRED | \
1289 TCPF_DELACK_TIMER_DEFERRED | \
1290 TCPF_MTU_REDUCED_DEFERRED | \
1291 TCPF_ACK_DEFERRED)
1292/**
1293 * tcp_release_cb - tcp release_sock() callback
1294 * @sk: socket
1295 *
1296 * called from release_sock() to perform protocol dependent
1297 * actions before socket release.
1298 */
1299void tcp_release_cb(struct sock *sk)
1300{
1301 unsigned long flags = smp_load_acquire(&sk->sk_tsq_flags);
1302 unsigned long nflags;
1303
1304 /* perform an atomic operation only if at least one flag is set */
1305 do {
1306 if (!(flags & TCP_DEFERRED_ALL))
1307 return;
1308 nflags = flags & ~TCP_DEFERRED_ALL;
1309 } while (!try_cmpxchg(&sk->sk_tsq_flags, &flags, nflags));
1310
1311 if (flags & TCPF_TSQ_DEFERRED) {
1312 tcp_tsq_write(sk);
1313 __sock_put(sk);
1314 }
1315
1316 if (flags & TCPF_WRITE_TIMER_DEFERRED) {
1317 tcp_write_timer_handler(sk);
1318 __sock_put(sk);
1319 }
1320 if (flags & TCPF_DELACK_TIMER_DEFERRED) {
1321 tcp_delack_timer_handler(sk);
1322 __sock_put(sk);
1323 }
1324 if (flags & TCPF_MTU_REDUCED_DEFERRED) {
1325 inet_csk(sk)->icsk_af_ops->mtu_reduced(sk);
1326 __sock_put(sk);
1327 }
1328 if ((flags & TCPF_ACK_DEFERRED) && inet_csk_ack_scheduled(sk))
1329 tcp_send_ack(sk);
1330}
1331EXPORT_IPV6_MOD(tcp_release_cb);
1332
1333void __init tcp_tsq_work_init(void)
1334{
1335 int i;
1336
1337 for_each_possible_cpu(i) {
1338 struct tsq_work *tsq = &per_cpu(tsq_work, i);
1339
1340 INIT_LIST_HEAD(list: &tsq->head);
1341 INIT_WORK(&tsq->work, tcp_tsq_workfn);
1342 }
1343}
1344
1345/*
1346 * Write buffer destructor automatically called from kfree_skb.
1347 * We can't xmit new skbs from this context, as we might already
1348 * hold qdisc lock.
1349 */
1350void tcp_wfree(struct sk_buff *skb)
1351{
1352 struct sock *sk = skb->sk;
1353 struct tcp_sock *tp = tcp_sk(sk);
1354 unsigned long flags, nval, oval;
1355 struct tsq_work *tsq;
1356 bool empty;
1357
1358 /* Keep one reference on sk_wmem_alloc.
1359 * Will be released by sk_free() from here or tcp_tsq_workfn()
1360 */
1361 WARN_ON(refcount_sub_and_test(skb->truesize - 1, &sk->sk_wmem_alloc));
1362
1363 /* If this softirq is serviced by ksoftirqd, we are likely under stress.
1364 * Wait until our queues (qdisc + devices) are drained.
1365 * This gives :
1366 * - less callbacks to tcp_write_xmit(), reducing stress (batches)
1367 * - chance for incoming ACK (processed by another cpu maybe)
1368 * to migrate this flow (skb->ooo_okay will be eventually set)
1369 */
1370 if (refcount_read(r: &sk->sk_wmem_alloc) >= SKB_TRUESIZE(1) && this_cpu_ksoftirqd() == current)
1371 goto out;
1372
1373 oval = smp_load_acquire(&sk->sk_tsq_flags);
1374 do {
1375 if (!(oval & TSQF_THROTTLED) || (oval & TSQF_QUEUED))
1376 goto out;
1377
1378 nval = (oval & ~TSQF_THROTTLED) | TSQF_QUEUED;
1379 } while (!try_cmpxchg(&sk->sk_tsq_flags, &oval, nval));
1380
1381 /* queue this socket to BH workqueue */
1382 local_irq_save(flags);
1383 tsq = this_cpu_ptr(&tsq_work);
1384 empty = list_empty(head: &tsq->head);
1385 list_add(new: &tp->tsq_node, head: &tsq->head);
1386 if (empty)
1387 queue_work(wq: system_bh_wq, work: &tsq->work);
1388 local_irq_restore(flags);
1389 return;
1390out:
1391 sk_free(sk);
1392}
1393
1394/* Note: Called under soft irq.
1395 * We can call TCP stack right away, unless socket is owned by user.
1396 */
1397enum hrtimer_restart tcp_pace_kick(struct hrtimer *timer)
1398{
1399 struct tcp_sock *tp = container_of(timer, struct tcp_sock, pacing_timer);
1400 struct sock *sk = (struct sock *)tp;
1401
1402 tcp_tsq_handler(sk);
1403 sock_put(sk);
1404
1405 return HRTIMER_NORESTART;
1406}
1407
1408static void tcp_update_skb_after_send(struct sock *sk, struct sk_buff *skb,
1409 u64 prior_wstamp)
1410{
1411 struct tcp_sock *tp = tcp_sk(sk);
1412
1413 if (sk->sk_pacing_status != SK_PACING_NONE) {
1414 unsigned long rate = READ_ONCE(sk->sk_pacing_rate);
1415
1416 /* Original sch_fq does not pace first 10 MSS
1417 * Note that tp->data_segs_out overflows after 2^32 packets,
1418 * this is a minor annoyance.
1419 */
1420 if (rate != ~0UL && rate && tp->data_segs_out >= 10) {
1421 u64 len_ns = div64_ul((u64)skb->len * NSEC_PER_SEC, rate);
1422 u64 credit = tp->tcp_wstamp_ns - prior_wstamp;
1423
1424 /* take into account OS jitter */
1425 len_ns -= min_t(u64, len_ns / 2, credit);
1426 tp->tcp_wstamp_ns += len_ns;
1427 }
1428 }
1429 list_move_tail(list: &skb->tcp_tsorted_anchor, head: &tp->tsorted_sent_queue);
1430}
1431
1432INDIRECT_CALLABLE_DECLARE(int ip_queue_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1433INDIRECT_CALLABLE_DECLARE(int inet6_csk_xmit(struct sock *sk, struct sk_buff *skb, struct flowi *fl));
1434INDIRECT_CALLABLE_DECLARE(void tcp_v4_send_check(struct sock *sk, struct sk_buff *skb));
1435
1436/* This routine actually transmits TCP packets queued in by
1437 * tcp_do_sendmsg(). This is used by both the initial
1438 * transmission and possible later retransmissions.
1439 * All SKB's seen here are completely headerless. It is our
1440 * job to build the TCP header, and pass the packet down to
1441 * IP so it can do the same plus pass the packet off to the
1442 * device.
1443 *
1444 * We are working here with either a clone of the original
1445 * SKB, or a fresh unique copy made by the retransmit engine.
1446 */
1447static int __tcp_transmit_skb(struct sock *sk, struct sk_buff *skb,
1448 int clone_it, gfp_t gfp_mask, u32 rcv_nxt)
1449{
1450 const struct inet_connection_sock *icsk = inet_csk(sk);
1451 struct inet_sock *inet;
1452 struct tcp_sock *tp;
1453 struct tcp_skb_cb *tcb;
1454 struct tcp_out_options opts;
1455 unsigned int tcp_options_size, tcp_header_size;
1456 struct sk_buff *oskb = NULL;
1457 struct tcp_key key;
1458 struct tcphdr *th;
1459 u64 prior_wstamp;
1460 int err;
1461
1462 BUG_ON(!skb || !tcp_skb_pcount(skb));
1463 tp = tcp_sk(sk);
1464 prior_wstamp = tp->tcp_wstamp_ns;
1465 tp->tcp_wstamp_ns = max(tp->tcp_wstamp_ns, tp->tcp_clock_cache);
1466 skb_set_delivery_time(skb, kt: tp->tcp_wstamp_ns, tstamp_type: SKB_CLOCK_MONOTONIC);
1467 if (clone_it) {
1468 oskb = skb;
1469
1470 tcp_skb_tsorted_save(oskb) {
1471 if (unlikely(skb_cloned(oskb)))
1472 skb = pskb_copy(skb: oskb, gfp_mask);
1473 else
1474 skb = skb_clone(skb: oskb, priority: gfp_mask);
1475 } tcp_skb_tsorted_restore(oskb);
1476
1477 if (unlikely(!skb))
1478 return -ENOBUFS;
1479 /* retransmit skbs might have a non zero value in skb->dev
1480 * because skb->dev is aliased with skb->rbnode.rb_left
1481 */
1482 skb->dev = NULL;
1483 }
1484
1485 inet = inet_sk(sk);
1486 tcb = TCP_SKB_CB(skb);
1487 memset(s: &opts, c: 0, n: sizeof(opts));
1488
1489 tcp_get_current_key(sk, out: &key);
1490 if (unlikely(tcb->tcp_flags & TCPHDR_SYN)) {
1491 tcp_options_size = tcp_syn_options(sk, skb, opts: &opts, key: &key);
1492 } else {
1493 tcp_options_size = tcp_established_options(sk, skb, opts: &opts, key: &key);
1494 /* Force a PSH flag on all (GSO) packets to expedite GRO flush
1495 * at receiver : This slightly improve GRO performance.
1496 * Note that we do not force the PSH flag for non GSO packets,
1497 * because they might be sent under high congestion events,
1498 * and in this case it is better to delay the delivery of 1-MSS
1499 * packets and thus the corresponding ACK packet that would
1500 * release the following packet.
1501 */
1502 if (tcp_skb_pcount(skb) > 1)
1503 tcb->tcp_flags |= TCPHDR_PSH;
1504 }
1505 tcp_header_size = tcp_options_size + sizeof(struct tcphdr);
1506
1507 /* We set skb->ooo_okay to one if this packet can select
1508 * a different TX queue than prior packets of this flow,
1509 * to avoid self inflicted reorders.
1510 * The 'other' queue decision is based on current cpu number
1511 * if XPS is enabled, or sk->sk_txhash otherwise.
1512 * We can switch to another (and better) queue if:
1513 * 1) No packet with payload is in qdisc/device queues.
1514 * Delays in TX completion can defeat the test
1515 * even if packets were already sent.
1516 * 2) Or rtx queue is empty.
1517 * This mitigates above case if ACK packets for
1518 * all prior packets were already processed.
1519 */
1520 skb->ooo_okay = sk_wmem_alloc_get(sk) < SKB_TRUESIZE(1) ||
1521 tcp_rtx_queue_empty(sk);
1522
1523 /* If we had to use memory reserve to allocate this skb,
1524 * this might cause drops if packet is looped back :
1525 * Other socket might not have SOCK_MEMALLOC.
1526 * Packets not looped back do not care about pfmemalloc.
1527 */
1528 skb->pfmemalloc = 0;
1529
1530 skb_push(skb, len: tcp_header_size);
1531 skb_reset_transport_header(skb);
1532
1533 skb_orphan(skb);
1534 skb->sk = sk;
1535 skb->destructor = skb_is_tcp_pure_ack(skb) ? __sock_wfree : tcp_wfree;
1536 refcount_add(i: skb->truesize, r: &sk->sk_wmem_alloc);
1537
1538 skb_set_dst_pending_confirm(skb, READ_ONCE(sk->sk_dst_pending_confirm));
1539
1540 /* Build TCP header and checksum it. */
1541 th = (struct tcphdr *)skb->data;
1542 th->source = inet->inet_sport;
1543 th->dest = inet->inet_dport;
1544 th->seq = htonl(tcb->seq);
1545 th->ack_seq = htonl(rcv_nxt);
1546 *(((__be16 *)th) + 6) = htons(((tcp_header_size >> 2) << 12) |
1547 (tcb->tcp_flags & TCPHDR_FLAGS_MASK));
1548
1549 th->check = 0;
1550 th->urg_ptr = 0;
1551
1552 /* The urg_mode check is necessary during a below snd_una win probe */
1553 if (unlikely(tcp_urg_mode(tp) && before(tcb->seq, tp->snd_up))) {
1554 if (before(seq1: tp->snd_up, seq2: tcb->seq + 0x10000)) {
1555 th->urg_ptr = htons(tp->snd_up - tcb->seq);
1556 th->urg = 1;
1557 } else if (after(tcb->seq + 0xFFFF, tp->snd_nxt)) {
1558 th->urg_ptr = htons(0xFFFF);
1559 th->urg = 1;
1560 }
1561 }
1562
1563 skb_shinfo(skb)->gso_type = sk->sk_gso_type;
1564 if (likely(!(tcb->tcp_flags & TCPHDR_SYN))) {
1565 th->window = htons(tcp_select_window(sk));
1566 tcp_ecn_send(sk, skb, th, tcp_header_len: tcp_header_size);
1567 } else {
1568 /* RFC1323: The window in SYN & SYN/ACK segments
1569 * is never scaled.
1570 */
1571 th->window = htons(min(tp->rcv_wnd, 65535U));
1572 }
1573
1574 tcp_options_write(th, tp, NULL, opts: &opts, key: &key);
1575
1576 if (tcp_key_is_md5(key: &key)) {
1577#ifdef CONFIG_TCP_MD5SIG
1578 /* Calculate the MD5 hash, as we have all we need now */
1579 sk_gso_disable(sk);
1580 tp->af_specific->calc_md5_hash(opts.hash_location,
1581 key.md5_key, sk, skb);
1582#endif
1583 } else if (tcp_key_is_ao(key: &key)) {
1584 int err;
1585
1586 err = tcp_ao_transmit_skb(sk, skb, key: key.ao_key, th,
1587 hash_location: opts.hash_location);
1588 if (err) {
1589 sk_skb_reason_drop(sk, skb, reason: SKB_DROP_REASON_NOT_SPECIFIED);
1590 return -ENOMEM;
1591 }
1592 }
1593
1594 /* BPF prog is the last one writing header option */
1595 bpf_skops_write_hdr_opt(sk, skb, NULL, NULL, synack_type: 0, opts: &opts);
1596
1597 INDIRECT_CALL_INET(icsk->icsk_af_ops->send_check,
1598 tcp_v6_send_check, tcp_v4_send_check,
1599 sk, skb);
1600
1601 if (likely(tcb->tcp_flags & TCPHDR_ACK))
1602 tcp_event_ack_sent(sk, rcv_nxt);
1603
1604 if (skb->len != tcp_header_size) {
1605 tcp_event_data_sent(tp, sk);
1606 tp->data_segs_out += tcp_skb_pcount(skb);
1607 tp->bytes_sent += skb->len - tcp_header_size;
1608 }
1609
1610 if (after(tcb->end_seq, tp->snd_nxt) || tcb->seq == tcb->end_seq)
1611 TCP_ADD_STATS(sock_net(sk), TCP_MIB_OUTSEGS,
1612 tcp_skb_pcount(skb));
1613
1614 tp->segs_out += tcp_skb_pcount(skb);
1615 skb_set_hash_from_sk(skb, sk);
1616 /* OK, its time to fill skb_shinfo(skb)->gso_{segs|size} */
1617 skb_shinfo(skb)->gso_segs = tcp_skb_pcount(skb);
1618 skb_shinfo(skb)->gso_size = tcp_skb_mss(skb);
1619
1620 /* Leave earliest departure time in skb->tstamp (skb->skb_mstamp_ns) */
1621
1622 /* Cleanup our debris for IP stacks */
1623 memset(s: skb->cb, c: 0, max(sizeof(struct inet_skb_parm),
1624 sizeof(struct inet6_skb_parm)));
1625
1626 tcp_add_tx_delay(skb, tp);
1627
1628 err = INDIRECT_CALL_INET(icsk->icsk_af_ops->queue_xmit,
1629 inet6_csk_xmit, ip_queue_xmit,
1630 sk, skb, &inet->cork.fl);
1631
1632 if (unlikely(err > 0)) {
1633 tcp_enter_cwr(sk);
1634 err = net_xmit_eval(err);
1635 }
1636 if (!err && oskb) {
1637 tcp_update_skb_after_send(sk, skb: oskb, prior_wstamp);
1638 tcp_rate_skb_sent(sk, skb: oskb);
1639 }
1640 return err;
1641}
1642
1643static int tcp_transmit_skb(struct sock *sk, struct sk_buff *skb, int clone_it,
1644 gfp_t gfp_mask)
1645{
1646 return __tcp_transmit_skb(sk, skb, clone_it, gfp_mask,
1647 tcp_sk(sk)->rcv_nxt);
1648}
1649
1650/* This routine just queues the buffer for sending.
1651 *
1652 * NOTE: probe0 timer is not checked, do not forget tcp_push_pending_frames,
1653 * otherwise socket can stall.
1654 */
1655static void tcp_queue_skb(struct sock *sk, struct sk_buff *skb)
1656{
1657 struct tcp_sock *tp = tcp_sk(sk);
1658
1659 /* Advance write_seq and place onto the write_queue. */
1660 WRITE_ONCE(tp->write_seq, TCP_SKB_CB(skb)->end_seq);
1661 __skb_header_release(skb);
1662 psp_enqueue_set_decrypted(sk, skb);
1663 tcp_add_write_queue_tail(sk, skb);
1664 sk_wmem_queued_add(sk, val: skb->truesize);
1665 sk_mem_charge(sk, size: skb->truesize);
1666}
1667
1668/* Initialize TSO segments for a packet. */
1669static int tcp_set_skb_tso_segs(struct sk_buff *skb, unsigned int mss_now)
1670{
1671 int tso_segs;
1672
1673 if (skb->len <= mss_now) {
1674 /* Avoid the costly divide in the normal
1675 * non-TSO case.
1676 */
1677 TCP_SKB_CB(skb)->tcp_gso_size = 0;
1678 tcp_skb_pcount_set(skb, segs: 1);
1679 return 1;
1680 }
1681 TCP_SKB_CB(skb)->tcp_gso_size = mss_now;
1682 tso_segs = DIV_ROUND_UP(skb->len, mss_now);
1683 tcp_skb_pcount_set(skb, segs: tso_segs);
1684 return tso_segs;
1685}
1686
1687/* Pcount in the middle of the write queue got changed, we need to do various
1688 * tweaks to fix counters
1689 */
1690static void tcp_adjust_pcount(struct sock *sk, const struct sk_buff *skb, int decr)
1691{
1692 struct tcp_sock *tp = tcp_sk(sk);
1693
1694 tp->packets_out -= decr;
1695
1696 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
1697 tp->sacked_out -= decr;
1698 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS)
1699 tp->retrans_out -= decr;
1700 if (TCP_SKB_CB(skb)->sacked & TCPCB_LOST)
1701 tp->lost_out -= decr;
1702
1703 /* Reno case is special. Sigh... */
1704 if (tcp_is_reno(tp) && decr > 0)
1705 tp->sacked_out -= min_t(u32, tp->sacked_out, decr);
1706
1707 tcp_verify_left_out(tp);
1708}
1709
1710static bool tcp_has_tx_tstamp(const struct sk_buff *skb)
1711{
1712 return TCP_SKB_CB(skb)->txstamp_ack ||
1713 (skb_shinfo(skb)->tx_flags & SKBTX_ANY_TSTAMP);
1714}
1715
1716static void tcp_fragment_tstamp(struct sk_buff *skb, struct sk_buff *skb2)
1717{
1718 struct skb_shared_info *shinfo = skb_shinfo(skb);
1719
1720 if (unlikely(tcp_has_tx_tstamp(skb)) &&
1721 !before(seq1: shinfo->tskey, TCP_SKB_CB(skb2)->seq)) {
1722 struct skb_shared_info *shinfo2 = skb_shinfo(skb2);
1723 u8 tsflags = shinfo->tx_flags & SKBTX_ANY_TSTAMP;
1724
1725 shinfo->tx_flags &= ~tsflags;
1726 shinfo2->tx_flags |= tsflags;
1727 swap(shinfo->tskey, shinfo2->tskey);
1728 TCP_SKB_CB(skb2)->txstamp_ack = TCP_SKB_CB(skb)->txstamp_ack;
1729 TCP_SKB_CB(skb)->txstamp_ack = 0;
1730 }
1731}
1732
1733static void tcp_skb_fragment_eor(struct sk_buff *skb, struct sk_buff *skb2)
1734{
1735 TCP_SKB_CB(skb2)->eor = TCP_SKB_CB(skb)->eor;
1736 TCP_SKB_CB(skb)->eor = 0;
1737}
1738
1739/* Insert buff after skb on the write or rtx queue of sk. */
1740static void tcp_insert_write_queue_after(struct sk_buff *skb,
1741 struct sk_buff *buff,
1742 struct sock *sk,
1743 enum tcp_queue tcp_queue)
1744{
1745 if (tcp_queue == TCP_FRAG_IN_WRITE_QUEUE)
1746 __skb_queue_after(list: &sk->sk_write_queue, prev: skb, newsk: buff);
1747 else
1748 tcp_rbtree_insert(root: &sk->tcp_rtx_queue, skb: buff);
1749}
1750
1751/* Function to create two new TCP segments. Shrinks the given segment
1752 * to the specified size and appends a new segment with the rest of the
1753 * packet to the list. This won't be called frequently, I hope.
1754 * Remember, these are still headerless SKBs at this point.
1755 */
1756int tcp_fragment(struct sock *sk, enum tcp_queue tcp_queue,
1757 struct sk_buff *skb, u32 len,
1758 unsigned int mss_now, gfp_t gfp)
1759{
1760 struct tcp_sock *tp = tcp_sk(sk);
1761 struct sk_buff *buff;
1762 int old_factor;
1763 long limit;
1764 u16 flags;
1765 int nlen;
1766
1767 if (WARN_ON(len > skb->len))
1768 return -EINVAL;
1769
1770 DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb));
1771
1772 /* tcp_sendmsg() can overshoot sk_wmem_queued by one full size skb.
1773 * We need some allowance to not penalize applications setting small
1774 * SO_SNDBUF values.
1775 * Also allow first and last skb in retransmit queue to be split.
1776 */
1777 limit = sk->sk_sndbuf + 2 * SKB_TRUESIZE(GSO_LEGACY_MAX_SIZE);
1778 if (unlikely((sk->sk_wmem_queued >> 1) > limit &&
1779 tcp_queue != TCP_FRAG_IN_WRITE_QUEUE &&
1780 skb != tcp_rtx_queue_head(sk) &&
1781 skb != tcp_rtx_queue_tail(sk))) {
1782 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPWQUEUETOOBIG);
1783 return -ENOMEM;
1784 }
1785
1786 if (skb_unclone_keeptruesize(skb, pri: gfp))
1787 return -ENOMEM;
1788
1789 /* Get a new skb... force flag on. */
1790 buff = tcp_stream_alloc_skb(sk, gfp, force_schedule: true);
1791 if (!buff)
1792 return -ENOMEM; /* We'll just try again later. */
1793 skb_copy_decrypted(to: buff, from: skb);
1794 mptcp_skb_ext_copy(to: buff, from: skb);
1795
1796 sk_wmem_queued_add(sk, val: buff->truesize);
1797 sk_mem_charge(sk, size: buff->truesize);
1798 nlen = skb->len - len;
1799 buff->truesize += nlen;
1800 skb->truesize -= nlen;
1801
1802 /* Correct the sequence numbers. */
1803 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
1804 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
1805 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
1806
1807 /* PSH and FIN should only be set in the second packet. */
1808 flags = TCP_SKB_CB(skb)->tcp_flags;
1809 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
1810 TCP_SKB_CB(buff)->tcp_flags = flags;
1811 TCP_SKB_CB(buff)->sacked = TCP_SKB_CB(skb)->sacked;
1812 tcp_skb_fragment_eor(skb, skb2: buff);
1813
1814 skb_split(skb, skb1: buff, len);
1815
1816 skb_set_delivery_time(skb: buff, kt: skb->tstamp, tstamp_type: SKB_CLOCK_MONOTONIC);
1817 tcp_fragment_tstamp(skb, skb2: buff);
1818
1819 old_factor = tcp_skb_pcount(skb);
1820
1821 /* Fix up tso_factor for both original and new SKB. */
1822 tcp_set_skb_tso_segs(skb, mss_now);
1823 tcp_set_skb_tso_segs(skb: buff, mss_now);
1824
1825 /* Update delivered info for the new segment */
1826 TCP_SKB_CB(buff)->tx = TCP_SKB_CB(skb)->tx;
1827
1828 /* If this packet has been sent out already, we must
1829 * adjust the various packet counters.
1830 */
1831 if (!before(seq1: tp->snd_nxt, TCP_SKB_CB(buff)->end_seq)) {
1832 int diff = old_factor - tcp_skb_pcount(skb) -
1833 tcp_skb_pcount(skb: buff);
1834
1835 if (diff)
1836 tcp_adjust_pcount(sk, skb, decr: diff);
1837 }
1838
1839 /* Link BUFF into the send queue. */
1840 __skb_header_release(skb: buff);
1841 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue);
1842 if (tcp_queue == TCP_FRAG_IN_RTX_QUEUE)
1843 list_add(new: &buff->tcp_tsorted_anchor, head: &skb->tcp_tsorted_anchor);
1844
1845 return 0;
1846}
1847
1848/* This is similar to __pskb_pull_tail(). The difference is that pulled
1849 * data is not copied, but immediately discarded.
1850 */
1851static int __pskb_trim_head(struct sk_buff *skb, int len)
1852{
1853 struct skb_shared_info *shinfo;
1854 int i, k, eat;
1855
1856 DEBUG_NET_WARN_ON_ONCE(skb_headlen(skb));
1857 eat = len;
1858 k = 0;
1859 shinfo = skb_shinfo(skb);
1860 for (i = 0; i < shinfo->nr_frags; i++) {
1861 int size = skb_frag_size(frag: &shinfo->frags[i]);
1862
1863 if (size <= eat) {
1864 skb_frag_unref(skb, f: i);
1865 eat -= size;
1866 } else {
1867 shinfo->frags[k] = shinfo->frags[i];
1868 if (eat) {
1869 skb_frag_off_add(frag: &shinfo->frags[k], delta: eat);
1870 skb_frag_size_sub(frag: &shinfo->frags[k], delta: eat);
1871 eat = 0;
1872 }
1873 k++;
1874 }
1875 }
1876 shinfo->nr_frags = k;
1877
1878 skb->data_len -= len;
1879 skb->len = skb->data_len;
1880 return len;
1881}
1882
1883/* Remove acked data from a packet in the transmit queue. */
1884int tcp_trim_head(struct sock *sk, struct sk_buff *skb, u32 len)
1885{
1886 u32 delta_truesize;
1887
1888 if (skb_unclone_keeptruesize(skb, GFP_ATOMIC))
1889 return -ENOMEM;
1890
1891 delta_truesize = __pskb_trim_head(skb, len);
1892
1893 TCP_SKB_CB(skb)->seq += len;
1894
1895 skb->truesize -= delta_truesize;
1896 sk_wmem_queued_add(sk, val: -delta_truesize);
1897 if (!skb_zcopy_pure(skb))
1898 sk_mem_uncharge(sk, size: delta_truesize);
1899
1900 /* Any change of skb->len requires recalculation of tso factor. */
1901 if (tcp_skb_pcount(skb) > 1)
1902 tcp_set_skb_tso_segs(skb, mss_now: tcp_skb_mss(skb));
1903
1904 return 0;
1905}
1906
1907/* Calculate MSS not accounting any TCP options. */
1908static inline int __tcp_mtu_to_mss(struct sock *sk, int pmtu)
1909{
1910 const struct tcp_sock *tp = tcp_sk(sk);
1911 const struct inet_connection_sock *icsk = inet_csk(sk);
1912 int mss_now;
1913
1914 /* Calculate base mss without TCP options:
1915 It is MMS_S - sizeof(tcphdr) of rfc1122
1916 */
1917 mss_now = pmtu - icsk->icsk_af_ops->net_header_len - sizeof(struct tcphdr);
1918
1919 /* Clamp it (mss_clamp does not include tcp options) */
1920 if (mss_now > tp->rx_opt.mss_clamp)
1921 mss_now = tp->rx_opt.mss_clamp;
1922
1923 /* Now subtract optional transport overhead */
1924 mss_now -= icsk->icsk_ext_hdr_len;
1925
1926 /* Then reserve room for full set of TCP options and 8 bytes of data */
1927 mss_now = max(mss_now,
1928 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_snd_mss));
1929 return mss_now;
1930}
1931
1932/* Calculate MSS. Not accounting for SACKs here. */
1933int tcp_mtu_to_mss(struct sock *sk, int pmtu)
1934{
1935 /* Subtract TCP options size, not including SACKs */
1936 return __tcp_mtu_to_mss(sk, pmtu) -
1937 (tcp_sk(sk)->tcp_header_len - sizeof(struct tcphdr));
1938}
1939EXPORT_IPV6_MOD(tcp_mtu_to_mss);
1940
1941/* Inverse of above */
1942int tcp_mss_to_mtu(struct sock *sk, int mss)
1943{
1944 const struct tcp_sock *tp = tcp_sk(sk);
1945 const struct inet_connection_sock *icsk = inet_csk(sk);
1946
1947 return mss +
1948 tp->tcp_header_len +
1949 icsk->icsk_ext_hdr_len +
1950 icsk->icsk_af_ops->net_header_len;
1951}
1952EXPORT_SYMBOL(tcp_mss_to_mtu);
1953
1954/* MTU probing init per socket */
1955void tcp_mtup_init(struct sock *sk)
1956{
1957 struct tcp_sock *tp = tcp_sk(sk);
1958 struct inet_connection_sock *icsk = inet_csk(sk);
1959 struct net *net = sock_net(sk);
1960
1961 icsk->icsk_mtup.enabled = READ_ONCE(net->ipv4.sysctl_tcp_mtu_probing) > 1;
1962 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp + sizeof(struct tcphdr) +
1963 icsk->icsk_af_ops->net_header_len;
1964 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, READ_ONCE(net->ipv4.sysctl_tcp_base_mss));
1965 icsk->icsk_mtup.probe_size = 0;
1966 if (icsk->icsk_mtup.enabled)
1967 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
1968}
1969
1970/* This function synchronize snd mss to current pmtu/exthdr set.
1971
1972 tp->rx_opt.user_mss is mss set by user by TCP_MAXSEG. It does NOT counts
1973 for TCP options, but includes only bare TCP header.
1974
1975 tp->rx_opt.mss_clamp is mss negotiated at connection setup.
1976 It is minimum of user_mss and mss received with SYN.
1977 It also does not include TCP options.
1978
1979 inet_csk(sk)->icsk_pmtu_cookie is last pmtu, seen by this function.
1980
1981 tp->mss_cache is current effective sending mss, including
1982 all tcp options except for SACKs. It is evaluated,
1983 taking into account current pmtu, but never exceeds
1984 tp->rx_opt.mss_clamp.
1985
1986 NOTE1. rfc1122 clearly states that advertised MSS
1987 DOES NOT include either tcp or ip options.
1988
1989 NOTE2. inet_csk(sk)->icsk_pmtu_cookie and tp->mss_cache
1990 are READ ONLY outside this function. --ANK (980731)
1991 */
1992unsigned int tcp_sync_mss(struct sock *sk, u32 pmtu)
1993{
1994 struct tcp_sock *tp = tcp_sk(sk);
1995 struct inet_connection_sock *icsk = inet_csk(sk);
1996 int mss_now;
1997
1998 if (icsk->icsk_mtup.search_high > pmtu)
1999 icsk->icsk_mtup.search_high = pmtu;
2000
2001 mss_now = tcp_mtu_to_mss(sk, pmtu);
2002 mss_now = tcp_bound_to_half_wnd(tp, pktsize: mss_now);
2003
2004 /* And store cached results */
2005 icsk->icsk_pmtu_cookie = pmtu;
2006 if (icsk->icsk_mtup.enabled)
2007 mss_now = min(mss_now, tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low));
2008 tp->mss_cache = mss_now;
2009
2010 return mss_now;
2011}
2012EXPORT_IPV6_MOD(tcp_sync_mss);
2013
2014/* Compute the current effective MSS, taking SACKs and IP options,
2015 * and even PMTU discovery events into account.
2016 */
2017unsigned int tcp_current_mss(struct sock *sk)
2018{
2019 const struct tcp_sock *tp = tcp_sk(sk);
2020 const struct dst_entry *dst = __sk_dst_get(sk);
2021 u32 mss_now;
2022 unsigned int header_len;
2023 struct tcp_out_options opts;
2024 struct tcp_key key;
2025
2026 mss_now = tp->mss_cache;
2027
2028 if (dst) {
2029 u32 mtu = dst_mtu(dst);
2030 if (mtu != inet_csk(sk)->icsk_pmtu_cookie)
2031 mss_now = tcp_sync_mss(sk, pmtu: mtu);
2032 }
2033 tcp_get_current_key(sk, out: &key);
2034 header_len = tcp_established_options(sk, NULL, opts: &opts, key: &key) +
2035 sizeof(struct tcphdr);
2036 /* The mss_cache is sized based on tp->tcp_header_len, which assumes
2037 * some common options. If this is an odd packet (because we have SACK
2038 * blocks etc) then our calculated header_len will be different, and
2039 * we have to adjust mss_now correspondingly */
2040 if (header_len != tp->tcp_header_len) {
2041 int delta = (int) header_len - tp->tcp_header_len;
2042 mss_now -= delta;
2043 }
2044
2045 return mss_now;
2046}
2047
2048/* RFC2861, slow part. Adjust cwnd, after it was not full during one rto.
2049 * As additional protections, we do not touch cwnd in retransmission phases,
2050 * and if application hit its sndbuf limit recently.
2051 */
2052static void tcp_cwnd_application_limited(struct sock *sk)
2053{
2054 struct tcp_sock *tp = tcp_sk(sk);
2055
2056 if (inet_csk(sk)->icsk_ca_state == TCP_CA_Open &&
2057 sk->sk_socket && !test_bit(SOCK_NOSPACE, &sk->sk_socket->flags)) {
2058 /* Limited by application or receiver window. */
2059 u32 init_win = tcp_init_cwnd(tp, dst: __sk_dst_get(sk));
2060 u32 win_used = max(tp->snd_cwnd_used, init_win);
2061 if (win_used < tcp_snd_cwnd(tp)) {
2062 tp->snd_ssthresh = tcp_current_ssthresh(sk);
2063 tcp_snd_cwnd_set(tp, val: (tcp_snd_cwnd(tp) + win_used) >> 1);
2064 }
2065 tp->snd_cwnd_used = 0;
2066 }
2067 tp->snd_cwnd_stamp = tcp_jiffies32;
2068}
2069
2070static void tcp_cwnd_validate(struct sock *sk, bool is_cwnd_limited)
2071{
2072 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2073 struct tcp_sock *tp = tcp_sk(sk);
2074
2075 /* Track the strongest available signal of the degree to which the cwnd
2076 * is fully utilized. If cwnd-limited then remember that fact for the
2077 * current window. If not cwnd-limited then track the maximum number of
2078 * outstanding packets in the current window. (If cwnd-limited then we
2079 * chose to not update tp->max_packets_out to avoid an extra else
2080 * clause with no functional impact.)
2081 */
2082 if (!before(seq1: tp->snd_una, seq2: tp->cwnd_usage_seq) ||
2083 is_cwnd_limited ||
2084 (!tp->is_cwnd_limited &&
2085 tp->packets_out > tp->max_packets_out)) {
2086 tp->is_cwnd_limited = is_cwnd_limited;
2087 tp->max_packets_out = tp->packets_out;
2088 tp->cwnd_usage_seq = tp->snd_nxt;
2089 }
2090
2091 if (tcp_is_cwnd_limited(sk)) {
2092 /* Network is feed fully. */
2093 tp->snd_cwnd_used = 0;
2094 tp->snd_cwnd_stamp = tcp_jiffies32;
2095 } else {
2096 /* Network starves. */
2097 if (tp->packets_out > tp->snd_cwnd_used)
2098 tp->snd_cwnd_used = tp->packets_out;
2099
2100 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_slow_start_after_idle) &&
2101 (s32)(tcp_jiffies32 - tp->snd_cwnd_stamp) >= inet_csk(sk)->icsk_rto &&
2102 !ca_ops->cong_control)
2103 tcp_cwnd_application_limited(sk);
2104
2105 /* The following conditions together indicate the starvation
2106 * is caused by insufficient sender buffer:
2107 * 1) just sent some data (see tcp_write_xmit)
2108 * 2) not cwnd limited (this else condition)
2109 * 3) no more data to send (tcp_write_queue_empty())
2110 * 4) application is hitting buffer limit (SOCK_NOSPACE)
2111 */
2112 if (tcp_write_queue_empty(sk) && sk->sk_socket &&
2113 test_bit(SOCK_NOSPACE, &sk->sk_socket->flags) &&
2114 (1 << sk->sk_state) & (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT))
2115 tcp_chrono_start(sk, type: TCP_CHRONO_SNDBUF_LIMITED);
2116 }
2117}
2118
2119/* Minshall's variant of the Nagle send check. */
2120static bool tcp_minshall_check(const struct tcp_sock *tp)
2121{
2122 return after(tp->snd_sml, tp->snd_una) &&
2123 !after(tp->snd_sml, tp->snd_nxt);
2124}
2125
2126/* Update snd_sml if this skb is under mss
2127 * Note that a TSO packet might end with a sub-mss segment
2128 * The test is really :
2129 * if ((skb->len % mss) != 0)
2130 * tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
2131 * But we can avoid doing the divide again given we already have
2132 * skb_pcount = skb->len / mss_now
2133 */
2134static void tcp_minshall_update(struct tcp_sock *tp, unsigned int mss_now,
2135 const struct sk_buff *skb)
2136{
2137 if (skb->len < tcp_skb_pcount(skb) * mss_now)
2138 tp->snd_sml = TCP_SKB_CB(skb)->end_seq;
2139}
2140
2141/* Return false, if packet can be sent now without violation Nagle's rules:
2142 * 1. It is full sized. (provided by caller in %partial bool)
2143 * 2. Or it contains FIN. (already checked by caller)
2144 * 3. Or TCP_CORK is not set, and TCP_NODELAY is set.
2145 * 4. Or TCP_CORK is not set, and all sent packets are ACKed.
2146 * With Minshall's modification: all sent small packets are ACKed.
2147 */
2148static bool tcp_nagle_check(bool partial, const struct tcp_sock *tp,
2149 int nonagle)
2150{
2151 return partial &&
2152 ((nonagle & TCP_NAGLE_CORK) ||
2153 (!nonagle && tp->packets_out && tcp_minshall_check(tp)));
2154}
2155
2156/* Return how many segs we'd like on a TSO packet,
2157 * depending on current pacing rate, and how close the peer is.
2158 *
2159 * Rationale is:
2160 * - For close peers, we rather send bigger packets to reduce
2161 * cpu costs, because occasional losses will be repaired fast.
2162 * - For long distance/rtt flows, we would like to get ACK clocking
2163 * with 1 ACK per ms.
2164 *
2165 * Use min_rtt to help adapt TSO burst size, with smaller min_rtt resulting
2166 * in bigger TSO bursts. We we cut the RTT-based allowance in half
2167 * for every 2^9 usec (aka 512 us) of RTT, so that the RTT-based allowance
2168 * is below 1500 bytes after 6 * ~500 usec = 3ms.
2169 */
2170static u32 tcp_tso_autosize(const struct sock *sk, unsigned int mss_now,
2171 int min_tso_segs)
2172{
2173 unsigned long bytes;
2174 u32 r;
2175
2176 bytes = READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift);
2177
2178 r = tcp_min_rtt(tcp_sk(sk)) >> READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_rtt_log);
2179 if (r < BITS_PER_TYPE(sk->sk_gso_max_size))
2180 bytes += sk->sk_gso_max_size >> r;
2181
2182 bytes = min_t(unsigned long, bytes, sk->sk_gso_max_size);
2183
2184 return max_t(u32, bytes / mss_now, min_tso_segs);
2185}
2186
2187/* Return the number of segments we want in the skb we are transmitting.
2188 * See if congestion control module wants to decide; otherwise, autosize.
2189 */
2190static u32 tcp_tso_segs(struct sock *sk, unsigned int mss_now)
2191{
2192 const struct tcp_congestion_ops *ca_ops = inet_csk(sk)->icsk_ca_ops;
2193 u32 min_tso, tso_segs;
2194
2195 min_tso = ca_ops->min_tso_segs ?
2196 ca_ops->min_tso_segs(sk) :
2197 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_min_tso_segs);
2198
2199 tso_segs = tcp_tso_autosize(sk, mss_now, min_tso_segs: min_tso);
2200 return min_t(u32, tso_segs, sk->sk_gso_max_segs);
2201}
2202
2203/* Returns the portion of skb which can be sent right away */
2204static unsigned int tcp_mss_split_point(const struct sock *sk,
2205 const struct sk_buff *skb,
2206 unsigned int mss_now,
2207 unsigned int max_segs,
2208 int nonagle)
2209{
2210 const struct tcp_sock *tp = tcp_sk(sk);
2211 u32 partial, needed, window, max_len;
2212
2213 window = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2214 max_len = mss_now * max_segs;
2215
2216 if (likely(max_len <= window && skb != tcp_write_queue_tail(sk)))
2217 return max_len;
2218
2219 needed = min(skb->len, window);
2220
2221 if (max_len <= needed)
2222 return max_len;
2223
2224 partial = needed % mss_now;
2225 /* If last segment is not a full MSS, check if Nagle rules allow us
2226 * to include this last segment in this skb.
2227 * Otherwise, we'll split the skb at last MSS boundary
2228 */
2229 if (tcp_nagle_check(partial: partial != 0, tp, nonagle))
2230 return needed - partial;
2231
2232 return needed;
2233}
2234
2235/* Can at least one segment of SKB be sent right now, according to the
2236 * congestion window rules? If so, return how many segments are allowed.
2237 */
2238static u32 tcp_cwnd_test(const struct tcp_sock *tp)
2239{
2240 u32 in_flight, cwnd, halfcwnd;
2241
2242 in_flight = tcp_packets_in_flight(tp);
2243 cwnd = tcp_snd_cwnd(tp);
2244 if (in_flight >= cwnd)
2245 return 0;
2246
2247 /* For better scheduling, ensure we have at least
2248 * 2 GSO packets in flight.
2249 */
2250 halfcwnd = max(cwnd >> 1, 1U);
2251 return min(halfcwnd, cwnd - in_flight);
2252}
2253
2254/* Initialize TSO state of a skb.
2255 * This must be invoked the first time we consider transmitting
2256 * SKB onto the wire.
2257 */
2258static int tcp_init_tso_segs(struct sk_buff *skb, unsigned int mss_now)
2259{
2260 int tso_segs = tcp_skb_pcount(skb);
2261
2262 if (!tso_segs || (tso_segs > 1 && tcp_skb_mss(skb) != mss_now))
2263 return tcp_set_skb_tso_segs(skb, mss_now);
2264
2265 return tso_segs;
2266}
2267
2268
2269/* Return true if the Nagle test allows this packet to be
2270 * sent now.
2271 */
2272static inline bool tcp_nagle_test(const struct tcp_sock *tp, const struct sk_buff *skb,
2273 unsigned int cur_mss, int nonagle)
2274{
2275 /* Nagle rule does not apply to frames, which sit in the middle of the
2276 * write_queue (they have no chances to get new data).
2277 *
2278 * This is implemented in the callers, where they modify the 'nonagle'
2279 * argument based upon the location of SKB in the send queue.
2280 */
2281 if (nonagle & TCP_NAGLE_PUSH)
2282 return true;
2283
2284 /* Don't use the nagle rule for urgent data (or for the final FIN). */
2285 if (tcp_urg_mode(tp) || (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN))
2286 return true;
2287
2288 if (!tcp_nagle_check(partial: skb->len < cur_mss, tp, nonagle))
2289 return true;
2290
2291 return false;
2292}
2293
2294/* Does at least the first segment of SKB fit into the send window? */
2295static bool tcp_snd_wnd_test(const struct tcp_sock *tp,
2296 const struct sk_buff *skb,
2297 unsigned int cur_mss)
2298{
2299 u32 end_seq = TCP_SKB_CB(skb)->end_seq;
2300
2301 if (skb->len > cur_mss)
2302 end_seq = TCP_SKB_CB(skb)->seq + cur_mss;
2303
2304 return !after(end_seq, tcp_wnd_end(tp));
2305}
2306
2307/* Trim TSO SKB to LEN bytes, put the remaining data into a new packet
2308 * which is put after SKB on the list. It is very much like
2309 * tcp_fragment() except that it may make several kinds of assumptions
2310 * in order to speed up the splitting operation. In particular, we
2311 * know that all the data is in scatter-gather pages, and that the
2312 * packet has never been sent out before (and thus is not cloned).
2313 */
2314static int tso_fragment(struct sock *sk, struct sk_buff *skb, unsigned int len,
2315 unsigned int mss_now, gfp_t gfp)
2316{
2317 int nlen = skb->len - len;
2318 struct sk_buff *buff;
2319 u16 flags;
2320
2321 /* All of a TSO frame must be composed of paged data. */
2322 DEBUG_NET_WARN_ON_ONCE(skb->len != skb->data_len);
2323
2324 buff = tcp_stream_alloc_skb(sk, gfp, force_schedule: true);
2325 if (unlikely(!buff))
2326 return -ENOMEM;
2327 skb_copy_decrypted(to: buff, from: skb);
2328 mptcp_skb_ext_copy(to: buff, from: skb);
2329
2330 sk_wmem_queued_add(sk, val: buff->truesize);
2331 sk_mem_charge(sk, size: buff->truesize);
2332 buff->truesize += nlen;
2333 skb->truesize -= nlen;
2334
2335 /* Correct the sequence numbers. */
2336 TCP_SKB_CB(buff)->seq = TCP_SKB_CB(skb)->seq + len;
2337 TCP_SKB_CB(buff)->end_seq = TCP_SKB_CB(skb)->end_seq;
2338 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(buff)->seq;
2339
2340 /* PSH and FIN should only be set in the second packet. */
2341 flags = TCP_SKB_CB(skb)->tcp_flags;
2342 TCP_SKB_CB(skb)->tcp_flags = flags & ~(TCPHDR_FIN | TCPHDR_PSH);
2343 TCP_SKB_CB(buff)->tcp_flags = flags;
2344
2345 tcp_skb_fragment_eor(skb, skb2: buff);
2346
2347 skb_split(skb, skb1: buff, len);
2348 tcp_fragment_tstamp(skb, skb2: buff);
2349
2350 /* Fix up tso_factor for both original and new SKB. */
2351 tcp_set_skb_tso_segs(skb, mss_now);
2352 tcp_set_skb_tso_segs(skb: buff, mss_now);
2353
2354 /* Link BUFF into the send queue. */
2355 __skb_header_release(skb: buff);
2356 tcp_insert_write_queue_after(skb, buff, sk, tcp_queue: TCP_FRAG_IN_WRITE_QUEUE);
2357
2358 return 0;
2359}
2360
2361/* Try to defer sending, if possible, in order to minimize the amount
2362 * of TSO splitting we do. View it as a kind of TSO Nagle test.
2363 *
2364 * This algorithm is from John Heffner.
2365 */
2366static bool tcp_tso_should_defer(struct sock *sk, struct sk_buff *skb,
2367 bool *is_cwnd_limited,
2368 bool *is_rwnd_limited,
2369 u32 max_segs)
2370{
2371 const struct inet_connection_sock *icsk = inet_csk(sk);
2372 u32 send_win, cong_win, limit, in_flight;
2373 struct tcp_sock *tp = tcp_sk(sk);
2374 struct sk_buff *head;
2375 int win_divisor;
2376 s64 delta;
2377
2378 if (icsk->icsk_ca_state >= TCP_CA_Recovery)
2379 goto send_now;
2380
2381 /* Avoid bursty behavior by allowing defer
2382 * only if the last write was recent (1 ms).
2383 * Note that tp->tcp_wstamp_ns can be in the future if we have
2384 * packets waiting in a qdisc or device for EDT delivery.
2385 */
2386 delta = tp->tcp_clock_cache - tp->tcp_wstamp_ns - NSEC_PER_MSEC;
2387 if (delta > 0)
2388 goto send_now;
2389
2390 in_flight = tcp_packets_in_flight(tp);
2391
2392 BUG_ON(tcp_skb_pcount(skb) <= 1);
2393 BUG_ON(tcp_snd_cwnd(tp) <= in_flight);
2394
2395 send_win = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
2396
2397 /* From in_flight test above, we know that cwnd > in_flight. */
2398 cong_win = (tcp_snd_cwnd(tp) - in_flight) * tp->mss_cache;
2399
2400 limit = min(send_win, cong_win);
2401
2402 /* If a full-sized TSO skb can be sent, do it. */
2403 if (limit >= max_segs * tp->mss_cache)
2404 goto send_now;
2405
2406 /* Middle in queue won't get any more data, full sendable already? */
2407 if ((skb != tcp_write_queue_tail(sk)) && (limit >= skb->len))
2408 goto send_now;
2409
2410 win_divisor = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_tso_win_divisor);
2411 if (win_divisor) {
2412 u32 chunk = min(tp->snd_wnd, tcp_snd_cwnd(tp) * tp->mss_cache);
2413
2414 /* If at least some fraction of a window is available,
2415 * just use it.
2416 */
2417 chunk /= win_divisor;
2418 if (limit >= chunk)
2419 goto send_now;
2420 } else {
2421 /* Different approach, try not to defer past a single
2422 * ACK. Receiver should ACK every other full sized
2423 * frame, so if we have space for more than 3 frames
2424 * then send now.
2425 */
2426 if (limit > tcp_max_tso_deferred_mss(tp) * tp->mss_cache)
2427 goto send_now;
2428 }
2429
2430 /* TODO : use tsorted_sent_queue ? */
2431 head = tcp_rtx_queue_head(sk);
2432 if (!head)
2433 goto send_now;
2434 delta = tp->tcp_clock_cache - head->tstamp;
2435 /* If next ACK is likely to come too late (half srtt), do not defer */
2436 if ((s64)(delta - (u64)NSEC_PER_USEC * (tp->srtt_us >> 4)) < 0)
2437 goto send_now;
2438
2439 /* Ok, it looks like it is advisable to defer.
2440 * Three cases are tracked :
2441 * 1) We are cwnd-limited
2442 * 2) We are rwnd-limited
2443 * 3) We are application limited.
2444 */
2445 if (cong_win < send_win) {
2446 if (cong_win <= skb->len) {
2447 *is_cwnd_limited = true;
2448 return true;
2449 }
2450 } else {
2451 if (send_win <= skb->len) {
2452 *is_rwnd_limited = true;
2453 return true;
2454 }
2455 }
2456
2457 /* If this packet won't get more data, do not wait. */
2458 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_FIN) ||
2459 TCP_SKB_CB(skb)->eor)
2460 goto send_now;
2461
2462 return true;
2463
2464send_now:
2465 return false;
2466}
2467
2468static inline void tcp_mtu_check_reprobe(struct sock *sk)
2469{
2470 struct inet_connection_sock *icsk = inet_csk(sk);
2471 struct tcp_sock *tp = tcp_sk(sk);
2472 struct net *net = sock_net(sk);
2473 u32 interval;
2474 s32 delta;
2475
2476 interval = READ_ONCE(net->ipv4.sysctl_tcp_probe_interval);
2477 delta = tcp_jiffies32 - icsk->icsk_mtup.probe_timestamp;
2478 if (unlikely(delta >= interval * HZ)) {
2479 int mss = tcp_current_mss(sk);
2480
2481 /* Update current search range */
2482 icsk->icsk_mtup.probe_size = 0;
2483 icsk->icsk_mtup.search_high = tp->rx_opt.mss_clamp +
2484 sizeof(struct tcphdr) +
2485 icsk->icsk_af_ops->net_header_len;
2486 icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
2487
2488 /* Update probe time stamp */
2489 icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
2490 }
2491}
2492
2493static bool tcp_can_coalesce_send_queue_head(struct sock *sk, int len)
2494{
2495 struct sk_buff *skb, *next;
2496
2497 skb = tcp_send_head(sk);
2498 tcp_for_write_queue_from_safe(skb, next, sk) {
2499 if (len <= skb->len)
2500 break;
2501
2502 if (tcp_has_tx_tstamp(skb) || !tcp_skb_can_collapse(to: skb, from: next))
2503 return false;
2504
2505 len -= skb->len;
2506 }
2507
2508 return true;
2509}
2510
2511static int tcp_clone_payload(struct sock *sk, struct sk_buff *to,
2512 int probe_size)
2513{
2514 skb_frag_t *lastfrag = NULL, *fragto = skb_shinfo(to)->frags;
2515 int i, todo, len = 0, nr_frags = 0;
2516 const struct sk_buff *skb;
2517
2518 if (!sk_wmem_schedule(sk, size: to->truesize + probe_size))
2519 return -ENOMEM;
2520
2521 skb_queue_walk(&sk->sk_write_queue, skb) {
2522 const skb_frag_t *fragfrom = skb_shinfo(skb)->frags;
2523
2524 if (skb_headlen(skb))
2525 return -EINVAL;
2526
2527 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++, fragfrom++) {
2528 if (len >= probe_size)
2529 goto commit;
2530 todo = min_t(int, skb_frag_size(fragfrom),
2531 probe_size - len);
2532 len += todo;
2533 if (lastfrag &&
2534 skb_frag_page(frag: fragfrom) == skb_frag_page(frag: lastfrag) &&
2535 skb_frag_off(frag: fragfrom) == skb_frag_off(frag: lastfrag) +
2536 skb_frag_size(frag: lastfrag)) {
2537 skb_frag_size_add(frag: lastfrag, delta: todo);
2538 continue;
2539 }
2540 if (unlikely(nr_frags == MAX_SKB_FRAGS))
2541 return -E2BIG;
2542 skb_frag_page_copy(fragto, fragfrom);
2543 skb_frag_off_copy(fragto, fragfrom);
2544 skb_frag_size_set(frag: fragto, size: todo);
2545 nr_frags++;
2546 lastfrag = fragto++;
2547 }
2548 }
2549commit:
2550 WARN_ON_ONCE(len != probe_size);
2551 for (i = 0; i < nr_frags; i++)
2552 skb_frag_ref(skb: to, f: i);
2553
2554 skb_shinfo(to)->nr_frags = nr_frags;
2555 to->truesize += probe_size;
2556 to->len += probe_size;
2557 to->data_len += probe_size;
2558 __skb_header_release(skb: to);
2559 return 0;
2560}
2561
2562/* tcp_mtu_probe() and tcp_grow_skb() can both eat an skb (src) if
2563 * all its payload was moved to another one (dst).
2564 * Make sure to transfer tcp_flags, eor, and tstamp.
2565 */
2566static void tcp_eat_one_skb(struct sock *sk,
2567 struct sk_buff *dst,
2568 struct sk_buff *src)
2569{
2570 TCP_SKB_CB(dst)->tcp_flags |= TCP_SKB_CB(src)->tcp_flags;
2571 TCP_SKB_CB(dst)->eor = TCP_SKB_CB(src)->eor;
2572 tcp_skb_collapse_tstamp(skb: dst, next_skb: src);
2573 tcp_unlink_write_queue(skb: src, sk);
2574 tcp_wmem_free_skb(sk, skb: src);
2575}
2576
2577/* Create a new MTU probe if we are ready.
2578 * MTU probe is regularly attempting to increase the path MTU by
2579 * deliberately sending larger packets. This discovers routing
2580 * changes resulting in larger path MTUs.
2581 *
2582 * Returns 0 if we should wait to probe (no cwnd available),
2583 * 1 if a probe was sent,
2584 * -1 otherwise
2585 */
2586static int tcp_mtu_probe(struct sock *sk)
2587{
2588 struct inet_connection_sock *icsk = inet_csk(sk);
2589 struct tcp_sock *tp = tcp_sk(sk);
2590 struct sk_buff *skb, *nskb, *next;
2591 struct net *net = sock_net(sk);
2592 int probe_size;
2593 int size_needed;
2594 int copy, len;
2595 int mss_now;
2596 int interval;
2597
2598 /* Not currently probing/verifying,
2599 * not in recovery,
2600 * have enough cwnd, and
2601 * not SACKing (the variable headers throw things off)
2602 */
2603 if (likely(!icsk->icsk_mtup.enabled ||
2604 icsk->icsk_mtup.probe_size ||
2605 inet_csk(sk)->icsk_ca_state != TCP_CA_Open ||
2606 tcp_snd_cwnd(tp) < 11 ||
2607 tp->rx_opt.num_sacks || tp->rx_opt.dsack))
2608 return -1;
2609
2610 /* Use binary search for probe_size between tcp_mss_base,
2611 * and current mss_clamp. if (search_high - search_low)
2612 * smaller than a threshold, backoff from probing.
2613 */
2614 mss_now = tcp_current_mss(sk);
2615 probe_size = tcp_mtu_to_mss(sk, pmtu: (icsk->icsk_mtup.search_high +
2616 icsk->icsk_mtup.search_low) >> 1);
2617 size_needed = probe_size + (tp->reordering + 1) * tp->mss_cache;
2618 interval = icsk->icsk_mtup.search_high - icsk->icsk_mtup.search_low;
2619 /* When misfortune happens, we are reprobing actively,
2620 * and then reprobe timer has expired. We stick with current
2621 * probing process by not resetting search range to its orignal.
2622 */
2623 if (probe_size > tcp_mtu_to_mss(sk, pmtu: icsk->icsk_mtup.search_high) ||
2624 interval < READ_ONCE(net->ipv4.sysctl_tcp_probe_threshold)) {
2625 /* Check whether enough time has elaplased for
2626 * another round of probing.
2627 */
2628 tcp_mtu_check_reprobe(sk);
2629 return -1;
2630 }
2631
2632 /* Have enough data in the send queue to probe? */
2633 if (tp->write_seq - tp->snd_nxt < size_needed)
2634 return -1;
2635
2636 if (tp->snd_wnd < size_needed)
2637 return -1;
2638 if (after(tp->snd_nxt + size_needed, tcp_wnd_end(tp)))
2639 return 0;
2640
2641 /* Do we need to wait to drain cwnd? With none in flight, don't stall */
2642 if (tcp_packets_in_flight(tp) + 2 > tcp_snd_cwnd(tp)) {
2643 if (!tcp_packets_in_flight(tp))
2644 return -1;
2645 else
2646 return 0;
2647 }
2648
2649 if (!tcp_can_coalesce_send_queue_head(sk, len: probe_size))
2650 return -1;
2651
2652 /* We're allowed to probe. Build it now. */
2653 nskb = tcp_stream_alloc_skb(sk, GFP_ATOMIC, force_schedule: false);
2654 if (!nskb)
2655 return -1;
2656
2657 /* build the payload, and be prepared to abort if this fails. */
2658 if (tcp_clone_payload(sk, to: nskb, probe_size)) {
2659 tcp_skb_tsorted_anchor_cleanup(skb: nskb);
2660 consume_skb(skb: nskb);
2661 return -1;
2662 }
2663 sk_wmem_queued_add(sk, val: nskb->truesize);
2664 sk_mem_charge(sk, size: nskb->truesize);
2665
2666 skb = tcp_send_head(sk);
2667 skb_copy_decrypted(to: nskb, from: skb);
2668 mptcp_skb_ext_copy(to: nskb, from: skb);
2669
2670 TCP_SKB_CB(nskb)->seq = TCP_SKB_CB(skb)->seq;
2671 TCP_SKB_CB(nskb)->end_seq = TCP_SKB_CB(skb)->seq + probe_size;
2672 TCP_SKB_CB(nskb)->tcp_flags = TCPHDR_ACK;
2673
2674 tcp_insert_write_queue_before(new: nskb, skb, sk);
2675 tcp_highest_sack_replace(sk, old: skb, new: nskb);
2676
2677 len = 0;
2678 tcp_for_write_queue_from_safe(skb, next, sk) {
2679 copy = min_t(int, skb->len, probe_size - len);
2680
2681 if (skb->len <= copy) {
2682 tcp_eat_one_skb(sk, dst: nskb, src: skb);
2683 } else {
2684 TCP_SKB_CB(nskb)->tcp_flags |= TCP_SKB_CB(skb)->tcp_flags &
2685 ~(TCPHDR_FIN|TCPHDR_PSH);
2686 __pskb_trim_head(skb, len: copy);
2687 tcp_set_skb_tso_segs(skb, mss_now);
2688 TCP_SKB_CB(skb)->seq += copy;
2689 }
2690
2691 len += copy;
2692
2693 if (len >= probe_size)
2694 break;
2695 }
2696 tcp_init_tso_segs(skb: nskb, mss_now: nskb->len);
2697
2698 /* We're ready to send. If this fails, the probe will
2699 * be resegmented into mss-sized pieces by tcp_write_xmit().
2700 */
2701 if (!tcp_transmit_skb(sk, skb: nskb, clone_it: 1, GFP_ATOMIC)) {
2702 /* Decrement cwnd here because we are sending
2703 * effectively two packets. */
2704 tcp_snd_cwnd_set(tp, val: tcp_snd_cwnd(tp) - 1);
2705 tcp_event_new_data_sent(sk, skb: nskb);
2706
2707 icsk->icsk_mtup.probe_size = tcp_mss_to_mtu(sk, nskb->len);
2708 tp->mtu_probe.probe_seq_start = TCP_SKB_CB(nskb)->seq;
2709 tp->mtu_probe.probe_seq_end = TCP_SKB_CB(nskb)->end_seq;
2710
2711 return 1;
2712 }
2713
2714 return -1;
2715}
2716
2717static bool tcp_pacing_check(struct sock *sk)
2718{
2719 struct tcp_sock *tp = tcp_sk(sk);
2720
2721 if (!tcp_needs_internal_pacing(sk))
2722 return false;
2723
2724 if (tp->tcp_wstamp_ns <= tp->tcp_clock_cache)
2725 return false;
2726
2727 if (!hrtimer_is_queued(timer: &tp->pacing_timer)) {
2728 hrtimer_start(timer: &tp->pacing_timer,
2729 tim: ns_to_ktime(ns: tp->tcp_wstamp_ns),
2730 mode: HRTIMER_MODE_ABS_PINNED_SOFT);
2731 sock_hold(sk);
2732 }
2733 return true;
2734}
2735
2736static bool tcp_rtx_queue_empty_or_single_skb(const struct sock *sk)
2737{
2738 const struct rb_node *node = sk->tcp_rtx_queue.rb_node;
2739
2740 /* No skb in the rtx queue. */
2741 if (!node)
2742 return true;
2743
2744 /* Only one skb in rtx queue. */
2745 return !node->rb_left && !node->rb_right;
2746}
2747
2748/* TCP Small Queues :
2749 * Control number of packets in qdisc/devices to two packets / or ~1 ms.
2750 * (These limits are doubled for retransmits)
2751 * This allows for :
2752 * - better RTT estimation and ACK scheduling
2753 * - faster recovery
2754 * - high rates
2755 * Alas, some drivers / subsystems require a fair amount
2756 * of queued bytes to ensure line rate.
2757 * One example is wifi aggregation (802.11 AMPDU)
2758 */
2759static bool tcp_small_queue_check(struct sock *sk, const struct sk_buff *skb,
2760 unsigned int factor)
2761{
2762 unsigned long limit;
2763
2764 limit = max_t(unsigned long,
2765 2 * skb->truesize,
2766 READ_ONCE(sk->sk_pacing_rate) >> READ_ONCE(sk->sk_pacing_shift));
2767 limit = min_t(unsigned long, limit,
2768 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_limit_output_bytes));
2769 limit <<= factor;
2770
2771 if (static_branch_unlikely(&tcp_tx_delay_enabled) &&
2772 tcp_sk(sk)->tcp_tx_delay) {
2773 u64 extra_bytes = (u64)READ_ONCE(sk->sk_pacing_rate) *
2774 tcp_sk(sk)->tcp_tx_delay;
2775
2776 /* TSQ is based on skb truesize sum (sk_wmem_alloc), so we
2777 * approximate our needs assuming an ~100% skb->truesize overhead.
2778 * USEC_PER_SEC is approximated by 2^20.
2779 * do_div(extra_bytes, USEC_PER_SEC/2) is replaced by a right shift.
2780 */
2781 extra_bytes >>= (20 - 1);
2782 limit += extra_bytes;
2783 }
2784 if (refcount_read(r: &sk->sk_wmem_alloc) > limit) {
2785 /* Always send skb if rtx queue is empty or has one skb.
2786 * No need to wait for TX completion to call us back,
2787 * after softirq schedule.
2788 * This helps when TX completions are delayed too much.
2789 */
2790 if (tcp_rtx_queue_empty_or_single_skb(sk))
2791 return false;
2792
2793 set_bit(nr: TSQ_THROTTLED, addr: &sk->sk_tsq_flags);
2794 /* It is possible TX completion already happened
2795 * before we set TSQ_THROTTLED, so we must
2796 * test again the condition.
2797 */
2798 smp_mb__after_atomic();
2799 if (refcount_read(r: &sk->sk_wmem_alloc) > limit)
2800 return true;
2801 }
2802 return false;
2803}
2804
2805static void tcp_chrono_set(struct tcp_sock *tp, const enum tcp_chrono new)
2806{
2807 const u32 now = tcp_jiffies32;
2808 enum tcp_chrono old = tp->chrono_type;
2809
2810 if (old > TCP_CHRONO_UNSPEC)
2811 tp->chrono_stat[old - 1] += now - tp->chrono_start;
2812 tp->chrono_start = now;
2813 tp->chrono_type = new;
2814}
2815
2816void tcp_chrono_start(struct sock *sk, const enum tcp_chrono type)
2817{
2818 struct tcp_sock *tp = tcp_sk(sk);
2819
2820 /* If there are multiple conditions worthy of tracking in a
2821 * chronograph then the highest priority enum takes precedence
2822 * over the other conditions. So that if something "more interesting"
2823 * starts happening, stop the previous chrono and start a new one.
2824 */
2825 if (type > tp->chrono_type)
2826 tcp_chrono_set(tp, new: type);
2827}
2828
2829void tcp_chrono_stop(struct sock *sk, const enum tcp_chrono type)
2830{
2831 struct tcp_sock *tp = tcp_sk(sk);
2832
2833
2834 /* There are multiple conditions worthy of tracking in a
2835 * chronograph, so that the highest priority enum takes
2836 * precedence over the other conditions (see tcp_chrono_start).
2837 * If a condition stops, we only stop chrono tracking if
2838 * it's the "most interesting" or current chrono we are
2839 * tracking and starts busy chrono if we have pending data.
2840 */
2841 if (tcp_rtx_and_write_queues_empty(sk))
2842 tcp_chrono_set(tp, new: TCP_CHRONO_UNSPEC);
2843 else if (type == tp->chrono_type)
2844 tcp_chrono_set(tp, new: TCP_CHRONO_BUSY);
2845}
2846
2847/* First skb in the write queue is smaller than ideal packet size.
2848 * Check if we can move payload from the second skb in the queue.
2849 */
2850static void tcp_grow_skb(struct sock *sk, struct sk_buff *skb, int amount)
2851{
2852 struct sk_buff *next_skb = skb->next;
2853 unsigned int nlen;
2854
2855 if (tcp_skb_is_last(sk, skb))
2856 return;
2857
2858 if (!tcp_skb_can_collapse(to: skb, from: next_skb))
2859 return;
2860
2861 nlen = min_t(u32, amount, next_skb->len);
2862 if (!nlen || !skb_shift(tgt: skb, skb: next_skb, shiftlen: nlen))
2863 return;
2864
2865 TCP_SKB_CB(skb)->end_seq += nlen;
2866 TCP_SKB_CB(next_skb)->seq += nlen;
2867
2868 if (!next_skb->len) {
2869 /* In case FIN is set, we need to update end_seq */
2870 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
2871
2872 tcp_eat_one_skb(sk, dst: skb, src: next_skb);
2873 }
2874}
2875
2876/* This routine writes packets to the network. It advances the
2877 * send_head. This happens as incoming acks open up the remote
2878 * window for us.
2879 *
2880 * LARGESEND note: !tcp_urg_mode is overkill, only frames between
2881 * snd_up-64k-mss .. snd_up cannot be large. However, taking into
2882 * account rare use of URG, this is not a big flaw.
2883 *
2884 * Send at most one packet when push_one > 0. Temporarily ignore
2885 * cwnd limit to force at most one packet out when push_one == 2.
2886
2887 * Returns true, if no segments are in flight and we have queued segments,
2888 * but cannot send anything now because of SWS or another problem.
2889 */
2890static bool tcp_write_xmit(struct sock *sk, unsigned int mss_now, int nonagle,
2891 int push_one, gfp_t gfp)
2892{
2893 struct tcp_sock *tp = tcp_sk(sk);
2894 struct sk_buff *skb;
2895 unsigned int tso_segs, sent_pkts;
2896 u32 cwnd_quota, max_segs;
2897 int result;
2898 bool is_cwnd_limited = false, is_rwnd_limited = false;
2899
2900 sent_pkts = 0;
2901
2902 tcp_mstamp_refresh(tp);
2903
2904 /* AccECN option beacon depends on mstamp, it may change mss */
2905 if (tcp_ecn_mode_accecn(tp) && tcp_accecn_option_beacon_check(sk))
2906 mss_now = tcp_current_mss(sk);
2907
2908 if (!push_one) {
2909 /* Do MTU probing. */
2910 result = tcp_mtu_probe(sk);
2911 if (!result) {
2912 return false;
2913 } else if (result > 0) {
2914 sent_pkts = 1;
2915 }
2916 }
2917
2918 max_segs = tcp_tso_segs(sk, mss_now);
2919 while ((skb = tcp_send_head(sk))) {
2920 unsigned int limit;
2921 int missing_bytes;
2922
2923 if (unlikely(tp->repair) && tp->repair_queue == TCP_SEND_QUEUE) {
2924 /* "skb_mstamp_ns" is used as a start point for the retransmit timer */
2925 tp->tcp_wstamp_ns = tp->tcp_clock_cache;
2926 skb_set_delivery_time(skb, kt: tp->tcp_wstamp_ns, tstamp_type: SKB_CLOCK_MONOTONIC);
2927 list_move_tail(list: &skb->tcp_tsorted_anchor, head: &tp->tsorted_sent_queue);
2928 tcp_init_tso_segs(skb, mss_now);
2929 goto repair; /* Skip network transmission */
2930 }
2931
2932 if (tcp_pacing_check(sk))
2933 break;
2934
2935 cwnd_quota = tcp_cwnd_test(tp);
2936 if (!cwnd_quota) {
2937 if (push_one == 2)
2938 /* Force out a loss probe pkt. */
2939 cwnd_quota = 1;
2940 else
2941 break;
2942 }
2943 cwnd_quota = min(cwnd_quota, max_segs);
2944 missing_bytes = cwnd_quota * mss_now - skb->len;
2945 if (missing_bytes > 0)
2946 tcp_grow_skb(sk, skb, amount: missing_bytes);
2947
2948 tso_segs = tcp_set_skb_tso_segs(skb, mss_now);
2949
2950 if (unlikely(!tcp_snd_wnd_test(tp, skb, mss_now))) {
2951 is_rwnd_limited = true;
2952 break;
2953 }
2954
2955 if (tso_segs == 1) {
2956 if (unlikely(!tcp_nagle_test(tp, skb, mss_now,
2957 (tcp_skb_is_last(sk, skb) ?
2958 nonagle : TCP_NAGLE_PUSH))))
2959 break;
2960 } else {
2961 if (!push_one &&
2962 tcp_tso_should_defer(sk, skb, is_cwnd_limited: &is_cwnd_limited,
2963 is_rwnd_limited: &is_rwnd_limited, max_segs))
2964 break;
2965 }
2966
2967 limit = mss_now;
2968 if (tso_segs > 1 && !tcp_urg_mode(tp))
2969 limit = tcp_mss_split_point(sk, skb, mss_now,
2970 max_segs: cwnd_quota,
2971 nonagle);
2972
2973 if (skb->len > limit &&
2974 unlikely(tso_fragment(sk, skb, limit, mss_now, gfp)))
2975 break;
2976
2977 if (tcp_small_queue_check(sk, skb, factor: 0))
2978 break;
2979
2980 /* Argh, we hit an empty skb(), presumably a thread
2981 * is sleeping in sendmsg()/sk_stream_wait_memory().
2982 * We do not want to send a pure-ack packet and have
2983 * a strange looking rtx queue with empty packet(s).
2984 */
2985 if (TCP_SKB_CB(skb)->end_seq == TCP_SKB_CB(skb)->seq)
2986 break;
2987
2988 if (unlikely(tcp_transmit_skb(sk, skb, 1, gfp)))
2989 break;
2990
2991repair:
2992 /* Advance the send_head. This one is sent out.
2993 * This call will increment packets_out.
2994 */
2995 tcp_event_new_data_sent(sk, skb);
2996
2997 tcp_minshall_update(tp, mss_now, skb);
2998 sent_pkts += tcp_skb_pcount(skb);
2999
3000 if (push_one)
3001 break;
3002 }
3003
3004 if (is_rwnd_limited)
3005 tcp_chrono_start(sk, type: TCP_CHRONO_RWND_LIMITED);
3006 else
3007 tcp_chrono_stop(sk, type: TCP_CHRONO_RWND_LIMITED);
3008
3009 is_cwnd_limited |= (tcp_packets_in_flight(tp) >= tcp_snd_cwnd(tp));
3010 if (likely(sent_pkts || is_cwnd_limited))
3011 tcp_cwnd_validate(sk, is_cwnd_limited);
3012
3013 if (likely(sent_pkts)) {
3014 if (tcp_in_cwnd_reduction(sk))
3015 tp->prr_out += sent_pkts;
3016
3017 /* Send one loss probe per tail loss episode. */
3018 if (push_one != 2)
3019 tcp_schedule_loss_probe(sk, advancing_rto: false);
3020 return false;
3021 }
3022 return !tp->packets_out && !tcp_write_queue_empty(sk);
3023}
3024
3025bool tcp_schedule_loss_probe(struct sock *sk, bool advancing_rto)
3026{
3027 struct inet_connection_sock *icsk = inet_csk(sk);
3028 struct tcp_sock *tp = tcp_sk(sk);
3029 u32 timeout, timeout_us, rto_delta_us;
3030 int early_retrans;
3031
3032 /* Don't do any loss probe on a Fast Open connection before 3WHS
3033 * finishes.
3034 */
3035 if (rcu_access_pointer(tp->fastopen_rsk))
3036 return false;
3037
3038 early_retrans = READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_early_retrans);
3039 /* Schedule a loss probe in 2*RTT for SACK capable connections
3040 * not in loss recovery, that are either limited by cwnd or application.
3041 */
3042 if ((early_retrans != 3 && early_retrans != 4) ||
3043 !tp->packets_out || !tcp_is_sack(tp) ||
3044 (icsk->icsk_ca_state != TCP_CA_Open &&
3045 icsk->icsk_ca_state != TCP_CA_CWR))
3046 return false;
3047
3048 /* Probe timeout is 2*rtt. Add minimum RTO to account
3049 * for delayed ack when there's one outstanding packet. If no RTT
3050 * sample is available then probe after TCP_TIMEOUT_INIT.
3051 */
3052 if (tp->srtt_us) {
3053 timeout_us = tp->srtt_us >> 2;
3054 if (tp->packets_out == 1)
3055 timeout_us += tcp_rto_min_us(sk);
3056 else
3057 timeout_us += TCP_TIMEOUT_MIN_US;
3058 timeout = usecs_to_jiffies(u: timeout_us);
3059 } else {
3060 timeout = TCP_TIMEOUT_INIT;
3061 }
3062
3063 /* If the RTO formula yields an earlier time, then use that time. */
3064 rto_delta_us = advancing_rto ?
3065 jiffies_to_usecs(inet_csk(sk)->icsk_rto) :
3066 tcp_rto_delta_us(sk); /* How far in future is RTO? */
3067 if (rto_delta_us > 0)
3068 timeout = min_t(u32, timeout, usecs_to_jiffies(rto_delta_us));
3069
3070 tcp_reset_xmit_timer(sk, ICSK_TIME_LOSS_PROBE, when: timeout, pace_delay: true);
3071 return true;
3072}
3073
3074/* Thanks to skb fast clones, we can detect if a prior transmit of
3075 * a packet is still in a qdisc or driver queue.
3076 * In this case, there is very little point doing a retransmit !
3077 */
3078static bool skb_still_in_host_queue(struct sock *sk,
3079 const struct sk_buff *skb)
3080{
3081 if (unlikely(skb_fclone_busy(sk, skb))) {
3082 set_bit(nr: TSQ_THROTTLED, addr: &sk->sk_tsq_flags);
3083 smp_mb__after_atomic();
3084 if (skb_fclone_busy(sk, skb)) {
3085 NET_INC_STATS(sock_net(sk),
3086 LINUX_MIB_TCPSPURIOUS_RTX_HOSTQUEUES);
3087 return true;
3088 }
3089 }
3090 return false;
3091}
3092
3093/* When probe timeout (PTO) fires, try send a new segment if possible, else
3094 * retransmit the last segment.
3095 */
3096void tcp_send_loss_probe(struct sock *sk)
3097{
3098 struct tcp_sock *tp = tcp_sk(sk);
3099 struct sk_buff *skb;
3100 int pcount;
3101 int mss = tcp_current_mss(sk);
3102
3103 /* At most one outstanding TLP */
3104 if (tp->tlp_high_seq)
3105 goto rearm_timer;
3106
3107 tp->tlp_retrans = 0;
3108 skb = tcp_send_head(sk);
3109 if (skb && tcp_snd_wnd_test(tp, skb, cur_mss: mss)) {
3110 pcount = tp->packets_out;
3111 tcp_write_xmit(sk, mss_now: mss, TCP_NAGLE_OFF, push_one: 2, GFP_ATOMIC);
3112 if (tp->packets_out > pcount)
3113 goto probe_sent;
3114 goto rearm_timer;
3115 }
3116 skb = skb_rb_last(&sk->tcp_rtx_queue);
3117 if (unlikely(!skb)) {
3118 tcp_warn_once(sk, cond: tp->packets_out, str: "invalid inflight: ");
3119 smp_store_release(&inet_csk(sk)->icsk_pending, 0);
3120 return;
3121 }
3122
3123 if (skb_still_in_host_queue(sk, skb))
3124 goto rearm_timer;
3125
3126 pcount = tcp_skb_pcount(skb);
3127 if (WARN_ON(!pcount))
3128 goto rearm_timer;
3129
3130 if ((pcount > 1) && (skb->len > (pcount - 1) * mss)) {
3131 if (unlikely(tcp_fragment(sk, TCP_FRAG_IN_RTX_QUEUE, skb,
3132 (pcount - 1) * mss, mss,
3133 GFP_ATOMIC)))
3134 goto rearm_timer;
3135 skb = skb_rb_next(skb);
3136 }
3137
3138 if (WARN_ON(!skb || !tcp_skb_pcount(skb)))
3139 goto rearm_timer;
3140
3141 if (__tcp_retransmit_skb(sk, skb, segs: 1))
3142 goto rearm_timer;
3143
3144 tp->tlp_retrans = 1;
3145
3146probe_sent:
3147 /* Record snd_nxt for loss detection. */
3148 tp->tlp_high_seq = tp->snd_nxt;
3149
3150 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPLOSSPROBES);
3151 /* Reset s.t. tcp_rearm_rto will restart timer from now */
3152 smp_store_release(&inet_csk(sk)->icsk_pending, 0);
3153rearm_timer:
3154 tcp_rearm_rto(sk);
3155}
3156
3157/* Push out any pending frames which were held back due to
3158 * TCP_CORK or attempt at coalescing tiny packets.
3159 * The socket must be locked by the caller.
3160 */
3161void __tcp_push_pending_frames(struct sock *sk, unsigned int cur_mss,
3162 int nonagle)
3163{
3164 /* If we are closed, the bytes will have to remain here.
3165 * In time closedown will finish, we empty the write queue and
3166 * all will be happy.
3167 */
3168 if (unlikely(sk->sk_state == TCP_CLOSE))
3169 return;
3170
3171 if (tcp_write_xmit(sk, mss_now: cur_mss, nonagle, push_one: 0,
3172 gfp: sk_gfp_mask(sk, GFP_ATOMIC)))
3173 tcp_check_probe_timer(sk);
3174}
3175
3176/* Send _single_ skb sitting at the send head. This function requires
3177 * true push pending frames to setup probe timer etc.
3178 */
3179void tcp_push_one(struct sock *sk, unsigned int mss_now)
3180{
3181 struct sk_buff *skb = tcp_send_head(sk);
3182
3183 BUG_ON(!skb || skb->len < mss_now);
3184
3185 tcp_write_xmit(sk, mss_now, TCP_NAGLE_PUSH, push_one: 1, gfp: sk->sk_allocation);
3186}
3187
3188/* This function returns the amount that we can raise the
3189 * usable window based on the following constraints
3190 *
3191 * 1. The window can never be shrunk once it is offered (RFC 793)
3192 * 2. We limit memory per socket
3193 *
3194 * RFC 1122:
3195 * "the suggested [SWS] avoidance algorithm for the receiver is to keep
3196 * RECV.NEXT + RCV.WIN fixed until:
3197 * RCV.BUFF - RCV.USER - RCV.WINDOW >= min(1/2 RCV.BUFF, MSS)"
3198 *
3199 * i.e. don't raise the right edge of the window until you can raise
3200 * it at least MSS bytes.
3201 *
3202 * Unfortunately, the recommended algorithm breaks header prediction,
3203 * since header prediction assumes th->window stays fixed.
3204 *
3205 * Strictly speaking, keeping th->window fixed violates the receiver
3206 * side SWS prevention criteria. The problem is that under this rule
3207 * a stream of single byte packets will cause the right side of the
3208 * window to always advance by a single byte.
3209 *
3210 * Of course, if the sender implements sender side SWS prevention
3211 * then this will not be a problem.
3212 *
3213 * BSD seems to make the following compromise:
3214 *
3215 * If the free space is less than the 1/4 of the maximum
3216 * space available and the free space is less than 1/2 mss,
3217 * then set the window to 0.
3218 * [ Actually, bsd uses MSS and 1/4 of maximal _window_ ]
3219 * Otherwise, just prevent the window from shrinking
3220 * and from being larger than the largest representable value.
3221 *
3222 * This prevents incremental opening of the window in the regime
3223 * where TCP is limited by the speed of the reader side taking
3224 * data out of the TCP receive queue. It does nothing about
3225 * those cases where the window is constrained on the sender side
3226 * because the pipeline is full.
3227 *
3228 * BSD also seems to "accidentally" limit itself to windows that are a
3229 * multiple of MSS, at least until the free space gets quite small.
3230 * This would appear to be a side effect of the mbuf implementation.
3231 * Combining these two algorithms results in the observed behavior
3232 * of having a fixed window size at almost all times.
3233 *
3234 * Below we obtain similar behavior by forcing the offered window to
3235 * a multiple of the mss when it is feasible to do so.
3236 *
3237 * Note, we don't "adjust" for TIMESTAMP or SACK option bytes.
3238 * Regular options like TIMESTAMP are taken into account.
3239 */
3240u32 __tcp_select_window(struct sock *sk)
3241{
3242 struct inet_connection_sock *icsk = inet_csk(sk);
3243 struct tcp_sock *tp = tcp_sk(sk);
3244 struct net *net = sock_net(sk);
3245 /* MSS for the peer's data. Previous versions used mss_clamp
3246 * here. I don't know if the value based on our guesses
3247 * of peer's MSS is better for the performance. It's more correct
3248 * but may be worse for the performance because of rcv_mss
3249 * fluctuations. --SAW 1998/11/1
3250 */
3251 int mss = icsk->icsk_ack.rcv_mss;
3252 int free_space = tcp_space(sk);
3253 int allowed_space = tcp_full_space(sk);
3254 int full_space, window;
3255
3256 if (sk_is_mptcp(sk))
3257 mptcp_space(ssk: sk, s: &free_space, fs: &allowed_space);
3258
3259 full_space = min_t(int, tp->window_clamp, allowed_space);
3260
3261 if (unlikely(mss > full_space)) {
3262 mss = full_space;
3263 if (mss <= 0)
3264 return 0;
3265 }
3266
3267 /* Only allow window shrink if the sysctl is enabled and we have
3268 * a non-zero scaling factor in effect.
3269 */
3270 if (READ_ONCE(net->ipv4.sysctl_tcp_shrink_window) && tp->rx_opt.rcv_wscale)
3271 goto shrink_window_allowed;
3272
3273 /* do not allow window to shrink */
3274
3275 if (free_space < (full_space >> 1)) {
3276 icsk->icsk_ack.quick = 0;
3277
3278 if (tcp_under_memory_pressure(sk))
3279 tcp_adjust_rcv_ssthresh(sk);
3280
3281 /* free_space might become our new window, make sure we don't
3282 * increase it due to wscale.
3283 */
3284 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
3285
3286 /* if free space is less than mss estimate, or is below 1/16th
3287 * of the maximum allowed, try to move to zero-window, else
3288 * tcp_clamp_window() will grow rcv buf up to tcp_rmem[2], and
3289 * new incoming data is dropped due to memory limits.
3290 * With large window, mss test triggers way too late in order
3291 * to announce zero window in time before rmem limit kicks in.
3292 */
3293 if (free_space < (allowed_space >> 4) || free_space < mss)
3294 return 0;
3295 }
3296
3297 if (free_space > tp->rcv_ssthresh)
3298 free_space = tp->rcv_ssthresh;
3299
3300 /* Don't do rounding if we are using window scaling, since the
3301 * scaled window will not line up with the MSS boundary anyway.
3302 */
3303 if (tp->rx_opt.rcv_wscale) {
3304 window = free_space;
3305
3306 /* Advertise enough space so that it won't get scaled away.
3307 * Import case: prevent zero window announcement if
3308 * 1<<rcv_wscale > mss.
3309 */
3310 window = ALIGN(window, (1 << tp->rx_opt.rcv_wscale));
3311 } else {
3312 window = tp->rcv_wnd;
3313 /* Get the largest window that is a nice multiple of mss.
3314 * Window clamp already applied above.
3315 * If our current window offering is within 1 mss of the
3316 * free space we just keep it. This prevents the divide
3317 * and multiply from happening most of the time.
3318 * We also don't do any window rounding when the free space
3319 * is too small.
3320 */
3321 if (window <= free_space - mss || window > free_space)
3322 window = rounddown(free_space, mss);
3323 else if (mss == full_space &&
3324 free_space > window + (full_space >> 1))
3325 window = free_space;
3326 }
3327
3328 return window;
3329
3330shrink_window_allowed:
3331 /* new window should always be an exact multiple of scaling factor */
3332 free_space = round_down(free_space, 1 << tp->rx_opt.rcv_wscale);
3333
3334 if (free_space < (full_space >> 1)) {
3335 icsk->icsk_ack.quick = 0;
3336
3337 if (tcp_under_memory_pressure(sk))
3338 tcp_adjust_rcv_ssthresh(sk);
3339
3340 /* if free space is too low, return a zero window */
3341 if (free_space < (allowed_space >> 4) || free_space < mss ||
3342 free_space < (1 << tp->rx_opt.rcv_wscale))
3343 return 0;
3344 }
3345
3346 if (free_space > tp->rcv_ssthresh) {
3347 free_space = tp->rcv_ssthresh;
3348 /* new window should always be an exact multiple of scaling factor
3349 *
3350 * For this case, we ALIGN "up" (increase free_space) because
3351 * we know free_space is not zero here, it has been reduced from
3352 * the memory-based limit, and rcv_ssthresh is not a hard limit
3353 * (unlike sk_rcvbuf).
3354 */
3355 free_space = ALIGN(free_space, (1 << tp->rx_opt.rcv_wscale));
3356 }
3357
3358 return free_space;
3359}
3360
3361void tcp_skb_collapse_tstamp(struct sk_buff *skb,
3362 const struct sk_buff *next_skb)
3363{
3364 if (unlikely(tcp_has_tx_tstamp(next_skb))) {
3365 const struct skb_shared_info *next_shinfo =
3366 skb_shinfo(next_skb);
3367 struct skb_shared_info *shinfo = skb_shinfo(skb);
3368
3369 shinfo->tx_flags |= next_shinfo->tx_flags & SKBTX_ANY_TSTAMP;
3370 shinfo->tskey = next_shinfo->tskey;
3371 TCP_SKB_CB(skb)->txstamp_ack |=
3372 TCP_SKB_CB(next_skb)->txstamp_ack;
3373 }
3374}
3375
3376/* Collapses two adjacent SKB's during retransmission. */
3377static bool tcp_collapse_retrans(struct sock *sk, struct sk_buff *skb)
3378{
3379 struct tcp_sock *tp = tcp_sk(sk);
3380 struct sk_buff *next_skb = skb_rb_next(skb);
3381 int next_skb_size;
3382
3383 next_skb_size = next_skb->len;
3384
3385 BUG_ON(tcp_skb_pcount(skb) != 1 || tcp_skb_pcount(next_skb) != 1);
3386
3387 if (next_skb_size && !tcp_skb_shift(to: skb, from: next_skb, pcount: 1, shiftlen: next_skb_size))
3388 return false;
3389
3390 tcp_highest_sack_replace(sk, old: next_skb, new: skb);
3391
3392 /* Update sequence range on original skb. */
3393 TCP_SKB_CB(skb)->end_seq = TCP_SKB_CB(next_skb)->end_seq;
3394
3395 /* Merge over control information. This moves PSH/FIN etc. over */
3396 TCP_SKB_CB(skb)->tcp_flags |= TCP_SKB_CB(next_skb)->tcp_flags;
3397
3398 /* All done, get rid of second SKB and account for it so
3399 * packet counting does not break.
3400 */
3401 TCP_SKB_CB(skb)->sacked |= TCP_SKB_CB(next_skb)->sacked & TCPCB_EVER_RETRANS;
3402 TCP_SKB_CB(skb)->eor = TCP_SKB_CB(next_skb)->eor;
3403
3404 /* changed transmit queue under us so clear hints */
3405 if (next_skb == tp->retransmit_skb_hint)
3406 tp->retransmit_skb_hint = skb;
3407
3408 tcp_adjust_pcount(sk, skb: next_skb, decr: tcp_skb_pcount(skb: next_skb));
3409
3410 tcp_skb_collapse_tstamp(skb, next_skb);
3411
3412 tcp_rtx_queue_unlink_and_free(skb: next_skb, sk);
3413 return true;
3414}
3415
3416/* Check if coalescing SKBs is legal. */
3417static bool tcp_can_collapse(const struct sock *sk, const struct sk_buff *skb)
3418{
3419 if (tcp_skb_pcount(skb) > 1)
3420 return false;
3421 if (skb_cloned(skb))
3422 return false;
3423 if (!skb_frags_readable(skb))
3424 return false;
3425 /* Some heuristics for collapsing over SACK'd could be invented */
3426 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_ACKED)
3427 return false;
3428
3429 return true;
3430}
3431
3432/* Collapse packets in the retransmit queue to make to create
3433 * less packets on the wire. This is only done on retransmission.
3434 */
3435static void tcp_retrans_try_collapse(struct sock *sk, struct sk_buff *to,
3436 int space)
3437{
3438 struct tcp_sock *tp = tcp_sk(sk);
3439 struct sk_buff *skb = to, *tmp;
3440 bool first = true;
3441
3442 if (!READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_retrans_collapse))
3443 return;
3444 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3445 return;
3446
3447 skb_rbtree_walk_from_safe(skb, tmp) {
3448 if (!tcp_can_collapse(sk, skb))
3449 break;
3450
3451 if (!tcp_skb_can_collapse(to, from: skb))
3452 break;
3453
3454 space -= skb->len;
3455
3456 if (first) {
3457 first = false;
3458 continue;
3459 }
3460
3461 if (space < 0)
3462 break;
3463
3464 if (after(TCP_SKB_CB(skb)->end_seq, tcp_wnd_end(tp)))
3465 break;
3466
3467 if (!tcp_collapse_retrans(sk, skb: to))
3468 break;
3469 }
3470}
3471
3472/* This retransmits one SKB. Policy decisions and retransmit queue
3473 * state updates are done by the caller. Returns non-zero if an
3474 * error occurred which prevented the send.
3475 */
3476int __tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3477{
3478 struct inet_connection_sock *icsk = inet_csk(sk);
3479 struct tcp_sock *tp = tcp_sk(sk);
3480 unsigned int cur_mss;
3481 int diff, len, err;
3482 int avail_wnd;
3483
3484 /* Inconclusive MTU probe */
3485 if (icsk->icsk_mtup.probe_size)
3486 icsk->icsk_mtup.probe_size = 0;
3487
3488 if (skb_still_in_host_queue(sk, skb)) {
3489 err = -EBUSY;
3490 goto out;
3491 }
3492
3493start:
3494 if (before(TCP_SKB_CB(skb)->seq, seq2: tp->snd_una)) {
3495 if (unlikely(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3496 TCP_SKB_CB(skb)->tcp_flags &= ~TCPHDR_SYN;
3497 TCP_SKB_CB(skb)->seq++;
3498 goto start;
3499 }
3500 if (unlikely(before(TCP_SKB_CB(skb)->end_seq, tp->snd_una))) {
3501 WARN_ON_ONCE(1);
3502 err = -EINVAL;
3503 goto out;
3504 }
3505 if (tcp_trim_head(sk, skb, len: tp->snd_una - TCP_SKB_CB(skb)->seq)) {
3506 err = -ENOMEM;
3507 goto out;
3508 }
3509 }
3510
3511 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) {
3512 err = -EHOSTUNREACH; /* Routing failure or similar. */
3513 goto out;
3514 }
3515
3516 cur_mss = tcp_current_mss(sk);
3517 avail_wnd = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
3518
3519 /* If receiver has shrunk his window, and skb is out of
3520 * new window, do not retransmit it. The exception is the
3521 * case, when window is shrunk to zero. In this case
3522 * our retransmit of one segment serves as a zero window probe.
3523 */
3524 if (avail_wnd <= 0) {
3525 if (TCP_SKB_CB(skb)->seq != tp->snd_una) {
3526 err = -EAGAIN;
3527 goto out;
3528 }
3529 avail_wnd = cur_mss;
3530 }
3531
3532 len = cur_mss * segs;
3533 if (len > avail_wnd) {
3534 len = rounddown(avail_wnd, cur_mss);
3535 if (!len)
3536 len = avail_wnd;
3537 }
3538 if (skb->len > len) {
3539 if (tcp_fragment(sk, tcp_queue: TCP_FRAG_IN_RTX_QUEUE, skb, len,
3540 mss_now: cur_mss, GFP_ATOMIC)) {
3541 err = -ENOMEM; /* We'll try again later. */
3542 goto out;
3543 }
3544 } else {
3545 if (skb_unclone_keeptruesize(skb, GFP_ATOMIC)) {
3546 err = -ENOMEM;
3547 goto out;
3548 }
3549
3550 diff = tcp_skb_pcount(skb);
3551 tcp_set_skb_tso_segs(skb, mss_now: cur_mss);
3552 diff -= tcp_skb_pcount(skb);
3553 if (diff)
3554 tcp_adjust_pcount(sk, skb, decr: diff);
3555 avail_wnd = min_t(int, avail_wnd, cur_mss);
3556 if (skb->len < avail_wnd)
3557 tcp_retrans_try_collapse(sk, to: skb, space: avail_wnd);
3558 }
3559
3560 /* RFC3168, section 6.1.1.1. ECN fallback
3561 * As AccECN uses the same SYN flags (+ AE), this check covers both
3562 * cases.
3563 */
3564 if ((TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN_ECN) == TCPHDR_SYN_ECN)
3565 tcp_ecn_clear_syn(sk, skb);
3566
3567 /* Update global and local TCP statistics. */
3568 segs = tcp_skb_pcount(skb);
3569 TCP_ADD_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS, segs);
3570 if (TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)
3571 __NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
3572 tp->total_retrans += segs;
3573 tp->bytes_retrans += skb->len;
3574
3575 /* make sure skb->data is aligned on arches that require it
3576 * and check if ack-trimming & collapsing extended the headroom
3577 * beyond what csum_start can cover.
3578 */
3579 if (unlikely((NET_IP_ALIGN && ((unsigned long)skb->data & 3)) ||
3580 skb_headroom(skb) >= 0xFFFF)) {
3581 struct sk_buff *nskb;
3582
3583 tcp_skb_tsorted_save(skb) {
3584 nskb = __pskb_copy(skb, MAX_TCP_HEADER, GFP_ATOMIC);
3585 if (nskb) {
3586 nskb->dev = NULL;
3587 err = tcp_transmit_skb(sk, skb: nskb, clone_it: 0, GFP_ATOMIC);
3588 } else {
3589 err = -ENOBUFS;
3590 }
3591 } tcp_skb_tsorted_restore(skb);
3592
3593 if (!err) {
3594 tcp_update_skb_after_send(sk, skb, prior_wstamp: tp->tcp_wstamp_ns);
3595 tcp_rate_skb_sent(sk, skb);
3596 }
3597 } else {
3598 err = tcp_transmit_skb(sk, skb, clone_it: 1, GFP_ATOMIC);
3599 }
3600
3601 if (BPF_SOCK_OPS_TEST_FLAG(tp, BPF_SOCK_OPS_RETRANS_CB_FLAG))
3602 tcp_call_bpf_3arg(sk, op: BPF_SOCK_OPS_RETRANS_CB,
3603 TCP_SKB_CB(skb)->seq, arg2: segs, arg3: err);
3604
3605 if (unlikely(err) && err != -EBUSY)
3606 NET_ADD_STATS(sock_net(sk), LINUX_MIB_TCPRETRANSFAIL, segs);
3607
3608 /* To avoid taking spuriously low RTT samples based on a timestamp
3609 * for a transmit that never happened, always mark EVER_RETRANS
3610 */
3611 TCP_SKB_CB(skb)->sacked |= TCPCB_EVER_RETRANS;
3612
3613out:
3614 trace_tcp_retransmit_skb(sk, skb, err);
3615 return err;
3616}
3617
3618int tcp_retransmit_skb(struct sock *sk, struct sk_buff *skb, int segs)
3619{
3620 struct tcp_sock *tp = tcp_sk(sk);
3621 int err = __tcp_retransmit_skb(sk, skb, segs);
3622
3623 if (err == 0) {
3624#if FASTRETRANS_DEBUG > 0
3625 if (TCP_SKB_CB(skb)->sacked & TCPCB_SACKED_RETRANS) {
3626 net_dbg_ratelimited("retrans_out leaked\n");
3627 }
3628#endif
3629 TCP_SKB_CB(skb)->sacked |= TCPCB_RETRANS;
3630 tp->retrans_out += tcp_skb_pcount(skb);
3631 }
3632
3633 /* Save stamp of the first (attempted) retransmit. */
3634 if (!tp->retrans_stamp)
3635 tp->retrans_stamp = tcp_skb_timestamp_ts(usec_ts: tp->tcp_usec_ts, skb);
3636
3637 if (tp->undo_retrans < 0)
3638 tp->undo_retrans = 0;
3639 tp->undo_retrans += tcp_skb_pcount(skb);
3640 return err;
3641}
3642
3643/* This gets called after a retransmit timeout, and the initially
3644 * retransmitted data is acknowledged. It tries to continue
3645 * resending the rest of the retransmit queue, until either
3646 * we've sent it all or the congestion window limit is reached.
3647 */
3648void tcp_xmit_retransmit_queue(struct sock *sk)
3649{
3650 const struct inet_connection_sock *icsk = inet_csk(sk);
3651 struct sk_buff *skb, *rtx_head, *hole = NULL;
3652 struct tcp_sock *tp = tcp_sk(sk);
3653 bool rearm_timer = false;
3654 u32 max_segs;
3655 int mib_idx;
3656
3657 if (!tp->packets_out)
3658 return;
3659
3660 rtx_head = tcp_rtx_queue_head(sk);
3661 skb = tp->retransmit_skb_hint ?: rtx_head;
3662 max_segs = tcp_tso_segs(sk, mss_now: tcp_current_mss(sk));
3663 skb_rbtree_walk_from(skb) {
3664 __u8 sacked;
3665 int segs;
3666
3667 if (tcp_pacing_check(sk))
3668 break;
3669
3670 /* we could do better than to assign each time */
3671 if (!hole)
3672 tp->retransmit_skb_hint = skb;
3673
3674 segs = tcp_snd_cwnd(tp) - tcp_packets_in_flight(tp);
3675 if (segs <= 0)
3676 break;
3677 sacked = TCP_SKB_CB(skb)->sacked;
3678 /* In case tcp_shift_skb_data() have aggregated large skbs,
3679 * we need to make sure not sending too bigs TSO packets
3680 */
3681 segs = min_t(int, segs, max_segs);
3682
3683 if (tp->retrans_out >= tp->lost_out) {
3684 break;
3685 } else if (!(sacked & TCPCB_LOST)) {
3686 if (!hole && !(sacked & (TCPCB_SACKED_RETRANS|TCPCB_SACKED_ACKED)))
3687 hole = skb;
3688 continue;
3689
3690 } else {
3691 if (icsk->icsk_ca_state != TCP_CA_Loss)
3692 mib_idx = LINUX_MIB_TCPFASTRETRANS;
3693 else
3694 mib_idx = LINUX_MIB_TCPSLOWSTARTRETRANS;
3695 }
3696
3697 if (sacked & (TCPCB_SACKED_ACKED|TCPCB_SACKED_RETRANS))
3698 continue;
3699
3700 if (tcp_small_queue_check(sk, skb, factor: 1))
3701 break;
3702
3703 if (tcp_retransmit_skb(sk, skb, segs))
3704 break;
3705
3706 NET_ADD_STATS(sock_net(sk), mib_idx, tcp_skb_pcount(skb));
3707
3708 if (tcp_in_cwnd_reduction(sk))
3709 tp->prr_out += tcp_skb_pcount(skb);
3710
3711 if (skb == rtx_head &&
3712 icsk->icsk_pending != ICSK_TIME_REO_TIMEOUT)
3713 rearm_timer = true;
3714
3715 }
3716 if (rearm_timer)
3717 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
3718 inet_csk(sk)->icsk_rto, pace_delay: true);
3719}
3720
3721/* We allow to exceed memory limits for FIN packets to expedite
3722 * connection tear down and (memory) recovery.
3723 * Otherwise tcp_send_fin() could be tempted to either delay FIN
3724 * or even be forced to close flow without any FIN.
3725 * In general, we want to allow one skb per socket to avoid hangs
3726 * with edge trigger epoll()
3727 */
3728void sk_forced_mem_schedule(struct sock *sk, int size)
3729{
3730 int delta, amt;
3731
3732 delta = size - sk->sk_forward_alloc;
3733 if (delta <= 0)
3734 return;
3735 amt = sk_mem_pages(amt: delta);
3736 sk_forward_alloc_add(sk, val: amt << PAGE_SHIFT);
3737 sk_memory_allocated_add(sk, val: amt);
3738
3739 if (mem_cgroup_sk_enabled(sk))
3740 mem_cgroup_sk_charge(sk, nr_pages: amt, gfp_mask: gfp_memcg_charge() | __GFP_NOFAIL);
3741}
3742
3743/* Send a FIN. The caller locks the socket for us.
3744 * We should try to send a FIN packet really hard, but eventually give up.
3745 */
3746void tcp_send_fin(struct sock *sk)
3747{
3748 struct sk_buff *skb, *tskb, *tail = tcp_write_queue_tail(sk);
3749 struct tcp_sock *tp = tcp_sk(sk);
3750
3751 /* Optimization, tack on the FIN if we have one skb in write queue and
3752 * this skb was not yet sent, or we are under memory pressure.
3753 * Note: in the latter case, FIN packet will be sent after a timeout,
3754 * as TCP stack thinks it has already been transmitted.
3755 */
3756 tskb = tail;
3757 if (!tskb && tcp_under_memory_pressure(sk))
3758 tskb = skb_rb_last(&sk->tcp_rtx_queue);
3759
3760 if (tskb) {
3761 TCP_SKB_CB(tskb)->tcp_flags |= TCPHDR_FIN;
3762 TCP_SKB_CB(tskb)->end_seq++;
3763 tp->write_seq++;
3764 if (!tail) {
3765 /* This means tskb was already sent.
3766 * Pretend we included the FIN on previous transmit.
3767 * We need to set tp->snd_nxt to the value it would have
3768 * if FIN had been sent. This is because retransmit path
3769 * does not change tp->snd_nxt.
3770 */
3771 WRITE_ONCE(tp->snd_nxt, tp->snd_nxt + 1);
3772 return;
3773 }
3774 } else {
3775 skb = alloc_skb_fclone(MAX_TCP_HEADER,
3776 priority: sk_gfp_mask(sk, GFP_ATOMIC |
3777 __GFP_NOWARN));
3778 if (unlikely(!skb))
3779 return;
3780
3781 INIT_LIST_HEAD(list: &skb->tcp_tsorted_anchor);
3782 skb_reserve(skb, MAX_TCP_HEADER);
3783 sk_forced_mem_schedule(sk, size: skb->truesize);
3784 /* FIN eats a sequence byte, write_seq advanced by tcp_queue_skb(). */
3785 tcp_init_nondata_skb(skb, sk, seq: tp->write_seq,
3786 TCPHDR_ACK | TCPHDR_FIN);
3787 tcp_queue_skb(sk, skb);
3788 }
3789 __tcp_push_pending_frames(sk, cur_mss: tcp_current_mss(sk), TCP_NAGLE_OFF);
3790}
3791
3792/* We get here when a process closes a file descriptor (either due to
3793 * an explicit close() or as a byproduct of exit()'ing) and there
3794 * was unread data in the receive queue. This behavior is recommended
3795 * by RFC 2525, section 2.17. -DaveM
3796 */
3797void tcp_send_active_reset(struct sock *sk, gfp_t priority,
3798 enum sk_rst_reason reason)
3799{
3800 struct sk_buff *skb;
3801
3802 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTRSTS);
3803
3804 /* NOTE: No TCP options attached and we never retransmit this. */
3805 skb = alloc_skb(MAX_TCP_HEADER, priority);
3806 if (!skb) {
3807 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3808 return;
3809 }
3810
3811 /* Reserve space for headers and prepare control bits. */
3812 skb_reserve(skb, MAX_TCP_HEADER);
3813 tcp_init_nondata_skb(skb, sk, seq: tcp_acceptable_seq(sk),
3814 TCPHDR_ACK | TCPHDR_RST);
3815 tcp_mstamp_refresh(tcp_sk(sk));
3816 /* Send it off. */
3817 if (tcp_transmit_skb(sk, skb, clone_it: 0, gfp_mask: priority))
3818 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTFAILED);
3819
3820 /* skb of trace_tcp_send_reset() keeps the skb that caused RST,
3821 * skb here is different to the troublesome skb, so use NULL
3822 */
3823 trace_tcp_send_reset(sk, NULL, reason);
3824}
3825
3826/* Send a crossed SYN-ACK during socket establishment.
3827 * WARNING: This routine must only be called when we have already sent
3828 * a SYN packet that crossed the incoming SYN that caused this routine
3829 * to get called. If this assumption fails then the initial rcv_wnd
3830 * and rcv_wscale values will not be correct.
3831 */
3832int tcp_send_synack(struct sock *sk)
3833{
3834 struct sk_buff *skb;
3835
3836 skb = tcp_rtx_queue_head(sk);
3837 if (!skb || !(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_SYN)) {
3838 pr_err("%s: wrong queue state\n", __func__);
3839 return -EFAULT;
3840 }
3841 if (!(TCP_SKB_CB(skb)->tcp_flags & TCPHDR_ACK)) {
3842 if (skb_cloned(skb)) {
3843 struct sk_buff *nskb;
3844
3845 tcp_skb_tsorted_save(skb) {
3846 nskb = skb_copy(skb, GFP_ATOMIC);
3847 } tcp_skb_tsorted_restore(skb);
3848 if (!nskb)
3849 return -ENOMEM;
3850 INIT_LIST_HEAD(list: &nskb->tcp_tsorted_anchor);
3851 tcp_highest_sack_replace(sk, old: skb, new: nskb);
3852 tcp_rtx_queue_unlink_and_free(skb, sk);
3853 __skb_header_release(skb: nskb);
3854 tcp_rbtree_insert(root: &sk->tcp_rtx_queue, skb: nskb);
3855 sk_wmem_queued_add(sk, val: nskb->truesize);
3856 sk_mem_charge(sk, size: nskb->truesize);
3857 skb = nskb;
3858 }
3859
3860 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_ACK;
3861 tcp_ecn_send_synack(sk, skb);
3862 }
3863 return tcp_transmit_skb(sk, skb, clone_it: 1, GFP_ATOMIC);
3864}
3865
3866/**
3867 * tcp_make_synack - Allocate one skb and build a SYNACK packet.
3868 * @sk: listener socket
3869 * @dst: dst entry attached to the SYNACK. It is consumed and caller
3870 * should not use it again.
3871 * @req: request_sock pointer
3872 * @foc: cookie for tcp fast open
3873 * @synack_type: Type of synack to prepare
3874 * @syn_skb: SYN packet just received. It could be NULL for rtx case.
3875 */
3876struct sk_buff *tcp_make_synack(const struct sock *sk, struct dst_entry *dst,
3877 struct request_sock *req,
3878 struct tcp_fastopen_cookie *foc,
3879 enum tcp_synack_type synack_type,
3880 struct sk_buff *syn_skb)
3881{
3882 struct inet_request_sock *ireq = inet_rsk(sk: req);
3883 const struct tcp_sock *tp = tcp_sk(sk);
3884 struct tcp_out_options opts;
3885 struct tcp_key key = {};
3886 struct sk_buff *skb;
3887 int tcp_header_size;
3888 struct tcphdr *th;
3889 int mss;
3890 u64 now;
3891
3892 skb = alloc_skb(MAX_TCP_HEADER, GFP_ATOMIC);
3893 if (unlikely(!skb)) {
3894 dst_release(dst);
3895 return NULL;
3896 }
3897 /* Reserve space for headers. */
3898 skb_reserve(skb, MAX_TCP_HEADER);
3899
3900 switch (synack_type) {
3901 case TCP_SYNACK_NORMAL:
3902 skb_set_owner_edemux(skb, sk: req_to_sk(req));
3903 break;
3904 case TCP_SYNACK_COOKIE:
3905 /* Under synflood, we do not attach skb to a socket,
3906 * to avoid false sharing.
3907 */
3908 break;
3909 case TCP_SYNACK_FASTOPEN:
3910 /* sk is a const pointer, because we want to express multiple
3911 * cpu might call us concurrently.
3912 * sk->sk_wmem_alloc in an atomic, we can promote to rw.
3913 */
3914 skb_set_owner_w(skb, sk: (struct sock *)sk);
3915 break;
3916 }
3917 skb_dst_set(skb, dst);
3918
3919 mss = tcp_mss_clamp(tp, mss: dst_metric_advmss(dst));
3920
3921 memset(s: &opts, c: 0, n: sizeof(opts));
3922 now = tcp_clock_ns();
3923#ifdef CONFIG_SYN_COOKIES
3924 if (unlikely(synack_type == TCP_SYNACK_COOKIE && ireq->tstamp_ok))
3925 skb_set_delivery_time(skb, kt: cookie_init_timestamp(req, now),
3926 tstamp_type: SKB_CLOCK_MONOTONIC);
3927 else
3928#endif
3929 {
3930 skb_set_delivery_time(skb, kt: now, tstamp_type: SKB_CLOCK_MONOTONIC);
3931 if (!tcp_rsk(req)->snt_synack) /* Timestamp first SYNACK */
3932 tcp_rsk(req)->snt_synack = tcp_skb_timestamp_us(skb);
3933 }
3934
3935#if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
3936 rcu_read_lock();
3937#endif
3938 if (tcp_rsk_used_ao(req)) {
3939#ifdef CONFIG_TCP_AO
3940 struct tcp_ao_key *ao_key = NULL;
3941 u8 keyid = tcp_rsk(req)->ao_keyid;
3942 u8 rnext = tcp_rsk(req)->ao_rcv_next;
3943
3944 ao_key = tcp_sk(sk)->af_specific->ao_lookup(sk, req_to_sk(req),
3945 keyid, -1);
3946 /* If there is no matching key - avoid sending anything,
3947 * especially usigned segments. It could try harder and lookup
3948 * for another peer-matching key, but the peer has requested
3949 * ao_keyid (RFC5925 RNextKeyID), so let's keep it simple here.
3950 */
3951 if (unlikely(!ao_key)) {
3952 trace_tcp_ao_synack_no_key(sk, keyid, rnext);
3953 rcu_read_unlock();
3954 kfree_skb(skb);
3955 net_warn_ratelimited("TCP-AO: the keyid %u from SYN packet is not present - not sending SYNACK\n",
3956 keyid);
3957 return NULL;
3958 }
3959 key.ao_key = ao_key;
3960 key.type = TCP_KEY_AO;
3961#endif
3962 } else {
3963#ifdef CONFIG_TCP_MD5SIG
3964 key.md5_key = tcp_rsk(req)->af_specific->req_md5_lookup(sk,
3965 req_to_sk(req));
3966 if (key.md5_key)
3967 key.type = TCP_KEY_MD5;
3968#endif
3969 }
3970 skb_set_hash(skb, READ_ONCE(tcp_rsk(req)->txhash), type: PKT_HASH_TYPE_L4);
3971 /* bpf program will be interested in the tcp_flags */
3972 TCP_SKB_CB(skb)->tcp_flags = TCPHDR_SYN | TCPHDR_ACK;
3973 tcp_header_size = tcp_synack_options(sk, req, mss, skb, opts: &opts,
3974 key: &key, foc, synack_type, syn_skb)
3975 + sizeof(*th);
3976
3977 skb_push(skb, len: tcp_header_size);
3978 skb_reset_transport_header(skb);
3979
3980 th = (struct tcphdr *)skb->data;
3981 memset(s: th, c: 0, n: sizeof(struct tcphdr));
3982 th->syn = 1;
3983 th->ack = 1;
3984 tcp_ecn_make_synack(req, th);
3985 th->source = htons(ireq->ir_num);
3986 th->dest = ireq->ir_rmt_port;
3987 skb->mark = ireq->ir_mark;
3988 skb->ip_summed = CHECKSUM_PARTIAL;
3989 th->seq = htonl(tcp_rsk(req)->snt_isn);
3990 /* XXX data is queued and acked as is. No buffer/window check */
3991 th->ack_seq = htonl(tcp_rsk(req)->rcv_nxt);
3992
3993 /* RFC1323: The window in SYN & SYN/ACK segments is never scaled. */
3994 th->window = htons(min(req->rsk_rcv_wnd, 65535U));
3995 tcp_options_write(th, NULL, tcprsk: tcp_rsk(req), opts: &opts, key: &key);
3996 th->doff = (tcp_header_size >> 2);
3997 TCP_INC_STATS(sock_net(sk), TCP_MIB_OUTSEGS);
3998
3999 /* Okay, we have all we need - do the md5 hash if needed */
4000 if (tcp_key_is_md5(key: &key)) {
4001#ifdef CONFIG_TCP_MD5SIG
4002 tcp_rsk(req)->af_specific->calc_md5_hash(opts.hash_location,
4003 key.md5_key, req_to_sk(req), skb);
4004#endif
4005 } else if (tcp_key_is_ao(key: &key)) {
4006#ifdef CONFIG_TCP_AO
4007 tcp_rsk(req)->af_specific->ao_synack_hash(opts.hash_location,
4008 key.ao_key, req, skb,
4009 opts.hash_location - (u8 *)th, 0);
4010#endif
4011 }
4012#if defined(CONFIG_TCP_MD5SIG) || defined(CONFIG_TCP_AO)
4013 rcu_read_unlock();
4014#endif
4015
4016 bpf_skops_write_hdr_opt(sk: (struct sock *)sk, skb, req, syn_skb,
4017 synack_type, opts: &opts);
4018
4019 skb_set_delivery_time(skb, kt: now, tstamp_type: SKB_CLOCK_MONOTONIC);
4020 tcp_add_tx_delay(skb, tp);
4021
4022 return skb;
4023}
4024EXPORT_IPV6_MOD(tcp_make_synack);
4025
4026static void tcp_ca_dst_init(struct sock *sk, const struct dst_entry *dst)
4027{
4028 struct inet_connection_sock *icsk = inet_csk(sk);
4029 const struct tcp_congestion_ops *ca;
4030 u32 ca_key = dst_metric(dst, RTAX_CC_ALGO);
4031
4032 if (ca_key == TCP_CA_UNSPEC)
4033 return;
4034
4035 rcu_read_lock();
4036 ca = tcp_ca_find_key(key: ca_key);
4037 if (likely(ca && bpf_try_module_get(ca, ca->owner))) {
4038 bpf_module_put(data: icsk->icsk_ca_ops, owner: icsk->icsk_ca_ops->owner);
4039 icsk->icsk_ca_dst_locked = tcp_ca_dst_locked(dst);
4040 icsk->icsk_ca_ops = ca;
4041 }
4042 rcu_read_unlock();
4043}
4044
4045/* Do all connect socket setups that can be done AF independent. */
4046static void tcp_connect_init(struct sock *sk)
4047{
4048 const struct dst_entry *dst = __sk_dst_get(sk);
4049 struct tcp_sock *tp = tcp_sk(sk);
4050 __u8 rcv_wscale;
4051 u16 user_mss;
4052 u32 rcv_wnd;
4053
4054 /* We'll fix this up when we get a response from the other end.
4055 * See tcp_input.c:tcp_rcv_state_process case TCP_SYN_SENT.
4056 */
4057 tp->tcp_header_len = sizeof(struct tcphdr);
4058 if (READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_timestamps))
4059 tp->tcp_header_len += TCPOLEN_TSTAMP_ALIGNED;
4060
4061 tcp_ao_connect_init(sk);
4062
4063 /* If user gave his TCP_MAXSEG, record it to clamp */
4064 user_mss = READ_ONCE(tp->rx_opt.user_mss);
4065 if (user_mss)
4066 tp->rx_opt.mss_clamp = user_mss;
4067 tp->max_window = 0;
4068 tcp_mtup_init(sk);
4069 tcp_sync_mss(sk, pmtu: dst_mtu(dst));
4070
4071 tcp_ca_dst_init(sk, dst);
4072
4073 if (!tp->window_clamp)
4074 WRITE_ONCE(tp->window_clamp, dst_metric(dst, RTAX_WINDOW));
4075 tp->advmss = tcp_mss_clamp(tp, mss: dst_metric_advmss(dst));
4076
4077 tcp_initialize_rcv_mss(sk);
4078
4079 /* limit the window selection if the user enforce a smaller rx buffer */
4080 if (sk->sk_userlocks & SOCK_RCVBUF_LOCK &&
4081 (tp->window_clamp > tcp_full_space(sk) || tp->window_clamp == 0))
4082 WRITE_ONCE(tp->window_clamp, tcp_full_space(sk));
4083
4084 rcv_wnd = tcp_rwnd_init_bpf(sk);
4085 if (rcv_wnd == 0)
4086 rcv_wnd = dst_metric(dst, RTAX_INITRWND);
4087
4088 tcp_select_initial_window(sk, space: tcp_full_space(sk),
4089 mss: tp->advmss - (tp->rx_opt.ts_recent_stamp ? tp->tcp_header_len - sizeof(struct tcphdr) : 0),
4090 rcv_wnd: &tp->rcv_wnd,
4091 window_clamp: &tp->window_clamp,
4092 READ_ONCE(sock_net(sk)->ipv4.sysctl_tcp_window_scaling),
4093 rcv_wscale: &rcv_wscale,
4094 init_rcv_wnd: rcv_wnd);
4095
4096 tp->rx_opt.rcv_wscale = rcv_wscale;
4097 tp->rcv_ssthresh = tp->rcv_wnd;
4098
4099 WRITE_ONCE(sk->sk_err, 0);
4100 sock_reset_flag(sk, flag: SOCK_DONE);
4101 tp->snd_wnd = 0;
4102 tcp_init_wl(tp, seq: 0);
4103 tcp_write_queue_purge(sk);
4104 tp->snd_una = tp->write_seq;
4105 tp->snd_sml = tp->write_seq;
4106 tp->snd_up = tp->write_seq;
4107 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
4108
4109 if (likely(!tp->repair))
4110 tp->rcv_nxt = 0;
4111 else
4112 tp->rcv_tstamp = tcp_jiffies32;
4113 tp->rcv_wup = tp->rcv_nxt;
4114 WRITE_ONCE(tp->copied_seq, tp->rcv_nxt);
4115
4116 inet_csk(sk)->icsk_rto = tcp_timeout_init(sk);
4117 WRITE_ONCE(inet_csk(sk)->icsk_retransmits, 0);
4118 tcp_clear_retrans(tp);
4119}
4120
4121static void tcp_connect_queue_skb(struct sock *sk, struct sk_buff *skb)
4122{
4123 struct tcp_sock *tp = tcp_sk(sk);
4124 struct tcp_skb_cb *tcb = TCP_SKB_CB(skb);
4125
4126 tcb->end_seq += skb->len;
4127 __skb_header_release(skb);
4128 sk_wmem_queued_add(sk, val: skb->truesize);
4129 sk_mem_charge(sk, size: skb->truesize);
4130 WRITE_ONCE(tp->write_seq, tcb->end_seq);
4131 tp->packets_out += tcp_skb_pcount(skb);
4132}
4133
4134/* Build and send a SYN with data and (cached) Fast Open cookie. However,
4135 * queue a data-only packet after the regular SYN, such that regular SYNs
4136 * are retransmitted on timeouts. Also if the remote SYN-ACK acknowledges
4137 * only the SYN sequence, the data are retransmitted in the first ACK.
4138 * If cookie is not cached or other error occurs, falls back to send a
4139 * regular SYN with Fast Open cookie request option.
4140 */
4141static int tcp_send_syn_data(struct sock *sk, struct sk_buff *syn)
4142{
4143 struct inet_connection_sock *icsk = inet_csk(sk);
4144 struct tcp_sock *tp = tcp_sk(sk);
4145 struct tcp_fastopen_request *fo = tp->fastopen_req;
4146 struct page_frag *pfrag = sk_page_frag(sk);
4147 struct sk_buff *syn_data;
4148 int space, err = 0;
4149
4150 tp->rx_opt.mss_clamp = tp->advmss; /* If MSS is not cached */
4151 if (!tcp_fastopen_cookie_check(sk, mss: &tp->rx_opt.mss_clamp, cookie: &fo->cookie))
4152 goto fallback;
4153
4154 /* MSS for SYN-data is based on cached MSS and bounded by PMTU and
4155 * user-MSS. Reserve maximum option space for middleboxes that add
4156 * private TCP options. The cost is reduced data space in SYN :(
4157 */
4158 tp->rx_opt.mss_clamp = tcp_mss_clamp(tp, mss: tp->rx_opt.mss_clamp);
4159 /* Sync mss_cache after updating the mss_clamp */
4160 tcp_sync_mss(sk, pmtu: icsk->icsk_pmtu_cookie);
4161
4162 space = __tcp_mtu_to_mss(sk, pmtu: icsk->icsk_pmtu_cookie) -
4163 MAX_TCP_OPTION_SPACE;
4164
4165 space = min_t(size_t, space, fo->size);
4166
4167 if (space &&
4168 !skb_page_frag_refill(min_t(size_t, space, PAGE_SIZE),
4169 pfrag, prio: sk->sk_allocation))
4170 goto fallback;
4171 syn_data = tcp_stream_alloc_skb(sk, gfp: sk->sk_allocation, force_schedule: false);
4172 if (!syn_data)
4173 goto fallback;
4174 memcpy(to: syn_data->cb, from: syn->cb, len: sizeof(syn->cb));
4175 if (space) {
4176 space = min_t(size_t, space, pfrag->size - pfrag->offset);
4177 space = tcp_wmem_schedule(sk, copy: space);
4178 }
4179 if (space) {
4180 space = copy_page_from_iter(page: pfrag->page, offset: pfrag->offset,
4181 bytes: space, i: &fo->data->msg_iter);
4182 if (unlikely(!space)) {
4183 tcp_skb_tsorted_anchor_cleanup(skb: syn_data);
4184 kfree_skb(skb: syn_data);
4185 goto fallback;
4186 }
4187 skb_fill_page_desc(skb: syn_data, i: 0, page: pfrag->page,
4188 off: pfrag->offset, size: space);
4189 page_ref_inc(page: pfrag->page);
4190 pfrag->offset += space;
4191 skb_len_add(skb: syn_data, delta: space);
4192 skb_zcopy_set(skb: syn_data, uarg: fo->uarg, NULL);
4193 }
4194 /* No more data pending in inet_wait_for_connect() */
4195 if (space == fo->size)
4196 fo->data = NULL;
4197 fo->copied = space;
4198
4199 tcp_connect_queue_skb(sk, skb: syn_data);
4200 if (syn_data->len)
4201 tcp_chrono_start(sk, type: TCP_CHRONO_BUSY);
4202
4203 err = tcp_transmit_skb(sk, skb: syn_data, clone_it: 1, gfp_mask: sk->sk_allocation);
4204
4205 skb_set_delivery_time(skb: syn, kt: syn_data->skb_mstamp_ns, tstamp_type: SKB_CLOCK_MONOTONIC);
4206
4207 /* Now full SYN+DATA was cloned and sent (or not),
4208 * remove the SYN from the original skb (syn_data)
4209 * we keep in write queue in case of a retransmit, as we
4210 * also have the SYN packet (with no data) in the same queue.
4211 */
4212 TCP_SKB_CB(syn_data)->seq++;
4213 TCP_SKB_CB(syn_data)->tcp_flags = TCPHDR_ACK | TCPHDR_PSH;
4214 if (!err) {
4215 tp->syn_data = (fo->copied > 0);
4216 tcp_rbtree_insert(root: &sk->tcp_rtx_queue, skb: syn_data);
4217 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPORIGDATASENT);
4218 goto done;
4219 }
4220
4221 /* data was not sent, put it in write_queue */
4222 __skb_queue_tail(list: &sk->sk_write_queue, newsk: syn_data);
4223 tp->packets_out -= tcp_skb_pcount(skb: syn_data);
4224
4225fallback:
4226 /* Send a regular SYN with Fast Open cookie request option */
4227 if (fo->cookie.len > 0)
4228 fo->cookie.len = 0;
4229 err = tcp_transmit_skb(sk, skb: syn, clone_it: 1, gfp_mask: sk->sk_allocation);
4230 if (err)
4231 tp->syn_fastopen = 0;
4232done:
4233 fo->cookie.len = -1; /* Exclude Fast Open option for SYN retries */
4234 return err;
4235}
4236
4237/* Build a SYN and send it off. */
4238int tcp_connect(struct sock *sk)
4239{
4240 struct tcp_sock *tp = tcp_sk(sk);
4241 struct sk_buff *buff;
4242 int err;
4243
4244 tcp_call_bpf(sk, op: BPF_SOCK_OPS_TCP_CONNECT_CB, nargs: 0, NULL);
4245
4246#if defined(CONFIG_TCP_MD5SIG) && defined(CONFIG_TCP_AO)
4247 /* Has to be checked late, after setting daddr/saddr/ops.
4248 * Return error if the peer has both a md5 and a tcp-ao key
4249 * configured as this is ambiguous.
4250 */
4251 if (unlikely(rcu_dereference_protected(tp->md5sig_info,
4252 lockdep_sock_is_held(sk)))) {
4253 bool needs_ao = !!tp->af_specific->ao_lookup(sk, sk, -1, -1);
4254 bool needs_md5 = !!tp->af_specific->md5_lookup(sk, sk);
4255 struct tcp_ao_info *ao_info;
4256
4257 ao_info = rcu_dereference_check(tp->ao_info,
4258 lockdep_sock_is_held(sk));
4259 if (ao_info) {
4260 /* This is an extra check: tcp_ao_required() in
4261 * tcp_v{4,6}_parse_md5_keys() should prevent adding
4262 * md5 keys on ao_required socket.
4263 */
4264 needs_ao |= ao_info->ao_required;
4265 WARN_ON_ONCE(ao_info->ao_required && needs_md5);
4266 }
4267 if (needs_md5 && needs_ao)
4268 return -EKEYREJECTED;
4269
4270 /* If we have a matching md5 key and no matching tcp-ao key
4271 * then free up ao_info if allocated.
4272 */
4273 if (needs_md5) {
4274 tcp_ao_destroy_sock(sk, false);
4275 } else if (needs_ao) {
4276 tcp_clear_md5_list(sk);
4277 kfree(rcu_replace_pointer(tp->md5sig_info, NULL,
4278 lockdep_sock_is_held(sk)));
4279 }
4280 }
4281#endif
4282#ifdef CONFIG_TCP_AO
4283 if (unlikely(rcu_dereference_protected(tp->ao_info,
4284 lockdep_sock_is_held(sk)))) {
4285 /* Don't allow connecting if ao is configured but no
4286 * matching key is found.
4287 */
4288 if (!tp->af_specific->ao_lookup(sk, sk, -1, -1))
4289 return -EKEYREJECTED;
4290 }
4291#endif
4292
4293 if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk))
4294 return -EHOSTUNREACH; /* Routing failure or similar. */
4295
4296 tcp_connect_init(sk);
4297
4298 if (unlikely(tp->repair)) {
4299 tcp_finish_connect(sk, NULL);
4300 return 0;
4301 }
4302
4303 buff = tcp_stream_alloc_skb(sk, gfp: sk->sk_allocation, force_schedule: true);
4304 if (unlikely(!buff))
4305 return -ENOBUFS;
4306
4307 /* SYN eats a sequence byte, write_seq updated by
4308 * tcp_connect_queue_skb().
4309 */
4310 tcp_init_nondata_skb(skb: buff, sk, seq: tp->write_seq, TCPHDR_SYN);
4311 tcp_mstamp_refresh(tp);
4312 tp->retrans_stamp = tcp_time_stamp_ts(tp);
4313 tcp_connect_queue_skb(sk, skb: buff);
4314 tcp_ecn_send_syn(sk, skb: buff);
4315 tcp_rbtree_insert(root: &sk->tcp_rtx_queue, skb: buff);
4316
4317 /* Send off SYN; include data in Fast Open. */
4318 err = tp->fastopen_req ? tcp_send_syn_data(sk, syn: buff) :
4319 tcp_transmit_skb(sk, skb: buff, clone_it: 1, gfp_mask: sk->sk_allocation);
4320 if (err == -ECONNREFUSED)
4321 return err;
4322
4323 /* We change tp->snd_nxt after the tcp_transmit_skb() call
4324 * in order to make this packet get counted in tcpOutSegs.
4325 */
4326 WRITE_ONCE(tp->snd_nxt, tp->write_seq);
4327 tp->pushed_seq = tp->write_seq;
4328 buff = tcp_send_head(sk);
4329 if (unlikely(buff)) {
4330 WRITE_ONCE(tp->snd_nxt, TCP_SKB_CB(buff)->seq);
4331 tp->pushed_seq = TCP_SKB_CB(buff)->seq;
4332 }
4333 TCP_INC_STATS(sock_net(sk), TCP_MIB_ACTIVEOPENS);
4334
4335 /* Timer for repeating the SYN until an answer. */
4336 tcp_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
4337 inet_csk(sk)->icsk_rto, pace_delay: false);
4338 return 0;
4339}
4340EXPORT_SYMBOL(tcp_connect);
4341
4342u32 tcp_delack_max(const struct sock *sk)
4343{
4344 u32 delack_from_rto_min = max(tcp_rto_min(sk), 2) - 1;
4345
4346 return min(READ_ONCE(inet_csk(sk)->icsk_delack_max), delack_from_rto_min);
4347}
4348
4349/* Send out a delayed ack, the caller does the policy checking
4350 * to see if we should even be here. See tcp_input.c:tcp_ack_snd_check()
4351 * for details.
4352 */
4353void tcp_send_delayed_ack(struct sock *sk)
4354{
4355 struct inet_connection_sock *icsk = inet_csk(sk);
4356 int ato = icsk->icsk_ack.ato;
4357 unsigned long timeout;
4358
4359 if (ato > TCP_DELACK_MIN) {
4360 const struct tcp_sock *tp = tcp_sk(sk);
4361 int max_ato = HZ / 2;
4362
4363 if (inet_csk_in_pingpong_mode(sk) ||
4364 (icsk->icsk_ack.pending & ICSK_ACK_PUSHED))
4365 max_ato = TCP_DELACK_MAX;
4366
4367 /* Slow path, intersegment interval is "high". */
4368
4369 /* If some rtt estimate is known, use it to bound delayed ack.
4370 * Do not use inet_csk(sk)->icsk_rto here, use results of rtt measurements
4371 * directly.
4372 */
4373 if (tp->srtt_us) {
4374 int rtt = max_t(int, usecs_to_jiffies(tp->srtt_us >> 3),
4375 TCP_DELACK_MIN);
4376
4377 if (rtt < max_ato)
4378 max_ato = rtt;
4379 }
4380
4381 ato = min(ato, max_ato);
4382 }
4383
4384 ato = min_t(u32, ato, tcp_delack_max(sk));
4385
4386 /* Stay within the limit we were given */
4387 timeout = jiffies + ato;
4388
4389 /* Use new timeout only if there wasn't a older one earlier. */
4390 if (icsk->icsk_ack.pending & ICSK_ACK_TIMER) {
4391 /* If delack timer is about to expire, send ACK now. */
4392 if (time_before_eq(icsk_delack_timeout(icsk), jiffies + (ato >> 2))) {
4393 tcp_send_ack(sk);
4394 return;
4395 }
4396
4397 if (!time_before(timeout, icsk_delack_timeout(icsk)))
4398 timeout = icsk_delack_timeout(icsk);
4399 }
4400 smp_store_release(&icsk->icsk_ack.pending,
4401 icsk->icsk_ack.pending | ICSK_ACK_SCHED | ICSK_ACK_TIMER);
4402 sk_reset_timer(sk, timer: &icsk->icsk_delack_timer, expires: timeout);
4403}
4404
4405/* This routine sends an ack and also updates the window. */
4406void __tcp_send_ack(struct sock *sk, u32 rcv_nxt, u16 flags)
4407{
4408 struct sk_buff *buff;
4409
4410 /* If we have been reset, we may not send again. */
4411 if (sk->sk_state == TCP_CLOSE)
4412 return;
4413
4414 /* We are not putting this on the write queue, so
4415 * tcp_transmit_skb() will set the ownership to this
4416 * sock.
4417 */
4418 buff = alloc_skb(MAX_TCP_HEADER,
4419 priority: sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4420 if (unlikely(!buff)) {
4421 struct inet_connection_sock *icsk = inet_csk(sk);
4422 unsigned long delay;
4423
4424 delay = TCP_DELACK_MAX << icsk->icsk_ack.retry;
4425 if (delay < tcp_rto_max(sk))
4426 icsk->icsk_ack.retry++;
4427 inet_csk_schedule_ack(sk);
4428 icsk->icsk_ack.ato = TCP_ATO_MIN;
4429 tcp_reset_xmit_timer(sk, ICSK_TIME_DACK, when: delay, pace_delay: false);
4430 return;
4431 }
4432
4433 /* Reserve space for headers and prepare control bits. */
4434 skb_reserve(skb: buff, MAX_TCP_HEADER);
4435 tcp_init_nondata_skb(skb: buff, sk,
4436 seq: tcp_acceptable_seq(sk), TCPHDR_ACK | flags);
4437
4438 /* We do not want pure acks influencing TCP Small Queues or fq/pacing
4439 * too much.
4440 * SKB_TRUESIZE(max(1 .. 66, MAX_TCP_HEADER)) is unfortunately ~784
4441 */
4442 skb_set_tcp_pure_ack(skb: buff);
4443
4444 /* Send it off, this clears delayed acks for us. */
4445 __tcp_transmit_skb(sk, skb: buff, clone_it: 0, gfp_mask: (__force gfp_t)0, rcv_nxt);
4446}
4447EXPORT_SYMBOL_GPL(__tcp_send_ack);
4448
4449void tcp_send_ack(struct sock *sk)
4450{
4451 __tcp_send_ack(sk, tcp_sk(sk)->rcv_nxt, 0);
4452}
4453
4454/* This routine sends a packet with an out of date sequence
4455 * number. It assumes the other end will try to ack it.
4456 *
4457 * Question: what should we make while urgent mode?
4458 * 4.4BSD forces sending single byte of data. We cannot send
4459 * out of window data, because we have SND.NXT==SND.MAX...
4460 *
4461 * Current solution: to send TWO zero-length segments in urgent mode:
4462 * one is with SEG.SEQ=SND.UNA to deliver urgent pointer, another is
4463 * out-of-date with SND.UNA-1 to probe window.
4464 */
4465static int tcp_xmit_probe_skb(struct sock *sk, int urgent, int mib)
4466{
4467 struct tcp_sock *tp = tcp_sk(sk);
4468 struct sk_buff *skb;
4469
4470 /* We don't queue it, tcp_transmit_skb() sets ownership. */
4471 skb = alloc_skb(MAX_TCP_HEADER,
4472 priority: sk_gfp_mask(sk, GFP_ATOMIC | __GFP_NOWARN));
4473 if (!skb)
4474 return -1;
4475
4476 /* Reserve space for headers and set control bits. */
4477 skb_reserve(skb, MAX_TCP_HEADER);
4478 /* Use a previous sequence. This should cause the other
4479 * end to send an ack. Don't queue or clone SKB, just
4480 * send it.
4481 */
4482 tcp_init_nondata_skb(skb, sk, seq: tp->snd_una - !urgent, TCPHDR_ACK);
4483 NET_INC_STATS(sock_net(sk), mib);
4484 return tcp_transmit_skb(sk, skb, clone_it: 0, gfp_mask: (__force gfp_t)0);
4485}
4486
4487/* Called from setsockopt( ... TCP_REPAIR ) */
4488void tcp_send_window_probe(struct sock *sk)
4489{
4490 if (sk->sk_state == TCP_ESTABLISHED) {
4491 tcp_sk(sk)->snd_wl1 = tcp_sk(sk)->rcv_nxt - 1;
4492 tcp_mstamp_refresh(tcp_sk(sk));
4493 tcp_xmit_probe_skb(sk, urgent: 0, mib: LINUX_MIB_TCPWINPROBE);
4494 }
4495}
4496
4497/* Initiate keepalive or window probe from timer. */
4498int tcp_write_wakeup(struct sock *sk, int mib)
4499{
4500 struct tcp_sock *tp = tcp_sk(sk);
4501 struct sk_buff *skb;
4502
4503 if (sk->sk_state == TCP_CLOSE)
4504 return -1;
4505
4506 skb = tcp_send_head(sk);
4507 if (skb && before(TCP_SKB_CB(skb)->seq, seq2: tcp_wnd_end(tp))) {
4508 int err;
4509 unsigned int mss = tcp_current_mss(sk);
4510 unsigned int seg_size = tcp_wnd_end(tp) - TCP_SKB_CB(skb)->seq;
4511
4512 if (before(seq1: tp->pushed_seq, TCP_SKB_CB(skb)->end_seq))
4513 tp->pushed_seq = TCP_SKB_CB(skb)->end_seq;
4514
4515 /* We are probing the opening of a window
4516 * but the window size is != 0
4517 * must have been a result SWS avoidance ( sender )
4518 */
4519 if (seg_size < TCP_SKB_CB(skb)->end_seq - TCP_SKB_CB(skb)->seq ||
4520 skb->len > mss) {
4521 seg_size = min(seg_size, mss);
4522 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4523 if (tcp_fragment(sk, tcp_queue: TCP_FRAG_IN_WRITE_QUEUE,
4524 skb, len: seg_size, mss_now: mss, GFP_ATOMIC))
4525 return -1;
4526 } else if (!tcp_skb_pcount(skb))
4527 tcp_set_skb_tso_segs(skb, mss_now: mss);
4528
4529 TCP_SKB_CB(skb)->tcp_flags |= TCPHDR_PSH;
4530 err = tcp_transmit_skb(sk, skb, clone_it: 1, GFP_ATOMIC);
4531 if (!err)
4532 tcp_event_new_data_sent(sk, skb);
4533 return err;
4534 } else {
4535 if (between(seq1: tp->snd_up, seq2: tp->snd_una + 1, seq3: tp->snd_una + 0xFFFF))
4536 tcp_xmit_probe_skb(sk, urgent: 1, mib);
4537 return tcp_xmit_probe_skb(sk, urgent: 0, mib);
4538 }
4539}
4540
4541/* A window probe timeout has occurred. If window is not closed send
4542 * a partial packet else a zero probe.
4543 */
4544void tcp_send_probe0(struct sock *sk)
4545{
4546 struct inet_connection_sock *icsk = inet_csk(sk);
4547 struct tcp_sock *tp = tcp_sk(sk);
4548 struct net *net = sock_net(sk);
4549 unsigned long timeout;
4550 int err;
4551
4552 err = tcp_write_wakeup(sk, mib: LINUX_MIB_TCPWINPROBE);
4553
4554 if (tp->packets_out || tcp_write_queue_empty(sk)) {
4555 /* Cancel probe timer, if it is not required. */
4556 WRITE_ONCE(icsk->icsk_probes_out, 0);
4557 icsk->icsk_backoff = 0;
4558 icsk->icsk_probes_tstamp = 0;
4559 return;
4560 }
4561
4562 WRITE_ONCE(icsk->icsk_probes_out, icsk->icsk_probes_out + 1);
4563 if (err <= 0) {
4564 if (icsk->icsk_backoff < READ_ONCE(net->ipv4.sysctl_tcp_retries2))
4565 icsk->icsk_backoff++;
4566 timeout = tcp_probe0_when(sk, max_when: tcp_rto_max(sk));
4567 } else {
4568 /* If packet was not sent due to local congestion,
4569 * Let senders fight for local resources conservatively.
4570 */
4571 timeout = TCP_RESOURCE_PROBE_INTERVAL;
4572 }
4573
4574 timeout = tcp_clamp_probe0_to_user_timeout(sk, when: timeout);
4575 tcp_reset_xmit_timer(sk, ICSK_TIME_PROBE0, when: timeout, pace_delay: true);
4576}
4577
4578int tcp_rtx_synack(const struct sock *sk, struct request_sock *req)
4579{
4580 const struct tcp_request_sock_ops *af_ops = tcp_rsk(req)->af_specific;
4581 struct flowi fl;
4582 int res;
4583
4584 /* Paired with WRITE_ONCE() in sock_setsockopt() */
4585 if (READ_ONCE(sk->sk_txrehash) == SOCK_TXREHASH_ENABLED)
4586 WRITE_ONCE(tcp_rsk(req)->txhash, net_tx_rndhash());
4587 res = af_ops->send_synack(sk, NULL, &fl, req, NULL, TCP_SYNACK_NORMAL,
4588 NULL);
4589 if (!res) {
4590 TCP_INC_STATS(sock_net(sk), TCP_MIB_RETRANSSEGS);
4591 NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPSYNRETRANS);
4592 if (unlikely(tcp_passive_fastopen(sk))) {
4593 /* sk has const attribute because listeners are lockless.
4594 * However in this case, we are dealing with a passive fastopen
4595 * socket thus we can change total_retrans value.
4596 */
4597 tcp_sk_rw(sk)->total_retrans++;
4598 }
4599 trace_tcp_retransmit_synack(sk, req);
4600 WRITE_ONCE(req->num_retrans, req->num_retrans + 1);
4601 }
4602 return res;
4603}
4604EXPORT_IPV6_MOD(tcp_rtx_synack);
4605